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::{Certificate, 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 ssl_ca_certs = env.ssl_ca_cert_path().map(|ssl_ca_file| {
147 0 : let buf = std::fs::read(ssl_ca_file).expect("SSL CA file should exist");
148 0 : Certificate::from_pem_bundle(&buf).expect("SSL CA file should be valid")
149 0 : });
150 0 :
151 0 : let mut http_client = reqwest::Client::builder();
152 0 : for ssl_ca_cert in ssl_ca_certs.unwrap_or_default() {
153 0 : http_client = http_client.add_root_certificate(ssl_ca_cert);
154 0 : }
155 0 : let http_client = http_client
156 0 : .build()
157 0 : .expect("HTTP client should construct with no error");
158 0 :
159 0 : Self {
160 0 : env: env.clone(),
161 0 : private_key,
162 0 : public_key,
163 0 : client: http_client,
164 0 : config: env.storage_controller.clone(),
165 0 : listen_port: OnceLock::default(),
166 0 : }
167 0 : }
168 :
169 0 : fn storage_controller_instance_dir(&self, instance_id: u8) -> PathBuf {
170 0 : self.env
171 0 : .base_data_dir
172 0 : .join(format!("storage_controller_{}", instance_id))
173 0 : }
174 :
175 0 : fn pid_file(&self, instance_id: u8) -> Utf8PathBuf {
176 0 : Utf8PathBuf::from_path_buf(
177 0 : self.storage_controller_instance_dir(instance_id)
178 0 : .join("storage_controller.pid"),
179 0 : )
180 0 : .expect("non-Unicode path")
181 0 : }
182 :
183 : /// Find the directory containing postgres subdirectories, such `bin` and `lib`
184 : ///
185 : /// This usually uses STORAGE_CONTROLLER_POSTGRES_VERSION of postgres, but will fall back
186 : /// to other versions if that one isn't found. Some automated tests create circumstances
187 : /// where only one version is available in pg_distrib_dir, such as `test_remote_extensions`.
188 0 : async fn get_pg_dir(&self, dir_name: &str) -> anyhow::Result<Utf8PathBuf> {
189 0 : let prefer_versions = [STORAGE_CONTROLLER_POSTGRES_VERSION, 16, 15, 14];
190 :
191 0 : for v in prefer_versions {
192 0 : let path = Utf8PathBuf::from_path_buf(self.env.pg_dir(v, dir_name)?).unwrap();
193 0 : if tokio::fs::try_exists(&path).await? {
194 0 : return Ok(path);
195 0 : }
196 : }
197 :
198 : // Fall through
199 0 : anyhow::bail!(
200 0 : "Postgres directory '{}' not found in {}",
201 0 : dir_name,
202 0 : self.env.pg_distrib_dir.display(),
203 0 : );
204 0 : }
205 :
206 0 : pub async fn get_pg_bin_dir(&self) -> anyhow::Result<Utf8PathBuf> {
207 0 : self.get_pg_dir("bin").await
208 0 : }
209 :
210 0 : pub async fn get_pg_lib_dir(&self) -> anyhow::Result<Utf8PathBuf> {
211 0 : self.get_pg_dir("lib").await
212 0 : }
213 :
214 : /// Readiness check for our postgres process
215 0 : async fn pg_isready(&self, pg_bin_dir: &Utf8Path, postgres_port: u16) -> anyhow::Result<bool> {
216 0 : let bin_path = pg_bin_dir.join("pg_isready");
217 0 : let args = [
218 0 : "-h",
219 0 : "localhost",
220 0 : "-U",
221 0 : &username(),
222 0 : "-d",
223 0 : DB_NAME,
224 0 : "-p",
225 0 : &format!("{}", postgres_port),
226 0 : ];
227 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
228 0 : let envs = [
229 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
230 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
231 0 : ];
232 0 : let exitcode = Command::new(bin_path)
233 0 : .args(args)
234 0 : .envs(envs)
235 0 : .spawn()?
236 0 : .wait()
237 0 : .await?;
238 :
239 0 : Ok(exitcode.success())
240 0 : }
241 :
242 : /// Create our database if it doesn't exist
243 : ///
244 : /// This function is equivalent to the `diesel setup` command in the diesel CLI. We implement
245 : /// the same steps by hand to avoid imposing a dependency on installing diesel-cli for developers
246 : /// who just want to run `cargo neon_local` without knowing about diesel.
247 : ///
248 : /// Returns the database url
249 0 : pub async fn setup_database(&self, postgres_port: u16) -> anyhow::Result<String> {
250 0 : let database_url = format!(
251 0 : "postgresql://{}@localhost:{}/{DB_NAME}",
252 0 : &username(),
253 0 : postgres_port
254 0 : );
255 :
256 0 : let pg_bin_dir = self.get_pg_bin_dir().await?;
257 0 : let createdb_path = pg_bin_dir.join("createdb");
258 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
259 0 : let envs = [
260 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
261 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
262 0 : ];
263 0 : let output = Command::new(&createdb_path)
264 0 : .args([
265 0 : "-h",
266 0 : "localhost",
267 0 : "-p",
268 0 : &format!("{}", postgres_port),
269 0 : "-U",
270 0 : &username(),
271 0 : "-O",
272 0 : &username(),
273 0 : DB_NAME,
274 0 : ])
275 0 : .envs(envs)
276 0 : .output()
277 0 : .await
278 0 : .expect("Failed to spawn createdb");
279 0 :
280 0 : if !output.status.success() {
281 0 : let stderr = String::from_utf8(output.stderr).expect("Non-UTF8 output from createdb");
282 0 : if stderr.contains("already exists") {
283 0 : tracing::info!("Database {DB_NAME} already exists");
284 : } else {
285 0 : anyhow::bail!("createdb failed with status {}: {stderr}", output.status);
286 : }
287 0 : }
288 :
289 0 : Ok(database_url)
290 0 : }
291 :
292 0 : pub async fn connect_to_database(
293 0 : &self,
294 0 : postgres_port: u16,
295 0 : ) -> anyhow::Result<(
296 0 : tokio_postgres::Client,
297 0 : tokio_postgres::Connection<tokio_postgres::Socket, tokio_postgres::tls::NoTlsStream>,
298 0 : )> {
299 0 : tokio_postgres::Config::new()
300 0 : .host("localhost")
301 0 : .port(postgres_port)
302 0 : // The user is the ambient operating system user name.
303 0 : // That is an impurity which we want to fix in => TODO https://github.com/neondatabase/neon/issues/8400
304 0 : //
305 0 : // Until we get there, use the ambient operating system user name.
306 0 : // Recent tokio-postgres versions default to this if the user isn't specified.
307 0 : // But tokio-postgres fork doesn't have this upstream commit:
308 0 : // https://github.com/sfackler/rust-postgres/commit/cb609be758f3fb5af537f04b584a2ee0cebd5e79
309 0 : // => we should rebase our fork => TODO https://github.com/neondatabase/neon/issues/8399
310 0 : .user(&username())
311 0 : .dbname(DB_NAME)
312 0 : .connect(tokio_postgres::NoTls)
313 0 : .await
314 0 : .map_err(anyhow::Error::new)
315 0 : }
316 :
317 : /// Wrapper for the pg_ctl binary, which we spawn as a short-lived subprocess when starting and stopping postgres
318 0 : async fn pg_ctl<I, S>(&self, args: I) -> ExitStatus
319 0 : where
320 0 : I: IntoIterator<Item = S>,
321 0 : S: AsRef<OsStr>,
322 0 : {
323 0 : let pg_bin_dir = self.get_pg_bin_dir().await.unwrap();
324 0 : let bin_path = pg_bin_dir.join("pg_ctl");
325 :
326 0 : let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
327 0 : let envs = [
328 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
329 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
330 0 : ];
331 0 :
332 0 : Command::new(bin_path)
333 0 : .args(args)
334 0 : .envs(envs)
335 0 : .spawn()
336 0 : .expect("Failed to spawn pg_ctl, binary_missing?")
