Line data Source code
1 : //! A stand-alone program that routes connections, e.g. from
2 : //! `aaa--bbb--1234.external.domain` to `aaa.bbb.internal.domain:1234`.
3 : //!
4 : //! This allows connecting to pods/services running in the same Kubernetes cluster from
5 : //! the outside. Similar to an ingress controller for HTTPS.
6 :
7 : use std::io;
8 : use std::net::SocketAddr;
9 : use std::path::Path;
10 : use std::sync::Arc;
11 :
12 : use anyhow::{Context, anyhow, bail, ensure};
13 : use clap::Arg;
14 : use futures::future::Either;
15 : use futures::{FutureExt, TryFutureExt};
16 : use itertools::Itertools;
17 : use rustls::crypto::ring;
18 : use rustls::pki_types::{DnsName, PrivateKeyDer};
19 : use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
20 : use tokio::net::TcpListener;
21 : use tokio_rustls::TlsConnector;
22 : use tokio_rustls::server::TlsStream;
23 : use tokio_util::sync::CancellationToken;
24 : use tracing::{Instrument, error, info};
25 : use utils::project_git_version;
26 : use utils::sentry_init::init_sentry;
27 :
28 : use crate::context::RequestContext;
29 : use crate::metrics::{Metrics, ThreadPoolMetrics};
30 : use crate::pglb::TlsRequired;
31 : use crate::pqproto::FeStartupPacket;
32 : use crate::protocol2::ConnectionInfo;
33 : use crate::proxy::{ErrorSource, copy_bidirectional_client_compute};
34 : use crate::stream::{PqStream, Stream};
35 : use crate::util::run_until_cancelled;
36 :
37 : project_git_version!(GIT_VERSION);
38 :
39 0 : fn cli() -> clap::Command {
40 0 : clap::Command::new("Neon proxy/router")
41 0 : .version(GIT_VERSION)
42 0 : .arg(
43 0 : Arg::new("listen")
44 0 : .short('l')
45 0 : .long("listen")
46 0 : .help("listen for incoming client connections on ip:port")
47 0 : .default_value("127.0.0.1:4432"),
48 : )
49 0 : .arg(
50 0 : Arg::new("listen-tls")
51 0 : .long("listen-tls")
52 0 : .help("listen for incoming client connections on ip:port, requiring TLS to compute")
53 0 : .default_value("127.0.0.1:4433"),
54 : )
55 0 : .arg(
56 0 : Arg::new("tls-key")
57 0 : .short('k')
58 0 : .long("tls-key")
59 0 : .help("path to TLS key for client postgres connections")
60 0 : .required(true),
61 : )
62 0 : .arg(
63 0 : Arg::new("tls-cert")
64 0 : .short('c')
65 0 : .long("tls-cert")
66 0 : .help("path to TLS cert for client postgres connections")
67 0 : .required(true),
68 : )
69 0 : .arg(
70 0 : Arg::new("dest")
71 0 : .short('d')
72 0 : .long("destination")
73 0 : .help("append this domain zone to the SNI hostname to get the destination address")
74 0 : .required(true),
75 : )
76 0 : }
77 :
78 0 : pub async fn run() -> anyhow::Result<()> {
79 0 : let _logging_guard = crate::logging::init().await?;
80 0 : let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
81 0 : let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
82 :
83 0 : Metrics::install(Arc::new(ThreadPoolMetrics::new(0)));
84 :
85 0 : let args = cli().get_matches();
86 0 : let destination: String = args
87 0 : .get_one::<String>("dest")
88 0 : .expect("string argument defined")
89 0 : .parse()?;
90 :
91 : // Configure TLS
92 0 : let tls_config = match (
93 0 : args.get_one::<String>("tls-key"),
94 0 : args.get_one::<String>("tls-cert"),
95 : ) {
96 0 : (Some(key_path), Some(cert_path)) => parse_tls(key_path.as_ref(), cert_path.as_ref())?,
97 0 : _ => bail!("tls-key and tls-cert must be specified"),
98 : };
99 :
100 0 : let compute_tls_config =
101 0 : Arc::new(crate::tls::client_config::compute_client_config_with_root_certs()?);
102 :
103 : // Start listening for incoming client connections
104 0 : let proxy_address: SocketAddr = args
105 0 : .get_one::<String>("listen")
106 0 : .expect("listen argument defined")
107 0 : .parse()?;
108 0 : let proxy_address_compute_tls: SocketAddr = args
109 0 : .get_one::<String>("listen-tls")
110 0 : .expect("listen-tls argument defined")
111 0 : .parse()?;
112 :
113 0 : info!("Starting sni router on {proxy_address}");
114 0 : info!("Starting sni router on {proxy_address_compute_tls}");
115 0 : let proxy_listener = TcpListener::bind(proxy_address).await?;
116 0 : let proxy_listener_compute_tls = TcpListener::bind(proxy_address_compute_tls).await?;
117 :
118 0 : let cancellation_token = CancellationToken::new();
119 0 : let dest = Arc::new(destination);
120 :
121 0 : let main = tokio::spawn(task_main(
122 0 : dest.clone(),
123 0 : tls_config.clone(),
124 0 : None,
125 0 : proxy_listener,
126 0 : cancellation_token.clone(),
127 : ))
128 0 : .map(crate::error::flatten_err);
129 :
130 0 : let main_tls = tokio::spawn(task_main(
131 0 : dest,
132 0 : tls_config,
133 0 : Some(compute_tls_config),
134 0 : proxy_listener_compute_tls,
135 0 : cancellation_token.clone(),
136 : ))
137 0 : .map(crate::error::flatten_err);
138 0 : let signals_task = tokio::spawn(crate::signals::handle(cancellation_token, || {}));
139 :
140 : // the signal task cant ever succeed.
