Line data Source code
1 : use std::fmt::Display;
2 : use std::net::{IpAddr, Ipv6Addr, SocketAddr};
3 : use std::sync::Arc;
4 : use std::time::Duration;
5 :
6 : use anyhow::Result;
7 : use axum::Router;
8 : use axum::extract::Request;
9 : use axum::middleware::{self, Next};
10 : use axum::response::{IntoResponse, Response};
11 : use axum::routing::{get, post};
12 : use http::StatusCode;
13 : use jsonwebtoken::jwk::JwkSet;
14 : use tokio::net::TcpListener;
15 : use tower::ServiceBuilder;
16 : use tower_http::{
17 : auth::AsyncRequireAuthorizationLayer, request_id::PropagateRequestIdLayer, trace::TraceLayer,
18 : };
19 : use tracing::{Span, error, info};
20 : use uuid::Uuid;
21 :
22 : use super::{
23 : headers::X_REQUEST_ID,
24 : middleware::authorize::Authorize,
25 : routes::{
26 : check_writability, configure, database_schema, dbs_and_roles, extension_server, extensions,
27 : grants, insights, metrics, metrics_json, status, terminate,
28 : },
29 : };
30 : use crate::compute::ComputeNode;
31 :
32 : /// `compute_ctl` has two servers: internal and external. The internal server
33 : /// binds to the loopback interface and handles communication from clients on
34 : /// the compute. The external server is what receives communication from the
35 : /// control plane, the metrics scraper, etc. We make the distinction because
36 : /// certain routes in `compute_ctl` only need to be exposed to local processes
37 : /// like Postgres via the neon extension and local_proxy.
38 : #[derive(Clone, Debug)]
39 : pub enum Server {
40 : Internal {
41 : port: u16,
42 : },
43 : External {
44 : port: u16,
45 : jwks: JwkSet,
46 : compute_id: String,
47 : },
48 : }
49 :
50 : impl Display for Server {
51 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
52 0 : match self {
53 0 : Server::Internal { .. } => f.write_str("internal"),
54 0 : Server::External { .. } => f.write_str("external"),
55 : }
56 0 : }
57 : }
58 :
59 : impl From<&Server> for Router<Arc<ComputeNode>> {
60 0 : fn from(server: &Server) -> Self {
61 0 : let mut router = Router::<Arc<ComputeNode>>::new();
62 :
63 0 : router = match server {
64 : Server::Internal { .. } => {
65 0 : router = router
66 0 : .route(
67 0 : "/extension_server/{*filename}",
68 0 : post(extension_server::download_extension),
69 0 : )
70 0 : .route("/extensions", post(extensions::install_extension))
71 0 : .route("/grants", post(grants::add_grant));
72 0 :
73 0 : // Add in any testing support
74 0 : if cfg!(feature = "testing") {
75 0 : use super::routes::failpoints;
76 0 :
77 0 : router = router.route("/failpoints", post(failpoints::configure_failpoints));
78 0 : }
79 :
80 0 : router
81 : }
82 : Server::External {
83 0 : jwks, compute_id, ..
84 0 : } => {
85 0 : let unauthenticated_router =
86 0 : Router::<Arc<ComputeNode>>::new().route("/metrics", get(metrics::get_metrics));
87 0 :
88 0 : let authenticated_router = Router::<Arc<ComputeNode>>::new()
89 0 : .route("/check_writability", post(check_writability::is_writable))
90 0 : .route("/configure", post(configure::configure))
91 0 : .route("/database_schema", get(database_schema::get_schema_dump))
92 0 : .route("/dbs_and_roles", get(dbs_and_roles::get_catalog_objects))
93 0 : .route("/insights", get(insights::get_insights))
94 0 : .route("/metrics.json", get(metrics_json::get_metrics))
95 0 : .route("/status", get(status::get_status))
96 0 : .route("/terminate", post(terminate::terminate))
97 0 : .layer(AsyncRequireAuthorizationLayer::new(Authorize::new(
98 0 : compute_id.clone(),
99 0 : jwks.clone(),
100 0 : )));
101 0 :
102 0 : router
103 0 : .merge(unauthenticated_router)
104 0 : .merge(authenticated_router)
105 : }
106 : };
107 :
108 0 : router
109 0 : .fallback(Server::handle_404)
110 0 : .method_not_allowed_fallback(Server::handle_405)
111 0 : .layer(
