Line data Source code
1 : //!
2 : //! Postgres wrapper (`compute_ctl`) is intended to be run as a Docker entrypoint or as a `systemd`
3 : //! `ExecStart` option. It will handle all the `Neon` specifics during compute node
4 : //! initialization:
5 : //! - `compute_ctl` accepts cluster (compute node) specification as a JSON file.
6 : //! - Every start is a fresh start, so the data directory is removed and
7 : //! initialized again on each run.
8 : //! - If remote_extension_config is provided, it will be used to fetch extensions list
9 : //! and download `shared_preload_libraries` from the remote storage.
10 : //! - Next it will put configuration files into the `PGDATA` directory.
11 : //! - Sync safekeepers and get commit LSN.
12 : //! - Get `basebackup` from pageserver using the returned on the previous step LSN.
13 : //! - Try to start `postgres` and wait until it is ready to accept connections.
14 : //! - Check and alter/drop/create roles and databases.
15 : //! - Hang waiting on the `postmaster` process to exit.
16 : //!
17 : //! Also `compute_ctl` spawns two separate service threads:
18 : //! - `compute-monitor` checks the last Postgres activity timestamp and saves it
19 : //! into the shared `ComputeNode`;
20 : //! - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
21 : //! last activity requests.
22 : //!
23 : //! If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
24 : //! `vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
25 : //! `vm-monitor` communicates with the VM autoscaling system. It coordinates
26 : //! downscaling and requests immediate upscaling under resource pressure.
27 : //!
28 : //! Usage example:
29 : //! ```sh
30 : //! compute_ctl -D /var/db/postgres/compute \
31 : //! -C 'postgresql://cloud_admin@localhost/postgres' \
32 : //! -S /var/db/postgres/specs/current.json \
33 : //! -b /usr/local/bin/postgres \
34 : //! -r http://pg-ext-s3-gateway \
35 : //! ```
36 : use std::collections::HashMap;
37 : use std::ffi::OsString;
38 : use std::fs::File;
39 : use std::path::Path;
40 : use std::process::exit;
41 : use std::str::FromStr;
42 : use std::sync::atomic::Ordering;
43 : use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
44 : use std::{thread, time::Duration};
45 :
46 : use anyhow::{Context, Result};
47 : use chrono::Utc;
48 : use clap::Parser;
49 : use compute_tools::disk_quota::set_disk_quota;
50 : use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
51 : use signal_hook::consts::{SIGQUIT, SIGTERM};
52 : use signal_hook::{consts::SIGINT, iterator::Signals};
53 : use tracing::{error, info, warn};
54 : use url::Url;
55 :
56 : use compute_api::responses::ComputeStatus;
57 : use compute_api::spec::ComputeSpec;
58 :
59 : use compute_tools::compute::{
60 : forward_termination_signal, ComputeNode, ComputeState, ParsedSpec, PG_PID,
61 : };
62 : use compute_tools::configurator::launch_configurator;
63 : use compute_tools::extension_server::get_pg_version_string;
64 : use compute_tools::http::launch_http_server;
65 : use compute_tools::logger::*;
66 : use compute_tools::monitor::launch_monitor;
67 : use compute_tools::params::*;
68 : use compute_tools::spec::*;
69 : use compute_tools::swap::resize_swap;
70 : use rlimit::{setrlimit, Resource};
71 : use utils::failpoint_support;
72 :
73 : // this is an arbitrary build tag. Fine as a default / for testing purposes
74 : // in-case of not-set environment var
75 : const BUILD_TAG_DEFAULT: &str = "latest";
76 :
77 : // Compatibility hack: if the control plane specified any remote-ext-config
78 : // use the default value for extension storage proxy gateway.
