[libvirt-users] PCI passthrough fails in virsh: iommu group is not viable

[Alex Holst]

Quoting Laine Stump (laine at laine.org):
> On 08/12/2015 02:34 PM, Alex Holst wrote:
> > I would really appreciate some pointers on what I am doing wrong here.
> >
> > I have a need to run multiple virtual guests which have each their own GPU and
> > some USB controllers passed-through. I am able to run one of the guests like
> > this (assuming vfio stuff has happened elsewhere), but I would prefer to use
> > virsh:
[..]

Thank you for your input. I have been working on this issue on and off
since my original mail to this list.

I have been unable to properly migrate the single VM from a shell
script, much less being able to run several VMs that each have a
pass-through.

As for details missing from my previous mail: This is an Ubuntu 15.04
host running several Windows 10 guests. The entire kvm command line I
have running is from this guide at Puget Systems:

https://www.pugetsystems.com/labs/articles/Multiheaded-NVIDIA-Gaming-using-Ubuntu-14-04-KVM-585/

I have discovered several problems with this guide, in particular that I
can remove the pci_stub ids from /etc/initramfs-tools/modules and the
virtual Windows host continues to work just fine.

So, now I'm back to scratch using virt-install and pointing to
the existing img file that works with the kvm shell script::

$ virt-install --name foo --memory 8192 --machine q35 \
	--host-device 02:00.0 --host-device 02:00.1 \
	--host-device 00:1a.0 --host-device 00:1d.0 \
	--disk /vm2/foo.img --boot menu=on

Starting install...
Creating domain...
Connected to domain foo
Escape character is ^]

Even though the 02:00.0 and 02:00.1 devices are the GPU and on-board
audio, the console remains in text mode and the actual guest OS is
nowhere to be seen.

$ lspci -nn | grep 02:00
02:00.0 VGA compatible controller [0300]: NVIDIA Corporation Device [10de:1401] (rev a1)
02:00.1 Audio device [0403]: NVIDIA Corporation Device [10de:0fba] (rev a1)
$ lspci -nn | egrep 00:1[a,d]
00:1a.0 USB controller [0c03]: Intel Corporation C600/X79 series chipset USB2 Enhanced Host Controller #2 [8086:1d2d] (rev 06)
00:1d.0 USB controller [0c03]: Intel Corporation C600/X79 series chipset USB2 Enhanced Host Controller #1 [8086:1d26] (rev 06)

Do you have any additional pointers for me on how to properly pass the
GPU through so the guest OS detects it and is able to make use of the
attached display?

Thanks,
Alex

> >
> > kvm -M q35 -m 8192 -cpu host,kvm=off \
> >   -smp 4,sockets=1,cores=4,threads=1 \
> >   -bios /usr/share/seabios/bios.bin -vga none \
> >   -device ioh3420,bus=pcie.0,addr=1c.0,multifunction=on,port=1,chassis=1,id=root.1 \
> >   -device vfio-pci,host=02:00.0,bus=root.1,addr=00.0,multifunction=on,x-vga=on \
> >   -device vfio-pci,host=02:00.1,bus=root.1,addr=00.1 \
> >   -device vfio-pci,host=00:1d.0,bus=pcie.0 \
> >   -device vfio-pci,host=00:1a.0,bus=pcie.0 \
> >   -nographic -boot menu=on /vm2/foo.img
> >
> > I found the hardware addresses using lspci.  When I invoke the same machine with virsh with what 
> > I believe are the same addresses,
> 
> 1) The XML that you provide at the end shows you only assigning 02:00.0,
> but not 02:00.1, and you are using managed='yes', so you are depending
> on libvirt to unbind any host driver and bind vfio-pci to the devices.
> If 02:00.0 and 02:00.1 are both in iommu group 18, but one of them isn't
> bound to vfio-pci, qemu will give "some kind of error" when it tries to
> assign any other device in the group.
> 
> 2) your example qemu commandline uses the q35 machinetype, but your
> example libvirt domain uses pc-i440fx-utopic machinetype, so the bus
> structure is completely different. Assuming a recent enough libvirt, you
> should be able to create a virtual machine based on q35 and attach the
> devices to pcie-root (bus 0). You wouldn't be able to add an ioh3420
> controller in unless you are running libvirt built from upstream master
> - support for that was just pushed last weekend. (it is a new controller
> defined like this:
> 
>   <controller type='pci' model='pcie-root-port'/>
> 
> 3) Doing the above *may* eliminate another problem that your current
> definition has - it has an emulated video device at 00:02.0, then the
> passthrough video device at 00:04.0. It's likely that even if you do get
> the domain started up, you'll end up with the passthrough vga as a
> secondary (which I'm guessing isn't what you want).
> 
> 4) I don't know what "x-vga=on" is, but libvirt doesn't directly support
> that. If it is necessary, you would need to add it using libvirt's "qemu
> commandline passthrough":
> 
>  
> http://blog.vmsplice.net/2011/04/how-to-pass-qemu-command-line-options.html
> 
> (you can add a "-set", "device.hostdev0.x-vga=on" I think)
> 
> 5) The libvirt definition you've provided add several other devices that
> you don't have in your qemu commandline which may get in the way of what
> you want to do; in particular the tablet, mouse, and keyboard. You'll
> probably want to trim those out (and if you're passing through USB,
> you'll likely want to replace all of the <controller type='usb'>
> elements with a single <controller type='usb' model='none'/>
> 
> 6) your qemu commandline specifies an exact bios location, which you
> haven't done in the libvirt definition. If that isn't the default bios,
> then you'll want to look into how to specify a non-default bios file
> here: http://www.libvirt.org/formatdomain.html
> 
> 7) If you're going to stick with the pc-ii440fx machinetype and your
> guest doesn't absolutely require a SATA disk, you'd probably be better
> off letting it attach to the 440fx's IDE controller instead (or even
> better, if your guest OS supports the virtio disk driver, use that since
> the performance will be *greatly* improved). (likewise, replace the
> network interface's "model type='e1000'" with "model type='virtio'" if
> you can).
> 
> 
-- 
Alex Holst