Re: [PATCH bpf-next v2] bpf: check flags in 'bpf_ringbuf_discard()' and 'bpf_ringbuf_submit()'

[Andrii Nakryiko]

On Sat, Apr 3, 2021 at 6:29 AM Pedro Tammela <pctammela@xxxxxxxxx> wrote:
>
> Em qua., 31 de mar. de 2021 às 04:02, Andrii Nakryiko
> <andrii.nakryiko@xxxxxxxxx> escreveu:
> >
> > On Tue, Mar 30, 2021 at 4:16 PM Alexei Starovoitov
> > <alexei.starovoitov@xxxxxxxxx> wrote:
> > >
> > > On Tue, Mar 30, 2021 at 3:54 PM Pedro Tammela <pctammela@xxxxxxxxx> wrote:
> > > >
> > > >  BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
> > > >  {
> > > > +       if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
> > > > +               return -EINVAL;
> > > > +
> > > >         bpf_ringbuf_commit(sample, flags, false /* discard */);
> > > > +
> > > >         return 0;
> > >
> > > I think ringbuf design was meant for bpf_ringbuf_submit to never fail.
> > > If we do flag validation it probably should be done at the verifier time.
> >
> > Oops, replied on another version already. But yes, BPF verifier relies
> > on it succeeding. I don't think we can do flags validation at BPF
> > verification time, though, because it is defined as non-const integer
> > and we do have valid cases where we dynamically determine whether to
> > FORCE_WAKEUP or NO_WAKEUP, based on application-driven criteria (e.g.,
> > amount of enqueued data).
>
> Then shouldn't we remove the flags check in 'bpf_ringbuf_output()'?

bpf_ringbuf_output() combines reserve + commit operations, so if it
performs checks before anything is reserved in ringbuf, it's ok for it
to fail and return error. So I don't see any problem there. But once
it internally reserves, it always proceeds to complete the commit.