79 : // Remove this once the control plane is updated to pass the gateway URL
80 0 : fn parse_remote_ext_config(arg: &str) -> Result<String> {
81 0 : if arg.starts_with("http") {
82 0 : Ok(arg.trim_end_matches('/').to_string())
83 : } else {
84 0 : Ok("http://pg-ext-s3-gateway".to_string())
85 : }
86 0 : }
87 :
88 : #[derive(Parser)]
89 : #[command(rename_all = "kebab-case")]
90 : struct Cli {
91 : #[arg(short = 'b', long, default_value = "postgres", env = "POSTGRES_PATH")]
92 0 : pub pgbin: String,
93 :
94 : #[arg(short = 'r', long, value_parser = parse_remote_ext_config)]
95 : pub remote_ext_config: Option<String>,
96 :
97 1 : #[arg(long, default_value_t = 3080)]
98 0 : pub http_port: u16,
99 :
100 : #[arg(short = 'D', long, value_name = "DATADIR")]
101 0 : pub pgdata: String,
102 :
103 : #[arg(short = 'C', long, value_name = "DATABASE_URL")]
104 0 : pub connstr: String,
105 :
106 : #[cfg(target_os = "linux")]
107 : #[arg(long, default_value = "neon-postgres")]
108 0 : pub cgroup: String,
109 :
110 : #[cfg(target_os = "linux")]
111 : #[arg(
112 : long,
113 : default_value = "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable application_name=vm-monitor"
114 : )]
115 0 : pub filecache_connstr: String,
116 :
117 : #[cfg(target_os = "linux")]
118 : #[arg(long, default_value = "0.0.0.0:10301")]
119 0 : pub vm_monitor_addr: String,
120 :
121 : #[arg(long, action = clap::ArgAction::SetTrue)]
122 0 : pub resize_swap_on_bind: bool,
123 :
124 : #[arg(long)]
125 : pub set_disk_quota_for_fs: Option<String>,
126 :
127 : #[arg(short = 's', long = "spec", group = "spec")]
128 : pub spec_json: Option<String>,
129 :
130 : #[arg(short = 'S', long, group = "spec-path")]
131 : pub spec_path: Option<OsString>,
132 :
133 : #[arg(short = 'i', long, group = "compute-id", conflicts_with_all = ["spec", "spec-path"])]
134 : pub compute_id: Option<String>,
135 :
136 : #[arg(short = 'p', long, conflicts_with_all = ["spec", "spec-path"], requires = "compute-id", value_name = "CONTROL_PLANE_API_BASE_URL")]
137 : pub control_plane_uri: Option<String>,
138 : }
139 :
140 0 : fn main() -> Result<()> {
141 0 : let cli = Cli::parse();
142 :
143 0 : let build_tag = init()?;
144 :
145 0 : let scenario = failpoint_support::init();
146 0 :
147 0 : // enable core dumping for all child processes
148 0 : setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
149 :
150 0 : let (pg_handle, start_pg_result) = {
151 : // Enter startup tracing context
152 0 : let _startup_context_guard = startup_context_from_env();
153 :
154 0 : let cli_spec = try_spec_from_cli(&cli)?;
155 :
156 0 : let compute = wait_spec(build_tag, &cli, cli_spec)?;
157 :
158 0 : start_postgres(&cli, compute)?
159 :
160 : // Startup is finished, exit the startup tracing span
161 : };
162 :
163 : // PostgreSQL is now running, if startup was successful. Wait until it exits.
164 0 : let wait_pg_result = wait_postgres(pg_handle)?;
165 :
166 0 : let delay_exit = cleanup_after_postgres_exit(start_pg_result)?;
167 :
168 0 : maybe_delay_exit(delay_exit);
169 0 :
170 0 : scenario.teardown();
171 0 :
172 0 : deinit_and_exit(wait_pg_result);
173 0 : }
174 :
175 0 : fn init() -> Result<String> {
176 0 : init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
177 :
178 0 : let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
179 0 : thread::spawn(move || {
180 0 : for sig in signals.forever() {
181 0 : handle_exit_signal(sig);
182 0 : }
183 0 : });
184 0 :
185 0 : let build_tag = option_env!("BUILD_TAG")
186 0 : .unwrap_or(BUILD_TAG_DEFAULT)
187 0 : .to_string();
188 0 : info!("build_tag: {build_tag}");
189 :
190 0 : Ok(build_tag)
191 0 : }
192 :
193 0 : fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
194 0 : // Extract OpenTelemetry context for the startup actions from the
195 0 : // TRACEPARENT and TRACESTATE env variables, and attach it to the current
196 0 : // tracing context.
