deploy/gitbay-runner.override.conf

  1# Drop-in for gitbay-runner.service, installed by `make deploy-runner` to
  2# /etc/systemd/system/gitbay-runner.service.d/override.conf.
  3#
  4# A build must never starve the host: the e2e suite alone starts sixty
  5# daemon instances, and with nothing holding it back a deploy's scp on
  6# the admin sshd stalled at 1%. Lower CPU and IO weight keep sshd,
  7# gitbayd and the backup timers responsive while a build runs.
  8#
  9# These weights are for the service, not per build, so `-jobs N` divides
 10# them among N builds rather than taking N times as much. Raising -jobs
 11# does not need them raised; it makes each build slower, not the host
 12# busier.
 13#
 14# A build runs whatever the repository's ci.yml says, as the runner's
 15# own user. Keep that user unprivileged: its key is added with
 16# `keys add --scope runner`, which confines it to the runner protocol
 17# and read-only git, and the sandboxing below keeps a step from
 18# touching the system outside its workspace.
 19#
 20# Delegate=yes and the storage path below are what rootless podman needs
 21# (#144): it manages its own cgroups for a container, and its image and
 22# container store lives under the runner's home, which ProtectSystem
 23# would otherwise make read-only. Prepare the host with
 24# deploy/runner-podman-setup.sh before deploying a runner that isolates.
 25[Service]
 26# The runner polls as a non-admin account with a runner-scoped key, and
 27# claims only the repositories that key is attached to (`repo runner
 28# add`): krz/gitbay and cmc/ci-smoke. The attachments are the boundary,
 29# so ExecStart names no -repos. cmc/ci-smoke is the nightly isolation
 30# canary; keep it attached or its scheduled build waits forever.
 31#
 32# Two layers of resource caps. MemoryMax and CPUQuota bound the unit —
 33# the runner and every build together — which is what keeps the forge
 34# alive when a build allocates without bound. -memory and -cpus on
 35# ExecStart cap each build's own cgroup, which the runner creates under
 36# this unit's delegated cgroup (Delegate=yes below); podman's own --memory
 37# and --cpus never applied here, since under rootless cgroupfs the
 38# container ran in the service cgroup itself (#188).
 39#
 40# 6G of the host's 7.7GB, no swap: the e2e suite peaks past 5GB, so the
 41# cap sits above that rather than at a fair share. CPUQuota=300% and
 42# -cpus 3 are three of the four cores, leaving one for gitbayd and sshd
 43# (#144, #184). OOMPolicy=continue: systemd's default stops the whole
 44# service when any process in it is OOM-killed, which would end the
 45# runner mid-build; the build fails and the runner carries on.
 46MemoryMax=6G
 47CPUQuota=300%
 48OOMPolicy=continue
 49#
 50# ExecStart is overridden here rather than left in the unit so the flags
 51# and the sandboxing that has to match them live in one file: -isolation
 52# podman needs NoNewPrivileges=no below, and -image needs an image the
 53# host has been given (deploy/runner-podman-setup.sh, Containerfile.ci).
 54# podman's run root is pinned under the runner's home by storage.conf
 55# (runner-podman-setup.sh), not taken from XDG_RUNTIME_DIR or /tmp: this
 56# unit has PrivateTmp, so a /tmp run root is a per-instance tmpfs.
 57#
 58# The cgroupfs manager puts podman's pause process under the user slice,
 59# outside this unit's cgroup, so a stop does not end it and the next
 60# start joins its namespaces — including a /tmp that no longer exists.
 61# End it with the service.
 62ExecStopPost=-/usr/bin/pkill -u ci-runner -x catatonit
 63# On stop the runner drains: it claims nothing more and finishes the
 64# build in flight, then exits. Give it long enough — the per-build limit
 65# is 45m plus half a minute of report retries — before systemd kills it.
 66# `make deploy-runner` therefore waits for a running build (#179).
 67TimeoutStopSec=50min
 68# The drain only works if the stop signal reaches the runner alone. The
 69# default control-group mode sends SIGTERM to every process in the
 70# cgroup at once — the ssh session streaming the log and the build's
 71# container with it — so the runner drained a build whose steps were
 72# already dead. mixed signals the main process only; whatever is left
 73# when it exits is killed.
 74KillMode=mixed
 75# -untrusted: this runner isolates in podman, so it takes merge request
 76# heads from forks; a runner without a container must not.
 77ExecStart=
 78ExecStart=/usr/local/bin/gitbay-runner -remote git@127.0.0.1 -workdir /var/lib/gitbay-runner/work -poll 5s -timeout 45m -isolation podman -image localhost/gitbay-ci:2 -cpus 3 -memory 6g -untrusted
 79Nice=10
 80CPUWeight=30
 81IOWeight=30
 82# NoNewPrivileges is off, and that is a deliberate trade (#144).
 83#
 84# Rootless podman sets up its user namespace with newuidmap, a setuid
 85# helper; NoNewPrivileges=yes blocks it and podman fails with
 86# "newuidmap: write to uid_map failed: Operation not permitted", so the
 87# runner refuses to start. The choice is between this flag and running
 88# builds in containers at all.
 89#
 90# Containers are the stronger boundary by a wide margin. NoNewPrivileges
 91# constrained a process that was already executing arbitrary repository
 92# code as this user; a container confines that code to an image and a
 93# bind-mounted workspace. What is lost is one hardening layer on the
 94# runner process itself, which is ours rather than a build's — a build no
 95# longer runs in this process's context at all.
 96#
 97# Under -isolation none there is no container, and this flag should be
 98# yes. Set it back if you run that way.
 99NoNewPrivileges=no
100ProtectSystem=full
101# ProtectKernelTunables is off, for the same reason NoNewPrivileges is
102# (#144). It overmounts /proc/sys and friends in this unit's namespace,
103# and the kernel then refuses a fresh proc mount in any child user
104# namespace ("mount too revealing"): crun fails with "mount `proc` to
105# `proc`: Operation not permitted". There is no podman setting for it.
106# What the flag protected — /proc/sys from a build running on the host
107# as this user — the container now covers: a build gets its own proc,
108# with those paths masked by the runtime. Under -isolation none, set it
109# back to yes.
110# ProtectControlGroups is off because the runner writes cgroups: it
111# creates one per build under this unit's delegated cgroup to carry
112# -memory and -cpus (#188). The flag mounts /sys/fs/cgroup read-only in
113# the unit's namespace, which makes even a delegated cgroup unwritable,
114# and the runner then refuses to start when a limit is set. Delegate=yes
115# already hands this unit its subtree; what the flag protected beyond
116# that is other units' cgroups, which are root-owned and not writable
117# by this user regardless.
118ProtectControlGroups=no
119RestrictSUIDSGID=yes
120Delegate=yes
121# The runner's home is /var/lib/gitbay-runner (see the Admin page), and
122# the leading - makes a missing path ignored rather than fatal: this
123# drop-in installs on hosts that have not been prepared for podman yet,
124# and a unit that refuses to start would stop every build on the
125# instance.
126ReadWritePaths=-/var/lib/gitbay-runner/.local/share/containers -/var/lib/gitbay-runner/.config/containers