xf86-video-intel: src/sna/sna_accel.c

[Chris Wilson]

 src/sna/sna_accel.c |   91 +++++++++++++++++++++++++++++++++++++---------------
 1 file changed, 66 insertions(+), 25 deletions(-)

New commits:
commit a3c9cbefcb9971900638144286b59112e3cfd404
Author: Chris Wilson <chris at chris-wilson.co.uk>
Date:   Mon Jun 23 12:27:30 2014 +0100

    sna: Force creation of GPU bo for drawing if need be
    
    Now that we are more eager to transfer to the GPU bo if we expect
    rendering to the CPU bo to be either slow or risk potential corruption,
    we can find ourselves in a recursive loop when we also decide not to
    create the GPU bo. Prevent that by explicitly creating the GPU bo if
    desired.
    
    Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>

diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index f362a9c..9752b18 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -3765,24 +3765,29 @@ use_gpu_bo:
 
 use_cpu_bo:
 	if (!USE_CPU_BO || priv->cpu_bo == NULL) {
-		if ((flags & FORCE_GPU) == 0)
+		if ((flags & FORCE_GPU) == 0) {
+			DBG(("%s: no CPU bo, and GPU not forced\n", __FUNCTION__));
 			return NULL;
+		}
+
+		flags &= ~FORCE_GPU;
+
+		region.extents = *box;
+		if (get_drawable_deltas(drawable, pixmap, &dx, &dy)) {
+			region.extents.x1 += dx;
+			region.extents.x2 += dx;
+			region.extents.y1 += dy;
+			region.extents.y2 += dy;
+		}
+		region.data = NULL;
 
 		if (!sna_drawable_move_region_to_cpu(&pixmap->drawable, &region,
-						     (flags & IGNORE_DAMAGE ? 0 : MOVE_READ) | MOVE_WRITE | MOVE_ASYNC_HINT)) {
+						     (flags & IGNORE_DAMAGE ? 0 : MOVE_READ) | MOVE_WRITE | MOVE_ASYNC_HINT) ||
+		    priv->cpu_bo == NULL) {
+			DBG(("%s: did not create CPU bo\n", __FUNCTION__));
 cpu_fail:
-			if (priv->gpu_bo) {
-				region.extents = *box;
-				if (get_drawable_deltas(drawable, pixmap, &dx, &dy)) {
-					region.extents.x1 += dx;
-					region.extents.x2 += dx;
-					region.extents.y1 += dy;
-					region.extents.y2 += dy;
-				}
-				region.data = NULL;
-
+			if (priv->gpu_bo)
 				goto move_to_gpu;
-			}
 
 			return NULL;
 		}
@@ -3798,7 +3803,6 @@ cpu_fail:
 		return NULL;
 	}
 
-
 	region.extents = *box;
 	if (get_drawable_deltas(drawable, pixmap, &dx, &dy)) {
 		region.extents.x1 += dx;
@@ -3808,27 +3812,64 @@ cpu_fail:
 	}
 	region.data = NULL;
 
-	/* Both CPU and GPU are busy, prefer to use the GPU */
-	if (priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo))
+	if (priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo)) {
+		DBG(("%s: both CPU and GPU are busy, prefer to use the GPU\n",
+		     __FUNCTION__));
 		goto move_to_gpu;
+	}
 
 	assert(priv->gpu_bo == NULL || priv->gpu_bo->proxy == NULL);
 
 	if (flags & RENDER_GPU) {
-		if ((flags & IGNORE_DAMAGE) == 0 && priv->gpu_damage)
-			goto move_to_gpu;
+		flags &= ~RENDER_GPU;
 
-		if (priv->gpu_bo && priv->gpu_bo->tiling)
-			goto move_to_gpu;
+		if ((flags & IGNORE_DAMAGE) == 0 && priv->gpu_damage) {
+			DBG(("%s: prefer to use GPU bo for rendering whilst reading from GPU damage\n", __FUNCTION__));
 
-		if (priv->cpu_bo->pitch >= 4096)
-			goto move_to_gpu;
+prefer_gpu_bo:
+			if (priv->gpu_bo == NULL) {
+				if ((flags & FORCE_GPU) == 0) {
+					DBG(("%s: untiled, will not force allocation\n",
+					     __FUNCTION__));
+					return NULL;
+				}
 
-		if ((flags & IGNORE_DAMAGE) == 0 && priv->cpu_bo->snoop)
-			goto move_to_gpu;
+				if (flags & IGNORE_DAMAGE) {
+					sna_damage_subtract(&priv->cpu_damage, &region);
+					if (priv->cpu_damage == NULL) {
+						list_del(&priv->flush_list);
+						priv->cpu = false;
+					}
+				}
 
-		if (!sna->kgem.can_blt_cpu)
+				if (!sna_pixmap_move_to_gpu(pixmap, MOVE_WRITE | MOVE_READ | __MOVE_FORCE))
+					return NULL;
+
+				DBG(("%s: allocated GPU bo for operation\n", __FUNCTION__));
+				goto done;
+			}
 			goto move_to_gpu;
+		}
+
+		if (priv->gpu_bo && priv->gpu_bo->tiling) {
+			DBG(("%s: prefer to use GPU bo for rendering large pixmaps\n", __FUNCTION__));
+			goto prefer_gpu_bo;
+		}
+
+		if (priv->cpu_bo->pitch >= 4096) {
+			DBG(("%s: prefer to use GPU bo for rendering wide pixmaps\n", __FUNCTION__));
+			goto prefer_gpu_bo;
+		}
+
+		if ((flags & IGNORE_DAMAGE) == 0 && priv->cpu_bo->snoop) {
+			DBG(("%s: prefer to use GPU bo for reading from snooped target bo\n", __FUNCTION__));
+			goto prefer_gpu_bo;
+		}
+
+		if (!sna->kgem.can_blt_cpu) {
+			DBG(("%s: can't render to CPU bo, try to use GPU bo\n", __FUNCTION__));
+			goto prefer_gpu_bo;
+		}
 	}
 
 	if (!sna->kgem.can_blt_cpu)