# The control plane as a systemd unit. One VM, systemd, binaries from CI — not Kubernetes: a tmctl
# child runs for HOURS and holds an exclusive lock on the book's files on the local disk, so any
# orchestrator that can move a pod mid-run is hostile to this workload (PLATFORM_DIRECTION §3).
#
# ⚠ A RUN IS NOT A CHILD OF THIS UNIT (D39.106, ratified). Each one is a transient unit in this
# user's OWN systemd manager, in tm-runs.slice, because a translation has to outlive a deploy or a
# crash of the control plane and a child cannot. What follows from that, and is easy to get wrong:
#
#   * `loginctl enable-linger tmplatform` is REQUIRED, once, at install: without it the user manager
#     does not exist outside a login session and no run can be created at all;
#   * this unit has to reach its own user bus, which is why ProtectHome= is `tmpfs` with a bind
#     rather than `yes` (measured: ProtectHome=yes makes /run/user/<uid> inaccessible and the bus
#     unreachable). DBUS_SESSION_BUS_ADDRESS alone is enough — XDG_RUNTIME_DIR is not needed;
#   * the limits at the bottom of this file bound THE CONTROL PLANE ONLY. A run is bounded by its
#     own cgroup in tm-runs.slice (TM_PLATFORM_RUN_MEMORY_MAX / _TASKS_MAX) — that is the answer to
#     PD-13, and it is measured rather than declared (STACK_DECISIONS §16).

[Unit]
Description=TextMachine control plane
After=network-online.target postgresql.service
Wants=network-online.target
# ⚠ ADD, at install, an ordering on this service user's own systemd manager:
#     After=user@<uid of tmplatform>.service
# BindPaths=/run/user/%U below needs that directory to exist at start, and it is logind that creates
# it. Without the line the first boot is a race; Restart=on-failure heals it after RestartSec, so the
# symptom is a service that comes up on its second try rather than one that stays down. The UID is
# site-specific, which is why it is not written here.

[Service]
# Type=exec, not notify: the binary does not speak sd_notify, and claiming it does would make
# systemd wait for a readiness signal that never comes.
Type=exec
ExecStart=/usr/local/bin/tmplatformd
User=tmplatform
Group=tmplatform

# KillMode=mixed: SIGTERM to the main process only, so the platform runs its own drain and stops
# its children the way the engine expects; SIGKILL to everything left when the timeout runs out.
KillMode=mixed
KillSignal=SIGTERM
# Longer than the platform's own drain (15s) plus the engine's stop grace (30s), or systemd would
# SIGKILL a tmctl mid-shutdown and leave its project lock behind.
TimeoutStopSec=90
Restart=on-failure
RestartSec=5s

# Secrets as credentials, not as environment: an environment variable is visible in
# /proc/<pid>/environ and is inherited by every tmctl child. The config reads *_FILE first.
LoadCredential=dsn:/etc/tmplatform/dsn
LoadCredential=oidc_client_secret:/etc/tmplatform/oidc_client_secret
Environment=TM_PLATFORM_DSN_FILE=%d/dsn
Environment=TM_PLATFORM_OIDC_CLIENT_SECRET_FILE=%d/oidc_client_secret
# The service's own user manager, which is where run units are created. %U is this unit's UID.
Environment=DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/%U/bus
EnvironmentFile=/etc/tmplatform/env

# Books live outside the repository and outside /var/lib; the unit gets the one directory it may
# write and nothing else. /srv/textmachine is the SERVER's books root — "~/books" is the owner's
# decision for a development machine, and under this unit a home directory is not reachable at all
# (see ProtectHome below).
#
# The directory must EXIST before the first start: with ProtectSystem=strict a ReadWritePaths= that
# names a missing path fails the mount-namespace setup and the unit does not come up. The deploy
# notes create it, and the "-" prefix that would make the line optional is deliberately not used —
# a service with no writable directory should fail at start, not at the first write hours in.
ReadWritePaths=/srv/textmachine
StateDirectory=tmplatform

# Sandboxing. Free, and it bounds what a compromised process reaches.
ProtectSystem=strict
# /home and /root stay invisible — the platform has no business in anyone's home, which is also why
# the books root is under /srv. An operator who must keep a library under a home directory adds it to
# BindPaths= below alongside the runtime directory.
#
# tmpfs plus a bind of the user runtime directory, NOT `yes`: /home and /root stay invisible, and
# the one thing that has to remain reachable — this user's systemd bus, through which every run is
# created — does. Measured, both ways: with ProtectHome=yes the directory is Permission denied and
# the bus is gone; with these two lines the socket is there.
# ⚠ The path must exist when this unit starts, which is what `enable-linger` guarantees.
ProtectHome=tmpfs
BindPaths=/run/user/%U
PrivateTmp=yes
PrivateDevices=yes
NoNewPrivileges=yes
ProtectKernelTunables=yes
ProtectKernelModules=yes
ProtectControlGroups=yes
RestrictSUIDSGID=yes
RestrictRealtime=yes
LockPersonality=yes
RestrictAddressFamilies=AF_INET AF_INET6 AF_UNIX

# These bound the unit's CGROUP, and by the argument at the top of this file every tmctl the
# platform spawns lives in it. So they do not bound "the control plane" — they bound the control
# plane plus every run in flight, together (PD-55). Two consequences follow, and both are decided
# here rather than discovered in production:
#
#  1. The ceiling is a machine backstop, not a service sizing. systemd.resource-control(5) calls
#     MemoryMax= "the last line of defense"; as a percentage it needs no knowledge of the box.
#  2. OOMPolicy is set explicitly. The system default is `stop`, which would take the control plane
#     down over one OOM inside it and land the unit in oom-kill failed state for Restart= to pick up.
#     `continue` logs the kill and keeps the service running.
#
# ⚠ These numbers bound the CONTROL PLANE and nothing else. Runs are not in this cgroup at all — each
# is a transient unit in tm-runs.slice with its own MemoryMax/MemorySwapMax/TasksMax
# (TM_PLATFORM_RUN_*). That is the answer to PD-13, and it is measured (STACK_DECISIONS §16).
MemoryMax=80%
OOMPolicy=continue
# Platform plus concurrent runs, each a Go process with a few dozen threads: room for roughly a
# dozen runs on one VM, which is more than a single box will carry.
TasksMax=512
LimitNOFILE=8192

[Install]
WantedBy=multi-user.target