client/dns/mgmt: bypass overlay for control-plane FQDN resolution

When an exit-node peer's network-map installs a 0.0.0.0/0 default route
on the overlay interface before that peer's WireGuard key material is
active, any UDP socket dialing an off-link address is routed into wt0
and the kernel returns ENOKEY.

Two places needed fixing:

 1. The mgmt cache refresh path. It reactively refreshes the
    control-plane FQDNs advertised by the mgmt (api/signal/stun/turn/
    the Relay pool root) after the daemon has installed its own
    resolv.conf pointing at the overlay listener. Previously the
    refresh dial followed the chain's upstream handler, which followed
    the overlay default route and deadlocked on ENOKEY.

 2. Foreign relay FQDN resolution. When a remote peer is homed on a
    different relay instance than us, we need to resolve a streamline-*
    subdomain that is not in the cache. That lookup went through the
    same overlay-routed upstream and failed identically, deadlocking
    the exit-node test whenever the relay LB put the two peers on
    different instances.

Fix both by giving the mgmt cache a dedicated net.Resolver that dials
the original pre-NetBird system nameservers through nbnet.NewDialer.
The dialer marks the socket as control-plane (SO_MARK on Linux,
IP_BOUND_IF on darwin, IP_UNICAST_IF on Windows); the routemanager's
policy rules keep those sockets on the underlay regardless of the
overlay default.

Pool-root domains (the Relay entries in ServerDomains) now register
through a subdomain-matching wrapper so that instance subdomains like
streamline-de-fra1-0.relay.netbird.io also hit the mgmt cache handler.
On cache miss under a pool root, ServeDNS resolves the FQDN on demand
through the bypass resolver, caches the result, and returns it.

Pool-root membership is derived dynamically from mgmt-advertised
ServerDomains.Relay[] — no hardcoded domain lists, no protocol change.
No hardcoded fallback nameservers: if the host had no original system
resolver at all, the bypass resolver stays nil and the stale-while-
revalidate cache keeps serving. The general upstream forwarder and
the user DNS path are unchanged.

client/internal/dns/mgmt/bypass_resolver.go

@@ -0,0 +1,55 @@
+package mgmt
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"net/netip"
+
+	nbnet "github.com/netbirdio/netbird/client/net"
+)
+
+// NewBypassResolver builds a *net.Resolver that sends queries directly to
+// the supplied nameservers through a socket that bypasses the NetBird
+// overlay interface. This lets the mgmt cache refresh control-plane
+// FQDNs (api/signal/relay/stun/turn) even when an exit-node default
+// route is installed on the overlay before its peer is live.
+//
+// Returns nil if nameservers is empty. The caller must not pass
+// loopback/overlay IPs (e.g. 127.0.0.1, the overlay listener address);
+// those would defeat the purpose of bypassing.
+func NewBypassResolver(nameservers []netip.Addr) *net.Resolver {
+	if len(nameservers) == 0 {
+		return nil
+	}
+
+	servers := make([]string, 0, len(nameservers))
+	for _, ns := range nameservers {
+		if !ns.IsValid() || ns.IsLoopback() || ns.IsUnspecified() {
+			continue
+		}
+		servers = append(servers, netip.AddrPortFrom(ns, 53).String())
+	}
+	if len(servers) == 0 {
+		return nil
+	}
+
+	return &net.Resolver{
+		PreferGo: true,
+		Dial: func(ctx context.Context, network, _ string) (net.Conn, error) {
+			nbDialer := nbnet.NewDialer()
+			var lastErr error
+			for _, ns := range servers {
+				conn, err := nbDialer.DialContext(ctx, network, ns)
+				if err == nil {
+					return conn, nil
+				}
+				lastErr = err
+			}
+			if lastErr == nil {
+				return nil, fmt.Errorf("no bypass nameservers configured")
+			}
+			return nil, fmt.Errorf("dial bypass nameservers: %w", lastErr)
+		},
+	}
+}

