Re: [PATCH 2/4] vfork: make it killable

[Oleg Nesterov <oleg@redhat.com>]

On 02/16, Andrew Morton wrote:
>
> On Thu, 16 Feb 2012 18:27:06 +0100
> Oleg Nesterov <oleg@redhat.com> wrote:
>
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -669,10 +669,34 @@ struct mm_struct *mm_access(struct task_struct *task, unsigned int mode)
> >
> >  void complete_vfork_done(struct task_struct *tsk)
> >  {
> > -	struct completion *vfork_done = tsk->vfork_done;
> > +	struct completion *vfork;
> >
> > -	tsk->vfork_done = NULL;
> > -	complete(vfork_done);
> > +	task_lock(tsk);
> > +	vfork = tsk->vfork_done;
> > +	if (likely(vfork)) {
> > +		tsk->vfork_done = NULL;
> > +		complete(vfork);
> > +	}
> > +	task_unlock(tsk);
> > +}
>
> OK, so now we don't need to test tsk->vfork_done in callers.  But
> mm_release() still does this, and it does it outside locks.

Yes, complete_vfork_done() can be called unconditionally,

> Mistake,
> or micro-optimisation?

micro-optimisation to avoid the unnecessary task_lock().

> If the latter, why is the lockless peek
> race-free?

If ->vfork_done != NULL, the child can never miss it. The parent
sets this pointer before the first wakeup.

However. The killed parent can clear ->vfork_done, see "if (killed)"
int wait_for_vfork_done(). That is why complete_vfork_done() should
re-check under task_lock().

Oleg.