On 25 Apr 2024, at 3:15, David Hildenbrand wrote:

> On 25.04.24 00:39, Zi Yan wrote:
>> On 24 Apr 2024, at 18:32, Yang Shi wrote:
>>
>>> On Wed, Apr 24, 2024 at 2:10 PM Zi Yan <zi.yan@sent.com> wrote:
>>>>
>>>> From: Zi Yan <ziy@nvidia.com>
>>>>
>>>> In __folio_remove_rmap(), a large folio is added to deferred split list
>>>> if any page in a folio loses its final mapping. It is possible that
>>>> the folio is unmapped fully, but it is unnecessary to add the folio
>>>> to deferred split list at all. Fix it by checking folio->_nr_pages_mapped
>>>> before adding a folio to deferred split list. If the folio is already
>>>> on the deferred split list, it will be skipped.
>>>>
>>>> Commit 98046944a159 ("mm: huge_memory: add the missing
>>>> folio_test_pmd_mappable() for THP split statistics") tried to exclude
>>>> mTHP deferred split stats from THP_DEFERRED_SPLIT_PAGE, but it does not
>>>> fix everything. A fully unmapped PTE-mapped order-9 THP was also added to
>>>> deferred split list and counted as THP_DEFERRED_SPLIT_PAGE, since nr is
>>>> 512 (non zero), level is RMAP_LEVEL_PTE, and inside deferred_split_folio()
>>>> the order-9 folio is folio_test_pmd_mappable(). However, this miscount
>>>> was present even earlier due to implementation, since PTEs are unmapped
>>>> individually and first PTE unmapping adds the THP into the deferred split
>>>> list.
>>>
>>> Shall you mention the miscounting for mTHP too? There is another patch
>>> series adding the counter support for mTHP.
>>
>> OK, will add it.
>
> I thought I made it clear: this patch won't "fix" it. Misaccounting will still happen. Just less frequently.
>
> Please spell that out.

Sure. Sorry I did not make that clear.


>
>>>
>>>>
>>>> With commit b06dc281aa99 ("mm/rmap: introduce
>>>> folio_remove_rmap_[pte|ptes|pmd]()"), kernel is able to unmap PTE-mapped
>>>> folios in one shot without causing the miscount, hence this patch.
>>>>
>>>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>>>> ---
>>>>   mm/rmap.c | 8 +++++---
>>>>   1 file changed, 5 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>> index a7913a454028..220ad8a83589 100644
>>>> --- a/mm/rmap.c
>>>> +++ b/mm/rmap.c
>>>> @@ -1553,9 +1553,11 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>>>>                   * page of the folio is unmapped and at least one page
>>>>                   * is still mapped.
>>>>                   */
>>>> -               if (folio_test_large(folio) && folio_test_anon(folio))
>>>> -                       if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
>>>> -                               deferred_split_folio(folio);
>>>> +               if (folio_test_large(folio) && folio_test_anon(folio) &&
>>>> +                   list_empty(&folio->_deferred_list) &&
>>>
>>> Do we really need this check? deferred_split_folio() does the same
>>> check too. Bailing out earlier sounds ok too, but there may not be too
>>> much gain.
>>
>> Sure, I can remove it.
>
> Please leave it. It's a function call that cannot be optimized out otherwise.

OK. If you think it is worth optimizing that function call, I will keep it.


--
Best Regards,
Yan, Zi