Line data Source code
1 : #![recursion_limit = "300"]
2 :
3 : //! Main entry point for the Page Server executable.
4 :
5 : use std::env;
6 : use std::env::{VarError, var};
7 : use std::io::Read;
8 : use std::str::FromStr;
9 : use std::sync::Arc;
10 : use std::time::Duration;
11 :
12 : use anyhow::{Context, anyhow};
13 : use camino::Utf8Path;
14 : use clap::{Arg, ArgAction, Command};
15 : use metrics::launch_timestamp::{LaunchTimestamp, set_launch_timestamp_metric};
16 : use metrics::set_build_info_metric;
17 : use nix::sys::socket::{setsockopt, sockopt};
18 : use pageserver::config::{PageServerConf, PageserverIdentity};
19 : use pageserver::controller_upcall_client::ControllerUpcallClient;
20 : use pageserver::deletion_queue::DeletionQueue;
21 : use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
22 : use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
23 : use pageserver::task_mgr::{
24 : BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME, WALRECEIVER_RUNTIME,
25 : };
26 : use pageserver::tenant::{TenantSharedResources, mgr, secondary};
27 : use pageserver::{
28 : CancellableTask, ConsumptionMetricsTasks, HttpEndpointListener, http, page_cache, page_service,
29 : task_mgr, virtual_file,
30 : };
31 : use postgres_backend::AuthType;
32 : use remote_storage::GenericRemoteStorage;
33 : use tokio::signal::unix::SignalKind;
34 : use tokio::time::Instant;
35 : use tokio_util::sync::CancellationToken;
36 : use tracing::*;
37 : use utils::auth::{JwtAuth, SwappableJwtAuth};
38 : use utils::crashsafe::syncfs;
39 : use utils::logging::TracingErrorLayerEnablement;
40 : use utils::sentry_init::init_sentry;
41 : use utils::{failpoint_support, logging, project_build_tag, project_git_version, tcp_listener};
42 :
43 : project_git_version!(GIT_VERSION);
44 : project_build_tag!(BUILD_TAG);
45 :
46 : #[global_allocator]
47 : static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
48 :
49 : /// Configure jemalloc to profile heap allocations by sampling stack traces every 2 MB (1 << 21).
50 : /// This adds roughly 3% overhead for allocations on average, which is acceptable considering
51 : /// performance-sensitive code will avoid allocations as far as possible anyway.
52 : #[allow(non_upper_case_globals)]
53 : #[unsafe(export_name = "malloc_conf")]
54 : pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
55 :
56 : const PID_FILE_NAME: &str = "pageserver.pid";
57 :
58 : const FEATURES: &[&str] = &[
59 : #[cfg(feature = "testing")]
60 : "testing",
61 : ];
62 :
63 4 : fn version() -> String {
64 4 : format!(
65 4 : "{GIT_VERSION} failpoints: {}, features: {:?}",
66 4 : fail::has_failpoints(),
67 4 : FEATURES,
68 4 : )
69 4 : }
70 :
71 0 : fn main() -> anyhow::Result<()> {
72 0 : let launch_ts = Box::leak(Box::new(LaunchTimestamp::generate()));
73 0 :
74 0 : let arg_matches = cli().get_matches();
75 0 :
76 0 : if arg_matches.get_flag("enabled-features") {
77 0 : println!("{{\"features\": {FEATURES:?} }}");
78 0 : return Ok(());
79 0 : }
80 0 :
81 0 : // Initialize up failpoints support
82 0 : let scenario = failpoint_support::init();
83 0 :
84 0 : let workdir = arg_matches
85 0 : .get_one::<String>("workdir")
86 0 : .map(Utf8Path::new)
87 0 : .unwrap_or_else(|| Utf8Path::new(".neon"));
88 0 : let workdir = workdir
89 0 : .canonicalize_utf8()
90 0 : .with_context(|| format!("Error opening workdir '{workdir}'"))?;
91 :
92 0 : let cfg_file_path = workdir.join("pageserver.toml");
93 0 : let identity_file_path = workdir.join("identity.toml");
94 0 :
95 0 : // Set CWD to workdir for non-daemon modes
96 0 : env::set_current_dir(&workdir)
97 0 : .with_context(|| format!("Failed to set application's current dir to '{workdir}'"))?;
98 :
99 0 : let conf = initialize_config(&identity_file_path, &cfg_file_path, &workdir)?;
100 :
101 : // Initialize logging.
