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// Re-exported so a downstream crate depending only on `tailscale` can name the auth-key secret type
142// for [`Device::new_with_secret`] without taking a separate, version-pinned dependency on `secrecy`
143// (which would risk a `SecretString`-type mismatch if the two `secrecy` majors diverged). Callers
144// pass `tailscale::SecretString`; `secrecy` is a pure-Rust wrapper (no aws-lc/openssl/ring).
145pub use secrecy::SecretString;
146#[doc(inline)]
147pub use ts_control::ExitNodeSelector;
148#[doc(inline)]
149pub use ts_control::Node as NodeInfo;
150#[doc(inline)]
151pub use ts_control::tls::{CertifiedKey, TlsAcceptor, TlsStream};
152#[doc(inline)]
153pub use ts_control::{CertError, MISSING_CERT_RPC, ServeConfig, ServeState, ServeTarget};
154/// The netmap DNS configuration returned by [`Device::dns_config`] (Go `netmap.NetworkMap.DNS`).
155#[doc(inline)]
156pub use ts_control::{DnsConfig, DnsResolver, ExtraRecord};
157#[doc(inline)]
158pub use ts_control::{ExitProxyConfig, ExitProxyScheme};
159pub use ts_control::{
160    IdTokenError, LogoutError, ServiceError, ServiceMode, SshAccept, SshAction, SshConnIdentity,
161    SshDecision, SshDenyReason, SshPolicy, SshPrincipal, SshRule, StableNodeId,
162};
163// Re-exported so the application data-path transport can be selected through the `tailscale`
164// facade alone: `Config::transport_mode` is `TransportMode` (default `Netstack`; `Tun(TunConfig {
165// name, mtu })` for a real kernel TUN interface). Both are `pub` in `ts_control` but were not
166// reachable through this facade, forcing downstream crates to depend on `ts_control` directly just
167// to name them.
168pub use ts_control::{TransportMode, TunConfig};
169#[doc(inline)]
170pub use ts_netstack_smoltcp::PingError;
171use ts_netstack_smoltcp::{CreateSocket, netcore::Channel};
172#[doc(inline)]
173pub use ts_runtime::fallback_tcp::{
174    FallbackConnFuture, FallbackConnHandler, FallbackDecision, FallbackTcpHandle,
175};
176#[doc(inline)]
177pub use ts_runtime::taildrop::WaitingFile;
178#[doc(inline)]
179pub use ts_runtime::{DeviceState, FileTarget, RegistrationError, Status, StatusNode, WhoIs};
180
181#[cfg(feature = "axum")]
182pub mod axum;
183pub mod config;
184mod dial;
185mod error;
186mod loopback;
187#[cfg(feature = "ssh")]
188pub mod ssh;
189
190#[doc(inline)]
191pub use dial::{ConnectedUdpSocket, DialConn};
192#[doc(inline)]
193pub use loopback::LoopbackHandle;
194
195/// How a program connects to a tailnet and communicates with peers.
196///
197/// The `Device` connects to the control plane, registers itself with the tailnet, and communicates
198/// with tailnet peers. Its tailnet identity is determined by the key state provided at
199/// construction-time.
200pub struct Device {
201    runtime: ts_runtime::Runtime,
202    /// Command channel to the application netstack. `None` in TUN transport mode, where there is
203    /// no userspace application netstack; the channel-driven socket APIs ([`Device::udp_bind`],
204    /// [`Device::tcp_listen`], [`Device::tcp_connect`], [`Device::ping`]) are unsupported there.
205    channel: Option<Channel>,
206    /// Whether IPv6 is enabled on the tailnet overlay (the `Config::enable_ipv6` gate, default
207    /// `false`). Captured at construction; used by [`Device::listen_service`] to decide whether an
208    /// IPv6 VIP-service address is bindable (the netstack only accepts IPv6 overlay addresses when
209    /// this is set).
210    enable_ipv6: bool,
211    /// The stored Serve config + its live per-port accept loops (`tsnet`'s `Get/SetServeConfig` +
212    /// serving runtime). Built lazily on the first [`Device::set_serve_config`] (it needs this
213    /// node's overlay IPv4, only known after registration). Held here so its accept loops abort when
214    /// the `Device` drops; `None` (empty config) until the first `set`.
215    serve: std::sync::Mutex<Option<ts_runtime::serve::ServeManager>>,
216    /// The live Funnel ingress manager (`tsnet`'s `ListenFunnel` data path), built on
217    /// [`Device::listen_funnel`](crate::Device::listen_funnel). Held here so its TLS-termination pump and the installed peerAPI
218    /// ingress sink stay alive for the device's life (and tear down when a new `listen_funnel`
219    /// replaces it, or the `Device` drops). `None` until the first `listen_funnel`.
220    funnel: std::sync::Mutex<Option<ts_runtime::funnel::FunnelManager>>,
221}
222
223/// Map a [`ts_runtime::taildrop::TaildropError`] to the device-facing [`Error`]. `Error` is a
224/// `Copy` enum with no payload, so the I/O detail string is dropped, but the *kind* is preserved so
225/// a caller can still distinguish the actionable cases: an invalid name →
226/// [`InternalErrorKind::BadRequest`], an in-progress conflict → [`InternalErrorKind::AlreadyExists`],
227/// a missing file → [`InternalErrorKind::NotFound`], and any other filesystem failure →
228/// [`InternalErrorKind::Io`].
229fn taildrop_err(e: ts_runtime::taildrop::TaildropError) -> Error {
230    use ts_runtime::taildrop::TaildropError;
231    match e {
232        TaildropError::InvalidFileName => Error::Internal(InternalErrorKind::BadRequest),
233        TaildropError::FileExists => Error::Internal(InternalErrorKind::AlreadyExists),
234        TaildropError::Io(io) if io.kind() == std::io::ErrorKind::NotFound => {
235            Error::Internal(InternalErrorKind::NotFound)
236        }
237        TaildropError::Io(_) => Error::Internal(InternalErrorKind::Io),
238    }
239}
240
241/// Map a [`ts_runtime::taildrop_send::TaildropSendError`] (the Taildrop *sender*) to the
242/// device-facing [`Error`]. The send-side conflict/forbidden/unexpected-status cases all reduce to
243/// `BadRequest` (the peer refused the transfer for a request-level reason), a dial failure or
244/// timeout to `Timeout`, an invalid name to `BadRequest`, and any stream I/O failure to `Io`.
245fn taildrop_send_err(e: ts_runtime::taildrop_send::TaildropSendError) -> Error {
246    use ts_runtime::taildrop_send::TaildropSendError;
247    match e {
248        TaildropSendError::Connect | TaildropSendError::Timeout => Error::Timeout,
249        TaildropSendError::InvalidName
250        | TaildropSendError::Forbidden
251        | TaildropSendError::Conflict
252        | TaildropSendError::UnexpectedStatus(_) => Error::Internal(InternalErrorKind::BadRequest),
253        TaildropSendError::Io => Error::Internal(InternalErrorKind::Io),
254    }
255}
256
257/// Resolve the effective registration auth key from `auth_key` plus the config's
258/// workload-identity-federation (WIF) / OAuth-client fields.
259///
260/// With the `identity-federation` feature enabled, an OAuth client secret (`tskey-client-…`) or a
261/// `client_id` + (`id_token` | `audience`) is exchanged for a Tailscale auth key against the SaaS
262/// admin API before registration (Go `tsnet.Server`'s `resolveAuthKey`). Without the feature this is
263/// a pure pass-through: `auth_key` is returned unchanged and the WIF config fields are ignored, so
264/// the default build is byte-identical to before.
265#[cfg(feature = "identity-federation")]
266async fn resolve_auth_key(
267    config: &Config,
268    auth_key: Option<String>,
269) -> Result<Option<String>, Error> {
270    let wif = ts_control::WifConfig {
271        auth_key,
272        client_id: config.client_id.clone(),
273        client_secret: config.client_secret.clone(),
274        id_token: config.id_token.clone(),
275        audience: config.audience.clone(),
276        tags: config.requested_tags.clone(),
277    };
278    ts_control::resolve_auth_key(&wif, &config.control_server_url)
279        .await
280        .map_err(|e| {
281            tracing::error!(error = %e, "resolving auth key via workload-identity federation");
282            Error::Internal(InternalErrorKind::BadRequest)
283        })
284}
285
286/// Pass-through when the `identity-federation` feature is disabled: the auth key is used as-is and
287/// the WIF config fields have no effect (matching Go, where the federation path is compiled out
288/// unless its optional feature is linked).
289#[cfg(not(feature = "identity-federation"))]
290async fn resolve_auth_key(
291    _config: &Config,
292    auth_key: Option<String>,
293) -> Result<Option<String>, Error> {
294    Ok(auth_key)
295}
296
297impl Device {
298    /// Create a device from the given [`Config`] and auth key.
299    ///
300    /// Internally, this will spawn multiple asynchronous actors onto a Tokio runtime.
301    ///
302    /// # Example
303    ///
304    /// ```rust,no_run
305    /// # #[tokio::main]
306    /// # async fn main() -> Result<(), Box<dyn std::error::Error>> {
307    /// # use tailscale::*;
308    /// let dev = Device::new(
309    ///     &Config::default_with_key_file("tsrs_keys.json").await?,
310    ///     Some("MY_AUTH_KEY".to_string()),
311    /// ).await?;
312    /// # Ok(()) }
313    /// ```
314    pub async fn new(config: &Config, auth_key: Option<String>) -> Result<Self, Error> {
315        check_magic_env()?;
316
317        // Resolve the effective registration auth key. The explicit `auth_key` argument wins; if it
318        // is `None`, fall back to `config.auth_key` (Go `tsnet.Server.AuthKey`). When the
319        // `identity-federation` feature is enabled, the resolved key is further passed through the
320        // WIF / OAuth-client bootstrap, which exchanges an OAuth client secret (`tskey-client-…`) or
321        // an IdP-issued OIDC token for a Tailscale auth key before registration (SaaS-only).
