fix(cluster): keepalived Split-Brain + Boot-Race behoben — v1.2.101

Ursache der „WireGuard reißt immer wieder ab"-Abrisse war NICHT die UniFi,
sondern keepalived-Flapping im HA-Cluster: die VIP 89.163.205.100 (an der die
UniFi-Site-to-Site hängt) wanderte bei ~17 VRRP-Wahlen/Tag zwischen utm-1/utm-2
→ Tunnel-Abriss bei jeder Wahl.

Drei Bugs:
1) Firewall ließ VRRP (IP-Proto 112) zwischen den Cluster-Peers NICHT zu
   (policy drop). Adverts überlebten nur via conntrack-Reverse-Matching → bei
   conntrack-Ablauf gedroppt → Peer promotet sich → Split-Brain.
   Fix: ruleset.nft.tpl erlaubt `ip/ip6 ... vrrp saddr @peer_ipv4/6`;
   firewall.go nimmt zusätzlich hb_src_ip/hb_peer_ip aus cluster_settings ins
   Peer-Set (deckt den Heartbeat-Pfad 169.254.0.x ab).
2) Kein nopreempt → erholter Node riss die VIP sofort zurück (Flap-Back);
   aggressiver gw-Check (fall 2 → 10s-Blip = Failover).
   Fix: keepalived.conf.tpl mit `nopreempt` in VI_1+VI_HB, chk_gateway fall 2→5;
   keepalived.go setzt State immer BACKUP (nopreempt wirkt nur in BACKUP),
   Priorität 200/100 aus pg_role bleibt → deckt sich mit „manuelles Promote".
3) keepalived-Boot-Race: Unit startete vor vlan500 (nur After=network-online)
   → „interface vlan500 doesn't exist" → permanenter CONFIG-Crash ohne Recovery
   (keepalived nach Reboot tot). Fix: postinst legt Drop-in mit
   After=/Wants=edgeguard-interfaces.service + Restart=on-failure an.

Neuer Test internal/keepalived/keepalived_test.go (nopreempt/BACKUP/fall).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

VERSION

@@ -1 +1 @@
-1.2.100
+1.2.101

internal/firewall/firewall.go

@@ -177,14 +177,14 @@ type RuleLeg struct {
 // template just emits one nft line per "leg" of the cross-product.
 type ResolvedRule struct {
 	ID       int64
-	Action   string  // accept | drop | reject
+	Action   string // accept | drop | reject
 	Log      bool
 	Name     string
 	Priority int
 
-	SrcIfaces []string  // empty = any
+	SrcIfaces []string // empty = any
-	DstIfaces []string  // empty = any
+	DstIfaces []string // empty = any
-	SrcAddrs  []string  // each is an nft expression like "1.2.3.4" or "10.0.0.0/24" or "{ 1.2.3.4, 5.6.7.8 }"
+	SrcAddrs  []string // each is an nft expression like "1.2.3.4" or "10.0.0.0/24" or "{ 1.2.3.4, 5.6.7.8 }"
 	DstAddrs  []string
 	Services  []ResolvedService // empty = any
 	Comment   string
@@ -192,16 +192,16 @@ type ResolvedRule struct {
 
 // ResolvedNATRule is one nat-rule joined with iface-sets.
 type ResolvedNATRule struct {
-	ID       int64
+	ID                             int64
-	Kind     string // dnat | snat | masquerade
+	Kind                           string // dnat | snat | masquerade
-	Priority int
+	Priority                       int
-	InIfaces []string
+	InIfaces                       []string
-	OutIfaces []string
+	OutIfaces                      []string
-	Proto    string  // empty = any
+	Proto                          string // empty = any
-	SrcCIDR  string
+	SrcCIDR                        string
-	DstCIDR  string
+	DstCIDR                        string
-	DPortStart, DPortEnd int
+	DPortStart, DPortEnd           int
-	TargetAddr string
+	TargetAddr                     string
 	TargetPortStart, TargetPortEnd int
 	// L3 ist "ip" oder "ip6" — Adressfamilie der Regel (aus SrcCIDR/
 	// DstCIDR/TargetAddr abgeleitet). TargetHost ist TargetAddr, bei
@@ -209,14 +209,14 @@ type ResolvedNATRule struct {
 	// nft-dnat-Syntax.
 	L3         string
 	TargetHost string
-	Comment  string
+	Comment    string
 }
 
