QEMU/KVM on Docker and CoreOS
docker pull ulexus/qemu
docker run --privileged ulexus/qemu QEMU_OPTIONS
--privileged is required in order to run with the kernel-level virtualization (kvm) optimization.
For the most part, it is fairly easy to run qemu within docker. The only real hiccup is that /dev/kvm (the device node for the kernel hypervisor access) isn't reissued (or permitted) within docker. That means we have to do two things for basic usage:
- Make the device node
- Execute the docker container with
While this is obviously not ideal, it isn't actually that bad, since you are running a full VM, in the container, which itself should isolate the client.
Basically, this allows me to run QEMU hosts with storage on RBD on my favorite server OS: http://coreos.com.
One of my gripes with Docker right now is that it's not easy for me to manage my own networking. Also, it has abysmal support for things like IPv6. The idea of NATing every connection is repugnant and backward to me. Never-the-less, I understand the motivations.
Still, for my use, I needed the ability to attach to my network bridge instead of the default
docker0 bridge. The entrypoint script allows you to pass the
$BRIDGE_IF environment variable. If set, it will add that bridge interface to the container's
/etc/qemu/bridge.conf file which, in turn, allow your qemu instance to attach to that bridge using the built-in
Note, however, that if you want to do this, you'll need to pass the
--net=host option to your
docker run command, in order to access the host's networking namespace.
Included in this image is support for Ceph/RBD volumes. In order to use Ceph, you should probably bind-mount your
/etc/ceph directory which contains your ceph.conf and client keyring. I use
docker run -v /etc/ceph:/etc/ceph for this purpose on my CoreOS boxes.
NOTE: using qemu's bridge networking with docker's
--net=host with RBD block storage creates switch loops for me, even with STP. Removing any one of those three seems to work fine. To work around this problem, you can also run with
--privileged --pid=host --ipc=host.
Also included in this repo is a service file, suitable for use with systemd (CoreOS and fleet), provided as an example. You'll need to fill in your own values, of course, and customize it to your liking.
The entrypoint script is site-specific for me, but you can override most of it simply by passing arguments to the execution of this container (which will, in turn, be passed as arguments to
If you intend to use the entrypoint script as is, it expects
etcd to be populated with some keys. (
%i below refers to the unit instance, such as
1 for the unit
/kvm/%i/host- Should match host's hostname; used as a mutex for this VM
/kvm/%i/ram- The amount of RAM to allocate to this VM
/kvm/%i/mac- The MAC address to assign to the NIC of this VM
/kvm/%i/rbd- The RBD image (of the form
/kvm/%i/spice_port- The TCP port to use for the spice server
/kvm/%i/extra_flags- (optional) Free-form qemu flags to append