[PATCH 5/7] drm/i915/execlists: Direct submit onto idle engines

[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_guc_submission.c | 10 ++--
 drivers/gpu/drm/i915/intel_lrc.c            | 56 +++++++++++++++++----
 drivers/gpu/drm/i915/intel_ringbuffer.h     |  7 +++
 3 files changed, 60 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
index 2feb65096966..d08ed64c5cec 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/intel_guc_submission.c
@@ -754,14 +754,18 @@ static bool __guc_dequeue(struct intel_engine_cs *engine)
 
 static void guc_dequeue(struct intel_engine_cs *engine)
 {
-	unsigned long flags;
+	unsigned long uninitialized_var(flags);
 	bool submit;
 
 	local_irq_save(flags);
 
-	spin_lock(&engine->timeline.lock);
+	if (!intel_engine_direct_submit(engine))
+		spin_lock(&engine->timeline.lock);
+
 	submit = __guc_dequeue(engine);
-	spin_unlock(&engine->timeline.lock);
+
+	if (!intel_engine_direct_submit(engine))
+		spin_unlock(&engine->timeline.lock);
 
 	if (submit)
 		guc_submit(engine);
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index ff7c6a741721..ee8d0c81bb64 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -599,6 +599,8 @@ static bool __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)
 			return false;
@@ -755,12 +757,16 @@ static bool __execlists_dequeue(struct intel_engine_cs *engine)
 static void execlists_dequeue(struct intel_engine_cs *engine)
 {
 	struct intel_engine_execlists * const execlists = &engine->execlists;
-	unsigned long flags;
+	unsigned long uninitialized_var(flags);
 	bool submit;
 
-	spin_lock_irqsave(&engine->timeline.lock, flags);
+	if (!intel_engine_direct_submit(engine))
+		spin_lock_irqsave(&engine->timeline.lock, flags);
+
 	submit = __execlists_dequeue(engine);
-	spin_unlock_irqrestore(&engine->timeline.lock, flags);
+
+	if (!intel_engine_direct_submit(engine))
+		spin_unlock_irqrestore(&engine->timeline.lock, flags);
 
 	if (submit)
 		execlists_submit_ports(engine);
@@ -1160,16 +1166,45 @@ static void queue_request(struct intel_engine_cs *engine,
 		      &lookup_priolist(engine, node, prio)->requests);
 }
 
-static void __submit_queue(struct intel_engine_cs *engine, int prio)
+static void __wakeup_queue(struct intel_engine_cs *engine, int prio)
 {
 	engine->execlists.queue_priority = prio;
+}
+
+static void __schedule_queue(struct intel_engine_cs *engine)
+{
 	tasklet_hi_schedule(&engine->execlists.tasklet);
 }
 
+static void __submit_queue(struct intel_engine_cs *engine)
+{
+	struct intel_engine_execlists * const execlists = &engine->execlists;
+	struct tasklet_struct * const t = &execlists->tasklet;
+
+	GEM_BUG_ON(!engine->i915->gt.awake);
+
+	/* If inside GPU reset, the tasklet will be queued automatically. */
+	if (unlikely(atomic_read(&t->count)))
+		return;
+
+	/* Directly submit the first request to reduce the initial latency */
+	if (!port_isset(execlists->port) && tasklet_trylock(t)) {
+		engine->flags |= I915_ENGINE_DIRECT_SUBMIT;
+		t->func(t->data);
+		engine->flags &= ~I915_ENGINE_DIRECT_SUBMIT;
+		tasklet_unlock(t);
+		return;
+	}
+
+	__schedule_queue(engine);
+}
+
 static void submit_queue(struct intel_engine_cs *engine, int prio)
 {
-	if (prio > engine->execlists.queue_priority)
-		__submit_queue(engine, prio);
+	if (prio > engine->execlists.queue_priority) {
+		__wakeup_queue(engine, prio);
+		__submit_queue(engine);
+	}
 }
 
 static void execlists_submit_request(struct i915_request *request)
@@ -1181,10 +1216,9 @@ 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);
 }
@@ -1306,8 +1340,10 @@ static void execlists_schedule(struct i915_request *request,
 		}
 
 		if (prio > engine->execlists.queue_priority &&
-		    i915_sw_fence_done(&sched_to_request(node)->submit))
-			__submit_queue(engine, prio);
+		    i915_sw_fence_done(&sched_to_request(node)->submit)) {
+			__wakeup_queue(engine, prio);
+			__schedule_queue(engine);
+		}
 	}
 
 	spin_unlock_irq(&engine->timeline.lock);
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index 010750e8ee44..f5545391d76a 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -569,6 +569,7 @@ struct intel_engine_cs {
 #define I915_ENGINE_NEEDS_CMD_PARSER BIT(0)
 #define I915_ENGINE_SUPPORTS_STATS   BIT(1)
 #define I915_ENGINE_HAS_PREEMPTION   BIT(2)
+#define I915_ENGINE_DIRECT_SUBMIT    BIT(3)
 	unsigned int flags;
 
 	/*
@@ -646,6 +647,12 @@ intel_engine_has_preemption(const struct intel_engine_cs *engine)
 	return engine->flags & I915_ENGINE_HAS_PREEMPTION;
 }
 
+static inline bool
+intel_engine_direct_submit(const struct intel_engine_cs *engine)
+{
+	return engine->flags & I915_ENGINE_DIRECT_SUBMIT;
+}
+
 static inline bool __execlists_need_preempt(int prio, int last)
 {
 	return prio > max(0, last);
-- 
2.17.0