client/internal/dns/mgmt/mgmt.go

@@ -71,6 +71,14 @@ type Resolver struct {
 	refreshing map[dns.Question]*atomic.Bool
 
 	cacheTTL time.Duration
+
+	// bypassResolver, when non-nil, is used by osLookup instead of
+	// net.DefaultResolver. It is constructed by the caller to dial the
+	// original (pre-NetBird) system nameservers through a socket that
+	// bypasses the overlay interface (control-plane fwmark / bound iface),
+	// so that when an exit-node default route is installed before a peer
+	// is handshaked the refresh does not fail with ENOKEY.
+	bypassResolver *net.Resolver
 }
 
 // NewResolver creates a new management domains cache resolver.
@@ -98,8 +106,28 @@ func (m *Resolver) SetChainResolver(chain ChainResolver, maxPriority int) {
 	m.mutex.Unlock()
 }
 
+// SetBypassResolver installs a resolver that osLookup uses instead of
+// net.DefaultResolver. It is intended to dial the original (pre-NetBird)
+// system nameservers through a socket that does not follow the overlay
+// default route, so that a refresh initiated while an exit node is active
+// but its WireGuard peer is not yet installed cannot deadlock on ENOKEY.
+// Passing nil restores use of net.DefaultResolver.
+func (m *Resolver) SetBypassResolver(r *net.Resolver) {
+	m.mutex.Lock()
+	m.bypassResolver = r
+	m.mutex.Unlock()
+}
+
 // ServeDNS serves cached A/AAAA records. Stale entries are returned
 // immediately and refreshed asynchronously (stale-while-revalidate).
+//
+// If the query name is not in the cache but falls under a pool-root
+// domain (a domain the mgmt advertised in ServerDomains.Relay, whose
+// instance subdomains like streamline-de-fra1-0.relay.netbird.io are
+// part of the relay pool), resolve it on demand through the bypass
+// resolver and cache the result. This is what lets the daemon reach
+// a foreign relay FQDN after an exit-node default route has been
+// installed on the overlay before its peer is live.
 func (m *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	if len(r.Question) == 0 {
 		m.continueToNext(w, r)
@@ -126,6 +154,10 @@ func (m *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	m.mutex.RUnlock()
 
 	if !found {
+		if m.isUnderPoolRoot(question.Name) {
+			m.resolveOnDemand(w, r, question)
+			return
+		}
 		m.continueToNext(w, r)
 		return
 	}
@@ -155,12 +187,87 @@ func (m *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	}
 }
 
-// MatchSubdomains returns false since this resolver only handles exact domain matches
-// for NetBird infrastructure domains (signal, relay, flow, etc.), not their subdomains.
+// MatchSubdomains returns false by default: the bare resolver is registered
+// against exact domains. Pool-root domains (currently Relay entries from
+// ServerDomains) are registered through a subdomain-matching wrapper at
+// the call site instead, so instance subdomains hit this handler and get
+// the on-demand resolve path in ServeDNS.
 func (m *Resolver) MatchSubdomains() bool {
 	return false
 }
 
