Line data Source code
1 : use std::sync::Arc;
2 :
3 : use futures::{FutureExt, TryFutureExt};
4 : use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
5 : use tokio_util::sync::CancellationToken;
6 : use tracing::{debug, error, info, Instrument};
7 :
8 : use crate::auth::backend::ConsoleRedirectBackend;
9 : use crate::cancellation::CancellationHandler;
10 : use crate::config::{ProxyConfig, ProxyProtocolV2};
11 : use crate::context::RequestContext;
12 : use crate::error::ReportableError;
13 : use crate::metrics::{Metrics, NumClientConnectionsGuard};
14 : use crate::protocol2::{read_proxy_protocol, ConnectHeader, ConnectionInfo};
15 : use crate::proxy::connect_compute::{connect_to_compute, TcpMechanism};
16 : use crate::proxy::handshake::{handshake, HandshakeData};
17 : use crate::proxy::passthrough::ProxyPassthrough;
18 : use crate::proxy::{
19 : prepare_client_connection, run_until_cancelled, ClientRequestError, ErrorSource,
20 : };
21 :
22 0 : pub async fn task_main(
23 0 : config: &'static ProxyConfig,
24 0 : backend: &'static ConsoleRedirectBackend,
25 0 : listener: tokio::net::TcpListener,
26 0 : cancellation_token: CancellationToken,
27 0 : cancellation_handler: Arc<CancellationHandler>,
28 0 : ) -> anyhow::Result<()> {
29 0 : scopeguard::defer! {
30 0 : info!("proxy has shut down");
31 0 : }
32 0 :
33 0 : // When set for the server socket, the keepalive setting
34 0 : // will be inherited by all accepted client sockets.
35 0 : socket2::SockRef::from(&listener).set_keepalive(true)?;
36 :
37 0 : let connections = tokio_util::task::task_tracker::TaskTracker::new();
38 0 : let cancellations = tokio_util::task::task_tracker::TaskTracker::new();
39 :
40 0 : while let Some(accept_result) =
41 0 : run_until_cancelled(listener.accept(), &cancellation_token).await
42 : {
43 0 : let (socket, peer_addr) = accept_result?;
44 :
45 0 : let conn_gauge = Metrics::get()
46 0 : .proxy
47 0 : .client_connections
48 0 : .guard(crate::metrics::Protocol::Tcp);
49 0 :
50 0 : let session_id = uuid::Uuid::new_v4();
51 0 : let cancellation_handler = Arc::clone(&cancellation_handler);
52 0 : let cancellations = cancellations.clone();
53 0 :
54 0 : debug!(protocol = "tcp", %session_id, "accepted new TCP connection");
55 :
56 0 : connections.spawn(async move {
57 0 : let (socket, peer_addr) = match read_proxy_protocol(socket).await {
58 0 : Err(e) => {
59 0 : error!("per-client task finished with an error: {e:#}");
60 0 : return;
61 : }
62 : // our load balancers will not send any more data. let's just exit immediately
63 0 : Ok((_socket, ConnectHeader::Local)) => {
64 0 : debug!("healthcheck received");
65 0 : return;
66 : }
67 0 : Ok((_socket, ConnectHeader::Missing)) if config.proxy_protocol_v2 == ProxyProtocolV2::Required => {
68 0 : error!("missing required proxy protocol header");
69 0 : return;
70 : }
71 0 : Ok((_socket, ConnectHeader::Proxy(_))) if config.proxy_protocol_v2 == ProxyProtocolV2::Rejected => {
72 0 : error!("proxy protocol header not supported");
73 0 : return;
74 : }
75 0 : Ok((socket, ConnectHeader::Proxy(info))) => (socket, info),
76 0 : Ok((socket, ConnectHeader::Missing)) => (socket, ConnectionInfo{ addr: peer_addr, extra: None }),
77 : };
78 :
79 0 : match socket.inner.set_nodelay(true) {
80 0 : Ok(()) => {}
81 0 : Err(e) => {
82 0 : error!("per-client task finished with an error: failed to set socket option: {e:#}");
83 0 : return;
84 : }
85 : }
86 :
87 0 : let ctx = RequestContext::new(
88 0 : session_id,
89 0 : peer_addr,
90 0 : crate::metrics::Protocol::Tcp,
91 0 : &config.region,
92 0 : );
93 :
94 0 : let res = handle_client(
95 0 : config,
96 0 : backend,
97 0 : &ctx,
98 0 : cancellation_handler,
99 0 : socket,
100 0 : conn_gauge,
101 0 : cancellations,
102 0 : )
103 0 : .instrument(ctx.span())
104 0 : .boxed()
105 0 : .await;
106 :
107 0 : match res {
108 0 : Err(e) => {
109 0 : ctx.set_error_kind(e.get_error_kind());
110 0 : error!(parent: &ctx.span(), "per-client task finished with an error: {e:#}");
111 : }
112 0 : Ok(None) => {
113 0 : ctx.set_success();
114 0 : }
115 0 : Ok(Some(p)) => {
116 0 : ctx.set_success();
117 0 : let _disconnect = ctx.log_connect();
118 0 : match p.proxy_pass(&config.connect_to_compute).await {
119 0 : Ok(()) => {}
120 0 : Err(ErrorSource::Client(e)) => {
121 0 : error!(?session_id, "per-client task finished with an IO error from the client: {e:#}");
