Skip to contentSkip to conversation

perf: Cure hotplug lock ordering issues

tip:smp/hotplug Peter Zijlstra 1 participant 1 message Apr 24, 2017 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 27 lines · tip-bot for Thomas Gleixner
Commit-ID:  14aff62ede39a7b069beb7b95c1d02edb8ee912d
Gitweb:     http://git.kernel.org/tip/14aff62ede39a7b069beb7b95c1d02edb8ee912d
Author:     Thomas Gleixner <[email protected]>
AuthorDate: Sun, 23 Apr 2017 12:13:09 +0200
Committer:  Thomas Gleixner <[email protected]>
CommitDate: Sun, 23 Apr 2017 19:11:27 +0200

perf: Cure hotplug lock ordering issues

The get_online_cpus() rework unearthed lock ordering issues.

Reorder hotpluglock and perf_sched_mutex() to avoid circular locking dependencies and convert static_branch_disable() to the cpuslocked version to avoid hotplug lock recursion.

---
 kernel/events/core.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
kernel/events/core.c +0 -0
index 8aa3063..b5b4f52f 100644
@@ -4016,10 +4016,12 @@ static void unaccount_event(struct perf_event *event)
 

kernel/events/core.c static void unaccount_event(struct perf_event *event) · line 4025

Quoted 10 lines
kernel/events/core.c +3 -1
 static void perf_sched_delayed(struct work_struct *work)
 {
+	get_online_cpus();
 	mutex_lock(&perf_sched_mutex);
 	if (atomic_dec_and_test(&perf_sched_count))
-		static_branch_disable(&perf_sched_events);
+		static_branch_disable_cpuslocked(&perf_sched_events);
 	mutex_unlock(&perf_sched_mutex);
+	put_online_cpus();
 }

So I think we can simply remove the perf_sched_mutex here. Its only purpose is to serialize the 0->1 state, such that nobody passes before we've done the additional sync_sched() call.

Once we're >=1, it really doesn't matter anymore. Not sure why I put it in here. Maybe extensive paranoia..