[PATCH 9/9] drm/i915/execlists: Direct submit onto idle engine

[Chris Wilson]

Bypass using the tasklet to submit the first request to HW, as the
tasklet may be deferred unto ksoftirqd and at a minimum will add in
excess of 10us (and maybe tens of milliseconds) to our execution
latency. This latency reduction is most notable when execution flows
between engines.

Suggested-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/intel_lrc.c | 39 +++++++++++++++++++++++++-------
 1 file changed, 31 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index f9f4064dec0e..7d6c5190bfd0 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -554,7 +554,7 @@ static void inject_preempt_context(struct intel_engine_cs *engine)
 	execlists_set_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT);
 }
 
-static void execlists_dequeue(struct intel_engine_cs *engine)
+static bool __execlists_dequeue(struct intel_engine_cs *engine)
 {
 	struct intel_engine_execlists * const execlists = &engine->execlists;
 	struct execlist_port *port = execlists->port;
@@ -585,7 +585,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 	 * and context switches) submission.
 	 */
 
-	spin_lock_irq(&engine->timeline.lock);
 	rb = execlists->first;
 	GEM_BUG_ON(rb_first(&execlists->queue) != rb);
 
@@ -598,6 +597,8 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 		 */
 		GEM_BUG_ON(!execlists_is_active(execlists,
 						EXECLISTS_ACTIVE_USER));
+		GEM_BUG_ON(execlists_is_active(execlists,
+					       EXECLISTS_ACTIVE_PREEMPT));
 		GEM_BUG_ON(!port_count(&port[0]));
 		if (port_count(&port[0]) > 1)
 			goto unlock;
@@ -745,12 +746,23 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 	GEM_BUG_ON(execlists->first && !port_isset(execlists->port));
 
 unlock:
+	if (last)
+		execlists_user_begin(execlists, execlists->port);
+
+	return submit;
+}
+
+static void execlists_dequeue(struct intel_engine_cs *engine)
+{
+	struct intel_engine_execlists * const execlists = &engine->execlists;
+	bool submit;
+
+	spin_lock_irq(&engine->timeline.lock);
+	submit = __execlists_dequeue(engine);
 	spin_unlock_irq(&engine->timeline.lock);
 
-	if (submit) {
-		execlists_user_begin(execlists, execlists->port);
+	if (submit)
 		execlists_submit_ports(engine);
-	}
 
 	GEM_BUG_ON(port_isset(execlists->port) &&
 		   !execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
@@ -1149,8 +1161,20 @@ static void queue_request(struct intel_engine_cs *engine,
 
 static void __submit_queue(struct intel_engine_cs *engine, int prio)
 {
+	struct intel_engine_execlists * const execlists = &engine->execlists;
+
+	GEM_BUG_ON(!engine->i915->gt.awake);
+
 	engine->execlists.queue_priority = prio;
-	tasklet_hi_schedule(&engine->execlists.tasklet);
+
+	if (!READ_ONCE(execlists->active) &&
+	    tasklet_trylock(&execlists->tasklet)) {
+		if (__execlists_dequeue(engine))
+			execlists_submit_ports(engine);
+		tasklet_unlock(&execlists->tasklet);
+	} else {
+		tasklet_hi_schedule(&engine->execlists.tasklet);
+	}
 }
 
 static void submit_queue(struct intel_engine_cs *engine, int prio)
@@ -1168,10 +1192,9 @@ static void execlists_submit_request(struct i915_request *request)
 	spin_lock_irqsave(&engine->timeline.lock, flags);
 
 	queue_request(engine, &request->sched, rq_prio(request));
-	submit_queue(engine, rq_prio(request));
-
 	GEM_BUG_ON(!engine->execlists.first);
 	GEM_BUG_ON(list_empty(&request->sched.link));
+	submit_queue(engine, rq_prio(request));
 
 	spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
-- 
2.17.0