337 0 : .wait()
338 0 : .await
339 0 : .expect("Failed to wait for pg_ctl termination")
340 0 : }
341 :
342 0 : pub async fn start(&self, start_args: NeonStorageControllerStartArgs) -> anyhow::Result<()> {
343 0 : let instance_dir = self.storage_controller_instance_dir(start_args.instance_id);
344 0 : if let Err(err) = tokio::fs::create_dir(&instance_dir).await {
345 0 : if err.kind() != std::io::ErrorKind::AlreadyExists {
346 0 : panic!("Failed to create instance dir {instance_dir:?}");
347 0 : }
348 0 : }
349 :
350 0 : if self.env.generate_local_ssl_certs {
351 0 : self.env.generate_ssl_cert(
352 0 : &instance_dir.join("server.crt"),
353 0 : &instance_dir.join("server.key"),
354 0 : )?;
355 0 : }
356 :
357 0 : let listen_url = &self.env.control_plane_api;
358 0 :
359 0 : let scheme = listen_url.scheme();
360 0 : let host = listen_url.host_str().unwrap();
361 :
362 0 : let (listen_port, postgres_port) = if let Some(base_port) = start_args.base_port {
363 0 : (
364 0 : base_port,
365 0 : self.config
366 0 : .database_url
367 0 : .expect("--base-port requires NeonStorageControllerConf::database_url")
368 0 : .port(),
369 0 : )
370 : } else {
371 0 : let port = listen_url.port().unwrap();
372 0 : (port, port + 1)
373 : };
374 :
375 0 : self.listen_port
376 0 : .set(listen_port)
377 0 : .expect("StorageController::listen_port is only set here");
378 :
379 : // Do we remove the pid file on stop?
380 0 : let pg_started = self.is_postgres_running().await?;
381 0 : let pg_lib_dir = self.get_pg_lib_dir().await?;
382 :
383 0 : if !pg_started {
384 : // Start a vanilla Postgres process used by the storage controller for persistence.
385 0 : let pg_data_path = Utf8PathBuf::from_path_buf(self.env.base_data_dir.clone())
386 0 : .unwrap()
387 0 : .join("storage_controller_db");
388 0 : let pg_bin_dir = self.get_pg_bin_dir().await?;
389 0 : let pg_log_path = pg_data_path.join("postgres.log");
390 0 :
391 0 : if !tokio::fs::try_exists(&pg_data_path).await? {
392 0 : let initdb_args = [
393 0 : "--pgdata",
394 0 : pg_data_path.as_ref(),
395 0 : "--username",
396 0 : &username(),
397 0 : "--no-sync",
398 0 : "--no-instructions",
399 0 : ];
400 0 : tracing::info!(
401 0 : "Initializing storage controller database with args: {:?}",
402 : initdb_args
403 : );
404 :
405 : // Initialize empty database
406 0 : let initdb_path = pg_bin_dir.join("initdb");
407 0 : let mut child = Command::new(&initdb_path)
408 0 : .envs(vec![
409 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
410 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
411 0 : ])
412 0 : .args(initdb_args)
413 0 : .spawn()
414 0 : .expect("Failed to spawn initdb");
415 0 : let status = child.wait().await?;
416 0 : if !status.success() {
417 0 : anyhow::bail!("initdb failed with status {status}");
418 0 : }
419 0 : };
420 :
421 : // Write a minimal config file:
422 : // - Specify the port, since this is chosen dynamically
423 : // - Switch off fsync, since we're running on lightweight test environments and when e.g. scale testing
424 : // the storage controller we don't want a slow local disk to interfere with that.
425 : //
426 : // NB: it's important that we rewrite this file on each start command so we propagate changes
427 : // from `LocalEnv`'s config file (`.neon/config`).
