policy: key autogroup:self invalidation on UserID not User view

invalidateAutogroupSelfCache derived the owning user from
node.User().ID(), dereferencing the User association view. The
NodeStore stores nodes by value with User as a *User pointer, and not
every write path hydrates that association, so a non-tagged node could
reach SetNodes with UserID set but User nil. UserView.ID() then
dereferenced a nil *User and panicked on /machine/map whenever an
autogroup:self policy was active and a node restarted tailscaled.

Group affected nodes by node.TypedUserID(), which reads the UserID
field directly. UserID is the authoritative ownership field for
non-tagged nodes, so this needs no hydrated association and fixes every
.User().ID() site in the function.

hscontrol/policy/v2/policy.go

@@ -1299,13 +1299,18 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 	// Tagged nodes don't participate in autogroup:self (identity is tag-based),
 	// so we skip them when collecting affected users, except when tag status changes
 	// (which affects the user's device set).
-	affectedUsers := make(map[uint]struct{})
+	//
+	// Ownership is keyed on TypedUserID (the UserID field), not the User
+	// association view: the NodeStore holds nodes by value with User as a
+	// *User pointer, and not every write path hydrates that association. A
+	// non-tagged node always has UserID set, so it is the reliable owner key.
+	affectedUsers := make(map[types.UserID]struct{})
 
 	// Check for removed nodes (only non-tagged nodes affect autogroup:self)
 	for nodeID, oldNode := range oldNodeMap {
 		if _, exists := newNodeMap[nodeID]; !exists {
 			if !oldNode.IsTagged() {
-				affectedUsers[oldNode.User().ID()] = struct{}{}
+				affectedUsers[oldNode.TypedUserID()] = struct{}{}
 			}
 		}
 	}
@@ -1314,7 +1319,7 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 	for nodeID, newNode := range newNodeMap {
 		if _, exists := oldNodeMap[nodeID]; !exists {
 			if !newNode.IsTagged() {
-				affectedUsers[newNode.User().ID()] = struct{}{}
+				affectedUsers[newNode.TypedUserID()] = struct{}{}
 			}
 		}
 	}
@@ -1327,10 +1332,10 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 			if oldNode.IsTagged() != newNode.IsTagged() {
 				if !oldNode.IsTagged() {
 					// Was untagged, now tagged: user lost a device
-					affectedUsers[oldNode.User().ID()] = struct{}{}
+					affectedUsers[oldNode.TypedUserID()] = struct{}{}
 				} else {
 					// Was tagged, now untagged: user gained a device
-					affectedUsers[newNode.User().ID()] = struct{}{}
+					affectedUsers[newNode.TypedUserID()] = struct{}{}
 				}
 
 				continue
@@ -1342,9 +1347,9 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 			}
 
 			// Check if user changed (both versions are non-tagged here)
-			if oldNode.User().ID() != newNode.User().ID() {
-				affectedUsers[oldNode.User().ID()] = struct{}{}
-				affectedUsers[newNode.User().ID()] = struct{}{}
+			if oldNode.TypedUserID() != newNode.TypedUserID() {
+				affectedUsers[oldNode.TypedUserID()] = struct{}{}
+				affectedUsers[newNode.TypedUserID()] = struct{}{}
 			}
 
 			// Check if IPs changed (simple check - could be more sophisticated)
@@ -1352,12 +1357,12 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 
 			newIPs := newNode.IPs()
 			if len(oldIPs) != len(newIPs) {
-				affectedUsers[newNode.User().ID()] = struct{}{}
+				affectedUsers[newNode.TypedUserID()] = struct{}{}
 			} else {
 				// Check if any IPs are different
 				for i, oldIP := range oldIPs {
 					if i >= len(newIPs) || oldIP != newIPs[i] {
-						affectedUsers[newNode.User().ID()] = struct{}{}
+						affectedUsers[newNode.TypedUserID()] = struct{}{}
 						break
 					}
 				}
@@ -1370,7 +1375,7 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 	// because autogroup:self rules depend on the entire user's device set.
 	for nodeID := range pm.filterRulesMap {
 		// Find the user for this cached node
-		var nodeUserID uint
+		var nodeUserID types.UserID
 
 		found := false
 
@@ -1384,7 +1389,7 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 					break
 				}
 
