[v3,1/9] mm: optimize do_wp_page() for exclusive pages in the swapcache

Commit Message

David Hildenbrand Jan. 31, 2022, 4:29 p.m. UTC

Liang Zhang reported [1] that the current COW logic in do_wp_page() is
sub-optimal when it comes to swap+read fault+write fault of anonymous
pages that have a single user, visible via a performance degradation in
the redis benchmark. Something similar was previously reported [2] by
Nadav with a simple reproducer.

After we put an anon page into the swapcache and unmapped it from a single
process, that process might read that page again and refault it read-only.
If that process then writes to that page, the process is actually the
exclusive user of the page, however, the COW logic in do_co_page() won't be
able to reuse it due to the additional reference from the swapcache.

Let's optimize for pages that have been added to the swapcache but only
have an exclusive user. Try removing the swapcache reference if there is
hope that we're the exclusive user.

We will fail removing the swapcache reference in two scenarios:
(1) There are additional swap entries referencing the page: copying
    instead of reusing is the right thing to do.
(2) The page is under writeback: theoretically we might be able to reuse
    in some cases, however, we cannot remove the additional reference
    and will have to copy.

Note that we'll only try removing the page from the swapcache when it's
highly likely that we'll be the exclusive owner after removing the
page from the swapache. As we're about to map that page writable and
redirty it, that should not affect reclaim but is rather the right thing
to do.

Further, we might have additional references from the LRU pagevecs,
which will force us to copy instead of being able to reuse. We'll try
handling such references for some scenarios next. Concurrent writeback
cannot be handled easily and we'll always have to copy.

While at it, remove the superfluous page_mapcount() check: it's
implicitly covered by the page_count() for ordinary anon pages.

[1] https://lkml.kernel.org/r/[e-mail]
[2] https://lkml.kernel.org/r/[e-mail]

Reported-by: Liang Zhang <[e-mail]>
Reported-by: Nadav Amit <[e-mail]>
Reviewed-by: Matthew Wilcox (Oracle) <[e-mail]>
Acked-by: Vlastimil Babka <[e-mail]>
Signed-off-by: David Hildenbrand <[e-mail]>
---
 mm/memory.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)

Patch

diff --git a/mm/memory.c b/mm/memory.c
index c125c4969913..bcd3b7c50891 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3291,19 +3291,27 @@  static vm_fault_t do_wp_page(struct vm_fault *vmf)
 	if (PageAnon(vmf->page)) {
 		struct page *page = vmf->page;
 
-		/* PageKsm() doesn't necessarily raise the page refcount */
-		if (PageKsm(page) || page_count(page) != 1)
+		/*
+		 * We have to verify under page lock: these early checks are
+		 * just an optimization to avoid locking the page and freeing
+		 * the swapcache if there is little hope that we can reuse.
+		 *
+		 * PageKsm() doesn't necessarily raise the page refcount.
+		 */
+		if (PageKsm(page) || page_count(page) > 1 + PageSwapCache(page))
 			goto copy;
 		if (!trylock_page(page))
 			goto copy;
-		if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
+		if (PageSwapCache(page))
+			try_to_free_swap(page);
+		if (PageKsm(page) || page_count(page) != 1) {
 			unlock_page(page);
 			goto copy;
 		}
 		/*
-		 * Ok, we've got the only map reference, and the only
-		 * page count reference, and the page is locked,
-		 * it's dark out, and we're wearing sunglasses. Hit it.
+		 * Ok, we've got the only page reference from our mapping
+		 * and the page is locked, it's dark out, and we're wearing
+		 * sunglasses. Hit it.
 		 */
 		unlock_page(page);
 		wp_page_reuse(vmf);