Ilija Hadzic's Virtual CRTCs feature discuss

[Daniel Vetter]

On Mon, Aug 18, 2014 at 03:35:16PM -0700, Derek wrote:
> Hi every one
> I'm currently working on VirtualMonitor in my leisure time. It allows you
> to use compute/tablet/smartphone as a second monitor for your primary
> computer. please refer to http://virtualmonitor.github.io for more
> information. Currently I have released very basic version for
> windows2000-windows7 to demonstrate the project is feasible. When I was
> trying to make a further step on windows, I realized it is difficult for an
> individual, as it is not open source and also without technical support
> from Microsoft.
> 
> Then I want to move to linux, and I found Ilija Hadzic's post about Virtual
> CRTCs. his post is here:
> http://lists.freedesktop.org/archives/dri-devel/2011-November/015975.html
> In his implementation, GPU driver can create arbitrary number of CRTCs
> (configurable by user) instead of only those CRTCs that represent real
> hardware.
> It is very useful not only for VirtualMonitor, but also
> VNC/Virtualization/USB display etc. based on this implementation those
> application will be able to take full advantage of the physical graphic
> card(3D acceleration).
> 
> I want raise Ilijia's original question agian, if anybody in this community
> think Virtual CRTCs is useful, and willing to work together to make a
> further progress.
> My thought is if can implement a driver independent layer between dri and
> vendor specific GPU driver, with some general API. maybe implement this
> based one vendor related GPU first? e.g. based on Ilijia's implementation
> for Radeon as a daemon?
> 
> GPU driver is not my expertise, If some expert from this community think
> this feature is interesting, and willing to initiate a project for this
> feature, that will be great. Then people can work and discuss together.
> 
> Any comments regarding to Virtual CRTC or VirtualMonitor are very welcome.
> Thanks.

I think the concept is overall sound (haven't looked at the old patches in
detail).For the actual implementation I think a separate virtual drm driver is now
the better approach, since with dma-buf and soon native fence support we
can do this now properly.

And especially now that we have multi-gpu support in X it should integrate
almost seamlessly into X (and other display managers with multi-gpu
support). Instead of requering special support in all drivers.

Another thing for which iirc no one ever proposed a decent solution is
synchronization to consumers of the frontbuffers for virtual gpus. So I
guess the driver-private ioctl interface to make that magic work will be
key.

A (configurable) virtual gpu should also be really useful for automated
testing, e.g. of hot-plug and unplug (both drm core and userspace). The
v4l folks have such a driver (recently massively revamped for 3.18) and it
looks extremely useful.

For the configuration interface I guess a few simple module options to get
started should be enough, eventually we can switch to configfs (or
similar) to be able to configure and create/destroy virtual gpus at
runtime.

Just my random thoughts, probably good to kickstart the discussion with
some quick patches and chat with people on #dri-devel on freenode irc.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch