Re: [PATCH -v4 04/10] futex: Use smp_store_release() in mark_wake_futex()

From: Darren Hart
Date: Thu Mar 16 2017 - 20:35:35 EST


On Wed, Feb 22, 2017 at 03:03:16PM +0100, Peter Zijlstra wrote:
> On Fri, Dec 16, 2016 at 04:50:45PM -0800, Darren Hart wrote:
> > On Tue, Dec 13, 2016 at 09:36:42AM +0100, Peter Zijlstra wrote:
> > > Since the futex_q can dissapear the instruction after assigning NULL,
> > > this really should be a RELEASE barrier. That stops loads from hitting
> > > dead memory too.
> > >
> >
> > +Paul McKenney
> >
> > Per the introduction of the comment below from:
> >
> > f1a11e0 futex: remove the wait queue
> >
> > I believe the intent was to ensure the plist_del in ... the previous
> > __unqueue_futex(q) ... from getting ahead of the smp_store_release added here,
> > which could result in q being destroyed by the waking task before plist_del can
> > act on it. Is that
> > right?
> >
> > The comment below predates the refactoring which hid plist_del under the
> > __unqueue_futex() making it a bit less clear as to the associated plist_del:
> >
> > However, since this comment, we have moved the wake-up out of wake_futex through
> > the use of wake queues (wake_up_q) which now happens after the hb lock is
> > released (see futex_wake, futex_wake_op, and futex_requeue). Is this race still
> > a valid concern?
>
> Yes I think so, since __unqueue_futex() dereferences lock_ptr and does
> stores in the memory it points to, those stores must not happen _after_
> we NULL lock_ptr itself.

Are you referring to the q->lock_ptr = NULL in mark_wake_futex()?
So the concern is parallel mark_wake_futex() calls on the same futex? But that
can't happen because the call is wrapped by the hb locks. In what scenario can
this occur?

> futex_wait(), which calls unqueue_me() could have had a spurious wakeup
> and observe our NULL store and 'free' the futex_q.

Urg. Spurious wakeups... yes... OK, still necessary. Gah. :-(

--
Darren Hart
VMware Open Source Technology Center