drm/i915: Implement inter-engine read-read optimisations

Currently, we only track the last request globally across all engines.
This prevents us from issuing concurrent read requests on e.g. the RCS
and BCS engines (or more likely the render and media engines). Without
semaphores, we incur costly stalls as we synchronise between rings -
greatly impacting the current performance of Broadwell versus Haswell in
certain workloads (like video decode). With the introduction of
reference counted requests, it is much easier to track the last request
per ring, as well as the last global write request so that we can
optimise inter-engine read read requests (as well as better optimise
certain CPU waits).

v2: Fix inverted readonly condition for nonblocking waits.
v3: Handle non-continguous engine array after waits
v4: Rebase, tidy, rewrite ring list debugging
v5: Use obj->active as a bitfield, it looks cool
v6: Micro-optimise, mostly involving moving code around
v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf)
v8: Rebase
v9: Refactor i915_gem_object_sync() to allow the compiler to better
optimise it.

Benchmark: igt/gem_read_read_speed
hsw:gt3e (with semaphores):
Before: Time to read-read 1024k:		275.794µs
After:  Time to read-read 1024k:		123.260µs

hsw:gt3e (w/o semaphores):
Before: Time to read-read 1024k:		230.433µs
After:  Time to read-read 1024k:		124.593µs

bdw-u (w/o semaphores):             Before          After
Time to read-read 1x1:            26.274µs       10.350µs
Time to read-read 128x128:        40.097µs       21.366µs
Time to read-read 256x256:        77.087µs       42.608µs
Time to read-read 512x512:       281.999µs      181.155µs
Time to read-read 1024x1024:    1196.141µs     1118.223µs
Time to read-read 2048x2048:    5639.072µs     5225.837µs
Time to read-read 4096x4096:   22401.662µs    21137.067µs
Time to read-read 8192x8192:   89617.735µs    85637.681µs

Testcase: igt/gem_concurrent_blit (read-read and friends)
Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com>
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> [v8]
[danvet: s/\<rq\>/req/g]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

1 files changed, 331 insertions, 209 deletions

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index fa4429144cb9..76ab4f18d618 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -38,11 +38,14 @@
 #include <linux/pci.h>
 #include <linux/dma-buf.h>
 
+#define RQ_BUG_ON(expr)
+
 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
 static void
-i915_gem_object_retire(struct drm_i915_gem_object *obj);
-
+i915_gem_object_retire__write(struct drm_i915_gem_object *obj);
+static void
+i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring);
 static void i915_gem_write_fence(struct drm_device *dev, int reg,
 				 struct drm_i915_gem_object *obj);
 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
@@ -515,8 +518,6 @@ int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
 		ret = i915_gem_object_wait_rendering(obj, true);
 		if (ret)
 			return ret;
-
-		i915_gem_object_retire(obj);
 	}
 
 	ret = i915_gem_object_get_pages(obj);
@@ -936,8 +937,6 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
 		ret = i915_gem_object_wait_rendering(obj, false);
 		if (ret)
 			return ret;
-
-		i915_gem_object_retire(obj);
 	}
 	/* Same trick applies to invalidate partially written cachelines read
 	 * before writing. */
@@ -1236,6 +1235,9 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 
 	WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
 
+	if (list_empty(&req->list))
+		return 0;
+
 	if (i915_gem_request_completed(req, true))
 		return 0;
 
