rock
ce154c55e6
Build and push runner images / build-runners (pull_request) Failing after 40s
ROOT CAUSE: Docker socket (/var/run/docker.sock) only existed inside the dind container — the runner container couldn't see it. The runner connected to dind via TCP (tcp://localhost:2376) with TLS. But workflow containers created by the runner had NO way to access the docker daemon: - unix socket not mounted (runner can't see it) - DOCKER_HOST env var not passed (runner.envs not configured) FIX: Share /var/run between dind and runner via emptyDir volume. When dind starts, it creates /var/run/docker.sock in the shared volume. Runner can now see the socket. docker_host: automount in config tells the runner to mount the socket into job containers automatically. Architecture after fix: dind container → creates /var/run/docker.sock → shared emptyDir runner container → sees /var/run/docker.sock → uses automount workflow container → gets /var/run/docker.sock mounted by runner Also removed runner.envs (TCP+TLS approach) — unix socket is simpler and works with automount.
40 lines
1.9 KiB
YAML
40 lines
1.9 KiB
YAML
# act_runner (the forgejo-runner binary) ships no config.yaml by default, so
|
|
# `forgejo-runner daemon` runs on its hardcoded defaults -- notably
|
|
# container.valid_volumes: [] ("if the sequence is empty, no volumes can be
|
|
# mounted"). Confirmed via `forgejo-runner generate-config` on this exact
|
|
# image (code.forgejo.org/forgejo/runner:6) and by running the daemon against
|
|
# a minimal override locally: a job container that requests any bind mount
|
|
# (e.g. the dind mTLS certs at /docker-certs/client, needed for
|
|
# `docker login`/build/push steps) is rejected outright with no default
|
|
# config in place.
|
|
#
|
|
# Scoped narrowly to exactly the certs path, read-only. Not a wildcard
|
|
# (valid_volumes: ['**']) -- that would let any workflow in any repo this
|
|
# runner serves bind-mount arbitrary paths off the runner pod's filesystem
|
|
# into a job container, which is a real widening of the CI trust boundary,
|
|
# not just a convenience.
|
|
#
|
|
# network: host is also only settable here, not per-workflow. A workflow's
|
|
# `container.options: --network host` is silently ignored -- confirmed live:
|
|
# every job's actual `docker create` call logged
|
|
# `network="FORGEJO-ACTIONS-TASK-N_..."`, an auto-generated per-job bridge,
|
|
# regardless of that options string. On that isolated bridge, DOCKER_HOST=
|
|
# tcp://localhost:2376 resolves to the job container itself (no daemon there),
|
|
# not to the dind sidecar, so any docker command that actually needs the
|
|
# daemon (build, push -- anything past docker login, which only talks to the
|
|
# registry over the network and never touches DOCKER_HOST) fails with "Cannot
|
|
# connect to the Docker daemon". host mode puts every job container in dind's
|
|
# own network namespace instead, where the daemon really is listening.
|
|
apiVersion: v1
|
|
kind: ConfigMap
|
|
metadata:
|
|
name: {{ .Release.Name }}-config
|
|
namespace: {{ .Release.Namespace }}
|
|
data:
|
|
config.yaml: |
|
|
container:
|
|
valid_volumes:
|
|
- /docker-certs/client
|
|
network: host
|
|
docker_host: automount
|