428 0 : tokio::fs::write(
429 0 : &pg_data_path.join("postgresql.conf"),
430 0 : format!("port = {}\nfsync=off\n", postgres_port),
431 0 : )
432 0 : .await?;
433 :
434 0 : println!("Starting storage controller database...");
435 0 : let db_start_args = [
436 0 : "-w",
437 0 : "-D",
438 0 : pg_data_path.as_ref(),
439 0 : "-l",
440 0 : pg_log_path.as_ref(),
441 0 : "-U",
442 0 : &username(),
443 0 : "start",
444 0 : ];
445 0 : tracing::info!(
446 0 : "Starting storage controller database with args: {:?}",
447 : db_start_args
448 : );
449 :
450 0 : let db_start_status = self.pg_ctl(db_start_args).await;
451 0 : let start_timeout: Duration = start_args.start_timeout.into();
452 0 : let db_start_deadline = Instant::now() + start_timeout;
453 0 : if !db_start_status.success() {
454 0 : return Err(anyhow::anyhow!(
455 0 : "Failed to start postgres {}",
456 0 : db_start_status.code().unwrap()
457 0 : ));
458 0 : }
459 :
460 : loop {
461 0 : if Instant::now() > db_start_deadline {
462 0 : return Err(anyhow::anyhow!("Timed out waiting for postgres to start"));
463 0 : }
464 0 :
465 0 : match self.pg_isready(&pg_bin_dir, postgres_port).await {
466 : Ok(true) => {
467 0 : tracing::info!("storage controller postgres is now ready");
468 0 : break;
469 : }
470 : Ok(false) => {
471 0 : tokio::time::sleep(Duration::from_millis(100)).await;
472 : }
473 0 : Err(e) => {
474 0 : tracing::warn!("Failed to check postgres status: {e}")
475 : }
476 : }
477 : }
478 :
479 0 : self.setup_database(postgres_port).await?;
480 0 : }
481 :
482 0 : let database_url = format!("postgresql://localhost:{}/{DB_NAME}", postgres_port);
483 0 :
484 0 : // We support running a startup SQL script to fiddle with the database before we launch storcon.
485 0 : // This is used by the test suite.
486 0 : let startup_script_path = self
487 0 : .env
488 0 : .base_data_dir
489 0 : .join("storage_controller_db.startup.sql");
490 0 : let startup_script = match tokio::fs::read_to_string(&startup_script_path).await {
491 0 : Ok(script) => {
492 0 : tokio::fs::remove_file(startup_script_path).await?;
493 0 : script
494 : }
495 0 : Err(e) => {
496 0 : if e.kind() == std::io::ErrorKind::NotFound {
497 : // always run some startup script so that this code path doesn't bit rot
498 0 : "BEGIN; COMMIT;".to_string()
499 : } else {
500 0 : anyhow::bail!("Failed to read startup script: {e}")
501 : }
502 : }
503 : };
504 0 : let (mut client, conn) = self.connect_to_database(postgres_port).await?;
505 0 : let conn = tokio::spawn(conn);
506 0 : let tx = client.build_transaction();
507 0 : let tx = tx.start().await?;
508 0 : tx.batch_execute(&startup_script).await?;
509 0 : tx.commit().await?;
510 0 : drop(client);
511 0 : conn.await??;
512 :
513 0 : let addr = format!("{}:{}", host, listen_port);
514 0 : let address_for_peers = Uri::builder()
515 0 : .scheme(scheme)
516 0 : .authority(addr.clone())
517 0 : .path_and_query("")
518 0 : .build()
519 0 : .unwrap();
520 0 :
521 0 : let mut args = vec![
522 0 : "--dev",
523 0 : "--database-url",
524 0 : &database_url,
525 0 : "--max-offline-interval",
526 0 : &humantime::Duration::from(self.config.max_offline).to_string(),
527 0 : "--max-warming-up-interval",
