Line data Source code
1 : #[cfg(test)]
2 : mod tests;
3 :
4 : pub mod connect_compute;
5 : mod copy_bidirectional;
6 : pub mod handshake;
7 : pub mod passthrough;
8 : pub mod retry;
9 : pub mod wake_compute;
10 :
11 : use crate::{
12 : auth,
13 : cancellation::{self, CancellationHandler},
14 : compute,
15 : config::{ProxyConfig, TlsConfig},
16 : context::RequestMonitoring,
17 : error::ReportableError,
18 : metrics::{NUM_CLIENT_CONNECTION_GAUGE, NUM_CONNECTION_REQUESTS_GAUGE},
19 : protocol2::WithClientIp,
20 : proxy::handshake::{handshake, HandshakeData},
21 : rate_limiter::EndpointRateLimiter,
22 : stream::{PqStream, Stream},
23 : EndpointCacheKey,
24 : };
25 : use anyhow::{bail, Context};
26 : use futures::TryFutureExt;
27 : use itertools::Itertools;
28 : use once_cell::sync::OnceCell;
29 : use pq_proto::{BeMessage as Be, StartupMessageParams};
30 : use regex::Regex;
31 : use smol_str::{format_smolstr, SmolStr};
32 : use std::sync::Arc;
33 : use thiserror::Error;
34 : use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
35 : use tokio_util::sync::CancellationToken;
36 : use tracing::{error, info, info_span, Instrument};
37 :
38 : use self::{
39 : connect_compute::{connect_to_compute, TcpMechanism},
40 : passthrough::ProxyPassthrough,
41 : };
42 :
43 : const ERR_INSECURE_CONNECTION: &str = "connection is insecure (try using `sslmode=require`)";
44 :
45 0 : pub async fn run_until_cancelled<F: std::future::Future>(
46 0 : f: F,
47 0 : cancellation_token: &CancellationToken,
48 0 : ) -> Option<F::Output> {
49 0 : match futures::future::select(
50 0 : std::pin::pin!(f),
51 0 : std::pin::pin!(cancellation_token.cancelled()),
52 0 : )
53 0 : .await
54 : {
55 0 : futures::future::Either::Left((f, _)) => Some(f),
56 0 : futures::future::Either::Right(((), _)) => None,
57 : }
58 0 : }
59 :
60 0 : pub async fn task_main(
61 0 : config: &'static ProxyConfig,
62 0 : listener: tokio::net::TcpListener,
63 0 : cancellation_token: CancellationToken,
64 0 : endpoint_rate_limiter: Arc<EndpointRateLimiter>,
65 0 : cancellation_handler: Arc<CancellationHandler>,
66 0 : ) -> anyhow::Result<()> {
67 0 : scopeguard::defer! {
68 0 : info!("proxy has shut down");
69 : }
70 :
71 : // When set for the server socket, the keepalive setting
72 : // will be inherited by all accepted client sockets.
73 0 : socket2::SockRef::from(&listener).set_keepalive(true)?;
74 :
75 0 : let connections = tokio_util::task::task_tracker::TaskTracker::new();
76 :
77 0 : while let Some(accept_result) =
78 0 : run_until_cancelled(listener.accept(), &cancellation_token).await
79 : {
80 0 : let (socket, peer_addr) = accept_result?;
81 :
82 0 : let session_id = uuid::Uuid::new_v4();
83 0 : let cancellation_handler = Arc::clone(&cancellation_handler);
84 0 : let endpoint_rate_limiter = endpoint_rate_limiter.clone();
85 :
86 0 : let session_span = info_span!(
87 0 : "handle_client",
88 0 : ?session_id,
89 0 : peer_addr = tracing::field::Empty,
90 0 : ep = tracing::field::Empty,
91 0 : );
92 :
93 0 : connections.spawn(
94 0 : async move {
95 0 : info!("accepted postgres client connection");
96 :
97 0 : let mut socket = WithClientIp::new(socket);
98 0 : let mut peer_addr = peer_addr.ip();
99 0 : if let Some(addr) = socket.wait_for_addr().await? {
100 0 : peer_addr = addr.ip();
101 0 : tracing::Span::current().record("peer_addr", &tracing::field::display(addr));
102 0 : } else if config.require_client_ip {
103 0 : bail!("missing required client IP");
104 0 : }
105 :
106 0 : socket
107 0 : .inner
108 0 : .set_nodelay(true)
109 0 : .context("failed to set socket option")?;
110 :
111 0 : let mut ctx = RequestMonitoring::new(session_id, peer_addr, "tcp", &config.region);
112 :
113 0 : let res = handle_client(
114 0 : config,
115 0 : &mut ctx,
116 0 : cancellation_handler,
117 0 : socket,
118 0 : ClientMode::Tcp,
119 0 : endpoint_rate_limiter,
120 0 : )
121 0 : .await;
122 :
123 0 : match res {
124 0 : Err(e) => {
125 0 : // todo: log and push to ctx the error kind
126 0 : ctx.set_error_kind(e.get_error_kind());
127 0 : ctx.log();
128 0 : Err(e.into())
129 : }
130 : Ok(None) => {
131 0 : ctx.set_success();
132 0 : ctx.log();
133 0 : Ok(())
134 : }
135 0 : Ok(Some(p)) => {
136 0 : ctx.set_success();
137 0 : ctx.log();
138 0 : p.proxy_pass().await
139 : }
140 : }
141 0 : }
142 0 : .unwrap_or_else(move |e| {
143 0 : // Acknowledge that the task has finished with an error.
