Improve the gradation when setting the cache_distance, taking into account
different cores and different SMT siblings.

Try harder to reschedule tasks on the last CPU when doing resched_best_idle,
by not treating all threads and siblings as the same CPU. This improves
throughput on lightly loaded threaded CPUs (eg i7).

Remove the last_task as it was not significantly improving performance.

-ck

Index: linux-2.6.36-ck2/kernel/sched_bfs.c
===================================================================
--- linux-2.6.36-ck2.orig/kernel/sched_bfs.c	2010-12-13 09:11:26.468427288 +1100
+++ linux-2.6.36-ck2/kernel/sched_bfs.c	2010-12-14 18:00:38.178783004 +1100
@@ -190,7 +190,6 @@
 	u64 nohz_stamp;
 	unsigned char in_nohz_recently;
 #endif
-	struct task_struct *last_task;
 #endif
 
 	struct task_struct *curr, *idle;
@@ -762,19 +761,12 @@
 
 static void resched_task(struct task_struct *p);
 
-/*
- * last_task stores the last non-idle task scheduled on the local rq for
- * cache warmth testing.
- */
-static inline void set_last_task(struct rq *rq, struct task_struct *p)
-{
-	rq->last_task = p;
-}
-
-#define CPUIDLE_CACHE_BUSY	(1)
-#define CPUIDLE_DIFF_CPU	(2)
-#define CPUIDLE_THREAD_BUSY	(4)
-#define CPUIDLE_DIFF_NODE	(8)
+#define CPUIDLE_DIFF_THREAD	(1)
+#define CPUIDLE_DIFF_CORE	(2)
+#define CPUIDLE_CACHE_BUSY	(4)
+#define CPUIDLE_DIFF_CPU	(8)
+#define CPUIDLE_THREAD_BUSY	(16)
+#define CPUIDLE_DIFF_NODE	(32)
 
 /*
  * The best idle CPU is chosen according to the CPUIDLE ranking above where the
@@ -828,26 +820,29 @@
 		tmp_rq = cpu_rq(cpu_tmp);
 
 		if (rq->cpu_locality[cpu_tmp]) {
-			/* Check rq->last_task hasn't been dereferenced */
-			if (rq->last_task && p != rq->last_task) {
 #ifdef CONFIG_NUMA
-				if (rq->cpu_locality[cpu_tmp] > 1)
-					ranking |= CPUIDLE_DIFF_NODE;
+			if (rq->cpu_locality[cpu_tmp] > 3)
+				ranking |= CPUIDLE_DIFF_NODE;
+			else
 #endif
+			if (rq->cpu_locality[cpu_tmp] > 2)
 				ranking |= CPUIDLE_DIFF_CPU;
-			}
 		}
 #ifdef CONFIG_SCHED_MC
+		if (rq->cpu_locality[cpu_tmp] == 2)
+			ranking |= CPUIDLE_DIFF_CORE;
 		if (!(tmp_rq->cache_idle(cpu_tmp)))
 			ranking |= CPUIDLE_CACHE_BUSY;
 #endif
 #ifdef CONFIG_SCHED_SMT
+		if (rq->cpu_locality[cpu_tmp] == 1)
+			ranking |= CPUIDLE_DIFF_THREAD;
 		if (!(tmp_rq->siblings_idle(cpu_tmp)))
 			ranking |= CPUIDLE_THREAD_BUSY;
 #endif
 		if (ranking < best_ranking) {
 			best_cpu = cpu_tmp;
-			if (ranking <= 1)
+			if (ranking == 0)
 				break;
 			best_ranking = ranking;
 		}
@@ -877,11 +872,6 @@
 static inline int
 cache_distance(struct rq *task_rq, struct rq *rq, struct task_struct *p)
 {
-	/* Check rq->last_task hasn't been dereferenced */
-	if (likely(rq->last_task)) {
-		if (rq->last_task == p)
-			return 0;
-	}
 	return rq->cpu_locality[cpu_of(task_rq)] * task_timeslice(p);
 }
 #else /* CONFIG_SMP */
@@ -920,10 +910,6 @@
 {
 	return 0;
 }
-
-static inline void set_last_task(struct rq *rq, struct task_struct *p)
-{
-}
 #endif /* CONFIG_SMP */
 
 /*
@@ -2813,8 +2799,6 @@
 		sched_info_switch(prev, next);
 		perf_event_task_sched_out(prev, next);
 
-		if (prev != idle)
-			set_last_task(rq, prev);
 		set_rq_task(rq, next);
 		grq.nr_switches++;
 		prev->oncpu = 0;
@@ -6566,10 +6550,12 @@
 					cpumask_set_cpu(other_cpu, &rq->cache_siblings);
 			}
 #endif
-			if (sd->level <= SD_LV_MC)
-				locality = 0;
-			else if (sd->level <= SD_LV_NODE)
+			if (sd->level <= SD_LV_SIBLING)
 				locality = 1;
+			else if (sd->level <= SD_LV_MC)
+				locality = 2;
+			else if (sd->level <= SD_LV_NODE)
+				locality = 3;
 			else
 				continue;
 
@@ -6675,7 +6661,7 @@
 			if (i == j)
 				rq->cpu_locality[j] = 0;
 			else
-				rq->cpu_locality[j] = 3;
+				rq->cpu_locality[j] = 4;
 		}
 	}
 #endif