# Drop-in for gitbay-runner.service, installed by `make deploy-runner` to
# /etc/systemd/system/gitbay-runner.service.d/override.conf.
#
# A build must never starve the host: the e2e suite alone starts sixty
# daemon instances, and with nothing holding it back a deploy's scp on
# the admin sshd stalled at 1%. Lower CPU and IO weight keep sshd,
# gitbayd and the backup timers responsive while a build runs.
#
# These weights are for the service, not per build, so `-jobs N` divides
# them among N builds rather than taking N times as much. Raising -jobs
# does not need them raised; it makes each build slower, not the host
# busier.
#
# A build runs whatever the repository's ci.yml says, as the runner's
# own user. Keep that user unprivileged: its key is added with
# `keys add --scope runner`, which confines it to the runner protocol
# and read-only git, and the sandboxing below keeps a step from
# touching the system outside its workspace.
#
# Delegate=yes and the storage path below are what rootless podman needs
# (#144): it manages its own cgroups for a container, and its image and
# container store lives under the runner's home, which ProtectSystem
# would otherwise make read-only. Prepare the host with
# deploy/runner-podman-setup.sh before deploying a runner that isolates.
[Service]
# cmc/ci-smoke is the nightly isolation canary; a runner scoped to named
# repositories never claims a build it is not scoped to, so the canary
# must be listed or its scheduled build waits forever.
#
# Two layers of resource caps. MemoryMax and CPUQuota bound the unit —
# the runner and every build together — which is what keeps the forge
# alive when a build allocates without bound. -memory and -cpus on
# ExecStart cap each build's own cgroup, which the runner creates under
# this unit's delegated cgroup (Delegate=yes below); podman's own --memory
# and --cpus never applied here, since under rootless cgroupfs the
# container ran in the service cgroup itself (#188).
#
# 6G of the host's 7.7GB, no swap: the e2e suite peaks past 5GB, so the
# cap sits above that rather than at a fair share. CPUQuota=300% and
# -cpus 3 are three of the four cores, leaving one for gitbayd and sshd
# (#144, #184). OOMPolicy=continue: systemd's default stops the whole
# service when any process in it is OOM-killed, which would end the
# runner mid-build; the build fails and the runner carries on.
MemoryMax=6G
CPUQuota=300%
OOMPolicy=continue
#
# ExecStart is overridden here rather than left in the unit so the flags
# and the sandboxing that has to match them live in one file: -isolation
# podman needs NoNewPrivileges=no below, and -image needs an image the
# host has been given (deploy/runner-podman-setup.sh, Containerfile.ci).
# podman's run root is pinned under the runner's home by storage.conf
# (runner-podman-setup.sh), not taken from XDG_RUNTIME_DIR or /tmp: this
# unit has PrivateTmp, so a /tmp run root is a per-instance tmpfs.
#
# The cgroupfs manager puts podman's pause process under the user slice,
# outside this unit's cgroup, so a stop does not end it and the next
# start joins its namespaces — including a /tmp that no longer exists.
# End it with the service.
ExecStopPost=-/usr/bin/pkill -u ci-runner -x catatonit
# On stop the runner drains: it claims nothing more and finishes the
# build in flight, then exits. Give it long enough — the per-build limit
# is 45m plus half a minute of report retries — before systemd kills it.
# `make deploy-runner` therefore waits for a running build (#179).
TimeoutStopSec=50min
# The drain only works if the stop signal reaches the runner alone. The
# default control-group mode sends SIGTERM to every process in the
# cgroup at once — the ssh session streaming the log and the build's
# container with it — so the runner drained a build whose steps were
# already dead. mixed signals the main process only; whatever is left
# when it exits is killed.
KillMode=mixed
# -untrusted: this runner isolates in podman, so it takes merge request
# heads from forks; a runner without a container must not.
ExecStart=
ExecStart=/usr/local/bin/gitbay-runner -remote git@127.0.0.1 -workdir /var/lib/gitbay-runner/work -poll 5s -timeout 45m -repos krz/gitbay,cmc/ci-smoke -isolation podman -image localhost/gitbay-ci:1 -cpus 3 -memory 6g -untrusted
Nice=10
CPUWeight=30
IOWeight=30
# NoNewPrivileges is off, and that is a deliberate trade (#144).
#
# Rootless podman sets up its user namespace with newuidmap, a setuid
# helper; NoNewPrivileges=yes blocks it and podman fails with
# "newuidmap: write to uid_map failed: Operation not permitted", so the
# runner refuses to start. The choice is between this flag and running
# builds in containers at all.
#
# Containers are the stronger boundary by a wide margin. NoNewPrivileges
# constrained a process that was already executing arbitrary repository
# code as this user; a container confines that code to an image and a
# bind-mounted workspace. What is lost is one hardening layer on the
# runner process itself, which is ours rather than a build's — a build no
# longer runs in this process's context at all.
#
# Under -isolation none there is no container, and this flag should be
# yes. Set it back if you run that way.
NoNewPrivileges=no
ProtectSystem=full
# ProtectKernelTunables is off, for the same reason NoNewPrivileges is
# (#144). It overmounts /proc/sys and friends in this unit's namespace,
# and the kernel then refuses a fresh proc mount in any child user
# namespace ("mount too revealing"): crun fails with "mount `proc` to
# `proc`: Operation not permitted". There is no podman setting for it.
# What the flag protected — /proc/sys from a build running on the host
# as this user — the container now covers: a build gets its own proc,
# with those paths masked by the runtime. Under -isolation none, set it
# back to yes.
# ProtectControlGroups is off because the runner writes cgroups: it
# creates one per build under this unit's delegated cgroup to carry
# -memory and -cpus (#188). The flag mounts /sys/fs/cgroup read-only in
# the unit's namespace, which makes even a delegated cgroup unwritable,
# and the runner then refuses to start when a limit is set. Delegate=yes
# already hands this unit its subtree; what the flag protected beyond
# that is other units' cgroups, which are root-owned and not writable
# by this user regardless.
ProtectControlGroups=no
RestrictSUIDSGID=yes
Delegate=yes
# The runner's home is /var/lib/gitbay-runner (see the Admin page), and
# the leading - makes a missing path ignored rather than fatal: this
# drop-in installs on hosts that have not been prepared for podman yet,
# and a unit that refuses to start would stop every build on the
# instance.
ReadWritePaths=-/var/lib/gitbay-runner/.local/share/containers -/var/lib/gitbay-runner/.config/containers