drm/i915: Pipelined fencing [infrastructure]

With this change, every batchbuffer can use all available fences (save
pinned and scanout, of course) without ever stalling the gpu!

In theory. Currently the actual pipelined update of the register is
disabled due to some stability issues. However, just the deferred update
is a significant win.

Based on a series of patches by Daniel Vetter.

The premise is that before every access to a buffer through the GTT we
have to declare whether we need a register or not. If the access is by
the GPU, a pipelined update to the register is made via the ringbuffer,
and we track the last seqno of the batches that access it. If by the
CPU we wait for the last GPU access and update the register (either
to clear or to set it for the current buffer).

One advantage of being able to pipeline changes is that we can defer the
actual updating of the fence register until we first need to access the
object through the GTT, i.e. we can eliminate the stall on set_tiling.
This is important as the userspace bo cache does not track the tiling
status of active buffers which generate frequent stalls on gen3 when
enabling tiling for an already bound buffer.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

6 files changed, 274 insertions, 166 deletions

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 7b37c198cb1..af9ff40b135 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -126,6 +126,7 @@ struct drm_i915_master_private {
 struct drm_i915_fence_reg {
 	struct list_head lru_list;
 	struct drm_i915_gem_object *obj;
+	uint32_t setup_seqno;
 };
 
 struct sdvo_device_mapping {
@@ -752,6 +753,7 @@ struct drm_i915_gem_object {
 	 * Current tiling mode for the object.
 	 */
 	unsigned int tiling_mode : 2;
+	unsigned int tiling_changed : 1;
 
 	/** How many users have pinned this object in GTT space. The following
 	 * users can each hold at most one reference: pwrite/pread, pin_ioctl
@@ -1121,10 +1123,10 @@ i915_gem_next_request_seqno(struct drm_device *dev,
 	return ring->outstanding_lazy_request = dev_priv->next_seqno;
 }
 
-int __must_check i915_gem_object_get_fence_reg(struct drm_i915_gem_object *obj,
-					       bool interruptible);
-int __must_check i915_gem_object_put_fence_reg(struct drm_i915_gem_object *obj,
-					       bool interruptible);
+int __must_check i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
+					   struct intel_ring_buffer *pipelined,
+					   bool interruptible);
+int __must_check i915_gem_object_put_fence(struct drm_i915_gem_object *obj);
 
 void i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_reset(struct drm_device *dev);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index c3e6d7bda6e..23d2417a358 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -47,7 +47,8 @@ static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_obje
 static int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
 				       unsigned alignment,
 				       bool map_and_fenceable);
-static void i915_gem_clear_fence_reg(struct drm_i915_gem_object *obj);
+static void i915_gem_clear_fence_reg(struct drm_device *dev,
+				     struct drm_i915_fence_reg *reg);
 static int i915_gem_phys_pwrite(struct drm_device *dev,
 				struct drm_i915_gem_object *obj,
 				struct drm_i915_gem_pwrite *args,
@@ -684,7 +685,11 @@ i915_gem_gtt_pwrite_slow(struct drm_device *dev,
 		goto out_unpin_pages;
 	}
 
-	ret = i915_gem_object_set_to_gtt_domain(obj, 1);
+	ret = i915_gem_object_set_to_gtt_domain(obj, true);
+	if (ret)
+		goto out_unpin_pages;
+
+	ret = i915_gem_object_put_fence(obj);
 	if (ret)
 		goto out_unpin_pages;
 
@@ -966,14 +971,17 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 	 */
 	if (obj->phys_obj)
 		ret = i915_gem_phys_pwrite(dev, obj, args, file);
-	else if (obj->tiling_mode == I915_TILING_NONE &&
-		 obj->gtt_space &&
+	else if (obj->gtt_space &&
 		 obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 		ret = i915_gem_object_pin(obj, 0, true);
 		if (ret)
 			goto out;
 
