Several perf metrics topdown related fixes
cover Several perf metrics topdown related fixes
From: Kan Liang <[email protected]>
The Patch 1 is a follow-up patch for Ian's ("Fix topdown event weak grouping")[1].
The patch 2 is to fix the perf metrics topdown events in a mixed group. It reuses the function introduced in [1].
Patch 1 & 2 should be on top of [1].
The patch 3 & 4 are to fix other perf metrics topdown related issues. They can be merged separately.
[1]: https://lore.kernel.org/all/[email protected]/
Kan Liang (4):
perf evsel: Fixes topdown events in a weak group for the hybrid
platform
perf stat: Always keep perf metrics topdown events in a group perf parse-events: Support different format of the topdown event name perf parse-events: Move slots event for hybrid platforms too
tools/perf/arch/x86/util/evlist.c | 7 ++++--- tools/perf/arch/x86/util/evsel.c | 5 +++-- tools/perf/arch/x86/util/topdown.c | 18 ++++++++++++++++++ tools/perf/arch/x86/util/topdown.h | 7 +++++++ tools/perf/builtin-stat.c | 7 +++++-- 5 files changed, 37 insertions(+), 7 deletions(-) create mode 100644 tools/perf/arch/x86/util/topdown.h
Signature · 1 line
2.35.1
PATCH 1/4 perf evsel: Fixes topdown events in a weak group for the hybrid platform
From: Kan Liang <[email protected]>
The patch ("perf evlist: Keep topdown counters in weak group") fixes the perf metrics topdown event issue when the topdown events are in a weak group on a non-hybrid platform. However, it doesn't work for the hybrid platform.
$./perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
751,765,068 cpu_core/slots/ (84.07%)
<not supported> cpu_core/topdown-bad-spec/
<not supported> cpu_core/topdown-be-bound/
<not supported> cpu_core/topdown-fe-bound/
<not supported> cpu_core/topdown-retiring/
12,398,197 cpu_core/branch-instructions/ (84.07%)
1,054,218 cpu_core/branch-misses/ (84.24%)
539,764,637 cpu_core/bus-cycles/ (84.64%)
14,683 cpu_core/cache-misses/ (84.87%)
7,277,809 cpu_core/cache-references/ (77.30%)
222,299,439 cpu_core/cpu-cycles/ (77.28%)
63,661,714 cpu_core/instructions/ (84.85%)
0 cpu_core/mem-loads/ (77.29%)
12,271,725 cpu_core/mem-stores/ (77.30%)
542,241,102 cpu_core/ref-cycles/ (84.85%)
8,854 cpu_core/cache-misses/ (76.71%)
7,179,013 cpu_core/cache-references/ (76.31%)1.003245250 seconds time elapsed
A hybrid platform has a different PMU name for the core PMUs, while the current perf hard code the PMU name "cpu".
The evsel->pmu_name can be used to replace the "cpu" to fix the issue. For a hybrid platform, the pmu_name must be non-NULL. Because there are at least two core PMUs. The PMU has to be specified.
For a non-hybrid platform, the pmu_name may be NULL. Because there is only one core PMU, "cpu". For a NULL pmu_name, we can safely assume that it is a "cpu" PMU.
With the patch,
$perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
766,620,266 cpu_core/slots/ (84.06%)
73,172,129 cpu_core/topdown-bad-spec/ # 9.5% bad speculation (84.06%)
193,443,341 cpu_core/topdown-be-bound/ # 25.0% backend bound (84.06%)
403,940,929 cpu_core/topdown-fe-bound/ # 52.3% frontend bound (84.06%)
102,070,237 cpu_core/topdown-retiring/ # 13.2% retiring (84.06%)
12,364,429 cpu_core/branch-instructions/ (84.03%)
1,080,124 cpu_core/branch-misses/ (84.24%)
564,120,383 cpu_core/bus-cycles/ (84.65%)
36,979 cpu_core/cache-misses/ (84.86%)
7,298,094 cpu_core/cache-references/ (77.30%)
227,174,372 cpu_core/cpu-cycles/ (77.31%)
63,886,523 cpu_core/instructions/ (84.87%)
0 cpu_core/mem-loads/ (77.31%)
12,208,782 cpu_core/mem-stores/ (77.31%)
566,409,738 cpu_core/ref-cycles/ (84.87%)
23,118 cpu_core/cache-misses/ (76.71%)
7,212,602 cpu_core/cache-references/ (76.29%)1.003228667 seconds time elapsed
--- tools/perf/arch/x86/util/evsel.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
tools/perf/arch/x86/util/evsel.c +3 -2
@@ -33,8 +33,9 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if ((evsel->pmu_name &&strcmp (evsel->pmu_name,"cpu" )) || - !pmu_have_event ("cpu" ,"slots" )) +const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; + +if (!pmu_have_event (pmu_name,"slots" ))return false; return evsel->name &&
Signature · 1 line
2.35.1
Ian Rogers
reviewerRe: [PATCH 1/4] perf evsel: Fixes topdown events in a weak group for the hybrid platform
Quoted 90 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The patch ("perf evlist: Keep topdown counters in weak group") fixes the perf metrics topdown event issue when the topdown events are in a weak group on a non-hybrid platform. However, it doesn't work for the hybrid platform.
$./perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
751,765,068 cpu_core/slots/ (84.07%) <not supported> cpu_core/topdown-bad-spec/ <not supported> cpu_core/topdown-be-bound/ <not supported> cpu_core/topdown-fe-bound/ <not supported> cpu_core/topdown-retiring/ 12,398,197 cpu_core/branch-instructions/ (84.07%) 1,054,218 cpu_core/branch-misses/ (84.24%) 539,764,637 cpu_core/bus-cycles/ (84.64%) 14,683 cpu_core/cache-misses/ (84.87%) 7,277,809 cpu_core/cache-references/ (77.30%) 222,299,439 cpu_core/cpu-cycles/ (77.28%) 63,661,714 cpu_core/instructions/ (84.85%) 0 cpu_core/mem-loads/ (77.29%) 12,271,725 cpu_core/mem-stores/ (77.30%) 542,241,102 cpu_core/ref-cycles/ (84.85%) 8,854 cpu_core/cache-misses/ (76.71%) 7,179,013 cpu_core/cache-references/ (76.31%)1.003245250 seconds time elapsedA hybrid platform has a different PMU name for the core PMUs, while the current perf hard code the PMU name "cpu".
