[PATCH v3 0/4] mm: arm64: bring up BATCHED_UNMAP_TLB_FLUSH

[PATCH v3 0/4] mm: arm64: bring up BATCHED_UNMAP_TLB_FLUSH

[Yicong Yang]

From: Yicong Yang <yangyicong@hisilicon.com>

Though ARM64 has the hardware to do tlb shootdown, the hardware
broadcasting is not free.
A simplest micro benchmark shows even on snapdragon 888 with only
8 cores, the overhead for ptep_clear_flush is huge even for paging
out one page mapped by only one process:
5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush

While pages are mapped by multiple processes or HW has more CPUs,
the cost should become even higher due to the bad scalability of
tlb shootdown.

The same benchmark can result in 16.99% CPU consumption on ARM64
server with around 100 cores according to Yicong's test on patch
4/4.

This patchset leverages the existing BATCHED_UNMAP_TLB_FLUSH by
1. only send tlbi instructions in the first stage -
	arch_tlbbatch_add_mm()
2. wait for the completion of tlbi by dsb while doing tlbbatch
	sync in arch_tlbbatch_flush()
My testing on snapdragon shows the overhead of ptep_clear_flush
is removed by the patchset. The micro benchmark becomes 5% faster
even for one page mapped by single process on snapdragon 888.

-v3:
1. Declare arch's tlbbatch defer support by arch_tlbbatch_should_defer() instead
   of ARCH_HAS_MM_CPUMASK, per Barry and Kefeng
2. Add Tested-by from Xin Hao
Link: https://lore.kernel.org/linux-mm/20220711034615.482895-1-21cnbao@gmail.com/

-v2:
1. Collected Yicong's test result on kunpeng920 ARM64 server;
2. Removed the redundant vma parameter in arch_tlbbatch_add_mm()
   according to the comments of Peter Zijlstra and Dave Hansen
3. Added ARCH_HAS_MM_CPUMASK rather than checking if mm_cpumask
   is empty according to the comments of Nadav Amit

Thanks, Peter, Dave and Nadav for your testing or reviewing
, and comments.

-v1:
https://lore.kernel.org/lkml/20220707125242.425242-1-21cnbao@gmail.com/

Anshuman Khandual (1):
  mm/tlbbatch: Introduce arch_tlbbatch_should_defer()

Barry Song (3):
  Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't
    apply to ARM64"
  mm: rmap: Extend tlbbatch APIs to fit new platforms
  arm64: support batched/deferred tlb shootdown during page reclamation

 Documentation/features/arch-support.txt       |  1 -
 .../features/vm/TLB/arch-support.txt          |  2 +-
 arch/arm64/Kconfig                            |  1 +
 arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
 arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
 arch/x86/include/asm/tlbflush.h               | 15 +++++++++-
 mm/rmap.c                                     | 19 +++++--------
 7 files changed, 61 insertions(+), 17 deletions(-)
 create mode 100644 arch/arm64/include/asm/tlbbatch.h

-- 
2.24.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH v3 1/4] Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't apply to ARM64"

[Yicong Yang]

From: Barry Song <v-songbaohua@oppo.com>

This reverts commit 6bfef171d0d74cb050112e0e49feb20bfddf7f42.

I was wrong. Though ARM64 has hardware TLB flush, but it is not free
and it is still expensive.
We still have a good chance to enable batched and deferred TLB flush
on ARM64 for memory reclamation. A possible way is that we only queue
tlbi instructions in hardware's queue. When we have to broadcast TLB,
we broadcast it by dsb. We just need to get adapted the existing
BATCHED_UNMAP_TLB_FLUSH.

Tested-by: Xin Hao <xhao@linux.alibaba.com>
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
---
 Documentation/features/arch-support.txt        | 1 -
 Documentation/features/vm/TLB/arch-support.txt | 2 +-
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/Documentation/features/arch-support.txt b/Documentation/features/arch-support.txt
index 118ae031840b..d22a1095e661 100644
--- a/Documentation/features/arch-support.txt
+++ b/Documentation/features/arch-support.txt
@@ -8,5 +8,4 @@ The meaning of entries in the tables is:
     | ok |  # feature supported by the architecture
     |TODO|  # feature not yet supported by the architecture
     | .. |  # feature cannot be supported by the hardware
-    | N/A|  # feature doesn't apply to the architecture
 
diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
index 039e4e91ada3..1c009312b9c1 100644
--- a/Documentation/features/vm/TLB/arch-support.txt
+++ b/Documentation/features/vm/TLB/arch-support.txt
@@ -9,7 +9,7 @@
     |       alpha: | TODO |
     |         arc: | TODO |
     |         arm: | TODO |
-    |       arm64: | N/A  |
+    |       arm64: | TODO |
     |        csky: | TODO |
     |     hexagon: | TODO |
     |        ia64: | TODO |
-- 
2.24.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 1/4] Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't apply to ARM64"

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> From: Barry Song <v-songbaohua@oppo.com>
> 
> This reverts commit 6bfef171d0d74cb050112e0e49feb20bfddf7f42.
> 
> I was wrong. Though ARM64 has hardware TLB flush, but it is not free
> and it is still expensive.
> We still have a good chance to enable batched and deferred TLB flush
> on ARM64 for memory reclamation. A possible way is that we only queue
> tlbi instructions in hardware's queue. When we have to broadcast TLB,
> we broadcast it by dsb. We just need to get adapted the existing
> BATCHED_UNMAP_TLB_FLUSH.
> 
> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> ---
>  Documentation/features/arch-support.txt        | 1 -
>  Documentation/features/vm/TLB/arch-support.txt | 2 +-
>  2 files changed, 1 insertion(+), 2 deletions(-)
> 
> diff --git a/Documentation/features/arch-support.txt b/Documentation/features/arch-support.txt
> index 118ae031840b..d22a1095e661 100644
> --- a/Documentation/features/arch-support.txt
> +++ b/Documentation/features/arch-support.txt
> @@ -8,5 +8,4 @@ The meaning of entries in the tables is:
>      | ok |  # feature supported by the architecture
>      |TODO|  # feature not yet supported by the architecture
>      | .. |  # feature cannot be supported by the hardware
> -    | N/A|  # feature doesn't apply to the architecture
>  
> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> index 039e4e91ada3..1c009312b9c1 100644
> --- a/Documentation/features/vm/TLB/arch-support.txt
> +++ b/Documentation/features/vm/TLB/arch-support.txt
> @@ -9,7 +9,7 @@
>      |       alpha: | TODO |
>      |         arc: | TODO |
>      |         arm: | TODO |
> -    |       arm64: | N/A  |
> +    |       arm64: | TODO |
>      |        csky: | TODO |
>      |     hexagon: | TODO |
>      |        ia64: | TODO |

I believe this patch is not needed, which explicitly reverts an
older commit. Instead when ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
framework gets enabled on arm64, the same patch could just drop
'NA' as possible values for arch support for a give feature in
file Documentation/features/arch-support.txt.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 1/4] Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't apply to ARM64"

[Barry Song]

On Fri, Sep 9, 2022 at 4:26 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 8/22/22 13:51, Yicong Yang wrote:
> > From: Barry Song <v-songbaohua@oppo.com>
> >
> > This reverts commit 6bfef171d0d74cb050112e0e49feb20bfddf7f42.
> >
> > I was wrong. Though ARM64 has hardware TLB flush, but it is not free
> > and it is still expensive.
> > We still have a good chance to enable batched and deferred TLB flush
> > on ARM64 for memory reclamation. A possible way is that we only queue
> > tlbi instructions in hardware's queue. When we have to broadcast TLB,
> > we broadcast it by dsb. We just need to get adapted the existing
> > BATCHED_UNMAP_TLB_FLUSH.
> >
> > Tested-by: Xin Hao <xhao@linux.alibaba.com>
> > Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> > Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> > ---
> >  Documentation/features/arch-support.txt        | 1 -
> >  Documentation/features/vm/TLB/arch-support.txt | 2 +-
> >  2 files changed, 1 insertion(+), 2 deletions(-)
> >
> > diff --git a/Documentation/features/arch-support.txt b/Documentation/features/arch-support.txt
> > index 118ae031840b..d22a1095e661 100644
> > --- a/Documentation/features/arch-support.txt
> > +++ b/Documentation/features/arch-support.txt
> > @@ -8,5 +8,4 @@ The meaning of entries in the tables is:
> >      | ok |  # feature supported by the architecture
> >      |TODO|  # feature not yet supported by the architecture
> >      | .. |  # feature cannot be supported by the hardware
> > -    | N/A|  # feature doesn't apply to the architecture
> >
> > diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> > index 039e4e91ada3..1c009312b9c1 100644
> > --- a/Documentation/features/vm/TLB/arch-support.txt
> > +++ b/Documentation/features/vm/TLB/arch-support.txt
> > @@ -9,7 +9,7 @@
> >      |       alpha: | TODO |
> >      |         arc: | TODO |
> >      |         arm: | TODO |
> > -    |       arm64: | N/A  |
> > +    |       arm64: | TODO |
> >      |        csky: | TODO |
> >      |     hexagon: | TODO |
> >      |        ia64: | TODO |
>
> I believe this patch is not needed, which explicitly reverts an
> older commit. Instead when ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> framework gets enabled on arm64, the same patch could just drop
> 'NA' as possible values for arch support for a give feature in
> file Documentation/features/arch-support.txt.

Sure. it is certainly ok to fix this in
arm64: support batched/deferred tlb shootdown during page reclamation

By a separate patch, I was trying to highlight that my previous patch was
wrong. but, yes. it is not fundamentally necessary.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH v3 2/4] mm/tlbbatch: Introduce arch_tlbbatch_should_defer()

[Yicong Yang]

From: Anshuman Khandual <khandual@linux.vnet.ibm.com>

The entire scheme of deferred TLB flush in reclaim path rests on the
fact that the cost to refill TLB entries is less than flushing out
individual entries by sending IPI to remote CPUs. But architecture
can have different ways to evaluate that. Hence apart from checking
TTU_BATCH_FLUSH in the TTU flags, rest of the decision should be
architecture specific.

Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
[https://lore.kernel.org/linuxppc-dev/20171101101735.2318-2-khandual@linux.vnet.ibm.com/]
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
[Rebase and fix incorrect return value type]
---
 arch/x86/include/asm/tlbflush.h | 12 ++++++++++++
 mm/rmap.c                       |  9 +--------
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index cda3118f3b27..8a497d902c16 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -240,6 +240,18 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a)
 	flush_tlb_mm_range(vma->vm_mm, a, a + PAGE_SIZE, PAGE_SHIFT, false);
 }
 
+static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
+{
+	bool should_defer = false;
+
+	/* If remote CPUs need to be flushed then defer batch the flush */
+	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
+		should_defer = true;
+	put_cpu();
+
+	return should_defer;
+}
+
 static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
 {
 	/*
diff --git a/mm/rmap.c b/mm/rmap.c
index edc06c52bc82..a17a004550c6 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -687,17 +687,10 @@ static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
  */
 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 {
-	bool should_defer = false;
-
 	if (!(flags & TTU_BATCH_FLUSH))
 		return false;
 
-	/* If remote CPUs need to be flushed then defer batch the flush */
-	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
-		should_defer = true;
-	put_cpu();
-
-	return should_defer;
+	return arch_tlbbatch_should_defer(mm);
 }
 
 /*
-- 
2.24.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 2/4] mm/tlbbatch: Introduce arch_tlbbatch_should_defer()

[Kefeng Wang]

On 2022/8/22 16:21, Yicong Yang wrote:
> From: Anshuman Khandual <khandual@linux.vnet.ibm.com>
>
> The entire scheme of deferred TLB flush in reclaim path rests on the
> fact that the cost to refill TLB entries is less than flushing out
> individual entries by sending IPI to remote CPUs. But architecture
> can have different ways to evaluate that. Hence apart from checking
> TTU_BATCH_FLUSH in the TTU flags, rest of the decision should be
> architecture specific.
>
> Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
> [https://lore.kernel.org/linuxppc-dev/20171101101735.2318-2-khandual@linux.vnet.ibm.com/]
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> [Rebase and fix incorrect return value type]

Reviewed-by: Kefeng Wang <wangkefeng.wang@huawei.com>

> ---
>   arch/x86/include/asm/tlbflush.h | 12 ++++++++++++
>   mm/rmap.c                       |  9 +--------
>   2 files changed, 13 insertions(+), 8 deletions(-)
>
> diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
> index cda3118f3b27..8a497d902c16 100644
> --- a/arch/x86/include/asm/tlbflush.h
> +++ b/arch/x86/include/asm/tlbflush.h
> @@ -240,6 +240,18 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a)
>   	flush_tlb_mm_range(vma->vm_mm, a, a + PAGE_SIZE, PAGE_SHIFT, false);
>   }
>   
> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> +{
> +	bool should_defer = false;
> +
> +	/* If remote CPUs need to be flushed then defer batch the flush */
> +	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
> +		should_defer = true;
> +	put_cpu();
> +
> +	return should_defer;
> +}
> +
>   static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
>   {
>   	/*
> diff --git a/mm/rmap.c b/mm/rmap.c
> index edc06c52bc82..a17a004550c6 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -687,17 +687,10 @@ static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
>    */
>   static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
>   {
> -	bool should_defer = false;
> -
>   	if (!(flags & TTU_BATCH_FLUSH))
>   		return false;
>   
> -	/* If remote CPUs need to be flushed then defer batch the flush */
> -	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
> -		should_defer = true;
> -	put_cpu();
> -
> -	return should_defer;
> +	return arch_tlbbatch_should_defer(mm);
>   }
>   
>   /*

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 2/4] mm/tlbbatch: Introduce arch_tlbbatch_should_defer()

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> From: Anshuman Khandual <khandual@linux.vnet.ibm.com>
> 
> The entire scheme of deferred TLB flush in reclaim path rests on the
> fact that the cost to refill TLB entries is less than flushing out
> individual entries by sending IPI to remote CPUs. But architecture
> can have different ways to evaluate that. Hence apart from checking
> TTU_BATCH_FLUSH in the TTU flags, rest of the decision should be
> architecture specific.
> 
> Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
> [https://lore.kernel.org/linuxppc-dev/20171101101735.2318-2-khandual@linux.vnet.ibm.com/]
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> [Rebase and fix incorrect return value type]

From semantics perspective, this patch still makes sense, even on its own.

Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>

> ---
>  arch/x86/include/asm/tlbflush.h | 12 ++++++++++++
>  mm/rmap.c                       |  9 +--------
>  2 files changed, 13 insertions(+), 8 deletions(-)
> 
> diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
> index cda3118f3b27..8a497d902c16 100644
> --- a/arch/x86/include/asm/tlbflush.h
> +++ b/arch/x86/include/asm/tlbflush.h
> @@ -240,6 +240,18 @@ static inline void flush_tlb_page(struct vm_area_struct *vma, unsigned long a)
>  	flush_tlb_mm_range(vma->vm_mm, a, a + PAGE_SIZE, PAGE_SHIFT, false);
>  }
>  
> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> +{
> +	bool should_defer = false;
> +
> +	/* If remote CPUs need to be flushed then defer batch the flush */
> +	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
> +		should_defer = true;
> +	put_cpu();
> +
> +	return should_defer;
> +}
> +
>  static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
>  {
>  	/*
> diff --git a/mm/rmap.c b/mm/rmap.c
> index edc06c52bc82..a17a004550c6 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -687,17 +687,10 @@ static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
>   */
>  static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
>  {
> -	bool should_defer = false;
> -
>  	if (!(flags & TTU_BATCH_FLUSH))
>  		return false;
>  
> -	/* If remote CPUs need to be flushed then defer batch the flush */
> -	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
> -		should_defer = true;
> -	put_cpu();
> -
> -	return should_defer;
> +	return arch_tlbbatch_should_defer(mm);
>  }
>  
>  /*

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH v3 3/4] mm: rmap: Extend tlbbatch APIs to fit new platforms

[Yicong Yang]

From: Barry Song <v-songbaohua@oppo.com>

Add uaddr to tlbbatch APIs so that platforms like ARM64 are
able to apply this on their specific hardware features. For
ARM64, this could be sending tlbi into hardware queues for
the page with this particular uaddr.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Mel Gorman <mgorman@suse.de>
Tested-by: Xin Hao <xhao@linux.alibaba.com>
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
---
 arch/x86/include/asm/tlbflush.h |  3 ++-
 mm/rmap.c                       | 10 ++++++----
 2 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index 8a497d902c16..5bd78ae55cd4 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -264,7 +264,8 @@ static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
 }
 
 static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
-					struct mm_struct *mm)
+					struct mm_struct *mm,
+					unsigned long uaddr)
 {
 	inc_mm_tlb_gen(mm);
 	cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
diff --git a/mm/rmap.c b/mm/rmap.c
index a17a004550c6..7187a72b63b1 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -642,12 +642,13 @@ void try_to_unmap_flush_dirty(void)
 #define TLB_FLUSH_BATCH_PENDING_LARGE			\
 	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
 
-static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
+static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
+				      unsigned long uaddr)
 {
 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 	int batch, nbatch;
 
-	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
+	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm, uaddr);
 	tlb_ubc->flush_required = true;
 
 	/*
@@ -725,7 +726,8 @@ void flush_tlb_batched_pending(struct mm_struct *mm)
 	}
 }
 #else
-static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
+static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
+				      unsigned long uaddr)
 {
 }
 
@@ -1587,7 +1589,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 				 */
 				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
 
-				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
+				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval), address);
 			} else {
 				pteval = ptep_clear_flush(vma, address, pvmw.pte);
 			}
-- 
2.24.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 3/4] mm: rmap: Extend tlbbatch APIs to fit new platforms

[Kefeng Wang]

On 2022/8/22 16:21, Yicong Yang wrote:
> From: Barry Song <v-songbaohua@oppo.com>
>
> Add uaddr to tlbbatch APIs so that platforms like ARM64 are
> able to apply this on their specific hardware features. For
> ARM64, this could be sending tlbi into hardware queues for
> the page with this particular uaddr.
>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Borislav Petkov <bp@alien8.de>
> Cc: Dave Hansen <dave.hansen@linux.intel.com>
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: Nadav Amit <namit@vmware.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
Reviewed-by: Kefeng Wang <wangkefeng.wang@huawei.com>
> ---
>   arch/x86/include/asm/tlbflush.h |  3 ++-
>   mm/rmap.c                       | 10 ++++++----
>   2 files changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
> index 8a497d902c16..5bd78ae55cd4 100644
> --- a/arch/x86/include/asm/tlbflush.h
> +++ b/arch/x86/include/asm/tlbflush.h
> @@ -264,7 +264,8 @@ static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
>   }
>   
>   static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> -					struct mm_struct *mm)
> +					struct mm_struct *mm,
> +					unsigned long uaddr)
>   {
>   	inc_mm_tlb_gen(mm);
>   	cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
> diff --git a/mm/rmap.c b/mm/rmap.c
> index a17a004550c6..7187a72b63b1 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -642,12 +642,13 @@ void try_to_unmap_flush_dirty(void)
>   #define TLB_FLUSH_BATCH_PENDING_LARGE			\
>   	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
>   
> -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> +				      unsigned long uaddr)
>   {
>   	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
>   	int batch, nbatch;
>   
> -	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
> +	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm, uaddr);
>   	tlb_ubc->flush_required = true;
>   
>   	/*
> @@ -725,7 +726,8 @@ void flush_tlb_batched_pending(struct mm_struct *mm)
>   	}
>   }
>   #else
> -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> +				      unsigned long uaddr)
>   {
>   }
>   
> @@ -1587,7 +1589,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
>   				 */
>   				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
>   
> -				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
> +				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval), address);
>   			} else {
>   				pteval = ptep_clear_flush(vma, address, pvmw.pte);
>   			}

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 3/4] mm: rmap: Extend tlbbatch APIs to fit new platforms

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> From: Barry Song <v-songbaohua@oppo.com>
> 
> Add uaddr to tlbbatch APIs so that platforms like ARM64 are

I guess 'uaddr' refers to a virtual address from the process address
space itself ? Please be more specific.