102 : //
103 : // It must be initialized before the custom panic hook is installed below.
104 : //
105 : // Regarding tracing_error enablement: at this time, we only use the
106 : // tracing_error crate to debug_assert that log spans contain tenant and timeline ids.
107 : // See `debug_assert_current_span_has_tenant_and_timeline_id` in the timeline module
108 0 : let tracing_error_layer_enablement = if cfg!(debug_assertions) {
109 0 : TracingErrorLayerEnablement::EnableWithRustLogFilter
110 : } else {
111 0 : TracingErrorLayerEnablement::Disabled
112 : };
113 0 : logging::init(
114 0 : conf.log_format,
115 0 : tracing_error_layer_enablement,
116 0 : logging::Output::Stdout,
117 0 : )?;
118 :
119 : // mind the order required here: 1. logging, 2. panic_hook, 3. sentry.
120 : // disarming this hook on pageserver, because we never tear down tracing.
121 0 : logging::replace_panic_hook_with_tracing_panic_hook().forget();
122 0 :
123 0 : // initialize sentry if SENTRY_DSN is provided
124 0 : let _sentry_guard = init_sentry(
125 0 : Some(GIT_VERSION.into()),
126 0 : &[("node_id", &conf.id.to_string())],
127 0 : );
128 0 :
129 0 : // after setting up logging, log the effective IO engine choice and read path implementations
130 0 : info!(?conf.virtual_file_io_engine, "starting with virtual_file IO engine");
131 0 : info!(?conf.virtual_file_io_mode, "starting with virtual_file IO mode");
132 0 : info!(?conf.wal_receiver_protocol, "starting with WAL receiver protocol");
133 0 : info!(?conf.validate_wal_contiguity, "starting with WAL contiguity validation");
134 0 : info!(?conf.page_service_pipelining, "starting with page service pipelining config");
135 0 : info!(?conf.get_vectored_concurrent_io, "starting with get_vectored IO concurrency config");
136 :
137 : // The tenants directory contains all the pageserver local disk state.
138 : // Create if not exists and make sure all the contents are durable before proceeding.
139 : // Ensuring durability eliminates a whole bug class where we come up after an unclean shutdown.
140 : // After unclea shutdown, we don't know if all the filesystem content we can read via syscalls is actually durable or not.
141 : // Examples for that: OOM kill, systemd killing us during shutdown, self abort due to unrecoverable IO error.
142 0 : let tenants_path = conf.tenants_path();
143 0 : {
144 0 : let open = || {
145 0 : nix::dir::Dir::open(
146 0 : tenants_path.as_std_path(),
147 0 : nix::fcntl::OFlag::O_DIRECTORY | nix::fcntl::OFlag::O_RDONLY,
148 0 : nix::sys::stat::Mode::empty(),
149 0 : )
150 0 : };
151 0 : let dirfd = match open() {
152 0 : Ok(dirfd) => dirfd,
153 0 : Err(e) => match e {
154 : nix::errno::Errno::ENOENT => {
155 0 : utils::crashsafe::create_dir_all(&tenants_path).with_context(|| {
156 0 : format!("Failed to create tenants root dir at '{tenants_path}'")
157 0 : })?;
158 0 : open().context("open tenants dir after creating it")?
