sna: On LLC systems quietly replace all linear mmappings using the CPU

If the GPU and CPU caches are shared and coherent, we can use a cached
mapping for linear bo in the CPU domain with no penalty and so avoid the
penalty of using WC/UC mappings through the GTT (and any aperture
pressure). We presume that the bo for such mappings are indeed LLC
cached...

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

4 files changed, 21 insertions, 9 deletions

diff --git a/src/sna/kgem.c b/src/sna/kgem.c
index 0075bedb..d119ae36 100644
--- a/src/sna/kgem.c
+++ b/src/sna/kgem.c
@@ -85,6 +85,7 @@ static inline void list_replace(struct list *old,
 #define DBG_NO_HW 0
 #define DBG_NO_TILING 0
 #define DBG_NO_VMAP 0
+#define DBG_NO_LLC 0
 #define DBG_NO_SEMAPHORES 0
 #define DBG_NO_MADV 0
 #define DBG_NO_MAP_UPLOAD 0
@@ -601,6 +602,13 @@ void kgem_init(struct kgem *kgem, int fd, struct pci_device *dev, int gen)
 	DBG(("%s: has relaxed fencing? %d\n", __FUNCTION__,
 	     kgem->has_relaxed_fencing));
 
+	kgem->has_llc = false;
+	if (!DBG_NO_LLC && gen >= 60)
+		kgem->has_llc = true;
+	kgem->has_cpu_bo = kgem->has_llc;
+	DBG(("%s: cpu bo enabled %d: llc? %d\n", __FUNCTION__,
+	     kgem->has_cpu_bo, kgem->has_llc));
+
 	kgem->has_semaphores = false;
 	if (gen >= 60 && semaphores_enabled())
 		kgem->has_semaphores = true;
@@ -2170,6 +2178,8 @@ struct kgem_bo *kgem_create_2d(struct kgem *kgem,
 
 	if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) {
 		int for_cpu = !!(flags & CREATE_CPU_MAP);
+		if (kgem->has_llc && tiling == I915_TILING_NONE)
+			for_cpu = 1;
 		/* We presume that we will need to upload to this bo,
 		 * and so would prefer to have an active VMA.
 		 */
@@ -2604,14 +2614,15 @@ void *kgem_bo_map(struct kgem *kgem, struct kgem_bo *bo)
 	assert(bo->exec == NULL);
 	assert(list_is_empty(&bo->list));
 
-	if (IS_CPU_MAP(bo->map)) {
-		if (bo->tiling == I915_TILING_NONE) {
-			kgem_bo_sync__cpu(kgem, bo);
-			return CPU_MAP(bo->map);
-		}
-		kgem_bo_release_map(kgem, bo);
+	if (kgem->has_llc && bo->tiling == I915_TILING_NONE) {
+		ptr = kgem_bo_map__cpu(kgem, bo);
+		kgem_bo_sync__cpu(kgem, bo);
+		return ptr;
 	}
 
+	if (IS_CPU_MAP(bo->map))
+		kgem_bo_release_map(kgem, bo);
+
 	ptr = bo->map;
 	if (ptr == NULL) {
 		kgem_trim_vma_cache(kgem, MAP_GTT, bo->bucket);
@@ -2958,7 +2969,7 @@ struct kgem_bo *kgem_create_buffer(struct kgem *kgem,
 	bo = NULL;
 
 #if !DBG_NO_MAP_UPLOAD
-	if (!DEBUG_NO_LLC && kgem->gen >= 60) {
+	if (kgem->has_cpu_bo) {
 		struct kgem_bo *old;
 
 		bo = malloc(sizeof(*bo));
diff --git a/src/sna/kgem.h b/src/sna/kgem.h
index 377d21d3..621e3cd5 100644
--- a/src/sna/kgem.h
+++ b/src/sna/kgem.h
@@ -146,6 +146,8 @@ struct kgem {
 	uint32_t has_vmap :1;
 	uint32_t has_relaxed_fencing :1;
 	uint32_t has_semaphores :1;
+	uint32_t has_llc :1;
+	uint32_t has_cpu_bo :1;
 
 	uint16_t fence_max;
 	uint16_t half_cpu_cache_pages;
diff --git a/src/sna/sna.h b/src/sna/sna.h
index 185bc1de..5153d339 100644
--- a/src/sna/sna.h
+++ b/src/sna/sna.h
@@ -93,7 +93,6 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define DEBUG_NO_RENDER 0
 #define DEBUG_NO_BLT 0
 #define DEBUG_NO_IO 0
-#define DEBUG_NO_LLC 0
 
 #define DEBUG_FLUSH_CACHE 0
 #define DEBUG_FLUSH_BATCH 0
diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 16657911..7774cba8 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -205,7 +205,7 @@ sna_pixmap_alloc_cpu(struct sna *sna,
 
 	assert(priv->stride);
 
-	if (!DEBUG_NO_LLC && sna->kgem.gen >= 60) {
+	if (sna->kgem.has_cpu_bo) {
 		DBG(("%s: allocating CPU buffer (%dx%d)\n", __FUNCTION__,
 		     pixmap->drawable.width, pixmap->drawable.height));