[PATCH v5 4/8] drm/i915/gvt: introduce the vGPU LRU list

[Ping Gao]

vGPU LRU list is introduced to record the schedule status of all the
vGPUs, the longest unscheduled vGPU get the highest priority to
schedule, it's mainly used to avoid guest driver timeout when its
vGPU is not scheduled for a long time. Besides that it keeps the
round-robin scheduling fairness when scheduler need choose a new
head beacause of vGPU stop/kill.

Signed-off-by: Ping Gao <ping.a.gao at intel.com>
---
 drivers/gpu/drm/i915/gvt/sched_policy.c | 76 ++++++++++++++++++++++++++++++---
 1 file changed, 71 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/gvt/sched_policy.c b/drivers/gpu/drm/i915/gvt/sched_policy.c
index c6523b9..5e167eb 100644
--- a/drivers/gpu/drm/i915/gvt/sched_policy.c
+++ b/drivers/gpu/drm/i915/gvt/sched_policy.c
@@ -49,6 +49,7 @@ static bool vgpu_has_pending_workload(struct intel_vgpu *vgpu)
 
 struct vgpu_sched_data {
 	struct list_head list;
+	struct list_head lru_list;
 	struct intel_vgpu *vgpu;
 
 	ktime_t sched_in_time;
@@ -65,6 +66,7 @@ struct gvt_sched_data {
 	struct hrtimer timer;
 	unsigned long period;
 	struct list_head runq_head;
+	struct list_head lru_vgpu_head;
 };
 
 static void try_to_schedule_next_vgpu(struct intel_gvt *gvt)
@@ -141,6 +143,36 @@ static struct intel_vgpu *find_busy_vgpu(struct list_head *head,
 	return vgpu;
 }
 
+/* 1.5 second */
+#define VGPU_TDR_THRES_MS 1500
+
+static struct intel_vgpu *find_longest_unsched_vgpu(struct list_head *lru_vgpu_head)
+{
+	struct list_head *pos;
+	struct vgpu_sched_data *vgpu_data;
+	struct intel_vgpu *vgpu = NULL;
+
+	/* Find out the first active vGPU which has
+	 * pending workload in the lru list, the longest
+	 * unsched vGPU is the first one after head.
+	 */
+	list_for_each(pos, lru_vgpu_head) {
+		vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
+		if (vgpu_has_pending_workload(vgpu_data->vgpu)) {
+			vgpu = vgpu_data->vgpu;
+			break;
+		}
+	}
+
+	/* Return the vGPU if it's pending time exceed the threshold */
+	if (vgpu && (ktime_to_ms(ktime_sub(ktime_get(),
+						vgpu->last_ctx_submit_time)) >
+							VGPU_TDR_THRES_MS))
+		return vgpu;
+
+	return NULL;
+}
+
 static struct list_head *get_sched_head(struct gvt_sched_data *sched_data)
 {
 	struct intel_gvt *gvt = sched_data->gvt;
@@ -152,8 +184,17 @@ static struct list_head *get_sched_head(struct gvt_sched_data *sched_data)
 		cur_vgpu_data = scheduler->current_vgpu->sched_data;
 		head = &cur_vgpu_data->list;
 	} else {
+		struct vgpu_sched_data *lru_vgpu_data;
+
 		gvt_dbg_sched("no current vgpu search from q head\n");
-		head = &sched_data->runq_head;
+
+		/* Choose new head according to lru list to avoid
+		 * unfairness when current_vgpu set to NULL because
+		 * of the vGPU stop/kill.
+		 */
+		lru_vgpu_data = container_of(sched_data->lru_vgpu_head.next,
+					struct vgpu_sched_data, lru_list);
+		head = lru_vgpu_data->list.prev;
 	}
 
 	return head;
@@ -161,8 +202,9 @@ static struct list_head *get_sched_head(struct gvt_sched_data *sched_data)
 
 static struct intel_vgpu *pickup_next_vgpu(struct gvt_sched_data *sched_data)
 {
-	struct intel_vgpu *next_vgpu = NULL;
-	struct list_head *head = NULL;
+	struct vgpu_sched_data *next_vgpu_data;
+	struct intel_vgpu *next_vgpu;
+	struct list_head *head;
 
