[PATCH 0/2] small fix for fix_small_imbalance

[PATCH 0/2] small fix for fix_small_imbalance

[Lei Wen]

Here is two patches which correct the scale usage in the fix_small_balance,
and give out comment over when fix_small_imbalance would cause load change.

Lei Wen (2):
  sched: reduce calculation effort in fix_small_imbalance
  sched: scale the busy and this queue's per-task load before compare

 kernel/sched/fair.c |   37 ++++++++++++++++++++++---------------
 1 file changed, 22 insertions(+), 15 deletions(-)

-- 
1.7.10.4

[PATCH 1/2] sched: reduce calculation effort in fix_small_imbalance

[Lei Wen]

Actually all below item could be repalced by scaled_busy_load_per_task
	(sds->busiest_load_per_task * SCHED_POWER_SCALE)
		/sds->busiest->sgp->power;

Signed-off-by: Lei Wen <leiwen@marvell.com>
---
 kernel/sched/fair.c |   19 ++++++++-----------
 1 file changed, 8 insertions(+), 11 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index c61a614..28052fa 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4727,20 +4727,17 @@ void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
 	pwr_now /= SCHED_POWER_SCALE;
 
 	/* Amount of load we'd subtract */
-	tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
-		sds->busiest->sgp->power;
-	if (sds->max_load > tmp)
+	if (sds->max_load > scaled_busy_load_per_task) {
 		pwr_move += sds->busiest->sgp->power *
-			min(sds->busiest_load_per_task, sds->max_load - tmp);
-
-	/* Amount of load we'd add */
-	if (sds->max_load * sds->busiest->sgp->power <
-		sds->busiest_load_per_task * SCHED_POWER_SCALE)
-		tmp = (sds->max_load * sds->busiest->sgp->power) /
-			sds->this->sgp->power;
-	else
+			min(sds->busiest_load_per_task,
+				sds->max_load - scaled_busy_load_per_task);
 		tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
 			sds->this->sgp->power;
+	} else
+		tmp = (sds->max_load * sds->busiest->sgp->power) /
+			sds->this->sgp->power;
+
+	/* Amount of load we'd add */
 	pwr_move += sds->this->sgp->power *
 			min(sds->this_load_per_task, sds->this_load + tmp);
 	pwr_move /= SCHED_POWER_SCALE;
-- 
1.7.10.4

[PATCH 2/2] sched: scale the busy and this queue's per-task load before compare

[Lei Wen]

Since for max_load and this_load, they are the value that already be
scaled. It is not reasonble to get a minimum value between the scaled
and non-scaled value, like below example.
	min(sds->busiest_load_per_task, sds->max_load);

Also add comment over in what condition, there would be cpu power gain
in move the load.

Signed-off-by: Lei Wen <leiwen@marvell.com>
---
 kernel/sched/fair.c |   28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 28052fa..77a149c 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4692,7 +4692,7 @@ void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
 {
 	unsigned long tmp, pwr_now = 0, pwr_move = 0;
 	unsigned int imbn = 2;
-	unsigned long scaled_busy_load_per_task;
+	unsigned long scaled_busy_load_per_task, scaled_this_load_per_task;
 
 	if (sds->this_nr_running) {
 		sds->this_load_per_task /= sds->this_nr_running;
@@ -4714,6 +4714,9 @@ void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
 		return;
 	}
 
+	scaled_this_load_per_task = sds->this_load_per_task
+					 * SCHED_POWER_SCALE;
+	scaled_this_load_per_task /= sds->this->sgp->power;
 	/*
 	 * OK, we don't have enough imbalance to justify moving tasks,
 	 * however we may be able to increase total CPU power used by
@@ -4721,28 +4724,35 @@ void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
 	 */
 
 	pwr_now += sds->busiest->sgp->power *
-			min(sds->busiest_load_per_task, sds->max_load);
+			min(scaled_busy_load_per_task, sds->max_load);
 	pwr_now += sds->this->sgp->power *
-			min(sds->this_load_per_task, sds->this_load);
+			min(scaled_this_load_per_task, sds->this_load);
 	pwr_now /= SCHED_POWER_SCALE;
 
 	/* Amount of load we'd subtract */
 	if (sds->max_load > scaled_busy_load_per_task) {
 		pwr_move += sds->busiest->sgp->power *
-			min(sds->busiest_load_per_task,
+			min(scaled_busy_load_per_task,
 				sds->max_load - scaled_busy_load_per_task);
-		tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
-			sds->this->sgp->power;
+		tmp = scaled_busy_load_per_task;
 	} else
-		tmp = (sds->max_load * sds->busiest->sgp->power) /
-			sds->this->sgp->power;
+		tmp = sds->max_load;
 
+	/* Scale to this queue from busiest queue */
+	tmp = (tmp * sds->busiest->sgp->power) /
+		sds->this->sgp->power;
 	/* Amount of load we'd add */
 	pwr_move += sds->this->sgp->power *
-			min(sds->this_load_per_task, sds->this_load + tmp);
+			min(scaled_this_load_per_task, sds->this_load + tmp);
 	pwr_move /= SCHED_POWER_SCALE;
 
 	/* Move if we gain throughput */
+	/*
+	 * The only possibilty for below statement be true, is:
+	 * sds->max_load is larger than scaled_busy_load_per_task, while,
+	 * scaled_this_load_per_task is larger than sds->this_load plus by
+	 * the scaled scaled_busy_load_per_task moved into this queue
+	 */
 	if (pwr_move > pwr_now)
 		env->imbalance = sds->busiest_load_per_task;
 }
-- 
1.7.10.4