fix: the web service starts once libpq can be found.

ADR.md

@@ -359,16 +359,22 @@ missing. PostgreSQL may also start after the services at boot.
 Decision: The web service makes its signing key and creates the refresh tokens' database itself,
 and the Router Manager its sessions database; neither `vpn-router-setup` nor a maintainer script
 creates them. Each service does it before it listens, retrying every 15 seconds and logging each
-failure, then listens. The first step is the PostgreSQL driver: psycopg finds libpq through the
+failure, then listens. `vpnrouter_auth` turns every failure of the database into
-linker cache, which the package manager refreshes only at the end of an installation, so a
-service started during the installation cannot import it yet. The Manager imports psycopg only in
-its retry loop. On the router, `vpnrouter_auth` alone imports it, in `driver()`, which the web
-service's retry loop calls first, and turns a failed import and every database failure into
 `DatabaseError`, the only database exception the web service and `vpn-router-user` handle.
 
+psycopg (Ubuntu ships only its pure-Python implementation) finds libpq through the dynamic linker
+cache, which dpkg refreshes with its deferred `ldconfig` only at the end of an installation, so a
+service started during the installation cannot import it yet; a failed import cannot be repeated
+in the same process either, as psycopg caches the lookup. The services therefore import psycopg
+plainly, and their units wait instead: `ExecStartPre=wait-for-libpq`, a shell script of each
+package, waits until `ldconfig -p` lists `libpq.so.5`, within `TimeoutStartSec=5min`. The
+maintainer scripts start the units with `systemctl --no-block` (debhelper's start is turned off),
+because a start that waited would hold dpkg before its `ldconfig`.
+
 Consequences: The order of installation and of the start of PostgreSQL and of the linker cache
-does not matter; a driver that cannot be loaded or a database that cannot be created is retried
+does not matter; a database that cannot be created is retried until the cause is gone, and the
-until the cause is gone, and the reason is in the service's log, without restarts. The web unit may write `/etc/vpn-router` for the key and the database password.
+reason is in the service's log. Outside an installation the wait ends at once. The web unit may
+write `/etc/vpn-router` for the key and the database password.
 
 ## ADR-17: Bootstrap diagnostics run only until the first setup succeeds
 

debian-package/debian/changelog

@@ -1,3 +1,14 @@
+vpn-router (0.9.2-1) unstable; urgency=medium
+
+  fix: the web service starts once libpq can be found.
+
+  * The web service unit waits, with wait-for-libpq, until libpq is in the
+    linker cache before it starts the service, for at most 5 minutes; the
+    service imports psycopg at start again.
+  * The package starts the web service without waiting for it.
+
+ -- Slawomir Koszewski <slawek@koszewscy.waw.pl>  Sun, 25 Oct 2026 06:00:00 +0200
+
 vpn-router (0.9.1-1) unstable; urgency=medium
 
   fix: the web service waits for the PostgreSQL driver instead of failing at start.

debian-package/debian/install

@@ -15,6 +15,7 @@ src/usr/lib/vpn-router/docker-rules                   usr/lib/vpn-router/
 src/usr/lib/vpn-router/setup                          usr/lib/vpn-router/
 src/usr/lib/vpn-router/web                            usr/lib/vpn-router/
 src/usr/lib/vpn-router/bootstrap                      usr/lib/vpn-router/
+src/usr/lib/vpn-router/wait-for-libpq                 usr/lib/vpn-router/
 src/usr/lib/vpn-router/generate-config                usr/lib/vpn-router/
 src/usr/lib/vpn-router/vpnrouter_platforms/*.py       usr/lib/vpn-router/vpnrouter_platforms/
 src/usr/share/vpn-router/templates/*                  usr/share/vpn-router/templates/

debian-package/debian/postinst

@@ -70,6 +70,12 @@ case "$1" in
         # The firewall is not touched here. Enabling it, allowing SSH and
         # setting the forward policy all belong to vpn-router-setup, which does
         # them only once the configuration says this host is a router.
+
+        # The web service waits for libpq in the linker cache, which dpkg
+        # refreshes after this script, so it is started without waiting for it.
+        if [ -d /run/systemd/system ]; then
+            systemctl --no-block restart vpn-router-web.service >/dev/null 2>&1 || true
+        fi
         ;;
 esac
 

debian-package/debian/rules

@@ -5,5 +5,5 @@
 override_dh_installsystemd:
 	dh_installsystemd --name=vpn-router-setup
 	dh_installsystemd --name=vpn-router-docker
-	dh_installsystemd --name=vpn-router-web
+	dh_installsystemd --name=vpn-router-web --no-start
 	dh_installsystemd --name=vpn-router-bootstrap --no-enable --no-start

debian-package/debian/vpn-router.vpn-router-web.service

@@ -6,7 +6,9 @@ Wants=network-online.target
 
 [Service]
 Type=simple
+ExecStartPre=/usr/lib/vpn-router/wait-for-libpq
 ExecStart=/usr/bin/python3 /usr/lib/vpn-router/web
+TimeoutStartSec=5min
 Restart=on-failure
 RestartSec=15
 NoNewPrivileges=yes

debian-package/src/usr/lib/vpn-router/vpnrouter_auth.py

@@ -5,9 +5,7 @@ router and that never leaves it. /auth issues an access token naming the user in
 a refresh token valid REFRESH_SECONDS; vpn-router-token issues tokens with the admin claim and
 any lifetime. A token without the admin claim must expire within MAX_LIFETIME seconds.
 
