Re: [bug] aarch64: userspace stalls on page fault after dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")

[Yang Shi <yang.shi@linux.alibaba.com>]

On 5/7/19 12:55 PM, Yang Shi wrote:
>
>
> On 5/7/19 10:19 AM, Jan Stancek wrote:
>> ----- Original Message -----
>>>
>>> On 5/7/19 4:51 AM, Jan Stancek wrote:
>>>> ----- Original Message -----
>>>>> On 5/6/19 2:35 PM, Jan Stancek wrote:
>>>>>> ----- Original Message -----
>>>>>>> On 5/5/19 7:10 AM, Jan Stancek wrote:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> I'm seeing userspace program getting stuck on aarch64, on 
>>>>>>>> kernels 4.20
>>>>>>>> and
>>>>>>>> newer.
>>>>>>>> It stalls from seconds to hours.
>>>>>>>>
>>>>>>>> I have simplified it to following scenario (reproducer linked 
>>>>>>>> below
>>>>>>>> [1]):
>>>>>>>>       while (1):
>>>>>>>>         spawn Thread 1: mmap, write, munmap
>>>>>>>>         spawn Thread 2: <nothing>
>>>>>>>>
>>>>>>>> Thread 1 is sporadically getting stuck on write to mapped area.
>>>>>>>> User-space
>>>>>>>> is not
>>>>>>>> moving forward - stdout output stops. Observed CPU usage is 
>>>>>>>> however
>>>>>>>> 100%.
>>>>>>>>
>>>>>>>> At this time, kernel appears to be busy handling page faults 
>>>>>>>> (~700k per
>>>>>>>> second):
>>>>>>>>
>>>>>>>> # perf top -a -g
>>>>>>>> -   98.97%     8.30%  a.out                     [.] 
>>>>>>>> map_write_unmap
>>>>>>>>        - 23.52% map_write_unmap
>>>>>>>>           - 24.29% el0_sync
>>>>>>>>              - 10.42% do_mem_abort
>>>>>>>>                 - 17.81% do_translation_fault
>>>>>>>>                    - 33.01% do_page_fault
>>>>>>>>                       - 56.18% handle_mm_fault
>>>>>>>>                            40.26% __handle_mm_fault
>>>>>>>>                            2.19% __ll_sc___cmpxchg_case_acq_4
>>>>>>>>                            0.87% mem_cgroup_from_task
>>>>>>>>                       - 6.18% find_vma
>>>>>>>>                            5.38% vmacache_find
>>>>>>>>                         1.35% __ll_sc___cmpxchg_case_acq_8
>>>>>>>>                         1.23% __ll_sc_atomic64_sub_return_release
>>>>>>>>                         0.78% down_read_trylock
>>>>>>>>                0.93% do_translation_fault
>>>>>>>>        + 8.30% thread_start
>>>>>>>>
>>>>>>>> #  perf stat -p 8189 -d
>>>>>>>> ^C
>>>>>>>>      Performance counter stats for process id '8189':
>>>>>>>>
>>>>>>>>             984.311350      task-clock (msec)         #    
>>>>>>>> 1.000 CPUs
>>>>>>>>             utilized
>>>>>>>>                      0 context-switches          #    0.000 K/sec
>>>>>>>>                      0 cpu-migrations            #    0.000 K/sec
>>>>>>>>                723,641 page-faults               #    0.735 M/sec
>>>>>>>>          2,559,199,434 cycles                    #    2.600 GHz
>>>>>>>>            711,933,112 instructions              #    0.28  insn
>>>>>>>>            per
>>>>>>>>            cycle
>>>>>>>>        <not supported>      branches
>>>>>>>>                757,658      branch-misses
>>>>>>>>            205,840,557 L1-dcache-loads           #  209.121 M/sec
>>>>>>>>             40,561,529 L1-dcache-load-misses     #   19.71% of all
>>>>>>>>             L1-dcache hits
>>>>>>>>        <not supported>      LLC-loads
>>>>>>>>        <not supported>      LLC-load-misses
>>>>>>>>
>>>>>>>>            0.984454892 seconds time elapsed
>>>>>>>>
>>>>>>>> With some extra traces, it appears looping in page fault for same
>>>>>>>> address,
>>>>>>>> over and over:
>>>>>>>>       do_page_fault // mm_flags: 0x55
>>>>>>>>         __do_page_fault
>>>>>>>>           __handle_mm_fault
>>>>>>>>             handle_pte_fault
>>>>>>>>               ptep_set_access_flags
>>>>>>>>                 if (pte_same(pte, entry))  // pte: e8000805060f53,
>>>>>>>>                 entry:
>>>>>>>>                 e8000805060f53
>>>>>>>>
>>>>>>>> I had traces in mmap() and munmap() as well, they don't get hit 
>>>>>>>> when
>>>>>>>> reproducer
>>>>>>>> hits the bad state.
