One of the really nice things about Docker containers is that the defaults mostly just work. One of those defaults is networking; docker run gives you a perfectly serviceable network experience with containers able to access the Internet, access each other, and expose services. runj is a much lower-level tool than Docker, so that sort of out-of-the-box network experience wouldn’t be something runj would directly provide. However, I recently added support to runj for some of the pieces that make a networking experience like that possible. Higher-level tools that use runj, like nerdctl, might use these pieces in the future.

One of the core use-cases for modern container systems is to run networked workloads, often across a group of machines deployed in a cluster. A variety of different networking models exist, but until now no networking at all was possible with runj. Now, after this change, runj has its first networking capability! The functionality that pull request enable jails to share the IPv4 network stack with the underlying FreeBSD host, similar to the “host networking” model common for Linux containers.

In March, I open-sourced runj, an experimental OCI-compatible runtime for FreeBSD jails. I started runj in order to teach myself more about FreeBSD in general and jails in particular, and the initial contribution policy I set was designed to give me the space to learn at my own pace. However, as I wrote in that first blog post, the amount of attention runj received on its first day really surprised me. The attention since then has continued, and I’ve had the opportunity to connect with members of the FreeBSD community who have shared my excitement about connecting FreeBSD with the broader container ecosystem. Everyone I’ve spoken with has been incredibly kind and respectful of the space I requested, which did give me the opportunity to learn on my own. I really appreciate their kindness, and now that I’ve achieved the first part of my learning goal I’m ready to move forward and work together rather than working alone.

containerd 1.5.0 was released today and now works on a new operating system: FreeBSD! This new release includes a series of patches (1, 2, 3, 4, 5, 6, 7, 8, 9, 10) which allow containerd to build, enable the native and zfs snapshotters, and use a compatible runtime like runj. I’m really excited about this! It’s awesome that only a small amount of work was needed to make containerd compatible with FreeBSD and that so much of it worked straight out of the box. And with a runtime for jails, containerd’s powerful APIs can now be used to manage FreeBSD’s native process isolation capability. In the rest of the post, we can take a look at how to use containerd on FreeBSD!

Today, I open-sourced runj, a new experimental, proof-of-concept OCI-compatible runtime for FreeBSD jails. For the past 6.5 years I’ve been working on Linux containers, but never really had much experience with FreeBSD jails. runj (pronounced “run jay”) is a vehicle for me to learn more about FreeBSD in general and jails in particular. With my position on the Technical Oversight Board of the Open Containers Initiative, I’m also interested in understanding how the OCI runtime specification can be adapted to other operating systems like FreeBSD.

Recently, I’ve been spending a fair amount of time looking at the OCI runtime specification and at the reference implementation, runc. I tend to learn best by doing, and the low-level bits of how containers work have interested me for a long while now, so I’ve started writing a new non-production OCI runtime to learn more about it. In the OCI spec, the docker run interface that folks familiar with Docker containers use has been broken up into multiple steps: create and start. But what’s interesting about runc’s implementation here is that the standard input/output streams (which I tend to refer to as STDIO) for the container’s main process are hooked up to the STDIO streams of runc create rather than that of runc start. This means that if you run runc create in one terminal, and then run runc start in a second terminal, the input and output of your container will be hooked up to the first terminal rather than the second! I’m not entirely clear on the history of why runc behaves this way, but I think the how is interesting on its own. And that how is through multiple processes synchronizing via a FIFO.