528 0 : &humantime::Duration::from(self.config.max_warming_up).to_string(),
529 0 : "--heartbeat-interval",
530 0 : &humantime::Duration::from(self.config.heartbeat_interval).to_string(),
531 0 : "--address-for-peers",
532 0 : &address_for_peers.to_string(),
533 0 : ]
534 0 : .into_iter()
535 0 : .map(|s| s.to_string())
536 0 : .collect::<Vec<_>>();
537 0 :
538 0 : match scheme {
539 0 : "http" => args.extend(["--listen".to_string(), addr]),
540 0 : "https" => args.extend(["--listen-https".to_string(), addr]),
541 : _ => {
542 0 : panic!("Unexpected url scheme in control_plane_api: {scheme}");
543 : }
544 : }
545 :
546 0 : if self.config.start_as_candidate {
547 0 : args.push("--start-as-candidate".to_string());
548 0 : }
549 :
550 0 : if self.config.use_https_pageserver_api {
551 0 : args.push("--use-https-pageserver-api".to_string());
552 0 : }
553 :
554 0 : if self.config.use_https_safekeeper_api {
555 0 : args.push("--use-https-safekeeper-api".to_string());
556 0 : }
557 :
558 0 : if self.config.use_local_compute_notifications {
559 0 : args.push("--use-local-compute-notifications".to_string());
560 0 : }
561 :
562 0 : if let Some(ssl_ca_file) = self.env.ssl_ca_cert_path() {
563 0 : args.push(format!("--ssl-ca-file={}", ssl_ca_file.to_str().unwrap()));
564 0 : }
565 :
566 0 : if let Some(private_key) = &self.private_key {
567 0 : let claims = Claims::new(None, Scope::PageServerApi);
568 0 : let jwt_token =
569 0 : encode_from_key_file(&claims, private_key).expect("failed to generate jwt token");
570 0 : args.push(format!("--jwt-token={jwt_token}"));
571 0 :
572 0 : let peer_claims = Claims::new(None, Scope::Admin);
573 0 : let peer_jwt_token = encode_from_key_file(&peer_claims, private_key)
574 0 : .expect("failed to generate jwt token");
575 0 : args.push(format!("--peer-jwt-token={peer_jwt_token}"));
576 0 : }
577 :
578 0 : if let Some(public_key) = &self.public_key {
579 0 : args.push(format!("--public-key=\"{public_key}\""));
580 0 : }
581 :
582 0 : if let Some(control_plane_hooks_api) = &self.env.control_plane_hooks_api {
583 0 : args.push(format!("--control-plane-url={control_plane_hooks_api}"));
584 0 : }
585 :
586 0 : if let Some(split_threshold) = self.config.split_threshold.as_ref() {
587 0 : args.push(format!("--split-threshold={split_threshold}"))
588 0 : }
589 :
590 0 : if let Some(max_split_shards) = self.config.max_split_shards.as_ref() {
591 0 : args.push(format!("--max-split-shards={max_split_shards}"))
592 0 : }
593 :
594 0 : if let Some(initial_split_threshold) = self.config.initial_split_threshold.as_ref() {
595 0 : args.push(format!(
596 0 : "--initial-split-threshold={initial_split_threshold}"
597 0 : ))
598 0 : }
599 :
600 0 : if let Some(initial_split_shards) = self.config.initial_split_shards.as_ref() {
601 0 : args.push(format!("--initial-split-shards={initial_split_shards}"))
602 0 : }
603 :
604 0 : if let Some(lag) = self.config.max_secondary_lag_bytes.as_ref() {
605 0 : args.push(format!("--max-secondary-lag-bytes={lag}"))
606 0 : }
607 :
608 0 : if let Some(threshold) = self.config.long_reconcile_threshold {
609 0 : args.push(format!(
610 0 : "--long-reconcile-threshold={}",
611 0 : humantime::Duration::from(threshold)
612 0 : ))
613 0 : }
614 :
615 0 : args.push(format!(
