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

[Ingo Molnar]

* Linus Torvalds <torvalds@xxxxxxxx> wrote:

> > > I think it is better to set EXIT_DEAD in do_exit(), along with PF_DEAD 
> > > flag.
> > 
> > nice idea - your patch looks good to me.
> 
> I'm not entirely convinced.
> 
> The thing is, We used to have DEAD in the task state flags, ie 
> TASK_ZOMBIE was it.
> 
> We started using PF_DEAD in 2003 with this commit message:
> 
> Author: Linus Torvalds <torvalds@xxxxxxxxxxxxx>  2003-10-26 03:16:23
> 
>     Add a sticky "PF_DEAD" task flag to keep track of dead processes.
>     
>     Use this to simplify 'finish_task_switch', but perhaps more
>     importantly we can use this to track down why some processes
>     seem to sometimes not die properly even after having been
>     marked as ZOMBIE. The "task->state" flags are too fluid to 
>     allow that well.
> 
> ie the PF_DEAD flag was never really about is _needing_ it: it was all 
> about being able to safely check it _without_ having to rely on 
> task->state.

but this was in times when we did alot of nontrivial operations after we 
marked a task "dead". Today we do this:

        /* PF_DEAD causes final put_task_struct after we schedule. */
        preempt_disable();
        BUG_ON(tsk->flags & PF_DEAD);
        tsk->flags |= PF_DEAD;

        schedule();
        BUG();

(i introduced the above changes to make more of the exit path 
preemptable.)

PF_DEAD has zero relevance by today, and Oleg's patches are perfectly 
correct and dont add the kind of risk that they'd have meant in 2003.

> So putting it back into task->state is not wrong per se, but it kind 
> of misses the point of why it was somewhere else in the first place 
> (or rather, why it was there in the _second_ place, since it was in 
> task->state in the first place and got moved out of there).

yeah, PF_DEAD had its purpose back when we first marked a task dead, 
then we did the release_task(). We used to have problems with 
proc_pid_flush() which could sleep, which would lose the TASK_ZOMBIE or 
TASK_DEAD flag and we'd come back from the 'final' schedule().

today that's impossible, due to the code above - we only mark it PF_DEAD 
straight before going into the final schedule().

	Ingo
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/