Line data Source code
1 : use std::collections::HashMap;
2 : use std::os::unix::fs::{PermissionsExt, symlink};
3 : use std::path::Path;
4 : use std::process::{Command, Stdio};
5 : use std::str::FromStr;
6 : use std::sync::atomic::{AtomicU32, Ordering};
7 : use std::sync::{Arc, Condvar, Mutex, RwLock};
8 : use std::time::{Duration, Instant};
9 : use std::{env, fs};
10 :
11 : use anyhow::{Context, Result};
12 : use chrono::{DateTime, Utc};
13 : use compute_api::privilege::Privilege;
14 : use compute_api::responses::{ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus};
15 : use compute_api::spec::{
16 : ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PgIdent,
17 : };
18 : use futures::StreamExt;
19 : use futures::future::join_all;
20 : use futures::stream::FuturesUnordered;
21 : use nix::sys::signal::{Signal, kill};
22 : use nix::unistd::Pid;
23 : use once_cell::sync::Lazy;
24 : use postgres;
25 : use postgres::NoTls;
26 : use postgres::error::SqlState;
27 : use remote_storage::{DownloadError, RemotePath};
28 : use tokio::spawn;
29 : use tracing::{Instrument, debug, error, info, instrument, warn};
30 : use utils::id::{TenantId, TimelineId};
31 : use utils::lsn::Lsn;
32 : use utils::measured_stream::MeasuredReader;
33 :
34 : use crate::configurator::launch_configurator;
35 : use crate::disk_quota::set_disk_quota;
36 : use crate::installed_extensions::get_installed_extensions;
37 : use crate::logger::startup_context_from_env;
38 : use crate::lsn_lease::launch_lsn_lease_bg_task_for_static;
39 : use crate::metrics::COMPUTE_CTL_UP;
40 : use crate::monitor::launch_monitor;
41 : use crate::pg_helpers::*;
42 : use crate::rsyslog::{
43 : PostgresLogsRsyslogConfig, configure_audit_rsyslog, configure_postgres_logs_export,
44 : launch_pgaudit_gc,
45 : };
46 : use crate::spec::*;
47 : use crate::swap::resize_swap;
48 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
49 : use crate::tls::watch_cert_for_changes;
50 : use crate::{config, extension_server, local_proxy};
51 :
52 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
53 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
54 : // This is an arbitrary build tag. Fine as a default / for testing purposes
55 : // in-case of not-set environment var
56 : const BUILD_TAG_DEFAULT: &str = "latest";
57 : /// Build tag/version of the compute node binaries/image. It's tricky and ugly
58 : /// to pass it everywhere as a part of `ComputeNodeParams`, so we use a
59 : /// global static variable.
60 0 : pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
61 0 : option_env!("BUILD_TAG")
62 0 : .unwrap_or(BUILD_TAG_DEFAULT)
63 0 : .to_string()
64 0 : });
65 :
66 : /// Static configuration params that don't change after startup. These mostly
67 : /// come from the CLI args, or are derived from them.
68 : pub struct ComputeNodeParams {
69 : /// The ID of the compute
70 : pub compute_id: String,
71 : // Url type maintains proper escaping
72 : pub connstr: url::Url,
73 :
74 : pub resize_swap_on_bind: bool,
75 : pub set_disk_quota_for_fs: Option<String>,
76 :
77 : // VM monitor parameters
78 : #[cfg(target_os = "linux")]
79 : pub filecache_connstr: String,
80 : #[cfg(target_os = "linux")]
81 : pub cgroup: String,
82 : #[cfg(target_os = "linux")]
83 : pub vm_monitor_addr: String,
84 :
85 : pub pgdata: String,
86 : pub pgbin: String,
87 : pub pgversion: String,
88 :
89 : /// The port that the compute's external HTTP server listens on
90 : pub external_http_port: u16,
91 : /// The port that the compute's internal HTTP server listens on
92 : pub internal_http_port: u16,
93 :
94 : /// the address of extension storage proxy gateway
95 : pub ext_remote_storage: Option<String>,
96 : }
97 :
98 : /// Compute node info shared across several `compute_ctl` threads.
99 : pub struct ComputeNode {
100 : pub params: ComputeNodeParams,
101 :
102 : // We connect to Postgres from many different places, so build configs once
103 : // and reuse them where needed. These are derived from 'params.connstr'
104 : pub conn_conf: postgres::config::Config,
105 : pub tokio_conn_conf: tokio_postgres::config::Config,
106 :
107 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
108 : /// To allow HTTP API server to serving status requests, while configuration
109 : /// is in progress, lock should be held only for short periods of time to do
110 : /// read/write, not the whole configuration process.
111 : pub state: Mutex<ComputeState>,
112 : /// `Condvar` to allow notifying waiters about state changes.
113 : pub state_changed: Condvar,
114 :
115 : // key: ext_archive_name, value: started download time, download_completed?
116 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
117 : pub compute_ctl_config: ComputeCtlConfig,
118 : }
119 :
120 : // store some metrics about download size that might impact startup time
121 : #[derive(Clone, Debug)]
122 : pub struct RemoteExtensionMetrics {
123 : num_ext_downloaded: u64,
124 : largest_ext_size: u64,
125 : total_ext_download_size: u64,
126 : }
127 :
128 : #[derive(Clone, Debug)]
129 : pub struct ComputeState {
130 : pub start_time: DateTime<Utc>,
131 : pub status: ComputeStatus,
132 : /// Timestamp of the last Postgres activity. It could be `None` if
133 : /// compute wasn't used since start.
134 : pub last_active: Option<DateTime<Utc>>,
135 : pub error: Option<String>,
136 :
137 : /// Compute spec. This can be received from the CLI or - more likely -
138 : /// passed by the control plane with a /configure HTTP request.
139 : pub pspec: Option<ParsedSpec>,
140 :
141 : /// If the spec is passed by a /configure request, 'startup_span' is the
142 : /// /configure request's tracing span. The main thread enters it when it
143 : /// processes the compute startup, so that the compute startup is considered
144 : /// to be part of the /configure request for tracing purposes.
145 : ///
146 : /// If the request handling thread/task called startup_compute() directly,
147 : /// it would automatically be a child of the request handling span, and we
148 : /// wouldn't need this. But because we use the main thread to perform the
149 : /// startup, and the /configure task just waits for it to finish, we need to
150 : /// set up the span relationship ourselves.
