[PATCH v11 00/26] Speculative page faults

[PATCH v11 00/26] Speculative page faults

[Vinayak Menon]

Hi Laurent,

We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).

Steps:

1) Run the program

2) The 2 threads will try lock/unlock and prints messages, and main thread does fork. In around 1 minute time, lock/unlock threads will hang on pthread_mutex_lock.


Initially the issue was reported when ptrace was tried on apps. Later it was discovered that the write protect done by fork is causing the issue and the below test code was created.


Observations:

1) We have tried disabling SPF and the issue disappears.

2) Adding this piece of code in __handle_speculative_fault also fixes the problem.

if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))

    return VM_FAULT_RETRY;

3) As an experiment we tried encapsulating handle_speculative_fault with down_read(mmap_sem) and that too fixes the problem.

4) It is observed that while in wp_page_copy, the contents of the old_page changes which should not ideally happen as the pte is !pte_write.

5) To prove that it is a race, we tried affining the threads to single core, and the issue disappears.

Let us know if you want us to try out any experiments.

Thanks,
Vinayak

/**test.c***/

#include <stdio.h>
#include <pthread.h>
#include <unistd.h>
#include <string.h>
#include <stdlib.h>

#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <errno.h>

#define UNUSED(expr) do { (void)(expr); } while(0)
#define print printf

pthread_mutex_t m;
pthread_cond_t c;

static void* cond_do(void* arg)
{
  UNUSED(arg);
  do {
    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
    pthread_mutex_lock(&m);
 
    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
    pthread_mutex_unlock(&m);
    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
  } while (true);

  return NULL;
}

static void* sig_do(void* arg)
{
  UNUSED(arg);
  do {
    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
    pthread_mutex_lock(&m);

    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
    pthread_mutex_unlock(&m);
    print("%s:%d state=%u addr: %lx\n", __func__, __LINE__, *(uint32_t*)(&m), (unsigned long)&m);
  } while (true);
  return NULL;
}

int main()
{
  pthread_t sig;
  pthread_t cond;
  pthread_mutex_init(&m, NULL);
 
  pthread_create(&cond, NULL, cond_do, NULL);
  pthread_create(&sig, NULL, sig_do, NULL);

  while(1) {
      if (!fork()) {
          usleep(500);
          abort();
      }
      usleep(550);
  }

  pthread_join(sig, NULL);
  pthread_join(cond, NULL);

  return 0;
}

Re: [PATCH v11 00/26] Speculative page faults

[Vinayak Menon]

On 1/11/2019 9:13 PM, Vinayak Menon wrote:
> Hi Laurent,
>
> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).


With the patch below, we don't hit the issue.

From: Vinayak Menon <vinmenon@codeaurora.org>
Date: Mon, 14 Jan 2019 16:06:34 +0530
Subject: [PATCH] mm: flush stale tlb entries on speculative write fault

It is observed that the following scenario results in
threads A and B of process 1 blocking on pthread_mutex_lock
forever after few iterations.

CPU 1                   CPU 2                    CPU 3
Process 1,              Process 1,               Process 1,
Thread A                Thread B                 Thread C

while (1) {             while (1) {              while(1) {
pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
}                       }

When from thread C, copy_one_pte write-protects the parent pte
(of lock l), stale tlb entries can exist with write permissions
on one of the CPUs at least. This can create a problem if one
of the threads A or B hits the write fault. Though dup_mmap calls
flush_tlb_mm after copy_page_range, since speculative page fault
does not take mmap_sem it can proceed further fixing a fault soon
after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
entry can still modify old_page even after it is copied to
new_page by wp_page_copy, thus causing a corruption.

Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
---
 mm/memory.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 52080e4..1ea168ff 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
                return VM_FAULT_RETRY;
        }

+       /*
+        * Discard tlb entries created before ptep_set_wrprotect
+        * in copy_one_pte
+        */
+       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
+               flush_tlb_page(vmf.vma, address);
+
        mem_cgroup_oom_enable();
        ret = handle_pte_fault(&vmf);
        mem_cgroup_oom_disable();
--
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

[-- Attachment #1: Type: text/plain, Size: 3581 bytes --]

Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>> Hi Laurent,
>>
>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
> 
> 
> With the patch below, we don't hit the issue.
> 
> From: Vinayak Menon <vinmenon@codeaurora.org>
> Date: Mon, 14 Jan 2019 16:06:34 +0530
> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
> 
> It is observed that the following scenario results in
> threads A and B of process 1 blocking on pthread_mutex_lock
> forever after few iterations.
> 
> CPU 1                   CPU 2                    CPU 3
> Process 1,              Process 1,               Process 1,
> Thread A                Thread B                 Thread C
> 
> while (1) {             while (1) {              while(1) {
> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
> }                       }
> 
> When from thread C, copy_one_pte write-protects the parent pte
> (of lock l), stale tlb entries can exist with write permissions
> on one of the CPUs at least. This can create a problem if one
> of the threads A or B hits the write fault. Though dup_mmap calls
> flush_tlb_mm after copy_page_range, since speculative page fault
> does not take mmap_sem it can proceed further fixing a fault soon
> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
> entry can still modify old_page even after it is copied to
> new_page by wp_page_copy, thus causing a corruption.

Nice catch and thanks for your investigation!

There is a real synchronization issue here between copy_page_range() and 
the speculative page fault handler. I didn't get it on PowerVM since the 
TLB are flushed when arch_exit_lazy_mode() is called in 
copy_page_range() but now, I can get it when running on x86_64.

> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
> ---
>   mm/memory.c | 7 +++++++
>   1 file changed, 7 insertions(+)
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 52080e4..1ea168ff 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>                  return VM_FAULT_RETRY;
>          }
> 
> +       /*
> +        * Discard tlb entries created before ptep_set_wrprotect
> +        * in copy_one_pte
> +        */
> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
> +               flush_tlb_page(vmf.vma, address);
> +
>          mem_cgroup_oom_enable();
>          ret = handle_pte_fault(&vmf);
>          mem_cgroup_oom_disable();

Your patch is fixing the race but I'm wondering about the cost of these 
tlb flushes. Here we are flushing on a per page basis (architecture like 
x86_64 are smarter and flush more pages) but there is a request to flush 
a range of tlb entries each time a cow page is newly touched. I think 
there could be some bad impact here.

Another option would be to flush the range in copy_pte_range() before 
unlocking the page table lock. This will flush entries flush_tlb_mm() 
would later handle in dup_mmap() but that will be called once per fork 
per cow VMA.

I tried the attached patch which seems to fix the issue on x86_64. Could 
you please give it a try on arm64 ?

Thanks,
Laurent.


[-- Attachment #2: 0001-mm-flush-TLB-once-pages-are-copied-when-SPF-is-on.patch --]
[-- Type: text/plain, Size: 2137 bytes --]

From 9847338187c5c7e2d387d14765452d00fa60981e Mon Sep 17 00:00:00 2001
From: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Date: Mon, 14 Jan 2019 18:35:39 +0100
Subject: [PATCH] mm: flush TLB once pages are copied when SPF is on

Vinayak Menon reported that the following scenario results in
threads A and B of process 1 blocking on pthread_mutex_lock
forever after few iterations.

CPU 1                   CPU 2                    CPU 3
Process 1,              Process 1,               Process 1,
Thread A                Thread B                 Thread C

while (1) {             while (1) {              while(1) {
pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
}                       }

When from thread C, copy_one_pte write-protects the parent pte
(of lock l), stale tlb entries can exist with write permissions
on one of the CPUs at least. This can create a problem if one
of the threads A or B hits the write fault. Though dup_mmap calls
flush_tlb_mm after copy_page_range, since speculative page fault
does not take mmap_sem it can proceed further fixing a fault soon
after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
entry can still modify old_page even after it is copied to
new_page by wp_page_copy, thus causing a corruption.

Reported-by: Vinayak Menon <vinmenon@codeaurora.org>
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 mm/memory.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 48e1cf0a54ef..b7501294e0a0 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1112,6 +1112,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 
 	arch_leave_lazy_mmu_mode();
+
+	/*
+	 * Prevent the page fault handler to copy the page while stale tlb entry
+	 * are still not flushed.
+	 */
+	if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
+	    is_cow_mapping(vma->vm_flags))
+		flush_tlb_range(vma, addr, end);
+
 	spin_unlock(src_ptl);
 	pte_unmap(orig_src_pte);
 	add_mm_rss_vec(dst_mm, rss);
-- 
2.20.1

Re: [PATCH v11 00/26] Speculative page faults

[Vinayak Menon]

On 1/15/2019 1:54 PM, Laurent Dufour wrote:
> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>> Hi Laurent,
>>>
>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>
>>
>> With the patch below, we don't hit the issue.
>>
>> From: Vinayak Menon <vinmenon@codeaurora.org>
>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>
>> It is observed that the following scenario results in
>> threads A and B of process 1 blocking on pthread_mutex_lock
>> forever after few iterations.
>>
>> CPU 1                   CPU 2                    CPU 3
>> Process 1,              Process 1,               Process 1,
>> Thread A                Thread B                 Thread C
>>
>> while (1) {             while (1) {              while(1) {
>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>> }                       }
>>
>> When from thread C, copy_one_pte write-protects the parent pte
>> (of lock l), stale tlb entries can exist with write permissions
>> on one of the CPUs at least. This can create a problem if one
>> of the threads A or B hits the write fault. Though dup_mmap calls
>> flush_tlb_mm after copy_page_range, since speculative page fault
>> does not take mmap_sem it can proceed further fixing a fault soon
>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>> entry can still modify old_page even after it is copied to
>> new_page by wp_page_copy, thus causing a corruption.
>
> Nice catch and thanks for your investigation!
>
> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>
>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>> ---
>>   mm/memory.c | 7 +++++++
>>   1 file changed, 7 insertions(+)
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 52080e4..1ea168ff 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>                  return VM_FAULT_RETRY;
>>          }
>>
>> +       /*
>> +        * Discard tlb entries created before ptep_set_wrprotect
>> +        * in copy_one_pte
>> +        */
>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>> +               flush_tlb_page(vmf.vma, address);
>> +
>>          mem_cgroup_oom_enable();
>>          ret = handle_pte_fault(&vmf);
>>          mem_cgroup_oom_disable();
>
> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>
> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.


But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?


>
> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>

Your patch works fine on arm64 with a minor change. Thanks Laurent.

diff --git a/mm/memory.c b/mm/memory.c
index 52080e4..4767095 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1087,6 +1087,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
        spinlock_t *src_ptl, *dst_ptl;
        int progress = 0;
        int rss[NR_MM_COUNTERS];
+       unsigned long orig_addr = addr;
        swp_entry_t entry = (swp_entry_t){0};

 again:
@@ -1125,6 +1126,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);

        arch_leave_lazy_mmu_mode();
+
+       /*
+        * Prevent the page fault handler to copy the page while stale tlb entry
+        * are still not flushed.
+        */
+       if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
+               is_cow_mapping(vma->vm_flags))
+                       flush_tlb_range(vma, orig_addr, end);
+
        spin_unlock(src_ptl);
        pte_unmap(orig_src_pte);
        add_mm_rss_vec(dst_mm, rss);

Thanks,

Vinayak

Re: [PATCH v11 00/26] Speculative page faults

[Vinayak Menon]

On 1/15/2019 1:54 PM, Laurent Dufour wrote:
> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>> Hi Laurent,
>>>
>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>
>>
>> With the patch below, we don't hit the issue.
>>
>> From: Vinayak Menon <vinmenon@codeaurora.org>
>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>
>> It is observed that the following scenario results in
>> threads A and B of process 1 blocking on pthread_mutex_lock
>> forever after few iterations.
>>
>> CPU 1                   CPU 2                    CPU 3
>> Process 1,              Process 1,               Process 1,
>> Thread A                Thread B                 Thread C
>>
>> while (1) {             while (1) {              while(1) {
>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>> }                       }
>>
>> When from thread C, copy_one_pte write-protects the parent pte
>> (of lock l), stale tlb entries can exist with write permissions
>> on one of the CPUs at least. This can create a problem if one
>> of the threads A or B hits the write fault. Though dup_mmap calls
>> flush_tlb_mm after copy_page_range, since speculative page fault
>> does not take mmap_sem it can proceed further fixing a fault soon
>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>> entry can still modify old_page even after it is copied to
>> new_page by wp_page_copy, thus causing a corruption.
>
> Nice catch and thanks for your investigation!
>
> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>
>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>> ---
>>   mm/memory.c | 7 +++++++
>>   1 file changed, 7 insertions(+)
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 52080e4..1ea168ff 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>                  return VM_FAULT_RETRY;
>>          }
>>
>> +       /*
>> +        * Discard tlb entries created before ptep_set_wrprotect
>> +        * in copy_one_pte
>> +        */
>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>> +               flush_tlb_page(vmf.vma, address);
>> +
>>          mem_cgroup_oom_enable();
>>          ret = handle_pte_fault(&vmf);
>>          mem_cgroup_oom_disable();
>
> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>
> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.


But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?


>
> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>

Your patch works fine on arm64 with a minor change. Thanks Laurent.

diff --git a/mm/memory.c b/mm/memory.c
index 52080e4..4767095 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1087,6 +1087,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
        spinlock_t *src_ptl, *dst_ptl;
        int progress = 0;
        int rss[NR_MM_COUNTERS];
+       unsigned long orig_addr = addr;
        swp_entry_t entry = (swp_entry_t){0};

 again:
@@ -1125,6 +1126,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);

        arch_leave_lazy_mmu_mode();
+
+       /*
+        * Prevent the page fault handler to copy the page while stale tlb entry
+        * are still not flushed.
+        */
+       if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
+               is_cow_mapping(vma->vm_flags))
+                       flush_tlb_range(vma, orig_addr, end);
+
        spin_unlock(src_ptl);
        pte_unmap(orig_src_pte);
        add_mm_rss_vec(dst_mm, rss);

Thanks,

Vinayak

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 16/01/2019 à 12:41, Vinayak Menon a écrit :
> 
> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>> Hi Laurent,
>>>>
>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>
>>>
>>> With the patch below, we don't hit the issue.
>>>
>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>
>>> It is observed that the following scenario results in
>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>> forever after few iterations.
>>>
>>> CPU 1                   CPU 2                    CPU 3
>>> Process 1,              Process 1,               Process 1,
>>> Thread A                Thread B                 Thread C
>>>
>>> while (1) {             while (1) {              while(1) {
>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>> }                       }
>>>
>>> When from thread C, copy_one_pte write-protects the parent pte
>>> (of lock l), stale tlb entries can exist with write permissions
>>> on one of the CPUs at least. This can create a problem if one
>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>> does not take mmap_sem it can proceed further fixing a fault soon
>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>> entry can still modify old_page even after it is copied to
>>> new_page by wp_page_copy, thus causing a corruption.
>>
>> Nice catch and thanks for your investigation!
>>
>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>
>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>> ---
>>>    mm/memory.c | 7 +++++++
>>>    1 file changed, 7 insertions(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 52080e4..1ea168ff 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>                   return VM_FAULT_RETRY;
>>>           }
>>>
>>> +       /*
>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>> +        * in copy_one_pte
>>> +        */
>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>> +               flush_tlb_page(vmf.vma, address);
>>> +
>>>           mem_cgroup_oom_enable();
>>>           ret = handle_pte_fault(&vmf);
>>>           mem_cgroup_oom_disable();
>>
>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>
>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
> 
> 
> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?

I think this should be less costly to do it per vma at the time of the 
fork instead of per page hit once the fork has been done, since this 
will happen in both the forked task and the forking one (the COW pages 
are concerning the both sides of the fork).

>>
>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>
> 
> Your patch works fine on arm64 with a minor change. Thanks Laurent.

Yup my mistake !
I tried to shrink the patch after testing it, sounds that I shrunk it 
far too much...

> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 52080e4..4767095 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1087,6 +1087,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>          spinlock_t *src_ptl, *dst_ptl;
>          int progress = 0;
>          int rss[NR_MM_COUNTERS];
> +       unsigned long orig_addr = addr;
>          swp_entry_t entry = (swp_entry_t){0};
> 
>   again:
> @@ -1125,6 +1126,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>          } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
> 
>          arch_leave_lazy_mmu_mode();
> +
> +       /*
> +        * Prevent the page fault handler to copy the page while stale tlb entry
> +        * are still not flushed.
> +        */
> +       if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
> +               is_cow_mapping(vma->vm_flags))
> +                       flush_tlb_range(vma, orig_addr, end);
> +
>          spin_unlock(src_ptl);
>          pte_unmap(orig_src_pte);
>          add_mm_rss_vec(dst_mm, rss);
> 
> Thanks,
> 
> Vinayak
> 

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/16 19:41, Vinayak Menon wrote:
> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>> Hi Laurent,
>>>>
>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>
>>> With the patch below, we don't hit the issue.
>>>
>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>
>>> It is observed that the following scenario results in
>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>> forever after few iterations.
>>>
>>> CPU 1                   CPU 2                    CPU 3
>>> Process 1,              Process 1,               Process 1,
>>> Thread A                Thread B                 Thread C
>>>
>>> while (1) {             while (1) {              while(1) {
>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>> }                       }
>>>
>>> When from thread C, copy_one_pte write-protects the parent pte
>>> (of lock l), stale tlb entries can exist with write permissions
>>> on one of the CPUs at least. This can create a problem if one
>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>> does not take mmap_sem it can proceed further fixing a fault soon
>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>> entry can still modify old_page even after it is copied to
>>> new_page by wp_page_copy, thus causing a corruption.
>> Nice catch and thanks for your investigation!
>>
>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>
>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>> ---
>>>   mm/memory.c | 7 +++++++
>>>   1 file changed, 7 insertions(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 52080e4..1ea168ff 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>                  return VM_FAULT_RETRY;
>>>          }
>>>
>>> +       /*
>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>> +        * in copy_one_pte
>>> +        */
>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>> +               flush_tlb_page(vmf.vma, address);
>>> +
>>>          mem_cgroup_oom_enable();
>>>          ret = handle_pte_fault(&vmf);
>>>          mem_cgroup_oom_disable();
>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>
>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>
> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>
>
>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>
> Your patch works fine on arm64 with a minor change. Thanks Laurent.
Hi, Vinayak and Laurent

I think the below change will impact the performance significantly. Becuase most of process has many
vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
call the flush_tlb_mm  later.

I think we can try the following way to do.

vm_write_begin(vma)
copy_pte_range
vm_write_end(vma)

The speculative page fault will return to grap the mmap_sem to run the nromal path.
Any thought?

Thanks,
zhong jiang
> diff --git a/mm/memory.c b/mm/memory.c
> index 52080e4..4767095 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1087,6 +1087,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>         spinlock_t *src_ptl, *dst_ptl;
>         int progress = 0;
>         int rss[NR_MM_COUNTERS];
> +       unsigned long orig_addr = addr;
>         swp_entry_t entry = (swp_entry_t){0};
>
>  again:
> @@ -1125,6 +1126,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
>
>         arch_leave_lazy_mmu_mode();
> +
> +       /*
> +        * Prevent the page fault handler to copy the page while stale tlb entry
> +        * are still not flushed.
> +        */
> +       if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
> +               is_cow_mapping(vma->vm_flags))
> +                       flush_tlb_range(vma, orig_addr, end);
> +
>         spin_unlock(src_ptl);
>         pte_unmap(orig_src_pte);
>         add_mm_rss_vec(dst_mm, rss);
>
> Thanks,
>
> Vinayak
>
>
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/16 19:41, Vinayak Menon wrote:
> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>> Hi Laurent,
>>>>
>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>
>>> With the patch below, we don't hit the issue.
>>>
>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>
>>> It is observed that the following scenario results in
>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>> forever after few iterations.
>>>
>>> CPU 1                   CPU 2                    CPU 3
>>> Process 1,              Process 1,               Process 1,
>>> Thread A                Thread B                 Thread C
>>>
>>> while (1) {             while (1) {              while(1) {
>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>> }                       }
>>>
>>> When from thread C, copy_one_pte write-protects the parent pte
>>> (of lock l), stale tlb entries can exist with write permissions
>>> on one of the CPUs at least. This can create a problem if one
>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>> does not take mmap_sem it can proceed further fixing a fault soon
>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>> entry can still modify old_page even after it is copied to
>>> new_page by wp_page_copy, thus causing a corruption.
>> Nice catch and thanks for your investigation!
>>
>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>
>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>> ---
>>>   mm/memory.c | 7 +++++++
>>>   1 file changed, 7 insertions(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 52080e4..1ea168ff 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>                  return VM_FAULT_RETRY;
>>>          }
>>>
>>> +       /*
>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>> +        * in copy_one_pte
>>> +        */
>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>> +               flush_tlb_page(vmf.vma, address);
>>> +
>>>          mem_cgroup_oom_enable();
>>>          ret = handle_pte_fault(&vmf);
>>>          mem_cgroup_oom_disable();
>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>
>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>
> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>
>
>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>
> Your patch works fine on arm64 with a minor change. Thanks Laurent.
Hi, Vinayak and Laurent

I think the below change will impact the performance significantly. Becuase most of process has many
vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
call the flush_tlb_mm  later.

I think we can try the following way to do.

vm_write_begin(vma)
copy_pte_range
vm_write_end(vma)

The speculative page fault will return to grap the mmap_sem to run the nromal path.
Any thought?

Thanks,
zhong jiang
> diff --git a/mm/memory.c b/mm/memory.c
> index 52080e4..4767095 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1087,6 +1087,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>         spinlock_t *src_ptl, *dst_ptl;
>         int progress = 0;
>         int rss[NR_MM_COUNTERS];
> +       unsigned long orig_addr = addr;
>         swp_entry_t entry = (swp_entry_t){0};
>
>  again:
> @@ -1125,6 +1126,15 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
>
>         arch_leave_lazy_mmu_mode();
> +
> +       /*
> +        * Prevent the page fault handler to copy the page while stale tlb entry
> +        * are still not flushed.
> +        */
> +       if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT) &&
> +               is_cow_mapping(vma->vm_flags))
> +                       flush_tlb_range(vma, orig_addr, end);
> +
>         spin_unlock(src_ptl);
>         pte_unmap(orig_src_pte);
>         add_mm_rss_vec(dst_mm, rss);
>
> Thanks,
>
> Vinayak
>
>
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 17/01/2019 à 16:51, zhong jiang a écrit :
> On 2019/1/16 19:41, Vinayak Menon wrote:
>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>
>>>> With the patch below, we don't hit the issue.
>>>>
>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>
>>>> It is observed that the following scenario results in
>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>> forever after few iterations.
>>>>
>>>> CPU 1                   CPU 2                    CPU 3
>>>> Process 1,              Process 1,               Process 1,
>>>> Thread A                Thread B                 Thread C
>>>>
>>>> while (1) {             while (1) {              while(1) {
>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>> }                       }
>>>>
>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>> (of lock l), stale tlb entries can exist with write permissions
>>>> on one of the CPUs at least. This can create a problem if one
>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>> entry can still modify old_page even after it is copied to
>>>> new_page by wp_page_copy, thus causing a corruption.
>>> Nice catch and thanks for your investigation!
>>>
>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>
>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>> ---
>>>>    mm/memory.c | 7 +++++++
>>>>    1 file changed, 7 insertions(+)
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 52080e4..1ea168ff 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>                   return VM_FAULT_RETRY;
>>>>           }
>>>>
>>>> +       /*
>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>> +        * in copy_one_pte
>>>> +        */
>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>> +               flush_tlb_page(vmf.vma, address);
>>>> +
>>>>           mem_cgroup_oom_enable();
>>>>           ret = handle_pte_fault(&vmf);
>>>>           mem_cgroup_oom_disable();
>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>
>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>
>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>
>>
>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>
>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
> Hi, Vinayak and Laurent
> 
> I think the below change will impact the performance significantly. Becuase most of process has many
> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
> call the flush_tlb_mm  later.
> 
> I think we can try the following way to do.
> 
> vm_write_begin(vma)
> copy_pte_range
> vm_write_end(vma)
> 
> The speculative page fault will return to grap the mmap_sem to run the nromal path.
> Any thought?

Hi Zhong,

I agree that flushing the TLB could have a bad impact on the 
performance, but tagging the VMA when copy_pte_range() is not fixing the 
issue as the VMA must be flagged until the PTE are flushed.

Here is what happens:

CPU A                CPU B                       CPU C
fork()
copy_pte_range()
   set PTE rdonly
got to next VMA...			
  .                   PTE is seen rdonly	         PTE still writable
  .                   thread is writing to page
  .                   -> page fault
  .                     copy the page             Thread writes to page
  .                      .                        -> no page fault
  .                     update the PTE
  .                     flush TLB for that PTE
flush TLB                                        PTE are now rdonly

So the write done by the CPU C is interfering with the page copy 
operation done by CPU B, leading to the data corruption.

Flushing the PTE in copy_pte_range() is fixing the issue as the CPU C is 
seeing the PTE as rdonly earlier. But this impacts performance.

Another option, I'll work on is to flag _all the COW eligible_ VMA 
before starting copying them and until the PTE are flushed on the CPU A.
This way when the CPU B will page fault the speculative handler will 
abort because the VMA is in the way to be touched.

But I need to ensure that all the calls to copy_pte_range() are handling 
this correctly.

Laurent.

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/18 17:29, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Hi Zhong,
>
> I agree that flushing the TLB could have a bad impact on the performance, but tagging the VMA when copy_pte_range() is not fixing the issue as the VMA must be flagged until the PTE are flushed.
>
> Here is what happens:
>
> CPU A                CPU B                       CPU C
> fork()
> copy_pte_range()
>   set PTE rdonly
> got to next VMA...           
>  .                   PTE is seen rdonly             PTE still writable
>  .                   thread is writing to page
>  .                   -> page fault
>  .                     copy the page             Thread writes to page
>  .                      .                        -> no page fault
>  .                     update the PTE
>  .                     flush TLB for that PTE
> flush TLB                                        PTE are now rdonly  
>
> So the write done by the CPU C is interfering with the page copy operation done by CPU B, leading to the data corruption.
>
I want to know the case if the CPU B has finished in front of the CPU C that the data still is vaild ?

This is to say, the old_page will be changed from other cpu because of the access from other cpu.

Maybe this is a stupid qestion :-)

Thanks,
zhong jiang.
> Flushing the PTE in copy_pte_range() is fixing the issue as the CPU C is seeing the PTE as rdonly earlier. But this impacts performance.
>
> Another option, I'll work on is to flag _all the COW eligible_ VMA before starting copying them and until the PTE are flushed on the CPU A.
> This way when the CPU B will page fault the speculative handler will abort because the VMA is in the way to be touched.
>
> But I need to ensure that all the calls to copy_pte_range() are handling this correctly.
>
> Laurent.
>
>
> .
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/18 17:29, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Hi Zhong,
>
> I agree that flushing the TLB could have a bad impact on the performance, but tagging the VMA when copy_pte_range() is not fixing the issue as the VMA must be flagged until the PTE are flushed.
>
> Here is what happens:
>
> CPU A                CPU B                       CPU C
> fork()
> copy_pte_range()
>   set PTE rdonly
> got to next VMA...           
>  .                   PTE is seen rdonly             PTE still writable
>  .                   thread is writing to page
>  .                   -> page fault
>  .                     copy the page             Thread writes to page
>  .                      .                        -> no page fault
>  .                     update the PTE
>  .                     flush TLB for that PTE
> flush TLB                                        PTE are now rdonly  
>
> So the write done by the CPU C is interfering with the page copy operation done by CPU B, leading to the data corruption.
>
I want to know the case if the CPU B has finished in front of the CPU C that the data still is vaild ?

This is to say, the old_page will be changed from other cpu because of the access from other cpu.

Maybe this is a stupid qestion :-)

Thanks,
zhong jiang.
> Flushing the PTE in copy_pte_range() is fixing the issue as the CPU C is seeing the PTE as rdonly earlier. But this impacts performance.
>
> Another option, I'll work on is to flag _all the COW eligible_ VMA before starting copying them and until the PTE are flushed on the CPU A.
> This way when the CPU B will page fault the speculative handler will abort because the VMA is in the way to be touched.
>
> But I need to ensure that all the calls to copy_pte_range() are handling this correctly.
>
> Laurent.
>
>
> .
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 18/01/2019 à 16:41, zhong jiang a écrit :
> On 2019/1/18 17:29, Laurent Dufour wrote:
>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>> Hi Laurent,
>>>>>>>
>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>
>>>>>> With the patch below, we don't hit the issue.
>>>>>>
>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>
>>>>>> It is observed that the following scenario results in
>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>> forever after few iterations.
>>>>>>
>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>> Process 1,              Process 1,               Process 1,
>>>>>> Thread A                Thread B                 Thread C
>>>>>>
>>>>>> while (1) {             while (1) {              while(1) {
>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>> }                       }
>>>>>>
>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>> entry can still modify old_page even after it is copied to
>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>> Nice catch and thanks for your investigation!
>>>>>
>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>
>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>> ---
>>>>>>     mm/memory.c | 7 +++++++
>>>>>>     1 file changed, 7 insertions(+)
>>>>>>
>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>> index 52080e4..1ea168ff 100644
>>>>>> --- a/mm/memory.c
>>>>>> +++ b/mm/memory.c
>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>                    return VM_FAULT_RETRY;
>>>>>>            }
>>>>>>
>>>>>> +       /*
>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>> +        * in copy_one_pte
>>>>>> +        */
>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>> +
>>>>>>            mem_cgroup_oom_enable();
>>>>>>            ret = handle_pte_fault(&vmf);
>>>>>>            mem_cgroup_oom_disable();
>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>
>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>
>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>
>>>>
>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>
>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>> Hi, Vinayak and Laurent
>>>
>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>> call the flush_tlb_mm  later.
>>>
>>> I think we can try the following way to do.
>>>
>>> vm_write_begin(vma)
>>> copy_pte_range
>>> vm_write_end(vma)
>>>
>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>> Any thought?
>>
>> Hi Zhong,
>>
>> I agree that flushing the TLB could have a bad impact on the performance, but tagging the VMA when copy_pte_range() is not fixing the issue as the VMA must be flagged until the PTE are flushed.
>>
>> Here is what happens:
>>
>> CPU A                CPU B                       CPU C
>> fork()
>> copy_pte_range()
>>    set PTE rdonly
>> got to next VMA...
>>   .                   PTE is seen rdonly             PTE still writable
>>   .                   thread is writing to page
>>   .                   -> page fault
>>   .                     copy the page             Thread writes to page
>>   .                      .                        -> no page fault
>>   .                     update the PTE
>>   .                     flush TLB for that PTE
>> flush TLB                                        PTE are now rdonly
>>
>> So the write done by the CPU C is interfering with the page copy operation done by CPU B, leading to the data corruption.
>>
> I want to know the case if the CPU B has finished in front of the CPU C that the data still is vaild ?

If the CPU B has done the flush TLB then the CPU C will write data to 
the right page. If the CPU B has not yet done the flush of the TLB as 
the time the CPU C is writing data, then this roughly the same issue.

Anyway this is fixed with the patch I'm about to sent for testing on arm64.

Cheers,
Laurent.

> 
> This is to say, the old_page will be changed from other cpu because of the access from other cpu.
> 
> Maybe this is a stupid qestion :-)
> 
> Thanks,
> zhong jiang.
>> Flushing the PTE in copy_pte_range() is fixing the issue as the CPU C is seeing the PTE as rdonly earlier. But this impacts performance.
>>
>> Another option, I'll work on is to flag _all the COW eligible_ VMA before starting copying them and until the PTE are flushed on the CPU A.
>> This way when the CPU B will page fault the speculative handler will abort because the VMA is in the way to be touched.
>>
>> But I need to ensure that all the calls to copy_pte_range() are handling this correctly.
>>
>> Laurent.
>>
>>
>> .
>>
> 
> 

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

[-- Attachment #1: Type: text/plain, Size: 4671 bytes --]

Le 17/01/2019 à 16:51, zhong jiang a écrit :
> On 2019/1/16 19:41, Vinayak Menon wrote:
>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>
>>>> With the patch below, we don't hit the issue.
>>>>
>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>
>>>> It is observed that the following scenario results in
>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>> forever after few iterations.
>>>>
>>>> CPU 1                   CPU 2                    CPU 3
>>>> Process 1,              Process 1,               Process 1,
>>>> Thread A                Thread B                 Thread C
>>>>
>>>> while (1) {             while (1) {              while(1) {
>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>> }                       }
>>>>
>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>> (of lock l), stale tlb entries can exist with write permissions
>>>> on one of the CPUs at least. This can create a problem if one
>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>> entry can still modify old_page even after it is copied to
>>>> new_page by wp_page_copy, thus causing a corruption.
>>> Nice catch and thanks for your investigation!
>>>
>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>
>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>> ---
>>>>    mm/memory.c | 7 +++++++
>>>>    1 file changed, 7 insertions(+)
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 52080e4..1ea168ff 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>                   return VM_FAULT_RETRY;
>>>>           }
>>>>
>>>> +       /*
>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>> +        * in copy_one_pte
>>>> +        */
>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>> +               flush_tlb_page(vmf.vma, address);
>>>> +
>>>>           mem_cgroup_oom_enable();
>>>>           ret = handle_pte_fault(&vmf);
>>>>           mem_cgroup_oom_disable();
>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>
>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>
>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>
>>
>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>
>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
> Hi, Vinayak and Laurent
> 
> I think the below change will impact the performance significantly. Becuase most of process has many
> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
> call the flush_tlb_mm  later.
> 
> I think we can try the following way to do.
> 
> vm_write_begin(vma)
> copy_pte_range
> vm_write_end(vma)
> 
> The speculative page fault will return to grap the mmap_sem to run the nromal path.
> Any thought?

Here is a new version of the patch fixing this issue. There is no 
additional TLB flush, all the fix is belonging on vm_write_{begin,end} 
calls.

I did some test on x86_64 and PowerPC but that needs to be double check 
on arm64.

Vinayak, Zhong, could you please give it a try ?

Thanks,
Laurent.


[-- Attachment #2: 0001-mm-protect-against-PTE-changes-done-by-dup_mmap.patch --]
[-- Type: text/plain, Size: 4050 bytes --]

From 3be977febb9ff93d516a2d222cca4b5a52472a9f Mon Sep 17 00:00:00 2001
From: Laurent Dufour <ldufour@linux.ibm.com>
Date: Fri, 18 Jan 2019 16:19:08 +0100
Subject: [PATCH] mm: protect against PTE changes done by dup_mmap()

Vinayak Menon and Ganesh Mahendran reported that the following scenario may
lead to thread being blocked due to data corruption:

    CPU 1                   CPU 2                    CPU 3
    Process 1,              Process 1,               Process 1,
    Thread A                Thread B                 Thread C

    while (1) {             while (1) {              while(1) {
    pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
    pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
    }                       }

In the details this happens because :

    CPU 1                CPU 2                       CPU 3
    fork()
    copy_pte_range()
      set PTE rdonly
    got to next VMA...
     .                   PTE is seen rdonly          PTE still writable
     .                   thread is writing to page
     .                   -> page fault
     .                     copy the page             Thread writes to page
     .                      .                        -> no page fault
     .                     update the PTE
     .                     flush TLB for that PTE
   flush TLB                                        PTE are now rdonly

So the write done by the CPU 3 is interfering with the page copy operation
done by CPU 2, leading to the data corruption.

To avoid this we mark all the VMA involved in the COW mechanism as changing
by calling vm_write_begin(). This ensures that the speculative page fault
handler will not try to handle a fault on these pages.
The marker is set until the TLB is flushed, ensuring that all the CPUs will
now see the PTE as not writable.
Once the TLB is flush, the marker is removed by calling vm_write_end().

The variable last is used to keep tracked of the latest VMA marked to
handle the error path where part of the VMA may have been marked.

Reported-by: Ganesh Mahendran <opensource.ganesh@gmail.com>
Reported-by: Vinayak Menon <vinmenon@codeaurora.org>
Signed-off-by: Laurent Dufour <ldufour@linux.ibm.com>
---
 kernel/fork.c | 30 ++++++++++++++++++++++++++++--
 1 file changed, 28 insertions(+), 2 deletions(-)

diff --git a/kernel/fork.c b/kernel/fork.c
index f1258c2ade09..39854b97d06a 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -395,7 +395,7 @@ EXPORT_SYMBOL(free_task);
 static __latent_entropy int dup_mmap(struct mm_struct *mm,
 					struct mm_struct *oldmm)
 {
-	struct vm_area_struct *mpnt, *tmp, *prev, **pprev;
+	struct vm_area_struct *mpnt, *tmp, *prev, **pprev, *last = NULL;
 	struct rb_node **rb_link, *rb_parent;
 	int retval;
 	unsigned long charge;
@@ -515,8 +515,18 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
 		rb_parent = &tmp->vm_rb;
 
 		mm->map_count++;
-		if (!(tmp->vm_flags & VM_WIPEONFORK))
+		if (!(tmp->vm_flags & VM_WIPEONFORK)) {
+			if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT)) {
+				/*
+				 * Mark this VMA as changing to prevent the
+				 * speculative page fault hanlder to process
+				 * it until the TLB are flushed below.
+				 */
+				last = mpnt;
+				vm_write_begin(mpnt);
+			}
 			retval = copy_page_range(mm, oldmm, mpnt);
+		}
 
 		if (tmp->vm_ops && tmp->vm_ops->open)
 			tmp->vm_ops->open(tmp);
@@ -530,6 +540,22 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
 out:
 	up_write(&mm->mmap_sem);
 	flush_tlb_mm(oldmm);
+
+	if (IS_ENABLED(CONFIG_SPECULATIVE_PAGE_FAULT)) {
+		/*
+		 * Since the TLB has been flush, we can safely unmark the
+		 * copied VMAs and allows the speculative page fault handler to
+		 * process them again.
+		 * Walk back the VMA list from the last marked VMA.
+		 */
+		for (; last; last = last->vm_prev) {
+			if (last->vm_flags & VM_DONTCOPY)
+				continue;
+			if (!(last->vm_flags & VM_WIPEONFORK))
+				vm_write_end(last);
+		}
+	}
+
 	up_write(&oldmm->mmap_sem);
 	dup_userfaultfd_complete(&uf);
 fail_uprobe_end:
-- 
2.20.1

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/19 0:24, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>
> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>
> Vinayak, Zhong, could you please give it a try ?
>
Thanks, look good to me. I will try it.

Sincerely,
zhong jiang
> Thanks,
> Laurent.
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/19 0:24, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>
> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>
> Vinayak, Zhong, could you please give it a try ?
>
Thanks, look good to me. I will try it.

Sincerely,
zhong jiang
> Thanks,
> Laurent.
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/19 0:24, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>
> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>
> Vinayak, Zhong, could you please give it a try ?
>
Hi Laurent

I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.

but It will better to filter the condition by is_cow_mapping. is it right?

for example:

if (is_cow_mapping(mnpt->vm_flags)) {
      
    ........
}
   
Thanks,
zhong jiang
> Thanks,
> Laurent.
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/19 0:24, Laurent Dufour wrote:
> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>
>>>>> With the patch below, we don't hit the issue.
>>>>>
>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>
>>>>> It is observed that the following scenario results in
>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>> forever after few iterations.
>>>>>
>>>>> CPU 1                   CPU 2                    CPU 3
>>>>> Process 1,              Process 1,               Process 1,
>>>>> Thread A                Thread B                 Thread C
>>>>>
>>>>> while (1) {             while (1) {              while(1) {
>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>> }                       }
>>>>>
>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>> on one of the CPUs at least. This can create a problem if one
>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>> entry can still modify old_page even after it is copied to
>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>> Nice catch and thanks for your investigation!
>>>>
>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>
>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>> ---
>>>>>    mm/memory.c | 7 +++++++
>>>>>    1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 52080e4..1ea168ff 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>                   return VM_FAULT_RETRY;
>>>>>           }
>>>>>
>>>>> +       /*
>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>> +        * in copy_one_pte
>>>>> +        */
>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>> +
>>>>>           mem_cgroup_oom_enable();
>>>>>           ret = handle_pte_fault(&vmf);
>>>>>           mem_cgroup_oom_disable();
>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>
>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>
>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>
>>>
>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>
>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>> Hi, Vinayak and Laurent
>>
>> I think the below change will impact the performance significantly. Becuase most of process has many
>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>> call the flush_tlb_mm  later.
>>
>> I think we can try the following way to do.
>>
>> vm_write_begin(vma)
>> copy_pte_range
>> vm_write_end(vma)
>>
>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>> Any thought?
>
> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>
> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>
> Vinayak, Zhong, could you please give it a try ?
>
Hi Laurent

I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.

but It will better to filter the condition by is_cow_mapping. is it right?

for example:

if (is_cow_mapping(mnpt->vm_flags)) {
      
    ........
}
   
Thanks,
zhong jiang
> Thanks,
> Laurent.
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 22/01/2019 à 17:22, zhong jiang a écrit :
> On 2019/1/19 0:24, Laurent Dufour wrote:
>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>> Hi Laurent,
>>>>>>>
>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>
>>>>>> With the patch below, we don't hit the issue.
>>>>>>
>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>
>>>>>> It is observed that the following scenario results in
>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>> forever after few iterations.
>>>>>>
>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>> Process 1,              Process 1,               Process 1,
>>>>>> Thread A                Thread B                 Thread C
>>>>>>
>>>>>> while (1) {             while (1) {              while(1) {
>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>> }                       }
>>>>>>
>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>> entry can still modify old_page even after it is copied to
>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>> Nice catch and thanks for your investigation!
>>>>>
>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>
>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>> ---
>>>>>>     mm/memory.c | 7 +++++++
>>>>>>     1 file changed, 7 insertions(+)
>>>>>>
>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>> index 52080e4..1ea168ff 100644
>>>>>> --- a/mm/memory.c
>>>>>> +++ b/mm/memory.c
>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>                    return VM_FAULT_RETRY;
>>>>>>            }
>>>>>>
>>>>>> +       /*
>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>> +        * in copy_one_pte
>>>>>> +        */
>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>> +
>>>>>>            mem_cgroup_oom_enable();
>>>>>>            ret = handle_pte_fault(&vmf);
>>>>>>            mem_cgroup_oom_disable();
>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>
>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>
>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>
>>>>
>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>
>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>> Hi, Vinayak and Laurent
>>>
>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>> call the flush_tlb_mm  later.
>>>
>>> I think we can try the following way to do.
>>>
>>> vm_write_begin(vma)
>>> copy_pte_range
>>> vm_write_end(vma)
>>>
>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>> Any thought?
>>
>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>
>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>
>> Vinayak, Zhong, could you please give it a try ?
>>
> Hi Laurent
> 
> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.

Good news !

> 
> but It will better to filter the condition by is_cow_mapping. is it right?
> 
> for example:
> 
> if (is_cow_mapping(mnpt->vm_flags)) {
>        
>      ........
> }

That's doable for sure but I don't think this has to be introduce in 
dup_mmap().
Unless there is a real performance benefit to do so, I don't think 
dup_mmap() has to mimic underlying checks done in copy_page_range().

Cheers,
Laurent.

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/24 16:20, Laurent Dufour wrote:
> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>> Hi Laurent,
>>>>>>>>
>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>
>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>
>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>
>>>>>>> It is observed that the following scenario results in
>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>> forever after few iterations.
>>>>>>>
>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>
>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>> }                       }
>>>>>>>
>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>> Nice catch and thanks for your investigation!
>>>>>>
>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>
>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>> ---
>>>>>>>     mm/memory.c | 7 +++++++
>>>>>>>     1 file changed, 7 insertions(+)
>>>>>>>
>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>> --- a/mm/memory.c
>>>>>>> +++ b/mm/memory.c
>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>                    return VM_FAULT_RETRY;
>>>>>>>            }
>>>>>>>
>>>>>>> +       /*
>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>> +        * in copy_one_pte
>>>>>>> +        */
>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>> +
>>>>>>>            mem_cgroup_oom_enable();
>>>>>>>            ret = handle_pte_fault(&vmf);
>>>>>>>            mem_cgroup_oom_disable();
>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>
>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>
>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>
>>>>>
>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>
>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>> Hi, Vinayak and Laurent
>>>>
>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>> call the flush_tlb_mm  later.
>>>>
>>>> I think we can try the following way to do.
>>>>
>>>> vm_write_begin(vma)
>>>> copy_pte_range
>>>> vm_write_end(vma)
>>>>
>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>> Any thought?
>>>
>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>
>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>
>>> Vinayak, Zhong, could you please give it a try ?
>>>
>> Hi Laurent
>>
>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>
> Good news !
>
>>
>> but It will better to filter the condition by is_cow_mapping. is it right?
>>
>> for example:
>>
>> if (is_cow_mapping(mnpt->vm_flags)) {
>>             ........
>> }
>
> That's doable for sure but I don't think this has to be introduce in dup_mmap().
> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>

Hi, Laurent

I test the performace with microbench after appling the patch. I find 
the page fault latency will increase about 8% than before.  I think we
should use is_cow_mapping to waken the impact and I will try it out.

or we can use the following solution to replace as Vinayak has said.
 
if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
    return VM_FAULT_RETRY;

Even though it will influence the performance of SPF, but at least it does
not bring in any negative impact. Any thought?

Thanks,


> Cheers,
> Laurent.
>
>
> .
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/24 16:20, Laurent Dufour wrote:
> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>> Hi Laurent,
>>>>>>>>
>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>
>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>
>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>
>>>>>>> It is observed that the following scenario results in
>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>> forever after few iterations.
>>>>>>>
>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>
>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>> }                       }
>>>>>>>
>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>> Nice catch and thanks for your investigation!
>>>>>>
>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>
>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>> ---
>>>>>>>     mm/memory.c | 7 +++++++
>>>>>>>     1 file changed, 7 insertions(+)
>>>>>>>
>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>> --- a/mm/memory.c
>>>>>>> +++ b/mm/memory.c
>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>                    return VM_FAULT_RETRY;
>>>>>>>            }
>>>>>>>
>>>>>>> +       /*
>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>> +        * in copy_one_pte
>>>>>>> +        */
>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>> +
>>>>>>>            mem_cgroup_oom_enable();
>>>>>>>            ret = handle_pte_fault(&vmf);
>>>>>>>            mem_cgroup_oom_disable();
>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>
>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>
>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>
>>>>>
>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>
>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>> Hi, Vinayak and Laurent
>>>>
>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>> call the flush_tlb_mm  later.
>>>>
>>>> I think we can try the following way to do.
>>>>
>>>> vm_write_begin(vma)
>>>> copy_pte_range
>>>> vm_write_end(vma)
>>>>
>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>> Any thought?
>>>
>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>
>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>
>>> Vinayak, Zhong, could you please give it a try ?
>>>
>> Hi Laurent
>>
>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>
> Good news !
>
>>
>> but It will better to filter the condition by is_cow_mapping. is it right?
>>
>> for example:
>>
>> if (is_cow_mapping(mnpt->vm_flags)) {
>>             ........
>> }
>
> That's doable for sure but I don't think this has to be introduce in dup_mmap().
> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>

Hi, Laurent

I test the performace with microbench after appling the patch. I find 
the page fault latency will increase about 8% than before.  I think we
should use is_cow_mapping to waken the impact and I will try it out.

or we can use the following solution to replace as Vinayak has said.
 
if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
    return VM_FAULT_RETRY;

Even though it will influence the performance of SPF, but at least it does
not bring in any negative impact. Any thought?

Thanks,


> Cheers,
> Laurent.
>
>
> .
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 25/01/2019 à 13:32, zhong jiang a écrit :
> On 2019/1/24 16:20, Laurent Dufour wrote:
>> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>>> Hi Laurent,
>>>>>>>>>
>>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>>
>>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>>
>>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>>
>>>>>>>> It is observed that the following scenario results in
>>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>>> forever after few iterations.
>>>>>>>>
>>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>>
>>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>>> }                       }
>>>>>>>>
>>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>>> Nice catch and thanks for your investigation!
>>>>>>>
>>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>>
>>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>> ---
>>>>>>>>      mm/memory.c | 7 +++++++
>>>>>>>>      1 file changed, 7 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>>> --- a/mm/memory.c
>>>>>>>> +++ b/mm/memory.c
>>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>>                     return VM_FAULT_RETRY;
>>>>>>>>             }
>>>>>>>>
>>>>>>>> +       /*
>>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>>> +        * in copy_one_pte
>>>>>>>> +        */
>>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>>> +
>>>>>>>>             mem_cgroup_oom_enable();
>>>>>>>>             ret = handle_pte_fault(&vmf);
>>>>>>>>             mem_cgroup_oom_disable();
>>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>>
>>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>>
>>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>>
>>>>>>
>>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>>
>>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>>> Hi, Vinayak and Laurent
>>>>>
>>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>>> call the flush_tlb_mm  later.
>>>>>
>>>>> I think we can try the following way to do.
>>>>>
>>>>> vm_write_begin(vma)
>>>>> copy_pte_range
>>>>> vm_write_end(vma)
>>>>>
>>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>>> Any thought?
>>>>
>>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>>
>>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>>
>>>> Vinayak, Zhong, could you please give it a try ?
>>>>
>>> Hi Laurent
>>>
>>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>>
>> Good news !
>>
>>>
>>> but It will better to filter the condition by is_cow_mapping. is it right?
>>>
>>> for example:
>>>
>>> if (is_cow_mapping(mnpt->vm_flags)) {
>>>              ........
>>> }
>>
>> That's doable for sure but I don't think this has to be introduce in dup_mmap().
>> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>>
> 
> Hi, Laurent
> 
> I test the performace with microbench after appling the patch. I find
> the page fault latency will increase about 8% than before.  I think we
> should use is_cow_mapping to waken the impact and I will try it out.

That's interesting,  I would not expect such a higher latency assuming 
that most of the area not in copied on write are also not managed by the 
speculative page fault handler (file mapping, etc.). Anyway I'm looking 
forward to see the result with additional is_cow_mapping() check.

> or we can use the following solution to replace as Vinayak has said.
> 
> if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>      return VM_FAULT_RETRY;
> 
> Even though it will influence the performance of SPF, but at least it does
> not bring in any negative impact. Any thought?

I don't agree, this checks will completely by pass the SPF handler for 
all the COW areas, even if there is no race situation.

Cheers,
Laurent.
> 
> Thanks,
> 
> 
>> Cheers,
>> Laurent.
>>
>>
>> .
>>
> 
> 

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/28 16:59, Laurent Dufour wrote:
> Le 25/01/2019 à 13:32, zhong jiang a écrit :
>> On 2019/1/24 16:20, Laurent Dufour wrote:
>>> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>>>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>>>> Hi Laurent,
>>>>>>>>>>
>>>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>>>
>>>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>>>
>>>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>>>
>>>>>>>>> It is observed that the following scenario results in
>>>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>>>> forever after few iterations.
>>>>>>>>>
>>>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>>>
>>>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>>>> }                       }
>>>>>>>>>
>>>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>>>> Nice catch and thanks for your investigation!
>>>>>>>>
>>>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>>>
>>>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>> ---
>>>>>>>>>      mm/memory.c | 7 +++++++
>>>>>>>>>      1 file changed, 7 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>>>> --- a/mm/memory.c
>>>>>>>>> +++ b/mm/memory.c
>>>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>>>                     return VM_FAULT_RETRY;
>>>>>>>>>             }
>>>>>>>>>
>>>>>>>>> +       /*
>>>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>>>> +        * in copy_one_pte
>>>>>>>>> +        */
>>>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>>>> +
>>>>>>>>>             mem_cgroup_oom_enable();
>>>>>>>>>             ret = handle_pte_fault(&vmf);
>>>>>>>>>             mem_cgroup_oom_disable();
>>>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>>>
>>>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>>>
>>>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>>>
>>>>>>>
>>>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>>>
>>>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>>>> Hi, Vinayak and Laurent
>>>>>>
>>>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>>>> call the flush_tlb_mm  later.
>>>>>>
>>>>>> I think we can try the following way to do.
>>>>>>
>>>>>> vm_write_begin(vma)
>>>>>> copy_pte_range
>>>>>> vm_write_end(vma)
>>>>>>
>>>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>>>> Any thought?
>>>>>
>>>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>>>
>>>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>>>
>>>>> Vinayak, Zhong, could you please give it a try ?
>>>>>
>>>> Hi Laurent
>>>>
>>>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>>>
>>> Good news !
>>>
>>>>
>>>> but It will better to filter the condition by is_cow_mapping. is it right?
>>>>
>>>> for example:
>>>>
>>>> if (is_cow_mapping(mnpt->vm_flags)) {
>>>>              ........
>>>> }
>>>
>>> That's doable for sure but I don't think this has to be introduce in dup_mmap().
>>> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>>>
>>
>> Hi, Laurent
>>
>> I test the performace with microbench after appling the patch. I find
>> the page fault latency will increase about 8% than before.  I think we
>> should use is_cow_mapping to waken the impact and I will try it out.
>
> That's interesting,  I would not expect such a higher latency assuming that most of the area not in copied on write are also not managed by the speculative page fault handler (file mapping, etc.). Anyway I'm looking forward to see the result with additional is_cow_mapping() check.
>
I test the performance again. It is the protect error access latency in lat_sig.c that it will result in a drop of 8% in that testcase.
The page fault latency, In fact, does not impact the performace. It seems to just the fluctuation.

Thanks,
zhong jiang
>> or we can use the following solution to replace as Vinayak has said.
>>
>> if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>      return VM_FAULT_RETRY;
>>
>> Even though it will influence the performance of SPF, but at least it does
>> not bring in any negative impact. Any thought?
>
> I don't agree, this checks will completely by pass the SPF handler for all the COW areas, even if there is no race situation.
>
> Cheers,
> Laurent.
>>
>> Thanks,
>>
>>
>>> Cheers,
>>> Laurent.
>>>
>>>
>>> .
>>>
>>
>>
>
>
>

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/28 16:59, Laurent Dufour wrote:
> Le 25/01/2019 à 13:32, zhong jiang a écrit :
>> On 2019/1/24 16:20, Laurent Dufour wrote:
>>> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>>>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>>>> Hi Laurent,
>>>>>>>>>>
>>>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>>>
>>>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>>>
>>>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>>>
>>>>>>>>> It is observed that the following scenario results in
>>>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>>>> forever after few iterations.
>>>>>>>>>
>>>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>>>
>>>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>>>> }                       }
>>>>>>>>>
>>>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>>>> Nice catch and thanks for your investigation!
>>>>>>>>
>>>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>>>
>>>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>> ---
>>>>>>>>>      mm/memory.c | 7 +++++++
>>>>>>>>>      1 file changed, 7 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>>>> --- a/mm/memory.c
>>>>>>>>> +++ b/mm/memory.c
>>>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>>>                     return VM_FAULT_RETRY;
>>>>>>>>>             }
>>>>>>>>>
>>>>>>>>> +       /*
>>>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>>>> +        * in copy_one_pte
>>>>>>>>> +        */
>>>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>>>> +
>>>>>>>>>             mem_cgroup_oom_enable();
>>>>>>>>>             ret = handle_pte_fault(&vmf);
>>>>>>>>>             mem_cgroup_oom_disable();
>>>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>>>
>>>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>>>
>>>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>>>
>>>>>>>
>>>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>>>
>>>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>>>> Hi, Vinayak and Laurent
>>>>>>
>>>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>>>> call the flush_tlb_mm  later.
>>>>>>
>>>>>> I think we can try the following way to do.
>>>>>>
>>>>>> vm_write_begin(vma)
>>>>>> copy_pte_range
>>>>>> vm_write_end(vma)
>>>>>>
>>>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>>>> Any thought?
>>>>>
>>>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>>>
>>>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>>>
>>>>> Vinayak, Zhong, could you please give it a try ?
>>>>>
>>>> Hi Laurent
>>>>
>>>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>>>
>>> Good news !
>>>
>>>>
>>>> but It will better to filter the condition by is_cow_mapping. is it right?
>>>>
>>>> for example:
>>>>
>>>> if (is_cow_mapping(mnpt->vm_flags)) {
>>>>              ........
>>>> }
>>>
>>> That's doable for sure but I don't think this has to be introduce in dup_mmap().
>>> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>>>
>>
>> Hi, Laurent
>>
>> I test the performace with microbench after appling the patch. I find
>> the page fault latency will increase about 8% than before.  I think we
>> should use is_cow_mapping to waken the impact and I will try it out.
>
> That's interesting,  I would not expect such a higher latency assuming that most of the area not in copied on write are also not managed by the speculative page fault handler (file mapping, etc.). Anyway I'm looking forward to see the result with additional is_cow_mapping() check.
>
I test the performance again. It is the protect error access latency in lat_sig.c that it will result in a drop of 8% in that testcase.
The page fault latency, In fact, does not impact the performace. It seems to just the fluctuation.

Thanks,
zhong jiang
>> or we can use the following solution to replace as Vinayak has said.
>>
>> if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>      return VM_FAULT_RETRY;
>>
>> Even though it will influence the performance of SPF, but at least it does
>> not bring in any negative impact. Any thought?
>
> I don't agree, this checks will completely by pass the SPF handler for all the COW areas, even if there is no race situation.
>
> Cheers,
> Laurent.
>>
>> Thanks,
>>
>>
>>> Cheers,
>>> Laurent.
>>>
>>>
>>> .
>>>
>>
>>
>
>
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 28/01/2019 à 15:09, zhong jiang a écrit :
> On 2019/1/28 16:59, Laurent Dufour wrote:
>> Le 25/01/2019 à 13:32, zhong jiang a écrit :
>>> On 2019/1/24 16:20, Laurent Dufour wrote:
>>>> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>>>>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>>>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>>>>> Hi Laurent,
>>>>>>>>>>>
>>>>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>>>>
>>>>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>>>>
>>>>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>>>>
>>>>>>>>>> It is observed that the following scenario results in
>>>>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>>>>> forever after few iterations.
>>>>>>>>>>
>>>>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>>>>
>>>>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>>>>> }                       }
>>>>>>>>>>
>>>>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>>>>> Nice catch and thanks for your investigation!
>>>>>>>>>
>>>>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>>>>
>>>>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>>> ---
>>>>>>>>>>       mm/memory.c | 7 +++++++
>>>>>>>>>>       1 file changed, 7 insertions(+)
>>>>>>>>>>
>>>>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>>>>> --- a/mm/memory.c
>>>>>>>>>> +++ b/mm/memory.c
>>>>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>>>>                      return VM_FAULT_RETRY;
>>>>>>>>>>              }
>>>>>>>>>>
>>>>>>>>>> +       /*
>>>>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>>>>> +        * in copy_one_pte
>>>>>>>>>> +        */
>>>>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>>>>> +
>>>>>>>>>>              mem_cgroup_oom_enable();
>>>>>>>>>>              ret = handle_pte_fault(&vmf);
>>>>>>>>>>              mem_cgroup_oom_disable();
>>>>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>>>>
>>>>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>>>>
>>>>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>>>>
>>>>>>>>
>>>>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>>>>
>>>>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>>>>> Hi, Vinayak and Laurent
>>>>>>>
>>>>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>>>>> call the flush_tlb_mm  later.
>>>>>>>
>>>>>>> I think we can try the following way to do.
>>>>>>>
>>>>>>> vm_write_begin(vma)
>>>>>>> copy_pte_range
>>>>>>> vm_write_end(vma)
>>>>>>>
>>>>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>>>>> Any thought?
>>>>>>
>>>>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>>>>
>>>>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>>>>
>>>>>> Vinayak, Zhong, could you please give it a try ?
>>>>>>
>>>>> Hi Laurent
>>>>>
>>>>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>>>>
>>>> Good news !
>>>>
>>>>>
>>>>> but It will better to filter the condition by is_cow_mapping. is it right?
>>>>>
>>>>> for example:
>>>>>
>>>>> if (is_cow_mapping(mnpt->vm_flags)) {
>>>>>               ........
>>>>> }
>>>>
>>>> That's doable for sure but I don't think this has to be introduce in dup_mmap().
>>>> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>>>>
>>>
>>> Hi, Laurent
>>>
>>> I test the performace with microbench after appling the patch. I find
>>> the page fault latency will increase about 8% than before.  I think we
>>> should use is_cow_mapping to waken the impact and I will try it out.
>>
>> That's interesting,  I would not expect such a higher latency assuming that most of the area not in copied on write are also not managed by the speculative page fault handler (file mapping, etc.). Anyway I'm looking forward to see the result with additional is_cow_mapping() check.
>>
> I test the performance again. It is the protect error access latency in lat_sig.c that it will result in a drop of 8% in that testcase.

What is that "protect error access latency in lat_sig.c" ?

> The page fault latency, In fact, does not impact the performace. It seems to just the fluctuation.
> 
> Thanks,
> zhong jiang
>>> or we can use the following solution to replace as Vinayak has said.
>>>
>>> if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>       return VM_FAULT_RETRY;
>>>
>>> Even though it will influence the performance of SPF, but at least it does
>>> not bring in any negative impact. Any thought?
>>
>> I don't agree, this checks will completely by pass the SPF handler for all the COW areas, even if there is no race situation.
>>
>> Cheers,
>> Laurent.
>>>
>>> Thanks,
>>>
>>>
>>>> Cheers,
>>>> Laurent.
>>>>
>>>>
>>>> .
>>>>
>>>
>>>
>>
>>
>>
> 
> 

Re: [PATCH v11 00/26] Speculative page faults

[zhong jiang]

On 2019/1/28 23:45, Laurent Dufour wrote:
> Le 28/01/2019 à 15:09, zhong jiang a écrit :
>> On 2019/1/28 16:59, Laurent Dufour wrote:
>>> Le 25/01/2019 à 13:32, zhong jiang a écrit :
>>>> On 2019/1/24 16:20, Laurent Dufour wrote:
>>>>> Le 22/01/2019 à 17:22, zhong jiang a écrit :
>>>>>> On 2019/1/19 0:24, Laurent Dufour wrote:
>>>>>>> Le 17/01/2019 à 16:51, zhong jiang a écrit :
>>>>>>>> On 2019/1/16 19:41, Vinayak Menon wrote:
>>>>>>>>> On 1/15/2019 1:54 PM, Laurent Dufour wrote:
>>>>>>>>>> Le 14/01/2019 à 14:19, Vinayak Menon a écrit :
>>>>>>>>>>> On 1/11/2019 9:13 PM, Vinayak Menon wrote:
>>>>>>>>>>>> Hi Laurent,
>>>>>>>>>>>>
>>>>>>>>>>>> We are observing an issue with speculative page fault with the following test code on ARM64 (4.14 kernel, 8 cores).
>>>>>>>>>>>
>>>>>>>>>>> With the patch below, we don't hit the issue.
>>>>>>>>>>>
>>>>>>>>>>> From: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>>>> Date: Mon, 14 Jan 2019 16:06:34 +0530
>>>>>>>>>>> Subject: [PATCH] mm: flush stale tlb entries on speculative write fault
>>>>>>>>>>>
>>>>>>>>>>> It is observed that the following scenario results in
>>>>>>>>>>> threads A and B of process 1 blocking on pthread_mutex_lock
>>>>>>>>>>> forever after few iterations.
>>>>>>>>>>>
>>>>>>>>>>> CPU 1                   CPU 2                    CPU 3
>>>>>>>>>>> Process 1,              Process 1,               Process 1,
>>>>>>>>>>> Thread A                Thread B                 Thread C
>>>>>>>>>>>
>>>>>>>>>>> while (1) {             while (1) {              while(1) {
>>>>>>>>>>> pthread_mutex_lock(l)   pthread_mutex_lock(l)    fork
>>>>>>>>>>> pthread_mutex_unlock(l) pthread_mutex_unlock(l)  }
>>>>>>>>>>> }                       }
>>>>>>>>>>>
>>>>>>>>>>> When from thread C, copy_one_pte write-protects the parent pte
>>>>>>>>>>> (of lock l), stale tlb entries can exist with write permissions
>>>>>>>>>>> on one of the CPUs at least. This can create a problem if one
>>>>>>>>>>> of the threads A or B hits the write fault. Though dup_mmap calls
>>>>>>>>>>> flush_tlb_mm after copy_page_range, since speculative page fault
>>>>>>>>>>> does not take mmap_sem it can proceed further fixing a fault soon
>>>>>>>>>>> after CPU 3 does ptep_set_wrprotect. But the CPU with stale tlb
>>>>>>>>>>> entry can still modify old_page even after it is copied to
>>>>>>>>>>> new_page by wp_page_copy, thus causing a corruption.
>>>>>>>>>> Nice catch and thanks for your investigation!
>>>>>>>>>>
>>>>>>>>>> There is a real synchronization issue here between copy_page_range() and the speculative page fault handler. I didn't get it on PowerVM since the TLB are flushed when arch_exit_lazy_mode() is called in copy_page_range() but now, I can get it when running on x86_64.
>>>>>>>>>>
>>>>>>>>>>> Signed-off-by: Vinayak Menon <vinmenon@codeaurora.org>
>>>>>>>>>>> ---
>>>>>>>>>>>       mm/memory.c | 7 +++++++
>>>>>>>>>>>       1 file changed, 7 insertions(+)
>>>>>>>>>>>
>>>>>>>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>>>>>>>> index 52080e4..1ea168ff 100644
>>>>>>>>>>> --- a/mm/memory.c
>>>>>>>>>>> +++ b/mm/memory.c
>>>>>>>>>>> @@ -4507,6 +4507,13 @@ int __handle_speculative_fault(struct mm_struct *mm, unsigned long address,
>>>>>>>>>>>                      return VM_FAULT_RETRY;
>>>>>>>>>>>              }
>>>>>>>>>>>
>>>>>>>>>>> +       /*
>>>>>>>>>>> +        * Discard tlb entries created before ptep_set_wrprotect
>>>>>>>>>>> +        * in copy_one_pte
>>>>>>>>>>> +        */
>>>>>>>>>>> +       if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>>>>>>>> +               flush_tlb_page(vmf.vma, address);
>>>>>>>>>>> +
>>>>>>>>>>>              mem_cgroup_oom_enable();
>>>>>>>>>>>              ret = handle_pte_fault(&vmf);
>>>>>>>>>>>              mem_cgroup_oom_disable();
>>>>>>>>>> Your patch is fixing the race but I'm wondering about the cost of these tlb flushes. Here we are flushing on a per page basis (architecture like x86_64 are smarter and flush more pages) but there is a request to flush a range of tlb entries each time a cow page is newly touched. I think there could be some bad impact here.
>>>>>>>>>>
>>>>>>>>>> Another option would be to flush the range in copy_pte_range() before unlocking the page table lock. This will flush entries flush_tlb_mm() would later handle in dup_mmap() but that will be called once per fork per cow VMA.
>>>>>>>>>
>>>>>>>>> But wouldn't this cause an unnecessary impact if most of the COW pages remain untouched (which I assume would be the usual case) and thus do not create a fault ?
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> I tried the attached patch which seems to fix the issue on x86_64. Could you please give it a try on arm64 ?
>>>>>>>>>>
>>>>>>>>> Your patch works fine on arm64 with a minor change. Thanks Laurent.
>>>>>>>> Hi, Vinayak and Laurent
>>>>>>>>
>>>>>>>> I think the below change will impact the performance significantly. Becuase most of process has many
>>>>>>>> vmas with cow flags. Flush the tlb in advance is not the better way to avoid the issue and it will
>>>>>>>> call the flush_tlb_mm  later.
>>>>>>>>
>>>>>>>> I think we can try the following way to do.
>>>>>>>>
>>>>>>>> vm_write_begin(vma)
>>>>>>>> copy_pte_range
>>>>>>>> vm_write_end(vma)
>>>>>>>>
>>>>>>>> The speculative page fault will return to grap the mmap_sem to run the nromal path.
>>>>>>>> Any thought?
>>>>>>>
>>>>>>> Here is a new version of the patch fixing this issue. There is no additional TLB flush, all the fix is belonging on vm_write_{begin,end} calls.
>>>>>>>
>>>>>>> I did some test on x86_64 and PowerPC but that needs to be double check on arm64.
>>>>>>>
>>>>>>> Vinayak, Zhong, could you please give it a try ?
>>>>>>>
>>>>>> Hi Laurent
>>>>>>
>>>>>> I apply the patch you had attached and none of any abnormal thing came in two days. It is feasible to fix the issue.
>>>>>
>>>>> Good news !
>>>>>
>>>>>>
>>>>>> but It will better to filter the condition by is_cow_mapping. is it right?
>>>>>>
>>>>>> for example:
>>>>>>
>>>>>> if (is_cow_mapping(mnpt->vm_flags)) {
>>>>>>               ........
>>>>>> }
>>>>>
>>>>> That's doable for sure but I don't think this has to be introduce in dup_mmap().
>>>>> Unless there is a real performance benefit to do so, I don't think dup_mmap() has to mimic underlying checks done in copy_page_range().
>>>>>
>>>>
>>>> Hi, Laurent
>>>>
>>>> I test the performace with microbench after appling the patch. I find
>>>> the page fault latency will increase about 8% than before.  I think we
>>>> should use is_cow_mapping to waken the impact and I will try it out.
>>>
>>> That's interesting,  I would not expect such a higher latency assuming that most of the area not in copied on write are also not managed by the speculative page fault handler (file mapping, etc.). Anyway I'm looking forward to see the result with additional is_cow_mapping() check.
>>>
>> I test the performance again. It is the protect error access latency in lat_sig.c that it will result in a drop of 8% in that testcase.
>
> What is that "protect error access latency in lat_sig.c" ?
>
It is the protect error access that the source code is lat_sig.c in microbench.
>> The page fault latency, In fact, does not impact the performace. It seems to just the fluctuation.
>>
>> Thanks,
>> zhong jiang
>>>> or we can use the following solution to replace as Vinayak has said.
>>>>
>>>> if (flags & FAULT_FLAG_WRITE && !pte_write(vmf.orig_pte))
>>>>       return VM_FAULT_RETRY;
>>>>
>>>> Even though it will influence the performance of SPF, but at least it does
>>>> not bring in any negative impact. Any thought?
>>>
>>> I don't agree, this checks will completely by pass the SPF handler for all the COW areas, even if there is no race situation.
>>>
>>> Cheers,
>>> Laurent.
>>>>
>>>> Thanks,
>>>>
>>>>
>>>>> Cheers,
>>>>> Laurent.
>>>>>
>>>>>
>>>>> .
>>>>>
>>>>
>>>>
>>>
>>>
>>>
>>
>>
>
>
> .
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Le 05/11/2018 A  11:42, Balbir Singh a A(C)critA :
> On Thu, May 17, 2018 at 01:06:07PM +0200, Laurent Dufour wrote:
>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>> page fault without holding the mm semaphore [1].
>>
>> The idea is to try to handle user space page faults without holding the
>> mmap_sem. This should allow better concurrency for massively threaded
> 
> Question -- I presume mmap_sem (rw_semaphore implementation tested against)
> was qrwlock?

