[v1] accel/amdxdna: Call dma_buf_vmap_unlocked() for imported object

[RESEND PATCH V1] accel/amdxdna: Call dma_buf_vmap_unlocked() for imported object

Posted by Lizhi Hou 2 weeks, 2 days ago

In amdxdna_gem_obj_vmap(), calling dma_buf_vmap() triggers a kernel
warning if LOCKDEP is enabled. So for imported object, use
dma_buf_vmap_unlocked(). Then, use drm_gem_vmap() for other objects.
The similar change applies to vunmap code.

Fixes: bd72d4acda10 ("accel/amdxdna: Support user space allocated buffer")
Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
---
 drivers/accel/amdxdna/amdxdna_gem.c | 38 +++++++++++------------------
 1 file changed, 14 insertions(+), 24 deletions(-)

diff --git a/drivers/accel/amdxdna/amdxdna_gem.c b/drivers/accel/amdxdna/amdxdna_gem.c
index d407a36eb412..50950be189ae 100644
--- a/drivers/accel/amdxdna/amdxdna_gem.c
+++ b/drivers/accel/amdxdna/amdxdna_gem.c
@@ -392,35 +392,25 @@ static const struct dma_buf_ops amdxdna_dmabuf_ops = {
 	.vunmap = drm_gem_dmabuf_vunmap,
 };
 
-static int amdxdna_gem_obj_vmap(struct drm_gem_object *obj, struct iosys_map *map)
+static int amdxdna_gem_obj_vmap(struct amdxdna_gem_obj *abo, struct iosys_map *map)
 {
-	struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
-
 	iosys_map_clear(map);
 
-	dma_resv_assert_held(obj->resv);
-
 	if (is_import_bo(abo))
-		dma_buf_vmap(abo->dma_buf, map);
-	else
-		drm_gem_shmem_object_vmap(obj, map);
-
-	if (!map->vaddr)
-		return -ENOMEM;
+		return dma_buf_vmap_unlocked(abo->dma_buf, map);
 
-	return 0;
+	return drm_gem_vmap(to_gobj(abo), map);
 }
 
-static void amdxdna_gem_obj_vunmap(struct drm_gem_object *obj, struct iosys_map *map)
+static void amdxdna_gem_obj_vunmap(struct amdxdna_gem_obj *abo, struct iosys_map *map)
 {
-	struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
-
-	dma_resv_assert_held(obj->resv);
+	if (iosys_map_is_null(map))
+		return;
 
 	if (is_import_bo(abo))
-		dma_buf_vunmap(abo->dma_buf, map);
-	else
-		drm_gem_shmem_object_vunmap(obj, map);
+		return dma_buf_vunmap_unlocked(abo->dma_buf, map);
+
+	return drm_gem_vunmap(to_gobj(abo), map);
 }
 
 static struct dma_buf *amdxdna_gem_prime_export(struct drm_gem_object *gobj, int flags)
@@ -468,7 +458,7 @@ static void amdxdna_gem_obj_free(struct drm_gem_object *gobj)
 	if (abo->type == AMDXDNA_BO_DEV_HEAP)
 		drm_mm_takedown(&abo->mm);
 
-	drm_gem_vunmap(gobj, &map);
+	amdxdna_gem_obj_vunmap(abo, &map);
 	mutex_destroy(&abo->lock);
 
 	if (is_import_bo(abo)) {
@@ -489,8 +479,8 @@ static const struct drm_gem_object_funcs amdxdna_gem_shmem_funcs = {
 	.pin = drm_gem_shmem_object_pin,
 	.unpin = drm_gem_shmem_object_unpin,
 	.get_sg_table = drm_gem_shmem_object_get_sg_table,
-	.vmap = amdxdna_gem_obj_vmap,
-	.vunmap = amdxdna_gem_obj_vunmap,
+	.vmap = drm_gem_shmem_object_vmap,
+	.vunmap = drm_gem_shmem_object_vunmap,
 	.mmap = amdxdna_gem_obj_mmap,
 	.vm_ops = &drm_gem_shmem_vm_ops,
 	.export = amdxdna_gem_prime_export,
@@ -692,7 +682,7 @@ amdxdna_drm_create_dev_heap(struct drm_device *dev,
 	abo->mem.dev_addr = client->xdna->dev_info->dev_mem_base;
 	drm_mm_init(&abo->mm, abo->mem.dev_addr, abo->mem.size);
 
