Re: [PATCH/RFC] sched: Remove SYSTEM_RUNNING checks from cond_resched*()

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 2009-07-08 at 02:50 +0200, Oleg Nesterov wrote:

> 	/*
> 	 * It is valid to assume CPU-locality during early bootup:
> 	 */
> 	if (system_state != SYSTEM_RUNNING)
> 		goto out;
> 
> this doesn't look right, smp_init() is called before we set
> SYSTEM_RUNNING.

The thing is, there's also ton's of code that might end up calling
cond_resched() and co before the scheduler is fully initialized. Doing
so would indeed mess things up.

Also, by definition we'd have to call smp_init() before SYSTEM_RUNNING,
because you simply cannot declare a system up and running when your core
functionality isn't initialized.

So I'd really rather preserve these checks -- I can even remember
running into some of these things a while back, but memory isn't
providing specific cases.

> Hmm, and
> 
> 	/*
> 	 * Kernel threads bound to a single CPU can safely use
> 	 * smp_processor_id():
> 	 */
> 	if (cpumask_equal(&current->cpus_allowed, cpumask_of(this_cpu)))
> 		goto out;
> 
> perhaps this should use PF_THREAD_BOUND ?

That might predate PF_THREAD_BOUND, also I think this is more generic,
and I think we used it for that set_affinity dance we did Rusty 'fixed'
a while back.