Re: CPU Hotplug rework

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 04/11/2012 06:30 AM, Paul E. McKenney wrote:

> On Tue, Apr 10, 2012 at 08:37:18PM -0400, Steven Rostedt wrote:
>> On Tue, 2012-04-10 at 17:28 -0700, Paul E. McKenney wrote:
>>
>>>> Just to throw out the stupid silly approach.
>>>>
>>>> What about creating a "__register_cpu_notifier()" that just does:
>>>>
>>>> int __ref __register_cpu_notifier(struct notifier_block *nb)
>>>> {
>>>> 	return raw_notifier_chain_register(&cpu_chain, nb);
>>>> }
>>>>
>>>> Also making cpu_maps_update_begin/done() global (and probably rename
>>>> them).
>>
>> I just noticed that the cpu_maps_update_begin/done() are already global.
>>
>>>>
>>>> and then in the above code do:
>>>>
>>>> 	cpu_maps_update_begin();
>>>> 	__register_cpu_notifier(nb);
>>>> 	do_setup();
>>>> 	cpu_maps_update_done();
>>>>
>>>>


Wow! Believe it or not, this is precisely the crux of the approach I was
suggesting all along!! :-) Just that when put to code, it looked slightly
different than this.. Sorry for not being clear.

So here is what I proposed, in a simplified form:

Modify the existing register_cpu_notifier() to this (by possibly giving
it a different name):

int __ref register_cpu_notifier(struct notifier_block *nb,
					int (*do_setup)(void))
{
	int ret;

	cpu_maps_update_begin();
	ret = raw_notifier_chain_register(&cpu_chain, nb);
	do_setup();
	cpu_maps_update_done();

	return ret;
}

and then, in the caller, do:

	register_cpu_notifier(nb, do_setup);

If the caller doesn't need any such extra setup, just do:

	register_cpu_notifier(nb, NULL);


Of course, register_cpu_notifier() should handle NULL properly.
(My patch [1] handles it, along with some other special cases.)
 
That's it!

Also, it is to be noted that cpu_maps_update_begin/done() are global, but
not exported symbols - so modules can't use them. With the above approach,
we need not make them exported symbols, since the caller need not care about
these locks at all.

>>>> Just saying,
>>>
>>> That does have some attractive properties, now that you mention it.  ;-)
>>
>> Which property? Stupid or Silly ;-)
> 
> As with any piece of software, no matter how small, both.  ;-)
> 
> Of course, __register_cpu_notifier() would need lockdep checking to make
> sure that it wasn't called without the benefit of cpu_maps_update_begin().


Not with my approach ;-) Its all automatically handled :-)

> I might be missing something, but as long as that was in place, seems
> like it is a lot simpler and easier to use than the alternatives that
> Srivatsa and I were kicking around.
> 


Hehe :-) Thanks for simplifying things, Steve!


[1]. https://lkml.org/lkml/2012/3/1/39

Regards,
Srivatsa S. Bhat