Re: Race condition observed between page migration and page fault handling on arm64 machines

[David Hildenbrand <david@redhat.com>]

On 05.08.24 11:51, Dev Jain wrote:
> 
> On 8/1/24 19:18, David Hildenbrand wrote:
>> On 01.08.24 15:43, Will Deacon wrote:
>>> On Thu, Aug 01, 2024 at 03:26:57PM +0200, David Hildenbrand wrote:
>>>> On 01.08.24 15:13, David Hildenbrand wrote:
>>>>>>>> To dampen the tradeoff, we could do this in shmem_fault()
>>>>>>>> instead? But
>>>>>>>> then, this would mean that we do this in all
>>>>>>>>
>>>>>>>> kinds of vma->vm_ops->fault, only when we discover another
>>>>>>>> reference
>>>>>>>> count race condition :) Doing this in do_fault()
>>>>>>>>
>>>>>>>> should solve this once and for all. In fact, do_pte_missing()
>>>>>>>> may call
>>>>>>>> do_anonymous_page() or do_fault(), and I just
>>>>>>>>
>>>>>>>> noticed that the former already checks this using
>>>>>>>> vmf_pte_changed().
>>>>>>>
>>>>>>> What I am still missing is why this is (a) arm64 only; and (b) if
>>>>>>> this
>>>>>>> is something we should really worry about. There are other reasons
>>>>>>> (e.g., speculative references) why migration could temporarily fail,
>>>>>>> does it happen that often that it is really something we have to
>>>>>>> worry
>>>>>>> about?
>>>>>>
>>>>>>
>>>>>> (a) See discussion at [1]; I guess it passes on x86, which is quite
>>>>>> strange since the race is clearly arch-independent.
>>>>>
>>>>> Yes, I think this is what we have to understand. Is the race simply
>>>>> less
>>>>> likely to trigger on x86?
>>>>>
>>>>> I would assume that it would trigger on any arch.
>>>>>
>>>>> I just ran it on a x86 VM with 2 NUMA nodes and it also seems to
>>>>> work here.
>>>>>
>>>>> Is this maybe related to deferred flushing? Such that the other CPU
>>>>> will
>>>>> by accident just observe the !pte_none a little less likely?
>>>>>
>>>>> But arm64 also usually defers flushes, right? At least unless
>>>>> ARM64_WORKAROUND_REPEAT_TLBI is around. With that we never do deferred
>>>>> flushes.
>>>>
>>>> Bingo!
>>>>
>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>> index e51ed44f8b53..ce94b810586b 100644
>>>> --- a/mm/rmap.c
>>>> +++ b/mm/rmap.c
>>>> @@ -718,10 +718,7 @@ static void set_tlb_ubc_flush_pending(struct
>>>> mm_struct
>>>> *mm, pte_t pteval,
>>>>     */
>>>>    static bool should_defer_flush(struct mm_struct *mm, enum
>>>> ttu_flags flags)
>>>>    {
>>>> -       if (!(flags & TTU_BATCH_FLUSH))
>>>> -               return false;
>>>> -
>>>> -       return arch_tlbbatch_should_defer(mm);
>>>> +       return false;
>>>>    }
>>>>
>>>>
>>>> On x86:
>>>>
>>>> # ./migration
>>>> TAP version 13
>>>> 1..1
>>>> # Starting 1 tests from 1 test cases.
>>>> #  RUN           migration.shared_anon ...
>>>> Didn't migrate 1 pages
>>>> # migration.c:170:shared_anon:Expected migrate(ptr, self->n1,
>>>> self->n2) (-2)
>>>> == 0 (0)
>>>> # shared_anon: Test terminated by assertion
>>>> #          FAIL  migration.shared_anon
>>>> not ok 1 migration.shared_anon
>>>>
>>>>
>>>> It fails all of the time!
>>>
>>> Nice work! I suppose that makes sense as, with the eager TLB
>>> invalidation, the window between the other CPU faulting and the
>>> migration entry being written is fairly wide.
>>>
>>> Not sure about a fix though :/ It feels a bit overkill to add a new
>>> invalid pte encoding just for this.
>>
>> Something like that might make the test happy in most cases:
>>
>> diff --git a/tools/testing/selftests/mm/migration.c
>> b/tools/testing/selftests/mm/migration.c
>> index 6908569ef406..4c18bfc13b94 100644
>> --- a/tools/testing/selftests/mm/migration.c
>> +++ b/tools/testing/selftests/mm/migration.c
>> @@ -65,6 +65,7 @@ int migrate(uint64_t *ptr, int n1, int n2)
>>          int ret, tmp;
>>          int status = 0;
>>          struct timespec ts1, ts2;
>> +       int errors = 0;
>>
>>          if (clock_gettime(CLOCK_MONOTONIC, &ts1))
>>                  return -1;
>> @@ -79,12 +80,17 @@ int migrate(uint64_t *ptr, int n1, int n2)
>>                  ret = move_pages(0, 1, (void **) &ptr, &n2, &status,
>>                                  MPOL_MF_MOVE_ALL);
>>                  if (ret) {
>> -                       if (ret > 0)
>> +                       if (ret > 0) {
>> +                               if (++errors < 100)
>> +                                       continue;
>>                                  printf("Didn't migrate %d pages\n", ret);
>> -                       else
>> +                       } else {
>>                                  perror("Couldn't migrate pages");
>> +                       }
>>                          return -2;
>>                  }
>> +               /* Progress! */
>> +               errors = 0;
>>
>>                  tmp = n2;
>>                  n2 = n1;
>>
>>
>> [root@localhost mm]# ./migration
>> TAP version 13
>> 1..1
>> # Starting 1 tests from 1 test cases.
>> #  RUN           migration.shared_anon ...
>> #            OK  migration.shared_anon
>> ok 1 migration.shared_anon
>> # PASSED: 1 / 1 tests passed.
>> # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
> 
> 
> This does make the test pass, to my surprise, since what you are doing
> from userspace
> 
> should have been done by the kernel, because it retries folio unmapping
> and moving
> 
> NR_MAX_MIGRATE_(A)SYNC_RETRY times; I had already tested pumping up these
> 
> macros and the original test was still failing. Now, I digged in more,
> and, if the
> 
> following assertion is correct:
> 
> 
> Any thread having a reference on a folio will end up calling folio_lock()
> 