141 : // the main task can error, or can succeed on cancellation.
142 : // we want to immediately exit on either of these cases
143 0 : let main = futures::future::try_join(main, main_tls);
144 0 : let signal = match futures::future::select(signals_task, main).await {
145 0 : Either::Left((res, _)) => crate::error::flatten_err(res)?,
146 0 : Either::Right((res, _)) => {
147 0 : res?;
148 0 : return Ok(());
149 : }
150 : };
151 :
152 : // maintenance tasks return `Infallible` success values, this is an impossible value
153 : // so this match statically ensures that there are no possibilities for that value
154 : match signal {}
155 0 : }
156 :
157 0 : pub(super) fn parse_tls(
158 0 : key_path: &Path,
159 0 : cert_path: &Path,
160 0 : ) -> anyhow::Result<Arc<rustls::ServerConfig>> {
161 0 : let key = {
162 0 : let key_bytes = std::fs::read(key_path).context("TLS key file")?;
163 :
164 0 : let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..]).collect_vec();
165 :
166 0 : ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
167 : PrivateKeyDer::Pkcs8(
168 0 : keys.pop()
169 0 : .expect("keys should not be empty")
170 0 : .context(format!(
171 0 : "Failed to read TLS keys at '{}'",
172 0 : key_path.display()
173 0 : ))?,
174 : )
175 : };
176 :
177 0 : let cert_chain_bytes = std::fs::read(cert_path).context(format!(
178 0 : "Failed to read TLS cert file at '{}.'",
179 0 : cert_path.display()
180 0 : ))?;
181 :
182 0 : let cert_chain: Vec<_> = {
183 0 : rustls_pemfile::certs(&mut &cert_chain_bytes[..])
184 0 : .try_collect()
185 0 : .with_context(|| {
186 0 : format!(
187 0 : "Failed to read TLS certificate chain from bytes from file at '{}'.",
188 0 : cert_path.display()
189 : )
190 0 : })?
191 : };
192 :
193 0 : let tls_config =
194 0 : rustls::ServerConfig::builder_with_provider(Arc::new(ring::default_provider()))
195 0 : .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])
196 0 : .context("ring should support TLS1.2 and TLS1.3")?
197 0 : .with_no_client_auth()
198 0 : .with_single_cert(cert_chain, key)?
199 0 : .into();
200 :
201 0 : Ok(tls_config)
202 0 : }
203 :
204 0 : pub(super) async fn task_main(
205 0 : dest_suffix: Arc<String>,
206 0 : tls_config: Arc<rustls::ServerConfig>,
207 0 : compute_tls_config: Option<Arc<rustls::ClientConfig>>,
208 0 : listener: tokio::net::TcpListener,
209 0 : cancellation_token: CancellationToken,
210 0 : ) -> anyhow::Result<()> {
211 : // When set for the server socket, the keepalive setting
212 : // will be inherited by all accepted client sockets.
213 0 : socket2::SockRef::from(&listener).set_keepalive(true)?;
214 :
215 0 : let connections = tokio_util::task::task_tracker::TaskTracker::new();
216 :
217 0 : while let Some(accept_result) =
218 0 : run_until_cancelled(listener.accept(), &cancellation_token).await
219 : {
220 0 : let (socket, peer_addr) = accept_result?;
221 :
222 0 : let session_id = uuid::Uuid::new_v4();
223 0 : let tls_config = Arc::clone(&tls_config);
224 0 : let dest_suffix = Arc::clone(&dest_suffix);
225 0 : let compute_tls_config = compute_tls_config.clone();
226 :
227 0 : connections.spawn(
228 0 : async move {
229 0 : socket
230 0 : .set_nodelay(true)
231 0 : .context("failed to set socket option")?;
232 :
233 0 : let ctx = RequestContext::new(
234 0 : session_id,
235 0 : ConnectionInfo {
236 0 : addr: peer_addr,
237 0 : extra: None,
238 0 : },
239 0 : crate::metrics::Protocol::SniRouter,
240 : );
241 0 : handle_client(ctx, dest_suffix, tls_config, compute_tls_config, socket).await
242 0 : }
243 0 : .unwrap_or_else(|e| {
244 0 : if let Some(FirstMessage(io_error)) = e.downcast_ref() {
245 : // this is noisy. if we get EOF on the very first message that's likely
246 : // just NLB doing a healthcheck.
247 0 : if io_error.kind() == io::ErrorKind::UnexpectedEof {
248 0 : return;
249 0 : }
250 0 : }
251 :
252 : // Acknowledge that the task has finished with an error.