> able to apply this on their specific hardware features. For
> ARM64, this could be sending tlbi into hardware queues for
> the page with this particular uaddr.

This subject line and commit message here are misleading. The patch
adds an address argument to arch callback arch_tlbbatch_add_mm() as
arm64 platform could use that to perform the TLB flush batching ?

This patch can be folded into the next one, so that the requirement
for an additional argument 'uaddr' in the arch callback will be self
evident. OR if this is going to be a preparatory patch, then it must
explain how 'uaddr' argument is helpful on platforms like arm64 while
performing TLB flush batching. But TBH, just folding it to next patch
explains the context better.

> 
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Borislav Petkov <bp@alien8.de>
> Cc: Dave Hansen <dave.hansen@linux.intel.com>
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: Nadav Amit <namit@vmware.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> ---
>  arch/x86/include/asm/tlbflush.h |  3 ++-
>  mm/rmap.c                       | 10 ++++++----
>  2 files changed, 8 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
> index 8a497d902c16..5bd78ae55cd4 100644
> --- a/arch/x86/include/asm/tlbflush.h
> +++ b/arch/x86/include/asm/tlbflush.h
> @@ -264,7 +264,8 @@ static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
>  }
>  
>  static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> -					struct mm_struct *mm)
> +					struct mm_struct *mm,
> +					unsigned long uaddr)
>  {
>  	inc_mm_tlb_gen(mm);
>  	cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
> diff --git a/mm/rmap.c b/mm/rmap.c
> index a17a004550c6..7187a72b63b1 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -642,12 +642,13 @@ void try_to_unmap_flush_dirty(void)
>  #define TLB_FLUSH_BATCH_PENDING_LARGE			\
>  	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
>  
> -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> +				      unsigned long uaddr)
>  {
>  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
>  	int batch, nbatch;
>  
> -	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
> +	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm, uaddr);
>  	tlb_ubc->flush_required = true;
>  
>  	/*
> @@ -725,7 +726,8 @@ void flush_tlb_batched_pending(struct mm_struct *mm)
>  	}
>  }
>  #else
> -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> +				      unsigned long uaddr)
>  {
>  }
>  
> @@ -1587,7 +1589,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
>  				 */
>  				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
>  
> -				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
> +				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval), address);
>  			} else {
>  				pteval = ptep_clear_flush(vma, address, pvmw.pte);
>  			}

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 3/4] mm: rmap: Extend tlbbatch APIs to fit new platforms

[Barry Song]

On Fri, Sep 9, 2022 at 4:51 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 8/22/22 13:51, Yicong Yang wrote:
> > From: Barry Song <v-songbaohua@oppo.com>
> >
> > Add uaddr to tlbbatch APIs so that platforms like ARM64 are
>
> I guess 'uaddr' refers to a virtual address from the process address
> space itself ? Please be more specific.
>
> > able to apply this on their specific hardware features. For
> > ARM64, this could be sending tlbi into hardware queues for
> > the page with this particular uaddr.
>
> This subject line and commit message here are misleading. The patch
> adds an address argument to arch callback arch_tlbbatch_add_mm() as
> arm64 platform could use that to perform the TLB flush batching ?
>
> This patch can be folded into the next one, so that the requirement
> for an additional argument 'uaddr' in the arch callback will be self
> evident. OR if this is going to be a preparatory patch, then it must
> explain how 'uaddr' argument is helpful on platforms like arm64 while
> performing TLB flush batching. But TBH, just folding it to next patch
> explains the context better.

The intention was to keep each change small, while still functionally
independent,
so that it was easier to be reviewed.

but yes, i agree in this particular case, if we fold this one to the
last one, we are
actually able to make the modification self-evident while the new patch seems
still small.

>
> >
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > Cc: Ingo Molnar <mingo@redhat.com>
> > Cc: Borislav Petkov <bp@alien8.de>
> > Cc: Dave Hansen <dave.hansen@linux.intel.com>
> > Cc: "H. Peter Anvin" <hpa@zytor.com>
> > Cc: Nadav Amit <namit@vmware.com>
> > Cc: Mel Gorman <mgorman@suse.de>
> > Tested-by: Xin Hao <xhao@linux.alibaba.com>
> > Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> > Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> > ---
> >  arch/x86/include/asm/tlbflush.h |  3 ++-
> >  mm/rmap.c                       | 10 ++++++----
> >  2 files changed, 8 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
> > index 8a497d902c16..5bd78ae55cd4 100644
> > --- a/arch/x86/include/asm/tlbflush.h
> > +++ b/arch/x86/include/asm/tlbflush.h
> > @@ -264,7 +264,8 @@ static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
> >  }
> >
> >  static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> > -                                     struct mm_struct *mm)
> > +                                     struct mm_struct *mm,
> > +                                     unsigned long uaddr)
> >  {
> >       inc_mm_tlb_gen(mm);
> >       cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
> > diff --git a/mm/rmap.c b/mm/rmap.c
> > index a17a004550c6..7187a72b63b1 100644
> > --- a/mm/rmap.c
> > +++ b/mm/rmap.c
> > @@ -642,12 +642,13 @@ void try_to_unmap_flush_dirty(void)
> >  #define TLB_FLUSH_BATCH_PENDING_LARGE                        \
> >       (TLB_FLUSH_BATCH_PENDING_MASK / 2)
> >
> > -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> > +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> > +                                   unsigned long uaddr)
> >  {
> >       struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
> >       int batch, nbatch;
> >
> > -     arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
> > +     arch_tlbbatch_add_mm(&tlb_ubc->arch, mm, uaddr);
> >       tlb_ubc->flush_required = true;
> >
> >       /*
> > @@ -725,7 +726,8 @@ void flush_tlb_batched_pending(struct mm_struct *mm)
> >       }
> >  }
> >  #else
> > -static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
> > +static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable,
> > +                                   unsigned long uaddr)
> >  {
> >  }
> >
> > @@ -1587,7 +1589,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
> >                                */
> >                               pteval = ptep_get_and_clear(mm, address, pvmw.pte);
> >
> > -                             set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
> > +                             set_tlb_ubc_flush_pending(mm, pte_dirty(pteval), address);
> >                       } else {
> >                               pteval = ptep_clear_flush(vma, address, pvmw.pte);
> >                       }

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

[PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Yicong Yang]

From: Barry Song <v-songbaohua@oppo.com>

on x86, batched and deferred tlb shootdown has lead to 90%
performance increase on tlb shootdown. on arm64, HW can do
tlb shootdown without software IPI. But sync tlbi is still
quite expensive.

Even running a simplest program which requires swapout can
prove this is true,
 #include <sys/types.h>
 #include <unistd.h>
 #include <sys/mman.h>
 #include <string.h>

 int main()
 {
 #define SIZE (1 * 1024 * 1024)
         volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
                                          MAP_SHARED | MAP_ANONYMOUS, -1, 0);

         memset(p, 0x88, SIZE);

         for (int k = 0; k < 10000; k++) {
                 /* swap in */
                 for (int i = 0; i < SIZE; i += 4096) {
                         (void)p[i];
                 }

                 /* swap out */
                 madvise(p, SIZE, MADV_PAGEOUT);
         }
 }

Perf result on snapdragon 888 with 8 cores by using zRAM
as the swap block device.

 ~ # perf record taskset -c 4 ./a.out
 [ perf record: Woken up 10 times to write data ]
 [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
 ~ # perf report
 # To display the perf.data header info, please use --header/--header-only options.
 # To display the perf.data header info, please use --header/--header-only options.
 #
 #
 # Total Lost Samples: 0
 #
 # Samples: 60K of event 'cycles'
 # Event count (approx.): 35706225414
 #
 # Overhead  Command  Shared Object      Symbol
 # ........  .......  .................  .............................................................................
 #
    21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
     8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
     6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
     6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
     5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
     3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
     3.49%  a.out    [kernel.kallsyms]  [k] memset64
     1.63%  a.out    [kernel.kallsyms]  [k] clear_page
     1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
     1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
     1.23%  a.out    [kernel.kallsyms]  [k] xas_load
     1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock

ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
swapping in/out a page mapped by only one process. If the
page is mapped by multiple processes, typically, like more
than 100 on a phone, the overhead would be much higher as
we have to run tlb flush 100 times for one single page.
Plus, tlb flush overhead will increase with the number
of CPU cores due to the bad scalability of tlb shootdown
in HW, so those ARM64 servers should expect much higher
overhead.

Further perf annonate shows 95% cpu time of ptep_clear_flush
is actually used by the final dsb() to wait for the completion
of tlb flush. This provides us a very good chance to leverage
the existing batched tlb in kernel. The minimum modification
is that we only send async tlbi in the first stage and we send
dsb while we have to sync in the second stage.

With the above simplest micro benchmark, collapsed time to
finish the program decreases around 5%.

Typical collapsed time w/o patch:
 ~ # time taskset -c 4 ./a.out
 0.21user 14.34system 0:14.69elapsed
w/ patch:
 ~ # time taskset -c 4 ./a.out
 0.22user 13.45system 0:13.80elapsed

Also, Yicong Yang added the following observation.
	Tested with benchmark in the commit on Kunpeng920 arm64 server,
	observed an improvement around 12.5% with command
	`time ./swap_bench`.
		w/o		w/
	real	0m13.460s	0m11.771s
	user	0m0.248s	0m0.279s
	sys	0m12.039s	0m11.458s

	Originally it's noticed a 16.99% overhead of ptep_clear_flush()
	which has been eliminated by this patch:

	[root@localhost yang]# perf record -- ./swap_bench && perf report
	[...]
	16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush

Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Nadav Amit <namit@vmware.com>
Cc: Mel Gorman <mgorman@suse.de>
Tested-by: Yicong Yang <yangyicong@hisilicon.com>
Tested-by: Xin Hao <xhao@linux.alibaba.com>
Signed-off-by: Barry Song <v-songbaohua@oppo.com>
Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
---
 .../features/vm/TLB/arch-support.txt          |  2 +-
 arch/arm64/Kconfig                            |  1 +
 arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
 arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
 4 files changed, 40 insertions(+), 3 deletions(-)
 create mode 100644 arch/arm64/include/asm/tlbbatch.h

diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
index 1c009312b9c1..2caf815d7c6c 100644
--- a/Documentation/features/vm/TLB/arch-support.txt
+++ b/Documentation/features/vm/TLB/arch-support.txt
@@ -9,7 +9,7 @@
     |       alpha: | TODO |
     |         arc: | TODO |
     |         arm: | TODO |
-    |       arm64: | TODO |
+    |       arm64: |  ok  |
     |        csky: | TODO |
     |     hexagon: | TODO |
     |        ia64: | TODO |
diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 571cc234d0b3..09d45cd6d665 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -93,6 +93,7 @@ config ARM64
 	select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
 	select ARCH_SUPPORTS_NUMA_BALANCING
 	select ARCH_SUPPORTS_PAGE_TABLE_CHECK
+	select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
 	select ARCH_WANT_DEFAULT_BPF_JIT
 	select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
new file mode 100644
index 000000000000..fedb0b87b8db
--- /dev/null
+++ b/arch/arm64/include/asm/tlbbatch.h
@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ARCH_ARM64_TLBBATCH_H
+#define _ARCH_ARM64_TLBBATCH_H
+
+struct arch_tlbflush_unmap_batch {
+	/*
+	 * For arm64, HW can do tlb shootdown, so we don't
+	 * need to record cpumask for sending IPI
+	 */
+};
+
+#endif /* _ARCH_ARM64_TLBBATCH_H */
diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
index 412a3b9a3c25..23cbc987321a 100644
--- a/arch/arm64/include/asm/tlbflush.h
+++ b/arch/arm64/include/asm/tlbflush.h
@@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
 	dsb(ish);
 }
 
-static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
+
+static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
 					 unsigned long uaddr)
 {
 	unsigned long addr;
 
 	dsb(ishst);
-	addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
+	addr = __TLBI_VADDR(uaddr, ASID(mm));
 	__tlbi(vale1is, addr);
 	__tlbi_user(vale1is, addr);
 }
 
+static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
+					 unsigned long uaddr)
+{
+	return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
+}
+
 static inline void flush_tlb_page(struct vm_area_struct *vma,
 				  unsigned long uaddr)
 {
@@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
 	dsb(ish);
 }
 
+static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
+{
+	return true;
+}
+
+static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
+					struct mm_struct *mm,
+					unsigned long uaddr)
+{
+	__flush_tlb_page_nosync(mm, uaddr);
+}
+
+static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
+{
+	dsb(ish);
+}
+
 /*
  * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
  * necessarily a performance improvement.
-- 
2.24.0


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Kefeng Wang]

On 2022/8/22 16:21, Yicong Yang wrote:
> From: Barry Song <v-songbaohua@oppo.com>
>
> on x86, batched and deferred tlb shootdown has lead to 90%
> performance increase on tlb shootdown. on arm64, HW can do
> tlb shootdown without software IPI. But sync tlbi is still
> quite expensive.
>
> Even running a simplest program which requires swapout can
> prove this is true,
>   #include <sys/types.h>
>   #include <unistd.h>
>   #include <sys/mman.h>
>   #include <string.h>
>
>   int main()
>   {
>   #define SIZE (1 * 1024 * 1024)
>           volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
>                                            MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>
>           memset(p, 0x88, SIZE);
>
>           for (int k = 0; k < 10000; k++) {
>                   /* swap in */
>                   for (int i = 0; i < SIZE; i += 4096) {
>                           (void)p[i];
>                   }
>
>                   /* swap out */
>                   madvise(p, SIZE, MADV_PAGEOUT);
>           }
>   }
>
> Perf result on snapdragon 888 with 8 cores by using zRAM
> as the swap block device.
>
>   ~ # perf record taskset -c 4 ./a.out
>   [ perf record: Woken up 10 times to write data ]
>   [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
>   ~ # perf report
>   # To display the perf.data header info, please use --header/--header-only options.
>   # To display the perf.data header info, please use --header/--header-only options.
>   #
>   #
>   # Total Lost Samples: 0
>   #
>   # Samples: 60K of event 'cycles'
>   # Event count (approx.): 35706225414
>   #
>   # Overhead  Command  Shared Object      Symbol
>   # ........  .......  .................  .............................................................................
>   #
>      21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
>       8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
>       6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
>       6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
>       5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>       3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
>       3.49%  a.out    [kernel.kallsyms]  [k] memset64
>       1.63%  a.out    [kernel.kallsyms]  [k] clear_page
>       1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
>       1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
>       1.23%  a.out    [kernel.kallsyms]  [k] xas_load
>       1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
>
> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> swapping in/out a page mapped by only one process. If the
> page is mapped by multiple processes, typically, like more
> than 100 on a phone, the overhead would be much higher as
> we have to run tlb flush 100 times for one single page.
> Plus, tlb flush overhead will increase with the number
> of CPU cores due to the bad scalability of tlb shootdown
> in HW, so those ARM64 servers should expect much higher
> overhead.
>
> Further perf annonate shows 95% cpu time of ptep_clear_flush
> is actually used by the final dsb() to wait for the completion
> of tlb flush. This provides us a very good chance to leverage
> the existing batched tlb in kernel. The minimum modification
> is that we only send async tlbi in the first stage and we send
> dsb while we have to sync in the second stage.
>
> With the above simplest micro benchmark, collapsed time to
> finish the program decreases around 5%.
>
> Typical collapsed time w/o patch:
>   ~ # time taskset -c 4 ./a.out
>   0.21user 14.34system 0:14.69elapsed
> w/ patch:
>   ~ # time taskset -c 4 ./a.out
>   0.22user 13.45system 0:13.80elapsed
>
> Also, Yicong Yang added the following observation.
> 	Tested with benchmark in the commit on Kunpeng920 arm64 server,
> 	observed an improvement around 12.5% with command
> 	`time ./swap_bench`.
> 		w/o		w/
> 	real	0m13.460s	0m11.771s
> 	user	0m0.248s	0m0.279s
> 	sys	0m12.039s	0m11.458s
>
> 	Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> 	which has been eliminated by this patch:
>
> 	[root@localhost yang]# perf record -- ./swap_bench && perf report
> 	[...]
> 	16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
>
> Cc: Jonathan Corbet <corbet@lwn.net>
> Cc: Nadav Amit <namit@vmware.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>

I tested on my kunpeng board too, looks good for now.

