Re: [PATCH] simplify/improve rcu batch tuning

[Oleg Nesterov <oleg@tv-sign.ru>]

On 09/08, Paul E. McKenney wrote:
>
> On Fri, Sep 08, 2006 at 03:39:30PM +0400, Oleg Nesterov wrote:
> > >
> > Actually I think it also makes sense to do tasklet_schedule(rcu_tasklet)
> > in call_rcu(), this way we can detect that we need to start the next batch
> > earlier.
> 
> As long as we don't do this too often...  One way to prevent doing this
> too often would be to check rcp->completed against rdp->batch similarly
> to __rcu_process_callbacks()'s checks.  In call_rcu(), perhaps something
> like the following inside the ->qlen check:
> 
> 	if (__rcu_pending(&rcu_ctrlblk, rdp) {
> 		tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
> 	}

Yes.

> > > > @@ -297,6 +294,7 @@ static void rcu_start_batch(struct rcu_c
> > > >  		smp_mb();
> > > >  		cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
> > > >
> > > > +		rcp->signaled = 0;
> > >
> > > Would it make sense to invoke force_quiescent_state() here in the
> > > case that rdp->qlen is still large?  The disadvantage is that qlen
> > > still counts the number of callbacks that are already slated for
> > > invocation.
> > 
> > This is not easy to do. rcu_start_batch() is "global", we need
> > to scan all per-cpu 'struct rcu_data' and check it's ->qlen.
> 
> My thought was that it might make sense to check only this CPU's struct
> rcu_data.  But I agree that the next approach seems more promising.

Yes, I understood this. And I don't say this is "bad", I just think this is
"not perfect". Because the CPU which actually starts a new grace period and
clears ->signaled may have ->qlen == 0, and the caller is cpu_quiet(), which
is a "response" to some other CPU's force_quiescent_state().

Oleg.