Skip to main content

tailscale/
lib.rs

1//! A work-in-progress [Tailscale](https://tailscale.com/blog/how-tailscale-works) library.
2//!
3//! `tailscale` allows Rust programs to connect to a tailnet and exchange traffic with peers over
4//! TCP and UDP. It can communicate with other `tailscale`-based peers, `tailscaled` (the Tailscale
5//! Go client), `tsnet`, and `libtailscale` via public DERP servers.
6//!
7//! <div class="warning">
8//! `tailscale` is unstable and insecure.
9//!
10//! We welcome enthusiasm and interest, but please **do not** build production software using these
11//! libraries or rely on it for data privacy until we have a chance to batten down some hatches and
12//! complete a third-party audit.
13//!
14//! See the [Caveats section](#caveats) for more details.
15//! </div>
16//!
17//! For language bindings, see the following crates:
18//!
19//! - C: [ts_ffi](https://docs.rs/ts_ffi)
20//! - Python: [ts_python](https://docs.rs/ts_python)
21//! - Elixir: [ts_elixir](https://docs.rs/ts_elixir)
22//!
23//! For instructions on how to run tests, lints, etc., see [CONTRIBUTING.md]. For the high-level
24//! architecture and repository layout, see [ARCHITECTURE.md].
25//!
26//! ## Code Sample
27//!
28//! A simple UDP client that periodically sends messages to a tailnet peer at `100.64.0.1:5678`:
29//!
30//! ```no_run
31//! # use std::{
32//! #     time::Duration,
33//! #     net::Ipv4Addr,
34//! #     error::Error,
35//! # };
36//! #
37//! # #[tokio::main]
38//! # async fn main() -> Result<(), Box<dyn Error>> {
39//! // Open a new connection to the tailnet
40//! let dev = tailscale::Device::new(
41//!     &tailscale::Config::default_with_key_file("tsrs_keys.json").await?,
42//!     Some("YOUR_AUTH_KEY_HERE".to_owned()),
43//! ).await?;
44//!
45//! // Bind a UDP socket on our tailnet IP, port 1234
46//! let sock = dev.udp_bind((dev.ipv4_addr().await?, 1234).into()).await?;
47//!
48//! // Send a packet containing "hello, world!" to 100.64.0.1:5678 once per second
49//! loop {
50//!     sock.send_to((Ipv4Addr::new(100, 64, 0, 1), 5678).into(), b"hello, world!").await?;
51//!     tokio::time::sleep(Duration::from_secs(1)).await;
52//! }
53//! # }
54//! ```
55//!
56//! Additional examples of using the `tailscale` crate can be found in the [`examples/`] directory.
57//!
58//! ## Using `tailscale`
59//!
60//! To use this crate or the language bindings, you will need to set the `TS_RS_EXPERIMENT` env var
61//! to `this_is_unstable_software`. We'll remove this requirement after a third-party code/cryptography
62//! audit and any necessary fixes.
63//!
64//! Under the hood, we use Tokio for our async runtime. You must also use Tokio, any kind and most
65//! configurations of Tokio runtimes should work, but there must be one available when you call any
66//! async API functions. The easiest way to do this is to use `#[tokio::main]`, see the
67//! [Tokio docs](https://docs.rs/tokio) for more information. In the future, we would like to limit
68//! our reliance on Tokio so that there are alternatives for users of other async runtimes.
69//!
70//! ## Caveats
71//!
72//! This software is still a work-in-progress! We are providing it in the open at this stage out of
73//! a belief in open-source and to see where the community runs with it, but please be aware of a
74//! few important considerations:
75//!
76//! - This implementation contains unaudited cryptography and hasn't undergone a comprehensive
77//!   security analysis. Conservatively, assume there could be a critical security hole meaning
78//!   anything you send or receive could be in the clear on the public Internet.
79//! - There are no compatibility guarantees at the moment. This is early-days software - we may
80//!   break dependent code in order to get things right.
81//! - Direct peer-to-peer connections via NAT traversal are implemented (STUN-discovered endpoints
82//!   and Disco, with `CallMeMaybe` hole-punching over DERP), with DERP relays as the fallback when
83//!   no direct path is available. Hard/symmetric NATs get the same single fixed-local-port candidate
84//!   (`EndpointSTUN4LocalPort`) Go Tailscale uses; behind a NAT with no static port mapping a flow
85//!   may still stay relayed through DERP, which caps its throughput. (Upstream Go does **not** do a
86//!   "256-port birthday-paradox spray" — that is a common misconception; the single-candidate guess
87//!   is the actual behavior, and this fork matches it.)
88//!
89//! ## Feature Flags
90//!
91//! - `axum`: enables the `axum` module, which enables you to run an `axum` HTTP server on top
92//!   of a [`netstack::TcpListener`].
93//! - `tsnet`: enables the `tsnet` module — a Go-idiomatic `tsnet.Server`-shaped facade over
94//!   [`Device`] and [`Config`]. Set fields, then call `up`/`listen`/`dial`/`loopback`/`close`/…, and
95//!   the wrapped [`Device`] is built and started lazily on the first call (Go's "fields may be
96//!   changed until the first method call"). A thin ergonomics layer only — same engine, same typed
97//!   returns. See the [`tsnet` module docs] for the full Go `tsnet.Server` → Rust API mapping.
98//!
99//! ## Platform Support
100//!
101//! `tailscale` currently supports the following platforms:
102//!
103//! - Linux (x86_64 and ARM64)
104//! - macOS (ARM64)
105//!
106//! ## Component crates
107//!
108//! The following crates are part of the tailscale-rs project and are dependencies of this one. For
109//! many tasks, just this crate should be sufficient and these other crates are an implementation detail.
110//! There are other crates too, see [ARCHITECTURE.md]
111//! or the [GitHub repo](https://github.com/tailscale/tailscale-rs).
112//!
113//! - [ts_runtime](https://docs.rs/ts_runtime): for each API-level `Device`, the runtime uses an actor
114//!   architecture to manage the lifecycle of the control client, data plane components, netstack, etc.
115//!   A message bus passes updates and communications between these top-level actors.
116//! - [ts_netcheck](https://docs.rs/ts_netcheck): checks network availability and reports latency to
117//!   DERP servers in different regions.
118//! - [ts_netstack_smoltcp](https://docs.rs/ts_netstack_smoltcp): a [smoltcp](https://docs.rs/smoltcp)-based
119//!   network stack that processes Layer 3+ packets to/from the overlay network.
120//! - [ts_control](https://docs.rs/ts_control): control plane client that handles registration,
121//!   authorization/authentication, configuration, and streaming updates.
122//! - [ts_dataplane](https://docs.rs/ts_dataplane): wires all the individual data plane functions together,
123//!   flowing inbound and outbound packets through the components in the correct order.
124//! - [ts_tunnel](https://docs.rs/ts_tunnel): a partial implementation of the WireGuard specification
125//!   that protects all data plane traffic, and is interoperable with other WireGuard clients, including Tailscale clients.
126//! - [ts_cli_util](https://docs.rs/ts_cli_util): helpers for writing command line tools and initializing
127//!   logging, used in examples.
128//! - [ts_disco_protocol](https://docs.rs/ts_disco_protocol): incomplete implementation of Tailscale's
129//!   discovery protocol (disco).
130//!
131//! [ARCHITECTURE.md]: https://github.com/tailscale/tailscale-rs/blob/main/ARCHITECTURE.md
132//! [CONTRIBUTING.md]: https://github.com/tailscale/tailscale-rs/blob/main/CONTRIBUTING.md
133//! [`examples/`]: https://github.com/tailscale/tailscale-rs/blob/main/examples/README.md
134//! [open an issue]: https://github.com/tailscale/tailscale-rs/issues
135//! [`axum` HTTP server]: https://docs.rs/axum/latest/axum/
136//! [`tsnet` module docs]: https://docs.rs/geiserx_tailscale/latest/tailscale/tsnet/
137
138use std::{
139    net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr},
140    time::Duration,
141};
142
143#[doc(inline)]
144pub use config::Config;
145#[doc(inline)]
146pub use error::{Error, InternalErrorKind};
147// Re-exported so a downstream crate depending only on `tailscale` can name the auth-key secret type
148// for [`Device::new_with_secret`] without taking a separate, version-pinned dependency on `secrecy`
149// (which would risk a `SecretString`-type mismatch if the two `secrecy` majors diverged). Callers
150// pass `tailscale::SecretString`; `secrecy` is a pure-Rust wrapper (no aws-lc/openssl/ring).
151pub use secrecy::SecretString;
152#[doc(inline)]
153pub use ts_control::ExitNodeSelector;
154#[doc(inline)]
155pub use ts_control::Node as NodeInfo;
156#[doc(inline)]
157pub use ts_control::tls::{CertifiedKey, TlsAcceptor, TlsStream};
158#[doc(inline)]
159pub use ts_control::{CertError, MISSING_CERT_RPC, ServeConfig, ServeState, ServeTarget};
160/// The netmap DNS configuration returned by [`Device::dns_config`] (Go `netmap.NetworkMap.DNS`).
161#[doc(inline)]
162pub use ts_control::{DnsConfig, DnsResolver, ExtraRecord};
163#[doc(inline)]
164pub use ts_control::{ExitProxyConfig, ExitProxyScheme};
165pub use ts_control::{
166    IdTokenError, LogoutError, ServiceError, ServiceMode, SetDnsError, SetDnsInternalErrorKind,
167    SshAccept, SshAction, SshConnIdentity, SshDecision, SshDenyReason, SshPolicy, SshPrincipal,
168    SshRule, StableNodeId,
169};
170// Re-exported so the application data-path transport can be selected through the `tailscale`
171// facade alone: `Config::transport_mode` is `TransportMode` (default `Netstack`; `Tun(TunConfig {
172// name, mtu })` for a real kernel TUN interface). Both are `pub` in `ts_control` but were not
173// reachable through this facade, forcing downstream crates to depend on `ts_control` directly just
174// to name them.
175pub use ts_control::{TransportMode, TunConfig};
176#[doc(inline)]
177pub use ts_netstack_smoltcp::PingError;
178use ts_netstack_smoltcp::{CreateSocket, netcore::Channel};
179#[doc(inline)]
180pub use ts_runtime::fallback_tcp::{
181    FallbackConnFuture, FallbackConnHandler, FallbackDecision, FallbackTcpHandle,
182};
183#[doc(inline)]
184pub use ts_runtime::taildrop::WaitingFile;
185#[doc(inline)]
186pub use ts_runtime::{
187    DeviceState, DnsQueryResult, ExitNodeSuggestion, FileTarget, IpnBusWatcher, NetcheckReport,
188    Notify, NotifyWatchOpt, RegionLatency, RegistrationError, Status, StatusNode, TkaLogEntry,
189    WhoIs,
190};
191/// The interactive-login URL type returned by [`Device::pop_browser_url`].
192#[doc(inline)]
193pub use url::Url;
194
195#[cfg(feature = "axum")]
196pub mod axum;
197pub mod config;
198mod dial;
199mod error;
200#[cfg(feature = "hyper")]
201pub mod http;
202mod loopback;
203#[cfg(feature = "ssh")]
204pub mod ssh;
205#[cfg(feature = "tsnet")]
206pub mod tsnet;
207
208#[doc(inline)]
209pub use dial::{ConnectedUdpSocket, DialConn};
210#[doc(inline)]
211pub use loopback::LoopbackHandle;
212
213/// How a program connects to a tailnet and communicates with peers.
214///
215/// The `Device` connects to the control plane, registers itself with the tailnet, and communicates
216/// with tailnet peers. Its tailnet identity is determined by the key state provided at
217/// construction-time.
218pub struct Device {
219    runtime: ts_runtime::Runtime,
220    /// Command channel to the application netstack. `None` in TUN transport mode, where there is
221    /// no userspace application netstack; the channel-driven socket APIs ([`Device::udp_bind`],
222    /// [`Device::tcp_listen`], [`Device::tcp_connect`], [`Device::ping`]) are unsupported there.
223    channel: Option<Channel>,
224    /// Whether IPv6 is enabled on the tailnet overlay (the `Config::enable_ipv6` gate, default
225    /// `false`). Captured at construction; used by [`Device::listen_service`] to decide whether an
226    /// IPv6 VIP-service address is bindable (the netstack only accepts IPv6 overlay addresses when
227    /// this is set).
228    enable_ipv6: bool,
229    /// The stored Serve config + its live per-port accept loops (`tsnet`'s `Get/SetServeConfig` +
230    /// serving runtime). Built lazily on the first [`Device::set_serve_config`] (it needs this
231    /// node's overlay IPv4, only known after registration). Held here so its accept loops abort when
232    /// the `Device` drops; `None` (empty config) until the first `set`.
233    serve: std::sync::Mutex<Option<ts_runtime::serve::ServeManager>>,
234    /// The live Funnel ingress manager (`tsnet`'s `ListenFunnel` data path), built on
235    /// [`Device::listen_funnel`](crate::Device::listen_funnel). Held here so its TLS-termination pump and the installed peerAPI
236    /// ingress sink stay alive for the device's life (and tear down when a new `listen_funnel`
237    /// replaces it, or the `Device` drops). `None` until the first `listen_funnel`.
238    funnel: std::sync::Mutex<Option<ts_runtime::funnel::FunnelManager>>,
239}
240
241/// Map a [`ts_runtime::taildrop::TaildropError`] to the device-facing [`Error`]. `Error` is a
242/// `Copy` enum with no payload, so the I/O detail string is dropped, but the *kind* is preserved so
243/// a caller can still distinguish the actionable cases: an invalid name →
244/// [`InternalErrorKind::BadRequest`], an in-progress conflict → [`InternalErrorKind::AlreadyExists`],
245/// a missing file → [`InternalErrorKind::NotFound`], and any other filesystem failure →
246/// [`InternalErrorKind::Io`].
247fn taildrop_err(e: ts_runtime::taildrop::TaildropError) -> Error {
248    use ts_runtime::taildrop::TaildropError;
249    match e {
250        TaildropError::InvalidFileName => Error::Internal(InternalErrorKind::BadRequest),
251        TaildropError::FileExists => Error::Internal(InternalErrorKind::AlreadyExists),
252        TaildropError::Io(io) if io.kind() == std::io::ErrorKind::NotFound => {
253            Error::Internal(InternalErrorKind::NotFound)
254        }
255        TaildropError::Io(_) => Error::Internal(InternalErrorKind::Io),
256    }
257}
258
259/// Map a [`ts_runtime::taildrop_send::TaildropSendError`] (the Taildrop *sender*) to the
260/// device-facing [`Error`]. The send-side conflict/forbidden/unexpected-status cases all reduce to
261/// `BadRequest` (the peer refused the transfer for a request-level reason), a dial failure or
262/// timeout to `Timeout`, an invalid name to `BadRequest`, and any stream I/O failure to `Io`.
