perf: x86 filter_events() - use hw event id ?

From: Sukadev Bhattiprolu <[email protected]>
Date: Tue, 30 Oct 2012 23:05:05 -0700
Subject: [PATCH] perf: x86 filter_events() - use hw event id ?

The ->event_map() operation expects to index through the _hardware event id_. But filter_events() function is calling ->event_map() with a loop index i.

Suppose a processor does not implement say PERF_COUNT_HW_BUS_CYCLES (6), but implements, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND (7).

->event_map() for "bus cycles" returns 0, we push the "stalled-cycles-frontend" event up by one in the events_attr[].

The index of "stalled-cycles-frontend" in the events_attr[] is 6 which is different from its hardware event id 7.

The next and subsequent calls to ->event_map() will use this index in the modified events_attr[] table. All subsequent event ids appear unimplemented because they continue to check the same index (6 in this case).

Should'nt we be checking the hardware event id instead ?

I have not tested this on x86 - only on my port of the sysfs events to Power.

---
 arch/x86/kernel/cpu/perf_event.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)

index 0a55ab2..ae88512 100644
@@ -1325,12 +1325,21 @@ struct perf_pmu_events_attr {
  * Remove all undefined events (x86_pmu.event_map(id) == 0)
  * out of events_attr attributes.
  */
+#define PMU_ATTR(a) container_of((a), struct perf_pmu_events_attr, attr)
+#define DEV_ATTR(a) container_of((a), struct device_attribute, attr)
+
 static void __init filter_events(struct attribute **attrs)
 {
 	int i, j;
+	struct perf_pmu_events_attr *pmu_attr;
+	struct device_attribute *dev_attr;
+	
 
 	for (i = 0; attrs[i]; i++) {
-		if (x86_pmu.event_map(i))
+		dev_attr = DEV_ATTR(attrs[i]);
+		pmu_attr = PMU_ATTR(dev_attr);
+
+		if (x86_pmu.event_map(pmu_attr->id))
 			continue;
 
 		for (j = i; attrs[j]; j++)

1.7.1