nvk/nvkmd: Rework memory placement flags

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30033>

src/nouveau/vulkan/nvk_device_memory.c

@@ -52,12 +52,6 @@ nvk_memory_type_flags(const VkMemoryType *type,
    if (type->propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
       flags |= NVKMD_MEM_CAN_MAP;
 
-   /* For dma-bufs, we have to allow them to live in GART because they might
-    * get forced there by the kernel if they're shared with another GPU.
-    */
-   if (handle_types & VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT)
-      flags |= NVKMD_MEM_GART;
-
    if (handle_types != 0)
       flags |= NVKMD_MEM_SHARED;
 

src/nouveau/vulkan/nvkmd/nouveau/nvkmd_nouveau_mem.c

@@ -82,15 +82,38 @@ nvkmd_nouveau_alloc_tiled_mem(struct nvkmd_dev *_dev,
                               struct nvkmd_mem **mem_out)
 {
    struct nvkmd_nouveau_dev *dev = nvkmd_nouveau_dev(_dev);
+   const struct nv_device_info *dev_info = &dev->base.pdev->dev_info;
 
-   enum nouveau_ws_bo_flags nouveau_flags = 0;
-   if (flags & NVKMD_MEM_LOCAL)
-      nouveau_flags |= NOUVEAU_WS_BO_LOCAL;
-   if (flags & NVKMD_MEM_GART)
-      nouveau_flags |= NOUVEAU_WS_BO_GART;
+   /* Only one placement flag may be specified */
+   assert(util_bitcount(flags & (NVKMD_MEM_LOCAL |
+                                 NVKMD_MEM_GART |
+                                 NVKMD_MEM_VRAM)) == 1);
+   enum nouveau_ws_bo_flags domains = 0;
+   if (flags & NVKMD_MEM_LOCAL) {
+      domains |= NOUVEAU_WS_BO_GART;
+      if (dev_info->vram_size_B > 0)
+         domains |= NOUVEAU_WS_BO_VRAM;
+   } else if (flags & NVKMD_MEM_GART) {
+      domains |= NOUVEAU_WS_BO_GART;
+   } else if (flags & NVKMD_MEM_VRAM) {
+      domains |= NOUVEAU_WS_BO_VRAM;
+   }
+
+   /* TODO:
+    *
+    * VRAM maps on Kepler appear to be broken and we don't really know why.
+    * My NVIDIA contact doesn't remember them not working so they probably
+    * should but they don't today.  Force everything that may be mapped to
+    * use GART for now.
+    */
+   if (dev_info->chipset < 0x110 && (flags & NOUVEAU_WS_BO_MAP)) {
+      assert(domains & NOUVEAU_WS_BO_GART);
+      domains = NOUVEAU_WS_BO_GART;
+   }
+
+   enum nouveau_ws_bo_flags nouveau_flags = domains;
    if (flags & NVKMD_MEM_CAN_MAP)
       nouveau_flags |= NOUVEAU_WS_BO_MAP;
-
    if (!(flags & NVKMD_MEM_SHARED))
       nouveau_flags |= NOUVEAU_WS_BO_NO_SHARE;
 
@@ -117,10 +140,12 @@ nvkmd_nouveau_import_dma_buf(struct nvkmd_dev *_dev,
       return vk_errorf(log_obj, VK_ERROR_INVALID_EXTERNAL_HANDLE, "%m");
 
    enum nvkmd_mem_flags flags = NVKMD_MEM_SHARED;
-   if (bo->flags & NOUVEAU_WS_BO_LOCAL)
-      flags |= NVKMD_MEM_LOCAL;
-   if (bo->flags & NOUVEAU_WS_BO_GART)
-      flags |= NVKMD_MEM_GART;
+
+   /* We always set LOCAL for shared things because we don't know where the
+    * kernel will place it.  The query only tells us where it is.
+    */
+   flags |= NVKMD_MEM_LOCAL;
+
    if (bo->flags & NOUVEAU_WS_BO_MAP)
       flags |= NVKMD_MEM_CAN_MAP;
 

src/nouveau/vulkan/nvkmd/nvkmd.h

@@ -32,15 +32,25 @@ struct vk_sync_signal;
  */
 
 enum nvkmd_mem_flags {
-   /** VRAM on discrete GPUs or GART on integrated */
+   /** Place memory as local as possible.
+    *
+    * This should be the default for most memory allocations.  On discrete
+    * GPUs, it will default to be placed in VRAM but may be paged out to GART,
+    * depending on system memory pressure.
+    */
    NVKMD_MEM_LOCAL      = 1 << 0,
+
+   /** Place the memory in GART */
    NVKMD_MEM_GART       = 1 << 1,
 
