perf_events: Restore sanity to scaling land
authorPeter Zijlstra <peterz@infradead.org>
Mon, 23 Nov 2009 14:00:36 +0000 (15:00 +0100)
committerIngo Molnar <mingo@elte.hu>
Mon, 23 Nov 2009 14:22:19 +0000 (15:22 +0100)
It is quite possible to call update_event_times() on a context
that isn't actually running and thereby confuse the thing.

perf stat was reporting !100% scale values for software counters
(2e2af50b perf_events: Disable events when we detach them,
solved the worst of that, but there was still some left).

The thing that happens is that because we are not self-reaping
(we have a caring parent) there is a time between the last
schedule (out) and having do_exit() called which will detach the
events.

This period would be accounted as enabled,!running because the
event->state==INACTIVE, even though !event->ctx->is_active.

Similar issues could have been observed by calling read() on a
event while the attached task was not scheduled in.

Solve this by teaching update_event_times() about
ctx->is_active.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <1258984836.4531.480.camel@laptop>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/perf_event.c

index 0b0d5f72fe7d83a3c6695ed9d589f090b96a7c3e..0aafe85362fd0b02098198b841e8667e70acb033 100644 (file)
@@ -274,7 +274,12 @@ static void update_event_times(struct perf_event *event)
            event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
                return;
 
-       event->total_time_enabled = ctx->time - event->tstamp_enabled;
+       if (ctx->is_active)
+               run_end = ctx->time;
+       else
+               run_end = event->tstamp_stopped;
+
+       event->total_time_enabled = run_end - event->tstamp_enabled;
 
        if (event->state == PERF_EVENT_STATE_INACTIVE)
                run_end = event->tstamp_stopped;