Skip to contentSkip to conversation

mm: unqueue deferred split folios outside the lruvec lock

PATCH Superseded Joanne Koong 3 participants 4 messages Oct 17 – Oct 19

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 mm: unqueue deferred split folios outside the lruvec lock

move_folios_to_lru(), folio_batch_move_lru() and folios_put_refs() call folio_unqueue_deferred_split() on folios where the refcount has dropped to zero. In these paths, folio_unqueue_deferred_split() is called with the lruvec lock often held (move_folios_to_lru() always holds it and folio_batch_move_lru()/folios_put_refs() usually hold it for every folio in the batch except the first). However, it's not needed to hold the lruvec lock.

The lruvec lock protects folio->lru and the reclaim LRU lists whereas the deferred split queue lives on folio->_deferred_list in deferred_split_lru. The deferred_split_lru is a list_lru with its own per-memcg per-node lock.

There are already other existing callers that call folio_unqueue_deferred_split() without holding the lruvec lock (eg __folio_put(), shrink_folio_list(), __folio_migrate_mapping(), ...).

Add folio_batch_unqueue_deferred_split() and use it in move_folios_to_lru(), folio_batch_move_lru(), and folios_put_refs() to unqueue each batch of zero-refcount folios from the deferred split list after the lruvec lock is released and before mem_cgroup_uncharge_folios(). This helps shorten the lruvec lock's critical sections, which run with IRQs disabled and are highly contended.

---
This patch came out of a discussion about splitting underused mTHP folios [1], where queueing small mTHP folios on the deferred split list was found to regress Barry's fault microbenchmark. The cost turned out to be the unqueue running under the lruvec lock rather than contention on the deferred split queue itself. Testing this patch on top of the series in [1] with the same microbenchmark, queueing every 16k folio went from a 15% runtime cost to noise, with lruvec lock wait dropping from 491 seconds to 38 seconds (most of that contention then moves to the deferred split list_lru lock, but runtime still drops back to the unqueued baseline).

[1] https://lore.kernel.org/linux-mm/CAJnrk1a2PpUGTsCgS7jct2KET4K2zAL+_h+te8G1MaU9bXs99w@mail.gmail.com/

v1 -> v2:
- Reword comment in folio_batch_move_lru() (David)
v1 link: https://lore.kernel.org/linux-mm/[email protected]/

 mm/folio.c    | 24 +++++++++++++++++++-----
 mm/internal.h |  1 +
 mm/vmscan.c   |  3 ++-
 3 files changed, 22 insertions(+), 6 deletions(-)
mm/folio.c +19 -5
index 35e242b48870..45c814f3857b 100644
@@ -109,6 +109,19 @@ EXPORT_SYMBOL(__folio_put);
 
 typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
 
+/*
+ * Unqueue a batch of folios from the deferred split queue. This must be
+ * called after the folios' refcounts have dropped to zero and before they
+ * are uncharged (see comment above __folio_unqueue_deferred_split() for why).
+ */
+void folio_batch_unqueue_deferred_split(struct folio_batch *fbatch)
+{
+	unsigned int i;
+
+	for (i = 0; i < folio_batch_count(fbatch); i++)
+		folio_unqueue_deferred_split(fbatch->folios[i]);
+}
+
 static void lru_add(struct lruvec *lruvec, struct folio *folio)
 {
 	int was_unevictable = folio_test_clear_unevictable(folio);
@@ -177,14 +190,10 @@ static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
 		 *
 		 * We're bypassing normal cleanup. Clear flags that are not
 		 * applicable to dead folios.
-		 *
-		 * Since the folio may be part of a huge page, unqueue from
-		 * deferred split list to avoid a dangling list entry.
 		 */
 		if (is_lru_add && folio_ref_freeze(folio, 1)) {
 			__folio_clear_active(folio);
 			__folio_clear_unevictable(folio);
-			folio_unqueue_deferred_split(folio);
 			fbatch->folios[i] = NULL;
 			folio_batch_add(&free_fbatch, folio);
 			continue;
@@ -201,6 +210,11 @@ static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
 
