Re: [PATCH 2/2] mm: drop hotplug lock from lru_add_drain_all

[Michal Hocko]

On Thu 02-11-17 14:02:53, Thomas Gleixner wrote:
> On Thu, 2 Nov 2017, Michal Hocko wrote:
> > On Thu 02-11-17 10:36:13, Michal Hocko wrote:
> > [...]
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 67330a438525..8c6e9c6d194c 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -6830,8 +6830,12 @@ void __init free_area_init(unsigned long *zones_size)
> > >  
> > >  static int page_alloc_cpu_dead(unsigned int cpu)
> > >  {
> > > +	unsigned long flags;
> > >  
> > > +	local_irq_save(flags);
> > >  	lru_add_drain_cpu(cpu);
> > > +	local_irq_restore(flags);
> > > +
> > >  	drain_pages(cpu);
> >   
> > I was staring into the hotplug code and tried to understand the context
> > this callback runs in and AFAIU IRQ disabling is not needed at all
> > because cpuhp_thread_fun runs with IRQ disabled when offlining an online
> > cpu. I have a bit hard time to follow the code due to all the
> > indirection so please correct me if I am wrong.
> 
> No. That function does neither run from the cpu hotplug thread of the
> outgoing CPU nor its called with interrupts disabled.
> 
> The callback is in the DEAD section, i.e. its called on the controlling CPU
> _after_ the hotplugged CPU vanished completely.

OK, so IIUC there is no race possible because kworkders simply do not
run on that cpu anymore.
-- 
Michal Hocko
SUSE Labs