mapper: deliver peer removals as separate deltas

Updates #3492

(cherry picked from commit eeaac680be)

hscontrol/app.go

@@ -14,6 +14,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"runtime"
+	"slices"
 	"strings"
 	"sync"
 	"syscall"
@@ -176,10 +177,8 @@ func NewHeadscale(cfg *types.Config) (*Headscale, error) {
 			return
 		}
 
-		policyChanged, err := app.state.DeleteNode(node)
-		if !policyChanged.IsEmpty() {
-			app.Change(policyChanged)
-		}
+		changes, err := app.state.DeleteNode(node)
+		app.Change(changes...)
 
 		if err != nil {
 			log.Error().Err(err).EmbedObject(node).Msg("ephemeral node deletion failed")
@@ -1168,7 +1167,7 @@ func readOrCreatePrivateKey(path string) (*key.MachinePrivate, error) {
 // All change should be enqueued here and empty will be automatically
 // ignored.
 func (h *Headscale) Change(cs ...change.Change) {
-	h.mapBatcher.AddWork(cs...)
+	h.mapBatcher.AddWork(slices.Concat(cs, h.state.DrainSelfRefreshes())...)
 }
 
 // HTTPHandler returns an [http.Handler] for the [Headscale] control server.

hscontrol/auth.go

@@ -217,10 +217,8 @@ func (h *Headscale) handleLogout(
 				EmbedObject(node).
 				Msg("Deleting ephemeral node during logout")
 
-			c, err := h.state.DeleteNode(node)
-			if !c.IsEmpty() {
-				h.Change(c)
-			}
+			changes, err := h.state.DeleteNode(node)
+			h.Change(changes...)
 
 			if err != nil {
 				return nil, fmt.Errorf("deleting ephemeral node: %w", err)

hscontrol/capver/capver.go

@@ -15,14 +15,22 @@ import (
 // minVersionParts is the minimum number of version parts needed for major.minor.
 const minVersionParts = 2
 
+// fullNetmapRemovalsCapVer is the capability version of Tailscale main when
+// clients started reporting peers a full netmap drops as removed
+// (tailscale/tailscale#15660); v1.104 is the first release at or above it.
+const fullNetmapRemovalsCapVer tailcfg.CapabilityVersion = 148
+
 // CanOldCodeBeCleanedUp is called at server startup to panic when
 // [MinSupportedCapabilityVersion] has crossed a threshold at which a
 // backwards-compat emit path can be deleted. Each entry pairs a
 // [tailcfg.CapabilityVersion] threshold with the message identifying
-// the code to remove; today there are none.
+// the code to remove.
 //
 // All capability-version-gated cleanups should be registered here.
 func CanOldCodeBeCleanedUp() {
+	if MinSupportedCapabilityVersion >= fullNetmapRemovalsCapVer {
+		panic("the tailscale/tailscale#15660 compat can be cleaned up: grep for it (hscontrol/mapper/batcher.go and its tests)")
+	}
 }
 
 func tailscaleVersSorted() []string {

hscontrol/grpcv1.go

@@ -438,10 +438,8 @@ func (api headscaleV1APIServer) DeleteNode(
 		return nil, status.Errorf(codes.NotFound, "node not found")
 	}
 
-	nodeChange, err := api.h.state.DeleteNode(node)
-	if !nodeChange.IsEmpty() {
-		api.h.Change(nodeChange)
-	}
+	changes, err := api.h.state.DeleteNode(node)
+	api.h.Change(changes...)
 
 	if err != nil {
 		return nil, err

hscontrol/mapper/batcher.go

@@ -78,13 +78,18 @@ type nodeConnection interface {
 	updateSentPeers(resp *tailcfg.MapResponse)
 }
 
-// generateMapResponse generates a [tailcfg.MapResponse] for the given [types.NodeID] based on the provided [change.Change].
-func generateMapResponse(nc nodeConnection, mapper *mapper, r change.Change) (*tailcfg.MapResponse, error) {
+// generateMapResponse generates the [tailcfg.MapResponse] values to send, in
+// order, for the given [types.NodeID] based on the provided [change.Change].
+func generateMapResponse(
+	nc nodeConnection,
+	mapper *mapper,
+	r change.Change,
+) ([]*tailcfg.MapResponse, error) {
 	nodeID := nc.nodeID()
 	version := nc.version()
 
 	if r.IsEmpty() {
-		return nil, nil //nolint:nilnil // Empty response means nothing to send
+		return nil, nil
 	}
 
 	if nodeID == 0 {
@@ -97,7 +102,7 @@ func generateMapResponse(nc nodeConnection, mapper *mapper, r change.Change) (*t
 
 	// Handle self-only responses
 	if r.IsSelfOnly() && r.TargetNode != nodeID {
-		return nil, nil //nolint:nilnil // No response needed for other nodes when self-only
+		return nil, nil
 	}
 
 	// Check if this is a self-update (the changed node is the receiving node).
@@ -106,7 +111,8 @@ func generateMapResponse(nc nodeConnection, mapper *mapper, r change.Change) (*t
 	isSelfUpdate := r.OriginNode != 0 && r.OriginNode == nodeID
 
 	var (
-		mapResp *tailcfg.MapResponse
+		resp    *tailcfg.MapResponse
+		removed []tailcfg.NodeID
 		err     error
 	)
 
@@ -122,51 +128,60 @@ func generateMapResponse(nc nodeConnection, mapper *mapper, r change.Change) (*t
 			currentPeerIDs = append(currentPeerIDs, peer.ID().NodeID())
 		}
 
-		removedPeers := nc.computePeerDiff(currentPeerIDs)
-		// Routes and policy changes also affect self. Each connection drops
-		// the self node when it already holds the same value.
-		mapResp, err = mapper.policyChangeResponse(nodeID, version, removedPeers, currentPeers, true)
+		removed = nc.computePeerDiff(currentPeerIDs)
+		// Include self node when this is a self-update (e.g., node's own tags changed)
+		// so the node sees its updated self info along with new packet filters.
+		resp, err = mapper.policyChangeResponse(nodeID, version, currentPeers, true)
 	} else if isSelfUpdate {
 		// Non-policy self-update: just send the self node info
-		mapResp, err = mapper.selfMapResponse(nodeID, version)
+		resp, err = mapper.selfMapResponse(nodeID, version)
 	} else {
-		mapResp, err = mapper.buildFromChange(nodeID, version, &r)
+		resp, err = mapper.buildFromChange(nodeID, version, &r)
 	}
 
 	if err != nil {
 		return nil, fmt.Errorf("generating map response for nodeID %d: %w", nodeID, err)
 	}
 
