Re: [PATCH v4 3/5] perf/core: Account dropped samples from BPF

[Andrii Nakryiko]

On Wed, Oct 23, 2024 at 11:47 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
> Hello,
>
> On Wed, Oct 23, 2024 at 09:12:52AM -0700, Andrii Nakryiko wrote:
> > On Tue, Oct 22, 2024 at 5:09 PM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> > >
> > > Like in the software events, the BPF overflow handler can drop samples
> > > by returning 0.  Let's count the dropped samples here too.
> > >
> > > Acked-by: Kyle Huey <me@xxxxxxxxxxxx>
> > > Cc: Alexei Starovoitov <ast@xxxxxxxxxx>
> > > Cc: Andrii Nakryiko <andrii@xxxxxxxxxx>
> > > Cc: Song Liu <song@xxxxxxxxxx>
> > > Cc: bpf@xxxxxxxxxxxxxxx
> > > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > > ---
> > >  kernel/events/core.c | 4 +++-
> > >  1 file changed, 3 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > index 5d24597180dec167..b41c17a0bc19f7c2 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -9831,8 +9831,10 @@ static int __perf_event_overflow(struct perf_event *event,
> > >         ret = __perf_event_account_interrupt(event, throttle);
> > >
> > >         if (event->prog && event->prog->type == BPF_PROG_TYPE_PERF_EVENT &&
> > > -           !bpf_overflow_handler(event, data, regs))
> > > +           !bpf_overflow_handler(event, data, regs)) {
> > > +               atomic64_inc(&event->dropped_samples);
> >
> > I don't see the full patch set (please cc relevant people and mailing
> > list on each patch in the patch set), but do we really want to pay the
>
> Sorry, you can find the whole series here.
>
> https://lore.kernel.org/lkml/20241023000928.957077-1-namhyung@xxxxxxxxxx
>
> I thought it's mostly for the perf part so I didn't CC bpf folks but
> I'll do in the next version.
>
>
> > price of atomic increment on what's the very typical situation of a
> > BPF program returning 0?
>
> Is it typical for BPF_PROG_TYPE_PERF_EVENT?  I guess TRACING programs
> usually return 0 but PERF_EVENT should care about the return values.
>

Yeah, it's pretty much always `return 0;` for perf_event-based BPF
profilers. It's rather unusual to return non-zero, actually.

> >
> > At least from a BPF perspective this is no "dropping sample", it's
> > just processing it in BPF and not paying the overhead of the perf
> > subsystem continuing processing it afterwards. So the dropping part is
> > also misleading, IMO.
>
> In the perf tools, we have a filtering logic in BPF to read sample
> values and to decide whether we want this sample or not.  In that case
> users would be interested in the exact number of samples.
>
> Thanks,
> Namhyung
>
> >
> > >                 return ret;
> > > +       }
> > >
> > >         /*
> > >          * XXX event_limit might not quite work as expected on inherited
> > > --
> > > 2.47.0.105.g07ac214952-goog
> > >