263fn taildrop_send_err(e: ts_runtime::taildrop_send::TaildropSendError) -> Error {
264    use ts_runtime::taildrop_send::TaildropSendError;
265    match e {
266        TaildropSendError::Connect | TaildropSendError::Timeout => Error::Timeout,
267        TaildropSendError::InvalidName
268        | TaildropSendError::Forbidden
269        | TaildropSendError::Conflict
270        | TaildropSendError::UnexpectedStatus(_) => Error::Internal(InternalErrorKind::BadRequest),
271        TaildropSendError::Io => Error::Internal(InternalErrorKind::Io),
272    }
273}
274
275/// Resolve the effective registration auth key from `auth_key` plus the config's
276/// workload-identity-federation (WIF) / OAuth-client fields.
277///
278/// With the `identity-federation` feature enabled, an OAuth client secret (`tskey-client-…`) or a
279/// `client_id` + (`id_token` | `audience`) is exchanged for a Tailscale auth key against the SaaS
280/// admin API before registration (Go `tsnet.Server`'s `resolveAuthKey`). Without the feature this is
281/// a pure pass-through: `auth_key` is returned unchanged and the WIF config fields are ignored, so
282/// the default build is byte-identical to before.
283#[cfg(feature = "identity-federation")]
284async fn resolve_auth_key(
285    config: &Config,
286    auth_key: Option<String>,
287) -> Result<Option<String>, Error> {
288    let wif = ts_control::WifConfig {
289        auth_key,
290        client_id: config.client_id.clone(),
291        client_secret: config.client_secret.clone(),
292        id_token: config.id_token.clone(),
293        audience: config.audience.clone(),
294        tags: config.requested_tags.clone(),
295    };
296    ts_control::resolve_auth_key(&wif, &config.control_server_url)
297        .await
298        .map_err(|e| {
299            tracing::error!(error = %e, "resolving auth key via workload-identity federation");
300            Error::Internal(InternalErrorKind::BadRequest)
301        })
302}
303
304/// Pass-through when the `identity-federation` feature is disabled: the auth key is used as-is and
305/// the WIF config fields have no effect (matching Go, where the federation path is compiled out
306/// unless its optional feature is linked).
307#[cfg(not(feature = "identity-federation"))]
308async fn resolve_auth_key(
309    _config: &Config,
310    auth_key: Option<String>,
311) -> Result<Option<String>, Error> {
312    Ok(auth_key)
313}
314
315impl Device {
316    /// Create a device from the given [`Config`] and auth key.
317    ///
318    /// Internally, this will spawn multiple asynchronous actors onto a Tokio runtime.
319    ///
320    /// # Example
321    ///
322    /// ```rust,no_run
323    /// # #[tokio::main]
324    /// # async fn main() -> Result<(), Box<dyn std::error::Error>> {
325    /// # use tailscale::*;
326    /// let dev = Device::new(
327    ///     &Config::default_with_key_file("tsrs_keys.json").await?,
328    ///     Some("MY_AUTH_KEY".to_string()),
329    /// ).await?;
330    /// # Ok(()) }
331    /// ```
332    pub async fn new(config: &Config, auth_key: Option<String>) -> Result<Self, Error> {
333        check_magic_env()?;
334
335        // Resolve the effective registration auth key. The explicit `auth_key` argument wins; if it
336        // is `None`, fall back to `config.auth_key` (Go `tsnet.Server.AuthKey`). When the
337        // `identity-federation` feature is enabled, the resolved key is further passed through the
338        // WIF / OAuth-client bootstrap, which exchanges an OAuth client secret (`tskey-client-…`) or
339        // an IdP-issued OIDC token for a Tailscale auth key before registration (SaaS-only).
340        let auth_key = auth_key.or_else(|| config.auth_key.clone());
341        let auth_key = resolve_auth_key(config, auth_key).await?;
342
343        let rt =
344            ts_runtime::Runtime::spawn(config.into(), auth_key, (&config.key_state).into()).await?;
345        // In TUN transport mode there is no application netstack, so the runtime has no command
346        // channel: that surfaces as `UnsupportedInTunMode`, which we map to a `None` channel rather
347        // than an error (the device is still usable for control-plane and peer-lookup APIs).
348        let channel = match rt.channel().await {
349            Ok(c) => Some(c),
350            Err(e) if e.kind == ts_runtime::ErrorKind::UnsupportedInTunMode => None,
351            Err(e) => return Err(e.into()),
352        };
353
354        Ok(Self {
355            runtime: rt,
356            channel,
357            enable_ipv6: config.enable_ipv6,
358            serve: std::sync::Mutex::new(None),
359            funnel: std::sync::Mutex::new(None),
360        })
361    }
362
363    /// Create a device from the given [`Config`] and a [`SecretString`] auth key.
364    ///
365    /// This is a back-compat-preserving convenience over [`new`](Self::new) for callers that already
366    /// hold the registration auth key as a [`secrecy::SecretString`] (e.g. a daemon that keeps the
367    /// pre-auth key wrapped end-to-end). It lets the caller avoid materializing a plain `String` at
368    /// the engine boundary: the secret is exposed only on the last inch, immediately before being
369    /// handed to [`new`](Self::new).
370    ///
371    /// # Honesty about the plaintext window
372    ///
373    /// This closes the *caller's* boundary, **not** the engine's internal handling. The engine still
374    /// resolves the auth key to a plain `String` internally for registration (the plaintext `String`
375    /// window inside the engine is identical to calling [`new`](Self::new) directly) — this method
376    /// does not make the engine itself secret-clean. If you call [`new`](Self::new) you create that
377    /// `String` yourself; if you call this you do not, but the engine creates one either way.
378    ///
379    /// Passing `None` is equivalent to `new(config, None)` (falls back to `config.auth_key`).
380    ///
381    /// # Example
382    ///
383    /// ```rust,no_run
384    /// # #[tokio::main]
385    /// # async fn main() -> Result<(), Box<dyn std::error::Error>> {
386    /// # use tailscale::*;
387    /// let dev = Device::new_with_secret(
388    ///     &Config::default_with_key_file("tsrs_keys.json").await?,
389    ///     Some(SecretString::from("MY_AUTH_KEY")),
390    /// ).await?;
391    /// # Ok(()) }
392    /// ```
393    pub async fn new_with_secret(
394        config: &Config,
395        auth_key: Option<SecretString>,
396    ) -> Result<Self, Error> {
397        use secrecy::ExposeSecret as _;
398
399        // Expose the secret on the last inch and delegate to `new`, so the spawn/registration path
400        // is shared verbatim (no duplicated runtime-spawn logic) and the engine-internal plaintext
401        // window is byte-for-byte identical to a direct `new` call.
402        let plain = auth_key.map(|s| s.expose_secret().to_string());
403        Self::new(config, plain).await
404    }
405
406    /// The application netstack command channel, or an error in TUN transport mode (no application
407    /// netstack exists).
408    fn channel(&self) -> Result<&Channel, Error> {
409        self.channel
410            .as_ref()
411            .ok_or(Error::Internal(InternalErrorKind::UnsupportedInTunMode))
412    }
413
414    /// Get this [`Device`]'s IPv4 tailnet address.
415    pub async fn ipv4_addr(&self) -> Result<Ipv4Addr, Error> {
416        self.runtime
417            .control
418            .ask(ts_runtime::control_runner::Ipv4)
419            .await
420            .map_err(ts_runtime::Error::from)?
421            .ok_or(Error::Internal(InternalErrorKind::Actor))
422    }
423
424    /// Get this [`Device`]'s IPv6 tailnet address.
425    pub async fn ipv6_addr(&self) -> Result<Ipv6Addr, Error> {
426        self.runtime
427            .control
428            .ask(ts_runtime::control_runner::Ipv6)
429            .await
430            .map_err(ts_runtime::Error::from)?
431            .ok_or(Error::Internal(InternalErrorKind::Actor))
432    }
433
434    /// This node's tailnet IPv4 and (when provisioned) IPv6 addresses as a pair — the Rust analog of
435    /// Go `tsnet.Server.TailscaleIPs() (ip4, ip6 netip.Addr)`.
436    ///
437    /// Reads the self node's assigned addresses (the same source Go splits by family). The tailnet
438    /// is IPv4-only unless [`Config::enable_ipv6`](crate::config::Config) is set, so the IPv6 half is
439    /// `None` when no v6 address is assigned — the Rust shape for Go returning the zero `netip.Addr`
440    /// in that case (Go's IPv6-absent sentinel). Errors until the first netmap is received (no self
441    /// node yet), matching Go returning invalid addresses before the node has joined.
442    pub async fn tailscale_ips(&self) -> Result<(Ipv4Addr, Option<Ipv6Addr>), Error> {
443        let me = self.self_node().await?;
444        let v4 = me.tailnet_address.ipv4.addr();
445        let v6 = me.tailnet_address.ipv6.addr();
446        // The decoder synthesizes the unspecified `::` placeholder on an IPv4-only tailnet; surface
447        // a real v6 only when IPv6 is enabled AND a non-placeholder address was assigned.
448        let v6 = (self.enable_ipv6 && !v6.is_unspecified()).then_some(v6);
449        Ok((v4, v6))
450    }
451
452    /// Bind a UDP socket to the specified [`SocketAddr`].
453    ///
454    /// Returns an error in TUN transport mode (there is no application netstack to bind on).
455    pub async fn udp_bind(&self, socket_addr: SocketAddr) -> Result<netstack::UdpSocket, Error> {
456        self.channel()?
457            .udp_bind(socket_addr)
458            .await
459            .map_err(Into::into)
460    }
461
462    /// Bind a TCP listener to the specified [`SocketAddr`].
463    ///
464    /// Returns an error in TUN transport mode (there is no application netstack to listen on).
465    pub async fn tcp_listen(
466        &self,
467        socket_addr: SocketAddr,
468    ) -> Result<netstack::TcpListener, Error> {
469        self.channel()?
470            .tcp_listen(socket_addr)
471            .await
472            .map_err(Into::into)
473    }
474
475    /// Register a fallback TCP handler (like `tsnet`'s `RegisterFallbackTCPHandler`).
476    ///
477    /// The callback is consulted for every inbound TCP flow that matches **no** explicit
478    /// [`Device::tcp_listen`] listener, with the flow's `(src, dst)` addresses. It returns
479    /// `(handler, intercept)`:
480    /// - `(_, false)` — decline; the next registered callback is tried.
481    /// - `(Some(h), true)` — claim the flow; `h` is handed the accepted [`netstack::TcpStream`].
482    /// - `(None, true)` — claim and reject the flow (the connection is closed).
483    ///
484    /// Multiple handlers may be registered; they are consulted in registration order and the first
485    /// to intercept wins. The returned [`FallbackTcpHandle`] deregisters the handler when dropped.
486    ///
487    /// Handlers serve flows over the overlay netstack only — never a host socket — and a flow no
488    /// handler claims is closed (fail-closed), never direct-dialed.
489    ///
490    /// Returns an error in TUN transport mode (there is no application netstack to attach to).
491    pub fn register_fallback_tcp_handler<F>(&self, cb: F) -> Result<FallbackTcpHandle, Error>
492    where
493        F: Fn(SocketAddr, SocketAddr) -> FallbackDecision + Send + Sync + 'static,
494    {
495        self.runtime
496            .register_fallback_tcp_handler(std::sync::Arc::new(cb))
497            .map_err(Into::into)
498    }
499
500    /// Resolve a tailnet peer (or this node) by MagicDNS name to its tailnet IPv4 address.
501    ///
502    /// This is an in-process lookup against the netmap we already hold — like `tsnet`'s in-memory
503    /// `dnsMap`, it does not query any DNS server (there is no `100.100.100.100` resolver). The
504    /// `name` may be a bare hostname or a fully-qualified MagicDNS name, with or without a trailing
505    /// dot, in any case (matching is case-insensitive). Returns `Ok(None)` if no tailnet node has
506    /// that name.
507    ///
508    /// Only MagicDNS names are resolved; names outside the tailnet are not looked up here, so the
509    /// caller's system resolver remains responsible for them. IPv6 is intentionally not resolved —
510    /// this fork operates IPv4-only on the tailnet.
511    pub async fn resolve(&self, name: &str) -> Result<Option<Ipv4Addr>, Error> {
512        if let Some(peer) = self.peer_by_name(name).await? {
513            return Ok(Some(peer.tailnet_address.ipv4.addr()));
514        }
515
516        // tsnet's dnsMap also resolves our own name; fall back to self when no peer matches.
517        let me = self.self_node().await?;
518        if me.matches_name(name) {
519            return Ok(Some(me.tailnet_address.ipv4.addr()));
520        }
521
522        Ok(None)
523    }
524
525    /// Run a real DNS query through the tailnet's MagicDNS responder (the `100.100.100.100`
526    /// forward path), returning the raw response, RCODE, and resolver(s) consulted — the analogue of
527    /// Go `LocalClient.QueryDNS`.
528    ///
529    /// Unlike [`resolve`](Self::resolve) (an in-memory netmap lookup that answers only MagicDNS
530    /// A-records), this issues an actual query of any `qtype` and runs it through the live
531    /// responder: an authoritative tailnet name is answered locally, anything else is forwarded to
532    /// the configured split-DNS / recursive upstreams (or delegated to the active exit node's DoH).
533    /// The response is returned as raw bytes (matching Go's `QueryDNS`), since this fork's DNS wire
534    /// codec has no answer-record decoder; the caller parses records itself if needed.
535    ///
536    /// `qtype` is the raw RFC 1035 TYPE value (`1`=A, `28`=AAAA, `12`=PTR, `16`=TXT, `33`=SRV,
537    /// `65`=HTTPS/SVCB, …). Anti-leak is inherited from the responder: a tailnet-suffix name never
538    /// egresses, recursive forwards delegate to the exit node when one is active, and only IPv4
539    /// upstreams are dialed.
540    ///
541    /// Returns an [`Error::Internal`] with `InternalErrorKind::UnsupportedInTunMode` in TUN
542    /// transport mode (MagicDNS there is an in-packet intercept, not a queryable responder).
543    pub async fn query_dns(&self, name: &str, qtype: u16) -> Result<DnsQueryResult, Error> {
544        self.runtime
545            .query_dns(name, qtype)
546            .await
547            .map_err(Into::into)
548    }
549
550    /// Connect to a tailnet peer by MagicDNS name and port over TCP.