-		ret = i915_gem_object_set_to_gtt_domain(obj, 1);
+		ret = i915_gem_object_set_to_gtt_domain(obj, true);
+		if (ret)
+			goto out_unpin;
+
+		ret = i915_gem_object_put_fence(obj);
 		if (ret)
 			goto out_unpin;
 
@@ -1205,12 +1213,12 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 	if (ret)
 		goto unlock;
 
-	/* Need a new fence register? */
-	if (obj->tiling_mode != I915_TILING_NONE) {
-		ret = i915_gem_object_get_fence_reg(obj, true);
-		if (ret)
-			goto unlock;
-	}
+	if (obj->tiling_mode == I915_TILING_NONE)
+		ret = i915_gem_object_put_fence(obj);
+	else
+		ret = i915_gem_object_get_fence(obj, NULL, true);
+	if (ret)
+		goto unlock;
 
 	if (i915_gem_object_is_inactive(obj))
 		list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
@@ -1608,7 +1616,6 @@ i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
 {
 	list_del_init(&obj->ring_list);
 	obj->last_rendering_seqno = 0;
-	obj->last_fenced_seqno = 0;
 }
 
 static void
@@ -1640,7 +1647,6 @@ i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
 
 	i915_gem_object_move_off_active(obj);
 	obj->fenced_gpu_access = false;
-	obj->last_fenced_ring = NULL;
 
 	obj->active = 0;
 	obj->pending_gpu_write = false;
@@ -1803,7 +1809,11 @@ static void i915_gem_reset_fences(struct drm_device *dev)
 		if (obj->tiling_mode)
 			i915_gem_release_mmap(obj);
 
-		i915_gem_clear_fence_reg(obj);
+		reg->obj->fence_reg = I915_FENCE_REG_NONE;
+		reg->obj->fenced_gpu_access = false;
+		reg->obj->last_fenced_seqno = 0;
+		reg->obj->last_fenced_ring = NULL;
+		i915_gem_clear_fence_reg(dev, reg);
 	}
 }
 
@@ -2114,8 +2124,9 @@ i915_gem_object_unbind(struct drm_i915_gem_object *obj)
 	}
 
 	/* release the fence reg _after_ flushing */
-	if (obj->fence_reg != I915_FENCE_REG_NONE)
-		i915_gem_clear_fence_reg(obj);
+	ret = i915_gem_object_put_fence(obj);
+	if (ret == -ERESTARTSYS)
+		return ret;
 
 	i915_gem_gtt_unbind_object(obj);
 	i915_gem_object_put_pages_gtt(obj);
@@ -2357,59 +2368,118 @@ static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
 	return 0;
 }
 
-static int i915_find_fence_reg(struct drm_device *dev,
-			       bool interruptible)
+static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
+{
+	return i915_seqno_passed(ring->get_seqno(ring), seqno);
+}
+
+static int
+i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
+			    struct intel_ring_buffer *pipelined,
+			    bool interruptible)
+{
+	int ret;
+
+	if (obj->fenced_gpu_access) {
+		if (obj->base.write_domain & I915_GEM_GPU_DOMAINS)
+			i915_gem_flush_ring(obj->base.dev,
+					    obj->last_fenced_ring,
+					    0, obj->base.write_domain);
+
+		obj->fenced_gpu_access = false;
+	}
+
+	if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
+		if (!ring_passed_seqno(obj->last_fenced_ring,
+				       obj->last_fenced_seqno)) {
+			ret = i915_do_wait_request(obj->base.dev,
+						   obj->last_fenced_seqno,
+						   interruptible,
+						   obj->last_fenced_ring);
+			if (ret)
+				return ret;
+		}
+
+		obj->last_fenced_seqno = 0;
+		obj->last_fenced_ring = NULL;
+	}
+
+	return 0;
+}
+
+int
+i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
+{
+	int ret;
+
+	if (obj->tiling_mode)
+		i915_gem_release_mmap(obj);
+
+	ret = i915_gem_object_flush_fence(obj, NULL, true);
+	if (ret)
+		return ret;
+
+	if (obj->fence_reg != I915_FENCE_REG_NONE) {
+		struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+		i915_gem_clear_fence_reg(obj->base.dev,
+					 &dev_priv->fence_regs[obj->fence_reg]);
+
+		obj->fence_reg = I915_FENCE_REG_NONE;
+	}
+
+	return 0;
+}
+
+static struct drm_i915_fence_reg *
+i915_find_fence_reg(struct drm_device *dev,
+		    struct intel_ring_buffer *pipelined)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct drm_i915_fence_reg *reg;
-	struct drm_i915_gem_object *obj = NULL;
-	int i, avail, ret;
+	struct drm_i915_fence_reg *reg, *first, *avail;
+	int i;
 
