[PATCH 03/22] perf stat: Switch aggregation to use for_each loop

[Ian Rogers]

On Sat, Dec 11, 2021 at 11:25 AM Jiri Olsa <jolsa at redhat.com> wrote:
>
> On Tue, Dec 07, 2021 at 06:45:48PM -0800, Ian Rogers wrote:
> > Tidy up the use of cpu and index to hopefully make the code less error
> > prone. Avoid unused warnings with (void) which will be removed in a
> > later patch.
> >
> > In aggr_update_shadow, the perf_cpu_map is switched from
> > the evlist to the counter's cpu map, so the index is appropriate. This
> > addresses a problem where uncore counts, with a cpumap like:
> > $ cat /sys/devices/uncore_imc_0/cpumask
> > 0,18
> > Don't aggregate counts in CPUs based on the index of those values in the
> > cpumap (0 and 1) but on the actual CPU (0 and 18). Thereby correcting
> > metric calculations in per-socket mode for counters with without a full
> > cpumask.
> >
> > Signed-off-by: Ian Rogers <irogers at google.com>
> > ---
> >  tools/perf/util/stat-display.c | 48 +++++++++++++++++++---------------
> >  1 file changed, 27 insertions(+), 21 deletions(-)
> >
> > diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c
> > index 588601000f3f..efab39a759ff 100644
> > --- a/tools/perf/util/stat-display.c
> > +++ b/tools/perf/util/stat-display.c
> > @@ -330,8 +330,8 @@ static void print_metric_header(struct perf_stat_config *config,
> >  static int first_shadow_cpu(struct perf_stat_config *config,
> >                           struct evsel *evsel, struct aggr_cpu_id id)
> >  {
> > -     struct evlist *evlist = evsel->evlist;
> > -     int i;
> > +     struct perf_cpu_map *cpus;
> > +     int cpu, idx;
> >
> >       if (config->aggr_mode == AGGR_NONE)
> >               return id.core;
> > @@ -339,14 +339,11 @@ static int first_shadow_cpu(struct perf_stat_config *config,
> >       if (!config->aggr_get_id)
> >               return 0;
> >
> > -     for (i = 0; i < evsel__nr_cpus(evsel); i++) {
> > -             int cpu2 = evsel__cpus(evsel)->map[i];
> > -
> > -             if (cpu_map__compare_aggr_cpu_id(
> > -                                     config->aggr_get_id(config, evlist->core.cpus, cpu2),
> > -                                     id)) {
> > -                     return cpu2;
> > -             }
> > +     cpus = evsel__cpus(evsel);
> > +     perf_cpu_map__for_each_cpu(cpu, idx, cpus) {
> > +             if (cpu_map__compare_aggr_cpu_id(config->aggr_get_id(config, cpus, idx),
> > +                                              id))
> > +                     return cpu;
>
> so this looks strange, you pass idx instead of cpu2 to aggr_get_id,
> which takes idx as 3rd argument, so it looks like it was broken now,
> should this be a separate fix?

Yep, I tried to cover this in the commit message, but agree a separate
patch would be clearer. The aggregation is currently broken on
anything other than CPU 0 or when the CPU mask covers every CPU - the
case for something like topdown, hence this not being spotted.

> also the original code for some reason passed evlist->core.cpus
> to aggr_get_id, which might differ rom evsel's cpus

Part of the same fix.

> same for aggr_update_shadow change

In this case the cpu is really an index and so the change is just
renaming one to the other for the sake of clarity.

Thanks,
Ian

> jirka
>