Re: [PATCH 1/1] mm: fix vma->anon_name memory leak for anonymous shmem VMAs

[Suren Baghdasaryan]

On Wed, Jan 4, 2023 at 10:24 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Wed, Jan 4, 2023 at 1:04 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >
> > On 03.01.23 20:53, Suren Baghdasaryan wrote:
> > > On Mon, Jan 2, 2023 at 4:00 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >>
> > >> On 28.12.22 20:42, Suren Baghdasaryan wrote:
> > >>> free_anon_vma_name() is missing a check for anonymous shmem VMA which
> > >>> leads to a memory leak due to refcount not being dropped. Fix this by
> > >>> adding the missing check.
> > >>>
> > >>> Fixes: d09e8ca6cb93 ("mm: anonymous shared memory naming")
> > >>> Reported-by: syzbot+91edf9178386a07d06a7@xxxxxxxxxxxxxxxxxxxxxxxxx
> > >>> Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > >>> ---
> > >>>    include/linux/mm_inline.h | 2 +-
> > >>>    1 file changed, 1 insertion(+), 1 deletion(-)
> > >>>
> > >>> diff --git a/include/linux/mm_inline.h b/include/linux/mm_inline.h
> > >>> index e8ed225d8f7c..d650ca2c5d29 100644
> > >>> --- a/include/linux/mm_inline.h
> > >>> +++ b/include/linux/mm_inline.h
> > >>> @@ -413,7 +413,7 @@ static inline void free_anon_vma_name(struct vm_area_struct *vma)
> > >>>         * Not using anon_vma_name because it generates a warning if mmap_lock
> > >>>         * is not held, which might be the case here.
> > >>>         */
> > >>> -     if (!vma->vm_file)
> > >>> +     if (!vma->vm_file || vma_is_anon_shmem(vma))
> > >>>                anon_vma_name_put(vma->anon_name);
> > >>
> > >> Wouldn't it be me more consistent to check for "vma->anon_name"?
> > >>
> > >> That's what dup_anon_vma_name() checks. And it's safe now because
> > >> anon_name is no longer overloaded in vm_area_struct.
> > >
> > > Thanks for the suggestion, David. Yes, with the recent change that
> > > does not overload anon_name, checking for "vma->anon_name" would be
> > > simpler. I think we can also drop anon_vma_name() function now
> > > (https://elixir.bootlin.com/linux/v6.2-rc2/source/mm/madvise.c#L94)
> > > since vma->anon_name does not depend on vma->vm_file anymore, remove
> > > the last part of this comment:
> > > https://elixir.bootlin.com/linux/v6.2-rc2/source/include/linux/mm_types.h#L584
> > > and use vma->anon_name directly going forward. If all that sounds
> > > good, I'll post a separate patch implementing all these changes.
> > > So, for this patch I would suggest keeping it as is because
> > > functionally it is correct and will change this check along with other
> > > corrections I mentioned above in a separate patch. Does that sound
> > > good?
> >
> > Works for me.
> >
> > Acked-by: David Hildenbrand <david@xxxxxxxxxx>
>
> Thank you! Will post the followup cleanup patch shorly.

Andrew, I posted v2 of this patch at
https://lore.kernel.org/all/20230105000241.1450843-1-surenb@xxxxxxxxxx/.
Please replace the original one in your tree. I had to follow David's
original suggestion instead of my planned cleanup because removing
anon_vma_name() would require us to add #ifdef CONFIG_ANON_VMA_NAME
everywhere we use vma->anon_name, which is quite ugly.
Thanks,
Suren.

>
> >
> > for this one, as it fixes the issue.
> >
> > --
> > Thanks,
> >
> > David / dhildenb
> >