Line data Source code
1 : //! Production console backend.
2 :
3 : use std::sync::Arc;
4 : use std::time::Duration;
5 :
6 : use ::http::header::AUTHORIZATION;
7 : use ::http::HeaderName;
8 : use futures::TryFutureExt;
9 : use tokio::time::Instant;
10 : use tokio_postgres::config::SslMode;
11 : use tracing::{debug, info, info_span, warn, Instrument};
12 :
13 : use super::super::messages::{ControlPlaneErrorMessage, GetRoleSecret, WakeCompute};
14 : use crate::auth::backend::jwt::AuthRule;
15 : use crate::auth::backend::ComputeUserInfo;
16 : use crate::cache::Cached;
17 : use crate::context::RequestMonitoring;
18 : use crate::control_plane::caches::ApiCaches;
19 : use crate::control_plane::errors::{
20 : ControlPlaneError, GetAuthInfoError, GetEndpointJwksError, WakeComputeError,
21 : };
22 : use crate::control_plane::locks::ApiLocks;
23 : use crate::control_plane::messages::{ColdStartInfo, EndpointJwksResponse, Reason};
24 : use crate::control_plane::{
25 : AuthInfo, AuthSecret, CachedAllowedIps, CachedNodeInfo, CachedRoleSecret, NodeInfo,
26 : };
27 : use crate::metrics::{CacheOutcome, Metrics};
28 : use crate::rate_limiter::WakeComputeRateLimiter;
29 : use crate::types::{EndpointCacheKey, EndpointId};
30 : use crate::{compute, http, scram};
31 :
32 : const X_REQUEST_ID: HeaderName = HeaderName::from_static("x-request-id");
33 :
34 : #[derive(Clone)]
35 : pub struct NeonControlPlaneClient {
36 : endpoint: http::Endpoint,
37 : pub caches: &'static ApiCaches,
38 : pub(crate) locks: &'static ApiLocks<EndpointCacheKey>,
39 : pub(crate) wake_compute_endpoint_rate_limiter: Arc<WakeComputeRateLimiter>,
40 : // put in a shared ref so we don't copy secrets all over in memory
41 : jwt: Arc<str>,
42 : }
43 :
44 : impl NeonControlPlaneClient {
45 : /// Construct an API object containing the auth parameters.
46 0 : pub fn new(
47 0 : endpoint: http::Endpoint,
48 0 : jwt: Arc<str>,
49 0 : caches: &'static ApiCaches,
50 0 : locks: &'static ApiLocks<EndpointCacheKey>,
51 0 : wake_compute_endpoint_rate_limiter: Arc<WakeComputeRateLimiter>,
52 0 : ) -> Self {
53 0 : Self {
54 0 : endpoint,
55 0 : caches,
56 0 : locks,
57 0 : wake_compute_endpoint_rate_limiter,
58 0 : jwt,
59 0 : }
60 0 : }
61 :
62 0 : pub(crate) fn url(&self) -> &str {
63 0 : self.endpoint.url().as_str()
64 0 : }
65 :
66 0 : async fn do_get_auth_info(
67 0 : &self,
68 0 : ctx: &RequestMonitoring,
69 0 : user_info: &ComputeUserInfo,
70 0 : ) -> Result<AuthInfo, GetAuthInfoError> {
71 0 : if !self
72 0 : .caches
73 0 : .endpoints_cache
74 0 : .is_valid(ctx, &user_info.endpoint.normalize())
75 : {
76 0 : info!("endpoint is not valid, skipping the request");
77 0 : return Ok(AuthInfo::default());
78 0 : }
79 0 : let request_id = ctx.session_id().to_string();
80 0 : let application_name = ctx.console_application_name();
81 0 : async {
82 0 : let request = self
83 0 : .endpoint
84 0 : .get_path("proxy_get_role_secret")
85 0 : .header(X_REQUEST_ID, &request_id)
86 0 : .header(AUTHORIZATION, format!("Bearer {}", &self.jwt))
87 0 : .query(&[("session_id", ctx.session_id())])
88 0 : .query(&[
89 0 : ("application_name", application_name.as_str()),
90 0 : ("project", user_info.endpoint.as_str()),
91 0 : ("role", user_info.user.as_str()),
92 0 : ])
93 0 : .build()?;
94 :
95 0 : info!(url = request.url().as_str(), "sending http request");
96 0 : let start = Instant::now();
97 0 : let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Cplane);
98 0 : let response = self.endpoint.execute(request).await?;
99 0 : drop(pause);
100 0 : info!(duration = ?start.elapsed(), "received http response");
101 0 : let body = match parse_body::<GetRoleSecret>(response).await {
102 0 : Ok(body) => body,
103 : // Error 404 is special: it's ok not to have a secret.
