Need an option to manage vgpu devices types

[Cedric Lemarchand]

There are 3 options to present a gpu to a guest:

- PCI passthrough
- SR-IOV
- vGPUs

The later is actually not usable with nova-compute charm, as the [devices] section of nova.conf is not handled (see [1] [2]).

A first step, which should be pretty strait forward, would be to allow a single value for "enabled_vgpu_types".

In terms of documentation, if the GPU driver is correctly installed on the hypervisor (qemu-kvm here), vgpu types can be discovered either with "mdevctl types" command or by listing "/sys/bus/pci/devices/0000:XX:YY.0/mdev_supported_types" path of the pci device.

[1] https://docs.openstack.org/nova/ussuri/admin/virtual-gpu.html
[2] https://docs.openstack.org/nova/ussuri/configuration/config.html#devices.enabled_vgpu_types

[Cedric Lemarchand (cedric-lemarchand)]

Also apparmor policy of nova-compute needs to be updated in order for the process to be able to create vgpu via the sysfs. Below a working example added to "/etc/apparmor.d/usr.bin.nova-compute" :

--- snip ---
/sys/devices/pci*/*/*/mdev_supported_types/*/create w,
--- snip ---