Re: [PATCH 1/1] drm/virtio: Implement device_attach

[Daniel Vetter <daniel@ffwll.ch>]

On Wed, Jan 10, 2024 at 05:56:28PM +0800, Julia Zhang wrote:
> drm_gem_map_attach() requires drm_gem_object_funcs.get_sg_table to be
> implemented, or else return ENOSYS. Virtio has no get_sg_table
> implemented for vram object. To fix this, add a new device_attach to
> call drm_gem_map_attach() for shmem object and return 0 for vram object
> instead of calling drm_gem_map_attach for both of these two kinds of
> object.
> 
> Signed-off-by: Julia Zhang <julia.zhang@amd.com>
> ---
>  drivers/gpu/drm/virtio/virtgpu_prime.c | 14 +++++++++++++-
>  1 file changed, 13 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/virtio/virtgpu_prime.c b/drivers/gpu/drm/virtio/virtgpu_prime.c
> index 44425f20d91a..f0b0ff6f3813 100644
> --- a/drivers/gpu/drm/virtio/virtgpu_prime.c
> +++ b/drivers/gpu/drm/virtio/virtgpu_prime.c
> @@ -71,6 +71,18 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
>  	drm_gem_unmap_dma_buf(attach, sgt, dir);
>  }
>  
> +static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
> +				     struct dma_buf_attachment *attach)
> +{
> +	struct drm_gem_object *obj = attach->dmabuf->priv;
> +	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
> +
> +	if (virtio_gpu_is_vram(bo))
> +		return 0;

You need to reject attach here because these vram buffer objects cannot be
used by any other driver. In that case dma_buf_attach _must_ fail, not
silently succeed.

Because if it silently succeeds then the subsequent dma_buf_map_attachment
will blow up because you don't have the ->get_sg_table hook implemented.

Per the documentation the error code for this case must be -EBUSY, see the
section for the attach hook here:

https://dri.freedesktop.org/docs/drm/driver-api/dma-buf.html#c.dma_buf_ops

Since you're looking into this area, please make sure there's not other
similar mistake in virtio code.

Also can you please make a kerneldoc patch for struct virtio_dma_buf_ops
to improve the documentation there? I think it would be good to move those
to the inline style and then at least put a kernel-doc hyperlink to struct
dma_buf_ops.attach and mention that attach must fail for non-shareable
buffers.

In general the virtio_dma_buf kerneldoc seems to be on the "too minimal,
explains nothing" side of things :-/

Cheers, Sima

> +
> +	return drm_gem_map_attach(dma_buf, attach);
> +}
> +
>  static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
>  	.ops = {
>  		.cache_sgt_mapping = true,
> @@ -83,7 +95,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
>  		.vmap = drm_gem_dmabuf_vmap,
>  		.vunmap = drm_gem_dmabuf_vunmap,
>  	},
> -	.device_attach = drm_gem_map_attach,
> +	.device_attach = virtgpu_gem_device_attach,
>  	.get_uuid = virtgpu_virtio_get_uuid,
>  };
>  
> -- 
> 2.34.1
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch