Line data Source code
1 : use std::ffi::OsStr;
2 : use std::fs;
3 : use std::path::PathBuf;
4 : use std::process::ExitStatus;
5 : use std::str::FromStr;
6 : use std::sync::OnceLock;
7 : use std::time::{Duration, Instant};
8 :
9 : use camino::{Utf8Path, Utf8PathBuf};
10 : use hyper0::Uri;
11 : use nix::unistd::Pid;
12 : use pageserver_api::controller_api::{
13 : NodeConfigureRequest, NodeDescribeResponse, NodeRegisterRequest, TenantCreateRequest,
14 : TenantCreateResponse, TenantLocateResponse,
15 : };
16 : use pageserver_api::models::{TenantConfigRequest, TimelineCreateRequest, TimelineInfo};
17 : use pageserver_api::shard::TenantShardId;
18 : use pageserver_client::mgmt_api::ResponseErrorMessageExt;
19 : use postgres_backend::AuthType;
20 : use reqwest::Method;
21 : use serde::de::DeserializeOwned;
22 : use serde::{Deserialize, Serialize};
23 : use tokio::process::Command;
24 : use tracing::instrument;
25 : use url::Url;
26 : use utils::auth::{Claims, Scope, encode_from_key_file};
27 : use utils::id::{NodeId, TenantId};
28 : use whoami::username;
29 :
30 : use crate::background_process;
31 : use crate::local_env::{LocalEnv, NeonStorageControllerConf};
32 :
33 : pub struct StorageController {
34 : env: LocalEnv,
35 : private_key: Option<Vec<u8>>,
36 : public_key: Option<String>,
37 : client: reqwest::Client,
38 : config: NeonStorageControllerConf,
39 :
40 : // The listen port is learned when starting the storage controller,
41 : // hence the use of OnceLock to init it at the right time.
42 : listen_port: OnceLock<u16>,
43 : }
44 :
45 : const COMMAND: &str = "storage_controller";
46 :
47 : const STORAGE_CONTROLLER_POSTGRES_VERSION: u32 = 16;
48 :
49 : const DB_NAME: &str = "storage_controller";
50 :
51 : pub struct NeonStorageControllerStartArgs {
52 : pub instance_id: u8,
53 : pub base_port: Option<u16>,
54 : pub start_timeout: humantime::Duration,
55 : }
56 :
57 : impl NeonStorageControllerStartArgs {
58 0 : pub fn with_default_instance_id(start_timeout: humantime::Duration) -> Self {
59 0 : Self {
60 0 : instance_id: 1,
61 0 : base_port: None,
62 0 : start_timeout,
63 0 : }
64 0 : }
65 : }
66 :
67 : pub struct NeonStorageControllerStopArgs {
68 : pub instance_id: u8,
69 : pub immediate: bool,
70 : }
71 :
72 : impl NeonStorageControllerStopArgs {
73 0 : pub fn with_default_instance_id(immediate: bool) -> Self {
74 0 : Self {
75 0 : instance_id: 1,
76 0 : immediate,
77 0 : }
78 0 : }
79 : }
80 :
81 0 : #[derive(Serialize, Deserialize)]
82 : pub struct AttachHookRequest {
83 : pub tenant_shard_id: TenantShardId,
84 : pub node_id: Option<NodeId>,
85 : pub generation_override: Option<i32>,
86 : }
87 :
88 0 : #[derive(Serialize, Deserialize)]
89 : pub struct AttachHookResponse {
90 : #[serde(rename = "gen")]
91 : pub generation: Option<u32>,
92 : }
93 :
94 0 : #[derive(Serialize, Deserialize)]
95 : pub struct InspectRequest {
96 : pub tenant_shard_id: TenantShardId,
97 : }
98 :
99 0 : #[derive(Serialize, Deserialize)]
100 : pub struct InspectResponse {
101 : pub attachment: Option<(u32, NodeId)>,
102 : }
103 :
104 : impl StorageController {
105 0 : pub fn from_env(env: &LocalEnv) -> Self {
106 0 : // Assume all pageservers have symmetric auth configuration: this service
107 0 : // expects to use one JWT token to talk to all of them.
108 0 : let ps_conf = env
109 0 : .pageservers
110 0 : .first()
111 0 : .expect("Config is validated to contain at least one pageserver");
112 0 : let (private_key, public_key) = match ps_conf.http_auth_type {
113 0 : AuthType::Trust => (None, None),
114 : AuthType::NeonJWT => {
115 0 : let private_key_path = env.get_private_key_path();
116 0 : let private_key = fs::read(private_key_path).expect("failed to read private key");
117 0 :
118 0 : // If pageserver auth is enabled, this implicitly enables auth for this service,
119 0 : // using the same credentials.
120 0 : let public_key_path =
121 0 : camino::Utf8PathBuf::try_from(env.base_data_dir.join("auth_public_key.pem"))
122 0 : .unwrap();
123 :
124 : // This service takes keys as a string rather than as a path to a file/dir: read the key into memory.
125 0 : let public_key = if std::fs::metadata(&public_key_path)
126 0 : .expect("Can't stat public key")
127 0 : .is_dir()
128 : {
129 : // Our config may specify a directory: this is for the pageserver's ability to handle multiple
130 : // keys. We only use one key at a time, so, arbitrarily load the first one in the directory.
131 0 : let mut dir =
132 0 : std::fs::read_dir(&public_key_path).expect("Can't readdir public key path");
133 0 : let dent = dir
134 0 : .next()
135 0 : .expect("Empty key dir")
136 0 : .expect("Error reading key dir");
137 0 :
138 0 : std::fs::read_to_string(dent.path()).expect("Can't read public key")
139 : } else {
140 0 : std::fs::read_to_string(&public_key_path).expect("Can't read public key")
141 : };
142 0 : (Some(private_key), Some(public_key))
143 : }
144 : };
145 :
146 0 : let mut http_client = reqwest::Client::builder();
147 0 : if let Some(ssl_ca_file) = env.ssl_ca_cert_path() {
148 0 : let buf = std::fs::read(ssl_ca_file).expect("SSL CA file should exist");
149 0 : let cert = reqwest::Certificate::from_pem(&buf).expect("SSL CA file should be valid");
150 0 : http_client = http_client.add_root_certificate(cert);
151 0 : }
152 0 : let http_client = http_client
153 0 : .build()
154 0 : .expect("HTTP client should construct with no error");
155 0 :
156 0 : Self {
157 0 : env: env.clone(),
158 0 : private_key,
159 0 : public_key,
160 0 : client: http_client,
161 0 : config: env.storage_controller.clone(),
162 0 : listen_port: OnceLock::default(),
163 0 : }
164 0 : }
165 :
166 0 : fn storage_controller_instance_dir(&self, instance_id: u8) -> PathBuf {
167 0 : self.env
168 0 : .base_data_dir
169 0 : .join(format!("storage_controller_{}", instance_id))
170 0 : }
171 :
172 0 : fn pid_file(&self, instance_id: u8) -> Utf8PathBuf {
173 0 : Utf8PathBuf::from_path_buf(
174 0 : self.storage_controller_instance_dir(instance_id)
175 0 : .join("storage_controller.pid"),
176 0 : )
177 0 : .expect("non-Unicode path")
178 0 : }
179 :
180 : /// Find the directory containing postgres subdirectories, such `bin` and `lib`
181 : ///
182 : /// This usually uses STORAGE_CONTROLLER_POSTGRES_VERSION of postgres, but will fall back
183 : /// to other versions if that one isn't found. Some automated tests create circumstances
184 : /// where only one version is available in pg_distrib_dir, such as `test_remote_extensions`.
185 0 : async fn get_pg_dir(&self, dir_name: &str) -> anyhow::Result<Utf8PathBuf> {
186 0 : let prefer_versions = [STORAGE_CONTROLLER_POSTGRES_VERSION, 16, 15, 14];
187 :
188 0 : for v in prefer_versions {
189 0 : let path = Utf8PathBuf::from_path_buf(self.env.pg_dir(v, dir_name)?).unwrap();
190 0 : if tokio::fs::try_exists(&path).await? {
191 0 : return Ok(path);
192 0 : }
193 : }
194 :
195 : // Fall through
196 0 : anyhow::bail!(
197 0 : "Postgres directory '{}' not found in {}",
198 0 : dir_name,
199 0 : self.env.pg_distrib_dir.display(),
200 0 : );
201 0 : }
202 :
203 0 : pub async fn get_pg_bin_dir(&self) -> anyhow::Result<Utf8PathBuf> {
204 0 : self.get_pg_dir("bin").await
205 0 : }
206 :
207 0 : pub async fn get_pg_lib_dir(&self) -> anyhow::Result<Utf8PathBuf> {
208 0 : self.get_pg_dir("lib").await
209 0 : }
210 :
211 : /// Readiness check for our postgres process
212 0 : async fn pg_isready(&self, pg_bin_dir: &Utf8Path, postgres_port: u16) -> anyhow::Result<bool> {
213 0 : let bin_path = pg_bin_dir.join("pg_isready");
214 0 : let args = [
215 0 : "-h",
216 0 : "localhost",
217 0 : "-U",
218 0 : &username(),
219 0 : "-d",
220 0 : DB_NAME,
221 0 : "-p",
222 0 : &format!("{}", postgres_port),
223 0 : ];
224 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
225 0 : let envs = [
226 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
227 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
228 0 : ];
229 0 : let exitcode = Command::new(bin_path)
230 0 : .args(args)
231 0 : .envs(envs)
232 0 : .spawn()?
