DINIT-AS-INIT.md 18 KB

Dinit as init: using Dinit as your Linux system's init

You can use Dinit, in conjunction with other software, to boot your system.

This document is intended as a guide for building a system around Dinit. The reader is assumed to be knowledgeable about how a Linux system works and the various components that it may be comprised of. If making changes to an existing system, please make backups and/or be prepared to recover a system that fails to boot. Support cannot be provided.

Converting an existing system

If running a Linux distribution, it is theoretically possible to replace your current init system (which on most main distributions is now Systemd, Sys V init, or OpenRC) with Dinit. However: Be warned that a modern Linux system is complex and changing your init system will require some effort and preparation. It is not a trivial task to take a system based on a typical distribution that uses some particular init system and make it instead boot with Dinit. You need to set up suitable service description files for your system; at present there are no automated conversion tools for converting service descriptions or startup scripts from other systems.

This guide is focused mainly on building a new system "from scratch" rather than for converting an existing system to use Dinit. With care and preparation it may, however, still be possible to convert a system based on a distribution to use Dinit. The steps required to accomplish this will depend on the specific details of how the system boots and what services it runs; this is largely beyond the scope of this guide. Note that altering a system in this way is likely to cause changes in system behaviour, potentially causing breakages, and should not be attempted for any critical systems.

The example services (see link below) are mostly designed for a system built "from scratch" rather than based on an existing distribution, and should be modified accordingly to keep existing functionality provided by your distribution and to work with its boot mechanism.

Once you have service descriptions ready, you can test Dinit by adding "init=/sbin/dinit" (for example) to the kernel command line when booting. To have Dinit run as your system init (once you are satisfied that the service descriptions are correct and that the system is bootable via Dinit), replace your system's /sbin/init with a link to the dinit executable.

Note: if your system is based on a distribution and boots via an "initrd"/"initramfs" (initial RAM-disk image or RAM-based filesystem), it may or may not honour kernel options such as "init=...", and it may or may not pass options such as "single" on to Dinit (which enables single-user mode). In order to be able to follow the advice/instructions in this guide, you might need to either adjust the ramdisk image or switch to mounting the root filesystem directly; consult kernel, bootloader and distribution documentation for details (which are beyond the scope of this guide).

General notes

For example service description files, please check the services subdirectory (and see descriptions of all of them below).

It is common to use "devtmpfs" on /dev, and the kernel can actually mount it there before it even starts the init process, which can be quite handy; for one thing it means that a range of device nodes will be available by default (including /dev/console, which dinit may need to display any output, as well as the block device used for the root mount, which must be available in order to perform the initial fsck). You must configure your kernel appropriately for this to happen.

(actually, it seems that Dinit manages output without /dev/console; probably the kernel is giving it appropriate stdin/out/err file descriptors. I'm not sure if this was the case for older kernels).

The /dev filesystem on linux after boot is usually managed by a "device node manager", such as Udev (which is now distributed only with Systemd) or Eudev. Even this is technically optional - you can still populate your root filesystem with device nodes directly - but I highly recommend using an automated system.

Various other virtual filesystems are mounted as standard on Linux these days. They include:

  • /sys - sysfs - representation of devices, buses, drivers etc; used by udev etc.
  • /sys/fs/cgroup - cgroupfs - control groups
  • /proc - procfs - information about running processes, and various kernel interfaces
  • /dev/shm - tmpfs - used for shared memory
  • /dev/pts - devpts - pseudoterminal devices
  • /run - tmpfs - storage for program state (replacement for /var/run); used by udev and some other programs

These filesystems (particularly /sys, /proc and /run) need to be mounted quite early as they will be used by early-boot processes. It is typical for some or all of them to be mounted by an initramfs/initrd-based initial boot, in which case it may not be necessary to mount them via Dinit services.

Many Linux distributions are now built around Systemd. Much of what Systemd manages was previously managed by other utilities/daemons (syslogd, inetd, cron, cgmanager, etc) and these can still be used to provide their original functionality, although at the cost of the losing automated integration.

Some packages may rely on the "logind" functionality of Systemd for session/seat management. This same functionality is also provided by Elogind and ConsoleKit2, though I'm not sure to what degree nor level of compatibility.

In general I've found it quite possible to run a desktop system with Dinit in place of SystemD, but my needs are minimal. If you're running a full-fledged desktop environment like Gnome or KDE you may experience problems (which, I believe, should not be intractable, but which may require implementation/shims of Systemd APIs in some cases).

