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 539fa82..16396e9 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1273,8 +1273,19 @@
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 @@
&perfclass_attr_group);
}
device_initcall(perf_counter_sysfs_init);
-
