On Sat, Sep 06, 2008 at 06:37:38PM +0200, Manfred Spraul wrote:
> Hi Paul,
>
> Paul E. McKenney wrote:
>> o	The rcu_pending() and rcu_needs_cpu() primitives are now
>> 	much more aggressive about permitting CPUs to enter dynticks
>> 	idle mode.  Only CPUs that have RCU callbacks are kept out
>> 	of dynticks idle mode.
>> 
>
> I've noticed that right now rcu_enter_nohz() can be nested within 
> rcu_irq_enter():
> irq_exit() first calls tick_nohz_stop_sched_tick(), then rcu_irq_exit().
> And tick_nohz_stop_sched_tick() can switch into nohz mode.
>
> Is that intentional? Does rcupreempt support that? It broke my rcustate 
> code on x86-64.
>
> I would prefer if something like the attached patch is applied. What do you 
> think?
> Do you need the patch as well?

Good question -- when I tried splitting irqs from NMIs, things broke
badly, and this might well explain it.  Thank you very much for the
hint!!!

							Thanx, Paul

> --
>    Manfred

> diff --git a/kernel/softirq.c b/kernel/softirq.c
> index ba20a90..cca5a83 100644
> --- a/kernel/softirq.c
> +++ b/kernel/softirq.c
> @@ -284,10 +284,10 @@ void irq_exit(void)
>  		invoke_softirq();
> 
>  #ifdef CONFIG_NO_HZ
> +	rcu_irq_exit(0);
>  	/* Make sure that timer wheel updates are propagated */
>  	if (!in_interrupt() && idle_cpu(smp_processor_id()) && !need_resched())
>  		tick_nohz_stop_sched_tick(0);
> -	rcu_irq_exit(0);
>  #endif
>  	preempt_enable_no_resched();
>  }