Good point. I suspect concurrent things like read/write would also be 
able to trigger this (did not check, though).

> 
> then it seems to me that the retry for loop wrapped around
> migrate_folio_move(), inside
> 
> migrate_pages_batch(), is useless; if migrate_folio_move() fails on the
> first iteration, it is
> 
> going to fail for all iterations since, if I am reading the code path
> correctly, the only way it
> 
> fails is when the actual refcount is not equal to expected refcount (in
> folio_migrate_mapping()),
> 
> and there is no way that the extra refcount is going to get released
> since the migration path
> 
> has the folio lock.
> 
> And therefore, this begs the question: isn't it logical to assert the
> actual refcount against the
> 
> expected refcount, just after we have changed the PTEs, so that if this
> assertion fails, we can
> 
> go to the next iteration of the for loop for migrate_folio_unmap()
> inside migrate_pages_batch()
> 
> by calling migrate_folio_undo_src()/dst() to restore the old state? I am
> trying to implement
> 
> this but is not as straightforward as it seemed to me this morning.

I agree with your assessment that migration code currently doesn't 
handle the case well when some other thread does an unconditional 
folio_lock(). folio_trylock() users would be handled, but that's not 
what we want with FGP_LOCK I assume.

So IIUC, your idea would be to unlock the folio in migration code and 
try again their. Sounds reasonable, without looking into the details :)

-- 
Cheers,

David / dhildenb