197 0 : //
198 0 : // This is used to propagate the context for the 'start_compute' operation
199 0 : // from the neon control plane. This allows linking together the wider
200 0 : // 'start_compute' operation that creates the compute container, with the
201 0 : // startup actions here within the container.
202 0 : //
203 0 : // There is no standard for passing context in env variables, but a lot of
204 0 : // tools use TRACEPARENT/TRACESTATE, so we use that convention too. See
205 0 : // https://github.com/open-telemetry/opentelemetry-specification/issues/740
206 0 : //
207 0 : // Switch to the startup context here, and exit it once the startup has
208 0 : // completed and Postgres is up and running.
209 0 : //
210 0 : // If this pod is pre-created without binding it to any particular endpoint
211 0 : // yet, this isn't the right place to enter the startup context. In that
212 0 : // case, the control plane should pass the tracing context as part of the
213 0 : // /configure API call.
214 0 : //
215 0 : // NOTE: This is supposed to only cover the *startup* actions. Once
216 0 : // postgres is configured and up-and-running, we exit this span. Any other
217 0 : // actions that are performed on incoming HTTP requests, for example, are
218 0 : // performed in separate spans.
219 0 : //
220 0 : // XXX: If the pod is restarted, we perform the startup actions in the same
221 0 : // context as the original startup actions, which probably doesn't make
222 0 : // sense.
223 0 : let mut startup_tracing_carrier: HashMap<String, String> = HashMap::new();
224 0 : if let Ok(val) = std::env::var("TRACEPARENT") {
225 0 : startup_tracing_carrier.insert("traceparent".to_string(), val);
226 0 : }
227 0 : if let Ok(val) = std::env::var("TRACESTATE") {
228 0 : startup_tracing_carrier.insert("tracestate".to_string(), val);
229 0 : }
230 0 : if !startup_tracing_carrier.is_empty() {
231 : use opentelemetry::propagation::TextMapPropagator;
232 : use opentelemetry_sdk::propagation::TraceContextPropagator;
233 0 : let guard = TraceContextPropagator::new()
234 0 : .extract(&startup_tracing_carrier)
235 0 : .attach();
236 0 : info!("startup tracing context attached");
237 0 : Some(guard)
238 : } else {
239 0 : None
240 : }
241 0 : }
242 :
243 0 : fn try_spec_from_cli(cli: &Cli) -> Result<CliSpecParams> {
244 : // First, try to get cluster spec from the cli argument
245 0 : if let Some(ref spec_json) = cli.spec_json {
246 0 : info!("got spec from cli argument {}", spec_json);
247 : return Ok(CliSpecParams {
248 0 : spec: Some(serde_json::from_str(spec_json)?),
249 : live_config_allowed: false,
250 : });
251 0 : }
252 :
253 : // Second, try to read it from the file if path is provided
254 0 : if let Some(ref spec_path) = cli.spec_path {
255 0 : let file = File::open(Path::new(spec_path))?;
256 : return Ok(CliSpecParams {
257 0 : spec: Some(serde_json::from_reader(file)?),
258 : live_config_allowed: true,
259 : });
260 0 : }
261 0 :
262 0 : if cli.compute_id.is_none() {
263 0 : panic!(
264 0 : "compute spec should be provided by one of the following ways: \
265 0 : --spec OR --spec-path OR --control-plane-uri and --compute-id"
266 0 : );
267 0 : };
268 0 : if cli.control_plane_uri.is_none() {
269 0 : panic!("must specify both --control-plane-uri and --compute-id or none");
270 0 : };
271 0 :
272 0 : match get_spec_from_control_plane(
273 0 : cli.control_plane_uri.as_ref().unwrap(),
274 0 : cli.compute_id.as_ref().unwrap(),
275 0 : ) {
276 0 : Ok(spec) => Ok(CliSpecParams {
277 0 : spec,
278 0 : live_config_allowed: true,
279 0 : }),
280 0 : Err(e) => {
281 0 : error!(
282 0 : "cannot get response from control plane: {}\n\
283 0 : neither spec nor confirmation that compute is in the Empty state was received",
284 : e
285 : );
286 0 : Err(e)
287 : }
288 : }
289 0 : }
290 :
291 : struct CliSpecParams {
292 : /// If a spec was provided via CLI or file, the [`ComputeSpec`]
293 : spec: Option<ComputeSpec>,
294 : live_config_allowed: bool,
295 : }
296 :
297 0 : fn wait_spec(
298 0 : build_tag: String,
299 0 : cli: &Cli,
300 0 : CliSpecParams {
301 0 : spec,
302 0 : live_config_allowed,
303 0 : }: CliSpecParams,
304 0 : ) -> Result<Arc<ComputeNode>> {
305 0 : let mut new_state = ComputeState::new();
306 : let spec_set;
307 :
308 0 : if let Some(spec) = spec {
309 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
310 0 : info!("new pspec.spec: {:?}", pspec.spec);
311 0 : new_state.pspec = Some(pspec);
312 0 : spec_set = true;
313 0 : } else {
314 0 : spec_set = false;
315 0 : }
316 0 : let connstr = Url::parse(&cli.connstr).context("cannot parse connstr as a URL")?;
317 0 : let conn_conf = postgres::config::Config::from_str(connstr.as_str())
318 0 : .context("cannot build postgres config from connstr")?;
319 0 : let tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr.as_str())
320 0 : .context("cannot build tokio postgres config from connstr")?;
321 0 : let compute_node = ComputeNode {
322 0 : connstr,
323 0 : conn_conf,
324 0 : tokio_conn_conf,
325 0 : pgdata: cli.pgdata.clone(),
326 0 : pgbin: cli.pgbin.clone(),
327 0 : pgversion: get_pg_version_string(&cli.pgbin),
328 0 : http_port: cli.http_port,
329 0 : live_config_allowed,
330 0 : state: Mutex::new(new_state),
331 0 : state_changed: Condvar::new(),
332 0 : ext_remote_storage: cli.remote_ext_config.clone(),
333 0 : ext_download_progress: RwLock::new(HashMap::new()),
334 0 : build_tag,
335 0 : };
336 0 : let compute = Arc::new(compute_node);
337 0 :
338 0 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
339 0 : // available for binding. Prewarming helps Postgres start quicker later,
340 0 : // because QEMU will already have its memory allocated from the host, and
341 0 : // the necessary binaries will already be cached.
342 0 : if !spec_set {
343 0 : compute.prewarm_postgres()?;
344 0 : }
345 :
346 : // Launch http service first, so that we can serve control-plane requests
347 : // while configuration is still in progress.
348 0 : let _http_handle =
349 0 : launch_http_server(cli.http_port, &compute).expect("cannot launch http endpoint thread");
350 0 :
351 0 : if !spec_set {
352 : // No spec provided, hang waiting for it.
353 0 : info!("no compute spec provided, waiting");
354 :
355 0 : let mut state = compute.state.lock().unwrap();
356 0 : while state.status != ComputeStatus::ConfigurationPending {
357 0 : state = compute.state_changed.wait(state).unwrap();
358 0 :
359 0 : if state.status == ComputeStatus::ConfigurationPending {
360 0 : info!("got spec, continue configuration");
361 : // Spec is already set by the http server handler.
362 0 : break;
363 0 : }
364 : }
365 :
366 : // Record for how long we slept waiting for the spec.
367 0 : let now = Utc::now();
368 0 : state.metrics.wait_for_spec_ms = now
369 0 : .signed_duration_since(state.start_time)
370 0 : .to_std()
371 0 : .unwrap()
372 0 : .as_millis() as u64;
373 0 :
374 0 : // Reset start time, so that the total startup time that is calculated later will
375 0 : // not include the time that we waited for the spec.
