Virtual Function I/O Mediated devices (vfio-mdev) allows a physical GPU to be split into multiple virtual GPUs (vGPU). Such a vGPU can then be assigned to a virtual machine or a container.
Contrary to paravirtualized GPUs (e.g. virtio-gpu), vGPUs can use the same driver as their parent GPU (e.g. a guest compatible with an Intel GPUs will be able to leverage an Intel-based vGPUs)
Although these instructions only cover Intel GPUs compatible with vfio-mdev (5th to 10th generation), Intel GVT-g, in Intel's lingo, it might work similarly with compatible Nvidia or AMD GPUs. On Intel GPUs, since generation 11th, vfio-mdev has been superseded by SR-IOV
Ensure that your GPU supports vfio-mdev
Install the Phyllome OS Desktop II
Make sure the GRUB has been updated after the first boot:
# grub2-mkconfig -o /boot/efi/EFI/fedora/grub.cfg
List all available vGPUs types:
$ mdevctl types
0000:00:02.0
i915-GVTg_V5_1
Available instances: 1
Device API: vfio-pci
Description: low_gm_size: 512MB, high_gm_size: 2048MB, fence: 4, resolution: 1920x1200, weight: 16
i915-GVTg_V5_2
Available instances: 2
Device API: vfio-pci
Description: low_gm_size: 256MB, high_gm_size: 1024MB, fence: 4, resolution: 1920x1200, weight: 8
i915-GVTg_V5_4
Available instances: 5
Device API: vfio-pci
Description: low_gm_size: 128MB, high_gm_size: 512MB, fence: 4, resolution: 1920x1200, weight: 4
i915-GVTg_V5_8
Available instances: 7
Device API: vfio-pci
Description: low_gm_size: 64MB, high_gm_size: 384MB, fence: 4, resolution: 1024x768, weight: 2
Allocating more memory to the GPU in the platform firmware may increase the number of vGPUs one can create. See Troubleshooting section below
In the example above, the i915-GVTg_V5_4 virtual type seems to offer the best trade-offs between the available resolution and the number of available instances.
$ uuidgen
7686131b-b229-4768-a02c-35d1dbed7c66
i915-GVTg_V5_4 using the previously generated UUID# mdevctl start --uuid 7686131b-b229-4768-a02c-35d1dbed7c66 -p 0000:00:02.0 --type i915-GVTg_V5_4
# mdevctl define --uuid 7686131b-b229-4768-a02c-35d1dbed7c66
# mdevctl modify --uuid 7686131b-b229-4768-a02c-35d1dbed7c66 --auto
$ mdevctl list --defined
7686131b-b229-4768-a02c-35d1dbed7c66 0000:00:02.0 i915-GVTg_V5_4 auto (active)
virtio-gpu and set the last one to the none.<domain type="kvm">
[...]
<device>
[...]
<video>
<model type="none"/>
</video>
[...]
</device>
[...]
</domain>
uuid matches the previously generated UUID.<domain type="kvm">
[...]
<device>
[...]
<hostdev mode="subsystem" type="mdev" managed="no" model="vfio-pci" display="off" ramfb="off">
<source>
<address uuid="7686131b-b229-4768-a02c-35d1dbed7c66"/>
</source>
</hostdev>
[...]
</device>
[...]
</domain>
<device>
[...]
<graphics type="spice">
<listen type="none"/>
<gl enable="yes" rendernode="/dev/dri/by-path/pci-0000:00:02.0-render"/>
</graphics>
[...]
</device>
When associated with OpenGL, Spice will only work locally
<domain type="kvm">
[...]
<device>
[...]
<hostdev mode="subsystem" type="mdev" managed="no" model="vfio-pci" display="on" ramfb="on">
<source>
<address uuid="7686131b-b229-4768-a02c-35d1dbed7c66"/>
</source>
</hostdev>
[...]
</device>
[...]
</domain>
RAMFB is set to on, which activates Drect Memory Access Buffers (DMA-BUFs), making the output of a virtual monitor available before the guest operating system takes over
$ mdevctl list -d
5d189f35-d83b-48a8-b8d1-ac1684873493 0000:00:02.0 i915-GVTg_V5_4 auto (active)
7686131b-b229-4768-a02c-35d1dbed7c66 0000:00:02.0 i915-GVTg_V5_4 auto (active)
90f22bd6-d269-4afd-9560-ece3c6e93b0b 0000:00:02.0 i915-GVTg_V5_4 auto (active)
7686131b-b229-4768-a02c-35d1dbed7c66# mdevctl stop --uuid 7686131b-b229-4768-a02c-35d1dbed7c66
# mdevctl undefine --uuid 7686131b-b229-4768-a02c-35d1dbed7c66
$ mdevctl list -d
5d189f35-d83b-48a8-b8d1-ac1684873493 0000:00:02.0 i915-GVTg_V5_4 auto (active)
90f22bd6-d269-4afd-9560-ece3c6e93b0b 0000:00:02.0 i915-GVTg_V5_4 auto (active)
On some computers, it is possible to increase the system memory allocated to the integrated GPU. By doing so, you may be able to create more vGPUs.
Before the host operating system boots up, enter the BIOS/UEFI and look for a setting called GPU aperture size or GPU shared memory
Use the highest possible value, but not higher than the available system memory
System memory will be reserved for the GPU, so make sure you have enough system memory to accomodate both the GPU and your operating system. For instance, if you have a total of 16GB of system memory, it is recommended to allocate no more than 4GB to the GPU
In the example below, two vGPUs are inactive. Attempts to enable them are failing.
$ mdevctl list -d
5d189f35-d83b-48a8-b8d1-ac1684873493 0000:00:02.0 i915-GVTg_V5_4 auto (active)
7686131b-b229-4768-a02c-35d1dbed7c66 0000:00:02.0 i915-GVTg_V5_4 auto (active)
8aec14d3-cb64-40ff-8347-9a872180f4f2 0000:00:02.0 i915-GVTg_V5_4 auto
90f22bd6-d269-4afd-9560-ece3c6e93b0b 0000:00:02.0 i915-GVTg_V5_4 auto (active)
a4f23c92-5889-4ce9-b4a6-4d2b786a6b85 0000:00:02.0 i915-GVTg_V5_4 auto
In the said configuration, there is 2GB of GPU shared allocated to the physical GPU. As each of such vGPUs type requires 512MB, it cannot equate or exceed the 2GB of shared GPU memory.