[PATCH] sched: fix uneven per-cpu task_group share distribution

[Ken Chen]

Impact: fix group scheduling behavior

While testing CFS scheduler on linux-2.6-tip tree:

  git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip

We found that task which is pinned to a CPU could be starved relative to its
allocated fair share.

The per-cpu sched_enetity load share calculation in tg_shares_up /
update_group_shares_cpu distributes total task_group's share among all CPUs
for a given SD domain, this would dilute task_group's per-cpu share because
it distributes share to CPU that even has no load.  The trapped share is now
un-consumable and it leads to fair share starvation on the runnable CPU.
Peter was right that it is still required for the low level function to make
distinction between a boosted share that don't have any load and actual tg
share that should be distributed among CPUs in which the tg is running.

Patch to add that boost and we think the scheduler should only boost one
times of tg shares over all empty CPU that don't have any load for the
specific task_group in order to bound maximum temporary boost that a given
task_group can have.

Signed-off-by: Ken Chen <kenchen@xxxxxxxxxx>
Signed-off-by: Ingo Molnar <mingo@xxxxxxx>
---
 kernel/sched.c |   49 ++++++++++++++++++++++++++++++-------------------
 1 files changed, 30 insertions(+), 19 deletions(-)

diff --git a/kernel/sched.c b/kernel/sched.c
index ae5ca3f..7d07c97 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -1475,24 +1475,34 @@ static void __set_se_shares(struct sched_entity *se, unsigned long shares);
  * Calculate and set the cpu's group shares.
  */
 static void
-update_group_shares_cpu(struct task_group *tg, int cpu,
+update_group_shares_cpu(struct task_group *tg, int cpu, unsigned long boost,
 			unsigned long sd_shares, unsigned long sd_rq_weight)
 {
-	unsigned long shares;
+	unsigned long shares, raw_shares;
 	unsigned long rq_weight;
 
 	if (!tg->se[cpu])
 		return;
 
 	rq_weight = tg->cfs_rq[cpu]->rq_weight;
-
-	/*
-	 *           \Sum shares * rq_weight
-	 * shares =  -----------------------
-	 *               \Sum rq_weight
-	 *
-	 */
-	shares = (sd_shares * rq_weight) / sd_rq_weight;
+	if (rq_weight && sd_rq_weight) {
+		/*
+		 *           \Sum shares * rq_weight
+		 * shares =  -----------------------
+		 *               \Sum rq_weight
+		 *
+		 */
+		raw_shares = (sd_shares * rq_weight) / sd_rq_weight;
+		shares = raw_shares;
+	} else {
+		/*
+		 * If there are currently no tasks on the cpu pretend there
+		 * is one of average load so that when a new task gets to
+		 * run here it will not get delayed by group starvation.
+		 */
+		raw_shares = 0;
+		shares = boost;
+	}
 	shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
 
 	if (abs(shares - tg->se[cpu]->load.weight) >
@@ -1501,7 +1511,7 @@ update_group_shares_cpu(struct task_group *tg, int cpu,
 		unsigned long flags;
 
 		spin_lock_irqsave(&rq->lock, flags);
-		tg->cfs_rq[cpu]->shares = shares;
+		tg->cfs_rq[cpu]->shares = raw_shares;
 
 		__set_se_shares(tg->se[cpu], shares);
 		spin_unlock_irqrestore(&rq->lock, flags);
@@ -1517,18 +1527,14 @@ static int tg_shares_up(struct task_group *tg, void *data)
 {
 	unsigned long weight, rq_weight = 0;
 	unsigned long shares = 0;
+	unsigned long boost;
 	struct sched_domain *sd = data;
-	int i;
+	int i, no_load_count = 0;
 
 	for_each_cpu(i, sched_domain_span(sd)) {
-		/*
-		 * If there are currently no tasks on the cpu pretend there
-		 * is one of average load so that when a new task gets to
-		 * run here it will not get delayed by group starvation.
-		 */
 		weight = tg->cfs_rq[i]->load.weight;
 		if (!weight)
-			weight = NICE_0_LOAD;
+			no_load_count++;
 
 		tg->cfs_rq[i]->rq_weight = weight;
 		rq_weight += weight;
@@ -1541,8 +1547,13 @@ static int tg_shares_up(struct task_group *tg, void *data)
 	if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
 		shares = tg->shares;
 
+	if (no_load_count)
+		boost = shares / no_load_count;
+	else
+		boost = shares / cpumask_weight(sched_domain_span(sd));
+
 	for_each_cpu(i, sched_domain_span(sd))
-		update_group_shares_cpu(tg, i, shares, rq_weight);
+		update_group_shares_cpu(tg, i, boost, shares, rq_weight);
 
 	return 0;
 }
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/