376 0 : state.start_time = now;
377 0 : }
378 :
379 0 : launch_lsn_lease_bg_task_for_static(&compute);
380 0 :
381 0 : Ok(compute)
382 0 : }
383 :
384 0 : fn start_postgres(
385 0 : cli: &Cli,
386 0 : compute: Arc<ComputeNode>,
387 0 : ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
388 0 : // We got all we need, update the state.
389 0 : let mut state = compute.state.lock().unwrap();
390 0 : state.set_status(ComputeStatus::Init, &compute.state_changed);
391 0 :
392 0 : info!(
393 0 : "running compute with features: {:?}",
394 0 : state.pspec.as_ref().unwrap().spec.features
395 : );
396 : // before we release the mutex, fetch some parameters for later.
397 : let &ComputeSpec {
398 0 : swap_size_bytes,
399 0 : disk_quota_bytes,
400 0 : #[cfg(target_os = "linux")]
401 0 : disable_lfc_resizing,
402 0 : ..
403 0 : } = &state.pspec.as_ref().unwrap().spec;
404 0 : drop(state);
405 0 :
406 0 : // Launch remaining service threads
407 0 : let _monitor_handle = launch_monitor(&compute);
408 0 : let _configurator_handle = launch_configurator(&compute);
409 0 :
410 0 : let mut prestartup_failed = false;
411 0 : let mut delay_exit = false;
412 :
413 : // Resize swap to the desired size if the compute spec says so
414 0 : if let (Some(size_bytes), true) = (swap_size_bytes, cli.resize_swap_on_bind) {
415 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
416 : // *before* starting postgres.
417 : //
418 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
419 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
420 : // OOM-killed during startup because swap wasn't available yet.
421 0 : match resize_swap(size_bytes) {
422 : Ok(()) => {
423 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
424 0 : info!(%size_bytes, %size_mib, "resized swap");
425 : }
426 0 : Err(err) => {
427 0 : let err = err.context("failed to resize swap");
428 0 : error!("{err:#}");
429 :
430 : // Mark compute startup as failed; don't try to start postgres, and report this
431 : // error to the control plane when it next asks.
432 0 : prestartup_failed = true;
433 0 : compute.set_failed_status(err);
434 0 : delay_exit = true;
435 : }
436 : }
437 0 : }
438 :
439 : // Set disk quota if the compute spec says so
440 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) =
441 0 : (disk_quota_bytes, cli.set_disk_quota_for_fs.as_ref())
442 : {
443 0 : match set_disk_quota(disk_quota_bytes, disk_quota_fs_mountpoint) {
444 : Ok(()) => {
445 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
446 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
447 : }
448 0 : Err(err) => {
449 0 : let err = err.context("failed to set disk quota");
450 0 : error!("{err:#}");
451 :
452 : // Mark compute startup as failed; don't try to start postgres, and report this
453 : // error to the control plane when it next asks.
454 0 : prestartup_failed = true;
455 0 : compute.set_failed_status(err);
456 0 : delay_exit = true;
457 : }
458 : }
459 0 : }
460 :
461 : // Start Postgres
462 0 : let mut pg = None;
463 0 : if !prestartup_failed {
464 0 : pg = match compute.start_compute() {
465 0 : Ok(pg) => {
466 0 : info!(postmaster_pid = %pg.0.id(), "Postgres was started");
467 0 : Some(pg)
468 : }
469 0 : Err(err) => {
470 0 : error!("could not start the compute node: {:#}", err);
471 0 : compute.set_failed_status(err);
472 0 : delay_exit = true;
473 0 : None
474 : }
475 : };
476 : } else {
477 0 : warn!("skipping postgres startup because pre-startup step failed");
478 : }
479 :
480 : // Start the vm-monitor if directed to. The vm-monitor only runs on linux
481 : // because it requires cgroups.
482 : cfg_if::cfg_if! {
483 : if #[cfg(target_os = "linux")] {
484 : use std::env;
485 : use tokio_util::sync::CancellationToken;
486 :
487 : // Note: it seems like you can make a runtime in an inner scope and
488 : // if you start a task in it it won't be dropped. However, make it
489 : // in the outermost scope just to be safe.