I don't think so, this series doesn't change the mmap_sem definition so 
it still belongs to the 'struct rw_semaphore'.

Laurent.

Re: [PATCH v11 00/26] Speculative page faults

[Balbir Singh]

On Thu, May 17, 2018 at 01:06:07PM +0200, Laurent Dufour wrote:
> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
> page fault without holding the mm semaphore [1].
> 
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded

Question -- I presume mmap_sem (rw_semaphore implementation tested against)
was qrwlock?

Balbir Singh.

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

[-- Attachment #1: Type: text/plain, Size: 46630 bytes --]

Hi Laurent,

I am sorry for replying you so late. 
The previous LKP test for this case are running on the same Intel skylake 4s platform, but it need maintain recently. 
So I changed to another test box to run the page_fault3 test case, it is Intel skylake 2s platform (nr_cpu: 104, memory: 64G).

I applied your patch to the SPF kernel (commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12), then triggered below 2 cases test.
a)  Turn on the SPF handler by below command, then run page_fault3-thp-always test.
echo 1 > /proc/sys/vm/speculative_page_fault

b) Turn off the SPF handler by below command, then run page_fault3-thp-always test.
 echo 0 > /proc/sys/vm/speculative_page_fault

Every test run 3 times, and then get test result and capture perf data. 
Here is average result for will-it-scale.per_thread_ops:                                                         
                                                                                          SPF_turn_off       SPF_turn_on
page_fault3-THP-Alwasys.will-it-scale.per_thread_ops    31963                  26285

Best regards,
Haiyan Song

________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Wednesday, August 22, 2018 10:23 PM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 03/08/2018 08:36, Song, HaiyanX wrote:
> Hi Laurent,

Hi Haiyan,

Sorry for the late answer, I was off a couple of days.

>
> Thanks for your analysis for the last perf results.
> Your mentioned ," the major differences at the head of the perf report is the 92% testcase which is weirdly not reported
> on the head side", which is a bug of 0-day,and it caused the item is not counted in perf.
>
> I've triggered the test page_fault2 and page_fault3 again only with thread mode of will-it-scale on 0-day (on the same test box,every case tested 3 times).
> I checked the perf report have no above mentioned problem.
>
> I have compared them, found some items have difference, such as below case:
>        page_fault2-thp-always: handle_mm_fault, base: 45.22%    head: 29.41%
>        page_fault3-thp-always: handle_mm_fault, base: 22.95%    head: 14.15%

These would mean that the system spends lees time running handle_mm_fault()
when SPF is in the picture in this 2 cases which is good. This should lead to
better results with the SPF series, and I can't find any values higher on the
head side.

>
> So i attached the perf result in mail again, could your have a look again for checking the difference between base and head commit.

I took a close look to all the perf result you sent, but I can't identify any
major difference. But the compiler optimization is getting rid of the
handle_pte_fault() symbol on the base kernel which add complexity to check the
differences.

To get rid of that, I'm proposing that you applied the attached patch to the
spf kernel. This patch is allowing to turn on/off the SPF handler through
/proc/sys/vm/speculative_page_fault.

This should ease the testing by limiting the reboot and avoid kernel's symbols
mismatch. Obviously there is still a small overhead due to the check but it
should not be viewable.

With this patch applied you can simply run
echo 1 > /proc/sys/vm/speculative_page_fault
to run a test with the speculative page fault handler activated. Or run
echo 0 > /proc/sys/vm/speculative_page_fault
to run a test without it.

I'm really sorry to asking that again, but could please run the test
page_fault3_base_THP-Always with and without SPF and capture the perf output.

I think we should focus on that test which showed the biggest regression.

Thanks,
Laurent.


>
> Thanks,
> Haiyan, Song
>
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Tuesday, July 17, 2018 5:36 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
>
> On 13/07/2018 05:56, Song, HaiyanX wrote:
>> Hi Laurent,
>
> Hi Haiyan,
>
> Thanks a lot for sharing this perf reports.
>
> I looked at them closely, and I've to admit that I was not able to found a
> major difference between the base and the head report, except that
> handle_pte_fault() is no more in-lined in the head one.
>
> As expected, __handle_speculative_fault() is never traced since these tests are
> dealing with file mapping, not handled in the speculative way.
>
> When running these test did you seen a major differences in the test's result
> between base and head ?
>
> From the number of cycles counted, the biggest difference is page_fault3 when
> run with the THP enabled:
>                                 BASE            HEAD            Delta
> page_fault2_base_thp_never      1142252426747   1065866197589   -6.69%
> page_fault2_base_THP-Alwasys    1124844374523   1076312228927   -4.31%
> page_fault3_base_thp_never      1099387298152   1134118402345   3.16%
> page_fault3_base_THP-Always     1059370178101   853985561949    -19.39%
>
>
> The very weird thing is the difference of the delta cycles reported between
> thp never and thp always, because the speculative way is aborted when checking
> for the vma->ops field, which is the same in both case, and the thp is never
> checked. So there is no code covering differnce, on the speculative path,
> between these 2 cases. This leads me to think that there are other interactions
> interfering in the measure.
>
> Looking at the perf-profile_page_fault3_*_THP-Always, the major differences at
> the head of the perf report is the 92% testcase which is weirdly not reported
> on the head side :
>     92.02%    22.33%  page_fault3_processes  [.] testcase
> 92.02% testcase
>
> Then the base reported 37.67% for __do_page_fault() where the head reported
> 48.41%, but the only difference in this function, between base and head, is the
> call to handle_speculative_fault(). But this is a macro checking for the fault
> flags, and mm->users and then calling __handle_speculative_fault() if needed.
> So this can't explain this difference, except if __handle_speculative_fault()
> is inlined in __do_page_fault().
> Is this the case on your build ?
>
> Haiyan, do you still have the output of the test to check those numbers too ?
>
> Cheers,
> Laurent
>
>> I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case.
>> Please help to check on these data if it can help you to find the higher change. Thanks.
>>
>> File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2
>>     tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.
>>
>> Best regards,
>> Haiyan Song
>>
>> ________________________________________
>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>> Sent: Thursday, July 12, 2018 1:05 AM
>> To: Song, HaiyanX
>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>
>> Hi Haiyan,
>>
>> Do you get a chance to capture some performance cycles on your system ?
>> I still can't get these numbers on my hardware.
>>
>> Thanks,
>> Laurent.
>>
>> On 04/07/2018 09:51, Laurent Dufour wrote:
>>> On 04/07/2018 05:23, Song, HaiyanX wrote:
>>>> Hi Laurent,
>>>>
>>>>
>>>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>>>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>>>
>>> Repeating the test only 3 times seems a bit too low to me.
>>>
>>> I'll focus on the higher change for the moment, but I don't have access to such
>>> a hardware.
>>>
>>> Is possible to provide a diff between base and SPF of the performance cycles
>>> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>>>
>>> Please stay focus on the test case process to see exactly where the series is
>>> impacting.
>>>
>>> Thanks,
>>> Laurent.
>>>
>>>>
>>>> And I did not find other high variation on test case result.
>>>>
>>>> a). Enable THP
>>>> testcase                          base     stddev       change      head     stddev         metric
>>>> page_fault3/enable THP           10519      ± 3%        -20.5%      8368      ±6%          will-it-scale.per_thread_ops
>>>> page_fault2/enalbe THP            8281      ± 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>>>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>>>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>>>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>>>
>>>> b). Disable THP
>>>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>>>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>>>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>>>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>>>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>>>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>>>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>>>
>>>>
>>>> Best regards,
>>>> Haiyan Song
>>>> ________________________________________
>>>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Monday, July 02, 2018 4:59 PM
>>>> To: Song, HaiyanX
>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>>>> V9 patch serials.
>>>>>
>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>>>> commit id:
>>>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>>>> Benchmark: will-it-scale
>>>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>>>
>>>>> Metrics:
>>>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>> THP: enable / disable
>>>>> nr_task:100%
>>>>>
>>>>> 1. Regressions:
>>>>>
>>>>> a). Enable THP
>>>>> testcase                          base           change      head           metric
>>>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>>>
>>>>> b). Disable THP
>>>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>>>
>>>>> Notes: for the above  values of test result, the higher is better.
>>>>
>>>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>>>> get reproducible results. The results have huge variation, even on the vanilla
>>>> kernel, and I can't state on any changes due to that.
>>>>
>>>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>>>> measure any changes between the vanilla and the SPF patched ones:
>>>>
>>>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>>>> page_fault3_threads             2697.7          2683.5          -0.53%
>>>> page_fault2_threads             170660.6        169574.1        -0.64%
>>>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>>>> context_switch1_processes       6478076.2       6529493.5       0.79%
>>>> brk1                            243391.2        238527.5        -2.00%
>>>>
>>>> Tests were run 10 times, no high variation detected.
>>>>
>>>> Did you see high variation on your side ? How many times the test were run to
>>>> compute the average values ?
>>>>
>>>> Thanks,
>>>> Laurent.
>>>>
>>>>
>>>>>
>>>>> 2. Improvement: not found improvement based on the selected test cases.
>>>>>
>>>>>
>>>>> Best regards
>>>>> Haiyan Song
>>>>> ________________________________________
>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>> Sent: Monday, May 28, 2018 4:54 PM
>>>>> To: Song, HaiyanX
>>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>>
>>>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> Yes, these tests are done on V9 patch.
>>>>>
>>>>> Do you plan to give this V11 a run ?
>>>>>
>>>>>>
>>>>>>
>>>>>> Best regards,
>>>>>> Haiyan Song
>>>>>>
>>>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>>>
>>>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>>>> tested on Intel 4s Skylake platform.
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>>>> series" while responding to the v11 header series...
>>>>>>> Were these tests done on v9 or v11 ?
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Laurent.
>>>>>>>
>>>>>>>>
>>>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>>>> Commit id:
>>>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>>>> Benchmark suite: will-it-scale
>>>>>>>> Download link:
>>>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>>>> Metrics:
>>>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>>>> THP: enable / disable
>>>>>>>> nr_task: 100%
>>>>>>>>
>>>>>>>> 1. Regressions:
>>>>>>>> a) THP enabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>>>
>>>>>>>> b) THP disabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>>>
>>>>>>>> 2. Improvements:
>>>>>>>> a) THP enabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>>>
>>>>>>>> b) THP disabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>>>
>>>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>>>
>>>>>>>>
>>>>>>>> Best regards
>>>>>>>> Haiyan Song
>>>>>>>>
>>>>>>>> ________________________________________
>>>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>>>
>>>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>>>> page fault without holding the mm semaphore [1].
>>>>>>>>
>>>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>>>
>>>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>>>> limiting the locking contention to these operations which are expected to
>>>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>>>> benchmark anymore.
>>>>>>>>
>>>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>>>> speculative page fault in that case.
>>>>>>>>
>>>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>>>> checked during the page fault are modified.
>>>>>>>>
>>>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>>>> parallel change is possible at this time.
>>>>>>>>
>>>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>>>> classic page fault handler will be called to handle the operation while
>>>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>>>> PTE.
>>>>>>>>
>>>>>>>> In pseudo code, this could be seen as:
>>>>>>>>     speculative_page_fault()
>>>>>>>>     {
>>>>>>>>             vma = get_vma()
>>>>>>>>             check vma sequence count
>>>>>>>>             check vma's support
>>>>>>>>             disable interrupt
>>>>>>>>                   check pgd,p4d,...,pte
>>>>>>>>                   save pmd and pte in vmf
>>>>>>>>                   save vma sequence counter in vmf
>>>>>>>>             enable interrupt
>>>>>>>>             check vma sequence count
>>>>>>>>             handle_pte_fault(vma)
>>>>>>>>                     ..
>>>>>>>>                     page = alloc_page()
>>>>>>>>                     pte_map_lock()
>>>>>>>>                             disable interrupt
>>>>>>>>                                     abort if sequence counter has changed
>>>>>>>>                                     abort if pmd or pte has changed
>>>>>>>>                                     pte map and lock
>>>>>>>>                             enable interrupt
>>>>>>>>                     if abort
>>>>>>>>                        free page
>>>>>>>>                        abort
>>>>>>>>                     ...
>>>>>>>>     }
>>>>>>>>
>>>>>>>>     arch_fault_handler()
>>>>>>>>     {
>>>>>>>>             if (speculative_page_fault(&vma))
>>>>>>>>                goto done
>>>>>>>>     again:
>>>>>>>>             lock(mmap_sem)
>>>>>>>>             vma = find_vma();
>>>>>>>>             handle_pte_fault(vma);
>>>>>>>>             if retry
>>>>>>>>                unlock(mmap_sem)
>>>>>>>>                goto again;
>>>>>>>>     done:
>>>>>>>>             handle fault error
>>>>>>>>     }
>>>>>>>>
>>>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>>>
>>>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>>>> the part of the faults processed speculatively.
>>>>>>>>
>>>>>>>> There are some trace events introduced by this series. They allow
>>>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>>>    back.
>>>>>>>>
>>>>>>>> To record all the related events, the easier is to run perf with the
>>>>>>>> following arguments :
>>>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>>>
>>>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>>>
>>>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>>>> on x86, PowerPC and arm64.
>>>>>>>>
>>>>>>>> ---------------------
>>>>>>>> Real Workload results
>>>>>>>>
>>>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>>>> this new version should not impact the performance boost seen.
>>>>>>>>
>>>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>>>> series:
>>>>>>>>                 vanilla         spf
>>>>>>>> faults          89.418          101.364         +13%
>>>>>>>> spf                n/a           97.989
>>>>>>>>
>>>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>>>> way.
>>>>>>>>
>>>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>>>> it a try on an android device. He reported that the application launch time
>>>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>>>> 20%.
>>>>>>>>
>>>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>>>
>>>>>>>> Application                             4.9     4.9+spf delta
>>>>>>>> com.tencent.mm                          416     389     -7%
>>>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>>>> com.tencent.mtt                         455     454     0%
>>>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>>>> com.immomo.momo                         501     487     -3%
>>>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>>>> com.sina.weibo                          907     906     0%
>>>>>>>> com.youku.phone                         816     731     -11%
>>>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>>>> com.UCMobile                            415     411     -1%
>>>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>>>
>>>>>>>> ------------------
>>>>>>>> Benchmarks results
>>>>>>>>
>>>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>>>> SPF is BASE + this series
>>>>>>>>
>>>>>>>> Kernbench:
>>>>>>>> ----------
>>>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>>>> kernel (kernel is build 5 times):
>>>>>>>>
>>>>>>>> Average Half load -j 8
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>>>
>>>>>>>> Average Optimal load -j 16
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>>>
>>>>>>>>
>>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>>          526743764      faults
>>>>>>>>                210      spf
>>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>>>> were created during the kernel build processing).
>>>>>>>>
>>>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>>>
>>>>>>>> Average Half load -j 40
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>>>
>>>>>>>> Average Optimal load -j 80
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>>>
>>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>>          116730856      faults
>>>>>>>>                  0      spf
>>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>>>> there is no impact on the performance.
>>>>>>>>
>>>>>>>> Ebizzy:
>>>>>>>> -------
>>>>>>>> The test is counting the number of records per second it can manage, the
>>>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>>>> result. The number is the record processes per second, the higher is the
>>>>>>>> best.
>>>>>>>>
>>>>>>>>                 BASE            SPF             delta
>>>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>>>
>>>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>>>            1706379      faults
>>>>>>>>            1674599      spf
>>>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>>>            1874773      faults
>>>>>>>>            1461153      spf
>>>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>>>> leading the ebizzy performance boost.
>>>>>>>>
>>>>>>>> ------------------
>>>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>>>    and Minchan Kim, hopefully.
>>>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>>>    __do_page_fault().
>>>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>>>    instead
>>>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>>>> useless
>>>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>>>
>>>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>>>
>>>>>>>>
>>>>>>>> Laurent Dufour (20):
>>>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>>>   mm: introduce INIT_VMA()
>>>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>>>   mm: introduce __vm_normal_page()
>>>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>>>   perf: add a speculative page fault sw event
>>>>>>>>   perf tools: add support for the SPF perf event
>>>>>>>>   mm: add speculative page fault vmstats
>>>>>>>>   powerpc/mm: add speculative page fault
>>>>>>>>
>>>>>>>> Mahendran Ganesh (2):
>>>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   arm64/mm: add speculative page fault
>>>>>>>>
>>>>>>>> Peter Zijlstra (4):
>>>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>>>   mm: VMA sequence count
>>>>>>>>   mm: provide speculative fault infrastructure
>>>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>>>
>>>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>>>  fs/exec.c                             |   2 +-
>>>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>>>  mm/internal.h                         |  20 ++
>>>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>>>  mm/mremap.c                           |  13 +
>>>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>>>  mm/swap.c                             |   6 +-
>>>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>>>
>>>>>>>> --
>>>>>>>> 2.7.4
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>


[-- Attachment #2: perf-profile_page_fault3-head-thp-always-SPF-off.gz --]
[-- Type: application/gzip, Size: 11278 bytes --]

[-- Attachment #3: perf-profile_page_fault3-head-thp-always-SPF-on.gz --]
[-- Type: application/gzip, Size: 11424 bytes --]

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

[-- Attachment #1: Type: text/plain, Size: 43707 bytes --]

On 03/08/2018 08:36, Song, HaiyanX wrote:
> Hi Laurent,

Hi Haiyan,

Sorry for the late answer, I was off a couple of days.

> 
> Thanks for your analysis for the last perf results.
> Your mentioned ," the major differences at the head of the perf report is the 92% testcase which is weirdly not reported
> on the head side", which is a bug of 0-day,and it caused the item is not counted in perf. 
> 
> I've triggered the test page_fault2 and page_fault3 again only with thread mode of will-it-scale on 0-day (on the same test box,every case tested 3 times).
> I checked the perf report have no above mentioned problem.
> 
> I have compared them, found some items have difference, such as below case:
>        page_fault2-thp-always: handle_mm_fault, base: 45.22%    head: 29.41%
>        page_fault3-thp-always: handle_mm_fault, base: 22.95%    head: 14.15%       

These would mean that the system spends lees time running handle_mm_fault()
when SPF is in the picture in this 2 cases which is good. This should lead to
better results with the SPF series, and I can't find any values higher on the
head side.

> 
> So i attached the perf result in mail again, could your have a look again for checking the difference between base and head commit.

I took a close look to all the perf result you sent, but I can't identify any
major difference. But the compiler optimization is getting rid of the
handle_pte_fault() symbol on the base kernel which add complexity to check the
differences.

To get rid of that, I'm proposing that you applied the attached patch to the
spf kernel. This patch is allowing to turn on/off the SPF handler through
/proc/sys/vm/speculative_page_fault.

This should ease the testing by limiting the reboot and avoid kernel's symbols
mismatch. Obviously there is still a small overhead due to the check but it
should not be viewable.

With this patch applied you can simply run
echo 1 > /proc/sys/vm/speculative_page_fault
to run a test with the speculative page fault handler activated. Or run
echo 0 > /proc/sys/vm/speculative_page_fault
to run a test without it.

I'm really sorry to asking that again, but could please run the test
page_fault3_base_THP-Always with and without SPF and capture the perf output.

I think we should focus on that test which showed the biggest regression.

Thanks,
Laurent.


> 
> Thanks,
> Haiyan, Song
>  
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Tuesday, July 17, 2018 5:36 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
> 
> On 13/07/2018 05:56, Song, HaiyanX wrote:
>> Hi Laurent,
> 
> Hi Haiyan,
> 
> Thanks a lot for sharing this perf reports.
> 
> I looked at them closely, and I've to admit that I was not able to found a
> major difference between the base and the head report, except that
> handle_pte_fault() is no more in-lined in the head one.
> 
> As expected, __handle_speculative_fault() is never traced since these tests are
> dealing with file mapping, not handled in the speculative way.
> 
> When running these test did you seen a major differences in the test's result
> between base and head ?
> 
> From the number of cycles counted, the biggest difference is page_fault3 when
> run with the THP enabled:
>                                 BASE            HEAD            Delta
> page_fault2_base_thp_never      1142252426747   1065866197589   -6.69%
> page_fault2_base_THP-Alwasys    1124844374523   1076312228927   -4.31%
> page_fault3_base_thp_never      1099387298152   1134118402345   3.16%
> page_fault3_base_THP-Always     1059370178101   853985561949    -19.39%
> 
> 
> The very weird thing is the difference of the delta cycles reported between
> thp never and thp always, because the speculative way is aborted when checking
> for the vma->ops field, which is the same in both case, and the thp is never
> checked. So there is no code covering differnce, on the speculative path,
> between these 2 cases. This leads me to think that there are other interactions
> interfering in the measure.
> 
> Looking at the perf-profile_page_fault3_*_THP-Always, the major differences at
> the head of the perf report is the 92% testcase which is weirdly not reported
> on the head side :
>     92.02%    22.33%  page_fault3_processes  [.] testcase
> 92.02% testcase
> 
> Then the base reported 37.67% for __do_page_fault() where the head reported
> 48.41%, but the only difference in this function, between base and head, is the
> call to handle_speculative_fault(). But this is a macro checking for the fault
> flags, and mm->users and then calling __handle_speculative_fault() if needed.
> So this can't explain this difference, except if __handle_speculative_fault()
> is inlined in __do_page_fault().
> Is this the case on your build ?
> 
> Haiyan, do you still have the output of the test to check those numbers too ?
> 
> Cheers,
> Laurent
> 
>> I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case.
>> Please help to check on these data if it can help you to find the higher change. Thanks.
>>
>> File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2
>>     tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.
>>
>> Best regards,
>> Haiyan Song
>>
>> ________________________________________
>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>> Sent: Thursday, July 12, 2018 1:05 AM
>> To: Song, HaiyanX
>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>
>> Hi Haiyan,
>>
>> Do you get a chance to capture some performance cycles on your system ?
>> I still can't get these numbers on my hardware.
>>
>> Thanks,
>> Laurent.
>>
>> On 04/07/2018 09:51, Laurent Dufour wrote:
>>> On 04/07/2018 05:23, Song, HaiyanX wrote:
>>>> Hi Laurent,
>>>>
>>>>
>>>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>>>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>>>
>>> Repeating the test only 3 times seems a bit too low to me.
>>>
>>> I'll focus on the higher change for the moment, but I don't have access to such
>>> a hardware.
>>>
>>> Is possible to provide a diff between base and SPF of the performance cycles
>>> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>>>
>>> Please stay focus on the test case process to see exactly where the series is
>>> impacting.
>>>
>>> Thanks,
>>> Laurent.
>>>
>>>>
>>>> And I did not find other high variation on test case result.
>>>>
>>>> a). Enable THP
>>>> testcase                          base     stddev       change      head     stddev         metric
>>>> page_fault3/enable THP           10519      +- 3%        -20.5%      8368      +-6%          will-it-scale.per_thread_ops
>>>> page_fault2/enalbe THP            8281      +- 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>>>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>>>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>>>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>>>
>>>> b). Disable THP
>>>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>>>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>>>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>>>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>>>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>>>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>>>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>>>
>>>>
>>>> Best regards,
>>>> Haiyan Song
>>>> ________________________________________
>>>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Monday, July 02, 2018 4:59 PM
>>>> To: Song, HaiyanX
>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>>>> V9 patch serials.
>>>>>
>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>>>> commit id:
>>>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>>>> Benchmark: will-it-scale
>>>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>>>
>>>>> Metrics:
>>>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>> THP: enable / disable
>>>>> nr_task:100%
>>>>>
>>>>> 1. Regressions:
>>>>>
>>>>> a). Enable THP
>>>>> testcase                          base           change      head           metric
>>>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>>>
>>>>> b). Disable THP
>>>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>>>
>>>>> Notes: for the above  values of test result, the higher is better.
>>>>
>>>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>>>> get reproducible results. The results have huge variation, even on the vanilla
>>>> kernel, and I can't state on any changes due to that.
>>>>
>>>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>>>> measure any changes between the vanilla and the SPF patched ones:
>>>>
>>>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>>>> page_fault3_threads             2697.7          2683.5          -0.53%
>>>> page_fault2_threads             170660.6        169574.1        -0.64%
>>>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>>>> context_switch1_processes       6478076.2       6529493.5       0.79%
>>>> brk1                            243391.2        238527.5        -2.00%
>>>>
>>>> Tests were run 10 times, no high variation detected.
>>>>
>>>> Did you see high variation on your side ? How many times the test were run to
>>>> compute the average values ?
>>>>
>>>> Thanks,
>>>> Laurent.
>>>>
>>>>
>>>>>
>>>>> 2. Improvement: not found improvement based on the selected test cases.
>>>>>
>>>>>
>>>>> Best regards
>>>>> Haiyan Song
>>>>> ________________________________________
>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>> Sent: Monday, May 28, 2018 4:54 PM
>>>>> To: Song, HaiyanX
>>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>>
>>>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>>>> Hi Laurent,
>>>>>>
>>>>>> Yes, these tests are done on V9 patch.
>>>>>
>>>>> Do you plan to give this V11 a run ?
>>>>>
>>>>>>
>>>>>>
>>>>>> Best regards,
>>>>>> Haiyan Song
>>>>>>
>>>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>>>
>>>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>>>> tested on Intel 4s Skylake platform.
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>>>> series" while responding to the v11 header series...
>>>>>>> Were these tests done on v9 or v11 ?
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Laurent.
>>>>>>>
>>>>>>>>
>>>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>>>> Commit id:
>>>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>>>> Benchmark suite: will-it-scale
>>>>>>>> Download link:
>>>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>>>> Metrics:
>>>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>>>> THP: enable / disable
>>>>>>>> nr_task: 100%
>>>>>>>>
>>>>>>>> 1. Regressions:
>>>>>>>> a) THP enabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>>>
>>>>>>>> b) THP disabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>>>
>>>>>>>> 2. Improvements:
>>>>>>>> a) THP enabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>>>
>>>>>>>> b) THP disabled:
>>>>>>>> testcase                        base            change          head       metric
>>>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>>>
>>>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>>>
>>>>>>>>
>>>>>>>> Best regards
>>>>>>>> Haiyan Song
>>>>>>>>
>>>>>>>> ________________________________________
>>>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>>>
>>>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>>>> page fault without holding the mm semaphore [1].
>>>>>>>>
>>>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>>>
>>>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>>>> limiting the locking contention to these operations which are expected to
>>>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>>>> benchmark anymore.
>>>>>>>>
>>>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>>>> speculative page fault in that case.
>>>>>>>>
>>>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>>>> checked during the page fault are modified.
>>>>>>>>
>>>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>>>> parallel change is possible at this time.
>>>>>>>>
>>>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>>>> classic page fault handler will be called to handle the operation while
>>>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>>>> PTE.
>>>>>>>>
>>>>>>>> In pseudo code, this could be seen as:
>>>>>>>>     speculative_page_fault()
>>>>>>>>     {
>>>>>>>>             vma = get_vma()
>>>>>>>>             check vma sequence count
>>>>>>>>             check vma's support
>>>>>>>>             disable interrupt
>>>>>>>>                   check pgd,p4d,...,pte
>>>>>>>>                   save pmd and pte in vmf
>>>>>>>>                   save vma sequence counter in vmf
>>>>>>>>             enable interrupt
>>>>>>>>             check vma sequence count
>>>>>>>>             handle_pte_fault(vma)
>>>>>>>>                     ..
>>>>>>>>                     page = alloc_page()
>>>>>>>>                     pte_map_lock()
>>>>>>>>                             disable interrupt
>>>>>>>>                                     abort if sequence counter has changed
>>>>>>>>                                     abort if pmd or pte has changed
>>>>>>>>                                     pte map and lock
>>>>>>>>                             enable interrupt
>>>>>>>>                     if abort
>>>>>>>>                        free page
>>>>>>>>                        abort
>>>>>>>>                     ...
>>>>>>>>     }
>>>>>>>>
>>>>>>>>     arch_fault_handler()
>>>>>>>>     {
>>>>>>>>             if (speculative_page_fault(&vma))
>>>>>>>>                goto done
>>>>>>>>     again:
>>>>>>>>             lock(mmap_sem)
>>>>>>>>             vma = find_vma();
>>>>>>>>             handle_pte_fault(vma);
>>>>>>>>             if retry
>>>>>>>>                unlock(mmap_sem)
>>>>>>>>                goto again;
>>>>>>>>     done:
>>>>>>>>             handle fault error
>>>>>>>>     }
>>>>>>>>
>>>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>>>
>>>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>>>> the part of the faults processed speculatively.
>>>>>>>>
>>>>>>>> There are some trace events introduced by this series. They allow
>>>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>>>    back.
>>>>>>>>
>>>>>>>> To record all the related events, the easier is to run perf with the
>>>>>>>> following arguments :
>>>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>>>
>>>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>>>
>>>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>>>> on x86, PowerPC and arm64.
>>>>>>>>
>>>>>>>> ---------------------
>>>>>>>> Real Workload results
>>>>>>>>
>>>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>>>> this new version should not impact the performance boost seen.
>>>>>>>>
>>>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>>>> series:
>>>>>>>>                 vanilla         spf
>>>>>>>> faults          89.418          101.364         +13%
>>>>>>>> spf                n/a           97.989
>>>>>>>>
>>>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>>>> way.
>>>>>>>>
>>>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>>>> it a try on an android device. He reported that the application launch time
>>>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>>>> 20%.
>>>>>>>>
>>>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>>>
>>>>>>>> Application                             4.9     4.9+spf delta
>>>>>>>> com.tencent.mm                          416     389     -7%
>>>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>>>> com.tencent.mtt                         455     454     0%
>>>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>>>> com.immomo.momo                         501     487     -3%
>>>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>>>> com.sina.weibo                          907     906     0%
>>>>>>>> com.youku.phone                         816     731     -11%
>>>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>>>> com.UCMobile                            415     411     -1%
>>>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>>>
>>>>>>>> ------------------
>>>>>>>> Benchmarks results
>>>>>>>>
>>>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>>>> SPF is BASE + this series
>>>>>>>>
>>>>>>>> Kernbench:
>>>>>>>> ----------
>>>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>>>> kernel (kernel is build 5 times):
>>>>>>>>
>>>>>>>> Average Half load -j 8
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>>>
>>>>>>>> Average Optimal load -j 16
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>>>
>>>>>>>>
>>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>>          526743764      faults
>>>>>>>>                210      spf
>>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>>>> were created during the kernel build processing).
>>>>>>>>
>>>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>>>
>>>>>>>> Average Half load -j 40
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>>>
>>>>>>>> Average Optimal load -j 80
>>>>>>>>                  Run    (std deviation)
>>>>>>>>                  BASE                   SPF
>>>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>>>
>>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>>          116730856      faults
>>>>>>>>                  0      spf
>>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>>>> there is no impact on the performance.
>>>>>>>>
>>>>>>>> Ebizzy:
>>>>>>>> -------
>>>>>>>> The test is counting the number of records per second it can manage, the
>>>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>>>> result. The number is the record processes per second, the higher is the
>>>>>>>> best.
>>>>>>>>
>>>>>>>>                 BASE            SPF             delta
>>>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>>>
>>>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>>>            1706379      faults
>>>>>>>>            1674599      spf
>>>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>>>            1874773      faults
>>>>>>>>            1461153      spf
>>>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>>
>>>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>>>> leading the ebizzy performance boost.
>>>>>>>>
>>>>>>>> ------------------
>>>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>>>    and Minchan Kim, hopefully.
>>>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>>>    __do_page_fault().
>>>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>>>    instead
>>>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>>>> useless
>>>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>>>
>>>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>>>
>>>>>>>>
>>>>>>>> Laurent Dufour (20):
>>>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>>>   mm: introduce INIT_VMA()
>>>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>>>   mm: introduce __vm_normal_page()
>>>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>>>   perf: add a speculative page fault sw event
>>>>>>>>   perf tools: add support for the SPF perf event
>>>>>>>>   mm: add speculative page fault vmstats
>>>>>>>>   powerpc/mm: add speculative page fault
>>>>>>>>
>>>>>>>> Mahendran Ganesh (2):
>>>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>>   arm64/mm: add speculative page fault
>>>>>>>>
>>>>>>>> Peter Zijlstra (4):
>>>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>>>   mm: VMA sequence count
>>>>>>>>   mm: provide speculative fault infrastructure
>>>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>>>
>>>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>>>  fs/exec.c                             |   2 +-
>>>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>>>  mm/internal.h                         |  20 ++
>>>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>>>  mm/mremap.c                           |  13 +
>>>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>>>  mm/swap.c                             |   6 +-
>>>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>>>
>>>>>>>> --
>>>>>>>> 2.7.4
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>
>>
> 


[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0001-mm-Add-a-speculative-page-fault-switch-in-sysctl.patch --]
[-- Type: text/x-patch; name="0001-mm-Add-a-speculative-page-fault-switch-in-sysctl.patch", Size: 0 bytes --]

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

[-- Attachment #1: Type: text/plain, Size: 43157 bytes --]

Add another 3 perf file.
________________________________________
From: Song, HaiyanX
Sent: Friday, August 03, 2018 2:36 PM
To: Laurent Dufour
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: RE: [PATCH v11 00/26] Speculative page faults

Hi Laurent,

Thanks for your analysis for the last perf results.
Your mentioned ," the major differences at the head of the perf report is the 92% testcase which is weirdly not reported
on the head side", which is a bug of 0-day,and it caused the item is not counted in perf.

I've triggered the test page_fault2 and page_fault3 again only with thread mode of will-it-scale on 0-day (on the same test box,every case tested 3 times).
I checked the perf report have no above mentioned problem.

I have compared them, found some items have difference, such as below case:
       page_fault2-thp-always: handle_mm_fault, base: 45.22%    head: 29.41%
       page_fault3-thp-always: handle_mm_fault, base: 22.95%    head: 14.15%

So i attached the perf result in mail again, could your have a look again for checking the difference between base and head commit.


Thanks,
Haiyan, Song

________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Tuesday, July 17, 2018 5:36 PM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 13/07/2018 05:56, Song, HaiyanX wrote:
> Hi Laurent,

Hi Haiyan,

Thanks a lot for sharing this perf reports.

I looked at them closely, and I've to admit that I was not able to found a
major difference between the base and the head report, except that
handle_pte_fault() is no more in-lined in the head one.

As expected, __handle_speculative_fault() is never traced since these tests are
dealing with file mapping, not handled in the speculative way.

When running these test did you seen a major differences in the test's result
between base and head ?

>From the number of cycles counted, the biggest difference is page_fault3 when
run with the THP enabled:
                                BASE            HEAD            Delta
page_fault2_base_thp_never      1142252426747   1065866197589   -6.69%
page_fault2_base_THP-Alwasys    1124844374523   1076312228927   -4.31%
page_fault3_base_thp_never      1099387298152   1134118402345   3.16%
page_fault3_base_THP-Always     1059370178101   853985561949    -19.39%


The very weird thing is the difference of the delta cycles reported between
thp never and thp always, because the speculative way is aborted when checking
for the vma->ops field, which is the same in both case, and the thp is never
checked. So there is no code covering differnce, on the speculative path,
between these 2 cases. This leads me to think that there are other interactions
interfering in the measure.

Looking at the perf-profile_page_fault3_*_THP-Always, the major differences at
the head of the perf report is the 92% testcase which is weirdly not reported
on the head side :
    92.02%    22.33%  page_fault3_processes  [.] testcase
92.02% testcase

Then the base reported 37.67% for __do_page_fault() where the head reported
48.41%, but the only difference in this function, between base and head, is the
call to handle_speculative_fault(). But this is a macro checking for the fault
flags, and mm->users and then calling __handle_speculative_fault() if needed.
So this can't explain this difference, except if __handle_speculative_fault()
is inlined in __do_page_fault().
Is this the case on your build ?

Haiyan, do you still have the output of the test to check those numbers too ?

Cheers,
Laurent

> I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case.
> Please help to check on these data if it can help you to find the higher change. Thanks.
>
> File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2
>     tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.
>
> Best regards,
> Haiyan Song
>
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Thursday, July 12, 2018 1:05 AM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
>
> Hi Haiyan,
>
> Do you get a chance to capture some performance cycles on your system ?
> I still can't get these numbers on my hardware.
>
> Thanks,
> Laurent.
>
> On 04/07/2018 09:51, Laurent Dufour wrote:
>> On 04/07/2018 05:23, Song, HaiyanX wrote:
>>> Hi Laurent,
>>>
>>>
>>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>>
>> Repeating the test only 3 times seems a bit too low to me.
>>
>> I'll focus on the higher change for the moment, but I don't have access to such
>> a hardware.
>>
>> Is possible to provide a diff between base and SPF of the performance cycles
>> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>>
>> Please stay focus on the test case process to see exactly where the series is
>> impacting.
>>
>> Thanks,
>> Laurent.
>>
>>>
>>> And I did not find other high variation on test case result.
>>>
>>> a). Enable THP
>>> testcase                          base     stddev       change      head     stddev         metric
>>> page_fault3/enable THP           10519      ± 3%        -20.5%      8368      ±6%          will-it-scale.per_thread_ops
>>> page_fault2/enalbe THP            8281      ± 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>>
>>> b). Disable THP
>>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>>
>>>
>>> Best regards,
>>> Haiyan Song
>>> ________________________________________
>>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Monday, July 02, 2018 4:59 PM
>>> To: Song, HaiyanX
>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>
>>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>>> Hi Laurent,
>>>>
>>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>>> V9 patch serials.
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>>> commit id:
>>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>>> Benchmark: will-it-scale
>>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>>
>>>> Metrics:
>>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task:100%
>>>>
>>>> 1. Regressions:
>>>>
>>>> a). Enable THP
>>>> testcase                          base           change      head           metric
>>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>>
>>>> b). Disable THP
>>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>>
>>>> Notes: for the above  values of test result, the higher is better.
>>>
>>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>>> get reproducible results. The results have huge variation, even on the vanilla
>>> kernel, and I can't state on any changes due to that.
>>>
>>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>>> measure any changes between the vanilla and the SPF patched ones:
>>>
>>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>>> page_fault3_threads             2697.7          2683.5          -0.53%
>>> page_fault2_threads             170660.6        169574.1        -0.64%
>>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>>> context_switch1_processes       6478076.2       6529493.5       0.79%
>>> brk1                            243391.2        238527.5        -2.00%
>>>
>>> Tests were run 10 times, no high variation detected.
>>>
>>> Did you see high variation on your side ? How many times the test were run to
>>> compute the average values ?
>>>
>>> Thanks,
>>> Laurent.
>>>
>>>
>>>>
>>>> 2. Improvement: not found improvement based on the selected test cases.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Monday, May 28, 2018 4:54 PM
>>>> To: Song, HaiyanX
>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> Yes, these tests are done on V9 patch.
>>>>
>>>> Do you plan to give this V11 a run ?
>>>>
>>>>>
>>>>>
>>>>> Best regards,
>>>>> Haiyan Song
>>>>>
>>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>>
>>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>>> tested on Intel 4s Skylake platform.
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>>> series" while responding to the v11 header series...
>>>>>> Were these tests done on v9 or v11 ?
>>>>>>
>>>>>> Cheers,
>>>>>> Laurent.
>>>>>>
>>>>>>>
>>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>>> Commit id:
>>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>>> Benchmark suite: will-it-scale
>>>>>>> Download link:
>>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>>> Metrics:
>>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>>> THP: enable / disable
>>>>>>> nr_task: 100%
>>>>>>>
>>>>>>> 1. Regressions:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> 2. Improvements:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>>
>>>>>>>
>>>>>>> Best regards
>>>>>>> Haiyan Song
>>>>>>>
>>>>>>> ________________________________________
>>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>>
>>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>>> page fault without holding the mm semaphore [1].
>>>>>>>
>>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>>
>>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>>> limiting the locking contention to these operations which are expected to
>>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>>> benchmark anymore.
>>>>>>>
>>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>>> speculative page fault in that case.
>>>>>>>
>>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>>> checked during the page fault are modified.
>>>>>>>
>>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>>> parallel change is possible at this time.
>>>>>>>
>>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>>> classic page fault handler will be called to handle the operation while
>>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>>> PTE.
>>>>>>>
>>>>>>> In pseudo code, this could be seen as:
>>>>>>>     speculative_page_fault()
>>>>>>>     {
>>>>>>>             vma = get_vma()
>>>>>>>             check vma sequence count
>>>>>>>             check vma's support
>>>>>>>             disable interrupt
>>>>>>>                   check pgd,p4d,...,pte
>>>>>>>                   save pmd and pte in vmf
>>>>>>>                   save vma sequence counter in vmf
>>>>>>>             enable interrupt
>>>>>>>             check vma sequence count
>>>>>>>             handle_pte_fault(vma)
>>>>>>>                     ..
>>>>>>>                     page = alloc_page()
>>>>>>>                     pte_map_lock()
>>>>>>>                             disable interrupt
>>>>>>>                                     abort if sequence counter has changed
>>>>>>>                                     abort if pmd or pte has changed
>>>>>>>                                     pte map and lock
>>>>>>>                             enable interrupt
>>>>>>>                     if abort
>>>>>>>                        free page
>>>>>>>                        abort
>>>>>>>                     ...
>>>>>>>     }
>>>>>>>
>>>>>>>     arch_fault_handler()
>>>>>>>     {
>>>>>>>             if (speculative_page_fault(&vma))
>>>>>>>                goto done
>>>>>>>     again:
>>>>>>>             lock(mmap_sem)
>>>>>>>             vma = find_vma();
>>>>>>>             handle_pte_fault(vma);
>>>>>>>             if retry
>>>>>>>                unlock(mmap_sem)
>>>>>>>                goto again;
>>>>>>>     done:
>>>>>>>             handle fault error
>>>>>>>     }
>>>>>>>
>>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>>
>>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>>> the part of the faults processed speculatively.
>>>>>>>
>>>>>>> There are some trace events introduced by this series. They allow
>>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>>    back.
>>>>>>>
>>>>>>> To record all the related events, the easier is to run perf with the
>>>>>>> following arguments :
>>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>>
>>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>>
>>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>>> on x86, PowerPC and arm64.
>>>>>>>
>>>>>>> ---------------------
>>>>>>> Real Workload results
>>>>>>>
>>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>>> this new version should not impact the performance boost seen.
>>>>>>>
>>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>>> series:
>>>>>>>                 vanilla         spf
>>>>>>> faults          89.418          101.364         +13%
>>>>>>> spf                n/a           97.989
>>>>>>>
>>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>>> way.
>>>>>>>
>>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>>> it a try on an android device. He reported that the application launch time
>>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>>> 20%.
>>>>>>>
>>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>>
>>>>>>> Application                             4.9     4.9+spf delta
>>>>>>> com.tencent.mm                          416     389     -7%
>>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>>> com.tencent.mtt                         455     454     0%
>>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>>> com.immomo.momo                         501     487     -3%
>>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>>> com.sina.weibo                          907     906     0%
>>>>>>> com.youku.phone                         816     731     -11%
>>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>>> com.UCMobile                            415     411     -1%
>>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>>
>>>>>>> ------------------
>>>>>>> Benchmarks results
>>>>>>>
>>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>>> SPF is BASE + this series
>>>>>>>
>>>>>>> Kernbench:
>>>>>>> ----------
>>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>>> kernel (kernel is build 5 times):
>>>>>>>
>>>>>>> Average Half load -j 8
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>>
>>>>>>> Average Optimal load -j 16
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>>
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          526743764      faults
>>>>>>>                210      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>>> were created during the kernel build processing).
>>>>>>>
>>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>>
>>>>>>> Average Half load -j 40
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>>
>>>>>>> Average Optimal load -j 80
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          116730856      faults
>>>>>>>                  0      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>>> there is no impact on the performance.
>>>>>>>
>>>>>>> Ebizzy:
>>>>>>> -------
>>>>>>> The test is counting the number of records per second it can manage, the
>>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>>> result. The number is the record processes per second, the higher is the
>>>>>>> best.
>>>>>>>
>>>>>>>                 BASE            SPF             delta
>>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>>
>>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>>            1706379      faults
>>>>>>>            1674599      spf
>>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>>            1874773      faults
>>>>>>>            1461153      spf
>>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>>> leading the ebizzy performance boost.
>>>>>>>
>>>>>>> ------------------
>>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>>    and Minchan Kim, hopefully.
>>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>>    __do_page_fault().
>>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>>    instead
>>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>>> useless
>>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>>
>>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>>
>>>>>>>
>>>>>>> Laurent Dufour (20):
>>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>>   mm: introduce INIT_VMA()
>>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>>   mm: introduce __vm_normal_page()
>>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>>   perf: add a speculative page fault sw event
>>>>>>>   perf tools: add support for the SPF perf event
>>>>>>>   mm: add speculative page fault vmstats
>>>>>>>   powerpc/mm: add speculative page fault
>>>>>>>
>>>>>>> Mahendran Ganesh (2):
>>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   arm64/mm: add speculative page fault
>>>>>>>
>>>>>>> Peter Zijlstra (4):
>>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: VMA sequence count
>>>>>>>   mm: provide speculative fault infrastructure
>>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>>
>>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>>  fs/exec.c                             |   2 +-
>>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>>  mm/internal.h                         |  20 ++
>>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>>  mm/mremap.c                           |  13 +
>>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>>  mm/swap.c                             |   6 +-
>>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>>
>>>>>>> --
>>>>>>> 2.7.4
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>>
>>
>


[-- Attachment #2: perf-profile_page_fault3_head_thp_always.gz --]
[-- Type: application/gzip, Size: 12909 bytes --]

[-- Attachment #3: perf-profile_page_fault3_head_thp_never.gz --]
[-- Type: application/gzip, Size: 12535 bytes --]

[-- Attachment #4: perf-profile_page_fault2_head_thp_never.gz --]
[-- Type: application/gzip, Size: 11782 bytes --]

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

[-- Attachment #1: Type: text/plain, Size: 42188 bytes --]

Hi Laurent,

Thanks for your analysis for the last perf results.
Your mentioned ," the major differences at the head of the perf report is the 92% testcase which is weirdly not reported
on the head side", which is a bug of 0-day,and it caused the item is not counted in perf. 

I've triggered the test page_fault2 and page_fault3 again only with thread mode of will-it-scale on 0-day (on the same test box,every case tested 3 times).
I checked the perf report have no above mentioned problem.

I have compared them, found some items have difference, such as below case:
       page_fault2-thp-always: handle_mm_fault, base: 45.22%    head: 29.41%
       page_fault3-thp-always: handle_mm_fault, base: 22.95%    head: 14.15%       

So i attached the perf result in mail again, could your have a look again for checking the difference between base and head commit.


Thanks,
Haiyan, Song
 
________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Tuesday, July 17, 2018 5:36 PM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 13/07/2018 05:56, Song, HaiyanX wrote:
> Hi Laurent,

Hi Haiyan,

Thanks a lot for sharing this perf reports.

I looked at them closely, and I've to admit that I was not able to found a
major difference between the base and the head report, except that
handle_pte_fault() is no more in-lined in the head one.

As expected, __handle_speculative_fault() is never traced since these tests are
dealing with file mapping, not handled in the speculative way.

When running these test did you seen a major differences in the test's result
between base and head ?

>From the number of cycles counted, the biggest difference is page_fault3 when
run with the THP enabled:
                                BASE            HEAD            Delta
page_fault2_base_thp_never      1142252426747   1065866197589   -6.69%
page_fault2_base_THP-Alwasys    1124844374523   1076312228927   -4.31%
page_fault3_base_thp_never      1099387298152   1134118402345   3.16%
page_fault3_base_THP-Always     1059370178101   853985561949    -19.39%


The very weird thing is the difference of the delta cycles reported between
thp never and thp always, because the speculative way is aborted when checking
for the vma->ops field, which is the same in both case, and the thp is never
checked. So there is no code covering differnce, on the speculative path,
between these 2 cases. This leads me to think that there are other interactions
interfering in the measure.

Looking at the perf-profile_page_fault3_*_THP-Always, the major differences at
the head of the perf report is the 92% testcase which is weirdly not reported
on the head side :
    92.02%    22.33%  page_fault3_processes  [.] testcase
92.02% testcase

Then the base reported 37.67% for __do_page_fault() where the head reported
48.41%, but the only difference in this function, between base and head, is the
call to handle_speculative_fault(). But this is a macro checking for the fault
flags, and mm->users and then calling __handle_speculative_fault() if needed.
So this can't explain this difference, except if __handle_speculative_fault()
is inlined in __do_page_fault().
Is this the case on your build ?

Haiyan, do you still have the output of the test to check those numbers too ?

Cheers,
Laurent

> I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case.
> Please help to check on these data if it can help you to find the higher change. Thanks.
>
> File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2
>     tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.
>
> Best regards,
> Haiyan Song
>
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Thursday, July 12, 2018 1:05 AM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
>
> Hi Haiyan,
>
> Do you get a chance to capture some performance cycles on your system ?
> I still can't get these numbers on my hardware.
>
> Thanks,
> Laurent.
>
> On 04/07/2018 09:51, Laurent Dufour wrote:
>> On 04/07/2018 05:23, Song, HaiyanX wrote:
>>> Hi Laurent,
>>>
>>>
>>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>>
>> Repeating the test only 3 times seems a bit too low to me.
>>
>> I'll focus on the higher change for the moment, but I don't have access to such
>> a hardware.
>>
>> Is possible to provide a diff between base and SPF of the performance cycles
>> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>>
>> Please stay focus on the test case process to see exactly where the series is
>> impacting.
>>
>> Thanks,
>> Laurent.
>>
>>>
>>> And I did not find other high variation on test case result.
>>>
>>> a). Enable THP
>>> testcase                          base     stddev       change      head     stddev         metric
>>> page_fault3/enable THP           10519      ± 3%        -20.5%      8368      ±6%          will-it-scale.per_thread_ops
>>> page_fault2/enalbe THP            8281      ± 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>>
>>> b). Disable THP
>>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>>
>>>
>>> Best regards,
>>> Haiyan Song
>>> ________________________________________
>>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Monday, July 02, 2018 4:59 PM
>>> To: Song, HaiyanX
>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>
>>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>>> Hi Laurent,
>>>>
>>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>>> V9 patch serials.
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>>> commit id:
>>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>>> Benchmark: will-it-scale
>>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>>
>>>> Metrics:
>>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task:100%
>>>>
>>>> 1. Regressions:
>>>>
>>>> a). Enable THP
>>>> testcase                          base           change      head           metric
>>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>>
>>>> b). Disable THP
>>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>>
>>>> Notes: for the above  values of test result, the higher is better.
>>>
>>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>>> get reproducible results. The results have huge variation, even on the vanilla
>>> kernel, and I can't state on any changes due to that.
>>>
>>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>>> measure any changes between the vanilla and the SPF patched ones:
>>>
>>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>>> page_fault3_threads             2697.7          2683.5          -0.53%
>>> page_fault2_threads             170660.6        169574.1        -0.64%
>>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>>> context_switch1_processes       6478076.2       6529493.5       0.79%
>>> brk1                            243391.2        238527.5        -2.00%
>>>
>>> Tests were run 10 times, no high variation detected.
>>>
>>> Did you see high variation on your side ? How many times the test were run to
>>> compute the average values ?
>>>
>>> Thanks,
>>> Laurent.
>>>
>>>
>>>>
>>>> 2. Improvement: not found improvement based on the selected test cases.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Monday, May 28, 2018 4:54 PM
>>>> To: Song, HaiyanX
>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> Yes, these tests are done on V9 patch.
>>>>
>>>> Do you plan to give this V11 a run ?
>>>>
>>>>>
>>>>>
>>>>> Best regards,
>>>>> Haiyan Song
>>>>>
>>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>>
>>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>>> tested on Intel 4s Skylake platform.
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>>> series" while responding to the v11 header series...
>>>>>> Were these tests done on v9 or v11 ?
>>>>>>
>>>>>> Cheers,
>>>>>> Laurent.
>>>>>>
>>>>>>>
>>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>>> Commit id:
>>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>>> Benchmark suite: will-it-scale
>>>>>>> Download link:
>>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>>> Metrics:
>>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>>> THP: enable / disable
>>>>>>> nr_task: 100%
>>>>>>>
>>>>>>> 1. Regressions:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> 2. Improvements:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>>
>>>>>>>
>>>>>>> Best regards
>>>>>>> Haiyan Song
>>>>>>>
>>>>>>> ________________________________________
>>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>>
>>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>>> page fault without holding the mm semaphore [1].
>>>>>>>
>>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>>
>>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>>> limiting the locking contention to these operations which are expected to
>>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>>> benchmark anymore.
>>>>>>>
>>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>>> speculative page fault in that case.
>>>>>>>
>>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>>> checked during the page fault are modified.
>>>>>>>
>>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>>> parallel change is possible at this time.
>>>>>>>
>>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>>> classic page fault handler will be called to handle the operation while
>>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>>> PTE.
>>>>>>>
>>>>>>> In pseudo code, this could be seen as:
>>>>>>>     speculative_page_fault()
>>>>>>>     {
>>>>>>>             vma = get_vma()
>>>>>>>             check vma sequence count
>>>>>>>             check vma's support
>>>>>>>             disable interrupt
>>>>>>>                   check pgd,p4d,...,pte
>>>>>>>                   save pmd and pte in vmf
>>>>>>>                   save vma sequence counter in vmf
>>>>>>>             enable interrupt
>>>>>>>             check vma sequence count
>>>>>>>             handle_pte_fault(vma)
>>>>>>>                     ..
>>>>>>>                     page = alloc_page()
>>>>>>>                     pte_map_lock()
>>>>>>>                             disable interrupt
>>>>>>>                                     abort if sequence counter has changed
>>>>>>>                                     abort if pmd or pte has changed
>>>>>>>                                     pte map and lock
>>>>>>>                             enable interrupt
>>>>>>>                     if abort
>>>>>>>                        free page
>>>>>>>                        abort
>>>>>>>                     ...
>>>>>>>     }
>>>>>>>
>>>>>>>     arch_fault_handler()
>>>>>>>     {
>>>>>>>             if (speculative_page_fault(&vma))
>>>>>>>                goto done
>>>>>>>     again:
>>>>>>>             lock(mmap_sem)
>>>>>>>             vma = find_vma();
>>>>>>>             handle_pte_fault(vma);
>>>>>>>             if retry
>>>>>>>                unlock(mmap_sem)
>>>>>>>                goto again;
>>>>>>>     done:
>>>>>>>             handle fault error
>>>>>>>     }
>>>>>>>
>>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>>
>>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>>> the part of the faults processed speculatively.
>>>>>>>
>>>>>>> There are some trace events introduced by this series. They allow
>>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>>    back.
>>>>>>>
>>>>>>> To record all the related events, the easier is to run perf with the
>>>>>>> following arguments :
>>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>>
>>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>>
>>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>>> on x86, PowerPC and arm64.
>>>>>>>
>>>>>>> ---------------------
>>>>>>> Real Workload results
>>>>>>>
>>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>>> this new version should not impact the performance boost seen.
>>>>>>>
>>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>>> series:
>>>>>>>                 vanilla         spf
>>>>>>> faults          89.418          101.364         +13%
>>>>>>> spf                n/a           97.989
>>>>>>>
>>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>>> way.
>>>>>>>
>>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>>> it a try on an android device. He reported that the application launch time
>>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>>> 20%.
>>>>>>>
>>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>>
>>>>>>> Application                             4.9     4.9+spf delta
>>>>>>> com.tencent.mm                          416     389     -7%
>>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>>> com.tencent.mtt                         455     454     0%
>>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>>> com.immomo.momo                         501     487     -3%
>>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>>> com.sina.weibo                          907     906     0%
>>>>>>> com.youku.phone                         816     731     -11%
>>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>>> com.UCMobile                            415     411     -1%
>>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>>
>>>>>>> ------------------
>>>>>>> Benchmarks results
>>>>>>>
>>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>>> SPF is BASE + this series
>>>>>>>
>>>>>>> Kernbench:
>>>>>>> ----------
>>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>>> kernel (kernel is build 5 times):
>>>>>>>
>>>>>>> Average Half load -j 8
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>>
>>>>>>> Average Optimal load -j 16
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>>
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          526743764      faults
>>>>>>>                210      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>>> were created during the kernel build processing).
>>>>>>>
>>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>>
>>>>>>> Average Half load -j 40
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>>
>>>>>>> Average Optimal load -j 80
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          116730856      faults
>>>>>>>                  0      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>>> there is no impact on the performance.
>>>>>>>
>>>>>>> Ebizzy:
>>>>>>> -------
>>>>>>> The test is counting the number of records per second it can manage, the
>>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>>> result. The number is the record processes per second, the higher is the
>>>>>>> best.
>>>>>>>
>>>>>>>                 BASE            SPF             delta
>>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>>
>>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>>            1706379      faults
>>>>>>>            1674599      spf
>>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>>            1874773      faults
>>>>>>>            1461153      spf
>>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>>> leading the ebizzy performance boost.
>>>>>>>
>>>>>>> ------------------
>>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>>    and Minchan Kim, hopefully.
>>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>>    __do_page_fault().
>>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>>    instead
>>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>>> useless
>>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>>
>>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>>
>>>>>>>
>>>>>>> Laurent Dufour (20):
>>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>>   mm: introduce INIT_VMA()
>>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>>   mm: introduce __vm_normal_page()
>>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>>   perf: add a speculative page fault sw event
>>>>>>>   perf tools: add support for the SPF perf event
>>>>>>>   mm: add speculative page fault vmstats
>>>>>>>   powerpc/mm: add speculative page fault
>>>>>>>
>>>>>>> Mahendran Ganesh (2):
>>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   arm64/mm: add speculative page fault
>>>>>>>
>>>>>>> Peter Zijlstra (4):
>>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: VMA sequence count
>>>>>>>   mm: provide speculative fault infrastructure
>>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>>
>>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>>  fs/exec.c                             |   2 +-
>>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>>  mm/internal.h                         |  20 ++
>>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>>  mm/mremap.c                           |  13 +
>>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>>  mm/swap.c                             |   6 +-
>>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>>
>>>>>>> --
>>>>>>> 2.7.4
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>>
>>
>


[-- Attachment #2: perf-profile_page_fault2_base_thp_always.gz --]
[-- Type: application/gzip, Size: 12167 bytes --]

[-- Attachment #3: perf-profile_page_fault2_base_thp_never.gz --]
[-- Type: application/gzip, Size: 11543 bytes --]

[-- Attachment #4: perf-profile_page_fault2_head_thp_always.gz --]
[-- Type: application/gzip, Size: 12019 bytes --]

[-- Attachment #5: perf-profile_page_fault3_base_thp_always.gz --]
[-- Type: application/gzip, Size: 12701 bytes --]

[-- Attachment #6: perf-profile_page_fault3_base_thp_always.gz --]
[-- Type: application/gzip, Size: 12701 bytes --]

[-- Attachment #7: perf-profile_page_fault3_base_thp_never.gz --]
[-- Type: application/gzip, Size: 12699 bytes --]

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

On 13/07/2018 05:56, Song, HaiyanX wrote:
> Hi Laurent,

Hi Haiyan,

Thanks a lot for sharing this perf reports.

I looked at them closely, and I've to admit that I was not able to found a
major difference between the base and the head report, except that
handle_pte_fault() is no more in-lined in the head one.

As expected, __handle_speculative_fault() is never traced since these tests are
dealing with file mapping, not handled in the speculative way.

When running these test did you seen a major differences in the test's result
between base and head ?

>From the number of cycles counted, the biggest difference is page_fault3 when
run with the THP enabled:
				BASE		HEAD		Delta
page_fault2_base_thp_never	1142252426747	1065866197589	-6.69%
page_fault2_base_THP-Alwasys	1124844374523	1076312228927	-4.31%
page_fault3_base_thp_never	1099387298152	1134118402345	3.16%
page_fault3_base_THP-Always	1059370178101	853985561949	-19.39%


The very weird thing is the difference of the delta cycles reported between
thp never and thp always, because the speculative way is aborted when checking
for the vma->ops field, which is the same in both case, and the thp is never
checked. So there is no code covering differnce, on the speculative path,
between these 2 cases. This leads me to think that there are other interactions
interfering in the measure.

Looking at the perf-profile_page_fault3_*_THP-Always, the major differences at
the head of the perf report is the 92% testcase which is weirdly not reported
on the head side :
    92.02%    22.33%  page_fault3_processes  [.] testcase
92.02% testcase

Then the base reported 37.67% for __do_page_fault() where the head reported
48.41%, but the only difference in this function, between base and head, is the
call to handle_speculative_fault(). But this is a macro checking for the fault
flags, and mm->users and then calling __handle_speculative_fault() if needed.
So this can't explain this difference, except if __handle_speculative_fault()
is inlined in __do_page_fault().
Is this the case on your build ?

Haiyan, do you still have the output of the test to check those numbers too ?

Cheers,
Laurent

> I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case. 
> Please help to check on these data if it can help you to find the higher change. Thanks.
> 
> File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2 
>     tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.
> 
> Best regards,
> Haiyan Song
> 
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Thursday, July 12, 2018 1:05 AM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
> 
> Hi Haiyan,
> 
> Do you get a chance to capture some performance cycles on your system ?
> I still can't get these numbers on my hardware.
> 
> Thanks,
> Laurent.
> 
> On 04/07/2018 09:51, Laurent Dufour wrote:
>> On 04/07/2018 05:23, Song, HaiyanX wrote:
>>> Hi Laurent,
>>>
>>>
>>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>>
>> Repeating the test only 3 times seems a bit too low to me.
>>
>> I'll focus on the higher change for the moment, but I don't have access to such
>> a hardware.
>>
>> Is possible to provide a diff between base and SPF of the performance cycles
>> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>>
>> Please stay focus on the test case process to see exactly where the series is
>> impacting.
>>
>> Thanks,
>> Laurent.
>>
>>>
>>> And I did not find other high variation on test case result.
>>>
>>> a). Enable THP
>>> testcase                          base     stddev       change      head     stddev         metric
>>> page_fault3/enable THP           10519      +- 3%        -20.5%      8368      +-6%          will-it-scale.per_thread_ops
>>> page_fault2/enalbe THP            8281      +- 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>>
>>> b). Disable THP
>>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>>
>>>
>>> Best regards,
>>> Haiyan Song
>>> ________________________________________
>>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Monday, July 02, 2018 4:59 PM
>>> To: Song, HaiyanX
>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>
>>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>>> Hi Laurent,
>>>>
>>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>>> V9 patch serials.
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>>> commit id:
>>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>>> Benchmark: will-it-scale
>>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>>
>>>> Metrics:
>>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task:100%
>>>>
>>>> 1. Regressions:
>>>>
>>>> a). Enable THP
>>>> testcase                          base           change      head           metric
>>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>>
>>>> b). Disable THP
>>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>>
>>>> Notes: for the above  values of test result, the higher is better.
>>>
>>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>>> get reproducible results. The results have huge variation, even on the vanilla
>>> kernel, and I can't state on any changes due to that.
>>>
>>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>>> measure any changes between the vanilla and the SPF patched ones:
>>>
>>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>>> page_fault3_threads             2697.7          2683.5          -0.53%
>>> page_fault2_threads             170660.6        169574.1        -0.64%
>>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>>> context_switch1_processes       6478076.2       6529493.5       0.79%
>>> brk1                            243391.2        238527.5        -2.00%
>>>
>>> Tests were run 10 times, no high variation detected.
>>>
>>> Did you see high variation on your side ? How many times the test were run to
>>> compute the average values ?
>>>
>>> Thanks,
>>> Laurent.
>>>
>>>
>>>>
>>>> 2. Improvement: not found improvement based on the selected test cases.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Monday, May 28, 2018 4:54 PM
>>>> To: Song, HaiyanX
>>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>>> Hi Laurent,
>>>>>
>>>>> Yes, these tests are done on V9 patch.
>>>>
>>>> Do you plan to give this V11 a run ?
>>>>
>>>>>
>>>>>
>>>>> Best regards,
>>>>> Haiyan Song
>>>>>
>>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>>
>>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>>> tested on Intel 4s Skylake platform.
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>>> series" while responding to the v11 header series...
>>>>>> Were these tests done on v9 or v11 ?
>>>>>>
>>>>>> Cheers,
>>>>>> Laurent.
>>>>>>
>>>>>>>
>>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>>> Commit id:
>>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>>> Benchmark suite: will-it-scale
>>>>>>> Download link:
>>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>>> Metrics:
>>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>>> THP: enable / disable
>>>>>>> nr_task: 100%
>>>>>>>
>>>>>>> 1. Regressions:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>>
>>>>>>> 2. Improvements:
>>>>>>> a) THP enabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> b) THP disabled:
>>>>>>> testcase                        base            change          head       metric
>>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>>
>>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>>
>>>>>>>
>>>>>>> Best regards
>>>>>>> Haiyan Song
>>>>>>>
>>>>>>> ________________________________________
>>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>>
>>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>>> page fault without holding the mm semaphore [1].
>>>>>>>
>>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>>
>>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>>> limiting the locking contention to these operations which are expected to
>>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>>> benchmark anymore.
>>>>>>>
>>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>>> speculative page fault in that case.
>>>>>>>
>>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>>> checked during the page fault are modified.
>>>>>>>
>>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>>> parallel change is possible at this time.
>>>>>>>
>>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>>> classic page fault handler will be called to handle the operation while
>>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>>> PTE.
>>>>>>>
>>>>>>> In pseudo code, this could be seen as:
>>>>>>>     speculative_page_fault()
>>>>>>>     {
>>>>>>>             vma = get_vma()
>>>>>>>             check vma sequence count
>>>>>>>             check vma's support
>>>>>>>             disable interrupt
>>>>>>>                   check pgd,p4d,...,pte
>>>>>>>                   save pmd and pte in vmf
>>>>>>>                   save vma sequence counter in vmf
>>>>>>>             enable interrupt
>>>>>>>             check vma sequence count
>>>>>>>             handle_pte_fault(vma)
>>>>>>>                     ..
>>>>>>>                     page = alloc_page()
>>>>>>>                     pte_map_lock()
>>>>>>>                             disable interrupt
>>>>>>>                                     abort if sequence counter has changed
>>>>>>>                                     abort if pmd or pte has changed
>>>>>>>                                     pte map and lock
>>>>>>>                             enable interrupt
>>>>>>>                     if abort
>>>>>>>                        free page
>>>>>>>                        abort
>>>>>>>                     ...
>>>>>>>     }
>>>>>>>
>>>>>>>     arch_fault_handler()
>>>>>>>     {
>>>>>>>             if (speculative_page_fault(&vma))
>>>>>>>                goto done
>>>>>>>     again:
>>>>>>>             lock(mmap_sem)
>>>>>>>             vma = find_vma();
>>>>>>>             handle_pte_fault(vma);
>>>>>>>             if retry
>>>>>>>                unlock(mmap_sem)
>>>>>>>                goto again;
>>>>>>>     done:
>>>>>>>             handle fault error
>>>>>>>     }
>>>>>>>
>>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>>
>>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>>> the part of the faults processed speculatively.
>>>>>>>
>>>>>>> There are some trace events introduced by this series. They allow
>>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>>    back.
>>>>>>>
>>>>>>> To record all the related events, the easier is to run perf with the
>>>>>>> following arguments :
>>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>>
>>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>>
>>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>>> on x86, PowerPC and arm64.
>>>>>>>
>>>>>>> ---------------------
>>>>>>> Real Workload results
>>>>>>>
>>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>>> this new version should not impact the performance boost seen.
>>>>>>>
>>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>>> series:
>>>>>>>                 vanilla         spf
>>>>>>> faults          89.418          101.364         +13%
>>>>>>> spf                n/a           97.989
>>>>>>>
>>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>>> way.
>>>>>>>
>>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>>> it a try on an android device. He reported that the application launch time
>>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>>> 20%.
>>>>>>>
>>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>>
>>>>>>> Application                             4.9     4.9+spf delta
>>>>>>> com.tencent.mm                          416     389     -7%
>>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>>> com.tencent.mtt                         455     454     0%
>>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>>> com.immomo.momo                         501     487     -3%
>>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>>> com.sina.weibo                          907     906     0%
>>>>>>> com.youku.phone                         816     731     -11%
>>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>>> com.UCMobile                            415     411     -1%
>>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>>
>>>>>>> ------------------
>>>>>>> Benchmarks results
>>>>>>>
>>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>>> SPF is BASE + this series
>>>>>>>
>>>>>>> Kernbench:
>>>>>>> ----------
>>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>>> kernel (kernel is build 5 times):
>>>>>>>
>>>>>>> Average Half load -j 8
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>>
>>>>>>> Average Optimal load -j 16
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>>
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          526743764      faults
>>>>>>>                210      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>>> were created during the kernel build processing).
>>>>>>>
>>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>>
>>>>>>> Average Half load -j 40
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>>
>>>>>>> Average Optimal load -j 80
>>>>>>>                  Run    (std deviation)
>>>>>>>                  BASE                   SPF
>>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>>
>>>>>>> During a run on the SPF, perf events were captured:
>>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>>          116730856      faults
>>>>>>>                  0      spf
>>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>>> there is no impact on the performance.
>>>>>>>
>>>>>>> Ebizzy:
>>>>>>> -------
>>>>>>> The test is counting the number of records per second it can manage, the
>>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>>> result. The number is the record processes per second, the higher is the
>>>>>>> best.
>>>>>>>
>>>>>>>                 BASE            SPF             delta
>>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>>
>>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>>            1706379      faults
>>>>>>>            1674599      spf
>>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>>            1874773      faults
>>>>>>>            1461153      spf
>>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>>                  0      pagefault:spf_vma_access
>>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>>
>>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>>> leading the ebizzy performance boost.
>>>>>>>
>>>>>>> ------------------
>>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>>    and Minchan Kim, hopefully.
>>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>>    __do_page_fault().
>>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>>    instead
>>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>>> useless
>>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>>
>>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>>
>>>>>>>
>>>>>>> Laurent Dufour (20):
>>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>>   mm: introduce INIT_VMA()
>>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>>   mm: introduce __vm_normal_page()
>>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>>   perf: add a speculative page fault sw event
>>>>>>>   perf tools: add support for the SPF perf event
>>>>>>>   mm: add speculative page fault vmstats
>>>>>>>   powerpc/mm: add speculative page fault
>>>>>>>
>>>>>>> Mahendran Ganesh (2):
>>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>>   arm64/mm: add speculative page fault
>>>>>>>
>>>>>>> Peter Zijlstra (4):
>>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>>   mm: VMA sequence count
>>>>>>>   mm: provide speculative fault infrastructure
>>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>>
>>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>>  fs/exec.c                             |   2 +-
>>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>>  mm/internal.h                         |  20 ++
>>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>>  mm/mremap.c                           |  13 +
>>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>>  mm/swap.c                             |   6 +-
>>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>>
>>>>>>> --
>>>>>>> 2.7.4
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>>
>>
> 

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

[-- Attachment #1: Type: text/plain, Size: 37281 bytes --]

Hi Laurent,

I attached the perf-profile.gz file for case page_fault2 and page_fault3. These files were captured during test the related test case. 
Please help to check on these data if it can help you to find the higher change. Thanks.

