state: fold primary route election into NodeStore snapshot

primaryRoutes lived as a separate write-then-read store guarded by an
external mutex. Each Connect, Disconnect, SetApprovedRoutes, and HA
prober write touched both NodeStore and primaryRoutes in sequence,
opening a TOCTOU window in which a stale Disconnect could overwrite a
fresh Connect's route assignment (issue #3203).

The election algorithm is a pure function of the snapshot: every node
contributes its IsOnline + AllApprovedRoutes + a runtime Unhealthy
bit. Move the computation into snapshotFromNodes so it runs in the
NodeStore writer goroutine and recomputes from immutable input on
every batch — no separate primaryRoutes mutation is required.

Caller-side changes are local. GetNodePrimaryRoutes, RoutesForPeer,
the HA prober and the debug endpoint all read PrimaryRoutesForNode /
HANodes / DebugRoutes off NodeStore. Health writes flow through
State.SetNodeUnhealthy, which captures the snapshot's primaries
before/after and signals back whether failover happened. The
SetApprovedRoutes and UpdateNodeFromMapRequest paths use the same
prevPrimaries / maps.Equal pattern to detect when announced/approved
changes shifted a primary.

The per-node lock and connectGen counter that closed the original
TOCTOU window are gone too: Connect bumps a SessionEpoch field on
the node inside its UpdateNode closure, Disconnect re-checks it in
its closure and no-ops on mismatch. The NodeStore writer goroutine
serialises both batches so the check + mutation are atomic by
construction.

Updates #3203

hscontrol/mapper/tail_test.go

@@ -9,7 +9,6 @@ import (
 
 	"github.com/google/go-cmp/cmp"
 	"github.com/google/go-cmp/cmp/cmpopts"
-	"github.com/juanfont/headscale/hscontrol/routes"
 	"github.com/juanfont/headscale/hscontrol/types"
 	"tailscale.com/net/tsaddr"
 	"tailscale.com/tailcfg"
@@ -209,25 +208,30 @@ func TestTailNode(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			primary := routes.New()
 			cfg := &types.Config{
 				BaseDomain:          tt.baseDomain,
 				TailcfgDNSConfig:    tt.dnsConfig,
 				RandomizeClientPort: false,
 				Taildrop:            types.TaildropConfig{Enabled: true},
 			}
-			_ = primary.SetRoutes(tt.node.ID, tt.node.SubnetRoutes()...)
 
-			// This is a hack to avoid having a second node to test the primary route.
+			// Stub primary-route lookup: tt.node owns its SubnetRoutes,
-			// This should be baked into the test case proper if it is extended in the future.
+			// node ID 2 owns 192.168.0.0/24 (a hack carried over from
-			_ = primary.SetRoutes(2, netip.MustParsePrefix("192.168.0.0/24"))
+			// the original routes-package-driven version of this test —
+			// avoids spinning up a second node just to validate that
+			// other nodes' primaries don't leak into tt.node's TailNode
+			// output).
+			primaries := map[types.NodeID][]netip.Prefix{
+				tt.node.ID: tt.node.SubnetRoutes(),
+				2:          {netip.MustParsePrefix("192.168.0.0/24")},
+			}
 			nv := tt.node.View()
 			got, err := nv.TailNode(
 				0,
 				func(id types.NodeID) []netip.Prefix {
 					// Route function returns primaries + exit routes
 					// (matching the real caller contract).
-					return slices.Concat(primary.PrimaryRoutes(id), nv.ExitRoutes())
+					return slices.Concat(primaries[id], nv.ExitRoutes())
 				},
 				cfg,
 			)

hscontrol/servertest/connect_race_test.go

@@ -141,8 +141,9 @@ func TestConnectDisconnectRace(t *testing.T) {
 			i, route, primary)
 
 		if t.Failed() {
-			t.Logf("primaryRoutes state at failure: %s",
+			t.Logf("primaryRoutes state at failure:\n%s",
-				srv.State().PrimaryRoutes())
+				srv.State().PrimaryRoutesString())
+
 			return
 		}
 	}

hscontrol/servertest/ha_dynamic_test.go

@@ -72,7 +72,7 @@ func TestHAFailover_ViewerSeesPrimaryFlip(t *testing.T) {
 			return hasPeerPrimaryRoute(nm, "dyn-flip-r1", route)
 		})
 