 	/* First try to find a free reg */
-	avail = 0;
+	avail = NULL;
 	for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
 		reg = &dev_priv->fence_regs[i];
 		if (!reg->obj)
-			return i;
+			return reg;
 
 		if (!reg->obj->pin_count)
-			avail++;
+			avail = reg;
 	}
 
-	if (avail == 0)
-		return -ENOSPC;
+	if (avail == NULL)
+		return NULL;
 
 	/* None available, try to steal one or wait for a user to finish */
-	avail = I915_FENCE_REG_NONE;
-	list_for_each_entry(reg, &dev_priv->mm.fence_list,
-			    lru_list) {
-		obj = reg->obj;
-		if (obj->pin_count)
+	avail = first = NULL;
+	list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
+		if (reg->obj->pin_count)
 			continue;
 
-		/* found one! */
-		avail = obj->fence_reg;
-		break;
-	}
+		if (first == NULL)
+			first = reg;
 
-	BUG_ON(avail == I915_FENCE_REG_NONE);
+		if (!pipelined ||
+		    !reg->obj->last_fenced_ring ||
+		    reg->obj->last_fenced_ring == pipelined) {
+			avail = reg;
+			break;
+		}
+	}
 
-	/* We only have a reference on obj from the active list. put_fence_reg
-	 * might drop that one, causing a use-after-free in it. So hold a
-	 * private reference to obj like the other callers of put_fence_reg
-	 * (set_tiling ioctl) do. */
-	drm_gem_object_reference(&obj->base);
-	ret = i915_gem_object_put_fence_reg(obj, interruptible);
-	drm_gem_object_unreference(&obj->base);
-	if (ret != 0)
-		return ret;
+	if (avail == NULL)
+		avail = first;
 
 	return avail;
 }
 
 /**
- * i915_gem_object_get_fence_reg - set up a fence reg for an object
+ * i915_gem_object_get_fence - set up a fence reg for an object
  * @obj: object to map through a fence reg
+ * @pipelined: ring on which to queue the change, or NULL for CPU access
+ * @interruptible: must we wait uninterruptibly for the register to retire?
  *
  * When mapping objects through the GTT, userspace wants to be able to write
  * to them without having to worry about swizzling if the object is tiled.
@@ -2421,52 +2491,119 @@ static int i915_find_fence_reg(struct drm_device *dev,
  * and tiling format.
  */
 int
-i915_gem_object_get_fence_reg(struct drm_i915_gem_object *obj,
-			      bool interruptible)
+i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
+			  struct intel_ring_buffer *pipelined,
+			  bool interruptible)
 {
 	struct drm_device *dev = obj->base.dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct drm_i915_fence_reg *reg = NULL;
-	struct intel_ring_buffer *pipelined = NULL;
+	struct drm_i915_fence_reg *reg;
 	int ret;
 