-	ret = drm_gem_vmap(to_gobj(abo), &map);
+	ret = amdxdna_gem_obj_vmap(abo, &map);
 	if (ret) {
 		XDNA_ERR(xdna, "Vmap heap bo failed, ret %d", ret);
 		goto release_obj;
@@ -770,7 +760,7 @@ amdxdna_drm_create_cmd_bo(struct drm_device *dev,
 	abo->type = AMDXDNA_BO_CMD;
 	abo->client = filp->driver_priv;
 
-	ret = drm_gem_vmap(to_gobj(abo), &map);
+	ret = amdxdna_gem_obj_vmap(abo, &map);
 	if (ret) {
 		XDNA_ERR(xdna, "Vmap cmd bo failed, ret %d", ret);
 		goto release_obj;
-- 
2.34.1

Re: [RESEND PATCH V1] accel/amdxdna: Call dma_buf_vmap_unlocked() for imported object

Posted by Falkowski, Maciej 2 weeks, 1 day ago

On 9/15/2025 6:10 PM, Lizhi Hou wrote:

> In amdxdna_gem_obj_vmap(), calling dma_buf_vmap() triggers a kernel
> warning if LOCKDEP is enabled. So for imported object, use
> dma_buf_vmap_unlocked(). Then, use drm_gem_vmap() for other objects.
> The similar change applies to vunmap code.
>
> Fixes: bd72d4acda10 ("accel/amdxdna: Support user space allocated buffer")
> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
> ---
>   drivers/accel/amdxdna/amdxdna_gem.c | 38 +++++++++++------------------
>   1 file changed, 14 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/accel/amdxdna/amdxdna_gem.c b/drivers/accel/amdxdna/amdxdna_gem.c
> index d407a36eb412..50950be189ae 100644
> --- a/drivers/accel/amdxdna/amdxdna_gem.c
> +++ b/drivers/accel/amdxdna/amdxdna_gem.c
> @@ -392,35 +392,25 @@ static const struct dma_buf_ops amdxdna_dmabuf_ops = {
>   	.vunmap = drm_gem_dmabuf_vunmap,
>   };
>   
> -static int amdxdna_gem_obj_vmap(struct drm_gem_object *obj, struct iosys_map *map)
> +static int amdxdna_gem_obj_vmap(struct amdxdna_gem_obj *abo, struct iosys_map *map)
>   {
> -	struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
> -
>   	iosys_map_clear(map);
>   
> -	dma_resv_assert_held(obj->resv);
> -
>   	if (is_import_bo(abo))
> -		dma_buf_vmap(abo->dma_buf, map);
> -	else
> -		drm_gem_shmem_object_vmap(obj, map);
> -
> -	if (!map->vaddr)
> -		return -ENOMEM;
> +		return dma_buf_vmap_unlocked(abo->dma_buf, map);
Hi,

The dma_buf_vmap_unlocked() will call iosys_map_clear at its start so that
in this case it will be called twice. Probably it will be optimize out, 
but maybe
its something to better omit.
>   
> -	return 0;
> +	return drm_gem_vmap(to_gobj(abo), map);
>   }
>   
> -static void amdxdna_gem_obj_vunmap(struct drm_gem_object *obj, struct iosys_map *map)
> +static void amdxdna_gem_obj_vunmap(struct amdxdna_gem_obj *abo, struct iosys_map *map)
>   {
> -	struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
> -
> -	dma_resv_assert_held(obj->resv);
> +	if (iosys_map_is_null(map))
> +		return;
>   
>   	if (is_import_bo(abo))
> -		dma_buf_vunmap(abo->dma_buf, map);
> -	else
> -		drm_gem_shmem_object_vunmap(obj, map);
> +		return dma_buf_vunmap_unlocked(abo->dma_buf, map);
I do also wonder what is the convention here on clearing iosys_map when 
returning.
The function drm_gem_vunmap will clear the map for callers while the 
other not.
I think at least comment explaining the logic will be necessary.