-	changed := srv.State().PrimaryRoutes().SetNodeHealthy(id1, false)
+	changed := srv.State().SetNodeUnhealthy(id1, true)
 	require.True(t, changed, "marking primary unhealthy should change primaries")
 
 	srv.App.Change(change.PolicyChange())

hscontrol/servertest/ha_health_test.go

@@ -80,7 +80,7 @@ func TestHAHealthProbe_HealthyNodes(t *testing.T) {
 	prober.ProbeOnce(ctx, srv.App.Change)
 
 	// Both nodes should be healthy, primary unchanged (node 1).
-	assert.True(t, srv.State().PrimaryRoutes().IsNodeHealthy(nodeID1))
+	assert.True(t, srv.State().IsNodeHealthy(nodeID1))
 
 	primaries := srv.State().GetNodePrimaryRoutes(nodeID1)
 	assert.Contains(t, primaries, route)
@@ -111,7 +111,7 @@ func TestHAHealthProbe_UnhealthyFailover(t *testing.T) {
 	require.Contains(t, primaries, route, "node 1 should be primary initially")
 
 	// Mark node 1 unhealthy — should failover to node 2.
-	changed := srv.State().PrimaryRoutes().SetNodeHealthy(nodeID1, false)
+	changed := srv.State().SetNodeUnhealthy(nodeID1, true)
 	assert.True(t, changed, "marking primary unhealthy should change primaries")
 
 	primaries2 := srv.State().GetNodePrimaryRoutes(nodeID2)
@@ -141,12 +141,12 @@ func TestHAHealthProbe_RecoveryNoFlap(t *testing.T) {
 	nodeID2 := advertiseAndApproveRoute(t, srv, c2, route)
 
 	// Failover: node 1 → node 2.
-	srv.State().PrimaryRoutes().SetNodeHealthy(nodeID1, false)
+	srv.State().SetNodeUnhealthy(nodeID1, true)
 	primaries := srv.State().GetNodePrimaryRoutes(nodeID2)
 	require.Contains(t, primaries, route, "node 2 should be primary")
 
 	// Recovery: node 1 healthy again. Node 2 should STAY primary.
-	changed := srv.State().PrimaryRoutes().SetNodeHealthy(nodeID1, true)
+	changed := srv.State().SetNodeUnhealthy(nodeID1, false)
 	assert.False(t, changed, "recovery should not change primaries (no flap)")
 
 	primaries = srv.State().GetNodePrimaryRoutes(nodeID2)
@@ -173,8 +173,8 @@ func TestHAHealthProbe_ConnectClearsUnhealthy(t *testing.T) {
 	advertiseAndApproveRoute(t, srv, c2, route)
 
 	// Mark unhealthy.
-	srv.State().PrimaryRoutes().SetNodeHealthy(nodeID1, false)
+	srv.State().SetNodeUnhealthy(nodeID1, true)
-	assert.False(t, srv.State().PrimaryRoutes().IsNodeHealthy(nodeID1))
+	assert.False(t, srv.State().IsNodeHealthy(nodeID1))
 
 	// Reconnect clears unhealthy via State.Connect → ClearUnhealthy.
 	c1.Disconnect(t)
@@ -182,7 +182,7 @@ func TestHAHealthProbe_ConnectClearsUnhealthy(t *testing.T) {
 
 	c1.WaitForPeers(t, 1, 10*time.Second)
 
-	assert.True(t, srv.State().PrimaryRoutes().IsNodeHealthy(nodeID1),
+	assert.True(t, srv.State().IsNodeHealthy(nodeID1),
 		"reconnect should clear unhealthy state")
 }
 

hscontrol/state/debug.go

@@ -2,11 +2,12 @@ package state
 
 import (
 	"fmt"
+	"net/netip"
+	"slices"
 	"strings"
 	"time"
 
 	hsdb "github.com/juanfont/headscale/hscontrol/db"
-	"github.com/juanfont/headscale/hscontrol/routes"
 	"github.com/juanfont/headscale/hscontrol/types"
 	"tailscale.com/tailcfg"
 )
@@ -132,8 +133,10 @@ func (s *State) DebugOverview() string {
 	sb.WriteString("\n")
 