@@ -1335,6 +1337,63 @@ out:
 	return ret;
 }
 
+static inline void
+i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
+{
+	struct drm_i915_file_private *file_priv = request->file_priv;
+
+	if (!file_priv)
+		return;
+
+	spin_lock(&file_priv->mm.lock);
+	list_del(&request->client_list);
+	request->file_priv = NULL;
+	spin_unlock(&file_priv->mm.lock);
+}
+
+static void i915_gem_request_retire(struct drm_i915_gem_request *request)
+{
+	trace_i915_gem_request_retire(request);
+
+	/* We know the GPU must have read the request to have
+	 * sent us the seqno + interrupt, so use the position
+	 * of tail of the request to update the last known position
+	 * of the GPU head.
+	 *
+	 * Note this requires that we are always called in request
+	 * completion order.
+	 */
+	request->ringbuf->last_retired_head = request->postfix;
+
+	list_del_init(&request->list);
+	i915_gem_request_remove_from_client(request);
+
+	put_pid(request->pid);
+
+	i915_gem_request_unreference(request);
+}
+
+static void
+__i915_gem_request_retire__upto(struct drm_i915_gem_request *req)
+{
+	struct intel_engine_cs *engine = req->ring;
+	struct drm_i915_gem_request *tmp;
+
+	lockdep_assert_held(&engine->dev->struct_mutex);
+
+	if (list_empty(&req->list))
+		return;
+
+	do {
+		tmp = list_first_entry(&engine->request_list,
+				       typeof(*tmp), list);
+
+		i915_gem_request_retire(tmp);
+	} while (tmp != req);
+
+	WARN_ON(i915_verify_lists(engine->dev));
+}
+
 /**
  * Waits for a request to be signaled, and cleans up the
  * request and object lists appropriately for that event.
@@ -1345,7 +1404,6 @@ i915_wait_request(struct drm_i915_gem_request *req)
 	struct drm_device *dev;
 	struct drm_i915_private *dev_priv;
 	bool interruptible;
-	unsigned reset_counter;
 	int ret;
 
 	BUG_ON(req == NULL);
@@ -1364,29 +1422,13 @@ i915_wait_request(struct drm_i915_gem_request *req)
 	if (ret)
 		return ret;
 
-	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
-	i915_gem_request_reference(req);
-	ret = __i915_wait_request(req, reset_counter,
+	ret = __i915_wait_request(req,
+				  atomic_read(&dev_priv->gpu_error.reset_counter),
 				  interruptible, NULL, NULL);
-	i915_gem_request_unreference(req);
-	return ret;
-}
-
-static int
-i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj)
-{
-	if (!obj->active)
-		return 0;
-
-	/* Manually manage the write flush as we may have not yet
-	 * retired the buffer.
-	 *
-	 * Note that the last_write_req is always the earlier of
-	 * the two (read/write) requests, so if we haved successfully waited,
-	 * we know we have passed the last write.
-	 */
-	i915_gem_request_assign(&obj->last_write_req, NULL);
+	if (ret)
+		return ret;
 
+	__i915_gem_request_retire__upto(req);
 	return 0;
 }
 
@@ -1398,18 +1440,52 @@ int
 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
 			       bool readonly)
 {
-	struct drm_i915_gem_request *req;
-	int ret;
+	int ret, i;
 
-	req = readonly ? obj->last_write_req : obj->last_read_req;
-	if (!req)
+	if (!obj->active)
 		return 0;
 
-	ret = i915_wait_request(req);
-	if (ret)
-		return ret;
+	if (readonly) {
+		if (obj->last_write_req != NULL) {
+			ret = i915_wait_request(obj->last_write_req);
+			if (ret)
+				return ret;
+
+			i = obj->last_write_req->ring->id;
+			if (obj->last_read_req[i] == obj->last_write_req)
+				i915_gem_object_retire__read(obj, i);
+			else
+				i915_gem_object_retire__write(obj);
+		}
+	} else {
+		for (i = 0; i < I915_NUM_RINGS; i++) {
+			if (obj->last_read_req[i] == NULL)
+				continue;
+
+			ret = i915_wait_request(obj->last_read_req[i]);
+			if (ret)
+				return ret;
+
+			i915_gem_object_retire__read(obj, i);
+		}
+		RQ_BUG_ON(obj->active);
+	}
+
+	return 0;
+}
+
+static void
+i915_gem_object_retire_request(struct drm_i915_gem_object *obj,
+			       struct drm_i915_gem_request *req)
+{
+	int ring = req->ring->id;
+
+	if (obj->last_read_req[ring] == req)
+		i915_gem_object_retire__read(obj, ring);
+	else if (obj->last_write_req == req)
+		i915_gem_object_retire__write(obj);
 
-	return i915_gem_object_wait_rendering__tail(obj);
+	__i915_gem_request_retire__upto(req);
 }
 
