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::path::Path;
8 : use std::{net::SocketAddr, sync::Arc};
9 :
10 : use anyhow::{Context, anyhow, bail, ensure};
11 : use clap::Arg;
12 : use futures::future::Either;
13 : use futures::{FutureExt, TryFutureExt};
14 : use itertools::Itertools;
15 : use rustls::crypto::ring;
16 : use rustls::pki_types::{DnsName, PrivateKeyDer};
17 : use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
18 : use tokio::net::TcpListener;
19 : use tokio_rustls::TlsConnector;
20 : use tokio_util::sync::CancellationToken;
21 : use tracing::{Instrument, error, info};
22 : use utils::project_git_version;
23 : use utils::sentry_init::init_sentry;
24 :
25 : use crate::context::RequestContext;
26 : use crate::metrics::{Metrics, ThreadPoolMetrics};
27 : use crate::protocol2::ConnectionInfo;
28 : use crate::proxy::{ErrorSource, copy_bidirectional_client_compute, run_until_cancelled};
29 : use crate::stream::{PqStream, Stream};
30 : use crate::tls::TlsServerEndPoint;
31 :
32 : project_git_version!(GIT_VERSION);
33 :
34 0 : fn cli() -> clap::Command {
35 0 : clap::Command::new("Neon proxy/router")
36 0 : .version(GIT_VERSION)
37 0 : .arg(
38 0 : Arg::new("listen")
39 0 : .short('l')
40 0 : .long("listen")
41 0 : .help("listen for incoming client connections on ip:port")
42 0 : .default_value("127.0.0.1:4432"),
43 0 : )
44 0 : .arg(
45 0 : Arg::new("listen-tls")
46 0 : .long("listen-tls")
47 0 : .help("listen for incoming client connections on ip:port, requiring TLS to compute")
48 0 : .default_value("127.0.0.1:4433"),
49 0 : )
50 0 : .arg(
51 0 : Arg::new("tls-key")
52 0 : .short('k')
53 0 : .long("tls-key")
54 0 : .help("path to TLS key for client postgres connections")
55 0 : .required(true),
56 0 : )
57 0 : .arg(
58 0 : Arg::new("tls-cert")
59 0 : .short('c')
60 0 : .long("tls-cert")
61 0 : .help("path to TLS cert for client postgres connections")
62 0 : .required(true),
63 0 : )
64 0 : .arg(
65 0 : Arg::new("dest")
66 0 : .short('d')
67 0 : .long("destination")
68 0 : .help("append this domain zone to the SNI hostname to get the destination address")
69 0 : .required(true),
70 0 : )
71 0 : }
72 :
73 0 : pub async fn run() -> anyhow::Result<()> {
74 0 : let _logging_guard = crate::logging::init().await?;
75 0 : let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
76 0 : let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
77 0 :
78 0 : Metrics::install(Arc::new(ThreadPoolMetrics::new(0)));
79 0 :
80 0 : let args = cli().get_matches();
81 0 : let destination: String = args
82 0 : .get_one::<String>("dest")
83 0 : .expect("string argument defined")
84 0 : .parse()?;
85 :
86 : // Configure TLS
87 0 : let (tls_config, tls_server_end_point): (Arc<rustls::ServerConfig>, TlsServerEndPoint) = match (
88 0 : args.get_one::<String>("tls-key"),
89 0 : args.get_one::<String>("tls-cert"),
90 : ) {
91 0 : (Some(key_path), Some(cert_path)) => parse_tls(key_path.as_ref(), cert_path.as_ref())?,
92 0 : _ => bail!("tls-key and tls-cert must be specified"),
93 : };
94 :
95 0 : let compute_tls_config =
96 0 : Arc::new(crate::tls::client_config::compute_client_config_with_root_certs()?);
97 :
98 : // Start listening for incoming client connections
99 0 : let proxy_address: SocketAddr = args
100 0 : .get_one::<String>("listen")
101 0 : .expect("listen argument defined")
102 0 : .parse()?;
103 0 : let proxy_address_compute_tls: SocketAddr = args
104 0 : .get_one::<String>("listen-tls")
