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::http::server::Server;
51 : use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
52 : use signal_hook::consts::{SIGQUIT, SIGTERM};
53 : use signal_hook::{consts::SIGINT, iterator::Signals};
54 : use tracing::{error, info, warn};
55 : use url::Url;
56 :
57 : use compute_api::responses::{ComputeCtlConfig, ComputeStatus};
58 : use compute_api::spec::ComputeSpec;
59 :
60 : use compute_tools::compute::{
61 : forward_termination_signal, ComputeNode, ComputeState, ParsedSpec, PG_PID,
62 : };
63 : use compute_tools::configurator::launch_configurator;
64 : use compute_tools::extension_server::get_pg_version_string;
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 : /// The port to bind the external listening HTTP server to. Clients running
98 : /// outside the compute will talk to the compute through this port. Keep
99 : /// the previous name for this argument around for a smoother release
100 : /// with the control plane.
101 1 : #[arg(long, default_value_t = 3080)]
102 0 : pub external_http_port: u16,
103 :
104 : /// The port to bind the internal listening HTTP server to. Clients include
105 : /// the neon extension (for installing remote extensions) and local_proxy.
106 1 : #[arg(long, default_value_t = 3081)]
107 0 : pub internal_http_port: u16,
108 :
109 : #[arg(short = 'D', long, value_name = "DATADIR")]
110 0 : pub pgdata: String,
111 :
112 : #[arg(short = 'C', long, value_name = "DATABASE_URL")]
113 0 : pub connstr: String,
114 :
115 : #[cfg(target_os = "linux")]
116 : #[arg(long, default_value = "neon-postgres")]
117 0 : pub cgroup: String,
118 :
119 : #[cfg(target_os = "linux")]
120 : #[arg(
121 : long,
122 : default_value = "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable application_name=vm-monitor"
123 : )]
124 0 : pub filecache_connstr: String,
125 :
126 : #[cfg(target_os = "linux")]
127 : #[arg(long, default_value = "0.0.0.0:10301")]
128 0 : pub vm_monitor_addr: String,
129 :
130 : #[arg(long, action = clap::ArgAction::SetTrue)]
131 0 : pub resize_swap_on_bind: bool,
132 :
133 : #[arg(long)]
134 : pub set_disk_quota_for_fs: Option<String>,
135 :
136 : #[arg(short = 's', long = "spec", group = "spec")]
137 : pub spec_json: Option<String>,
138 :
139 : #[arg(short = 'S', long, group = "spec-path")]
140 : pub spec_path: Option<OsString>,
141 :
142 : #[arg(short = 'i', long, group = "compute-id")]
143 0 : pub compute_id: String,
144 :
145 : #[arg(short = 'p', long, conflicts_with_all = ["spec", "spec-path"], value_name = "CONTROL_PLANE_API_BASE_URL")]
146 : pub control_plane_uri: Option<String>,
147 : }
148 :
149 0 : fn main() -> Result<()> {
150 0 : let cli = Cli::parse();
151 :
152 : // For historical reasons, the main thread that processes the spec and launches postgres
153 : // is synchronous, but we always have this tokio runtime available and we "enter" it so
154 : // that you can use tokio::spawn() and tokio::runtime::Handle::current().block_on(...)
155 : // from all parts of compute_ctl.
156 0 : let runtime = tokio::runtime::Builder::new_multi_thread()
157 0 : .enable_all()
158 0 : .build()?;
159 0 : let _rt_guard = runtime.enter();
160 :
161 0 : let build_tag = runtime.block_on(init())?;
162 :
163 0 : let scenario = failpoint_support::init();
164 0 :
165 0 : // enable core dumping for all child processes
166 0 : setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
167 :
168 0 : let (pg_handle, start_pg_result) = {
169 : // Enter startup tracing context
170 0 : let _startup_context_guard = startup_context_from_env();
171 :
172 0 : let cli_spec = try_spec_from_cli(&cli)?;
173 :
174 0 : let compute = wait_spec(build_tag, &cli, cli_spec)?;
175 :
176 0 : start_postgres(&cli, compute)?
