Re: [PATCH v2 1/5] perf record: Set PERF_RECORD_PERIOD if attr->freq is set.

[Ian Rogers]

On Wed, Jul 29, 2020 at 2:43 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Wed, Jul 29, 2020 at 11:52 AM Arnaldo Carvalho de Melo
> <acme@xxxxxxxxxx> wrote:
> >
> > Em Tue, Jul 28, 2020 at 01:57:30AM -0700, Ian Rogers escreveu:
> > > From: David Sharp <dhsharp@xxxxxxxxxx>
> > >
> > > evsel__config() would only set PERF_RECORD_PERIOD if it set attr->freq
> >
> > There is no such thing as 'PERF_RECORD_PERIOD', its PERF_SAMPLE_PERIOD,
> > also...
> >
> > > from perf record options. When it is set by libpfm events, it would not
> > > get set. This changes evsel__config to see if attr->freq is set outside of
> > > whether or not it changes attr->freq itself.
> > >
> > > Signed-off-by: David Sharp <dhsharp@xxxxxxxxxx>
> > > Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> > > ---
> > >  tools/perf/util/evsel.c | 7 ++++++-
> > >  1 file changed, 6 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > > index ef802f6d40c1..811f538f7d77 100644
> > > --- a/tools/perf/util/evsel.c
> > > +++ b/tools/perf/util/evsel.c
> > > @@ -979,13 +979,18 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
> > >       if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
> > >                                    opts->user_interval != ULLONG_MAX)) {
> > >               if (opts->freq) {
> > > -                     evsel__set_sample_bit(evsel, PERIOD);
> > >                       attr->freq              = 1;
> > >                       attr->sample_freq       = opts->freq;
> > >               } else {
> > >                       attr->sample_period = opts->default_interval;
> > >               }
> > >       }
> > > +     /*
> > > +      * If attr->freq was set (here or earlier), ask for period
> > > +      * to be sampled.
> > > +      */
> > > +     if (attr->freq)
> > > +             evsel__set_sample_bit(evsel, PERIOD);
> >
> > Why can't the libpfm code set opts?
> >
> > With this patch we will end up calling evsel__set_sample_bit(evsel,
> > PERIOD) twice, which isn't a problem but looks strange.
>
> Thanks Arnaldo! The case I was looking at was something like:
> perf record --pfm-events cycles:freq=1000
>
> For regular events this would be:
> perf record -e cycles/freq=1000/
>
> With libpfm4 events the perf_event_attr is set up (a public API in
> linux/perf_event.h) and then parse_events__add_event is used (an
> internal API) to make the evsel and this added to the evlist
> (parse_libpfm_events_option). This is similar to the parse_events
> function except rather than set up a perf_event_attr the regular
> parsing sets up config terms that are then applied to evsel and attr
> later in evsel__config, via evsel__apply_config_terms.
>
> I think we can  update this change so that in pfm.c after
> parse_events__add_event we do:
> if (attr.freq)
>   evsel__set_sample_bit(evsel, PERIOD);
>
> This code could also be part of parse_events__add_event. I think the
> intent in placing this code here was that it is close to the similar
> evsel__apply_config_terms and setting of sample bits in the evsel. The
> logic here is already dependent on reading the attr->sample_period.
>
> I'm not sure I follow the double setting case - I think that is only
> possible with a config term or with period_set (-P).
>
> Thanks,
> Ian

Polite ping. Thanks,
Ian

>
> > - Arnaldo
> >
> > >
> > >       if (opts->no_samples)
> > >               attr->sample_freq = 0;
> > > --
> > > 2.28.0.163.g6104cc2f0b6-goog
> > >
> >
> > --
> >
> > - Arnaldo