 	// Route information
-	routeCount := len(strings.Split(strings.TrimSpace(s.primaryRoutes.String()), "\n"))
+	primaryStr := s.PrimaryRoutesString()
-	if s.primaryRoutes.String() == "" {
+
+	routeCount := len(strings.Split(strings.TrimSpace(primaryStr), "\n"))
+	if primaryStr == "" {
 		routeCount = 0
 	}
 
@@ -253,9 +256,55 @@ func (s *State) DebugFilter() ([]tailcfg.FilterRule, error) {
 	return filter, nil
 }
 
-// DebugRoutes returns the current primary routes information as a structured object.
+// DebugRoutes returns the current primary routes information as a
-func (s *State) DebugRoutes() routes.DebugRoutes {
+// structured object built from the NodeStore snapshot.
-	return s.primaryRoutes.DebugJSON()
+func (s *State) DebugRoutes() types.DebugRoutes {
+	debug := types.DebugRoutes{
+		AvailableRoutes: make(map[types.NodeID][]netip.Prefix),
+		PrimaryRoutes:   make(map[string]types.NodeID),
+	}
+
+	for _, nv := range s.nodeStore.ListNodes().All() {
+		if !nv.Valid() {
+			continue
+		}
+
+		online, known := nv.IsOnline().GetOk()
+		if !known || !online {
+			continue
+		}
+
+		approved := nv.AllApprovedRoutes()
+		if len(approved) == 0 {
+			continue
+		}
+
+		slices.SortFunc(approved, netip.Prefix.Compare)
+		debug.AvailableRoutes[nv.ID()] = approved
+	}
+
+	for prefix, id := range s.nodeStore.PrimaryRoutes() {
+		debug.PrimaryRoutes[prefix.String()] = id
+	}
+
+	var unhealthy []types.NodeID
+
+	for _, nv := range s.nodeStore.ListNodes().All() {
+		if !nv.Valid() {
+			continue
+		}
+
+		if !s.nodeStore.IsNodeHealthy(nv.ID()) {
+			unhealthy = append(unhealthy, nv.ID())
+		}
+	}
+
+	if len(unhealthy) > 0 {
+		slices.Sort(unhealthy)
+		debug.UnhealthyNodes = unhealthy
+	}
+
+	return debug
 }
 
 // DebugRoutesString returns the current primary routes information as a string.
@@ -322,8 +371,10 @@ func (s *State) DebugOverviewJSON() DebugOverviewInfo {
 	}
 
 	// Route information
-	routeCount := len(strings.Split(strings.TrimSpace(s.primaryRoutes.String()), "\n"))
+	primaryStr := s.PrimaryRoutesString()
-	if s.primaryRoutes.String() == "" {
+
+	routeCount := len(strings.Split(strings.TrimSpace(primaryStr), "\n"))
+	if primaryStr == "" {
 		routeCount = 0
 	}
 

hscontrol/state/ha_health.go

@@ -46,7 +46,7 @@ func (p *HAHealthProber) ProbeOnce(
 	ctx context.Context,
 	dispatch func(...change.Change),
 ) {
-	haNodes := p.state.primaryRoutes.HANodes()
+	haNodes := p.state.nodeStore.HANodes()
 	if len(haNodes) == 0 {
 		return
 	}
@@ -97,7 +97,7 @@ func (p *HAHealthProber) ProbeOnce(
 					Dur("latency", latency).
 					Msg("HA probe: node responded")
 
-				if p.state.primaryRoutes.SetNodeHealthy(id, true) {
+				if p.state.SetNodeUnhealthy(id, false) {
 					dispatch(change.PolicyChange())
 
 					log.Info().
@@ -121,7 +121,7 @@ func (p *HAHealthProber) ProbeOnce(
 					Dur("timeout", p.cfg.ProbeTimeout).
 					Msg("HA probe: node did not respond")
 
-				if p.state.primaryRoutes.SetNodeHealthy(id, false) {
+				if p.state.SetNodeUnhealthy(id, true) {
 					dispatch(change.PolicyChange())
 
 					log.Info().

hscontrol/state/node_store.go

@@ -916,6 +916,36 @@ func (s *NodeStore) IsNodeHealthy(id types.NodeID) bool {
 	return !n.Unhealthy
 }
 
