Re: Workqueues splat due to ending up on wrong CPU

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Dec 02, 2019 at 12:13:38PM -0800, Tejun Heo wrote:
> Hello, Paul.
> 
> (cc'ing scheduler folks - workqueue rescuer is very occassionally
> triggering a warning which says that it isn't on the cpu it should be
> on under rcu cpu hotplug torture test.  It's checking smp_processor_id
> is the expected one after a successful set_cpus_allowed_ptr() call.)
> 
> On Sun, Dec 01, 2019 at 05:55:48PM -0800, Paul E. McKenney wrote:
> > > And hyperthreading seems to have done the trick!  One splat thus far,
> > > shown below.  The run should complete this evening, Pacific Time.
> > 
> > That was the only one for that run, but another 24*56-hour run got three
> > more.  All of them expected to be on CPU 0 (which never goes offline, so
> > why?) and the "XXX" diagnostic never did print.
> 
> Heh, I didn't expect that, so maybe set_cpus_allowed_ptr() is
> returning 0 while not migrating the rescuer task to the target cpu for
> some reason?
> 
> The rescuer is always calling to migrate itself, so it must always be
> running.  set_cpus_allowed_ptr() migrates live ones by calling
> stop_one_cpu() which schedules a migration function which runs from a
> highpri task on the target cpu.  Please take a look at the following.
> 
>   static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
>   {
>           ...
> 	  enabled = stopper->enabled;
> 	  if (enabled)
> 		  __cpu_stop_queue_work(stopper, work, &wakeq);
> 	  else if (work->done)
> 		  cpu_stop_signal_done(work->done);
>           ...
>   }
> 
> So, if stopper->enabled is clear, it'll signal completion without
> running the work.

Is there ever a valid case for this? That is, why isn't that a WARN()?

> stopper->enabled is cleared during cpu hotunplug
> and restored from bringup_cpu() while cpu is being brought back up.
> 
>   static int bringup_wait_for_ap(unsigned int cpu)
>   {
>           ...
> 	  stop_machine_unpark(cpu);
>           ....
>   }
> 
>   static int bringup_cpu(unsigned int cpu)
>   {
> 	  ...
> 	  ret = __cpu_up(cpu, idle);
>           ...
> 	  return bringup_wait_for_ap(cpu);
>   }
> 
> __cpu_up() is what marks the cpu online and once the cpu is online,
> kthreads are free to migrate into the cpu, so it looks like there's a
> brief window where a cpu is marked online but the stopper thread is
> still disabled meaning that a kthread may schedule into the cpu but
> not out of it, which would explain the symptom that you were seeing.

Yes.

> It could be that I'm misreading the code.  What do you guys think?

The below seems to not insta explode...

---
 kernel/cpu.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/kernel/cpu.c b/kernel/cpu.c
index a59cc980adad..9eaedd002f41 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -525,8 +525,7 @@ static int bringup_wait_for_ap(unsigned int cpu)
 	if (WARN_ON_ONCE((!cpu_online(cpu))))
 		return -ECANCELED;
 
-	/* Unpark the stopper thread and the hotplug thread of the target cpu */
-	stop_machine_unpark(cpu);
+	/* Unpark the hotplug thread of the target cpu */
 	kthread_unpark(st->thread);
 
 	/*
@@ -1089,8 +1088,8 @@ void notify_cpu_starting(unsigned int cpu)
 
 /*
  * Called from the idle task. Wake up the controlling task which brings the
- * stopper and the hotplug thread of the upcoming CPU up and then delegates
- * the rest of the online bringup to the hotplug thread.
+ * hotplug thread of the upcoming CPU up and then delegates the rest of the
+ * online bringup to the hotplug thread.
  */
 void cpuhp_online_idle(enum cpuhp_state state)
 {
@@ -1100,6 +1099,12 @@ void cpuhp_online_idle(enum cpuhp_state state)
 	if (state != CPUHP_AP_ONLINE_IDLE)
 		return;
 
+	/*
+	 * Unpark the stopper thread before we start the idle thread; this
+	 * ensures the stopper is always available.
+	 */
+	stop_machine_unpark(smp_processor_id());
+
 	st->state = CPUHP_AP_ONLINE_IDLE;
 	complete_ap_thread(st, true);
 }