Skip to contentSkip to conversation

sched/fair: Restore env status before goto redo in load_balance()

PATCH In review Xunlei Pang 3 participants 4 messages Mar 18, 2015 – Apr 1, 2015 Scheduler

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched/fair: Restore env status before goto redo in load_balance()

From: Xunlei Pang <[email protected]>

In load_balance(), some members of lb_env will be assigned with new values in LBF_DST_PINNED case. But lb_env::flags may still retain LBF_ALL_PINNED if no proper tasks were found afterwards due to another balance, task affinity changing, etc, which can really happen because busiest rq lock has already been released.

This is wrong, for example with env.dst_cpu assigned new_dst_cpu when going back to "redo" label, it may cause should_we_balance() to return false which is unreasonable.

This patch restores proper status of env before "goto redo", and improves "out_all_pinned" and "out_one_pinned" labels.

---
 kernel/sched/fair.c | 35 ++++++++++++++++++++---------------
 1 file changed, 20 insertions(+), 15 deletions(-)
kernel/sched/fair.c +20 -15
index ee595ef..45bbda1 100644
@@ -6843,6 +6843,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
 		.dst_cpu	= this_cpu,
 		.dst_rq		= this_rq,
 		.dst_grpmask    = sched_group_cpus(sd->groups),
+		.new_dst_cpu	= -1,
 		.idle		= idle,
 		.loop_break	= sched_nr_migrate_break,
 		.cpus		= cpus,
@@ -6977,12 +6978,19 @@ more_balance:
 		/* All tasks on this runqueue were pinned by CPU affinity */
 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
 			cpumask_clear_cpu(cpu_of(busiest), cpus);
-			if (!cpumask_empty(cpus)) {
-				env.loop = 0;
-				env.loop_break = sched_nr_migrate_break;
-				goto redo;
+			if (env.new_dst_cpu != -1) {
+				env.new_dst_cpu = -1;
+				cpumask_or(cpus, cpus,
+					sched_group_cpus(sd->groups));
+				cpumask_and(cpus, cpus, cpu_active_mask);
+
+				env.dst_cpu = this_cpu;
+				env.dst_rq = this_rq;
 			}
-			goto out_all_pinned;
+			env.flags &= ~LBF_SOME_PINNED;
+			env.loop = 0;
+			env.loop_break = sched_nr_migrate_break;
+			goto redo;
 		}
 	}
 
@@ -7009,7 +7017,7 @@ more_balance:
 				raw_spin_unlock_irqrestore(&busiest->lock,
 							    flags);
 				env.flags |= LBF_ALL_PINNED;
-				goto out_one_pinned;
+				goto out_active_balanced;
 			}
 
 			/*
@@ -7058,26 +7066,23 @@ more_balance:
 out_balanced:
 	/*
 	 * We reach balance although we may have faced some affinity
-	 * constraints. Clear the imbalance flag if it was set.
+	 * constraints.
+	 *
+	 * When LBF_ALL_PINNED was not set, clear the imbalance flag
+	 * if it was set.
 	 */
-	if (sd_parent) {
+	if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
 
 		if (*group_imbalance)
 			*group_imbalance = 0;
 	}
 
-out_all_pinned:
-	/*
-	 * We reach balance because all tasks are pinned at this level so
-	 * we can't migrate them. Let the imbalance flag set so parent level
-	 * can try to migrate them.
-	 */
 	schedstat_inc(sd, lb_balanced[idle]);
 
 	sd->nr_balance_failed = 0;
 
-out_one_pinned:
+out_active_balanced:
 	/* tune up the balancing interval */
 	if (((env.flags & LBF_ALL_PINNED) &&
 			sd->balance_interval < MAX_PINNED_INTERVAL) ||
Signature · 1 line
1.9.1
#1 lore raw

Ping Peter

Quoted 103 lines

From: Xunlei Pang <[email protected]>

In load_balance(), some members of lb_env will be assigned with new values in LBF_DST_PINNED case. But lb_env::flags may still retain LBF_ALL_PINNED if no proper tasks were found afterwards due to another balance, task affinity changing, etc, which can really happen because busiest rq lock has already been released.

This is wrong, for example with env.dst_cpu assigned new_dst_cpu when going back to "redo" label, it may cause should_we_balance() to return false which is unreasonable.

This patch restores proper status of env before "goto redo", and improves "out_all_pinned" and "out_one_pinned" labels.

---
 kernel/sched/fair.c | 35 ++++++++++++++++++++---------------
 1 file changed, 20 insertions(+), 15 deletions(-)
kernel/sched/fair.c +20 -16
index ee595ef..45bbda1 100644
@@ -6843,6 +6843,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
                .dst_cpu        = this_cpu,
                .dst_rq         = this_rq,
                .dst_grpmask    = sched_group_cpus(sd->groups),
+               .new_dst_cpu    = -1,
                .idle           = idle,
                .loop_break     = sched_nr_migrate_break,
                .cpus           = cpus,
@@ -6977,12 +6978,19 @@ more_balance:
                /* All tasks on this runqueue were pinned by CPU affinity */
                if (unlikely(env.flags & LBF_ALL_PINNED)) {
                        cpumask_clear_cpu(cpu_of(busiest), cpus);
-                       if (!cpumask_empty(cpus)) {
-                               env.loop = 0;
-                               env.loop_break = sched_nr_migrate_break;
-                               goto redo;
+                       if (env.new_dst_cpu != -1) {
+                               env.new_dst_cpu = -1;
+                               cpumask_or(cpus, cpus,
+                                       sched_group_cpus(sd->groups));
+                               cpumask_and(cpus, cpus, cpu_active_mask);
+
+                               env.dst_cpu = this_cpu;
+                               env.dst_rq = this_rq;
                        }
-                       goto out_all_pinned;
+                       env.flags &= ~LBF_SOME_PINNED;
+                       env.loop = 0;
+                       env.loop_break = sched_nr_migrate_break;
+                       goto redo;
                }
        }
 
@@ -7009,7 +7017,7 @@ more_balance:
                                raw_spin_unlock_irqrestore(&busiest->lock,
                                                            flags);
                                env.flags |= LBF_ALL_PINNED;
-                               goto out_one_pinned;
+                               goto out_active_balanced;
                        }
 
                        /*
@@ -7058,26 +7066,23 @@ more_balance:
 out_balanced:
        /*
         * We reach balance although we may have faced some affinity
-        * constraints. Clear the imbalance flag if it was set.
+        * constraints.
+        *
+        * When LBF_ALL_PINNED was not set, clear the imbalance flag
+        * if it was set.
         */
-       if (sd_parent) {
+       if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
                int *group_imbalance = &sd_parent->groups->sgc->imbalance;
 
                if (*group_imbalance)
                        *group_imbalance = 0;
        }
 
-out_all_pinned:
-       /*
-        * We reach balance because all tasks are pinned at this level so
-        * we can't migrate them. Let the imbalance flag set so parent level
-        * can try to migrate them.
-        */
        schedstat_inc(sd, lb_balanced[idle]);
 
        sd->nr_balance_failed = 0;
 
-out_one_pinned:
+out_active_balanced:
        /* tune up the balancing interval */
        if (((env.flags & LBF_ALL_PINNED) &&
                        sd->balance_interval < MAX_PINNED_INTERVAL) ||
--

1.9.1

maintainer
#2 lore raw
In reply to Xunlei Pang
Quoted 7 lines · Xunlei Pang

From: Xunlei Pang <[email protected]>

In load_balance(), some members of lb_env will be assigned with new values in LBF_DST_PINNED case. But lb_env::flags may still retain LBF_ALL_PINNED if no proper tasks were found afterwards due to another balance, task affinity changing, etc, which can really happen because busiest rq lock has already been released.

