mirror of
https://github.com/pollen-robotics/microduck.git
synced 2026-11-22 06:29:20 +08:00
The last item of the plan in install-path-gap.md, and the one everything else is
blind to: every other check in the tree drives the engine against a stub
`systemctl`, so none of them can see whether a restart *happened*.
`scripts/systemd-test.sh` boots systemd as pid 1, mints three signed releases
carrying real units and a real `updaterd`, installs one, applies the next through
the running daemon, and then applies one shipping a unit that cannot start. It
observes four things nothing else can:
- `on_apply` really restarts what the release ships — asserted on the unit's main
PID changing, not on a command having been issued;
- the deferred transient timer really fires and really replaces `updaterd`, in
about four seconds, and the successor runs the new release. A child process
could not: it would sit in the cgroup being killed. That was the reason for
`systemd-run` and, until now, an untested claim;
- `hooks/postinstall` really installs, enables and starts a unit the board has
never had;
- a unit that installs cleanly and cannot start fails the update and names itself,
rather than reverting with "not healthy within 30s: unreachable" — bug 1, and the
only class the cheaper items cannot reach.
Docker rather than systemd-nspawn, which the plan proposed. The objection to
privileged containers *in CI* stands and this is not CI; what decided it is the
plan's own argument for ranking this last — a check that can only run on a machine
nobody develops on stops being read. Docker Desktop runs systemd as pid 1 on the
laptop this was written on. nspawn does not.
The fixture is its own example rather than a flag on `fake-release`, which
deliberately mints releases with no units because it exists to drive the engine
where systemd is not present.
Two injections are left, both cheap now the harness exists: `systemd-run` failing,
and `enable --now` on a unit whose `User=` does not exist.
Not in CI: it needs --privileged and the host's cgroup filesystem, and the `board`
job is already the slowest thing in the loop. Run it when `on_apply`, the deferred
restarts or the hook change.
Assisted-by: Claude:claude-opus-5
22 lines
1.1 KiB
Docker
22 lines
1.1 KiB
Docker
# The image `scripts/systemd-test.sh` boots, and the only place in this repository where systemd
|
|
# runs as pid 1.
|
|
#
|
|
# Trixie, unlike `dev-build.Dockerfile`: this one *runs* what a board runs rather than building it,
|
|
# so it should be the userland we ship rather than the older one binaries are built against.
|
|
#
|
|
# `systemd` and `dbus` only. The point is a real init with real cgroups and real transient timers,
|
|
# and everything else the update needs — the binaries, the units, the hook — arrives inside the
|
|
# signed releases the harness mints, exactly as it does on a board.
|
|
FROM debian:trixie-slim
|
|
RUN apt-get update \
|
|
&& apt-get install -y --no-install-recommends systemd dbus \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
# systemd's own convention for "stop, cleanly". Without it `docker stop` sends SIGTERM, which pid 1
|
|
# systemd reads as a request to re-exec rather than to shut down.
|
|
STOPSIGNAL SIGRTMIN+3
|
|
|
|
# `/lib/systemd/systemd`, because Debian's slim images ship no `/sbin/init` symlink — the harness
|
|
# passes it explicitly and this only records why.
|
|
CMD ["/lib/systemd/systemd"]
|