Reviewed-by: Kefeng Wang <wangkefeng.wang@huawei.com>

> ---
>   .../features/vm/TLB/arch-support.txt          |  2 +-
>   arch/arm64/Kconfig                            |  1 +
>   arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>   arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>   4 files changed, 40 insertions(+), 3 deletions(-)
>   create mode 100644 arch/arm64/include/asm/tlbbatch.h
>
> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> index 1c009312b9c1..2caf815d7c6c 100644
> --- a/Documentation/features/vm/TLB/arch-support.txt
> +++ b/Documentation/features/vm/TLB/arch-support.txt
> @@ -9,7 +9,7 @@
>       |       alpha: | TODO |
>       |         arc: | TODO |
>       |         arm: | TODO |
> -    |       arm64: | TODO |
> +    |       arm64: |  ok  |
>       |        csky: | TODO |
>       |     hexagon: | TODO |
>       |        ia64: | TODO |
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 571cc234d0b3..09d45cd6d665 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -93,6 +93,7 @@ config ARM64
>   	select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
>   	select ARCH_SUPPORTS_NUMA_BALANCING
>   	select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> +	select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
>   	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
>   	select ARCH_WANT_DEFAULT_BPF_JIT
>   	select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> new file mode 100644
> index 000000000000..fedb0b87b8db
> --- /dev/null
> +++ b/arch/arm64/include/asm/tlbbatch.h
> @@ -0,0 +1,12 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ARCH_ARM64_TLBBATCH_H
> +#define _ARCH_ARM64_TLBBATCH_H
> +
> +struct arch_tlbflush_unmap_batch {
> +	/*
> +	 * For arm64, HW can do tlb shootdown, so we don't
> +	 * need to record cpumask for sending IPI
> +	 */
> +};
> +
> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> index 412a3b9a3c25..23cbc987321a 100644
> --- a/arch/arm64/include/asm/tlbflush.h
> +++ b/arch/arm64/include/asm/tlbflush.h
> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
>   	dsb(ish);
>   }
>   
> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> +
> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
>   					 unsigned long uaddr)
>   {
>   	unsigned long addr;
>   
>   	dsb(ishst);
> -	addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> +	addr = __TLBI_VADDR(uaddr, ASID(mm));
>   	__tlbi(vale1is, addr);
>   	__tlbi_user(vale1is, addr);
>   }
>   
> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> +					 unsigned long uaddr)
> +{
> +	return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> +}
> +
>   static inline void flush_tlb_page(struct vm_area_struct *vma,
>   				  unsigned long uaddr)
>   {
> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
>   	dsb(ish);
>   }
>   
> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> +{
> +	return true;
> +}
> +
> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> +					struct mm_struct *mm,
> +					unsigned long uaddr)
> +{
> +	__flush_tlb_page_nosync(mm, uaddr);
> +}
> +
> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> +{
> +	dsb(ish);
> +}
> +
>   /*
>    * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
>    * necessarily a performance improvement.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> From: Barry Song <v-songbaohua@oppo.com>
> 
> on x86, batched and deferred tlb shootdown has lead to 90%
> performance increase on tlb shootdown. on arm64, HW can do
> tlb shootdown without software IPI. But sync tlbi is still
> quite expensive.
> 
> Even running a simplest program which requires swapout can
> prove this is true,
>  #include <sys/types.h>
>  #include <unistd.h>
>  #include <sys/mman.h>
>  #include <string.h>
> 
>  int main()
>  {
>  #define SIZE (1 * 1024 * 1024)
>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> 
>          memset(p, 0x88, SIZE);
> 
>          for (int k = 0; k < 10000; k++) {
>                  /* swap in */
>                  for (int i = 0; i < SIZE; i += 4096) {
>                          (void)p[i];
>                  }
> 
>                  /* swap out */
>                  madvise(p, SIZE, MADV_PAGEOUT);
>          }
>  }
> 
> Perf result on snapdragon 888 with 8 cores by using zRAM
> as the swap block device.
> 
>  ~ # perf record taskset -c 4 ./a.out
>  [ perf record: Woken up 10 times to write data ]
>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
>  ~ # perf report
>  # To display the perf.data header info, please use --header/--header-only options.
>  # To display the perf.data header info, please use --header/--header-only options.
>  #
>  #
>  # Total Lost Samples: 0
>  #
>  # Samples: 60K of event 'cycles'
>  # Event count (approx.): 35706225414
>  #
>  # Overhead  Command  Shared Object      Symbol
>  # ........  .......  .................  .............................................................................
>  #
>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
> 
> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> swapping in/out a page mapped by only one process. If the
> page is mapped by multiple processes, typically, like more
> than 100 on a phone, the overhead would be much higher as
> we have to run tlb flush 100 times for one single page.
> Plus, tlb flush overhead will increase with the number
> of CPU cores due to the bad scalability of tlb shootdown
> in HW, so those ARM64 servers should expect much higher
> overhead.
> 
> Further perf annonate shows 95% cpu time of ptep_clear_flush
> is actually used by the final dsb() to wait for the completion
> of tlb flush. This provides us a very good chance to leverage
> the existing batched tlb in kernel. The minimum modification
> is that we only send async tlbi in the first stage and we send
> dsb while we have to sync in the second stage.
> 
> With the above simplest micro benchmark, collapsed time to
> finish the program decreases around 5%.
> 
> Typical collapsed time w/o patch:
>  ~ # time taskset -c 4 ./a.out
>  0.21user 14.34system 0:14.69elapsed
> w/ patch:
>  ~ # time taskset -c 4 ./a.out
>  0.22user 13.45system 0:13.80elapsed
> 
> Also, Yicong Yang added the following observation.
> 	Tested with benchmark in the commit on Kunpeng920 arm64 server,
> 	observed an improvement around 12.5% with command
> 	`time ./swap_bench`.
> 		w/o		w/
> 	real	0m13.460s	0m11.771s
> 	user	0m0.248s	0m0.279s
> 	sys	0m12.039s	0m11.458s
> 
> 	Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> 	which has been eliminated by this patch:
> 
> 	[root@localhost yang]# perf record -- ./swap_bench && perf report
> 	[...]
> 	16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
> 
> Cc: Jonathan Corbet <corbet@lwn.net>
> Cc: Nadav Amit <namit@vmware.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> ---
>  .../features/vm/TLB/arch-support.txt          |  2 +-
>  arch/arm64/Kconfig                            |  1 +
>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>  4 files changed, 40 insertions(+), 3 deletions(-)
>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> 
> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> index 1c009312b9c1..2caf815d7c6c 100644
> --- a/Documentation/features/vm/TLB/arch-support.txt
> +++ b/Documentation/features/vm/TLB/arch-support.txt
> @@ -9,7 +9,7 @@
>      |       alpha: | TODO |
>      |         arc: | TODO |
>      |         arm: | TODO |
> -    |       arm64: | TODO |
> +    |       arm64: |  ok  |
>      |        csky: | TODO |
>      |     hexagon: | TODO |
>      |        ia64: | TODO |
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 571cc234d0b3..09d45cd6d665 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -93,6 +93,7 @@ config ARM64
>  	select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
>  	select ARCH_SUPPORTS_NUMA_BALANCING
>  	select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> +	select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
>  	select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
>  	select ARCH_WANT_DEFAULT_BPF_JIT
>  	select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> new file mode 100644
> index 000000000000..fedb0b87b8db
> --- /dev/null
> +++ b/arch/arm64/include/asm/tlbbatch.h
> @@ -0,0 +1,12 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ARCH_ARM64_TLBBATCH_H
> +#define _ARCH_ARM64_TLBBATCH_H
> +
> +struct arch_tlbflush_unmap_batch {
> +	/*
> +	 * For arm64, HW can do tlb shootdown, so we don't
> +	 * need to record cpumask for sending IPI
> +	 */
> +};
> +
> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> index 412a3b9a3c25..23cbc987321a 100644
> --- a/arch/arm64/include/asm/tlbflush.h
> +++ b/arch/arm64/include/asm/tlbflush.h
> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
>  	dsb(ish);
>  }
>  
> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> +
> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
>  					 unsigned long uaddr)
>  {
>  	unsigned long addr;
>  
>  	dsb(ishst);
> -	addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> +	addr = __TLBI_VADDR(uaddr, ASID(mm));
>  	__tlbi(vale1is, addr);
>  	__tlbi_user(vale1is, addr);
>  }
>  
> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> +					 unsigned long uaddr)
> +{
> +	return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> +}
> +
>  static inline void flush_tlb_page(struct vm_area_struct *vma,
>  				  unsigned long uaddr)
>  {
> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
>  	dsb(ish);
>  }
>  
> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> +{
> +	return true;
> +}

Always defer and batch up TLB flush, unconditionally ?

> +
> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> +					struct mm_struct *mm,
> +					unsigned long uaddr)
> +{
> +	__flush_tlb_page_nosync(mm, uaddr);
> +}
> +
> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> +{
> +	dsb(ish);
> +}

Adding up __flush_tlb_page_nosync() without a corresponding dsb(ish) and
then doing once via arch_tlbbatch_flush() will have the same effect from
an architecture perspective ?

> +
>  /*
>   * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
>   * necessarily a performance improvement.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 8/22/22 13:51, Yicong Yang wrote:
> > From: Barry Song <v-songbaohua@oppo.com>
> >
> > on x86, batched and deferred tlb shootdown has lead to 90%
> > performance increase on tlb shootdown. on arm64, HW can do
> > tlb shootdown without software IPI. But sync tlbi is still
> > quite expensive.
> >
> > Even running a simplest program which requires swapout can
> > prove this is true,
> >  #include <sys/types.h>
> >  #include <unistd.h>
> >  #include <sys/mman.h>
> >  #include <string.h>
> >
> >  int main()
> >  {
> >  #define SIZE (1 * 1024 * 1024)
> >          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
> >                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> >
> >          memset(p, 0x88, SIZE);
> >
> >          for (int k = 0; k < 10000; k++) {
> >                  /* swap in */
> >                  for (int i = 0; i < SIZE; i += 4096) {
> >                          (void)p[i];
> >                  }
> >
> >                  /* swap out */
> >                  madvise(p, SIZE, MADV_PAGEOUT);
> >          }
> >  }
> >
> > Perf result on snapdragon 888 with 8 cores by using zRAM
> > as the swap block device.
> >
> >  ~ # perf record taskset -c 4 ./a.out
> >  [ perf record: Woken up 10 times to write data ]
> >  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
> >  ~ # perf report
> >  # To display the perf.data header info, please use --header/--header-only options.
> >  # To display the perf.data header info, please use --header/--header-only options.
> >  #
> >  #
> >  # Total Lost Samples: 0
> >  #
> >  # Samples: 60K of event 'cycles'
> >  # Event count (approx.): 35706225414
> >  #
> >  # Overhead  Command  Shared Object      Symbol
> >  # ........  .......  .................  .............................................................................
> >  #
> >     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
> >      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
> >      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
> >      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
> >      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
> >      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
> >      3.49%  a.out    [kernel.kallsyms]  [k] memset64
> >      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
> >      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
> >      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
> >      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
> >      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
> >
> > ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> > swapping in/out a page mapped by only one process. If the
> > page is mapped by multiple processes, typically, like more
> > than 100 on a phone, the overhead would be much higher as
> > we have to run tlb flush 100 times for one single page.
> > Plus, tlb flush overhead will increase with the number
> > of CPU cores due to the bad scalability of tlb shootdown
> > in HW, so those ARM64 servers should expect much higher
> > overhead.
> >
> > Further perf annonate shows 95% cpu time of ptep_clear_flush
> > is actually used by the final dsb() to wait for the completion
> > of tlb flush. This provides us a very good chance to leverage
> > the existing batched tlb in kernel. The minimum modification
> > is that we only send async tlbi in the first stage and we send
> > dsb while we have to sync in the second stage.
> >
> > With the above simplest micro benchmark, collapsed time to
> > finish the program decreases around 5%.
> >
> > Typical collapsed time w/o patch:
> >  ~ # time taskset -c 4 ./a.out
> >  0.21user 14.34system 0:14.69elapsed
> > w/ patch:
> >  ~ # time taskset -c 4 ./a.out
> >  0.22user 13.45system 0:13.80elapsed
> >
> > Also, Yicong Yang added the following observation.
> >       Tested with benchmark in the commit on Kunpeng920 arm64 server,
> >       observed an improvement around 12.5% with command
> >       `time ./swap_bench`.
> >               w/o             w/
> >       real    0m13.460s       0m11.771s
> >       user    0m0.248s        0m0.279s
> >       sys     0m12.039s       0m11.458s
> >
> >       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> >       which has been eliminated by this patch:
> >
> >       [root@localhost yang]# perf record -- ./swap_bench && perf report
> >       [...]
> >       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
> >
> > Cc: Jonathan Corbet <corbet@lwn.net>
> > Cc: Nadav Amit <namit@vmware.com>
> > Cc: Mel Gorman <mgorman@suse.de>
> > Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> > Tested-by: Xin Hao <xhao@linux.alibaba.com>
> > Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> > Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> > ---
> >  .../features/vm/TLB/arch-support.txt          |  2 +-
> >  arch/arm64/Kconfig                            |  1 +
> >  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
> >  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
> >  4 files changed, 40 insertions(+), 3 deletions(-)
> >  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> >
> > diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> > index 1c009312b9c1..2caf815d7c6c 100644
> > --- a/Documentation/features/vm/TLB/arch-support.txt
> > +++ b/Documentation/features/vm/TLB/arch-support.txt
> > @@ -9,7 +9,7 @@
> >      |       alpha: | TODO |
> >      |         arc: | TODO |
> >      |         arm: | TODO |
> > -    |       arm64: | TODO |
> > +    |       arm64: |  ok  |
> >      |        csky: | TODO |
> >      |     hexagon: | TODO |
> >      |        ia64: | TODO |
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 571cc234d0b3..09d45cd6d665 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -93,6 +93,7 @@ config ARM64
> >       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
> >       select ARCH_SUPPORTS_NUMA_BALANCING
> >       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> > +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> >       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
> >       select ARCH_WANT_DEFAULT_BPF_JIT
> >       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> > diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> > new file mode 100644
> > index 000000000000..fedb0b87b8db
> > --- /dev/null
> > +++ b/arch/arm64/include/asm/tlbbatch.h
> > @@ -0,0 +1,12 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +#ifndef _ARCH_ARM64_TLBBATCH_H
> > +#define _ARCH_ARM64_TLBBATCH_H
> > +
> > +struct arch_tlbflush_unmap_batch {
> > +     /*
> > +      * For arm64, HW can do tlb shootdown, so we don't
> > +      * need to record cpumask for sending IPI
> > +      */
> > +};
> > +
> > +#endif /* _ARCH_ARM64_TLBBATCH_H */
> > diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> > index 412a3b9a3c25..23cbc987321a 100644
> > --- a/arch/arm64/include/asm/tlbflush.h
> > +++ b/arch/arm64/include/asm/tlbflush.h
> > @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> >       dsb(ish);
> >  }
> >
> > -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> > +
> > +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
> >                                        unsigned long uaddr)
> >  {
> >       unsigned long addr;
> >
> >       dsb(ishst);
> > -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> > +     addr = __TLBI_VADDR(uaddr, ASID(mm));
> >       __tlbi(vale1is, addr);
> >       __tlbi_user(vale1is, addr);
> >  }
> >
> > +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> > +                                      unsigned long uaddr)
> > +{
> > +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> > +}
> > +
> >  static inline void flush_tlb_page(struct vm_area_struct *vma,
> >                                 unsigned long uaddr)
> >  {
> > @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
> >       dsb(ish);
> >  }
> >
> > +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> > +{
> > +     return true;
> > +}
>
> Always defer and batch up TLB flush, unconditionally ?

My understanding is we actually don't need tlbbatch for a machine with one
or two cores as the tlb flush is not expensive. even for a system with four
cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
when we have 8 cores, we see the obvious cost of tlb flush. for a server with
100 crores, the cost is incredibly huge.

But, we can hardly write source code to differentiate machines according to
how many cores a machine has, especially when cores can be hot-plugged.