112 0 : ServiceBuilder::new()
113 0 : .layer(tower_otel::trace::HttpLayer::server(tracing::Level::INFO))
114 0 : // Add this middleware since we assume the request ID exists
115 0 : .layer(middleware::from_fn(maybe_add_request_id_header))
116 0 : .layer(
117 0 : TraceLayer::new_for_http()
118 0 : .on_request(|request: &http::Request<_>, _span: &Span| {
119 0 : let request_id = request
120 0 : .headers()
121 0 : .get(X_REQUEST_ID)
122 0 : .unwrap()
123 0 : .to_str()
124 0 : .unwrap();
125 0 :
126 0 : info!(%request_id, "{} {}", request.method(), request.uri());
127 0 : })
128 0 : .on_response(
129 0 : |response: &http::Response<_>, latency: Duration, _span: &Span| {
130 0 : let request_id = response
131 0 : .headers()
132 0 : .get(X_REQUEST_ID)
133 0 : .unwrap()
134 0 : .to_str()
135 0 : .unwrap();
136 0 :
137 0 : info!(
138 : %request_id,
139 0 : code = response.status().as_u16(),
140 0 : latency = latency.as_millis()
141 : );
142 0 : },
143 0 : ),
144 0 : )
145 0 : .layer(PropagateRequestIdLayer::x_request_id()),
146 0 : )
147 0 : }
148 : }
149 :
150 : impl Server {
151 0 : async fn handle_404() -> impl IntoResponse {
152 0 : StatusCode::NOT_FOUND
153 0 : }
154 :
155 0 : async fn handle_405() -> impl IntoResponse {
156 0 : StatusCode::METHOD_NOT_ALLOWED
157 0 : }
158 :
159 0 : async fn listener(&self) -> Result<TcpListener> {
160 0 : let addr = SocketAddr::new(self.ip(), self.port());
161 0 : let listener = TcpListener::bind(&addr).await?;
162 :
163 0 : Ok(listener)
164 0 : }
165 :
166 0 : fn ip(&self) -> IpAddr {
167 0 : match self {
168 : // TODO: Change this to Ipv6Addr::LOCALHOST when the GitHub runners
169 : // allow binding to localhost
170 0 : Server::Internal { .. } => IpAddr::from(Ipv6Addr::UNSPECIFIED),
171 0 : Server::External { .. } => IpAddr::from(Ipv6Addr::UNSPECIFIED),
172 : }
173 0 : }
174 :
175 0 : fn port(&self) -> u16 {
176 0 : match self {
177 0 : Server::Internal { port, .. } => *port,
178 0 : Server::External { port, .. } => *port,
179 : }
180 0 : }
181 :
182 0 : async fn serve(self, compute: Arc<ComputeNode>) {
183 0 : let listener = self.listener().await.unwrap_or_else(|e| {
184 0 : // If we can't bind, the compute cannot operate correctly
185 0 : panic!(
186 0 : "failed to bind the compute_ctl {} HTTP server to {}: {}",
187 0 : self,
188 0 : SocketAddr::new(self.ip(), self.port()),
189 0 : e
190 0 : );
191 0 : });
192 0 :
193 0 : if tracing::enabled!(tracing::Level::INFO) {
194 0 : let local_addr = match listener.local_addr() {
195 0 : Ok(local_addr) => local_addr,
196 0 : Err(_) => SocketAddr::new(self.ip(), self.port()),
197 : };
198 :
199 0 : info!(
200 0 : "compute_ctl {} HTTP server listening at {}",
201 : self, local_addr
202 : );
203 0 : }
204 :
205 0 : let router = Router::from(&self)
206 0 : .with_state(compute)
207 0 : .into_make_service_with_connect_info::<SocketAddr>();
208 :
209 0 : if let Err(e) = axum::serve(listener, router).await {
210 0 : error!("compute_ctl {} HTTP server error: {}", self, e);
211 0 : }
212 0 : }
213 :
214 0 : pub fn launch(self, compute: &Arc<ComputeNode>) {
215 0 : let state = Arc::clone(compute);
216 0 :
217 0 : info!("Launching the {} server", self);
218 :
219 0 : tokio::spawn(self.serve(state));
220 0 : }
221 : }
222 :
223 : /// This middleware function allows compute_ctl to generate its own request ID
224 : /// if one isn't supplied. The control plane will always send one as a UUID. The
225 : /// neon Postgres extension on the other hand does not send one.
226 0 : async fn maybe_add_request_id_header(mut request: Request, next: Next) -> Response {
227 0 : let headers = request.headers_mut();
228 0 : if headers.get(X_REQUEST_ID).is_none() {
229 0 : headers.append(X_REQUEST_ID, Uuid::new_v4().to_string().parse().unwrap());
230 0 : }
231 :
232 0 : next.run(request).await
233 0 : }
|
