drm/i915: Preallocate requests

By allocating the request prior to writing to the ringbuffer, we can
abort the operation without leaving the GPU in an inconsistent state.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers/gpu/drm/i915/i915_drv.h

@@ -993,6 +993,7 @@ int i915_gpu_idle(struct drm_device *dev);
 int i915_gem_idle(struct drm_device *dev);
 uint32_t i915_add_request(struct drm_device *dev,
 			  struct drm_file *file_priv,
+			  struct drm_i915_gem_request *request,
 			  struct intel_ring_buffer *ring);
 int i915_do_wait_request(struct drm_device *dev,
 			 uint32_t seqno,

drivers/gpu/drm/i915/i915_gem.c

@@ -1609,20 +1609,22 @@ i915_gem_process_flushing_list(struct drm_device *dev,
 uint32_t
 i915_add_request(struct drm_device *dev,
 		 struct drm_file *file_priv,
+		 struct drm_i915_gem_request *request,
 		 struct intel_ring_buffer *ring)
 {
 	drm_i915_private_t *dev_priv = dev->dev_private;
 	struct drm_i915_file_private *i915_file_priv = NULL;
-	struct drm_i915_gem_request *request;
 	uint32_t seqno;
 	int was_empty;
 
 	if (file_priv != NULL)
 		i915_file_priv = file_priv->driver_priv;
 
+	if (request == NULL) {
 		request = kzalloc(sizeof(*request), GFP_KERNEL);
 		if (request == NULL)
 			return 0;
+	}
 
 	seqno = ring->add_request(dev, ring, file_priv, 0);
 
@@ -1839,7 +1841,7 @@ i915_do_wait_request(struct drm_device *dev, uint32_t seqno,
 	BUG_ON(seqno == 0);
 
 	if (seqno == dev_priv->next_seqno) {
-		seqno = i915_add_request(dev, NULL, ring);
+		seqno = i915_add_request(dev, NULL, NULL, ring);
 		if (seqno == 0)
 			return -ENOMEM;
 	}
@@ -3505,8 +3507,7 @@ i915_gem_wait_for_pending_flip(struct drm_device *dev,
 	return ret;
 }
 
-
-int
+static int
 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 		       struct drm_file *file_priv,
 		       struct drm_i915_gem_execbuffer2 *args,
@@ -3518,6 +3519,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	struct drm_i915_gem_object *obj_priv;
 	struct drm_clip_rect *cliprects = NULL;
 	struct drm_i915_gem_relocation_entry *relocs = NULL;
+	struct drm_i915_gem_request *request = NULL;
 	int ret = 0, ret2, i, pinned = 0;
 	uint64_t exec_offset;
 	uint32_t seqno, reloc_index;
@@ -3571,6 +3573,12 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 		}
 	}
 
+	request = kzalloc(sizeof(*request), GFP_KERNEL);
+	if (request == NULL) {
+		ret = -ENOMEM;
+		goto pre_mutex_err;
+	}
+
 	ret = i915_gem_get_relocs_from_user(exec_list, args->buffer_count,
 					    &relocs);
 	if (ret != 0)
@@ -3736,11 +3744,11 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	}
 
 	if (dev_priv->render_ring.outstanding_lazy_request) {
-		(void)i915_add_request(dev, file_priv, &dev_priv->render_ring);
+		(void)i915_add_request(dev, file_priv, NULL, &dev_priv->render_ring);
 		dev_priv->render_ring.outstanding_lazy_request = false;
 	}
 	if (dev_priv->bsd_ring.outstanding_lazy_request) {
-		(void)i915_add_request(dev, file_priv, &dev_priv->bsd_ring);
+		(void)i915_add_request(dev, file_priv, NULL, &dev_priv->bsd_ring);
 		dev_priv->bsd_ring.outstanding_lazy_request = false;
 	}
 
@@ -3810,7 +3818,8 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 	 * *some* interrupts representing completion of buffers that we can
 	 * wait on when trying to clear up gtt space).
 	 */
-	seqno = i915_add_request(dev, file_priv, ring);
+	seqno = i915_add_request(dev, file_priv, request, ring);
+	request = NULL;
 
 #if WATCH_LRU
 	i915_dump_lru(dev, __func__);
@@ -3849,6 +3858,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 
 	drm_free_large(object_list);
 	kfree(cliprects);
+	kfree(request);
 
 	return ret;
 }
@@ -4199,7 +4209,7 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data,
 		 */
 		if (obj->write_domain) {
 			i915_gem_flush(dev, 0, obj->write_domain);
-			(void)i915_add_request(dev, file_priv, obj_priv->ring);
+			(void)i915_add_request(dev, file_priv, NULL, obj_priv->ring);
 		}
 
 		/* Update the active list for the hardware's current position.

