Trust wg interface in firewalld to bypass owner-flagged chains

WGIface.Close() took w.mu and held it across w.tun.Close(). The
underlying wireguard-go device waits for its send/receive goroutines to
drain before Close() returns, and some of those goroutines re-enter
WGIface during shutdown. In particular, the userspace packet filter DNS
hook in client/internal/dns.ServiceViaMemory.filterDNSTraffic calls
s.wgInterface.GetDevice() on every packet, which also needs w.mu. With
the Close-side holding the mutex, the read goroutine blocks in
GetDevice and Close waits forever for that goroutine to exit:

  goroutine N (TestDNSPermanent_updateUpstream):
    WGIface.Close -> holds w.mu -> tun.Close -> sync.WaitGroup.Wait
  goroutine M (wireguard read routine):
    FilteredDevice.Read -> filterOutbound -> udpHooksDrop ->
    filterDNSTraffic.func1 -> WGIface.GetDevice -> sync.Mutex.Lock

This surfaces as a 5 minute test timeout on the macOS Client/Unit
CI job (panic: test timed out after 5m0s, running tests:
TestDNSPermanent_updateUpstream).

Release w.mu before calling w.tun.Close(). The other Close steps
(wgProxyFactory.Free, waitUntilRemoved, Destroy) do not mutate any
fields guarded by w.mu beyond what Free() already does, so the lock
is not needed once the tun has started shutting down. A new unit test
in iface_close_test.go uses a fake WGTunDevice to reproduce the
deadlock deterministically without requiring CAP_NET_ADMIN.

client/firewall/firewalld/firewalld.go

@@ -0,0 +1,11 @@
+// Package firewalld integrates with the firewalld daemon so NetBird can place
+// its wg interface into firewalld's "trusted" zone. This is required because
+// firewalld's nftables chains are created with NFT_CHAIN_OWNER on recent
+// versions, which returns EPERM to any other process that tries to insert
+// rules into them. The workaround mirrors what Tailscale does: let firewalld
+// itself add the accept rules to its own chains by trusting the interface.
+package firewalld
+
+// TrustedZone is the firewalld zone name used for interfaces whose traffic
+// should bypass firewalld filtering.
+const TrustedZone = "trusted"

client/firewall/firewalld/firewalld_linux.go

@@ -0,0 +1,260 @@
+//go:build linux
+
+package firewalld
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os/exec"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/godbus/dbus/v5"
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	dbusDest      = "org.fedoraproject.FirewallD1"
+	dbusPath      = "/org/fedoraproject/FirewallD1"
+	dbusRootIface = "org.fedoraproject.FirewallD1"
+	dbusZoneIface = "org.fedoraproject.FirewallD1.zone"
+
+	errZoneAlreadySet = "ZONE_ALREADY_SET"
+	errAlreadyEnabled = "ALREADY_ENABLED"
+	errUnknownIface   = "UNKNOWN_INTERFACE"
+	errNotEnabled     = "NOT_ENABLED"
+
+	// callTimeout bounds each individual DBus or firewall-cmd invocation.
+	// A fresh context is created for each call so a slow DBus probe can't
+	// exhaust the deadline before the firewall-cmd fallback gets to run.
+	callTimeout = 3 * time.Second
+)
+
+var (
+	errDBusUnavailable = errors.New("firewalld dbus unavailable")
+
+	// trustLogOnce ensures the "added to trusted zone" message is logged at
+	// Info level only for the first successful add per process; repeat adds
+	// from other init paths are quieter.
+	trustLogOnce sync.Once
+
+	parentCtxMu sync.RWMutex
+	parentCtx   context.Context = context.Background()
+)
+
+// SetParentContext installs a parent context whose cancellation aborts any
+// in-flight TrustInterface call. It does not affect UntrustInterface, which
+// always uses a fresh Background-rooted timeout so cleanup can still run
+// during engine shutdown when the engine context is already cancelled.
+func SetParentContext(ctx context.Context) {
+	parentCtxMu.Lock()
+	parentCtx = ctx
+	parentCtxMu.Unlock()
+}
+
+func getParentContext() context.Context {
+	parentCtxMu.RLock()
+	defer parentCtxMu.RUnlock()
+	return parentCtx
+}
+
+// TrustInterface places iface into firewalld's trusted zone if firewalld is
+// running. It is idempotent and best-effort: errors are returned so callers
+// can log, but a non-running firewalld is not an error. Only the first
+// successful call per process logs at Info. Respects the parent context set
+// via SetParentContext so startup-time cancellation unblocks it.
+func TrustInterface(iface string) error {
+	parent := getParentContext()
+	if !isRunning(parent) {
+		return nil
+	}
+	if err := addTrusted(parent, iface); err != nil {
+		return fmt.Errorf("add %s to firewalld trusted zone: %w", iface, err)
+	}
+	trustLogOnce.Do(func() {
+		log.Infof("added %s to firewalld trusted zone", iface)
+	})
+	log.Debugf("firewalld: ensured %s is in trusted zone", iface)
+	return nil
+}
+
+// UntrustInterface removes iface from firewalld's trusted zone if firewalld
+// is running. Idempotent. Uses a Background-rooted timeout so it still runs
+// during shutdown after the engine context has been cancelled.
+func UntrustInterface(iface string) error {
+	if !isRunning(context.Background()) {
+		return nil
+	}
+	if err := removeTrusted(context.Background(), iface); err != nil {
+		return fmt.Errorf("remove %s from firewalld trusted zone: %w", iface, err)
+	}
+	return nil
+}
+
+func newCallContext(parent context.Context) (context.Context, context.CancelFunc) {
+	return context.WithTimeout(parent, callTimeout)
+}
+
+func isRunning(parent context.Context) bool {
+	ctx, cancel := newCallContext(parent)
+	ok, err := isRunningDBus(ctx)
+	cancel()
+	if err == nil {
+		return ok
+	}
+	if errors.Is(err, errDBusUnavailable) || errors.Is(err, context.DeadlineExceeded) {
+		ctx, cancel = newCallContext(parent)
+		defer cancel()
+		return isRunningCLI(ctx)
+	}
+	return false
+}
+
+func addTrusted(parent context.Context, iface string) error {
+	ctx, cancel := newCallContext(parent)
+	err := addDBus(ctx, iface)
+	cancel()
+	if err == nil {
+		return nil
+	}
+	if !errors.Is(err, errDBusUnavailable) {
+		log.Debugf("firewalld: dbus add failed, falling back to firewall-cmd: %v", err)
+	}
+	ctx, cancel = newCallContext(parent)
+	defer cancel()
+	return addCLI(ctx, iface)
+}
+
+func removeTrusted(parent context.Context, iface string) error {
+	ctx, cancel := newCallContext(parent)
+	err := removeDBus(ctx, iface)
+	cancel()
+	if err == nil {
+		return nil
+	}
+	if !errors.Is(err, errDBusUnavailable) {
+		log.Debugf("firewalld: dbus remove failed, falling back to firewall-cmd: %v", err)
+	}
+	ctx, cancel = newCallContext(parent)
+	defer cancel()
+	return removeCLI(ctx, iface)
+}
+
+func isRunningDBus(ctx context.Context) (bool, error) {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return false, fmt.Errorf("%w: %v", errDBusUnavailable, err)
+	}
+	obj := conn.Object(dbusDest, dbusPath)
+
+	var zone string
+	if err := obj.CallWithContext(ctx, dbusRootIface+".getDefaultZone", 0).Store(&zone); err != nil {
+		return false, fmt.Errorf("firewalld getDefaultZone: %w", err)
+	}
+	return true, nil
+}
+
+func isRunningCLI(ctx context.Context) bool {
+	if _, err := exec.LookPath("firewall-cmd"); err != nil {
+		return false
+	}
+	return exec.CommandContext(ctx, "firewall-cmd", "--state").Run() == nil
+}
+
+func addDBus(ctx context.Context, iface string) error {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return fmt.Errorf("%w: %v", errDBusUnavailable, err)
+	}
+	obj := conn.Object(dbusDest, dbusPath)
+
+	call := obj.CallWithContext(ctx, dbusZoneIface+".addInterface", 0, TrustedZone, iface)
+	if call.Err == nil {
+		return nil
+	}
+
+	if dbusErrContains(call.Err, errAlreadyEnabled) {
+		return nil
+	}
+
+	if dbusErrContains(call.Err, errZoneAlreadySet) {
+		move := obj.CallWithContext(ctx, dbusZoneIface+".changeZoneOfInterface", 0, TrustedZone, iface)
+		if move.Err != nil {
+			return fmt.Errorf("firewalld changeZoneOfInterface: %w", move.Err)
+		}
+		return nil
+	}
+
+	return fmt.Errorf("firewalld addInterface: %w", call.Err)
+}
+
+func removeDBus(ctx context.Context, iface string) error {
+	conn, err := dbus.SystemBus()
+	if err != nil {
+		return fmt.Errorf("%w: %v", errDBusUnavailable, err)
+	}
+	obj := conn.Object(dbusDest, dbusPath)
+
+	call := obj.CallWithContext(ctx, dbusZoneIface+".removeInterface", 0, TrustedZone, iface)
+	if call.Err == nil {
+		return nil
+	}
+
+	if dbusErrContains(call.Err, errUnknownIface) || dbusErrContains(call.Err, errNotEnabled) {
+		return nil
+	}
+
+	return fmt.Errorf("firewalld removeInterface: %w", call.Err)
+}
+
+func addCLI(ctx context.Context, iface string) error {
+	if _, err := exec.LookPath("firewall-cmd"); err != nil {
+		return fmt.Errorf("firewall-cmd not available: %w", err)
+	}
+
+	// --change-interface (no --permanent) binds the interface for the
+	// current runtime only; we do not want membership to persist across
+	// reboots because netbird re-asserts it on every startup.
+	out, err := exec.CommandContext(ctx,
+		"firewall-cmd", "--zone="+TrustedZone, "--change-interface="+iface,
+	).CombinedOutput()
+	if err != nil {
+		return fmt.Errorf("firewall-cmd change-interface: %w: %s", err, strings.TrimSpace(string(out)))
+	}
+	return nil
+}
+
+func removeCLI(ctx context.Context, iface string) error {
+	if _, err := exec.LookPath("firewall-cmd"); err != nil {
+		return fmt.Errorf("firewall-cmd not available: %w", err)
+	}
+
+	out, err := exec.CommandContext(ctx,
+		"firewall-cmd", "--zone="+TrustedZone, "--remove-interface="+iface,
+	).CombinedOutput()
+	if err != nil {
+		msg := strings.TrimSpace(string(out))
+		if strings.Contains(msg, errUnknownIface) || strings.Contains(msg, errNotEnabled) {
+			return nil
+		}
+		return fmt.Errorf("firewall-cmd remove-interface: %w: %s", err, msg)
+	}
+	return nil
+}
+
+func dbusErrContains(err error, code string) bool {
+	if err == nil {
+		return false
+	}
+	var de dbus.Error
+	if errors.As(err, &de) {
+		for _, b := range de.Body {
+			if s, ok := b.(string); ok && strings.Contains(s, code) {
+				return true
+			}
+		}
+	}
+	return strings.Contains(err.Error(), code)
+}