490 0 : let rt = if env::var_os("AUTOSCALING").is_some() {
491 0 : Some(
492 0 : tokio::runtime::Builder::new_multi_thread()
493 0 : .worker_threads(4)
494 0 : .enable_all()
495 0 : .build()
496 0 : .expect("failed to create tokio runtime for monitor")
497 0 : )
498 : } else {
499 0 : None
500 : };
501 :
502 : // This token is used internally by the monitor to clean up all threads
503 0 : let token = CancellationToken::new();
504 :
505 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
506 0 : let pgconnstr = if disable_lfc_resizing.unwrap_or(false) {
507 0 : None
508 : } else {
509 0 : Some(cli.filecache_connstr.clone())
510 : };
511 :
512 0 : let vm_monitor = rt.as_ref().map(|rt| {
513 0 : rt.spawn(vm_monitor::start(
514 0 : Box::leak(Box::new(vm_monitor::Args {
515 0 : cgroup: Some(cli.cgroup.clone()),
516 0 : pgconnstr,
517 0 : addr: cli.vm_monitor_addr.clone(),
518 0 : })),
519 0 : token.clone(),
520 0 : ))
521 0 : });
522 0 : }
523 0 : }
524 0 :
525 0 : Ok((
526 0 : pg,
527 0 : StartPostgresResult {
528 0 : delay_exit,
529 0 : compute,
530 0 : #[cfg(target_os = "linux")]
531 0 : rt,
532 0 : #[cfg(target_os = "linux")]
533 0 : token,
534 0 : #[cfg(target_os = "linux")]
535 0 : vm_monitor,
536 0 : },
537 0 : ))
538 0 : }
539 :
540 : type PostgresHandle = (std::process::Child, std::thread::JoinHandle<()>);
541 :
542 : struct StartPostgresResult {
543 : delay_exit: bool,
544 : // passed through from WaitSpecResult
545 : compute: Arc<ComputeNode>,
546 :
547 : #[cfg(target_os = "linux")]
548 : rt: Option<tokio::runtime::Runtime>,
549 : #[cfg(target_os = "linux")]
550 : token: tokio_util::sync::CancellationToken,
551 : #[cfg(target_os = "linux")]
552 : vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
553 : }
554 :
555 0 : fn wait_postgres(pg: Option<PostgresHandle>) -> Result<WaitPostgresResult> {
556 0 : // Wait for the child Postgres process forever. In this state Ctrl+C will
557 0 : // propagate to Postgres and it will be shut down as well.
558 0 : let mut exit_code = None;
559 0 : if let Some((mut pg, logs_handle)) = pg {
560 0 : info!(postmaster_pid = %pg.id(), "Waiting for Postgres to exit");
561 :
562 0 : let ecode = pg
563 0 : .wait()
564 0 : .expect("failed to start waiting on Postgres process");
565 0 : PG_PID.store(0, Ordering::SeqCst);
566 0 :
567 0 : // Process has exited, so we can join the logs thread.
568 0 : let _ = logs_handle
569 0 : .join()
570 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
571 0 :
572 0 : info!("Postgres exited with code {}, shutting down", ecode);
573 0 : exit_code = ecode.code()
574 0 : }
575 :
576 0 : Ok(WaitPostgresResult { exit_code })
577 0 : }
578 :
579 : struct WaitPostgresResult {
580 : exit_code: Option<i32>,
581 : }
582 :
583 0 : fn cleanup_after_postgres_exit(
584 0 : StartPostgresResult {
585 0 : mut delay_exit,
586 0 : compute,
587 0 : #[cfg(target_os = "linux")]
588 0 : vm_monitor,
589 0 : #[cfg(target_os = "linux")]
590 0 : token,
591 0 : #[cfg(target_os = "linux")]
592 0 : rt,
593 0 : }: StartPostgresResult,
594 0 : ) -> Result<bool> {
595 : // Terminate the vm_monitor so it releases the file watcher on
596 : // /sys/fs/cgroup/neon-postgres.
