drm/nouveau: create temp vmas for both src and dst of bo moves
Greatly simplifies a number of things, particularly once per-client GPU
address spaces are involved.
May add this back later once I know what things'll look like.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 71d01ce..a049984 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -496,19 +496,12 @@
nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
{
- struct nouveau_mem *old_node = old_mem->mm_node;
- struct nouveau_mem *new_node = new_mem->mm_node;
- struct nouveau_bo *nvbo = nouveau_bo(bo);
+ struct nouveau_mem *node = old_mem->mm_node;
+ u64 src_offset = node->vma[0].offset;
+ u64 dst_offset = node->vma[1].offset;
u32 page_count = new_mem->num_pages;
- u64 src_offset, dst_offset;
int ret;
- src_offset = old_node->tmp_vma.offset;
- if (new_node->tmp_vma.node)
- dst_offset = new_node->tmp_vma.offset;
- else
- dst_offset = nvbo->vma.offset;
-
page_count = new_mem->num_pages;
while (page_count) {
int line_count = (page_count > 2047) ? 2047 : page_count;
@@ -542,19 +535,13 @@
nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
{
- struct nouveau_mem *old_node = old_mem->mm_node;
- struct nouveau_mem *new_node = new_mem->mm_node;
+ struct nouveau_mem *node = old_mem->mm_node;
struct nouveau_bo *nvbo = nouveau_bo(bo);
u64 length = (new_mem->num_pages << PAGE_SHIFT);
- u64 src_offset, dst_offset;
+ u64 src_offset = node->vma[0].offset;
+ u64 dst_offset = node->vma[1].offset;
int ret;
- src_offset = old_node->tmp_vma.offset;
- if (new_node->tmp_vma.node)
- dst_offset = new_node->tmp_vma.offset;
- else
- dst_offset = nvbo->vma.offset;
-
while (length) {
u32 amount, stride, height;
@@ -690,6 +677,27 @@
}
static int
+nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
+ struct ttm_mem_reg *mem, struct nouveau_vma *vma)
+{
+ struct nouveau_mem *node = mem->mm_node;
+ int ret;
+
+ ret = nouveau_vm_get(chan->vm, mem->num_pages << PAGE_SHIFT,
+ node->page_shift, NV_MEM_ACCESS_RO, vma);
+ if (ret)
+ return ret;
+
+ if (mem->mem_type == TTM_PL_VRAM)
+ nouveau_vm_map(vma, node);
+ else
+ nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT,
+ node, node->pages);
+
+ return 0;
+}
+
+static int
nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
bool no_wait_reserve, bool no_wait_gpu,
struct ttm_mem_reg *new_mem)
@@ -706,31 +714,20 @@
mutex_lock_nested(&chan->mutex, NOUVEAU_KCHANNEL_MUTEX);
}
- /* create temporary vma for old memory, this will get cleaned
- * up after ttm destroys the ttm_mem_reg
+ /* create temporary vmas for the transfer and attach them to the
+ * old nouveau_mem node, these will get cleaned up after ttm has
+ * destroyed the ttm_mem_reg
*/
if (dev_priv->card_type >= NV_50) {
struct nouveau_mem *node = old_mem->mm_node;
- if (!node->tmp_vma.node) {
- u32 page_shift = nvbo->vma.node->type;
- if (old_mem->mem_type == TTM_PL_TT)
- page_shift = nvbo->vma.vm->spg_shift;
- ret = nouveau_vm_get(chan->vm,
- old_mem->num_pages << PAGE_SHIFT,
- page_shift, NV_MEM_ACCESS_RO,
- &node->tmp_vma);
- if (ret)
- goto out;
- }
+ ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
+ if (ret)
+ goto out;
- if (old_mem->mem_type == TTM_PL_VRAM)
- nouveau_vm_map(&node->tmp_vma, node);
- else {
- nouveau_vm_map_sg(&node->tmp_vma, 0,
- old_mem->num_pages << PAGE_SHIFT,
