Re: [RFC] locking/rwsem: Avoid issuing wakeup before setting the reader waiter to nil

[Will Deacon]

On Thu, Nov 29, 2018 at 01:26:34PM -0500, Waiman Long wrote:
> On 11/29/2018 01:08 PM, Peter Zijlstra wrote:
> > Hmm, I think we're missing a barrier in wake_q_add(); when cmpxchg()
> > fails we still need an smp_mb().
> >
> > Something like so.
> >
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 3d87a28da378..69def558edf6 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -400,6 +400,13 @@ void wake_q_add(struct wake_q_head *head, struct task_struct *task)
> >  {
> >  	struct wake_q_node *node = &task->wake_q;
> >  
> > +	/*
> > +	 * Ensure, that when the below cmpxchg() fails, the corresponding
> > +	 * wake_up_q() will observe our prior state.
> > +	 *
> > +	 * Pairs with the smp_mb() from wake_up_q()'s wake_up_process().
> > +	 */
> > +	smp_mb();
> >  	/*
> >  	 * Atomically grab the task, if ->wake_q is !nil already it means
> >  	 * its already queued (either by us or someone else) and will get the
> > @@ -408,7 +415,7 @@ void wake_q_add(struct wake_q_head *head, struct task_struct *task)
> >  	 * This cmpxchg() executes a full barrier, which pairs with the full
> >  	 * barrier executed by the wakeup in wake_up_q().
> >  	 */
> > -	if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
> > +	if (cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))
> >  		return;
> >  
> >  	get_task_struct(task);
> 
> That can be costly for x86 which will now have 2 locked instructions.
> Should we introduce a kind of special cmpxchg (e.g. cmpxchg_mb) that
> will guarantee a memory barrier whether the operation fails or not?

I thought smp_mb__before_atomic() was designed for this sort of thing?

Will