322        let auth_key = auth_key.or_else(|| config.auth_key.clone());
323        let auth_key = resolve_auth_key(config, auth_key).await?;
324
325        let rt =
326            ts_runtime::Runtime::spawn(config.into(), auth_key, (&config.key_state).into()).await?;
327        // In TUN transport mode there is no application netstack, so the runtime has no command
328        // channel: that surfaces as `UnsupportedInTunMode`, which we map to a `None` channel rather
329        // than an error (the device is still usable for control-plane and peer-lookup APIs).
330        let channel = match rt.channel().await {
331            Ok(c) => Some(c),
332            Err(e) if e.kind == ts_runtime::ErrorKind::UnsupportedInTunMode => None,
333            Err(e) => return Err(e.into()),
334        };
335
336        Ok(Self {
337            runtime: rt,
338            channel,
339            enable_ipv6: config.enable_ipv6,
340            serve: std::sync::Mutex::new(None),
341            funnel: std::sync::Mutex::new(None),
342        })
343    }
344
345    /// Create a device from the given [`Config`] and a [`SecretString`] auth key.
346    ///
347    /// This is a back-compat-preserving convenience over [`new`](Self::new) for callers that already
348    /// hold the registration auth key as a [`secrecy::SecretString`] (e.g. a daemon that keeps the
349    /// pre-auth key wrapped end-to-end). It lets the caller avoid materializing a plain `String` at
350    /// the engine boundary: the secret is exposed only on the last inch, immediately before being
351    /// handed to [`new`](Self::new).
352    ///
353    /// # Honesty about the plaintext window
354    ///
355    /// This closes the *caller's* boundary, **not** the engine's internal handling. The engine still
356    /// resolves the auth key to a plain `String` internally for registration (the plaintext `String`
357    /// window inside the engine is identical to calling [`new`](Self::new) directly) — this method
358    /// does not make the engine itself secret-clean. If you call [`new`](Self::new) you create that
359    /// `String` yourself; if you call this you do not, but the engine creates one either way.
360    ///
361    /// Passing `None` is equivalent to `new(config, None)` (falls back to `config.auth_key`).
362    ///
363    /// # Example
364    ///
365    /// ```rust,no_run
366    /// # #[tokio::main]
367    /// # async fn main() -> Result<(), Box<dyn std::error::Error>> {
368    /// # use tailscale::*;
369    /// let dev = Device::new_with_secret(
370    ///     &Config::default_with_key_file("tsrs_keys.json").await?,
371    ///     Some(SecretString::from("MY_AUTH_KEY")),
372    /// ).await?;
373    /// # Ok(()) }
374    /// ```
375    pub async fn new_with_secret(
376        config: &Config,
377        auth_key: Option<SecretString>,
378    ) -> Result<Self, Error> {
379        use secrecy::ExposeSecret as _;
380
381        // Expose the secret on the last inch and delegate to `new`, so the spawn/registration path
382        // is shared verbatim (no duplicated runtime-spawn logic) and the engine-internal plaintext
383        // window is byte-for-byte identical to a direct `new` call.
384        let plain = auth_key.map(|s| s.expose_secret().to_string());
385        Self::new(config, plain).await
386    }
387
388    /// The application netstack command channel, or an error in TUN transport mode (no application
389    /// netstack exists).
390    fn channel(&self) -> Result<&Channel, Error> {
391        self.channel
392            .as_ref()
393            .ok_or(Error::Internal(InternalErrorKind::UnsupportedInTunMode))
394    }
395
396    /// Get this [`Device`]'s IPv4 tailnet address.
397    pub async fn ipv4_addr(&self) -> Result<Ipv4Addr, Error> {
398        self.runtime
399            .control
400            .ask(ts_runtime::control_runner::Ipv4)
401            .await
402            .map_err(ts_runtime::Error::from)?
403            .ok_or(Error::Internal(InternalErrorKind::Actor))
404    }
405
406    /// Get this [`Device`]'s IPv6 tailnet address.
407    pub async fn ipv6_addr(&self) -> Result<Ipv6Addr, Error> {
408        self.runtime
409            .control
410            .ask(ts_runtime::control_runner::Ipv6)
411            .await
412            .map_err(ts_runtime::Error::from)?
413            .ok_or(Error::Internal(InternalErrorKind::Actor))
414    }
415
416    /// This node's tailnet IPv4 and (when provisioned) IPv6 addresses as a pair — the Rust analog of
417    /// Go `tsnet.Server.TailscaleIPs() (ip4, ip6 netip.Addr)`.
418    ///
419    /// Reads the self node's assigned addresses (the same source Go splits by family). The tailnet
420    /// is IPv4-only unless [`Config::enable_ipv6`](crate::config::Config) is set, so the IPv6 half is
421    /// `None` when no v6 address is assigned — the Rust shape for Go returning the zero `netip.Addr`
422    /// in that case (Go's IPv6-absent sentinel). Errors until the first netmap is received (no self
423    /// node yet), matching Go returning invalid addresses before the node has joined.
424    pub async fn tailscale_ips(&self) -> Result<(Ipv4Addr, Option<Ipv6Addr>), Error> {
425        let me = self.self_node().await?;
426        let v4 = me.tailnet_address.ipv4.addr();
427        let v6 = me.tailnet_address.ipv6.addr();
428        // The decoder synthesizes the unspecified `::` placeholder on an IPv4-only tailnet; surface
429        // a real v6 only when IPv6 is enabled AND a non-placeholder address was assigned.
430        let v6 = (self.enable_ipv6 && !v6.is_unspecified()).then_some(v6);
431        Ok((v4, v6))
432    }
433
434    /// Bind a UDP socket to the specified [`SocketAddr`].
435    ///
436    /// Returns an error in TUN transport mode (there is no application netstack to bind on).
437    pub async fn udp_bind(&self, socket_addr: SocketAddr) -> Result<netstack::UdpSocket, Error> {
438        self.channel()?
439            .udp_bind(socket_addr)
440            .await
441            .map_err(Into::into)
442    }
443
444    /// Bind a TCP listener to the specified [`SocketAddr`].
445    ///
446    /// Returns an error in TUN transport mode (there is no application netstack to listen on).
447    pub async fn tcp_listen(
448        &self,
449        socket_addr: SocketAddr,
450    ) -> Result<netstack::TcpListener, Error> {
451        self.channel()?
452            .tcp_listen(socket_addr)
453            .await
454            .map_err(Into::into)
455    }
456
457    /// Register a fallback TCP handler (like `tsnet`'s `RegisterFallbackTCPHandler`).
458    ///
459    /// The callback is consulted for every inbound TCP flow that matches **no** explicit
460    /// [`Device::tcp_listen`] listener, with the flow's `(src, dst)` addresses. It returns
461    /// `(handler, intercept)`:
462    /// - `(_, false)` — decline; the next registered callback is tried.
463    /// - `(Some(h), true)` — claim the flow; `h` is handed the accepted [`netstack::TcpStream`].
464    /// - `(None, true)` — claim and reject the flow (the connection is closed).
465    ///
466    /// Multiple handlers may be registered; they are consulted in registration order and the first
467    /// to intercept wins. The returned [`FallbackTcpHandle`] deregisters the handler when dropped.
468    ///
469    /// Handlers serve flows over the overlay netstack only — never a host socket — and a flow no
470    /// handler claims is closed (fail-closed), never direct-dialed.
471    ///
472    /// Returns an error in TUN transport mode (there is no application netstack to attach to).
473    pub fn register_fallback_tcp_handler<F>(&self, cb: F) -> Result<FallbackTcpHandle, Error>
474    where
475        F: Fn(SocketAddr, SocketAddr) -> FallbackDecision + Send + Sync + 'static,
476    {
477        self.runtime
478            .register_fallback_tcp_handler(std::sync::Arc::new(cb))
479            .map_err(Into::into)
480    }
481
482    /// Resolve a tailnet peer (or this node) by MagicDNS name to its tailnet IPv4 address.
483    ///
484    /// This is an in-process lookup against the netmap we already hold — like `tsnet`'s in-memory
485    /// `dnsMap`, it does not query any DNS server (there is no `100.100.100.100` resolver). The
486    /// `name` may be a bare hostname or a fully-qualified MagicDNS name, with or without a trailing
487    /// dot, in any case (matching is case-insensitive). Returns `Ok(None)` if no tailnet node has
488    /// that name.
489    ///
490    /// Only MagicDNS names are resolved; names outside the tailnet are not looked up here, so the
491    /// caller's system resolver remains responsible for them. IPv6 is intentionally not resolved —
492    /// this fork operates IPv4-only on the tailnet.
493    pub async fn resolve(&self, name: &str) -> Result<Option<Ipv4Addr>, Error> {
494        if let Some(peer) = self.peer_by_name(name).await? {
495            return Ok(Some(peer.tailnet_address.ipv4.addr()));
496        }
497
498        // tsnet's dnsMap also resolves our own name; fall back to self when no peer matches.
499        let me = self.self_node().await?;
500        if me.matches_name(name) {
501            return Ok(Some(me.tailnet_address.ipv4.addr()));
502        }
503
504        Ok(None)
505    }
506
507    /// Connect to a tailnet peer by MagicDNS name and port over TCP.
508    ///
509    /// Resolves `name` via [`Device::resolve`] (an in-process netmap lookup, no DNS server), then
510    /// dials the resulting tailnet IPv4 address. Returns [`InternalErrorKind::BadRequest`] if the
511    /// name does not resolve to a tailnet node.
512    pub async fn connect_by_name(
513        &self,
514        name: &str,
515        port: u16,
516    ) -> Result<netstack::TcpStream, Error> {
517        let addr = self
518            .resolve(name)
519            .await?