File name perf-profile_page_fault2_head_THP-Always.gz, means the perf-profile result get from page_fault2 
    tested for head commit (a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12) with THP_always configuration.

Best regards,
Haiyan Song

________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Thursday, July 12, 2018 1:05 AM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

Hi Haiyan,

Do you get a chance to capture some performance cycles on your system ?
I still can't get these numbers on my hardware.

Thanks,
Laurent.

On 04/07/2018 09:51, Laurent Dufour wrote:
> On 04/07/2018 05:23, Song, HaiyanX wrote:
>> Hi Laurent,
>>
>>
>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
>
> Repeating the test only 3 times seems a bit too low to me.
>
> I'll focus on the higher change for the moment, but I don't have access to such
> a hardware.
>
> Is possible to provide a diff between base and SPF of the performance cycles
> measured when running page_fault3 and page_fault2 when the 20% change is detected.
>
> Please stay focus on the test case process to see exactly where the series is
> impacting.
>
> Thanks,
> Laurent.
>
>>
>> And I did not find other high variation on test case result.
>>
>> a). Enable THP
>> testcase                          base     stddev       change      head     stddev         metric
>> page_fault3/enable THP           10519      ± 3%        -20.5%      8368      ±6%          will-it-scale.per_thread_ops
>> page_fault2/enalbe THP            8281      ± 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>
>> b). Disable THP
>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>
>>
>> Best regards,
>> Haiyan Song
>> ________________________________________
>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>> Sent: Monday, July 02, 2018 4:59 PM
>> To: Song, HaiyanX
>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>
>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>> Hi Laurent,
>>>
>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>> V9 patch serials.
>>>
>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>> commit id:
>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>> Benchmark: will-it-scale
>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>
>>> Metrics:
>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>> THP: enable / disable
>>> nr_task:100%
>>>
>>> 1. Regressions:
>>>
>>> a). Enable THP
>>> testcase                          base           change      head           metric
>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>
>>> b). Disable THP
>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>
>>> Notes: for the above  values of test result, the higher is better.
>>
>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>> get reproducible results. The results have huge variation, even on the vanilla
>> kernel, and I can't state on any changes due to that.
>>
>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>> measure any changes between the vanilla and the SPF patched ones:
>>
>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>> page_fault3_threads             2697.7          2683.5          -0.53%
>> page_fault2_threads             170660.6        169574.1        -0.64%
>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>> context_switch1_processes       6478076.2       6529493.5       0.79%
>> brk1                            243391.2        238527.5        -2.00%
>>
>> Tests were run 10 times, no high variation detected.
>>
>> Did you see high variation on your side ? How many times the test were run to
>> compute the average values ?
>>
>> Thanks,
>> Laurent.
>>
>>
>>>
>>> 2. Improvement: not found improvement based on the selected test cases.
>>>
>>>
>>> Best regards
>>> Haiyan Song
>>> ________________________________________
>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Monday, May 28, 2018 4:54 PM
>>> To: Song, HaiyanX
>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>
>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>> Hi Laurent,
>>>>
>>>> Yes, these tests are done on V9 patch.
>>>
>>> Do you plan to give this V11 a run ?
>>>
>>>>
>>>>
>>>> Best regards,
>>>> Haiyan Song
>>>>
>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>
>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>> tested on Intel 4s Skylake platform.
>>>>>
>>>>> Hi,
>>>>>
>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>> series" while responding to the v11 header series...
>>>>> Were these tests done on v9 or v11 ?
>>>>>
>>>>> Cheers,
>>>>> Laurent.
>>>>>
>>>>>>
>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>> Commit id:
>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>> Benchmark suite: will-it-scale
>>>>>> Download link:
>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>> Metrics:
>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>> THP: enable / disable
>>>>>> nr_task: 100%
>>>>>>
>>>>>> 1. Regressions:
>>>>>> a) THP enabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>
>>>>>> b) THP disabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>
>>>>>> 2. Improvements:
>>>>>> a) THP enabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>
>>>>>> b) THP disabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>
>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>
>>>>>>
>>>>>> Best regards
>>>>>> Haiyan Song
>>>>>>
>>>>>> ________________________________________
>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>
>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>> page fault without holding the mm semaphore [1].
>>>>>>
>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>
>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>> limiting the locking contention to these operations which are expected to
>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>> benchmark anymore.
>>>>>>
>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>> speculative page fault in that case.
>>>>>>
>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>> checked during the page fault are modified.
>>>>>>
>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>> parallel change is possible at this time.
>>>>>>
>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>> classic page fault handler will be called to handle the operation while
>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>> PTE.
>>>>>>
>>>>>> In pseudo code, this could be seen as:
>>>>>>     speculative_page_fault()
>>>>>>     {
>>>>>>             vma = get_vma()
>>>>>>             check vma sequence count
>>>>>>             check vma's support
>>>>>>             disable interrupt
>>>>>>                   check pgd,p4d,...,pte
>>>>>>                   save pmd and pte in vmf
>>>>>>                   save vma sequence counter in vmf
>>>>>>             enable interrupt
>>>>>>             check vma sequence count
>>>>>>             handle_pte_fault(vma)
>>>>>>                     ..
>>>>>>                     page = alloc_page()
>>>>>>                     pte_map_lock()
>>>>>>                             disable interrupt
>>>>>>                                     abort if sequence counter has changed
>>>>>>                                     abort if pmd or pte has changed
>>>>>>                                     pte map and lock
>>>>>>                             enable interrupt
>>>>>>                     if abort
>>>>>>                        free page
>>>>>>                        abort
>>>>>>                     ...
>>>>>>     }
>>>>>>
>>>>>>     arch_fault_handler()
>>>>>>     {
>>>>>>             if (speculative_page_fault(&vma))
>>>>>>                goto done
>>>>>>     again:
>>>>>>             lock(mmap_sem)
>>>>>>             vma = find_vma();
>>>>>>             handle_pte_fault(vma);
>>>>>>             if retry
>>>>>>                unlock(mmap_sem)
>>>>>>                goto again;
>>>>>>     done:
>>>>>>             handle fault error
>>>>>>     }
>>>>>>
>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>
>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>> the part of the faults processed speculatively.
>>>>>>
>>>>>> There are some trace events introduced by this series. They allow
>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>    back.
>>>>>>
>>>>>> To record all the related events, the easier is to run perf with the
>>>>>> following arguments :
>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>
>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>
>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>> on x86, PowerPC and arm64.
>>>>>>
>>>>>> ---------------------
>>>>>> Real Workload results
>>>>>>
>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>> this new version should not impact the performance boost seen.
>>>>>>
>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>> series:
>>>>>>                 vanilla         spf
>>>>>> faults          89.418          101.364         +13%
>>>>>> spf                n/a           97.989
>>>>>>
>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>> way.
>>>>>>
>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>> it a try on an android device. He reported that the application launch time
>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>> 20%.
>>>>>>
>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>
>>>>>> Application                             4.9     4.9+spf delta
>>>>>> com.tencent.mm                          416     389     -7%
>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>> com.tencent.mtt                         455     454     0%
>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>> com.immomo.momo                         501     487     -3%
>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>> com.sina.weibo                          907     906     0%
>>>>>> com.youku.phone                         816     731     -11%
>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>> com.UCMobile                            415     411     -1%
>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>
>>>>>> ------------------
>>>>>> Benchmarks results
>>>>>>
>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>> SPF is BASE + this series
>>>>>>
>>>>>> Kernbench:
>>>>>> ----------
>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>> kernel (kernel is build 5 times):
>>>>>>
>>>>>> Average Half load -j 8
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>
>>>>>> Average Optimal load -j 16
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>
>>>>>>
>>>>>> During a run on the SPF, perf events were captured:
>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>          526743764      faults
>>>>>>                210      spf
>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>> were created during the kernel build processing).
>>>>>>
>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>
>>>>>> Average Half load -j 40
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>
>>>>>> Average Optimal load -j 80
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>
>>>>>> During a run on the SPF, perf events were captured:
>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>          116730856      faults
>>>>>>                  0      spf
>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>> there is no impact on the performance.
>>>>>>
>>>>>> Ebizzy:
>>>>>> -------
>>>>>> The test is counting the number of records per second it can manage, the
>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>> result. The number is the record processes per second, the higher is the
>>>>>> best.
>>>>>>
>>>>>>                 BASE            SPF             delta
>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>
>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>            1706379      faults
>>>>>>            1674599      spf
>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>            1874773      faults
>>>>>>            1461153      spf
>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>> leading the ebizzy performance boost.
>>>>>>
>>>>>> ------------------
>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>    and Minchan Kim, hopefully.
>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>    __do_page_fault().
>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>    instead
>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>> useless
>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>
>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>
>>>>>>
>>>>>> Laurent Dufour (20):
>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>   mm: introduce INIT_VMA()
>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>   mm: introduce __vm_normal_page()
>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>   perf: add a speculative page fault sw event
>>>>>>   perf tools: add support for the SPF perf event
>>>>>>   mm: add speculative page fault vmstats
>>>>>>   powerpc/mm: add speculative page fault
>>>>>>
>>>>>> Mahendran Ganesh (2):
>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   arm64/mm: add speculative page fault
>>>>>>
>>>>>> Peter Zijlstra (4):
>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>   mm: VMA sequence count
>>>>>>   mm: provide speculative fault infrastructure
>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>
>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>  fs/exec.c                             |   2 +-
>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>  mm/internal.h                         |  20 ++
>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>  mm/mremap.c                           |  13 +
>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>  mm/swap.c                             |   6 +-
>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>
>>>>>> --
>>>>>> 2.7.4
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>>
>


[-- Attachment #2: perf-profile_page_fault2_base_THP-Alwasys.gz --]
[-- Type: application/gzip, Size: 10171 bytes --]

[-- Attachment #3: perf-profile_page_fault2_base_thp_never.gz --]
[-- Type: application/gzip, Size: 11474 bytes --]

[-- Attachment #4: perf-profile_page_fault2_head_THP-Always.gz --]
[-- Type: application/gzip, Size: 10374 bytes --]

[-- Attachment #5: perf-profile_page_fault2_head_thp_never.gz --]
[-- Type: application/gzip, Size: 11327 bytes --]

[-- Attachment #6: perf-profile_page_fault3_base_THP-Always.gz --]
[-- Type: application/gzip, Size: 9503 bytes --]

[-- Attachment #7: perf-profile_page_fault3_base_thp_never.gz --]
[-- Type: application/gzip, Size: 9843 bytes --]

[-- Attachment #8: perf-profile_page_fault3_head_THP-Always.gz --]
[-- Type: application/gzip, Size: 9596 bytes --]

[-- Attachment #9: perf-profile_page_fault3_head_thp_never.gz --]
[-- Type: application/gzip, Size: 10137 bytes --]

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Hi Haiyan,

Do you get a chance to capture some performance cycles on your system ?
I still can't get these numbers on my hardware.

Thanks,
Laurent.

On 04/07/2018 09:51, Laurent Dufour wrote:
> On 04/07/2018 05:23, Song, HaiyanX wrote:
>> Hi Laurent,
>>
>>
>> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
>> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.
> 
> Repeating the test only 3 times seems a bit too low to me.
> 
> I'll focus on the higher change for the moment, but I don't have access to such
> a hardware.
> 
> Is possible to provide a diff between base and SPF of the performance cycles
> measured when running page_fault3 and page_fault2 when the 20% change is detected.
> 
> Please stay focus on the test case process to see exactly where the series is
> impacting.
> 
> Thanks,
> Laurent.
> 
>>
>> And I did not find other high variation on test case result.
>>
>> a). Enable THP
>> testcase                          base     stddev       change      head     stddev         metric
>> page_fault3/enable THP           10519      A+- 3%        -20.5%      8368      A+-6%          will-it-scale.per_thread_ops
>> page_fault2/enalbe THP            8281      A+- 2%        -18.8%      6728                   will-it-scale.per_thread_ops
>> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
>> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
>> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
>>
>> b). Disable THP
>> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
>> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
>> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
>> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
>> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
>> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
>> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
>>
>>
>> Best regards,
>> Haiyan Song
>> ________________________________________
>> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
>> Sent: Monday, July 02, 2018 4:59 PM
>> To: Song, HaiyanX
>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>
>> On 11/06/2018 09:49, Song, HaiyanX wrote:
>>> Hi Laurent,
>>>
>>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>>> V9 patch serials.
>>>
>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>>> commit id:
>>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>>> Benchmark: will-it-scale
>>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>>
>>> Metrics:
>>>   will-it-scale.per_process_ops=processes/nr_cpu
>>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>> THP: enable / disable
>>> nr_task:100%
>>>
>>> 1. Regressions:
>>>
>>> a). Enable THP
>>> testcase                          base           change      head           metric
>>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>>
>>> b). Disable THP
>>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>>
>>> Notes: for the above  values of test result, the higher is better.
>>
>> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
>> get reproducible results. The results have huge variation, even on the vanilla
>> kernel, and I can't state on any changes due to that.
>>
>> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
>> measure any changes between the vanilla and the SPF patched ones:
>>
>> test THP enabled                4.17.0-rc4-mm1  spf             delta
>> page_fault3_threads             2697.7          2683.5          -0.53%
>> page_fault2_threads             170660.6        169574.1        -0.64%
>> context_switch1_threads         6915269.2       6877507.3       -0.55%
>> context_switch1_processes       6478076.2       6529493.5       0.79%
>> brk1                            243391.2        238527.5        -2.00%
>>
>> Tests were run 10 times, no high variation detected.
>>
>> Did you see high variation on your side ? How many times the test were run to
>> compute the average values ?
>>
>> Thanks,
>> Laurent.
>>
>>
>>>
>>> 2. Improvement: not found improvement based on the selected test cases.
>>>
>>>
>>> Best regards
>>> Haiyan Song
>>> ________________________________________
>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Monday, May 28, 2018 4:54 PM
>>> To: Song, HaiyanX
>>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>>
>>> On 28/05/2018 10:22, Haiyan Song wrote:
>>>> Hi Laurent,
>>>>
>>>> Yes, these tests are done on V9 patch.
>>>
>>> Do you plan to give this V11 a run ?
>>>
>>>>
>>>>
>>>> Best regards,
>>>> Haiyan Song
>>>>
>>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>>
>>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>>> tested on Intel 4s Skylake platform.
>>>>>
>>>>> Hi,
>>>>>
>>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>>> series" while responding to the v11 header series...
>>>>> Were these tests done on v9 or v11 ?
>>>>>
>>>>> Cheers,
>>>>> Laurent.
>>>>>
>>>>>>
>>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>>> Commit id:
>>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>>> Benchmark suite: will-it-scale
>>>>>> Download link:
>>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>>> Metrics:
>>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>>> THP: enable / disable
>>>>>> nr_task: 100%
>>>>>>
>>>>>> 1. Regressions:
>>>>>> a) THP enabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>>
>>>>>> b) THP disabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>>
>>>>>> 2. Improvements:
>>>>>> a) THP enabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>>
>>>>>> b) THP disabled:
>>>>>> testcase                        base            change          head       metric
>>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>>
>>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>>
>>>>>>
>>>>>> Best regards
>>>>>> Haiyan Song
>>>>>>
>>>>>> ________________________________________
>>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>>
>>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>>> page fault without holding the mm semaphore [1].
>>>>>>
>>>>>> The idea is to try to handle user space page faults without holding the
>>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>>> process since the page fault handler will not wait for other threads memory
>>>>>> layout change to be done, assuming that this change is done in another part
>>>>>> of the process's memory space. This type page fault is named speculative
>>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>>> is failing its processing and a classic page fault is then tried.
>>>>>>
>>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>>> limiting the locking contention to these operations which are expected to
>>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>>> benchmark anymore.
>>>>>>
>>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>>> have to be protected against parallel changes. This is done by using a per
>>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>>> speculative page fault in that case.
>>>>>>
>>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>>> checked during the page fault are modified.
>>>>>>
>>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>>> parallel change is possible at this time.
>>>>>>
>>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>>> classic page fault handler will be called to handle the operation while
>>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>>> PTE.
>>>>>>
>>>>>> In pseudo code, this could be seen as:
>>>>>>     speculative_page_fault()
>>>>>>     {
>>>>>>             vma = get_vma()
>>>>>>             check vma sequence count
>>>>>>             check vma's support
>>>>>>             disable interrupt
>>>>>>                   check pgd,p4d,...,pte
>>>>>>                   save pmd and pte in vmf
>>>>>>                   save vma sequence counter in vmf
>>>>>>             enable interrupt
>>>>>>             check vma sequence count
>>>>>>             handle_pte_fault(vma)
>>>>>>                     ..
>>>>>>                     page = alloc_page()
>>>>>>                     pte_map_lock()
>>>>>>                             disable interrupt
>>>>>>                                     abort if sequence counter has changed
>>>>>>                                     abort if pmd or pte has changed
>>>>>>                                     pte map and lock
>>>>>>                             enable interrupt
>>>>>>                     if abort
>>>>>>                        free page
>>>>>>                        abort
>>>>>>                     ...
>>>>>>     }
>>>>>>
>>>>>>     arch_fault_handler()
>>>>>>     {
>>>>>>             if (speculative_page_fault(&vma))
>>>>>>                goto done
>>>>>>     again:
>>>>>>             lock(mmap_sem)
>>>>>>             vma = find_vma();
>>>>>>             handle_pte_fault(vma);
>>>>>>             if retry
>>>>>>                unlock(mmap_sem)
>>>>>>                goto again;
>>>>>>     done:
>>>>>>             handle fault error
>>>>>>     }
>>>>>>
>>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>>
>>>>>> This series add a new software performance event named 'speculative-faults'
>>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>>> the part of the faults processed speculatively.
>>>>>>
>>>>>> There are some trace events introduced by this series. They allow
>>>>>> identifying why the page faults were not processed speculatively. This
>>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>>> grouped in a system named 'pagefault', they are:
>>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>>    back.
>>>>>>
>>>>>> To record all the related events, the easier is to run perf with the
>>>>>> following arguments :
>>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>>
>>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>>> page fault handled speculatively. I can be seen this way:
>>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>>
>>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>>> on x86, PowerPC and arm64.
>>>>>>
>>>>>> ---------------------
>>>>>> Real Workload results
>>>>>>
>>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>>> this new version should not impact the performance boost seen.
>>>>>>
>>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>>> series:
>>>>>>                 vanilla         spf
>>>>>> faults          89.418          101.364         +13%
>>>>>> spf                n/a           97.989
>>>>>>
>>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>>> way.
>>>>>>
>>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>>> it a try on an android device. He reported that the application launch time
>>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>>> 20%.
>>>>>>
>>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>>
>>>>>> Application                             4.9     4.9+spf delta
>>>>>> com.tencent.mm                          416     389     -7%
>>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>>> com.tencent.mtt                         455     454     0%
>>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>>> com.autonavi.minimap                    711     701     -1%
>>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>>> com.immomo.momo                         501     487     -3%
>>>>>> com.tencent.peng                        2145    2112    -2%
>>>>>> com.smile.gifmaker                      491     461     -6%
>>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>>> com.baidu.searchbox                     333     314     -6%
>>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>>> com.sina.weibo                          907     906     0%
>>>>>> com.youku.phone                         816     731     -11%
>>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>>> com.UCMobile                            415     411     -1%
>>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>>
>>>>>> ------------------
>>>>>> Benchmarks results
>>>>>>
>>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>>> SPF is BASE + this series
>>>>>>
>>>>>> Kernbench:
>>>>>> ----------
>>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>>> kernel (kernel is build 5 times):
>>>>>>
>>>>>> Average Half load -j 8
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>>
>>>>>> Average Optimal load -j 16
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>>
>>>>>>
>>>>>> During a run on the SPF, perf events were captured:
>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>          526743764      faults
>>>>>>                210      spf
>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>               2278      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> Very few speculative page faults were recorded as most of the processes
>>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>>> were created during the kernel build processing).
>>>>>>
>>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>>
>>>>>> Average Half load -j 40
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>>
>>>>>> Average Optimal load -j 80
>>>>>>                  Run    (std deviation)
>>>>>>                  BASE                   SPF
>>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>>
>>>>>> During a run on the SPF, perf events were captured:
>>>>>>  Performance counter stats for '../kernbench -M':
>>>>>>          116730856      faults
>>>>>>                  0      spf
>>>>>>                  3      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                476      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>>> there is no impact on the performance.
>>>>>>
>>>>>> Ebizzy:
>>>>>> -------
>>>>>> The test is counting the number of records per second it can manage, the
>>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>>> consistent result I repeated the test 100 times and measure the average
>>>>>> result. The number is the record processes per second, the higher is the
>>>>>> best.
>>>>>>
>>>>>>                 BASE            SPF             delta
>>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>>
>>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>>            1706379      faults
>>>>>>            1674599      spf
>>>>>>              30588      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                363      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>>            1874773      faults
>>>>>>            1461153      spf
>>>>>>             413293      pagefault:spf_vma_changed
>>>>>>                  0      pagefault:spf_vma_noanon
>>>>>>                200      pagefault:spf_vma_notsup
>>>>>>                  0      pagefault:spf_vma_access
>>>>>>                  0      pagefault:spf_pmd_changed
>>>>>>
>>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>>> leading the ebizzy performance boost.
>>>>>>
>>>>>> ------------------
>>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>>    and Minchan Kim, hopefully.
>>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>>    __do_page_fault().
>>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>>    instead
>>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>>> useless
>>>>>>    trace event pagefault:spf_pte_lock.
>>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>>
>>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>>
>>>>>>
>>>>>> Laurent Dufour (20):
>>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>>   mm: introduce INIT_VMA()
>>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>>   mm: protect mremap() against SPF hanlder
>>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>>   mm: introduce __vm_normal_page()
>>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>>   mm: adding speculative page fault failure trace events
>>>>>>   perf: add a speculative page fault sw event
>>>>>>   perf tools: add support for the SPF perf event
>>>>>>   mm: add speculative page fault vmstats
>>>>>>   powerpc/mm: add speculative page fault
>>>>>>
>>>>>> Mahendran Ganesh (2):
>>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>>   arm64/mm: add speculative page fault
>>>>>>
>>>>>> Peter Zijlstra (4):
>>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>>   mm: VMA sequence count
>>>>>>   mm: provide speculative fault infrastructure
>>>>>>   x86/mm: add speculative pagefault handling
>>>>>>
>>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>>  fs/exec.c                             |   2 +-
>>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>>  include/linux/migrate.h               |   4 +-
>>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>>  include/linux/mm_types.h              |   7 +
>>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>>  include/linux/swap.h                  |  10 +-
>>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>>  kernel/fork.c                         |   5 +-
>>>>>>  mm/Kconfig                            |  22 ++
>>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>>  mm/hugetlb.c                          |   2 +
>>>>>>  mm/init-mm.c                          |   3 +
>>>>>>  mm/internal.h                         |  20 ++
>>>>>>  mm/khugepaged.c                       |   5 +
>>>>>>  mm/madvise.c                          |   6 +-
>>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>>  mm/migrate.c                          |   6 +-
>>>>>>  mm/mlock.c                            |  13 +-
>>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>>  mm/mprotect.c                         |   4 +-
>>>>>>  mm/mremap.c                           |  13 +
>>>>>>  mm/nommu.c                            |   2 +-
>>>>>>  mm/rmap.c                             |   5 +-
>>>>>>  mm/swap.c                             |   6 +-
>>>>>>  mm/swap_state.c                       |   8 +-
>>>>>>  mm/vmstat.c                           |   5 +-
>>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>>  tools/perf/util/python.c              |   1 +
>>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>>
>>>>>> --
>>>>>> 2.7.4
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>>
> 

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

On 04/07/2018 05:23, Song, HaiyanX wrote:
> Hi Laurent,
> 
> 
> For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
> I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.

Repeating the test only 3 times seems a bit too low to me.

I'll focus on the higher change for the moment, but I don't have access to such
a hardware.

Is possible to provide a diff between base and SPF of the performance cycles
measured when running page_fault3 and page_fault2 when the 20% change is detected.

Please stay focus on the test case process to see exactly where the series is
impacting.

Thanks,
Laurent.

> 
> And I did not find other high variation on test case result.
> 
> a). Enable THP
> testcase                          base     stddev       change      head     stddev         metric
> page_fault3/enable THP           10519      A+- 3%        -20.5%      8368      A+-6%          will-it-scale.per_thread_ops
> page_fault2/enalbe THP            8281      A+- 2%        -18.8%      6728                   will-it-scale.per_thread_ops
> brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
> context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
> context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops
> 
> b). Disable THP
> page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
> page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
> brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
> context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
> brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
> page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
> context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops
> 
> 
> Best regards,
> Haiyan Song
> ________________________________________
> From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Monday, July 02, 2018 4:59 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
> 
> On 11/06/2018 09:49, Song, HaiyanX wrote:
>> Hi Laurent,
>>
>> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
>> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
>> V9 patch serials.
>>
>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
>> commit id:
>>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
>> Benchmark: will-it-scale
>> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>>
>> Metrics:
>>   will-it-scale.per_process_ops=processes/nr_cpu
>>   will-it-scale.per_thread_ops=threads/nr_cpu
>>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>> THP: enable / disable
>> nr_task:100%
>>
>> 1. Regressions:
>>
>> a). Enable THP
>> testcase                          base           change      head           metric
>> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
>> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
>> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
>> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
>> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>>
>> b). Disable THP
>> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
>> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
>> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
>> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
>> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
>> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
>> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>>
>> Notes: for the above  values of test result, the higher is better.
> 
> I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
> get reproducible results. The results have huge variation, even on the vanilla
> kernel, and I can't state on any changes due to that.
> 
> I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
> measure any changes between the vanilla and the SPF patched ones:
> 
> test THP enabled                4.17.0-rc4-mm1  spf             delta
> page_fault3_threads             2697.7          2683.5          -0.53%
> page_fault2_threads             170660.6        169574.1        -0.64%
> context_switch1_threads         6915269.2       6877507.3       -0.55%
> context_switch1_processes       6478076.2       6529493.5       0.79%
> brk1                            243391.2        238527.5        -2.00%
> 
> Tests were run 10 times, no high variation detected.
> 
> Did you see high variation on your side ? How many times the test were run to
> compute the average values ?
> 
> Thanks,
> Laurent.
> 
> 
>>
>> 2. Improvement: not found improvement based on the selected test cases.
>>
>>
>> Best regards
>> Haiyan Song
>> ________________________________________
>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>> Sent: Monday, May 28, 2018 4:54 PM
>> To: Song, HaiyanX
>> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>> Subject: Re: [PATCH v11 00/26] Speculative page faults
>>
>> On 28/05/2018 10:22, Haiyan Song wrote:
>>> Hi Laurent,
>>>
>>> Yes, these tests are done on V9 patch.
>>
>> Do you plan to give this V11 a run ?
>>
>>>
>>>
>>> Best regards,
>>> Haiyan Song
>>>
>>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>>
>>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>>> tested on Intel 4s Skylake platform.
>>>>
>>>> Hi,
>>>>
>>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>>> series" while responding to the v11 header series...
>>>> Were these tests done on v9 or v11 ?
>>>>
>>>> Cheers,
>>>> Laurent.
>>>>
>>>>>
>>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>>> Commit id:
>>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>>> Benchmark suite: will-it-scale
>>>>> Download link:
>>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>>> Metrics:
>>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>>> THP: enable / disable
>>>>> nr_task: 100%
>>>>>
>>>>> 1. Regressions:
>>>>> a) THP enabled:
>>>>> testcase                        base            change          head       metric
>>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>>
>>>>> b) THP disabled:
>>>>> testcase                        base            change          head       metric
>>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>>
>>>>> 2. Improvements:
>>>>> a) THP enabled:
>>>>> testcase                        base            change          head       metric
>>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>>
>>>>> b) THP disabled:
>>>>> testcase                        base            change          head       metric
>>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>>
>>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>>> on head commit is better than that on base commit for this benchmark.
>>>>>
>>>>>
>>>>> Best regards
>>>>> Haiyan Song
>>>>>
>>>>> ________________________________________
>>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>>
>>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>>> page fault without holding the mm semaphore [1].
>>>>>
>>>>> The idea is to try to handle user space page faults without holding the
>>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>>> process since the page fault handler will not wait for other threads memory
>>>>> layout change to be done, assuming that this change is done in another part
>>>>> of the process's memory space. This type page fault is named speculative
>>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>>> is failing its processing and a classic page fault is then tried.
>>>>>
>>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>>> freeing operation which was hitting the performance by 20% as reported by
>>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>>> limiting the locking contention to these operations which are expected to
>>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>>> our back a reference count is added and 2 services (get_vma() and
>>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>>> benchmark anymore.
>>>>>
>>>>> The VMA's attributes checked during the speculative page fault processing
>>>>> have to be protected against parallel changes. This is done by using a per
>>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>>> handler to fast check for parallel changes in progress and to abort the
>>>>> speculative page fault in that case.
>>>>>
>>>>> Once the VMA has been found, the speculative page fault handler would check
>>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>>> checked during the page fault are modified.
>>>>>
>>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>>> leading to touching this PTE will need to lock the page table, so no
>>>>> parallel change is possible at this time.
>>>>>
>>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>>> different from the one recorded at the beginning of the SPF operation, the
>>>>> classic page fault handler will be called to handle the operation while
>>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>>> PTE.
>>>>>
>>>>> In pseudo code, this could be seen as:
>>>>>     speculative_page_fault()
>>>>>     {
>>>>>             vma = get_vma()
>>>>>             check vma sequence count
>>>>>             check vma's support
>>>>>             disable interrupt
>>>>>                   check pgd,p4d,...,pte
>>>>>                   save pmd and pte in vmf
>>>>>                   save vma sequence counter in vmf
>>>>>             enable interrupt
>>>>>             check vma sequence count
>>>>>             handle_pte_fault(vma)
>>>>>                     ..
>>>>>                     page = alloc_page()
>>>>>                     pte_map_lock()
>>>>>                             disable interrupt
>>>>>                                     abort if sequence counter has changed
>>>>>                                     abort if pmd or pte has changed
>>>>>                                     pte map and lock
>>>>>                             enable interrupt
>>>>>                     if abort
>>>>>                        free page
>>>>>                        abort
>>>>>                     ...
>>>>>     }
>>>>>
>>>>>     arch_fault_handler()
>>>>>     {
>>>>>             if (speculative_page_fault(&vma))
>>>>>                goto done
>>>>>     again:
>>>>>             lock(mmap_sem)
>>>>>             vma = find_vma();
>>>>>             handle_pte_fault(vma);
>>>>>             if retry
>>>>>                unlock(mmap_sem)
>>>>>                goto again;
>>>>>     done:
>>>>>             handle fault error
>>>>>     }
>>>>>
>>>>> Support for THP is not done because when checking for the PMD, we can be
>>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>>
>>>>> This series add a new software performance event named 'speculative-faults'
>>>>> or 'spf'. It counts the number of successful page fault event handled
>>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>>> counting the total number of page fault events while 'spf' is only counting
>>>>> the part of the faults processed speculatively.
>>>>>
>>>>> There are some trace events introduced by this series. They allow
>>>>> identifying why the page faults were not processed speculatively. This
>>>>> doesn't take in account the faults generated by a monothreaded process
>>>>> which directly processed while holding the mmap_sem. This trace events are
>>>>> grouped in a system named 'pagefault', they are:
>>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>>    back.
>>>>>
>>>>> To record all the related events, the easier is to run perf with the
>>>>> following arguments :
>>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>>
>>>>> There is also a dedicated vmstat counter showing the number of successful
>>>>> page fault handled speculatively. I can be seen this way:
>>>>> $ grep speculative_pgfault /proc/vmstat
>>>>>
>>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>>> on x86, PowerPC and arm64.
>>>>>
>>>>> ---------------------
>>>>> Real Workload results
>>>>>
>>>>> As mentioned in previous email, we did non official runs using a "popular
>>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>>> which showed a 30% improvements in the number of transaction processed per
>>>>> second. This run has been done on the v6 series, but changes introduced in
>>>>> this new version should not impact the performance boost seen.
>>>>>
>>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>>> series:
>>>>>                 vanilla         spf
>>>>> faults          89.418          101.364         +13%
>>>>> spf                n/a           97.989
>>>>>
>>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>>> way.
>>>>>
>>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>>> it a try on an android device. He reported that the application launch time
>>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>>> 20%.
>>>>>
>>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>>
>>>>> Application                             4.9     4.9+spf delta
>>>>> com.tencent.mm                          416     389     -7%
>>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>>> com.tencent.mtt                         455     454     0%
>>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>>> com.autonavi.minimap                    711     701     -1%
>>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>>> com.immomo.momo                         501     487     -3%
>>>>> com.tencent.peng                        2145    2112    -2%
>>>>> com.smile.gifmaker                      491     461     -6%
>>>>> com.baidu.BaiduMap                      479     366     -23%
>>>>> com.taobao.taobao                       1341    1198    -11%
>>>>> com.baidu.searchbox                     333     314     -6%
>>>>> com.tencent.mobileqq                    394     384     -3%
>>>>> com.sina.weibo                          907     906     0%
>>>>> com.youku.phone                         816     731     -11%
>>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>>> com.UCMobile                            415     411     -1%
>>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>>> com.tencent.qqmusic                     336     329     -2%
>>>>> com.sankuai.meituan                     1661    1302    -22%
>>>>> com.netease.cloudmusic                  1193    1200    1%
>>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>>
>>>>> ------------------
>>>>> Benchmarks results
>>>>>
>>>>> Base kernel is v4.17.0-rc4-mm1
>>>>> SPF is BASE + this series
>>>>>
>>>>> Kernbench:
>>>>> ----------
>>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>>> kernel (kernel is build 5 times):
>>>>>
>>>>> Average Half load -j 8
>>>>>                  Run    (std deviation)
>>>>>                  BASE                   SPF
>>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>>
>>>>> Average Optimal load -j 16
>>>>>                  Run    (std deviation)
>>>>>                  BASE                   SPF
>>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>>
>>>>>
>>>>> During a run on the SPF, perf events were captured:
>>>>>  Performance counter stats for '../kernbench -M':
>>>>>          526743764      faults
>>>>>                210      spf
>>>>>                  3      pagefault:spf_vma_changed
>>>>>                  0      pagefault:spf_vma_noanon
>>>>>               2278      pagefault:spf_vma_notsup
>>>>>                  0      pagefault:spf_vma_access
>>>>>                  0      pagefault:spf_pmd_changed
>>>>>
>>>>> Very few speculative page faults were recorded as most of the processes
>>>>> involved are monothreaded (sounds that on this architecture some threads
>>>>> were created during the kernel build processing).
>>>>>
>>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>>
>>>>> Average Half load -j 40
>>>>>                  Run    (std deviation)
>>>>>                  BASE                   SPF
>>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>>
>>>>> Average Optimal load -j 80
>>>>>                  Run    (std deviation)
>>>>>                  BASE                   SPF
>>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>>
>>>>> During a run on the SPF, perf events were captured:
>>>>>  Performance counter stats for '../kernbench -M':
>>>>>          116730856      faults
>>>>>                  0      spf
>>>>>                  3      pagefault:spf_vma_changed
>>>>>                  0      pagefault:spf_vma_noanon
>>>>>                476      pagefault:spf_vma_notsup
>>>>>                  0      pagefault:spf_vma_access
>>>>>                  0      pagefault:spf_pmd_changed
>>>>>
>>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>>> there is no impact on the performance.
>>>>>
>>>>> Ebizzy:
>>>>> -------
>>>>> The test is counting the number of records per second it can manage, the
>>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>>> consistent result I repeated the test 100 times and measure the average
>>>>> result. The number is the record processes per second, the higher is the
>>>>> best.
>>>>>
>>>>>                 BASE            SPF             delta
>>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>>
>>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>>            1706379      faults
>>>>>            1674599      spf
>>>>>              30588      pagefault:spf_vma_changed
>>>>>                  0      pagefault:spf_vma_noanon
>>>>>                363      pagefault:spf_vma_notsup
>>>>>                  0      pagefault:spf_vma_access
>>>>>                  0      pagefault:spf_pmd_changed
>>>>>
>>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>>            1874773      faults
>>>>>            1461153      spf
>>>>>             413293      pagefault:spf_vma_changed
>>>>>                  0      pagefault:spf_vma_noanon
>>>>>                200      pagefault:spf_vma_notsup
>>>>>                  0      pagefault:spf_vma_access
>>>>>                  0      pagefault:spf_pmd_changed
>>>>>
>>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>>> leading the ebizzy performance boost.
>>>>>
>>>>> ------------------
>>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>>    and Minchan Kim, hopefully.
>>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>>    __do_page_fault().
>>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>>    instead
>>>>>    of aborting the speculative page fault handling. Dropping the now
>>>>> useless
>>>>>    trace event pagefault:spf_pte_lock.
>>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>>    additional tests done didn't show a significant performance improvement.
>>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>>
>>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>>
>>>>>
>>>>> Laurent Dufour (20):
>>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>>   mm: make pte_unmap_same compatible with SPF
>>>>>   mm: introduce INIT_VMA()
>>>>>   mm: protect VMA modifications using VMA sequence count
>>>>>   mm: protect mremap() against SPF hanlder
>>>>>   mm: protect SPF handler against anon_vma changes
>>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>>   mm: introduce __vm_normal_page()
>>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>>   mm: protect mm_rb tree with a rwlock
>>>>>   mm: adding speculative page fault failure trace events
>>>>>   perf: add a speculative page fault sw event
>>>>>   perf tools: add support for the SPF perf event
>>>>>   mm: add speculative page fault vmstats
>>>>>   powerpc/mm: add speculative page fault
>>>>>
>>>>> Mahendran Ganesh (2):
>>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>>   arm64/mm: add speculative page fault
>>>>>
>>>>> Peter Zijlstra (4):
>>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>>   mm: VMA sequence count
>>>>>   mm: provide speculative fault infrastructure
>>>>>   x86/mm: add speculative pagefault handling
>>>>>
>>>>>  arch/arm64/Kconfig                    |   1 +
>>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>>  arch/x86/Kconfig                      |   1 +
>>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>>  fs/exec.c                             |   2 +-
>>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>>  fs/userfaultfd.c                      |  17 +-
>>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>>  include/linux/migrate.h               |   4 +-
>>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>>  include/linux/mm_types.h              |   7 +
>>>>>  include/linux/pagemap.h               |   4 +-
>>>>>  include/linux/rmap.h                  |  12 +-
>>>>>  include/linux/swap.h                  |  10 +-
>>>>>  include/linux/vm_event_item.h         |   3 +
>>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>>  kernel/fork.c                         |   5 +-
>>>>>  mm/Kconfig                            |  22 ++
>>>>>  mm/huge_memory.c                      |   6 +-
>>>>>  mm/hugetlb.c                          |   2 +
>>>>>  mm/init-mm.c                          |   3 +
>>>>>  mm/internal.h                         |  20 ++
>>>>>  mm/khugepaged.c                       |   5 +
>>>>>  mm/madvise.c                          |   6 +-
>>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>>  mm/mempolicy.c                        |  51 ++-
>>>>>  mm/migrate.c                          |   6 +-
>>>>>  mm/mlock.c                            |  13 +-
>>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>>  mm/mprotect.c                         |   4 +-
>>>>>  mm/mremap.c                           |  13 +
>>>>>  mm/nommu.c                            |   2 +-
>>>>>  mm/rmap.c                             |   5 +-
>>>>>  mm/swap.c                             |   6 +-
>>>>>  mm/swap_state.c                       |   8 +-
>>>>>  mm/vmstat.c                           |   5 +-
>>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>>  tools/perf/util/evsel.c               |   1 +
>>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>>  tools/perf/util/python.c              |   1 +
>>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>>
>>>>> --
>>>>> 2.7.4
>>>>>
>>>>>
>>>>
>>>
>>
> 
> 

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

Hi Laurent,


For the test result on Intel 4s skylake platform (192 CPUs, 768G Memory), the below test cases all were run 3 times.
I check the test results, only page_fault3_thread/enable THP have 6% stddev for head commit, other tests have lower stddev.

And I did not find other high variation on test case result.

a). Enable THP
testcase                          base     stddev       change      head     stddev         metric
page_fault3/enable THP           10519      ± 3%        -20.5%      8368      ±6%          will-it-scale.per_thread_ops
page_fault2/enalbe THP            8281      ± 2%        -18.8%      6728                   will-it-scale.per_thread_ops
brk1/eanble THP                 998475                   -2.2%    976893                   will-it-scale.per_process_ops
context_switch1/enable THP      223910                   -1.3%    220930                   will-it-scale.per_process_ops
context_switch1/enable THP      233722                   -1.0%    231288                   will-it-scale.per_thread_ops

b). Disable THP
page_fault3/disable THP          10856                  -23.1%      8344                   will-it-scale.per_thread_ops
page_fault2/disable THP           8147                  -18.8%      6613                   will-it-scale.per_thread_ops
brk1/disable THP                   957                    -7.9%      881                   will-it-scale.per_thread_ops
context_switch1/disable THP     237006                    -2.2%    231907                  will-it-scale.per_thread_ops
brk1/disable THP                997317                    -2.0%    977778                  will-it-scale.per_process_ops
page_fault3/disable THP         467454                    -1.8%    459251                  will-it-scale.per_process_ops
context_switch1/disable THP     224431                    -1.3%    221567                  will-it-scale.per_process_ops


Best regards,
Haiyan Song
________________________________________
From: Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Monday, July 02, 2018 4:59 PM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 11/06/2018 09:49, Song, HaiyanX wrote:
> Hi Laurent,
>
> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
> V9 patch serials.
>
> The regression result is sorted by the metric will-it-scale.per_thread_ops.
> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
> commit id:
>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
> Benchmark: will-it-scale
> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
>
> Metrics:
>   will-it-scale.per_process_ops=processes/nr_cpu
>   will-it-scale.per_thread_ops=threads/nr_cpu
>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> THP: enable / disable
> nr_task:100%
>
> 1. Regressions:
>
> a). Enable THP
> testcase                          base           change      head           metric
> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
>
> b). Disable THP
> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
>
> Notes: for the above  values of test result, the higher is better.

I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
get reproducible results. The results have huge variation, even on the vanilla
kernel, and I can't state on any changes due to that.

I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
measure any changes between the vanilla and the SPF patched ones:

test THP enabled                4.17.0-rc4-mm1  spf             delta
page_fault3_threads             2697.7          2683.5          -0.53%
page_fault2_threads             170660.6        169574.1        -0.64%
context_switch1_threads         6915269.2       6877507.3       -0.55%
context_switch1_processes       6478076.2       6529493.5       0.79%
brk1                            243391.2        238527.5        -2.00%

Tests were run 10 times, no high variation detected.

Did you see high variation on your side ? How many times the test were run to
compute the average values ?

Thanks,
Laurent.


>
> 2. Improvement: not found improvement based on the selected test cases.
>
>
> Best regards
> Haiyan Song
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Monday, May 28, 2018 4:54 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
>
> On 28/05/2018 10:22, Haiyan Song wrote:
>> Hi Laurent,
>>
>> Yes, these tests are done on V9 patch.
>
> Do you plan to give this V11 a run ?
>
>>
>>
>> Best regards,
>> Haiyan Song
>>
>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>
>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>> tested on Intel 4s Skylake platform.
>>>
>>> Hi,
>>>
>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>> series" while responding to the v11 header series...
>>> Were these tests done on v9 or v11 ?
>>>
>>> Cheers,
>>> Laurent.
>>>
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>> Commit id:
>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>> Benchmark suite: will-it-scale
>>>> Download link:
>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>> Metrics:
>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task: 100%
>>>>
>>>> 1. Regressions:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>
>>>> 2. Improvements:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>
>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>> on head commit is better than that on base commit for this benchmark.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>>
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>> page fault without holding the mm semaphore [1].
>>>>
>>>> The idea is to try to handle user space page faults without holding the
>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>> process since the page fault handler will not wait for other threads memory
>>>> layout change to be done, assuming that this change is done in another part
>>>> of the process's memory space. This type page fault is named speculative
>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>> is failing its processing and a classic page fault is then tried.
>>>>
>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>> freeing operation which was hitting the performance by 20% as reported by
>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>> limiting the locking contention to these operations which are expected to
>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>> our back a reference count is added and 2 services (get_vma() and
>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>> benchmark anymore.
>>>>
>>>> The VMA's attributes checked during the speculative page fault processing
>>>> have to be protected against parallel changes. This is done by using a per
>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>> handler to fast check for parallel changes in progress and to abort the
>>>> speculative page fault in that case.
>>>>
>>>> Once the VMA has been found, the speculative page fault handler would check
>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>> checked during the page fault are modified.
>>>>
>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>> leading to touching this PTE will need to lock the page table, so no
>>>> parallel change is possible at this time.
>>>>
>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>> different from the one recorded at the beginning of the SPF operation, the
>>>> classic page fault handler will be called to handle the operation while
>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>> PTE.
>>>>
>>>> In pseudo code, this could be seen as:
>>>>     speculative_page_fault()
>>>>     {
>>>>             vma = get_vma()
>>>>             check vma sequence count
>>>>             check vma's support
>>>>             disable interrupt
>>>>                   check pgd,p4d,...,pte
>>>>                   save pmd and pte in vmf
>>>>                   save vma sequence counter in vmf
>>>>             enable interrupt
>>>>             check vma sequence count
>>>>             handle_pte_fault(vma)
>>>>                     ..
>>>>                     page = alloc_page()
>>>>                     pte_map_lock()
>>>>                             disable interrupt
>>>>                                     abort if sequence counter has changed
>>>>                                     abort if pmd or pte has changed
>>>>                                     pte map and lock
>>>>                             enable interrupt
>>>>                     if abort
>>>>                        free page
>>>>                        abort
>>>>                     ...
>>>>     }
>>>>
>>>>     arch_fault_handler()
>>>>     {
>>>>             if (speculative_page_fault(&vma))
>>>>                goto done
>>>>     again:
>>>>             lock(mmap_sem)
>>>>             vma = find_vma();
>>>>             handle_pte_fault(vma);
>>>>             if retry
>>>>                unlock(mmap_sem)
>>>>                goto again;
>>>>     done:
>>>>             handle fault error
>>>>     }
>>>>
>>>> Support for THP is not done because when checking for the PMD, we can be
>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>
>>>> This series add a new software performance event named 'speculative-faults'
>>>> or 'spf'. It counts the number of successful page fault event handled
>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>> counting the total number of page fault events while 'spf' is only counting
>>>> the part of the faults processed speculatively.
>>>>
>>>> There are some trace events introduced by this series. They allow
>>>> identifying why the page faults were not processed speculatively. This
>>>> doesn't take in account the faults generated by a monothreaded process
>>>> which directly processed while holding the mmap_sem. This trace events are
>>>> grouped in a system named 'pagefault', they are:
>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>    back.
>>>>
>>>> To record all the related events, the easier is to run perf with the
>>>> following arguments :
>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>
>>>> There is also a dedicated vmstat counter showing the number of successful
>>>> page fault handled speculatively. I can be seen this way:
>>>> $ grep speculative_pgfault /proc/vmstat
>>>>
>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>> on x86, PowerPC and arm64.
>>>>
>>>> ---------------------
>>>> Real Workload results
>>>>
>>>> As mentioned in previous email, we did non official runs using a "popular
>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>> which showed a 30% improvements in the number of transaction processed per
>>>> second. This run has been done on the v6 series, but changes introduced in
>>>> this new version should not impact the performance boost seen.
>>>>
>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>> series:
>>>>                 vanilla         spf
>>>> faults          89.418          101.364         +13%
>>>> spf                n/a           97.989
>>>>
>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>> way.
>>>>
>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>> it a try on an android device. He reported that the application launch time
>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>> 20%.
>>>>
>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>
>>>> Application                             4.9     4.9+spf delta
>>>> com.tencent.mm                          416     389     -7%
>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>> com.tencent.mtt                         455     454     0%
>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>> com.autonavi.minimap                    711     701     -1%
>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>> com.immomo.momo                         501     487     -3%
>>>> com.tencent.peng                        2145    2112    -2%
>>>> com.smile.gifmaker                      491     461     -6%
>>>> com.baidu.BaiduMap                      479     366     -23%
>>>> com.taobao.taobao                       1341    1198    -11%
>>>> com.baidu.searchbox                     333     314     -6%
>>>> com.tencent.mobileqq                    394     384     -3%
>>>> com.sina.weibo                          907     906     0%
>>>> com.youku.phone                         816     731     -11%
>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>> com.UCMobile                            415     411     -1%
>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>> com.tencent.qqmusic                     336     329     -2%
>>>> com.sankuai.meituan                     1661    1302    -22%
>>>> com.netease.cloudmusic                  1193    1200    1%
>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>
>>>> ------------------
>>>> Benchmarks results
>>>>
>>>> Base kernel is v4.17.0-rc4-mm1
>>>> SPF is BASE + this series
>>>>
>>>> Kernbench:
>>>> ----------
>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>> kernel (kernel is build 5 times):
>>>>
>>>> Average Half load -j 8
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>
>>>> Average Optimal load -j 16
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          526743764      faults
>>>>                210      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>               2278      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Very few speculative page faults were recorded as most of the processes
>>>> involved are monothreaded (sounds that on this architecture some threads
>>>> were created during the kernel build processing).
>>>>
>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>
>>>> Average Half load -j 40
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>
>>>> Average Optimal load -j 80
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          116730856      faults
>>>>                  0      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                476      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>> there is no impact on the performance.
>>>>
>>>> Ebizzy:
>>>> -------
>>>> The test is counting the number of records per second it can manage, the
>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>> consistent result I repeated the test 100 times and measure the average
>>>> result. The number is the record processes per second, the higher is the
>>>> best.
>>>>
>>>>                 BASE            SPF             delta
>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>
>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>            1706379      faults
>>>>            1674599      spf
>>>>              30588      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                363      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>            1874773      faults
>>>>            1461153      spf
>>>>             413293      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                200      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>> leading the ebizzy performance boost.
>>>>
>>>> ------------------
>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>    and Minchan Kim, hopefully.
>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>    __do_page_fault().
>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>    instead
>>>>    of aborting the speculative page fault handling. Dropping the now
>>>> useless
>>>>    trace event pagefault:spf_pte_lock.
>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>    additional tests done didn't show a significant performance improvement.
>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>
>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>
>>>>
>>>> Laurent Dufour (20):
>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>   mm: make pte_unmap_same compatible with SPF
>>>>   mm: introduce INIT_VMA()
>>>>   mm: protect VMA modifications using VMA sequence count
>>>>   mm: protect mremap() against SPF hanlder
>>>>   mm: protect SPF handler against anon_vma changes
>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>   mm: introduce __vm_normal_page()
>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>   mm: protect mm_rb tree with a rwlock
>>>>   mm: adding speculative page fault failure trace events
>>>>   perf: add a speculative page fault sw event
>>>>   perf tools: add support for the SPF perf event
>>>>   mm: add speculative page fault vmstats
>>>>   powerpc/mm: add speculative page fault
>>>>
>>>> Mahendran Ganesh (2):
>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   arm64/mm: add speculative page fault
>>>>
>>>> Peter Zijlstra (4):
>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>   mm: VMA sequence count
>>>>   mm: provide speculative fault infrastructure
>>>>   x86/mm: add speculative pagefault handling
>>>>
>>>>  arch/arm64/Kconfig                    |   1 +
>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>  arch/x86/Kconfig                      |   1 +
>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>  fs/exec.c                             |   2 +-
>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>  fs/userfaultfd.c                      |  17 +-
>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>  include/linux/migrate.h               |   4 +-
>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>  include/linux/mm_types.h              |   7 +
>>>>  include/linux/pagemap.h               |   4 +-
>>>>  include/linux/rmap.h                  |  12 +-
>>>>  include/linux/swap.h                  |  10 +-
>>>>  include/linux/vm_event_item.h         |   3 +
>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>  kernel/fork.c                         |   5 +-
>>>>  mm/Kconfig                            |  22 ++
>>>>  mm/huge_memory.c                      |   6 +-
>>>>  mm/hugetlb.c                          |   2 +
>>>>  mm/init-mm.c                          |   3 +
>>>>  mm/internal.h                         |  20 ++
>>>>  mm/khugepaged.c                       |   5 +
>>>>  mm/madvise.c                          |   6 +-
>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>  mm/mempolicy.c                        |  51 ++-
>>>>  mm/migrate.c                          |   6 +-
>>>>  mm/mlock.c                            |  13 +-
>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>  mm/mprotect.c                         |   4 +-
>>>>  mm/mremap.c                           |  13 +
>>>>  mm/nommu.c                            |   2 +-
>>>>  mm/rmap.c                             |   5 +-
>>>>  mm/swap.c                             |   6 +-
>>>>  mm/swap_state.c                       |   8 +-
>>>>  mm/vmstat.c                           |   5 +-
>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>  tools/perf/util/evsel.c               |   1 +
>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>  tools/perf/util/python.c              |   1 +
>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>
>>>> --
>>>> 2.7.4
>>>>
>>>>
>>>
>>
>

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

On 11/06/2018 09:49, Song, HaiyanX wrote:
> Hi Laurent,
> 
> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
> V9 patch serials.
> 
> The regression result is sorted by the metric will-it-scale.per_thread_ops.
> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
> commit id:
>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
> Benchmark: will-it-scale
> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
> 
> Metrics:
>   will-it-scale.per_process_ops=processes/nr_cpu
>   will-it-scale.per_thread_ops=threads/nr_cpu
>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> THP: enable / disable
> nr_task:100%
> 
> 1. Regressions:
> 
> a). Enable THP
> testcase                          base           change      head           metric
> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
> 
> b). Disable THP
> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
> 
> Notes: for the above  values of test result, the higher is better.

I tried the same tests on my PowerPC victim VM (1024 CPUs, 11TB) and I can't
get reproducible results. The results have huge variation, even on the vanilla
kernel, and I can't state on any changes due to that.

I tried on smaller node (80 CPUs, 32G), and the tests ran better, but I didn't
measure any changes between the vanilla and the SPF patched ones:

test THP enabled		4.17.0-rc4-mm1	spf		delta
page_fault3_threads		2697.7		2683.5		-0.53%
page_fault2_threads		170660.6	169574.1	-0.64%
context_switch1_threads		6915269.2	6877507.3	-0.55%
context_switch1_processes	6478076.2	6529493.5	0.79%
brk1				243391.2	238527.5	-2.00%

Tests were run 10 times, no high variation detected.

Did you see high variation on your side ? How many times the test were run to
compute the average values ?

Thanks,
Laurent.


> 
> 2. Improvement: not found improvement based on the selected test cases.
> 
> 
> Best regards
> Haiyan Song
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Monday, May 28, 2018 4:54 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
> 
> On 28/05/2018 10:22, Haiyan Song wrote:
>> Hi Laurent,
>>
>> Yes, these tests are done on V9 patch.
> 
> Do you plan to give this V11 a run ?
> 
>>
>>
>> Best regards,
>> Haiyan Song
>>
>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>
>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>> tested on Intel 4s Skylake platform.
>>>
>>> Hi,
>>>
>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>> series" while responding to the v11 header series...
>>> Were these tests done on v9 or v11 ?
>>>
>>> Cheers,
>>> Laurent.
>>>
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>> Commit id:
>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>> Benchmark suite: will-it-scale
>>>> Download link:
>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>> Metrics:
>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task: 100%
>>>>
>>>> 1. Regressions:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>
>>>> 2. Improvements:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>
>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>> on head commit is better than that on base commit for this benchmark.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>>
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>> page fault without holding the mm semaphore [1].
>>>>
>>>> The idea is to try to handle user space page faults without holding the
>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>> process since the page fault handler will not wait for other threads memory
>>>> layout change to be done, assuming that this change is done in another part
>>>> of the process's memory space. This type page fault is named speculative
>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>> is failing its processing and a classic page fault is then tried.
>>>>
>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>> freeing operation which was hitting the performance by 20% as reported by
>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>> limiting the locking contention to these operations which are expected to
>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>> our back a reference count is added and 2 services (get_vma() and
>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>> benchmark anymore.
>>>>
>>>> The VMA's attributes checked during the speculative page fault processing
>>>> have to be protected against parallel changes. This is done by using a per
>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>> handler to fast check for parallel changes in progress and to abort the
>>>> speculative page fault in that case.
>>>>
>>>> Once the VMA has been found, the speculative page fault handler would check
>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>> checked during the page fault are modified.
>>>>
>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>> leading to touching this PTE will need to lock the page table, so no
>>>> parallel change is possible at this time.
>>>>
>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>> different from the one recorded at the beginning of the SPF operation, the
>>>> classic page fault handler will be called to handle the operation while
>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>> PTE.
>>>>
>>>> In pseudo code, this could be seen as:
>>>>     speculative_page_fault()
>>>>     {
>>>>             vma = get_vma()
>>>>             check vma sequence count
>>>>             check vma's support
>>>>             disable interrupt
>>>>                   check pgd,p4d,...,pte
>>>>                   save pmd and pte in vmf
>>>>                   save vma sequence counter in vmf
>>>>             enable interrupt
>>>>             check vma sequence count
>>>>             handle_pte_fault(vma)
>>>>                     ..
>>>>                     page = alloc_page()
>>>>                     pte_map_lock()
>>>>                             disable interrupt
>>>>                                     abort if sequence counter has changed
>>>>                                     abort if pmd or pte has changed
>>>>                                     pte map and lock
>>>>                             enable interrupt
>>>>                     if abort
>>>>                        free page
>>>>                        abort
>>>>                     ...
>>>>     }
>>>>
>>>>     arch_fault_handler()
>>>>     {
>>>>             if (speculative_page_fault(&vma))
>>>>                goto done
>>>>     again:
>>>>             lock(mmap_sem)
>>>>             vma = find_vma();
>>>>             handle_pte_fault(vma);
>>>>             if retry
>>>>                unlock(mmap_sem)
>>>>                goto again;
>>>>     done:
>>>>             handle fault error
>>>>     }
>>>>
>>>> Support for THP is not done because when checking for the PMD, we can be
>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>
>>>> This series add a new software performance event named 'speculative-faults'
>>>> or 'spf'. It counts the number of successful page fault event handled
>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>> counting the total number of page fault events while 'spf' is only counting
>>>> the part of the faults processed speculatively.
>>>>
>>>> There are some trace events introduced by this series. They allow
>>>> identifying why the page faults were not processed speculatively. This
>>>> doesn't take in account the faults generated by a monothreaded process
>>>> which directly processed while holding the mmap_sem. This trace events are
>>>> grouped in a system named 'pagefault', they are:
>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>    back.
>>>>
>>>> To record all the related events, the easier is to run perf with the
>>>> following arguments :
>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>
>>>> There is also a dedicated vmstat counter showing the number of successful
>>>> page fault handled speculatively. I can be seen this way:
>>>> $ grep speculative_pgfault /proc/vmstat
>>>>
>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>> on x86, PowerPC and arm64.
>>>>
>>>> ---------------------
>>>> Real Workload results
>>>>
>>>> As mentioned in previous email, we did non official runs using a "popular
>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>> which showed a 30% improvements in the number of transaction processed per
>>>> second. This run has been done on the v6 series, but changes introduced in
>>>> this new version should not impact the performance boost seen.
>>>>
>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>> series:
>>>>                 vanilla         spf
>>>> faults          89.418          101.364         +13%
>>>> spf                n/a           97.989
>>>>
>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>> way.
>>>>
>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>> it a try on an android device. He reported that the application launch time
>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>> 20%.
>>>>
>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>
>>>> Application                             4.9     4.9+spf delta
>>>> com.tencent.mm                          416     389     -7%
>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>> com.tencent.mtt                         455     454     0%
>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>> com.autonavi.minimap                    711     701     -1%
>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>> com.immomo.momo                         501     487     -3%
>>>> com.tencent.peng                        2145    2112    -2%
>>>> com.smile.gifmaker                      491     461     -6%
>>>> com.baidu.BaiduMap                      479     366     -23%
>>>> com.taobao.taobao                       1341    1198    -11%
>>>> com.baidu.searchbox                     333     314     -6%
>>>> com.tencent.mobileqq                    394     384     -3%
>>>> com.sina.weibo                          907     906     0%
>>>> com.youku.phone                         816     731     -11%
>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>> com.UCMobile                            415     411     -1%
>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>> com.tencent.qqmusic                     336     329     -2%
>>>> com.sankuai.meituan                     1661    1302    -22%
>>>> com.netease.cloudmusic                  1193    1200    1%
>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>
>>>> ------------------
>>>> Benchmarks results
>>>>
>>>> Base kernel is v4.17.0-rc4-mm1
>>>> SPF is BASE + this series
>>>>
>>>> Kernbench:
>>>> ----------
>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>> kernel (kernel is build 5 times):
>>>>
>>>> Average Half load -j 8
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>
>>>> Average Optimal load -j 16
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          526743764      faults
>>>>                210      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>               2278      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Very few speculative page faults were recorded as most of the processes
>>>> involved are monothreaded (sounds that on this architecture some threads
>>>> were created during the kernel build processing).
>>>>
>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>
>>>> Average Half load -j 40
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>
>>>> Average Optimal load -j 80
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          116730856      faults
>>>>                  0      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                476      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>> there is no impact on the performance.
>>>>
>>>> Ebizzy:
>>>> -------
>>>> The test is counting the number of records per second it can manage, the
>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>> consistent result I repeated the test 100 times and measure the average
>>>> result. The number is the record processes per second, the higher is the
>>>> best.
>>>>
>>>>                 BASE            SPF             delta
>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>
>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>            1706379      faults
>>>>            1674599      spf
>>>>              30588      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                363      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>            1874773      faults
>>>>            1461153      spf
>>>>             413293      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                200      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>> leading the ebizzy performance boost.
>>>>
>>>> ------------------
>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>    and Minchan Kim, hopefully.
>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>    __do_page_fault().
>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>    instead
>>>>    of aborting the speculative page fault handling. Dropping the now
>>>> useless
>>>>    trace event pagefault:spf_pte_lock.
>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>    additional tests done didn't show a significant performance improvement.
>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>
>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>
>>>>
>>>> Laurent Dufour (20):
>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>   mm: make pte_unmap_same compatible with SPF
>>>>   mm: introduce INIT_VMA()
>>>>   mm: protect VMA modifications using VMA sequence count
>>>>   mm: protect mremap() against SPF hanlder
>>>>   mm: protect SPF handler against anon_vma changes
>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>   mm: introduce __vm_normal_page()
>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>   mm: protect mm_rb tree with a rwlock
>>>>   mm: adding speculative page fault failure trace events
>>>>   perf: add a speculative page fault sw event
>>>>   perf tools: add support for the SPF perf event
>>>>   mm: add speculative page fault vmstats
>>>>   powerpc/mm: add speculative page fault
>>>>
>>>> Mahendran Ganesh (2):
>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   arm64/mm: add speculative page fault
>>>>
>>>> Peter Zijlstra (4):
>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>   mm: VMA sequence count
>>>>   mm: provide speculative fault infrastructure
>>>>   x86/mm: add speculative pagefault handling
>>>>
>>>>  arch/arm64/Kconfig                    |   1 +
>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>  arch/x86/Kconfig                      |   1 +
>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>  fs/exec.c                             |   2 +-
>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>  fs/userfaultfd.c                      |  17 +-
>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>  include/linux/migrate.h               |   4 +-
>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>  include/linux/mm_types.h              |   7 +
>>>>  include/linux/pagemap.h               |   4 +-
>>>>  include/linux/rmap.h                  |  12 +-
>>>>  include/linux/swap.h                  |  10 +-
>>>>  include/linux/vm_event_item.h         |   3 +
>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>  kernel/fork.c                         |   5 +-
>>>>  mm/Kconfig                            |  22 ++
>>>>  mm/huge_memory.c                      |   6 +-
>>>>  mm/hugetlb.c                          |   2 +
>>>>  mm/init-mm.c                          |   3 +
>>>>  mm/internal.h                         |  20 ++
>>>>  mm/khugepaged.c                       |   5 +
>>>>  mm/madvise.c                          |   6 +-
>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>  mm/mempolicy.c                        |  51 ++-
>>>>  mm/migrate.c                          |   6 +-
>>>>  mm/mlock.c                            |  13 +-
>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>  mm/mprotect.c                         |   4 +-
>>>>  mm/mremap.c                           |  13 +
>>>>  mm/nommu.c                            |   2 +-
>>>>  mm/rmap.c                             |   5 +-
>>>>  mm/swap.c                             |   6 +-
>>>>  mm/swap_state.c                       |   8 +-
>>>>  mm/vmstat.c                           |   5 +-
>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>  tools/perf/util/evsel.c               |   1 +
>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>  tools/perf/util/python.c              |   1 +
>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>
>>>> --
>>>> 2.7.4
>>>>
>>>>
>>>
>>
> 

Re: [PATCH v11 00/26] Speculative page faults

[Haiyan Song]

[-- Attachment #1: Type: text/plain, Size: 31691 bytes --]

On Mon, Jun 11, 2018 at 05:15:22PM +0200, Laurent Dufour wrote:

Hi Laurent,

For perf date tested on Intel 4s Skylake platform, here attached the compare result
between base and head commit in attachment, which include the perf-profile comparision information.

