diff --git a/src/sna/sna_blt.c b/src/sna/sna_blt.c
index 7efbcf90..b6749ece 100644
--- a/src/sna/sna_blt.c
+++ b/src/sna/sna_blt.c
@@ -153,6 +153,7 @@ static bool sna_blt_fill_init(struct sna *sna,
 	if (!kgem_check_bo_fenced(kgem, bo, NULL) ||
 	    !kgem_check_batch(kgem, 12)) {
 		_kgem_submit(kgem);
+		assert(kgem_check_bo_fenced(kgem, bo, NULL));
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
@@ -290,6 +291,8 @@ static Bool sna_blt_copy_init(struct sna *sna,
 	kgem_set_mode(kgem, KGEM_BLT);
 	if (!kgem_check_bo_fenced(kgem, src, dst, NULL)) {
 		_kgem_submit(kgem);
+		if (!kgem_check_bo_fenced(kgem, src, dst, NULL))
+			return FALSE;
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
@@ -337,6 +340,8 @@ static Bool sna_blt_alpha_fixup_init(struct sna *sna,
 	kgem_set_mode(kgem, KGEM_BLT);
 	if (!kgem_check_bo_fenced(kgem, src, dst, NULL)) {
 		_kgem_submit(kgem);
+		if (!kgem_check_bo_fenced(kgem, src, dst, NULL))
+			return FALSE;
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
@@ -1109,8 +1114,11 @@ prepare_blt_copy(struct sna *sna,
 	if (!kgem_bo_can_blt(&sna->kgem, priv->gpu_bo))
 		return FALSE;
 
-	if (!kgem_check_bo_fenced(&sna->kgem, priv->gpu_bo, NULL)) {
+	if (!kgem_check_bo_fenced(&sna->kgem, op->dst.bo, priv->gpu_bo, NULL)) {
 		_kgem_submit(&sna->kgem);
+		if (!kgem_check_bo_fenced(&sna->kgem,
+					  op->dst.bo, priv->gpu_bo, NULL))
+			return FALSE;
 		_kgem_set_mode(&sna->kgem, KGEM_BLT);
 	}
 
@@ -1594,6 +1602,7 @@ sna_blt_composite(struct sna *sna,
 
 	if (!kgem_check_bo_fenced(&sna->kgem, priv->gpu_bo, NULL)) {
 		_kgem_submit(&sna->kgem);
+		assert(kgem_check_bo_fenced(&sna->kgem, priv->gpu_bo, NULL));
 		_kgem_set_mode(&sna->kgem, KGEM_BLT);
 	}
 
@@ -1891,6 +1900,7 @@ static bool sna_blt_fill_box(struct sna *sna, uint8_t alu,
 	    !kgem_check_reloc(kgem, 1) ||
 	    !kgem_check_bo_fenced(kgem, bo, NULL)) {
 		_kgem_submit(kgem);
+		assert(kgem_check_bo_fenced(&sna->kgem, bo, NULL));
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
@@ -1964,6 +1974,7 @@ Bool sna_blt_fill_boxes(struct sna *sna, uint8_t alu,
 	if (!kgem_check_bo_fenced(kgem, bo, NULL) ||
 	    !kgem_check_batch(kgem, 12)) {
 		_kgem_submit(kgem);
+		assert(kgem_check_bo_fenced(&sna->kgem, bo, NULL));
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
@@ -2127,6 +2138,11 @@ Bool sna_blt_copy_boxes(struct sna *sna, uint8_t alu,
 	    !kgem_check_reloc(kgem, 2) ||
 	    !kgem_check_bo_fenced(kgem, dst_bo, src_bo, NULL)) {
 		_kgem_submit(kgem);
+		if (!kgem_check_bo_fenced(kgem, dst_bo, src_bo, NULL))
+			return sna_tiling_copy_boxes(sna, alu,
+						     src_bo, src_dx, src_dy,
+						     dst_bo, dst_dx, dst_dy,
+						     bpp, box, nbox);
 		_kgem_set_mode(kgem, KGEM_BLT);
 	}
 
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index 5df53dc0..94c2744e 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -507,6 +507,10 @@ Bool sna_tiling_fill_boxes(struct sna *sna,
 			   const xRenderColor *color,
 			   PixmapPtr dst, struct kgem_bo *dst_bo,
 			   const BoxRec *box, int n);
+Bool sna_tiling_copy_boxes(struct sna *sna, uint8_t alu,
+			   struct kgem_bo *src_bo, int16_t src_dx, int16_t src_dy,
+			   struct kgem_bo *dst_bo, int16_t dst_dx, int16_t dst_dy,
+			   int bpp, const BoxRec *box, int nbox);
 
