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::fs::File;
38 : use std::path::Path;
39 : use std::process::exit;
40 : use std::sync::atomic::Ordering;
41 : use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
42 : use std::{thread, time::Duration};
43 :
44 : use anyhow::{Context, Result};
45 : use chrono::Utc;
46 : use clap::Arg;
47 : use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
48 : use signal_hook::consts::{SIGQUIT, SIGTERM};
49 : use signal_hook::{consts::SIGINT, iterator::Signals};
50 : use tracing::{error, info, warn};
51 : use url::Url;
52 :
53 : use compute_api::responses::ComputeStatus;
54 : use compute_api::spec::ComputeSpec;
55 :
56 : use compute_tools::compute::{
57 : forward_termination_signal, ComputeNode, ComputeState, ParsedSpec, PG_PID,
58 : };
59 : use compute_tools::configurator::launch_configurator;
60 : use compute_tools::extension_server::get_pg_version;
61 : use compute_tools::http::api::launch_http_server;
62 : use compute_tools::logger::*;
63 : use compute_tools::monitor::launch_monitor;
64 : use compute_tools::params::*;
65 : use compute_tools::spec::*;
66 : use compute_tools::swap::resize_swap;
67 : use rlimit::{setrlimit, Resource};
68 :
69 : // this is an arbitrary build tag. Fine as a default / for testing purposes
70 : // in-case of not-set environment var
71 : const BUILD_TAG_DEFAULT: &str = "latest";
72 :
73 0 : fn main() -> Result<()> {
74 0 : let (build_tag, clap_args) = init()?;
75 :
76 : // enable core dumping for all child processes
77 0 : setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
78 :
79 0 : let (pg_handle, start_pg_result) = {
80 : // Enter startup tracing context
81 0 : let _startup_context_guard = startup_context_from_env();
82 :
83 0 : let cli_args = process_cli(&clap_args)?;
84 :
85 0 : let cli_spec = try_spec_from_cli(&clap_args, &cli_args)?;
86 :
87 0 : let wait_spec_result = wait_spec(build_tag, cli_args, cli_spec)?;
88 :
89 0 : start_postgres(&clap_args, wait_spec_result)?
90 :
91 : // Startup is finished, exit the startup tracing span
92 : };
93 :
94 : // PostgreSQL is now running, if startup was successful. Wait until it exits.
95 0 : let wait_pg_result = wait_postgres(pg_handle)?;
96 :
97 0 : let delay_exit = cleanup_after_postgres_exit(start_pg_result)?;
98 :
99 0 : maybe_delay_exit(delay_exit);
100 0 :
101 0 : deinit_and_exit(wait_pg_result);
102 0 : }
103 :
104 0 : fn init() -> Result<(String, clap::ArgMatches)> {
105 0 : init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
106 :
107 0 : let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
108 0 : thread::spawn(move || {
109 0 : for sig in signals.forever() {
110 0 : handle_exit_signal(sig);
111 0 : }
112 0 : });
113 0 :
114 0 : let build_tag = option_env!("BUILD_TAG")
115 0 : .unwrap_or(BUILD_TAG_DEFAULT)
116 0 : .to_string();
117 0 : info!("build_tag: {build_tag}");
118 :
119 0 : Ok((build_tag, cli().get_matches()))
120 0 : }
121 :
122 0 : fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
123 0 : let pgbin_default = "postgres";
124 0 : let pgbin = matches
125 0 : .get_one::<String>("pgbin")
126 0 : .map(|s| s.as_str())
127 0 : .unwrap_or(pgbin_default);
128 0 :
129 0 : let ext_remote_storage = matches
130 0 : .get_one::<String>("remote-ext-config")
131 0 : // Compatibility hack: if the control plane specified any remote-ext-config
132 0 : // use the default value for extension storage proxy gateway.
133 0 : // Remove this once the control plane is updated to pass the gateway URL
134 0 : .map(|conf| {
135 0 : if conf.starts_with("http") {
136 0 : conf.trim_end_matches('/')
137 : } else {
138 0 : "http://pg-ext-s3-gateway"
139 : }
140 0 : });
141 0 :
142 0 : let http_port = *matches
143 0 : .get_one::<u16>("http-port")
144 0 : .expect("http-port is required");
145 0 : let pgdata = matches
146 0 : .get_one::<String>("pgdata")
147 0 : .expect("PGDATA path is required");
148 0 : let connstr = matches
149 0 : .get_one::<String>("connstr")
150 0 : .expect("Postgres connection string is required");
151 0 : let spec_json = matches.get_one::<String>("spec");
152 0 : let spec_path = matches.get_one::<String>("spec-path");
153 0 : let resize_swap_on_bind = matches.get_flag("resize-swap-on-bind");
154 0 :
155 0 : Ok(ProcessCliResult {
156 0 : connstr,
157 0 : pgdata,
158 0 : pgbin,
159 0 : ext_remote_storage,
160 0 : http_port,
161 0 : spec_json,
162 0 : spec_path,
163 0 : resize_swap_on_bind,
164 0 : })
165 0 : }
166 :
167 : struct ProcessCliResult<'clap> {
168 : connstr: &'clap str,
169 : pgdata: &'clap str,
170 : pgbin: &'clap str,
171 : ext_remote_storage: Option<&'clap str>,
172 : http_port: u16,
173 : spec_json: Option<&'clap String>,
174 : spec_path: Option<&'clap String>,
175 : resize_swap_on_bind: bool,
176 : }
177 :
178 0 : fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
179 0 : // Extract OpenTelemetry context for the startup actions from the
180 0 : // TRACEPARENT and TRACESTATE env variables, and attach it to the current
181 0 : // tracing context.
182 0 : //
183 0 : // This is used to propagate the context for the 'start_compute' operation
184 0 : // from the neon control plane. This allows linking together the wider
185 0 : // 'start_compute' operation that creates the compute container, with the
186 0 : // startup actions here within the container.
187 0 : //
188 0 : // There is no standard for passing context in env variables, but a lot of
189 0 : // tools use TRACEPARENT/TRACESTATE, so we use that convention too. See
190 0 : // https://github.com/open-telemetry/opentelemetry-specification/issues/740
191 0 : //
192 0 : // Switch to the startup context here, and exit it once the startup has
193 0 : // completed and Postgres is up and running.
194 0 : //
195 0 : // If this pod is pre-created without binding it to any particular endpoint
196 0 : // yet, this isn't the right place to enter the startup context. In that
197 0 : // case, the control plane should pass the tracing context as part of the
198 0 : // /configure API call.
199 0 : //
200 0 : // NOTE: This is supposed to only cover the *startup* actions. Once
201 0 : // postgres is configured and up-and-running, we exit this span. Any other
202 0 : // actions that are performed on incoming HTTP requests, for example, are
203 0 : // performed in separate spans.
204 0 : //
205 0 : // XXX: If the pod is restarted, we perform the startup actions in the same
206 0 : // context as the original startup actions, which probably doesn't make
207 0 : // sense.
208 0 : let mut startup_tracing_carrier: HashMap<String, String> = HashMap::new();
209 0 : if let Ok(val) = std::env::var("TRACEPARENT") {
210 0 : startup_tracing_carrier.insert("traceparent".to_string(), val);
211 0 : }
212 0 : if let Ok(val) = std::env::var("TRACESTATE") {
213 0 : startup_tracing_carrier.insert("tracestate".to_string(), val);
214 0 : }
215 0 : if !startup_tracing_carrier.is_empty() {
216 : use opentelemetry::propagation::TextMapPropagator;
217 : use opentelemetry::sdk::propagation::TraceContextPropagator;
218 0 : let guard = TraceContextPropagator::new()
219 0 : .extract(&startup_tracing_carrier)
220 0 : .attach();
221 0 : info!("startup tracing context attached");
222 0 : Some(guard)
223 : } else {
224 0 : None
225 : }
226 0 : }
227 :
228 0 : fn try_spec_from_cli(
229 0 : matches: &clap::ArgMatches,
230 0 : ProcessCliResult {
231 0 : spec_json,
232 0 : spec_path,
233 0 : ..