-	// When a full update (SendAllPeers=true) produces zero visible peers
-	// (e.g., a restrictive policy isolates this node), the resulting
-	// [tailcfg.MapResponse] has Peers: []*tailcfg.Node{} (empty non-nil slice).
-	//
-	// The Tailscale client only treats Peers as a full authoritative
-	// replacement when len(Peers) > 0 (controlclient/map.go:462).
-	// An empty Peers slice is indistinguishable from a delta response,
-	// so the client silently preserves its existing peer state.
-	//
-	// This matters when a [change.FullUpdate] replaces a pending
-	// [change.PolicyChange] in the batcher ([Batcher.addToBatch]
-	// short-circuits on [change.HasFull]). The [change.PolicyChange]
-	// would have computed PeersRemoved via
-	// [multiChannelNodeConn.computePeerDiff], but the [change.FullUpdate]
-	// path uses [MapResponseBuilder.WithPeers] which sets Peers: [].
-	//
-	// Fix: when a full update results in zero peers, compute the diff
-	// against lastSentPeers and add explicit PeersRemoved entries so
-	// the client correctly clears its stale peer state.
-	if mapResp != nil && r.SendAllPeers && len(mapResp.Peers) == 0 {
-		removedPeers := nc.computePeerDiff(nil)
-		if len(removedPeers) > 0 {
-			mapResp.PeersRemoved = removedPeers
-		}
+	if resp == nil {
+		return nil, nil
 	}
 
-	return mapResp, nil
+	// A full peer list replaces the node's peer set, so peers missing from
+	// it are removals. Clients take Peers as authoritative only when it is
+	// non-empty, so a full update that isolates a node (e.g.
+	// [Batcher.addToBatch] replacing a pending [change.PolicyChange] with a
+	// [change.FullUpdate]) needs them sent explicitly.
+	if resp.Peers != nil {
+		peerIDs := make([]tailcfg.NodeID, 0, len(resp.Peers))
+		for _, peer := range resp.Peers {
+			peerIDs = append(peerIDs, peer.ID)
+		}
+
+		removed = nc.computePeerDiff(peerIDs)
+	}
+
+	return withRemovals(resp, removed), nil
 }
 
-// handleNodeChange generates and sends a [tailcfg.MapResponse] for a given node and [change.Change].
+// withRemovals returns resp and the peers removed from the node's view as the
+// responses to send. The removals lead, in a response of their own: clients
+// before Tailscale v1.104 only report removed peers to IPN bus watchers opted
+// out of full netmaps (the Android app since 1.100) from responses they apply
+// as deltas, and resp can carry DNSConfig, SSHPolicy, Node or Peers, which
+// force a full rebuild.
+//
+// TODO(kradalby): when the tailscale/tailscale#15660 compat goes (see
+// capver.CanOldCodeBeCleanedUp), set resp.PeersRemoved and return resp alone.
+func withRemovals(resp *tailcfg.MapResponse, removed []tailcfg.NodeID) []*tailcfg.MapResponse {
+	if len(removed) == 0 {
+		return []*tailcfg.MapResponse{resp}
+	}
+
+	return []*tailcfg.MapResponse{{PeersRemoved: removed}, resp}
+}
+
+// handleNodeChange generates and sends the [tailcfg.MapResponse] values for a given node and [change.Change].
 func handleNodeChange(nc nodeConnection, mapper *mapper, r change.Change) error {
 	if nc == nil {
 		return ErrNodeConnectionNil
@@ -176,35 +191,31 @@ func handleNodeChange(nc nodeConnection, mapper *mapper, r change.Change) error
 
 	log.Debug().Caller().Uint64(zf.NodeID, nodeID.Uint64()).Str(zf.Reason, r.Reason).Msg("node change processing started")
 
-	data, err := generateMapResponse(nc, mapper, r)
+	resps, err := generateMapResponse(nc, mapper, r)
 	if err != nil {
 		return fmt.Errorf("generating map response for node %d: %w", nodeID, err)
 	}
 
-	if data == nil {
-		// No data to send is valid for some response types
-		return nil
-	}
+	for _, resp := range resps {
+		err = nc.send(resp)
+		if err != nil {
+			// If the node has no active connections, the data was not
+			// delivered. Do not update lastSentPeers — recording phantom
+			// peer state would corrupt future computePeerDiff calculations,
+			// causing the node to miss peer additions or removals after
+			// reconnection.
+			if errors.Is(err, errNoActiveConnections) {
+				return nil
+			}
 
-	// Send the map response
-	err = nc.send(data)
-	if err != nil {
-		// If the node has no active connections, the data was not
-		// delivered. Do not update lastSentPeers — recording phantom
-		// peer state would corrupt future computePeerDiff calculations,
-		// causing the node to miss peer additions or removals after
-		// reconnection.
-		if errors.Is(err, errNoActiveConnections) {
-			return nil
+			return fmt.Errorf("sending map response to node %d: %w", nodeID, err)
 		}
 
-		return fmt.Errorf("sending map response to node %d: %w", nodeID, err)
+		// Update peer tracking only after confirmed delivery to at
+		// least one active connection.
+		nc.updateSentPeers(resp)
 	}
 
-	// Update peer tracking only after confirmed delivery to at
-	// least one active connection.
-	nc.updateSentPeers(data)
-
 	return nil
 }
 
@@ -341,6 +352,29 @@ func (b *Batcher) AddNode(
 		// path, and under workMu so a concurrent async bundle for this node
 		// cannot interleave its own lastSentPeers update.
 		nodeConn.workMu.Lock()
+
+		// Peers removed while the node had no stream: the initial map is a
+		// full rebuild, so follow it with the removal as a delta, see
+		// [withRemovals].
+		//
+		// TODO(kradalby): delete with the tailscale/tailscale#15660 compat, see
+		// capver.CanOldCodeBeCleanedUp.
+		peerIDs := make([]tailcfg.NodeID, 0, len(initialMap.Peers))
+		for _, peer := range initialMap.Peers {
+			peerIDs = append(peerIDs, peer.ID)
+		}
+
+		removed := make([]types.NodeID, 0)
+		for _, peerID := range nodeConn.computePeerDiff(peerIDs) {
+			removed = append(removed, types.NodeID(peerID)) //nolint:gosec // NodeID types are equivalent
+		}
+
+		if len(removed) > 0 {
+			rm := change.PeersRemoved(removed...)
+			rm.TargetNode = id
+			b.AddWork(rm)
+		}
+
 		nodeConn.updateSentPeers(initialMap)
 		nodeConn.workMu.Unlock()
 
@@ -525,9 +559,12 @@ func (b *Batcher) worker(workerID int) {
 					// node waits until the initial map is sent.
 					nc.workMu.Lock()
 
