Merge branch 'perf/urgent' into perf/core, to pick up dependencies
[cascardo/linux.git] / kernel / events / core.c
index 5650f53..ca4fde5 100644 (file)
@@ -1832,6 +1832,8 @@ group_sched_out(struct perf_event *group_event,
        struct perf_event *event;
        int state = group_event->state;
 
+       perf_pmu_disable(ctx->pmu);
+
        event_sched_out(group_event, cpuctx, ctx);
 
        /*
@@ -1840,6 +1842,8 @@ group_sched_out(struct perf_event *group_event,
        list_for_each_entry(event, &group_event->sibling_list, group_entry)
                event_sched_out(event, cpuctx, ctx);
 
+       perf_pmu_enable(ctx->pmu);
+
        if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
                cpuctx->exclusive = 0;
 }
@@ -2837,19 +2841,36 @@ unlock:
        }
 }
 
+static DEFINE_PER_CPU(struct list_head, sched_cb_list);
+
 void perf_sched_cb_dec(struct pmu *pmu)
 {
+       struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
+
        this_cpu_dec(perf_sched_cb_usages);
+
+       if (!--cpuctx->sched_cb_usage)
+               list_del(&cpuctx->sched_cb_entry);
 }
 
+
 void perf_sched_cb_inc(struct pmu *pmu)
 {
+       struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
+
+       if (!cpuctx->sched_cb_usage++)
+               list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
+
        this_cpu_inc(perf_sched_cb_usages);
 }
 
 /*
  * This function provides the context switch callback to the lower code
  * layer. It is invoked ONLY when the context switch callback is enabled.
+ *
+ * This callback is relevant even to per-cpu events; for example multi event
+ * PEBS requires this to provide PID/TID information. This requires we flush
+ * all queued PEBS records before we context switch to a new task.
  */
 static void perf_pmu_sched_task(struct task_struct *prev,
                                struct task_struct *next,
@@ -2857,34 +2878,24 @@ static void perf_pmu_sched_task(struct task_struct *prev,
 {
        struct perf_cpu_context *cpuctx;
        struct pmu *pmu;
-       unsigned long flags;
 
        if (prev == next)
                return;
 
-       local_irq_save(flags);
+       list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
+               pmu = cpuctx->unique_pmu; /* software PMUs will not have sched_task */
 
-       rcu_read_lock();
-
-       list_for_each_entry_rcu(pmu, &pmus, entry) {
-               if (pmu->sched_task) {
-                       cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
-
-                       perf_ctx_lock(cpuctx, cpuctx->task_ctx);
-
-                       perf_pmu_disable(pmu);
+               if (WARN_ON_ONCE(!pmu->sched_task))
+                       continue;
 
-                       pmu->sched_task(cpuctx->task_ctx, sched_in);
+               perf_ctx_lock(cpuctx, cpuctx->task_ctx);
+               perf_pmu_disable(pmu);
 
-                       perf_pmu_enable(pmu);
+               pmu->sched_task(cpuctx->task_ctx, sched_in);
 
-                       perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
-               }
+               perf_pmu_enable(pmu);
+               perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
        }
-
-       rcu_read_unlock();
-
-       local_irq_restore(flags);
 }
 
 static void perf_event_switch(struct task_struct *task,
@@ -10426,6 +10437,8 @@ static void __init perf_event_init_all_cpus(void)
 
                INIT_LIST_HEAD(&per_cpu(pmu_sb_events.list, cpu));
                raw_spin_lock_init(&per_cpu(pmu_sb_events.lock, cpu));
+
+               INIT_LIST_HEAD(&per_cpu(sched_cb_list, cpu));
        }
 }