The
evsel->pmu_namecan be used to replace the "cpu" to fix the issue. For a hybrid platform, thepmu_namemust be non-NULL. Because there are at least two core PMUs. The PMU has to be specified.
For a non-hybrid platform, thepmu_namemay be NULL. Because there is only one core PMU, "cpu". For a NULLpmu_name, we can safely assume that it is a "cpu" PMU.With the patch,
$perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
766,620,266 cpu_core/slots/ (84.06%) 73,172,129 cpu_core/topdown-bad-spec/ # 9.5% bad speculation (84.06%) 193,443,341 cpu_core/topdown-be-bound/ # 25.0% backend bound (84.06%) 403,940,929 cpu_core/topdown-fe-bound/ # 52.3% frontend bound (84.06%) 102,070,237 cpu_core/topdown-retiring/ # 13.2% retiring (84.06%) 12,364,429 cpu_core/branch-instructions/ (84.03%) 1,080,124 cpu_core/branch-misses/ (84.24%) 564,120,383 cpu_core/bus-cycles/ (84.65%) 36,979 cpu_core/cache-misses/ (84.86%) 7,298,094 cpu_core/cache-references/ (77.30%) 227,174,372 cpu_core/cpu-cycles/ (77.31%) 63,886,523 cpu_core/instructions/ (84.87%) 0 cpu_core/mem-loads/ (77.31%) 12,208,782 cpu_core/mem-stores/ (77.31%) 566,409,738 cpu_core/ref-cycles/ (84.87%) 23,118 cpu_core/cache-misses/ (76.71%) 7,212,602 cpu_core/cache-references/ (76.29%)1.003228667 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/arch/x86/util/evsel.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)tools/perf/arch/x86/util/evsel.c +0
-0
tools/perf/arch/x86/util/evsel.c void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) · line 38
Quoted 8 lines
tools/perf/arch/x86/util/evsel.c +3
-2@@ -33,8 +33,9 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if ((evsel->pmu_name &&strcmp (evsel->pmu_name,"cpu" )) || - !pmu_have_event ("cpu" ,"slots" )) +const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; + +if (!pmu_have_event (pmu_name,"slots" ))
Playing devil's advocate, if I have a PMU for my network accelerator and it has an event called "slots" then this test will also be true.
The property that is being tested here is "does this CPU have topdown events" and so allowing any PMU removes the "does this CPU" part of the equation. I think ideally we'd have an arch functions something like:
bool arch_pmu__has_intel_topdown_events(void)
{
static bool has_topdown_events = pmu_have_event("cpu", "slots") ||pmu_have_event("cpu_core", "slots");
return has_topdown_events;
}
bool arch_pmu__supports_intel_topdown_events(const char *pmu_name) {
if (!pmu_name)
return false;
return arch_pmu__has_intel_topdown_events() && (!strncmp(pmu_name,
"cpu") || !strncmp(pmu_name, "cpu_core"));
}
bool arch_evsel__is_intel_topdown_event(struct evsel *evsel) {
if (!arch_pmu__supports_intel_topdown_events(evsel->pmu))
return false;
return strcasestr(evsel->name, "slots") || strcasestr(evsel->name, "topdown");
}
This then gives us:
bool arch_evsel__must_be_in_group(const struct evsel *evsel) {
return arch_evsel__is_intel_topdown_event(evsel);
}
These functions can then be reused for the arch_evlist topdown code, etc. What I don't see in these functions is use of any hybrid abstraction and so it isn't clear to me how with hybrid something like this would be plumbed in.
Thanks,
Ian
Quoted 6 lines
tools/perf/arch/x86/util/evsel.c +0
-1return false; return evsel->name && --2.35.1
Re: [PATCH 1/4] perf evsel: Fixes topdown events in a weak group for the hybrid platform
Quoted 92 lines · Ian Rogers
Quoted 91 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The patch ("perf evlist: Keep topdown counters in weak group") fixes the perf metrics topdown event issue when the topdown events are in a weak group on a non-hybrid platform. However, it doesn't work for the hybrid platform.
$./perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
751,765,068 cpu_core/slots/ (84.07%) <not supported> cpu_core/topdown-bad-spec/ <not supported> cpu_core/topdown-be-bound/ <not supported> cpu_core/topdown-fe-bound/ <not supported> cpu_core/topdown-retiring/ 12,398,197 cpu_core/branch-instructions/ (84.07%) 1,054,218 cpu_core/branch-misses/ (84.24%) 539,764,637 cpu_core/bus-cycles/ (84.64%) 14,683 cpu_core/cache-misses/ (84.87%) 7,277,809 cpu_core/cache-references/ (77.30%) 222,299,439 cpu_core/cpu-cycles/ (77.28%) 63,661,714 cpu_core/instructions/ (84.85%) 0 cpu_core/mem-loads/ (77.29%) 12,271,725 cpu_core/mem-stores/ (77.30%) 542,241,102 cpu_core/ref-cycles/ (84.85%) 8,854 cpu_core/cache-misses/ (76.71%) 7,179,013 cpu_core/cache-references/ (76.31%)1.003245250 seconds time elapsedA hybrid platform has a different PMU name for the core PMUs, while the current perf hard code the PMU name "cpu".
The
evsel->pmu_namecan be used to replace the "cpu" to fix the issue. For a hybrid platform, thepmu_namemust be non-NULL. Because there are at least two core PMUs. The PMU has to be specified.
For a non-hybrid platform, thepmu_namemay be NULL. Because there is only one core PMU, "cpu". For a NULLpmu_name, we can safely assume that it is a "cpu" PMU.With the patch,
$perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
766,620,266 cpu_core/slots/ (84.06%) 73,172,129 cpu_core/topdown-bad-spec/ # 9.5% bad speculation (84.06%) 193,443,341 cpu_core/topdown-be-bound/ # 25.0% backend bound (84.06%) 403,940,929 cpu_core/topdown-fe-bound/ # 52.3% frontend bound (84.06%) 102,070,237 cpu_core/topdown-retiring/ # 13.2% retiring (84.06%) 12,364,429 cpu_core/branch-instructions/ (84.03%) 1,080,124 cpu_core/branch-misses/ (84.24%) 564,120,383 cpu_core/bus-cycles/ (84.65%) 36,979 cpu_core/cache-misses/ (84.86%) 7,298,094 cpu_core/cache-references/ (77.30%) 227,174,372 cpu_core/cpu-cycles/ (77.31%) 63,886,523 cpu_core/instructions/ (84.87%) 0 cpu_core/mem-loads/ (77.31%) 12,208,782 cpu_core/mem-stores/ (77.31%) 566,409,738 cpu_core/ref-cycles/ (84.87%) 23,118 cpu_core/cache-misses/ (76.71%) 7,212,602 cpu_core/cache-references/ (76.29%)1.003228667 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/arch/x86/util/evsel.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)tools/perf/arch/x86/util/evsel.c +0
-0@@ -33,8 +33,9 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)
Quoted 11 lines
tools/perf/arch/x86/util/evsel.c void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) · line 38
Quoted 7 lines
tools/perf/arch/x86/util/evsel.c +3
-2bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if ((evsel->pmu_name &&strcmp (evsel->pmu_name,"cpu" )) || - !pmu_have_event ("cpu" ,"slots" )) +const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; + +if (!pmu_have_event (pmu_name,"slots" ))Playing devil's advocate, if I have a PMU for my network accelerator and it has an event called "slots" then this test will also be true.
IIRC, the pmu_have_event should only check the event which is exposed by the kernel. It's very unlikely that another PMU expose the exact same name.
If you still worry about it, I think we can check the PMU type PERF_TYPE_RAW here, which is reserved for the core PMU. Others cannot use it.
It looks like arch_evsel__must_be_in_group() is the only user for the evsel__sys_has_perf_metrics() for now, so I make it static.
The other pmu_have_event("cpu", "slots") is in evlist.c. topdown_sys_has_perf_metrics() in patch 4 should be used to replace it. I think Zhengjun will post patches for the changes for the evlist.c
tools/perf/arch/x86/util/evsel.c +11 -2
@@ -31,11 +31,20 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)free (env.cpuid); } -bool arch_evsel__must_be_in_group (const struct evsel *evsel) +static bool evsel__sys_has_perf_metrics (const struct evsel *evsel) {const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; -if (!pmu_have_event (pmu_name,"slots" )) +if ((evsel->core.attr.type ==PERF_TYPE_RAW ) && +pmu_have_event (pmu_name,"slots" )) +return true; + +return false; +} + +bool arch_evsel__must_be_in_group (const struct evsel *evsel) +{ +if (!evsel__sys_has_perf_metrics (evsel))return false; return evsel->name &&
Thanks,
Kan
Quoted 50 lines
The property that is being tested here is "does this CPU have topdown events" and so allowing any PMU removes the "does this CPU" part of the equation. I think ideally we'd have an arch functions something like:
bool
arch_pmu__has_intel_topdown_events(void){ static bool has_topdown_events = pmu_have_event("cpu", "slots") ||
pmu_have_event("cpu_core", "slots");return has_topdown_events; }bool
arch_pmu__supports_intel_topdown_events(const char *pmu_name){ if (!pmu_name) return false; return arch_pmu__has_intel_topdown_events() && (!strncmp(pmu_name,"cpu") || !strncmp(pmu_name, "
cpu_core"));
}bool
arch_evsel__is_intel_topdown_event(struct evsel *evsel){ if (!arch_pmu__supports_intel_topdown_events(evsel->pmu)) return false; return strcasestr(evsel->name, "slots") || strcasestr(evsel->name, "topdown"); }This then gives us:
bool
arch_evsel__must_be_in_group(const struct evsel *evsel){ return arch_evsel__is_intel_topdown_event(evsel); }These functions can then be reused for the
arch_evlisttopdown code, etc. What I don't see in these functions is use of any hybrid abstraction and so it isn't clear to me how with hybrid something like this would be plumbed in.Thanks,
IanQuoted 6 lines
tools/perf/arch/x86/util/evsel.c +0
-1return false; return evsel->name && --2.35.1
Ian Rogers
reviewerRe: [PATCH 1/4] perf evsel: Fixes topdown events in a weak group for the hybrid platform
Quoted 108 lines · Kan
Quoted 92 lines · Ian Rogers
Quoted 91 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The patch ("perf evlist: Keep topdown counters in weak group") fixes the perf metrics topdown event issue when the topdown events are in a weak group on a non-hybrid platform. However, it doesn't work for the hybrid platform.
$./perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
751,765,068 cpu_core/slots/ (84.07%) <not supported> cpu_core/topdown-bad-spec/ <not supported> cpu_core/topdown-be-bound/ <not supported> cpu_core/topdown-fe-bound/ <not supported> cpu_core/topdown-retiring/ 12,398,197 cpu_core/branch-instructions/ (84.07%) 1,054,218 cpu_core/branch-misses/ (84.24%) 539,764,637 cpu_core/bus-cycles/ (84.64%) 14,683 cpu_core/cache-misses/ (84.87%) 7,277,809 cpu_core/cache-references/ (77.30%) 222,299,439 cpu_core/cpu-cycles/ (77.28%) 63,661,714 cpu_core/instructions/ (84.85%) 0 cpu_core/mem-loads/ (77.29%) 12,271,725 cpu_core/mem-stores/ (77.30%) 542,241,102 cpu_core/ref-cycles/ (84.85%) 8,854 cpu_core/cache-misses/ (76.71%) 7,179,013 cpu_core/cache-references/ (76.31%)1.003245250 seconds time elapsedA hybrid platform has a different PMU name for the core PMUs, while the current perf hard code the PMU name "cpu".
The
evsel->pmu_namecan be used to replace the "cpu" to fix the issue. For a hybrid platform, thepmu_namemust be non-NULL. Because there are at least two core PMUs. The PMU has to be specified.
For a non-hybrid platform, thepmu_namemay be NULL. Because there is only one core PMU, "cpu". For a NULLpmu_name, we can safely assume that it is a "cpu" PMU.With the patch,
$perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
766,620,266 cpu_core/slots/ (84.06%) 73,172,129 cpu_core/topdown-bad-spec/ # 9.5% bad speculation (84.06%) 193,443,341 cpu_core/topdown-be-bound/ # 25.0% backend bound (84.06%) 403,940,929 cpu_core/topdown-fe-bound/ # 52.3% frontend bound (84.06%) 102,070,237 cpu_core/topdown-retiring/ # 13.2% retiring (84.06%) 12,364,429 cpu_core/branch-instructions/ (84.03%) 1,080,124 cpu_core/branch-misses/ (84.24%) 564,120,383 cpu_core/bus-cycles/ (84.65%) 36,979 cpu_core/cache-misses/ (84.86%) 7,298,094 cpu_core/cache-references/ (77.30%) 227,174,372 cpu_core/cpu-cycles/ (77.31%) 63,886,523 cpu_core/instructions/ (84.87%) 0 cpu_core/mem-loads/ (77.31%) 12,208,782 cpu_core/mem-stores/ (77.31%) 566,409,738 cpu_core/ref-cycles/ (84.87%) 23,118 cpu_core/cache-misses/ (76.71%) 7,212,602 cpu_core/cache-references/ (76.29%)1.003228667 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/arch/x86/util/evsel.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)tools/perf/arch/x86/util/evsel.c +0
-0@@ -33,8 +33,9 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)Quoted 11 lines
tools/perf/arch/x86/util/evsel.c void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) · line 38
Quoted 7 lines
tools/perf/arch/x86/util/evsel.c +3
-2bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if ((evsel->pmu_name &&strcmp (evsel->pmu_name,"cpu" )) || - !pmu_have_event ("cpu" ,"slots" )) +const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; + +if (!pmu_have_event (pmu_name,"slots" ))Playing devil's advocate, if I have a PMU for my network accelerator and it has an event called "slots" then this test will also be true.
Quoted 6 lines
IIRC, the
pmu_have_eventshould only check the event which is exposed by the kernel. It's very unlikely that another PMU expose the exact same name.If you still worry about it, I think we can check the PMU type
PERF_TYPE_RAWhere, which is reserved for the core PMU. Others cannot use it.
That's cool, this isn't documented behavior though:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/include/uapi/linux/perf_event.h?h=perf/core#n34
and PERF_TYPE_HARDWARE wouldn't seem a wholly unreasonable type. It kind of feels like depending on a quirk, and so we should bury the quirk in a helper function and document it :-)
Quoted 6 lines
It looks like
arch_evsel__must_be_in_group()is the only user for theevsel__sys_has_perf_metrics()for now, so I make it static.The other
pmu_have_event("cpu", "slots")is inevlist.c.topdown_sys_has_perf_metrics()in patch 4 should be used to replace it. I think Zhengjun will post patches for the changes for theevlist.c
Ok, is Zhengjun putting his changes on top of this and fixing up the APIs or is he waiting on these changes landing? Let me know how to help. I'm guessing landing my changes is the first step.
Thanks,
Ian
Quoted 83 lines
tools/perf/arch/x86/util/evsel.c +11
-2@@ -31,11 +31,20 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) free(env.cpuid); } -bool arch_evsel__must_be_in_group (const struct evsel *evsel) +static bool evsel__sys_has_perf_metrics (const struct evsel *evsel) {const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; -if (!pmu_have_event (pmu_name,"slots" )) +if ((evsel->core.attr.type ==PERF_TYPE_RAW ) && +pmu_have_event (pmu_name,"slots" )) +return true; + +return false; +} + +bool arch_evsel__must_be_in_group (const struct evsel *evsel) +{ +if (!evsel__sys_has_perf_metrics (evsel))return false;return evsel->name &&Thanks,
KanQuoted 50 lines
The property that is being tested here is "does this CPU have topdown events" and so allowing any PMU removes the "does this CPU" part of the equation. I think ideally we'd have an arch functions something like:
bool
arch_pmu__has_intel_topdown_events(void){ static bool has_topdown_events = pmu_have_event("cpu", "slots") ||
pmu_have_event("cpu_core", "slots");return has_topdown_events; }bool
arch_pmu__supports_intel_topdown_events(const char *pmu_name){ if (!pmu_name) return false; return arch_pmu__has_intel_topdown_events() && (!strncmp(pmu_name,"cpu") || !strncmp(pmu_name, "
cpu_core"));
}bool
arch_evsel__is_intel_topdown_event(struct evsel *evsel){ if (!arch_pmu__supports_intel_topdown_events(evsel->pmu)) return false; return strcasestr(evsel->name, "slots") || strcasestr(evsel->name, "topdown"); }This then gives us:
bool
arch_evsel__must_be_in_group(const struct evsel *evsel){ return arch_evsel__is_intel_topdown_event(evsel); }These functions can then be reused for the
arch_evlisttopdown code, etc. What I don't see in these functions is use of any hybrid abstraction and so it isn't clear to me how with hybrid something like this would be plumbed in.Thanks,
IanQuoted 6 lines
tools/perf/arch/x86/util/evsel.c +0
-1return false; return evsel->name && --2.35.1
Re: [PATCH 1/4] perf evsel: Fixes topdown events in a weak group for the hybrid platform
Quoted 116 lines · Ian Rogers
Quoted 108 lines · Kan
Quoted 92 lines · Ian Rogers
Quoted 91 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The patch ("perf evlist: Keep topdown counters in weak group") fixes the perf metrics topdown event issue when the topdown events are in a weak group on a non-hybrid platform. However, it doesn't work for the hybrid platform.
