Line data Source code
1 : //! Routers for our serverless APIs
2 : //!
3 : //! Handles both SQL over HTTP and SQL over Websockets.
4 :
5 : mod backend;
6 : pub mod cancel_set;
7 : mod conn_pool;
8 : mod conn_pool_lib;
9 : mod error;
10 : mod http_conn_pool;
11 : mod http_util;
12 : mod json;
13 : mod local_conn_pool;
14 : mod sql_over_http;
15 : mod websocket;
16 :
17 : use std::net::{IpAddr, SocketAddr};
18 : use std::pin::{pin, Pin};
19 : use std::sync::Arc;
20 :
21 : use anyhow::Context;
22 : use async_trait::async_trait;
23 : use atomic_take::AtomicTake;
24 : use bytes::Bytes;
25 : pub use conn_pool_lib::GlobalConnPoolOptions;
26 : use futures::future::{select, Either};
27 : use futures::TryFutureExt;
28 : use http::{Method, Response, StatusCode};
29 : use http_body_util::combinators::BoxBody;
30 : use http_body_util::{BodyExt, Empty};
31 : use hyper::body::Incoming;
32 : use hyper_util::rt::TokioExecutor;
33 : use hyper_util::server::conn::auto::Builder;
34 : use rand::rngs::StdRng;
35 : use rand::SeedableRng;
36 : use sql_over_http::{uuid_to_header_value, NEON_REQUEST_ID};
37 : use tokio::io::{AsyncRead, AsyncWrite};
38 : use tokio::net::{TcpListener, TcpStream};
39 : use tokio::time::timeout;
40 : use tokio_rustls::TlsAcceptor;
41 : use tokio_util::sync::CancellationToken;
42 : use tokio_util::task::TaskTracker;
43 : use tracing::{info, warn, Instrument};
44 : use utils::http::error::ApiError;
45 :
46 : use crate::cancellation::CancellationHandlerMain;
47 : use crate::config::{ProxyConfig, ProxyProtocolV2};
48 : use crate::context::RequestMonitoring;
49 : use crate::metrics::Metrics;
50 : use crate::protocol2::{read_proxy_protocol, ChainRW, ConnectHeader, ConnectionInfo};
51 : use crate::proxy::run_until_cancelled;
52 : use crate::rate_limiter::EndpointRateLimiter;
53 : use crate::serverless::backend::PoolingBackend;
54 : use crate::serverless::http_util::{api_error_into_response, json_response};
55 :
56 : pub(crate) const SERVERLESS_DRIVER_SNI: &str = "api";
57 :
58 0 : pub async fn task_main(
59 0 : config: &'static ProxyConfig,
60 0 : auth_backend: &'static crate::auth::Backend<'static, ()>,
61 0 : ws_listener: TcpListener,
62 0 : cancellation_token: CancellationToken,
63 0 : cancellation_handler: Arc<CancellationHandlerMain>,
64 0 : endpoint_rate_limiter: Arc<EndpointRateLimiter>,
65 0 : ) -> anyhow::Result<()> {
66 0 : scopeguard::defer! {
67 0 : info!("websocket server has shut down");
68 0 : }
69 0 :
70 0 : let local_pool = local_conn_pool::LocalConnPool::new(&config.http_config);
71 0 : let conn_pool = conn_pool_lib::GlobalConnPool::new(&config.http_config);
72 0 : {
73 0 : let conn_pool = Arc::clone(&conn_pool);
74 0 : tokio::spawn(async move {
75 0 : conn_pool.gc_worker(StdRng::from_entropy()).await;
76 0 : });
77 0 : }
78 0 :
79 0 : // shutdown the connection pool
80 0 : tokio::spawn({
81 0 : let cancellation_token = cancellation_token.clone();
82 0 : let conn_pool = conn_pool.clone();
83 0 : async move {
84 0 : cancellation_token.cancelled().await;
85 0 : tokio::task::spawn_blocking(move || conn_pool.shutdown())
86 0 : .await
87 0 : .unwrap();
88 0 : }
89 0 : });
90 0 :
91 0 : let http_conn_pool = http_conn_pool::GlobalConnPool::new(&config.http_config);
92 0 : {
93 0 : let http_conn_pool = Arc::clone(&http_conn_pool);
94 0 : tokio::spawn(async move {
95 0 : http_conn_pool.gc_worker(StdRng::from_entropy()).await;
96 0 : });
97 0 : }
98 0 :
99 0 : // shutdown the connection pool
100 0 : tokio::spawn({
101 0 : let cancellation_token = cancellation_token.clone();
102 0 : let http_conn_pool = http_conn_pool.clone();
103 0 : async move {
104 0 : cancellation_token.cancelled().await;
105 0 : tokio::task::spawn_blocking(move || http_conn_pool.shutdown())
106 0 : .await
107 0 : .unwrap();
108 0 : }
109 0 : });
110 0 :
111 0 : let backend = Arc::new(PoolingBackend {