233 0 : .wait()
234 0 : .await?;
235 :
236 0 : Ok(exitcode.success())
237 0 : }
238 :
239 : /// Create our database if it doesn't exist
240 : ///
241 : /// This function is equivalent to the `diesel setup` command in the diesel CLI. We implement
242 : /// the same steps by hand to avoid imposing a dependency on installing diesel-cli for developers
243 : /// who just want to run `cargo neon_local` without knowing about diesel.
244 : ///
245 : /// Returns the database url
246 0 : pub async fn setup_database(&self, postgres_port: u16) -> anyhow::Result<String> {
247 0 : let database_url = format!(
248 0 : "postgresql://{}@localhost:{}/{DB_NAME}",
249 0 : &username(),
250 0 : postgres_port
251 0 : );
252 :
253 0 : let pg_bin_dir = self.get_pg_bin_dir().await?;
254 0 : let createdb_path = pg_bin_dir.join("createdb");
255 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
256 0 : let envs = [
257 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
258 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
259 0 : ];
260 0 : let output = Command::new(&createdb_path)
261 0 : .args([
262 0 : "-h",
263 0 : "localhost",
264 0 : "-p",
265 0 : &format!("{}", postgres_port),
266 0 : "-U",
267 0 : &username(),
268 0 : "-O",
269 0 : &username(),
270 0 : DB_NAME,
271 0 : ])
272 0 : .envs(envs)
273 0 : .output()
274 0 : .await
275 0 : .expect("Failed to spawn createdb");
276 0 :
277 0 : if !output.status.success() {
278 0 : let stderr = String::from_utf8(output.stderr).expect("Non-UTF8 output from createdb");
279 0 : if stderr.contains("already exists") {
280 0 : tracing::info!("Database {DB_NAME} already exists");
281 : } else {
282 0 : anyhow::bail!("createdb failed with status {}: {stderr}", output.status);
283 : }
284 0 : }
285 :
286 0 : Ok(database_url)
287 0 : }
288 :
289 0 : pub async fn connect_to_database(
290 0 : &self,
291 0 : postgres_port: u16,
292 0 : ) -> anyhow::Result<(
293 0 : tokio_postgres::Client,
294 0 : tokio_postgres::Connection<tokio_postgres::Socket, tokio_postgres::tls::NoTlsStream>,
295 0 : )> {
296 0 : tokio_postgres::Config::new()
297 0 : .host("localhost")
298 0 : .port(postgres_port)
299 0 : // The user is the ambient operating system user name.
300 0 : // That is an impurity which we want to fix in => TODO https://github.com/neondatabase/neon/issues/8400
301 0 : //
302 0 : // Until we get there, use the ambient operating system user name.
303 0 : // Recent tokio-postgres versions default to this if the user isn't specified.
304 0 : // But tokio-postgres fork doesn't have this upstream commit:
305 0 : // https://github.com/sfackler/rust-postgres/commit/cb609be758f3fb5af537f04b584a2ee0cebd5e79
306 0 : // => we should rebase our fork => TODO https://github.com/neondatabase/neon/issues/8399
307 0 : .user(&username())
308 0 : .dbname(DB_NAME)
309 0 : .connect(tokio_postgres::NoTls)
310 0 : .await
311 0 : .map_err(anyhow::Error::new)
312 0 : }
313 :
314 : /// Wrapper for the pg_ctl binary, which we spawn as a short-lived subprocess when starting and stopping postgres
315 0 : async fn pg_ctl<I, S>(&self, args: I) -> ExitStatus
316 0 : where
317 0 : I: IntoIterator<Item = S>,
318 0 : S: AsRef<OsStr>,
319 0 : {
320 0 : let pg_bin_dir = self.get_pg_bin_dir().await.unwrap();
321 0 : let bin_path = pg_bin_dir.join("pg_ctl");
322 :
323 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
324 0 : let envs = [
325 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
326 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
327 0 : ];
328 0 :
329 0 : Command::new(bin_path)
330 0 : .args(args)
331 0 : .envs(envs)
332 0 : .spawn()
333 0 : .expect("Failed to spawn pg_ctl, binary_missing?")
