On Tue, Apr 08, 2025 at 11:10:27PM -0700, Ian Rogers wrote:
> The patches add support to computing the min, max, mean or last
> retirement latency and then using that value as the basis for metrics.
> When values aren't available, support is added to use the retirement
> latency as recorded for an event in the perf json.
>
> Support is added for reading the retirement latency from the forked
> perf command more than once. To avoid killing the process commands are
> sent through the control fd. Some name handling is changed to make it
> more robust.
>
> Rather than retirement latency events having issues with perf record,
> make it so that the retirement latency modifier enables sample
> weights.
>
> v4: Don't use json min/max in retirement latency stats as they will
> never update afterwards. Warn once if json data is used when TPEBS
> recording was requested.
>
> v3: Two fixes from Kan Liang. Ensure min/max statistics don't vary
> when real samples are encountered.
>
> v2: Addition of header cleanup patch originally posted:
> https://lore.kernel.org/lkml/20241210191823.612631-1-irogers@google.com/
> as there are no arch specific reasons not to build this code.
> Fix bug in "perf pmu-events: Add retirement latency to JSON events
> inside of perf" where "t->stats.n != 0" should have been
> "t->stats.n == 0".
> Add patch so that perf record of a retirement latency event
> doesn't crash but instead enables sample weights for the event.
>
> Ian Rogers (16):
> perf intel-tpebs: Cleanup header
> perf intel-tpebs: Simplify tpebs_cmd
> perf intel-tpebs: Rename tpebs_start to evsel__tpebs_open
> perf intel-tpebs: Separate evsel__tpebs_prepare out of
> evsel__tpebs_open
> perf intel-tpebs: Move cpumap_buf out of evsel__tpebs_open
> perf intel-tpebs: Reduce scope of tpebs_events_size
> perf intel-tpebs: Inline get_perf_record_args
> perf intel-tpebs: Ensure events are opened, factor out finding
> perf intel-tpebs: Refactor tpebs_results list
> perf intel-tpebs: Add support for updating counts in evsel__tpebs_read
> perf intel-tpebs: Add mutex for tpebs_results
> perf intel-tpebs: Don't close record on read
> perf intel-tpebs: Use stats for retirement latency statistics
> perf stat: Add mean, min, max and last --tpebs-mode options
> perf pmu-events: Add retirement latency to JSON events inside of perf
> perf record: Retirement latency cleanup in evsel__config
I have a nitpick but otherwise looks good to me.
Acked-by: Namhyung Kim <namhyung@kernel.org>
Thanks,
Namhyung
>
> tools/perf/Documentation/perf-stat.txt | 7 +
> tools/perf/builtin-stat.c | 29 +-
> tools/perf/pmu-events/empty-pmu-events.c | 216 +++----
> tools/perf/pmu-events/jevents.py | 6 +
> tools/perf/pmu-events/pmu-events.h | 3 +
> tools/perf/util/Build | 2 +-
> tools/perf/util/evlist.c | 1 -
> tools/perf/util/evsel.c | 22 +-
> tools/perf/util/evsel.h | 6 +
> tools/perf/util/intel-tpebs.c | 682 ++++++++++++++---------
> tools/perf/util/intel-tpebs.h | 40 +-
> tools/perf/util/parse-events.c | 4 +
> tools/perf/util/pmu.c | 52 +-
> tools/perf/util/pmu.h | 3 +
> 14 files changed, 666 insertions(+), 407 deletions(-)
>
> --
> 2.49.0.504.g3bcea36a83-goog
>