Re: [PATCH bpf v2] bpf: fix nested bpf tracepoints with per-cpu data

[Alexei Starovoitov]

On Thu, Jun 13, 2019 at 5:52 PM Matt Mullins <mmullins@xxxxxx> wrote:
>
> On Fri, 2019-06-14 at 00:47 +0200, Daniel Borkmann wrote:
> > On 06/12/2019 07:00 AM, Andrii Nakryiko wrote:
> > > On Tue, Jun 11, 2019 at 8:48 PM Matt Mullins <mmullins@xxxxxx> wrote:
> > > >
> > > > BPF_PROG_TYPE_RAW_TRACEPOINTs can be executed nested on the same CPU, as
> > > > they do not increment bpf_prog_active while executing.
> > > >
> > > > This enables three levels of nesting, to support
> > > >   - a kprobe or raw tp or perf event,
> > > >   - another one of the above that irq context happens to call, and
> > > >   - another one in nmi context
> > > > (at most one of which may be a kprobe or perf event).
> > > >
> > > > Fixes: 20b9d7ac4852 ("bpf: avoid excessive stack usage for perf_sample_data")
> >
> > Generally, looks good to me. Two things below:
> >
> > Nit, for stable, shouldn't fixes tag be c4f6699dfcb8 ("bpf: introduce BPF_RAW_TRACEPOINT")
> > instead of the one you currently have?
>
> Ah, yeah, that's probably more reasonable; I haven't managed to come up
> with a scenario where one could hit this without raw tracepoints.  I'll
> fix up the nits that've accumulated since v2.
>
> > One more question / clarification: we have __bpf_trace_run() vs trace_call_bpf().
> >
> > Only raw tracepoints can be nested since the rest has the bpf_prog_active per-CPU
> > counter via trace_call_bpf() and would bail out otherwise, iiuc. And raw ones use
> > the __bpf_trace_run() added in c4f6699dfcb8 ("bpf: introduce BPF_RAW_TRACEPOINT").
> >
> > 1) I tried to recall and find a rationale for mentioned trace_call_bpf() split in
> > the c4f6699dfcb8 log, but couldn't find any. Is the raison d'Ãtre purely because of
> > performance overhead (and desire to not miss events as a result of nesting)? (This
> > also means we're not protected by bpf_prog_active in all the map ops, of course.)
> > 2) Wouldn't this also mean that we only need to fix the raw tp programs via
> > get_bpf_raw_tp_regs() / put_bpf_raw_tp_regs() and won't need this duplication for
> > the rest which relies upon trace_call_bpf()? I'm probably missing something, but
> > given they have separate pt_regs there, how could they be affected then?
>
> For the pt_regs, you're correct: I only used get/put_raw_tp_regs for
> the _raw_tp variants.  However, consider the following nesting:
>
>                                     trace_nest_level raw_tp_nest_level
>   (kprobe) bpf_perf_event_output            1               0
>   (raw_tp) bpf_perf_event_output_raw_tp     2               1
>   (raw_tp) bpf_get_stackid_raw_tp           2               2
>
> I need to increment a nest level (and ideally increment it only once)
> between the kprobe and the first raw_tp, because they would otherwise
> share the struct perf_sample_data.  But I also need to increment a nest
> level between the two raw_tps, since they share the pt_regs -- I can't
> use trace_nest_level for everything because it's not used by
> get_stackid, and I can't use raw_tp_nest_level for everything because
> it's not incremented by kprobes.
>
> If raw tracepoints were to bump bpf_prog_active, then I could get away
> with just using that count in these callsites -- I'm reluctant to do
> that, though, since it would prevent kprobes from ever running inside a
> raw_tp.  I'd like to retain the ability to (e.g.)
>   trace.py -K htab_map_update_elem
> and get some stack traces from at least within raw tracepoints.
>
> That said, as I wrote up this example, bpf_trace_nest_level seems to be
> wildly misnamed; I should name those after the structure they're
> protecting...

I still don't get what's wrong with the previous approach.
Didn't I manage to convince both of you that perf_sample_data
inside bpf_perf_event_array doesn't have any issue that Daniel brought up?
I think this refcnting approach is inferior.