Re: [PATCH] kthread: Prevent unpark race which puts threads on the wrong cpu

[Thomas Gleixner <tglx@linutronix.de>]

On Thu, 11 Apr 2013, Srivatsa S. Bhat wrote:
> On 04/11/2013 04:18 PM, Srivatsa S. Bhat wrote:
> > On 04/11/2013 03:49 PM, Thomas Gleixner wrote:
> >> Dave,
> >>
> >> On Wed, 10 Apr 2013, Dave Hansen wrote:
> >>
> >>> I think I got a full trace this time:
> >>>
> >>> 	http://sr71.net/~dave/linux/bigbox-trace.1365621899.txt.gz
> >>>
> >>> The last timestamp is pretty close to the timestamp on the console:
> >>>
> >>> [ 2071.033434] smpboot_thread_fn():
> >>> [ 2071.033455] smpboot_thread_fn() cpu: 22 159
> >>> [ 2071.033470] td->cpu: 22
> >>> [ 2071.033475] smp_processor_id(): 21
> >>> [ 2071.033486] comm: migration/%u
> >>
> >> Yes, that's helpful. Though it makes my mind boggle:
> >>
> >>     migration/22-4335  [021] d.h.  2071.020530: sched_wakeup: comm=migration/21 pid=4323 prio=0 success=1 target_cpu=021^M
> >>     migration/22-4335  [021] d...  2071.020541: sched_switch: prev_comm=migration/22 prev_pid=4335 prev_prio=0 prev_state=0x200 ==> next_comm=migration/21 next_pid=4323 next_prio=0^M
> >>     migration/21-4323  [021] d...  2071.026110: sched_switch: prev_comm=migration/21 prev_pid=4323 prev_prio=0 prev_state=S ==> next_comm=migration/22 next_pid=4335 next_prio=0^M
> >>     migration/22-4335  [021] ....  2071.033422: smpboot_thread_fn <-kthread^M
> >>
> >> So the migration thread schedules out with state TASK_PARKED and gets
> >> scheduled back in right away without a wakeup. Srivatsa was about
> >> right, that this might be related to the sched_set_stop_task() call,
> >> but the changelog led completely down the wrong road.
> >>
> >> So the issue is:
> >>
> >> CPU14				CPU21
> >> create_thread(for CPU22)
> >>  park_thread()	
> >>    wait_for_completion()	park_me()
> >> 				 complete()
> >>    sched_set_stop_task()
> >> 				 schedule(TASK_PARKED)
> >>
> >> The sched_set_stop_task() call is issued while the task is on the
> >> runqueue of CPU21 and that confuses the hell out of the stop_task
> >> class on that cpu. So as we have to synchronize the state for the
> >> bind call (the issue observed by Borislav) we need to do the same
> >> before issuing sched_set_stop_task(). Delta patch below.
> >>
> > 
> > In that case, why not just apply this 2 line patch on mainline?
> > The patch I sent yesterday was more elaborate because I wrongly assumed
> > that kthread_bind() can cause a wakeup. But now, I feel the patch shown
> > below should work just fine too. Yeah, it binds the task during creation
> > as well as during unpark, but that should be ok (see below).
> > 
> > Somehow, I believe we can handle this issue without introducing that
> > whole TASK_PARKED thing..
> > 
> > 
> > diff --git a/kernel/kthread.c b/kernel/kthread.c
> > index 691dc2e..74af4a8 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;
> >  }
> >  
> 
> Oh wait a minute, its simpler than even that I believe! We don't need that
> extra kthread_bind() also.
> 
> So, here is what is happening, from what I understand:
> sched_set_stop_task() will make the newly created migration thread as the
> stop task, during ht->create(). But looking into __sched_setscheduler(),
> we observe that if p->on_rq is true (where p is the migration thread),
> the task is enqueued on the rq by calling ->enqueue_task() of stop_task
> sched class, which simply increments the rq->nr_running.
> 
> So, on the next call to schedule(), the pick_next_task() in core.c sees
> that rq->nr_running is not equal to rq->cfs.h_nr_running and hence invokes
> the stop_task sched class' ->pick_next_task(). And that will return the
> stop task, which is nothing but this migration thread that we just created.
> That's why the migration thread starts running prematurely, even before
> it is bound, and even before its corresponding CPU is brought online.
> 
> So, if we just wait till p->on_rq is reset to 0 by __schedule(), before
> calling ht->create(), the task doesn't get queued to the runqueue with
> stop task priority. So that's it - we simply wait till it gets off the
> rq and everything should be fine. And during the subsequent online, during
> the unpark phase, the task gets bound to the appropriate cpu and only
> then woken up, as desired.

Right, but I want to avoid that wait for the common case. In most of
the cases the newly created thread reaches PARKED and scheduled out
before we do bind or any other stuff.
 
Thanks,

	tglx