Re: On migrate_disable() and latencies

On Thu, 28 Jul 2011, Yong Zhang wrote:

On Wed, Jul 27, 2011 at 11:30:08AM -0700, Paul E. McKenney wrote:

quoted

o	Tasks awakening outside of migrate-disable regions will pick
	the CPU running the lowest-priority task, whether or not this
	task is in migrate-disable state.  (At least I don't see
	anything in 3.0-rt3 that looks like a scheduling decision
	based on ->migrate_disable, perhaps due to blindness.)

I'm also confused here, seems we just disable migration for RT task.
migrate_disable()
{
	...
		if (p->sched_class->set_cpus_allowed)
			p->sched_class->set_cpus_allowed(p, mask);
		p->rt.nr_cpus_allowed =	cpumask_weight(mask);
	...
}

Shouldn't we also forbid migration on !RT task?

We do. Just RT is the only sched class which has a set_cpus_allowed()
callback implemented and want's an update to its rt.nr_cpus_allowed
field.

       if (!p->migrate_disable) {
                if (p->sched_class && p->sched_class->set_cpus_allowed)
                        p->sched_class->set_cpus_allowed(p, new_mask);
                p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
        }

The general part is here:

       cpumask_copy(&p->cpus_allowed, new_mask);

And tsk_cpus_allowed() does:
{
        if (p->migrate_disable)
                return cpumask_of(task_cpu(p));

        return &p->cpus_allowed;
}

which is the relevant information to keep any task independent of it's
sched class pinned.

Thanks,

	tglx