Re: [PATCH] drm/virtio: do not enforce blob_alignment in cross-domain contexts

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Critical] drm/virtio: NULL pointer dereference in verify_blob() when 3D features are disabled
- [Medium] drm/virtio: data race when reading vfpriv->context_init in verify_blob()
--

commit e614dba4f51f24b9effac7458783e310e89be777
Author: Val Packett <val@invisiblethingslab.com>
drm/virtio: do not enforce blob_alignment in cross-domain contexts

The commit prevents enforcing blob_alignment requirements in cross-domain
contexts, as resources created there are arbitrary shared system memory files
like Wayland keymaps.

> diff --git a/drivers/gpu/drm/virtio/virtgpu_ioctl.c b/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> --- a/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> +++ b/drivers/gpu/drm/virtio/virtgpu_ioctl.c
> @@ -540,6 +540,8 @@ static int verify_blob(struct virtio_gpu_device *vgdev,
>  	params->blob_hints = rc_blob->blob_hints;
>  
>  	if (vgdev->has_blob_alignment &&
> +	    (vfpriv->context_init & VIRTIO_GPU_CONTEXT_INIT_CAPSET_ID_MASK) !=
> +	    VIRTIO_GPU_CAPSET_CROSS_DOMAIN &&

[Severity: Critical]
If a virtio-gpu device advertises VIRTIO_GPU_F_RESOURCE_BLOB and
VIRTIO_GPU_F_BLOB_ALIGNMENT but not VIRTIO_GPU_F_VIRGL, could this cause a
NULL pointer dereference?

In this scenario, virtio_gpu_driver_open() returns successfully but leaves
file->driver_priv (which is passed as vfpriv) as NULL. When userspace then
calls DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB with VIRTGPU_BLOB_MEM_GUEST,
verify_blob() is reached without aborting on the has_virgl_3d check.

This newly added code will then unconditionally dereference
vfpriv->context_init because vgdev->has_blob_alignment is true, causing a
kernel panic. Is there a way to verify vfpriv is not NULL before accessing
context_init?

[Severity: Medium]
Does this introduce a data race when reading context_init?

This accesses the lock-protected context_init field without acquiring
vfpriv->context_lock. A lockless read of context_init could read a partially
written bitmask if updated concurrently, leading to incorrectly enforcing or
bypassing blob alignment checks and returning spurious -EINVAL to userspace.

This could happen if userspace calls DRM_IOCTL_VIRTGPU_CONTEXT_INIT and
DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB concurrently on the same file
descriptor.

>  	    !IS_ALIGNED(params->size, vgdev->blob_alignment))
>  		return -EINVAL;

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261006004044.2242154-1-val@invisiblethingslab.com?part=1