+// isUnderPoolRoot reports whether fqdn is an instance subdomain under any
+// pool-root domain advertised by the mgmt (currently ServerDomains.Relay),
+// e.g. "streamline-de-fra1-0.relay.netbird.io." is under "relay.netbird.io".
+// The pool-root itself is not considered a subdomain (it matches the exact
+// cache entry populated by AddDomain instead).
+func (m *Resolver) isUnderPoolRoot(fqdn string) bool {
+	m.mutex.RLock()
+	defer m.mutex.RUnlock()
+	if m.serverDomains == nil {
+		return false
+	}
+	fqdn = strings.ToLower(strings.TrimSuffix(fqdn, "."))
+	for _, root := range m.serverDomains.Relay {
+		r := strings.ToLower(strings.TrimSuffix(root.PunycodeString(), "."))
+		if r == "" || fqdn == r {
+			continue
+		}
+		if strings.HasSuffix(fqdn, "."+r) {
+			return true
+		}
+	}
+	return false
+}
+
+// resolveOnDemand resolves an uncached pool-root subdomain (e.g. a relay
+// instance FQDN) through the bypass resolver path, caches the result, and
+// writes it back to w. Falls through to the next handler on error so the
+// normal chain can still attempt the resolve.
+func (m *Resolver) resolveOnDemand(w dns.ResponseWriter, r *dns.Msg, question dns.Question) {
+	d, err := domain.FromString(strings.TrimSuffix(question.Name, "."))
+	if err != nil {
+		log.Debugf("on-demand resolve: parse domain %q: %v", question.Name, err)
+		m.continueToNext(w, r)
+		return
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), dnsTimeout)
+	defer cancel()
+
+	records, err := m.lookupRecords(ctx, d, question)
+	if err != nil {
+		log.Debugf("on-demand resolve %s type=%s: %v",
+			d.SafeString(), dns.TypeToString[question.Qtype], err)
+		m.continueToNext(w, r)
+		return
+	}
+	if len(records) == 0 {
+		m.continueToNext(w, r)
+		return
+	}
+
+	now := time.Now()
+	m.mutex.Lock()
+	if _, exists := m.records[question]; !exists {
+		m.records[question] = &cachedRecord{records: records, cachedAt: now}
+	}
+	m.mutex.Unlock()
+
+	resp := &dns.Msg{}
+	resp.SetReply(r)
+	resp.Authoritative = false
+	resp.RecursionAvailable = true
+	resp.Answer = cloneRecordsWithTTL(records, uint32(m.cacheTTL.Seconds()))
+
+	log.Debugf("on-demand resolved %d records for domain=%s", len(resp.Answer), question.Name)
+
+	if err := w.WriteMsg(resp); err != nil {
+		log.Errorf("failed to write on-demand response: %v", err)
+	}
+}
+
+
 // continueToNext signals the handler chain to continue to the next handler.
 func (m *Resolver) continueToNext(w dns.ResponseWriter, r *dns.Msg) {
 	resp := &dns.Msg{}
@@ -315,14 +422,29 @@ func (m *Resolver) markRefreshFailed(question dns.Question, expected *cachedReco
 	return c.consecFailures
 }
 
-// lookupBoth resolves A and AAAA via chain or OS. Per-family errors let
-// callers tell records, NODATA (nil err, no records), and failure apart.
+// lookupBoth resolves A and AAAA via bypass resolver, chain, or OS.
+// Per-family errors let callers tell records, NODATA (nil err, no records),
+// and failure apart.
+//
+// Preference order:
+//  1. bypassResolver (direct, overlay-bypassing dial to original system
+//     nameservers; immune to the exit-node ENOKEY race).
+//  2. chain (handler chain; used when NetBird is the system resolver and
+//     no bypass resolver is installed).
+//  3. net.DefaultResolver via osLookup (legacy fallback).
 func (m *Resolver) lookupBoth(ctx context.Context, d domain.Domain, dnsName string) (aRecords, aaaaRecords []dns.RR, errA, errAAAA error) {
 	m.mutex.RLock()
 	chain := m.chain
 	maxPriority := m.chainMaxPriority
+	bypass := m.bypassResolver
 	m.mutex.RUnlock()
 
+	if bypass != nil {
+		aRecords, errA = m.osLookup(ctx, d, dnsName, dns.TypeA)
+		aaaaRecords, errAAAA = m.osLookup(ctx, d, dnsName, dns.TypeAAAA)
+		return
+	}
+
 	if chain != nil && chain.HasRootHandlerAtOrBelow(maxPriority) {
 		aRecords, errA = m.lookupViaChain(ctx, chain, maxPriority, dnsName, dns.TypeA)
 		aaaaRecords, errAAAA = m.lookupViaChain(ctx, chain, maxPriority, dnsName, dns.TypeAAAA)
@@ -337,15 +459,22 @@ func (m *Resolver) lookupBoth(ctx context.Context, d domain.Domain, dnsName stri
 	return
 }
 