551    ///
552    /// Resolves `name` via [`Device::resolve`] (an in-process netmap lookup, no DNS server), then
553    /// dials the resulting tailnet IPv4 address. Returns [`InternalErrorKind::BadRequest`] if the
554    /// name does not resolve to a tailnet node.
555    pub async fn connect_by_name(
556        &self,
557        name: &str,
558        port: u16,
559    ) -> Result<netstack::TcpStream, Error> {
560        let addr = self
561            .resolve(name)
562            .await?
563            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
564
565        self.tcp_connect((addr, port).into()).await
566    }
567
568    /// Resolve a `host:port` string to a tailnet [`SocketAddr`], honoring the family forced by a
569    /// `network` suffix. The host may be an IP literal (parsed directly) or a MagicDNS name
570    /// (resolved via [`Device::resolve`], which yields a tailnet IPv4). Shared by [`Device::dial`]
571    /// and [`Device::dial_tcp`]. The IPv4-only invariant is enforced here: a `…6` network, or any v6
572    /// destination, requires `Config::enable_ipv6` and otherwise returns
573    /// [`InternalErrorKind::BadRequest`] (a clean typed error rather than a downstream actor error).
574    async fn resolve_dial_addr(
575        &self,
576        network: dial::Network,
577        addr: &str,
578    ) -> Result<SocketAddr, Error> {
579        let (host, port) = dial::split_host_port(addr)?;
580
581        // An IP literal is used directly; otherwise resolve the MagicDNS name (IPv4 only).
582        let ip: IpAddr = if let Ok(ip) = host.parse::<IpAddr>() {
583            ip
584        } else {
585            self.resolve(host)
586                .await?
587                .ok_or(Error::Internal(InternalErrorKind::BadRequest))?
588                .into()
589        };
590
591        dial::check_family(network.family, ip)?;
592
593        // IPv4-only invariant: a v6 destination is only reachable when IPv6 is provisioned.
594        if ip.is_ipv6() && !self.enable_ipv6 {
595            return Err(Error::Internal(InternalErrorKind::BadRequest));
596        }
597
598        Ok((ip, port).into())
599    }
600
601    /// Connect to a tailnet address over TCP or UDP, the Rust analog of Go
602    /// `tsnet.Server.Dial(ctx, network, address)`.
603    ///
604    /// `network` is one of `"tcp"`, `"tcp4"`, `"tcp6"`, `"udp"`, `"udp4"`, `"udp6"`; `addr` is a
605    /// `host:port` string where `host` is a MagicDNS name, an IPv4 literal, or a bracketed IPv6
606    /// literal (`[2001:db8::1]:443`). The host is resolved in-process via [`Device::resolve`] (no DNS
607    /// server). Returns a [`DialConn`] whose arm matches the transport — use [`Device::dial_tcp`]
608    /// when you want the TCP stream directly.
609    ///
610    /// Differences from Go (documented for parity): ports must be **numeric** (Go's `LookupPort`
611    /// also resolves named ports like `"http"`; this fork avoids a services-file dependency), and
612    /// `…6`/v6 destinations require `Config::enable_ipv6` (the tailnet is IPv4-only by default).
613    ///
614    /// # Errors
615    /// [`InternalErrorKind::BadRequest`] for an unsupported `network`, a malformed/portless `addr`,
616    /// an unresolvable name, or a v6 destination while IPv6 is disabled; otherwise the transport's
617    /// own connect error.
618    pub async fn dial(&self, network: &str, addr: &str) -> Result<DialConn, Error> {
619        let net = dial::parse_network(network)?;
620        let remote = self.resolve_dial_addr(net, addr).await?;
621
622        match net.transport {
623            dial::Transport::Tcp => Ok(DialConn::Tcp(self.tcp_connect(remote).await?)),
624            dial::Transport::Udp => {
625                // Bind an ephemeral local UDP socket on this node's tailnet address of the SAME
626                // family as the remote, then connect it (Go's `Dial("udp", …)` returns a connected
627                // UDP `net.Conn`, with the local source picked by `IfElse(dst.Is6(), v6, v4)`). A v4
628                // local socket cannot send to a v6 peer, so the family must match `remote`. (TCP gets
629                // this for free: `tcp_connect` already picks the source family from `remote`.)
630                let local_ip: IpAddr = if remote.is_ipv6() {
631                    self.ipv6_addr().await?.into()
632                } else {
633                    self.ipv4_addr().await?.into()
634                };
635                let sock = self.udp_bind((local_ip, 0).into()).await?;
636                Ok(DialConn::Udp(ConnectedUdpSocket::new(sock, remote)))
637            }
638        }
639    }
640
641    /// Connect to a tailnet address over TCP, returning the stream directly — the common case of
642    /// [`Device::dial`] for `"tcp"`. `addr` is a `host:port` string (MagicDNS name or IP literal).
643    /// This is the building block for HTTP-over-tailnet: an embedder's `hyper`/`reqwest` client can
644    /// route requests by calling `dial_tcp(&format!("{host}:{port}"))` from its connector, mirroring
645    /// how Go `tsnet.Server.HTTPClient` sets `http.Transport.DialContext = Server.Dial`.
646    ///
647    /// # Errors
648    /// As [`Device::dial`] for the `"tcp"` network.
649    pub async fn dial_tcp(&self, addr: &str) -> Result<netstack::TcpStream, Error> {
650        let remote = self
651            .resolve_dial_addr(
652                dial::Network {
653                    transport: dial::Transport::Tcp,
654                    family: dial::Family::Any,
655                },
656                addr,
657            )
658            .await?;
659        self.tcp_connect(remote).await
660    }
661
662    /// Connect to a tailnet address over UDP, returning a connected socket directly — the `"udp"`
663    /// sibling of [`dial_tcp`](Device::dial_tcp) and the common case of [`Device::dial`] for
664    /// `"udp"`. `addr` is a `host:port` string (MagicDNS name or IP literal).
665    ///
666    /// Returns a [`ConnectedUdpSocket`] (`send`/`recv` against a fixed peer), the connected
667    /// UDP-`net.Conn` shape Go's `tsnet.Server.Dial("udp", …)` returns — as opposed to
668    /// [`listen_packet`](Device::listen_packet), which yields an unconnected `net.PacketConn`. An
669    /// ephemeral local UDP socket is bound on this node's tailnet address of the same family as the
670    /// resolved remote (a v4 local socket cannot send to a v6 peer).
671    ///
672    /// # Errors
673    /// As [`Device::dial`] for the `"udp"` network (name resolution, the IPv4-only / `enable_ipv6`
674    /// family invariant, or TUN transport mode having no application netstack to bind on).
675    pub async fn dial_udp(&self, addr: &str) -> Result<ConnectedUdpSocket, Error> {
676        let remote = self
677            .resolve_dial_addr(
678                dial::Network {
679                    transport: dial::Transport::Udp,
680                    family: dial::Family::Any,
681                },
682                addr,
683            )
684            .await?;
685        let local_ip: IpAddr = if remote.is_ipv6() {
686            self.ipv6_addr().await?.into()
687        } else {
688            self.ipv4_addr().await?.into()
689        };
690        let sock = self.udp_bind((local_ip, 0).into()).await?;
691        Ok(ConnectedUdpSocket::new(sock, remote))
692    }
693
694    /// Bind a UDP socket from a `host:port` string, the Rust analog of Go
695    /// `tsnet.Server.ListenPacket(network, addr)`.
696    ///
697    /// `network` is one of `"udp"`, `"udp4"`, `"udp6"`; `addr` must be a **valid IP literal**
698    /// `host:port` (Go's `ListenPacket` rejects a name or empty host — unlike `Listen`). An
699    /// unspecified host (`0.0.0.0`/`[::]`) binds on this node's tailnet address. Returns the
700    /// unconnected [`netstack::UdpSocket`] (a `net.PacketConn`).
701    ///
702    /// # Errors
703    /// [`InternalErrorKind::BadRequest`] for a non-UDP/unsupported `network`, a malformed addr, a
704    /// non-IP host, a family mismatch, or a v6 bind while IPv6 is disabled.
705    pub async fn listen_packet(
706        &self,
707        network: &str,
708        addr: &str,
709    ) -> Result<netstack::UdpSocket, Error> {
710        let net = dial::parse_network(network)?;
711        if net.transport != dial::Transport::Udp {
712            return Err(Error::Internal(InternalErrorKind::BadRequest));
713        }
714        let (host, port) = dial::split_host_port(addr)?;
715
716        // ListenPacket requires a valid IP host (Go rejects a name here).
717        let ip: IpAddr = host
718            .parse()
719            .map_err(|_| Error::Internal(InternalErrorKind::BadRequest))?;
720        dial::check_family(net.family, ip)?;
721
722        // A v6 bind (whether an explicit literal or an unspecified `[::]`) requires IPv6 to be
723        // provisioned — enforce the gate for BOTH cases (the unspecified `[::]` path used to skip it).
724        if ip.is_ipv6() && !self.enable_ipv6 {
725            return Err(Error::Internal(InternalErrorKind::BadRequest));
726        }
727
728        // An unspecified bind host (`0.0.0.0` / `[::]`) means "this node's tailnet address" — of the
729        // SAME family as the requested address, so a `udp6` `[::]:0` binds a v6 socket (it used to
730        // fall through to the v4 address regardless, silently yielding an IPv4 socket for a v6 listen).
731        let bind_ip: IpAddr = if ip.is_unspecified() {
732            if ip.is_ipv6() {
733                self.ipv6_addr().await?.into()
734            } else {
735                self.ipv4_addr().await?.into()
736            }
737        } else {
738            ip
739        };
740
741        self.udp_bind((bind_ip, port).into()).await
742    }
743
744    /// Connect to a TCP socket at the remote address.
745    ///
746    /// Returns an error in TUN transport mode (there is no application netstack to dial from).
747    pub async fn tcp_connect(&self, remote: SocketAddr) -> Result<netstack::TcpStream, Error> {
748        let channel = self.channel()?;
749
750        let ip: IpAddr = match remote.is_ipv4() {
751            true => self.ipv4_addr().await?.into(),
752            false => self.ipv6_addr().await?.into(),
753        };
754
755        // TODO(npry): collision checking
756        let ephemeral_port = rand::random_range(49152..=u16::MAX);
757
758        channel
759            .tcp_connect((ip, ephemeral_port).into(), remote)
760            .await
761            .map_err(Into::into)
762    }
763
764    /// Start a SOCKS5 proxy on a host loopback address that dials into the tailnet (Go
765    /// `tsnet.Server.Loopback`, SOCKS5 half).
766    ///
767    /// Binds a TCP listener on `127.0.0.1:0` (host loopback only — never an external interface) and
768    /// serves SOCKS5 (RFC 1928) with required username/password auth (RFC 1929): username `tsnet`,
769    /// password = the returned `proxy_cred`. Each `CONNECT` is dialed INTO the overlay via
770    /// [`Device::connect_by_name`] / [`Device::tcp_connect`] and spliced to the accepted host socket, so
771    /// a non-Rust host process can reach tailnet peers through the proxy. Returns the bound address, the
772    /// proxy credential, and a [`LoopbackHandle`] whose drop stops the listener.
773    ///
774    /// Anti-leak: the listener is loopback-only and every connection egresses over the overlay, never a
775    /// host socket — the host's real origin IP is never used to reach the destination. Unlike Go, the
776    /// LocalAPI HTTP surface is not served (this fork exposes status/whois/id-token natively on
777    /// `Device`); only the SOCKS5 proxy is provided.
778    ///
779    /// Returns an error in TUN transport mode (no application netstack to dial from).
780    pub async fn loopback(&self) -> Result<(std::net::SocketAddr, String, LoopbackHandle), Error> {
781        loopback::start(self.overlay_dialer().await?).await
782    }
783
784    /// Build an [`OverlayDialer`](loopback::OverlayDialer): the cloneable, `&Device`-free dialer that
785    /// resolves a MagicDNS name (or takes an IPv4 literal) and `tcp_connect`s it into the overlay,
786    /// reused by [`Device::loopback`] (SOCKS5) and the `hyper` [`http_connector`](Device::http_connector).
787    ///
788    /// Captures only cloneable pieces — never `&self` — so the dialer (and anything built on it, like a
789    /// spawned accept loop or an HTTP connector) carries no borrow of the `Device`: a clone of the
790    /// netstack command channel, this device's own overlay IPv4 (fetched once), and a boxed resolver
791    /// closure over clones of the control + peer-tracker actor refs. The resolver replicates
792    /// [`Device::resolve`] (peer-by-name, falling back to this node's own name).
793    async fn overlay_dialer(&self) -> Result<loopback::OverlayDialer, Error> {
794        let channel = self.channel()?.clone();
795        let self_ipv4 = self.ipv4_addr().await?;
796
797        let control = self.runtime.control.clone();
798        let peer_tracker = self.runtime.peer_tracker.clone();
799        let resolve: loopback::Resolver = std::sync::Arc::new(move |name: String| {
800            let control = control.clone();
801            let peer_tracker = peer_tracker.clone();
802            Box::pin(async move {
803                let pt = peer_tracker
804                    .upgrade()
805                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
806                let peer = pt
807                    .ask(ts_runtime::peer_tracker::PeerByName { name: name.clone() })
808                    .await
809                    .map_err(ts_runtime::Error::from)?;
810                if let Some(peer) = peer {
811                    return Ok(Some(peer.tailnet_address.ipv4.addr()));
812                }
813                // tsnet's dnsMap also resolves our own name; fall back to self.
814                let me = control
815                    .ask(ts_runtime::control_runner::SelfNode)
816                    .await
817                    .map_err(ts_runtime::Error::from)?
818                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
819                if me.matches_name(&name) {
820                    Ok(Some(me.tailnet_address.ipv4.addr()))
821                } else {
822                    Ok(None)
823                }
824            }) as std::pin::Pin<Box<dyn std::future::Future<Output = _> + Send>>
825        });
826
827        Ok(loopback::OverlayDialer::new(channel, self_ipv4, resolve))
828    }
829
830    /// Build a [`hyper`-compatible connector](crate::http::TailnetConnector) that routes outbound HTTP
831    /// requests over the tailnet — the analog of Go `tsnet.Server.HTTPClient`, whose mechanism is
832    /// simply `http.Transport{DialContext: s.Dial}`.