>
> > +
> > +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> > +                                     struct mm_struct *mm,
> > +                                     unsigned long uaddr)
> > +{
> > +     __flush_tlb_page_nosync(mm, uaddr);
> > +}
> > +
> > +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> > +{
> > +     dsb(ish);
> > +}
>
> Adding up __flush_tlb_page_nosync() without a corresponding dsb(ish) and
> then doing once via arch_tlbbatch_flush() will have the same effect from
> an architecture perspective ?

The difference is we drop the cost of lots of single tlb flush. we
only need to sync
when we have to sync. dsb(ish) guarantees the completion of previous
multiple tlb
flush instructions.

>
> > +
> >  /*
> >   * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
> >   * necessarily a performance improvement.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Yicong Yang]

On 2022/9/9 13:35, Barry Song wrote:
> On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
>>
>>
>>
>> On 8/22/22 13:51, Yicong Yang wrote:
>>> From: Barry Song <v-songbaohua@oppo.com>
>>>
>>> on x86, batched and deferred tlb shootdown has lead to 90%
>>> performance increase on tlb shootdown. on arm64, HW can do
>>> tlb shootdown without software IPI. But sync tlbi is still
>>> quite expensive.
>>>
>>> Even running a simplest program which requires swapout can
>>> prove this is true,
>>>  #include <sys/types.h>
>>>  #include <unistd.h>
>>>  #include <sys/mman.h>
>>>  #include <string.h>
>>>
>>>  int main()
>>>  {
>>>  #define SIZE (1 * 1024 * 1024)
>>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
>>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>>>
>>>          memset(p, 0x88, SIZE);
>>>
>>>          for (int k = 0; k < 10000; k++) {
>>>                  /* swap in */
>>>                  for (int i = 0; i < SIZE; i += 4096) {
>>>                          (void)p[i];
>>>                  }
>>>
>>>                  /* swap out */
>>>                  madvise(p, SIZE, MADV_PAGEOUT);
>>>          }
>>>  }
>>>
>>> Perf result on snapdragon 888 with 8 cores by using zRAM
>>> as the swap block device.
>>>
>>>  ~ # perf record taskset -c 4 ./a.out
>>>  [ perf record: Woken up 10 times to write data ]
>>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
>>>  ~ # perf report
>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>  #
>>>  #
>>>  # Total Lost Samples: 0
>>>  #
>>>  # Samples: 60K of event 'cycles'
>>>  # Event count (approx.): 35706225414
>>>  #
>>>  # Overhead  Command  Shared Object      Symbol
>>>  # ........  .......  .................  .............................................................................
>>>  #
>>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
>>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
>>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
>>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
>>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
>>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
>>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
>>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
>>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
>>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
>>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
>>>
>>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
>>> swapping in/out a page mapped by only one process. If the
>>> page is mapped by multiple processes, typically, like more
>>> than 100 on a phone, the overhead would be much higher as
>>> we have to run tlb flush 100 times for one single page.
>>> Plus, tlb flush overhead will increase with the number
>>> of CPU cores due to the bad scalability of tlb shootdown
>>> in HW, so those ARM64 servers should expect much higher
>>> overhead.
>>>
>>> Further perf annonate shows 95% cpu time of ptep_clear_flush
>>> is actually used by the final dsb() to wait for the completion
>>> of tlb flush. This provides us a very good chance to leverage
>>> the existing batched tlb in kernel. The minimum modification
>>> is that we only send async tlbi in the first stage and we send
>>> dsb while we have to sync in the second stage.
>>>
>>> With the above simplest micro benchmark, collapsed time to
>>> finish the program decreases around 5%.
>>>
>>> Typical collapsed time w/o patch:
>>>  ~ # time taskset -c 4 ./a.out
>>>  0.21user 14.34system 0:14.69elapsed
>>> w/ patch:
>>>  ~ # time taskset -c 4 ./a.out
>>>  0.22user 13.45system 0:13.80elapsed
>>>
>>> Also, Yicong Yang added the following observation.
>>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
>>>       observed an improvement around 12.5% with command
>>>       `time ./swap_bench`.
>>>               w/o             w/
>>>       real    0m13.460s       0m11.771s
>>>       user    0m0.248s        0m0.279s
>>>       sys     0m12.039s       0m11.458s
>>>
>>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
>>>       which has been eliminated by this patch:
>>>
>>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
>>>       [...]
>>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
>>>
>>> Cc: Jonathan Corbet <corbet@lwn.net>
>>> Cc: Nadav Amit <namit@vmware.com>
>>> Cc: Mel Gorman <mgorman@suse.de>
>>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
>>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
>>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
>>> ---
>>>  .../features/vm/TLB/arch-support.txt          |  2 +-
>>>  arch/arm64/Kconfig                            |  1 +
>>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>>>  4 files changed, 40 insertions(+), 3 deletions(-)
>>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
>>>
>>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
>>> index 1c009312b9c1..2caf815d7c6c 100644
>>> --- a/Documentation/features/vm/TLB/arch-support.txt
>>> +++ b/Documentation/features/vm/TLB/arch-support.txt
>>> @@ -9,7 +9,7 @@
>>>      |       alpha: | TODO |
>>>      |         arc: | TODO |
>>>      |         arm: | TODO |
>>> -    |       arm64: | TODO |
>>> +    |       arm64: |  ok  |
>>>      |        csky: | TODO |
>>>      |     hexagon: | TODO |
>>>      |        ia64: | TODO |
>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>>> index 571cc234d0b3..09d45cd6d665 100644
>>> --- a/arch/arm64/Kconfig
>>> +++ b/arch/arm64/Kconfig
>>> @@ -93,6 +93,7 @@ config ARM64
>>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
>>>       select ARCH_SUPPORTS_NUMA_BALANCING
>>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
>>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
>>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
>>>       select ARCH_WANT_DEFAULT_BPF_JIT
>>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
>>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
>>> new file mode 100644
>>> index 000000000000..fedb0b87b8db
>>> --- /dev/null
>>> +++ b/arch/arm64/include/asm/tlbbatch.h
>>> @@ -0,0 +1,12 @@
>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>> +#ifndef _ARCH_ARM64_TLBBATCH_H
>>> +#define _ARCH_ARM64_TLBBATCH_H
>>> +
>>> +struct arch_tlbflush_unmap_batch {
>>> +     /*
>>> +      * For arm64, HW can do tlb shootdown, so we don't
>>> +      * need to record cpumask for sending IPI
>>> +      */
>>> +};
>>> +
>>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
>>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
>>> index 412a3b9a3c25..23cbc987321a 100644
>>> --- a/arch/arm64/include/asm/tlbflush.h
>>> +++ b/arch/arm64/include/asm/tlbflush.h
>>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
>>>       dsb(ish);
>>>  }
>>>
>>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>> +
>>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
>>>                                        unsigned long uaddr)
>>>  {
>>>       unsigned long addr;
>>>
>>>       dsb(ishst);
>>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
>>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
>>>       __tlbi(vale1is, addr);
>>>       __tlbi_user(vale1is, addr);
>>>  }
>>>
>>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>> +                                      unsigned long uaddr)
>>> +{
>>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
>>> +}
>>> +
>>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>                                 unsigned long uaddr)
>>>  {
>>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>       dsb(ish);
>>>  }
>>>
>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>>> +{
>>> +     return true;
>>> +}
>>
>> Always defer and batch up TLB flush, unconditionally ?
> 
> My understanding is we actually don't need tlbbatch for a machine with one
> or two cores as the tlb flush is not expensive. even for a system with four
> cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> 100 crores, the cost is incredibly huge.
> 
> But, we can hardly write source code to differentiate machines according to
> how many cores a machine has, especially when cores can be hot-plugged.
> 

Another thing is that we're not recording mm_cpumask() on arm64 so for now we cannot do
the check like x86 and others.

>>
>>> +
>>> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
>>> +                                     struct mm_struct *mm,
>>> +                                     unsigned long uaddr)
>>> +{
>>> +     __flush_tlb_page_nosync(mm, uaddr);
>>> +}
>>> +
>>> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
>>> +{
>>> +     dsb(ish);
>>> +}
>>
>> Adding up __flush_tlb_page_nosync() without a corresponding dsb(ish) and
>> then doing once via arch_tlbbatch_flush() will have the same effect from
>> an architecture perspective ?
> 
> The difference is we drop the cost of lots of single tlb flush. we
> only need to sync
> when we have to sync. dsb(ish) guarantees the completion of previous
> multiple tlb
> flush instructions.
> 
>>
>>> +
>>>  /*
>>>   * This is meant to avoid soft lock-ups on large TLB flushing ranges and not
>>>   * necessarily a performance improvement.
> 
> Thanks
> Barry
> .
> 

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/9/22 11:05, Barry Song wrote:
> On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
>>
>>
>>
>> On 8/22/22 13:51, Yicong Yang wrote:
>>> From: Barry Song <v-songbaohua@oppo.com>
>>>
>>> on x86, batched and deferred tlb shootdown has lead to 90%
>>> performance increase on tlb shootdown. on arm64, HW can do
>>> tlb shootdown without software IPI. But sync tlbi is still
>>> quite expensive.
>>>
>>> Even running a simplest program which requires swapout can
>>> prove this is true,
>>>  #include <sys/types.h>
>>>  #include <unistd.h>
>>>  #include <sys/mman.h>
>>>  #include <string.h>
>>>
>>>  int main()
>>>  {
>>>  #define SIZE (1 * 1024 * 1024)
>>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
>>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>>>
>>>          memset(p, 0x88, SIZE);
>>>
>>>          for (int k = 0; k < 10000; k++) {
>>>                  /* swap in */
>>>                  for (int i = 0; i < SIZE; i += 4096) {
>>>                          (void)p[i];
>>>                  }
>>>
>>>                  /* swap out */
>>>                  madvise(p, SIZE, MADV_PAGEOUT);
>>>          }
>>>  }
>>>
>>> Perf result on snapdragon 888 with 8 cores by using zRAM
>>> as the swap block device.
>>>
>>>  ~ # perf record taskset -c 4 ./a.out
>>>  [ perf record: Woken up 10 times to write data ]
>>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
>>>  ~ # perf report
>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>  #
>>>  #
>>>  # Total Lost Samples: 0
>>>  #
>>>  # Samples: 60K of event 'cycles'
>>>  # Event count (approx.): 35706225414
>>>  #
>>>  # Overhead  Command  Shared Object      Symbol
>>>  # ........  .......  .................  .............................................................................
>>>  #
>>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
>>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
>>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
>>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
>>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
>>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
>>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
>>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
>>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
>>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
>>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
>>>
>>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
>>> swapping in/out a page mapped by only one process. If the
>>> page is mapped by multiple processes, typically, like more
>>> than 100 on a phone, the overhead would be much higher as
>>> we have to run tlb flush 100 times for one single page.
>>> Plus, tlb flush overhead will increase with the number
>>> of CPU cores due to the bad scalability of tlb shootdown
>>> in HW, so those ARM64 servers should expect much higher
>>> overhead.
>>>
>>> Further perf annonate shows 95% cpu time of ptep_clear_flush
>>> is actually used by the final dsb() to wait for the completion
>>> of tlb flush. This provides us a very good chance to leverage
>>> the existing batched tlb in kernel. The minimum modification
>>> is that we only send async tlbi in the first stage and we send
>>> dsb while we have to sync in the second stage.
>>>
>>> With the above simplest micro benchmark, collapsed time to
>>> finish the program decreases around 5%.
>>>
>>> Typical collapsed time w/o patch:
>>>  ~ # time taskset -c 4 ./a.out
>>>  0.21user 14.34system 0:14.69elapsed
>>> w/ patch:
>>>  ~ # time taskset -c 4 ./a.out
>>>  0.22user 13.45system 0:13.80elapsed
>>>
>>> Also, Yicong Yang added the following observation.
>>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
>>>       observed an improvement around 12.5% with command
>>>       `time ./swap_bench`.
>>>               w/o             w/
>>>       real    0m13.460s       0m11.771s
>>>       user    0m0.248s        0m0.279s
>>>       sys     0m12.039s       0m11.458s
>>>
>>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
>>>       which has been eliminated by this patch:
>>>
>>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
>>>       [...]
>>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
>>>
>>> Cc: Jonathan Corbet <corbet@lwn.net>
>>> Cc: Nadav Amit <namit@vmware.com>
>>> Cc: Mel Gorman <mgorman@suse.de>
>>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
>>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
>>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
>>> ---
>>>  .../features/vm/TLB/arch-support.txt          |  2 +-
>>>  arch/arm64/Kconfig                            |  1 +
>>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>>>  4 files changed, 40 insertions(+), 3 deletions(-)
>>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
>>>
>>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
>>> index 1c009312b9c1..2caf815d7c6c 100644
>>> --- a/Documentation/features/vm/TLB/arch-support.txt
>>> +++ b/Documentation/features/vm/TLB/arch-support.txt
>>> @@ -9,7 +9,7 @@
>>>      |       alpha: | TODO |
>>>      |         arc: | TODO |
>>>      |         arm: | TODO |
>>> -    |       arm64: | TODO |
>>> +    |       arm64: |  ok  |
>>>      |        csky: | TODO |
>>>      |     hexagon: | TODO |
>>>      |        ia64: | TODO |
>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>>> index 571cc234d0b3..09d45cd6d665 100644
>>> --- a/arch/arm64/Kconfig
>>> +++ b/arch/arm64/Kconfig
>>> @@ -93,6 +93,7 @@ config ARM64
>>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
>>>       select ARCH_SUPPORTS_NUMA_BALANCING
>>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
>>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
>>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
>>>       select ARCH_WANT_DEFAULT_BPF_JIT
>>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
>>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
>>> new file mode 100644
>>> index 000000000000..fedb0b87b8db
>>> --- /dev/null
>>> +++ b/arch/arm64/include/asm/tlbbatch.h
>>> @@ -0,0 +1,12 @@
>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>> +#ifndef _ARCH_ARM64_TLBBATCH_H
>>> +#define _ARCH_ARM64_TLBBATCH_H
>>> +
>>> +struct arch_tlbflush_unmap_batch {
>>> +     /*
>>> +      * For arm64, HW can do tlb shootdown, so we don't
>>> +      * need to record cpumask for sending IPI
>>> +      */
>>> +};
>>> +
>>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
>>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
>>> index 412a3b9a3c25..23cbc987321a 100644
>>> --- a/arch/arm64/include/asm/tlbflush.h
>>> +++ b/arch/arm64/include/asm/tlbflush.h
>>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
>>>       dsb(ish);
>>>  }
>>>
>>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>> +
>>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
>>>                                        unsigned long uaddr)
>>>  {
>>>       unsigned long addr;
>>>
>>>       dsb(ishst);
>>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
>>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
>>>       __tlbi(vale1is, addr);
>>>       __tlbi_user(vale1is, addr);
>>>  }
>>>
>>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>> +                                      unsigned long uaddr)
>>> +{
>>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
>>> +}
>>> +
>>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>                                 unsigned long uaddr)
>>>  {
>>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>       dsb(ish);
>>>  }
>>>
>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>>> +{
>>> +     return true;
>>> +}
>>
>> Always defer and batch up TLB flush, unconditionally ?
> 
> My understanding is we actually don't need tlbbatch for a machine with one
> or two cores as the tlb flush is not expensive. even for a system with four
> cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> 100 crores, the cost is incredibly huge.

Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
expensive as well, while queuing up individual TLB flushes ?