112 0 : http_conn_pool: Arc::clone(&http_conn_pool),
113 0 : local_pool,
114 0 : pool: Arc::clone(&conn_pool),
115 0 : config,
116 0 : auth_backend,
117 0 : endpoint_rate_limiter: Arc::clone(&endpoint_rate_limiter),
118 0 : });
119 0 : let tls_acceptor: Arc<dyn MaybeTlsAcceptor> = match config.tls_config.as_ref() {
120 0 : Some(config) => {
121 0 : let mut tls_server_config = rustls::ServerConfig::clone(&config.to_server_config());
122 0 : // prefer http2, but support http/1.1
123 0 : tls_server_config.alpn_protocols = vec![b"h2".to_vec(), b"http/1.1".to_vec()];
124 0 : Arc::new(tls_server_config)
125 : }
126 : None => {
127 0 : warn!("TLS config is missing");
128 0 : Arc::new(NoTls)
129 : }
130 : };
131 :
132 0 : let connections = tokio_util::task::task_tracker::TaskTracker::new();
133 0 : connections.close(); // allows `connections.wait to complete`
134 :
135 0 : while let Some(res) = run_until_cancelled(ws_listener.accept(), &cancellation_token).await {
136 0 : let (conn, peer_addr) = res.context("could not accept TCP stream")?;
137 0 : if let Err(e) = conn.set_nodelay(true) {
138 0 : tracing::error!("could not set nodelay: {e}");
139 0 : continue;
140 0 : }
141 0 : let conn_id = uuid::Uuid::new_v4();
142 0 : let http_conn_span = tracing::info_span!("http_conn", ?conn_id);
143 :
144 0 : let n_connections = Metrics::get()
145 0 : .proxy
146 0 : .client_connections
147 0 : .sample(crate::metrics::Protocol::Http);
148 0 : tracing::trace!(?n_connections, threshold = ?config.http_config.client_conn_threshold, "check");
149 0 : if n_connections > config.http_config.client_conn_threshold {
150 0 : tracing::trace!("attempting to cancel a random connection");
151 0 : if let Some(token) = config.http_config.cancel_set.take() {
152 0 : tracing::debug!("cancelling a random connection");
153 0 : token.cancel();
154 0 : }
155 0 : }
156 :
157 0 : let conn_token = cancellation_token.child_token();
158 0 : let tls_acceptor = tls_acceptor.clone();
159 0 : let backend = backend.clone();
160 0 : let connections2 = connections.clone();
161 0 : let cancellation_handler = cancellation_handler.clone();
162 0 : let endpoint_rate_limiter = endpoint_rate_limiter.clone();
163 0 : connections.spawn(
164 0 : async move {
165 0 : let conn_token2 = conn_token.clone();
166 0 : let _cancel_guard = config.http_config.cancel_set.insert(conn_id, conn_token2);
167 0 :
168 0 : let session_id = uuid::Uuid::new_v4();
169 0 :
170 0 : let _gauge = Metrics::get()
171 0 : .proxy
172 0 : .client_connections
173 0 : .guard(crate::metrics::Protocol::Http);
174 :
175 0 : let startup_result = Box::pin(connection_startup(
176 0 : config,
177 0 : tls_acceptor,
178 0 : session_id,
179 0 : conn,
180 0 : peer_addr,
181 0 : ))
182 0 : .await;
183 0 : let Some((conn, conn_info)) = startup_result else {
184 0 : return;
185 : };
186 :
187 0 : Box::pin(connection_handler(
188 0 : config,
189 0 : backend,
190 0 : connections2,
191 0 : cancellation_handler,
192 0 : endpoint_rate_limiter,
193 0 : conn_token,
194 0 : conn,
195 0 : conn_info,
196 0 : session_id,
197 0 : ))
198 0 : .await;
199 0 : }
200 0 : .instrument(http_conn_span),
201 0 : );
202 : }
203 :
204 0 : connections.wait().await;
205 :
206 0 : Ok(())
207 0 : }
208 :
209 : pub(crate) trait AsyncReadWrite: AsyncRead + AsyncWrite + Send + 'static {}
210 : impl<T: AsyncRead + AsyncWrite + Send + 'static> AsyncReadWrite for T {}
211 : pub(crate) type AsyncRW = Pin<Box<dyn AsyncReadWrite>>;
212 :
213 : #[async_trait]
214 : trait MaybeTlsAcceptor: Send + Sync + 'static {
215 : async fn accept(self: Arc<Self>, conn: ChainRW<TcpStream>) -> std::io::Result<AsyncRW>;
216 : }
217 :
218 : #[async_trait]
219 : impl MaybeTlsAcceptor for rustls::ServerConfig {
220 0 : async fn accept(self: Arc<Self>, conn: ChainRW<TcpStream>) -> std::io::Result<AsyncRW> {
221 0 : Ok(Box::pin(TlsAcceptor::from(self).accept(conn).await?))
