Files
picpeak/backend/wait-for-db.sh
T
Luca 7223118b89 feat(deploy): make the all-in-one image installable without a shell (#1124)
The all-in-one image could not be installed from a GUI at all — the deployment it
exists for. validateEnv treats a missing JWT_SECRET as critical and exits, and the
documented run command supplies it with `openssl rand`, a shell command a Synology
Container Manager or QNAP Container Station form cannot run.

wait-for-db.sh now generates one on first start and persists it next to the database,
extending the existing /run/secrets hydration rather than adding a second mechanism.
Explicit env still wins, then /run/secrets, then the generated file. The write is
load-bearing: JWT_SECRET is exported only when the file actually persisted, because an
unpersisted secret would mint a new one every restart and sign every session out.

Creation writes to a private temp file and hard-links it into place — atomic, fails with
EEXIST when another container won, and the loser adopts the winner's value. Non-regular
paths are rejected before the link, since POSIX ln links INTO a directory rather than
failing, which would make a mistyped -v target unrecoverable.

Also repairs the onboarding paths a new install actually walks: the installer no longer
rotates the secrets of a running install on re-run, deprecates the dead scripts/install.sh
in place, corrects the CONTRIBUTING dev loop, and fixes the vite proxy target that had
been pointing at a stray local port since 0da45e69.

Reviewed over three rounds. Co-authored by @Luca-Timo.
2026-08-22 18:37:12 +02:00

364 lines
18 KiB
Bash
Executable File

#!/bin/sh
# wait-for-db.sh - Wait for PostgreSQL to be ready before starting the application
set -e
# Machine secrets (JWT/DB/Redis): if not supplied via the environment, read them
# from the generated secret files that the compose `secrets-init` service writes
# to /run/secrets. Explicit env ALWAYS wins, so installs that set
# JWT_SECRET/DB_PASSWORD/REDIS_PASSWORD in .env are unaffected. Runs before the
# root -> nodejs re-exec so the exported values survive su-exec.
for _pair in JWT_SECRET:jwt_secret DB_PASSWORD:db_password REDIS_PASSWORD:redis_password; do
_var="${_pair%%:*}"
_file="/run/secrets/${_pair##*:}"
eval "_cur=\${$_var:-}"
if [ -z "$_cur" ] && [ -s "$_file" ]; then
export "$_var=$(cat "$_file")"
fi
done
unset _pair _var _file _cur
# Permission handling (#484): the image starts as root so this script can
# chown bind-mounted host volumes to UID 1001 (nodejs) before dropping
# privileges via su-exec. This avoids the fresh-install restart loop where
# the host directory's UID (commonly 1000) didn't match the container's
# hard-coded nodejs user. Compose deployments that pin `user:` to something
# other than root skip this branch — they own permissions themselves and hit
# the preflight check below instead.
# The writable roots. Defaults are the compose layout; the all-in-one image
# (#1042) points all of them under one mounted volume, so these must follow the
# same env vars the app itself reads rather than hard-coding /app.
DATA_DIRS="${STORAGE_PATH:-/app/storage} ${DATA_DIR:-/app/data} ${LOG_DIR:-/app/logs}"
# The backup root is adopted when explicitly configured, but never gates boot:
# docker-compose.production.yml does not mount /backup, so a hardened non-root
# deployment would fail `mkdir -p /backup` against a root-owned / and refuse to
# start over a directory it never needed.
if [ -n "${BACKUP_DIR:-}" ]; then
DATA_DIRS="$DATA_DIRS $BACKUP_DIR"