The very idea behind TLB deferral is the opportunity it (might) provide
to accumulate address ranges and cpu masks so that individual TLB flush
can be replaced with a more cost effective range based TLB flush. Hence
I guess unless address range or cpumask based cost effective TLB flush
is available, deferral does not improve the unmap performance as much.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Thu, Sep 15, 2022 at 6:07 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 9/9/22 11:05, Barry Song wrote:
> > On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> > <anshuman.khandual@arm.com> wrote:
> >>
> >>
> >>
> >> On 8/22/22 13:51, Yicong Yang wrote:
> >>> From: Barry Song <v-songbaohua@oppo.com>
> >>>
> >>> on x86, batched and deferred tlb shootdown has lead to 90%
> >>> performance increase on tlb shootdown. on arm64, HW can do
> >>> tlb shootdown without software IPI. But sync tlbi is still
> >>> quite expensive.
> >>>
> >>> Even running a simplest program which requires swapout can
> >>> prove this is true,
> >>>  #include <sys/types.h>
> >>>  #include <unistd.h>
> >>>  #include <sys/mman.h>
> >>>  #include <string.h>
> >>>
> >>>  int main()
> >>>  {
> >>>  #define SIZE (1 * 1024 * 1024)
> >>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
> >>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> >>>
> >>>          memset(p, 0x88, SIZE);
> >>>
> >>>          for (int k = 0; k < 10000; k++) {
> >>>                  /* swap in */
> >>>                  for (int i = 0; i < SIZE; i += 4096) {
> >>>                          (void)p[i];
> >>>                  }
> >>>
> >>>                  /* swap out */
> >>>                  madvise(p, SIZE, MADV_PAGEOUT);
> >>>          }
> >>>  }
> >>>
> >>> Perf result on snapdragon 888 with 8 cores by using zRAM
> >>> as the swap block device.
> >>>
> >>>  ~ # perf record taskset -c 4 ./a.out
> >>>  [ perf record: Woken up 10 times to write data ]
> >>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
> >>>  ~ # perf report
> >>>  # To display the perf.data header info, please use --header/--header-only options.
> >>>  # To display the perf.data header info, please use --header/--header-only options.
> >>>  #
> >>>  #
> >>>  # Total Lost Samples: 0
> >>>  #
> >>>  # Samples: 60K of event 'cycles'
> >>>  # Event count (approx.): 35706225414
> >>>  #
> >>>  # Overhead  Command  Shared Object      Symbol
> >>>  # ........  .......  .................  .............................................................................
> >>>  #
> >>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
> >>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
> >>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
> >>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
> >>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
> >>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
> >>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
> >>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
> >>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
> >>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
> >>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
> >>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
> >>>
> >>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> >>> swapping in/out a page mapped by only one process. If the
> >>> page is mapped by multiple processes, typically, like more
> >>> than 100 on a phone, the overhead would be much higher as
> >>> we have to run tlb flush 100 times for one single page.
> >>> Plus, tlb flush overhead will increase with the number
> >>> of CPU cores due to the bad scalability of tlb shootdown
> >>> in HW, so those ARM64 servers should expect much higher
> >>> overhead.
> >>>
> >>> Further perf annonate shows 95% cpu time of ptep_clear_flush
> >>> is actually used by the final dsb() to wait for the completion
> >>> of tlb flush. This provides us a very good chance to leverage
> >>> the existing batched tlb in kernel. The minimum modification
> >>> is that we only send async tlbi in the first stage and we send
> >>> dsb while we have to sync in the second stage.
> >>>
> >>> With the above simplest micro benchmark, collapsed time to
> >>> finish the program decreases around 5%.
> >>>
> >>> Typical collapsed time w/o patch:
> >>>  ~ # time taskset -c 4 ./a.out
> >>>  0.21user 14.34system 0:14.69elapsed
> >>> w/ patch:
> >>>  ~ # time taskset -c 4 ./a.out
> >>>  0.22user 13.45system 0:13.80elapsed
> >>>
> >>> Also, Yicong Yang added the following observation.
> >>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
> >>>       observed an improvement around 12.5% with command
> >>>       `time ./swap_bench`.
> >>>               w/o             w/
> >>>       real    0m13.460s       0m11.771s
> >>>       user    0m0.248s        0m0.279s
> >>>       sys     0m12.039s       0m11.458s
> >>>
> >>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> >>>       which has been eliminated by this patch:
> >>>
> >>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
> >>>       [...]
> >>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
> >>>
> >>> Cc: Jonathan Corbet <corbet@lwn.net>
> >>> Cc: Nadav Amit <namit@vmware.com>
> >>> Cc: Mel Gorman <mgorman@suse.de>
> >>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> >>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> >>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> >>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> >>> ---
> >>>  .../features/vm/TLB/arch-support.txt          |  2 +-
> >>>  arch/arm64/Kconfig                            |  1 +
> >>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
> >>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
> >>>  4 files changed, 40 insertions(+), 3 deletions(-)
> >>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> >>>
> >>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> >>> index 1c009312b9c1..2caf815d7c6c 100644
> >>> --- a/Documentation/features/vm/TLB/arch-support.txt
> >>> +++ b/Documentation/features/vm/TLB/arch-support.txt
> >>> @@ -9,7 +9,7 @@
> >>>      |       alpha: | TODO |
> >>>      |         arc: | TODO |
> >>>      |         arm: | TODO |
> >>> -    |       arm64: | TODO |
> >>> +    |       arm64: |  ok  |
> >>>      |        csky: | TODO |
> >>>      |     hexagon: | TODO |
> >>>      |        ia64: | TODO |
> >>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> >>> index 571cc234d0b3..09d45cd6d665 100644
> >>> --- a/arch/arm64/Kconfig
> >>> +++ b/arch/arm64/Kconfig
> >>> @@ -93,6 +93,7 @@ config ARM64
> >>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
> >>>       select ARCH_SUPPORTS_NUMA_BALANCING
> >>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> >>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> >>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
> >>>       select ARCH_WANT_DEFAULT_BPF_JIT
> >>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> >>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> >>> new file mode 100644
> >>> index 000000000000..fedb0b87b8db
> >>> --- /dev/null
> >>> +++ b/arch/arm64/include/asm/tlbbatch.h
> >>> @@ -0,0 +1,12 @@
> >>> +/* SPDX-License-Identifier: GPL-2.0 */
> >>> +#ifndef _ARCH_ARM64_TLBBATCH_H
> >>> +#define _ARCH_ARM64_TLBBATCH_H
> >>> +
> >>> +struct arch_tlbflush_unmap_batch {
> >>> +     /*
> >>> +      * For arm64, HW can do tlb shootdown, so we don't
> >>> +      * need to record cpumask for sending IPI
> >>> +      */
> >>> +};
> >>> +
> >>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> >>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> >>> index 412a3b9a3c25..23cbc987321a 100644
> >>> --- a/arch/arm64/include/asm/tlbflush.h
> >>> +++ b/arch/arm64/include/asm/tlbflush.h
> >>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> >>>       dsb(ish);
> >>>  }
> >>>
> >>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>> +
> >>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
> >>>                                        unsigned long uaddr)
> >>>  {
> >>>       unsigned long addr;
> >>>
> >>>       dsb(ishst);
> >>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> >>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
> >>>       __tlbi(vale1is, addr);
> >>>       __tlbi_user(vale1is, addr);
> >>>  }
> >>>
> >>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>> +                                      unsigned long uaddr)
> >>> +{
> >>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> >>> +}
> >>> +
> >>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>>                                 unsigned long uaddr)
> >>>  {
> >>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>>       dsb(ish);
> >>>  }
> >>>
> >>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> >>> +{
> >>> +     return true;
> >>> +}
> >>
> >> Always defer and batch up TLB flush, unconditionally ?
> >
> > My understanding is we actually don't need tlbbatch for a machine with one
> > or two cores as the tlb flush is not expensive. even for a system with four
> > cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> > when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> > 100 crores, the cost is incredibly huge.
>
> Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
> one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
> expensive as well, while queuing up individual TLB flushes ?

This one is much much cheaper as it is not waiting for the
completion of tlbi. waiting for the completion of tlbi is a big
deal in arm64, thus, similar optimization can be seen here

3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
when page aging").


>
> The very idea behind TLB deferral is the opportunity it (might) provide
> to accumulate address ranges and cpu masks so that individual TLB flush
> can be replaced with a more cost effective range based TLB flush. Hence
> I guess unless address range or cpumask based cost effective TLB flush
> is available, deferral does not improve the unmap performance as much.


After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
from all cpus in the system, tlbi is not scalable. The point here is that we
avoid waiting for each individual TLBi. Alternatively, they are batched. If
you read the benchmark in the commit log, you can find the great decline
in the cost to swap out a page.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Nadav Amit]

> On Sep 14, 2022, at 11:42 PM, Barry Song <21cnbao@gmail.com> wrote:
> 
>> 
>> The very idea behind TLB deferral is the opportunity it (might) provide
>> to accumulate address ranges and cpu masks so that individual TLB flush
>> can be replaced with a more cost effective range based TLB flush. Hence
>> I guess unless address range or cpumask based cost effective TLB flush
>> is available, deferral does not improve the unmap performance as much.
> 
> 
> After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
> from all cpus in the system, tlbi is not scalable. The point here is that we
> avoid waiting for each individual TLBi. Alternatively, they are batched. If
> you read the benchmark in the commit log, you can find the great decline
> in the cost to swap out a page.

Just a minor correction: arch_tlbbatch_flush() does not collect ranges.
On x86 it only accumulate CPU mask.


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/15/22 20:01, Nadav Amit wrote:
> 
> 
>> On Sep 14, 2022, at 11:42 PM, Barry Song <21cnbao@gmail.com> wrote:
>>
>>>
>>> The very idea behind TLB deferral is the opportunity it (might) provide
>>> to accumulate address ranges and cpu masks so that individual TLB flush
>>> can be replaced with a more cost effective range based TLB flush. Hence
>>> I guess unless address range or cpumask based cost effective TLB flush
>>> is available, deferral does not improve the unmap performance as much.
>>
>>
>> After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
>> from all cpus in the system, tlbi is not scalable. The point here is that we
>> avoid waiting for each individual TLBi. Alternatively, they are batched. If
>> you read the benchmark in the commit log, you can find the great decline
>> in the cost to swap out a page.
> 
> Just a minor correction: arch_tlbbatch_flush() does not collect ranges.
> On x86 it only accumulate CPU mask.

Thanks Nadav for the clarification.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/15/22 12:12, Barry Song wrote:
> On Thu, Sep 15, 2022 at 6:07 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
>>
>>
>>
>> On 9/9/22 11:05, Barry Song wrote:
>>> On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
>>> <anshuman.khandual@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 8/22/22 13:51, Yicong Yang wrote:
>>>>> From: Barry Song <v-songbaohua@oppo.com>
>>>>>
>>>>> on x86, batched and deferred tlb shootdown has lead to 90%
>>>>> performance increase on tlb shootdown. on arm64, HW can do
>>>>> tlb shootdown without software IPI. But sync tlbi is still
>>>>> quite expensive.
>>>>>
>>>>> Even running a simplest program which requires swapout can
>>>>> prove this is true,
>>>>>  #include <sys/types.h>
>>>>>  #include <unistd.h>
>>>>>  #include <sys/mman.h>
>>>>>  #include <string.h>
>>>>>
>>>>>  int main()
>>>>>  {
>>>>>  #define SIZE (1 * 1024 * 1024)
>>>>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
>>>>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>>>>>
>>>>>          memset(p, 0x88, SIZE);
>>>>>
>>>>>          for (int k = 0; k < 10000; k++) {
>>>>>                  /* swap in */
>>>>>                  for (int i = 0; i < SIZE; i += 4096) {
>>>>>                          (void)p[i];
>>>>>                  }
>>>>>
>>>>>                  /* swap out */
>>>>>                  madvise(p, SIZE, MADV_PAGEOUT);
>>>>>          }
>>>>>  }
>>>>>
>>>>> Perf result on snapdragon 888 with 8 cores by using zRAM
>>>>> as the swap block device.
>>>>>
>>>>>  ~ # perf record taskset -c 4 ./a.out
>>>>>  [ perf record: Woken up 10 times to write data ]
>>>>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
>>>>>  ~ # perf report
>>>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>>>  # To display the perf.data header info, please use --header/--header-only options.
>>>>>  #
>>>>>  #
>>>>>  # Total Lost Samples: 0
>>>>>  #
>>>>>  # Samples: 60K of event 'cycles'
>>>>>  # Event count (approx.): 35706225414
>>>>>  #
>>>>>  # Overhead  Command  Shared Object      Symbol
>>>>>  # ........  .......  .................  .............................................................................
>>>>>  #
>>>>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
>>>>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
>>>>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
>>>>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
>>>>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>>>>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
>>>>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
>>>>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
>>>>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
>>>>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
>>>>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
>>>>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
>>>>>
>>>>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
>>>>> swapping in/out a page mapped by only one process. If the
>>>>> page is mapped by multiple processes, typically, like more
>>>>> than 100 on a phone, the overhead would be much higher as
>>>>> we have to run tlb flush 100 times for one single page.
>>>>> Plus, tlb flush overhead will increase with the number
>>>>> of CPU cores due to the bad scalability of tlb shootdown
>>>>> in HW, so those ARM64 servers should expect much higher
>>>>> overhead.
>>>>>
>>>>> Further perf annonate shows 95% cpu time of ptep_clear_flush
>>>>> is actually used by the final dsb() to wait for the completion
>>>>> of tlb flush. This provides us a very good chance to leverage
>>>>> the existing batched tlb in kernel. The minimum modification
>>>>> is that we only send async tlbi in the first stage and we send
>>>>> dsb while we have to sync in the second stage.
>>>>>
>>>>> With the above simplest micro benchmark, collapsed time to
>>>>> finish the program decreases around 5%.
>>>>>
>>>>> Typical collapsed time w/o patch:
>>>>>  ~ # time taskset -c 4 ./a.out
>>>>>  0.21user 14.34system 0:14.69elapsed
>>>>> w/ patch:
>>>>>  ~ # time taskset -c 4 ./a.out
>>>>>  0.22user 13.45system 0:13.80elapsed
>>>>>
>>>>> Also, Yicong Yang added the following observation.
>>>>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
>>>>>       observed an improvement around 12.5% with command
>>>>>       `time ./swap_bench`.
>>>>>               w/o             w/
>>>>>       real    0m13.460s       0m11.771s
>>>>>       user    0m0.248s        0m0.279s
>>>>>       sys     0m12.039s       0m11.458s
>>>>>
>>>>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
>>>>>       which has been eliminated by this patch:
>>>>>
>>>>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
>>>>>       [...]
>>>>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
>>>>>
>>>>> Cc: Jonathan Corbet <corbet@lwn.net>
>>>>> Cc: Nadav Amit <namit@vmware.com>
>>>>> Cc: Mel Gorman <mgorman@suse.de>
>>>>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
>>>>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
>>>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
>>>>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
>>>>> ---
>>>>>  .../features/vm/TLB/arch-support.txt          |  2 +-
>>>>>  arch/arm64/Kconfig                            |  1 +
>>>>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>>>>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>>>>>  4 files changed, 40 insertions(+), 3 deletions(-)
>>>>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
>>>>>
>>>>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
>>>>> index 1c009312b9c1..2caf815d7c6c 100644
>>>>> --- a/Documentation/features/vm/TLB/arch-support.txt
>>>>> +++ b/Documentation/features/vm/TLB/arch-support.txt
>>>>> @@ -9,7 +9,7 @@
>>>>>      |       alpha: | TODO |
>>>>>      |         arc: | TODO |
>>>>>      |         arm: | TODO |
>>>>> -    |       arm64: | TODO |
>>>>> +    |       arm64: |  ok  |
>>>>>      |        csky: | TODO |
>>>>>      |     hexagon: | TODO |
>>>>>      |        ia64: | TODO |
>>>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>>>>> index 571cc234d0b3..09d45cd6d665 100644
>>>>> --- a/arch/arm64/Kconfig
>>>>> +++ b/arch/arm64/Kconfig
>>>>> @@ -93,6 +93,7 @@ config ARM64
>>>>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
>>>>>       select ARCH_SUPPORTS_NUMA_BALANCING
>>>>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
>>>>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
>>>>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
>>>>>       select ARCH_WANT_DEFAULT_BPF_JIT
>>>>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
>>>>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
>>>>> new file mode 100644
>>>>> index 000000000000..fedb0b87b8db
>>>>> --- /dev/null
>>>>> +++ b/arch/arm64/include/asm/tlbbatch.h
>>>>> @@ -0,0 +1,12 @@
>>>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>>>> +#ifndef _ARCH_ARM64_TLBBATCH_H
>>>>> +#define _ARCH_ARM64_TLBBATCH_H
>>>>> +
>>>>> +struct arch_tlbflush_unmap_batch {
>>>>> +     /*
>>>>> +      * For arm64, HW can do tlb shootdown, so we don't
>>>>> +      * need to record cpumask for sending IPI
>>>>> +      */
>>>>> +};
>>>>> +
>>>>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
>>>>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
>>>>> index 412a3b9a3c25..23cbc987321a 100644
>>>>> --- a/arch/arm64/include/asm/tlbflush.h
>>>>> +++ b/arch/arm64/include/asm/tlbflush.h
>>>>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
>>>>>       dsb(ish);
>>>>>  }
>>>>>
>>>>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>>>> +
>>>>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
>>>>>                                        unsigned long uaddr)
>>>>>  {
>>>>>       unsigned long addr;
>>>>>
>>>>>       dsb(ishst);
>>>>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
>>>>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
>>>>>       __tlbi(vale1is, addr);
>>>>>       __tlbi_user(vale1is, addr);
>>>>>  }
>>>>>
>>>>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
>>>>> +                                      unsigned long uaddr)
>>>>> +{
>>>>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
>>>>> +}
>>>>> +
>>>>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>>>                                 unsigned long uaddr)
>>>>>  {
>>>>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
>>>>>       dsb(ish);
>>>>>  }
>>>>>
>>>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>>>>> +{
>>>>> +     return true;
>>>>> +}
>>>>
>>>> Always defer and batch up TLB flush, unconditionally ?
>>>
>>> My understanding is we actually don't need tlbbatch for a machine with one
>>> or two cores as the tlb flush is not expensive. even for a system with four
>>> cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
>>> when we have 8 cores, we see the obvious cost of tlb flush. for a server with
>>> 100 crores, the cost is incredibly huge.
>>
>> Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
>> one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
>> expensive as well, while queuing up individual TLB flushes ?
> 
> This one is much much cheaper as it is not waiting for the
> completion of tlbi. waiting for the completion of tlbi is a big
> deal in arm64, thus, similar optimization can be seen here
> 
> 3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
> when page aging").
> 
> 
>>
>> The very idea behind TLB deferral is the opportunity it (might) provide
>> to accumulate address ranges and cpu masks so that individual TLB flush
>> can be replaced with a more cost effective range based TLB flush. Hence
>> I guess unless address range or cpumask based cost effective TLB flush
>> is available, deferral does not improve the unmap performance as much.
> 
> 
> After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
> from all cpus in the system, tlbi is not scalable. The point here is that we
> avoid waiting for each individual TLBi. Alternatively, they are batched. If
> you read the benchmark in the commit log, you can find the great decline
> in the cost to swap out a page.

Alright, although collecting and deferring 'dsb(ish)' to the very end, does
not feel like a direct fit case for ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH but I
guess it can be used to improve unmap performance on arm64.

But is this 'dsb(ish)' deferral architecturally valid ?

