Re: [Lse-tech] [RFC, PATCH] 5/5 rcu lock update: Hierarchical rcu_cpu_mask bitmap

[Manfred Spraul <manfred@colorfullife.com>]

Paul E. McKenney wrote:

>I don't understand how the following helps, given that the CPU being
>offlined is supposed to be completely dead by the time we get here.
>Can't see anything that it really hurts other than a bit of delay in
>an extremely infrequent operation, but...
>
>  
>
>>+	spin_unlock_wait(&rcu_state.mutex);
>> 	cpu_quiet(cpu, 1);
>>    
>>

I was concerned about a race of cpu_quiet for the dead cpu (last cpu in 
a group) with a concurrent cpu_quiet for another cpu in another cpu 
group. A stale bit in either rcu_state.outstanding or 
rcu_groups[].outstanding would lock up the rcu subsystem. I've thought 
about it again and I agree with you - there is no race.

I agree with most of your other points - the patch is a proof of 
concept, it definitively needs a big cleanup before is should be 
considered for merging [+ a system that needs it]

--
    Manfred