-	/* Just update our place in the LRU if our fence is getting used. */
+	/* Just update our place in the LRU if our fence is getting reused. */
 	if (obj->fence_reg != I915_FENCE_REG_NONE) {
 		reg = &dev_priv->fence_regs[obj->fence_reg];
 		list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
+
+		if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
+			pipelined = NULL;
+
+		if (!pipelined) {
+			if (reg->setup_seqno) {
+				if (!ring_passed_seqno(obj->last_fenced_ring,
+						       reg->setup_seqno)) {
+					ret = i915_do_wait_request(obj->base.dev,
+								   reg->setup_seqno,
+								   interruptible,
+								   obj->last_fenced_ring);
+					if (ret)
+						return ret;
+				}
+
+				reg->setup_seqno = 0;
+			}
+		} else if (obj->last_fenced_ring &&
+			   obj->last_fenced_ring != pipelined) {
+			ret = i915_gem_object_flush_fence(obj,
+							  pipelined,
+							  interruptible);
+			if (ret)
+				return ret;
+		} else if (obj->tiling_changed) {
+			if (obj->fenced_gpu_access) {
+				if (obj->base.write_domain & I915_GEM_GPU_DOMAINS)
+					i915_gem_flush_ring(obj->base.dev, obj->ring,
+							    0, obj->base.write_domain);
+
+				obj->fenced_gpu_access = false;
+			}
+		}
+
+		if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
+			pipelined = NULL;
+		BUG_ON(!pipelined && reg->setup_seqno);
+
+		if (obj->tiling_changed) {
+			if (pipelined) {
+				reg->setup_seqno =
+					i915_gem_next_request_seqno(dev, pipelined);
+				obj->last_fenced_seqno = reg->setup_seqno;
+				obj->last_fenced_ring = pipelined;
+			}
+			goto update;
+		}
+
 		return 0;
 	}
 
-	switch (obj->tiling_mode) {
-	case I915_TILING_NONE:
-		WARN(1, "allocating a fence for non-tiled object?\n");
-		break;
-	case I915_TILING_X:
-		if (!obj->stride)
-			return -EINVAL;
-		WARN((obj->stride & (512 - 1)),
-		     "object 0x%08x is X tiled but has non-512B pitch\n",
-		     obj->gtt_offset);
-		break;
-	case I915_TILING_Y:
-		if (!obj->stride)
-			return -EINVAL;
-		WARN((obj->stride & (128 - 1)),
-		     "object 0x%08x is Y tiled but has non-128B pitch\n",
-		     obj->gtt_offset);
-		break;
-	}
+	reg = i915_find_fence_reg(dev, pipelined);
+	if (reg == NULL)
+		return -ENOSPC;
 
-	ret = i915_find_fence_reg(dev, interruptible);
-	if (ret < 0)
+	ret = i915_gem_object_flush_fence(obj, pipelined, interruptible);
+	if (ret)
 		return ret;
 
-	obj->fence_reg = ret;
-	reg = &dev_priv->fence_regs[obj->fence_reg];
-	list_add_tail(&reg->lru_list, &dev_priv->mm.fence_list);
+	if (reg->obj) {
+		struct drm_i915_gem_object *old = reg->obj;
+
+		drm_gem_object_reference(&old->base);
+
+		if (old->tiling_mode)
+			i915_gem_release_mmap(old);
+
+		/* XXX The pipelined change over appears to be incoherent. */
+		ret = i915_gem_object_flush_fence(old,
+						  NULL, //pipelined,
+						  interruptible);
+		if (ret) {
+			drm_gem_object_unreference(&old->base);
+			return ret;
+		}
+
+		if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
+			pipelined = NULL;
+
+		old->fence_reg = I915_FENCE_REG_NONE;
+		old->last_fenced_ring = pipelined;
+		old->last_fenced_seqno =
+			pipelined ? i915_gem_next_request_seqno(dev, pipelined) : 0;
+
+		drm_gem_object_unreference(&old->base);
+	} else if (obj->last_fenced_seqno == 0)
+		pipelined = NULL;
 
 	reg->obj = obj;
+	list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
+	obj->fence_reg = reg - dev_priv->fence_regs;
+	obj->last_fenced_ring = pipelined;
 
