Re: [PATCH] NMI request/release, version 4

[Corey Minyard <cminyard@mvista.com>]

Dipankar Sarma wrote:

>On Wed, Oct 23, 2002 at 04:53:34PM -0500, Corey Minyard wrote:
>  
>
>>>After local_irq_count() went away, the idle CPU check was broken
>>>and that meant that if you had an idle CPU, it could hold up RCU
>>>grace period completion.
>>>
>>Ah, much better.  That seems to fix it.
>>    
>>
>Great! Do you have any latency numbers ? Just curious.
>
Unfortunately not.  3 seconds is well within the realm of human 
observation with printk.

>>>It might just be simpler to use completions instead -
>>>
>>>	call_rcu(&(handler->rcu), free_nmi_handler, handler);
>>>	init_completion(&handler->completion);
>>>	spin_unlock_irqrestore(&nmi_handler_lock, flags);
>>>	wait_for_completion(&handler->completion);
>>>
>>>and do
>>>
>>>	complete(&handler->completion);
>>>
>>>in the  the RCU callback.
>>>
>>>      
>>>
>>I was working under the assumption that the spinlocks were needed.  But 
>>now I see that there are spinlocks in wait_for_completion.  You did get 
>>init_completion() and call_rcu() backwards, they would need to be the 
>>opposite order, I think.
>>    
>>
>
>AFAICS, the ordering of call_rcu() and init_completion should not matter
>because the CPU that is executing them would not have gone
>through a quiescent state and thus the RCU callback cannot happen.
>Only after a context swtich in wait_for_completion(), the callback
>is possible.
>
Yes, I think you are right.  I'm still not used to the RCUs :-).

-Corey