234 0 : }: &ProcessCliResult,
235 0 : ) -> Result<CliSpecParams> {
236 0 : let compute_id = matches.get_one::<String>("compute-id");
237 0 : let control_plane_uri = matches.get_one::<String>("control-plane-uri");
238 0 :
239 0 : let spec;
240 0 : let mut live_config_allowed = false;
241 0 : match spec_json {
242 : // First, try to get cluster spec from the cli argument
243 0 : Some(json) => {
244 0 : info!("got spec from cli argument {}", json);
245 0 : spec = Some(serde_json::from_str(json)?);
246 : }
247 : None => {
248 : // Second, try to read it from the file if path is provided
249 0 : if let Some(sp) = spec_path {
250 0 : let path = Path::new(sp);
251 0 : let file = File::open(path)?;
252 0 : spec = Some(serde_json::from_reader(file)?);
253 0 : live_config_allowed = true;
254 0 : } else if let Some(id) = compute_id {
255 0 : if let Some(cp_base) = control_plane_uri {
256 0 : live_config_allowed = true;
257 0 : spec = match get_spec_from_control_plane(cp_base, id) {
258 0 : Ok(s) => s,
259 0 : Err(e) => {
260 0 : error!("cannot get response from control plane: {}", e);
261 0 : panic!("neither spec nor confirmation that compute is in the Empty state was received");
262 : }
263 : };
264 : } else {
265 0 : panic!("must specify both --control-plane-uri and --compute-id or none");
266 : }
267 : } else {
268 0 : panic!(
269 0 : "compute spec should be provided by one of the following ways: \
270 0 : --spec OR --spec-path OR --control-plane-uri and --compute-id"
271 0 : );
272 : }
273 : }
274 : };
275 :
276 0 : Ok(CliSpecParams {
277 0 : spec,
278 0 : live_config_allowed,
279 0 : })
280 0 : }
281 :
282 : struct CliSpecParams {
283 : /// If a spec was provided via CLI or file, the [`ComputeSpec`]
284 : spec: Option<ComputeSpec>,
285 : live_config_allowed: bool,
286 : }
287 :
288 0 : fn wait_spec(
289 0 : build_tag: String,
290 0 : ProcessCliResult {
291 0 : connstr,
292 0 : pgdata,
293 0 : pgbin,
294 0 : ext_remote_storage,
295 0 : resize_swap_on_bind,
296 0 : http_port,
297 0 : ..
298 0 : }: ProcessCliResult,
299 0 : CliSpecParams {
300 0 : spec,
301 0 : live_config_allowed,
302 0 : }: CliSpecParams,
303 0 : ) -> Result<WaitSpecResult> {
304 0 : let mut new_state = ComputeState::new();
305 : let spec_set;
306 :
307 0 : if let Some(spec) = spec {
308 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
309 0 : info!("new pspec.spec: {:?}", pspec.spec);
310 0 : new_state.pspec = Some(pspec);
311 0 : spec_set = true;
312 0 : } else {
313 0 : spec_set = false;
314 0 : }
315 0 : let compute_node = ComputeNode {
316 0 : connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
317 0 : pgdata: pgdata.to_string(),
318 0 : pgbin: pgbin.to_string(),
319 0 : pgversion: get_pg_version(pgbin),
320 0 : live_config_allowed,
321 0 : state: Mutex::new(new_state),
322 0 : state_changed: Condvar::new(),
323 0 : ext_remote_storage: ext_remote_storage.map(|s| s.to_string()),
324 0 : ext_download_progress: RwLock::new(HashMap::new()),
325 0 : build_tag,
326 0 : };
327 0 : let compute = Arc::new(compute_node);
328 0 :
329 0 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
330 0 : // available for binding. Prewarming helps Postgres start quicker later,
331 0 : // because QEMU will already have its memory allocated from the host, and
332 0 : // the necessary binaries will already be cached.
333 0 : if !spec_set {
334 0 : compute.prewarm_postgres()?;
335 0 : }
336 :
337 : // Launch http service first, so that we can serve control-plane requests
338 : // while configuration is still in progress.
339 0 : let _http_handle =
340 0 : launch_http_server(http_port, &compute).expect("cannot launch http endpoint thread");
341 0 :
342 0 : if !spec_set {
343 : // No spec provided, hang waiting for it.
344 0 : info!("no compute spec provided, waiting");
345 :
346 0 : let mut state = compute.state.lock().unwrap();
347 0 : while state.status != ComputeStatus::ConfigurationPending {
348 0 : state = compute.state_changed.wait(state).unwrap();
349 0 :
350 0 : if state.status == ComputeStatus::ConfigurationPending {
351 0 : info!("got spec, continue configuration");
352 : // Spec is already set by the http server handler.
353 0 : break;
354 0 : }
355 : }
356 :
357 : // Record for how long we slept waiting for the spec.
358 0 : let now = Utc::now();
359 0 : state.metrics.wait_for_spec_ms = now
360 0 : .signed_duration_since(state.start_time)
361 0 : .to_std()
362 0 : .unwrap()
363 0 : .as_millis() as u64;
364 0 :
365 0 : // Reset start time, so that the total startup time that is calculated later will
366 0 : // not include the time that we waited for the spec.
367 0 : state.start_time = now;
368 0 : }
369 :
370 0 : launch_lsn_lease_bg_task_for_static(&compute);
371 0 :
372 0 : Ok(WaitSpecResult {
373 0 : compute,
374 0 : http_port,
375 0 : resize_swap_on_bind,
376 0 : })
377 0 : }
378 :
379 : struct WaitSpecResult {
380 : compute: Arc<ComputeNode>,
381 : // passed through from ProcessCliResult
382 : http_port: u16,
383 : resize_swap_on_bind: bool,
384 : }
385 :
386 0 : fn start_postgres(
387 0 : // need to allow unused because `matches` is only used if target_os = "linux"
388 0 : #[allow(unused_variables)] matches: &clap::ArgMatches,
389 0 : WaitSpecResult {
390 0 : compute,
391 0 : http_port,
392 0 : resize_swap_on_bind,
393 0 : }: WaitSpecResult,
394 0 : ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
395 0 : // We got all we need, update the state.
396 0 : let mut state = compute.state.lock().unwrap();
397 0 : state.status = ComputeStatus::Init;
398 0 : compute.state_changed.notify_all();
399 0 :
400 0 : info!(
401 0 : "running compute with features: {:?}",
402 0 : state.pspec.as_ref().unwrap().spec.features
403 : );
404 : // before we release the mutex, fetch the swap size (if any) for later.
405 0 : let swap_size_bytes = state.pspec.as_ref().unwrap().spec.swap_size_bytes;
406 0 : drop(state);
407 0 :
408 0 : // Launch remaining service threads
409 0 : let _monitor_handle = launch_monitor(&compute);
410 0 : let _configurator_handle = launch_configurator(&compute);
411 0 :
412 0 : let mut prestartup_failed = false;
413 0 : let mut delay_exit = false;
414 :
415 : // Resize swap to the desired size if the compute spec says so
416 0 : if let (Some(size_bytes), true) = (swap_size_bytes, resize_swap_on_bind) {
417 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
418 : // *before* starting postgres.
419 : //
420 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
421 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
422 : // OOM-killed during startup because swap wasn't available yet.
423 0 : match resize_swap(size_bytes) {
424 : Ok(()) => {
425 0 : let size_gib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
426 0 : info!(%size_bytes, %size_gib, "resized swap");
427 : }
428 0 : Err(err) => {
429 0 : let err = err.context("failed to resize swap");
430 0 : error!("{err:#}");
431 :
432 : // Mark compute startup as failed; don't try to start postgres, and report this
433 : // error to the control plane when it next asks.
434 0 : prestartup_failed = true;
435 0 : let mut state = compute.state.lock().unwrap();
436 0 : state.error = Some(format!("{err:?}"));
437 0 : state.status = ComputeStatus::Failed;
438 0 : compute.state_changed.notify_all();
439 0 : delay_exit = true;
440 : }
441 : }
442 0 : }
443 :
444 0 : let extension_server_port: u16 = http_port;
445 0 :
446 0 : // Start Postgres
447 0 : let mut pg = None;
448 0 : if !prestartup_failed {
449 0 : pg = match compute.start_compute(extension_server_port) {
450 0 : Ok(pg) => Some(pg),
451 0 : Err(err) => {
452 0 : error!("could not start the compute node: {:#}", err);
453 0 : let mut state = compute.state.lock().unwrap();
454 0 : state.error = Some(format!("{:?}", err));
455 0 : state.status = ComputeStatus::Failed;
456 0 : // Notify others that Postgres failed to start. In case of configuring the
457 0 : // empty compute, it's likely that API handler is still waiting for compute
458 0 : // state change. With this we will notify it that compute is in Failed state,
459 0 : // so control plane will know about it earlier and record proper error instead
460 0 : // of timeout.
