[PATCH v2 6/9] drm/i915/gvt: introduce the vGPU LRU list

[Ping Gao]

vGPU LRU list used to record the schedule status of all the vGPUs, the
longest unsched vGPU get the highest priority to schedule, it's
mainly used to void guest driver trigger TDR because of vGPU service
is non-available for a long time. Besides that it keep 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 | 75 ++++++++++++++++++++++++++++++---
 1 file changed, 70 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/gvt/sched_policy.c b/drivers/gpu/drm/i915/gvt/sched_policy.c
index c174ce6..4dcbdc3 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 tbs_vgpu_data {
 	struct list_head list;
+	struct list_head lru_list;
 	struct intel_vgpu *vgpu;
 	/* put some per-vgpu sched stats here */
 	int64_t sched_in_time;
@@ -65,6 +66,7 @@ struct tbs_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 +143,35 @@ static struct intel_vgpu *get_vgpu_timeslice_left(struct list_head *head,
 	return vgpu;
 }
 
+/* 1.5 second */
+#define VGPU_TDR_THRES_MS  ((cycles_t)tsc_khz * 1500)
+
+static struct intel_vgpu *get_vgpu_longest_unsched(struct list_head *lru_vgpu_head)
+{
+	struct list_head *pos;
+	struct tbs_vgpu_data *vgpu_data;
+	struct intel_vgpu *vgpu = NULL;
+
+	/* Find out the first real vGPU which has
+	 * pending workload in the lru list.
+	 */
+	list_for_each(pos, lru_vgpu_head) {
+		vgpu_data = container_of(pos, struct tbs_vgpu_data, lru_list);
+		if (vgpu_data->vgpu->id != 0 &&
+				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 && (get_cycles() - vgpu->last_ctx_submit_time >
+			VGPU_TDR_THRES_MS))
+		return vgpu;
+
+	return NULL;
+}
+
 static struct list_head *get_sched_head(struct tbs_sched_data *sched_data)
 {
 	struct intel_gvt *gvt = sched_data->gvt;
@@ -152,8 +183,17 @@ static struct list_head *get_sched_head(struct tbs_sched_data *sched_data)
 		cur_vgpu_data = scheduler->current_vgpu->sched_data;
 		head = &cur_vgpu_data->list;
 	} else {
+		struct tbs_vgpu_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 tbs_vgpu_data, lru_list);
+		head = lru_vgpu_data->list.prev;
 	}
 
 	return head;
@@ -161,16 +201,31 @@ static struct list_head *get_sched_head(struct tbs_sched_data *sched_data)
 
 static struct intel_vgpu *pickup_next_vgpu(struct tbs_sched_data *sched_data)
 {
-	struct intel_vgpu *next_vgpu = NULL;
-	struct list_head *head = NULL;
+	struct tbs_vgpu_data *next_vgpu_data;
+	struct intel_vgpu *next_vgpu;
+	struct list_head *head;
 
 	/* The scheduler is follow round-robin style, sched
 	 * head means where start to choose next vGPU, that's
 	 * important to keep fairness. */
 	head = get_sched_head(sched_data);
 
-	/* Choose the vGPU which has timeslice left */
-	next_vgpu = get_vgpu_timeslice_left(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 = get_vgpu_longest_unsched(&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 timeslice left */
+		next_vgpu = get_vgpu_timeslice_left(head, sched_data);
+	}
 
 	return next_vgpu;
 }
@@ -181,6 +236,7 @@ static void tbs_sched_func(struct tbs_sched_data *sched_data)
 {
 	struct intel_gvt *gvt = sched_data->gvt;
 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
+	struct tbs_vgpu_data *vgpu_data;
 	struct intel_vgpu *vgpu = NULL;
 
 	/* no vgpu or has already had a target */
@@ -192,6 +248,11 @@ static void tbs_sched_func(struct tbs_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) {
@@ -235,6 +296,7 @@ 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;
@@ -267,6 +329,7 @@ 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;
 	return 0;
@@ -287,6 +350,7 @@ 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);
 
 	vgpu->gvt->num_vgpu_sched++;
 	if (vgpu->gvt->num_vgpu_sched == 2)
@@ -303,6 +367,7 @@ static void tbs_sched_stop_schedule(struct intel_vgpu *vgpu)
 		return;
 
 	list_del_init(&vgpu_data->list);
+	list_del_init(&vgpu_data->lru_list);
 
 	vgpu->gvt->num_vgpu_sched--;
 	if (vgpu->gvt->num_vgpu_sched == 1)
-- 
2.7.4