833    ///
834    /// Hand the returned connector to `hyper_util::client::legacy::Client::builder(...).build(conn)`;
835    /// each request's `Uri` host is resolved as a MagicDNS name (or IPv4 literal) and dialed into the
836    /// overlay (default port 80 for `http`, 443 for `https`), so the request egresses over the tailnet
837    /// rather than the host's network. TLS, redirects, and pooling are the hyper client's concern — the
838    /// connector only supplies the transport, exactly like Go's bare `DialContext` injection.
839    ///
840    /// Available only with the **`hyper`** crate feature.
841    ///
842    /// # Errors
843    /// Fails for the same reasons as [`Device::loopback`]'s setup: TUN transport mode (no application
844    /// netstack) or the node not yet having an overlay IPv4.
845    #[cfg(feature = "hyper")]
846    pub async fn http_connector(&self) -> Result<crate::http::TailnetConnector, Error> {
847        Ok(crate::http::TailnetConnector::new(
848            self.overlay_dialer().await?,
849        ))
850    }
851
852    /// Get our node info.
853    pub async fn self_node(&self) -> Result<NodeInfo, Error> {
854        self.runtime
855            .control
856            .ask(ts_runtime::control_runner::SelfNode)
857            .await
858            .map_err(ts_runtime::Error::from)?
859            .ok_or(Error::Internal(InternalErrorKind::Actor))
860    }
861
862    /// The DNS names this node can obtain TLS certificates for — Go `tsnet.Server.CertDomains()`.
863    ///
864    /// These are the `CertDomains` control pushed in the netmap DNS config: the names a TLS-serving
865    /// consumer (e.g. a `ListenTLS`/`GetCertificate`-style caller) should request a cert for. Returns
866    /// an empty `Vec` before the first netmap, or when control granted none — mirroring Go returning a
867    /// clone of `nm.DNS.CertDomains` (empty/`nil` when absent).
868    pub async fn cert_domains(&self) -> Result<Vec<String>, Error> {
869        self.runtime
870            .control
871            .ask(ts_runtime::control_runner::CertDomains)
872            .await
873            .map_err(ts_runtime::Error::from)
874            .map_err(Into::into)
875    }
876
877    /// The DNS configuration control pushed in the latest netmap — Go `tsnet`'s view of
878    /// `netmap.NetworkMap.DNS` (what `tailscale dns status` reports).
879    ///
880    /// Returns the full [`DnsConfig`] — MagicDNS on/off, search domains, global + fallback resolvers,
881    /// split-DNS routes, extra records, cert domains — or `None` before the first netmap / when
882    /// control has sent no DNS config. A superset of [`cert_domains`](Device::cert_domains), which
883    /// remains a separate narrower accessor for the TLS-cert use. Mirrors Go reading a clone of
884    /// `nm.DNS` (absent ⇒ `None`).
885    pub async fn dns_config(&self) -> Result<Option<DnsConfig>, Error> {
886        self.runtime
887            .control
888            .ask(ts_runtime::control_runner::DnsConfig)
889            .await
890            .map_err(ts_runtime::Error::from)
891            .map_err(Into::into)
892    }
893
894    /// The URL control last asked this node to open in a browser (`MapResponse.PopBrowserURL`), or
895    /// `None` if control has never sent one.
896    ///
897    /// This is the interactive-login / consent URL an embedder driving a non-authkey (interactive)
898    /// login must surface to the user — the Rust analog of Go `ipn` delivering `BrowseToURL` through
899    /// the notification bus. A daemon polls this after starting an interactive login to obtain the
900    /// auth URL to present.
901    ///
902    /// **Sticky semantics** (Go `controlclient`'s `sess.lastPopBrowserURL`): once control sends a
903    /// URL it remains the returned value until control sends a *different* non-empty one — it is
904    /// **never cleared back to `None`** (control sends `PopBrowserURL` empty on nearly every netmap
905    /// tick; those empty updates are ignored, not treated as "clear"). So a non-`None` result does
906    /// **not** signal "control is asking *right now*" vs. "already handled" — it is the last URL
907    /// seen this session. A consumer that acts on it should de-duplicate on the URL value rather than
908    /// re-acting on every poll. For a push stream of *new* consent URLs (rather than polling this
909    /// sticky value), subscribe to [`watch_ipn_bus`](Self::watch_ipn_bus) and react to
910    /// [`Notify::browse_to_url`](crate::Notify::browse_to_url).
911    pub async fn pop_browser_url(&self) -> Result<Option<Url>, Error> {
912        self.runtime
913            .control
914            .ask(ts_runtime::control_runner::PopBrowserUrl)
915            .await
916            .map_err(ts_runtime::Error::from)
917            .map_err(Into::into)
918    }
919
920    /// This node's latest network-conditions report — the Rust analog of Go's `netcheck.Report` as
921    /// `tailscale netcheck` surfaces it.
922    ///
923    /// Returns the [`NetcheckReport`]: the preferred (lowest-latency) DERP region and the per-region
924    /// latency map this node last measured. Empty (default) before the first measurement. This fork's
925    /// net-report path measures only DERP-region latency, so the report carries that subset rather
926    /// than fabricating the UDP/port-mapping fields Go also reports (see [`NetcheckReport`]).
927    pub async fn netcheck(&self) -> Result<NetcheckReport, Error> {
928        self.runtime
929            .control
930            .ask(ts_runtime::control_runner::Netcheck)
931            .await
932            .map_err(ts_runtime::Error::from)
933            .map_err(Into::into)
934    }
935
936    /// Suggest a reasonably good exit node to use, based on this node's current netmap and latest
937    /// network-conditions report — Go `tailscale exit-node suggest` / `LocalBackend.SuggestExitNode`.
938    ///
939    /// Returns the suggested exit node's stable id + name as an [`ExitNodeSuggestion`]; engage it by
940    /// passing the id to [`Config::exit_node`](crate::config::Config) /
941    /// [`Device::set_exit_node`](crate::Device::set_exit_node) as a stable-id selector. The
942    /// suggestion uses the classic DERP-region-latency algorithm: among peers control marked
943    /// suggestable (the `suggest-exit-node` capability) that advertise an exit route and are online,
944    /// it prefers the one whose home DERP region this node measured as lowest-latency, and is
945    /// **sticky** — a prior suggestion that is still a good candidate is kept across calls, so
946    /// repeated calls don't flap between equally-good options.
947    ///
948    /// Outcomes (mirroring Go):
949    /// - `Ok(Some(suggestion))` — a node was suggested.
950    /// - `Ok(None)` — no eligible candidate (no suggestion); **not** an error.
951    /// - `Err(`[`Error::NoPreferredDerp`]`)` — no netcheck has completed yet, so no preferred DERP
952    ///   region is known; retry once connectivity has been measured.
953    ///
954    /// ## Scope (Phase 1)
955    /// This ports Go's classic DERP path only. The traffic-steering path and the Mullvad
956    /// geographic-distance ranking (for exit nodes with no DERP home) are not yet implemented, and
957    /// the suggestion does not carry a `Location` (Go's `omitempty` field) — this fork's peer model
958    /// has none yet. The candidate exit-route check accepts a peer advertising `0.0.0.0/0` (this
959    /// fork is IPv4-only), rather than Go's both-`0.0.0.0/0`-and-`::/0` requirement.
960    pub async fn suggest_exit_node(&self) -> Result<Option<ExitNodeSuggestion>, Error> {
961        // The runtime returns the actor-gather outcome (outer) wrapping the algorithm outcome
962        // (inner: `Ok(None)` empty, or the `NoPreferredDerp` domain error). Flatten both into the
963        // device-facing `Error`.
964        self.runtime.suggest_exit_node().await?.map_err(Into::into)
965    }
966
967    /// This node's key-expiry instant as Unix seconds (`Node.KeyExpiry` in Go), or `Ok(None)` if
968    /// the key never expires.
969    ///
970    /// Like Go, this fork is **reactive** about key expiry — it reports it rather than rotating the
971    /// node key in the background. A caller can schedule re-authentication around this time; on
972    /// expiry, re-create the [`Device`] (which re-registers), supplying a fresh node key + the prior
973    /// `old_node_key` to rotate, or the same key to refresh.
974    pub async fn self_key_expiry_unix(&self) -> Result<Option<i64>, Error> {
975        Ok(self.self_node().await?.key_expiry_unix())
976    }
977
978    /// Whether this node's key has expired as of now (`!KeyExpiry.IsZero() && KeyExpiry.Before(now)`
979    /// in Go). A key with no expiry is never expired. See [`Device::self_key_expiry_unix`] for the
980    /// reactive-rotation note.
981    pub async fn self_key_expired(&self) -> Result<bool, Error> {
982        let now = std::time::SystemTime::now()
983            .duration_since(std::time::UNIX_EPOCH)
984            .map(|d| d.as_secs() as i64)
985            // An unreadable clock (pre-epoch) is treated as the far future so a time-limited key
986            // looks expired — fail-safe toward prompting re-auth rather than trusting a stale key.
987            .unwrap_or(i64::MAX);
988        Ok(self.self_node().await?.key_expired_at_unix(now))
989    }
990
991    /// Fetch the current Tailscale SSH policy pushed by control, if any.
992    ///
993    /// Returns `Ok(None)` when control has not sent an SSH policy. The SSH server treats an absent
994    /// or empty policy as **deny-all** (fail-closed). Used by the SSH auth path
995    /// ([`SshPolicy::evaluate`][ts_control::SshPolicy::evaluate]) to authorize incoming
996    /// connections.
997    pub async fn ssh_policy(&self) -> Result<Option<ts_control::SshPolicy>, Error> {
998        self.runtime
999            .control
1000            .ask(ts_runtime::control_runner::CurrentSshPolicy)
1001            .await
1002            .map_err(ts_runtime::Error::from)
1003            .map_err(Into::into)
1004    }
1005
1006    /// Look up a peer by name.
1007    pub async fn peer_by_name(&self, name: &str) -> Result<Option<NodeInfo>, Error> {
1008        let pt = self
1009            .runtime
1010            .peer_tracker
1011            .upgrade()
1012            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
1013
1014        pt.ask(ts_runtime::peer_tracker::PeerByName {
1015            name: name.to_string(),
1016        })
1017        .await
1018        .map_err(ts_runtime::Error::from)
1019        .map_err(Into::into)
1020    }
1021
1022    /// Look up a peer by ip.
1023    pub async fn peer_by_tailnet_ip(&self, ip: IpAddr) -> Result<Option<NodeInfo>, Error> {
1024        let pt = self
1025            .runtime
1026            .peer_tracker
1027            .upgrade()
1028            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
1029
1030        pt.ask(ts_runtime::peer_tracker::PeerByTailnetIp { ip })
1031            .await
1032            .map_err(ts_runtime::Error::from)
1033            .map_err(Into::into)
1034    }
1035
1036    /// Look up the peer(s) with the most-specific route matches for `ip`.
1037    ///
1038    /// This reports which peers *advertise* a route covering `ip`, independent of this device's
1039    /// `accept_routes` setting — analogous to the Go client's informational `PrimaryRoutes`. It is
1040    /// not a reachability oracle: with `accept_routes` off, the dataplane will not actually route
1041    /// to (or accept return traffic from) advertised subnet routes even if this returns a peer.
1042    pub async fn peers_with_route(&self, ip: IpAddr) -> Result<Vec<NodeInfo>, Error> {
1043        let pt = self
1044            .runtime
1045            .peer_tracker
1046            .upgrade()
1047            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
1048
1049        pt.ask(ts_runtime::peer_tracker::PeerByAcceptedRoute { ip })
1050            .await
1051            .map_err(ts_runtime::Error::from)
1052            .map_err(Into::into)
1053    }
1054
1055    /// List the Taildrop files this device has fully received and not yet consumed (Go LocalAPI
1056    /// `WaitingFiles`).
1057    ///
1058    /// Returns the files waiting under the configured `taildrop_dir`, sorted by name. Returns an
1059    /// empty list when Taildrop is disabled (`Config::taildrop_dir` unset) — fail-closed, never an
1060    /// error for the disabled case. A filesystem error while listing surfaces as
1061    /// [`InternalErrorKind::Actor`].
1062    pub fn taildrop_waiting_files(&self) -> Result<Vec<WaitingFile>, Error> {
1063        let Some(store) = self.runtime.taildrop_store() else {
1064            return Ok(Vec::new());
1065        };
1066        store
1067            .waiting_files()
1068            .map_err(|_| Error::Internal(InternalErrorKind::Actor))
1069    }
1070
1071    /// Open a received Taildrop file by name for reading, returning the handle and its size (Go
1072    /// LocalAPI `OpenFile`).
1073    ///
1074    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
1075    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
1076    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
1077    pub fn taildrop_open_file(&self, name: &str) -> Result<(std::fs::File, u64), Error> {
1078        let store = self
1079            .runtime
1080            .taildrop_store()
1081            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
1082        store.open_file(name).map_err(taildrop_err)
1083    }
1084
1085    /// Delete a received Taildrop file by name (Go LocalAPI `DeleteFile`).
1086    ///
1087    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
1088    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
1089    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
1090    pub fn taildrop_delete_file(&self, name: &str) -> Result<(), Error> {
1091        let store = self
1092            .runtime
1093            .taildrop_store()
1094            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
1095        store.delete_file(name).map_err(taildrop_err)
1096    }
1097
1098    /// Send a local file to a tailnet `peer` via Taildrop (Go `PushFile` / `tailscale file cp`).
1099    ///
1100    /// Pushes `content_length` bytes from `reader` to the peer's peerAPI as
1101    /// `PUT /v0/put/<name>` over the overlay netstack — the sending counterpart to the receive store
1102    /// surfaced by [`Device::taildrop_waiting_files`]. The transfer rides the encrypted WireGuard
1103    /// overlay, never a host socket. The body is streamed from offset 0 (no resume).
1104    ///
1105    /// The destination is derived **solely from `peer`'s own node record**
1106    /// ([`NodeInfo::peerapi_addr`][ts_control::Node::peerapi_addr]): its advertised tailnet IPv4 and
1107    /// `peerapi4` port. The caller obtains `peer` from [`Device::peer_by_name`] /
1108    /// [`Device::peer_by_tailnet_ip`], so it is always a current netmap peer — a raw control-supplied
1109    /// or attacker-chosen address can never be targeted. As defense in depth, the resolved address is
1110    /// additionally asserted to be a Tailscale CGNAT IP before dialing.
1111    ///
1112    /// Returns [`InternalErrorKind::BadRequest`] when the peer advertises no IPv4 peerAPI (so it
1113    /// cannot receive files), when the name is invalid, or when the peer refuses the transfer
1114    /// (`403`/`409`/unexpected status); [`Error::Timeout`] on a dial failure or timeout; and
1115    /// [`InternalErrorKind::Io`] on a mid-transfer stream error.
