Awesome

Introduction

Fast init for Linux systems. Reverse engineered from the EeePC fastinit by Claudio Matsuoka — "gaps filled with frog DNA …"

Alpine screenshot
Figure 1: Screenshot of Finit booting Alpine Linux (HowTo).

Features include:

Runlevels, defined per service
One-shot tasks, services (daemons), or SysV init start/stop scripts
Runparts and /etc/rc.local support
Process supervision similar to systemd
Sourcing environment files
Conditions for network/process/custom dependencies
Readiness notification; PID files (native) for synchronizing system startup, support for systemd sd_notify(), or s6 style too
Limited support for tmpfiles.d(5) (no aging, attributes, or subvolumes)
Pre/Post script actions
Rudimentary templating support
Tooling to enable/disable services
Built-in getty
Built-in watchdog, with support for hand-over to watchdogd
Built-in support for Debian/BusyBox /etc/network/interfaces
Cgroups v2, both configuration and monitoring in initctl top
Plugin support for customization
Proper rescue mode with bundled sulogin for protected maintenance shell
Integration with watchdogd for full system supervision
Logging to kernel ring buffer before syslogd has started, see the recommended sysklogd project for complete logging integration and how to log to the kernel ring buffer from scripts using logger

Focus is on small and embedded systems, although Finit is fully usable on server and desktop systems as well. For working examples, see the contrib/ section with tutorials for the following Linux distributions:

Void Linux,
Alpine Linux, and
Debian GNU/Linux, also works on Ubuntu/Linux Mint

Note: support for various Linux distributions does not mean Finit installs easily on all architectures. The bundled install scripts are examples for standard installations, tested on amd64 (x86_64) systems. Custom setups, e.g., for embedded systems, can be found in any of the following Buildroot based examples: myLinux, Infix, or the plain br2-finit-demo.

Example

This example /etc/finit.conf can also be split up in multiple .conf files in /etc/finit.d. Available, but not yet enabled, services can be placed in /etc/finit.d/available and enabled by an operator using the initctl tool. See the above mentioned Linux distributions, or myLinux.

Note: as of Finit v4.4, .conf lines can be broken up using the standard UNIX continuation character (\), also trailing comments are now supported. The latter means you need to escape any hashes used in directives and descriptions (\#). For more on this and examples, see the finit.conf(5) manual or doc/config.md.

# Fallback if /etc/hostname is missing
host default

# Runlevel to start after bootstrap, 'S', default: 2
#runlevel 2

# Support for setting global environment variables, using foo=bar syntax
# be careful though with variables like PATH, SHELL, LOGNAME, etc.
#PATH=/usr/bin:/bin:/usr/sbin:/sbin

# Max file size for each log file: 100 kiB, rotate max 4 copies:
# log => log.1 => log.2.gz => log.3.gz => log.4.gz
log size=100k count=4

# Services to be monitored and respawned as needed
service [S12345] env:-/etc/conf.d/watchdog watchdog $WATCHDOG_OPTS $WATCHDOG_DEV -- System watchdog daemon
service [S12345] env:-/etc/conf.d/syslog syslogd -n $SYSLOGD_OPTS          -- System log daemon
service [S12345] <pid/syslogd> env:-/etc/conf.d/klogd klogd -n $KLOGD_OPTS -- Kernel log daemon
service   [2345] env:-/etc/conf.d/lldpd lldpd -d $LLDPD_OPTS               -- LLDP daemon (IEEE 802.1ab)

# The BusyBox ntpd does not use syslog when running in the foreground
# So we use this trick to redirect stdout/stderr to a log file.  The
# log file is rotated with the above settings.  The condition declares
# a dependency on a system default route (gateway) to be set.  A single
# <!> at the beginning means ntpd does not respect SIGHUP for restart.
service [2345] log:/var/log/ntpd.log <!net/route/default> ntpd -n -l -I eth0 -- NTP daemon

# For multiple instances of the same service, add :ID somewhere between
# the service/run/task keyword and the command.
service :80   [2345] merecat -n -p 80   /var/www -- Web server
service :8080 [2345] merecat -n -p 8080 /var/www -- Old web server

