Re: mm: possible deadlock in mm_take_all_locks

From: Kirill A. Shutemov
Date: Sun Jan 10 2016 - 15:39:16 EST


On Sun, Jan 10, 2016 at 09:05:32AM +0100, Dmitry Vyukov wrote:
> On Sat, Jan 9, 2016 at 12:23 AM, Kirill A. Shutemov
> <kirill@xxxxxxxxxxxxx> wrote:
> > On Fri, Jan 08, 2016 at 05:58:33PM +0100, Dmitry Vyukov wrote:
> >> Hello,
> >>
> >> I've hit the following deadlock warning while running syzkaller fuzzer
> >> on commit b06f3a168cdcd80026276898fd1fee443ef25743. As far as I
> >> understand this is a false positive, because both call stacks are
> >> protected by mm_all_locks_mutex.
> >
> > +Michal
> >
> > I don't think it's false positive.
> >
> > The reason we don't care about order of taking i_mmap_rwsem is that we
> > never takes i_mmap_rwsem under other i_mmap_rwsem, but that's not true for
> > i_mmap_rwsem vs. hugetlbfs_i_mmap_rwsem_key. That's why we have the
> > annotation in the first place.
> >
> > See commit b610ded71918 ("hugetlb: fix lockdep splat caused by pmd
> > sharing").
>
> Description of b610ded71918 suggests that that code takes hugetlb
> mutex first and them normal page mutex. In this patch you take them in
> the opposite order: normal mutex, then hugetlb mutex. Won't this patch
> only increase probability of deadlocks? Shouldn't you take them in the
> opposite order?

You are right. I got it wrong. Conditions should be reversed.

The comment around hugetlbfs_i_mmap_rwsem_key definition is somewhat
confusing:

"This needs an annotation because huge_pmd_share() does an allocation
under i_mmap_rwsem."

I read this as we do hugetlb allocation when i_mmap_rwsem already taken
and made locking order respectively. I guess i_mmap_rwsem should be
replaced with hugetlbfs_i_mmap_rwsem_key in the comment.

--
Kirill A. Shutemov