Re: [PATCH v3 01/18] perf metricgroup: Add care to picking the evsel for displaying a metric
From: Ian Rogers
Date: Tue Nov 11 2025 - 12:20:42 EST
On Tue, Nov 11, 2025 at 12:15 AM Mi, Dapeng <dapeng1.mi@xxxxxxxxxxxxxxx> wrote:
>
>
> On 11/11/2025 12:04 PM, Ian Rogers wrote:
> > Rather than using the first evsel in the matched events, try to find
> > the least shared non-tool evsel. The aim is to pick the first evsel
> > that typifies the metric within the list of metrics.
> >
> > This addresses an issue where Default metric group metrics may lose
> > their counter value due to how the stat displaying hides counters for
> > default event/metric output.
> >
> > For a metricgroup like TopdownL1 on an Intel Alderlake the change is,
> > before there are 4 events with metrics:
> > ```
> > $ perf stat -M topdownL1 -a sleep 1
> >
> > Performance counter stats for 'system wide':
> >
> > 7,782,334,296 cpu_core/TOPDOWN.SLOTS/ # 10.4 % tma_bad_speculation
> > # 19.7 % tma_frontend_bound
> > 2,668,927,977 cpu_core/topdown-retiring/ # 35.7 % tma_backend_bound
> > # 34.1 % tma_retiring
> > 803,623,987 cpu_core/topdown-bad-spec/
> > 167,514,386 cpu_core/topdown-heavy-ops/
> > 1,555,265,776 cpu_core/topdown-fe-bound/
> > 2,792,733,013 cpu_core/topdown-be-bound/
> > 279,769,310 cpu_atom/TOPDOWN_RETIRING.ALL/ # 12.2 % tma_retiring
> > # 15.1 % tma_bad_speculation
> > 457,917,232 cpu_atom/CPU_CLK_UNHALTED.CORE/ # 38.4 % tma_backend_bound
> > # 34.2 % tma_frontend_bound
> > 783,519,226 cpu_atom/TOPDOWN_FE_BOUND.ALL/
> > 10,790,192 cpu_core/INT_MISC.UOP_DROPPING/
> > 879,845,633 cpu_atom/TOPDOWN_BE_BOUND.ALL/
> > ```
> >
> > After there are 6 events with metrics:
> > ```
> > $ perf stat -M topdownL1 -a sleep 1
> >
> > Performance counter stats for 'system wide':
> >
> > 2,377,551,258 cpu_core/TOPDOWN.SLOTS/ # 7.9 % tma_bad_speculation
> > # 36.4 % tma_frontend_bound
> > 480,791,142 cpu_core/topdown-retiring/ # 35.5 % tma_backend_bound
> > 186,323,991 cpu_core/topdown-bad-spec/
> > 65,070,590 cpu_core/topdown-heavy-ops/ # 20.1 % tma_retiring
> > 871,733,444 cpu_core/topdown-fe-bound/
> > 848,286,598 cpu_core/topdown-be-bound/
> > 260,936,456 cpu_atom/TOPDOWN_RETIRING.ALL/ # 12.4 % tma_retiring
> > # 17.6 % tma_bad_speculation
> > 419,576,513 cpu_atom/CPU_CLK_UNHALTED.CORE/
> > 797,132,597 cpu_atom/TOPDOWN_FE_BOUND.ALL/ # 38.0 % tma_frontend_bound
> > 3,055,447 cpu_core/INT_MISC.UOP_DROPPING/
> > 671,014,164 cpu_atom/TOPDOWN_BE_BOUND.ALL/ # 32.0 % tma_backend_bound
> > ```
>
> It looks the output of cpu_core and cpu_atom events are mixed together,
> like the "cpu_core/INT_MISC.UOP_DROPPING/". Could we resort the events and
> separate the cpu_core and cpu_atom events output? It would make the output
> more read-friendly. Thanks.