Best regards,
Maciej

> +
> +	return drm_gem_vunmap(to_gobj(abo), map);
>   }
>   
>   static struct dma_buf *amdxdna_gem_prime_export(struct drm_gem_object *gobj, int flags)
> @@ -468,7 +458,7 @@ static void amdxdna_gem_obj_free(struct drm_gem_object *gobj)
>   	if (abo->type == AMDXDNA_BO_DEV_HEAP)
>   		drm_mm_takedown(&abo->mm);
>   
> -	drm_gem_vunmap(gobj, &map);
> +	amdxdna_gem_obj_vunmap(abo, &map);
>   	mutex_destroy(&abo->lock);
>   
>   	if (is_import_bo(abo)) {
> @@ -489,8 +479,8 @@ static const struct drm_gem_object_funcs amdxdna_gem_shmem_funcs = {
>   	.pin = drm_gem_shmem_object_pin,
>   	.unpin = drm_gem_shmem_object_unpin,
>   	.get_sg_table = drm_gem_shmem_object_get_sg_table,
> -	.vmap = amdxdna_gem_obj_vmap,
> -	.vunmap = amdxdna_gem_obj_vunmap,
> +	.vmap = drm_gem_shmem_object_vmap,
> +	.vunmap = drm_gem_shmem_object_vunmap,
>   	.mmap = amdxdna_gem_obj_mmap,
>   	.vm_ops = &drm_gem_shmem_vm_ops,
>   	.export = amdxdna_gem_prime_export,
> @@ -692,7 +682,7 @@ amdxdna_drm_create_dev_heap(struct drm_device *dev,
>   	abo->mem.dev_addr = client->xdna->dev_info->dev_mem_base;
>   	drm_mm_init(&abo->mm, abo->mem.dev_addr, abo->mem.size);
>   
> -	ret = drm_gem_vmap(to_gobj(abo), &map);
> +	ret = amdxdna_gem_obj_vmap(abo, &map);
>   	if (ret) {
>   		XDNA_ERR(xdna, "Vmap heap bo failed, ret %d", ret);
>   		goto release_obj;
> @@ -770,7 +760,7 @@ amdxdna_drm_create_cmd_bo(struct drm_device *dev,
>   	abo->type = AMDXDNA_BO_CMD;
>   	abo->client = filp->driver_priv;
>   
> -	ret = drm_gem_vmap(to_gobj(abo), &map);
> +	ret = amdxdna_gem_obj_vmap(abo, &map);
>   	if (ret) {
>   		XDNA_ERR(xdna, "Vmap cmd bo failed, ret %d", ret);
>   		goto release_obj;

Re: [RESEND PATCH V1] accel/amdxdna: Call dma_buf_vmap_unlocked() for imported object

Posted by Lizhi Hou 2 weeks, 1 day ago

On 9/16/25 08:55, Falkowski, Maciej wrote:
> On 9/15/2025 6:10 PM, Lizhi Hou wrote:
>
>> In amdxdna_gem_obj_vmap(), calling dma_buf_vmap() triggers a kernel
>> warning if LOCKDEP is enabled. So for imported object, use
>> dma_buf_vmap_unlocked(). Then, use drm_gem_vmap() for other objects.
>> The similar change applies to vunmap code.
>>
>> Fixes: bd72d4acda10 ("accel/amdxdna: Support user space allocated 
>> buffer")
>> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
>> ---
>>   drivers/accel/amdxdna/amdxdna_gem.c | 38 +++++++++++------------------
>>   1 file changed, 14 insertions(+), 24 deletions(-)
>>
>> diff --git a/drivers/accel/amdxdna/amdxdna_gem.c 
>> b/drivers/accel/amdxdna/amdxdna_gem.c
>> index d407a36eb412..50950be189ae 100644
>> --- a/drivers/accel/amdxdna/amdxdna_gem.c
>> +++ b/drivers/accel/amdxdna/amdxdna_gem.c
>> @@ -392,35 +392,25 @@ static const struct dma_buf_ops 
>> amdxdna_dmabuf_ops = {
>>       .vunmap = drm_gem_dmabuf_vunmap,
>>   };
>>   -static int amdxdna_gem_obj_vmap(struct drm_gem_object *obj, struct 
>> iosys_map *map)
>> +static int amdxdna_gem_obj_vmap(struct amdxdna_gem_obj *abo, struct 
>> iosys_map *map)
>>   {
>> -    struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
>> -
>>       iosys_map_clear(map);
>>   -    dma_resv_assert_held(obj->resv);
>> -
>>       if (is_import_bo(abo))
>> -        dma_buf_vmap(abo->dma_buf, map);
>> -    else
>> -        drm_gem_shmem_object_vmap(obj, map);
>> -
>> -    if (!map->vaddr)
>> -        return -ENOMEM;
>> +        return dma_buf_vmap_unlocked(abo->dma_buf, map);
> Hi,
>
> The dma_buf_vmap_unlocked() will call iosys_map_clear at its start so 
> that
> in this case it will be called twice. Probably it will be optimize 
> out, but maybe
> its something to better omit.
Ok. I will remove iosys_map_clear. The caller uses cleared map even 
drm_gem_vmap might not clear it.
>>   -    return 0;
>> +    return drm_gem_vmap(to_gobj(abo), map);
>>   }
>>   -static void amdxdna_gem_obj_vunmap(struct drm_gem_object *obj, 
>> struct iosys_map *map)
>> +static void amdxdna_gem_obj_vunmap(struct amdxdna_gem_obj *abo, 
>> struct iosys_map *map)
>>   {
>> -    struct amdxdna_gem_obj *abo = to_xdna_obj(obj);
>> -
>> -    dma_resv_assert_held(obj->resv);
>> +    if (iosys_map_is_null(map))
>> +        return;
>>         if (is_import_bo(abo))
>> -        dma_buf_vunmap(abo->dma_buf, map);
>> -    else
>> -        drm_gem_shmem_object_vunmap(obj, map);
>> +        return dma_buf_vunmap_unlocked(abo->dma_buf, map);
> I do also wonder what is the convention here on clearing iosys_map 
> when returning.
> The function drm_gem_vunmap will clear the map for callers while the 
> other not.
> I think at least comment explaining the logic will be necessary.