520            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
521
522        self.tcp_connect((addr, port).into()).await
523    }
524
525    /// Resolve a `host:port` string to a tailnet [`SocketAddr`], honoring the family forced by a
526    /// `network` suffix. The host may be an IP literal (parsed directly) or a MagicDNS name
527    /// (resolved via [`Device::resolve`], which yields a tailnet IPv4). Shared by [`Device::dial`]
528    /// and [`Device::dial_tcp`]. The IPv4-only invariant is enforced here: a `…6` network, or any v6
529    /// destination, requires `Config::enable_ipv6` and otherwise returns
530    /// [`InternalErrorKind::BadRequest`] (a clean typed error rather than a downstream actor error).
531    async fn resolve_dial_addr(
532        &self,
533        network: dial::Network,
534        addr: &str,
535    ) -> Result<SocketAddr, Error> {
536        let (host, port) = dial::split_host_port(addr)?;
537
538        // An IP literal is used directly; otherwise resolve the MagicDNS name (IPv4 only).
539        let ip: IpAddr = if let Ok(ip) = host.parse::<IpAddr>() {
540            ip
541        } else {
542            self.resolve(host)
543                .await?
544                .ok_or(Error::Internal(InternalErrorKind::BadRequest))?
545                .into()
546        };
547
548        dial::check_family(network.family, ip)?;
549
550        // IPv4-only invariant: a v6 destination is only reachable when IPv6 is provisioned.
551        if ip.is_ipv6() && !self.enable_ipv6 {
552            return Err(Error::Internal(InternalErrorKind::BadRequest));
553        }
554
555        Ok((ip, port).into())
556    }
557
558    /// Connect to a tailnet address over TCP or UDP, the Rust analog of Go
559    /// `tsnet.Server.Dial(ctx, network, address)`.
560    ///
561    /// `network` is one of `"tcp"`, `"tcp4"`, `"tcp6"`, `"udp"`, `"udp4"`, `"udp6"`; `addr` is a
562    /// `host:port` string where `host` is a MagicDNS name, an IPv4 literal, or a bracketed IPv6
563    /// literal (`[2001:db8::1]:443`). The host is resolved in-process via [`Device::resolve`] (no DNS
564    /// server). Returns a [`DialConn`] whose arm matches the transport — use [`Device::dial_tcp`]
565    /// when you want the TCP stream directly.
566    ///
567    /// Differences from Go (documented for parity): ports must be **numeric** (Go's `LookupPort`
568    /// also resolves named ports like `"http"`; this fork avoids a services-file dependency), and
569    /// `…6`/v6 destinations require `Config::enable_ipv6` (the tailnet is IPv4-only by default).
570    ///
571    /// # Errors
572    /// [`InternalErrorKind::BadRequest`] for an unsupported `network`, a malformed/portless `addr`,
573    /// an unresolvable name, or a v6 destination while IPv6 is disabled; otherwise the transport's
574    /// own connect error.
575    pub async fn dial(&self, network: &str, addr: &str) -> Result<DialConn, Error> {
576        let net = dial::parse_network(network)?;
577        let remote = self.resolve_dial_addr(net, addr).await?;
578
579        match net.transport {
580            dial::Transport::Tcp => Ok(DialConn::Tcp(self.tcp_connect(remote).await?)),
581            dial::Transport::Udp => {
582                // Bind an ephemeral local UDP socket on this node's tailnet address of the SAME
583                // family as the remote, then connect it (Go's `Dial("udp", …)` returns a connected
584                // UDP `net.Conn`, with the local source picked by `IfElse(dst.Is6(), v6, v4)`). A v4
585                // local socket cannot send to a v6 peer, so the family must match `remote`. (TCP gets
586                // this for free: `tcp_connect` already picks the source family from `remote`.)
587                let local_ip: IpAddr = if remote.is_ipv6() {
588                    self.ipv6_addr().await?.into()
589                } else {
590                    self.ipv4_addr().await?.into()
591                };
592                let sock = self.udp_bind((local_ip, 0).into()).await?;
593                Ok(DialConn::Udp(ConnectedUdpSocket::new(sock, remote)))
594            }
595        }
596    }
597
598    /// Connect to a tailnet address over TCP, returning the stream directly — the common case of
599    /// [`Device::dial`] for `"tcp"`. `addr` is a `host:port` string (MagicDNS name or IP literal).
600    /// This is the building block for HTTP-over-tailnet: an embedder's `hyper`/`reqwest` client can
601    /// route requests by calling `dial_tcp(&format!("{host}:{port}"))` from its connector, mirroring
602    /// how Go `tsnet.Server.HTTPClient` sets `http.Transport.DialContext = Server.Dial`.
603    ///
604    /// # Errors
605    /// As [`Device::dial`] for the `"tcp"` network.
606    pub async fn dial_tcp(&self, addr: &str) -> Result<netstack::TcpStream, Error> {
607        let remote = self
608            .resolve_dial_addr(
609                dial::Network {
610                    transport: dial::Transport::Tcp,
611                    family: dial::Family::Any,
612                },
613                addr,
614            )
615            .await?;
616        self.tcp_connect(remote).await
617    }
618
619    /// Bind a UDP socket from a `host:port` string, the Rust analog of Go
620    /// `tsnet.Server.ListenPacket(network, addr)`.
621    ///
622    /// `network` is one of `"udp"`, `"udp4"`, `"udp6"`; `addr` must be a **valid IP literal**
623    /// `host:port` (Go's `ListenPacket` rejects a name or empty host — unlike `Listen`). An
624    /// unspecified host (`0.0.0.0`/`[::]`) binds on this node's tailnet address. Returns the
625    /// unconnected [`netstack::UdpSocket`] (a `net.PacketConn`).
626    ///
627    /// # Errors
628    /// [`InternalErrorKind::BadRequest`] for a non-UDP/unsupported `network`, a malformed addr, a
629    /// non-IP host, a family mismatch, or a v6 bind while IPv6 is disabled.
630    pub async fn listen_packet(
631        &self,
632        network: &str,
633        addr: &str,
634    ) -> Result<netstack::UdpSocket, Error> {
635        let net = dial::parse_network(network)?;
636        if net.transport != dial::Transport::Udp {
637            return Err(Error::Internal(InternalErrorKind::BadRequest));
638        }
639        let (host, port) = dial::split_host_port(addr)?;
640
641        // ListenPacket requires a valid IP host (Go rejects a name here).
642        let ip: IpAddr = host
643            .parse()
644            .map_err(|_| Error::Internal(InternalErrorKind::BadRequest))?;
645        dial::check_family(net.family, ip)?;
646
647        // A v6 bind (whether an explicit literal or an unspecified `[::]`) requires IPv6 to be
648        // provisioned — enforce the gate for BOTH cases (the unspecified `[::]` path used to skip it).
649        if ip.is_ipv6() && !self.enable_ipv6 {
650            return Err(Error::Internal(InternalErrorKind::BadRequest));
651        }
652
653        // An unspecified bind host (`0.0.0.0` / `[::]`) means "this node's tailnet address" — of the
654        // SAME family as the requested address, so a `udp6` `[::]:0` binds a v6 socket (it used to
655        // fall through to the v4 address regardless, silently yielding an IPv4 socket for a v6 listen).
656        let bind_ip: IpAddr = if ip.is_unspecified() {
657            if ip.is_ipv6() {
658                self.ipv6_addr().await?.into()
659            } else {
660                self.ipv4_addr().await?.into()
661            }
662        } else {
663            ip
664        };
665
666        self.udp_bind((bind_ip, port).into()).await
667    }
668
669    /// Connect to a TCP socket at the remote address.
670    ///
671    /// Returns an error in TUN transport mode (there is no application netstack to dial from).
672    pub async fn tcp_connect(&self, remote: SocketAddr) -> Result<netstack::TcpStream, Error> {
673        let channel = self.channel()?;
674
675        let ip: IpAddr = match remote.is_ipv4() {
676            true => self.ipv4_addr().await?.into(),
677            false => self.ipv6_addr().await?.into(),
678        };
679
680        // TODO(npry): collision checking
681        let ephemeral_port = rand::random_range(49152..=u16::MAX);
682
683        channel
684            .tcp_connect((ip, ephemeral_port).into(), remote)
685            .await
686            .map_err(Into::into)
687    }
688
689    /// Start a SOCKS5 proxy on a host loopback address that dials into the tailnet (Go
690    /// `tsnet.Server.Loopback`, SOCKS5 half).
691    ///
692    /// Binds a TCP listener on `127.0.0.1:0` (host loopback only — never an external interface) and
693    /// serves SOCKS5 (RFC 1928) with required username/password auth (RFC 1929): username `tsnet`,
694    /// password = the returned `proxy_cred`. Each `CONNECT` is dialed INTO the overlay via
695    /// [`Device::connect_by_name`] / [`Device::tcp_connect`] and spliced to the accepted host socket, so
696    /// a non-Rust host process can reach tailnet peers through the proxy. Returns the bound address, the
697    /// proxy credential, and a [`LoopbackHandle`] whose drop stops the listener.
698    ///
699    /// Anti-leak: the listener is loopback-only and every connection egresses over the overlay, never a
700    /// host socket — the host's real origin IP is never used to reach the destination. Unlike Go, the
701    /// LocalAPI HTTP surface is not served (this fork exposes status/whois/id-token natively on
702    /// `Device`); only the SOCKS5 proxy is provided.
703    ///
704    /// Returns an error in TUN transport mode (no application netstack to dial from).
705    pub async fn loopback(&self) -> Result<(std::net::SocketAddr, String, LoopbackHandle), Error> {
706        // Capture only cloneable pieces — never `&self` — for the spawned accept loop: a clone of the
707        // netstack command channel, this device's own overlay IPv4 (fetched once), and a boxed
708        // resolver closure over clones of the control + peer-tracker actor refs. The resolver
709        // replicates `Device::resolve` (peer-by-name, falling back to this node's own name).