159 : }
160 0 : e => anyhow::bail!(e),
161 : },
162 : };
163 :
164 0 : if conf.no_sync {
165 0 : info!("Skipping syncfs on startup");
166 : } else {
167 0 : let started = Instant::now();
168 0 : syncfs(dirfd)?;
169 0 : let elapsed = started.elapsed();
170 0 : info!(
171 0 : elapsed_ms = elapsed.as_millis(),
172 0 : "made tenant directory contents durable"
173 : );
174 : }
175 : }
176 :
177 : // Basic initialization of things that don't change after startup
178 0 : tracing::info!("Initializing virtual_file...");
179 : virtual_file::init(
180 0 : conf.max_file_descriptors,
181 0 : conf.virtual_file_io_engine,
182 0 : conf.virtual_file_io_mode,
183 0 : if conf.no_sync {
184 0 : virtual_file::SyncMode::UnsafeNoSync
185 : } else {
186 0 : virtual_file::SyncMode::Sync
187 : },
188 : );
189 0 : tracing::info!("Initializing page_cache...");
190 0 : page_cache::init(conf.page_cache_size);
191 0 :
192 0 : start_pageserver(launch_ts, conf).context("Failed to start pageserver")?;
193 :
194 0 : scenario.teardown();
195 0 : Ok(())
196 0 : }
197 :
198 0 : fn initialize_config(
199 0 : identity_file_path: &Utf8Path,
200 0 : cfg_file_path: &Utf8Path,
201 0 : workdir: &Utf8Path,
202 0 : ) -> anyhow::Result<&'static PageServerConf> {
203 : // The deployment orchestrator writes out an indentity file containing the node id
204 : // for all pageservers. This file is the source of truth for the node id. In order
205 : // to allow for rolling back pageserver releases, the node id is also included in
206 : // the pageserver config that the deployment orchestrator writes to disk for the pageserver.
207 : // A rolled back version of the pageserver will get the node id from the pageserver.toml
208 : // config file.
209 0 : let identity = match std::fs::File::open(identity_file_path) {
210 0 : Ok(mut f) => {
211 0 : let md = f.metadata().context("stat config file")?;
212 0 : if !md.is_file() {
213 0 : anyhow::bail!(
214 0 : "Pageserver found identity file but it is a dir entry: {identity_file_path}. Aborting start up ..."
215 0 : );
216 0 : }
217 0 :
218 0 : let mut s = String::new();
219 0 : f.read_to_string(&mut s).context("read identity file")?;
220 0 : toml_edit::de::from_str::<PageserverIdentity>(&s)?
221 : }
222 0 : Err(e) => {
223 0 : anyhow::bail!(
224 0 : "Pageserver could not read identity file: {identity_file_path}: {e}. Aborting start up ..."
225 0 : );
226 : }
227 : };
228 :
229 0 : let config_file_contents =
230 0 : std::fs::read_to_string(cfg_file_path).context("read config file from filesystem")?;
231 0 : let config_toml = serde_path_to_error::deserialize(
232 0 : toml_edit::de::Deserializer::from_str(&config_file_contents)
233 0 : .context("build toml deserializer")?,
234 : )
235 0 : .context("deserialize config toml")?;
236 0 : let conf = PageServerConf::parse_and_validate(identity.id, config_toml, workdir)
237 0 : .context("runtime-validation of config toml")?;
238 :
239 0 : Ok(Box::leak(Box::new(conf)))
240 0 : }
241 :
242 : struct WaitForPhaseResult<F: std::future::Future + Unpin> {
243 : timeout_remaining: Duration,
244 : skipped: Option<F>,
245 : }
246 :
247 : /// During startup, we apply a timeout to our waits for readiness, to avoid
248 : /// stalling the whole service if one Tenant experiences some problem. Each
249 : /// phase may consume some of the timeout: this function returns the updated
250 : /// timeout for use in the next call.
251 0 : async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
252 0 : where
253 0 : F: std::future::Future + Unpin,
254 0 : {
255 0 : let initial_t = Instant::now();
256 0 : let skipped = match tokio::time::timeout(timeout, &mut fut).await {
257 0 : Ok(_) => None,
258 : Err(_) => {
259 0 : tracing::info!(
260 0 : timeout_millis = timeout.as_millis(),
261 0 : %phase,
262 0 : "Startup phase timed out, proceeding anyway"
263 : );
264 0 : Some(fut)
265 : }
266 : };
267 :
268 0 : WaitForPhaseResult {
269 0 : timeout_remaining: timeout
270 0 : .checked_sub(Instant::now().duration_since(initial_t))
271 0 : .unwrap_or(Duration::ZERO),
272 0 : skipped,
273 0 : }
274 0 : }
275 :
276 0 : fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
277 0 : let elapsed = started_at.elapsed();
278 0 : let secs = elapsed.as_secs_f64();
279 0 : STARTUP_DURATION.with_label_values(&[phase]).set(secs);
280 0 :
281 0 : info!(
282 0 : elapsed_ms = elapsed.as_millis(),
283 0 : "{human_phase} ({secs:.3}s since start)"
284 : )
285 0 : }
286 :
287 0 : fn start_pageserver(
288 0 : launch_ts: &'static LaunchTimestamp,
289 0 : conf: &'static PageServerConf,
290 0 : ) -> anyhow::Result<()> {
291 0 : // Monotonic time for later calculating startup duration
292 0 : let started_startup_at = Instant::now();
293 0 :
294 0 : // Print version and launch timestamp to the log,
295 0 : // and expose them as prometheus metrics.