The input map will not be re-used, so it does not mater if it is cleared 
or not when returning. I will add a comment.

Thanks,

Lizhi

>
> Best regards,
> Maciej
>
>> +
>> +    return drm_gem_vunmap(to_gobj(abo), map);
>>   }
>>     static struct dma_buf *amdxdna_gem_prime_export(struct 
>> drm_gem_object *gobj, int flags)
>> @@ -468,7 +458,7 @@ static void amdxdna_gem_obj_free(struct 
>> drm_gem_object *gobj)
>>       if (abo->type == AMDXDNA_BO_DEV_HEAP)
>>           drm_mm_takedown(&abo->mm);
>>   -    drm_gem_vunmap(gobj, &map);
>> +    amdxdna_gem_obj_vunmap(abo, &map);
>>       mutex_destroy(&abo->lock);
>>         if (is_import_bo(abo)) {
>> @@ -489,8 +479,8 @@ static const struct drm_gem_object_funcs 
>> amdxdna_gem_shmem_funcs = {
>>       .pin = drm_gem_shmem_object_pin,
>>       .unpin = drm_gem_shmem_object_unpin,
>>       .get_sg_table = drm_gem_shmem_object_get_sg_table,
>> -    .vmap = amdxdna_gem_obj_vmap,
>> -    .vunmap = amdxdna_gem_obj_vunmap,
>> +    .vmap = drm_gem_shmem_object_vmap,
>> +    .vunmap = drm_gem_shmem_object_vunmap,
>>       .mmap = amdxdna_gem_obj_mmap,
>>       .vm_ops = &drm_gem_shmem_vm_ops,
>>       .export = amdxdna_gem_prime_export,
>> @@ -692,7 +682,7 @@ amdxdna_drm_create_dev_heap(struct drm_device *dev,
>>       abo->mem.dev_addr = client->xdna->dev_info->dev_mem_base;
>>       drm_mm_init(&abo->mm, abo->mem.dev_addr, abo->mem.size);
>>   -    ret = drm_gem_vmap(to_gobj(abo), &map);
>> +    ret = amdxdna_gem_obj_vmap(abo, &map);
>>       if (ret) {
>>           XDNA_ERR(xdna, "Vmap heap bo failed, ret %d", ret);
>>           goto release_obj;
>> @@ -770,7 +760,7 @@ amdxdna_drm_create_cmd_bo(struct drm_device *dev,
>>       abo->type = AMDXDNA_BO_CMD;
>>       abo->client = filp->driver_priv;
>>   -    ret = drm_gem_vmap(to_gobj(abo), &map);
>> +    ret = amdxdna_gem_obj_vmap(abo, &map);
>>       if (ret) {
>>           XDNA_ERR(xdna, "Vmap cmd bo failed, ret %d", ret);
>>           goto release_obj;