Re: Q: wait_task_inactive() and !CONFIG_SMP && CONFIG_PREEMPT

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

On 07/28, Roland McGrath wrote:
>
> I can't speak to the kthread case.  I suspect that set_task_cpu() is always
> safe on !SMP PREEMPT, and that's why it's fine.

Yes, kthread_bind() is fine, it changes nothing in *k if !SMP.

> > I refer to this patch of the comment:
> > 
> > 	If a second call a short while later returns the same number, the
> > 	caller can be sure that @p has remained unscheduled the whole time.
> > 
> > The dummy version always returns the same number == 1.
> 
> Right.  For the general case where this is the contract wait_task_inactive
> is expected to meet, it does matter.  I think task_current_syscall() does
> want this checked for the preempted uniprocessor case, for example.
> 
> > So. I think that wait_task_inactive() needs "defined(SMP) || defined(PREEMPT)"
> > and the dummy version should return ->nvcsw too.
> 
> Is this what we want?
> 
> 	#ifdef CONFIG_SMP
> 	extern unsigned long wait_task_inactive(struct task_struct *, long);
> 	#else
> 	static inline unsigned long wait_task_inactive(struct task_struct *p,
> 						       long match_state)
> 	{
> 		unsigned long ret = 0;
> 		if (match_state) {
> 			preempt_disable();
> 			if (p->state == match_state)
> 				ret = (p->nvcsw << 1) | 1;
> 			preempt_enable();
> 		}
> 		return ret;
> 	}
> 	#endif

I dont think this is right.

Firstly, the above always fails if match_state == 0, this is not right.

But more importantly, we can't just check ->state == match_state. And
preempt_disable() buys nothing.

Let's look at task_current_syscall(). The "target" can set, say,
TASK_UNINTERRUPTIBLE many times, do a lot of syscalls, and not once
call schedule().

And the task remains fully preemptible even if it runs in
TASK_UNINTERRUPTIBLE state.


Let's suppose we implement kthread_set_nice() in the same manner
as kthread_bind(),

	kthread_set_nice(struct task_struct *k, long nice)
	{
		wait_task_inactive(k, TASK_UNINTERRUPTIBLE);

		... just change ->prio/static_prio ...
	}

the above is ugly of course, but should be correct correct even
with !SMP && PREEMPT.


I think we need

	#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
	extern unsigned long wait_task_inactive(struct task_struct *, long);
	#else
	static inline unsigned long wait_task_inactive(struct task_struct *p,
							long match_state)
	{
		if (match_state && p->state != match_state)
			return 0;
		return p->nvcsw | (LONG_MAX + 1); // the same in sched.c
	}

Oleg.