Re: Patch: linux-2.5.8-pre1/kernel/exit.c change caused BUG() atboot time

[Andrew Morton <akpm@zip.com.au>]

Linus Torvalds wrote:
> 
> On Thu, 4 Apr 2002, Andrew Morton wrote:
> >  Another approach would be:
> >
> > preempt_schedule()
> > {
> >       current->state2 = current->state;
> >       current->state = TASK_RUNNING;
> >       schedule();
> >       current->state = current->state2;
> > }
> 
> Yes, but please no.
> 
> My current tree says
> 
>         asmlinkage void preempt_schedule(void)
>         {
>                 if (unlikely(preempt_get_count()))
>                         return;
>                 if (current->state != TASK_RUNNING)
>                         return;
>                 schedule();
>         }
> 
> and if people start getting latency problems due to loops with state !=
> TASK_RUNNING, then I suspect we might just make "set_current_state()"
> check that case explicitly and do a conditional reschedule (ie make it the
> same as if we released a lock). That would be a hell of a lot cleaner, in
> my opinion.
> 

That would work.  And would also fix my "spin_unlock sometimes
stomps on TASK_INTERRUPTIBLE" problem.

It does mean that we'll need to convert many open-coded

	current->state = whatever;

instances to use [__]set_current_state().  But that's not
a bad thing.  Janitorial patches for this are already
floating about.

Robert, do you have time to do the code-and-test thing?

-