296 0 : // A changed version string indicates changed software.
297 0 : // A changed launch timestamp indicates a pageserver restart.
298 0 : info!(
299 0 : "version: {} launch_timestamp: {} build_tag: {}",
300 0 : version(),
301 0 : launch_ts.to_string(),
302 : BUILD_TAG,
303 : );
304 0 : set_build_info_metric(GIT_VERSION, BUILD_TAG);
305 0 : set_launch_timestamp_metric(launch_ts);
306 0 : #[cfg(target_os = "linux")]
307 0 : metrics::register_internal(Box::new(metrics::more_process_metrics::Collector::new())).unwrap();
308 0 : metrics::register_internal(Box::new(
309 0 : pageserver::metrics::tokio_epoll_uring::Collector::new(),
310 0 : ))
311 0 : .unwrap();
312 0 : pageserver::preinitialize_metrics(conf);
313 0 :
314 0 : // If any failpoints were set from FAILPOINTS environment variable,
315 0 : // print them to the log for debugging purposes
316 0 : let failpoints = fail::list();
317 0 : if !failpoints.is_empty() {
318 0 : info!(
319 0 : "started with failpoints: {}",
320 0 : failpoints
321 0 : .iter()
322 0 : .map(|(name, actions)| format!("{name}={actions}"))
323 0 : .collect::<Vec<String>>()
324 0 : .join(";")
325 : )
326 0 : }
327 :
328 : // Create and lock PID file. This ensures that there cannot be more than one
329 : // pageserver process running at the same time.
330 0 : let lock_file_path = conf.workdir.join(PID_FILE_NAME);
331 0 : info!("Claiming pid file at {lock_file_path:?}...");
332 0 : let lock_file =
333 0 : utils::pid_file::claim_for_current_process(&lock_file_path).context("claim pid file")?;
334 0 : info!("Claimed pid file at {lock_file_path:?}");
335 :
336 : // Ensure that the lock file is held even if the main thread of the process panics.
337 : // We need to release the lock file only when the process exits.
338 0 : std::mem::forget(lock_file);
339 0 :
340 0 : // Bind the HTTP and libpq ports early, so that if they are in use by some other
341 0 : // process, we error out early.
342 0 : let http_addr = &conf.listen_http_addr;
343 0 : info!("Starting pageserver http handler on {http_addr}");
344 0 : let http_listener = tcp_listener::bind(http_addr)?;
345 :
346 0 : let pg_addr = &conf.listen_pg_addr;
347 0 :
348 0 : info!("Starting pageserver pg protocol handler on {pg_addr}");
349 0 : let pageserver_listener = tcp_listener::bind(pg_addr)?;
350 :
351 : // Enable SO_KEEPALIVE on the socket, to detect dead connections faster.
352 : // These are configured via net.ipv4.tcp_keepalive_* sysctls.
353 : //
354 : // TODO: also set this on the walreceiver socket, but tokio-postgres doesn't
355 : // support enabling keepalives while using the default OS sysctls.
356 0 : setsockopt(&pageserver_listener, sockopt::KeepAlive, &true)?;
357 :
358 : // Launch broker client
359 : // The storage_broker::connect call needs to happen inside a tokio runtime thread.
360 0 : let broker_client = WALRECEIVER_RUNTIME
361 0 : .block_on(async {
362 0 : // Note: we do not attempt connecting here (but validate endpoints sanity).
