perfcounters: release CPU context when exiting task counters

If counters are exiting via do_exit() not via filp close, then
the CPU context needs to be released - otherwise future percpu
counter creations might fail.

Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 539fa8283a06f34e81ead1b6618e3d9f07cd9d38..16396e9406fa8b6eb445fbb3cd82134e5c428511 100644 (file)
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1273,8 +1273,19 @@ __perf_counter_exit_task(struct task_struct *child,
        local_irq_disable();
        perf_flags = hw_perf_save_disable();
 
-       if (child_counter->state == PERF_COUNTER_STATE_ACTIVE)
+       if (child_counter->state == PERF_COUNTER_STATE_ACTIVE) {
+               struct perf_cpu_context *cpuctx;
+
+               cpuctx = &__get_cpu_var(perf_cpu_context);
+
                child_counter->hw_ops->hw_perf_counter_disable(child_counter);
+               child_counter->state = PERF_COUNTER_STATE_INACTIVE;
+               child_counter->oncpu = -1;
+
+               cpuctx->active_oncpu--;
+               child_ctx->nr_active--;
+       }
+
        list_del_init(&child_counter->list_entry);
 
        hw_perf_restore(perf_flags);
@@ -1539,4 +1550,3 @@ static int __init perf_counter_sysfs_init(void)
                                  &perfclass_attr_group);
 }
 device_initcall(perf_counter_sysfs_init);
-