The image every tenant boots lived under a personal account, gitea.mynodes.xyz/ilker/hpc-node. It is now hpcod/node-image, with a second organisation, hpcapps, for the application images a tenant can request. Both are public, because this registry exists to distribute code and images and is meant to be readable. That last sentence is a correction. This change was started on the strength of a measurement I read as an exposure: an unrelated Gitea account could fetch the node image manifest. It is not an exposure, it is the intended posture - and the measurement was wrong in the other direction too. Testing a container registry with plain curl gets a 401 because the client is expected to fetch a bearer token first; doing the token exchange shows the image is anonymously pullable, more open than I had reported. The namespace move is still worth having, but for the ordinary reason: platform infrastructure should not sit under one person's username, and applications should have a namespace of their own. The useful consequence is that the registry credential can go. If the image pulls anonymously, nothing in the provisioning path needs a token: registry_pull_user is empty, the credential tasks skip, and a template builds and imports the image with no secret involved - verified, the run skipped three tasks and still pulled. The mechanism stays for a site whose registry is private. The best-handled secret is the one that is not in the path. Also verified on the way: a tenant provisions from the new location, the import taking 27s, with no other change.
70 lines
2.8 KiB
Bash
Executable File
70 lines
2.8 KiB
Bash
Executable File
#!/bin/bash
|
|
# Build the compute node image and publish it.
|
|
#
|
|
# Run on the control plane host. The tag is the cache key a tenant request
|
|
# resolves to, so bump it whenever the stack changes.
|
|
#
|
|
# Rootless podman as a service account needs two things that a normal login
|
|
# gets for free, and fails differently for each:
|
|
#
|
|
# /etc/subuid, /etc/subgid a range for the account, or unpacking a layer
|
|
# dies on "lchown: invalid argument" - the build
|
|
# cannot map uid 0 inside the namespace.
|
|
# loginctl enable-linger a persistent user session, or runc cannot create
|
|
# the build container for want of a delegated
|
|
# cgroup.
|
|
#
|
|
# Neither can be set with usermod while a service is running as the account,
|
|
# so /etc/subuid and /etc/subgid are edited directly. See PREFLIGHT below.
|
|
set -euo pipefail
|
|
|
|
REGISTRY=${REGISTRY:-gitea.mynodes.xyz}
|
|
# The platform organisation rather than a personal account. Organisational,
|
|
# not a security boundary - this registry is public on purpose.
|
|
OWNER=${OWNER:-hpcod}
|
|
NAME=${NAME:-node-image}
|
|
TAG=${TAG:-rocky9-ohpc3}
|
|
IMAGE="${REGISTRY}/${OWNER}/${NAME}:${TAG}"
|
|
|
|
# Keep the registry credential somewhere stable. Podman defaults to
|
|
# $XDG_RUNTIME_DIR/containers/auth.json, so a login performed under one runtime
|
|
# directory is invisible to a build run under another - which shows up as
|
|
# "unauthorized" on push, long after a successful login.
|
|
export REGISTRY_AUTH_FILE="${REGISTRY_AUTH_FILE:-$HOME/.config/containers/auth.json}"
|
|
|
|
cd "$(dirname "$0")"
|
|
|
|
: "${XDG_RUNTIME_DIR:?set XDG_RUNTIME_DIR=/run/user/\$(id -u) - rootless podman needs it}"
|
|
|
|
# PREFLIGHT: fail with the actual cause rather than a layer-unpacking error.
|
|
if ! grep -q "^$(id -un):" /etc/subuid 2>/dev/null; then
|
|
echo "no subuid range for $(id -un). As root:" >&2
|
|
echo " echo '$(id -un):400000:65536' >> /etc/subuid" >&2
|
|
echo " echo '$(id -un):400000:65536' >> /etc/subgid" >&2
|
|
echo " loginctl enable-linger $(id -u)" >&2
|
|
exit 1
|
|
fi
|
|
if [ ! -d "/run/user/$(id -u)" ]; then
|
|
echo "no user runtime dir. As root: loginctl enable-linger $(id -u)" >&2
|
|
exit 1
|
|
fi
|
|
|
|
echo "==> building ${IMAGE}"
|
|
# --squash-all, not --squash: --squash keeps the base image layers, so files
|
|
# deleted in our layers still ship in the layers underneath. Only squashing
|
|
# everything into a single layer actually shrinks what gets pulled.
|
|
podman build --squash-all -t "${IMAGE}" .
|
|
|
|
echo "==> pushing"
|
|
if ! podman push "${IMAGE}"; then
|
|
echo "push failed. If this says unauthorized, log in with the same" >&2
|
|
echo "REGISTRY_AUTH_FILE this script uses:" >&2
|
|
echo " REGISTRY_AUTH_FILE=${REGISTRY_AUTH_FILE} podman login ${REGISTRY} -u <user>" >&2
|
|
exit 1
|
|
fi
|
|
|
|
echo
|
|
echo "IMAGE PUBLISHED: ${IMAGE}"
|
|
echo "Tenants import it with:"
|
|
echo " wwctl image import docker://${IMAGE} <name>"
|