- node, node->pages);
- }
+ ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
+ if (ret)
+ goto out;
}
if (dev_priv->card_type < NV_50)
@@ -757,7 +754,6 @@
bool no_wait_reserve, bool no_wait_gpu,
struct ttm_mem_reg *new_mem)
{
- struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
struct ttm_placement placement;
struct ttm_mem_reg tmp_mem;
@@ -777,23 +773,7 @@
if (ret)
goto out;
- if (dev_priv->card_type >= NV_50) {
- struct nouveau_bo *nvbo = nouveau_bo(bo);
- struct nouveau_mem *node = tmp_mem.mm_node;
- struct nouveau_vma *vma = &nvbo->vma;
- if (vma->node->type != vma->vm->spg_shift)
- vma = &node->tmp_vma;
- nouveau_vm_map_sg(vma, 0, tmp_mem.num_pages << PAGE_SHIFT,
- node, node->pages);
- }
-
ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
-
- if (dev_priv->card_type >= NV_50) {
- struct nouveau_bo *nvbo = nouveau_bo(bo);
- nouveau_vm_unmap(&nvbo->vma);
- }
-
if (ret)
goto out;
@@ -846,21 +826,15 @@
if (!vma->vm)
return;
- switch (new_mem->mem_type) {
- case TTM_PL_VRAM:
- nouveau_vm_map(vma, node);
- break;
- case TTM_PL_TT:
- if (vma->node->type != vma->vm->spg_shift) {
- nouveau_vm_unmap(vma);
- vma = &node->tmp_vma;
- }
- nouveau_vm_map_sg(vma, 0, new_mem->num_pages << PAGE_SHIFT,
- node, node->pages);
- break;
- default:
+ if (new_mem->mem_type == TTM_PL_VRAM) {
+ nouveau_vm_map(&nvbo->vma, new_mem->mm_node);
+ } else
+ if (new_mem->mem_type == TTM_PL_TT &&
+ nvbo->page_shift == nvbo->vma.vm->spg_shift) {
+ nouveau_vm_map_sg(&nvbo->vma, 0, new_mem->
+ num_pages << PAGE_SHIFT, node, node->pages);
+ } else {
nouveau_vm_unmap(&nvbo->vma);
- break;
}
}
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 1439188..d7083d5 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -77,7 +77,7 @@
struct drm_device *dev;
struct nouveau_vma bar_vma;
- struct nouveau_vma tmp_vma;
+ struct nouveau_vma vma[2];
u8 page_shift;
struct drm_mm_node *tag;
diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index f55b51b..9d96056 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -762,20 +762,29 @@
return 0;
}
+static inline void
+nouveau_mem_node_cleanup(struct nouveau_mem *node)
+{
+ if (node->vma[0].node) {
+ nouveau_vm_unmap(&node->vma[0]);
+ nouveau_vm_put(&node->vma[0]);
+ }
+
+ if (node->vma[1].node) {
+ nouveau_vm_unmap(&node->vma[1]);
+ nouveau_vm_put(&node->vma[1]);
+ }
+}
+
static void
nouveau_vram_manager_del(struct ttm_mem_type_manager *man,
struct ttm_mem_reg *mem)
{
struct drm_nouveau_private *dev_priv = nouveau_bdev(man->bdev);
struct nouveau_vram_engine *vram = &dev_priv->engine.vram;
- struct nouveau_mem *node = mem->mm_node;
struct drm_device *dev = dev_priv->dev;
- if (node->tmp_vma.node) {
- nouveau_vm_unmap(&node->tmp_vma);
- nouveau_vm_put(&node->tmp_vma);
- }
-
+ nouveau_mem_node_cleanup(mem->mm_node);
vram->put(dev, (struct nouveau_mem **)&mem->mm_node);
}
@@ -860,15 +869,9 @@
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);
- }
-
+ nouveau_mem_node_cleanup(mem->mm_node);
mem->mm_node = NULL;
- kfree(node);
+ kfree(mem->mm_node);
}
static int
@@ -878,11 +881,7 @@
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))
@@ -891,24 +890,8 @@
node = kzalloc(sizeof(*node), GFP_KERNEL);
if (!node)
return -ENOMEM;
+ node->page_shift = 12;
- /* 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;