177 :
178 : // Startup is finished, exit the startup tracing span
179 : };
180 :
181 : // PostgreSQL is now running, if startup was successful. Wait until it exits.
182 0 : let wait_pg_result = wait_postgres(pg_handle)?;
183 :
184 0 : let delay_exit = cleanup_after_postgres_exit(start_pg_result)?;
185 :
186 0 : maybe_delay_exit(delay_exit);
187 0 :
188 0 : scenario.teardown();
189 0 :
190 0 : deinit_and_exit(wait_pg_result);
191 0 : }
192 :
193 0 : async fn init() -> Result<String> {
194 0 : init_tracing_and_logging(DEFAULT_LOG_LEVEL).await?;
195 :
196 0 : let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
197 0 : thread::spawn(move || {
198 0 : for sig in signals.forever() {
199 0 : handle_exit_signal(sig);
200 0 : }
201 0 : });
202 0 :
203 0 : let build_tag = option_env!("BUILD_TAG")
204 0 : .unwrap_or(BUILD_TAG_DEFAULT)
205 0 : .to_string();
206 0 : info!("build_tag: {build_tag}");
207 :
208 0 : Ok(build_tag)
209 0 : }
210 :
211 0 : fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
212 0 : // Extract OpenTelemetry context for the startup actions from the
213 0 : // TRACEPARENT and TRACESTATE env variables, and attach it to the current
214 0 : // tracing context.
215 0 : //
216 0 : // This is used to propagate the context for the 'start_compute' operation
217 0 : // from the neon control plane. This allows linking together the wider
218 0 : // 'start_compute' operation that creates the compute container, with the
219 0 : // startup actions here within the container.
220 0 : //
221 0 : // There is no standard for passing context in env variables, but a lot of
222 0 : // tools use TRACEPARENT/TRACESTATE, so we use that convention too. See
223 0 : // https://github.com/open-telemetry/opentelemetry-specification/issues/740
224 0 : //
225 0 : // Switch to the startup context here, and exit it once the startup has
226 0 : // completed and Postgres is up and running.
227 0 : //
228 0 : // If this pod is pre-created without binding it to any particular endpoint
229 0 : // yet, this isn't the right place to enter the startup context. In that
230 0 : // case, the control plane should pass the tracing context as part of the
231 0 : // /configure API call.
232 0 : //
233 0 : // NOTE: This is supposed to only cover the *startup* actions. Once
234 0 : // postgres is configured and up-and-running, we exit this span. Any other
235 0 : // actions that are performed on incoming HTTP requests, for example, are
236 0 : // performed in separate spans.
237 0 : //
238 0 : // XXX: If the pod is restarted, we perform the startup actions in the same
239 0 : // context as the original startup actions, which probably doesn't make
240 0 : // sense.
241 0 : let mut startup_tracing_carrier: HashMap<String, String> = HashMap::new();
242 0 : if let Ok(val) = std::env::var("TRACEPARENT") {
243 0 : startup_tracing_carrier.insert("traceparent".to_string(), val);
244 0 : }
245 0 : if let Ok(val) = std::env::var("TRACESTATE") {
246 0 : startup_tracing_carrier.insert("tracestate".to_string(), val);
247 0 : }
248 0 : if !startup_tracing_carrier.is_empty() {
249 : use opentelemetry::propagation::TextMapPropagator;
250 : use opentelemetry_sdk::propagation::TraceContextPropagator;
251 0 : let guard = TraceContextPropagator::new()
252 0 : .extract(&startup_tracing_carrier)
253 0 : .attach();
254 0 : info!("startup tracing context attached");
255 0 : Some(guard)
256 : } else {
257 0 : None
258 : }
259 0 : }
260 :
261 0 : fn try_spec_from_cli(cli: &Cli) -> Result<CliSpecParams> {
262 : // First, try to get cluster spec from the cli argument
263 0 : if let Some(ref spec_json) = cli.spec_json {