-					var err error
-
-					result.mapResponse, err = generateMapResponse(nc, b.mapper, w.changes[0])
+					resps, err := generateMapResponse(nc, b.mapper, w.changes[0])
+					if len(resps) > 0 {
+						// The initial map must be the stream's first frame; AddNode
+						// sends any removals ahead of it after it instead.
+						result.mapResponse = resps[len(resps)-1]
+					}
 
 					result.err = err
 					if result.err != nil {

hscontrol/mapper/batcher_test.go

@@ -19,6 +19,7 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 	"tailscale.com/tailcfg"
+	"tailscale.com/types/netmap"
 )
 
 var errNodeNotFoundAfterAdd = errors.New("node not found after adding to batcher")
@@ -2159,9 +2160,8 @@ func TestNodeDeletedWhileChangesPending(t *testing.T) {
 
 			// Delete the node from state - this returns a NodeRemoved change
 			// In production, this change is sent to batcher via app.Change()
-			nodeChange, err := st.DeleteNode(nodeToDelete)
+			nodeChanges, err := st.DeleteNode(nodeToDelete)
 			require.NoError(t, err, "should be able to delete node from state")
-			t.Logf("Deleted node %d from state, change: %s", node3.n.ID, nodeChange.Reason)
 
 			// Verify node is deleted from state
 			_, exists := st.GetNodeByID(node3.n.ID)
@@ -2169,7 +2169,7 @@ func TestNodeDeletedWhileChangesPending(t *testing.T) {
 
 			// Send the NodeRemoved change to batcher (this is what app.Change() does)
 			// With the fix, this should clean up node3 from batcher's internal state
-			batcher.AddWork(nodeChange)
+			batcher.AddWork(nodeChanges...)
 
 			// Wait for the batcher to process the removal and clean up the node
 			assert.EventuallyWithT(t, func(c *assert.CollectT) {
@@ -2415,3 +2415,116 @@ func TestSelfSentOnlyWhenChanged(t *testing.T) {
 	assert.Contains(t, frames[0].Node.Name, "renamed-again")
 	assert.Nil(t, frames[1].Node, "second connection already holds the new self")
 }
+
+// TestHandleNodeChangeSendsRemovalsAsDelta checks that peers missing from a
+// response listing the node's complete peer set go out first as a response
+// of their own that clients can apply as a delta, and that the full-set
+// response never carries them. Clients only report removals to IPN bus
+// watchers opted out of full netmaps from delta responses.
+//
+// TODO(kradalby): with the tailscale/tailscale#15660 compat gone, removals
+// ride the full-set response; adapt this and TestHandleNodeChangeRetryAfterRemoval.
+func TestHandleNodeChangeSendsRemovalsAsDelta(t *testing.T) {
+	tests := []struct {
+		name  string
+		nodes int
+		ch    change.Change
+	}{
+		{"policy_change", 2, change.PolicyChange()},
+		{"full_update", 2, change.FullUpdate()},
+		// Clients ignore an empty Peers list, so an isolating full update
+		// relies on the removal entirely.
+		{"full_update_isolated", 1, change.FullUpdate()},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			testData, cleanup := setupBatcherWithTestData(t, NewBatcherAndMapper, 1, tt.nodes, normalBufferSize)
+			defer cleanup()
+
+			gone := tailcfg.NodeID(999)
+
+			mc := newMockNodeConnection(testData.Nodes[0].n.ID)
+			mc.peers.Store(gone, struct{}{})
+
+			for i := range testData.Nodes[1:] {
+				mc.peers.Store(testData.Nodes[i+1].n.ID.NodeID(), struct{}{})
+			}
+
+			require.NoError(t, handleNodeChange(mc, testData.Batcher.mapper, tt.ch))
+
+			sent := mc.getSent()
+			require.Len(t, sent, 2)
+
+			assert.Equal(t, []tailcfg.NodeID{gone}, sent[0].PeersRemoved)
+			_, ok := netmap.MutationsFromMapResponse(sent[0], time.Time{})
+			assert.True(t, ok, "removal must be applicable as a delta: %+v", sent[0])
+
+			assert.Empty(t, sent[1].PeersRemoved, "full-set response must not carry removals")
+
+			_, tracked := mc.peers.Load(gone)
+			assert.False(t, tracked, "removed peer must leave lastSentPeers")
+		})
+	}
+}
+
+// TestHandleNodeChangeRetryAfterRemoval checks that a change retried after
+// its removal was delivered but its content was not does not repeat the
+// removal.
+func TestHandleNodeChangeRetryAfterRemoval(t *testing.T) {
+	testData, cleanup := setupBatcherWithTestData(t, NewBatcherAndMapper, 1, 2, normalBufferSize)
+	defer cleanup()
+
+	gone := tailcfg.NodeID(999)
+
+	mc := newMockNodeConnection(testData.Nodes[0].n.ID)
+	mc.peers.Store(gone, struct{}{})
+
+	var sends int
+
+	mc.sendFn = func(resp *tailcfg.MapResponse) error {
+		sends++
+		if sends == 2 {
+			return errNoReadyConnections
+		}
+
+		mc.sent = append(mc.sent, resp)
+
+		return nil
+	}
+
+	err := handleNodeChange(mc, testData.Batcher.mapper, change.PolicyChange())
+	require.ErrorIs(t, err, errNoReadyConnections)
+	require.NoError(t, handleNodeChange(mc, testData.Batcher.mapper, change.PolicyChange()))
+
+	sent := mc.getSent()
+	require.Len(t, sent, 2, "removal once, then the retried content")
+	assert.Equal(t, []tailcfg.NodeID{gone}, sent[0].PeersRemoved)
+	assert.Empty(t, sent[1].PeersRemoved)
+	assert.NotEmpty(t, sent[1].PacketFilters)
+}
+
+// TestDNSConfigOnlyWithSelfRefresh checks policy responses leave DNSConfig
+// out, which clients read as unchanged, while self refreshes carry it: a
+// node's DNS config derives from its own CapMap and Hostinfo, and a
+// DNSConfig forces clients into a full netmap rebuild.
+func TestDNSConfigOnlyWithSelfRefresh(t *testing.T) {
+	testData, cleanup := setupBatcherWithTestData(t, NewBatcherAndMapper, 1, 2, normalBufferSize)
+	defer cleanup()
+
+	testData.Config.TailcfgDNSConfig = &tailcfg.DNSConfig{
+		Proxied: true,
+		Domains: []string{"headscale.test"},
+	}
+
+	self := testData.Nodes[0].n.ID
+	mc := newMockNodeConnection(self)
+
+	require.NoError(t, handleNodeChange(mc, testData.Batcher.mapper, change.PolicyChange()))
+	require.NoError(t, handleNodeChange(mc, testData.Batcher.mapper, change.SelfUpdate(self)))
+
+	sent := mc.getSent()
+	require.Len(t, sent, 2)
+	assert.Nil(t, sent[0].DNSConfig, "policy response must not carry DNSConfig")
+	assert.NotNil(t, sent[1].DNSConfig, "self refresh must carry DNSConfig")
+}

