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//!
94//! ## Platform Support
95//!
96//! `tailscale` currently supports the following platforms:
97//!
98//! - Linux (x86_64 and ARM64)
99//! - macOS (ARM64)
100//!
101//! ## Component crates
102//!
103//! The following crates are part of the tailscale-rs project and are dependencies of this one. For
104//! many tasks, just this crate should be sufficient and these other crates are an implementation detail.
105//! There are other crates too, see [ARCHITECTURE.md]
106//! or the [GitHub repo](https://github.com/tailscale/tailscale-rs).
107//!
108//! - [ts_runtime](https://docs.rs/ts_runtime): for each API-level `Device`, the runtime uses an actor
109//!   architecture to manage the lifecycle of the control client, data plane components, netstack, etc.
110//!   A message bus passes updates and communications between these top-level actors.
111//! - [ts_netcheck](https://docs.rs/ts_netcheck): checks network availability and reports latency to
112//!   DERP servers in different regions.
113//! - [ts_netstack_smoltcp](https://docs.rs/ts_netstack_smoltcp): a [smoltcp](https://docs.rs/smoltcp)-based
114//!   network stack that processes Layer 3+ packets to/from the overlay network.
115//! - [ts_control](https://docs.rs/ts_control): control plane client that handles registration,
116//!   authorization/authentication, configuration, and streaming updates.
117//! - [ts_dataplane](https://docs.rs/ts_dataplane): wires all the individual data plane functions together,
118//!   flowing inbound and outbound packets through the components in the correct order.
119//! - [ts_tunnel](https://docs.rs/ts_tunnel): a partial implementation of the WireGuard specification
120//!   that protects all data plane traffic, and is interoperable with other WireGuard clients, including Tailscale clients.
121//! - [ts_cli_util](https://docs.rs/ts_cli_util): helpers for writing command line tools and initializing
122//!   logging, used in examples.
123//! - [ts_disco_protocol](https://docs.rs/ts_disco_protocol): incomplete implementation of Tailscale's
124//!   discovery protocol (disco).
125//!
126//! [ARCHITECTURE.md]: https://github.com/tailscale/tailscale-rs/blob/main/ARCHITECTURE.md
127//! [CONTRIBUTING.md]: https://github.com/tailscale/tailscale-rs/blob/main/CONTRIBUTING.md
128//! [`examples/`]: https://github.com/tailscale/tailscale-rs/blob/main/examples/README.md
129//! [open an issue]: https://github.com/tailscale/tailscale-rs/issues
130//! [`axum` HTTP server]: https://docs.rs/axum/latest/axum/
131
132use std::{
133    net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr},
134    time::Duration,
135};
136
137#[doc(inline)]
138pub use config::Config;
139#[doc(inline)]
140pub use error::{Error, InternalErrorKind};
141#[doc(inline)]
142pub use ts_control::ExitNodeSelector;
143#[doc(inline)]
144pub use ts_control::Node as NodeInfo;
145#[doc(inline)]
146pub use ts_control::tls::{CertifiedKey, TlsAcceptor, TlsStream};
147#[doc(inline)]
148pub use ts_control::{CertError, MISSING_CERT_RPC, ServeConfig, ServeState, ServeTarget};
149#[doc(inline)]
150pub use ts_control::{ExitProxyConfig, ExitProxyScheme};
151pub use ts_control::{
152    IdTokenError, LogoutError, ServiceError, ServiceMode, SshAccept, SshAction, SshConnIdentity,
153    SshDecision, SshDenyReason, SshPolicy, SshPrincipal, SshRule, StableNodeId,
154};
155#[doc(inline)]
156pub use ts_netstack_smoltcp::PingError;
157use ts_netstack_smoltcp::{CreateSocket, netcore::Channel};
158#[doc(inline)]
159pub use ts_runtime::fallback_tcp::{
160    FallbackConnFuture, FallbackConnHandler, FallbackDecision, FallbackTcpHandle,
161};
162#[doc(inline)]
163pub use ts_runtime::taildrop::WaitingFile;
164#[doc(inline)]
165pub use ts_runtime::{DeviceState, RegistrationError, Status, StatusNode, WhoIs};
166
167#[cfg(feature = "axum")]
168pub mod axum;
169pub mod config;
170mod error;
171mod loopback;
172#[cfg(feature = "ssh")]
173pub mod ssh;
174
175#[doc(inline)]
176pub use loopback::LoopbackHandle;
177
178/// How a program connects to a tailnet and communicates with peers.
179///
180/// The `Device` connects to the control plane, registers itself with the tailnet, and communicates
181/// with tailnet peers. Its tailnet identity is determined by the key state provided at
182/// construction-time.
183pub struct Device {
184    runtime: ts_runtime::Runtime,
185    /// Command channel to the application netstack. `None` in TUN transport mode, where there is
186    /// no userspace application netstack; the channel-driven socket APIs ([`Device::udp_bind`],
187    /// [`Device::tcp_listen`], [`Device::tcp_connect`], [`Device::ping`]) are unsupported there.
188    channel: Option<Channel>,
189    /// Whether IPv6 is enabled on the tailnet overlay (the `Config::enable_ipv6` gate, default
190    /// `false`). Captured at construction; used by [`Device::listen_service`] to decide whether an
191    /// IPv6 VIP-service address is bindable (the netstack only accepts IPv6 overlay addresses when
192    /// this is set).
193    enable_ipv6: bool,
194    /// The stored Serve config + its live per-port accept loops (`tsnet`'s `Get/SetServeConfig` +
195    /// serving runtime). Built lazily on the first [`Device::set_serve_config`] (it needs this
196    /// node's overlay IPv4, only known after registration). Held here so its accept loops abort when
197    /// the `Device` drops; `None` (empty config) until the first `set`.
198    serve: std::sync::Mutex<Option<ts_runtime::serve::ServeManager>>,
199    /// The live Funnel ingress manager (`tsnet`'s `ListenFunnel` data path), built on
200    /// [`Device::listen_funnel`](crate::Device::listen_funnel). Held here so its TLS-termination pump and the installed peerAPI
201    /// ingress sink stay alive for the device's life (and tear down when a new `listen_funnel`
202    /// replaces it, or the `Device` drops). `None` until the first `listen_funnel`.
203    funnel: std::sync::Mutex<Option<ts_runtime::funnel::FunnelManager>>,
204}
205
206/// Map a [`ts_runtime::taildrop::TaildropError`] to the device-facing [`Error`]. `Error` is a
207/// `Copy` enum with no payload, so the I/O detail string is dropped, but the *kind* is preserved so
208/// a caller can still distinguish the actionable cases: an invalid name →
209/// [`InternalErrorKind::BadRequest`], an in-progress conflict → [`InternalErrorKind::AlreadyExists`],
210/// a missing file → [`InternalErrorKind::NotFound`], and any other filesystem failure →
211/// [`InternalErrorKind::Io`].
212fn taildrop_err(e: ts_runtime::taildrop::TaildropError) -> Error {
213    use ts_runtime::taildrop::TaildropError;
214    match e {
215        TaildropError::InvalidFileName => Error::Internal(InternalErrorKind::BadRequest),
216        TaildropError::FileExists => Error::Internal(InternalErrorKind::AlreadyExists),
217        TaildropError::Io(io) if io.kind() == std::io::ErrorKind::NotFound => {
218            Error::Internal(InternalErrorKind::NotFound)
219        }
220        TaildropError::Io(_) => Error::Internal(InternalErrorKind::Io),
221    }
222}
223
224/// Map a [`ts_runtime::taildrop_send::TaildropSendError`] (the Taildrop *sender*) to the
225/// device-facing [`Error`]. The send-side conflict/forbidden/unexpected-status cases all reduce to
226/// `BadRequest` (the peer refused the transfer for a request-level reason), a dial failure or
227/// timeout to `Timeout`, an invalid name to `BadRequest`, and any stream I/O failure to `Io`.
