Date:   Fri, 22 Sep 2023 12:19:19 -0400
From:   "Liam R. Howlett" <Liam.Howlett@...cle.com>
To:     Jann Horn <jannh@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        kernel list <linux-kernel@...r.kernel.org>,
        Linux-MM <linux-mm@...ck.org>
Cc:     Lorenzo Stoakes <lstoakes@...il.com>,
        Vlastimil Babka <vbabka@...e.cz>
Subject: Re: maple tree change made it possible for VMA iteration to see same
 VMA twice due to late vma_merge() failure

* Liam R. Howlett <Liam.Howlett@...cle.com> [230816 15:18]:
> * Jann Horn <jannh@...gle.com> [230816 13:13]:
> > On Wed, Aug 16, 2023 at 6:18 PM Liam R. Howlett <Liam.Howlett@...cle.com> wrote:
> > > * Jann Horn <jannh@...gle.com> [230815 15:37]:
> > > > commit 18b098af2890 ("vma_merge: set vma iterator to correct
> > > > position.") added a vma_prev(vmi) call to vma_merge() at a point where
> > > > it's still possible to bail out. My understanding is that this moves
> > > > the VMA iterator back by one VMA.
> > > >
> > > > If you patch some extra logging into the kernel and inject a fake
> > > > out-of-memory error at the vma_iter_prealloc() call in vma_split() (a
> > > > real out-of-memory error there is very unlikely to happen in practice,
> > > > I think - my understanding is that the kernel will basically kill
> > > > every process on the system except for init before it starts failing
> > > > GFP_KERNEL allocations that fit within a single slab, unless the
> > > > allocation uses GFP_ACCOUNT or stuff like that, which the maple tree
> > > > doesn't):
> > [...]
> > > > then you'll get this fun log output, showing that the same VMA
> > > > (ffff88810c0b5e00) was visited by two iterations of the VMA iteration
> > > > loop, and on the second iteration, prev==vma:
> > > >
> > > > [  326.765586] userfaultfd_register: begin vma iteration
> > > > [  326.766985] userfaultfd_register: prev=ffff88810c0b5ef0,
> > > > vma=ffff88810c0b5e00 (0000000000101000-0000000000102000)
> > > > [  326.768786] userfaultfd_register: vma_merge returned 0000000000000000
> > > > [  326.769898] userfaultfd_register: prev=ffff88810c0b5e00,
> > > > vma=ffff88810c0b5e00 (0000000000101000-0000000000102000)
> > > >
> > > > I don't know if this can lead to anything bad but it seems pretty
> > > > clearly unintended?
> > >
> > > Yes, unintended.
> > >
> > > So we are running out of memory, but since vma_merge() doesn't
> > > differentiate between failure and 'nothing to merge', we end up in a
> > > situation that we will revisit the same VMA.
> > >
> > > I've been thinking about a way to work this into the interface and I
> > > don't see a clean way because we (could) do different things before the
> > > call depending on the situation.
> > >
> > > I think we need to undo any vma iterator changes in the failure
> > > scenarios if there is a chance of the iterator continuing to be used,
> > > which is probably not limited to just this case.
> > 
> > I don't fully understand the maple tree interface - in the specific
> > case of vma_merge(), could you move the vma_prev() call down below the
> > point of no return, after vma_iter_prealloc()? Or does
> > vma_iter_prealloc() require that the iterator is already in the insert
> > position?
> 
> Yes, but maybe it shouldn't.  I detect a write going beyond the end of a
> node and take corrective action, but not to the front of a node.
> 
> If I change the internal code to figure out the preallocations without
> being pointed at the insert location, I still cannot take corrective
> action on failure since I don't know where I should have been within the
> tree structure, that is, I have lost the original range.
> 
> I'm still looking at this, but I'm wondering if I should change my
> interface for preallocations so I can handle this internally.  That
> would be a bigger change.
> 
> > 
> > > I will audit these areas and CC you on the result.

Looking at this, I think it's best to make a label and undo the
vma_prev() with a vma_next() - at least for now.

I'm also reading this for the error path on dup_anon_vma() failure, and
it appears to also have an issue which I'd like to point out here before
I send the fix for the first issue.

-----------
                vma_start_write(next);                                                                                  
                remove = next;                          /* case 1 */                                                    
                vma_end = next->vm_end;                                                                                 
                err = dup_anon_vma(prev, next);                                                                         
                if (curr) {                             /* case 6 */                                                    
                        vma_start_write(curr);                                                                          
                        remove = curr;                                                                                  
                        remove2 = next;                                                                                 
                        if (!next->anon_vma)                                                                            
                                err = dup_anon_vma(prev, curr);  
-----------

Since dup_anon_vma() can fail, I think here in case 6 we could overwrite
the failure.

That is, we will fail to clone the anon vma and mask the failure if we
are running case 6 with an anon in next.  Once the first dup_anon_vma()
returns error, the next call to clone curr vma may return 0 if there is
no anon vma (this, I think _must_ be the case). Then we are in a
situation where we will be removing next and expanding prev over curr
and next, but have not dup'ed the anon vma from next.


Thanks,
Liam

Powered by blists - more mailing lists