Re: [PATCH v3] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Jason Low <jason.low2@hp.com>]

On Tue, 2015-04-28 at 10:50 -0700, Jason Low wrote:
> On Tue, 2015-04-28 at 19:17 +0200, Peter Zijlstra wrote:
> 
> > To me it makes more sense to reverse these two branches (identical code
> > wise of course) and put the special case first.
> > 
> > Alternatively we could also do something like the below, which to my
> > eyes looks a little better still, but I don't care too much.
> > 
> > 	if (rwsem_has_spinner(sem)) {
> > 		/*
> > 		 * comment ...
> > 		 */
> > 		 smp_rmb();
> > 		 if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
> > 			return sem;
> > 		 goto locked;
> > 	}
> > 
> > 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
> > locked:
> 
> How about putting this into its own function:
> 
> static inline bool __rwsem_wake_acquire_wait_lock(sem)
> {
> 	/*
> 	 *
> 	 * Comments
> 	 *
> 	 */
> 	if (unlikely(rwsem_has_spinner(sem))) {
> 		/*
> 		 * Comments
> 		 */
> 		smp_rmb();
> 		if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
> 			return false;
> 	}
> 
> 	return true;
> }

That is, with the raw_spin_lock_irqsave() too :)

static inline bool __rwsem_wake_acquire_wait_lock(sem)
{
 	/*
 	 *
 	 * Comments
 	 *
 	 */
 	if (unlikely(rwsem_has_spinner(sem))) {
 		/*
 		 * Comments
 		 */
 		smp_rmb();
 		if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
 			return false;

		/* trylock successful */
		return true;
 	}

	raw_spin_lock_irqsave(&sem->wait_lock, flags);
 	return true;
}