Re: [tip:sched/core] sched/rt: Simplify the IPI based RT balancing logic

[Pavan Kondeti <pkondeti@codeaurora.org>]

On Fri, Jan 19, 2018 at 10:03:53AM -0500, Steven Rostedt wrote:
> On Fri, 19 Jan 2018 14:53:05 +0530
> Pavan Kondeti <pkondeti@codeaurora.org> wrote:
> 
> > I am seeing "spinlock already unlocked" BUG for rd->rto_lock on a 4.9
> > stable kernel based system. This issue is observed only after
> > inclusion of this patch. It appears to me that rq->rd can change
> > between spinlock is acquired and released in rto_push_irq_work_func()
> > IRQ work if hotplug is in progress. It was only reported couple of
> > times during long stress testing. The issue can be easily reproduced
> > if an artificial delay is introduced between lock and unlock of
> > rto_lock. The rq->rd is changed under rq->lock, so we can protect this
> > race with rq->lock. The below patch solved the problem. we are taking
> > rq->lock in pull_rt_task()->tell_cpu_to_push(), so I extended the same
> > here. Please let me know your thoughts on this.
> 
> As so rq->rd can change. Interesting.
> 
> > 
> > diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
> > index d863d39..478192b 100644
> > --- a/kernel/sched/rt.c
> > +++ b/kernel/sched/rt.c
> > @@ -2284,6 +2284,7 @@ void rto_push_irq_work_func(struct irq_work *work)
> >                 raw_spin_unlock(&rq->lock);
> >         }
> > 
> > +       raw_spin_lock(&rq->lock);
> 
> 
> What about just saving the rd then?
> 
> 	struct root_domain *rd;
> 
> 	rd = READ_ONCE(rq->rd);
> 
> then use that. Then we don't need to worry about it changing.
> 

I am thinking of another problem because of the race between
rto_push_irq_work_func() and rq_attach_root() where rq->rd is modified.

Lets say, we cache the rq->rd here and queued the IRQ work on a remote
CPU. In the mean time, the rq_attach_root() might drop all the references
to this cached (old) rd and wants to free it. The rq->rd is freed in
RCU-sched callback. If that remote CPU is in RCU quiescent state, the rq->rd
can get freed before the IRQ work is executed. This results in the corruption
of the remote  CPU's IRQ work list. Right?

Taking rq->lock in rto_push_irq_work_func() also does not help here. Probably
we have to wait for the IRQ work to finish before freeing the older root domain
in RCU-sched callback.

Thanks,
Pavan
-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.