>>>>>>>>
>>>>>>>> Notes:
>>>>>>>> - I'm not able to reproduce this on x86.
>>>>>>>> - Attaching GDB or strace immediatelly recovers application 
>>>>>>>> from stall.
>>>>>>>> - It also seems to recover faster when system is busy with 
>>>>>>>> other tasks.
>>>>>>>> - MAP_SHARED vs. MAP_PRIVATE makes no difference.
>>>>>>>> - Turning off THP makes no difference.
>>>>>>>> - Reproducer [1] usually hits it within ~minute on HW described 
>>>>>>>> below.
>>>>>>>> - Longman mentioned that "When the rwsem becomes reader-owned, it
>>>>>>>> causes
>>>>>>>>       all the spinning writers to go to sleep adding wakeup 
>>>>>>>> latency to
>>>>>>>>       the time required to finish the critical sections", but 
>>>>>>>> this looks
>>>>>>>>       like busy loop, so I'm not sure if it's related to rwsem 
>>>>>>>> issues
>>>>>>>>       identified
>>>>>>>>       in:
>>>>>>>> https://lore.kernel.org/lkml/20190428212557.13482-2-longman@redhat.com/ 
>>>>>>>>
>>>>>>> It sounds possible to me. What the optimization done by the 
>>>>>>> commit ("mm:
>>>>>>> mmap: zap pages with read mmap_sem in munmap") is to downgrade 
>>>>>>> write
>>>>>>> rwsem to read when zapping pages and page table in munmap() 
>>>>>>> after the
>>>>>>> vmas have been detached from the rbtree.
>>>>>>>
>>>>>>> So the mmap(), which is writer, in your test may steal the lock and
>>>>>>> execute with the munmap(), which is the reader after the 
>>>>>>> downgrade, in
>>>>>>> parallel to break the mutual exclusion.
>>>>>>>
>>>>>>> In this case, the parallel mmap() may map to the same area since 
>>>>>>> vmas
>>>>>>> have been detached by munmap(), then mmap() may create the 
>>>>>>> complete same
>>>>>>> vmas, and page fault happens on the same vma at the same address.
>>>>>>>
>>>>>>> I'm not sure why gdb or strace could recover this, but they use 
>>>>>>> ptrace
>>>>>>> which may acquire mmap_sem to break the parallel inadvertently.
>>>>>>>
>>>>>>> May you please try Waiman's patch to see if it makes any 
>>>>>>> difference?
>>>>>> I don't see any difference in behaviour after applying:
>>>>>>      [PATCH-tip v7 01/20] locking/rwsem: Prevent decrement of 
>>>>>> reader count
>>>>>>      before increment
>>>>>> Issue is still easily reproducible for me.
>>>>>>
>>>>>> I'm including output of mem_abort_decode() / show_pte() for 
>>>>>> sample PTE,
>>>>>> that
>>>>>> I see in page fault loop. (I went through all bits, but couldn't 
>>>>>> find
>>>>>> anything invalid about it)
>>>>>>
>>>>>>      mem_abort_decode: Mem abort info:
>>>>>>      mem_abort_decode:   ESR = 0x92000047
>>>>>>      mem_abort_decode:   Exception class = DABT (lower EL), IL = 
>>>>>> 32 bits
>>>>>>      mem_abort_decode:   SET = 0, FnV = 0
>>>>>>      mem_abort_decode:   EA = 0, S1PTW = 0
>>>>>>      mem_abort_decode: Data abort info:
>>>>>>      mem_abort_decode:   ISV = 0, ISS = 0x00000047
>>>>>>      mem_abort_decode:   CM = 0, WnR = 1
>>>>>>      show_pte: user pgtable: 64k pages, 48-bit VAs, pgdp =
>>>>>>      0000000067027567
>>>>>>      show_pte: [0000ffff6dff0000] pgd=000000176bae0003
>>>>>>      show_pte: , pud=000000176bae0003
>>>>>>      show_pte: , pmd=000000174ad60003
>>>>>>      show_pte: , pte=00e80008023a0f53
>>>>>>      show_pte: , pte_pfn: 8023a
>>>>>>
>>>>>>      >>> print bin(0x47)
>>>>>>      0b1000111
>>>>>>
>>>>>>      Per D12-2779 (ARM Architecture Reference Manual),
>>>>>>          ISS encoding for an exception from an Instruction Abort:
>>>>>>        IFSC, bits [5:0], Instruction Fault Status Code
>>>>>>        0b000111 Translation fault, level 3
>>>>>>
>>>>>> ---
>>>>>>
>>>>>> My theory is that TLB is getting broken.
>>>> Theory continued:
>>>>
>>>> unmap_region() is batching updates to TLB (for vmas and page tables).