1116    pub async fn send_file<R>(
1117        &self,
1118        peer: &NodeInfo,
1119        name: &str,
1120        content_length: u64,
1121        reader: R,
1122    ) -> Result<(), Error>
1123    where
1124        R: tokio::io::AsyncRead + Unpin,
1125    {
1126        let channel = self.channel()?;
1127
1128        // Destination comes only from the peer's own node record — never an arbitrary address.
1129        let dst = peer
1130            .peerapi_addr()
1131            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
1132        // Defense in depth: refuse to dial anything outside the Tailscale CGNAT range, so a
1133        // malformed node record can't steer the PUT at a non-tailnet host.
1134        if !ts_control::is_tailscale_ip(dst.ip()) {
1135            return Err(Error::Internal(InternalErrorKind::BadRequest));
1136        }
1137
1138        let self_ipv4 = self.ipv4_addr().await?;
1139
1140        ts_runtime::taildrop_send::send_file(channel, self_ipv4, dst, name, content_length, reader)
1141            .await
1142            .map_err(taildrop_send_err)
1143    }
1144
1145    /// List the tailnet peers this node can Taildrop a file *to* — the Rust analog of Go's LocalAPI
1146    /// `FileTargets`.
1147    ///
1148    /// Each [`FileTarget`] pairs a peer's node record with the `http://ip:port` base of its peerAPI;
1149    /// pass `target.node` straight to [`Device::send_file`]. A peer qualifies when it advertises a
1150    /// reachable IPv4 peerAPI **and** is either owned by the same user as this node **or** explicitly
1151    /// granted the file-sharing-target capability — mirroring upstream's send-path filter. The list is
1152    /// gated on this node holding the file-sharing capability (control grants it when the admin
1153    /// enables Taildrop); absent that, the result is empty (fail-closed, not an error). Sorted by the
1154    /// peer's MagicDNS name. Targets are listed regardless of online state (matching upstream — an
1155    /// offline target's [`send_file`](Device::send_file) simply times out). Empty before the first
1156    /// netmap.
1157    pub async fn file_targets(&self) -> Result<Vec<FileTarget>, Error> {
1158        self.runtime.file_targets().await.map_err(Into::into)
1159    }
1160
1161    /// Begin a debug packet capture, streaming a pcap of every packet crossing the dataplane to
1162    /// `writer` (Go `tsnet.Server.CapturePcap`).
1163    ///
1164    /// Installs a capture hook on the running dataplane: from now until [`Device::stop_capture`] is
1165    /// called (or another capture replaces this one), a copy of every plaintext IP packet on the
1166    /// datapath — outbound (pre-encrypt) and inbound (post-decrypt) — is framed and written to
1167    /// `writer`. The 24-byte pcap global header is written immediately on success.
1168    ///
1169    /// The format is byte-faithful classic pcap with Tailscale's `LINKTYPE_USER0` + 4-byte path
1170    /// preamble per record (see [`ts_runtime::capture`]); a resulting file opens in Wireshark, and
1171    /// with Tailscale's `ts-dissector.lua` the direction/path of each packet decodes.
1172    ///
1173    /// The hook runs **inline on the single-threaded dataplane step**, so `writer` must not block for
1174    /// long — a slow writer back-pressures the datapath. Records are **not** flushed per packet (that
1175    /// would be a syscall on every packet on the dataplane thread); buffered bytes are flushed when
1176    /// the writer is dropped on [`Device::stop_capture`]. Wrap `writer` in a [`std::io::BufWriter`] if
1177    /// you want buffering. A write error is swallowed per-packet (the capture silently drops that
1178    /// record) rather than tearing down the datapath; call [`Device::stop_capture`] to end it. Returns
1179    /// an error only if the dataplane actor is unreachable or the initial global-header write fails.
1180    pub async fn capture_pcap<W>(&self, writer: W) -> Result<(), Error>
1181    where
1182        W: std::io::Write + Send + 'static,
1183    {
1184        let sink = std::sync::Arc::new(std::sync::Mutex::new(
1185            ts_runtime::capture::PcapSink::new(writer)
1186                .map_err(|_| Error::Internal(InternalErrorKind::Io))?,
1187        ));
1188        let hook: ts_runtime::CaptureHook = std::sync::Arc::new(move |path, pkt: &[u8]| {
1189            if let Ok(mut sink) = sink.lock() {
1190                // A per-packet write failure (e.g. a closed pipe) silently drops that record rather
1191                // than tearing down the datapath; the caller ends capture via `stop_capture`.
1192                drop(sink.log_packet(path.code(), pkt));
1193            }
1194        });
1195        self.runtime.install_capture(Some(hook)).await?;
1196        Ok(())
1197    }
1198
1199    /// Stop a debug packet capture started by [`Device::capture_pcap`] (Go `ClearCaptureSink`).
1200    ///
1201    /// Clears the dataplane capture hook; the writer is dropped (its remaining buffered bytes are
1202    /// flushed by its own `Drop`). Idempotent — clearing when no capture is installed is a no-op.
1203    /// Returns an error only if the dataplane actor is unreachable.
1204    pub async fn stop_capture(&self) -> Result<(), Error> {
1205        self.runtime.install_capture(None).await?;
1206        Ok(())
1207    }
1208
1209    /// Snapshot of this device and its tailnet peers (like `tailscale status`).
1210    ///
1211    /// Combines this node's self info with the current peer set: each [`StatusNode`] reports the
1212    /// stable id, display name, tailnet IPs, advertised routes, and exit-node flag. (Per-peer
1213    /// `online`/user/capabilities are honestly `None`/empty in this fork — the domain node model
1214    /// does not yet carry the wire-level liveness/login fields; see `ts_runtime::status` docs.)
1215    pub async fn status(&self) -> Result<Status, Error> {
1216        self.runtime.status().await.map_err(Into::into)
1217    }
1218
1219    /// Fetch the current Tailnet Lock (TKA) status pushed by control, if any.
1220    ///
1221    /// Returns `Ok(None)` when control has sent no `TKAInfo` (tailnet lock not in use, or no change
1222    /// observed yet). The returned [`TkaStatus`][ts_control::TkaStatus] carries the authority head
1223    /// (a base32 `AUMHash`, decode with [`tka::AumHash::from_base32`][ts_tka::AumHash::from_base32])
1224    /// and the disablement signal. Signature verification of a peer's node-key signature against the
1225    /// authority is performed with the [`tka`] module's [`tka::Authority`].
1226    pub async fn tka_status(&self) -> Result<Option<ts_control::TkaStatus>, Error> {
1227        self.runtime
1228            .control
1229            .ask(ts_runtime::control_runner::CurrentTkaStatus)
1230            .await
1231            .map_err(ts_runtime::Error::from)
1232            .map_err(Into::into)
1233    }
1234
1235    /// Read the Tailnet Lock update-chain history — the Rust analog of Go
1236    /// `LocalClient.NetworkLockLog`.
1237    ///
1238    /// Returns up to `limit` [`TkaLogEntry`] rows of the AUM chain **head-first** (newest first,
1239    /// walking back toward the genesis), read **locally** from this node's synced + verified chain —
1240    /// a pure read with no control round-trip. The list is empty when no lock is synced (lock not in
1241    /// use, or control hasn't pushed a chain yet). Each entry carries the AUM's chain-link hash, its
1242    /// change kind (`"add-key"` / `"remove-key"` / `"checkpoint"` / …), the ids of the keys that
1243    /// signed it, and the raw CBOR (Go `NetworkLockUpdate.Raw`) for a faithful full decode.
1244    pub async fn tka_log(&self, limit: usize) -> Result<Vec<TkaLogEntry>, Error> {
1245        self.runtime.tka_log(limit).await.map_err(Into::into)
1246    }
1247
1248    /// Sign a peer's `node_key` with this node's network-lock key and submit the signature to
1249    /// control — the Rust analog of Go `LocalClient.NetworkLockSign` for the Direct case.
1250    ///
1251    /// Builds a `Direct` [`NodeKeySignature`][ts_tka::NodeKeySignature] authorizing `node_key`, signed
1252    /// by this node's network-lock private key, and POSTs it to `/machine/tka/sign`. The signing node
1253    /// must itself be trusted under the current authority for control to accept the signature.
1254    ///
1255    /// **This only *submits* the signature; it does not mutate this node's local
1256    /// [`Authority`][ts_tka::Authority].** The local trusted-key state advances solely through the
1257    /// verified netmap-driven sync path (every applied AUM passes
1258    /// [`VerifiedAumChain::verify`][ts_tka::VerifiedAumChain::verify]), so a successful `tka_sign` is
1259    /// reflected locally on the next sync — the active fail-closed enforcement posture is unchanged.
1260    ///
1261    /// # Errors
1262    /// [`ts_control::TkaSyncError::Unsupported`] if control has no TKA endpoint (no lock / control too
1263    /// old), [`ts_control::TkaSyncError::NetworkError`] on a transient failure, or a coarse
1264    /// `Internal` for other RPC failures.
1265    pub async fn tka_sign(
1266        &self,
1267        node_key: &ts_keys::NodePublicKey,
1268    ) -> Result<(), ts_control::TkaSyncError> {
1269        self.runtime.tka_sign(node_key.to_bytes()).await
1270    }
1271
1272    /// Disable Tailnet Lock by presenting the `disablement_secret` to control — the Rust analog of
1273    /// Go `LocalClient.NetworkLockDisable`.
1274    ///
1275    /// Targets this node's current authority head (from the cached [`tka_status`](Device::tka_status));
1276    /// the `disablement_secret` is the operator-held capability (one of the lock's
1277    /// `DisablementValues`) that authorizes turning the lock off. Control verifies the secret against
1278    /// the authority's disablement set and, if valid, disables the lock for the tailnet.
1279    ///
1280    /// **Submit-only:** this POSTs the disablement; it does not mutate this node's local
1281    /// [`Authority`][ts_tka::Authority]. The disablement is reflected locally through the existing
1282    /// verified netmap-driven sync — which then clears enforcement to admit-all. The active
1283    /// fail-closed enforcement posture (until that sync lands) is unchanged.
1284    ///
1285    /// # Errors
1286    /// [`ts_control::TkaSyncError::Unsupported`] when there is no known TKA head to disable (lock not
1287    /// in use / control hasn't pushed a status) or control has no TKA endpoint;
1288    /// [`ts_control::TkaSyncError::NetworkError`] on a transient failure; a coarse `Internal` for
1289    /// other RPC failures (incl. control rejecting an invalid secret).
1290    pub async fn tka_disable(
1291        &self,
1292        disablement_secret: Vec<u8>,
1293    ) -> Result<(), ts_control::TkaSyncError> {
1294        self.runtime.tka_disable(disablement_secret).await
1295    }
1296
1297    /// Initialize Tailnet Lock for this tailnet with this node as the sole initial trusted key — the
1298    /// Rust analog of Go `LocalClient.NetworkLockInit` for the single-node "lock yourself in" case.
1299    ///
1300    /// Builds and signs a genesis Checkpoint AUM trusting only this node's network-lock key and
1301    /// gated by `disablement_secret` (stored as its Argon2i [`disablement_value`][ts_tka::disablement_value]
1302    /// in the lock; the raw secret is the operator-held capability that later disables it via
1303    /// [`tka_disable`](Device::tka_disable)), then drives control's two-phase
1304    /// `/machine/tka/init/{begin,finish}`.
1305    ///
1306    /// **Single-node only (for now):** if control reports that other nodes must be (re)signed under
1307    /// the new lock (a multi-node tailnet), this returns [`ts_control::TkaSyncError::Unsupported`] —
1308    /// the multi-node init (re-signing each node, incl. rotation keys) is a deferred follow-up.
1309    ///
1310    /// **Submit-only:** this creates the lock at control and does not seed this node's local
1311    /// [`Authority`][ts_tka::Authority]; the lock is reflected locally through the verified
1312    /// netmap-driven sync (every applied AUM passes
1313    /// [`VerifiedAumChain::verify`][ts_tka::VerifiedAumChain::verify]). Verify-and-log posture is
1314    /// unchanged.
1315    ///
1316    /// # Errors
1317    /// [`ts_control::TkaSyncError::Unsupported`] if control has no TKA endpoint or requires re-signing
1318    /// other nodes; [`ts_control::TkaSyncError::NetworkError`] on a transient failure; a coarse
1319    /// `Internal` for a malformed genesis or other RPC failure (incl. control rejecting the init,
1320    /// e.g. a lock already exists).
1321    pub async fn tka_init(
1322        &self,
1323        disablement_secret: Vec<u8>,
1324    ) -> Result<(), ts_control::TkaSyncError> {
1325        self.runtime.tka_init(disablement_secret).await
1326    }
1327
1328    /// Request an OIDC **ID token** from control for this node, scoped to `audience` (workload-
1329    /// identity federation, like `tailscale`'s `id-token` LocalAPI).
1330    ///
1331    /// Returns a signed JWT whose `sub` claim is this node's MagicDNS name and whose `aud` claim is
1332    /// `audience`, suitable for presenting to a third-party relying party (e.g. AWS/GCP
1333    /// workload-identity federation). The node is the token *subject*, not the authenticator — this
1334    /// is token issuance over the Noise transport (`POST /machine/id-token`), not a login path.
1335    /// Requires the control plane to support capability version ≥ 30.
1336    pub async fn fetch_id_token(&self, audience: &str) -> Result<String, ts_control::IdTokenError> {
1337        self.runtime.fetch_id_token(audience.to_string()).await
1338    }
1339
1340    /// Publish a `TXT` DNS record for this node into the tailnet's `ts.net` zone via control's
1341    /// `/machine/set-dns` RPC — the Rust analog of Go `tailscale.com/client/tailscale`'s
1342    /// `LocalClient.SetDNS(ctx, name, value)`.
1343    ///
1344    /// `name` is the full record name (e.g. `_acme-challenge.host.tailnet.ts.net`) and `value` is
1345    /// the record value (e.g. the base64url DNS-01 digest). Like Go's `SetDNS`, this publishes a
1346    /// `TXT` record specifically — its canonical use is satisfying an ACME DNS-01 challenge so a CA
1347    /// can verify control of a `*.ts.net` name. Issuance over the Noise transport (`POST
1348    /// /machine/set-dns`), not a login path.
1349    pub async fn set_dns(&self, name: &str, value: &str) -> Result<(), ts_control::SetDnsError> {
1350        self.runtime
1351            .set_dns(name.to_string(), value.to_string())
1352            .await
1353    }
1354
1355    /// Log this node out of the tailnet — deregister it from the control plane (the equivalent of
1356    /// Go `tsnet`'s `LocalClient.Logout`).