597 : // Note: the vm-monitor only runs on linux because it requires cgroups.
598 : cfg_if::cfg_if! {
599 : if #[cfg(target_os = "linux")] {
600 0 : if let Some(handle) = vm_monitor {
601 0 : // Kills all threads spawned by the monitor
602 0 : token.cancel();
603 0 : // Kills the actual task running the monitor
604 0 : handle.abort();
605 0 :
606 0 : // If handle is some, rt must have been used to produce it, and
607 0 : // hence is also some
608 0 : rt.unwrap().shutdown_timeout(Duration::from_secs(2));
609 0 : }
610 : }
611 : }
612 :
613 : // Maybe sync safekeepers again, to speed up next startup
614 0 : let compute_state = compute.state.lock().unwrap().clone();
615 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
616 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
617 0 : info!("syncing safekeepers on shutdown");
618 0 : let storage_auth_token = pspec.storage_auth_token.clone();
619 0 : let lsn = compute.sync_safekeepers(storage_auth_token)?;
620 0 : info!("synced safekeepers at lsn {lsn}");
621 0 : }
622 :
623 0 : let mut state = compute.state.lock().unwrap();
624 0 : if state.status == ComputeStatus::TerminationPending {
625 0 : state.status = ComputeStatus::Terminated;
626 0 : compute.state_changed.notify_all();
627 0 : // we were asked to terminate gracefully, don't exit to avoid restart
628 0 : delay_exit = true
629 0 : }
630 0 : drop(state);
631 :
632 0 : if let Err(err) = compute.check_for_core_dumps() {
633 0 : error!("error while checking for core dumps: {err:?}");
634 0 : }
635 :
636 0 : Ok(delay_exit)
637 0 : }
638 :
639 0 : fn maybe_delay_exit(delay_exit: bool) {
640 0 : // If launch failed, keep serving HTTP requests for a while, so the cloud
641 0 : // control plane can get the actual error.
642 0 : if delay_exit {
643 0 : info!("giving control plane 30s to collect the error before shutdown");
644 0 : thread::sleep(Duration::from_secs(30));
645 0 : }
646 0 : }
647 :
648 0 : fn deinit_and_exit(WaitPostgresResult { exit_code }: WaitPostgresResult) -> ! {
649 0 : // Shutdown trace pipeline gracefully, so that it has a chance to send any
650 0 : // pending traces before we exit. Shutting down OTEL tracing provider may
651 0 : // hang for quite some time, see, for example:
652 0 : // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
653 0 : // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
654 0 : //
655 0 : // Yet, we want computes to shut down fast enough, as we may need a new one
656 0 : // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
657 0 : // complete, then just error out and exit the main thread.
658 0 : info!("shutting down tracing");
659 0 : let (sender, receiver) = mpsc::channel();
660 0 : let _ = thread::spawn(move || {
661 0 : tracing_utils::shutdown_tracing();
662 0 : sender.send(()).ok()
663 0 : });
664 0 : let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
665 0 : if shutdown_res.is_err() {
666 0 : error!("timed out while shutting down tracing, exiting anyway");
667 0 : }
668 :
669 0 : info!("shutting down");
670 0 : exit(exit_code.unwrap_or(1))
671 : }
672 :
673 : /// When compute_ctl is killed, send also termination signal to sync-safekeepers
674 : /// to prevent leakage. TODO: it is better to convert compute_ctl to async and
675 : /// wait for termination which would be easy then.
676 0 : fn handle_exit_signal(sig: i32) {
677 0 : info!("received {sig} termination signal");
678 0 : forward_termination_signal();
679 0 : exit(1);
680 : }
681 :
682 : #[cfg(test)]
683 : mod test {
684 : use clap::CommandFactory;
685 :
686 : use super::Cli;
687 :
688 : #[test]
689 1 : fn verify_cli() {
690 1 : Cli::command().debug_assert()
691 1 : }
692 : }
|