228fn taildrop_send_err(e: ts_runtime::taildrop_send::TaildropSendError) -> Error {
229    use ts_runtime::taildrop_send::TaildropSendError;
230    match e {
231        TaildropSendError::Connect | TaildropSendError::Timeout => Error::Timeout,
232        TaildropSendError::InvalidName
233        | TaildropSendError::Forbidden
234        | TaildropSendError::Conflict
235        | TaildropSendError::UnexpectedStatus(_) => Error::Internal(InternalErrorKind::BadRequest),
236        TaildropSendError::Io => Error::Internal(InternalErrorKind::Io),
237    }
238}
239
240/// Resolve the effective registration auth key from `auth_key` plus the config's
241/// workload-identity-federation (WIF) / OAuth-client fields.
242///
243/// With the `identity-federation` feature enabled, an OAuth client secret (`tskey-client-…`) or a
244/// `client_id` + (`id_token` | `audience`) is exchanged for a Tailscale auth key against the SaaS
245/// admin API before registration (Go `tsnet.Server`'s `resolveAuthKey`). Without the feature this is
246/// a pure pass-through: `auth_key` is returned unchanged and the WIF config fields are ignored, so
247/// the default build is byte-identical to before.
248#[cfg(feature = "identity-federation")]
249async fn resolve_auth_key(
250    config: &Config,
251    auth_key: Option<String>,
252) -> Result<Option<String>, Error> {
253    let wif = ts_control::WifConfig {
254        auth_key,
255        client_id: config.client_id.clone(),
256        client_secret: config.client_secret.clone(),
257        id_token: config.id_token.clone(),
258        audience: config.audience.clone(),
259        tags: config.requested_tags.clone(),
260    };
261    ts_control::resolve_auth_key(&wif, &config.control_server_url)
262        .await
263        .map_err(|e| {
264            tracing::error!(error = %e, "resolving auth key via workload-identity federation");
265            Error::Internal(InternalErrorKind::BadRequest)
266        })
267}
268
269/// Pass-through when the `identity-federation` feature is disabled: the auth key is used as-is and
270/// the WIF config fields have no effect (matching Go, where the federation path is compiled out
271/// unless its optional feature is linked).
272#[cfg(not(feature = "identity-federation"))]
273async fn resolve_auth_key(
274    _config: &Config,
275    auth_key: Option<String>,
276) -> Result<Option<String>, Error> {
277    Ok(auth_key)
278}
279
280impl Device {
281    /// Create a device from the given [`Config`] and auth key.
282    ///
283    /// Internally, this will spawn multiple asynchronous actors onto a Tokio runtime.
284    ///
285    /// # Example
286    ///
287    /// ```rust,no_run
288    /// # #[tokio::main]
289    /// # async fn main() -> Result<(), Box<dyn std::error::Error>> {
290    /// # use tailscale::*;
291    /// let dev = Device::new(
292    ///     &Config::default_with_key_file("tsrs_keys.json").await?,
293    ///     Some("MY_AUTH_KEY".to_string()),
294    /// ).await?;
295    /// # Ok(()) }
296    /// ```
297    pub async fn new(config: &Config, auth_key: Option<String>) -> Result<Self, Error> {
298        check_magic_env()?;
299
300        // Resolve the effective registration auth key. The explicit `auth_key` argument wins; if it
301        // is `None`, fall back to `config.auth_key` (Go `tsnet.Server.AuthKey`). When the
302        // `identity-federation` feature is enabled, the resolved key is further passed through the
303        // WIF / OAuth-client bootstrap, which exchanges an OAuth client secret (`tskey-client-…`) or
304        // an IdP-issued OIDC token for a Tailscale auth key before registration (SaaS-only).
305        let auth_key = auth_key.or_else(|| config.auth_key.clone());
306        let auth_key = resolve_auth_key(config, auth_key).await?;
307
308        let rt =
309            ts_runtime::Runtime::spawn(config.into(), auth_key, (&config.key_state).into()).await?;
310        // In TUN transport mode there is no application netstack, so the runtime has no command
311        // channel: that surfaces as `UnsupportedInTunMode`, which we map to a `None` channel rather
312        // than an error (the device is still usable for control-plane and peer-lookup APIs).
313        let channel = match rt.channel().await {
314            Ok(c) => Some(c),
315            Err(e) if e.kind == ts_runtime::ErrorKind::UnsupportedInTunMode => None,
316            Err(e) => return Err(e.into()),
317        };
318
319        Ok(Self {
320            runtime: rt,
321            channel,
322            enable_ipv6: config.enable_ipv6,
323            serve: std::sync::Mutex::new(None),
324            funnel: std::sync::Mutex::new(None),
325        })
326    }
327
328    /// The application netstack command channel, or an error in TUN transport mode (no application
329    /// netstack exists).
330    fn channel(&self) -> Result<&Channel, Error> {
331        self.channel
332            .as_ref()
333            .ok_or(Error::Internal(InternalErrorKind::UnsupportedInTunMode))
334    }
335
336    /// Get this [`Device`]'s IPv4 tailnet address.
337    pub async fn ipv4_addr(&self) -> Result<Ipv4Addr, Error> {
338        self.runtime
339            .control
340            .ask(ts_runtime::control_runner::Ipv4)
341            .await
342            .map_err(ts_runtime::Error::from)?
343            .ok_or(Error::Internal(InternalErrorKind::Actor))
344    }
345
346    /// Get this [`Device`]'s IPv6 tailnet address.
347    pub async fn ipv6_addr(&self) -> Result<Ipv6Addr, Error> {
348        self.runtime
349            .control
350            .ask(ts_runtime::control_runner::Ipv6)
351            .await
352            .map_err(ts_runtime::Error::from)?
353            .ok_or(Error::Internal(InternalErrorKind::Actor))
354    }
355
356    /// Bind a UDP socket to the specified [`SocketAddr`].
357    ///
358    /// Returns an error in TUN transport mode (there is no application netstack to bind on).
359    pub async fn udp_bind(&self, socket_addr: SocketAddr) -> Result<netstack::UdpSocket, Error> {
360        self.channel()?
361            .udp_bind(socket_addr)
362            .await
363            .map_err(Into::into)
364    }
365
366    /// Bind a TCP listener to the specified [`SocketAddr`].
367    ///
368    /// Returns an error in TUN transport mode (there is no application netstack to listen on).
369    pub async fn tcp_listen(
370        &self,
371        socket_addr: SocketAddr,
372    ) -> Result<netstack::TcpListener, Error> {
373        self.channel()?
374            .tcp_listen(socket_addr)
375            .await
376            .map_err(Into::into)
377    }
378
379    /// Register a fallback TCP handler (like `tsnet`'s `RegisterFallbackTCPHandler`).
380    ///
381    /// The callback is consulted for every inbound TCP flow that matches **no** explicit
382    /// [`Device::tcp_listen`] listener, with the flow's `(src, dst)` addresses. It returns
383    /// `(handler, intercept)`:
384    /// - `(_, false)` — decline; the next registered callback is tried.
385    /// - `(Some(h), true)` — claim the flow; `h` is handed the accepted [`netstack::TcpStream`].
386    /// - `(None, true)` — claim and reject the flow (the connection is closed).
387    ///
388    /// Multiple handlers may be registered; they are consulted in registration order and the first
389    /// to intercept wins. The returned [`FallbackTcpHandle`] deregisters the handler when dropped.
390    ///
391    /// Handlers serve flows over the overlay netstack only — never a host socket — and a flow no
392    /// handler claims is closed (fail-closed), never direct-dialed.
