[PATCH] mm/mmu_notifier: Fix mmget() assert in __mmu_interval_notifier_insert

[PATCH] mm/mmu_notifier: Fix mmget() assert in __mmu_interval_notifier_insert

[Jann Horn]

The comment talks about having to hold mmget() (which means mm_users), but
the actual check is on mm_count (which would be mmgrab()).

Given that MMU notifiers are torn down in
mmput() -> __mmput() -> exit_mmap() -> mmu_notifier_release(),
I believe that the comment is correct and the check should be on
mm->mm_users. Fix it up accordingly.

Fixes: 99cb252f5e68 ("mm/mmu_notifier: add an interval tree notifier")
Signed-off-by: Jann Horn <jannh@google.com>
---
Can someone please double-check this? I'm like 90% sure that I fixed
this the right way around, but it'd be good if someone more familiar
with mmu notifiers could confirm.

 mm/mmu_notifier.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/mmu_notifier.c b/mm/mmu_notifier.c
index 4fc918163dd3..5654dd19addc 100644
--- a/mm/mmu_notifier.c
+++ b/mm/mmu_notifier.c
@@ -913,7 +913,7 @@ static int __mmu_interval_notifier_insert(
 		return -EOVERFLOW;
 
 	/* Must call with a mmget() held */
-	if (WARN_ON(atomic_read(&mm->mm_count) <= 0))
+	if (WARN_ON(atomic_read(&mm->mm_users) <= 0))
 		return -EINVAL;
 
 	/* pairs with mmdrop in mmu_interval_notifier_remove() */

base-commit: 15bc20c6af4ceee97a1f90b43c0e386643c071b4
-- 
2.28.0.402.g5ffc5be6b7-goog

Re: [PATCH] mm/mmu_notifier: Fix mmget() assert in __mmu_interval_notifier_insert

[Jason Gunthorpe]

On Tue, Sep 01, 2020 at 02:01:43AM +0200, Jann Horn wrote:
> The comment talks about having to hold mmget() (which means mm_users), but
> the actual check is on mm_count (which would be mmgrab()).
> 
> Given that MMU notifiers are torn down in
> mmput() -> __mmput() -> exit_mmap() -> mmu_notifier_release(),
> I believe that the comment is correct and the check should be on
> mm->mm_users. Fix it up accordingly.
> 
> Fixes: 99cb252f5e68 ("mm/mmu_notifier: add an interval tree notifier")
> Signed-off-by: Jann Horn <jannh@google.com>
> ---
> Can someone please double-check this? I'm like 90% sure that I fixed
> this the right way around, but it'd be good if someone more familiar
> with mmu notifiers could confirm.

Hmm.. I added the comment but copied the expression from older code.

The intent is certainly that the caller must have a mmget, this
function must not race with __mmu_notifier_release() and the mmget is
what prevents that.

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason