load_balance() fixes for affinity

We noticed in testing that affined workloads do not provide consistent performance under certain circumstances. To isolate the issue, we began testing with a representative CPU workload. In some cases the CPU workload results would vary by a large percentage from run to run. We used JTAG and scheduler trace to try and profile the issue and noticed that at least one core was left idle for the duration of the test in the low score runs.

The specific test methodology used in our investigation involved using an executable that was compiled to fork a specific number of workers. The executable was then affined to a number of cores equal to the number of workers forked, using the userspace taskset utility. The expectation was that all cores in the mask of runnable CPUs would run a process until the work was complete. In practice, for several affinity masks, one to two cores would not receive work until late in the test or not at all.

Our first observation was that it appeared initial cpu assignment of the forked threads appeared suboptimal. We observed that it was highly probable that many or all of the threads would be initially assigned to the same CPU. Our limited investigation into this indicated that the forking of the threads occurs fast enough that load estimates for CPUs have not adjusted to the work they have been assigned. The function select_task_rq_fair() does not take number of running process into consideration and thus can easily assign several tasks spawned close together to the same CPU within the allowed CPUs for a workload. While this seems suboptimal, the load_balance() algorithm run on other idle cores in the affinity mask should resolve imbalances that result from it.

As our test case was leaving cores idle for many seconds at a time, we dug into load_balance() code and specifically the group_imbalance path. We added ftrace to the path in question and noticed that two branches in load_balance() conflict in their assessments of the ability of the algorithm to migrate load. The group_imbalance path correctly sets the flag to indicate the group can not be properly balanced due to affinity, but the redo condition right after this branch incorrectly assumes that there may be other cores with work to be pulled by considering cores outside of the scheduling domain in question.

Consider the following circumstance where (T) represents tasks and (*) represents the allowed CPUs:

A B C
{ 0 1 } { 2 3 } { 4 5 }
T T T T
T
* * * * *

In this case, core 0 in scheduling group 'A' should run load balance on its local scheduling domain and try to pull work from core 1. When it fails to move any work due to affinity, it will mark scheduling group 'A' as "group_imbalance". However, right after doing so, core 0 load_balance() evaluates the redo path with core 1 removed from consideration. This should leave no other cores to consider and result in a jump to "out_all_pinned", leaving the "group_imbalance" flag to be seen by future runs of the algorithm on core 3. As the code stands currently, core 0 evaluates all cores instead of those in the scheduling domain under consideration. It incorrectly sees other cores in the system and attempts a redo. With core 1 removed from the mask, the redo identifies no load imbalance and results in a jump to "out_balanced", clearing the "group_imbalance" flag.

We propose correcting this logic with patch 1 of this series.

Making this correction revealed an issue in the calculate_imbalance() path. Patch 2 removes a branch that does not make sense with the current load_balance() algorithm because it has no scenario where it benifits the "group_imbalance" case in calculate_imbalance() and which causes problems in systems with affined workloads and many idle or lightly loaded CPUs.

[V2]
-Corrected use of Signed-off-by tags
-Removed temp cpumask variable from stack

Jeffrey Hugo (2):
sched/fair: Fix load_balance() affinity redo path sched/fair: Remove group imbalance from calculate_imbalance()

 kernel/sched/fair.c | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

The group_imbalance path in calculate_imbalance() made sense when it was added back in 2007 with commit 908a7c1b9b80 ("sched: fix improper load balance across sched domain") because busiest->load_per_task factored into the amount of imbalance that was calculated. That is not the case today.

The group_imbalance path can only affect the outcome of calculate_imbalance() when the average load of the domain is less than the original busiest->load_per_task. In this case, busiest->load_per_task is overwritten with the scheduling domain load average. Thus busiest->load_per_task no longer represents actual load that can be moved.

At the final comparison between env->imbalance and busiest->load_per_task, imbalance may be larger than the new busiest->load_per_task causing the check to fail under the assumption that there is a task that could be migrated to satisfy the imbalance. However env->imbalance may still be smaller than the original busiest->load_per_task, thus it is unlikely that there is a task that can be migrated to satisfy the imbalance. Calculate_imbalance() would not choose to run fix_small_imbalance() when we expect it should. In the worst case, this can result in idle cpus.

Since the group imbalance path in calculate_imbalance() is at best a NOP but otherwise harmful, remove it.

---
 kernel/sched/fair.c | 9 ---------
 1 file changed, 9 deletions(-)

index 8f783ba..3283561 100644
@@ -7760,15 +7760,6 @@ static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *s
 	local = &sds->local_stat;
 	busiest = &sds->busiest_stat;
 
-	if (busiest->group_type == group_imbalanced) {
-		/*
-		 * In the group_imb case we cannot rely on group-wide averages
-		 * to ensure cpu-load equilibrium, look at wider averages. XXX
-		 */
-		busiest->load_per_task =
-			min(busiest->load_per_task, sds->avg_load);
-	}
-
 	/*
 	 * Avg load of busiest sg can be less and avg load of local sg can
 	 * be greater than avg load across all sgs of sd because avg load

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source cpu from consideration and attempts to redo and balance among the new subset of cpus.

There is a bug in this code path where the algorithm considers all active cpus in the system (minus the source that was just masked out). This is not valid for two reasons: some active cpus may not be in the current scheduling domain and one of the active cpus is dst_cpu. These cpus should not be considered, as we cannot pull load from them.

Instead of failing out of load_balance(), we may end up redoing the search with no valid cpus and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other cpus in future load_balance() runs as that algorithm intends.

Fix the check by removing cpus not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining cpus from which load might be migrated.

---
 kernel/sched/fair.c | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

index d711093..a5d41b1 100644
@@ -8220,7 +8220,24 @@ static int load_balance(int this_cpu, struct rq *this_rq,
 		/* 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)) {
+			/*
+			 * dst_cpu is not a valid busiest cpu in the following
+			 * check since load cannot be pulled from dst_cpu to be
+			 * put on dst_cpu.
+			 */
+			cpumask_clear_cpu(env.dst_cpu, cpus);
+			/*
+			 * Go back to "redo" iff the load-balance cpumask
+			 * contains other potential busiest cpus for the
+			 * current sched domain.
+			 */
+			if (cpumask_intersects(cpus, sched_domain_span(env.sd))) {
+				/*
+				 * Now that the check has passed, reenable
+				 * dst_cpu so that load can be calculated on
+				 * it in the redo path.
+				 */
+				cpumask_set_cpu(env.dst_cpu, cpus);
 				env.loop = 0;
 				env.loop_break = sched_nr_migrate_break;
 				goto redo;

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.