710        let channel = self.channel()?.clone();
711        let self_ipv4 = self.ipv4_addr().await?;
712
713        let control = self.runtime.control.clone();
714        let peer_tracker = self.runtime.peer_tracker.clone();
715        let resolve: loopback::Resolver = std::sync::Arc::new(move |name: String| {
716            let control = control.clone();
717            let peer_tracker = peer_tracker.clone();
718            Box::pin(async move {
719                let pt = peer_tracker
720                    .upgrade()
721                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
722                let peer = pt
723                    .ask(ts_runtime::peer_tracker::PeerByName { name: name.clone() })
724                    .await
725                    .map_err(ts_runtime::Error::from)?;
726                if let Some(peer) = peer {
727                    return Ok(Some(peer.tailnet_address.ipv4.addr()));
728                }
729                // tsnet's dnsMap also resolves our own name; fall back to self.
730                let me = control
731                    .ask(ts_runtime::control_runner::SelfNode)
732                    .await
733                    .map_err(ts_runtime::Error::from)?
734                    .ok_or(Error::Internal(InternalErrorKind::Actor))?;
735                if me.matches_name(&name) {
736                    Ok(Some(me.tailnet_address.ipv4.addr()))
737                } else {
738                    Ok(None)
739                }
740            }) as std::pin::Pin<Box<dyn std::future::Future<Output = _> + Send>>
741        });
742
743        let dialer = loopback::OverlayDialer::new(channel, self_ipv4, resolve);
744        loopback::start(dialer).await
745    }
746
747    /// Get our node info.
748    pub async fn self_node(&self) -> Result<NodeInfo, Error> {
749        self.runtime
750            .control
751            .ask(ts_runtime::control_runner::SelfNode)
752            .await
753            .map_err(ts_runtime::Error::from)?
754            .ok_or(Error::Internal(InternalErrorKind::Actor))
755    }
756
757    /// The DNS names this node can obtain TLS certificates for — Go `tsnet.Server.CertDomains()`.
758    ///
759    /// These are the `CertDomains` control pushed in the netmap DNS config: the names a TLS-serving
760    /// consumer (e.g. a `ListenTLS`/`GetCertificate`-style caller) should request a cert for. Returns
761    /// an empty `Vec` before the first netmap, or when control granted none — mirroring Go returning a
762    /// clone of `nm.DNS.CertDomains` (empty/`nil` when absent).
763    pub async fn cert_domains(&self) -> Result<Vec<String>, Error> {
764        self.runtime
765            .control
766            .ask(ts_runtime::control_runner::CertDomains)
767            .await
768            .map_err(ts_runtime::Error::from)
769            .map_err(Into::into)
770    }
771
772    /// The DNS configuration control pushed in the latest netmap — Go `tsnet`'s view of
773    /// `netmap.NetworkMap.DNS` (what `tailscale dns status` reports).
774    ///
775    /// Returns the full [`DnsConfig`] — MagicDNS on/off, search domains, global + fallback resolvers,
776    /// split-DNS routes, extra records, cert domains — or `None` before the first netmap / when
777    /// control has sent no DNS config. A superset of [`cert_domains`](Device::cert_domains), which
778    /// remains a separate narrower accessor for the TLS-cert use. Mirrors Go reading a clone of
779    /// `nm.DNS` (absent ⇒ `None`).
780    pub async fn dns_config(&self) -> Result<Option<DnsConfig>, Error> {
781        self.runtime
782            .control
783            .ask(ts_runtime::control_runner::DnsConfig)
784            .await
785            .map_err(ts_runtime::Error::from)
786            .map_err(Into::into)
787    }
788
789    /// This node's key-expiry instant as Unix seconds (`Node.KeyExpiry` in Go), or `Ok(None)` if
790    /// the key never expires.
791    ///
792    /// Like Go, this fork is **reactive** about key expiry — it reports it rather than rotating the
793    /// node key in the background. A caller can schedule re-authentication around this time; on
794    /// expiry, re-create the [`Device`] (which re-registers), supplying a fresh node key + the prior
795    /// `old_node_key` to rotate, or the same key to refresh.
796    pub async fn self_key_expiry_unix(&self) -> Result<Option<i64>, Error> {
797        Ok(self.self_node().await?.key_expiry_unix())
798    }
799
800    /// Whether this node's key has expired as of now (`!KeyExpiry.IsZero() && KeyExpiry.Before(now)`
801    /// in Go). A key with no expiry is never expired. See [`Device::self_key_expiry_unix`] for the
802    /// reactive-rotation note.
803    pub async fn self_key_expired(&self) -> Result<bool, Error> {
804        let now = std::time::SystemTime::now()
805            .duration_since(std::time::UNIX_EPOCH)
806            .map(|d| d.as_secs() as i64)
807            // An unreadable clock (pre-epoch) is treated as the far future so a time-limited key
808            // looks expired — fail-safe toward prompting re-auth rather than trusting a stale key.
809            .unwrap_or(i64::MAX);
810        Ok(self.self_node().await?.key_expired_at_unix(now))
811    }
812
813    /// Fetch the current Tailscale SSH policy pushed by control, if any.
814    ///
815    /// Returns `Ok(None)` when control has not sent an SSH policy. The SSH server treats an absent
816    /// or empty policy as **deny-all** (fail-closed). Used by the SSH auth path
817    /// ([`SshPolicy::evaluate`][ts_control::SshPolicy::evaluate]) to authorize incoming
818    /// connections.
819    pub async fn ssh_policy(&self) -> Result<Option<ts_control::SshPolicy>, Error> {
820        self.runtime
821            .control
822            .ask(ts_runtime::control_runner::CurrentSshPolicy)
823            .await
824            .map_err(ts_runtime::Error::from)
825            .map_err(Into::into)
826    }
827
828    /// Look up a peer by name.
829    pub async fn peer_by_name(&self, name: &str) -> Result<Option<NodeInfo>, Error> {
830        let pt = self
831            .runtime
832            .peer_tracker
833            .upgrade()
834            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
835
836        pt.ask(ts_runtime::peer_tracker::PeerByName {
837            name: name.to_string(),
838        })
839        .await
840        .map_err(ts_runtime::Error::from)
841        .map_err(Into::into)
842    }
843
844    /// Look up a peer by ip.
845    pub async fn peer_by_tailnet_ip(&self, ip: IpAddr) -> Result<Option<NodeInfo>, Error> {
846        let pt = self
847            .runtime
848            .peer_tracker
849            .upgrade()
850            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
851
852        pt.ask(ts_runtime::peer_tracker::PeerByTailnetIp { ip })
853            .await
854            .map_err(ts_runtime::Error::from)
855            .map_err(Into::into)
856    }
857
858    /// Look up the peer(s) with the most-specific route matches for `ip`.
859    ///
860    /// This reports which peers *advertise* a route covering `ip`, independent of this device's
861    /// `accept_routes` setting — analogous to the Go client's informational `PrimaryRoutes`. It is
862    /// not a reachability oracle: with `accept_routes` off, the dataplane will not actually route
863    /// to (or accept return traffic from) advertised subnet routes even if this returns a peer.
864    pub async fn peers_with_route(&self, ip: IpAddr) -> Result<Vec<NodeInfo>, Error> {
865        let pt = self
866            .runtime
867            .peer_tracker
868            .upgrade()
869            .ok_or(Error::Internal(InternalErrorKind::Actor))?;
870
871        pt.ask(ts_runtime::peer_tracker::PeerByAcceptedRoute { ip })
872            .await
873            .map_err(ts_runtime::Error::from)
874            .map_err(Into::into)
875    }
876
877    /// List the Taildrop files this device has fully received and not yet consumed (Go LocalAPI
878    /// `WaitingFiles`).
879    ///
880    /// Returns the files waiting under the configured `taildrop_dir`, sorted by name. Returns an
881    /// empty list when Taildrop is disabled (`Config::taildrop_dir` unset) — fail-closed, never an
882    /// error for the disabled case. A filesystem error while listing surfaces as
883    /// [`InternalErrorKind::Actor`].
884    pub fn taildrop_waiting_files(&self) -> Result<Vec<WaitingFile>, Error> {
885        let Some(store) = self.runtime.taildrop_store() else {
886            return Ok(Vec::new());
887        };
888        store
889            .waiting_files()
890            .map_err(|_| Error::Internal(InternalErrorKind::Actor))
891    }
892
893    /// Open a received Taildrop file by name for reading, returning the handle and its size (Go
894    /// LocalAPI `OpenFile`).
895    ///
896    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
897    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
898    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
899    pub fn taildrop_open_file(&self, name: &str) -> Result<(std::fs::File, u64), Error> {
900        let store = self
901            .runtime
902            .taildrop_store()
903            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
904        store.open_file(name).map_err(taildrop_err)
905    }
906
907    /// Delete a received Taildrop file by name (Go LocalAPI `DeleteFile`).
908    ///
909    /// The `name` is validated (path-traversal-safe) inside the store before any path is built.
910    /// Returns [`InternalErrorKind::BadRequest`] when Taildrop is disabled or the name is invalid,
911    /// and [`InternalErrorKind::Actor`] for a filesystem error (e.g. the file does not exist).
912    pub fn taildrop_delete_file(&self, name: &str) -> Result<(), Error> {
913        let store = self
914            .runtime
915            .taildrop_store()
916            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
917        store.delete_file(name).map_err(taildrop_err)
918    }
919
920    /// Send a local file to a tailnet `peer` via Taildrop (Go `PushFile` / `tailscale file cp`).
921    ///
922    /// Pushes `content_length` bytes from `reader` to the peer's peerAPI as
923    /// `PUT /v0/put/<name>` over the overlay netstack — the sending counterpart to the receive store
924    /// surfaced by [`Device::taildrop_waiting_files`]. The transfer rides the encrypted WireGuard
925    /// overlay, never a host socket. The body is streamed from offset 0 (no resume).
