Skip to contentSkip to conversation

perf stat: Do not delay the workload with --delay

PATCH In review Namhyung Kim 6 participants 6 messages Dec 12, 2022 – Dec 15, 2022 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH 1/1 perf stat: Do not delay the workload with --delay

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled

Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed

real 0m7.474s
user 0m0.356s
sys 0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled

Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed

real 0m4.484s
user 0m0.385s
sys 0m0.086s

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>
---
 tools/perf/builtin-stat.c | 33 +++++++++++++++++----------------
 1 file changed, 17 insertions(+), 16 deletions(-)
tools/perf/builtin-stat.c +17 -16
index d040fbcdcc5a..b39bf785a16e 100644
@@ -540,26 +540,14 @@ static int enable_counters(void)
 			return err;
 	}
 
-	if (stat_config.initial_delay < 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		return 0;
-	}
-
-	if (stat_config.initial_delay > 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		usleep(stat_config.initial_delay * USEC_PER_MSEC);
-	}
-
 	/*
 	 * We need to enable counters only if:
 	 * - we don't have tracee (attaching to task or cpu)
 	 * - we have initial delay configured
 	 */
-	if (!target__none(&target) || stat_config.initial_delay) {
+	if (!target__none(&target)) {
 		if (!all_counters_use_bpf)
 			evlist__enable(evsel_list);
-		if (stat_config.initial_delay > 0)
-			pr_info(EVLIST_ENABLED_MSG);
 	}
 	return 0;
 }
@@ -930,14 +918,27 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			return err;
 	}
 
-	err = enable_counters();
-	if (err)
-		return -1;
+	if (stat_config.initial_delay) {
+		pr_info(EVLIST_DISABLED_MSG);
+	} else {
+		err = enable_counters();
+		if (err)
+			return -1;
+	}
 
 	/* Exec the command, if any */
 	if (forks)
 		evlist__start_workload(evsel_list);
 
+	if (stat_config.initial_delay > 0) {
+		usleep(stat_config.initial_delay * USEC_PER_MSEC);
+		err = enable_counters();
+		if (err)
+			return -1;
+
+		pr_info(EVLIST_ENABLED_MSG);
+	}
+
 	t0 = rdclock();
 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Acked-by on v1 1/1
#1 lore raw
Quoted 108 lines · Namhyung Kim

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed

real 0m7.474s
user 0m0.356s
sys 0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed

real 0m4.484s
user 0m0.385s
sys 0m0.086s

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>
---
 tools/perf/builtin-stat.c | 33 +++++++++++++++++----------------
 1 file changed, 17 insertions(+), 16 deletions(-)
tools/perf/builtin-stat.c +10 -16
index d040fbcdcc5a..b39bf785a16e 100644
@@ -540,26 +540,14 @@ static int enable_counters(void)
 			return err;
 	}
 
-	if (stat_config.initial_delay < 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		return 0;
-	}
-
-	if (stat_config.initial_delay > 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		usleep(stat_config.initial_delay * USEC_PER_MSEC);
-	}
-
 	/*
 	 * We need to enable counters only if:
 	 * - we don't have tracee (attaching to task or cpu)
 	 * - we have initial delay configured
 	 */
-	if (!target__none(&target) || stat_config.initial_delay) {
+	if (!target__none(&target)) {
 		if (!all_counters_use_bpf)
 			evlist__enable(evsel_list);
-		if (stat_config.initial_delay > 0)
-			pr_info(EVLIST_ENABLED_MSG);
 	}
 	return 0;
 }
@@ -930,14 +918,27 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			return err;
 	}
 
-	err = enable_counters();
-	if (err)
-		return -1;
+	if (stat_config.initial_delay) {
+		pr_info(EVLIST_DISABLED_MSG);
+	} else {
+		err = enable_counters();
+		if (err)
+			return -1;
+	}
 
 	/* Exec the command, if any */
 	if (forks)
 		evlist__start_workload(evsel_list);
 
+	if (stat_config.initial_delay > 0) {
+		usleep(stat_config.initial_delay * USEC_PER_MSEC);

tools/perf/builtin-stat.c static int __run_perf_stat(int argc, const char **argv, int run_idx) · line 944

Quoted 10 lines
tools/perf/builtin-stat.c +7 -0
+		err = enable_counters();
+		if (err)
+			return -1;
+
+		pr_info(EVLIST_ENABLED_MSG);
+	}
+
 	t0 = rdclock();
 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 