 /* A nonblocking variant of the above wait. This is a highly dangerous routine
@@ -1420,37 +1496,66 @@ i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
 					    struct drm_i915_file_private *file_priv,
 					    bool readonly)
 {
-	struct drm_i915_gem_request *req;
 	struct drm_device *dev = obj->base.dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct drm_i915_gem_request *requests[I915_NUM_RINGS];
 	unsigned reset_counter;
-	int ret;
+	int ret, i, n = 0;
 
 	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
 	BUG_ON(!dev_priv->mm.interruptible);
 
-	req = readonly ? obj->last_write_req : obj->last_read_req;
-	if (!req)
+	if (!obj->active)
 		return 0;
 
 	ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
 	if (ret)
 		return ret;
 
-	ret = i915_gem_check_olr(req);
-	if (ret)
-		return ret;
-
 	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
-	i915_gem_request_reference(req);
+
+	if (readonly) {
+		struct drm_i915_gem_request *req;
+
+		req = obj->last_write_req;
+		if (req == NULL)
+			return 0;
+
+		ret = i915_gem_check_olr(req);
+		if (ret)
+			goto err;
+
+		requests[n++] = i915_gem_request_reference(req);
+	} else {
+		for (i = 0; i < I915_NUM_RINGS; i++) {
+			struct drm_i915_gem_request *req;
+
+			req = obj->last_read_req[i];
+			if (req == NULL)
+				continue;
+
+			ret = i915_gem_check_olr(req);
+			if (ret)
+				goto err;
+
+			requests[n++] = i915_gem_request_reference(req);
+		}
+	}
+
 	mutex_unlock(&dev->struct_mutex);
-	ret = __i915_wait_request(req, reset_counter, true, NULL, file_priv);
+	for (i = 0; ret == 0 && i < n; i++)
+		ret = __i915_wait_request(requests[i], reset_counter, true,
+					  NULL, file_priv);
 	mutex_lock(&dev->struct_mutex);
-	i915_gem_request_unreference(req);
-	if (ret)
-		return ret;
 
-	return i915_gem_object_wait_rendering__tail(obj);
+err:
+	for (i = 0; i < n; i++) {
+		if (ret == 0)
+			i915_gem_object_retire_request(obj, requests[i]);
+		i915_gem_request_unreference(requests[i]);
+	}
+
+	return ret;
 }
 
 /**
@@ -2231,78 +2336,58 @@ i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
 	return 0;
 }
 
-static void
-i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
-			       struct intel_engine_cs *ring)
+void i915_vma_move_to_active(struct i915_vma *vma,
+			     struct intel_engine_cs *ring)
 {
-	struct drm_i915_gem_request *req;
-	struct intel_engine_cs *old_ring;
-
-	BUG_ON(ring == NULL);
-
-	req = intel_ring_get_request(ring);
-	old_ring = i915_gem_request_get_ring(obj->last_read_req);
-
-	if (old_ring != ring && obj->last_write_req) {
-		/* Keep the request relative to the current ring */
-		i915_gem_request_assign(&obj->last_write_req, req);
-	}
+	struct drm_i915_gem_object *obj = vma->obj;
 
 	/* Add a reference if we're newly entering the active list. */
-	if (!obj->active) {
+	if (obj->active == 0)
 		drm_gem_object_reference(&obj->base);
-		obj->active = 1;
-	}
+	obj->active |= intel_ring_flag(ring);
 
-	list_move_tail(&obj->ring_list, &ring->active_list);
+	list_move_tail(&obj->ring_list[ring->id], &ring->active_list);
+	i915_gem_request_assign(&obj->last_read_req[ring->id],
+				intel_ring_get_request(ring));
 
-	i915_gem_request_assign(&obj->last_read_req, req);
+	list_move_tail(&vma->mm_list, &vma->vm->active_list);
 }
 