151 : pub startup_span: Option<tracing::span::Span>,
152 :
153 : pub metrics: ComputeMetrics,
154 : }
155 :
156 : impl ComputeState {
157 0 : pub fn new() -> Self {
158 0 : Self {
159 0 : start_time: Utc::now(),
160 0 : status: ComputeStatus::Empty,
161 0 : last_active: None,
162 0 : error: None,
163 0 : pspec: None,
164 0 : startup_span: None,
165 0 : metrics: ComputeMetrics::default(),
166 0 : }
167 0 : }
168 :
169 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
170 0 : let prev = self.status;
171 0 : info!("Changing compute status from {} to {}", prev, status);
172 0 : self.status = status;
173 0 : state_changed.notify_all();
174 0 :
175 0 : COMPUTE_CTL_UP.reset();
176 0 : COMPUTE_CTL_UP
177 0 : .with_label_values(&[&BUILD_TAG, status.to_string().as_str()])
178 0 : .set(1);
179 0 : }
180 :
181 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
182 0 : self.error = Some(format!("{err:?}"));
183 0 : self.set_status(ComputeStatus::Failed, state_changed);
184 0 : }
185 : }
186 :
187 : impl Default for ComputeState {
188 0 : fn default() -> Self {
189 0 : Self::new()
190 0 : }
191 : }
192 :
193 : #[derive(Clone, Debug)]
194 : pub struct ParsedSpec {
195 : pub spec: ComputeSpec,
196 : pub tenant_id: TenantId,
197 : pub timeline_id: TimelineId,
198 : pub pageserver_connstr: String,
199 : pub safekeeper_connstrings: Vec<String>,
200 : pub storage_auth_token: Option<String>,
201 : }
202 :
203 : impl TryFrom<ComputeSpec> for ParsedSpec {
204 : type Error = String;
205 0 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
206 : // Extract the options from the spec file that are needed to connect to
207 : // the storage system.
208 : //
209 : // For backwards-compatibility, the top-level fields in the spec file
210 : // may be empty. In that case, we need to dig them from the GUCs in the
211 : // cluster.settings field.
212 0 : let pageserver_connstr = spec
213 0 : .pageserver_connstring
214 0 : .clone()
215 0 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
216 0 : .ok_or("pageserver connstr should be provided")?;
217 0 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
218 0 : if matches!(spec.mode, ComputeMode::Primary) {
219 0 : spec.cluster
220 0 : .settings
221 0 : .find("neon.safekeepers")
222 0 : .ok_or("safekeeper connstrings should be provided")?
223 0 : .split(',')
224 0 : .map(|str| str.to_string())
225 0 : .collect()
226 : } else {
227 0 : vec![]
228 : }
229 : } else {
230 0 : spec.safekeeper_connstrings.clone()
231 : };
232 0 : let storage_auth_token = spec.storage_auth_token.clone();
233 0 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
234 0 : tenant_id
235 : } else {
236 0 : spec.cluster
237 0 : .settings
238 0 : .find("neon.tenant_id")
239 0 : .ok_or("tenant id should be provided")
240 0 : .map(|s| TenantId::from_str(&s))?
241 0 : .or(Err("invalid tenant id"))?
242 : };
243 0 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
244 0 : timeline_id
245 : } else {
246 0 : spec.cluster
247 0 : .settings
248 0 : .find("neon.timeline_id")
249 0 : .ok_or("timeline id should be provided")
250 0 : .map(|s| TimelineId::from_str(&s))?
251 0 : .or(Err("invalid timeline id"))?
252 : };
253 :
254 0 : Ok(ParsedSpec {
255 0 : spec,
256 0 : pageserver_connstr,
257 0 : safekeeper_connstrings,
258 0 : storage_auth_token,
259 0 : tenant_id,
260 0 : timeline_id,
261 0 : })
262 0 : }
263 : }
264 :
265 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
266 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
267 : ///
268 : /// This function should be used to start postgres, as it will start it in the
269 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
270 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
271 : /// creates it during the sysinit phase of its inittab.
272 0 : fn maybe_cgexec(cmd: &str) -> Command {
273 0 : // The cplane sets this env var for autoscaling computes.
274 0 : // use `var_os` so we don't have to worry about the variable being valid
275 0 : // unicode. Should never be an concern . . . but just in case
276 0 : if env::var_os("AUTOSCALING").is_some() {
277 0 : let mut command = Command::new("cgexec");
278 0 : command.args(["-g", "memory:neon-postgres"]);
279 0 : command.arg(cmd);
280 0 : command
281 : } else {
282 0 : Command::new(cmd)
283 : }
284 0 : }
285 :
286 : struct PostgresHandle {
287 : postgres: std::process::Child,
288 : log_collector: tokio::task::JoinHandle<Result<()>>,
289 : }
290 :
291 : impl PostgresHandle {
292 : /// Return PID of the postgres (postmaster) process
293 0 : fn pid(&self) -> Pid {
294 0 : Pid::from_raw(self.postgres.id() as i32)
295 0 : }
296 : }
297 :
298 : struct StartVmMonitorResult {
299 : #[cfg(target_os = "linux")]
300 : token: tokio_util::sync::CancellationToken,
301 : #[cfg(target_os = "linux")]
302 : vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
303 : }
304 :
305 : impl ComputeNode {
306 0 : pub fn new(params: ComputeNodeParams, config: ComputeConfig) -> Result<Self> {
307 0 : let connstr = params.connstr.as_str();
308 0 : let conn_conf = postgres::config::Config::from_str(connstr)
309 0 : .context("cannot build postgres config from connstr")?;
310 0 : let tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr)
311 0 : .context("cannot build tokio postgres config from connstr")?;
312 :
313 0 : let mut new_state = ComputeState::new();
314 0 : if let Some(spec) = config.spec {
315 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
316 0 : new_state.pspec = Some(pspec);
317 0 : }
318 :
319 0 : Ok(ComputeNode {
320 0 : params,
321 0 : conn_conf,
322 0 : tokio_conn_conf,
323 0 : state: Mutex::new(new_state),
324 0 : state_changed: Condvar::new(),
325 0 : ext_download_progress: RwLock::new(HashMap::new()),
326 0 : compute_ctl_config: config.compute_ctl_config,
327 0 : })
328 0 : }
329 :
330 : /// Top-level control flow of compute_ctl. Returns a process exit code we should
331 : /// exit with.
