* Paul Mackerras <paulus@samba.org> wrote:

quoted text
> Ingo Molnar writes:
> 
> > Yeah, indeed that box has a CPU hotplug testcase - sets cpu1 to 
> > offline then online.
> > 
> > There should be no counters active anywhere during that.
> 
> OK, I can't reproduce this on powerpc.  I guess you have dynamic 
> per-cpu patches in there, and per-cpu areas are getting 
> reinitialized when cpus come up.  That, combined with the fact 
> that the migration_notifier in kernel/sched.c puts itself at 
> priority 10, means that we're getting a call to 
> perf_counter_task_migration() for a newly-added CPU before 
> perf_cpu_notify() has been called for that CPU, and so we're 
> trying to use an uninitialized perf_cpu_context and we go boom.

Sounds very plausible.

quoted text
> Could you try the same test with this patch?  If this fixes it, 
> then that's what the problem is.  It's up to you whether 
> increasing the priority on perf_cpu_nb is the right solution or 
> whether we should solve the problem some other way.
> 
> Paul.
> 
> diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
> --- a/kernel/perf_counter.c
> +++ b/kernel/perf_counter.c
> @@ -3902,8 +3902,12 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
>  	return NOTIFY_OK;
>  }
>  
> +/*
> + * This has to have a higher priority than migration_notifier in sched.c.
> + */
>  static struct notifier_block __cpuinitdata perf_cpu_nb = {
>  	.notifier_call		= perf_cpu_notify,
> +	.priority		= 20,
>  };

Makes sense. Mind doing a full patch with a changelog, and with a 
comment that explains what the priority rules are? Perhaps add a 
comment to the counterpart in sched.c too.

	Ingo
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/