hscontrol/mapper/mapper.go

@@ -318,19 +318,17 @@ func (m *mapper) selfMapResponse(
 }
 
 // policyChangeResponse creates a [tailcfg.MapResponse] for policy changes.
-// It sends:
-//   - PeersRemoved for peers that are no longer visible after the policy change
+// Peers no longer visible after the change are sent separately, see
+// [handleNodeChange]. It sends:
 //   - PeersChanged for remaining peers (their AllowedIPs may have changed due to policy)
 //   - Updated PacketFilters
 //   - Updated SSHPolicy (SSH rules may reference users/groups that changed)
-//   - DNSConfig so the client's resolver state stays anchored even when a
-//     policy-triggered wgengine reconfigure races a netmon LinkChange (the
-//     LinkChange handler reapplies dns.Manager.Set with the engine's
-//     lastDNSConfig; if that snapshot is stale, the OS resolver loses the
-//     MagicDNS reverse-DNS routes and Nameservers and curl-by-FQDN stops
-//     resolving for the rest of the policy window).
 //   - Optionally, the node's own self info (when includeSelf is true)
 //
+// DNSConfig is left out: it forces clients into a full netmap rebuild, and
+// the node's DNS config inputs arrive with its own [change.SelfUpdate], see
+// [state.State.DrainSelfRefreshes].
+//
 // This avoids the issue where an empty Peers slice is interpreted by Tailscale
 // clients as "no change" rather than "no peers".
 // When includeSelf is true, the node's self info is included so that a node
@@ -339,14 +337,12 @@ func (m *mapper) selfMapResponse(
 func (m *mapper) policyChangeResponse(
 	nodeID types.NodeID,
 	capVer tailcfg.CapabilityVersion,
-	removedPeers []tailcfg.NodeID,
 	currentPeers views.Slice[types.NodeView],
 	includeSelf bool,
 ) (*tailcfg.MapResponse, error) {
 	builder := m.NewMapResponseBuilder(nodeID).
 		WithDebugType(policyResponseDebug).
 		WithCapabilityVersion(capVer).
-		WithDNSConfig().
 		WithPacketFilters().
 		WithSSHPolicy()
 
@@ -354,16 +350,6 @@ func (m *mapper) policyChangeResponse(
 		builder = builder.WithSelfNode()
 	}
 
-	if len(removedPeers) > 0 {
-		// Convert [tailcfg.NodeID] to [types.NodeID] for [MapResponseBuilder.WithPeersRemoved]
-		removedIDs := make([]types.NodeID, len(removedPeers))
-		for i, id := range removedPeers {
-			removedIDs[i] = types.NodeID(id) //nolint:gosec // NodeID types are equivalent
-		}
-
-		builder.WithPeersRemoved(removedIDs...)
-	}
-
 	// Send remaining peers in PeersChanged - their AllowedIPs may have
 	// changed due to the policy update (e.g., different routes allowed).
 	// Cross-user peers must also carry their user profile, otherwise the

hscontrol/policy/v2/policy.go

@@ -10,6 +10,7 @@ import (
 	"slices"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/juanfont/headscale/hscontrol/policy/matcher"
@@ -88,6 +89,9 @@ type PolicyManager struct {
 	nodeAttrsMap     map[types.NodeID]tailcfg.NodeCapMap
 	nodeAttrsHashes  map[types.NodeID]deephash.Sum
 	nodeAttrsChanged []types.NodeID
+
+	// Skip the policy lock when no node-attribute refresh is pending.
+	nodeAttrsPending atomic.Bool
 }
 
 // filterAndPolicy combines the compiled filter rules with policy content for hashing.
@@ -1877,6 +1881,10 @@ func (pm *PolicyManager) refreshNodeAttrsLocked() error {
 	pm.nodeAttrsHashes = newHashes
 	pm.nodeAttrsChanged = append(pm.nodeAttrsChanged, changed...)
 
+	if len(pm.nodeAttrsChanged) > 0 {
+		pm.nodeAttrsPending.Store(true)
+	}
+
 	return nil
 }
 
@@ -1924,9 +1932,9 @@ func (pm *PolicyManager) NodeCapMaps() map[types.NodeID]tailcfg.NodeCapMap {
 
 // NodesWithChangedCapMap returns the IDs of nodes whose nodeAttrs
 // CapMap shifted across one or more [PolicyManager.updateLocked] calls
-// since the last drain. The buffer drains on return. The mapper calls
-// this once per [state.State.ReloadPolicy] to decide which nodes need
-// a [change.SelfUpdate].
+// since the last drain. The buffer drains on return.
+// [state.State.DrainSelfRefreshes] calls this whenever changes are
+// dispatched to decide which nodes need a [change.SelfUpdate].
 //
 // [PolicyManager.refreshNodeAttrsLocked] APPENDS to the buffer; the drain
 // returns the union of every change since the previous read. A concurrent
@@ -1934,7 +1942,7 @@ func (pm *PolicyManager) NodeCapMaps() map[types.NodeID]tailcfg.NodeCapMap {
 // [PolicyManager.SetPolicy] and a drain cannot silently lose the
 // policy-reload diff.
 func (pm *PolicyManager) NodesWithChangedCapMap() []types.NodeID {
-	if pm == nil {
+	if pm == nil || !pm.nodeAttrsPending.Load() {
 		return nil
 	}
 
@@ -1943,6 +1951,7 @@ func (pm *PolicyManager) NodesWithChangedCapMap() []types.NodeID {
 
 	out := pm.nodeAttrsChanged
 	pm.nodeAttrsChanged = nil
+	pm.nodeAttrsPending.Store(false)
 
 	return out
 }

hscontrol/poll_test.go

@@ -596,9 +596,9 @@ func TestDeletedNodeEndsLongPoll(t *testing.T) {
 		t.Fatal("expected the initial map write to start")
 	}
 
-	c, err := app.state.DeleteNode(nodeView)
+	cs, err := app.state.DeleteNode(nodeView)
 	require.NoError(t, err)
-	app.Change(c)
+	app.Change(cs...)
 