 // ResolvedService is one nft (proto, dport-spec) tuple.
 type ResolvedService struct {
-	Proto      string // tcp|udp|icmp|icmpv6
+	Proto     string // tcp|udp|icmp|icmpv6
-	PortStart  int    // 0 = no port match
+	PortStart int    // 0 = no port match
-	PortEnd    int
+	PortEnd   int
 }
 
 func (g *Generator) loadView(ctx context.Context) (*View, error) {
@@ -269,6 +269,31 @@ func (g *Generator) loadView(ctx context.Context) (*View, error) {
 	}
 	peerRows.Close()
 
+	// ── Heartbeat-IPs aus cluster_settings ins Peer-Set ──
+	// Der VRRP-Heartbeat (VI_HB) läuft über hb_src_ip/hb_peer_ip (z.B.
+	// 169.254.0.1/.2) — diese stehen NICHT in ha_nodes. Ohne sie würde die
+	// VRRP-Accept-Regel den Heartbeat-Pfad nicht abdecken. Best-effort:
+	// fehlt cluster_settings (Single-Node), bleibt es bei den ha_nodes-IPs.
+	var hbSrc, hbPeer *string
+	if err := g.Pool.QueryRow(ctx,
+		`SELECT hb_src_ip, hb_peer_ip FROM cluster_settings WHERE id = 1`).
+		Scan(&hbSrc, &hbPeer); err == nil {
+		for _, ip := range []*string{hbSrc, hbPeer} {
+			if ip == nil {
+				continue
+			}
+			parsed := net.ParseIP(*ip)
+			if parsed == nil {
+				continue
+			}
+			if parsed.To4() != nil {
+				view.PeerIPv4 = append(view.PeerIPv4, parsed.String())
+			} else {
+				view.PeerIPv6 = append(view.PeerIPv6, parsed.String())
+			}
+		}
+	}
+
 	// ── Lade Address-Objects + Groups → ID → ResolvedAddr-list ──
 	addrObjs, err := g.loadAddrObjects(ctx)
 	if err != nil {
@@ -777,15 +802,15 @@ ORDER BY priority DESC, id ASC`)
 	out := []ResolvedRule{}
 	for rows.Next() {
 		var (
-			id                   int64
+			id                 int64
-			name, action, com    string
+			name, action, com  string
-			pr                   int
+			pr                 int
-			log                  bool
+			log                bool
-			srcZone, dstZone     string
+			srcZone, dstZone   string
-			srcObjID, srcGrpID   *int64
+			srcObjID, srcGrpID *int64
-			dstObjID, dstGrpID   *int64
+			dstObjID, dstGrpID *int64
-			srcCIDR, dstCIDR     *string
+			srcCIDR, dstCIDR   *string
-			svcObjID, svcGrpID   *int64
+			svcObjID, svcGrpID *int64
 		)
 		if err := rows.Scan(
 			&id, &name, &pr, &action, &log, &com,
@@ -842,12 +867,12 @@ ORDER BY priority DESC, id ASC`)
 	out := []ResolvedNATRule{}
 	for rows.Next() {
 		var (
-			id          int64
+			id                                       int64
-			pr          int
+			pr                                       int
-			kind, com   string
+			kind, com                                string
 			inZone, outZone, proto, srcCIDR, dstCIDR *string
-			dpStart, dpEnd, tpStart, tpEnd int
+			dpStart, dpEnd, tpStart, tpEnd           int
-			targetAddr  string
+			targetAddr                               string
 		)
 		if err := rows.Scan(
 			&id, &pr, &kind, &com,
@@ -861,7 +886,7 @@ ORDER BY priority DESC, id ASC`)
 		r := ResolvedNATRule{
 			ID: id, Kind: kind, Priority: pr, Comment: com,
 			DPortStart: dpStart, DPortEnd: dpEnd,
-			TargetAddr: targetAddr,
+			TargetAddr:      targetAddr,
 			TargetPortStart: tpStart, TargetPortEnd: tpEnd,
 		}
 		if proto != nil {

internal/firewall/ruleset.nft.tpl

@@ -55,6 +55,13 @@ table inet edgeguard {
         tcp dport 5432 ip6 saddr @peer_ipv6 accept
         tcp dport 6379 ip  saddr @peer_ipv4 accept
         tcp dport 6379 ip6 saddr @peer_ipv6 accept
+        # Cluster-internal: VRRP-Advertisements (keepalived VIP-Failover, Proto 112).
+        # OHNE diese Regel überleben Adverts nur via conntrack-Reverse-Matching —
+        # läuft ein conntrack-Eintrag ab/wird geflusht, werden Adverts gedroppt →
+        # der Peer promotet sich → VIP-Flapping/Split-Brain. peer_ipv4/6 enthält
+        # Public- UND Heartbeat-IPs (ha_nodes + cluster_settings.hb_*).
+        ip  protocol vrrp ip  saddr @peer_ipv4 accept
+        ip6 nexthdr  vrrp ip6 saddr @peer_ipv6 accept
 
         # ── Service-Auto-Rules (DNS/Squid/WG/...) ──
         # Aus dem laufenden Service-State abgeleitet — Operator

internal/keepalived/keepalived.conf.tpl

@@ -16,7 +16,7 @@ vrrp_script chk_gateway {
     script "/usr/lib/edgeguard/keepalived-gw-check.sh {{ .GWCheckIP }}"
     interval 5
     weight -110
-    fall 2
+    fall 5
     rise 2
 }
 {{ end }}
@@ -34,6 +34,7 @@ vrrp_instance VI_1 {
     virtual_router_id {{ .RouterID }}
     priority {{ .Priority }}
     advert_int 1
+    nopreempt
 {{ if .SrcIP }}    unicast_src_ip {{ .SrcIP }}
     unicast_peer {
         {{ .PeerIP }}
@@ -60,6 +61,7 @@ vrrp_instance VI_HB {
     virtual_router_id {{ .HBRouterID }}
     priority {{ .Priority }}
     advert_int 1
+    nopreempt
 {{ if .HBSrcIP }}    unicast_src_ip {{ .HBSrcIP }}
     unicast_peer {
         {{ .HBPeerIP }}

internal/keepalived/keepalived.go

@@ -167,17 +167,19 @@ func (g *generator) buildView(cs *models.ClusterSettings, vips []VIPEntry, local
 		v.HBRouterID = 52
 	}
 
-	// pg_role=standby ist das härtere Signal — ein Standby-Node ist niemals
+	// State IMMER BACKUP: das Template setzt `nopreempt`, und nopreempt wirkt
-	// MASTER, auch wenn role='primary' noch aus dem Join-Prozess stammt.
+	// in keepalived NUR, wenn die Instanz im BACKUP-Zustand startet (bei state
-	// Reihenfolge: standby → BACKUP; sonst primary-Check.
+	// MASTER wird nopreempt ignoriert). Die Priorität entscheidet weiterhin die
+	// Initial-Election (primary=200 gewinnt), aber ein erholter Node reißt die
+	// VIP NICHT mehr zurück → kein Flap-Back / Split-Brain. Deckt sich mit der
+	// "kein Auto-Promote"-Philosophie: Promotion bleibt manuell.
+	// pg_role=standby ist das härtere Signal (Standby ist nie bevorzugter Node).
+	v.State = "BACKUP"
 	if local.PGRole == "standby" {
-		v.State = "BACKUP"
 		v.Priority = 100
 	} else if local.PGRole == "primary" || local.Role == "primary" {
-		v.State = "MASTER"
 		v.Priority = 200
 	} else {
-		v.State = "BACKUP"
 		v.Priority = 100
 	}
 

