Skip to contentSkip to conversation

perf stat: Skip duration_time in setup_system_wide

PATCH In review Jin Yao 2 participants 4 messages Sep 22, 2020 – Sep 23, 2020 perf events

lore
Share & tools
lore

Jin Yao

author
#0 lore raw

PATCH 1/1 perf stat: Skip duration_time in setup_system_wide

Some metrics (such as DRAM_BW_Use) consists of uncore events and duration_time. For uncore events, counter->core.system_wide is true. But for duration_time, counter->core.system_wide is false so target.system_wide is set to false.

Then 'enable_on_exec' is set in perf_event_attr of uncore event. Kernel will return error when trying to open the uncore event.

This patch skips the duration_time in setup_system_wide then target.system_wide will be set to true for the evlist of uncore events + duration_time.

Before (tested on skylake desktop):

 # perf stat -M DRAM_BW_Use -- sleep 1
 Error:
 The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/).
 /bin/dmesg | grep -i perf may provide additional information.

After:

 # perf stat -M DRAM_BW_Use -- sleep 1

Performance counter stats for 'system wide':

                169      arb/event=0x84,umask=0x1/ #     0.00 DRAM_BW_Use
             40,427      arb/event=0x81,umask=0x1/
      1,000,902,197 ns   duration_time
        1.000902197 seconds time elapsed
Fixes 648b5af3f3ae ("libperf: Move 'system_wide' from 'struct evsel' to 'struct perf_evsel'")
---
 tools/perf/builtin-stat.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +3 -1
index 7f8d756d9408..9bcc93bc0973 100644
@@ -2047,8 +2047,10 @@ static void setup_system_wide(int forks)
 		struct evsel *counter;
 
 		evlist__for_each_entry(evsel_list, counter) {
-			if (!counter->core.system_wide)
+			if (!counter->core.system_wide &&
+			    strcmp(counter->name, "duration_time")) {
 				return;
+			}
 		}
 
 		if (evsel_list->core.nr_entries)
Signature · 1 line
2.17.1
#1 lore raw

Em Tue, Sep 22, 2020 at 09:50:04AM +0800, Jin Yao escreveu:

Quoted 25 lines

Some metrics (such as DRAM_BW_Use) consists of uncore events and duration_time. For uncore events, counter->core.system_wide is true. But for duration_time, counter->core.system_wide is false so target.system_wide is set to false.

Then 'enable_on_exec' is set in perf_event_attr of uncore event. Kernel will return error when trying to open the uncore event.

This patch skips the duration_time in setup_system_wide then target.system_wide will be set to true for the evlist of uncore events + duration_time.

Before (tested on skylake desktop):

 # perf stat -M DRAM_BW_Use -- sleep 1
 Error:
 The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/).
 /bin/dmesg | grep -i perf may provide additional information.

After:

 # perf stat -M DRAM_BW_Use -- sleep 1

Performance counter stats for 'system wide':

Quoted 7 lines
                169      arb/event=0x84,umask=0x1/ #     0.00 DRAM_BW_Use
             40,427      arb/event=0x81,umask=0x1/
      1,000,902,197 ns   duration_time
        1.000902197 seconds time elapsed
Fixes 648b5af3f3ae ("libperf: Move 'system_wide' from 'struct evsel' to 'struct perf_evsel'")

Humm, what makes you think that this cset was the one introducing this problem? It just moves evsel->system_wide to evsel->core.system_wide.

- Arnaldo

Quoted 24 lines
---
 tools/perf/builtin-stat.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +3 -1
index 7f8d756d9408..9bcc93bc0973 100644
@@ -2047,8 +2047,10 @@ static void setup_system_wide(int forks)
 		struct evsel *counter;
 
 		evlist__for_each_entry(evsel_list, counter) {
-			if (!counter->core.system_wide)
+			if (!counter->core.system_wide &&
+			    strcmp(counter->name, "duration_time")) {
 				return;
+			}
 		}
 
 		if (evsel_list->core.nr_entries)
Signature · 1 line
2.17.1
Signature · 1 line
- Arnaldo
#2 lore raw

Em Tue, Sep 22, 2020 at 02:56:30PM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 30 lines

Em Tue, Sep 22, 2020 at 09:50:04AM +0800, Jin Yao escreveu:

Quoted 29 lines

Some metrics (such as DRAM_BW_Use) consists of uncore events and duration_time. For uncore events, counter->core.system_wide is true. But for duration_time, counter->core.system_wide is false so target.system_wide is set to false.

Then 'enable_on_exec' is set in perf_event_attr of uncore event. Kernel will return error when trying to open the uncore event.

This patch skips the duration_time in setup_system_wide then target.system_wide will be set to true for the evlist of uncore events + duration_time.

Before (tested on skylake desktop):

 # perf stat -M DRAM_BW_Use -- sleep 1
 Error:
 The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/).
 /bin/dmesg | grep -i perf may provide additional information.

After:

 # perf stat -M DRAM_BW_Use -- sleep 1

Performance counter stats for 'system wide':

                169      arb/event=0x84,umask=0x1/ #     0.00 DRAM_BW_Use
             40,427      arb/event=0x81,umask=0x1/
      1,000,902,197 ns   duration_time
Quoted 6 lines
Quoted 3 lines
        1.000902197 seconds time elapsed
Fixes 648b5af3f3ae ("libperf: Move 'system_wide' from 'struct evsel' to 'struct perf_evsel'")

Humm, what makes you think that this cset was the one introducing this problem? It just moves evsel->system_wide to evsel->core.system_wide.