$./perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide':
751,765,068 cpu_core/slots/ (84.07%) <not supported> cpu_core/topdown-bad-spec/ <not supported> cpu_core/topdown-be-bound/ <not supported> cpu_core/topdown-fe-bound/ <not supported> cpu_core/topdown-retiring/ 12,398,197 cpu_core/branch-instructions/ (84.07%) 1,054,218 cpu_core/branch-misses/ (84.24%) 539,764,637 cpu_core/bus-cycles/ (84.64%) 14,683 cpu_core/cache-misses/ (84.87%) 7,277,809 cpu_core/cache-references/ (77.30%) 222,299,439 cpu_core/cpu-cycles/ (77.28%) 63,661,714 cpu_core/instructions/ (84.85%) 0 cpu_core/mem-loads/ (77.29%) 12,271,725 cpu_core/mem-stores/ (77.30%) 542,241,102 cpu_core/ref-cycles/ (84.85%) 8,854 cpu_core/cache-misses/ (76.71%) 7,179,013 cpu_core/cache-references/ (76.31%)1.003245250 seconds time elapsedA hybrid platform has a different PMU name for the core PMUs, while the current perf hard code the PMU name "cpu".
The
evsel->pmu_namecan be used to replace the "cpu" to fix the issue. For a hybrid platform, thepmu_namemust be non-NULL. Because there are at least two core PMUs. The PMU has to be specified.
For a non-hybrid platform, thepmu_namemay be NULL. Because there is only one core PMU, "cpu". For a NULLpmu_name, we can safely assume that it is a "cpu" PMU.With the patch,
$perf stat -e '{cpu_core/slots/,cpu_core/topdown-bad-spec/, cpu_core/topdown-be-bound/,cpu_core/topdown-fe-bound/, cpu_core/topdown-retiring/,cpu_core/branch-instructions/, cpu_core/branch-misses/,cpu_core/bus-cycles/,cpu_core/cache-misses/, cpu_core/cache-references/,cpu_core/cpu-cycles/,cpu_core/instructions/, cpu_core/mem-loads/,cpu_core/mem-stores/,cpu_core/ref-cycles/, cpu_core/cache-misses/,cpu_core/cache-references/}:W' -a sleep 1
Performance counter stats for 'system wide': 766,620,266 cpu_core/slots/ (84.06%) 73,172,129 cpu_core/topdown-bad-spec/ # 9.5% bad speculation (84.06%) 193,443,341 cpu_core/topdown-be-bound/ # 25.0% backend bound (84.06%) 403,940,929 cpu_core/topdown-fe-bound/ # 52.3% frontend bound (84.06%) 102,070,237 cpu_core/topdown-retiring/ # 13.2% retiring (84.06%) 12,364,429 cpu_core/branch-instructions/ (84.03%) 1,080,124 cpu_core/branch-misses/ (84.24%) 564,120,383 cpu_core/bus-cycles/ (84.65%) 36,979 cpu_core/cache-misses/ (84.86%) 7,298,094 cpu_core/cache-references/ (77.30%) 227,174,372 cpu_core/cpu-cycles/ (77.31%) 63,886,523 cpu_core/instructions/ (84.87%) 0 cpu_core/mem-loads/ (77.31%) 12,208,782 cpu_core/mem-stores/ (77.31%) 566,409,738 cpu_core/ref-cycles/ (84.87%) 23,118 cpu_core/cache-misses/ (76.71%) 7,212,602 cpu_core/cache-references/ (76.29%) 1.003228667 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/arch/x86/util/evsel.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)tools/perf/arch/x86/util/evsel.c +0
-0@@ -33,8 +33,9 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr)Quoted 11 lines
tools/perf/arch/x86/util/evsel.c void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) · line 38
Quoted 7 lines
tools/perf/arch/x86/util/evsel.c +3
-2bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if ((evsel->pmu_name &&strcmp (evsel->pmu_name,"cpu" )) || - !pmu_have_event ("cpu" ,"slots" )) +const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; + +if (!pmu_have_event (pmu_name,"slots" ))Playing devil's advocate, if I have a PMU for my network accelerator and it has an event called "slots" then this test will also be true.
Quoted 6 lines
IIRC, the
pmu_have_eventshould only check the event which is exposed by the kernel. It's very unlikely that another PMU expose the exact same name.If you still worry about it, I think we can check the PMU type
PERF_TYPE_RAWhere, which is reserved for the core PMU. Others cannot use it.
Quoted 5 lines
That's cool, this isn't documented behavior though:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/include/uapi/linux/perf_event.h?h=perf/core#n34
andPERF_TYPE_HARDWAREwouldn't seem a wholly unreasonable type. It kind of feels like depending on a quirk, and so we should bury the quirk in a helper function and document it :-)
The PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE are aliases for the PERF_TYPE_RAW. The PERF_TYPE_HARDWARE is used for the 10 common hardware events and The PERF_TYPE_HW_CACHE is used for the hardware cache events. Other core events (not include the atom core events in a hybrid machine) should have the PERF_TYPE_RAW type.
Since the perf metrics is a big core only feature, checking both the PERF_TYPE_RAW type and the slots event should be good enough here.
I will add more comments in V2.
Quoted 11 lines
Quoted 7 lines
It looks like
arch_evsel__must_be_in_group()is the only user for theevsel__sys_has_perf_metrics()for now, so I make it static.The other
pmu_have_event("cpu", "slots")is inevlist.c.topdown_sys_has_perf_metrics()in patch 4 should be used to replace it. I think Zhengjun will post patches for the changes for theevlist.cOk, is Zhengjun putting his changes on top of this and fixing up the APIs or is he waiting on these changes landing? Let me know how to help. I'm guessing landing my changes is the first step.
Right. Your changes is the first step. Then this patch set. Zhengjun's will be on top of us.
Thanks,
Kan
Quoted 87 lines
Thanks,
IanQuoted 83 lines
tools/perf/arch/x86/util/evsel.c +11
-2@@ -31,11 +31,20 @@ void arch_evsel__fixup_new_cycles(struct perf_event_attr *attr) free(env.cpuid); } -bool arch_evsel__must_be_in_group (const struct evsel *evsel) +static bool evsel__sys_has_perf_metrics (const struct evsel *evsel) {const char *pmu_name = evsel->pmu_name ? evsel->pmu_name :"cpu" ; -if (!pmu_have_event (pmu_name,"slots" )) +if ((evsel->core.attr.type ==PERF_TYPE_RAW ) && +pmu_have_event (pmu_name,"slots" )) +return true; + +return false; +} + +bool arch_evsel__must_be_in_group (const struct evsel *evsel) +{ +if (!evsel__sys_has_perf_metrics (evsel))return false;return evsel->name &&Thanks,
KanQuoted 50 lines
The property that is being tested here is "does this CPU have topdown events" and so allowing any PMU removes the "does this CPU" part of the equation. I think ideally we'd have an arch functions something like:
bool
arch_pmu__has_intel_topdown_events(void){ static bool has_topdown_events = pmu_have_event("cpu", "slots") ||
pmu_have_event("cpu_core", "slots");return has_topdown_events; }bool
arch_pmu__supports_intel_topdown_events(const char *pmu_name){ if (!pmu_name) return false; return arch_pmu__has_intel_topdown_events() && (!strncmp(pmu_name,"cpu") || !strncmp(pmu_name, "
cpu_core"));
}bool
arch_evsel__is_intel_topdown_event(struct evsel *evsel){ if (!arch_pmu__supports_intel_topdown_events(evsel->pmu)) return false; return strcasestr(evsel->name, "slots") || strcasestr(evsel->name, "topdown"); }This then gives us:
bool
arch_evsel__must_be_in_group(const struct evsel *evsel){ return arch_evsel__is_intel_topdown_event(evsel); }These functions can then be reused for the
arch_evlisttopdown code, etc. What I don't see in these functions is use of any hybrid abstraction and so it isn't clear to me how with hybrid something like this would be plumbed in.Thanks,
IanQuoted 6 lines
tools/perf/arch/x86/util/evsel.c +0
-1return false; return evsel->name && --2.35.1
PATCH 2/4 perf stat: Always keep perf metrics topdown events in a group
In reply to [email protected]From: Kan Liang <[email protected]>
If any member in a group has a different cpu mask than the other members, the current perf stat disables group. when the perf metrics topdown events are part of the group, the below <not supported> error will be triggered.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1
WARNING: grouped events cpus do not match, disabling group:
anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }Performance counter stats for 'system wide':
141,465,174 slots
<not supported> topdown-retiring
1,605,330,334 uncore_imc_free_running_0/dclk/The perf metrics topdown events must always be grouped with a slots event as leader.
With the patch, the topdown events aren't broken from the group for the splitting.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1
WARNING: grouped events cpus do not match, disabling group:
anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }Performance counter stats for 'system wide':
346,110,588 slots
124,608,256 topdown-retiring
1,606,869,976 uncore_imc_free_running_0/dclk/1.003877592 seconds time elapsed
--- tools/perf/builtin-stat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +5 -2
@@ -272,8 +272,11 @@ static void evlist__check_cpu_maps(struct evlist *evlist) }for_each_group_evsel (pos, leader) { -evsel__set_leader (pos, pos); - pos->core.nr_members =0 ; +if (!evsel__must_be_in_group (pos) && pos != leader) { +evsel__set_leader (pos, pos); + pos->core.nr_members =0 ; + leader->core.nr_members--; + } } evsel->core.leader->nr_members =0 ; }
Signature · 1 line
2.35.1
Ian Rogers
reviewerRe: [PATCH 2/4] perf stat: Always keep perf metrics topdown events in a group
Quoted 28 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
If any member in a group has a different cpu mask than the other members, the current perf stat disables group. when the perf metrics topdown events are part of the group, the below <not supported> error will be triggered.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1 WARNING: grouped events cpus do not match, disabling group: anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }Performance counter stats for 'system wide':
141,465,174 slots <not supported> topdown-retiring 1,605,330,334 uncore_imc_free_running_0/dclk/The perf metrics topdown events must always be grouped with a slots event as leader.
With the patch, the topdown events aren't broken from the group for the splitting.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1 WARNING: grouped events cpus do not match, disabling group: anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }
Quoted 7 lines
Performance counter stats for 'system wide':
346,110,588 slots 124,608,256 topdown-retiring 1,606,869,976 uncore_imc_free_running_0/dclk/1.003877592 seconds time elapsed
Nice! This is based on:
https://lore.kernel.org/lkml/[email protected]/
You may end up with a group with the leader having a group count of 1 (itself). I explicitly zeroed that in the change above, but this may be unnecessary. Maybe we should move this code to helper functions for sharing and consistency on what the leader count should be.
Thanks,
Ian
Quoted 27 lines
Fixes a9a1790247bd ("perf stat: Ensure group is defined on top of the same cpu mask")Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-stat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)tools/perf/builtin-stat.c +5
-3@@ -272,8 +272,11 @@ static void evlist__check_cpu_maps(struct evlist *evlist) }for_each_group_evsel (pos, leader) { -evsel__set_leader (pos, pos); - pos->core.nr_members =0 ; +if (!evsel__must_be_in_group (pos) && pos != leader) { +evsel__set_leader (pos, pos); + pos->core.nr_members =0 ; + leader->core.nr_members--; + } } evsel->core.leader->nr_members =0 ; } --2.35.1
Re: [PATCH 2/4] perf stat: Always keep perf metrics topdown events in a group
Quoted 35 lines · Ian Rogers
Quoted 34 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
If any member in a group has a different cpu mask than the other members, the current perf stat disables group. when the perf metrics topdown events are part of the group, the below <not supported> error will be triggered.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1 WARNING: grouped events cpus do not match, disabling group: anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }Performance counter stats for 'system wide':
141,465,174 slots <not supported> topdown-retiring 1,605,330,334 uncore_imc_free_running_0/dclk/The perf metrics topdown events must always be grouped with a slots event as leader.
With the patch, the topdown events aren't broken from the group for the splitting.
$ perf stat -e "{slots,topdown-retiring,uncore_imc_free_running_0/dclk/}" -a sleep 1 WARNING: grouped events cpus do not match, disabling group: anon group { slots, topdown-retiring, uncore_imc_free_running_0/dclk/ }Performance counter stats for 'system wide':
346,110,588 slots 124,608,256 topdown-retiring 1,606,869,976 uncore_imc_free_running_0/dclk/
Quoted 9 lines
Quoted 1 line
1.003877592 seconds time elapsedNice! This is based on:
https://lore.kernel.org/lkml/[email protected]/
You may end up with a group with the leader having a group count of 1 (itself). I explicitly zeroed that in the change above, but this may be unnecessary. Maybe we should move this code to helper functions for sharing and consistency on what the leader count should be.