393    ///
394    /// Returns an error in TUN transport mode (there is no application netstack to attach to).
395    pub fn register_fallback_tcp_handler<F>(&self, cb: F) -> Result<FallbackTcpHandle, Error>
396    where
397        F: Fn(SocketAddr, SocketAddr) -> FallbackDecision + Send + Sync + 'static,
398    {
399        self.runtime
400            .register_fallback_tcp_handler(std::sync::Arc::new(cb))
401            .map_err(Into::into)
402    }
403
404    /// Resolve a tailnet peer (or this node) by MagicDNS name to its tailnet IPv4 address.
405    ///
406    /// This is an in-process lookup against the netmap we already hold — like `tsnet`'s in-memory
407    /// `dnsMap`, it does not query any DNS server (there is no `100.100.100.100` resolver). The
408    /// `name` may be a bare hostname or a fully-qualified MagicDNS name, with or without a trailing
409    /// dot, in any case (matching is case-insensitive). Returns `Ok(None)` if no tailnet node has
410    /// that name.
411    ///
412    /// Only MagicDNS names are resolved; names outside the tailnet are not looked up here, so the
413    /// caller's system resolver remains responsible for them. IPv6 is intentionally not resolved —
414    /// this fork operates IPv4-only on the tailnet.
415    pub async fn resolve(&self, name: &str) -> Result<Option<Ipv4Addr>, Error> {
416        if let Some(peer) = self.peer_by_name(name).await? {
417            return Ok(Some(peer.tailnet_address.ipv4.addr()));
418        }
419
420        // tsnet's dnsMap also resolves our own name; fall back to self when no peer matches.
421        let me = self.self_node().await?;
422        if me.matches_name(name) {
423            return Ok(Some(me.tailnet_address.ipv4.addr()));
424        }
425
426        Ok(None)
427    }
428
429    /// Connect to a tailnet peer by MagicDNS name and port over TCP.
430    ///
431    /// Resolves `name` via [`Device::resolve`] (an in-process netmap lookup, no DNS server), then
432    /// dials the resulting tailnet IPv4 address. Returns [`InternalErrorKind::BadRequest`] if the
433    /// name does not resolve to a tailnet node.
434    pub async fn connect_by_name(
435        &self,
436        name: &str,
437        port: u16,
438    ) -> Result<netstack::TcpStream, Error> {
439        let addr = self
440            .resolve(name)
441            .await?
442            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
443
444        self.tcp_connect((addr, port).into()).await
445    }
446
447    /// Connect to a TCP socket at the remote address.
448    ///
449    /// Returns an error in TUN transport mode (there is no application netstack to dial from).
450    pub async fn tcp_connect(&self, remote: SocketAddr) -> Result<netstack::TcpStream, Error> {
451        let channel = self.channel()?;
452
453        let ip: IpAddr = match remote.is_ipv4() {
454            true => self.ipv4_addr().await?.into(),
455            false => self.ipv6_addr().await?.into(),
456        };
457
458        // TODO(npry): collision checking
459        let ephemeral_port = rand::random_range(49152..=u16::MAX);
460
461        channel
462            .tcp_connect((ip, ephemeral_port).into(), remote)
463            .await
464            .map_err(Into::into)
465    }
466
467    /// Start a SOCKS5 proxy on a host loopback address that dials into the tailnet (Go
468    /// `tsnet.Server.Loopback`, SOCKS5 half).
469    ///
470    /// Binds a TCP listener on `127.0.0.1:0` (host loopback only — never an external interface) and
471    /// serves SOCKS5 (RFC 1928) with required username/password auth (RFC 1929): username `tsnet`,
472    /// password = the returned `proxy_cred`. Each `CONNECT` is dialed INTO the overlay via
473    /// [`Device::connect_by_name`] / [`Device::tcp_connect`] and spliced to the accepted host socket, so
474    /// a non-Rust host process can reach tailnet peers through the proxy. Returns the bound address, the
475    /// proxy credential, and a [`LoopbackHandle`] whose drop stops the listener.
476    ///
477    /// Anti-leak: the listener is loopback-only and every connection egresses over the overlay, never a
478    /// host socket — the host's real origin IP is never used to reach the destination. Unlike Go, the
479    /// LocalAPI HTTP surface is not served (this fork exposes status/whois/id-token natively on
480    /// `Device`); only the SOCKS5 proxy is provided.
481    ///
482    /// Returns an error in TUN transport mode (no application netstack to dial from).
483    pub async fn loopback(&self) -> Result<(std::net::SocketAddr, String, LoopbackHandle), Error> {
484        // Capture only cloneable pieces — never `&self` — for the spawned accept loop: a clone of the
485        // netstack command channel, this device's own overlay IPv4 (fetched once), and a boxed
486        // resolver closure over clones of the control + peer-tracker actor refs. The resolver
487        // replicates `Device::resolve` (peer-by-name, falling back to this node's own name).
488        let channel = self.channel()?.clone();
489        let self_ipv4 = self.ipv4_addr().await?;
490
491        let control = self.runtime.control.clone();
492        let peer_tracker = self.runtime.peer_tracker.clone();
493        let resolve: loopback::Resolver = std::sync::Arc::new(move |name: String| {
494            let control = control.clone();
495            let peer_tracker = peer_tracker.clone();
496            Box::pin(async move {
497                let pt = peer_tracker
498                    .upgrade()
499                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
500                let peer = pt
501                    .ask(ts_runtime::peer_tracker::PeerByName { name: name.clone() })
502                    .await
503                    .map_err(ts_runtime::Error::from)?;
504                if let Some(peer) = peer {
505                    return Ok(Some(peer.tailnet_address.ipv4.addr()));
506                }
507                // tsnet's dnsMap also resolves our own name; fall back to self.
508                let me = control
509                    .ask(ts_runtime::control_runner::SelfNode)
510                    .await
511                    .map_err(ts_runtime::Error::from)?
512                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
513                if me.matches_name(&name) {
514                    Ok(Some(me.tailnet_address.ipv4.addr()))
515                } else {
516                    Ok(None)
517                }
518            }) as std::pin::Pin<Box<dyn std::future::Future<Output = _> + Send>>
519        });
520
521        let dialer = loopback::OverlayDialer::new(channel, self_ipv4, resolve);
522        loopback::start(dialer).await
523    }
524
525    /// Get our node info.
526    pub async fn self_node(&self) -> Result<NodeInfo, Error> {
527        self.runtime
528            .control
529            .ask(ts_runtime::control_runner::SelfNode)
530            .await
531            .map_err(ts_runtime::Error::from)?
532            .ok_or(Error::Internal(InternalErrorKind::Actor))
533    }
534
535    /// This node's key-expiry instant as Unix seconds (`Node.KeyExpiry` in Go), or `Ok(None)` if
536    /// the key never expires.
537    ///
538    /// Like Go, this fork is **reactive** about key expiry — it reports it rather than rotating the
539    /// node key in the background. A caller can schedule re-authentication around this time; on
540    /// expiry, re-create the [`Device`] (which re-registers), supplying a fresh node key + the prior
541    /// `old_node_key` to rotate, or the same key to refresh.
542    pub async fn self_key_expiry_unix(&self) -> Result<Option<i64>, Error> {
543        Ok(self.self_node().await?.key_expiry_unix())
544    }
545
546    /// Whether this node's key has expired as of now (`!KeyExpiry.IsZero() && KeyExpiry.Before(now)`
547    /// in Go). A key with no expiry is never expired. See [`Device::self_key_expiry_unix`] for the
548    /// reactive-rotation note.