drivers/gpu/drm/i915/intel_overlay.c

@@ -218,6 +218,7 @@ static void intel_overlay_unmap_regs(struct intel_overlay *overlay,
 }
 
 static int intel_overlay_do_wait_request(struct intel_overlay *overlay,
+					 struct drm_i915_gem_request *request,
 					 bool interruptible,
 					 int stage)
 {
@@ -226,7 +227,7 @@ static int intel_overlay_do_wait_request(struct intel_overlay *overlay,
 	int ret;
 
 	overlay->last_flip_req =
-		i915_add_request(dev, NULL, &dev_priv->render_ring);
+		i915_add_request(dev, NULL, request, &dev_priv->render_ring);
 	if (overlay->last_flip_req == 0)
 		return -ENOMEM;
 
@@ -246,11 +247,15 @@ static int intel_overlay_do_wait_request(struct intel_overlay *overlay,
 static int intel_overlay_on(struct intel_overlay *overlay)
 {
 	struct drm_device *dev = overlay->dev;
+	struct drm_i915_gem_request *request;
 
 	BUG_ON(overlay->active);
-
 	overlay->active = 1;
 
+	request = kzalloc(sizeof(*request), GFP_KERNEL);
+	if (request == NULL)
+		return -ENOMEM;
+
 	BEGIN_LP_RING(4);
 	OUT_RING(MI_OVERLAY_FLIP | MI_OVERLAY_ON);
 	OUT_RING(overlay->flip_addr | OFC_UPDATE);
@@ -258,21 +263,26 @@ static int intel_overlay_on(struct intel_overlay *overlay)
 	OUT_RING(MI_NOOP);
 	ADVANCE_LP_RING();
 
-	return intel_overlay_do_wait_request(overlay, true,
+	return intel_overlay_do_wait_request(overlay, request, true,
 					     NEEDS_WAIT_FOR_FLIP);
 }
 
 /* overlay needs to be enabled in OCMD reg */
-static void intel_overlay_continue(struct intel_overlay *overlay,
+static int intel_overlay_continue(struct intel_overlay *overlay,
 				  bool load_polyphase_filter)
 {
 	struct drm_device *dev = overlay->dev;
         drm_i915_private_t *dev_priv = dev->dev_private;
+	struct drm_i915_gem_request *request;
 	u32 flip_addr = overlay->flip_addr;
 	u32 tmp;
 
 	BUG_ON(!overlay->active);
 
+	request = kzalloc(sizeof(*request), GFP_KERNEL);
+	if (request == NULL)
+		return -ENOMEM;
+
 	if (load_polyphase_filter)
 		flip_addr |= OFC_UPDATE;
 
@@ -287,17 +297,23 @@ static void intel_overlay_continue(struct intel_overlay *overlay,
         ADVANCE_LP_RING();
 
 	overlay->last_flip_req =
-		i915_add_request(dev, NULL, &dev_priv->render_ring);
+		i915_add_request(dev, NULL, request, &dev_priv->render_ring);
+	return 0;
 }
 
 /* overlay needs to be disabled in OCMD reg */
 static int intel_overlay_off(struct intel_overlay *overlay)
 {
-	u32 flip_addr = overlay->flip_addr;
 	struct drm_device *dev = overlay->dev;
+	u32 flip_addr = overlay->flip_addr;
+	struct drm_i915_gem_request *request;
 
 	BUG_ON(!overlay->active);
 
+	request = kzalloc(sizeof(*request), GFP_KERNEL);
+	if (request == NULL)
+		return -ENOMEM;
+
 	/* According to intel docs the overlay hw may hang (when switching
 	 * off) without loading the filter coeffs. It is however unclear whether
 	 * this applies to the disabling of the overlay or to the switching off
@@ -315,7 +331,8 @@ static int intel_overlay_off(struct intel_overlay *overlay)
 	OUT_RING(MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
 	ADVANCE_LP_RING();
 
-	return intel_overlay_do_wait_request(overlay, true, SWITCH_OFF);
+	return intel_overlay_do_wait_request(overlay, request, true,
+					     SWITCH_OFF);
 }
 
 static void intel_overlay_release_old_vid_tail(struct intel_overlay *overlay)
@@ -397,13 +414,19 @@ static int intel_overlay_release_old_vid(struct intel_overlay *overlay)
 		return 0;
 
 	if (I915_READ(ISR) & I915_OVERLAY_PLANE_FLIP_PENDING_INTERRUPT) {
+		struct drm_i915_gem_request *request;
+
 		/* synchronous slowpath */
+		request = kzalloc(sizeof(*request), GFP_KERNEL);
+		if (request == NULL)
+			return -ENOMEM;
+
 		BEGIN_LP_RING(2);
 		OUT_RING(MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
 		OUT_RING(MI_NOOP);
 		ADVANCE_LP_RING();
 
-		ret = intel_overlay_do_wait_request(overlay, true,
+		ret = intel_overlay_do_wait_request(overlay, request, true,
 						    RELEASE_OLD_VID);
 		if (ret)
 			return ret;
@@ -755,7 +778,9 @@ int intel_overlay_do_put_image(struct intel_overlay *overlay,
 
 	intel_overlay_unmap_regs(overlay, regs);
 
-	intel_overlay_continue(overlay, scale_changed);
+	ret = intel_overlay_continue(overlay, scale_changed);
+	if (ret)
+		goto out_unpin;
 
 	overlay->old_vid_bo = overlay->vid_bo;
 	overlay->vid_bo = to_intel_bo(new_bo);