-psycopg is imported by driver() when the database is used, not when the module is loaded, so the
+Every failure of the database raises DatabaseError, the only database exception callers see.
-module loads while libpq cannot be found yet. A driver that cannot be imported and every
-failure of the database raise DatabaseError, the only database exception callers see.
 
 Refresh tokens are random strings kept in PostgreSQL as their SHA-256, in the database of
 web.database, by default the vpn-router database on this host that provision() creates on the
@@ -24,6 +22,8 @@ import secrets
 import subprocess
 import time
 
+import psycopg
+
 WEB_KEY = pathlib.Path('/etc/vpn-router/web.key')
 DATABASE_SECRET = pathlib.Path('/etc/vpn-router/database.secret')
 
@@ -44,7 +44,7 @@ SCHEMA = 'CREATE TABLE IF NOT EXISTS refresh_tokens (hash text PRIMARY KEY, name
 
 
 class DatabaseError(Exception):
-    """The PostgreSQL driver cannot be loaded, or the database of the refresh tokens failed."""
+    """The database of the refresh tokens failed."""
 
 
 def key():
@@ -117,15 +117,6 @@ def admin_token(seconds):
     return {'token': sign({'admin': True, 'iat': now, 'exp': now + seconds}, secret), 'expires': now + seconds}
 
 
-def driver():
-    """The psycopg module; DatabaseError while libpq cannot be found."""
-    try:
-        import psycopg
-    except ImportError as exc:
-        raise DatabaseError(f'the PostgreSQL driver cannot be loaded: {exc}') from exc
-    return psycopg
-
-
 def connection_string(cfg):
     """web.database, or the database provision() creates; its password, of URL-safe characters, is written unquoted."""
     configured = cfg['web']['database'].strip()
@@ -138,7 +129,6 @@ def connection_string(cfg):
 @contextlib.contextmanager
 def database(cfg):
     """A connection to the database of the refresh tokens, with their table, closed at the end; DatabaseError on any failure."""
-    psycopg = driver()
     try:
         with psycopg.connect(connection_string(cfg), autocommit=True) as connection:
             connection.execute(SCHEMA)

debian-package/src/usr/lib/vpn-router/wait-for-libpq

@@ -0,0 +1,12 @@
+#!/bin/sh
+# Wait until libpq is in the dynamic linker cache, where psycopg looks it up.
+# ExecStartPre= of vpn-router-web.service; TimeoutStartSec= of the unit bounds the wait.
+
+reported=
+until ldconfig -p | grep -q '^[[:space:]]*libpq\.so\.5 '; do
+    if [ -z "$reported" ]; then
+        echo 'waiting for libpq.so.5 in the linker cache'
+        reported=yes
+    fi
+    sleep 2
+done

debian-package/src/usr/lib/vpn-router/web

@@ -30,11 +30,10 @@ default the address the service listens on, is the last entry of its X-Forwarded
 Docker is installed, and the state of site-to-site, road-warrior (IPsec P2S), OpenVPN and
 WireGuard. No configuration data, key or address is shown.
 
-The service first initializes, without listening: it loads the PostgreSQL driver
+The service first initializes, without listening: it makes the signing key and the refresh
-(vpnrouter_auth.driver; psycopg finds libpq through the linker cache, which the package manager
-refreshes only at the end of an installation), then it makes the signing key and the refresh
 tokens' database (vpnrouter_auth), retrying every INIT_RETRY_SECONDS until both are ready and
-logging each failure. Then it listens.
+logging each failure. Then it listens. Its unit starts it only once wait-for-libpq has found libpq
+in the linker cache, where psycopg looks it up.
 
 The service needs nothing else to start and does not end on its own. Every DISCOVERY_SECONDS it
 works out the internal address from /etc/vpn-router/vpn-router.json as general.mode directs,
@@ -697,13 +696,11 @@ def log(message):
 
 
 def initialize():
-    """Load the PostgreSQL driver, make the signing key and the refresh tokens' database, retrying every INIT_RETRY_SECONDS until all are ready."""
+    """Make the signing key and the refresh tokens' database, retrying every INIT_RETRY_SECONDS until both are ready."""
     reported = None
     while True:
-        phase = 'loading the PostgreSQL driver'
-        try:
-            vpnrouter_auth.driver()
         phase = 'reading the configuration'
+        try:
             cfg = vpnrouter.load_config()
             phase = 'making the signing key'
             vpnrouter_auth.ensure_key()