549    pub async fn self_key_expired(&self) -> Result<bool, Error> {
550        let now = std::time::SystemTime::now()
551            .duration_since(std::time::UNIX_EPOCH)
552            .map(|d| d.as_secs() as i64)
553            // An unreadable clock (pre-epoch) is treated as the far future so a time-limited key
554            // looks expired — fail-safe toward prompting re-auth rather than trusting a stale key.
555            .unwrap_or(i64::MAX);
556        Ok(self.self_node().await?.key_expired_at_unix(now))
557    }
558
559    /// Fetch the current Tailscale SSH policy pushed by control, if any.
560    ///
561    /// Returns `Ok(None)` when control has not sent an SSH policy. The SSH server treats an absent
562    /// or empty policy as **deny-all** (fail-closed). Used by the SSH auth path
563    /// ([`SshPolicy::evaluate`][ts_control::SshPolicy::evaluate]) to authorize incoming
564    /// connections.
565    pub async fn ssh_policy(&self) -> Result<Option<ts_control::SshPolicy>, Error> {
566        self.runtime
567            .control
568            .ask(ts_runtime::control_runner::CurrentSshPolicy)
569            .await
570            .map_err(ts_runtime::Error::from)
571            .map_err(Into::into)
572    }
573
574    /// Look up a peer by name.
575    pub async fn peer_by_name(&self, name: &str) -> Result<Option<NodeInfo>, Error> {
576        let pt = self
577            .runtime
578            .peer_tracker
579            .upgrade()
580            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
581
582        pt.ask(ts_runtime::peer_tracker::PeerByName {
583            name: name.to_string(),
584        })
585        .await
586        .map_err(ts_runtime::Error::from)
587        .map_err(Into::into)
588    }
589
590    /// Look up a peer by ip.
591    pub async fn peer_by_tailnet_ip(&self, ip: IpAddr) -> Result<Option<NodeInfo>, Error> {
592        let pt = self
593            .runtime
594            .peer_tracker
595            .upgrade()
596            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
597
598        pt.ask(ts_runtime::peer_tracker::PeerByTailnetIp { ip })
599            .await
600            .map_err(ts_runtime::Error::from)
601            .map_err(Into::into)
602    }
603
604    /// Look up the peer(s) with the most-specific route matches for `ip`.
605    ///
606    /// This reports which peers *advertise* a route covering `ip`, independent of this device's
607    /// `accept_routes` setting — analogous to the Go client's informational `PrimaryRoutes`. It is
608    /// not a reachability oracle: with `accept_routes` off, the dataplane will not actually route
609    /// to (or accept return traffic from) advertised subnet routes even if this returns a peer.
610    pub async fn peers_with_route(&self, ip: IpAddr) -> Result<Vec<NodeInfo>, Error> {
611        let pt = self
612            .runtime
613            .peer_tracker
614            .upgrade()
615            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
616
617        pt.ask(ts_runtime::peer_tracker::PeerByAcceptedRoute { ip })
618            .await
619            .map_err(ts_runtime::Error::from)
620            .map_err(Into::into)
621    }
622
623    /// List the Taildrop files this device has fully received and not yet consumed (Go LocalAPI
624    /// `WaitingFiles`).
625    ///
626    /// Returns the files waiting under the configured `taildrop_dir`, sorted by name. Returns an
627    /// empty list when Taildrop is disabled (`Config::taildrop_dir` unset) — fail-closed, never an
628    /// error for the disabled case. A filesystem error while listing surfaces as
629    /// [`InternalErrorKind::Actor`].
630    pub fn taildrop_waiting_files(&self) -> Result<Vec<WaitingFile>, Error> {
631        let Some(store) = self.runtime.taildrop_store() else {
632            return Ok(Vec::new());
633        };
634        store
635            .waiting_files()
636            .map_err(|_| Error::Internal(InternalErrorKind::Actor))
637    }
638
639    /// Open a received Taildrop file by name for reading, returning the handle and its size (Go
640    /// LocalAPI `OpenFile`).
641    ///
642    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
643    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
644    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
645    pub fn taildrop_open_file(&self, name: &str) -> Result<(std::fs::File, u64), Error> {
646        let store = self
647            .runtime
648            .taildrop_store()
649            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
650        store.open_file(name).map_err(taildrop_err)
651    }
652
653    /// Delete a received Taildrop file by name (Go LocalAPI `DeleteFile`).
654    ///
655    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
656    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
657    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
658    pub fn taildrop_delete_file(&self, name: &str) -> Result<(), Error> {
659        let store = self
660            .runtime
661            .taildrop_store()
662            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
663        store.delete_file(name).map_err(taildrop_err)
664    }
665
666    /// Send a local file to a tailnet `peer` via Taildrop (Go `PushFile` / `tailscale file cp`).
667    ///
668    /// Pushes `content_length` bytes from `reader` to the peer's peerAPI as
669    /// `PUT /v0/put/<name>` over the overlay netstack — the sending counterpart to the receive store
670    /// surfaced by [`Device::taildrop_waiting_files`]. The transfer rides the encrypted WireGuard
671    /// overlay, never a host socket. The body is streamed from offset 0 (no resume).
672    ///
673    /// The destination is derived **solely from `peer`'s own node record**
674    /// ([`NodeInfo::peerapi_addr`][ts_control::Node::peerapi_addr]): its advertised tailnet IPv4 and
675    /// `peerapi4` port. The caller obtains `peer` from [`Device::peer_by_name`] /
676    /// [`Device::peer_by_tailnet_ip`], so it is always a current netmap peer — a raw control-supplied
677    /// or attacker-chosen address can never be targeted. As defense in depth, the resolved address is
678    /// additionally asserted to be a Tailscale CGNAT IP before dialing.
679    ///
680    /// Returns [`InternalErrorKind::BadRequest`] when the peer advertises no IPv4 peerAPI (so it
681    /// cannot receive files), when the name is invalid, or when the peer refuses the transfer
682    /// (`403`/`409`/unexpected status); [`Error::Timeout`] on a dial failure or timeout; and
683    /// [`InternalErrorKind::Io`] on a mid-transfer stream error.
684    pub async fn send_file<R>(
685        &self,
686        peer: &NodeInfo,
687        name: &str,
688        content_length: u64,
689        reader: R,
690    ) -> Result<(), Error>
691    where
692        R: tokio::io::AsyncRead + Unpin,
693    {
694        let channel = self.channel()?;
695
696        // Destination comes only from the peer's own node record — never an arbitrary address.
697        let dst = peer
698            .peerapi_addr()
699            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
700        // Defense in depth: refuse to dial anything outside the Tailscale CGNAT range, so a
701        // malformed node record can't steer the PUT at a non-tailnet host.
702        if !ts_control::is_tailscale_ip(dst.ip()) {
703            return Err(Error::Internal(InternalErrorKind::BadRequest));
704        }
705
706        let self_ipv4 = self.ipv4_addr().await?;
707
708        ts_runtime::taildrop_send::send_file(channel, self_ipv4, dst, name, content_length, reader)
709            .await
710            .map_err(taildrop_send_err)
711    }
712
713    /// Begin a debug packet capture, streaming a pcap of every packet crossing the dataplane to
714    /// `writer` (Go `tsnet.Server.CapturePcap`).
715    ///
716    /// Installs a capture hook on the running dataplane: from now until [`Device::stop_capture`] is
717    /// called (or another capture replaces this one), a copy of every plaintext IP packet on the
718    /// datapath — outbound (pre-encrypt) and inbound (post-decrypt) — is framed and written to
719    /// `writer`. The 24-byte pcap global header is written immediately on success.
720    ///
721    /// The format is byte-faithful classic pcap with Tailscale's `LINKTYPE_USER0` + 4-byte path
722    /// preamble per record (see [`ts_runtime::capture`]); a resulting file opens in Wireshark, and
723    /// with Tailscale's `ts-dissector.lua` the direction/path of each packet decodes.
724    ///
725    /// The hook runs **inline on the single-threaded dataplane step**, so `writer` must not block for
726    /// long — a slow writer back-pressures the datapath. Records are **not** flushed per packet (that
727    /// would be a syscall on every packet on the dataplane thread); buffered bytes are flushed when
728    /// the writer is dropped on [`Device::stop_capture`]. Wrap `writer` in a [`std::io::BufWriter`] if
729    /// you want buffering. A write error is swallowed per-packet (the capture silently drops that
730    /// record) rather than tearing down the datapath; call [`Device::stop_capture`] to end it. Returns
731    /// an error only if the dataplane actor is unreachable or the initial global-header write fails.
732    pub async fn capture_pcap<W>(&self, writer: W) -> Result<(), Error>
733    where
734        W: std::io::Write + Send + 'static,
735    {
736        let sink = std::sync::Arc::new(std::sync::Mutex::new(
737            ts_runtime::capture::PcapSink::new(writer)
738                .map_err(|_| Error::Internal(InternalErrorKind::Io))?,
739        ));
740        let hook: ts_runtime::CaptureHook = std::sync::Arc::new(move |path, pkt: &[u8]| {
741            if let Ok(mut sink) = sink.lock() {
742                // A per-packet write failure (e.g. a closed pipe) silently drops that record rather
743                // than tearing down the datapath; the caller ends capture via `stop_capture`.
744                drop(sink.log_packet(path.code(), pkt));
745            }
746        });
747        self.runtime.install_capture(Some(hook)).await?;
748        Ok(())
749    }
750
751    /// Stop a debug packet capture started by [`Device::capture_pcap`] (Go `ClearCaptureSink`).
752    ///
753    /// Clears the dataplane capture hook; the writer is dropped (its remaining buffered bytes are
754    /// flushed by its own `Drop`). Idempotent — clearing when no capture is installed is a no-op.
755    /// Returns an error only if the dataplane actor is unreachable.
756    pub async fn stop_capture(&self) -> Result<(), Error> {
757        self.runtime.install_capture(None).await?;
758        Ok(())
759    }
760
761    /// Snapshot of this device and its tailnet peers (like `tailscale status`).
762    ///
763    /// Combines this node's self info with the current peer set: each [`StatusNode`] reports the
764    /// stable id, display name, tailnet IPs, advertised routes, and exit-node flag. (Per-peer
765    /// `online`/user/capabilities are honestly `None`/empty in this fork — the domain node model
766    /// does not yet carry the wire-level liveness/login fields; see `ts_runtime::status` docs.)
767    pub async fn status(&self) -> Result<Status, Error> {
768        self.runtime.status().await.map_err(Into::into)
769    }
770
771    /// Fetch the current Tailnet Lock (TKA) status pushed by control, if any.
772    ///
773    /// Returns `Ok(None)` when control has sent no `TKAInfo` (tailnet lock not in use, or no change
774    /// observed yet). The returned [`TkaStatus`][ts_control::TkaStatus] carries the authority head
775    /// (a base32 `AUMHash`, decode with [`tka::AumHash::from_base32`][ts_tka::AumHash::from_base32])
776    /// and the disablement signal. Signature verification of a peer's node-key signature against the
777    /// authority is performed with the [`tka`] module's [`tka::Authority`][ts_tka::Authority].
778    pub async fn tka_status(&self) -> Result<Option<ts_control::TkaStatus>, Error> {
779        self.runtime
780            .control
781            .ask(ts_runtime::control_runner::CurrentTkaStatus)
782            .await
783            .map_err(ts_runtime::Error::from)
784            .map_err(Into::into)
785    }
786
787    /// Request an OIDC **ID token** from control for this node, scoped to `audience` (workload-
788    /// identity federation, like `tailscale`'s `id-token` LocalAPI).
789    ///
790    /// Returns a signed JWT whose `sub` claim is this node's MagicDNS name and whose `aud` claim is
791    /// `audience`, suitable for presenting to a third-party relying party (e.g. AWS/GCP
792    /// workload-identity federation). The node is the token *subject*, not the authenticator — this
793    /// is token issuance over the Noise transport (`POST /machine/id-token`), not a login path.
794    /// Requires the control plane to support capability version ≥ 30.
795    pub async fn fetch_id_token(&self, audience: &str) -> Result<String, ts_control::IdTokenError> {
796        self.runtime.fetch_id_token(audience.to_string()).await
797    }
798
799    /// Log this node out of the tailnet — deregister it from the control plane (the equivalent of
800    /// Go `tsnet`'s `LocalClient.Logout`).
801    ///
802    /// Re-`POST`s `/machine/register` with this node's current node key and a past expiry, which the
803    /// control plane honors by **expiring the node now**: it drops out of every peer's netmap and
804    /// must re-register (re-authenticate) to rejoin.
805    ///
806    /// This is primarily for **non-ephemeral** nodes. An ephemeral node is garbage-collected by
807    /// control shortly after it disconnects, but a persistent node lingers in the tailnet
808    /// (visible to peers, counting against the machine limit) for up to ~24h after the process exits
809    /// unless explicitly logged out. Call this before [`shutdown`](Self::shutdown) to deregister
810    /// immediately. Calling it on an ephemeral node simply brings the GC forward; it is idempotent,
811    /// so logging out an already-gone node is not an error.
812    ///
813    /// This is a **control-plane state change only**: it does not tear down the local datapath (do
814    /// that via [`shutdown`](Self::shutdown)), and it does not delete or rotate the on-disk node key
815    /// — re-registering with the same key (a fresh [`Device::new`]) is the re-login path.
816    pub async fn logout(&self) -> Result<(), ts_control::LogoutError> {
817        self.runtime.logout().await
818    }
819
820    /// Snapshot this node's client metrics in Prometheus text exposition format.
821    ///
822    /// Mirrors Go Tailscale's `clientmetric` registry: process-global counters/gauges incremented
823    /// on the datapath hot loops (e.g. `magicsock_send_udp`, `magicsock_recv_data_bytes_udp`),
824    /// rendered as `# TYPE <name> <kind>\n<name> <value>\n` per metric, sorted by name. (Go `tsnet`
825    /// exposes no metrics method of its own, so this is the fork's clean public surface.) The
826    /// registry is process-global, so the output covers every `Device` in the process.
827    pub fn metrics(&self) -> String {
828        ts_metrics::write_prometheus()
829    }
830
831    /// Map a tailnet source `addr` to the node that owns its IP (like `tsnet`'s `WhoIs`).
832    ///
833    /// Only the IP of `addr` is used; the port is ignored. Returns `Ok(None)` if no tailnet node
834    /// owns that address.
835    pub async fn whois(&self, addr: SocketAddr) -> Result<Option<WhoIs>, Error> {
836        self.runtime.whois(addr).await.map_err(Into::into)
837    }
838
839    /// Change the selected exit node at runtime, without recreating the [`Device`] — the equivalent
840    /// of Go `tsnet`'s `LocalClient.EditPrefs(ExitNodeID/ExitNodeIP)`.
841    ///
842    /// The peer may be named by stable node ID, tailnet IP, or MagicDNS name via
843    /// [`ExitNodeSelector`] (a bare IP or name parses with `selector.parse()`); this is the same
844    /// selector type as [`Config::exit_node`](crate::Config::exit_node), so the construction-time
845    /// and runtime paths are identical. Passing `None` clears the exit node — internet-bound traffic
846    /// is then dropped (fail-closed) unless this node egresses directly.
847    ///
848    /// The change is applied immediately: the new selector is re-resolved against the live peer set
849    /// and the outbound route + inbound source filter are recomputed at once. A selector for a peer
850    /// not yet in the netmap simply takes effect once that peer appears.