The basic procedure for boot (to be implemented by services) is as follows:

  • mount early virtual filesystems
  • start device node manager
  • trigger device node manager (udevadm trigger --action=add) to add boot-time device nodes (or run additional actions for nodes already created if using kernel-mounted devtmpfs)
  • set the system time from the hardware realtime clock
  • run root filesystem check
  • remount root filesystem read-write
  • start syslog deamon
  • various miscellaneous tasks: seed the random number generator, configure the loopback interface, cleanup files in /tmp, /var/run and /var/lock
  • start other daemons as appropriate (dhcpcd, any networking daemons)
  • start getty instances on virtual terminals

The service description files and scripts in the services subdirectory provide a template for accomplishing the above, but may need some adjustment for your particular configuration.

The additional software required can be broken into essential and optional packages, which are detailed in following sections.

Essential packages for building a Dinit-based system

Other than the obvious system C library and C++ library, you'll need a range of packages to create a functional Dinit-based system.

First, a device node manager. I recommend "Eudev".

  • Eudev - the Gentoo fork of Udev; https://github.com/gentoo/eudev
  • Mdev may also be an option; it is part of the "busybox" utility suite. I don't personally have any experience with it.

Then, a "getty" and "login" program. Both can be found in the util-linux package, at: https://www.kernel.org/pub/linux/utils/util-linux

Also provided in the util-linux package are standard utilities such as fsck and mount. You'll probably want e2fsprogs (or the equivalent for your chosen filesystem): http://e2fsprogs.sourceforge.net/

There are plenty of syslog daemons; the one I recommend is troglobit's enhanced version of sysklogd. The syslog daemon from GNU Inetutils is another option.

You will need a shell script interpreter / command line, for which you have a range of options. A common choice is GNU Bash, but many distributions are using Dash as the /bin/sh shell because it is significantly faster (affecting boot time) although it is basically unusable as an interactive shell.

Optional packages for building a Dinit-based system

elogind, to act as seat/session manager (extracted from Systemd's logind): https://github.com/elogind/elogind

Alternatively, ConsoleKit2: https://github.com/ConsoleKit2/ConsoleKit2

cgmanager, the control group manager; you probably want this if you use ConsoleKit2, and maybe if you want to use containers: https://github.com/lxc/cgmanager

(However, I believe that cgmanager works with the old v1 cgroups interface. I expect that v2 cgroups together with cgroup namespaces as found in newer kernels will render it obsolete).

The above use D-Bus: https://dbus.freedesktop.org/

Another implementation of D-Bus is dbus-broker: https://github.com/bus1/dbus-broker

Explanation of example services

A set of example service description files can be found in the services subdirectory; these can be used to boot a real system, assuming the appropriate package dependencies are in place. Here are explanations for each package:

  • boot - the internal service which is started on boot. Its dependencies are mostly listed in the boot.d directory. However, the ttyX services are directly listed as depends-ms type dependencies in the service description file. The boot.d directory simplifies enabling services for the package manager, and enables the use of dinitctl enable and dinitclt disable commands to enable or disable particular services.

Having covered boot, we'll go through the other services in roughly the order in which they are expected to start:

  • early-filesystems - this service has no dependencies and so is one of the earliest to start. It mounts virtual filesystems including sysfs, devtmpfs (on /dev) and proc, via the early-filesystems.sh shell script. Note that if startup is via an initial ram disk (initrd, initramfs) as is now common, these early filesystems are most likely already mounted by that, so this service may not be needed or could be edited to remove initrd-mounted filesystems.
  • udevd - this services starts the device node manager, udevd (from the eudev package). This daemon receives notification of hotplug events from the kernel, and creates device nodes (in /dev) according to its configuration. Note that "hotplug" events includes initialisation of devices even when they are not "hot-pluggable".
  • udev-trigger - this is a scripted service which triggers device add actions for all currently present devices. This is required for udevd to process devices which already existed when it started.
  • udev-settle - this is a scripted service which waits until udevd judges that the device list has "settled", that is, there are no more attached devices which are still potentially to be found and reported by the kernel. This is something of a hack, and should not really be relied on, but is convenient to keep our example scripts reasonably simple. A less hacky alternative would be to have triggered services representing particular devices (disks, network interfaces, etc) that other services require; see the 'netdev-enp3s0' service for example.
  • hwclock - this sets the current time (according to the kernel) from the hardware clock. It depends on udevd as it needs the hardware clock device node to be present.
  • modules - this service runs the modules.sh script, which checks whether the kernel supports modules via the proc filesystem, so it depends-on the early-filesystems service. If modules are supported by the kernel, the /etc/modules file is read; each line can contain the name of a kernel module (which will then be loaded) and arguments.
  • rootfscheck - via the rootfscheck.sh script, this service runs a filesystem check on the root filesystem (if it is marked dirty). The script runs "on the console" so that output is visible during boot, and is marked start-interruptible and skippable so that pressing Ctrl+C can skip the check. The default start-timeout of 60 seconds is overridden to 0 (no start timeout), since a filesystem check may take some time. If the filesystem check requires manual intervention, the user is prompted to enter the root password and a maintenance shell is spawned (once it is exited, the system is rebooted). The system is also rebooted if the filesystem check makes automatic changes that require it.
  • rootrw - once the root filesystem has been checked, it can be mounted read-write (the kernel normally mounts root as read-only).
  • auxfscheck - runs fsck for the auxillary filesystems (apart from the root filesystem) which are needed for general system operation. Any filesystems listed in /etc/fstab will be checked, depending on how they are configured.
  • filesystems - this service mounts any auxillary filesystems. It also enables the swap (the example script expects a swapfile at /swapfile). It depends on auxfscheck, i.e. it does not mount filesystems before they have been checked.
  • rcboot - this service runs the rcboot.sh script, which performs a number of basic functions:
    • cleans up the /tmp directory, the /var/lock directory and the /var/run directory
    • creates directories that may be needed under /var/run
    • copies saved entropy to the /dev/urandom device
    • configures the "lo" (loopback) network device (relies on ifconfig from the GNU inetutils package)
    • sets the hostname On shutdown, it saves entropy from /dev/urandom so that it can be restored next boot.

