[PATCH v2 1/3] sched/fair: Record the task sleeping time as the cache hot duration

[Chen Yu <yu.c.chen@intel.com>]

The cache hot duration is calculated by the average sleeping
time of a task, which is the time delta between the task
being dequeued and enqueued.

The cache hot duration of a task is introduced to describe
how soon this dequeue task could be woken up. During this
cache hot period, the task's previous CPU is regarded as
still cache-hot for the task. This information will be used
by SIS_CACHE to improve cache locality for short-sleeping tasks.

Suggested-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Suggested-by: Aaron Lu <aaron.lu@intel.com>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
---
 include/linux/sched.h |  4 ++++
 kernel/sched/fair.c   | 39 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 43 insertions(+)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 8d258162deb0..7d0fafd29345 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1347,6 +1347,10 @@ struct task_struct {
 	struct callback_head		cid_work;
 #endif
 
+	u64				last_dequeue_time;
+	u64				avg_hot_dur;	/* Average cache hot duration */
+	int				last_dequeue_cpu;
+
 	struct tlbflush_unmap_batch	tlb_ubc;
 
 	/* Cache last used pipe for splice(): */
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 53e7bf2ccc44..672616503e35 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6667,6 +6667,36 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
 	struct sched_entity *se = &p->se;
 	int idle_h_nr_running = task_has_idle_policy(p);
 	int task_new = !(flags & ENQUEUE_WAKEUP);
+	u64 last_dequeue = p->last_dequeue_time;
+
+	if ((flags & ENQUEUE_WAKEUP) && last_dequeue &&
+	    cpu_online(p->last_dequeue_cpu)) {
+		/*
+		 * The enqueue task_cpu(p) has already been assigned
+		 * with a new one. Need to calculate the task's sleeping
+		 * time based on its previous running CPU.
+		 */
+		u64 now = sched_clock_cpu(p->last_dequeue_cpu);
+
+		/*
+		 * Record the task's short sleep time. This sleep time
+		 * indicates how soon this task might be woken up again.
+		 * The task's previous running CPU is regarded as cache-hot
+		 * in the sleep time. So, define the average sleep time of
+		 * the task as its cache-hot duration. The SIS could leverage
+		 * the cache-hot duration for better idle CPU selection.
+		 * This improves cache locality for short-sleeping tasks.
+		 *
+		 * If the sleep time is longer than sysctl_sched_migration_cost,
+		 * give the cache hot duration a penalty by cutting it to half.
+		 */
+		if (now > last_dequeue) {
+			if (now - last_dequeue < sysctl_sched_migration_cost)
+				update_avg(&p->avg_hot_dur, now - last_dequeue);
+			else
+				p->avg_hot_dur >>= 1;
+		}
+	}
 
 	/*
 	 * The code below (indirectly) updates schedutil which looks at
@@ -6821,6 +6851,15 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
 
 dequeue_throttle:
 	util_est_update(&rq->cfs, p, task_sleep);
+
+	if (task_sleep) {
+		p->last_dequeue_time = sched_clock_cpu(cpu_of(rq));
+		p->last_dequeue_cpu = cpu_of(rq);
+	} else {
+		/* 0 indicates the dequeue is not caused by sleep */
+		p->last_dequeue_time = 0;
+	}
+
 	hrtick_update(rq);
 }
 
-- 
2.25.1