Re: [PATCH v2 4/4] virtio-gpu: Don't require udmabuf when blob support is enabled

[Dmitry Osipenko <dmitry.osipenko@collabora.com>]

On 9/27/22 11:32, Gerd Hoffmann wrote:
> On Mon, Sep 26, 2022 at 09:32:40PM +0300, Dmitry Osipenko wrote:
>> On 9/23/22 15:32, Gerd Hoffmann wrote:
>>> On Tue, Sep 13, 2022 at 12:50:22PM +0200, Antonio Caggiano wrote:
>>>> From: Dmitry Osipenko <dmitry.osipenko@collabora.com>
>>>>
>>>> Host blobs don't need udmabuf, it's only needed by guest blobs. The host
>>>> blobs are utilized by the Mesa virgl driver when persistent memory mapping
>>>> is needed by a GL buffer, otherwise virgl driver doesn't use blobs.
>>>> Persistent mapping support bumps GL version from 4.3 to 4.5 in guest.
>>>> Relax the udmabuf requirement.
>>>
>>> What about blob=on,virgl=off?
>>>
>>> In that case qemu manages the resources and continued to require
>>> udmabuf.
>>
>> The udmabuf is used only by the blob resource-creation command in Qemu.
>> I couldn't find when we could hit that udmabuf code path in Qemu because
>> BLOB_MEM_GUEST resource type is used only by crosvm+Venus when crosvm
>> uses a dedicated render-server for virglrenderer.
> 
> Recent enough linux guest driver will use BLOB_MEM_GUEST resources
> with blob=on + virgl=off

I reproduced this case today using "-device
virtio-gpu-device,blob=true". You're right, it doesn't work without udmabuf.

[    8.369306] virtio_gpu_dequeue_ctrl_func: 90 callbacks suppressed
[    8.369311] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1200 (command 0x105)
[    8.371848] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1205 (command 0x104)
[    8.373085] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1200 (command 0x105)
[    8.376273] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1205 (command 0x104)
[    8.416972] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1200 (command 0x105)
[    8.418841] [drm:virtio_gpu_dequeue_ctrl_func] *ERROR* response
0x1205 (command 0x104)

I see now why you're wanting to keep the udmabuf requirement for
blob=on,virgl=off.


>>   - /dev/udmabuf isn't accessible by normal user
>>   - udmabuf driver isn't shipped by all of the popular Linux distros,
>> for example Debian doesn't ship it
> 
> That's why blob resources are off by default.
> 
>> Because of all of the above, I don't think it makes sense to
>> hard-require udmabuf at the start of Qemu. It's much better to fail
>> resource creation dynamically.
> 
> Disagree.  When virgl/venus is enabled, then yes, qemu would let
> virglrenderer manage resources and I'm ok with whatever requirements
> virglrenderer has.  When qemu manages resources by itself udmabuf is
> a hard requirement for blob support though.

Let's try to relax the udmabuf requirement only for blob=on,virgl=on.
I'll update this patch.

Thank you very much for the review!

-- 
Best regards,
Dmitry