Intel GVT-g (formerly XenGT)

[je]

Hello,

Intel GVT-g is Intels mediated pass-through technology for graphics virtualization. Mediated pass-through allows to share a graphics card with multiple guests [1]. Intel GVT-g was formerly known as XenGT. However, GVT-g was added to the mainline kernel 4.10 [2] and have since seen improvements [3].

Are there any plans to support GVT-g in upcoming Qubes OS releases? Did you do any experiments with GVT-g and QubesOS?

If you did not do any experiments with GVT-g and QubesOS, than I would really like to see a GSoC project which evaluates the Intel GVT-g technology in Qubes OS. Because Intel GVT-g could be maybe used to enable Android and WebGL development in Qubes OS. Furthermore, proper GPU acceleration support could improve or enable many use cases which require GPU acceleration.

[1] https://01.org/igvt-g

[2] http://www.phoronix.com/scan.php?page=news_item&px=Intel-GVT-G-Linux-4.10-State

[3] http://www.phoronix.com/scan.php?page=news_item&px=Intel-GVT-g-Linux-4.12-Slated

[Marek Marczykowski-Górecki]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

On Thu, Apr 20, 2017 at 07:05:40AM -0700, je wrote:
> Hello,
>
> Intel GVT-g is Intels mediated pass-through technology for graphics
> virtualization. Mediated pass-through allows to share a graphics card with
> multiple guests [1]. Intel GVT-g was formerly known as XenGT. However,
> GVT-g was added to the mainline kernel 4.10 [2] and have since seen
> improvements [3].
>
> Are there any plans to support GVT-g in upcoming Qubes OS releases? Did you
> do any experiments with GVT-g and QubesOS?

Currently not. While this may look very attractive, it also have a huge
attack surface - especially the mediating part running in device model.
There are at least two things to be concerned about:
- exploiting some bug there to break out into dom0,
- exploiting some bug there to steal/subvert data of other VM using the
same GPU

The first one could be somehow mitigated by sandboxing it in a separate
VM - like we do with qemu for HVM domains. But it wouldn't prevent the
second kind of attack, which is especially severe if you have only one
GPU.
This also require a lot of research what other parts of the system could
be affected by such a complex feature.

This all doesn't mean we will never add such a feature - using this
technology, or some other future one. But probably it will not be
enabled by default. And surely it will not be in the near future - since
our resources are limited, we focus on things improving security of
Qubes OS, not loosening it.

> If you did not do any experiments with GVT-g and QubesOS, than I would
> really like to see a GSoC project which evaluates the Intel GVT-g
> technology in Qubes OS. Because Intel GVT-g could be maybe used to enable
> Android and WebGL development in Qubes OS. Furthermore, proper GPU
> acceleration support could improve or enable many use cases which require
> GPU acceleration.

Well, this indeed may be a good candidate for GSoC project. I'm somehow
sceptical if this as a whole could be framed as such (IMO it's much more
than 3 months of work), but some parts probably yes. Anyway for this is
for the next year - deadline for projects submission for this year
already have passed.

> [1] https://01.org/igvt-g
>
> [2]
> http://www.phoronix.com/scan.php?page=news_item&px=Intel-GVT-G-Linux-4.10-State
>
> [3]
> http://www.phoronix.com/scan.php?page=news_item&px=Intel-GVT-g-Linux-4.12-Slated
>

- --
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2

iQEcBAEBCAAGBQJY+MrMAAoJENuP0xzK19csm6EH/jhI3O7/l0byXOZdV1jt/Q/f
OpaVZo7DbHLgV6lyQpt++N56jlk+4MBJo1o26qfi6JhItBAMjtS8CGEzWtWVGCiX
/4BQw02rAEqJMhAFE+H7jRamf9TDUCw17PVS//hPs1btW9ccPRls/x7zaHSgRsQ+
Y2GLOGGXvkFygiixzPhR8Bf3pzngWESznF2qRxGMCkycs+068efRTshJ+bhmZkOq
fB5Ml1BO7xdCUWOhgMZmaIRKiiIKqeHVpDpOjrRQXeh6Q1Hx1wRU8bXXZu6GFFdx
LzRfj6464Wr1iKeceZ4hEj2/fjmGH8zyzE/RJiu/GvNhkOGfjCCSsR0BhI1XvnU=
=OCFo
-----END PGP SIGNATURE-----

[je]

I think that GPU virtualization is a very new field. I would not consider this as a feature anytime soon in Qubes OS. However, I think Qubes OS should explore new technologies in this area soon as they appear. Currently it is not really possible to play games, use 3D rendering applications or use WebGL. I tried once to play a simple WebGL based tower game on Qubes OS. It was just painful. Whereas the same game was running very fast on an Alcatel OneTouch FirefoxOS phone (http://www.gsmarena.com/alcatel_one_touch_fire-5319.php). My Intel i7-3x running Qubes OS should have been able to surpass the performance of my phone by far.

What I was thinking about is, to have a page for QubesOS with proposals which can be used for GSoC students, students or researches which have to write a thesis. We could call it Qubes OS Research Lab. The proposals should be around interesting and novel research topics in virtualization, such as GPU virtualization, Unikernels, separation of Desktop Environment and Dom0, introspection/forensics and many more. I think students would be interested to work on topics which allow them to contribute to an open source project and work together with a community.

[Jean-Philippe Ouellet]

On Thu, May 25, 2017 at 9:58 AM, je <j.ep...@openmailbox.org> wrote:
> On Thursday, April 20, 2017 at 10:50:58 AM UTC-4, Marek Marczykowski-Górecki

There is the GSoC Ideas List page [1] which already lists many
projects similar to those you describe. Feel free to suggest more
ideas there in the form of pull requests against [2]. If you think it
should be split into categories (perhaps: "implementation
improvements", and "exploratory work" or so) then feel free to submit
a PR doing that as well.

Cheers,
Jean-Philippe (a person who started contributing to Qubes as a student)

[1]: https://www.qubes-os.org/gsoc/
[2]: https://github.com/QubesOS/qubesos.github.io/blob/master/pages/gsoc.md