fi
# When every root lives under ONE mounted volume (the all-in-one image, #1042),
# the mount point itself must be adopted too. Chowning only the children leaves
# a host directory created with 0700/0750 and a foreign owner untraversable by
# UID 1001 after the su-exec drop, so the preflight below rejects children the
# script just created. Docker Desktop's permissive bind mounts hide this; a NAS
# share does not.
#
# It is deliberately kept out of DATA_DIRS: everything below it is already
# chowned recursively, so adding it there would walk the whole photo library a
# second time on every restart — minutes of startup delay on exactly the large
# NAS libraries this image targets. The mount point needs its own ownership
# fixed, nothing more, so it gets a shallow chown of its own below.
DATA_ROOT_DIR="${DATA_ROOT:-}"
# Create the roots before touching them. With the compose layout each is its own
# mount point so they always exist — but the AIO image mounts ONE volume at
# /data, and a bind-mounted host directory hides the tree baked into the image.
# chown would then fail on paths that do not exist and report "the filesystem
# rejects chown", which is both wrong and a dead end for NAS users.
# shellcheck disable=SC2086 — intentional word-splitting over the roots
mkdir -p $DATA_ROOT_DIR $DATA_DIRS 2>/dev/null || true
# Last resort for JWT_SECRET, after the /run/secrets pass above: generate one
# and persist it next to the database. Deliberately placed here rather than
# beside that loop because it needs DATA_DIR to exist, and before the chown
# below so the new file is adopted along with everything else.
#
# The compose stack never reaches this — `secrets-init` has already written
# /run/secrets/jwt_secret. It exists for deployments with no shell: the
# all-in-one image mounts no secrets volume, and the documented one-liner asks
# the user to run `openssl rand` on the HOST, which a NAS Container Manager or
# Container Station form cannot do. Without this, JWT_SECRET is unset,
# validateEnv treats it as critical, and the container exits — so the whole
# GUI-driven install path is a dead end (#705).
#
# JWT_SECRET only. DB_PASSWORD must match whatever the Postgres volume was
# initialised with, so inventing one at boot would break the connection rather
# than fix it; the AIO default engine is SQLite and Redis is not used here.
#
# Losing the file invalidates every session and gallery link. Note the built-in
# backup does NOT carry it — backupService archives STORAGE_PATH and the
# database dump, not DATA_DIR — so only a copy of the whole volume preserves it.
if [ -z "${JWT_SECRET:-}" ]; then
_jwt_file="${DATA_DIR:-/app/data}/jwt.secret"
# Read, trim, and length-check rather than testing that the file merely
# exists. A short write (ENOSPC, a SIGKILL mid-boot) leaves a partial file
# that a `-s` test accepts on every later boot, and validateEnv only WARNS
# below 32 characters — so the install would sign with a truncated key
# indefinitely and nothing would say so. Whitespace is stripped first because
# validateEnv rejects an all-blank secret as critical, and an all-blank file
# long enough to pass a naive length test would boot-loop forever.
_jwt_read() {
[ -f "$1" ] || return 0
tr -d '\r\n\t ' < "$1" 2>/dev/null || true
}
# Loop rather than a single pass, because the two failure causes look
# identical on the first read and need opposite responses: a file that is
# absent-then-present is a concurrent boot winning the race (adopt it), while
# a file that is still unusable after we have tried to link and re-read is
# genuine corruption (clear it and retry). Removing on the first pass would
# delete a perfectly good secret another container had just created.
_jwt_made=""
_jwt_cur="$(_jwt_read "$_jwt_file")"
_jwt_try=0
while [ "${#_jwt_cur}" -lt 32 ] && [ "$_jwt_try" -lt 2 ]; do
_jwt_try=$((_jwt_try + 1))
if [ -e "$_jwt_file" ] && [ ! -f "$_jwt_file" ]; then
# A directory (or anything non-regular) has to go BEFORE the link is
# attempted, not after: POSIX `ln src dir` links src INTO the directory
# instead of failing, so the first attempt would create
# jwt.secret/jwt.secret.<sfx>.tmp, leave the directory non-empty, and make
# rmdir impossible from then on — the container could never recover from a
# mistyped `-v` target without host access. Verified against the runtime
# image; busybox ln does exactly this.
echo "[secrets] $_jwt_file is not a regular file — removing it." >&2
rmdir "$_jwt_file" 2>/dev/null || rm -f "$_jwt_file" 2>/dev/null || true
elif [ "$_jwt_try" -gt 1 ] && [ -f "$_jwt_file" ]; then
# Re-read immediately before unlinking. Our earlier read may be stale:
# another container can have linked a perfectly good secret in between,
# and deleting it would leave the two of us signing with different keys.
_jwt_cur="$(_jwt_read "$_jwt_file")"
[ "${#_jwt_cur}" -ge 32 ] && break
echo "[secrets] $_jwt_file is unusable — regenerating." >&2
rm -f "$_jwt_file" 2>/dev/null || true
fi
# Still not a regular file (a non-empty directory we refuse to delete):
# nothing further will work, so stop and let the warning below explain.
if [ -e "$_jwt_file" ] && [ ! -f "$_jwt_file" ]; then
break
fi
# openssl is not in the runtime image; /dev/urandom + base64 always are.
# 48 bytes -> 64 base64 chars, comfortably past the 32 above.
_jwt_new="$(head -c 48 /dev/urandom | base64 | tr -d '\n' || true)"
# Two properties are needed at once, and each idiom alone gives only one:
#
# * the file must never be READABLE half-written, or a concurrent boot
# reads a partial secret — a plain `set -C` redirect creates the inode
# first and writes after, so a loser can read it in between
# * the create must not CLOBBER, or two containers each believe they own
# the secret and sign with different keys — which is what `mv -f` does
#
# Writing the full content to a private temp file and hard-linking it into
# place gives both: the link is atomic, fails with EEXIST when another
# container already won, and the content is complete before the name exists.
#
# The suffix must be random rather than $$: the PID namespace makes $$ equal
# to 1 in every container, so containers sharing a volume would all pick the
# same temp path — and since a hard link shares the INODE, a second process
# truncating that path writes straight through the linked secret.
if [ -n "$_jwt_new" ]; then
_jwt_sfx="$(head -c 12 /dev/urandom | base64 | tr -dc 'a-z0-9' | cut -c1-10 || true)"
[ -n "$_jwt_sfx" ] || _jwt_sfx="$$"
_jwt_tmp="$_jwt_file.$_jwt_sfx.tmp"
if (umask 077; printf '%s\n' "$_jwt_new" > "$_jwt_tmp") 2>/dev/null \
&& ln "$_jwt_tmp" "$_jwt_file" 2>/dev/null; then
_jwt_made=yes
fi
rm -f "$_jwt_tmp" 2>/dev/null || true
fi
# Trust the file, never the value in hand — whoever won the link is the
# source of truth for every container on this volume, and the loser adopts
# it rather than signing with the value it happened to generate.
_jwt_cur="$(_jwt_read "$_jwt_file")"
done
if [ "${#_jwt_cur}" -ge 32 ]; then
export JWT_SECRET="$_jwt_cur"
if [ -n "$_jwt_made" ]; then
echo "[secrets] No JWT_SECRET supplied — generated one and stored it in $_jwt_file."
echo "[secrets] Back up the whole data volume to keep it; the built-in backup does"
echo "[secrets] not include it, and losing it signs every admin session and gallery"
echo "[secrets] link out."
fi
else
echo "WARNING: no JWT_SECRET supplied and $_jwt_file could not be created." >&2
echo " Startup will fail validation. Pass -e JWT_SECRET=... , or check that the" >&2
echo " data directory is writable and that $_jwt_file is not a directory." >&2
fi
unset _jwt_new _jwt_tmp _jwt_sfx _jwt_made _jwt_try
unset _jwt_file _jwt_cur
fi
if [ "$(id -u)" = "0" ]; then
if [ -n "$DATA_ROOT_DIR" ] && ! chown nodejs:nodejs "$DATA_ROOT_DIR" 2>/dev/null; then
echo "ERROR: failed to chown $DATA_ROOT_DIR to nodejs (UID 1001)." >&2
echo " The mounted volume root must be traversable by UID 1001 after the privilege drop." >&2
echo " Workaround: chown 1001:1001 the host directory you mounted at $DATA_ROOT_DIR." >&2
exit 1
fi
if ! chown -R nodejs:nodejs $DATA_DIRS 2>/dev/null; then
echo "ERROR: failed to chown $DATA_DIRS to nodejs (UID 1001)." >&2
echo " This usually means the host filesystem rejects chown (e.g. NFS without root squash" >&2
echo " disabled, or a SELinux/AppArmor policy blocking the operation)." >&2
echo " Workaround: pre-chown the host directories to 1001:1001 and pin 'user: \"1001:1001\"'" >&2
echo " in your compose file so this script never tries to chown them itself." >&2
echo " See https://docs.picpeak.app/deployment/docker#permissions" >&2
exit 1
fi
exec su-exec nodejs:nodejs "$0" "$@"