105 0 : .expect("listen-tls argument defined")
106 0 : .parse()?;
107 :
108 0 : info!("Starting sni router on {proxy_address}");
109 0 : info!("Starting sni router on {proxy_address_compute_tls}");
110 0 : let proxy_listener = TcpListener::bind(proxy_address).await?;
111 0 : let proxy_listener_compute_tls = TcpListener::bind(proxy_address_compute_tls).await?;
112 :
113 0 : let cancellation_token = CancellationToken::new();
114 0 : let dest = Arc::new(destination);
115 0 :
116 0 : let main = tokio::spawn(task_main(
117 0 : dest.clone(),
118 0 : tls_config.clone(),
119 0 : None,
120 0 : tls_server_end_point,
121 0 : proxy_listener,
122 0 : cancellation_token.clone(),
123 0 : ))
124 0 : .map(crate::error::flatten_err);
125 0 :
126 0 : let main_tls = tokio::spawn(task_main(
127 0 : dest,
128 0 : tls_config,
129 0 : Some(compute_tls_config),
130 0 : tls_server_end_point,
131 0 : proxy_listener_compute_tls,
132 0 : cancellation_token.clone(),
133 0 : ))
134 0 : .map(crate::error::flatten_err);
135 0 : let signals_task = tokio::spawn(crate::signals::handle(cancellation_token, || {}));
136 0 :
137 0 : // the signal task cant ever succeed.
138 0 : // the main task can error, or can succeed on cancellation.
139 0 : // we want to immediately exit on either of these cases
140 0 : let main = futures::future::try_join(main, main_tls);
141 0 : let signal = match futures::future::select(signals_task, main).await {
142 0 : Either::Left((res, _)) => crate::error::flatten_err(res)?,
143 0 : Either::Right((res, _)) => {
144 0 : res?;
145 0 : return Ok(());
146 : }
147 : };
148 :
149 : // maintenance tasks return `Infallible` success values, this is an impossible value
150 : // so this match statically ensures that there are no possibilities for that value
151 : match signal {}
152 0 : }
153 :
154 0 : pub(super) fn parse_tls(
155 0 : key_path: &Path,
156 0 : cert_path: &Path,
157 0 : ) -> anyhow::Result<(Arc<rustls::ServerConfig>, TlsServerEndPoint)> {
158 0 : let key = {
159 0 : let key_bytes = std::fs::read(key_path).context("TLS key file")?;
160 :
161 0 : let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..]).collect_vec();
162 0 :
163 0 : ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
164 : PrivateKeyDer::Pkcs8(
165 0 : keys.pop()
166 0 : .expect("keys should not be empty")
167 0 : .context(format!(
168 0 : "Failed to read TLS keys at '{}'",
169 0 : key_path.display()
170 0 : ))?,
171 : )
172 : };
173 :
174 0 : let cert_chain_bytes = std::fs::read(cert_path).context(format!(
175 0 : "Failed to read TLS cert file at '{}.'",
176 0 : cert_path.display()
177 0 : ))?;
178 :
179 0 : let cert_chain: Vec<_> = {
180 0 : rustls_pemfile::certs(&mut &cert_chain_bytes[..])
181 0 : .try_collect()
182 0 : .with_context(|| {
183 0 : format!(
184 0 : "Failed to read TLS certificate chain from bytes from file at '{}'.",
185 0 : cert_path.display()
186 0 : )
187 0 : })?
188 : };
189 :
190 : // needed for channel bindings
191 0 : let first_cert = cert_chain.first().context("missing certificate")?;
192 0 : let tls_server_end_point = TlsServerEndPoint::new(first_cert)?;
193 :
194 0 : let tls_config =
195 0 : rustls::ServerConfig::builder_with_provider(Arc::new(ring::default_provider()))
196 0 : .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])
197 0 : .context("ring should support TLS1.2 and TLS1.3")?
198 0 : .with_no_client_auth()
199 0 : .with_single_cert(cert_chain, key)?
