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# cmc/ci-smoke is the nightly isolation canary; a runner scoped to named
 27# repositories never claims a build it is not scoped to, so the canary
 28# must be listed or its scheduled build waits forever.
 29#
 30# -cpus 3 of the host's 4 leaves a core for gitbayd and sshd while a build
 31# runs. No -memory: the e2e suite peaks past 5GB of the 7GB, and a cap
 32# that kills it is an outage rather than a limit (#144).
 33#
 34# ExecStart is overridden here rather than left in the unit so the flags
 35# and the sandboxing that has to match them live in one file: -isolation
 36# podman needs NoNewPrivileges=no below, and -image needs an image the
 37# host has been given (deploy/runner-podman-setup.sh, Containerfile.ci).
 38# podman's run root is pinned under the runner's home by storage.conf
 39# (runner-podman-setup.sh), not taken from XDG_RUNTIME_DIR or /tmp: this
 40# unit has PrivateTmp, so a /tmp run root is a per-instance tmpfs.
 41#
 42# The cgroupfs manager puts podman's pause process under the user slice,
 43# outside this unit's cgroup, so a stop does not end it and the next
 44# start joins its namespaces — including a /tmp that no longer exists.
 45# End it with the service.
 46ExecStopPost=-/usr/bin/pkill -u ci-runner -x catatonit
 47# On stop the runner drains: it claims nothing more and finishes the
 48# build in flight, then exits. Give it long enough — the per-build limit
 49# is 45m plus half a minute of report retries — before systemd kills it.
 50# `make deploy-runner` therefore waits for a running build (#179).
 51TimeoutStopSec=50min
 52# The drain only works if the stop signal reaches the runner alone. The
 53# default control-group mode sends SIGTERM to every process in the
 54# cgroup at once — the ssh session streaming the log and the build's
 55# container with it — so the runner drained a build whose steps were
 56# already dead. mixed signals the main process only; whatever is left
 57# when it exits is killed.
 58KillMode=mixed
 59ExecStart=
 60ExecStart=/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
 61Nice=10
 62CPUWeight=30
 63IOWeight=30
 64# NoNewPrivileges is off, and that is a deliberate trade (#144).
 65#
 66# Rootless podman sets up its user namespace with newuidmap, a setuid
 67# helper; NoNewPrivileges=yes blocks it and podman fails with
 68# "newuidmap: write to uid_map failed: Operation not permitted", so the
 69# runner refuses to start. The choice is between this flag and running
 70# builds in containers at all.
 71#
 72# Containers are the stronger boundary by a wide margin. NoNewPrivileges
 73# constrained a process that was already executing arbitrary repository
 74# code as this user; a container confines that code to an image and a
 75# bind-mounted workspace. What is lost is one hardening layer on the
 76# runner process itself, which is ours rather than a build's — a build no
 77# longer runs in this process's context at all.
 78#
 79# Under -isolation none there is no container, and this flag should be
 80# yes. Set it back if you run that way.
 81NoNewPrivileges=no
 82ProtectSystem=full
 83# ProtectKernelTunables is off, for the same reason NoNewPrivileges is
 84# (#144). It overmounts /proc/sys and friends in this unit's namespace,
 85# and the kernel then refuses a fresh proc mount in any child user
 86# namespace ("mount too revealing"): crun fails with "mount `proc` to
 87# `proc`: Operation not permitted". There is no podman setting for it.
 88# What the flag protected — /proc/sys from a build running on the host
 89# as this user — the container now covers: a build gets its own proc,
 90# with those paths masked by the runtime. Under -isolation none, set it
 91# back to yes.
 92ProtectControlGroups=yes
 93RestrictSUIDSGID=yes
 94Delegate=yes
 95# The runner's home is /var/lib/gitbay-runner (see the Admin page), and
 96# the leading - makes a missing path ignored rather than fatal: this
 97# drop-in installs on hosts that have not been prepared for podman yet,
 98# and a unit that refuses to start would stop every build on the
 99# instance.
100ReadWritePaths=-/var/lib/gitbay-runner/.local/share/containers -/var/lib/gitbay-runner/.config/containers