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[Unit]
 26# The host egress rule (#260, gitbay-runner-egress.nft) limits what this
 27# unit's user reaches on the host: 127.0.0.1:22 for the runner, the
 28# forge's public 22, 80 and 443 for builds, nothing else. Required, so
 29# the runner does not start without it: a table that failed to load must
 30# not mean builds reach the admin sshd.
 31Requires=gitbay-runner-egress.service
 32After=gitbay-runner-egress.service
 33
 34[Service]
 35# The runner polls as a non-admin account with a runner-scoped key, and
 36# claims only the repositories that key is attached to (`repo runner
 37# add`): krz/gitbay and cmc/ci-smoke. The attachments are the boundary,
 38# so ExecStart names no -repos. cmc/ci-smoke is the nightly isolation
 39# canary; keep it attached or its scheduled build waits forever.
 40#
 41# Two layers of resource caps. MemoryMax and CPUQuota bound the unit —
 42# the runner and every build together — which is what keeps the forge
 43# alive when a build allocates without bound. -memory and -cpus on
 44# ExecStart cap each build's own cgroup, which the runner creates under
 45# this unit's delegated cgroup (Delegate=yes below); podman's own --memory
 46# and --cpus never applied here, since under rootless cgroupfs the
 47# container ran in the service cgroup itself (#188).
 48#
 49# 6G of the host's 7.7GB, no swap: the e2e suite peaks past 5GB, so the
 50# cap sits above that rather than at a fair share. CPUQuota=300% and
 51# -cpus 3 are three of the four cores, leaving one for gitbayd and sshd
 52# (#144, #184). OOMPolicy=continue: systemd's default stops the whole
 53# service when any process in it is OOM-killed, which would end the
 54# runner mid-build; the build fails and the runner carries on.
 55MemoryMax=6G
 56CPUQuota=300%
 57OOMPolicy=continue
 58#
 59# ExecStart is overridden here rather than left in the unit so the flags
 60# and the sandboxing that has to match them live in one file: -isolation
 61# podman needs NoNewPrivileges=no below, and -image needs an image the
 62# host has been given (deploy/runner-podman-setup.sh, Containerfile.ci).
 63# podman's run root is pinned under the runner's home by storage.conf
 64# (runner-podman-setup.sh), not taken from XDG_RUNTIME_DIR or /tmp: this
 65# unit has PrivateTmp, so a /tmp run root is a per-instance tmpfs.
 66#
 67# The cgroupfs manager puts podman's pause process under the user slice,
 68# outside this unit's cgroup, so a stop does not end it and the next
 69# start joins its namespaces — including a /tmp that no longer exists.
 70# End it with the service.
 71ExecStopPost=-/usr/bin/pkill -u ci-runner -x catatonit
 72# On stop the runner drains: it claims nothing more and finishes the
 73# build in flight, then exits. Give it long enough — the per-build limit
 74# is 45m plus half a minute of report retries — before systemd kills it.
 75# `make deploy-runner` therefore waits for a running build (#179).
 76TimeoutStopSec=50min
 77# The drain only works if the stop signal reaches the runner alone. The
 78# default control-group mode sends SIGTERM to every process in the
 79# cgroup at once — the ssh session streaming the log and the build's
 80# container with it — so the runner drained a build whose steps were
 81# already dead. mixed signals the main process only; whatever is left
 82# when it exits is killed.
 83KillMode=mixed
 84# -untrusted: this runner isolates in podman, so it takes merge request
 85# heads from forks; a runner without a container must not.
 86ExecStart=
 87ExecStart=/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
 88Nice=10
 89CPUWeight=30
 90IOWeight=30
 91# NoNewPrivileges is off, and that is a deliberate trade (#144).
 92#
 93# Rootless podman sets up its user namespace with newuidmap, a setuid
 94# helper; NoNewPrivileges=yes blocks it and podman fails with
 95# "newuidmap: write to uid_map failed: Operation not permitted", so the
 96# runner refuses to start. The choice is between this flag and running
 97# builds in containers at all.
 98#
 99# Containers are the stronger boundary by a wide margin. NoNewPrivileges
100# constrained a process that was already executing arbitrary repository
101# code as this user; a container confines that code to an image and a
102# bind-mounted workspace. What is lost is one hardening layer on the
103# runner process itself, which is ours rather than a build's — a build no
104# longer runs in this process's context at all.
105#
106# Under -isolation none there is no container, and this flag should be
107# yes. Set it back if you run that way.
108NoNewPrivileges=no
109ProtectSystem=full
110# ProtectKernelTunables is off, for the same reason NoNewPrivileges is
111# (#144). It overmounts /proc/sys and friends in this unit's namespace,
112# and the kernel then refuses a fresh proc mount in any child user
113# namespace ("mount too revealing"): crun fails with "mount `proc` to
114# `proc`: Operation not permitted". There is no podman setting for it.
115# What the flag protected — /proc/sys from a build running on the host
116# as this user — the container now covers: a build gets its own proc,
117# with those paths masked by the runtime. Under -isolation none, set it
118# back to yes.
119# ProtectControlGroups is off because the runner writes cgroups: it
120# creates one per build under this unit's delegated cgroup to carry
121# -memory and -cpus (#188). The flag mounts /sys/fs/cgroup read-only in
122# the unit's namespace, which makes even a delegated cgroup unwritable,
123# and the runner then refuses to start when a limit is set. Delegate=yes
124# already hands this unit its subtree; what the flag protected beyond
125# that is other units' cgroups, which are root-owned and not writable
126# by this user regardless.
127ProtectControlGroups=no
128RestrictSUIDSGID=yes
129Delegate=yes
130# The runner's home is /var/lib/gitbay-runner (see the Admin page), and
131# the leading - makes a missing path ignored rather than fatal: this
132# drop-in installs on hosts that have not been prepared for podman yet,
133# and a unit that refuses to start would stop every build on the
134# instance.
135ReadWritePaths=-/var/lib/gitbay-runner/.local/share/containers -/var/lib/gitbay-runner/.config/containers