On Fri, Jul 29, 2016 at 11:41:11PM +0200, Vegard Nossum wrote:
> Digging a bit deeper into this, it seems the event itself is getting
> created by perf_event_open() and it gets added to the pmu_event_list
> through:
>
> perf_event_open()
> - perf_event_alloc()
> - account_event()
> - account_pmu_sb_event()
> - attach_sb_event()
>
> so at this point the event is being attached but its ->ctx is still
> NULL. It seems like ->ctx is set just a bit later in
> perf_event_open(), though.
>
> But before that, __schedule() comes along and creates a stack trace
> similar to the one above:
>
> __schedule()
> - __perf_event_task_sched_out()
> - perf_iterate_sb()
> - perf_iterate_sb_cpu()
> - event_filter_match()
> - perf_cgroup_match()
> - __get_cpu_context()
> - (dereference ctx which is NULL)
>
> So I guess the question is... should the event be attached (= put on
> the list) before ->ctx gets set? Or should the cgroup code check for a
> NULL ->ctx?
Does this fix it? Ordering is a bit of a mess, adding the events to the
list _after_ they've been installed has the risk of missing things I
think, nor does that result in particularly nice code.
Then again, this isn't pretty either.
---
kernel/events/core.c | 23 ++++++++++++++++++-----
1 file changed, 18 insertions(+), 5 deletions(-)
diff --git a/kernel/events/core.c b/kernel/events/core.c
index a19550d80ab1..87d02b8cb87e 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -1716,8 +1716,8 @@ static inline int pmu_filter_match(struct perf_event
*event)
static inline int
event_filter_match(struct perf_event *event)
{
- return (event->cpu == -1 || event->cpu == smp_processor_id())
- && perf_cgroup_match(event) && pmu_filter_match(event);
+ return (event->cpu == -1 || event->cpu == smp_processor_id()) &&
+ perf_cgroup_match(event) && pmu_filter_match(event);
}
static void
@@ -1737,8 +1737,8 @@ event_sched_out(struct perf_event *event,
* maintained, otherwise bogus information is return
* via read() for time_enabled, time_running:
*/
- if (event->state == PERF_EVENT_STATE_INACTIVE
- && !event_filter_match(event)) {
+ if (event->state == PERF_EVENT_STATE_INACTIVE &&
+ !event_filter_match(event)) {
delta = tstamp - event->tstamp_stopped;
event->tstamp_running += delta;
event->tstamp_stopped = tstamp;
@@ -2236,10 +2236,15 @@ perf_install_in_context(struct perf_event_context *ctx,
lockdep_assert_held(&ctx->mutex);
- event->ctx = ctx;
if (event->cpu != -1)
event->cpu = cpu;
+ /*
+ * Ensures that if we can observe event->ctx, both the event and ctx
+ * will be 'complete'. See perf_iterate_sb_cpu().
+ */
+ smp_store_release(&event->ctx, ctx);
+
if (!task) {
cpu_function_call(cpu, __perf_install_in_context, event);
return;
@@ -5969,6 +5974,14 @@ static void perf_iterate_sb_cpu(perf_iterate_f output,
void *data)
struct perf_event *event;
list_for_each_entry_rcu(event, &pel->list, sb_list) {
+ /*
+ * Skip events that are not fully formed yet; ensure that
+ * if we observe event->ctx, both event and ctx will be
+ * complete enough. See perf_install_in_context().
+ */
+ if (!smp_load_acquire(&event->ctx))
+ continue;
+
if (event->state < PERF_EVENT_STATE_INACTIVE)
continue;
if (!event_filter_match(event))