363 0 : storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)
364 0 : })
365 0 : .with_context(|| {
366 0 : format!(
367 0 : "create broker client for uri={:?} keepalive_interval={:?}",
368 0 : &conf.broker_endpoint, conf.broker_keepalive_interval,
369 0 : )
370 0 : })?;
371 :
372 : // Initialize authentication for incoming connections
373 : let http_auth;
374 : let pg_auth;
375 0 : if conf.http_auth_type == AuthType::NeonJWT || conf.pg_auth_type == AuthType::NeonJWT {
376 : // unwrap is ok because check is performed when creating config, so path is set and exists
377 0 : let key_path = conf.auth_validation_public_key_path.as_ref().unwrap();
378 0 : info!("Loading public key(s) for verifying JWT tokens from {key_path:?}");
379 :
380 0 : let jwt_auth = JwtAuth::from_key_path(key_path)?;
381 0 : let auth: Arc<SwappableJwtAuth> = Arc::new(SwappableJwtAuth::new(jwt_auth));
382 :
383 0 : http_auth = match &conf.http_auth_type {
384 0 : AuthType::Trust => None,
385 0 : AuthType::NeonJWT => Some(auth.clone()),
386 : };
387 0 : pg_auth = match &conf.pg_auth_type {
388 0 : AuthType::Trust => None,
389 0 : AuthType::NeonJWT => Some(auth),
390 : };
391 0 : } else {
392 0 : http_auth = None;
393 0 : pg_auth = None;
394 0 : }
395 0 : info!("Using auth for http API: {:#?}", conf.http_auth_type);
396 0 : info!("Using auth for pg connections: {:#?}", conf.pg_auth_type);
397 :
398 0 : match var("NEON_AUTH_TOKEN") {
399 0 : Ok(v) => {
400 0 : info!("Loaded JWT token for authentication with Safekeeper");
401 0 : pageserver::config::SAFEKEEPER_AUTH_TOKEN
402 0 : .set(Arc::new(v))
403 0 : .map_err(|_| anyhow!("Could not initialize SAFEKEEPER_AUTH_TOKEN"))?;
404 : }
405 : Err(VarError::NotPresent) => {
406 0 : info!("No JWT token for authentication with Safekeeper detected");
407 : }
408 0 : Err(e) => return Err(e).with_context(
409 0 : || "Failed to either load to detect non-present NEON_AUTH_TOKEN environment variable",
410 0 : ),
411 : };
412 :
413 : // Top-level cancellation token for the process
414 0 : let shutdown_pageserver = tokio_util::sync::CancellationToken::new();
415 :
416 : // Set up remote storage client
417 0 : let remote_storage = BACKGROUND_RUNTIME.block_on(create_remote_storage_client(conf))?;
418 :
419 : // Set up deletion queue
420 0 : let (deletion_queue, deletion_workers) = DeletionQueue::new(
421 0 : remote_storage.clone(),
422 0 : ControllerUpcallClient::new(conf, &shutdown_pageserver),
423 0 : conf,
424 0 : );
425 0 : deletion_workers.spawn_with(BACKGROUND_RUNTIME.handle());
426 0 :
427 0 : // Up to this point no significant I/O has been done: this should have been fast. Record
428 0 : // duration prior to starting I/O intensive phase of startup.
429 0 : startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
430 0 : STARTUP_IS_LOADING.set(1);
431 0 :
432 0 : // Startup staging or optimizing:
433 0 : //
434 0 : // We want to minimize downtime for `page_service` connections, and trying not to overload
435 0 : // BACKGROUND_RUNTIME by doing initial compactions and initial logical sizes at the same time.
436 0 : //
437 0 : // init_done_rx will notify when all initial load operations have completed.
438 0 : //
439 0 : // background_jobs_can_start (same name used to hold off background jobs from starting at
440 0 : // consumer side) will be dropped once we can start the background jobs. Currently it is behind
441 0 : // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
442 0 : // (background_task_maximum_delay).