 	/* Cleanup filtered dead folios. */
 	if (is_lru_add) {
+		/*
+		 * Unqueue any large folios from the deferred split queue to
+		 * avoid a dangling list entry once the folios are freed.
+		 */
+		folio_batch_unqueue_deferred_split(&free_fbatch);
 		mem_cgroup_uncharge_folios(&free_fbatch);
 		free_unref_folios(&free_fbatch);
 	}
@@ -1020,7 +1034,6 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 			free_huge_folio(folio);
 			continue;
 		}
-		folio_unqueue_deferred_split(folio);
 		__page_cache_release(folio, &lruvec, &flags);
 
 		if (j != i)
@@ -1035,6 +1048,7 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 	}
 
 	folios->nr = j;
+	folio_batch_unqueue_deferred_split(folios);
 	mem_cgroup_uncharge_folios(folios);
 	free_unref_folios(folios);
 }
mm/internal.h +1 -0
index d497c35ee783..33ae388feb33 100644
@@ -761,6 +761,7 @@ static inline bool folio_unqueue_deferred_split(struct folio *folio)
 
 	return __folio_unqueue_deferred_split(folio);
 }
+void folio_batch_unqueue_deferred_split(struct folio_batch *fbatch);
 
 static inline struct folio *page_rmappable_folio(struct page *page)
 {
mm/vmscan.c +2 -1
index daa7e37c2236..9c71a50309e8 100644
@@ -2007,9 +2007,9 @@ static unsigned int move_folios_to_lru(struct list_head *list, bool do_rotate)
 		if (unlikely(folio_put_testzero(folio))) {
 			__folio_clear_lru_flags(folio);
 
-			folio_unqueue_deferred_split(folio);
 			if (folio_batch_add(&free_folios, folio) == 0) {
 				lruvec_unlock_irq(lruvec);
+				folio_batch_unqueue_deferred_split(&free_folios);
 				mem_cgroup_uncharge_folios(&free_folios);
 				free_unref_folios(&free_folios);
 				lruvec = NULL;
@@ -2041,6 +2041,7 @@ static unsigned int move_folios_to_lru(struct list_head *list, bool do_rotate)
 		lruvec_unlock_irq(lruvec);
 
 	if (free_folios.nr) {
+		folio_batch_unqueue_deferred_split(&free_folios);
 		mem_cgroup_uncharge_folios(&free_folios);
 		free_unref_folios(&free_folios);
 	}
Signature · 1 line
2.52.0
Reviewed-by on v2 1/1
#1 lore raw
Quoted 26 lines · Joanne Koong

move_folios_to_lru(), folio_batch_move_lru() and folios_put_refs() call folio_unqueue_deferred_split() on folios where the refcount has dropped to zero. In these paths, folio_unqueue_deferred_split() is called with the lruvec lock often held (move_folios_to_lru() always holds it and folio_batch_move_lru()/folios_put_refs() usually hold it for every folio in the batch except the first). However, it's not needed to hold the lruvec lock.

The lruvec lock protects folio->lru and the reclaim LRU lists whereas the deferred split queue lives on folio->_deferred_list in deferred_split_lru. The deferred_split_lru is a list_lru with its own per-memcg per-node lock.

There are already other existing callers that call folio_unqueue_deferred_split() without holding the lruvec lock (eg __folio_put(), shrink_folio_list(), __folio_migrate_mapping(), ...).

Add folio_batch_unqueue_deferred_split() and use it in move_folios_to_lru(), folio_batch_move_lru(), and folios_put_refs() to unqueue each batch of zero-refcount folios from the deferred split list after the lruvec lock is released and before mem_cgroup_uncharge_folios(). This helps shorten the lruvec lock's critical sections, which run with IRQs disabled and are highly contended.

Quoted 10 lines

---
This patch came out of a discussion about splitting underused mTHP folios [1], where queueing small mTHP folios on the deferred split list was found to regress Barry's fault microbenchmark. The cost turned out to be the unqueue running under the lruvec lock rather than contention on the deferred split queue itself. Testing this patch on top of the series in [1] with the same microbenchmark, queueing every 16k folio went from a 15% runtime cost to noise, with lruvec lock wait dropping from 491 seconds to 38 seconds (most of that contention then moves to the deferred split list_lru lock, but runtime still drops back to the unqueued baseline).

Looks great :) Could you also keep the benchmark results above the '---'? They'd be useful to have in the changelog, along with the test setup.

With that, LGTM!

reviewer
#2 lore raw
Quoted 41 lines · Lance Yang
Quoted 26 lines · Joanne Koong

move_folios_to_lru(), folio_batch_move_lru() and folios_put_refs() call folio_unqueue_deferred_split() on folios where the refcount has dropped to zero. In these paths, folio_unqueue_deferred_split() is called with the lruvec lock often held (move_folios_to_lru() always holds it and folio_batch_move_lru()/folios_put_refs() usually hold it for every folio in the batch except the first). However, it's not needed to hold the lruvec lock.