104 : // TODO(anna): retry
105 0 : Err(e) => {
106 0 : return if e.get_reason().is_not_found() {
107 0 : Ok(AuthInfo::default())
108 : } else {
109 0 : Err(e.into())
110 : }
111 : }
112 : };
113 :
114 0 : let secret = if body.role_secret.is_empty() {
115 0 : None
116 : } else {
117 0 : let secret = scram::ServerSecret::parse(&body.role_secret)
118 0 : .map(AuthSecret::Scram)
119 0 : .ok_or(GetAuthInfoError::BadSecret)?;
120 0 : Some(secret)
121 : };
122 0 : let allowed_ips = body.allowed_ips.unwrap_or_default();
123 0 : Metrics::get()
124 0 : .proxy
125 0 : .allowed_ips_number
126 0 : .observe(allowed_ips.len() as f64);
127 0 : Ok(AuthInfo {
128 0 : secret,
129 0 : allowed_ips,
130 0 : project_id: body.project_id,
131 0 : })
132 0 : }
133 0 : .map_err(crate::error::log_error)
134 0 : .instrument(info_span!("http", id = request_id))
135 0 : .await
136 0 : }
137 :
138 0 : async fn do_get_endpoint_jwks(
139 0 : &self,
140 0 : ctx: &RequestMonitoring,
141 0 : endpoint: EndpointId,
142 0 : ) -> Result<Vec<AuthRule>, GetEndpointJwksError> {
143 0 : if !self
144 0 : .caches
145 0 : .endpoints_cache
146 0 : .is_valid(ctx, &endpoint.normalize())
147 : {
148 0 : return Err(GetEndpointJwksError::EndpointNotFound);
149 0 : }
150 0 : let request_id = ctx.session_id().to_string();
151 0 : async {
152 0 : let request = self
153 0 : .endpoint
154 0 : .get_with_url(|url| {
155 0 : url.path_segments_mut()
156 0 : .push("endpoints")
157 0 : .push(endpoint.as_str())
158 0 : .push("jwks");
159 0 : })
160 0 : .header(X_REQUEST_ID, &request_id)
161 0 : .header(AUTHORIZATION, format!("Bearer {}", &self.jwt))
162 0 : .query(&[("session_id", ctx.session_id())])
163 0 : .build()
164 0 : .map_err(GetEndpointJwksError::RequestBuild)?;
165 :
166 0 : info!(url = request.url().as_str(), "sending http request");
167 0 : let start = Instant::now();
168 0 : let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Cplane);
169 0 : let response = self
170 0 : .endpoint
171 0 : .execute(request)
172 0 : .await
173 0 : .map_err(GetEndpointJwksError::RequestExecute)?;
174 0 : drop(pause);
175 0 : info!(duration = ?start.elapsed(), "received http response");
176 :
177 0 : let body = parse_body::<EndpointJwksResponse>(response).await?;
178 :
179 0 : let rules = body
180 0 : .jwks
181 0 : .into_iter()
182 0 : .map(|jwks| AuthRule {
183 0 : id: jwks.id,
184 0 : jwks_url: jwks.jwks_url,
185 0 : audience: jwks.jwt_audience,
186 0 : role_names: jwks.role_names,
187 0 : })
188 0 : .collect();
189 0 :
190 0 : Ok(rules)
191 0 : }
192 0 : .map_err(crate::error::log_error)
193 0 : .instrument(info_span!("http", id = request_id))
194 0 : .await
195 0 : }
196 :
197 0 : async fn do_wake_compute(
198 0 : &self,
199 0 : ctx: &RequestMonitoring,
200 0 : user_info: &ComputeUserInfo,
201 0 : ) -> Result<NodeInfo, WakeComputeError> {
202 0 : let request_id = ctx.session_id().to_string();
203 0 : let application_name = ctx.console_application_name();
204 0 : async {
205 0 : let mut request_builder = self
206 0 : .endpoint