client/firewall/firewalld/firewalld_linux_test.go

@@ -0,0 +1,49 @@
+//go:build linux
+
+package firewalld
+
+import (
+	"errors"
+	"testing"
+
+	"github.com/godbus/dbus/v5"
+)
+
+func TestDBusErrContains(t *testing.T) {
+	tests := []struct {
+		name string
+		err  error
+		code string
+		want bool
+	}{
+		{"nil error", nil, errZoneAlreadySet, false},
+		{"plain error match", errors.New("ZONE_ALREADY_SET: wt0"), errZoneAlreadySet, true},
+		{"plain error miss", errors.New("something else"), errZoneAlreadySet, false},
+		{
+			"dbus.Error body match",
+			dbus.Error{Name: "org.fedoraproject.FirewallD1.Exception", Body: []any{"ZONE_ALREADY_SET: wt0"}},
+			errZoneAlreadySet,
+			true,
+		},
+		{
+			"dbus.Error body miss",
+			dbus.Error{Name: "org.fedoraproject.FirewallD1.Exception", Body: []any{"INVALID_INTERFACE"}},
+			errAlreadyEnabled,
+			false,
+		},
+		{
+			"dbus.Error non-string body falls back to Error()",
+			dbus.Error{Name: "x", Body: []any{123}},
+			"x",
+			true,
+		},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := dbusErrContains(tc.err, tc.code)
+			if got != tc.want {
+				t.Fatalf("dbusErrContains(%v, %q) = %v; want %v", tc.err, tc.code, got, tc.want)
+			}
+		})
+	}
+}

client/firewall/firewalld/firewalld_other.go

@@ -0,0 +1,25 @@
+//go:build !linux
+
+package firewalld
+
+import "context"
+
+// SetParentContext is a no-op on non-Linux platforms because firewalld only
+// runs on Linux.
+func SetParentContext(context.Context) {
+	// intentionally empty: firewalld is a Linux-only daemon
+}
+
+// TrustInterface is a no-op on non-Linux platforms because firewalld only
+// runs on Linux.
+func TrustInterface(string) error {
+	// intentionally empty: firewalld is a Linux-only daemon
+	return nil
+}
+
+// UntrustInterface is a no-op on non-Linux platforms because firewalld only
+// runs on Linux.
+func UntrustInterface(string) error {
+	// intentionally empty: firewalld is a Linux-only daemon
+	return nil
+}

client/firewall/iptables/manager_linux.go

@@ -12,6 +12,7 @@ import (
 	log "github.com/sirupsen/logrus"
 
 	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/client/firewall/firewalld"
 	firewall "github.com/netbirdio/netbird/client/firewall/manager"
 	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
@@ -86,6 +87,12 @@ func (m *Manager) Init(stateManager *statemanager.Manager) error {
 		log.Warnf("raw table not available, notrack rules will be disabled: %v", err)
 	}
 
+	// Trust after all fatal init steps so a later failure doesn't leave the
+	// interface in firewalld's trusted zone without a corresponding Close.
+	if err := firewalld.TrustInterface(m.wgIface.Name()); err != nil {
+		log.Warnf("failed to trust interface in firewalld: %v", err)
+	}
+
 	// persist early to ensure cleanup of chains
 	go func() {
 		if err := stateManager.PersistState(context.Background()); err != nil {
@@ -191,6 +198,12 @@ func (m *Manager) Close(stateManager *statemanager.Manager) error {
 		merr = multierror.Append(merr, fmt.Errorf("reset router: %w", err))
 	}
 
+	// Appending to merr intentionally blocks DeleteState below so ShutdownState
+	// stays persisted and the crash-recovery path retries firewalld cleanup.
+	if err := firewalld.UntrustInterface(m.wgIface.Name()); err != nil {
+		merr = multierror.Append(merr, err)
+	}
+
 	// attempt to delete state only if all other operations succeeded
 	if merr == nil {
 		if err := stateManager.DeleteState(&ShutdownState{}); err != nil {
@@ -217,6 +230,11 @@ func (m *Manager) AllowNetbird() error {
 	if err != nil {
 		return fmt.Errorf("allow netbird interface traffic: %w", err)
 	}
+
+	if err := firewalld.TrustInterface(m.wgIface.Name()); err != nil {
+		log.Warnf("failed to trust interface in firewalld: %v", err)
+	}
+
 	return nil
 }
 

client/firewall/nftables/manager_linux.go

@@ -14,6 +14,7 @@ import (
 	log "github.com/sirupsen/logrus"
 	"golang.org/x/sys/unix"
 
+	"github.com/netbirdio/netbird/client/firewall/firewalld"
 	firewall "github.com/netbirdio/netbird/client/firewall/manager"
 	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
@@ -217,6 +218,10 @@ func (m *Manager) AllowNetbird() error {
 		return fmt.Errorf("flush allow input netbird rules: %w", err)
 	}
 
+	if err := firewalld.TrustInterface(m.wgIface.Name()); err != nil {
+		log.Warnf("failed to trust interface in firewalld: %v", err)
+	}
+
 	return nil
 }
 

client/firewall/nftables/router_linux.go

@@ -19,6 +19,7 @@ import (
 	"golang.org/x/sys/unix"
 
 	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/client/firewall/firewalld"
 	firewall "github.com/netbirdio/netbird/client/firewall/manager"
 	nbid "github.com/netbirdio/netbird/client/internal/acl/id"
 	"github.com/netbirdio/netbird/client/internal/routemanager/ipfwdstate"
@@ -40,6 +41,8 @@ const (
 	chainNameForward       = "FORWARD"
 	chainNameMangleForward = "netbird-mangle-forward"
 
