Up until the middle of 2019 it was very unusual to even expect that any CI service would allow nested virtualization. Those who required such functionality had to maintain their own CI runners on their own infrastructure. Things changed when Google Cloud introduced nested KVM support.
Cirrus CI was probably the first CI service to officially support nested virtualization in free tier. There are reports that Travis CI currently also provides such feature but no public announcement has been made yet.
It turns out I was the first person to try using Libvirt in Cirrus CI (I've even hit a previously unknown bug which was promptly fixed by their staff). Since the process has some subtle differences to the popular documented use cases I've decided to describe it here.
Cirrus CI uses Docker images as environment for their runners. It significantly simplifies the setup and enables efficient caching between runs.
Since popular Docker images do not include any hypervisor packages we need to
build our own image. I've decided to add the required packages to Debian base
image. The whole Dockerfile is essentially one
Keep in mind that libvirt package in Debian drops root privileges when
qemu-kvm. You'll either need to disable that in
/etc/libvirt/qemu.conf (as I did) or to change permissions for
allow access by
Required system services
Default entry point for CI runner is not customizable in Cirrus CI - it's an agent process that communicates with CI service and sends progress reports you see in web interface. Because of that no systemd units are started automatically as it would have been the case on a normal system. More than that, starting systemd manually also looks impossible.
That means all the daemons required by libvirt must be started manually (see documentation on background_script syntax):
# .cirrus.yml dbus_background_script: - mkdir -p /var/run/dbus - /usr/bin/dbus-daemon --system --nofork --nopidfile virtlogd_background_script: - /usr/sbin/virtlogd libvirtd_background_script: - sleep 2 && /usr/sbin/libvirtd
Hypervisor kernel is provided as is, and it currently runs legacy iptables firewall. Trying to use iptables-nft (which is the default in current Debian) produces a misconfigured guest network that is hard to debug.
That's why we need to tell Debian to use legacy iptables interface across the whole system:
# .cirrus.yml iptables_legacy_script: - update-alternatives --set iptables /usr/sbin/iptables-legacy
That's it! Following these steps I was able to execute Libvirt (via Vagrant-Libvirt, via Molecule) in Cirrus CI environment. Full configuration is available here, it includes some extra caching steps and many debug statements that helped me to implement this process in the first place.