332 0 : pub fn run(self) -> Result<Option<i32>> {
333 0 : let this = Arc::new(self);
334 0 :
335 0 : let cli_spec = this.state.lock().unwrap().pspec.clone();
336 0 :
337 0 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
338 0 : // available for binding. Prewarming helps Postgres start quicker later,
339 0 : // because QEMU will already have its memory allocated from the host, and
340 0 : // the necessary binaries will already be cached.
341 0 : if cli_spec.is_none() {
342 0 : this.prewarm_postgres()?;
343 0 : }
344 :
345 : // Set the up metric with Empty status before starting the HTTP server.
346 : // That way on the first metric scrape, an external observer will see us
347 : // as 'up' and 'empty' (unless the compute was started with a spec or
348 : // already configured by control plane).
349 0 : COMPUTE_CTL_UP
350 0 : .with_label_values(&[&BUILD_TAG, ComputeStatus::Empty.to_string().as_str()])
351 0 : .set(1);
352 0 :
353 0 : // Launch the external HTTP server first, so that we can serve control plane
354 0 : // requests while configuration is still in progress.
355 0 : crate::http::server::Server::External {
356 0 : port: this.params.external_http_port,
357 0 : config: this.compute_ctl_config.clone(),
358 0 : compute_id: this.params.compute_id.clone(),
359 0 : }
360 0 : .launch(&this);
361 0 :
362 0 : // The internal HTTP server could be launched later, but there isn't much
363 0 : // sense in waiting.
364 0 : crate::http::server::Server::Internal {
365 0 : port: this.params.internal_http_port,
366 0 : }
367 0 : .launch(&this);
368 :
369 : // If we got a spec from the CLI already, use that. Otherwise wait for the
370 : // control plane to pass it to us with a /configure HTTP request
371 0 : let pspec = if let Some(cli_spec) = cli_spec {
372 0 : cli_spec
373 : } else {
374 0 : this.wait_spec()?
375 : };
376 :
377 0 : launch_lsn_lease_bg_task_for_static(&this);
378 0 :
379 0 : // We have a spec, start the compute
380 0 : let mut delay_exit = false;
381 0 : let mut vm_monitor = None;
382 0 : let mut pg_process: Option<PostgresHandle> = None;
383 0 :
384 0 : match this.start_compute(&mut pg_process) {
385 0 : Ok(()) => {
386 0 : // Success! Launch remaining services (just vm-monitor currently)
387 0 : vm_monitor =
388 0 : Some(this.start_vm_monitor(pspec.spec.disable_lfc_resizing.unwrap_or(false)));
389 0 : }
390 0 : Err(err) => {
391 0 : // Something went wrong with the startup. Log it and expose the error to
392 0 : // HTTP status requests.
393 0 : error!("could not start the compute node: {:#}", err);
394 0 : this.set_failed_status(err);
395 0 : delay_exit = true;
396 :
397 : // If the error happened after starting PostgreSQL, kill it
398 0 : if let Some(ref pg_process) = pg_process {
399 0 : kill(pg_process.pid(), Signal::SIGQUIT).ok();
400 0 : }
401 : }
402 : }
403 :
404 : // If startup was successful, or it failed in the late stages,
405 : // PostgreSQL is now running. Wait until it exits.
406 0 : let exit_code = if let Some(pg_handle) = pg_process {
407 0 : let exit_status = this.wait_postgres(pg_handle);
408 0 : info!("Postgres exited with code {}, shutting down", exit_status);
409 0 : exit_status.code()
410 : } else {
411 0 : None
412 : };
413 :
414 : // Terminate the vm_monitor so it releases the file watcher on
415 : // /sys/fs/cgroup/neon-postgres.
416 : // Note: the vm-monitor only runs on linux because it requires cgroups.
417 0 : if let Some(vm_monitor) = vm_monitor {
418 : cfg_if::cfg_if! {
419 : if #[cfg(target_os = "linux")] {
420 : // Kills all threads spawned by the monitor
421 0 : vm_monitor.token.cancel();
422 0 : if let Some(handle) = vm_monitor.vm_monitor {
423 0 : // Kills the actual task running the monitor
424 0 : handle.abort();
425 0 : }
426 : } else {
427 : _ = vm_monitor; // appease unused lint on macOS
428 : }
429 : }
430 0 : }
431 :
432 : // Reap the postgres process
433 0 : delay_exit |= this.cleanup_after_postgres_exit()?;
434 :
435 : // If launch failed, keep serving HTTP requests for a while, so the cloud
436 : // control plane can get the actual error.
437 0 : if delay_exit {
438 0 : info!("giving control plane 30s to collect the error before shutdown");
439 0 : std::thread::sleep(Duration::from_secs(30));
440 0 : }
441 0 : Ok(exit_code)
442 0 : }
443 :
444 0 : pub fn wait_spec(&self) -> Result<ParsedSpec> {
445 0 : info!("no compute spec provided, waiting");
446 0 : let mut state = self.state.lock().unwrap();
447 0 : while state.status != ComputeStatus::ConfigurationPending {
448 0 : state = self.state_changed.wait(state).unwrap();
449 0 : }
450 :
451 0 : info!("got spec, continue configuration");
452 0 : let spec = state.pspec.as_ref().unwrap().clone();
453 0 :
454 0 : // Record for how long we slept waiting for the spec.
455 0 : let now = Utc::now();
456 0 : state.metrics.wait_for_spec_ms = now
457 0 : .signed_duration_since(state.start_time)
458 0 : .to_std()
459 0 : .unwrap()
460 0 : .as_millis() as u64;
461 0 :
462 0 : // Reset start time, so that the total startup time that is calculated later will
463 0 : // not include the time that we waited for the spec.
464 0 : state.start_time = now;
465 0 :
466 0 : Ok(spec)
467 0 : }
468 :
469 : /// Start compute.
470 : ///
471 : /// Prerequisites:
472 : /// - the compute spec has been placed in self.state.pspec
473 : ///
474 : /// On success:
475 : /// - status is set to ComputeStatus::Running
476 : /// - self.running_postgres is set
477 : ///
478 : /// On error:
479 : /// - status is left in ComputeStatus::Init. The caller is responsible for setting it to Failed
480 : /// - if Postgres was started before the fatal error happened, self.running_postgres is
481 : /// set. The caller is responsible for killing it.
482 : ///
483 : /// Note that this is in the critical path of a compute cold start. Keep this fast.
484 : /// Try to do things concurrently, to hide the latencies.
