Re: [patch] sched: fix set_task_cpu() and provide an unlocked runqueue variant

[Mike Galbraith <efault@gmx.de>]

On Thu, 2009-11-26 at 10:35 +0100, Peter Zijlstra wrote:
> On Thu, 2009-11-26 at 02:31 +0100, Mike Galbraith wrote:
> > On Thu, 2009-11-26 at 02:01 +0100, Mike Galbraith wrote:
> > > On Wed, 2009-11-25 at 19:27 +0100, Peter Zijlstra wrote:
> > 
> > > > I've got to ask, what's that barrier for?
> > > 
> > > It's a leftover from frustrated bug hunting.
> > 
> > To be more specific, I put it there to ensure that min_vruntimes are
> > stable. 
> 
> min_vruntime should only ever be poked at when holding the respective
> rq->lock, even with a barrier a 64bit read on a 32bit machine can go all
> funny.

Yeah, but we're looking at an unlocked runqueue.  But never mind...

> >  I figured that if try_to_wake_up() needs a barrier to look at
> > task->state, I had better do the same for the runqueues.
> 
> Ah, ttwu() has that barrier for another reason. The wmb in ttwu() is to
> ensure the wakee sees the state of the waker at the time of waking.
> 
> That is, its about ordering things like:
> 
> 
>    A				   B
> 
> 
> my_cond = true;
> wake_process(my_friend);
> 
> 				while (!my_cond)
> 				  schedule();
> 
> 
> So that you can actually observe my_cond being true once you wakeup
> (schedule acts as a mb() when it actually schedules).

(Ah!  I think you actually made a wrinkle in grey-ware)

...ATM, I kinda wish I'd not gone off and read barriers.txt.  It hasn't
sunk in yet, but certainly has made me _paranoid as hell_ :-)

	-Mike