perf_counter: Report the cloning task as parent on perf_counter_fork()
authorPeter Zijlstra <peterz@infradead.org>
Thu, 13 Aug 2009 14:14:42 +0000 (16:14 +0200)
committerIngo Molnar <mingo@elte.hu>
Thu, 13 Aug 2009 14:17:15 +0000 (16:17 +0200)
A bug in (9f498cc: perf_counter: Full task tracing) makes
profiling multi-threaded apps it go belly up.

[ output as: (PID:TID):(PPID:PTID) ]

 # ./perf report -D | grep FORK
0x4b0 [0x18]: PERF_EVENT_FORK: (3237:3237):(3236:3236)
0xa10 [0x18]: PERF_EVENT_FORK: (3237:3238):(3236:3236)
0xa70 [0x18]: PERF_EVENT_FORK: (3237:3239):(3236:3236)
0xad0 [0x18]: PERF_EVENT_FORK: (3237:3240):(3236:3236)
0xb18 [0x18]: PERF_EVENT_FORK: (3237:3241):(3236:3236)

Shows us that the test (27d028d perf report: Update for the new
FORK/EXIT events) in builtin-report.c:

        /*
         * A thread clone will have the same PID for both
         * parent and child.
         */
        if (thread == parent)
                return 0;

Will clearly fail.

The problem is that perf_counter_fork() reports the actual
parent, instead of the cloning thread.

Fixing that (with the below patch), yields:

 # ./perf report -D | grep FORK
0x4c8 [0x18]: PERF_EVENT_FORK: (1590:1590):(1589:1589)
0xbd8 [0x18]: PERF_EVENT_FORK: (1590:1591):(1590:1590)
0xc80 [0x18]: PERF_EVENT_FORK: (1590:1592):(1590:1590)
0x3338 [0x18]: PERF_EVENT_FORK: (1590:1593):(1590:1590)
0x66b0 [0x18]: PERF_EVENT_FORK: (1590:1594):(1590:1590)

Which both makes more sense and doesn't confuse perf report
anymore.

Reported-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: paulus@samba.org
Cc: Anton Blanchard <anton@samba.org>
Cc: Arjan van de Ven <arjan@infradead.org>
LKML-Reference: <1250172882.5241.62.camel@twins>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/perf_counter.c

index 3f841beefc7b2869d9b92a3341dca0620f5938b0..534e20d14d631b44cefc136f99200b2257af378f 100644 (file)
@@ -3028,10 +3028,10 @@ static void perf_counter_task_output(struct perf_counter *counter,
                return;
 
        task_event->event.pid = perf_counter_pid(counter, task);
-       task_event->event.ppid = perf_counter_pid(counter, task->real_parent);
+       task_event->event.ppid = perf_counter_pid(counter, current);
 
        task_event->event.tid = perf_counter_tid(counter, task);
-       task_event->event.ptid = perf_counter_tid(counter, task->real_parent);
+       task_event->event.ptid = perf_counter_tid(counter, current);
 
        perf_output_put(&handle, task_event->event);
        perf_output_end(&handle);