From: stephane eranian on
Hi,

I believe there is another bug related to timing and I am not
sure if this patch fixes it too.

If you add TIME_ENABLED|TIME_RUNNING in samples, they
will get zero all the time (ENA, RUN):

$ libpfm/perf_examples/task_smpl noploop 4
period=240000000 freq=0
58 PERF_COUNT_HW_CPU_CYCLES
59 PERF_COUNT_HW_INSTRUCTIONS
noploop for 4 seconds

IIP:0x0000000000400651 PID:10613 TID:10613 TIME:27480525003765
STREAM_ID:58 PERIOD:240000000 ENA=0 RUN=0 NR=2
240000257 PERF_COUNT_HW_CPU_CYCLES (58)
239740143 PERF_COUNT_HW_INSTRUCTIONS (59)

IP:0x0000000000400651 PID:10613 TID:10613 TIME:27480626252349
STREAM_ID:58 PERIOD:240000000 ENA=0 RUN=0 NR=2
480000257 PERF_COUNT_HW_CPU_CYCLES (58)
479688363 PERF_COUNT_HW_INSTRUCTIONS (59)

Events are also grouped in this example. Both this same issue exists
also when only
one event is used. I suspect an update_event_times() or
update_group_times() is also
missing on the sampling path in perf_overflow_handler().


On Tue, May 11, 2010 at 4:42 PM, Peter Zijlstra <peterz(a)infradead.org> wrote:
> On Fri, 2010-05-07 at 18:56 -0700, Corey Ashford wrote:
>> Hi,
>>
>> There appears to be a bug in the kernel related to reading up the time
>> running and enabled values for events that are in a group.  The group
>> leader's time running and time enable values look correct, but all of
>> the other group members have a zero value for their time running and
>> time enabled fields.
>>
>> This happens only when remote monitoring a process (perhaps only after
>> it has terminated)... when self monitoring, the time running/enabled
>> values come out non-zero, and the values are the same for all of the
>> counters (as one would expect since they can be enabled/disabled
>> simultaneously).
>>
>> I've attached a test case which you can place in the tools/perf
>> subdirectory and compile with just:
>>
>> gcc -o show_re_bug show_re_bug.c
>
> The below seems to fix this for me.
>
> ---
> Subject: perf: Fix exit() vs event-groups
> From: Peter Zijlstra <a.p.zijlstra(a)chello.nl>
> Date: Tue May 11 16:19:10 CEST 2010
>
> Corey reported that the value scale times of group siblings are not
> updated when the monitored task dies.
>
> The problem appears to be that we only update the group leader's
> time values, fix it by updating the whole group.
>
> Reported-by: Corey Ashford <cjashfor(a)linux.vnet.ibm.com>
> Signed-off-by: Peter Zijlstra <a.p.zijlstra(a)chello.nl>
> Cc: Paul Mackerras <paulus(a)samba.org>
> Cc: stable(a)kernel.org
> ---
>  kernel/perf_event.c |   44 +++++++++++++++++++++++---------------------
>  1 file changed, 23 insertions(+), 21 deletions(-)
>
> Index: linux-2.6/kernel/perf_event.c
> ===================================================================
> --- linux-2.6.orig/kernel/perf_event.c
> +++ linux-2.6/kernel/perf_event.c
> @@ -255,6 +255,18 @@ static void update_event_times(struct pe
>        event->total_time_running = run_end - event->tstamp_running;
>  }
>
> +/*
> + * Update total_time_enabled and total_time_running for all events in a group.
> + */
> +static void update_group_times(struct perf_event *leader)
> +{
> +       struct perf_event *event;
> +
> +       update_event_times(leader);
> +       list_for_each_entry(event, &leader->sibling_list, group_entry)
> +               update_event_times(event);
> +}
> +
>  static struct list_head *
>  ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
>  {
> @@ -320,7 +332,7 @@ list_del_event(struct perf_event *event,
>        if (event->group_leader != event)
>                event->group_leader->nr_siblings--;
>
> -       update_event_times(event);
> +       update_group_times(event);
>
>        /*
>         * If event was in error state, then keep it
> @@ -502,18 +514,6 @@ retry:
>  }
>
>  /*
> - * Update total_time_enabled and total_time_running for all events in a group.
> - */
> -static void update_group_times(struct perf_event *leader)
> -{
> -       struct perf_event *event;
> -
> -       update_event_times(leader);
> -       list_for_each_entry(event, &leader->sibling_list, group_entry)
> -               update_event_times(event);
> -}
> -
> -/*
>  * Cross CPU call to disable a performance event
>  */
>  static void __perf_event_disable(void *info)
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Stephane Eranian on
On Tue, May 11, 2010 at 10:11 PM, Peter Zijlstra <peterz(a)infradead.org> wrote:
> On Tue, 2010-05-11 at 21:55 +0200, stephane eranian wrote:
>> Events are also grouped in this example. Both this same issue exists
>> also when only
>> one event is used. I suspect an update_event_times() or
>> update_group_times() is also
>> missing on the sampling path in perf_overflow_handler().
>
> Hrmm, you're right, except fixing that is non-trivial.
> update_event_times() wants to be called with ctx->lock held and we can't
> do that from NMI context.
>
Why do you need the lock when the event is running on a CPU in NMI context?
What action issued remotely could modify the data used by this function without
the need for IPI?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/