 	/* The scheduler follows round-robin style, sched
 	 * head means where start to choose next vGPU, it's
@@ -172,8 +214,22 @@ static struct intel_vgpu *pickup_next_vgpu(struct gvt_sched_data *sched_data)
 	 */
 	head = get_sched_head(sched_data);
 
-	/* Choose the vGPU which has pending workload. */
-	next_vgpu = find_busy_vgpu(head, sched_data);
+	/* Find out the vGPU which have not been scheduled near a max
+	 * threshold, then execute it immediately to avoid guest TDR.
+	 */
+	next_vgpu = find_longest_unsched_vgpu(&sched_data->lru_vgpu_head);
+	if (next_vgpu) {
+		next_vgpu_data = next_vgpu->sched_data;
+
+		/* Move the longest unsched vGPU after
+		 * current vGPU to keep fairness round-robin.
+		 */
+		list_del_init(&next_vgpu_data->list);
+		list_add(&next_vgpu_data->list, head);
+	} else {
+		/* Choose the vGPU which has pending workload */
+		next_vgpu = find_busy_vgpu(head, sched_data);
+	}
 
 	return next_vgpu;
 }
@@ -184,6 +240,7 @@ static void tbs_sched_func(struct gvt_sched_data *sched_data)
 {
 	struct intel_gvt *gvt = sched_data->gvt;
 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
+	struct vgpu_sched_data *vgpu_data;
 	struct intel_vgpu *vgpu = NULL;
 
 	/* no active vgpu or has already had a target */
@@ -195,6 +252,11 @@ static void tbs_sched_func(struct gvt_sched_data *sched_data)
 	if (vgpu) {
 		scheduler->next_vgpu = vgpu;
 		gvt_dbg_sched("pick next vgpu %d\n", vgpu->id);
+
+		/* Move the last used vGPU to the tail of lru_list */
+		vgpu_data = vgpu->sched_data;
+		list_del_init(&vgpu_data->lru_list);
+		list_add_tail(&vgpu_data->lru_list, &sched_data->lru_vgpu_head);
 	}
 out:
 	if (scheduler->next_vgpu) {
@@ -238,6 +300,7 @@ static int tbs_sched_init(struct intel_gvt *gvt)
 		return -ENOMEM;
 
 	INIT_LIST_HEAD(&data->runq_head);
+	INIT_LIST_HEAD(&data->lru_vgpu_head);
 	hrtimer_init(&data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
 	data->timer.function = tbs_timer_fn;
 	data->period = GVT_DEFAULT_TIME_SLICE;
@@ -270,6 +333,7 @@ static int tbs_sched_init_vgpu(struct intel_vgpu *vgpu)
 
 	data->vgpu = vgpu;
 	INIT_LIST_HEAD(&data->list);
+	INIT_LIST_HEAD(&data->lru_list);
 
 	vgpu->sched_data = data;
 
@@ -291,6 +355,7 @@ static void tbs_sched_start_schedule(struct intel_vgpu *vgpu)
 		return;
 
 	list_add_tail(&vgpu_data->list, &sched_data->runq_head);
+	list_add_tail(&vgpu_data->lru_list, &sched_data->lru_vgpu_head);
 
 	if (!hrtimer_active(&sched_data->timer))
 		hrtimer_start(&sched_data->timer, ktime_add_ns(ktime_get(),
@@ -302,6 +367,7 @@ static void tbs_sched_stop_schedule(struct intel_vgpu *vgpu)
 	struct vgpu_sched_data *vgpu_data = vgpu->sched_data;
 
 	list_del_init(&vgpu_data->list);
+	list_del_init(&vgpu_data->lru_list);
 }
 
 static struct intel_gvt_sched_policy_ops tbs_schedule_ops = {
-- 
2.7.4