mkdocs: bump version

Signed-off-by: Kristoffer Dalby <kristoffer@tailscale.com>
(cherry picked from commit 565fd254d0)

.github/workflows/test-integration.yaml

@@ -340,6 +340,8 @@ jobs:
           - TestTagsAuthKeyWithTagCannotAddViaCLI
           - TestTagsAuthKeyWithTagCannotChangeViaCLI
           - TestTagsAuthKeyWithTagAdminOverrideReauthPreserves
+          - TestTagsReauthDifferentKeyRetagsNode
+          - TestTagsReauthDifferentKeyRemovesTag
           - TestTagsAuthKeyWithTagCLICannotModifyAdminTags
           - TestTagsAuthKeyWithoutTagCannotRequestTags
           - TestTagsAuthKeyWithoutTagRegisterNoTags
@@ -367,6 +369,12 @@ jobs:
           - TestTagsAuthKeyWithoutUserInheritsTags
           - TestTagsAuthKeyWithoutUserRejectsAdvertisedTags
           - TestTagsAuthKeyConvertToUserViaCLIRegister
+          - TestTaggedNodeLogoutReloginSingleUseKeyOnline
+          - TestTaggedNodeLogoutReloginReusableKeyOnline
+          - TestTagsOIDCReauthAddOwnedTag
+          - TestTagsReauthEmptyTagsReturnsToUserSurvives
+          - TestTS2021WebSocketGET
+          - TestTS2021WASMClientUnderNode
           - TestTailscaleRustAxum
     uses: ./.github/workflows/integration-test-template.yml
     secrets: inherit

CHANGELOG.md

@@ -1,8 +1,36 @@
 # CHANGELOG
 
-## 0.30.0 (202x-xx-xx)
+## 0.29.3 (2026-07-29)
 
-**Minimum supported Tailscale client version: v1.xx.0**
+**Minimum supported Tailscale client version: v1.80.0**
+
+### Changes
+
+- Fix tagged node stuck expired after `tailscale logout`, unable to re-authenticate [#3394](https://github.com/juanfont/headscale/pull/3394)
+- Re-registering a tagged node with a different pre-auth key now applies the new key's tags instead of silently keeping the old ones [#3394](https://github.com/juanfont/headscale/pull/3394)
+- Fix re-authenticating an already-tagged node with `--advertise-tags` being rejected when the authenticating user owns the tags [#3394](https://github.com/juanfont/headscale/pull/3394)
+- Fix ephemeral nodes lingering as disconnected after reconnect churn [#3383](https://github.com/juanfont/headscale/pull/3383)
+- Fix node registration falsely returning `401 registration timed out` when auth completes as the request context expires [#3392](https://github.com/juanfont/headscale/pull/3392)
+- Check the machine key on the followup registration poll so a leaked auth ID cannot return the registering user's identity [#3393](https://github.com/juanfont/headscale/pull/3393)
+- Reject `/key` requests below the supported capability version floor, matching `/ts2021` [#3391](https://github.com/juanfont/headscale/pull/3391)
+
+## 0.29.2 (2026-07-01)
+
+**Minimum supported Tailscale client version: v1.80.0**
+
+### Changes
+
+- Fix map generation serializing on the policy lock, so a mass reconnect on `autogroup:self`, via or relay policies no longer stalls clients into `unexpected EOF` retry loops [#3358](https://github.com/juanfont/headscale/pull/3358)
+- Fix `/ts2021` rejecting the WebSocket `GET` upgrade with 405, which prevented Tailscale JS/WASM control clients from connecting [#3359](https://github.com/juanfont/headscale/pull/3359)
+- Gracefully handle nodes with an invalid FQDN (empty or too long) instead of failing map delivery; offending names are logged at startup with the fix command [#3349](https://github.com/juanfont/headscale/pull/3349)
+
+## 0.29.1 (2026-06-18)
+
+**Minimum supported Tailscale client version: v1.80.0**
+
+### Changes
+
+- Fix nodes with `tags='null'` losing their assigned user on upgrade [#3325](https://github.com/juanfont/headscale/pull/3325)
 
 ## 0.29.0 (2026-06-17)
 

Dockerfile.derper

@@ -1,6 +1,6 @@
 # For testing purposes only
 
-FROM golang:1.26.4-alpine AS build-env
+FROM golang:1.26.5-alpine AS build-env
 
 WORKDIR /go/src
 

Dockerfile.integration

@@ -2,7 +2,7 @@
 # and are in no way endorsed by Headscale's maintainers as an
 # official nor supported release or distribution.
 
-FROM docker.io/golang:1.26.4-trixie AS builder
+FROM docker.io/golang:1.26.5-trixie AS builder
 ARG VERSION=dev
 ENV GOPATH /go
 WORKDIR /go/src/headscale

Dockerfile.tailscale-HEAD

@@ -4,7 +4,7 @@
 # This Dockerfile is more or less lifted from tailscale/tailscale
 # to ensure a similar build process when testing the HEAD of tailscale.
 
-FROM golang:1.26.4-alpine AS build-env
+FROM golang:1.26.5-alpine AS build-env
 
 WORKDIR /go/src
 

Dockerfile.wasmclient

@@ -0,0 +1,30 @@
+# For integration testing only.
+#
+# Builds the Tailscale control client (integration/wasmic/wasmclient) for
+# GOOS=js/GOARCH=wasm and packages it with Go's wasm_exec Node runner. The
+# container idles; the integration test execs
+#   node /app/wasm_exec_node.js /app/client.wasm <control-url>
+# to drive a real browser-style WebSocket GET against headscale's /ts2021,
+# guarding the regression in issue #3357.
+
+FROM golang:1.26.5-alpine AS build
+
+WORKDIR /src
+
+# Only the module metadata and the wasm client package are needed to build the
+# js/wasm binary; its imports (tailscale.com/control/controlhttp, ...) resolve
+# from the module proxy.
+COPY go.mod go.sum ./
+COPY integration/wasmic/wasmclient ./integration/wasmic/wasmclient
+
+RUN GOOS=js GOARCH=wasm go build -o /out/client.wasm ./integration/wasmic/wasmclient \
+    && cp "$(go env GOROOT)/lib/wasm/wasm_exec.js" /out/wasm_exec.js \
+    && cp "$(go env GOROOT)/lib/wasm/wasm_exec_node.js" /out/wasm_exec_node.js
+
+FROM node:24-alpine
+
+WORKDIR /app
+COPY --from=build /out/ /app/
+
+# Idle; the test execs the client on demand with the headscale control URL.
+ENTRYPOINT ["tail", "-f", "/dev/null"]

docs/ref/policy.md

@@ -158,17 +158,17 @@ devices. Can only be used in policy destinations.
         {
           "src": ["boss@"],
           "dst": ["boss@"],
-          "ip": "*"
+          "ip": ["*"]
         },
         {
           "src": ["dev1@"],
           "dst": ["dev1@"],
-          "ip": "*"
+          "ip": ["*"]
         },
         {
           "src": ["intern1@"],
           "dst": ["intern1@"],
-          "ip": "*"
+          "ip": ["*"]
         }
       ]
     }

flake.lock

@@ -1,5 +1,27 @@
 {
   "nodes": {
+    "flake-checks": {
+      "inputs": {
+        "flake-utils": "flake-utils",
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "treefmt-nix": "treefmt-nix"
+      },
+      "locked": {
+        "lastModified": 1783015821,
+        "narHash": "sha256-vmzYTZxIAy3OYXwkRL5GQYk8fNnh8TLWcxQRIO3ywOU=",
+        "owner": "kradalby",
+        "repo": "flake-checks",
+        "rev": "3c821706eb0bd07f515f9a7726b650ccae927433",
+        "type": "github"
+      },
+      "original": {
+        "owner": "kradalby",
+        "repo": "flake-checks",
+        "type": "github"
+      }
+    },
     "flake-utils": {
       "inputs": {
         "systems": "systems"
@@ -20,11 +42,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1781153106,
-        "narHash": "sha256-yzsroLCcuRG4KdGMxWt0eXKOrRSgQT8/xjYngeq9ujU=",
+        "lastModified": 1784421286,
+        "narHash": "sha256-YKn3t7FvaargtyNHbrOw+qzDM+93q5kSCREUPNxeZQs=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9ee75f111a06d7ab2b2f729698a8eff53d54e070",
+        "rev": "62eaae30cd05c04718500ddb44e7ef85cd45ddd5",
         "type": "github"
       },
       "original": {

flake.nix

@@ -2,11 +2,12 @@
   description = "headscale - Open Source Tailscale Control server";
 
   inputs = {
-    # Pinned to staging-next-26.05 for Go 1.26.4 (security fix GO-2026-5037/5039):
-    # nixpkgs-unstable still ships 1.26.3 — the bump is merged to nixpkgs staging
-    # but the large-rebuild staging->unstable pipeline lags. The 26.05 line is
-    # otherwise current (dev tools match unstable). Switch back to nixpkgs-unstable
-    # once it ships go_1_26 >= 1.26.4.
+    # Pinned to staging-next-26.05 for Go 1.26.5: the Tailscale HEAD build
+    # (Dockerfile.tailscale-HEAD) requires go >= 1.26.5, and nixpkgs-unstable
+    # still ships 1.26.4 — the bump is merged to nixpkgs staging but the
+    # large-rebuild staging->unstable pipeline lags. The 26.05 line is otherwise
+    # current (dev tools match unstable). Switch back to nixpkgs-unstable once it
+    # ships go_1_26 >= 1.26.5.
     nixpkgs.url = "github:NixOS/nixpkgs/staging-next-26.05";
     flake-utils.url = "github:numtide/flake-utils";
   };
@@ -31,7 +32,7 @@
       overlays.default = _: prev:
         let
           pkgs = nixpkgs.legacyPackages.${prev.stdenv.hostPlatform.system};
-          # Go 1.26 builder; resolves to Go 1.26.4 from the pinned nixpkgs.
+          # Go 1.26 builder; resolves to Go 1.26.5 from the pinned nixpkgs.
           buildGo = pkgs.buildGo126Module;
           vendorHash = (builtins.fromJSON (builtins.readFile ./flakehashes.json)).vendor.sri;
         in
@@ -101,7 +102,7 @@
           };
 
           # Build golangci-lint with stock Go 1.26 (upstream uses hardcoded Go
-          # version); it does not build against the pinned 1.26.4.
+          # version); it does not build against the pinned 1.26.5.
           golangci-lint = buildGo rec {
             pname = "golangci-lint";
             version = "2.12.2";

flakehashes.json

@@ -1,6 +1,6 @@
 {
   "vendor": {
-    "goModSum": "sha256-csVm5v6HZ49PBp/FCX+yK1sjV8/nuUQz3GKN21Ne1mg=",
+    "goModSum": "sha256-GFiMC2ktrQnpSVzJHabVP1TYwJvn5+a/BDLhvmGTUWc=",
     "sri": "sha256-fzKyXNMw/2yAEhaTZu0n1NXatPO2IP0HFA2ey1vZIYM="
   }
 }

go.mod

@@ -1,6 +1,6 @@
 module github.com/juanfont/headscale
 
-go 1.26.4
+go 1.26.5
 
 require (
 	github.com/arl/statsviz v0.8.0

hscontrol/app.go

@@ -544,6 +544,11 @@ func (h *Headscale) createRouter(grpcMux *grpcRuntime.ServeMux) *chi.Mux {
 	r.Use(middleware.Recoverer)
 	r.Use(securityHeaders)
 
+	// TS2021 accepts both the native client's HTTP POST upgrade and the
+	// browser/WASM client's WebSocket GET upgrade; NoiseUpgradeHandler
+	// dispatches on the Upgrade header, not the method. Registering GET as
+	// well keeps the router from rejecting the WebSocket handshake with 405.
+	r.Get(ts2021UpgradePath, h.NoiseUpgradeHandler)
 	r.Post(ts2021UpgradePath, h.NoiseUpgradeHandler)
 
 	r.Get("/robots.txt", h.RobotsHandler)

hscontrol/auth.go

@@ -23,6 +23,18 @@ type AuthProvider interface {
 	AuthURL(authID types.AuthID) string
 }
 
+// machineKeyMismatch fails closed when a node looked up by NodeKey was started
+// in a Noise session with a different machine key. Without this anyone holding a
+// target's NodeKey could open a session with a throwaway machine key and act on
+// the owner's node. Returns a 401 [HTTPError] on mismatch, nil otherwise.
+func machineKeyMismatch(node types.NodeView, machineKey key.MachinePublic) error {
+	if node.MachineKey() != machineKey {
+		return NewHTTPError(http.StatusUnauthorized, "node exists with a different machine key", nil)
+	}
+
+	return nil
+}
+
 func (h *Headscale) handleRegister(
 	ctx context.Context,
 	req tailcfg.RegisterRequest,
@@ -223,6 +235,20 @@ func (h *Headscale) handleLogout(
 			Msg("Node is not ephemeral, setting expiry instead of deleting")
 	}
 
+	// Tagged nodes have key expiry permanently disabled (they are owned by
+	// their tags, not a user, and never expire - KB 1068). Logging one out has
+	// no expiry semantics, so do not stamp an expiry on it: doing so leaves the
+	// node IsExpired() forever and it can never re-authenticate (#3371). The
+	// admin path `headscale nodes expire` remains free to set a deliberate
+	// expiry via SetNodeExpiry; only the logout path is guarded here.
+	if node.IsTagged() {
+		log.Debug().
+			EmbedObject(node).
+			Msg("Tagged node logout: not stamping expiry (tagged nodes never expire)")
+
+		return nodeToRegisterResponse(node), nil
+	}
+
 	// Update the internal state with the nodes new expiry, meaning it is
 	// logged out.
 	//
@@ -291,19 +317,41 @@ func (h *Headscale) waitForFollowup(
 	}
 
 	if reg, ok := h.state.GetAuthCacheEntry(followupReg); ok {
+		var verdict types.AuthVerdict
 		select {
-		case <-ctx.Done():
-			return nil, NewHTTPError(http.StatusUnauthorized, "registration timed out", err)
-		case verdict := <-reg.WaitForAuth():
-			if verdict.Accept() {
-				if !verdict.Node.Valid() {
-					// registration is expired in the cache, instruct the client to try a new registration
-					return h.reqToNewRegisterResponse(req, machineKey)
-				}
-
-				return nodeToRegisterResponse(verdict.Node), nil
+		// Prefer a completed registration even if the context has also
+		// expired. When both are ready, a plain select picks at random and
+		// would discard a successful registration as a spurious timeout
+		// (issue #3385).
+		case verdict = <-reg.WaitForAuth():
+		default:
+			select {
+			case <-ctx.Done():
+				return nil, NewHTTPError(http.StatusUnauthorized, "registration timed out", ctx.Err())
+			case verdict = <-reg.WaitForAuth():
 			}
 		}
+
+		if verdict.Accept() {
+			if !verdict.Node.Valid() {
+				// registration is expired in the cache, instruct the client to try a new registration
+				return h.reqToNewRegisterResponse(req, machineKey)
+			}
+
+			// The followup poll is only authenticated by the auth ID in the
+			// URL, so fail closed unless the Noise session asking for the
+			// result was started with the same machine key that opened the
+			// registration. [State.HandleNodeFromAuthPath] resolves the node
+			// from the cached [types.RegistrationData.MachineKey], so the two
+			// match on the normal path. [Headscale.handleRegister] and
+			// [Headscale.handleLogout] apply the same check.
+			err := machineKeyMismatch(verdict.Node, machineKey)
+			if err != nil {
+				return nil, err
+			}
+
+			return nodeToRegisterResponse(verdict.Node), nil
+		}
 	}
 
 	// if the follow-up registration isn't found anymore, instruct the client to try a new registration

hscontrol/auth_tags_test.go

@@ -1338,3 +1338,392 @@ func TestReregistrationZeroExpiryStaysNil(t *testing.T) {
 	assert.False(t, node2.Expiry().Valid(),
 		"re-registration with zero client expiry and no default should leave expiry nil, not pointer to zero time")
 }
+
+// tsLogoutSentinelExpiry is the past expiry a real tailscale client sends on
+// `tailscale logout`: time.Unix(123, 0) (controlclient/direct.go). The issue
+// #3371 trace shows it verbatim as `expiry=123`. Using it here (rather than a
+// generic time.Now().Add(-1h)) keeps the reproduction faithful to the wire
+// behaviour: handleRegister must classify this as a logout, and handleLogout
+// must clamp it to now.
+func tsLogoutSentinelExpiry() time.Time {
+	return time.Unix(123, 0)
+}
+
+// TestIssue3371_TaggedNodeLogoutReloginSingleUseKey reproduces
+// https://github.com/juanfont/headscale/issues/3371 through the real
+// register/logout HTTP-handler path (handleRegister -> handleLogout ->
+// handleRegister), not by poking SetNodeExpiry directly.
+//
+// Root cause (a): `tailscale logout` sends a past expiry; handleLogout stamps
+// it on the node via SetNodeExpiry with no IsTagged guard, so a tagged node —
+// which must have key-expiry disabled — becomes Expired.
+//
+// Root cause (b): on the next `tailscale up --auth-key <fresh key>`,
+// HandleNodeFromPreAuthKey sees an expired node, takes the expired-node
+// validation path, consumes the fresh single-use key on the in-place
+// re-registration, yet leaves the node expired (the expiry-refresh block is
+// gated `!node.IsTagged()`). The response carries NodeKeyExpired=true, so the
+// client rotates its node key and retries with the now-spent key, which is
+// rejected with "authkey already used" forever.
+//
+// Faithful to the artifacts: the client rotates its NodeKey on relogin, the
+// logout carries BOTH a past expiry AND an auth key, and a BRAND NEW key is
+// presented for the relogin (the trace shows tag:tag2 keys burned while the
+// node kept tag:tag1).
+func TestIssue3371_TaggedNodeLogoutReloginSingleUseKey(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("tag-logout-user")
+	tags := []string{"tag:tag1"}
+
+	// `headscale preauthkeys create --tags tag:tag1` (single-use).
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), false, false, nil, tags)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	// `tailscale up --auth-key $KEY1`: initial join.
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "headscale-debug"},
+	}
+
+	resp, err := app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+	require.True(t, resp.MachineAuthorized)
+	require.False(t, resp.NodeKeyExpired)
+
+	node, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.True(t, node.IsTagged(), "precondition: node is tagged")
+	require.False(t, node.Expiry().Valid(), "precondition: tagged node has expiry disabled")
+
+	// `tailscale logout`: client sends a past-expiry register with the auth key
+	// still attached (handleRegister must treat past expiry as logout regardless
+	// of Auth). Reuse the same node key: logout does not rotate it.
+	logoutReq := tailcfg.RegisterRequest{
+		Auth:    &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey: nodeKey.Public(),
+		Expiry:  tsLogoutSentinelExpiry(),
+	}
+
+	_, err = app.handleRegister(context.Background(), logoutReq, machineKey.Public())
+	require.NoError(t, err)
+
+	// A tagged node must NOT be expired by logout — tagged nodes never expire.
+	// This is root cause (a); it fails before the fix.
+	nodeAfterLogout, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	assert.False(t, nodeAfterLogout.IsExpired(),
+		"issue #3371 root cause (a): logout must not expire a tagged node")
+	assert.False(t, nodeAfterLogout.Expiry().Valid(),
+		"issue #3371 root cause (a): tagged node must keep key-expiry disabled after logout")
+
+	// `tailscale up --auth-key $KEY2`: a BRAND NEW single-use key, and the client
+	// rotates its node key (as the real client does on relogin).
+	pak2, err := app.state.CreatePreAuthKey(user.TypedID(), false, false, nil, tags)
+	require.NoError(t, err)
+
+	nodeKey2 := key.NewNode()
+	reloginReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak2.Key},
+		NodeKey:  nodeKey2.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "headscale-debug"},
+	}
+
+	reloginResp, err := app.handleRegister(context.Background(), reloginReq, machineKey.Public())
+	require.NoError(t, err,
+		"issue #3371: a fresh valid key must re-authenticate the tagged node after logout")
+	require.NotNil(t, reloginResp)
+
+	// The whole point: the node comes back online, not stuck expired.
+	assert.False(t, reloginResp.NodeKeyExpired,
+		"issue #3371: relogin response must not report the node key as expired")
+	assert.True(t, reloginResp.MachineAuthorized)
+
+	relogged, found := app.state.GetNodeByNodeKey(nodeKey2.Public())
+	require.True(t, found)
+	assert.True(t, relogged.IsTagged(), "node stays tagged after relogin")
+	assert.False(t, relogged.IsExpired(),
+		"issue #3371: tagged node must be online (not expired) after relogin")
+	assert.False(t, relogged.Expiry().Valid(),
+		"issue #3371: tagged node must have key-expiry disabled after relogin")
+	assert.Equal(t, node.ID(), relogged.ID(), "must re-use the same node, not duplicate")
+	assert.Equal(t, 1, app.state.ListNodes().Len(), "machine maps to exactly one node")
+}
+
+// TestIssue3371_TaggedNodeLogoutReloginReusableKey is the reusable-key variant
+// from the issue ("tailscale up then hangs indefinitely instead of erroring").
+// With a reusable key the relogin does not hit "authkey already used", but the
+// node still stays expired without the fix — so the client never observes a
+// non-expired node and hangs. The observable failure here is the persisted
+// expired state after relogin.
+func TestIssue3371_TaggedNodeLogoutReloginReusableKey(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("tag-logout-reusable")
+	tags := []string{"tag:tag1"}
+
+	// `headscale preauthkeys create --reusable --tags tag:tag1`.
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), true, false, nil, tags)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "reusable-tagged"},
+	}
+
+	_, err = app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+
+	node, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.True(t, node.IsTagged())
+	require.False(t, node.Expiry().Valid())
+
+	// Logout.
+	logoutReq := tailcfg.RegisterRequest{
+		Auth:    &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey: nodeKey.Public(),
+		Expiry:  tsLogoutSentinelExpiry(),
+	}
+	_, err = app.handleRegister(context.Background(), logoutReq, machineKey.Public())
+	require.NoError(t, err)
+
+	// Relogin with the same reusable key, rotating the node key.
+	nodeKey2 := key.NewNode()
+	reloginReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey2.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "reusable-tagged"},
+	}
+	reloginResp, err := app.handleRegister(context.Background(), reloginReq, machineKey.Public())
+	require.NoError(t, err)
+	require.NotNil(t, reloginResp)
+
+	assert.False(t, reloginResp.NodeKeyExpired,
+		"issue #3371: reusable-key relogin must not report node key expired")
+
+	relogged, found := app.state.GetNodeByNodeKey(nodeKey2.Public())
+	require.True(t, found)
+	assert.True(t, relogged.IsTagged())
+	assert.False(t, relogged.IsExpired(),
+		"issue #3371: tagged node must be online (not expired) after reusable-key relogin")
+	assert.False(t, relogged.Expiry().Valid(),
+		"issue #3371: tagged node must have key-expiry disabled after reusable-key relogin")
+}
+
+// TestIssue3371_TaggedNodeLogoutDoesNotSetExpiry pins the deepest root cause
+// (a) in isolation: `tailscale logout` on a tagged node must not stamp an
+// expiry at all. This is the assertion PR #3372 does not make — it leaves
+// handleLogout expiring tagged nodes and only unwinds the damage on the next
+// registration. Keeping this separate from the relogin tests means a
+// regression that re-introduces logout-sets-expiry is caught even if the
+// re-registration cleanup masks it.
+func TestIssue3371_TaggedNodeLogoutDoesNotSetExpiry(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("tag-logout-noexpiry")
+	tags := []string{"tag:tag1"}
+
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), true, false, nil, tags)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "noexpiry-tagged"},
+	}
+	_, err = app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+
+	logoutReq := tailcfg.RegisterRequest{
+		Auth:    &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey: nodeKey.Public(),
+		Expiry:  tsLogoutSentinelExpiry(),
+	}
+	_, err = app.handleRegister(context.Background(), logoutReq, machineKey.Public())
+	require.NoError(t, err)
+
+	nodeAfterLogout, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	assert.True(t, nodeAfterLogout.IsTagged(), "node stays tagged through logout")
+	assert.False(t, nodeAfterLogout.Expiry().Valid(),
+		"issue #3371 root cause (a): logout must not set an expiry on a tagged node")
+	assert.False(t, nodeAfterLogout.IsExpired(),
+		"issue #3371 root cause (a): tagged node must not be expired by logout")
+
+	// The database column must be NULL, not a clamped 'now' timestamp — a
+	// persisted expiry survives restart and re-triggers the lockout.
+	var dbNode types.Node
+	require.NoError(t,
+		app.state.DB().DB.First(&dbNode, nodeAfterLogout.ID().Uint64()).Error)
+	assert.Nil(t, dbNode.Expiry,
+		"issue #3371 root cause (a): tagged node's DB expiry must remain NULL after logout")
+}
+
+// TestIssue3371_UserOwnedNodeLogoutStillExpires is the guard rail: the fix for
+// tagged nodes must not change logout for ordinary user-owned nodes. A
+// user-owned node that logs out MUST still be expired (that is what logout
+// means for it).
+func TestIssue3371_UserOwnedNodeLogoutStillExpires(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("user-logout")
+
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), true, false, nil, nil)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "user-node"},
+		Expiry:   time.Now().Add(24 * time.Hour),
+	}
+	_, err = app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+
+	node, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.False(t, node.IsTagged(), "precondition: user-owned node")
+
+	logoutReq := tailcfg.RegisterRequest{
+		Auth:    &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey: nodeKey.Public(),
+		Expiry:  tsLogoutSentinelExpiry(),
+	}
+	logoutResp, err := app.handleRegister(context.Background(), logoutReq, machineKey.Public())
+	require.NoError(t, err)
+	require.NotNil(t, logoutResp)
+
+	nodeAfterLogout, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	assert.True(t, nodeAfterLogout.IsExpired(),
+		"user-owned node must still be expired by logout (fix must not regress this)")
+	assert.True(t, logoutResp.NodeKeyExpired,
+		"logout response for a user-owned node must report the key expired")
+}
+
+// TestIssue3371_TaggedNodeFutureExpirySurvivesRelogin is the discriminator
+// guard rail for the fix. A tagged node may carry a DELIBERATE future expiry
+// set by an admin (`headscale nodes expire`); TestTaggedNodeCanHaveKeyExpiry
+// establishes that is legal. The #3371 fix clears only a STALE PAST expiry (the
+// logout stamp) on re-registration — it must NOT wipe a future expiry. This
+// test locks that boundary: without care, a "tagged => clear expiry" fix would
+// silently destroy the admin's setting.
+//
+// Passes before the fix (re-registration currently never touches a tagged
+// node's expiry) and must keep passing after.
+func TestIssue3371_TaggedNodeFutureExpirySurvivesRelogin(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("tag-future-expiry")
+	tags := []string{"tag:tag1"}
+
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), true, false, nil, tags)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "future-expiry-tagged"},
+	}
+	_, err = app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+
+	node, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.True(t, node.IsTagged())
+
+	// Admin sets a deliberate future expiry (`headscale nodes expire`).
+	future := time.Now().Add(24 * time.Hour)
+	_, _, err = app.state.SetNodeExpiry(node.ID(), &future)
+	require.NoError(t, err)
+
+	withFuture, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.True(t, withFuture.Expiry().Valid(), "precondition: future expiry set")
+	require.False(t, withFuture.IsExpired(), "precondition: future expiry is not expired")
+
+	// Node re-registers (rotating its node key). The future expiry must survive.
+	nodeKey2 := key.NewNode()
+	reregReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey2.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "future-expiry-tagged"},
+	}
+	_, err = app.handleRegister(context.Background(), reregReq, machineKey.Public())
+	require.NoError(t, err)
+
+	after, found := app.state.GetNodeByNodeKey(nodeKey2.Public())
+	require.True(t, found)
+	require.True(t, after.IsTagged(), "node stays tagged")
+	assert.True(t, after.Expiry().Valid(),
+		"deliberate future expiry must survive re-registration (not cleared by #3371 fix)")
+	assert.WithinDuration(t, future, after.Expiry().Get(), 5*time.Second,
+		"the surviving expiry must be the admin-set future value, unchanged")
+}
+
+// TestIssue3371_EphemeralTaggedNodeLogoutDeletes is a regression guard for the
+// ephemeral+tagged combination. A tagged pre-auth key can also be ephemeral.
+// handleLogout deletes ephemeral nodes (before any expiry stamp), so the #3371
+// fix (which suppresses the expiry stamp for tagged nodes) must not divert an
+// ephemeral tagged node away from deletion.
+//
+// Passes before the fix and must keep passing after.
+func TestIssue3371_EphemeralTaggedNodeLogoutDeletes(t *testing.T) {
+	app := createTestApp(t)
+
+	user := app.state.CreateUserForTest("tag-ephemeral")
+	tags := []string{"tag:tag1"}
+
+	// Ephemeral + tagged key.
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), true, true, nil, tags)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	regReq := tailcfg.RegisterRequest{
+		Auth:     &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey:  nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{Hostname: "ephemeral-tagged"},
+	}
+	_, err = app.handleRegister(context.Background(), regReq, machineKey.Public())
+	require.NoError(t, err)
+
+	node, found := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, found)
+	require.True(t, node.IsTagged(), "precondition: node is tagged")
+	require.True(t, node.IsEphemeral(), "precondition: node is ephemeral")
+
+	// Logout: an ephemeral node is deleted, not expired.
+	logoutReq := tailcfg.RegisterRequest{
+		Auth:    &tailcfg.RegisterResponseAuth{AuthKey: pak.Key},
+		NodeKey: nodeKey.Public(),
+		Expiry:  tsLogoutSentinelExpiry(),
+	}
+	_, err = app.handleRegister(context.Background(), logoutReq, machineKey.Public())
+	require.NoError(t, err)
+
+	require.EventuallyWithT(t, func(c *assert.CollectT) {
+		_, stillThere := app.state.GetNodeByNodeKey(nodeKey.Public())
+		assert.False(c, stillThere,
+			"ephemeral tagged node must be deleted on logout, not expired")
+	}, 2*time.Second, 50*time.Millisecond, "waiting for ephemeral node deletion")
+}

