Re: [PATCH] mm: huge_memory: add folio_mark_accessed() when zapping file THP

[David Hildenbrand <david@redhat.com>]

On 09.04.25 11:49, Baolin Wang wrote:
> 
> 
> On 2025/4/9 17:46, David Hildenbrand wrote:
>> On 09.04.25 11:38, Baolin Wang wrote:
>>> When investigating performance issues during file folio unmap, I
>>> noticed some
>>> behavioral differences in handling non-PMD-sized folios and PMD-sized
>>> folios.
>>> For non-PMD-sized file folios, it will call folio_mark_accessed() to
>>> mark the
>>> folio as having seen activity, but this is not done for PMD-sized folios.
>>>
>>> This might not cause obvious issues, but a potential problem could be
>>> that,
>>> it might lead to reclaim hot file folios under memory pressure, as quoted
>>> from Johannes:
>>>
>>> "
>>> Sometimes file contents are only accessed through relatively short-lived
>>> mappings. But they can nevertheless be accessed a lot and be hot. It's
>>> important to not lose that information on unmap, and end up kicking out a
>>> frequently used cache page.
>>> "
>>>
>>> Therefore, we should also add folio_mark_accessed() for PMD-sized file
>>> folios when unmapping.
>>>
>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>> Acked-by: Johannes Weiner <hannes@cmpxchg.org>
>>> Acked-by: Zi Yan <ziy@nvidia.com>
>>> ---
>>> Changes from RFC:
>>>    - Update the commit message, per Johannes.
>>>    - Collect Acked tags from Johannes and Zi. Thanks.
>>> ---
>>>    mm/huge_memory.c | 4 ++++
>>>    1 file changed, 4 insertions(+)
>>>
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index 2a47682d1ab7..955781b4e946 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -2260,6 +2260,10 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct
>>> vm_area_struct *vma,
>>>                    zap_deposited_table(tlb->mm, pmd);
>>>                add_mm_counter(tlb->mm, mm_counter_file(folio),
>>>                           -HPAGE_PMD_NR);
>>> +
>>> +            if (flush_needed && pmd_young(orig_pmd) &&
>>> +                likely(vma_has_recency(vma)))
>>> +                folio_mark_accessed(folio);
>>
>> So the flush_needed check is really just a pmd_present() check. (the
>> latter would be clearer, but I don't mind)
> 
> Yes, we've already checked pmd_present() before, so I assume the
> flush_needed check is cheaper:)

Maybe :)

-- 
Cheers,

David / dhildenb