+// PrimaryRoutes returns the snapshot's prefix→primary map. The map is
+// owned by the snapshot and must not be mutated; it is safe to read
+// concurrently because snapshots are immutable once published.
+func (s *NodeStore) PrimaryRoutes() map[netip.Prefix]types.NodeID {
+	return s.data.Load().routes
+}
+
+// PrimaryRoutesString renders the snapshot's prefix→primary map for
+// debug output and test diagnostics.
+func (s *NodeStore) PrimaryRoutesString() string {
+	snap := s.data.Load()
+	if len(snap.routes) == 0 {
+		return ""
+	}
+
+	prefixes := make([]netip.Prefix, 0, len(snap.routes))
+	for p := range snap.routes {
+		prefixes = append(prefixes, p)
+	}
+
+	slices.SortFunc(prefixes, netip.Prefix.Compare)
+
+	var b strings.Builder
+	for _, p := range prefixes {
+		fmt.Fprintf(&b, "%s: %d\n", p, snap.routes[p])
+	}
+
+	return b.String()
+}
+
 // RebuildPeerMaps rebuilds the peer relationship map using the current peersFunc.
 // This must be called after policy changes because peersFunc uses PolicyManager's
 // filters to determine which nodes can see each other. Without rebuilding, the

hscontrol/state/state.go

@@ -15,6 +15,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"maps"
 	"net/netip"
 	"slices"
 	"strconv"
@@ -27,7 +28,6 @@ import (
 	hsdb "github.com/juanfont/headscale/hscontrol/db"
 	"github.com/juanfont/headscale/hscontrol/policy"
 	"github.com/juanfont/headscale/hscontrol/policy/matcher"
-	"github.com/juanfont/headscale/hscontrol/routes"
 	"github.com/juanfont/headscale/hscontrol/types"
 	"github.com/juanfont/headscale/hscontrol/types/change"
 	"github.com/juanfont/headscale/hscontrol/util"
@@ -120,14 +120,6 @@ type sshCheckPair struct {
 	Dst types.NodeID
 }
 
-// nodeLock pairs a per-node Connect generation counter with a mutex
-// that serialises Connect/Disconnect transitions for that node.
-// See the State.nodeLocks field for the rationale (issue #3203).
-type nodeLock struct {
-	mu  sync.Mutex
-	gen atomic.Uint64
-}
-
 // State manages Headscale's core state, coordinating between database, policy management,
 // IP allocation, and DERP routing. All methods are thread-safe.
 type State struct {
@@ -154,25 +146,6 @@ type State struct {
 	// via the eviction callback so any waiting goroutines wake.
 	authCache *expirable.LRU[types.AuthID, *types.AuthRequest]
 
-	// primaryRoutes tracks primary route assignments for nodes
-	primaryRoutes *routes.PrimaryRoutes
-
-	// nodeLocks pairs a per-node monotonic Connect generation counter with
-	// a mutex that serialises Connect/Disconnect transitions for that
-	// node. Connect bumps gen and rewrites primaryRoutes for id;
-	// Disconnect re-checks gen UNDER the lock and only proceeds when its
-	// captured generation still matches. This closes the TOCTOU window
-	// between the gen check and the NodeStore + primaryRoutes mutations
-	// that allowed a stale Disconnect to overwrite a fresh Connect's
-	// primary route assignment (issue #3203).
-	//
-	// TODO(#3203 follow-up): if a similar race is found in
-	// SetApprovedRoutes, UpdateNodeFromMapRequest's route path, or
-	// SetNodeHealthy, those mutators should acquire this same per-node
-	// lock. Each requires its own reproduction test before being brought
-	// under the lock.
-	nodeLocks sync.Map // types.NodeID → *nodeLock
-
 	// pings tracks pending ping requests and their response channels.
 	pings *pingTracker
 
@@ -280,13 +253,12 @@ func NewState(cfg *types.Config) (*State, error) {
 	return &State{
 		cfg: cfg,
 
