[PATCH -mm 0/2] fix do_wait(!__WALL) hang (Was: mmotm 2009-08-24-16-24 uploaded)

[Oleg Nesterov <oleg@redhat.com>]

On 08/28, Roland McGrath wrote:
>
> > eligible_child:
> >
> > 	/* Wait for all children (clone and not) if __WALL is set;
> > 	 * otherwise, wait for clone children *only* if __WCLONE is
> > 	 * set; otherwise, wait for non-clone children *only*.  (Note:
> > 	 * A "clone" child here is one that reports to its parent
> > 	 * using a signal other than SIGCHLD.) */
> >
> > 	if (((p->exit_signal != SIGCHLD) ^ !!(wo->wo_flags & __WCLONE))
> > 	    && !(wo->wo_flags & __WALL))
> > 		return 0;
> >
> > I just can't understand what is the supposed behaviour when p is
> > sub-thread and p->exit_signal == -1.
>
> As you say, you're not even supposed to be here when exit_signal = -1,
> except for the ptrace case.  This logic exists for the non-CLONE_THREAD
> clone case, i.e. ->exit_signal != SIGCHLD and != -1.
>
> > IOW, perhaps this check should be
> >
> > 	if (!task_detached(p) && !(wo->wo_flags & __WALL) &&
> > 	    (p->exit_signal != SIGCHLD) ^ !!(wo->wo_flags & __WCLONE))
> > 		return 0;
>
> That seems OK to me.

And this looks more right...

> > In short. If ptracer calls wait4(ptraced_sub_thread), is it really
> > supposed it must use __WCLONE || __WALL?
>
> I suspect not, but I'm not quite sure.  That is, it makes sense to me that
> a ptracer should always get its tracees in all waits.  That is consistent
> with not having to use WUNTRACED, for example.  But I'm not really sure any
> more what the historical behavior of this has been.

Yes. Let's not change this code. This is user-visible, might break something.
I think it is better to make more "safe" fix.

Oleg.