334 0 : .wait()
335 0 : .await
336 0 : .expect("Failed to wait for pg_ctl termination")
337 0 : }
338 :
339 0 : pub async fn start(&self, start_args: NeonStorageControllerStartArgs) -> anyhow::Result<()> {
340 0 : let instance_dir = self.storage_controller_instance_dir(start_args.instance_id);
341 0 : if let Err(err) = tokio::fs::create_dir(&instance_dir).await {
342 0 : if err.kind() != std::io::ErrorKind::AlreadyExists {
343 0 : panic!("Failed to create instance dir {instance_dir:?}");
344 0 : }
345 0 : }
346 :
347 0 : if self.env.generate_local_ssl_certs {
348 0 : self.env.generate_ssl_cert(
349 0 : &instance_dir.join("server.crt"),
350 0 : &instance_dir.join("server.key"),
351 0 : )?;
352 0 : }
353 :
354 0 : let listen_url = &self.env.control_plane_api;
355 0 :
356 0 : let scheme = listen_url.scheme();
357 0 : let host = listen_url.host_str().unwrap();
358 :
359 0 : let (listen_port, postgres_port) = if let Some(base_port) = start_args.base_port {
360 0 : (
361 0 : base_port,
362 0 : self.config
363 0 : .database_url
364 0 : .expect("--base-port requires NeonStorageControllerConf::database_url")
365 0 : .port(),
366 0 : )
367 : } else {
368 0 : let port = listen_url.port().unwrap();
369 0 : (port, port + 1)
370 : };
371 :
372 0 : self.listen_port
373 0 : .set(listen_port)
374 0 : .expect("StorageController::listen_port is only set here");
375 :
376 : // Do we remove the pid file on stop?
377 0 : let pg_started = self.is_postgres_running().await?;
378 0 : let pg_lib_dir = self.get_pg_lib_dir().await?;
379 :
380 0 : if !pg_started {
381 : // Start a vanilla Postgres process used by the storage controller for persistence.
382 0 : let pg_data_path = Utf8PathBuf::from_path_buf(self.env.base_data_dir.clone())
383 0 : .unwrap()
384 0 : .join("storage_controller_db");
385 0 : let pg_bin_dir = self.get_pg_bin_dir().await?;
386 0 : let pg_log_path = pg_data_path.join("postgres.log");
387 0 :
388 0 : if !tokio::fs::try_exists(&pg_data_path).await? {
389 0 : let initdb_args = [
390 0 : "--pgdata",
391 0 : pg_data_path.as_ref(),
392 0 : "--username",
393 0 : &username(),
394 0 : "--no-sync",
395 0 : "--no-instructions",
396 0 : ];
397 0 : tracing::info!(
398 0 : "Initializing storage controller database with args: {:?}",
399 : initdb_args
400 : );
401 :
402 : // Initialize empty database
403 0 : let initdb_path = pg_bin_dir.join("initdb");
404 0 : let mut child = Command::new(&initdb_path)
405 0 : .envs(vec![
406 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
407 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
408 0 : ])
409 0 : .args(initdb_args)
410 0 : .spawn()
411 0 : .expect("Failed to spawn initdb");
412 0 : let status = child.wait().await?;
413 0 : if !status.success() {
414 0 : anyhow::bail!("initdb failed with status {status}");
415 0 : }
416 0 : };
417 :
418 : // Write a minimal config file:
419 : // - Specify the port, since this is chosen dynamically
420 : // - Switch off fsync, since we're running on lightweight test environments and when e.g. scale testing
421 : // the storage controller we don't want a slow local disk to interfere with that.
422 : //
423 : // NB: it's important that we rewrite this file on each start command so we propagate changes
424 : // from `LocalEnv`'s config file (`.neon/config`).
425 0 : tokio::fs::write(
426 0 : &pg_data_path.join("postgresql.conf"),
427 0 : format!("port = {}\nfsync=off\n", postgres_port),
428 0 : )
429 0 : .await?;
430 :
431 0 : println!("Starting storage controller database...");
432 0 : let db_start_args = [
433 0 : "-w",
434 0 : "-D",
435 0 : pg_data_path.as_ref(),
436 0 : "-l",
437 0 : pg_log_path.as_ref(),
438 0 : "-U",
439 0 : &username(),
440 0 : "start",
441 0 : ];
442 0 : tracing::info!(
443 0 : "Starting storage controller database with args: {:?}",
444 : db_start_args
445 : );
446 :
447 0 : let db_start_status = self.pg_ctl(db_start_args).await;
448 0 : let start_timeout: Duration = start_args.start_timeout.into();
449 0 : let db_start_deadline = Instant::now() + start_timeout;
450 0 : if !db_start_status.success() {
451 0 : return Err(anyhow::anyhow!(
452 0 : "Failed to start postgres {}",
453 0 : db_start_status.code().unwrap()
454 0 : ));
455 0 : }
456 :
457 : loop {
458 0 : if Instant::now() > db_start_deadline {
459 0 : return Err(anyhow::anyhow!("Timed out waiting for postgres to start"));
460 0 : }
461 0 :
462 0 : match self.pg_isready(&pg_bin_dir, postgres_port).await {
463 : Ok(true) => {
464 0 : tracing::info!("storage controller postgres is now ready");
465 0 : break;
466 : }
467 : Ok(false) => {
468 0 : tokio::time::sleep(Duration::from_millis(100)).await;
469 : }
470 0 : Err(e) => {
471 0 : tracing::warn!("Failed to check postgres status: {e}")
472 : }
473 : }
474 : }
475 :
476 0 : self.setup_database(postgres_port).await?;
477 0 : }
478 :
479 0 : let database_url = format!("postgresql://localhost:{}/{DB_NAME}", postgres_port);
480 0 :
481 0 : // We support running a startup SQL script to fiddle with the database before we launch storcon.
482 0 : // This is used by the test suite.
483 0 : let startup_script_path = self
484 0 : .env
485 0 : .base_data_dir
486 0 : .join("storage_controller_db.startup.sql");
487 0 : let startup_script = match tokio::fs::read_to_string(&startup_script_path).await {
488 0 : Ok(script) => {
489 0 : tokio::fs::remove_file(startup_script_path).await?;
490 0 : script
491 : }
492 0 : Err(e) => {
493 0 : if e.kind() == std::io::ErrorKind::NotFound {
494 : // always run some startup script so that this code path doesn't bit rot
495 0 : "BEGIN; COMMIT;".to_string()
496 : } else {
497 0 : anyhow::bail!("Failed to read startup script: {e}")
498 : }
499 : }
500 : };
501 0 : let (mut client, conn) = self.connect_to_database(postgres_port).await?;
502 0 : let conn = tokio::spawn(conn);
503 0 : let tx = client.build_transaction();
504 0 : let tx = tx.start().await?;
505 0 : tx.batch_execute(&startup_script).await?;
506 0 : tx.commit().await?;
507 0 : drop(client);
508 0 : conn.await??;
509 :
510 0 : let addr = format!("{}:{}", host, listen_port);
511 0 : let address_for_peers = Uri::builder()
512 0 : .scheme(scheme)
513 0 : .authority(addr.clone())
514 0 : .path_and_query("")
515 0 : .build()
516 0 : .unwrap();
517 0 :
518 0 : let mut args = vec![
519 0 : "--dev",
520 0 : "--database-url",
521 0 : &database_url,
522 0 : "--max-offline-interval",
523 0 : &humantime::Duration::from(self.config.max_offline).to_string(),
524 0 : "--max-warming-up-interval",
525 0 : &humantime::Duration::from(self.config.max_warming_up).to_string(),
526 0 : "--heartbeat-interval",
527 0 : &humantime::Duration::from(self.config.heartbeat_interval).to_string(),
528 0 : "--address-for-peers",
529 0 : &address_for_peers.to_string(),
530 0 : ]
531 0 : .into_iter()
532 0 : .map(|s| s.to_string())
533 0 : .collect::<Vec<_>>();
534 0 :
535 0 : match scheme {
536 0 : "http" => args.extend(["--listen".to_string(), addr]),
537 0 : "https" => args.extend(["--listen-https".to_string(), addr]),
538 : _ => {
539 0 : panic!("Unexpected url scheme in control_plane_api: {scheme}");
540 : }
541 : }
542 :
543 0 : if self.config.start_as_candidate {
544 0 : args.push("--start-as-candidate".to_string());
545 0 : }
546 :
547 0 : if self.config.use_https_pageserver_api {
548 0 : args.push("--use-https-pageserver-api".to_string());
549 0 : }
550 :
551 0 : if self.config.use_https_safekeeper_api {
552 0 : args.push("--use-https-safekeeper-api".to_string());
553 0 : }
554 :
555 0 : if let Some(ssl_ca_file) = self.env.ssl_ca_cert_path() {
556 0 : args.push(format!("--ssl-ca-file={}", ssl_ca_file.to_str().unwrap()));
557 0 : }
558 :
559 0 : if let Some(private_key) = &self.private_key {
560 0 : let claims = Claims::new(None, Scope::PageServerApi);
561 0 : let jwt_token =
562 0 : encode_from_key_file(&claims, private_key).expect("failed to generate jwt token");
563 0 : args.push(format!("--jwt-token={jwt_token}"));
564 0 :
565 0 : let peer_claims = Claims::new(None, Scope::Admin);
566 0 : let peer_jwt_token = encode_from_key_file(&peer_claims, private_key)
567 0 : .expect("failed to generate jwt token");
568 0 : args.push(format!("--peer-jwt-token={peer_jwt_token}"));
569 0 : }
570 :
571 0 : if let Some(public_key) = &self.public_key {
572 0 : args.push(format!("--public-key=\"{public_key}\""));
