diff --git a/src/sna/gen6_common.h b/src/sna/gen6_common.h
index 5da739e2..4bf69b11 100644
--- a/src/sna/gen6_common.h
+++ b/src/sna/gen6_common.h
@@ -52,6 +52,9 @@ inline static bool can_switch_to_blt(struct sna *sna,
 	if (bo && RQ_IS_BLT(bo->rq))
 		return true;
 
+	if (bo && bo->tiling == I915_TILING_Y)
+		return false;
+
 	if (sna->render_state.gt < 2)
 		return true;
 
diff --git a/src/sna/gen6_render.c b/src/sna/gen6_render.c
index 7878e7bc..312d6afd 100644
--- a/src/sna/gen6_render.c
+++ b/src/sna/gen6_render.c
@@ -1965,54 +1965,77 @@ gen6_composite_set_target(struct sna *sna,
 
 static bool
 try_blt(struct sna *sna,
-	PicturePtr dst, PicturePtr src,
-	int width, int height)
+	uint8_t op,
+	PicturePtr src,
+	PicturePtr mask,
+	PicturePtr dst,
+	int16_t src_x, int16_t src_y,
+	int16_t msk_x, int16_t msk_y,
+	int16_t dst_x, int16_t dst_y,
+	int16_t width, int16_t height,
+	unsigned flags,
+	struct sna_composite_op *tmp)
 {
 	struct kgem_bo *bo;
 
 	if (sna->kgem.mode == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
 	if (too_large(width, height)) {
 		DBG(("%s: operation too large for 3D pipe (%d, %d)\n",
 		     __FUNCTION__, width, height));
-		return true;
+		goto execute;
 	}
 
 	bo = __sna_drawable_peek_bo(dst->pDrawable);
 	if (bo == NULL)
-		return true;
+		goto execute;
 
 	if (untiled_tlb_miss(bo))
-		return true;
+		goto execute;
 
-	if (bo->rq)
-		return RQ_IS_BLT(bo->rq);
+	if (bo->rq) {
+		if (RQ_IS_BLT(bo->rq))
+			goto execute;
+
+		return false;
+	}
+
+	if (bo->tiling == I915_TILING_Y)
+		goto upload;
 
 	if (src->pDrawable == dst->pDrawable &&
 	    can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (sna_picture_is_solid(src, NULL) && can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (src->pDrawable) {
 		struct kgem_bo *s = __sna_drawable_peek_bo(src->pDrawable);
 		if (s == NULL)
-			return true;
+			goto execute;
 
 		if (prefer_blt_bo(sna, s, bo))
-			return true;
+			goto execute;
 	}
 
 	if (sna->kgem.ring == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
-	return false;
+upload:
+	flags |= COMPOSITE_UPLOAD;
+execute:
+	return sna_blt_composite(sna, op,
+				 src, dst,
+				 src_x, src_y,
+				 dst_x, dst_y,
+				 width, height,
+				 flags, tmp);
 }
 
 static bool
@@ -2242,13 +2265,13 @@ gen6_render_composite(struct sna *sna,
 	     width, height, sna->kgem.ring));
 
 	if (mask == NULL &&
-	    try_blt(sna, dst, src, width, height) &&
-	    sna_blt_composite(sna, op,
-			      src, dst,
-			      src_x, src_y,
-			      dst_x, dst_y,
-			      width, height,
-			      flags, tmp))
+	    try_blt(sna, op,
+		    src, mask, dst,
+		    src_x, src_y,
+		    msk_x, msk_y,
+		    dst_x, dst_y,
+		    width, height,
+		    flags, tmp))
 		return true;
 
 	if (gen6_composite_fallback(sna, src, mask, dst))
diff --git a/src/sna/gen7_render.c b/src/sna/gen7_render.c
index 39f0d385..d1de7522 100644
--- a/src/sna/gen7_render.c
+++ b/src/sna/gen7_render.c
@@ -2184,55 +2184,78 @@ gen7_composite_set_target(struct sna *sna,
 