144 0 : error!("per-client task finished with an error: {e:#}");
145 0 : })
146 0 : .instrument(session_span),
147 0 : );
148 : }
149 :
150 0 : connections.close();
151 0 : drop(listener);
152 0 :
153 0 : // Drain connections
154 0 : connections.wait().await;
155 :
156 0 : Ok(())
157 0 : }
158 :
159 : pub enum ClientMode {
160 : Tcp,
161 : Websockets { hostname: Option<String> },
162 : }
163 :
164 : /// Abstracts the logic of handling TCP vs WS clients
165 : impl ClientMode {
166 0 : pub fn allow_cleartext(&self) -> bool {
167 0 : match self {
168 0 : ClientMode::Tcp => false,
169 0 : ClientMode::Websockets { .. } => true,
170 : }
171 0 : }
172 :
173 0 : pub fn allow_self_signed_compute(&self, config: &ProxyConfig) -> bool {
174 0 : match self {
175 0 : ClientMode::Tcp => config.allow_self_signed_compute,
176 0 : ClientMode::Websockets { .. } => false,
177 : }
178 0 : }
179 :
180 0 : fn hostname<'a, S>(&'a self, s: &'a Stream<S>) -> Option<&'a str> {
181 0 : match self {
182 0 : ClientMode::Tcp => s.sni_hostname(),
183 0 : ClientMode::Websockets { hostname } => hostname.as_deref(),
184 : }
185 0 : }
186 :
187 0 : fn handshake_tls<'a>(&self, tls: Option<&'a TlsConfig>) -> Option<&'a TlsConfig> {
188 0 : match self {
189 0 : ClientMode::Tcp => tls,
190 : // TLS is None here if using websockets, because the connection is already encrypted.
191 0 : ClientMode::Websockets { .. } => None,
192 : }
193 0 : }
194 : }
195 :
196 0 : #[derive(Debug, Error)]
197 : // almost all errors should be reported to the user, but there's a few cases where we cannot
198 : // 1. Cancellation: we are not allowed to tell the client any cancellation statuses for security reasons
199 : // 2. Handshake: handshake reports errors if it can, otherwise if the handshake fails due to protocol violation,
200 : // we cannot be sure the client even understands our error message
201 : // 3. PrepareClient: The client disconnected, so we can't tell them anyway...
