Re: [PATCH RFC tip/core/rcu] v2 accelerate grace period if last non-dynticked CPU

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sun, Jan 31, 2010 at 08:24:04PM -0800, Paul E. McKenney wrote:
> There is nothing illegal about the following:
> 
> 	static void my_rcu_callback(struct rcu_head *rcu)
> 	{
> 		struct foo *fp = container_of(rcu, struct foo, rcu_head);
> 
> 		if (fp->refcnt != 0) {
> 			call_rcu(rcu);
> 			return;
> 		}
> 		kfree(fp);
> 	}
> 
> And allowing RCU_NEEDS_CPU_FLUSHES of infinity would work correctly in
> some sense, but would be a massive power inefficiency.
> 
> The choice of "5" allows a callback that posts one other callback,
> which happens often enough to be worth the extra iterations.  It is
> necessary to budget two passes through the loop per level of RCU
> callback, one for the current CPU to start the grace period and another
> for it to end it.
> 
> Seem reasonable?


Yeah ok, I did not considered situations like the above example.

Thanks.