Re: kernel BUG at mm/page-writeback.c:LINE!

[Linus Torvalds <torvalds@linux-foundation.org>]

[-- Attachment #1: Type: text/plain, Size: 1644 bytes --]

On Tue, Jan 5, 2021 at 11:53 AM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> I took your "way to go" statement as an ack, and made it all be commit
> c2407cf7d22d ("mm: make wait_on_page_writeback() wait for multiple
> pending writebacks").

Oh, and Michael Larabel (of phoronix) reports that that one-liner does
something bad to a few PostgreSQL tests, on the order of 5-10%
regression on some machines (but apparently not others).

I suspect that's a sign of instability in the benchmark numbers, but
it probably also means that we have some silly condition where
multiple threads want to clean the same page.

I sent him a patch to try if it ends up being better to just not wake
things up early at all (instead of the "if" -> "while") conversion.
That trivial patch appended here in case anybody has comments.

Just the fact that that one-liner made a performance impact makes me
go "hmm", though. Michael didn't see the BUG_ON(), so it's presumably
some _other_ user of wait_on_page_writeback() than the
write_cache_pages() one that causes issues.

Anybody got any suspicions? Honestly, when working on the page wait
queues, I was working under the assumption that it's really just the
page lock that truly matters.

I'm thinking things like __filemap_fdatawait_range(), which doesn't
hold the page lock at all, so it's all kinds of non-serialized, and
could now be waiting for any number of IO's ro complete..

Oh well. This email doesn't really have a point, it's more of a
heads-up that that "wait to see one or multiple writebacks" thing
seems to matter more than I would have expected for some loads..

            Linus

[-- Attachment #2: patch --]
[-- Type: application/octet-stream, Size: 1292 bytes --]

 mm/filemap.c        | 6 ++++--
 mm/page-writeback.c | 2 +-
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 5c9d564317a5..031d19d0f4b2 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1052,14 +1052,16 @@ static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync,
 	if (!wake_page_match(wait_page, key))
 		return 0;
 
+	/* Stop walking if the bit was already taken */
+	if (test_bit(key->bit_nr, &key->page->flags))
+		return -1;
+
 	/*
 	 * If it's a lock handoff wait, we get the bit for it, and
 	 * stop walking (and do not wake it up) if we can't.
 	 */
 	flags = wait->flags;
 	if (flags & WQ_FLAG_EXCLUSIVE) {
-		if (test_bit(key->bit_nr, &key->page->flags))
-			return -1;
 		if (flags & WQ_FLAG_CUSTOM) {
 			if (test_and_set_bit(key->bit_nr, &key->page->flags))
 				return -1;
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index eb34d204d4ee..586042472ac9 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -2826,7 +2826,7 @@ EXPORT_SYMBOL(__test_set_page_writeback);
  */
 void wait_on_page_writeback(struct page *page)
 {
-	while (PageWriteback(page)) {
+	if (PageWriteback(page)) {
 		trace_wait_on_page_writeback(page, page_mapping(page));
 		wait_on_page_bit(page, PG_writeback);
 	}