# Alternative method instead of below runparts, can also use /etc/rc.local
#sysv [S] /etc/init.d/keyboard-setup       -- Setting up preliminary keymap
#sysv [S] /etc/init.d/acpid                -- Starting ACPI Daemon
#task [S] /etc/init.d/kbd                  -- Preparing console

# Hidden from boot progress, using empty `--` description
#sysv [S] /etc/init.d/keyboard-setup       --
#sysv [S] /etc/init.d/acpid                --
#task [S] /etc/init.d/kbd                  --

# Run start scripts from this directory
# runparts /etc/start.d

# Virtual consoles run BusyBox getty, keep kernel default speed
tty [12345] /sbin/getty -L 0 /dev/tty1  linux nowait noclear
tty [2345]  /sbin/getty -L 0 /dev/tty2  linux nowait noclear
tty [2345]  /sbin/getty -L 0 /dev/tty3  linux nowait noclear

# Use built-in getty for serial port and USB serial
#tty [12345] /dev/ttyAMA0 noclear nowait
#tty [12345] /dev/ttyUSB0 noclear

# Just give me a shell, I need to debug this embedded system!
#tty [12345] console noclear nologin

The service stanza, as well as task, run and others are described in full in doc/config.md. Here's a quick overview of some of the most common components needed to start a UNIX daemon:

service [LVLS] <COND> log env:[-]/etc/default/daemon daemon ARGS -- Daemon daemon
^       ^      ^      ^   ^                          ^      ^       ^
|       |      |      |   |                          |      |        `---------- Optional description
|       |      |      |   |                          |       `------------------ Daemon arguments
|       |      |      |   |                           `------------------------- Path to daemon
|       |      |      |    `---------------------------------------------------- Optional env. file
|       |      |       `-------------------------------------------------------- Redirect output to log
|       |       `--------------------------------------------------------------- Optional conditions
|        `---------------------------------------------------------------------- Optional Runlevels
 `------------------------------------------------------------------------------ Monitored application

Some components are optional: runlevel(s), condition(s) and description, making it easy to create simple start scripts and still possible for more advanced uses as well:

service /usr/sbin/sshd -D

Dependencies are handled using conditions. One of the most common conditions is to wait for basic networking to become available:

service <net/route/default> nginx -- High performance HTTP server

Here is another example where we instruct Finit to not start BusyBox ntpd until syslogd has started properly. Finit waits for syslogd to create its PID file, by default /var/run/syslogd.pid.

service [2345] log <!pid/syslogd> ntpd -n -N -p pool.ntp.org
service [S12345] syslogd -n -- Syslog daemon

Notice the log keyword, BusyBox ntpd uses stderr for logging when run in the foreground. With log Finit redirects stdout + stderr to the system log daemon using the command line logger(1) tool.

A service, or task, can have multiple dependencies listed. Here we wait for both syslogd to have started and basic networking to be up:

service [2345] log <pid/syslogd,net/route/default> ntpd -n -N -p pool.ntp.org

If either condition fails, e.g. loss of networking, ntpd is stopped and as soon as it comes back up again ntpd is restarted automatically.

Note: Make sure daemons do not fork and detach themselves from the controlling TTY, usually an -n or -f flag, or -D as in the case of OpenSSH above. If it detaches itself, Finit cannot monitor it and will instead try to restart it.

Features

Process Supervision

Start, monitor and restart services should they fail.

Getty

Finit supports external getty but also comes with a limited built-in Getty, useful for really small systems. A getty sets up the TTY and waits for user input before handing over to /bin/login, which is responsible for handling the actual authentication.

tty [12345] /dev/tty1    nowait  linux
tty [12345] /dev/ttyAMA0 noclear vt100
tty [12345] /sbin/getty  -L /dev/ttyAMA0 vt100

Users of embedded systems may want to enable automatic serial console with the special @console device. This works regardless weather the system uses ttyS0, ttyAMA0, ttyMXC0, or anything else. Finit figures it out by querying sysfs: /sys/class/tty/console/active.

tty [12345] @console linux noclear

Notice the optional noclear, nowait, and nologin flags. The latter is for skipping the login process entirely. For more information, see doc/config.md.