202 : pub enum ClientRequestError {
203 : #[error("{0}")]
204 : Cancellation(#[from] cancellation::CancelError),
205 : #[error("{0}")]
206 : Handshake(#[from] handshake::HandshakeError),
207 : #[error("{0}")]
208 : HandshakeTimeout(#[from] tokio::time::error::Elapsed),
209 : #[error("{0}")]
210 : PrepareClient(#[from] std::io::Error),
211 : #[error("{0}")]
212 : ReportedError(#[from] crate::stream::ReportedError),
213 : }
214 :
215 : impl ReportableError for ClientRequestError {
216 0 : fn get_error_kind(&self) -> crate::error::ErrorKind {
217 0 : match self {
218 0 : ClientRequestError::Cancellation(e) => e.get_error_kind(),
219 0 : ClientRequestError::Handshake(e) => e.get_error_kind(),
220 0 : ClientRequestError::HandshakeTimeout(_) => crate::error::ErrorKind::RateLimit,
221 0 : ClientRequestError::ReportedError(e) => e.get_error_kind(),
222 0 : ClientRequestError::PrepareClient(_) => crate::error::ErrorKind::ClientDisconnect,
223 : }
224 0 : }
225 : }
226 :
227 0 : pub async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
228 0 : config: &'static ProxyConfig,
229 0 : ctx: &mut RequestMonitoring,
230 0 : cancellation_handler: Arc<CancellationHandler>,
231 0 : stream: S,
232 0 : mode: ClientMode,
233 0 : endpoint_rate_limiter: Arc<EndpointRateLimiter>,
234 0 : ) -> Result<Option<ProxyPassthrough<S>>, ClientRequestError> {
235 0 : info!(
236 0 : protocol = ctx.protocol,
237 0 : "handling interactive connection from client"
238 0 : );
239 :
240 0 : let proto = ctx.protocol;
241 0 : let _client_gauge = NUM_CLIENT_CONNECTION_GAUGE
242 0 : .with_label_values(&[proto])
243 0 : .guard();
244 0 : let _request_gauge = NUM_CONNECTION_REQUESTS_GAUGE
245 0 : .with_label_values(&[proto])
246 0 : .guard();
247 0 :
248 0 : let tls = config.tls_config.as_ref();
249 0 :
250 0 : let pause = ctx.latency_timer.pause();
251 0 : let do_handshake = handshake(stream, mode.handshake_tls(tls));
252 0 : let (mut stream, params) =
253 0 : match tokio::time::timeout(config.handshake_timeout, do_handshake).await?? {
254 0 : HandshakeData::Startup(stream, params) => (stream, params),
255 0 : HandshakeData::Cancel(cancel_key_data) => {
256 0 : return Ok(cancellation_handler
257 0 : .cancel_session(cancel_key_data, ctx.session_id)
258 0 : .await
259 0 : .map(|()| None)?)
260 : }
261 : };
262 0 : drop(pause);
263 0 :
264 0 : let hostname = mode.hostname(stream.get_ref());
265 0 :
266 0 : let common_names = tls.map(|tls| &tls.common_names);
267 0 :
268 0 : // Extract credentials which we're going to use for auth.
269 0 : let result = config
270 0 : .auth_backend
271 0 : .as_ref()
272 0 : .map(|_| auth::ComputeUserInfoMaybeEndpoint::parse(ctx, ¶ms, hostname, common_names))
273 0 : .transpose();
274 :
275 0 : let user_info = match result {
276 0 : Ok(user_info) => user_info,
277 0 : Err(e) => stream.throw_error(e).await?,
278 : };
279 :
280 : // check rate limit
281 0 : if let Some(ep) = user_info.get_endpoint() {
282 0 : if !endpoint_rate_limiter.check(ep) {
283 0 : return stream
284 0 : .throw_error(auth::AuthError::too_many_connections())
285 0 : .await?;
286 0 : }
287 0 : }
288 :
289 0 : let user = user_info.get_user().to_owned();
290 0 : let user_info = match user_info
291 0 : .authenticate(
292 0 : ctx,
293 0 : &mut stream,
294 0 : mode.allow_cleartext(),
295 0 : &config.authentication_config,
296 0 : )
297 0 : .await
298 : {
299 0 : Ok(auth_result) => auth_result,
300 0 : Err(e) => {
301 0 : let db = params.get("database");
302 0 : let app = params.get("application_name");
303 0 : let params_span = tracing::info_span!("", ?user, ?db, ?app);
304 :
305 0 : return stream.throw_error(e).instrument(params_span).await?;
306 : }
307 : };
308 :
309 0 : let mut node = connect_to_compute(
310 0 : ctx,
311 0 : &TcpMechanism { params: ¶ms },
312 0 : &user_info,
313 0 : mode.allow_self_signed_compute(config),
314 0 : )
315 0 : .or_else(|e| stream.throw_error(e))
316 0 : .await?;
317 :
318 0 : let session = cancellation_handler.get_session();
319 0 : prepare_client_connection(&node, &session, &mut stream).await?;
320 :
321 : // Before proxy passing, forward to compute whatever data is left in the
322 : // PqStream input buffer. Normally there is none, but our serverless npm
323 : // driver in pipeline mode sends startup, password and first query
324 : // immediately after opening the connection.