485 0 : fn start_compute(self: &Arc<Self>, pg_handle: &mut Option<PostgresHandle>) -> Result<()> {
486 0 : let compute_state: ComputeState;
487 0 :
488 0 : let start_compute_span;
489 0 : let _this_entered;
490 0 : {
491 0 : let mut state_guard = self.state.lock().unwrap();
492 :
493 : // Create a tracing span for the startup operation.
494 : //
495 : // We could otherwise just annotate the function with #[instrument], but if
496 : // we're being configured from a /configure HTTP request, we want the
497 : // startup to be considered part of the /configure request.
498 : //
499 : // Similarly, if a trace ID was passed in env variables, attach it to the span.
500 0 : start_compute_span = {
501 : // Temporarily enter the parent span, so that the new span becomes its child.
502 0 : if let Some(p) = state_guard.startup_span.take() {
503 0 : let _parent_entered = p.entered();
504 0 : tracing::info_span!("start_compute")
505 0 : } else if let Some(otel_context) = startup_context_from_env() {
506 : use tracing_opentelemetry::OpenTelemetrySpanExt;
507 0 : let span = tracing::info_span!("start_compute");
508 0 : span.set_parent(otel_context);
509 0 : span
510 : } else {
511 0 : tracing::info_span!("start_compute")
512 : }
513 : };
514 0 : _this_entered = start_compute_span.enter();
515 0 :
516 0 : state_guard.set_status(ComputeStatus::Init, &self.state_changed);
517 0 : compute_state = state_guard.clone()
518 0 : }
519 0 :
520 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
521 0 : info!(
522 0 : "starting compute for project {}, operation {}, tenant {}, timeline {}, project {}, branch {}, endpoint {}, features {:?}, spec.remote_extensions {:?}",
523 0 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
524 0 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
525 0 : pspec.tenant_id,
526 0 : pspec.timeline_id,
527 0 : pspec.spec.project_id.as_deref().unwrap_or("None"),
528 0 : pspec.spec.branch_id.as_deref().unwrap_or("None"),
529 0 : pspec.spec.endpoint_id.as_deref().unwrap_or("None"),
530 : pspec.spec.features,
531 : pspec.spec.remote_extensions,
532 : );
533 :
534 : ////// PRE-STARTUP PHASE: things that need to be finished before we start the Postgres process
535 :
536 : // Collect all the tasks that must finish here
537 0 : let mut pre_tasks = tokio::task::JoinSet::new();
538 0 :
539 0 : // Make sure TLS certificates are properly loaded and in the right place.
540 0 : if self.compute_ctl_config.tls.is_some() {
541 0 : let this = self.clone();
542 0 : pre_tasks.spawn(async move {
543 0 : this.watch_cert_for_changes().await;
544 :
545 0 : Ok::<(), anyhow::Error>(())
546 0 : });
547 0 : }
548 :
549 : // If there are any remote extensions in shared_preload_libraries, start downloading them
550 0 : if pspec.spec.remote_extensions.is_some() {
551 0 : let (this, spec) = (self.clone(), pspec.spec.clone());
552 0 : pre_tasks.spawn(async move {
553 0 : this.download_preload_extensions(&spec)
554 0 : .in_current_span()
555 0 : .await
556 0 : });
557 0 : }
558 :
559 : // Prepare pgdata directory. This downloads the basebackup, among other things.
560 0 : {
561 0 : let (this, cs) = (self.clone(), compute_state.clone());
562 0 : pre_tasks.spawn_blocking_child(move || this.prepare_pgdata(&cs));
563 0 : }
564 :
565 : // Resize swap to the desired size if the compute spec says so
566 0 : if let (Some(size_bytes), true) =
567 0 : (pspec.spec.swap_size_bytes, self.params.resize_swap_on_bind)
568 0 : {
569 0 : pre_tasks.spawn_blocking_child(move || {
570 0 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
571 0 : // *before* starting postgres.
572 0 : //
573 0 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
574 0 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
575 0 : // OOM-killed during startup because swap wasn't available yet.
576 0 : resize_swap(size_bytes).context("failed to resize swap")?;
577 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
578 0 : info!(%size_bytes, %size_mib, "resized swap");
579 :
580 0 : Ok::<(), anyhow::Error>(())
581 0 : });
582 0 : }
583 :
584 : // Set disk quota if the compute spec says so
585 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) = (
586 0 : pspec.spec.disk_quota_bytes,
587 0 : self.params.set_disk_quota_for_fs.as_ref(),
588 0 : ) {
589 0 : let disk_quota_fs_mountpoint = disk_quota_fs_mountpoint.clone();
590 0 : pre_tasks.spawn_blocking_child(move || {
591 0 : set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint)
592 0 : .context("failed to set disk quota")?;
593 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
594 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
595 :
596 0 : Ok::<(), anyhow::Error>(())
597 0 : });
598 0 : }
599 :
600 : // tune pgbouncer
601 0 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
602 0 : info!("tuning pgbouncer");
603 :
604 0 : let pgbouncer_settings = pgbouncer_settings.clone();
605 0 : let tls_config = self.compute_ctl_config.tls.clone();
606 0 :
607 0 : // Spawn a background task to do the tuning,
608 0 : // so that we don't block the main thread that starts Postgres.
609 0 : let _handle = tokio::spawn(async move {
610 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
611 0 : if let Err(err) = res {
612 0 : error!("error while tuning pgbouncer: {err:?}");
613 : // Continue with the startup anyway
614 0 : }
615 0 : });
616 0 : }
617 :
618 : // configure local_proxy
619 0 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
620 0 : info!("configuring local_proxy");
621 :
622 : // Spawn a background task to do the configuration,
623 : // so that we don't block the main thread that starts Postgres.
624 0 : let local_proxy = local_proxy.clone();
625 0 : let _handle = tokio::spawn(async move {
626 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
627 0 : error!("error while configuring local_proxy: {err:?}");
628 : // Continue with the startup anyway
629 0 : }
630 0 : });
631 0 : }
632 :
633 : // Configure and start rsyslog for compliance audit logging
634 0 : match pspec.spec.audit_log_level {
635 : ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
636 0 : let remote_endpoint =
637 0 : std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
638 0 : if remote_endpoint.is_empty() {
639 0 : anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
640 0 : }
641 0 :
642 0 : let log_directory_path = Path::new(&self.params.pgdata).join("log");
643 0 : let log_directory_path = log_directory_path.to_string_lossy().to_string();
644 :
645 : // Add project_id,endpoint_id tag to identify the logs.
