On Wed, Apr 24, 2019 at 09:49:58PM +0200, Paul Chaignon wrote:
> In case of a null check on a pointer inside a subprog, we should mark all
> registers with this pointer as either safe or unknown, in both the current
> and previous frames. Currently, only spilled registers and registers in
> the current frame are marked. Packet bound checks in subprogs have the
> same issue. The first patch fixes it to mark registers in previous frames
> as well.
>
> A good reproducer for null checks looks as follow:
>
> 1: ptr = bpf_map_lookup_elem(map, &key);
> 2: ret = subprog(ptr) {
> 3: return ptr != NULL;
> 4: }
> 5: if (ret)
> 6: value = *ptr;
>
> With the above, the verifier will complain on line 6 because it sees ptr
> as map_value_or_null despite the null check in subprog 1. The second
> patch implements the above as a new test case.
>
> Note that this patch fixes another resulting bug when using
> bpf_sk_release():
>
> 1: sk = bpf_sk_lookup_tcp(...);
> 2: subprog(sk) {
> 3: if (sk)
> 4: bpf_sk_release(sk);
> 5: }
> 6: if (!sk)
> 7: return 0;
> 8: return 1;
>
> In the above, mark_ptr_or_null_regs will warn on line 6 because it will
> try to free the reference state, even though it was already freed on
> line 3.
>
> Changelogs:
> Changes in v3:
> - Fix same issue in find_good_pkt_pointers().
> - Add test case for find_good_pkt_pointers() issue.
> - Change title to account for the above change. Old title was
> "bpf: mark registers as safe or unknown in all frames".
> - Refactor find_good_pkt_pointers and mark_ptr_or_null_regs.
> - I did not keep Yonghong's ack because of the above changes.
> Changes in v2:
> - Fix example codes in commit message.
Applied to bpf tree, thanks!