-void i915_vma_move_to_active(struct i915_vma *vma,
-			     struct intel_engine_cs *ring)
+static void
+i915_gem_object_retire__write(struct drm_i915_gem_object *obj)
 {
-	list_move_tail(&vma->mm_list, &vma->vm->active_list);
-	return i915_gem_object_move_to_active(vma->obj, ring);
+	RQ_BUG_ON(obj->last_write_req == NULL);
+	RQ_BUG_ON(!(obj->active & intel_ring_flag(obj->last_write_req->ring)));
+
+	i915_gem_request_assign(&obj->last_write_req, NULL);
+	intel_fb_obj_flush(obj, true);
 }
 
 static void
-i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
+i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
 {
 	struct i915_vma *vma;
 
-	BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
-	BUG_ON(!obj->active);
+	RQ_BUG_ON(obj->last_read_req[ring] == NULL);
+	RQ_BUG_ON(!(obj->active & (1 << ring)));
+
+	list_del_init(&obj->ring_list[ring]);
+	i915_gem_request_assign(&obj->last_read_req[ring], NULL);
+
+	if (obj->last_write_req && obj->last_write_req->ring->id == ring)
+		i915_gem_object_retire__write(obj);
+
+	obj->active &= ~(1 << ring);
+	if (obj->active)
+		return;
 
 	list_for_each_entry(vma, &obj->vma_list, vma_link) {
 		if (!list_empty(&vma->mm_list))
 			list_move_tail(&vma->mm_list, &vma->vm->inactive_list);
 	}
 
-	intel_fb_obj_flush(obj, true);
-
-	list_del_init(&obj->ring_list);
-
-	i915_gem_request_assign(&obj->last_read_req, NULL);
-	i915_gem_request_assign(&obj->last_write_req, NULL);
-	obj->base.write_domain = 0;
-
 	i915_gem_request_assign(&obj->last_fenced_req, NULL);
-
-	obj->active = 0;
 	drm_gem_object_unreference(&obj->base);
-
-	WARN_ON(i915_verify_lists(dev));
-}
-
-static void
-i915_gem_object_retire(struct drm_i915_gem_object *obj)
-{
-	if (obj->last_read_req == NULL)
-		return;
-
-	if (i915_gem_request_completed(obj->last_read_req, true))
-		i915_gem_object_move_to_inactive(obj);
 }
 
 static int
@@ -2479,20 +2564,6 @@ int __i915_add_request(struct intel_engine_cs *ring,
 	return 0;
 }
 
-static inline void
-i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
-{
-	struct drm_i915_file_private *file_priv = request->file_priv;
-
-	if (!file_priv)
-		return;
-
-	spin_lock(&file_priv->mm.lock);
-	list_del(&request->client_list);
-	request->file_priv = NULL;
-	spin_unlock(&file_priv->mm.lock);
-}
-
 static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
 				   const struct intel_context *ctx)
 {
@@ -2538,16 +2609,6 @@ static void i915_set_reset_status(struct drm_i915_private *dev_priv,
 	}
 }
 
-static void i915_gem_free_request(struct drm_i915_gem_request *request)
-{
-	list_del(&request->list);
-	i915_gem_request_remove_from_client(request);
-
-	put_pid(request->pid);
-
-	i915_gem_request_unreference(request);
-}
-
 void i915_gem_request_free(struct kref *req_ref)
 {
 	struct drm_i915_gem_request *req = container_of(req_ref,
@@ -2648,9 +2709,9 @@ static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
 
 		obj = list_first_entry(&ring->active_list,
 				       struct drm_i915_gem_object,
-				       ring_list);
+				       ring_list[ring->id]);
 
-		i915_gem_object_move_to_inactive(obj);
+		i915_gem_object_retire__read(obj, ring->id);
 	}
 
 	/*
@@ -2686,7 +2747,7 @@ static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
 					   struct drm_i915_gem_request,
 					   list);
 
-		i915_gem_free_request(request);
+		i915_gem_request_retire(request);
 	}
 
 	/* This may not have been flushed before the reset, so clean it now */
@@ -2734,6 +2795,8 @@ void i915_gem_reset(struct drm_device *dev)
 	i915_gem_context_reset(dev);
 
 	i915_gem_restore_fences(dev);
+
+	WARN_ON(i915_verify_lists(dev));
 }
 
 /**
@@ -2742,11 +2805,11 @@ void i915_gem_reset(struct drm_device *dev)
 void
 i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
 {
-	if (list_empty(&ring->request_list))
-		return;
-
 	WARN_ON(i915_verify_lists(ring->dev));
 