+	firewalldTableName = "firewalld"
+
 	userDataAcceptForwardRuleIif = "frwacceptiif"
 	userDataAcceptForwardRuleOif = "frwacceptoif"
 	userDataAcceptInputRule      = "inputaccept"
@@ -133,6 +136,10 @@ func (r *router) Reset() error {
 		merr = multierror.Append(merr, fmt.Errorf("remove accept filter rules: %w", err))
 	}
 
+	if err := firewalld.UntrustInterface(r.wgIface.Name()); err != nil {
+		merr = multierror.Append(merr, err)
+	}
+
 	if err := r.removeNatPreroutingRules(); err != nil {
 		merr = multierror.Append(merr, fmt.Errorf("remove filter prerouting rules: %w", err))
 	}
@@ -280,6 +287,10 @@ func (r *router) createContainers() error {
 		log.Errorf("failed to add accept rules for the forward chain: %s", err)
 	}
 
+	if err := firewalld.TrustInterface(r.wgIface.Name()); err != nil {
+		log.Warnf("failed to trust interface in firewalld: %v", err)
+	}
+
 	if err := r.refreshRulesMap(); err != nil {
 		log.Errorf("failed to refresh rules: %s", err)
 	}
@@ -1319,6 +1330,13 @@ func (r *router) isExternalChain(chain *nftables.Chain) bool {
 		return false
 	}
 
+	// Skip firewalld-owned chains. Firewalld creates its chains with the
+	// NFT_CHAIN_OWNER flag, so inserting rules into them returns EPERM.
+	// We delegate acceptance to firewalld by trusting the interface instead.
+	if chain.Table.Name == firewalldTableName {
+		return false
+	}
+
 	// Skip all iptables-managed tables in the ip family
 	if chain.Table.Family == nftables.TableFamilyIPv4 && isIptablesTable(chain.Table.Name) {
 		return false

client/firewall/uspfilter/allow_netbird.go

@@ -3,6 +3,9 @@
 package uspfilter
 
 import (
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/firewall/firewalld"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
 )
 
@@ -16,6 +19,9 @@ func (m *Manager) Close(stateManager *statemanager.Manager) error {
 	if m.nativeFirewall != nil {
 		return m.nativeFirewall.Close(stateManager)
 	}
+	if err := firewalld.UntrustInterface(m.wgIface.Name()); err != nil {
+		log.Warnf("failed to untrust interface in firewalld: %v", err)
+	}
 	return nil
 }
 
@@ -24,5 +30,8 @@ func (m *Manager) AllowNetbird() error {
 	if m.nativeFirewall != nil {
 		return m.nativeFirewall.AllowNetbird()
 	}
+	if err := firewalld.TrustInterface(m.wgIface.Name()); err != nil {
+		log.Warnf("failed to trust interface in firewalld: %v", err)
+	}
 	return nil
 }

client/firewall/uspfilter/common/iface.go

@@ -9,6 +9,7 @@ import (
 
 // IFaceMapper defines subset methods of interface required for manager
 type IFaceMapper interface {
+	Name() string
 	SetFilter(device.PacketFilter) error
 	Address() wgaddr.Address
 	GetWGDevice() *wgdevice.Device

client/firewall/uspfilter/filter_test.go

@@ -31,12 +31,20 @@ var logger = log.NewFromLogrus(logrus.StandardLogger())
 var flowLogger = netflow.NewManager(nil, []byte{}, nil).GetLogger()
 
 type IFaceMock struct {
+	NameFunc        func() string
 	SetFilterFunc   func(device.PacketFilter) error
 	AddressFunc     func() wgaddr.Address
 	GetWGDeviceFunc func() *wgdevice.Device
 	GetDeviceFunc   func() *device.FilteredDevice
 }
 
+func (i *IFaceMock) Name() string {
+	if i.NameFunc == nil {
+		return "wgtest"
+	}
+	return i.NameFunc()
+}
+
 func (i *IFaceMock) GetWGDevice() *wgdevice.Device {
 	if i.GetWGDeviceFunc == nil {
 		return nil

client/iface/iface.go

@@ -217,7 +217,6 @@ func (w *WGIface) RemoveAllowedIP(peerKey string, allowedIP netip.Prefix) error
 // Close closes the tunnel interface
 func (w *WGIface) Close() error {
 	w.mu.Lock()
-	defer w.mu.Unlock()
 
 	var result *multierror.Error
 
@@ -225,7 +224,15 @@ func (w *WGIface) Close() error {
 		result = multierror.Append(result, fmt.Errorf("failed to free WireGuard proxy: %w", err))
 	}
 
-	if err := w.tun.Close(); err != nil {
+	// Release w.mu before calling w.tun.Close(): the underlying
+	// wireguard-go device.Close() waits for its send/receive goroutines
+	// to drain. Some of those goroutines re-enter WGIface methods that
+	// take w.mu (e.g. the packet filter DNS hook calls GetDevice()), so
+	// holding the mutex here would deadlock the shutdown path.
+	tun := w.tun
+	w.mu.Unlock()
+
+	if err := tun.Close(); err != nil {
 		result = multierror.Append(result, fmt.Errorf("failed to close wireguard interface %s: %w", w.Name(), err))
 	}
 

client/iface/iface_close_test.go

@@ -0,0 +1,113 @@
+//go:build !android
+
+package iface
+
+import (
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	wgdevice "golang.zx2c4.com/wireguard/device"
+	"golang.zx2c4.com/wireguard/tun/netstack"
+
+	"github.com/netbirdio/netbird/client/iface/device"
+	"github.com/netbirdio/netbird/client/iface/udpmux"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/iface/wgproxy"
+)
+
+// fakeTunDevice implements WGTunDevice and lets the test control when
+// Close() returns. It mimics the wireguard-go shutdown path, which blocks
+// until its goroutines drain. Some of those goroutines (e.g. the packet
+// filter DNS hook in client/internal/dns) call back into WGIface, so if
+// WGIface.Close() held w.mu across tun.Close() the shutdown would
+// deadlock.
+type fakeTunDevice struct {
+	closeStarted chan struct{}
+	unblockClose chan struct{}
+}
+
+func (f *fakeTunDevice) Create() (device.WGConfigurer, error) {
+	return nil, errors.New("not implemented")
+}
+func (f *fakeTunDevice) Up() (*udpmux.UniversalUDPMuxDefault, error) {
+	return nil, errors.New("not implemented")
+}
+func (f *fakeTunDevice) UpdateAddr(wgaddr.Address) error      { return nil }
+func (f *fakeTunDevice) WgAddress() wgaddr.Address            { return wgaddr.Address{} }
+func (f *fakeTunDevice) MTU() uint16                          { return DefaultMTU }
+func (f *fakeTunDevice) DeviceName() string                   { return "nb-close-test" }
+func (f *fakeTunDevice) FilteredDevice() *device.FilteredDevice { return nil }
+func (f *fakeTunDevice) Device() *wgdevice.Device             { return nil }
+func (f *fakeTunDevice) GetNet() *netstack.Net                { return nil }
+func (f *fakeTunDevice) GetICEBind() device.EndpointManager   { return nil }
+
+func (f *fakeTunDevice) Close() error {
+	close(f.closeStarted)
+	<-f.unblockClose
+	return nil
+}
+
+type fakeProxyFactory struct{}
+
+func (fakeProxyFactory) GetProxy() wgproxy.Proxy { return nil }
+func (fakeProxyFactory) GetProxyPort() uint16    { return 0 }
+func (fakeProxyFactory) Free() error             { return nil }
+
+// TestWGIface_CloseReleasesMutexBeforeTunClose guards against a deadlock
+// that surfaces as a macOS test-timeout in
+// TestDNSPermanent_updateUpstream: WGIface.Close() used to hold w.mu
+// while waiting for the wireguard-go device goroutines to finish, and
+// one of those goroutines (the DNS filter hook) calls back into
+// WGIface.GetDevice() which needs the same mutex. The fix is to drop
+// the lock before tun.Close() returns control.
+func TestWGIface_CloseReleasesMutexBeforeTunClose(t *testing.T) {
+	tun := &fakeTunDevice{
+		closeStarted: make(chan struct{}),
+		unblockClose: make(chan struct{}),
+	}
+	w := &WGIface{
+		tun:            tun,
+		wgProxyFactory: fakeProxyFactory{},
+	}
+
+	closeDone := make(chan error, 1)
+	go func() {
+		closeDone <- w.Close()
+	}()
+
+	select {
+	case <-tun.closeStarted:
+	case <-time.After(2 * time.Second):
+		close(tun.unblockClose)
+		t.Fatal("tun.Close() was never invoked")
+	}
+
+	// Simulate the WireGuard read goroutine calling back into WGIface
+	// via the packet filter's DNS hook. If Close() still held w.mu
+	// during tun.Close(), this would block until the test timeout.
+	getDeviceDone := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
+		_ = w.GetDevice()
+		close(getDeviceDone)
+	}()
+
+	select {
+	case <-getDeviceDone:
+	case <-time.After(2 * time.Second):
+		close(tun.unblockClose)
+		wg.Wait()
+		t.Fatal("GetDevice() deadlocked while WGIface.Close was closing the tun")
+	}
+
+	close(tun.unblockClose)
+	select {
+	case <-closeDone:
+	case <-time.After(2 * time.Second):
+		t.Fatal("WGIface.Close() never returned after the tun was unblocked")
+	}
+}