 	// The reconnect grace period is 10s, so a generous bound here still fails
 	// if teardown waits for a node that can never come back.
@@ -677,9 +677,9 @@ func TestDeletedNodeInterruptsBlockedWrite(t *testing.T) {
 		t.Fatal("expected the initial map write to block")
 	}
 
-	c, err := app.state.DeleteNode(nodeView)
+	cs, err := app.state.DeleteNode(nodeView)
 	require.NoError(t, err)
-	app.Change(c)
+	app.Change(cs...)
 
 	select {
 	case <-serveDone:

hscontrol/servertest/client.go

@@ -3,6 +3,7 @@ package servertest
 import (
 	"context"
 	"fmt"
+	"slices"
 	"sync"
 	"testing"
 	"time"
@@ -16,7 +17,9 @@ import (
 	"tailscale.com/types/key"
 	"tailscale.com/types/netmap"
 	"tailscale.com/types/persist"
+	"tailscale.com/types/views"
 	"tailscale.com/util/eventbus"
+	"tailscale.com/wgengine/filter"
 )
 
 // TestClient wraps a Tailscale [controlclient.Direct] connected to a
@@ -42,6 +45,10 @@ type TestClient struct {
 	netmap  *netmap.NetworkMap
 	history []*netmap.NetworkMap
 
+	// deltaUpdates and deltaRemoved: see [WithDeltaUpdates].
+	deltaUpdates bool
+	deltaRemoved []tailcfg.NodeID
+
 	// updates is a buffered channel that receives a signal
 	// each time a new [netmap.NetworkMap] arrives.
 	updates chan *netmap.NetworkMap
@@ -55,10 +62,19 @@ type TestClient struct {
 type ClientOption func(*clientConfig)
 
 type clientConfig struct {
-	ephemeral bool
-	hostname  string
-	tags      []string
-	user      *types.User
+	ephemeral    bool
+	hostname     string
+	tags         []string
+	user         *types.User
+	deltaUpdates bool
+}
+
+// WithDeltaUpdates makes the client accept incremental map updates the way
+// a real tailscaled does, and record the peer removals it could apply that
+// way ([TestClient.DeltaRemovedPeers]). Only those removals reach IPN bus
+// watchers opted out of full netmaps, such as the Android app.
+func WithDeltaUpdates() ClientOption {
+	return func(c *clientConfig) { c.deltaUpdates = true }
 }
 
 // WithEphemeral makes the client register as an ephemeral node.
@@ -156,6 +172,8 @@ func NewClient(tb testing.TB, server *TestServer, name string, opts ...ClientOpt
 		bus:     bus,
 		dialer:  dialer,
 		tracker: tracker,
+
+		deltaUpdates: cc.deltaUpdates,
 	}
 
 	tb.Cleanup(func() {
@@ -197,9 +215,13 @@ func (c *TestClient) startPoll(tb testing.TB) {
 
 	go func() {
 		defer close(c.pollDone)
-		// [controlclient.Direct.PollNetMap] blocks until ctx is cancelled or the server closes
-		// the connection.
-		_ = c.direct.PollNetMap(c.pollCtx, c)
+
+		var updater controlclient.NetmapUpdater = c
+		if c.deltaUpdates {
+			updater = deltaRecorder{c}
+		}
+
+		_ = c.direct.PollNetMap(c.pollCtx, updater)
 	}()
 }
 
@@ -218,6 +240,43 @@ func (c *TestClient) UpdateFullNetmap(nm *netmap.NetworkMap) {
 	}
 }
 
+// deltaRecorder is the [controlclient.NetmapUpdater] of a [WithDeltaUpdates]
+// client. Like tailscaled it takes packet filter and user profile updates
+// narrowly, which lets [controlclient.Direct] try the delta path. It
+// records removals and declines the mutations, so the full netmap is
+// still delivered and the client's peer state stays exact.
+type deltaRecorder struct{ *TestClient }
+
+func (d deltaRecorder) UpdatePacketFilter(views.Slice[tailcfg.FilterRule], []filter.Match) bool {
+	return true
+}
+
+func (d deltaRecorder) UpdateUserProfiles(map[tailcfg.UserID]tailcfg.UserProfileView) bool {
+	return true
+}
+
+func (d deltaRecorder) UpdateNetmapDelta(muts []netmap.NodeMutation) bool {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	for _, m := range muts {
+		if _, ok := m.(netmap.NodeMutationRemove); ok {
+			d.deltaRemoved = append(d.deltaRemoved, m.NodeIDBeingMutated())
+		}
+	}
+
+	return false
+}
+
+// DeltaRemovedPeers returns the peers whose removal arrived in a response
+// the client could apply incrementally; see [WithDeltaUpdates].
+func (c *TestClient) DeltaRemovedPeers() []tailcfg.NodeID {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+
+	return slices.Clone(c.deltaRemoved)
+}
+
 // cleanup releases all resources.
 func (c *TestClient) cleanup() {
 	if c.pollCancel != nil {

hscontrol/servertest/issues_test.go

@@ -648,7 +648,7 @@ func TestIssuesServerMutations(t *testing.T) {
 
 		deleteChange, err := srv.State().DeleteNode(node2View)
 		require.NoError(t, err)
-		srv.App.Change(deleteChange)
+		srv.App.Change(deleteChange...)
 
 		c1.WaitForCondition(t, "deleted peer gone", 10*time.Second,
 			func(nm *netmap.NetworkMap) bool {
@@ -1087,6 +1087,100 @@ func TestIssuesIdentity(t *testing.T) {
 	})
 }
 