122 : }
123 0 : Err(ErrorSource::Compute(e)) => {
124 0 : error!(?session_id, "per-client task finished with an IO error from the compute: {e:#}");
125 : }
126 : }
127 : }
128 : }
129 0 : });
130 : }
131 :
132 0 : connections.close();
133 0 : cancellations.close();
134 0 : drop(listener);
135 0 :
136 0 : // Drain connections
137 0 : connections.wait().await;
138 0 : cancellations.wait().await;
139 :
140 0 : Ok(())
141 0 : }
142 :
143 : #[allow(clippy::too_many_arguments)]
144 0 : pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
145 0 : config: &'static ProxyConfig,
146 0 : backend: &'static ConsoleRedirectBackend,
147 0 : ctx: &RequestContext,
148 0 : cancellation_handler: Arc<CancellationHandler>,
149 0 : stream: S,
150 0 : conn_gauge: NumClientConnectionsGuard<'static>,
151 0 : cancellations: tokio_util::task::task_tracker::TaskTracker,
152 0 : ) -> Result<Option<ProxyPassthrough<S>>, ClientRequestError> {
153 0 : debug!(
154 0 : protocol = %ctx.protocol(),
155 0 : "handling interactive connection from client"
156 : );
157 :
158 0 : let metrics = &Metrics::get().proxy;
159 0 : let proto = ctx.protocol();
160 0 : let request_gauge = metrics.connection_requests.guard(proto);
161 0 :
162 0 : let tls = config.tls_config.as_ref();
163 0 :
164 0 : let record_handshake_error = !ctx.has_private_peer_addr();
165 0 : let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
166 0 : let do_handshake = handshake(ctx, stream, tls, record_handshake_error);
167 :
168 0 : let (mut stream, params) = match tokio::time::timeout(config.handshake_timeout, do_handshake)
169 0 : .await??
170 : {
171 0 : HandshakeData::Startup(stream, params) => (stream, params),
172 0 : HandshakeData::Cancel(cancel_key_data) => {
173 0 : // spawn a task to cancel the session, but don't wait for it
174 0 : cancellations.spawn({
175 0 : let cancellation_handler_clone = Arc::clone(&cancellation_handler);
176 0 : let ctx = ctx.clone();
177 0 : let cancel_span = tracing::span!(parent: None, tracing::Level::INFO, "cancel_session", session_id = ?ctx.session_id());
178 0 : cancel_span.follows_from(tracing::Span::current());
179 0 : async move {
180 0 : cancellation_handler_clone
181 0 : .cancel_session(
182 0 : cancel_key_data,
183 0 : ctx,
184 0 : config.authentication_config.ip_allowlist_check_enabled,
185 0 : config.authentication_config.is_vpc_acccess_proxy,
186 0 : backend.get_api(),
187 0 : )
188 0 : .await
189 0 : .inspect_err(|e | debug!(error = ?e, "cancel_session failed")).ok();
190 0 : }.instrument(cancel_span)
191 0 : });
192 0 :
193 0 : return Ok(None);
194 : }
195 : };
196 0 : drop(pause);
197 0 :
198 0 : ctx.set_db_options(params.clone());
199 :
200 0 : let (node_info, user_info, _ip_allowlist) = match backend
201 0 : .authenticate(ctx, &config.authentication_config, &mut stream)
202 0 : .await
203 : {
204 0 : Ok(auth_result) => auth_result,
205 0 : Err(e) => {
206 0 : return stream.throw_error(e).await?;
207 : }
208 : };
209 :
210 0 : let mut node = connect_to_compute(
211 0 : ctx,
212 0 : &TcpMechanism {
213 0 : user_info,
214 0 : params_compat: true,
215 0 : params: ¶ms,
216 0 : locks: &config.connect_compute_locks,
217 0 : },
218 0 : &node_info,
219 0 : config.wake_compute_retry_config,
220 0 : &config.connect_to_compute,
221 0 : )
222 0 : .or_else(|e| stream.throw_error(e))
223 0 : .await?;
224 :
225 0 : let cancellation_handler_clone = Arc::clone(&cancellation_handler);
226 0 : let session = cancellation_handler_clone.get_key();
227 0 :
228 0 : session
229 0 : .write_cancel_key(node.cancel_closure.clone())
230 0 : .await?;
231 :
232 0 : prepare_client_connection(&node, *session.key(), &mut stream).await?;
233 :
234 : // Before proxy passing, forward to compute whatever data is left in the
235 : // PqStream input buffer. Normally there is none, but our serverless npm
236 : // driver in pipeline mode sends startup, password and first query
237 : // immediately after opening the connection.
238 0 : let (stream, read_buf) = stream.into_inner();
239 0 : node.stream.write_all(&read_buf).await?;
240 :
241 0 : Ok(Some(ProxyPassthrough {
242 0 : client: stream,
243 0 : aux: node.aux.clone(),
244 0 : compute: node,
245 0 : session_id: ctx.session_id(),
246 0 : cancel: session,
247 0 : _req: request_gauge,
248 0 : _conn: conn_gauge,
249 0 : }))
250 0 : }
|