Re: [PATCH RFC] rcu: Limit GP initialization to CPUs that have been online

[Mike Galbraith <efault@gmx.de>]

On Fri, 2012-04-13 at 11:42 -0700, Paul E. McKenney wrote: 
> On Wed, Apr 11, 2012 at 01:04:36PM +0200, Mike Galbraith wrote:

> > Hi Paul (and Dimitri),
> > 
> > Just got back to this.  I changed the patch around to check for a
> > hotplug event in force_qs_rnp(), and should that happen, process any
> > freshly added CPUs immediately rather than tell the caller to restart
> > from scratch.  The rest of the delta is cosmetic, and there should be
> > zero performance change.
> > 
> > Does this change address any of your concerns? 
> 
> Apologies for being slow to respond...

Hey, I'm grateful for any time you have to share.

> One of my main concerns was present in my original patch:  I now believe
> that a given grace period needs to operate on the set of rcu_node
> structures (and CPUs) that were present at the beginning of the grace
> period.  Otherwise, things could get confused if a given CPU participated
> in an later force_quiescent_state() state, but not in an earlier one.
> 
> I believe that the correct way to handle this is to squirrel the maximum
> number of CPUs away in the rcu_state structure at the beginning of each
> grace period and use that as the limit.
> 
> I call out a few more below.

Read it, got it.  The last node being partially initialized still causes
some head scratching, I'll ponder that more between explosions.  Thanks.

-Mike