Re: [PATCH] CPU hotplug, smpboot: Fix crash in smpboot_thread_fn()

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 04/11/2013 01:40 PM, Thomas Gleixner wrote:
> On Wed, 10 Apr 2013, Srivatsa S. Bhat wrote:
> 
>> Interestingly, in every single stack trace, the crashing task is the migration
>> thread. Now, migration thread belongs to the highest priority stop_task sched
>> class, and this particular sched class is very unique in the way it implements
>> its internal sched class functions, and I suspect this has a lot of bearing
>> on how functions like kthread_bind(), wake_up_process() etc react with it
>> (by looking at how it implements its functions such as select_task_rq(),
>> enqueue_task(), dequeue_task() etc).
> 
> I don't think that's relevant. The migration thread can only be woken
> via try_to_wakeup and my previous patch which implements a separate
> task state makes sure that it cannot be woken accidentaly by anything
> else than unpark.
> 

Hmm, but it got to be simpler than that, no? Given that it used to work fine
before...

>> But note that __kthread_bind() can wake up the task if the task is an RT
>> task. So it can be called only when the CPU (to which we want to bind the task)
> 
> kthread_bind() does NOT wakeup anything. It merily sets the cpus
> allowed ptr without further ado.
> 

Sorry, I was mistaken and was carried away by a bug in the code I was testing.
I had intended to move kthread_bind() to the body of kthread_create_on_cpu()
and place it after the call to kthread_park(), as shown below:

diff --git a/kernel/kthread.c b/kernel/kthread.c
index 691dc2e..b485fc0 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -308,6 +308,9 @@ struct task_struct *kthread_create_on_cpu(int (*threadfn)(void *data),
 	to_kthread(p)->cpu = cpu;
 	/* Park the thread to get it out of TASK_UNINTERRUPTIBLE state */
 	kthread_park(p);
+
+	wait_task_inactive(p, TASK_INTERRUPTIBLE);
+	__kthread_bind(p, cpu);
 	return p;
 }
 
But by mistake, I had written the code as:

diff --git a/kernel/kthread.c b/kernel/kthread.c
index 691dc2e..b485fc0 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -308,6 +308,9 @@ struct task_struct *kthread_create_on_cpu(int (*threadfn)(void *data),
 	to_kthread(p)->cpu = cpu;
 	/* Park the thread to get it out of TASK_UNINTERRUPTIBLE state */
 	kthread_park(p);
+
+	if (!wait_task_inactive(p, TASK_INTERRUPTIBLE))
+		__kthread_bind(p, cpu);
 	return p;
 }
 
So, no wonder it never actually bound the task to the CPU. So when I gave this a run,
I saw watchdog threads hitting the same BUG_ON(), and since watchdog threads are
of RT priority, and RT is the only class that implements ->set_cpus_allowed(), I
thought that those threads got woken up due to the bind. But I was mistaken of
course, because I had checked for the wrong return value of wait_task_inactive().

Regards,
Srivatsa S. Bhat