851    ///
852    /// Only NEW flows use the changed exit; in-flight connections are not torn down and continue
853    /// egressing via the previously-selected exit until they close.
854    pub async fn set_exit_node(&self, exit_node: Option<ExitNodeSelector>) -> Result<(), Error> {
855        self.runtime
856            .set_exit_node(exit_node)
857            .await
858            .map_err(Into::into)
859    }
860
861    /// The currently-selected exit node, or `None` if none is selected.
862    pub fn exit_node(&self) -> Option<ExitNodeSelector> {
863        self.runtime.exit_node()
864    }
865
866    /// The stable id of the exit node traffic is **currently** egressing through, or `None` if none
867    /// is engaged (the equivalent of Go `tsnet`'s `Status.ExitNodeStatus.ID`).
868    ///
869    /// This differs from [`exit_node`](Self::exit_node), which returns the *configured* selector:
870    /// the active exit node is the route updater's resolved, fail-closed answer. It is `None` when
871    /// no exit node is configured, the configured selector matches no current peer, or the matched
872    /// peer no longer advertises a default route (egress is then dropped, fail-closed). Match the id
873    /// against [`Status::peers`](crate::Status::peers) (via [`status`](Self::status)) for details.
874    pub fn active_exit_node(&self) -> Option<ts_control::StableNodeId> {
875        self.runtime.active_exit_node()
876    }
877
878    /// Watch for netmap changes: the returned receiver's value is the current set of peer
879    /// [`StatusNode`]s and updates on every netmap change (like subscribing to `ipn` notifications).
880    pub async fn watch_netmap(
881        &self,
882    ) -> Result<tokio::sync::watch::Receiver<Vec<StatusNode>>, Error> {
883        self.runtime.watch_netmap().await.map_err(Into::into)
884    }
885
886    /// The current device connection-[`DeviceState`] (`Connecting` / `Running` / `NeedsLogin` /
887    /// `Expired` / `Failed`).
888    pub fn device_state(&self) -> DeviceState {
889        self.runtime.device_state()
890    }
891
892    /// Watch the device connection-[`DeviceState`], reacting push-style to control connection
893    /// transitions instead of polling [`status`](Self::status).
894    ///
895    /// Returns a [`tokio::sync::watch::Receiver`]; await its
896    /// [`changed`](tokio::sync::watch::Receiver::changed) to be woken on each transition. The
897    /// initial value is the current state.
898    pub fn watch_state(&self) -> tokio::sync::watch::Receiver<DeviceState> {
899        self.runtime.watch_state()
900    }
901
902    /// Wait until the device finishes registering, returning a typed outcome — the clean
903    /// replacement for polling [`ipv4_addr`](Self::ipv4_addr) in a loop.
904    ///
905    /// Resolves `Ok(())` once the device is [`DeviceState::Running`]. On a non-running outcome it
906    /// returns a typed [`RegistrationError`]:
907    /// - [`AuthRejected`](RegistrationError::AuthRejected) — bad/expired/unknown auth key;
908    ///   **permanent** (re-pair).
909    /// - [`NeedsLogin`](RegistrationError::NeedsLogin) — interactive authorization required;
910    ///   **not permanent** (the runtime keeps retrying and reaches `Running` once the user
911    ///   authorizes). Auth-key callers treat this as failure; interactive callers should ignore it
912    ///   and drive the flow via [`watch_state`](Self::watch_state).
913    /// - [`NetworkUnreachable`](RegistrationError::NetworkUnreachable) — **transient** (retry).
914    /// - [`Timeout`](RegistrationError::Timeout) — no settled state within `timeout` (`None` waits
915    ///   indefinitely).
916    ///
917    /// [`KeyExpired`](RegistrationError::KeyExpired) is not produced here (a key expires only after
918    /// the node is up); observe it via [`watch_state`](Self::watch_state). Use
919    /// [`RegistrationError::is_permanent`] to branch "re-pair" vs. "retry / drive login".
920    pub async fn wait_until_running(
921        &self,
922        timeout: Option<Duration>,
923    ) -> Result<(), RegistrationError> {
924        self.runtime.wait_until_running(timeout).await
925    }
926
927    /// Ping a tailnet peer over the overlay with an ICMPv4 echo, returning the round-trip time
928    /// (like `tailscale ping`).
929    ///
930    /// The echo is sent from this device's own tailnet IPv4 over the overlay netstack — never a
931    /// host socket. IPv6 destinations return [`PingError::Ipv6Unsupported`] (this fork is
932    /// IPv4-only on the tailnet). A peer answers from its own OS stack; this netstack does not
933    /// auto-reply to echo requests.
934    ///
935    /// In TUN transport mode there is no application netstack to ping from; this surfaces as
936    /// [`PingError::Timeout`] (the same error this method already uses for an unavailable source
937    /// address — `PingError` carries no dedicated "unsupported" variant).
938    pub async fn ping(&self, dst: IpAddr, timeout: Duration) -> Result<Duration, PingError> {
939        let channel = self.channel().map_err(|_| PingError::Timeout)?;
940        let src = self.ipv4_addr().await.map_err(|_| PingError::Timeout)?;
941        ts_netstack_smoltcp::ping(channel, src, dst, timeout).await
942    }
943
944    /// Obtain a TLS certificate for a node's MagicDNS `name` (like `tsnet`'s `GetCertificate`).
945    ///
946    /// **Fail-closed without the `acme` feature.** By default this fork has no client-side ACME
947    /// engine wired in, so this returns [`ts_control::CertError::Unimplemented`] (after a
948    /// tailnet-name check) — it NEVER self-signs and NEVER returns a placeholder certificate
949    /// ([`ts_control::MISSING_CERT_RPC`] names what is missing).
950    ///
951    /// **With the `acme` feature** this instead drives the client-side ACME DNS-01 engine to issue a
952    /// real Let's Encrypt certificate for `name`, publishing the challenge TXT via the node's
953    /// `POST /machine/set-dns` RPC (routed through the control runner). SaaS-only: a self-hosted
954    /// control plane may 501 on set-dns, surfaced as [`ts_control::CertError::Acme`].
955    #[cfg(not(feature = "acme"))]
956    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
957        ts_control::get_certificate(name).await
958    }
959
960    /// See the no-`acme` variant for the contract; with `acme` this issues a real cert via the
961    /// runtime's ACME engine (`Device → Runtime → ControlRunner → issue_certificate_via_setdns`).
962    #[cfg(feature = "acme")]
963    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
964        self.runtime.get_certificate(name.to_string()).await
965    }
966
967    /// Build a [`TlsAcceptor`] terminating TLS for `cfg.name` on the overlay (like `tsnet`'s
968    /// `ListenTLS`).
969    ///
970    /// Obtains the certificate via [`Device::get_certificate`] — so with the `acme` feature this
971    /// issues a real Let's Encrypt cert (when the control plane answers `set-dns`), and without it
972    /// (or when issuance is unavailable) it surfaces the same fail-closed
973    /// [`ts_control::CertError`] rather than ever serving a self-signed cert or downgrading to
974    /// plaintext. Terminate accepted overlay streams with [`ts_control::accept_tls`].
975    pub async fn listen_tls(
976        &self,
977        cfg: &ts_control::ServeConfig,
978    ) -> Result<TlsAcceptor, ts_control::CertError> {
979        // Route through Device::get_certificate (the acme-aware issuance path) rather than
980        // ts_control::listen_tls, which only knows the non-acme stub. Validate the serve config
981        // first (same fail-closed checks ts_control::listen_tls applies), then assemble the acceptor.
982        cfg.validate()?;
983        let cert = self.get_certificate(&cfg.name).await?;
984        ts_control::tls_acceptor(cert)
985    }
986
987    /// The currently-stored Serve config (like `tsnet`'s `GetServeConfig`).