 static bool
 try_blt(struct sna *sna,
-	PicturePtr dst, PicturePtr src,
-	int width, int height)
+	uint8_t op,
+	PicturePtr src,
+	PicturePtr mask,
+	PicturePtr dst,
+	int16_t src_x, int16_t src_y,
+	int16_t msk_x, int16_t msk_y,
+	int16_t dst_x, int16_t dst_y,
+	int16_t width, int16_t height,
+	unsigned flags,
+	struct sna_composite_op *tmp)
 {
 	struct kgem_bo *bo;
 
 	if (sna->kgem.mode == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
 	if (too_large(width, height)) {
 		DBG(("%s: operation too large for 3D pipe (%d, %d)\n",
 		     __FUNCTION__, width, height));
-		return true;
+		goto execute;
 	}
 
 	bo = __sna_drawable_peek_bo(dst->pDrawable);
 	if (bo == NULL)
-		return true;
+		goto execute;
 
 	if (untiled_tlb_miss(bo))
-		return true;
+		goto execute;
 
-	if (bo->rq)
-		return RQ_IS_BLT(bo->rq);
+	if (bo->rq) {
+		if (RQ_IS_BLT(bo->rq))
+			goto execute;
+
+		return false;
+	}
+
+	if (bo->tiling == I915_TILING_Y)
+		goto upload;
 
 	if (src->pDrawable == dst->pDrawable &&
 	    (sna->render_state.gt < 3 || width*height < 1024) &&
 	    can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (sna_picture_is_solid(src, NULL) && can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (src->pDrawable) {
 		struct kgem_bo *s = __sna_drawable_peek_bo(src->pDrawable);
 		if (s == NULL)
-			return true;
+			goto upload;
 
 		if (prefer_blt_bo(sna, s, bo))
-			return true;
+			goto execute;
 	}
 
 	if (sna->kgem.ring == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
-	return false;
+upload:
+	flags |= COMPOSITE_UPLOAD;
+execute:
+	return sna_blt_composite(sna, op,
+				 src, dst,
+				 src_x, src_y,
+				 dst_x, dst_y,
+				 width, height,
+				 flags, tmp);
 }
 
 static bool
@@ -2462,13 +2485,13 @@ gen7_render_composite(struct sna *sna,
 	     width, height, sna->kgem.mode, sna->kgem.ring));
 
 	if (mask == NULL &&
-	    try_blt(sna, dst, src, width, height) &&
-	    sna_blt_composite(sna, op,
-			      src, dst,
-			      src_x, src_y,
-			      dst_x, dst_y,
-			      width, height,
-			      flags, tmp))
+	    try_blt(sna, op,
+		    src, mask, dst,
+		    src_x, src_y,
+		    msk_x, msk_y,
+		    dst_x, dst_y,
+		    width, height,
+		    flags, tmp))
 		return true;
 
 	if (gen7_composite_fallback(sna, src, mask, dst))
diff --git a/src/sna/gen8_render.c b/src/sna/gen8_render.c
index c0fe9f9c..b40c582a 100644
--- a/src/sna/gen8_render.c
+++ b/src/sna/gen8_render.c
@@ -2029,55 +2029,78 @@ gen8_composite_set_target(struct sna *sna,
 
 static bool
 try_blt(struct sna *sna,
-	PicturePtr dst, PicturePtr src,
-	int width, int height)
+	uint8_t op,
+	PicturePtr src,
+	PicturePtr mask,
+	PicturePtr dst,
+	int16_t src_x, int16_t src_y,
+	int16_t msk_x, int16_t msk_y,
+	int16_t dst_x, int16_t dst_y,
+	int16_t width, int16_t height,
+	unsigned flags,
+	struct sna_composite_op *tmp)
 {
 	struct kgem_bo *bo;
 
 	if (sna->kgem.mode == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
 	if (too_large(width, height)) {
 		DBG(("%s: operation too large for 3D pipe (%d, %d)\n",
 		     __FUNCTION__, width, height));
-		return true;
+		goto execute;
 	}
 
 	bo = __sna_drawable_peek_bo(dst->pDrawable);
 	if (bo == NULL)
-		return true;
+		goto execute;
 
 	if (untiled_tlb_miss(bo))
-		return true;
+		goto execute;
 