646 : //
647 : // These ids are passed from cplane,
648 : // for backwards compatibility (old computes that don't have them),
649 : // we set them to None.
650 : // TODO: Clean up this code when all computes have them.
651 0 : let tag: Option<String> = match (
652 0 : pspec.spec.project_id.as_deref(),
653 0 : pspec.spec.endpoint_id.as_deref(),
654 : ) {
655 0 : (Some(project_id), Some(endpoint_id)) => {
656 0 : Some(format!("{project_id}/{endpoint_id}"))
657 : }
658 0 : (Some(project_id), None) => Some(format!("{project_id}/None")),
659 0 : (None, Some(endpoint_id)) => Some(format!("None,{endpoint_id}")),
660 0 : (None, None) => None,
661 : };
662 :
663 0 : configure_audit_rsyslog(log_directory_path.clone(), tag, &remote_endpoint)?;
664 :
665 : // Launch a background task to clean up the audit logs
666 0 : launch_pgaudit_gc(log_directory_path);
667 : }
668 0 : _ => {}
669 : }
670 :
671 : // Configure and start rsyslog for Postgres logs export
672 0 : let conf = PostgresLogsRsyslogConfig::new(pspec.spec.logs_export_host.as_deref());
673 0 : configure_postgres_logs_export(conf)?;
674 :
675 : // Launch remaining service threads
676 0 : let _monitor_handle = launch_monitor(self);
677 0 : let _configurator_handle = launch_configurator(self);
678 0 :
679 0 : // Wait for all the pre-tasks to finish before starting postgres
680 0 : let rt = tokio::runtime::Handle::current();
681 0 : while let Some(res) = rt.block_on(pre_tasks.join_next()) {
682 0 : res??;
683 : }
684 :
685 : ////// START POSTGRES
686 0 : let start_time = Utc::now();
687 0 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
688 0 : let postmaster_pid = pg_process.pid();
689 0 : *pg_handle = Some(pg_process);
690 0 :
691 0 : // If this is a primary endpoint, perform some post-startup configuration before
692 0 : // opening it up for the world.
693 0 : let config_time = Utc::now();
694 0 : if pspec.spec.mode == ComputeMode::Primary {
695 0 : self.configure_as_primary(&compute_state)?;
696 :
697 0 : let conf = self.get_tokio_conn_conf(None);
698 0 : tokio::task::spawn(async {
699 0 : let res = get_installed_extensions(conf).await;
700 0 : match res {
701 0 : Ok(extensions) => {
702 0 : info!(
703 0 : "[NEON_EXT_STAT] {}",
704 0 : serde_json::to_string(&extensions)
705 0 : .expect("failed to serialize extensions list")
706 : );
707 : }
708 0 : Err(err) => error!("could not get installed extensions: {err:?}"),
709 : }
710 0 : });
711 0 : }
712 :
713 : // All done!
714 0 : let startup_end_time = Utc::now();
715 0 : let metrics = {
716 0 : let mut state = self.state.lock().unwrap();
717 0 : state.metrics.start_postgres_ms = config_time
718 0 : .signed_duration_since(start_time)
719 0 : .to_std()
720 0 : .unwrap()
721 0 : .as_millis() as u64;
722 0 : state.metrics.config_ms = startup_end_time
723 0 : .signed_duration_since(config_time)
724 0 : .to_std()
725 0 : .unwrap()
726 0 : .as_millis() as u64;
727 0 : state.metrics.total_startup_ms = startup_end_time
728 0 : .signed_duration_since(compute_state.start_time)
729 0 : .to_std()
730 0 : .unwrap()
731 0 : .as_millis() as u64;
732 0 : state.metrics.clone()
733 0 : };
734 0 : self.set_status(ComputeStatus::Running);
735 0 :
736 0 : // Log metrics so that we can search for slow operations in logs
737 0 : info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
738 :
739 0 : Ok(())
740 0 : }
741 :
742 : #[instrument(skip_all)]
743 : async fn download_preload_extensions(&self, spec: &ComputeSpec) -> Result<()> {
744 : let remote_extensions = if let Some(remote_extensions) = &spec.remote_extensions {
745 : remote_extensions
746 : } else {
747 : return Ok(());
748 : };
749 :
750 : // First, create control files for all available extensions
751 : extension_server::create_control_files(remote_extensions, &self.params.pgbin);
752 :
753 : let library_load_start_time = Utc::now();
754 : let remote_ext_metrics = self.prepare_preload_libraries(spec).await?;
755 :
756 : let library_load_time = Utc::now()
757 : .signed_duration_since(library_load_start_time)
758 : .to_std()
759 : .unwrap()
760 : .as_millis() as u64;
761 : let mut state = self.state.lock().unwrap();
762 : state.metrics.load_ext_ms = library_load_time;
763 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
764 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
765 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
766 : info!(
767 : "Loading shared_preload_libraries took {:?}ms",
768 : library_load_time
769 : );
770 : info!("{:?}", remote_ext_metrics);
771 :
772 : Ok(())
773 : }
774 :
775 : /// Start the vm-monitor if directed to. The vm-monitor only runs on linux
776 : /// because it requires cgroups.