926    ///
927    /// The destination is derived **solely from `peer`'s own node record**
928    /// ([`NodeInfo::peerapi_addr`][ts_control::Node::peerapi_addr]): its advertised tailnet IPv4 and
929    /// `peerapi4` port. The caller obtains `peer` from [`Device::peer_by_name`] /
930    /// [`Device::peer_by_tailnet_ip`], so it is always a current netmap peer — a raw control-supplied
931    /// or attacker-chosen address can never be targeted. As defense in depth, the resolved address is
932    /// additionally asserted to be a Tailscale CGNAT IP before dialing.
933    ///
934    /// Returns [`InternalErrorKind::BadRequest`] when the peer advertises no IPv4 peerAPI (so it
935    /// cannot receive files), when the name is invalid, or when the peer refuses the transfer
936    /// (`403`/`409`/unexpected status); [`Error::Timeout`] on a dial failure or timeout; and
937    /// [`InternalErrorKind::Io`] on a mid-transfer stream error.
938    pub async fn send_file<R>(
939        &self,
940        peer: &NodeInfo,
941        name: &str,
942        content_length: u64,
943        reader: R,
944    ) -> Result<(), Error>
945    where
946        R: tokio::io::AsyncRead + Unpin,
947    {
948        let channel = self.channel()?;
949
950        // Destination comes only from the peer's own node record — never an arbitrary address.
951        let dst = peer
952            .peerapi_addr()
953            .ok_or(Error::Internal(InternalErrorKind::BadRequest))?;
954        // Defense in depth: refuse to dial anything outside the Tailscale CGNAT range, so a
955        // malformed node record can't steer the PUT at a non-tailnet host.
956        if !ts_control::is_tailscale_ip(dst.ip()) {
957            return Err(Error::Internal(InternalErrorKind::BadRequest));
958        }
959
960        let self_ipv4 = self.ipv4_addr().await?;
961
962        ts_runtime::taildrop_send::send_file(channel, self_ipv4, dst, name, content_length, reader)
963            .await
964            .map_err(taildrop_send_err)
965    }
966
967    /// List the tailnet peers this node can Taildrop a file *to* — the Rust analog of Go's LocalAPI
968    /// `FileTargets`.
969    ///
970    /// Each [`FileTarget`] pairs a peer's node record with the `http://ip:port` base of its peerAPI;
971    /// pass `target.node` straight to [`Device::send_file`]. A peer qualifies when it advertises a
972    /// reachable IPv4 peerAPI **and** is either owned by the same user as this node **or** explicitly
973    /// granted the file-sharing-target capability — mirroring upstream's send-path filter. The list is
974    /// gated on this node holding the file-sharing capability (control grants it when the admin
975    /// enables Taildrop); absent that, the result is empty (fail-closed, not an error). Sorted by the
976    /// peer's MagicDNS name. Targets are listed regardless of online state (matching upstream — an
977    /// offline target's [`send_file`](Device::send_file) simply times out). Empty before the first
978    /// netmap.
979    pub async fn file_targets(&self) -> Result<Vec<FileTarget>, Error> {
980        self.runtime.file_targets().await.map_err(Into::into)
981    }
982
983    /// Begin a debug packet capture, streaming a pcap of every packet crossing the dataplane to
984    /// `writer` (Go `tsnet.Server.CapturePcap`).
985    ///
986    /// Installs a capture hook on the running dataplane: from now until [`Device::stop_capture`] is
987    /// called (or another capture replaces this one), a copy of every plaintext IP packet on the
988    /// datapath — outbound (pre-encrypt) and inbound (post-decrypt) — is framed and written to
989    /// `writer`. The 24-byte pcap global header is written immediately on success.
990    ///
991    /// The format is byte-faithful classic pcap with Tailscale's `LINKTYPE_USER0` + 4-byte path
992    /// preamble per record (see [`ts_runtime::capture`]); a resulting file opens in Wireshark, and
993    /// with Tailscale's `ts-dissector.lua` the direction/path of each packet decodes.
994    ///
995    /// The hook runs **inline on the single-threaded dataplane step**, so `writer` must not block for
996    /// long — a slow writer back-pressures the datapath. Records are **not** flushed per packet (that
997    /// would be a syscall on every packet on the dataplane thread); buffered bytes are flushed when
998    /// the writer is dropped on [`Device::stop_capture`]. Wrap `writer` in a [`std::io::BufWriter`] if
999    /// you want buffering. A write error is swallowed per-packet (the capture silently drops that
1000    /// record) rather than tearing down the datapath; call [`Device::stop_capture`] to end it. Returns
1001    /// an error only if the dataplane actor is unreachable or the initial global-header write fails.
1002    pub async fn capture_pcap<W>(&self, writer: W) -> Result<(), Error>
1003    where
1004        W: std::io::Write + Send + 'static,
1005    {
1006        let sink = std::sync::Arc::new(std::sync::Mutex::new(
1007            ts_runtime::capture::PcapSink::new(writer)
1008                .map_err(|_| Error::Internal(InternalErrorKind::Io))?,
1009        ));
1010        let hook: ts_runtime::CaptureHook = std::sync::Arc::new(move |path, pkt: &[u8]| {
1011            if let Ok(mut sink) = sink.lock() {
1012                // A per-packet write failure (e.g. a closed pipe) silently drops that record rather
1013                // than tearing down the datapath; the caller ends capture via `stop_capture`.
1014                drop(sink.log_packet(path.code(), pkt));
1015            }
1016        });
1017        self.runtime.install_capture(Some(hook)).await?;
1018        Ok(())
1019    }
1020
1021    /// Stop a debug packet capture started by [`Device::capture_pcap`] (Go `ClearCaptureSink`).
1022    ///
1023    /// Clears the dataplane capture hook; the writer is dropped (its remaining buffered bytes are
1024    /// flushed by its own `Drop`). Idempotent — clearing when no capture is installed is a no-op.
1025    /// Returns an error only if the dataplane actor is unreachable.
1026    pub async fn stop_capture(&self) -> Result<(), Error> {
1027        self.runtime.install_capture(None).await?;
1028        Ok(())
1029    }
1030
1031    /// Snapshot of this device and its tailnet peers (like `tailscale status`).
1032    ///
1033    /// Combines this node's self info with the current peer set: each [`StatusNode`] reports the
1034    /// stable id, display name, tailnet IPs, advertised routes, and exit-node flag. (Per-peer
1035    /// `online`/user/capabilities are honestly `None`/empty in this fork — the domain node model
1036    /// does not yet carry the wire-level liveness/login fields; see `ts_runtime::status` docs.)
1037    pub async fn status(&self) -> Result<Status, Error> {
1038        self.runtime.status().await.map_err(Into::into)
1039    }
1040
1041    /// Fetch the current Tailnet Lock (TKA) status pushed by control, if any.
1042    ///
1043    /// Returns `Ok(None)` when control has sent no `TKAInfo` (tailnet lock not in use, or no change
1044    /// observed yet). The returned [`TkaStatus`][ts_control::TkaStatus] carries the authority head
1045    /// (a base32 `AUMHash`, decode with [`tka::AumHash::from_base32`][ts_tka::AumHash::from_base32])
1046    /// and the disablement signal. Signature verification of a peer's node-key signature against the
1047    /// authority is performed with the [`tka`] module's [`tka::Authority`][ts_tka::Authority].
1048    pub async fn tka_status(&self) -> Result<Option<ts_control::TkaStatus>, Error> {
1049        self.runtime
1050            .control
1051            .ask(ts_runtime::control_runner::CurrentTkaStatus)
1052            .await
1053            .map_err(ts_runtime::Error::from)
1054            .map_err(Into::into)
1055    }
1056
1057    /// Request an OIDC **ID token** from control for this node, scoped to `audience` (workload-
1058    /// identity federation, like `tailscale`'s `id-token` LocalAPI).
1059    ///
1060    /// Returns a signed JWT whose `sub` claim is this node's MagicDNS name and whose `aud` claim is
1061    /// `audience`, suitable for presenting to a third-party relying party (e.g. AWS/GCP
1062    /// workload-identity federation). The node is the token *subject*, not the authenticator — this
1063    /// is token issuance over the Noise transport (`POST /machine/id-token`), not a login path.
1064    /// Requires the control plane to support capability version ≥ 30.
1065    pub async fn fetch_id_token(&self, audience: &str) -> Result<String, ts_control::IdTokenError> {
1066        self.runtime.fetch_id_token(audience.to_string()).await
1067    }
1068
1069    /// Log this node out of the tailnet — deregister it from the control plane (the equivalent of
1070    /// Go `tsnet`'s `LocalClient.Logout`).
1071    ///
1072    /// Re-`POST`s `/machine/register` with this node's current node key and a past expiry, which the
1073    /// control plane honors by **expiring the node now**: it drops out of every peer's netmap and
1074    /// must re-register (re-authenticate) to rejoin.
1075    ///
1076    /// This is primarily for **non-ephemeral** nodes. An ephemeral node is garbage-collected by
1077    /// control shortly after it disconnects, but a persistent node lingers in the tailnet
1078    /// (visible to peers, counting against the machine limit) for up to ~24h after the process exits
1079    /// unless explicitly logged out. Call this before [`shutdown`](Self::shutdown) to deregister
1080    /// immediately. Calling it on an ephemeral node simply brings the GC forward; it is idempotent,
1081    /// so logging out an already-gone node is not an error.
1082    ///
1083    /// This is a **control-plane state change only**: it does not tear down the local datapath (do
1084    /// that via [`shutdown`](Self::shutdown)), and it does not delete or rotate the on-disk node key
1085    /// — re-registering with the same key (a fresh [`Device::new`]) is the re-login path.
1086    pub async fn logout(&self) -> Result<(), ts_control::LogoutError> {
1087        self.runtime.logout().await
1088    }
1089
1090    /// Snapshot this node's client metrics in Prometheus text exposition format.