Tested successfully on s390

Signature · 5 lines
Thomas Richter, Dept 3303, IBM s390 Linux Development, Boeblingen, Germany
--
Vorsitzender des Aufsichtsrats: Gregor Pillen
Geschäftsführung: David Faller
Sitz der Gesellschaft: Böblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294
#2 lore raw
In reply to Namhyung Kim
Quoted 41 lines · Namhyung Kim

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed

real 0m7.474s
user 0m0.356s
sys 0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed

real 0m4.484s
user 0m0.385s
sys 0m0.086s

Quoted 6 lines

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")

Looks like the same fix as here, but it was only partially applied:

https://lore.kernel.org/linux-perf-users/20220801112726.GA73228@leoy-ThinkPad-X240s/T/#mce4ad6b17a5fa15970765bb01e7b691c7326fbb0

Quoted 71 lines
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>
---
 tools/perf/builtin-stat.c | 33 +++++++++++++++++----------------
 1 file changed, 17 insertions(+), 16 deletions(-)
tools/perf/builtin-stat.c +17 -16
index d040fbcdcc5a..b39bf785a16e 100644
@@ -540,26 +540,14 @@ static int enable_counters(void)
 			return err;
 	}
 
-	if (stat_config.initial_delay < 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		return 0;
-	}
-
-	if (stat_config.initial_delay > 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		usleep(stat_config.initial_delay * USEC_PER_MSEC);
-	}
-
 	/*
 	 * We need to enable counters only if:
 	 * - we don't have tracee (attaching to task or cpu)
 	 * - we have initial delay configured
 	 */
-	if (!target__none(&target) || stat_config.initial_delay) {
+	if (!target__none(&target)) {
 		if (!all_counters_use_bpf)
 			evlist__enable(evsel_list);
-		if (stat_config.initial_delay > 0)
-			pr_info(EVLIST_ENABLED_MSG);
 	}
 	return 0;
 }
@@ -930,14 +918,27 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			return err;
 	}
 
-	err = enable_counters();
-	if (err)
-		return -1;
+	if (stat_config.initial_delay) {
+		pr_info(EVLIST_DISABLED_MSG);
+	} else {
+		err = enable_counters();
+		if (err)
+			return -1;
+	}
 
 	/* Exec the command, if any */
 	if (forks)
 		evlist__start_workload(evsel_list);
 
+	if (stat_config.initial_delay > 0) {
+		usleep(stat_config.initial_delay * USEC_PER_MSEC);
+		err = enable_counters();
+		if (err)
+			return -1;
+
+		pr_info(EVLIST_ENABLED_MSG);
+	}
+
 	t0 = rdclock();
 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 
reviewer
#3 lore raw
In reply to Namhyung Kim
Quoted 38 lines · Namhyung Kim

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed
  real  0m7.474s
  user  0m0.356s
  sys   0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed
Quoted 3 lines
  real  0m4.484s
  user  0m0.385s
  sys   0m0.086s

The commit message feels like it could almost be turned into a shell test. The test would need some fudge factors in case of load on the test system. Any thoughts if we could add this? We wouldn't need to rely on 'time' as we have tool events of user_time, system_time, etc.

Thanks,
Ian

Quoted 80 lines

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>
---
 tools/perf/builtin-stat.c | 33 +++++++++++++++++----------------
 1 file changed, 17 insertions(+), 16 deletions(-)
tools/perf/builtin-stat.c +17 -17
index d040fbcdcc5a..b39bf785a16e 100644
@@ -540,26 +540,14 @@ static int enable_counters(void)
                        return err;
        }
 
-       if (stat_config.initial_delay < 0) {
-               pr_info(EVLIST_DISABLED_MSG);
-               return 0;
-       }
-
-       if (stat_config.initial_delay > 0) {
-               pr_info(EVLIST_DISABLED_MSG);
-               usleep(stat_config.initial_delay * USEC_PER_MSEC);
-       }
-
        /*
         * We need to enable counters only if:
         * - we don't have tracee (attaching to task or cpu)
         * - we have initial delay configured
         */
-       if (!target__none(&target) || stat_config.initial_delay) {
+       if (!target__none(&target)) {
                if (!all_counters_use_bpf)
                        evlist__enable(evsel_list);
-               if (stat_config.initial_delay > 0)
-                       pr_info(EVLIST_ENABLED_MSG);
        }
        return 0;
 }
@@ -930,14 +918,27 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
                        return err;
        }
 