777 0 : fn start_vm_monitor(&self, disable_lfc_resizing: bool) -> StartVmMonitorResult {
778 : cfg_if::cfg_if! {
779 : if #[cfg(target_os = "linux")] {
780 : use std::env;
781 : use tokio_util::sync::CancellationToken;
782 :
783 : // This token is used internally by the monitor to clean up all threads
784 0 : let token = CancellationToken::new();
785 :
786 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
787 0 : let pgconnstr = if disable_lfc_resizing {
788 0 : None
789 : } else {
790 0 : Some(self.params.filecache_connstr.clone())
791 : };
792 :
793 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
794 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
795 0 : Box::leak(Box::new(vm_monitor::Args {
796 0 : cgroup: Some(self.params.cgroup.clone()),
797 0 : pgconnstr,
798 0 : addr: self.params.vm_monitor_addr.clone(),
799 0 : })),
800 0 : token.clone(),
801 0 : ));
802 0 : Some(vm_monitor)
803 : } else {
804 0 : None
805 : };
806 0 : StartVmMonitorResult { token, vm_monitor }
807 0 : } else {
808 0 : _ = disable_lfc_resizing; // appease unused lint on macOS
809 0 : StartVmMonitorResult { }
810 0 : }
811 0 : }
812 0 : }
813 :
814 0 : fn cleanup_after_postgres_exit(&self) -> Result<bool> {
815 0 : // Maybe sync safekeepers again, to speed up next startup
816 0 : let compute_state = self.state.lock().unwrap().clone();
817 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
818 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
819 0 : info!("syncing safekeepers on shutdown");
820 0 : let storage_auth_token = pspec.storage_auth_token.clone();
821 0 : let lsn = self.sync_safekeepers(storage_auth_token)?;
822 0 : info!("synced safekeepers at lsn {lsn}");
823 0 : }
824 :
825 0 : let mut delay_exit = false;
826 0 : let mut state = self.state.lock().unwrap();
827 0 : if state.status == ComputeStatus::TerminationPending {
828 0 : state.status = ComputeStatus::Terminated;
829 0 : self.state_changed.notify_all();
830 0 : // we were asked to terminate gracefully, don't exit to avoid restart
831 0 : delay_exit = true
832 0 : }
833 0 : drop(state);
834 :
835 0 : if let Err(err) = self.check_for_core_dumps() {
836 0 : error!("error while checking for core dumps: {err:?}");
837 0 : }
838 :
839 0 : Ok(delay_exit)
840 0 : }
841 :
842 : /// Check that compute node has corresponding feature enabled.
843 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
844 0 : let state = self.state.lock().unwrap();
845 :
846 0 : if let Some(s) = state.pspec.as_ref() {
847 0 : s.spec.features.contains(&feature)
848 : } else {
849 0 : false
850 : }
851 0 : }
852 :
853 0 : pub fn set_status(&self, status: ComputeStatus) {
854 0 : let mut state = self.state.lock().unwrap();
855 0 : state.set_status(status, &self.state_changed);
856 0 : }
857 :
858 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
859 0 : let mut state = self.state.lock().unwrap();
860 0 : state.set_failed_status(err, &self.state_changed);
861 0 : }
862 :
863 0 : pub fn get_status(&self) -> ComputeStatus {
864 0 : self.state.lock().unwrap().status
865 0 : }
866 :
867 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
868 0 : self.state
869 0 : .lock()
870 0 : .unwrap()
871 0 : .pspec
872 0 : .as_ref()
873 0 : .map(|s| s.timeline_id)
874 0 : }
875 :
876 : // Remove `pgdata` directory and create it again with right permissions.
877 0 : fn create_pgdata(&self) -> Result<()> {
878 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
879 0 : // If it is something different then create_dir() will error out anyway.
880 0 : let pgdata = &self.params.pgdata;
881 0 : let _ok = fs::remove_dir_all(pgdata);
882 0 : fs::create_dir(pgdata)?;
883 0 : fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
884 :
885 0 : Ok(())
886 0 : }
887 :
888 : // Get basebackup from the libpq connection to pageserver using `connstr` and
889 : // unarchive it to `pgdata` directory overriding all its previous content.
890 : #[instrument(skip_all, fields(%lsn))]
891 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
892 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
893 : let start_time = Instant::now();
894 :
895 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
896 : let mut config = postgres::Config::from_str(shard0_connstr)?;
897 :
898 : // Use the storage auth token from the config file, if given.
899 : // Note: this overrides any password set in the connection string.
900 : if let Some(storage_auth_token) = &spec.storage_auth_token {
901 : info!("Got storage auth token from spec file");
902 : config.password(storage_auth_token);
903 : } else {
904 : info!("Storage auth token not set");
905 : }
906 :
907 : config.application_name("compute_ctl");
908 : if let Some(spec) = &compute_state.pspec {
909 : config.options(&format!(
910 : "-c neon.compute_mode={}",
911 : spec.spec.mode.to_type_str()
912 : ));
913 : }
914 :
915 : // Connect to pageserver
916 : let mut client = config.connect(NoTls)?;
917 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
918 :
919 : let basebackup_cmd = match lsn {
920 : Lsn(0) => {
921 : if spec.spec.mode != ComputeMode::Primary {
922 : format!(
923 : "basebackup {} {} --gzip --replica",
924 : spec.tenant_id, spec.timeline_id
925 : )
926 : } else {
927 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
928 : }
929 : }
930 : _ => {
931 : if spec.spec.mode != ComputeMode::Primary {
932 : format!(
933 : "basebackup {} {} {} --gzip --replica",
934 : spec.tenant_id, spec.timeline_id, lsn
935 : )
936 : } else {
937 : format!(
938 : "basebackup {} {} {} --gzip",
939 : spec.tenant_id, spec.timeline_id, lsn
940 : )
941 : }
942 : }
943 : };
944 :
945 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
946 : let mut measured_reader = MeasuredReader::new(copyreader);
947 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
948 :
949 : // Read the archive directly from the `CopyOutReader`
950 : //
951 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
952 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
953 : // sends an Error after finishing the tarball, we will not notice it.
