s6-linux-init
Software
skarnet.org

Quickstart and FAQ for s6-linux-init

Quickstart

1. Install all the s6-linux-init dependencies:
   • skalibs
   • execline
   • s6-portable-utils
   • s6-linux-utils
   • s6
2. Install s6-linux-init itself.
3. Save your old /sbin/init binary.
4. Save and remove your old /etc/s6-linux-init directory, if you have one.
5. Make sure you have a /run directory.
6. Write a machine initialization script in /etc/rc.init and a machine shutdown script in /etc/rc.shutdown. Make sure they are executable. See below for more information on how to write these scripts.
7. Check that your devtmpfs is automounted by your kernel at boot time. If it is not, add the -d 1 option to the s6-linux-init-maker command line below.
8. As root, run:

        rm -rf /tmp/s6-linux-init /tmp/init
        s6-linux-init-maker /tmp/s6-linux-init
        mv /tmp/s6-linux-init /etc/
        ln -sf /etc/s6-linux-init/init /sbin/init 

9. Reboot.
10. Congratulations! your machine is now running a s6-based init system.
11. To shut the machine down, use the s6-halt, s6-poweroff or s6-reboot command as appropriate.

What should go into /etc/rc.init and /etc/rc.shutdown ?

/etc/rc.init

This script will be run after s6-linux-init has done is job, i.e. s6-svscan is running as process 1, and it is now up to /etc/rc.init to get the machine to its usable state. It normally contains a call to the service manager to bring up all the services; for instance, if you're using s6-rc as your service manager, and your top bundle (containing all the services you want to bring up) is named ok-all, a proper /etc/rc.init could look like this:

#!/bin/sh
s6-rc-init /run/service && exec s6-rc -u change ok-all

The script can assume that:

• There is a tmpfs partition, only writable by root, mounted on /run
• There is a s6 supervision tree running on /run/service
• /dev is mounted, but /proc and /sys are not

/etc/rc.shutdown

This script is spawned by s6-svscan when the administrator calls s6-halt, s6-poweroff or s6-reboot. When this script exits, the final shutdown sequence is run, which means that the supervision tree is dismantled, all processes are killed, the file systems are umounted and the system undergoes a hardware shutdown or reboot. So the goal of this script is to bring services down in an orderly fashion and perform all the necessary cleanups before all remaining processes are summarily killed.

If you're using s6-rc as your service manager, a proper /etc/rc.shutdown could look like this:

#!/bin/sh
exec s6-rc -da change

FAQ

Why is it so complicated to use s6 as an init process? It's much simpler with runit.

Yes, runit is simpler, because it provides a simple runit binary suitable as a /sbin/init program and calls scripts to handle the three stages of init. However, the runit design has a few perfectible points:

• The one-time initialization is performed in /etc/runit/1, but the supervision tree is not run until /etc/runit/2, which means means that it is impossible to start supervised services during the one-time initialization. Early daemons such as udevd, for instance, have to remain unsupervised.
• runit runs with its descriptors pointing to /dev/console, which means that error messages from the supervision tree, and uncaught logs, will be displayed on the system console; they are not saved beyond the console buffer capabilities.
• The runit supervision tree is of height 3 (runit, runsvdir, runsv), when height 2 is enough - some init systems, like sysvinit, systemd or launchd, even provide a supervision tree of height 1! (At the expense of complexity in the init process, of course.) Height 3 is a bit redundant, because the supervision capabilities of the root will be redundant with either those of the trunk or those of the branches. Its display is also aesthetically less pleasing than height 2: try out ps afuxww on a runit-based system. Yes, this point is extremely minor, but still deserves a mention. :-)

Running a s6-based init addresses those issues:

• Save for the initial tmpfs mount, all of the machine initialization runs in the stage 2 script, i.e. /etc/rc.init, and the supervision tree is already available at that point. This makes it possible to start one-shot services as well as long-run services in the desired order while ensuring that every long-run service is properly supervised, i.e. it lays the ground for a proper dependency management system.
• s6-linux-init solves the problem of uncaught logs in a clean way, and any error message from any process in the system is guaranteed to end up in a logging directory. The only exception is error messages from the catch-all logger process itself: those naturally go to /dev/console.
• When s6-svscan runs as process 1, the supervision tree is of height 2, and ps afuxww looks clean.

To sum up, a s6-based init is cleaner than a runit-based init; it's a bit more complex to set up, but it organizes the system in a better way, without using more resources. And the goal of s6-linux-init is to make the setup more accessible.

My /etc/rc.init script is not printing anything!

You probably gave the -r option to s6-linux-init-maker, and your /etc/rc.init's output is being logged into the /run/uncaught-logs directory instead of printed to /dev/console.

I want to run s6 in a container, and I just want to log to stdout/stderr, without this tmpfs and /dev/console stuff and without having a catch-all logger inside the container. Is it possible ?

Yes, it is possible, but then s6-linux-init may not be what you are looking for. For your case, it will be simpler to run s6-svscan directly!

If you are using Docker, there is a s6-overlay project specifically made for integrating s6 into Docker images.