+	reg->setup_seqno =
+		pipelined ? i915_gem_next_request_seqno(dev, pipelined) : 0;
+	obj->last_fenced_seqno = reg->setup_seqno;
+
+update:
+	obj->tiling_changed = false;
 	switch (INTEL_INFO(dev)->gen) {
 	case 6:
 		ret = sandybridge_write_fence_reg(obj, pipelined);
@@ -2497,87 +2634,34 @@ i915_gem_object_get_fence_reg(struct drm_i915_gem_object *obj,
  * data structures in dev_priv and obj.
  */
 static void
-i915_gem_clear_fence_reg(struct drm_i915_gem_object *obj)
+i915_gem_clear_fence_reg(struct drm_device *dev,
+			 struct drm_i915_fence_reg *reg)
 {
-	struct drm_device *dev = obj->base.dev;
 	drm_i915_private_t *dev_priv = dev->dev_private;
-	struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[obj->fence_reg];
-	uint32_t fence_reg;
+	uint32_t fence_reg = reg - dev_priv->fence_regs;
 
 	switch (INTEL_INFO(dev)->gen) {
 	case 6:
-		I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 +
-			     (obj->fence_reg * 8), 0);
+		I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
 		break;
 	case 5:
 	case 4:
-		I915_WRITE64(FENCE_REG_965_0 + (obj->fence_reg * 8), 0);
+		I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
 		break;
 	case 3:
-		if (obj->fence_reg >= 8)
-			fence_reg = FENCE_REG_945_8 + (obj->fence_reg - 8) * 4;
+		if (fence_reg >= 8)
+			fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
 		else
 	case 2:
-			fence_reg = FENCE_REG_830_0 + obj->fence_reg * 4;
+			fence_reg = FENCE_REG_830_0 + fence_reg * 4;
 
 		I915_WRITE(fence_reg, 0);
 		break;
 	}
 
-	reg->obj = NULL;
-	obj->fence_reg = I915_FENCE_REG_NONE;
 	list_del_init(&reg->lru_list);
-}
-
-/**
- * i915_gem_object_put_fence_reg - waits on outstanding fenced access
- * to the buffer to finish, and then resets the fence register.
- * @obj: tiled object holding a fence register.
- * @bool: whether the wait upon the fence is interruptible
- *
- * Zeroes out the fence register itself and clears out the associated
- * data structures in dev_priv and obj.
- */
-int
-i915_gem_object_put_fence_reg(struct drm_i915_gem_object *obj,
-			      bool interruptible)
-{
-	struct drm_device *dev = obj->base.dev;
-	int ret;
-
-	if (obj->fence_reg == I915_FENCE_REG_NONE)
-		return 0;
-
-	/* If we've changed tiling, GTT-mappings of the object
-	 * need to re-fault to ensure that the correct fence register
-	 * setup is in place.
-	 */
-	i915_gem_release_mmap(obj);
-
-	/* On the i915, GPU access to tiled buffers is via a fence,
-	 * therefore we must wait for any outstanding access to complete
-	 * before clearing the fence.
-	 */
-	if (obj->fenced_gpu_access) {
-		i915_gem_object_flush_gpu_write_domain(obj);
-		obj->fenced_gpu_access = false;
-	}
-
-	if (obj->last_fenced_seqno) {
-		ret = i915_do_wait_request(dev,
-					   obj->last_fenced_seqno,
-					   interruptible,
-					   obj->last_fenced_ring);
-		if (ret)
-			return ret;
-
-		obj->last_fenced_seqno = false;
-	}
-
-	i915_gem_object_flush_gtt_write_domain(obj);
-	i915_gem_clear_fence_reg(obj);
-
-	return 0;
+	reg->obj = NULL;
+	reg->setup_seqno = 0;
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index af01a58a643..9bdc495e17b 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -424,7 +424,7 @@ i915_gem_execbuffer_relocate(struct drm_device *dev,
 }
 
 static int
-i915_gem_execbuffer_reserve(struct drm_device *dev,
+i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
 			    struct drm_file *file,
 			    struct list_head *objects,
 			    struct drm_i915_gem_exec_object2 *exec)
@@ -499,10 +499,15 @@ i915_gem_execbuffer_reserve(struct drm_device *dev,
 			}
 
 			if (need_fence) {
-				ret = i915_gem_object_get_fence_reg(obj, true);
+				ret = i915_gem_object_get_fence(obj, ring, 1);
+				if (ret)
+					break;
+			} else if (entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
+				   obj->tiling_mode == I915_TILING_NONE) {
+				/* XXX pipelined! */
+				ret = i915_gem_object_put_fence(obj);
 				if (ret)
 					break;
-
 			}
 			obj->pending_fenced_gpu_access = need_fence;
 