200 0 : .into();
201 0 :
202 0 : Ok((tls_config, tls_server_end_point))
203 0 : }
204 :
205 0 : pub(super) async fn task_main(
206 0 : dest_suffix: Arc<String>,
207 0 : tls_config: Arc<rustls::ServerConfig>,
208 0 : compute_tls_config: Option<Arc<rustls::ClientConfig>>,
209 0 : tls_server_end_point: TlsServerEndPoint,
210 0 : listener: tokio::net::TcpListener,
211 0 : cancellation_token: CancellationToken,
212 0 : ) -> anyhow::Result<()> {
213 0 : // When set for the server socket, the keepalive setting
214 0 : // will be inherited by all accepted client sockets.
215 0 : socket2::SockRef::from(&listener).set_keepalive(true)?;
216 :
217 0 : let connections = tokio_util::task::task_tracker::TaskTracker::new();
218 :
219 0 : while let Some(accept_result) =
220 0 : run_until_cancelled(listener.accept(), &cancellation_token).await
221 : {
222 0 : let (socket, peer_addr) = accept_result?;
223 :
224 0 : let session_id = uuid::Uuid::new_v4();
225 0 : let tls_config = Arc::clone(&tls_config);
226 0 : let dest_suffix = Arc::clone(&dest_suffix);
227 0 : let compute_tls_config = compute_tls_config.clone();
228 0 :
229 0 : connections.spawn(
230 0 : async move {
231 0 : socket
232 0 : .set_nodelay(true)
233 0 : .context("failed to set socket option")?;
234 :
235 0 : info!(%peer_addr, "serving");
236 0 : let ctx = RequestContext::new(
237 0 : session_id,
238 0 : ConnectionInfo {
239 0 : addr: peer_addr,
240 0 : extra: None,
241 0 : },
242 0 : crate::metrics::Protocol::SniRouter,
243 0 : "sni",
244 0 : );
245 0 : handle_client(
246 0 : ctx,
247 0 : dest_suffix,
248 0 : tls_config,
249 0 : compute_tls_config,
250 0 : tls_server_end_point,
251 0 : socket,
252 0 : )
253 0 : .await
254 0 : }
255 0 : .unwrap_or_else(|e| {
256 0 : // Acknowledge that the task has finished with an error.
257 0 : error!("per-client task finished with an error: {e:#}");
258 0 : })
259 0 : .instrument(tracing::info_span!("handle_client", ?session_id)),
260 : );
261 : }
262 :
263 0 : connections.close();
264 0 : drop(listener);
265 0 :
266 0 : connections.wait().await;
267 :
268 0 : info!("all client connections have finished");
269 0 : Ok(())
270 0 : }
271 :
272 : const ERR_INSECURE_CONNECTION: &str = "connection is insecure (try using `sslmode=require`)";
273 :
274 0 : async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
275 0 : ctx: &RequestContext,
276 0 : raw_stream: S,
277 0 : tls_config: Arc<rustls::ServerConfig>,
278 0 : tls_server_end_point: TlsServerEndPoint,
279 0 : ) -> anyhow::Result<Stream<S>> {
280 0 : let mut stream = PqStream::new(Stream::from_raw(raw_stream));
281 :
282 0 : let msg = stream.read_startup_packet().await?;
283 : use pq_proto::FeStartupPacket::SslRequest;
284 :
285 0 : match msg {
286 : SslRequest { direct: false } => {
287 0 : stream
288 0 : .write_message(&pq_proto::BeMessage::EncryptionResponse(true))
289 0 : .await?;
290 :
291 : // Upgrade raw stream into a secure TLS-backed stream.
292 : // NOTE: We've consumed `tls`; this fact will be used later.
293 :
294 0 : let (raw, read_buf) = stream.into_inner();
295 0 : // TODO: Normally, client doesn't send any data before
296 0 : // server says TLS handshake is ok and read_buf is empty.
297 0 : // However, you could imagine pipelining of postgres
298 0 : // SSLRequest + TLS ClientHello in one hunk similar to
299 0 : // pipelining in our node js driver. We should probably
300 0 : // support that by chaining read_buf with the stream.
