Re: [RFC PATCH] Dynamic sched domains (v0.5)

[Nick Piggin]

Dinakar Guniguntala wrote:
> On Mon, May 02, 2005 at 07:44:05PM +1000, Nick Piggin wrote:
>
> > Dinakar Guniguntala wrote:
> >
> >
> > > +void rebuild_sched_domains(cpumask_t span1, cpumask_t span2)
> > > +{
> > > +	cpumask_t change_map;
> > > +
> > > +	cpus_or(change_map, span1, span2);
> > > +
> > > +	preempt_disable();
> >
> > Oh, you can't do this here, attach_domains does a synchronize_kernel.
> > So take it out, it doesn't do anything anyway, does it?
>
>
> I put that in to prevent hangs with CONFIG_PREEMPT turned on, but
> clearly didn't test it with preempt turned on. Looks like all I need to 
> do here is a local_irq_disable

What are you protecting against, though? synchroinze_kernel can
sleep, so local_irq_disable is probably the wrong thing to do as well.

AFAIKS, you don't need anything here - so long as you have mutual
exclusion from other sched-domain building then this can take as long
as it wants / be preempted as many times as we like.

> > I suggest you also use some sort of locking to prevent concurrent rebuilds
> > and rebuilds racing with cpu hotplug. You could probably have a static
> > semaphore around rebuild_sched_domains, and take lock_cpu_hotplug here too.
>
>
> I already do a lock_cpu_hotplug() in cpuset.c before calling 
> rebuild_sched_domains and also am holding cpuset_sem, so that should take
> care of both hotplug and concurrent rebuilds

OK.

But if we want this to be a respectable interface (possibly for more than
just cpusets) then it should probably do some locking itself. It isn't
performance critical, so I think taking a semaphore wouldn't hurt.

--
SUSE Labs, Novell Inc.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/