325 0 : let (stream, read_buf) = stream.into_inner();
326 0 : node.stream.write_all(&read_buf).await?;
327 :
328 0 : Ok(Some(ProxyPassthrough {
329 0 : client: stream,
330 0 : aux: node.aux.clone(),
331 0 : compute: node,
332 0 : req: _request_gauge,
333 0 : conn: _client_gauge,
334 0 : cancel: session,
335 0 : }))
336 0 : }
337 :
338 : /// Finish client connection initialization: confirm auth success, send params, etc.
339 0 : #[tracing::instrument(skip_all)]
340 : async fn prepare_client_connection(
341 : node: &compute::PostgresConnection,
342 : session: &cancellation::Session,
343 : stream: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
344 : ) -> Result<(), std::io::Error> {
345 : // Register compute's query cancellation token and produce a new, unique one.
346 : // The new token (cancel_key_data) will be sent to the client.
347 : let cancel_key_data = session.enable_query_cancellation(node.cancel_closure.clone());
348 :
349 : // Forward all postgres connection params to the client.
350 : // Right now the implementation is very hacky and inefficent (ideally,
351 : // we don't need an intermediate hashmap), but at least it should be correct.
352 : for (name, value) in &node.params {
353 : // TODO: Theoretically, this could result in a big pile of params...
354 : stream.write_message_noflush(&Be::ParameterStatus {
355 : name: name.as_bytes(),
356 : value: value.as_bytes(),
357 : })?;
358 : }
359 :
360 : stream
361 : .write_message_noflush(&Be::BackendKeyData(cancel_key_data))?
362 : .write_message(&Be::ReadyForQuery)
363 : .await?;
364 :
365 : Ok(())
366 : }
367 :
368 26 : #[derive(Debug, Clone, PartialEq, Eq, Default)]
369 : pub struct NeonOptions(Vec<(SmolStr, SmolStr)>);
370 :
371 : impl NeonOptions {
372 22 : pub fn parse_params(params: &StartupMessageParams) -> Self {
373 22 : params
374 22 : .options_raw()
375 22 : .map(Self::parse_from_iter)
376 22 : .unwrap_or_default()
377 22 : }
378 14 : pub fn parse_options_raw(options: &str) -> Self {
379 14 : Self::parse_from_iter(StartupMessageParams::parse_options_raw(options))
380 14 : }
381 :
382 26 : fn parse_from_iter<'a>(options: impl Iterator<Item = &'a str>) -> Self {
383 26 : let mut options = options
384 26 : .filter_map(neon_option)
385 26 : .map(|(k, v)| (k.into(), v.into()))
386 26 : .collect_vec();
387 26 : options.sort();
388 26 : Self(options)
389 26 : }
390 :
391 8 : pub fn get_cache_key(&self, prefix: &str) -> EndpointCacheKey {
392 8 : // prefix + format!(" {k}:{v}")
393 8 : // kinda jank because SmolStr is immutable
394 8 : std::iter::once(prefix)
395 8 : .chain(self.0.iter().flat_map(|(k, v)| [" ", &**k, ":", &**v]))
396 8 : .collect::<SmolStr>()
397 8 : .into()
398 8 : }
399 :
400 : /// <https://swagger.io/docs/specification/serialization/> DeepObject format
401 : /// `paramName[prop1]=value1¶mName[prop2]=value2&...`
402 0 : pub fn to_deep_object(&self) -> Vec<(SmolStr, SmolStr)> {
403 0 : self.0
404 0 : .iter()
405 0 : .map(|(k, v)| (format_smolstr!("options[{}]", k), v.clone()))
406 0 : .collect()
407 0 : }
408 : }
409 :
410 64 : pub fn neon_option(bytes: &str) -> Option<(&str, &str)> {
411 64 : static RE: OnceCell<Regex> = OnceCell::new();
412 64 : let re = RE.get_or_init(|| Regex::new(r"^neon_(\w+):(.+)").unwrap());
413 :
414 64 : let cap = re.captures(bytes)?;
415 8 : let (_, [k, v]) = cap.extract();
416 8 : Some((k, v))
417 64 : }
|