Let's examine single page unmap path via try_to_unmap_one(). 

should_defer_flush() {
	ptep_get_and_clear()		
	set_tlb_ubc_flush_pending()
		arch_tlbbatch_add_mm()
			__flush_tlb_page_nosync()
} else {
	ptep_clear_flush()
		ptep_get_and_clear()
		flush_tlb_page()
			flush_tlb_page_nosync()
				__flush_tlb_page_nosync()
			dsb(ish)
}

__flush_tlb_page_nosync()
{
	dsb(ishst);
	addr = __TLBI_VADDR(uaddr, ASID(mm));
 	__tlbi(vale1is, addr);
 	__tlbi_user(vale1is, addr);
}

Currently without TLB deferral, 'dsb(ish)' gets executed just after __tlbi()
and __tlbi_user(), because __flush_tlb_page_nosync() is an inline function.

#define __TLBI_0(op, arg) asm (ARM64_ASM_PREAMBLE                              \
                               "tlbi " #op "\n"                                \
                   ALTERNATIVE("nop\n                   nop",                  \
                               "dsb ish\n               tlbi " #op,            \
                               ARM64_WORKAROUND_REPEAT_TLBI,                   \
                               CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
                            : : )

#define __TLBI_1(op, arg) asm (ARM64_ASM_PREAMBLE                              \
                               "tlbi " #op ", %0\n"                            \
                   ALTERNATIVE("nop\n                   nop",                  \
                               "dsb ish\n               tlbi " #op ", %0",     \
                               ARM64_WORKAROUND_REPEAT_TLBI,                   \
                               CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
                            : : "r" (arg))

#define __TLBI_N(op, arg, n, ...) __TLBI_##n(op, arg)

#define __tlbi(op, ...)         __TLBI_N(op, ##__VA_ARGS__, 1, 0)

#define __tlbi_user(op, arg) do {                                               \
        if (arm64_kernel_unmapped_at_el0())                                     \
                __tlbi(op, (arg) | USER_ASID_FLAG);                             \
} while (0)

There is already a 'dsb(ish)' in between two subsequent TLB operations in
case ARM64_WORKAROUND_REPEAT_TLBI is detected on the system. Hence I guess
deferral should not enabled on such systems ?

But with deferral enabled, 'dsb(ish)' will be executed in arch_tlbbatch_flush()
via try_to_unmap_flush[_dirty](). There might be random number of instructions
in between __tlbi()/__tlbi_user() i.e 'tlbi' instructions and final 'dsb(ish)'.
Just wondering, if such 'detached in time with other instructions in between'
'tlbi' and 'dsb(ish)', is architecturally valid ?

There is a comment in 'struct tlbflush_unmap_batch'.

        /*
         * The arch code makes the following promise: generic code can modify a
         * PTE, then call arch_tlbbatch_add_mm() (which internally provides all
         * needed barriers), then call arch_tlbbatch_flush(), and the entries
         * will be flushed on all CPUs by the time that arch_tlbbatch_flush()
         * returns.
         */


It expects arch_tlbbatch_add_mm() to provide all barriers, hence wondering if
that would include just the first 'dsb(isht)' not the subsequent 'dsb(ish)' ?

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Mon, Sep 19, 2022 at 4:24 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 9/15/22 12:12, Barry Song wrote:
> > On Thu, Sep 15, 2022 at 6:07 PM Anshuman Khandual
> > <anshuman.khandual@arm.com> wrote:
> >>
> >>
> >>
> >> On 9/9/22 11:05, Barry Song wrote:
> >>> On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> >>> <anshuman.khandual@arm.com> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 8/22/22 13:51, Yicong Yang wrote:
> >>>>> From: Barry Song <v-songbaohua@oppo.com>
> >>>>>
> >>>>> on x86, batched and deferred tlb shootdown has lead to 90%
> >>>>> performance increase on tlb shootdown. on arm64, HW can do
> >>>>> tlb shootdown without software IPI. But sync tlbi is still
> >>>>> quite expensive.
> >>>>>
> >>>>> Even running a simplest program which requires swapout can
> >>>>> prove this is true,
> >>>>>  #include <sys/types.h>
> >>>>>  #include <unistd.h>
> >>>>>  #include <sys/mman.h>
> >>>>>  #include <string.h>
> >>>>>
> >>>>>  int main()
> >>>>>  {
> >>>>>  #define SIZE (1 * 1024 * 1024)
> >>>>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
> >>>>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> >>>>>
> >>>>>          memset(p, 0x88, SIZE);
> >>>>>
> >>>>>          for (int k = 0; k < 10000; k++) {
> >>>>>                  /* swap in */
> >>>>>                  for (int i = 0; i < SIZE; i += 4096) {
> >>>>>                          (void)p[i];
> >>>>>                  }
> >>>>>
> >>>>>                  /* swap out */
> >>>>>                  madvise(p, SIZE, MADV_PAGEOUT);
> >>>>>          }
> >>>>>  }
> >>>>>
> >>>>> Perf result on snapdragon 888 with 8 cores by using zRAM
> >>>>> as the swap block device.
> >>>>>
> >>>>>  ~ # perf record taskset -c 4 ./a.out
> >>>>>  [ perf record: Woken up 10 times to write data ]
> >>>>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
> >>>>>  ~ # perf report
> >>>>>  # To display the perf.data header info, please use --header/--header-only options.
> >>>>>  # To display the perf.data header info, please use --header/--header-only options.
> >>>>>  #
> >>>>>  #
> >>>>>  # Total Lost Samples: 0
> >>>>>  #
> >>>>>  # Samples: 60K of event 'cycles'
> >>>>>  # Event count (approx.): 35706225414
> >>>>>  #
> >>>>>  # Overhead  Command  Shared Object      Symbol
> >>>>>  # ........  .......  .................  .............................................................................
> >>>>>  #
> >>>>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
> >>>>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
> >>>>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
> >>>>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
> >>>>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
> >>>>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
> >>>>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
> >>>>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
> >>>>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
> >>>>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
> >>>>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
> >>>>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
> >>>>>
> >>>>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> >>>>> swapping in/out a page mapped by only one process. If the
> >>>>> page is mapped by multiple processes, typically, like more
> >>>>> than 100 on a phone, the overhead would be much higher as
> >>>>> we have to run tlb flush 100 times for one single page.
> >>>>> Plus, tlb flush overhead will increase with the number
> >>>>> of CPU cores due to the bad scalability of tlb shootdown
> >>>>> in HW, so those ARM64 servers should expect much higher
> >>>>> overhead.
> >>>>>
> >>>>> Further perf annonate shows 95% cpu time of ptep_clear_flush
> >>>>> is actually used by the final dsb() to wait for the completion
> >>>>> of tlb flush. This provides us a very good chance to leverage
> >>>>> the existing batched tlb in kernel. The minimum modification
> >>>>> is that we only send async tlbi in the first stage and we send
> >>>>> dsb while we have to sync in the second stage.
> >>>>>
> >>>>> With the above simplest micro benchmark, collapsed time to
> >>>>> finish the program decreases around 5%.
> >>>>>
> >>>>> Typical collapsed time w/o patch:
> >>>>>  ~ # time taskset -c 4 ./a.out
> >>>>>  0.21user 14.34system 0:14.69elapsed
> >>>>> w/ patch:
> >>>>>  ~ # time taskset -c 4 ./a.out
> >>>>>  0.22user 13.45system 0:13.80elapsed
> >>>>>
> >>>>> Also, Yicong Yang added the following observation.
> >>>>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
> >>>>>       observed an improvement around 12.5% with command
> >>>>>       `time ./swap_bench`.
> >>>>>               w/o             w/
> >>>>>       real    0m13.460s       0m11.771s
> >>>>>       user    0m0.248s        0m0.279s
> >>>>>       sys     0m12.039s       0m11.458s
> >>>>>
> >>>>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> >>>>>       which has been eliminated by this patch:
> >>>>>
> >>>>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
> >>>>>       [...]
> >>>>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
> >>>>>
> >>>>> Cc: Jonathan Corbet <corbet@lwn.net>
> >>>>> Cc: Nadav Amit <namit@vmware.com>
> >>>>> Cc: Mel Gorman <mgorman@suse.de>
> >>>>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> >>>>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> >>>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> >>>>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> >>>>> ---
> >>>>>  .../features/vm/TLB/arch-support.txt          |  2 +-
> >>>>>  arch/arm64/Kconfig                            |  1 +
> >>>>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
> >>>>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
> >>>>>  4 files changed, 40 insertions(+), 3 deletions(-)
> >>>>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> >>>>>
> >>>>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> >>>>> index 1c009312b9c1..2caf815d7c6c 100644
> >>>>> --- a/Documentation/features/vm/TLB/arch-support.txt
> >>>>> +++ b/Documentation/features/vm/TLB/arch-support.txt
> >>>>> @@ -9,7 +9,7 @@
> >>>>>      |       alpha: | TODO |
> >>>>>      |         arc: | TODO |
> >>>>>      |         arm: | TODO |
> >>>>> -    |       arm64: | TODO |
> >>>>> +    |       arm64: |  ok  |
> >>>>>      |        csky: | TODO |
> >>>>>      |     hexagon: | TODO |
> >>>>>      |        ia64: | TODO |
> >>>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> >>>>> index 571cc234d0b3..09d45cd6d665 100644
> >>>>> --- a/arch/arm64/Kconfig
> >>>>> +++ b/arch/arm64/Kconfig
> >>>>> @@ -93,6 +93,7 @@ config ARM64
> >>>>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
> >>>>>       select ARCH_SUPPORTS_NUMA_BALANCING
> >>>>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> >>>>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> >>>>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
> >>>>>       select ARCH_WANT_DEFAULT_BPF_JIT
> >>>>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> >>>>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> >>>>> new file mode 100644
> >>>>> index 000000000000..fedb0b87b8db
> >>>>> --- /dev/null
> >>>>> +++ b/arch/arm64/include/asm/tlbbatch.h
> >>>>> @@ -0,0 +1,12 @@
> >>>>> +/* SPDX-License-Identifier: GPL-2.0 */
> >>>>> +#ifndef _ARCH_ARM64_TLBBATCH_H
> >>>>> +#define _ARCH_ARM64_TLBBATCH_H
> >>>>> +
> >>>>> +struct arch_tlbflush_unmap_batch {
> >>>>> +     /*
> >>>>> +      * For arm64, HW can do tlb shootdown, so we don't
> >>>>> +      * need to record cpumask for sending IPI
> >>>>> +      */
> >>>>> +};
> >>>>> +
> >>>>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> >>>>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> >>>>> index 412a3b9a3c25..23cbc987321a 100644
> >>>>> --- a/arch/arm64/include/asm/tlbflush.h
> >>>>> +++ b/arch/arm64/include/asm/tlbflush.h
> >>>>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> >>>>>       dsb(ish);
> >>>>>  }
> >>>>>
> >>>>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>>>> +
> >>>>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
> >>>>>                                        unsigned long uaddr)
> >>>>>  {
> >>>>>       unsigned long addr;
> >>>>>
> >>>>>       dsb(ishst);
> >>>>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> >>>>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
> >>>>>       __tlbi(vale1is, addr);
> >>>>>       __tlbi_user(vale1is, addr);
> >>>>>  }
> >>>>>
> >>>>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>>>> +                                      unsigned long uaddr)
> >>>>> +{
> >>>>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> >>>>> +}
> >>>>> +
> >>>>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>>>>                                 unsigned long uaddr)
> >>>>>  {
> >>>>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>>>>       dsb(ish);
> >>>>>  }
> >>>>>
> >>>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> >>>>> +{
> >>>>> +     return true;
> >>>>> +}
> >>>>
> >>>> Always defer and batch up TLB flush, unconditionally ?
> >>>
> >>> My understanding is we actually don't need tlbbatch for a machine with one
> >>> or two cores as the tlb flush is not expensive. even for a system with four
> >>> cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> >>> when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> >>> 100 crores, the cost is incredibly huge.
> >>
> >> Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
> >> one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
> >> expensive as well, while queuing up individual TLB flushes ?
> >
> > This one is much much cheaper as it is not waiting for the
> > completion of tlbi. waiting for the completion of tlbi is a big
> > deal in arm64, thus, similar optimization can be seen here
> >
> > 3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
> > when page aging").
> >
> >
> >>
> >> The very idea behind TLB deferral is the opportunity it (might) provide
> >> to accumulate address ranges and cpu masks so that individual TLB flush
> >> can be replaced with a more cost effective range based TLB flush. Hence
> >> I guess unless address range or cpumask based cost effective TLB flush
> >> is available, deferral does not improve the unmap performance as much.
> >
> >
> > After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
> > from all cpus in the system, tlbi is not scalable. The point here is that we
> > avoid waiting for each individual TLBi. Alternatively, they are batched. If
> > you read the benchmark in the commit log, you can find the great decline
> > in the cost to swap out a page.
>
> Alright, although collecting and deferring 'dsb(ish)' to the very end, does
> not feel like a direct fit case for ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH but I
> guess it can be used to improve unmap performance on arm64.
>
> But is this 'dsb(ish)' deferral architecturally valid ?

yes as dsb(ish) ensures the completion of tlbi.
https://developer.arm.com/documentation/den0024/a/Memory-Ordering/Barriers

We are even depending on the dsb(ish) during context switch in commit:
3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
when page aging").

Before the context switch, lots of tlbi could have been sent.

>
> Let's examine single page unmap path via try_to_unmap_one().
>
> should_defer_flush() {
>         ptep_get_and_clear()
>         set_tlb_ubc_flush_pending()
>                 arch_tlbbatch_add_mm()
>                         __flush_tlb_page_nosync()
> } else {
>         ptep_clear_flush()
>                 ptep_get_and_clear()
>                 flush_tlb_page()
>                         flush_tlb_page_nosync()
>                                 __flush_tlb_page_nosync()
>                         dsb(ish)
> }
>
> __flush_tlb_page_nosync()
> {
>         dsb(ishst);
>         addr = __TLBI_VADDR(uaddr, ASID(mm));
>         __tlbi(vale1is, addr);
>         __tlbi_user(vale1is, addr);
> }
>
> Currently without TLB deferral, 'dsb(ish)' gets executed just after __tlbi()
> and __tlbi_user(), because __flush_tlb_page_nosync() is an inline function.
>
> #define __TLBI_0(op, arg) asm (ARM64_ASM_PREAMBLE                              \
>                                "tlbi " #op "\n"                                \
>                    ALTERNATIVE("nop\n                   nop",                  \
>                                "dsb ish\n               tlbi " #op,            \
>                                ARM64_WORKAROUND_REPEAT_TLBI,                   \
>                                CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
>                             : : )
>
> #define __TLBI_1(op, arg) asm (ARM64_ASM_PREAMBLE                              \
>                                "tlbi " #op ", %0\n"                            \
>                    ALTERNATIVE("nop\n                   nop",                  \
>                                "dsb ish\n               tlbi " #op ", %0",     \
>                                ARM64_WORKAROUND_REPEAT_TLBI,                   \
>                                CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
>                             : : "r" (arg))
>
> #define __TLBI_N(op, arg, n, ...) __TLBI_##n(op, arg)
>
> #define __tlbi(op, ...)         __TLBI_N(op, ##__VA_ARGS__, 1, 0)
>
> #define __tlbi_user(op, arg) do {                                               \
>         if (arm64_kernel_unmapped_at_el0())                                     \
>                 __tlbi(op, (arg) | USER_ASID_FLAG);                             \
> } while (0)
>
> There is already a 'dsb(ish)' in between two subsequent TLB operations in
> case ARM64_WORKAROUND_REPEAT_TLBI is detected on the system. Hence I guess
> deferral should not enabled on such systems ?
>
> But with deferral enabled, 'dsb(ish)' will be executed in arch_tlbbatch_flush()
> via try_to_unmap_flush[_dirty](). There might be random number of instructions
> in between __tlbi()/__tlbi_user() i.e 'tlbi' instructions and final 'dsb(ish)'.
> Just wondering, if such 'detached in time with other instructions in between'
> 'tlbi' and 'dsb(ish)', is architecturally valid ?

yes. I think so, arm64 even depends on the implicit dsb in context switch.