461 0 : compute.state_changed.notify_all();
462 0 : drop(state); // unlock
463 0 : delay_exit = true;
464 0 : None
465 : }
466 : };
467 : } else {
468 0 : warn!("skipping postgres startup because pre-startup step failed");
469 : }
470 :
471 : // Start the vm-monitor if directed to. The vm-monitor only runs on linux
472 : // because it requires cgroups.
473 : cfg_if::cfg_if! {
474 : if #[cfg(target_os = "linux")] {
475 : use std::env;
476 : use tokio_util::sync::CancellationToken;
477 0 : let vm_monitor_addr = matches
478 0 : .get_one::<String>("vm-monitor-addr")
479 0 : .expect("--vm-monitor-addr should always be set because it has a default arg");
480 0 : let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
481 0 : let cgroup = matches.get_one::<String>("cgroup");
482 :
483 : // Only make a runtime if we need to.
484 : // Note: it seems like you can make a runtime in an inner scope and
485 : // if you start a task in it it won't be dropped. However, make it
486 : // in the outermost scope just to be safe.
487 0 : let rt = if env::var_os("AUTOSCALING").is_some() {
488 0 : Some(
489 0 : tokio::runtime::Builder::new_multi_thread()
490 0 : .worker_threads(4)
491 0 : .enable_all()
492 0 : .build()
493 0 : .expect("failed to create tokio runtime for monitor")
494 0 : )
495 : } else {
496 0 : None
497 : };
498 :
499 : // This token is used internally by the monitor to clean up all threads
500 0 : let token = CancellationToken::new();
501 0 :
502 0 : let vm_monitor = rt.as_ref().map(|rt| {
503 0 : rt.spawn(vm_monitor::start(
504 0 : Box::leak(Box::new(vm_monitor::Args {
505 0 : cgroup: cgroup.cloned(),
506 0 : pgconnstr: file_cache_connstr.cloned(),
507 0 : addr: vm_monitor_addr.clone(),
508 0 : })),
509 0 : token.clone(),
510 0 : ))
511 0 : });
512 0 : }
513 0 : }
514 0 :
515 0 : Ok((
516 0 : pg,
517 0 : StartPostgresResult {
518 0 : delay_exit,
519 0 : compute,
520 0 : #[cfg(target_os = "linux")]
521 0 : rt,
522 0 : #[cfg(target_os = "linux")]
523 0 : token,
524 0 : #[cfg(target_os = "linux")]
525 0 : vm_monitor,
526 0 : },
527 0 : ))
528 0 : }
529 :
530 : type PostgresHandle = (std::process::Child, std::thread::JoinHandle<()>);
531 :
532 : struct StartPostgresResult {
533 : delay_exit: bool,
534 : // passed through from WaitSpecResult
535 : compute: Arc<ComputeNode>,
536 :
537 : #[cfg(target_os = "linux")]
538 : rt: Option<tokio::runtime::Runtime>,
539 : #[cfg(target_os = "linux")]
540 : token: tokio_util::sync::CancellationToken,
541 : #[cfg(target_os = "linux")]
542 : vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
543 : }
544 :
545 0 : fn wait_postgres(pg: Option<PostgresHandle>) -> Result<WaitPostgresResult> {
546 0 : // Wait for the child Postgres process forever. In this state Ctrl+C will
547 0 : // propagate to Postgres and it will be shut down as well.
548 0 : let mut exit_code = None;
549 0 : if let Some((mut pg, logs_handle)) = pg {
550 0 : let ecode = pg
551 0 : .wait()
552 0 : .expect("failed to start waiting on Postgres process");
553 0 : PG_PID.store(0, Ordering::SeqCst);
554 0 :
555 0 : // Process has exited, so we can join the logs thread.