And also attached some perf-profile.json captured from test result for page_fault2 and page_fault3 for
checking the regression, thanks.


Best regards,
Haiyan Song



> Hi Haiyan,
> 
> I don't have access to the same hardware you ran the test on, but I give a try
> to those test on a Power8 system (2 sockets, 5 cores/s, 8 threads/c, 80 CPUs 32G).
> I run each will-it-scale test 10 times and compute the average.
> 
> test THP enabled		4.17.0-rc4-mm1	spf		delta
> page_fault3_threads		2697.7		2683.5		-0.53%
> page_fault2_threads		170660.6	169574.1	-0.64%
> context_switch1_threads		6915269.2	6877507.3	-0.55%
> context_switch1_processes	6478076.2	6529493.5	0.79%
> rk1				243391.2	238527.5	-2.00%
> 
> Test were launched with the arguments '-t 80 -s 5', only the average report is
> taken in account. Note that page size is 64K by default on ppc64.
> 
> It would be nice if you could capture some perf data to figure out why the
> page_fault2/3 are showing such a performance regression.
> 
> Thanks,
> Laurent.
> 
> On 11/06/2018 09:49, Song, HaiyanX wrote:
> > Hi Laurent,
> > 
> > Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
> > tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
> > V9 patch serials.
> > 
> > The regression result is sorted by the metric will-it-scale.per_thread_ops.
> > branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
> > commit id:
> >   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
> >   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
> > Benchmark: will-it-scale
> > Download link: https://github.com/antonblanchard/will-it-scale/tree/master
> > 
> > Metrics:
> >   will-it-scale.per_process_ops=processes/nr_cpu
> >   will-it-scale.per_thread_ops=threads/nr_cpu
> >   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> > THP: enable / disable
> > nr_task:100%
> > 
> > 1. Regressions:
> > 
> > a). Enable THP
> > testcase                          base           change      head           metric
> > page_fault3/enable THP           10519          -20.5%       8368      will-it-scale.per_thread_ops
> > page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
> > brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
> > context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
> > context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
> > 
> > b). Disable THP
> > page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
> > page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
> > brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
> > context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
> > brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
> > page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
> > context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
> > 
> > Notes: for the above  values of test result, the higher is better.
> > 
> > 2. Improvement: not found improvement based on the selected test cases.
> > 
> > 
> > Best regards
> > Haiyan Song
> > ________________________________________
> > From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> > Sent: Monday, May 28, 2018 4:54 PM
> > To: Song, HaiyanX
> > Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> > Subject: Re: [PATCH v11 00/26] Speculative page faults
> > 
> > On 28/05/2018 10:22, Haiyan Song wrote:
> >> Hi Laurent,
> >>
> >> Yes, these tests are done on V9 patch.
> > 
> > Do you plan to give this V11 a run ?
> > 
> >>
> >>
> >> Best regards,
> >> Haiyan Song
> >>
> >> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
> >>> On 28/05/2018 07:23, Song, HaiyanX wrote:
> >>>>
> >>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
> >>>> tested on Intel 4s Skylake platform.
> >>>
> >>> Hi,
> >>>
> >>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
> >>> series" while responding to the v11 header series...
> >>> Were these tests done on v9 or v11 ?
> >>>
> >>> Cheers,
> >>> Laurent.
> >>>
> >>>>
> >>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
> >>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
> >>>> Commit id:
> >>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
> >>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
> >>>> Benchmark suite: will-it-scale
> >>>> Download link:
> >>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
> >>>> Metrics:
> >>>>     will-it-scale.per_process_ops=processes/nr_cpu
> >>>>     will-it-scale.per_thread_ops=threads/nr_cpu
> >>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> >>>> THP: enable / disable
> >>>> nr_task: 100%
> >>>>
> >>>> 1. Regressions:
> >>>> a) THP enabled:
> >>>> testcase                        base            change          head       metric
> >>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
> >>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
> >>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
> >>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
> >>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
> >>>>
> >>>> b) THP disabled:
> >>>> testcase                        base            change          head       metric
> >>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
> >>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
> >>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
> >>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
> >>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
> >>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
> >>>>
> >>>> 2. Improvements:
> >>>> a) THP enabled:
> >>>> testcase                        base            change          head       metric
> >>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
> >>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
> >>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
> >>>>
> >>>> b) THP disabled:
> >>>> testcase                        base            change          head       metric
> >>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
> >>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
> >>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
> >>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
> >>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
> >>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
> >>>>
> >>>> Notes: for above values in column "change", the higher value means that the related testcase result
> >>>> on head commit is better than that on base commit for this benchmark.
> >>>>
> >>>>
> >>>> Best regards
> >>>> Haiyan Song
> >>>>
> >>>> ________________________________________
> >>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> >>>> Sent: Thursday, May 17, 2018 7:06 PM
> >>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
> >>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> >>>> Subject: [PATCH v11 00/26] Speculative page faults
> >>>>
> >>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
> >>>> page fault without holding the mm semaphore [1].
> >>>>
> >>>> The idea is to try to handle user space page faults without holding the
> >>>> mmap_sem. This should allow better concurrency for massively threaded
> >>>> process since the page fault handler will not wait for other threads memory
> >>>> layout change to be done, assuming that this change is done in another part
> >>>> of the process's memory space. This type page fault is named speculative
> >>>> page fault. If the speculative page fault fails because of a concurrency is
> >>>> detected or because underlying PMD or PTE tables are not yet allocating, it
> >>>> is failing its processing and a classic page fault is then tried.
> >>>>
> >>>> The speculative page fault (SPF) has to look for the VMA matching the fault
> >>>> address without holding the mmap_sem, this is done by introducing a rwlock
> >>>> which protects the access to the mm_rb tree. Previously this was done using
> >>>> SRCU but it was introducing a lot of scheduling to process the VMA's
> >>>> freeing operation which was hitting the performance by 20% as reported by
> >>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
> >>>> limiting the locking contention to these operations which are expected to
> >>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
> >>>> our back a reference count is added and 2 services (get_vma() and
> >>>> put_vma()) are introduced to handle the reference count. Once a VMA is
> >>>> fetched from the RB tree using get_vma(), it must be later freed using
> >>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
> >>>> benchmark anymore.
> >>>>
> >>>> The VMA's attributes checked during the speculative page fault processing
> >>>> have to be protected against parallel changes. This is done by using a per
> >>>> VMA sequence lock. This sequence lock allows the speculative page fault
> >>>> handler to fast check for parallel changes in progress and to abort the
> >>>> speculative page fault in that case.
> >>>>
> >>>> Once the VMA has been found, the speculative page fault handler would check
> >>>> for the VMA's attributes to verify that the page fault has to be handled
> >>>> correctly or not. Thus, the VMA is protected through a sequence lock which
> >>>> allows fast detection of concurrent VMA changes. If such a change is
> >>>> detected, the speculative page fault is aborted and a *classic* page fault
> >>>> is tried.  VMA sequence lockings are added when VMA attributes which are
> >>>> checked during the page fault are modified.
> >>>>
> >>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
> >>>> so once the page table is locked, the VMA is valid, so any other changes
> >>>> leading to touching this PTE will need to lock the page table, so no
> >>>> parallel change is possible at this time.
> >>>>
> >>>> The locking of the PTE is done with interrupts disabled, this allows
> >>>> checking for the PMD to ensure that there is not an ongoing collapsing
> >>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
> >>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
> >>>> valid at the time the PTE is locked, we have the guarantee that the
> >>>> collapsing operation will have to wait on the PTE lock to move forward.
> >>>> This allows the SPF handler to map the PTE safely. If the PMD value is
> >>>> different from the one recorded at the beginning of the SPF operation, the
> >>>> classic page fault handler will be called to handle the operation while
> >>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
> >>>> the lock is done using spin_trylock() to avoid dead lock when handling a
> >>>> page fault while a TLB invalidate is requested by another CPU holding the
> >>>> PTE.
> >>>>
> >>>> In pseudo code, this could be seen as:
> >>>>     speculative_page_fault()
> >>>>     {
> >>>>             vma = get_vma()
> >>>>             check vma sequence count
> >>>>             check vma's support
> >>>>             disable interrupt
> >>>>                   check pgd,p4d,...,pte
> >>>>                   save pmd and pte in vmf
> >>>>                   save vma sequence counter in vmf
> >>>>             enable interrupt
> >>>>             check vma sequence count
> >>>>             handle_pte_fault(vma)
> >>>>                     ..
> >>>>                     page = alloc_page()
> >>>>                     pte_map_lock()
> >>>>                             disable interrupt
> >>>>                                     abort if sequence counter has changed
> >>>>                                     abort if pmd or pte has changed
> >>>>                                     pte map and lock
> >>>>                             enable interrupt
> >>>>                     if abort
> >>>>                        free page
> >>>>                        abort
> >>>>                     ...
> >>>>     }
> >>>>
> >>>>     arch_fault_handler()
> >>>>     {
> >>>>             if (speculative_page_fault(&vma))
> >>>>                goto done
> >>>>     again:
> >>>>             lock(mmap_sem)
> >>>>             vma = find_vma();
> >>>>             handle_pte_fault(vma);
> >>>>             if retry
> >>>>                unlock(mmap_sem)
> >>>>                goto again;
> >>>>     done:
> >>>>             handle fault error
> >>>>     }
> >>>>
> >>>> Support for THP is not done because when checking for the PMD, we can be
> >>>> confused by an in progress collapsing operation done by khugepaged. The
> >>>> issue is that pmd_none() could be true either if the PMD is not already
> >>>> populated or if the underlying PTE are in the way to be collapsed. So we
> >>>> cannot safely allocate a PMD if pmd_none() is true.
> >>>>
> >>>> This series add a new software performance event named 'speculative-faults'
> >>>> or 'spf'. It counts the number of successful page fault event handled
> >>>> speculatively. When recording 'faults,spf' events, the faults one is
> >>>> counting the total number of page fault events while 'spf' is only counting
> >>>> the part of the faults processed speculatively.
> >>>>
> >>>> There are some trace events introduced by this series. They allow
> >>>> identifying why the page faults were not processed speculatively. This
> >>>> doesn't take in account the faults generated by a monothreaded process
> >>>> which directly processed while holding the mmap_sem. This trace events are
> >>>> grouped in a system named 'pagefault', they are:
> >>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
> >>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
> >>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
> >>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
> >>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
> >>>>    back.
> >>>>
> >>>> To record all the related events, the easier is to run perf with the
> >>>> following arguments :
> >>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
> >>>>
> >>>> There is also a dedicated vmstat counter showing the number of successful
> >>>> page fault handled speculatively. I can be seen this way:
> >>>> $ grep speculative_pgfault /proc/vmstat
> >>>>
> >>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
> >>>> on x86, PowerPC and arm64.
> >>>>
> >>>> ---------------------
> >>>> Real Workload results
> >>>>
> >>>> As mentioned in previous email, we did non official runs using a "popular
> >>>> in memory multithreaded database product" on 176 cores SMT8 Power system
> >>>> which showed a 30% improvements in the number of transaction processed per
> >>>> second. This run has been done on the v6 series, but changes introduced in
> >>>> this new version should not impact the performance boost seen.
> >>>>
> >>>> Here are the perf data captured during 2 of these runs on top of the v8
> >>>> series:
> >>>>                 vanilla         spf
> >>>> faults          89.418          101.364         +13%
> >>>> spf                n/a           97.989
> >>>>
> >>>> With the SPF kernel, most of the page fault were processed in a speculative
> >>>> way.
> >>>>
> >>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
> >>>> it a try on an android device. He reported that the application launch time
> >>>> was improved in average by 6%, and for large applications (~100 threads) by
> >>>> 20%.
> >>>>
> >>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
> >>>> MSM845 (8 cores) with 6GB (the less is better):
> >>>>
> >>>> Application                             4.9     4.9+spf delta
> >>>> com.tencent.mm                          416     389     -7%
> >>>> com.eg.android.AlipayGphone             1135    986     -13%
> >>>> com.tencent.mtt                         455     454     0%
> >>>> com.qqgame.hlddz                        1497    1409    -6%
> >>>> com.autonavi.minimap                    711     701     -1%
> >>>> com.tencent.tmgp.sgame                  788     748     -5%
> >>>> com.immomo.momo                         501     487     -3%
> >>>> com.tencent.peng                        2145    2112    -2%
> >>>> com.smile.gifmaker                      491     461     -6%
> >>>> com.baidu.BaiduMap                      479     366     -23%
> >>>> com.taobao.taobao                       1341    1198    -11%
> >>>> com.baidu.searchbox                     333     314     -6%
> >>>> com.tencent.mobileqq                    394     384     -3%
> >>>> com.sina.weibo                          907     906     0%
> >>>> com.youku.phone                         816     731     -11%
> >>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
> >>>> com.UCMobile                            415     411     -1%
> >>>> com.tencent.tmgp.ak                     1464    1431    -2%
> >>>> com.tencent.qqmusic                     336     329     -2%
> >>>> com.sankuai.meituan                     1661    1302    -22%
> >>>> com.netease.cloudmusic                  1193    1200    1%
> >>>> air.tv.douyu.android                    4257    4152    -2%
> >>>>
> >>>> ------------------
> >>>> Benchmarks results
> >>>>
> >>>> Base kernel is v4.17.0-rc4-mm1
> >>>> SPF is BASE + this series
> >>>>
> >>>> Kernbench:
> >>>> ----------
> >>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
> >>>> kernel (kernel is build 5 times):
> >>>>
> >>>> Average Half load -j 8
> >>>>                  Run    (std deviation)
> >>>>                  BASE                   SPF
> >>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
> >>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
> >>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
> >>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
> >>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
> >>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
> >>>>
> >>>> Average Optimal load -j 16
> >>>>                  Run    (std deviation)
> >>>>                  BASE                   SPF
> >>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
> >>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
> >>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
> >>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
> >>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
> >>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
> >>>>
> >>>>
> >>>> During a run on the SPF, perf events were captured:
> >>>>  Performance counter stats for '../kernbench -M':
> >>>>          526743764      faults
> >>>>                210      spf
> >>>>                  3      pagefault:spf_vma_changed
> >>>>                  0      pagefault:spf_vma_noanon
> >>>>               2278      pagefault:spf_vma_notsup
> >>>>                  0      pagefault:spf_vma_access
> >>>>                  0      pagefault:spf_pmd_changed
> >>>>
> >>>> Very few speculative page faults were recorded as most of the processes
> >>>> involved are monothreaded (sounds that on this architecture some threads
> >>>> were created during the kernel build processing).
> >>>>
> >>>> Here are the kerbench results on a 80 CPUs Power8 system:
> >>>>
> >>>> Average Half load -j 40
> >>>>                  Run    (std deviation)
> >>>>                  BASE                   SPF
> >>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
> >>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
> >>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
> >>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
> >>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
> >>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
> >>>>
> >>>> Average Optimal load -j 80
> >>>>                  Run    (std deviation)
> >>>>                  BASE                   SPF
> >>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
> >>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
> >>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
> >>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
> >>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
> >>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
> >>>>
> >>>> During a run on the SPF, perf events were captured:
> >>>>  Performance counter stats for '../kernbench -M':
> >>>>          116730856      faults
> >>>>                  0      spf
> >>>>                  3      pagefault:spf_vma_changed
> >>>>                  0      pagefault:spf_vma_noanon
> >>>>                476      pagefault:spf_vma_notsup
> >>>>                  0      pagefault:spf_vma_access
> >>>>                  0      pagefault:spf_pmd_changed
> >>>>
> >>>> Most of the processes involved are monothreaded so SPF is not activated but
> >>>> there is no impact on the performance.
> >>>>
> >>>> Ebizzy:
> >>>> -------
> >>>> The test is counting the number of records per second it can manage, the
> >>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
> >>>> consistent result I repeated the test 100 times and measure the average
> >>>> result. The number is the record processes per second, the higher is the
> >>>> best.
> >>>>
> >>>>                 BASE            SPF             delta
> >>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
> >>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
> >>>>
> >>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
> >>>>  Performance counter stats for './ebizzy -mTt 16':
> >>>>            1706379      faults
> >>>>            1674599      spf
> >>>>              30588      pagefault:spf_vma_changed
> >>>>                  0      pagefault:spf_vma_noanon
> >>>>                363      pagefault:spf_vma_notsup
> >>>>                  0      pagefault:spf_vma_access
> >>>>                  0      pagefault:spf_pmd_changed
> >>>>
> >>>> And the ones captured during a run on a 80 CPUs Power node:
> >>>>  Performance counter stats for './ebizzy -mTt 80':
> >>>>            1874773      faults
> >>>>            1461153      spf
> >>>>             413293      pagefault:spf_vma_changed
> >>>>                  0      pagefault:spf_vma_noanon
> >>>>                200      pagefault:spf_vma_notsup
> >>>>                  0      pagefault:spf_vma_access
> >>>>                  0      pagefault:spf_pmd_changed
> >>>>
> >>>> In ebizzy's case most of the page fault were handled in a speculative way,
> >>>> leading the ebizzy performance boost.
> >>>>
> >>>> ------------------
> >>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
> >>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
> >>>>    and Minchan Kim, hopefully.
> >>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
> >>>>    __do_page_fault().
> >>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
> >>>>    instead
> >>>>    of aborting the speculative page fault handling. Dropping the now
> >>>> useless
> >>>>    trace event pagefault:spf_pte_lock.
> >>>>  - No more try to reuse the fetched VMA during the speculative page fault
> >>>>    handling when retrying is needed. This adds a lot of complexity and
> >>>>    additional tests done didn't show a significant performance improvement.
> >>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
> >>>>
> >>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
> >>>> [2] https://patchwork.kernel.org/patch/9999687/
> >>>>
> >>>>
> >>>> Laurent Dufour (20):
> >>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
> >>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
> >>>>   mm: make pte_unmap_same compatible with SPF
> >>>>   mm: introduce INIT_VMA()
> >>>>   mm: protect VMA modifications using VMA sequence count
> >>>>   mm: protect mremap() against SPF hanlder
> >>>>   mm: protect SPF handler against anon_vma changes
> >>>>   mm: cache some VMA fields in the vm_fault structure
> >>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
> >>>>   mm: introduce __lru_cache_add_active_or_unevictable
> >>>>   mm: introduce __vm_normal_page()
> >>>>   mm: introduce __page_add_new_anon_rmap()
> >>>>   mm: protect mm_rb tree with a rwlock
> >>>>   mm: adding speculative page fault failure trace events
> >>>>   perf: add a speculative page fault sw event
> >>>>   perf tools: add support for the SPF perf event
> >>>>   mm: add speculative page fault vmstats
> >>>>   powerpc/mm: add speculative page fault
> >>>>
> >>>> Mahendran Ganesh (2):
> >>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >>>>   arm64/mm: add speculative page fault
> >>>>
> >>>> Peter Zijlstra (4):
> >>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
> >>>>   mm: VMA sequence count
> >>>>   mm: provide speculative fault infrastructure
> >>>>   x86/mm: add speculative pagefault handling
> >>>>
> >>>>  arch/arm64/Kconfig                    |   1 +
> >>>>  arch/arm64/mm/fault.c                 |  12 +
> >>>>  arch/powerpc/Kconfig                  |   1 +
> >>>>  arch/powerpc/mm/fault.c               |  16 +
> >>>>  arch/x86/Kconfig                      |   1 +
> >>>>  arch/x86/mm/fault.c                   |  27 +-
> >>>>  fs/exec.c                             |   2 +-
> >>>>  fs/proc/task_mmu.c                    |   5 +-
> >>>>  fs/userfaultfd.c                      |  17 +-
> >>>>  include/linux/hugetlb_inline.h        |   2 +-
> >>>>  include/linux/migrate.h               |   4 +-
> >>>>  include/linux/mm.h                    | 136 +++++++-
> >>>>  include/linux/mm_types.h              |   7 +
> >>>>  include/linux/pagemap.h               |   4 +-
> >>>>  include/linux/rmap.h                  |  12 +-
> >>>>  include/linux/swap.h                  |  10 +-
> >>>>  include/linux/vm_event_item.h         |   3 +
> >>>>  include/trace/events/pagefault.h      |  80 +++++
> >>>>  include/uapi/linux/perf_event.h       |   1 +
> >>>>  kernel/fork.c                         |   5 +-
> >>>>  mm/Kconfig                            |  22 ++
> >>>>  mm/huge_memory.c                      |   6 +-
> >>>>  mm/hugetlb.c                          |   2 +
> >>>>  mm/init-mm.c                          |   3 +
> >>>>  mm/internal.h                         |  20 ++
> >>>>  mm/khugepaged.c                       |   5 +
> >>>>  mm/madvise.c                          |   6 +-
> >>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
> >>>>  mm/mempolicy.c                        |  51 ++-
> >>>>  mm/migrate.c                          |   6 +-
> >>>>  mm/mlock.c                            |  13 +-
> >>>>  mm/mmap.c                             | 229 ++++++++++---
> >>>>  mm/mprotect.c                         |   4 +-
> >>>>  mm/mremap.c                           |  13 +
> >>>>  mm/nommu.c                            |   2 +-
> >>>>  mm/rmap.c                             |   5 +-
> >>>>  mm/swap.c                             |   6 +-
> >>>>  mm/swap_state.c                       |   8 +-
> >>>>  mm/vmstat.c                           |   5 +-
> >>>>  tools/include/uapi/linux/perf_event.h |   1 +
> >>>>  tools/perf/util/evsel.c               |   1 +
> >>>>  tools/perf/util/parse-events.c        |   4 +
> >>>>  tools/perf/util/parse-events.l        |   1 +
> >>>>  tools/perf/util/python.c              |   1 +
> >>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
> >>>>  create mode 100644 include/trace/events/pagefault.h
> >>>>
> >>>> --
> >>>> 2.7.4
> >>>>
> >>>>
> >>>
> >>
> > 
> 

[-- Attachment #2: compare-result.txt --]
[-- Type: text/plain, Size: 185207 bytes --]

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/always/page_fault3/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
         44:3          -13%          43:3     perf-profile.calltrace.cycles-pp.error_entry
         22:3           -6%          22:3     perf-profile.calltrace.cycles-pp.sync_regs.error_entry
         44:3          -13%          44:3     perf-profile.children.cycles-pp.error_entry
         21:3           -7%          21:3     perf-profile.self.cycles-pp.error_entry
         %stddev     %change         %stddev
             \          |                \  
     10519 +-  3%     -20.5%       8368 +-  6%  will-it-scale.per_thread_ops
    118098           +11.2%     131287 +-  2%  will-it-scale.time.involuntary_context_switches
 6.084e+08 +-  3%     -20.4%  4.845e+08 +-  6%  will-it-scale.time.minor_page_faults
      7467            +5.0%       7841        will-it-scale.time.percent_of_cpu_this_job_got
     44922            +5.0%      47176        will-it-scale.time.system_time
   7126337 +-  3%     -15.4%    6025689 +-  6%  will-it-scale.time.voluntary_context_switches
  91905646            -1.3%   90673935        will-it-scale.workload
     27.15 +-  6%      -8.7%      24.80 +- 10%  boot-time.boot
   2516213 +-  6%      +8.3%    2726303        interrupts.CAL:Function_call_interrupts
    388.00 +-  9%     +60.2%     621.67 +- 20%  irq_exception_noise.softirq_nr
     11.28 +-  2%      -1.9        9.37 +-  4%  mpstat.cpu.idle%
     10065 +-140%    +243.4%      34559 +-  4%  numa-numastat.node0.other_node
     18739           -11.6%      16573 +-  3%  uptime.idle
     29406 +-  2%     -11.8%      25929 +-  5%  vmstat.system.cs
    329614 +-  8%     +17.0%     385618 +- 10%  meminfo.DirectMap4k
    237851           +21.2%     288160 +-  5%  meminfo.Inactive
    237615           +21.2%     287924 +-  5%  meminfo.Inactive(anon)
   7917847           -10.7%    7071860        softirqs.RCU
   4784181 +-  3%     -14.5%    4089039 +-  4%  softirqs.SCHED
  45666107 +-  7%     +12.9%   51535472 +-  3%  softirqs.TIMER
 2.617e+09 +-  2%     -13.9%  2.253e+09 +-  6%  cpuidle.C1E.time
   6688774 +-  2%     -12.8%    5835101 +-  5%  cpuidle.C1E.usage
 1.022e+10 +-  2%     -18.0%  8.376e+09 +-  3%  cpuidle.C6.time
  13440993 +-  2%     -16.3%   11243794 +-  4%  cpuidle.C6.usage
     54781 +- 16%     +37.5%      75347 +- 12%  numa-meminfo.node0.Inactive
     54705 +- 16%     +37.7%      75347 +- 12%  numa-meminfo.node0.Inactive(anon)
     52522           +35.0%      70886 +-  6%  numa-meminfo.node2.Inactive
     52443           +34.7%      70653 +-  6%  numa-meminfo.node2.Inactive(anon)
     31046 +-  6%     +30.3%      40457 +- 11%  numa-meminfo.node2.SReclaimable
     58563           +21.1%      70945 +-  6%  proc-vmstat.nr_inactive_anon
     58564           +21.1%      70947 +-  6%  proc-vmstat.nr_zone_inactive_anon
  69701118            -1.2%   68842151        proc-vmstat.pgalloc_normal
 2.765e+10            -1.3%  2.729e+10        proc-vmstat.pgfault
  69330418            -1.2%   68466824        proc-vmstat.pgfree
    118098           +11.2%     131287 +-  2%  time.involuntary_context_switches
 6.084e+08 +-  3%     -20.4%  4.845e+08 +-  6%  time.minor_page_faults
      7467            +5.0%       7841        time.percent_of_cpu_this_job_got
     44922            +5.0%      47176        time.system_time
   7126337 +-  3%     -15.4%    6025689 +-  6%  time.voluntary_context_switches
     13653 +- 16%     +33.5%      18225 +- 12%  numa-vmstat.node0.nr_inactive_anon
     13651 +- 16%     +33.5%      18224 +- 12%  numa-vmstat.node0.nr_zone_inactive_anon
     13069 +-  3%     +30.1%      17001 +-  4%  numa-vmstat.node2.nr_inactive_anon
    134.67 +- 42%     -49.5%      68.00 +- 31%  numa-vmstat.node2.nr_mlock
      7758 +-  6%     +30.4%      10112 +- 11%  numa-vmstat.node2.nr_slab_reclaimable
     13066 +-  3%     +30.1%      16998 +-  4%  numa-vmstat.node2.nr_zone_inactive_anon
      1039 +- 11%     -17.5%     857.33        slabinfo.Acpi-ParseExt.active_objs
      1039 +- 11%     -17.5%     857.33        slabinfo.Acpi-ParseExt.num_objs
      2566 +-  6%      -8.8%       2340 +-  5%  slabinfo.biovec-64.active_objs
      2566 +-  6%      -8.8%       2340 +-  5%  slabinfo.biovec-64.num_objs
    898.33 +-  3%      -9.5%     813.33 +-  3%  slabinfo.kmem_cache_node.active_objs
      1066 +-  2%      -8.0%     981.33 +-  3%  slabinfo.kmem_cache_node.num_objs
      1940            +2.3%       1984        turbostat.Avg_MHz
   6679037 +-  2%     -12.7%    5830270 +-  5%  turbostat.C1E
      2.25 +-  2%      -0.3        1.94 +-  6%  turbostat.C1E%
  13418115           -16.3%   11234510 +-  4%  turbostat.C6
      8.75 +-  2%      -1.6        7.18 +-  3%  turbostat.C6%
      5.99 +-  2%     -14.4%       5.13 +-  4%  turbostat.CPU%c1
      5.01 +-  3%     -20.1%       4.00 +-  4%  turbostat.CPU%c6
      1.77 +-  3%     -34.7%       1.15        turbostat.Pkg%pc2
 1.378e+13            +1.2%  1.394e+13        perf-stat.branch-instructions
      0.98            -0.0        0.94        perf-stat.branch-miss-rate%
 1.344e+11            -2.3%  1.313e+11        perf-stat.branch-misses
 1.076e+11            -1.8%  1.057e+11        perf-stat.cache-misses
 2.258e+11            -2.1%   2.21e+11        perf-stat.cache-references
  17788064 +-  2%     -11.9%   15674207 +-  6%  perf-stat.context-switches
 2.241e+14            +2.4%  2.294e+14        perf-stat.cpu-cycles
 1.929e+13            +2.2%  1.971e+13        perf-stat.dTLB-loads
      4.01            -0.2        3.83        perf-stat.dTLB-store-miss-rate%
 4.519e+11            -1.3%  4.461e+11        perf-stat.dTLB-store-misses
 1.082e+13            +3.6%  1.121e+13        perf-stat.dTLB-stores
  3.02e+10           +23.2%  3.721e+10 +-  3%  perf-stat.iTLB-load-misses
 2.721e+08 +-  8%      -8.8%  2.481e+08 +-  3%  perf-stat.iTLB-loads
 6.985e+13            +1.8%  7.111e+13        perf-stat.instructions
      2313           -17.2%       1914 +-  3%  perf-stat.instructions-per-iTLB-miss
 2.764e+10            -1.3%  2.729e+10        perf-stat.minor-faults
 1.421e+09 +-  2%     -16.4%  1.188e+09 +-  9%  perf-stat.node-load-misses
 1.538e+10            -9.3%  1.395e+10        perf-stat.node-loads
      9.75            +1.4       11.10        perf-stat.node-store-miss-rate%
 3.012e+09           +14.1%  3.437e+09        perf-stat.node-store-misses
 2.789e+10            -1.3%  2.753e+10        perf-stat.node-stores
 2.764e+10            -1.3%  2.729e+10        perf-stat.page-faults
    760059            +3.2%     784235        perf-stat.path-length
    193545 +- 25%     -57.8%      81757 +- 46%  sched_debug.cfs_rq:/.MIN_vruntime.avg
  26516863 +- 19%     -49.7%   13338070 +- 33%  sched_debug.cfs_rq:/.MIN_vruntime.max
   2202271 +- 21%     -53.2%    1029581 +- 38%  sched_debug.cfs_rq:/.MIN_vruntime.stddev
    193545 +- 25%     -57.8%      81757 +- 46%  sched_debug.cfs_rq:/.max_vruntime.avg
  26516863 +- 19%     -49.7%   13338070 +- 33%  sched_debug.cfs_rq:/.max_vruntime.max
   2202271 +- 21%     -53.2%    1029581 +- 38%  sched_debug.cfs_rq:/.max_vruntime.stddev
      0.32 +- 70%    +253.2%       1.14 +- 54%  sched_debug.cfs_rq:/.removed.load_avg.avg
      4.44 +- 70%    +120.7%       9.80 +- 27%  sched_debug.cfs_rq:/.removed.load_avg.stddev
     14.90 +- 70%    +251.0%      52.31 +- 53%  sched_debug.cfs_rq:/.removed.runnable_sum.avg
    205.71 +- 70%    +119.5%     451.60 +- 27%  sched_debug.cfs_rq:/.removed.runnable_sum.stddev
      0.16 +- 70%    +237.9%       0.54 +- 50%  sched_debug.cfs_rq:/.removed.util_avg.avg
      2.23 +- 70%    +114.2%       4.77 +- 24%  sched_debug.cfs_rq:/.removed.util_avg.stddev
    573.70 +-  5%      -9.7%     518.06 +-  6%  sched_debug.cfs_rq:/.util_avg.min
    114.87 +-  8%     +14.1%     131.04 +- 10%  sched_debug.cfs_rq:/.util_est_enqueued.avg
     64.42 +- 54%     -63.9%      23.27 +- 68%  sched_debug.cpu.cpu_load[1].max
      5.05 +- 48%     -55.2%       2.26 +- 51%  sched_debug.cpu.cpu_load[1].stddev
     57.58 +- 59%     -60.3%      22.88 +- 70%  sched_debug.cpu.cpu_load[2].max
     21019 +-  3%     -15.1%      17841 +-  6%  sched_debug.cpu.nr_switches.min
     20797 +-  3%     -15.0%      17670 +-  6%  sched_debug.cpu.sched_count.min
     10287 +-  3%     -15.1%       8736 +-  6%  sched_debug.cpu.sched_goidle.avg
     13693 +-  2%     -10.7%      12233 +-  5%  sched_debug.cpu.sched_goidle.max
      9976 +-  3%     -16.0%       8381 +-  7%  sched_debug.cpu.sched_goidle.min
      0.00 +- 26%     +98.9%       0.00 +- 28%  sched_debug.rt_rq:/.rt_time.min
      4230 +-141%    -100.0%       0.00        latency_stats.avg.trace_module_notify.notifier_call_chain.blocking_notifier_call_chain.do_init_module.load_module.__do_sys_finit_module.do_syscall_64.entry_SYSCALL_64_after_hwframe
     28498 +-141%    -100.0%       0.00        latency_stats.avg.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4065 +-138%     -92.2%     315.33 +- 91%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
      0.00       +3.6e+105%       3641 +-141%  latency_stats.avg.down.console_lock.console_device.tty_lookup_driver.tty_open.chrdev_open.do_dentry_open.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +2.5e+106%      25040 +-141%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +3.4e+106%      34015 +-141%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_get_acl.get_acl.posix_acl_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open
      0.00       +4.8e+106%      47686 +-141%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
      4230 +-141%    -100.0%       0.00        latency_stats.max.trace_module_notify.notifier_call_chain.blocking_notifier_call_chain.do_init_module.load_module.__do_sys_finit_module.do_syscall_64.entry_SYSCALL_64_after_hwframe
     28498 +-141%    -100.0%       0.00        latency_stats.max.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4065 +-138%     -92.2%     315.33 +- 91%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
      4254 +-134%     -88.0%     511.67 +- 90%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
     43093 +- 35%     +76.6%      76099 +-115%  latency_stats.max.blk_execute_rq.scsi_execute.ioctl_internal_command.scsi_set_medium_removal.cdrom_release.[cdrom].sr_block_release.[sr_mod].__blkdev_put.blkdev_close.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64
     24139 +- 70%    +228.5%      79285 +-105%  latency_stats.max.blk_execute_rq.scsi_execute.scsi_test_unit_ready.sr_check_events.[sr_mod].cdrom_check_events.[cdrom].sr_block_check_events.[sr_mod].disk_check_events.disk_clear_events.check_disk_change.sr_block_open.[sr_mod].__blkdev_get.blkdev_get
      0.00       +3.6e+105%       3641 +-141%  latency_stats.max.down.console_lock.console_device.tty_lookup_driver.tty_open.chrdev_open.do_dentry_open.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +2.5e+106%      25040 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +3.4e+106%      34015 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_get_acl.get_acl.posix_acl_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open
      0.00       +6.5e+106%      64518 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
      4230 +-141%    -100.0%       0.00        latency_stats.sum.trace_module_notify.notifier_call_chain.blocking_notifier_call_chain.do_init_module.load_module.__do_sys_finit_module.do_syscall_64.entry_SYSCALL_64_after_hwframe
     28498 +-141%    -100.0%       0.00        latency_stats.sum.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4065 +-138%     -92.2%     315.33 +- 91%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
     57884 +-  9%     +47.3%      85264 +-118%  latency_stats.sum.blk_execute_rq.scsi_execute.ioctl_internal_command.scsi_set_medium_removal.cdrom_release.[cdrom].sr_block_release.[sr_mod].__blkdev_put.blkdev_close.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64
      0.00       +3.6e+105%       3641 +-141%  latency_stats.sum.down.console_lock.console_device.tty_lookup_driver.tty_open.chrdev_open.do_dentry_open.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +2.5e+106%      25040 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +3.4e+106%      34015 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_get_acl.get_acl.posix_acl_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open
      0.00       +9.5e+106%      95373 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
     11.70           -11.7        0.00        perf-profile.calltrace.cycles-pp.__do_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
     11.52           -11.5        0.00        perf-profile.calltrace.cycles-pp.shmem_fault.__do_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
     10.44           -10.4        0.00        perf-profile.calltrace.cycles-pp.shmem_getpage_gfp.shmem_fault.__do_fault.__handle_mm_fault.handle_mm_fault
      9.83            -9.8        0.00        perf-profile.calltrace.cycles-pp.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault.__handle_mm_fault
      9.55            -9.5        0.00        perf-profile.calltrace.cycles-pp.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      9.35            -9.3        0.00        perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      6.81            -6.8        0.00        perf-profile.calltrace.cycles-pp.page_add_file_rmap.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault
      7.71            -0.3        7.45        perf-profile.calltrace.cycles-pp.find_get_entry.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault
      0.59 +-  7%      -0.2        0.35 +- 70%  perf-profile.calltrace.cycles-pp.smp_apic_timer_interrupt.apic_timer_interrupt.__do_page_fault.do_page_fault.page_fault
      0.59 +-  7%      -0.2        0.35 +- 70%  perf-profile.calltrace.cycles-pp.apic_timer_interrupt.__do_page_fault.do_page_fault.page_fault
     10.41            -0.2       10.24        perf-profile.calltrace.cycles-pp.native_irq_return_iret
      7.68            -0.1        7.60        perf-profile.calltrace.cycles-pp.swapgs_restore_regs_and_return_to_usermode
      0.76            -0.1        0.70        perf-profile.calltrace.cycles-pp.down_read_trylock.__do_page_fault.do_page_fault.page_fault
      1.38            -0.0        1.34        perf-profile.calltrace.cycles-pp.do_page_fault
      1.05            -0.0        1.02        perf-profile.calltrace.cycles-pp.trace_graph_entry.do_page_fault
      0.92            +0.0        0.94        perf-profile.calltrace.cycles-pp.find_vma.__do_page_fault.do_page_fault.page_fault
      0.91            +0.0        0.93        perf-profile.calltrace.cycles-pp.vmacache_find.find_vma.__do_page_fault.do_page_fault.page_fault
      0.65            +0.0        0.67        perf-profile.calltrace.cycles-pp.set_page_dirty.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      0.62            +0.0        0.66        perf-profile.calltrace.cycles-pp.page_mapping.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault
      4.15            +0.1        4.27        perf-profile.calltrace.cycles-pp.page_remove_rmap.unmap_page_range.unmap_vmas.unmap_region.do_munmap
     10.17            +0.2       10.39        perf-profile.calltrace.cycles-pp.munmap
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.entry_SYSCALL_64_after_hwframe.munmap
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.do_syscall_64.entry_SYSCALL_64_after_hwframe.munmap
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.unmap_region.do_munmap.vm_munmap.__x64_sys_munmap.do_syscall_64
      9.54            +0.2        9.76        perf-profile.calltrace.cycles-pp.unmap_page_range.unmap_vmas.unmap_region.do_munmap.vm_munmap
      9.54            +0.2        9.76        perf-profile.calltrace.cycles-pp.unmap_vmas.unmap_region.do_munmap.vm_munmap.__x64_sys_munmap
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.do_munmap.vm_munmap.__x64_sys_munmap.do_syscall_64.entry_SYSCALL_64_after_hwframe
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.vm_munmap.__x64_sys_munmap.do_syscall_64.entry_SYSCALL_64_after_hwframe.munmap
      9.56            +0.2        9.78        perf-profile.calltrace.cycles-pp.__x64_sys_munmap.do_syscall_64.entry_SYSCALL_64_after_hwframe.munmap
      0.00            +0.6        0.56 +-  2%  perf-profile.calltrace.cycles-pp.lock_page_memcg.page_add_file_rmap.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +0.6        0.59        perf-profile.calltrace.cycles-pp.page_mapping.set_page_dirty.fault_dirty_shared_page.handle_pte_fault.__handle_mm_fault
      0.00            +0.6        0.60        perf-profile.calltrace.cycles-pp.current_time.file_update_time.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +0.7        0.68        perf-profile.calltrace.cycles-pp.___might_sleep.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault
      0.00            +0.7        0.74        perf-profile.calltrace.cycles-pp.unlock_page.fault_dirty_shared_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +0.8        0.80        perf-profile.calltrace.cycles-pp.set_page_dirty.fault_dirty_shared_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +0.9        0.88        perf-profile.calltrace.cycles-pp._raw_spin_lock.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +0.9        0.91        perf-profile.calltrace.cycles-pp.__set_page_dirty_no_writeback.fault_dirty_shared_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +1.3        1.27        perf-profile.calltrace.cycles-pp.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +1.3        1.30        perf-profile.calltrace.cycles-pp.file_update_time.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +2.8        2.76        perf-profile.calltrace.cycles-pp.fault_dirty_shared_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +6.8        6.81        perf-profile.calltrace.cycles-pp.page_add_file_rmap.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +9.4        9.39        perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +9.6        9.59        perf-profile.calltrace.cycles-pp.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +9.8        9.77        perf-profile.calltrace.cycles-pp.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault
      0.00           +10.4       10.37        perf-profile.calltrace.cycles-pp.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault
      0.00           +11.5       11.46        perf-profile.calltrace.cycles-pp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00           +11.6       11.60        perf-profile.calltrace.cycles-pp.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00           +26.6       26.62        perf-profile.calltrace.cycles-pp.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      7.88            -0.3        7.61        perf-profile.children.cycles-pp.find_get_entry
      1.34 +-  8%      -0.2        1.16 +-  2%  perf-profile.children.cycles-pp.hrtimer_interrupt
     10.41            -0.2       10.24        perf-profile.children.cycles-pp.native_irq_return_iret
      0.38 +- 28%      -0.1        0.26 +-  4%  perf-profile.children.cycles-pp.tick_sched_timer
     11.80            -0.1       11.68        perf-profile.children.cycles-pp.__do_fault
      0.55 +- 15%      -0.1        0.43 +-  2%  perf-profile.children.cycles-pp.__hrtimer_run_queues
      0.60            -0.1        0.51        perf-profile.children.cycles-pp.pmd_devmap_trans_unstable
      0.38 +- 13%      -0.1        0.29 +-  4%  perf-profile.children.cycles-pp.ktime_get
      7.68            -0.1        7.60        perf-profile.children.cycles-pp.swapgs_restore_regs_and_return_to_usermode
      5.18            -0.1        5.12        perf-profile.children.cycles-pp.trace_graph_entry
      0.79            -0.1        0.73        perf-profile.children.cycles-pp.down_read_trylock
      7.83            -0.1        7.76        perf-profile.children.cycles-pp.sync_regs
      3.01            -0.1        2.94        perf-profile.children.cycles-pp.fault_dirty_shared_page
      1.02            -0.1        0.96        perf-profile.children.cycles-pp._raw_spin_lock
      4.66            -0.1        4.61        perf-profile.children.cycles-pp.prepare_ftrace_return
      0.37 +-  8%      -0.1        0.32 +-  3%  perf-profile.children.cycles-pp.current_kernel_time64
      5.26            -0.1        5.21        perf-profile.children.cycles-pp.ftrace_graph_caller
      0.66 +-  5%      -0.1        0.61        perf-profile.children.cycles-pp.current_time
      0.18 +-  5%      -0.0        0.15 +-  3%  perf-profile.children.cycles-pp.update_process_times
      0.27            -0.0        0.26        perf-profile.children.cycles-pp._cond_resched
      0.16            -0.0        0.15 +-  3%  perf-profile.children.cycles-pp.rcu_all_qs
      0.94            +0.0        0.95        perf-profile.children.cycles-pp.vmacache_find
      0.48            +0.0        0.50        perf-profile.children.cycles-pp.__mod_node_page_state
      0.17            +0.0        0.19 +-  2%  perf-profile.children.cycles-pp.__unlock_page_memcg
      1.07            +0.0        1.10        perf-profile.children.cycles-pp.find_vma
      0.79 +-  3%      +0.1        0.86 +-  2%  perf-profile.children.cycles-pp.lock_page_memcg
      4.29            +0.1        4.40        perf-profile.children.cycles-pp.page_remove_rmap
      1.39 +-  2%      +0.1        1.52        perf-profile.children.cycles-pp.file_update_time
      0.00            +0.2        0.16        perf-profile.children.cycles-pp.__vm_normal_page
      9.63            +0.2        9.84        perf-profile.children.cycles-pp.entry_SYSCALL_64_after_hwframe
      9.63            +0.2        9.84        perf-profile.children.cycles-pp.do_syscall_64
      9.63            +0.2        9.84        perf-profile.children.cycles-pp.unmap_page_range
     10.17            +0.2       10.39        perf-profile.children.cycles-pp.munmap
      9.56            +0.2        9.78        perf-profile.children.cycles-pp.unmap_region
      9.56            +0.2        9.78        perf-profile.children.cycles-pp.do_munmap
      9.56            +0.2        9.78        perf-profile.children.cycles-pp.vm_munmap
      9.56            +0.2        9.78        perf-profile.children.cycles-pp.__x64_sys_munmap
      9.54            +0.2        9.77        perf-profile.children.cycles-pp.unmap_vmas
      1.01            +0.2        1.25        perf-profile.children.cycles-pp.___might_sleep
      0.00            +1.6        1.59        perf-profile.children.cycles-pp.pte_map_lock
      0.00           +26.9       26.89        perf-profile.children.cycles-pp.handle_pte_fault
      4.25            -1.0        3.24        perf-profile.self.cycles-pp.__handle_mm_fault
      1.42            -0.3        1.11        perf-profile.self.cycles-pp.alloc_set_pte
      4.87            -0.3        4.59        perf-profile.self.cycles-pp.find_get_entry
     10.41            -0.2       10.24        perf-profile.self.cycles-pp.native_irq_return_iret
      0.37 +- 13%      -0.1        0.28 +-  4%  perf-profile.self.cycles-pp.ktime_get
      0.60            -0.1        0.51        perf-profile.self.cycles-pp.pmd_devmap_trans_unstable
      7.50            -0.1        7.42        perf-profile.self.cycles-pp.swapgs_restore_regs_and_return_to_usermode
      7.83            -0.1        7.76        perf-profile.self.cycles-pp.sync_regs
      4.85            -0.1        4.79        perf-profile.self.cycles-pp.trace_graph_entry
      1.01            -0.1        0.95        perf-profile.self.cycles-pp._raw_spin_lock
      0.78            -0.1        0.73        perf-profile.self.cycles-pp.down_read_trylock
      0.36 +-  9%      -0.1        0.31 +-  4%  perf-profile.self.cycles-pp.current_kernel_time64
      0.28            -0.0        0.23 +-  2%  perf-profile.self.cycles-pp.__do_fault
      1.04            -0.0        1.00        perf-profile.self.cycles-pp.find_lock_entry
      0.30            -0.0        0.28 +-  3%  perf-profile.self.cycles-pp.fault_dirty_shared_page
      0.70            -0.0        0.67        perf-profile.self.cycles-pp.prepare_ftrace_return
      0.44            -0.0        0.42        perf-profile.self.cycles-pp.do_page_fault
      0.16            -0.0        0.14        perf-profile.self.cycles-pp.rcu_all_qs
      0.78            -0.0        0.77        perf-profile.self.cycles-pp.shmem_getpage_gfp
      0.20            -0.0        0.19        perf-profile.self.cycles-pp._cond_resched
      0.50            +0.0        0.51        perf-profile.self.cycles-pp.set_page_dirty
      0.93            +0.0        0.95        perf-profile.self.cycles-pp.vmacache_find
      0.36 +-  2%      +0.0        0.38        perf-profile.self.cycles-pp.__might_sleep
      0.47            +0.0        0.50        perf-profile.self.cycles-pp.__mod_node_page_state
      0.17            +0.0        0.19 +-  2%  perf-profile.self.cycles-pp.__unlock_page_memcg
      2.34            +0.0        2.38        perf-profile.self.cycles-pp.unmap_page_range
      0.78 +-  3%      +0.1        0.85 +-  2%  perf-profile.self.cycles-pp.lock_page_memcg
      2.17            +0.1        2.24        perf-profile.self.cycles-pp.__do_page_fault
      0.00            +0.2        0.16 +-  3%  perf-profile.self.cycles-pp.__vm_normal_page
      1.00            +0.2        1.24        perf-profile.self.cycles-pp.___might_sleep
      0.00            +0.7        0.70        perf-profile.self.cycles-pp.pte_map_lock
      0.00            +1.4        1.42 +-  2%  perf-profile.self.cycles-pp.handle_pte_fault

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/never/context_switch1/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
           :3           33%           1:3     dmesg.WARNING:at#for_ip_interrupt_entry/0x
          2:3          -67%            :3     kmsg.pstore:crypto_comp_decompress_failed,ret=
          2:3          -67%            :3     kmsg.pstore:decompression_failed
         %stddev     %change         %stddev
             \          |                \  
    224431            -1.3%     221567        will-it-scale.per_process_ops
    237006            -2.2%     231907        will-it-scale.per_thread_ops
 1.601e+09 +- 29%     -46.9%  8.501e+08 +- 12%  will-it-scale.time.involuntary_context_switches
      5429            -1.6%       5344        will-it-scale.time.user_time
  88596221            -1.7%   87067269        will-it-scale.workload
      6863 +-  6%      -9.7%       6200        boot-time.idle
    144908 +- 40%     -66.8%      48173 +- 93%  meminfo.CmaFree
      0.00 +- 70%      +0.0        0.00        mpstat.cpu.iowait%
    448336 +- 14%     -34.8%     292125 +-  3%  turbostat.C1
      7684 +-  6%      -9.5%       6957        uptime.idle
 1.601e+09 +- 29%     -46.9%  8.501e+08 +- 12%  time.involuntary_context_switches
      5429            -1.6%       5344        time.user_time
  44013162            -1.7%   43243125        vmstat.system.cs
    207684            -1.1%     205485        vmstat.system.in
   2217033 +- 15%     -15.8%    1866876 +-  2%  cpuidle.C1.time
    451218 +- 14%     -34.7%     294841 +-  2%  cpuidle.C1.usage
     24839 +- 10%     -19.9%      19896        cpuidle.POLL.time
      7656 +- 11%     -38.9%       4676 +-  8%  cpuidle.POLL.usage
      5.48 +- 49%     -67.3%       1.79 +-100%  irq_exception_noise.__do_page_fault.95th
      9.46 +- 21%     -58.2%       3.95 +- 64%  irq_exception_noise.__do_page_fault.99th
     35.67 +-  8%   +1394.4%     533.00 +- 96%  irq_exception_noise.irq_nr
     52109 +-  3%     -16.0%      43784 +-  4%  irq_exception_noise.softirq_time
     36226 +- 40%     -66.7%      12048 +- 93%  proc-vmstat.nr_free_cma
     25916            -1.0%      25659        proc-vmstat.nr_slab_reclaimable
     16279 +-  8%   +2646.1%     447053 +- 82%  proc-vmstat.pgalloc_movable
   2231117           -18.4%    1820828 +- 20%  proc-vmstat.pgalloc_normal
   1109316 +- 46%     -86.9%     145207 +-109%  numa-numastat.node1.local_node
   1114700 +- 45%     -84.5%     172877 +- 85%  numa-numastat.node1.numa_hit
      5523 +-140%    +402.8%      27768 +- 39%  numa-numastat.node1.other_node
     29013 +- 29%   +3048.1%     913379 +- 73%  numa-numastat.node3.local_node
     65032 +- 13%   +1335.1%     933270 +- 70%  numa-numastat.node3.numa_hit
     36018           -44.8%      19897 +- 75%  numa-numastat.node3.other_node
     12.79 +- 21%   +7739.1%       1002 +-136%  sched_debug.cpu.cpu_load[1].max
      1.82 +- 10%   +3901.1%      72.92 +-135%  sched_debug.cpu.cpu_load[1].stddev
      1.71 +-  4%   +5055.8%      88.08 +-137%  sched_debug.cpu.cpu_load[2].stddev
     12.33 +- 23%   +9061.9%       1129 +-139%  sched_debug.cpu.cpu_load[3].max
      1.78 +- 10%   +4514.8%      82.18 +-137%  sched_debug.cpu.cpu_load[3].stddev
      4692 +- 72%    +154.5%      11945 +- 29%  sched_debug.cpu.max_idle_balance_cost.stddev
     23979            -8.3%      21983        slabinfo.kmalloc-96.active_objs
      1358 +-  6%     -17.9%       1114 +-  3%  slabinfo.nsproxy.active_objs
      1358 +-  6%     -17.9%       1114 +-  3%  slabinfo.nsproxy.num_objs
     15229           +12.4%      17119        slabinfo.pde_opener.active_objs
     15229           +12.4%      17119        slabinfo.pde_opener.num_objs
     59541 +-  8%     -10.1%      53537 +-  8%  slabinfo.vm_area_struct.active_objs
     59612 +-  8%     -10.1%      53604 +-  8%  slabinfo.vm_area_struct.num_objs
 4.163e+13            -1.4%  4.105e+13        perf-stat.branch-instructions
 6.537e+11            -1.2%  6.459e+11        perf-stat.branch-misses
 2.667e+10            -1.7%  2.621e+10        perf-stat.context-switches
      1.21            +1.3%       1.22        perf-stat.cpi
    150508            -9.8%     135825 +-  3%  perf-stat.cpu-migrations
      5.75 +- 33%      +5.4       11.11 +- 26%  perf-stat.iTLB-load-miss-rate%
 3.619e+09 +- 36%    +100.9%  7.272e+09 +- 30%  perf-stat.iTLB-load-misses
 2.089e+14            -1.3%  2.062e+14        perf-stat.instructions
     64607 +- 29%     -50.5%      31964 +- 37%  perf-stat.instructions-per-iTLB-miss
      0.83            -1.3%       0.82        perf-stat.ipc
      3972 +-  4%     -14.7%       3388 +-  8%  numa-meminfo.node0.PageTables
    207919 +- 25%     -57.2%      88989 +- 74%  numa-meminfo.node1.Active
    207715 +- 26%     -57.3%      88785 +- 74%  numa-meminfo.node1.Active(anon)
    356529           -34.3%     234069 +-  2%  numa-meminfo.node1.FilePages
    789129 +-  5%     -19.8%     633161 +- 12%  numa-meminfo.node1.MemUsed
     34777 +-  8%     -48.2%      18010 +- 30%  numa-meminfo.node1.SReclaimable
     69641 +-  4%     -20.7%      55250 +- 12%  numa-meminfo.node1.SUnreclaim
    125526 +-  4%     -96.3%       4602 +- 41%  numa-meminfo.node1.Shmem
    104419           -29.8%      73261 +- 16%  numa-meminfo.node1.Slab
    103661 +- 17%     -72.0%      29029 +- 99%  numa-meminfo.node2.Active
    103661 +- 17%     -72.2%      28829 +-101%  numa-meminfo.node2.Active(anon)
    103564 +- 18%     -72.0%      29007 +-100%  numa-meminfo.node2.AnonPages
    671654 +-  7%     -14.6%     573598 +-  4%  numa-meminfo.node2.MemUsed
     44206 +-127%    +301.4%     177465 +- 42%  numa-meminfo.node3.Active
     44206 +-127%    +301.0%     177263 +- 42%  numa-meminfo.node3.Active(anon)
      8738           +12.2%       9805 +-  8%  numa-meminfo.node3.KernelStack
    603605 +-  9%     +27.8%     771554 +- 14%  numa-meminfo.node3.MemUsed
     14438 +-  6%    +122.9%      32181 +- 42%  numa-meminfo.node3.SReclaimable
      2786 +-137%   +3302.0%      94792 +- 71%  numa-meminfo.node3.Shmem
     71461 +-  7%     +45.2%     103771 +- 29%  numa-meminfo.node3.Slab
    247197 +-  4%      -7.8%     227843        numa-meminfo.node3.Unevictable
    991.67 +-  4%     -14.7%     846.00 +-  8%  numa-vmstat.node0.nr_page_table_pages
     51926 +- 26%     -57.3%      22196 +- 74%  numa-vmstat.node1.nr_active_anon
     89137           -34.4%      58516 +-  2%  numa-vmstat.node1.nr_file_pages
      1679 +-  5%     -10.8%       1498 +-  4%  numa-vmstat.node1.nr_mapped
     31386 +-  4%     -96.3%       1150 +- 41%  numa-vmstat.node1.nr_shmem
      8694 +-  8%     -48.2%       4502 +- 30%  numa-vmstat.node1.nr_slab_reclaimable
     17410 +-  4%     -20.7%      13812 +- 12%  numa-vmstat.node1.nr_slab_unreclaimable
     51926 +- 26%     -57.3%      22196 +- 74%  numa-vmstat.node1.nr_zone_active_anon
   1037174 +- 24%     -57.0%     446205 +- 35%  numa-vmstat.node1.numa_hit
    961611 +- 26%     -65.8%     328687 +- 50%  numa-vmstat.node1.numa_local
     75563 +- 44%     +55.5%     117517 +-  9%  numa-vmstat.node1.numa_other
     25914 +- 17%     -72.2%       7206 +-101%  numa-vmstat.node2.nr_active_anon
     25891 +- 18%     -72.0%       7251 +-100%  numa-vmstat.node2.nr_anon_pages
     25914 +- 17%     -72.2%       7206 +-101%  numa-vmstat.node2.nr_zone_active_anon
     11051 +-127%    +301.0%      44309 +- 42%  numa-vmstat.node3.nr_active_anon
     36227 +- 40%     -66.7%      12049 +- 93%  numa-vmstat.node3.nr_free_cma
      0.33 +-141%  +25000.0%      83.67 +- 81%  numa-vmstat.node3.nr_inactive_file
      8739           +12.2%       9806 +-  8%  numa-vmstat.node3.nr_kernel_stack
    696.67 +-137%   +3299.7%      23684 +- 71%  numa-vmstat.node3.nr_shmem
      3609 +-  6%    +122.9%       8044 +- 42%  numa-vmstat.node3.nr_slab_reclaimable
     61799 +-  4%      -7.8%      56960        numa-vmstat.node3.nr_unevictable
     11053 +-127%    +301.4%      44361 +- 42%  numa-vmstat.node3.nr_zone_active_anon
      0.33 +-141%  +25000.0%      83.67 +- 81%  numa-vmstat.node3.nr_zone_inactive_file
     61799 +-  4%      -7.8%      56960        numa-vmstat.node3.nr_zone_unevictable
    217951 +-  8%    +280.8%     829976 +- 65%  numa-vmstat.node3.numa_hit
     91303 +- 19%    +689.3%     720647 +- 77%  numa-vmstat.node3.numa_local
    126648           -13.7%     109329 +- 13%  numa-vmstat.node3.numa_other
      8.54            -0.1        8.40        perf-profile.calltrace.cycles-pp.dequeue_task_fair.__schedule.schedule.pipe_wait.pipe_read
      5.04            -0.1        4.94        perf-profile.calltrace.cycles-pp.__switch_to.read
      3.43            -0.1        3.35        perf-profile.calltrace.cycles-pp.syscall_return_via_sysret.write
      2.77            -0.1        2.72        perf-profile.calltrace.cycles-pp.reweight_entity.enqueue_task_fair.ttwu_do_activate.try_to_wake_up.autoremove_wake_function
      1.99            -0.0        1.94        perf-profile.calltrace.cycles-pp.copy_page_to_iter.pipe_read.__vfs_read.vfs_read.ksys_read
      0.60 +-  2%      -0.0        0.57 +-  2%  perf-profile.calltrace.cycles-pp.find_next_bit.cpumask_next_wrap.select_idle_sibling.select_task_rq_fair.try_to_wake_up
      0.81            -0.0        0.78        perf-profile.calltrace.cycles-pp.___perf_sw_event.__schedule.schedule.pipe_wait.pipe_read
      0.78            +0.0        0.80        perf-profile.calltrace.cycles-pp.__fdget_pos.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe.write
      0.73            +0.0        0.75        perf-profile.calltrace.cycles-pp.__fget_light.__fdget_pos.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.92            +0.0        0.95        perf-profile.calltrace.cycles-pp.check_preempt_wakeup.check_preempt_curr.ttwu_do_wakeup.try_to_wake_up.autoremove_wake_function
      2.11            +0.0        2.15        perf-profile.calltrace.cycles-pp.security_file_permission.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      7.00            -0.1        6.86        perf-profile.children.cycles-pp.syscall_return_via_sysret
      5.26            -0.1        5.14        perf-profile.children.cycles-pp.__switch_to
      5.65            -0.1        5.56        perf-profile.children.cycles-pp.reweight_entity
      2.17            -0.1        2.12        perf-profile.children.cycles-pp.copy_page_to_iter
      2.94            -0.0        2.90        perf-profile.children.cycles-pp.update_cfs_group
      3.11            -0.0        3.07        perf-profile.children.cycles-pp.pick_next_task_fair
      2.59            -0.0        2.55        perf-profile.children.cycles-pp.load_new_mm_cr3
      1.92            -0.0        1.88        perf-profile.children.cycles-pp._raw_spin_lock_irqsave
      1.11            -0.0        1.08 +-  2%  perf-profile.children.cycles-pp.find_next_bit
      0.59            -0.0        0.56        perf-profile.children.cycles-pp.finish_task_switch
      0.14 +- 15%      -0.0        0.11 +- 16%  perf-profile.children.cycles-pp.write@plt
      1.21            -0.0        1.18        perf-profile.children.cycles-pp.set_next_entity
      0.85            -0.0        0.82        perf-profile.children.cycles-pp.___perf_sw_event
      0.13 +-  3%      -0.0        0.11 +-  4%  perf-profile.children.cycles-pp.timespec_trunc
      0.47 +-  2%      -0.0        0.45        perf-profile.children.cycles-pp.anon_pipe_buf_release
      0.38 +-  2%      -0.0        0.36        perf-profile.children.cycles-pp.file_update_time
      0.74            -0.0        0.73        perf-profile.children.cycles-pp.copyout
      0.41 +-  2%      -0.0        0.39        perf-profile.children.cycles-pp.copy_user_enhanced_fast_string
      0.32            -0.0        0.30        perf-profile.children.cycles-pp.__x64_sys_read
      0.14            -0.0        0.12 +-  3%  perf-profile.children.cycles-pp.current_kernel_time64
      0.91            +0.0        0.92        perf-profile.children.cycles-pp.touch_atime
      0.40            +0.0        0.41        perf-profile.children.cycles-pp._cond_resched
      0.18 +-  2%      +0.0        0.20        perf-profile.children.cycles-pp.activate_task
      0.05            +0.0        0.07 +-  6%  perf-profile.children.cycles-pp.default_wake_function
      0.24            +0.0        0.27 +-  3%  perf-profile.children.cycles-pp.rcu_all_qs
      0.60 +-  2%      +0.0        0.64 +-  2%  perf-profile.children.cycles-pp.update_min_vruntime
      0.42 +-  4%      +0.0        0.46 +-  4%  perf-profile.children.cycles-pp.probe_sched_switch
      1.33            +0.0        1.38        perf-profile.children.cycles-pp.__fget_light
      0.53 +-  2%      +0.1        0.58        perf-profile.children.cycles-pp.entry_SYSCALL_64_stage2
      0.31            +0.1        0.36 +-  2%  perf-profile.children.cycles-pp.generic_pipe_buf_confirm
      4.35            +0.1        4.41        perf-profile.children.cycles-pp.switch_mm_irqs_off
      2.52            +0.1        2.58        perf-profile.children.cycles-pp.selinux_file_permission
      0.00            +0.1        0.07 +- 11%  perf-profile.children.cycles-pp.hrtick_update
      7.00            -0.1        6.86        perf-profile.self.cycles-pp.syscall_return_via_sysret
      5.26            -0.1        5.14        perf-profile.self.cycles-pp.__switch_to
      0.29            -0.1        0.19 +-  2%  perf-profile.self.cycles-pp.ksys_read
      1.49            -0.1        1.43        perf-profile.self.cycles-pp.dequeue_task_fair
      2.41            -0.1        2.35        perf-profile.self.cycles-pp.__schedule
      1.46            -0.0        1.41        perf-profile.self.cycles-pp.select_task_rq_fair
      2.94            -0.0        2.90        perf-profile.self.cycles-pp.update_cfs_group
      0.44            -0.0        0.40        perf-profile.self.cycles-pp.dequeue_entity
      0.48            -0.0        0.44        perf-profile.self.cycles-pp.finish_task_switch
      2.59            -0.0        2.55        perf-profile.self.cycles-pp.load_new_mm_cr3
      1.11            -0.0        1.08 +-  2%  perf-profile.self.cycles-pp.find_next_bit
      1.91            -0.0        1.88        perf-profile.self.cycles-pp._raw_spin_lock_irqsave
      0.78            -0.0        0.75        perf-profile.self.cycles-pp.___perf_sw_event
      0.14 +- 15%      -0.0        0.11 +- 16%  perf-profile.self.cycles-pp.write@plt
      0.37            -0.0        0.35 +-  2%  perf-profile.self.cycles-pp.__wake_up_common_lock
      0.20 +-  2%      -0.0        0.17 +-  2%  perf-profile.self.cycles-pp.__fdget_pos
      0.47 +-  2%      -0.0        0.44        perf-profile.self.cycles-pp.anon_pipe_buf_release
      0.87            -0.0        0.85        perf-profile.self.cycles-pp.copy_user_generic_unrolled
      0.13 +-  3%      -0.0        0.11 +-  4%  perf-profile.self.cycles-pp.timespec_trunc
      0.41 +-  2%      -0.0        0.39        perf-profile.self.cycles-pp.copy_user_enhanced_fast_string
      0.38            -0.0        0.36        perf-profile.self.cycles-pp.__wake_up_common
      0.32            -0.0        0.30        perf-profile.self.cycles-pp.__x64_sys_read
      0.14 +-  3%      -0.0        0.12 +-  3%  perf-profile.self.cycles-pp.current_kernel_time64
      0.30            -0.0        0.28        perf-profile.self.cycles-pp.set_next_entity
      0.28 +-  3%      +0.0        0.30        perf-profile.self.cycles-pp._cond_resched
      0.18 +-  2%      +0.0        0.20        perf-profile.self.cycles-pp.activate_task
      0.17 +-  2%      +0.0        0.19        perf-profile.self.cycles-pp.__might_fault
      0.05            +0.0        0.07 +-  6%  perf-profile.self.cycles-pp.default_wake_function
      0.17 +-  2%      +0.0        0.20        perf-profile.self.cycles-pp.ttwu_do_activate
      0.66            +0.0        0.69        perf-profile.self.cycles-pp.write
      0.24            +0.0        0.27 +-  3%  perf-profile.self.cycles-pp.rcu_all_qs
      0.67            +0.0        0.70        perf-profile.self.cycles-pp.entry_SYSCALL_64_after_hwframe
      0.60 +-  2%      +0.0        0.64 +-  2%  perf-profile.self.cycles-pp.update_min_vruntime
      0.42 +-  4%      +0.0        0.46 +-  4%  perf-profile.self.cycles-pp.probe_sched_switch
      1.33            +0.0        1.37        perf-profile.self.cycles-pp.__fget_light
      1.61            +0.0        1.66        perf-profile.self.cycles-pp.pipe_read
      0.53 +-  2%      +0.1        0.58        perf-profile.self.cycles-pp.entry_SYSCALL_64_stage2
      0.31            +0.1        0.36 +-  2%  perf-profile.self.cycles-pp.generic_pipe_buf_confirm
      1.04            +0.1        1.11        perf-profile.self.cycles-pp.pipe_write
      0.00            +0.1        0.07 +- 11%  perf-profile.self.cycles-pp.hrtick_update
      2.00            +0.1        2.08        perf-profile.self.cycles-pp.switch_mm_irqs_off