1357    ///
1358    /// Re-`POST`s `/machine/register` with this node's current node key and a past expiry, which the
1359    /// control plane honors by **expiring the node now**: it drops out of every peer's netmap and
1360    /// must re-register (re-authenticate) to rejoin.
1361    ///
1362    /// This is primarily for **non-ephemeral** nodes. An ephemeral node is garbage-collected by
1363    /// control shortly after it disconnects, but a persistent node lingers in the tailnet
1364    /// (visible to peers, counting against the machine limit) for up to ~24h after the process exits
1365    /// unless explicitly logged out. Call this before [`shutdown`](Self::shutdown) to deregister
1366    /// immediately. Calling it on an ephemeral node simply brings the GC forward; it is idempotent,
1367    /// so logging out an already-gone node is not an error.
1368    ///
1369    /// This is a **control-plane state change only**: it does not tear down the local datapath (do
1370    /// that via [`shutdown`](Self::shutdown)), and it does not delete or rotate the on-disk node key
1371    /// — re-registering with the same key (a fresh [`Device::new`]) is the re-login path.
1372    pub async fn logout(&self) -> Result<(), ts_control::LogoutError> {
1373        self.runtime.logout().await
1374    }
1375
1376    /// Snapshot this node's client metrics in Prometheus text exposition format.
1377    ///
1378    /// Mirrors Go Tailscale's `clientmetric` registry: process-global counters/gauges incremented
1379    /// on the datapath hot loops (e.g. `magicsock_send_udp`, `magicsock_recv_data_bytes_udp`),
1380    /// rendered as `# TYPE <name> <kind>\n<name> <value>\n` per metric, sorted by name. (Go `tsnet`
1381    /// exposes no metrics method of its own, so this is the fork's clean public surface.) The
1382    /// registry is process-global, so the output covers every `Device` in the process.
1383    pub fn metrics(&self) -> String {
1384        ts_metrics::write_prometheus()
1385    }
1386
1387    /// Map a tailnet source `addr` to the node that owns its IP (like `tsnet`'s `WhoIs`).
1388    ///
1389    /// Only the IP of `addr` is used; the port is ignored. Returns `Ok(None)` if no tailnet node
1390    /// owns that address.
1391    pub async fn whois(&self, addr: SocketAddr) -> Result<Option<WhoIs>, Error> {
1392        self.runtime.whois(addr).await.map_err(Into::into)
1393    }
1394
1395    /// Change the selected exit node at runtime, without recreating the [`Device`] — the equivalent
1396    /// of Go `tsnet`'s `LocalClient.EditPrefs(ExitNodeID/ExitNodeIP)`.
1397    ///
1398    /// The peer may be named by stable node ID, tailnet IP, or MagicDNS name via
1399    /// [`ExitNodeSelector`] (a bare IP or name parses with `selector.parse()`); this is the same
1400    /// selector type as [`Config::exit_node`](crate::Config::exit_node), so the construction-time
1401    /// and runtime paths are identical. Passing `None` clears the exit node — internet-bound traffic
1402    /// is then dropped (fail-closed) unless this node egresses directly.
1403    ///
1404    /// The change is applied immediately: the new selector is re-resolved against the live peer set
1405    /// and the outbound route + inbound source filter are recomputed at once. A selector for a peer
1406    /// not yet in the netmap simply takes effect once that peer appears.
1407    ///
1408    /// Only NEW flows use the changed exit; in-flight connections are not torn down and continue
1409    /// egressing via the previously-selected exit until they close.
1410    pub async fn set_exit_node(&self, exit_node: Option<ExitNodeSelector>) -> Result<(), Error> {
1411        self.runtime
1412            .set_exit_node(exit_node)
1413            .await
1414            .map_err(Into::into)
1415    }
1416
1417    /// The currently-selected exit node, or `None` if none is selected.
1418    pub fn exit_node(&self) -> Option<ExitNodeSelector> {
1419        self.runtime.exit_node()
1420    }
1421
1422    /// Toggle whether this node accepts peer-advertised subnet routes at runtime, without recreating
1423    /// the [`Device`] — the equivalent of Go `tsnet`'s `LocalClient.EditPrefs(RouteAll)` /
1424    /// `tailscale set --accept-routes`.
1425    ///
1426    /// This is a purely **local** preference: unlike [`set_advertise_routes`](Self::set_advertise_routes)
1427    /// it is never reported to control, so it only changes which peer-advertised subnet routes *this*
1428    /// node installs. The change is applied immediately — the outbound route table and the inbound
1429    /// source filter are recomputed together against the live peer set, so turning it on installs (and
1430    /// accepts traffic from) newly-accepted subnets and turning it off removes them from both in
1431    /// lock-step. A peer's own tailnet address is always reachable regardless; the exit-node default
1432    /// route is governed by [`set_exit_node`](Self::set_exit_node), not this flag.
1433    ///
1434    /// Only NEW flows are affected; in-flight connections are not torn down. In TUN transport mode the
1435    /// netstack data path honors the toggle immediately, but the host routing table is not re-steered
1436    /// until the device is rebuilt.
1437    pub async fn set_accept_routes(&self, accept: bool) -> Result<(), Error> {
1438        self.runtime
1439            .set_accept_routes(accept)
1440            .await
1441            .map_err(Into::into)
1442    }
1443
1444    /// Whether this node currently accepts peer-advertised subnet routes (`--accept-routes`).
1445    pub fn accept_routes(&self) -> bool {
1446        self.runtime.accept_routes()
1447    }
1448
1449    /// Toggle whether this node accepts the tailnet's DNS configuration at runtime, without
1450    /// recreating the [`Device`] — the equivalent of Go `tsnet`'s `LocalClient.EditPrefs(CorpDNS)` /
1451    /// `tailscale set --accept-dns`.
1452    ///
1453    /// Like [`set_accept_routes`](Self::set_accept_routes) this is a purely **local** preference,
1454    /// never reported to control. When `false`, the MagicDNS responder ignores the control-pushed DNS
1455    /// configuration and answers every query `REFUSED` (mirroring Go applying an empty `dns.Config`
1456    /// when `CorpDNS` is off), so the node can join the tailnet for connectivity without taking over
1457    /// its DNS. The change is applied immediately to the netstack responder and the peerAPI DoH server
1458    /// that shares its view; flipping it back to `true` restores serving from the still-current config
1459    /// (the config is only gated at the read site, never destroyed), so the OFF→ON restore is
1460    /// automatic.
1461    ///
1462    /// In TUN transport mode the in-datapath responder honors the toggle immediately, but the host
1463    /// resolver/route programming (which points the host at `100.100.100.100`) is applied once at
1464    /// device build and is not re-steered until the device is rebuilt.
1465    pub async fn set_accept_dns(&self, accept: bool) -> Result<(), Error> {
1466        self.runtime
1467            .set_accept_dns(accept)
1468            .await
1469            .map_err(Into::into)
1470    }
1471
1472    /// Whether this node currently accepts the tailnet's DNS configuration (`--accept-dns` / `CorpDNS`).
1473    pub fn accept_dns(&self) -> bool {
1474        self.runtime.accept_dns()
1475    }
1476
1477    /// Change the subnet routes this node advertises at runtime — Go `tailscale set
1478    /// --advertise-routes`. This is the runtime equivalent of
1479    /// [`Config::advertise_routes`](crate::Config::advertise_routes): the node re-advertises the
1480    /// prefixes to control (so it is granted the subnet-router role for them) AND starts forwarding
1481    /// them on the data path, applied together so the two never disagree.
1482    ///
1483    /// `routes` is filtered to the IPv4-only, deduplicated set this fork honors (IPv6 prefixes are
1484    /// dropped under the IPv6-off posture). This sets the explicit subnet prefixes only; it does not
1485    /// affect the exit-node `0.0.0.0/0` advertisement. Only NEW forwarded flows use the changed set;
1486    /// in-flight flows keep their existing routing until they close.
1487    pub async fn set_advertise_routes(&self, routes: Vec<ipnet::IpNet>) -> Result<(), Error> {
1488        self.runtime
1489            .set_advertise_routes(routes)
1490            .await
1491            .map_err(Into::into)
1492    }
1493
1494    /// Advertise (or stop advertising) this node as an **exit node** at runtime — Go `tailscale set
1495    /// --advertise-exit-node`. The runtime equivalent of
1496    /// [`Config::advertise_exit_node`](crate::Config::advertise_exit_node): when `enable` it adds the
1497    /// `0.0.0.0/0` default route to what this node advertises (and forwards), when `false` it removes
1498    /// it.
1499    ///
1500    /// Composes with [`set_advertise_routes`](Device::set_advertise_routes): the explicit subnet
1501    /// routes and the exit-node advertisement are independent — toggling one preserves the other.
1502    /// Advertising an exit node only makes this node *eligible*; control + the peer still decide
1503    /// whether to route through it. Only NEW forwarded flows see the change; in-flight flows keep
1504    /// their routing.
1505    pub async fn set_advertise_exit_node(&self, enable: bool) -> Result<(), Error> {
1506        self.runtime
1507            .set_advertise_exit_node(enable)
1508            .await
1509            .map_err(Into::into)
1510    }
1511
1512    /// Change this node's hostname at runtime — Go `tailscale set --hostname`. Re-reports
1513    /// `Hostinfo.Hostname` to control on the live connection (no rebuild, no reconnect); control
1514    /// reflects the new name in the netmap (it drives the node's MagicDNS name / `tailscale status`
1515    /// display). Hostname is display metadata, so there is no data-path effect. The new value also
1516    /// persists across a later re-registration.
1517    pub async fn set_hostname(&self, hostname: String) -> Result<(), Error> {
1518        self.runtime
1519            .set_hostname(hostname)
1520            .await
1521            .map_err(Into::into)
1522    }
1523
1524    /// Re-bind the underlay UDP socket after a **network/link change** — Wi-Fi switch, sleep/wake,
1525    /// or any event that invalidates the device's local address/NAT mapping. This is the Rust
1526    /// analog of Go magicsock's `Conn.Rebind()`.
1527    ///
1528    /// The embedder owns deciding *when* to call this (it watches the OS for link changes — there is
1529    /// no built-in network monitor); `rebind` is the engine half that does the socket work:
1530    /// - Re-binds the underlay UDP socket, preferring the same local port (so the advertised
1531    ///   endpoint stays stable) and falling back to an ephemeral port. The IPv4-only-by-default
1532    ///   invariant is preserved.
1533    /// - Invalidates the now-stale local mapping: learned reflexive (STUN) addresses and every
1534    ///   peer's *confirmed* direct path are cleared, while candidate endpoints are kept — so peers
1535    ///   are re-probed over the new socket and **relay over DERP (never a direct host dial) until a
1536    ///   path re-confirms**. Endpoint discovery re-runs on its normal cadence.
1537    /// - Leaves peers, control, the netmap, disco keys, and DERP connections untouched; existing
1538    ///   WireGuard sessions survive (they ride whatever underlay carries them).
1539    ///
1540    /// A no-op if the underlay socket failed to bind at startup (the device is DERP-only). Existing
1541    /// connectivity is preserved on a re-bind error (the old socket is kept; the error is returned).
1542    pub async fn rebind(&self) -> Result<(), Error> {
1543        self.runtime.rebind().await.map_err(Into::into)
1544    }
1545
1546    /// Force an immediate STUN re-probe / endpoint re-derivation **without** rebinding the underlay
1547    /// socket — the Rust analog of Go magicsock's `Conn.ReSTUN("debug")` (what `tailscale debug
1548    /// restun` triggers).
1549    ///
1550    /// Unlike [`rebind`](Self::rebind), this does **not** swap the socket or disturb any learned
1551    /// path: it keeps the existing UDP socket and its NAT mapping and only re-runs the STUN sweep
1552    /// now (re-learning this node's reflexive/public address) instead of waiting out the periodic
1553    /// (~23s, jittered) prober. Use it when this node's public endpoint may have changed (e.g. a NAT
1554    /// rebinding) but the socket itself is still fine — it is strictly lighter than a rebind.
1555    ///
1556    /// Peers, control, the netmap, disco keys, and DERP are untouched, and there is **no control
1557    /// round-trip**. A no-op if the underlay socket failed to bind at startup (the device is
1558    /// DERP-only) or while no peer is configured (matching the periodic prober's gate).
1559    pub async fn re_stun(&self) -> Result<(), Error> {
1560        self.runtime.re_stun().await.map_err(Into::into)
1561    }
1562
1563    /// The stable id of the exit node traffic is **currently** egressing through, or `None` if none
1564    /// is engaged (the equivalent of Go `tsnet`'s `Status.ExitNodeStatus.ID`).
1565    ///
1566    /// This differs from [`exit_node`](Self::exit_node), which returns the *configured* selector:
1567    /// the active exit node is the route updater's resolved, fail-closed answer. It is `None` when
1568    /// no exit node is configured, the configured selector matches no current peer, or the matched
1569    /// peer no longer advertises a default route (egress is then dropped, fail-closed). Match the id
1570    /// against [`Status::peers`](crate::Status::peers) (via [`status`](Self::status)) for details.
1571    pub fn active_exit_node(&self) -> Option<ts_control::StableNodeId> {
1572        self.runtime.active_exit_node()
1573    }
1574
1575    /// Watch for netmap changes: the returned receiver's value is the current set of peer
1576    /// [`StatusNode`]s and updates on every netmap change. This is the narrow peer-only view; for
1577    /// the unified Go-`WatchIPNBus` feed (peers + device-state + login URL in one stream) use
1578    /// [`watch_ipn_bus`](Self::watch_ipn_bus).
1579    pub async fn watch_netmap(
1580        &self,
1581    ) -> Result<tokio::sync::watch::Receiver<Vec<StatusNode>>, Error> {
1582        self.runtime.watch_netmap().await.map_err(Into::into)
1583    }
1584
1585    /// The current device connection-[`DeviceState`] (`Connecting` / `Running` / `NeedsLogin` /
1586    /// `Expired` / `Failed`).
1587    pub fn device_state(&self) -> DeviceState {
1588        self.runtime.device_state()
1589    }
1590
1591    /// Watch the device connection-[`DeviceState`], reacting push-style to control connection
1592    /// transitions instead of polling [`status`](Self::status).
1593    ///
1594    /// Returns a [`tokio::sync::watch::Receiver`]; await its
1595    /// [`changed`](tokio::sync::watch::Receiver::changed) to be woken on each transition. The
1596    /// initial value is the current state.
