Finit v5.0-rc1

Changes

• keventd is now a full device manager, promoted from the power supply
  monitor it was in v4, and the device management defaults have swapped:
  keventd is enabled by default (--without-keventd opts out, libblkid is
  required to build it), while the hotplug plugin, which runs an external
  udevd, mdevd, or mdev, is now opt-in. With the plugin enabled,
  keventd steps back to passive mode and monitors only power supply
  events, as before. As the system device manager it replaces mdev,
  mdevd, and udevd, and handles:

  • udev rules: an engine covering most of the udev(7) grammar --
    parent-chain matching, PROGRAM/RESULT, IMPORT, RUN,
    GOTO/LABEL, and the standard substitutions -- reading the usual
    udev/rules.d directories under /lib, /run, and /etc with
    udev's precedence and masking. A curated ruleset derived from eudev
    3.2.14 is installed by default, --without-udev-rules skips it.
    See Device Manager for the supported subset and the
    few deliberate divergences

  • builtins used by the ruleset: blkid, hwdb, input_id, kmod,
    net_id, path_id, and usb_id

  • device nodes with per-subsystem default permissions, persistent
    symlinks (/dev/disk/by-uuid, by-label, by-id, by-path, plus
    by-id/by-path for input, serial, sound, v4l, and tape), kernel
    firmware load requests, and module loading from MODALIAS

  • predictable network interface names via the eudev ruleset,
    net.ifnames=0 on the kernel command line keeps classic names.
    Note: switching a system over to keventd can rename its NICs

  • coldplug: as device manager keventd starts with -c, replaying add
    events for hardware present at boot. The pid/keventd condition
    is deferred until the coldplug queue has drained, so services
    gating on it can assume /dev is populated and symlinks are live

  • keventd -S, the udevadm settle equivalent, asks the running
    daemon over D-Bus and falls back to polling the kernel's uevent
    sequence number

  • libudev compatibility: per-device properties in /run/udev/data
    and processed uevents rebroadcast to netlink group 0x4, so
    libudev-zero consumers, e.g. compositors and libinput, see hotplug
    events

  • new conditions for the service engine: dev/<node> for device
    nodes, class/<subsystem>/<name> for sysfs class devices without a
    node (LEDs, IIO, DSA ports), and driver/<name>, asserted while a
    driver is bound to at least one device

• Generated /run/finit/system/*.conf files for bundled services (keventd,
  watchdogd, runparts) are now read before /lib/finit/system/*.conf instead
  of after. Built-ins can now be referenced in if: statements from system
  .conf files and overridden by name from both /lib/finit/system and
  /etc/finit.d, the latter is still always read last

• The dbus.so plugin, which starts an external dbus-daemon, is now enabled
  by default. It does nothing on systems without a dbus-daemon installed,
  and --disable-dbus-plugin opts out. The daemon it used to register from C
  now lives in 20-dbus.conf, and the directories it needs in
  tmpfiles.d/dbus.conf, so both can be overridden from /etc like any other
  system file. Those directories are no longer chowned to messagebus,
  matching how Finit ships directories for other daemons

• The bundled sulogin and watchdogd are now built and installed by
  default, like keventd and the D-Bus support. --without-sulogin and
  --without-watchdog opt out. Note, Finit starts watchdogd by itself
  when the WDT device node exists

• The tty block takes passenv, which the line-based format has had since
  v4.4 (issue #286) and the block format was missing

• New provides setting for run/task/service/sysv blocks, naming conditions
  the service asserts in addition to its own pid/<ident>. Lets variants of
  one service, qualified with if, share the barrier condition downstream
  services wait for, which a block title alone could not express. Any
  namespace is allowed, and a second claim on the same condition is refused
  with a warning naming the owner

• The command setting in a block takes a list of candidates for the same
  service, command = { "/lib/systemd/systemd-udevd", "-udevd" }, and Finit
  starts the first one it finds. The line-based format could only express
  this by repeating the whole stanza per candidate