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/never/page_fault3/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
          1:3          -33%            :3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=file_update_time/0x
           :3           33%           1:3     stderr.mount.nfs:Connection_timed_out
         34:3         -401%          22:3     perf-profile.calltrace.cycles-pp.error_entry.testcase
         17:3         -207%          11:3     perf-profile.calltrace.cycles-pp.sync_regs.error_entry.testcase
         34:3         -404%          22:3     perf-profile.children.cycles-pp.error_entry
          0:3           -2%           0:3     perf-profile.children.cycles-pp.error_exit
         16:3         -196%          11:3     perf-profile.self.cycles-pp.error_entry
          0:3           -2%           0:3     perf-profile.self.cycles-pp.error_exit
         %stddev     %change         %stddev
             \          |                \  
    467454            -1.8%     459251        will-it-scale.per_process_ops
     10856 +-  4%     -23.1%       8344 +-  7%  will-it-scale.per_thread_ops
    118134 +-  2%     +11.7%     131943        will-it-scale.time.involuntary_context_switches
 6.277e+08 +-  4%     -23.1%  4.827e+08 +-  7%  will-it-scale.time.minor_page_faults
      7406            +5.8%       7839        will-it-scale.time.percent_of_cpu_this_job_got
     44526            +5.8%      47106        will-it-scale.time.system_time
   7351468 +-  5%     -18.3%    6009014 +-  7%  will-it-scale.time.voluntary_context_switches
  91835846            -2.2%   89778599        will-it-scale.workload
   2534640            +4.3%    2643005 +-  2%  interrupts.CAL:Function_call_interrupts
      2819 +-  5%     +22.9%       3464 +- 18%  kthread_noise.total_time
     30273 +-  4%     -12.7%      26415 +-  5%  vmstat.system.cs
      1.52 +-  2%     +15.2%       1.75 +-  2%  irq_exception_noise.__do_page_fault.99th
    296.67 +- 12%     -36.7%     187.67 +- 12%  irq_exception_noise.softirq_time
    230900 +-  3%     +30.3%     300925 +-  3%  meminfo.Inactive
    230184 +-  3%     +30.4%     300180 +-  3%  meminfo.Inactive(anon)
     11.62 +-  3%      -2.2        9.40 +-  5%  mpstat.cpu.idle%
      0.00 +- 14%      -0.0        0.00 +-  4%  mpstat.cpu.iowait%
   7992174           -11.1%    7101976 +-  3%  softirqs.RCU
   4973624 +-  2%     -12.9%    4333370 +-  2%  softirqs.SCHED
    118134 +-  2%     +11.7%     131943        time.involuntary_context_switches
 6.277e+08 +-  4%     -23.1%  4.827e+08 +-  7%  time.minor_page_faults
      7406            +5.8%       7839        time.percent_of_cpu_this_job_got
     44526            +5.8%      47106        time.system_time
   7351468 +-  5%     -18.3%    6009014 +-  7%  time.voluntary_context_switches
 2.702e+09 +-  5%     -16.7%  2.251e+09 +-  7%  cpuidle.C1E.time
   6834329 +-  5%     -15.8%    5756243 +-  7%  cpuidle.C1E.usage
 1.046e+10 +-  3%     -19.8%  8.389e+09 +-  4%  cpuidle.C6.time
  13961845 +-  3%     -19.3%   11265555 +-  4%  cpuidle.C6.usage
   1309307 +-  7%     -14.8%    1116168 +-  8%  cpuidle.POLL.time
     19774 +-  6%     -13.7%      17063 +-  7%  cpuidle.POLL.usage
      2523 +-  4%     -11.1%       2243 +-  4%  slabinfo.biovec-64.active_objs
      2523 +-  4%     -11.1%       2243 +-  4%  slabinfo.biovec-64.num_objs
      2610 +-  8%     -33.7%       1731 +- 22%  slabinfo.dmaengine-unmap-16.active_objs
      2610 +-  8%     -33.7%       1731 +- 22%  slabinfo.dmaengine-unmap-16.num_objs
      5118 +- 17%     -22.6%       3962 +-  9%  slabinfo.eventpoll_pwq.active_objs
      5118 +- 17%     -22.6%       3962 +-  9%  slabinfo.eventpoll_pwq.num_objs
      4583 +-  3%     -14.0%       3941 +-  4%  slabinfo.sock_inode_cache.active_objs
      4583 +-  3%     -14.0%       3941 +-  4%  slabinfo.sock_inode_cache.num_objs
      1933            +2.6%       1984        turbostat.Avg_MHz
   6832021 +-  5%     -15.8%    5754156 +-  7%  turbostat.C1E
      2.32 +-  5%      -0.4        1.94 +-  7%  turbostat.C1E%
  13954211 +-  3%     -19.3%   11259436 +-  4%  turbostat.C6
      8.97 +-  3%      -1.8        7.20 +-  4%  turbostat.C6%
      6.18 +-  4%     -17.1%       5.13 +-  5%  turbostat.CPU%c1
      5.12 +-  3%     -21.7%       4.01 +-  4%  turbostat.CPU%c6
      1.76 +-  2%     -34.7%       1.15 +-  2%  turbostat.Pkg%pc2
     57314 +-  4%     +30.4%      74717 +-  4%  proc-vmstat.nr_inactive_anon
     57319 +-  4%     +30.4%      74719 +-  4%  proc-vmstat.nr_zone_inactive_anon
     24415 +- 19%     -62.2%       9236 +-  7%  proc-vmstat.numa_hint_faults
  69661453            -1.8%   68405712        proc-vmstat.numa_hit
  69553390            -1.8%   68297790        proc-vmstat.numa_local
      8792 +- 29%     -92.6%     654.33 +- 23%  proc-vmstat.numa_pages_migrated
     40251 +- 32%     -76.5%       9474 +-  3%  proc-vmstat.numa_pte_updates
  69522532            -1.6%   68383074        proc-vmstat.pgalloc_normal
 2.762e+10            -2.2%  2.701e+10        proc-vmstat.pgfault
  68825100            -1.5%   67772256        proc-vmstat.pgfree
      8792 +- 29%     -92.6%     654.33 +- 23%  proc-vmstat.pgmigrate_success
     57992 +-  6%     +56.2%      90591 +-  3%  numa-meminfo.node0.Inactive
     57916 +-  6%     +56.3%      90513 +-  3%  numa-meminfo.node0.Inactive(anon)
     37285 +- 12%     +36.0%      50709 +-  5%  numa-meminfo.node0.SReclaimable
    110971 +-  8%     +22.7%     136209 +-  8%  numa-meminfo.node0.Slab
     23601 +- 55%    +559.5%     155651 +- 36%  numa-meminfo.node1.AnonPages
     62484 +- 12%     +17.5%      73417 +-  3%  numa-meminfo.node1.Inactive
     62323 +- 12%     +17.2%      73023 +-  4%  numa-meminfo.node1.Inactive(anon)
    109714 +- 63%     -85.6%      15832 +- 96%  numa-meminfo.node2.AnonPages
     52236 +- 13%     +22.7%      64074 +-  3%  numa-meminfo.node2.Inactive
     51922 +- 12%     +23.2%      63963 +-  3%  numa-meminfo.node2.Inactive(anon)
     60241 +- 11%     +21.9%      73442 +-  8%  numa-meminfo.node3.Inactive
     60077 +- 12%     +22.0%      73279 +-  8%  numa-meminfo.node3.Inactive(anon)
     14093 +-  6%     +55.9%      21977 +-  3%  numa-vmstat.node0.nr_inactive_anon
      9321 +- 12%     +36.0%      12675 +-  5%  numa-vmstat.node0.nr_slab_reclaimable
     14090 +-  6%     +56.0%      21977 +-  3%  numa-vmstat.node0.nr_zone_inactive_anon
      5900 +- 55%    +559.4%      38909 +- 36%  numa-vmstat.node1.nr_anon_pages
     15413 +- 12%     +14.8%      17688 +-  4%  numa-vmstat.node1.nr_inactive_anon
     15413 +- 12%     +14.8%      17688 +-  4%  numa-vmstat.node1.nr_zone_inactive_anon
     27430 +- 63%     -85.6%       3960 +- 96%  numa-vmstat.node2.nr_anon_pages
     12928 +- 12%     +20.0%      15508 +-  3%  numa-vmstat.node2.nr_inactive_anon
     12927 +- 12%     +20.0%      15507 +-  3%  numa-vmstat.node2.nr_zone_inactive_anon
      6229 +- 10%    +117.5%      13547 +- 44%  numa-vmstat.node3
     14669 +- 11%     +19.6%      17537 +-  7%  numa-vmstat.node3.nr_inactive_anon
     14674 +- 11%     +19.5%      17541 +-  7%  numa-vmstat.node3.nr_zone_inactive_anon
     24617 +-141%    -100.0%       0.00        latency_stats.avg.io_schedule.nfs_lock_and_join_requests.nfs_updatepage.nfs_write_end.generic_perform_write.nfs_file_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      5049 +-105%     -99.4%      28.33 +- 82%  latency_stats.avg.call_rwsem_down_write_failed.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
    152457 +- 27%    +233.6%     508656 +- 92%  latency_stats.avg.max
      0.00       +3.9e+107%     390767 +-141%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_openat
     24617 +-141%    -100.0%       0.00        latency_stats.max.io_schedule.nfs_lock_and_join_requests.nfs_updatepage.nfs_write_end.generic_perform_write.nfs_file_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4240 +-141%    -100.0%       0.00        latency_stats.max.call_rwsem_down_write_failed.do_unlinkat.do_syscall_64.entry_SYSCALL_64_after_hwframe
      8565 +- 70%     -99.1%      80.33 +-115%  latency_stats.max.call_rwsem_down_write_failed.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
    204835 +-  6%    +457.6%    1142244 +-114%  latency_stats.max.max
      0.00       +5.1e+105%       5057 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_access.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_openat.do_filp_open
      0.00         +1e+108%     995083 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_openat
     13175 +-  4%    -100.0%       0.00        latency_stats.sum.io_schedule.__lock_page_or_retry.filemap_fault.__do_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault.page_fault
     24617 +-141%    -100.0%       0.00        latency_stats.sum.io_schedule.nfs_lock_and_join_requests.nfs_updatepage.nfs_write_end.generic_perform_write.nfs_file_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4260 +-141%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_write_failed.do_unlinkat.do_syscall_64.entry_SYSCALL_64_after_hwframe
      8640 +- 70%     -97.5%     216.33 +-108%  latency_stats.sum.call_rwsem_down_write_failed.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      6673 +- 89%     -92.8%     477.67 +- 74%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
      0.00       +4.2e+105%       4228 +-130%  latency_stats.sum.io_schedule.__lock_page_killable.__lock_page_or_retry.filemap_fault.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault.page_fault
      0.00       +7.5e+105%       7450 +- 98%  latency_stats.sum.io_schedule.__lock_page_or_retry.filemap_fault.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault.page_fault
      0.00       +1.3e+106%      13050 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_access.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_openat.do_filp_open
      0.00       +1.5e+110%  1.508e+08 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_openat
      0.97            -0.0        0.94        perf-stat.branch-miss-rate%
 1.329e+11            -2.6%  1.294e+11        perf-stat.branch-misses
 2.254e+11            -1.9%   2.21e+11        perf-stat.cache-references
  18308779 +-  4%     -12.8%   15969618 +-  5%  perf-stat.context-switches
      3.20            +1.8%       3.26        perf-stat.cpi
 2.233e+14            +2.7%  2.293e+14        perf-stat.cpu-cycles
      4.01            -0.2        3.83        perf-stat.dTLB-store-miss-rate%
  4.51e+11            -2.2%   4.41e+11        perf-stat.dTLB-store-misses
  1.08e+13            +2.6%  1.109e+13        perf-stat.dTLB-stores
 3.158e+10 +-  5%     +16.8%  3.689e+10 +-  2%  perf-stat.iTLB-load-misses
      2214 +-  5%     -13.8%       1907 +-  2%  perf-stat.instructions-per-iTLB-miss
      0.31            -1.8%       0.31        perf-stat.ipc
 2.762e+10            -2.2%  2.701e+10        perf-stat.minor-faults
 1.535e+10           -11.2%  1.362e+10        perf-stat.node-loads
      9.75            +1.1       10.89        perf-stat.node-store-miss-rate%
 3.012e+09           +10.6%  3.332e+09 +-  2%  perf-stat.node-store-misses
 2.787e+10            -2.2%  2.725e+10        perf-stat.node-stores
 2.762e+10            -2.2%  2.701e+10        perf-stat.page-faults
    759458            +3.2%     783404        perf-stat.path-length
    246.39 +- 15%     -20.4%     196.12 +-  6%  sched_debug.cfs_rq:/.load_avg.max
      0.21 +-  3%      +9.0%       0.23 +-  4%  sched_debug.cfs_rq:/.nr_running.stddev
     16.64 +- 27%     +61.0%      26.79 +- 17%  sched_debug.cfs_rq:/.nr_spread_over.max
     75.15           -14.4%      64.30 +-  4%  sched_debug.cfs_rq:/.util_avg.stddev
    178.80 +-  3%     +25.4%     224.12 +-  7%  sched_debug.cfs_rq:/.util_est_enqueued.avg
      1075 +-  5%     -12.3%     943.36 +-  2%  sched_debug.cfs_rq:/.util_est_enqueued.max
   2093630 +- 27%     -36.1%    1337941 +- 16%  sched_debug.cpu.avg_idle.max
    297057 +- 11%     +37.8%     409294 +- 14%  sched_debug.cpu.avg_idle.min
    293240 +- 55%     -62.3%     110571 +- 13%  sched_debug.cpu.avg_idle.stddev
    770075 +-  9%     -19.3%     621136 +- 12%  sched_debug.cpu.max_idle_balance_cost.max
     48919 +- 46%     -66.9%      16190 +- 81%  sched_debug.cpu.max_idle_balance_cost.stddev
     21716 +-  5%     -16.8%      18061 +-  7%  sched_debug.cpu.nr_switches.min
     21519 +-  5%     -17.7%      17700 +-  7%  sched_debug.cpu.sched_count.min
     10586 +-  5%     -18.1%       8669 +-  7%  sched_debug.cpu.sched_goidle.avg
     14183 +-  3%     -17.6%      11693 +-  5%  sched_debug.cpu.sched_goidle.max
     10322 +-  5%     -18.6%       8407 +-  7%  sched_debug.cpu.sched_goidle.min
    400.99 +-  8%     -13.0%     348.75 +-  3%  sched_debug.cpu.sched_goidle.stddev
      5459 +-  8%     +10.0%       6006 +-  3%  sched_debug.cpu.ttwu_local.avg
      8.47 +- 42%    +345.8%      37.73 +- 77%  sched_debug.rt_rq:/.rt_time.max
      0.61 +- 42%    +343.0%       2.72 +- 77%  sched_debug.rt_rq:/.rt_time.stddev
     91.98           -30.9       61.11 +- 70%  perf-profile.calltrace.cycles-pp.testcase
      9.05            -9.1        0.00        perf-profile.calltrace.cycles-pp.__do_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      8.91            -8.9        0.00        perf-profile.calltrace.cycles-pp.shmem_fault.__do_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      8.06            -8.1        0.00        perf-profile.calltrace.cycles-pp.shmem_getpage_gfp.shmem_fault.__do_fault.__handle_mm_fault.handle_mm_fault
      7.59            -7.6        0.00        perf-profile.calltrace.cycles-pp.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault.__handle_mm_fault
      7.44            -7.4        0.00        perf-profile.calltrace.cycles-pp.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      7.28            -7.3        0.00        perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      5.31            -5.3        0.00        perf-profile.calltrace.cycles-pp.page_add_file_rmap.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault
      8.08            -2.8        5.30 +- 70%  perf-profile.calltrace.cycles-pp.native_irq_return_iret.testcase
      5.95            -2.1        3.83 +- 70%  perf-profile.calltrace.cycles-pp.find_get_entry.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault
      5.95            -2.0        3.93 +- 70%  perf-profile.calltrace.cycles-pp.swapgs_restore_regs_and_return_to_usermode.testcase
      3.10            -1.1        2.01 +- 70%  perf-profile.calltrace.cycles-pp.__perf_sw_event.__do_page_fault.do_page_fault.page_fault.testcase
      2.36            -0.8        1.55 +- 70%  perf-profile.calltrace.cycles-pp.___perf_sw_event.__perf_sw_event.__do_page_fault.do_page_fault.page_fault
      1.08            -0.4        0.70 +- 70%  perf-profile.calltrace.cycles-pp.do_page_fault.testcase
      0.82            -0.3        0.54 +- 70%  perf-profile.calltrace.cycles-pp.trace_graph_entry.do_page_fault.testcase
      0.77            -0.3        0.50 +- 70%  perf-profile.calltrace.cycles-pp.ftrace_graph_caller.__do_page_fault.do_page_fault.page_fault.testcase
      0.59            -0.2        0.37 +- 70%  perf-profile.calltrace.cycles-pp.down_read_trylock.__do_page_fault.do_page_fault.page_fault.testcase
     91.98           -30.9       61.11 +- 70%  perf-profile.children.cycles-pp.testcase
      9.14            -3.2        5.99 +- 70%  perf-profile.children.cycles-pp.__do_fault
      8.20            -2.8        5.40 +- 70%  perf-profile.children.cycles-pp.shmem_getpage_gfp
      8.08            -2.8        5.31 +- 70%  perf-profile.children.cycles-pp.native_irq_return_iret
      6.08            -2.2        3.92 +- 70%  perf-profile.children.cycles-pp.find_get_entry
      6.08            -2.1        3.96 +- 70%  perf-profile.children.cycles-pp.sync_regs
      5.95            -2.0        3.93 +- 70%  perf-profile.children.cycles-pp.swapgs_restore_regs_and_return_to_usermode
      4.12            -1.4        2.73 +- 70%  perf-profile.children.cycles-pp.ftrace_graph_caller
      3.65            -1.2        2.42 +- 70%  perf-profile.children.cycles-pp.prepare_ftrace_return
      3.18            -1.1        2.07 +- 70%  perf-profile.children.cycles-pp.__perf_sw_event
      2.34            -0.8        1.52 +- 70%  perf-profile.children.cycles-pp.fault_dirty_shared_page
      0.80            -0.3        0.50 +- 70%  perf-profile.children.cycles-pp._raw_spin_lock
      0.76            -0.3        0.50 +- 70%  perf-profile.children.cycles-pp.tlb_flush_mmu_free
      0.61            -0.2        0.39 +- 70%  perf-profile.children.cycles-pp.down_read_trylock
      0.48 +-  2%      -0.2        0.28 +- 70%  perf-profile.children.cycles-pp.pmd_devmap_trans_unstable
      0.26 +-  6%      -0.1        0.15 +- 71%  perf-profile.children.cycles-pp.ktime_get
      0.20 +-  2%      -0.1        0.12 +- 70%  perf-profile.children.cycles-pp.perf_exclude_event
      0.22 +-  2%      -0.1        0.13 +- 70%  perf-profile.children.cycles-pp._cond_resched
      0.17            -0.1        0.11 +- 70%  perf-profile.children.cycles-pp.page_rmapping
      0.13            -0.1        0.07 +- 70%  perf-profile.children.cycles-pp.rcu_all_qs
      0.07            -0.0        0.04 +- 70%  perf-profile.children.cycles-pp.ftrace_lookup_ip
     22.36            -7.8       14.59 +- 70%  perf-profile.self.cycles-pp.testcase
      8.08            -2.8        5.31 +- 70%  perf-profile.self.cycles-pp.native_irq_return_iret
      6.08            -2.1        3.96 +- 70%  perf-profile.self.cycles-pp.sync_regs
      5.81            -2.0        3.84 +- 70%  perf-profile.self.cycles-pp.swapgs_restore_regs_and_return_to_usermode
      3.27            -1.6        1.65 +- 70%  perf-profile.self.cycles-pp.__handle_mm_fault
      3.79            -1.4        2.36 +- 70%  perf-profile.self.cycles-pp.find_get_entry
      3.80            -1.3        2.53 +- 70%  perf-profile.self.cycles-pp.trace_graph_entry
      1.10            -0.5        0.57 +- 70%  perf-profile.self.cycles-pp.alloc_set_pte
      1.24            -0.4        0.81 +- 70%  perf-profile.self.cycles-pp.shmem_fault
      0.80            -0.3        0.50 +- 70%  perf-profile.self.cycles-pp._raw_spin_lock
      0.81            -0.3        0.51 +- 70%  perf-profile.self.cycles-pp.find_lock_entry
      0.80 +-  2%      -0.3        0.51 +- 70%  perf-profile.self.cycles-pp.__perf_sw_event
      0.61            -0.2        0.38 +- 70%  perf-profile.self.cycles-pp.down_read_trylock
      0.60            -0.2        0.39 +- 70%  perf-profile.self.cycles-pp.shmem_getpage_gfp
      0.48            -0.2        0.27 +- 70%  perf-profile.self.cycles-pp.pmd_devmap_trans_unstable
      0.47            -0.2        0.30 +- 70%  perf-profile.self.cycles-pp.file_update_time
      0.34            -0.1        0.22 +- 70%  perf-profile.self.cycles-pp.do_page_fault
      0.22 +-  4%      -0.1        0.11 +- 70%  perf-profile.self.cycles-pp.__do_fault
      0.25 +-  5%      -0.1        0.14 +- 71%  perf-profile.self.cycles-pp.ktime_get
      0.21 +-  2%      -0.1        0.12 +- 70%  perf-profile.self.cycles-pp.finish_fault
      0.23 +-  2%      -0.1        0.14 +- 70%  perf-profile.self.cycles-pp.fault_dirty_shared_page
      0.22 +-  2%      -0.1        0.14 +- 70%  perf-profile.self.cycles-pp.prepare_exit_to_usermode
      0.20 +-  2%      -0.1        0.12 +- 70%  perf-profile.self.cycles-pp.perf_exclude_event
      0.16            -0.1        0.10 +- 70%  perf-profile.self.cycles-pp._cond_resched
      0.13            -0.1        0.07 +- 70%  perf-profile.self.cycles-pp.rcu_all_qs
      0.07            -0.0        0.04 +- 70%  perf-profile.self.cycles-pp.ftrace_lookup_ip

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/always/context_switch1/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
           :3           33%           1:3     dmesg.WARNING:at#for_ip_interrupt_entry/0x
           :3           33%           1:3     dmesg.WARNING:at#for_ip_ret_from_intr/0x
           :3           67%           2:3     kmsg.pstore:crypto_comp_decompress_failed,ret=
           :3           67%           2:3     kmsg.pstore:decompression_failed
         %stddev     %change         %stddev
             \          |                \  
    223910            -1.3%     220930        will-it-scale.per_process_ops
    233722            -1.0%     231288        will-it-scale.per_thread_ops
 6.001e+08 +- 13%     +31.4%  7.887e+08 +-  4%  will-it-scale.time.involuntary_context_switches
     18003 +-  4%     +10.9%      19956        will-it-scale.time.minor_page_faults
  1.29e+10            -2.5%  1.258e+10        will-it-scale.time.voluntary_context_switches
  87865617            -1.2%   86826277        will-it-scale.workload
   2880329 +-  2%      +5.4%    3034904        interrupts.CAL:Function_call_interrupts
   7695018           -23.3%    5905066 +-  8%  meminfo.DirectMap2M
      0.00 +- 39%      -0.0        0.00 +- 78%  mpstat.cpu.iowait%
      4621 +- 12%     +13.4%       5241        proc-vmstat.numa_hint_faults_local
    715714           +27.6%     913142 +- 13%  softirqs.SCHED
    515653 +-  6%     -20.0%     412650 +- 15%  turbostat.C1
  43643516            -1.2%   43127031        vmstat.system.cs
   2893393 +-  4%     -23.6%    2210524 +- 10%  cpuidle.C1.time
    518051 +-  6%     -19.9%     415081 +- 15%  cpuidle.C1.usage
     23.10           +22.9%      28.38 +-  9%  boot-time.boot
     18.38           +23.2%      22.64 +- 12%  boot-time.dhcp
      5216            +5.0%       5478 +-  2%  boot-time.idle
    963.76 +- 44%    +109.7%       2021 +- 34%  irq_exception_noise.__do_page_fault.sum
      6.33 +- 14%    +726.3%      52.33 +- 62%  irq_exception_noise.irq_time
     56524 +-  7%     -18.8%      45915 +-  4%  irq_exception_noise.softirq_time
 6.001e+08 +- 13%     +31.4%  7.887e+08 +-  4%  time.involuntary_context_switches
     18003 +-  4%     +10.9%      19956        time.minor_page_faults
  1.29e+10            -2.5%  1.258e+10        time.voluntary_context_switches
      1386 +-  7%     +15.4%       1600 +- 11%  slabinfo.scsi_sense_cache.active_objs
      1386 +-  7%     +15.4%       1600 +- 11%  slabinfo.scsi_sense_cache.num_objs
      1427 +-  5%      -8.9%       1299 +-  2%  slabinfo.task_group.active_objs
      1427 +-  5%      -8.9%       1299 +-  2%  slabinfo.task_group.num_objs
     65519 +- 12%     +20.6%      79014 +- 16%  numa-meminfo.node0.SUnreclaim
      8484           -11.9%       7475 +-  7%  numa-meminfo.node1.KernelStack
      9264 +- 26%     -33.7%       6146 +-  7%  numa-meminfo.node1.Mapped
      2138 +- 61%    +373.5%      10127 +- 92%  numa-meminfo.node3.Inactive
      2059 +- 61%    +387.8%      10046 +- 93%  numa-meminfo.node3.Inactive(anon)
     16379 +- 12%     +20.6%      19752 +- 16%  numa-vmstat.node0.nr_slab_unreclaimable
      8483           -11.9%       7474 +-  7%  numa-vmstat.node1.nr_kernel_stack
      6250 +- 29%     -42.8%       3575 +- 24%  numa-vmstat.node2
      3798 +- 17%     +63.7%       6218 +-  5%  numa-vmstat.node3
    543.00 +- 61%    +368.1%       2541 +- 91%  numa-vmstat.node3.nr_inactive_anon
    543.33 +- 61%    +367.8%       2541 +- 91%  numa-vmstat.node3.nr_zone_inactive_anon
 4.138e+13            -1.1%   4.09e+13        perf-stat.branch-instructions
 6.569e+11            -2.0%  6.441e+11        perf-stat.branch-misses
 2.645e+10            -1.2%  2.613e+10        perf-stat.context-switches
      1.21            +1.2%       1.23        perf-stat.cpi
    153343 +-  2%     -12.1%     134776        perf-stat.cpu-migrations
 5.966e+13            -1.3%  5.889e+13        perf-stat.dTLB-loads
 3.736e+13            -1.2%   3.69e+13        perf-stat.dTLB-stores
      5.85 +- 15%      +8.8       14.67 +-  9%  perf-stat.iTLB-load-miss-rate%
 3.736e+09 +- 17%    +161.3%   9.76e+09 +- 11%  perf-stat.iTLB-load-misses
 5.987e+10            -5.4%  5.667e+10        perf-stat.iTLB-loads
 2.079e+14            -1.2%  2.054e+14        perf-stat.instructions
     57547 +- 18%     -62.9%      21340 +- 11%  perf-stat.instructions-per-iTLB-miss
      0.82            -1.2%       0.81        perf-stat.ipc
  27502531 +-  8%      +9.5%   30122136 +-  3%  perf-stat.node-store-misses
      1449 +- 27%     -34.6%     948.85        sched_debug.cfs_rq:/.load.min
    319416 +-115%    -188.5%    -282549        sched_debug.cfs_rq:/.spread0.avg
    657044 +- 55%     -88.3%      76887 +- 23%  sched_debug.cfs_rq:/.spread0.max
  -1525243           +54.6%   -2357898        sched_debug.cfs_rq:/.spread0.min
    101614 +-  6%     +30.6%     132713 +- 19%  sched_debug.cpu.avg_idle.stddev
     11.54 +- 41%     -61.2%       4.48        sched_debug.cpu.cpu_load[1].avg
      1369 +- 67%     -98.5%      20.67 +- 48%  sched_debug.cpu.cpu_load[1].max
     99.29 +- 67%     -97.6%       2.35 +- 26%  sched_debug.cpu.cpu_load[1].stddev
      9.58 +- 38%     -55.2%       4.29        sched_debug.cpu.cpu_load[2].avg
      1024 +- 68%     -98.5%      15.27 +- 36%  sched_debug.cpu.cpu_load[2].max
     74.51 +- 67%     -97.3%       1.99 +- 15%  sched_debug.cpu.cpu_load[2].stddev
      7.37 +- 29%     -42.0%       4.28        sched_debug.cpu.cpu_load[3].avg
    600.58 +- 68%     -97.9%      12.48 +- 20%  sched_debug.cpu.cpu_load[3].max
     43.98 +- 66%     -95.8%       1.83 +-  5%  sched_debug.cpu.cpu_load[3].stddev
      5.95 +- 19%     -28.1%       4.28        sched_debug.cpu.cpu_load[4].avg
    325.39 +- 67%     -96.4%      11.67 +- 10%  sched_debug.cpu.cpu_load[4].max
     24.19 +- 65%     -92.5%       1.81 +-  3%  sched_debug.cpu.cpu_load[4].stddev
    907.23 +-  4%     -14.1%     779.70 +- 10%  sched_debug.cpu.nr_load_updates.stddev
      0.00 +- 83%    +122.5%       0.00        sched_debug.rt_rq:/.rt_time.min
      8.49 +-  2%      -0.3        8.21 +-  2%  perf-profile.calltrace.cycles-pp.dequeue_task_fair.__schedule.schedule.pipe_wait.pipe_read
     57.28            -0.3       57.01        perf-profile.calltrace.cycles-pp.read
      5.06            -0.2        4.85        perf-profile.calltrace.cycles-pp.select_task_rq_fair.try_to_wake_up.autoremove_wake_function.__wake_up_common.__wake_up_common_lock
      4.98            -0.2        4.78        perf-profile.calltrace.cycles-pp.__switch_to.read
      3.55            -0.2        3.39 +-  2%  perf-profile.calltrace.cycles-pp.syscall_return_via_sysret.read
      2.72            -0.1        2.60        perf-profile.calltrace.cycles-pp.reweight_entity.enqueue_task_fair.ttwu_do_activate.try_to_wake_up.autoremove_wake_function
      2.67            -0.1        2.57 +-  2%  perf-profile.calltrace.cycles-pp.reweight_entity.dequeue_task_fair.__schedule.schedule.pipe_wait
      3.40            -0.1        3.31        perf-profile.calltrace.cycles-pp.syscall_return_via_sysret.write
      3.77            -0.1        3.68        perf-profile.calltrace.cycles-pp.select_idle_sibling.select_task_rq_fair.try_to_wake_up.autoremove_wake_function.__wake_up_common
      1.95            -0.1        1.88        perf-profile.calltrace.cycles-pp.copy_page_to_iter.pipe_read.__vfs_read.vfs_read.ksys_read
      2.19            -0.1        2.13        perf-profile.calltrace.cycles-pp.__switch_to_asm.read
      1.30            -0.1        1.25        perf-profile.calltrace.cycles-pp.update_curr.reweight_entity.enqueue_task_fair.ttwu_do_activate.try_to_wake_up
      1.27            -0.1        1.22 +-  2%  perf-profile.calltrace.cycles-pp.update_curr.reweight_entity.dequeue_task_fair.__schedule.schedule
      2.29            -0.0        2.24        perf-profile.calltrace.cycles-pp.load_new_mm_cr3.switch_mm_irqs_off.__schedule.schedule.pipe_wait
      0.96            -0.0        0.92        perf-profile.calltrace.cycles-pp.__calc_delta.update_curr.reweight_entity.dequeue_task_fair.__schedule
      0.85            -0.0        0.81 +-  3%  perf-profile.calltrace.cycles-pp.cpumask_next_wrap.select_idle_sibling.select_task_rq_fair.try_to_wake_up.autoremove_wake_function
      1.63            -0.0        1.59        perf-profile.calltrace.cycles-pp.native_write_msr.read
      0.72            -0.0        0.69        perf-profile.calltrace.cycles-pp.copyout.copy_page_to_iter.pipe_read.__vfs_read.vfs_read
      0.65 +-  2%      -0.0        0.62        perf-profile.calltrace.cycles-pp._raw_spin_lock_irqsave.try_to_wake_up.autoremove_wake_function.__wake_up_common.__wake_up_common_lock
      0.61            -0.0        0.58 +-  2%  perf-profile.calltrace.cycles-pp.find_next_bit.cpumask_next_wrap.select_idle_sibling.select_task_rq_fair.try_to_wake_up
      0.88            -0.0        0.85        perf-profile.calltrace.cycles-pp.touch_atime.pipe_read.__vfs_read.vfs_read.ksys_read
      0.80            -0.0        0.77 +-  2%  perf-profile.calltrace.cycles-pp.___perf_sw_event.__schedule.schedule.pipe_wait.pipe_read
      0.82            -0.0        0.79        perf-profile.calltrace.cycles-pp.prepare_to_wait.pipe_wait.pipe_read.__vfs_read.vfs_read
      0.72            -0.0        0.70        perf-profile.calltrace.cycles-pp.mutex_lock.pipe_write.__vfs_write.vfs_write.ksys_write
      0.56 +-  2%      -0.0        0.53        perf-profile.calltrace.cycles-pp.update_rq_clock.try_to_wake_up.autoremove_wake_function.__wake_up_common.__wake_up_common_lock
      0.83            -0.0        0.81        perf-profile.calltrace.cycles-pp.__wake_up_common_lock.pipe_read.__vfs_read.vfs_read.ksys_read
     42.40            +0.3       42.69        perf-profile.calltrace.cycles-pp.write
     31.80            +0.4       32.18        perf-profile.calltrace.cycles-pp.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
     24.35            +0.5       24.84        perf-profile.calltrace.cycles-pp.pipe_wait.pipe_read.__vfs_read.vfs_read.ksys_read
     20.36            +0.6       20.92 +-  2%  perf-profile.calltrace.cycles-pp.try_to_wake_up.autoremove_wake_function.__wake_up_common.__wake_up_common_lock.pipe_write
     22.01            +0.6       22.58        perf-profile.calltrace.cycles-pp.schedule.pipe_wait.pipe_read.__vfs_read.vfs_read
     21.87            +0.6       22.46        perf-profile.calltrace.cycles-pp.__schedule.schedule.pipe_wait.pipe_read.__vfs_read
      3.15 +- 11%      +1.0        4.12 +- 14%  perf-profile.calltrace.cycles-pp.ttwu_do_wakeup.try_to_wake_up.autoremove_wake_function.__wake_up_common.__wake_up_common_lock
      1.07 +- 34%      +1.1        2.12 +- 31%  perf-profile.calltrace.cycles-pp.tracing_record_taskinfo_sched_switch.__schedule.schedule.pipe_wait.pipe_read
      0.66 +- 75%      +1.1        1.72 +- 37%  perf-profile.calltrace.cycles-pp.trace_save_cmdline.tracing_record_taskinfo.ttwu_do_wakeup.try_to_wake_up.autoremove_wake_function
      0.75 +- 74%      +1.1        1.88 +- 34%  perf-profile.calltrace.cycles-pp.tracing_record_taskinfo.ttwu_do_wakeup.try_to_wake_up.autoremove_wake_function.__wake_up_common
      0.69 +- 76%      +1.2        1.85 +- 36%  perf-profile.calltrace.cycles-pp.trace_save_cmdline.tracing_record_taskinfo_sched_switch.__schedule.schedule.pipe_wait
      8.73 +-  2%      -0.3        8.45        perf-profile.children.cycles-pp.dequeue_task_fair
     57.28            -0.3       57.01        perf-profile.children.cycles-pp.read
      6.95            -0.2        6.70        perf-profile.children.cycles-pp.syscall_return_via_sysret
      5.57            -0.2        5.35        perf-profile.children.cycles-pp.reweight_entity
      5.26            -0.2        5.05        perf-profile.children.cycles-pp.select_task_rq_fair
      5.19            -0.2        4.99        perf-profile.children.cycles-pp.__switch_to
      4.90            -0.2        4.73 +-  2%  perf-profile.children.cycles-pp.update_curr
      1.27            -0.1        1.13 +-  8%  perf-profile.children.cycles-pp.fsnotify
      3.92            -0.1        3.83        perf-profile.children.cycles-pp.select_idle_sibling
      2.01            -0.1        1.93        perf-profile.children.cycles-pp.__calc_delta
      2.14            -0.1        2.06        perf-profile.children.cycles-pp.copy_page_to_iter
      1.58            -0.1        1.51        perf-profile.children.cycles-pp._raw_spin_unlock_irqrestore
      2.90            -0.1        2.84        perf-profile.children.cycles-pp.update_cfs_group
      1.93            -0.1        1.87        perf-profile.children.cycles-pp._raw_spin_lock_irqsave
      2.35            -0.1        2.29        perf-profile.children.cycles-pp.__switch_to_asm
      1.33            -0.1        1.27 +-  3%  perf-profile.children.cycles-pp.cpumask_next_wrap
      2.57            -0.1        2.52        perf-profile.children.cycles-pp.load_new_mm_cr3
      1.53            -0.1        1.47 +-  2%  perf-profile.children.cycles-pp.__fdget_pos
      1.11            -0.0        1.07 +-  2%  perf-profile.children.cycles-pp.find_next_bit
      1.18            -0.0        1.14        perf-profile.children.cycles-pp.update_rq_clock
      0.88            -0.0        0.83        perf-profile.children.cycles-pp.copy_user_generic_unrolled
      1.70            -0.0        1.65        perf-profile.children.cycles-pp.native_write_msr
      0.97            -0.0        0.93 +-  2%  perf-profile.children.cycles-pp.account_entity_dequeue
      0.59            -0.0        0.56        perf-profile.children.cycles-pp.finish_task_switch
      0.91            -0.0        0.88        perf-profile.children.cycles-pp.touch_atime
      0.69            -0.0        0.65        perf-profile.children.cycles-pp.account_entity_enqueue
      2.13            -0.0        2.09        perf-profile.children.cycles-pp.mutex_lock
      0.32 +-  3%      -0.0        0.29 +-  4%  perf-profile.children.cycles-pp.__sb_start_write
      0.84            -0.0        0.81 +-  2%  perf-profile.children.cycles-pp.___perf_sw_event
      0.89            -0.0        0.87        perf-profile.children.cycles-pp.prepare_to_wait
      0.73            -0.0        0.71        perf-profile.children.cycles-pp.copyout
      0.31 +-  2%      -0.0        0.28 +-  3%  perf-profile.children.cycles-pp.__list_del_entry_valid
      0.46 +-  2%      -0.0        0.44        perf-profile.children.cycles-pp.anon_pipe_buf_release
      0.38            -0.0        0.36 +-  3%  perf-profile.children.cycles-pp.idle_cpu
      0.32            -0.0        0.30 +-  2%  perf-profile.children.cycles-pp.__x64_sys_read
      0.21 +-  2%      -0.0        0.20 +-  2%  perf-profile.children.cycles-pp.deactivate_task
      0.13            -0.0        0.12 +-  4%  perf-profile.children.cycles-pp.timespec_trunc
      0.09            -0.0        0.08        perf-profile.children.cycles-pp.iov_iter_init
      0.08            -0.0        0.07        perf-profile.children.cycles-pp.native_load_tls
      0.11 +-  4%      +0.0        0.12        perf-profile.children.cycles-pp.tick_sched_timer
      0.08 +-  5%      +0.0        0.10 +-  4%  perf-profile.children.cycles-pp.finish_wait
      0.38 +-  2%      +0.0        0.40 +-  2%  perf-profile.children.cycles-pp.file_update_time
      0.31            +0.0        0.33 +-  2%  perf-profile.children.cycles-pp.smp_apic_timer_interrupt
      0.24 +-  3%      +0.0        0.26 +-  3%  perf-profile.children.cycles-pp.rcu_all_qs
      0.39            +0.0        0.41        perf-profile.children.cycles-pp._cond_resched
      0.05            +0.0        0.07 +-  6%  perf-profile.children.cycles-pp.default_wake_function
      0.23 +-  2%      +0.0        0.26 +-  3%  perf-profile.children.cycles-pp.current_time
      0.30            +0.0        0.35 +-  2%  perf-profile.children.cycles-pp.generic_pipe_buf_confirm
      0.52            +0.1        0.58        perf-profile.children.cycles-pp.entry_SYSCALL_64_stage2
      0.00            +0.1        0.08 +-  5%  perf-profile.children.cycles-pp.hrtick_update
     42.40            +0.3       42.69        perf-profile.children.cycles-pp.write
     31.86            +0.4       32.26        perf-profile.children.cycles-pp.__vfs_read
     24.40            +0.5       24.89        perf-profile.children.cycles-pp.pipe_wait
     20.40            +0.6       20.96 +-  2%  perf-profile.children.cycles-pp.try_to_wake_up
     22.30            +0.6       22.89        perf-profile.children.cycles-pp.schedule
     22.22            +0.6       22.84        perf-profile.children.cycles-pp.__schedule
      0.99 +- 36%      +0.9        1.94 +- 32%  perf-profile.children.cycles-pp.tracing_record_taskinfo
      3.30 +- 10%      +1.0        4.27 +- 13%  perf-profile.children.cycles-pp.ttwu_do_wakeup
      1.14 +- 31%      +1.1        2.24 +- 29%  perf-profile.children.cycles-pp.tracing_record_taskinfo_sched_switch
      1.59 +- 46%      +2.0        3.60 +- 36%  perf-profile.children.cycles-pp.trace_save_cmdline
      6.95            -0.2        6.70        perf-profile.self.cycles-pp.syscall_return_via_sysret
      5.19            -0.2        4.99        perf-profile.self.cycles-pp.__switch_to
      1.27            -0.1        1.12 +-  8%  perf-profile.self.cycles-pp.fsnotify
      1.49            -0.1        1.36        perf-profile.self.cycles-pp.select_task_rq_fair
      2.47            -0.1        2.37 +-  2%  perf-profile.self.cycles-pp.reweight_entity
      0.29            -0.1        0.19 +-  2%  perf-profile.self.cycles-pp.ksys_read
      1.50            -0.1        1.42        perf-profile.self.cycles-pp._raw_spin_unlock_irqrestore
      2.01            -0.1        1.93        perf-profile.self.cycles-pp.__calc_delta
      1.93            -0.1        1.86        perf-profile.self.cycles-pp._raw_spin_lock_irqsave
      1.47            -0.1        1.40        perf-profile.self.cycles-pp.dequeue_task_fair
      2.90            -0.1        2.84        perf-profile.self.cycles-pp.update_cfs_group
      1.29            -0.1        1.23        perf-profile.self.cycles-pp.do_syscall_64
      2.57            -0.1        2.52        perf-profile.self.cycles-pp.load_new_mm_cr3
      2.28            -0.1        2.23        perf-profile.self.cycles-pp.__switch_to_asm
      1.80            -0.1        1.75        perf-profile.self.cycles-pp.select_idle_sibling
      1.11            -0.0        1.07 +-  2%  perf-profile.self.cycles-pp.find_next_bit
      0.87            -0.0        0.83        perf-profile.self.cycles-pp.copy_user_generic_unrolled
      0.43            -0.0        0.39 +-  2%  perf-profile.self.cycles-pp.dequeue_entity
      1.70            -0.0        1.65        perf-profile.self.cycles-pp.native_write_msr
      0.92            -0.0        0.88 +-  2%  perf-profile.self.cycles-pp.account_entity_dequeue
      0.48            -0.0        0.44        perf-profile.self.cycles-pp.finish_task_switch
      0.77            -0.0        0.74        perf-profile.self.cycles-pp.___perf_sw_event
      0.66            -0.0        0.63        perf-profile.self.cycles-pp.account_entity_enqueue
      0.46 +-  2%      -0.0        0.43 +-  2%  perf-profile.self.cycles-pp.anon_pipe_buf_release
      0.32 +-  3%      -0.0        0.29 +-  4%  perf-profile.self.cycles-pp.__sb_start_write
      0.31 +-  2%      -0.0        0.28 +-  3%  perf-profile.self.cycles-pp.__list_del_entry_valid
      0.38            -0.0        0.36 +-  3%  perf-profile.self.cycles-pp.idle_cpu
      0.19 +-  4%      -0.0        0.17 +-  2%  perf-profile.self.cycles-pp.__fdget_pos
      0.50            -0.0        0.48        perf-profile.self.cycles-pp.__atime_needs_update
      0.23 +-  2%      -0.0        0.21 +-  3%  perf-profile.self.cycles-pp.touch_atime
      0.31            -0.0        0.30        perf-profile.self.cycles-pp.__x64_sys_read
      0.21 +-  2%      -0.0        0.20 +-  2%  perf-profile.self.cycles-pp.deactivate_task
      0.21 +-  2%      -0.0        0.19        perf-profile.self.cycles-pp.check_preempt_curr
      0.40            -0.0        0.39        perf-profile.self.cycles-pp.autoremove_wake_function
      0.40            -0.0        0.38        perf-profile.self.cycles-pp.copy_user_enhanced_fast_string
      0.27            -0.0        0.26        perf-profile.self.cycles-pp.pipe_wait
      0.13            -0.0        0.12 +-  4%  perf-profile.self.cycles-pp.timespec_trunc
      0.22 +-  2%      -0.0        0.20 +-  2%  perf-profile.self.cycles-pp.put_prev_entity
      0.09            -0.0        0.08        perf-profile.self.cycles-pp.iov_iter_init
      0.08            -0.0        0.07        perf-profile.self.cycles-pp.native_load_tls
      0.11            -0.0        0.10        perf-profile.self.cycles-pp.schedule
      0.12 +-  4%      +0.0        0.13        perf-profile.self.cycles-pp.copyin
      0.08 +-  5%      +0.0        0.10 +-  4%  perf-profile.self.cycles-pp.finish_wait
      0.18            +0.0        0.20 +-  2%  perf-profile.self.cycles-pp.ttwu_do_activate
      0.28 +-  2%      +0.0        0.30 +-  2%  perf-profile.self.cycles-pp._cond_resched
      0.24 +-  3%      +0.0        0.26 +-  3%  perf-profile.self.cycles-pp.rcu_all_qs
      0.05            +0.0        0.07 +-  6%  perf-profile.self.cycles-pp.default_wake_function
      0.08 +- 14%      +0.0        0.11 +- 14%  perf-profile.self.cycles-pp.tracing_record_taskinfo_sched_switch
      0.51            +0.0        0.55 +-  4%  perf-profile.self.cycles-pp.vfs_write
      0.30            +0.0        0.35 +-  2%  perf-profile.self.cycles-pp.generic_pipe_buf_confirm
      0.52            +0.1        0.58        perf-profile.self.cycles-pp.entry_SYSCALL_64_stage2
      0.00            +0.1        0.08 +-  5%  perf-profile.self.cycles-pp.hrtick_update
      1.97            +0.1        2.07 +-  2%  perf-profile.self.cycles-pp.switch_mm_irqs_off
      1.59 +- 46%      +2.0        3.60 +- 36%  perf-profile.self.cycles-pp.trace_save_cmdline

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/never/brk1/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
           :3           33%           1:3     kmsg.pstore:crypto_comp_decompress_failed,ret=
           :3           33%           1:3     kmsg.pstore:decompression_failed
         %stddev     %change         %stddev
             \          |                \  
    997317            -2.0%     977778        will-it-scale.per_process_ops
    957.00            -7.9%     881.00 +-  3%  will-it-scale.per_thread_ops
     18.42 +-  3%      -8.2%      16.90        will-it-scale.time.user_time
 1.917e+08            -2.0%  1.879e+08        will-it-scale.workload
     18.42 +-  3%      -8.2%      16.90        time.user_time
      0.30 +- 11%     -36.7%       0.19 +- 11%  turbostat.Pkg%pc2
     57539 +- 51%    +140.6%     138439 +- 31%  meminfo.CmaFree
    410877 +- 11%     -22.1%     320082 +- 22%  meminfo.DirectMap4k
    343575 +- 27%     +71.3%     588703 +- 31%  numa-numastat.node0.local_node
    374176 +- 24%     +63.3%     611007 +- 27%  numa-numastat.node0.numa_hit
   1056347 +-  4%     -39.9%     634843 +- 38%  numa-numastat.node3.local_node
   1060682 +-  4%     -39.0%     646862 +- 35%  numa-numastat.node3.numa_hit
     14383 +- 51%    +140.6%      34608 +- 31%  proc-vmstat.nr_free_cma
    179.00            +2.4%     183.33        proc-vmstat.nr_inactive_file
    179.00            +2.4%     183.33        proc-vmstat.nr_zone_inactive_file
    564483 +-  3%     -38.0%     350064 +- 36%  proc-vmstat.pgalloc_movable
   1811959           +10.8%    2008488 +-  5%  proc-vmstat.pgalloc_normal
      7153 +- 42%     -94.0%     431.33 +-119%  latency_stats.max.pipe_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      6627 +-141%    +380.5%      31843 +-110%  latency_stats.max.call_rwsem_down_write_failed_killable.do_mprotect_pkey.__x64_sys_mprotect.do_syscall_64.entry_SYSCALL_64_after_hwframe
     15244 +- 31%     -99.9%      15.00 +-141%  latency_stats.sum.call_rwsem_down_read_failed.__do_page_fault.do_page_fault.page_fault.__get_user_8.exit_robust_list.mm_release.do_exit.do_group_exit.get_signal.do_signal.exit_to_usermode_loop
      4301 +-117%     -83.7%     700.33 +-  6%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
     12153 +- 28%     -83.1%       2056 +- 70%  latency_stats.sum.pipe_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      6772 +-141%   +1105.8%      81665 +-127%  latency_stats.sum.call_rwsem_down_write_failed_killable.do_mprotect_pkey.__x64_sys_mprotect.do_syscall_64.entry_SYSCALL_64_after_hwframe
 2.465e+13            -1.3%  2.434e+13        perf-stat.branch-instructions
 2.691e+11            -2.1%  2.635e+11        perf-stat.branch-misses
 3.402e+13            -1.4%  3.355e+13        perf-stat.dTLB-loads
 1.694e+13            +1.4%  1.718e+13        perf-stat.dTLB-stores
      1.75 +- 50%      +4.7        6.45 +- 11%  perf-stat.iTLB-load-miss-rate%
 4.077e+08 +- 48%    +232.3%  1.355e+09 +- 11%  perf-stat.iTLB-load-misses
  2.31e+10 +-  2%     -14.9%  1.965e+10 +-  3%  perf-stat.iTLB-loads
 1.163e+14            -1.6%  1.144e+14        perf-stat.instructions
    346171 +- 36%     -75.3%      85575 +- 11%  perf-stat.instructions-per-iTLB-miss
 6.174e+08 +-  2%      -9.5%  5.589e+08        perf-stat.node-store-misses
    595.00 +- 10%     +31.4%     782.00 +-  3%  slabinfo.Acpi-State.active_objs
    595.00 +- 10%     +31.4%     782.00 +-  3%  slabinfo.Acpi-State.num_objs
      2831 +-  3%     -14.0%       2434 +-  5%  slabinfo.avtab_node.active_objs
      2831 +-  3%     -14.0%       2434 +-  5%  slabinfo.avtab_node.num_objs
    934.00           -10.9%     832.33 +-  5%  slabinfo.inotify_inode_mark.active_objs
    934.00           -10.9%     832.33 +-  5%  slabinfo.inotify_inode_mark.num_objs
      1232 +-  4%     +13.4%       1397 +-  6%  slabinfo.nsproxy.active_objs
      1232 +-  4%     +13.4%       1397 +-  6%  slabinfo.nsproxy.num_objs
    499.67 +- 12%     +24.8%     623.67 +- 10%  slabinfo.secpath_cache.active_objs
    499.67 +- 12%     +24.8%     623.67 +- 10%  slabinfo.secpath_cache.num_objs
     31393 +- 84%    +220.1%     100477 +- 21%  numa-meminfo.node0.Active
     31393 +- 84%    +220.1%     100477 +- 21%  numa-meminfo.node0.Active(anon)
     30013 +- 85%    +232.1%      99661 +- 21%  numa-meminfo.node0.AnonPages
     21603 +- 34%     -85.0%       3237 +-100%  numa-meminfo.node0.Inactive
     21528 +- 34%     -85.0%       3237 +-100%  numa-meminfo.node0.Inactive(anon)
     10247 +- 35%     -46.4%       5495        numa-meminfo.node0.Mapped
     35388 +- 14%     -41.6%      20670 +- 15%  numa-meminfo.node0.SReclaimable
     22911 +- 29%     -82.3%       4057 +- 84%  numa-meminfo.node0.Shmem
    117387 +-  9%     -22.5%      90986 +- 12%  numa-meminfo.node0.Slab
     68863 +- 67%     +77.7%     122351 +- 13%  numa-meminfo.node1.Active
     68863 +- 67%     +77.7%     122351 +- 13%  numa-meminfo.node1.Active(anon)
    228376           +22.3%     279406 +- 17%  numa-meminfo.node1.FilePages
      1481 +-116%   +1062.1%      17218 +- 39%  numa-meminfo.node1.Inactive
      1481 +-116%   +1062.0%      17216 +- 39%  numa-meminfo.node1.Inactive(anon)
      6593 +-  2%     +11.7%       7367 +-  3%  numa-meminfo.node1.KernelStack
    596227 +-  8%     +18.0%     703748 +-  4%  numa-meminfo.node1.MemUsed
     15298 +- 12%     +88.5%      28843 +- 36%  numa-meminfo.node1.SReclaimable
     52718 +-  9%     +21.0%      63810 +- 11%  numa-meminfo.node1.SUnreclaim
      1808 +- 97%   +2723.8%      51054 +- 97%  numa-meminfo.node1.Shmem
     68017 +-  5%     +36.2%      92654 +- 18%  numa-meminfo.node1.Slab
    125541 +- 29%     -64.9%      44024 +- 98%  numa-meminfo.node3.Active
    125137 +- 29%     -65.0%      43823 +- 98%  numa-meminfo.node3.Active(anon)
     93173 +- 25%     -87.8%      11381 +- 20%  numa-meminfo.node3.AnonPages
      9150 +-  5%      -9.3%       8301 +-  8%  numa-meminfo.node3.KernelStack
      7848 +- 84%    +220.0%      25118 +- 21%  numa-vmstat.node0.nr_active_anon
      7503 +- 85%    +232.1%      24914 +- 21%  numa-vmstat.node0.nr_anon_pages
      5381 +- 34%     -85.0%     809.00 +-100%  numa-vmstat.node0.nr_inactive_anon
      2559 +- 35%     -46.4%       1372        numa-vmstat.node0.nr_mapped
      5727 +- 29%     -82.3%       1014 +- 84%  numa-vmstat.node0.nr_shmem
      8846 +- 14%     -41.6%       5167 +- 15%  numa-vmstat.node0.nr_slab_reclaimable
      7848 +- 84%    +220.0%      25118 +- 21%  numa-vmstat.node0.nr_zone_active_anon
      5381 +- 34%     -85.0%     809.00 +-100%  numa-vmstat.node0.nr_zone_inactive_anon
      4821 +-  2%     +30.3%       6283 +- 15%  numa-vmstat.node1
     17215 +- 67%     +77.7%      30591 +- 13%  numa-vmstat.node1.nr_active_anon
     57093           +22.3%      69850 +- 17%  numa-vmstat.node1.nr_file_pages
    370.00 +-116%   +1061.8%       4298 +- 39%  numa-vmstat.node1.nr_inactive_anon
      6593 +-  2%     +11.7%       7366 +-  3%  numa-vmstat.node1.nr_kernel_stack
    451.67 +- 97%   +2725.6%      12762 +- 97%  numa-vmstat.node1.nr_shmem
      3824 +- 12%     +88.6%       7211 +- 36%  numa-vmstat.node1.nr_slab_reclaimable
     13179 +-  9%     +21.0%      15952 +- 11%  numa-vmstat.node1.nr_slab_unreclaimable
     17215 +- 67%     +77.7%      30591 +- 13%  numa-vmstat.node1.nr_zone_active_anon
    370.00 +-116%   +1061.8%       4298 +- 39%  numa-vmstat.node1.nr_zone_inactive_anon
    364789 +- 12%     +62.8%     593926 +- 34%  numa-vmstat.node1.numa_hit
    239539 +- 19%     +95.4%     468113 +- 43%  numa-vmstat.node1.numa_local
     71.00 +- 28%     +42.3%     101.00        numa-vmstat.node2.nr_mlock
     31285 +- 29%     -65.0%      10960 +- 98%  numa-vmstat.node3.nr_active_anon
     23292 +- 25%     -87.8%       2844 +- 19%  numa-vmstat.node3.nr_anon_pages
     14339 +- 52%    +141.1%      34566 +- 32%  numa-vmstat.node3.nr_free_cma
      9151 +-  5%      -9.3%       8299 +-  8%  numa-vmstat.node3.nr_kernel_stack
     31305 +- 29%     -64.9%      10975 +- 98%  numa-vmstat.node3.nr_zone_active_anon
    930131 +-  3%     -35.9%     596006 +- 34%  numa-vmstat.node3.numa_hit
    836455 +-  3%     -40.9%     493947 +- 44%  numa-vmstat.node3.numa_local
     75182 +- 58%     -83.8%      12160 +-  2%  sched_debug.cfs_rq:/.load.max
      6.65 +-  5%     -10.6%       5.94 +-  6%  sched_debug.cfs_rq:/.load_avg.avg
      0.16 +-  7%     +22.6%       0.20 +- 12%  sched_debug.cfs_rq:/.nr_running.stddev
      5.58 +- 24%    +427.7%      29.42 +- 93%  sched_debug.cfs_rq:/.nr_spread_over.max
      0.54 +- 15%    +306.8%       2.19 +- 86%  sched_debug.cfs_rq:/.nr_spread_over.stddev
      1.05 +- 25%     -65.1%       0.37 +- 71%  sched_debug.cfs_rq:/.removed.load_avg.avg
      9.62 +- 11%     -50.7%       4.74 +- 70%  sched_debug.cfs_rq:/.removed.load_avg.stddev
     48.70 +- 25%     -65.1%      17.02 +- 71%  sched_debug.cfs_rq:/.removed.runnable_sum.avg
    444.31 +- 11%     -50.7%     219.26 +- 70%  sched_debug.cfs_rq:/.removed.runnable_sum.stddev
      0.47 +- 13%     -60.9%       0.19 +- 71%  sched_debug.cfs_rq:/.removed.util_avg.avg
      4.47 +-  4%     -46.5%       2.39 +- 70%  sched_debug.cfs_rq:/.removed.util_avg.stddev
      1.64 +-  7%     +22.1%       2.00 +- 13%  sched_debug.cfs_rq:/.runnable_load_avg.stddev
     74653 +- 59%     -84.4%      11676        sched_debug.cfs_rq:/.runnable_weight.max
   -119169          -491.3%     466350 +- 27%  sched_debug.cfs_rq:/.spread0.avg
    517161 +- 30%    +145.8%    1271292 +- 23%  sched_debug.cfs_rq:/.spread0.max
    624.79 +-  5%     -14.2%     535.76 +-  7%  sched_debug.cfs_rq:/.util_est_enqueued.avg
    247.91 +- 32%     -99.8%       0.48 +-  8%  sched_debug.cfs_rq:/.util_est_enqueued.min
    179704 +-  3%     +30.4%     234297 +- 16%  sched_debug.cpu.avg_idle.stddev
      1.56 +-  9%     +24.4%       1.94 +- 14%  sched_debug.cpu.cpu_load[0].stddev
      1.50 +-  6%     +27.7%       1.91 +- 14%  sched_debug.cpu.cpu_load[1].stddev
      1.45 +-  3%     +30.8%       1.90 +- 14%  sched_debug.cpu.cpu_load[2].stddev
      1.43 +-  3%     +36.1%       1.95 +- 11%  sched_debug.cpu.cpu_load[3].stddev
      1.55 +-  7%     +43.5%       2.22 +-  7%  sched_debug.cpu.cpu_load[4].stddev
     10004 +-  3%     -11.6%       8839 +-  3%  sched_debug.cpu.curr->pid.avg
      1146 +- 26%     +52.2%       1745 +-  7%  sched_debug.cpu.curr->pid.min
      3162 +-  6%     +25.4%       3966 +- 11%  sched_debug.cpu.curr->pid.stddev
    403738 +-  3%     -11.7%     356696 +-  7%  sched_debug.cpu.nr_switches.max
      0.08 +- 21%     +78.2%       0.14 +- 14%  sched_debug.cpu.nr_uninterruptible.avg
    404435 +-  3%     -11.8%     356732 +-  7%  sched_debug.cpu.sched_count.max
      4.17            -0.3        3.87        perf-profile.calltrace.cycles-pp.kmem_cache_alloc.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      2.40            -0.2        2.17        perf-profile.calltrace.cycles-pp.vma_compute_subtree_gap.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk
      7.58            -0.2        7.36        perf-profile.calltrace.cycles-pp.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     15.00            -0.2       14.81        perf-profile.calltrace.cycles-pp.syscall_return_via_sysret.brk
      7.83            -0.2        7.66        perf-profile.calltrace.cycles-pp.unmap_vmas.unmap_region.do_munmap.__x64_sys_brk.do_syscall_64
     28.66            -0.1       28.51        perf-profile.calltrace.cycles-pp.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      2.15            -0.1        2.03        perf-profile.calltrace.cycles-pp.vma_compute_subtree_gap.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.07            -0.1        0.99        perf-profile.calltrace.cycles-pp.memcpy_erms.strlcpy.perf_event_mmap.do_brk_flags.__x64_sys_brk
      1.03            -0.1        0.95        perf-profile.calltrace.cycles-pp.kmem_cache_free.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64
      7.33            -0.1        7.25        perf-profile.calltrace.cycles-pp.unmap_page_range.unmap_vmas.unmap_region.do_munmap.__x64_sys_brk
      0.76            -0.1        0.69        perf-profile.calltrace.cycles-pp.__vm_enough_memory.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     11.85            -0.1       11.77        perf-profile.calltrace.cycles-pp.unmap_region.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.64            -0.1        1.57        perf-profile.calltrace.cycles-pp.strlcpy.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64
      1.06            -0.1        0.99        perf-profile.calltrace.cycles-pp.__indirect_thunk_start.brk
      0.73            -0.1        0.67        perf-profile.calltrace.cycles-pp.sync_mm_rss.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      4.59            -0.1        4.52        perf-profile.calltrace.cycles-pp.security_vm_enough_memory_mm.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      2.82            -0.1        2.76        perf-profile.calltrace.cycles-pp.selinux_vm_enough_memory.security_vm_enough_memory_mm.do_brk_flags.__x64_sys_brk.do_syscall_64
      2.89            -0.1        2.84        perf-profile.calltrace.cycles-pp.down_write_killable.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      3.37            -0.1        3.32        perf-profile.calltrace.cycles-pp.get_unmapped_area.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.99            -0.0        1.94        perf-profile.calltrace.cycles-pp.cred_has_capability.selinux_vm_enough_memory.security_vm_enough_memory_mm.do_brk_flags.__x64_sys_brk
      2.32            -0.0        2.27        perf-profile.calltrace.cycles-pp.perf_iterate_sb.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64
      1.88            -0.0        1.84        perf-profile.calltrace.cycles-pp.security_mmap_addr.get_unmapped_area.do_brk_flags.__x64_sys_brk.do_syscall_64
      0.77            -0.0        0.73        perf-profile.calltrace.cycles-pp._raw_spin_lock.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      1.62            -0.0        1.59        perf-profile.calltrace.cycles-pp.memset_erms.kmem_cache_alloc.do_brk_flags.__x64_sys_brk.do_syscall_64
      0.81            -0.0        0.79        perf-profile.calltrace.cycles-pp.___might_sleep.down_write_killable.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.66            -0.0        0.64        perf-profile.calltrace.cycles-pp.arch_get_unmapped_area_topdown.brk
      0.72            +0.0        0.74        perf-profile.calltrace.cycles-pp.do_munmap.brk
      0.90            +0.0        0.93        perf-profile.calltrace.cycles-pp.___might_sleep.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      4.40            +0.1        4.47        perf-profile.calltrace.cycles-pp.find_vma.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.96            +0.1        2.09        perf-profile.calltrace.cycles-pp.vmacache_find.find_vma.do_munmap.__x64_sys_brk.do_syscall_64
      0.52 +-  2%      +0.2        0.68        perf-profile.calltrace.cycles-pp.__vma_link_rb.brk
      0.35 +- 70%      +0.2        0.54 +-  2%  perf-profile.calltrace.cycles-pp.find_vma.brk
      2.20            +0.3        2.50        perf-profile.calltrace.cycles-pp.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     64.62            +0.3       64.94        perf-profile.calltrace.cycles-pp.entry_SYSCALL_64_after_hwframe.brk
     60.53            +0.4       60.92        perf-profile.calltrace.cycles-pp.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
     63.20            +0.4       63.60        perf-profile.calltrace.cycles-pp.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      3.73            +0.5        4.26        perf-profile.calltrace.cycles-pp.vma_link.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00            +0.6        0.56        perf-profile.calltrace.cycles-pp.free_pgtables.unmap_region.do_munmap.__x64_sys_brk.do_syscall_64
     24.54            +0.6       25.14        perf-profile.calltrace.cycles-pp.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      0.00            +0.6        0.64        perf-profile.calltrace.cycles-pp.put_vma.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64
      0.71            +0.6        1.36        perf-profile.calltrace.cycles-pp.__vma_rb_erase.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00            +0.7        0.70        perf-profile.calltrace.cycles-pp._raw_write_lock.__vma_rb_erase.do_munmap.__x64_sys_brk.do_syscall_64
      3.10            +0.7        3.82        perf-profile.calltrace.cycles-pp.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk.do_syscall_64
      0.00            +0.8        0.76        perf-profile.calltrace.cycles-pp._raw_write_lock.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk
      0.00            +0.8        0.85        perf-profile.calltrace.cycles-pp.__vma_merge.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      5.09            -0.5        4.62        perf-profile.children.cycles-pp.vma_compute_subtree_gap
      4.54            -0.3        4.21        perf-profile.children.cycles-pp.kmem_cache_alloc
      8.11            -0.2        7.89        perf-profile.children.cycles-pp.perf_event_mmap
      8.05            -0.2        7.85        perf-profile.children.cycles-pp.unmap_vmas
     15.01            -0.2       14.81        perf-profile.children.cycles-pp.syscall_return_via_sysret
     29.20            -0.1       29.06        perf-profile.children.cycles-pp.do_brk_flags
      1.11            -0.1        1.00        perf-profile.children.cycles-pp.kmem_cache_free
     12.28            -0.1       12.17        perf-profile.children.cycles-pp.unmap_region
      7.83            -0.1        7.74        perf-profile.children.cycles-pp.unmap_page_range
      0.87 +-  3%      -0.1        0.79        perf-profile.children.cycles-pp.__vm_enough_memory
      1.29            -0.1        1.22        perf-profile.children.cycles-pp.__indirect_thunk_start
      1.81            -0.1        1.74        perf-profile.children.cycles-pp.strlcpy
      4.65            -0.1        4.58        perf-profile.children.cycles-pp.security_vm_enough_memory_mm
      3.08            -0.1        3.02        perf-profile.children.cycles-pp.down_write_killable
      2.88            -0.1        2.82        perf-profile.children.cycles-pp.selinux_vm_enough_memory
      0.73            -0.1        0.67        perf-profile.children.cycles-pp.sync_mm_rss
      3.65            -0.1        3.59        perf-profile.children.cycles-pp.get_unmapped_area
      2.26            -0.1        2.20        perf-profile.children.cycles-pp.cred_has_capability
      1.12            -0.1        1.07        perf-profile.children.cycles-pp.memcpy_erms
      0.39            -0.0        0.35        perf-profile.children.cycles-pp.__rb_insert_augmented
      2.52            -0.0        2.48        perf-profile.children.cycles-pp.perf_iterate_sb
      2.13            -0.0        2.09        perf-profile.children.cycles-pp.security_mmap_addr
      0.55 +-  2%      -0.0        0.52        perf-profile.children.cycles-pp.unmap_single_vma
      1.62            -0.0        1.59        perf-profile.children.cycles-pp.memset_erms
      0.13 +-  3%      -0.0        0.11 +-  4%  perf-profile.children.cycles-pp.__vma_link_file
      0.80            -0.0        0.77        perf-profile.children.cycles-pp._raw_spin_lock
      0.43            -0.0        0.41        perf-profile.children.cycles-pp.strlen
      0.07 +-  6%      -0.0        0.06 +-  8%  perf-profile.children.cycles-pp.should_failslab
      0.43            -0.0        0.42        perf-profile.children.cycles-pp.may_expand_vm
      0.15            +0.0        0.16        perf-profile.children.cycles-pp.__vma_link_list
      0.45            +0.0        0.47        perf-profile.children.cycles-pp.rcu_all_qs
      0.81            +0.1        0.89        perf-profile.children.cycles-pp.free_pgtables
      6.35            +0.1        6.49        perf-profile.children.cycles-pp.find_vma
      2.28            +0.2        2.45        perf-profile.children.cycles-pp.vmacache_find
     64.66            +0.3       64.98        perf-profile.children.cycles-pp.entry_SYSCALL_64_after_hwframe
      2.42            +0.3        2.76        perf-profile.children.cycles-pp.remove_vma
     61.77            +0.4       62.13        perf-profile.children.cycles-pp.__x64_sys_brk
     63.40            +0.4       63.79        perf-profile.children.cycles-pp.do_syscall_64
      1.27            +0.4        1.72        perf-profile.children.cycles-pp.__vma_rb_erase
      4.02            +0.5        4.53        perf-profile.children.cycles-pp.vma_link
     25.26            +0.6       25.89        perf-profile.children.cycles-pp.do_munmap
      0.00            +0.7        0.70        perf-profile.children.cycles-pp.put_vma
      3.80            +0.7        4.53        perf-profile.children.cycles-pp.__vma_link_rb
      0.00            +1.2        1.24        perf-profile.children.cycles-pp.__vma_merge
      0.00            +1.5        1.51        perf-profile.children.cycles-pp._raw_write_lock
      5.07            -0.5        4.60        perf-profile.self.cycles-pp.vma_compute_subtree_gap
      0.59            -0.2        0.38        perf-profile.self.cycles-pp.remove_vma
     15.01            -0.2       14.81        perf-profile.self.cycles-pp.syscall_return_via_sysret
      3.15            -0.2        2.96        perf-profile.self.cycles-pp.do_munmap
      0.98            -0.1        0.87        perf-profile.self.cycles-pp.__vma_rb_erase
      1.10            -0.1        0.99        perf-profile.self.cycles-pp.kmem_cache_free
      0.68            -0.1        0.58        perf-profile.self.cycles-pp.__vm_enough_memory
      0.42            -0.1        0.33        perf-profile.self.cycles-pp.unmap_vmas
      3.62            -0.1        3.53        perf-profile.self.cycles-pp.perf_event_mmap
      1.41            -0.1        1.34        perf-profile.self.cycles-pp.entry_SYSCALL_64_after_hwframe
      1.29            -0.1        1.22        perf-profile.self.cycles-pp.__indirect_thunk_start
      0.73            -0.1        0.66        perf-profile.self.cycles-pp.sync_mm_rss
      2.96            -0.1        2.90        perf-profile.self.cycles-pp.__x64_sys_brk
      3.24            -0.1        3.19        perf-profile.self.cycles-pp.brk
      1.11            -0.0        1.07        perf-profile.self.cycles-pp.memcpy_erms
      0.53 +-  3%      -0.0        0.49 +-  2%  perf-profile.self.cycles-pp.vma_link
      0.73            -0.0        0.69        perf-profile.self.cycles-pp.unmap_region
      1.66            -0.0        1.61        perf-profile.self.cycles-pp.down_write_killable
      0.39            -0.0        0.35        perf-profile.self.cycles-pp.__rb_insert_augmented
      1.74            -0.0        1.71        perf-profile.self.cycles-pp.kmem_cache_alloc
      0.55 +-  2%      -0.0        0.52        perf-profile.self.cycles-pp.unmap_single_vma
      1.61            -0.0        1.59        perf-profile.self.cycles-pp.memset_erms
      0.80            -0.0        0.77        perf-profile.self.cycles-pp._raw_spin_lock
      0.13            -0.0        0.11 +-  4%  perf-profile.self.cycles-pp.__vma_link_file
      0.43            -0.0        0.41        perf-profile.self.cycles-pp.strlen
      0.07 +-  6%      -0.0        0.06 +-  8%  perf-profile.self.cycles-pp.should_failslab
      0.81            -0.0        0.79        perf-profile.self.cycles-pp.tlb_finish_mmu
      0.15            +0.0        0.16        perf-profile.self.cycles-pp.__vma_link_list
      0.45            +0.0        0.47        perf-profile.self.cycles-pp.rcu_all_qs
      0.71            +0.0        0.72        perf-profile.self.cycles-pp.strlcpy
      0.51            +0.1        0.56        perf-profile.self.cycles-pp.free_pgtables
      1.41            +0.1        1.48        perf-profile.self.cycles-pp.__vma_link_rb
      2.27            +0.2        2.44        perf-profile.self.cycles-pp.vmacache_find
      0.00            +0.7        0.69        perf-profile.self.cycles-pp.put_vma
      0.00            +1.2        1.23        perf-profile.self.cycles-pp.__vma_merge
      0.00            +1.5        1.50        perf-profile.self.cycles-pp._raw_write_lock