client/iface/udpmux/universal.go

@@ -171,7 +171,7 @@ func (u *UDPConn) performFilterCheck(addr net.Addr) error {
 	}
 
 	if u.address.Network.Contains(a) {
-		log.Warnf("Address %s is part of the NetBird network %s, refusing to write", addr, u.address)
+		log.Warnf("address %s is part of the NetBird network %s, refusing to write", addr, u.address)
 		return fmt.Errorf("address %s is part of the NetBird network %s, refusing to write", addr, u.address)
 	}
 
@@ -181,7 +181,7 @@ func (u *UDPConn) performFilterCheck(addr net.Addr) error {
 		u.addrCache.Store(addr.String(), isRouted)
 		if isRouted {
 			// Extra log, as the error only shows up with ICE logging enabled
-			log.Infof("Address %s is part of routed network %s, refusing to write", addr, prefix)
+			log.Infof("address %s is part of routed network %s, refusing to write", addr, prefix)
 			return fmt.Errorf("address %s is part of routed network %s, refusing to write", addr, prefix)
 		}
 	}

client/internal/engine.go

@@ -26,6 +26,7 @@ import (
 
 	nberrors "github.com/netbirdio/netbird/client/errors"
 	"github.com/netbirdio/netbird/client/firewall"
+	"github.com/netbirdio/netbird/client/firewall/firewalld"
 	firewallManager "github.com/netbirdio/netbird/client/firewall/manager"
 	"github.com/netbirdio/netbird/client/iface"
 	"github.com/netbirdio/netbird/client/iface/device"
@@ -603,6 +604,8 @@ func (e *Engine) createFirewall() error {
 		return nil
 	}
 
+	firewalld.SetParentContext(e.ctx)
+
 	var err error
 	e.firewall, err = firewall.NewFirewall(e.wgInterface, e.stateManager, e.flowManager.GetLogger(), e.config.DisableServerRoutes, e.config.MTU)
 	if err != nil {

client/internal/routemanager/systemops/systemops_bsd_other.go

@@ -0,0 +1,10 @@
+//go:build (dragonfly || freebsd || netbsd || openbsd) && !darwin
+
+package systemops
+
+// Non-darwin BSDs don't support the IP_BOUND_IF + scoped default model. They
+// always fall through to the ref-counter exclusion-route path; these stubs
+// exist only so systemops_unix.go compiles.
+func (r *SysOps) setupAdvancedRouting() error   { return nil }
+func (r *SysOps) cleanupAdvancedRouting() error { return nil }
+func (r *SysOps) flushPlatformExtras() error    { return nil }

client/internal/routemanager/systemops/systemops_darwin.go

@@ -0,0 +1,241 @@
+//go:build darwin && !ios
+
+package systemops
+
+import (
+	"errors"
+	"fmt"
+	"net/netip"
+	"os"
+	"time"
+
+	"github.com/hashicorp/go-multierror"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/net/route"
+	"golang.org/x/sys/unix"
+
+	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/client/internal/routemanager/vars"
+	nbnet "github.com/netbirdio/netbird/client/net"
+)
+
+// scopedRouteBudget bounds retries for the scoped default route. Installing or
+// deleting it matters enough that we're willing to spend longer waiting for the
+// kernel reply than for per-prefix exclusion routes.
+const scopedRouteBudget = 5 * time.Second
+
+// setupAdvancedRouting installs an RTF_IFSCOPE default route per address family
+// pinned to the current physical egress, so IP_BOUND_IF scoped lookups can
+// resolve gateway'd destinations while the VPN's split default owns the
+// unscoped table.
+//
+// Timing note: this runs during routeManager.Init, which happens before the
+// VPN interface is created and before any peer routes propagate. The initial
+// mgmt / signal / relay TCP dials always fire before this runs, so those
+// sockets miss the IP_BOUND_IF binding and rely on the kernel's normal route
+// lookup, which at that point correctly picks the physical default. Those
+// already-established TCP flows keep their originally-selected interface for
+// their lifetime on Darwin because the kernel caches the egress route
+// per-socket at connect time; adding the VPN's 0/1 + 128/1 split default
+// afterwards does not migrate them since the original en0 default stays in
+// the table. Any subsequent reconnect via nbnet.NewDialer picks up the
+// populated bound-iface cache and gets IP_BOUND_IF set cleanly.
+func (r *SysOps) setupAdvancedRouting() error {
+	// Drop any previously-cached egress interface before reinstalling. On a
+	// refresh, a family that no longer resolves would otherwise keep the stale
+	// binding, causing new sockets to scope to an interface without a matching
+	// scoped default.
+	nbnet.ClearBoundInterfaces()
+
+	if err := r.flushScopedDefaults(); err != nil {
+		log.Warnf("flush residual scoped defaults: %v", err)
+	}
+
+	var merr *multierror.Error
+	installed := 0
+
+	for _, unspec := range []netip.Addr{netip.IPv4Unspecified(), netip.IPv6Unspecified()} {
+		ok, err := r.installScopedDefaultFor(unspec)
+		if err != nil {
+			merr = multierror.Append(merr, err)
+			continue
+		}
+		if ok {
+			installed++
+		}
+	}
+
+	if installed == 0 && merr != nil {
+		return nberrors.FormatErrorOrNil(merr)
+	}
+	if merr != nil {
+		log.Warnf("advanced routing setup partially succeeded: %v", nberrors.FormatErrorOrNil(merr))
+	}
+	return nil
+}
+
+// installScopedDefaultFor resolves the physical default nexthop for the given
+// address family, installs a scoped default via it, and caches the iface for
+// subsequent IP_BOUND_IF / IPV6_BOUND_IF socket binds.
+func (r *SysOps) installScopedDefaultFor(unspec netip.Addr) (bool, error) {
+	nexthop, err := GetNextHop(unspec)
+	if err != nil {
+		if errors.Is(err, vars.ErrRouteNotFound) {
+			return false, nil
+		}
+		return false, fmt.Errorf("get default nexthop for %s: %w", unspec, err)
+	}
+	if nexthop.Intf == nil {
+		return false, fmt.Errorf("unusable default nexthop for %s (no interface)", unspec)
+	}
+
+	if err := r.addScopedDefault(unspec, nexthop); err != nil {
+		return false, fmt.Errorf("add scoped default on %s: %w", nexthop.Intf.Name, err)
+	}
+
+	af := unix.AF_INET
+	if unspec.Is6() {
+		af = unix.AF_INET6
+	}
+	nbnet.SetBoundInterface(af, nexthop.Intf)
+	via := "point-to-point"
+	if nexthop.IP.IsValid() {
+		via = nexthop.IP.String()
+	}
+	log.Infof("installed scoped default route via %s on %s for %s", via, nexthop.Intf.Name, afOf(unspec))
+	return true, nil
+}
+
+func (r *SysOps) cleanupAdvancedRouting() error {
+	nbnet.ClearBoundInterfaces()
+	return r.flushScopedDefaults()
+}
+
+// flushPlatformExtras runs darwin-specific residual cleanup hooked into the
+// generic FlushMarkedRoutes path, so a crashed daemon's scoped defaults get
+// removed on the next boot regardless of whether a profile is brought up.
+func (r *SysOps) flushPlatformExtras() error {
+	return r.flushScopedDefaults()
+}
+
+// flushScopedDefaults removes any scoped default routes tagged with routeProtoFlag.
+// Safe to call at startup to clear residual entries from a prior session.
+func (r *SysOps) flushScopedDefaults() error {
+	rib, err := retryFetchRIB()
+	if err != nil {
+		return fmt.Errorf("fetch routing table: %w", err)
+	}
+
+	msgs, err := route.ParseRIB(route.RIBTypeRoute, rib)
+	if err != nil {
+		return fmt.Errorf("parse routing table: %w", err)
+	}
+
+	var merr *multierror.Error
+	removed := 0
+
+	for _, msg := range msgs {
+		rtMsg, ok := msg.(*route.RouteMessage)
+		if !ok {
+			continue
+		}
+		if rtMsg.Flags&routeProtoFlag == 0 {
+			continue
+		}
+		if rtMsg.Flags&unix.RTF_IFSCOPE == 0 {
+			continue
+		}
+
+		info, err := MsgToRoute(rtMsg)
+		if err != nil {
+			log.Debugf("skip scoped flush: %v", err)
+			continue
+		}
+		if !info.Dst.IsValid() || info.Dst.Bits() != 0 {
+			continue
+		}
+
+		if err := r.deleteScopedRoute(rtMsg); err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("delete scoped default %s on index %d: %w",
+				info.Dst, rtMsg.Index, err))
+			continue
+		}
+		removed++
+		log.Debugf("flushed residual scoped default %s on index %d", info.Dst, rtMsg.Index)
+	}
+
+	if removed > 0 {
+		log.Infof("flushed %d residual scoped default route(s)", removed)
+	}
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+func (r *SysOps) addScopedDefault(unspec netip.Addr, nexthop Nexthop) error {
+	return r.scopedRouteSocket(unix.RTM_ADD, unspec, nexthop)
+}
+
+func (r *SysOps) deleteScopedRoute(rtMsg *route.RouteMessage) error {
+	// Preserve identifying flags from the stored route (including RTF_GATEWAY
+	// only if present); kernel-set bits like RTF_DONE don't belong on RTM_DELETE.
+	keep := unix.RTF_UP | unix.RTF_STATIC | unix.RTF_GATEWAY | unix.RTF_IFSCOPE | routeProtoFlag
+	del := &route.RouteMessage{
+		Type:    unix.RTM_DELETE,
+		Flags:   rtMsg.Flags & keep,
+		Version: unix.RTM_VERSION,
+		Seq:     r.getSeq(),
+		Index:   rtMsg.Index,
+		Addrs:   rtMsg.Addrs,
+	}
+	return r.writeRouteMessage(del, scopedRouteBudget)
+}
+
+func (r *SysOps) scopedRouteSocket(action int, unspec netip.Addr, nexthop Nexthop) error {
+	flags := unix.RTF_UP | unix.RTF_STATIC | unix.RTF_IFSCOPE | routeProtoFlag
+
+	msg := &route.RouteMessage{
+		Type:    action,
+		Flags:   flags,
+		Version: unix.RTM_VERSION,
+		ID:      uintptr(os.Getpid()),
+		Seq:     r.getSeq(),
+		Index:   nexthop.Intf.Index,
+	}
+
+	const numAddrs = unix.RTAX_NETMASK + 1
+	addrs := make([]route.Addr, numAddrs)
+
+	dst, err := addrToRouteAddr(unspec)
+	if err != nil {
+		return fmt.Errorf("build destination: %w", err)
+	}
+	mask, err := prefixToRouteNetmask(netip.PrefixFrom(unspec, 0))
+	if err != nil {
+		return fmt.Errorf("build netmask: %w", err)
+	}
+	addrs[unix.RTAX_DST] = dst
+	addrs[unix.RTAX_NETMASK] = mask
+
+	if nexthop.IP.IsValid() {
+		msg.Flags |= unix.RTF_GATEWAY
+		gw, err := addrToRouteAddr(nexthop.IP.Unmap())
+		if err != nil {
+			return fmt.Errorf("build gateway: %w", err)
+		}
+		addrs[unix.RTAX_GATEWAY] = gw
+	} else {
+		addrs[unix.RTAX_GATEWAY] = &route.LinkAddr{
+			Index: nexthop.Intf.Index,
+			Name:  nexthop.Intf.Name,
+		}
+	}
+	msg.Addrs = addrs
+
+	return r.writeRouteMessage(msg, scopedRouteBudget)
+}
+
+func afOf(a netip.Addr) string {
+	if a.Is4() {
+		return "IPv4"
+	}
+	return "IPv6"
+}

client/internal/routemanager/systemops/systemops_generic.go

@@ -21,6 +21,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/routemanager/util"
 	"github.com/netbirdio/netbird/client/internal/routemanager/vars"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
+	nbnet "github.com/netbirdio/netbird/client/net"
 	"github.com/netbirdio/netbird/client/net/hooks"
 )
 
@@ -31,8 +32,6 @@ var splitDefaultv4_2 = netip.PrefixFrom(netip.AddrFrom4([4]byte{128}), 1)
 var splitDefaultv6_1 = netip.PrefixFrom(netip.IPv6Unspecified(), 1)
 var splitDefaultv6_2 = netip.PrefixFrom(netip.AddrFrom16([16]byte{0x80}), 1)
 
-var ErrRoutingIsSeparate = errors.New("routing is separate")
-
 func (r *SysOps) setupRefCounter(initAddresses []net.IP, stateManager *statemanager.Manager) error {
 	stateManager.RegisterState(&ShutdownState{})
 
@@ -397,12 +396,16 @@ func ipToAddr(ip net.IP, intf *net.Interface) (netip.Addr, error) {
 }
 
 // IsAddrRouted checks if the candidate address would route to the vpn, in which case it returns true and the matched prefix.
+// When advanced routing is active the WG socket is bound to the physical interface (fwmark on linux,
+// IP_UNICAST_IF on windows, IP_BOUND_IF on darwin) and bypasses the main routing table, so the check is skipped.
 func IsAddrRouted(addr netip.Addr, vpnRoutes []netip.Prefix) (bool, netip.Prefix) {
-	localRoutes, err := hasSeparateRouting()
+	if nbnet.AdvancedRouting() {
+		return false, netip.Prefix{}
+	}
+
+	localRoutes, err := GetRoutesFromTable()
 	if err != nil {
-		if !errors.Is(err, ErrRoutingIsSeparate) {
-			log.Errorf("Failed to get routes: %v", err)
-		}
+		log.Errorf("Failed to get routes: %v", err)
 		return false, netip.Prefix{}
 	}
 

client/internal/routemanager/systemops/systemops_js.go

@@ -22,10 +22,6 @@ func GetRoutesFromTable() ([]netip.Prefix, error) {
 	return []netip.Prefix{}, nil
 }
 
-func hasSeparateRouting() ([]netip.Prefix, error) {
-	return []netip.Prefix{}, nil
-}
-
 // GetDetailedRoutesFromTable returns empty routes for WASM.
 func GetDetailedRoutesFromTable() ([]DetailedRoute, error) {
 	return []DetailedRoute{}, nil

client/internal/routemanager/systemops/systemops_linux.go

@@ -894,13 +894,6 @@ func getAddressFamily(prefix netip.Prefix) int {
 	return netlink.FAMILY_V6
 }
 
-func hasSeparateRouting() ([]netip.Prefix, error) {
-	if !nbnet.AdvancedRouting() {
-		return GetRoutesFromTable()
-	}
-	return nil, ErrRoutingIsSeparate
-}
-
 func isOpErr(err error) bool {
 	// EAFTNOSUPPORT when ipv6 is disabled via sysctl, EOPNOTSUPP when disabled in boot options or otherwise not supported
 	if errors.Is(err, syscall.EAFNOSUPPORT) || errors.Is(err, syscall.EOPNOTSUPP) {

client/internal/routemanager/systemops/systemops_nonlinux.go

@@ -48,10 +48,6 @@ func EnableIPForwarding() error {
 	return nil
 }
 
-func hasSeparateRouting() ([]netip.Prefix, error) {
-	return GetRoutesFromTable()
-}
-
 // GetIPRules returns IP rules for debugging (not supported on non-Linux platforms)
 func GetIPRules() ([]IPRule, error) {
 	log.Infof("IP rules collection is not supported on %s", runtime.GOOS)

client/internal/routemanager/systemops/systemops_unix.go

@@ -25,6 +25,9 @@ import (
 
 const (
 	envRouteProtoFlag = "NB_ROUTE_PROTO_FLAG"
+
+	// routeBudget bounds retries for per-prefix exclusion route programming.
+	routeBudget = 1 * time.Second
 )
 
 var routeProtoFlag int
@@ -41,26 +44,42 @@ func init() {
 }
 
 func (r *SysOps) SetupRouting(initAddresses []net.IP, stateManager *statemanager.Manager, advancedRouting bool) error {
+	if advancedRouting {
+		return r.setupAdvancedRouting()
+	}
+
+	log.Infof("Using legacy routing setup with ref counters")
 	return r.setupRefCounter(initAddresses, stateManager)
 }
 
 func (r *SysOps) CleanupRouting(stateManager *statemanager.Manager, advancedRouting bool) error {
+	if advancedRouting {
+		return r.cleanupAdvancedRouting()
+	}
+
 	return r.cleanupRefCounter(stateManager)
 }
 
 // FlushMarkedRoutes removes single IP exclusion routes marked with the configured RTF_PROTO flag.
+// On darwin it also flushes residual RTF_IFSCOPE scoped default routes so a
+// crashed prior session can't leave crud in the table.
 func (r *SysOps) FlushMarkedRoutes() error {
+	var merr *multierror.Error
+
+	if err := r.flushPlatformExtras(); err != nil {
+		merr = multierror.Append(merr, fmt.Errorf("flush platform extras: %w", err))
+	}
+
 	rib, err := retryFetchRIB()
 	if err != nil {
-		return fmt.Errorf("fetch routing table: %w", err)
+		return nberrors.FormatErrorOrNil(multierror.Append(merr, fmt.Errorf("fetch routing table: %w", err)))
 	}
 
 	msgs, err := route.ParseRIB(route.RIBTypeRoute, rib)
 	if err != nil {
-		return fmt.Errorf("parse routing table: %w", err)
+		return nberrors.FormatErrorOrNil(multierror.Append(merr, fmt.Errorf("parse routing table: %w", err)))
 	}
 
-	var merr *multierror.Error
 	flushedCount := 0
 
 	for _, msg := range msgs {
@@ -117,12 +136,12 @@ func (r *SysOps) routeSocket(action int, prefix netip.Prefix, nexthop Nexthop) e
 		return fmt.Errorf("invalid prefix: %s", prefix)
 	}
 
-	expBackOff := backoff.NewExponentialBackOff()
-	expBackOff.InitialInterval = 50 * time.Millisecond
-	expBackOff.MaxInterval = 500 * time.Millisecond
-	expBackOff.MaxElapsedTime = 1 * time.Second
+	msg, err := r.buildRouteMessage(action, prefix, nexthop)
+	if err != nil {
+		return fmt.Errorf("build route message: %w", err)
+	}
 
-	if err := backoff.Retry(r.routeOp(action, prefix, nexthop), expBackOff); err != nil {
+	if err := r.writeRouteMessage(msg, routeBudget); err != nil {
 		a := "add"
 		if action == unix.RTM_DELETE {
 			a = "remove"
@@ -132,50 +151,91 @@ func (r *SysOps) routeSocket(action int, prefix netip.Prefix, nexthop Nexthop) e
 	return nil
 }
 
-func (r *SysOps) routeOp(action int, prefix netip.Prefix, nexthop Nexthop) func() error {
-	operation := func() error {
-		fd, err := unix.Socket(syscall.AF_ROUTE, syscall.SOCK_RAW, syscall.AF_UNSPEC)
-		if err != nil {
-			return fmt.Errorf("open routing socket: %w", err)
+// writeRouteMessage sends a route message over AF_ROUTE and waits for the
+// kernel's matching reply, retrying transient failures until budget elapses.
+// Callers do not need to manage sockets or seq numbers themselves.
+func (r *SysOps) writeRouteMessage(msg *route.RouteMessage, budget time.Duration) error {
+	expBackOff := backoff.NewExponentialBackOff()
+	expBackOff.InitialInterval = 50 * time.Millisecond
+	expBackOff.MaxInterval = 500 * time.Millisecond
+	expBackOff.MaxElapsedTime = budget
+
+	return backoff.Retry(func() error { return routeMessageRoundtrip(msg) }, expBackOff)
+}
+
+func routeMessageRoundtrip(msg *route.RouteMessage) error {
+	fd, err := unix.Socket(syscall.AF_ROUTE, syscall.SOCK_RAW, syscall.AF_UNSPEC)
+	if err != nil {
+		return fmt.Errorf("open routing socket: %w", err)
+	}
+	defer func() {
+		if err := unix.Close(fd); err != nil && !errors.Is(err, unix.EBADF) {
+			log.Warnf("close routing socket: %v", err)
 		}
-		defer func() {
-			if err := unix.Close(fd); err != nil && !errors.Is(err, unix.EBADF) {
-				log.Warnf("failed to close routing socket: %v", err)
+	}()
+
+	tv := unix.Timeval{Sec: 1}
+	if err := unix.SetsockoptTimeval(fd, unix.SOL_SOCKET, unix.SO_RCVTIMEO, &tv); err != nil {
+		return backoff.Permanent(fmt.Errorf("set recv timeout: %w", err))
+	}
+
+	// AF_ROUTE is a broadcast channel: every route socket on the host sees
+	// every RTM_* event. With concurrent route programming the default
+	// per-socket queue overflows and our own reply gets dropped.
+	if err := unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_RCVBUF, 1<<20); err != nil {
+		log.Debugf("set SO_RCVBUF on route socket: %v", err)
+	}
+
+	bytes, err := msg.Marshal()
+	if err != nil {
+		return backoff.Permanent(fmt.Errorf("marshal: %w", err))
+	}
+
+	if _, err = unix.Write(fd, bytes); err != nil {
+		if errors.Is(err, unix.ENOBUFS) || errors.Is(err, unix.EAGAIN) {
+			return fmt.Errorf("write: %w", err)
+		}
+		return backoff.Permanent(fmt.Errorf("write: %w", err))
+	}
+	return readRouteResponse(fd, msg.Type, msg.Seq)
+}
+
+// readRouteResponse reads from the AF_ROUTE socket until it sees a reply
+// matching our write (same type, seq, and pid). AF_ROUTE SOCK_RAW is a
+// broadcast channel: interface up/down, third-party route changes and neighbor
+// discovery events can all land between our write and read, so we must filter.
+func readRouteResponse(fd, wantType, wantSeq int) error {
+	pid := int32(os.Getpid())
+	resp := make([]byte, 2048)
+	deadline := time.Now().Add(time.Second)
+	for {
+		if time.Now().After(deadline) {
+			// Transient: under concurrent pressure the kernel can drop our reply
+			// from the socket buffer. Let backoff.Retry re-send with a fresh seq.
+			return fmt.Errorf("read: timeout waiting for route reply type=%d seq=%d", wantType, wantSeq)
+		}
+		n, err := unix.Read(fd, resp)
+		if err != nil {
+			if errors.Is(err, unix.EAGAIN) || errors.Is(err, unix.EWOULDBLOCK) {
+				// SO_RCVTIMEO fired while waiting; loop to re-check the absolute deadline.
+				continue
 			}
-		}()
-
-		msg, err := r.buildRouteMessage(action, prefix, nexthop)
-		if err != nil {
-			return backoff.Permanent(fmt.Errorf("build route message: %w", err))
+			return backoff.Permanent(fmt.Errorf("read: %w", err))
 		}
-
-		msgBytes, err := msg.Marshal()
-		if err != nil {
-			return backoff.Permanent(fmt.Errorf("marshal route message: %w", err))
+		if n < int(unsafe.Sizeof(unix.RtMsghdr{})) {
+			continue
 		}
-
-		if _, err = unix.Write(fd, msgBytes); err != nil {
-			if errors.Is(err, unix.ENOBUFS) || errors.Is(err, unix.EAGAIN) {
-				return fmt.Errorf("write: %w", err)
-			}
-			return backoff.Permanent(fmt.Errorf("write: %w", err))
+		hdr := (*unix.RtMsghdr)(unsafe.Pointer(&resp[0]))
+		// Darwin reflects the sender's pid on replies; matching (Type, Seq, Pid)
+		// uniquely identifies our own reply among broadcast traffic.
+		if int(hdr.Type) != wantType || int(hdr.Seq) != wantSeq || hdr.Pid != pid {
+			continue
 		}
-
-		respBuf := make([]byte, 2048)
-		n, err := unix.Read(fd, respBuf)
-		if err != nil {
-			return backoff.Permanent(fmt.Errorf("read route response: %w", err))
+		if hdr.Errno != 0 {
+			return backoff.Permanent(fmt.Errorf("kernel: %w", syscall.Errno(hdr.Errno)))
 		}
-
-		if n > 0 {
-			if err := r.parseRouteResponse(respBuf[:n]); err != nil {
-				return backoff.Permanent(err)
-			}
-		}
-
 		return nil
 	}
-	return operation
 }
 
