Skip to contentSkip to conversation

perf, tools: Stop perf stat -p when profiled process exits v3

PATCH In review Andi Kleen 2 participants 3 messages Sep 13, 2012 perf events

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

PATCH 1/1 perf, tools: Stop perf stat -p when profiled process exits v3

From: Andi Kleen <[email protected]>

When counting a process with perf stat -p check if the process died and exit collection if yes.

v2: Add more checks, handle non -p again. Handle /proc not there.
v3: Handle multi pid case. Fix non /proc error path

---
 tools/perf/builtin-stat.c |   14 +++++++++++++-
 1 files changed, 13 insertions(+), 1 deletions(-)
tools/perf/builtin-stat.c +13 -1
index 861f0ae..b5e7df2 100644
@@ -536,7 +536,19 @@ static int run_perf_stat(int argc __used, const char **argv)
 		if (WIFSIGNALED(status))
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
-		while(!done) sleep(1);
+		char piddir[40];
+		int check_proc = target.pid &&
+				 access("/proc", X_OK) == 0 &&
+				 !strchr(target.pid, ',');
+		if (check_proc)
+			snprintf(piddir, sizeof piddir, "/proc/%d", 
+				 atoi(target.pid));
+		while(!done) {
+			sleep(1);
+			if (check_proc && access(piddir, X_OK) < 0 && 
+			    errno == ENOENT)
+				break;
+		}
 	}
 
 	t1 = rdclock();
Signature · 1 line
1.7.7.6
#1 lore raw
Quoted 25 lines · Andi Kleen

From: Andi Kleen <[email protected]>

When counting a process with perf stat -p check if the process died and exit collection if yes.

v2: Add more checks, handle non -p again. Handle /proc not there.
v3: Handle multi pid case. Fix non /proc error path

---
  tools/perf/builtin-stat.c |   14 +++++++++++++-
  1 files changed, 13 insertions(+), 1 deletions(-)
tools/perf/builtin-stat.c +4 -1
index 861f0ae..b5e7df2 100644
@@ -536,7 +536,19 @@ static int run_perf_stat(int argc __used, const char **argv)
  		if (WIFSIGNALED(status))
  			psignal(WTERMSIG(status), argv[0]);
  	} else {
-		while(!done) sleep(1);
+		char piddir[40];
+		int check_proc = target.pid &&
+				 access("/proc", X_OK) == 0 &&
+				 !strchr(target.pid, ',');

tools/perf/builtin-stat.c static int run_perf_stat(int argc __used, const char **argv) · line 552

Quoted 10 lines
tools/perf/builtin-stat.c +9 -0
+		if (check_proc)
+			snprintf(piddir, sizeof piddir, "/proc/%d",
+				 atoi(target.pid));
+		while(!done) {
+			sleep(1);
+			if (check_proc && access(piddir, X_OK) < 0 &&
+			    errno == ENOENT)
+				break;
+		}
  	}

I still think this is the wrong approach. A more complete solution would monitor all of the targets given by the user. e.g.,

1. create a perf_target__parse_pid - all it needs to do is:
struct strlist *slist = strlist__new(false, pid_str);
and save that into the target struct. That line is currently done in thread_map.c, thread_map__new_by_pid_str() and thread_map__new_by_tid_str. Those functions can be altered to take the slist instead of the string.

2. create a perf_target__is_alive that runs through each entry in the slist (strlist__for_each(pos, slist)) and checks that it still exists. As long as 1 target is alive it returns true. If no pid or tid is specified it always returns true.

The is_alive function can be used by perf-top, perf-stat and perf-record to know that its target has died so it can stop monitoring.

David

author
#2 lore raw
Quoted 2 lines

The is_alive function can be used by perf-top, perf-stat and perf-record to know that its target has died so it can stop monitoring.

Sorry that's to complicated and overengineered for me.
My simple patch works for me at least. It's not solving world hunger, but it wasn't intended to do that.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.