bugfix/1025-local-link-lost-connection (#34)

Fixes the bug that local link loses connection after a few minutes.

Openproject:
#1025
#608

Reviewed-on: #34

src/service/10-link-local-mgmt

@@ -5,6 +5,8 @@
 # using nmcli device modify (active session only, NOT saved to the profile).
 # The persistent profile always keeps ipv4.link-local=enabled so that
 # direct-connect (no DHCP) plug-ins always activate and trigger events.
+# Avahi is reloaded on each event — no /etc/avahi/hosts file, avahi uses
+# natural per-interface advertisement so each segment gets the right IP.
 #
 # Triggers: up, down, dhcp4-change on ethernet interfaces
 # Install to: /etc/NetworkManager/dispatcher.d/10-link-local-mgmt
@@ -12,47 +14,31 @@
 
 INTERFACE="$1"
 ACTION="$2"
-CONNECTION_NAME="${CONNECTION_ID:-}"
-
 # Only handle ethernet interfaces
 if [[ ! "$INTERFACE" =~ ^eth ]]; then
     exit 0
 fi
 
-# If CONNECTION_ID env var is not set, look up the active connection for this interface
-if [ -z "$CONNECTION_NAME" ]; then
-    CONNECTION_NAME=$(nmcli -t -f NAME,DEVICE connection show --active 2>/dev/null \
-        | grep ":${INTERFACE}$" | cut -d: -f1 | head -n1)
-    [ -z "$CONNECTION_NAME" ] && exit 0
-fi
-
-# Update /etc/avahi/hosts to point mDNS hostname at the best available DHCP address
-# across all ethernet interfaces (so Avahi doesn't advertise a link-local address).
-update_avahi() {
-    local hostname
-    hostname=$(hostname)
-    # Find first non-link-local IPv4 across all ethernet interfaces
-    local dhcp_ip
-    dhcp_ip=$(ip -4 addr show 2>/dev/null \
-        | grep -A5 ': eth' \
-        | grep -oP '(?<=inet\s)\d+(\.\d+){3}' \
-        | grep -v '^127\.' \
-        | grep -v '^169\.254\.' \
-        | head -n1)
-
-    if [ -n "$dhcp_ip" ]; then
-        mkdir -p /etc/avahi
-        echo "$dhcp_ip $hostname $hostname.local" > /etc/avahi/hosts
-        logger -t nm-link-local "Avahi: pinned $hostname -> $dhcp_ip"
-    else
-        rm -f /etc/avahi/hosts
-        logger -t nm-link-local "Avahi: removed hosts pin, using all addresses"
-    fi
-    systemctl restart avahi-daemon 2>/dev/null
+reload_avahi() {
+    systemctl reload avahi-daemon 2>/dev/null || systemctl restart avahi-daemon 2>/dev/null
+    logger -t nm-link-local "[$INTERFACE] $ACTION — avahi reloaded"
 }
 
 case "$ACTION" in
-    up|dhcp4-change)
+    up)
+        # On 'up' the interface may still carry a stale DHCP address from the previous
+        # session (NM hasn't cleaned it up yet). Reading ip-addr here is unreliable.
+        # Always re-enable link-local as a clean slate; let dhcp4-change suppress it
+        # later if a real DHCP lease is obtained.
+        logger -t nm-link-local "[$INTERFACE] Up — ensuring link-local active (clean slate)"
+        (sleep 2 && nmcli device modify "$INTERFACE" ipv4.link-local enabled 2>/dev/null \
+            && logger -t nm-link-local "[$INTERFACE] Link-local explicitly enabled on up") &
+        reload_avahi
+        ;;
+
+    dhcp4-change)
+        # dhcp4-change fires only when DHCP actually succeeds (new/renewed lease).
+        # At this point the DHCP IP is reliably present — safe to read and suppress link-local.
         DHCP_IP=$(ip -4 addr show "$INTERFACE" 2>/dev/null \
             | grep -oP '(?<=inet\s)\d+(\.\d+){3}' \
             | grep -v '^127\.' \
@@ -60,24 +46,19 @@ case "$ACTION" in
             | head -n1)
 
         if [ -n "$DHCP_IP" ]; then
-            logger -t nm-link-local "[$INTERFACE] DHCP $DHCP_IP detected — suppressing link-local (session only)"
-            # Use device modify (not connection modify) so the persistent profile keeps
-            # ipv4.link-local=enabled. This ensures direct-connect plug-ins always activate.
+            logger -t nm-link-local "[$INTERFACE] DHCP $DHCP_IP confirmed — suppressing link-local (session only)"
             # Run in background after a delay — nmcli blocks on NM, which is waiting for
             # this dispatcher to return, causing a deadlock if called synchronously.
             (sleep 2 && nmcli device modify "$INTERFACE" ipv4.link-local disabled 2>/dev/null \
                 && logger -t nm-link-local "[$INTERFACE] Link-local suppressed for current session") &
-        else
-            logger -t nm-link-local "[$INTERFACE] No DHCP on $INTERFACE — keeping link-local active"
         fi
-        update_avahi
+        reload_avahi
         ;;
 
     down)
-        # Profile always has ipv4.link-local=enabled so no action needed here.
-        # The suppression from device modify was session-only and is gone when the
-        # connection goes down.
-        logger -t nm-link-local "[$INTERFACE] Down — link-local will be active on next connect"
-        update_avahi
+        # NOTE: a carrier-change does NOT fully reset session-level 'device modify' state.
+        # The re-enable is therefore handled in the 'up' handler when no DHCP is detected.
+        logger -t nm-link-local "[$INTERFACE] Down — link-local will be re-enabled on next up without DHCP"
+        reload_avahi
         ;;
 esac

src/service/update_and_run_server_and_frontend.sh

@@ -8,13 +8,28 @@ set -e
 # Enable link-local for all wired ethernet connections
 while IFS=: read -r name type; do
     if [[ "$type" == *"ethernet"* ]]; then
-        echo "Enabling IPv4 link-local for connection: $name"
+        echo "Configuring connection: $name"
+        # link-local: always enabled so direct-connect (no DHCP) works immediately
         sudo nmcli connection modify "$name" ipv4.link-local enabled 2>/dev/null || echo "Failed to modify $name"
+        # may-fail=yes: do NOT tear down the connection when DHCP times out.
+        # Without this, NM declares ip-config-unavailable after the 45s DHCP timeout
+        # and enters a reconnect loop that causes ~1.5 min outages every ~45 seconds.
+        sudo nmcli connection modify "$name" ipv4.may-fail yes 2>/dev/null || echo "Failed to set may-fail on $name"
+        # Infinite DHCP timeout: NM keeps retrying DHCP in the background but never
+        # declares ip-config-unavailable. This prevents the 45s reconnect loop that
+        # kills the link-local address in direct-connect (no DHCP server) scenarios.
+        sudo nmcli connection modify "$name" ipv4.dhcp-timeout infinity 2>/dev/null || echo "Failed to set dhcp-timeout on $name"
         sudo nmcli connection up "$name" 2>/dev/null || echo "Failed to bring up $name"
     fi
 done < <(nmcli -t -f NAME,TYPE connection show)
 
 
+# Remove stale avahi hosts pin — this file overrides per-interface advertisement
+# and causes mDNS to always resolve to eth0's IP regardless of which interface
+# the query arrived on, breaking eth1 mDNS entirely.
+sudo rm -f /etc/avahi/hosts
+sudo systemctl restart avahi-daemon
+
 # Ensure Loopback is loaded with a fixed name and index 
 # Needed for dante
 # TODO image when we create the next image this should be part of it