 func (r *SysOps) buildRouteMessage(action int, prefix netip.Prefix, nexthop Nexthop) (msg *route.RouteMessage, err error) {
@@ -183,6 +243,7 @@ func (r *SysOps) buildRouteMessage(action int, prefix netip.Prefix, nexthop Next
 		Type:    action,
 		Flags:   unix.RTF_UP | routeProtoFlag,
 		Version: unix.RTM_VERSION,
+		ID:      uintptr(os.Getpid()),
 		Seq:     r.getSeq(),
 	}
 
@@ -221,19 +282,6 @@ func (r *SysOps) buildRouteMessage(action int, prefix netip.Prefix, nexthop Next
 	return msg, nil
 }
 
-func (r *SysOps) parseRouteResponse(buf []byte) error {
-	if len(buf) < int(unsafe.Sizeof(unix.RtMsghdr{})) {
-		return nil
-	}
-
-	rtMsg := (*unix.RtMsghdr)(unsafe.Pointer(&buf[0]))
-	if rtMsg.Errno != 0 {
-		return fmt.Errorf("parse: %d", rtMsg.Errno)
-	}
-
-	return nil
-}
-
 // addrToRouteAddr converts a netip.Addr to the appropriate route.Addr (*route.Inet4Addr or *route.Inet6Addr).
 func addrToRouteAddr(addr netip.Addr) (route.Addr, error) {
 	if addr.Is4() {

client/net/dialer_init_darwin.go

@@ -0,0 +1,5 @@
+package net
+
+func (d *Dialer) init() {
+	d.Dialer.Control = applyBoundIfToSocket
+}

client/net/dialer_init_generic.go

@@ -1,4 +1,4 @@
-//go:build !linux && !windows
+//go:build !linux && !windows && !darwin
 
 package net
 

client/net/env_android.go

@@ -1,24 +0,0 @@
-//go:build android
-
-package net
-
-// Init initializes the network environment for Android
-func Init() {
-	// No initialization needed on Android
-}
-
-// AdvancedRouting reports whether routing loops can be avoided without using exclusion routes.
-// Always returns true on Android since we cannot handle routes dynamically.
-func AdvancedRouting() bool {
-	return true
-}
-
-// SetVPNInterfaceName is a no-op on Android
-func SetVPNInterfaceName(name string) {
-	// No-op on Android - not needed for Android VPN service
-}
-
-// GetVPNInterfaceName returns empty string on Android
-func GetVPNInterfaceName() string {
-	return ""
-}

client/net/env_bound_iface.go

@@ -1,4 +1,4 @@
-//go:build windows
+//go:build (darwin && !ios) || windows
 
 package net
 
@@ -24,17 +24,22 @@ func Init() {
 }
 
 func checkAdvancedRoutingSupport() bool {
-	var err error
-	var legacyRouting bool
+	legacyRouting := false
 	if val := os.Getenv(envUseLegacyRouting); val != "" {
-		legacyRouting, err = strconv.ParseBool(val)
+		parsed, err := strconv.ParseBool(val)
 		if err != nil {
-			log.Warnf("failed to parse %s: %v", envUseLegacyRouting, err)
+			log.Warnf("ignoring unparsable %s=%q: %v", envUseLegacyRouting, val, err)
+		} else {
+			legacyRouting = parsed
 		}
 	}
 
-	if legacyRouting || netstack.IsEnabled() {
-		log.Info("advanced routing has been requested to be disabled")
+	if legacyRouting {
+		log.Infof("advanced routing disabled: legacy routing requested via %s", envUseLegacyRouting)
+		return false
+	}
+	if netstack.IsEnabled() {
+		log.Info("advanced routing disabled: netstack mode is enabled")
 		return false
 	}
 

client/net/env_generic.go

@@ -1,4 +1,4 @@
-//go:build !linux && !windows && !android
+//go:build !linux && !windows && !darwin
 
 package net
 

client/net/env_mobile.go

@@ -0,0 +1,25 @@
+//go:build ios || android
+
+package net
+
+// Init initializes the network environment for mobile platforms.
+func Init() {
+	// no-op on mobile: routing scope is owned by the VPN extension.
+}
+
+// AdvancedRouting reports whether routing loops can be avoided without using exclusion routes.
+// Always returns true on mobile since routes cannot be handled dynamically and the VPN extension
+// owns the routing scope.
+func AdvancedRouting() bool {
+	return true
+}
+
+// SetVPNInterfaceName is a no-op on mobile.
+func SetVPNInterfaceName(string) {
+	// no-op on mobile: the VPN extension manages the interface.
+}
+
+// GetVPNInterfaceName returns an empty string on mobile.
+func GetVPNInterfaceName() string {
+	return ""
+}

client/net/listener_init_darwin.go

@@ -0,0 +1,5 @@
+package net
+
+func (l *ListenerConfig) init() {
+	l.ListenConfig.Control = applyBoundIfToSocket
+}

client/net/listener_init_generic.go

@@ -1,4 +1,4 @@
-//go:build !linux && !windows
+//go:build !linux && !windows && !darwin
 