-// lookupRecords resolves a single record type via chain or OS. The OS branch
-// arms the loop detector for the duration of its call so that ServeDNS can
-// spot the OS resolver routing the recursive query back to us.
+// lookupRecords resolves a single record type. See lookupBoth for the
+// preference order. The OS branch arms the loop detector for the duration
+// of its call so that ServeDNS can spot the OS resolver routing the
+// recursive query back to us; the bypass branch skips the loop detector
+// because its dial does not enter the system resolver.
 func (m *Resolver) lookupRecords(ctx context.Context, d domain.Domain, q dns.Question) ([]dns.RR, error) {
 	m.mutex.RLock()
 	chain := m.chain
 	maxPriority := m.chainMaxPriority
+	bypass := m.bypassResolver
 	m.mutex.RUnlock()
 
+	if bypass != nil {
+		return m.osLookup(ctx, d, q.Name, q.Qtype)
+	}
+
 	if chain != nil && chain.HasRootHandlerAtOrBelow(maxPriority) {
 		return m.lookupViaChain(ctx, chain, maxPriority, q.Name, q.Qtype)
 	}
@@ -394,9 +523,9 @@ func (m *Resolver) lookupViaChain(ctx context.Context, chain ChainResolver, maxP
 	return filtered, nil
 }
 
-// osLookup resolves a single family via net.DefaultResolver using resutil,
-// which disambiguates NODATA from NXDOMAIN and Unmaps v4-mapped-v6. NODATA
-// returns (nil, nil).
+// osLookup resolves a single family via the bypass resolver (if configured)
+// or net.DefaultResolver using resutil, which disambiguates NODATA from
+// NXDOMAIN and Unmaps v4-mapped-v6. NODATA returns (nil, nil).
 func (m *Resolver) osLookup(ctx context.Context, d domain.Domain, dnsName string, qtype uint16) ([]dns.RR, error) {
 	network := resutil.NetworkForQtype(qtype)
 	if network == "" {
@@ -406,7 +535,14 @@ func (m *Resolver) osLookup(ctx context.Context, d domain.Domain, dnsName string
 	log.Infof("looking up IP for mgmt domain=%s type=%s", d.SafeString(), dns.TypeToString[qtype])
 	defer log.Infof("done looking up IP for mgmt domain=%s type=%s", d.SafeString(), dns.TypeToString[qtype])
 
-	result := resutil.LookupIP(ctx, net.DefaultResolver, network, d.PunycodeString(), qtype)
+	m.mutex.RLock()
+	resolver := m.bypassResolver
+	m.mutex.RUnlock()
+	if resolver == nil {
+		resolver = net.DefaultResolver
+	}
+
+	result := resutil.LookupIP(ctx, resolver, network, d.PunycodeString(), qtype)
 	if result.Rcode == dns.RcodeSuccess {
 		return resutil.IPsToRRs(dnsName, result.IPs, uint32(m.cacheTTL.Seconds())), nil
 	}
@@ -467,6 +603,24 @@ func (m *Resolver) RemoveDomain(d domain.Domain) error {
 	return nil
 }
 
+// GetPoolRootDomains returns the set of domains that should be registered
+// with subdomain matching (currently the Relay entries from ServerDomains).
+// Instance subdomains under these roots are resolved on demand in ServeDNS.
+func (m *Resolver) GetPoolRootDomains() domain.List {
+	m.mutex.RLock()
+	defer m.mutex.RUnlock()
+	if m.serverDomains == nil {
+		return nil
+	}
+	out := make(domain.List, 0, len(m.serverDomains.Relay))
+	for _, d := range m.serverDomains.Relay {
+		if d != "" {
+			out = append(out, d)
+		}
+	}
+	return out
+}
+
 // GetCachedDomains returns a list of all cached domains.
 func (m *Resolver) GetCachedDomains() domain.List {
 	m.mutex.RLock()