Sure..

Quoted 3 lines

This is wrong, for example with env.dst_cpu assigned new_dst_cpu when going back to "redo" label, it may cause should_we_balance() to return false which is unreasonable.

Why? You've got a very unlikely, very hard case, its unlikely that anything we do will substantially improve the situation, but you make the code uglier for it.

Quoted 2 lines

This patch restores proper status of env before "goto redo", and improves "out_all_pinned" and "out_one_pinned" labels.

That doesn't even begin to explain half of what the patch does.

kernel/sched/fair.c more_balance: · line 6981

Quoted 9 lines
kernel/sched/fair.c +1 -4
@@ -6977,12 +6978,19 @@ more_balance:
 		/* All tasks on this runqueue were pinned by CPU affinity */
 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
 			cpumask_clear_cpu(cpu_of(busiest), cpus);
-			if (!cpumask_empty(cpus)) {
-				env.loop = 0;
-				env.loop_break = sched_nr_migrate_break;
-				goto redo;
+			if (env.new_dst_cpu != -1) {

I really don't get this, how can this not be?

kernel/sched/fair.c more_balance: · line 6985

Quoted 4 lines
kernel/sched/fair.c +4 -0
+				env.new_dst_cpu = -1;
+				cpumask_or(cpus, cpus,
+					sched_group_cpus(sd->groups));
+				cpumask_and(cpus, cpus, cpu_active_mask);

More unexplained magic, why is this right?

The rest of the patch isn't much better.

#3 lore raw
Quoted 36 lines · Peter Zijlstra
Quoted 7 lines · Xunlei Pang

From: Xunlei Pang <[email protected]>

In load_balance(), some members of lb_env will be assigned with new values in LBF_DST_PINNED case. But lb_env::flags may still retain LBF_ALL_PINNED if no proper tasks were found afterwards due to another balance, task affinity changing, etc, which can really happen because busiest rq lock has already been released.

Sure..

Quoted 3 lines

This is wrong, for example with env.dst_cpu assigned new_dst_cpu when going back to "redo" label, it may cause should_we_balance() to return false which is unreasonable.

Why? You've got a very unlikely, very hard case, its unlikely that anything we do will substantially improve the situation, but you make the code uglier for it.

Quoted 2 lines

This patch restores proper status of env before "goto redo", and improves "out_all_pinned" and "out_one_pinned" labels.

That doesn't even begin to explain half of what the patch does.

kernel/sched/fair.c more_balance: · line 6981

Quoted 9 lines
kernel/sched/fair.c +1 -4
@@ -6977,12 +6978,19 @@ more_balance:
              /* All tasks on this runqueue were pinned by CPU affinity */
              if (unlikely(env.flags & LBF_ALL_PINNED)) {
                      cpumask_clear_cpu(cpu_of(busiest), cpus);
-                     if (!cpumask_empty(cpus)) {
-                             env.loop = 0;
-                             env.loop_break = sched_nr_migrate_break;
-                             goto redo;
+                     if (env.new_dst_cpu != -1) {

I really don't get this, how can this not be?

Quoted 6 lines

kernel/sched/fair.c more_balance: · line 6985

Quoted 4 lines
kernel/sched/fair.c +4 -0
+                             env.new_dst_cpu = -1;
+                             cpumask_or(cpus, cpus,
+                                     sched_group_cpus(sd->groups));
+                             cpumask_and(cpus, cpus, cpu_active_mask);

More unexplained magic, why is this right?

When LBF_DST_PINNED was set, after going back to "more_balance", things may change as the changelog describes, so it can hit LBF_ALL_PINNED afterwards. Then env.cpus, env.dst_rq, env.dst_cpu held the values assigned in the LBF_DST_PINNED case which is unreasonable. When we want to redo, we must reset those values.

Quoted 2 lines

The rest of the patch isn't much better.