Re: hotplug thread issues

[Thomas Gleixner <tglx@linutronix.de>]

On Thu, 6 Nov 2014, Peter Zijlstra wrote:

> Hi Thomas,
> 
> So there have been some reports on hitting:
> 
>   BUG_ON(td->cpu != smp_processor_id());
> 
> in smpboot_thread_fn.
> 
> Now I've been staring at this for a wee bit today and I've found two
> issues, but I'm not sure either are enough to explain the observed.
> 
> 1) smpboot_register_percpu_thread() seems to lack serialization against
>    hotplug. It has a for_each_online() loop, but no get_online_cpus() --
>    unlike smpboot_unregister_percpu_thread, which does.
> 
>    Typical usage like spawn_ksoftirqd() should be fine, they're early
>    init calls and those run before we bring up the other CPUs. Therefore
>    this does not explain the observation that its ksoftirqd/n triggering
>    the BUG.
> 
>    However, the usage in proc_dowatchdog() is susceptible to this race
>    and its entirely possible to go wrong there.

Hmm. Need to have a look.
 
> 
> 2) the usage of __set_current_state(TASK_PARKED) in __kthread_parkme()
>    is wrong AFAICT, one should always use set_current_state() for
>    setting !TASK_RUNNING state. The comment with set_current_state()
>    explains why.
> 
>    This would've allowed the test_bit(KTHREAD_SHOULD_PARK) load to have
>    been satisfied before the store of TASK_PARKED.

My bad. Can you send a proper patch addressing that issue please? That
should be tagged stable as well I guess.

Thanks,

	tglx