Skip to contentSkip to conversation

perf stat cleanups

PATCH In review David Ahern 5 participants 17 messages Sep 28, 2013 – Oct 15, 2013 perf events

lore
Share & tools
lore
#0 lore raw

cover perf stat cleanups

Arnaldo:

A few perf-stat cleanups.

David Ahern (3):
perf stat: Fix misleading message when specifying cpu list or system wide perf stat: Don't require a workload when using system wide or CPU options perf stat: Add units to nanosec-based counters

 tools/perf/builtin-stat.c |   15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)
Signature · 1 line
1.7.10.1
#1 lore raw

PATCH 1/3 perf stat: Fix misleading message when specifying cpu list or system wide

perf-stat displays the command run in its summary output which is misleading when using a cpu list or system wide collection.

Before:

perf stat -a -- sleep 1

Performance counter stats for 'sleep 1':

16152.670249 task-clock                #   16.132 CPUs utilized
         417 context-switches          #    0.002 M/sec
           7 cpu-migrations            #    0.030 K/sec

...

After:

perf stat -a -- sleep 1

Performance counter stats for 'system wide':

16206.931120 task-clock                #   16.144 CPUs utilized
         395 context-switches          #    0.002 M/sec
           5 cpu-migrations            #    0.030 K/sec

...

or

perf stat -C1 -- sleep 1

Performance counter stats for 'CPU(s) 1':

   1001.669257 task-clock                #    1.000 CPUs utilized
         4,264 context-switches          #    0.004 M/sec
             3 cpu-migrations            #    0.003 K/sec

...

Cc Ingo Molnar <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 tools/perf/builtin-stat.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +5 -1
index f686d5f..6cc0aa2 100644
@@ -1115,7 +1115,11 @@ static void print_stat(int argc, const char **argv)
 	if (!csv_output) {
 		fprintf(output, "\n");
 		fprintf(output, " Performance counter stats for ");
-		if (!perf_target__has_task(&target)) {
+		if (target.system_wide)
+			fprintf(output, "\'system wide");
+		else if (target.cpu_list)
+			fprintf(output, "\'CPU(s) %s", target.cpu_list);
+		else if (!perf_target__has_task(&target)) {
 			fprintf(output, "\'%s", argv[0]);
 			for (i = 1; i < argc; i++)
 				fprintf(output, " %s", argv[i]);
Signature · 1 line
1.7.10.1
#2 lore raw

[tip:perf/core] perf stat: Fix misleading message when specifying cpu list or system wide

perf stat: Fix misleading message when specifying cpu list or system wide
Commit-ID:  62d3b617c02785a4f1fbde8d93ca77a0b33d8454
Gitweb:     http://git.kernel.org/tip/62d3b617c02785a4f1fbde8d93ca77a0b33d8454
Author:     David Ahern <[email protected]>
AuthorDate: Sat, 28 Sep 2013 14:27:58 -0600
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Fri, 11 Oct 2013 12:17:42 -0300

perf stat: Fix misleading message when specifying cpu list or system wide

The "perf stat" tool displays the command run in its summary output which is misleading when using a cpu list or system wide collection.

Before:

perf stat -a -- sleep 1

Performance counter stats for 'sleep 1':

16152.670249 task-clock                #   16.132 CPUs utilized
         417 context-switches          #    0.002 M/sec
           7 cpu-migrations            #    0.030 K/sec

...

After:

perf stat -a -- sleep 1

Performance counter stats for 'system wide':

16206.931120 task-clock                #   16.144 CPUs utilized
         395 context-switches          #    0.002 M/sec
           5 cpu-migrations            #    0.030 K/sec

...

or

perf stat -C1 -- sleep 1

Performance counter stats for 'CPU(s) 1':

   1001.669257 task-clock                #    1.000 CPUs utilized
         4,264 context-switches          #    0.004 M/sec
             3 cpu-migrations            #    0.003 K/sec

...

---
 tools/perf/builtin-stat.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +5 -1
index fb02b53..c8a2662 100644
@@ -1229,7 +1229,11 @@ static void print_stat(int argc, const char **argv)
 	if (!csv_output) {
 		fprintf(output, "\n");
 		fprintf(output, " Performance counter stats for ");
-		if (!perf_target__has_task(&target)) {
+		if (target.system_wide)
+			fprintf(output, "\'system wide");
+		else if (target.cpu_list)
+			fprintf(output, "\'CPU(s) %s", target.cpu_list);
+		else if (!perf_target__has_task(&target)) {
 			fprintf(output, "\'%s", argv[0]);
 			for (i = 1; i < argc; i++)
 				fprintf(output, " %s", argv[i]);
#3 lore raw

PATCH 2/3 perf stat: Don't require a workload when using system wide or CPU options

In reply to David Ahern

perf-stat can do system wide counters or one or more cpus. For these options do not require a workload to be specified.

