perf stat: Fix duplicate output with --for-each-cgroup
Namhyung Kim
maintainerPATCH 1/2 perf stat: Fix duplicate output with --for-each-cgroup
Currently it produces following output with duplicate events when
--for-each-cgroup option is used. It seems perf stat adds them when
it handles default events but didn't copy some fields in evsel__clone().
$ sudo perf stat -a --for-each-cgroup / true
Performance counter stats for 'system wide':
8,440,165 duration_time /
8,439,895 duration_time /
8,440,015 duration_time /
8,440,024 duration_time /
8,440,075 duration_time /
8,440,095 duration_time /
330 context-switches / # 679.4 cs/sec cs_per_second
485.69 msec cpu-clock / # 57.5 CPUs CPUs_utilized
70 cpu-migrations / # 144.1 migrations/sec migrations_per_second
71 page-faults / # 146.2 faults/sec page_faults_per_second
12,183,711 branch-misses / # 10.9 % branch_miss_rate (5.15%)
111,981,297 branches / (5.15%)
95,844,809 branches / # 197.3 M/sec branch_frequency (35.49%)
65,611,429 cpu-cycles / # 0.1 GHz cycles_frequency (98.32%)
24,170,987 cpu-cycles / (95.12%)
18,552,509 instructions / # 0.8 instructions insn_per_cycle (95.12%)
22,405,293 cpu-cycles / (64.78%)
6,840,383 stalled-cycles-frontend / # 0.31 frontend_cycles_idle (64.78%)
<not counted> cpu-cycles /
<not supported> stalled-cycles-backend / # nan backend_cycles_idle
<not supported> stalled-cycles-backend / # nan stalled_cycles_per_instruction
<not supported> instructions /
<not supported> stalled-cycles-frontend /0.006546057 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
But I'm worrying about opening same events multiple times. Probably due to grouping, but I'm not sure if it's beneficial in the end. Without duplication, it seems it won't cause multiplexing (assuming no other users at the same time).
--- tools/perf/util/evsel.c | 6 ++++++ 1 file changed, 6 insertions(+)
tools/perf/util/evsel.c +6 -0
@@ -612,12 +612,18 @@ struct evsel *evsel__clone(struct evsel *orig) evsel->percore = orig->percore; evsel->precise_max = orig->precise_max; evsel->is_libpfm_event = orig->is_libpfm_event; + evsel->skippable = orig->skippable; + evsel->retire_lat = orig->retire_lat; + evsel->dont_regroup = orig->dont_regroup; evsel->exclude_GH = orig->exclude_GH; evsel->sample_read = orig->sample_read; evsel->collect_stat = orig->collect_stat; evsel->weak_group = orig->weak_group; evsel->use_config_name = orig->use_config_name; + evsel->default_metricgroup = orig->default_metricgroup; + evsel->default_show_events = orig->default_show_events; + evsel->pmu = orig->pmu; evsel->first_wildcard_match = orig->first_wildcard_match;
Signature · 1 line
2.55.0.795.g602f6c329a-goog
Namhyung Kim
maintainerPATCH 2/2 perf evsel: Arrange some fields that should be cloned
In the evsel, there's an internal struct to put fields need copy when the evsel is cloned. This is purely to make it easier track those fields even if it sometimes failed to do so. :)
--- tools/perf/util/evsel.h | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)
tools/perf/util/evsel.h +6 -5
@@ -74,6 +74,10 @@ struct evsel {const char *unit;struct cgroup *cgrp;const char *metric_id; + +/* The PMU the event is from. Used for missing_features, PMU name, etc. */ +struct perf_pmu *pmu; +/* * This point to the first evsel with the same name, intended to store the * aggregated counts in aggregation mode.
@@ -94,6 +98,8 @@ struct evsel {bool skippable;bool retire_lat;bool dont_regroup; +bool default_metricgroup;/* A member of the Default metricgroup */ +bool default_show_events;/* If a default group member, show the event */ int bpf_fd;struct bpf_object *bpf_obj;struct list_head config_terms;
@@ -125,8 +131,6 @@ struct evsel {bool cmdline_group_boundary;bool reset_group;bool needs_auxtrace_mmap; -bool default_metricgroup;/* A member of the Default metricgroup */ -bool default_show_events;/* If a default group member, show the event */ bool needs_uniquify;bool fallenback_eacces;bool fallenback_eopnotsupp;
@@ -182,9 +186,6 @@ struct evsel {unsigned long open_flags;int precise_ip_original; -/* The PMU the event is from. Used for missing_features, PMU name, etc. */ -struct perf_pmu *pmu; -/* For tool events */ /* Beginning time subtracted when the counter is read. */ union {
Signature · 1 line
2.55.0.795.g602f6c329a-goog
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Incomplete state cloning in `evsel__clone()` breaks TPEBS latency fallback metrics and BPF…
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Incomplete state cloning in evsel__clone() breaks TPEBS latency fallback metrics and BPF counters for duplicated events. --
--- Patch [1]: [PATCH v1 1/2] perf stat: Fix duplicate output with --for-each-cgroup --- commit 10359224b851cd21a00edee0e074af745faf63be
Author: Namhyung Kim <[email protected]>
perf stat: Fix duplicate output with --for-each-cgroup
This commit fixes duplicate event output in perf stat when using the --for-each-cgroup option. It addresses the issue by copying several fields that were previously missed during evsel__clone().
tools/perf/util/evsel.c struct evsel *evsel__clone(struct evsel *dest, struct evsel *orig) · line 614
Quoted 10 lines
tools/perf/util/evsel.c +2
-0@@ -610,12 +610,18 @@ struct evsel *evsel__clone(struct evsel *dest, struct evsel *orig) evsel->percore = orig->percore; evsel->precise_max = orig->precise_max; evsel->is_libpfm_event = orig->is_libpfm_event; + evsel->skippable = orig->skippable; + evsel->retire_lat = orig->retire_lat;
[Severity: Medium]
When copying the retire_lat flag, do we also need to copy the associated retirement_latency struct and the bpf_counter flag to avoid a regression?
Without copying the retirement_latency struct (mean, min, max), a cloned event falling back to TPEBS might use a zeroed struct and incorrectly calculate 0.0 latency instead of using the parsed values.
Similarly, omitting the bpf_counter flag might cause cloned BPF events to silently bypass BPF counting.
Quoted 4 lines
tools/perf/util/evsel.c +1
-0+ evsel->dont_regroup = orig->dont_regroup; evsel->exclude_GH = orig->exclude_GH; evsel->sample_read = orig->sample_read;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1