>
> There is a comment in 'struct tlbflush_unmap_batch'.
>
>         /*
>          * The arch code makes the following promise: generic code can modify a
>          * PTE, then call arch_tlbbatch_add_mm() (which internally provides all
>          * needed barriers), then call arch_tlbbatch_flush(), and the entries
>          * will be flushed on all CPUs by the time that arch_tlbbatch_flush()
>          * returns.
>          */
>
>
> It expects arch_tlbbatch_add_mm() to provide all barriers, hence wondering if
> that would include just the first 'dsb(isht)' not the subsequent 'dsb(ish)' ?

yes. include/asm/tlbflush.h, we see the below comments:

 *      TLB Invalidation
 *      ================
 *
 *      This header file implements the low-level TLB invalidation routines
 *      (sometimes referred to as "flushing" in the kernel) for arm64.
 *
 *      Every invalidation operation uses the following template:
 *
 *      DSB ISHST       // Ensure prior page-table updates have completed
 *      TLBI ...        // Invalidate the TLB
 *      DSB ISH         // Ensure the TLB invalidation has completed
 *      if (invalidated kernel mappings)
 *              ISB     // Discard any instructions fetched from the old mapping
 *

Clearly dsb(ishst) has ensured page-table updates are visible to all CPUs.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Mon, Sep 19, 2022 at 4:53 PM Barry Song <21cnbao@gmail.com> wrote:
>
> On Mon, Sep 19, 2022 at 4:24 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
> >
> >
> >
> > On 9/15/22 12:12, Barry Song wrote:
> > > On Thu, Sep 15, 2022 at 6:07 PM Anshuman Khandual
> > > <anshuman.khandual@arm.com> wrote:
> > >>
> > >>
> > >>
> > >> On 9/9/22 11:05, Barry Song wrote:
> > >>> On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> > >>> <anshuman.khandual@arm.com> wrote:
> > >>>>
> > >>>>
> > >>>>
> > >>>> On 8/22/22 13:51, Yicong Yang wrote:
> > >>>>> From: Barry Song <v-songbaohua@oppo.com>
> > >>>>>
> > >>>>> on x86, batched and deferred tlb shootdown has lead to 90%
> > >>>>> performance increase on tlb shootdown. on arm64, HW can do
> > >>>>> tlb shootdown without software IPI. But sync tlbi is still
> > >>>>> quite expensive.
> > >>>>>
> > >>>>> Even running a simplest program which requires swapout can
> > >>>>> prove this is true,
> > >>>>>  #include <sys/types.h>
> > >>>>>  #include <unistd.h>
> > >>>>>  #include <sys/mman.h>
> > >>>>>  #include <string.h>
> > >>>>>
> > >>>>>  int main()
> > >>>>>  {
> > >>>>>  #define SIZE (1 * 1024 * 1024)
> > >>>>>          volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
> > >>>>>                                           MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> > >>>>>
> > >>>>>          memset(p, 0x88, SIZE);
> > >>>>>
> > >>>>>          for (int k = 0; k < 10000; k++) {
> > >>>>>                  /* swap in */
> > >>>>>                  for (int i = 0; i < SIZE; i += 4096) {
> > >>>>>                          (void)p[i];
> > >>>>>                  }
> > >>>>>
> > >>>>>                  /* swap out */
> > >>>>>                  madvise(p, SIZE, MADV_PAGEOUT);
> > >>>>>          }
> > >>>>>  }
> > >>>>>
> > >>>>> Perf result on snapdragon 888 with 8 cores by using zRAM
> > >>>>> as the swap block device.
> > >>>>>
> > >>>>>  ~ # perf record taskset -c 4 ./a.out
> > >>>>>  [ perf record: Woken up 10 times to write data ]
> > >>>>>  [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
> > >>>>>  ~ # perf report
> > >>>>>  # To display the perf.data header info, please use --header/--header-only options.
> > >>>>>  # To display the perf.data header info, please use --header/--header-only options.
> > >>>>>  #
> > >>>>>  #
> > >>>>>  # Total Lost Samples: 0
> > >>>>>  #
> > >>>>>  # Samples: 60K of event 'cycles'
> > >>>>>  # Event count (approx.): 35706225414
> > >>>>>  #
> > >>>>>  # Overhead  Command  Shared Object      Symbol
> > >>>>>  # ........  .......  .................  .............................................................................
> > >>>>>  #
> > >>>>>     21.07%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irq
> > >>>>>      8.23%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore
> > >>>>>      6.67%  a.out    [kernel.kallsyms]  [k] filemap_map_pages
> > >>>>>      6.16%  a.out    [kernel.kallsyms]  [k] __zram_bvec_write
> > >>>>>      5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
> > >>>>>      3.71%  a.out    [kernel.kallsyms]  [k] _raw_spin_lock
> > >>>>>      3.49%  a.out    [kernel.kallsyms]  [k] memset64
> > >>>>>      1.63%  a.out    [kernel.kallsyms]  [k] clear_page
> > >>>>>      1.42%  a.out    [kernel.kallsyms]  [k] _raw_spin_unlock
> > >>>>>      1.26%  a.out    [kernel.kallsyms]  [k] mod_zone_state.llvm.8525150236079521930
> > >>>>>      1.23%  a.out    [kernel.kallsyms]  [k] xas_load
> > >>>>>      1.15%  a.out    [kernel.kallsyms]  [k] zram_slot_lock
> > >>>>>
> > >>>>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> > >>>>> swapping in/out a page mapped by only one process. If the
> > >>>>> page is mapped by multiple processes, typically, like more
> > >>>>> than 100 on a phone, the overhead would be much higher as
> > >>>>> we have to run tlb flush 100 times for one single page.
> > >>>>> Plus, tlb flush overhead will increase with the number
> > >>>>> of CPU cores due to the bad scalability of tlb shootdown
> > >>>>> in HW, so those ARM64 servers should expect much higher
> > >>>>> overhead.
> > >>>>>
> > >>>>> Further perf annonate shows 95% cpu time of ptep_clear_flush
> > >>>>> is actually used by the final dsb() to wait for the completion
> > >>>>> of tlb flush. This provides us a very good chance to leverage
> > >>>>> the existing batched tlb in kernel. The minimum modification
> > >>>>> is that we only send async tlbi in the first stage and we send
> > >>>>> dsb while we have to sync in the second stage.
> > >>>>>
> > >>>>> With the above simplest micro benchmark, collapsed time to
> > >>>>> finish the program decreases around 5%.
> > >>>>>
> > >>>>> Typical collapsed time w/o patch:
> > >>>>>  ~ # time taskset -c 4 ./a.out
> > >>>>>  0.21user 14.34system 0:14.69elapsed
> > >>>>> w/ patch:
> > >>>>>  ~ # time taskset -c 4 ./a.out
> > >>>>>  0.22user 13.45system 0:13.80elapsed
> > >>>>>
> > >>>>> Also, Yicong Yang added the following observation.
> > >>>>>       Tested with benchmark in the commit on Kunpeng920 arm64 server,
> > >>>>>       observed an improvement around 12.5% with command
> > >>>>>       `time ./swap_bench`.
> > >>>>>               w/o             w/
> > >>>>>       real    0m13.460s       0m11.771s
> > >>>>>       user    0m0.248s        0m0.279s
> > >>>>>       sys     0m12.039s       0m11.458s
> > >>>>>
> > >>>>>       Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> > >>>>>       which has been eliminated by this patch:
> > >>>>>
> > >>>>>       [root@localhost yang]# perf record -- ./swap_bench && perf report
> > >>>>>       [...]
> > >>>>>       16.99%  swap_bench  [kernel.kallsyms]  [k] ptep_clear_flush
> > >>>>>
> > >>>>> Cc: Jonathan Corbet <corbet@lwn.net>
> > >>>>> Cc: Nadav Amit <namit@vmware.com>
> > >>>>> Cc: Mel Gorman <mgorman@suse.de>
> > >>>>> Tested-by: Yicong Yang <yangyicong@hisilicon.com>
> > >>>>> Tested-by: Xin Hao <xhao@linux.alibaba.com>
> > >>>>> Signed-off-by: Barry Song <v-songbaohua@oppo.com>
> > >>>>> Signed-off-by: Yicong Yang <yangyicong@hisilicon.com>
> > >>>>> ---
> > >>>>>  .../features/vm/TLB/arch-support.txt          |  2 +-
> > >>>>>  arch/arm64/Kconfig                            |  1 +
> > >>>>>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
> > >>>>>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
> > >>>>>  4 files changed, 40 insertions(+), 3 deletions(-)
> > >>>>>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> > >>>>>
> > >>>>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> > >>>>> index 1c009312b9c1..2caf815d7c6c 100644
> > >>>>> --- a/Documentation/features/vm/TLB/arch-support.txt
> > >>>>> +++ b/Documentation/features/vm/TLB/arch-support.txt
> > >>>>> @@ -9,7 +9,7 @@
> > >>>>>      |       alpha: | TODO |
> > >>>>>      |         arc: | TODO |
> > >>>>>      |         arm: | TODO |
> > >>>>> -    |       arm64: | TODO |
> > >>>>> +    |       arm64: |  ok  |
> > >>>>>      |        csky: | TODO |
> > >>>>>      |     hexagon: | TODO |
> > >>>>>      |        ia64: | TODO |
> > >>>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > >>>>> index 571cc234d0b3..09d45cd6d665 100644
> > >>>>> --- a/arch/arm64/Kconfig
> > >>>>> +++ b/arch/arm64/Kconfig
> > >>>>> @@ -93,6 +93,7 @@ config ARM64
> > >>>>>       select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
> > >>>>>       select ARCH_SUPPORTS_NUMA_BALANCING
> > >>>>>       select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> > >>>>> +     select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> > >>>>>       select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
> > >>>>>       select ARCH_WANT_DEFAULT_BPF_JIT
> > >>>>>       select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> > >>>>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> > >>>>> new file mode 100644
> > >>>>> index 000000000000..fedb0b87b8db
> > >>>>> --- /dev/null
> > >>>>> +++ b/arch/arm64/include/asm/tlbbatch.h
> > >>>>> @@ -0,0 +1,12 @@
> > >>>>> +/* SPDX-License-Identifier: GPL-2.0 */
> > >>>>> +#ifndef _ARCH_ARM64_TLBBATCH_H
> > >>>>> +#define _ARCH_ARM64_TLBBATCH_H
> > >>>>> +
> > >>>>> +struct arch_tlbflush_unmap_batch {
> > >>>>> +     /*
> > >>>>> +      * For arm64, HW can do tlb shootdown, so we don't
> > >>>>> +      * need to record cpumask for sending IPI
> > >>>>> +      */
> > >>>>> +};
> > >>>>> +
> > >>>>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> > >>>>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> > >>>>> index 412a3b9a3c25..23cbc987321a 100644
> > >>>>> --- a/arch/arm64/include/asm/tlbflush.h
> > >>>>> +++ b/arch/arm64/include/asm/tlbflush.h
> > >>>>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> > >>>>>       dsb(ish);
> > >>>>>  }
> > >>>>>
> > >>>>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> > >>>>> +
> > >>>>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
> > >>>>>                                        unsigned long uaddr)
> > >>>>>  {
> > >>>>>       unsigned long addr;
> > >>>>>
> > >>>>>       dsb(ishst);
> > >>>>> -     addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> > >>>>> +     addr = __TLBI_VADDR(uaddr, ASID(mm));
> > >>>>>       __tlbi(vale1is, addr);
> > >>>>>       __tlbi_user(vale1is, addr);
> > >>>>>  }
> > >>>>>
> > >>>>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> > >>>>> +                                      unsigned long uaddr)
> > >>>>> +{
> > >>>>> +     return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> > >>>>> +}
> > >>>>> +
> > >>>>>  static inline void flush_tlb_page(struct vm_area_struct *vma,
> > >>>>>                                 unsigned long uaddr)
> > >>>>>  {
> > >>>>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
> > >>>>>       dsb(ish);
> > >>>>>  }
> > >>>>>
> > >>>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> > >>>>> +{
> > >>>>> +     return true;
> > >>>>> +}
> > >>>>
> > >>>> Always defer and batch up TLB flush, unconditionally ?
> > >>>
> > >>> My understanding is we actually don't need tlbbatch for a machine with one
> > >>> or two cores as the tlb flush is not expensive. even for a system with four
> > >>> cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> > >>> when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> > >>> 100 crores, the cost is incredibly huge.
> > >>
> > >> Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
> > >> one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
> > >> expensive as well, while queuing up individual TLB flushes ?
> > >
> > > This one is much much cheaper as it is not waiting for the
> > > completion of tlbi. waiting for the completion of tlbi is a big
> > > deal in arm64, thus, similar optimization can be seen here
> > >
> > > 3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
> > > when page aging").
> > >
> > >
> > >>
> > >> The very idea behind TLB deferral is the opportunity it (might) provide
> > >> to accumulate address ranges and cpu masks so that individual TLB flush
> > >> can be replaced with a more cost effective range based TLB flush. Hence
> > >> I guess unless address range or cpumask based cost effective TLB flush
> > >> is available, deferral does not improve the unmap performance as much.
> > >
> > >
> > > After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
> > > from all cpus in the system, tlbi is not scalable. The point here is that we
> > > avoid waiting for each individual TLBi. Alternatively, they are batched. If
> > > you read the benchmark in the commit log, you can find the great decline
> > > in the cost to swap out a page.
> >
> > Alright, although collecting and deferring 'dsb(ish)' to the very end, does
> > not feel like a direct fit case for ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH but I
> > guess it can be used to improve unmap performance on arm64.
> >
> > But is this 'dsb(ish)' deferral architecturally valid ?
>
> yes as dsb(ish) ensures the completion of tlbi.
> https://developer.arm.com/documentation/den0024/a/Memory-Ordering/Barriers
>
> We are even depending on the dsb(ish) during context switch in commit:
> 3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
> when page aging").
>
> Before the context switch, lots of tlbi could have been sent.
>
> >
> > Let's examine single page unmap path via try_to_unmap_one().
> >
> > should_defer_flush() {
> >         ptep_get_and_clear()
> >         set_tlb_ubc_flush_pending()
> >                 arch_tlbbatch_add_mm()
> >                         __flush_tlb_page_nosync()
> > } else {
> >         ptep_clear_flush()
> >                 ptep_get_and_clear()
> >                 flush_tlb_page()
> >                         flush_tlb_page_nosync()
> >                                 __flush_tlb_page_nosync()
> >                         dsb(ish)
> > }
> >
> > __flush_tlb_page_nosync()
> > {
> >         dsb(ishst);
> >         addr = __TLBI_VADDR(uaddr, ASID(mm));
> >         __tlbi(vale1is, addr);
> >         __tlbi_user(vale1is, addr);
> > }
> >
> > Currently without TLB deferral, 'dsb(ish)' gets executed just after __tlbi()
> > and __tlbi_user(), because __flush_tlb_page_nosync() is an inline function.
> >
> > #define __TLBI_0(op, arg) asm (ARM64_ASM_PREAMBLE                              \
> >                                "tlbi " #op "\n"                                \
> >                    ALTERNATIVE("nop\n                   nop",                  \
> >                                "dsb ish\n               tlbi " #op,            \
> >                                ARM64_WORKAROUND_REPEAT_TLBI,                   \
> >                                CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
> >                             : : )
> >
> > #define __TLBI_1(op, arg) asm (ARM64_ASM_PREAMBLE                              \
> >                                "tlbi " #op ", %0\n"                            \
> >                    ALTERNATIVE("nop\n                   nop",                  \
> >                                "dsb ish\n               tlbi " #op ", %0",     \
> >                                ARM64_WORKAROUND_REPEAT_TLBI,                   \
> >                                CONFIG_ARM64_WORKAROUND_REPEAT_TLBI)            \
> >                             : : "r" (arg))
> >
> > #define __TLBI_N(op, arg, n, ...) __TLBI_##n(op, arg)
> >
> > #define __tlbi(op, ...)         __TLBI_N(op, ##__VA_ARGS__, 1, 0)
> >
> > #define __tlbi_user(op, arg) do {                                               \
> >         if (arm64_kernel_unmapped_at_el0())                                     \
> >                 __tlbi(op, (arg) | USER_ASID_FLAG);                             \
> > } while (0)
> >
> > There is already a 'dsb(ish)' in between two subsequent TLB operations in
> > case ARM64_WORKAROUND_REPEAT_TLBI is detected on the system. Hence I guess
> > deferral should not enabled on such systems ?
> >
> > But with deferral enabled, 'dsb(ish)' will be executed in arch_tlbbatch_flush()
> > via try_to_unmap_flush[_dirty](). There might be random number of instructions
> > in between __tlbi()/__tlbi_user() i.e 'tlbi' instructions and final 'dsb(ish)'.
> > Just wondering, if such 'detached in time with other instructions in between'
> > 'tlbi' and 'dsb(ish)', is architecturally valid ?
>
> yes. I think so, arm64 even depends on the implicit dsb in context switch.

Please note we are not leveraging the time windows between tlbi and
dsb(ish) to improve performance.
we are actually shrinking the number of dsb(isb). In memory
reclamation, we usually unmap 32 or more
pages, then call try_to_unmap_flush(). that is why we are batching dsb(isb).
so we are reducing 31 or more dsb(ish) for each memory reclamation,
that is the point.


Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> +{
> +	return true;
> +}

This needs to be conditional on systems, where there will be performance
improvements, and should not just be enabled all the time on all systems.
num_online_cpus() > X, which does not hold any cpu hotplug lock would be
a good metric ?

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Tue, Sep 20, 2022 at 3:00 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
> On 8/22/22 13:51, Yicong Yang wrote:
> > +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> > +{
> > +     return true;
> > +}
>
> This needs to be conditional on systems, where there will be performance
> improvements, and should not just be enabled all the time on all systems.
> num_online_cpus() > X, which does not hold any cpu hotplug lock would be
> a good metric ?

for a small system, i don't see how this patch will help, e.g. cpus <= 4;
so we can actually disable tlb-batch on small systems.
just need to check if we will have any race condition since hotplug will
make the condition true and false dynamically.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/20/22 09:09, Barry Song wrote:
> On Tue, Sep 20, 2022 at 3:00 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
>>
>>
>> On 8/22/22 13:51, Yicong Yang wrote:
>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>>> +{
>>> +     return true;
>>> +}
>>
>> This needs to be conditional on systems, where there will be performance
>> improvements, and should not just be enabled all the time on all systems.
>> num_online_cpus() > X, which does not hold any cpu hotplug lock would be
>> a good metric ?
> 
> for a small system, i don't see how this patch will help, e.g. cpus <= 4;
> so we can actually disable tlb-batch on small systems.

Do not subscribe ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH based on NR_CPUS ?
That might not help much as the default value is 256 for NR_CPUS.

OR

arch_tlbbatch_should_defer() checks on

1. online cpus			(dont enable batched TLB if <= X)
2. ARM64_WORKAROUND_REPEAT_TLBI (dont enable batched TLB)

> just need to check if we will have any race condition since hotplug will
> make the condition true and false dynamically.

If should_defer_flush() evaluate to be false, then ptep_clear_flush()
clears and flushes the entry right away. This should not race with other
queued up TLBI requests, which will be flushed separately. Wondering how
there can be a race here !