-		db:            db,
+		db:        db,
-		ipAlloc:       ipAlloc,
+		ipAlloc:   ipAlloc,
-		polMan:        polMan,
+		polMan:    polMan,
-		authCache:     authCache,
+		authCache: authCache,
-		primaryRoutes: routes.New(),
+		nodeStore: nodeStore,
-		nodeStore:     nodeStore,
+		pings:     newPingTracker(),
-		pings:         newPingTracker(),
 
 		sshCheckAuth: make(map[sshCheckPair]time.Time),
 	}, nil
@@ -577,51 +549,31 @@ func (s *State) DeleteNode(node types.NodeView) (change.Change, error) {
 	return c, nil
 }
 
-// Connect marks a node as connected and updates its primary routes in the state.
+// Connect marks a node connected and returns the resulting changes
-// It returns the list of changes and a generation number. The generation number
+// plus a session epoch. The caller must pass the epoch back to
-// must be passed to Disconnect() so that stale disconnects from old poll sessions
+// Disconnect so deferred grace-period disconnects from a previous
-// are rejected (see the grace period logic in poll.go).
+// poll session are dropped (see poll.go).
-//
-// The per-node lock is held across the entire body so that Connect's
-// gen bump, NodeStore mutation, and primaryRoutes.SetRoutes are atomic
-// with respect to a concurrent Disconnect on the same node. Without
-// this, a stale Disconnect that passed its gen check could overwrite
-// the fresh Connect's online status and approved-route assignment
-// (issue #3203).
 func (s *State) Connect(id types.NodeID) ([]change.Change, uint64) {
-	nl := s.nodeLockFor(id)
+	prevRoutes := s.nodeStore.PrimaryRoutes()
-	nl.mu.Lock()
-	defer nl.mu.Unlock()
 
-	// Increment the connect generation for this node. This ensures that any
+	// Reconnecting clears Unhealthy: the node just proved basic
-	// in-flight Disconnect() from a previous session will see a stale generation
+	// connectivity by completing the Noise handshake.
-	// and become a no-op.
+	var epoch uint64
-	gen := nl.gen.Add(1)
-
-	// Update online status in NodeStore before creating change notification
-	// so the NodeStore already reflects the correct state when other nodes
-	// process the NodeCameOnline change for full map generation.
 	node, ok := s.nodeStore.UpdateNode(id, func(n *types.Node) {
+		n.SessionEpoch++
+		epoch = n.SessionEpoch
 		n.IsOnline = new(true)
-		// n.LastSeen = ptr.To(now)
+		n.Unhealthy = false
 	})
 	if !ok {
-		return nil, gen
+		return nil, 0
 	}
 
 	c := []change.Change{change.NodeOnlineFor(node)}
 
 	log.Info().EmbedObject(node).Msg("node connected")
 
-	// Reconnecting clears any prior unhealthy state — the node proved
+	if !maps.Equal(prevRoutes, s.nodeStore.PrimaryRoutes()) {
-	// basic connectivity by establishing the Noise session.
-	s.primaryRoutes.ClearUnhealthy(id)
-
-	// Use the node's current routes for primary route update.
-	// AllApprovedRoutes() returns only the intersection of announced and approved routes.
-	routeChange := s.primaryRoutes.SetRoutes(id, node.AllApprovedRoutes()...)
-
-	if routeChange {
 		c = append(c, change.NodeAdded(id))
 	}
 
@@ -629,50 +581,21 @@ func (s *State) Connect(id types.NodeID) ([]change.Change, uint64) {
 	// aliases targeting this node, so peers need a fresh netmap.
 	c = append(c, change.PolicyChange())
 
-	return c, gen
+	return c, epoch
 }
 
-// nodeLockFor returns the per-node lock for id, creating it on first
+// Disconnect marks the node offline. epoch must match the value
-// use. The returned *nodeLock is stable for the lifetime of the State —
+// Connect returned for this session; otherwise the call no-ops so a
-// callers may cache the reference across operations on the same node.
+// deferred disconnect from an older session cannot overwrite state set
-func (s *State) nodeLockFor(id types.NodeID) *nodeLock {
+// by a newer Connect. The check and the IsOnline write share an
-	if v, ok := s.nodeLocks.Load(id); ok {
+// UpdateNode closure, making them atomic against concurrent Connects.
-		return v.(*nodeLock)
+func (s *State) Disconnect(id types.NodeID, epoch uint64) ([]change.Change, error) {
-	}
+	var stale bool
-	v, _ := s.nodeLocks.LoadOrStore(id, &nodeLock{})
-	return v.(*nodeLock)
-}
-
-// Disconnect marks a node as disconnected and updates its primary routes in the state.
-// The gen parameter is the generation returned by Connect(). If a newer Connect() has
-// been called since the session that is disconnecting, the generation will not match
-// and this call becomes a no-op, preventing stale disconnects from overwriting the
-// online status set by a newer session.
-//
-// The per-node lock is acquired before the gen check so that the check
-// and the subsequent NodeStore + primaryRoutes mutations are atomic
-// with respect to a concurrent Connect on the same node. A stale
-// Disconnect that observed an out-of-date gen must not be allowed to
-// proceed even partially — see issue #3203.
-func (s *State) Disconnect(id types.NodeID, gen uint64) ([]change.Change, error) {
-	nl := s.nodeLockFor(id)
-	nl.mu.Lock()
-	defer nl.mu.Unlock()
-
-	// Check if this disconnect is stale. A newer Connect() will have incremented
-	// the generation, so if ours doesn't match, a newer session owns this node.
-	// The check happens UNDER the lock so a concurrent Connect cannot bump gen
-	// between this read and the mutations below.
-	if current := nl.gen.Load(); current != gen {
-		log.Debug().
-			Uint64("disconnect_gen", gen).
-			Uint64("current_gen", current).
-			Msg("stale disconnect rejected, newer session active")
-
-		return nil, nil
-	}
-
 	node, ok := s.nodeStore.UpdateNode(id, func(n *types.Node) {
+		if n.SessionEpoch != epoch {
+			stale = true
+			return
+		}
 		now := time.Now()
 		n.LastSeen = &now
 		n.IsOnline = new(false)
@@ -682,6 +605,15 @@ func (s *State) Disconnect(id types.NodeID, gen uint64) ([]change.Change, error)
 		return nil, fmt.Errorf("%w: %d", ErrNodeNotFound, id)
 	}
 
+	if stale {
+		log.Debug().
+			Uint64("disconnect_epoch", epoch).
+			Uint64("current_epoch", node.SessionEpoch()).
+			Msg("stale disconnect rejected, newer session active")
+
+		return nil, nil
+	}
+
 	log.Info().EmbedObject(node).Msg("node disconnected")
 
 	// Persist LastSeen best-effort: NodeStore already reflects offline
@@ -693,16 +625,9 @@ func (s *State) Disconnect(id types.NodeID, gen uint64) ([]change.Change, error)
 		c = change.Change{}
 	}
 
-	// The node is disconnecting so make sure that none of the routes it
+	// Going offline can affect policy compilation beyond subnet routes
-	// announced are served to any nodes.
+	// (cap/relay grants, tag/group aliases, via routes), so peers need
-	s.primaryRoutes.SetRoutes(id)
+	// a fresh netmap regardless of whether the primary moved.
-
-	// A node going offline can affect policy compilation in ways beyond
-	// subnet routes: cap/relay grants targeting this node, identity-based
-	// aliases (tags, groups, users) that reference its tags/IPs, via
-	// routes steered through it, and so on. Always trigger a PolicyChange
-	// so peers receive a recomputed netmap and drop any cached state
-	// derived from this node (including peer relay allocations).
 	//
 	// TODO(kradalby): fires one full netmap recompute per peer on
 	// every connect/disconnect. Coalesce in mapper/batcher.go:addToBatch.
@@ -970,6 +895,8 @@ func (s *State) SetApprovedRoutes(nodeID types.NodeID, routes []netip.Prefix) (t
 	// TODO(kradalby): In principle we should call the AutoApprove logic here
 	// because even if the CLI removes an auto-approved route, it will be added
 	// back automatically.
+	prevRoutes := s.nodeStore.PrimaryRoutes()
+
 	n, ok := s.nodeStore.UpdateNode(nodeID, func(node *types.Node) {
 		node.ApprovedRoutes = routes
 	})
@@ -984,13 +911,9 @@ func (s *State) SetApprovedRoutes(nodeID types.NodeID, routes []netip.Prefix) (t
 		return types.NodeView{}, change.Change{}, err
 	}
 