hscontrol/auth_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"net/http"
 	"net/url"
 	"strings"
 	"testing"
@@ -692,6 +693,11 @@ func TestAuthenticationFlows(t *testing.T) {
 					user := app.state.CreateUserForTest("followup-user")
 
 					node := app.state.CreateNodeForTest(user, "followup-success-node")
+					// [State.HandleNodeFromAuthPath] resolves the node from the
+					// machine key cached when the registration was opened, so on
+					// the real path the node carries the polling session's
+					// machine key. CreateNodeForTest picks a random one.
+					node.MachineKey = machineKey1.Public()
 					nodeToRegister.FinishAuth(types.AuthVerdict{Node: node.View()})
 				}()
 
@@ -4101,3 +4107,150 @@ func TestHandleNodeFromAuthPath_OldUserNil_NoPanic(t *testing.T) {
 	assert.NotEqual(t, types.NodeID(99002), node.ID(), "new node, not orphan")
 	assert.Equal(t, userB.ID, node.UserID().Get(), "new node belongs to userB")
 }
+
+// TestWaitForFollowupMachineKeyMismatch covers the followup poll in
+// [Headscale.waitForFollowup]. That poll is authenticated only by the auth ID
+// embedded in the followup URL, so without a machine-key check anyone who
+// learns an ID gets the registering user's User/Login back in the
+// [tailcfg.RegisterResponse].
+//
+// [Headscale.handleRegister] and [Headscale.handleLogout] already fail closed
+// here; see the "existing_node_machine_key_mismatch" case in
+// [TestAuthenticationFlows] for the equivalent assertion on that path.
+//
+// The nodes are given the registering session's machine key because that is
+// what production produces: [State.HandleNodeFromAuthPath] resolves the node
+// from the machine key cached in [types.RegistrationData] when the
+// registration was opened.
+func TestWaitForFollowupMachineKeyMismatch(t *testing.T) {
+	app := createTestApp(t)
+
+	victimMachineKey := key.NewMachine()
+	attackerMachineKey := key.NewMachine()
+
+	// Park a completed registration in the auth cache, as a node that is
+	// already polling for its verdict would see it.
+	newPendingFollowup := func(hostname string) string {
+		authID := types.MustAuthID()
+		regEntry := types.NewRegisterAuthRequest(&types.RegistrationData{
+			MachineKey: victimMachineKey.Public(),
+			NodeKey:    key.NewNode().Public(),
+			Hostname:   hostname,
+		})
+		app.state.SetAuthCacheEntry(authID, regEntry)
+
+		user := app.state.CreateUserForTest(hostname + "-user")
+		node := app.state.CreateNodeForTest(user, hostname)
+		node.MachineKey = victimMachineKey.Public()
+		// CreateNodeForTest only sets UserID, but nodeToRegisterResponse reads
+		// the owner, and the owner's identity is exactly what must not leak.
+		node.User = user
+		regEntry.FinishAuth(types.AuthVerdict{Node: node.View()})
+
+		return fmt.Sprintf("http://localhost:8080/register/%s", authID)
+	}
+
+	followup := func(url string, machineKey key.MachinePublic) (*tailcfg.RegisterResponse, error) {
+		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+		defer cancel()
+
+		return app.handleRegister(ctx, tailcfg.RegisterRequest{
+			Followup: url,
+			NodeKey:  key.NewNode().Public(),
+		}, machineKey)
+	}
+
+	t.Run("mismatched machine key is rejected", func(t *testing.T) {
+		resp, err := followup(newPendingFollowup("followup-mismatch"), attackerMachineKey.Public())
+
+		require.Error(t, err, "followup with a foreign machine key must not succeed")
+		assert.Nil(t, resp, "no registration details should be returned")
+
+		var httpErr HTTPError
+		require.ErrorAs(t, err, &httpErr)
+		assert.Equal(t, http.StatusUnauthorized, httpErr.Code)
+	})
+
+	// Positive control. Without it a regression that stops the poll from
+	// finding the cache entry at all would still pass the case above, because
+	// waitForFollowup falls back to handing out a fresh AuthURL.
+	t.Run("matching machine key still completes", func(t *testing.T) {
+		resp, err := followup(newPendingFollowup("followup-match"), victimMachineKey.Public())
+
+		require.NoError(t, err)
+		require.NotNil(t, resp)
+		assert.True(t, resp.MachineAuthorized)
+		assert.NotEmpty(t, resp.User.DisplayName, "the owner's identity is returned on the legitimate path")
+	})
+}
+
+// TestFollowupWaitPrefersCompletedAuthOverExpiredContext reproduces
+// https://github.com/juanfont/headscale/issues/3385.
+//
+// Root cause: [Headscale.waitForFollowup] selects on ctx.Done() and the auth
+// verdict channel with equal priority. When the registration has ALREADY
+// completed (verdict buffered) but the request context has ALSO expired, Go's
+// select picks a ready case at random, so roughly half the time it returns
+// "registration timed out" and discards a successful registration.
+//
+// The v0.28.0 hscontrol test suite hit this because the followup context
+// timeout was only 100ms while the setup goroutine (create user + node in
+// SQLite) frequently took longer on slower/constrained builders (ppc64le,
+// Alpine CI). Both channels ended up ready at once and the flake surfaced as
+// TestAuthenticationFlows/followup_registration_success failing with
+// "http error[401]: registration timed out".
+//
+// The fix must give the completed-auth case priority over context
+// cancellation. This test forces both cases ready on every iteration; it must
+// never report a timeout.
+func TestFollowupWaitPrefersCompletedAuthOverExpiredContext(t *testing.T) {
+	app := createTestApp(t)
+
+	machineKey := key.NewMachine().Public()
+	nodeKey := key.NewNode().Public()
+
+	const iterations = 300
+
+	timeouts, authorized := 0, 0
+
+	for i := range iterations {
+		regID, err := types.NewAuthID()
+		require.NoError(t, err)
+
+		authReq := types.NewRegisterAuthRequest(&types.RegistrationData{
+			Hostname: "followup-race-node",
+		})
+		app.state.SetAuthCacheEntry(regID, authReq)
+
+		// Registration completes BEFORE we wait: verdict is buffered.
+		user := app.state.CreateUserForTest(fmt.Sprintf("followup-race-user-%d", i))
+		node := app.state.CreateNodeForTest(user, "followup-race-node")
+		// waitForFollowup fails closed unless the node carries the polling
+		// session's machine key; production sets this via the cached
+		// RegistrationData. CreateNodeForTest picks a random one.
+		node.MachineKey = machineKey
+		authReq.FinishAuth(types.AuthVerdict{Node: node.View()})
+
+		// Context is expired BEFORE we wait: both select cases are ready.
+		ctx, cancel := context.WithCancel(context.Background())
+		cancel()
+
+		req := tailcfg.RegisterRequest{
+			Followup: fmt.Sprintf("http://localhost:8080/register/%s", regID),
+			NodeKey:  nodeKey,
+		}
+
+		resp, err := app.waitForFollowup(ctx, req, machineKey)
+		switch {
+		case err != nil:
+			timeouts++
+		case resp != nil && resp.MachineAuthorized:
+			authorized++
+		}
+	}
+
+	assert.Zero(t, timeouts,
+		"waitForFollowup must never report a timeout when auth has already completed; got %d/%d timeouts",
+		timeouts, iterations)
+	assert.Equal(t, iterations, authorized, "every completed registration must be returned as authorized")
+}

hscontrol/db/db.go

@@ -706,13 +706,20 @@ AND auth_key_id NOT IN (
 				// but this prevents deleting users whose nodes have been
 				// tagged, and the ON DELETE CASCADE FK would destroy the
 				// tagged nodes if the user were deleted.
+				//
+				// A nil tags slice marshals to the JSON literal 'null', so
+				// untagged nodes can carry tags='null'. That spelling must be
+				// excluded alongside '[]' and '' or untagged nodes lose their
+				// user. Nodes already detached by the earlier version of this
+				// migration are repaired by the recovery migration below.
 				// Fixes: https://github.com/juanfont/headscale/issues/3077
+				// Fixes: https://github.com/juanfont/headscale/issues/3323
 				ID: "202602201200-clear-tagged-node-user-id",
 				Migrate: func(tx *gorm.DB) error {
 					err := tx.Exec(`
 UPDATE nodes
 SET user_id = NULL
-WHERE tags IS NOT NULL AND tags != '[]' AND tags != '';
+WHERE tags IS NOT NULL AND tags != '[]' AND tags != '' AND tags != 'null';
 						`).Error
 					if err != nil {
 						return fmt.Errorf("clearing user_id on tagged nodes: %w", err)
@@ -744,6 +751,63 @@ WHERE expiry IS NOT NULL AND expiry < '1900-01-01';
 				},
 				Rollback: func(db *gorm.DB) error { return nil },
 			},
+			{
+				// Recover user_id on untagged nodes detached by the earlier
+				// version of 202602201200-clear-tagged-node-user-id, which
+				// treated tags='null' as tagged and cleared the user. This
+				// repairs databases that already upgraded to 0.29.0; fresh
+				// upgrades are protected by the fixed migration above and find
+				// nothing to repair. Recovery is best-effort: the owner is
+				// re-derived from the node's pre-auth key, so nodes registered
+				// via CLI/OIDC (no pre-auth key) cannot be recovered and must
+				// be reassigned manually.
+				// Fixes: https://github.com/juanfont/headscale/issues/3323
+				ID: "202606181200-recover-null-tags-node-user-id",
+				Migrate: func(tx *gorm.DB) error {
+					err := tx.Exec(`
+UPDATE nodes
+SET user_id = (
+	SELECT pak.user_id FROM pre_auth_keys pak WHERE pak.id = nodes.auth_key_id
+)
+WHERE user_id IS NULL
+	AND auth_key_id IS NOT NULL
+	AND (tags IS NULL OR tags = '' OR tags = '[]' OR tags = 'null');
+						`).Error
+					if err != nil {
+						return fmt.Errorf("recovering user_id on untagged nodes: %w", err)
+					}
+
+					return nil
+				},
+				Rollback: func(db *gorm.DB) error { return nil },
+			},
+			{
+				// Clear stale key expiry on tagged nodes. A tagged node is
+				// owned by its tags and never expires (KB 1068), but a buggy
+				// handleLogout stamped a past expiry on it, leaving it
+				// permanently Expired and unable to re-authenticate. The
+				// buggy writer is fixed, so this only repairs rows written
+				// before the upgrade; a fixed server cannot recreate them.
+				// Match the tagged-node predicate the earlier
+				// clear-tagged-node-user-id migration uses (a nil tags slice
+				// marshals to 'null', so exclude it).
+				// Fixes: https://github.com/juanfont/headscale/issues/3371
+				ID: "202607241200-clear-tagged-node-expiry",
+				Migrate: func(tx *gorm.DB) error {
+					err := tx.Exec(`
+UPDATE nodes
+SET expiry = NULL
+WHERE tags IS NOT NULL AND tags != '[]' AND tags != '' AND tags != 'null'
+	AND expiry IS NOT NULL;
+						`).Error
+					if err != nil {
+						return fmt.Errorf("clearing expiry on tagged nodes: %w", err)
+					}
+
+					return nil
+				},
+				Rollback: func(db *gorm.DB) error { return nil },
+			},
 		},
 	)
 

hscontrol/db/db_test.go

@@ -198,6 +198,161 @@ func TestSQLiteMigrationAndDataValidation(t *testing.T) {
 				assert.False(t, node5.IsExpired(), "node5 should not be reported as expired")
 			},
 		},
+		// Test for the clear-tagged-node-user-id migration
+		// (202602201200-clear-tagged-node-user-id). A nil tags slice
+		// marshals to the JSON literal 'null', so untagged nodes can carry
+		// tags='null' in the database. The migration must only clear
+		// user_id on genuinely tagged nodes, not on these untagged ones.
+		// Fixes: https://github.com/juanfont/headscale/issues/3323
+		{
+			dbPath: "testdata/sqlite/null_tags_user_id_migration_test.sql",
+			wantFunc: func(t *testing.T, hsdb *HSDatabase) {
+				t.Helper()
+
+				nodes, err := Read(hsdb.DB, func(rx *gorm.DB) (types.Nodes, error) {
+					return ListNodes(rx)
+				})
+				require.NoError(t, err)
+				require.Len(t, nodes, 4, "should have all 4 nodes")
+
+				byHostname := make(map[string]*types.Node, len(nodes))
+				for _, n := range nodes {
+					byHostname[n.Hostname] = n
+				}
+
+				// Node 1 had tags='null' (untagged) and belonged to user2.
+				// The migration must NOT clear its user_id.
+				node1 := byHostname["node1"]
+				require.NotNil(t, node1, "node1 should exist")
+				assert.False(t, node1.IsTagged(), "node1 with tags='null' should be untagged")
+				require.NotNil(t, node1.UserID, "node1 should keep its user assigned")
+				assert.Equal(t, uint(2), *node1.UserID, "node1 should still belong to user2")
+
+				// Node 2 is genuinely tagged; user_id must be cleared.
+				node2 := byHostname["node2"]
+				require.NotNil(t, node2, "node2 should exist")
+				assert.True(t, node2.IsTagged(), "node2 should be tagged")
+				assert.Nil(t, node2.UserID, "node2 (tagged) should have user_id cleared")
+
+				// Node 3 had tags='[]' (untagged); user_id preserved.
+				node3 := byHostname["node3"]
+				require.NotNil(t, node3, "node3 should exist")
+				assert.False(t, node3.IsTagged(), "node3 with tags='[]' should be untagged")
+				require.NotNil(t, node3.UserID, "node3 should keep its user assigned")
+				assert.Equal(t, uint(1), *node3.UserID, "node3 should still belong to user1")
+
+				// Node 4 had tags='' (untagged); user_id preserved.
+				node4 := byHostname["node4"]
+				require.NotNil(t, node4, "node4 should exist")
+				assert.False(t, node4.IsTagged(), "node4 with tags='' should be untagged")
+				require.NotNil(t, node4.UserID, "node4 should keep its user assigned")
+				assert.Equal(t, uint(1), *node4.UserID, "node4 should still belong to user1")
+			},
+		},
+		// Test for the null-tags user_id recovery migration. Databases that
+		// already upgraded to 0.29.0 had user_id wrongly cleared on untagged
+		// nodes with tags='null'. The recovery migration re-derives user_id
+		// from the node's pre-auth key where one exists.
+		// Fixes: https://github.com/juanfont/headscale/issues/3323
+		{
+			dbPath: "testdata/sqlite/recover_null_tags_user_id_migration_test.sql",
+			wantFunc: func(t *testing.T, hsdb *HSDatabase) {
+				t.Helper()
+
+				nodes, err := Read(hsdb.DB, func(rx *gorm.DB) (types.Nodes, error) {
+					return ListNodes(rx)
+				})
+				require.NoError(t, err)
+				require.Len(t, nodes, 4, "should have all 4 nodes")
+
+				byHostname := make(map[string]*types.Node, len(nodes))
+				for _, n := range nodes {
+					byHostname[n.Hostname] = n
+				}
+
+				// Node 1: authkey-registered, orphaned by the bug. The recovery
+				// migration restores user_id from its pre-auth key (user2).
+				node1 := byHostname["node1"]
+				require.NotNil(t, node1, "node1 should exist")
+				require.NotNil(t, node1.UserID, "node1 user_id should be recovered")
+				assert.Equal(t, uint(2), *node1.UserID, "node1 should be recovered to user2")
+
+				// Node 2: genuinely tagged, correctly cleared. Must stay cleared.
+				node2 := byHostname["node2"]
+				require.NotNil(t, node2, "node2 should exist")
+				assert.True(t, node2.IsTagged(), "node2 should be tagged")
+				assert.Nil(t, node2.UserID, "node2 (tagged) must remain cleared")
+
+				// Node 3: CLI-registered, no pre-auth key. Unrecoverable.
+				node3 := byHostname["node3"]
+				require.NotNil(t, node3, "node3 should exist")
+				assert.Nil(t, node3.UserID, "node3 has no pre-auth key to recover from")
+
+				// Node 4: never orphaned; user_id must be untouched.
+				node4 := byHostname["node4"]
+				require.NotNil(t, node4, "node4 should exist")
+				require.NotNil(t, node4.UserID, "node4 user_id should be untouched")
+				assert.Equal(t, uint(1), *node4.UserID, "node4 should still belong to user1")
+			},
+		},
+		// Test for the clear-tagged-node-expiry migration
+		// (202607241200-clear-tagged-node-expiry). A buggy handleLogout stamped
+		// a key expiry on tagged nodes, which never expire (KB 1068), leaving
+		// them permanently Expired. The migration clears expiry on tagged rows
+		// only, preserving user-owned nodes' expiry.
+		// Fixes: https://github.com/juanfont/headscale/issues/3371
+		{
+			dbPath: "testdata/sqlite/clear_tagged_node_expiry_migration_test.sql",
+			wantFunc: func(t *testing.T, hsdb *HSDatabase) {
+				t.Helper()
+
+				nodes, err := Read(hsdb.DB, func(rx *gorm.DB) (types.Nodes, error) {
+					return ListNodes(rx)
+				})
+				require.NoError(t, err)
+				require.Len(t, nodes, 5, "should have all 5 nodes")
+
+				byHostname := make(map[string]*types.Node, len(nodes))
+				for _, n := range nodes {
+					byHostname[n.Hostname] = n
+				}
+
+				// Node 1: tagged with a stale PAST expiry (the bug). Cleared.
+				node1 := byHostname["node1"]
+				require.NotNil(t, node1, "node1 should exist")
+				assert.True(t, node1.IsTagged(), "node1 should be tagged")
+				assert.Nil(t, node1.Expiry, "node1 (tagged) stale expiry should be cleared")
+				assert.False(t, node1.IsExpired(), "node1 must not be reported expired")
+
+				// Node 2: tagged with a FUTURE expiry. Tagged nodes never expire,
+				// so this is cleared too.
+				node2 := byHostname["node2"]
+				require.NotNil(t, node2, "node2 should exist")
+				assert.True(t, node2.IsTagged(), "node2 should be tagged")
+				assert.Nil(t, node2.Expiry, "node2 (tagged) expiry should be cleared")
+
+				// Node 3: tagged, expiry already NULL. Stays NULL.
+				node3 := byHostname["node3"]
+				require.NotNil(t, node3, "node3 should exist")
+				assert.True(t, node3.IsTagged(), "node3 should be tagged")
+				assert.Nil(t, node3.Expiry, "node3 (tagged) NULL expiry should be preserved")
+
+				// Node 4: untagged (tags='null') with a PAST expiry. PRESERVED —
+				// the migration must not touch user-owned nodes.
+				node4 := byHostname["node4"]
+				require.NotNil(t, node4, "node4 should exist")
+				assert.False(t, node4.IsTagged(), "node4 (tags='null') should be untagged")
+				require.NotNil(t, node4.Expiry, "node4 (user-owned) expiry must be preserved")
+				assert.Equal(t, 2020, node4.Expiry.UTC().Year(), "node4 past expiry preserved")
+
+				// Node 5: untagged (tags='[]') with a FUTURE expiry. PRESERVED.
+				node5 := byHostname["node5"]
+				require.NotNil(t, node5, "node5 should exist")
+				assert.False(t, node5.IsTagged(), "node5 (tags='[]') should be untagged")
+				require.NotNil(t, node5.Expiry, "node5 (user-owned) expiry must be preserved")
+				assert.Equal(t, 2099, node5.Expiry.UTC().Year(), "node5 future expiry preserved")
+			},
+		},
 	}
 
 	for _, tt := range tests {

hscontrol/db/node.go

@@ -496,6 +496,16 @@ func (e *EphemeralGarbageCollector) Cancel(nodeID types.NodeID) {
 	}
 }
 
