Docker in QEMU/KVM

Docker in QEMU/KVM #

Some applications may require a properly isolated Docker engine where users of the API have every freedom but when they must not be able to compromise the host security. Since access to the Docker socket is equivalent to being root (or worse) we must preferably run the engine on a seperate machine.

Long story short: virtualization with QEMU/KVM provides all the required isolation and CoreOS is easy to deploy and bundles Docker by default.

The following steps are designed for a CentOS 7 hypervisor.

Prerequisites #

First of all, we need to prepare our hypervisor, so install QEMU and libvirt.

yum install qemu-kvm libvirt virt-install
modprobe kvm
systemctl enable --now libvirtd

Make sure you have hardware virtualization available. If you’re running in a virtual machine already you may need to enable passthrough explicitly. On an Intel machine you should have a module kvm_intel loaded as well.

We are going to use virt-install as well, however the version in EPEL is not recent enough to use the kernel= andinitrd= arguments with --location. Thus prefer a local manager and append --connect qemu+ssh://root@hypervisor/system to virsh or virt-install commands.

Boot a CoreOS Virtual Machine #

There is a guide on how to boot CoreOS with libvirt but I prefer to perform a clean installation to disk. Therefore we need to boot CoreOS to RAM and deploy using an Ignition configuration. Preferably, this is done with the provided PXE images.

Depending on your version of virt-install there are different installation methods available: in the terminal via text console or remotely over VNC.

Via Text Console (modern virt-install) #

If you don’t want to bother with VNC connections and would prefer to install via a text console on the hypervisor itself, you can download and run the CoreOS vmlinuz and cpio.gz directly by specifying them in the --location argument:

virt-install --name core --memory 2048 --vcpus 2 \
  --accelerate --rng /dev/urandom --autostart --graphics none \
  --disk size=20,bus=virtio --os-variant virtio26 \
  --location ",kernel=coreos_production_pxe.vmlinuz,initrd=coreos_production_pxe_image.cpio.gz" \
  --extra-args "coreos.autologin console=ttyS0"

If you prefer to download and verify an ISO locally instead, you can substitute the --location argument:

  --location ../path/to/coreos.iso,kernel=/coreos/vmlinuz,initrd=/coreos/cpio.gz \

This is useful when you’re doing many installs to avoid the repeated downloads.

You can find files inside an ISO with isoinfo -Jf -i /path/to/disc.iso.

Via Text Console (older virt-install) #

Older versions of virt-install – among them version 1.5.0 that is shipped with CentOS 7 – do not support the --location ...,kernel=...,initrd=... syntax and complain about unreachable URLs. In this case you can download the files and fake a Debain installation directory that is autodetected simply by passing the directory path to virt-install.

Download and verify the PXE image as per the CoreOS docs:

cd /var/lib/libvirt/images
mkdir -p coreos && cd coreos
wget $stable/coreos_production_pxe.vmlinuz
wget $stable/coreos_production_pxe.vmlinuz.sig
wget $stable/coreos_production_pxe_image.cpio.gz
wget $stable/coreos_production_pxe_image.cpio.gz.sig
gpg --verify coreos_production_pxe.vmlinuz.sig
gpg --verify coreos_production_pxe_image.cpio.gz.sig

Create a fake MANIFEST and a directory structure that mimics a Debian netboot installer:

mkdir -p amd64/current/images/netboot/debian-installer/amd64/
echo debian-installer > amd64/current/images/MANIFEST
ln -sr coreos_production_pxe.vmlinuz amd64/current/images/netboot/debian-installer/amd64/linux
ln -sr coreos_production_pxe_image.cpio.gz amd64/current/images/netboot/debian-installer/amd64/initrd.gz

Pass the amd64 subdirectory as the installer location:

virt-install --name core --memory 2048 --vcpus 2 \
  --accelerate --rng /dev/urandom --autostart --graphics none \
  --disk size=20,bus=virtio --os-variant virtio26 \
  --location /var/lib/libvirt/images/coreos/amd64 \
  --extra-args "coreos.autologin console=ttyS0"

This method is probably useful for other distributions that don’t get detected automatically either as well.

Via VNC Viewer #

Sometimes an installer may just refuse to start on the serial console or you’re more confident in a graphical installer. This method also applies when you want to use an ISO image without specifying additional kernel parameters. As an example, this section uses an image of, which can be used to interactively boot many different distributions.

First, download the image:

cd /var/lib/libvirt/boot
curl -LO

Now create the virtual machine with virt-install, specifying the ISO with the --cdrom argument:

virt-install --name core --memory 2048 --vcpus 2 \
  --accelerate --rng /dev/urandom --autostart \
  --disk size=20,bus=virtio --os-variant virtio26 \
  --graphics vnc,listen= --noautoconsole \
  --cdrom /var/lib/libvirt/boot/

This should start the installation process and enable a VNC console. You can check the port with virsh vncdisplay runner and verify with ss -tln if in doubt. In my case a default of :0 corresponds to port 5900 on the host, so temporarily open that port in the firewall:

firewall-cmd --add-port 5900/tcp

Connect with your favourite VNC client and complete the installation.

You can’t currently change the keyboard map on the console. Set a password with sudo passwd core and connect with ssh instead if you run into problems.

Install CoreOS to Disk #

Prepare Ignition #

By now you should have prepared an Ignition configuration. There is of course a lot of variation possible here but most importantly you should enable rngd.service and docker.service and make sure that you can connect with SSH public keys. Mine looks somewhat like this:

# enable docker service
    - name: rngd.service
      enabled: yes
    - name: docker.service
      enabled: yes

# ssh public keys
    - name: core
        - # add your keys here

# automatic updates during maintenance window
  reboot_strategy: reboot
  window_start: 04:00
  window_length: 3h

# enable console autologin
    - name: OEM
        device: /dev/disk/by-label/OEM
        format: ext4
    - filesystem: OEM
      path: /grub.cfg
      mode: 0644
      append: true
        inline: |
          set linux_append="$linux_append coreos.autologin"

Installation #

After transpiling, I am using to host small static files quickly. Download the configuration and finally install CoreOS to disk:

curl -LO ""
sudo coreos-install -d /dev/vda -i docker.json
sudo udevadm settle
sudo reboot

Miscellaneous #

Fixed DHCP Address #

You can add a fixed address for this virtual machine by creating an IP assignment for its MAC address with virsh:

virsh net-dhcp-leases default   # see current leases
virsh net-update default add-last ip-dhcp-host \
  --xml "<host mac='52:54:00:e7:b6:4d' ip='' />" \
  --live --config

SSH Client Configuration #

Add an appropriate SSH config on the hypervisor:

Host runner
  User core
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null