Re: [PATCH] perf/bpf: Don't call bpf_overflow_handler() for tracing events

[Andrii Nakryiko]

On Fri, Jul 26, 2024 at 5:37 AM Kyle Huey <me@xxxxxxxxxxxx> wrote:
>
> On Fri, Jul 19, 2024 at 11:26 AM Andrii Nakryiko
> <andrii.nakryiko@xxxxxxxxx> wrote:
> >
> > On Tue, Jul 16, 2024 at 12:25 AM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
> > >
> > > On Mon, Jul 15, 2024 at 09:48:58AM -0700, Kyle Huey wrote:
> > > > On Mon, Jul 15, 2024 at 9:30 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, Jul 15, 2024 at 08:19:44AM -0700, Kyle Huey wrote:
> > > > >
> > > > > > I think this would probably work but stealing the bit seems far more
> > > > > > complicated than just gating on perf_event_is_tracing().
> > > > >
> > > > > perf_event_is_tracing() is something like 3 branches. It is not a simple
> > > > > conditional. Combined with that re-load and the wrong return value, this
> > > > > all wants a cleanup.
> > > > >
> > > > > Using that LSB works, it's just that the code aint pretty.
> > > >
> > > > Maybe we could gate on !event->tp_event instead. Somebody who is more
> > > > familiar with this code than me should probably confirm that tp_event
> > > > being non-null and perf_event_is_tracing() being true are equivalent
> > > > though.
> > > >
> > >
> > > it looks like that's the case, AFAICS tracepoint/kprobe/uprobe events
> > > are the only ones having the tp_event pointer set, Masami?
> > >
> > > fwiw I tried to run bpf selftests with that and it's fine
> >
> > Why can't we do the most straightforward thing in this case?
> >
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index ab6c4c942f79..cf4645b26c90 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -9707,7 +9707,8 @@ static int __perf_event_overflow(struct perf_event *event,
> >
> >         ret = __perf_event_account_interrupt(event, throttle);
> >
> > -       if (event->prog && !bpf_overflow_handler(event, data, regs))
> > +       if (event->prog && event->prog->type == BPF_PROG_TYPE_PERF_EVENT &&
> > +           !bpf_overflow_handler(event, data, regs))
> >                 return ret;
> >
> >
> > >
> > > jirka
> > >
>
> Yes, that's effectively equivalent to calling perf_event_is_tracing()
> and would work too. Do you want to land that patch? It needs to go to
> 6.10 stable too.

I'd appreciate it if you can just incorporate that into your patch and
resend it, thank you!

>
> - Kyle