perf stat: Check aliases in should_skip_zero_counter
PATCH 1/1 perf stat: Check aliases in should_skip_zero_counter
Aggregation IDs (e.g. sockets, caches) with event zero counts are hidden in aggregated modes (like --per-socket) because alias merging aggregates counts but not CPU maps. This causes display logic to skip aggregation IDs not present in the leader's map.
Instead of mutating CPU maps in evsel__merge_aggr_counters (which could have side effects on shared CPU maps), check aliases in should_skip_zero_counter to ensure all applicable aggregation IDs are displayed.
Before the fix:
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1
Performance counter stats for 'system wide':
S0 12 0 amd_umc/config=0xff/
1.000721604 seconds time elapsed
After the fix:
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1
Performance counter stats for 'system wide':
S0 12 0 amd_umc/config=0xff/ S1 12 0 amd_umc/config=0xff/
1.000879376 seconds time elapsed
perf test looks good too.
--- tools/perf/util/stat-display.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+)
tools/perf/util/stat-display.c +21 -0
@@ -945,6 +945,27 @@ static bool should_skip_zero_counter(struct perf_stat_config *config,if (aggr_cpu_id__equal (id, &own_id))return false; } + +/* +* If the counter is a leader in alias merging, check if the aggr id +* matches any of its aliases. +*/ +if (config->aggr_mode !=AGGR_NONE && counter->first_wildcard_match ==NULL ) { +struct evsel *alias; + +evlist__for_each_entry (counter->evlist, alias) { +if (alias->first_wildcard_match == counter) { +perf_cpu_map__for_each_cpu (cpu, idx, alias->core.cpus) { +struct aggr_cpu_id own_id = + config->aggr_get_id (config, cpu); + +if (aggr_cpu_id__equal (id, &own_id)) +return false; + } + } + } + } +return true; }
Signature · 1 line
2.54.0.545.g6539524ca2-goog
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] The alias checking loop introduces an O(M * N^2) time complexity during the display phase…
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The alias checking loop introduces an O(M * N^2) time complexity during the display phase, causing unnecessary overhead. --
commit 1ba80061bda839b47a8c18bdc8aa600c735044d6
Author: Chun-Tse Shao <[email protected]>
perf stat: Check aliases in should_skip_zero_counter
This commit checks aliases directly in should_skip_zero_counter instead of mutating CPU maps in evsel__merge_aggr_counters. This is done to prevent zero count aggregation IDs from being hidden in aggregated modes like --per-socket.
Quoted 7 lines
tools/perf/util/stat-display.c +0
-0@@ -945,6 +945,27 @@ static bool should_skip_zero_counter(struct perf_stat_config *config,if (aggr_cpu_id__equal (id, &own_id))return false;
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 956
Quoted 10 lines
tools/perf/util/stat-display.c +9
-0} + +/* +* If the counter is a leader in alias merging, check if the aggr id +* matches any of its aliases. +*/ +if (config->aggr_mode !=AGGR_NONE && counter->first_wildcard_match ==NULL ) { +struct evsel *alias; + +evlist__for_each_entry (counter->evlist, alias) {
Does this introduce an O(M * N^2) time complexity during the display phase?
Since should_skip_zero_counter() is called for each aggregation unit (M) and each event (N), the function is called O(M * N) times.
If the event is a standalone event or an alias leader, this block iterates over the entire evlist (N events) to find its aliases.
Could this scale poorly for very large event lists by introducing unnecessary iterations, leading to an overall O(M * N^2) complexity?
Quoted 14 lines
tools/perf/util/stat-display.c +12
-0+if (alias->first_wildcard_match == counter) { +perf_cpu_map__for_each_cpu (cpu, idx, alias->core.cpus) { +struct aggr_cpu_id own_id = + config->aggr_get_id (config, cpu); + +if (aggr_cpu_id__equal (id, &own_id)) +return false; + } + } + } + } +return true; }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Ian Rogers
reviewerQuoted 21 lines · Chun-Tse Shao
Aggregation IDs (e.g. sockets, caches) with event zero counts are hidden in aggregated modes (like
--per-socket) because alias merging aggregates counts but not CPU maps. This causes display logic to skip aggregation IDs not present in the leader's map.Instead of mutating CPU maps in
evsel__merge_aggr_counters(which could have side effects on shared CPU maps), check aliases inshould_skip_zero_counterto ensure all applicable aggregation IDs are displayed.Before the fix:
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1Performance counter stats for 'system wide':
S0 12 0 amd_umc/config=0xff/
1.000721604 seconds time elapsedAfter the fix:
Quoted 8 lines
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1Performance counter stats for 'system wide':
S0 12 0 amd_umc/config=0xff/ S1 12 0 amd_umc/config=0xff/1.000879376 seconds time elapsed
Does this change the behavior without --per-socket when the uncore PMUs have different cpumasks?
