Re: [PATCH 3/8] vfork: make it killable

[Oleg Nesterov <oleg@redhat.com>]

On 07/29, Matt Fleming wrote:
>
> On Wed, 2011-07-27 at 18:32 +0200, Oleg Nesterov wrote:
>
> [...]
>
> >  static long clone_vfork_finish(struct task_struct *child,
> >  				struct completion *vfork_done, long pid)
> >  {
> > -	freezer_do_not_count();
> > -	wait_for_completion(vfork_done);
> > -	freezer_count();
> > +	int killed = wait_for_completion_killable(vfork_done);
> > +
> > +	if (killed) {
> > +		struct completion *steal = xchg(&child->vfork_done, NULL);
> > +		/* if we race with complete_vfork_done() we have to wait */
> > +		if (unlikely(!steal))
> > +			wait_for_completion(vfork_done);
> > +
> > +		return -EINTR;
> > +	}
>
> Hmm.. isn't this inherently racy anyway? Why go to the trouble of trying
> to handle this race at all?

Suppose the child does xchg() and sees vfork_done != NULL. In this
case the parent shouldn't return from do_fork() until the child
does complete(), this "struct completion" was allocated on parent's
stack.

OK, I am starting to agree this looks overcomplicated, task_lock()
can make the code look simpler (see 0/8).

Oleg.