-	// Update primary routes table based on SubnetRoutes (intersection of announced and approved).
+	// PolicyChange fans out a fresh netmap whenever the new approved
-	// The primary routes table is what the mapper uses to generate network maps, so updating it
+	// set shifted a primary advertiser.
-	// here ensures that route changes are distributed to peers.
+	routeChange := !maps.Equal(prevRoutes, s.nodeStore.PrimaryRoutes())
-	routeChange := s.primaryRoutes.SetRoutes(nodeID, nodeView.AllApprovedRoutes()...)
-
-	// If routes changed or the changeset isn't already a full update, trigger a policy change
-	// to ensure all nodes get updated network maps
 	if routeChange || !c.IsFull() {
 		c = change.PolicyChange()
 	}
@@ -1176,20 +1099,9 @@ func (s *State) SetPolicyInDB(data string) (*types.Policy, error) {
 	return s.db.SetPolicy(data)
 }
 
-// SetNodeRoutes sets the primary routes for a node.
-func (s *State) SetNodeRoutes(nodeID types.NodeID, routes ...netip.Prefix) change.Change {
-	if s.primaryRoutes.SetRoutes(nodeID, routes...) {
-		// Route changes affect packet filters for all nodes, so trigger a policy change
-		// to ensure filters are regenerated across the entire network
-		return change.PolicyChange()
-	}
-
-	return change.Change{}
-}
-
 // GetNodePrimaryRoutes returns the primary routes for a node.
 func (s *State) GetNodePrimaryRoutes(nodeID types.NodeID) []netip.Prefix {
-	return s.primaryRoutes.PrimaryRoutes(nodeID)
+	return s.nodeStore.PrimaryRoutesForNode(nodeID)
 }
 
 // RoutesForPeer computes the routes a peer should advertise in a viewer's
@@ -1204,7 +1116,7 @@ func (s *State) RoutesForPeer(
 	matchers []matcher.Match,
 ) []netip.Prefix {
 	viaResult := s.polMan.ViaRoutesForPeer(viewer, peer)
-	globalPrimaries := s.primaryRoutes.PrimaryRoutes(peer.ID())
+	globalPrimaries := s.nodeStore.PrimaryRoutesForNode(peer.ID())
 	exitRoutes := peer.ExitRoutes()
 
 	var reduced []netip.Prefix
@@ -1264,14 +1176,47 @@ func (s *State) RoutesForPeer(
 	return reduced
 }
 
-// PrimaryRoutes returns the primary routes tracker.
+// PrimaryRoutesString renders the current prefix→primary assignment
-func (s *State) PrimaryRoutes() *routes.PrimaryRoutes {
+// for diagnostics.
-	return s.primaryRoutes
+func (s *State) PrimaryRoutesString() string {
+	return s.nodeStore.PrimaryRoutesString()
 }
 
-// PrimaryRoutesString returns a string representation of all primary routes.
+// IsNodeHealthy reports the HA prober's view of id. Unknown nodes
-func (s *State) PrimaryRoutesString() string {
+// report healthy.
-	return s.primaryRoutes.String()
+func (s *State) IsNodeHealthy(id types.NodeID) bool {
+	return s.nodeStore.IsNodeHealthy(id)
+}
+
+// SetNodeUnhealthy flips the runtime Unhealthy bit and reports whether
+// the resulting primary route assignment changed, so the HA prober can
+// decide whether to fan out a PolicyChange.
+//
+// A request to mark a node Unhealthy is dropped if the node is no
+// longer an HA candidate (offline or no approved routes). The prober
+// reads HANodes() at the start of a probe cycle and writes back after
+// the timeout fires; in that window the node may have left the
+// candidate set, and the bit would just be stale. The check happens
+// inside the writer goroutine so it serialises against the
+// SetApprovedRoutes / Disconnect that removed candidacy.
+func (s *State) SetNodeUnhealthy(id types.NodeID, unhealthy bool) bool {
+	prevRoutes := s.nodeStore.PrimaryRoutes()
+
+	_, ok := s.nodeStore.UpdateNode(id, func(n *types.Node) {
+		if unhealthy {
+			online := n.IsOnline != nil && *n.IsOnline
+			if !online || len(n.AllApprovedRoutes()) == 0 {
+				return
+			}
+		}
+
+		n.Unhealthy = unhealthy
+	})
+	if !ok {
+		return false
+	}
+
+	return !maps.Equal(prevRoutes, s.nodeStore.PrimaryRoutes())
 }
 
