Hi,
On 2023/6/26 20:32, Maxime Ripard wrote:
Hi,
On Tue, Jun 20, 2023 at 06:18:31PM +0200, Thomas Zimmermann wrote:
Am 20.06.23 um 18:06 schrieb Sui Jingfeng:
Hi,
On 2023/6/20 22:43, Thomas Zimmermann wrote:
Hi
Am 20.06.23 um 06:08 schrieb Sui Jingfeng:
ping ?
On 2023/3/14 20:53, Sui Jingfeng wrote:
else is not generally useful after return
No indention please.
OK, will be fixed at the next version.
Signed-off-by: Sui Jingfeng <15330273260@xxxxxx>
---
drivers/gpu/drm/drm_gem.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index a6208e2c089b..364e3733af98 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -1150,8 +1150,8 @@ int drm_gem_pin(struct drm_gem_object *obj)
{
if (obj->funcs->pin)
return obj->funcs->pin(obj);
- else
- return 0;
+
+ return 0;
This change is ok.
}
void drm_gem_unpin(struct drm_gem_object *obj)
@@ -1172,7 +1172,8 @@ int drm_gem_vmap(struct drm_gem_object
*obj, struct iosys_map *map)
ret = obj->funcs->vmap(obj, map);
if (ret)
return ret;
- else if (iosys_map_is_null(map))
+
+ if (iosys_map_is_null(map))
return -ENOMEM;
This is not correct. Calling iosys_map_is_null() is part of handling
the return value from vmap, so the else is fine.
Are you serious ?
1. Before apply this patch:
If the 'ret' is 0, it stand for obj->funcs->vmap() is successful, then
if (iosys_map_is_null(map)) will be run.
If the 'ret' is NOT 0, then it return immediately.
2. After apply this patch:
If the 'ret' is NOT 0, it stand for obj->funcs->vmap() failed, then it
return immediately.
If the 'ret' is 0, it stand for obj->funcs->vmap() is successful, then
the check if (iosys_map_is_null(map))
will be run!
I feel strange about the core here, I think the check ' if
(iosys_map_is_null(map))' is not needed,
the implement should responsible to handle all of possible errors.
The ->vmap() callback can succeed with ret=0, but we still have no memory.
Then we return -ENOMEM. The call to _is_null(map) is part of the error
handling for ->vmap(). That is a bit strange, but it as always worked like
that. Keeping all error handling in the same if-else block make all this
more obvious.
Reading that patch, it wasn't obvious to me at all and could have made
the same patch.
Could we add a comment maybe to make it more obvious?
It could, but what should we to do ?
It seems that it is true the ->vmap() callback can succeed with ret=0,
But I think this break the *convention*,
the vmap() function at mm/vmalloc.c already said "Return: the address of
the area or %NULL on failure."
The kernel's vmap() function return NULL on failure(may because the
space is not enough on 32-bit ARM SoCs).
But the drm core's vmap hook just don't honor this.
Take the drm/tegra as an example:
```
static void *tegra_bo_mmap(struct host1x_bo *bo)
{
struct tegra_bo *obj = host1x_to_tegra_bo(bo);
struct iosys_map map;
int ret;
if (obj->vaddr) {
return obj->vaddr;
} else if (obj->gem.import_attach) {
ret = dma_buf_vmap_unlocked(obj->gem.import_attach->dmabuf, &map);
return ret ? NULL : map.vaddr;
} else {
return vmap(obj->pages, obj->num_pages, VM_MAP,
pgprot_writecombine(PAGE_KERNEL));
}
}
static int tegra_gem_prime_vmap(struct dma_buf *buf, struct iosys_map *map)
{
struct drm_gem_object *gem = buf->priv;
struct tegra_bo *bo = to_tegra_bo(gem);
void *vaddr;
vaddr = tegra_bo_mmap(&bo->base);
if (IS_ERR(vaddr))
return PTR_ERR(vaddr);
iosys_map_set_vaddr(map, vaddr);
return 0;
}
```
On one of the all code path, tegra_gem_prime_vmap() call vmap() fucntion
to give the vmap the caller wanted.
but the tegra_gem_prime_vmap() function *only* think 'error' as error(by
calling IS_ERR(vaddr)),
But I think the 'NULL' should also be counted as error.
For my patch(this patch), but the ultimate results is same for the case
before apply this patch and after apply this patch.
Their are also many other drivers have the same problem.
But I don't have their hardware to do the basic test.
Maxime
--
Jingfeng