>>>> And at the same time another thread handles page fault for same mm,
>>>> which increases "tlb_flush_pending".
>>>>
>>>> tlb_finish_mmu() called from unmap_region() will thus set 'force = 1'.
>>>> And arch_tlb_finish_mmu() will in turn reset TLB range, presumably 
>>>> making
>>>> it smaller then it would be if force == 0.
>>>>
>>>> Change below appears to fix it:
>>>>
>>>> diff --git a/mm/mmu_gather.c b/mm/mmu_gather.c
>>>> index f2f03c655807..a4cef21bd62b 100644
>>>> --- a/mm/mmu_gather.c
>>>> +++ b/mm/mmu_gather.c
>>>> @@ -93,7 +93,7 @@ void arch_tlb_finish_mmu(struct mmu_gather *tlb,
>>>>           struct mmu_gather_batch *batch, *next;
>>>>              if (force) {
>>>> -               __tlb_reset_range(tlb);
>>>>                   __tlb_adjust_range(tlb, start, end - start);
>>> I don't get why the change could fix it?
>> My guess is that reset clears "tlb->freed_tables", which changes how
>> tlb_flush() operates, see "bool last_level = !tlb->freed_tables;" in
>> arch/arm64/include/asm/tlb.h. Maybe that doesn't clear some intermediate
>> entries? No clue.
>>
>> If I let it reset the range, but preserve "freed_tables", it also
>> seems to solve the problem:
>
> This makes sense. munmap() does free page tables, so "freed_tables" 
> should be 1 instead of 0. So, in this case, __tlb_reset_range() should 
> not be called.

Further investigation shows the __tlb_reset_range() call is added by 
commit a6d60245d6d9 ("asm-generic/tlb: Track which levels of the page 
tables have been cleared"), which is aimed to make TLB flush smarter for 
huge page.

The commit was merged by 4.20 too with my munmap optimization. So, it 
looks nobody noticed that munmap() could run into "force" TLB flush 
other than MADV_DONTNEED. And, it does remove page tables. So, I believe 
removing __tlb_reset_range() should be the right fix.

>
>>
>> diff --git a/mm/mmu_gather.c b/mm/mmu_gather.c
>> index f2f03c655807..17fb0d7edc03 100644
>> --- a/mm/mmu_gather.c
>> +++ b/mm/mmu_gather.c
>> @@ -93,8 +93,20 @@ void arch_tlb_finish_mmu(struct mmu_gather *tlb,
>>          struct mmu_gather_batch *batch, *next;
>>            if (force) {
>> -               __tlb_reset_range(tlb);
>> +               if (tlb->fullmm) {
>> +                       tlb->start = tlb->end = ~0;
>> +               } else {
>> +                       tlb->start = TASK_SIZE;
>> +                       tlb->end = 0;
>> +               }
>>                  __tlb_adjust_range(tlb, start, end - start);
>>
>>> __tlb_reset_range() just reset
>>> start and end to TASK_SIZE and 0, then __tlb_adjust_range() set proper
>>> start and end. I don't get why "force" flush smaller range?
>> I'm still trying to understand this part. It's actually not smaller, 
>> but it changes:
>>
>> unmap_region()
>>    # vm_start: ffff49bd0000 vm_end: ffff49be0000
>>    ...
>>    # tlb.start, tlb.end: 1000000000000 0
>>    free_pgtables()
>>    # tlb.start, tlb.end: ffff40000000 ffff40010000
>>    tlb_finish_mmu()
>>      arch_tlb_finish_mmu()
>>        # will see force == 1
>>        # resets tlb.start, tlb.end to: ffff49bd0000 ffff49be0000
>>
>>>>           }
>>>>
>>>>>> I made a dummy kernel module that exports debugfs file, which on 
>>>>>> read
>>>>>> triggers:
>>>>>>      flush_tlb_all();
>>>>>>
>>>>>> Any time reproducer stalls and I read debugfs file, it recovers
>>>>>> immediately and resumes printing to stdout.
>>>>> That commit doesn't change anything about TLB flush, just move 
>>>>> zapping
>>>>> pages under read mmap_sem as what MADV_DONTNEED does.
>>>>>
>>>>> I don't have aarch64 board to reproduce and debug it. And, I'm not
>>>>> familiar with aarch64 architecture either. But, some history told 
>>>>> me the
>>>>> parallel zapping page may run into stale TLB and defer a flush 
>>>>> meaning
>>>>> that this call may observe pte_none and fails to flush the TLB. But,
>>>>> this has been solved by commit 56236a59556c ("mm: refactor TLB 
>>>>> gathering
>>>>> API") and 99baac21e458 ("mm: fix MADV_[FREE|DONTNEED] TLB flush miss
>>>>> problem").