Runlevels

Support for SysV init-style runlevels is available, in the same minimal style as everything else in Finit. The [2345] syntax can be applied to service, task, run, and TTY stanzas.

Reserved runlevels are 0 and 6, halt and reboot, respectively just like SysV init. Runlevel 1 can be configured freely, but is recommended to be kept as the system single-user runlevel since Finit will not start networking here. The configured runlevel NUM from /etc/finit.conf is what Finit changes to after bootstrap, unless 'single' (or 'S') is given on the kernel cmdline, in which case runlevel 1 is started.

All services in runlevel S) are started first, followed by the desired run-time runlevel. Run tasks in runlevel S can be started in sequence by using run [S] cmd. Changing runlevels at runtime is done like any other init, e.g. <kbd>init 4</kbd>, but also using the more advanced intictl tool.

Conditions

As mentioned previously, Finit has an advanced dependency system to handle synchronization, called conditions. It can be used in many ways; depend on another service, network availability, etc.

One really cool example useful for embedded systems is to run certain scripts if a board has a certain feature encoded in its device tree. At bootstrap we run the following ident script:

#!/bin/sh
conddir=/var/run/finit/cond/hw/model
dtmodel=/sys/firmware/devicetree/base/model

if ! test -e $dtmodel; then
    exit 0
fi

model=$(cat $dtmodel | tr "[A-Z] " "[a-z]-")
mkdir -p $conddir && ln -s ../../reconf $conddir/$model

Provided the device tree node exists, and is a string, we can then use the condition <hw/model/foo> when starting other scripts. Here is an example:

run  [S]                /path/to/ident    --
task [2] <hw/model/foo> /path/to/foo-init -- Initializing Foo board

Notice the trick with an empty description to hide the call to ident in the Finit progress output.

Plugins

Plugins can extend the functionality of Finit and hook into the different stages of the boot process and at runtime. Plugins are written in C and compiled into a dynamic library loaded automatically by finit at boot. A basic set of plugins are bundled in the plugins/ directory.

Capabilities:

Hooks
Hook into the boot at predefined points to extend Finit
I/O
Listen to external events and control Finit behavior/services

Extensions and functionality not purely related to what an /sbin/init needs to start a system are available as a set of plugins that either hook into the boot process or respond to various I/O.

For more information, see doc/plugins.md.

Automatic Reload

By default, Finit monitors /etc/finit.d/ and /etc/finit.d/enabled/ registering any changes to .conf files. To activate a change the user must call initctl reload, which reloads all modified files, stops any removed services, starts new ones, and restarts any modified ones. If the command line arguments of a service have changed, the process will be terminated and then started again with the updated arguments. If the arguments have not been modified and the process supports SIGHUP, the process will receive a SIGHUP rather than being terminated and started.

For some use-cases the extra step of calling initctl reload creates an unnecessary overhead, which can be removed at build-time using:

configure --enable-auto-reload

Cgroups

Finit supports cgroups v2 and comes with the following default groups in which services and user sessions are placed in:

 /sys/fs/cgroup
   |-- init/               # cpu.weight:100
   |-- system/             # cpu.weight:9800
   `-- user/               # cpu.weight:100

Finit itself and its helper scripts and services are placed in the top-level leaf-node group init/, which also is reserved.

All run/task/service/sysv processes are placed in their own sub-group in system/. The name of each sub-group is taken from the respective .conf file from /etc/finit.d.

All getty/tty processes are placed in their own sub-group in user/. The name of each sub-group is taken from the username.

A fourth group also exists, the root group. It is also reserved and primarily intended for RT tasks. If you have RT tasks they need to be declared as such in their service stanza like this:

service [...] <...> cgroup.root /path/to/foo args -- description

cgroup.root
service [...] <...> /path/to/foo args -- description
service [...] <...> /path/to/bar args -- description

See doc/config.md for more information, e.g., how to configure per-group limits.