-       err = enable_counters();
-       if (err)
-               return -1;
+       if (stat_config.initial_delay) {
+               pr_info(EVLIST_DISABLED_MSG);
+       } else {
+               err = enable_counters();
+               if (err)
+                       return -1;
+       }
 
        /* Exec the command, if any */
        if (forks)
                evlist__start_workload(evsel_list);
 
+       if (stat_config.initial_delay > 0) {
+               usleep(stat_config.initial_delay * USEC_PER_MSEC);
+               err = enable_counters();
+               if (err)
+                       return -1;
+
+               pr_info(EVLIST_ENABLED_MSG);
+       }
+
        t0 = rdclock();
        clock_gettime(CLOCK_MONOTONIC, &ref_time);
 
--

2.39.0.rc1.256.g54fd8350bd-goog

#4 lore raw
In reply to Namhyung Kim

Em Mon, Dec 12, 2022 at 03:08:20PM -0800, Namhyung Kim escreveu:

Quoted 37 lines

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed

real 0m7.474s
user 0m0.356s
sys 0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed
Quoted 8 lines

real 0m4.484s
user 0m0.385s
sys 0m0.086s

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Ok, I took up this one instead of Adrian's.

- Arnaldo

Quoted 74 lines
Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>
---
 tools/perf/builtin-stat.c | 33 +++++++++++++++++----------------
 1 file changed, 17 insertions(+), 16 deletions(-)
tools/perf/builtin-stat.c +17 -16
index d040fbcdcc5a..b39bf785a16e 100644
@@ -540,26 +540,14 @@ static int enable_counters(void)
 			return err;
 	}
 
-	if (stat_config.initial_delay < 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		return 0;
-	}
-
-	if (stat_config.initial_delay > 0) {
-		pr_info(EVLIST_DISABLED_MSG);
-		usleep(stat_config.initial_delay * USEC_PER_MSEC);
-	}
-
 	/*
 	 * We need to enable counters only if:
 	 * - we don't have tracee (attaching to task or cpu)
 	 * - we have initial delay configured
 	 */
-	if (!target__none(&target) || stat_config.initial_delay) {
+	if (!target__none(&target)) {
 		if (!all_counters_use_bpf)
 			evlist__enable(evsel_list);
-		if (stat_config.initial_delay > 0)
-			pr_info(EVLIST_ENABLED_MSG);
 	}
 	return 0;
 }
@@ -930,14 +918,27 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			return err;
 	}
 
-	err = enable_counters();
-	if (err)
-		return -1;
+	if (stat_config.initial_delay) {
+		pr_info(EVLIST_DISABLED_MSG);
+	} else {
+		err = enable_counters();
+		if (err)
+			return -1;
+	}
 
 	/* Exec the command, if any */
 	if (forks)
 		evlist__start_workload(evsel_list);
 
+	if (stat_config.initial_delay > 0) {
+		usleep(stat_config.initial_delay * USEC_PER_MSEC);
+		err = enable_counters();
+		if (err)
+			return -1;
+
+		pr_info(EVLIST_ENABLED_MSG);
+	}
+
 	t0 = rdclock();
 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 
Signature · 1 line
2.39.0.rc1.256.g54fd8350bd-goog
Signature · 1 line
- Arnaldo
Reviewed-by on v1 1/1
#5 lore raw
In reply to Namhyung Kim
Quoted 46 lines · Namhyung Kim

The -D/--delay option is to delay the measure after the program starts. But the current code goes to sleep before starting the program so the program is delayed too. This is not the intention, let's fix it.

Before:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       4,326,949,337      cycles

         4.007494118 seconds time elapsed

real 0m7.474s
user 0m0.356s
sys 0m0.120s

It ran the workload for 4 seconds and gave the 3 second delay. So it should skip the first 3 second and measure the last 1 second only. But as you can see, it delays 3 seconds and ran the workload after that for 4 seconds. So the total time (real) was 7 seconds.

After:

  $ time sudo ./perf stat -a -e cycles -D 3000 sleep 4
  Events disabled
  Events enabled
   Performance counter stats for 'system wide':

       1,063,551,013      cycles

         1.002769510 seconds time elapsed

real 0m4.484s
user 0m0.385s
sys 0m0.086s

The bug was introduced when it changed enablement of system-wide events with a command line workload. But it should've considered the initial delay case. The code was reworked since then (in bb8bc52e7578) so I'm afraid it won't be applied cleanly.

Quoted 5 lines
Fixes d0a0a511493d ("perf stat: Fix forked applications enablement of counters")
Cc Sumanth Korikkar <[email protected]>
Cc Thomas Richter <[email protected]>