Re: [PATCH] mm/rmap: do not add fully unmapped large folio to deferred split list

[David Hildenbrand <david@redhat.com>]

On 12.04.24 20:29, Yang Shi wrote:
> On Fri, Apr 12, 2024 at 7:31 AM Zi Yan <ziy@nvidia.com> wrote:
>>
>> On 12 Apr 2024, at 10:21, Zi Yan wrote:
>>
>>> On 11 Apr 2024, at 17:59, Yang Shi wrote:
>>>
>>>> On Thu, Apr 11, 2024 at 2:15 PM David Hildenbrand <david@redhat.com> wrote:
>>>>>
>>>>> On 11.04.24 21:01, Yang Shi wrote:
>>>>>> On Thu, Apr 11, 2024 at 8:46 AM David Hildenbrand <david@redhat.com> wrote:
>>>>>>>
>>>>>>> On 11.04.24 17:32, Zi Yan 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 mapcount before
>>>>>>>> adding a folio to deferred split list.
>>>>>>>>
>>>>>>>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>>>>>>>> ---
>>>>>>>>     mm/rmap.c | 9 ++++++---
>>>>>>>>     1 file changed, 6 insertions(+), 3 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>>>>>>> index 2608c40dffad..d599a772e282 100644
>>>>>>>> --- a/mm/rmap.c
>>>>>>>> +++ b/mm/rmap.c
>>>>>>>> @@ -1494,7 +1494,7 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>>>>>>>>                 enum rmap_level level)
>>>>>>>>     {
>>>>>>>>         atomic_t *mapped = &folio->_nr_pages_mapped;
>>>>>>>> -     int last, nr = 0, nr_pmdmapped = 0;
>>>>>>>> +     int last, nr = 0, nr_pmdmapped = 0, mapcount = 0;
>>>>>>>>         enum node_stat_item idx;
>>>>>>>>
>>>>>>>>         __folio_rmap_sanity_checks(folio, page, nr_pages, level);
>>>>>>>> @@ -1506,7 +1506,8 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>>>>>>>>                         break;
>>>>>>>>                 }
>>>>>>>>
>>>>>>>> -             atomic_sub(nr_pages, &folio->_large_mapcount);
>>>>>>>> +             mapcount = atomic_sub_return(nr_pages,
>>>>>>>> +                                          &folio->_large_mapcount) + 1;
>>>>>>>
>>>>>>> That becomes a new memory barrier on some archs. Rather just re-read it
>>>>>>> below. Re-reading should be fine here.
>>>>>>>
>>>>>>>>                 do {
>>>>>>>>                         last = atomic_add_negative(-1, &page->_mapcount);
>>>>>>>>                         if (last) {
>>>>>>>> @@ -1554,7 +1555,9 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>>>>>>>>                  * is still mapped.
>>>>>>>>                  */
>>>>>>>>                 if (folio_test_large(folio) && folio_test_anon(folio))
>>>>>>>> -                     if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
>>>>>>>> +                     if ((level == RMAP_LEVEL_PTE &&
>>>>>>>> +                          mapcount != 0) ||
>>>>>>>> +                         (level == RMAP_LEVEL_PMD && nr < nr_pmdmapped))
>>>>>>>>                                 deferred_split_folio(folio);
>>>>>>>>         }
>>>>>>>
>>>>>>> But I do wonder if we really care? Usually the folio will simply get
>>>>>>> freed afterwards, where we simply remove it from the list.
>>>>>>>
>>>>>>> If it's pinned, we won't be able to free or reclaim, but it's rather a
>>>>>>> corner case ...
>>>>>>>
>>>>>>> Is it really worth the added code? Not convinced.
>>>>>>
>>>>>> It is actually not only an optimization, but also fixed the broken
>>>>>> thp_deferred_split_page counter in /proc/vmstat.
>>>>>>
>>>>>> The counter actually counted the partially unmapped huge pages (so
>>>>>> they are on deferred split queue), but it counts the fully unmapped
>>>>>> mTHP as well now. For example, when a 64K THP is fully unmapped, the
>>>>>> thp_deferred_split_page is not supposed to get inc'ed, but it does
>>>>>> now.
>>>>>>
>>>>>> The counter is also useful for performance analysis, for example,
