Skip to contentSkip to conversation

perf tools: ignore zombie process for user profile

PATCH In review [email protected] 3 participants 5 messages Dec 14, 2016 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf tools: ignore zombie process for user profile

From: Kan Liang <[email protected]>

If user has zombie process, the perf record -u will error out. Here is an example.

 $ ./testd &
 [1] 23796
 $ sudo perf record -e cycles -u kan
 Error:
 The sys_perf_event_open() syscall returned with 3 (No such process) for
 event (cycles).
 /bin/dmesg may provide additional information.
 No CONFIG_PERF_EVENTS=y kernel support configured?

The source code of testd is as below.
int main() {

	if (fork())
	{
		while (1);
	}
	return 0;

}

Zombie process is dead process. It is meaningless to profile it. It's better to ignore it for user profile.

---
 tools/perf/util/thread_map.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)
tools/perf/util/thread_map.c +36 -0
index 40585f5..4643b69 100644
@@ -89,6 +89,39 @@ struct thread_map *thread_map__new_by_tid(pid_t tid)
 	return threads;
 }
 
+static bool is_zombie_process(pid_t pid)
+{
+	char filename[PATH_MAX];
+	char comm[PATH_MAX];
+	char bf[BUFSIZ];
+	char s_state;
+	int s_pid;
+	FILE *fp;
+	ssize_t n;
+
+	snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
+	fp = fopen(filename, "r");
+	if (!fp) {
+		pr_warning("couldn't open %s\n", filename);
+		return false;
+	}
+	if (fgets(bf, sizeof(bf), fp) == NULL) {
+		pr_warning("Couldn't read stat for pid %d\n", pid);
+		fclose(fp);
+		return false;
+	}
+	fclose(fp);
+
+	n = sscanf(bf, "%d %s %c ", &s_pid, comm, &s_state);
+	if (n < 3)
+		return false;
+
+	if (s_state == 'Z')
+		return true;
+
+	return false;
+}
+
 struct thread_map *thread_map__new_by_uid(uid_t uid)
 {
 	DIR *proc;
@@ -124,6 +157,9 @@ struct thread_map *thread_map__new_by_uid(uid_t uid)
 		if (st.st_uid != uid)
 			continue;
 
+		if (is_zombie_process(pid))
+			continue;
+
 		snprintf(path, sizeof(path), "/proc/%d/task", pid);
 		items = scandir(path, &namelist, filter, NULL);
 		if (items <= 0)
Signature · 1 line
2.4.3
#1 lore raw
Quoted 21 lines · On Wed, Dec 14, 2016 at 12:48:05PM -0500, [email protected] wrote:

From: Kan Liang <[email protected]>

If user has zombie process, the perf record -u will error out. Here is an example.

 $ ./testd &
 [1] 23796
 $ sudo perf record -e cycles -u kan
 Error:
 The sys_perf_event_open() syscall returned with 3 (No such process) for
 event (cycles).
 /bin/dmesg may provide additional information.
 No CONFIG_PERF_EVENTS=y kernel support configured?

The source code of testd is as below.
int main() {

	if (fork())
	{
		while (1);
	}
	return 0;
Quoted 4 lines

}

Zombie process is dead process. It is meaningless to profile it. It's better to ignore it for user profile.

I recently posted different patch for same issue:
http://marc.info/?l=linux-kernel&m=148153895827359&w=2

jirka

author
#2 lore raw
Quoted 27 lines · On Wed, Dec 14, 2016 at 12:48:05PM -0500, [email protected] wrote:
Quoted 23 lines

From: Kan Liang <[email protected]>

If user has zombie process, the perf record -u will error out. Here is an example.

 $ ./testd &
 [1] 23796
 $ sudo perf record -e cycles -u kan
 Error:
 The sys_perf_event_open() syscall returned with 3 (No such process)
for  event (cycles).
 /bin/dmesg may provide additional information.
 No CONFIG_PERF_EVENTS=y kernel support configured?

The source code of testd is as below.
int main() {

	if (fork())
	{
		while (1);
	}
	return 0;

}

Quoted 5 lines
Quoted 2 lines

Zombie process is dead process. It is meaningless to profile it. It's better to ignore it for user profile.

I recently posted different patch for same issue:
http://marc.info/?l=linux-kernel&m=148153895827359&w=2

The change as below make me confuse.
+ /* The system wide setup does not work with threads. */
+ if (!evsel->system_wide)
+ return false;
It looks the meaning of the comments is inconsistent with the code.

Your original patch doesn't work well with the issue. But if I change the above code as below, the issue is fixed.

	if (evsel->system_wide)
		return false;

Thanks,
Kan

#3 lore raw

Em Wed, Dec 14, 2016 at 06:26:02PM +0000, Liang, Kan escreveu:

Quoted 14 lines
Quoted 7 lines · On Wed, Dec 14, 2016 at 12:48:05PM -0500, [email protected] wrote:
Quoted 3 lines

From: Kan Liang <[email protected]>
Zombie process is dead process. It is meaningless to profile it. It's better to ignore it for user profile.

I recently posted different patch for same issue:
http://marc.info/?l=linux-kernel&m=148153895827359&w=2

The change as below make me confuse.

(quoted diff) +3 -0
+	/* The system wide setup does not work with threads. */
+	if (!evsel->system_wide)
+		return false;

It looks the meaning of the comments is inconsistent with the code.

Quoted 4 lines

Your original patch doesn't work well with the issue. But if I change the above code as below, the issue is fixed.

	if (evsel->system_wide)
		return false;

yeah, Namhyung noticed that, Jiri sent a fixed patch, I have it already in my perf/core branch.

- Arnaldo

author
#4 lore raw

.

Quoted 15 lines

Em Wed, Dec 14, 2016 at 06:26:02PM +0000, Liang, Kan escreveu:

Quoted 13 lines
Quoted 7 lines · On Wed, Dec 14, 2016 at 12:48:05PM -0500, [email protected] wrote:
Quoted 3 lines

From: Kan Liang <[email protected]> Zombie process is dead process. It is meaningless to profile it.
It's better to ignore it for user profile.

I recently posted different patch for same issue:
http://marc.info/?l=linux-kernel&m=148153895827359&w=2

The change as below make me confuse.

(quoted diff) +3 -0
+	/* The system wide setup does not work with threads. */
+	if (!evsel->system_wide)
+		return false;

It looks the meaning of the comments is inconsistent with the code.

Quoted 8 lines
Quoted 4 lines

Your original patch doesn't work well with the issue. But if I change the above code as below, the issue is fixed.

	if (evsel->system_wide)
		return false;

yeah, Namhyung noticed that, Jiri sent a fixed patch, I have it already in my perf/core branch.

That's great. Then Jiri's patch would work on my case. Please ignore this patch.

Thanks,
Kan