sched: use highest_prio.next to optimize pull operations

We currently take the rq->lock for every cpu in an overload state during
pull_rt_tasks().  However, we now have enough information via the
highest_prio.[curr|next] fields to determine if there is any tasks of
interest to warrant the overhead of the rq->lock, before we actually take
it.  So we use this information to reduce lock contention during the
pull for the case where the source-rq doesnt have tasks that preempt
the current task.
Signed-off-by: NGregory Haskins <ghaskins@novell.com>

kernel/sched_rt.c

@@ -1218,6 +1218,18 @@ static int pull_rt_task(struct rq *this_rq)
 			continue;
 
 		src_rq = cpu_rq(cpu);
+
+		/*
+		 * Don't bother taking the src_rq->lock if the next highest
+		 * task is known to be lower-priority than our current task.
+		 * This may look racy, but if this value is about to go
+		 * logically higher, the src_rq will push this task away.
+		 * And if its going logically lower, we do not care
+		 */
+		if (src_rq->rt.highest_prio.next >=
+		    this_rq->rt.highest_prio.curr)
+			continue;
+
 		/*
 		 * We can potentially drop this_rq's lock in
 		 * double_lock_balance, and another CPU could