drm/nv50-nvc0: delay GART binding until move_notify time

The immediate benefit of doing this is that on NV50 and up, the GPU
virtual address of any buffer is now constant, regardless of what
memtype they're placed in.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

1 files changed, 81 insertions, 0 deletions

diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index ff5fe28b467b..73f37bd0adfa 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -785,3 +785,84 @@ const struct ttm_mem_type_manager_func nouveau_vram_manager = {
         nouveau_vram_manager_del,
         nouveau_vram_manager_debug
 };
+
+static int
+nouveau_gart_manager_init(struct ttm_mem_type_manager *man, unsigned long psize)
+{
+        return 0;
+}
+
+static int
+nouveau_gart_manager_fini(struct ttm_mem_type_manager *man)
+{
+        return 0;
+}
+
+static void
+nouveau_gart_manager_del(struct ttm_mem_type_manager *man,
+                         struct ttm_mem_reg *mem)
+{
+        struct nouveau_mem *node = mem->mm_node;
+
+        if (node->tmp_vma.node) {
+                nouveau_vm_unmap(&node->tmp_vma);
+                nouveau_vm_put(&node->tmp_vma);
+        }
+        mem->mm_node = NULL;
+}
+
+static int
+nouveau_gart_manager_new(struct ttm_mem_type_manager *man,
+                         struct ttm_buffer_object *bo,
+                         struct ttm_placement *placement,
+                         struct ttm_mem_reg *mem)
+{
+        struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
+        struct nouveau_bo *nvbo = nouveau_bo(bo);
+        struct nouveau_vma *vma = &nvbo->vma;
+        struct nouveau_vm *vm = vma->vm;
+        struct nouveau_mem *node;
+        int ret;
+
+        if (unlikely((mem->num_pages << PAGE_SHIFT) >=
+                     dev_priv->gart_info.aper_size))
+                return -ENOMEM;
+
+        node = kzalloc(sizeof(*node), GFP_KERNEL);
+        if (!node)
+                return -ENOMEM;
+
+        /* This node must be for evicting large-paged VRAM
+         * to system memory.  Due to a nv50 limitation of
+         * not being able to mix large/small pages within
+         * the same PDE, we need to create a temporary
+         * small-paged VMA for the eviction.
+         */
+        if (vma->node->type != vm->spg_shift) {
+                ret = nouveau_vm_get(vm, (u64)vma->node->length << 12,
+                                     vm->spg_shift, NV_MEM_ACCESS_RW,
+                                     &node->tmp_vma);
+                if (ret) {
+                        kfree(node);
+                        return ret;
+                }
+        }
+
+        node->page_shift = nvbo->vma.node->type;
+        mem->mm_node = node;
+        mem->start   = 0;
+        return 0;
+}
+
+void
+nouveau_gart_manager_debug(struct ttm_mem_type_manager *man, const char *prefix)
+{
+}
+
+const struct ttm_mem_type_manager_func nouveau_gart_manager = {
+        nouveau_gart_manager_init,
+        nouveau_gart_manager_fini,
+        nouveau_gart_manager_new,
+        nouveau_gart_manager_del,
+        nouveau_gart_manager_debug
+};