1597    pub fn watch_state(&self) -> tokio::sync::watch::Receiver<DeviceState> {
1598        self.runtime.watch_state()
1599    }
1600
1601    /// Subscribe to the unified IPN notification bus (Go `ipn`'s `WatchIPNBus`).
1602    ///
1603    /// Returns an [`IpnBusWatcher`]; await [`next`](IpnBusWatcher::next) to receive [`Notify`]
1604    /// events that merge device-[`DeviceState`] transitions (with the interactive-login URL surfaced
1605    /// as [`Notify::browse_to_url`]) and netmap peer-set changes into one feed — the single stream a
1606    /// consumer porting from Go's `WatchNotifications` expects, instead of composing
1607    /// [`watch_state`](Self::watch_state) and [`watch_netmap`](Self::watch_netmap) by hand. `mask`
1608    /// ([`NotifyWatchOpt`]) front-loads the current state as an initial snapshot on subscribe
1609    /// (`INITIAL_STATE` / `INITIAL_NETMAP`), mirroring Go's `NotifyInitialState` /
1610    /// `NotifyInitialNetMap`. Delivery is best-effort (a slow consumer drops notifications rather
1611    /// than stalling the runtime); the stream ends when the device shuts down.
1612    pub async fn watch_ipn_bus(&self, mask: NotifyWatchOpt) -> Result<IpnBusWatcher, Error> {
1613        self.runtime.watch_ipn_bus(mask).await.map_err(Into::into)
1614    }
1615
1616    /// Wait until the device finishes registering, returning a typed outcome — the clean
1617    /// replacement for polling [`ipv4_addr`](Self::ipv4_addr) in a loop.
1618    ///
1619    /// Resolves `Ok(())` once the device is [`DeviceState::Running`]. On a non-running outcome it
1620    /// returns a typed [`RegistrationError`]:
1621    /// - [`AuthRejected`](RegistrationError::AuthRejected) — bad/expired/unknown auth key;
1622    ///   **permanent** (re-pair).
1623    /// - [`NeedsLogin`](RegistrationError::NeedsLogin) — interactive authorization required;
1624    ///   **not permanent** (the runtime keeps retrying and reaches `Running` once the user
1625    ///   authorizes). Auth-key callers treat this as failure; interactive callers should ignore it
1626    ///   and drive the flow via [`watch_state`](Self::watch_state).
1627    /// - [`NetworkUnreachable`](RegistrationError::NetworkUnreachable) — **transient** (retry).
1628    /// - [`Timeout`](RegistrationError::Timeout) — no settled state within `timeout` (`None` waits
1629    ///   indefinitely).
1630    ///
1631    /// [`KeyExpired`](RegistrationError::KeyExpired) is not produced here (a key expires only after
1632    /// the node is up); observe it via [`watch_state`](Self::watch_state). Use
1633    /// [`RegistrationError::is_permanent`] to branch "re-pair" vs. "retry / drive login".
1634    pub async fn wait_until_running(
1635        &self,
1636        timeout: Option<Duration>,
1637    ) -> Result<(), RegistrationError> {
1638        self.runtime.wait_until_running(timeout).await
1639    }
1640
1641    /// Ping a tailnet peer over the overlay with an ICMPv4 echo, returning the round-trip time
1642    /// (like `tailscale ping`).
1643    ///
1644    /// The echo is sent from this device's own tailnet IPv4 over the overlay netstack — never a
1645    /// host socket. IPv6 destinations return [`PingError::Ipv6Unsupported`] (this fork is
1646    /// IPv4-only on the tailnet). A peer answers from its own OS stack; this netstack does not
1647    /// auto-reply to echo requests.
1648    ///
1649    /// In TUN transport mode there is no application netstack to ping from; this surfaces as
1650    /// [`PingError::Timeout`] (the same error this method already uses for an unavailable source
1651    /// address — `PingError` carries no dedicated "unsupported" variant).
1652    pub async fn ping(&self, dst: IpAddr, timeout: Duration) -> Result<Duration, PingError> {
1653        let channel = self.channel().map_err(|_| PingError::Timeout)?;
1654        let src = self.ipv4_addr().await.map_err(|_| PingError::Timeout)?;
1655        ts_netstack_smoltcp::ping(channel, src, dst, timeout).await
1656    }
1657
1658    /// The current **direct path** to the peer at tailnet IP `dst`: its confirmed direct UDP
1659    /// endpoint and that path's last-measured round-trip latency, or `None` when traffic to the peer
1660    /// is **relayed via DERP** (no trusted direct path right now), the peer is unknown, or it has no
1661    /// disco key.
1662    ///
1663    /// This is the direct-path analog of Go's `tailscale ping`/`PeerStatus` connectivity: a present
1664    /// result means packets reach the peer directly at the returned address, with roughly the
1665    /// returned RTT. The latency is a live snapshot taken from the most recent disco ping/pong that
1666    /// confirmed the path (up to one probe interval stale) — not a fresh on-demand round-trip. Unlike
1667    /// [`ping`](Device::ping) (an ICMP echo over the netstack), this reports the *underlay* path the
1668    /// data plane actually uses, distinguishing a direct connection from a DERP-relayed one.
1669    pub async fn direct_path(&self, dst: IpAddr) -> Result<Option<(SocketAddr, Duration)>, Error> {
1670        self.runtime.direct_path(dst).await.map_err(Into::into)
1671    }
1672
1673    /// Send a disco ping to the peer at tailnet IP `dst` **now** and await the pong — a fresh,
1674    /// on-demand round-trip measurement (Go's `tailscale ping`, `PingType::Disco`). Returns the
1675    /// endpoint that answered and the measured RTT, or `None` if no pong arrives within `timeout`
1676    /// (or the peer is unknown / has no candidate direct path).
1677    ///
1678    /// Unlike [`direct_path`](Device::direct_path) — which reports the *last periodic probe's* RTT
1679    /// from cache — this actively sends a ping and waits for the reply, so the latency is current. A
1680    /// `None` here means "no direct path confirmed within the timeout" (the peer may still be
1681    /// reachable via DERP). Unlike [`ping`](Device::ping) (an ICMP echo over the netstack), this
1682    /// measures the disco/underlay path the data plane uses for direct connections.
1683    pub async fn ping_disco(
1684        &self,
1685        dst: IpAddr,
1686        timeout: Duration,
1687    ) -> Result<Option<(SocketAddr, Duration)>, Error> {
1688        self.runtime
1689            .ping_disco(dst, timeout)
1690            .await
1691            .map_err(Into::into)
1692    }
1693
1694    /// Obtain a TLS certificate for a node's MagicDNS `name` (like `tsnet`'s `GetCertificate`).
1695    ///
1696    /// **Fail-closed without the `acme` feature.** By default this fork has no client-side ACME
1697    /// engine wired in, so this returns [`ts_control::CertError::Unimplemented`] (after a
1698    /// tailnet-name check) — it NEVER self-signs and NEVER returns a placeholder certificate
1699    /// ([`ts_control::MISSING_CERT_RPC`] names what is missing).
1700    ///
1701    /// **With the `acme` feature** this instead drives the client-side ACME DNS-01 engine to issue a
1702    /// real Let's Encrypt certificate for `name`, publishing the challenge TXT via the node's
1703    /// `POST /machine/set-dns` RPC (routed through the control runner). SaaS-only: a self-hosted
1704    /// control plane may 501 on set-dns, surfaced as [`ts_control::CertError::Acme`].
1705    #[cfg(not(feature = "acme"))]
1706    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
1707        ts_control::get_certificate(name).await
1708    }
1709
1710    /// See the no-`acme` variant for the contract; with `acme` this issues a real cert via the
1711    /// runtime's ACME engine (`Device → Runtime → ControlRunner → issue_certificate_via_setdns`).
1712    #[cfg(feature = "acme")]
1713    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
1714        self.runtime.get_certificate(name.to_string()).await
1715    }
1716
1717    /// Issue a real Let's Encrypt certificate for a node's MagicDNS `name` and return the **PEM
1718    /// pair** `(cert_chain_pem, key_pem)` — the analog of Go's `LocalClient.CertPairWithValidity`,
1719    /// for writing the daemon's on-disk `.crt` + `.key` (`tnet cert`). **`acme` feature only.**
1720    ///
1721    /// This drives the same client-side ACME DNS-01 issuance as [`Device::get_certificate`] (one
1722    /// order, the challenge TXT published via the node's `POST /machine/set-dns` RPC, routed through
1723    /// the runtime → control runner); it differs only in returning the raw leaf+chain PEM and the
1724    /// leaf private-key PEM instead of the opaque [`CertifiedKey`]. The second tuple element is
1725    /// **secret key material**: it is never logged anywhere on this path — persist it to a `0600`
1726    /// file and never trace it.
1727    ///
1728    /// **`min_validity` (honest "always fresh").** Go's `CertPairWithValidity` reuses a cached cert
1729    /// when it has at least `min_validity` of its lifetime remaining, re-issuing otherwise. This
1730    /// fork keeps **no cert cache** — every call issues fresh — so `min_validity` is accepted for
1731    /// signature compatibility but does not alter behavior: a freshly issued (full-lifetime) cert
1732    /// satisfies any `min_validity`. A reuse cache is separate future work; this does NOT fake one.
1733    ///
1734    /// Fail-closed: returns a [`ts_control::CertError`] (never a self-signed or partial pair) on any
1735    /// ACME/HTTP failure. SaaS-only: a self-hosted control plane may 501 on set-dns, surfaced as
1736    /// [`ts_control::CertError::Acme`].
1737    #[cfg(feature = "acme")]
1738    pub async fn cert_pair(
1739        &self,
1740        name: &str,
1741        min_validity: Option<Duration>,
1742    ) -> Result<(String, String), ts_control::CertError> {
1743        self.runtime.cert_pair(name.to_string(), min_validity).await
1744    }
1745
1746    /// Build a [`TlsAcceptor`] terminating TLS for `cfg.name` on the overlay (like `tsnet`'s
1747    /// `ListenTLS`).
1748    ///
1749    /// Obtains the certificate via [`Device::get_certificate`] — so with the `acme` feature this
1750    /// issues a real Let's Encrypt cert (when the control plane answers `set-dns`), and without it
1751    /// (or when issuance is unavailable) it surfaces the same fail-closed
1752    /// [`ts_control::CertError`] rather than ever serving a self-signed cert or downgrading to
1753    /// plaintext. Terminate accepted overlay streams with [`ts_control::accept_tls`].
1754    pub async fn listen_tls(
1755        &self,
1756        cfg: &ts_control::ServeConfig,
1757    ) -> Result<TlsAcceptor, ts_control::CertError> {
1758        // Route through Device::get_certificate (the acme-aware issuance path) rather than
1759        // ts_control::listen_tls, which only knows the non-acme stub. Validate the serve config
1760        // first (same fail-closed checks ts_control::listen_tls applies), then assemble the acceptor.
1761        cfg.validate()?;
1762        let cert = self.get_certificate(&cfg.name).await?;
1763        ts_control::tls_acceptor(cert)
1764    }
1765
1766    /// The currently-stored Serve config (like `tsnet`'s `GetServeConfig`).
1767    ///
1768    /// Returns the config last passed to [`Device::set_serve_config`], or an empty
1769    /// [`ts_control::ServeState`] (no ports) if none was ever set. Pure read — does not touch the
1770    /// network.
1771    pub fn get_serve_config(&self) -> ts_control::ServeState {
1772        match &*self.serve.lock().unwrap_or_else(|e| e.into_inner()) {
1773            Some(mgr) => mgr.get(),
1774            None => ts_control::ServeState::default(),
1775        }
1776    }
1777
1778    /// Replace this node's Serve config and (re)bind its tailnet ports (like `tsnet`'s
1779    /// `SetServeConfig`, REPLACE semantics).
1780    ///
1781    /// `state` becomes the **whole** config (full-replace reconcile: every previously-bound serve
1782    /// port's accept loop is torn down and the new config's ports are bound from scratch). For each
1783    /// configured port the manager binds an overlay listener on this node's tailnet IPv4 and
1784    /// dispatches per [`ts_control::ServeTarget`]:
1785    /// - [`Accept`](ts_control::ServeTarget::Accept) — the TLS-terminated stream is handed back over
1786    ///   the returned [`ServeAcceptedReceiver`](ts_runtime::serve::ServeAcceptedReceiver) (the
1787    ///   in-process stand-in for `ListenTLS`'s `net.Listener`).
1788    /// - [`Proxy`](ts_control::ServeTarget::Proxy) — reverse-proxy the decrypted stream to a local
1789    ///   host backend.
1790    /// - [`Text`](ts_control::ServeTarget::Text) — write a fixed body and close.
1791    /// - [`TcpForward`](ts_control::ServeTarget::TcpForward) — forward the **raw** (non-TLS) stream
1792    ///   to a local host backend.
1793    ///
1794    /// **Fail-closed.** `state.validate()` runs first. Every TLS-terminating port's acceptor is
1795    /// obtained up-front via [`Device::listen_tls`] (the ACME-aware cert path); if any cert cannot be
1796    /// issued the whole call fails with that [`ts_control::CertError`] and **nothing is bound** — a
1797    /// TLS port never downgrades to plaintext.
1798    ///
1799    /// **Anti-leak.** Listeners bind the overlay netstack only (never a host socket). The
1800    /// `Proxy`/`TcpForward` backend dial is a local host socket to the embedder's own backend (like
1801    /// Go's reverse-proxy to `127.0.0.1`), intentionally NOT routed through the exit-egress
1802    /// forwarder. A backend dial failure drops that connection; it never falls back.
1803    ///
1804    /// Returns an error in TUN transport mode (there is no application netstack to bind on). The
1805    /// previous config's accept loops (and any earlier `ServeAcceptedReceiver`) stop when this
1806    /// returns; the new receiver delivers every `Accept`-port connection.
1807    pub async fn set_serve_config(
1808        &self,
1809        state: ts_control::ServeState,
1810    ) -> Result<ts_runtime::serve::ServeAcceptedReceiver, Error> {
1811        state
1812            .validate()
1813            .map_err(|_| Error::Internal(InternalErrorKind::BadRequest))?;
1814
1815        // Fail-closed: build every TLS-terminating port's acceptor up-front via the ACME-aware cert
1816        // path. If any cert can't be issued, return before binding anything (no plaintext downgrade).
