diff --git a/kernel/sched/bfs.c b/kernel/sched/bfs.c
index e430d062d2df..85bb250eca52 100644
--- a/kernel/sched/bfs.c
+++ b/kernel/sched/bfs.c
@@ -2929,11 +2929,9 @@ static void attach_task(struct rq *rq, struct task_struct *p)
  */
 static void attach_one_task(struct rq *rq, struct task_struct *p)
 {
-	unsigned long flags;
-
-	raw_spin_lock_irqsave(&rq->lock, flags);
+	raw_spin_lock(&rq->lock);
 	attach_task(rq, p);
-	raw_spin_unlock_irqrestore(&rq->lock, flags);
+	raw_spin_unlock(&rq->lock);
 }
 
 static int active_load_balance_cpu_stop(void *data)
@@ -2942,8 +2940,11 @@ static int active_load_balance_cpu_stop(void *data)
 	unsigned long flags;
 	struct task_struct *p;
 	int target_cpu = rq->push_cpu;
+	bool migration = false;
 
-	raw_spin_lock_irqsave(&rq->lock, flags);
+	local_irq_save(flags);
+
+	raw_spin_lock(&rq->lock);
 
 	p = rq_first_queued_task(rq);
 	if (unlikely(NULL == p))
@@ -2954,16 +2955,32 @@ static int active_load_balance_cpu_stop(void *data)
 		goto unlock_out;
 	}
 
-	detach_task(rq, p, target_cpu);
+	raw_spin_unlock(&rq->lock);
+	raw_spin_lock(&p->pi_lock);
+	raw_spin_lock(&rq->lock);
+
+	/*
+	 * _something_ may have changed the task, double check again
+	 */
+	if (likely(rq_first_queued_task(rq) == p &&
+		   rq == task_rq(p) &&
+		   cpumask_test_cpu(target_cpu, &p->cpus_allowed))) {
+		detach_task(rq, p, target_cpu);
+		migration = true;
+	}
 
 unlock_out:
 	rq->active_balance = 0;
-	raw_spin_unlock_irqrestore(&rq->lock, flags);
+	raw_spin_unlock(&rq->lock);
 
-	if (p) {
+	if (migration) {
 		struct rq *target_rq = cpu_rq(target_cpu);
 		attach_one_task(target_rq, p);
 	}
+	if (p)
+		raw_spin_unlock(&p->pi_lock);
+
+	local_irq_restore(flags);
 
 	return 0;
 }