drm/i915: Push pipelining of display plane flushes to the caller

This ensures that we do wait upon the flushes to complete if necessary
and avoid the visual tears, whilst enabling pipelined page-flips.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drivers/gpu/drm/i915/i915_drv.h

@@ -1013,7 +1013,8 @@ void i915_gem_process_flushing_list(struct drm_device *dev,
 				    struct intel_ring_buffer *ring);
 int i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj,
 				      int write);
-int i915_gem_object_set_to_display_plane(struct drm_gem_object *obj);
+int i915_gem_object_set_to_display_plane(struct drm_gem_object *obj,
+					 bool pipelined);
 int i915_gem_attach_phys_object(struct drm_device *dev,
 				struct drm_gem_object *obj,
 				int id,

drivers/gpu/drm/i915/i915_gem.c

@@ -2597,6 +2597,7 @@ i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj,
 	/* Queue the GPU write cache flushing we need. */
 	old_write_domain = obj->write_domain;
 	i915_gem_flush(dev, 0, obj->write_domain);
+	BUG_ON(obj->write_domain);
 
 	trace_i915_gem_object_change_domain(obj,
 					    obj->read_domains,
@@ -2704,7 +2705,8 @@ i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
  * wait, as in modesetting process we're not supposed to be interrupted.
  */
 int
-i915_gem_object_set_to_display_plane(struct drm_gem_object *obj)
+i915_gem_object_set_to_display_plane(struct drm_gem_object *obj,
+				     bool pipelined)
 {
 	struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 	uint32_t old_read_domains;
@@ -2714,8 +2716,8 @@ i915_gem_object_set_to_display_plane(struct drm_gem_object *obj)
 	if (obj_priv->gtt_space == NULL)
 		return -EINVAL;
 
-	ret = i915_gem_object_flush_gpu_write_domain(obj, true);
-	if (ret != 0)
+	ret = i915_gem_object_flush_gpu_write_domain(obj, pipelined);
+	if (ret)
 		return ret;
 
 	i915_gem_object_flush_cpu_write_domain(obj);

drivers/gpu/drm/i915/intel_display.c

@@ -1417,7 +1417,9 @@ static void intel_update_fbc(struct drm_device *dev)
 }
 
 int
-intel_pin_and_fence_fb_obj(struct drm_device *dev, struct drm_gem_object *obj)
+intel_pin_and_fence_fb_obj(struct drm_device *dev,
+			   struct drm_gem_object *obj,
+			   bool pipelined)
 {
 	struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 	u32 alignment;
@@ -1445,14 +1447,12 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev, struct drm_gem_object *obj)
 	}
 
 	ret = i915_gem_object_pin(obj, alignment);
-	if (ret != 0)
+	if (ret)
 		return ret;
 
-	ret = i915_gem_object_set_to_display_plane(obj);
-	if (ret != 0) {
-		i915_gem_object_unpin(obj);
-		return ret;
-	}
+	ret = i915_gem_object_set_to_display_plane(obj, pipelined);
+	if (ret)
+		goto err_unpin;
 
 	/* Install a fence for tiled scan-out. Pre-i965 always needs a
 	 * fence, whereas 965+ only requires a fence if using
@@ -1462,13 +1462,15 @@ intel_pin_and_fence_fb_obj(struct drm_device *dev, struct drm_gem_object *obj)
 	if (obj_priv->fence_reg == I915_FENCE_REG_NONE &&
 	    obj_priv->tiling_mode != I915_TILING_NONE) {
 		ret = i915_gem_object_get_fence_reg(obj);
-		if (ret != 0) {
-			i915_gem_object_unpin(obj);
-			return ret;
-		}
+		if (ret)
+			goto err_unpin;
 	}
 
 	return 0;
+
+err_unpin:
+	i915_gem_object_unpin(obj);
+	return ret;
 }
 
 /* Assume fb object is pinned & idle & fenced and just update base pointers */
@@ -1589,7 +1591,7 @@ intel_pipe_set_base(struct drm_crtc *crtc, int x, int y,
 	obj_priv = to_intel_bo(obj);
 
 	mutex_lock(&dev->struct_mutex);
-	ret = intel_pin_and_fence_fb_obj(dev, obj);
+	ret = intel_pin_and_fence_fb_obj(dev, obj, false);
 	if (ret != 0) {
 		mutex_unlock(&dev->struct_mutex);
 		return ret;
@@ -5004,7 +5006,7 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
 	struct intel_unpin_work *work;
 	unsigned long flags, offset;
 	int pipe = intel_crtc->pipe;
-	u32 pf, pipesrc;
+	u32 was_dirty, pf, pipesrc;
 	int ret;
 
 	work = kzalloc(sizeof *work, GFP_KERNEL);
@@ -5033,7 +5035,8 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
 	obj = intel_fb->obj;
 
 	mutex_lock(&dev->struct_mutex);
-	ret = intel_pin_and_fence_fb_obj(dev, obj);
+	was_dirty = obj->write_domain & I915_GEM_GPU_DOMAINS;
+	ret = intel_pin_and_fence_fb_obj(dev, obj, true);
 	if (ret)
 		goto cleanup_work;
 
@@ -5051,17 +5054,24 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
 	atomic_inc(&obj_priv->pending_flip);
 	work->pending_flip_obj = obj;
 
-	if (IS_GEN3(dev) || IS_GEN2(dev)) {
-		u32 flip_mask;
+	if (was_dirty || IS_GEN3(dev) || IS_GEN2(dev)) {
+		BEGIN_LP_RING(2);
+		if (IS_GEN3(dev) || IS_GEN2(dev)) {
+			u32 flip_mask;
 
-		if (intel_crtc->plane)
-			flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
-		else
-			flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
+			/* Can't queue multiple flips, so wait for the previous
+			 * one to finish before executing the next.
+			 */
 
-		BEGIN_LP_RING(2);
-		OUT_RING(MI_WAIT_FOR_EVENT | flip_mask);
-		OUT_RING(0);
+			if (intel_crtc->plane)
+				flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
+			else
+				flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
+
+			OUT_RING(MI_WAIT_FOR_EVENT | flip_mask);
+		} else
+			OUT_RING(MI_NOOP);
+		OUT_RING(MI_FLUSH);
 		ADVANCE_LP_RING();
 	}
 

drivers/gpu/drm/i915/intel_drv.h

@@ -281,7 +281,8 @@ extern void ironlake_enable_drps(struct drm_device *dev);
 extern void ironlake_disable_drps(struct drm_device *dev);
 
 extern int intel_pin_and_fence_fb_obj(struct drm_device *dev,
-				      struct drm_gem_object *obj);
+				      struct drm_gem_object *obj,
+				      bool pipelined);
 
 extern int intel_framebuffer_init(struct drm_device *dev,
 				  struct intel_framebuffer *ifb,

drivers/gpu/drm/i915/intel_fb.c

@@ -94,7 +94,7 @@ static int intelfb_create(struct intel_fbdev *ifbdev,
 	mutex_lock(&dev->struct_mutex);
 
 	/* Flush everything out, we'll be doing GTT only from now on */
-	ret = intel_pin_and_fence_fb_obj(dev, fbo);
+	ret = intel_pin_and_fence_fb_obj(dev, fbo, false);
 	if (ret) {
 		DRM_ERROR("failed to pin fb: %d\n", ret);
 		goto out_unref;