>>>>>
>>>>> For more detail, please refer to commit 4647706ebeee ("mm: always 
>>>>> flush
>>>>> VMA ranges affected by zap_page_range"). Copied Mel and Rik in this
>>>>> thread. Also added Will Deacon and Catalin Marinas, who are aarch64
>>>>> maintainers, in this loop
>>>> Thanks
>>>>
>>>>> But, your test (triggering TLB flush) does demonstrate TLB flush is
>>>>> *not* done properly at some point as expected for aarch64. Could you
>>>>> please give the below patch a try?
>>>> Your patch also fixes my reproducer.
>>> Thanks for testing it.
>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index ab650c2..ef41ad5 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -1336,8 +1336,10 @@ void unmap_vmas(struct mmu_gather *tlb,
>>>>>
>>>>>            mmu_notifier_range_init(&range, vma->vm_mm, start_addr,
>>>>>            end_addr);
>>>>> mmu_notifier_invalidate_range_start(&range);
>>>>> -       for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
>>>>> +       for ( ; vma && vma->vm_start < end_addr; vma = 
>>>>> vma->vm_next) {
>>>>>                    unmap_single_vma(tlb, vma, start_addr, 
>>>>> end_addr, NULL);
>>>>> +               flush_tlb_range(vma, start_addr, end_addr);
>>>>> +       }
>>>>>            mmu_notifier_invalidate_range_end(&range);
>>>>>     }
>>>>>
>>>>>>>> - I tried 2 different aarch64 systems so far: APM X-Gene CPU 
>>>>>>>> Potenza A3
>>>>>>>> and
>>>>>>>>       Qualcomm 65-LA-115-151.
>>>>>>>>       I can reproduce it on both with v5.1-rc7. It's easier to 
>>>>>>>> reproduce
>>>>>>>>       on latter one (for longer periods of time), which has 46 
>>>>>>>> CPUs.
>>>>>>>> - Sample output of reproducer on otherwise idle system:
>>>>>>>>       # ./a.out
>>>>>>>>       [00000314] map_write_unmap took: 26305 ms
>>>>>>>>       [00000867] map_write_unmap took: 13642 ms
>>>>>>>>       [00002200] map_write_unmap took: 44237 ms
>>>>>>>>       [00002851] map_write_unmap took: 992 ms
>>>>>>>>       [00004725] map_write_unmap took: 542 ms
>>>>>>>>       [00006443] map_write_unmap took: 5333 ms
>>>>>>>>       [00006593] map_write_unmap took: 21162 ms
>>>>>>>>       [00007435] map_write_unmap took: 16982 ms
>>>>>>>>       [00007488] map_write unmap took: 13 ms^C
>>>>>>>>
>>>>>>>> I ran a bisect, which identified following commit as first bad 
>>>>>>>> one:
>>>>>>>>       dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in 
>>>>>>>> munmap")
>>>>>>>>
>>>>>>>> I can also make the issue go away with following change:
>>>>>>>> diff --git a/mm/mmap.c b/mm/mmap.c
>>>>>>>> index 330f12c17fa1..13ce465740e2 100644
>>>>>>>> --- a/mm/mmap.c
>>>>>>>> +++ b/mm/mmap.c
>>>>>>>> @@ -2844,7 +2844,7 @@ EXPORT_SYMBOL(vm_munmap);
>>>>>>>>      SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
>>>>>>>>      {
>>>>>>>>             profile_munmap(addr);
>>>>>>>> -       return __vm_munmap(addr, len, true);
>>>>>>>> +       return __vm_munmap(addr, len, false);
>>>>>>>>      }
>>>>>>>>
>>>>>>>> # cat /proc/cpuinfo  | head
>>>>>>>> processor       : 0
>>>>>>>> BogoMIPS        : 40.00
>>>>>>>> Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32 cpuid
>>>>>>>> asimdrdm
>>>>>>>> CPU implementer : 0x51
>>>>>>>> CPU architecture: 8
>>>>>>>> CPU variant     : 0x0
>>>>>>>> CPU part        : 0xc00
>>>>>>>> CPU revision    : 1
>>>>>>>>
>>>>>>>> # numactl -H
>>>>>>>> available: 1 nodes (0)
>>>>>>>> node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 
>>>>>>>> 20 21 22
>>>>>>>> 23
>>>>>>>> 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
>>>>>>>> node 0 size: 97938 MB
>>>>>>>> node 0 free: 95732 MB
>>>>>>>> node distances:
>>>>>>>> node   0
>>>>>>>>       0:  10
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> Jan
>>>>>>>>
>>>>>>>> [1]
>>>>>>>> https://github.com/jstancek/reproducers/blob/master/kernel/page_fault_stall/mmap5.c 
>>>>>>>>
>>>>>>>> [2]
>>>>>>>> https://github.com/jstancek/reproducers/blob/master/kernel/page_fault_stall/config 
>>>>>>>>
>>>
>