fi
# Belt-and-suspenders: if we got here as non-root (compose `user:` override),
# verify the bind mounts are actually writable before proceeding. Failing
# loud here beats the previous behavior — silent mkdir-||-true at line 69
# followed by a confusing migration error and a restart loop.
_uid="$(id -u)"
_gid="$(id -g)"
for _dir in $DATA_ROOT_DIR $DATA_DIRS; do
if [ ! -w "$_dir" ]; then
echo "ERROR: $_dir is not writable by UID $_uid." >&2
echo " Either drop the 'user:' override from your compose file so the container starts as" >&2
echo " root and can self-fix permissions, or run on the host:" >&2
echo " chown -R $_uid:$_gid <host-mount-for-$_dir>" >&2
echo " See https://docs.picpeak.app/deployment/docker#permissions" >&2
exit 1
fi
done
# Explicit SQLite boots (DATABASE_CLIENT=sqlite3 — the all-in-one image's
# default, #1042) have no Postgres to wait for: skip the whole readiness/
# create/verify section below. The engine resolver further down still runs,
# still logs the resolved engine, and still refuses the populated-both
# conflict (#1038). Compose deployments pin DATABASE_CLIENT=pg and knexfile's
# production block defaults to pg when unset, so nothing changes for them.
if [ "${DATABASE_CLIENT:-}" != "sqlite3" ]; then
host="${DB_HOST:-postgres}"
port="${DB_PORT:-5432}"
user="${DB_USER:-picpeak}"
target_db="${DB_NAME:-picpeak}"
# Hand the app EXACTLY the connection this script verified. knexfile's
# production block defaults DB_HOST to `db` while this script defaults to
# `postgres`, so a bare `docker run` with no DB_HOST would have had the
# readiness check pass against one host and the app then dial another (#1038
# review). Compose sets DB_HOST explicitly and is unaffected.
export DB_HOST="$host"
export DB_PORT="$port"
export DB_USER="$user"
export DB_NAME="$target_db"
# Use target database for checks - the picpeak user may not have access to 'postgres' database
default_db="${DB_CHECK_DB:-$target_db}"
sanitize_identifier() {
printf '%s' "$1" | sed "s/'/''/g"
}
echo "Waiting for PostgreSQL at $host:$port..."
# First, wait for PostgreSQL server to be reachable
max_attempts=30
attempt=0
while [ $attempt -lt $max_attempts ]; do
if PGPASSWORD="$DB_PASSWORD" psql -h "$host" -p "$port" -U "$user" -d "$target_db" -c '\q' >/dev/null 2>&1; then
>&2 echo "PostgreSQL is up - database \"$target_db\" is accessible."
break
fi
# If target DB doesn't work, try connecting to 'postgres' or 'template1' to create it
if PGPASSWORD="$DB_PASSWORD" psql -h "$host" -p "$port" -U "$user" -d "template1" -c '\q' >/dev/null 2>&1; then
>&2 echo "PostgreSQL is up - checking if database \"$target_db\" needs to be created..."
