[PATCH] sched: Don't push a higher priority class in get_push_task()

[Qais Yousef]

Commit a7c81556ec4d ("sched: Fix migrate_disable() vs rt/dl balancing")
will attempt to push/pull a higher priority task if the candidate task
is in migrate_disable() section. This is an attempt to prevent
starvation of these lower priority task that, in theory at least, could
end up in a situation where they're forever in migrate disable section
with no CPU time to run.

One issue with that is get_push_task() assumes rq->curr is of the same
sched_class, which AFAICT is not guaranteed to be true.

This patch adds extra bolts and braces to ensure that this voluntary
push operation is performed on a task of the same scheduling class only.

Otherwise an RT task could end up causing a DL task to be pushed away.
Which breaks the strict priority between sched classes.

We could also end up trying to push the migration task. Which I think is
harmless and is nothing but a wasted effort.

Fixes: a7c81556ec4d ("sched: Fix migrate_disable() vs rt/dl balancing")
Signed-off-by: Qais Yousef <qais.yousef@xxxxxxx>
---
 kernel/sched/deadline.c |  2 +-
 kernel/sched/rt.c       |  4 ++--
 kernel/sched/sched.h    | 17 ++++++++++++++++-
 3 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index aac3539aa0fe..afadc7e1f968 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -2276,7 +2276,7 @@ static void pull_dl_task(struct rq *this_rq)
 				goto skip;
 
 			if (is_migration_disabled(p)) {
-				push_task = get_push_task(src_rq);
+				push_task = get_push_task(src_rq, SCHED_DEADLINE);
 			} else {
 				deactivate_task(src_rq, p, 0);
 				set_task_cpu(p, this_cpu);
diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 8f720b71d13d..c2c5c08e3030 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -1892,7 +1892,7 @@ static int push_rt_task(struct rq *rq, bool pull)
 		 * to this other CPU, instead attempt to push the current
 		 * running task on this CPU away.
 		 */
-		push_task = get_push_task(rq);
+		push_task = get_push_task(rq, SCHED_FIFO);
 		if (push_task) {
 			raw_spin_unlock(&rq->lock);
 			stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
@@ -2225,7 +2225,7 @@ static void pull_rt_task(struct rq *this_rq)
 				goto skip;
 
 			if (is_migration_disabled(p)) {
-				push_task = get_push_task(src_rq);
+				push_task = get_push_task(src_rq, SCHED_FIFO);
 			} else {
 				deactivate_task(src_rq, p, 0);
 				set_task_cpu(p, this_cpu);
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 10a1522b1e30..4e156f008d22 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1954,12 +1954,27 @@ extern void trigger_load_balance(struct rq *rq);
 
 extern void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask, u32 flags);
 
-static inline struct task_struct *get_push_task(struct rq *rq)
+static inline struct task_struct *get_push_task(struct rq *rq, int policy)
 {
 	struct task_struct *p = rq->curr;
 
 	lockdep_assert_held(&rq->lock);
 
+	switch(policy) {
+	case SCHED_FIFO:
+	case SCHED_RR:
+		if (!rt_task(p))
+			return NULL;
+		break;
+	case SCHED_DEADLINE:
+		if (!dl_task(p))
+			return NULL;
+		break;
+	default:
+		WARN_ON_ONCE(1);
+		return NULL;
+	}
+
 	if (rq->push_busy)
 		return NULL;
 
-- 
2.25.1