264 0 : info!("got spec from cli argument {}", spec_json);
265 : return Ok(CliSpecParams {
266 0 : spec: Some(serde_json::from_str(spec_json)?),
267 0 : compute_ctl_config: ComputeCtlConfig::default(),
268 : live_config_allowed: false,
269 : });
270 0 : }
271 :
272 : // Second, try to read it from the file if path is provided
273 0 : if let Some(ref spec_path) = cli.spec_path {
274 0 : let file = File::open(Path::new(spec_path))?;
275 : return Ok(CliSpecParams {
276 0 : spec: Some(serde_json::from_reader(file)?),
277 0 : compute_ctl_config: ComputeCtlConfig::default(),
278 : live_config_allowed: true,
279 : });
280 0 : }
281 0 :
282 0 : if cli.control_plane_uri.is_none() {
283 0 : panic!("must specify --control-plane-uri");
284 0 : };
285 0 :
286 0 : match get_spec_from_control_plane(cli.control_plane_uri.as_ref().unwrap(), &cli.compute_id) {
287 0 : Ok(resp) => Ok(CliSpecParams {
288 0 : spec: resp.0,
289 0 : compute_ctl_config: resp.1,
290 0 : live_config_allowed: true,
291 0 : }),
292 0 : Err(e) => {
293 0 : error!(
294 0 : "cannot get response from control plane: {}\n\
295 0 : neither spec nor confirmation that compute is in the Empty state was received",
296 : e
297 : );
298 0 : Err(e)
299 : }
300 : }
301 0 : }
302 :
303 : struct CliSpecParams {
304 : /// If a spec was provided via CLI or file, the [`ComputeSpec`]
305 : spec: Option<ComputeSpec>,
306 : #[allow(dead_code)]
307 : compute_ctl_config: ComputeCtlConfig,
308 : live_config_allowed: bool,
309 : }
310 :
311 0 : fn wait_spec(
312 0 : build_tag: String,
313 0 : cli: &Cli,
314 0 : CliSpecParams {
315 0 : spec,
316 0 : live_config_allowed,
317 0 : compute_ctl_config: _,
318 0 : }: CliSpecParams,
319 0 : ) -> Result<Arc<ComputeNode>> {
320 0 : let mut new_state = ComputeState::new();
321 : let spec_set;
322 :
323 0 : if let Some(spec) = spec {
324 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
325 0 : info!("new pspec.spec: {:?}", pspec.spec);
326 0 : new_state.pspec = Some(pspec);
327 0 : spec_set = true;
328 0 : } else {
329 0 : spec_set = false;
330 0 : }
331 0 : let connstr = Url::parse(&cli.connstr).context("cannot parse connstr as a URL")?;
332 0 : let conn_conf = postgres::config::Config::from_str(connstr.as_str())
333 0 : .context("cannot build postgres config from connstr")?;
334 0 : let tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr.as_str())
335 0 : .context("cannot build tokio postgres config from connstr")?;
336 0 : let compute_node = ComputeNode {
337 0 : compute_id: cli.compute_id.clone(),
338 0 : connstr,
339 0 : conn_conf,
340 0 : tokio_conn_conf,
341 0 : pgdata: cli.pgdata.clone(),
342 0 : pgbin: cli.pgbin.clone(),
343 0 : pgversion: get_pg_version_string(&cli.pgbin),
344 0 : external_http_port: cli.external_http_port,
345 0 : internal_http_port: cli.internal_http_port,
346 0 : live_config_allowed,
347 0 : state: Mutex::new(new_state),
348 0 : state_changed: Condvar::new(),
349 0 : ext_remote_storage: cli.remote_ext_config.clone(),
350 0 : ext_download_progress: RwLock::new(HashMap::new()),
351 0 : build_tag,
352 0 : };
353 0 : let compute = Arc::new(compute_node);
354 0 :
355 0 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
356 0 : // available for binding. Prewarming helps Postgres start quicker later,
357 0 : // because QEMU will already have its memory allocated from the host, and
358 0 : // the necessary binaries will already be cached.
359 0 : if !spec_set {
360 0 : compute.prewarm_postgres()?;
361 0 : }
362 :
363 : // Launch the external HTTP server first, so that we can serve control plane
364 : // requests while configuration is still in progress.