556 0 : let _ = logs_handle
557 0 : .join()
558 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
559 0 :
560 0 : info!("Postgres exited with code {}, shutting down", ecode);
561 0 : exit_code = ecode.code()
562 0 : }
563 :
564 0 : Ok(WaitPostgresResult { exit_code })
565 0 : }
566 :
567 : struct WaitPostgresResult {
568 : exit_code: Option<i32>,
569 : }
570 :
571 0 : fn cleanup_after_postgres_exit(
572 0 : StartPostgresResult {
573 0 : mut delay_exit,
574 0 : compute,
575 0 : #[cfg(target_os = "linux")]
576 0 : vm_monitor,
577 0 : #[cfg(target_os = "linux")]
578 0 : token,
579 0 : #[cfg(target_os = "linux")]
580 0 : rt,
581 0 : }: StartPostgresResult,
582 0 : ) -> Result<bool> {
583 : // Terminate the vm_monitor so it releases the file watcher on
584 : // /sys/fs/cgroup/neon-postgres.
585 : // Note: the vm-monitor only runs on linux because it requires cgroups.
586 : cfg_if::cfg_if! {
587 : if #[cfg(target_os = "linux")] {
588 0 : if let Some(handle) = vm_monitor {
589 0 : // Kills all threads spawned by the monitor
590 0 : token.cancel();
591 0 : // Kills the actual task running the monitor
592 0 : handle.abort();
593 0 :
594 0 : // If handle is some, rt must have been used to produce it, and
595 0 : // hence is also some
596 0 : rt.unwrap().shutdown_timeout(Duration::from_secs(2));
597 0 : }
598 : }
599 : }
600 :
601 : // Maybe sync safekeepers again, to speed up next startup
602 0 : let compute_state = compute.state.lock().unwrap().clone();
603 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
604 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
605 0 : info!("syncing safekeepers on shutdown");
606 0 : let storage_auth_token = pspec.storage_auth_token.clone();
607 0 : let lsn = compute.sync_safekeepers(storage_auth_token)?;
608 0 : info!("synced safekeepers at lsn {lsn}");
609 0 : }
610 :
611 0 : let mut state = compute.state.lock().unwrap();
612 0 : if state.status == ComputeStatus::TerminationPending {
613 0 : state.status = ComputeStatus::Terminated;
614 0 : compute.state_changed.notify_all();
615 0 : // we were asked to terminate gracefully, don't exit to avoid restart
616 0 : delay_exit = true
617 0 : }
618 0 : drop(state);
619 :
620 0 : if let Err(err) = compute.check_for_core_dumps() {
621 0 : error!("error while checking for core dumps: {err:?}");
622 0 : }
623 :
624 0 : Ok(delay_exit)
625 0 : }
626 :
627 0 : fn maybe_delay_exit(delay_exit: bool) {
628 0 : // If launch failed, keep serving HTTP requests for a while, so the cloud
629 0 : // control plane can get the actual error.
630 0 : if delay_exit {
631 0 : info!("giving control plane 30s to collect the error before shutdown");
632 0 : thread::sleep(Duration::from_secs(30));
633 0 : }
634 0 : }
635 :
636 0 : fn deinit_and_exit(WaitPostgresResult { exit_code }: WaitPostgresResult) -> ! {
637 0 : // Shutdown trace pipeline gracefully, so that it has a chance to send any
638 0 : // pending traces before we exit. Shutting down OTEL tracing provider may
639 0 : // hang for quite some time, see, for example:
640 0 : // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
641 0 : // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
642 0 : //
643 0 : // Yet, we want computes to shut down fast enough, as we may need a new one
644 0 : // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
645 0 : // complete, then just error out and exit the main thread.
