Merge "sched: break the forever prev_cpu selection preference"

include/linux/sched.h

@@ -1482,6 +1482,7 @@ struct task_struct {
 	u32 init_load_pct;
 	u64 last_wake_ts;
 	u64 last_switch_out_ts;
+	u64 last_cpu_selected_ts;
 	struct related_thread_group *grp;
 	struct list_head grp_list;
 	u64 cpu_cycles;

kernel/sched/core.c

@@ -3056,6 +3056,7 @@ static inline void mark_task_starting(struct task_struct *p)
 
 	wallclock = sched_ktime_clock();
 	p->ravg.mark_start = p->last_wake_ts = wallclock;
+	p->last_cpu_selected_ts = wallclock;
 	p->last_switch_out_ts = 0;
 	update_task_cpu_cycles(p, cpu_of(rq));
 }

kernel/sched/fair.c

@@ -2678,6 +2678,9 @@ static unsigned int __read_mostly
 sched_short_sleep_task_threshold = 2000 * NSEC_PER_USEC;
 unsigned int __read_mostly sysctl_sched_select_prev_cpu_us = 2000;
 
+static unsigned int __read_mostly
+sched_long_cpu_selection_threshold = 100 * NSEC_PER_MSEC;
+
 unsigned int __read_mostly sysctl_sched_restrict_cluster_spill;
 
 void update_up_down_migrate(void)
@@ -3455,6 +3458,7 @@ bias_to_prev_cpu(struct cpu_select_env *env, struct cluster_cpu_stats *stats)
 	struct sched_cluster *cluster;
 
 	if (env->boost_type != SCHED_BOOST_NONE || env->reason ||
+	    !task->ravg.mark_start ||
 	    env->need_idle || !sched_short_sleep_task_threshold)
 		return false;
 
@@ -3463,6 +3467,10 @@ bias_to_prev_cpu(struct cpu_select_env *env, struct cluster_cpu_stats *stats)
 					unlikely(!cpu_active(prev_cpu)))
 		return false;
 
+	if (task->ravg.mark_start - task->last_cpu_selected_ts >=
+				sched_long_cpu_selection_threshold)
+		return false;
+
 	/*
 	 * This function should be used by task wake up path only as it's
 	 * assuming p->last_switch_out_ts as last sleep time.
@@ -3604,6 +3612,7 @@ retry:
 		if (stats.best_capacity_cpu >= 0)
 			target = stats.best_capacity_cpu;
 	}
+	p->last_cpu_selected_ts = sched_ktime_clock();
 
 out:
 	rcu_read_unlock();