Re: [PATCH 3/4 V2] implement per-domain single-thread state machine call_srcu()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2012-04-10 at 16:18 -0700, Paul E. McKenney wrote:
> > +static void srcu_invoke_callbacks(struct srcu_struct *sp)
> > +{
> > +     int i;
> > +     struct rcu_head *head;
> > +
> > +     for (i = 0; i < SRCU_CALLBACK_BATCH; i++) {
> 
> If there really can be thousands of callbacks dumped into SRCU, a more
> adaptive strategy might be needed.  In the meantime, I am hoping that
> the fact that the workqueue is retriggered in this case suffices.
> 
> Note that this function is preemptible, so there is less penalty for
> running a very long batch.

With just the ->func() invocation below non-preemptible, I really don't
see a point in having this loop limit.

> Which reminds me...  An srcu_struct structure with a large pile of
> SRCU callbacks won't react very quickly in response to an invocation of
> synchronize_srcu_expedited().  This is why the other RCU implementations
> have a non-callback codepath for expedited grace periods.
> 
> Or am I missing something here?

I would suggest adding that extra complexity when we need it ;-)

> > +             head = rcu_batch_dequeue(&sp->batch_done);
> > +             if (!head)
> > +                     break;
> > +             head->func(head);
> 
> I have surrounded this with local_bh_disable() and local_bh_enable()
> in order to enforce the no-sleeping-in-callbacks rule.  Please let me
> know if I missed some other enforcement mechanism.

Is that -rt inspired hackery? Otherwise I would simply suggest
preempt_disable/enable(), they do pretty much the same and are less
confusing.

> > +     }