Re: Deadlock between cpu_hotplug_begin and cpu_add_remove_lock

[Oleg Nesterov <oleg@redhat.com>]

On 01/23, Srivatsa S. Bhat wrote:
>
> On 01/23/2014 12:48 AM, Oleg Nesterov wrote:
> > On 01/22, Srivatsa S. Bhat wrote:
> >>
> >> Wait a min, that _will_ actually work for all cases because I have provided
> >> an option to invoke _any_ arbitrary function as the "setup" routine.
> >
> > And probably the generic solution makes sense. I am not sure I actually
> > understand the semantics of register_allcpu_notifier(), but the problem
> > it tries to solve looks clear/valid.
> >
>
> Thank you. But I was wondering whether its usage is a bit unintuitive/
> convoluted. So I was contemplating between going with that solution or the
> below one, where the call-sites are expected to do:
>
> cpu_maps_update_begin();
> for_each_online_cpu(cpu) {
> 	...
> }
> __register_cpu_notifier(); //use the __reg() variant, which doesn't take locks
> cpu_maps_update_done();
>
> Of course, that requires exporting the functions cpu_maps_update_begin/done(),
> but this latter form of callback registration might look more natural.

Yes, I thought about this too ;)

> But for some of the other call-sites, we might have to use one
> of the solutions mentioned above.

Yes, yes, sure, I agree.

I suggested this change only for discussion, for the case we need
an "urgent" fix without changes outside of drivers/md/. The generic
solution is better.

Oleg.