Apart from that I reproduced the problem and after applying your patch it seems cured:

[acme@quaco perf]$ grep 'model name' -m1 /proc/cpuinfo
model name : Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz

Before (with -v to see details):

[root@quaco ~]# perf stat -v -M DRAM_BW_Use -- sleep 1
Using CPUID GenuineIntel-6-8E-A
metric expr 64 * ( arb@event\=0x81\,umask\=0x1@ + arb@event\=0x84\,umask\=0x1@ ) / 1000000 / duration_time / 1000 for DRAM_BW_Use found event duration_time
found event arb/event=0x84,umask=0x1/
found event arb/event=0x81,umask=0x1/
adding {arb/event=0x84,umask=0x1/,arb/event=0x81,umask=0x1/}:W,duration_time
Control descriptor is not initialized
Warning:
arb/event=0x84,umask=0x1/ event is not supported by the kernel.
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/). /bin/dmesg | grep -i perf may provide additional information.

[root@quaco ~]#

After:

[root@quaco ~]# perf stat -M DRAM_BW_Use -- sleep 1

Performance counter stats for 'system wide':

               2,806      arb/event=0x84,umask=0x1/ #     0.63 DRAM_BW_Use
          10,001,820      arb/event=0x81,umask=0x1/
       1,016,875,686 ns   duration_time
         1.016875686 seconds time elapsed

[root@quaco ~]#

So I'm removing that fixes and adding this one, that I think is where "duration_time" was being considered...

Fixes e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring")

Also, wouldn't it be better to have the duration_time event with its evsel->core.system_wide set to true?

- Arnaldo

author
#3 lore raw

Hi Arnaldo,

Quoted 76 lines · Arnaldo Carvalho de Melo

Em Tue, Sep 22, 2020 at 02:56:30PM -0300, Arnaldo Carvalho de Melo escreveu:

Quoted 30 lines

Em Tue, Sep 22, 2020 at 09:50:04AM +0800, Jin Yao escreveu:

Quoted 29 lines

Some metrics (such as DRAM_BW_Use) consists of uncore events and duration_time. For uncore events, counter->core.system_wide is true. But for duration_time, counter->core.system_wide is false so target.system_wide is set to false.

Then 'enable_on_exec' is set in perf_event_attr of uncore event. Kernel will return error when trying to open the uncore event.

This patch skips the duration_time in setup_system_wide then target.system_wide will be set to true for the evlist of uncore events + duration_time.

Before (tested on skylake desktop):

  # perf stat -M DRAM_BW_Use -- sleep 1
  Error:
  The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/).
  /bin/dmesg | grep -i perf may provide additional information.

After:

  # perf stat -M DRAM_BW_Use -- sleep 1
   Performance counter stats for 'system wide':

                 169      arb/event=0x84,umask=0x1/ #     0.00 DRAM_BW_Use
              40,427      arb/event=0x81,umask=0x1/
       1,000,902,197 ns   duration_time
Quoted 6 lines
Quoted 3 lines
         1.000902197 seconds time elapsed
Fixes 648b5af3f3ae ("libperf: Move 'system_wide' from 'struct evsel' to 'struct perf_evsel'")

Humm, what makes you think that this cset was the one introducing this problem? It just moves evsel->system_wide to evsel->core.system_wide.

Apart from that I reproduced the problem and after applying your patch it seems cured:

[acme@quaco perf]$ grep 'model name' -m1 /proc/cpuinfo
model name : Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz

Before (with -v to see details):

[root@quaco ~]# perf stat -v -M DRAM_BW_Use -- sleep 1
Using CPUID GenuineIntel-6-8E-A
metric expr 64 * ( arb@event\=0x81\,umask\=0x1@ + arb@event\=0x84\,umask\=0x1@ ) / 1000000 / duration_time / 1000 for DRAM_BW_Use found event duration_time
found event arb/event=0x84,umask=0x1/
found event arb/event=0x81,umask=0x1/
adding {arb/event=0x84,umask=0x1/,arb/event=0x81,umask=0x1/}:W,duration_time
Control descriptor is not initialized
Warning:
arb/event=0x84,umask=0x1/ event is not supported by the kernel.
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (arb/event=0x84,umask=0x1/). /bin/dmesg | grep -i perf may provide additional information.

   [root@quaco ~]#

After:

   [root@quaco ~]# perf stat -M DRAM_BW_Use -- sleep 1
   
    Performance counter stats for 'system wide':
   
                2,806      arb/event=0x84,umask=0x1/ #     0.63 DRAM_BW_Use
           10,001,820      arb/event=0x81,umask=0x1/
        1,016,875,686 ns   duration_time
   
          1.016875686 seconds time elapsed
   
   [root@quaco ~]#
Quoted 5 lines

So I'm removing that fixes and adding this one, that I think is where "duration_time" was being considered...

Fixes e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring")

Yes, this fixes is much better, thanks.

Quoted 3 lines

Also, wouldn't it be better to have the duration_time event with its evsel->core.system_wide set to true?

That looks to be another solution, should be OK too I think. :)

But anyway we need a test.

Thanks
Jin Yao

Quoted 2 lines

- Arnaldo