Re: [PATCH v4] mm: use per_vma lock for MADV_DONTNEED

[Barry Song <21cnbao@gmail.com>]

On Wed, Jun 18, 2025 at 6:33 PM Lorenzo Stoakes
<lorenzo.stoakes@oracle.com> wrote:
>
> On Wed, Jun 18, 2025 at 06:11:26PM +0800, Barry Song wrote:
> [sip]
> > > ----8<----
> > > From 1ffcaea75ebdaffe15805386f6d7733883d265a5 Mon Sep 17 00:00:00 2001
> > > From: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> > > Date: Tue, 17 Jun 2025 14:35:13 +0100
> > > Subject: [PATCH] mm/madvise: avoid any chance of uninitialised pointer deref
> > >
> > > If we were to extend madvise() to support more operations under VMA lock,
> > > we could potentially dereference prev to uninitialised state in
> > > madvise_update_vma().
> > >
> > > Avoid this by explicitly setting prev to vma before invoking the visit()
> > > function.
> > >
> > > This has no impact on behaviour, as all visitors compatible with a VMA lock
> > > do not require prev to be set to the previous VMA and at any rate we only
> > > examine a single VMA in VMA lock mode.
> > >
> > > Reported-by: Lance Yang <ioworker0@gmail.com>
> > > Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> > > ---
> > >  mm/madvise.c | 3 +++
> > >  1 file changed, 3 insertions(+)
> > >
> > > diff --git a/mm/madvise.c b/mm/madvise.c
> > > index efe5d64e1175..0970623a0e98 100644
> > > --- a/mm/madvise.c
> > > +++ b/mm/madvise.c
> > > @@ -1333,6 +1333,8 @@ static int madvise_vma_behavior(struct vm_area_struct *vma,
> > >                 return madvise_guard_remove(vma, prev, start, end);
> > >         }
> > >
> > > +       /* We cannot provide prev in this lock mode. */
> > > +       VM_WARN_ON_ONCE(arg->lock_mode == MADVISE_VMA_READ_LOCK);
> >
> > Thanks, Lorenzo.
> > Do we even reach this point for MADVISE_MMAP_READ_LOCK cases?
> > madvise_update_vma() attempts to merge or split VMAs—wouldn't that be
> > a scenario that requires a write lock?
>
> Well we're relying on happening to reach here with the correct lock afaict.
>
> I'm going to be doing some follow-up series to clean all this up!
>
> I'd rather keep this in here for now just to ensure we don't miss some stupidity
> here.

I have no objection to keeping this as-is—just curious if using
VM_WARN_ON_ONCE(arg->lock_mode != MADVISE_MMAP_WRITE_LOCK)
would be more accurate.

In any case, your cleanup series will address this, so it's probably
not something we need to handle right now.

>
> Thanks!
>
> >
> > The prerequisite for using a VMA read lock is that the operation must
> > be safe under an mmap read lock as well.
> >
> > >         anon_name = anon_vma_name(vma);
> > >         anon_vma_name_get(anon_name);
> > >         error = madvise_update_vma(vma, prev, start, end, new_flags,
> > > @@ -1549,6 +1551,7 @@ int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
> > >         if (madv_behavior && madv_behavior->lock_mode == MADVISE_VMA_READ_LOCK) {
> > >                 vma = try_vma_read_lock(mm, madv_behavior, start, end);
> > >                 if (vma) {
> > > +                       prev = vma;
> > >                         error = visit(vma, &prev, start, end, arg);
> > >                         vma_end_read(vma);
> > >                         return error;
> > > --
> > > 2.49.0
> >

Thanks
Barry