988    ///
989    /// Returns the config last passed to [`Device::set_serve_config`], or an empty
990    /// [`ts_control::ServeState`] (no ports) if none was ever set. Pure read — does not touch the
991    /// network.
992    pub fn get_serve_config(&self) -> ts_control::ServeState {
993        match &*self.serve.lock().unwrap_or_else(|e| e.into_inner()) {
994            Some(mgr) => mgr.get(),
995            None => ts_control::ServeState::default(),
996        }
997    }
998
999    /// Replace this node's Serve config and (re)bind its tailnet ports (like `tsnet`'s
1000    /// `SetServeConfig`, REPLACE semantics).
1001    ///
1002    /// `state` becomes the **whole** config (full-replace reconcile: every previously-bound serve
1003    /// port's accept loop is torn down and the new config's ports are bound from scratch). For each
1004    /// configured port the manager binds an overlay listener on this node's tailnet IPv4 and
1005    /// dispatches per [`ts_control::ServeTarget`]:
1006    /// - [`Accept`](ts_control::ServeTarget::Accept) — the TLS-terminated stream is handed back over
1007    ///   the returned [`ServeAcceptedReceiver`](ts_runtime::serve::ServeAcceptedReceiver) (the
1008    ///   in-process stand-in for `ListenTLS`'s `net.Listener`).
1009    /// - [`Proxy`](ts_control::ServeTarget::Proxy) — reverse-proxy the decrypted stream to a local
1010    ///   host backend.
1011    /// - [`Text`](ts_control::ServeTarget::Text) — write a fixed body and close.
1012    /// - [`TcpForward`](ts_control::ServeTarget::TcpForward) — forward the **raw** (non-TLS) stream
1013    ///   to a local host backend.
1014    ///
1015    /// **Fail-closed.** `state.validate()` runs first. Every TLS-terminating port's acceptor is
1016    /// obtained up-front via [`Device::listen_tls`] (the ACME-aware cert path); if any cert cannot be
1017    /// issued the whole call fails with that [`ts_control::CertError`] and **nothing is bound** — a
1018    /// TLS port never downgrades to plaintext.
1019    ///
1020    /// **Anti-leak.** Listeners bind the overlay netstack only (never a host socket). The
1021    /// `Proxy`/`TcpForward` backend dial is a local host socket to the embedder's own backend (like
1022    /// Go's reverse-proxy to `127.0.0.1`), intentionally NOT routed through the exit-egress
1023    /// forwarder. A backend dial failure drops that connection; it never falls back.
1024    ///
1025    /// Returns an error in TUN transport mode (there is no application netstack to bind on). The
1026    /// previous config's accept loops (and any earlier `ServeAcceptedReceiver`) stop when this
1027    /// returns; the new receiver delivers every `Accept`-port connection.
1028    pub async fn set_serve_config(
1029        &self,
1030        state: ts_control::ServeState,
1031    ) -> Result<ts_runtime::serve::ServeAcceptedReceiver, Error> {
1032        state
1033            .validate()
1034            .map_err(|_| Error::Internal(InternalErrorKind::BadRequest))?;
1035
1036        // Fail-closed: build every TLS-terminating port's acceptor up-front via the ACME-aware cert
1037        // path. If any cert can't be issued, return before binding anything (no plaintext downgrade).
1038        let mut resolved = std::collections::BTreeMap::new();
1039        for (port, target) in &state.ports {
1040            let acceptor = if target.terminates_tls() {
1041                let cfg = ts_control::ServeConfig {
1042                    name: state.name.clone(),
1043                    port: *port,
1044                    target: target.clone(),
1045                };
1046                Some(self.listen_tls(&cfg).await.map_err(|_| {
1047                    // Cert issuance is fail-closed in this fork; surface as a request error rather
1048                    // than ever binding a plaintext TLS port.
1049                    Error::Internal(InternalErrorKind::BadRequest)
1050                })?)
1051            } else {
1052                None
1053            };
1054            resolved.insert(
1055                *port,
1056                ts_runtime::serve::ResolvedPort {
1057                    target: target.clone(),
1058                    acceptor,
1059                },
1060            );
1061        }
1062
1063        // The manager binds the OVERLAY netstack on this node's own tailnet IPv4.
1064        let self_ipv4 = self.ipv4_addr().await?;
1065        let channel = self.channel()?.clone();
1066
1067        let mut slot = self.serve.lock().unwrap_or_else(|e| e.into_inner());
1068        let mgr =
1069            slot.get_or_insert_with(|| ts_runtime::serve::ServeManager::new(channel, self_ipv4));
1070        Ok(mgr.set(state, resolved))
1071    }
1072
1073    /// Expose a tailnet TLS service to the public internet via Tailscale Funnel (like `tsnet`'s
1074    /// `ListenFunnel`), returning a [`FunnelAcceptedReceiver`](ts_runtime::funnel::FunnelAcceptedReceiver)
1075    /// that delivers each TLS-terminated public connection.
1076    ///
1077    /// **Two fail-closed gates, then the live ingress listener.** First the node-attribute gate is
1078    /// fully enforced from this node's own capability map (mirroring Go `ipn.NodeCanFunnel` +
1079    /// `ipn.CheckFunnelPort`): the tailnet admin must have enabled HTTPS and granted the `funnel`
1080    /// node attribute, and `cfg.port` must be in the set the `funnel-ports` capability allows —
1081    /// otherwise this returns [`ts_control::FunnelError::NotAllowed`] /
1082    /// [`ts_control::FunnelError::PortNotAllowed`] before touching any cert or network. Then the
1083    /// node's `*.ts.net` certificate is obtained via the ACME-aware [`Device::get_certificate`] (the
1084    /// Funnel hostname *is* the node's MagicDNS name, so its DNS-01 cert matches); fail-closed on
1085    /// [`ts_control::FunnelError::Cert`] — no self-signed or plaintext fallback.
1086    ///
1087    /// On success a [`FunnelManager`](ts_runtime::funnel::FunnelManager) is registered: its ingress
1088    /// sink is installed into the runtime's peerAPI `/v0/ingress` slot (making that route live without
1089    /// restarting the peerAPI server), and the `HostInfo.IngressEnabled` map-request signal is set so
1090    /// control routes Funnel traffic to this node. Public Funnel bytes arrive as a relay POST to
1091    /// `/v0/ingress`, are membership-gated + `101`-hijacked into a raw stream, TLS-terminated by the
1092    /// manager, and delivered over the returned receiver.
1093    ///
1094    /// **Where the relay comes from.** The public ingress **relay + DNS mapping** that feed
1095    /// `/v0/ingress` are Tailscale infrastructure ([`ts_control::MISSING_FUNNEL_RELAY`]), provisioned
1096    /// automatically against real Tailscale SaaS with a Funnel-enabled ACL; against a self-hosted
1097    /// control plane no relay exists, so the listener is correct but never fed.
1098    ///
1099    /// Anti-leak: Funnel TLS terminates only on the overlay netstack (the hijacked ingress stream
1100    /// arrives on the overlay peerAPI listener), never a host socket; there is no self-signed or
1101    /// plaintext fallback. A new `listen_funnel` replaces the previous manager (its pump + sink tear
1102    /// down); dropping the `Device` tears it down too.
1103    pub async fn listen_funnel(
1104        &self,
1105        cfg: &ts_control::ServeConfig,
1106        opts: ts_control::FunnelOptions,
1107    ) -> Result<ts_runtime::funnel::FunnelAcceptedReceiver, ts_control::FunnelError> {
1108        // Gate 1 (fail-closed, no network): node-attribute + funnel-port access from our cap map.