Cc Ingo Molnar <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 tools/perf/builtin-stat.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +3 -1
index 6cc0aa2..60239fe 100644
@@ -1517,8 +1517,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && !perf_target__has_task(&target) &&
+	    !perf_target__has_cpu(&target))
 		usage_with_options(stat_usage, options);
+
 	if (run_count < 0) {
 		usage_with_options(stat_usage, options);
 	} else if (run_count == 0) {
Signature · 1 line
1.7.10.1
maintainer
#4 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Hi David,

Quoted 19 lines · David Ahern

perf-stat can do system wide counters or one or more cpus. For these options do not require a workload to be specified.

Cc Ingo Molnar <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 tools/perf/builtin-stat.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +0 -0
index 6cc0aa2..60239fe 100644

tools/perf/builtin-stat.c int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) · line 1521

Quoted 6 lines
tools/perf/builtin-stat.c +2 -1
@@ -1517,8 +1517,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && !perf_target__has_task(&target) &&
+	    !perf_target__has_cpu(&target))

You can use perf_target__none() for this.

Btw I found a bug in setting child_pid (which introduced by me ;) during this review. Will send a fix.

Thanks,
Namhyung

Quoted 5 lines
tools/perf/builtin-stat.c +1 -0
 		usage_with_options(stat_usage, options);
+
 	if (run_count < 0) {
 		usage_with_options(stat_usage, options);
 	} else if (run_count == 0) {
#5 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Quoted 9 lines · Namhyung Kim

Review int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) · line 1521

Quoted 7 lines
(quoted diff) +2 -1
@@ -1517,8 +1517,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
  	} else if (big_num_opt == 0) /* User passed --no-big-num */
  		big_num = false;
 
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && !perf_target__has_task(&target) &&
+	    !perf_target__has_cpu(&target))

You can use perf_target__none() for this.

Indeed. Updated patch attached.

David

Quoted 1 line

From a7018902283b3cfc5cae8a1876e2c14f3c7c9c7e Mon Sep 17 00:00:00 2001

From: David Ahern <[email protected]>
Date: Mon, 30 Sep 2013 07:37:37 -0600
Subject: [PATCH] perf stat: Don't require a workload when using system wide or CPU options -v2

perf-stat can do system wide counters or one or more cpus. For these options do not require a workload to be specified.

v2: use perf_target__none per Namhyung's comment

Cc Ingo Molnar <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 tools/perf/builtin-stat.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +2 -1
index 6cc0aa2..db55b91 100644
@@ -1517,8 +1517,9 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && perf_target__none(&target))
 		usage_with_options(stat_usage, options);
+
 	if (run_count < 0) {
 		usage_with_options(stat_usage, options);
 	} else if (run_count == 0) {
Signature · 1 line
1.7.10.1
#6 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Em Mon, Sep 30, 2013 at 07:40:12AM -0600, David Ahern escreveu:

Quoted 5 lines · Namhyung Kim
Quoted 4 lines
Quoted 4 lines
tools/perf/builtin-stat.c +0 -0
@@ -1517,8 +1517,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
Quoted 7 lines
Quoted 5 lines

Review int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) · line 1521

Quoted 3 lines
(quoted diff) +2 -1
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && !perf_target__has_task(&target) &&
+	    !perf_target__has_cpu(&target))

You can use perf_target__none() for this.

Indeed. Updated patch attached.

Cool patch, applying to acme/perf/core.

While trying it noticed this, that should also be fixed eventually:

[root@zoo ~]# perf stat -c C 0
C: No such file or directory

Performance counter stats for 'C 0':

<not counted> task-clock
<not counted> context-switches
<not counted> cpu-migrations
<SNIP>

- Arnaldo

#7 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Quoted 22 lines · Arnaldo Carvalho de Melo

Em Mon, Sep 30, 2013 at 07:40:12AM -0600, David Ahern escreveu:

Quoted 5 lines · Namhyung Kim
Quoted 4 lines
Quoted 4 lines
tools/perf/builtin-stat.c +0 -0
@@ -1517,8 +1517,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
Quoted 7 lines
Quoted 5 lines

Review int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused) · line 1521

Quoted 3 lines
(quoted diff) +2 -1
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && !perf_target__has_task(&target) &&
+	    !perf_target__has_cpu(&target))

You can use perf_target__none() for this.