-	if (bo->rq)
-		return RQ_IS_BLT(bo->rq);
+	if (bo->rq) {
+		if (RQ_IS_BLT(bo->rq))
+			goto execute;
+
+		return false;
+	}
+
+	if (bo->tiling == I915_TILING_Y)
+		goto upload;
 
 	if (sna_picture_is_solid(src, NULL) && can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (src->pDrawable == dst->pDrawable &&
 	    (sna->render_state.gt < 3 || width*height < 1024) &&
 	    can_switch_to_blt(sna, bo, 0))
-		return true;
+		goto execute;
 
 	if (src->pDrawable) {
 		struct kgem_bo *s = __sna_drawable_peek_bo(src->pDrawable);
 		if (s == NULL)
-			return true;
+			goto upload;
 
 		if (prefer_blt_bo(sna, s, bo))
-			return true;
+			goto execute;
 	}
 
 	if (sna->kgem.ring == KGEM_BLT) {
 		DBG(("%s: already performing BLT\n", __FUNCTION__));
-		return true;
+		goto execute;
 	}
 
-	return false;
+upload:
+	flags |= COMPOSITE_UPLOAD;
+execute:
+	return sna_blt_composite(sna, op,
+				 src, dst,
+				 src_x, src_y,
+				 dst_x, dst_y,
+				 width, height,
+				 flags, tmp);
 }
 
 static bool
@@ -2307,13 +2330,13 @@ gen8_render_composite(struct sna *sna,
 	     width, height, sna->kgem.mode, sna->kgem.ring));
 
 	if (mask == NULL &&
-	    try_blt(sna, dst, src, width, height) &&
-	    sna_blt_composite(sna, op,
-			      src, dst,
-			      src_x, src_y,
-			      dst_x, dst_y,
-			      width, height,
-			      flags, tmp))
+	    try_blt(sna, op,
+		    src, mask, dst,
+		    src_x, src_y,
+		    msk_x, msk_y,
+		    dst_x, dst_y,
+		    width, height,
+		    flags, tmp))
 		return true;
 
 	if (gen8_composite_fallback(sna, src, mask, dst))
diff --git a/src/sna/sna_blt.c b/src/sna/sna_blt.c
index 9bb32674..1b82abcb 100644
--- a/src/sna/sna_blt.c
+++ b/src/sna/sna_blt.c
@@ -2606,6 +2606,8 @@ clear:
 			}
 			if (hint & REPLACES)
 				kgem_bo_undo(&sna->kgem, tmp->dst.bo);
+			if (flags & COMPOSITE_UPLOAD)
+				return false;
 		} else {
 			RegionRec region;
 
@@ -2692,6 +2694,8 @@ fill:
 			}
 			if (hint & REPLACES)
 				kgem_bo_undo(&sna->kgem, tmp->dst.bo);
+			if (flags & COMPOSITE_UPLOAD)
+				return false;
 		} else {
 			RegionRec region;
 
@@ -2830,7 +2834,7 @@ fill:
 		if (src_pixmap->drawable.width  <= sna->render.max_3d_size &&
 		    src_pixmap->drawable.height <= sna->render.max_3d_size &&
 		    bo->pitch <= sna->render.max_3d_pitch &&
-		    (flags & COMPOSITE_FALLBACK) == 0)
+		    (flags & (COMPOSITE_UPLOAD | COMPOSITE_FALLBACK)) == 0)
 		{
 			return false;
 		}
@@ -2881,7 +2885,7 @@ fallback:
 			DBG(("%s: fallback -- unaccelerated upload\n",
 			     __FUNCTION__));
 			goto fallback;
-		} else {
+		} else if ((flags & COMPOSITE_UPLOAD) == 0) {
 			ret = prepare_blt_copy(sna, tmp, bo, alpha_fixup);
 			if (!ret)
 				goto fallback;
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index 723de80b..1baf9455 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -238,8 +238,9 @@ struct sna_render {
 			  int16_t w, int16_t h,
 			  unsigned flags,
 			  struct sna_composite_op *tmp);
-#define COMPOSITE_PARTIAL 0x1
-#define COMPOSITE_FALLBACK 0x80000000
+#define COMPOSITE_PARTIAL	0x1
+#define COMPOSITE_UPLOAD	0x40000000
+#define COMPOSITE_FALLBACK	0x80000000
 
 	bool (*check_composite_spans)(struct sna *sna, uint8_t op,
 				      PicturePtr dst, PicturePtr src,