The initctl tool has three commands to help debug and optimize the setup and monitoring of cgroups. See the ps, top, and cgroup commands for details.

Note: systems that do not support cgroups, specifically version 2, are automatically detected. On such systems the above functionality is disabled early at boot.

Runparts & /etc/rc.local

At the end of the boot, when all bootstrap (S) tasks and services have started, but not networking, Finit calls its built-in run-parts(8) command on any configured runparts <DIR> directory. This happens just before changing to the configured runlevel (default 2). (Networking is enabled just prior to changing from single user mode.)

runparts /etc/rc.d/

Right after the runlevel change when all services have started properly, /etc/rc.local is called.

No configuration stanza in /etc/finit.conf is required for rc.local. If it exists and is an executable shell script Finit calls it at the very end of the boot, before calling the HOOK_SYSTEM_UP. See more on hooks in doc/plugins.md.

Limitations

It is not possible to call Finit via signals or use initctl in any runparts or /etc/rc.local script. This because Finit is single threaded and is calling these scripts in a blocking fashion at the end of runlevel S, at which point the event loop has not yet been started.

The event loop is the whole thing which Finit is built around, except for runlevel S, which remains a slow procession through a lot of set up, with a few hooks and blocking call outs to external scripts.

However, not all initctl commands are prohibited. Supported commands:

inictl cond: only operate of files in /run/finit/cond
initctl enable/disable: enabled run/task/service is activated on the runlevel change from S to 2
initctl touch/show/create/delete/list: create, provided the non-interactive mode is used, again changes take effect in the runlevel change directly after bootstrap
initctl -f reboot/poweroff/halt: provided the -f flag is used to force direct kernel commands

Example: you can set a usr/ condition in /etc/rc.local and have a service/task in runlevel 2 depend on it to execute.

Runlevels

Basic support for runlevels is included in Finit from v1.8. By default all services, tasks, run commands and TTYs listed without a set of runlevels get a default set [234] assigned. The default runlevel after boot is 2.

Finit supports runlevels 0-9, and S, with 0 reserved for halt, 6 reboot and S for services to only run at bootstrap. Runlevel 1 is the single user level, where usually no networking is enabled. In Finit this is more of a policy for the user to define. Normally only runlevels 1-6 are used, and even more commonly, only the default runlevel is used.

To specify an allowed set of runlevels for a service, run command, task, or tty, add [NNN] to your /etc/finit.conf, like this:

service [S12345] syslogd -n -x             -- System log daemon
run     [S]      /etc/init.d/acpid start   -- Starting ACPI Daemon
task    [S]      /etc/init.d/kbd start     -- Preparing console
service [S12345] <pid/syslogd> klogd -n -x -- Kernel log daemon

tty     [12345]  /dev/tty1
tty     [2]      /dev/tty2
tty     [2]      /dev/tty3
tty     [2]      /dev/tty4
tty     [2]      /dev/tty5
tty     [2]      /dev/tty6

In this example syslogd is first started, in parallel, and then acpid is called using a conventional SysV init script. It is called with the run command, meaning the following task command to start the kbd script is not called until the acpid init script has fully completed. Then the keyboard setup script is called in parallel with klogd as a monitored service.

Again, tasks and services are started in parallel, while run commands are called in the order listed and subsequent commands are not started until a run command has completed. Also, task and run commands are run in a shell, so pipes and redirects can be used.

The following examples illustrate this. Bootstrap task and run commands are also removed when they have completed, initctl show will not list them.

task [S] echo "foo" | cat >/tmp/bar
run  [S] echo "$HOME" >/tmp/secret

Switching between runlevels can be done by calling init with a single argument, e.g. <kbd>init 5</kbd>, or using initctl runlevel 5, both switch to runlevel 5. When changing runlevels Finit also automatically reloads all .conf files in the /etc/finit.d/ directory. So if you want to set a new system config, switch to runlevel 1, change all config files in the system, and touch all .conf files in /etc/finit.d before switching back to the previous runlevel again — that way Finit can both stop old services and start any new ones for you, without rebooting the system.

Rebooting & Halting

Traditionally, rebooting and halting a UNIX system is done by changing its runlevel. Finit comes with its own tooling providing: shutdown, reboot, poweroff, and suspend, but also the initctl tool, detailed in the next section.

For compatibility reasons Finit listens to the same set of signals as BusyBox init. This is not 100% compatible with SysV init, but clearly the more common combination for Finit. For more details, see doc/signals.md.

Commands & Status

Finit also implements a modern API to query status, and start/stop services, called initctl. Unlike telinit the initctl tool does not return until the given command has fully completed.

Usage: initctl [OPTIONS] [COMMAND]

Options:
  -b, --batch               Batch mode, no screen size probing
  -c, --create              Create missing paths (and files) as needed
  -f, --force               Ignore missing files and arguments, never prompt
  -h, --help                This help text
  -j, --json                JSON output in 'status' and 'cond' commands
  -1, --once                Only one lap in commands like 'top'
  -p, --plain               Use plain table headings, no ctrl chars
  -q, --quiet               Silent, only return status of command
  -t, --no-heading          Skip table headings
  -v, --verbose             Verbose output
  -V, --version             Show program version

Commands:
  debug                     Toggle Finit (daemon) debug
  help                      This help text
  version                   Show program version

  ls | list                 List all .conf in /etc/finit.d
  create   <CONF>           Create   .conf in /etc/finit.d/available
  delete   <CONF>           Delete   .conf in /etc/finit.d/available
  show     <CONF>           Show     .conf in /etc/finit.d/available
  edit     <CONF>           Edit     .conf in /etc/finit.d/available
  touch    <CONF>           Change   .conf in /etc/finit.d/available
  enable   <CONF>           Enable   .conf in /etc/finit.d/available
  disable  <CONF>           Disable  .conf in /etc/finit.d/enabled
  reload                    Reload  *.conf in /etc/finit.d (activate changes)

  cond     set   <COND>     Set (assert) user-defined conditions     +usr/COND
  cond     get   <COND>     Get status of user-defined condition, see $? and -v
  cond     clear <COND>     Clear (deassert) user-defined conditions -usr/COND
  cond     status           Show condition status, default cond command
  cond     dump  [TYPE]     Dump all, or a type of, conditions and their status

  log      [NAME]           Show ten last Finit, or NAME, messages from syslog
  start    <NAME>[:ID]      Start service by name, with optional ID
  stop     <NAME>[:ID]      Stop/Pause a running service by name
  reload   <NAME>[:ID]      Reload service as if .conf changed (SIGHUP or restart)
                            This allows restart of run/tasks that have already run
                            Note: Finit .conf file(s) are *not* reloaded!
  restart  <NAME>[:ID]      Restart (stop/start) service by name
  signal   <NAME>[:ID] <S>  Send signal S to service by name, with optional ID
  ident    [NAME]           Show matching identities for NAME, or all
  status   <NAME>[:ID]      Show service status, by name
  status                    Show status of services, default command

  cgroup                    List cgroup config overview
  ps                        List processes based on cgroups
  top                       Show top-like listing based on cgroups

  plugins                   List installed plugins

  runlevel [0-9]            Show or set runlevel: 0 halt, 6 reboot
  reboot                    Reboot system
  halt                      Halt system
  poweroff                  Halt and power off system
  suspend                   Suspend system

  utmp     show             Raw dump of UTMP/WTMP db

For services not supporting SIGHUP the <!> notation in the .conf file must be used to tell Finit to stop and start it on reload and runlevel changes. If <> holds more conditions, these will also affect how a service is maintained.

Note: even though it is possible to start services not belonging in the current runlevel these services will not be respawned automatically by Finit if they exit (crash). Hence, if the runlevel is 2, the below Dropbear SSH service will not be restarted if it is killed or exits.

The status command is the default, it displays a quick overview of all monitored run/task/services. Here we call initctl -p, suitable for scripting and documentation:

alpine:~# initctl -p
PID   IDENT     STATUS   RUNLEVELS     DESCRIPTION
======================================================================
1506  acpid     running  [---2345----] ACPI daemon
1509  crond     running  [---2345----] Cron daemon
1489  dropbear  running  [---2345----] Dropbear SSH daemon
1511  klogd     running  [S-12345----] Kernel log daemon
1512  ntpd      running  [---2345----] NTP daemon
1473  syslogd   running  [S-12345----] Syslog daemon

alpine:~# initctl -pv
PID   IDENT     STATUS   RUNLEVELS     COMMAND
======================================================================
1506  acpid     running  [---2345----] acpid -f
1509  crond     running  [---2345----] crond -f -S $CRON_OPTS
1489  dropbear  running  [---2345----] dropbear -R -F $DROPBEAR_OPTS
1511  klogd     running  [S-12345----] klogd -n $KLOGD_OPTS
1512  ntpd      running  [---2345----] ntpd -n $NTPD_OPTS
1473  syslogd   running  [S-12345----] syslogd -n

The environment variables to each of the services above are read from, in the case of Alpine Linux, /etc/conf.d/. Other distributions may have other directories, e.g., Debian use /etc/default/.

The status command takes an optional NAME:ID argument. Here we check the status of dropbear, which only has one instance in this system:

alpine:~# initctl -p status dropbear
     Status : running
   Identity : dropbear
Description : Dropbear SSH daemon
     Origin : /etc/finit.d/enabled/dropbear.conf
Environment : -/etc/conf.d/dropbear
Condition(s):
    Command : dropbear -R -F $DROPBEAR_OPTS
   PID file : !/run/dropbear.pid
        PID : 1485
       User : root
      Group : root
     Uptime : 2 hour 46 min 56 sec
  Runlevels : [---2345----]
     Memory : 1.2M
     CGroup : /system/dropbear cpu 0 [100, max] mem [--.--, max]
              |- 1485 dropbear -R -F
              |- 2634 dropbear -R -F
              |- 2635 ash
              `- 2652 initctl -p status dropbear

Apr  8 12:19:49 alpine authpriv.info dropbear[1485]: Not backgrounding
Apr  8 12:37:45 alpine authpriv.info dropbear[2300]: Child connection from 192.168.121.1:47834
Apr  8 12:37:46 alpine authpriv.notice dropbear[2300]: Password auth succeeded for 'root' from 192.168.121.1:47834
Apr  8 12:37:46 alpine authpriv.info dropbear[2300]: Exit (root) from <192.168.121.1:47834>: Disconnect received
Apr  8 15:02:11 alpine authpriv.info dropbear[2634]: Child connection from 192.168.121.1:48576
Apr  8 15:02:12 alpine authpriv.notice dropbear[2634]: Password auth succeeded for 'root' from 192.168.121.1:48576

Requirements

Finit is capable of running on both desktop/server systems with udev and embedded systems that usually come with BusyBox mdev. Some systems have systemd-udev or eudev today instead of the original udev, Finit probes for all of them at runtime and expects /dev/ to be a writable file system using devtmpfs. It is also possible to run on a statically set up /dev if needed. It is however not a good idea to have both udev and mdev installed at the same time, this will lead to unpredictable results.

At boot Finit calls either mdev or udevd to populate /dev, this is done slightly differently and on systems with udev you might want to add the following one-shot task early in your /etc/finit.conf:

run [S] udevadm settle --timeout=120 -- Waiting for udev

Finit has a built-in Getty for TTYs, but requires a working /bin/login or /bin/sh, if no TTYs are configured in /etc/finit.conf.

For a fully operational system /var, /run and /tmp must be set up properly in /etc/fstab -- which is iterated over at boot.

Origin & References

This project is based on the original finit by Claudio Matsuoka which was reverse engineered from syscalls of the EeePC fastinit — "gaps filled with frog DNA …"

Finit is developed and maintained by Joachim Wiberg at GitHub. Please file bug reports, clone it, or send pull requests for bug fixes and proposed extensions.