Handle perf counter's inability to setup counters before cpu is brought online

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

diff --git a/lttng-context-perf-counters.c b/lttng-context-perf-counters.c
index 2e194991daf79f24d4bb9f255e99f7f1919cf2d5..56e0d25035ae7f6a09eec09114ced55524b81159 100644 (file)
--- a/lttng-context-perf-counters.c
+++ b/lttng-context-perf-counters.c
@@ -36,8 +36,19 @@ void perf_counter_record(struct lttng_ctx_field *field,
        uint64_t value;
 
        event = field->u.perf_counter.e[ctx->cpu];
-       event->pmu->read(event);
-       value = local64_read(&event->count);
+       if (likely(event)) {
+               event->pmu->read(event);
+               value = local64_read(&event->count);
+       } else {
+               /*
+                * Perf chooses not to be clever and not to support enabling a
+                * perf counter before the cpu is brought up. Therefore, we need
+                * to support having events coming (e.g. scheduler events)
+                * before the counter is setup. Write an arbitrary 0 in this
+                * case.
+                */
+               value = 0;
+       }
        lib_ring_buffer_align_ctx(ctx, ltt_alignof(value));
        chan->ops->event_write(ctx, &value, sizeof(value));
 }
@@ -90,7 +101,7 @@ int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
                container_of(nb, struct lttng_ctx_field, u.perf_counter.nb);
        struct perf_event **events = field->u.perf_counter.e;
        struct perf_event_attr *attr = field->u.perf_counter.attr;
-       
+       struct perf_event *pevent;
 
        if (!field->u.perf_counter.hp_enable)
                return NOTIFY_OK;
@@ -98,16 +109,21 @@ int __cpuinit lttng_perf_counter_cpu_hp_callback(struct notifier_block *nb,
        switch (action) {
        case CPU_ONLINE:
        case CPU_ONLINE_FROZEN:
-               events[cpu] = perf_event_create_kernel_counter(attr,
+               pevent = perf_event_create_kernel_counter(attr,
                                cpu, NULL, overflow_callback);
-               if (!events[cpu])
+               if (!pevent)
                        return NOTIFY_BAD;
+               barrier();      /* Create perf counter before setting event */
+               events[cpu] = pevent;
                break;
        case CPU_UP_CANCELED:
        case CPU_UP_CANCELED_FROZEN:
        case CPU_DEAD:
        case CPU_DEAD_FROZEN:
-               perf_event_release_kernel(events[cpu]);
+               pevent = events[cpu];
+               events[cpu] = NULL;
+               barrier();      /* NULLify event before perf counter teardown */
+               perf_event_release_kernel(pevent);
                break;
        }
        return NOTIFY_OK;