646 0 : info!("shutting down tracing");
647 0 : let (sender, receiver) = mpsc::channel();
648 0 : let _ = thread::spawn(move || {
649 0 : tracing_utils::shutdown_tracing();
650 0 : sender.send(()).ok()
651 0 : });
652 0 : let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
653 0 : if shutdown_res.is_err() {
654 0 : error!("timed out while shutting down tracing, exiting anyway");
655 0 : }
656 :
657 0 : info!("shutting down");
658 0 : exit(exit_code.unwrap_or(1))
659 : }
660 :
661 1 : fn cli() -> clap::Command {
662 1 : // Env variable is set by `cargo`
663 1 : let version = option_env!("CARGO_PKG_VERSION").unwrap_or("unknown");
664 1 : clap::Command::new("compute_ctl")
665 1 : .version(version)
666 1 : .arg(
667 1 : Arg::new("http-port")
668 1 : .long("http-port")
669 1 : .value_name("HTTP_PORT")
670 1 : .default_value("3080")
671 1 : .value_parser(clap::value_parser!(u16))
672 1 : .required(false),
673 1 : )
674 1 : .arg(
675 1 : Arg::new("connstr")
676 1 : .short('C')
677 1 : .long("connstr")
678 1 : .value_name("DATABASE_URL")
679 1 : .required(true),
680 1 : )
681 1 : .arg(
682 1 : Arg::new("pgdata")
683 1 : .short('D')
684 1 : .long("pgdata")
685 1 : .value_name("DATADIR")
686 1 : .required(true),
687 1 : )
688 1 : .arg(
689 1 : Arg::new("pgbin")
690 1 : .short('b')
691 1 : .long("pgbin")
692 1 : .default_value("postgres")
693 1 : .value_name("POSTGRES_PATH"),
694 1 : )
695 1 : .arg(
696 1 : Arg::new("spec")
697 1 : .short('s')
698 1 : .long("spec")
699 1 : .value_name("SPEC_JSON"),
700 1 : )
701 1 : .arg(
702 1 : Arg::new("spec-path")
703 1 : .short('S')
704 1 : .long("spec-path")
705 1 : .value_name("SPEC_PATH"),
706 1 : )
707 1 : .arg(
708 1 : Arg::new("compute-id")
709 1 : .short('i')
710 1 : .long("compute-id")
711 1 : .value_name("COMPUTE_ID"),
712 1 : )
713 1 : .arg(
714 1 : Arg::new("control-plane-uri")
715 1 : .short('p')
716 1 : .long("control-plane-uri")
717 1 : .value_name("CONTROL_PLANE_API_BASE_URI"),
718 1 : )
719 1 : .arg(
720 1 : Arg::new("remote-ext-config")
721 1 : .short('r')
722 1 : .long("remote-ext-config")
723 1 : .value_name("REMOTE_EXT_CONFIG"),
724 1 : )
725 1 : // TODO(fprasx): we currently have default arguments because the cloud PR
726 1 : // to pass them in hasn't been merged yet. We should get rid of them once
727 1 : // the PR is merged.
728 1 : .arg(
729 1 : Arg::new("vm-monitor-addr")
730 1 : .long("vm-monitor-addr")
731 1 : .default_value("0.0.0.0:10301")
732 1 : .value_name("VM_MONITOR_ADDR"),
733 1 : )
734 1 : .arg(
735 1 : Arg::new("cgroup")
736 1 : .long("cgroup")
737 1 : .default_value("neon-postgres")
738 1 : .value_name("CGROUP"),
739 1 : )
740 1 : .arg(
741 1 : Arg::new("filecache-connstr")
742 1 : .long("filecache-connstr")
743 1 : .default_value(
744 1 : "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable application_name=vm-monitor",
745 1 : )
746 1 : .value_name("FILECACHE_CONNSTR"),
747 1 : )
748 1 : .arg(
749 1 : Arg::new("resize-swap-on-bind")
750 1 : .long("resize-swap-on-bind")
751 1 : .action(clap::ArgAction::SetTrue),
752 1 : )
753 1 : }
754 :
755 : /// When compute_ctl is killed, send also termination signal to sync-safekeepers
756 : /// to prevent leakage. TODO: it is better to convert compute_ctl to async and
757 : /// wait for termination which would be easy then.
758 0 : fn handle_exit_signal(sig: i32) {
759 0 : info!("received {sig} termination signal");
760 0 : forward_termination_signal();
761 0 : exit(1);
762 : }
763 :
764 : #[test]
765 1 : fn verify_cli() {
766 1 : cli().debug_assert()
767 1 : }
|