[PATCH v4 0/7] Event parsing fixes
Ian Rogers
irogers at google.com
Thu Aug 22 08:10:13 PDT 2024
On Thu, Aug 22, 2024 at 7:32 AM Liang, Kan <kan.liang at linux.intel.com> wrote:
>
>
>
> On 2024-08-22 9:24 a.m., James Clark wrote:
> > I rebased this one and made some other fixes so that I could test it,
> > so I thought I'd repost it here in case it's helpful. I also added a
> > new test.
> >
> > But for the testing it all looks ok.
> >
> > There is one small difference where it now shows "stalled-cycles-..."
> > as <not supported> events, when before it just didn't show them at all when
> > they weren't supported:
> >
> > $ perf stat -- true
> >
> > Performance counter stats for 'true':
> >
> > 0.66 msec task-clock # 0.384 CPUs utilized
> > 0 context-switches # 0.000 /sec
> > 0 cpu-migrations # 0.000 /sec
> > 52 page-faults # 78.999 K/sec
> > <not counted> cpu_atom/instructions/ (0.00%)
> > 978,399 cpu_core/instructions/ # 1.02 insn per cycle
> > <not counted> cpu_atom/cycles/ (0.00%)
> > 959,722 cpu_core/cycles/ # 1.458 GHz
> > <not supported> cpu_atom/stalled-cycles-frontend/
> > <not supported> cpu_core/stalled-cycles-frontend/
> >
>
> Intel didn't support the events for a very long time. It would impact
> many existing generations and all future generations.
> The current method is to hide the non-exist events. The TopdownL1 is an
> example. If it doesn't exist in the json file, perf stat will not
> display it.
> I don't think it's a good idea to disclose non-exist events in the perf
> stat default.
>
> The <not supported> doesn't help here, since there could be many reasons
> that the perf tool fails to open a counter. It just provides a
> misleading message for an event that never existed.
The list of "default" events, not metrics, similarly has "<not
supported>" in many configurations with "-dd" or "-ddd" on AMD. I'm
not sure the set of default events, at different detail levels, is
necessarily the best. The default events can also be a source of
multiplexing, for example, showing branch miss rate alongside topdown
metrics. Anyway, for the "<not supported>" we should probably be able
to tweak should_skip_zero_counter that is in stat-display.c and tag
these default events as "skippable".
Thanks,
Ian
> Thanks,
> Kan
> > I don't think that's a big deal though and could probably be fixed up
> > later if we really want to.
> >
> > Tested on Raptor Lake, Juno, N1, Ampere (with the DSU cycles PMU) and
> > I also faked an Apple M on Juno.
> >
> > Changes since v3:
> >
> > * Rebase onto perf-tools-next 6236ebe07
> > * Fix Intel TPEBS counting mode test
> > * Fix arm-spe build
> > * Add support for DT devices in stat test
> > * Add a new test for hybrid perf stat default arguments
> >
> > Ian Rogers (5):
> > perf evsel: Add alternate_hw_config and use in evsel__match
> > perf stat: Uniquify event name improvements
> > perf stat: Remove evlist__add_default_attrs use strings
> > perf evsel x86: Make evsel__has_perf_metrics work for legacy events
> > perf evsel: Remove pmu_name
> >
> > James Clark (2):
> > perf test: Make stat test work on DT devices
> > perf test: Add a test for default perf stat command
> >
> > tools/perf/arch/arm64/util/arm-spe.c | 4 +-
> > tools/perf/arch/x86/util/evlist.c | 74 +----
> > tools/perf/arch/x86/util/evsel.c | 35 ++-
> > tools/perf/builtin-diff.c | 6 +-
> > tools/perf/builtin-stat.c | 291 +++++++-----------
> > tools/perf/tests/parse-events.c | 2 +-
> > tools/perf/tests/shell/stat.sh | 33 +-
> > .../perf/tests/shell/test_stat_intel_tpebs.sh | 11 +-
> > tools/perf/util/evlist.c | 46 +--
> > tools/perf/util/evlist.h | 12 -
> > tools/perf/util/evsel.c | 28 +-
> > tools/perf/util/evsel.h | 22 +-
> > tools/perf/util/metricgroup.c | 4 +-
> > tools/perf/util/parse-events.c | 58 ++--
> > tools/perf/util/parse-events.h | 8 +-
> > tools/perf/util/parse-events.y | 2 +-
> > tools/perf/util/pmu.c | 6 +-
> > tools/perf/util/pmu.h | 2 +-
> > tools/perf/util/stat-display.c | 101 ++++--
> > tools/perf/util/stat-shadow.c | 14 +-
> > tools/perf/util/stat.c | 2 +-
> > 21 files changed, 348 insertions(+), 413 deletions(-)
> >
More information about the linux-arm-kernel
mailing list