diff --git a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
index d65dcaee38323..2158d7a61c5e5 100644
--- a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
+++ b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
@@ -37,6 +37,7 @@
  * @xstride: value to add to the pixel pointer between each line
  * @pstride: value to add to the pixel pointer between each pixel
  * @nplanes: number of planes (deduced from pixel_format)
+ * @prepared: plane update has been prepared
  */
 struct atmel_hlcdc_plane_state {
 	struct drm_plane_state base;
@@ -64,6 +65,7 @@ struct atmel_hlcdc_plane_state {
 	int xstride[ATMEL_HLCDC_MAX_PLANES];
 	int pstride[ATMEL_HLCDC_MAX_PLANES];
 	int nplanes;
+	bool prepared;
 };
 
 static inline struct atmel_hlcdc_plane_state *
@@ -714,12 +716,54 @@ static int atmel_hlcdc_plane_atomic_check(struct drm_plane *p,
 static int atmel_hlcdc_plane_prepare_fb(struct drm_plane *p,
 					const struct drm_plane_state *new_state)
 {
+	/*
+	 * FIXME: we should avoid this const -> non-const cast but it's
+	 * currently the only solution we have to modify the ->prepared
+	 * state and rollback the update request.
+	 * Ideally, we should rework the code to attach all the resources
+	 * to atmel_hlcdc_plane_state (including the DMA desc allocation),
+	 * but this require a complete rework of the atmel_hlcdc_layer
+	 * code.
+	 */
+	struct drm_plane_state *s = (struct drm_plane_state *)new_state;
 	struct atmel_hlcdc_plane *plane = drm_plane_to_atmel_hlcdc_plane(p);
+	struct atmel_hlcdc_plane_state *state =
+			drm_plane_state_to_atmel_hlcdc_plane_state(s);
+	int ret;
 
-	if (!new_state->fb)
-		return 0;
+	ret = atmel_hlcdc_layer_update_start(&plane->layer);
+	if (!ret)
+		state->prepared = true;
+
+	return ret;
+}
+
+static void atmel_hlcdc_plane_cleanup_fb(struct drm_plane *p,
+				const struct drm_plane_state *old_state)
+{
+	/*
+	 * FIXME: we should avoid this const -> non-const cast but it's
+	 * currently the only solution we have to modify the ->prepared
+	 * state and rollback the update request.
+	 * Ideally, we should rework the code to attach all the resources
+	 * to atmel_hlcdc_plane_state (including the DMA desc allocation),
+	 * but this require a complete rework of the atmel_hlcdc_layer
+	 * code.
+	 */
+	struct drm_plane_state *s = (struct drm_plane_state *)old_state;
+	struct atmel_hlcdc_plane *plane = drm_plane_to_atmel_hlcdc_plane(p);
+	struct atmel_hlcdc_plane_state *state =
+			drm_plane_state_to_atmel_hlcdc_plane_state(s);
+
+	/*
+	 * The Request has already been applied or cancelled, nothing to do
+	 * here.
+	 */
+	if (!state->prepared)
+		return;
 
-	return atmel_hlcdc_layer_update_start(&plane->layer);
+	atmel_hlcdc_layer_update_rollback(&plane->layer);
+	state->prepared = false;
 }
 
 static void atmel_hlcdc_plane_atomic_update(struct drm_plane *p,
@@ -844,6 +888,7 @@ static void atmel_hlcdc_plane_init_properties(struct atmel_hlcdc_plane *plane,
 
 static struct drm_plane_helper_funcs atmel_hlcdc_layer_plane_helper_funcs = {
 	.prepare_fb = atmel_hlcdc_plane_prepare_fb,
+	.cleanup_fb = atmel_hlcdc_plane_cleanup_fb,
 	.atomic_check = atmel_hlcdc_plane_atomic_check,
 	.atomic_update = atmel_hlcdc_plane_atomic_update,
 	.atomic_disable = atmel_hlcdc_plane_atomic_disable,
@@ -883,6 +928,7 @@ atmel_hlcdc_plane_atomic_duplicate_state(struct drm_plane *p)
 		return NULL;
 
 	copy->disc_updated = false;
+	copy->prepared = false;
 
 	if (copy->base.fb)
 		drm_framebuffer_reference(copy->base.fb);