perf_events: Fix for transaction recovery in group_sched_in()

This new version (see commit 8e5fc1a) is much simpler and ensures that
in case of error in group_sched_in() during event_sched_in(), the
events up to the failed event go through regular event_sched_out().
But the failed event and the remaining events in the group have their
timings adjusted as if they had also gone through event_sched_in() and
event_sched_out(). This ensures timing uniformity across all events in
a group. This also takes care of the tstamp_stopped problem in case
the group could never be scheduled. The tstamp_stopped is updated as
if the event had actually run.

With this patch, the following now reports correct time_enabled,
in case the NMI watchdog is active:

$ task -e unhalted_core_cycles,instructions_retired,baclears,baclears
noploop 1
noploop for 1 seconds

0 unhalted_core_cycles (100.00% scaling, ena=997,552,872, run=0)
0 instructions_retired (100.00% scaling, ena=997,552,872, run=0)
0 baclears (100.00% scaling, ena=997,552,872, run=0)
0 baclears (100.00% scaling, ena=997,552,872, run=0)

And the older test case also works:

$ task -einstructions_retired,baclears,baclears -e
unhalted_core_cycles,baclears,baclears sleep 5

1680885 instructions_retired (69.39% scaling, ena=950756, run=291006)
  10735 baclears (69.39% scaling, ena=950756, run=291006)
  10735 baclears (69.39% scaling, ena=950756, run=291006)

      0 unhalted_core_cycles (100.00% scaling, ena=817932, run=0)
      0 baclears (100.00% scaling, ena=817932, run=0)
      0 baclears (100.00% scaling, ena=817932, run=0)

Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <4cbeeebc.8ee7d80a.5a28.0d5f@mx.google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 20 insertions, 2 deletions

diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index 39afdb07d75..517d827f498 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -691,6 +691,8 @@ group_sched_in(struct perf_event *group_event,
 {
 	struct perf_event *event, *partial_group = NULL;
 	struct pmu *pmu = group_event->pmu;
+	u64 now = ctx->time;
+	bool simulate = false;
 
 	if (group_event->state == PERF_EVENT_STATE_OFF)
 		return 0;
@@ -719,11 +721,27 @@ group_error:
 	/*
 	 * Groups can be scheduled in as one unit only, so undo any
 	 * partial group before returning:
+	 * The events up to the failed event are scheduled out normally,
+	 * tstamp_stopped will be updated.
+	 *
+	 * The failed events and the remaining siblings need to have
+	 * their timings updated as if they had gone thru event_sched_in()
+	 * and event_sched_out(). This is required to get consistent timings
+	 * across the group. This also takes care of the case where the group
+	 * could never be scheduled by ensuring tstamp_stopped is set to mark
+	 * the time the event was actually stopped, such that time delta
+	 * calculation in update_event_times() is correct.
 	 */
 	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
 		if (event == partial_group)
-			break;
-		event_sched_out(event, cpuctx, ctx);
+			simulate = true;
+
+		if (simulate) {
+			event->tstamp_running += now - event->tstamp_stopped;
+			event->tstamp_stopped = now;
+		} else {
+			event_sched_out(event, cpuctx, ctx);
+		}
 	}
 	event_sched_out(group_event, cpuctx, ctx);