perf: Fix throttle logic

It was possible to call pmu::start() on an already running event. In
particular this lead so some wreckage as the hrtimer events would
re-initialize active timers.

This was due to throttled events being activated again by scheduling.
Scheduling in a context would add and force start events, resulting in
running events with a possible throttle status. The next tick to hit
that task will then try to unthrottle the event and call ->start() on
an already running event.

Reported-by: Jeff Moyer <jmoyer@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index 999835b6112bc0705e6a511a63889a89c920b9d7..656222fcf767e4442acaf291a9bca566f9332d00 100644 (file)
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -782,6 +782,10 @@ retry:
        raw_spin_unlock_irq(&ctx->lock);
 }
 
+#define MAX_INTERRUPTS (~0ULL)
+
+static void perf_log_throttle(struct perf_event *event, int enable);
+
 static int
 event_sched_in(struct perf_event *event,
                 struct perf_cpu_context *cpuctx,
@@ -794,6 +798,17 @@ event_sched_in(struct perf_event *event,
 
        event->state = PERF_EVENT_STATE_ACTIVE;
        event->oncpu = smp_processor_id();
+
+       /*
+        * Unthrottle events, since we scheduled we might have missed several
+        * ticks already, also for a heavily scheduling task there is little
+        * guarantee it'll get a tick in a timely manner.
+        */
+       if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
+               perf_log_throttle(event, 1);
+               event->hw.interrupts = 0;
+       }
+
        /*
         * The new state must be visible before we turn it on in the hardware:
         */
@@ -1596,10 +1611,6 @@ void __perf_event_task_sched_in(struct task_struct *task)
        }
 }
 
-#define MAX_INTERRUPTS (~0ULL)
-
-static void perf_log_throttle(struct perf_event *event, int enable);
-
 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
 {
        u64 frequency = event->attr.sample_freq;