[PATCH 4/6] drm/ttm: Support kmap for single-page mappings in ttm_bo_vmap()
Thomas Zimmermann
tzimmermann at suse.de
Fri Jun 14 13:21:58 UTC 2024
In ttm_bo_vmap(), set up single-page mappings with kmap() in certain
cases. The feature is already present in ttm_bo_kmap().
This functionality is require by DRM's xe driver, which claims that
using kmap() is an optimization over vmap(). [1] Reading the commit
at [2] indicates otherwise. It is not possible to use kmap_local_page()
and kunmap_local_page(), as TTM cannot guarantee the requirements for
ordering these calls. [3]
Signed-off-by: Thomas Zimmermann <tzimmermann at suse.de>
Link: https://elixir.bootlin.com/linux/v6.9/source/drivers/gpu/drm/xe/xe_bo.c#L1870 # 1
Link: https://lore.kernel.org/all/20220813220034.806698-1-ira.weiny@intel.com/T/#u # 2
Link: https://elixir.bootlin.com/linux/v6.9/source/include/linux/highmem.h#L70 # 3
---
drivers/gpu/drm/ttm/ttm_bo_util.c | 33 ++++++++++++++++++++++---------
1 file changed, 24 insertions(+), 9 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index 31f9772f05dac..c06cfccace39d 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -516,6 +516,8 @@ int ttm_bo_vmap(struct ttm_buffer_object *bo,
.no_wait_gpu = false
};
struct ttm_tt *ttm = bo->ttm;
+ struct ttm_resource_manager *man =
+ ttm_manager_type(bo->bdev, bo->resource->mem_type);
unsigned long start_page = offset >> PAGE_SHIFT;
unsigned long aligned_size = size + (offset - (start_page << PAGE_SHIFT));
unsigned long num_pages = DIV_ROUND_UP(aligned_size, PAGE_SIZE);
@@ -527,15 +529,25 @@ int ttm_bo_vmap(struct ttm_buffer_object *bo,
if (ret)
return ret;
- /*
- * We need to use vmap to get the desired page protection
- * or to make the buffer object look contiguous.
- */
- prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
- vaddr = vmap(ttm->pages + start_page, num_pages, 0, prot);
- if (!vaddr)
- return -ENOMEM;
- alloc_flags = ttm_bo_map_vmap;
+ if (num_pages == 1 && ttm->caching == ttm_cached &&
+ !(man->use_tt && (ttm->page_flags & TTM_TT_FLAG_DECRYPTED))) {
+ /*
+ * We're mapping a single page, and the desired
+ * page protection is consistent with the bo.
+ */
+ vaddr = kmap(ttm->pages[start_page]);
+ alloc_flags = ttm_bo_map_kmap;
+ } else {
+ /*
+ * We need to use vmap to get the desired page protection
+ * or to make the buffer object look contiguous.
+ */
+ prot = ttm_io_prot(bo, mem, PAGE_KERNEL);
+ vaddr = vmap(ttm->pages + start_page, num_pages, 0, prot);
+ if (!vaddr)
+ return -ENOMEM;
+ alloc_flags = ttm_bo_map_vmap;
+ }
iosys_map_set_vaddr(map, vaddr);
map->alloc_flags = alloc_flags;
@@ -567,6 +579,9 @@ void ttm_bo_vunmap(struct ttm_buffer_object *bo, struct iosys_map *map)
case ttm_bo_map_vmap:
vunmap(map->vaddr);
break;
+ case ttm_bo_map_kmap:
+ kunmap(kmap_to_page(map->vaddr));
+ break;
case ttm_bo_map_premapped:
break;
default:
--
2.45.2
More information about the Intel-gfx
mailing list