libvirtd #

SR-IOV Virtual Interfaces #

I’ve experimented with the SR-IOV functionality of my Supermicro X10SDV’s network interfaces a little. Virtual machines using these virtual function interfaces do not reach the same speeds between different virtual machines on the same hypervisor as is the case with macvtap networking. In return however, host to guest networking just works without any strage workarounds.

Enable #

First you’ll need to enable this function in the BIOS and in the kernel. I am not sure if the following kernel arguments are only necessary for classic PCI-E passthrough or for SR-IOV in general. But I’ve added these to /etc/defaults/grub:

intel_iommu=on iommu=pt

You can now add virtual functions by writing a number to /sys/class/net/<interface>/device/sriov_numvfs.

Tame NetworkManager #

By default, NetworkManager will try to manage each new virtual function automatically so you will receive lots and lots of IP adresses on interfaces like eno4v1 etc. In order to prevent that, you can add a [keyfile] section in /etc/NetworkManager/NetworkManager.conf:

# prevent networkmanager on sr-iov virtual functions

Alternatively, you can prevent “driver probing” on the new virtual functions, so they will not appear as network interfaces to the host at all. This needs to be done before adding virtual functions, obviously.

echo 0 > /sys/class/net/<interface>/device/sriov_drivers_autoprobe

Add on Startup #

Somehow, there does not appear to be a method to add functions on startup besides manually scripting writes to the sysfs. Since rc.local is not reliable in times of systemd – it especially lacks any ordering guarantees – I wrote a simple oneshot service file interface-vf@.service:

Description = initialize sr-iov virtual functions on %i interface
After = systemd-udevd.service

Type = oneshot
RemainAfterExit = true

# disable driver probing and add virtual functions
ExecStart = sh -c "echo  0 >/sys/class/net/%i/device/sriov_drivers_autoprobe"
ExecStart = sh -c "echo 10 >/sys/class/net/%i/device/sriov_numvfs"

# remove virtual functions
ExecStop  = sh -c "echo  0 >/sys/class/net/%i/device/sriov_numvfs"

This service can be required by libvirtd.service with an override, so the virtual functions are created before starting the virtual machine manager.

Shared Filesystems with virtio-fs #

After moving away from an ESXi hypervisor to a Linux box with KVM, I looked for possibilities of sharing parts of the host filesystem with the guest machines. virtio-fs is a relatively new approach, which replaces shared folders based on the 9p filesystem. Instead of using network filesystem semantics over a virtualized pipe, virtio-fs uses shared memory directly.

In order to use it, you need very recent versions of QEMU and libvirt. Neither CentOS 8 nor Ubuntu 20.04 provide sufficiently recent packages at the moment. However, there is a PPA for Ubuntu at ppa:jacob/virtualisation.

An upstream guide is available on

The XML snippet required in virt-manager for this new filesystem share looks like this:

<filesystem type="mount" accessmode="passthrough">
  <driver type="virtiofs"/>
  <source dir="/path/to/shared/directory"/>
  <target dir="hostshare"/>

You also need to add a NUMA node for the shared memory region. I think there is a better / more automatic way. But this works for now:

<cpu mode="custom" match="exact" check="none">
  <model fallback="forbid">qemu64</model>
  <topology sockets="1" dies="1" cores="2" threads="1"/>
    <cell id="0" cpus="0-1" memory="2097152" unit="KiB" memAccess="shared"/>

Inside the virtual machine, you can mount the share with:

mount -t virtiofs hostshare /mnt