I think the current code has already did evsel->core.leader->nr_members = 0; for the leader. So I don't change it.
Yes, a helper function seems reasonable. I will add it in V2.
Thanks,
Kan
Quoted 30 lines
Thanks,
IanQuoted 27 lines
Fixes a9a1790247bd ("perf stat: Ensure group is defined on top of the same cpu mask")Signed-off-by Kan Liang <[email protected]>--- tools/perf/builtin-stat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)tools/perf/builtin-stat.c +5
-3@@ -272,8 +272,11 @@ static void evlist__check_cpu_maps(struct evlist *evlist) }for_each_group_evsel (pos, leader) { -evsel__set_leader (pos, pos); - pos->core.nr_members =0 ; +if (!evsel__must_be_in_group (pos) && pos != leader) { +evsel__set_leader (pos, pos); + pos->core.nr_members =0 ; + leader->core.nr_members--; + } } evsel->core.leader->nr_members =0 ; } --2.35.1
PATCH 3/4 perf parse-events: Support different format of the topdown event name
In reply to [email protected]From: Kan Liang <[email protected]>
The evsel->name may have a different format for a topdown event, a pure topdown name (e.g., topdown-fe-bound), or a PMU name + a topdown name (e.g., cpu/topdown-fe-bound/). The cpu/topdown-fe-bound/ kind format isn't supported by the arch_evlist__leader(). This format is a very common format for a hybrid platform, which requires specifying the PMU name for each event.
Without the patch,
$perf stat -e '{instructions,slots,cpu/topdown-fe-bound/}' -a sleep 1
Performance counter stats for 'system wide':
<not counted> instructions
<not counted> slots
<not supported> cpu/topdown-fe-bound/1.003482041 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
The events in group usually have to be from the same PMU. Try reorganizing the group.
With the patch,
perf stat -e '{instructions,slots,cpu/topdown-fe-bound/}' -a sleep 1
Performance counter stats for 'system wide':
157,383,996 slots
25,011,711 instructions
27,441,686 cpu/topdown-fe-bound/1.003530890 seconds time elapsed
--- tools/perf/arch/x86/util/evlist.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/arch/x86/util/evlist.c +1 -1
@@ -36,7 +36,7 @@ struct evsel *arch_evlist__leader(struct list_head *list)if (slots == first)return first; } -if (!strncasecmp (evsel->name,"topdown" ,7 )) +if (strcasestr (evsel->name,"topdown" )) has_topdown = true;if (slots && has_topdown)return slots;
Signature · 1 line
2.35.1
Re: [PATCH 3/4] perf parse-events: Support different format of the topdown event name
Quoted 35 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The
evsel->namemay have a different format for a topdown event, a pure topdown name (e.g., topdown-fe-bound), or a PMU name + a topdown name (e.g., cpu/topdown-fe-bound/). The cpu/topdown-fe-bound/ kind format isn't supported by thearch_evlist__leader(). This format is a very common format for a hybrid platform, which requires specifying the PMU name for each event.Without the patch,
$perf stat -e '{instructions,slots,cpu/topdown-fe-bound/}' -a sleep 1
Performance counter stats for 'system wide':
<not counted> instructions <not counted> slots <not supported> cpu/topdown-fe-bound/1.003482041 seconds time elapsedSome 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
The events in group usually have to be from the same PMU. Try reorganizing the group.With the patch,
perf stat -e '{instructions,slots,cpu/topdown-fe-bound/}' -a sleep 1
Performance counter stats for 'system wide':
Quoted 8 lines
157,383,996 slots 25,011,711 instructions 27,441,686 cpu/topdown-fe-bound/1.003530890 seconds time elapsedFixes bc355822f0d9 ("perf parse-events: Move slots only with topdown")Signed-off-by Kan Liang <[email protected]>
Thanks,
Ian
Quoted 20 lines
--- tools/perf/arch/x86/util/evlist.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)tools/perf/arch/x86/util/evlist.c +1
-2@@ -36,7 +36,7 @@ struct evsel *arch_evlist__leader(struct list_head *list)if (slots == first)return first; } -if (!strncasecmp (evsel->name,"topdown" ,7 )) +if (strcasestr (evsel->name,"topdown" )) has_topdown = true;if (slots && has_topdown)return slots; --2.35.1
PATCH 4/4 perf parse-events: Move slots event for the hybrid platform too
In reply to [email protected]From: Kan Liang <[email protected]>
The commit 94dbfd6781a0 ("perf parse-events: Architecture specific leader override") introduced a feature to reorder the slots event to fulfill the restriction of the perf metrics topdown group. But the feature doesn't work on the hybrid machine.
$perf stat -e "{cpu_core/instructions/,cpu_core/slots/,cpu_core/topdown-retiring/}" -a sleep 1
Performance counter stats for 'system wide':
<not counted> cpu_core/instructions/
<not counted> cpu_core/slots/
<not supported> cpu_core/topdown-retiring/1.002871801 seconds time elapsed
A hybrid platform has a different PMU name for the core PMUs, while current perf hard code the PMU name "cpu".
Introduce a new function to check whether the system supports the perf metrics feature. The result is cached for the future usage.
For X86, the core PMU name always has "cpu" prefix.