 Bool sna_blt_composite(struct sna *sna,
 		       uint32_t op,
diff --git a/src/sna/sna_tiling.c b/src/sna/sna_tiling.c
index eac664e3..0bc45390 100644
--- a/src/sna/sna_tiling.c
+++ b/src/sna/sna_tiling.c
@@ -417,3 +417,99 @@ done:
 	pixman_region_fini(&region);
 	return ret;
 }
+
+Bool sna_tiling_copy_boxes(struct sna *sna, uint8_t alu,
+			   struct kgem_bo *src_bo, int16_t src_dx, int16_t src_dy,
+			   struct kgem_bo *dst_bo, int16_t dst_dx, int16_t dst_dy,
+			   int bpp, const BoxRec *box, int nbox)
+{
+	RegionRec region, tile, this;
+	struct kgem_bo *bo;
+	int step;
+	Bool ret = FALSE;
+
+	if (!kgem_bo_can_blt(&sna->kgem, src_bo) ||
+	    !kgem_bo_can_blt(&sna->kgem, dst_bo)) {
+		/* XXX */
+		DBG(("%s: tiling blt fail: src?=%d, dst?=%d\n",
+		     __FUNCTION__,
+		     kgem_bo_can_blt(&sna->kgem, src_bo),
+		     kgem_bo_can_blt(&sna->kgem, dst_bo)));
+		return FALSE;
+	}
+
+	pixman_region_init_rects(&region, box, nbox);
+
+	step = sna->render.max_3d_size;
+	while (step * step * 4 > sna->kgem.max_tile_size)
+		step /= 2;
+
+	DBG(("%s (alu=%d), tile.size=%d, box=%dx[(%d, %d), (%d, %d)])\n",
+	     __FUNCTION__, alu, step, nbox,
+	     region.extents.x1, region.extents.y1,
+	     region.extents.x2, region.extents.y2));
+
+	for (tile.extents.y1 = tile.extents.y2 = region.extents.y1;
+	     tile.extents.y2 < region.extents.y2;
+	     tile.extents.y1 = tile.extents.y2) {
+		tile.extents.y2 = tile.extents.y1 + step;
+		if (tile.extents.y2 > region.extents.y2)
+			tile.extents.y2 = region.extents.y2;
+
+		for (tile.extents.x1 = tile.extents.x2 = region.extents.x1;
+		     tile.extents.x2 < region.extents.x2;
+		     tile.extents.x1 = tile.extents.x2) {
+			int w, h;
+
+			tile.extents.x2 = tile.extents.x1 + step;
+			if (tile.extents.x2 > region.extents.x2)
+				tile.extents.x2 = region.extents.x2;
+
+			tile.data = NULL;
+
+			RegionNull(&this);
+			RegionIntersect(&this, &region, &tile);
+			if (!RegionNotEmpty(&this))
+				continue;
+
+			w = this.extents.x2 - this.extents.x1;
+			h = this.extents.y2 - this.extents.y1;
+			bo = kgem_create_2d(&sna->kgem, w, h, bpp,
+					    kgem_choose_tiling(&sna->kgem,
+							       I915_TILING_X,
+							       w, h, bpp),
+					    0);
+			if (bo) {
+				int16_t dx = this.extents.x1;
+				int16_t dy = this.extents.y1;
+
+				assert(bo->pitch <= 8192);
+				assert(bo->tiling != I915_TILING_Y);
+
+				if (!sna_blt_copy_boxes(sna, alu,
+							src_bo, src_dx, src_dy,
+							bo, -dx, -dy,
+							bpp, REGION_RECTS(&this), REGION_NUM_RECTS(&this)))
+					goto err;
+
+				if (!sna_blt_copy_boxes(sna, alu,
+							bo, -dx, -dy,
+							dst_bo, dst_dx, dst_dy,
+							bpp, REGION_RECTS(&this), REGION_NUM_RECTS(&this)))
+					goto err;
+
+				kgem_bo_destroy(&sna->kgem, bo);
+			}
+			RegionUninit(&this);
+		}
+	}
+
+	ret = TRUE;
+	goto done;
+err:
+	kgem_bo_destroy(&sna->kgem, bo);
+	RegionUninit(&this);
+done:
+	pixman_region_fini(&region);
+	return ret;
+}