-				nodeUserID = node.User().ID()
+				nodeUserID = node.TypedUserID()
 				found = true
 
 				break
@@ -1400,7 +1405,7 @@ func (pm *PolicyManager) invalidateAutogroupSelfCache(oldNodes, newNodes views.S
 						break
 					}
 
-					nodeUserID = node.User().ID()
+					nodeUserID = node.TypedUserID()
 					found = true
 
 					break

hscontrol/policy/v2/policy_test.go

@@ -226,6 +226,72 @@ func TestInvalidateAutogroupSelfCache(t *testing.T) {
 	}
 }
 
+// TestSetNodesAutogroupSelfUnhydratedUser reproduces the panic seen on
+// /machine/map when an autogroup:self policy is active and a non-tagged
+// node reaches the policy manager with its UserID set but the User
+// association left unhydrated (User pointer nil). The NodeStore stores
+// nodes by value with User as a *User; not every write path hydrates the
+// association, so the autogroup:self cache invalidation must derive the
+// owning user from UserID, not from the User view.
+func TestSetNodesAutogroupSelfUnhydratedUser(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"},
+	}
+
+	policy := `{
+		"acls": [
+			{
+				"action": "accept",
+				"src": ["autogroup:member"],
+				"dst": ["autogroup:self:*"]
+			}
+		]
+	}`
+
+	// unhydratedNode mirrors a NodeStore snapshot entry whose UserID is
+	// set (so it is unambiguously user-owned, not tagged) but whose User
+	// association was never loaded.
+	unhydratedNode := func(name, ipv4, ipv6 string, userID uint) *types.Node {
+		return &types.Node{
+			Hostname: name,
+			IPv4:     ap(ipv4),
+			IPv6:     ap(ipv6),
+			UserID:   new(userID),
+			User:     nil,
+		}
+	}
+
+	initialNodes := types.Nodes{
+		node("user1-node1", "100.64.0.1", "fd7a:115c:a1e0::1", users[0]),
+		node("user2-node1", "100.64.0.2", "fd7a:115c:a1e0::2", users[1]),
+	}
+	for i, n := range initialNodes {
+		n.ID = types.NodeID(i + 1) //nolint:gosec // safe conversion in test
+	}
+
+	pm, err := NewPolicyManager([]byte(policy), users, initialNodes.ViewSlice())
+	require.NoError(t, err)
+
+	require.False(t, initialNodes[0].IsTagged(), "node must be user-owned for autogroup:self")
+
+	// Simulate a node restarting tailscaled: the same node is pushed back
+	// into the policy manager, but the snapshot version has no hydrated
+	// User association. This is the exact shape that crashed beta.1.
+	updatedNodes := types.Nodes{
+		unhydratedNode("user1-node1", "100.64.0.1", "fd7a:115c:a1e0::1", users[0].ID),
+		node("user2-node1", "100.64.0.2", "fd7a:115c:a1e0::2", users[1]),
+	}
+	for i, n := range updatedNodes {
+		n.ID = types.NodeID(i + 1) //nolint:gosec // safe conversion in test
+	}
+
+	require.NotPanics(t, func() {
+		_, err = pm.SetNodes(updatedNodes.ViewSlice())
+	}, "SetNodes must not panic when a non-tagged node has an unhydrated User")
+	require.NoError(t, err)
+}
+
 // TestInvalidateGlobalPolicyCache tests the cache invalidation logic for global policies.
 func TestInvalidateGlobalPolicyCache(t *testing.T) {
 	mustIPPtr := func(s string) *netip.Addr {