finit/TODO.md

Goals of Finit

The intent of finit is to monitor not just processes, but to supervise an entire system and the behavior of certain processes that declare themselves to be finit compliant. Similar to launchd, the goal of finit is to be a replacement for:

• init
• inetd
• crond
• watchdogd

A small and simple replacement, primarily for embedded systems, with a very low dependency on external packages.

General

• Add compile-time supoprt for running /bin/sh instead of getty
• Add support for init <q | reload-configuration> and SIGHUP to have finit reload its finit.conf, but also
• Add support for inotify to automatically reload finit.conf
• Implement initctl stop|start|restart|reload|status <SVC> and service <SVC> stop|start|restart|reload|status on top
• Add PRE and POST hooks for when switching between runlevels
• Add support for init -v,--version | version
• Cleanup move sources from top-level directory to src/ and include/
• Make sure to install queue.h to $(PREFIX)/include/finit/queue.h
• Move finit.conf "check" command to plugin which checks /etc/fstab instead. This is the de-facto practice. But keep the check command for really low-end systems w/o /etc/fstab.

Init

Finit in itself is not SysV Init compatible, it doesn't use /etc/rc.d and /etc/inet.d scripts to boot the system. However, a plausible way to achieve compatibility would be to add a plugin to finit which reads /etc/inittab to be able to start standard Linux systems.

Inetd

Simple way of running Internet services on demand, like the old inetd:

inetd PORT/PROTO <wait|nowait> [@USER[:GROUP]] /path/to/daemon args -- desc

Examples:

# Inetd services, launched on demand
inetd ssh/tcp nowait [2345] @root:root /usr/sbin/sshd -i -- SSH daemon

In keeping with the finit tradition, an optional callback can be setup to each inetd service. When a client connects the finit will call the callback with a struct in_pktinfo argument which the callback then can allowed or deny.

Alternatively, borrowing slightly from TCP wrappers by Wietse Venema, a set of allow and deny directives can be added to finit.conf to control inetd services. It could be what IP addreses, networks or interfaces are allowed to connect to the given inetd services. To start with the easiest (and what we need for WeOS at Westermo) is allow/deny per iface.

# Disable access to SSH
deny  ssh all
# Allow SSH connections originating from eth0
allow ssh eth0

Another option, to further extend the Inetd capabilities, is a separate finetd.conf to list services. This would be beneficial when running an SSH service on different ports on different interfaces:

# Run SSH on port 22 on all interfaces except upstream
tcp ssh {
   listen = *, !wan0
   exec   = dropbear -i
}

# Run SSH on non-standard port 222 for upstream interface
tcp ssh {
   port   = 222
   listen = wan0
   exec   = dropbear -i
   wait   = yes
}

# Run TFTP service on all interfaces except upstream interface
udp tftp {
   listen = *, !wan0
   exec   = uftpd -i -t
   user   = root
}

or, more concise

# System has four interfaces: wan0, eth0, eth1, eth2
# This runs SSH on only three, with various ports
tcp ssh {
   listen = wan0:222 eth0:22 eth1:22
   exec   = /usr/sbin/dropbear -i
   wait   = yes
}

this would be equivalent

tcp ssh {
   listen = wan0:222 !eth2
   exec   = /usr/sbin/dropbear -i
   wait   = yes
}

Syntax similar to xinetd

Custom Inet Port for Common Services

• How to do this, without an obtrusive syntax?

Maybe like this? Then we can still refer to it as the time protocol, and have multiple entries.

inetd time:8037/udp wait [2345] @root:root internal
inetd time:3737/udp wait [2345] @root:root internal

But what if we want to have different allow/deny rules per service? Here's a proposal, similar to user/pass encoding in URI's. This variant could potentially be a lot easier to implement and also to maintain for users, compared to the above allow/deny/xinetd variants.

inetd ssh@eth0:222/tcp nowait [2345] /usr/sbin/sshd -i
inetd ssh@eth1:22/tcp  nowait [2345] /usr/sbin/sshd -i

or

inetd ssh@eth0:222/tcp nowait [2345] /usr/sbin/sshd -i
inetd ssh@*:22/tcp     nowait [2345] /usr/sbin/sshd -i

or

inetd ssh@eth0:222/tcp nowait [2345] /usr/sbin/sshd -i
inetd ssh/tcp          nowait [2345] /usr/sbin/sshd -i

All three versions would be equivalent, except for the first one when there are more interfaces than just eth0 and eth1. The latter two would however be equivalent.

