Re: [PATCH] perf stat: Hide invalid uncore event output for aggr mode

[Namhyung Kim]

Hi Ian,

On Wed, Jan 25, 2023 at 9:33 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Wed, Jan 25, 2023 at 12:12 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> >
> > The current display code for perf stat iterates given cpus and build the
> > aggr map to collect the event data for the aggregation mode.
> >
> > But uncore events have their own cpu maps and it won't guarantee that
> > it'd match to the aggr map.  For example, per-package uncore events
> > would generate a single value for each socket.  When user asks per-core
> > aggregation mode, the output would contain 0 values for other cores.
> >
> > Thus it needs to check the uncore PMU's cpumask and if it matches to the
> > current aggregation id.
> >
> > Before:
> >   $ sudo ./perf stat -a --per-core -e power/energy-pkg/ sleep 1
> >
> >    Performance counter stats for 'system wide':
> >
> >   S0-D0-C0              1               3.73 Joules power/energy-pkg/
> >   S0-D0-C1              0      <not counted> Joules power/energy-pkg/
> >   S0-D0-C2              0      <not counted> Joules power/energy-pkg/
> >   S0-D0-C3              0      <not counted> Joules power/energy-pkg/
> >
> >          1.001404046 seconds time elapsed
> >
> >   Some events weren't counted. Try disabling the NMI watchdog:
> >         echo 0 > /proc/sys/kernel/nmi_watchdog
> >         perf stat ...
> >         echo 1 > /proc/sys/kernel/nmi_watchdog
> >
> > The core 1, 2 and 3 should not be printed because the event is handled
> > in a cpu in the core 0 only.  With this change, the output becomes like
> > below.
> >
> > After:
> >   $ sudo ./perf stat -a --per-core -e power/energy-pkg/ sleep 1
> >
> >    Performance counter stats for 'system wide':
> >
> >   S0-D0-C0              1               2.09 Joules power/energy-pkg/
> >
> > Fixes: b89761351089 ("perf stat: Update event skip condition for system-wide per-thread mode and merged uncore and hybrid events")
>
> Thanks! Tested further with mixed core and uncore events like:
> $ perf stat -A -a -e power/energy-pkg/,cycles sleep 1
> and the "<not counted>" are now nicely gone.

Thanks for the test!

>
> > Cc: Athira Jajeev <atrajeev@xxxxxxxxxxxxxxxxxx>
> > Cc: Michael Petlan <mpetlan@xxxxxxxxxx>
> > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > ---
> >  tools/perf/util/stat-display.c | 37 ++++++++++++++++++++++++++++------
> >  1 file changed, 31 insertions(+), 6 deletions(-)
> >
> > diff --git a/tools/perf/util/stat-display.c b/tools/perf/util/stat-display.c
> > index 8bd8b0142630..b9dcb13650d0 100644
> > --- a/tools/perf/util/stat-display.c
> > +++ b/tools/perf/util/stat-display.c
> > @@ -787,6 +787,22 @@ static void uniquify_counter(struct perf_stat_config *config, struct evsel *coun
> >                 uniquify_event_name(counter);
> >  }
> >
> > +static bool check_uncore_event_aggr(struct perf_stat_config *config,
> > +                                   struct evsel *counter,
> > +                                   struct aggr_cpu_id *id)
>
> nit: const *s
> nit: check_uncore_event_aggr isn't particularly intention revealing,
> perhaps something more like should_print_counter. Perhaps some
> kernel-doc like:

It sounds like a generic name.  My intention was to check the uncore events
specifically.  But maybe we can add other conditions like global per-thread
aggregation mode.

How about this?

/*
 * When the event count is zero, check if the event should not be printed.
 * For example, uncore events have dedicated CPUs to manage them,
 * result for other CPUs should be zero and skipped.
 */
static bool should_skip_zero_counter(...)


> /**
>  * should_print_counter() - Based on id, should the current counter's
> value be printed.
>  * @config: The perf stat configuration with knowledge of the aggregation mode.
>  * @counter: The counter with its associated cpumap.
>  * @id: The aggregation type that is being queried for printing.
>  *
>  * An evlist can have evsels with different cpumaps, for example, by
> mixing core and uncore events.
>  * When displaying one counter the other counter shouldn't be printed.
> Check the counter's cpumap
>  * to see whether for any CPU it is associated with the counter should
> be printed.
>  *
>  * Return: true for counters that should be printed.
>  */
>
> > +{
> > +       struct perf_cpu cpu;
> > +       int idx;
> > +
> > +       perf_cpu_map__for_each_cpu(cpu, idx, counter->core.own_cpus) {
>
> I know this is pre-existing, sorry to whine. I think we need to
> document cpus vs own_cpus in the perf_evsel. Normally own_cpus ==
> cpus, but in a case like this the difference matters and I have a hard
> time understanding what "own" is supposed to be conveying, and why
> here we don't use cpus. I also lose what the connection is with
> perf_evlist all_cpus, does that union own_cpus or cpus? At least there
> is a comment there :-D Honestly, why do we need to even have 2 cpu
> maps in an evsel?

Right, maybe we can use pmu->cpus instead of evsel->core.own_cpus.
IIUC ->cpus is from the user and ->own_cpus is from the hardware.
I agree with you that having 2 cpu maps is confusing and it's been a
source of subtle bugs.

Let me see what I can do..

Thanks,
Namhyung

>
> > +               struct aggr_cpu_id own_id = config->aggr_get_id(config, cpu);
> > +
> > +               if (aggr_cpu_id__equal(id, &own_id))
> > +                       return true;
> > +       }
> > +       return false;
> > +}
> > +
> >  static void print_counter_aggrdata(struct perf_stat_config *config,
> >                                    struct evsel *counter, int s,
> >                                    struct outstate *os)
> > @@ -814,12 +830,21 @@ static void print_counter_aggrdata(struct perf_stat_config *config,
> >         ena = aggr->counts.ena;
> >         run = aggr->counts.run;
> >
> > -       /*
> > -        * Skip value 0 when enabling --per-thread globally, otherwise it will
> > -        * have too many 0 output.
> > -        */
> > -       if (val == 0 && config->aggr_mode == AGGR_THREAD && config->system_wide)
> > -               return;
> > +       if (val == 0) {
> > +               /*
> > +                * Skip value 0 when enabling --per-thread globally,
> > +                * otherwise it will have too many 0 output.
> > +                */
> > +               if (config->aggr_mode == AGGR_THREAD && config->system_wide)
> > +                       return;
> > +               /*
> > +                * Skip value 0 when it's an uncore event and the given aggr id
> > +                * does not belong to the PMU cpumask.
> > +                */
> > +               if (counter->core.requires_cpu &&
> > +                   !check_uncore_event_aggr(config, counter, &id))
> > +                       return;
> > +       }
> >
> >         if (!metric_only) {
> >                 if (config->json_output)
> > --
> > 2.39.1.405.gd4c25cc71f-goog
> >