Re: perf event group siblings not counting in mainline

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2012-01-10 at 11:09 +0000, Will Deacon wrote:
> On Sat, Dec 24, 2011 at 12:54:15PM +0000, Will Deacon wrote:
> > If I try using perf with event groups on a mainline kernel, only the group
> > leader counts:
> > 
> > 
> > linaro@mashed-potato:~$ perf stat --no-scale -g -e task-clock -e cpu-clock -e cs -- ls
> > hwb  linux
> > 
> >  Performance counter stats for 'ls':
> > 
> >          17.545541 task-clock                #    0.634 CPUs utilized
> >           0.000000 cpu-clock
> >                  0 cs                        #    0.000 M/sec
> > 
> >        0.027684984 seconds time elapsed
> > 
> > 
> > This is fixed by b79387ef ("perf: Fix enable_on_exec for sibling events")
> > in -next (20111222) but I can't see this queued anywhere for 3.2 (I thought
> > I might see it in -rc7).
> > 
> > Is this going in for the final cut, or will groups just be borked until 3.3?
> 
> Now that 3.2 is out with this regression, would it be possible to get this
> (b79387ef ("perf: Fix enable_on_exec for sibling events")) into -stable please?

How can this be a regression? Its always been like that. At worst its a
broken new feature.

Also, you really shouldn't set enable_on_exec on non group leaders,
that's just daft. But yes that patch sorts it. Something like the below
should too.

---
 tools/perf/builtin-stat.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index f5d2a63..623519f 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -298,7 +298,7 @@ static int create_perf_stat_counter(struct perf_evsel *evsel,
 						group, group_fd);
 	if (target_pid == -1 && target_tid == -1) {
 		attr->disabled = 1;
-		attr->enable_on_exec = 1;
+		attr->enable_on_exec = !group_fd;
 	}
 
 	return perf_evsel__open_per_thread(evsel, evsel_list->threads,