• Finit now ships with a built-in brokerless D-Bus implementation, libink,
  exposing the running init system as a peer on its own private bus at
  /run/finit/bus, and -- opportunistically -- registering org.finit on the
  standard system bus when a dbus-daemon is reachable. No external
  libdbus/sd-bus/GIO dependency. The bus implements the stock
  org.freedesktop.DBus, Peer, Introspectable, and Properties
  interfaces. See D-Bus Integration for the full surface, build
  flag, and dbus-send/dbus-monitor examples

• New org.finit.Manager1 interface at /org/finit/manager: ListServices,
  GetService, Start/Stop/Restart/Reload, SetRunlevel, SetDebug,
  Signal, Suspend, and the Reboot/Halt/Poweroff triplet, which take
  a timeout argument arming the emergency shutdown bypass like initctl -t.
  Read-only properties Runlevel, PrevRunlevel, Version; signals
  ServiceStateChanged (sss) and RunlevelChanged (ss)

• New org.finit.Service1 interface at /org/finit/service/<encoded>, one
  object per loaded service, with Start/Stop/Restart/Reload for
  working off an object handle rather than passing the identity string around

• New org.finit.Cond1 interface at /org/finit/cond: Get, Set, Clear,
  List, Dump for user-defined conditions, with a
  ConditionChanged (ss) signal

• keventd serves org.finit.Device1 on its own bus at /run/keventd/bus:
  Settle, Trigger, Info, RulesReload, queue-state properties, and a
  DeviceProcessed (ss) signal. The udevadm settle/trigger/info
  equivalents are thereby bus methods and keventd -S asks the running daemon
  first. Finit also gained a Manager1.ConfigReloaded signal, fired when
  initctl reload completes, for external condition providers and monitoring

• Privileged D-Bus methods accept root and members of the --with-group
  group; everyone else is refused. On the local bus the caller's uid and
  groups come straight from the kernel (SO_PEERCRED + SO_PEERGROUPS), so
  the check never blocks PID 1 on an NSS lookup. Over the system bus the uid
  is resolved from the broker (GetConnectionUnixUser) and privileged methods
  are root-only. Read-only methods are open

• initctl now transparently routes through D-Bus when the bus is reachable,
  with the legacy INIT_SOCKET transport as a fallback: start, stop,
  restart, reload, reload <svc>, reboot, halt, poweroff,
  suspend, debug, signal, runlevel, and cond {get,set,clr} all use
  the new path. Two new subcommands show up that have no legacy equivalent:

  • initctl monitor -- streams every signal on the bus to the terminal,
    one line per delivery (HH:MM:SS iface.member(args)), until
    interrupted. Same idea as dbus-monitor, but scoped to Finit and with
    no address plumbing required.

  • Issuing initctl cond set/clr over D-Bus also fires the
    Cond1.ConditionChanged signal, so observers see user-driven state
    changes the same way they see service-driven ones.

• Restart log now spells out the signal name and flags core dumps,
  e.g. killed by SIGKILL or killed by SIGSEGV, core dumped, in place of
  the bare numeric by signal: N. Gives operators a much stronger breadcrumb
  when a daemon dies unexpectedly

Fixes

• Reject a .conf file that declares the same block title twice. The title is
  the service identity and libConfuse merges two sections sharing one, so the
  file loaded as a single service holding a mix of both declarations,
  silently. This hit system/10-hotplug.conf, where the two udevd blocks
  left only the second command, dropping the service on systems that only have
  systemd-udevd. The same title in another file is still how a system .conf
  is overridden
• Remove stale daemon-owned (pid:!) pidfiles after unclean exits. When a
  daemon dies via SIGKILL/OOM/segfault, or exits early during startup, its
  pidfile lingers and can prevent the next instance from starting --
  dbus-daemon for example refuses to start when its pidfile already exists,
  so Finit's restart loop gives up. Finit now drops the file when it still
  names the just-reaped PID and that PID is no longer alive (the liveness
  check guards against PID reuse). This is a controlled exception to the
  long-standing rule that Finit does not touch service-owned pidfiles
• Fix misspelled SIGUNKOWN returned by sig_name() for unknown signal
  numbers, now spelled correctly as SIGUNKNOWN. Surfaced by the new restart
  log above