@@ -522,7 +527,7 @@ i915_gem_execbuffer_reserve(struct drm_device *dev,
 		/* First attempt, just clear anything that is purgeable.
 		 * Second attempt, clear the entire GTT.
 		 */
-		ret = i915_gem_evict_everything(dev, retry == 0);
+		ret = i915_gem_evict_everything(ring->dev, retry == 0);
 		if (ret)
 			return ret;
 
@@ -548,6 +553,7 @@ err:
 static int
 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 				  struct drm_file *file,
+				  struct intel_ring_buffer *ring,
 				  struct list_head *objects,
 				  struct drm_i915_gem_exec_object2 *exec,
 				  int count)
@@ -590,7 +596,7 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 		goto err;
 	}
 
-	ret = i915_gem_execbuffer_reserve(dev, file, objects, exec);
+	ret = i915_gem_execbuffer_reserve(ring, file, objects, exec);
 	if (ret)
 		goto err;
 
@@ -930,7 +936,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	}
 
 	/* Move the objects en-masse into the GTT, evicting if necessary. */
-	ret = i915_gem_execbuffer_reserve(dev, file, &objects, exec);
+	ret = i915_gem_execbuffer_reserve(ring, file, &objects, exec);
 	if (ret)
 		goto err;
 
@@ -938,7 +944,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	ret = i915_gem_execbuffer_relocate(dev, file, &objects, exec);
 	if (ret) {
 		if (ret == -EFAULT) {
-			ret = i915_gem_execbuffer_relocate_slow(dev, file,
+			ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
 								&objects, exec,
 								args->buffer_count);
 			BUG_ON(!mutex_is_locked(&dev->struct_mutex));
diff --git a/drivers/gpu/drm/i915/i915_gem_tiling.c b/drivers/gpu/drm/i915/i915_gem_tiling.c
index 1c5fdb30f27..22a32b9932c 100644
--- a/drivers/gpu/drm/i915/i915_gem_tiling.c
+++ b/drivers/gpu/drm/i915/i915_gem_tiling.c
@@ -244,9 +244,6 @@ i915_gem_object_fence_ok(struct drm_i915_gem_object *obj, int tiling_mode)
 	if (INTEL_INFO(obj->base.dev)->gen >= 4)
 		return true;
 
-	if (!obj->gtt_space)
-		return true;
-
 	if (INTEL_INFO(obj->base.dev)->gen == 3) {
 		if (obj->gtt_offset & ~I915_FENCE_START_MASK)
 			return false;
@@ -345,27 +342,21 @@ i915_gem_set_tiling(struct drm_device *dev, void *data,
 		 * tiling mode. Otherwise we can just leave it alone, but
 		 * need to ensure that any fence register is cleared.
 		 */
-		if (!i915_gem_object_fence_ok(obj, args->tiling_mode))
-			ret = i915_gem_object_unbind(obj);
-		else if (obj->fence_reg != I915_FENCE_REG_NONE)
-			ret = i915_gem_object_put_fence_reg(obj, true);
-		else
-			i915_gem_release_mmap(obj);
+		i915_gem_release_mmap(obj);
 
-		if (ret != 0) {
-			args->tiling_mode = obj->tiling_mode;
-			args->stride = obj->stride;
-			goto err;
-		}
+		obj->map_and_fenceable =
+			obj->gtt_space == NULL ||
+			(obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end &&
+			 i915_gem_object_fence_ok(obj, args->tiling_mode));
 
+		obj->tiling_changed = true;
 		obj->tiling_mode = args->tiling_mode;
 		obj->stride = args->stride;
 	}
-err:
 	drm_gem_object_unreference(&obj->base);
 	mutex_unlock(&dev->struct_mutex);
 
