Re: [PATCH] mm: migrate: record the mlocked page status to remove unnecessary lru drain

["Yin, Fengwei" <fengwei.yin@intel.com>]

On 10/20/2023 10:45 AM, Baolin Wang wrote:
> 
> 
> On 10/20/2023 10:30 AM, Yin, Fengwei wrote:
>>
>>
>> On 10/20/2023 10:09 AM, Baolin Wang wrote:
>>>
>>>
>>> On 10/19/2023 8:07 PM, Yin, Fengwei wrote:
>>>>
>>>>
>>>> On 10/19/2023 4:51 PM, Baolin Wang wrote:
>>>>>
>>>>>
>>>>> On 10/19/2023 4:22 PM, Yin Fengwei wrote:
>>>>>> Hi Baolin,
>>>>>>
>>>>>> On 10/19/23 15:25, Baolin Wang wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 10/19/2023 2:09 PM, Huang, Ying wrote:
>>>>>>>> Zi Yan <ziy@nvidia.com> writes:
>>>>>>>>
>>>>>>>>> On 18 Oct 2023, at 9:04, Baolin Wang wrote:
>>>>>>>>>
>>>>>>>>>> When doing compaction, I found the lru_add_drain() is an obvious hotspot
>>>>>>>>>> when migrating pages. The distribution of this hotspot is as follows:
>>>>>>>>>>        - 18.75% compact_zone
>>>>>>>>>>           - 17.39% migrate_pages
>>>>>>>>>>              - 13.79% migrate_pages_batch
>>>>>>>>>>                 - 11.66% migrate_folio_move
>>>>>>>>>>                    - 7.02% lru_add_drain
>>>>>>>>>>                       + 7.02% lru_add_drain_cpu
>>>>>>>>>>                    + 3.00% move_to_new_folio
>>>>>>>>>>                      1.23% rmap_walk
>>>>>>>>>>                 + 1.92% migrate_folio_unmap
>>>>>>>>>>              + 3.20% migrate_pages_sync
>>>>>>>>>>           + 0.90% isolate_migratepages
>>>>>>>>>>
>>>>>>>>>> The lru_add_drain() was added by commit c3096e6782b7 ("mm/migrate:
>>>>>>>>>> __unmap_and_move() push good newpage to LRU") to drain the newpage to LRU
>>>>>>>>>> immediately, to help to build up the correct newpage->mlock_count in
>>>>>>>>>> remove_migration_ptes() for mlocked pages. However, if there are no mlocked
>>>>>>>>>> pages are migrating, then we can avoid this lru drain operation, especailly
>>>>>>>>>> for the heavy concurrent scenarios.
>>>>>>>>>
>>>>>>>>> lru_add_drain() is also used to drain pages out of folio_batch. Pages in folio_batch
>>>>>>>>> have an additional pin to prevent migration. See folio_get(folio); in folio_add_lru().
>>>>>>>>
>>>>>>>> lru_add_drain() is called after the page reference count checking in
>>>>>>>> move_to_new_folio().  So, I don't this is an issue.
>>>>>>>
>>>>>>> Agree. The purpose of adding lru_add_drain() is to address the 'mlock_count' issue for mlocked pages. Please see commit c3096e6782b7 and related comments. Moreover I haven't seen an increase in the number of page migration failures due to page reference count checking after this patch.
>>>>>>
>>>>>> I agree with your. My understanding also is that the lru_add_drain() is only needed
>>>>>> for mlocked folio to correct mlock_count. Like to hear the confirmation from Huge.
>>>>>>
>>>>>>
>>>>>> But I have question: why do we need use page_was_mlocked instead of check
>>>>>> folio_test_mlocked(src)? Does page migration clear the mlock flag? Thanks.
>>>>>
>>>>> Yes, please see the call trace: try_to_migrate_one() ---> page_remove_rmap() ---> munlock_vma_folio().
>>>>
>>>> Yes. This will clear mlock bit.
>>>>
>>>> What about set dst folio mlocked if source is before try_to_migrate_one()? And
>>>> then check whether dst folio is mlocked after? And need clear mlocked if migration
>>>> fails. I suppose the change is minor. Just a thought. Thanks.
>>>
>>> IMO, this will break the mlock related statistics in mlock_folio() when the remove_migration_pte() rebuilds the mlock status and mlock count.
>>>
>>> Another concern I can see is that, during the page migration, a concurrent munlock() can be called to clean the VM_LOCKED flags for the VMAs, so the remove_migration_pte() should not rebuild the mlock status and mlock count. But the dst folio's mlcoked status is still remained, which is wrong.
>>>
>>> So your suggested apporach seems not easy, and I think my patch is simple with re-using existing __migrate_folio_record() and __migrate_folio_extract() :)
>>
>> Can these concerns be addressed by clear dst mlocked after lru_add_drain() but before
>> remove_migration_pte()?
> 
> IMHO, that seems too hacky to me. I still prefer to rely on the migration process of the mlcock pages.

BTW, Yosry tried to address the overlap of field lru and mlock_count:
https://lore.kernel.org/lkml/20230618065719.1363271-1-yosryahmed@google.com/
But the lore doesn't group all the patches.


Regards
Yin, Fengwei