1091    ///
1092    /// Mirrors Go Tailscale's `clientmetric` registry: process-global counters/gauges incremented
1093    /// on the datapath hot loops (e.g. `magicsock_send_udp`, `magicsock_recv_data_bytes_udp`),
1094    /// rendered as `# TYPE <name> <kind>\n<name> <value>\n` per metric, sorted by name. (Go `tsnet`
1095    /// exposes no metrics method of its own, so this is the fork's clean public surface.) The
1096    /// registry is process-global, so the output covers every `Device` in the process.
1097    pub fn metrics(&self) -> String {
1098        ts_metrics::write_prometheus()
1099    }
1100
1101    /// Map a tailnet source `addr` to the node that owns its IP (like `tsnet`'s `WhoIs`).
1102    ///
1103    /// Only the IP of `addr` is used; the port is ignored. Returns `Ok(None)` if no tailnet node
1104    /// owns that address.
1105    pub async fn whois(&self, addr: SocketAddr) -> Result<Option<WhoIs>, Error> {
1106        self.runtime.whois(addr).await.map_err(Into::into)
1107    }
1108
1109    /// Change the selected exit node at runtime, without recreating the [`Device`] — the equivalent
1110    /// of Go `tsnet`'s `LocalClient.EditPrefs(ExitNodeID/ExitNodeIP)`.
1111    ///
1112    /// The peer may be named by stable node ID, tailnet IP, or MagicDNS name via
1113    /// [`ExitNodeSelector`] (a bare IP or name parses with `selector.parse()`); this is the same
1114    /// selector type as [`Config::exit_node`](crate::Config::exit_node), so the construction-time
1115    /// and runtime paths are identical. Passing `None` clears the exit node — internet-bound traffic
1116    /// is then dropped (fail-closed) unless this node egresses directly.
1117    ///
1118    /// The change is applied immediately: the new selector is re-resolved against the live peer set
1119    /// and the outbound route + inbound source filter are recomputed at once. A selector for a peer
1120    /// not yet in the netmap simply takes effect once that peer appears.
1121    ///
1122    /// Only NEW flows use the changed exit; in-flight connections are not torn down and continue
1123    /// egressing via the previously-selected exit until they close.
1124    pub async fn set_exit_node(&self, exit_node: Option<ExitNodeSelector>) -> Result<(), Error> {
1125        self.runtime
1126            .set_exit_node(exit_node)
1127            .await
1128            .map_err(Into::into)
1129    }
1130
1131    /// The currently-selected exit node, or `None` if none is selected.
1132    pub fn exit_node(&self) -> Option<ExitNodeSelector> {
1133        self.runtime.exit_node()
1134    }
1135
1136    /// Re-bind the underlay UDP socket after a **network/link change** — Wi-Fi switch, sleep/wake,
1137    /// or any event that invalidates the device's local address/NAT mapping. This is the Rust
1138    /// analog of Go magicsock's `Conn.Rebind()`.
1139    ///
1140    /// The embedder owns deciding *when* to call this (it watches the OS for link changes — there is
1141    /// no built-in network monitor); `rebind` is the engine half that does the socket work:
1142    /// - Re-binds the underlay UDP socket, preferring the same local port (so the advertised
1143    ///   endpoint stays stable) and falling back to an ephemeral port. The IPv4-only-by-default
1144    ///   invariant is preserved.
1145    /// - Invalidates the now-stale local mapping: learned reflexive (STUN) addresses and every
1146    ///   peer's *confirmed* direct path are cleared, while candidate endpoints are kept — so peers
1147    ///   are re-probed over the new socket and **relay over DERP (never a direct host dial) until a
1148    ///   path re-confirms**. Endpoint discovery re-runs on its normal cadence.
1149    /// - Leaves peers, control, the netmap, disco keys, and DERP connections untouched; existing
1150    ///   WireGuard sessions survive (they ride whatever underlay carries them).
1151    ///
1152    /// A no-op if the underlay socket failed to bind at startup (the device is DERP-only). Existing
1153    /// connectivity is preserved on a re-bind error (the old socket is kept; the error is returned).
1154    pub async fn rebind(&self) -> Result<(), Error> {
1155        self.runtime.rebind().await.map_err(Into::into)
1156    }
1157
1158    /// The stable id of the exit node traffic is **currently** egressing through, or `None` if none
1159    /// is engaged (the equivalent of Go `tsnet`'s `Status.ExitNodeStatus.ID`).
1160    ///
1161    /// This differs from [`exit_node`](Self::exit_node), which returns the *configured* selector:
1162    /// the active exit node is the route updater's resolved, fail-closed answer. It is `None` when
1163    /// no exit node is configured, the configured selector matches no current peer, or the matched
1164    /// peer no longer advertises a default route (egress is then dropped, fail-closed). Match the id
1165    /// against [`Status::peers`](crate::Status::peers) (via [`status`](Self::status)) for details.
1166    pub fn active_exit_node(&self) -> Option<ts_control::StableNodeId> {
1167        self.runtime.active_exit_node()
1168    }
1169
1170    /// Watch for netmap changes: the returned receiver's value is the current set of peer
1171    /// [`StatusNode`]s and updates on every netmap change (like subscribing to `ipn` notifications).
1172    pub async fn watch_netmap(
1173        &self,
1174    ) -> Result<tokio::sync::watch::Receiver<Vec<StatusNode>>, Error> {
1175        self.runtime.watch_netmap().await.map_err(Into::into)
1176    }
1177
1178    /// The current device connection-[`DeviceState`] (`Connecting` / `Running` / `NeedsLogin` /
1179    /// `Expired` / `Failed`).
1180    pub fn device_state(&self) -> DeviceState {
1181        self.runtime.device_state()
1182    }
1183
1184    /// Watch the device connection-[`DeviceState`], reacting push-style to control connection
1185    /// transitions instead of polling [`status`](Self::status).
1186    ///
1187    /// Returns a [`tokio::sync::watch::Receiver`]; await its
1188    /// [`changed`](tokio::sync::watch::Receiver::changed) to be woken on each transition. The
1189    /// initial value is the current state.
1190    pub fn watch_state(&self) -> tokio::sync::watch::Receiver<DeviceState> {
1191        self.runtime.watch_state()
1192    }
1193
1194    /// Wait until the device finishes registering, returning a typed outcome — the clean
1195    /// replacement for polling [`ipv4_addr`](Self::ipv4_addr) in a loop.
1196    ///
1197    /// Resolves `Ok(())` once the device is [`DeviceState::Running`]. On a non-running outcome it
1198    /// returns a typed [`RegistrationError`]:
1199    /// - [`AuthRejected`](RegistrationError::AuthRejected) — bad/expired/unknown auth key;
1200    ///   **permanent** (re-pair).
1201    /// - [`NeedsLogin`](RegistrationError::NeedsLogin) — interactive authorization required;
1202    ///   **not permanent** (the runtime keeps retrying and reaches `Running` once the user
1203    ///   authorizes). Auth-key callers treat this as failure; interactive callers should ignore it
1204    ///   and drive the flow via [`watch_state`](Self::watch_state).
1205    /// - [`NetworkUnreachable`](RegistrationError::NetworkUnreachable) — **transient** (retry).
1206    /// - [`Timeout`](RegistrationError::Timeout) — no settled state within `timeout` (`None` waits
1207    ///   indefinitely).
1208    ///
1209    /// [`KeyExpired`](RegistrationError::KeyExpired) is not produced here (a key expires only after
1210    /// the node is up); observe it via [`watch_state`](Self::watch_state). Use
1211    /// [`RegistrationError::is_permanent`] to branch "re-pair" vs. "retry / drive login".
1212    pub async fn wait_until_running(
1213        &self,
1214        timeout: Option<Duration>,
1215    ) -> Result<(), RegistrationError> {
1216        self.runtime.wait_until_running(timeout).await
1217    }
1218
1219    /// Ping a tailnet peer over the overlay with an ICMPv4 echo, returning the round-trip time
1220    /// (like `tailscale ping`).
1221    ///
1222    /// The echo is sent from this device's own tailnet IPv4 over the overlay netstack — never a
1223    /// host socket. IPv6 destinations return [`PingError::Ipv6Unsupported`] (this fork is
1224    /// IPv4-only on the tailnet). A peer answers from its own OS stack; this netstack does not
1225    /// auto-reply to echo requests.
1226    ///
1227    /// In TUN transport mode there is no application netstack to ping from; this surfaces as
1228    /// [`PingError::Timeout`] (the same error this method already uses for an unavailable source
1229    /// address — `PingError` carries no dedicated "unsupported" variant).
1230    pub async fn ping(&self, dst: IpAddr, timeout: Duration) -> Result<Duration, PingError> {
1231        let channel = self.channel().map_err(|_| PingError::Timeout)?;
1232        let src = self.ipv4_addr().await.map_err(|_| PingError::Timeout)?;
1233        ts_netstack_smoltcp::ping(channel, src, dst, timeout).await
1234    }
1235
1236    /// Obtain a TLS certificate for a node's MagicDNS `name` (like `tsnet`'s `GetCertificate`).
1237    ///
1238    /// **Fail-closed without the `acme` feature.** By default this fork has no client-side ACME
1239    /// engine wired in, so this returns [`ts_control::CertError::Unimplemented`] (after a
1240    /// tailnet-name check) — it NEVER self-signs and NEVER returns a placeholder certificate
1241    /// ([`ts_control::MISSING_CERT_RPC`] names what is missing).
1242    ///
1243    /// **With the `acme` feature** this instead drives the client-side ACME DNS-01 engine to issue a
1244    /// real Let's Encrypt certificate for `name`, publishing the challenge TXT via the node's
1245    /// `POST /machine/set-dns` RPC (routed through the control runner). SaaS-only: a self-hosted
1246    /// control plane may 501 on set-dns, surfaced as [`ts_control::CertError::Acme`].
1247    #[cfg(not(feature = "acme"))]
1248    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
1249        ts_control::get_certificate(name).await
1250    }
1251
1252    /// See the no-`acme` variant for the contract; with `acme` this issues a real cert via the
1253    /// runtime's ACME engine (`Device → Runtime → ControlRunner → issue_certificate_via_setdns`).