222 0 : }
223 : }
224 :
225 : struct NoTls;
226 :
227 : #[async_trait]
228 : impl MaybeTlsAcceptor for NoTls {
229 0 : async fn accept(self: Arc<Self>, conn: ChainRW<TcpStream>) -> std::io::Result<AsyncRW> {
230 0 : Ok(Box::pin(conn))
231 0 : }
232 : }
233 :
234 : /// Handles the TCP startup lifecycle.
235 : /// 1. Parses PROXY protocol V2
236 : /// 2. Handles TLS handshake
237 0 : async fn connection_startup(
238 0 : config: &ProxyConfig,
239 0 : tls_acceptor: Arc<dyn MaybeTlsAcceptor>,
240 0 : session_id: uuid::Uuid,
241 0 : conn: TcpStream,
242 0 : peer_addr: SocketAddr,
243 0 : ) -> Option<(AsyncRW, ConnectionInfo)> {
244 : // handle PROXY protocol
245 0 : let (conn, peer) = match read_proxy_protocol(conn).await {
246 0 : Ok(c) => c,
247 0 : Err(e) => {
248 0 : tracing::warn!(?session_id, %peer_addr, "failed to accept TCP connection: invalid PROXY protocol V2 header: {e:#}");
249 0 : return None;
250 : }
251 : };
252 :
253 0 : let conn_info = match peer {
254 : // our load balancers will not send any more data. let's just exit immediately
255 : ConnectHeader::Local => {
256 0 : tracing::debug!("healthcheck received");
257 0 : return None;
258 : }
259 0 : ConnectHeader::Missing if config.proxy_protocol_v2 == ProxyProtocolV2::Required => {
260 0 : tracing::warn!("missing required proxy protocol header");
261 0 : return None;
262 : }
263 0 : ConnectHeader::Proxy(_) if config.proxy_protocol_v2 == ProxyProtocolV2::Rejected => {
264 0 : tracing::warn!("proxy protocol header not supported");
265 0 : return None;
266 : }
267 0 : ConnectHeader::Proxy(info) => info,
268 0 : ConnectHeader::Missing => ConnectionInfo {
269 0 : addr: peer_addr,
270 0 : extra: None,
271 0 : },
272 : };
273 :
274 0 : let has_private_peer_addr = match conn_info.addr.ip() {
275 0 : IpAddr::V4(ip) => ip.is_private(),
276 0 : IpAddr::V6(_) => false,
277 : };
278 0 : info!(?session_id, %conn_info, "accepted new TCP connection");
279 :
280 : // try upgrade to TLS, but with a timeout.
281 0 : let conn = match timeout(config.handshake_timeout, tls_acceptor.accept(conn)).await {
282 0 : Ok(Ok(conn)) => {
283 0 : info!(?session_id, %conn_info, "accepted new TLS connection");
284 0 : conn
285 : }
286 : // The handshake failed
287 0 : Ok(Err(e)) => {
288 0 : if !has_private_peer_addr {
289 0 : Metrics::get().proxy.tls_handshake_failures.inc();
290 0 : }
291 0 : warn!(?session_id, %conn_info, "failed to accept TLS connection: {e:?}");
292 0 : return None;
293 : }
294 : // The handshake timed out
295 0 : Err(e) => {
296 0 : if !has_private_peer_addr {
297 0 : Metrics::get().proxy.tls_handshake_failures.inc();
298 0 : }
299 0 : warn!(?session_id, %conn_info, "failed to accept TLS connection: {e:?}");
300 0 : return None;
301 : }
302 : };
303 :
304 0 : Some((conn, conn_info))
305 0 : }
306 :
307 : /// Handles HTTP connection
308 : /// 1. With graceful shutdowns
309 : /// 2. With graceful request cancellation with connection failure
310 : /// 3. With websocket upgrade support.