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Tue, Sep 20, 2022 at 8:45 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
>
> On 9/20/22 09:09, Barry Song wrote:
> > On Tue, Sep 20, 2022 at 3:00 PM Anshuman Khandual
> > <anshuman.khandual@arm.com> wrote:
> >>
> >>
> >> On 8/22/22 13:51, Yicong Yang wrote:
> >>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> >>> +{
> >>> +     return true;
> >>> +}
> >>
> >> This needs to be conditional on systems, where there will be performance
> >> improvements, and should not just be enabled all the time on all systems.
> >> num_online_cpus() > X, which does not hold any cpu hotplug lock would be
> >> a good metric ?
> >
> > for a small system, i don't see how this patch will help, e.g. cpus <= 4;
> > so we can actually disable tlb-batch on small systems.
>
> Do not subscribe ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH based on NR_CPUS ?
> That might not help much as the default value is 256 for NR_CPUS.
>
> OR
>
> arch_tlbbatch_should_defer() checks on
>
> 1. online cpus                  (dont enable batched TLB if <= X)
> 2. ARM64_WORKAROUND_REPEAT_TLBI (dont enable batched TLB)
>
> > just need to check if we will have any race condition since hotplug will
> > make the condition true and false dynamically.
>
> If should_defer_flush() evaluate to be false, then ptep_clear_flush()
> clears and flushes the entry right away. This should not race with other
> queued up TLBI requests, which will be flushed separately. Wondering how
> there can be a race here !

Right. How about we make something as below?

static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
{
    /* for a small system very small number of CPUs, TLB shootdown is cheap */
    if (num_online_cpus() <= 4 ||
unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
          return false;

#ifdef CONFIG_ARM64_WORKAROUND_REPEAT_TLBI
    if (unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
         return false;
#endif

    return true;
}

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Wed, Sep 21, 2022 at 1:50 PM Barry Song <21cnbao@gmail.com> wrote:
>
> On Tue, Sep 20, 2022 at 8:45 PM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
> >
> >
> >
> > On 9/20/22 09:09, Barry Song wrote:
> > > On Tue, Sep 20, 2022 at 3:00 PM Anshuman Khandual
> > > <anshuman.khandual@arm.com> wrote:
> > >>
> > >>
> > >> On 8/22/22 13:51, Yicong Yang wrote:
> > >>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> > >>> +{
> > >>> +     return true;
> > >>> +}
> > >>
> > >> This needs to be conditional on systems, where there will be performance
> > >> improvements, and should not just be enabled all the time on all systems.
> > >> num_online_cpus() > X, which does not hold any cpu hotplug lock would be
> > >> a good metric ?
> > >
> > > for a small system, i don't see how this patch will help, e.g. cpus <= 4;
> > > so we can actually disable tlb-batch on small systems.
> >
> > Do not subscribe ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH based on NR_CPUS ?
> > That might not help much as the default value is 256 for NR_CPUS.
> >
> > OR
> >
> > arch_tlbbatch_should_defer() checks on
> >
> > 1. online cpus                  (dont enable batched TLB if <= X)
> > 2. ARM64_WORKAROUND_REPEAT_TLBI (dont enable batched TLB)
> >
> > > just need to check if we will have any race condition since hotplug will
> > > make the condition true and false dynamically.
> >
> > If should_defer_flush() evaluate to be false, then ptep_clear_flush()
> > clears and flushes the entry right away. This should not race with other
> > queued up TLBI requests, which will be flushed separately. Wondering how
> > there can be a race here !
>
> Right. How about we make something as below?
>
> static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> {
>     /* for a small system very small number of CPUs, TLB shootdown is cheap */
>     if (num_online_cpus() <= 4 ||
> unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
>           return false;
>
> #ifdef CONFIG_ARM64_WORKAROUND_REPEAT_TLBI
>     if (unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
>          return false;
> #endif
>
>     return true;
> }

sorry, i mean

static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
{
    /* for a small system very small number of CPUs, TLB shootdown is cheap */
    if (num_online_cpus() <= 4)
            return false;

#ifdef CONFIG_ARM64_WORKAROUND_REPEAT_TLBI
    if (unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
            return false;
#endif

    return true;
}

>
> Thanks
> Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/21/22 07:21, Barry Song wrote:
> On Wed, Sep 21, 2022 at 1:50 PM Barry Song <21cnbao@gmail.com> wrote:
>>
>> On Tue, Sep 20, 2022 at 8:45 PM Anshuman Khandual
>> <anshuman.khandual@arm.com> wrote:
>>>
>>>
>>>
>>> On 9/20/22 09:09, Barry Song wrote:
>>>> On Tue, Sep 20, 2022 at 3:00 PM Anshuman Khandual
>>>> <anshuman.khandual@arm.com> wrote:
>>>>>
>>>>>
>>>>> On 8/22/22 13:51, Yicong Yang wrote:
>>>>>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>>>>>> +{
>>>>>> +     return true;
>>>>>> +}
>>>>>
>>>>> This needs to be conditional on systems, where there will be performance
>>>>> improvements, and should not just be enabled all the time on all systems.
>>>>> num_online_cpus() > X, which does not hold any cpu hotplug lock would be
>>>>> a good metric ?
>>>>
>>>> for a small system, i don't see how this patch will help, e.g. cpus <= 4;
>>>> so we can actually disable tlb-batch on small systems.
>>>
>>> Do not subscribe ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH based on NR_CPUS ?
>>> That might not help much as the default value is 256 for NR_CPUS.
>>>
>>> OR
>>>
>>> arch_tlbbatch_should_defer() checks on
>>>
>>> 1. online cpus                  (dont enable batched TLB if <= X)
>>> 2. ARM64_WORKAROUND_REPEAT_TLBI (dont enable batched TLB)
>>>
>>>> just need to check if we will have any race condition since hotplug will
>>>> make the condition true and false dynamically.
>>>
>>> If should_defer_flush() evaluate to be false, then ptep_clear_flush()
>>> clears and flushes the entry right away. This should not race with other
>>> queued up TLBI requests, which will be flushed separately. Wondering how
>>> there can be a race here !
>>
>> Right. How about we make something as below?
>>
>> static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
>> {
>>     /* for a small system very small number of CPUs, TLB shootdown is cheap */
>>     if (num_online_cpus() <= 4 ||
>> unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
>>           return false;
>>
>> #ifdef CONFIG_ARM64_WORKAROUND_REPEAT_TLBI
>>     if (unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
>>          return false;
>> #endif
>>
>>     return true;
>> }
> 
> sorry, i mean
> 
> static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> {
>     /* for a small system very small number of CPUs, TLB shootdown is cheap */
>     if (num_online_cpus() <= 4)
>             return false;
> 
> #ifdef CONFIG_ARM64_WORKAROUND_REPEAT_TLBI
>     if (unlikely(this_cpu_has_cap(ARM64_WORKAROUND_REPEAT_TLBI)))
>             return false;
> #endif
> 
>     return true;
> }

This is a good starting point.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 8/22/22 13:51, Yicong Yang wrote:
> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> +					struct mm_struct *mm,
> +					unsigned long uaddr)
> +{
> +	__flush_tlb_page_nosync(mm, uaddr);
> +}
> +
> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> +{
> +	dsb(ish);
> +}

Just wondering if arch_tlbbatch_add_mm() could also detect continuous mapping
TLB invalidation requests on a given mm and try to generate a range based TLB
invalidation such as flush_tlb_range().

struct arch_tlbflush_unmap_batch via task->tlb_ubc->arch can track continuous
ranges while being queued up via arch_tlbbatch_add_mm(), any range formed can
later be flushed in subsequent arch_tlbbatch_flush() ?

OR

It might not be worth the effort and complexity, in comparison to performance
improvement, TLB range flush brings in ?

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Barry Song]

On Wed, Sep 21, 2022 at 6:53 PM Anshuman Khandual
<anshuman.khandual@arm.com> wrote:
>
>
> On 8/22/22 13:51, Yicong Yang wrote:
> > +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
> > +                                     struct mm_struct *mm,
> > +                                     unsigned long uaddr)
> > +{
> > +     __flush_tlb_page_nosync(mm, uaddr);
> > +}
> > +
> > +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
> > +{
> > +     dsb(ish);
> > +}
>
> Just wondering if arch_tlbbatch_add_mm() could also detect continuous mapping
> TLB invalidation requests on a given mm and try to generate a range based TLB
> invalidation such as flush_tlb_range().
>
> struct arch_tlbflush_unmap_batch via task->tlb_ubc->arch can track continuous
> ranges while being queued up via arch_tlbbatch_add_mm(), any range formed can
> later be flushed in subsequent arch_tlbbatch_flush() ?
>
> OR
>
> It might not be worth the effort and complexity, in comparison to performance
> improvement, TLB range flush brings in ?

Probably it is not worth the complexity as perf annotate shows
"
Further perf annonate shows 95% cpu time of ptep_clear_flush
is actually used by the final dsb() to wait for the completion
of tlb flush."

so any further optimization before dsb(ish) might bring some improvement
but seems minor.

Thanks
Barry

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Nadav Amit]

On Sep 20, 2022, at 11:53 PM, Anshuman Khandual <anshuman.khandual@arm.com> wrote:

> ⚠ External Email
> 
> On 8/22/22 13:51, Yicong Yang wrote:
>> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
>> +                                     struct mm_struct *mm,
>> +                                     unsigned long uaddr)
>> +{
>> +     __flush_tlb_page_nosync(mm, uaddr);
>> +}
>> +
>> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
>> +{
>> +     dsb(ish);
>> +}
> 
> Just wondering if arch_tlbbatch_add_mm() could also detect continuous mapping
> TLB invalidation requests on a given mm and try to generate a range based TLB
> invalidation such as flush_tlb_range().
> 
> struct arch_tlbflush_unmap_batch via task->tlb_ubc->arch can track continuous
> ranges while being queued up via arch_tlbbatch_add_mm(), any range formed can
> later be flushed in subsequent arch_tlbbatch_flush() ?
> 
> OR
> 
> It might not be worth the effort and complexity, in comparison to performance
> improvement, TLB range flush brings in ?

So here are my 2 cents, based on my experience with Intel-x86. It is likely
different on arm64, but perhaps it can provide you some insight into what
parameters you should measure and consider.

In general there is a tradeoff between full TLB flushes and entry-specific
ones. Flushing specific entries takes more time than flushing the entire
TLB, but sade TLB refills.

Dave Hansen made some calculations in the past and came up with 33 as a
magic cutoff number, i.e., if you need to flush more than 33 entries, just
flush the entire TLB. I am not sure that this exact number is very
meaningful, since one might argue that it should’ve taken PTI into account
(which might require twice as many TLB invalidations).

Anyhow, back to arch_tlbbatch_add_mm(). It may be possible to track ranges,
but the question is whether you would actually succeed in forming continuous
ranges that are eventually (on x86) smaller than the full TLB flush cutoff
(=33). Questionable (perhaps better with MGLRU?).

Then, you should remember that tracking should be very efficient, since even
few cache misses might have greater cost than what you save by
selective-flushing. Finally, on x86 you would need to invoke the smp/IPI
layer multiple times to send different cores the relevant range they need to
flush.

IOW: It is somewhat complicated to implement efficeintly. On x86, and
probably other IPI-based TLB shootdown systems, does not have clear
performance benefit (IMHO).

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 4/4] arm64: support batched/deferred tlb shootdown during page reclamation

[Anshuman Khandual]

On 9/21/22 12:47, Nadav Amit wrote:
> On Sep 20, 2022, at 11:53 PM, Anshuman Khandual <anshuman.khandual@arm.com> wrote:
> 
>> ⚠ External Email
>>
>> On 8/22/22 13:51, Yicong Yang wrote:
>>> +static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
>>> +                                     struct mm_struct *mm,
>>> +                                     unsigned long uaddr)
>>> +{
>>> +     __flush_tlb_page_nosync(mm, uaddr);
>>> +}
>>> +
>>> +static inline void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
>>> +{
>>> +     dsb(ish);
>>> +}
>>
>> Just wondering if arch_tlbbatch_add_mm() could also detect continuous mapping
>> TLB invalidation requests on a given mm and try to generate a range based TLB
>> invalidation such as flush_tlb_range().
>>
>> struct arch_tlbflush_unmap_batch via task->tlb_ubc->arch can track continuous
>> ranges while being queued up via arch_tlbbatch_add_mm(), any range formed can
>> later be flushed in subsequent arch_tlbbatch_flush() ?
>>
>> OR
>>
>> It might not be worth the effort and complexity, in comparison to performance
>> improvement, TLB range flush brings in ?
> 
> So here are my 2 cents, based on my experience with Intel-x86. It is likely
> different on arm64, but perhaps it can provide you some insight into what
> parameters you should measure and consider.
> 
> In general there is a tradeoff between full TLB flushes and entry-specific
> ones. Flushing specific entries takes more time than flushing the entire
> TLB, but sade TLB refills.

Right.

> 
> Dave Hansen made some calculations in the past and came up with 33 as a
> magic cutoff number, i.e., if you need to flush more than 33 entries, just
> flush the entire TLB. I am not sure that this exact number is very
> meaningful, since one might argue that it should’ve taken PTI into account
> (which might require twice as many TLB invalidations).

Okay.

> 
> Anyhow, back to arch_tlbbatch_add_mm(). It may be possible to track ranges,
> but the question is whether you would actually succeed in forming continuous
> ranges that are eventually (on x86) smaller than the full TLB flush cutoff
> (=33). Questionable (perhaps better with MGLRU?).

This proposal here for arm64 does not cause a full TLB flush ever. It creates
individual TLB flushes all the time. Hence the choice here is not between full
TLB flush and possible range flushes. Choice is actually between individual
TLB flushes and range/full TLB flushes.

> 
> Then, you should remember that tracking should be very efficient, since even
> few cache misses might have greater cost than what you save by
> selective-flushing. Finally, on x86 you would need to invoke the smp/IPI
> layer multiple times to send different cores the relevant range they need to
> flush.

Agreed, these reasons make it much difficult to gain any more performance.

> 
> IOW: It is somewhat complicated to implement efficeintly. On x86, and
> probably other IPI-based TLB shootdown systems, does not have clear
> performance benefit (IMHO).

Agreed, thanks for such a detailed explanation, appreciate it.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [PATCH v3 0/4] mm: arm64: bring up BATCHED_UNMAP_TLB_FLUSH

[Yicong Yang]

Hi mm and arm64 maintainers,

a gentle ping for this..

Thanks.

On 2022/8/22 16:21, Yicong Yang wrote:
> From: Yicong Yang <yangyicong@hisilicon.com>
> 
> Though ARM64 has the hardware to do tlb shootdown, the hardware
> broadcasting is not free.
> A simplest micro benchmark shows even on snapdragon 888 with only
> 8 cores, the overhead for ptep_clear_flush is huge even for paging
> out one page mapped by only one process:
> 5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
> 
> While pages are mapped by multiple processes or HW has more CPUs,
> the cost should become even higher due to the bad scalability of
> tlb shootdown.
> 
> The same benchmark can result in 16.99% CPU consumption on ARM64
> server with around 100 cores according to Yicong's test on patch
> 4/4.
> 
> This patchset leverages the existing BATCHED_UNMAP_TLB_FLUSH by
> 1. only send tlbi instructions in the first stage -
> 	arch_tlbbatch_add_mm()
> 2. wait for the completion of tlbi by dsb while doing tlbbatch
> 	sync in arch_tlbbatch_flush()
> My testing on snapdragon shows the overhead of ptep_clear_flush
> is removed by the patchset. The micro benchmark becomes 5% faster
> even for one page mapped by single process on snapdragon 888.
> 
> -v3:
> 1. Declare arch's tlbbatch defer support by arch_tlbbatch_should_defer() instead
>    of ARCH_HAS_MM_CPUMASK, per Barry and Kefeng
> 2. Add Tested-by from Xin Hao
> Link: https://lore.kernel.org/linux-mm/20220711034615.482895-1-21cnbao@gmail.com/
> 
> -v2:
> 1. Collected Yicong's test result on kunpeng920 ARM64 server;
> 2. Removed the redundant vma parameter in arch_tlbbatch_add_mm()
>    according to the comments of Peter Zijlstra and Dave Hansen
> 3. Added ARCH_HAS_MM_CPUMASK rather than checking if mm_cpumask
>    is empty according to the comments of Nadav Amit
> 
> Thanks, Peter, Dave and Nadav for your testing or reviewing
> , and comments.
> 
> -v1:
> https://lore.kernel.org/lkml/20220707125242.425242-1-21cnbao@gmail.com/
> 
> Anshuman Khandual (1):
>   mm/tlbbatch: Introduce arch_tlbbatch_should_defer()
> 
> Barry Song (3):
>   Revert "Documentation/features: mark BATCHED_UNMAP_TLB_FLUSH doesn't
>     apply to ARM64"
>   mm: rmap: Extend tlbbatch APIs to fit new platforms
>   arm64: support batched/deferred tlb shootdown during page reclamation
> 
>  Documentation/features/arch-support.txt       |  1 -
>  .../features/vm/TLB/arch-support.txt          |  2 +-
>  arch/arm64/Kconfig                            |  1 +
>  arch/arm64/include/asm/tlbbatch.h             | 12 ++++++++
>  arch/arm64/include/asm/tlbflush.h             | 28 +++++++++++++++++--
>  arch/x86/include/asm/tlbflush.h               | 15 +++++++++-
>  mm/rmap.c                                     | 19 +++++--------
>  7 files changed, 61 insertions(+), 17 deletions(-)
>  create mode 100644 arch/arm64/include/asm/tlbbatch.h
> 

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel