FelisMC/Felis
2
0 Fork 0
Code Issues Pull Requests 6 Actions Packages Projects Releases Wiki Activity
Files
b7fdef7522283b02087810036560517d55ca3011
Felis/deploy/bootstrap.sh
T

4219 lines
198 KiB
Bash
Raw Blame History

#!/usr/bin/env bash
#
# Felis one-line bootstrap installer.
#
# curl -fsSL <raw-url>/deploy/bootstrap.sh | sudo bash
#
# Brings a fresh single-node Linux host from nothing to a running Felis control
# plane: it installs whatever is missing (picking apt/dnf/yum/zypper/pacman by OS), provisions a
# swap file on tiny hosts, then configures Docker, k3s and PostgreSQL, builds and
# imports the felis image, runs database migrations and applies the rendered
# install bundle (CRD + namespaces + RBAC + NetworkPolicies + control-plane
# Deployments + in-cluster registry).
#
# The recommended entrypoint is now `sudo felis setup`, which wraps this
# bootstrap in a TUI and then continues to the Owner/edge setup. This script
# remains usable directly for raw host provisioning.
#
# At the start it asks what to install:
# [1] Felis — the full control plane described above.
# [2] Felis-nano — ONLY the Yggdrasil hasJoined multiplexer (`felis nano`) as a
# systemd service: no k3s, no Postgres, no bundle. For a
# third-party server operator who just wants multi-Yggdrasil
# auth federation. Preselect non-interactively with
# FELIS_INSTALL_MODE=nano.
#
# The script is idempotent: re-running it converges rather than duplicating, and
# generated secrets are persisted to /etc/felis/secrets.env so reruns reuse them.
#
# Tunables override the demo defaults. sudo resets the environment, so a variable exported
# before `curl ... | sudo bash` never arrives. Name it on the sudo line, or keep it with -E:
# curl -fsSL <raw-url>/deploy/bootstrap.sh | sudo FELIS_INSTALL_MODE=nano FELIS_NANO_LISTEN=10.0.0.5:8081 bash
# export FELIS_INSTALL_MODE=nano; curl -fsSL <raw-url>/deploy/bootstrap.sh | sudo -E bash
# FELIS_INSTALL_MODE full|nano — skip the prompt (default: ask on a tty, else full; nano
# instead on a host that runs felis-nano and no full install)
# FELIS_NANO_LISTEN listen addr for `felis nano` (default: the address an installed
# felis-nano already uses, else 127.0.0.1:8081 — loopback only; set a
# private-network IP to serve an off-host proxy)
# FELIS_NANO_PROXY_CIDR the proxy allowed to reach a non-loopback nano bind, as an address
# with a prefix length (for example 10.0.0.7/32). firewalld opens the
# port to that source only; unset, it opens nothing
# FELIS_LEGACY_FORWARDING_SERVERS comma-separated backends that receive their identity
# through the handshake address instead of modern forwarding
# (default: legacy18). Read once at Velocity start, so changing it
# means re-running this script and restarting the proxy.
# FELIS_VELOCITY_XMX maximum heap of the Velocity proxy, as <n>M or <n>G (default: 1G;
# at least 256M). docs/operations.md sizes it by player count.
# FELIS_VELOCITY_FORK_JAR path to a Felis-Legacy Velocity fork build to install as the
# proxy instead of the stock download (default: unset, stock).
# FELIS_VELOCITY_FORK_JAR_SHA256 expected sha256 of that jar. REQUIRED whenever the jar
# above is set; the install refuses on a mismatch.
# FELIS_GAME_STACK pinned|latest — which Limbo, Paper, LuckPerms and Velocity builds to
# install (default: pinned, the builds deploy/game-stack.lock names,
# each checked against its sha256). latest resolves upstream's newest
# builds on every run; the Minecraft version then follows Limbo's CI.
# FELIS_JRE_VERSION Temurin feature version for the proxy (default: 25, whose build and
# digests are pinned; a rerun moves an installer-managed JRE to the
# pinned build). Another feature version is checked against the
# digest Adoptium's API publishes for it.
# FELIS_GO_VERSION Go toolchain used to build the nano binary (default: 1.26.8)
# FELIS_GO_SHA256 sha256 of that version's linux tarball for this host's architecture.
# REQUIRED for a non-default FELIS_GO_VERSION; the default's is pinned.
# FELIS_K3S_VERSION k3s release a fresh install gets (default: v1.36.4+k3s1). An
# installed k3s is left alone unless FELIS_UPGRADE_DEPS=1.
# FELIS_CLOUDFLARED_VERSION / FELIS_CLOUDFLARED_SHA256 cloudflared release installed
# when none is present (default: 2026.9.1, digests pinned); the
# sha256 is REQUIRED for any other version
# FELIS_UPGRADE_DEPS 1 moves an installed k3s and cloudflared to the versions above
# (k3s one minor version at a time; neither is ever downgraded) and
# restarts cloudflared-felis onto the new binary (default: 0)
# FELIS_REPO_URL git URL to build from (raw script mode only)
# FELIS_VERSION_BOOTSTRAP release|dev — which version to install (default: release).
# release DOWNLOADS the prebuilt felis binary published for the newest
# tag (panel included — it is go:embed'ed into that same binary) and
# builds only a thin image around it; dev clones and compiles. If the
# asset is missing or this architecture has none, release warns and falls
# back to compiling the SAME tag. The downloaded binary must match
# the release's SHA256SUMS before it is run; a release without one
# is compiled from source too. The game stack is always built here.
# FELIS_GITHUB_TOKEN GitHub token; REQUIRED while the repo is private
# FELIS_REF branch/tag/sha — pins the build, overrides the channel, and forces a
# source build (naming a ref asks for that tree, not a published asset)
# FELIS_IMAGE control-plane image ref (default:
# registry.felis.svc:5000/felis/felis:<the felis version>, so each
# release has its own tag and `kubectl rollout undo` returns to the
# previous one; :demo when the version is unknown — never :latest;
# anything not under the registry is used as-is but is NOT
# mirrored into it, so it has no pull source after an image GC)
# FELIS_ROOT_DOMAIN deployment root domain (default: <node-ip>.nip.io)
# FELIS_PANEL_NODEPORT local HTTPS panel/API NodePort (default: 30443)
# FELIS_EGRESS_MODE loadbalancer|nodeport (default: nodeport — no MetalLB on a demo box)
# FELIS_BACKUP_PVC world-archive PVC the installer renders and felis-api hands to its
# backup/restore Jobs (default: felis-backups; empty string disables
# backups — the endpoints answer 503)
# FELIS_ARCHIVE_LOCAL_PATH path that PVC is mounted at inside those Jobs; written into
# felis.toml [archive] local_path (default: /var/lib/felis/archives)
# FELIS_WORLDS_HOST_PATH node directory holding the world volumes (on the k3s this
# installer provisions: /var/lib/rancher/k3s/storage). Setting it
# lets the daily reaper also archive and then delete worlds idle
# for 15 days (without it the reaper only deletes backups past
# their expiry and leaves every world); the reaper reads that
# root as root, so it keeps k3s's own 0700 root:root
# (default: unset = worlds are never reaped)
# FELIS_REGISTRY_STORAGE / FELIS_UPLOADS_STORAGE / FELIS_BACKUP_STORAGE capacity the
# registry, uploads and world-archive PVCs request on first install
# (defaults: 10Gi, 5Gi, 10Gi). An existing claim keeps its size; on
# k3s local-path the number is not enforced, see troubleshooting §9
# PKG_LOCK_TIMEOUT seconds to wait for package-manager locks (default: 900)
# APT_LOCK_TIMEOUT legacy alias for PKG_LOCK_TIMEOUT
set -Eeuo pipefail
# ---------------------------------------------------------------------------
# Configuration & constants
# ---------------------------------------------------------------------------
FELIS_REPO_URL="${FELIS_REPO_URL:-https://github.com/FelisMC/Felis.git}"
# Which version to install. "release" builds the newest published GitHub release;
# "dev" builds the tip of main. Release is the default because an installer that
# tracks a moving branch by default hands every new host a different, untested
# commit — the version an operator reports in a bug is then meaningless.
FELIS_VERSION_BOOTSTRAP="${FELIS_VERSION_BOOTSTRAP:-release}"
# Empty by default and resolved from the channel below. Setting it explicitly pins the
# build to that ref and skips resolution entirely: naming a ref IS asking for a
# development build, so it takes the dev-FORM stamp regardless of the channel. Skipping
# resolution also skips the tag lookup, so the base stays v0.0.0 and the stamp is
# v0.0.0+g<sha> rather than <latest-tag>+g<sha> — deliberate, so pinning a ref costs no
# network call the operator did not ask for.
FELIS_REF="${FELIS_REF:-}"
# Recorded HERE because resolve_install_ref overwrites FELIS_REF on both channels — after it
# runs, "did the operator pin a ref?" is unanswerable. Naming a ref asks for THAT tree to be
# built, so it takes the source path even on the release channel.
FELIS_REF_PINNED=""
if [ -n "$FELIS_REF" ]; then FELIS_REF_PINNED=1; fi
# Set once a prebuilt felis binary is installed at HOST_BIN, by either the TUI hand-off or a
# release download. It is what the image build, the CRD apply and the game stack key off:
# all three only need "is there a binary and no checkout", never "which route got us here".
HAVE_PREBUILT_BINARY=""
# The image the control plane ran before this run moved it (deploy_bundle), for the rollback
# hint in summary.
PREVIOUS_FELIS_IMAGE=""
# Optional GitHub credential, needed while this repository is private: GitHub answers
# 404 (not 403) for a repo the caller cannot see, so without it both the release lookup
# and the clone fail as "not found". Exported because git's credential helper below runs
# as a child process and reads it from the environment — which is also why it is never
# interpolated into the clone URL. A token in the URL is written verbatim into
# .git/config and survives the install; an environment variable does not.
FELIS_GITHUB_TOKEN="${FELIS_GITHUB_TOKEN:-}"
export FELIS_GITHUB_TOKEN
# Set by resolve_install_ref/stamp_version and linked into the binary as main.version.
FELIS_VERSION=""
FELIS_VERSION_BASE=""
# Where the installer parks every image it builds, so kubelet can re-pull one the
# image GC has collected (the disk-pressure drill's dead end: ImagePullBackOff with
# nothing to pull from). The node pulls through the loopback hostPort the registry
# Deployment binds (configure_registry_mirror below); pushes go through
# REGISTRY_PUSH_HOST — docker treats 127.0.0.1 as insecure by default, so the
# daemon needs no insecure-registries entry for it.
REGISTRY_URL="registry.felis.svc:5000"
REGISTRY_PUSH_HOST="127.0.0.1:${REGISTRY_URL##*:}"
# Empty unless the operator names one: resolve_felis_image derives the tag from the version
# this run installs, which is only known once the binary or the checkout is.
FELIS_IMAGE="${FELIS_IMAGE:-}"
FELIS_EGRESS_MODE="${FELIS_EGRESS_MODE:-nodeport}"
FELIS_PANEL_NODEPORT="${FELIS_PANEL_NODEPORT:-30443}"
# World-archive storage. The installer renders this PVC (minecraft namespace) and felis-api
# advertises it to its backup/restore Jobs; emptying it disables backups (503). The archive
# path is written into felis.toml so the Jobs' mount and [archive] local_path agree by
# construction — a mismatch would leave tarLocal's absolute archive refs unresolvable.
FELIS_BACKUP_PVC="${FELIS_BACKUP_PVC:-felis-backups}"
FELIS_ARCHIVE_LOCAL_PATH="${FELIS_ARCHIVE_LOCAL_PATH:-/var/lib/felis/archives}"
# PVC capacities; empty keeps `felis manifests`' defaults.
FELIS_REGISTRY_STORAGE="${FELIS_REGISTRY_STORAGE:-}"
FELIS_UPLOADS_STORAGE="${FELIS_UPLOADS_STORAGE:-}"
FELIS_BACKUP_STORAGE="${FELIS_BACKUP_STORAGE:-}"
# Retention is opt-in because it DELETES worlds (after a verified archive): point this at the
# node directory the world volumes live under. On the k3s this installer provisions that is
# /var/lib/rancher/k3s/storage — the reaper resolves each PVC's local-path directory exactly
# from its volumeName. Left unset, the reaper CronJob renders retention-only: backups past
# their expiry are still deleted daily, and no world is ever archived or deleted.
FELIS_WORLDS_HOST_PATH="${FELIS_WORLDS_HOST_PATH:-}"
# k3s's local-path provisioner root. It appears with the first volume the provisioner
# creates, which on a fresh install is after the reaper's PV has been applied.
K3S_STORAGE_ROOT="/var/lib/rancher/k3s/storage"
# Control-plane database backups (felis db backup): a daily timer bundles pg_dump with the
# /etc/felis state a rebuild needs, and every upgrade that has migrations to apply snapshots
# the database first (felis migrate up). The directory sits outside /var/lib/rancher on
# purpose: reinstalling k3s must not take the database backups with it. Copy it off the
# host for anything beyond "undo a bad upgrade or a mistaken delete" (troubleshooting §16).
FELIS_DB_BACKUP_DIR="${FELIS_DB_BACKUP_DIR:-/var/lib/felis/db-backups}"
FELIS_DB_BACKUP_KEEP="${FELIS_DB_BACKUP_KEEP:-14}"
FELIS_DB_BACKUP_TIME="${FELIS_DB_BACKUP_TIME:-*-*-* 03:30:00}"
# node-exporter textfile collector target; FelisDBBackupStale (deploy/alerts) reads it.
FELIS_DB_BACKUP_METRICS="${FELIS_DB_BACKUP_METRICS:-/var/lib/node_exporter/textfile_collector/felis_db_backup.prom}"
# 0 migrates without the pre-migration snapshot, e.g. against an external database newer
# than this host's pg_dump. The upgrade stops if the snapshot fails and this is not set.
FELIS_PRE_MIGRATE_BACKUP="${FELIS_PRE_MIGRATE_BACKUP:-1}"
# The off-site copy (felis offsite, troubleshooting §16). Every hour the host encrypts each
# world archive and the newest database bundles and copies them to an S3-compatible bucket,
# and the reaper deletes an idle world only once its archive is there. Without a bucket
# every backup lives on this one machine, and losing its disk loses them all. Set the
# bucket and its endpoint to turn it on; FELIS_OFFSITE_ACCESS_KEY / FELIS_OFFSITE_SECRET_KEY
# (and optionally a FELIS_OFFSITE_KEY from `felis offsite keygen`) go into
# /etc/felis/offsite.env, mode 0600, with the encryption key generated when there is none.
# A re-run without these keeps the [offsite] felis.host.toml already has.
FELIS_OFFSITE_ENDPOINT="${FELIS_OFFSITE_ENDPOINT:-}"
FELIS_OFFSITE_BUCKET="${FELIS_OFFSITE_BUCKET:-}"
FELIS_OFFSITE_REGION="${FELIS_OFFSITE_REGION:-}"
FELIS_OFFSITE_PREFIX="${FELIS_OFFSITE_PREFIX:-}"
FELIS_OFFSITE_DB_KEEP="${FELIS_OFFSITE_DB_KEEP:-}"
INSTALL_MODE="${FELIS_INSTALL_MODE:-}"
# Loopback by default: hasJoined is an unauthenticated endpoint by protocol (Velocity
# sends no token), so a public bind is a free auth relay — anyone can point their own
# proxy at it and spend YOUR egress IP on Mojang, until Mojang rate-limits you and your
# own players stop getting in. Same-host Velocity reaches 127.0.0.1 fine; a proxy on
# another machine must opt in explicitly with FELIS_NANO_LISTEN=<private-ip>:8081.
# Left empty here: resolve_nano_listen applies that default only after an existing unit's
# address has had its say.
FELIS_NANO_LISTEN="${FELIS_NANO_LISTEN:-}"
FELIS_NANO_PROXY_CIDR="${FELIS_NANO_PROXY_CIDR:-}"
# Backends that take their forwarded identity through the handshake address instead of
# proxy-wide modern forwarding. See write_velocity_service for why a protocol-47 backend
# needs this. Overridable because adding a second 1.8 backend otherwise means editing this
# script; it is still a restart-time list, not one that follows the CRs.
FELIS_LEGACY_FORWARDING_SERVERS="${FELIS_LEGACY_FORWARDING_SERVERS:-legacy18}"
FELIS_VELOCITY_XMX="${FELIS_VELOCITY_XMX:-1G}"
# The Go tarball is unpacked and run as root, so the default version is pinned by the sha256
# go.dev/dl publishes for each architecture install_go_toolchain handles. Move all three
# together; any other FELIS_GO_VERSION has to bring its own FELIS_GO_SHA256.
GO_PINNED_VERSION="1.26.8"
GO_PINNED_SHA256_AMD64="d0f743b33e8d8945e6b1f432edd15785c70507121d6e2a723b21285eddf8b57b"
GO_PINNED_SHA256_ARM64="211ffced9dcb9633a55eac6364816ec0ddd951389a740e88fa8b3337971bdda0"
FELIS_GO_VERSION="${FELIS_GO_VERSION:-$GO_PINNED_VERSION}"
FELIS_GO_SHA256="${FELIS_GO_SHA256:-}"
# cloudflared runs as root on the edge, so it gets the same treatment: a pinned release and
# the sha256 GitHub lists for each asset. A different FELIS_CLOUDFLARED_VERSION has to bring
# its own FELIS_CLOUDFLARED_SHA256. An installed binary is replaced only under
# FELIS_UPGRADE_DEPS=1; `felis update --cloudflared` reports when that would change it.
CLOUDFLARED_PINNED_VERSION="2026.9.1"
CLOUDFLARED_PINNED_SHA256_AMD64="03f1f25d1cc93b9ad6c60569d44060bc4f17ed97075760ed8cfca4b12dcd68cc"
CLOUDFLARED_PINNED_SHA256_ARM64="3d97437c71848bd8df68041e12436b484a661d95073ea1937f01a845ce88faa3"
CLOUDFLARED_PINNED_SHA256_ARM="093ffa3638ab2b636de63c43a8c68f96a69cf71f9699dd8277a91b160b0f4fc0"
FELIS_CLOUDFLARED_VERSION="${FELIS_CLOUDFLARED_VERSION:-$CLOUDFLARED_PINNED_VERSION}"
FELIS_CLOUDFLARED_SHA256="${FELIS_CLOUDFLARED_SHA256:-}"
CLOUDFLARED_BIN=/usr/local/bin/cloudflared
# The k3s release a fresh install gets, and the tag its install script is read from. The
# script checks the k3s binary against that release's sha256sum file, so pinning the tag
# pins both. An installed k3s moves only under FELIS_UPGRADE_DEPS=1.
FELIS_K3S_VERSION="${FELIS_K3S_VERSION:-v1.36.4+k3s1}"
FELIS_UPGRADE_DEPS="${FELIS_UPGRADE_DEPS:-0}"
# The in-cluster registry's image, by digest. It must equal platform.defaultRegistryImage
# (internal/platform/identities.go, TestBootstrapPinsTheRegistryImage): the renderer puts
# that ref in the Deployment, and this script caches and pins the same ref in containerd.
REGISTRY_IMAGE="docker.io/library/registry:2.8.3@sha256:a3d8aaa63ed8681a604f1dea0aa03f100d5895b6a58ace528858a7b332415373"
PKG_LOCK_TIMEOUT="${PKG_LOCK_TIMEOUT:-${APT_LOCK_TIMEOUT:-900}}"
APT_LOCK_TIMEOUT="${APT_LOCK_TIMEOUT:-$PKG_LOCK_TIMEOUT}"
# --- the game stack: proxy on the host, the two always-on backends in k3s ---
FELIS_LIMBO_IMAGE="${FELIS_LIMBO_IMAGE:-${REGISTRY_URL}/felis/limbo:demo}"
FELIS_LOBBY_IMAGE="${FELIS_LOBBY_IMAGE:-${REGISTRY_URL}/felis/lobby:demo}"
# Plain Paper base recommended for a user's own server (deploy/paper). Not a system
# server — forwarding is applied by the operator's init-forwarding initContainer, so it
# needs no secret. Seeded recommended in 0019_recommended_paper.sql.
FELIS_PAPER_IMAGE="${FELIS_PAPER_IMAGE:-${REGISTRY_URL}/felis/paper:demo}"
# The Velocity build comes from deploy/game-stack.lock (VELOCITY_VERSION and its jar digest).
# Setting FELIS_VELOCITY_VERSION to another version, or FELIS_GAME_STACK=latest, installs
# the newest BUILD of that minor instead. The minor itself is never discovered: PaperMC's
# Fill v3 groups velocity builds by version group, and "newest across all groups" can mean
# an unreleased 4.x SNAPSHOT that needs another runtime. Crossing a major is a deliberate
# code change.
FELIS_VELOCITY_VERSION="${FELIS_VELOCITY_VERSION:-}"
VELOCITY_LATEST_MINOR="3.5.1"
# pinned installs the builds deploy/game-stack.lock names (resolve_game_jars); latest asks
# upstream for its newest ones, which is how that lock file gets refreshed.
FELIS_GAME_STACK="${FELIS_GAME_STACK:-pinned}"
# Path to a Felis-Legacy Velocity fork build, installed as the proxy in place of the
# stock download. Unset — the default — changes nothing.
#
# Stock Velocity will not offer the login-plugin-message exchange below 1.13, so a 1.8
# client reaching a modern-forwarding backend today is a side effect of Via replacing
# the channel initializers before that check runs. It works, and nobody designed it.
# The fork registers the login packets on 1.7.2 and drops the gate, which makes the
# same outcome deliberate.
#
# Opt-in because it is unmeasured where it counts: FL-008's probe runs offline-mode
# against a stub, and this jar would carry every real Mojang session on the server.
FELIS_VELOCITY_FORK_JAR="${FELIS_VELOCITY_FORK_JAR:-}"
# Expected sha256 of that jar, REQUIRED whenever it is set. Case and internal spaces are
# ignored, so whatever sha256sum, Get-FileHash or certutil printed can be pasted as-is.
# No digest is hardcoded here:
# the build lives in Felis-Legacy and has never been reproduced on a second machine, so
# any constant this script carried would pin one machine's output rather than the fork.
#
# So this is not a supply-chain signature and does not pretend to be one — an operator
# who can write the jar can write this value too. What it does buy: a path is not an
# identity, and every re-run of this script re-checks it. A truncated copy, a stale build
# left at the same path, or the two-patch jar where the three-patch one was meant all
# change the digest and stop the install. Naming the digest once is what turns "whatever
# is at that path today" into one specific build.
FELIS_VELOCITY_FORK_JAR_SHA256="${FELIS_VELOCITY_FORK_JAR_SHA256:-}"
# Temurin 25: Velocity 3.5 needs 21+, and 25 is also what a future Velocity 4 requires,
# so the runtime does not have to move again when the pin does. Distro JDK packaging is
# a lottery across four package managers — a tarball is one code path everywhere (same
# reasoning as install_go_toolchain).
FELIS_JRE_VERSION="${FELIS_JRE_VERSION:-25}"
# The JRE the proxy runs on is unpacked as root and carries every player session, so the
# default feature version is pinned to one build and the sha256 Adoptium publishes for each
# architecture. Move the three together (the Adoptium API lists them:
# /v3/assets/latest/25/hotspot?image_type=jre&os=linux).
JRE_PINNED_FEATURE="25"
JRE_PINNED_RELEASE="25.0.4.1+1"
JRE_PINNED_SHA256_X64="1731a34baadec5479258ea0202e4d5d865d2efeee60cb0c7d7eb056fe96ca219"
JRE_PINNED_SHA256_AARCH64="34828cbb93ed31c281c84ecb31ddab655d11a802f263c1fc019d42e9e0230fed"
# The port the proxy listens on: the ONLY Minecraft port players ever touch. Backends
# are ClusterIP-only, verify the modern-forwarding HMAC, and use NetworkPolicy to limit
# non-node ingress to the declared proxy CIDRs.
FELIS_GAME_PORT="${FELIS_GAME_PORT:-25565}"
# Velocity server names — must match internal/naming (SystemLoginServer/SystemLobbyServer);
# felis-api hands the proxy backends under exactly these names.
LOGIN_SERVER="login"
LOBBY_SERVER="lobby"
CONTROL_NS="felis"
MINECRAFT_NS="minecraft"
BUILD_NS="felis-build"
POD_CIDR="10.42.0.0/16" # k3s default cluster CIDR
SERVICE_CIDR="10.43.0.0/16" # k3s default service CIDR
DB_NAME="felis"
DB_USER="felis"
STATE_DIR="/etc/felis"
SECRETS_ENV="${STATE_DIR}/secrets.env"
BOOTSTRAP_DONE="${STATE_DIR}/bootstrap.done"
PANEL_TLS_CERT="${STATE_DIR}/panel-tls.crt"
PANEL_TLS_KEY="${STATE_DIR}/panel-tls.key"
SRC_DIR="/opt/felis/src"
HOST_BIN="/usr/local/bin/felis"
# The binary this run replaced (keep_previous_host_binary), and whether the new one has
# been put to use: once migrations start, or the nano service restarts onto it, the old
# one no longer matches what is running and a failed run keeps the new one.
HOST_BIN_PREV=""
HOST_BIN_KEPT=0
HOST_BIN_IN_USE=0
# Felis's own build toolchain, not /usr/local/go: install_go_toolchain replaces whatever
# version sits here, and an operator's Go at the conventional path is not ours to swap.
GOROOT_DIR="/opt/felis/go"
NANO_SERVICE="/etc/systemd/system/felis-nano.service"
DB_BACKUP_SERVICE="/etc/systemd/system/felis-db-backup.service"
DB_BACKUP_TIMER="/etc/systemd/system/felis-db-backup.timer"
WATCHDOG_SERVICE="/etc/systemd/system/felis-watchdog.service"
WATCHDOG_TIMER="/etc/systemd/system/felis-watchdog.timer"
WATCHDOG_STATE="/var/lib/felis/watchdog/state.json"
OFFSITE_ENV="${STATE_DIR}/offsite.env"
OFFSITE_SERVICE="/etc/systemd/system/felis-offsite.service"
OFFSITE_TIMER="/etc/systemd/system/felis-offsite.timer"
BUILD_TOOLS_SERVICE="/etc/systemd/system/felis-build-tools.service"
BUILD_TOOLS_TIMER="/etc/systemd/system/felis-build-tools.timer"
BUILD_TOOLS_STATUS="/var/lib/felis/build-tools/status.json"
# While this marker holds a future Unix time, felis watchdog mails nothing: an install
# restarts the control plane and the system servers on purpose. cleanup removes it; the
# time in it is the backstop for an installer killed before its EXIT trap runs.
WATCHDOG_QUIET_FILE="/run/felis/watchdog-quiet-until"
VELOCITY_DIR="/opt/felis/velocity"
VELOCITY_USER="felis-velocity"
VELOCITY_SERVICE="/etc/systemd/system/felis-velocity.service"
# What the running proxy was last started from (velocity_fingerprint), and the builds the
# login and lobby pods were last started on: a rerun restarts only what actually changed,
# since each of those restarts disconnects every player online.
VELOCITY_FINGERPRINT="${STATE_DIR}/velocity.fingerprint"
SYSTEM_SERVER_IMAGES="${STATE_DIR}/system-server-images"
# The nftables rules that keep PostgreSQL's port to this host and its pods, on a host
# without firewalld (configure_postgres_firewall).
PG_FIREWALL_RULES="${STATE_DIR}/postgres-firewall.nft"
PG_FIREWALL_SERVICE="/etc/systemd/system/felis-postgres-firewall.service"
JRE_DIR="/opt/felis/jre"
K3S_BIN_DIR="${K3S_BIN_DIR:-/usr/local/bin}"
K3S_BIN="${K3S_BIN_DIR}/k3s"
# k3s's containerd mirror config, written by configure_registry_mirror. A variable
# (not just the literal path) so bootstrap_test.sh can point the writer at a
# scratch file.
K3S_REGISTRIES_FILE="/etc/rancher/k3s/registries.yaml"
APT_LOCK_FILES=(
/var/lib/dpkg/lock-frontend
/var/lib/dpkg/lock
/var/cache/apt/archives/lock
/var/lib/apt/lists/lock
)
APT_BACKGROUND_TIMERS=(
apt-daily.timer
apt-daily-upgrade.timer
)
APT_BACKGROUND_SERVICES=(
apt-daily.service
apt-daily-upgrade.service
unattended-upgrades.service
)
DNF_BACKGROUND_TIMERS=(
dnf-makecache.timer
dnf-automatic.timer
)
DNF_BACKGROUND_SERVICES=(
dnf-makecache.service
dnf-automatic.service
)
YUM_BACKGROUND_TIMERS=(
yum-cron.timer
)
YUM_BACKGROUND_SERVICES=(
yum-cron.service
)
ZYPPER_BACKGROUND_TIMERS=(
packagekit-background.timer
)
ZYPPER_BACKGROUND_SERVICES=(
packagekit.service
)
# ---------------------------------------------------------------------------
# Clean PATH (sudo may strip /usr/local/bin)
# ---------------------------------------------------------------------------
PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
export PATH
# BuildKit attaches a provenance attestation to every build by default, and it records the
# build's start time. On Docker's containerd image store the image id is the digest of the
# index that carries it, so an unchanged rebuild would get a new id each run: the login and
# lobby pods would restart on every rerun, and each run would push another versioned tag.
# Without it the id is the manifest digest, which an all-cached build reproduces.
export BUILDX_NO_DEFAULT_ATTESTATIONS=1
# ---------------------------------------------------------------------------
# Logging
# ---------------------------------------------------------------------------
log() { printf '\033[1;36m[felis]\033[0m %s\n' "$*"; }
ok() { printf '\033[1;32m[ ok ]\033[0m %s\n' "$*"; }
warn() { printf '\033[1;33m[warn]\033[0m %s\n' "$*" >&2; }
die() { printf '\033[1;31m[fail]\033[0m %s\n' "$*" >&2; exit 1; }
TEMP_PATHS=()
DOCKER_CONTAINERS=()
REGISTRY_DOCKER_CONFIG=""
PKG_TIMERS_TO_RESTORE=()
on_error() {
local line="$1" code="$2"
warn "bootstrap failed near line ${line} (exit ${code})"
}
cleanup() {
local status=$? id path unit
restore_previous_host_binary "$status"
for unit in "${PKG_TIMERS_TO_RESTORE[@]-}"; do
[ -n "$unit" ] || continue
systemctl start "$unit" >/dev/null 2>&1 || true
done
if command -v docker >/dev/null 2>&1; then
for id in "${DOCKER_CONTAINERS[@]-}"; do
[ -n "$id" ] && docker rm "$id" >/dev/null 2>&1 || true
done
fi
for path in "${TEMP_PATHS[@]-}"; do
[ -n "$path" ] && rm -rf -- "$path" || true
done
# A failed install leaves something broken the owners should hear about, so the
# watchdog speaks again the moment the installer exits, however it exits.
rm -f -- "$WATCHDOG_QUIET_FILE" 2>/dev/null || true
}
# keep_previous_host_binary copies the felis binary this run is about to replace, once. A
# run that fails before the new binary is in use puts it back (restore_previous_host_binary):
# until then the old binary, the old cluster and the unmigrated database still agree, while
# a new binary left behind runs the host timers against a schema it was not built for, and
# `felis setup` with it would migrate the database under the old control plane.
keep_previous_host_binary() {
[ "$HOST_BIN_KEPT" = 0 ] || return 0
HOST_BIN_KEPT=1
[ -x "$HOST_BIN" ] || return 0
HOST_BIN_PREV="${HOST_BIN}.prev"
rm -f "$HOST_BIN_PREV"
cp "$HOST_BIN" "$HOST_BIN_PREV"
}
restore_previous_host_binary() { # exit-status
[ -n "$HOST_BIN_PREV" ] && [ -f "$HOST_BIN_PREV" ] || return 0
if [ "$1" -ne 0 ] && [ "$HOST_BIN_IN_USE" != 1 ]; then
# install(1) onto a fresh file, as everywhere else HOST_BIN is written (SELinux label).
rm -f "$HOST_BIN"
if install -m 0755 "$HOST_BIN_PREV" "$HOST_BIN"; then
command -v restorecon >/dev/null 2>&1 && restorecon "$HOST_BIN" >/dev/null 2>&1 || true
warn "restored the previous felis binary at ${HOST_BIN}; the database was not migrated, so rerunning the installer picks up where this run stopped"
else
warn "could not restore the previous felis binary; it is at ${HOST_BIN_PREV}"
return 0
fi
fi
rm -f -- "$HOST_BIN_PREV"
}
remember_temp() { TEMP_PATHS+=("$1"); }
remember_container() { DOCKER_CONTAINERS+=("$1"); }
trap 'on_error "$LINENO" "$?"' ERR
trap cleanup EXIT
k3s_cmd() { [ -x "$K3S_BIN" ] || die "k3s binary not found at ${K3S_BIN}"; "$K3S_BIN" "$@"; }
kube() { k3s_cmd kubectl "$@"; }
# Create or update a single-key Secret without putting the value in kubectl's argv.
# The temporary file is mode 0600 and is also registered with the EXIT cleanup path.
apply_literal_secret() {
local namespace="$1" name="$2" key="$3" value="$4" tmp
tmp="$(umask 077; mktemp)"
remember_temp "$tmp"
printf '%s' "$value" > "$tmp"
kube -n "$namespace" create secret generic "$name" \
--from-file="${key}=${tmp}" \
--dry-run=client -o yaml | kube apply -f -
rm -f "$tmp"
}
# The control plane mounts felis-config from its own namespace; the workload
# namespace's backup/restore/fileedit Jobs and the reaper mount a local copy (a
# secretKeyRef is namespace-local). The installer owns the rendered config, so both
# copies are (re)applied on every run — unlike the create-if-absent credential
# replicas `felis setup` makes, because a stale config copy keeps an old database URL
# or archive policy after an upgrade or a credential rotation.
apply_felis_config_secrets() {
kube -n "$CONTROL_NS" create secret generic felis-config \
--from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
--dry-run=client -o yaml | kube apply -f -
kube -n "$MINECRAFT_NS" create secret generic felis-config \
--from-file=felis.toml="${STATE_DIR}/felis.pod.toml" \
--dry-run=client -o yaml | kube apply -f -
}
# The registry gate reads one token file per principal from felis-registry-auth
# (registry namespace = control namespace); a build Job's push container reads the
# build principal's username/password from felis-registry-push in the build
# namespace. Values go through 0600 temp files, never kubectl's argv.
apply_registry_secrets() {
local dir
dir="$(umask 077; mktemp -d)"
remember_temp "$dir"
printf '%s' "$REGISTRY_PLATFORM_TOKEN" > "${dir}/platform"
printf '%s' "$REGISTRY_BUILD_TOKEN" > "${dir}/build"
printf '%s' "$REGISTRY_PRUNE_TOKEN" > "${dir}/prune"
printf '%s' build > "${dir}/username"
kube -n "$CONTROL_NS" create secret generic felis-registry-auth \
--from-file=platform="${dir}/platform" \
--from-file=build="${dir}/build" \
--from-file=prune="${dir}/prune" \
--dry-run=client -o yaml | kube apply -f -
kube -n "$BUILD_NS" create secret generic felis-registry-push \
--from-file=username="${dir}/username" \
--from-file=password="${dir}/build" \
--dry-run=client -o yaml | kube apply -f -
rm -rf -- "$dir"
}
# node_global_cidrs prints one host-length CIDR per global address on this node.
# Game server egress already excludes every private range; this adds the node's
# public addresses, which would otherwise let a server dial the panel NodePort,
# SSH, or anything else the host serves on them.
node_global_cidrs() {
command -v ip >/dev/null 2>&1 || return 0
ip -o addr show scope global 2>/dev/null | awk '
$3 == "inet" { split($4, a, "/"); print a[1] "/32" }
$3 == "inet6" { split($4, a, "/"); print a[1] "/128" }
' | sort -u
}
as_postgres() {
if command -v runuser >/dev/null 2>&1; then
runuser -u postgres -- "$@"
else
sudo -u postgres "$@"
fi
}
bootstrap_from_tui() {
[ "${FELIS_BOOTSTRAP_FROM_TUI:-}" = "1" ]
}
pause_package_background_timers() {
command -v systemctl >/dev/null 2>&1 || return 0
local active=0 timers=() services=() unit
case "${PKG:-}" in
apt) timers=("${APT_BACKGROUND_TIMERS[@]}"); services=("${APT_BACKGROUND_SERVICES[@]}") ;;
dnf) timers=("${DNF_BACKGROUND_TIMERS[@]}"); services=("${DNF_BACKGROUND_SERVICES[@]}") ;;
yum) timers=("${YUM_BACKGROUND_TIMERS[@]}"); services=("${YUM_BACKGROUND_SERVICES[@]}") ;;
zypper) timers=("${ZYPPER_BACKGROUND_TIMERS[@]}"); services=("${ZYPPER_BACKGROUND_SERVICES[@]}") ;;
*) return 0 ;;
esac
for unit in "${timers[@]}"; do
if systemctl is-active --quiet "$unit"; then
PKG_TIMERS_TO_RESTORE+=("$unit")
active=1
fi
done
if [ "$active" -eq 1 ]; then
log "pausing package-manager timers during bootstrap: ${PKG_TIMERS_TO_RESTORE[*]}"
systemctl stop "${PKG_TIMERS_TO_RESTORE[@]}" || warn "could not stop package-manager timers; package operations may need to wait"
fi
for unit in "${services[@]}"; do
if systemctl is-active --quiet "$unit"; then
log "stopping package-manager background service during bootstrap: ${unit}"
systemctl stop "$unit" || warn "could not stop ${unit}; package operations may need to wait"
fi
done
}
pkg_lock_files() {
case "${PKG:-}" in
apt) printf '%s\n' "${APT_LOCK_FILES[@]}" ;;
dnf|yum)
printf '%s\n' \
/var/lib/rpm/.rpm.lock \
/var/lib/dnf/rpmdb_lock.pid \
/var/cache/dnf/metadata_lock.pid \
/run/dnf.pid \
/var/run/dnf.pid
;;
zypper)
printf '%s\n' \
/var/lib/rpm/.rpm.lock \
/run/zypp.pid \
/var/run/zypp.pid
;;
pacman) printf '%s\n' /var/lib/pacman/db.lck ;;
esac
}
pkg_lock_process_names() {
case "${PKG:-}" in
dnf) printf '%s\n' dnf dnf5 rpm ;;
yum) printf '%s\n' yum rpm ;;
zypper) printf '%s\n' zypper rpm ;;
pacman) printf '%s\n' pacman ;;
esac
}
pkg_busy_pids() {
local file file_count name
{
if command -v fuser >/dev/null 2>&1; then
local files=()
file_count=0
while IFS= read -r file; do
if [ -e "$file" ]; then
files+=("$file")
file_count=$((file_count + 1))
fi
done < <(pkg_lock_files)
[ "$file_count" -eq 0 ] || fuser "${files[@]}" 2>/dev/null | tr ' ' '\n'
fi
if command -v pgrep >/dev/null 2>&1; then
while IFS= read -r name; do
[ -n "$name" ] && pgrep -x "$name" 2>/dev/null || true
done < <(pkg_lock_process_names)
fi
} | awk 'NF && !seen[$1]++'
}
pkg_lock_busy() {
[ -n "$(pkg_busy_pids)" ]
}
pkg_lock_holders() {
local pids
pids="$(pkg_busy_pids | paste -sd, - || true)"
[ -n "$pids" ] || return 0
ps -o pid=,comm= -p "$pids" 2>/dev/null | awk '{$1=$1; print}' | paste -sd ';' -
}
wait_for_pkg_locks() {
local deadline holders next_notice
deadline=$((SECONDS + PKG_LOCK_TIMEOUT))
next_notice=0
while pkg_lock_busy; do
if [ "$SECONDS" -ge "$next_notice" ]; then
holders="$(pkg_lock_holders)"
if [ -n "$holders" ]; then
log "waiting for ${PKG} package locks to clear (timeout ${PKG_LOCK_TIMEOUT}s; holders: ${holders})"
else
log "waiting for ${PKG} package locks to clear (timeout ${PKG_LOCK_TIMEOUT}s)"
fi
next_notice=$((SECONDS + 30))
fi
[ "$SECONDS" -lt "$deadline" ] || die "${PKG} package manager is still busy after ${PKG_LOCK_TIMEOUT}s; wait for the current package operation to finish, then retry"
sleep 5
done
}
# NEEDRESTART_SUSPEND keeps Ubuntu's needrestart hook from restarting services after
# the install's own apt runs. It would restart felis-velocity whenever a rerun
# happens to install or upgrade a package (the running JVM maps files the rerun
# replaces), dropping every player on a rerun that changed nothing the proxy runs.
# The installer restarts what it changes itself.
apt_get() {
wait_for_pkg_locks
NEEDRESTART_SUSPEND=1 DEBIAN_FRONTEND=noninteractive apt-get \
-o DPkg::Lock::Timeout="$PKG_LOCK_TIMEOUT" \
"$@"
}
validate_timeout() {
local name="$1" value="$2"
case "$value" in
''|*[!0-9]*) die "${name} must be a non-negative integer (seconds), got: ${value}" ;;
esac
}
validate_nodeport() {
local name="$1" value="$2"
case "$value" in
''|*[!0-9]*) die "${name} must be a Kubernetes NodePort integer, got: ${value}" ;;
esac
if [ "$value" -lt 30000 ] || [ "$value" -gt 32767 ]; then
die "${name} must be in Kubernetes NodePort range 30000-32767, got: ${value}"
fi
}
# A value without a usable port would reach the firewall and the summary as-is: 8081 opens
# port 8081 while nano binds nothing, 127.0.0.1 prints http://127.0.0.1:127.0.0.1/..., and
# the unit crash-loops either way.
validate_listen() {
local name="$1" value="$2" port="${2##*:}" host="${2%:*}"
case "$value" in *:*) ;; *) port="" ;; esac
case "$port" in
''|*[!0-9]*) die "${name} must be host:port (for example 127.0.0.1:8081), got: ${value}" ;;
esac
[ "$port" -ge 1 ] && [ "$port" -le 65535 ] || die "${name} port must be 1-65535, got: ${value}"
# Go takes a colon in the host only inside brackets; ::1:8081 would crash-loop the unit.
case "$host" in
*:*) case "$host" in "["*"]") ;; *) die "${name} needs an IPv6 host in brackets (for example [::1]:8081), got: ${value}" ;; esac ;;
esac
}
# The value lands inside a firewalld rich rule, so anything but address characters and one
# prefix length is refused here rather than handed to firewall-cmd.
validate_cidr() {
case "$2" in
"") return 0 ;;
*[!0-9A-Fa-f.:/]*|*/*/*|*/|/*) ;;
*/[0-9]*) return 0 ;;
esac
die "$1 must be an address with a prefix length (for example 10.0.0.7/32 or fd00::7/128), got: $2"
}
validate_settings() {
validate_timeout PKG_LOCK_TIMEOUT "$PKG_LOCK_TIMEOUT"
validate_timeout APT_LOCK_TIMEOUT "$APT_LOCK_TIMEOUT"
validate_nodeport FELIS_PANEL_NODEPORT "$FELIS_PANEL_NODEPORT"
validate_listen FELIS_NANO_LISTEN "$FELIS_NANO_LISTEN"
validate_cidr FELIS_NANO_PROXY_CIDR "$FELIS_NANO_PROXY_CIDR"
validate_offsite_settings
case "$FELIS_GAME_STACK" in
pinned|latest) ;;
*) die "FELIS_GAME_STACK must be pinned or latest (got '${FELIS_GAME_STACK}')" ;;
esac
[ "$(heap_megabytes "$FELIS_VELOCITY_XMX")" -ge 256 ] \
|| die "FELIS_VELOCITY_XMX must be a heap size of at least 256M, written <n>M or <n>G (got '${FELIS_VELOCITY_XMX}')"
case "$FELIS_UPGRADE_DEPS" in
0|1) ;;
*) die "FELIS_UPGRADE_DEPS must be 0 or 1 (got '${FELIS_UPGRADE_DEPS}')" ;;
esac
}
# version_newer reports whether version $1 sorts after $2 (a leading v is ignored).
version_newer() {
local a="${1#v}" b="${2#v}"
[ "$a" != "$b" ] && [ "$(printf '%s\n%s\n' "$a" "$b" | sort -V | tail -n 1)" = "$a" ]
}
# heap_megabytes prints a JVM heap size written <n>M or <n>G in megabytes, or 0 for any
# other spelling.
heap_megabytes() {
local n="${1%?}"
case "$n" in
""|*[!0-9]*|0*) echo 0; return ;;
esac
# Six digits of gigabytes is far past any host; the cap keeps the arithmetic in range.
[ "${#n}" -le 6 ] || { echo 0; return; }
case "$1" in
*[Mm]) echo "$n" ;;
*[Gg]) echo "$((n * 1024))" ;;
*) echo 0 ;;
esac
}
# validate_offsite_settings checks the FELIS_OFFSITE_* inputs before anything is
# installed. They are written into felis.toml as TOML strings and the secrets into a
# single-quoted env file, so a quote, a backslash or a line break is refused outright.
validate_offsite_settings() {
local name value
for name in FELIS_OFFSITE_ENDPOINT FELIS_OFFSITE_BUCKET FELIS_OFFSITE_REGION FELIS_OFFSITE_PREFIX \
FELIS_OFFSITE_ACCESS_KEY FELIS_OFFSITE_SECRET_KEY FELIS_OFFSITE_KEY; do
value="${!name:-}"
case "$value" in
*\"* | *\'* | *\\* | *$'\n'* | *$'\r'*) die "${name} must not contain quotes, backslashes or line breaks" ;;
esac
done
if [ -z "$FELIS_OFFSITE_BUCKET" ]; then
if [ -n "$FELIS_OFFSITE_ENDPOINT$FELIS_OFFSITE_REGION$FELIS_OFFSITE_PREFIX$FELIS_OFFSITE_DB_KEEP" ]; then
die "FELIS_OFFSITE_* is set without FELIS_OFFSITE_BUCKET; name the bucket too"
fi
return 0
fi
[ -n "$FELIS_OFFSITE_ENDPOINT" ] || die "FELIS_OFFSITE_BUCKET needs FELIS_OFFSITE_ENDPOINT (https://host[:port] of the S3-compatible store)"
case "$FELIS_OFFSITE_BUCKET" in
*/* | *' '*) die "FELIS_OFFSITE_BUCKET is a bucket name; put a path inside it in FELIS_OFFSITE_PREFIX" ;;
esac
if [ -n "$FELIS_OFFSITE_DB_KEEP" ] && ! [[ "$FELIS_OFFSITE_DB_KEEP" =~ ^[1-9][0-9]*$ ]]; then
die "FELIS_OFFSITE_DB_KEEP must be a positive number of bundles, got: ${FELIS_OFFSITE_DB_KEEP}"
fi
}
# ---------------------------------------------------------------------------
# 0. Privilege & host facts
# ---------------------------------------------------------------------------
if [ "$(id -u)" -ne 0 ]; then
if [ -r "$0" ]; then
log "re-executing under sudo"
exec sudo -E bash "$0" "$@"
fi
die "must run as root (for a piped installer, use: curl -fsSL <url> | sudo bash)"
fi
detect_os() {
[ -r /etc/os-release ] || die "cannot read /etc/os-release; unsupported host"
# shellcheck disable=SC1091
. /etc/os-release
OS_ID="${ID:-unknown}"
OS_VERSION="${VERSION_ID:-unknown}"
OS_ID_LIKE="${ID_LIKE:-}"
OS_CODENAME="${VERSION_CODENAME:-${UBUNTU_CODENAME:-}}"
if command -v apt-get >/dev/null 2>&1; then
PKG="apt"
elif command -v dnf >/dev/null 2>&1; then
PKG="dnf"
elif command -v yum >/dev/null 2>&1; then
PKG="yum"
elif command -v zypper >/dev/null 2>&1; then
PKG="zypper"
elif command -v pacman >/dev/null 2>&1; then
PKG="pacman"
else
die "no supported package manager (apt/dnf/yum/zypper/pacman) found on ${OS_ID} ${OS_VERSION}"
fi
log "host: ${PRETTY_NAME:-$OS_ID $OS_VERSION} (package manager: ${PKG})"
}
# persisted_root_domain echoes the root_domain an earlier run wrote, or nothing. The
# generated toml is the only durable record of it: nothing else on the host stores the
# domain, and it is written on every successful install.
persisted_root_domain() {
local f="${STATE_DIR}/felis.host.toml"
[ -r "$f" ] || return 0
awk -F'"' '/^[[:space:]]*root_domain[[:space:]]*=/ { print $2; exit }' "$f"
}
detect_node_ip() {
NODE_IP="$(ip -4 route get 1.1.1.1 2>/dev/null | awk '{for(i=1;i<=NF;i++) if($i=="src"){print $(i+1); exit}}')"
[ -n "${NODE_IP:-}" ] || NODE_IP="$(hostname -I 2>/dev/null | awk '{print $1}')"
[ -n "${NODE_IP:-}" ] || die "could not determine this host's primary IPv4 address"
# Precedence: an explicit FELIS_ROOT_DOMAIN, then whatever the last run persisted, then
# the nip.io default. The middle step is what makes a re-run idempotent. Without it this
# installer re-derived the domain from scratch every time and defaulted to nip.io, so
# re-running it on a live install -- the only way to move felis-api to a newer release,
# and what `felis update` points operators at -- rewrote root_domain, panel_hostname and
# admin_hostname to nip.io names. ensure_panel_tls_cert is write-once and kept serving a
# certificate for the OLD hostnames, so the console stopped matching its own cert, with
# no re-domain flow to recover through. Secrets never had this problem:
# load_or_make_secrets has always sourced secrets.env before generating anything.
local persisted
persisted="$(persisted_root_domain)"
if [ -n "${FELIS_ROOT_DOMAIN:-}" ] && [ -n "$persisted" ] && [ "$FELIS_ROOT_DOMAIN" != "$persisted" ]; then
# Deliberate re-domain. Allowed -- there is no other route to it -- but it is not a
# thing this script finishes: the panel certificate, the two secrets, the velocity
# config and the login CR all still carry the old name.
warn "FELIS_ROOT_DOMAIN (${FELIS_ROOT_DOMAIN}) differs from the installed ${persisted}."
warn "This re-domains the install. The write-once panel certificate is NOT reissued and"
warn "will keep the old hostnames; the proxy and login config need the same treatment."
fi
FELIS_ROOT_DOMAIN="${FELIS_ROOT_DOMAIN:-${persisted:-${NODE_IP}.nip.io}}"
if [ -n "$persisted" ] && [ "$FELIS_ROOT_DOMAIN" = "$persisted" ]; then
log "node IP: ${NODE_IP} root domain: ${FELIS_ROOT_DOMAIN} (reusing the installed domain)"
else
log "node IP: ${NODE_IP} root domain: ${FELIS_ROOT_DOMAIN}"
fi
}
pkg_install() {
case "$PKG" in
apt) apt_get install -y "$@" ;;
dnf) wait_for_pkg_locks; dnf install -y "$@" ;;
yum) wait_for_pkg_locks; yum install -y "$@" ;;
zypper) wait_for_pkg_locks; zypper --non-interactive install -y "$@" ;;
pacman) wait_for_pkg_locks; pacman -S --noconfirm --needed "$@" ;;
esac
}
pkg_refresh_once() {
[ -n "${_PKG_REFRESHED:-}" ] && return 0
case "$PKG" in
apt) apt_get update -y ;;
dnf|yum) : ;; # dnf/yum refresh metadata on demand
zypper) wait_for_pkg_locks; zypper --non-interactive refresh ;;
# Arch supports only whole-system upgrades (-Sy alone leaves a partial upgrade), so the
# refresh stays -Syu, with PostgreSQL held back once a cluster exists: a new major
# version cannot open the old data directory, and the upgrade would take the platform's
# database down on an unrelated rerun. check_postgres_major explains the way forward.
pacman)
wait_for_pkg_locks
if [ -f "$(postgres_data_dir)/PG_VERSION" ]; then
pacman -Syu --noconfirm --ignore postgresql
else
pacman -Syu --noconfirm
fi
;;
esac
_PKG_REFRESHED=1
}
# ---------------------------------------------------------------------------
# 1. Swap — k3s + Postgres + a Go build will OOM on a <2 GiB box without it
# ---------------------------------------------------------------------------
ensure_swap() {
local mem_kb swap_kb
mem_kb="$(awk '/^MemTotal:/{print $2}' /proc/meminfo)"
swap_kb="$(awk '/^SwapTotal:/{print $2}' /proc/meminfo)"
if [ "${swap_kb:-0}" -gt 0 ]; then
ok "swap already present ($((swap_kb/1024)) MiB)"
return 0
fi
if [ "${mem_kb:-0}" -ge 2097152 ]; then
ok "RAM $((mem_kb/1024)) MiB is sufficient; skipping swap"
return 0
fi
log "low RAM ($((mem_kb/1024)) MiB) and no swap — creating a 2 GiB swap file"
if ! fallocate -l 2G /swapfile 2>/dev/null; then
dd if=/dev/zero of=/swapfile bs=1M count=2048 status=none
fi
chmod 600 /swapfile
mkswap /swapfile >/dev/null
swapon /swapfile
grep -q '^/swapfile ' /etc/fstab || echo '/swapfile none swap sw 0 0' >> /etc/fstab
ok "2 GiB swap active"
}
# ---------------------------------------------------------------------------
# 2. Base packages
# ---------------------------------------------------------------------------
install_base() {
local packages=(ca-certificates openssl)
pkg_refresh_once
command -v curl >/dev/null 2>&1 || packages+=(curl)
command -v tar >/dev/null 2>&1 || packages+=(tar)
if ! bootstrap_from_tui && ! command -v git >/dev/null 2>&1; then
packages+=(git)
fi
pkg_install "${packages[@]}"
ok "base tools present"
}
install_cloudflared() {
local current="" path
if path="$(command -v cloudflared 2>/dev/null)"; then
current="$(cloudflared --version 2>/dev/null | awk '{ for (i = 1; i < NF; i++) if ($i == "version") { print $(i + 1); exit } }')"
if [ "$current" = "$FELIS_CLOUDFLARED_VERSION" ]; then
ok "cloudflared ${current} already installed"
return 0
fi
if [ "$FELIS_UPGRADE_DEPS" != 1 ]; then
ok "cloudflared ${current:-(version unreadable)} already installed; this release pins ${FELIS_CLOUDFLARED_VERSION} (FELIS_UPGRADE_DEPS=1 moves it)"
return 0
fi
if [ "$path" != "$CLOUDFLARED_BIN" ]; then
warn "cloudflared at ${path} was not installed by Felis; upgrade it the way it was installed"
return 0
fi
if [ -n "$current" ] && version_newer "$current" "$FELIS_CLOUDFLARED_VERSION"; then
ok "cloudflared ${current} is newer than the pinned ${FELIS_CLOUDFLARED_VERSION}; left as it is"
return 0
fi
fi
local machine arch url tmp want have
machine="$(uname -m)"
case "$machine" in
x86_64|amd64) arch="amd64"; want="$CLOUDFLARED_PINNED_SHA256_AMD64" ;;
aarch64|arm64) arch="arm64"; want="$CLOUDFLARED_PINNED_SHA256_ARM64" ;;
armv7l|armv6l) arch="arm"; want="$CLOUDFLARED_PINNED_SHA256_ARM" ;;
*) die "unsupported architecture for cloudflared: ${machine}" ;;
esac
[ "$FELIS_CLOUDFLARED_VERSION" = "$CLOUDFLARED_PINNED_VERSION" ] || want="$FELIS_CLOUDFLARED_SHA256"
[ -n "$want" ] || die "no pinned sha256 for cloudflared ${FELIS_CLOUDFLARED_VERSION}; set FELIS_CLOUDFLARED_SHA256 to the digest of cloudflared-linux-${arch} on that GitHub release"
url="https://github.com/cloudflare/cloudflared/releases/download/${FELIS_CLOUDFLARED_VERSION}/cloudflared-linux-${arch}"
tmp="$(mktemp)"
remember_temp "$tmp"
log "installing cloudflared ${FELIS_CLOUDFLARED_VERSION} (${arch})"
curl -fsSL --retry 5 --retry-delay 2 "$url" -o "$tmp"
have="$(sha256sum <"$tmp" | cut -d' ' -f1)"
if [ "$have" != "$(printf '%s' "$want" | tr 'A-Z' 'a-z')" ]; then
rm -f "$tmp"
die "cloudflared-linux-${arch} ${FELIS_CLOUDFLARED_VERSION} hashes to ${have}, expected ${want}; refusing to install it"
fi
install -m 0755 "$tmp" "$CLOUDFLARED_BIN"
rm -f "$tmp"
ok "cloudflared installed ($(cloudflared --version | head -n 1))"
# The running tunnel keeps the old binary mapped until it restarts.
if [ -n "$current" ] && systemctl is-active --quiet cloudflared-felis 2>/dev/null; then
systemctl restart cloudflared-felis
ok "cloudflared-felis restarted onto ${FELIS_CLOUDFLARED_VERSION}"
fi
}
# ---------------------------------------------------------------------------
# 3. Docker (used only to build & export the felis image; k3s uses containerd)
# ---------------------------------------------------------------------------
docker_apt_repo_os() {
case "$OS_ID" in
debian|ubuntu)
printf '%s\n' "$OS_ID"
;;
*)
die "Docker apt repository is not configured for ${OS_ID} ${OS_VERSION}"
;;
esac
}
install_docker_apt() {
local arch keyring repo_os
repo_os="$(docker_apt_repo_os)"
[ -n "$OS_CODENAME" ] || die "cannot determine apt codename for ${OS_ID} ${OS_VERSION}"
arch="$(dpkg --print-architecture)"
keyring="/etc/apt/keyrings/docker.asc"
log "installing docker apt repository"
install -m 0755 -d /etc/apt/keyrings
curl -fsSL "https://download.docker.com/linux/${repo_os}/gpg" -o "$keyring" \
|| die "failed to download Docker GPG key for ${repo_os}"
chmod a+r "$keyring"
cat > /etc/apt/sources.list.d/docker.list <<EOF
deb [arch=${arch} signed-by=${keyring}] https://download.docker.com/linux/${repo_os} ${OS_CODENAME} stable
EOF
apt_get update -y
pkg_install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
}
docker_rpm_repo_url() {
case "$OS_ID" in
fedora)
printf '%s\n' "https://download.docker.com/linux/fedora/docker-ce.repo"
;;
rhel)
printf '%s\n' "https://download.docker.com/linux/rhel/docker-ce.repo"
;;
centos|almalinux|rocky)
printf '%s\n' "https://download.docker.com/linux/centos/docker-ce.repo"
;;
*)
case " ${OS_ID_LIKE} " in
*" rhel "*|*" centos "*)
printf '%s\n' "https://download.docker.com/linux/centos/docker-ce.repo"
;;
*)
die "Docker rpm repository is not configured for ${OS_ID} ${OS_VERSION}"
;;
esac
;;
esac
}
install_docker_rpm() {
local repo_file repo_url
repo_file="/etc/yum.repos.d/docker-ce.repo"
repo_url="$(docker_rpm_repo_url)"
log "installing docker rpm repository"
mkdir -p "$(dirname "$repo_file")"
curl -fsSL "$repo_url" -o "$repo_file" \
|| die "failed to download Docker repo file: ${repo_url}"
pkg_install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
}
install_docker_zypper() {
log "installing docker from the distribution repositories"
pkg_install docker
}
install_docker_pacman() {
log "installing docker from the Arch repositories"
pkg_install docker docker-buildx
}
install_docker() {
if command -v docker >/dev/null 2>&1; then
ok "docker already installed"
else
case "$PKG" in
apt) install_docker_apt ;;
dnf|yum) install_docker_rpm ;;
zypper) install_docker_zypper ;;
pacman) install_docker_pacman ;;
*) die "Docker installation is not supported with package manager: ${PKG}" ;;
esac
fi
systemctl enable --now docker
ok "docker running"
}
# ---------------------------------------------------------------------------
# 4. k3s — single node, trimmed for RAM. NetworkPolicy stays ENABLED on purpose:
# Felis's minecraft fence (default-deny + allow-rcon/allow-game) is a core
# security claim, so we must NOT pass --disable-network-policy.
# On SUSE-family hosts firewalld ships active by default; open the required
# rules rather than disabling the firewall.
# ---------------------------------------------------------------------------
configure_k3s_firewall() {
command -v firewall-cmd >/dev/null 2>&1 || return 0
systemctl is-active --quiet firewalld || return 0
log "configuring firewalld for k3s"
firewall-cmd --permanent --add-port=6443/tcp
firewall-cmd --permanent --add-port="${FELIS_PANEL_NODEPORT}/tcp"
firewall-cmd --permanent --zone=trusted --add-source="$POD_CIDR"
firewall-cmd --permanent --zone=trusted --add-source="$SERVICE_CIDR"
firewall-cmd --reload
}
install_k3s() {
configure_k3s_firewall
if [ -x "$K3S_BIN" ]; then
local current
current="$("$K3S_BIN" --version 2>/dev/null | awk 'NR == 1 { print $3 }')"
if [ "$current" = "$FELIS_K3S_VERSION" ]; then
ok "k3s ${current} already installed at ${K3S_BIN}"
elif [ "$FELIS_UPGRADE_DEPS" != 1 ]; then
ok "k3s ${current:-(version unreadable)} already installed at ${K3S_BIN}; this release pins ${FELIS_K3S_VERSION} (FELIS_UPGRADE_DEPS=1 moves it)"
elif k3s_upgrade_allowed "$current" "$FELIS_K3S_VERSION"; then
log "upgrading k3s ${current} to ${FELIS_K3S_VERSION}; running pods keep running while it restarts"
run_k3s_installer
fi
else
log "installing k3s ${FELIS_K3S_VERSION} into ${K3S_BIN_DIR} (no traefik/servicelb/metrics-server)"
run_k3s_installer
fi
[ -x "$K3S_BIN" ] || die "k3s installation completed but ${K3S_BIN} is missing"
systemctl enable --now k3s
export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
log "waiting for the node to become Ready"
wait_for_node_ready
}
# The script from the release's own tag rather than get.k3s.io, which serves whatever
# master holds today. '+' is literal in a URL path, so the tag needs no escaping. On an
# installed k3s the same script replaces the binary in place and restarts the service.
run_k3s_installer() {
curl -sfL --retry 5 --retry-delay 2 "https://raw.githubusercontent.com/k3s-io/k3s/${FELIS_K3S_VERSION}/install.sh" | \
INSTALL_K3S_VERSION="$FELIS_K3S_VERSION" \
INSTALL_K3S_BIN_DIR="$K3S_BIN_DIR" \
INSTALL_K3S_EXEC="--disable traefik --disable servicelb --disable metrics-server --write-kubeconfig-mode 644" \
sh -
}
# k3s_upgrade_allowed decides whether an installed k3s ($1) may move to $2. Kubernetes
# supports upgrading one minor version at a time, so a larger jump stops the install
# before anything changed; a newer installed k3s is left as it is.
k3s_upgrade_allowed() {
local current="$1" want="$2" cur_major cur_minor want_major want_minor rest
IFS=. read -r cur_major cur_minor rest <<<"${current#v}"
IFS=. read -r want_major want_minor rest <<<"${want#v}"
case "${cur_major}${cur_minor}${want_major}${want_minor}" in
""|*[!0-9]*) die "cannot compare the installed k3s '${current}' with ${want}; upgrade it by hand (docs/operations.md §4)" ;;
esac
if version_newer "$current" "$want"; then
ok "k3s ${current} is newer than the pinned ${want}; left as it is"
return 1
fi
if [ "$cur_major" != "$want_major" ] || [ "$((want_minor - cur_minor))" -gt 1 ]; then
die "k3s ${current} -> ${want} skips a minor version, and Kubernetes upgrades one minor at a time. Rerun with FELIS_K3S_VERSION set to the newest v${cur_major}.$((cur_minor + 1)).x+k3sN release first (https://github.com/k3s-io/k3s/releases)"
fi
return 0
}
# Waits for the (single) node to report Ready. Shared by the k3s install and the
# registry-mirror restart below: both restart the agent, and a bootstrap that
# proceeds early fails later with a misleading "not found"/timeout instead.
wait_for_node_ready() {
local _
for _ in $(seq 1 60); do
if kube get nodes 2>/dev/null | grep -q ' Ready '; then
ok "k3s node Ready"
return 0
fi
sleep 5
done
kube get nodes || true
die "k3s node did not become Ready in time"
}
# ---------------------------------------------------------------------------
# 4b. The in-cluster registry: the node-side pull path, the registry's own
# image, and the hosting of every image this installer builds.
#
# Kubelet's image GC collects an unused image under disk pressure (drilled:
# the game images were collected and ImagePullBackOff had nothing to pull
# from). The fix is a pull source that is always there — the registry the
# bundle already renders. Two node-level facts make that work:
# * kubelet cannot reach the registry Service VIP (the live stack answered
# "Empty reply"), so containerd is told to go through the loopback
# hostPort the registry Deployment binds (the Deployment renders it) —
# that is configure_registry_mirror below;
# * the registry's own image (REGISTRY_IMAGE) must already be in containerd
# before the registry Deployment can start at all —
# import_registry_image below caches it.
# After deploy_bundle, push_images_to_registry mirrors the built images into
# the registry, so containerd's imported copies are a first-boot cache
# rather than the only copy.
# ---------------------------------------------------------------------------
# The node's containerd cannot dial the registry Service VIP, so pulls arrive
# over the loopback hostPort the registry Deployment binds. k3s reads this file
# when the agent starts and regenerates containerd's certs.d from it — no
# restart, no effect — so a CONTENT change restarts k3s; an identical file
# (every re-run) restarts nothing. K3S_REGISTRIES_FILE is a variable so
# bootstrap_test.sh can point the function at a scratch file.
configure_registry_mirror() {
local file="$K3S_REGISTRIES_FILE" tmp
tmp="$(mktemp)"
remember_temp "$tmp"
cat > "$tmp" <<EOF
mirrors:
"${REGISTRY_URL}":
endpoint:
- "http://${REGISTRY_PUSH_HOST}"
EOF
if [ -f "$file" ] && cmp -s "$tmp" "$file"; then
rm -f "$tmp"
ok "registry mirror already configured (${REGISTRY_URL} -> http://${REGISTRY_PUSH_HOST})"
return 0
fi
mkdir -p "$(dirname "$file")"
mv "$tmp" "$file"
log "restarting k3s to load the registry mirror (${REGISTRY_URL} -> http://${REGISTRY_PUSH_HOST})"
systemctl restart k3s
export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
wait_for_node_ready
}
# The registry Deployment runs REGISTRY_IMAGE (platform.defaultRegistryImage; the
# renderer's default — this script never passes --registry-image). On a box
# that cannot reach Docker Hub the Deployment can never start without a local
# copy, so the installer caches one whenever it can. Best-effort by design: if
# the pull fails the registry rollout still fails loudly at deploy_bundle, with
# the regular diagnostics — but for every box that CAN pull, the image is
# fetched exactly once, here, instead of at first pod start.
#
# crictl pulls through CRI, the same call kubelet makes, so containerd records the
# digest ref the Deployment names and kubelet finds it. A docker save/import round
# trip rewrites the manifest and would leave a copy the digest ref never matches.
import_registry_image() {
local images
# Read the list whole before matching; see postgres_installed for the SIGPIPE.
images="$(k3s_cmd ctr images ls -q 2>/dev/null || true)"
if grep -qxF "$(registry_image_containerd_ref)" <<<"$images"; then
ok "registry image ${REGISTRY_IMAGE} already in k3s containerd"
return 0
fi
log "pulling the registry's own image (${REGISTRY_IMAGE}) into k3s containerd"
if k3s_cmd crictl pull "$REGISTRY_IMAGE" >/dev/null; then
ok "registry image ${REGISTRY_IMAGE} pulled"
else
warn "could not pull ${REGISTRY_IMAGE}: the in-cluster registry will start only if the node can pull it from Docker Hub; on an air-gapped box import it by hand (docs/troubleshooting.md §8e)"
fi
}
# registry_image_containerd_ref prints the name containerd lists REGISTRY_IMAGE under
# once CRI has pulled it: the repository and the digest, with the tag dropped.
registry_image_containerd_ref() {
printf '%s@%s\n' "${REGISTRY_IMAGE%%:*}" "${REGISTRY_IMAGE#*@}"
}
# The registry pod runs REGISTRY_IMAGE and, as its registry-gate sidecar, the felis
# image — neither of which can be pulled from the registry they make up. A kubelet
# image GC that collected either would leave the registry, and every pull through
# it, dead until someone re-imported by hand. containerd reports an image labelled
# io.cri-containerd.pinned=pinned as pinned over CRI, and kubelet's image GC never
# removes a pinned image. Older felis/felis tags are unpinned first, so upgrades
# do not pile up pinned images forever.
pin_registry_images() {
local ref registry_ref
registry_ref="$(registry_image_containerd_ref)"
while read -r ref; do
case "$ref" in
"$FELIS_IMAGE"|"$registry_ref") ;;
*/felis/felis:*|docker.io/library/registry[:@]*)
k3s_cmd ctr images label "$ref" io.cri-containerd.pinned= >/dev/null 2>&1 || true ;;
esac
done < <(k3s_cmd ctr images ls -q 2>/dev/null || true)
for ref in "$FELIS_IMAGE" "$registry_ref"; do
if k3s_cmd ctr images label "$ref" io.cri-containerd.pinned=pinned >/dev/null 2>&1; then
ok "pinned ${ref} in containerd (exempt from kubelet image GC)"
else
warn "could not pin ${ref} in containerd: if the kubelet's image GC collects it, the registry pod cannot restart until it is re-imported (docs/troubleshooting.md §8e)"
fi
done
}
# pin_user_server_images fixes every user server still naming a tag in the platform
# registry (felis/paper:demo) to the digest that tag names now. It has to run before
# build_game_stack and push_images_to_registry put new builds under those tags: a
# server left on the bare tag would boot the new build on its next wake and open its
# world with a newer Minecraft version, and chunk upgrades cannot be undone. felis-api
# pins every server it creates; this catches the ones created before it did. A fresh
# install has no CRD, so nothing to pin; a running server restarts once onto the
# build it already runs.
pin_user_server_images() {
kube get crd minecraftservers.felis.lolicon.best >/dev/null 2>&1 || return 0
# The registry answers the lookups, and a k3s restart above may have left its pod
# still starting. A registry that never comes up fails the lookups below, loudly.
kube -n "$CONTROL_NS" rollout status deployment/registry --timeout=180s >/dev/null 2>&1 || true
if "$HOST_BIN" pin-images --namespace "$MINECRAFT_NS" \
--registry "$REGISTRY_URL" --endpoint "$REGISTRY_PUSH_HOST"; then
ok "user servers pinned to the builds they run"
else
warn "could not pin every user server listed above to its current build: each one still names a tag this run is about to point at a new build, so its next start may open its world with a newer Minecraft version. Pin it before starting it again: sudo felis pin-images (docs/troubleshooting.md §15b)"
fi
}
# ---------------------------------------------------------------------------
# 5. Source/binary + image build + containerd import
# ---------------------------------------------------------------------------
# repo_slug prints the "owner/name" of FELIS_REPO_URL, for the REST API.
repo_slug() {
printf '%s\n' "$FELIS_REPO_URL" | sed -e 's#^.*github\.com[:/]##' -e 's#\.git$##'
}
# github_api GETs a REST path and prints the body.
#
# The token goes in through `curl --config -` rather than `-H "Authorization: ..."`
# because argv is world-readable via /proc while this runs. Same reason git_auth uses a
# credential helper instead of a URL: a bootstrap that leaks its own credential to any
# local user has not really installed anything privately.
github_api() {
local url="https://api.github.com/$1"
local ua="felis-bootstrap (+${FELIS_REPO_URL})"
if [ -n "$FELIS_GITHUB_TOKEN" ]; then
printf 'header = "Authorization: Bearer %s"\n' "$FELIS_GITHUB_TOKEN" \
| curl -fsSL --retry 5 --retry-delay 2 --config - \
-A "$ua" -H "Accept: application/vnd.github+json" "$url"
else
curl -fsSL --retry 5 --retry-delay 2 \
-A "$ua" -H "Accept: application/vnd.github+json" "$url"
fi
}
# github_latest_tag prints the tag of the newest published stable release, or fails.
# Same endpoint internal/updater/github.go polls, so `felis update` and the installer
# can never disagree about what "latest" means.
github_latest_tag() {
local json tag
# Fetch first, filter second — the SIGPIPE reason documented on resolve_latest_game_jars.
json="$(github_api "repos/$(repo_slug)/releases/latest")" || return 1
tag="$(printf '%s' "$json" | grep -o '"tag_name"[[:space:]]*:[[:space:]]*"[^"]*"' || true)"
tag="${tag%%$'\n'*}"
tag="${tag#*:}" # "v1.2.3"
tag="${tag#*\"}" # v1.2.3"
tag="${tag%\"}" # v1.2.3
[ -n "$tag" ] || return 1
printf '%s\n' "$tag"
}
# felis_asset_arch prints the release-asset suffix for this host, or fails. Unlike the
# cloudflared/JRE/Go mappers just below, this one does NOT die on an unmapped architecture:
# those have no local alternative, whereas a missing prebuilt binary only means we compile
# the same tag on this host — slower, byte-for-byte the same result.
felis_asset_arch() {
case "$(uname -m)" in
x86_64|amd64) printf 'amd64\n' ;;
aarch64|arm64) printf 'arm64\n' ;;
*) return 1 ;;
esac
}
# github_asset_id prints the numeric id of the asset named $2 on release tag $1.
#
# Two details here are load-bearing and both are wrong in the obvious version. The newline
# flatten comes FIRST: api.github.com pretty-prints, so "name" and "url" land on different
# lines and splitting on "{" alone matches nothing at all. And the id is read out of the
# asset's own url, never from a bare "id" key — the payload carries a release id, an author
# id, one id per asset AND one per nested uploader, so matching "id" silently resolves the
# wrong file. Anchoring on the CLOSING quote keeps felis-linux-amd64 from also matching a
# sibling like felis-linux-amd64.sha256.
#
# The segment this greps ends at the nested "uploader" object, which is fine because "url"
# precedes it. Anything published after uploader (size, digest, browser_download_url) is NOT
# reachable this way — fetch releases/assets/<id> with the JSON Accept if that is ever needed.
github_asset_id() {
local tag="$1" name="$2" json id
# Fetch first, filter second — the SIGPIPE reason documented on resolve_latest_game_jars.
json="$(github_api "repos/$(repo_slug)/releases/tags/${tag}")" || return 1
id="$(printf '%s' "$json" | tr -d '\n' | tr '{' '\n' \
| grep "\"name\":[[:space:]]*\"${name}\"" \
| grep -o 'releases/assets/[0-9]\{1,\}' | head -1)" || true
id="${id##*/}"
[ -n "$id" ] || return 1
printf '%s\n' "$id"
}
# download_release_asset fetches ONE asset of a published release into $3.
#
# It RETURNS non-zero rather than dying: every caller falls back to building the same tag
# from source, so a tag whose release workflow has not finished uploading yet — a real window,
# since that job runs vet, tests and a full image build first — still installs.
#
# This cannot ride the github_api helper above. That helper sends
# Accept: application/vnd.github+json, and with it this endpoint returns the asset's METADATA
# as JSON under HTTP 200: a "successful" download of a text blob that passes every check and
# only surfaces much later, inside docker build, as an exec format error.
#
# Plain -L, never --location-trusted. GitHub 302s to a DIFFERENT host whose signed URL carries
# its own credentials in the query string; curl drops Authorization across that hop, and the
# drop is REQUIRED — forwarding the token makes the storage backend reject the request with
# 400 while leaking the credential to a third host for nothing.
#
# -f is not cosmetic: without it curl writes GitHub's error JSON into the output file and
# still exits 0.
download_release_asset() {
local tag="$1" name="$2" dest="$3" id ua url rc=0
ua="felis-bootstrap (+${FELIS_REPO_URL})"
id="$(github_asset_id "$tag" "$name")" || return 1
url="https://api.github.com/repos/$(repo_slug)/releases/assets/${id}"
log "downloading ${name} from release ${tag}"
if [ -n "$FELIS_GITHUB_TOKEN" ]; then
printf 'header = "Authorization: Bearer %s"\n' "$FELIS_GITHUB_TOKEN" \
| curl -fsSL --retry 5 --retry-delay 2 --config - \
-A "$ua" -H "Accept: application/octet-stream" -o "$dest" "$url" || rc=$?
else
curl -fsSL --retry 5 --retry-delay 2 \
-A "$ua" -H "Accept: application/octet-stream" -o "$dest" "$url" || rc=$?
fi
# curl -f leaves a PARTIAL file behind when a transfer dies mid-stream, so a failed
# download must not hand the caller something it could mistake for a complete one.
if [ "$rc" -ne 0 ] || [ ! -s "$dest" ]; then
rm -f "$dest"
return 1
fi
}
# verify_release_checksum checks the downloaded asset $2 (at $3) against the SHA256SUMS
# file release.yml publishes next to it on tag $1. It returns non-zero, with a warning,
# when the release has no SHA256SUMS (a tag cut before release.yml wrote one, or a release
# still uploading), when the file does not list the asset, or when the hash differs.
verify_release_checksum() {
local tag="$1" name="$2" file="$3" sums want have
sums="$(mktemp)"
remember_temp "$sums"
if ! download_release_asset "$tag" SHA256SUMS "$sums"; then
rm -f "$sums"
warn "release ${tag} publishes no SHA256SUMS, so ${name} cannot be verified; building ${tag} from source on this host instead"
return 1
fi
# sha256sum's text-mode line is "<hash> <name>", binary mode "<hash> *<name>".
want="$(awk -v n="$name" '$2 == n || $2 == "*" n { print $1; exit }' "$sums")"
rm -f "$sums"
if [ -z "$want" ]; then
warn "release ${tag}'s SHA256SUMS does not list ${name}; building ${tag} from source on this host instead"
return 1
fi
# Hash stdin, never the path: the same sha256sum escaping install_via_plugins avoids.
have="$(sha256sum <"$file" | cut -d' ' -f1)"
if [ "$have" != "$want" ]; then
warn "downloaded ${name} hashes to ${have}, but release ${tag}'s SHA256SUMS says ${want}; discarding it and building ${tag} from source on this host instead"
return 1
fi
ok "${name} matches release ${tag}'s SHA256SUMS"
}
# use_release_binary reports whether this run should install a prebuilt binary instead of
# compiling one. Only the plain release channel qualifies: FELIS_SKIP_FETCH means "build
# exactly what I staged" and a pinned FELIS_REF means "build that tree", both of which are
# explicit requests for a source build, and the dev channel has no release to download.
use_release_binary() {
[ -z "${FELIS_SKIP_FETCH:-}" ] || return 1
[ -z "$FELIS_REF_PINNED" ] || return 1
[ "$FELIS_VERSION_BOOTSTRAP" = "release" ]
}
# download_release_binary installs the prebuilt felis binary for $FELIS_REF onto the host.
# It returns non-zero to ask the caller to build that same tag from source instead; it never
# dies, because no reachable failure here is worth aborting an install over.
#
# One asset covers the panel too: internal/panel/panel.go go:embeds internal/panel/static, and
# release.yml builds through the repo Dockerfile so that tree holds the real npm output rather
# than the tracked placeholder. There is nothing else to fetch.
download_release_binary() {
local arch asset tmp got
if ! arch="$(felis_asset_arch)"; then
warn "no prebuilt felis binary for architecture $(uname -m); building ${FELIS_REF} from source on this host instead"
return 1
fi
asset="felis-linux-${arch}"
# Convergence check, and the cheapest one available: no API call, no download, and it asks
# the exact question that matters. Reruns are the common case for this installer.
if [ -x "$HOST_BIN" ] && [ "$("$HOST_BIN" version 2>/dev/null | head -n 1)" = "felis ${FELIS_REF}" ]; then
HAVE_PREBUILT_BINARY=1
ok "host binary is already ${FELIS_REF}; skipping the download"
return 0
fi
# Staged next to HOST_BIN rather than in TMPDIR, because the validation below EXECUTES it
# and /tmp is mounted noexec on CIS-hardened images. There the exec dies 126, the check
# reads it as a bad asset, and every such host silently falls back to the full on-host
# compile this path exists to avoid. /usr/local/bin has to be exec for felis to run at
# all, so validating there tests the binary instead of the mount — and it keeps the
# private-repo artifact out of a world-readable 1777 directory on the way in.
mkdir -p "$(dirname "$HOST_BIN")"
tmp="$(mktemp "$(dirname "$HOST_BIN")/.felis-download.XXXXXX")"
remember_temp "$tmp"
if ! download_release_asset "$FELIS_REF" "$asset" "$tmp"; then
rm -f "$tmp"
# Deliberately NOT "set FELIS_VERSION_BOOTSTRAP=dev": that channel builds main, a
# different commit than the tag the operator asked for. Falling back keeps the tag and
# changes only how it is obtained, so no operator action is needed at all.
warn "release ${FELIS_REF} publishes no usable ${asset}; building ${FELIS_REF} from source on this host instead"
return 1
fi
# The hash comes BEFORE the exec below: until the file matches the release's SHA256SUMS it
# is unverified bytes, and running it as root to ask its version would hand root to
# whoever could swap the asset or sit on the download path. Missing or unmatched both
# fall back to the source build of the same tag, which trusts only the git fetch.
if ! verify_release_checksum "$FELIS_REF" "$asset" "$tmp"; then
rm -f "$tmp"
return 1
fi
chmod 0755 "$tmp"
# Run it once. This single exec subsumes three checks that would otherwise each need their
# own code: a truncated download segfaults (curl reports success for an EOF-delimited body,
# so nothing earlier catches it), a wrong-architecture asset fails to exec, and an unstamped
# or mis-stamped build reports the wrong string here — rather than silently disabling
# `felis update` for the entire life of the install.
got="$("$tmp" version 2>/dev/null | head -n 1 || true)"
if [ "$got" != "felis ${FELIS_REF}" ]; then
rm -f "$tmp"
warn "downloaded ${asset} reports '${got:-nothing}' rather than 'felis ${FELIS_REF}'; building ${FELIS_REF} from source on this host instead"
return 1
fi
# install(1) onto a freshly created destination, matching install_embedded_binary. A rename
# would carry the source SELinux label instead of type-transitioning to bin_t — see
# build_nano_binary for the 203/EXEC this shape avoids. Same-directory staging does not
# change that: install(1) still creates the destination and copies.
keep_previous_host_binary
rm -f "$HOST_BIN"
install -m 0755 "$tmp" "$HOST_BIN"
rm -f "$tmp"
command -v restorecon >/dev/null 2>&1 && restorecon "$HOST_BIN" >/dev/null 2>&1 || true
HAVE_PREBUILT_BINARY=1
ok "installed ${asset} ${FELIS_REF} at ${HOST_BIN}"
}
# git_auth runs git with the token supplied by an inline credential helper. The helper
# is a shell snippet that READS the exported variable when git asks; the token is never
# in argv and never reaches .git/config, so it does not outlive the process.
#
# The EMPTY credential.helper in front is load-bearing, not a typo. credential.helper is a
# MULTI-valued config: a bare `-c credential.helper=...` APPENDS to whatever the host has
# configured, it does not replace it. On a host with `credential.helper=store` git then runs
# both — ours answers the prompt, and store writes the token to ~/.git-credentials on the
# approve that follows a successful auth, so it outlives the install after all. An empty
# value is git's documented list reset. Verified on Fedora: with store configured the single
# -c form persists the token to disk, the reset form writes nothing.
#
# GIT_TERMINAL_PROMPT=0 on both arms: GitHub answers a private repo with no or a bad token
# by asking for credentials, and git would put that prompt on /dev/tty, where a piped
# install sits waiting instead of failing with the FELIS_GITHUB_TOKEN hint.
git_auth() {
if [ -n "$FELIS_GITHUB_TOKEN" ]; then
GIT_TERMINAL_PROMPT=0 git -c 'credential.helper=' \
-c 'credential.helper=!f() { printf "username=x-access-token\npassword=%s\n" "$FELIS_GITHUB_TOKEN"; }; f' "$@"
else
GIT_TERMINAL_PROMPT=0 git "$@"
fi
}
# resolve_install_ref decides WHICH commit to build and, on the release channel, what to
# stamp it as. Runs before fetch_source because it chooses that fetch's ref.
resolve_install_ref() {
# Idempotent: main calls it early to fail fast on a missing token, and fetch_source
# calls it again because the nano path reaches fetch_source by its own route.
[ -n "${REF_RESOLVED:-}" ] && return 0
REF_RESOLVED=1
if [ -n "$FELIS_REF" ]; then
ok "building the pinned ref ${FELIS_REF}"
return 0
fi
case "$FELIS_VERSION_BOOTSTRAP" in
release)
log "resolving the newest published Felis release"
FELIS_REF="$(github_latest_tag)" || die "could not resolve the newest Felis release.
If the repository is private, set FELIS_GITHUB_TOKEN to a token with read access to it.
If no release has been published yet, set FELIS_VERSION_BOOTSTRAP=dev to build main instead."
# A release IS its tag, so the stamp is final here and stamp_version leaves it be.
FELIS_VERSION="$FELIS_REF"
ok "release channel: ${FELIS_REF}"
;;
dev)
FELIS_REF="main"
# Best effort: the tag only names what this build is AHEAD of, and a repo with no
# release yet is exactly when someone reaches for the dev channel. v0.0.0 keeps the
# stamp parseable so `felis update` still compares rather than refusing.
FELIS_VERSION_BASE="$(github_latest_tag 2>/dev/null || true)"
[ -n "$FELIS_VERSION_BASE" ] || FELIS_VERSION_BASE="v0.0.0"
ok "dev channel: main (ahead of ${FELIS_VERSION_BASE})"
;;
*)
die "FELIS_VERSION_BOOTSTRAP must be 'release' or 'dev', got '${FELIS_VERSION_BOOTSTRAP}'"
;;
esac
}
# stamp_version finalises the dev/pinned stamp once a checkout exists to read the SHA
# from. Release builds are already stamped and return untouched.
#
# The form is "<tag>+g<sha>", NOT `git describe`'s "<tag>-<n>-g<sha>", and the difference
# is load-bearing. updates.Parse reads a "-" tail as a PRERELEASE, which sorts BELOW the
# plain tag: a dev build 14 commits past v1.2.3 would compare as older than v1.2.3, and
# `felis update` would propose "upgrading" onto the release it already contains. After
# "+" the tail is build metadata, ignored for ordering, so the build reads as current
# against v1.2.3 and as behind against v1.3.0 — both correct.
#
# `git describe` is also unreliable here: the primary clone is --depth 1 and carries no
# tags, so describe falls back to a bare SHA, which updates.Parse rejects outright (it
# fails closed on a non-numeric core). fetch_source does retry with a full clone when the
# shallow one fails, which WOULD carry tags — that is exactly the point: the stamp must not
# depend on which arm happened to win. rev-parse needs no history at all.
stamp_version() {
local sha
[ -n "$FELIS_VERSION" ] && return 0
sha="$(git -C "$SRC_DIR" rev-parse --short HEAD 2>/dev/null || true)"
[ -n "$sha" ] || sha="unknown"
[ -n "$FELIS_VERSION_BASE" ] || FELIS_VERSION_BASE="v0.0.0"
FELIS_VERSION="${FELIS_VERSION_BASE}+g${sha}"
ok "build stamp: ${FELIS_VERSION}"
}
fetch_source() {
if [ -n "${FELIS_SKIP_FETCH:-}" ]; then
[ -d "$SRC_DIR" ] || die "FELIS_SKIP_FETCH set but ${SRC_DIR} does not exist"
ok "skipping fetch; using pre-staged source at ${SRC_DIR}"
# Whatever is staged is what gets built, so the channel has no say here and no
# release lookup is made. stamp_version reads the staged checkout's own SHA.
stamp_version
return 0
fi
resolve_install_ref
if [ -d "${SRC_DIR}/.git" ]; then
log "updating source in ${SRC_DIR}"
git_auth -C "$SRC_DIR" fetch --depth 1 origin "$FELIS_REF" \
|| die "could not fetch ${FELIS_REF} from ${FELIS_REPO_URL}; if the repository is private, set FELIS_GITHUB_TOKEN to a token with read access to it"
git -C "$SRC_DIR" checkout -f FETCH_HEAD
else
log "cloning ${FELIS_REPO_URL} (${FELIS_REF})"
mkdir -p "$(dirname "$SRC_DIR")"
# The fallback checks the ref out explicitly. It used to be a bare full clone, which
# silently landed on the default branch: harmless when FELIS_REF was always "main",
# but the release channel now asks for a tag, and a build stamped v1.2.3 that
# actually contains main is worse than a failed install.
git_auth clone --depth 1 --branch "$FELIS_REF" "$FELIS_REPO_URL" "$SRC_DIR" 2>/dev/null \
|| { git_auth clone "$FELIS_REPO_URL" "$SRC_DIR" \
&& git_auth -C "$SRC_DIR" checkout -f "$FELIS_REF"; } \
|| die "could not check out ${FELIS_REF} from ${FELIS_REPO_URL}; if the repository is private, set FELIS_GITHUB_TOKEN to a token with read access to it"
fi
stamp_version
ok "source ready at ${SRC_DIR}"
}
install_embedded_binary() {
local src
src="${FELIS_BOOTSTRAP_BINARY:-}"
[ -n "$src" ] || die "FELIS_BOOTSTRAP_BINARY is not set; cannot install the embedded setup binary"
[ -x "$src" ] || die "FELIS_BOOTSTRAP_BINARY is not executable: ${src}"
mkdir -p "$(dirname "$HOST_BIN")"
if [ "$(readlink -f "$src")" != "$(readlink -f "$HOST_BIN" 2>/dev/null || true)" ]; then
log "installing current felis binary onto the host (${HOST_BIN})"
keep_previous_host_binary
install -m 0755 "$src" "$HOST_BIN"
else
ok "host binary already installed at ${HOST_BIN}"
fi
HAVE_PREBUILT_BINARY=1
}
build_image_from_binary() {
local tmp
tmp="$(mktemp -d)"
remember_temp "$tmp"
cp "$HOST_BIN" "${tmp}/felis"
# static-debian12, matching the repo Dockerfile's final stage. Every binary that can reach
# HOST_BIN traces back to that Dockerfile's CGO_ENABLED=0 build: the downloaded CI asset,
# the binary the TUI is already running, and the one build_image_from_source docker-cp's
# out of the image it just built. None of them link glibc, so the larger base-debian12
# bought nothing and only widened the runtime surface. It mattered little while this was
# the rare fallback; now that the release channel downloads a binary and wraps it here,
# this is the image most installs actually run, and it should be the one CI publishes.
cat > "${tmp}/Dockerfile" <<'EOF'
FROM gcr.io/distroless/static-debian12:nonroot
ENV PATH=/usr/local/bin:/usr/bin:/bin
COPY felis /usr/local/bin/felis
USER 65532:65532
ENTRYPOINT ["/usr/local/bin/felis"]
EOF
chmod 0755 "${tmp}/felis"
log "building ${FELIS_IMAGE} from the current felis binary"
docker build -t "$FELIS_IMAGE" "$tmp"
rm -rf "$tmp"
}
verify_image_starts() {
log "verifying ${FELIS_IMAGE} starts"
docker run --rm --user 1000:1000 --entrypoint /usr/local/bin/felis "$FELIS_IMAGE" help >/dev/null
}
build_image_from_source() {
log "building ${FELIS_IMAGE} ${FELIS_VERSION:+(${FELIS_VERSION}) }(this compiles the Go binary; first run is slow)"
# Without the stamp main.version stays "dev", and `felis update` refuses to compare a
# "dev" build against upstream rather than treating it as 0.0.0. So an unstamped image
# is not a cosmetic problem: it silently disables update reporting for the install.
docker build -t "$FELIS_IMAGE" \
--build-arg FELIS_VERSION="${FELIS_VERSION:-dev}" "$SRC_DIR"
log "extracting the felis binary onto the host (${HOST_BIN})"
local cid
cid="$(docker create "$FELIS_IMAGE")"
remember_container "$cid"
keep_previous_host_binary
docker cp "${cid}:/usr/local/bin/felis" "$HOST_BIN"
docker rm "$cid" >/dev/null
chmod 0755 "$HOST_BIN"
}
# resolve_felis_image names the control-plane image after the release it carries
# (registry.felis.svc:5000/felis/felis:v1.2.3) unless FELIS_IMAGE was given. One tag per
# release is what makes `kubectl rollout undo` a rollback: the previous ReplicaSet names the
# previous tag, and the registry keeps the newest five of them (its pruner, §9). Under one
# mutable tag the undo re-created the pods on the image the upgrade had just written over it.
#
# The version is the release tag on the download path, the stamp on a source build
# (v1.2.3+gabc1234 becomes the tag v1.2.3-gabc1234: '+' is not allowed in a tag), and the
# binary's own report on the setup-console path, which skips both. A rerun of the same version
# reuses its tag; deploy_bundle restarts the pods onto the rebuilt image then.
resolve_felis_image() {
local v
[ -z "$FELIS_IMAGE" ] || return 0
v="$FELIS_VERSION"
if [ -z "$v" ] && [ -n "$HAVE_PREBUILT_BINARY" ]; then
v="$("$HOST_BIN" version 2>/dev/null | head -n 1 || true)"
v="${v#felis }"
fi
FELIS_IMAGE="${REGISTRY_URL}/felis/felis:$(image_tag_for_version "$v")"
ok "control-plane image: ${FELIS_IMAGE}"
}
# image_tag_for_version turns a felis version into an image tag: every character a tag may not
# hold becomes '-'. An unknown version ("dev" is what an unstamped binary reports) is :demo.
image_tag_for_version() {
local v
v="$(printf '%s' "$1" | tr -c 'A-Za-z0-9_.-' '-')"
case "$v" in
""|dev|[!A-Za-z0-9_]*) printf 'demo' ;;
*) printf '%s' "${v:0:128}" ;;
esac
}
build_image() {
systemctl start docker
# Keyed on the binary, not on the route that produced it: the TUI hand-off and a release
# download both land on exactly the same state (a felis binary at HOST_BIN, no checkout),
# and a release download that fell back to source has cleared this so the source build runs.
if [ -n "$HAVE_PREBUILT_BINARY" ]; then
build_image_from_binary
else
build_image_from_source
fi
verify_image_starts
log "importing ${FELIS_IMAGE} into k3s containerd"
remove_k3s_image "$FELIS_IMAGE"
docker save "$FELIS_IMAGE" | k3s_cmd ctr images import -
# Reclaim the ~150 MiB the docker daemon holds; reruns restart it on demand.
systemctl stop docker docker.socket 2>/dev/null || true
ok "image built, binary on host, image imported"
}
remove_k3s_image() {
local image="$1"
k3s_cmd ctr images rm "$image" >/dev/null 2>&1 || true
case "$image" in
*/*) ;;
*) k3s_cmd ctr images rm "docker.io/library/${image}" >/dev/null 2>&1 || true ;;
esac
}
# ---------------------------------------------------------------------------
# 5b. The game stack: the login limbo + lobby images, and the Velocity proxy.
#
# Without this, `felis setup` asks the Owner to bind by joining Minecraft while
# no Minecraft server exists — the whole point of installing is a joinable,
# Mojang-authenticating server, so the installer produces one.
#
# The trust chain, end to end:
# player --(Mojang auth)--> Velocity --(modern forwarding, HMAC)--> login limbo
# Velocity is the ONLY thing that talks to Mojang; the backends run offline-mode and
# trust the forwarded profile, which is exactly why the forwarding secret is the
# identity boundary. NetworkPolicy narrows reachability but cannot block the node
# hosting a pod, so it is defense in depth rather than a substitute for the HMAC.
# ---------------------------------------------------------------------------
# game_stack_source sets GAME_STACK_DIR to a docker build context holding
# deploy/{limbo,lobby} and plugins/. The TUI path pipes this script in over stdin and
# has no checkout on disk, so there the sources come out of the felis binary itself.
#
# Keyed on HAVE_PREBUILT_BINARY, the same flag build_image uses, because the question is
# identical: a prebuilt binary means fetch_source never ran, so SRC_DIR is whatever an
# EARLIER install left behind. Trusting it there would build felis-paper and the Velocity
# plugin from the old commit while the control plane is the freshly downloaded release —
# a silent version skew across the plugin/API boundary. The embedded tar always matches
# the binary it came out of, so it is the correct source on every prebuilt path.
game_stack_source() {
if [ -z "$HAVE_PREBUILT_BINARY" ] && [ -f "${SRC_DIR}/deploy/limbo/Dockerfile" ]; then
GAME_STACK_DIR="$SRC_DIR"
ok "game-stack sources: ${SRC_DIR}"
return 0
fi
GAME_STACK_DIR="$(mktemp -d)"
remember_temp "$GAME_STACK_DIR"
log "unpacking the embedded game-stack sources (no checkout on this host)"
"$HOST_BIN" bootstrap-assets game-stack | tar -x -C "$GAME_STACK_DIR" \
|| die "could not unpack the embedded game-stack sources"
[ -f "${GAME_STACK_DIR}/deploy/limbo/Dockerfile" ] \
|| die "embedded game-stack tar is missing deploy/limbo/Dockerfile"
ok "game-stack sources unpacked to ${GAME_STACK_DIR}"
}
# meta_get prints a small metadata document. curl's --retry covers transient HTTP
# statuses and timeouts; a TLS handshake cut mid-way (exit 35, seen against Fill over
# a flaky IPv6 path) is outside its retry set, so the outer loop retries every failure
# twice more. --retry-all-errors would say the same but needs curl 7.71+.
meta_get() {
local url="$1" out attempt
for attempt in 1 2 3; do
if out="$(curl -fsSL --retry 5 --retry-delay 2 \
-A "felis-bootstrap (+https://github.com/FelisMC/Felis)" "$url")"; then
printf '%s' "$out"
return 0
fi
if [ "$attempt" -lt 3 ]; then
warn "fetching ${url} failed (attempt ${attempt}/3); retrying"
sleep 5
fi
done
return 1
}
# resolve_game_jars sets the artifacts the three game images and the proxy are built from:
# the builds deploy/game-stack.lock names (FELIS_GAME_STACK=pinned, the default), or
# upstream's newest ones (latest). Pinned is what makes an install reproducible: every host
# installing one release gets the same login gate, lobby and proxy, each download is
# checked against the lock's sha256, and a rerun's docker builds hit their cache, so
# restart_existing_system_servers leaves the login and lobby pods running.
resolve_game_jars() {
if [ "$FELIS_GAME_STACK" = "latest" ]; then
resolve_latest_game_jars
return 0
fi
load_game_stack_lock "${GAME_STACK_DIR}/deploy/game-stack.lock"
ok "Limbo ${LIMBO_VERSION} + Paper on Minecraft ${MC_VERSION}, LuckPerms and Velocity ${VELOCITY_VERSION}: the builds game-stack.lock pins"
}
GAME_STACK_LOCK_KEYS="MC_VERSION LIMBO_VERSION LIMBO_JAR_URL LIMBO_JAR_SHA256 LIMBO_SCHEM_URL LIMBO_SCHEM_SHA256 PAPER_JAR_URL PAPER_JAR_SHA256 LUCKPERMS_JAR_URL LUCKPERMS_JAR_SHA256 VELOCITY_VERSION VELOCITY_JAR_URL VELOCITY_JAR_SHA256"
# load_game_stack_lock reads the lock's KEY=value lines into the globals of the same names.
# It never sources the file: only the keys above are accepted, every one has to be set, the
# values are limited to URL and version characters (they reach docker build-args), and each
# *_SHA256 has to be a sha256.
load_game_stack_lock() {
local file="$1" line key value
[ -f "$file" ] || die "no game-stack lock at ${file}; FELIS_GAME_STACK=latest resolves upstream's newest builds instead"
for key in $GAME_STACK_LOCK_KEYS; do printf -v "$key" '%s' ""; done
while IFS= read -r line || [ -n "$line" ]; do
case "$line" in ''|'#'*) continue ;; esac
key="${line%%=*}"
value="${line#*=}"
[ "$key" != "$line" ] || die "${file}: not a KEY=value line: ${line}"
case " ${GAME_STACK_LOCK_KEYS} " in
*" ${key} "*) ;;
*) die "${file}: unknown key ${key}" ;;
esac
case "$value" in
''|*[!A-Za-z0-9._:/+%-]*) die "${file}: ${key} has an unexpected value: ${value}" ;;
esac
printf -v "$key" '%s' "$value"
done < "$file"
for key in $GAME_STACK_LOCK_KEYS; do
[ -n "${!key}" ] || die "${file} does not set ${key}"
case "$key" in
*_SHA256) [[ "${!key}" =~ ^[0-9a-f]{64}$ ]] || die "${file}: ${key} is not a lowercase sha256" ;;
esac
done
}
# url_sha256 prints the sha256 of what $1 serves.
url_sha256() {
local sum
sum="$(curl -fsSL --retry 5 --retry-delay 2 -A "felis-bootstrap (+https://github.com/FelisMC/Felis)" "$1" | sha256sum)" || return 1
printf '%s\n' "${sum%% *}"
}
# resolve_latest_game_jars pins Limbo and Paper to the SAME Minecraft version. LOOHP/Limbo
# speaks exactly one protocol per build, so the login gate dictates the version and Paper
# follows — a client that can pass the gate must also be able to reach the lobby.
# MC_VERSION is read off Limbo's CI artifact name (Limbo-<limbo-ver>-<mc-ver>.jar), which
# is the only place the pairing is published.
#
# Limbo's CI and LuckPerms publish no digest, so latest hashes their downloads as it finds
# them: the image builds still check that they receive those bytes, but nothing vouches for
# the bytes themselves. That is what the lock file adds.
resolve_latest_game_jars() {
local ci="https://ci.loohpjames.com/job/Limbo/lastSuccessfulBuild" meta build file base rest paper
log "resolving the newest LOOHP/Limbo CI build"
# Fetch first, filter second: `curl | grep | head` dies of SIGPIPE under `set -o pipefail`
# the moment head closes the pipe early. Same shape everywhere below.
meta="$(meta_get "${ci}/api/json")" \
|| die "could not read the LOOHP/Limbo CI build metadata"
file="$(printf '%s' "$meta" | grep -o 'Limbo-[0-9A-Za-z._-]*\.jar' || true)"
file="${file%%$'\n'*}"
[ -n "$file" ] || die "no Limbo jar in the LOOHP/Limbo CI artifact list"
# The numbered build, so the jar hashed below is the jar the image build downloads even
# if CI finishes another build in between.
build="$(printf '%s' "$meta" | grep -o '"number":[0-9]*' || true)"
build="${build%%$'\n'*}"
build="${build#*:}"
[ -n "$build" ] || die "no build number in the LOOHP/Limbo CI metadata"
ci="https://ci.loohpjames.com/job/Limbo/${build}"
base="${file%.jar}" # Limbo-2026.0.2-ALPHA-26.2
MC_VERSION="${base##*-}" # 26.2
rest="${base%-*}" # Limbo-2026.0.2-ALPHA
LIMBO_VERSION="${rest#Limbo-}"
[ -n "$MC_VERSION" ] && [ -n "$LIMBO_VERSION" ] || die "cannot parse Limbo artifact name: ${file}"
LIMBO_JAR_URL="${ci}/artifact/target/${file}"
LIMBO_SCHEM_URL="${ci}/artifact/spawn.schem"
# PaperMC Fill v3. The old api.papermc.io v2 has returned HTTP 410 since 2026-07-01 and
# is never coming back; Fill wants a descriptive User-Agent.
log "resolving the newest Paper ${MC_VERSION} build"
paper="$(papermc_latest_jar paper "$MC_VERSION")" \
|| die "could not resolve a Paper build for Minecraft ${MC_VERSION} (the login gate pins this protocol; the build likely exists — Fill upstream is down, flapping, or no longer content-addressed)"
PAPER_JAR_URL="${paper% *}"
PAPER_JAR_SHA256="${paper##* }"
# LuckPerms is not version-matched to MC_VERSION the way Paper is: it ships one
# current Bukkit build that supports the whole supported Minecraft range, so there is
# no per-version endpoint to ask.
log "resolving the newest LuckPerms build"
LUCKPERMS_JAR_URL="$(luckperms_latest_jar)" \
|| die "could not resolve a LuckPerms build (metadata.luckperms.net is down or flapping); the lobby needs it for the panel's permission controls"
log "hashing the Limbo and LuckPerms downloads (their upstreams publish no digest)"
LIMBO_JAR_SHA256="$(url_sha256 "$LIMBO_JAR_URL")" || die "could not download ${LIMBO_JAR_URL}"
LIMBO_SCHEM_SHA256="$(url_sha256 "$LIMBO_SCHEM_URL")" || die "could not download ${LIMBO_SCHEM_URL}"
LUCKPERMS_JAR_SHA256="$(url_sha256 "$LUCKPERMS_JAR_URL")" || die "could not download ${LUCKPERMS_JAR_URL}"
VELOCITY_VERSION="$VELOCITY_LATEST_MINOR"
VELOCITY_JAR_URL=""
VELOCITY_JAR_SHA256=""
ok "Limbo ${LIMBO_VERSION} (CI build ${build}) + Paper, both on Minecraft ${MC_VERSION}; LuckPerms resolved"
warn "FELIS_GAME_STACK=latest: these are upstream's builds as of now, not the ones this release pins"
}
# luckperms_latest_jar prints the download URL of the current LuckPerms Bukkit build.
# Bukkit, not bukkit-legacy: legacy targets Minecraft 1.8-1.12, and Paper 26.2 is far
# past that. The same fetch-then-grep shape (and --retry rationale) as
# papermc_latest_jar; the metadata endpoint hands back every platform's URL at once, so
# the grep has to pin the /bukkit/ path segment or it would just as happily return the
# Fabric or Velocity jar, neither of which Paper can load.
luckperms_latest_jar() {
local json url
json="$(meta_get "https://metadata.luckperms.net/data/all")" || return 1
url="$(printf '%s' "$json" \
| grep -o 'https://download\.luckperms\.net/[0-9]\{1,\}/bukkit/loader/[^"]*\.jar' || true)"
url="${url%%$'\n'*}"
[ -n "$url" ] || return 1
printf '%s\n' "$url"
}
# papermc_latest_jar prints "<url> <sha256>" for the newest build of <project> <version>.
# --retry rides out Fill's transient gateway errors (502/503/504 are in curl's retry
# set): a single blip must not abort the whole bootstrap claiming the build is missing.
# Plain --retry only, deliberately: --retry-connrefused needs curl 7.52+, which the yum
# (el7) path does not have, and it would only add ECONNREFUSED to an already-covered set.
# The digest is not fished out of the JSON separately: Fill's download URLs are
# content-addressed (/v1/objects/<sha256>/<name>.jar), so the path segment names the
# bytes the URL serves and both halves come from the same grep of the same response. A
# URL without that shape fails the resolve rather than waving the download through
# unchecked.
papermc_latest_jar() {
local project="$1" version="$2" json urls url sha
json="$(meta_get "https://fill.papermc.io/v3/projects/${project}/versions/${version}/builds/latest")" || return 1
urls="$(printf '%s' "$json" | grep -o 'https://fill-data\.papermc\.io/[^"]*\.jar' || true)"
url="${urls%%$'\n'*}"
[ -n "$url" ] || return 1
sha="${url#*/objects/}"
sha="${sha%%/*}"
case "$sha" in *[!0-9a-f]*|"") return 1 ;; esac
[ "${#sha}" -eq 64 ] || return 1
printf '%s %s\n' "$url" "$sha"
}
build_game_stack() {
systemctl start docker
game_stack_source
resolve_game_jars
log "building ${FELIS_LIMBO_IMAGE} (LOOHP/Limbo ${LIMBO_VERSION}, Minecraft ${MC_VERSION})"
docker build -f "${GAME_STACK_DIR}/deploy/limbo/Dockerfile" \
--build-arg LIMBO_JAR_URL="$LIMBO_JAR_URL" \
--build-arg LIMBO_JAR_SHA256="$LIMBO_JAR_SHA256" \
--build-arg LIMBO_SCHEM_URL="$LIMBO_SCHEM_URL" \
--build-arg LIMBO_SCHEM_SHA256="$LIMBO_SCHEM_SHA256" \
--build-arg LIMBO_VERSION="$LIMBO_VERSION" \
-t "$FELIS_LIMBO_IMAGE" "$GAME_STACK_DIR"
log "building ${FELIS_LOBBY_IMAGE} (Paper ${MC_VERSION} + felis-paper /menu + LuckPerms)"
docker build -f "${GAME_STACK_DIR}/deploy/lobby/Dockerfile" \
--build-arg PAPER_JAR_URL="$PAPER_JAR_URL" \
--build-arg PAPER_JAR_SHA256="$PAPER_JAR_SHA256" \
--build-arg LUCKPERMS_JAR_URL="$LUCKPERMS_JAR_URL" \
--build-arg LUCKPERMS_JAR_SHA256="$LUCKPERMS_JAR_SHA256" \
-t "$FELIS_LOBBY_IMAGE" "$GAME_STACK_DIR"
# Plain Paper, same MC_VERSION and PAPER_JAR_URL (no new dependency). Forwarding is the
# operator initContainer's job, so this image carries no /menu plugin and no secret gate.
log "building ${FELIS_PAPER_IMAGE} (plain Paper ${MC_VERSION}, forwarding via the operator initContainer)"
docker build -f "${GAME_STACK_DIR}/deploy/paper/Dockerfile" \
--build-arg PAPER_JAR_URL="$PAPER_JAR_URL" \
--build-arg PAPER_JAR_SHA256="$PAPER_JAR_SHA256" \
-t "$FELIS_PAPER_IMAGE" "$GAME_STACK_DIR"
# The builds the system servers run, for restart_existing_system_servers. Docker's layer
# cache gives an unchanged build the same id, so a rerun that rebuilt nothing leaves the
# login and lobby pods (and every player on them) alone.
LIMBO_IMAGE_ID="$(docker image inspect -f '{{.Id}}' "$FELIS_LIMBO_IMAGE")"
LOBBY_IMAGE_ID="$(docker image inspect -f '{{.Id}}' "$FELIS_LOBBY_IMAGE")"
local img
for img in "$FELIS_LIMBO_IMAGE" "$FELIS_LOBBY_IMAGE" "$FELIS_PAPER_IMAGE"; do
log "importing ${img} into k3s containerd"
remove_k3s_image "$img"
docker save "$img" | k3s_cmd ctr images import -
done
build_velocity_plugin
systemctl stop docker docker.socket 2>/dev/null || true
ok "login + lobby images imported; felis-velocity.jar staged"
}
ensure_velocity_directory() {
local path="$1" mode="$2" owner="$3" group="$4"
if [ -L "$path" ]; then
rm -f -- "$path"
elif [ -e "$path" ] && [ ! -d "$path" ]; then
die "velocity path exists but is not a directory: ${path}"
fi
install -d -o "$owner" -g "$group" -m "$mode" "$path"
}
# Velocity manages its own working directory: on startup it migrates velocity.toml to
# the running config-version, extracts localizations, and creates plugins/bStats — all
# fatal or noisy if it cannot write. So the service owns the tree and velocity.toml. The
# immutable artifacts (velocity.jar, felis-velocity.jar) and the forwarding secret stay
# root-owned and read-only to the proxy; ProtectSystem=strict confines writes to VELOCITY_DIR.
prepare_velocity_layout() {
id -u "$VELOCITY_USER" >/dev/null 2>&1 \
|| useradd --system --home-dir "$VELOCITY_DIR" --shell /usr/sbin/nologin "$VELOCITY_USER"
ensure_velocity_directory "$VELOCITY_DIR" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
ensure_velocity_directory "${VELOCITY_DIR}/plugins" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
ensure_velocity_directory "${VELOCITY_DIR}/plugins/felis-link" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
ensure_velocity_directory "${VELOCITY_DIR}/logs" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
ensure_velocity_directory "${VELOCITY_DIR}/crash-reports" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
}
atomic_install_file() {
local source="$1" target="$2" mode="$3" owner="$4" group="$5"
local parent base staged
parent="$(dirname "$target")"
base="$(basename "$target")"
[ -d "$parent" ] && [ ! -L "$parent" ] \
|| die "refusing to install through a non-directory/symlink parent: ${parent}"
if [ -d "$target" ] && [ ! -L "$target" ]; then
die "refusing to replace directory with file: ${target}"
fi
staged="$(mktemp "${parent}/.${base}.XXXXXX")"
remember_temp "$staged"
install -o "$owner" -g "$group" -m "$mode" "$source" "$staged"
mv -fT "$staged" "$target"
}
# build_velocity_plugin compiles plugins/velocity in the same gradle image the two
# Dockerfiles use, and drops the jar where Velocity will look for it. Docker is the
# toolchain here on purpose: the host needs no JDK and no gradle, only a JRE.
build_velocity_plugin() {
log "building felis-velocity.jar (gradle in a container; the host gets no JDK)"
prepare_velocity_layout
# :z relabels the bind mount for SELinux (Fedora/EL enforce it; elsewhere it is a no-op).
docker run --rm \
-v "${GAME_STACK_DIR}:/src:z" \
-w /src/plugins/velocity \
gradle:8.14-jdk21 gradle --no-daemon clean build \
|| die "felis-velocity plugin build failed"
local -a jars=( "${GAME_STACK_DIR}"/plugins/velocity/build/libs/felis-velocity-*.jar )
[ "${#jars[@]}" -eq 1 ] && [ -f "${jars[0]}" ] \
|| die "felis-velocity build must produce exactly one plugin jar"
atomic_install_file "${jars[0]}" "${VELOCITY_DIR}/plugins/felis-velocity.jar" 0644 root root
}
# install_via_plugins stages ViaVersion + ViaBackwards + ViaRewind so players on clients
# older than the proxy can still join.
#
# Modern forwarding nominally refuses anything below 1.13: HandshakeSessionHandler#handleLogin
# reads the handshake protocol version and disconnects with
# velocity.error.modern-forwarding-needs-new-client. That gate stops firing once Via is
# present — it logs "Replacing channel initializers" during startup, so the version reaching
# the check is plausibly already the rewritten one. The 1.13 floor is a property of the
# UNASSISTED proxy pipeline, not of the forwarding protocol, so nothing here changes
# player-info-forwarding-mode and no backend is patched or downgraded.
#
# Measured end to end rather than assumed (Felis-Legacy FL-007, cell modern-via121): a
# protocol-47 client joined a stock Paper 1.21.11 backend through a modern-forwarding proxy.
# The proof is the join itself, not the log line — that backend ran velocity.enabled=true with
# a shared secret, and Paper in that state rejects any login not carrying forwarding data
# signed with a matching HMAC. Only protocol 47 was measured; the rest of Via's 1.7-1.12 range
# is its own documented support.
#
# Pinned by hash and not by "latest" on purpose. These three jars sit in front of every packet
# on the proxy, and they are the exact bytes FL-007 measured — a moving tag would quietly make
# this an unmeasured configuration. Bumping a version means bumping its checksum here.
#
# The three versions are a set, not three independent pins. ViaRewind is the component that
# carries 1.8/1.7 support, and 4.1.2 against ViaVersion/ViaBackwards 5.11.0 fails to load
# Protocol1_9To1_8 — the single protocol every 1.8 client needs — with "Invalid version: 1"
# at proxy startup. 4.1.3 is the release that adds 5.11.0 compatibility; a two-arm run of the
# same proxy image logs that error three times on 4.1.2 and not at all on 4.1.3. Read the
# ViaRewind release notes before moving ViaVersion or ViaBackwards.
install_via_plugins() {
prepare_velocity_layout
local name version want target url tmp have
while read -r name version want; do
[ -n "$name" ] || continue
target="${VELOCITY_DIR}/plugins/${name}.jar"
# Hash stdin, never the path: sha256sum escapes its output line when the filename
# carries a backslash or a newline, and a leading "\" on the digest silently fails
# every comparison below.
have=""
[ -f "$target" ] && have="$(sha256sum <"$target" | cut -d' ' -f1)"
if [ "$have" = "$want" ]; then
ok "${name} ${version} already staged"
continue
fi
url="https://github.com/ViaVersion/${name}/releases/download/${version}/${name}-${version}.jar"
log "downloading ${name} ${version}"
tmp="$(mktemp "${VELOCITY_DIR}/.${name}.jar.XXXXXX")"
remember_temp "$tmp"
curl -fsSL "$url" -o "$tmp" || die "failed to download ${name} ${version}: ${url}"
have="$(sha256sum <"$tmp" | cut -d' ' -f1)"
[ "$have" = "$want" ] \
|| die "${name} ${version} checksum mismatch: got ${have}, expected ${want}"
atomic_install_file "$tmp" "$target" 0644 root root
done <<'EOF'
ViaVersion 5.11.0 18d19e90fc9467d68128c076630ae8700449c901402a3ef421837ce006bc8cae
ViaBackwards 5.11.0 b21983d561e3f92df257683f0133ab6c68ec68175e8acfd82c6231723bf83587
ViaRewind 4.1.3 2d5970d22b4711c9ab2800932326c7b08acdace25ed7c6bbb8f6ea81054962b4
EOF
pin_via_block_connections
ok "Via staged; clients from 1.8 up can join under modern forwarding"
}
# pin_via_block_connections turns ViaVersion's serverside block-connection tracking off.
#
# ConnectionData.init() only builds its block-connection provider when Via's lowest supported
# protocol is below 1.13. Under modern forwarding the Velocity injector reports 393 (1.13), so
# init() returns early, blockConnectionProvider stays null, and the first 1.12.2->1.13 chunk
# rewrite dereferences it. A 1.8 client on a protocol-47 backend takes an NPE on the first chunk
# it is sent and never finishes joining. Every call site in protocols/v1_12_2to1_13 is behind
# isServersideBlockConnections(), so switching the option off skips all of them. The cost is
# cosmetic and pre-1.13 only: fences and glass panes stop being drawn connected.
#
# ViaVersion's default is true, so a fresh install ships that NPE unless it is corrected here.
# Seeding a file with this one key is enough: Config#loadConfig parses the bundled
# assets/viaversion/config.yml as the base map and merges the on-disk file over it, so every
# other option still comes from the shipped default and stays current across version bumps.
#
# Do not read the absence of "Loading block connection mappings" from the log as proof this
# worked. init() gates on the protocol version as well, and under modern forwarding that half
# fails on its own — the line is missing either way. The config value is the only evidence.
pin_via_block_connections() {
local dir="${VELOCITY_DIR}/plugins/viaversion" config tmp
config="${dir}/config.yml"
ensure_velocity_directory "$dir" 0750 "$VELOCITY_USER" "$VELOCITY_USER"
tmp="$(mktemp "${VELOCITY_DIR}/.viaversion-config.XXXXXX")"
remember_temp "$tmp"
if [ ! -f "$config" ]; then
printf 'serverside-blockconnections: false\n' > "$tmp"
elif grep -qE '^serverside-blockconnections:' "$config"; then
sed -E 's/^serverside-blockconnections:.*/serverside-blockconnections: false/' "$config" > "$tmp"
else
{ cat "$config"; printf 'serverside-blockconnections: false\n'; } > "$tmp"
fi
# Via rewrites this file itself on every load, so the proxy user has to own it.
atomic_install_file "$tmp" "$config" 0640 "$VELOCITY_USER" "$VELOCITY_USER"
}
install_jre() {
local arch want url release json
case "$(uname -m)" in
x86_64|amd64) arch="x64"; want="$JRE_PINNED_SHA256_X64" ;;
aarch64|arm64) arch="aarch64"; want="$JRE_PINNED_SHA256_AARCH64" ;;
*)
if [ -x "${JRE_DIR}/bin/java" ]; then
ok "JRE already installed at ${JRE_DIR}"
return 0
fi
die "no Temurin JRE build for architecture $(uname -m); pre-stage one at ${JRE_DIR}"
;;
esac
if [ "$FELIS_JRE_VERSION" = "$JRE_PINNED_FEATURE" ]; then
release="$JRE_PINNED_RELEASE"
url="https://github.com/adoptium/temurin${JRE_PINNED_FEATURE}-binaries/releases/download/jdk-${release/+/%2B}/OpenJDK${JRE_PINNED_FEATURE}U-jre_${arch}_linux_hotspot_${release/+/_}.tar.gz"
else
# Another feature version is installed once and then left alone; its digest is the
# one Adoptium's API publishes next to the link.
if [ -x "${JRE_DIR}/bin/java" ]; then
ok "JRE already installed at ${JRE_DIR}"
return 0
fi
json="$(meta_get "https://api.adoptium.net/v3/assets/latest/${FELIS_JRE_VERSION}/hotspot?architecture=${arch}&image_type=jre&os=linux&vendor=eclipse")" \
|| die "could not ask the Adoptium API for a Temurin ${FELIS_JRE_VERSION} JRE"
url="$(printf '%s' "$json" | grep -o '"link": *"[^"]*\.tar\.gz"' || true)"
url="${url%%$'\n'*}"
url="${url%\"}"
url="${url##*\"}"
want="$(printf '%s' "$json" | grep -o '"checksum": *"[0-9a-f]\{64\}"' || true)"
want="${want%%$'\n'*}"
want="${want%\"}"
want="${want##*\"}"
release="$(printf '%s' "$json" | grep -o '"release_name": *"jdk-[^"]*"' || true)"
release="${release%%$'\n'*}"
release="${release%\"}"
release="${release##*\"jdk-}"
[ -n "$url" ] && [ -n "$want" ] && [ -n "$release" ] \
|| die "the Adoptium API lists no Temurin ${FELIS_JRE_VERSION} JRE for linux/${arch}"
fi
# A rerun moves the installer's own JRE to the pinned build, which is how a JRE security
# release reaches the proxy: bump the pin, rerun, and install_velocity_service restarts
# the proxy because the JRE's release file changed. A JRE someone else put here (another
# vendor, or pre-staged for an architecture Temurin does not build) is left alone.
if [ -x "${JRE_DIR}/bin/java" ]; then
if grep -qxF "IMPLEMENTOR_VERSION=\"Temurin-${release}\"" "${JRE_DIR}/release" 2>/dev/null; then
ok "Temurin ${release} JRE already installed at ${JRE_DIR}"
return 0
fi
if ! grep -qxF 'IMPLEMENTOR="Eclipse Adoptium"' "${JRE_DIR}/release" 2>/dev/null; then
ok "JRE at ${JRE_DIR} is not a Temurin build this installer put there; left as is"
return 0
fi
log "moving the proxy's JRE to Temurin ${release}"
fi
log "installing Temurin ${release} JRE (${arch}) to ${JRE_DIR}"
local tmp have
tmp="$(mktemp -d)"
remember_temp "$tmp"
curl -fsSL --retry 5 --retry-delay 2 "$url" -o "${tmp}/jre.tar.gz" \
|| die "failed to download the Temurin JRE: ${url}"
have="$(sha256sum <"${tmp}/jre.tar.gz" | cut -d' ' -f1)"
[ "$have" = "$want" ] \
|| die "Temurin ${release} JRE (${arch}) hashes to ${have}, expected ${want}; refusing to install it"
# Unpacked beside the live one and swapped in with two renames, so a failed unpack leaves
# the proxy's runtime untouched. The running proxy keeps the files it has open.
rm -rf "${JRE_DIR}.new" "${JRE_DIR}.old"
mkdir -p "${JRE_DIR}.new"
# The tarball has a single versioned top-level directory (jdk-25+36-jre/); strip it so
# the path in the systemd unit never carries a build number.
tar -C "${JRE_DIR}.new" --strip-components=1 -xzf "${tmp}/jre.tar.gz" || die "failed to unpack the JRE"
[ -x "${JRE_DIR}.new/bin/java" ] || die "unpacked JRE has no bin/java"
[ ! -e "$JRE_DIR" ] || mv "$JRE_DIR" "${JRE_DIR}.old"
mv "${JRE_DIR}.new" "$JRE_DIR"
rm -rf "${JRE_DIR}.old"
ok "JRE at ${JRE_DIR}/bin/java (Temurin ${release})"
}
install_velocity() {
install_jre
local url tmp have want resolved
prepare_velocity_layout
if [ -n "$FELIS_VELOCITY_FORK_JAR" ]; then
[ -f "$FELIS_VELOCITY_FORK_JAR" ] \
|| die "FELIS_VELOCITY_FORK_JAR is not a readable file: ${FELIS_VELOCITY_FORK_JAR}"
# Hash stdin, never the path — same reason as install_via_plugins: sha256sum escapes its
# output line for a filename carrying a backslash or a newline, and the leading "\" that
# adds would fail every comparison below.
have="$(sha256sum <"$FELIS_VELOCITY_FORK_JAR" | cut -d' ' -f1)"
# Refuse rather than warn. This jar is the proxy every player connects through, and a
# warning in an install log is not a gate. The digest is printed so the first run after
# a deliberate rebuild is one copy-paste, not an investigation.
[ -n "$FELIS_VELOCITY_FORK_JAR_SHA256" ] || die \
"FELIS_VELOCITY_FORK_JAR_SHA256 is required whenever FELIS_VELOCITY_FORK_JAR is set.
The jar at that path hashes to ${have}.
Check that against the build you meant to install, then re-run with
FELIS_VELOCITY_FORK_JAR_SHA256=${have}"
# Normalise the operator's digest before comparing. sha256sum prints lowercase, but the
# build host is often Windows, where Get-FileHash prints uppercase and certutil has
# shipped both with and without spaces between the bytes. All three name the same jar,
# so comparing raw would refuse two of the three spellings and word it as tampering.
want="$(printf '%s' "$FELIS_VELOCITY_FORK_JAR_SHA256" | tr -d '[:space:]' | tr 'A-Z' 'a-z')"
[ "$have" = "$want" ] || die \
"FELIS_VELOCITY_FORK_JAR checksum mismatch: got ${have}, expected ${want}"
log "installing the Felis-Legacy Velocity fork from ${FELIS_VELOCITY_FORK_JAR} (sha256 ${have})"
atomic_install_file "$FELIS_VELOCITY_FORK_JAR" "${VELOCITY_DIR}/velocity.jar" 0644 root root
else
local version="${FELIS_VELOCITY_VERSION:-${VELOCITY_VERSION:-$VELOCITY_LATEST_MINOR}}"
if [ "$FELIS_GAME_STACK" = "pinned" ] && [ "$version" = "${VELOCITY_VERSION:-}" ]; then
url="$VELOCITY_JAR_URL"
want="$VELOCITY_JAR_SHA256"
else
log "resolving the newest Velocity ${version} build"
resolved="$(papermc_latest_jar velocity "$version")" \
|| die "no Velocity build for ${version} (override with FELIS_VELOCITY_VERSION)"
url="${resolved% *}"
want="${resolved##* }"
fi
stage_velocity_jar "$url" "$want" "$version"
fi
install_via_plugins
write_velocity_config
install_velocity_service
configure_velocity_firewall
}
# stage_velocity_jar installs the stock proxy from $1 unless velocity.jar already hashes to
# $2, so a rerun of the same build neither downloads nor touches the file the proxy runs.
stage_velocity_jar() {
local url="$1" want="$2" version="$3" have="" tmp
[ -f "${VELOCITY_DIR}/velocity.jar" ] && have="$(sha256sum <"${VELOCITY_DIR}/velocity.jar" | cut -d' ' -f1)"
if [ "$have" = "$want" ]; then
ok "Velocity ${version} already staged"
return 0
fi
log "downloading Velocity ${version}"
tmp="$(mktemp "${VELOCITY_DIR}/.velocity.jar.XXXXXX")"
remember_temp "$tmp"
curl -fsSL --retry 5 --retry-delay 2 "$url" -o "$tmp" || die "failed to download Velocity: ${url}"
# The same gate the Via plugins and the fork jar pass: this jar is the proxy every
# player connects through, and the lock file (or Fill's content-addressed URL) names its
# digest — a truncated or tampered download becomes a refusal here, not a proxy that
# won't boot.
have="$(sha256sum <"$tmp" | cut -d' ' -f1)"
[ "$have" = "$want" ] \
|| die "Velocity ${version} checksum mismatch: got ${have}, expected ${want}"
atomic_install_file "$tmp" "${VELOCITY_DIR}/velocity.jar" 0644 root root
}
# felis_internal_ip echoes the felis-api-internal Service ClusterIP. Cluster DNS does not
# resolve from the host, but a Service ClusterIP DOES route from the node (kube-proxy programs
# the host netns) — the same trick the on-node break-glass console uses. The internal face is
# deliberately ClusterIP-only: it is service-token authenticated and must never be published on
# a node's external IP. Both the felis-link plugin config and the Velocity sessionserver
# override (install_velocity_service) point at it, so the lookup lives here once.
felis_internal_ip() {
local ip
ip="$(kube -n "$CONTROL_NS" get svc felis-api-internal -o jsonpath='{.spec.clusterIP}')" \
|| die "could not resolve the felis-api-internal ClusterIP"
[ -n "$ip" ] || die "felis-api-internal has no ClusterIP"
printf '%s' "$ip"
}
write_velocity_config() {
local api_ip tmp
api_ip="$(felis_internal_ip)"
prepare_velocity_layout
tmp="$(mktemp -d)"
remember_temp "$tmp"
# The forwarding key. Velocity refuses to start on an empty one ("The forwarding-secret
# file must not be empty."), which is the failure mode we want if this ever goes wrong.
(umask 077; printf '%s' "$FORWARDING_SECRET" > "${tmp}/forwarding.secret")
cat > "${tmp}/velocity.toml" <<EOF
# Generated by deploy/bootstrap.sh — do not edit by hand; rerun the installer.
config-version = "2.7"
bind = "0.0.0.0:${FELIS_GAME_PORT}"
motd = "A Felis server"
show-max-players = 100
# The crown jewel. THIS is the process that talks to Mojang. Every backend runs
# offline-mode and trusts the profile forwarded from here, so online-mode = false would
# not "relax auth" — it would let anyone join as anyone, the Owner's account included.
online-mode = true
force-key-authentication = true
prevent-client-proxy-connections = false
# modern forwarding hands the Mojang-verified profile to the backend under an HMAC keyed
# by forwarding.secret. The legacy (BungeeCord) mode carries no secret and fails OPEN, so
# it is not an option at any price. Mode is proxy-wide, hence every backend gets the key.
player-info-forwarding-mode = "modern"
forwarding-secret-file = "forwarding.secret"
announce-forge = false
kick-existing-players = false
ping-passthrough = "disabled"
enable-player-address-logging = true
[servers]
# A dead address on purpose. felis-velocity re-registers "${LOGIN_SERVER}" with the real
# backend endpoint it reads from felis-api (ServerRegistry) a few seconds after start, and
# until it does, the proxy must point somewhere that REFUSES rather than somewhere that
# admits. It cannot simply be omitted: Velocity's config validation rejects a \`try\` entry
# that is not in [servers] ("Fallback server ... is not registered in your configuration!").
${LOGIN_SERVER} = "127.0.0.1:1"
# The login gate is the only fallback, and it authenticates. Never add the lobby here:
# a fallback to "${LOBBY_SERVER}" would route an unauthenticated player straight past the gate.
try = ["${LOGIN_SERVER}"]
[forced-hosts]
# Empty on purpose — felis-velocity does host-based routing itself from the subdomains
# felis-api reports (spec §11), so static entries here would only go stale.
[advanced]
haproxy-protocol = false
# Off on purpose. Velocity answers bungeecord:main itself, before any plugin event, so
# with it on EVERY backend — including each user's own server and whatever plugins its
# owner installed — can KickPlayer or ConnectOther anyone on the network, and no plugin
# can restrict that to one server. The login gate releases players over felis:control
# instead, which felis-velocity accepts only from the login server.
bungee-plugin-message-channel = false
[query]
enabled = false
EOF
(umask 077; cat > "${tmp}/felis-link.properties" <<EOF
# Generated by deploy/bootstrap.sh — do not edit by hand; rerun the installer.
api-base-url=http://${api_ip}:8081
service-token=${SERVICE_TOKEN}
root-domain=${FELIS_ROOT_DOMAIN}
panel-hostname=console.${FELIS_ROOT_DOMAIN}
admin-hostname=op.console.${FELIS_ROOT_DOMAIN}
login-server=${LOGIN_SERVER}
lobby-server=${LOBBY_SERVER}
EOF
)
atomic_install_file "${tmp}/forwarding.secret" "${VELOCITY_DIR}/forwarding.secret" 0640 root "$VELOCITY_USER"
atomic_install_file "${tmp}/velocity.toml" "${VELOCITY_DIR}/velocity.toml" 0640 "$VELOCITY_USER" "$VELOCITY_USER"
atomic_install_file "${tmp}/felis-link.properties" \
"${VELOCITY_DIR}/plugins/felis-link/felis-link.properties" 0640 root "$VELOCITY_USER"
ok "velocity.toml + forwarding secret + felis-link.properties written (${VELOCITY_DIR})"
}
install_velocity_service() {
local api_ip
# Point Velocity (-Dmojang.sessionserver) at the felis-api hasJoined multiplexer so a
# full install federates Mojang + the configured [[auth_source]] set (LittleSkin by default)
# out of the box — not just the standalone `felis nano`. felis-api enforces the reclaim
# blacklist on this route; a loopback nano would bypass it.
api_ip="$(felis_internal_ip)"
# Servers that receive their forwarded identity through the handshake address (BungeeCord/legacy
# style) instead of the proxy-wide modern+secret forwarding. A protocol-47 (1.8.x) backend sits
# behind ViaVersion, which strips modern forwarding's login-plugin-message when it down-translates
# the proxy->backend pipeline to protocol 47; only the handshake field survives Via. The Felis
# fork reads this list from -Dfelis.legacy-forwarding.servers and forwards those servers legacy;
# every other backend keeps modern+secret untouched.
#
# The list is a JVM system property, so it is fixed for the life of the proxy process and a
# change needs a Velocity restart. FELIS_LEGACY_FORWARDING_SERVERS makes that reachable
# without editing this script, which is as far as a startup property can go. Having it follow
# the MinecraftServer CRs instead is a larger change: the forwarding decision lives in the
# fork's patch to Velocity core, not in the Felis plugin, so core would need to read state the
# plugin owns and refreshes.
#
# The -D below is double-quoted in ExecStart on purpose. The fork trims each element, so it
# accepts "legacy18, legacy112", but systemd splits ExecStart on whitespace before java ever
# sees it -- unquoted, that spelling would hand java a stray "legacy112" argument and the unit
# would not start. Quoting keeps the whole property one argv item.
local legacy_forwarding_servers="${FELIS_LEGACY_FORWARDING_SERVERS}"
# -Xms stays at 512M so a small proxy does not reserve its whole ceiling up front, unless
# the ceiling itself is lower (the JVM refuses an initial heap above the maximum).
local xmx="$FELIS_VELOCITY_XMX" xms="512M"
[ "$(heap_megabytes "$xmx")" -ge 512 ] || xms="$xmx"
cat > "$VELOCITY_SERVICE" <<EOF
[Unit]
Description=Felis Velocity proxy (Mojang authentication + modern forwarding)
After=network-online.target k3s.service
Wants=network-online.target
[Service]
Type=simple
User=${VELOCITY_USER}
Group=${VELOCITY_USER}
WorkingDirectory=${VELOCITY_DIR}
ExecStart=${JRE_DIR}/bin/java -Xms${xms} -Xmx${xmx} -XX:+UseG1GC -XX:+ParallelRefProcEnabled -XX:+AlwaysPreTouch -Dmojang.sessionserver=http://${api_ip}:8081/session/minecraft/hasJoined "-Dfelis.legacy-forwarding.servers=${legacy_forwarding_servers}" -jar ${VELOCITY_DIR}/velocity.jar
Restart=on-failure
RestartSec=5
NoNewPrivileges=yes
ProtectSystem=strict
ProtectHome=yes
PrivateTmp=yes
ReadWritePaths=${VELOCITY_DIR}
[Install]
WantedBy=multi-user.target
EOF
systemctl daemon-reload
systemctl enable felis-velocity
# A restart disconnects every player on the network, so a rerun that changed nothing the
# proxy runs leaves it up. Otherwise restart, not `enable --now`: on a re-run the old proxy
# is already up and --now would leave it running against the new config. The fingerprint is
# recorded only after the restart, so a run that died between writing and restarting still
# restarts next time.
local fp
fp="$(velocity_fingerprint)"
if systemctl is-active --quiet felis-velocity \
&& [ "$fp" = "$(cat "$VELOCITY_FINGERPRINT" 2>/dev/null || true)" ]; then
ok "felis-velocity unchanged; left running (0.0.0.0:${FELIS_GAME_PORT})"
return 0
fi
systemctl restart felis-velocity
printf '%s\n' "$fp" > "$VELOCITY_FINGERPRINT"
ok "felis-velocity.service enabled and started (0.0.0.0:${FELIS_GAME_PORT})"
}
# velocity_fingerprint hashes what the proxy process runs: its unit (JVM flags and system
# properties), the JRE, the jars and the files the installer writes for it. The Via config
# and whatever else plugins write at runtime stay out; Via rewrites its config on every load.
velocity_fingerprint() {
local f
for f in "$VELOCITY_SERVICE" "${JRE_DIR}/release" "${VELOCITY_DIR}/velocity.jar" \
"${VELOCITY_DIR}/velocity.toml" "${VELOCITY_DIR}/forwarding.secret" \
"${VELOCITY_DIR}/plugins/felis-link/felis-link.properties" "${VELOCITY_DIR}"/plugins/*.jar; do
[ -f "$f" ] || continue
printf '%s %s\n' "$(sha256sum <"$f" | cut -d' ' -f1)" "$f"
done | sha256sum | cut -d' ' -f1
}
configure_velocity_firewall() {
command -v firewall-cmd >/dev/null 2>&1 || return 0
systemctl is-active --quiet firewalld || return 0
log "opening firewalld port ${FELIS_GAME_PORT}/tcp for the Minecraft proxy"
firewall-cmd --permanent --add-port="${FELIS_GAME_PORT}/tcp"
firewall-cmd --reload
}
# ---------------------------------------------------------------------------
# 6. PostgreSQL on the host. felis-api pods reach it at <node-ip>:5432;
# migrations run from the host binary against 127.0.0.1.
# ---------------------------------------------------------------------------
write_pg_hba_block() {
local hba="$1" tmp tmp_new node_cidr
node_cidr="${NODE_IP}/32"
tmp="$(mktemp)"
tmp_new="${tmp}.new"
remember_temp "$tmp"
remember_temp "$tmp_new"
awk \
-v db="$DB_NAME" \
-v user="$DB_USER" \
-v pod="$POD_CIDR" \
-v node="$node_cidr" '
$0 == "# BEGIN FELIS MANAGED HBA" { skip = 1; next }
$0 == "# END FELIS MANAGED HBA" { skip = 0; next }
skip { next }
# Clean up rules appended by older bootstrap versions.
$1 == "host" && $2 == db && $3 == user && $5 == "scram-sha-256" &&
($4 == "127.0.0.1/32" || $4 == pod || $4 == node) { next }
{ print }
' "$hba" > "$tmp"
{
printf "# BEGIN FELIS MANAGED HBA\n"
printf "# Felis rules must precede distro defaults such as 127.0.0.1 ident.\n"
printf "host %s %s 127.0.0.1/32 scram-sha-256\n" "$DB_NAME" "$DB_USER"
printf "host %s %s %s scram-sha-256\n" "$DB_NAME" "$DB_USER" "$POD_CIDR"
printf "host %s %s %s scram-sha-256\n" "$DB_NAME" "$DB_USER" "$node_cidr"
printf "# END FELIS MANAGED HBA\n"
printf "\n"
cat "$tmp"
} > "$tmp_new"
cat "$tmp_new" > "$hba"
rm -f "$tmp" "$tmp_new"
}
postgres_data_dir() {
case "$PKG" in
pacman) printf '%s\n' /var/lib/postgres/data ;;
*) printf '%s\n' /var/lib/pgsql/data ;;
esac
}
init_postgres_data_dir() {
local data_dir
data_dir="$(postgres_data_dir)"
[ "$PKG" = "apt" ] && return 0
[ -f "${data_dir}/PG_VERSION" ] && return 0
log "initialising postgresql data directory at ${data_dir}"
if command -v postgresql-setup >/dev/null 2>&1; then
postgresql-setup --initdb || /usr/bin/postgresql-setup initdb
elif command -v initdb >/dev/null 2>&1; then
install -d -o postgres -g postgres -m 0700 "$data_dir"
as_postgres initdb -D "$data_dir"
else
die "cannot initialise postgresql data directory: postgresql-setup/initdb not found"
fi
}
# postgres_installed reports whether a PostgreSQL client and server unit are present.
# The unit list is read into a variable before matching: `systemctl list-unit-files |
# grep -q` dies of SIGPIPE under pipefail once grep stops reading a list longer than
# one write, and the install then took the "not installed" branch on a host that has it.
postgres_installed() {
local units
command -v psql >/dev/null 2>&1 || return 1
units="$(systemctl list-unit-files 2>/dev/null)" || return 1
grep -q '^postgresql' <<<"$units"
}
install_postgres() {
if postgres_installed; then
ok "postgresql already installed"
else
log "installing postgresql"
case "$PKG" in
apt) pkg_install postgresql ;;
dnf) pkg_install postgresql-server postgresql ;;
yum) pkg_install postgresql-server postgresql ;;
zypper) pkg_install postgresql-server postgresql ;;
pacman) pkg_install postgresql ;;
esac
fi
init_postgres_data_dir
check_postgres_major
systemctl enable --now postgresql
ok "postgresql running"
}
# check_postgres_major refuses to start a PostgreSQL server whose major version differs from
# the one that created the data directory. The server would not start anyway; this says why
# and what to do, instead of a failed unit in the middle of the install. Debian and Ubuntu
# keep one cluster per version under /var/lib/postgresql/<major> and upgrade with
# pg_upgradecluster, so they are left to their own tooling.
check_postgres_major() {
local data_dir have want
[ "$PKG" = "apt" ] && return 0
data_dir="$(postgres_data_dir)"
[ -f "${data_dir}/PG_VERSION" ] || return 0
have="$(tr -d '[:space:]' < "${data_dir}/PG_VERSION")"
want="$( (postgres --version 2>/dev/null || psql --version 2>/dev/null) | head -n 1 \
| sed -nE 's/^[^0-9]*([0-9]+)\..*/\1/p')"
[ -n "$want" ] && [ "$have" != "$want" ] || return 0
die "PostgreSQL ${want} is installed, but ${data_dir} holds a PostgreSQL ${have} cluster.
The server cannot open it. Upgrade the cluster first (pg_upgrade, with the ${have} binaries
still installed), or reinstall PostgreSQL ${have}, then rerun the installer. Take a
database bundle before either: sudo felis db backup"
}
configure_postgres() {
local cfg hba listen
cfg="$(as_postgres psql -tAc 'SHOW config_file;' 2>/dev/null || true)"
hba="$(as_postgres psql -tAc 'SHOW hba_file;' 2>/dev/null || true)"
[ -n "$cfg" ] && [ -n "$hba" ] || die "could not query postgresql config/hba file paths"
listen="$(as_postgres psql -tAc 'SHOW listen_addresses;' 2>/dev/null || true)"
# Listen on all interfaces (applied on restart). ALTER SYSTEM is idempotent. Pods reach the
# database at the node IP, and configure_postgres_firewall keeps the port from everyone else.
#
# The connection itself is not encrypted (sslmode=disable in felis.toml). On this
# single-node shape it never leaves the host: pods reach the node IP over their veth pair
# and the host binary uses loopback, so TLS would guard a path no other machine is on.
as_postgres psql -v ON_ERROR_STOP=1 -c "ALTER SYSTEM SET listen_addresses = '*';" >/dev/null
# Allow the host loopback, the pod CIDR, and the node IP before broader distro defaults.
write_pg_hba_block "$hba"
# Role + database (idempotent), and (re)set the password to our generated one.
as_postgres psql -v ON_ERROR_STOP=1 <<SQL >/dev/null
SET password_encryption = 'scram-sha-256';
DO \$\$
BEGIN
IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = '${DB_USER}') THEN
CREATE ROLE ${DB_USER} LOGIN PASSWORD '${DB_PASSWORD}';
END IF;
END
\$\$;
ALTER ROLE ${DB_USER} WITH LOGIN PASSWORD '${DB_PASSWORD}';
SQL
if ! as_postgres psql -tAc "SELECT 1 FROM pg_database WHERE datname='${DB_NAME}'" | grep -q 1; then
as_postgres createdb -O "$DB_USER" "$DB_NAME"
fi
# listen_addresses is the only setting here that needs a restart, and a restart cuts
# every connection felis-api holds mid-transaction. pg_hba.conf and the role are live
# after a reload.
if [ "$listen" = "*" ]; then
as_postgres psql -v ON_ERROR_STOP=1 -tAc 'SELECT pg_reload_conf();' >/dev/null
else
systemctl restart postgresql
fi
configure_postgres_firewall
ok "postgresql configured (listen=*, role/db '${DB_NAME}', pg_hba opened to pods)"
}
# configure_postgres_firewall keeps 5432 to this host and its pods. PostgreSQL listens on
# every address (pods dial the node IP), and pg_hba.conf only refuses a connection after the
# server has spoken to it. firewalld's default zone does not open 5432 (configure_k3s_firewall
# trusts only the pod and service CIDRs), so that host needs nothing more; on any other host a
# small nftables table of our own drops 5432 from everywhere but loopback, the pod CIDR and the
# node's own address, loaded at boot by a oneshot unit ordered before PostgreSQL.
configure_postgres_firewall() {
if command -v firewall-cmd >/dev/null 2>&1 && systemctl is-active --quiet firewalld; then
return 0
fi
command -v nft >/dev/null 2>&1 || pkg_install nftables
command -v nft >/dev/null 2>&1 || { warn "nft is not available; PostgreSQL's 5432 is reachable from the network (pg_hba still refuses other hosts)"; return 0; }
local node_rule
case "$NODE_IP" in
*:*) node_rule="ip6 saddr ${NODE_IP}" ;;
*) node_rule="ip saddr ${NODE_IP}" ;;
esac
install -d -m 0700 "$STATE_DIR"
cat > "$PG_FIREWALL_RULES" <<EOF
# Generated by deploy/bootstrap.sh — do not edit by hand; rerun the installer.
# The first two lines make a reload replace the table instead of failing on it.
table inet felis_postgres
delete table inet felis_postgres
table inet felis_postgres {
chain input {
type filter hook input priority filter; policy accept;
tcp dport 5432 iif "lo" accept
tcp dport 5432 ip saddr ${POD_CIDR} accept
tcp dport 5432 ${node_rule} accept
tcp dport 5432 drop
}
}
EOF
chmod 0600 "$PG_FIREWALL_RULES"
cat > "$PG_FIREWALL_SERVICE" <<EOF
[Unit]
Description=Felis: keep PostgreSQL's port to this host and its pods
DefaultDependencies=no
After=nftables.service
# nftables.service starts from a flushed ruleset; a restart of it reloads this table too.
PartOf=nftables.service
Before=network-pre.target postgresql.service
Wants=network-pre.target
[Service]
Type=oneshot
RemainAfterExit=yes
ExecStart=$(command -v nft) -f ${PG_FIREWALL_RULES}
ExecStop=$(command -v nft) delete table inet felis_postgres
[Install]
WantedBy=multi-user.target
EOF
systemctl daemon-reload
systemctl enable felis-postgres-firewall.service >/dev/null 2>&1
systemctl restart felis-postgres-firewall.service
ok "5432 accepts loopback, ${POD_CIDR} and ${NODE_IP} only (nftables table felis_postgres)"
}
# ---------------------------------------------------------------------------
# 7. Secrets + felis.toml (pod variant reaches Postgres at the node IP; host
# variant at 127.0.0.1 for migrations)
# ---------------------------------------------------------------------------
load_or_make_secrets() {
mkdir -p "$STATE_DIR"
chmod 0700 "$STATE_DIR"
if [ -f "$SECRETS_ENV" ]; then
# shellcheck disable=SC1090
. "$SECRETS_ENV"
ok "reusing persisted secrets from ${SECRETS_ENV}"
fi
DB_PASSWORD="${DB_PASSWORD:-$(openssl rand -hex 24)}"
# One felis-api internal token per caller (naming.CallerTokens), so each is scoped
# to its own routes and a leak is contained to that caller: SERVICE_TOKEN is the
# proxy's (felis-link.properties), LIMBO_TOKEN the login gate's, BUILD_TOKEN what a
# build Job fetches its context with, OPS_TOKEN what `felis backup-now` presents.
# `felis rotate-token <caller>` rewrites the matching line here.
SERVICE_TOKEN="${SERVICE_TOKEN:-$(openssl rand -hex 32)}"
LIMBO_TOKEN="${LIMBO_TOKEN:-$(openssl rand -hex 32)}"
BUILD_TOKEN="${BUILD_TOKEN:-$(openssl rand -hex 32)}"
OPS_TOKEN="${OPS_TOKEN:-$(openssl rand -hex 32)}"
SESSION_SECRET="${SESSION_SECRET:-$(openssl rand -hex 32)}"
# The Velocity modern-forwarding key. It is what makes a backend's UUID trustworthy:
# the proxy does the Mojang handshake and HMACs the resulting profile with this key,
# and a backend that cannot verify it would fall back to an offline UUID derived from
# the username — i.e. anyone could join as anyone, the Owner included. Same value on
# the proxy (forwarding.secret) and in every backend pod (felis-forwarding-secret).
FORWARDING_SECRET="${FORWARDING_SECRET:-$(openssl rand -hex 32)}"
# Registry write credentials, one per principal the registry gate knows
# (internal/registrygate): platform pushes the installer's own images and the
# Trivy DB mirrors, build is what a build Job's push container presents and may
# never write under felis/ or mirror/, prune is felis-api deleting manifests
# nothing references (internal/registryprune). Reads stay anonymous.
REGISTRY_PLATFORM_TOKEN="${REGISTRY_PLATFORM_TOKEN:-$(openssl rand -hex 32)}"
REGISTRY_BUILD_TOKEN="${REGISTRY_BUILD_TOKEN:-$(openssl rand -hex 32)}"
REGISTRY_PRUNE_TOKEN="${REGISTRY_PRUNE_TOKEN:-$(openssl rand -hex 32)}"
(
umask 077
cat > "$SECRETS_ENV" <<EOF
DB_PASSWORD=${DB_PASSWORD}
SERVICE_TOKEN=${SERVICE_TOKEN}
LIMBO_TOKEN=${LIMBO_TOKEN}
BUILD_TOKEN=${BUILD_TOKEN}
OPS_TOKEN=${OPS_TOKEN}
SESSION_SECRET=${SESSION_SECRET}
FORWARDING_SECRET=${FORWARDING_SECRET}
REGISTRY_PLATFORM_TOKEN=${REGISTRY_PLATFORM_TOKEN}
REGISTRY_BUILD_TOKEN=${REGISTRY_BUILD_TOKEN}
REGISTRY_PRUNE_TOKEN=${REGISTRY_PRUNE_TOKEN}
EOF
)
chmod 0600 "$SECRETS_ENV"
}
ensure_panel_tls_cert() {
mkdir -p "$STATE_DIR"
chmod 0700 "$STATE_DIR"
if [ -s "$PANEL_TLS_CERT" ] && [ -s "$PANEL_TLS_KEY" ]; then
ok "panel TLS certificate already present"
return 0
fi
local cn conf
cn="op.console.${FELIS_ROOT_DOMAIN}"
conf="$(mktemp)"
remember_temp "$conf"
cat > "$conf" <<EOF
[req]
default_bits = 2048
distinguished_name = dn
x509_extensions = v3_req
prompt = no
[dn]
CN = ${cn}
[v3_req]
subjectAltName = @alt_names
[alt_names]
DNS.1 = op.console.${FELIS_ROOT_DOMAIN}
DNS.2 = console.${FELIS_ROOT_DOMAIN}
DNS.3 = localhost
IP.1 = 127.0.0.1
IP.2 = ${NODE_IP}
EOF
log "generating self-signed panel TLS certificate"
openssl req -x509 -newkey rsa:2048 -sha256 -days 825 -nodes \
-keyout "$PANEL_TLS_KEY" \
-out "$PANEL_TLS_CERT" \
-subj "/CN=${cn}" \
-config "$conf" >/dev/null 2>&1
chmod 0600 "$PANEL_TLS_KEY"
chmod 0644 "$PANEL_TLS_CERT"
ok "panel TLS certificate ready (${PANEL_TLS_CERT})"
}
# persisted_smtp_block echoes the [smtp] section an earlier run left behind, or
# nothing. Unlike every other value in the generated toml, [smtp] is not derived
# from this script's inputs -- `felis setup`'s SMTP screen writes it, after
# proving the relay works. A wholesale `cat >` therefore erases it on every
# re-run, and since re-running the installer is the documented way to update
# felis-api, an operator who updates loses mail: OTP delivery silently reverts
# to the no-Mailer path and every code is logged instead of sent. Same defect
# family as the root_domain loss fixed in ecbeb20 -- generated file, hand-set
# value, no carry-forward.
#
# The carry is the section's header and key lines only. Printing every line up to
# the next section header hoarded the generated [[auth_source]] comment block
# that sits below [smtp] into this carry: each re-run then re-emitted the hoard
# plus a fresh template copy, growing both config files by one comment block per
# run (audit #50). The extraction is idempotent, which is also why re-reading the
# freshly rewritten host file on the pod pass is safe. The pod toml remains the
# fallback for a host file with no [smtp] section at all.
persisted_smtp_block() {
local f out
for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
[ -r "$f" ] || continue
out="$(awk '
/^[[:space:]]*\[/ {
if (insmtp) exit
insmtp = ($0 ~ /^[[:space:]]*\[smtp\][[:space:]]*$/)
if (insmtp) print
next
}
insmtp && /^[[:space:]]*("[A-Za-z_][A-Za-z0-9_]*"|[A-Za-z_][A-Za-z0-9_]*)[[:space:]]*=/ { print }
' "$f")"
[ -n "$out" ] || continue
printf '%s' "$out"
return 0
done
}
# persisted_auth_source_blocks echoes the [[auth_source]] tables an earlier run left
# behind, or the LittleSkin default when there is no earlier felis.toml at all. The
# list is the operator's: it is the only way to add or drop a Yggdrasil root on a full
# install, and nothing in this script's inputs derives it. Without the carry-forward a
# re-run would put LittleSkin back after the operator removed it and silently drop any
# root they added. An earlier file with no tables stays that way — that is a Mojang-only
# server, not a missing value. Same first-readable-file rule as persisted_smtp_block.
persisted_auth_source_blocks() {
local f
for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
[ -r "$f" ] || continue
# Every [[auth_source]] table, up to (not including) the next other section header.
# TOML also accepts [[ auth_source ]] and a quoted key; a header this does not
# recognise would silently drop that table.
awk '/^[[:space:]]*\[/ { f = /^[[:space:]]*\[\[[[:space:]]*["\047]?auth_source["\047]?[[:space:]]*\]\]/ }
f { print }' "$f"
return 0
done
printf '%s\n' '[[auth_source]]' 'tag = "littleskin"' 'prefix = "LS"' \
'url = "https://littleskin.cn/api/yggdrasil/sessionserver/session/minecraft/hasJoined"'
}
# persisted_archive_block echoes the operator-owned [archive] keys an earlier run
# left behind — the retention window, the pre-reap warn offsets, the local cap,
# the on-demand backup retention/count/cooldown — so a re-run does not silently
# revert them to the built-ins (felis reaper, felis backup and felis api read
# these from the config Secret; defaults: 90d retention, 3d/1d warnings, no cap,
# manual backups kept 30d, 5 per server, one per 10m). store and local_path are
# NOT carried: this script owns them (FELIS_ARCHIVE_LOCAL_PATH must equal the
# mount). Same first-readable-file rule as persisted_smtp_block; warn_before must
# be a single-line TOML array (the shape every writer here emits).
persisted_archive_block() {
local f out
for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
[ -r "$f" ] || continue
out="$(awk '
/^[[:space:]]*\[/ { sect = $0; next }
sect ~ /^[[:space:]]*\[archive\][[:space:]]*$/ &&
/^[[:space:]]*(retention|warn_before|max_local_bytes|manual_retention|manual_keep|manual_cooldown)[[:space:]]*=/ { print }
' "$f")"
[ -n "$out" ] || continue
printf '%s\n' "$out"
return 0
done
}
# persisted_registry_block echoes the operator-owned [registry] keys an earlier run
# left behind — the build-lane executor mirrors, the resource caps, the uploads
# backend and its [registry.s3] subtable — so §15's upgrade path (re-run the
# installer) does not silently revert them. Nothing in this script's inputs
# derives these: they are hand-written per docs/troubleshooting.md §8e or stamped
# by the storage wizard. url and build_namespace are NOT carried: this script
# owns them (they must match REGISTRY_URL / BUILD_NS). Same first-readable-file
# rule as persisted_smtp_block.
persisted_registry_block() {
local f out
for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
[ -r "$f" ] || continue
out="$(awk '
/^[[:space:]]*\[/ { sect = $0; next }
sect ~ /^[[:space:]]*\[registry\][[:space:]]*$/ &&
/^[[:space:]]*(kaniko_image|trivy_image|trivy_db_repository|trivy_java_db_repository|build_cpu_limit|build_mem_limit|build_disk_limit|build_user_namespaces|build_runtime_class|max_concurrent_builds|user_uploads_context|user_uploads_max_bytes|context_max_bytes)[[:space:]]*=/ { print }
sect ~ /^[[:space:]]*\[registry\.s3\][[:space:]]*$/ && /^[[:space:]]*[A-Za-z_]+[[:space:]]*=/ {
if (!s3hdr) { printf "[registry.s3]\n"; s3hdr = 1 }
print
}
' "$f")"
[ -n "$out" ] || continue
printf '%s\n' "$out"
return 0
done
}
# persisted_offsite_block echoes the [offsite] section an earlier run (or the operator)
# left behind: after the first install the bucket is configured by editing felis.host.toml
# or by re-running with FELIS_OFFSITE_*, and a plain re-run must keep it. Deleting the
# section and re-running is how the off-site copy is turned off. Same first-readable-file
# rule as persisted_smtp_block.
persisted_offsite_block() {
local f out
for f in "${STATE_DIR}/felis.host.toml" "${STATE_DIR}/felis.pod.toml"; do
[ -r "$f" ] || continue
out="$(awk '
/^[[:space:]]*\[/ {
if (inoff) exit
inoff = ($0 ~ /^[[:space:]]*\[offsite\][[:space:]]*$/)
if (inoff) print
next
}
inoff && /^[[:space:]]*(endpoint|region|bucket|prefix|access_key_ref|secret_key_ref|key_ref|db_keep)[[:space:]]*=/ { print }
' "$f")"
[ -n "$out" ] || continue
printf '%s\n' "$out"
return 0
done
}
# offsite_block is the [offsite] section this run writes: from FELIS_OFFSITE_* when the
# bucket is given, else the one an earlier run left.
offsite_block() {
if [ -z "$FELIS_OFFSITE_BUCKET" ]; then
persisted_offsite_block
return 0
fi
printf '[offsite]\n'
printf 'endpoint = "%s"\n' "$FELIS_OFFSITE_ENDPOINT"
printf 'bucket = "%s"\n' "$FELIS_OFFSITE_BUCKET"
if [ -n "$FELIS_OFFSITE_REGION" ]; then printf 'region = "%s"\n' "$FELIS_OFFSITE_REGION"; fi
if [ -n "$FELIS_OFFSITE_PREFIX" ]; then printf 'prefix = "%s"\n' "$FELIS_OFFSITE_PREFIX"; fi
if [ -n "$FELIS_OFFSITE_DB_KEEP" ]; then printf 'db_keep = %s\n' "$FELIS_OFFSITE_DB_KEEP"; fi
}
# offsite_enabled: the [offsite] section this run writes names a bucket.
offsite_enabled() {
offsite_block | grep -Eq '^[[:space:]]*bucket[[:space:]]*=[[:space:]]*"[^"]+"'
}
write_felis_toml() {
local target="$1" db_host="$2" smtp_block auth_source_blocks registry_block archive_block offsite_section
smtp_block="$(persisted_smtp_block)"
if [ -n "$smtp_block" ]; then
log "carrying forward the configured [smtp] relay"
smtp_block="${smtp_block}"$'\n' # keep a blank line before the next section
fi
auth_source_blocks="$(persisted_auth_source_blocks)"
registry_block="$(persisted_registry_block)"
if [ -n "$registry_block" ]; then
log "carrying forward the configured [registry] overrides"
registry_block="${registry_block}"$'\n' # keep a blank line before the next section
fi
archive_block="$(persisted_archive_block)"
if [ -n "$archive_block" ]; then
log "carrying forward the configured [archive] overrides"
archive_block="${archive_block}"$'\n' # keep a blank line before the next section
fi
# The pod copy needs it as much as the host one: the reaper reads [offsite] from the
# config Secret to know it must wait for each archive's off-site copy.
offsite_section="$(offsite_block)"
if [ -n "$offsite_section" ]; then
offsite_section="${offsite_section}"$'\n\n' # keep a blank line before the next section
fi
cat > "$target" <<EOF
# Generated by deploy/bootstrap.sh; rerun the installer to regenerate. Hand edits are
# overwritten, except [smtp], [[auth_source]], [offsite], and the operator-owned
# [registry] / [archive] overrides, which carry forward.
[server]
listen = "0.0.0.0:8080"
root_domain = "${FELIS_ROOT_DOMAIN}"
[database]
url = "postgres://${DB_USER}:${DB_PASSWORD}@${db_host}:5432/${DB_NAME}?sslmode=disable"
[k8s]
namespace = "${MINECRAFT_NS}"
egress_mode = "${FELIS_EGRESS_MODE}"
[velocity]
# The two always-on system servers that felis setup provisions. They are built and imported
# into k3s by build_game_stack below, so setup never has to be told "build these first".
login_image = "${FELIS_LIMBO_IMAGE}"
lobby_image = "${FELIS_LOBBY_IMAGE}"
# The public port players connect on; the panel shows it in server addresses.
game_port = ${FELIS_GAME_PORT}
[registry]
url = "${REGISTRY_URL}"
build_namespace = "${BUILD_NS}"
${registry_block}
[archive]
store = "tarLocal"
local_path = "${FELIS_ARCHIVE_LOCAL_PATH}"
${archive_block}
${offsite_section}[auth]
admin_hostname = "op.console.${FELIS_ROOT_DOMAIN}"
panel_hostname = "console.${FELIS_ROOT_DOMAIN}"
${smtp_block}
# Third-party Yggdrasil sources federated by the hasJoined multiplexer. Mojang is
# always the code-owned identity anchor (premium-first), prepended in Go; sources here
# append as namespace-rewritten guests. A fresh install federates LittleSkin. Edit the
# list in ${STATE_DIR}/felis.host.toml and rerun the installer; re-runs keep it as it
# is, and with no [[auth_source]] at all the server is Mojang-only.
# Order is trust: the first source that answers 200 wins, so list the most trusted roots
# first, and remove a compromised root rather than just moving it down.
# A tag is permanent: it is hashed into every player UUID of its source, so changing it
# (even its case) gives all of them new UUIDs and orphans their data, links and bans.
${auth_source_blocks}
EOF
}
ensure_default_config() {
local target="${STATE_DIR}/felis.toml" backup
if [ -L "$target" ] && [ "$(readlink "$target")" = "${STATE_DIR}/felis.host.toml" ]; then
ok "default host config already points at ${STATE_DIR}/felis.host.toml"
return 0
fi
if [ -e "$target" ] || [ -L "$target" ]; then
if bootstrap_from_tui; then
backup="${target}.bak.$(date -u +%Y%m%d%H%M%S).$$"
warn "replacing existing ${target}; backup saved at ${backup}"
mv "$target" "$backup"
ln -s "${STATE_DIR}/felis.host.toml" "$target"
ok "default host config: ${target} -> ${STATE_DIR}/felis.host.toml"
return 0
fi
warn "leaving existing ${target}; setup can use -config ${STATE_DIR}/felis.host.toml if needed"
return 0
fi
ln -s "${STATE_DIR}/felis.host.toml" "$target"
ok "default host config: ${target} -> ${STATE_DIR}/felis.host.toml"
}
# ---------------------------------------------------------------------------
# 8. Migrate + deploy bundle
# ---------------------------------------------------------------------------
run_migrations() {
local backup_flags=(-backup-dir "$FELIS_DB_BACKUP_DIR")
write_felis_toml "${STATE_DIR}/felis.host.toml" "127.0.0.1"
ensure_default_config
if [ "$FELIS_PRE_MIGRATE_BACKUP" = 0 ]; then
warn "FELIS_PRE_MIGRATE_BACKUP=0: pending migrations run without a database snapshot"
backup_flags=(-no-backup)
fi
# Migrations only roll forward. On an existing database with migrations pending, the
# binary bundles the database into FELIS_DB_BACKUP_DIR first and refuses to migrate
# if that fails; a fresh database has nothing to protect and is migrated directly.
log "running database migrations (host binary -> 127.0.0.1)"
# From here the database may move forward, and the binary that moved it stays.
HOST_BIN_IN_USE=1
"$HOST_BIN" migrate up -config "${STATE_DIR}/felis.host.toml" "${backup_flags[@]}"
ok "migrations applied"
}
# Silence felis watchdog's mail for the rest of this install (see WATCHDOG_QUIET_FILE).
# Two hours covers a slow source build; cleanup lifts it as soon as the installer exits.
quiet_watchdog() {
install -d -m 0755 "$(dirname "$WATCHDOG_QUIET_FILE")"
printf '%s\n' "$(( $(date +%s) + 7200 ))" > "$WATCHDOG_QUIET_FILE"
}
# The hourly off-site copy. The bucket is checked now, in the install, so wrong
# credentials or an unreachable endpoint show up here; the first copy itself runs in the
# background, since a host with many archives can take a long while to upload them.
# With no bucket configured the timer is removed (the operator deleted [offsite]) and the
# install says loudly that every backup is on this machine only.
install_offsite_timer() {
if [ "${OFFSITE_ENABLED:-0}" != 1 ]; then
if [ -e "$OFFSITE_TIMER" ] || [ -e "$OFFSITE_SERVICE" ]; then
systemctl disable --now felis-offsite.timer >/dev/null 2>&1 || true
rm -f "$OFFSITE_TIMER" "$OFFSITE_SERVICE"
systemctl daemon-reload
warn "no [offsite] bucket is configured any more; the off-site copy timer was removed"
fi
return 0
fi
cat > "$OFFSITE_SERVICE" <<EOF
[Unit]
Description=Felis off-site copy (world archives, database bundles, user registry images and submission uploads, encrypted, to the [offsite] bucket)
After=network-online.target k3s.service postgresql.service felis-db-backup.service
Wants=network-online.target
[Service]
Type=oneshot
EnvironmentFile=${OFFSITE_ENV}
ExecStart=${HOST_BIN} offsite sync -config ${STATE_DIR}/felis.host.toml -env-file ${OFFSITE_ENV} -db-dir ${FELIS_DB_BACKUP_DIR} -backup-pvc "${FELIS_BACKUP_PVC}"
TimeoutStartSec=55min
Nice=10
IOSchedulingClass=idle
PrivateTmp=yes
NoNewPrivileges=yes
EOF
cat > "$OFFSITE_TIMER" <<EOF
[Unit]
Description=Hourly Felis off-site copy
[Timer]
OnCalendar=hourly
RandomizedDelaySec=10min
Persistent=true
[Install]
WantedBy=timers.target
EOF
systemctl daemon-reload
systemctl enable --now felis-offsite.timer
if "$HOST_BIN" offsite list -config "${STATE_DIR}/felis.host.toml" -env-file "$OFFSITE_ENV" >/dev/null; then
systemctl start --no-block felis-offsite.service
ok "off-site copy: hourly to the [offsite] bucket, first copy started (sudo felis offsite status; journalctl -u felis-offsite)"
else
warn "the [offsite] bucket did not answer (error above); nothing is copied off this machine until it does: fix ${OFFSITE_ENV} or [offsite] in ${STATE_DIR}/felis.host.toml, then sudo systemctl start felis-offsite.service"
fi
}
# summary_offsite is the installer's last word on where the backups live.
summary_offsite() {
if [ "${OFFSITE_ENABLED:-0}" != 1 ]; then
warn "NO OFF-SITE COPY: every world archive and database backup is on this machine only."
warn "Losing its disk loses them all. Set FELIS_OFFSITE_BUCKET, FELIS_OFFSITE_ENDPOINT,"
warn "FELIS_OFFSITE_ACCESS_KEY and FELIS_OFFSITE_SECRET_KEY and re-run (docs/troubleshooting.md §16)."
return 0
fi
if [ "${OFFSITE_KEY_NEW:-0}" = 1 ]; then
echo
warn "================================================================================"
warn "The off-site copies are encrypted with this key. Store it NOW somewhere other than"
warn "this machine (a password manager): without it nothing in the bucket can be read."
warn ""
warn " FELIS_OFFSITE_KEY=${FELIS_OFFSITE_KEY}"
warn ""
warn "It is also in ${OFFSITE_ENV}, which is lost with this machine."
warn "================================================================================"
else
log "Off-site copy: the encryption key is in ${OFFSITE_ENV}; keep a copy of it off this machine."
fi
}
# The platform watchdog: every two minutes it checks the control plane, the login gate,
# the fleet, PostgreSQL, the game proxy, the database backups and the host's disks and
# memory, and mails the owners (their verified addresses, over the [smtp] relay) what
# has stayed wrong long enough to matter. It runs on the host so a k3s that is down is
# still reported. The first run happens now, so a broken unit shows up in this install.
install_watchdog_timer() {
local disks="/,/var/lib/rancher/k3s,/var/lib/postgresql,/var/lib/felis" path
for path in "$FELIS_WORLDS_HOST_PATH" "$FELIS_ARCHIVE_LOCAL_PATH" "$FELIS_DB_BACKUP_DIR"; do
if [ -n "$path" ]; then disks="${disks},${path}"; fi
done
install -d -m 0700 "$(dirname "$WATCHDOG_STATE")"
cat > "$WATCHDOG_SERVICE" <<EOF
[Unit]
Description=Felis platform watchdog (health checks, owner alert mail)
After=network-online.target k3s.service postgresql.service
[Service]
Type=oneshot
ExecStart=${HOST_BIN} watchdog -config ${STATE_DIR}/felis.host.toml -state ${WATCHDOG_STATE} -quiet-file ${WATCHDOG_QUIET_FILE} -backup-dir ${FELIS_DB_BACKUP_DIR} -proxy-addr 127.0.0.1:${FELIS_GAME_PORT} -disk-paths ${disks}
TimeoutStartSec=3min
Nice=5
PrivateTmp=yes
NoNewPrivileges=yes
ProtectSystem=full
EOF
cat > "$WATCHDOG_TIMER" <<EOF
[Unit]
Description=Felis platform watchdog, every two minutes
[Timer]
OnBootSec=3min
OnUnitActiveSec=2min
AccuracySec=15s
[Install]
WantedBy=timers.target
EOF
systemctl daemon-reload
systemctl enable --now felis-watchdog.timer
if systemctl start felis-watchdog.service; then
ok "watchdog: checks every 2 minutes and mails the owners' verified addresses (journalctl -u felis-watchdog)"
else
journalctl -u felis-watchdog.service -n 20 --no-pager >&2 || true
warn "the first watchdog run failed (log above); nothing will be mailed until it runs: sudo systemctl start felis-watchdog.service"
fi
}
# The build lane's tools: kaniko and trivy (pinned by digest in internal/build/tools.go)
# and Trivy's vulnerability and Java DBs, copied into the registry's mirror/ where build
# Jobs pull them; the build namespace has no internet egress. The timer refreshes the DBs
# twice a day (upstream publishes every six hours) and the watchdog warns when three days
# pass without a clean run. The first copy starts now in the background: the Java DB
# alone is several hundred MB.
install_build_tools_timer() {
install -d -m 0755 "$(dirname "$BUILD_TOOLS_STATUS")"
cat > "$BUILD_TOOLS_SERVICE" <<EOF
[Unit]
Description=Copy the Felis build tools and Trivy's vulnerability DBs into the registry
After=network-online.target k3s.service
Wants=network-online.target
[Service]
Type=oneshot
ExecStart=${HOST_BIN} mirror-build-tools -endpoint ${REGISTRY_PUSH_HOST} -status ${BUILD_TOOLS_STATUS} -secrets-env ${SECRETS_ENV}
TimeoutStartSec=1h
Nice=10
PrivateTmp=yes
NoNewPrivileges=yes
ProtectSystem=full
EOF
cat > "$BUILD_TOOLS_TIMER" <<EOF
[Unit]
Description=Refresh the Felis build tools and Trivy DBs twice a day
[Timer]
OnCalendar=*-*-* 04,16:00:00
RandomizedDelaySec=1h
Persistent=true
[Install]
WantedBy=timers.target
EOF
systemctl daemon-reload
systemctl enable --now felis-build-tools.timer
systemctl start --no-block felis-build-tools.service
ok "build tools: kaniko, trivy and the Trivy DBs are being copied into the registry (journalctl -u felis-build-tools); refreshed twice a day"
}
# The daily database backup. The first run happens now, so a broken pipeline (pg_dump
# missing, directory unwritable) shows up in this install rather than in the first
# restore someone needs.
install_db_backup_timer() {
install -d -m 0700 "$FELIS_DB_BACKUP_DIR"
cat > "$DB_BACKUP_SERVICE" <<EOF
[Unit]
Description=Felis control-plane database backup (pg_dump + /etc/felis state)
After=postgresql.service k3s.service
Wants=postgresql.service
[Service]
Type=oneshot
ExecStart=${HOST_BIN} db backup -config ${STATE_DIR}/felis.host.toml -dir ${FELIS_DB_BACKUP_DIR} -label daily -keep ${FELIS_DB_BACKUP_KEEP} -metrics-file ${FELIS_DB_BACKUP_METRICS}
Nice=10
IOSchedulingClass=idle
PrivateTmp=yes
NoNewPrivileges=yes
EOF
cat > "$DB_BACKUP_TIMER" <<EOF
[Unit]
Description=Daily Felis control-plane database backup
[Timer]
OnCalendar=${FELIS_DB_BACKUP_TIME}
RandomizedDelaySec=15min
Persistent=true
[Install]
WantedBy=timers.target
EOF
systemctl daemon-reload
systemctl enable --now felis-db-backup.timer
if systemctl start felis-db-backup.service; then
ok "database backups: daily at ${FELIS_DB_BACKUP_TIME}, newest ${FELIS_DB_BACKUP_KEEP} kept in ${FELIS_DB_BACKUP_DIR} (first one taken now)"
else
journalctl -u felis-db-backup.service -n 20 --no-pager >&2 || true
warn "the first database backup failed (log above); fix it before relying on the daily timer: sudo systemctl start felis-db-backup.service"
fi
}
# configure_offsite writes OFFSITE_ENV, the secrets behind [offsite]: the bucket's access
# keys and the key every off-site object is sealed with. Values in this run's environment
# replace the file's (rotating the bucket credentials is a re-run); the encryption key is
# generated when neither has one. A different key than the file's is refused: every object
# already in the bucket is sealed with the old one, and swapping it would make them
# unreadable without a word.
configure_offsite() {
OFFSITE_ENABLED=0
OFFSITE_KEY_NEW=0
offsite_enabled || return 0
OFFSITE_ENABLED=1
local env_ak="${FELIS_OFFSITE_ACCESS_KEY:-}" env_sk="${FELIS_OFFSITE_SECRET_KEY:-}" env_key="${FELIS_OFFSITE_KEY:-}"
local file_key=""
FELIS_OFFSITE_ACCESS_KEY="" FELIS_OFFSITE_SECRET_KEY="" FELIS_OFFSITE_KEY=""
if [ -f "$OFFSITE_ENV" ]; then
# shellcheck disable=SC1090
. "$OFFSITE_ENV"
file_key="$FELIS_OFFSITE_KEY"
fi
FELIS_OFFSITE_ACCESS_KEY="${env_ak:-$FELIS_OFFSITE_ACCESS_KEY}"
FELIS_OFFSITE_SECRET_KEY="${env_sk:-$FELIS_OFFSITE_SECRET_KEY}"
if [ -n "$env_key" ] && [ -n "$file_key" ] && [ "$env_key" != "$file_key" ]; then
die "FELIS_OFFSITE_KEY differs from the key in ${OFFSITE_ENV}; the objects already in the bucket are sealed with that one. Unset FELIS_OFFSITE_KEY to keep it (docs/troubleshooting.md §16)"
fi
FELIS_OFFSITE_KEY="${env_key:-$file_key}"
if [ -z "$FELIS_OFFSITE_ACCESS_KEY" ] || [ -z "$FELIS_OFFSITE_SECRET_KEY" ]; then
die "[offsite] names a bucket but there are no credentials for it: set FELIS_OFFSITE_ACCESS_KEY and FELIS_OFFSITE_SECRET_KEY (they are kept in ${OFFSITE_ENV})"
fi
if [ -z "$FELIS_OFFSITE_KEY" ]; then
FELIS_OFFSITE_KEY="$(openssl rand -base64 32)"
OFFSITE_KEY_NEW=1
fi
(
umask 077
cat > "$OFFSITE_ENV" <<EOF
# The off-site copy's secrets (felis offsite, docs/troubleshooting.md §16). Keep a copy of
# FELIS_OFFSITE_KEY somewhere other than this machine: without it the copies in the
# bucket cannot be read, and this file goes with the machine.
FELIS_OFFSITE_ACCESS_KEY='${FELIS_OFFSITE_ACCESS_KEY}'
FELIS_OFFSITE_SECRET_KEY='${FELIS_OFFSITE_SECRET_KEY}'
FELIS_OFFSITE_KEY='${FELIS_OFFSITE_KEY}'
EOF
)
chmod 0600 "$OFFSITE_ENV"
ok "off-site copy: secrets in ${OFFSITE_ENV}"
}
# Releases before the reaper ran as root granted uid 1000 traverse on the worlds root: an
# ACL entry, or o+x where the host had no setfacl. uid 1000 is now the game servers' uid,
# and the per-volume directories below that root are 0777, so the grant let a game process
# (and, for o+x, every local account) reach any world by its directory name. Every run
# takes it back: the ACL entry from any worlds root, the other-bits only from k3s's storage
# root, which k3s ships 0700 root:root. A custom root keeps its mode, which may be the
# operator's own.
revoke_worlds_root_grant() {
local dir
for dir in "$K3S_STORAGE_ROOT" "$FELIS_WORLDS_HOST_PATH"; do
[ -n "$dir" ] && [ -d "$dir" ] || continue
if command -v getfacl >/dev/null 2>&1 && getfacl -cpn "$dir" 2>/dev/null | grep -q '^user:1000:'; then
if setfacl -x u:1000 "$dir"; then
log "revoked the old uid-1000 traverse grant on ${dir}"
else
warn "could not revoke the old uid-1000 traverse grant on ${dir}; remove it with: setfacl -x u:1000 ${dir}"
fi
fi
done
if [ -d "$K3S_STORAGE_ROOT" ] && [ -n "$(find "$K3S_STORAGE_ROOT" -maxdepth 0 -perm -o=x)" ]; then
if chmod o-rwx "$K3S_STORAGE_ROOT"; then
log "revoked the old world-traversable mode on ${K3S_STORAGE_ROOT} (back to k3s's 0700)"
else
warn "could not restore ${K3S_STORAGE_ROOT} to 0700; any local account can reach the world volumes below it: chmod o-rwx ${K3S_STORAGE_ROOT}"
fi
fi
}
deploy_bundle() {
local prev_api prev_operator prev_gate
export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
write_felis_toml "${STATE_DIR}/felis.pod.toml" "${NODE_IP}"
prev_api="$(deployment_image felis-api api)"
prev_operator="$(deployment_image felis-operator operator)"
prev_gate="$(deployment_image registry registry-gate)"
if [ -n "$prev_api" ] && [ "$prev_api" != "$FELIS_IMAGE" ]; then
PREVIOUS_FELIS_IMAGE="$prev_api"
printf '%s\n' "$prev_api" > "${STATE_DIR}/previous-felis-image"
fi
# Always the embedded copy. It is byte-identical to deploy/crd/ (bootstrap_asset.go embeds
# that very file), it needs no checkout — which the release-download path does not have —
# and one source beats a branch whose two arms have to be kept in agreement by hand.
log "applying MinecraftServer CRD"
"$HOST_BIN" bootstrap-assets crd | kube apply -f -
log "ensuring namespaces"
local ns
for ns in "$CONTROL_NS" "$MINECRAFT_NS" "$BUILD_NS"; do
kube create namespace "$ns" --dry-run=client -o yaml | kube apply -f -
done
log "provisioning felis-config + internal caller tokens + felis-forwarding-secret + registry credentials + panel TLS secrets (out-of-band, never in the bundle)"
apply_felis_config_secrets
# felis-api mounts all four caller tokens from the control namespace. The login
# gate's and the build Job's are also applied into the namespace their pods run in
# (a secretKeyRef is namespace-local); applying them here rather than leaving it to
# `felis setup` means an upgrade has them in place before the new operator points
# the login pod at felis-limbo-token. The proxy's and the ops token stay here only.
apply_literal_secret "$CONTROL_NS" felis-service-token token "$SERVICE_TOKEN"
apply_literal_secret "$CONTROL_NS" felis-limbo-token token "$LIMBO_TOKEN"
apply_literal_secret "$MINECRAFT_NS" felis-limbo-token token "$LIMBO_TOKEN"
apply_literal_secret "$CONTROL_NS" felis-build-token token "$BUILD_TOKEN"
apply_literal_secret "$BUILD_NS" felis-build-token token "$BUILD_TOKEN"
apply_literal_secret "$CONTROL_NS" felis-ops-token token "$OPS_TOKEN"
# The forwarding key every backend verifies the proxy's handshake with. `felis setup`
# replicates it into the minecraft namespace (ensureSecretReplica) before it creates
# the pods that mount it; the operator injects it into EVERY backend, because Velocity's
# forwarding mode is one proxy-wide setting — a backend that does not speak it is not
# "less secure", it is unjoinable.
apply_literal_secret "$CONTROL_NS" felis-forwarding-secret secret "$FORWARDING_SECRET"
apply_registry_secrets
kube -n "$CONTROL_NS" create secret tls felis-api-tls \
--cert="$PANEL_TLS_CERT" \
--key="$PANEL_TLS_KEY" \
--dry-run=client -o yaml | kube apply -f -
revoke_worlds_root_grant
log "rendering + applying the control-plane bundle"
local -a manifest_args=(
--felis-image "$FELIS_IMAGE"
--panel-node-port "$FELIS_PANEL_NODEPORT"
--velocity-cidr "${NODE_IP}/32"
)
local cidr
while read -r cidr; do
[ -n "$cidr" ] && manifest_args+=(--server-egress-deny-cidr "$cidr")
done < <(node_global_cidrs)
# Backups are on by default (the renderer's own default names felis-backups); an emptied
# FELIS_BACKUP_PVC asks for the no-backup shape explicitly, and a custom name must be
# passed through or the api would advertise a PVC the bundle never created.
if [ -n "$FELIS_BACKUP_PVC" ]; then
manifest_args+=(--backup-pvc "$FELIS_BACKUP_PVC")
else
manifest_args+=(--backup-pvc=)
fi
# With an archive store the reaper always renders, since backups past their expiry have
# to leave it; it reaps idle worlds only when the operator names where the worlds live.
# The archive path must equal the [archive] local_path written above.
if [ -n "$FELIS_BACKUP_PVC" ]; then
manifest_args+=(--archive-local-path "$FELIS_ARCHIVE_LOCAL_PATH")
if [ -z "$FELIS_WORLDS_HOST_PATH" ]; then
log "idle-world retention is off: the daily reaper deletes expired backups and keeps every world (set FELIS_WORLDS_HOST_PATH=${K3S_STORAGE_ROOT} to reap worlds idle for 15 days)"
fi
fi
if [ -n "$FELIS_WORLDS_HOST_PATH" ]; then
log "retention enabled: the daily reaper will read worlds from ${FELIS_WORLDS_HOST_PATH}"
# The reaper reads this root as root with DAC_OVERRIDE (platform.reaperPodSecurityContext)
# through a static hostPath PV, so the host directory keeps k3s's own 0700 root:root and
# needs no extra grant. It must exist, though: the PV declares type Directory.
if [ ! -d "$FELIS_WORLDS_HOST_PATH" ]; then
if [ "$FELIS_WORLDS_HOST_PATH" = "$K3S_STORAGE_ROOT" ]; then
# The provisioner would create it moments later with this same 0700 root:root;
# creating it now keeps a fresh install from warning about its own default.
install -d -m 0700 -o root -g root "$FELIS_WORLDS_HOST_PATH"
else
warn "worlds root ${FELIS_WORLDS_HOST_PATH} does not exist yet; the reaper CronJob cannot start until it does (hostPath type Directory)"
fi
fi
manifest_args+=(--worlds-host-path "$FELIS_WORLDS_HOST_PATH")
fi
local size
size="$(pvc_size "$CONTROL_NS" registry "$FELIS_REGISTRY_STORAGE" FELIS_REGISTRY_STORAGE)"
if [ -n "$size" ]; then manifest_args+=(--registry-storage "$size"); fi
size="$(pvc_size "$CONTROL_NS" felis-uploads "$FELIS_UPLOADS_STORAGE" FELIS_UPLOADS_STORAGE)"
if [ -n "$size" ]; then manifest_args+=(--uploads-storage "$size"); fi
if [ -n "$FELIS_BACKUP_PVC" ]; then
size="$(pvc_size "$MINECRAFT_NS" "$FELIS_BACKUP_PVC" "$FELIS_BACKUP_STORAGE" FELIS_BACKUP_STORAGE)"
if [ -n "$size" ]; then manifest_args+=(--backup-storage "$size"); fi
fi
"$HOST_BIN" manifests "${manifest_args[@]}" | kube apply -f -
restart_existing_control_plane "$prev_api" "$prev_operator" "$prev_gate"
log "waiting for control-plane rollouts"
local d
for d in $(kube -n "$CONTROL_NS" get deploy -o name); do
if ! kube -n "$CONTROL_NS" rollout status "$d" --timeout=180s; then
diagnose_rollout "$d"
die "control-plane rollout did not complete: ${d}"
fi
done
# Before per-caller tokens the proxy's token was replicated into the workload
# namespaces for the login gate and the build Jobs. The new operator and api no
# longer reference those copies; leaving them would keep the proxy's credential
# readable from namespaces that have no business with it.
kube -n "$MINECRAFT_NS" delete secret felis-service-token --ignore-not-found
kube -n "$BUILD_NS" delete secret felis-service-token --ignore-not-found
}
# pvc_size <namespace> <claim> <wanted> <env name> prints the size to render the claim
# with: its current request when it exists, else the wanted size (empty = the renderer's
# default). A claim's request can only grow, and only on a storage class that allows
# expansion (k3s local-path does not), so re-applying a different size would fail the
# whole apply; a mismatch is reported and left to the operator.
pvc_size() {
local ns="$1" claim="$2" want="$3" env="$4" have
have="$(kube -n "$ns" get pvc "$claim" -o jsonpath='{.spec.resources.requests.storage}' 2>/dev/null || true)"
if [ -z "$have" ]; then
printf '%s' "$want"
return 0
fi
if [ -n "$want" ] && [ "$want" != "$have" ]; then
warn "PVC ${ns}/${claim} already requests ${have}; keeping it (${env}=${want} applies to a new claim; grow this one with kubectl patch where its storage class allows expansion)"
fi
printf '%s' "$have"
}
# deployment_image <deployment> <container> prints the image that container of a
# control-plane Deployment runs now, or nothing when the Deployment does not exist yet.
deployment_image() {
kube -n "$CONTROL_NS" get deployment "$1" \
-o "jsonpath={.spec.template.spec.containers[?(@.name==\"$2\")].image}" 2>/dev/null || true
}
# restart_existing_control_plane restarts the Deployments the bundle apply left as they were:
# those that already ran FELIS_IMAGE, whose tag now names a rebuilt image (a rerun of the same
# version, or a FELIS_IMAGE the operator reuses). A Deployment whose image changed is rolling
# from the apply already, and must not be restarted on top: the restart is a second template
# change, so `rollout undo` would step back to the new image instead of the previous release.
#
# The registry pod runs the same binary in its registry-gate and registry-gc containers, so it
# follows the same rule; the rollout wait below covers it before anything is pushed.
restart_existing_control_plane() {
local prev_api="$1" prev_operator="$2" prev_gate="${3:-}"
# `if`, not `[ test ] && cmd`: as the LAST command of the function the and-list returns 1
# when the test is false, which becomes the function's exit status and kills the whole
# install under `set -Eeuo pipefail` — right after the bundle is applied and before the
# rollout wait.
if [ "$prev_api" = "$FELIS_IMAGE" ]; then
log "restarting felis-api onto the rebuilt ${FELIS_IMAGE}"
kube -n "$CONTROL_NS" rollout restart deployment/felis-api
fi
if [ "$prev_operator" = "$FELIS_IMAGE" ]; then
log "restarting felis-operator onto the rebuilt ${FELIS_IMAGE}"
kube -n "$CONTROL_NS" rollout restart deployment/felis-operator
fi
if [ "$prev_gate" = "$FELIS_IMAGE" ]; then
log "restarting the registry's gate onto the rebuilt ${FELIS_IMAGE}"
kube -n "$CONTROL_NS" rollout restart deployment/registry
fi
}
# push_image_to_registry <ref> re-tags a locally built image for the node's
# loopback push endpoint and uploads it. The registry keys a repository by the
# path AFTER the host, so pushing 127.0.0.1:5000/felis/felis:demo lands exactly
# where a later kubelet pull of registry.felis.svc:5000/felis/felis:demo (the
# mirror rewrites the host) will look. A ref not under REGISTRY_URL is not
# mirrored — warn, don't fail: the install is still self-consistent, that image
# just has no pull source once GC collects its containerd copy.
push_image_to_registry() {
local ref="$1" push_ref
case "$ref" in
"${REGISTRY_URL}/"*)
push_ref="${REGISTRY_PUSH_HOST}/${ref#"${REGISTRY_URL}/"}"
;;
*)
warn "not mirroring ${ref} into the internal registry: it is not under ${REGISTRY_URL}; once the image GC collects that tag, nothing can re-pull it"
return 0
;;
esac
log "mirroring ${ref} into the internal registry"
docker tag "$ref" "$push_ref" || die "could not tag ${ref} as ${push_ref} — is docker healthy?"