311 : #[allow(clippy::too_many_arguments)]
312 0 : async fn connection_handler(
313 0 : config: &'static ProxyConfig,
314 0 : backend: Arc<PoolingBackend>,
315 0 : connections: TaskTracker,
316 0 : cancellation_handler: Arc<CancellationHandlerMain>,
317 0 : endpoint_rate_limiter: Arc<EndpointRateLimiter>,
318 0 : cancellation_token: CancellationToken,
319 0 : conn: AsyncRW,
320 0 : conn_info: ConnectionInfo,
321 0 : session_id: uuid::Uuid,
322 0 : ) {
323 0 : let session_id = AtomicTake::new(session_id);
324 0 :
325 0 : // Cancel all current inflight HTTP requests if the HTTP connection is closed.
326 0 : let http_cancellation_token = CancellationToken::new();
327 0 : let _cancel_connection = http_cancellation_token.clone().drop_guard();
328 0 :
329 0 : let conn_info2 = conn_info.clone();
330 0 : let server = Builder::new(TokioExecutor::new());
331 0 : let conn = server.serve_connection_with_upgrades(
332 0 : hyper_util::rt::TokioIo::new(conn),
333 0 : hyper::service::service_fn(move |req: hyper::Request<Incoming>| {
334 0 : // First HTTP request shares the same session ID
335 0 : let mut session_id = session_id.take().unwrap_or_else(uuid::Uuid::new_v4);
336 :
337 0 : if matches!(backend.auth_backend, crate::auth::Backend::Local(_)) {
338 : // take session_id from request, if given.
339 0 : if let Some(id) = req
340 0 : .headers()
341 0 : .get(&NEON_REQUEST_ID)
342 0 : .and_then(|id| uuid::Uuid::try_parse_ascii(id.as_bytes()).ok())
343 0 : {
344 0 : session_id = id;
345 0 : }
346 0 : }
347 :
348 : // Cancel the current inflight HTTP request if the requets stream is closed.
349 : // This is slightly different to `_cancel_connection` in that
350 : // h2 can cancel individual requests with a `RST_STREAM`.
351 0 : let http_request_token = http_cancellation_token.child_token();
352 0 : let cancel_request = http_request_token.clone().drop_guard();
353 0 :
354 0 : // `request_handler` is not cancel safe. It expects to be cancelled only at specific times.
355 0 : // By spawning the future, we ensure it never gets cancelled until it decides to.
356 0 : let handler = connections.spawn(
357 0 : request_handler(
358 0 : req,
359 0 : config,
360 0 : backend.clone(),
361 0 : connections.clone(),
362 0 : cancellation_handler.clone(),
363 0 : session_id,
364 0 : conn_info2.clone(),
365 0 : http_request_token,
366 0 : endpoint_rate_limiter.clone(),
367 0 : )
368 0 : .in_current_span()
369 0 : .map_ok_or_else(api_error_into_response, |r| r),
370 0 : );
371 0 : async move {
372 0 : let mut res = handler.await;
373 0 : cancel_request.disarm();
374 :
375 : // add the session ID to the response
376 0 : if let Ok(resp) = &mut res {
377 0 : resp.headers_mut()
378 0 : .append(&NEON_REQUEST_ID, uuid_to_header_value(session_id));
379 0 : }
380 :
381 0 : res
382 0 : }
383 0 : }),
384 0 : );
385 :
386 : // On cancellation, trigger the HTTP connection handler to shut down.
