Experimenting With Virtual GPU Support On Linux 4.10 + Libvirt

[boxie]

As far as I understand, the AMD solution uses sriov and this requires hardware support (you can check it with `lspci`).

Currently, only the FirePro S series have the needed hardware bits.

Pretty sure it is Tonga*(GCN1.2) hardware that has the required sriov bits. Technically I should be able to run it on my R9 380 (and Michael on his R9 285).

The feature is being marketed on pro level hardware (i.e. workstation and server) for VDI style applications.

In a different thread bridgman has said that whether or not consumer level hardware gets it is a AMD marketing team decision.

[JTRealms]

AMD critically need to catch up here, its really the last piece of the virtualization puzzle.

Valve should be putting the pressure on too. A great stand out feature of SteamOS could be, barebones VM's with virtualized GPU's, combined with their in-home streaming stuff, the possiblities are endless.

I guess its just the normies who will miss out anyway, virgl is probably a more likely situation.

[wpoely86]

Pretty sure it is Tonga*(GCN1.2) hardware that has the required sriov bits. Technically I should be able to run it on my R9 380 (and Michael on his R9 285).

Well, it's easy to check. Run `lspci -vvv` (as root) and look for a line like:

        Capabilities: [160 v1] Single Root I/O Virtualization (SR-IOV)
                IOVCap: Migration-, Interrupt Message Number: 000
                IOVCtl: Enable- Migration- Interrupt- MSE- ARIHierarchy+
                IOVSta: Migration-
                Initial VFs: 8, Total VFs: 8, Number of VFs: 0, Function Dependency Link: 00
                VF offset: 128, stride: 4, Device ID: 1520
                Supported Page Size: 00000553, System Page Size: 00000001
                Region 0: Memory at 0000030000000000 (64-bit, prefetchable)
                Region 3: Memory at 0000030000020000 (64-bit, prefetchable)
                VF Migration: offset: 00000000, BIR: 0

If you find that for your AMD gpu, you have the needed bits. At least my R470 doesn't have it.

[boxie]

Well, it's easy to check. Run `lspci -vvv` (as root) and look for a line like:

 Capabilities: [160 v1] Single Root I/O Virtualization (SR-IOV)
IOVCap: Migration-, Interrupt Message Number: 000
IOVCtl: Enable- Migration- Interrupt- MSE- ARIHierarchy+
IOVSta: Migration-
Initial VFs: 8, Total VFs: 8, Number of VFs: 0, Function Dependency Link: 00
VF offset: 128, stride: 4, Device ID: 1520
Supported Page Size: 00000553, System Page Size: 00000001
Region 0: Memory at 0000030000000000 (64-bit, prefetchable)
Region 3: Memory at 0000030000020000 (64-bit, prefetchable)
VF Migration: offset: 00000000, BIR: 0

If you find that for your AMD gpu, you have the needed bits. At least my R470 doesn't have it.

Which kernel version do you need for that capability to be exposed?

[wpoely86]

Which kernel version do you need for that capability to be exposed?

`lspci` just shows you what the pcie card can do. It's independent of the kernel. Unless you have a really old version of `lspci` you should see it (if it's there).

The output I've shown comes from a RHEL 7 box.

[boxie]

`lspci` just shows you what the pcie card can do. It's independent of the kernel. Unless you have a really old version of `lspci` you should see it (if it's there).

The output I've shown comes from a RHEL 7 box.

not showing anything for me. and here I was getting my hopes up

[dagger]

Has anyone tried that setup yet?

Not sure what am I missing:

uname -r
4.10.0-0.rc6.git2.2.fc26.x86_64

lsmod | grep vfio
vfio_mdev              16384  0
mdev                   20480  2 kvmgt,vfio_mdev
vfio_pci               45056  0
vfio_iommu_type1       24576  0
vfio_virqfd            16384  1 vfio_pci
vfio                   28672  4 vfio_iommu_type1,kvmgt,vfio_mdev,vfio_pci
irqbypass              16384  2 kvm,vfio_pci

ls -la /sys/class/mdev_bus    
ls: cannot access '/sys/class/mdev_bus': No such file or directory