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/always/brk1/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
           :3           33%           1:3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=schedule_tail/0x
           :3           33%           1:3     kmsg.DHCP/BOOTP:Reply_not_for_us_on_eth#,op[#]xid[#]
         %stddev     %change         %stddev
             \          |                \  
    998475            -2.2%     976893        will-it-scale.per_process_ops
    625.87            -2.3%     611.42        will-it-scale.time.elapsed_time
    625.87            -2.3%     611.42        will-it-scale.time.elapsed_time.max
      8158            -1.9%       8000        will-it-scale.time.maximum_resident_set_size
     18.42 +-  2%     -11.9%      16.24        will-it-scale.time.user_time
  34349225 +- 13%     -14.5%   29371024 +- 17%  will-it-scale.time.voluntary_context_switches
 1.919e+08            -2.2%  1.877e+08        will-it-scale.workload
      1639 +- 23%     -18.4%       1337 +- 30%  meminfo.Mlocked
     17748 +- 82%    +103.1%      36051        numa-numastat.node3.other_node
  33410486 +- 14%     -14.8%   28449258 +- 18%  cpuidle.C1.usage
    698749 +- 15%     -18.0%     573307 +- 20%  cpuidle.POLL.usage
   3013702 +- 14%     -15.1%    2559405 +- 17%  softirqs.SCHED
  54361293 +-  2%     -19.0%   44044816 +-  2%  softirqs.TIMER
  33408303 +- 14%     -14.9%   28447123 +- 18%  turbostat.C1
      0.34 +- 16%     -52.0%       0.16 +- 15%  turbostat.Pkg%pc2
      1310 +- 74%    +412.1%       6710 +- 58%  irq_exception_noise.__do_page_fault.samples
      3209 +- 74%    +281.9%      12258 +- 53%  irq_exception_noise.__do_page_fault.sum
    600.67 +-132%     -96.0%      24.00 +- 23%  irq_exception_noise.irq_nr
     99557 +-  7%     -24.0%      75627 +-  7%  irq_exception_noise.softirq_nr
     41424 +-  9%     -24.6%      31253 +-  6%  irq_exception_noise.softirq_time
    625.87            -2.3%     611.42        time.elapsed_time
    625.87            -2.3%     611.42        time.elapsed_time.max
      8158            -1.9%       8000        time.maximum_resident_set_size
     18.42 +-  2%     -11.9%      16.24        time.user_time
  34349225 +- 13%     -14.5%   29371024 +- 17%  time.voluntary_context_switches
    988.00 +-  8%     +14.5%       1131 +-  2%  slabinfo.Acpi-ParseExt.active_objs
    988.00 +-  8%     +14.5%       1131 +-  2%  slabinfo.Acpi-ParseExt.num_objs
      2384 +-  3%     +21.1%       2888 +- 11%  slabinfo.pool_workqueue.active_objs
      2474 +-  2%     +20.4%       2979 +- 11%  slabinfo.pool_workqueue.num_objs
    490.33 +- 10%     -19.2%     396.00 +- 11%  slabinfo.secpath_cache.active_objs
    490.33 +- 10%     -19.2%     396.00 +- 11%  slabinfo.secpath_cache.num_objs
      1123 +-  7%     +14.2%       1282 +-  3%  slabinfo.skbuff_fclone_cache.active_objs
      1123 +-  7%     +14.2%       1282 +-  3%  slabinfo.skbuff_fclone_cache.num_objs
      1.09            -0.0        1.07        perf-stat.branch-miss-rate%
 2.691e+11            -2.4%  2.628e+11        perf-stat.branch-misses
  71981351 +- 12%     -13.8%   62013509 +- 16%  perf-stat.context-switches
 1.697e+13            +1.1%  1.715e+13        perf-stat.dTLB-stores
      2.36 +- 29%      +4.4        6.76 +- 11%  perf-stat.iTLB-load-miss-rate%
  5.21e+08 +- 28%    +194.8%  1.536e+09 +- 10%  perf-stat.iTLB-load-misses
    239983 +- 24%     -68.4%      75819 +- 11%  perf-stat.instructions-per-iTLB-miss
   3295653 +-  2%      -6.3%    3088753 +-  3%  perf-stat.node-stores
    606239            +1.1%     612799        perf-stat.path-length
      3755 +- 28%     -37.5%       2346 +- 52%  sched_debug.cfs_rq:/.exec_clock.stddev
     10.45 +-  4%     +24.3%      12.98 +- 18%  sched_debug.cfs_rq:/.load_avg.stddev
      6243 +- 46%     -38.6%       3831 +- 78%  sched_debug.cpu.load.stddev
    867.80 +-  7%     +25.3%       1087 +-  6%  sched_debug.cpu.nr_load_updates.stddev
    395898 +-  3%     -11.1%     352071 +-  7%  sched_debug.cpu.nr_switches.max
    -13.33           -21.1%     -10.52        sched_debug.cpu.nr_uninterruptible.min
    395674 +-  3%     -11.1%     351762 +-  7%  sched_debug.cpu.sched_count.max
     33152 +-  4%     -12.8%      28899        sched_debug.cpu.ttwu_count.min
      0.03 +- 20%     +77.7%       0.05 +- 15%  sched_debug.rt_rq:/.rt_time.max
     89523            +1.8%      91099        proc-vmstat.nr_active_anon
    409.67 +- 23%     -18.4%     334.33 +- 30%  proc-vmstat.nr_mlock
     89530            +1.8%      91117        proc-vmstat.nr_zone_active_anon
   2337130            -2.2%    2286775        proc-vmstat.numa_hit
   2229090            -2.3%    2178626        proc-vmstat.numa_local
      8460 +- 39%     -75.5%       2076 +- 53%  proc-vmstat.numa_pages_migrated
     28643 +- 55%     -83.5%       4727 +- 58%  proc-vmstat.numa_pte_updates
   2695806            -1.8%    2646639        proc-vmstat.pgfault
   2330191            -2.1%    2281197        proc-vmstat.pgfree
      8460 +- 39%     -75.5%       2076 +- 53%  proc-vmstat.pgmigrate_success
    237651 +-  2%     +31.3%     312092 +- 16%  numa-meminfo.node0.FilePages
      8059 +-  2%     +10.7%       8925 +-  7%  numa-meminfo.node0.KernelStack
      6830 +- 25%     +48.8%      10164 +- 35%  numa-meminfo.node0.Mapped
      1612 +- 21%     +70.0%       2740 +- 19%  numa-meminfo.node0.PageTables
     10772 +- 65%    +679.4%      83962 +- 59%  numa-meminfo.node0.Shmem
    163195 +- 15%     -36.9%     103036 +- 32%  numa-meminfo.node1.Active
    163195 +- 15%     -36.9%     103036 +- 32%  numa-meminfo.node1.Active(anon)
      1730 +-  4%     +33.9%       2317 +- 14%  numa-meminfo.node1.PageTables
     55778 +- 19%     +32.5%      73910 +-  8%  numa-meminfo.node1.SUnreclaim
      2671 +- 16%     -45.0%       1469 +- 15%  numa-meminfo.node2.PageTables
     61537 +- 13%     -17.7%      50647 +-  3%  numa-meminfo.node2.SUnreclaim
     48644 +- 94%    +149.8%     121499 +- 11%  numa-meminfo.node3.Active
     48440 +- 94%    +150.4%     121295 +- 11%  numa-meminfo.node3.Active(anon)
     11832 +- 79%     -91.5%       1008 +- 67%  numa-meminfo.node3.Inactive
     11597 +- 82%     -93.3%     772.00 +- 82%  numa-meminfo.node3.Inactive(anon)
     10389 +- 32%     -43.0%       5921 +-  6%  numa-meminfo.node3.Mapped
     33704 +- 24%     -44.2%      18792 +- 15%  numa-meminfo.node3.SReclaimable
    104733 +- 14%     -25.3%      78275 +-  8%  numa-meminfo.node3.Slab
    139329 +-133%     -99.8%     241.67 +- 79%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
      5403 +-139%     -97.5%     137.67 +- 71%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
    165968 +-101%     -61.9%      63304 +- 58%  latency_stats.avg.max
     83.00        +12810.4%      10715 +-140%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_access.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat.filename_lookup
    102.67 +-  6%  +18845.5%      19450 +-140%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    136.33 +- 16%  +25043.5%      34279 +-141%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.__lookup_slow.lookup_slow.walk_component.path_lookupat.filename_lookup
     18497 +-141%    -100.0%       0.00        latency_stats.max.call_rwsem_down_write_failed_killable.vm_mmap_pgoff.ksys_mmap_pgoff.do_syscall_64.entry_SYSCALL_64_after_hwframe
    140500 +-131%     -99.8%     247.00 +- 78%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
      5403 +-139%     -97.5%     137.67 +- 71%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
     87.33 +-  5%  +23963.0%      21015 +-140%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_access.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat.filename_lookup
    136.33 +- 16%  +25043.5%      34279 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.__lookup_slow.lookup_slow.walk_component.path_lookupat.filename_lookup
    149.33 +- 14%  +25485.9%      38208 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
     18761 +-141%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_write_failed_killable.vm_mmap_pgoff.ksys_mmap_pgoff.do_syscall_64.entry_SYSCALL_64_after_hwframe
     23363 +-114%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_read_failed.__do_page_fault.do_page_fault.page_fault.__get_user_8.exit_robust_list.mm_release.do_exit.do_group_exit.get_signal.do_signal.exit_to_usermode_loop
    144810 +-125%     -99.8%     326.67 +- 70%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_do_create.nfs3_proc_create.nfs_create.path_openat.do_filp_open.do_sys_open.do_syscall_64
      5403 +-139%     -97.5%     137.67 +- 71%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.path_openat.do_filp_open.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
     59698 +- 98%     -78.0%      13110 +-141%  latency_stats.sum.call_rwsem_down_read_failed.do_exit.do_group_exit.get_signal.do_signal.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
    166.33        +12768.5%      21404 +-140%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_access.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat.filename_lookup
    825.00 +-  6%  +18761.7%     155609 +-140%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    136.33 +- 16%  +25043.5%      34279 +-141%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup.__lookup_slow.lookup_slow.walk_component.path_lookupat.filename_lookup
     59412 +-  2%     +31.3%      78021 +- 16%  numa-vmstat.node0.nr_file_pages
      8059 +-  2%     +10.7%       8923 +-  7%  numa-vmstat.node0.nr_kernel_stack
      1701 +- 25%     +49.1%       2536 +- 35%  numa-vmstat.node0.nr_mapped
    402.33 +- 21%     +70.0%     684.00 +- 19%  numa-vmstat.node0.nr_page_table_pages
      2692 +- 65%    +679.5%      20988 +- 59%  numa-vmstat.node0.nr_shmem
    622587 +- 36%     +37.7%     857545 +- 13%  numa-vmstat.node0.numa_local
     40797 +- 15%     -36.9%      25757 +- 32%  numa-vmstat.node1.nr_active_anon
    432.00 +-  4%     +33.9%     578.33 +- 14%  numa-vmstat.node1.nr_page_table_pages
     13944 +- 19%     +32.5%      18477 +-  8%  numa-vmstat.node1.nr_slab_unreclaimable
     40797 +- 15%     -36.9%      25757 +- 32%  numa-vmstat.node1.nr_zone_active_anon
    625073 +- 26%     +29.4%     808657 +- 18%  numa-vmstat.node1.numa_hit
    503969 +- 34%     +39.2%     701446 +- 23%  numa-vmstat.node1.numa_local
    137.33 +- 40%     -49.0%      70.00 +- 29%  numa-vmstat.node2.nr_mlock
    667.67 +- 17%     -45.1%     366.33 +- 15%  numa-vmstat.node2.nr_page_table_pages
     15384 +- 13%     -17.7%      12662 +-  3%  numa-vmstat.node2.nr_slab_unreclaimable
     12114 +- 94%    +150.3%      30326 +- 11%  numa-vmstat.node3.nr_active_anon
      2887 +- 83%     -93.4%     190.00 +- 82%  numa-vmstat.node3.nr_inactive_anon
      2632 +- 30%     -39.2%       1600 +-  5%  numa-vmstat.node3.nr_mapped
    101.00           -30.0%      70.67 +- 29%  numa-vmstat.node3.nr_mlock
      8425 +- 24%     -44.2%       4697 +- 15%  numa-vmstat.node3.nr_slab_reclaimable
     12122 +- 94%    +150.3%      30346 +- 11%  numa-vmstat.node3.nr_zone_active_anon
      2887 +- 83%     -93.4%     190.00 +- 82%  numa-vmstat.node3.nr_zone_inactive_anon
    106945 +- 13%     +17.4%     125554        numa-vmstat.node3.numa_other
      4.17            -0.3        3.82        perf-profile.calltrace.cycles-pp.kmem_cache_alloc.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     15.02            -0.3       14.77        perf-profile.calltrace.cycles-pp.syscall_return_via_sysret.brk
      2.42            -0.2        2.18        perf-profile.calltrace.cycles-pp.vma_compute_subtree_gap.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk
      7.60            -0.2        7.39        perf-profile.calltrace.cycles-pp.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      7.79            -0.2        7.63        perf-profile.calltrace.cycles-pp.unmap_vmas.unmap_region.do_munmap.__x64_sys_brk.do_syscall_64
      0.82 +-  9%      -0.1        0.68        perf-profile.calltrace.cycles-pp.__vm_enough_memory.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      2.13            -0.1        2.00        perf-profile.calltrace.cycles-pp.vma_compute_subtree_gap.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.05            -0.1        0.95        perf-profile.calltrace.cycles-pp.kmem_cache_free.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64
      7.31            -0.1        7.21        perf-profile.calltrace.cycles-pp.unmap_page_range.unmap_vmas.unmap_region.do_munmap.__x64_sys_brk
      0.74            -0.1        0.67        perf-profile.calltrace.cycles-pp.sync_mm_rss.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      1.06            -0.1        1.00        perf-profile.calltrace.cycles-pp.memcpy_erms.strlcpy.perf_event_mmap.do_brk_flags.__x64_sys_brk
      3.38            -0.1        3.33        perf-profile.calltrace.cycles-pp.get_unmapped_area.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.05            -0.0        1.00 +-  2%  perf-profile.calltrace.cycles-pp.__indirect_thunk_start.brk
      2.34            -0.0        2.29        perf-profile.calltrace.cycles-pp.perf_iterate_sb.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64
      1.64            -0.0        1.59        perf-profile.calltrace.cycles-pp.strlcpy.perf_event_mmap.do_brk_flags.__x64_sys_brk.do_syscall_64
      1.89            -0.0        1.86        perf-profile.calltrace.cycles-pp.security_mmap_addr.get_unmapped_area.do_brk_flags.__x64_sys_brk.do_syscall_64
      0.76            -0.0        0.73        perf-profile.calltrace.cycles-pp._raw_spin_lock.unmap_page_range.unmap_vmas.unmap_region.do_munmap
      0.57 +-  2%      -0.0        0.55        perf-profile.calltrace.cycles-pp.selinux_mmap_addr.security_mmap_addr.get_unmapped_area.do_brk_flags.__x64_sys_brk
      0.54 +-  2%      +0.0        0.56        perf-profile.calltrace.cycles-pp.do_brk_flags.brk
      0.72            +0.0        0.76 +-  2%  perf-profile.calltrace.cycles-pp.do_munmap.brk
      4.38            +0.1        4.43        perf-profile.calltrace.cycles-pp.find_vma.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      1.96            +0.1        2.04        perf-profile.calltrace.cycles-pp.vmacache_find.find_vma.do_munmap.__x64_sys_brk.do_syscall_64
      0.53            +0.2        0.68        perf-profile.calltrace.cycles-pp.__vma_link_rb.brk
      2.21            +0.3        2.51        perf-profile.calltrace.cycles-pp.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     64.44            +0.5       64.90        perf-profile.calltrace.cycles-pp.entry_SYSCALL_64_after_hwframe.brk
     63.04            +0.5       63.54        perf-profile.calltrace.cycles-pp.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
     60.37            +0.5       60.88        perf-profile.calltrace.cycles-pp.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      3.75            +0.5        4.29        perf-profile.calltrace.cycles-pp.vma_link.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00            +0.6        0.57        perf-profile.calltrace.cycles-pp.free_pgtables.unmap_region.do_munmap.__x64_sys_brk.do_syscall_64
      0.00            +0.6        0.64        perf-profile.calltrace.cycles-pp.put_vma.remove_vma.do_munmap.__x64_sys_brk.do_syscall_64
      0.72            +0.7        1.37        perf-profile.calltrace.cycles-pp.__vma_rb_erase.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
     24.42            +0.7       25.08        perf-profile.calltrace.cycles-pp.do_munmap.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe.brk
      0.00            +0.7        0.71        perf-profile.calltrace.cycles-pp._raw_write_lock.__vma_rb_erase.do_munmap.__x64_sys_brk.do_syscall_64
      3.12            +0.7        3.84        perf-profile.calltrace.cycles-pp.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk.do_syscall_64
      0.00            +0.8        0.77        perf-profile.calltrace.cycles-pp._raw_write_lock.__vma_link_rb.vma_link.do_brk_flags.__x64_sys_brk
      0.00            +0.9        0.85        perf-profile.calltrace.cycles-pp.__vma_merge.do_brk_flags.__x64_sys_brk.do_syscall_64.entry_SYSCALL_64_after_hwframe
      5.10            -0.5        4.60        perf-profile.children.cycles-pp.vma_compute_subtree_gap
      4.53            -0.3        4.18        perf-profile.children.cycles-pp.kmem_cache_alloc
     15.03            -0.3       14.77        perf-profile.children.cycles-pp.syscall_return_via_sysret
      8.13            -0.2        7.92        perf-profile.children.cycles-pp.perf_event_mmap
      8.01            -0.2        7.81        perf-profile.children.cycles-pp.unmap_vmas
      0.97 +- 14%      -0.2        0.78        perf-profile.children.cycles-pp.__vm_enough_memory
      1.13            -0.1        1.00        perf-profile.children.cycles-pp.kmem_cache_free
      7.82            -0.1        7.70        perf-profile.children.cycles-pp.unmap_page_range
     12.23            -0.1       12.13        perf-profile.children.cycles-pp.unmap_region
      0.74            -0.1        0.67        perf-profile.children.cycles-pp.sync_mm_rss
      3.06            -0.1        3.00        perf-profile.children.cycles-pp.down_write_killable
      0.40 +-  2%      -0.1        0.34        perf-profile.children.cycles-pp.__rb_insert_augmented
      1.29            -0.1        1.23        perf-profile.children.cycles-pp.__indirect_thunk_start
      2.54            -0.1        2.49        perf-profile.children.cycles-pp.perf_iterate_sb
      3.66            -0.0        3.61        perf-profile.children.cycles-pp.get_unmapped_area
      1.80            -0.0        1.75        perf-profile.children.cycles-pp.strlcpy
      0.53 +-  2%      -0.0        0.49 +-  2%  perf-profile.children.cycles-pp.cap_capable
      1.57            -0.0        1.53        perf-profile.children.cycles-pp.arch_get_unmapped_area_topdown
      1.11            -0.0        1.08        perf-profile.children.cycles-pp.memcpy_erms
      0.13            -0.0        0.10        perf-profile.children.cycles-pp.__vma_link_file
      0.55            -0.0        0.52        perf-profile.children.cycles-pp.unmap_single_vma
      1.47            -0.0        1.44        perf-profile.children.cycles-pp.cap_vm_enough_memory
      2.14            -0.0        2.12        perf-profile.children.cycles-pp.security_mmap_addr
      0.32            -0.0        0.30        perf-profile.children.cycles-pp.userfaultfd_unmap_complete
      1.25            -0.0        1.23        perf-profile.children.cycles-pp.up_write
      0.50            -0.0        0.49        perf-profile.children.cycles-pp.userfaultfd_unmap_prep
      0.27            -0.0        0.26        perf-profile.children.cycles-pp.tlb_flush_mmu_free
      1.14            -0.0        1.12        perf-profile.children.cycles-pp.__might_sleep
      0.07            -0.0        0.06        perf-profile.children.cycles-pp.should_failslab
      0.72            +0.0        0.74        perf-profile.children.cycles-pp._cond_resched
      0.45            +0.0        0.47        perf-profile.children.cycles-pp.rcu_all_qs
      0.15 +-  3%      +0.0        0.17 +-  4%  perf-profile.children.cycles-pp.__vma_link_list
      0.15 +-  5%      +0.0        0.18 +-  5%  perf-profile.children.cycles-pp.tick_sched_timer
      0.05 +-  8%      +0.1        0.12 +- 17%  perf-profile.children.cycles-pp.perf_mux_hrtimer_handler
      0.80            +0.1        0.89        perf-profile.children.cycles-pp.free_pgtables
      0.22 +-  7%      +0.1        0.31 +-  9%  perf-profile.children.cycles-pp.__hrtimer_run_queues
      0.00            +0.1        0.11 +- 15%  perf-profile.children.cycles-pp.clockevents_program_event
      6.34            +0.1        6.47        perf-profile.children.cycles-pp.find_vma
      2.27            +0.1        2.40        perf-profile.children.cycles-pp.vmacache_find
      0.40 +-  4%      +0.2        0.58 +-  5%  perf-profile.children.cycles-pp.apic_timer_interrupt
      0.40 +-  4%      +0.2        0.58 +-  5%  perf-profile.children.cycles-pp.smp_apic_timer_interrupt
      0.37 +-  4%      +0.2        0.54 +-  5%  perf-profile.children.cycles-pp.hrtimer_interrupt
      0.00            +0.2        0.19 +- 12%  perf-profile.children.cycles-pp.ktime_get
      2.42            +0.3        2.77        perf-profile.children.cycles-pp.remove_vma
     64.49            +0.5       64.94        perf-profile.children.cycles-pp.entry_SYSCALL_64_after_hwframe
      1.27            +0.5        1.73        perf-profile.children.cycles-pp.__vma_rb_erase
     61.62            +0.5       62.10        perf-profile.children.cycles-pp.__x64_sys_brk
     63.24            +0.5       63.74        perf-profile.children.cycles-pp.do_syscall_64
      4.03            +0.5        4.56        perf-profile.children.cycles-pp.vma_link
      0.00            +0.7        0.69        perf-profile.children.cycles-pp.put_vma
     25.13            +0.7       25.84        perf-profile.children.cycles-pp.do_munmap
      3.83            +0.7        4.56        perf-profile.children.cycles-pp.__vma_link_rb
      0.00            +1.2        1.25        perf-profile.children.cycles-pp.__vma_merge
      0.00            +1.5        1.53        perf-profile.children.cycles-pp._raw_write_lock
      5.08            -0.5        4.58        perf-profile.self.cycles-pp.vma_compute_subtree_gap
     15.03            -0.3       14.77        perf-profile.self.cycles-pp.syscall_return_via_sysret
      0.59            -0.2        0.39        perf-profile.self.cycles-pp.remove_vma
      0.72 +-  7%      -0.1        0.58        perf-profile.self.cycles-pp.__vm_enough_memory
      1.12            -0.1        0.99        perf-profile.self.cycles-pp.kmem_cache_free
      3.11            -0.1        2.99        perf-profile.self.cycles-pp.do_munmap
      0.99            -0.1        0.88        perf-profile.self.cycles-pp.__vma_rb_erase
      3.63            -0.1        3.52        perf-profile.self.cycles-pp.perf_event_mmap
      3.26            -0.1        3.17        perf-profile.self.cycles-pp.brk
      0.41 +-  2%      -0.1        0.33        perf-profile.self.cycles-pp.unmap_vmas
      0.74            -0.1        0.67        perf-profile.self.cycles-pp.sync_mm_rss
      1.75            -0.1        1.68        perf-profile.self.cycles-pp.kmem_cache_alloc
      0.40 +-  2%      -0.1        0.34        perf-profile.self.cycles-pp.__rb_insert_augmented
      1.29 +-  2%      -0.1        1.23        perf-profile.self.cycles-pp.__indirect_thunk_start
      0.73            -0.0        0.68 +-  2%  perf-profile.self.cycles-pp.unmap_region
      0.53            -0.0        0.49        perf-profile.self.cycles-pp.vma_link
      1.40            -0.0        1.35        perf-profile.self.cycles-pp.entry_SYSCALL_64_after_hwframe
      5.22            -0.0        5.18        perf-profile.self.cycles-pp.unmap_page_range
      0.53 +-  2%      -0.0        0.49 +-  2%  perf-profile.self.cycles-pp.cap_capable
      1.11            -0.0        1.07        perf-profile.self.cycles-pp.memcpy_erms
      1.86            -0.0        1.82        perf-profile.self.cycles-pp.perf_iterate_sb
      1.30            -0.0        1.27        perf-profile.self.cycles-pp.arch_get_unmapped_area_topdown
      0.13            -0.0        0.10        perf-profile.self.cycles-pp.__vma_link_file
      0.55            -0.0        0.52        perf-profile.self.cycles-pp.unmap_single_vma
      0.74            -0.0        0.72        perf-profile.self.cycles-pp.selinux_mmap_addr
      0.32            -0.0        0.30        perf-profile.self.cycles-pp.userfaultfd_unmap_complete
      1.13            -0.0        1.12        perf-profile.self.cycles-pp.__might_sleep
      1.24            -0.0        1.23        perf-profile.self.cycles-pp.up_write
      0.50            -0.0        0.49        perf-profile.self.cycles-pp.userfaultfd_unmap_prep
      0.27            -0.0        0.26        perf-profile.self.cycles-pp.tlb_flush_mmu_free
      0.07            -0.0        0.06        perf-profile.self.cycles-pp.should_failslab
      0.45            +0.0        0.47        perf-profile.self.cycles-pp.rcu_all_qs
      0.71            +0.0        0.73        perf-profile.self.cycles-pp.strlcpy
      0.15 +-  3%      +0.0        0.17 +-  4%  perf-profile.self.cycles-pp.__vma_link_list
      0.51            +0.1        0.57        perf-profile.self.cycles-pp.free_pgtables
      1.40            +0.1        1.49        perf-profile.self.cycles-pp.__vma_link_rb
      2.27            +0.1        2.39        perf-profile.self.cycles-pp.vmacache_find
      0.00            +0.2        0.18 +- 12%  perf-profile.self.cycles-pp.ktime_get
      0.00            +0.7        0.69        perf-profile.self.cycles-pp.put_vma
      0.00            +1.2        1.24        perf-profile.self.cycles-pp.__vma_merge
      0.00            +1.5        1.52        perf-profile.self.cycles-pp._raw_write_lock

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/always/page_fault2/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
           :3           33%           1:3     dmesg.WARNING:at#for_ip_native_iret/0x
          1:3          -33%            :3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=__schedule/0x
           :3           33%           1:3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=__slab_free/0x
          1:3          -33%            :3     kmsg.DHCP/BOOTP:Reply_not_for_us_on_eth#,op[#]xid[#]
          3:3         -100%            :3     kmsg.pstore:crypto_comp_decompress_failed,ret=
          3:3         -100%            :3     kmsg.pstore:decompression_failed
          2:3            4%           2:3     perf-profile.calltrace.cycles-pp.sync_regs.error_entry
          5:3            7%           5:3     perf-profile.calltrace.cycles-pp.error_entry
          5:3            7%           5:3     perf-profile.children.cycles-pp.error_entry
          2:3            3%           2:3     perf-profile.self.cycles-pp.error_entry
         %stddev     %change         %stddev
             \          |                \  
      8281 +-  2%     -18.8%       6728        will-it-scale.per_thread_ops
     92778 +-  2%     +17.6%     109080        will-it-scale.time.involuntary_context_switches
  21954366 +-  3%      +4.1%   22857988 +-  2%  will-it-scale.time.maximum_resident_set_size
  4.81e+08 +-  2%     -18.9%  3.899e+08        will-it-scale.time.minor_page_faults
      5804           +12.2%       6512        will-it-scale.time.percent_of_cpu_this_job_got
     34918           +12.2%      39193        will-it-scale.time.system_time
   5638528 +-  2%     -15.3%    4778392        will-it-scale.time.voluntary_context_switches
  15846405            -2.0%   15531034        will-it-scale.workload
   2818137            +1.5%    2861500        interrupts.CAL:Function_call_interrupts
      3.33 +- 28%     -60.0%       1.33 +- 93%  irq_exception_noise.irq_time
      2866           +23.9%       3552 +-  2%  kthread_noise.total_time
   5589674 +- 14%     +31.4%    7344810 +-  6%  meminfo.DirectMap2M
     31169           -16.9%      25906        uptime.idle
     25242 +-  4%     -14.2%      21654 +-  6%  vmstat.system.cs
      7055           -11.6%       6237        boot-time.idle
     21.12           +19.3%      25.19 +-  9%  boot-time.kernel_boot
     20.03 +-  2%      -3.7       16.38        mpstat.cpu.idle%
      0.00 +-  8%      -0.0        0.00 +-  4%  mpstat.cpu.iowait%
   7284147 +-  2%     -16.4%    6092495        softirqs.RCU
   5350756 +-  2%     -10.9%    4769417 +-  4%  softirqs.SCHED
     42933 +- 21%     -28.2%      30807 +-  7%  numa-meminfo.node2.SReclaimable
     63219 +- 13%     -16.6%      52717 +-  6%  numa-meminfo.node2.SUnreclaim
    106153 +- 16%     -21.3%      83525 +-  5%  numa-meminfo.node2.Slab
    247154 +-  4%      -7.6%     228415        numa-meminfo.node3.Unevictable
     11904 +-  4%     +17.1%      13945 +-  8%  numa-vmstat.node0
      2239 +- 22%     -26.6%       1644 +-  2%  numa-vmstat.node2.nr_mapped
     10728 +- 21%     -28.2%       7701 +-  7%  numa-vmstat.node2.nr_slab_reclaimable
     15803 +- 13%     -16.6%      13179 +-  6%  numa-vmstat.node2.nr_slab_unreclaimable
     61788 +-  4%      -7.6%      57103        numa-vmstat.node3.nr_unevictable
     61788 +-  4%      -7.6%      57103        numa-vmstat.node3.nr_zone_unevictable
     92778 +-  2%     +17.6%     109080        time.involuntary_context_switches
  21954366 +-  3%      +4.1%   22857988 +-  2%  time.maximum_resident_set_size
  4.81e+08 +-  2%     -18.9%  3.899e+08        time.minor_page_faults
      5804           +12.2%       6512        time.percent_of_cpu_this_job_got
     34918           +12.2%      39193        time.system_time
   5638528 +-  2%     -15.3%    4778392        time.voluntary_context_switches
   3942289 +-  2%     -10.5%    3528902 +-  2%  cpuidle.C1.time
    242290           -14.2%     207992        cpuidle.C1.usage
  1.64e+09 +-  2%     -15.7%  1.381e+09        cpuidle.C1E.time
   4621281 +-  2%     -14.7%    3939757        cpuidle.C1E.usage
 2.115e+10 +-  2%     -18.5%  1.723e+10        cpuidle.C6.time
  24771099 +-  2%     -18.0%   20305766        cpuidle.C6.usage
   1210810 +-  4%     -17.6%     997270 +-  2%  cpuidle.POLL.time
     18742 +-  3%     -17.0%      15559 +-  2%  cpuidle.POLL.usage
      4135 +-141%    -100.0%       0.00        latency_stats.avg.x86_reserve_hardware.x86_pmu_event_init.perf_try_init_event.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
     33249 +-129%    -100.0%       0.00        latency_stats.max.call_rwsem_down_read_failed.m_start.seq_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4135 +-141%    -100.0%       0.00        latency_stats.max.x86_reserve_hardware.x86_pmu_event_init.perf_try_init_event.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
     65839 +-116%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_read_failed.m_start.seq_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4135 +-141%    -100.0%       0.00        latency_stats.sum.x86_reserve_hardware.x86_pmu_event_init.perf_try_init_event.perf_event_alloc.__do_sys_perf_event_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      8387 +-122%     -90.9%     767.00 +- 13%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    263970 +- 10%     -68.6%      82994 +-  3%  latency_stats.sum.do_syslog.kmsg_read.proc_reg_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      6173 +- 77%    +173.3%      16869 +- 98%  latency_stats.sum.pipe_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
    101.33            -4.6%      96.67        proc-vmstat.nr_anon_transparent_hugepages
     39967            -1.8%      39241        proc-vmstat.nr_slab_reclaimable
     67166            -2.4%      65522        proc-vmstat.nr_slab_unreclaimable
    237743            -3.9%     228396        proc-vmstat.nr_unevictable
    237743            -3.9%     228396        proc-vmstat.nr_zone_unevictable
 4.807e+09            -2.0%   4.71e+09        proc-vmstat.numa_hit
 4.807e+09            -2.0%   4.71e+09        proc-vmstat.numa_local
 4.791e+09            -2.1%   4.69e+09        proc-vmstat.pgalloc_normal
 4.783e+09            -2.0%  4.685e+09        proc-vmstat.pgfault
 4.807e+09            -2.0%  4.709e+09        proc-vmstat.pgfree
      1753            +4.6%       1833        turbostat.Avg_MHz
    239445           -14.1%     205783        turbostat.C1
   4617105 +-  2%     -14.8%    3934693        turbostat.C1E
      1.40 +-  2%      -0.2        1.18        turbostat.C1E%
  24764661 +-  2%     -18.0%   20297643        turbostat.C6
     18.09 +-  2%      -3.4       14.74        turbostat.C6%
      7.53 +-  2%     -17.1%       6.24        turbostat.CPU%c1
     11.88 +-  2%     -19.1%       9.61        turbostat.CPU%c6
      7.62 +-  3%     -20.8%       6.04        turbostat.Pkg%pc2
    388.30            +1.5%     393.93        turbostat.PkgWatt
    390974 +-  8%     +35.8%     530867 +- 11%  sched_debug.cfs_rq:/.min_vruntime.stddev
  -1754042           +75.7%   -3081270        sched_debug.cfs_rq:/.spread0.min
    388140 +-  8%     +36.2%     528494 +- 11%  sched_debug.cfs_rq:/.spread0.stddev
    542.30 +-  3%     -10.0%     488.21 +-  3%  sched_debug.cfs_rq:/.util_avg.min
     53.35 +- 16%     +48.7%      79.35 +- 12%  sched_debug.cfs_rq:/.util_est_enqueued.avg
     30520 +-  6%     -15.2%      25883 +- 12%  sched_debug.cpu.nr_switches.avg
    473770 +- 27%     -37.4%     296623 +- 32%  sched_debug.cpu.nr_switches.max
     17077 +-  2%     -15.1%      14493        sched_debug.cpu.nr_switches.min
     30138 +-  6%     -15.0%      25606 +- 12%  sched_debug.cpu.sched_count.avg
    472345 +- 27%     -37.2%     296419 +- 32%  sched_debug.cpu.sched_count.max
     16858 +-  2%     -15.2%      14299        sched_debug.cpu.sched_count.min
      8358 +-  2%     -15.5%       7063        sched_debug.cpu.sched_goidle.avg
     12225           -13.6%      10565        sched_debug.cpu.sched_goidle.max
      8032 +-  2%     -16.0%       6749        sched_debug.cpu.sched_goidle.min
     14839 +-  6%     -15.3%      12568 +- 12%  sched_debug.cpu.ttwu_count.avg
    235115 +- 28%     -38.3%     145175 +- 31%  sched_debug.cpu.ttwu_count.max
      7627 +-  3%     -15.9%       6413 +-  2%  sched_debug.cpu.ttwu_count.min
    226299 +- 29%     -39.5%     136827 +- 32%  sched_debug.cpu.ttwu_local.max
      0.85            -0.0        0.81        perf-stat.branch-miss-rate%
 3.675e+10            -4.1%  3.523e+10        perf-stat.branch-misses
 4.052e+11            -2.3%  3.958e+11        perf-stat.cache-misses
 7.008e+11            -2.5%  6.832e+11        perf-stat.cache-references
  15320995 +-  4%     -14.3%   13136557 +-  6%  perf-stat.context-switches
      9.16            +4.8%       9.59        perf-stat.cpi
  2.03e+14            +4.6%  2.124e+14        perf-stat.cpu-cycles
     44508            -1.7%      43743        perf-stat.cpu-migrations
      1.30            -0.1        1.24        perf-stat.dTLB-store-miss-rate%
 4.064e+10            -3.5%  3.922e+10        perf-stat.dTLB-store-misses
 3.086e+12            +1.1%  3.119e+12        perf-stat.dTLB-stores
 3.611e+08 +-  6%      -8.5%  3.304e+08 +-  5%  perf-stat.iTLB-loads
      0.11            -4.6%       0.10        perf-stat.ipc
 4.783e+09            -2.0%  4.685e+09        perf-stat.minor-faults
      1.53 +-  2%      -0.3        1.22 +-  8%  perf-stat.node-load-miss-rate%
 1.389e+09 +-  3%     -22.1%  1.083e+09 +-  9%  perf-stat.node-load-misses
 8.922e+10            -1.9%   8.75e+10        perf-stat.node-loads
      5.06            +1.7        6.77 +-  3%  perf-stat.node-store-miss-rate%
 1.204e+09           +29.3%  1.556e+09 +-  3%  perf-stat.node-store-misses
 2.256e+10            -5.1%  2.142e+10 +-  2%  perf-stat.node-stores
 4.783e+09            -2.0%  4.685e+09        perf-stat.page-faults
   1399242            +1.9%    1425404        perf-stat.path-length
      1144 +-  8%     -13.6%     988.00 +-  8%  slabinfo.Acpi-ParseExt.active_objs
      1144 +-  8%     -13.6%     988.00 +-  8%  slabinfo.Acpi-ParseExt.num_objs
      1878 +- 17%     +29.0%       2422 +- 16%  slabinfo.dmaengine-unmap-16.active_objs
      1878 +- 17%     +29.0%       2422 +- 16%  slabinfo.dmaengine-unmap-16.num_objs
      1085 +-  5%     -24.1%     823.33 +-  9%  slabinfo.file_lock_cache.active_objs
      1085 +-  5%     -24.1%     823.33 +-  9%  slabinfo.file_lock_cache.num_objs
     61584 +-  4%     -16.6%      51381 +-  5%  slabinfo.filp.active_objs
    967.00 +-  4%     -16.5%     807.67 +-  5%  slabinfo.filp.active_slabs
     61908 +-  4%     -16.5%      51713 +-  5%  slabinfo.filp.num_objs
    967.00 +-  4%     -16.5%     807.67 +-  5%  slabinfo.filp.num_slabs
      1455           -15.4%       1232 +-  4%  slabinfo.nsproxy.active_objs
      1455           -15.4%       1232 +-  4%  slabinfo.nsproxy.num_objs
     84720 +-  6%     -18.3%      69210 +-  4%  slabinfo.pid.active_objs
      1324 +-  6%     -18.2%       1083 +-  4%  slabinfo.pid.active_slabs
     84820 +-  5%     -18.2%      69386 +-  4%  slabinfo.pid.num_objs
      1324 +-  6%     -18.2%       1083 +-  4%  slabinfo.pid.num_slabs
      2112 +- 18%     -26.3%       1557 +-  5%  slabinfo.scsi_sense_cache.active_objs
      2112 +- 18%     -26.3%       1557 +-  5%  slabinfo.scsi_sense_cache.num_objs
      5018 +-  5%      -7.6%       4635 +-  4%  slabinfo.sock_inode_cache.active_objs
      5018 +-  5%      -7.6%       4635 +-  4%  slabinfo.sock_inode_cache.num_objs
      1193 +-  4%     +13.8%       1358 +-  4%  slabinfo.task_group.active_objs
      1193 +-  4%     +13.8%       1358 +-  4%  slabinfo.task_group.num_objs
     62807 +-  3%     -14.4%      53757 +-  3%  slabinfo.vm_area_struct.active_objs
      1571 +-  3%     -12.1%       1381 +-  3%  slabinfo.vm_area_struct.active_slabs
     62877 +-  3%     -14.3%      53880 +-  3%  slabinfo.vm_area_struct.num_objs
      1571 +-  3%     -12.1%       1381 +-  3%  slabinfo.vm_area_struct.num_slabs
     47.45           -47.4        0.00        perf-profile.calltrace.cycles-pp.alloc_pages_vma.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
     47.16           -47.2        0.00        perf-profile.calltrace.cycles-pp.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault.handle_mm_fault.__do_page_fault
     46.99           -47.0        0.00        perf-profile.calltrace.cycles-pp.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault.handle_mm_fault
     44.95           -44.9        0.00        perf-profile.calltrace.cycles-pp._raw_spin_lock.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault
      7.42 +-  2%      -7.4        0.00        perf-profile.calltrace.cycles-pp.copy_page.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      6.32 +- 10%      -6.3        0.00        perf-profile.calltrace.cycles-pp.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      6.28 +- 10%      -6.3        0.00        perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +0.9        0.85 +- 11%  perf-profile.calltrace.cycles-pp._raw_spin_lock.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +0.9        0.92 +-  4%  perf-profile.calltrace.cycles-pp.__list_del_entry_valid.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault
      0.00            +1.1        1.13 +-  7%  perf-profile.calltrace.cycles-pp.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault
      0.00            +1.2        1.19 +-  7%  perf-profile.calltrace.cycles-pp.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault
      0.00            +1.2        1.22 +-  5%  perf-profile.calltrace.cycles-pp.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +1.3        1.34 +-  7%  perf-profile.calltrace.cycles-pp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +1.4        1.36 +-  7%  perf-profile.calltrace.cycles-pp.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +4.5        4.54 +- 19%  perf-profile.calltrace.cycles-pp.pagevec_lru_move_fn.__lru_cache_add.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +4.6        4.64 +- 19%  perf-profile.calltrace.cycles-pp.__lru_cache_add.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +6.6        6.64 +- 15%  perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +6.7        6.68 +- 15%  perf-profile.calltrace.cycles-pp.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +7.5        7.54 +-  5%  perf-profile.calltrace.cycles-pp.copy_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00           +44.6       44.55 +-  3%  perf-profile.calltrace.cycles-pp._raw_spin_lock.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault
      0.00           +46.6       46.63 +-  3%  perf-profile.calltrace.cycles-pp.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault.__handle_mm_fault
      0.00           +46.8       46.81 +-  3%  perf-profile.calltrace.cycles-pp.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00           +47.1       47.10 +-  3%  perf-profile.calltrace.cycles-pp.alloc_pages_vma.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00           +63.1       63.15        perf-profile.calltrace.cycles-pp.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      0.39 +-  3%      +0.0        0.42 +-  3%  perf-profile.children.cycles-pp.radix_tree_lookup_slot
      0.21 +-  3%      +0.0        0.25 +-  5%  perf-profile.children.cycles-pp.__mod_node_page_state
      0.00            +0.1        0.06 +-  8%  perf-profile.children.cycles-pp.get_vma_policy
      0.00            +0.1        0.08 +-  5%  perf-profile.children.cycles-pp.__lru_cache_add_active_or_unevictable
      0.00            +0.2        0.18 +-  6%  perf-profile.children.cycles-pp.__page_add_new_anon_rmap
      0.00            +1.4        1.35 +-  5%  perf-profile.children.cycles-pp.pte_map_lock
      0.00           +63.2       63.21        perf-profile.children.cycles-pp.handle_pte_fault
      1.40 +-  2%      -0.4        1.03 +- 10%  perf-profile.self.cycles-pp._raw_spin_lock
      0.56 +-  3%      -0.2        0.35 +-  6%  perf-profile.self.cycles-pp.__handle_mm_fault
      0.22 +-  3%      -0.0        0.18 +-  7%  perf-profile.self.cycles-pp.alloc_set_pte
      0.09            +0.0        0.10 +-  4%  perf-profile.self.cycles-pp.vmacache_find
      0.39 +-  2%      +0.0        0.41 +-  3%  perf-profile.self.cycles-pp.__radix_tree_lookup
      0.18            +0.0        0.20 +-  6%  perf-profile.self.cycles-pp.mem_cgroup_charge_statistics
      0.17 +-  2%      +0.0        0.20 +-  7%  perf-profile.self.cycles-pp.___might_sleep
      0.33 +-  2%      +0.0        0.36 +-  6%  perf-profile.self.cycles-pp.handle_mm_fault
      0.20 +-  2%      +0.0        0.24 +-  3%  perf-profile.self.cycles-pp.__mod_node_page_state
      0.00            +0.1        0.05        perf-profile.self.cycles-pp.finish_fault
      0.00            +0.1        0.05        perf-profile.self.cycles-pp.get_vma_policy
      0.00            +0.1        0.08 +- 10%  perf-profile.self.cycles-pp.__lru_cache_add_active_or_unevictable
      0.00            +0.2        0.25 +-  5%  perf-profile.self.cycles-pp.handle_pte_fault
      0.00            +0.5        0.49 +-  8%  perf-profile.self.cycles-pp.pte_map_lock

=========================================================================================
tbox_group/testcase/rootfs/kconfig/compiler/nr_task/thp_enabled/test/cpufreq_governor:
  lkp-skl-4sp1/will-it-scale/debian-x86_64-2018-04-03.cgz/x86_64-rhel-7.2/gcc-7/100%/never/page_fault2/performance

commit: 
  ba98a1cdad71d259a194461b3a61471b49b14df1
  a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12

ba98a1cdad71d259 a7a8993bfe3ccb54ad468b9f17 
---------------- -------------------------- 
       fail:runs  %reproduction    fail:runs
           |             |             |    
          1:3          -33%            :3     kmsg.DHCP/BOOTP:Reply_not_for_us_on_eth#,op[#]xid[#]
           :3           33%           1:3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=sched_slice/0x
          1:3          -33%            :3     dmesg.WARNING:stack_going_in_the_wrong_direction?ip=schedule_tail/0x
          1:3           24%           2:3     perf-profile.calltrace.cycles-pp.sync_regs.error_entry
          3:3           46%           5:3     perf-profile.calltrace.cycles-pp.error_entry
          5:3           -9%           5:3     perf-profile.children.cycles-pp.error_entry
          2:3           -4%           2:3     perf-profile.self.cycles-pp.error_entry
         %stddev     %change         %stddev
             \          |                \  
      8147           -18.8%       6613        will-it-scale.per_thread_ops
     93113           +17.0%     108982        will-it-scale.time.involuntary_context_switches
 4.732e+08           -19.0%  3.833e+08        will-it-scale.time.minor_page_faults
      5854           +12.0%       6555        will-it-scale.time.percent_of_cpu_this_job_got
     35247           +12.1%      39495        will-it-scale.time.system_time
   5546661           -15.5%    4689314        will-it-scale.time.voluntary_context_switches
  15801637            -1.9%   15504487        will-it-scale.workload
      1.43 +- 11%     -59.7%       0.58 +- 28%  irq_exception_noise.__do_page_fault.min
      2811 +-  3%     +23.7%       3477 +-  3%  kthread_noise.total_time
    292776 +-  5%     +39.6%     408829 +- 21%  meminfo.DirectMap4k
     19.80            -3.7       16.12        mpstat.cpu.idle%
     29940           -14.5%      25593        uptime.idle
     24064 +-  3%      -8.5%      22016        vmstat.system.cs
     34.86            -1.9%      34.19        boot-time.boot
     26.95            -2.8%      26.19 +-  2%  boot-time.kernel_boot
   7190569 +-  2%     -15.2%    6100136 +-  3%  softirqs.RCU
   5513663           -13.8%    4751548        softirqs.SCHED
     18064 +-  2%     +24.3%      22461 +-  7%  numa-vmstat.node0.nr_slab_unreclaimable
      8507 +- 12%     -16.8%       7075 +-  4%  numa-vmstat.node2.nr_slab_reclaimable
     18719 +-  9%     -19.6%      15043 +-  4%  numa-vmstat.node3.nr_slab_unreclaimable
     72265 +-  2%     +24.3%      89855 +-  7%  numa-meminfo.node0.SUnreclaim
    115980 +-  4%     +22.6%     142233 +- 12%  numa-meminfo.node0.Slab
     34035 +- 12%     -16.8%      28307 +-  4%  numa-meminfo.node2.SReclaimable
     74888 +-  9%     -19.7%      60162 +-  4%  numa-meminfo.node3.SUnreclaim
     93113           +17.0%     108982        time.involuntary_context_switches
 4.732e+08           -19.0%  3.833e+08        time.minor_page_faults
      5854           +12.0%       6555        time.percent_of_cpu_this_job_got
     35247           +12.1%      39495        time.system_time
   5546661           -15.5%    4689314        time.voluntary_context_switches
 4.792e+09            -1.9%  4.699e+09        proc-vmstat.numa_hit
 4.791e+09            -1.9%  4.699e+09        proc-vmstat.numa_local
     40447 +- 11%     +13.2%      45804 +-  6%  proc-vmstat.pgactivate
 4.778e+09            -1.9%  4.688e+09        proc-vmstat.pgalloc_normal
 4.767e+09            -1.9%  4.675e+09        proc-vmstat.pgfault
 4.791e+09            -1.9%  4.699e+09        proc-vmstat.pgfree
    230178 +-  2%     -10.1%     206883 +-  3%  cpuidle.C1.usage
 1.617e+09           -15.0%  1.375e+09        cpuidle.C1E.time
   4514401           -14.1%    3878206        cpuidle.C1E.usage
 2.087e+10           -18.5%  1.701e+10        cpuidle.C6.time
  24458365           -18.0%   20045336        cpuidle.C6.usage
   1163758           -16.1%     976094 +-  4%  cpuidle.POLL.time
     17907           -14.6%      15294 +-  4%  cpuidle.POLL.usage
      1758            +4.5%       1838        turbostat.Avg_MHz
    227522 +-  2%     -10.2%     204426 +-  3%  turbostat.C1
   4512700           -14.2%    3873264        turbostat.C1E
      1.39            -0.2        1.18        turbostat.C1E%
  24452583           -18.0%   20039031        turbostat.C6
     17.85            -3.3       14.55        turbostat.C6%
      7.44           -16.8%       6.19        turbostat.CPU%c1
     11.72           -19.3%       9.45        turbostat.CPU%c6
      7.51           -21.3%       5.91        turbostat.Pkg%pc2
    389.33            +1.6%     395.59        turbostat.PkgWatt
    559.33 +- 13%     -17.9%     459.33 +- 20%  slabinfo.dmaengine-unmap-128.active_objs
    559.33 +- 13%     -17.9%     459.33 +- 20%  slabinfo.dmaengine-unmap-128.num_objs
     57734 +-  3%      -5.7%      54421 +-  4%  slabinfo.filp.active_objs
    905.67 +-  3%      -5.6%     854.67 +-  4%  slabinfo.filp.active_slabs
     57981 +-  3%      -5.6%      54720 +-  4%  slabinfo.filp.num_objs
    905.67 +-  3%      -5.6%     854.67 +-  4%  slabinfo.filp.num_slabs
      1378           -12.0%       1212 +-  7%  slabinfo.nsproxy.active_objs
      1378           -12.0%       1212 +-  7%  slabinfo.nsproxy.num_objs
    507.33 +-  7%     -26.8%     371.33 +-  2%  slabinfo.secpath_cache.active_objs
    507.33 +-  7%     -26.8%     371.33 +-  2%  slabinfo.secpath_cache.num_objs
      4788 +-  5%      -8.3%       4391 +-  2%  slabinfo.sock_inode_cache.active_objs
      4788 +-  5%      -8.3%       4391 +-  2%  slabinfo.sock_inode_cache.num_objs
      1431 +-  8%     -12.3%       1255 +-  3%  slabinfo.task_group.active_objs
      1431 +-  8%     -12.3%       1255 +-  3%  slabinfo.task_group.num_objs
      4.27 +- 17%     +27.0%       5.42 +-  7%  sched_debug.cfs_rq:/.runnable_load_avg.avg
     13.44 +- 62%     +73.6%      23.33 +- 24%  sched_debug.cfs_rq:/.runnable_load_avg.stddev
    772.55 +- 21%     -32.7%     520.27 +-  4%  sched_debug.cfs_rq:/.util_est_enqueued.max
      4.39 +- 15%     +29.0%       5.66 +- 11%  sched_debug.cpu.cpu_load[0].avg
    152.09 +- 72%     +83.9%     279.67 +- 33%  sched_debug.cpu.cpu_load[0].max
     13.84 +- 58%     +78.7%      24.72 +- 29%  sched_debug.cpu.cpu_load[0].stddev
      4.53 +- 14%     +25.8%       5.70 +- 10%  sched_debug.cpu.cpu_load[1].avg
    156.58 +- 66%     +76.6%     276.58 +- 33%  sched_debug.cpu.cpu_load[1].max
     14.02 +- 55%     +72.4%      24.17 +- 28%  sched_debug.cpu.cpu_load[1].stddev
      4.87 +- 11%     +17.3%       5.72 +-  9%  sched_debug.cpu.cpu_load[2].avg
      1.58 +-  2%     +13.5%       1.79 +-  6%  sched_debug.cpu.nr_running.max
     16694           -14.6%      14259        sched_debug.cpu.nr_switches.min
     31989 +- 13%     +20.6%      38584 +-  6%  sched_debug.cpu.nr_switches.stddev
     16505           -14.8%      14068        sched_debug.cpu.sched_count.min
     32084 +- 13%     +19.9%      38482 +-  6%  sched_debug.cpu.sched_count.stddev
      8185           -15.0%       6957        sched_debug.cpu.sched_goidle.avg
     12151 +-  2%     -13.5%      10507        sched_debug.cpu.sched_goidle.max
      7867           -15.7%       6631        sched_debug.cpu.sched_goidle.min
      7595           -16.1%       6375        sched_debug.cpu.ttwu_count.min
     15873 +- 13%     +21.2%      19239 +-  6%  sched_debug.cpu.ttwu_count.stddev
      5244 +- 17%     +17.0%       6134 +-  5%  sched_debug.cpu.ttwu_local.avg
     15646 +- 12%     +21.5%      19008 +-  6%  sched_debug.cpu.ttwu_local.stddev
      0.85            -0.0        0.81        perf-stat.branch-miss-rate%
 3.689e+10            -4.6%  3.518e+10        perf-stat.branch-misses
     57.39            +0.6       58.00        perf-stat.cache-miss-rate%
 4.014e+11            -1.2%  3.967e+11        perf-stat.cache-misses
 6.994e+11            -2.2%   6.84e+11        perf-stat.cache-references
  14605393 +-  3%      -8.5%   13369913        perf-stat.context-switches
      9.21            +4.5%       9.63        perf-stat.cpi
 2.037e+14            +4.6%   2.13e+14        perf-stat.cpu-cycles
     44424            -2.0%      43541        perf-stat.cpu-migrations
      1.29            -0.1        1.24        perf-stat.dTLB-store-miss-rate%
 4.018e+10            -2.8%  3.905e+10        perf-stat.dTLB-store-misses
 3.071e+12            +1.4%  3.113e+12        perf-stat.dTLB-stores
     93.04            +1.5       94.51        perf-stat.iTLB-load-miss-rate%
 4.946e+09           +19.3%  5.903e+09 +-  5%  perf-stat.iTLB-load-misses
 3.702e+08            -7.5%  3.423e+08 +-  2%  perf-stat.iTLB-loads
      4470           -15.9%       3760 +-  5%  perf-stat.instructions-per-iTLB-miss
      0.11            -4.3%       0.10        perf-stat.ipc
 4.767e+09            -1.9%  4.675e+09        perf-stat.minor-faults
      1.46 +-  4%      -0.1        1.33 +-  9%  perf-stat.node-load-miss-rate%
      4.91            +1.7        6.65 +-  2%  perf-stat.node-store-miss-rate%
 1.195e+09           +32.8%  1.587e+09 +-  2%  perf-stat.node-store-misses
 2.313e+10            -3.7%  2.227e+10        perf-stat.node-stores
 4.767e+09            -1.9%  4.675e+09        perf-stat.page-faults
   1399047            +2.0%    1427115        perf-stat.path-length
      8908 +- 73%    -100.0%       0.00        latency_stats.avg.call_rwsem_down_read_failed.m_start.seq_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      3604 +-141%    -100.0%       0.00        latency_stats.avg.call_rwsem_down_write_failed.do_unlinkat.do_syscall_64.entry_SYSCALL_64_after_hwframe
     61499 +-130%     -92.6%       4534 +- 16%  latency_stats.avg.expand_files.__alloc_fd.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4391 +-138%     -70.9%       1277 +-129%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
     67311 +-112%     -48.5%      34681 +- 36%  latency_stats.avg.max
      3956 +-138%    +320.4%      16635 +-140%  latency_stats.avg.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    164.67 +- 30%   +7264.0%      12126 +-138%  latency_stats.avg.flush_work.fsnotify_destroy_group.inotify_release.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +5.4e+105%       5367 +-141%  latency_stats.avg.call_rwsem_down_write_failed.unlink_file_vma.free_pgtables.exit_mmap.mmput.flush_old_exec.load_elf_binary.search_binary_handler.do_execveat_common.__x64_sys_execve.do_syscall_64.entry_SYSCALL_64_after_hwframe
     36937 +-119%    -100.0%       0.00        latency_stats.max.call_rwsem_down_read_failed.m_start.seq_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      3604 +-141%    -100.0%       0.00        latency_stats.max.call_rwsem_down_write_failed.do_unlinkat.do_syscall_64.entry_SYSCALL_64_after_hwframe
     84146 +-107%     -72.5%      23171 +- 31%  latency_stats.max.expand_files.__alloc_fd.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4391 +-138%     -70.9%       1277 +-129%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
      5817 +- 83%     -69.7%       1760 +- 67%  latency_stats.max.pipe_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
      6720 +-137%   +1628.2%     116147 +-141%  latency_stats.max.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    164.67 +- 30%   +7264.0%      12126 +-138%  latency_stats.max.flush_work.fsnotify_destroy_group.inotify_release.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +1.2e+106%      12153 +-141%  latency_stats.max.call_rwsem_down_write_failed.unlink_file_vma.free_pgtables.exit_mmap.mmput.flush_old_exec.load_elf_binary.search_binary_handler.do_execveat_common.__x64_sys_execve.do_syscall_64.entry_SYSCALL_64_after_hwframe
    110122 +-120%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_read_failed.m_start.seq_read.__vfs_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
      3604 +-141%    -100.0%       0.00        latency_stats.sum.call_rwsem_down_write_failed.do_unlinkat.do_syscall_64.entry_SYSCALL_64_after_hwframe
  12078828 +-139%     -99.3%      89363 +- 29%  latency_stats.sum.expand_files.__alloc_fd.do_sys_open.do_syscall_64.entry_SYSCALL_64_after_hwframe
    144453 +-120%     -80.9%      27650 +- 19%  latency_stats.sum.poll_schedule_timeout.do_sys_poll.__x64_sys_poll.do_syscall_64.entry_SYSCALL_64_after_hwframe
      4391 +-138%     -70.9%       1277 +-129%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_lookup.nfs_lookup_revalidate.lookup_fast.walk_component.link_path_walk.path_lookupat.filename_lookup
      9438 +- 86%     -68.4%       2980 +- 35%  latency_stats.sum.pipe_write.__vfs_write.vfs_write.ksys_write.do_syscall_64.entry_SYSCALL_64_after_hwframe
     31656 +-138%    +320.4%     133084 +-140%  latency_stats.sum.rpc_wait_bit_killable.__rpc_execute.rpc_run_task.rpc_call_sync.nfs3_rpc_wrapper.nfs3_proc_getattr.__nfs_revalidate_inode.nfs_do_access.nfs_permission.inode_permission.link_path_walk.path_lookupat
    164.67 +- 30%   +7264.0%      12126 +-138%  latency_stats.sum.flush_work.fsnotify_destroy_group.inotify_release.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +8.8e+105%       8760 +-141%  latency_stats.sum.msleep_interruptible.uart_wait_until_sent.tty_wait_until_sent.tty_port_close_start.tty_port_close.tty_release.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +1.3e+106%      12897 +-141%  latency_stats.sum.tty_wait_until_sent.tty_port_close_start.tty_port_close.tty_release.__fput.task_work_run.exit_to_usermode_loop.do_syscall_64.entry_SYSCALL_64_after_hwframe
      0.00       +3.2e+106%      32207 +-141%  latency_stats.sum.call_rwsem_down_write_failed.unlink_file_vma.free_pgtables.exit_mmap.mmput.flush_old_exec.load_elf_binary.search_binary_handler.do_execveat_common.__x64_sys_execve.do_syscall_64.entry_SYSCALL_64_after_hwframe
     44.43 +-  3%     -44.4        0.00        perf-profile.calltrace.cycles-pp.alloc_pages_vma.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
     44.13 +-  3%     -44.1        0.00        perf-profile.calltrace.cycles-pp.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault.handle_mm_fault.__do_page_fault
     43.95 +-  3%     -43.9        0.00        perf-profile.calltrace.cycles-pp.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault.handle_mm_fault
     41.85 +-  4%     -41.9        0.00        perf-profile.calltrace.cycles-pp._raw_spin_lock.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.__handle_mm_fault
      7.74 +-  8%      -7.7        0.00        perf-profile.calltrace.cycles-pp.copy_page.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      7.19 +-  4%      -7.2        0.00        perf-profile.calltrace.cycles-pp.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      7.15 +-  4%      -7.2        0.00        perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      5.09 +-  3%      -5.1        0.00        perf-profile.calltrace.cycles-pp.__lru_cache_add.alloc_set_pte.finish_fault.__handle_mm_fault.handle_mm_fault
      4.99 +-  3%      -5.0        0.00        perf-profile.calltrace.cycles-pp.pagevec_lru_move_fn.__lru_cache_add.alloc_set_pte.finish_fault.__handle_mm_fault
      0.93 +-  6%      -0.1        0.81 +-  2%  perf-profile.calltrace.cycles-pp.find_get_entry.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault
      0.00            +0.8        0.84        perf-profile.calltrace.cycles-pp._raw_spin_lock.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +0.9        0.92 +-  3%  perf-profile.calltrace.cycles-pp.__list_del_entry_valid.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault
      0.00            +1.1        1.08        perf-profile.calltrace.cycles-pp.find_lock_entry.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault
      0.00            +1.1        1.14        perf-profile.calltrace.cycles-pp.shmem_getpage_gfp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault
      0.00            +1.2        1.17        perf-profile.calltrace.cycles-pp.pte_map_lock.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +1.3        1.29        perf-profile.calltrace.cycles-pp.shmem_fault.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +1.3        1.31        perf-profile.calltrace.cycles-pp.__do_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
     61.62            +1.7       63.33        perf-profile.calltrace.cycles-pp.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault.page_fault
     41.73 +-  4%      +3.0       44.75        perf-profile.calltrace.cycles-pp.native_queued_spin_lock_slowpath._raw_spin_lock.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma
      0.00            +4.6        4.55 +- 15%  perf-profile.calltrace.cycles-pp.pagevec_lru_move_fn.__lru_cache_add.alloc_set_pte.finish_fault.handle_pte_fault
      0.00            +4.6        4.65 +- 14%  perf-profile.calltrace.cycles-pp.__lru_cache_add.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault
      0.00            +6.6        6.57 +- 10%  perf-profile.calltrace.cycles-pp.alloc_set_pte.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00            +6.6        6.61 +- 10%  perf-profile.calltrace.cycles-pp.finish_fault.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00            +7.2        7.25 +-  2%  perf-profile.calltrace.cycles-pp.copy_page.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
     41.41 +- 70%     +22.3       63.67        perf-profile.calltrace.cycles-pp.handle_mm_fault.__do_page_fault.do_page_fault.page_fault
     42.19 +- 70%     +22.6       64.75        perf-profile.calltrace.cycles-pp.__do_page_fault.do_page_fault.page_fault
     42.20 +- 70%     +22.6       64.76        perf-profile.calltrace.cycles-pp.do_page_fault.page_fault
     42.27 +- 70%     +22.6       64.86        perf-profile.calltrace.cycles-pp.page_fault
      0.00           +44.9       44.88        perf-profile.calltrace.cycles-pp._raw_spin_lock.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault
      0.00           +46.9       46.92        perf-profile.calltrace.cycles-pp.get_page_from_freelist.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault.__handle_mm_fault
      0.00           +47.1       47.10        perf-profile.calltrace.cycles-pp.__alloc_pages_nodemask.alloc_pages_vma.handle_pte_fault.__handle_mm_fault.handle_mm_fault
      0.00           +47.4       47.37        perf-profile.calltrace.cycles-pp.alloc_pages_vma.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault
      0.00           +63.0       63.00        perf-profile.calltrace.cycles-pp.handle_pte_fault.__handle_mm_fault.handle_mm_fault.__do_page_fault.do_page_fault
      0.97 +-  6%      -0.1        0.84 +-  2%  perf-profile.children.cycles-pp.find_get_entry
      1.23 +-  6%      -0.1        1.11        perf-profile.children.cycles-pp.find_lock_entry
      0.09 +- 10%      -0.0        0.07 +-  6%  perf-profile.children.cycles-pp.unlock_page
      0.19 +-  4%      +0.0        0.21 +-  2%  perf-profile.children.cycles-pp.mem_cgroup_charge_statistics
      0.21 +-  2%      +0.0        0.25        perf-profile.children.cycles-pp.__mod_node_page_state
      0.00            +0.1        0.05 +-  8%  perf-profile.children.cycles-pp.get_vma_policy
      0.00            +0.1        0.08        perf-profile.children.cycles-pp.__lru_cache_add_active_or_unevictable
      0.00            +0.2        0.18 +-  2%  perf-profile.children.cycles-pp.__page_add_new_anon_rmap
      0.00            +1.3        1.30        perf-profile.children.cycles-pp.pte_map_lock
     63.40            +1.6       64.97        perf-profile.children.cycles-pp.__do_page_fault
     63.19            +1.6       64.83        perf-profile.children.cycles-pp.do_page_fault
     61.69            +1.7       63.36        perf-profile.children.cycles-pp.__handle_mm_fault
     63.19            +1.7       64.86        perf-profile.children.cycles-pp.page_fault
     61.99            +1.7       63.70        perf-profile.children.cycles-pp.handle_mm_fault
     72.27            +2.2       74.52        perf-profile.children.cycles-pp.native_queued_spin_lock_slowpath
     67.51            +2.4       69.87        perf-profile.children.cycles-pp._raw_spin_lock
     44.49 +-  3%      +3.0       47.45        perf-profile.children.cycles-pp.alloc_pages_vma
     44.28 +-  3%      +3.0       47.26        perf-profile.children.cycles-pp.__alloc_pages_nodemask
     44.13 +-  3%      +3.0       47.12        perf-profile.children.cycles-pp.get_page_from_freelist
      0.00           +63.1       63.06        perf-profile.children.cycles-pp.handle_pte_fault
      1.46 +-  7%      -0.5        1.01        perf-profile.self.cycles-pp._raw_spin_lock
      0.58 +-  6%      -0.2        0.34        perf-profile.self.cycles-pp.__handle_mm_fault
      0.55 +-  6%      -0.1        0.44 +-  2%  perf-profile.self.cycles-pp.find_get_entry
      0.22 +-  5%      -0.1        0.16 +-  2%  perf-profile.self.cycles-pp.alloc_set_pte
      0.10 +-  8%      -0.0        0.08        perf-profile.self.cycles-pp.down_read_trylock
      0.09 +-  5%      -0.0        0.07        perf-profile.self.cycles-pp.unlock_page
      0.06            -0.0        0.05        perf-profile.self.cycles-pp.pmd_devmap_trans_unstable
      0.20 +-  2%      +0.0        0.24 +-  3%  perf-profile.self.cycles-pp.__mod_node_page_state
      0.00            +0.1        0.05        perf-profile.self.cycles-pp.finish_fault
      0.00            +0.1        0.05        perf-profile.self.cycles-pp.get_vma_policy
      0.00            +0.1        0.08 +-  6%  perf-profile.self.cycles-pp.__lru_cache_add_active_or_unevictable
      0.00            +0.2        0.25        perf-profile.self.cycles-pp.handle_pte_fault
      0.00            +0.5        0.46 +-  7%  perf-profile.self.cycles-pp.pte_map_lock
     72.26            +2.3       74.52        perf-profile.self.cycles-pp.native_queued_spin_lock_slowpath

[-- Attachment #3: perf-profile.zip --]
[-- Type: application/zip, Size: 19025 bytes --]

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

Hi Haiyan,

I don't have access to the same hardware you ran the test on, but I give a try
to those test on a Power8 system (2 sockets, 5 cores/s, 8 threads/c, 80 CPUs 32G).
I run each will-it-scale test 10 times and compute the average.

test THP enabled		4.17.0-rc4-mm1	spf		delta
page_fault3_threads		2697.7		2683.5		-0.53%
page_fault2_threads		170660.6	169574.1	-0.64%
context_switch1_threads		6915269.2	6877507.3	-0.55%
context_switch1_processes	6478076.2	6529493.5	0.79%
brk1				243391.2	238527.5	-2.00%

Test were launched with the arguments '-t 80 -s 5', only the average report is
taken in account. Note that page size is 64K by default on ppc64.

It would be nice if you could capture some perf data to figure out why the
page_fault2/3 are showing such a performance regression.

Thanks,
Laurent.

On 11/06/2018 09:49, Song, HaiyanX wrote:
> Hi Laurent,
> 
> Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
> tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
> V9 patch serials.
> 
> The regression result is sorted by the metric will-it-scale.per_thread_ops.
> branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
> commit id:
>   head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
>   base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
> Benchmark: will-it-scale
> Download link: https://github.com/antonblanchard/will-it-scale/tree/master
> 
> Metrics:
>   will-it-scale.per_process_ops=processes/nr_cpu
>   will-it-scale.per_thread_ops=threads/nr_cpu
>   test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> THP: enable / disable
> nr_task:100%
> 
> 1. Regressions:
> 
> a). Enable THP
> testcase                          base           change      head           metric
> page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
> page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
> brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
> context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
> context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops
> 
> b). Disable THP
> page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
> page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
> brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
> context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
> brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
> page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
> context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops
> 
> Notes: for the above  values of test result, the higher is better.
> 
> 2. Improvement: not found improvement based on the selected test cases.
> 
> 
> Best regards
> Haiyan Song
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Monday, May 28, 2018 4:54 PM
> To: Song, HaiyanX
> Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: Re: [PATCH v11 00/26] Speculative page faults
> 
> On 28/05/2018 10:22, Haiyan Song wrote:
>> Hi Laurent,
>>
>> Yes, these tests are done on V9 patch.
> 
> Do you plan to give this V11 a run ?
> 
>>
>>
>> Best regards,
>> Haiyan Song
>>
>> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>>
>>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>>> tested on Intel 4s Skylake platform.
>>>
>>> Hi,
>>>
>>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>>> series" while responding to the v11 header series...
>>> Were these tests done on v9 or v11 ?
>>>
>>> Cheers,
>>> Laurent.
>>>
>>>>
>>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>>> Commit id:
>>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>>> Benchmark suite: will-it-scale
>>>> Download link:
>>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>>> Metrics:
>>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>>> THP: enable / disable
>>>> nr_task: 100%
>>>>
>>>> 1. Regressions:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>>
>>>> 2. Improvements:
>>>> a) THP enabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>>
>>>> b) THP disabled:
>>>> testcase                        base            change          head       metric
>>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>>
>>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>>> on head commit is better than that on base commit for this benchmark.
>>>>
>>>>
>>>> Best regards
>>>> Haiyan Song
>>>>
>>>> ________________________________________
>>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>>> Sent: Thursday, May 17, 2018 7:06 PM
>>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>>
>>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>>> page fault without holding the mm semaphore [1].
>>>>
>>>> The idea is to try to handle user space page faults without holding the
>>>> mmap_sem. This should allow better concurrency for massively threaded
>>>> process since the page fault handler will not wait for other threads memory
>>>> layout change to be done, assuming that this change is done in another part
>>>> of the process's memory space. This type page fault is named speculative
>>>> page fault. If the speculative page fault fails because of a concurrency is
>>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>>> is failing its processing and a classic page fault is then tried.
>>>>
>>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>>> which protects the access to the mm_rb tree. Previously this was done using
>>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>>> freeing operation which was hitting the performance by 20% as reported by
>>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>>> limiting the locking contention to these operations which are expected to
>>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>>> our back a reference count is added and 2 services (get_vma() and
>>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>>> fetched from the RB tree using get_vma(), it must be later freed using
>>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>>> benchmark anymore.
>>>>
>>>> The VMA's attributes checked during the speculative page fault processing
>>>> have to be protected against parallel changes. This is done by using a per
>>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>>> handler to fast check for parallel changes in progress and to abort the
>>>> speculative page fault in that case.
>>>>
>>>> Once the VMA has been found, the speculative page fault handler would check
>>>> for the VMA's attributes to verify that the page fault has to be handled
>>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>>> allows fast detection of concurrent VMA changes. If such a change is
>>>> detected, the speculative page fault is aborted and a *classic* page fault
>>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>>> checked during the page fault are modified.
>>>>
>>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>>> so once the page table is locked, the VMA is valid, so any other changes
>>>> leading to touching this PTE will need to lock the page table, so no
>>>> parallel change is possible at this time.
>>>>
>>>> The locking of the PTE is done with interrupts disabled, this allows
>>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>>> valid at the time the PTE is locked, we have the guarantee that the
>>>> collapsing operation will have to wait on the PTE lock to move forward.
>>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>>> different from the one recorded at the beginning of the SPF operation, the
>>>> classic page fault handler will be called to handle the operation while
>>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>>> page fault while a TLB invalidate is requested by another CPU holding the
>>>> PTE.
>>>>
>>>> In pseudo code, this could be seen as:
>>>>     speculative_page_fault()
>>>>     {
>>>>             vma = get_vma()
>>>>             check vma sequence count
>>>>             check vma's support
>>>>             disable interrupt
>>>>                   check pgd,p4d,...,pte
>>>>                   save pmd and pte in vmf
>>>>                   save vma sequence counter in vmf
>>>>             enable interrupt
>>>>             check vma sequence count
>>>>             handle_pte_fault(vma)
>>>>                     ..
>>>>                     page = alloc_page()
>>>>                     pte_map_lock()
>>>>                             disable interrupt
>>>>                                     abort if sequence counter has changed
>>>>                                     abort if pmd or pte has changed
>>>>                                     pte map and lock
>>>>                             enable interrupt
>>>>                     if abort
>>>>                        free page
>>>>                        abort
>>>>                     ...
>>>>     }
>>>>
>>>>     arch_fault_handler()
>>>>     {
>>>>             if (speculative_page_fault(&vma))
>>>>                goto done
>>>>     again:
>>>>             lock(mmap_sem)
>>>>             vma = find_vma();
>>>>             handle_pte_fault(vma);
>>>>             if retry
>>>>                unlock(mmap_sem)
>>>>                goto again;
>>>>     done:
>>>>             handle fault error
>>>>     }
>>>>
>>>> Support for THP is not done because when checking for the PMD, we can be
>>>> confused by an in progress collapsing operation done by khugepaged. The
>>>> issue is that pmd_none() could be true either if the PMD is not already
>>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>>> cannot safely allocate a PMD if pmd_none() is true.
>>>>
>>>> This series add a new software performance event named 'speculative-faults'
>>>> or 'spf'. It counts the number of successful page fault event handled
>>>> speculatively. When recording 'faults,spf' events, the faults one is
>>>> counting the total number of page fault events while 'spf' is only counting
>>>> the part of the faults processed speculatively.
>>>>
>>>> There are some trace events introduced by this series. They allow
>>>> identifying why the page faults were not processed speculatively. This
>>>> doesn't take in account the faults generated by a monothreaded process
>>>> which directly processed while holding the mmap_sem. This trace events are
>>>> grouped in a system named 'pagefault', they are:
>>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>>    back.
>>>>
>>>> To record all the related events, the easier is to run perf with the
>>>> following arguments :
>>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>>
>>>> There is also a dedicated vmstat counter showing the number of successful
>>>> page fault handled speculatively. I can be seen this way:
>>>> $ grep speculative_pgfault /proc/vmstat
>>>>
>>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>>> on x86, PowerPC and arm64.
>>>>
>>>> ---------------------
>>>> Real Workload results
>>>>
>>>> As mentioned in previous email, we did non official runs using a "popular
>>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>>> which showed a 30% improvements in the number of transaction processed per
>>>> second. This run has been done on the v6 series, but changes introduced in
>>>> this new version should not impact the performance boost seen.
>>>>
>>>> Here are the perf data captured during 2 of these runs on top of the v8
>>>> series:
>>>>                 vanilla         spf
>>>> faults          89.418          101.364         +13%
>>>> spf                n/a           97.989
>>>>
>>>> With the SPF kernel, most of the page fault were processed in a speculative
>>>> way.
>>>>
>>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>>> it a try on an android device. He reported that the application launch time
>>>> was improved in average by 6%, and for large applications (~100 threads) by
>>>> 20%.
>>>>
>>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>>> MSM845 (8 cores) with 6GB (the less is better):
>>>>
>>>> Application                             4.9     4.9+spf delta
>>>> com.tencent.mm                          416     389     -7%
>>>> com.eg.android.AlipayGphone             1135    986     -13%
>>>> com.tencent.mtt                         455     454     0%
>>>> com.qqgame.hlddz                        1497    1409    -6%
>>>> com.autonavi.minimap                    711     701     -1%
>>>> com.tencent.tmgp.sgame                  788     748     -5%
>>>> com.immomo.momo                         501     487     -3%
>>>> com.tencent.peng                        2145    2112    -2%
>>>> com.smile.gifmaker                      491     461     -6%
>>>> com.baidu.BaiduMap                      479     366     -23%
>>>> com.taobao.taobao                       1341    1198    -11%
>>>> com.baidu.searchbox                     333     314     -6%
>>>> com.tencent.mobileqq                    394     384     -3%
>>>> com.sina.weibo                          907     906     0%
>>>> com.youku.phone                         816     731     -11%
>>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>>> com.UCMobile                            415     411     -1%
>>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>>> com.tencent.qqmusic                     336     329     -2%
>>>> com.sankuai.meituan                     1661    1302    -22%
>>>> com.netease.cloudmusic                  1193    1200    1%
>>>> air.tv.douyu.android                    4257    4152    -2%
>>>>
>>>> ------------------
>>>> Benchmarks results
>>>>
>>>> Base kernel is v4.17.0-rc4-mm1
>>>> SPF is BASE + this series
>>>>
>>>> Kernbench:
>>>> ----------
>>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>>> kernel (kernel is build 5 times):
>>>>
>>>> Average Half load -j 8
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>>
>>>> Average Optimal load -j 16
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>>
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          526743764      faults
>>>>                210      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>               2278      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Very few speculative page faults were recorded as most of the processes
>>>> involved are monothreaded (sounds that on this architecture some threads
>>>> were created during the kernel build processing).
>>>>
>>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>>
>>>> Average Half load -j 40
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>>
>>>> Average Optimal load -j 80
>>>>                  Run    (std deviation)
>>>>                  BASE                   SPF
>>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>>
>>>> During a run on the SPF, perf events were captured:
>>>>  Performance counter stats for '../kernbench -M':
>>>>          116730856      faults
>>>>                  0      spf
>>>>                  3      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                476      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> Most of the processes involved are monothreaded so SPF is not activated but
>>>> there is no impact on the performance.
>>>>
>>>> Ebizzy:
>>>> -------
>>>> The test is counting the number of records per second it can manage, the
>>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>>> consistent result I repeated the test 100 times and measure the average
>>>> result. The number is the record processes per second, the higher is the
>>>> best.
>>>>
>>>>                 BASE            SPF             delta
>>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>>
>>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>>  Performance counter stats for './ebizzy -mTt 16':
>>>>            1706379      faults
>>>>            1674599      spf
>>>>              30588      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                363      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> And the ones captured during a run on a 80 CPUs Power node:
>>>>  Performance counter stats for './ebizzy -mTt 80':
>>>>            1874773      faults
>>>>            1461153      spf
>>>>             413293      pagefault:spf_vma_changed
>>>>                  0      pagefault:spf_vma_noanon
>>>>                200      pagefault:spf_vma_notsup
>>>>                  0      pagefault:spf_vma_access
>>>>                  0      pagefault:spf_pmd_changed
>>>>
>>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>>> leading the ebizzy performance boost.
>>>>
>>>> ------------------
>>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>>    and Minchan Kim, hopefully.
>>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>>    __do_page_fault().
>>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>>    instead
>>>>    of aborting the speculative page fault handling. Dropping the now
>>>> useless
>>>>    trace event pagefault:spf_pte_lock.
>>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>>    handling when retrying is needed. This adds a lot of complexity and
>>>>    additional tests done didn't show a significant performance improvement.
>>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>>
>>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>>
>>>>
>>>> Laurent Dufour (20):
>>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>>   mm: make pte_unmap_same compatible with SPF
>>>>   mm: introduce INIT_VMA()
>>>>   mm: protect VMA modifications using VMA sequence count
>>>>   mm: protect mremap() against SPF hanlder
>>>>   mm: protect SPF handler against anon_vma changes
>>>>   mm: cache some VMA fields in the vm_fault structure
>>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>>   mm: introduce __vm_normal_page()
>>>>   mm: introduce __page_add_new_anon_rmap()
>>>>   mm: protect mm_rb tree with a rwlock
>>>>   mm: adding speculative page fault failure trace events
>>>>   perf: add a speculative page fault sw event
>>>>   perf tools: add support for the SPF perf event
>>>>   mm: add speculative page fault vmstats
>>>>   powerpc/mm: add speculative page fault
>>>>
>>>> Mahendran Ganesh (2):
>>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>>   arm64/mm: add speculative page fault
>>>>
>>>> Peter Zijlstra (4):
>>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>>   mm: VMA sequence count
>>>>   mm: provide speculative fault infrastructure
>>>>   x86/mm: add speculative pagefault handling
>>>>
>>>>  arch/arm64/Kconfig                    |   1 +
>>>>  arch/arm64/mm/fault.c                 |  12 +
>>>>  arch/powerpc/Kconfig                  |   1 +
>>>>  arch/powerpc/mm/fault.c               |  16 +
>>>>  arch/x86/Kconfig                      |   1 +
>>>>  arch/x86/mm/fault.c                   |  27 +-
>>>>  fs/exec.c                             |   2 +-
>>>>  fs/proc/task_mmu.c                    |   5 +-
>>>>  fs/userfaultfd.c                      |  17 +-
>>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>>  include/linux/migrate.h               |   4 +-
>>>>  include/linux/mm.h                    | 136 +++++++-
>>>>  include/linux/mm_types.h              |   7 +
>>>>  include/linux/pagemap.h               |   4 +-
>>>>  include/linux/rmap.h                  |  12 +-
>>>>  include/linux/swap.h                  |  10 +-
>>>>  include/linux/vm_event_item.h         |   3 +
>>>>  include/trace/events/pagefault.h      |  80 +++++
>>>>  include/uapi/linux/perf_event.h       |   1 +
>>>>  kernel/fork.c                         |   5 +-
>>>>  mm/Kconfig                            |  22 ++
>>>>  mm/huge_memory.c                      |   6 +-
>>>>  mm/hugetlb.c                          |   2 +
>>>>  mm/init-mm.c                          |   3 +
>>>>  mm/internal.h                         |  20 ++
>>>>  mm/khugepaged.c                       |   5 +
>>>>  mm/madvise.c                          |   6 +-
>>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>>  mm/mempolicy.c                        |  51 ++-
>>>>  mm/migrate.c                          |   6 +-
>>>>  mm/mlock.c                            |  13 +-
>>>>  mm/mmap.c                             | 229 ++++++++++---
>>>>  mm/mprotect.c                         |   4 +-
>>>>  mm/mremap.c                           |  13 +
>>>>  mm/nommu.c                            |   2 +-
>>>>  mm/rmap.c                             |   5 +-
>>>>  mm/swap.c                             |   6 +-
>>>>  mm/swap_state.c                       |   8 +-
>>>>  mm/vmstat.c                           |   5 +-
>>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>>  tools/perf/util/evsel.c               |   1 +
>>>>  tools/perf/util/parse-events.c        |   4 +
>>>>  tools/perf/util/parse-events.l        |   1 +
>>>>  tools/perf/util/python.c              |   1 +
>>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>>  create mode 100644 include/trace/events/pagefault.h
>>>>
>>>> --
>>>> 2.7.4
>>>>
>>>>
>>>
>>
> 

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

Hi Laurent,

Regression test for v11 patch serials have been run, some regression is found by LKP-tools (linux kernel performance)
tested on Intel 4s skylake platform. This time only test the cases which have been run and found regressions on
V9 patch serials.

The regression result is sorted by the metric will-it-scale.per_thread_ops.
branch: Laurent-Dufour/Speculative-page-faults/20180520-045126
commit id:
  head commit : a7a8993bfe3ccb54ad468b9f1799649e4ad1ff12
  base commit : ba98a1cdad71d259a194461b3a61471b49b14df1
Benchmark: will-it-scale
Download link: https://github.com/antonblanchard/will-it-scale/tree/master

Metrics:
  will-it-scale.per_process_ops=processes/nr_cpu
  will-it-scale.per_thread_ops=threads/nr_cpu
  test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
THP: enable / disable
nr_task:100%

1. Regressions:

a). Enable THP
testcase                          base           change      head           metric
page_fault3/enable THP           10519          -20.5%        836      will-it-scale.per_thread_ops
page_fault2/enalbe THP            8281          -18.8%       6728      will-it-scale.per_thread_ops
brk1/eanble THP                 998475           -2.2%     976893      will-it-scale.per_process_ops
context_switch1/enable THP      223910           -1.3%     220930      will-it-scale.per_process_ops
context_switch1/enable THP      233722           -1.0%     231288      will-it-scale.per_thread_ops

b). Disable THP
page_fault3/disable THP          10856          -23.1%       8344      will-it-scale.per_thread_ops
page_fault2/disable THP           8147          -18.8%       6613      will-it-scale.per_thread_ops
brk1/disable THP                   957           -7.9%        881      will-it-scale.per_thread_ops
context_switch1/disable THP     237006           -2.2%     231907      will-it-scale.per_thread_ops
brk1/disable THP                997317           -2.0%     977778      will-it-scale.per_process_ops
page_fault3/disable THP         467454           -1.8%     459251      will-it-scale.per_process_ops
context_switch1/disable THP     224431           -1.3%     221567      will-it-scale.per_process_ops

Notes: for the above  values of test result, the higher is better.

2. Improvement: not found improvement based on the selected test cases.


Best regards
Haiyan Song
________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Monday, May 28, 2018 4:54 PM
To: Song, HaiyanX
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 28/05/2018 10:22, Haiyan Song wrote:
> Hi Laurent,
>
> Yes, these tests are done on V9 patch.

Do you plan to give this V11 a run ?

>
>
> Best regards,
> Haiyan Song
>
> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>
>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>> tested on Intel 4s Skylake platform.
>>
>> Hi,
>>
>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>> series" while responding to the v11 header series...
>> Were these tests done on v9 or v11 ?
>>
>> Cheers,
>> Laurent.
>>
>>>
>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>> Commit id:
>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>> Benchmark suite: will-it-scale
>>> Download link:
>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>> Metrics:
>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>> THP: enable / disable
>>> nr_task: 100%
>>>
>>> 1. Regressions:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>
>>> 2. Improvements:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>
>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>> on head commit is better than that on base commit for this benchmark.
>>>
>>>
>>> Best regards
>>> Haiyan Song
>>>
>>> ________________________________________
>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Thursday, May 17, 2018 7:06 PM
>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>
>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>> page fault without holding the mm semaphore [1].
>>>
>>> The idea is to try to handle user space page faults without holding the
>>> mmap_sem. This should allow better concurrency for massively threaded
>>> process since the page fault handler will not wait for other threads memory
>>> layout change to be done, assuming that this change is done in another part
>>> of the process's memory space. This type page fault is named speculative
>>> page fault. If the speculative page fault fails because of a concurrency is
>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>> is failing its processing and a classic page fault is then tried.
>>>
>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>> which protects the access to the mm_rb tree. Previously this was done using
>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>> freeing operation which was hitting the performance by 20% as reported by
>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>> limiting the locking contention to these operations which are expected to
>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>> our back a reference count is added and 2 services (get_vma() and
>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>> fetched from the RB tree using get_vma(), it must be later freed using
>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>> benchmark anymore.
>>>
>>> The VMA's attributes checked during the speculative page fault processing
>>> have to be protected against parallel changes. This is done by using a per
>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>> handler to fast check for parallel changes in progress and to abort the
>>> speculative page fault in that case.
>>>
>>> Once the VMA has been found, the speculative page fault handler would check
>>> for the VMA's attributes to verify that the page fault has to be handled
>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>> allows fast detection of concurrent VMA changes. If such a change is
>>> detected, the speculative page fault is aborted and a *classic* page fault
>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>> checked during the page fault are modified.
>>>
>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>> so once the page table is locked, the VMA is valid, so any other changes
>>> leading to touching this PTE will need to lock the page table, so no
>>> parallel change is possible at this time.
>>>
>>> The locking of the PTE is done with interrupts disabled, this allows
>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>> valid at the time the PTE is locked, we have the guarantee that the
>>> collapsing operation will have to wait on the PTE lock to move forward.
>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>> different from the one recorded at the beginning of the SPF operation, the
>>> classic page fault handler will be called to handle the operation while
>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>> page fault while a TLB invalidate is requested by another CPU holding the
>>> PTE.
>>>
>>> In pseudo code, this could be seen as:
>>>     speculative_page_fault()
>>>     {
>>>             vma = get_vma()
>>>             check vma sequence count
>>>             check vma's support
>>>             disable interrupt
>>>                   check pgd,p4d,...,pte
>>>                   save pmd and pte in vmf
>>>                   save vma sequence counter in vmf
>>>             enable interrupt
>>>             check vma sequence count
>>>             handle_pte_fault(vma)
>>>                     ..
>>>                     page = alloc_page()
>>>                     pte_map_lock()
>>>                             disable interrupt
>>>                                     abort if sequence counter has changed
>>>                                     abort if pmd or pte has changed
>>>                                     pte map and lock
>>>                             enable interrupt
>>>                     if abort
>>>                        free page
>>>                        abort
>>>                     ...
>>>     }
>>>
>>>     arch_fault_handler()
>>>     {
>>>             if (speculative_page_fault(&vma))
>>>                goto done
>>>     again:
>>>             lock(mmap_sem)
>>>             vma = find_vma();
>>>             handle_pte_fault(vma);
>>>             if retry
>>>                unlock(mmap_sem)
>>>                goto again;
>>>     done:
>>>             handle fault error
>>>     }
>>>
>>> Support for THP is not done because when checking for the PMD, we can be
>>> confused by an in progress collapsing operation done by khugepaged. The
>>> issue is that pmd_none() could be true either if the PMD is not already
>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>> cannot safely allocate a PMD if pmd_none() is true.
>>>
>>> This series add a new software performance event named 'speculative-faults'
>>> or 'spf'. It counts the number of successful page fault event handled
>>> speculatively. When recording 'faults,spf' events, the faults one is
>>> counting the total number of page fault events while 'spf' is only counting
>>> the part of the faults processed speculatively.
>>>
>>> There are some trace events introduced by this series. They allow
>>> identifying why the page faults were not processed speculatively. This
>>> doesn't take in account the faults generated by a monothreaded process
>>> which directly processed while holding the mmap_sem. This trace events are
>>> grouped in a system named 'pagefault', they are:
>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>    back.
>>>
>>> To record all the related events, the easier is to run perf with the
>>> following arguments :
>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>
>>> There is also a dedicated vmstat counter showing the number of successful
>>> page fault handled speculatively. I can be seen this way:
>>> $ grep speculative_pgfault /proc/vmstat
>>>
>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>> on x86, PowerPC and arm64.
>>>
>>> ---------------------
>>> Real Workload results
>>>
>>> As mentioned in previous email, we did non official runs using a "popular
>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>> which showed a 30% improvements in the number of transaction processed per
>>> second. This run has been done on the v6 series, but changes introduced in
>>> this new version should not impact the performance boost seen.
>>>
>>> Here are the perf data captured during 2 of these runs on top of the v8
>>> series:
>>>                 vanilla         spf
>>> faults          89.418          101.364         +13%
>>> spf                n/a           97.989
>>>
>>> With the SPF kernel, most of the page fault were processed in a speculative
>>> way.
>>>
>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>> it a try on an android device. He reported that the application launch time
>>> was improved in average by 6%, and for large applications (~100 threads) by
>>> 20%.
>>>
>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>> MSM845 (8 cores) with 6GB (the less is better):
>>>
>>> Application                             4.9     4.9+spf delta
>>> com.tencent.mm                          416     389     -7%
>>> com.eg.android.AlipayGphone             1135    986     -13%
>>> com.tencent.mtt                         455     454     0%
>>> com.qqgame.hlddz                        1497    1409    -6%
>>> com.autonavi.minimap                    711     701     -1%
>>> com.tencent.tmgp.sgame                  788     748     -5%
>>> com.immomo.momo                         501     487     -3%
>>> com.tencent.peng                        2145    2112    -2%
>>> com.smile.gifmaker                      491     461     -6%
>>> com.baidu.BaiduMap                      479     366     -23%
>>> com.taobao.taobao                       1341    1198    -11%
>>> com.baidu.searchbox                     333     314     -6%
>>> com.tencent.mobileqq                    394     384     -3%
>>> com.sina.weibo                          907     906     0%
>>> com.youku.phone                         816     731     -11%
>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>> com.UCMobile                            415     411     -1%
>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>> com.tencent.qqmusic                     336     329     -2%
>>> com.sankuai.meituan                     1661    1302    -22%
>>> com.netease.cloudmusic                  1193    1200    1%
>>> air.tv.douyu.android                    4257    4152    -2%
>>>
>>> ------------------
>>> Benchmarks results
>>>
>>> Base kernel is v4.17.0-rc4-mm1
>>> SPF is BASE + this series
>>>
>>> Kernbench:
>>> ----------
>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>> kernel (kernel is build 5 times):
>>>
>>> Average Half load -j 8
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>
>>> Average Optimal load -j 16
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          526743764      faults
>>>                210      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>               2278      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Very few speculative page faults were recorded as most of the processes
>>> involved are monothreaded (sounds that on this architecture some threads
>>> were created during the kernel build processing).
>>>
>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>
>>> Average Half load -j 40
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>
>>> Average Optimal load -j 80
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          116730856      faults
>>>                  0      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                476      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Most of the processes involved are monothreaded so SPF is not activated but
>>> there is no impact on the performance.
>>>
>>> Ebizzy:
>>> -------
>>> The test is counting the number of records per second it can manage, the
>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>> consistent result I repeated the test 100 times and measure the average
>>> result. The number is the record processes per second, the higher is the
>>> best.
>>>
>>>                 BASE            SPF             delta
>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>
>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>  Performance counter stats for './ebizzy -mTt 16':
>>>            1706379      faults
>>>            1674599      spf
>>>              30588      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                363      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> And the ones captured during a run on a 80 CPUs Power node:
>>>  Performance counter stats for './ebizzy -mTt 80':
>>>            1874773      faults
>>>            1461153      spf
>>>             413293      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                200      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>> leading the ebizzy performance boost.
>>>
>>> ------------------
>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>    and Minchan Kim, hopefully.
>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>    __do_page_fault().
>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>    instead
>>>    of aborting the speculative page fault handling. Dropping the now
>>> useless
>>>    trace event pagefault:spf_pte_lock.
>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>    handling when retrying is needed. This adds a lot of complexity and
>>>    additional tests done didn't show a significant performance improvement.
>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>
>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>
>>>
>>> Laurent Dufour (20):
>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>   mm: make pte_unmap_same compatible with SPF
>>>   mm: introduce INIT_VMA()
>>>   mm: protect VMA modifications using VMA sequence count
>>>   mm: protect mremap() against SPF hanlder
>>>   mm: protect SPF handler against anon_vma changes
>>>   mm: cache some VMA fields in the vm_fault structure
>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>   mm: introduce __vm_normal_page()
>>>   mm: introduce __page_add_new_anon_rmap()
>>>   mm: protect mm_rb tree with a rwlock
>>>   mm: adding speculative page fault failure trace events
>>>   perf: add a speculative page fault sw event
>>>   perf tools: add support for the SPF perf event
>>>   mm: add speculative page fault vmstats
>>>   powerpc/mm: add speculative page fault
>>>
>>> Mahendran Ganesh (2):
>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   arm64/mm: add speculative page fault
>>>
>>> Peter Zijlstra (4):
>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>   mm: VMA sequence count
>>>   mm: provide speculative fault infrastructure
>>>   x86/mm: add speculative pagefault handling
>>>
>>>  arch/arm64/Kconfig                    |   1 +
>>>  arch/arm64/mm/fault.c                 |  12 +
>>>  arch/powerpc/Kconfig                  |   1 +
>>>  arch/powerpc/mm/fault.c               |  16 +
>>>  arch/x86/Kconfig                      |   1 +
>>>  arch/x86/mm/fault.c                   |  27 +-
>>>  fs/exec.c                             |   2 +-
>>>  fs/proc/task_mmu.c                    |   5 +-
>>>  fs/userfaultfd.c                      |  17 +-
>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>  include/linux/migrate.h               |   4 +-
>>>  include/linux/mm.h                    | 136 +++++++-
>>>  include/linux/mm_types.h              |   7 +
>>>  include/linux/pagemap.h               |   4 +-
>>>  include/linux/rmap.h                  |  12 +-
>>>  include/linux/swap.h                  |  10 +-
>>>  include/linux/vm_event_item.h         |   3 +
>>>  include/trace/events/pagefault.h      |  80 +++++
>>>  include/uapi/linux/perf_event.h       |   1 +
>>>  kernel/fork.c                         |   5 +-
>>>  mm/Kconfig                            |  22 ++
>>>  mm/huge_memory.c                      |   6 +-
>>>  mm/hugetlb.c                          |   2 +
>>>  mm/init-mm.c                          |   3 +
>>>  mm/internal.h                         |  20 ++
>>>  mm/khugepaged.c                       |   5 +
>>>  mm/madvise.c                          |   6 +-
>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>  mm/mempolicy.c                        |  51 ++-
>>>  mm/migrate.c                          |   6 +-
>>>  mm/mlock.c                            |  13 +-
>>>  mm/mmap.c                             | 229 ++++++++++---
>>>  mm/mprotect.c                         |   4 +-
>>>  mm/mremap.c                           |  13 +
>>>  mm/nommu.c                            |   2 +-
>>>  mm/rmap.c                             |   5 +-
>>>  mm/swap.c                             |   6 +-
>>>  mm/swap_state.c                       |   8 +-
>>>  mm/vmstat.c                           |   5 +-
>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>  tools/perf/util/evsel.c               |   1 +
>>>  tools/perf/util/parse-events.c        |   4 +
>>>  tools/perf/util/parse-events.l        |   1 +
>>>  tools/perf/util/python.c              |   1 +
>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>  create mode 100644 include/trace/events/pagefault.h
>>>
>>> --
>>> 2.7.4
>>>
>>>
>>
>

RE: [PATCH v11 00/26] Speculative page faults

[Wang, Kemi]

Full run would take one or two weeks depended on our resource available. Could you pick some ones up, e.g. those have performance regression?

-----Original Message-----
From: owner-linux-mm@kvack.org [mailto:owner-linux-mm@kvack.org] On Behalf Of Laurent Dufour
Sent: Monday, May 28, 2018 4:55 PM
To: Song, HaiyanX <haiyanx.song@intel.com>
Cc: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox <willy@infradead.org>; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner <tglx@linutronix.de>; Ingo Molnar <mingo@redhat.com>; hpa@zytor.com; Will Deacon <will.deacon@arm.com>; Sergey Senozhatsky <sergey.senozhatsky@gmail.com>; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli <aarcange@redhat.com>; Alexei Starovoitov <alexei.starovoitov@gmail.com>; Wang, Kemi <kemi.wang@intel.com>; Daniel Jordan <daniel.m.jordan@oracle.com>; David Rientjes <rientjes@google.com>; Jerome Glisse <jglisse@redhat.com>; Ganesh Mahendran <opensource.ganesh@gmail.com>; Minchan Kim <minchan@kernel.org>; Punit Agrawal <punitagrawal@gmail.com>; vinayak menon <vinayakm.list@gmail.com>; Yang Shi <yang.shi@linux.alibaba.com>; linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen <tim.c.chen@linux.intel.com>; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: Re: [PATCH v11 00/26] Speculative page faults

On 28/05/2018 10:22, Haiyan Song wrote:
> Hi Laurent,
> 
> Yes, these tests are done on V9 patch.

Do you plan to give this V11 a run ?

> 
> 
> Best regards,
> Haiyan Song
> 
> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>
>>> Some regression and improvements is found by LKP-tools(linux kernel 
>>> performance) on V9 patch series tested on Intel 4s Skylake platform.
>>
>> Hi,
>>
>> Thanks for reporting this benchmark results, but you mentioned the 
>> "V9 patch series" while responding to the v11 header series...
>> Were these tests done on v9 or v11 ?
>>
>> Cheers,
>> Laurent.
>>
>>>
>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 
>>> patch series) Commit id:
>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>> Benchmark suite: will-it-scale
>>> Download link:
>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>> Metrics:
>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>> THP: enable / disable
>>> nr_task: 100%
>>>
>>> 1. Regressions:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>
>>> 2. Improvements:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>
>>> Notes: for above values in column "change", the higher value means 
>>> that the related testcase result on head commit is better than that on base commit for this benchmark.
>>>
>>>
>>> Best regards
>>> Haiyan Song
>>>
>>> ________________________________________
>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf 
>>> of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Thursday, May 17, 2018 7:06 PM
>>> To: akpm@linux-foundation.org; mhocko@kernel.org; 
>>> peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; 
>>> dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; 
>>> khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; 
>>> benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; 
>>> Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey 
>>> Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; 
>>> Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; 
>>> Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak 
>>> menon; Yang Shi
>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; 
>>> haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; 
>>> paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; 
>>> x86@kernel.org
>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>
>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to 
>>> handle page fault without holding the mm semaphore [1].
>>>
>>> The idea is to try to handle user space page faults without holding 
>>> the mmap_sem. This should allow better concurrency for massively 
>>> threaded process since the page fault handler will not wait for 
>>> other threads memory layout change to be done, assuming that this 
>>> change is done in another part of the process's memory space. This 
>>> type page fault is named speculative page fault. If the speculative 
>>> page fault fails because of a concurrency is detected or because 
>>> underlying PMD or PTE tables are not yet allocating, it is failing its processing and a classic page fault is then tried.
>>>
>>> The speculative page fault (SPF) has to look for the VMA matching 
>>> the fault address without holding the mmap_sem, this is done by 
>>> introducing a rwlock which protects the access to the mm_rb tree. 
>>> Previously this was done using SRCU but it was introducing a lot of 
>>> scheduling to process the VMA's freeing operation which was hitting 
>>> the performance by 20% as reported by Kemi Wang [2]. Using a rwlock 
>>> to protect access to the mm_rb tree is limiting the locking 
>>> contention to these operations which are expected to be in a O(log 
>>> n) order. In addition to ensure that the VMA is not freed in our 
>>> back a reference count is added and 2 services (get_vma() and
>>> put_vma()) are introduced to handle the reference count. Once a VMA 
>>> is fetched from the RB tree using get_vma(), it must be later freed 
>>> using put_vma(). I can't see anymore the overhead I got while 
>>> will-it-scale benchmark anymore.
>>>
>>> The VMA's attributes checked during the speculative page fault 
>>> processing have to be protected against parallel changes. This is 
>>> done by using a per VMA sequence lock. This sequence lock allows the 
>>> speculative page fault handler to fast check for parallel changes in 
>>> progress and to abort the speculative page fault in that case.
>>>
>>> Once the VMA has been found, the speculative page fault handler 
>>> would check for the VMA's attributes to verify that the page fault 
>>> has to be handled correctly or not. Thus, the VMA is protected 
>>> through a sequence lock which allows fast detection of concurrent 
>>> VMA changes. If such a change is detected, the speculative page 
>>> fault is aborted and a *classic* page fault is tried.  VMA sequence 
>>> lockings are added when VMA attributes which are checked during the page fault are modified.
>>>
>>> When the PTE is fetched, the VMA is checked to see if it has been 
>>> changed, so once the page table is locked, the VMA is valid, so any 
>>> other changes leading to touching this PTE will need to lock the 
>>> page table, so no parallel change is possible at this time.
>>>
>>> The locking of the PTE is done with interrupts disabled, this allows 
>>> checking for the PMD to ensure that there is not an ongoing 
>>> collapsing operation. Since khugepaged is firstly set the PMD to 
>>> pmd_none and then is waiting for the other CPU to have caught the 
>>> IPI interrupt, if the pmd is valid at the time the PTE is locked, we 
>>> have the guarantee that the collapsing operation will have to wait on the PTE lock to move forward.
>>> This allows the SPF handler to map the PTE safely. If the PMD value 
>>> is different from the one recorded at the beginning of the SPF 
>>> operation, the classic page fault handler will be called to handle 
>>> the operation while holding the mmap_sem. As the PTE lock is done 
>>> with the interrupts disabled, the lock is done using spin_trylock() 
>>> to avoid dead lock when handling a page fault while a TLB invalidate 
>>> is requested by another CPU holding the PTE.
>>>
>>> In pseudo code, this could be seen as:
>>>     speculative_page_fault()
>>>     {
>>>             vma = get_vma()
>>>             check vma sequence count
>>>             check vma's support
>>>             disable interrupt
>>>                   check pgd,p4d,...,pte
>>>                   save pmd and pte in vmf
>>>                   save vma sequence counter in vmf
>>>             enable interrupt
>>>             check vma sequence count
>>>             handle_pte_fault(vma)
>>>                     ..
>>>                     page = alloc_page()
>>>                     pte_map_lock()
>>>                             disable interrupt
>>>                                     abort if sequence counter has changed
>>>                                     abort if pmd or pte has changed
>>>                                     pte map and lock
>>>                             enable interrupt
>>>                     if abort
>>>                        free page
>>>                        abort
>>>                     ...
>>>     }
>>>
>>>     arch_fault_handler()
>>>     {
>>>             if (speculative_page_fault(&vma))
>>>                goto done
>>>     again:
>>>             lock(mmap_sem)
>>>             vma = find_vma();
>>>             handle_pte_fault(vma);
>>>             if retry
>>>                unlock(mmap_sem)
>>>                goto again;
>>>     done:
>>>             handle fault error
>>>     }
>>>
>>> Support for THP is not done because when checking for the PMD, we 
>>> can be confused by an in progress collapsing operation done by 
>>> khugepaged. The issue is that pmd_none() could be true either if the 
>>> PMD is not already populated or if the underlying PTE are in the way 
>>> to be collapsed. So we cannot safely allocate a PMD if pmd_none() is true.
>>>
>>> This series add a new software performance event named 'speculative-faults'
>>> or 'spf'. It counts the number of successful page fault event 
>>> handled speculatively. When recording 'faults,spf' events, the 
>>> faults one is counting the total number of page fault events while 
>>> 'spf' is only counting the part of the faults processed speculatively.
>>>
>>> There are some trace events introduced by this series. They allow 
>>> identifying why the page faults were not processed speculatively. 
>>> This doesn't take in account the faults generated by a monothreaded 
>>> process which directly processed while holding the mmap_sem. This 
>>> trace events are grouped in a system named 'pagefault', they are:
>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our 
>>> back
>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>  - pagefault:spf_vma_access : the VMA's access right are not 
>>> respected
>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>    back.
>>>
>>> To record all the related events, the easier is to run perf with the 
>>> following arguments :
>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>
>>> There is also a dedicated vmstat counter showing the number of 
>>> successful page fault handled speculatively. I can be seen this way:
>>> $ grep speculative_pgfault /proc/vmstat
>>>
>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is 
>>> functional on x86, PowerPC and arm64.
>>>
>>> ---------------------
>>> Real Workload results
>>>
>>> As mentioned in previous email, we did non official runs using a 
>>> "popular in memory multithreaded database product" on 176 cores SMT8 
>>> Power system which showed a 30% improvements in the number of 
>>> transaction processed per second. This run has been done on the v6 
>>> series, but changes introduced in this new version should not impact the performance boost seen.
>>>
>>> Here are the perf data captured during 2 of these runs on top of the 
>>> v8
>>> series:
>>>                 vanilla         spf
>>> faults          89.418          101.364         +13%
>>> spf                n/a           97.989
>>>
>>> With the SPF kernel, most of the page fault were processed in a 
>>> speculative way.
>>>
>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel 
>>> and gave it a try on an android device. He reported that the 
>>> application launch time was improved in average by 6%, and for large 
>>> applications (~100 threads) by 20%.
>>>
>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a 
>>> Qcom
>>> MSM845 (8 cores) with 6GB (the less is better):
>>>
>>> Application                             4.9     4.9+spf delta
>>> com.tencent.mm                          416     389     -7%
>>> com.eg.android.AlipayGphone             1135    986     -13%
>>> com.tencent.mtt                         455     454     0%
>>> com.qqgame.hlddz                        1497    1409    -6%
>>> com.autonavi.minimap                    711     701     -1%
>>> com.tencent.tmgp.sgame                  788     748     -5%
>>> com.immomo.momo                         501     487     -3%
>>> com.tencent.peng                        2145    2112    -2%
>>> com.smile.gifmaker                      491     461     -6%
>>> com.baidu.BaiduMap                      479     366     -23%
>>> com.taobao.taobao                       1341    1198    -11%
>>> com.baidu.searchbox                     333     314     -6%
>>> com.tencent.mobileqq                    394     384     -3%
>>> com.sina.weibo                          907     906     0%
>>> com.youku.phone                         816     731     -11%
>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>> com.UCMobile                            415     411     -1%
>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>> com.tencent.qqmusic                     336     329     -2%
>>> com.sankuai.meituan                     1661    1302    -22%
>>> com.netease.cloudmusic                  1193    1200    1%
>>> air.tv.douyu.android                    4257    4152    -2%
>>>
>>> ------------------
>>> Benchmarks results
>>>
>>> Base kernel is v4.17.0-rc4-mm1
>>> SPF is BASE + this series
>>>
>>> Kernbench:
>>> ----------
>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 
>>> 4.15 kernel (kernel is build 5 times):
>>>
>>> Average Half load -j 8
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>
>>> Average Optimal load -j 16
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          526743764      faults
>>>                210      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>               2278      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Very few speculative page faults were recorded as most of the 
>>> processes involved are monothreaded (sounds that on this 
>>> architecture some threads were created during the kernel build processing).
>>>
>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>
>>> Average Half load -j 40
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>
>>> Average Optimal load -j 80
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          116730856      faults
>>>                  0      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                476      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Most of the processes involved are monothreaded so SPF is not 
>>> activated but there is no impact on the performance.
>>>
>>> Ebizzy:
>>> -------
>>> The test is counting the number of records per second it can manage, 
>>> the higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. 
>>> To get consistent result I repeated the test 100 times and measure 
>>> the average result. The number is the record processes per second, 
>>> the higher is the best.
>>>
>>>                 BASE            SPF             delta
>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>
>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>  Performance counter stats for './ebizzy -mTt 16':
>>>            1706379      faults
>>>            1674599      spf
>>>              30588      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                363      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> And the ones captured during a run on a 80 CPUs Power node:
>>>  Performance counter stats for './ebizzy -mTt 80':
>>>            1874773      faults
>>>            1461153      spf
>>>             413293      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                200      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> In ebizzy's case most of the page fault were handled in a 
>>> speculative way, leading the ebizzy performance boost.
>>>
>>> ------------------
>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>    and Minchan Kim, hopefully.
>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>    __do_page_fault().
>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>    instead
>>>    of aborting the speculative page fault handling. Dropping the now 
>>> useless
>>>    trace event pagefault:spf_pte_lock.
>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>    handling when retrying is needed. This adds a lot of complexity and
>>>    additional tests done didn't show a significant performance improvement.
>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>
>>> [1] 
>>> http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-s
>>> peculative-page-faults-tt965642.html#none
>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>
>>>
>>> Laurent Dufour (20):
>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>   mm: make pte_unmap_same compatible with SPF
>>>   mm: introduce INIT_VMA()
>>>   mm: protect VMA modifications using VMA sequence count
>>>   mm: protect mremap() against SPF hanlder
>>>   mm: protect SPF handler against anon_vma changes
>>>   mm: cache some VMA fields in the vm_fault structure
>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>   mm: introduce __vm_normal_page()
>>>   mm: introduce __page_add_new_anon_rmap()
>>>   mm: protect mm_rb tree with a rwlock
>>>   mm: adding speculative page fault failure trace events
>>>   perf: add a speculative page fault sw event
>>>   perf tools: add support for the SPF perf event
>>>   mm: add speculative page fault vmstats
>>>   powerpc/mm: add speculative page fault
>>>
>>> Mahendran Ganesh (2):
>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   arm64/mm: add speculative page fault
>>>
>>> Peter Zijlstra (4):
>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>   mm: VMA sequence count
>>>   mm: provide speculative fault infrastructure
>>>   x86/mm: add speculative pagefault handling
>>>
>>>  arch/arm64/Kconfig                    |   1 +
>>>  arch/arm64/mm/fault.c                 |  12 +
>>>  arch/powerpc/Kconfig                  |   1 +
>>>  arch/powerpc/mm/fault.c               |  16 +
>>>  arch/x86/Kconfig                      |   1 +
>>>  arch/x86/mm/fault.c                   |  27 +-
>>>  fs/exec.c                             |   2 +-
>>>  fs/proc/task_mmu.c                    |   5 +-
>>>  fs/userfaultfd.c                      |  17 +-
>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>  include/linux/migrate.h               |   4 +-
>>>  include/linux/mm.h                    | 136 +++++++-
>>>  include/linux/mm_types.h              |   7 +
>>>  include/linux/pagemap.h               |   4 +-
>>>  include/linux/rmap.h                  |  12 +-
>>>  include/linux/swap.h                  |  10 +-
>>>  include/linux/vm_event_item.h         |   3 +
>>>  include/trace/events/pagefault.h      |  80 +++++
>>>  include/uapi/linux/perf_event.h       |   1 +
>>>  kernel/fork.c                         |   5 +-
>>>  mm/Kconfig                            |  22 ++
>>>  mm/huge_memory.c                      |   6 +-
>>>  mm/hugetlb.c                          |   2 +
>>>  mm/init-mm.c                          |   3 +
>>>  mm/internal.h                         |  20 ++
>>>  mm/khugepaged.c                       |   5 +
>>>  mm/madvise.c                          |   6 +-
>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>  mm/mempolicy.c                        |  51 ++-
>>>  mm/migrate.c                          |   6 +-
>>>  mm/mlock.c                            |  13 +-
>>>  mm/mmap.c                             | 229 ++++++++++---
>>>  mm/mprotect.c                         |   4 +-
>>>  mm/mremap.c                           |  13 +
>>>  mm/nommu.c                            |   2 +-
>>>  mm/rmap.c                             |   5 +-
>>>  mm/swap.c                             |   6 +-
>>>  mm/swap_state.c                       |   8 +-
>>>  mm/vmstat.c                           |   5 +-
>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>  tools/perf/util/evsel.c               |   1 +
>>>  tools/perf/util/parse-events.c        |   4 +
>>>  tools/perf/util/parse-events.l        |   1 +
>>>  tools/perf/util/python.c              |   1 +
>>>  44 files changed, 1161 insertions(+), 211 deletions(-)  create mode 
>>> 100644 include/trace/events/pagefault.h
>>>
>>> --
>>> 2.7.4
>>>
>>>
>>
> 

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

On 28/05/2018 10:22, Haiyan Song wrote:
> Hi Laurent,
> 
> Yes, these tests are done on V9 patch.

Do you plan to give this V11 a run ?

> 
> 
> Best regards,
> Haiyan Song
> 
> On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
>> On 28/05/2018 07:23, Song, HaiyanX wrote:
>>>
>>> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
>>> tested on Intel 4s Skylake platform.
>>
>> Hi,
>>
>> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
>> series" while responding to the v11 header series...
>> Were these tests done on v9 or v11 ?
>>
>> Cheers,
>> Laurent.
>>
>>>
>>> The regression result is sorted by the metric will-it-scale.per_thread_ops.
>>> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
>>> Commit id:
>>>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>>>     head commit: 0355322b3577eeab7669066df42c550a56801110
>>> Benchmark suite: will-it-scale
>>> Download link:
>>> https://github.com/antonblanchard/will-it-scale/tree/master/tests
>>> Metrics:
>>>     will-it-scale.per_process_ops=processes/nr_cpu
>>>     will-it-scale.per_thread_ops=threads/nr_cpu
>>> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
>>> THP: enable / disable
>>> nr_task: 100%
>>>
>>> 1. Regressions:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
>>> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
>>> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
>>> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
>>> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
>>> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
>>> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
>>> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
>>> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
>>> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
>>>
>>> 2. Improvements:
>>> a) THP enabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
>>> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
>>> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
>>>
>>> b) THP disabled:
>>> testcase                        base            change          head       metric
>>> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
>>> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
>>> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
>>> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
>>> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
>>> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
>>>
>>> Notes: for above values in column "change", the higher value means that the related testcase result
>>> on head commit is better than that on base commit for this benchmark.
>>>
>>>
>>> Best regards
>>> Haiyan Song
>>>
>>> ________________________________________
>>> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
>>> Sent: Thursday, May 17, 2018 7:06 PM
>>> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
>>> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
>>> Subject: [PATCH v11 00/26] Speculative page faults
>>>
>>> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
>>> page fault without holding the mm semaphore [1].
>>>
>>> The idea is to try to handle user space page faults without holding the
>>> mmap_sem. This should allow better concurrency for massively threaded
>>> process since the page fault handler will not wait for other threads memory
>>> layout change to be done, assuming that this change is done in another part
>>> of the process's memory space. This type page fault is named speculative
>>> page fault. If the speculative page fault fails because of a concurrency is
>>> detected or because underlying PMD or PTE tables are not yet allocating, it
>>> is failing its processing and a classic page fault is then tried.
>>>
>>> The speculative page fault (SPF) has to look for the VMA matching the fault
>>> address without holding the mmap_sem, this is done by introducing a rwlock
>>> which protects the access to the mm_rb tree. Previously this was done using
>>> SRCU but it was introducing a lot of scheduling to process the VMA's
>>> freeing operation which was hitting the performance by 20% as reported by
>>> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
>>> limiting the locking contention to these operations which are expected to
>>> be in a O(log n) order. In addition to ensure that the VMA is not freed in
>>> our back a reference count is added and 2 services (get_vma() and
>>> put_vma()) are introduced to handle the reference count. Once a VMA is
>>> fetched from the RB tree using get_vma(), it must be later freed using
>>> put_vma(). I can't see anymore the overhead I got while will-it-scale
>>> benchmark anymore.
>>>
>>> The VMA's attributes checked during the speculative page fault processing
>>> have to be protected against parallel changes. This is done by using a per
>>> VMA sequence lock. This sequence lock allows the speculative page fault
>>> handler to fast check for parallel changes in progress and to abort the
>>> speculative page fault in that case.
>>>
>>> Once the VMA has been found, the speculative page fault handler would check
>>> for the VMA's attributes to verify that the page fault has to be handled
>>> correctly or not. Thus, the VMA is protected through a sequence lock which
>>> allows fast detection of concurrent VMA changes. If such a change is
>>> detected, the speculative page fault is aborted and a *classic* page fault
>>> is tried.  VMA sequence lockings are added when VMA attributes which are
>>> checked during the page fault are modified.
>>>
>>> When the PTE is fetched, the VMA is checked to see if it has been changed,
>>> so once the page table is locked, the VMA is valid, so any other changes
>>> leading to touching this PTE will need to lock the page table, so no
>>> parallel change is possible at this time.
>>>
>>> The locking of the PTE is done with interrupts disabled, this allows
>>> checking for the PMD to ensure that there is not an ongoing collapsing
>>> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
>>> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
>>> valid at the time the PTE is locked, we have the guarantee that the
>>> collapsing operation will have to wait on the PTE lock to move forward.
>>> This allows the SPF handler to map the PTE safely. If the PMD value is
>>> different from the one recorded at the beginning of the SPF operation, the
>>> classic page fault handler will be called to handle the operation while
>>> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
>>> the lock is done using spin_trylock() to avoid dead lock when handling a
>>> page fault while a TLB invalidate is requested by another CPU holding the
>>> PTE.
>>>
>>> In pseudo code, this could be seen as:
>>>     speculative_page_fault()
>>>     {
>>>             vma = get_vma()
>>>             check vma sequence count
>>>             check vma's support
>>>             disable interrupt
>>>                   check pgd,p4d,...,pte
>>>                   save pmd and pte in vmf
>>>                   save vma sequence counter in vmf
>>>             enable interrupt
>>>             check vma sequence count
>>>             handle_pte_fault(vma)
>>>                     ..
>>>                     page = alloc_page()
>>>                     pte_map_lock()
>>>                             disable interrupt
>>>                                     abort if sequence counter has changed
>>>                                     abort if pmd or pte has changed
>>>                                     pte map and lock
>>>                             enable interrupt
>>>                     if abort
>>>                        free page
>>>                        abort
>>>                     ...
>>>     }
>>>
>>>     arch_fault_handler()
>>>     {
>>>             if (speculative_page_fault(&vma))
>>>                goto done
>>>     again:
>>>             lock(mmap_sem)
>>>             vma = find_vma();
>>>             handle_pte_fault(vma);
>>>             if retry
>>>                unlock(mmap_sem)
>>>                goto again;
>>>     done:
>>>             handle fault error
>>>     }
>>>
>>> Support for THP is not done because when checking for the PMD, we can be
>>> confused by an in progress collapsing operation done by khugepaged. The
>>> issue is that pmd_none() could be true either if the PMD is not already
>>> populated or if the underlying PTE are in the way to be collapsed. So we
>>> cannot safely allocate a PMD if pmd_none() is true.
>>>
>>> This series add a new software performance event named 'speculative-faults'
>>> or 'spf'. It counts the number of successful page fault event handled
>>> speculatively. When recording 'faults,spf' events, the faults one is
>>> counting the total number of page fault events while 'spf' is only counting
>>> the part of the faults processed speculatively.
>>>
>>> There are some trace events introduced by this series. They allow
>>> identifying why the page faults were not processed speculatively. This
>>> doesn't take in account the faults generated by a monothreaded process
>>> which directly processed while holding the mmap_sem. This trace events are
>>> grouped in a system named 'pagefault', they are:
>>>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>>>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>>>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>>>  - pagefault:spf_vma_access : the VMA's access right are not respected
>>>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>>>    back.
>>>
>>> To record all the related events, the easier is to run perf with the
>>> following arguments :
>>> $ perf stat -e 'faults,spf,pagefault:*' <command>
>>>
>>> There is also a dedicated vmstat counter showing the number of successful
>>> page fault handled speculatively. I can be seen this way:
>>> $ grep speculative_pgfault /proc/vmstat
>>>
>>> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
>>> on x86, PowerPC and arm64.
>>>
>>> ---------------------
>>> Real Workload results
>>>
>>> As mentioned in previous email, we did non official runs using a "popular
>>> in memory multithreaded database product" on 176 cores SMT8 Power system
>>> which showed a 30% improvements in the number of transaction processed per
>>> second. This run has been done on the v6 series, but changes introduced in
>>> this new version should not impact the performance boost seen.
>>>
>>> Here are the perf data captured during 2 of these runs on top of the v8
>>> series:
>>>                 vanilla         spf
>>> faults          89.418          101.364         +13%
>>> spf                n/a           97.989
>>>
>>> With the SPF kernel, most of the page fault were processed in a speculative
>>> way.
>>>
>>> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
>>> it a try on an android device. He reported that the application launch time
>>> was improved in average by 6%, and for large applications (~100 threads) by
>>> 20%.
>>>
>>> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
>>> MSM845 (8 cores) with 6GB (the less is better):
>>>
>>> Application                             4.9     4.9+spf delta
>>> com.tencent.mm                          416     389     -7%
>>> com.eg.android.AlipayGphone             1135    986     -13%
>>> com.tencent.mtt                         455     454     0%
>>> com.qqgame.hlddz                        1497    1409    -6%
>>> com.autonavi.minimap                    711     701     -1%
>>> com.tencent.tmgp.sgame                  788     748     -5%
>>> com.immomo.momo                         501     487     -3%
>>> com.tencent.peng                        2145    2112    -2%
>>> com.smile.gifmaker                      491     461     -6%
>>> com.baidu.BaiduMap                      479     366     -23%
>>> com.taobao.taobao                       1341    1198    -11%
>>> com.baidu.searchbox                     333     314     -6%
>>> com.tencent.mobileqq                    394     384     -3%
>>> com.sina.weibo                          907     906     0%
>>> com.youku.phone                         816     731     -11%
>>> com.happyelements.AndroidAnimal.qq      763     717     -6%
>>> com.UCMobile                            415     411     -1%
>>> com.tencent.tmgp.ak                     1464    1431    -2%
>>> com.tencent.qqmusic                     336     329     -2%
>>> com.sankuai.meituan                     1661    1302    -22%
>>> com.netease.cloudmusic                  1193    1200    1%
>>> air.tv.douyu.android                    4257    4152    -2%
>>>
>>> ------------------
>>> Benchmarks results
>>>
>>> Base kernel is v4.17.0-rc4-mm1
>>> SPF is BASE + this series
>>>
>>> Kernbench:
>>> ----------
>>> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
>>> kernel (kernel is build 5 times):
>>>
>>> Average Half load -j 8
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
>>> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
>>> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
>>> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
>>> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
>>> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
>>>
>>> Average Optimal load -j 16
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
>>> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
>>> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
>>> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
>>> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
>>> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
>>>
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          526743764      faults
>>>                210      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>               2278      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Very few speculative page faults were recorded as most of the processes
>>> involved are monothreaded (sounds that on this architecture some threads
>>> were created during the kernel build processing).
>>>
>>> Here are the kerbench results on a 80 CPUs Power8 system:
>>>
>>> Average Half load -j 40
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
>>> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
>>> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
>>> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
>>> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
>>> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
>>>
>>> Average Optimal load -j 80
>>>                  Run    (std deviation)
>>>                  BASE                   SPF
>>> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
>>> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
>>> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
>>> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
>>> Context Switches 223861  (138865)       225032  (139632)        0.52%
>>> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
>>>
>>> During a run on the SPF, perf events were captured:
>>>  Performance counter stats for '../kernbench -M':
>>>          116730856      faults
>>>                  0      spf
>>>                  3      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                476      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> Most of the processes involved are monothreaded so SPF is not activated but
>>> there is no impact on the performance.
>>>
>>> Ebizzy:
>>> -------
>>> The test is counting the number of records per second it can manage, the
>>> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
>>> consistent result I repeated the test 100 times and measure the average
>>> result. The number is the record processes per second, the higher is the
>>> best.
>>>
>>>                 BASE            SPF             delta
>>> 16 CPUs x86 VM  742.57          1490.24         100.69%
>>> 80 CPUs P8 node 13105.4         24174.23        84.46%
>>>
>>> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>>>  Performance counter stats for './ebizzy -mTt 16':
>>>            1706379      faults
>>>            1674599      spf
>>>              30588      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                363      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> And the ones captured during a run on a 80 CPUs Power node:
>>>  Performance counter stats for './ebizzy -mTt 80':
>>>            1874773      faults
>>>            1461153      spf
>>>             413293      pagefault:spf_vma_changed
>>>                  0      pagefault:spf_vma_noanon
>>>                200      pagefault:spf_vma_notsup
>>>                  0      pagefault:spf_vma_access
>>>                  0      pagefault:spf_pmd_changed
>>>
>>> In ebizzy's case most of the page fault were handled in a speculative way,
>>> leading the ebizzy performance boost.
>>>
>>> ------------------
>>> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>>>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>>>    and Minchan Kim, hopefully.
>>>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>>>    __do_page_fault().
>>>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>>>    instead
>>>    of aborting the speculative page fault handling. Dropping the now
>>> useless
>>>    trace event pagefault:spf_pte_lock.
>>>  - No more try to reuse the fetched VMA during the speculative page fault
>>>    handling when retrying is needed. This adds a lot of complexity and
>>>    additional tests done didn't show a significant performance improvement.
>>>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
>>>
>>> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
>>> [2] https://patchwork.kernel.org/patch/9999687/
>>>
>>>
>>> Laurent Dufour (20):
>>>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>>>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>>>   mm: make pte_unmap_same compatible with SPF
>>>   mm: introduce INIT_VMA()
>>>   mm: protect VMA modifications using VMA sequence count
>>>   mm: protect mremap() against SPF hanlder
>>>   mm: protect SPF handler against anon_vma changes
>>>   mm: cache some VMA fields in the vm_fault structure
>>>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>>>   mm: introduce __lru_cache_add_active_or_unevictable
>>>   mm: introduce __vm_normal_page()
>>>   mm: introduce __page_add_new_anon_rmap()
>>>   mm: protect mm_rb tree with a rwlock
>>>   mm: adding speculative page fault failure trace events
>>>   perf: add a speculative page fault sw event
>>>   perf tools: add support for the SPF perf event
>>>   mm: add speculative page fault vmstats
>>>   powerpc/mm: add speculative page fault
>>>
>>> Mahendran Ganesh (2):
>>>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>>>   arm64/mm: add speculative page fault
>>>
>>> Peter Zijlstra (4):
>>>   mm: prepare for FAULT_FLAG_SPECULATIVE
>>>   mm: VMA sequence count
>>>   mm: provide speculative fault infrastructure
>>>   x86/mm: add speculative pagefault handling
>>>
>>>  arch/arm64/Kconfig                    |   1 +
>>>  arch/arm64/mm/fault.c                 |  12 +
>>>  arch/powerpc/Kconfig                  |   1 +
>>>  arch/powerpc/mm/fault.c               |  16 +
>>>  arch/x86/Kconfig                      |   1 +
>>>  arch/x86/mm/fault.c                   |  27 +-
>>>  fs/exec.c                             |   2 +-
>>>  fs/proc/task_mmu.c                    |   5 +-
>>>  fs/userfaultfd.c                      |  17 +-
>>>  include/linux/hugetlb_inline.h        |   2 +-
>>>  include/linux/migrate.h               |   4 +-
>>>  include/linux/mm.h                    | 136 +++++++-
>>>  include/linux/mm_types.h              |   7 +
>>>  include/linux/pagemap.h               |   4 +-
>>>  include/linux/rmap.h                  |  12 +-
>>>  include/linux/swap.h                  |  10 +-
>>>  include/linux/vm_event_item.h         |   3 +
>>>  include/trace/events/pagefault.h      |  80 +++++
>>>  include/uapi/linux/perf_event.h       |   1 +
>>>  kernel/fork.c                         |   5 +-
>>>  mm/Kconfig                            |  22 ++
>>>  mm/huge_memory.c                      |   6 +-
>>>  mm/hugetlb.c                          |   2 +
>>>  mm/init-mm.c                          |   3 +
>>>  mm/internal.h                         |  20 ++
>>>  mm/khugepaged.c                       |   5 +
>>>  mm/madvise.c                          |   6 +-
>>>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>>>  mm/mempolicy.c                        |  51 ++-
>>>  mm/migrate.c                          |   6 +-
>>>  mm/mlock.c                            |  13 +-
>>>  mm/mmap.c                             | 229 ++++++++++---
>>>  mm/mprotect.c                         |   4 +-
>>>  mm/mremap.c                           |  13 +
>>>  mm/nommu.c                            |   2 +-
>>>  mm/rmap.c                             |   5 +-
>>>  mm/swap.c                             |   6 +-
>>>  mm/swap_state.c                       |   8 +-
>>>  mm/vmstat.c                           |   5 +-
>>>  tools/include/uapi/linux/perf_event.h |   1 +
>>>  tools/perf/util/evsel.c               |   1 +
>>>  tools/perf/util/parse-events.c        |   4 +
>>>  tools/perf/util/parse-events.l        |   1 +
>>>  tools/perf/util/python.c              |   1 +
>>>  44 files changed, 1161 insertions(+), 211 deletions(-)
>>>  create mode 100644 include/trace/events/pagefault.h
>>>
>>> --
>>> 2.7.4
>>>
>>>
>>
> 

Re: [PATCH v11 00/26] Speculative page faults

[Haiyan Song]

Hi Laurent,

Yes, these tests are done on V9 patch.