954 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
955 : ar.set_ignore_zeros(true);
956 : ar.unpack(&self.params.pgdata)?;
957 :
958 : // Report metrics
959 : let mut state = self.state.lock().unwrap();
960 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
961 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
962 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
963 : Ok(())
964 : }
965 :
966 : // Gets the basebackup in a retry loop
967 : #[instrument(skip_all, fields(%lsn))]
968 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
969 : let mut retry_period_ms = 500.0;
970 : let mut attempts = 0;
971 : const DEFAULT_ATTEMPTS: u16 = 10;
972 : #[cfg(feature = "testing")]
973 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
974 : u16::from_str(&v).unwrap()
975 : } else {
976 : DEFAULT_ATTEMPTS
977 : };
978 : #[cfg(not(feature = "testing"))]
979 : let max_attempts = DEFAULT_ATTEMPTS;
980 : loop {
981 : let result = self.try_get_basebackup(compute_state, lsn);
982 : match result {
983 : Ok(_) => {
984 : return result;
985 : }
986 : Err(ref e) if attempts < max_attempts => {
987 : warn!(
988 : "Failed to get basebackup: {} (attempt {}/{})",
989 : e, attempts, max_attempts
990 : );
991 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
992 : retry_period_ms *= 1.5;
993 : }
994 : Err(_) => {
995 : return result;
996 : }
997 : }
998 : attempts += 1;
999 : }
1000 : }
1001 :
1002 0 : pub async fn check_safekeepers_synced_async(
1003 0 : &self,
1004 0 : compute_state: &ComputeState,
1005 0 : ) -> Result<Option<Lsn>> {
1006 0 : // Construct a connection config for each safekeeper
1007 0 : let pspec: ParsedSpec = compute_state
1008 0 : .pspec
1009 0 : .as_ref()
1010 0 : .expect("spec must be set")
1011 0 : .clone();
1012 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
1013 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
1014 0 : // Format connstr
1015 0 : let id = connstr.clone();
1016 0 : let connstr = format!("postgresql://no_user@{}", connstr);
1017 0 : let options = format!(
1018 0 : "-c timeline_id={} tenant_id={}",
1019 0 : pspec.timeline_id, pspec.tenant_id
1020 0 : );
1021 0 :
1022 0 : // Construct client
1023 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
1024 0 : config.options(&options);
1025 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
1026 0 : config.password(storage_auth_token);
1027 0 : }
1028 :
1029 0 : (id, config)
1030 0 : });
1031 0 :
1032 0 : // Create task set to query all safekeepers
1033 0 : let mut tasks = FuturesUnordered::new();
1034 0 : let quorum = sk_configs.len() / 2 + 1;
1035 0 : for (id, config) in sk_configs {
1036 0 : let timeout = tokio::time::Duration::from_millis(100);
1037 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
1038 0 : tasks.push(tokio::spawn(task));
1039 0 : }
1040 :
1041 : // Get a quorum of responses or errors
1042 0 : let mut responses = Vec::new();
1043 0 : let mut join_errors = Vec::new();
1044 0 : let mut task_errors = Vec::new();
1045 0 : let mut timeout_errors = Vec::new();
1046 0 : while let Some(response) = tasks.next().await {
1047 0 : match response {
1048 0 : Ok(Ok(Ok(r))) => responses.push(r),
1049 0 : Ok(Ok(Err(e))) => task_errors.push(e),
1050 0 : Ok(Err(e)) => timeout_errors.push(e),
1051 0 : Err(e) => join_errors.push(e),
1052 : };
1053 0 : if responses.len() >= quorum {
1054 0 : break;
1055 0 : }
1056 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
1057 0 : break;
1058 0 : }
1059 : }
1060 :
1061 : // In case of error, log and fail the check, but don't crash.
1062 : // We're playing it safe because these errors could be transient
1063 : // and we don't yet retry. Also being careful here allows us to
1064 : // be backwards compatible with safekeepers that don't have the
1065 : // TIMELINE_STATUS API yet.
1066 0 : if responses.len() < quorum {
1067 0 : error!(
1068 0 : "failed sync safekeepers check {:?} {:?} {:?}",
1069 : join_errors, task_errors, timeout_errors
1070 : );
1071 0 : return Ok(None);
1072 0 : }
1073 0 :
1074 0 : Ok(check_if_synced(responses))
1075 0 : }
1076 :
1077 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
1078 : // in one roundtrip. If not, we should do a full sync_safekeepers.
1079 : #[instrument(skip_all)]
1080 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
1081 : let start_time = Utc::now();
1082 :
1083 : let rt = tokio::runtime::Handle::current();
1084 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
1085 :
1086 : // Record runtime
1087 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
1088 : .signed_duration_since(start_time)
1089 : .to_std()
1090 : .unwrap()
1091 : .as_millis() as u64;
1092 : result
1093 : }
1094 :
1095 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
1096 : // and return the reported LSN back to the caller.
1097 : #[instrument(skip_all)]
1098 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
1099 : let start_time = Utc::now();
1100 :
1101 : let mut sync_handle = maybe_cgexec(&self.params.pgbin)
1102 : .args(["--sync-safekeepers"])
1103 : .env("PGDATA", &self.params.pgdata) // we cannot use -D in this mode
1104 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1105 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1106 : } else {
1107 : vec![]
1108 : })
1109 : .stdout(Stdio::piped())
1110 : .stderr(Stdio::piped())
1111 : .spawn()
1112 : .expect("postgres --sync-safekeepers failed to start");
1113 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
1114 :
1115 : // `postgres --sync-safekeepers` will print all log output to stderr and
1116 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
1117 : // will be collected in a child thread.
1118 : let stderr = sync_handle
1119 : .stderr
1120 : .take()
1121 : .expect("stderr should be captured");
1122 : let logs_handle = handle_postgres_logs(stderr);
1123 :
1124 : let sync_output = sync_handle
1125 : .wait_with_output()
1126 : .expect("postgres --sync-safekeepers failed");
1127 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
1128 :
1129 : // Process has exited, so we can join the logs thread.
1130 : let _ = tokio::runtime::Handle::current()
1131 : .block_on(logs_handle)
1132 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1133 :
1134 : if !sync_output.status.success() {
1135 : anyhow::bail!(
1136 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
1137 : sync_output.status,
1138 : String::from_utf8(sync_output.stdout)
1139 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
1140 : );
1141 : }
1142 :
1143 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
1144 : .signed_duration_since(start_time)
1145 : .to_std()
1146 : .unwrap()
1147 : .as_millis() as u64;
1148 :
1149 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
1150 :
1151 : Ok(lsn)
1152 : }
1153 :
1154 : /// Do all the preparations like PGDATA directory creation, configuration,
1155 : /// safekeepers sync, basebackup, etc.
1156 : #[instrument(skip_all)]
1157 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
1158 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1159 : let spec = &pspec.spec;
1160 : let pgdata_path = Path::new(&self.params.pgdata);
1161 :
1162 : // Remove/create an empty pgdata directory and put configuration there.
1163 : self.create_pgdata()?;
1164 : config::write_postgres_conf(
1165 : pgdata_path,
1166 : &pspec.spec,
1167 : self.params.internal_http_port,
1168 : &self.compute_ctl_config.tls,
1169 : )?;
1170 :
1171 : // Syncing safekeepers is only safe with primary nodes: if a primary
1172 : // is already connected it will be kicked out, so a secondary (standby)
1173 : // cannot sync safekeepers.
1174 : let lsn = match spec.mode {
1175 : ComputeMode::Primary => {
1176 : info!("checking if safekeepers are synced");
1177 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
1178 : lsn
1179 : } else {
1180 : info!("starting safekeepers syncing");
1181 : self.sync_safekeepers(pspec.storage_auth_token.clone())
1182 0 : .with_context(|| "failed to sync safekeepers")?