+// IsScheduled reports whether a deletion timer is currently armed for nodeID.
+func (e *EphemeralGarbageCollector) IsScheduled(nodeID types.NodeID) bool {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	_, ok := e.toBeDeleted[nodeID]
+
+	return ok
+}
+
 // Start starts the garbage collector.
 func (e *EphemeralGarbageCollector) Start() {
 	for {

hscontrol/db/testdata/sqlite/clear_tagged_node_expiry_migration_test.sql

@@ -0,0 +1,85 @@
+-- Test SQL dump for the clear-tagged-node-expiry migration
+-- (202607241200-clear-tagged-node-expiry)
+--
+-- A buggy handleLogout stamped a past key expiry on tagged nodes. Tagged
+-- nodes are owned by their tags and never expire (KB 1068), so such a row is
+-- reported as permanently Expired and can never re-authenticate. The migration
+-- clears expiry on tagged rows while leaving user-owned nodes untouched.
+-- Fixes: https://github.com/juanfont/headscale/issues/3371
+
+PRAGMA foreign_keys=OFF;
+BEGIN TRANSACTION;
+
+-- Migrations table: entries applied up to (but not including) the fix. The
+-- intervening expiry migrations (clear-zero-time) also run against this dump;
+-- their predicates (expiry < 1900) do not match the post-2000 dates below, so
+-- they leave these rows for the new migration to handle.
+CREATE TABLE `migrations` (`id` text,PRIMARY KEY (`id`));
+INSERT INTO migrations VALUES('202312101416');
+INSERT INTO migrations VALUES('202312101430');
+INSERT INTO migrations VALUES('202402151347');
+INSERT INTO migrations VALUES('2024041121742');
+INSERT INTO migrations VALUES('202406021630');
+INSERT INTO migrations VALUES('202409271400');
+INSERT INTO migrations VALUES('202407191627');
+INSERT INTO migrations VALUES('202408181235');
+INSERT INTO migrations VALUES('202501221827');
+INSERT INTO migrations VALUES('202501311657');
+INSERT INTO migrations VALUES('202502070949');
+INSERT INTO migrations VALUES('202502131714');
+INSERT INTO migrations VALUES('202502171819');
+INSERT INTO migrations VALUES('202505091439');
+INSERT INTO migrations VALUES('202505141324');
+INSERT INTO migrations VALUES('202507021200');
+INSERT INTO migrations VALUES('202510311551');
+INSERT INTO migrations VALUES('202511101554-drop-old-idx');
+INSERT INTO migrations VALUES('202511011637-preauthkey-bcrypt');
+INSERT INTO migrations VALUES('202511122344-remove-newline-index');
+INSERT INTO migrations VALUES('202511131445-node-forced-tags-to-tags');
+INSERT INTO migrations VALUES('202601121700-migrate-hostinfo-request-tags');
+INSERT INTO migrations VALUES('202602201200-clear-tagged-node-user-id');
+
+-- Users table
+CREATE TABLE `users` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`name` text,`display_name` text,`email` text,`provider_identifier` text,`provider` text,`profile_pic_url` text);
+INSERT INTO users VALUES(1,'2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL,'user1','User One','user1@example.com',NULL,NULL,NULL);
+
+-- Pre-auth keys table
+CREATE TABLE `pre_auth_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`key` text,`user_id` integer,`reusable` numeric,`ephemeral` numeric DEFAULT false,`used` numeric DEFAULT false,`tags` text,`created_at` datetime,`expiration` datetime,`prefix` text,`hash` blob,CONSTRAINT `fk_pre_auth_keys_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE SET NULL);
+
+-- API keys table
+CREATE TABLE `api_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`prefix` text,`hash` blob,`created_at` datetime,`expiration` datetime,`last_seen` datetime);
+
+-- Nodes table - current schema (after the tags rename + last_seen/expiry reordering)
+CREATE TABLE IF NOT EXISTS "nodes" (`id` integer PRIMARY KEY AUTOINCREMENT,`machine_key` text,`node_key` text,`disco_key` text,`endpoints` text,`host_info` text,`ipv4` text,`ipv6` text,`hostname` text,`given_name` varchar(63),`user_id` integer,`register_method` text,`tags` text,`auth_key_id` integer,`last_seen` datetime,`expiry` datetime,`approved_routes` text,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,CONSTRAINT `fk_nodes_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE CASCADE,CONSTRAINT `fk_nodes_auth_key` FOREIGN KEY (`auth_key_id`) REFERENCES `pre_auth_keys`(`id`));
+
+-- Node 1: TAGGED, user_id NULL, stale PAST expiry (the #3371 bug). After migration: expiry NULL.
+INSERT INTO nodes VALUES(1,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e01','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605501','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57701','[]','{}','100.64.0.1','fd7a:115c:a1e0::1','node1','node1',NULL,'authkey','["tag:foo"]',NULL,'2024-01-01 00:00:00+00:00','2020-06-01 00:00:00+00:00','[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 2: TAGGED, user_id NULL, FUTURE expiry. Tagged nodes never expire, so cleared to NULL too.
+INSERT INTO nodes VALUES(2,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e02','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605502','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57702','[]','{}','100.64.0.2','fd7a:115c:a1e0::2','node2','node2',NULL,'authkey','["tag:foo"]',NULL,'2024-01-01 00:00:00+00:00','2099-01-01 00:00:00+00:00','[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 3: TAGGED, user_id NULL, expiry already NULL. After migration: still NULL.
+INSERT INTO nodes VALUES(3,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e03','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605503','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57703','[]','{}','100.64.0.3','fd7a:115c:a1e0::3','node3','node3',NULL,'authkey','["tag:foo"]',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 4: UNTAGGED user-owned (tags='null'), PAST expiry. Must be PRESERVED (guard against
+-- the #3323-class over-match: a nil tags slice marshals to the literal 'null').
+INSERT INTO nodes VALUES(4,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e04','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605504','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57704','[]','{}','100.64.0.4','fd7a:115c:a1e0::4','node4','node4',1,'cli','null',NULL,'2024-01-01 00:00:00+00:00','2020-06-01 00:00:00+00:00','[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 5: UNTAGGED user-owned (tags='[]'), FUTURE expiry. Must be PRESERVED.
+INSERT INTO nodes VALUES(5,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e05','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605505','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57705','[]','{}','100.64.0.5','fd7a:115c:a1e0::5','node5','node5',1,'cli','[]',NULL,'2024-01-01 00:00:00+00:00','2099-01-01 00:00:00+00:00','[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Policies table (empty)
+CREATE TABLE `policies` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`data` text);
+
+DELETE FROM sqlite_sequence;
+INSERT INTO sqlite_sequence VALUES('users',1);
+INSERT INTO sqlite_sequence VALUES('nodes',5);
+CREATE INDEX idx_users_deleted_at ON users(deleted_at);
+CREATE UNIQUE INDEX idx_api_keys_prefix ON api_keys(prefix);
+CREATE INDEX idx_policies_deleted_at ON policies(deleted_at);
+CREATE UNIQUE INDEX idx_provider_identifier ON users(provider_identifier) WHERE provider_identifier IS NOT NULL;
+CREATE UNIQUE INDEX idx_name_provider_identifier ON users(name, provider_identifier);
+CREATE UNIQUE INDEX idx_name_no_provider_identifier ON users(name) WHERE provider_identifier IS NULL;
+CREATE UNIQUE INDEX IF NOT EXISTS idx_pre_auth_keys_prefix ON pre_auth_keys(prefix) WHERE prefix IS NOT NULL AND prefix != '';
+
+COMMIT;

hscontrol/db/testdata/sqlite/null_tags_user_id_migration_test.sql

@@ -0,0 +1,85 @@
+-- Test SQL dump for the clear-tagged-node-user-id migration
+-- (202602201200-clear-tagged-node-user-id) against nodes whose tags
+-- column holds the JSON literal 'null'.
+--
+-- A nil Strings slice marshals to the JSON literal `null`, so pre-0.29
+-- databases contain untagged nodes with tags='null'. The migration's
+-- WHERE clause (tags IS NOT NULL AND tags != '[]' AND tags != '') treats
+-- the 4-character string 'null' as "tagged" and wrongly clears user_id,
+-- detaching the node from its owning user on upgrade.
+-- Fixes: https://github.com/juanfont/headscale/issues/3323
+
+PRAGMA foreign_keys=OFF;
+BEGIN TRANSACTION;
+
+-- Migrations table: every entry BEFORE clear-tagged-node-user-id has been
+-- applied. That migration is intentionally absent so it runs against this dump.
+CREATE TABLE `migrations` (`id` text,PRIMARY KEY (`id`));
+INSERT INTO migrations VALUES('202312101416');
+INSERT INTO migrations VALUES('202312101430');
+INSERT INTO migrations VALUES('202402151347');
+INSERT INTO migrations VALUES('2024041121742');
+INSERT INTO migrations VALUES('202406021630');
+INSERT INTO migrations VALUES('202409271400');
+INSERT INTO migrations VALUES('202407191627');
+INSERT INTO migrations VALUES('202408181235');
+INSERT INTO migrations VALUES('202501221827');
+INSERT INTO migrations VALUES('202501311657');
+INSERT INTO migrations VALUES('202502070949');
+INSERT INTO migrations VALUES('202502131714');
+INSERT INTO migrations VALUES('202502171819');
+INSERT INTO migrations VALUES('202505091439');
+INSERT INTO migrations VALUES('202505141324');
+INSERT INTO migrations VALUES('202507021200');
+INSERT INTO migrations VALUES('202510311551');
+INSERT INTO migrations VALUES('202511101554-drop-old-idx');
+INSERT INTO migrations VALUES('202511011637-preauthkey-bcrypt');
+INSERT INTO migrations VALUES('202511122344-remove-newline-index');
+INSERT INTO migrations VALUES('202511131445-node-forced-tags-to-tags');
+INSERT INTO migrations VALUES('202601121700-migrate-hostinfo-request-tags');
+
+-- Users table
+CREATE TABLE `users` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`name` text,`display_name` text,`email` text,`provider_identifier` text,`provider` text,`profile_pic_url` text);
+INSERT INTO users VALUES(1,'2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL,'user1','User One','user1@example.com',NULL,NULL,NULL);
+INSERT INTO users VALUES(2,'2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL,'user2','User Two','user2@example.com',NULL,NULL,NULL);
+
+-- Pre-auth keys table
+CREATE TABLE `pre_auth_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`key` text,`user_id` integer,`reusable` numeric,`ephemeral` numeric DEFAULT false,`used` numeric DEFAULT false,`tags` text,`created_at` datetime,`expiration` datetime,`prefix` text,`hash` blob,CONSTRAINT `fk_pre_auth_keys_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE SET NULL);
+
+-- API keys table
+CREATE TABLE `api_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`prefix` text,`hash` blob,`created_at` datetime,`expiration` datetime,`last_seen` datetime);
+
+-- Nodes table - current schema (after the tags rename + last_seen/expiry reordering)
+CREATE TABLE IF NOT EXISTS "nodes" (`id` integer PRIMARY KEY AUTOINCREMENT,`machine_key` text,`node_key` text,`disco_key` text,`endpoints` text,`host_info` text,`ipv4` text,`ipv6` text,`hostname` text,`given_name` varchar(63),`user_id` integer,`register_method` text,`tags` text,`auth_key_id` integer,`last_seen` datetime,`expiry` datetime,`approved_routes` text,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,CONSTRAINT `fk_nodes_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE CASCADE,CONSTRAINT `fk_nodes_auth_key` FOREIGN KEY (`auth_key_id`) REFERENCES `pre_auth_keys`(`id`));
+
+-- Node 1: tags='null' (untagged, nil slice marshalled to JSON null), owned by user2.
+-- After migration: user_id MUST be preserved (this is the bug).
+INSERT INTO nodes VALUES(1,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e01','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605501','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57701','[]','{}','100.64.0.1','fd7a:115c:a1e0::1','node1','node1',2,'cli','null',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 2: genuinely tagged, owned by user1.
+-- After migration: user_id MUST be cleared to NULL (tagged nodes are owned by tags).
+INSERT INTO nodes VALUES(2,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e02','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605502','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57702','[]','{}','100.64.0.2','fd7a:115c:a1e0::2','node2','node2',1,'cli','["tag:server"]',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 3: empty-array tags (untagged), owned by user1.
+-- After migration: user_id MUST be preserved.
+INSERT INTO nodes VALUES(3,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e03','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605503','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57703','[]','{}','100.64.0.3','fd7a:115c:a1e0::3','node3','node3',1,'cli','[]',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 4: empty-string tags (untagged), owned by user1.
+-- After migration: user_id MUST be preserved.
+INSERT INTO nodes VALUES(4,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e04','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605504','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57704','[]','{}','100.64.0.4','fd7a:115c:a1e0::4','node4','node4',1,'cli','',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Policies table (empty)
+CREATE TABLE `policies` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`data` text);
+
+DELETE FROM sqlite_sequence;
+INSERT INTO sqlite_sequence VALUES('users',2);
+INSERT INTO sqlite_sequence VALUES('nodes',4);
+CREATE INDEX idx_users_deleted_at ON users(deleted_at);
+CREATE UNIQUE INDEX idx_api_keys_prefix ON api_keys(prefix);
+CREATE INDEX idx_policies_deleted_at ON policies(deleted_at);
+CREATE UNIQUE INDEX idx_provider_identifier ON users(provider_identifier) WHERE provider_identifier IS NOT NULL;
+CREATE UNIQUE INDEX idx_name_provider_identifier ON users(name, provider_identifier);
+CREATE UNIQUE INDEX idx_name_no_provider_identifier ON users(name) WHERE provider_identifier IS NULL;
+CREATE UNIQUE INDEX IF NOT EXISTS idx_pre_auth_keys_prefix ON pre_auth_keys(prefix) WHERE prefix IS NOT NULL AND prefix != '';
+
+COMMIT;

hscontrol/db/testdata/sqlite/recover_null_tags_user_id_migration_test.sql

@@ -0,0 +1,87 @@
+-- Test SQL dump for the null-tags user_id RECOVERY migration.
+--
+-- Represents a database that already upgraded to 0.29.0, where the buggy
+-- clear-tagged-node-user-id migration (202602201200) already cleared
+-- user_id on untagged nodes whose tags column held 'null'. The recovery
+-- migration runs against this state and re-derives user_id from the node's
+-- pre-auth key where possible.
+-- Fixes: https://github.com/juanfont/headscale/issues/3323
+
+PRAGMA foreign_keys=OFF;
+BEGIN TRANSACTION;
+
+-- Migrations table: everything through the current last migration has been
+-- applied (this DB already ran the buggy clear-tagged migration). The new
+-- recovery migration is intentionally absent so it runs against this dump.
+CREATE TABLE `migrations` (`id` text,PRIMARY KEY (`id`));
+INSERT INTO migrations VALUES('202312101416');
+INSERT INTO migrations VALUES('202312101430');
+INSERT INTO migrations VALUES('202402151347');
+INSERT INTO migrations VALUES('2024041121742');
+INSERT INTO migrations VALUES('202406021630');
+INSERT INTO migrations VALUES('202409271400');
+INSERT INTO migrations VALUES('202407191627');
+INSERT INTO migrations VALUES('202408181235');
+INSERT INTO migrations VALUES('202501221827');
+INSERT INTO migrations VALUES('202501311657');
+INSERT INTO migrations VALUES('202502070949');
+INSERT INTO migrations VALUES('202502131714');
+INSERT INTO migrations VALUES('202502171819');
+INSERT INTO migrations VALUES('202505091439');
+INSERT INTO migrations VALUES('202505141324');
+INSERT INTO migrations VALUES('202507021200');
+INSERT INTO migrations VALUES('202510311551');
+INSERT INTO migrations VALUES('202511101554-drop-old-idx');
+INSERT INTO migrations VALUES('202511011637-preauthkey-bcrypt');
+INSERT INTO migrations VALUES('202511122344-remove-newline-index');
+INSERT INTO migrations VALUES('202511131445-node-forced-tags-to-tags');
+INSERT INTO migrations VALUES('202601121700-migrate-hostinfo-request-tags');
+INSERT INTO migrations VALUES('202602201200-clear-tagged-node-user-id');
+INSERT INTO migrations VALUES('202605221435-clear-zero-time-node-expiry');
+
+-- Users table
+CREATE TABLE `users` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`name` text,`display_name` text,`email` text,`provider_identifier` text,`provider` text,`profile_pic_url` text);
+INSERT INTO users VALUES(1,'2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL,'user1','User One','user1@example.com',NULL,NULL,NULL);
+INSERT INTO users VALUES(2,'2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL,'user2','User Two','user2@example.com',NULL,NULL,NULL);
+
+-- Pre-auth keys table. Key 1 belongs to user2, key 2 to user1.
+CREATE TABLE `pre_auth_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`key` text,`user_id` integer,`reusable` numeric,`ephemeral` numeric DEFAULT false,`used` numeric DEFAULT false,`tags` text,`created_at` datetime,`expiration` datetime,`prefix` text,`hash` blob,CONSTRAINT `fk_pre_auth_keys_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE SET NULL);
+INSERT INTO pre_auth_keys VALUES(1,NULL,2,1,false,true,NULL,'2024-01-01 00:00:00+00:00',NULL,'pak1',NULL);
+INSERT INTO pre_auth_keys VALUES(2,NULL,1,1,false,true,NULL,'2024-01-01 00:00:00+00:00',NULL,'pak2',NULL);
+
+-- API keys table
+CREATE TABLE `api_keys` (`id` integer PRIMARY KEY AUTOINCREMENT,`prefix` text,`hash` blob,`created_at` datetime,`expiration` datetime,`last_seen` datetime);
+
+-- Nodes table
+CREATE TABLE IF NOT EXISTS "nodes" (`id` integer PRIMARY KEY AUTOINCREMENT,`machine_key` text,`node_key` text,`disco_key` text,`endpoints` text,`host_info` text,`ipv4` text,`ipv6` text,`hostname` text,`given_name` varchar(63),`user_id` integer,`register_method` text,`tags` text,`auth_key_id` integer,`last_seen` datetime,`expiry` datetime,`approved_routes` text,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,CONSTRAINT `fk_nodes_user` FOREIGN KEY (`user_id`) REFERENCES `users`(`id`) ON DELETE CASCADE,CONSTRAINT `fk_nodes_auth_key` FOREIGN KEY (`auth_key_id`) REFERENCES `pre_auth_keys`(`id`));
+
+-- Node 1: authkey-registered, tags='null', already orphaned (user_id NULL) by
+-- the buggy migration. auth_key_id=1 (user2). Recovery: user_id -> 2.
+INSERT INTO nodes VALUES(1,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e01','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605501','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57701','[]','{}','100.64.0.1','fd7a:115c:a1e0::1','node1','node1',NULL,'authkey','null',1,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 2: genuinely tagged, user_id correctly cleared. Must stay NULL.
+INSERT INTO nodes VALUES(2,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e02','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605502','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57702','[]','{}','100.64.0.2','fd7a:115c:a1e0::2','node2','node2',NULL,'authkey','["tag:server"]',2,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 3: CLI-registered, tags='null', orphaned, no auth_key_id.
+-- Unrecoverable: must stay NULL.
+INSERT INTO nodes VALUES(3,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e03','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605503','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57703','[]','{}','100.64.0.3','fd7a:115c:a1e0::3','node3','node3',NULL,'cli','null',NULL,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Node 4: authkey-registered, untouched (user_id still set). Must stay user1.
+INSERT INTO nodes VALUES(4,'mkey:a0ab77456320823945ae0331823e3c0d516fae9585bd42698dfa1ac3d7679e04','nodekey:7c84167ab68f494942de14deb83587fd841843de2bac105b6c670048c1605504','discokey:53075b3c6cad3b62a2a29caea61beeb93f66b8c75cb89dac465236a5bbf57704','[]','{}','100.64.0.4','fd7a:115c:a1e0::4','node4','node4',1,'authkey','null',2,'2024-01-01 00:00:00+00:00',NULL,'[]','2024-01-01 00:00:00+00:00','2024-01-01 00:00:00+00:00',NULL);
+
+-- Policies table (empty)
+CREATE TABLE `policies` (`id` integer PRIMARY KEY AUTOINCREMENT,`created_at` datetime,`updated_at` datetime,`deleted_at` datetime,`data` text);
+
+DELETE FROM sqlite_sequence;
+INSERT INTO sqlite_sequence VALUES('users',2);
+INSERT INTO sqlite_sequence VALUES('pre_auth_keys',2);
+INSERT INTO sqlite_sequence VALUES('nodes',4);
+CREATE INDEX idx_users_deleted_at ON users(deleted_at);
+CREATE UNIQUE INDEX idx_api_keys_prefix ON api_keys(prefix);
+CREATE INDEX idx_policies_deleted_at ON policies(deleted_at);
+CREATE UNIQUE INDEX idx_provider_identifier ON users(provider_identifier) WHERE provider_identifier IS NOT NULL;
+CREATE UNIQUE INDEX idx_name_provider_identifier ON users(name, provider_identifier);
+CREATE UNIQUE INDEX idx_name_no_provider_identifier ON users(name) WHERE provider_identifier IS NULL;
+CREATE UNIQUE INDEX IF NOT EXISTS idx_pre_auth_keys_prefix ON pre_auth_keys(prefix) WHERE prefix IS NOT NULL AND prefix != '';
+
+COMMIT;

hscontrol/handlers.go

@@ -19,17 +19,6 @@ import (
 )
 
 const (
-	// NoiseCapabilityVersion is used by Tailscale clients to indicate
-	// their codebase version. Tailscale clients can communicate over TS2021
-	// from CapabilityVersion 28, but we only have good support for it
-	// since https://github.com/tailscale/tailscale/pull/4323 (Noise in any HTTPS port).
-	//
-	// Related to this change, there is https://github.com/tailscale/tailscale/pull/5379,
-	// where CapabilityVersion 39 is introduced to indicate #4323 was merged.
-	//
-	// See also https://github.com/tailscale/tailscale/blob/main/tailcfg/tailcfg.go
-	NoiseCapabilityVersion = 39
-
 	reservedResponseHeaderSize = 4
 )
 
@@ -202,21 +191,28 @@ func (h *Headscale) KeyHandler(
 		return
 	}
 
-	// TS2021 (Tailscale v2 protocol) requires to have a different key
-	if capVer >= NoiseCapabilityVersion {
-		resp := tailcfg.OverTLSPublicKeyResponse{
-			PublicKey: h.noisePrivateKey.Public(),
-		}
-
-		writer.Header().Set("Content-Type", "application/json")
-
-		err := json.NewEncoder(writer).Encode(resp)
-		if err != nil {
-			log.Error().Err(err).Msg("failed to encode public key response")
-		}
-
+	// Only disclose the Noise public key to clients this server can
+	// actually complete a handshake with. Gating on the same floor the
+	// Noise handshake enforces (capver.MinSupportedCapabilityVersion, see
+	// isSupportedVersion in noise.go) keeps /key consistent with /ts2021:
+	// versions the handshake would reject get a clear rejection here
+	// instead of a key that only serves as a version-boundary oracle.
+	// See https://github.com/juanfont/headscale/issues/3380.
+	if !isSupportedVersion(capVer) {
+		httpError(writer, NewHTTPError(http.StatusBadRequest, "unsupported client version", unsupportedClientError(capVer)))
 		return
 	}
+
+	resp := tailcfg.OverTLSPublicKeyResponse{
+		PublicKey: h.noisePrivateKey.Public(),
+	}
+
+	writer.Header().Set("Content-Type", "application/json")
+
+	err = json.NewEncoder(writer).Encode(resp)
+	if err != nil {
+		log.Error().Err(err).Msg("failed to encode public key response")
+	}
 }
 
 func (h *Headscale) HealthHandler(

hscontrol/handlers_test.go

@@ -4,12 +4,15 @@ import (
 	"bytes"
 	"context"
 	"errors"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
 
+	"github.com/juanfont/headscale/hscontrol/capver"
 	"github.com/stretchr/testify/assert"
+	"tailscale.com/types/key"
 )
 
 var errTestUnexpected = errors.New("unexpected failure")
@@ -47,6 +50,52 @@ func TestHandleVerifyRequest_OversizedBodyRejected(t *testing.T) {
 		"oversized body must surface 413")
 }
 
+// TestKeyHandler_UnsupportedCapVerDoesNotLeakKey reproduces
+// https://github.com/juanfont/headscale/issues/3380. The /key handler
+// must gate key disclosure on the same floor the Noise handshake
+// enforces (capver.MinSupportedCapabilityVersion). A capability version
+// below that floor can never complete a handshake, so it must be
+// rejected rather than handed the server's Noise public key, which would
+// otherwise serve only as a fingerprint / version-boundary oracle.
+func TestKeyHandler_UnsupportedCapVerDoesNotLeakKey(t *testing.T) {
+	t.Parallel()
+
+	noise := key.NewMachine()
+	h := &Headscale{noisePrivateKey: &noise}
+
+	unsupported := capver.MinSupportedCapabilityVersion - 1
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequestWithContext(
+		context.Background(),
+		http.MethodGet,
+		fmt.Sprintf("/key?v=%d", unsupported),
+		nil,
+	)
+
+	h.KeyHandler(rec, req)
+
+	assert.Equal(t, http.StatusBadRequest, rec.Code,
+		"a client below the supported floor must be rejected")
+	assert.NotContains(t, rec.Body.String(), noise.Public().String(),
+		"must not disclose Noise public key to a client below the supported floor")
+
+	// A supported client still receives the key.
+	recOK := httptest.NewRecorder()
+	reqOK := httptest.NewRequestWithContext(
+		context.Background(),
+		http.MethodGet,
+		fmt.Sprintf("/key?v=%d", capver.MinSupportedCapabilityVersion),
+		nil,
+	)
+
+	h.KeyHandler(recOK, reqOK)
+
+	assert.Equal(t, http.StatusOK, recOK.Code)
+	assert.Contains(t, recOK.Body.String(), noise.Public().String(),
+		"a supported client must receive the Noise public key")
+}
+
 // errorAsHTTPError is a small local helper that unwraps an [HTTPError]
 // from an error chain.
 func errorAsHTTPError(err error) (HTTPError, bool) {

hscontrol/mapper/builder.go

@@ -9,6 +9,8 @@ import (
 	"github.com/juanfont/headscale/hscontrol/policy"
 	policyv2 "github.com/juanfont/headscale/hscontrol/policy/v2"
 	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/juanfont/headscale/hscontrol/util/zlog/zf"
+	"github.com/rs/zerolog/log"
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/views"
 	"tailscale.com/util/multierr"
@@ -144,7 +146,15 @@ func (b *MapResponseBuilder) WithSSHPolicy() *MapResponseBuilder {
 
 	sshPolicy, err := b.mapper.state.SSHPolicy(node)
 	if err != nil {
-		b.addError(err)
+		// SSH policy is optional for a node to function. Rather than fail the
+		// whole map (leaving the node unable to connect), log and continue
+		// without it; the node still receives a usable netmap.
+		log.Warn().Caller().
+			Err(err).
+			Uint64(zf.NodeID, node.ID().Uint64()).
+			Str(zf.NodeHostname, node.Hostname()).
+			Msg("building map response: skipping SSH policy for node; node will receive a map without SSH rules")
+
 		return b
 	}
 
@@ -279,7 +289,19 @@ func (b *MapResponseBuilder) buildTailPeers(peers views.Slice[types.NodeView]) (
 			return b.mapper.state.RoutesForPeer(node, peer, matchers)
 		}, b.mapper.cfg, allCapMaps[peer.ID()])
 		if err != nil {
-			return nil, err
+			// One peer with invalid data (e.g. an empty or over-long
+			// GivenName that fails GetFQDN) must not blank out the map for
+			// every node that can see it. Drop the offending peer, log it
+			// with the identity an operator needs to fix it, and keep
+			// building from the remaining valid peers.
+			log.Warn().Caller().
+				Err(err).
+				Uint64(zf.NodeID, peer.ID().Uint64()).
+				Str(zf.NodeHostname, peer.Hostname()).
+				Uint64("map.viewer.node.id", b.nodeID.Uint64()).
+				Msgf("dropping peer %d from map response: invalid node data; fix with `headscale nodes rename %d <name>`", peer.ID(), peer.ID())
+
+			continue
 		}
 
 		// [tailcfg.Node.CapMap] on a peer carries the small set of

hscontrol/mapper/initialmap_storm_test.go

@@ -0,0 +1,189 @@
+//go:build !race
+
+// This is a timing-sensitive performance regression test; the race detector's
+// ~10x slowdown makes its wall-clock assertion meaningless, so it is excluded
+// from -race builds. The concurrency correctness of the policy lock change it
+// guards is covered under -race by TestPolicyManagerConcurrentReads in
+// hscontrol/policy/v2.
+
+package mapper
+
+import (
+	"net/netip"
+	"runtime"
+	"slices"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/juanfont/headscale/hscontrol/db"
+	"github.com/juanfont/headscale/hscontrol/derp"
+	"github.com/juanfont/headscale/hscontrol/state"
+	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"tailscale.com/tailcfg"
+)
+
+// setupStormBatcher builds a real state+batcher with production-default
+// NodeStore batching so the reconnect-storm contention is realistic. It mirrors
+// setupBatcherWithTestData but lets the test control BatcherWorkers and the
+// policy.
+func setupStormBatcher(tb testing.TB, nodeCount, workers int, policy string) (*TestData, func()) {
+	tb.Helper()
+
+	tmpDir := tb.TempDir()
+	prefixV4 := netip.MustParsePrefix("100.64.0.0/10")
+	prefixV6 := netip.MustParsePrefix("fd7a:115c:a1e0::/48")
+
+	cfg := &types.Config{
+		Database: types.DatabaseConfig{
+			Type:   types.DatabaseSqlite,
+			Sqlite: types.SqliteConfig{Path: tmpDir + "/headscale_test.db"},
+		},
+		PrefixV4:     &prefixV4,
+		PrefixV6:     &prefixV6,
+		IPAllocation: types.IPAllocationStrategySequential,
+		BaseDomain:   "headscale.test",
+		Policy:       types.PolicyConfig{Mode: types.PolicyModeDB},
+		DERP: types.DERPConfig{
+			ServerEnabled: false,
+			DERPMap: &tailcfg.DERPMap{
+				Regions: map[int]*tailcfg.DERPRegion{999: {RegionID: 999}},
+			},
+		},
+		Tuning: types.Tuning{
+			BatchChangeDelay: 10 * time.Millisecond,
+			BatcherWorkers:   workers,
+			// Production defaults: coalesce writes so the storm is not
+			// exaggerated by an unrealistically small NodeStore batch.
+			NodeStoreBatchSize:    100,
+			NodeStoreBatchTimeout: 500 * time.Millisecond,
+		},
+	}
+
+	database, err := db.NewHeadscaleDatabase(cfg)
+	require.NoError(tb, err)
+
+	users := database.CreateUsersForTest(1, "testuser")
+	dbNodes := database.CreateRegisteredNodesForTest(users[0], nodeCount, "node")
+
+	allNodes := make([]node, 0, nodeCount)
+	for i := range dbNodes {
+		allNodes = append(allNodes, node{
+			n:  dbNodes[i],
+			ch: make(chan *tailcfg.MapResponse, normalBufferSize),
+		})
+	}
+
+	st, err := state.NewState(cfg)
+	require.NoError(tb, err)
+
+	derpMap, err := derp.GetDERPMap(cfg.DERP)
+	require.NoError(tb, err)
+	st.SetDERPMap(derpMap)
+
+	_, err = st.SetPolicy([]byte(policy))
+	require.NoError(tb, err)
+
+	batcher := wrapBatcherForTest(NewBatcherAndMapper(cfg, st), st)
+	batcher.Start()
+
+	td := &TestData{
+		Database: database,
+		Users:    users,
+		Nodes:    allNodes,
+		State:    st,
+		Config:   cfg,
+		Batcher:  batcher,
+	}
+
+	return td, func() {
+		batcher.Close()
+		st.Close()
+		database.Close()
+	}
+}
+
+// TestInitialMapNotStarvedByReconnectStorm reproduces juanfont/headscale#3346.
+//
+// When every node redials at once (e.g. after a server upgrade restart), each
+// connection writes the NodeStore (UpdateNodeFromMapRequest + Connect) and the
+// batcher generates its initial map. All of that reads the policy through the
+// PolicyManager. Before the fix the PolicyManager guarded every read with a
+// single exclusive mutex, so the NodeStore writer's O(n^2) BuildPeerMap and
+// every node's FilterForNode serialised against each other. On a per-node
+// filter policy (autogroup:self, via, relay grants) each hold is expensive, so
+// under the storm time-to-initial-map grew without bound.
+//
+// On the production server in #3346 this drove the batcher's per-node
+// total.duration from ~4s to ~76s; tailscale clients aborted the map POST
+// first and reported
+//
+//	PollNetMap: Post ".../machine/map": unexpected EOF
+//
+// then redialled, feeding the storm so it never converged. An allow-all policy
+// does NOT reproduce this — BuildPeerMap is cheap there; the per-node filter
+// path is what makes it expensive, matching a real deployment's ACLs.
+//
+// The fix makes PolicyManager reads take a shared RLock so map generation runs
+// concurrently. AddNode blocks until the initial map is generated and handed to
+// the node channel, so its wall-clock duration is the time-to-initial-map the
+// client experiences. Without the fix this test's slowest node takes ~10s+ at
+// this scale (lock-bound, and more workers do not help); with it, generation
+// parallelises across workers and stays well within a client's patience.
+func TestInitialMapNotStarvedByReconnectStorm(t *testing.T) {
+	if testing.Short() {
+		t.Skip("timing-sensitive storm regression; skipped in -short")
+	}
+
+	const (
+		nodeCount = 300
+
+		// A per-node-filter policy: forces BuildPeerMap and FilterForNode onto
+		// the slow path that recompiles filter rules per node, the same shape
+		// as a real ACL using autogroup:self / via / relay grants.
+		perNodeFilterPolicy = `{"acls":[{"action":"accept","src":["autogroup:member"],"dst":["autogroup:self:*"]}]}`
+
+		// Deliberately roomy so it passes on CI's few-core runners, where the
+		// single-writer BuildPeerMap sets the floor (~10s) whatever the reads
+		// do. It still trips on a hang or a return to the ~76s serialised
+		// behaviour; the fine-grained concurrency is verified separately by
+		// TestPolicyManagerConcurrentReads under -race.
+		maxAcceptableLatency = 30 * time.Second
+	)
+
+	// Use the real available parallelism, as production does.
+	workers := runtime.NumCPU()
+
+	td, cleanup := setupStormBatcher(t, nodeCount, workers, perNodeFilterPolicy)
+	defer cleanup()
+
+	latencies := make([]time.Duration, nodeCount)
+
+	var wg sync.WaitGroup
+
+	for i := range td.Nodes {
+		wg.Go(func() {
+			n := &td.Nodes[i]
+
+			start := time.Now()
+			err := td.Batcher.AddNode(n.n.ID, n.ch, tailcfg.CapabilityVersion(100), nil)
+			latencies[i] = time.Since(start)
+
+			assert.NoError(t, err) //nolint:testifylint // assert (not require) is correct off the test goroutine
+		})
+	}
+
+	wg.Wait()
+
+	slices.Sort(latencies)
+	p50 := latencies[len(latencies)/2]
+	p95 := latencies[len(latencies)*95/100]
+	maxLatency := latencies[len(latencies)-1]
+	t.Logf("initial-map latency over %d nodes (workers=%d): p50=%s p95=%s max=%s",
+		nodeCount, workers, p50, p95, maxLatency)
+
+	require.Less(t, maxLatency, maxAcceptableLatency,
+		"slowest initial map took %s: policy reads are serialising instead of running concurrently (issue #3346)", maxLatency)
+}

hscontrol/mapper/mapper_test.go

@@ -3,6 +3,7 @@ package mapper
 import (
 	"fmt"
 	"net/netip"
+	"strings"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
@@ -536,6 +537,94 @@ func TestBuildFromChangeVisibilityMatchesFullMap(t *testing.T) {
 	}
 }
 
+// TestFullMapResponseSurvivesPeerWithInvalidName proves a single node with an
+// FQDN-invalid GivenName must not break map generation for its peers.
+//
+// A node whose stored GivenName is empty (ErrNodeHasNoGivenName) or yields an
+// FQDN longer than MaxHostnameLength (ErrHostnameTooLong) makes GetFQDN, and
+// therefore TailNode, return an error. buildTailPeers used to abort the entire
+// peer list on the first such error, so MapResponseBuilder.Build() failed for
+// every node that could see the bad peer; on the initial-connection path that
+// surfaced as "PollNetMap: ... unexpected EOF" and the "Unable to connect to
+// the Tailscale coordination server" health warning. A legacy DB row loads
+// verbatim (NewNodeStore reads db.ListNodes() without re-sanitising names), so
+// the bad peer persists across restart. The build for an unaffected viewer
+// must succeed: the bad peer is dropped, valid peers and self survive.
+func TestFullMapResponseSurvivesPeerWithInvalidName(t *testing.T) {
+	for _, tt := range []struct {
+		name    string
+		badName string
+	}{
+		{"empty given name", ""},
+		{"over-long fqdn", strings.Repeat("a", types.MaxHostnameLength+1)},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			tmp := t.TempDir()
+			p4 := netip.MustParsePrefix("100.64.0.0/10")
+			p6 := netip.MustParsePrefix("fd7a:115c:a1e0::/48")
+			cfg := &types.Config{
+				Database: types.DatabaseConfig{
+					Type:   types.DatabaseSqlite,
+					Sqlite: types.SqliteConfig{Path: tmp + "/h.db"},
+				},
+				PrefixV4:     &p4,
+				PrefixV6:     &p6,
+				IPAllocation: types.IPAllocationStrategySequential,
+				BaseDomain:   "headscale.test",
+				Policy:       types.PolicyConfig{Mode: types.PolicyModeDB},
+				DERP: types.DERPConfig{
+					DERPMap: &tailcfg.DERPMap{
+						Regions: map[int]*tailcfg.DERPRegion{999: {RegionID: 999}},
+					},
+				},
+				Tuning: types.Tuning{
+					NodeStoreBatchSize:    state.TestBatchSize,
+					NodeStoreBatchTimeout: state.TestBatchTimeout,
+				},
+			}
+
+			database, err := db.NewHeadscaleDatabase(cfg)
+			require.NoError(t, err)
+
+			user := database.CreateUserForTest("u1")
+			n1 := database.CreateRegisteredNodeForTest(user, "n1")     // viewer, valid
+			bad := database.CreateRegisteredNodeForTest(user, "bad")   // peer, name corrupted below
+			good := database.CreateRegisteredNodeForTest(user, "good") // peer, valid control
+
+			// Simulate a legacy/corrupt row that v29 loads verbatim.
+			require.NoError(t, database.DB.
+				Model(&types.Node{}).
+				Where("id = ?", bad.ID).
+				Update("given_name", tt.badName).Error)
+			require.NoError(t, database.Close())
+
+			s, err := state.NewState(cfg)
+			require.NoError(t, err)
+			t.Cleanup(func() { _ = s.Close() })
+
+			// Allow-all so n1 sees both peers; the bad one must still be dropped.
+			_, err = s.SetPolicy([]byte(`{"acls":[{"action":"accept","src":["*"],"dst":["*:*"]}]}`))
+			require.NoError(t, err)
+
+			m := &mapper{state: s, cfg: cfg}
+			capVer := tailcfg.CurrentCapabilityVersion
+
+			resp, err := m.fullMapResponse(n1.ID, capVer)
+			require.NoError(t, err, "n1's map must build despite a peer with an invalid name")
+			require.NotNil(t, resp)
+			require.NotNil(t, resp.Node, "n1 must receive its own self node")
+
+			peers := map[tailcfg.NodeID]bool{}
+			for _, p := range resp.Peers {
+				peers[p.ID] = true
+			}
+
+			assert.False(t, peers[bad.ID.NodeID()], "the peer with an invalid name must be dropped")
+			assert.True(t, peers[good.ID.NodeID()], "valid peers must remain in the map")
+		})
+	}
+}
+
 // TestGenerateDNSConfigNilHostinfoNoPanic proves generateDNSConfig does not
 // panic when a node's Hostinfo is nil (e.g. a legacy DB row with a NULL
 // host_info column). addNextDNSMetadata dereferenced node.Hostinfo().OS()

hscontrol/noise_test.go

@@ -459,6 +459,71 @@ func TestSSHActionHandler_RejectsMissingSessionWithoutCheck(t *testing.T) {
 		"a bogus auth_id with no active check must be rejected, body=%s", rec.Body.String())
 }
 
+// TestTS2021Route_AcceptsGETAndPOST reproduces a regression where the
+// browser/WASM control client could not connect. Tailscale's JS/WASM control
+// client opens /ts2021 as a WebSocket, which is an HTTP GET upgrade; the native
+// Go client uses an HTTP POST upgrade. The gorilla->chi router migration
+// registered /ts2021 for POST only, so the GET WebSocket handshake was rejected
+// with 405 Method Not Allowed by the router before it could reach
+// NoiseUpgradeHandler. Both methods must route to the handler.
+//
+// NoiseUpgradeHandler dispatches on the Upgrade header, not the HTTP method, so
+// once the route is reachable the handler handles both upgrade styles. The
+// httptest recorder is not an http.Hijacker, so the upgrade itself fails past
+// the router (501 for the WebSocket path, 400 for the native path) — the point
+// is only that neither is 405, i.e. the router no longer rejects GET early.
+func TestTS2021Route_AcceptsGETAndPOST(t *testing.T) {
+	t.Parallel()
+
+	handler := createTestApp(t).HTTPHandler()
+
+	tests := []struct {
+		name    string
+		method  string
+		headers map[string]string
+	}{
+		{
+			name:   "websocket_get_from_wasm_client",
+			method: http.MethodGet,
+			headers: map[string]string{
+				"Connection":             "Upgrade",
+				"Upgrade":                "websocket",
+				"Sec-WebSocket-Version":  "13",
+				"Sec-WebSocket-Key":      "dGhlIHNhbXBsZSBub25jZQ==",
+				"Sec-WebSocket-Protocol": "tailscale-control-protocol",
+			},
+		},
+		{
+			name:   "native_post_upgrade",
+			method: http.MethodPost,
+			headers: map[string]string{
+				"Connection":            "upgrade",
+				"Upgrade":               "tailscale-control-protocol",
+				"X-Tailscale-Handshake": "AAAA",
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			req := httptest.NewRequestWithContext(context.Background(), tt.method,
+				"/ts2021?X-Tailscale-Handshake=AAAA", nil)
+			for k, v := range tt.headers {
+				req.Header.Set(k, v)
+			}
+
+			rec := httptest.NewRecorder()
+			handler.ServeHTTP(rec, req)
+
+			assert.NotEqual(t, http.StatusMethodNotAllowed, rec.Code,
+				"%s /ts2021 must reach NoiseUpgradeHandler, not be rejected by the router with 405",
+				tt.method)
+		})
+	}
+}
+
 // newSSHActionFollowUpRequest is like newSSHActionRequest but carries the
 // auth_id query parameter that marks a follow-up poll.
 func newSSHActionFollowUpRequest(t *testing.T, src, dst types.NodeID, authID types.AuthID) *http.Request {

hscontrol/policy/pm.go

@@ -30,6 +30,12 @@ type PolicyManager interface {
 	// NodeCanHaveTag reports whether the given node can have the given tag.
 	NodeCanHaveTag(node types.NodeView, tag string) bool
 
+	// UserCanHaveTag reports whether the given user owns the given tag, i.e.
+	// is listed (directly or via a group) in the tag's tagOwners. This is the
+	// user half of NodeCanHaveTag, used to authorise re-auth tag changes
+	// against the authenticating user rather than the node's stale ownership.
+	UserCanHaveTag(user types.UserView, tag string) bool
+
 	// TagExists reports whether the given tag is defined in the policy.
 	TagExists(tag string) bool
 

hscontrol/policy/v2/policy.go

@@ -15,6 +15,7 @@ import (
 	"github.com/juanfont/headscale/hscontrol/policy/matcher"
 	"github.com/juanfont/headscale/hscontrol/policy/policyutil"
 	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/puzpuzpuz/xsync/v4"
 	"github.com/rs/zerolog/log"
 	"go4.org/netipx"
 	"tailscale.com/net/tsaddr"
@@ -28,7 +29,10 @@ import (
 var ErrInvalidTagOwner = errors.New("tag owner is not an Alias")
 
 type PolicyManager struct {
-	mu    sync.Mutex
+	// RWMutex, not Mutex, so concurrent map generation does not serialise on
+	// reads. The per-node caches are xsync.Maps so a read can fill them without
+	// taking the write lock.
+	mu    sync.RWMutex
 	pol   *Policy
 	users []types.User
 	nodes views.Slice[types.NodeView]
@@ -55,7 +59,7 @@ type PolicyManager struct {
 	viaTargetTags  map[Tag]struct{}
 
 	// Lazy map of SSH policies
-	sshPolicyMap map[types.NodeID]*tailcfg.SSHPolicy
+	sshPolicyMap *xsync.Map[types.NodeID, *tailcfg.SSHPolicy]
 
 	// compiledGrants are the grants with sources pre-resolved.
 	// The single source of truth for filter compilation. Both
@@ -64,12 +68,12 @@ type PolicyManager struct {
 	userNodeIdx    userNodeIndex
 
 	// Lazy map of per-node filter rules (reduced, for packet filters)
-	filterRulesMap map[types.NodeID][]tailcfg.FilterRule
+	filterRulesMap *xsync.Map[types.NodeID, []tailcfg.FilterRule]
 
 	// Lazy map of per-node matchers derived from UNREDUCED filter
 	// rules. Only populated on the slow path when needsPerNodeFilter
 	// is true; the fast path returns pm.matchers directly.
-	matchersForNodeMap map[types.NodeID][]matcher.Match
+	matchersForNodeMap *xsync.Map[types.NodeID, []matcher.Match]
 
 	// needsPerNodeFilter is true when any compiled grant requires
 	// per-node work (autogroup:self or via grants).
@@ -197,9 +201,9 @@ func NewPolicyManager(b []byte, users []types.User, nodes views.Slice[types.Node
 		pol:                policy,
 		users:              users,
 		nodes:              nodes,
-		sshPolicyMap:       make(map[types.NodeID]*tailcfg.SSHPolicy, nodes.Len()),
-		filterRulesMap:     make(map[types.NodeID][]tailcfg.FilterRule, nodes.Len()),
-		matchersForNodeMap: make(map[types.NodeID][]matcher.Match, nodes.Len()),
+		sshPolicyMap:       xsync.NewMap[types.NodeID, *tailcfg.SSHPolicy](),
+		filterRulesMap:     xsync.NewMap[types.NodeID, []tailcfg.FilterRule](),
+		matchersForNodeMap: xsync.NewMap[types.NodeID, []matcher.Match](),
 	}
 
 	_, err = pm.updateLocked()
@@ -354,9 +358,9 @@ func (pm *PolicyManager) updateLocked() (bool, error) {
 		// TODO(kradalby): This could potentially be optimized by only clearing the
 		// policies for nodes that have changed. Particularly if the only difference is
 		// that nodes has been added or removed.
-		clear(pm.sshPolicyMap)
-		clear(pm.filterRulesMap)
-		clear(pm.matchersForNodeMap)
+		pm.sshPolicyMap.Clear()
+		pm.filterRulesMap.Clear()
+		pm.matchersForNodeMap.Clear()
 	}
 
 	// If nothing changed, no need to update nodes
@@ -400,8 +404,8 @@ func (pm *PolicyManager) NodeNeedsPeerRecompute(node types.NodeView) bool {
 		return true
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	if pm.relayTargetIPs != nil && node.InIPSet(pm.relayTargetIPs) {
 		return true
@@ -422,10 +426,10 @@ func (pm *PolicyManager) NodeNeedsPeerRecompute(node types.NodeView) bool {
 // /machine/ssh/action/{src}/to/{dst}?local_user={local_user} per the
 // SaaS wire format. Cache is invalidated on policy reload.
 func (pm *PolicyManager) SSHPolicy(baseURL string, node types.NodeView) (*tailcfg.SSHPolicy, error) {
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
-	if sshPol, ok := pm.sshPolicyMap[node.ID()]; ok {
+	if sshPol, ok := pm.sshPolicyMap.Load(node.ID()); ok {
 		return sshPol, nil
 	}
 
@@ -434,7 +438,7 @@ func (pm *PolicyManager) SSHPolicy(baseURL string, node types.NodeView) (*tailcf
 		return nil, fmt.Errorf("compiling SSH policy: %w", err)
 	}
 
-	pm.sshPolicyMap[node.ID()] = sshPol
+	pm.sshPolicyMap.Store(node.ID(), sshPol)
 
 	return sshPol, nil
 }
@@ -450,8 +454,8 @@ func (pm *PolicyManager) SSHPolicy(baseURL string, node types.NodeView) (*tailcf
 func (pm *PolicyManager) SSHCheckParams(
 	srcNodeID, dstNodeID types.NodeID,
 ) (time.Duration, bool) {
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	if pm.pol == nil || len(pm.pol.SSHs) == 0 {
 		return 0, false
@@ -584,8 +588,8 @@ func (pm *PolicyManager) Filter() ([]tailcfg.FilterRule, []matcher.Match) {
 		return nil, nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	return pm.filter, pm.matchers
 }
@@ -604,8 +608,8 @@ func (pm *PolicyManager) BuildPeerMap(nodes views.Slice[types.NodeView]) map[typ
 		return nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// Precompute each node's subnet routes and exit-node status once; the
 	// O(n^2) pair scans below would otherwise recompute them for every pair.
@@ -726,7 +730,7 @@ func (pm *PolicyManager) filterForNodeLocked(
 		return nil
 	}
 
-	if rules, ok := pm.filterRulesMap[node.ID()]; ok {
+	if rules, ok := pm.filterRulesMap.Load(node.ID()); ok {
 		return rules
 	}
 
@@ -738,7 +742,7 @@ func (pm *PolicyManager) filterForNodeLocked(
 	}
 
 	reduced := policyutil.ReduceFilterRules(node, unreduced)
-	pm.filterRulesMap[node.ID()] = reduced
+	pm.filterRulesMap.Store(node.ID(), reduced)
 
 	return reduced
 }
@@ -755,8 +759,8 @@ func (pm *PolicyManager) FilterForNode(node types.NodeView) ([]tailcfg.FilterRul
 		return nil, nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	return pm.filterForNodeLocked(node), nil
 }
@@ -776,8 +780,8 @@ func (pm *PolicyManager) MatchersForNode(node types.NodeView) ([]matcher.Match,
 		return nil, nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// For global policies, return the shared global matchers.
 	// Via grants require per-node matchers because the global matchers
@@ -786,7 +790,7 @@ func (pm *PolicyManager) MatchersForNode(node types.NodeView) ([]matcher.Match,
 		return pm.matchers, nil
 	}
 
-	if cached, ok := pm.matchersForNodeMap[node.ID()]; ok {
+	if cached, ok := pm.matchersForNodeMap.Load(node.ID()); ok {
 		return cached, nil
 	}
 
@@ -794,7 +798,7 @@ func (pm *PolicyManager) MatchersForNode(node types.NodeView) ([]matcher.Match,
 	// the stored compiled grants for this specific node.
 	unreduced := pm.filterRulesForNodeLocked(node)
 	matchers := matcher.MatchesFromFilterRules(unreduced)
-	pm.matchersForNodeMap[node.ID()] = matchers
+	pm.matchersForNodeMap.Store(node.ID(), matchers)
 
 	return matchers, nil
 }
@@ -813,7 +817,7 @@ func (pm *PolicyManager) SetUsers(users []types.User) (bool, error) {
 	// Clear SSH policy map when users change to force SSH policy recomputation
 	// This ensures that if SSH policy compilation previously failed due to missing users,
 	// it will be retried with the new user list
-	clear(pm.sshPolicyMap)
+	pm.sshPolicyMap.Clear()
 
 	changed, err := pm.updateLocked()
 	if err != nil {
@@ -866,9 +870,9 @@ func (pm *PolicyManager) SetNodes(nodes views.Slice[types.NodeView]) (bool, erro
 
 		if !needsUpdate {
 			// This ensures fresh filter rules are generated for all nodes
-			clear(pm.sshPolicyMap)
-			clear(pm.filterRulesMap)
-			clear(pm.matchersForNodeMap)
+			pm.sshPolicyMap.Clear()
+			pm.filterRulesMap.Clear()
+			pm.matchersForNodeMap.Clear()
 		}
 		// Always return true when nodes changed, even if filter hash didn't change
 		// (can happen with autogroup:self or when nodes are added but don't affect rules)
@@ -921,8 +925,8 @@ func (pm *PolicyManager) NodeCanHaveTag(node types.NodeView, tag string) bool {
 		return false
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// pm.pol is written by SetPolicy under pm.mu; reading it before the
 	// lock races with concurrent policy reloads.
@@ -960,6 +964,37 @@ func (pm *PolicyManager) NodeCanHaveTag(node types.NodeView, tag string) bool {
 	return false
 }
 
+// UserCanHaveTag reports whether the given user is one of the tag's owners
+// (directly or via a group). It is the user half of [PolicyManager.NodeCanHaveTag]:
+// re-authentication authorises requested tags against the authenticating user,
+// because a tag-owned node carries no user and its IP is not in any owner set,
+// so only the user presenting the credential can prove ownership.
+func (pm *PolicyManager) UserCanHaveTag(user types.UserView, tag string) bool {
+	if pm == nil || !user.Valid() {
+		return false
+	}
+
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
+
+	if pm.pol == nil {
+		return false
+	}
+
+	owners, exists := pm.pol.TagOwners[Tag(tag)]
+	if !exists {
+		return false
+	}
+
+	for _, owner := range owners {
+		if pm.userMatchesOwner(user, owner) {
+			return true
+		}
+	}
+
+	return false
+}
+
 // userMatchesOwner checks if a user matches a tag owner entry.
 // This is used as a fallback when the node's IP is not in the [PolicyManager.tagOwnerMap].
 func (pm *PolicyManager) userMatchesOwner(user types.UserView, owner Owner) bool {
@@ -1010,8 +1045,8 @@ func (pm *PolicyManager) TagExists(tag string) bool {
 		return false
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// pm.pol is written by SetPolicy under pm.mu; reading it before the
 	// lock races with concurrent policy reloads.
@@ -1029,8 +1064,8 @@ func (pm *PolicyManager) NodeCanApproveRoute(node types.NodeView, route netip.Pr
 		return false
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// If the route to-be-approved is an exit route, then we need to check
 	// if the node is in allowed to approve it. This is treated differently
@@ -1092,8 +1127,8 @@ func (pm *PolicyManager) ViaRoutesForPeer(viewer, peer types.NodeView) types.Via
 		return result
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	// pm.pol is written by SetPolicy under pm.mu; reading it before the
 	// lock races with concurrent policy reloads.
@@ -1311,8 +1346,8 @@ func (pm *PolicyManager) DebugString() string {
 
 	// pm.pol, filter, matchers, and the derived maps are all written
 	// under pm.mu by SetPolicy/SetUsers/SetNodes.
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	var sb strings.Builder
 
@@ -1475,7 +1510,7 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 	// Clear cache entries for affected users only.
 	// For autogroup:self, we need to clear all nodes belonging to affected users
 	// because autogroup:self rules depend on the entire user's device set.
-	for nodeID := range pm.filterRulesMap {
+	pm.filterRulesMap.Range(func(nodeID types.NodeID, _ []tailcfg.FilterRule) bool {
 		// Find the user for this cached node
 		var nodeUserID types.UserID
 
@@ -1518,20 +1553,22 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 		// If we found the user and they're affected, clear this cache entry
 		if found {
 			if _, affected := affectedUsers[nodeUserID]; affected {
-				delete(pm.filterRulesMap, nodeID)
-				delete(pm.matchersForNodeMap, nodeID)
+				pm.filterRulesMap.Delete(nodeID)
+				pm.matchersForNodeMap.Delete(nodeID)
 			}
 		} else {
 			// Node not found in either old or new list, clear it
-			delete(pm.filterRulesMap, nodeID)
-			delete(pm.matchersForNodeMap, nodeID)
+			pm.filterRulesMap.Delete(nodeID)
+			pm.matchersForNodeMap.Delete(nodeID)
 		}
-	}
+
+		return true
+	})
 
 	if len(affectedUsers) > 0 {
 		log.Debug().
 			Int("affected_users", len(affectedUsers)).
-			Int("remaining_cache_entries", len(pm.filterRulesMap)).
+			Int("remaining_cache_entries", pm.filterRulesMap.Size()).
 			Msg("Selectively cleared autogroup:self cache for affected users")
 	}
 }
@@ -1572,23 +1609,27 @@ func (pm *PolicyManager) invalidateGlobalPolicyCache(newNodes views.Slice[types.
 		}
 
 		if newNode.HasNetworkChanges(oldNode) {
-			delete(pm.filterRulesMap, nodeID)
-			delete(pm.matchersForNodeMap, nodeID)
+			pm.filterRulesMap.Delete(nodeID)
+			pm.matchersForNodeMap.Delete(nodeID)
 		}
 	}
 
 	// Remove deleted nodes from cache
-	for nodeID := range pm.filterRulesMap {
+	pm.filterRulesMap.Range(func(nodeID types.NodeID, _ []tailcfg.FilterRule) bool {
 		if _, exists := newNodeMap[nodeID]; !exists {
-			delete(pm.filterRulesMap, nodeID)
+			pm.filterRulesMap.Delete(nodeID)
 		}
-	}
 
-	for nodeID := range pm.matchersForNodeMap {
+		return true
+	})
+
+	pm.matchersForNodeMap.Range(func(nodeID types.NodeID, _ []matcher.Match) bool {
 		if _, exists := newNodeMap[nodeID]; !exists {
-			delete(pm.matchersForNodeMap, nodeID)
+			pm.matchersForNodeMap.Delete(nodeID)
 		}
-	}
+
+		return true
+	})
 }
 
 // flattenTags resolves nested tag-owner references. Cycles
@@ -1770,8 +1811,8 @@ func (pm *PolicyManager) NodeCapMap(id types.NodeID) tailcfg.NodeCapMap {
 		return nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	src := pm.nodeAttrsMap[id]
 	if len(src) == 0 {
@@ -1794,8 +1835,8 @@ func (pm *PolicyManager) NodeCapMaps() map[types.NodeID]tailcfg.NodeCapMap {
 		return nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	out := make(map[types.NodeID]tailcfg.NodeCapMap, len(pm.nodeAttrsMap))
 	maps.Copy(out, pm.nodeAttrsMap)

hscontrol/policy/v2/policy_concurrency_test.go

@@ -0,0 +1,103 @@
+package v2
+
+import (
+	"fmt"
+	"sync"
+	"testing"
+
+	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"gorm.io/gorm"
+)
+
+// TestPolicyManagerConcurrentReads is the correctness guard for the #3346 fix:
+// PolicyManager read methods take a shared RLock and populate their per-node
+// caches (filterRulesMap, matchersForNodeMap) concurrently. This test hammers
+// those reads from many goroutines while a writer mutates the node set, so the
+// race detector catches any unsafe access to the shared caches or policy state.
+//
+// It uses an autogroup:self policy so reads take the per-node filter slow path
+// — the same path that made #3346's reconnect storm expensive — which is where
+// the lazy caches are written.
+func TestPolicyManagerConcurrentReads(t *testing.T) {
+	users := types.Users{
+		{Model: gorm.Model{ID: 1}, Name: "user1", Email: "user1@headscale.net"},
+		{Model: gorm.Model{ID: 2}, Name: "user2", Email: "user2@headscale.net"},
+		{Model: gorm.Model{ID: 3}, Name: "user3", Email: "user3@headscale.net"},
+	}
+
+	policy := `{
+		"acls": [
+			{
+				"action": "accept",
+				"src": ["autogroup:member"],
+				"dst": ["autogroup:self:*"]
+			}
+		]
+	}`
+
+	const nodeCount = 60
+
+	nodes := make(types.Nodes, 0, nodeCount)
+	for i := range nodeCount {
+		n := node(
+			fmt.Sprintf("node%d", i),
+			fmt.Sprintf("100.64.0.%d", i+1),
+			fmt.Sprintf("fd7a:115c:a1e0::%d", i+1),
+			users[i%len(users)],
+		)
+		n.ID = types.NodeID(i + 1) //nolint:gosec // safe in test
+		nodes = append(nodes, n)
+	}
+
+	pm, err := NewPolicyManager([]byte(policy), users, nodes.ViewSlice())
+	require.NoError(t, err)
+
+	const (
+		readers        = 16
+		iterations     = 60
+		mutatorReloads = 30
+	)
+
+	var wg sync.WaitGroup
+
+	// Concurrent readers exercise every converted RLock read path, including
+	// the two lazily populated per-node caches. Assertions inside the
+	// goroutines use assert (not require) so a failure does not call
+	// t.FailNow from a non-test goroutine.
+	for r := range readers {
+		wg.Go(func() {
+			for i := range iterations {
+				nv := nodes[(r+i)%len(nodes)].View()
+
+				rules, err := pm.FilterForNode(nv)
+				assert.NoError(t, err) //nolint:testifylint // assert (not require) is correct off the test goroutine
+				assert.NotNil(t, rules)
+
+				_, err = pm.MatchersForNode(nv)
+				assert.NoError(t, err) //nolint:testifylint // assert (not require) is correct off the test goroutine
+
+				pm.Filter()
+				pm.NodeCapMap(nv.ID())
+
+				// BuildPeerMap is the O(n^2) writer-side read; exercise it
+				// under RLock too, but not every iteration.
+				if i%8 == 0 {
+					assert.NotNil(t, pm.BuildPeerMap(nodes.ViewSlice()))
+				}
+			}
+		})
+	}
+
+	// A writer repeatedly re-sets the node set, invalidating and racing the
+	// caches the readers are populating.
+	wg.Go(func() {
+		for range mutatorReloads {
+			_, err := pm.SetNodes(nodes.ViewSlice())
+			assert.NoError(t, err) //nolint:testifylint // assert (not require) is correct off the test goroutine
+		}
+	})
+
+	wg.Wait()
+}

hscontrol/policy/v2/policy_test.go

@@ -8,6 +8,7 @@ import (
 	"github.com/google/go-cmp/cmp"
 	"github.com/juanfont/headscale/hscontrol/policy/matcher"
 	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/puzpuzpuz/xsync/v4"
 	"github.com/stretchr/testify/require"
 	"gorm.io/gorm"
 	"tailscale.com/net/tsaddr"
@@ -122,7 +123,7 @@ func TestInvalidateAutogroupSelfCache(t *testing.T) {
 		require.NoError(t, err)
 	}
 
-	require.Len(t, pm.filterRulesMap, len(initialNodes))
+	require.Equal(t, len(initialNodes), pm.filterRulesMap.Size())
 
 	tests := []struct {
 		name            string
@@ -207,19 +208,20 @@ func TestInvalidateAutogroupSelfCache(t *testing.T) {
 				}
 			}
 
-			pm.filterRulesMap = make(map[types.NodeID][]tailcfg.FilterRule)
+			pm.filterRulesMap.Clear()
+
 			for _, n := range initialNodes {
 				_, err := pm.FilterForNode(n.View())
 				require.NoError(t, err)
 			}
 
-			initialCacheSize := len(pm.filterRulesMap)
+			initialCacheSize := pm.filterRulesMap.Size()
 			require.Equal(t, len(initialNodes), initialCacheSize)
 
 			pm.invalidateAutogroupSelfCache(initialNodes.ViewSlice(), tt.newNodes.ViewSlice())
 
 			// Verify the expected number of cache entries were cleared
-			finalCacheSize := len(pm.filterRulesMap)
+			finalCacheSize := pm.filterRulesMap.Size()
 			clearedEntries := initialCacheSize - finalCacheSize
 			require.Equal(t, tt.expectedCleared, clearedEntries, tt.description)
 		})
@@ -498,15 +500,19 @@ func TestInvalidateGlobalPolicyCache(t *testing.T) {
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			pm := &PolicyManager{
-				nodes:          tt.oldNodes.ViewSlice(),
-				filterRulesMap: tt.initialCache,
+				nodes:              tt.oldNodes.ViewSlice(),
+				filterRulesMap:     xsync.NewMap[types.NodeID, []tailcfg.FilterRule](),
+				matchersForNodeMap: xsync.NewMap[types.NodeID, []matcher.Match](),
+			}
+			for id, rules := range tt.initialCache {
+				pm.filterRulesMap.Store(id, rules)
 			}
 
 			pm.invalidateGlobalPolicyCache(tt.newNodes.ViewSlice())
 
 			// Verify cache state
 			for nodeID, shouldExist := range tt.expectedCacheAfter {
-				_, exists := pm.filterRulesMap[nodeID]
+				_, exists := pm.filterRulesMap.Load(nodeID)
 				require.Equal(t, shouldExist, exists, "node %d cache existence mismatch", nodeID)
 			}
 		})

hscontrol/policy/v2/sshtest.go

@@ -128,8 +128,8 @@ func (pm *PolicyManager) RunSSHTests() error {
 		return nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	cache := make(map[types.NodeID]*tailcfg.SSHPolicy)
 	results := runSSHPolicyTests(pm.pol, pm.users, pm.nodes, cache)

hscontrol/policy/v2/test.go

@@ -204,8 +204,8 @@ func (pm *PolicyManager) RunTests() error {
 		return nil
 	}
 
-	pm.mu.Lock()
-	defer pm.mu.Unlock()
+	pm.mu.RLock()
+	defer pm.mu.RUnlock()
 
 	results := runPolicyTests(pm.pol, pm.filter, pm.users, pm.nodes)
 	if results.AllPassed {

hscontrol/poll.go

@@ -96,12 +96,6 @@ func (m *mapSession) stopFromBatcher() {
 	}
 }
 
-func (m *mapSession) beforeServeLongPoll() {
-	if m.node.IsEphemeral() {
-		m.h.ephemeralGC.Cancel(m.node.ID)
-	}
-}
-
 // afterServeLongPoll is called when a long-polling session ends and the node
 // is disconnected.
 func (m *mapSession) afterServeLongPoll() {
@@ -144,8 +138,6 @@ func (m *mapSession) serve() {
 //
 //nolint:gocyclo
 func (m *mapSession) serveLongPoll() {
-	m.beforeServeLongPoll()
-
 	m.log.Trace().Caller().Msg("long poll session started")
 
 	// connectGen is set by [state.State.Connect] below and captured by the deferred cleanup closure.
@@ -230,6 +222,11 @@ func (m *mapSession) serveLongPoll() {
 	mapReqChange, err := m.h.state.UpdateNodeFromMapRequest(m.node.ID, m.req)
 	if err != nil {
 		m.log.Error().Caller().Err(err).Msg("failed to update node from initial MapRequest")
+		// Write an explicit error rather than returning silently: a bare
+		// return leaves net/http to send an empty 200, which the client
+		// reads as "unexpected EOF" and retries forever (issue #3346).
+		httpError(m.w, err)
+
 		return
 	}
 
@@ -243,6 +240,13 @@ func (m *mapSession) serveLongPoll() {
 
 	connectChanges, connectGen = m.h.state.Connect(m.node.ID)
 
+	// Cancel ephemeral GC only after Connect succeeds. Cancelling at the start
+	// of serveLongPoll left departed nodes without a deletion timer when a
+	// reconnect attempt failed before Connect (issue #3382).
+	if m.node.IsEphemeral() {
+		m.h.ephemeralGC.Cancel(m.node.ID)
+	}
+
 	m.log.Info().Caller().Str(zf.Chan, fmt.Sprintf("%p", m.ch)).Msg("node has connected")
 
 	// TODO(kradalby): Redo the comments here
@@ -252,6 +256,11 @@ func (m *mapSession) serveLongPoll() {
 	// time between the node connecting and the batcher being ready.
 	if err := m.h.mapBatcher.AddNode(m.node.ID, m.ch, m.capVer, m.stopFromBatcher); err != nil { //nolint:noinlineerr
 		m.log.Error().Caller().Err(err).Msg("failed to add node to batcher")
+		// Write an explicit error rather than returning silently: a bare
+		// return leaves net/http to send an empty 200, which the client
+		// reads as "unexpected EOF" and retries forever (issue #3346).
+		httpError(m.w, err)
+
 		return
 	}
 

hscontrol/poll_test.go

@@ -7,12 +7,15 @@ import (
 	"testing"
 	"time"
 
+	"github.com/juanfont/headscale/hscontrol/db"
 	"github.com/juanfont/headscale/hscontrol/mapper"
 	"github.com/juanfont/headscale/hscontrol/state"
+	"github.com/juanfont/headscale/hscontrol/types"
 	"github.com/juanfont/headscale/hscontrol/types/change"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 	"tailscale.com/tailcfg"
+	"tailscale.com/types/key"
 )
 
 type delayedSuccessResponseWriter struct {
@@ -89,6 +92,192 @@ func (w *delayedSuccessResponseWriter) WriteCount() int {
 	return w.writeCount
 }
 
+// recordingResponseWriter records the status code and whether anything was
+// written, so a test can tell an explicit error response apart from a handler
+// that returned without writing (which net/http turns into an empty 200 the
+// client reads as "unexpected EOF").
+type recordingResponseWriter struct {
+	mu     sync.Mutex
+	header http.Header
+	status int
+	writes int
+}
+
+func (w *recordingResponseWriter) Header() http.Header {
+	if w.header == nil {
+		w.header = make(http.Header)
+	}
+
+	return w.header
+}
+
+func (w *recordingResponseWriter) WriteHeader(code int) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	if w.status == 0 {
+		w.status = code
+	}
+}
+
+func (w *recordingResponseWriter) Write(data []byte) (int, error) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	if w.status == 0 {
+		w.status = http.StatusOK
+	}
+
+	w.writes++
+
+	return len(data), nil
+}
+
+func (w *recordingResponseWriter) Flush() {}
+
+func (w *recordingResponseWriter) statusCode() int {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	return w.status
+}
+
+// TestServeLongPollWritesErrorWhenInitialMapFails proves that when the initial
+// map cannot be generated (here: the node's own GivenName is invalid, so
+// WithSelfNode fails and AddNode errors), serveLongPoll writes an explicit HTTP
+// error instead of returning with no body. Returning empty leaves net/http to
+// send an empty 200, which the Tailscale client reports as
+// "PollNetMap: ... unexpected EOF" and retries forever (issue #3346).
+func TestServeLongPollWritesErrorWhenInitialMapFails(t *testing.T) {
+	app := createTestApp(t)
+	user := app.state.CreateUserForTest("self-bad-name-user")
+	createdNode := app.state.CreateRegisteredNodeForTest(user, "self-bad-name-node")
+
+	// Corrupt the node's stored name to empty so GetFQDN fails for itself,
+	// then reload state so the bad row enters the NodeStore verbatim.
+	app.mapBatcher.Close()
+	require.NoError(t, app.state.Close())
+
+	database, err := db.NewHeadscaleDatabase(app.cfg)
+	require.NoError(t, err)
+	require.NoError(t, database.DB.
+		Model(&types.Node{}).
+		Where("id = ?", createdNode.ID).
+		Update("given_name", "").Error)
+	require.NoError(t, database.Close())
+
+	app.state, err = state.NewState(app.cfg)
+	require.NoError(t, err)
+
+	app.mapBatcher = mapper.NewBatcherAndMapper(app.cfg, app.state)
+	app.mapBatcher.Start()
+
+	t.Cleanup(func() {
+		app.mapBatcher.Close()
+		require.NoError(t, app.state.Close())
+	})
+
+	nodeView, ok := app.state.GetNodeByID(createdNode.ID)
+	require.True(t, ok)
+
+	node := nodeView.AsStruct()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	writer := &recordingResponseWriter{}
+	session := app.newMapSession(ctx, tailcfg.MapRequest{
+		Stream:  true,
+		Version: tailcfg.CapabilityVersion(100),
+	}, writer, node)
+
+	serveDone := make(chan struct{})
+
+	go func() {
+		session.serveLongPoll()
+		close(serveDone)
+	}()
+
+	t.Cleanup(func() {
+		// Break the post-disconnect reconnect wait so the goroutine exits.
+		dummyCh := make(chan *tailcfg.MapResponse, 1)
+		_ = app.mapBatcher.AddNode(node.ID, dummyCh, tailcfg.CapabilityVersion(100), nil)
+
+		cancel()
+
+		select {
+		case <-serveDone:
+		case <-time.After(2 * time.Second):
+		}
+
+		_ = app.mapBatcher.RemoveNode(node.ID, dummyCh)
+	})
+
+	assert.Eventually(t, func() bool {
+		return writer.statusCode() >= http.StatusInternalServerError
+	}, 2*time.Second, 10*time.Millisecond,
+		"serveLongPoll must write an HTTP error response when the initial map cannot be built, not an empty 200")
+}
+
+// TestFailedReconnectDoesNotCancelEphemeralGC proves that a
+// long-poll reconnect attempt which fails before [state.State.Connect] must
+// not cancel a previously armed ephemeral GC timer. Cancelling at the start of
+// [mapSession.serveLongPoll] left departed ephemeral nodes stuck offline with
+// no deletion scheduled (https://github.com/juanfont/headscale/issues/3382).
+func TestFailedReconnectDoesNotCancelEphemeralGC(t *testing.T) {
+	t.Parallel()
+
+	app := createTestApp(t)
+	app.StartEphemeralGCForTest(t)
+
+	user := app.state.CreateUserForTest("eph-gc-cancel-user")
+	pak, err := app.state.CreatePreAuthKey(user.TypedID(), false, true, nil, nil)
+	require.NoError(t, err)
+
+	machineKey := key.NewMachine()
+	nodeKey := key.NewNode()
+
+	_, err = app.handleRegister(context.Background(), tailcfg.RegisterRequest{
+		Auth: &tailcfg.RegisterResponseAuth{
+			AuthKey: pak.Key,
+		},
+		NodeKey: nodeKey.Public(),
+		Hostinfo: &tailcfg.Hostinfo{
+			Hostname: "eph-gc-cancel-node",
+		},
+		Expiry: time.Now().Add(24 * time.Hour),
+	}, machineKey.Public())
+	require.NoError(t, err)
+
+	nodeView, ok := app.state.GetNodeByNodeKey(nodeKey.Public())
+	require.True(t, ok)
+	require.True(t, nodeView.IsEphemeral(), "node must be ephemeral so Cancel would arm on long-poll")
+
+	node := nodeView.AsStruct()
+
+	// Arm a long-lived deletion timer — the state after a normal disconnect
+	// has called afterServeLongPoll. A long expiry avoids racing the
+	// fail-before-Connect path below.
+	app.ephemeralGC.Schedule(node.ID, time.Hour)
+	require.True(t, app.ephemeralGC.IsScheduled(node.ID), "test sanity: GC timer must be armed")
+
+	// Drop the node from the NodeStore so UpdateNodeFromMapRequest fails before
+	// Connect, while the session still carries an ephemeral AuthKey (so the
+	// old Cancel-on-entry path would clear the timer).
+	app.state.DeleteNodeFromStoreForTest(node.ID)
+
+	writer := &recordingResponseWriter{}
+	session := app.newMapSession(context.Background(), tailcfg.MapRequest{
+		Stream:  true,
+		Version: tailcfg.CapabilityVersion(100),
+	}, writer, node)
+
+	session.serveLongPoll()
+
+	assert.GreaterOrEqual(t, writer.statusCode(), http.StatusInternalServerError,
+		"failed reconnect must write an HTTP error before Connect")
+	assert.True(t, app.ephemeralGC.IsScheduled(node.ID),
+		"failed reconnect must not cancel the ephemeral GC timer (issue #3382)")
+}
+
 // TestGitHubIssue3129_TransientlyBlockedWriteDoesNotLeaveLiveStaleSession
 // tests the scenario reported in
 // https://github.com/juanfont/headscale/issues/3129.

hscontrol/state/node_health.go

@@ -0,0 +1,92 @@
+package state
+
+import (
+	"fmt"
+
+	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/juanfont/headscale/hscontrol/util/zlog/zf"
+	"github.com/rs/zerolog/log"
+)
+
+// nodeHealthCheck names a class of stored-node-data defect that breaks normal
+// operation and explains how to fix it. ok == true means the node passes the
+// check. This is the extension point for node-data validation: add a check
+// here as new corrupt-data classes surface (nil hostinfo, invalid IPs,
+// tags-XOR-user violations, ...) and both the boot scan and any future caller
+// run the whole set.
+type nodeHealthCheck struct {
+	name  string
+	check func(nv types.NodeView, cfg *types.Config) (problem, fixHint string, ok bool)
+}
+
+// nodeHealthChecks is the registry of node-data health checks. Today it carries
+// the one issue #3346 needs; append to it rather than reshaping callers.
+var nodeHealthChecks = []nodeHealthCheck{givenNameMapsToValidFQDN}
+
+// givenNameMapsToValidFQDN flags a node whose stored GivenName cannot produce a
+// valid FQDN (empty, or longer than MaxHostnameLength once base_domain is
+// applied). Such a node cannot be rendered into a netmap — neither its own nor
+// any peer's — so it must be renamed to recover.
+var givenNameMapsToValidFQDN = nodeHealthCheck{
+	name: "given-name-maps-to-valid-fqdn",
+	check: func(nv types.NodeView, cfg *types.Config) (string, string, bool) {
+		err := types.ValidateGivenName(nv.GivenName(), cfg.BaseDomain)
+		if err != nil {
+			return err.Error(), fmt.Sprintf("headscale nodes rename %d <name>", nv.ID()), false
+		}
+
+		return "", "", true
+	},
+}
+
+// nodeHealthFinding is a single failed check for a single node.
+type nodeHealthFinding struct {
+	nodeID   types.NodeID
+	hostname string
+	check    string
+	problem  string
+	fixHint  string
+}
+
+// scanNodeHealth runs every registered check against every node in the store
+// and returns one finding per failure. It only reports — it never mutates a
+// node — so an operator can repair the underlying data without the server
+// silently rewriting a user-visible name.
+func (s *State) scanNodeHealth() []nodeHealthFinding {
+	var findings []nodeHealthFinding
+
+	for _, nv := range s.nodeStore.ListNodes().All() {
+		for _, c := range nodeHealthChecks {
+			problem, fixHint, ok := c.check(nv, s.cfg)
+			if ok {
+				continue
+			}
+
+			findings = append(findings, nodeHealthFinding{
+				nodeID:   nv.ID(),
+				hostname: nv.Hostname(),
+				check:    c.name,
+				problem:  problem,
+				fixHint:  fixHint,
+			})
+		}
+	}
+
+	return findings
+}
+
+// logNodeHealth scans the store once and logs an actionable warning per
+// finding. Called at startup so an operator learns — by node id and fix
+// command — about stored data that will break map generation, without the
+// server changing anything itself.
+func (s *State) logNodeHealth() {
+	for _, f := range s.scanNodeHealth() {
+		log.Warn().
+			Uint64(zf.NodeID, f.nodeID.Uint64()).
+			Str(zf.NodeHostname, f.hostname).
+			Str("check", f.check).
+			Str("problem", f.problem).
+			Str("fix", f.fixHint).
+			Msg("node has invalid data that breaks map generation; rename it to restore connectivity")
+	}
+}

hscontrol/state/node_health_test.go

@@ -0,0 +1,67 @@
+package state
+
+import (
+	"testing"
+
+	"github.com/juanfont/headscale/hscontrol/db"
+	"github.com/juanfont/headscale/hscontrol/types"
+	"github.com/stretchr/testify/require"
+)
+
+func TestGivenNameMapsToValidFQDNCheck(t *testing.T) {
+	cfg := &types.Config{BaseDomain: "example.com"}
+
+	_, _, ok := givenNameMapsToValidFQDN.check((&types.Node{ID: 1, GivenName: "valid"}).View(), cfg)
+	require.True(t, ok, "a valid given name must pass the check")
+
+	problem, fixHint, ok := givenNameMapsToValidFQDN.check((&types.Node{ID: 7, GivenName: ""}).View(), cfg)
+	require.False(t, ok, "an empty given name must fail the check")
+	require.NotEmpty(t, problem)
+	require.Contains(t, fixHint, "rename 7", "fix hint must name the offending node")
+}
+
+// TestScanNodeHealthReportsInvalidNameWithoutMutating proves the boot scan
+// reports a node whose stored name would break map generation (issue #3346)
+// with an actionable fix, and that it never rewrites the stored name — the
+// maintainer's decision is log-only, no silent mutation.
+func TestScanNodeHealthReportsInvalidNameWithoutMutating(t *testing.T) {
+	dbPath := t.TempDir() + "/headscale.db"
+	cfg := persistTestConfig(dbPath)
+
+	database, err := db.NewHeadscaleDatabase(cfg)
+	require.NoError(t, err)
+
+	user := database.CreateUserForTest("scan-user")
+	bad := database.CreateRegisteredNodeForTest(user, "scan-bad")
+	good := database.CreateRegisteredNodeForTest(user, "scan-good")
+
+	require.NoError(t, database.DB.
+		Model(&types.Node{}).
+		Where("id = ?", bad.ID).
+		Update("given_name", "").Error)
+	require.NoError(t, database.Close())
+
+	s, err := NewState(cfg)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = s.Close() })
+
+	findings := s.scanNodeHealth()
+
+	var badFinding *nodeHealthFinding
+
+	for i := range findings {
+		require.NotEqual(t, good.ID, findings[i].nodeID, "a valid node must not be reported")
+
+		if findings[i].nodeID == bad.ID {
+			badFinding = &findings[i]
+		}
+	}
+
+	require.NotNil(t, badFinding, "a node with an invalid name must be reported")
+	require.Contains(t, badFinding.fixHint, "rename", "finding must carry an actionable fix")
+
+	// Log-only: neither the scan nor boot may rewrite the stored name.
+	nv, ok := s.GetNodeByID(bad.ID)
+	require.True(t, ok)
+	require.Empty(t, nv.GivenName(), "boot scan must not mutate the stored name")
+}

hscontrol/state/rename_test.go

@@ -0,0 +1,40 @@
+package state
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/juanfont/headscale/hscontrol/db"
+	"github.com/stretchr/testify/require"
+)
+
+// TestRenameNodeRejectsNameExceedingFQDNLimit proves RenameNode rejects a name
+// that is a valid DNS label but whose FQDN, under the configured base_domain,
+// exceeds MaxHostnameLength. Without the FQDN-length gate such a name persists
+// and then breaks map generation for the node and its peers (issue #3346):
+// admin-facing writes must not be able to introduce an unmappable name.
+func TestRenameNodeRejectsNameExceedingFQDNLimit(t *testing.T) {
+	dbPath := t.TempDir() + "/headscale.db"
+	cfg := persistTestConfig(dbPath)
+	// A long base domain so a 63-char label overflows the 255-char FQDN bound.
+	cfg.BaseDomain = strings.Repeat("b", 200) + ".example.com"
+
+	database, err := db.NewHeadscaleDatabase(cfg)
+	require.NoError(t, err)
+
+	user := database.CreateUserForTest("rename-user")
+	node := database.CreateRegisteredNodeForTest(user, "rename-node")
+	require.NoError(t, database.Close())
+
+	s, err := NewState(cfg)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = s.Close() })
+
+	// Valid 63-char DNS label, but the resulting FQDN exceeds 255 chars.
+	_, _, err = s.RenameNode(node.ID, strings.Repeat("a", 63))
+	require.Error(t, err, "rename to a name whose FQDN exceeds the limit must be rejected")
+
+	// A short, valid name is still accepted.
+	_, _, err = s.RenameNode(node.ID, "short")
+	require.NoError(t, err)
+}

hscontrol/state/state.go

@@ -286,7 +286,7 @@ func NewState(cfg *types.Config) (*State, error) {
 	)
 	nodeStore.Start()
 
-	return &State{
+	s := &State{
 		cfg: cfg,
 
 		db:        db,
@@ -298,7 +298,14 @@ func NewState(cfg *types.Config) (*State, error) {
 
 		sshCheckAuth:  make(map[sshCheckPair]time.Time),
 		registerLocks: xsync.NewMap[key.MachinePublic, *sync.Mutex](),
-	}, nil
+	}
+
+	// Surface nodes whose stored data would break map generation (e.g. an
+	// invalid given name from a legacy row) so an operator can fix them. This
+	// only logs; it never mutates a node's stored name at boot.
+	s.logNodeHealth()
+
+	return s, nil
 }
 
 // Close gracefully shuts down the [State] instance and releases all resources.
@@ -1031,7 +1038,10 @@ func (s *State) SetApprovedRoutes(nodeID types.NodeID, routes []netip.Prefix) (t
 // auto-sanitisation) and collisions error out rather than silently
 // bumping a user-facing label. See HOSTNAME.md for the CLI contract.
 func (s *State) RenameNode(nodeID types.NodeID, newName string) (types.NodeView, change.Change, error) {
-	err := dnsname.ValidLabel(newName)
+	// Validate the label AND that the resulting FQDN fits MaxHostnameLength:
+	// a valid 63-char label can still overflow under a long base_domain, and
+	// an unmappable name would break this node and its peers (issue #3346).
+	err := types.ValidateGivenName(newName, s.cfg.BaseDomain)
 	if err != nil {
 		return types.NodeView{}, change.Change{}, fmt.Errorf("renaming node: %w", err)
 	}
@@ -1436,6 +1446,13 @@ func (s *State) PutNodeInStoreForTest(node types.Node) types.NodeView {
 	return s.nodeStore.PutNode(node)
 }
 
+// DeleteNodeFromStoreForTest removes a node from the in-memory [NodeStore]
+// without touching the database. Used to force [State.UpdateNodeFromMapRequest]
+// failures in poll-session tests while keeping the DB row intact for later restore.
+func (s *State) DeleteNodeFromStoreForTest(id types.NodeID) {
+	s.nodeStore.DeleteNode(id)
+}
+
 // CreateRegisteredNodeForTest creates a test node with allocated IPs. This is a convenience wrapper around the database layer.
 func (s *State) CreateRegisteredNodeForTest(user *types.User, hostname ...string) *types.Node {
 	return s.db.CreateRegisteredNodeForTest(user, hostname...)
@@ -1605,7 +1622,17 @@ func (s *State) applyAuthNodeUpdate(params authNodeUpdateParams) (types.NodeView
 	// Validate tags BEFORE calling [NodeStore.UpdateNode] to ensure we don't modify
 	// [NodeStore] if validation fails. This maintains consistency between [NodeStore]
 	// and database.
-	rejectedTags := s.validateRequestTags(params.ExistingNode, requestTags)
+	//
+	// A tag-owned node carries no user and its IP is not in any tag owner's set,
+	// so checking the node alone rejects every tag on re-auth (#3374). Authorise
+	// against the authenticating user too: they are the one presenting the
+	// credential and may own the requested tags.
+	var authUser types.UserView
+	if params.User != nil {
+		authUser = params.User.View()
+	}
+
+	rejectedTags := s.validateRequestTagsForReauth(params.ExistingNode, authUser, requestTags)
 	if len(rejectedTags) > 0 {
 		return types.NodeView{}, fmt.Errorf(
 			"%w %v are invalid or not permitted",
@@ -1699,8 +1726,14 @@ func (s *State) applyAuthNodeUpdate(params authNodeUpdateParams) (types.NodeView
 			} else {
 				node.Expiry = regData.Expiry
 			}
+		case isTagged && node.IsExpired():
+			// Tagged → Tagged, but carrying a stale PAST expiry from an older
+			// headscale's logout stamp (#3371). Tagged nodes never expire, so
+			// clear it; a deliberate future expiry has IsExpired() == false and
+			// falls through to the no-op below.
+			node.Expiry = nil
 		}
-		// Tagged → Tagged: keep existing expiry (nil) - no action needed
+		// Tagged → Tagged with no stale expiry: keep existing expiry - no action.
 
 		// Apply default node expiry for non-tagged nodes when the
 		// resolved expiry is still nil or zero (e.g., CLI registration
@@ -1721,8 +1754,21 @@ func (s *State) applyAuthNodeUpdate(params authNodeUpdateParams) (types.NodeView
 	// Persist to database.
 	// Explicitly select all node columns so GORM includes nil/zero-value fields
 	// (see nodeUpdateColumns comment).
+	//
+	// AuthKeyID is excluded from nodeUpdateColumns (#2862: never persist a
+	// possibly-deleted key's stale reference on the shared update path). But
+	// when a re-auth untags a node it clears AuthKeyID to nil (above), and that
+	// must persist or the node reloads as tagged/ephemeral after a restart and
+	// is garbage-collected. Writing NULL can never cause an FK error, so include
+	// the column only in that clearing case; the other transitions keep the
+	// #2862-safe column set untouched.
+	updateColumns := nodeUpdateColumns
+	if !updatedNodeView.AuthKeyID().Valid() {
+		updateColumns = append(slices.Clone(nodeUpdateColumns), "AuthKeyID")
+	}
+
 	_, err := hsdb.Write(s.db.DB, func(tx *gorm.DB) (*types.Node, error) {
-		err := tx.Select(nodeUpdateColumns).Updates(updatedNodeView.AsStruct()).Error
+		err := tx.Select(updateColumns).Updates(updatedNodeView.AsStruct()).Error
 		if err != nil {
 			return nil, fmt.Errorf("saving node: %w", err)
 		}
@@ -1926,6 +1972,34 @@ func (s *State) validateRequestTags(node types.NodeView, requestTags []string) [
 	return rejectedTags
 }
 
+// validateRequestTagsForReauth authorises re-auth request tags against the
+// existing node OR the authenticating user. A tag-owned node (#3374) has no
+// user and its IP is in no owner set, so NodeCanHaveTag alone rejects every
+// tag; the authenticating user who owns the tags must also be consulted.
+// Tags neither the node nor the user owns are still rejected, so this
+// authorises the user, it does not skip authorisation.
+func (s *State) validateRequestTagsForReauth(node types.NodeView, authUser types.UserView, requestTags []string) []string {
+	if len(requestTags) == 0 {
+		return nil
+	}
+
+	var rejectedTags []string
+
+	for _, tag := range requestTags {
+		if s.polMan.NodeCanHaveTag(node, tag) {
+			continue
+		}
+
+		if authUser.Valid() && s.polMan.UserCanHaveTag(authUser, tag) {
+			continue
+		}
+
+		rejectedTags = append(rejectedTags, tag)
+	}
+
+	return rejectedTags
+}
+
 // processReauthTags handles tag changes during node re-authentication.
 // It processes RequestTags from the client and updates node tags accordingly.
 // Returns rejected tags (if any) for post-validation error handling.
@@ -1960,16 +2034,34 @@ func (s *State) processReauthTags(
 			node.Tags = []string{}
 			node.UserID = &user.ID
 			node.User = user
+
+			// The node is no longer tagged, so it must not keep a reference to
+			// the tagged auth key. Leaving AuthKey set means a node created by a
+			// tagged+ephemeral key stays IsEphemeral() after converting to
+			// user-owned and is garbage-collected on its next disconnect,
+			// silently deleting the user's just-claimed device. Clearing the
+			// reference is persisted via the AuthKeyID column added to this
+			// path's write below.
+			node.AuthKey = nil
+			node.AuthKeyID = nil
 		}
 
 		return nil
 	}
 
-	// Non-empty RequestTags: validate and apply
+	// Non-empty RequestTags: validate and apply. Authorise each tag against the
+	// node OR the authenticating user, matching the pre-check in
+	// validateRequestTagsForReauth. Without the user half, a tag-owned node
+	// (#3374) has every tag rejected here even after the pre-check passed, so
+	// this returns the tags as rejected and the re-advertised tag is silently
+	// dropped despite a success response.
+	authUser := user.View()
+
 	var approvedTags, rejectedTags []string
 
 	for _, tag := range requestTags {
-		if s.polMan.NodeCanHaveTag(node.View(), tag) {
+		if s.polMan.NodeCanHaveTag(node.View(), tag) ||
+			(authUser.Valid() && s.polMan.UserCanHaveTag(authUser, tag)) {
 			approvedTags = append(approvedTags, tag)
 		} else {
 			rejectedTags = append(rejectedTags, tag)
@@ -2319,7 +2411,14 @@ func (s *State) HandleNodeFromPreAuthKey(
 	// must present a valid key. Without this a node that re-uses its NodeKey
 	// after expiry would skip validation and be re-authorised with a spent or
 	// expired key; the boundary must not depend on the client rotating its key.
+	//
+	// Tagged nodes are excluded: they never expire (KB 1068), so an
+	// IsExpired() tagged node only reflects a stale logout stamp left by an
+	// older headscale (#3371). Forcing it down the re-validation path burns its
+	// fresh key and blocks re-auth forever; treat it as a plain re-registration
+	// and clear the stale expiry in the update below.
 	isExpired := existsSameUser && existingNodeSameUser.Valid() &&
+		!existingNodeSameUser.IsTagged() &&
 		existingNodeSameUser.IsExpired()
 
 	// A tagged key presented for a currently user-owned node converts that node
@@ -2328,7 +2427,17 @@ func (s *State) HandleNodeFromPreAuthKey(
 	isOwnershipConversion := existsSameUser && existingNodeSameUser.Valid() &&
 		pak.IsTagged() && !existingNodeSameUser.IsTagged()
 
-	if isExistingNodeReregistering && !isNodeKeyRotation && !isExpired && !isOwnershipConversion {
+	// A tagged key that differs from the one the node last authed with retags
+	// the node (see the in-place update below). Applying a key's tags is an
+	// authorisation decision, so the key must be validated rather than ride the
+	// skip-validation fast-path; otherwise a spent, revoked or expired tagged
+	// key could still retag a node that reuses its node key. Like isExpired,
+	// this boundary must not depend on the client rotating its key.
+	isRetag := existsSameUser && existingNodeSameUser.Valid() &&
+		pak.IsTagged() && existingNodeSameUser.IsTagged() &&
+		(!existingNodeSameUser.AuthKeyID().Valid() || existingNodeSameUser.AuthKeyID().Get() != pak.ID)
+
+	if isExistingNodeReregistering && !isNodeKeyRotation && !isExpired && !isOwnershipConversion && !isRetag {
 		// Existing, still-valid node re-registering with same NodeKey: skip
 		// validation. Pre-auth keys are only needed for initial authentication.
 		// Critical for containers that run "tailscale up --authkey=KEY" on every
@@ -2376,8 +2485,10 @@ func (s *State) HandleNodeFromPreAuthKey(
 
 	var finalNode types.NodeView
 
-	// If this node exists for this user, update the node in place.
-	// Note: For tags-only keys (pak.User == nil), existsSameUser is always false.
+	// If this node exists for this user, update the node in place. For a
+	// tags-only key (pak.User == nil) this is true when the machine already has
+	// a tagged node (findExistingNodeForPAK matches it under UserID 0); for a
+	// user-owned key it is true when the same user already has the node.
 	if existsSameUser && existingNodeSameUser.Valid() {
 		log.Trace().
 			Caller().
@@ -2419,17 +2530,37 @@ func (s *State) HandleNodeFromPreAuthKey(
 
 			node.RegisterMethod = util.RegisterMethodAuthKey
 
-			// Tags from PreAuthKey are only applied during initial registration.
-			// On re-registration the node keeps its existing tags and ownership,
-			// except when a tagged key converts a user-owned node: that adopts
-			// the key's tags and drops user ownership (tagged nodes are
-			// user-less and never expire). Only update AuthKey reference
-			// otherwise.
-			if pak.IsTagged() && !node.IsTagged() {
-				node.Tags = pak.Proto().GetAclTags()
+			// Tags from a PreAuthKey are applied on initial registration and
+			// re-applied whenever a *different* key is presented on
+			// re-registration: re-keying is Tailscale's documented way to change
+			// an auth-key device's tags (KB 1068 - "generate a new auth key with
+			// the new set of tags ... doing so replaces the device's existing
+			// tags"). Presenting the SAME key again (container restart,
+			// #2830/#3312) preserves the node's current tags and any admin
+			// override. A tagged key presented for a user-owned node also converts
+			// it, dropping user ownership.
+			//
+			// node.AuthKeyID still holds the prior key's ID here (it is reassigned
+			// below), and SetNodeTags leaves AuthKeyID intact, so an admin retag
+			// cannot masquerade as a new key.
+			keyChanged := node.AuthKeyID == nil || *node.AuthKeyID != pak.ID
+			if pak.IsTagged() && (!node.IsTagged() || keyChanged) {
+				wasUserOwned := !node.IsTagged()
+
+				node.Tags = pak.Tags
 				node.UserID = nil
 				node.User = nil
-				node.Expiry = nil
+
+				// Converting a user-owned node to tagged drops the user's key
+				// expiry (tagged nodes never expire). But retagging an
+				// already-tagged node must preserve a deliberate FUTURE expiry
+				// set via `headscale nodes expire` - that is a node property, not
+				// tied to the auth key - and only clear a stale PAST expiry. This
+				// keeps the retag path symmetric with the same-key relogin path
+				// (#3371) rather than silently overriding an admin decision.
+				if wasUserOwned || node.IsExpired() {
+					node.Expiry = nil
+				}
 			}
 			node.AuthKey = pak
 			node.AuthKeyID = &pak.ID
@@ -2454,6 +2585,13 @@ func (s *State) HandleNodeFromPreAuthKey(
 				} else {
 					node.Expiry = nil
 				}
+			} else if node.IsExpired() {
+				// #3371: a tagged node must never carry key expiry. Clear a
+				// stale PAST expiry left by a logout (older headscale) so
+				// re-auth is not permanently blocked. A deliberate future
+				// expiry (headscale nodes expire) has IsExpired() == false and
+				// is left untouched.
+				node.Expiry = nil
 			}
 		})
 
@@ -2463,8 +2601,14 @@ func (s *State) HandleNodeFromPreAuthKey(
 
 		_, err = hsdb.Write(s.db.DB, func(tx *gorm.DB) (*types.Node, error) {
 			// Explicitly select all node columns so GORM includes nil/zero-value fields
-			// (see nodeUpdateColumns comment).
-			err := tx.Select(nodeUpdateColumns).Updates(updatedNodeView.AsStruct()).Error
+			// (see nodeUpdateColumns comment). AuthKeyID is normally excluded to
+			// avoid persisting a deleted key's stale reference on MapRequest
+			// (#2862), but re-registration presents a freshly-validated key, so
+			// its ID must be persisted here — otherwise a restart reloads the old
+			// key and any key-scoped property (e.g. Ephemeral) silently reverts.
+			reregColumns := append(slices.Clone(nodeUpdateColumns), "AuthKeyID")
+
+			err := tx.Select(reregColumns).Updates(updatedNodeView.AsStruct()).Error
 			if err != nil {
 				return nil, fmt.Errorf("saving node: %w", err)
 			}

hscontrol/types/node.go

@@ -21,6 +21,7 @@ import (
 	"tailscale.com/tailcfg"
 	"tailscale.com/types/key"
 	"tailscale.com/types/views"
+	"tailscale.com/util/dnsname"
 )
 
 var (
@@ -568,6 +569,28 @@ func (node *Node) GetFQDN(baseDomain string) (string, error) {
 	return hostname, nil
 }
 
+// ValidateGivenName reports whether givenName is usable as a node's DNS label:
+// a valid DNS label that, combined with baseDomain, yields an FQDN within
+// MaxHostnameLength. Admin-facing write paths (e.g. node rename) reject names
+// that fail this, since the mapper cannot build a map for a node — or any of
+// its peers — whose GetFQDN fails. Derived paths sanitise/coerce instead.
+func ValidateGivenName(givenName, baseDomain string) error {
+	err := dnsname.ValidLabel(givenName)
+	if err != nil {
+		return fmt.Errorf("%q is not a valid DNS label: %w", givenName, err)
+	}
+
+	// Reuse GetFQDN so the length bound stays identical to what the mapper
+	// enforces; a valid 63-char label can still overflow under a long
+	// base_domain.
+	_, err = (&Node{GivenName: givenName}).GetFQDN(baseDomain)
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
 // AnnouncedRoutes returns the list of routes the node announces, as
 // reported by the client in [tailcfg.Hostinfo.RoutableIPs]. Announcement alone
 // does not grant visibility — see [Node.SubnetRoutes] for approval-gated

hscontrol/types/node_test.go

@@ -418,6 +418,33 @@ func TestNodeFQDN(t *testing.T) {
 	}
 }
 
+func TestValidateGivenName(t *testing.T) {
+	tests := []struct {
+		name       string
+		givenName  string
+		baseDomain string
+		wantErr    bool
+	}{
+		{"valid", "test", "example.com", false},
+		{"empty", "", "example.com", true},
+		{"invalid label chars", "not valid", "example.com", true},
+		{"label too long", strings.Repeat("a", 64), "example.com", true},
+		// A valid 63-char label whose FQDN overflows only because the base
+		// domain is long: ValidLabel passes, the FQDN-length bound rejects it.
+		{"fqdn too long under long base domain", strings.Repeat("a", 63), strings.Repeat("b", 200) + ".example.com", true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			err := ValidateGivenName(tc.givenName, tc.baseDomain)
+			if (err != nil) != tc.wantErr {
+				t.Errorf("ValidateGivenName(%q, %q) error = %v, wantErr %v",
+					tc.givenName, tc.baseDomain, err, tc.wantErr)
+			}
+		})
+	}
+}
+
 func TestPeerChangeFromMapRequest(t *testing.T) {
 	nKeys := []key.NodePublic{
 		key.NewNode().Public(),

integration/tags_test.go

@@ -11,6 +11,7 @@ import (
 	"github.com/juanfont/headscale/integration/hsic"
 	"github.com/juanfont/headscale/integration/integrationutil"
 	"github.com/juanfont/headscale/integration/tsic"
+	"github.com/oauth2-proxy/mockoidc"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 	"tailscale.com/tailcfg"
@@ -579,6 +580,213 @@ func TestTagsAuthKeyWithTagAdminOverrideReauthPreserves(t *testing.T) {
 	t.Logf("Test 2.5 PASS: Admin tags preserved through reauth (admin decisions are authoritative)")
 }
 
+// TestTagsReauthDifferentKeyRetagsNode reproduces issue #3370 end-to-end with a
+// real tailscale client: a node registered with a single-use tag:valid-owned
+// key is re-authenticated via `tailscale up --force-reauth` with a *fresh*
+// single-use tag:second key. Tailscale's documented behaviour (KB 1068) is that
+// re-keying replaces the device's tags, verified by the reporter against SaaS on
+// the same node/IP. Before the fix headscale consumes the new key but keeps the
+// old tag; after the fix the node retags in place.
+//
+// Unlike Test 2.5 (same reusable key + admin override -> tags preserved), this
+// presents a *different* key, so it exercises the opposite arm of the retag
+// discriminator. It also asserts what a state-unit test cannot: the new tag
+// propagates to the node's own Self view and netmap, and the node ID and IPs are
+// unchanged.
+//
+// https://github.com/juanfont/headscale/issues/3370
+func TestTagsReauthDifferentKeyRetagsNode(t *testing.T) {
+	IntegrationSkip(t)
+
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		Users:        []string{tagTestUser},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnv(
+		[]tsic.Option{},
+		hsic.WithACLPolicy(tagsTestPolicy()),
+		hsic.WithTestName("tags-rekey-retag"),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	userMap, err := headscale.MapUsers()
+	require.NoError(t, err)
+
+	userID := userMap[tagTestUser].GetId()
+
+	// KEY1: single-use tag:valid-owned.
+	key1, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"head",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+	)
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key1.GetKey())
+	require.NoError(t, err)
+
+	var (
+		initialNodeID uint64
+		initialIPs    []string
+	)
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			initialIPs = nodes[0].GetIpAddresses()
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial registration")
+
+	t.Logf("Step 1: node %d registered with tag:valid-owned, IPs %v", initialNodeID, initialIPs)
+
+	// KEY2: fresh single-use tag:second. Re-key via --force-reauth.
+	key2, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:second"})
+	require.NoError(t, err)
+
+	//nolint:errcheck // result is verified via EventuallyWithT below
+	client.Execute([]string{
+		"tailscale", "up",
+		"--login-server=" + headscale.GetEndpoint(),
+		"--hostname=" + client.Hostname(),
+		"--authkey=" + key2.GetKey(),
+		"--force-reauth",
+	})
+
+	// Server-side: node retagged in place, same node ID and IPs, no duplicate.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "must not duplicate the node")
+
+		if len(nodes) == 1 {
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:second"})
+			assert.Equal(c, initialNodeID, nodes[0].GetId(), "node ID must be unchanged")
+			assert.ElementsMatch(c, initialIPs, nodes[0].GetIpAddresses(), "IPs must be preserved")
+		}
+	}, integrationutil.ScaledTimeout(20*time.Second), integrationutil.SlowPoll, "server must reflect the retag")
+
+	// Node self view: the new tag propagates to the client (issue #2978 surface).
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		assertNodeSelfHasTagsWithCollect(c, client, []string{"tag:second"})
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "node self view must reflect the retag")
+
+	// Netmap: independent serialization surface.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		assertNetmapSelfHasTagsWithCollect(c, client, []string{"tag:second"})
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "netmap self must reflect the retag")
+
+	t.Logf("Test #3370 PASS: re-keying with a different tagged key retagged the node in place")
+}
+
+// TestTagsReauthDifferentKeyRemovesTag is the sharpest proof of the KB 1068
+// "replaces, not merges" rule at the integration level: a node registered with
+// a two-tag key is re-keyed with a single-tag key, and the second tag must be
+// *removed*, not retained. Tag removal is the highest-risk propagation path
+// (peers must stop seeing the removed tag), so it is worth a real-client test.
+//
+// https://github.com/juanfont/headscale/issues/3370
+func TestTagsReauthDifferentKeyRemovesTag(t *testing.T) {
+	IntegrationSkip(t)
+
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		Users:        []string{tagTestUser},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnv(
+		[]tsic.Option{},
+		hsic.WithACLPolicy(tagsTestPolicy()),
+		hsic.WithTestName("tags-rekey-remove"),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	userMap, err := headscale.MapUsers()
+	require.NoError(t, err)
+
+	userID := userMap[tagTestUser].GetId()
+
+	// KEY1: single-use with BOTH tags.
+	key1, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:valid-owned", "tag:second"})
+	require.NoError(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"head",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+	)
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key1.GetKey())
+	require.NoError(t, err)
+
+	var initialNodeID uint64
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:second", "tag:valid-owned"})
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial registration")
+
+	// KEY2: fresh single-use with ONLY tag:valid-owned. Re-key.
+	key2, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	//nolint:errcheck // result is verified via EventuallyWithT below
+	client.Execute([]string{
+		"tailscale", "up",
+		"--login-server=" + headscale.GetEndpoint(),
+		"--hostname=" + client.Hostname(),
+		"--authkey=" + key2.GetKey(),
+		"--force-reauth",
+	})
+
+	// tag:second must be gone on every surface.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+			assert.Equal(c, initialNodeID, nodes[0].GetId())
+		}
+	}, integrationutil.ScaledTimeout(20*time.Second), integrationutil.SlowPoll, "re-keying must replace (remove tag:second), not merge")
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		assertNodeSelfHasTagsWithCollect(c, client, []string{"tag:valid-owned"})
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "removed tag must clear from node self view")
+
+	t.Logf("Test #3370 PASS: re-keying replaced the tag set (tag:second removed)")
+}
+
 // TestTagsAuthKeyWithTagCLICannotModifyAdminTags tests that the client CLI
 // cannot modify admin-assigned tags.
 //
@@ -3203,3 +3411,449 @@ func TestTagsAuthKeyConvertToUserViaCLIRegister(t *testing.T) {
 		}
 	}, integrationutil.HAConvergeTimeout, 1*time.Second, "node should be user-owned after conversion via CLI register")
 }
+
+// TestTaggedNodeLogoutReloginSingleUseKeyOnline reproduces issue #3371
+// end-to-end with a real tailscale client: a tagged node registered with a
+// single-use key logs out (`tailscale logout`) and re-authenticates with a
+// FRESH single-use tagged key. Tagged nodes never expire (KB 1068), so logout
+// must not stamp an expiry; before the fix the node was left permanently
+// expired and the fresh key was consumed on a re-registration that still
+// reported NodeKeyExpired, locking the node out forever.
+//
+// The observable proof at the integration level is that after relogin the node
+// is back online with a NULL expiry and the same node ID — not stuck expired.
+//
+// https://github.com/juanfont/headscale/issues/3371
+func TestTaggedNodeLogoutReloginSingleUseKeyOnline(t *testing.T) {
+	IntegrationSkip(t)
+
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		Users:        []string{tagTestUser},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnv(
+		[]tsic.Option{},
+		hsic.WithACLPolicy(tagsTestPolicy()),
+		hsic.WithTestName("tags-logout-single"),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	userMap, err := headscale.MapUsers()
+	require.NoError(t, err)
+
+	userID := userMap[tagTestUser].GetId()
+
+	// KEY1: single-use tag:valid-owned. Initial join.
+	key1, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"head",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+	)
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key1.GetKey())
+	require.NoError(t, err)
+
+	var initialNodeID uint64
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+			assert.Nil(c, nodes[0].GetExpiry(), "tagged node must have no expiry")
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial registration")
+
+	// `tailscale logout`. A tagged node must not be expired by this.
+	err = client.Logout()
+	require.NoError(t, err)
+
+	err = client.WaitForNeedsLogin(integrationutil.ScaledTimeout(60 * time.Second))
+	require.NoError(t, err)
+
+	// The node must remain in the DB, tagged, and crucially NOT carry a
+	// stale expiry. This is the #3371 root cause (a) surface.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "node must persist through logout")
+
+		if len(nodes) == 1 {
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+			assert.Nil(c, nodes[0].GetExpiry(), "#3371: logout must not stamp expiry on a tagged node")
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "tagged node must survive logout without expiry")
+
+	// KEY2: a FRESH single-use tagged key. Relogin.
+	key2, err := scenario.CreatePreAuthKeyWithTags(userID, false, false, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key2.GetKey())
+	require.NoError(t, err,
+		"#3371: a fresh key must re-authenticate the tagged node after logout")
+
+	// Back online, same node, still tagged, still no expiry.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "must not duplicate the node")
+
+		if len(nodes) == 1 {
+			assert.Equal(c, initialNodeID, nodes[0].GetId(), "node ID must be unchanged")
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+			assert.Nil(c, nodes[0].GetExpiry(), "#3371: tagged node must have no expiry after relogin")
+			assert.True(c, nodes[0].GetOnline(), "#3371: tagged node must be online after relogin, not stuck expired")
+		}
+	}, integrationutil.ScaledTimeout(60*time.Second), integrationutil.SlowPoll, "tagged node must come back online after relogin")
+
+	t.Logf("Test #3371 PASS: tagged node logged out and re-authenticated online with a fresh single-use key")
+}
+
+// TestTaggedNodeLogoutReloginReusableKeyOnline is the reusable-key variant of
+// issue #3371 (the "tailscale up hangs indefinitely" report). With a reusable
+// key the relogin does not hit "authkey already used", but before the fix the
+// node still stayed expired, so the client never observed a non-expired node.
+// The observable proof is the same: online with NULL expiry after relogin.
+//
+// https://github.com/juanfont/headscale/issues/3371
+func TestTaggedNodeLogoutReloginReusableKeyOnline(t *testing.T) {
+	IntegrationSkip(t)
+
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		Users:        []string{tagTestUser},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnv(
+		[]tsic.Option{},
+		hsic.WithACLPolicy(tagsTestPolicy()),
+		hsic.WithTestName("tags-logout-reuse"),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	userMap, err := headscale.MapUsers()
+	require.NoError(t, err)
+
+	userID := userMap[tagTestUser].GetId()
+
+	// A single REUSABLE tag:valid-owned key used for both login and relogin.
+	key, err := scenario.CreatePreAuthKeyWithTags(userID, true, false, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"head",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+	)
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key.GetKey())
+	require.NoError(t, err)
+
+	var initialNodeID uint64
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			assert.Nil(c, nodes[0].GetExpiry(), "tagged node must have no expiry")
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial registration")
+
+	err = client.Logout()
+	require.NoError(t, err)
+
+	err = client.WaitForNeedsLogin(integrationutil.ScaledTimeout(60 * time.Second))
+	require.NoError(t, err)
+
+	// Relogin with the SAME reusable key.
+	err = client.Login(headscale.GetEndpoint(), key.GetKey())
+	require.NoError(t, err)
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "must not duplicate the node")
+
+		if len(nodes) == 1 {
+			assert.Equal(c, initialNodeID, nodes[0].GetId(), "node ID must be unchanged")
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+			assert.Nil(c, nodes[0].GetExpiry(), "#3371: tagged node must have no expiry after reusable-key relogin")
+			assert.True(c, nodes[0].GetOnline(), "#3371: tagged node must be online after reusable-key relogin")
+		}
+	}, integrationutil.ScaledTimeout(60*time.Second), integrationutil.SlowPoll, "tagged node must come back online after reusable-key relogin")
+
+	t.Logf("Test #3371 PASS: tagged node logged out and re-authenticated online with a reusable key")
+}
+
+// TestTagsOIDCReauthAddOwnedTag reproduces issue #3374 through the interactive
+// OIDC path with a real client. A node is registered tag-owned (no user) via
+// --advertise-tags, then re-authenticates via OIDC advertising an ADDITIONAL
+// owned tag. Because a tag-owned node has no user and its IP is in no owner
+// set, the pre-fix authorization (which asked only "can this NODE have the
+// tag") rejected the whole set, leaving the node logged out. The fix also
+// authorises against the authenticating user, so the added owned tag is
+// accepted.
+//
+// This is the OIDC/interactive twin of the #3370 PAK-path retag tests, and it
+// hard-asserts the resulting tag SET (the pre-existing web-auth add-tag test
+// only logs), so it catches the silent-drop where the pre-check passes but the
+// apply-time re-check in processReauthTags still rejects.
+//
+// https://github.com/juanfont/headscale/issues/3374
+func TestTagsOIDCReauthAddOwnedTag(t *testing.T) {
+	IntegrationSkip(t)
+
+	oidcUser := "oidcuser"
+
+	// Each OIDC login consumes one mock user, so the same identity must be
+	// listed twice: once for the initial login and once for the reauth.
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		OIDCUsers: []mockoidc.MockUser{
+			oidcMockUser(oidcUser, true),
+			oidcMockUser(oidcUser, true),
+		},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	oidcMap := map[string]string{
+		"HEADSCALE_OIDC_ISSUER":             scenario.mockOIDC.Issuer(),
+		"HEADSCALE_OIDC_CLIENT_ID":          scenario.mockOIDC.ClientID(),
+		"CREDENTIALS_DIRECTORY_TEST":        "/tmp",
+		"HEADSCALE_OIDC_CLIENT_SECRET_PATH": "${CREDENTIALS_DIRECTORY_TEST}/hs_client_oidc_secret",
+	}
+
+	// The OIDC user owns both tags. Ownership is what authorises the reauth tag
+	// change once the node is tag-owned. Reference the user by email; it already
+	// contains an "@", so no trailing "@" is added (that suffix is only for
+	// non-email usernames).
+	owner := new(policyv2.Username(oidcUser + "@headscale.net"))
+	policy := &policyv2.Policy{
+		TagOwners: policyv2.TagOwners{
+			"tag:valid-owned": policyv2.Owners{owner},
+			"tag:second":      policyv2.Owners{owner},
+		},
+		ACLs: []policyv2.ACL{
+			{
+				Action:       "accept",
+				Sources:      []policyv2.Alias{policyv2.Wildcard},
+				Destinations: []policyv2.AliasWithPorts{{Alias: policyv2.Wildcard, Ports: []tailcfg.PortRange{tailcfg.PortRangeAny}}},
+			},
+		},
+	}
+
+	err = scenario.CreateHeadscaleEnvWithLoginURL(
+		[]tsic.Option{
+			tsic.WithExtraLoginArgs([]string{"--advertise-tags=tag:valid-owned"}),
+		},
+		hsic.WithTestName("tags-oidc-addtag"),
+		hsic.WithConfigEnv(oidcMap),
+		hsic.WithFileInContainer("/tmp/hs_client_oidc_secret", []byte(scenario.mockOIDC.ClientSecret())),
+		hsic.WithACLPolicy(policy),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"unstable",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+		tsic.WithExtraLoginArgs([]string{"--advertise-tags=tag:valid-owned"}),
+	)
+	require.NoError(t, err)
+
+	// Initial OIDC login advertising tag:valid-owned.
+	u, err := client.LoginWithURL(headscale.GetEndpoint())
+	require.NoError(t, err)
+
+	_, err = doLoginURL(client.Hostname(), u)
+	require.NoError(t, err)
+
+	var initialNodeID uint64
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial tag-owned registration")
+
+	// Re-authenticate advertising an ADDITIONAL owned tag via --force-reauth,
+	// which drives the register/auth path (HandleNodeFromAuthPath), not a poll.
+	// Parse the login URL from this command's own output; issuing a separate
+	// `tailscale up` while a reauth is pending errors server-side.
+	command := []string{
+		"tailscale", "up",
+		"--login-server=" + headscale.GetEndpoint(),
+		"--hostname=" + client.Hostname(),
+		"--advertise-tags=tag:valid-owned,tag:second",
+		"--force-reauth",
+	}
+
+	stdout, stderr, _ := client.Execute(command)
+	t.Logf("reauth command output: stdout=%s stderr=%s", stdout, stderr)
+
+	loginURL, err := util.ParseLoginURLFromCLILogin(stdout + stderr)
+	require.NoError(t, err, "failed to parse login URL from reauth command")
+
+	_, err = doLoginURL(client.Hostname(), loginURL)
+	require.NoError(t, err)
+
+	// Both tags must be present — not rejected, not silently dropped.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "must not duplicate the node")
+
+		if len(nodes) == 1 {
+			assert.Equal(c, initialNodeID, nodes[0].GetId(), "node ID must be unchanged")
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned", "tag:second"})
+		}
+	}, integrationutil.ScaledTimeout(30*time.Second), integrationutil.SlowPoll, "#3374: added owned tag must be accepted on OIDC reauth")
+
+	t.Logf("Test #3374 PASS: OIDC reauth added an owned tag to a tag-owned node")
+}
+
+// TestTagsReauthEmptyTagsReturnsToUserSurvives covers the #3374 untag path with
+// a real client: a tag-owned node created by a tagged+ephemeral key
+// re-authenticates via user login with an EMPTY tag set. It must return to the
+// user, and — crucially — survive: clearing the tags must also clear the
+// node's reference to the ephemeral auth key, or the node stays IsEphemeral()
+// and is garbage-collected on its next disconnect, silently deleting the
+// user's just-claimed device.
+//
+// https://github.com/juanfont/headscale/issues/3374
+func TestTagsReauthEmptyTagsReturnsToUserSurvives(t *testing.T) {
+	IntegrationSkip(t)
+
+	spec := ScenarioSpec{
+		NodesPerUser: 0,
+		Users:        []string{tagTestUser},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoError(t, err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnvWithLoginURL(
+		[]tsic.Option{},
+		hsic.WithACLPolicy(tagsTestPolicy()),
+		hsic.WithTestName("tags-untag-survive"),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	userMap, err := headscale.MapUsers()
+	require.NoError(t, err)
+
+	userID := userMap[tagTestUser].GetId()
+
+	// A tagged + EPHEMERAL key: the node is tag-owned and ephemeral.
+	key, err := scenario.CreatePreAuthKeyWithTags(userID, false, true, []string{"tag:valid-owned"})
+	require.NoError(t, err)
+
+	client, err := scenario.CreateTailscaleNode(
+		"head",
+		tsic.WithNetwork(scenario.networks[scenario.testDefaultNetwork]),
+	)
+	require.NoError(t, err)
+
+	err = client.Login(headscale.GetEndpoint(), key.GetKey())
+	require.NoError(t, err)
+
+	err = client.WaitForRunning(integrationutil.PeerSyncTimeout())
+	require.NoError(t, err)
+
+	var initialNodeID uint64
+
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1)
+
+		if len(nodes) == 1 {
+			initialNodeID = nodes[0].GetId()
+			assertNodeHasTagsWithCollect(c, nodes[0], []string{"tag:valid-owned"})
+		}
+	}, integrationutil.StatusReadyTimeout, integrationutil.SlowPoll, "waiting for initial tag-owned ephemeral registration")
+
+	// Re-authenticate with an EMPTY tag set via --force-reauth. An
+	// already-authenticated node only emits a fresh login URL when forced, so
+	// parse the URL from this command's own output rather than issuing a
+	// second `tailscale up` (which would error with "no URL found").
+	command := []string{
+		"tailscale", "up",
+		"--login-server=" + headscale.GetEndpoint(),
+		"--hostname=" + client.Hostname(),
+		"--advertise-tags=",
+		"--force-reauth",
+	}
+
+	stdout, stderr, _ := client.Execute(command)
+	t.Logf("reauth command output: stdout=%s stderr=%s", stdout, stderr)
+
+	loginURL, err := util.ParseLoginURLFromCLILogin(stdout + stderr)
+	require.NoError(t, err, "failed to parse login URL from reauth command")
+
+	body, err := doLoginURL(client.Hostname(), loginURL)
+	require.NoError(t, err)
+
+	// CLI user-login registration untags the node and returns it to the user.
+	err = scenario.runHeadscaleRegister(tagTestUser, body)
+	require.NoError(t, err)
+
+	// Node returns to the user, keeps its ID, and does NOT vanish.
+	assert.EventuallyWithT(t, func(c *assert.CollectT) {
+		nodes, err := headscale.ListNodes()
+		assert.NoError(c, err)
+		assert.Len(c, nodes, 1, "#3374: untagged node must survive, not be GC'd as ephemeral")
+
+		if len(nodes) == 1 {
+			assert.Equal(c, initialNodeID, nodes[0].GetId(), "node ID must be unchanged")
+			assert.Empty(c, nodes[0].GetTags(), "#3374: node must have no tags after untag")
+			// A user-owned node reports its real user; a tagged node would
+			// report the special "tagged-devices" user instead.
+			assert.Equal(c, tagTestUser, nodes[0].GetUser().GetName(), "#3374: untagged node must return to the authenticating user")
+		}
+	}, integrationutil.ScaledTimeout(30*time.Second), integrationutil.SlowPoll, "#3374: empty-tags reauth returns node to user and it survives")
+
+	t.Logf("Test #3374 PASS: empty-tags reauth returned the ephemeral tag-owned node to its user and it survived")
+}