>>>>>> whether a workload did a lot of partial unmap or not. So fixing the
>>>>>> counter seems worthy. Zi Yan should have mentioned this in the commit
>>>>>> log.
>>>>>
>>>>> Yes, all that is information that is missing from the patch description.
>>>>> If it's a fix, there should be a "Fixes:".
>>>>>
>>>>> Likely we want to have a folio_large_mapcount() check in the code below.
>>>>> (I yet have to digest the condition where this happens -- can we have an
>>>>> example where we'd use to do the wrong thing and now would do the right
>>>>> thing as well?)
>>>>
>>>> For example, map 1G memory with 64K mTHP, then unmap the whole 1G or
>>>> some full 64K areas, you will see thp_deferred_split_page increased,
>>>> but it shouldn't.
>>>>
>>>> It looks __folio_remove_rmap() incorrectly detected whether the mTHP
>>>> is fully unmapped or partially unmapped by comparing the number of
>>>> still-mapped subpages to ENTIRELY_MAPPED, which should just work for
>>>> PMD-mappable THP.
>>>>
>>>> However I just realized this problem was kind of workaround'ed by commit:
>>>>
>>>> commit 98046944a1597f3a02b792dbe9665e9943b77f28
>>>> Author: Baolin Wang <baolin.wang@linux.alibaba.com>
>>>> Date:   Fri Mar 29 14:59:33 2024 +0800
>>>>
>>>>      mm: huge_memory: add the missing folio_test_pmd_mappable() for THP
>>>> split statistics
>>>>
>>>>      Now the mTHP can also be split or added into the deferred list, so add
>>>>      folio_test_pmd_mappable() validation for PMD mapped THP, to avoid
>>>>      confusion with PMD mapped THP related statistics.
>>>>
>>>>      Link: https://lkml.kernel.org/r/a5341defeef27c9ac7b85c97f030f93e4368bbc1.1711694852.git.baolin.wang@linux.alibaba.com
>>>>      Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>>>      Acked-by: David Hildenbrand <david@redhat.com>
>>>>      Cc: Muchun Song <muchun.song@linux.dev>
>>>>      Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
>>>>
>>>> This commit made thp_deferred_split_page didn't count mTHP anymore, it
>>>> also made thp_split_page didn't count mTHP anymore.
>>>>
>>>> However Zi Yan's patch does make the code more robust and we don't
>>>> need to worry about the miscounting issue anymore if we will add
>>>> deferred_split_page and split_page counters for mTHP in the future.
>>>
>>> Actually, the patch above does not fix everything. A fully unmapped
>>> PTE-mapped order-9 THP is also added to deferred split list and
>>> counted as THP_DEFERRED_SPLIT_PAGE without my patch, 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().
>>>
>>> I will add this information in the next version.
>>
>> It might
>> Fixes: b06dc281aa99 ("mm/rmap: introduce folio_remove_rmap_[pte|ptes|pmd]()"),
>> but before this commit fully unmapping a PTE-mapped order-9 THP still increased
>> THP_DEFERRED_SPLIT_PAGE, because PTEs are unmapped individually and first PTE
>> unmapping adds the THP into the deferred split list. This means commit b06dc281aa99
>> did not change anything and before that THP_DEFERRED_SPLIT_PAGE increase is
>> due to implementation. I will add this to the commit log as well without Fixes
>> tag.
> 
> Thanks for digging deeper. The problem may be not that obvious before
> mTHP because PMD-mappable THP is converted to PTE-mapped due to
> partial unmap in most cases. But mTHP is always PTE-mapped in the
> first place. The other reason is batched rmap remove was not supported
> before David's optimization.

Yes.

> 
> Now we do have reasonable motivation to make it precise and it is also
> easier to do so than before.

If by "precise" you mean "less unreliable in some cases", yes. See my 
other mail.

-- 
Cheers,

David / dhildenb