Re: [patch 2/2] cpu hotplug notifier for updating sched domains

[Nick Piggin <piggin@cyberone.com.au>]

Nathan Lynch wrote:

>On Tue, 2004-09-07 at 19:44, Nick Piggin wrote:
>
>>I think the next step is to now make the setup code only use cpu_online_map
>>

^^^ OK I see you've already done that. Sorry I should have looked at the
patches a bit closer. Your implementation looks very nice.

>>and get rid of everywhere I had been doing cpus_and(tmp, ..., 
>>cpu_online_map).
>>

^^^ This should still be done, of course. That can come later.

>>This may also make your patch 1/2 unnecessary? What do you think?
>>
>
>Well, we have to "lie" to arch_init_sched_domains a little bit when
>bringing a cpu online, by setting the soon-to-be-online cpu's bit in the
>argument mask.  So I think the first patch is still necessary.
>
>

Can't we do everything in the CPU_UP_ONLINE case though?


One other thing:

void __init sched_init_smp(void)
 {
+	lock_cpu_hotplug();
 	arch_init_sched_domains(cpu_online_map);
 	sched_domain_debug();
+	unlock_cpu_hotplug();
+
+	hotcpu_notifier(update_sched_domains, 0);
 }

Do you have a theoretical race here? Can we hotplug a CPU before the notifier
is registered? (I know we *can't* because it is still earlyish boot).

Can you move hotcpu_notifier under the cpu_hotplug lock? Seems not because
register_cpu_notifier takes the lock itself. Seems like a flaw in the API to
me. Probably the notifier chain should be protected by a lock that nests
inside the cpucontrol lock. Rusty?