207 0 : .get_path("proxy_wake_compute")
208 0 : .header("X-Request-ID", &request_id)
209 0 : .header("Authorization", format!("Bearer {}", &self.jwt))
210 0 : .query(&[("session_id", ctx.session_id())])
211 0 : .query(&[
212 0 : ("application_name", application_name.as_str()),
213 0 : ("project", user_info.endpoint.as_str()),
214 0 : ]);
215 0 :
216 0 : let options = user_info.options.to_deep_object();
217 0 : if !options.is_empty() {
218 0 : request_builder = request_builder.query(&options);
219 0 : }
220 :
221 0 : let request = request_builder.build()?;
222 :
223 0 : info!(url = request.url().as_str(), "sending http request");
224 0 : let start = Instant::now();
225 0 : let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Cplane);
226 0 : let response = self.endpoint.execute(request).await?;
227 0 : drop(pause);
228 0 : info!(duration = ?start.elapsed(), "received http response");
229 0 : let body = parse_body::<WakeCompute>(response).await?;
230 :
231 : // Unfortunately, ownership won't let us use `Option::ok_or` here.
232 0 : let (host, port) = match parse_host_port(&body.address) {
233 0 : None => return Err(WakeComputeError::BadComputeAddress(body.address)),
234 0 : Some(x) => x,
235 0 : };
236 0 :
237 0 : // Don't set anything but host and port! This config will be cached.
238 0 : // We'll set username and such later using the startup message.
239 0 : // TODO: add more type safety (in progress).
240 0 : let mut config = compute::ConnCfg::new();
241 0 : config.host(host).port(port).ssl_mode(SslMode::Disable); // TLS is not configured on compute nodes.
242 0 :
243 0 : let node = NodeInfo {
244 0 : config,
245 0 : aux: body.aux,
246 0 : allow_self_signed_compute: false,
247 0 : };
248 0 :
249 0 : Ok(node)
250 0 : }
251 0 : .map_err(crate::error::log_error)
252 0 : .instrument(info_span!("http", id = request_id))
253 0 : .await
254 0 : }
255 : }
256 :
257 : impl super::ControlPlaneApi for NeonControlPlaneClient {
258 0 : #[tracing::instrument(skip_all)]
259 : async fn get_role_secret(
260 : &self,
261 : ctx: &RequestMonitoring,
262 : user_info: &ComputeUserInfo,
263 : ) -> Result<CachedRoleSecret, GetAuthInfoError> {
264 : let normalized_ep = &user_info.endpoint.normalize();
265 : let user = &user_info.user;
266 : if let Some(role_secret) = self
267 : .caches
268 : .project_info
269 : .get_role_secret(normalized_ep, user)
270 : {
271 : return Ok(role_secret);
272 : }
273 : let auth_info = self.do_get_auth_info(ctx, user_info).await?;
274 : if let Some(project_id) = auth_info.project_id {
275 : let normalized_ep_int = normalized_ep.into();
276 : self.caches.project_info.insert_role_secret(
277 : project_id,
278 : normalized_ep_int,
279 : user.into(),
280 : auth_info.secret.clone(),
281 : );
282 : self.caches.project_info.insert_allowed_ips(
283 : project_id,
284 : normalized_ep_int,
285 : Arc::new(auth_info.allowed_ips),
286 : );
287 : ctx.set_project_id(project_id);
288 : }
289 : // When we just got a secret, we don't need to invalidate it.