443 0 : let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
444 0 : let (init_done_tx, init_done_rx) = utils::completion::channel();
445 0 :
446 0 : let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
447 0 :
448 0 : let order = pageserver::InitializationOrder {
449 0 : initial_tenant_load_remote: Some(init_done_tx),
450 0 : initial_tenant_load: Some(init_remote_done_tx),
451 0 : background_jobs_can_start: background_jobs_barrier.clone(),
452 0 : };
453 0 :
454 0 : info!(config=?conf.l0_flush, "using l0_flush config");
455 0 : let l0_flush_global_state =
456 0 : pageserver::l0_flush::L0FlushGlobalState::new(conf.l0_flush.clone());
457 0 :
458 0 : // Scan the local 'tenants/' directory and start loading the tenants
459 0 : let deletion_queue_client = deletion_queue.new_client();
460 0 : let background_purges = mgr::BackgroundPurges::default();
461 0 : let tenant_manager = BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
462 0 : conf,
463 0 : background_purges.clone(),
464 0 : TenantSharedResources {
465 0 : broker_client: broker_client.clone(),
466 0 : remote_storage: remote_storage.clone(),
467 0 : deletion_queue_client,
468 0 : l0_flush_global_state,
469 0 : },
470 0 : order,
471 0 : shutdown_pageserver.clone(),
472 0 : ))?;
473 0 : let tenant_manager = Arc::new(tenant_manager);
474 0 :
475 0 : BACKGROUND_RUNTIME.spawn({
476 0 : let shutdown_pageserver = shutdown_pageserver.clone();
477 0 : let drive_init = async move {
478 0 : // NOTE: unlike many futures in pageserver, this one is cancellation-safe
479 0 : let guard = scopeguard::guard_on_success((), |_| {
480 0 : tracing::info!("Cancelled before initial load completed")
481 0 : });
482 0 :
483 0 : let timeout = conf.background_task_maximum_delay;
484 0 :
485 0 : let init_remote_done = std::pin::pin!(async {
486 0 : init_remote_done_rx.wait().await;
487 0 : startup_checkpoint(
488 0 : started_startup_at,
489 0 : "initial_tenant_load_remote",
490 0 : "Remote part of initial load completed",
491 0 : );
492 0 : });
493 :
494 : let WaitForPhaseResult {
495 0 : timeout_remaining: timeout,
496 0 : skipped: init_remote_skipped,
497 0 : } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
498 :
499 0 : let init_load_done = std::pin::pin!(async {
500 0 : init_done_rx.wait().await;
501 0 : startup_checkpoint(
502 0 : started_startup_at,
503 0 : "initial_tenant_load",
504 0 : "Initial load completed",
505 0 : );
506 0 : STARTUP_IS_LOADING.set(0);
507 0 : });
508 :
509 : let WaitForPhaseResult {
510 0 : timeout_remaining: _timeout,
511 0 : skipped: init_load_skipped,
512 0 : } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
513 :
514 : // initial logical sizes can now start, as they were waiting on init_done_rx.
515 :
516 0 : scopeguard::ScopeGuard::into_inner(guard);
517 0 :
518 0 : // allow background jobs to start: we either completed prior stages, or they reached timeout
519 0 : // and were skipped. It is important that we do not let them block background jobs indefinitely,
520 0 : // because things like consumption metrics for billing are blocked by this barrier.
521 0 : drop(background_jobs_can_start);
522 0 : startup_checkpoint(
523 0 : started_startup_at,
524 0 : "background_jobs_can_start",
525 0 : "Starting background jobs",
526 0 : );
527 0 :
528 0 : // We are done. If we skipped any phases due to timeout, run them to completion here so that
529 0 : // they will eventually update their startup_checkpoint, and so that we do not declare the
530 0 : // 'complete' stage until all the other stages are really done.
531 0 : let guard = scopeguard::guard_on_success((), |_| {
532 0 : tracing::info!("Cancelled before waiting for skipped phases done")
533 0 : });
534 0 : if let Some(f) = init_remote_skipped {
535 0 : f.await;
536 0 : }
537 0 : if let Some(f) = init_load_skipped {
538 0 : f.await;
539 0 : }
540 0 : scopeguard::ScopeGuard::into_inner(guard);
541 0 :
542 0 : startup_checkpoint(started_startup_at, "complete", "Startup complete");
543 0 : };
544 0 :
545 0 : async move {
546 0 : let mut drive_init = std::pin::pin!(drive_init);
547 0 : // just race these tasks
548 0 : tokio::select! {
549 0 : _ = shutdown_pageserver.cancelled() => {},
550 0 : _ = &mut drive_init => {},
551 : }
552 0 : }
553 0 : });
554 0 :
555 0 : let (secondary_controller, secondary_controller_tasks) = secondary::spawn_tasks(
556 0 : tenant_manager.clone(),
557 0 : remote_storage.clone(),
558 0 : background_jobs_barrier.clone(),
559 0 : shutdown_pageserver.clone(),
560 0 : );
561 0 :
562 0 : // shared state between the disk-usage backed eviction background task and the http endpoint
563 0 : // that allows triggering disk-usage based eviction manually. note that the http endpoint
564 0 : // is still accessible even if background task is not configured as long as remote storage has
565 0 : // been configured.
566 0 : let disk_usage_eviction_state: Arc<disk_usage_eviction_task::State> = Arc::default();
567 0 :
568 0 : let disk_usage_eviction_task = launch_disk_usage_global_eviction_task(
569 0 : conf,
570 0 : remote_storage.clone(),
571 0 : disk_usage_eviction_state.clone(),
572 0 : tenant_manager.clone(),
573 0 : background_jobs_barrier.clone(),
574 0 : );
575 :
576 : // Start up the service to handle HTTP mgmt API request. We created the
577 : // listener earlier already.
578 0 : let http_endpoint_listener = {
579 0 : let _rt_guard = MGMT_REQUEST_RUNTIME.enter(); // for hyper
580 0 : let cancel = CancellationToken::new();
581 :
582 0 : let router_state = Arc::new(
583 0 : http::routes::State::new(
584 0 : conf,
585 0 : tenant_manager.clone(),
586 0 : http_auth.clone(),
587 0 : remote_storage.clone(),
588 0 : broker_client.clone(),
589 0 : disk_usage_eviction_state,
590 0 : deletion_queue.new_client(),
591 0 : secondary_controller,
592 0 : )
593 0 : .context("Failed to initialize router state")?,
594 : );
595 0 : let router = http::make_router(router_state, launch_ts, http_auth.clone())?
596 0 : .build()
597 0 : .map_err(|err| anyhow!(err))?;
598 0 : let service = http_utils::RouterService::new(router).unwrap();
599 0 : let server = hyper0::Server::from_tcp(http_listener)?
600 0 : .serve(service)
601 0 : .with_graceful_shutdown({
602 0 : let cancel = cancel.clone();
603 0 : async move { cancel.clone().cancelled().await }
604 0 : });
605 0 :
606 0 : let task = MGMT_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
607 0 : "http endpoint listener",
608 0 : server,
609 0 : ));
610 0 : HttpEndpointListener(CancellableTask { task, cancel })
611 0 : };
612 0 :
613 0 : let consumption_metrics_tasks = {
614 0 : let cancel = shutdown_pageserver.child_token();
615 0 : let task = crate::BACKGROUND_RUNTIME.spawn({
616 0 : let tenant_manager = tenant_manager.clone();
617 0 : let cancel = cancel.clone();
618 0 : async move {
619 0 : // first wait until background jobs are cleared to launch.
620 0 : //
621 0 : // this is because we only process active tenants and timelines, and the
622 0 : // Timeline::get_current_logical_size will spawn the logical size calculation,
623 0 : // which will not be rate-limited.
624 0 : tokio::select! {
625 0 : _ = cancel.cancelled() => { return; },
626 0 : _ = background_jobs_barrier.wait() => {}
627 0 : };
628 0 :
629 0 : pageserver::consumption_metrics::run(conf, tenant_manager, cancel).await;
630 0 : }
631 0 : });
632 0 : ConsumptionMetricsTasks(CancellableTask { task, cancel })
633 : };
634 :
635 : // Spawn a task to listen for libpq connections. It will spawn further tasks
636 : // for each connection. We created the listener earlier already.
637 0 : let page_service = page_service::spawn(conf, tenant_manager.clone(), pg_auth, {
638 0 : let _entered = COMPUTE_REQUEST_RUNTIME.enter(); // TcpListener::from_std requires it
639 0 : pageserver_listener
640 0 : .set_nonblocking(true)
641 0 : .context("set listener to nonblocking")?;
642 0 : tokio::net::TcpListener::from_std(pageserver_listener).context("create tokio listener")?
643 : });
644 :
645 : // All started up! Now just sit and wait for shutdown signal.
646 0 : BACKGROUND_RUNTIME.block_on(async move {
647 0 : let signal_token = CancellationToken::new();
648 0 : let signal_cancel = signal_token.child_token();
649 0 :
650 0 : // Spawn signal handlers. Runs in a loop since we want to be responsive to multiple signals
651 0 : // even after triggering shutdown (e.g. a SIGQUIT after a slow SIGTERM shutdown). See:
652 0 : // https://github.com/neondatabase/neon/issues/9740.