+	if (list_empty(&ring->active_list))
+		return;
+
 	/* Retire requests first as we use it above for the early return.
 	 * If we retire requests last, we may use a later seqno and so clear
 	 * the requests lists without clearing the active list, leading to
@@ -2762,16 +2825,7 @@ i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
 		if (!i915_gem_request_completed(request, true))
 			break;
 
-		trace_i915_gem_request_retire(request);
-
-		/* We know the GPU must have read the request to have
-		 * sent us the seqno + interrupt, so use the position
-		 * of tail of the request to update the last known position
-		 * of the GPU head.
-		 */
-		request->ringbuf->last_retired_head = request->postfix;
-
-		i915_gem_free_request(request);
+		i915_gem_request_retire(request);
 	}
 
 	/* Move any buffers on the active list that are no longer referenced
@@ -2783,12 +2837,12 @@ i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
 
 		obj = list_first_entry(&ring->active_list,
 				      struct drm_i915_gem_object,
-				      ring_list);
+				      ring_list[ring->id]);
 
-		if (!i915_gem_request_completed(obj->last_read_req, true))
+		if (!list_empty(&obj->last_read_req[ring->id]->list))
 			break;
 
-		i915_gem_object_move_to_inactive(obj);
+		i915_gem_object_retire__read(obj, ring->id);
 	}
 
 	if (unlikely(ring->trace_irq_req &&
@@ -2883,17 +2937,30 @@ i915_gem_idle_work_handler(struct work_struct *work)
 static int
 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
 {
-	struct intel_engine_cs *ring;
-	int ret;
+	int ret, i;
+
+	if (!obj->active)
+		return 0;
 
-	if (obj->active) {
-		ring = i915_gem_request_get_ring(obj->last_read_req);
+	for (i = 0; i < I915_NUM_RINGS; i++) {
+		struct drm_i915_gem_request *req;
 
-		ret = i915_gem_check_olr(obj->last_read_req);
+		req = obj->last_read_req[i];
+		if (req == NULL)
+			continue;
+
+		if (list_empty(&req->list))
+			goto retire;
+
+		ret = i915_gem_check_olr(req);
 		if (ret)
 			return ret;
 
-		i915_gem_retire_requests_ring(ring);
+		if (i915_gem_request_completed(req, true)) {
+			__i915_gem_request_retire__upto(req);
+retire:
+			i915_gem_object_retire__read(obj, i);
+		}
 	}
 
 	return 0;
@@ -2927,9 +2994,10 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_wait *args = data;
 	struct drm_i915_gem_object *obj;
-	struct drm_i915_gem_request *req;
+	struct drm_i915_gem_request *req[I915_NUM_RINGS];
 	unsigned reset_counter;
-	int ret = 0;
+	int i, n = 0;
+	int ret;
 
 	if (args->flags != 0)
 		return -EINVAL;
@@ -2949,11 +3017,9 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	if (ret)
 		goto out;
 
-	if (!obj->active || !obj->last_read_req)
+	if (!obj->active)
 		goto out;
 
-	req = obj->last_read_req;
-
 	/* Do this after OLR check to make sure we make forward progress polling
 	 * on this IOCTL with a timeout == 0 (like busy ioctl)
 	 */
@@ -2964,13 +3030,23 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 
 	drm_gem_object_unreference(&obj->base);
 	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
-	i915_gem_request_reference(req);
+
+	for (i = 0; i < I915_NUM_RINGS; i++) {
+		if (obj->last_read_req[i] == NULL)
+			continue;
+
+		req[n++] = i915_gem_request_reference(obj->last_read_req[i]);
+	}
+
 	mutex_unlock(&dev->struct_mutex);
 
-	ret = __i915_wait_request(req, reset_counter, true,
-				  args->timeout_ns > 0 ? &args->timeout_ns : NULL,
-				  file->driver_priv);
-	i915_gem_request_unreference__unlocked(req);
+	for (i = 0; i < n; i++) {
+		if (ret == 0)
+			ret = __i915_wait_request(req[i], reset_counter, true,
+						  args->timeout_ns > 0 ? &args->timeout_ns : NULL,
+						  file->driver_priv);
+		i915_gem_request_unreference__unlocked(req[i]);
+	}
 	return ret;
 
 out:
@@ -2979,6 +3055,56 @@ out:
 	return ret;
 }
 
+static int
+__i915_gem_object_sync(struct drm_i915_gem_object *obj,
+		       struct intel_engine_cs *to,
+		       struct drm_i915_gem_request *req)
+{
+	struct intel_engine_cs *from;
+	int ret;
+
+	from = i915_gem_request_get_ring(req);
+	if (to == from)
+		return 0;
+
+	if (i915_gem_request_completed(req, true))
+		return 0;
+
+	ret = i915_gem_check_olr(req);
+	if (ret)
+		return ret;
+
+	if (!i915_semaphore_is_enabled(obj->base.dev)) {
+		ret = __i915_wait_request(req,
+					  atomic_read(&to_i915(obj->base.dev)->gpu_error.reset_counter),
+					  to_i915(obj->base.dev)->mm.interruptible, NULL, NULL);
+		if (ret)
+			return ret;
+
+		i915_gem_object_retire_request(obj, req);
+	} else {
+		int idx = intel_ring_sync_index(from, to);
+		u32 seqno = i915_gem_request_get_seqno(req);
+
+		if (seqno <= from->semaphore.sync_seqno[idx])
+			return 0;
+
+		trace_i915_gem_ring_sync_to(from, to, req);
+		ret = to->semaphore.sync_to(to, from, seqno);
+		if (ret)
+			return ret;
+
+		/* We use last_read_req because sync_to()
+		 * might have just caused seqno wrap under
+		 * the radar.
+		 */
+		from->semaphore.sync_seqno[idx] =
+			i915_gem_request_get_seqno(obj->last_read_req[from->id]);
+	}
+
+	return 0;
+}
+
 /**
  * i915_gem_object_sync - sync an object to a ring.
  *
@@ -2987,7 +3113,17 @@ out:
  *
  * This code is meant to abstract object synchronization with the GPU.
  * Calling with NULL implies synchronizing the object with the CPU
- * rather than a particular GPU ring.
+ * rather than a particular GPU ring. Conceptually we serialise writes
+ * between engines inside the GPU. We only allow on engine to write
+ * into a buffer at any time, but multiple readers. To ensure each has
+ * a coherent view of memory, we must:
+ *
+ * - If there is an outstanding write request to the object, the new
+ *   request must wait for it to complete (either CPU or in hw, requests
+ *   on the same ring will be naturally ordered).
+ *
+ * - If we are a write request (pending_write_domain is set), the new
+ *   request must wait for outstanding read requests to complete.
  *
  * Returns 0 if successful, else propagates up the lower layer error.
  */
@@ -2995,41 +3131,32 @@ int
 i915_gem_object_sync(struct drm_i915_gem_object *obj,
 		     struct intel_engine_cs *to)
 {
-	struct intel_engine_cs *from;
-	u32 seqno;
-	int ret, idx;
+	const bool readonly = obj->base.pending_write_domain == 0;
+	struct drm_i915_gem_request *req[I915_NUM_RINGS];
+	int ret, i, n;
 
-	from = i915_gem_request_get_ring(obj->last_read_req);
-
-	if (from == NULL || to == from)
-		return 0;
-
-	if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
-		return i915_gem_object_wait_rendering(obj, false);
-
-	idx = intel_ring_sync_index(from, to);
-
-	seqno = i915_gem_request_get_seqno(obj->last_read_req);
-	/* Optimization: Avoid semaphore sync when we are sure we already
-	 * waited for an object with higher seqno */
-	if (seqno <= from->semaphore.sync_seqno[idx])
+	if (!obj->active)
 		return 0;
 
-	ret = i915_gem_check_olr(obj->last_read_req);
-	if (ret)
-		return ret;
+	if (to == NULL)
+		return i915_gem_object_wait_rendering(obj, readonly);
 