365 0 : Server::External(cli.external_http_port).launch(&compute);
366 0 :
367 0 : // The internal HTTP server could be launched later, but there isn't much
368 0 : // sense in waiting.
369 0 : Server::Internal(cli.internal_http_port).launch(&compute);
370 0 :
371 0 : if !spec_set {
372 : // No spec provided, hang waiting for it.
373 0 : info!("no compute spec provided, waiting");
374 :
375 0 : let mut state = compute.state.lock().unwrap();
376 0 : while state.status != ComputeStatus::ConfigurationPending {
377 0 : state = compute.state_changed.wait(state).unwrap();
378 0 :
379 0 : if state.status == ComputeStatus::ConfigurationPending {
380 0 : info!("got spec, continue configuration");
381 : // Spec is already set by the http server handler.
382 0 : break;
383 0 : }
384 : }
385 :
386 : // Record for how long we slept waiting for the spec.
387 0 : let now = Utc::now();
388 0 : state.metrics.wait_for_spec_ms = now
389 0 : .signed_duration_since(state.start_time)
390 0 : .to_std()
391 0 : .unwrap()
392 0 : .as_millis() as u64;
393 0 :
394 0 : // Reset start time, so that the total startup time that is calculated later will
395 0 : // not include the time that we waited for the spec.
396 0 : state.start_time = now;
397 0 : }
398 :
399 0 : launch_lsn_lease_bg_task_for_static(&compute);
400 0 :
401 0 : Ok(compute)
402 0 : }
403 :
404 0 : fn start_postgres(
405 0 : cli: &Cli,
406 0 : compute: Arc<ComputeNode>,
407 0 : ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
408 0 : // We got all we need, update the state.
409 0 : let mut state = compute.state.lock().unwrap();
410 0 : state.set_status(ComputeStatus::Init, &compute.state_changed);
411 0 :
412 0 : info!(
413 0 : "running compute with features: {:?}",
414 0 : state.pspec.as_ref().unwrap().spec.features
415 : );
416 : // before we release the mutex, fetch some parameters for later.
417 : let &ComputeSpec {
418 0 : swap_size_bytes,
419 0 : disk_quota_bytes,
420 0 : #[cfg(target_os = "linux")]
421 0 : disable_lfc_resizing,
422 0 : ..
423 0 : } = &state.pspec.as_ref().unwrap().spec;
424 0 : drop(state);
425 0 :
426 0 : // Launch remaining service threads
427 0 : let _monitor_handle = launch_monitor(&compute);
428 0 : let _configurator_handle = launch_configurator(&compute);
429 0 :
430 0 : let mut prestartup_failed = false;
431 0 : let mut delay_exit = false;
432 :
433 : // Resize swap to the desired size if the compute spec says so
434 0 : if let (Some(size_bytes), true) = (swap_size_bytes, cli.resize_swap_on_bind) {
435 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
436 : // *before* starting postgres.
437 : //
438 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
439 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
440 : // OOM-killed during startup because swap wasn't available yet.
441 0 : match resize_swap(size_bytes) {
442 : Ok(()) => {
443 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
444 0 : info!(%size_bytes, %size_mib, "resized swap");
445 : }
446 0 : Err(err) => {
447 0 : let err = err.context("failed to resize swap");
448 0 : error!("{err:#}");
449 :
450 : // Mark compute startup as failed; don't try to start postgres, and report this
451 : // error to the control plane when it next asks.
452 0 : prestartup_failed = true;
453 0 : compute.set_failed_status(err);
454 0 : delay_exit = true;
455 : }
456 : }
457 0 : }
458 :
459 : // Set disk quota if the compute spec says so
460 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) =
461 0 : (disk_quota_bytes, cli.set_disk_quota_for_fs.as_ref())
462 : {
463 0 : match set_disk_quota(disk_quota_bytes, disk_quota_fs_mountpoint) {
464 : Ok(()) => {
465 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
466 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
467 : }
468 0 : Err(err) => {
469 0 : let err = err.context("failed to set disk quota");
470 0 : error!("{err:#}");
471 :
472 : // Mark compute startup as failed; don't try to start postgres, and report this
473 : // error to the control plane when it next asks.
474 0 : prestartup_failed = true;
475 0 : compute.set_failed_status(err);
476 0 : delay_exit = true;
477 : }
478 : }
479 0 : }
480 :
481 : // Start Postgres
482 0 : let mut pg = None;
483 0 : if !prestartup_failed {
484 0 : pg = match compute.start_compute() {
485 0 : Ok(pg) => {
486 0 : info!(postmaster_pid = %pg.0.id(), "Postgres was started");
487 0 : Some(pg)
488 : }
489 0 : Err(err) => {
490 0 : error!("could not start the compute node: {:#}", err);
491 0 : compute.set_failed_status(err);
492 0 : delay_exit = true;
493 0 : None
494 : }
495 : };
496 : } else {
497 0 : warn!("skipping postgres startup because pre-startup step failed");
498 : }
499 :
500 : // Start the vm-monitor if directed to. The vm-monitor only runs on linux
501 : // because it requires cgroups.
502 : cfg_if::cfg_if! {
503 : if #[cfg(target_os = "linux")] {
504 : use std::env;
505 : use tokio_util::sync::CancellationToken;
506 :
507 : // This token is used internally by the monitor to clean up all threads
508 0 : let token = CancellationToken::new();
509 :
510 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
511 0 : let pgconnstr = if disable_lfc_resizing.unwrap_or(false) {
512 0 : None
513 : } else {
514 0 : Some(cli.filecache_connstr.clone())
515 : };
516 :
517 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
518 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
519 0 : Box::leak(Box::new(vm_monitor::Args {
520 0 : cgroup: Some(cli.cgroup.clone()),
521 0 : pgconnstr,
522 0 : addr: cli.vm_monitor_addr.clone(),
523 0 : })),
524 0 : token.clone(),
525 0 : ));
526 0 : Some(vm_monitor)
527 : } else {
528 0 : None
529 : };
530 : }
531 : }
532 :
533 0 : Ok((
534 0 : pg,
535 0 : StartPostgresResult {
536 0 : delay_exit,
537 0 : compute,
538 0 : #[cfg(target_os = "linux")]
539 0 : token,
540 0 : #[cfg(target_os = "linux")]
541 0 : vm_monitor,
542 0 : },
543 0 : ))
544 0 : }
545 :
546 : type PostgresHandle = (std::process::Child, tokio::task::JoinHandle<Result<()>>);
547 :
548 : struct StartPostgresResult {
549 : delay_exit: bool,
550 : // passed through from WaitSpecResult
551 : compute: Arc<ComputeNode>,
552 :
553 : #[cfg(target_os = "linux")]
554 : token: tokio_util::sync::CancellationToken,
555 : #[cfg(target_os = "linux")]
556 : vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
557 : }
558 :
559 0 : fn wait_postgres(pg: Option<PostgresHandle>) -> Result<WaitPostgresResult> {
560 0 : // Wait for the child Postgres process forever. In this state Ctrl+C will
561 0 : // propagate to Postgres and it will be shut down as well.
562 0 : let mut exit_code = None;
563 0 : if let Some((mut pg, logs_handle)) = pg {
564 0 : info!(postmaster_pid = %pg.id(), "Waiting for Postgres to exit");
565 :
566 0 : let ecode = pg
567 0 : .wait()
568 0 : .expect("failed to start waiting on Postgres process");
569 0 : PG_PID.store(0, Ordering::SeqCst);
570 0 :
571 0 : // Process has exited. Wait for the log collecting task to finish.
572 0 : let _ = tokio::runtime::Handle::current()
573 0 : .block_on(logs_handle)
574 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
575 0 :
576 0 : info!("Postgres exited with code {}, shutting down", ecode);
577 0 : exit_code = ecode.code()
578 0 : }
579 :
580 0 : Ok(WaitPostgresResult { exit_code })
581 0 : }
582 :
583 : struct WaitPostgresResult {
584 : exit_code: Option<i32>,
585 : }
586 :
587 0 : fn cleanup_after_postgres_exit(
588 0 : StartPostgresResult {
589 0 : mut delay_exit,
590 0 : compute,
591 0 : #[cfg(target_os = "linux")]
592 0 : vm_monitor,
593 0 : #[cfg(target_os = "linux")]
594 0 : token,
595 0 : }: StartPostgresResult,
596 0 : ) -> Result<bool> {
597 : // Terminate the vm_monitor so it releases the file watcher on
598 : // /sys/fs/cgroup/neon-postgres.
599 : // Note: the vm-monitor only runs on linux because it requires cgroups.
600 : cfg_if::cfg_if! {
601 : if #[cfg(target_os = "linux")] {
602 0 : if let Some(handle) = vm_monitor {
603 0 : // Kills all threads spawned by the monitor
604 0 : token.cancel();
605 0 : // Kills the actual task running the monitor
606 0 : handle.abort();
607 0 : }
608 : }
609 : }
610 :
611 : // Maybe sync safekeepers again, to speed up next startup
612 0 : let compute_state = compute.state.lock().unwrap().clone();
613 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
614 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
615 0 : info!("syncing safekeepers on shutdown");
616 0 : let storage_auth_token = pspec.storage_auth_token.clone();
617 0 : let lsn = compute.sync_safekeepers(storage_auth_token)?;
618 0 : info!("synced safekeepers at lsn {lsn}");
619 0 : }
620 :
621 0 : let mut state = compute.state.lock().unwrap();
622 0 : if state.status == ComputeStatus::TerminationPending {
623 0 : state.status = ComputeStatus::Terminated;
624 0 : compute.state_changed.notify_all();
625 0 : // we were asked to terminate gracefully, don't exit to avoid restart
626 0 : delay_exit = true
627 0 : }
628 0 : drop(state);
629 :
630 0 : if let Err(err) = compute.check_for_core_dumps() {
631 0 : error!("error while checking for core dumps: {err:?}");
632 0 : }
633 :
634 0 : Ok(delay_exit)
635 0 : }
636 :
637 0 : fn maybe_delay_exit(delay_exit: bool) {
638 0 : // If launch failed, keep serving HTTP requests for a while, so the cloud
639 0 : // control plane can get the actual error.
640 0 : if delay_exit {
641 0 : info!("giving control plane 30s to collect the error before shutdown");
642 0 : thread::sleep(Duration::from_secs(30));
643 0 : }
644 0 : }
645 :
646 0 : fn deinit_and_exit(WaitPostgresResult { exit_code }: WaitPostgresResult) -> ! {
647 0 : // Shutdown trace pipeline gracefully, so that it has a chance to send any
648 0 : // pending traces before we exit. Shutting down OTEL tracing provider may
649 0 : // hang for quite some time, see, for example:
650 0 : // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
651 0 : // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
652 0 : //
653 0 : // Yet, we want computes to shut down fast enough, as we may need a new one
654 0 : // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
655 0 : // complete, then just error out and exit the main thread.
656 0 : info!("shutting down tracing");
657 0 : let (sender, receiver) = mpsc::channel();
658 0 : let _ = thread::spawn(move || {
659 0 : tracing_utils::shutdown_tracing();
660 0 : sender.send(()).ok()
661 0 : });
662 0 : let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
663 0 : if shutdown_res.is_err() {
664 0 : error!("timed out while shutting down tracing, exiting anyway");
665 0 : }
666 :
667 0 : info!("shutting down");
668 0 : exit(exit_code.unwrap_or(1))
669 : }
670 :
671 : /// When compute_ctl is killed, send also termination signal to sync-safekeepers
672 : /// to prevent leakage. TODO: it is better to convert compute_ctl to async and
673 : /// wait for termination which would be easy then.
674 0 : fn handle_exit_signal(sig: i32) {
675 0 : info!("received {sig} termination signal");
676 0 : forward_termination_signal();
677 0 : exit(1);
678 : }
679 :
680 : #[cfg(test)]
681 : mod test {
682 : use clap::CommandFactory;
683 :
684 : use super::Cli;
685 :
686 : #[test]
687 1 : fn verify_cli() {
688 1 : Cli::command().debug_assert()
689 1 : }
690 : }
|