Best regards,
Haiyan Song

On Mon, May 28, 2018 at 09:51:34AM +0200, Laurent Dufour wrote:
> On 28/05/2018 07:23, Song, HaiyanX wrote:
> > 
> > Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
> > tested on Intel 4s Skylake platform.
> 
> Hi,
> 
> Thanks for reporting this benchmark results, but you mentioned the "V9 patch
> series" while responding to the v11 header series...
> Were these tests done on v9 or v11 ?
> 
> Cheers,
> Laurent.
> 
> > 
> > The regression result is sorted by the metric will-it-scale.per_thread_ops.
> > Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
> > Commit id:
> >     base commit: d55f34411b1b126429a823d06c3124c16283231f
> >     head commit: 0355322b3577eeab7669066df42c550a56801110
> > Benchmark suite: will-it-scale
> > Download link:
> > https://github.com/antonblanchard/will-it-scale/tree/master/tests
> > Metrics:
> >     will-it-scale.per_process_ops=processes/nr_cpu
> >     will-it-scale.per_thread_ops=threads/nr_cpu
> > test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> > THP: enable / disable
> > nr_task: 100%
> > 
> > 1. Regressions:
> > a) THP enabled:
> > testcase                        base            change          head       metric
> > page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
> > page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
> > brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
> > page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
> > signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
> > 
> > b) THP disabled:
> > testcase                        base            change          head       metric
> > page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
> > page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
> > context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
> > brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
> > page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
> > signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
> > 
> > 2. Improvements:
> > a) THP enabled:
> > testcase                        base            change          head       metric
> > malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
> > writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
> > signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
> > 
> > b) THP disabled:
> > testcase                        base            change          head       metric
> > malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
> > read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
> > page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
> > read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
> > writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
> > signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
> > 
> > Notes: for above values in column "change", the higher value means that the related testcase result
> > on head commit is better than that on base commit for this benchmark.
> > 
> > 
> > Best regards
> > Haiyan Song
> > 
> > ________________________________________
> > From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> > Sent: Thursday, May 17, 2018 7:06 PM
> > To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
> > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> > Subject: [PATCH v11 00/26] Speculative page faults
> > 
> > This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
> > page fault without holding the mm semaphore [1].
> > 
> > The idea is to try to handle user space page faults without holding the
> > mmap_sem. This should allow better concurrency for massively threaded
> > process since the page fault handler will not wait for other threads memory
> > layout change to be done, assuming that this change is done in another part
> > of the process's memory space. This type page fault is named speculative
> > page fault. If the speculative page fault fails because of a concurrency is
> > detected or because underlying PMD or PTE tables are not yet allocating, it
> > is failing its processing and a classic page fault is then tried.
> > 
> > The speculative page fault (SPF) has to look for the VMA matching the fault
> > address without holding the mmap_sem, this is done by introducing a rwlock
> > which protects the access to the mm_rb tree. Previously this was done using
> > SRCU but it was introducing a lot of scheduling to process the VMA's
> > freeing operation which was hitting the performance by 20% as reported by
> > Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
> > limiting the locking contention to these operations which are expected to
> > be in a O(log n) order. In addition to ensure that the VMA is not freed in
> > our back a reference count is added and 2 services (get_vma() and
> > put_vma()) are introduced to handle the reference count. Once a VMA is
> > fetched from the RB tree using get_vma(), it must be later freed using
> > put_vma(). I can't see anymore the overhead I got while will-it-scale
> > benchmark anymore.
> > 
> > The VMA's attributes checked during the speculative page fault processing
> > have to be protected against parallel changes. This is done by using a per
> > VMA sequence lock. This sequence lock allows the speculative page fault
> > handler to fast check for parallel changes in progress and to abort the
> > speculative page fault in that case.
> > 
> > Once the VMA has been found, the speculative page fault handler would check
> > for the VMA's attributes to verify that the page fault has to be handled
> > correctly or not. Thus, the VMA is protected through a sequence lock which
> > allows fast detection of concurrent VMA changes. If such a change is
> > detected, the speculative page fault is aborted and a *classic* page fault
> > is tried.  VMA sequence lockings are added when VMA attributes which are
> > checked during the page fault are modified.
> > 
> > When the PTE is fetched, the VMA is checked to see if it has been changed,
> > so once the page table is locked, the VMA is valid, so any other changes
> > leading to touching this PTE will need to lock the page table, so no
> > parallel change is possible at this time.
> > 
> > The locking of the PTE is done with interrupts disabled, this allows
> > checking for the PMD to ensure that there is not an ongoing collapsing
> > operation. Since khugepaged is firstly set the PMD to pmd_none and then is
> > waiting for the other CPU to have caught the IPI interrupt, if the pmd is
> > valid at the time the PTE is locked, we have the guarantee that the
> > collapsing operation will have to wait on the PTE lock to move forward.
> > This allows the SPF handler to map the PTE safely. If the PMD value is
> > different from the one recorded at the beginning of the SPF operation, the
> > classic page fault handler will be called to handle the operation while
> > holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
> > the lock is done using spin_trylock() to avoid dead lock when handling a
> > page fault while a TLB invalidate is requested by another CPU holding the
> > PTE.
> > 
> > In pseudo code, this could be seen as:
> >     speculative_page_fault()
> >     {
> >             vma = get_vma()
> >             check vma sequence count
> >             check vma's support
> >             disable interrupt
> >                   check pgd,p4d,...,pte
> >                   save pmd and pte in vmf
> >                   save vma sequence counter in vmf
> >             enable interrupt
> >             check vma sequence count
> >             handle_pte_fault(vma)
> >                     ..
> >                     page = alloc_page()
> >                     pte_map_lock()
> >                             disable interrupt
> >                                     abort if sequence counter has changed
> >                                     abort if pmd or pte has changed
> >                                     pte map and lock
> >                             enable interrupt
> >                     if abort
> >                        free page
> >                        abort
> >                     ...
> >     }
> > 
> >     arch_fault_handler()
> >     {
> >             if (speculative_page_fault(&vma))
> >                goto done
> >     again:
> >             lock(mmap_sem)
> >             vma = find_vma();
> >             handle_pte_fault(vma);
> >             if retry
> >                unlock(mmap_sem)
> >                goto again;
> >     done:
> >             handle fault error
> >     }
> > 
> > Support for THP is not done because when checking for the PMD, we can be
> > confused by an in progress collapsing operation done by khugepaged. The
> > issue is that pmd_none() could be true either if the PMD is not already
> > populated or if the underlying PTE are in the way to be collapsed. So we
> > cannot safely allocate a PMD if pmd_none() is true.
> > 
> > This series add a new software performance event named 'speculative-faults'
> > or 'spf'. It counts the number of successful page fault event handled
> > speculatively. When recording 'faults,spf' events, the faults one is
> > counting the total number of page fault events while 'spf' is only counting
> > the part of the faults processed speculatively.
> > 
> > There are some trace events introduced by this series. They allow
> > identifying why the page faults were not processed speculatively. This
> > doesn't take in account the faults generated by a monothreaded process
> > which directly processed while holding the mmap_sem. This trace events are
> > grouped in a system named 'pagefault', they are:
> >  - pagefault:spf_vma_changed : if the VMA has been changed in our back
> >  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
> >  - pagefault:spf_vma_notsup : the VMA's type is not supported
> >  - pagefault:spf_vma_access : the VMA's access right are not respected
> >  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
> >    back.
> > 
> > To record all the related events, the easier is to run perf with the
> > following arguments :
> > $ perf stat -e 'faults,spf,pagefault:*' <command>
> > 
> > There is also a dedicated vmstat counter showing the number of successful
> > page fault handled speculatively. I can be seen this way:
> > $ grep speculative_pgfault /proc/vmstat
> > 
> > This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
> > on x86, PowerPC and arm64.
> > 
> > ---------------------
> > Real Workload results
> > 
> > As mentioned in previous email, we did non official runs using a "popular
> > in memory multithreaded database product" on 176 cores SMT8 Power system
> > which showed a 30% improvements in the number of transaction processed per
> > second. This run has been done on the v6 series, but changes introduced in
> > this new version should not impact the performance boost seen.
> > 
> > Here are the perf data captured during 2 of these runs on top of the v8
> > series:
> >                 vanilla         spf
> > faults          89.418          101.364         +13%
> > spf                n/a           97.989
> > 
> > With the SPF kernel, most of the page fault were processed in a speculative
> > way.
> > 
> > Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
> > it a try on an android device. He reported that the application launch time
> > was improved in average by 6%, and for large applications (~100 threads) by
> > 20%.
> > 
> > Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
> > MSM845 (8 cores) with 6GB (the less is better):
> > 
> > Application                             4.9     4.9+spf delta
> > com.tencent.mm                          416     389     -7%
> > com.eg.android.AlipayGphone             1135    986     -13%
> > com.tencent.mtt                         455     454     0%
> > com.qqgame.hlddz                        1497    1409    -6%
> > com.autonavi.minimap                    711     701     -1%
> > com.tencent.tmgp.sgame                  788     748     -5%
> > com.immomo.momo                         501     487     -3%
> > com.tencent.peng                        2145    2112    -2%
> > com.smile.gifmaker                      491     461     -6%
> > com.baidu.BaiduMap                      479     366     -23%
> > com.taobao.taobao                       1341    1198    -11%
> > com.baidu.searchbox                     333     314     -6%
> > com.tencent.mobileqq                    394     384     -3%
> > com.sina.weibo                          907     906     0%
> > com.youku.phone                         816     731     -11%
> > com.happyelements.AndroidAnimal.qq      763     717     -6%
> > com.UCMobile                            415     411     -1%
> > com.tencent.tmgp.ak                     1464    1431    -2%
> > com.tencent.qqmusic                     336     329     -2%
> > com.sankuai.meituan                     1661    1302    -22%
> > com.netease.cloudmusic                  1193    1200    1%
> > air.tv.douyu.android                    4257    4152    -2%
> > 
> > ------------------
> > Benchmarks results
> > 
> > Base kernel is v4.17.0-rc4-mm1
> > SPF is BASE + this series
> > 
> > Kernbench:
> > ----------
> > Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
> > kernel (kernel is build 5 times):
> > 
> > Average Half load -j 8
> >                  Run    (std deviation)
> >                  BASE                   SPF
> > Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
> > User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
> > System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
> > Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
> > Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
> > Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
> > 
> > Average Optimal load -j 16
> >                  Run    (std deviation)
> >                  BASE                   SPF
> > Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
> > User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
> > System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
> > Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
> > Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
> > Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
> > 
> > 
> > During a run on the SPF, perf events were captured:
> >  Performance counter stats for '../kernbench -M':
> >          526743764      faults
> >                210      spf
> >                  3      pagefault:spf_vma_changed
> >                  0      pagefault:spf_vma_noanon
> >               2278      pagefault:spf_vma_notsup
> >                  0      pagefault:spf_vma_access
> >                  0      pagefault:spf_pmd_changed
> > 
> > Very few speculative page faults were recorded as most of the processes
> > involved are monothreaded (sounds that on this architecture some threads
> > were created during the kernel build processing).
> > 
> > Here are the kerbench results on a 80 CPUs Power8 system:
> > 
> > Average Half load -j 40
> >                  Run    (std deviation)
> >                  BASE                   SPF
> > Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
> > User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
> > System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
> > Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
> > Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
> > Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
> > 
> > Average Optimal load -j 80
> >                  Run    (std deviation)
> >                  BASE                   SPF
> > Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
> > User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
> > System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
> > Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
> > Context Switches 223861  (138865)       225032  (139632)        0.52%
> > Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
> > 
> > During a run on the SPF, perf events were captured:
> >  Performance counter stats for '../kernbench -M':
> >          116730856      faults
> >                  0      spf
> >                  3      pagefault:spf_vma_changed
> >                  0      pagefault:spf_vma_noanon
> >                476      pagefault:spf_vma_notsup
> >                  0      pagefault:spf_vma_access
> >                  0      pagefault:spf_pmd_changed
> > 
> > Most of the processes involved are monothreaded so SPF is not activated but
> > there is no impact on the performance.
> > 
> > Ebizzy:
> > -------
> > The test is counting the number of records per second it can manage, the
> > higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
> > consistent result I repeated the test 100 times and measure the average
> > result. The number is the record processes per second, the higher is the
> > best.
> > 
> >                 BASE            SPF             delta
> > 16 CPUs x86 VM  742.57          1490.24         100.69%
> > 80 CPUs P8 node 13105.4         24174.23        84.46%
> > 
> > Here are the performance counter read during a run on a 16 CPUs x86 VM:
> >  Performance counter stats for './ebizzy -mTt 16':
> >            1706379      faults
> >            1674599      spf
> >              30588      pagefault:spf_vma_changed
> >                  0      pagefault:spf_vma_noanon
> >                363      pagefault:spf_vma_notsup
> >                  0      pagefault:spf_vma_access
> >                  0      pagefault:spf_pmd_changed
> > 
> > And the ones captured during a run on a 80 CPUs Power node:
> >  Performance counter stats for './ebizzy -mTt 80':
> >            1874773      faults
> >            1461153      spf
> >             413293      pagefault:spf_vma_changed
> >                  0      pagefault:spf_vma_noanon
> >                200      pagefault:spf_vma_notsup
> >                  0      pagefault:spf_vma_access
> >                  0      pagefault:spf_pmd_changed
> > 
> > In ebizzy's case most of the page fault were handled in a speculative way,
> > leading the ebizzy performance boost.
> > 
> > ------------------
> > Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
> >  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
> >    and Minchan Kim, hopefully.
> >  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
> >    __do_page_fault().
> >  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
> >    instead
> >    of aborting the speculative page fault handling. Dropping the now
> > useless
> >    trace event pagefault:spf_pte_lock.
> >  - No more try to reuse the fetched VMA during the speculative page fault
> >    handling when retrying is needed. This adds a lot of complexity and
> >    additional tests done didn't show a significant performance improvement.
> >  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
> > 
> > [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
> > [2] https://patchwork.kernel.org/patch/9999687/
> > 
> > 
> > Laurent Dufour (20):
> >   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
> >   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
> >   mm: make pte_unmap_same compatible with SPF
> >   mm: introduce INIT_VMA()
> >   mm: protect VMA modifications using VMA sequence count
> >   mm: protect mremap() against SPF hanlder
> >   mm: protect SPF handler against anon_vma changes
> >   mm: cache some VMA fields in the vm_fault structure
> >   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
> >   mm: introduce __lru_cache_add_active_or_unevictable
> >   mm: introduce __vm_normal_page()
> >   mm: introduce __page_add_new_anon_rmap()
> >   mm: protect mm_rb tree with a rwlock
> >   mm: adding speculative page fault failure trace events
> >   perf: add a speculative page fault sw event
> >   perf tools: add support for the SPF perf event
> >   mm: add speculative page fault vmstats
> >   powerpc/mm: add speculative page fault
> > 
> > Mahendran Ganesh (2):
> >   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
> >   arm64/mm: add speculative page fault
> > 
> > Peter Zijlstra (4):
> >   mm: prepare for FAULT_FLAG_SPECULATIVE
> >   mm: VMA sequence count
> >   mm: provide speculative fault infrastructure
> >   x86/mm: add speculative pagefault handling
> > 
> >  arch/arm64/Kconfig                    |   1 +
> >  arch/arm64/mm/fault.c                 |  12 +
> >  arch/powerpc/Kconfig                  |   1 +
> >  arch/powerpc/mm/fault.c               |  16 +
> >  arch/x86/Kconfig                      |   1 +
> >  arch/x86/mm/fault.c                   |  27 +-
> >  fs/exec.c                             |   2 +-
> >  fs/proc/task_mmu.c                    |   5 +-
> >  fs/userfaultfd.c                      |  17 +-
> >  include/linux/hugetlb_inline.h        |   2 +-
> >  include/linux/migrate.h               |   4 +-
> >  include/linux/mm.h                    | 136 +++++++-
> >  include/linux/mm_types.h              |   7 +
> >  include/linux/pagemap.h               |   4 +-
> >  include/linux/rmap.h                  |  12 +-
> >  include/linux/swap.h                  |  10 +-
> >  include/linux/vm_event_item.h         |   3 +
> >  include/trace/events/pagefault.h      |  80 +++++
> >  include/uapi/linux/perf_event.h       |   1 +
> >  kernel/fork.c                         |   5 +-
> >  mm/Kconfig                            |  22 ++
> >  mm/huge_memory.c                      |   6 +-
> >  mm/hugetlb.c                          |   2 +
> >  mm/init-mm.c                          |   3 +
> >  mm/internal.h                         |  20 ++
> >  mm/khugepaged.c                       |   5 +
> >  mm/madvise.c                          |   6 +-
> >  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
> >  mm/mempolicy.c                        |  51 ++-
> >  mm/migrate.c                          |   6 +-
> >  mm/mlock.c                            |  13 +-
> >  mm/mmap.c                             | 229 ++++++++++---
> >  mm/mprotect.c                         |   4 +-
> >  mm/mremap.c                           |  13 +
> >  mm/nommu.c                            |   2 +-
> >  mm/rmap.c                             |   5 +-
> >  mm/swap.c                             |   6 +-
> >  mm/swap_state.c                       |   8 +-
> >  mm/vmstat.c                           |   5 +-
> >  tools/include/uapi/linux/perf_event.h |   1 +
> >  tools/perf/util/evsel.c               |   1 +
> >  tools/perf/util/parse-events.c        |   4 +
> >  tools/perf/util/parse-events.l        |   1 +
> >  tools/perf/util/python.c              |   1 +
> >  44 files changed, 1161 insertions(+), 211 deletions(-)
> >  create mode 100644 include/trace/events/pagefault.h
> > 
> > --
> > 2.7.4
> > 
> > 
> 

Re: [PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

On 28/05/2018 07:23, Song, HaiyanX wrote:
> 
> Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
> tested on Intel 4s Skylake platform.

Hi,

Thanks for reporting this benchmark results, but you mentioned the "V9 patch
series" while responding to the v11 header series...
Were these tests done on v9 or v11 ?

Cheers,
Laurent.

> 
> The regression result is sorted by the metric will-it-scale.per_thread_ops.
> Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
> Commit id:
>     base commit: d55f34411b1b126429a823d06c3124c16283231f
>     head commit: 0355322b3577eeab7669066df42c550a56801110
> Benchmark suite: will-it-scale
> Download link:
> https://github.com/antonblanchard/will-it-scale/tree/master/tests
> Metrics:
>     will-it-scale.per_process_ops=processes/nr_cpu
>     will-it-scale.per_thread_ops=threads/nr_cpu
> test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
> THP: enable / disable
> nr_task: 100%
> 
> 1. Regressions:
> a) THP enabled:
> testcase                        base            change          head       metric
> page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
> page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
> brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
> page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
> signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops
> 
> b) THP disabled:
> testcase                        base            change          head       metric
> page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
> page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
> context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
> brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
> page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
> signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops
> 
> 2. Improvements:
> a) THP enabled:
> testcase                        base            change          head       metric
> malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
> writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
> signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops
> 
> b) THP disabled:
> testcase                        base            change          head       metric
> malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
> read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
> page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
> read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
> writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
> signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops
> 
> Notes: for above values in column "change", the higher value means that the related testcase result
> on head commit is better than that on base commit for this benchmark.
> 
> 
> Best regards
> Haiyan Song
> 
> ________________________________________
> From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
> Sent: Thursday, May 17, 2018 7:06 PM
> To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
> Subject: [PATCH v11 00/26] Speculative page faults
> 
> This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
> page fault without holding the mm semaphore [1].
> 
> The idea is to try to handle user space page faults without holding the
> mmap_sem. This should allow better concurrency for massively threaded
> process since the page fault handler will not wait for other threads memory
> layout change to be done, assuming that this change is done in another part
> of the process's memory space. This type page fault is named speculative
> page fault. If the speculative page fault fails because of a concurrency is
> detected or because underlying PMD or PTE tables are not yet allocating, it
> is failing its processing and a classic page fault is then tried.
> 
> The speculative page fault (SPF) has to look for the VMA matching the fault
> address without holding the mmap_sem, this is done by introducing a rwlock
> which protects the access to the mm_rb tree. Previously this was done using
> SRCU but it was introducing a lot of scheduling to process the VMA's
> freeing operation which was hitting the performance by 20% as reported by
> Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
> limiting the locking contention to these operations which are expected to
> be in a O(log n) order. In addition to ensure that the VMA is not freed in
> our back a reference count is added and 2 services (get_vma() and
> put_vma()) are introduced to handle the reference count. Once a VMA is
> fetched from the RB tree using get_vma(), it must be later freed using
> put_vma(). I can't see anymore the overhead I got while will-it-scale
> benchmark anymore.
> 
> The VMA's attributes checked during the speculative page fault processing
> have to be protected against parallel changes. This is done by using a per
> VMA sequence lock. This sequence lock allows the speculative page fault
> handler to fast check for parallel changes in progress and to abort the
> speculative page fault in that case.
> 
> Once the VMA has been found, the speculative page fault handler would check
> for the VMA's attributes to verify that the page fault has to be handled
> correctly or not. Thus, the VMA is protected through a sequence lock which
> allows fast detection of concurrent VMA changes. If such a change is
> detected, the speculative page fault is aborted and a *classic* page fault
> is tried.  VMA sequence lockings are added when VMA attributes which are
> checked during the page fault are modified.
> 
> When the PTE is fetched, the VMA is checked to see if it has been changed,
> so once the page table is locked, the VMA is valid, so any other changes
> leading to touching this PTE will need to lock the page table, so no
> parallel change is possible at this time.
> 
> The locking of the PTE is done with interrupts disabled, this allows
> checking for the PMD to ensure that there is not an ongoing collapsing
> operation. Since khugepaged is firstly set the PMD to pmd_none and then is
> waiting for the other CPU to have caught the IPI interrupt, if the pmd is
> valid at the time the PTE is locked, we have the guarantee that the
> collapsing operation will have to wait on the PTE lock to move forward.
> This allows the SPF handler to map the PTE safely. If the PMD value is
> different from the one recorded at the beginning of the SPF operation, the
> classic page fault handler will be called to handle the operation while
> holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
> the lock is done using spin_trylock() to avoid dead lock when handling a
> page fault while a TLB invalidate is requested by another CPU holding the
> PTE.
> 
> In pseudo code, this could be seen as:
>     speculative_page_fault()
>     {
>             vma = get_vma()
>             check vma sequence count
>             check vma's support
>             disable interrupt
>                   check pgd,p4d,...,pte
>                   save pmd and pte in vmf
>                   save vma sequence counter in vmf
>             enable interrupt
>             check vma sequence count
>             handle_pte_fault(vma)
>                     ..
>                     page = alloc_page()
>                     pte_map_lock()
>                             disable interrupt
>                                     abort if sequence counter has changed
>                                     abort if pmd or pte has changed
>                                     pte map and lock
>                             enable interrupt
>                     if abort
>                        free page
>                        abort
>                     ...
>     }
> 
>     arch_fault_handler()
>     {
>             if (speculative_page_fault(&vma))
>                goto done
>     again:
>             lock(mmap_sem)
>             vma = find_vma();
>             handle_pte_fault(vma);
>             if retry
>                unlock(mmap_sem)
>                goto again;
>     done:
>             handle fault error
>     }
> 
> Support for THP is not done because when checking for the PMD, we can be
> confused by an in progress collapsing operation done by khugepaged. The
> issue is that pmd_none() could be true either if the PMD is not already
> populated or if the underlying PTE are in the way to be collapsed. So we
> cannot safely allocate a PMD if pmd_none() is true.
> 
> This series add a new software performance event named 'speculative-faults'
> or 'spf'. It counts the number of successful page fault event handled
> speculatively. When recording 'faults,spf' events, the faults one is
> counting the total number of page fault events while 'spf' is only counting
> the part of the faults processed speculatively.
> 
> There are some trace events introduced by this series. They allow
> identifying why the page faults were not processed speculatively. This
> doesn't take in account the faults generated by a monothreaded process
> which directly processed while holding the mmap_sem. This trace events are
> grouped in a system named 'pagefault', they are:
>  - pagefault:spf_vma_changed : if the VMA has been changed in our back
>  - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
>  - pagefault:spf_vma_notsup : the VMA's type is not supported
>  - pagefault:spf_vma_access : the VMA's access right are not respected
>  - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
>    back.
> 
> To record all the related events, the easier is to run perf with the
> following arguments :
> $ perf stat -e 'faults,spf,pagefault:*' <command>
> 
> There is also a dedicated vmstat counter showing the number of successful
> page fault handled speculatively. I can be seen this way:
> $ grep speculative_pgfault /proc/vmstat
> 
> This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
> on x86, PowerPC and arm64.
> 
> ---------------------
> Real Workload results
> 
> As mentioned in previous email, we did non official runs using a "popular
> in memory multithreaded database product" on 176 cores SMT8 Power system
> which showed a 30% improvements in the number of transaction processed per
> second. This run has been done on the v6 series, but changes introduced in
> this new version should not impact the performance boost seen.
> 
> Here are the perf data captured during 2 of these runs on top of the v8
> series:
>                 vanilla         spf
> faults          89.418          101.364         +13%
> spf                n/a           97.989
> 
> With the SPF kernel, most of the page fault were processed in a speculative
> way.
> 
> Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
> it a try on an android device. He reported that the application launch time
> was improved in average by 6%, and for large applications (~100 threads) by
> 20%.
> 
> Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
> MSM845 (8 cores) with 6GB (the less is better):
> 
> Application                             4.9     4.9+spf delta
> com.tencent.mm                          416     389     -7%
> com.eg.android.AlipayGphone             1135    986     -13%
> com.tencent.mtt                         455     454     0%
> com.qqgame.hlddz                        1497    1409    -6%
> com.autonavi.minimap                    711     701     -1%
> com.tencent.tmgp.sgame                  788     748     -5%
> com.immomo.momo                         501     487     -3%
> com.tencent.peng                        2145    2112    -2%
> com.smile.gifmaker                      491     461     -6%
> com.baidu.BaiduMap                      479     366     -23%
> com.taobao.taobao                       1341    1198    -11%
> com.baidu.searchbox                     333     314     -6%
> com.tencent.mobileqq                    394     384     -3%
> com.sina.weibo                          907     906     0%
> com.youku.phone                         816     731     -11%
> com.happyelements.AndroidAnimal.qq      763     717     -6%
> com.UCMobile                            415     411     -1%
> com.tencent.tmgp.ak                     1464    1431    -2%
> com.tencent.qqmusic                     336     329     -2%
> com.sankuai.meituan                     1661    1302    -22%
> com.netease.cloudmusic                  1193    1200    1%
> air.tv.douyu.android                    4257    4152    -2%
> 
> ------------------
> Benchmarks results
> 
> Base kernel is v4.17.0-rc4-mm1
> SPF is BASE + this series
> 
> Kernbench:
> ----------
> Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
> kernel (kernel is build 5 times):
> 
> Average Half load -j 8
>                  Run    (std deviation)
>                  BASE                   SPF
> Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
> User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
> System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
> Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
> Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
> Sleeps           105064  (1240.96)      105074  (337.612)       0.01%
> 
> Average Optimal load -j 16
>                  Run    (std deviation)
>                  BASE                   SPF
> Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
> User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
> System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
> Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
> Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
> Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%
> 
> 
> During a run on the SPF, perf events were captured:
>  Performance counter stats for '../kernbench -M':
>          526743764      faults
>                210      spf
>                  3      pagefault:spf_vma_changed
>                  0      pagefault:spf_vma_noanon
>               2278      pagefault:spf_vma_notsup
>                  0      pagefault:spf_vma_access
>                  0      pagefault:spf_pmd_changed
> 
> Very few speculative page faults were recorded as most of the processes
> involved are monothreaded (sounds that on this architecture some threads
> were created during the kernel build processing).
> 
> Here are the kerbench results on a 80 CPUs Power8 system:
> 
> Average Half load -j 40
>                  Run    (std deviation)
>                  BASE                   SPF
> Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
> User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
> System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
> Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
> Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
> Sleeps           317923  (652.499)      318469  (1255.59)       0.17%
> 
> Average Optimal load -j 80
>                  Run    (std deviation)
>                  BASE                   SPF
> Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
> User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
> System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
> Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
> Context Switches 223861  (138865)       225032  (139632)        0.52%
> Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%
> 
> During a run on the SPF, perf events were captured:
>  Performance counter stats for '../kernbench -M':
>          116730856      faults
>                  0      spf
>                  3      pagefault:spf_vma_changed
>                  0      pagefault:spf_vma_noanon
>                476      pagefault:spf_vma_notsup
>                  0      pagefault:spf_vma_access
>                  0      pagefault:spf_pmd_changed
> 
> Most of the processes involved are monothreaded so SPF is not activated but
> there is no impact on the performance.
> 
> Ebizzy:
> -------
> The test is counting the number of records per second it can manage, the
> higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
> consistent result I repeated the test 100 times and measure the average
> result. The number is the record processes per second, the higher is the
> best.
> 
>                 BASE            SPF             delta
> 16 CPUs x86 VM  742.57          1490.24         100.69%
> 80 CPUs P8 node 13105.4         24174.23        84.46%
> 
> Here are the performance counter read during a run on a 16 CPUs x86 VM:
>  Performance counter stats for './ebizzy -mTt 16':
>            1706379      faults
>            1674599      spf
>              30588      pagefault:spf_vma_changed
>                  0      pagefault:spf_vma_noanon
>                363      pagefault:spf_vma_notsup
>                  0      pagefault:spf_vma_access
>                  0      pagefault:spf_pmd_changed
> 
> And the ones captured during a run on a 80 CPUs Power node:
>  Performance counter stats for './ebizzy -mTt 80':
>            1874773      faults
>            1461153      spf
>             413293      pagefault:spf_vma_changed
>                  0      pagefault:spf_vma_noanon
>                200      pagefault:spf_vma_notsup
>                  0      pagefault:spf_vma_access
>                  0      pagefault:spf_pmd_changed
> 
> In ebizzy's case most of the page fault were handled in a speculative way,
> leading the ebizzy performance boost.
> 
> ------------------
> Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
>  - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
>    and Minchan Kim, hopefully.
>  - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
>    __do_page_fault().
>  - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
>    instead
>    of aborting the speculative page fault handling. Dropping the now
> useless
>    trace event pagefault:spf_pte_lock.
>  - No more try to reuse the fetched VMA during the speculative page fault
>    handling when retrying is needed. This adds a lot of complexity and
>    additional tests done didn't show a significant performance improvement.
>  - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.
> 
> [1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
> [2] https://patchwork.kernel.org/patch/9999687/
> 
> 
> Laurent Dufour (20):
>   mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
>   x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>   powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>   mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
>   mm: make pte_unmap_same compatible with SPF
>   mm: introduce INIT_VMA()
>   mm: protect VMA modifications using VMA sequence count
>   mm: protect mremap() against SPF hanlder
>   mm: protect SPF handler against anon_vma changes
>   mm: cache some VMA fields in the vm_fault structure
>   mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
>   mm: introduce __lru_cache_add_active_or_unevictable
>   mm: introduce __vm_normal_page()
>   mm: introduce __page_add_new_anon_rmap()
>   mm: protect mm_rb tree with a rwlock
>   mm: adding speculative page fault failure trace events
>   perf: add a speculative page fault sw event
>   perf tools: add support for the SPF perf event
>   mm: add speculative page fault vmstats
>   powerpc/mm: add speculative page fault
> 
> Mahendran Ganesh (2):
>   arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
>   arm64/mm: add speculative page fault
> 
> Peter Zijlstra (4):
>   mm: prepare for FAULT_FLAG_SPECULATIVE
>   mm: VMA sequence count
>   mm: provide speculative fault infrastructure
>   x86/mm: add speculative pagefault handling
> 
>  arch/arm64/Kconfig                    |   1 +
>  arch/arm64/mm/fault.c                 |  12 +
>  arch/powerpc/Kconfig                  |   1 +
>  arch/powerpc/mm/fault.c               |  16 +
>  arch/x86/Kconfig                      |   1 +
>  arch/x86/mm/fault.c                   |  27 +-
>  fs/exec.c                             |   2 +-
>  fs/proc/task_mmu.c                    |   5 +-
>  fs/userfaultfd.c                      |  17 +-
>  include/linux/hugetlb_inline.h        |   2 +-
>  include/linux/migrate.h               |   4 +-
>  include/linux/mm.h                    | 136 +++++++-
>  include/linux/mm_types.h              |   7 +
>  include/linux/pagemap.h               |   4 +-
>  include/linux/rmap.h                  |  12 +-
>  include/linux/swap.h                  |  10 +-
>  include/linux/vm_event_item.h         |   3 +
>  include/trace/events/pagefault.h      |  80 +++++
>  include/uapi/linux/perf_event.h       |   1 +
>  kernel/fork.c                         |   5 +-
>  mm/Kconfig                            |  22 ++
>  mm/huge_memory.c                      |   6 +-
>  mm/hugetlb.c                          |   2 +
>  mm/init-mm.c                          |   3 +
>  mm/internal.h                         |  20 ++
>  mm/khugepaged.c                       |   5 +
>  mm/madvise.c                          |   6 +-
>  mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
>  mm/mempolicy.c                        |  51 ++-
>  mm/migrate.c                          |   6 +-
>  mm/mlock.c                            |  13 +-
>  mm/mmap.c                             | 229 ++++++++++---
>  mm/mprotect.c                         |   4 +-
>  mm/mremap.c                           |  13 +
>  mm/nommu.c                            |   2 +-
>  mm/rmap.c                             |   5 +-
>  mm/swap.c                             |   6 +-
>  mm/swap_state.c                       |   8 +-
>  mm/vmstat.c                           |   5 +-
>  tools/include/uapi/linux/perf_event.h |   1 +
>  tools/perf/util/evsel.c               |   1 +
>  tools/perf/util/parse-events.c        |   4 +
>  tools/perf/util/parse-events.l        |   1 +
>  tools/perf/util/python.c              |   1 +
>  44 files changed, 1161 insertions(+), 211 deletions(-)
>  create mode 100644 include/trace/events/pagefault.h
> 
> --
> 2.7.4
> 
> 

RE: [PATCH v11 00/26] Speculative page faults

[Song, HaiyanX]

Some regression and improvements is found by LKP-tools(linux kernel performance) on V9 patch series
tested on Intel 4s Skylake platform.

The regression result is sorted by the metric will-it-scale.per_thread_ops.
Branch: Laurent-Dufour/Speculative-page-faults/20180316-151833 (V9 patch series)
Commit id:
    base commit: d55f34411b1b126429a823d06c3124c16283231f
    head commit: 0355322b3577eeab7669066df42c550a56801110
Benchmark suite: will-it-scale
Download link:
https://github.com/antonblanchard/will-it-scale/tree/master/tests
Metrics:
    will-it-scale.per_process_ops=processes/nr_cpu
    will-it-scale.per_thread_ops=threads/nr_cpu
test box: lkp-skl-4sp1(nr_cpu=192,memory=768G)
THP: enable / disable
nr_task: 100%

1. Regressions:
a) THP enabled:
testcase                        base            change          head       metric
page_fault3/ enable THP         10092           -17.5%          8323       will-it-scale.per_thread_ops
page_fault2/ enable THP          8300           -17.2%          6869       will-it-scale.per_thread_ops
brk1/ enable THP                  957.67         -7.6%           885       will-it-scale.per_thread_ops
page_fault3/ enable THP        172821            -5.3%        163692       will-it-scale.per_process_ops
signal1/ enable THP              9125            -3.2%          8834       will-it-scale.per_process_ops

b) THP disabled:
testcase                        base            change          head       metric
page_fault3/ disable THP        10107           -19.1%          8180       will-it-scale.per_thread_ops
page_fault2/ disable THP         8432           -17.8%          6931       will-it-scale.per_thread_ops
context_switch1/ disable THP   215389            -6.8%        200776       will-it-scale.per_thread_ops
brk1/ disable THP                 939.67         -6.6%           877.33    will-it-scale.per_thread_ops
page_fault3/ disable THP       173145            -4.7%        165064       will-it-scale.per_process_ops
signal1/ disable THP             9162            -3.9%          8802       will-it-scale.per_process_ops

2. Improvements:
a) THP enabled:
testcase                        base            change          head       metric
malloc1/ enable THP               66.33        +469.8%           383.67    will-it-scale.per_thread_ops
writeseek3/ enable THP          2531             +4.5%          2646       will-it-scale.per_thread_ops
signal1/ enable THP              989.33          +2.8%          1016       will-it-scale.per_thread_ops

b) THP disabled:
testcase                        base            change          head       metric
malloc1/ disable THP              90.33        +417.3%           467.33    will-it-scale.per_thread_ops
read2/ disable THP             58934            +39.2%         82060       will-it-scale.per_thread_ops
page_fault1/ disable THP        8607            +36.4%         11736       will-it-scale.per_thread_ops
read1/ disable THP            314063            +12.7%        353934       will-it-scale.per_thread_ops
writeseek3/ disable THP         2452            +12.5%          2759       will-it-scale.per_thread_ops
signal1/ disable THP             971.33          +5.5%          1024       will-it-scale.per_thread_ops

Notes: for above values in column "change", the higher value means that the related testcase result
on head commit is better than that on base commit for this benchmark.


Best regards
Haiyan Song

________________________________________
From: owner-linux-mm@kvack.org [owner-linux-mm@kvack.org] on behalf of Laurent Dufour [ldufour@linux.vnet.ibm.com]
Sent: Thursday, May 17, 2018 7:06 PM
To: akpm@linux-foundation.org; mhocko@kernel.org; peterz@infradead.org; kirill@shutemov.name; ak@linux.intel.com; dave@stgolabs.net; jack@suse.cz; Matthew Wilcox; khandual@linux.vnet.ibm.com; aneesh.kumar@linux.vnet.ibm.com; benh@kernel.crashing.org; mpe@ellerman.id.au; paulus@samba.org; Thomas Gleixner; Ingo Molnar; hpa@zytor.com; Will Deacon; Sergey Senozhatsky; sergey.senozhatsky.work@gmail.com; Andrea Arcangeli; Alexei Starovoitov; Wang, Kemi; Daniel Jordan; David Rientjes; Jerome Glisse; Ganesh Mahendran; Minchan Kim; Punit Agrawal; vinayak menon; Yang Shi
Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org; haren@linux.vnet.ibm.com; npiggin@gmail.com; bsingharora@gmail.com; paulmck@linux.vnet.ibm.com; Tim Chen; linuxppc-dev@lists.ozlabs.org; x86@kernel.org
Subject: [PATCH v11 00/26] Speculative page faults

This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
page fault without holding the mm semaphore [1].

The idea is to try to handle user space page faults without holding the
mmap_sem. This should allow better concurrency for massively threaded
process since the page fault handler will not wait for other threads memory
layout change to be done, assuming that this change is done in another part
of the process's memory space. This type page fault is named speculative
page fault. If the speculative page fault fails because of a concurrency is
detected or because underlying PMD or PTE tables are not yet allocating, it
is failing its processing and a classic page fault is then tried.

The speculative page fault (SPF) has to look for the VMA matching the fault
address without holding the mmap_sem, this is done by introducing a rwlock
which protects the access to the mm_rb tree. Previously this was done using
SRCU but it was introducing a lot of scheduling to process the VMA's
freeing operation which was hitting the performance by 20% as reported by
Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
limiting the locking contention to these operations which are expected to
be in a O(log n) order. In addition to ensure that the VMA is not freed in
our back a reference count is added and 2 services (get_vma() and
put_vma()) are introduced to handle the reference count. Once a VMA is
fetched from the RB tree using get_vma(), it must be later freed using
put_vma(). I can't see anymore the overhead I got while will-it-scale
benchmark anymore.

The VMA's attributes checked during the speculative page fault processing
have to be protected against parallel changes. This is done by using a per
VMA sequence lock. This sequence lock allows the speculative page fault
handler to fast check for parallel changes in progress and to abort the
speculative page fault in that case.

Once the VMA has been found, the speculative page fault handler would check
for the VMA's attributes to verify that the page fault has to be handled
correctly or not. Thus, the VMA is protected through a sequence lock which
allows fast detection of concurrent VMA changes. If such a change is
detected, the speculative page fault is aborted and a *classic* page fault
is tried.  VMA sequence lockings are added when VMA attributes which are
checked during the page fault are modified.

When the PTE is fetched, the VMA is checked to see if it has been changed,
so once the page table is locked, the VMA is valid, so any other changes
leading to touching this PTE will need to lock the page table, so no
parallel change is possible at this time.

The locking of the PTE is done with interrupts disabled, this allows
checking for the PMD to ensure that there is not an ongoing collapsing
operation. Since khugepaged is firstly set the PMD to pmd_none and then is
waiting for the other CPU to have caught the IPI interrupt, if the pmd is
valid at the time the PTE is locked, we have the guarantee that the
collapsing operation will have to wait on the PTE lock to move forward.
This allows the SPF handler to map the PTE safely. If the PMD value is
different from the one recorded at the beginning of the SPF operation, the
classic page fault handler will be called to handle the operation while
holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
the lock is done using spin_trylock() to avoid dead lock when handling a
page fault while a TLB invalidate is requested by another CPU holding the
PTE.

In pseudo code, this could be seen as:
    speculative_page_fault()
    {
            vma = get_vma()
            check vma sequence count
            check vma's support
            disable interrupt
                  check pgd,p4d,...,pte
                  save pmd and pte in vmf
                  save vma sequence counter in vmf
            enable interrupt
            check vma sequence count
            handle_pte_fault(vma)
                    ..
                    page = alloc_page()
                    pte_map_lock()
                            disable interrupt
                                    abort if sequence counter has changed
                                    abort if pmd or pte has changed
                                    pte map and lock
                            enable interrupt
                    if abort
                       free page
                       abort
                    ...
    }

    arch_fault_handler()
    {
            if (speculative_page_fault(&vma))
               goto done
    again:
            lock(mmap_sem)
            vma = find_vma();
            handle_pte_fault(vma);
            if retry
               unlock(mmap_sem)
               goto again;
    done:
            handle fault error
    }

Support for THP is not done because when checking for the PMD, we can be
confused by an in progress collapsing operation done by khugepaged. The
issue is that pmd_none() could be true either if the PMD is not already
populated or if the underlying PTE are in the way to be collapsed. So we
cannot safely allocate a PMD if pmd_none() is true.

This series add a new software performance event named 'speculative-faults'
or 'spf'. It counts the number of successful page fault event handled
speculatively. When recording 'faults,spf' events, the faults one is
counting the total number of page fault events while 'spf' is only counting
the part of the faults processed speculatively.

There are some trace events introduced by this series. They allow
identifying why the page faults were not processed speculatively. This
doesn't take in account the faults generated by a monothreaded process
which directly processed while holding the mmap_sem. This trace events are
grouped in a system named 'pagefault', they are:
 - pagefault:spf_vma_changed : if the VMA has been changed in our back
 - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
 - pagefault:spf_vma_notsup : the VMA's type is not supported
 - pagefault:spf_vma_access : the VMA's access right are not respected
 - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
   back.

To record all the related events, the easier is to run perf with the
following arguments :
$ perf stat -e 'faults,spf,pagefault:*' <command>

There is also a dedicated vmstat counter showing the number of successful
page fault handled speculatively. I can be seen this way:
$ grep speculative_pgfault /proc/vmstat

This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
on x86, PowerPC and arm64.

---------------------
Real Workload results

As mentioned in previous email, we did non official runs using a "popular
in memory multithreaded database product" on 176 cores SMT8 Power system
which showed a 30% improvements in the number of transaction processed per
second. This run has been done on the v6 series, but changes introduced in
this new version should not impact the performance boost seen.

Here are the perf data captured during 2 of these runs on top of the v8
series:
                vanilla         spf
faults          89.418          101.364         +13%
spf                n/a           97.989

With the SPF kernel, most of the page fault were processed in a speculative
way.

Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
it a try on an android device. He reported that the application launch time
was improved in average by 6%, and for large applications (~100 threads) by
20%.

Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
MSM845 (8 cores) with 6GB (the less is better):

Application                             4.9     4.9+spf delta
com.tencent.mm                          416     389     -7%
com.eg.android.AlipayGphone             1135    986     -13%
com.tencent.mtt                         455     454     0%
com.qqgame.hlddz                        1497    1409    -6%
com.autonavi.minimap                    711     701     -1%
com.tencent.tmgp.sgame                  788     748     -5%
com.immomo.momo                         501     487     -3%
com.tencent.peng                        2145    2112    -2%
com.smile.gifmaker                      491     461     -6%
com.baidu.BaiduMap                      479     366     -23%
com.taobao.taobao                       1341    1198    -11%
com.baidu.searchbox                     333     314     -6%
com.tencent.mobileqq                    394     384     -3%
com.sina.weibo                          907     906     0%
com.youku.phone                         816     731     -11%
com.happyelements.AndroidAnimal.qq      763     717     -6%
com.UCMobile                            415     411     -1%
com.tencent.tmgp.ak                     1464    1431    -2%
com.tencent.qqmusic                     336     329     -2%
com.sankuai.meituan                     1661    1302    -22%
com.netease.cloudmusic                  1193    1200    1%
air.tv.douyu.android                    4257    4152    -2%

------------------
Benchmarks results

Base kernel is v4.17.0-rc4-mm1
SPF is BASE + this series

Kernbench:
----------
Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
kernel (kernel is build 5 times):

Average Half load -j 8
                 Run    (std deviation)
                 BASE                   SPF
Elapsed Time     1448.65 (5.72312)      1455.84 (4.84951)       0.50%
User    Time     10135.4 (30.3699)      10148.8 (31.1252)       0.13%
System  Time     900.47  (2.81131)      923.28  (7.52779)       2.53%
Percent CPU      761.4   (1.14018)      760.2   (0.447214)      -0.16%
Context Switches 85380   (3419.52)      84748   (1904.44)       -0.74%
Sleeps           105064  (1240.96)      105074  (337.612)       0.01%

Average Optimal load -j 16
                 Run    (std deviation)
                 BASE                   SPF
Elapsed Time     920.528 (10.1212)      927.404 (8.91789)       0.75%
User    Time     11064.8 (981.142)      11085   (990.897)       0.18%
System  Time     979.904 (84.0615)      1001.14 (82.5523)       2.17%
Percent CPU      1089.5  (345.894)      1086.1  (343.545)       -0.31%
Context Switches 159488  (78156.4)      158223  (77472.1)       -0.79%
Sleeps           110566  (5877.49)      110388  (5617.75)       -0.16%


During a run on the SPF, perf events were captured:
 Performance counter stats for '../kernbench -M':
         526743764      faults
               210      spf
                 3      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
              2278      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed

Very few speculative page faults were recorded as most of the processes
involved are monothreaded (sounds that on this architecture some threads
were created during the kernel build processing).

Here are the kerbench results on a 80 CPUs Power8 system:

Average Half load -j 40
                 Run    (std deviation)
                 BASE                   SPF
Elapsed Time     117.152 (0.774642)     117.166 (0.476057)      0.01%
User    Time     4478.52 (24.7688)      4479.76 (9.08555)       0.03%
System  Time     131.104 (0.720056)     134.04  (0.708414)      2.24%
Percent CPU      3934    (19.7104)      3937.2  (19.0184)       0.08%
Context Switches 92125.4 (576.787)      92581.6 (198.622)       0.50%
Sleeps           317923  (652.499)      318469  (1255.59)       0.17%

Average Optimal load -j 80
                 Run    (std deviation)
                 BASE                   SPF
Elapsed Time     107.73  (0.632416)     107.31  (0.584936)      -0.39%
User    Time     5869.86 (1466.72)      5871.71 (1467.27)       0.03%
System  Time     153.728 (23.8573)      157.153 (24.3704)       2.23%
Percent CPU      5418.6  (1565.17)      5436.7  (1580.91)       0.33%
Context Switches 223861  (138865)       225032  (139632)        0.52%
Sleeps           330529  (13495.1)      332001  (14746.2)       0.45%

During a run on the SPF, perf events were captured:
 Performance counter stats for '../kernbench -M':
         116730856      faults
                 0      spf
                 3      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               476      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed

Most of the processes involved are monothreaded so SPF is not activated but
there is no impact on the performance.

Ebizzy:
-------
The test is counting the number of records per second it can manage, the
higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
consistent result I repeated the test 100 times and measure the average
result. The number is the record processes per second, the higher is the
best.

                BASE            SPF             delta
16 CPUs x86 VM  742.57          1490.24         100.69%
80 CPUs P8 node 13105.4         24174.23        84.46%

Here are the performance counter read during a run on a 16 CPUs x86 VM:
 Performance counter stats for './ebizzy -mTt 16':
           1706379      faults
           1674599      spf
             30588      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               363      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed

And the ones captured during a run on a 80 CPUs Power node:
 Performance counter stats for './ebizzy -mTt 80':
           1874773      faults
           1461153      spf
            413293      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               200      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed

In ebizzy's case most of the page fault were handled in a speculative way,
leading the ebizzy performance boost.

------------------
Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
 - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
   and Minchan Kim, hopefully.
 - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
   __do_page_fault().
 - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
   instead
   of aborting the speculative page fault handling. Dropping the now
useless
   trace event pagefault:spf_pte_lock.
 - No more try to reuse the fetched VMA during the speculative page fault
   handling when retrying is needed. This adds a lot of complexity and
   additional tests done didn't show a significant performance improvement.
 - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.

[1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
[2] https://patchwork.kernel.org/patch/9999687/


Laurent Dufour (20):
  mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
  x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
  mm: make pte_unmap_same compatible with SPF
  mm: introduce INIT_VMA()
  mm: protect VMA modifications using VMA sequence count
  mm: protect mremap() against SPF hanlder
  mm: protect SPF handler against anon_vma changes
  mm: cache some VMA fields in the vm_fault structure
  mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
  mm: introduce __lru_cache_add_active_or_unevictable
  mm: introduce __vm_normal_page()
  mm: introduce __page_add_new_anon_rmap()
  mm: protect mm_rb tree with a rwlock
  mm: adding speculative page fault failure trace events
  perf: add a speculative page fault sw event
  perf tools: add support for the SPF perf event
  mm: add speculative page fault vmstats
  powerpc/mm: add speculative page fault

Mahendran Ganesh (2):
  arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  arm64/mm: add speculative page fault

Peter Zijlstra (4):
  mm: prepare for FAULT_FLAG_SPECULATIVE
  mm: VMA sequence count
  mm: provide speculative fault infrastructure
  x86/mm: add speculative pagefault handling

 arch/arm64/Kconfig                    |   1 +
 arch/arm64/mm/fault.c                 |  12 +
 arch/powerpc/Kconfig                  |   1 +
 arch/powerpc/mm/fault.c               |  16 +
 arch/x86/Kconfig                      |   1 +
 arch/x86/mm/fault.c                   |  27 +-
 fs/exec.c                             |   2 +-
 fs/proc/task_mmu.c                    |   5 +-
 fs/userfaultfd.c                      |  17 +-
 include/linux/hugetlb_inline.h        |   2 +-
 include/linux/migrate.h               |   4 +-
 include/linux/mm.h                    | 136 +++++++-
 include/linux/mm_types.h              |   7 +
 include/linux/pagemap.h               |   4 +-
 include/linux/rmap.h                  |  12 +-
 include/linux/swap.h                  |  10 +-
 include/linux/vm_event_item.h         |   3 +
 include/trace/events/pagefault.h      |  80 +++++
 include/uapi/linux/perf_event.h       |   1 +
 kernel/fork.c                         |   5 +-
 mm/Kconfig                            |  22 ++
 mm/huge_memory.c                      |   6 +-
 mm/hugetlb.c                          |   2 +
 mm/init-mm.c                          |   3 +
 mm/internal.h                         |  20 ++
 mm/khugepaged.c                       |   5 +
 mm/madvise.c                          |   6 +-
 mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
 mm/mempolicy.c                        |  51 ++-
 mm/migrate.c                          |   6 +-
 mm/mlock.c                            |  13 +-
 mm/mmap.c                             | 229 ++++++++++---
 mm/mprotect.c                         |   4 +-
 mm/mremap.c                           |  13 +
 mm/nommu.c                            |   2 +-
 mm/rmap.c                             |   5 +-
 mm/swap.c                             |   6 +-
 mm/swap_state.c                       |   8 +-
 mm/vmstat.c                           |   5 +-
 tools/include/uapi/linux/perf_event.h |   1 +
 tools/perf/util/evsel.c               |   1 +
 tools/perf/util/parse-events.c        |   4 +
 tools/perf/util/parse-events.l        |   1 +
 tools/perf/util/python.c              |   1 +
 44 files changed, 1161 insertions(+), 211 deletions(-)
 create mode 100644 include/trace/events/pagefault.h

--
2.7.4

[PATCH v11 00/26] Speculative page faults

[Laurent Dufour]

This is a port on kernel 4.17 of the work done by Peter Zijlstra to handle
page fault without holding the mm semaphore [1].

The idea is to try to handle user space page faults without holding the
mmap_sem. This should allow better concurrency for massively threaded
process since the page fault handler will not wait for other threads memory
layout change to be done, assuming that this change is done in another part
of the process's memory space. This type page fault is named speculative
page fault. If the speculative page fault fails because of a concurrency is
detected or because underlying PMD or PTE tables are not yet allocating, it
is failing its processing and a classic page fault is then tried.

The speculative page fault (SPF) has to look for the VMA matching the fault
address without holding the mmap_sem, this is done by introducing a rwlock
which protects the access to the mm_rb tree. Previously this was done using
SRCU but it was introducing a lot of scheduling to process the VMA's
freeing operation which was hitting the performance by 20% as reported by
Kemi Wang [2]. Using a rwlock to protect access to the mm_rb tree is
limiting the locking contention to these operations which are expected to
be in a O(log n) order. In addition to ensure that the VMA is not freed in
our back a reference count is added and 2 services (get_vma() and
put_vma()) are introduced to handle the reference count. Once a VMA is
fetched from the RB tree using get_vma(), it must be later freed using
put_vma(). I can't see anymore the overhead I got while will-it-scale
benchmark anymore.

The VMA's attributes checked during the speculative page fault processing
have to be protected against parallel changes. This is done by using a per
VMA sequence lock. This sequence lock allows the speculative page fault
handler to fast check for parallel changes in progress and to abort the
speculative page fault in that case.

Once the VMA has been found, the speculative page fault handler would check
for the VMA's attributes to verify that the page fault has to be handled
correctly or not. Thus, the VMA is protected through a sequence lock which
allows fast detection of concurrent VMA changes. If such a change is
detected, the speculative page fault is aborted and a *classic* page fault
is tried.  VMA sequence lockings are added when VMA attributes which are
checked during the page fault are modified.

When the PTE is fetched, the VMA is checked to see if it has been changed,
so once the page table is locked, the VMA is valid, so any other changes
leading to touching this PTE will need to lock the page table, so no
parallel change is possible at this time.

The locking of the PTE is done with interrupts disabled, this allows
checking for the PMD to ensure that there is not an ongoing collapsing
operation. Since khugepaged is firstly set the PMD to pmd_none and then is
waiting for the other CPU to have caught the IPI interrupt, if the pmd is
valid at the time the PTE is locked, we have the guarantee that the
collapsing operation will have to wait on the PTE lock to move forward.
This allows the SPF handler to map the PTE safely. If the PMD value is
different from the one recorded at the beginning of the SPF operation, the
classic page fault handler will be called to handle the operation while
holding the mmap_sem. As the PTE lock is done with the interrupts disabled,
the lock is done using spin_trylock() to avoid dead lock when handling a
page fault while a TLB invalidate is requested by another CPU holding the
PTE.

In pseudo code, this could be seen as:
    speculative_page_fault()
    {
	    vma = get_vma()
	    check vma sequence count
	    check vma's support
	    disable interrupt
		  check pgd,p4d,...,pte
		  save pmd and pte in vmf
		  save vma sequence counter in vmf
	    enable interrupt
	    check vma sequence count
	    handle_pte_fault(vma)
		    ..
		    page = alloc_page()
		    pte_map_lock()
			    disable interrupt
				    abort if sequence counter has changed
				    abort if pmd or pte has changed
				    pte map and lock
			    enable interrupt
		    if abort
		       free page
		       abort
		    ...
    }
    
    arch_fault_handler()
    {
	    if (speculative_page_fault(&vma))
	       goto done
    again:
	    lock(mmap_sem)
	    vma = find_vma();
	    handle_pte_fault(vma);
	    if retry
	       unlock(mmap_sem)
	       goto again;
    done:
	    handle fault error
    }

Support for THP is not done because when checking for the PMD, we can be
confused by an in progress collapsing operation done by khugepaged. The
issue is that pmd_none() could be true either if the PMD is not already
populated or if the underlying PTE are in the way to be collapsed. So we
cannot safely allocate a PMD if pmd_none() is true.

This series add a new software performance event named 'speculative-faults'
or 'spf'. It counts the number of successful page fault event handled
speculatively. When recording 'faults,spf' events, the faults one is
counting the total number of page fault events while 'spf' is only counting
the part of the faults processed speculatively.

There are some trace events introduced by this series. They allow
identifying why the page faults were not processed speculatively. This
doesn't take in account the faults generated by a monothreaded process
which directly processed while holding the mmap_sem. This trace events are
grouped in a system named 'pagefault', they are:
 - pagefault:spf_vma_changed : if the VMA has been changed in our back
 - pagefault:spf_vma_noanon : the vma->anon_vma field was not yet set.
 - pagefault:spf_vma_notsup : the VMA's type is not supported
 - pagefault:spf_vma_access : the VMA's access right are not respected
 - pagefault:spf_pmd_changed : the upper PMD pointer has changed in our
   back.

To record all the related events, the easier is to run perf with the
following arguments : 
$ perf stat -e 'faults,spf,pagefault:*' <command>

There is also a dedicated vmstat counter showing the number of successful
page fault handled speculatively. I can be seen this way: 
$ grep speculative_pgfault /proc/vmstat

This series builds on top of v4.16-mmotm-2018-04-13-17-28 and is functional
on x86, PowerPC and arm64.

---------------------
Real Workload results

As mentioned in previous email, we did non official runs using a "popular
in memory multithreaded database product" on 176 cores SMT8 Power system
which showed a 30% improvements in the number of transaction processed per
second. This run has been done on the v6 series, but changes introduced in
this new version should not impact the performance boost seen.

Here are the perf data captured during 2 of these runs on top of the v8
series:
		vanilla		spf
faults		89.418		101.364		+13%	
spf                n/a		 97.989

With the SPF kernel, most of the page fault were processed in a speculative
way.

Ganesh Mahendran had backported the series on top of a 4.9 kernel and gave
it a try on an android device. He reported that the application launch time
was improved in average by 6%, and for large applications (~100 threads) by
20%.

Here are the launch time Ganesh mesured on Android 8.0 on top of a Qcom
MSM845 (8 cores) with 6GB (the less is better):

Application				4.9	4.9+spf	delta
com.tencent.mm 				416	389 	-7%
com.eg.android.AlipayGphone 		1135 	986 	-13%
com.tencent.mtt 			455	454	0%
com.qqgame.hlddz 			1497	1409	-6%
com.autonavi.minimap 			711	701	-1%
com.tencent.tmgp.sgame 			788	748	-5%
com.immomo.momo 			501	487	-3%
com.tencent.peng 			2145	2112	-2%
com.smile.gifmaker 			491	461	-6%
com.baidu.BaiduMap 			479	366	-23%
com.taobao.taobao 			1341	1198	-11%
com.baidu.searchbox 			333	314 	-6%
com.tencent.mobileqq 			394	384	-3%
com.sina.weibo 				907	906	0%
com.youku.phone 			816	731	-11%
com.happyelements.AndroidAnimal.qq 	763	717	-6%
com.UCMobile 				415	411	-1%
com.tencent.tmgp.ak			1464	1431	-2%
com.tencent.qqmusic			336	329	-2%
com.sankuai.meituan			1661	1302	-22%
com.netease.cloudmusic			1193	1200	1%
air.tv.douyu.android			4257	4152	-2%

------------------
Benchmarks results

Base kernel is v4.17.0-rc4-mm1
SPF is BASE + this series

Kernbench:
----------
Here are the results on a 16 CPUs X86 guest using kernbench on a 4.15
kernel (kernel is build 5 times):

Average	Half load -j 8
		 Run	(std deviation)
		 BASE			SPF
Elapsed	Time	 1448.65 (5.72312)	1455.84	(4.84951)	0.50%
User	Time	 10135.4 (30.3699)	10148.8	(31.1252)	0.13%
System	Time	 900.47	 (2.81131)	923.28	(7.52779)	2.53%
Percent	CPU	 761.4	 (1.14018)	760.2	(0.447214)	-0.16%
Context	Switches 85380	 (3419.52)	84748	(1904.44)	-0.74%
Sleeps		 105064	 (1240.96)	105074	(337.612)	0.01%
						
Average	Optimal	load -j	16
		 Run	(std deviation)
		 BASE			SPF
Elapsed	Time	 920.528 (10.1212)	927.404	(8.91789)	0.75%
User	Time	 11064.8 (981.142)	11085	(990.897)	0.18%
System	Time	 979.904 (84.0615)	1001.14	(82.5523)	2.17%
Percent	CPU	 1089.5	 (345.894)	1086.1	(343.545)	-0.31%
Context	Switches 159488	 (78156.4)	158223	(77472.1)	-0.79%
Sleeps		 110566	 (5877.49)	110388	(5617.75)	-0.16%


During a run on the SPF, perf events were captured:
 Performance counter stats for '../kernbench -M':
         526743764      faults                   
               210      spf                      
                 3      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon 
              2278      pagefault:spf_vma_notsup 
                 0      pagefault:spf_vma_access 
                 0      pagefault:spf_pmd_changed

Very few speculative page faults were recorded as most of the processes
involved are monothreaded (sounds that on this architecture some threads
were created during the kernel build processing).

Here are the kerbench results on a 80 CPUs Power8 system:

Average	Half load -j 40
		 Run	(std deviation)
		 BASE			SPF
Elapsed	Time	 117.152 (0.774642)	117.166	(0.476057)	0.01%
User	Time	 4478.52 (24.7688)	4479.76	(9.08555)	0.03%
System	Time	 131.104 (0.720056)	134.04	(0.708414)	2.24%
Percent	CPU	 3934	 (19.7104)	3937.2	(19.0184)	0.08%
Context	Switches 92125.4 (576.787)	92581.6	(198.622)	0.50%
Sleeps		 317923	 (652.499)	318469	(1255.59)	0.17%
						
Average	Optimal	load -j	80
		 Run	(std deviation)
		 BASE			SPF
Elapsed	Time	 107.73	 (0.632416)	107.31	(0.584936)	-0.39%
User	Time	 5869.86 (1466.72)	5871.71	(1467.27)	0.03%
System	Time	 153.728 (23.8573)	157.153	(24.3704)	2.23%
Percent	CPU	 5418.6	 (1565.17)	5436.7	(1580.91)	0.33%
Context	Switches 223861	 (138865)	225032	(139632)	0.52%
Sleeps		 330529	 (13495.1)	332001	(14746.2)	0.45%

During a run on the SPF, perf events were captured:
 Performance counter stats for '../kernbench -M':
         116730856      faults
                 0      spf
                 3      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               476      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed

Most of the processes involved are monothreaded so SPF is not activated but
there is no impact on the performance.

Ebizzy:
-------
The test is counting the number of records per second it can manage, the
higher is the best. I run it like this 'ebizzy -mTt <nrcpus>'. To get
consistent result I repeated the test 100 times and measure the average
result. The number is the record processes per second, the higher is the
best.

  		BASE		SPF		delta	
16 CPUs x86 VM	742.57		1490.24		100.69%
80 CPUs P8 node 13105.4		24174.23	84.46%

Here are the performance counter read during a run on a 16 CPUs x86 VM:
 Performance counter stats for './ebizzy -mTt 16':
           1706379      faults
           1674599      spf
             30588      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               363      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access 
                 0      pagefault:spf_pmd_changed

And the ones captured during a run on a 80 CPUs Power node:
 Performance counter stats for './ebizzy -mTt 80':
           1874773      faults
           1461153      spf
            413293      pagefault:spf_vma_changed
                 0      pagefault:spf_vma_noanon
               200      pagefault:spf_vma_notsup
                 0      pagefault:spf_vma_access
                 0      pagefault:spf_pmd_changed 

In ebizzy's case most of the page fault were handled in a speculative way,
leading the ebizzy performance boost.

------------------
Changes since v10 (https://lkml.org/lkml/2018/4/17/572):
 - Accounted for all review feedbacks from Punit Agrawal, Ganesh Mahendran
   and Minchan Kim, hopefully.
 - Remove unneeded check on CONFIG_SPECULATIVE_PAGE_FAULT in
   __do_page_fault().
 - Loop in pte_spinlock() and pte_map_lock() when pte try lock fails
   instead
   of aborting the speculative page fault handling. Dropping the now
useless
   trace event pagefault:spf_pte_lock.
 - No more try to reuse the fetched VMA during the speculative page fault
   handling when retrying is needed. This adds a lot of complexity and
   additional tests done didn't show a significant performance improvement.
 - Convert IS_ENABLED(CONFIG_NUMA) back to #ifdef due to build error.

[1] http://linux-kernel.2935.n7.nabble.com/RFC-PATCH-0-6-Another-go-at-speculative-page-faults-tt965642.html#none
[2] https://patchwork.kernel.org/patch/9999687/


Laurent Dufour (20):
  mm: introduce CONFIG_SPECULATIVE_PAGE_FAULT
  x86/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  powerpc/mm: set ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  mm: introduce pte_spinlock for FAULT_FLAG_SPECULATIVE
  mm: make pte_unmap_same compatible with SPF
  mm: introduce INIT_VMA()
  mm: protect VMA modifications using VMA sequence count
  mm: protect mremap() against SPF hanlder
  mm: protect SPF handler against anon_vma changes
  mm: cache some VMA fields in the vm_fault structure
  mm/migrate: Pass vm_fault pointer to migrate_misplaced_page()
  mm: introduce __lru_cache_add_active_or_unevictable
  mm: introduce __vm_normal_page()
  mm: introduce __page_add_new_anon_rmap()
  mm: protect mm_rb tree with a rwlock
  mm: adding speculative page fault failure trace events
  perf: add a speculative page fault sw event
  perf tools: add support for the SPF perf event
  mm: add speculative page fault vmstats
  powerpc/mm: add speculative page fault

Mahendran Ganesh (2):
  arm64/mm: define ARCH_SUPPORTS_SPECULATIVE_PAGE_FAULT
  arm64/mm: add speculative page fault

Peter Zijlstra (4):
  mm: prepare for FAULT_FLAG_SPECULATIVE
  mm: VMA sequence count
  mm: provide speculative fault infrastructure
  x86/mm: add speculative pagefault handling

 arch/arm64/Kconfig                    |   1 +
 arch/arm64/mm/fault.c                 |  12 +
 arch/powerpc/Kconfig                  |   1 +
 arch/powerpc/mm/fault.c               |  16 +
 arch/x86/Kconfig                      |   1 +
 arch/x86/mm/fault.c                   |  27 +-
 fs/exec.c                             |   2 +-
 fs/proc/task_mmu.c                    |   5 +-
 fs/userfaultfd.c                      |  17 +-
 include/linux/hugetlb_inline.h        |   2 +-
 include/linux/migrate.h               |   4 +-
 include/linux/mm.h                    | 136 +++++++-
 include/linux/mm_types.h              |   7 +
 include/linux/pagemap.h               |   4 +-
 include/linux/rmap.h                  |  12 +-
 include/linux/swap.h                  |  10 +-
 include/linux/vm_event_item.h         |   3 +
 include/trace/events/pagefault.h      |  80 +++++
 include/uapi/linux/perf_event.h       |   1 +
 kernel/fork.c                         |   5 +-
 mm/Kconfig                            |  22 ++
 mm/huge_memory.c                      |   6 +-
 mm/hugetlb.c                          |   2 +
 mm/init-mm.c                          |   3 +
 mm/internal.h                         |  20 ++
 mm/khugepaged.c                       |   5 +
 mm/madvise.c                          |   6 +-
 mm/memory.c                           | 612 +++++++++++++++++++++++++++++-----
 mm/mempolicy.c                        |  51 ++-
 mm/migrate.c                          |   6 +-
 mm/mlock.c                            |  13 +-
 mm/mmap.c                             | 229 ++++++++++---
 mm/mprotect.c                         |   4 +-
 mm/mremap.c                           |  13 +
 mm/nommu.c                            |   2 +-
 mm/rmap.c                             |   5 +-
 mm/swap.c                             |   6 +-
 mm/swap_state.c                       |   8 +-
 mm/vmstat.c                           |   5 +-
 tools/include/uapi/linux/perf_event.h |   1 +
 tools/perf/util/evsel.c               |   1 +
 tools/perf/util/parse-events.c        |   4 +
 tools/perf/util/parse-events.l        |   1 +
 tools/perf/util/python.c              |   1 +
 44 files changed, 1161 insertions(+), 211 deletions(-)
 create mode 100644 include/trace/events/pagefault.h

-- 
2.7.4