616 0 : "--neon-local-repo-dir={}",
617 0 : self.env.base_data_dir.display()
618 0 : ));
619 0 :
620 0 : if self.config.timelines_onto_safekeepers {
621 0 : args.push("--timelines-onto-safekeepers".to_string());
622 0 : }
623 :
624 0 : println!("Starting storage controller");
625 0 :
626 0 : background_process::start_process(
627 0 : COMMAND,
628 0 : &instance_dir,
629 0 : &self.env.storage_controller_bin(),
630 0 : args,
631 0 : vec![
632 0 : ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
633 0 : ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
634 0 : ],
635 0 : background_process::InitialPidFile::Create(self.pid_file(start_args.instance_id)),
636 0 : &start_args.start_timeout,
637 0 : || async {
638 0 : match self.ready().await {
639 0 : Ok(_) => Ok(true),
640 0 : Err(_) => Ok(false),
641 : }
642 0 : },
643 0 : )
644 0 : .await?;
645 :
646 0 : Ok(())
647 0 : }
648 :
649 0 : pub async fn stop(&self, stop_args: NeonStorageControllerStopArgs) -> anyhow::Result<()> {
650 0 : background_process::stop_process(
651 0 : stop_args.immediate,
652 0 : COMMAND,
653 0 : &self.pid_file(stop_args.instance_id),
654 0 : )?;
655 :
656 0 : let storcon_instances = self.env.storage_controller_instances().await?;
657 0 : for (instance_id, instanced_dir_path) in storcon_instances {
658 0 : if instance_id == stop_args.instance_id {
659 0 : continue;
660 0 : }
661 0 :
662 0 : let pid_file = instanced_dir_path.join("storage_controller.pid");
663 0 : let pid = tokio::fs::read_to_string(&pid_file)
664 0 : .await
665 0 : .map_err(|err| {
666 0 : anyhow::anyhow!("Failed to read storcon pid file at {pid_file:?}: {err}")
667 0 : })?
668 0 : .parse::<i32>()
669 0 : .expect("pid is valid i32");
670 :
671 0 : let other_proc_alive = !background_process::process_has_stopped(Pid::from_raw(pid))?;
672 0 : if other_proc_alive {
673 : // There is another storage controller instance running, so we return
674 : // and leave the database running.
675 0 : return Ok(());
676 0 : }
677 : }
678 :
679 0 : let pg_data_path = self.env.base_data_dir.join("storage_controller_db");
680 0 :
681 0 : println!("Stopping storage controller database...");
682 0 : let pg_stop_args = ["-D", &pg_data_path.to_string_lossy(), "stop"];
683 0 : let stop_status = self.pg_ctl(pg_stop_args).await;
684 0 : if !stop_status.success() {
685 0 : match self.is_postgres_running().await {
686 : Ok(false) => {
687 0 : println!("Storage controller database is already stopped");
688 0 : return Ok(());
689 : }
690 : Ok(true) => {
691 0 : anyhow::bail!("Failed to stop storage controller database");
692 : }
693 0 : Err(err) => {
694 0 : anyhow::bail!("Failed to stop storage controller database: {err}");
695 : }
696 : }
697 0 : }
698 0 :
699 0 : Ok(())
700 0 : }
701 :
702 0 : async fn is_postgres_running(&self) -> anyhow::Result<bool> {
703 0 : let pg_data_path = self.env.base_data_dir.join("storage_controller_db");
704 0 :
705 0 : let pg_status_args = ["-D", &pg_data_path.to_string_lossy(), "status"];
706 0 : let status_exitcode = self.pg_ctl(pg_status_args).await;
707 :
708 : // pg_ctl status returns this exit code if postgres is not running: in this case it is
709 : // fine that stop failed. Otherwise it is an error that stop failed.
