Re: [PATCH v5 00/11] Better fixes for grouping of events

[Arnaldo Carvalho de Melo]

Em Sat, Mar 11, 2023 at 06:15:32PM -0800, Ian Rogers escreveu:
> The rules for grouping events have grown more complex. Topdown events
> must be grouped, but flags like --metric-no-group and flags on metrics
> don't respect this. Uncore events may be expanded using wild cards for
> PMU names, but then the events need reordering so the group members
> are adjacent. Rather than fixing metrics, this change fixes the main
> event parsing code to first sort and then regroup evsels.
> 
> As this is shared functionality changes to it should cause
> concern. The change is done with the intent of simplifying and making
> more robust the grouping logic, examples are given. If additional
> changes are necessary, they are most likely necessary to the
> evsel__group_pmu_name logic as the code avoids breaking groups that
> are on the same "group" PMU. The group_pmu_name is a variant of the
> pmu_name tweaked in the case of software and aux events, that use
> groups in a slightly different manner to conventional events.
> 
> The code was manually tested as well as passing perf test on a Intel
> tigerlake CPU with intel-pt.
> 
> v5. Add Adrian's Acked-by on patch 3, add a fix to the warning logic
>     for multiple event groups as reported by Adrian.


Appled locally, testing.

- Arnaldo

> v4. Move the Intel pmu->auxtrace initialization to the existing
>     perf_pmu__get_default_config as suggested by Adrian Hunter.
> v3. Rename pmu_name to group_pmu_name and add patch to warn when
>     events are regrouped as requested by Namhyung.
> v2. Fix up the commit message on 4/10 (thanks Arnaldo). Drop
>     unnecessary v1 5/10 (thanks Kan). evlist->core.nr_groups wasn't
>     being correctly maintained after the sort/regrouping and so the
>     new patch 10/10 removes that variable and computes it from the
>     evlist when necessary, generally just tests.
> 
> Ian Rogers (11):
>   libperf evlist: Avoid a use of evsel idx
>   perf stat: Don't remove all grouped events when CPU maps disagree
>   perf pmu: Earlier PMU auxtrace initialization
>   perf stat: Modify the group test
>   perf evsel: Allow const evsel for certain accesses
>   perf evsel: Add function to compute group PMU name
>   perf parse-events: Pass ownership of the group name
>   perf parse-events: Sort and group parsed events
>   perf evsel: Remove use_uncore_alias
>   perf evlist: Remove nr_groups
>   perf parse-events: Warn when events are regrouped
> 
>  tools/lib/perf/evlist.c                  |  31 ++-
>  tools/lib/perf/include/internal/evlist.h |   1 -
>  tools/lib/perf/include/perf/evlist.h     |   1 +
>  tools/perf/arch/x86/util/auxtrace.c      |   4 -
>  tools/perf/arch/x86/util/evlist.c        |  39 ++--
>  tools/perf/arch/x86/util/pmu.c           |   8 +-
>  tools/perf/builtin-record.c              |   2 +-
>  tools/perf/builtin-report.c              |   2 +-
>  tools/perf/builtin-stat.c                |  24 +-
>  tools/perf/tests/bpf.c                   |   1 -
>  tools/perf/tests/parse-events.c          |  24 +-
>  tools/perf/tests/pfm.c                   |  12 +-
>  tools/perf/tests/pmu-events.c            |   2 +-
>  tools/perf/util/evlist.c                 |   2 +-
>  tools/perf/util/evlist.h                 |   8 +-
>  tools/perf/util/evsel.c                  |  27 ++-
>  tools/perf/util/evsel.h                  |   8 +-
>  tools/perf/util/header.c                 |   3 +-
>  tools/perf/util/metricgroup.c            |   3 +-
>  tools/perf/util/parse-events.c           | 275 ++++++++++++-----------
>  tools/perf/util/parse-events.h           |  14 +-
>  tools/perf/util/parse-events.y           |  28 +--
>  tools/perf/util/pfm.c                    |   1 -
>  tools/perf/util/pmu.c                    |   6 +-
>  tools/perf/util/python.c                 |   2 +-
>  tools/perf/util/stat-shadow.c            |   2 +-
>  26 files changed, 284 insertions(+), 246 deletions(-)
> 
> -- 
> 2.40.0.rc1.284.g88254d51c5-goog
> 

-- 

- Arnaldo