Xen project Mailing List

Thank you very much for all your answers.

Среда, 8 января 2020, 10:54 +03:00 от Oleksandr Andrushchenko <oleksandr_andrushchenko@xxxxxxxx>:

On 1/6/20 10:38 AM, Jürgen Groß wrote:
> On 06.01.20 08:56, Santucco wrote:
>> Hello,
>>
>> I’m trying to use vdispl interface from PV OS, it doesn’t work.
>> Configuration details:
>>      Xen 4.12.1
>>      Dom0: Linux 4.20.17-gentoo #13 SMP Sat Dec 28 11:12:24 MSK 2019
>> x86_64 Intel(R) Celeron(R) CPU N3050 @ 1.60GHz GenuineIntel GNU/Linux
>>      DomU: x86 Plan9, PV
>>      displ_be as a backend for vdispl and vkb
>>
>> when VM starts, displ_be reports about an error:
>> gnttab: error: ioctl DMABUF_EXP_FROM_REFS failed: Invalid argument
>> (displ_be.log:221)
>>
>> related Dom0 output is:
>> [ 191.579278] Cannot provide dma-buf: use_ptemode 1
>> (dmesg.create.log:123)
>
> This seems to be a limitation of the xen dma-buf driver. It was written
> for being used on ARM initially where PV is not available.
This is true and we never tried/targeted PV domains with this
implementation,
so if there is a need for that someone has to take a look on the proper
implementation for PV…

Have I got your right and there is no the proper implementation :-)?

>
> CC-ing Oleksandr Andrushchenko who is the author of that driver. He
> should be able to tell us what would be needed to enable PV dom0.
>
> Depending on your use case it might be possible to use PVH dom0, but
> support for this mode is "experimental" only and some features are not
> yet working.
>
Well, one of the workarounds possible is to drop zero-copying use-case
(this is why display backend tries to create dmu-bufs from grants passed
by the guest domain and fails because of "Cannot provide dma-buf:
use_ptemode 1")
So, in this case display backend will do memory copying for the incoming
frames
and won't touch DMABUF_EXP_FROM_REFS ioctl.
To do so just disable zero-copying while building the backend [1]

Thanks, I have just tried the workaround. The backend has failed in an other place not corresponding with dma_buf.
Anyway it is enough to continue debugging my frontend implementation.

Do you know how big is performance penalty in comparison with the zero-copy variant?

Does it make a sense if I make a dedicated HVM domain with linux only for the purpose of vdispl and vkbd backends? Is there a hope this approach will work?

>
> Juergen
>
[1] https://github.com/xen-troops/displ_be/blob/master/CMakeLists.txt#L12

Best regards,

Alexander Sychev

Re: [Xen-devel] PV DRM doesn't work without auto_translated_physmap feature in Dom0