-	return ret;
+	return 0;
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index c2c94a26f92..e141dd2e46e 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -1474,7 +1474,7 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev,
 	 * a fence as the cost is not that onerous.
 	 */
 	if (obj->tiling_mode != I915_TILING_NONE) {
-		ret = i915_gem_object_get_fence_reg(obj, false);
+		ret = i915_gem_object_get_fence(obj, pipelined, false);
 		if (ret)
 			goto err_unpin;
 	}
@@ -4370,6 +4370,12 @@ static int intel_crtc_cursor_set(struct drm_crtc *crtc,
 	/* we only need to pin inside GTT if cursor is non-phy */
 	mutex_lock(&dev->struct_mutex);
 	if (!dev_priv->info->cursor_needs_physical) {
+		if (obj->tiling_mode) {
+			DRM_ERROR("cursor cannot be tiled\n");
+			ret = -EINVAL;
+			goto fail_locked;
+		}
+
 		ret = i915_gem_object_pin(obj, PAGE_SIZE, true);
 		if (ret) {
 			DRM_ERROR("failed to pin cursor bo\n");
@@ -4382,6 +4388,12 @@ static int intel_crtc_cursor_set(struct drm_crtc *crtc,
 			goto fail_unpin;
 		}
 
+		ret = i915_gem_object_put_fence(obj);
+		if (ret) {
+			DRM_ERROR("failed to move cursor bo into the GTT\n");
+			goto fail_unpin;
+		}
+
 		addr = obj->gtt_offset;
 	} else {
 		int align = IS_I830(dev) ? 16 * 1024 : 256;
@@ -4966,6 +4978,7 @@ static void intel_unpin_work_fn(struct work_struct *__work)
 	i915_gem_object_unpin(work->old_fb_obj);
 	drm_gem_object_unreference(&work->pending_flip_obj->base);
 	drm_gem_object_unreference(&work->old_fb_obj->base);
+
 	mutex_unlock(&work->dev->struct_mutex);
 	kfree(work);
 }
@@ -5009,10 +5022,12 @@ static void do_intel_finish_page_flip(struct drm_device *dev,
 	spin_unlock_irqrestore(&dev->event_lock, flags);
 
 	obj = work->old_fb_obj;
+
 	atomic_clear_mask(1 << intel_crtc->plane,
 			  &obj->pending_flip.counter);
 	if (atomic_read(&obj->pending_flip) == 0)
 		wake_up(&dev_priv->pending_flip_queue);
+
 	schedule_work(&work->work);
 
 	trace_i915_flip_complete(intel_crtc->plane, work->pending_flip_obj);
diff --git a/drivers/gpu/drm/i915/intel_overlay.c b/drivers/gpu/drm/i915/intel_overlay.c
index af715cc03ee..d0c1add393a 100644
--- a/drivers/gpu/drm/i915/intel_overlay.c
+++ b/drivers/gpu/drm/i915/intel_overlay.c
@@ -787,6 +787,10 @@ static int intel_overlay_do_put_image(struct intel_overlay *overlay,
 	if (ret != 0)
 		goto out_unpin;
 
+	ret = i915_gem_object_put_fence(new_bo);
+	if (ret)
+		goto out_unpin;
+
 	if (!overlay->active) {
 		regs = intel_overlay_map_regs(overlay);
 		if (!regs) {
@@ -1161,6 +1165,12 @@ int intel_overlay_put_image(struct drm_device *dev, void *data,
 	mutex_lock(&dev->mode_config.mutex);
 	mutex_lock(&dev->struct_mutex);
 
+	if (new_bo->tiling_mode) {
+		DRM_ERROR("buffer used for overlay image can not be tiled\n");
+		ret = -EINVAL;
+		goto out_unlock;
+	}
+
 	ret = intel_overlay_recover_from_interrupt(overlay, true);
 	if (ret != 0)
 		goto out_unlock;