+   /** Place the memory in VRAM */
+   NVKMD_MEM_VRAM       = 1 << 2,
+
    /** This memory object may be mapped */
-   NVKMD_MEM_CAN_MAP    = 1 << 2,
+   NVKMD_MEM_CAN_MAP    = 1 << 3,
 
    /** This memory object may be shared with other processes */
-   NVKMD_MEM_SHARED     = 1 << 3,
+   NVKMD_MEM_SHARED     = 1 << 4,
 };
 
 enum nvkmd_mem_map_flags {

src/nouveau/winsys/nouveau_bo.c

@@ -100,24 +100,14 @@ nouveau_ws_bo_new_tiled_locked(struct nouveau_ws_device *dev,
    req.info.domain = 0;
 
    /* It needs to live somewhere */
-   assert((flags & NOUVEAU_WS_BO_LOCAL) || (flags & NOUVEAU_WS_BO_GART));
+   assert((flags & NOUVEAU_WS_BO_VRAM) || (flags & NOUVEAU_WS_BO_GART));
 
-   if (flags & NOUVEAU_WS_BO_LOCAL)
-      req.info.domain |= dev->local_mem_domain;
+   if (flags & NOUVEAU_WS_BO_VRAM)
+      req.info.domain |= NOUVEAU_GEM_DOMAIN_VRAM;
 
    if (flags & NOUVEAU_WS_BO_GART)
       req.info.domain |= NOUVEAU_GEM_DOMAIN_GART;
 
-   /* TODO:
-    *
-    * VRAM maps on Kepler appear to be broken and we don't really know why.
-    * My NVIDIA contact doesn't remember them not working so they probably
-    * should but they don't today.  Force everything that may be mapped to
-    * use GART for now.
-    */
-   else if (dev->info.chipset < 0x110 && (flags & NOUVEAU_WS_BO_MAP))
-      req.info.domain |= NOUVEAU_GEM_DOMAIN_GART;
-
    if (flags & NOUVEAU_WS_BO_MAP)
       req.info.domain |= NOUVEAU_GEM_DOMAIN_MAPPABLE;
 
@@ -201,8 +191,8 @@ nouveau_ws_bo_from_dma_buf_locked(struct nouveau_ws_device *dev, int fd)
       goto fail_fd_to_handle;
 
    enum nouveau_ws_bo_flags flags = 0;
-   if (info.domain & dev->local_mem_domain)
-      flags |= NOUVEAU_WS_BO_LOCAL;
+   if (info.domain & NOUVEAU_GEM_DOMAIN_VRAM)
+      flags |= NOUVEAU_WS_BO_VRAM;
    if (info.domain & NOUVEAU_GEM_DOMAIN_GART)
       flags |= NOUVEAU_WS_BO_GART;
    if (info.map_handle)

src/nouveau/winsys/nouveau_bo.h

@@ -20,7 +20,7 @@ extern "C" {
 
 enum nouveau_ws_bo_flags {
    /* vram or gart depending on GPU */
-   NOUVEAU_WS_BO_LOCAL = 1 << 0,
+   NOUVEAU_WS_BO_VRAM  = 1 << 0,
    NOUVEAU_WS_BO_GART  = 1 << 1,
    NOUVEAU_WS_BO_MAP   = 1 << 2,
    NOUVEAU_WS_BO_NO_SHARE = 1 << 3,

src/nouveau/winsys/nouveau_device.c

@@ -337,11 +337,6 @@ nouveau_ws_device_new(drmDevicePtr drm_device)
    else
       device->max_push = value;
 
-   if (device->info.vram_size_B == 0)
-      device->local_mem_domain = NOUVEAU_GEM_DOMAIN_GART;
-   else
-      device->local_mem_domain = NOUVEAU_GEM_DOMAIN_VRAM;
-
    if (drm_device->bustype == DRM_BUS_PCI &&
        !nouveau_ws_param(fd, NOUVEAU_GETPARAM_VRAM_BAR_SIZE, &value))
       device->info.bar_size_B = value;

src/nouveau/winsys/nouveau_device.h

@@ -22,7 +22,6 @@ struct nouveau_ws_device {
    struct nv_device_info info;
 
    uint32_t max_push;
-   uint32_t local_mem_domain;
 
    simple_mtx_t bos_lock;
    struct hash_table *bos;