Re: UBSAN: array-index-out-of-bounds in kernel/bpf/arraymap.c:177

[Qian Cai]

On Tue, May 19, 2020 at 3:30 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Tue, May 19, 2020 at 8:00 AM Qian Cai <cai@xxxxxx> wrote:
> >
> > On Mon, May 18, 2020 at 8:25 PM Andrii Nakryiko
> > <andrii.nakryiko@xxxxxxxxx> wrote:
> > >
> > > On Mon, May 18, 2020 at 5:09 PM Qian Cai <cai@xxxxxx> wrote:
> > > >
> > > > On Mon, May 18, 2020 at 7:55 PM Andrii Nakryiko
> > > > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > > >
> > > > > On Sun, May 17, 2020 at 7:45 PM Qian Cai <cai@xxxxxx> wrote:
> > > > > >
> > > > > > With Clang 9.0.1,
> > > > > >
> > > > > > return array->value + array->elem_size * (index & array->index_mask);
> > > > > >
> > > > > > but array->value is,
> > > > > >
> > > > > > char value[0] __aligned(8);
> > > > >
> > > > > This, and ptrs and pptrs, should be flexible arrays. But they are in a
> > > > > union, and unions don't support flexible arrays. Putting each of them
> > > > > into anonymous struct field also doesn't work:
> > > > >
> > > > > /data/users/andriin/linux/include/linux/bpf.h:820:18: error: flexible
> > > > > array member in a struct with no named members
> > > > >    struct { void *ptrs[] __aligned(8); };
> > > > >
> > > > > So it probably has to stay this way. Is there a way to silence UBSAN
> > > > > for this particular case?
> > > >
> > > > I am not aware of any way to disable a particular function in UBSAN
> > > > except for the whole file in kernel/bpf/Makefile,
> > > >
> > > > UBSAN_SANITIZE_arraymap.o := n
> > > >
> > > > If there is no better way to do it, I'll send a patch for it.
> > >
> > >
> > > That's probably going to be too drastic, we still would want to
> > > validate the rest of arraymap.c code, probably. Not sure, maybe
> > > someone else has better ideas.
> >
> > This works although it might makes sense to create a pair of
> > ubsan_disable_current()/ubsan_enable_current() for it.
> >
> > diff --git a/kernel/bpf/arraymap.c b/kernel/bpf/arraymap.c
> > index 11584618e861..6415b089725e 100644
> > --- a/kernel/bpf/arraymap.c
> > +++ b/kernel/bpf/arraymap.c
> > @@ -170,11 +170,16 @@ static void *array_map_lookup_elem(struct
> > bpf_map *map, void *key)
> >  {
> >         struct bpf_array *array = container_of(map, struct bpf_array, map);
> >         u32 index = *(u32 *)key;
> > +       void *elem;
> >
> >         if (unlikely(index >= array->map.max_entries))
> >                 return NULL;
> >
> > -       return array->value + array->elem_size * (index & array->index_mask);
> > +       current->in_ubsan++;
> > +       elem = array->value + array->elem_size * (index & array->index_mask);
> > +       current->in_ubsan--;
>
> This is an unnecessary performance hit for silencing what is clearly a
> false positive. I'm not sure that's the right solution here. It seems
> like something that's lacking on the tooling side instead. C language
> doesn't allow to express the intent here using flexible array
> approach. That doesn't mean that what we are doing here is wrong or
> undefined.

Oh, so you worry about this ++ and -- hurt the performance? If so, how
about this?

ubsan_disable_current();
elem = array->value + array->elem_size * (index & array->index_mask);
ubsan_enable_current();

#ifdef UBSAN
ubsan_disable_current()
{
      current->in_ubsan++;
}
#else
ubsan_disable_current() {}
#endif

etc

Production kernel would normally have UBSAN=n, so it is an noop.

Leaving this false positive unsilenced may also waste many people's
time over and over again, and increase the noisy level. Especially, it
seems this is one-off (not seen other parts of kernel doing like this)
and rather expensive to silence it in the UBSAN or/and compilers.