710 : const PG_STATUS_NOT_RUNNING: i32 = 3;
711 : const PG_NO_DATA_DIR: i32 = 4;
712 : const PG_STATUS_RUNNING: i32 = 0;
713 0 : match status_exitcode.code() {
714 0 : Some(PG_STATUS_NOT_RUNNING) => Ok(false),
715 0 : Some(PG_NO_DATA_DIR) => Ok(false),
716 0 : Some(PG_STATUS_RUNNING) => Ok(true),
717 0 : Some(code) => Err(anyhow::anyhow!(
718 0 : "pg_ctl status returned unexpected status code: {:?}",
719 0 : code
720 0 : )),
721 0 : None => Err(anyhow::anyhow!("pg_ctl status returned no status code")),
722 : }
723 0 : }
724 :
725 0 : fn get_claims_for_path(path: &str) -> anyhow::Result<Option<Claims>> {
726 0 : let category = match path.find('/') {
727 0 : Some(idx) => &path[..idx],
728 0 : None => path,
729 : };
730 :
731 0 : match category {
732 0 : "status" | "ready" => Ok(None),
733 0 : "control" | "debug" => Ok(Some(Claims::new(None, Scope::Admin))),
734 0 : "v1" => Ok(Some(Claims::new(None, Scope::PageServerApi))),
735 0 : _ => Err(anyhow::anyhow!("Failed to determine claims for {}", path)),
736 : }
737 0 : }
738 :
739 : /// Simple HTTP request wrapper for calling into storage controller
740 0 : async fn dispatch<RQ, RS>(
741 0 : &self,
742 0 : method: reqwest::Method,
743 0 : path: String,
744 0 : body: Option<RQ>,
745 0 : ) -> anyhow::Result<RS>
746 0 : where
747 0 : RQ: Serialize + Sized,
748 0 : RS: DeserializeOwned + Sized,
749 0 : {
750 : // In the special case of the `storage_controller start` subcommand, we wish
751 : // to use the API endpoint of the newly started storage controller in order
752 : // to pass the readiness check. In this scenario [`Self::listen_port`] will
753 : // be set (see [`Self::start`]).
754 : //
755 : // Otherwise, we infer the storage controller api endpoint from the configured
756 : // control plane API.
757 0 : let port = if let Some(port) = self.listen_port.get() {
758 0 : *port
759 : } else {
760 0 : self.env.control_plane_api.port().unwrap()
761 : };
762 :
763 : // The configured URL has the /upcall path prefix for pageservers to use: we will strip that out
764 : // for general purpose API access.
765 0 : let url = Url::from_str(&format!(
766 0 : "{}://{}:{port}/{path}",
767 0 : self.env.control_plane_api.scheme(),
768 0 : self.env.control_plane_api.host_str().unwrap(),
769 0 : ))
770 0 : .unwrap();
771 0 :
772 0 : let mut builder = self.client.request(method, url);
773 0 : if let Some(body) = body {
774 0 : builder = builder.json(&body)
775 0 : }
776 0 : if let Some(private_key) = &self.private_key {
777 0 : println!("Getting claims for path {}", path);
778 0 : if let Some(required_claims) = Self::get_claims_for_path(&path)? {
779 0 : println!("Got claims {:?} for path {}", required_claims, path);
780 0 : let jwt_token = encode_from_key_file(&required_claims, private_key)?;
781 0 : builder = builder.header(
782 0 : reqwest::header::AUTHORIZATION,
783 0 : format!("Bearer {jwt_token}"),
784 0 : );
785 0 : }
786 0 : }
787 :
788 0 : let response = builder.send().await?;
789 0 : let response = response.error_from_body().await?;
790 :
791 0 : Ok(response
792 0 : .json()
793 0 : .await
794 0 : .map_err(pageserver_client::mgmt_api::Error::ReceiveBody)?)
