[PATCH] rcu-tasks: Delay rcu_tasks_verify_self_tests() to avoid missed callbacks

[Waiman Long <longman@redhat.com>]

Even though rcu_tasks selftest is initiated early in the boot process,
the verification done at late initcall time may not be late enough to
catch all the callbacks especially on systems with just a few cpus and
small memory.

After 12 bootup's On a s390x system, 1 of them had failed rcu_tasks
verification test.

[    8.183013] call_rcu_tasks() has been failed.
[    8.183041] WARNING: CPU: 0 PID: 1 at kernel/rcu/tasks.h:1696 rcu_tasks_verify_self_tests+0x64/0xd0
[    8.203246] Callback from call_rcu_tasks() invoked.

In this particular case, the callback missed the check by about
20ms. Similar rcu_tasks selftest failures are also seen in ppc64le
systems.

[    0.313391] call_rcu_tasks() has been failed.
[    0.313407] WARNING: CPU: 0 PID: 1 at kernel/rcu/tasks.h:1696 rcu_tasks_verify_self_tests+0x5c/0xa0
[    0.335569] Callback from call_rcu_tasks() invoked.

Avoid this missed callback by delaying the verification using
delayed_work. The delay is set to be about 0.1s which hopefully will
be long enough to catch all the callbacks on systems with few cpus and
small memory.

Fixes: bfba7ed084f8 ("rcu-tasks: Add RCU-tasks self tests")
Signed-off-by: Waiman Long <longman@redhat.com>
---
 kernel/rcu/tasks.h | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
index 3925e32159b5..25f964a671ba 100644
--- a/kernel/rcu/tasks.h
+++ b/kernel/rcu/tasks.h
@@ -1735,7 +1735,7 @@ static void rcu_tasks_initiate_self_tests(void)
 #endif
 }
 
-static int rcu_tasks_verify_self_tests(void)
+static void rcu_tasks_verify_self_tests(struct work_struct *work __maybe_unused)
 {
 	int ret = 0;
 	int i;
@@ -1749,10 +1749,23 @@ static int rcu_tasks_verify_self_tests(void)
 
 	if (ret)
 		WARN_ON(1);
+}
+
+static struct delayed_work rcu_tasks_verify_work;
 
-	return ret;
+/*
+ * The rcu_tasks verification is done indirectly via the work queue to
+ * introduce an additional 0.1s delay to catch all the callbacks before
+ * the verification is done as late_initcall time may not be late enough
+ * to have all the callbacks fired.
+ */
+static int rcu_tasks_verify_schedule_work(void)
+{
+	INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_self_tests);
+	schedule_delayed_work(&rcu_tasks_verify_work, HZ/10);
+	return 0;
 }
-late_initcall(rcu_tasks_verify_self_tests);
+late_initcall(rcu_tasks_verify_schedule_work);
 #else /* #ifdef CONFIG_PROVE_RCU */
 static void rcu_tasks_initiate_self_tests(void) { }
 #endif /* #else #ifdef CONFIG_PROVE_RCU */
-- 
2.31.1