sched/fair: Improve the ->group_imb logic

Group imbalance is meant to deal with situations where affinity masks
and sched domains don't align well, such as 3 cpus from one group and
6 from another. In this case the domain based balancer will want to
put an equal amount of tasks on each side even though they don't have
equal cpus.

Currently group_imb is set whenever two cpus of a group have a weight
difference of at least one avg task and the heaviest cpu has at least
two tasks. A group with imbalance set will always be picked as busiest
and a balance pass will be forced.

The problem is that even if there are no affinity masks this stuff can
trigger and cause weird balancing decisions, eg. the observed
behaviour was that of 6 cpus, 5 had 2 and 1 had 3 tasks, due to the
difference of 1 avg load (they all had the same weight) and nr_running
being >1 the group_imbalance logic triggered and did the weird thing
of pulling more load instead of trying to move the 1 excess task to
the other domain of 6 cpus that had 5 cpu with 2 tasks and 1 cpu with
1 task.

Curb the group_imbalance stuff by making the nr_running condition
weaker by also tracking the min_nr_running and using the difference in
nr_running over the set instead of the absolute max nr_running.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-9s7dedozxo8kjsb9kqlrukkf@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>

kernel/sched/fair.c

@@ -3775,7 +3775,8 @@ static inline void update_sg_lb_stats(struct lb_env *env,
 			int local_group, const struct cpumask *cpus,
 			int *balance, struct sg_lb_stats *sgs)
 {
-	unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
+	unsigned long nr_running, max_nr_running, min_nr_running;
+	unsigned long load, max_cpu_load, min_cpu_load;
 	unsigned int balance_cpu = -1, first_idle_cpu = 0;
 	unsigned long avg_load_per_task = 0;
 	int i;
@@ -3787,10 +3788,13 @@ static inline void update_sg_lb_stats(struct lb_env *env,
 	max_cpu_load = 0;
 	min_cpu_load = ~0UL;
 	max_nr_running = 0;
+	min_nr_running = ~0UL;
 
 	for_each_cpu_and(i, sched_group_cpus(group), cpus) {
 		struct rq *rq = cpu_rq(i);
 
+		nr_running = rq->nr_running;
+
 		/* Bias balancing toward cpus of our domain */
 		if (local_group) {
 			if (idle_cpu(i) && !first_idle_cpu) {
@@ -3801,16 +3805,19 @@ static inline void update_sg_lb_stats(struct lb_env *env,
 			load = target_load(i, load_idx);
 		} else {
 			load = source_load(i, load_idx);
-			if (load > max_cpu_load) {
+			if (load > max_cpu_load)
 				max_cpu_load = load;
-				max_nr_running = rq->nr_running;
-			}
 			if (min_cpu_load > load)
 				min_cpu_load = load;
+
+			if (nr_running > max_nr_running)
+				max_nr_running = nr_running;
+			if (min_nr_running > nr_running)
+				min_nr_running = nr_running;
 		}
 
 		sgs->group_load += load;
-		sgs->sum_nr_running += rq->nr_running;
+		sgs->sum_nr_running += nr_running;
 		sgs->sum_weighted_load += weighted_cpuload(i);
 		if (idle_cpu(i))
 			sgs->idle_cpus++;
@@ -3848,7 +3855,8 @@ static inline void update_sg_lb_stats(struct lb_env *env,
 	if (sgs->sum_nr_running)
 		avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
 
-	if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
+	if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
+	    (max_nr_running - min_nr_running) > 1)
 		sgs->group_imb = 1;
 
 	sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,