integration/ts2021_websocket_test.go

@@ -0,0 +1,315 @@
+package integration
+
+import (
+	"context"
+	"crypto/tls"
+	"crypto/x509"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/netip"
+	"net/url"
+	"testing"
+	"time"
+
+	"github.com/coder/websocket"
+	"github.com/juanfont/headscale/integration/dockertestutil"
+	"github.com/juanfont/headscale/integration/hsic"
+	"github.com/juanfont/headscale/integration/tsic"
+	"github.com/ory/dockertest/v3"
+	"github.com/samber/lo"
+	"github.com/stretchr/testify/require"
+	"tailscale.com/control/controlbase"
+	"tailscale.com/control/controlhttp/controlhttpcommon"
+	"tailscale.com/net/wsconn"
+	"tailscale.com/tailcfg"
+	"tailscale.com/types/key"
+	"tailscale.com/util/rands"
+)
+
+// Tailscale's JS/WASM control client opens /ts2021 as a browser WebSocket — an
+// HTTP GET upgrade — rather than the native client's HTTP POST upgrade. A router
+// that registers /ts2021 for POST only rejects that GET with 405 before the
+// Noise handshake starts, which breaks every WASM client (issue #3357).
+//
+// These two tests guard that path against real headscale:
+//
+//   - TestTS2021WebSocketGET dials the WebSocket GET directly from the test
+//     process using the same coder/websocket + controlbase primitives the WASM
+//     client uses. It is fast and always on.
+//   - TestTS2021WASMClientUnderNode runs the *actual* tailscale.com js/wasm
+//     control dial (integration/wasmic/wasmclient, built for GOOS=js) inside a
+//     Node container, alongside normal Tailscale clients, and asserts it
+//     completes the Noise handshake with headscale over the WebSocket.
+//
+// The server cannot tell the two apart: both send GET /ts2021 with
+// Sec-WebSocket-Protocol: tailscale-control-protocol. Before the fix both fail
+// with 405; after it, both complete the handshake.
+
+// TestTS2021WebSocketGET connects to /ts2021 over a WebSocket GET from the test
+// process and completes the Noise handshake, exactly as a browser/WASM client
+// would.
+func TestTS2021WebSocketGET(t *testing.T) {
+	IntegrationSkip(t)
+	t.Parallel()
+
+	spec := ScenarioSpec{
+		NodesPerUser: 1,
+		Users:        []string{"user1"},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoErrorf(t, err, "failed to create scenario: %s", err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	err = scenario.CreateHeadscaleEnv([]tsic.Option{}, hsic.WithTestName("ts2021ws"))
+	requireNoErrHeadscaleEnv(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	conn, err := dialTS2021WebSocket(t, headscale.GetEndpoint(), headscale.GetCert())
+	require.NoError(t, err,
+		"WebSocket GET to /ts2021 must reach NoiseUpgradeHandler, not be rejected by the router with 405")
+	require.NotNil(t, conn)
+	t.Cleanup(func() { _ = conn.Close() })
+
+	t.Logf("noise established over websocket, protocol version %d", conn.ProtocolVersion())
+}
+
+// TestTS2021WASMClientUnderNode runs the real tailscale.com js/wasm control dial
+// inside a Node container against real headscale, next to normal Tailscale
+// clients, and asserts the WASM client completes the /ts2021 handshake.
+func TestTS2021WASMClientUnderNode(t *testing.T) {
+	IntegrationSkip(t)
+	t.Parallel()
+
+	spec := ScenarioSpec{
+		NodesPerUser: 2,
+		Users:        []string{"user1"},
+		Networks: map[string]NetworkSpec{
+			"wasmnet": {Users: []string{"user1"}},
+		},
+		ExtraService: map[string][]extraServiceFunc{
+			"wasmnet": {wasmClientService},
+		},
+		// The wasm client image builds from this module; pair it with the
+		// head Tailscale clients so the whole environment is current.
+		Versions: []string{"head"},
+	}
+
+	scenario, err := NewScenario(spec)
+
+	require.NoErrorf(t, err, "failed to create scenario: %s", err)
+	defer scenario.ShutdownAssertNoPanics(t)
+
+	// The Tailscale JS/WASM client dials the control server as a WebSocket.
+	// client_js.go only honours a custom port for ws:// (plain HTTP); over
+	// wss:// it always targets :443, so it cannot reach a TLS control server on
+	// :8080. Run headscale without TLS, matching the http:// setup in the issue.
+	err = scenario.CreateHeadscaleEnv(
+		[]tsic.Option{},
+		hsic.WithTestName("ts2021wasm"),
+		hsic.WithoutTLS(),
+	)
+	requireNoErrHeadscaleEnv(t, err)
+
+	allClients, err := scenario.ListTailscaleClients()
+	requireNoErrListClients(t, err)
+
+	allIPs, err := scenario.ListTailscaleClientsIPs()
+	requireNoErrListClientIPs(t, err)
+
+	// Normal Tailscale clients come up and form a working tailnet alongside the
+	// WASM control client.
+	err = scenario.WaitForTailscaleSync()
+	requireNoErrSync(t, err)
+
+	headscale, err := scenario.Headscale()
+	requireNoErrGetHeadscale(t, err)
+
+	// Sanity-check the tailnet the WASM client is joining: the normal clients
+	// must be able to reach each other.
+	allAddrs := lo.Map(allIPs, func(x netip.Addr, _ int) string { return x.String() })
+	assertPingAll(t, allClients, allAddrs)
+
+	services, err := scenario.Services("wasmnet")
+	require.NoError(t, err)
+	require.Len(t, services, 1, "expected the wasm client container")
+
+	wasm := services[0]
+	controlURL := headscale.GetEndpoint()
+
+	// Fetch the server's Noise key here and pass it to the WASM client: Go's
+	// net/http DNS resolver is unavailable under GOOS=js, so the client can only
+	// use the JS WebSocket transport, not an HTTP GET to /key.
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
+	defer cancel()
+
+	controlKey, err := fetchServerNoiseKey(ctx, &http.Client{Timeout: 15 * time.Second}, controlURL)
+	require.NoError(t, err)
+	controlKeyText, err := controlKey.MarshalText()
+	require.NoError(t, err)
+
+	// Run the real js/wasm control client under Node: it dials /ts2021 as a
+	// WebSocket GET; success means the Noise handshake completed.
+	stdout, stderr, err := dockertestutil.ExecuteCommand(
+		wasm,
+		[]string{"node", "/app/wasm_exec_node.js", "/app/client.wasm", controlURL, string(controlKeyText)},
+		[]string{},
+		dockertestutil.ExecuteCommandTimeout(60*time.Second),
+	)
+	t.Logf("wasm client stdout:\n%s", stdout)
+	t.Logf("wasm client stderr:\n%s", stderr)
+
+	require.NoError(t, err,
+		"wasm control client must connect to /ts2021 over websocket (405 means the router rejected the GET)")
+	require.Contains(t, stdout, "WASM_TS2021_OK",
+		"wasm control client should report a completed Noise handshake")
+}
+
+// wasmClientService builds and starts the Node + js/wasm control-client
+// container (Dockerfile.wasmclient) on the given network so it can reach
+// headscale by hostname. It idles; the test execs the client on demand.
+func wasmClientService(s *Scenario, networkName string) (*dockertest.Resource, error) {
+	hash := rands.HexString(hsicOIDCMockHashLength)
+	hostname := "hs-wasmclient-" + hash
+
+	network, ok := s.networks[s.prefixedNetworkName(networkName)]
+	if !ok {
+		return nil, fmt.Errorf("network does not exist: %s", networkName) //nolint:err113
+	}
+
+	runOpts := &dockertest.RunOptions{
+		Name:     hostname,
+		Networks: []*dockertest.Network{network},
+		Env:      []string{},
+	}
+	dockertestutil.DockerAddIntegrationLabels(runOpts, "wasmclient")
+
+	buildOpts := &dockertest.BuildOptions{
+		Dockerfile: "Dockerfile.wasmclient",
+		ContextDir: dockerContextPath,
+	}
+
+	resource, err := s.pool.BuildAndRunWithBuildOptions(
+		buildOpts,
+		runOpts,
+		dockertestutil.DockerRestartPolicy,
+	)
+	if err != nil {
+		return nil, fmt.Errorf("building wasm client container: %w", err)
+	}
+
+	return resource, nil
+}
+
+// dialTS2021WebSocket opens /ts2021 as a WebSocket GET (subprotocol
+// tailscale-control-protocol) and completes the Noise handshake, mirroring what
+// tailscale.com/control/controlhttp/client_js.go does in a browser. It returns
+// the established Noise connection, or an error (a router that only allows POST
+// returns 405 here).
+func dialTS2021WebSocket(t *testing.T, endpoint string, caCert []byte) (*controlbase.Conn, error) {
+	t.Helper()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	t.Cleanup(cancel)
+
+	u, err := url.Parse(endpoint)
+	require.NoError(t, err)
+
+	httpClient := &http.Client{Timeout: 15 * time.Second}
+
+	if u.Scheme == "https" {
+		pool := x509.NewCertPool()
+		pool.AppendCertsFromPEM(caCert)
+		httpClient.Transport = &http.Transport{
+			TLSClientConfig: &tls.Config{RootCAs: pool, MinVersion: tls.VersionTLS12},
+		}
+	}
+
+	controlKey, err := fetchServerNoiseKey(ctx, httpClient, endpoint)
+	require.NoError(t, err)
+
+	init, cont, err := controlbase.ClientDeferred(
+		key.NewMachine(),
+		controlKey,
+		uint16(tailcfg.CurrentCapabilityVersion),
+	)
+	require.NoError(t, err)
+
+	wsScheme := "ws"
+	if u.Scheme == "https" {
+		wsScheme = "wss"
+	}
+
+	wsURL := &url.URL{
+		Scheme: wsScheme,
+		Host:   u.Host,
+		Path:   "/ts2021",
+		RawQuery: url.Values{
+			controlhttpcommon.HandshakeHeaderName: []string{base64.StdEncoding.EncodeToString(init)},
+		}.Encode(),
+	}
+
+	wsConn, resp, err := websocket.Dial(ctx, wsURL.String(), &websocket.DialOptions{
+		Subprotocols: []string{controlhttpcommon.UpgradeHeaderValue},
+		HTTPClient:   httpClient,
+	})
+	if resp != nil && resp.Body != nil {
+		_ = resp.Body.Close()
+	}
+
+	if err != nil {
+		return nil, err
+	}
+
+	netConn := wsconn.NetConn(ctx, wsConn, websocket.MessageBinary, wsURL.String())
+
+	cbConn, err := cont(ctx, netConn)
+	if err != nil {
+		_ = netConn.Close()
+		return nil, fmt.Errorf("noise handshake over websocket: %w", err)
+	}
+
+	return cbConn, nil
+}
+
+// fetchServerNoiseKey retrieves headscale's Noise public key from /key, the same
+// endpoint a real client consults before dialing /ts2021.
+func fetchServerNoiseKey(
+	ctx context.Context,
+	client *http.Client,
+	endpoint string,
+) (key.MachinePublic, error) {
+	var zero key.MachinePublic
+
+	keyURL := fmt.Sprintf("%s/key?v=%d", endpoint, tailcfg.CurrentCapabilityVersion)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, keyURL, nil)
+	if err != nil {
+		return zero, err
+	}
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return zero, err
+	}
+	defer resp.Body.Close()
+
+	var k tailcfg.OverTLSPublicKeyResponse
+
+	err = json.NewDecoder(resp.Body).Decode(&k)
+	if err != nil {
+		return zero, fmt.Errorf("decoding /key response: %w", err)
+	}
+
+	if k.PublicKey.IsZero() {
+		return zero, errors.New("server returned zero Noise public key") //nolint:err113
+	}
+
+	return k.PublicKey, nil
+}

integration/wasmic/wasmclient/main.go

@@ -0,0 +1,103 @@
+//go:build js
+
+// Command wasmclient is a minimal Tailscale control client compiled to
+// GOOS=js/GOARCH=wasm and run under Node. It exercises the real
+// tailscale.com/control/controlhttp js/wasm dial path
+// (control/controlhttp/client_js.go), which opens /ts2021 as a browser-style
+// WebSocket GET — the exact transport a Tailscale JS/WASM client uses.
+//
+// It is the container-side half of the integration test guarding issue #3357:
+// headscale must register /ts2021 for GET, not POST only, or the WebSocket
+// upgrade is rejected with 405 before the Noise handshake can start.
+//
+// It is intentionally not the full tsconnect IPN — the regression is entirely
+// in the control-connection upgrade, and this drives the real upgrade code with
+// the smallest possible harness. On success it prints wasmSuccessMarker and
+// exits 0; on any failure it prints wasmFailureMarker and exits non-zero.
+package main
+
+import (
+	"context"
+	"fmt"
+	"net/url"
+	"os"
+	"time"
+
+	"tailscale.com/control/controlhttp"
+	"tailscale.com/tailcfg"
+	"tailscale.com/types/key"
+)
+
+// These markers are matched by the integration test on the client's stdout.
+const (
+	wasmSuccessMarker = "WASM_TS2021_OK"
+	wasmFailureMarker = "WASM_TS2021_FAIL"
+)
+
+func main() {
+	if len(os.Args) < 3 {
+		fmt.Printf("%s: usage: wasmclient <control-url> <noise-key>\n", wasmFailureMarker)
+		os.Exit(2)
+	}
+
+	if err := run(os.Args[1], os.Args[2]); err != nil {
+		fmt.Printf("%s: %v\n", wasmFailureMarker, err)
+		os.Exit(1)
+	}
+}
+
+// run dials /ts2021 exactly as tailscale.com/control/controlhttp/client_js.go
+// does in a browser: a WebSocket GET via the JS/undici WebSocket. The server's
+// Noise key is passed in (the test fetches /key) rather than fetched here,
+// because Go's net/http DNS resolver is unavailable under GOOS=js — only the
+// WebSocket transport, which runs through the JS host, works.
+func run(controlURL, noiseKeyText string) error {
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	u, err := url.Parse(controlURL)
+	if err != nil {
+		return fmt.Errorf("parse control url %q: %w", controlURL, err)
+	}
+
+	var controlKey key.MachinePublic
+	if err := controlKey.UnmarshalText([]byte(noiseKeyText)); err != nil {
+		return fmt.Errorf("parse noise key %q: %w", noiseKeyText, err)
+	}
+
+	port := u.Port()
+	if port == "" {
+		if u.Scheme == "https" {
+			port = "443"
+		} else {
+			port = "80"
+		}
+	}
+
+	// client_js.go selects ws:// (and appends the port) only when HTTPPort is a
+	// custom non-80 port and HTTPS is 443 or disabled; otherwise it dials wss://
+	// on the default port. Set the fields to match the server's actual scheme.
+	d := &controlhttp.Dialer{
+		Hostname:        u.Hostname(),
+		MachineKey:      key.NewMachine(),
+		ControlKey:      controlKey,
+		ProtocolVersion: uint16(tailcfg.CurrentCapabilityVersion),
+	}
+	if u.Scheme == "https" {
+		d.HTTPSPort = port
+	} else {
+		d.HTTPPort = port
+		d.HTTPSPort = controlhttp.NoPort
+	}
+
+	conn, err := d.Dial(ctx)
+	if err != nil {
+		return fmt.Errorf("ts2021 websocket dial: %w", err)
+	}
+	defer conn.Close()
+
+	fmt.Printf("%s: noise established over websocket, protocol version %d\n",
+		wasmSuccessMarker, conn.ProtocolVersion())
+
+	return nil
+}

integration/wasmic/wasmclient/stub.go

@@ -0,0 +1,8 @@
+//go:build !js
+
+// This package only does something when built for GOOS=js (see main.go). The
+// stub exists so `go build ./...` and `go vet ./...` on the host don't fail with
+// "build constraints exclude all Go files" for this directory.
+package main
+
+func main() {}

mkdocs.yml

@@ -68,7 +68,9 @@ exclude_docs: |
 # Plugins
 plugins:
   - search:
-      separator: '[\s\-,:!=\[\]()"`/]+|\.(?!\d)|&[lg]t;|(?!\b)(?=[A-Z][a-z])'
+      provider: lunr
+      lunr:
+        separator: '[\s\-,:!=\[\]()"`/]+|\.(?!\d)|&[lg]t;|(?!\b)(?=[A-Z][a-z])'
   - macros:
   - include-markdown:
   - minify:
@@ -111,7 +113,7 @@ extra:
     - icon: fontawesome/brands/discord
       link: https://discord.gg/c84AZQhmpx
   headscale:
-    version: 0.29.0
+    version: 0.29.3
 
 # Extensions
 markdown_extensions: