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