On 1/31/22 12:35, Will McVicker wrote:
This fixes commit 54d516b1d62f ("mm/gup: small refactoring: simplify
try_grab_page()") which refactors try_grab_page() to call
try_grab_compound_head() with refs=1. The refactor commit is causing
pin_user_pages() to return -ENOMEM when we try to pin one user page that
is migratable and not in the movable zone. Previously, try_grab_page()
didn't check if the page was pinnable for FOLL_PIN. To match the same
functionality, this fix adds the check `refs > 1 &&` to skip the call to
is_pinnable_page().
That's a clear write-up of what you're seeing, what caused it, and how
you'd like to correct it. The previous code had a loophole, and you want
to keep that loophole. More below...
This issue is reproducible with the Pixel 6 on the 5.15 LTS kernel. Here
is the call stack to reproduce the -ENOMEM error:
...
Fixes: 54d516b1d62f ("mm/gup: small refactoring: simplify try_grab_page()")
Cc: John Hubbard <jhubbard@xxxxxxxxxx>
Cc: Minchan Kim <minchan@xxxxxxxxxx>
Signed-off-by: Will McVicker <willmcvicker@xxxxxxxxxx>
---
mm/gup.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mm/gup.c b/mm/gup.c
index f0af462ac1e2..0509c49c46a3 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -135,7 +135,7 @@ struct page *try_grab_compound_head(struct page *page,
* right zone, so fail and let the caller fall back to the slow
* path.
*/
- if (unlikely((flags & FOLL_LONGTERM) &&
+ if (refs > 1 && unlikely((flags & FOLL_LONGTERM) &&
!is_pinnable_page(page)))
return NULL;
...but are you really sure that this is the best way to "fix" the
problem? This trades correctness for "bug-for-bug compatibility" with
the previous code. It says, "it's OK to violate the pinnable and
longterm checks, as long as you do it one page at a time, rather than in
larger chunks.
Wouldn't it be better to try to fix up the calling code so that it's
not in violation of these zone rules?
thanks,
--
John Hubbard
NVIDIA
