Re: [PATCH v5 2/4] rcu: Reduce synchronize_rcu() latency

[Uladzislau Rezki]

On Mon, Mar 04, 2024 at 11:56:19PM +0100, Frederic Weisbecker wrote:
> Le Mon, Mar 04, 2024 at 05:23:13PM +0100, Uladzislau Rezki a écrit :
> > On Mon, Mar 04, 2024 at 12:55:47PM +0100, Frederic Weisbecker wrote:
> > The easiest way is to drop the patch. To address it we can go with:
> > 
> > <snip>
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 31f3a61f9c38..9aa2cd46583e 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -1661,16 +1661,8 @@ static void rcu_sr_normal_gp_cleanup(void)
> >  	 * wait-head is released if last. The worker is not kicked.
> >  	 */
> >  	llist_for_each_safe(rcu, next, wait_tail->next) {
> > -		if (rcu_sr_is_wait_head(rcu)) {
> > -			if (!rcu->next) {
> > -				rcu_sr_put_wait_head(rcu);
> > -				wait_tail->next = NULL;
> > -			} else {
> > -				wait_tail->next = rcu;
> > -			}
> > -
> > +		if (rcu_sr_is_wait_head(rcu))
> >  			break;
> > -		}
> >  
> >  		rcu_sr_normal_complete(rcu);
> >  		// It can be last, update a next on this step.
> > <snip>
> > 
> > i.e. the process of users from GP is still there. The work is triggered
> > to perform a final complete(if there are users) + releasing wait-heads
> > so we do not race anymore.
> 
> It's worth mentioning that this doesn't avoid scheduling the workqueue.
> Except perhaps for the very first time rcu_sr_normal_gp_cleanup() is called,
> the workqueue will always have to be scheduled at least in order to release the
> wait_tail of the previous rcu_sr_normal_gp_cleanup() call.
> 
No, it does not avoid for sure :) I will add more explanation.

> But indeed you keep the optimization that performs the completions themselves
> synchronously from the GP kthread if there aren't too many of them (which
> probably is the case most of the time).
> 
> > I am OK with both cases. Dropping the patch will make it more simple
> > for sure.
> 
> I am ok with both cases as well :-)
> 
> You choose. But note that the time spent doing the completions from the GP
> kthread may come at the expense of delaying the start of the next grace period,
> on which further synchronous RCU calls may in turn depend on...
> 
That is a true point. Therefore we do it with a fixed number which should not
influence on a GP.

--
Uladzislau Rezki