# Check if database exists
db_exists=$(PGPASSWORD="$DB_PASSWORD" psql -h "$host" -p "$port" -U "$user" -d "template1" -tAc "SELECT 1 FROM pg_database WHERE datname = '$(sanitize_identifier "$target_db")'" 2>/dev/null || echo 0)
if [ "$db_exists" != "1" ]; then
>&2 echo "Database \"$target_db\" not found. Attempting to create..."
if PGPASSWORD="$DB_PASSWORD" psql -h "$host" -p "$port" -U "$user" -d "template1" -c "CREATE DATABASE \"$target_db\";" >/dev/null 2>&1; then
>&2 echo "Database \"$target_db\" created successfully."
else
>&2 echo "Warning: Could not create database. It may already exist or user lacks permissions."
fi
fi
break
fi
attempt=$((attempt + 1))
>&2 echo "PostgreSQL is unavailable - sleeping (attempt $attempt/$max_attempts)"
sleep 2
done
if [ $attempt -eq $max_attempts ]; then
>&2 echo "Failed to connect to PostgreSQL after $max_attempts attempts."
exit 1
fi
# Final verification - wait for target database to accept connections
until PGPASSWORD="$DB_PASSWORD" psql -h "$host" -p "$port" -U "$user" -d "$target_db" -c '\q' >/dev/null 2>&1; do
>&2 echo "Waiting for database \"$target_db\" to accept connections..."
sleep 2
done
>&2 echo "Target database \"$target_db\" is ready."
else
>&2 echo "DATABASE_CLIENT=sqlite3 — skipping the PostgreSQL readiness wait."
fi # end Postgres wait (skipped for explicit sqlite3 boots)
# Ensure storage directories exist with proper permissions (Issue #67 fix)
# When host directories are bind-mounted, the container's built-in directories are overridden
# This ensures the required directory structure exists before the application starts
echo "Ensuring storage directories exist..."
STORAGE_BASE="${STORAGE_PATH:-/app/storage}"
mkdir -p "$STORAGE_BASE/events/active" "$STORAGE_BASE/events/archived" "$STORAGE_BASE/thumbnails" 2>/dev/null || true
# Backup destinations seeded by migrations 029 + 030 (/backup/picpeak and
# /backup/database). Creating the root alone is not enough: on a bind mount the
# subdirectories baked into the image are hidden and the backup services do not
# create them, so a backup would fail with ENOENT.
BACKUP_BASE="${BACKUP_DIR:-/backup}"
if [ -d "$BACKUP_BASE" ]; then
mkdir -p "$BACKUP_BASE/picpeak" "$BACKUP_BASE/database" 2>/dev/null || true
fi
# Resolve which database engine this boot should use (#1038) BEFORE migrations
# run, while the Postgres target is still untouched. An install that has been
# unknowingly running on SQLite (the image used to leave NODE_ENV unset, so
# knexfile.js fell back to sqlite3 and ignored DB_HOST/DB_USER/DB_PASSWORD)
# keeps serving from its SQLite file instead of coming up against an empty
# Postgres. The exported value survives the `exec` below, so the migration
# runner and the server agree on the engine.
RESOLVED_DB_CLIENT="$(node scripts/resolve-db-engine.js)"
RESOLVER_STATUS=$?
# Exit 3 means two populated databases with no record of which is current
# (#1038). Starting either would hide the other's data, so stop here — the
# resolver has already printed what to do.
if [ "$RESOLVER_STATUS" = "3" ]; then
exit 1
fi
# Validate rather than trust: anything unexpected on stdout (a stray log line
# from a library that writes to the console) must not become DATABASE_CLIENT,
# which would break knexfile for every process that follows.
case "$RESOLVED_DB_CLIENT" in
pg|sqlite3)
export DATABASE_CLIENT="$RESOLVED_DB_CLIENT"
;;
"")
>&2 echo "Database engine resolver returned nothing; falling back to the configured client."
;;
*)
>&2 echo "Database engine resolver returned an unexpected value; ignoring it and falling back to the configured client."
;;
esac
# Run migrations (use safe runner in production). Invoked via node directly —
# the runtime image no longer ships npm (see Dockerfile: its bundled deps kept
# tripping CVE scanners while npm itself never runs in production).
echo "Running database migrations..."
if [ "$NODE_ENV" = "production" ]; then
node migrations/run-migrations-safe.js
else
node migrations/run-migrations.js
fi
# Execute the main command
exec "$@"