internal/keepalived/keepalived_test.go

@@ -0,0 +1,79 @@
+package keepalived
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+
+	"git.netcell-it.de/projekte/edgeguard-native/internal/models"
+)
+
+func render(t *testing.T, v View) string {
+	t.Helper()
+	var buf bytes.Buffer
+	if err := tpl.Execute(&buf, v); err != nil {
+		t.Fatalf("template execute: %v", err)
+	}
+	return buf.String()
+}
+
+// Split-Brain-Schutz: jede vrrp_instance MUSS `nopreempt` tragen, sonst reißt
+// ein erholter Node die VIP zurück → Flapping. nopreempt wirkt nur bei state
+// BACKUP — also muss auch der bevorzugte Node BACKUP starten.
+func testView() View {
+	return View{
+		State: "BACKUP", Interface: "eth0", RouterID: 51, Priority: 200,
+		SrcIP: "89.163.205.6", PeerIP: "89.163.205.8", AuthPass: "edgeguard",
+		VIPs:        []VIPEntry{{Address: "89.163.205.100", Prefix: 24, Device: "eth0"}},
+		HBInterface: "ens19", HBSrcIP: "169.254.0.1", HBPeerIP: "169.254.0.2", HBRouterID: 52,
+		GWCheckIP: "89.163.205.1",
+	}
+}
+
+func TestTemplateNopreemptOnBothInstances(t *testing.T) {
+	out := render(t, testView())
+	if n := strings.Count(out, "nopreempt"); n != 2 {
+		t.Fatalf("erwarte nopreempt in VI_1 UND VI_HB (2×), gefunden: %d\n%s", n, out)
+	}
+	if strings.Contains(out, "state MASTER") {
+		t.Fatalf("kein Node darf state MASTER starten (nopreempt würde ignoriert):\n%s", out)
+	}
+	if c := strings.Count(out, "state BACKUP"); c != 2 {
+		t.Fatalf("erwarte state BACKUP in beiden Instanzen, gefunden: %d", c)
+	}
+}
+
+// gw-Check darf nicht zu zucken (fall 5, nicht fall 2) — ein kurzer Upstream-
+// Blip soll keinen Failover erzwingen.
+func TestTemplateGatewayCheckNotTwitchy(t *testing.T) {
+	out := render(t, testView())
+	if !strings.Contains(out, "fall 5") {
+		t.Fatalf("chk_gateway sollte fall 5 nutzen:\n%s", out)
+	}
+}
+
+// buildView: State immer BACKUP, Priorität aus pg_role.
+func TestBuildViewStateAlwaysBackup(t *testing.T) {
+	g := &generator{localID: "n1"}
+	cs := &models.ClusterSettings{VRRPRouterID: 51}
+	pub := "89.163.205.6"
+	cases := []struct {
+		pgRole, role string
+		wantPrio     int
+	}{
+		{"primary", "primary", 200},
+		{"standby", "primary", 100},
+		{"", "primary", 200},
+		{"", "", 100},
+	}
+	for _, c := range cases {
+		local := &models.HANode{ID: "n1", PGRole: c.pgRole, Role: c.role, PublicIP: &pub}
+		v := g.buildView(cs, nil, local, nil)
+		if v.State != "BACKUP" {
+			t.Errorf("pg_role=%q role=%q: State=%q, erwarte immer BACKUP (nopreempt)", c.pgRole, c.role, v.State)
+		}
+		if v.Priority != c.wantPrio {
+			t.Errorf("pg_role=%q role=%q: Priority=%d, erwarte %d", c.pgRole, c.role, v.Priority, c.wantPrio)
+		}
+	}
+}

packaging/debian/edgeguard-api/DEBIAN/postinst

@@ -845,6 +845,25 @@ EOSQL
         # atomic-write (tempfile → rename) durchführen kann.
         install -d -m 0755 /etc/keepalived
         chown "$EG_USER":"$EG_USER" /etc/keepalived
+        # ── keepalived systemd drop-in: Boot-Race-Fix ──────────────────
+        # keepalived referenziert VIP-Devices (vlanXXX), die erst von
+        # edgeguard-interfaces.service angelegt werden. Ohne Ordering startet
+        # keepalived vor den VLANs (beide nur After=network-online.target) →
+        # "interface vlanX doesn't exist" → permanenter CONFIG-Crash OHNE
+        # Auto-Recovery (keepalived bleibt nach Reboot tot). After=/Wants=
+        # wartet auf die Interfaces; Restart=on-failure ist das Sicherheitsnetz.
+        install -d /etc/systemd/system/keepalived.service.d
+        cat > /etc/systemd/system/keepalived.service.d/10-edgeguard.conf <<'KEEPALIVEDDROPIN'
+[Unit]
+After=edgeguard-interfaces.service
+Wants=edgeguard-interfaces.service
+StartLimitIntervalSec=0
+
+[Service]
+Restart=on-failure
+RestartSec=3
+KEEPALIVEDDROPIN
+        systemctl daemon-reload
         sudo -n -u "$EG_USER" /usr/bin/edgeguard-ctl render-config --only=keepalived || true
         if [ -f /etc/keepalived/keepalived.conf ]; then
             systemctl enable keepalived >/dev/null 2>&1 || true