1254    #[cfg(feature = "acme")]
1255    pub async fn get_certificate(&self, name: &str) -> Result<CertifiedKey, ts_control::CertError> {
1256        self.runtime.get_certificate(name.to_string()).await
1257    }
1258
1259    /// Build a [`TlsAcceptor`] terminating TLS for `cfg.name` on the overlay (like `tsnet`'s
1260    /// `ListenTLS`).
1261    ///
1262    /// Obtains the certificate via [`Device::get_certificate`] — so with the `acme` feature this
1263    /// issues a real Let's Encrypt cert (when the control plane answers `set-dns`), and without it
1264    /// (or when issuance is unavailable) it surfaces the same fail-closed
1265    /// [`ts_control::CertError`] rather than ever serving a self-signed cert or downgrading to
1266    /// plaintext. Terminate accepted overlay streams with [`ts_control::accept_tls`].
1267    pub async fn listen_tls(
1268        &self,
1269        cfg: &ts_control::ServeConfig,
1270    ) -> Result<TlsAcceptor, ts_control::CertError> {
1271        // Route through Device::get_certificate (the acme-aware issuance path) rather than
1272        // ts_control::listen_tls, which only knows the non-acme stub. Validate the serve config
1273        // first (same fail-closed checks ts_control::listen_tls applies), then assemble the acceptor.
1274        cfg.validate()?;
1275        let cert = self.get_certificate(&cfg.name).await?;
1276        ts_control::tls_acceptor(cert)
1277    }
1278
1279    /// The currently-stored Serve config (like `tsnet`'s `GetServeConfig`).
1280    ///
1281    /// Returns the config last passed to [`Device::set_serve_config`], or an empty
1282    /// [`ts_control::ServeState`] (no ports) if none was ever set. Pure read — does not touch the
1283    /// network.
1284    pub fn get_serve_config(&self) -> ts_control::ServeState {
1285        match &*self.serve.lock().unwrap_or_else(|e| e.into_inner()) {
1286            Some(mgr) => mgr.get(),
1287            None => ts_control::ServeState::default(),
1288        }
1289    }
1290
1291    /// Replace this node's Serve config and (re)bind its tailnet ports (like `tsnet`'s
1292    /// `SetServeConfig`, REPLACE semantics).
1293    ///
1294    /// `state` becomes the **whole** config (full-replace reconcile: every previously-bound serve
1295    /// port's accept loop is torn down and the new config's ports are bound from scratch). For each
1296    /// configured port the manager binds an overlay listener on this node's tailnet IPv4 and
1297    /// dispatches per [`ts_control::ServeTarget`]:
1298    /// - [`Accept`](ts_control::ServeTarget::Accept) — the TLS-terminated stream is handed back over
1299    ///   the returned [`ServeAcceptedReceiver`](ts_runtime::serve::ServeAcceptedReceiver) (the
1300    ///   in-process stand-in for `ListenTLS`'s `net.Listener`).
1301    /// - [`Proxy`](ts_control::ServeTarget::Proxy) — reverse-proxy the decrypted stream to a local
1302    ///   host backend.
1303    /// - [`Text`](ts_control::ServeTarget::Text) — write a fixed body and close.
1304    /// - [`TcpForward`](ts_control::ServeTarget::TcpForward) — forward the **raw** (non-TLS) stream
1305    ///   to a local host backend.
1306    ///
1307    /// **Fail-closed.** `state.validate()` runs first. Every TLS-terminating port's acceptor is
1308    /// obtained up-front via [`Device::listen_tls`] (the ACME-aware cert path); if any cert cannot be
1309    /// issued the whole call fails with that [`ts_control::CertError`] and **nothing is bound** — a
1310    /// TLS port never downgrades to plaintext.
1311    ///
1312    /// **Anti-leak.** Listeners bind the overlay netstack only (never a host socket). The
1313    /// `Proxy`/`TcpForward` backend dial is a local host socket to the embedder's own backend (like
1314    /// Go's reverse-proxy to `127.0.0.1`), intentionally NOT routed through the exit-egress
1315    /// forwarder. A backend dial failure drops that connection; it never falls back.
1316    ///
1317    /// Returns an error in TUN transport mode (there is no application netstack to bind on). The
1318    /// previous config's accept loops (and any earlier `ServeAcceptedReceiver`) stop when this
1319    /// returns; the new receiver delivers every `Accept`-port connection.
1320    pub async fn set_serve_config(
1321        &self,
1322        state: ts_control::ServeState,
1323    ) -> Result<ts_runtime::serve::ServeAcceptedReceiver, Error> {
1324        state
1325            .validate()
1326            .map_err(|_| Error::Internal(InternalErrorKind::BadRequest))?;
1327
1328        // Fail-closed: build every TLS-terminating port's acceptor up-front via the ACME-aware cert
1329        // path. If any cert can't be issued, return before binding anything (no plaintext downgrade).
1330        let mut resolved = std::collections::BTreeMap::new();
1331        for (port, target) in &state.ports {
1332            let acceptor = if target.terminates_tls() {
1333                let cfg = ts_control::ServeConfig {
1334                    name: state.name.clone(),
1335                    port: *port,
1336                    target: target.clone(),
1337                };
1338                Some(self.listen_tls(&cfg).await.map_err(|_| {
1339                    // Cert issuance is fail-closed in this fork; surface as a request error rather
1340                    // than ever binding a plaintext TLS port.
1341                    Error::Internal(InternalErrorKind::BadRequest)
1342                })?)
1343            } else {
1344                None
1345            };
1346            resolved.insert(
1347                *port,
1348                ts_runtime::serve::ResolvedPort {
1349                    target: target.clone(),
1350                    acceptor,
1351                },
1352            );
1353        }
1354
1355        // The manager binds the OVERLAY netstack on this node's own tailnet IPv4.
1356        let self_ipv4 = self.ipv4_addr().await?;
1357        let channel = self.channel()?.clone();
1358
1359        let mut slot = self.serve.lock().unwrap_or_else(|e| e.into_inner());
1360        let mgr =
1361            slot.get_or_insert_with(|| ts_runtime::serve::ServeManager::new(channel, self_ipv4));
1362        Ok(mgr.set(state, resolved))
1363    }
1364
1365    /// Expose a tailnet TLS service to the public internet via Tailscale Funnel (like `tsnet`'s
1366    /// `ListenFunnel`), returning a [`FunnelAcceptedReceiver`](ts_runtime::funnel::FunnelAcceptedReceiver)
1367    /// that delivers each TLS-terminated public connection.
1368    ///
1369    /// **Two fail-closed gates, then the live ingress listener.** First the node-attribute gate is
1370    /// fully enforced from this node's own capability map (mirroring Go `ipn.NodeCanFunnel` +
1371    /// `ipn.CheckFunnelPort`): the tailnet admin must have enabled HTTPS and granted the `funnel`
1372    /// node attribute, and `cfg.port` must be in the set the `funnel-ports` capability allows —
1373    /// otherwise this returns [`ts_control::FunnelError::NotAllowed`] /
1374    /// [`ts_control::FunnelError::PortNotAllowed`] before touching any cert or network. Then the
1375    /// node's `*.ts.net` certificate is obtained via the ACME-aware [`Device::get_certificate`] (the
1376    /// Funnel hostname *is* the node's MagicDNS name, so its DNS-01 cert matches); fail-closed on
1377    /// [`ts_control::FunnelError::Cert`] — no self-signed or plaintext fallback.
1378    ///
1379    /// On success a [`FunnelManager`](ts_runtime::funnel::FunnelManager) is registered: its ingress
1380    /// sink is installed into the runtime's peerAPI `/v0/ingress` slot (making that route live without
1381    /// restarting the peerAPI server), and the `HostInfo.IngressEnabled` map-request signal is set so
1382    /// control routes Funnel traffic to this node. Public Funnel bytes arrive as a relay POST to
1383    /// `/v0/ingress`, are membership-gated + `101`-hijacked into a raw stream, TLS-terminated by the
1384    /// manager, and delivered over the returned receiver.
1385    ///
1386    /// **Where the relay comes from.** The public ingress **relay + DNS mapping** that feed
1387    /// `/v0/ingress` are Tailscale infrastructure ([`ts_control::MISSING_FUNNEL_RELAY`]), provisioned
1388    /// automatically against real Tailscale SaaS with a Funnel-enabled ACL; against a self-hosted
1389    /// control plane no relay exists, so the listener is correct but never fed.
1390    ///
1391    /// Anti-leak: Funnel TLS terminates only on the overlay netstack (the hijacked ingress stream
1392    /// arrives on the overlay peerAPI listener), never a host socket; there is no self-signed or
1393    /// plaintext fallback. A new `listen_funnel` replaces the previous manager (its pump + sink tear
1394    /// down); dropping the `Device` tears it down too.
1395    pub async fn listen_funnel(
1396        &self,
1397        cfg: &ts_control::ServeConfig,
1398        opts: ts_control::FunnelOptions,
1399    ) -> Result<ts_runtime::funnel::FunnelAcceptedReceiver, ts_control::FunnelError> {
1400        // Gate 1 (fail-closed, no network): node-attribute + funnel-port access from our cap map.
1401        let me = self
1402            .self_node()
1403            .await
1404            .map_err(|_| ts_control::FunnelError::NotAllowed)?;
1405        cfg.validate()?;
1406        ts_control::funnel_access(&me, cfg.port)?;
1407
1408        // Gate 2 (fail-closed): obtain the node's `*.ts.net` cert via the ACME-aware path and build
1409        // the TLS acceptor. A cert failure surfaces as FunnelError::Cert — never a plaintext listener.
1410        let cert = self
1411            .get_certificate(&cfg.name)
1412            .await
1413            .map_err(ts_control::FunnelError::Cert)?;
1414        let acceptor = ts_control::tls_acceptor(cert).map_err(ts_control::FunnelError::Cert)?;
1415
1416        // `opts.funnel_only` (reject tailnet-internal connections) is accepted for surface stability;
1417        // the ingress data path only ever carries relay-delivered public traffic, so there is no
1418        // tailnet-internal leg on this listener to reject. Documented as a no-op here for now.
1419        let _ = opts;
1420
1421        // Build the funnel manager + its ingress sink + the hand-back receiver, install the sink into
1422        // the runtime's shared peerAPI `/v0/ingress` slot (making the route live), and flip the
1423        // IngressEnabled map signal. Hold the manager on the device so its pump/sink live as long as
1424        // the listener; replacing a prior manager tears the old one down on drop at end of scope.
1425        let (manager, sink, receiver) = ts_runtime::funnel::FunnelManager::new(acceptor);
1426        {
1427            let slot = self.runtime.funnel_ingress_slot();
1428            *slot.lock().unwrap_or_else(|e| e.into_inner()) = Some(sink);
1429        }
1430        self.runtime
1431            .ingress_active_flag()
1432            .store(true, std::sync::atomic::Ordering::Relaxed);
1433
1434        let old = {
1435            let mut held = self.funnel.lock().unwrap_or_else(|e| e.into_inner());
1436            held.replace(manager)
1437        };
1438        drop(old);
1439
1440        Ok(receiver)
1441    }
1442
1443    /// Host a Tailscale **VIP service** (`svc:<label>`) by binding an overlay listener on the
1444    /// service's control-assigned virtual IP (like `tsnet`'s `ListenService`).
1445    ///
1446    /// **Fail-closed.** Mirrors Go `tsnet.Server.ListenService`'s preconditions, enforced from this
1447    /// node's own netmap state ([`ts_control::resolve_service_listen`]): the `name` must be a valid
1448    /// `svc:<dns-label>`, this node must be **tagged** (Go `ErrUntaggedServiceHost`), and control
1449    /// must have assigned the service a VIP address on this node (delivered via the `service-host`
1450    /// node-capability — see [`ts_control::Node::service_addresses`]). Any unmet precondition
1451    /// returns a typed [`ts_control::ServiceError`] before binding anything.
1452    ///
1453    /// When all hold, this binds a [`tcp_listen`][Device::tcp_listen] on the service VIP and the
1454    /// configured `mode` port over the **overlay netstack** (never a host socket) and returns the
1455    /// listener. The netstack already accepts packets for control-assigned VIPs (they are injected
1456    /// alongside the node's own tailnet address), so the listener is reachable by tailnet peers.
1457    ///
1458    /// The `Tun`/L3 service mode is unsupported (a TODO in upstream tsnet); only TCP/HTTP modes
1459    /// (which bind the same VIP:port at the listen layer) are offered. Returns an error in TUN
1460    /// transport mode (there is no application netstack to bind on).
1461    pub async fn listen_service(
1462        &self,
1463        name: &str,
1464        mode: ts_control::ServiceMode,
1465    ) -> Result<netstack::TcpListener, ts_control::ServiceError> {
1466        let me = self
1467            .self_node()
1468            .await
1469            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))?;
1470        let listen_addr = ts_control::resolve_service_listen(&me, name, mode, self.enable_ipv6)?;
1471        self.tcp_listen(listen_addr)
1472            .await
1473            .map_err(|e| ts_control::ServiceError::Listen(e.to_string()))
1474    }
1475
1476    /// Attempt to gracefully shut down this device's runtime.
1477    ///
1478    /// Reports whether the device was fully shut down before the timeout. It is still shut
1479    /// down if it timed out, just more violently and with potential resource leaks.
1480    ///
1481    /// If `timeout` is `None`, then shutdown will never time-out.
1482    pub async fn shutdown(self, timeout: Option<Duration>) -> bool {
1483        self.runtime.graceful_shutdown(timeout).await
1484    }
1485}
1486
1487/// Command-channel-driven userspace network stack.
1488///
1489/// This is an opinionated wrapper around [smoltcp](https://docs.rs/smoltcp) that provides an
1490/// easier-to-integrate, more-portable API.
1491pub mod netstack {
1492    #[doc(inline)]
1493    pub use ts_netstack_smoltcp::netcore::Error;
1494    #[doc(inline)]
1495    pub use ts_netstack_smoltcp::netcore::InternalErrorKind;
1496    #[doc(inline)]
1497    pub use ts_netstack_smoltcp::netsock::{TcpListener, TcpStream, UdpSocket};
1498}
1499
1500/// Geneve (RFC 8926) framing for Tailscale **peer-relay** traffic. A peer that advertises
1501/// [`NodeInfo::is_peer_relay`] runs a UDP relay server; relayed disco + WireGuard frames are
1502/// Geneve-encapsulated with a VNI. This module exposes the header codec so the framing is
1503/// recognizable. NOTE: the active relay *data path* (the relay-allocation handshake +
1504/// magicsock integration) is **not yet implemented** in this fork — this is the wire-aware slice.
1505pub mod geneve {
1506    #[doc(inline)]
1507    pub use ts_packet::geneve::{
1508        GENEVE_FIXED_HEADER_LEN, GENEVE_PROTOCOL_DISCO, GENEVE_PROTOCOL_WIREGUARD, GeneveError,
1509        GeneveHeader,
1510    };
1511}
1512
1513/// Tailnet Lock (TKA) verification: the [`tka::Authority`] checks a peer's node-key signature
1514/// against the trusted-key state, mirroring Go's `tka` package. Pair with [`Device::tka_status`]
1515/// (the control-pushed head/disablement signal).
1516pub mod tka {
1517    #[doc(inline)]
1518    pub use ts_tka::{
1519        AumHash, AumKind, Authority, Key, KeyKind, NodeKeySignature, SigKind, State, TkaError,
1520        aum_hash,
1521    };
1522}
1523
1524/// Tailscale cryptographic key types.
1525pub mod keys {
1526    #[doc(inline)]
1527    pub use ts_keys::{
1528        DiscoKeyPair, DiscoPrivateKey, DiscoPublicKey, MachineKeyPair, MachinePrivateKey,
1529        MachinePublicKey, NetworkLockKeyPair, NetworkLockPrivateKey, NetworkLockPublicKey,
1530        NodeKeyPair, NodePrivateKey, NodePublicKey, NodeState, PersistState,
1531    };
1532}
1533
1534const ENV_MAGIC_VAR: &str = "TS_RS_EXPERIMENT";
1535const ENV_MAGIC_VALUE: &str = "this_is_unstable_software";
1536
1537fn check_magic_env() -> Result<(), Error> {
1538    if std::env::var(ENV_MAGIC_VAR).as_deref() != Ok(ENV_MAGIC_VALUE) {
1539        let warning = format!(
1540            "
1541check failed: set {ENV_MAGIC_VAR}={ENV_MAGIC_VALUE} to acknowledge that tailscale-rs is early-days
1542experimental software containing bugs, unvalidated cryptography, and no stability or compatibility
1543guarantees.
1544            "
1545        );
1546
1547        eprintln!("{}", warning.trim());
1548
1549        return Err(Error::UnstableEnvVar);
1550    };
1551
1552    Ok(())
1553}
1554
1555#[cfg(test)]
1556mod tests {
1557    use secrecy::ExposeSecret as _;
1558
1559    use super::*;
1560
1561    // `Device::new`/`new_with_secret` cannot be unit-tested end-to-end without a live control
1562    // server (registration). The only behavioral difference `new_with_secret` introduces over `new`
1563    // is exposing the `SecretString` to a plain `String` on the last inch; everything after is the
1564    // shared `new` path. So we assert that equivalence at the auth-key-resolution level: the secret
1565    // path must resolve to the exact same key the plain path feeds into `resolve_auth_key`.
1566    const SAMPLE_KEY: &str = "tskey-auth-koCgSLP5R811CNTRL-EXAMPLEEXAMPLEEXAMPLEEXAMPLE";
1567
1568    // The mapping `new_with_secret` applies (`Option<SecretString>` -> `Option<String>`) must be a
1569    // byte-for-byte round-trip, so the spawn arg is identical to a direct `new(config, Some(..))`.
1570    #[test]
1571    fn secret_exposes_to_identical_string() {
1572        let plain = Some(SAMPLE_KEY.to_string());
1573        let from_secret =
1574            Some(SecretString::from(SAMPLE_KEY)).map(|s| s.expose_secret().to_string());
1575        assert_eq!(from_secret, plain);
1576
1577        // `None` must pass through unchanged (so it falls back to `config.auth_key` exactly as `new`).
1578        let none_secret: Option<SecretString> = None;
1579        assert_eq!(
1580            none_secret.map(|s| s.expose_secret().to_string()),
1581            None::<String>
1582        );
1583    }
1584
1585    // End-to-end equivalence at the resolve layer: feeding the exposed secret through
1586    // `resolve_auth_key` yields the same `Option<String>` as feeding the plain string — i.e. both
1587    // constructors reach the same spawn argument, without registering against a control server.
1588    #[tokio::test]
1589    async fn new_with_secret_resolves_same_as_new() {
1590        let config = Config::default();
1591
1592        let via_plain = resolve_auth_key(&config, Some(SAMPLE_KEY.to_string()))
1593            .await
1594            .expect("plain auth key resolves");
1595
1596        let exposed = Some(SecretString::from(SAMPLE_KEY)).map(|s| s.expose_secret().to_string());
1597        let via_secret = resolve_auth_key(&config, exposed)
1598            .await
1599            .expect("secret-derived auth key resolves");
1600
1601        assert_eq!(via_plain, via_secret);
1602        // Without the `identity-federation` feature `resolve_auth_key` is a pass-through, so the
1603        // resolved key is the input verbatim; assert that too to pin the default-build behavior.
1604        #[cfg(not(feature = "identity-federation"))]
1605        assert_eq!(via_secret, Some(SAMPLE_KEY.to_string()));
1606    }
1607}