notsystemd

A pod-native stand-in for systemd, for software that expects to run under it.

Some vendor software ships as a systemd-booted image: a Debian userland with /sbin/init as PID 1 and a dozen units that notify readiness, ping a watchdog, and expect pre-start hooks to have run. Splitting that into one container per service means each container needs whatever systemd did for its unit. notsystemd is that: run as a container's entrypoint, it reads the vendor's own .service file and runs it the way systemd would, so the image definition is the unit file and nothing more.

$ notsystemd run /lib/systemd/system/unifi-core.service

Commands

• notsystemd run [--drop-in-dir DIR]... UNIT_FILE runs a service unit until it exits or notsystemd gets SIGTERM or SIGINT. SIGHUP runs ExecReload=. It exits with the service's status.
• notsystemd is-active UNIT exits 0 while the unit is active, for a readiness probe.
• notsystemd show UNIT prints the unit's recorded state as JSON.
• notsystemd notify [--ready] [--status TEXT] [KEY=VALUE]... sends a notification to $NOTIFY_SOCKET, like systemd-notify.
• notsystemd sudo, systemctl, systemd-run and timedatectl stand in for those commands; see Shims.

Every command takes --state-dir (default /run/notsystemd). run writes <state-dir>/<unit>.json on every state change, with the unit's ActiveState, SubState, StatusText, MainPID, result and its own PID, and puts the notify socket beside it. Mount the directory as an emptyDir shared by the pod's containers and each can read the others' state.

What run supports

• Type=simple, exec, notify, forking (with or without PIDFile=) and oneshot (with RemainAfterExit=). notify-reload runs as notify; dbus and idle run as simple.
• ExecStartPre=, ExecStart=, ExecStartPost=, ExecReload=, ExecStop= and ExecStopPost=, with the -, +, !, : and @ prefixes, quoting, ;-separated command lists, and $VAR/${VAR} expansion. ExecStop= and ExecStopPost= see MAINPID, SERVICE_RESULT, EXIT_CODE and EXIT_STATUS.
• User=, Group=, WorkingDirectory= (including - and ~), UMask= and PermissionsStartOnly=.
• Environment= and EnvironmentFile= (including -), layered over the container's own environment.
• RuntimeDirectory=, StateDirectory=, CacheDirectory=, LogsDirectory=, ConfigurationDirectory= and their *Mode=, created and owned by the unit's user; the runtime directory is removed on stop.
• $NOTIFY_SOCKET with READY=1, STATUS=, MAINPID=, RELOADING=1, STOPPING=1, WATCHDOG=1, WATCHDOG=trigger and WATCHDOG_USEC=, accepted from any process (NotifyAccess=all).
• WatchdogSec= and WatchdogSignal=: a missed ping signals the main process and fails the unit.
• TimeoutStartSec=, TimeoutStopSec=, TimeoutSec=, KillMode= and KillSignal=, escalating to SIGKILL when a stop times out.
• SuccessExitStatus=, with exit statuses and signal names.
• LimitNOFILE=, set on notsystemd itself so every command the unit runs inherits it, and lowered to what the kernel allows as systemd does.
• Drop-ins from <unit>.d/*.conf and from --drop-in-dir, and the %n, %N, %p, %i, %t, %S, %C, %L and %E specifiers.

As PID 1 it reaps every orphan in the container, and treats the whole container as the unit's control group when stopping it. Anywhere else it makes itself a child subreaper, so a daemon that double-forks stays visible.

Restart= is not implemented on purpose: the container exits with the service and Kubernetes restarts it. Dependencies between units (After=, Requires=, Wants=) are the pod's business too.

Shims

The software notsystemd exists for does more than get started by systemd: it runs systemctl, timedatectl and systemd-run, through sudo. Invoked under one of those names, notsystemd answers as that command, so an image links them to it; the flake's shims package holds the links. The shims find unit state in $NOTSYSTEMD_STATE_DIR, default /run/notsystemd.

• systemctl answers is-active, is-failed, is-enabled, show (with -p/--property and --value), status and is-system-running from the state directory. A unit the pod does not run reads as inactive, not enabled, and LoadState=not-found. stop, restart and their variants send the unit's notsystemd SIGTERM, so its container exits and Kubernetes starts it again; reload sends it SIGHUP, and kill signals the main process. Signalling a unit in another container needs shareProcessNamespace: true on the pod. start, enable, disable, mask, unmask, daemon-reload and reset-failed succeed without doing anything, since the pod spec decides which units run.
• timedatectl reports NTP as enabled and synchronised, because a container's clock is the node's, and accepts set-* without changing anything.
• systemd-run runs its command in place, as --uid/--gid with --setenv, waiting for it under --wait or --pipe and starting it in the background otherwise.
• sudo runs its command as -u/-g, keeping the environment apart from HOME, USER and LOGNAME, and asks for no password. Only root can switch users, as with sudo's -n.

Planned

A D-Bus facade serving org.freedesktop.systemd1 from the state directory, for software that watches units' ActiveState and StatusText over the system bus.

Development

nix develop (or direnv) provides Go, ginkgo and gomod2nix.

make test   # go tool ginkgo run -r
make build  # nix build .#, which also runs the specs
make check  # nix flake check
make tidy   # go mod tidy, then regenerate nix/gomod2nix.toml