795 0 : }
796 :
797 : /// Call into the attach_hook API, for use before handing out attachments to pageservers
798 : #[instrument(skip(self))]
799 : pub async fn attach_hook(
800 : &self,
801 : tenant_shard_id: TenantShardId,
802 : pageserver_id: NodeId,
803 : ) -> anyhow::Result<Option<u32>> {
804 : let request = AttachHookRequest {
805 : tenant_shard_id,
806 : node_id: Some(pageserver_id),
807 : generation_override: None,
808 : };
809 :
810 : let response = self
811 : .dispatch::<_, AttachHookResponse>(
812 : Method::POST,
813 : "debug/v1/attach-hook".to_string(),
814 : Some(request),
815 : )
816 : .await?;
817 :
818 : Ok(response.generation)
819 : }
820 :
821 : #[instrument(skip(self))]
822 : pub async fn inspect(
823 : &self,
824 : tenant_shard_id: TenantShardId,
825 : ) -> anyhow::Result<Option<(u32, NodeId)>> {
826 : let request = InspectRequest { tenant_shard_id };
827 :
828 : let response = self
829 : .dispatch::<_, InspectResponse>(
830 : Method::POST,
831 : "debug/v1/inspect".to_string(),
832 : Some(request),
833 : )
834 : .await?;
835 :
836 : Ok(response.attachment)
837 : }
838 :
839 : #[instrument(skip(self))]
840 : pub async fn tenant_create(
841 : &self,
842 : req: TenantCreateRequest,
843 : ) -> anyhow::Result<TenantCreateResponse> {
844 : self.dispatch(Method::POST, "v1/tenant".to_string(), Some(req))
845 : .await
846 : }
847 :
848 : #[instrument(skip(self))]
849 : pub async fn tenant_import(&self, tenant_id: TenantId) -> anyhow::Result<TenantCreateResponse> {
850 : self.dispatch::<(), TenantCreateResponse>(
851 : Method::POST,
852 : format!("debug/v1/tenant/{tenant_id}/import"),
853 : None,
854 : )
855 : .await
856 : }
857 :
858 : #[instrument(skip(self))]
859 : pub async fn tenant_locate(&self, tenant_id: TenantId) -> anyhow::Result<TenantLocateResponse> {
860 : self.dispatch::<(), _>(
861 : Method::GET,
862 : format!("debug/v1/tenant/{tenant_id}/locate"),
863 : None,
864 : )
865 : .await
866 : }
867 :
868 : #[instrument(skip_all, fields(node_id=%req.node_id))]
869 : pub async fn node_register(&self, req: NodeRegisterRequest) -> anyhow::Result<()> {
870 : self.dispatch::<_, ()>(Method::POST, "control/v1/node".to_string(), Some(req))
871 : .await
872 : }
873 :
874 : #[instrument(skip_all, fields(node_id=%req.node_id))]
875 : pub async fn node_configure(&self, req: NodeConfigureRequest) -> anyhow::Result<()> {
876 : self.dispatch::<_, ()>(
877 : Method::PUT,
878 : format!("control/v1/node/{}/config", req.node_id),
879 : Some(req),
880 : )
881 : .await
882 : }
883 :
884 0 : pub async fn node_list(&self) -> anyhow::Result<Vec<NodeDescribeResponse>> {
885 0 : self.dispatch::<(), Vec<NodeDescribeResponse>>(
886 0 : Method::GET,
887 0 : "control/v1/node".to_string(),
888 0 : None,
889 0 : )
890 0 : .await
891 0 : }
892 :
893 : #[instrument(skip(self))]
894 : pub async fn ready(&self) -> anyhow::Result<()> {
895 : self.dispatch::<(), ()>(Method::GET, "ready".to_string(), None)
896 : .await
897 : }
898 :
899 : #[instrument(skip_all, fields(%tenant_id, timeline_id=%req.new_timeline_id))]
900 : pub async fn tenant_timeline_create(
901 : &self,
902 : tenant_id: TenantId,
903 : req: TimelineCreateRequest,
904 : ) -> anyhow::Result<TimelineInfo> {
905 : self.dispatch(
906 : Method::POST,
907 : format!("v1/tenant/{tenant_id}/timeline"),
908 : Some(req),
909 : )
910 : .await
911 : }
912 :
913 0 : pub async fn set_tenant_config(&self, req: &TenantConfigRequest) -> anyhow::Result<()> {
914 0 : self.dispatch(Method::PUT, "v1/tenant/config".to_string(), Some(req))
915 0 : .await
916 0 : }
917 : }
|
