Moving Onto Mirage
Git Your Unikernel Here!

For a little while I’ve had this site running as a MirageOS unikernel, shadowing the main site hosted on GitHub. I’ve finally decided to make the switch, as part of moving over to take advantage of Mirage’s DNS and TLS libraries.

Following the usual pattern, as previously explained by Amir, Mindy and others, the process is:

  • Construct a static Jekyll site.
  • Write a Travis YAML file to cause Travis to build the unikernel image and commit it back to the deployment repository.
  • Write a Git post-merge hook for the deployment repository, so that the latest unikernel is automatically booted when a merge is detected, i.e., there is a new unikernel image.
  • Write a cron job that periodically polls the deployment repository, pulling any changes.

Building a Jekyll site is well-documented – I did find that I had to tweak my _config.yml so as to make sure my local toolchain matched the one used by Github, ensuring consistency between versions of the site. For convenience:

make site

Bringing up the network

The .travis.yml file then specifies the three main targets for the CI test build to carry out: Unix with a standard sockets backed (MIRAGE_BACKEND=unix, MIRAGE_NET=socket) and with the Mirage network stack (MIRAGE_BACKEND=unix, MIRAGE_NET=direct), and with the Xen backend (MIRAGE_BACKEND=xen). For the latter case, we must also specify the static IP configuration to be used (MIRAGE_ADDR, ..._GWS, and ..._MASK). The script then calls the standard skeleton script after first building the site content using Jekyll.

This tests the three basic combinations of network backend for a Mirage appliance:

$ make configure.socket build
  • UNIX/socket requires no configuration. The network device is configured with the loopback address, Appliances can be run without requiring root privileges, assuming they only bind to non-privileged ports.
$ make build
  • UNIX/direct/dhcp requires no configuration if a DHCP server is running and can respond. The appliance must be run with root privileges to use the new network bridging capability of OSX 10.10, whereupon the DHCP client in the appliance follows the usual protocol.
$ make configure.xen build \
  ADDR="" GWS="" MASK=""
  • Xen uses the Mirage network stack and expects static configuration of the network device.

Using Travis CI

Of course, all that is for local development – for the live site, this is actually all wrapped up using Travis CI. Due to a small pull request waiting on the OCaml Travis CI skeleton scripts and a few Mirage releases currently being readied, this looks a little more complex than it needs to (the FORK_USER and DEV_REMOTE variables shouldn’t need to be specified in the long run) but anyway:

language: c
script: bash -ex
  - FORK_USER=mor1 DEV_REMOTE=git://
  - FORK_USER=mor1 DEV_REMOTE=git://
  - FORK_USER=mor1 DEV_REMOTE=git://
    XENIMG=mortio MIRDIR=_mirage DEPLOY=1

This uses the local .travis-sh script to build the three versions of the site, using the Mirage development OPAM repository so as to pick up the latest versions of all the various packages, and updating the Travis gcc and binutils to ensure the stubs for a couple of packages (notably mirage-entropy-xen) build.

Next stop: adding TLS and DNS support…

Copyright © 2009–2016 Richard Mortier