On Thu, Mar 02, 2023 at 06:41:55PM -0500, Dave Marchevsky wrote:
> On 3/2/23 6:29 PM, David Vernet wrote:
> > On Thu, Mar 02, 2023 at 03:23:22PM -0800, Alexei Starovoitov wrote:
> >> On Thu, Mar 2, 2023 at 3:19 PM Dave Marchevsky <davemarchevsky@xxxxxx> wrote:
> >>>
> >>> --- a/tools/testing/selftests/bpf/progs/rbtree_fail.c
> >>> +++ b/tools/testing/selftests/bpf/progs/rbtree_fail.c
> >>> @@ -232,8 +232,9 @@ long rbtree_api_first_release_unlock_escape(void *ctx)
> >>>
> >>> bpf_spin_lock(&glock);
> >>> res = bpf_rbtree_first(&groot);
> >>> - if (res)
> >>> - n = container_of(res, struct node_data, node);
> >>> + if (!res)
> >>> + return -1;
> >>
> >> The verifier cannot be ok with this return... I hope...
> >
> > This is a negative testcase which correctly fails, though the error
> > message wasn't what I was expecting to see:
> >
> > __failure __msg("rbtree_remove node input must be non-owning ref")
> >
> > Something about the lock still being held seems much more intuitive.
> >
>
> It's necessary to call bpf_rbtree_remove w/ lock held. This test expects
> to fail because non-owning ref "n" is clobbered after the critical
> section where it's returned by bpf_rbtree_first ends.
Oh, I see. I think that would arguably be a bit more clear if we added a
bpf_spin_unlock() to that return case then. Ideally for a negative test
we can keep the number of bugs being tested to 1. I assume that was
Alexei's point, which clearly went over my head.
>
> >>
> >>> + n = container_of(res, struct node_data, node);
> >>> bpf_spin_unlock(&glock);