1109        let me = self
1110            .self_node()
1111            .await
1112            .map_err(|_| ts_control::FunnelError::NotAllowed)?;
1113        cfg.validate()?;
1114        ts_control::funnel_access(&me, cfg.port)?;
1115
1116        // Gate 2 (fail-closed): obtain the node's `*.ts.net` cert via the ACME-aware path and build
1117        // the TLS acceptor. A cert failure surfaces as FunnelError::Cert — never a plaintext listener.
1118        let cert = self
1119            .get_certificate(&cfg.name)
1120            .await
1121            .map_err(ts_control::FunnelError::Cert)?;
1122        let acceptor = ts_control::tls_acceptor(cert).map_err(ts_control::FunnelError::Cert)?;
1123
1124        // `opts.funnel_only` (reject tailnet-internal connections) is accepted for surface stability;
1125        // the ingress data path only ever carries relay-delivered public traffic, so there is no
1126        // tailnet-internal leg on this listener to reject. Documented as a no-op here for now.
1127        let _ = opts;
1128
1129        // Build the funnel manager + its ingress sink + the hand-back receiver, install the sink into
1130        // the runtime's shared peerAPI `/v0/ingress` slot (making the route live), and flip the
1131        // IngressEnabled map signal. Hold the manager on the device so its pump/sink live as long as
1132        // the listener; replacing a prior manager tears the old one down on drop at end of scope.
1133        let (manager, sink, receiver) = ts_runtime::funnel::FunnelManager::new(acceptor);
1134        {
1135            let slot = self.runtime.funnel_ingress_slot();
1136            *slot.lock().unwrap_or_else(|e| e.into_inner()) = Some(sink);
1137        }
1138        self.runtime
1139            .ingress_active_flag()
1140            .store(true, std::sync::atomic::Ordering::Relaxed);
1141
1142        let old = {
1143            let mut held = self.funnel.lock().unwrap_or_else(|e| e.into_inner());
1144            held.replace(manager)
1145        };
1146        drop(old);
1147
1148        Ok(receiver)
1149    }
1150
1151    /// Host a Tailscale **VIP service** (`svc:<label>`) by binding an overlay listener on the
1152    /// service's control-assigned virtual IP (like `tsnet`'s `ListenService`).
1153    ///
1154    /// **Fail-closed.** Mirrors Go `tsnet.Server.ListenService`'s preconditions, enforced from this
1155    /// node's own netmap state ([`ts_control::resolve_service_listen`]): the `name` must be a valid
1156    /// `svc:<dns-label>`, this node must be **tagged** (Go `ErrUntaggedServiceHost`), and control
1157    /// must have assigned the service a VIP address on this node (delivered via the `service-host`
1158    /// node-capability — see [`ts_control::Node::service_addresses`]). Any unmet precondition
1159    /// returns a typed [`ts_control::ServiceError`] before binding anything.
1160    ///
1161    /// When all hold, this binds a [`tcp_listen`][Device::tcp_listen] on the service VIP and the
1162    /// configured `mode` port over the **overlay netstack** (never a host socket) and returns the
1163    /// listener. The netstack already accepts packets for control-assigned VIPs (they are injected
1164    /// alongside the node's own tailnet address), so the listener is reachable by tailnet peers.
1165    ///
1166    /// The `Tun`/L3 service mode is unsupported (a TODO in upstream tsnet); only TCP/HTTP modes
1167    /// (which bind the same VIP:port at the listen layer) are offered. Returns an error in TUN
1168    /// transport mode (there is no application netstack to bind on).
1169    pub async fn listen_service(
1170        &self,
1171        name: &str,
1172        mode: ts_control::ServiceMode,
1173    ) -> Result<netstack::TcpListener, ts_control::ServiceError> {
1174        let me = self
1175            .self_node()
1176            .await
1177            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))?;
1178        let listen_addr = ts_control::resolve_service_listen(&me, name, mode, self.enable_ipv6)?;
1179        self.tcp_listen(listen_addr)
1180            .await
1181            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))
1182    }
1183
1184    /// Attempt to gracefully shut down this device's runtime.
1185    ///
1186    /// Reports whether the device was fully shut down before the timeout. It is still shut
1187    /// down if it timed out, just more violently and with potential resource leaks.
1188    ///
1189    /// If `timeout` is `None`, then shutdown will never time-out.
1190    pub async fn shutdown(self, timeout: Option<Duration>) -> bool {
1191        self.runtime.graceful_shutdown(timeout).await
1192    }
1193}
1194
1195/// Command-channel-driven userspace network stack.
1196///
1197/// This is an opinionated wrapper around [smoltcp](https://docs.rs/smoltcp) that provides an
1198/// easier-to-integrate, more-portable API.
1199pub mod netstack {
1200    #[doc(inline)]
1201    pub use ts_netstack_smoltcp::netcore::Error;
1202    #[doc(inline)]
1203    pub use ts_netstack_smoltcp::netcore::InternalErrorKind;
1204    #[doc(inline)]
1205    pub use ts_netstack_smoltcp::netsock::{TcpListener, TcpStream, UdpSocket};
1206}
1207
1208/// Geneve (RFC 8926) framing for Tailscale **peer-relay** traffic. A peer that advertises
1209/// [`NodeInfo::is_peer_relay`] runs a UDP relay server; relayed disco + WireGuard frames are
1210/// Geneve-encapsulated with a VNI. This module exposes the header codec so the framing is
1211/// recognizable. NOTE: the active relay *data path* (the relay-allocation handshake +
1212/// magicsock integration) is **not yet implemented** in this fork — this is the wire-aware slice.
1213pub mod geneve {
1214    #[doc(inline)]
1215    pub use ts_packet::geneve::{
1216        GENEVE_FIXED_HEADER_LEN, GENEVE_PROTOCOL_DISCO, GENEVE_PROTOCOL_WIREGUARD, GeneveError,
1217        GeneveHeader,
1218    };
1219}
1220
1221/// Tailnet Lock (TKA) verification: the [`tka::Authority`] checks a peer's node-key signature
1222/// against the trusted-key state, mirroring Go's `tka` package. Pair with [`Device::tka_status`]
1223/// (the control-pushed head/disablement signal).
1224pub mod tka {
1225    #[doc(inline)]
1226    pub use ts_tka::{
1227        AumHash, AumKind, Authority, Key, KeyKind, NodeKeySignature, SigKind, State, TkaError,
1228        aum_hash,
1229    };
1230}
1231
1232/// Tailscale cryptographic key types.
1233pub mod keys {
1234    #[doc(inline)]
1235    pub use ts_keys::{
1236        DiscoKeyPair, DiscoPrivateKey, DiscoPublicKey, MachineKeyPair, MachinePrivateKey,
1237        MachinePublicKey, NetworkLockKeyPair, NetworkLockPrivateKey, NetworkLockPublicKey,
1238        NodeKeyPair, NodePrivateKey, NodePublicKey, NodeState, PersistState,
1239    };
1240}
1241
1242const ENV_MAGIC_VAR: &str = "TS_RS_EXPERIMENT";
1243const ENV_MAGIC_VALUE: &str = "this_is_unstable_software";
1244
1245fn check_magic_env() -> Result<(), Error> {
1246    if std::env::var(ENV_MAGIC_VAR).as_deref() != Ok(ENV_MAGIC_VALUE) {
1247        let warning = format!(
1248            "
1249check failed: set {ENV_MAGIC_VAR}={ENV_MAGIC_VALUE} to acknowledge that tailscale-rs is early-days
1250experimental software containing bugs, unvalidated cryptography, and no stability or compatibility
1251guarantees.
1252            "
1253        );
1254
1255        eprintln!("{}", warning.trim());
1256
1257        return Err(Error::UnstableEnvVar);
1258    };
1259
1260    Ok(())
1261}