Line data Source code
1 : //! Main entry point for the Page Server executable.
2 :
3 : use std::env::{var, VarError};
4 : use std::sync::Arc;
5 : use std::time::Duration;
6 : use std::{env, ops::ControlFlow, str::FromStr};
7 :
8 : use anyhow::{anyhow, Context};
9 : use camino::Utf8Path;
10 : use clap::{Arg, ArgAction, Command};
11 :
12 : use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
13 : use pageserver::control_plane_client::ControlPlaneClient;
14 : use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
15 : use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
16 : use pageserver::task_mgr::WALRECEIVER_RUNTIME;
17 : use pageserver::tenant::{secondary, TenantSharedResources};
18 : use remote_storage::GenericRemoteStorage;
19 : use tokio::time::Instant;
20 : use tracing::*;
21 :
22 : use metrics::set_build_info_metric;
23 : use pageserver::{
24 : config::{defaults::*, PageServerConf},
25 : context::{DownloadBehavior, RequestContext},
26 : deletion_queue::DeletionQueue,
27 : http, page_cache, page_service, task_mgr,
28 : task_mgr::TaskKind,
29 : task_mgr::{BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME},
30 : tenant::mgr,
31 : virtual_file,
32 : };
33 : use postgres_backend::AuthType;
34 : use utils::failpoint_support;
35 : use utils::logging::TracingErrorLayerEnablement;
36 : use utils::{
37 : auth::{JwtAuth, SwappableJwtAuth},
38 : logging, project_build_tag, project_git_version,
39 : sentry_init::init_sentry,
40 : tcp_listener,
41 : };
42 :
43 : project_git_version!(GIT_VERSION);
44 : project_build_tag!(BUILD_TAG);
45 :
46 : const PID_FILE_NAME: &str = "pageserver.pid";
47 :
48 : const FEATURES: &[&str] = &[
49 : #[cfg(feature = "testing")]
50 : "testing",
51 : ];
52 :
53 1990 : fn version() -> String {
54 1990 : format!(
55 1990 : "{GIT_VERSION} failpoints: {}, features: {:?}",
56 1990 : fail::has_failpoints(),
57 1990 : FEATURES,
58 1990 : )
59 1990 : }
60 :
61 1384 : fn main() -> anyhow::Result<()> {
62 1384 : let launch_ts = Box::leak(Box::new(LaunchTimestamp::generate()));
63 1384 :
64 1384 : let arg_matches = cli().get_matches();
65 1384 :
66 1384 : if arg_matches.get_flag("enabled-features") {
67 0 : println!("{{\"features\": {FEATURES:?} }}");
68 0 : return Ok(());
69 1384 : }
70 1384 :
71 1384 : let workdir = arg_matches
72 1384 : .get_one::<String>("workdir")
73 1384 : .map(Utf8Path::new)
74 1384 : .unwrap_or_else(|| Utf8Path::new(".neon"));
75 1384 : let workdir = workdir
76 1384 : .canonicalize_utf8()
77 1384 : .with_context(|| format!("Error opening workdir '{workdir}'"))?;
78 :
79 1384 : let cfg_file_path = workdir.join("pageserver.toml");
80 1384 :
81 1384 : // Set CWD to workdir for non-daemon modes
82 1384 : env::set_current_dir(&workdir)
83 1384 : .with_context(|| format!("Failed to set application's current dir to '{workdir}'"))?;
84 :
85 1384 : let conf = match initialize_config(&cfg_file_path, arg_matches, &workdir)? {
86 604 : ControlFlow::Continue(conf) => conf,
87 : ControlFlow::Break(()) => {
88 389 : info!("Pageserver config init successful");
89 389 : return Ok(());
90 : }
91 : };
92 :
93 : // Initialize logging.
94 : //
95 : // It must be initialized before the custom panic hook is installed below.
96 : //
97 : // Regarding tracing_error enablement: at this time, we only use the
98 : // tracing_error crate to debug_assert that log spans contain tenant and timeline ids.
99 : // See `debug_assert_current_span_has_tenant_and_timeline_id` in the timeline module
100 604 : let tracing_error_layer_enablement = if cfg!(debug_assertions) {
101 604 : TracingErrorLayerEnablement::EnableWithRustLogFilter
102 : } else {
103 0 : TracingErrorLayerEnablement::Disabled
104 : };
105 604 : logging::init(
106 604 : conf.log_format,
107 604 : tracing_error_layer_enablement,
108 604 : logging::Output::Stdout,
109 604 : )?;
110 :
111 : // mind the order required here: 1. logging, 2. panic_hook, 3. sentry.
112 : // disarming this hook on pageserver, because we never tear down tracing.
113 604 : logging::replace_panic_hook_with_tracing_panic_hook().forget();
114 604 :
115 604 : // initialize sentry if SENTRY_DSN is provided
116 604 : let _sentry_guard = init_sentry(
117 604 : Some(GIT_VERSION.into()),
118 604 : &[("node_id", &conf.id.to_string())],
119 604 : );
120 604 :
121 604 : let tenants_path = conf.tenants_path();
122 604 : if !tenants_path.exists() {
123 387 : utils::crashsafe::create_dir_all(conf.tenants_path())
124 387 : .with_context(|| format!("Failed to create tenants root dir at '{tenants_path}'"))?;
125 217 : }
126 :
127 : // Initialize up failpoints support
128 604 : let scenario = failpoint_support::init();
129 604 :
130 604 : // Basic initialization of things that don't change after startup
131 604 : virtual_file::init(conf.max_file_descriptors, conf.virtual_file_io_engine);
132 604 : page_cache::init(conf.page_cache_size);
133 604 :
134 604 : start_pageserver(launch_ts, conf).context("Failed to start pageserver")?;
135 :
136 0 : scenario.teardown();
137 0 : Ok(())
138 392 : }
139 :
140 996 : fn initialize_config(
141 996 : cfg_file_path: &Utf8Path,
142 996 : arg_matches: clap::ArgMatches,
143 996 : workdir: &Utf8Path,
144 996 : ) -> anyhow::Result<ControlFlow<(), &'static PageServerConf>> {
145 996 : let init = arg_matches.get_flag("init");
146 996 : let update_config = init || arg_matches.get_flag("update-config");
147 :
148 996 : let (mut toml, config_file_exists) = if cfg_file_path.is_file() {
149 606 : if init {
150 1 : anyhow::bail!(
151 1 : "Config file '{cfg_file_path}' already exists, cannot init it, use --update-config to update it",
152 1 : );
153 605 : }
154 : // Supplement the CLI arguments with the config file
155 605 : let cfg_file_contents = std::fs::read_to_string(cfg_file_path)
156 605 : .with_context(|| format!("Failed to read pageserver config at '{cfg_file_path}'"))?;
157 : (
158 605 : cfg_file_contents
159 605 : .parse::<toml_edit::Document>()
160 605 : .with_context(|| {
161 0 : format!("Failed to parse '{cfg_file_path}' as pageserver config")
162 605 : })?,
163 : true,
164 : )
165 390 : } else if cfg_file_path.exists() {
166 0 : anyhow::bail!("Config file '{cfg_file_path}' exists but is not a regular file");
167 : } else {
168 : // We're initializing the tenant, so there's no config file yet
169 : (
170 390 : DEFAULT_CONFIG_FILE
171 390 : .parse::<toml_edit::Document>()
172 390 : .context("could not parse built-in config file")?,
173 : false,
174 : )
175 : };
176 :
177 995 : if let Some(values) = arg_matches.get_many::<String>("config-override") {
178 10136 : for option_line in values {
179 9142 : let doc = toml_edit::Document::from_str(option_line).with_context(|| {
180 0 : format!("Option '{option_line}' could not be parsed as a toml document")
181 9142 : })?;
182 :
183 9168 : for (key, item) in doc.iter() {
184 9168 : if config_file_exists && update_config && key == "id" && toml.contains_key(key) {
185 1 : anyhow::bail!("Pageserver config file exists at '{cfg_file_path}' and has node id already, it cannot be overridden");
186 9167 : }
187 9167 : toml.insert(key, item.clone());
188 : }
189 : }
190 0 : }
191 :
192 994 : debug!("Resulting toml: {toml}");
193 994 : let conf = PageServerConf::parse_and_validate(&toml, workdir)
194 994 : .context("Failed to parse pageserver configuration")?;
195 :
196 993 : if update_config {
197 389 : info!("Writing pageserver config to '{cfg_file_path}'");
198 :
199 389 : std::fs::write(cfg_file_path, toml.to_string())
200 389 : .with_context(|| format!("Failed to write pageserver config to '{cfg_file_path}'"))?;
201 389 : info!("Config successfully written to '{cfg_file_path}'")
202 604 : }
203 :
204 993 : Ok(if init {
205 389 : ControlFlow::Break(())
206 : } else {
207 604 : ControlFlow::Continue(Box::leak(Box::new(conf)))
208 : })
209 996 : }
210 :
211 : struct WaitForPhaseResult<F: std::future::Future + Unpin> {
212 : timeout_remaining: Duration,
213 : skipped: Option<F>,
214 : }
215 :
216 : /// During startup, we apply a timeout to our waits for readiness, to avoid
217 : /// stalling the whole service if one Tenant experiences some problem. Each
218 : /// phase may consume some of the timeout: this function returns the updated
219 : /// timeout for use in the next call.
220 1201 : async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
221 1201 : where
222 1201 : F: std::future::Future + Unpin,
223 1201 : {
224 1201 : let initial_t = Instant::now();
225 1201 : let skipped = match tokio::time::timeout(timeout, &mut fut).await {
226 1190 : Ok(_) => None,
227 : Err(_) => {
228 4 : tracing::info!(
229 4 : timeout_millis = timeout.as_millis(),
230 4 : %phase,
231 4 : "Startup phase timed out, proceeding anyway"
232 4 : );
233 4 : Some(fut)
234 : }
235 : };
236 :
237 1194 : WaitForPhaseResult {
238 1194 : timeout_remaining: timeout
239 1194 : .checked_sub(Instant::now().duration_since(initial_t))
240 1194 : .unwrap_or(Duration::ZERO),
241 1194 : skipped,
242 1194 : }
243 1194 : }
244 :
245 2990 : fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
246 2990 : let elapsed = started_at.elapsed();
247 2990 : let secs = elapsed.as_secs_f64();
248 2990 : STARTUP_DURATION.with_label_values(&[phase]).set(secs);
249 2990 :
250 2990 : info!(
251 2990 : elapsed_ms = elapsed.as_millis(),
252 2990 : "{human_phase} ({secs:.3}s since start)"
253 2990 : )
254 2990 : }
255 :
256 604 : fn start_pageserver(
257 604 : launch_ts: &'static LaunchTimestamp,
258 604 : conf: &'static PageServerConf,
259 604 : ) -> anyhow::Result<()> {
260 604 : // Monotonic time for later calculating startup duration
261 604 : let started_startup_at = Instant::now();
262 604 :
263 604 : // Print version and launch timestamp to the log,
264 604 : // and expose them as prometheus metrics.
265 604 : // A changed version string indicates changed software.
266 604 : // A changed launch timestamp indicates a pageserver restart.
267 604 : info!(
268 604 : "version: {} launch_timestamp: {} build_tag: {}",
269 604 : version(),
270 604 : launch_ts.to_string(),
271 604 : BUILD_TAG,
272 604 : );
273 604 : set_build_info_metric(GIT_VERSION, BUILD_TAG);
274 604 : set_launch_timestamp_metric(launch_ts);
275 604 : pageserver::preinitialize_metrics();
276 604 :
277 604 : // If any failpoints were set from FAILPOINTS environment variable,
278 604 : // print them to the log for debugging purposes
279 604 : let failpoints = fail::list();
280 604 : if !failpoints.is_empty() {
281 8 : info!(
282 8 : "started with failpoints: {}",
283 8 : failpoints
284 8 : .iter()
285 9 : .map(|(name, actions)| format!("{name}={actions}"))
286 8 : .collect::<Vec<String>>()
287 8 : .join(";")
288 8 : )
289 596 : }
290 :
291 : // Create and lock PID file. This ensures that there cannot be more than one
292 : // pageserver process running at the same time.
293 604 : let lock_file_path = conf.workdir.join(PID_FILE_NAME);
294 604 : let lock_file =
295 604 : utils::pid_file::claim_for_current_process(&lock_file_path).context("claim pid file")?;
296 604 : info!("Claimed pid file at {lock_file_path:?}");
297 :
298 : // Ensure that the lock file is held even if the main thread of the process panics.
299 : // We need to release the lock file only when the process exits.
300 604 : std::mem::forget(lock_file);
301 604 :
302 604 : // Bind the HTTP and libpq ports early, so that if they are in use by some other
303 604 : // process, we error out early.
304 604 : let http_addr = &conf.listen_http_addr;
305 604 : info!("Starting pageserver http handler on {http_addr}");
306 604 : let http_listener = tcp_listener::bind(http_addr)?;
307 :
308 604 : let pg_addr = &conf.listen_pg_addr;
309 604 : info!("Starting pageserver pg protocol handler on {pg_addr}");
310 604 : let pageserver_listener = tcp_listener::bind(pg_addr)?;
311 :
312 : // Launch broker client
313 : // The storage_broker::connect call needs to happen inside a tokio runtime thread.
314 604 : let broker_client = WALRECEIVER_RUNTIME
315 604 : .block_on(async {
316 604 : // Note: we do not attempt connecting here (but validate endpoints sanity).
317 604 : storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)
318 604 : })
319 604 : .with_context(|| {
320 0 : format!(
321 0 : "create broker client for uri={:?} keepalive_interval={:?}",
322 0 : &conf.broker_endpoint, conf.broker_keepalive_interval,
323 0 : )
324 604 : })?;
325 :
326 : // Initialize authentication for incoming connections
327 : let http_auth;
328 : let pg_auth;
329 604 : if conf.http_auth_type == AuthType::NeonJWT || conf.pg_auth_type == AuthType::NeonJWT {
330 : // unwrap is ok because check is performed when creating config, so path is set and exists
331 11 : let key_path = conf.auth_validation_public_key_path.as_ref().unwrap();
332 11 : info!("Loading public key(s) for verifying JWT tokens from {key_path:?}");
333 :
334 11 : let jwt_auth = JwtAuth::from_key_path(key_path)?;
335 11 : let auth: Arc<SwappableJwtAuth> = Arc::new(SwappableJwtAuth::new(jwt_auth));
336 :
337 11 : http_auth = match &conf.http_auth_type {
338 0 : AuthType::Trust => None,
339 11 : AuthType::NeonJWT => Some(auth.clone()),
340 : };
341 11 : pg_auth = match &conf.pg_auth_type {
342 0 : AuthType::Trust => None,
343 11 : AuthType::NeonJWT => Some(auth),
344 : };
345 593 : } else {
346 593 : http_auth = None;
347 593 : pg_auth = None;
348 593 : }
349 604 : info!("Using auth for http API: {:#?}", conf.http_auth_type);
350 604 : info!("Using auth for pg connections: {:#?}", conf.pg_auth_type);
351 :
352 593 : match var("NEON_AUTH_TOKEN") {
353 11 : Ok(v) => {
354 11 : info!("Loaded JWT token for authentication with Safekeeper");
355 11 : pageserver::config::SAFEKEEPER_AUTH_TOKEN
356 11 : .set(Arc::new(v))
357 11 : .map_err(|_| anyhow!("Could not initialize SAFEKEEPER_AUTH_TOKEN"))?;
358 : }
359 : Err(VarError::NotPresent) => {
360 593 : info!("No JWT token for authentication with Safekeeper detected");
361 : }
362 0 : Err(e) => {
363 0 : return Err(e).with_context(|| {
364 0 : "Failed to either load to detect non-present NEON_AUTH_TOKEN environment variable"
365 0 : })
366 : }
367 : };
368 :
369 : // Top-level cancellation token for the process
370 604 : let shutdown_pageserver = tokio_util::sync::CancellationToken::new();
371 :
372 : // Set up remote storage client
373 604 : let remote_storage = create_remote_storage_client(conf)?;
374 :
375 : // Set up deletion queue
376 604 : let (deletion_queue, deletion_workers) = DeletionQueue::new(
377 604 : remote_storage.clone(),
378 604 : ControlPlaneClient::new(conf, &shutdown_pageserver),
379 604 : conf,
380 604 : );
381 604 : if let Some(deletion_workers) = deletion_workers {
382 604 : deletion_workers.spawn_with(BACKGROUND_RUNTIME.handle());
383 604 : }
384 :
385 : // Up to this point no significant I/O has been done: this should have been fast. Record
386 : // duration prior to starting I/O intensive phase of startup.
387 604 : startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
388 604 : STARTUP_IS_LOADING.set(1);
389 604 :
390 604 : // Startup staging or optimizing:
391 604 : //
392 604 : // We want to minimize downtime for `page_service` connections, and trying not to overload
393 604 : // BACKGROUND_RUNTIME by doing initial compactions and initial logical sizes at the same time.
394 604 : //
395 604 : // init_done_rx will notify when all initial load operations have completed.
396 604 : //
397 604 : // background_jobs_can_start (same name used to hold off background jobs from starting at
398 604 : // consumer side) will be dropped once we can start the background jobs. Currently it is behind
399 604 : // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
400 604 : // (background_task_maximum_delay).
401 604 : let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
402 604 : let (init_done_tx, init_done_rx) = utils::completion::channel();
403 604 :
404 604 : let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
405 604 :
406 604 : let order = pageserver::InitializationOrder {
407 604 : initial_tenant_load_remote: Some(init_done_tx),
408 604 : initial_tenant_load: Some(init_remote_done_tx),
409 604 : background_jobs_can_start: background_jobs_barrier.clone(),
410 604 : };
411 604 :
412 604 : // Scan the local 'tenants/' directory and start loading the tenants
413 604 : let deletion_queue_client = deletion_queue.new_client();
414 604 : let tenant_manager = BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
415 604 : conf,
416 604 : TenantSharedResources {
417 604 : broker_client: broker_client.clone(),
418 604 : remote_storage: remote_storage.clone(),
419 604 : deletion_queue_client,
420 604 : },
421 604 : order,
422 604 : shutdown_pageserver.clone(),
423 604 : ))?;
424 604 : let tenant_manager = Arc::new(tenant_manager);
425 604 :
426 604 : BACKGROUND_RUNTIME.spawn({
427 604 : let shutdown_pageserver = shutdown_pageserver.clone();
428 604 : let drive_init = async move {
429 604 : // NOTE: unlike many futures in pageserver, this one is cancellation-safe
430 604 : let guard = scopeguard::guard_on_success((), |_| {
431 0 : tracing::info!("Cancelled before initial load completed")
432 604 : });
433 604 :
434 604 : let timeout = conf.background_task_maximum_delay;
435 604 :
436 604 : let init_remote_done = std::pin::pin!(async {
437 604 : init_remote_done_rx.wait().await;
438 596 : startup_checkpoint(
439 596 : started_startup_at,
440 596 : "initial_tenant_load_remote",
441 596 : "Remote part of initial load completed",
442 596 : );
443 604 : });
444 :
445 : let WaitForPhaseResult {
446 597 : timeout_remaining: timeout,
447 597 : skipped: init_remote_skipped,
448 604 : } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
449 :
450 597 : let init_load_done = std::pin::pin!(async {
451 597 : init_done_rx.wait().await;
452 597 : startup_checkpoint(
453 597 : started_startup_at,
454 597 : "initial_tenant_load",
455 597 : "Initial load completed",
456 597 : );
457 597 : STARTUP_IS_LOADING.set(0);
458 597 : });
459 :
460 : let WaitForPhaseResult {
461 597 : timeout_remaining: _timeout,
462 597 : skipped: init_load_skipped,
463 597 : } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
464 :
465 : // initial logical sizes can now start, as they were waiting on init_done_rx.
466 :
467 597 : scopeguard::ScopeGuard::into_inner(guard);
468 597 :
469 597 : // allow background jobs to start: we either completed prior stages, or they reached timeout
470 597 : // and were skipped. It is important that we do not let them block background jobs indefinitely,
471 597 : // because things like consumption metrics for billing are blocked by this barrier.
472 597 : drop(background_jobs_can_start);
473 597 : startup_checkpoint(
474 597 : started_startup_at,
475 597 : "background_jobs_can_start",
476 597 : "Starting background jobs",
477 597 : );
478 597 :
479 597 : // We are done. If we skipped any phases due to timeout, run them to completion here so that
480 597 : // they will eventually update their startup_checkpoint, and so that we do not declare the
481 597 : // 'complete' stage until all the other stages are really done.
482 597 : let guard = scopeguard::guard_on_success((), |_| {
483 0 : tracing::info!("Cancelled before waiting for skipped phases done")
484 597 : });
485 597 : if let Some(f) = init_remote_skipped {
486 4 : f.await;
487 593 : }
488 596 : if let Some(f) = init_load_skipped {
489 0 : f.await;
490 596 : }
491 596 : scopeguard::ScopeGuard::into_inner(guard);
492 596 :
493 596 : startup_checkpoint(started_startup_at, "complete", "Startup complete");
494 604 : };
495 604 :
496 604 : async move {
497 604 : let mut drive_init = std::pin::pin!(drive_init);
498 604 : // just race these tasks
499 775 : tokio::select! {
500 775 : _ = shutdown_pageserver.cancelled() => {},
501 775 : _ = &mut drive_init => {},
502 775 : }
503 604 : }
504 604 : });
505 :
506 604 : let secondary_controller = if let Some(remote_storage) = &remote_storage {
507 604 : secondary::spawn_tasks(
508 604 : tenant_manager.clone(),
509 604 : remote_storage.clone(),
510 604 : background_jobs_barrier.clone(),
511 604 : shutdown_pageserver.clone(),
512 604 : )
513 : } else {
514 0 : secondary::null_controller()
515 : };
516 :
517 : // shared state between the disk-usage backed eviction background task and the http endpoint
518 : // that allows triggering disk-usage based eviction manually. note that the http endpoint
519 : // is still accessible even if background task is not configured as long as remote storage has
520 : // been configured.
521 604 : let disk_usage_eviction_state: Arc<disk_usage_eviction_task::State> = Arc::default();
522 :
523 604 : if let Some(remote_storage) = &remote_storage {
524 604 : launch_disk_usage_global_eviction_task(
525 604 : conf,
526 604 : remote_storage.clone(),
527 604 : disk_usage_eviction_state.clone(),
528 604 : tenant_manager.clone(),
529 604 : background_jobs_barrier.clone(),
530 604 : )?;
531 0 : }
532 :
533 : // Start up the service to handle HTTP mgmt API request. We created the
534 : // listener earlier already.
535 : {
536 604 : let _rt_guard = MGMT_REQUEST_RUNTIME.enter();
537 :
538 604 : let router_state = Arc::new(
539 604 : http::routes::State::new(
540 604 : conf,
541 604 : tenant_manager,
542 604 : http_auth.clone(),
543 604 : remote_storage.clone(),
544 604 : broker_client.clone(),
545 604 : disk_usage_eviction_state,
546 604 : deletion_queue.new_client(),
547 604 : secondary_controller,
548 604 : )
549 604 : .context("Failed to initialize router state")?,
550 : );
551 604 : let router = http::make_router(router_state, launch_ts, http_auth.clone())?
552 604 : .build()
553 604 : .map_err(|err| anyhow!(err))?;
554 604 : let service = utils::http::RouterService::new(router).unwrap();
555 604 : let server = hyper::Server::from_tcp(http_listener)?
556 604 : .serve(service)
557 604 : .with_graceful_shutdown(task_mgr::shutdown_watcher());
558 604 :
559 604 : task_mgr::spawn(
560 604 : MGMT_REQUEST_RUNTIME.handle(),
561 604 : TaskKind::HttpEndpointListener,
562 604 : None,
563 604 : None,
564 604 : "http endpoint listener",
565 604 : true,
566 604 : async {
567 4477 : server.await?;
568 172 : Ok(())
569 604 : },
570 604 : );
571 : }
572 :
573 604 : if let Some(metric_collection_endpoint) = &conf.metric_collection_endpoint {
574 6 : let metrics_ctx = RequestContext::todo_child(
575 6 : TaskKind::MetricsCollection,
576 6 : // This task itself shouldn't download anything.
577 6 : // The actual size calculation does need downloads, and
578 6 : // creates a child context with the right DownloadBehavior.
579 6 : DownloadBehavior::Error,
580 6 : );
581 6 :
582 6 : let local_disk_storage = conf.workdir.join("last_consumption_metrics.json");
583 6 :
584 6 : task_mgr::spawn(
585 6 : crate::BACKGROUND_RUNTIME.handle(),
586 6 : TaskKind::MetricsCollection,
587 6 : None,
588 6 : None,
589 6 : "consumption metrics collection",
590 6 : true,
591 6 : async move {
592 6 : // first wait until background jobs are cleared to launch.
593 6 : //
594 6 : // this is because we only process active tenants and timelines, and the
595 6 : // Timeline::get_current_logical_size will spawn the logical size calculation,
596 6 : // which will not be rate-limited.
597 6 : let cancel = task_mgr::shutdown_token();
598 6 :
599 9 : tokio::select! {
600 9 : _ = cancel.cancelled() => { return Ok(()); },
601 9 : _ = background_jobs_barrier.wait() => {}
602 9 : };
603 :
604 6 : pageserver::consumption_metrics::collect_metrics(
605 6 : metric_collection_endpoint,
606 6 : conf.metric_collection_interval,
607 6 : conf.cached_metric_collection_interval,
608 6 : conf.synthetic_size_calculation_interval,
609 6 : conf.id,
610 6 : local_disk_storage,
611 6 : cancel,
612 6 : metrics_ctx,
613 6 : )
614 6 : .instrument(info_span!("metrics_collection"))
615 242 : .await?;
616 3 : Ok(())
617 6 : },
618 6 : );
619 598 : }
620 :
621 : // Spawn a task to listen for libpq connections. It will spawn further tasks
622 : // for each connection. We created the listener earlier already.
623 604 : {
624 604 : let libpq_ctx = RequestContext::todo_child(
625 604 : TaskKind::LibpqEndpointListener,
626 604 : // listener task shouldn't need to download anything. (We will
627 604 : // create a separate sub-contexts for each connection, with their
628 604 : // own download behavior. This context is used only to listen and
629 604 : // accept connections.)
630 604 : DownloadBehavior::Error,
631 604 : );
632 604 : task_mgr::spawn(
633 604 : COMPUTE_REQUEST_RUNTIME.handle(),
634 604 : TaskKind::LibpqEndpointListener,
635 604 : None,
636 604 : None,
637 604 : "libpq endpoint listener",
638 604 : true,
639 604 : async move {
640 604 : page_service::libpq_listener_main(
641 604 : conf,
642 604 : broker_client,
643 604 : pg_auth,
644 604 : pageserver_listener,
645 604 : conf.pg_auth_type,
646 604 : libpq_ctx,
647 604 : task_mgr::shutdown_token(),
648 604 : )
649 10595 : .await
650 604 : },
651 604 : );
652 604 : }
653 604 :
654 604 : let mut shutdown_pageserver = Some(shutdown_pageserver.drop_guard());
655 604 :
656 604 : // All started up! Now just sit and wait for shutdown signal.
657 604 : {
658 604 : use signal_hook::consts::*;
659 604 : let signal_handler = BACKGROUND_RUNTIME.spawn_blocking(move || {
660 604 : let mut signals =
661 604 : signal_hook::iterator::Signals::new([SIGINT, SIGTERM, SIGQUIT]).unwrap();
662 604 : return signals
663 604 : .forever()
664 604 : .next()
665 604 : .expect("forever() never returns None unless explicitly closed");
666 604 : });
667 604 : let signal = BACKGROUND_RUNTIME
668 604 : .block_on(signal_handler)
669 604 : .expect("join error");
670 604 : match signal {
671 : SIGQUIT => {
672 432 : info!("Got signal {signal}. Terminating in immediate shutdown mode",);
673 428 : std::process::exit(111);
674 : }
675 : SIGINT | SIGTERM => {
676 172 : info!("Got signal {signal}. Terminating gracefully in fast shutdown mode",);
677 :
678 : // This cancels the `shutdown_pageserver` cancellation tree.