1817        let mut resolved = std::collections::BTreeMap::new();
1818        for (port, target) in &state.ports {
1819            let acceptor = if target.terminates_tls() {
1820                let cfg = ts_control::ServeConfig {
1821                    name: state.name.clone(),
1822                    port: *port,
1823                    target: target.clone(),
1824                };
1825                Some(self.listen_tls(&cfg).await.map_err(|_| {
1826                    // Cert issuance is fail-closed in this fork; surface as a request error rather
1827                    // than ever binding a plaintext TLS port.
1828                    Error::Internal(InternalErrorKind::BadRequest)
1829                })?)
1830            } else {
1831                None
1832            };
1833            resolved.insert(
1834                *port,
1835                ts_runtime::serve::ResolvedPort {
1836                    target: target.clone(),
1837                    acceptor,
1838                },
1839            );
1840        }
1841
1842        // The manager binds the OVERLAY netstack on this node's own tailnet IPv4.
1843        let self_ipv4 = self.ipv4_addr().await?;
1844        let channel = self.channel()?.clone();
1845
1846        let mut slot = self.serve.lock().unwrap_or_else(|e| e.into_inner());
1847        let mgr =
1848            slot.get_or_insert_with(|| ts_runtime::serve::ServeManager::new(channel, self_ipv4));
1849        Ok(mgr.set(state, resolved))
1850    }
1851
1852    /// Expose a tailnet TLS service to the public internet via Tailscale Funnel (like `tsnet`'s
1853    /// `ListenFunnel`), returning a [`FunnelAcceptedReceiver`](ts_runtime::funnel::FunnelAcceptedReceiver)
1854    /// that delivers each TLS-terminated public connection.
1855    ///
1856    /// **Two fail-closed gates, then the live ingress listener.** First the node-attribute gate is
1857    /// fully enforced from this node's own capability map (mirroring Go `ipn.NodeCanFunnel` +
1858    /// `ipn.CheckFunnelPort`): the tailnet admin must have enabled HTTPS and granted the `funnel`
1859    /// node attribute, and `cfg.port` must be in the set the `funnel-ports` capability allows —
1860    /// otherwise this returns [`ts_control::FunnelError::NotAllowed`] /
1861    /// [`ts_control::FunnelError::PortNotAllowed`] before touching any cert or network. Then the
1862    /// node's `*.ts.net` certificate is obtained via the ACME-aware [`Device::get_certificate`] (the
1863    /// Funnel hostname *is* the node's MagicDNS name, so its DNS-01 cert matches); fail-closed on
1864    /// [`ts_control::FunnelError::Cert`] — no self-signed or plaintext fallback.
1865    ///
1866    /// On success a [`FunnelManager`](ts_runtime::funnel::FunnelManager) is registered: its ingress
1867    /// sink is installed into the runtime's peerAPI `/v0/ingress` slot (making that route live without
1868    /// restarting the peerAPI server), and the `HostInfo.IngressEnabled` map-request signal is set so
1869    /// control routes Funnel traffic to this node. Public Funnel bytes arrive as a relay POST to
1870    /// `/v0/ingress`, are membership-gated + `101`-hijacked into a raw stream, TLS-terminated by the
1871    /// manager, and delivered over the returned receiver.
1872    ///
1873    /// **Where the relay comes from.** The public ingress **relay + DNS mapping** that feed
1874    /// `/v0/ingress` are Tailscale infrastructure ([`ts_control::MISSING_FUNNEL_RELAY`]), provisioned
1875    /// automatically against real Tailscale SaaS with a Funnel-enabled ACL; against a self-hosted
1876    /// control plane no relay exists, so the listener is correct but never fed.
1877    ///
1878    /// Anti-leak: Funnel TLS terminates only on the overlay netstack (the hijacked ingress stream
1879    /// arrives on the overlay peerAPI listener), never a host socket; there is no self-signed or
1880    /// plaintext fallback. A new `listen_funnel` replaces the previous manager (its pump + sink tear
1881    /// down); dropping the `Device` tears it down too.
1882    pub async fn listen_funnel(
1883        &self,
1884        cfg: &ts_control::ServeConfig,
1885        opts: ts_control::FunnelOptions,
1886    ) -> Result<ts_runtime::funnel::FunnelAcceptedReceiver, ts_control::FunnelError> {
1887        // Gate 1 (fail-closed, no network): node-attribute + funnel-port access from our cap map.
1888        let me = self
1889            .self_node()
1890            .await
1891            .map_err(|_| ts_control::FunnelError::NotAllowed)?;
1892        cfg.validate()?;
1893        ts_control::funnel_access(&me, cfg.port)?;
1894
1895        // Gate 2 (fail-closed): obtain the node's `*.ts.net` cert via the ACME-aware path and build
1896        // the TLS acceptor. A cert failure surfaces as FunnelError::Cert — never a plaintext listener.
1897        let cert = self
1898            .get_certificate(&cfg.name)
1899            .await
1900            .map_err(ts_control::FunnelError::Cert)?;
1901        let acceptor = ts_control::tls_acceptor(cert).map_err(ts_control::FunnelError::Cert)?;
1902
1903        // `opts.funnel_only` (reject tailnet-internal connections) is accepted for surface stability;
1904        // the ingress data path only ever carries relay-delivered public traffic, so there is no
1905        // tailnet-internal leg on this listener to reject. Documented as a no-op here for now.
1906        let _ = opts;
1907
1908        // Build the funnel manager + its ingress sink + the hand-back receiver, install the sink into
1909        // the runtime's shared peerAPI `/v0/ingress` slot (making the route live), and flip the
1910        // IngressEnabled map signal. Hold the manager on the device so its pump/sink live as long as
1911        // the listener; replacing a prior manager tears the old one down on drop at end of scope.
1912        let (manager, sink, receiver) = ts_runtime::funnel::FunnelManager::new(acceptor);
1913        {
1914            let slot = self.runtime.funnel_ingress_slot();
1915            *slot.lock().unwrap_or_else(|e| e.into_inner()) = Some(sink);
1916        }
1917        self.runtime
1918            .ingress_active_flag()
1919            .store(true, std::sync::atomic::Ordering::Relaxed);
1920
1921        let old = {
1922            let mut held = self.funnel.lock().unwrap_or_else(|e| e.into_inner());
1923            held.replace(manager)
1924        };
1925        drop(old);
1926
1927        Ok(receiver)
1928    }
1929
1930    /// Host a Tailscale **VIP service** (`svc:<label>`) by binding an overlay listener on the
1931    /// service's control-assigned virtual IP (like `tsnet`'s `ListenService`).
1932    ///
1933    /// **Fail-closed.** Mirrors Go `tsnet.Server.ListenService`'s preconditions, enforced from this
1934    /// node's own netmap state ([`ts_control::resolve_service_listen`]): the `name` must be a valid
1935    /// `svc:<dns-label>`, this node must be **tagged** (Go `ErrUntaggedServiceHost`), and control
1936    /// must have assigned the service a VIP address on this node (delivered via the `service-host`
1937    /// node-capability — see [`ts_control::Node::service_addresses`]). Any unmet precondition
1938    /// returns a typed [`ts_control::ServiceError`] before binding anything.
1939    ///
1940    /// When all hold, this binds a [`tcp_listen`][Device::tcp_listen] on the service VIP and the
1941    /// configured `mode` port over the **overlay netstack** (never a host socket) and returns the
1942    /// listener. The netstack already accepts packets for control-assigned VIPs (they are injected
1943    /// alongside the node's own tailnet address), so the listener is reachable by tailnet peers.
1944    ///
1945    /// The `Tun`/L3 service mode is unsupported (a TODO in upstream tsnet); only TCP/HTTP modes
1946    /// (which bind the same VIP:port at the listen layer) are offered. Returns an error in TUN
1947    /// transport mode (there is no application netstack to bind on).
1948    pub async fn listen_service(
1949        &self,
1950        name: &str,
1951        mode: ts_control::ServiceMode,
1952    ) -> Result<netstack::TcpListener, ts_control::ServiceError> {
1953        let me = self
1954            .self_node()
1955            .await
1956            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))?;
1957        let listen_addr = ts_control::resolve_service_listen(&me, name, mode, self.enable_ipv6)?;
1958        self.tcp_listen(listen_addr)
1959            .await
1960            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))
1961    }
1962
1963    /// Attempt to gracefully shut down this device's runtime.
1964    ///
1965    /// Reports whether the device was fully shut down before the timeout. It is still shut
1966    /// down if it timed out, just more violently and with potential resource leaks.
1967    ///
1968    /// If `timeout` is `None`, then shutdown will never time-out.
1969    pub async fn shutdown(self, timeout: Option<Duration>) -> bool {
1970        self.runtime.graceful_shutdown(timeout).await
1971    }
1972}
1973
1974/// Command-channel-driven userspace network stack.
1975///
1976/// This is an opinionated wrapper around [smoltcp](https://docs.rs/smoltcp) that provides an
1977/// easier-to-integrate, more-portable API.
1978pub mod netstack {
1979    #[doc(inline)]
1980    pub use ts_netstack_smoltcp::netcore::Error;
1981    #[doc(inline)]
1982    pub use ts_netstack_smoltcp::netcore::InternalErrorKind;
1983    #[doc(inline)]
1984    pub use ts_netstack_smoltcp::netsock::{TcpListener, TcpStream, UdpSocket};
1985}
1986
1987/// Geneve (RFC 8926) framing for Tailscale **peer-relay** traffic. A peer that advertises
1988/// [`NodeInfo::is_peer_relay`] runs a UDP relay server; relayed disco + WireGuard frames are
1989/// Geneve-encapsulated with a VNI. This module exposes the header codec so the framing is
1990/// recognizable. NOTE: the active relay *data path* (the relay-allocation handshake +
1991/// magicsock integration) is **not yet implemented** in this fork — this is the wire-aware slice.
1992pub mod geneve {
1993    #[doc(inline)]
1994    pub use ts_packet::geneve::{
1995        GENEVE_FIXED_HEADER_LEN, GENEVE_PROTOCOL_DISCO, GENEVE_PROTOCOL_WIREGUARD, GeneveError,
1996        GeneveHeader,
1997    };
1998}
1999
2000/// Tailnet Lock (TKA) verification: the [`tka::Authority`] checks a peer's node-key signature
2001/// against the trusted-key state, mirroring Go's `tka` package. Pair with [`Device::tka_status`]
2002/// (the control-pushed head/disablement signal).
2003pub mod tka {
2004    #[doc(inline)]
2005    pub use ts_tka::{
2006        AumHash, AumKind, Authority, Key, KeyKind, NodeKeySignature, SigKind, State, TkaError,
2007        aum_hash,
2008    };
2009}
2010
2011/// Tailscale cryptographic key types.
2012pub mod keys {
2013    #[doc(inline)]
2014    pub use ts_keys::{
2015        DiscoKeyPair, DiscoPrivateKey, DiscoPublicKey, MachineKeyPair, MachinePrivateKey,
2016        MachinePublicKey, NetworkLockKeyPair, NetworkLockPrivateKey, NetworkLockPublicKey,
2017        NodeKeyPair, NodePrivateKey, NodePublicKey, NodeState, PersistState,
2018    };
2019}
2020
2021const ENV_MAGIC_VAR: &str = "TS_RS_EXPERIMENT";
2022const ENV_MAGIC_VALUE: &str = "this_is_unstable_software";
2023
2024fn check_magic_env() -> Result<(), Error> {
2025    if std::env::var(ENV_MAGIC_VAR).as_deref() != Ok(ENV_MAGIC_VALUE) {
2026        let warning = format!(
2027            "
2028check failed: set {ENV_MAGIC_VAR}={ENV_MAGIC_VALUE} to acknowledge that tailscale-rs is early-days
2029experimental software containing bugs, unvalidated cryptography, and no stability or compatibility
2030guarantees.
2031            "
2032        );
2033
2034        eprintln!("{}", warning.trim());
2035
2036        return Err(Error::UnstableEnvVar);
2037    };
2038
2039    Ok(())
2040}
2041
2042#[cfg(test)]
2043mod tests {
2044    use secrecy::ExposeSecret as _;
2045
2046    use super::*;
2047
2048    // `Device::new`/`new_with_secret` cannot be unit-tested end-to-end without a live control
2049    // server (registration). The only behavioral difference `new_with_secret` introduces over `new`
2050    // is exposing the `SecretString` to a plain `String` on the last inch; everything after is the
2051    // shared `new` path. So we assert that equivalence at the auth-key-resolution level: the secret
2052    // path must resolve to the exact same key the plain path feeds into `resolve_auth_key`.
2053    const SAMPLE_KEY: &str = "tskey-auth-koCgSLP5R811CNTRL-EXAMPLEEXAMPLEEXAMPLEEXAMPLE";
2054
2055    // The mapping `new_with_secret` applies (`Option<SecretString>` -> `Option<String>`) must be a
2056    // byte-for-byte round-trip, so the spawn arg is identical to a direct `new(config, Some(..))`.
2057    #[test]
2058    fn secret_exposes_to_identical_string() {
2059        let plain = Some(SAMPLE_KEY.to_string());
2060        let from_secret =
2061            Some(SecretString::from(SAMPLE_KEY)).map(|s| s.expose_secret().to_string());
2062        assert_eq!(from_secret, plain);
2063
2064        // `None` must pass through unchanged (so it falls back to `config.auth_key` exactly as `new`).
2065        let none_secret: Option<SecretString> = None;
2066        assert_eq!(
2067            none_secret.map(|s| s.expose_secret().to_string()),
2068            None::<String>
2069        );
2070    }
2071
2072    // End-to-end equivalence at the resolve layer: feeding the exposed secret through
2073    // `resolve_auth_key` yields the same `Option<String>` as feeding the plain string — i.e. both
2074    // constructors reach the same spawn argument, without registering against a control server.
2075    #[tokio::test]
2076    async fn new_with_secret_resolves_same_as_new() {
2077        let config = Config::default();
2078
2079        let via_plain = resolve_auth_key(&config, Some(SAMPLE_KEY.to_string()))
2080            .await
2081            .expect("plain auth key resolves");
2082
2083        let exposed = Some(SecretString::from(SAMPLE_KEY)).map(|s| s.expose_secret().to_string());
2084        let via_secret = resolve_auth_key(&config, exposed)
2085            .await
2086            .expect("secret-derived auth key resolves");
2087
2088        assert_eq!(via_plain, via_secret);
2089        // Without the `identity-federation` feature `resolve_auth_key` is a pass-through, so the
2090        // resolved key is the input verbatim; assert that too to pin the default-build behavior.
2091        #[cfg(not(feature = "identity-federation"))]
2092        assert_eq!(via_secret, Some(SAMPLE_KEY.to_string()));
2093    }
2094}