local attempt=1
until docker --config "$REGISTRY_DOCKER_CONFIG" push "$push_ref"; do
# The gate answers writes 503 while the registry-gc sidecar sweeps; wait
# that out, and fail at once on anything else.
if [ "$attempt" -ge 40 ] || ! registry_read_only; then
die "could not mirror ${ref} into the internal registry — check the registry Deployment/pod (the registry, registry-gate and registry-gc containers) and its PVC"
fi
warn "the registry is read-only for garbage collection; retrying the push of ${push_ref} in 30s (${attempt}/40)"
attempt=$((attempt + 1))
sleep 30
done
docker rmi "$push_ref" >/dev/null 2>&1 || true
}
# registry_read_only asks the gate, over its pod-loopback maintenance listener,
# whether a garbage-collection window is open.
registry_read_only() {
kubectl -n "$CONTROL_NS" exec deploy/registry -c registry-gc -- \
wget -q -O /dev/null "http://127.0.0.1:$((${REGISTRY_URL##*:} + 2))/readonly" >/dev/null 2>&1
}
# registry_docker_login logs a throwaway docker config into the registry gate as
# the platform principal: writes are refused anonymously, and this identity is
# the only one allowed under felis/. The config lives in a 0700 temp dir that the
# EXIT trap removes, so the token never lands in root's ~/.docker.
registry_docker_login() {
REGISTRY_DOCKER_CONFIG="$(umask 077; mktemp -d)"
remember_temp "$REGISTRY_DOCKER_CONFIG"
printf '%s' "$REGISTRY_PLATFORM_TOKEN" | docker --config "$REGISTRY_DOCKER_CONFIG" \
login --username platform --password-stdin "$REGISTRY_PUSH_HOST" >/dev/null \
|| die "could not log in to the internal registry at ${REGISTRY_PUSH_HOST} as platform — check the registry-gate container's log and the felis-registry-auth Secret"
}
# Every image this installer builds is hosted in the registry, so the copies it
# imported into containerd are a first-boot cache, not the only copy: kubelet
# re-pulls from the registry after any image GC. Runs AFTER deploy_bundle — the
# registry it pushes into does not exist before that.
#
# Docker is started once for the whole batch and stopped once at the end. A
# start/stop pair per image trips systemd's start rate limit — observed live on
# a re-run: three fast pushes, then "Start request repeated too quickly /
# start-limit-hit" and the fourth image never got mirrored. docker.service is
# socket-triggered, so each cycle counts twice against the burst limit.
push_images_to_registry() {
local img
systemctl start docker
registry_docker_login
for img in "$FELIS_IMAGE" "$FELIS_LIMBO_IMAGE" "$FELIS_LOBBY_IMAGE" "$FELIS_PAPER_IMAGE"; do
[ -n "$img" ] || continue
push_image_to_registry "$img"
done
for img in "$FELIS_LIMBO_IMAGE" "$FELIS_LOBBY_IMAGE" "$FELIS_PAPER_IMAGE"; do
[ -n "$img" ] || continue
push_version_tag "$img"
done
systemctl stop docker docker.socket 2>/dev/null || true
}
# push_version_tag mirrors a game image a second time under a tag no later run
# rewrites: <Minecraft version>-<first 12 hex of the image id>, e.g.
# felis/paper:26.2-3f9c0a1b2c4d. The :demo tag moves with every run, and servers are
# pinned to the digest it named when they were created, so this is the readable name
# for each build: an admin can whitelist it to create servers on that exact
# Minecraft version long after :demo has moved on.
push_version_tag() {
local ref="$1" id versioned
[ -n "${MC_VERSION:-}" ] || return 0
id="$(docker image inspect -f '{{.Id}}' "$ref" 2>/dev/null)" || return 0
id="${id#sha256:}"
versioned="${ref%:*}:${MC_VERSION}-${id:0:12}"
docker tag "$ref" "$versioned" || die "could not tag ${ref} as ${versioned} — is docker healthy?"
push_image_to_registry "$versioned"
docker rmi "$versioned" >/dev/null 2>&1 || true
}
# The login/lobby images use mutable :demo tags. Importing/pushing a replacement
# updates containerd, but an existing StatefulSet template is byte-for-byte
# unchanged and Kubernetes will not roll it. Recreate the two always-on system
# pods so a convergent bootstrap actually starts the images it just built — but
# only when the build changed: every player online is on one of these two, and a
# rerun that rebuilt nothing has nothing to start. SYSTEM_SERVER_IMAGES records the
# build each was last started on; it is written after the restarts, so a run that
# died in between restarts them next time.
restart_existing_system_servers() {
local name id pods next=""
while read -r name id; do
[ -n "$name" ] || continue
next="${next}${name} ${id}"$'\n'
pods="$(kube -n "$MINECRAFT_NS" get pod \
-l "felis.lolicon.best/server=${name}" -o name 2>/dev/null || true)"
[ -n "$pods" ] || continue
if [ -n "$id" ] && grep -qxF "${name} ${id}" "$SYSTEM_SERVER_IMAGES" 2>/dev/null; then
ok "${name} system server already runs this build; left running"
continue
fi
log "restarting existing ${name} system server to pick up its imported image"
kube -n "$MINECRAFT_NS" delete pod \
-l "felis.lolicon.best/server=${name}" --wait=false
done <<EOF
${LOGIN_SERVER} ${LIMBO_IMAGE_ID:-}
${LOBBY_SERVER} ${LOBBY_IMAGE_ID:-}
EOF
printf '%s' "$next" > "$SYSTEM_SERVER_IMAGES"
}
diagnose_rollout() {
local deploy="$1" name selector pod
name="${deploy##*/}"
warn "rollout not complete: ${deploy}"
kube -n "$CONTROL_NS" describe "$deploy" || true
case "$name" in
felis-api) selector='app.kubernetes.io/name=felis,app.kubernetes.io/component=api' ;;
felis-operator) selector='app.kubernetes.io/name=felis,app.kubernetes.io/component=operator' ;;
registry) selector='app.kubernetes.io/name=felis,app.kubernetes.io/component=registry' ;;
*) selector='' ;;
esac
[ -n "$selector" ] || return 0
kube -n "$CONTROL_NS" get pods -l "$selector" -o wide || true
for pod in $(kube -n "$CONTROL_NS" get pods -l "$selector" -o name 2>/dev/null); do
warn "pod detail: ${pod}"
kube -n "$CONTROL_NS" describe "$pod" || true
warn "recent logs: ${pod}"
kube -n "$CONTROL_NS" logs "$pod" --all-containers --tail=120 || true
kube -n "$CONTROL_NS" logs "$pod" --all-containers --previous --tail=120 || true
done
}
mark_bootstrap_done() {
date -u +%Y-%m-%dT%H:%M:%SZ > "$BOOTSTRAP_DONE"
chmod 0644 "$BOOTSTRAP_DONE"
}
# ---------------------------------------------------------------------------
# 9. Summary
# ---------------------------------------------------------------------------
summary() {
export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
echo
ok "Felis control plane deployed."
echo
kube -n "$CONTROL_NS" get pods -o wide || true
echo
systemctl --no-pager --full status felis-velocity 2>/dev/null | head -n 4 || true
echo
log "Player panel: https://console.${FELIS_ROOT_DOMAIN} — served on 443 once your edge/Cloudflare Tunnel routes it here."
log "Operator console (Op/Admin/Owner): https://op.console.${FELIS_ROOT_DOMAIN} — the Owner runs 'felis setup' and onboards here."
log "Before the edge is ready: direct + self-signed at https://${NODE_IP}:${FELIS_PANEL_NODEPORT} (browser will warn on first visit)."
log "Minecraft address: ${NODE_IP}:${FELIS_GAME_PORT} (point mc.${FELIS_ROOT_DOMAIN} here)"
log "The proxy authenticates against Mojang and forwards the verified profile to the"
log "login gate; the backends are reachable in-cluster only. Follow it with:"
log " sudo journalctl -u felis-velocity -f"
if [ "${FELIS_BOOTSTRAP_FROM_TUI:-}" = "1" ]; then
log "Returning to the setup console to create the Owner account and verify panel access."
else
log "Next: run 'sudo felis setup' on this host to create the Owner account."
fi
log "setup provisions the login/lobby servers, then asks the Owner to bind by joining"
log "the proxy in Minecraft — that is what makes the Owner's admin identity a real"
log "Mojang account rather than a password."
log "Use 'sudo felis breakGlass' only for emergency local Owner recovery/reset."
if [ -n "${PREVIOUS_FELIS_IMAGE:-}" ]; then
echo
log "The control plane moved from ${PREVIOUS_FELIS_IMAGE} to ${FELIS_IMAGE}"
log "(also recorded in ${STATE_DIR}/previous-felis-image). To go back to it:"
log " kubectl -n ${CONTROL_NS} rollout undo deployment/felis-api deployment/felis-operator"
log "The database stays migrated; docs/troubleshooting.md §16 has the full rollback."
fi
echo
summary_offsite
echo
}
# ---------------------------------------------------------------------------
# 10. Felis-nano install path — the auth multiplexer only: felis binary + a
# minimal [[auth_source]] config + a systemd unit running `felis nano`.
# No k3s, no Postgres, no control-plane bundle. Chosen at the top-of-run
# prompt (or FELIS_INSTALL_MODE=nano).
# ---------------------------------------------------------------------------
prompt_install_mode() {
# felis setup carries on to the Owner and edge setup, which needs the control plane, so
# a nano install under it could only end in a setup error.
if bootstrap_from_tui; then
[ "$INSTALL_MODE" != nano ] || die "felis setup installs the full control plane; for Felis-nano run deploy/bootstrap.sh with FELIS_INSTALL_MODE=nano"
INSTALL_MODE="full"
log "install mode: full (felis setup)"
return 0
fi
case "$INSTALL_MODE" in
full|nano) log "install mode: ${INSTALL_MODE} (from FELIS_INSTALL_MODE)"; return 0 ;;
"") ;;
*) die "FELIS_INSTALL_MODE must be 'full' or 'nano', got: ${INSTALL_MODE}" ;;
esac
# A felis-nano unit with no full install beside it makes this re-run a nano update;
# defaulting to full there would put k3s and Postgres on a host that asked for neither.
local def=full n=1 reply
if [ -e "$NANO_SERVICE" ] && [ ! -e "$BOOTSTRAP_DONE" ]; then def=nano n=2; fi
# No override: ask on the controlling terminal. Under `curl | sudo bash` stdin
# is the script, so we must read /dev/tty, not stdin. No tty (CI/cloud-init) →
# take the default. Open it to find out: /dev/tty is mode 0666 on every Linux host, so
# `-r` passes even when there is no controlling terminal and only the open fails.
if ! (: </dev/tty) 2>/dev/null; then
INSTALL_MODE="$def"
log "no terminal for a prompt; defaulting to a ${def} install (set FELIS_INSTALL_MODE=full or nano to override)"
return 0
fi
printf '\n'
printf 'What do you want to install on this host?\n'
printf ' [1] Felis — full control plane (k3s + Postgres + panel; orchestrates Minecraft servers)\n'
printf ' [2] Felis-nano — auth multiplexer only (federates Mojang + third-party Yggdrasil; no k3s/DB)\n'
while :; do
printf 'Choose [1/2] (default %s): ' "$n"
IFS= read -r reply </dev/tty || reply=""
case "$reply" in
"") INSTALL_MODE="$def"; break ;;
1|full|Felis|felis) INSTALL_MODE="full"; break ;;
2|nano|felis-nano|Felis-nano) INSTALL_MODE="nano"; break ;;
*) printf 'Please enter 1 or 2.\n' ;;
esac
done
log "install mode: ${INSTALL_MODE}"
}
install_go_toolchain() {
local arch tarball url tmp want have
if [ -x "${GOROOT_DIR}/bin/go" ] && "${GOROOT_DIR}/bin/go" version | grep -q "go${FELIS_GO_VERSION} "; then
ok "go ${FELIS_GO_VERSION} already installed at ${GOROOT_DIR}"
return 0
fi
case "$(uname -m)" in
x86_64|amd64) arch="amd64"; want="$GO_PINNED_SHA256_AMD64" ;;
aarch64|arm64) arch="arm64"; want="$GO_PINNED_SHA256_ARM64" ;;
*) die "no Go toolchain build for architecture $(uname -m); set FELIS_GO_VERSION or pre-stage ${GOROOT_DIR}" ;;
esac
[ "$FELIS_GO_VERSION" = "$GO_PINNED_VERSION" ] || want="$FELIS_GO_SHA256"
[ -n "$want" ] || die "no pinned sha256 for Go ${FELIS_GO_VERSION}; set FELIS_GO_SHA256 to the linux-${arch} digest https://go.dev/dl/ lists for it, or pre-stage ${GOROOT_DIR}"
tarball="go${FELIS_GO_VERSION}.linux-${arch}.tar.gz"
url="https://go.dev/dl/${tarball}"
log "installing Go ${FELIS_GO_VERSION} (${arch}) to ${GOROOT_DIR}"
# A private directory, not a fixed /tmp name another local user could have planted first.
tmp="$(mktemp -d)"
remember_temp "$tmp"
curl -fsSL "$url" -o "${tmp}/${tarball}" || die "failed to download the Go toolchain: ${url}"
# Checked before the old toolchain is removed, so a refusal leaves the host as it was.
# Hash stdin, never the path — same reason as install_via_plugins.
have="$(sha256sum <"${tmp}/${tarball}" | cut -d' ' -f1)"
[ "$have" = "$want" ] || die "Go ${FELIS_GO_VERSION} (${arch}) checksum mismatch: got ${have}, expected ${want}"
rm -rf "$GOROOT_DIR"
mkdir -p "$(dirname "$GOROOT_DIR")"
tar -C "$(dirname "$GOROOT_DIR")" -xzf "${tmp}/${tarball}" || die "failed to unpack ${tarball}"
ok "go toolchain at ${GOROOT_DIR}/bin/go"
}
build_nano_binary() {
local staged="${SRC_DIR}/.felis-nano-build"
log "building felis from source (${SRC_DIR})${FELIS_VERSION:+ — ${FELIS_VERSION}}"
# Stamped for the same reason the image build is: a nano host runs `felis version` and
# `felis update` too, and this path used to pass no ldflags at all.
( cd "$SRC_DIR" \
&& PATH="${GOROOT_DIR}/bin:${PATH}" CGO_ENABLED=0 go build -trimpath \
-ldflags "-X main.version=${FELIS_VERSION:-dev}" -o "$staged" ./cmd/felis ) \
|| die "go build ./cmd/felis failed"
# The whole point of this path is the nano subcommand; a binary without it would
# only surface as a crash-looping felis-nano.service, so fail loudly here instead.
"$staged" -h 2>&1 | grep -qw nano || die "built binary has no 'nano' subcommand"
# Stage-then-install, never `go build -o ${HOST_BIN}` directly: the Go linker renames
# its output out of $TMPDIR, and a same-filesystem rename CARRIES THE SOURCE SELinux
# label — the binary lands in /usr/local/bin still labelled user_tmp_t. Root (being
# unconfined) can still run it, so it looks fine by hand, but the DynamicUser service
# cannot exec it and felis-nano dies with 203/EXEC. Creating the file fresh at the
# destination lets the policy's type transition label it bin_t; restorecon is the belt.
mkdir -p "$(dirname "$HOST_BIN")"
keep_previous_host_binary
rm -f "$HOST_BIN"
install -m 0755 "$staged" "$HOST_BIN"
rm -f "$staged"
command -v restorecon >/dev/null 2>&1 && restorecon "$HOST_BIN" >/dev/null 2>&1 || true
ok "felis binary on host at ${HOST_BIN}"
}
acquire_nano_binary() {
# The release channel takes the same prebuilt binary the control plane does. This is the
# biggest win on this path: a host that only wants the auth multiplexer stops needing a Go
# toolchain and a checkout at all.
if use_release_binary; then
resolve_install_ref
if download_release_binary; then
return 0
fi
fi
# Raw curl|bash with no usable release: build it straight from source with a pinned Go
# toolchain — nano needs one static binary, not an image, so dragging in docker
# (as the full control-plane path does) buys nothing and costs a daemon that must
# start. It does not start on EL10: the docker-ce el10 rpms install but dockerd
# fails, which used to kill the whole nano install at `systemctl enable --now docker`.
fetch_source
install_go_toolchain
build_nano_binary
}
write_nano_config() {
local target="${STATE_DIR}/felis.toml"
# The unit is a DynamicUser, so it can read felis.toml only if it can search this
# directory. The mode is explicit because a hardened root umask (027) would leave it 0750,
# which is what older installers did to nano-only hosts. Only the full install keeps its
# own 0700: that directory holds secrets, and install_nano_service reports the lockout
# rather than this widening it.
if [ ! -d "$STATE_DIR" ]; then
mkdir -p "$STATE_DIR"
chmod 0755 "$STATE_DIR"
elif [ ! -e "$SECRETS_ENV" ] && [ ! -e "$BOOTSTRAP_DONE" ]; then
chmod 0755 "$STATE_DIR"
fi
if [ -e "$target" ]; then
ok "config already present at ${target}; leaving it (edit it to add [[auth_source]] roots)"
return 0
fi
cat > "$target" <<'EOF'
# Felis-nano — Yggdrasil hasJoined multiplexer.
# Mojang is always the first (identity) source, added in code — do NOT list it here.
# Add each third-party Yggdrasil root below (priority = order). url is the FULL
# hasJoined endpoint. After editing: sudo systemctl restart felis-nano
#
# Order is trust: the first source that answers 200 wins, so list the most trusted roots
# first, and remove a compromised root rather than just moving it down.
# tag is permanent: it is hashed into every player UUID of its source, so changing it
# (even its case) gives all of them new UUIDs and orphans their data, links and bans.
#
# prefix is required, 1-4 letters/digits, unique per source. A player of this source
# whose name belongs to a Mojang account joins as PREFIX_name (LS_steve) instead —
# otherwise the proxy, which keys its player list on the NAME, refuses to have both
# online at once ("You are already connected to this proxy!"). Everyone else keeps
# their own name.
#
# [[auth_source]]
# tag = "littleskin"
# prefix = "LS"
# url = "https://littleskin.cn/api/yggdrasil/sessionserver/session/minecraft/hasJoined"
EOF
chmod 0644 "$target"
ok "wrote nano config template ${target} (edit it to add your Yggdrasil sources)"
}
# resolve_nano_listen settles FELIS_NANO_LISTEN: the operator's value, else the address the
# installed felis-nano unit listens on, else loopback. Re-running this script is how a nano
# host updates, and without the middle step that re-run moved an off-host proxy's endpoint
# back to 127.0.0.1, so every login through it failed.
resolve_nano_listen() {
if [ -z "$FELIS_NANO_LISTEN" ] && [ -r "$NANO_SERVICE" ]; then
FELIS_NANO_LISTEN="$(sed -n 's/^ExecStart=.* -listen \([^ ]*\).*$/\1/p' "$NANO_SERVICE")"
fi
FELIS_NANO_LISTEN="${FELIS_NANO_LISTEN:-127.0.0.1:8081}"
}
nano_listen_is_loopback() {
case "${FELIS_NANO_LISTEN%:*}" in
127.*|localhost|"[::1]") return 0 ;;
*) return 1 ;;
esac
}
configure_nano_firewall() {
# A loopback bind is unreachable from off-host by construction, so opening the port
# would advertise a hole nothing answers on. Only punch it for a routable bind.
if nano_listen_is_loopback; then
ok "nano listens on ${FELIS_NANO_LISTEN} (loopback); no firewall port opened"
return 0
fi
command -v firewall-cmd >/dev/null 2>&1 || return 0
systemctl is-active --quiet firewalld || return 0
local port="${FELIS_NANO_LISTEN##*:}" family=ipv4
# hasJoined takes no token, so the port is opened to the proxy alone. Earlier installers
# opened it to every source, and a re-run must not leave that behind. A rule for a previous
# FELIS_NANO_PROXY_CIDR is not tracked; it stays until removed by hand.
if firewall-cmd --permanent --query-port="${port}/tcp" >/dev/null 2>&1; then
log "closing firewalld port ${port}/tcp, which an earlier install opened to every source"
firewall-cmd --permanent --remove-port="${port}/tcp"
fi
if [ -n "$FELIS_NANO_PROXY_CIDR" ]; then
case "$FELIS_NANO_PROXY_CIDR" in *:*) family=ipv6 ;; esac
log "opening firewalld port ${port}/tcp to ${FELIS_NANO_PROXY_CIDR} only"
firewall-cmd --permanent --add-rich-rule="rule family=\"${family}\" source address=\"${FELIS_NANO_PROXY_CIDR}\" port port=\"${port}\" protocol=\"tcp\" accept"
else
warn "no FELIS_NANO_PROXY_CIDR, so firewalld keeps ${port}/tcp closed; the summary shows how to admit your proxy"
fi
firewall-cmd --reload
}
install_nano_service() {
# DynamicUser: no static account, ephemeral UID. nano writes nothing (logs go to
# journald) and only reads the world-readable felis.toml, so strict sandboxing fits.
cat > "$NANO_SERVICE" <<EOF
[Unit]
Description=Felis-nano Yggdrasil hasJoined multiplexer
After=network-online.target
Wants=network-online.target
[Service]
Type=simple
ExecStart=${HOST_BIN} nano -config ${STATE_DIR}/felis.toml -listen ${FELIS_NANO_LISTEN}
Restart=on-failure
RestartSec=5
DynamicUser=yes
NoNewPrivileges=yes
ProtectSystem=strict
ProtectHome=yes
[Install]
WantedBy=multi-user.target
EOF
systemctl daemon-reload
systemctl enable felis-nano
# restart, not `enable --now`: on a re-run the service is already active and --now would
# leave the OLD binary running against the NEW unit. Converge means converge.
HOST_BIN_IN_USE=1
systemctl restart felis-nano
# restart returns as soon as the process is forked. A config the new binary rejects, or a
# file it cannot open, only shows once it has exited and the unit sits in auto-restart.
sleep 2
if ! systemctl is-active --quiet felis-nano; then
journalctl -u felis-nano -n 20 --no-pager || true
die "felis-nano did not stay up; its last log lines are above"
fi
ok "felis-nano.service enabled and started (listen ${FELIS_NANO_LISTEN})"
}
summary_nano() {
local port="${FELIS_NANO_LISTEN##*:}" host="${FELIS_NANO_LISTEN%:*}"
# A wildcard bind names no address a proxy could dial; the node's own is the useful one.
case "$host" in ""|0.0.0.0|"[::]") host="${NODE_IP}" ;; esac
echo
ok "Felis-nano deployed."
echo
systemctl --no-pager --full status felis-nano 2>/dev/null | head -n 6 || true
echo
log "hasJoined endpoint: http://${host}:${port}/session/minecraft/hasJoined"
log "Point Velocity at it — add to the proxy JVM startup flags (between java and -jar):"
log " -Dmojang.sessionserver=http://${host}:${port}/session/minecraft/hasJoined"
log " (the FULL endpoint URL, path included — Velocity's default for this property is"
log " the full https://sessionserver.mojang.com/session/minecraft/hasJoined)"
if nano_listen_is_loopback; then
log "Bound to loopback: reachable from Velocity on THIS host, and from nowhere else."
log "Proxy on another machine? Re-run with the address on the sudo line (sudo drops"
log "exported variables):"
log " curl -fsSL <raw-url>/deploy/bootstrap.sh | sudo FELIS_NANO_LISTEN=<private-ip>:${port} FELIS_NANO_PROXY_CIDR=<proxy-ip>/32 bash"
log "firewalld then admits ${port}/tcp ONLY from that proxy — hasJoined takes no auth token,"
log "so an internet-facing one is a free auth relay burning your Mojang egress IP."
elif [ -n "$FELIS_NANO_PROXY_CIDR" ]; then
log "Bound to ${FELIS_NANO_LISTEN}. firewalld, where it runs, admits ${port}/tcp only from"
log "${FELIS_NANO_PROXY_CIDR}; any other firewall in front of this host must do the same."
else
log "WARNING: bound to ${FELIS_NANO_LISTEN} with no FELIS_NANO_PROXY_CIDR. hasJoined takes no auth"
log "token, so admit ${port}/tcp from your proxy alone, or anyone can relay their logins"
log "through you. firewalld, where it runs, keeps the port closed until you add:"
log " firewall-cmd --permanent --add-rich-rule='rule family=\"ipv4\" source address=\"<proxy-ip>/32\" port port=\"${port}\" protocol=\"tcp\" accept' && firewall-cmd --reload"
fi
log "Then edit ${STATE_DIR}/felis.toml to add your [[auth_source]] roots and run:"
log " sudo systemctl restart felis-nano"
log "Follow live login traffic with: sudo journalctl -u felis-nano -f"
echo
}
main_nano() {
detect_node_ip
install_base
acquire_nano_binary
write_nano_config
install_nano_service
configure_nano_firewall
summary_nano
}
main() {
resolve_nano_listen
validate_settings
detect_os
prompt_install_mode
if [ "$INSTALL_MODE" = "nano" ]; then
main_nano
return
fi
quiet_watchdog
pause_package_background_timers
detect_node_ip
ensure_swap
install_base
# Right after install_base because it is the first point curl exists, and well before
# docker and k3s: a missing FELIS_GITHUB_TOKEN or an unpublished release should cost
# the operator seconds, not a k3s install they then have to unwind. This is purely
# fail-fast — fetch_source resolves again itself — so it must skip on exactly the paths
# that never consume the result, or it invents a network dependency and a version they
# do not have: the TUI rebuilds the binary it is already running, and FELIS_SKIP_FETCH
# builds whatever is staged, which stamp_version reads the SHA off. Resolving anyway
# would set FELIS_VERSION to the newest tag and stamp a staged tree as that release.
bootstrap_from_tui || [ -n "${FELIS_SKIP_FETCH:-}" ] || resolve_install_ref
install_cloudflared
load_or_make_secrets
configure_offsite
ensure_panel_tls_cert
install_docker
install_k3s
# The registry mirror must exist before the bundle's pods start pulling (and
# before any re-run's rollouts); the registry's own image must be in containerd
# before its Deployment can start at all.
configure_registry_mirror
import_registry_image
# Three ways to end up with a felis binary, in preference order. The release download is
# the only one that skips compiling: it is the CI artifact for this exact tag, panel
# included. Both other arms leave HAVE_PREBUILT_BINARY unset where a source build is what
# actually happens, which is what routes build_image below.
if bootstrap_from_tui; then
install_embedded_binary
elif use_release_binary && download_release_binary; then
:
else
fetch_source
fi
resolve_felis_image
build_image
# After build_image imported the felis image: the registry pod's gate runs it.
pin_registry_images
# Before build_game_stack: the builds user servers run must be read off the
# registry's tags before new ones replace them.
pin_user_server_images
build_game_stack
install_postgres
configure_postgres
run_migrations
deploy_bundle
# AFTER deploy_bundle: the registry the built images are mirrored into is part
# of that bundle.
push_images_to_registry
# After deploy_bundle, like the pushes: it writes into the registry.
install_build_tools_timer
restart_existing_system_servers
# After deploy_bundle: the proxy dials felis-api's internal ClusterIP, which does not
# exist until the bundle is applied.
install_velocity
# After deploy_bundle: the bundle's MinecraftServer export reads the cluster.
install_db_backup_timer
# After the backup timer: its first bundle is part of the first copy.
install_offsite_timer
# Last: its first run should see the platform as this install leaves it.
install_watchdog_timer
mark_bootstrap_done
summary
}
main "$@"
Reference in new issue View Git Blame Copy Permalink