573 0 : }
574 :
575 0 : if let Some(control_plane_hooks_api) = &self.env.control_plane_hooks_api {
576 0 : args.push(format!("--control-plane-url={control_plane_hooks_api}"));
577 0 : }
578 :
579 0 : if let Some(split_threshold) = self.config.split_threshold.as_ref() {
580 0 : args.push(format!("--split-threshold={split_threshold}"))
581 0 : }
582 :
583 0 : if let Some(max_split_shards) = self.config.max_split_shards.as_ref() {
584 0 : args.push(format!("--max-split-shards={max_split_shards}"))
585 0 : }
586 :
587 0 : if let Some(initial_split_threshold) = self.config.initial_split_threshold.as_ref() {
588 0 : args.push(format!(
589 0 : "--initial-split-threshold={initial_split_threshold}"
590 0 : ))
591 0 : }
592 :
593 0 : if let Some(initial_split_shards) = self.config.initial_split_shards.as_ref() {
594 0 : args.push(format!("--initial-split-shards={initial_split_shards}"))
595 0 : }
596 :
597 0 : if let Some(lag) = self.config.max_secondary_lag_bytes.as_ref() {
598 0 : args.push(format!("--max-secondary-lag-bytes={lag}"))
599 0 : }
600 :
601 0 : if let Some(threshold) = self.config.long_reconcile_threshold {
602 0 : args.push(format!(
603 0 : "--long-reconcile-threshold={}",
604 0 : humantime::Duration::from(threshold)
605 0 : ))
606 0 : }
607 :
608 0 : args.push(format!(
609 0 : "--neon-local-repo-dir={}",
610 0 : self.env.base_data_dir.display()
611 0 : ));
612 0 :
613 0 : if self.config.timelines_onto_safekeepers {
614 0 : args.push("--timelines-onto-safekeepers".to_string());
615 0 : }
616 :
617 0 : println!("Starting storage controller");
618 0 :
619 0 : background_process::start_process(
620 0 : COMMAND,
621 0 : &instance_dir,
622 0 : &self.env.storage_controller_bin(),
623 0 : args,
624 0 : vec![
625 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
626 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
627 0 : ],
628 0 : background_process::InitialPidFile::Create(self.pid_file(start_args.instance_id)),
629 0 : &start_args.start_timeout,
630 0 : || async {
631 0 : match self.ready().await {
632 0 : Ok(_) => Ok(true),
633 0 : Err(_) => Ok(false),
634 : }
635 0 : },
636 0 : )
637 0 : .await?;
638 :
639 0 : Ok(())
640 0 : }
641 :
642 0 : pub async fn stop(&self, stop_args: NeonStorageControllerStopArgs) -> anyhow::Result<()> {
643 0 : background_process::stop_process(
644 0 : stop_args.immediate,
645 0 : COMMAND,
646 0 : &self.pid_file(stop_args.instance_id),
647 0 : )?;
648 :
649 0 : let storcon_instances = self.env.storage_controller_instances().await?;
650 0 : for (instance_id, instanced_dir_path) in storcon_instances {
651 0 : if instance_id == stop_args.instance_id {
652 0 : continue;
653 0 : }
654 0 :
655 0 : let pid_file = instanced_dir_path.join("storage_controller.pid");
656 0 : let pid = tokio::fs::read_to_string(&pid_file)
657 0 : .await
658 0 : .map_err(|err| {
659 0 : anyhow::anyhow!("Failed to read storcon pid file at {pid_file:?}: {err}")
660 0 : })?
661 0 : .parse::<i32>()
662 0 : .expect("pid is valid i32");
663 :
664 0 : let other_proc_alive = !background_process::process_has_stopped(Pid::from_raw(pid))?;
665 0 : if other_proc_alive {
666 : // There is another storage controller instance running, so we return
667 : // and leave the database running.
668 0 : return Ok(());
669 0 : }
670 : }
671 :
672 0 : let pg_data_path = self.env.base_data_dir.join("storage_controller_db");
673 0 :
674 0 : println!("Stopping storage controller database...");
675 0 : let pg_stop_args = ["-D", &pg_data_path.to_string_lossy(), "stop"];
676 0 : let stop_status = self.pg_ctl(pg_stop_args).await;
677 0 : if !stop_status.success() {
678 0 : match self.is_postgres_running().await {
679 : Ok(false) => {
680 0 : println!("Storage controller database is already stopped");
681 0 : return Ok(());
682 : }
683 : Ok(true) => {
684 0 : anyhow::bail!("Failed to stop storage controller database");
685 : }
686 0 : Err(err) => {
687 0 : anyhow::bail!("Failed to stop storage controller database: {err}");
688 : }
689 : }
690 0 : }
691 0 :
692 0 : Ok(())
693 0 : }
694 :
695 0 : async fn is_postgres_running(&self) -> anyhow::Result<bool> {
696 0 : let pg_data_path = self.env.base_data_dir.join("storage_controller_db");
697 0 :
698 0 : let pg_status_args = ["-D", &pg_data_path.to_string_lossy(), "status"];
699 0 : let status_exitcode = self.pg_ctl(pg_status_args).await;
700 :
701 : // pg_ctl status returns this exit code if postgres is not running: in this case it is
702 : // fine that stop failed. Otherwise it is an error that stop failed.
703 : const PG_STATUS_NOT_RUNNING: i32 = 3;
704 : const PG_NO_DATA_DIR: i32 = 4;
705 : const PG_STATUS_RUNNING: i32 = 0;
706 0 : match status_exitcode.code() {
707 0 : Some(PG_STATUS_NOT_RUNNING) => Ok(false),
708 0 : Some(PG_NO_DATA_DIR) => Ok(false),
709 0 : Some(PG_STATUS_RUNNING) => Ok(true),
710 0 : Some(code) => Err(anyhow::anyhow!(
711 0 : "pg_ctl status returned unexpected status code: {:?}",
712 0 : code
713 0 : )),
714 0 : None => Err(anyhow::anyhow!("pg_ctl status returned no status code")),
715 : }
716 0 : }
717 :
718 0 : fn get_claims_for_path(path: &str) -> anyhow::Result<Option<Claims>> {
719 0 : let category = match path.find('/') {
720 0 : Some(idx) => &path[..idx],
721 0 : None => path,
722 : };
723 :
724 0 : match category {
725 0 : "status" | "ready" => Ok(None),
726 0 : "control" | "debug" => Ok(Some(Claims::new(None, Scope::Admin))),
727 0 : "v1" => Ok(Some(Claims::new(None, Scope::PageServerApi))),
728 0 : _ => Err(anyhow::anyhow!("Failed to determine claims for {}", path)),
729 : }
730 0 : }
731 :
732 : /// Simple HTTP request wrapper for calling into storage controller
733 0 : async fn dispatch<RQ, RS>(
734 0 : &self,
735 0 : method: reqwest::Method,
736 0 : path: String,
737 0 : body: Option<RQ>,
738 0 : ) -> anyhow::Result<RS>
739 0 : where
740 0 : RQ: Serialize + Sized,
741 0 : RS: DeserializeOwned + Sized,
742 0 : {
743 : // In the special case of the `storage_controller start` subcommand, we wish
744 : // to use the API endpoint of the newly started storage controller in order
745 : // to pass the readiness check. In this scenario [`Self::listen_port`] will
746 : // be set (see [`Self::start`]).
747 : //
748 : // Otherwise, we infer the storage controller api endpoint from the configured
749 : // control plane API.
750 0 : let port = if let Some(port) = self.listen_port.get() {
751 0 : *port
752 : } else {
753 0 : self.env.control_plane_api.port().unwrap()
754 : };
755 :
756 : // The configured URL has the /upcall path prefix for pageservers to use: we will strip that out
757 : // for general purpose API access.
758 0 : let url = Url::from_str(&format!(
759 0 : "{}://{}:{port}/{path}",
760 0 : self.env.control_plane_api.scheme(),
761 0 : self.env.control_plane_api.host_str().unwrap(),
762 0 : ))
763 0 : .unwrap();
764 0 :
765 0 : let mut builder = self.client.request(method, url);
766 0 : if let Some(body) = body {
767 0 : builder = builder.json(&body)
768 0 : }
769 0 : if let Some(private_key) = &self.private_key {
770 0 : println!("Getting claims for path {}", path);
771 0 : if let Some(required_claims) = Self::get_claims_for_path(&path)? {
772 0 : println!("Got claims {:?} for path {}", required_claims, path);
773 0 : let jwt_token = encode_from_key_file(&required_claims, private_key)?;
774 0 : builder = builder.header(
775 0 : reqwest::header::AUTHORIZATION,
776 0 : format!("Bearer {jwt_token}"),
777 0 : );
778 0 : }
779 0 : }
780 :
781 0 : let response = builder.send().await?;
782 0 : let response = response.error_from_body().await?;
783 :
784 0 : Ok(response
785 0 : .json()
786 0 : .await
787 0 : .map_err(pageserver_client::mgmt_api::Error::ReceiveBody)?)