So the metrics are tagged as to not group the events:
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/arch/x86/alderlake/adl-metrics.json?h=perf-tools-next#n117
Running with each metric causes the output to be:
```
$ perf stat -M tma_bad_speculation,tma_backend_bound,tma_frontend_bound,tma_retiring
-a sleep 1
Performance counter stats for 'system wide':
1,615,145,897 cpu_core/TOPDOWN.SLOTS/ # 8.1 %
tma_bad_speculation
# 42.5 %
tma_frontend_bound (49.89%)
243,037,087 cpu_core/topdown-retiring/ # 34.5 %
tma_backend_bound (49.89%)
129,341,306 cpu_core/topdown-bad-spec/
(49.89%)
2,679,894 cpu_core/INT_MISC.UOP_DROPPING/
(49.89%)
696,940,348 cpu_core/topdown-fe-bound/
(49.89%)
563,319,011 cpu_core/topdown-be-bound/
(49.89%)
1,795,034,847 cpu_core/slots/
(50.11%)
262,140,961 cpu_core/topdown-retiring/
(50.11%)
44,589,349 cpu_core/topdown-heavy-ops/ # 14.4 %
tma_retiring (50.11%)
160,987,341 cpu_core/topdown-bad-spec/
(50.11%)
778,250,364 cpu_core/topdown-fe-bound/
(50.11%)
622,499,674 cpu_core/topdown-be-bound/
(50.11%)
90,849,750 cpu_atom/TOPDOWN_RETIRING.ALL/ # 8.1 %
tma_retiring
# 17.2 %
tma_bad_speculation
223,878,243 cpu_atom/CPU_CLK_UNHALTED.CORE/
423,068,733 cpu_atom/TOPDOWN_FE_BOUND.ALL/ # 37.8 %
tma_frontend_bound
413,413,499 cpu_atom/TOPDOWN_BE_BOUND.ALL/ # 36.9 %
tma_backend_bound
```
so you can see that it is the effect of not grouping the events that
leads to the cpu_core and cpu_atom split.
The code that does sorting/fixing/adding of events, primarily to fix
topdown, is parse_events__sort_events_and_fix_groups:
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/parse-events.c?h=perf-tools-next#n2030
but I've tried to make that code respect the incoming evsel list order
because if a user specifies an order then they generally expect it to
be respected (unless invalid or because of topdown events). For
--metric-only the event order doesn't really matter.
Anyway, I think trying to fix this is out of scope for this patch
series, although I agree with you about the readability. The behavior
here matches old behavior such as:
```
$ perf --version
perf version 6.16.12
$ perf stat -M TopdownL1 -a sleep 1
Performance counter stats for 'system wide':
11,086,754,658 cpu_core/TOPDOWN.SLOTS/ # 27.1 %
tma_backend_bound
# 7.5 %
tma_bad_speculation
# 36.5 %
tma_frontend_bound
# 28.9 %
tma_retiring
3,219,475,010 cpu_core/topdown-retiring/
820,655,931 cpu_core/topdown-bad-spec/
418,883,912 cpu_core/topdown-heavy-ops/
4,082,884,459 cpu_core/topdown-fe-bound/
3,012,532,414 cpu_core/topdown-be-bound/
1,030,171,196 cpu_atom/TOPDOWN_RETIRING.ALL/ # 17.4 %
tma_retiring
# 16.5 %
tma_bad_speculation
1,185,093,601 cpu_atom/CPU_CLK_UNHALTED.CORE/ # 29.8 %
tma_backend_bound
# 36.4 %
tma_frontend_bound
2,154,914,153 cpu_atom/TOPDOWN_FE_BOUND.ALL/
14,988,684 cpu_core/INT_MISC.UOP_DROPPING/
1,763,486,868 cpu_atom/TOPDOWN_BE_BOUND.ALL/
1.004103365 seconds time elapsed
```
ie the cpu_core and cpu_atom mixing of events isn't a regression
introduced here. There isn't a simple fix for the ordering, as we
don't want to mess up the non-metric cases. I'm happy if you think
things can be otherwise to make a change.
Thanks,
Ian