perf/core: Fix group scheduling with mixed hw and sw events

When hw and sw events are mixed in the same group, they are all attached
to the hw perf_event_context. This sometimes requires moving group of
perf_event to a different context.

We found a bug in how the kernel handles this, for example if we do:

   perf stat -e '{faults,ref-cycles,faults}'  -I 1000

     1.005591180              1,297      faults
     1.005591180        457,476,576      ref-cycles
     1.005591180    <not supported>      faults

First, sw event "faults" is attached to the sw context, and becomes the
group leader. Then, hw event "ref-cycles" is attached, so both events
are moved to the hw context. Last, another sw "faults" tries to attach,
but it fails because of mismatch between the new target ctx (from sw
pmu) and the group_leader's ctx (hw context, same as ref-cycles).

The broken condition is:
   group_leader is sw event;
   group_leader is on hw context;
   add a sw event to the group.

Fix this scenario by checking group_leader's context (instead of just
event type). If group_leader is on hw context, use the ->pmu of this
context to look up context for the new event.

Signed-off-by: Song Liu <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Cc: <[email protected]>
Cc: Alexander Shishkin <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>
Cc: Jiri Olsa <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Stephane Eranian <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Vince Weaver <[email protected]>
Fixes: b04243e ("perf: Complete software pmu grouping")
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Ingo Molnar <[email protected]>

Author: liu-song-6

include/linux/perf_event.h

@@ -1016,6 +1016,14 @@ static inline int is_software_event(struct perf_event *event)
 	return event->event_caps & PERF_EV_CAP_SOFTWARE;
 }
 
+/*
+ * Return 1 for event in sw context, 0 for event in hw context
+ */
+static inline int in_software_context(struct perf_event *event)
+{
+	return event->ctx->pmu->task_ctx_nr == perf_sw_context;
+}
+
 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 
 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);

kernel/events/core.c

@@ -10521,19 +10521,20 @@ SYSCALL_DEFINE5(perf_event_open,
 	if (pmu->task_ctx_nr == perf_sw_context)
 		event->event_caps |= PERF_EV_CAP_SOFTWARE;
 
-	if (group_leader &&
-	    (is_software_event(event) != is_software_event(group_leader))) {
-		if (is_software_event(event)) {
+	if (group_leader) {
+		if (is_software_event(event) &&
+		    !in_software_context(group_leader)) {
 			/*
-			 * If event and group_leader are not both a software
-			 * event, and event is, then group leader is not.
+			 * If the event is a sw event, but the group_leader
+			 * is on hw context.
 			 *
-			 * Allow the addition of software events to !software
-			 * groups, this is safe because software events never
-			 * fail to schedule.
+			 * Allow the addition of software events to hw
+			 * groups, this is safe because software events
+			 * never fail to schedule.
 			 */
-			pmu = group_leader->pmu;
-		} else if (is_software_event(group_leader) &&
+			pmu = group_leader->ctx->pmu;
+		} else if (!is_software_event(event) &&
+			   is_software_event(group_leader) &&
 			   (group_leader->group_caps & PERF_EV_CAP_SOFTWARE)) {
 			/*
 			 * In case the group is a pure software group, and we