Systemd gets a lot of hate. There's a lot of heat and very little light in those discussions, in my opinion, and I don't expect that this post will change the mind of anyone who has already decided to hate Systemd. My goal here is far more modest. I want to share a feature of the new init system that I find really compelling, and that I hadn't seen discussed pretty much anywhere: Systemd's native ability to leverage the Linux kernel's namespacing features to run services in a sandboxed environment.

Linux namespaces are the kernel feature that enables partitioning everything from processes, to the filesystem, to the network stack. A process operating in one namespace will have a different view of the system's resources than a process operating in another namespace. Probably the best known application of Linux namespaces are container platforms such as Docker. While Docker is ostensibly a devops platform enabling rapid deployment of applications, the underlying kernel namespace features can be applied to security as well, allowing processes to be effectively partitioned off from one another and, to varying degrees, the underlying system. That's what folks mean when they talk about running a process or service in a sandboxed environment.

So what if I'm a sysadmin who wants to run a service in a sandbox? This would traditionally be done by setting up a chroot environment. But another option, one that offers a bit more flexibility, would be to run the service in a mount namespace, and then reconfigure the existing filesystem within the namespace to apply least-privilege to data the services needs and hide data the service doesn't need access to.

With Systemd, you can configure your service according to either of the above scenarios by simply adding a couple of lines to the service file. Say I want my service to run within a chroot located at /srv/http. Assuming the chroot is set up appropriately so that the service has access to all of the data it needs within its folder hierarchy, then all I need to do is add the line RootDirectory=/srv/http to the [Service] section of the Systemd service file.

The second scenario is a bit more interesting. Say I'm running a web front-end for my Git service, and that my service needs access to /dev/urandom /tmp and read-only access to /home/git. Systemd offers several options that allow you to do this in a way that exposes little else to the service. Take the below service file:

...
[Service]
PrivateDevices=yes
PrivateTmp=yes
ProtectHome=tmpfs
BindReadOnlyPaths=/home/git
...

These options implicitly place the service within a mount namespace, meaning we set up our file hierarchy however we like. In the above example, PrivateDevices creates a private /dev structure that only provides access to pseudo-devices like random null, etc. Critically, disk devices are not visible to the service when the PrivateDevices option is used. Likewise, PrivateTmp creates a /tmp folder that is empty except for what the service itself writes. ProtectHome has a few options, but the tmpfs option, according to the documentation, is designed for pairing with the BindPaths/BindReadOnlyPaths options in order to selectively provide access to folders within /home. Since all we need there is read-only access to the git user, that's exactly what we provide, nothing more and nothing less.

This is all great, but it admittedly only provides mount namespacing for the service. This is probably sufficient in most cases, but Systemd does offer options for network and user namespacing. Readers looking to utilize those, or looking for a comprehensive description of the mount namespacing options, are encouraged to read the systemd.exec man page.