perf stat: Fix uncore event mixed metric with workload error issue
PATCH 1/1 perf stat: Fix uncore event mixed metric with workload error issue
The metric mixed with uncore event is failed with workload.
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use -- sleep 1
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (uncore_imc/cas_count_read/). /bin/dmesg | grep -i perf may provide additional information.
But it works without workload.
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use
^C
Performance counter stats for 'system wide':
5.38 MiB uncore_imc/cas_count_read/
3.98 MiB uncore_imc/cas_count_write/
1550285007 ns duration_time1.550285007 seconds time elapsed
The uncore events (e.g. uncore_imc/cas_count_read/ and uncore_imc/cas_count_write/) have been forced to system wide but duration_time is not.
Currently, the target system wide is set according to:
1. there's no workload specified
2. there is workload specified but all requested events are system wide events.
For DRAM_BW_Use, the target is not set to system wide since 2 is not met. Next, the counter creation on CPU will run into some issues.
It should make sense to set the target system wide if there is workload specified and at least one required event is system wide event.
With this patch,
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use -- sleep 1
Performance counter stats for 'system wide':
4.81 MiB uncore_imc/cas_count_read/
2.89 MiB uncore_imc/cas_count_write/
1002019036 ns duration_time1.002019036 seconds time elapsed
--- tools/perf/builtin-stat.c | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-)
tools/perf/builtin-stat.c +8 -5
@@ -1841,8 +1841,8 @@ static void setup_system_wide(int forks)* conditions is met: * * - there's no workload specified -* - there is workload specified but all requested -* events are system wide events +* - there is workload specified but at least one requested +* event is system wide event */ if (!target__none (&target))return ;
@@ -1851,13 +1851,16 @@ static void setup_system_wide(int forks) target.system_wide = true; else {struct evsel *counter; +bool system_wide = false;evlist__for_each_entry (evsel_list, counter) { -if (!counter->core.system_wide) -return ; +if (counter->core.system_wide) { + system_wide = true; +break ; + } } -if (evsel_list->core.nr_entries) +if (evsel_list->core.nr_entries && system_wide) target.system_wide = true; } }
Signature · 1 line
2.17.1
On Mon, Apr 27, 2020 at 10:41:16PM +0800, Jin Yao wrote:
SNIP
Quoted 17 lines
index
9207b6c45475..b01ee06b1965100644tools/perf/builtin-stat.c +2
-2@@ -1841,8 +1841,8 @@ static void setup_system_wide(int forks)* conditions is met: * * - there's no workload specified -* - there is workload specified but all requested -* events are system wide events +* - there is workload specified but at least one requested +* event is system wide event */ if (!target__none (&target))return ;@@ -1851,13 +1851,16 @@ static void setup_system_wide(int forks) target.system_wide = true; else {
tools/perf/builtin-stat.c static void setup_system_wide(int forks) · line 1860
Quoted 10 lines
tools/perf/builtin-stat.c +5
-2struct evsel *counter; +bool system_wide = false;evlist__for_each_entry (evsel_list, counter) { -if (!counter->core.system_wide) -return ; +if (counter->core.system_wide) { + system_wide = true; +break ; + }
I wonder this would break some expectations.. would it be more safe to detect duration event and bypass it from the decission? but maybe the case I'm worried about is not a problem at all.. Andi?
jirka
Quoted 10 lines
tools/perf/builtin-stat.c +1
-1} -if (evsel_list->core.nr_entries) +if (evsel_list->core.nr_entries && system_wide) target.system_wide = true; } }Signature · 1 line
2.17.1
Quoted 4 lines
I wonder this would break some expectations.. would it be more safe to detect duration event and bypass it from the decission? but maybe the case I'm worried about is not a problem at all.. Andi?
Don't see what it would break.
Yes maybe we need to special case duration_time more, but that would be a much bigger patch.
-Andi
Quoted 9 lines · Andi Kleen
Quoted 4 lines
I wonder this would break some expectations.. would it be more safe to detect duration event and bypass it from the decission? but maybe the case I'm worried about is not a problem at all.. Andi?
Don't see what it would break.
Yes maybe we need to special case
duration_timemore, but that would be a much bigger patch.
would below change work for you? if duration_time is the only case, I'd rather go with the special case for it
jirka
---
tools/perf/builtin-stat.c +2 -0
@@ -1853,6 +1853,8 @@ static void setup_system_wide(int forks)struct evsel *counter;evlist__for_each_entry (evsel_list, counter) { +if (counter->tool_event ==PERF_TOOL_DURATION_TIME ) +continue ;if (!counter->core.system_wide)return ; }
Hi Jiri,
Quoted 8 lines · Jiri Olsa
Quoted 7 lines · Andi Kleen
Quoted 4 lines
I wonder this would break some expectations.. would it be more safe to detect duration event and bypass it from the decission? but maybe the case I'm worried about is not a problem at all.. Andi?
Don't see what it would break.
Quoted 9 lines
Quoted 2 lines
Yes maybe we need to special case
duration_timemore, but that would be a much bigger patch.would below change work for you? if
duration_timeis the only case, I'd rather go with the special case for itjirka
Just tested, for the case of DRAM_BW_Use, it can work.
root@lkp-csl-2sp5 ~# perf stat -M DRAM_BW_Use -- sleep 1
Performance counter stats for 'system wide':
9.54 MiB uncore_imc/cas_count_read/
3.26 MiB uncore_imc/cas_count_write/
1002109793 ns duration_time1.002109793 seconds time elapsed
Thanks
Jin Yao
Quoted 15 lines
---
tools/perf/builtin-stat.c +2
-0@@ -1853,6 +1853,8 @@ static void setup_system_wide(int forks)struct evsel *counter;evlist__for_each_entry (evsel_list, counter) { +if (counter->tool_event ==PERF_TOOL_DURATION_TIME ) +continue ;if (!counter->core.system_wide)return ; }