Re: stack access issue. Re: [syzbot] [bpf?] UBSAN: array-index-out-of-bounds in check_stack_range_initialized

[Andrei Matei <andreimatei1@xxxxxxxxx>]

Fixing in https://lore.kernel.org/bpf/20240324230323.1097685-1-andreimatei1@xxxxxxxxx/

FWIW, I managed to decode the BPF program that syzkaller used:

0: (18) r0 = 0x0
2: (18) r1 = map[id:4]
4: (b7) r8 = 0
5: (7b) *(u64 *)(r10 -8) = r8
6: (bf) r2 = r10
7: (07) r2 += -8
8: (b7) r3 = 8
9: (b7) r4 = 0
10: (85) call bloom_map_peek_elem#322320
11: (95) exit

Where the map is a bloom filter (as Alexei somehow already knew on the patch
thread) with a humongous value size.

4: type 30  flags 0x0
        key 0B  value 2147483649B  max_entries 255  memlock 720B

On Sat, Mar 23, 2024 at 10:55 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Sat, Mar 23, 2024 at 7:12 PM Andrei Matei <andreimatei1@xxxxxxxxx> wrote:
> >
> > On Sat, Mar 23, 2024 at 8:52 PM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > On Sat, Mar 23, 2024 at 5:50 PM Andrei Matei <andreimatei1@xxxxxxxxx> wrote:
> > > >
> > > > + Edward
> > > >
> > > > On Thu, Mar 21, 2024 at 3:33 AM Alexei Starovoitov
> > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > >
> > > > > Hi Andrei,
> > > > >
> > > > > looks like the refactoring of stack access introduced a bug.
> > > > > See the reproducer below.
> > > > > positive offsets are not caught by check_stack_access_within_bounds().
> > > >
> > > > check_stack_access_within_bounds() tries to catch positive offsets;
> > > > It does: [1]
> > > >
> > > > err = check_stack_slot_within_bounds(env, min_off, state, type);
> > > > if (!err && max_off > 0)
> > > >   err = -EINVAL; /* out of stack access into non-negative offsets */
> > > >
> > > > Notice the max_off > 0 in there.
> > > > And we have various tests that seem to check that positive offsets are
> > > > rejected. Do you know what the bug is?
> > > > I'm thinking maybe there's some overflow going on, except that UBSAN
> > > > reported an index of -1 as being the problem.
> > > >
> > > > Edward, I see that you've been tickling the robot trying to narrow the issue;
> > > > perhaps you've figured it out?
> > > >
> > > > If the bug is not immediately apparent to anyone, I would really appreciate a
> > > > bit of tutoring around how to reproduce and get verifier logs.
> > >
> > > The repro is right there in the email I forwarded:
> > >
> > > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=15c38711180000
> >
> > I understand, but how does one go from this to either BPF assembly,
> > or to running it in such a way that you also get verifier logs?
>
> Adding logs to repro.c is too hard, but you can
> hack the kernel with printk-s.
>
> Like the following:
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index de7813947981..d158b83ed16c 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -7179,6 +7179,7 @@ static int check_stack_range_initialized(
>                         return -EFAULT;
>                 }
>
> +               printk("slot %d %d spi %d\n", slot, slot % BPF_REG_SIZE, spi);
>                 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
>
>
> shows that spi and slot get negative: -1, -2, ...