Re: [PATCH v5 1/2] kasan: catch invalid free before SLUB reinitializes the object

[Andrey Konovalov <andreyknvl@gmail.com>]

On Thu, Aug 1, 2024 at 6:01 AM Jann Horn <jannh@google.com> wrote:
>
> > > @@ -503,15 +509,22 @@ bool __kasan_mempool_poison_object(void *ptr, unsigned long ip)
> > >                 kasan_poison(ptr, folio_size(folio), KASAN_PAGE_FREE, false);
> > >                 return true;
> > >         }
> > >
> > >         if (is_kfence_address(ptr))
> > >                 return false;
> > > +       if (!kasan_arch_is_ready())
> > > +               return true;
> >
> > Hm, I think we had a bug here: the function should return true in both
> > cases. This seems reasonable: if KASAN is not checking the object, the
> > caller can do whatever they want with it.
>
> But if the object is a kfence allocation, we maybe do want the caller
> to free it quickly so that kfence can catch potential UAF access? So
> "return false" in that case seems appropriate.

Return false would mean: allocation is buggy, do not use it and do not
free it (note that the return value meaning here is inverse compared
to the newly added check_slab_allocation()). And this doesn't seem
like something we want for KFENCE-managed objects. But regardless of
the return value here, the callers tend not to free these allocations
to the slab allocator, that's the point of mempools. So KFENCE won't
catch a UAF either way.

> Or maybe we don't
> because that makes the probability of catching an OOB access much
> lower if the mempool is going to always return non-kfence allocations
> as long as the pool isn't empty? Also I guess whether kfence vetoes
> reuse of kfence objects probably shouldn't depend on whether the
> kernel is built with KASAN... so I guess it would be more consistent
> to either put "return true" there or change the !KASAN stub of this to
> check for kfence objects or something like that? Honestly I think the
> latter would be most appropriate, though then maybe the hook shouldn't
> have "kasan" in its name...

Yeah, we could add some custom handling of mempool to KFENCE as well.
But that would be a separate effort.

> Either way, I agree that the current situation wrt mempools and kfence
> is inconsistent, but I think I should probably leave that as-is in my
> series for now, and the kfence mempool issue can be addressed
> separately afterwards? I also would like to avoid changing kfence
> behavior as part of this patch.

Sure, sounds good to me.

> If you want, I can add a comment above the "if (is_kfence_address())"
> that notes the inconsistency?

Up to you, I'll likely add a note to the bug tracker to fix this once
the patch lands anyway.

Thanks!