301 0 : if !read_buf.is_empty() {
302 0 : bail!("data is sent before server replied with EncryptionResponse");
303 0 : }
304 0 :
305 0 : Ok(Stream::Tls {
306 0 : tls: Box::new(
307 0 : raw.upgrade(tls_config, !ctx.has_private_peer_addr())
308 0 : .await?,
309 : ),
310 0 : tls_server_end_point,
311 : })
312 : }
313 0 : unexpected => {
314 0 : info!(
315 : ?unexpected,
316 0 : "unexpected startup packet, rejecting connection"
317 : );
318 0 : stream
319 0 : .throw_error_str(ERR_INSECURE_CONNECTION, crate::error::ErrorKind::User, None)
320 0 : .await?
321 : }
322 : }
323 0 : }
324 :
325 0 : async fn handle_client(
326 0 : ctx: RequestContext,
327 0 : dest_suffix: Arc<String>,
328 0 : tls_config: Arc<rustls::ServerConfig>,
329 0 : compute_tls_config: Option<Arc<rustls::ClientConfig>>,
330 0 : tls_server_end_point: TlsServerEndPoint,
331 0 : stream: impl AsyncRead + AsyncWrite + Unpin,
332 0 : ) -> anyhow::Result<()> {
333 0 : let mut tls_stream = ssl_handshake(&ctx, stream, tls_config, tls_server_end_point).await?;
334 :
335 : // Cut off first part of the SNI domain
336 : // We receive required destination details in the format of
337 : // `{k8s_service_name}--{k8s_namespace}--{port}.non-sni-domain`
338 0 : let sni = tls_stream.sni_hostname().ok_or(anyhow!("SNI missing"))?;
339 0 : let dest: Vec<&str> = sni
340 0 : .split_once('.')
341 0 : .context("invalid SNI")?
342 : .0
343 0 : .splitn(3, "--")
344 0 : .collect();
345 0 : let port = dest[2].parse::<u16>().context("invalid port")?;
346 0 : let destination = format!("{}.{}.{}:{}", dest[0], dest[1], dest_suffix, port);
347 0 :
348 0 : info!("destination: {}", destination);
349 :
350 0 : let mut client = tokio::net::TcpStream::connect(&destination).await?;
351 :
352 0 : let client = if let Some(compute_tls_config) = compute_tls_config {
353 0 : info!("upgrading TLS");
354 :
355 : // send SslRequest
356 0 : client
357 0 : .write_all(b"\x00\x00\x00\x08\x04\xd2\x16\x2f")
358 0 : .await?;
359 :
360 : // wait for S/N respons
361 0 : let mut resp = b'N';
362 0 : client.read_exact(std::slice::from_mut(&mut resp)).await?;
363 :
364 : // error if not S
365 0 : ensure!(resp == b'S', "compute refused TLS");
366 :
367 : // upgrade to TLS.
368 0 : let domain = DnsName::try_from(destination)?;
369 0 : let domain = rustls::pki_types::ServerName::DnsName(domain);
370 0 : let client = TlsConnector::from(compute_tls_config)
371 0 : .connect(domain, client)
372 0 : .await?;
373 0 : Connection::Tls(client)
374 : } else {
375 0 : Connection::Raw(client)
376 : };
377 :
378 : // doesn't yet matter as pg-sni-router doesn't report analytics logs
379 0 : ctx.set_success();
380 0 : ctx.log_connect();
381 0 :
382 0 : // Starting from here we only proxy the client's traffic.
383 0 : info!("performing the proxy pass...");
384 :
385 0 : let res = match client {
386 0 : Connection::Raw(mut c) => copy_bidirectional_client_compute(&mut tls_stream, &mut c).await,
387 0 : Connection::Tls(mut c) => copy_bidirectional_client_compute(&mut tls_stream, &mut c).await,
388 : };
389 :
390 0 : match res {
391 0 : Ok(_) => Ok(()),
392 0 : Err(ErrorSource::Client(err)) => Err(err).context("client"),
393 0 : Err(ErrorSource::Compute(err)) => Err(err).context("compute"),
394 : }
395 0 : }
396 :
397 : #[allow(clippy::large_enum_variant)]
398 : enum Connection {
399 : Raw(tokio::net::TcpStream),
400 : Tls(tokio_rustls::client::TlsStream<tokio::net::TcpStream>),
401 : }
|