679 : // Right now that tree doesn't reach very far, and `task_mgr` is used instead.
680 : // The plan is to change that over time.
681 172 : shutdown_pageserver.take();
682 172 : let bg_remote_storage = remote_storage.clone();
683 172 : let bg_deletion_queue = deletion_queue.clone();
684 172 : BACKGROUND_RUNTIME.block_on(pageserver::shutdown_pageserver(
685 172 : bg_remote_storage.map(|_| bg_deletion_queue),
686 172 : 0,
687 172 : ));
688 172 : unreachable!()
689 : }
690 0 : _ => unreachable!(),
691 : }
692 : }
693 0 : }
694 :
695 604 : fn create_remote_storage_client(
696 604 : conf: &'static PageServerConf,
697 604 : ) -> anyhow::Result<Option<GenericRemoteStorage>> {
698 604 : let config = if let Some(config) = &conf.remote_storage_config {
699 604 : config
700 : } else {
701 0 : tracing::warn!("no remote storage configured, this is a deprecated configuration");
702 0 : return Ok(None);
703 : };
704 :
705 : // Create the client
706 604 : let mut remote_storage = GenericRemoteStorage::from_config(config)?;
707 :
708 : // If `test_remote_failures` is non-zero, wrap the client with a
709 : // wrapper that simulates failures.
710 604 : if conf.test_remote_failures > 0 {
711 63 : if !cfg!(feature = "testing") {
712 0 : anyhow::bail!("test_remote_failures option is not available because pageserver was compiled without the 'testing' feature");
713 63 : }
714 63 : info!(
715 63 : "Simulating remote failures for first {} attempts of each op",
716 63 : conf.test_remote_failures
717 63 : );
718 63 : remote_storage =
719 63 : GenericRemoteStorage::unreliable_wrapper(remote_storage, conf.test_remote_failures);
720 541 : }
721 :
722 604 : Ok(Some(remote_storage))
723 604 : }
724 :
725 1386 : fn cli() -> Command {
726 1386 : Command::new("Neon page server")
727 1386 : .about("Materializes WAL stream to pages and serves them to the postgres")
728 1386 : .version(version())
729 1386 : .arg(
730 1386 : Arg::new("init")
731 1386 : .long("init")
732 1386 : .action(ArgAction::SetTrue)
733 1386 : .help("Initialize pageserver with all given config overrides"),
734 1386 : )
735 1386 : .arg(
736 1386 : Arg::new("workdir")
737 1386 : .short('D')
738 1386 : .long("workdir")
739 1386 : .help("Working directory for the pageserver"),
740 1386 : )
741 1386 : // See `settings.md` for more details on the extra configuration patameters pageserver can process
742 1386 : .arg(
743 1386 : Arg::new("config-override")
744 1386 : .short('c')
745 1386 : .num_args(1)
746 1386 : .action(ArgAction::Append)
747 1386 : .help("Additional configuration overrides of the ones from the toml config file (or new ones to add there). \
748 1386 : Any option has to be a valid toml document, example: `-c=\"foo='hey'\"` `-c=\"foo={value=1}\"`"),
749 1386 : )
750 1386 : .arg(
751 1386 : Arg::new("update-config")
752 1386 : .long("update-config")
753 1386 : .action(ArgAction::SetTrue)
754 1386 : .help("Update the config file when started"),
755 1386 : )
756 1386 : .arg(
757 1386 : Arg::new("enabled-features")
758 1386 : .long("enabled-features")
759 1386 : .action(ArgAction::SetTrue)
760 1386 : .help("Show enabled compile time features"),
761 1386 : )
762 1386 : }
763 :
764 2 : #[test]
765 2 : fn verify_cli() {
766 2 : cli().debug_assert();
767 2 : }
|