Quoted 13 lines
perf testlooks good too.Assisted-by Gemini:gemini-3.1-pro-previewSigned-off-by Chun-Tse Shao <[email protected]>--- tools/perf/util/stat-display.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+)tools/perf/util/stat-display.c +0
-0
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 951
Quoted 7 lines
tools/perf/util/stat-display.c +4
-0@@ -945,6 +945,27 @@ static bool should_skip_zero_counter(struct perf_stat_config *config,if (aggr_cpu_id__equal (id, &own_id))return false; } + +/* +* If the counter is a leader in alias merging, check if the aggr id +* matches any of its aliases.
Could you expand why in this case we want the zero count?
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 957
Quoted 6 lines
tools/perf/util/stat-display.c +6
-0+*/ +if (config->aggr_mode !=AGGR_NONE && counter->first_wildcard_match ==NULL ) { +struct evsel *alias; + +evlist__for_each_entry (counter->evlist, alias) { +if (alias->first_wildcard_match == counter) {
Could this also have a "&& !perf_cpu_map__equal(counter->core.cpu, alias->core.cpus)" ? ie if the alias' cpumask matches that of the counter, then the loop above covers it. Perhaps the cpumaps of each alias could be merged and the test only run if their intersection is not empty.
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 960
Quoted 3 lines
tools/perf/util/stat-display.c +3
-0+perf_cpu_map__for_each_cpu (cpu, idx, alias->core.cpus) { +struct aggr_cpu_id own_id = + config->aggr_get_id (config, cpu);
Could this fit on one line?
Thanks,
Ian
Quoted 14 lines
tools/perf/util/stat-display.c +8
-1+ +if (aggr_cpu_id__equal (id, &own_id)) +return false; + } + } + } + } +return true; } --2.54.0.545.g6539524ca2-goog
Hi Ian, thanks for your review.
Quoted 27 lines · Ian Rogers
Quoted 25 lines · Chun-Tse Shao
Aggregation IDs (e.g. sockets, caches) with event zero counts are hidden in aggregated modes (like
--per-socket) because alias merging aggregates counts but not CPU maps. This causes display logic to skip aggregation IDs not present in the leader's map.Instead of mutating CPU maps in
evsel__merge_aggr_counters(which could have side effects on shared CPU maps), check aliases inshould_skip_zero_counterto ensure all applicable aggregation IDs are displayed.Before the fix:
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1Performance counter stats for 'system wide':
S0 12 0 amd_umc/config=0xff/
1.000721604 seconds time elapsedAfter the fix:
$ perf stat -e amd_umc/config=0xff/ --per-socket -a -- sleep 1Performance counter stats for 'system wide':
Quoted 7 lines
Quoted 4 lines
S0 12 0 amd_umc/config=0xff/ S1 12 0 amd_umc/config=0xff/1.000879376 seconds time elapsedDoes this change the behavior without --per-socket when the uncore PMUs have different cpumasks?
Nope, the behavior is unchanged for -a, since counts aggregate to the leader's aggr[0] and all CPUs's aggr_id is 0 in AGGR_GLOBAL mode.
Quoted 9 lines
Quoted 9 lines
perf testlooks good too.Assisted-by Gemini:gemini-3.1-pro-previewSigned-off-by Chun-Tse Shao <[email protected]>--- tools/perf/util/stat-display.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+)
Quoted 14 lines
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 951
Quoted 12 lines
tools/perf/util/stat-display.c +4
-0@@ -945,6 +945,27 @@ static bool should_skip_zero_counter(struct perf_stat_config *config,if (aggr_cpu_id__equal (id, &own_id))return false; } + +/* +* If the counter is a leader in alias merging, check if the aggr id +* matches any of its aliases. Could you expand why in this case we want the zero count?
We don't want to hide counts if it is not in leader's cpumask even though it is 0. It would cause confusion to users/parsers. In my scenario, I tested an event that was all zeros on S1 and was hidden by perf. Initially, I thought perf wasn't collecting the event on S1, then realized collecting was working but hidden.
Quoted 8 lines
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 957
Quoted 7 lines
tools/perf/util/stat-display.c +6
-0+*/ +if (config->aggr_mode !=AGGR_NONE && counter->first_wildcard_match ==NULL ) { +struct evsel *alias; + +evlist__for_each_entry (counter->evlist, alias) { +if (alias->first_wildcard_match == counter) {
Quoted 5 lines
Could this also have a "&& !perf_cpu_map__equal(counter->core.cpu,
alias->core.cpus)" ? ie if the alias' cpumask matches that of the counter, then the loop above covers it. Perhaps the cpumaps of each alias could be merged and the test only run if their intersection is not empty.
Thanks for the suggestion, I will add it in the next patch.
Quoted 6 lines
tools/perf/util/stat-display.c static bool should_skip_zero_counter(struct perf_stat_config *config, · line 960
Quoted 4 lines
tools/perf/util/stat-display.c +3
-0+perf_cpu_map__for_each_cpu (cpu, idx, alias->core.cpus) { +struct aggr_cpu_id own_id = + config->aggr_get_id (config, cpu);Could this fit on one line?
Then it will make the line 101 characters long...
Quoted 18 lines
Thanks,
IanQuoted 14 lines
tools/perf/util/stat-display.c +8
-1+ +if (aggr_cpu_id__equal (id, &own_id)) +return false; + } + } + } + } +return true; } --2.54.0.545.g6539524ca2-goog