290 : Ok(Cached::new_uncached(auth_info.secret))
291 : }
292 :
293 0 : async fn get_allowed_ips_and_secret(
294 0 : &self,
295 0 : ctx: &RequestMonitoring,
296 0 : user_info: &ComputeUserInfo,
297 0 : ) -> Result<(CachedAllowedIps, Option<CachedRoleSecret>), GetAuthInfoError> {
298 0 : let normalized_ep = &user_info.endpoint.normalize();
299 0 : if let Some(allowed_ips) = self.caches.project_info.get_allowed_ips(normalized_ep) {
300 0 : Metrics::get()
301 0 : .proxy
302 0 : .allowed_ips_cache_misses
303 0 : .inc(CacheOutcome::Hit);
304 0 : return Ok((allowed_ips, None));
305 0 : }
306 0 : Metrics::get()
307 0 : .proxy
308 0 : .allowed_ips_cache_misses
309 0 : .inc(CacheOutcome::Miss);
310 0 : let auth_info = self.do_get_auth_info(ctx, user_info).await?;
311 0 : let allowed_ips = Arc::new(auth_info.allowed_ips);
312 0 : let user = &user_info.user;
313 0 : if let Some(project_id) = auth_info.project_id {
314 0 : let normalized_ep_int = normalized_ep.into();
315 0 : self.caches.project_info.insert_role_secret(
316 0 : project_id,
317 0 : normalized_ep_int,
318 0 : user.into(),
319 0 : auth_info.secret.clone(),
320 0 : );
321 0 : self.caches.project_info.insert_allowed_ips(
322 0 : project_id,
323 0 : normalized_ep_int,
324 0 : allowed_ips.clone(),
325 0 : );
326 0 : ctx.set_project_id(project_id);
327 0 : }
328 0 : Ok((
329 0 : Cached::new_uncached(allowed_ips),
330 0 : Some(Cached::new_uncached(auth_info.secret)),
331 0 : ))
332 0 : }
333 :
334 0 : #[tracing::instrument(skip_all)]
335 : async fn get_endpoint_jwks(
336 : &self,
337 : ctx: &RequestMonitoring,
338 : endpoint: EndpointId,
339 : ) -> Result<Vec<AuthRule>, GetEndpointJwksError> {
340 : self.do_get_endpoint_jwks(ctx, endpoint).await
341 : }
342 :
343 0 : #[tracing::instrument(skip_all)]
344 : async fn wake_compute(
345 : &self,
346 : ctx: &RequestMonitoring,
347 : user_info: &ComputeUserInfo,
348 : ) -> Result<CachedNodeInfo, WakeComputeError> {
349 : let key = user_info.endpoint_cache_key();
350 :
351 : macro_rules! check_cache {
352 : () => {
353 : if let Some(cached) = self.caches.node_info.get(&key) {
354 : let (cached, info) = cached.take_value();
355 0 : let info = info.map_err(|c| {
356 0 : info!(key = &*key, "found cached wake_compute error");
357 0 : WakeComputeError::ControlPlane(ControlPlaneError::Message(Box::new(*c)))
358 0 : })?;
359 :
360 : debug!(key = &*key, "found cached compute node info");
361 : ctx.set_project(info.aux.clone());
362 0 : return Ok(cached.map(|()| info));
363 : }
364 : };
365 : }
366 :
367 : // Every time we do a wakeup http request, the compute node will stay up
368 : // for some time (highly depends on the console's scale-to-zero policy);
369 : // The connection info remains the same during that period of time,
370 : // which means that we might cache it to reduce the load and latency.
371 : check_cache!();
372 :
373 : let permit = self.locks.get_permit(&key).await?;
374 :
375 : // after getting back a permit - it's possible the cache was filled
376 : // double check
377 : if permit.should_check_cache() {
378 : check_cache!();
379 : }
380 :
381 : // check rate limit
382 : if !self
383 : .wake_compute_endpoint_rate_limiter
384 : .check(user_info.endpoint.normalize_intern(), 1)
385 : {
386 : return Err(WakeComputeError::TooManyConnections);
387 : }
388 :
389 : let node = permit.release_result(self.do_wake_compute(ctx, user_info).await);
390 : match node {
391 : Ok(node) => {
392 : ctx.set_project(node.aux.clone());
393 : debug!(key = &*key, "created a cache entry for woken compute node");
394 :
395 : let mut stored_node = node.clone();
396 : // store the cached node as 'warm_cached'
397 : stored_node.aux.cold_start_info = ColdStartInfo::WarmCached;
398 :
399 : let (_, cached) = self.caches.node_info.insert_unit(key, Ok(stored_node));
400 :
401 0 : Ok(cached.map(|()| node))
402 : }
403 : Err(err) => match err {
404 : WakeComputeError::ControlPlane(ControlPlaneError::Message(err)) => {
405 : let Some(status) = &err.status else {
406 : return Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
407 : err,
408 : )));
409 : };
410 :
411 : let reason = status
412 : .details
413 : .error_info
414 0 : .map_or(Reason::Unknown, |x| x.reason);
415 :
416 : // if we can retry this error, do not cache it.
