[PATCH 4/5] drm/i915/execlists: Direct submission from irq handler

[Chris Wilson]

Continuing the themem of bypassing ksoftirqd latency, also first try to
directly submit from the CS interrupt handler to clear the ELSP and
queue the next.

In the past, we have been hesitant to do this as the context switch
processing has been quite heavy, requiring forcewaked mmio. However, as
we now can read the GPU state from the cacheable HWSP, it is relatively
cheap!

v2: Explain why we test_bit(IRQ_EXECLIST) after doing notify_ring (it's
because the notify_ring() may itself trigger direct submission clearing
the bit)

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/i915_irq.c             | 21 ++++++++++++++-------
 drivers/gpu/drm/i915/i915_tasklet.h         | 19 +++++++++++++++++++
 drivers/gpu/drm/i915/intel_guc_submission.c |  2 ++
 3 files changed, 35 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index f8aff5a5aa83..e1b3a7575fe7 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1465,19 +1465,26 @@ gen8_cs_irq_handler(struct intel_engine_cs *engine, u32 iir)
 	struct intel_engine_execlists * const execlists = &engine->execlists;
 	bool tasklet = false;
 
-	if (iir & GT_CONTEXT_SWITCH_INTERRUPT) {
-		if (READ_ONCE(engine->execlists.active))
-			tasklet = !test_and_set_bit(ENGINE_IRQ_EXECLIST,
-						    &engine->irq_posted);
-	}
+	if (iir & GT_CONTEXT_SWITCH_INTERRUPT && READ_ONCE(execlists->active))
+		tasklet = !test_and_set_bit(ENGINE_IRQ_EXECLIST,
+					    &engine->irq_posted);
 
 	if (iir & GT_RENDER_USER_INTERRUPT) {
 		notify_ring(engine);
-		tasklet |= USES_GUC_SUBMISSION(engine->i915);
+		/*
+		 * notify_ring() may trigger direct submission onto this
+		 * engine, clearing the ENGINE_IRQ_EXECLIST bit. In that
+		 * case, we don't want to resubmit and so clear the tasklet
+		 * boolean. GuC never sets the ENGINE_IRQ_EXECLIST bit and
+		 * so when using the GuC this equates to an unconditional
+		 * setting of tasklet to true.
+		 */
+		if (!test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted))
+			tasklet = USES_GUC_SUBMISSION(engine->i915);
 	}
 
 	if (tasklet)
-		i915_tasklet_schedule(&execlists->tasklet);
+		i915_tasklet(&execlists->tasklet);
 }
 
 static void gen8_gt_irq_ack(struct drm_i915_private *i915,
diff --git a/drivers/gpu/drm/i915/i915_tasklet.h b/drivers/gpu/drm/i915/i915_tasklet.h
index 37341f6c5cbe..f1c3a0779cef 100644
--- a/drivers/gpu/drm/i915/i915_tasklet.h
+++ b/drivers/gpu/drm/i915/i915_tasklet.h
@@ -97,4 +97,23 @@ static inline bool i915_tasklet_direct_submit(const struct i915_tasklet *t)
 	return t->flags & I915_TASKLET_DIRECT_SUBMIT;
 }
 
+static inline bool i915_tasklet_try(struct i915_tasklet *t)
+{
+	if (unlikely(!tasklet_trylock(&t->base)))
+		return false;
+
+	t->base.func(t->base.data);
+	tasklet_unlock(&t->base);
+	return true;
+}
+
+static inline void i915_tasklet(struct i915_tasklet *t)
+{
+	if (!i915_tasklet_is_enabled(t)) /* GPU reset active */
+		return;
+
+	if (!i915_tasklet_try(t))
+		i915_tasklet_schedule(t);
+}
+
 #endif /* _I915_TASKLET_H_ */
diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
index fb8560a84f4f..4607cac87980 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/intel_guc_submission.c
@@ -780,6 +780,8 @@ static void guc_submission_tasklet(unsigned long data)
 	struct execlist_port *port = execlists->port;
 	struct i915_request *rq;
 
+	clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
+
 	rq = port_request(port);
 	while (rq && i915_request_completed(rq)) {
 		trace_i915_request_out(rq);
-- 
2.17.0