[libvirt] [PATCH v4 3/3] hw/vfio/display: add ramfb support

[Erik Skultety]

On Thu, Jun 14, 2018 at 12:36:25AM +0200, Gerd Hoffmann wrote:
> On Wed, Jun 13, 2018 at 01:50:47PM -0600, Alex Williamson wrote:
> > On Wed, 13 Jun 2018 10:41:49 +0200
> > Gerd Hoffmann <kraxel at redhat.com> wrote:
> >
> > > So we have a boot display when using a vgpu as primary display.
> > >
> > > Use vfio-pci-ramfb instead of vfio-pci to enable it.
> >
> > Using a different device here seems like it almost guarantees a very
> > complicated path to support under libvirt.  What necessitates this
> > versus a simple ramfb=on option to vfio-pci?
>
> Well, it's simliar to qxl vs. qxl-vga.  It's not qxl,vga={on,off} and
> libvirt has no problems to deal with that ...
>
> Another more technical reason is (again) hotplug.  ramfb needs an fw_cfg
> entry for configuration, and fw_cfg entries can't be hotplugged.  So
> hotplugging vfio-pci with ramfb=on isn't going to fly.  So we need a
> separate device with hotplug turned off.

Well if that's not supposed to work ever, libvirt's hotplug code could format
the following FWIW:
"-device vfio-pci [opts],ramfb=off"

As such, new device wouldn't be that of an issue for libvirt if vfio-pci and
vfio-pci-ramfb are back to back compatible with all the device options that are
available for vfio-pci (I mean in terms of using an mdev). Because in that
case, what libvirt could do is to look whether we're supposed to turn on the
display, if so, then we need support for this in capabilities to query and then
we could prefer this new device over the "legacy" vfio-pci one. However, if we
expect a case where QEMU would succeed to start with an mdev mapped to this
new ramfb device but not with vfio-pci, then that's an issue. Otherwise I don't
necessarily see a problem, if QEMU supports this new device and we need
display, let's use that, otherwise let's use the old vfio-pci device. But I'm
still curious about the ramfb=off possibility I asked above for hotplug
nonetheless.

Thanks,
Erik

>
> > I'm also not sure I understand the usage model, SeaBIOS and OVMF know
> > how to write to this display, but it seems that the guest does not.
>
> Yes.
>
> > I suppose in the UEFI case runtime services can be used to continue
> > writing this display,
>
> Yes.
>
> > but BIOS doesn't have such an option, unless we're somehow emulating
> > VGA here.
>
> vgabios support is in the pipeline, including text mode emulation (at
> vgabios level, direct access to vga window @ 0xa0000 doesn't work).
>
> > So for UEFI, I can imagine this
> > covers us from power on through firmware boot and up to guest drivers
> > initializing the GPU (assuming the vGPU supports a kernel mode driver,
> > does NVIDIA?),
>
> Yes.  Shouldn't matter whenever the driver is kernel or userspace.
>
> > but for BIOS it seems we likely still have a break from
> > the bootloader to GPU driver initialization.
>
> Depends.  vgacon (text mode console) doesn't work.  fbcon @ vesafb works.
>
> > For instance, what driver
> > is used to draw the boot animation (or blue screen) on SeaBIOS Windows
> > VM?
>
> Windows depends on vgabios for that and it works fine.
>
> > I'm assuming that this display and the vGPU display are one in the
> > same, so there's some cut from one to the other.
>
> Yes.  If the vfio query plane ioctl reports a valid guest video mode
> configuration the vgpu display will be used, ramfb otherwise.
>
> cheers,
>   Gerd
>