253 0 : error!("per-client task finished with an error: {e:#}");
254 0 : })
255 0 : .instrument(tracing::info_span!("handle_client", ?session_id)),
256 : );
257 : }
258 :
259 0 : connections.close();
260 0 : drop(listener);
261 :
262 0 : connections.wait().await;
263 :
264 0 : info!("all client connections have finished");
265 0 : Ok(())
266 0 : }
267 :
268 : #[derive(Debug, thiserror::Error)]
269 : #[error(transparent)]
270 : struct FirstMessage(io::Error);
271 :
272 0 : async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
273 0 : ctx: &RequestContext,
274 0 : raw_stream: S,
275 0 : tls_config: Arc<rustls::ServerConfig>,
276 0 : ) -> anyhow::Result<TlsStream<S>> {
277 0 : let (mut stream, msg) = PqStream::parse_startup(Stream::from_raw(raw_stream))
278 0 : .await
279 0 : .map_err(FirstMessage)?;
280 :
281 0 : match msg {
282 : FeStartupPacket::SslRequest { direct: None } => {
283 0 : let raw = stream.accept_tls().await?;
284 :
285 0 : Ok(raw
286 0 : .upgrade(tls_config, !ctx.has_private_peer_addr())
287 0 : .await?)
288 : }
289 0 : unexpected => {
290 0 : info!(
291 : ?unexpected,
292 0 : "unexpected startup packet, rejecting connection"
293 : );
294 0 : Err(stream.throw_error(TlsRequired, None).await)?
295 : }
296 : }
297 0 : }
298 :
299 0 : async fn handle_client(
300 0 : ctx: RequestContext,
301 0 : dest_suffix: Arc<String>,
302 0 : tls_config: Arc<rustls::ServerConfig>,
303 0 : compute_tls_config: Option<Arc<rustls::ClientConfig>>,
304 0 : stream: impl AsyncRead + AsyncWrite + Unpin,
305 0 : ) -> anyhow::Result<()> {
306 0 : let mut tls_stream = ssl_handshake(&ctx, stream, tls_config).await?;
307 :
308 : // Cut off first part of the SNI domain
309 : // We receive required destination details in the format of
310 : // `{k8s_service_name}--{k8s_namespace}--{port}.non-sni-domain`
311 0 : let sni = tls_stream
312 0 : .get_ref()
313 0 : .1
314 0 : .server_name()
315 0 : .ok_or(anyhow!("SNI missing"))?;
316 0 : let dest: Vec<&str> = sni
317 0 : .split_once('.')
318 0 : .context("invalid SNI")?
319 : .0
320 0 : .splitn(3, "--")
321 0 : .collect();
322 0 : let port = dest[2].parse::<u16>().context("invalid port")?;
323 0 : let destination = format!("{}.{}.{}:{}", dest[0], dest[1], dest_suffix, port);
324 :
325 0 : info!("destination: {}", destination);
326 :
327 0 : let mut client = tokio::net::TcpStream::connect(&destination).await?;
328 :
329 0 : let client = if let Some(compute_tls_config) = compute_tls_config {
330 0 : info!("upgrading TLS");
331 :
332 : // send SslRequest
333 0 : client
334 0 : .write_all(b"\x00\x00\x00\x08\x04\xd2\x16\x2f")
335 0 : .await?;
336 :
337 : // wait for S/N respons
338 0 : let mut resp = b'N';
339 0 : client.read_exact(std::slice::from_mut(&mut resp)).await?;
340 :
341 : // error if not S
342 0 : ensure!(resp == b'S', "compute refused TLS");
343 :
344 : // upgrade to TLS.
345 0 : let domain = DnsName::try_from(destination)?;
346 0 : let domain = rustls::pki_types::ServerName::DnsName(domain);
347 0 : let client = TlsConnector::from(compute_tls_config)
348 0 : .connect(domain, client)
349 0 : .await?;
350 0 : Connection::Tls(client)
351 : } else {
352 0 : Connection::Raw(client)
353 : };
354 :
355 : // doesn't yet matter as pg-sni-router doesn't report analytics logs
356 0 : ctx.set_success();
357 0 : ctx.log_connect();
358 :
359 : // Starting from here we only proxy the client's traffic.
360 0 : info!("performing the proxy pass...");
361 :
362 0 : let res = match client {
363 0 : Connection::Raw(mut c) => copy_bidirectional_client_compute(&mut tls_stream, &mut c).await,
364 0 : Connection::Tls(mut c) => copy_bidirectional_client_compute(&mut tls_stream, &mut c).await,
365 : };
366 :
367 0 : match res {
368 0 : Ok(_) => Ok(()),
369 0 : Err(ErrorSource::Client(err)) => Err(err).context("client"),
370 0 : Err(ErrorSource::Compute(err)) => Err(err).context("compute"),
371 : }
372 0 : }
373 :
374 : #[allow(clippy::large_enum_variant)]
375 : enum Connection {
376 : Raw(tokio::net::TcpStream),
377 : Tls(tokio_rustls::client::TlsStream<tokio::net::TcpStream>),
378 : }
|