788 0 : }
789 :
790 : /// Call into the attach_hook API, for use before handing out attachments to pageservers
791 : #[instrument(skip(self))]
792 : pub async fn attach_hook(
793 : &self,
794 : tenant_shard_id: TenantShardId,
795 : pageserver_id: NodeId,
796 : ) -> anyhow::Result<Option<u32>> {
797 : let request = AttachHookRequest {
798 : tenant_shard_id,
799 : node_id: Some(pageserver_id),
800 : generation_override: None,
801 : };
802 :
803 : let response = self
804 : .dispatch::<_, AttachHookResponse>(
805 : Method::POST,
806 : "debug/v1/attach-hook".to_string(),
807 : Some(request),
808 : )
809 : .await?;
810 :
811 : Ok(response.generation)
812 : }
813 :
814 : #[instrument(skip(self))]
815 : pub async fn inspect(
816 : &self,
817 : tenant_shard_id: TenantShardId,
818 : ) -> anyhow::Result<Option<(u32, NodeId)>> {
819 : let request = InspectRequest { tenant_shard_id };
820 :
821 : let response = self
822 : .dispatch::<_, InspectResponse>(
823 : Method::POST,
824 : "debug/v1/inspect".to_string(),
825 : Some(request),
826 : )
827 : .await?;
828 :
829 : Ok(response.attachment)
830 : }
831 :
832 : #[instrument(skip(self))]
833 : pub async fn tenant_create(
834 : &self,
835 : req: TenantCreateRequest,
836 : ) -> anyhow::Result<TenantCreateResponse> {
837 : self.dispatch(Method::POST, "v1/tenant".to_string(), Some(req))
838 : .await
839 : }
840 :
841 : #[instrument(skip(self))]
842 : pub async fn tenant_import(&self, tenant_id: TenantId) -> anyhow::Result<TenantCreateResponse> {
843 : self.dispatch::<(), TenantCreateResponse>(
844 : Method::POST,
845 : format!("debug/v1/tenant/{tenant_id}/import"),
846 : None,
847 : )
848 : .await
849 : }
850 :
851 : #[instrument(skip(self))]
852 : pub async fn tenant_locate(&self, tenant_id: TenantId) -> anyhow::Result<TenantLocateResponse> {
853 : self.dispatch::<(), _>(
854 : Method::GET,
855 : format!("debug/v1/tenant/{tenant_id}/locate"),
856 : None,
857 : )
858 : .await
859 : }
860 :
861 : #[instrument(skip_all, fields(node_id=%req.node_id))]
862 : pub async fn node_register(&self, req: NodeRegisterRequest) -> anyhow::Result<()> {
863 : self.dispatch::<_, ()>(Method::POST, "control/v1/node".to_string(), Some(req))
864 : .await
865 : }
866 :
867 : #[instrument(skip_all, fields(node_id=%req.node_id))]
868 : pub async fn node_configure(&self, req: NodeConfigureRequest) -> anyhow::Result<()> {
869 : self.dispatch::<_, ()>(
870 : Method::PUT,
871 : format!("control/v1/node/{}/config", req.node_id),
872 : Some(req),
873 : )
874 : .await
875 : }
876 :
877 0 : pub async fn node_list(&self) -> anyhow::Result<Vec<NodeDescribeResponse>> {
878 0 : self.dispatch::<(), Vec<NodeDescribeResponse>>(
879 0 : Method::GET,
880 0 : "control/v1/node".to_string(),
881 0 : None,
882 0 : )
883 0 : .await
884 0 : }
885 :
886 : #[instrument(skip(self))]
887 : pub async fn ready(&self) -> anyhow::Result<()> {
888 : self.dispatch::<(), ()>(Method::GET, "ready".to_string(), None)
889 : .await
890 : }
891 :
892 : #[instrument(skip_all, fields(%tenant_id, timeline_id=%req.new_timeline_id))]
893 : pub async fn tenant_timeline_create(
894 : &self,
895 : tenant_id: TenantId,
896 : req: TimelineCreateRequest,
897 : ) -> anyhow::Result<TimelineInfo> {
898 : self.dispatch(
899 : Method::POST,
900 : format!("v1/tenant/{tenant_id}/timeline"),
901 : Some(req),
902 : )
903 : .await
904 : }
905 :
906 0 : pub async fn set_tenant_config(&self, req: &TenantConfigRequest) -> anyhow::Result<()> {
907 0 : self.dispatch(Method::PUT, "v1/tenant/config".to_string(), Some(req))
908 0 : .await
909 0 : }
910 : }
|