653 0 : tokio::spawn(async move {
654 0 : let mut sigint = tokio::signal::unix::signal(SignalKind::interrupt()).unwrap();
655 0 : let mut sigterm = tokio::signal::unix::signal(SignalKind::terminate()).unwrap();
656 0 : let mut sigquit = tokio::signal::unix::signal(SignalKind::quit()).unwrap();
657 :
658 : loop {
659 0 : let signal = tokio::select! {
660 0 : _ = sigquit.recv() => {
661 0 : info!("Got signal SIGQUIT. Terminating in immediate shutdown mode.");
662 0 : std::process::exit(111);
663 : }
664 0 : _ = sigint.recv() => "SIGINT",
665 0 : _ = sigterm.recv() => "SIGTERM",
666 : };
667 :
668 0 : if !signal_token.is_cancelled() {
669 0 : info!("Got signal {signal}. Terminating gracefully in fast shutdown mode.");
670 0 : signal_token.cancel();
671 : } else {
672 0 : info!("Got signal {signal}. Already shutting down.");
673 : }
674 : }
675 0 : });
676 0 :
677 0 : // Wait for cancellation signal and shut down the pageserver.
678 0 : //
679 0 : // This cancels the `shutdown_pageserver` cancellation tree. Right now that tree doesn't
680 0 : // reach very far, and `task_mgr` is used instead. The plan is to change that over time.
681 0 : signal_cancel.cancelled().await;
682 :
683 0 : shutdown_pageserver.cancel();
684 0 : pageserver::shutdown_pageserver(
685 0 : http_endpoint_listener,
686 0 : page_service,
687 0 : consumption_metrics_tasks,
688 0 : disk_usage_eviction_task,
689 0 : &tenant_manager,
690 0 : background_purges,
691 0 : deletion_queue.clone(),
692 0 : secondary_controller_tasks,
693 0 : 0,
694 0 : )
695 0 : .await;
696 0 : unreachable!();
697 0 : })
698 0 : }
699 :
700 0 : async fn create_remote_storage_client(
701 0 : conf: &'static PageServerConf,
702 0 : ) -> anyhow::Result<GenericRemoteStorage> {
703 0 : let config = if let Some(config) = &conf.remote_storage_config {
704 0 : config
705 : } else {
706 0 : anyhow::bail!("no remote storage configured, this is a deprecated configuration");
707 : };
708 :
709 : // Create the client
710 0 : let mut remote_storage = GenericRemoteStorage::from_config(config).await?;
711 :
712 : // If `test_remote_failures` is non-zero, wrap the client with a
713 : // wrapper that simulates failures.
714 0 : if conf.test_remote_failures > 0 {
715 0 : if !cfg!(feature = "testing") {
716 0 : anyhow::bail!(
717 0 : "test_remote_failures option is not available because pageserver was compiled without the 'testing' feature"
718 0 : );
719 0 : }
720 0 : info!(
721 0 : "Simulating remote failures for first {} attempts of each op",
722 : conf.test_remote_failures
723 : );
724 0 : remote_storage =
725 0 : GenericRemoteStorage::unreliable_wrapper(remote_storage, conf.test_remote_failures);
726 0 : }
727 :
728 0 : Ok(remote_storage)
729 0 : }
730 :
731 4 : fn cli() -> Command {
732 4 : Command::new("Neon page server")
733 4 : .about("Materializes WAL stream to pages and serves them to the postgres")
734 4 : .version(version())
735 4 : .arg(
736 4 : Arg::new("workdir")
737 4 : .short('D')
738 4 : .long("workdir")
739 4 : .help("Working directory for the pageserver"),
740 4 : )
741 4 : .arg(
742 4 : Arg::new("enabled-features")
743 4 : .long("enabled-features")
744 4 : .action(ArgAction::SetTrue)
745 4 : .help("Show enabled compile time features"),
746 4 : )
747 4 : }
748 :
749 : #[test]
750 4 : fn verify_cli() {
751 4 : cli().debug_assert();
752 4 : }
|
