Re: [PATCH 2/3] rcu: Create rcu_sync infrastructure

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Oct 04, 2013 at 01:29:45PM +0200, Oleg Nesterov wrote:
> On 10/03, Paul E. McKenney wrote:
> >
> > On Thu, Oct 03, 2013 at 11:10:09PM +0200, Oleg Nesterov wrote:
> > >
> > > rcusync: introduce rcu_sync_struct->exclusive mode
> > >
> > > CHANGELOG.
> >
> > Should the changelog really be in all caps?  (Sorry, couldn't resist...)
> 
> Apparently you do not realize it is going to be an EXCELLENT changelog!

;-)

> > > @@ -53,9 +55,13 @@ void rcu_sync_enter(struct rcu_sync_struct *rss)
> > >  	if (need_sync) {
> > >  		rss->ops->sync();
> > >  		rss->gp_state = GP_PASSED;
> > > -		wake_up_all(&rss->gp_wait);
> > > +		if (!rss->exclusive)
> > > +			wake_up_all(&rss->gp_comp.wait);
> >
> > Not sure about the wake_up_all() on a completion,
> 
> Yes, we reuse completion->wait in the !exclusive case. Like we reuse
> its spinlock as rss_lock.
> 
> We can add a completion/complete union, but this will complicate the
> code a bit and imo doesn't make sense.

Fair enough!

							Thanx, Paul

> > but if we are exclusive,
> > don't we need to complete() the completion here?
> 
> No, if we are exclusive we should delay the "wake up the next writer"
> till rcu_sync_exit().
> 
> > Oh, I guess gp_comp.wait is exactly a wait_queue_head_t, so I guess
> > you can do wake_up_all() on it...
> 
> Yes, and we never "mix" completion/wait_queue_head_t operations/members.
> IOW, we always use ->gp_comp if "exclusive", and only ->gp_comp.wait is
> used otherwise.
> 
> > Never mind!!!
> 
> Agreed ;)
> 
> Oleg.
>