drm/i915: Only enqueue already completed requests

If we are asked to submit a completed request, just move it onto the
active-list without modifying it's payload. If we try to emit the
modified payload of a completed request, we risk racing with the
ring->head update during retirement which may advance the head past our
breadcrumb and so we generate a warning for the emission being behind
the RING_HEAD.

v2: Commentary for the sneaky, shared responsibility between functions.
v3: Spelling mistakes and bonus assertion

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190923110056.15176-3-chris@chris-wilson.co.uk
(cherry picked from commit c0bb487dc19fc45dbeede7dcf8f513df51a3cd33)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>

3 files changed, 74 insertions, 38 deletions

diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index 21f06d5db9dd7..b9520319684aa 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -796,6 +796,17 @@ static bool can_merge_rq(const struct i915_request *prev,
 	GEM_BUG_ON(prev == next);
 	GEM_BUG_ON(!assert_priority_queue(prev, next));
 
+	/*
+	 * We do not submit known completed requests. Therefore if the next
+	 * request is already completed, we can pretend to merge it in
+	 * with the previous context (and we will skip updating the ELSP
+	 * and tracking). Thus hopefully keeping the ELSP full with active
+	 * contexts, despite the best efforts of preempt-to-busy to confuse
+	 * us.
+	 */
+	if (i915_request_completed(next))
+		return true;
+
 	if (!can_merge_ctx(prev->hw_context, next->hw_context))
 		return false;
 
@@ -1171,21 +1182,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				continue;
 			}
 
-			if (i915_request_completed(rq)) {
-				ve->request = NULL;
-				ve->base.execlists.queue_priority_hint = INT_MIN;
-				rb_erase_cached(rb, &execlists->virtual);
-				RB_CLEAR_NODE(rb);
-
-				rq->engine = engine;
-				__i915_request_submit(rq);
-
-				spin_unlock(&ve->base.active.lock);
-
-				rb = rb_first_cached(&execlists->virtual);
-				continue;
-			}
-
 			if (last && !can_merge_rq(last, rq)) {
 				spin_unlock(&ve->base.active.lock);
 				return; /* leave this for another */
@@ -1236,11 +1232,23 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				GEM_BUG_ON(ve->siblings[0] != engine);
 			}
 
-			__i915_request_submit(rq);
-			if (!i915_request_completed(rq)) {
+			if (__i915_request_submit(rq)) {
 				submit = true;
 				last = rq;
 			}
+
+			/*
+			 * Hmm, we have a bunch of virtual engine requests,
+			 * but the first one was already completed (thanks
+			 * preempt-to-busy!). Keep looking at the veng queue
+			 * until we have no more relevant requests (i.e.
+			 * the normal submit queue has higher priority).
+			 */
+			if (!submit) {
+				spin_unlock(&ve->base.active.lock);
+				rb = rb_first_cached(&execlists->virtual);
+				continue;
+			}
 		}
 
 		spin_unlock(&ve->base.active.lock);
@@ -1253,8 +1261,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 		int i;
 
 		priolist_for_each_request_consume(rq, rn, p, i) {
-			if (i915_request_completed(rq))
-				goto skip;
+			bool merge = true;
 
 			/*
 			 * Can we combine this request with the current port?
@@ -1295,14 +1302,23 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				    ctx_single_port_submission(rq->hw_context))
 					goto done;
 
-				*port = execlists_schedule_in(last, port - execlists->pending);
-				port++;
+				merge = false;
 			}
 
-			last = rq;
-			submit = true;
-skip:
-			__i915_request_submit(rq);
+			if (__i915_request_submit(rq)) {
+				if (!merge) {
+					*port = execlists_schedule_in(last, port - execlists->pending);
+					port++;
+					last = NULL;
+				}
+
+				GEM_BUG_ON(last &&
+					   !can_merge_ctx(last->hw_context,
+							  rq->hw_context));
+
+				submit = true;
+				last = rq;
+			}
 		}
 
 		rb_erase_cached(&p->node, &execlists->queue);
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 8f88851a9306e..1c5506822dc72 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -377,9 +377,10 @@ __i915_request_await_execution(struct i915_request *rq,
 	return 0;
 }
 
-void __i915_request_submit(struct i915_request *request)
+bool __i915_request_submit(struct i915_request *request)
 {
 	struct intel_engine_cs *engine = request->engine;
+	bool result = false;
 
 	GEM_TRACE("%s fence %llx:%lld, current %d\n",
 		  engine->name,
@@ -389,6 +390,25 @@ void __i915_request_submit(struct i915_request *request)
 	GEM_BUG_ON(!irqs_disabled());
 	lockdep_assert_held(&engine->active.lock);
 
+	/*
+	 * With the advent of preempt-to-busy, we frequently encounter
+	 * requests that we have unsubmitted from HW, but left running
+	 * until the next ack and so have completed in the meantime. On
+	 * resubmission of that completed request, we can skip
+	 * updating the payload, and execlists can even skip submitting
+	 * the request.
+	 *
+	 * We must remove the request from the caller's priority queue,
+	 * and the caller must only call us when the request is in their
+	 * priority queue, under the active.lock. This ensures that the
+	 * request has *not* yet been retired and we can safely move
+	 * the request into the engine->active.list where it will be
+	 * dropped upon retiring. (Otherwise if resubmit a *retired*
+	 * request, this would be a horrible use-after-free.)
+	 */
+	if (i915_request_completed(request))
+		goto xfer;
+
 	if (i915_gem_context_is_banned(request->gem_context))
 		i915_request_skip(request, -EIO);
 
@@ -412,13 +432,18 @@ void __i915_request_submit(struct i915_request *request)
 	    i915_sw_fence_signaled(&request->semaphore))
 		engine->saturated |= request->sched.semaphores;
 
-	/* We may be recursing from the signal callback of another i915 fence */
-	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
+	engine->emit_fini_breadcrumb(request,
+				     request->ring->vaddr + request->postfix);
+
+	trace_i915_request_execute(request);
+	engine->serial++;
+	result = true;
 
-	list_move_tail(&request->sched.link, &engine->active.requests);
+xfer:	/* We may be recursing from the signal callback of another i915 fence */
+	spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 
-	GEM_BUG_ON(test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
-	set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
+	if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags))
+		list_move_tail(&request->sched.link, &engine->active.requests);
 
 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
 	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) &&
@@ -429,12 +454,7 @@ void __i915_request_submit(struct i915_request *request)
 
 	spin_unlock(&request->lock);
 
-	engine->emit_fini_breadcrumb(request,
-				     request->ring->vaddr + request->postfix);
-
-	engine->serial++;
-
-	trace_i915_request_execute(request);
+	return result;
 }
 
 void i915_request_submit(struct i915_request *request)
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h
index 8ac6e1226a561..e4dd013761e81 100644
--- a/drivers/gpu/drm/i915/i915_request.h
+++ b/drivers/gpu/drm/i915/i915_request.h
@@ -292,7 +292,7 @@ int i915_request_await_execution(struct i915_request *rq,
 
 void i915_request_add(struct i915_request *rq);
 
-void __i915_request_submit(struct i915_request *request);
+bool __i915_request_submit(struct i915_request *request);
 void i915_request_submit(struct i915_request *request);
 
 void i915_request_skip(struct i915_request *request, int error);