Multiple inetd ssh entries with the same port will not be unique, they will reuse all settings of the first listed entry. However, an entry with a unique port and interface will be registered as a unique service.

Note: Also, there is currently no verification that the same port is used more than once. So a standard http service will clash with an ssh entry ssh@*:80/tcp ... why you would ever want that?

Inetd callbacks are no different from other callbacks. They match to an inetd line using the command. So multiple /usr/bin/sshd inetd lines in finit.conf will all use the same callback. However, the callback will be called with the Internet port and interface as arguments. Like this:

svc_cmd_t sshd_callback(svc_t *svc, int port, char *ifname)
{
    if (!cfg_is_ssh_enabled(ifname, port))
        return SVC_STOP;

    return SVC_START;
}

As usual, the callback is run in a separate process context and it is up to the implementer to return SVC_STOP or SVC_START (SVC_RELOAD is not applicable for inetd services).

Internal inetd services

There are a few standard services that inetd usually had built-in:

• time
• echo
• chargen
• discard

At least 'time' will be implemented in finit, as a plugin, to serve as a simple means of testing inetd functionality stand-alone, but also as a very rudimentary time server for rdate clients.

Example:

# Built-in inetd time service, launched on demand
inetd time/udp wait [2345] @root:root internal

Crond

Requirements are quite rudimentary, basic cron functionality, mainly system-level timed periodic tasks:

• Periodically run a task, in a matching runlevel
• Optionally run as non-root (crontab -e as operator)

Proposed syntax:

# Crontab
cron @YY-mm-ddTHH:MM [LVLS] /path/to/cmd [ARGS] -- Optional descr

# One-shot 'at' command
at @YY-mm-ddTHH:MM [LVLS] /path/to/cmd [ARGS] -- Optional descr

Notice the ISO date in notation. The runlevels is extra filtering sugar. E.g., if cron service is only allowed in runlevels two or three it will not start if system currently is in runlevel four.

To run a command as another user the .conf file must have a that owner. E.g., /etc/finit.d/extra.conf may be owned by operator and only hold cron ... lines.

What would be extremely useful is if initctl, or a homegrown at, could support the basic functionality of the at command directly from the shell -- that way setting up one-time jobs would not entail re-reading /etc/finit.conf

Watchdog

Support for monitoring the health of the system and its processes.

• Add support for a /dev/watchdog built-in to replace both the BusyBox watchdogd and my own refreshed uClinux-dist adaptation, watchdogd
• Add a system supervisor (pmon) to optionally reboot the system when a process stops responding, or when a respawn limit has been reached, as well as when system load gets too high. For details on UNIX loadavg, see http://stackoverflow.com/questions/11987495/linux-proc-loadavg
  • Default to 0.7 as MAX recommended load before warning and 0.9 reboot
• API for processes to register with the watchdog, libwdt
  • Deeper monitoring of a process' main loop by instrumenting
  • If a process is not heard from within its subscribed period time reboot system or restart process.
  • Add client libwdt library, inspired by old Westermo API and the libwdt API published in the Fritz!Box source dump, take for instance the fritzbox7170-source-files-04.87-tar.gz drop and see the GPL-release_kernel.tgz in drivers/char/avm_new/ for the kernel driver, and the LGPL-GPL-release_target_tools.tgz in wdt/ for the userland API.
  • When a client process (a standard app/daemon instrumented with libwdt calls in its main loop) registers with the watchdog, we raise the RT priority to 98 (just below the kernel watchdog in prio). This to ensure that system monitoring goes before anything else in the system.
• Separate /etc/watchdog.conf configuration file, or a perhaps support for /etc/finit.d/PLUGIN.conf?
• Support enable/disable watchdog features:
  • Supervise processes,
  • CPU loadavg,
  • Watch for file descriptor leaks,
  • etc.

Documentation

• Write man pages for finit and finit.conf, steal from the excellent pimd man pages ...
• Update Debian finit.conf example with runlevels, tty/console and dependency handling for Debian 8 (the dreaded systemd release).

Investigation

• FHS changes affecting runtime status, plugins, etc.