Re: [PATCH 2/2] locking/rwsem: Use a return variable in rwsem_spin_on_owner()

[Ingo Molnar <mingo@kernel.org>]

* Jason Low <jason.low2@hp.com> wrote:

> On Thu, 2015-04-09 at 07:37 +0200, Ingo Molnar wrote:
> 
> > The 'break' path does not seem to be equivalent, we used to do:
> > 
> > > -			rcu_read_unlock();
> > > -			return false;
> > 
> > and now we'll do:
> > 
> > > +			ret = false;
> > ...
> > > +	if (!READ_ONCE(sem->owner)) {
> > > +		long count = READ_ONCE(sem->count);
> > 
> > it's harmless (we do one more round of checking), but that's not an 
> > equivalent transformation and slows down the preemption trigger a 
> > (tiny) bit, because the chance that we actually catch the lock when 
> > breaking out early is vanishingly small. (It might in fact do the 
> > wrong thing in returning true if need_resched() is set and we've 
> > switched owners in that small window.)
> > 
> > Given how dissimilar the return path is in this case, I'm not sure 
> > it's worth sharing it. This might be one of the few cases where 
> > separate return statements is the better solution.
> 
> I also preferred the multiple returns for the rwsem variant to avoid
> needing to check sem->owner when it should go to slowpath, as you
> mentioned.
> 
> Now that I think of it though, if we want to have just one return path,
> we can still do that if we add an "out" label.

That's the usual pattern we use, but:

> -		/* abort spinning when need_resched or owner is not running */
> +		/* Abort spinning when need_resched or owner is not running. */
>  		if (!owner->on_cpu || need_resched()) {
>  			rcu_read_unlock();
> -			return false;
> +			ret = false;
> +			goto out;
>  		}

The point is to generally unify the 'out' paths - i.e. to merge it 
with the rcu_read_unlock() as well, so that we have really simple 
gotos and only a single exit path.
 
That's not really doable here without extra overhead AFAICS, so I'd 
suggest we leave it alone ...

I have applied your other patch.

Thanks,

	Ingo