417 : if reason.can_retry() {
418 : return Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
419 : err,
420 : )));
421 : }
422 :
423 : // at this point, we should only have quota errors.
424 : debug!(
425 : key = &*key,
426 : "created a cache entry for the wake compute error"
427 : );
428 :
429 : self.caches.node_info.insert_ttl(
430 : key,
431 : Err(err.clone()),
432 : Duration::from_secs(30),
433 : );
434 :
435 : Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
436 : err,
437 : )))
438 : }
439 : err => return Err(err),
440 : },
441 : }
442 : }
443 : }
444 :
445 : /// Parse http response body, taking status code into account.
446 0 : async fn parse_body<T: for<'a> serde::Deserialize<'a>>(
447 0 : response: http::Response,
448 0 : ) -> Result<T, ControlPlaneError> {
449 0 : let status = response.status();
450 0 : if status.is_success() {
451 : // We shouldn't log raw body because it may contain secrets.
452 0 : info!("request succeeded, processing the body");
453 0 : return Ok(response.json().await?);
454 0 : }
455 0 : let s = response.bytes().await?;
456 : // Log plaintext to be able to detect, whether there are some cases not covered by the error struct.
457 0 : info!("response_error plaintext: {:?}", s);
458 :
459 : // Don't throw an error here because it's not as important
460 : // as the fact that the request itself has failed.
461 0 : let mut body = serde_json::from_slice(&s).unwrap_or_else(|e| {
462 0 : warn!("failed to parse error body: {e}");
463 0 : ControlPlaneErrorMessage {
464 0 : error: "reason unclear (malformed error message)".into(),
465 0 : http_status_code: status,
466 0 : status: None,
467 0 : }
468 0 : });
469 0 : body.http_status_code = status;
470 0 :
471 0 : warn!("console responded with an error ({status}): {body:?}");
472 0 : Err(ControlPlaneError::Message(Box::new(body)))
473 0 : }
474 :
475 3 : fn parse_host_port(input: &str) -> Option<(&str, u16)> {
476 3 : let (host, port) = input.rsplit_once(':')?;
477 3 : let ipv6_brackets: &[_] = &['[', ']'];
478 3 : Some((host.trim_matches(ipv6_brackets), port.parse().ok()?))
479 3 : }
480 :
481 : #[cfg(test)]
482 : mod tests {
483 : use super::*;
484 :
485 : #[test]
486 1 : fn test_parse_host_port_v4() {
487 1 : let (host, port) = parse_host_port("127.0.0.1:5432").expect("failed to parse");
488 1 : assert_eq!(host, "127.0.0.1");
489 1 : assert_eq!(port, 5432);
490 1 : }
491 :
492 : #[test]
493 1 : fn test_parse_host_port_v6() {
494 1 : let (host, port) = parse_host_port("[2001:db8::1]:5432").expect("failed to parse");
495 1 : assert_eq!(host, "2001:db8::1");
496 1 : assert_eq!(port, 5432);
497 1 : }
498 :
499 : #[test]
500 1 : fn test_parse_host_port_url() {
501 1 : let (host, port) = parse_host_port("compute-foo-bar-1234.default.svc.cluster.local:5432")
502 1 : .expect("failed to parse");
503 1 : assert_eq!(host, "compute-foo-bar-1234.default.svc.cluster.local");
504 1 : assert_eq!(port, 5432);
505 1 : }
506 : }
|
