Re: [PATCH 1/1] do_exit(): Solve possibility of BUG() due to race with try_to_wake_up()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Sep 04, 2014 at 09:15:26AM +0200, Peter Zijlstra wrote:
> On Wed, Sep 03, 2014 at 05:18:48PM +0200, Oleg Nesterov wrote:
> > On 09/03, Peter Zijlstra wrote:
> > >
> > > On Wed, Sep 03, 2014 at 03:36:40PM +0200, Oleg Nesterov wrote:
> > > >
> > > > 	// Ensure that the previous __set_current_state(RUNNING) can't
> > > > 	// leak after spin_unlock_wait()
> > > > 	smp_mb();
> > > > 	spin_unlock_wait();
> > > > 	// Another mb to ensure this too can't be reordered with unlock_wait
> > > > 	set_current_state(TASK_DEAD);
> > > >
> > > > What do you think looks better?
> > >
> > > spin_unlock_wait() would be a control dependency right? Therefore that
> > > store could not creep up anyhow.
> > 
> > Hmm. indeed, thanks! This probably means that task_work_run() can use
> > rmb() instead of mb().
> > 
> > What I can't understand is do we still need a compiler barrier or not.
> > Probably "in theory yes" ?
> 
> Yes, this is where I'm forever in doubt as well. The worry is the
> compiler reordering things, but I'm not sure how it would do that in
> this case, then again, I've been shown to not be creative enough in
> these cases many times before.
> 
> Paul might know, he's had much more exposure to compiler people.

Well, if we are talking about the code sequence above, spin_unlock_wait()
does reads followed by a conditional.  And set_current_state() does
a write.  The one thing that might be missing is that for this to work is
that Alpha might need an ACCESS_ONCE() to avoid read reordering.

(Yes, ACCESS_ONCE() is required for the other architectures to meet
the letter of the law in memory-barriers.txt, but if you know that
your particular architecture and compiler won't mess you up, you
have more freedom in your arch-specific code.)

							Thanx, Paul