powerpc/perf_event: Fix for power_pmu_disable()
authorMatt Evans <matt@ozlabs.org>
Mon, 5 Jul 2010 17:36:32 +0000 (17:36 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Thu, 8 Jul 2010 08:11:37 +0000 (18:11 +1000)
When power_pmu_disable() removes the given event from a particular index into
cpuhw->event[], it shuffles down higher event[] entries.  But, this array is
paired with cpuhw->events[] and cpuhw->flags[] so should shuffle them
similarly.

If these arrays get out of sync, code such as power_check_constraints() will
fail.  This caused a bug where events were temporarily disabled and then failed
to be re-enabled; subsequent code tried to write_pmc() with its (disabled) idx
of 0, causing a message "oops trying to write PMC0".  This triggers this bug on
POWER7, running a miss-heavy test:

  perf record -e L1-dcache-load-misses -e L1-dcache-store-misses ./misstest

Signed-off-by: Matt Evans <matt@ozlabs.org>
Acked-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/kernel/perf_event.c

index 43b83c35cf54b0dfb07bc5abef8ee558d5f3cd9c..5c14ffe5125813d19e6793ebf69e132fa19dbffe 100644 (file)
@@ -791,8 +791,11 @@ static void power_pmu_disable(struct perf_event *event)
        cpuhw = &__get_cpu_var(cpu_hw_events);
        for (i = 0; i < cpuhw->n_events; ++i) {
                if (event == cpuhw->event[i]) {
-                       while (++i < cpuhw->n_events)
+                       while (++i < cpuhw->n_events) {
                                cpuhw->event[i-1] = cpuhw->event[i];
+                               cpuhw->events[i-1] = cpuhw->events[i];
+                               cpuhw->flags[i-1] = cpuhw->flags[i];
+                       }
                        --cpuhw->n_events;
                        ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
                        if (event->hw.idx) {