Re: [PATCH 02/11] wait: Provide infrastructure to deal with nested blocking

[Oleg Nesterov <oleg@redhat.com>]

On 10/02, Peter Zijlstra wrote:
>
> On Mon, Sep 29, 2014 at 11:02:21PM +0200, Oleg Nesterov wrote:
> > On 09/24, Peter Zijlstra wrote:
> > >
> > > +int woken_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
> > > +{
> > > +	/*
> > > +	 * Although this function is called under waitqueue lock, LOCK
> > > +	 * doesn't imply write barrier and the users expects write
> > > +	 * barrier semantics on wakeup functions.  The following
> > > +	 * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
> > > +	 * and is paired with set_mb() in wait_woken().
> > > +	 */
> > > +	smp_wmb(); /* C */
> > > +	wait->flags |= WQ_FLAG_WOKEN;
> >
> > Perhaps it is just me, but I was a bit confused by the comment above wmb().
> > Afaics, it is not that "users expects write barrier semantics", just we
> > need to ensure that
> >
> > 	CONDITION = true;
> > 	wait->flags |= WQ_FLAG_WOKEN;
> >
> > can't be reordered (and this differs from smp_wmb() in try_to_wake_up()).
> > Otherwise we can obviously race with
> >
> > 	// wait_woken() -> set_mb()
> > 	wait->flags &= ~WQ_FLAG_WOKEN;
> > 	mb();
> >
> > 	if (CONDITION)
> > 		break;
> >
>
> Yes, that comment could be clearer. It is however, to me, the 'same' as
> a regular wakeup in that we need to separate whatever state changes
> before the wakeup (CONDITION=true typically) from whatever writes are
> required to affect the wakeup (->state=TASK_RUNNING typically,

Not really, ttwu() needs to serialize CONDITION=true and the reading of
task->state. And for the waiter its state is write only, it doesn't need
to check it.

While in this case we need to separate CONDITION and WQ_FLAG_WOKEN, and
the waiter needs to check them in the right order.

But please forget, the code looks clear with or without the comment, and
"paired with set_mb() in wait_woken()" should explain the intent anyway.

Oleg.