-	trace_i915_gem_ring_sync_to(from, to, obj->last_read_req);
-	ret = to->semaphore.sync_to(to, from, seqno);
-	if (!ret)
-		/* We use last_read_req because sync_to()
-		 * might have just caused seqno wrap under
-		 * the radar.
-		 */
-		from->semaphore.sync_seqno[idx] =
-				i915_gem_request_get_seqno(obj->last_read_req);
+	n = 0;
+	if (readonly) {
+		if (obj->last_write_req)
+			req[n++] = obj->last_write_req;
+	} else {
+		for (i = 0; i < I915_NUM_RINGS; i++)
+			if (obj->last_read_req[i])
+				req[n++] = obj->last_read_req[i];
+	}
+	for (i = 0; i < n; i++) {
+		ret = __i915_gem_object_sync(obj, to, req[i]);
+		if (ret)
+			return ret;
+	}
 
-	return ret;
+	return 0;
 }
 
 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
@@ -3115,10 +3242,6 @@ int i915_vma_unbind(struct i915_vma *vma)
 	/* Since the unbound list is global, only move to that list if
 	 * no more VMAs exist. */
 	if (list_empty(&obj->vma_list)) {
-		/* Throw away the active reference before
-		 * moving to the unbound list. */
-		i915_gem_object_retire(obj);
-
 		i915_gem_gtt_finish_object(obj);
 		list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
 	}
@@ -3151,6 +3274,7 @@ int i915_gpu_idle(struct drm_device *dev)
 			return ret;
 	}
 
+	WARN_ON(i915_verify_lists(dev));
 	return 0;
 }
 
@@ -3773,8 +3897,6 @@ i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
 	if (ret)
 		return ret;
 
-	i915_gem_object_retire(obj);
-
 	/* Flush and acquire obj->pages so that we are coherent through
 	 * direct access in memory with previous cached writes through
 	 * shmemfs and that our cache domain tracking remains valid.
@@ -3972,11 +4094,9 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	u32 old_read_domains, old_write_domain;
 	int ret;
 
-	if (pipelined != i915_gem_request_get_ring(obj->last_read_req)) {
-		ret = i915_gem_object_sync(obj, pipelined);
-		if (ret)
-			return ret;
-	}
+	ret = i915_gem_object_sync(obj, pipelined);
+	if (ret)
+		return ret;
 
 	/* Mark the pin_display early so that we account for the
 	 * display coherency whilst setting up the cache domains.
@@ -4060,7 +4180,6 @@ i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
 	if (ret)
 		return ret;
 
-	i915_gem_object_retire(obj);
 	i915_gem_object_flush_gtt_write_domain(obj);
 
 	old_write_domain = obj->base.write_domain;
@@ -4349,15 +4468,15 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data,
 	 * necessary flushes here.
 	 */
 	ret = i915_gem_object_flush_active(obj);
+	if (ret)
+		goto unref;
 
-	args->busy = obj->active;
-	if (obj->last_read_req) {
-		struct intel_engine_cs *ring;
-		BUILD_BUG_ON(I915_NUM_RINGS > 16);
-		ring = i915_gem_request_get_ring(obj->last_read_req);
-		args->busy |= intel_ring_flag(ring) << 16;
-	}
+	BUILD_BUG_ON(I915_NUM_RINGS > 16);
+	args->busy = obj->active << 16;
+	if (obj->last_write_req)
+		args->busy |= obj->last_write_req->ring->id;
 
+unref:
 	drm_gem_object_unreference(&obj->base);
 unlock:
 	mutex_unlock(&dev->struct_mutex);
@@ -4431,8 +4550,11 @@ unlock:
 void i915_gem_object_init(struct drm_i915_gem_object *obj,
 			  const struct drm_i915_gem_object_ops *ops)
 {
+	int i;
+
 	INIT_LIST_HEAD(&obj->global_list);
-	INIT_LIST_HEAD(&obj->ring_list);
+	for (i = 0; i < I915_NUM_RINGS; i++)
+		INIT_LIST_HEAD(&obj->ring_list[i]);
 	INIT_LIST_HEAD(&obj->obj_exec_link);
 	INIT_LIST_HEAD(&obj->vma_list);
 	INIT_LIST_HEAD(&obj->batch_pool_link);