 // ValidateAPIKey checks if an API key is valid and active.
@@ -2495,6 +2440,10 @@ func (s *State) UpdateNodeFromMapRequest(id types.NodeID, req tailcfg.MapRequest
 		endpointChanged    bool
 		derpChanged        bool
 	)
+	// Snapshot the primary assignment so we can tell whether the
+	// Hostinfo + auto-approval that follows shifted any prefix.
+	prevRoutes := s.nodeStore.PrimaryRoutes()
+
 	// We need to ensure we update the node as it is in the NodeStore at
 	// the time of the request.
 	updatedNode, ok := s.nodeStore.UpdateNode(id, func(currentNode *types.Node) {
@@ -2626,10 +2575,22 @@ func (s *State) UpdateNodeFromMapRequest(id types.NodeID, req tailcfg.MapRequest
 		}
 	} // Continue with the rest of the processing using the updated node
 
-	// Handle route changes after NodeStore update.
+	// SubnetRoutes = announced ∩ approved, so a Hostinfo update can
-	// Update routes if announced routes changed (even if approved routes stayed the same)
+	// move a primary without ever touching ApprovedRoutes. The pre/post
-	// because SubnetRoutes is the intersection of announced AND approved routes.
+	// snapshot diff catches that.
-	nodeRouteChange := s.maybeUpdateNodeRoutes(id, updatedNode, hostinfoChanged, needsRouteApproval, routeChange, req.Hostinfo)
+	nodeRouteChange := change.Change{}
+
+	if !maps.Equal(prevRoutes, s.nodeStore.PrimaryRoutes()) {
+		log.Debug().
+			Caller().
+			Uint64(zf.NodeID, id.Uint64()).
+			Strs(zf.RoutesAnnounced, util.PrefixesToString(updatedNode.AnnouncedRoutes())).
+			Strs(zf.ApprovedRoutes, util.PrefixesToString(updatedNode.ApprovedRoutes().AsSlice())).
+			Strs(zf.AllApprovedRoutes, util.PrefixesToString(updatedNode.AllApprovedRoutes())).
+			Msg("primary route assignment shifted after MapRequest")
+
+		nodeRouteChange = change.PolicyChange()
+	}
 
 	_, policyChange, err := s.persistNodeToDB(updatedNode)
 	if err != nil {
@@ -2723,34 +2684,3 @@ func peerChangeEmpty(peerChange tailcfg.PeerChange) bool {
 		peerChange.LastSeen == nil &&
 		peerChange.KeyExpiry == nil
 }
-
-// maybeUpdateNodeRoutes updates node routes if announced routes changed but approved routes didn't.
-// This is needed because SubnetRoutes is the intersection of announced AND approved routes.
-func (s *State) maybeUpdateNodeRoutes(
-	id types.NodeID,
-	node types.NodeView,
-	hostinfoChanged, needsRouteApproval, routeChange bool,
-	hostinfo *tailcfg.Hostinfo,
-) change.Change {
-	// Only update if announced routes changed without approval change
-	if !hostinfoChanged || !needsRouteApproval || routeChange || hostinfo == nil {
-		return change.Change{}
-	}
-
-	log.Debug().
-		Caller().
-		Uint64(zf.NodeID, id.Uint64()).
-		Msg("updating routes because announced routes changed but approved routes did not")
-
-	// SetNodeRoutes sets the active/distributed routes using AllApprovedRoutes()
-	// which returns only the intersection of announced AND approved routes.
-	log.Debug().
-		Caller().
-		Uint64(zf.NodeID, id.Uint64()).
-		Strs(zf.RoutesAnnounced, util.PrefixesToString(node.AnnouncedRoutes())).
-		Strs(zf.ApprovedRoutes, util.PrefixesToString(node.ApprovedRoutes().AsSlice())).
-		Strs(zf.AllApprovedRoutes, util.PrefixesToString(node.AllApprovedRoutes())).
-		Msg("updating node routes for distribution")
-
-	return s.SetNodeRoutes(id, node.AllApprovedRoutes()...)
-}