Re: [PATCH 1/2] PF_DEAD: cleanup usage

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

Linus Torvalds wrote:
> 
> HOWEVER. I just noticed something strange. EXIT_DEAD should be in
> "task->exit_state", not in "task->state". So there's something strange
> going on in that neck of the woods _anyway_. That whole
>
> 	...
>         if (unlikely(prev->flags & PF_DEAD))
>                 prev->state = EXIT_DEAD;
> 	...
>
> in kernel/sched.c seems totally bogus.

We can use any random value instead of EXIT_DEAD (except RUNNING,INTERRUPTIBLE,
and UNINTERRUPTIBLE). The only reason for changing the state's value is this
check below:

	if (->state && !PREEMPT_ACTIVE) {
		...
		deactivate_task();
	}

This state is invisible to proc/array.c because yes, we already have something
in ->exit_state.

We can add new TASK_DEAD or something for that. Or we can do:

-	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
+	// PF_DEAD means that preemption was disabled
+	if (PF_DEAD || (prev->state && !(preempt_count() & PREEMPT_ACTIVE))) {

but this is so ugly.

> I think we should remove that thing entirely, since it's actually a total
> no-op, apart from doing something that is actively wrong. The "exit_state"
> flag should already _be_ EXIT_DEAD, no?

No, it could be EXIT_ZOMBIE if the task was not auto-reaped. So the task
should be deactivated when ->exit_state != 0. But see below.

> And now that "exit_state" is already separate from "state", the reason for
> PF_DEAD really is gone, and it could be replaced with a
>
> 	tsk->exit_state & EXIT_DEAD
>
> test instead.

No, I don't think this is right. After exit_notify() sets ->exit_state this
process is still valid from the scheduler POV, it may be preempted, may sleep.

So I beleive we really need to do something special when exiting process does
it's last context switch from do_exit(). If the process enters schedule() in
TASK_RUNNUNG, we should pass the hint to scheduler - PF_DEAD. Or we can change
the ->state before calling schedule().

Oleg.