On 12/11/25 01:11, Val Packett wrote:
>
> On 12/10/25 1:51 PM, Dmitry Osipenko wrote:
>> Hi,
>>
>> On 12/10/25 18:39, Val Packett wrote:
>>> This functionality was added for using a KMS-only virtgpu with a
>>> physical
>>> (or SR-IOV) headless GPU in passthrough, but it should not be restricted
>>> to KMS-only mode. It can be used with cross-domain to pass guest memfds
>>> to the host compositor with zero copies (using udmabuf on both sides).
>>>
>>> Drop the check for the absence of virgl_3d to allow for more use cases.
>>>
>>> Fixes: ca77f27a2665 ("drm/virtio: Import prime buffers from other
>>> devices as guest blobs")
>>> Signed-off-by: Val Packett <[email protected]>
>>> ---
>>>
>>> Hi. I couldn't find any comments on that line in the reviews (on
>>> patchwork), so I assume
>>> there was never a specific technical reason for that check, just an
>>> abundance of caution?
>>>
>>> BTW, while here.. The drm_gem_prime_import "fallback" seems pretty
>>> much equivalent to
>>> `return (-ENODEV)`, as drm_gem_prime_import(_dev) just translates the
>>> call to
>>> gem_prime_import_sg_table which we don't use. Should it be replaced
>>> with `return (-ENODEV)`?
>> Returning -ENODEV should break dmabuf self-importing where virtio-gpu
>> driver export dmabuf and then imports to itself.
>
> Hm, I don't think so because the self-import case (for when `buf->ops ==
> &virtgpu_dmabuf_ops`) is handled right here, right above this check.
>
> drm_gem_prime_import would handle the self-import for `buf->ops ==
> &drm_gem_prime_dmabuf_ops` which shouldn't be the case since we have
> `virtgpu_dmabuf_ops`..
Indeed, then it should be redundant. Practically, there should be no
difference with removal of drm_gem_prime_import() and
virtgpu_gem_prime_import_sg_table() from the code. Not sure if changing
it will be worthwhile.
--
Best regards,
Dmitry