Indeed. Updated patch attached.

Cool patch, applying to acme/perf/core.

While trying it noticed this, that should also be fixed eventually:

[root@zoo ~]# perf stat -c C 0
C: No such file or directory

Quoted 6 lines

Performance counter stats for 'C 0':

<not counted> task-clock
<not counted> context-switches
<not counted> cpu-migrations
<SNIP>

Btw., would anyone be interested in adding CPU and node binding options to perf stat?

There's code to do something like that in tools/perf/bench/numa.c:

        /* Special option string parsing callbacks: */
        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
                        "bind the first N tasks to these specific cpus (the rest is unbound)",
                        parse_cpus_opt),
        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
                        "bind the first N tasks to these specific memory nodes (the rest is unbound)",
                        parse_nodes_opt),

Thanks,

Ingo

#8 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Em Tue, Oct 08, 2013 at 02:51:13PM +0200, Ingo Molnar escreveu:

Quoted 5 lines

Btw., would anyone be interested in adding CPU and node binding options to perf stat?

There's code to do something like that in tools/perf/bench/numa.c:

Quoted 7 lines
        /* Special option string parsing callbacks: */
        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
                        "bind the first N tasks to these specific cpus (the rest is unbound)",
                        parse_cpus_opt),
        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
                        "bind the first N tasks to these specific memory nodes (the rest is unbound)",
                        parse_nodes_opt),

As a convenience to using:

taskset -c cpu-list perf stat -C cpu-list workload

?

- Arnaldo

#9 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Quoted 14 lines · Arnaldo Carvalho de Melo

Em Tue, Oct 08, 2013 at 02:51:13PM +0200, Ingo Molnar escreveu:

Quoted 5 lines

Btw., would anyone be interested in adding CPU and node binding options to perf stat?

There's code to do something like that in tools/perf/bench/numa.c:

Quoted 7 lines
        /* Special option string parsing callbacks: */
        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
                        "bind the first N tasks to these specific cpus (the rest is unbound)",
                        parse_cpus_opt),
        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
                        "bind the first N tasks to these specific memory nodes (the rest is unbound)",
                        parse_nodes_opt),
Quoted 5 lines

As a convenience to using:

taskset -c cpu-list perf stat -C cpu-list workload

?

Yeah, something like that would be handy I think.

Ingo

#10 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

In reply to Arnaldo Carvalho de Melo
Quoted 5 lines · Arnaldo Carvalho de Melo

While trying it noticed this, that should also be fixed eventually:

[root@zoo ~]# perf stat -c C 0
C: No such file or directory

you are missing '-' before the C: perf stat -C 0. The -C = cpus. -c for stat is scale.

David

#11 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Em Tue, Oct 08, 2013 at 07:25:39AM -0600, David Ahern escreveu:

Quoted 9 lines · Arnaldo Carvalho de Melo
Quoted 5 lines

While trying it noticed this, that should also be fixed eventually:

[root@zoo ~]# perf stat -c C 0
C: No such file or directory

you are missing '-' before the C: perf stat -C 0. The -C = cpus. -c for stat is scale.

Nah, I mean:

[acme@zoo linux]$ perf stat non-existent-proggie
non-existent-proggie: No such file or directory

Performance counter stats for 'non-existent-proggie':

<not counted> task-clock
<not counted> context-switches
<not counted> cpu-migrations
<not counted> page-faults
<not counted> cycles
<not counted> stalled-cycles-frontend
<not supported> stalled-cycles-backend
<not counted> instructions
<not counted> branches
<not counted> branch-misses

0.001823577 seconds time elapsed

[acme@zoo linux]$

Why should we report all those counters for something that wasn't found in PATH?

We either report just the first line, with the errno, or that and the last, with how long it took to try to execute it, but that probably shouldn't be printed, as it is supposed to be how long the workload ran, right?

- Arnaldo

#12 lore raw

Re: [PATCH 2/3] perf stat: Don't require a workload when using system wide or CPU options

Quoted 2 lines · Arnaldo Carvalho de Melo

[acme@zoo linux]$ perf stat non-existent-proggie
non-existent-proggie: No such file or directory

ok. so this is a different error. need to detect the workload does not exist.

Quoted 7 lines

Why should we report all those counters for something that wasn't found in PATH?

We either report just the first line, with the errno, or that and the last, with how long it took to try to execute it, but that probably shouldn't be printed, as it is supposed to be how long the workload ran, right?

