Re: [PATCH] sched: correct how RT task is picked

[Steven Rostedt <rostedt@goodmis.org>]

On Thu, May 12, 2011 at 08:06:06PM +0800, Yong Zhang wrote:
> On Wed, May 11, 2011 at 09:44:04PM +0800, Hillf Danton wrote:
> > On Wed, May 11, 2011 at 4:43 PM, Yong Zhang <yong.zhang0@gmail.com> wrote:
> > > On Tue, May 10, 2011 at 9:04 PM, Hillf Danton <dhillf@gmail.com> wrote:
> > >> When picking RT task for given CPU,
> > >> [1] if the cpu is invalid for cpumask test, right result could not be
> > >
> > > 'cpu is invalid' means weather we care it or not, it's not real 'invalid'
> > >
> > If cpu is not cared, how to determine whether it is allowed for task to run?
> 
> pick_next_highest_task_rt() can be used to get the next highest pullable
> task on a certain rq(regradless on which cpu that task could run). but
> currently we have no such kind of caller.
> 
> > 
> > >> reached even by further checking nr_cpus_allowed,
> > >> on the other hand, the input cpu is valid in two cases that
> > >> pick_next_highest_task_rt() is called, thus the invalid input cpu
> > >> looks over-concern.
> > >> [2] if the cpu is valid for cpumask test, further checking
> > >> nr_cpus_allowed looks overwork, since it is computed based on
> > >> cpus_allowed,
> > >
> > > No, cpumask_test_cpu(cpu, &p->cpus_allowed) doesn't mean
> > > p->rt.nr_cpus_allowed > 1.
> > >
> > If cpu is allowed for task to run, then why more cpus are enforced?
> 
> I think you can take a look at next_prio(), it just calculate the
> next highest task on the current cpu; in this case,
> cpumask_test_cpu(cpu, &p->cpus_allowed) will be true for the most
> of time, but maybe that task is bound to this cpu.

I've been looking at the history here, and I think that '-1' is a relic.

If you look at sched_rt.c in f65eda4f789168ba5ff3fa75546c29efeed19f58:

$ git show f65eda4f:kernel/sched_rt.c 

You'll see that push_rt_task calls pick_next_highest_task_rt() with a
-1. That code has long been replaced.

I'm a bit nervous about taking Hillf's patch as is. But a little more
reviewing and testing may prove that it is legit.

-- Steve