Re: mm: deadlock between get_online_cpus/pcpu_alloc

[Vlastimil Babka]

On 02/07/2017 01:37 PM, Michal Hocko wrote:
> > > @@ -6711,7 +6714,16 @@ static int page_alloc_cpu_dead(unsigned int cpu)
> > >  {
> > >
> > >  	lru_add_drain_cpu(cpu);
> > > +
> > > +	/*
> > > +	 * A per-cpu drain via a workqueue from drain_all_pages can be
> > > +	 * rescheduled onto an unrelated CPU. That allows the hotplug
> > > +	 * operation and the drain to potentially race on the same
> > > +	 * CPU. Serialise hotplug versus drain using pcpu_drain_mutex
> > > +	 */
> > > +	mutex_lock(&pcpu_drain_mutex);
> > >  	drain_pages(cpu);
> > > +	mutex_unlock(&pcpu_drain_mutex);
> >
> > You cannot put sleepable lock inside the preempt disbaled section...
> > We can make it a spinlock right?
>
> Scratch that! For some reason I thought that cpu notifiers are run in an
> atomic context. Now that I am checking the code again it turns out I was
> wrong. __cpu_notify uses __raw_notifier_call_chain so this is not an
> atomic context.

Good.

> Anyway, shouldn't be it sufficient to disable preemption
> on drain_local_pages_wq? The CPU hotplug callback will not preempt us
> and so we cannot work on the same cpus, right?

I thought the problem here was that the callback races with the work item that 
has been migrated to a different cpu. Once we are not working on the local cpu, 
disabling preempt/irq's won't help?