Right. need to feed/detect the exec failure back to the parent.

David

#13 lore raw

[tip:perf/core] perf stat: Don' t require a workload when using system wide or CPU options

In reply to David Ahern
perf stat: Don't require a workload when using system wide or CPU options
Commit-ID:  ac3063bd4725689f39d7a23fdfca2e034c73dcac
Gitweb:     http://git.kernel.org/tip/ac3063bd4725689f39d7a23fdfca2e034c73dcac
Author:     David Ahern <[email protected]>
AuthorDate: Mon, 30 Sep 2013 07:37:37 -0600
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Fri, 11 Oct 2013 12:17:44 -0300

perf stat: Don't require a workload when using system wide or CPU options

The "perf stat" command can do system wide counters or one or more cpus. For these options do not require a workload to be specified.

v2: use perf_target__none per Namhyung's comment.

---
 tools/perf/builtin-stat.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +2 -1
index c8a2662..2178e66 100644
@@ -1659,8 +1659,9 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	} else if (big_num_opt == 0) /* User passed --no-big-num */
 		big_num = false;
 
-	if (!argc && !perf_target__has_task(&target))
+	if (!argc && perf_target__none(&target))
 		usage_with_options(stat_usage, options);
+
 	if (run_count < 0) {
 		usage_with_options(stat_usage, options);
 	} else if (run_count == 0) {
#14 lore raw

PATCH 3/3 perf stat: Add units to nanosec-based counters

In reply to David Ahern

Ingo pointed out that the task-clock counter should have the units explicitly stated since it is not a counter.

Before:

perf stat -a -- sleep 1

Performance counter stats for 'sleep 1':

      16186.874834 task-clock          #   16.154 CPUs utilized

...

After:

perf stat -a -- sleep 1

Performance counter stats for 'system wide':

      16146.402138 task-clock (msec)   #   16.125 CPUs utilized

...

Cc Ingo Molnar <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 tools/perf/builtin-stat.c |    5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +4 -1
index 60239fe..746a6db 100644
@@ -628,10 +628,13 @@ static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double msecs = avg / 1e6;
 	const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
+	char name[25];
 
 	aggr_printout(evsel, cpu, nr);
 
-	fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
+	scnprintf(name, sizeof(name), "%s%s",
+		  perf_evsel__name(evsel), csv_output ? "" : " (msec)");
+	fprintf(output, fmt, msecs, csv_sep, name);
 
 	if (evsel->cgrp)
 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
Signature · 1 line
1.7.10.1
#15 lore raw

[tip:perf/core] perf stat: Add units to nanosec-based counters

perf stat: Add units to nanosec-based counters
Commit-ID:  4bbe5a61f29b13437a6a16467328d3bae8fce9e7
Gitweb:     http://git.kernel.org/tip/4bbe5a61f29b13437a6a16467328d3bae8fce9e7
Author:     David Ahern <[email protected]>
AuthorDate: Sat, 28 Sep 2013 14:28:00 -0600
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Fri, 11 Oct 2013 12:17:46 -0300

perf stat: Add units to nanosec-based counters

Ingo pointed out that the task-clock counter should have the units explicitly stated since it is not a counter.

Before:

perf stat -a -- sleep 1

Performance counter stats for 'sleep 1':

      16186.874834 task-clock          #   16.154 CPUs utilized

...

After:

perf stat -a -- sleep 1

Performance counter stats for 'system wide':

      16146.402138 task-clock (msec)   #   16.125 CPUs utilized

...

---
 tools/perf/builtin-stat.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +4 -1
index 2178e66..1a9c95d 100644
@@ -706,10 +706,13 @@ static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double msecs = avg / 1e6;
 	const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
+	char name[25];
 
 	aggr_printout(evsel, cpu, nr);
 
-	fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
+	scnprintf(name, sizeof(name), "%s%s",
+		  perf_evsel__name(evsel), csv_output ? "" : " (msec)");
+	fprintf(output, fmt, msecs, csv_sep, name);
 
 	if (evsel->cgrp)
 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
#16 lore raw
In reply to David Ahern

ping

Quoted 12 lines · David Ahern

Arnaldo:

A few perf-stat cleanups.

David Ahern (3):
perf stat: Fix misleading message when specifying cpu list or system wide perf stat: Don't require a workload when using system wide or CPU options perf stat: Add units to nanosec-based counters

  tools/perf/builtin-stat.c |   15 ++++++++++++---
  1 file changed, 12 insertions(+), 3 deletions(-)