Re: [RFC PATCH] mm: filemap: avoid unnecessary major faults in filemap_fault()

[Yin Fengwei]

On 11/23/23 15:57, zhangpeng (AS) wrote:
> On 2023/11/23 13:26, Yin Fengwei wrote:
> 
>> On 11/23/23 12:12, zhangpeng (AS) wrote:
>>> On 2023/11/23 9:09, Yin Fengwei wrote:
>>>
>>>> Hi Peng,
>>>>
>>>> On 11/22/23 22:00, Peng Zhang wrote:
>>>>> From: ZhangPeng <zhangpeng362@xxxxxxxxxx>
>>>>>
>>>>> The major fault occurred when using mlockall(MCL_CURRENT | MCL_FUTURE)
>>>>> in application, which leading to an unexpected performance issue[1].
>>>>>
>>>>> This caused by temporarily cleared pte during a read/modify/write update
>>>>> of the pte, eg, do_numa_page()/change_pte_range().
>>>>>
>>>>> For the data segment of the user-mode program, the global variable area
>>>>> is a private mapping. After the pagecache is loaded, the private anonymous
>>>>> page is generated after the COW is triggered. Mlockall can lock COW pages
>>>>> (anonymous pages), but the original file pages cannot be locked and may
>>>>> be reclaimed. If the global variable (private anon page) is accessed when
>>>>> vmf->pte is zeroed in numa fault, a file page fault will be triggered.
>>>>>
>>>>> At this time, the original private file page may have been reclaimed.
>>>>> If the page cache is not available at this time, a major fault will be
>>>>> triggered and the file will be read, causing additional overhead.
>>>>>
>>>>> Fix this by rechecking the pte by holding ptl in filemap_fault() before
>>>>> triggering a major fault.
>>>>>
>>>>> [1] https://lore.kernel.org/linux-mm/9e62fd9a-bee0-52bf-50a7-498fa17434ee@xxxxxxxxxx/
>>>>>
>>>>> Signed-off-by: ZhangPeng <zhangpeng362@xxxxxxxxxx>
>>>>> Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
>>>>> ---
>>>>>    mm/filemap.c | 14 ++++++++++++++
>>>>>    1 file changed, 14 insertions(+)
>>>>>
>>>>> diff --git a/mm/filemap.c b/mm/filemap.c
>>>>> index 71f00539ac00..bb5e6a2790dc 100644
>>>>> --- a/mm/filemap.c
>>>>> +++ b/mm/filemap.c
>>>>> @@ -3226,6 +3226,20 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
>>>>>                mapping_locked = true;
>>>>>            }
>>>>>        } else {
>>>>> +        pte_t *ptep = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
>>>>> +                          vmf->address, &vmf->ptl);
>>>>> +        if (ptep) {
>>>>> +            /*
>>>>> +             * Recheck pte with ptl locked as the pte can be cleared
>>>>> +             * temporarily during a read/modify/write update.
>>>>> +             */
>>>>> +            if (unlikely(!pte_none(ptep_get(ptep))))
>>>>> +                ret = VM_FAULT_NOPAGE;
>>>>> +            pte_unmap_unlock(ptep, vmf->ptl);
>>>>> +            if (unlikely(ret))
>>>>> +                return ret;
>>>>> +        }
>>>> I am curious. Did you try not to take PTL here and just check whether PTE is not NONE?
>>> Thank you for your reply.
>>>
>>> If we don't take PTL, the current use case won't trigger this issue either.
>> Is this verified by testing or just in theory?
> 
> If we add a delay between ptep_modify_prot_start() and ptep_modify_prot_commit(),
> this issue will also trigger. Without delay, we haven't reproduced this problem
> so far.
Thanks for the testing.

> 
>>> In most cases, if we don't take PTL, this issue won't be triggered. However,
>>> there is still a possibility of triggering this issue. The corner case is that
>>> task 2 triggers a page fault when task 1 is between ptep_modify_prot_start()
>>> and ptep_modify_prot_commit() in do_numa_page(). Furthermore,task 2 passes the
>>> check whether the PTE is not NONE before task 1 updates PTE in
>>> ptep_modify_prot_commit() without taking PTL.
>> There is very limited operations between ptep_modify_prot_start() and
>> ptep_modify_prot_commit(). While the code path from page fault to this check is
>> long. My understanding is it's very likely the PTE is not NONE when do PTE check
>> here without hold PTL (This is my theory. :)).
> 
> Yes, there is a high probability that this issue won't occur without taking PTL.
> 
>> In the other side, acquiring/releasing PTL may bring performance impaction. It may
>> not be big deal because the IO operations in this code path. But it's better to
>> collect some performance data IMHO.
> 
> We tested the performance of file private mapping page fault (page_fault2.c of
> will-it-scale [1]) and file shared mapping page fault (page_fault3.c of will-it-scale).
> The difference in performance (in operations per second) before and after patch
> applied is about 0.7% on a x86 physical machine.
> 
> [1] https://github.com/antonblanchard/will-it-scale/tree/master
Maybe include this performance related information to commit message?

For the code change, looks good to me.

Reviewed-by: Yin Fengwei <fengwei.yin@xxxxxxxxx>


Regards
Yin, Fengwei

> 
>>
>> Regards
>> Yin, Fengwei
>>
>>>> Regards
>>>> Yin, Fengwei
>>>>
>>>>> +
>>>>>            /* No page in the page cache at all */
>>>>>            count_vm_event(PGMAJFAULT);
>>>>>            count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
>