By the time rcboot has started, the system is quite functional. The following additional services can then start:

  • syslogd - the logging daemon. This service has the starts-log option set, so that Dinit will commence logging (from its buffer) once the service starts. The example service relies on troglobit's sysklogd. Unfortunately it must be a bgprocess as it does not support signalling readiness via a file descriptor.
  • late-filesystems - check and mount any filesystems which are not needed for general system operation (i.e. "user" filesystems). It's not expected that other services will depend on this service. This service uses the late-filesystems.sh script; configure late filesystems via that script.
  • dbusd - starts the DBus daemon (system instance), which is used by other services to provide an interface to user processes
  • dhcpcd - starts a DHCP client daemon on a network interface (the example uses enp3s0).
  • 'netdev-enp3s0' - a triggered service representing the availablility of the enp3s0 network interface. See the service description file for details. Note that the 'udev-settle` service somewhat makes this redundant, as would use of a suitable network manager; it is provided for example purpsoses.
  • sshd - starts the SSH daemon.

We want most of the preceding services to be started before we allow a user to login. To that end, we have:

  • loginready - an internal service, which depends on rcboot, dbusd, udevd and syslogd. This service has option runs-on-console set, to prevent Dinit from outputting service status messages to the console once login is possible.
  • ttyX where X is 1-6 - a service which starts a login prompt on the corresponding virtual terminal (and which depends on loginready). Note that the tty services do not have the runs-on-console option, since that would conflict with loginready (and with each other) and ultimately prevent the tty services running, as only one service can run on the console at a time.

There are two additional services, which are not depended on by any other service, and so do not normally start at all:

  • recovery - this service is started by Dinit if boot fails (and if the user when prompted then chooses the recovery option). It prompts for the root password and then provides a shell.
  • single - this is a "single user mode" startup service which simply runs a shell. An unprivileged user cannot normally start this; doing so requires putting "single" on the kernel command line. When the shell exits, the chain-to setting will cause normal startup to resume (i.e. via the boot service).

While they are a little rough around the edges, these service definitions demonstrate the essentials of getting a system up and running.

Testing and debugging tips

You can pass arbitrary arguments to dinit by using a shell script in the place of /sbin/init, which should exec dinit (so as to give it the same PID). Don't forget to make the script executable and to include the shebang line (#!/bin/sh or similar).

You can run a shell directly on a virtual terminal by adding a ttyN service or modifying one of the existing ones (see the example services). You'll still need getty to setup the terminal for the shell; an example setting:

command = /sbin/agetty tty6 linux-c -n -l /bin/bash

You can remove most or all dependencies from this service so that it starts early, and set the term-signal = none, as well as setting stop-timeout = 0 (i.e. disabling stop timeout), so that it will not be killed at shutdown (you will need to manually exit the shell to complete shutdown). This means you always have a shell available to check system state when something is going wrong. While this is not something you want to enable permanently, it can be a good tool to debug a reproducible boot issue or shutdown issue.

Caveats

For services which specify a logfile, the location must be writable when the service starts (otherwise the service will fail to start). Typically this means that nearly all services must depend on the service that makes the root filesystem writable. For services that must start before the root filesystem becomes writable, it may be possible to log in /run or another directory that is mounted with a RAM-based filesystem; alternatively, the shares-console option can be used for these services so that their output is visible at startup. There is also the possibility of using the "log-type = buffer" setting to keep output buffered in memory instead.