Re: [PATCH] mm/mmap/vma_merge: always check invariants

[Lorenzo Stoakes <lstoakes@gmail.com>]

On Thu, May 11, 2023 at 11:08:41AM -0700, Mike Rapoport wrote:
> (adding Peter)
>
> On Wed, May 10, 2023 at 09:26:10AM -0700, Lorenzo Stoakes wrote:
> > On Wed, May 10, 2023 at 05:17:49PM +0100, Mark Rutland wrote:
> > > On Wed, May 10, 2023 at 09:04:44AM -0700, Lorenzo Stoakes wrote:
> > > > On Wed, May 10, 2023 at 03:15:51PM +0100, Mark Rutland wrote:
> > > > > Hi,
> > > > >
> > > > > On Sun, Apr 30, 2023 at 09:19:17PM +0100, Lorenzo Stoakes wrote:
> > > > > > We may still have inconsistent input parameters even if we choose not to
> > > > > > merge and the vma_merge() invariant checks are useful for checking this
> > > > > > with no production runtime cost (these are only relevant when
> > > > > > CONFIG_DEBUG_VM is specified).
> > > > > >
> > > > > > Therefore, perform these checks regardless of whether we merge.
> > > > > >
> > > > > > This is relevant, as a recent issue (addressed in commit "mm/mempolicy:
> > > > > > Correctly update prev when policy is equal on mbind") in the mbind logic
> > > > > > was only picked up in the 6.2.y stable branch where these assertions are
> > > > > > performed prior to determining mergeability.
> > > > > >
> > > > > > Had this remained the same in mainline this issue may have been picked up
> > > > > > faster, so moving forward let's always check them.
> > > > > >
> > > > > > Signed-off-by: Lorenzo Stoakes <lstoakes@gmail.com>
> > > > > > ---
> > > > > >  mm/mmap.c | 10 +++++-----
> > > > > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > > > > >
> > > > > > diff --git a/mm/mmap.c b/mm/mmap.c
> > > > > > index 5522130ae606..13678edaa22c 100644
> > > > > > --- a/mm/mmap.c
> > > > > > +++ b/mm/mmap.c
> > > > > > @@ -960,17 +960,17 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
> > > > > >  		merge_next = true;
> > > > > >  	}
> > > > > >
> > > > > > +	/* Verify some invariant that must be enforced by the caller. */
> > > > > > +	VM_WARN_ON(prev && addr <= prev->vm_start);
> > > > > > +	VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
> > > > > > +	VM_WARN_ON(addr >= end);
> > > > > > +
> > > > >
> > > > > I'm seeing this fire a lot when fuzzing v6.4-rc1 on arm64 using Syzkaller.
> > > > >
> > > >
> > > > Thanks, from the line I suspect addr != curr->vm_start, but need to look
> > > > into the repro, at lsf/mm so a bit time lagged :)
> > >
> > > No problem; FWIW I can confirm your theory, the reproducer is causing:
> > >
> > > 	addr > curr->vm_start
> > >
> > > ... confirmed the the following hack, log below.
> >
> > Awesome thanks for that! Just been firing up qemu to do this.
> >
> > Cases 5-8 should really have addr == curr->vm_start, I wonder if it's
> > another case but curr is being set incorrectly, it should in theory not be
> > the case.
>
> AFAIU, it's a case of "adjust vma, but don't merge, because prev is not
> compatible". Looks like uffd first attempts to merge compatible the newly
> registered range with adjacent vmas relying on that there won't be no merge
> when addr != curr->vm_start and only after the merge attempt it splits the
> edges.
>
> I think that moving the split in fs/userfaultfd.c:1495 (as of v6.4-rc1)
> before vma_merge() will be the right fix.
>

Indeed it was, patch at
https://lore.kernel.org/all/20230514172731.134188-1-lstoakes@gmail.com/

[snip]