The lruvec lock protects folio->lru and the reclaim LRU lists whereas the deferred split queue lives on folio->_deferred_list in deferred_split_lru. The deferred_split_lru is a list_lru with its own per-memcg per-node lock.

There are already other existing callers that call folio_unqueue_deferred_split() without holding the lruvec lock (eg __folio_put(), shrink_folio_list(), __folio_migrate_mapping(), ...).

Add folio_batch_unqueue_deferred_split() and use it in move_folios_to_lru(), folio_batch_move_lru(), and folios_put_refs() to unqueue each batch of zero-refcount folios from the deferred split list after the lruvec lock is released and before mem_cgroup_uncharge_folios(). This helps shorten the lruvec lock's critical sections, which run with IRQs disabled and are highly contended.

Quoted 10 lines

---
This patch came out of a discussion about splitting underused mTHP folios [1], where queueing small mTHP folios on the deferred split list was found to regress Barry's fault microbenchmark. The cost turned out to be the unqueue running under the lruvec lock rather than contention on the deferred split queue itself. Testing this patch on top of the series in [1] with the same microbenchmark, queueing every 16k folio went from a 15% runtime cost to noise, with lruvec lock wait dropping from 491 seconds to 38 seconds (most of that contention then moves to the deferred split list_lru lock, but runtime still drops back to the unqueued baseline).

Quoted 2 lines

Looks great :) Could you also keep the benchmark results above the '---'? They'd be useful to have in the changelog, along with the test setup.

+1. I guess Joanne didn't include it in the changelog because the benchmark depends on her patchset, which hasn't been merged into mainline yet. But it should be fine to include the relevant details in the changelog since there is already a link to Joanne's patchset.

Joanne, could you also link your patchset and explain which patchset you used to collect the data?

Quoted 4 lines

With that, LGTM!

#3 lore raw
Quoted 46 lines · Barry Song
Quoted 41 lines · Lance Yang
Quoted 26 lines · Joanne Koong

move_folios_to_lru(), folio_batch_move_lru() and folios_put_refs() call folio_unqueue_deferred_split() on folios where the refcount has dropped to zero. In these paths, folio_unqueue_deferred_split() is called with the lruvec lock often held (move_folios_to_lru() always holds it and folio_batch_move_lru()/folios_put_refs() usually hold it for every folio in the batch except the first). However, it's not needed to hold the lruvec lock.

The lruvec lock protects folio->lru and the reclaim LRU lists whereas the deferred split queue lives on folio->_deferred_list in deferred_split_lru. The deferred_split_lru is a list_lru with its own per-memcg per-node lock.

There are already other existing callers that call folio_unqueue_deferred_split() without holding the lruvec lock (eg __folio_put(), shrink_folio_list(), __folio_migrate_mapping(), ...).

Add folio_batch_unqueue_deferred_split() and use it in move_folios_to_lru(), folio_batch_move_lru(), and folios_put_refs() to unqueue each batch of zero-refcount folios from the deferred split list after the lruvec lock is released and before mem_cgroup_uncharge_folios(). This helps shorten the lruvec lock's critical sections, which run with IRQs disabled and are highly contended.

Quoted 10 lines

---
This patch came out of a discussion about splitting underused mTHP folios [1], where queueing small mTHP folios on the deferred split list was found to regress Barry's fault microbenchmark. The cost turned out to be the unqueue running under the lruvec lock rather than contention on the deferred split queue itself. Testing this patch on top of the series in [1] with the same microbenchmark, queueing every 16k folio went from a 15% runtime cost to noise, with lruvec lock wait dropping from 491 seconds to 38 seconds (most of that contention then moves to the deferred split list_lru lock, but runtime still drops back to the unqueued baseline).

Quoted 2 lines

Looks great :) Could you also keep the benchmark results above the '---'? They'd be useful to have in the changelog, along with the test setup.

Quoted 7 lines

+1. I guess Joanne didn't include it in the changelog because the benchmark depends on her patchset, which hasn't been merged into mainline yet. But it should be fine to include the relevant details in the changelog since there is already a link to Joanne's patchset.

Joanne, could you also link your patchset and explain which patchset you used to collect the data?

Thanks for taking a look at this patch, Lance and Barry!

I will move this whole paragraph up into the commit message along with the link to that discussion thread.

The benchmarks were for the scenario where any folio down to 16k was queued onto the deferred split list to stress the unqueue path, but this will not be what happens in actuality (there will be an enforced minimum folio size of >= 2 MB for queueing folios onto the deferred split list).

Thanks,
Joanne

Quoted 5 lines
Quoted 5 lines

With that, LGTM!