With the patch,
$perf stat -e "{cpu_core/instructions/,cpu_core/slots/,cpu_core/topdown-retiring/}" -a sleep 1
Performance counter stats for 'system wide':
76,337,010 cpu_core/slots/
10,416,809 cpu_core/instructions/
11,692,372 cpu_core/topdown-retiring/1.002805453 seconds time elapsed
--- tools/perf/arch/x86/util/evlist.c | 5 +++-- tools/perf/arch/x86/util/topdown.c | 18 ++++++++++++++++++ tools/perf/arch/x86/util/topdown.h | 7 +++++++ 3 files changed, 28 insertions(+), 2 deletions(-) create mode 100644 tools/perf/arch/x86/util/topdown.h
tools/perf/arch/x86/util/evlist.c +3 -2
@@ -3,6 +3,7 @@#include "util/pmu.h" #include "util/evlist.h" #include "util/parse-events.h" +#include "topdown.h" #define TOPDOWN_L1_EVENTS "{slots,topdown-retiring,topdown-bad-spec,topdown-fe-bound,topdown-be-bound}"#define TOPDOWN_L2_EVENTS "{slots,topdown-retiring,topdown-bad-spec,topdown-fe-bound,topdown-be-bound,topdown-heavy-ops,topdown-br-mispredict,topdown-fetch-lat,topdown-mem-bound}"
@@ -25,12 +26,12 @@ struct evsel *arch_evlist__leader(struct list_head *list) first = list_first_entry(list, struct evsel, core.node); -if (!pmu_have_event ("cpu" ,"slots" )) +if (!topdown_sys_has_perf_metrics ())return first;/* If there is a slots event and a topdown event then the slots event comes first. */ __evlist__for_each_entry (list, evsel) { -if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && evsel->name) { +if (evsel->pmu_name && !strncmp (evsel->pmu_name,"cpu" ,3 ) && evsel->name) {if (strcasestr (evsel->name,"slots" )) { slots = evsel;if (slots == first)
tools/perf/arch/x86/util/topdown.c +18 -0
@@ -3,6 +3,24 @@#include "api/fs/fs.h" #include "util/pmu.h" #include "util/topdown.h" +#include "topdown.h" + +bool topdown_sys_has_perf_metrics (void ) +{ +static bool has_perf_metrics; +static bool cached; +struct perf_pmu *pmu; + +if (cached) +return has_perf_metrics; + + pmu =perf_pmu__find_by_type (PERF_TYPE_RAW ); +if (pmu &&pmu_have_event (pmu->name,"slots" )) + has_perf_metrics = true; + + cached = true; +return has_perf_metrics; +}/* * Check whether we can use a group for top down.
tools/perf/arch/x86/util/topdown.h new file +7 -0
@@ -0,0 +1,7 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _TOPDOWN_H +#define _TOPDOWN_H 1 + +bool topdown_sys_has_perf_metrics (void ); + +#endif
Signature · 1 line
2.35.1
Ian Rogers
reviewerRe: [PATCH 4/4] perf parse-events: Move slots event for the hybrid platform too
Quoted 82 lines · On Fri, May 13, 2022 at 8:16 AM <[email protected]> wrote:
From: Kan Liang <[email protected]>
The commit
94dbfd6781a0("perf parse-events: Architecture specific leader override") introduced a feature to reorder the slots event to fulfill the restriction of the perf metrics topdown group. But the feature doesn't work on the hybrid machine.$perf stat -e "{cpu_core/instructions/,cpu_core/slots/,cpu_core/topdown-retiring/}" -a sleep 1
Performance counter stats for 'system wide':
<not counted> cpu_core/instructions/ <not counted> cpu_core/slots/ <not supported> cpu_core/topdown-retiring/1.002871801 seconds time elapsedA hybrid platform has a different PMU name for the core PMUs, while current perf hard code the PMU name "cpu".
Introduce a new function to check whether the system supports the perf metrics feature. The result is cached for the future usage.
For X86, the core PMU name always has "cpu" prefix.
With the patch,
$perf stat -e "{cpu_core/instructions/,cpu_core/slots/,cpu_core/topdown-retiring/}" -a sleep 1
Performance counter stats for 'system wide':
76,337,010 cpu_core/slots/ 10,416,809 cpu_core/instructions/ 11,692,372 cpu_core/topdown-retiring/1.002805453 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 5 +++-- tools/perf/arch/x86/util/topdown.c | 18 ++++++++++++++++++ tools/perf/arch/x86/util/topdown.h | 7 +++++++ 3 files changed, 28 insertions(+), 2 deletions(-) create mode 100644 tools/perf/arch/x86/util/topdown.htools/perf/arch/x86/util/evlist.c +3
-2@@ -3,6 +3,7 @@#include "util/pmu.h" #include "util/evlist.h" #include "util/parse-events.h" +#include "topdown.h" #define TOPDOWN_L1_EVENTS "{slots,topdown-retiring,topdown-bad-spec,topdown-fe-bound,topdown-be-bound}"#define TOPDOWN_L2_EVENTS "{slots,topdown-retiring,topdown-bad-spec,topdown-fe-bound,topdown-be-bound,topdown-heavy-ops,topdown-br-mispredict,topdown-fetch-lat,topdown-mem-bound}"@@ -25,12 +26,12 @@ struct evsel *arch_evlist__leader(struct list_head *list) first = list_first_entry(list, struct evsel, core.node); -if (!pmu_have_event ("cpu" ,"slots" )) +if (!topdown_sys_has_perf_metrics ())return first;/* If there is a slots event and a topdown event then the slots event comes first. */ __evlist__for_each_entry (list, evsel) { -if (evsel->pmu_name && !strcmp (evsel->pmu_name,"cpu" ) && evsel->name) { +if (evsel->pmu_name && !strncmp (evsel->pmu_name,"cpu" ,3 ) && evsel->name) {if (strcasestr (evsel->name,"slots" )) { slots = evsel;if (slots == first)tools/perf/arch/x86/util/topdown.c +1
-0@@ -3,6 +3,24 @@#include "api/fs/fs.h" #include "util/pmu.h" #include "util/topdown.h" +#include "topdown.h"
tools/perf/arch/x86/util/topdown.c line 16
Quoted 10 lines
tools/perf/arch/x86/util/topdown.c +10
-0+ +bool topdown_sys_has_perf_metrics (void ) +{ +static bool has_perf_metrics; +static bool cached; +struct perf_pmu *pmu; + +if (cached) +return has_perf_metrics; +
Worth a comment here for the meaning of looking up PERF_TYPE_RAW.
Thanks,
Ian
Quoted 26 lines
tools/perf/arch/x86/util/topdown.c +7
-0+ pmu =perf_pmu__find_by_type (PERF_TYPE_RAW ); +if (pmu &&pmu_have_event (pmu->name,"slots" )) + has_perf_metrics = true; + + cached = true; +return has_perf_metrics; +}/* * Check whether we can use a group for top down. tools/perf/arch/x86/util/topdown.h new file +7
-1@@ -0,0 +1,7 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _TOPDOWN_H +#define _TOPDOWN_H 1 + +bool topdown_sys_has_perf_metrics (void ); + +#endif --2.35.1