Re: [PATCH 5/8] locking, mutex: Cancelable MCS lock for adaptive spinning

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

On Mon, 2014-02-10 at 22:32 +0100, Peter Zijlstra wrote:
> On Mon, Feb 10, 2014 at 01:15:59PM -0800, Jason Low wrote:
> > On Mon, 2014-02-10 at 20:58 +0100, Peter Zijlstra wrote:
> > > +void osq_unlock(struct optimistic_spin_queue **lock)
> > > +{
> > > +	struct optimistic_spin_queue *node = this_cpu_ptr(&osq_node);
> > > +	struct optimistic_spin_queue *next;
> > > +
> > > +	/*
> > > +	 * Fast path for the uncontended case.
> > > +	 */
> > > +	if (likely(cmpxchg(lock, node, NULL) == node))
> > > +		return;
> > 
> > Can we can also add the following code here as I'm noticing next != NULL
> > is the much more likely scenario on my box:
> > 
> >         next = xchg(&node->next, NULL);
> >         if (next) {
> >                 ACCESS_ONCE(next->locked) = 1;
> >                 return;
> 
> Is adding that really much faster than the relatively straight path
> oqs_wait_next() would walk to bit the same exit?
> 
> The only reason I pulled out the above cmpxchg() is because its the
> uncontended fast path, which seems like a special enough case.

So it would avoid 2 extra checks (*lock == node) and (node->next) in the
oqs_wait_next() path, which aren't necessary when node->next != NULL.

And I think node->next != NULL can be considered a special enough case
after the cmpxchg() fails because in the contended case, we're expecting
the node->next to be pointing at something. The only times node->next is
NULL after cmpxchg() fails are during a very small race window with the
osq_lock(), and when the next node is unqueuing due to need_resched,
which is also a very small window.