 package net
 

client/net/net_darwin.go

@@ -0,0 +1,160 @@
+package net
+
+import (
+	"fmt"
+	"net"
+	"net/netip"
+	"strconv"
+	"strings"
+	"sync"
+	"syscall"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/unix"
+)
+
+// On darwin IPV6_BOUND_IF also scopes v4-mapped egress from dual-stack
+// (IPV6_V6ONLY=0) AF_INET6 sockets, so a single setsockopt on "udp6"/"tcp6"
+// covers both families. Setting IP_BOUND_IF on an AF_INET6 socket returns
+// EINVAL regardless of V6ONLY because the IPPROTO_IP ctloutput path is
+// dispatched by socket domain (AF_INET only) not by inp_vflag.
+
+// boundIface holds the physical interface chosen at routing setup time. Sockets
+// created via nbnet.NewDialer / nbnet.NewListener bind to it via IP_BOUND_IF
+// (IPv4) or IPV6_BOUND_IF (IPv6 / dual-stack) so their scoped route lookup
+// hits the RTF_IFSCOPE default installed by the routemanager, rather than
+// following the VPN's split default.
+var (
+	boundIfaceMu sync.RWMutex
+	boundIface4  *net.Interface
+	boundIface6  *net.Interface
+)
+
+// SetBoundInterface records the egress interface for an address family. Called
+// by the routemanager after a scoped default route has been installed.
+// af must be unix.AF_INET or unix.AF_INET6; other values are ignored.
+// nil iface is rejected — use ClearBoundInterfaces to clear all slots.
+func SetBoundInterface(af int, iface *net.Interface) {
+	if iface == nil {
+		log.Warnf("SetBoundInterface: nil iface for AF %d, ignored", af)
+		return
+	}
+	boundIfaceMu.Lock()
+	defer boundIfaceMu.Unlock()
+	switch af {
+	case unix.AF_INET:
+		boundIface4 = iface
+	case unix.AF_INET6:
+		boundIface6 = iface
+	default:
+		log.Warnf("SetBoundInterface: unsupported address family %d", af)
+	}
+}
+
+// ClearBoundInterfaces resets the cached egress interfaces. Called by the
+// routemanager during cleanup.
+func ClearBoundInterfaces() {
+	boundIfaceMu.Lock()
+	defer boundIfaceMu.Unlock()
+	boundIface4 = nil
+	boundIface6 = nil
+}
+
+// boundInterfaceFor returns the cached egress interface for a socket's address
+// family, falling back to the other family if the preferred slot is empty.
+// The kernel stores both IP_BOUND_IF and IPV6_BOUND_IF in inp_boundifp, so
+// either setsockopt scopes the socket; preferring same-family still matters
+// when v4 and v6 defaults egress different NICs.
+func boundInterfaceFor(network, address string) *net.Interface {
+	if iface := zoneInterface(address); iface != nil {
+		return iface
+	}
+
+	boundIfaceMu.RLock()
+	defer boundIfaceMu.RUnlock()
+
+	primary, secondary := boundIface4, boundIface6
+	if isV6Network(network) {
+		primary, secondary = boundIface6, boundIface4
+	}
+	if primary != nil {
+		return primary
+	}
+	return secondary
+}
+
+func isV6Network(network string) bool {
+	return strings.HasSuffix(network, "6")
+}
+
+// zoneInterface extracts an explicit interface from an IPv6 link-local zone (e.g. fe80::1%en0).
+func zoneInterface(address string) *net.Interface {
+	if address == "" {
+		return nil
+	}
+	addr, err := netip.ParseAddrPort(address)
+	if err != nil {
+		a, err := netip.ParseAddr(address)
+		if err != nil {
+			return nil
+		}
+		addr = netip.AddrPortFrom(a, 0)
+	}
+	zone := addr.Addr().Zone()
+	if zone == "" {
+		return nil
+	}
+	if iface, err := net.InterfaceByName(zone); err == nil {
+		return iface
+	}
+	if idx, err := strconv.Atoi(zone); err == nil {
+		if iface, err := net.InterfaceByIndex(idx); err == nil {
+			return iface
+		}
+	}
+	return nil
+}
+
+func setIPv4BoundIf(fd uintptr, iface *net.Interface) error {
+	if err := unix.SetsockoptInt(int(fd), unix.IPPROTO_IP, unix.IP_BOUND_IF, iface.Index); err != nil {
+		return fmt.Errorf("set IP_BOUND_IF: %w (interface: %s, index: %d)", err, iface.Name, iface.Index)
+	}
+	return nil
+}
+
+func setIPv6BoundIf(fd uintptr, iface *net.Interface) error {
+	if err := unix.SetsockoptInt(int(fd), unix.IPPROTO_IPV6, unix.IPV6_BOUND_IF, iface.Index); err != nil {
+		return fmt.Errorf("set IPV6_BOUND_IF: %w (interface: %s, index: %d)", err, iface.Name, iface.Index)
+	}
+	return nil
+}
+
+// applyBoundIfToSocket binds the socket to the cached physical egress interface
+// so scoped route lookup avoids the VPN utun and egresses the underlay directly.
+func applyBoundIfToSocket(network, address string, c syscall.RawConn) error {
+	if !AdvancedRouting() {
+		return nil
+	}
+
+	iface := boundInterfaceFor(network, address)
+	if iface == nil {
+		log.Debugf("no bound iface cached for %s to %s, skipping BOUND_IF", network, address)
+		return nil
+	}
+
+	isV6 := isV6Network(network)
+	var controlErr error
+	if err := c.Control(func(fd uintptr) {
+		if isV6 {
+			controlErr = setIPv6BoundIf(fd, iface)
+		} else {
+			controlErr = setIPv4BoundIf(fd, iface)
+		}
+		if controlErr == nil {
+			log.Debugf("set BOUND_IF=%d on %s for %s to %s", iface.Index, iface.Name, network, address)
+		}
+	}); err != nil {
+		return fmt.Errorf("control: %w", err)
+	}
+	return controlErr
+}

client/server/state.go

@@ -12,7 +12,6 @@ import (
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/routemanager/systemops"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
-	nbnet "github.com/netbirdio/netbird/client/net"
 	"github.com/netbirdio/netbird/client/proto"
 )
 
@@ -138,10 +137,8 @@ func restoreResidualState(ctx context.Context, statePath string) error {
 	}
 
 	// clean up any remaining routes independently of the state file
-	if !nbnet.AdvancedRouting() {
-		if err := systemops.New(nil, nil).FlushMarkedRoutes(); err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("flush marked routes: %w", err))
-		}
+	if err := systemops.New(nil, nil).FlushMarkedRoutes(); err != nil {
+		merr = multierror.Append(merr, fmt.Errorf("flush marked routes: %w", err))
 	}
 
 	return nberrors.FormatErrorOrNil(merr)

client/ssh/config/manager.go

@@ -187,24 +187,23 @@ func (m *Manager) buildPeerConfig(allHostPatterns []string) (string, error) {
 		return "", fmt.Errorf("get NetBird executable path: %w", err)
 	}
 
-	hostLine := strings.Join(deduplicatedPatterns, " ")
-	config := fmt.Sprintf("Host %s\n", hostLine)
-	config += fmt.Sprintf("    Match exec \"%s ssh detect %%h %%p\"\n", execPath)
-	config += "        PreferredAuthentications password,publickey,keyboard-interactive\n"
-	config += "        PasswordAuthentication yes\n"
-	config += "        PubkeyAuthentication yes\n"
-	config += "        BatchMode no\n"
-	config += fmt.Sprintf("        ProxyCommand %s ssh proxy %%h %%p\n", execPath)
-	config += "        StrictHostKeyChecking no\n"
+	hostList := strings.Join(deduplicatedPatterns, ",")
+	config := fmt.Sprintf("Match host \"%s\" exec \"%s ssh detect %%h %%p\"\n", hostList, execPath)
+	config += "    PreferredAuthentications password,publickey,keyboard-interactive\n"
+	config += "    PasswordAuthentication yes\n"
+	config += "    PubkeyAuthentication yes\n"
+	config += "    BatchMode no\n"
+	config += fmt.Sprintf("    ProxyCommand %s ssh proxy %%h %%p\n", execPath)
+	config += "    StrictHostKeyChecking no\n"
 
 	if runtime.GOOS == "windows" {
-		config += "        UserKnownHostsFile NUL\n"
+		config += "    UserKnownHostsFile NUL\n"
 	} else {
-		config += "        UserKnownHostsFile /dev/null\n"
+		config += "    UserKnownHostsFile /dev/null\n"
 	}
 
-	config += "        CheckHostIP no\n"
-	config += "        LogLevel ERROR\n\n"
+	config += "    CheckHostIP no\n"
+	config += "    LogLevel ERROR\n\n"
 
 	return config, nil
 }

client/ssh/config/manager_test.go

@@ -116,6 +116,37 @@ func TestManager_PeerLimit(t *testing.T) {
 	assert.True(t, os.IsNotExist(err), "SSH config should not be created with too many peers")
 }
 
+func TestManager_MatchHostFormat(t *testing.T) {
+	tempDir, err := os.MkdirTemp("", "netbird-ssh-config-test")
+	require.NoError(t, err)
+	defer func() { assert.NoError(t, os.RemoveAll(tempDir)) }()
+
+	manager := &Manager{
+		sshConfigDir:  filepath.Join(tempDir, "ssh_config.d"),
+		sshConfigFile: "99-netbird.conf",
+	}
+
+	peers := []PeerSSHInfo{
+		{Hostname: "peer1", IP: "100.125.1.1", FQDN: "peer1.nb.internal"},
+		{Hostname: "peer2", IP: "100.125.1.2", FQDN: "peer2.nb.internal"},
+	}
+
+	err = manager.SetupSSHClientConfig(peers)
+	require.NoError(t, err)
+
+	configPath := filepath.Join(manager.sshConfigDir, manager.sshConfigFile)
+	content, err := os.ReadFile(configPath)
+	require.NoError(t, err)
+	configStr := string(content)
+
+	// Must use "Match host" with comma-separated patterns, not a bare "Host" directive.
+	// A bare "Host" followed by "Match exec" is incorrect per ssh_config(5): the Host block
+	// ends at the next Match keyword, making it a no-op and leaving the Match exec unscoped.
+	assert.NotContains(t, configStr, "\nHost ", "should not use bare Host directive")
+	assert.Contains(t, configStr, "Match host \"100.125.1.1,peer1.nb.internal,peer1,100.125.1.2,peer2.nb.internal,peer2\"",
+		"should use Match host with comma-separated patterns")
+}
+
 func TestManager_ForcedSSHConfig(t *testing.T) {
 	// Set force environment variable
 	t.Setenv(EnvForceSSHConfig, "true")