+// TestPeerRemovedAsDelta checks a peer leaving a client's view reaches it
+// in a response it can apply as a delta. Bundled with DNSConfig or
+// SSHPolicy, the removal forces a full netmap rebuild, which never tells IPN
+// bus watchers opted out of full netmaps (the Android app since 1.100) that
+// the peer is gone, so they keep showing it.
+//
+// TODO(kradalby): with the tailscale/tailscale#15660 compat gone, removals
+// ride full rebuilds; assert they reach the netmap instead.
+func TestPeerRemovedAsDelta(t *testing.T) {
+	t.Parallel()
+
+	// MagicDNS puts DNSConfig in policy responses, making them full rebuilds.
+	setup := func(t *testing.T) (*servertest.TestServer, *servertest.TestClient, types.NodeID) {
+		t.Helper()
+
+		srv := servertest.NewServer(t, servertest.WithMagicDNS("delta.example.com"))
+		user1 := srv.CreateUser(t, "delta-user1")
+		user2 := srv.CreateUser(t, "delta-user2")
+
+		c1 := servertest.NewClient(t, srv, "delta-node1",
+			servertest.WithUser(user1), servertest.WithDeltaUpdates())
+		servertest.NewClient(t, srv, "delta-node2", servertest.WithUser(user2))
+
+		c1.WaitForPeers(t, 1, 15*time.Second)
+
+		return srv, c1, findNodeID(t, srv, "delta-node2")
+	}
+
+	assertRemovedAsDelta := func(t *testing.T, c1 *servertest.TestClient, id types.NodeID) {
+		t.Helper()
+
+		assert.EventuallyWithT(t, func(c *assert.CollectT) {
+			assert.Contains(c, c1.DeltaRemovedPeers(), id.NodeID())
+		}, 10*time.Second, 50*time.Millisecond, "peer removal never applied as a delta")
+	}
+
+	t.Run("node_deleted", func(t *testing.T) {
+		t.Parallel()
+
+		srv, c1, nodeID2 := setup(t)
+
+		node2, ok := srv.State().GetNodeByID(nodeID2)
+		require.True(t, ok)
+
+		changes, err := srv.State().DeleteNode(node2)
+		require.NoError(t, err)
+		srv.App.Change(changes...)
+
+		assertRemovedAsDelta(t, c1, nodeID2)
+	})
+
+	// The new stream's initial map is a full rebuild, so a removal missed
+	// while disconnected must still follow as a delta.
+	t.Run("deleted_while_disconnected", func(t *testing.T) {
+		t.Parallel()
+
+		srv, c1, nodeID2 := setup(t)
+
+		node2, ok := srv.State().GetNodeByID(nodeID2)
+		require.True(t, ok)
+
+		c1.Disconnect(t)
+
+		changes, err := srv.State().DeleteNode(node2)
+		require.NoError(t, err)
+		srv.App.Change(changes...)
+
+		c1.Reconnect(t)
+
+		assertRemovedAsDelta(t, c1, nodeID2)
+	})
+
+	t.Run("hidden_by_policy", func(t *testing.T) {
+		t.Parallel()
+
+		srv, c1, nodeID2 := setup(t)
+
+		changed, err := srv.State().SetPolicy([]byte(`{
+			"acls": [
+				{"action": "accept", "src": ["delta-user1@"], "dst": ["delta-user1@:*"]},
+				{"action": "accept", "src": ["delta-user2@"], "dst": ["delta-user2@:*"]}
+			]
+		}`))
+		require.NoError(t, err)
+		require.True(t, changed)
+
+		changes, err := srv.State().ReloadPolicy()
+		require.NoError(t, err)
+		srv.App.Change(changes...)
+
+		assertRemovedAsDelta(t, c1, nodeID2)
+	})
+}
+
 func findNodeID(tb testing.TB, srv *servertest.TestServer, hostname string) types.NodeID {
 	tb.Helper()
 

hscontrol/servertest/nodeattrs_test.go

@@ -3,6 +3,7 @@ package servertest_test
 import (
 	"context"
 	"net/netip"
+	"strings"
 	"testing"
 	"time"
 
@@ -462,3 +463,96 @@ func TestSuggestExitNodeDefaultOnPeerCapMap(t *testing.T) {
 	viewer.WaitForCondition(t, "peer suggest-exit-node gone after unapprove",
 		10*time.Second, peerHasCap(false))
 }
+
+// firstResolver returns the address of the first DNS resolver in nm.
+func firstResolver(nm *netmap.NetworkMap) string {
+	if nm == nil || len(nm.DNS.Resolvers) == 0 {
+		return ""
+	}
+
+	return nm.DNS.Resolvers[0].Addr
+}
+
+// TestNodeAttrsNextDNS checks a node's DNS config follows each of its
+// inputs: the NextDNS profile from nodeAttrs, whether reached through a
+// policy reload or a tag change, and the device metadata from its Hostinfo.
+// Policy responses do not carry DNSConfig, so each must arrive on its own.
+func TestNodeAttrsNextDNS(t *testing.T) {
+	t.Parallel()
+
+	setup := func(t *testing.T) (*servertest.TestServer, *servertest.TestClient) {
+		t.Helper()
+
+		srv := servertest.NewServer(t, servertest.WithDNSResolvers("https://dns.nextdns.io/base"))
+		user := srv.CreateUser(t, "nd-user")
+		c := servertest.NewClient(t, srv, "nd-node", servertest.WithUser(user))
+
+		c.WaitForCondition(t, "base resolver", 10*time.Second,
+			func(nm *netmap.NetworkMap) bool {
+				return strings.HasPrefix(firstResolver(nm), "https://dns.nextdns.io/base?")
+			})
+
+		return srv, c
+	}
+
+	waitResolver := func(t *testing.T, c *servertest.TestClient, prefix string) {
+		t.Helper()
+
+		c.WaitForCondition(t, "resolver "+prefix, 10*time.Second,
+			func(nm *netmap.NetworkMap) bool {
+				return strings.HasPrefix(firstResolver(nm), prefix)
+			})
+	}
+
+	t.Run("policy_reload", func(t *testing.T) {
+		t.Parallel()
+
+		srv, c := setup(t)
+
+		reloadPolicy(t, srv, `{
+			"acls":      [{"action": "accept", "src": ["*"], "dst": ["*:*"]}],
+			"nodeAttrs": [{"target": ["nd-user@"], "attr": ["nextdns:userprof"]}]
+		}`)
+
+		waitResolver(t, c, "https://dns.nextdns.io/userprof?")
+	})
+
+	t.Run("tag_change", func(t *testing.T) {
+		t.Parallel()
+
+		srv, c := setup(t)
+
+		reloadPolicy(t, srv, `{
+			"tagOwners": {"tag:dns": ["nd-user@"]},
+			"acls":      [{"action": "accept", "src": ["*"], "dst": ["*:*"]}],
+			"nodeAttrs": [{"target": ["tag:dns"], "attr": ["nextdns:tagprof"]}]
+		}`)
+
+		_, tagChange, err := srv.State().SetNodeTags(findNodeID(t, srv, "nd-node"), []string{"tag:dns"})
+		require.NoError(t, err)
+		srv.App.Change(tagChange)
+
+		waitResolver(t, c, "https://dns.nextdns.io/tagprof?")
+	})
+
+	t.Run("hostname_change", func(t *testing.T) {
+		t.Parallel()
+
+		_, c := setup(t)
+
+		c.Direct().SetHostinfo(&tailcfg.Hostinfo{
+			BackendLogID: "servertest-nd-node",
+			Hostname:     "nd-renamed",
+		})
+
+		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+
+		_ = c.Direct().SendUpdate(ctx)
+
+		c.WaitForCondition(t, "renamed device_name", 10*time.Second,
+			func(nm *netmap.NetworkMap) bool {
+				return strings.Contains(firstResolver(nm), "device_name=nd-renamed")
+			})
+	})
+}

hscontrol/servertest/race_test.go

@@ -197,7 +197,7 @@ func TestRaceConcurrentServerMutations(t *testing.T) {
 		wg.Go(func() {
 			delChange, err := srv.State().DeleteNode(nv2)
 			if err == nil {
-				srv.App.Change(delChange)
+				srv.App.Change(delChange...)
 			}
 		})
 

hscontrol/servertest/server.go

@@ -16,6 +16,7 @@ import (
 	"github.com/juanfont/headscale/hscontrol/types"
 	"tailscale.com/net/memnet"
 	"tailscale.com/tailcfg"
+	"tailscale.com/types/dnstype"
 )
 