387 0 : let res = match select(pin!(cancellation_token.cancelled()), pin!(conn)).await {
388 0 : Either::Left((_cancelled, mut conn)) => {
389 0 : tracing::debug!(%conn_info, "cancelling connection");
390 0 : conn.as_mut().graceful_shutdown();
391 0 : conn.await
392 : }
393 0 : Either::Right((res, _)) => res,
394 : };
395 :
396 0 : match res {
397 0 : Ok(()) => tracing::info!(%conn_info, "HTTP connection closed"),
398 0 : Err(e) => tracing::warn!(%conn_info, "HTTP connection error {e}"),
399 : }
400 0 : }
401 :
402 : #[allow(clippy::too_many_arguments)]
403 0 : async fn request_handler(
404 0 : mut request: hyper::Request<Incoming>,
405 0 : config: &'static ProxyConfig,
406 0 : backend: Arc<PoolingBackend>,
407 0 : ws_connections: TaskTracker,
408 0 : cancellation_handler: Arc<CancellationHandlerMain>,
409 0 : session_id: uuid::Uuid,
410 0 : conn_info: ConnectionInfo,
411 0 : // used to cancel in-flight HTTP requests. not used to cancel websockets
412 0 : http_cancellation_token: CancellationToken,
413 0 : endpoint_rate_limiter: Arc<EndpointRateLimiter>,
414 0 : ) -> Result<Response<BoxBody<Bytes, hyper::Error>>, ApiError> {
415 0 : let host = request
416 0 : .headers()
417 0 : .get("host")
418 0 : .and_then(|h| h.to_str().ok())
419 0 : .and_then(|h| h.split(':').next())
420 0 : .map(|s| s.to_string());
421 0 :
422 0 : // Check if the request is a websocket upgrade request.
423 0 : if config.http_config.accept_websockets
424 0 : && framed_websockets::upgrade::is_upgrade_request(&request)
425 : {
426 0 : let ctx = RequestMonitoring::new(
427 0 : session_id,
428 0 : conn_info,
429 0 : crate::metrics::Protocol::Ws,
430 0 : &config.region,
431 0 : );
432 0 :
433 0 : let span = ctx.span();
434 0 : info!(parent: &span, "performing websocket upgrade");
435 :
436 0 : let (response, websocket) = framed_websockets::upgrade::upgrade(&mut request)
437 0 : .map_err(|e| ApiError::BadRequest(e.into()))?;
438 :
439 0 : ws_connections.spawn(
440 0 : async move {
441 0 : if let Err(e) = websocket::serve_websocket(
442 0 : config,
443 0 : backend.auth_backend,
444 0 : ctx,
445 0 : websocket,
446 0 : cancellation_handler,
447 0 : endpoint_rate_limiter,
448 0 : host,
449 0 : )
450 0 : .await
451 : {
452 0 : warn!("error in websocket connection: {e:#}");
453 0 : }
454 0 : }
455 0 : .instrument(span),
456 0 : );
457 0 :
458 0 : // Return the response so the spawned future can continue.
459 0 : Ok(response.map(|b| b.map_err(|x| match x {}).boxed()))
460 0 : } else if request.uri().path() == "/sql" && *request.method() == Method::POST {
461 0 : let ctx = RequestMonitoring::new(
462 0 : session_id,
463 0 : conn_info,
464 0 : crate::metrics::Protocol::Http,
465 0 : &config.region,
466 0 : );
467 0 : let span = ctx.span();
468 0 :
469 0 : sql_over_http::handle(config, ctx, request, backend, http_cancellation_token)
470 0 : .instrument(span)
471 0 : .await
472 0 : } else if request.uri().path() == "/sql" && *request.method() == Method::OPTIONS {
473 0 : Response::builder()
474 0 : .header("Allow", "OPTIONS, POST")
475 0 : .header("Access-Control-Allow-Origin", "*")
476 0 : .header(
477 0 : "Access-Control-Allow-Headers",
478 0 : "Authorization, Neon-Connection-String, Neon-Raw-Text-Output, Neon-Array-Mode, Neon-Pool-Opt-In, Neon-Batch-Read-Only, Neon-Batch-Isolation-Level",
479 0 : )
480 0 : .header("Access-Control-Max-Age", "86400" /* 24 hours */)
481 0 : .status(StatusCode::OK) // 204 is also valid, but see: https://developer.mozilla.org/en-US/docs/Web/HTTP/Methods/OPTIONS#status_code
482 0 : .body(Empty::new().map_err(|x| match x {}).boxed())
483 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
484 : } else {
485 0 : json_response(StatusCode::BAD_REQUEST, "query is not supported")
486 : }
487 0 : }
|
