On Tue, Dec 10, 2024 at 08:35:30PM +0100, David Hildenbrand wrote:
> On 10.12.24 18:24, Lorenzo Stoakes wrote:
> > If dup_mmap() encounters an issue, currently uprobe is able to access the
> > relevant mm via the reverse mapping (in build_map_info()), and if we are
> > very unlucky with a race window, observe invalid XA_ZERO_ENTRY state which
> > we establish as part of the fork error path.
> >
> > This occurs because uprobe_write_opcode() invokes anon_vma_prepare() which
> > in turn invokes find_mergeable_anon_vma() that uses a VMA iterator,
> > invoking vma_iter_load() which uses the advanced maple tree API and thus is
> > able to observe XA_ZERO_ENTRY entries added to dup_mmap() in commit
> > d24062914837 ("fork: use __mt_dup() to duplicate maple tree in
> > dup_mmap()").
> >
> > This change was made on the assumption that only process tear-down code
> > would actually observe (and make use of) these values. However this very
> > unlikely but still possible edge case with uprobes exists and unfortunately
> > does make these observable.
> >
> > The uprobe operation prevents races against the dup_mmap() operation via
> > the dup_mmap_sem semaphore, which is acquired via uprobe_start_dup_mmap()
> > and dropped via uprobe_end_dup_mmap(), and held across
> > register_for_each_vma() prior to invoking build_map_info() which does the
> > reverse mapping lookup.
> >
> > Currently these are acquired and dropped within dup_mmap(), which exposes
> > the race window prior to error handling in the invoking dup_mm() which
> > tears down the mm.
> >
> > We can avoid all this by just moving the invocation of
> > uprobe_start_dup_mmap() and uprobe_end_dup_mmap() up a level to dup_mm()
> > and only release this lock once the dup_mmap() operation succeeds or clean
> > up is done.
>
> What I understand is: we need to perform the uprobe_end_dup_mmap() after the
> mmput().
Ack yes.
>
> I assume/hope that we cannot see another mmget() before we return here. In
> that case, this LGTM.
We are dealing with a tiny time window and brief rmap availability, so it's hard
to say that's impossible. You also have to have failed to allocate really very
small amounts of memory, so we are talking lottery odds for this to even happen
in the first instance :)
I mean the syzkaller report took a year or so to hit it, and had to do
fault injection to do so.
Also it's not impossible that there are other means of accessing the mm
contianing XA_ZERO_ENTRY items through other means (I believe Liam was looking
into this).
However this patch is intended to at least eliminate the most proximate obvious
case with as simple a code change as possible.
Ideally we'd somehow mark the mm as being inaccessible somehow, but MMF_ flags
are out, and the obvious one to extend to mean this here, MMF_UNSTABLE, may
interact with oomk logic in some horrid way.
>
> --
> Cheers,
>
> David / dhildenb
>
So overall this patch is a relatively benign attempt to deal with the most
obvious issue with no apparent cost, but doesn't really rule out the need
to do more going forward...