 // TestServer is an in-process Headscale control server suitable for
@@ -43,6 +44,8 @@ type serverConfig struct {
 	nodeExpiry       time.Duration
 	batcherWorkers   int
 	taildropEnabled  bool
+	magicDNSDomain   string
+	dnsResolvers     []string
 }
 
 func defaultServerConfig() *serverConfig {
@@ -91,6 +94,18 @@ func WithTaildropEnabled(enabled bool) ServerOption {
 	return func(c *serverConfig) { c.taildropEnabled = enabled }
 }
 
+// WithMagicDNS enables MagicDNS under domain, so map responses carry a
+// [tailcfg.DNSConfig].
+func WithMagicDNS(domain string) ServerOption {
+	return func(c *serverConfig) { c.magicDNSDomain = domain }
+}
+
+// WithDNSResolvers sets the global DNS resolvers, so map responses carry a
+// [tailcfg.DNSConfig].
+func WithDNSResolvers(addrs ...string) ServerOption {
+	return func(c *serverConfig) { c.dnsResolvers = addrs }
+}
+
 // NewServer creates and starts a Headscale test server.
 // The server is fully functional and accepts real Tailscale control
 // protocol connections over Noise.
@@ -137,6 +152,20 @@ func NewServer(tb testing.TB, opts ...ServerOption) *TestServer {
 		},
 	}
 
+	if sc.magicDNSDomain != "" || len(sc.dnsResolvers) > 0 {
+		cfg.TailcfgDNSConfig = &tailcfg.DNSConfig{}
+	}
+
+	if sc.magicDNSDomain != "" {
+		cfg.BaseDomain = sc.magicDNSDomain
+		cfg.TailcfgDNSConfig.Proxied = true
+		cfg.TailcfgDNSConfig.Domains = []string{sc.magicDNSDomain}
+	}
+
+	for _, addr := range sc.dnsResolvers {
+		cfg.TailcfgDNSConfig.Resolvers = append(cfg.TailcfgDNSConfig.Resolvers, &dnstype.Resolver{Addr: addr})
+	}
+
 	app, err := hscontrol.NewHeadscale(&cfg)
 	if err != nil {
 		tb.Fatalf("servertest: NewHeadscale: %v", err)

hscontrol/servertest/stress_test.go

@@ -121,7 +121,7 @@ func TestStressConnectDisconnect(t *testing.T) {
 
 		deleteChange, err := srv.State().DeleteNode(nv)
 		require.NoError(t, err)
-		srv.App.Change(deleteChange)
+		srv.App.Change(deleteChange...)
 
 		// c2 should see c1 disappear.
 		c2.WaitForCondition(t, "deleted node gone", 10*time.Second,

hscontrol/state/maprequest.go

@@ -36,6 +36,10 @@ type mapRequestDelta struct {
 	// changed (see [peerHostinfo]). Only that forces a whole-node resend.
 	peerHostinfoChanged bool
 
+	// dnsMetadataChanged reports whether a Hostinfo field feeding the
+	// node's NextDNS device metadata (Hostname, OS) changed.
+	dnsMetadataChanged bool
+
 	// routesChanged reports whether announced routes (RoutableIPs)
 	// changed. Routes are policy and election inputs, so they are tracked
 	// on their own.

hscontrol/state/persist_test.go

@@ -510,7 +510,7 @@ func TestDeleteNodeKeepsStoreOnDBFailure(t *testing.T) {
 	c, err := s.DeleteNode(node)
 	require.NoError(t, s.db.DB.Callback().Delete().Remove("fail_node_delete"))
 	require.ErrorIs(t, err, errInjectedNodeDelete)
-	assert.True(t, c.IsEmpty(), "an uncommitted deletion must not stop the node's session")
+	assert.Empty(t, c, "an uncommitted deletion must not stop the node's session")
 
 	_, ok = s.GetNodeByID(nodeID)
 	assert.True(t, ok, "a database failure must leave the in-memory node available")
@@ -528,11 +528,12 @@ func TestDeleteNodeReturnsRemovalOnPolicyFailure(t *testing.T) {
 
 	s.polMan = failingSetNodesPolicyManager{PolicyManager: s.polMan}
 
-	c, err := s.DeleteNode(node)
+	cs, err := s.DeleteNode(node)
 	require.ErrorIs(t, err, errInjectedPolicyNodeUpdate)
-	assert.Equal(t, []types.NodeID{nodeID}, c.PeersRemoved,
+	require.NotEmpty(t, cs)
+	assert.Equal(t, []types.NodeID{nodeID}, cs[0].PeersRemoved,
 		"a committed deletion must still notify peers and stop the node's session")
-	assert.Equal(t, []types.NodeID{nodeID}, c.DeletedNodes,
+	assert.Equal(t, []types.NodeID{nodeID}, cs[0].DeletedNodes,
 		"a committed deletion must identify the session to stop")
 
 	_, ok = s.GetNodeByID(nodeID)

hscontrol/state/state.go

@@ -191,6 +191,11 @@ type State struct {
 	// caller snapshot or resurrected by an update racing with deletion.
 	persistMu sync.Mutex
 
+	// selfRefresh holds nodes whose NextDNS device metadata changed, drained
+	// by [State.DrainSelfRefreshes].
+	selfRefresh   []types.NodeID
+	selfRefreshMu sync.Mutex
+
 	// registerLocks serialises registration per machine key so concurrent
 	// registrations of the same machine resolve to a single node instead of
 	// racing the find-then-create section and each creating their own.
@@ -357,21 +362,11 @@ func (s *State) ReloadPolicy() ([]change.Change, error) {
 	// policies to not propagate correctly when switching between policy types.
 	s.nodeStore.RebuildPeerMaps()
 
+	// Nodes whose CapMap shifted get their self refresh from
+	// [State.DrainSelfRefreshes] when these changes are dispatched.
 	//nolint:prealloc // cs starts with one element and may grow
 	cs := []change.Change{change.PolicyChange()}
 
-	// Per-node selective self refresh for nodeAttrs. A broadcast
-	// [change.PolicyChange] re-renders peer lists and packet filters
-	// but never repopulates a node's own [tailcfg.Node.CapMap]; that
-	// lives on the self entry only. The drain returns every node ID
-	// whose cap output shifted across recent updateLocked calls —
-	// refreshNodeAttrsLocked appends rather than overwrites so a
-	// concurrent SetUsers/SetNodes between SetPolicy and the drain
-	// cannot silently lose the policy-reload diff.
-	for _, id := range s.polMan.NodesWithChangedCapMap() {
-		cs = append(cs, change.SelfUpdate(id))
-	}
-
 	// Always call autoApproveNodes during policy reload, regardless of whether
 	// the policy content has changed. This ensures that routes are re-evaluated
 	// when they might have been manually disabled but could now be auto-approved
@@ -604,18 +599,18 @@ func (s *State) SaveNode(node types.NodeView) (types.NodeView, change.Change, er
 }
 
 // DeleteNode permanently removes a node and cleans up associated resources.
-// Once the database deletion commits, the returned change always contains the
-// node-removal notification, even if a later policy refresh fails. Callers must
-// publish a non-empty change before handling the error so live sessions are
+// Once the database deletion commits, the returned changes always start with
+// the node-removal notification, even if a later policy refresh fails. Callers
+// must publish the changes before handling the error so live sessions are
 // still torn down after a committed deletion.
-func (s *State) DeleteNode(node types.NodeView) (change.Change, error) {
+func (s *State) DeleteNode(node types.NodeView) ([]change.Change, error) {
 	s.persistMu.Lock()
 
 	err := s.db.DeleteNode(node.AsStruct())
 	if err != nil {
 		s.persistMu.Unlock()
 
-		return change.Change{}, err
+		return nil, err
 	}
 
 	// The database is the durable source of truth. Only remove the in-memory
@@ -626,21 +621,16 @@ func (s *State) DeleteNode(node types.NodeView) (change.Change, error) {
 
 	s.ipAlloc.FreeIPs(node.IPs())
 
-	c := change.NodeRemoved(node.ID())
+	// An explicit removal of its own, ahead of the policy refresh, so peers
+	// learn of the deletion without depending on their sent-peers tracking.
+	removed := change.NodeRemoved(node.ID())
 
-	// Check if policy manager needs updating after node deletion
 	policyChange, err := s.updatePolicyManagerNodes()
 	if err != nil {
-		return c, fmt.Errorf("updating policy manager after node deletion: %w", err)
+		return []change.Change{removed, policyChange}, fmt.Errorf("updating policy manager after node deletion: %w", err)
 	}
 
-	if !policyChange.IsEmpty() {
-		// Merge policy change with NodeRemoved to preserve PeersRemoved info
-		// This ensures the batcher cleans up the deleted node from its state
-		c = c.Merge(policyChange)
-	}
-
-	return c, nil
+	return []change.Change{removed, policyChange}, nil
 }
 
 // Connect acquires a control session and returns the resulting changes
@@ -2878,6 +2868,29 @@ func (s *State) UpdatePolicyManagerUsersForTest() error {
 	return err
 }
 
+// DrainSelfRefreshes returns a [change.SelfUpdate] for every node whose own
+// entry or DNS config changed since the last call: its policy CapMap, from
+// any policy, user or node update, or its NextDNS device metadata. Drained
+// where changes are dispatched, so no path has to return the refresh itself.
+func (s *State) DrainSelfRefreshes() []change.Change {
+	ids := s.polMan.NodesWithChangedCapMap()
+
+	s.selfRefreshMu.Lock()
+	ids = append(ids, s.selfRefresh...)
+	s.selfRefresh = nil
+	s.selfRefreshMu.Unlock()
+
+	slices.Sort(ids)
+	ids = slices.Compact(ids)
+
+	cs := make([]change.Change, 0, len(ids))
+	for _, id := range ids {
+		cs = append(cs, change.SelfUpdate(id))
+	}
+
+	return cs
+}
+
 // updatePolicyManagerNodes updates the policy manager with current nodes.
 // Returns true if the policy changed and notifications should be sent.
 // TODO(kradalby): This is a temporary stepping stone, ultimately we should
@@ -3092,6 +3105,10 @@ func (s *State) UpdateNodeFromMapRequest(id types.NodeID, req tailcfg.MapRequest
 			!hostinfoEqual(currentNode.Hostinfo, newHostinfo)
 		delta.peerHostinfoChanged = newHostinfo != nil &&
 			!peerHostinfoEqual(currentNode.Hostinfo, newHostinfo)
+		delta.dnsMetadataChanged = newHostinfo != nil &&
+			(currentNode.Hostinfo == nil ||
+				currentNode.Hostinfo.Hostname != newHostinfo.Hostname ||
+				currentNode.Hostinfo.OS != newHostinfo.OS)
 
 		// A change carrying only an updated LastSeen is not worth a
 		// full-row database UPDATE plus the O(n) policy rescan
@@ -3193,6 +3210,12 @@ func (s *State) UpdateNodeFromMapRequest(id types.NodeID, req tailcfg.MapRequest
 		return change.Change{}, fmt.Errorf("%w: %d", ErrNodeNotInNodeStore, id)
 	}
 
+	if delta.dnsMetadataChanged {
+		s.selfRefreshMu.Lock()
+		s.selfRefresh = append(s.selfRefresh, id)
+		s.selfRefreshMu.Unlock()
+	}
+
 	if routeChange {
 		log.Debug().
 			Uint64(zf.NodeID, id.Uint64()).

hscontrol/types/change/change.go

@@ -337,11 +337,14 @@ func FullSelf(nodeID types.NodeID) Change {
 	}
 }
 
+// SelfUpdate sends a node its own entry and the DNS config derived from it
+// (NextDNS profile and device metadata).
 func SelfUpdate(nodeID types.NodeID) Change {
 	return Change{
 		Reason:      "self update",
 		TargetNode:  nodeID,
 		IncludeSelf: true,
+		IncludeDNS:  true,
 	}
 }
 

hscontrol/types/change/change_test.go

@@ -464,6 +464,7 @@ func TestSelfUpdate(t *testing.T) {
 	assert.Equal(t, "self update", r.Reason)
 	assert.Equal(t, types.NodeID(42), r.TargetNode)
 	assert.True(t, r.IncludeSelf)
+	assert.True(t, r.IncludeDNS, "DNS config derives from the node's own CapMap and Hostinfo")
 	assert.True(t, r.IsSelfOnly())
 }