Re: [PATCH v4] mm/migrate: split source folio if it is on deferred split list

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 4367 bytes --]

On 20 Mar 2024, at 12:02, Matthew Wilcox wrote:

> On Tue, Mar 19, 2024 at 09:45:11PM -0400, Zi Yan wrote:
>> +++ b/mm/migrate.c
>> @@ -1654,25 +1654,65 @@ static int migrate_pages_batch(struct list_head *from,
>>
>>  			/*
>>  			 * Large folio migration might be unsupported or
>> -			 * the allocation might be failed so we should retry
>> -			 * on the same folio with the large folio split
>> +			 * the folio is on deferred split list so we should
>> +			 * retry on the same folio with the large folio split
>>  			 * to normal folios.
>>  			 *
>>  			 * Split folios are put in split_folios, and
>>  			 * we will migrate them after the rest of the
>>  			 * list is processed.
>>  			 */
>> -			if (!thp_migration_supported() && is_thp) {
>> -				nr_failed++;
>> -				stats->nr_thp_failed++;
>> -				if (!try_split_folio(folio, split_folios)) {
>> -					stats->nr_thp_split++;
>> -					stats->nr_split++;
>> +			if (is_thp) {
>> +				bool is_on_deferred_list = false;
>> +
>> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
>> +				/*
>> +				 * Check without taking split_queue_lock to
>> +				 * reduce locking overheads. The worst case is
>> +				 * that if the folio is put on the deferred
>> +				 * split list after the check, it will be
>> +				 * migrated and not put back on the list.
>> +				 * The migrated folio will not be split
>> +				 * via shrinker during memory pressure.
>> +				 */
>> +				if (!data_race(list_empty(&folio->_deferred_list))) {
>> +					struct deferred_split *ds_queue;
>> +					unsigned long flags;
>> +
>> +					ds_queue =
>> +						get_deferred_split_queue(folio);
>> +					spin_lock_irqsave(&ds_queue->split_queue_lock,
>> +							  flags);
>> +					/*
>> +					 * Only check if the folio is on
>> +					 * deferred split list without removing
>> +					 * it. Since the folio can be on
>> +					 * deferred_split_scan() local list and
>> +					 * removing it can cause the local list
>> +					 * corruption. Folio split process
>> +					 * below can handle it with the help of
>> +					 * folio_ref_freeze().
>> +					 */
>> +					is_on_deferred_list =
>> +						!list_empty(&folio->_deferred_list);
>> +					spin_unlock_irqrestore(&ds_queue->split_queue_lock,
>> +							       flags);
>> +				}
>> +#endif
>> +				if (!thp_migration_supported() ||
>> +						is_on_deferred_list) {
>> +					nr_failed++;
>> +					stats->nr_thp_failed++;
>> +					if (!try_split_folio(folio,
>> +							     split_folios)) {
>> +						stats->nr_thp_split++;
>> +						stats->nr_split++;
>> +						continue;
>> +					}
>> +					stats->nr_failed_pages += nr_pages;
>> +					list_move_tail(&folio->lru, ret_folios);
>>  					continue;
>>  				}
>> -				stats->nr_failed_pages += nr_pages;
>> -				list_move_tail(&folio->lru, ret_folios);
>> -				continue;
>>  			}
>
> I don't think we need to try quite this hard.  I don't think we need
> to take the lock to be certain if it's on the deferred list -- is
> there anything preventing the folio being added to the deferred list
> after we drop the lock?

No. OK, I will use the less hard version.

>
> I also don't think we should account this as a thp split since those
> are treated by callers as failures.  So maybe this?

I think we need to match the stats with code behavior, otherwise userspace
caller can get confused by the results, where only a subset of a folio
is migrated and split stats and failure stats are not bumped accordingly.

>
> +++ b/mm/migrate.c
> @@ -1652,6 +1652,17 @@ static int migrate_pages_batch(struct list_head *from,
>
>                         cond_resched();
>
> +                       /*
> +                        * The rare folio on the deferred split list should
> +                        * be split now.  It should not count as a failure.
> +                        */
> +                       if (nr_pages > 2 &&
> +                           !list_empty(&folio->_deferred_list)) {
> +                               if (try_split_folio(folio, from) == 0) {
> +                                       is_large = is_thp = false;
> +                                       nr_pages = 1;
> +                               }
> +                       }
>                         /*
>                          * Large folio migration might be unsupported or
>                          * the allocation might be failed so we should retry


--
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]