1183 : };
1184 : info!("safekeepers synced at LSN {}", lsn);
1185 : lsn
1186 : }
1187 : ComputeMode::Static(lsn) => {
1188 : info!("Starting read-only node at static LSN {}", lsn);
1189 : lsn
1190 : }
1191 : ComputeMode::Replica => {
1192 : info!("Initializing standby from latest Pageserver LSN");
1193 : Lsn(0)
1194 : }
1195 : };
1196 :
1197 : info!(
1198 : "getting basebackup@{} from pageserver {}",
1199 : lsn, &pspec.pageserver_connstr
1200 : );
1201 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
1202 0 : format!(
1203 0 : "failed to get basebackup@{} from pageserver {}",
1204 0 : lsn, &pspec.pageserver_connstr
1205 0 : )
1206 0 : })?;
1207 :
1208 : // Update pg_hba.conf received with basebackup.
1209 : update_pg_hba(pgdata_path)?;
1210 :
1211 : // Place pg_dynshmem under /dev/shm. This allows us to use
1212 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
1213 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
1214 : //
1215 : // Why on earth don't we just stick to the 'posix' default, you might
1216 : // ask. It turns out that making large allocations with 'posix' doesn't
1217 : // work very well with autoscaling. The behavior we want is that:
1218 : //
1219 : // 1. You can make large DSM allocations, larger than the current RAM
1220 : // size of the VM, without errors
1221 : //
1222 : // 2. If the allocated memory is really used, the VM is scaled up
1223 : // automatically to accommodate that
1224 : //
1225 : // We try to make that possible by having swap in the VM. But with the
1226 : // default 'posix' DSM implementation, we fail step 1, even when there's
1227 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
1228 : // the shmem segment, which is really just a file in /dev/shm in Linux,
1229 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
1230 : // than available RAM.
1231 : //
1232 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
1233 : // the Postgres 'mmap' DSM implementation doesn't use
1234 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
1235 : // fill the file with zeros. It's weird that that differs between
1236 : // 'posix' and 'mmap', but we take advantage of it. When the file is
1237 : // filled slowly with write(2), the kernel allows it to grow larger, as
1238 : // long as there's swap available.
1239 : //
1240 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
1241 : // segment to be larger than currently available RAM. But because we
1242 : // don't want to store it on a real file, which the kernel would try to
1243 : // flush to disk, so symlink pg_dynshm to /dev/shm.
1244 : //
1245 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
1246 : // control plane control that option. If 'mmap' is not used, this
1247 : // symlink doesn't affect anything.
1248 : //
1249 : // See https://github.com/neondatabase/autoscaling/issues/800
1250 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
1251 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
1252 :
1253 : match spec.mode {
1254 : ComputeMode::Primary => {}
1255 : ComputeMode::Replica | ComputeMode::Static(..) => {
1256 : add_standby_signal(pgdata_path)?;
1257 : }
1258 : }
1259 :
1260 : Ok(())
1261 : }
1262 :
1263 : /// Start and stop a postgres process to warm up the VM for startup.
1264 0 : pub fn prewarm_postgres(&self) -> Result<()> {
1265 0 : info!("prewarming");
1266 :
1267 : // Create pgdata
1268 0 : let pgdata = &format!("{}.warmup", self.params.pgdata);
1269 0 : create_pgdata(pgdata)?;
1270 :
1271 : // Run initdb to completion
1272 0 : info!("running initdb");
1273 0 : let initdb_bin = Path::new(&self.params.pgbin)
1274 0 : .parent()
1275 0 : .unwrap()
1276 0 : .join("initdb");
1277 0 : Command::new(initdb_bin)
1278 0 : .args(["--pgdata", pgdata])
1279 0 : .output()
1280 0 : .expect("cannot start initdb process");
1281 :
1282 : // Write conf
1283 : use std::io::Write;
1284 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
1285 0 : let mut file = std::fs::File::create(conf_path)?;
1286 0 : writeln!(file, "shared_buffers=65536")?;
1287 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
1288 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
1289 :
1290 : // Start postgres
1291 0 : info!("starting postgres");
1292 0 : let mut pg = maybe_cgexec(&self.params.pgbin)
1293 0 : .args(["-D", pgdata])
1294 0 : .spawn()
1295 0 : .expect("cannot start postgres process");
1296 0 :
1297 0 : // Stop it when it's ready
1298 0 : info!("waiting for postgres");
1299 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
1300 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
1301 : // it to avoid orphaned processes prowling around while datadir is
1302 : // wiped.
1303 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
1304 0 : kill(pm_pid, Signal::SIGQUIT)?;
1305 0 : info!("sent SIGQUIT signal");
1306 0 : pg.wait()?;
1307 0 : info!("done prewarming");
1308 :
1309 : // clean up
1310 0 : let _ok = fs::remove_dir_all(pgdata);
1311 0 : Ok(())
1312 0 : }
1313 :
1314 : /// Start Postgres as a child process and wait for it to start accepting
1315 : /// connections.
1316 : ///
1317 : /// Returns a handle to the child process and a handle to the logs thread.
1318 : #[instrument(skip_all)]
1319 : pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
1320 : let pgdata_path = Path::new(&self.params.pgdata);
1321 :
1322 : // Run postgres as a child process.
1323 : let mut pg = maybe_cgexec(&self.params.pgbin)
1324 : .args(["-D", &self.params.pgdata])
1325 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1326 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1327 : } else {
1328 : vec![]
1329 : })
1330 : .stderr(Stdio::piped())
1331 : .spawn()
1332 : .expect("cannot start postgres process");
1333 : PG_PID.store(pg.id(), Ordering::SeqCst);
1334 :
1335 : // Start a task to collect logs from stderr.
1336 : let stderr = pg.stderr.take().expect("stderr should be captured");
1337 : let logs_handle = handle_postgres_logs(stderr);
1338 :
1339 : wait_for_postgres(&mut pg, pgdata_path)?;
1340 :
1341 : Ok(PostgresHandle {
1342 : postgres: pg,
1343 : log_collector: logs_handle,
1344 : })
1345 : }
1346 :
1347 : /// Wait for the child Postgres process forever. In this state Ctrl+C will
1348 : /// propagate to Postgres and it will be shut down as well.
1349 0 : fn wait_postgres(&self, mut pg_handle: PostgresHandle) -> std::process::ExitStatus {
1350 0 : info!(postmaster_pid = %pg_handle.postgres.id(), "Waiting for Postgres to exit");
1351 :
1352 0 : let ecode = pg_handle
1353 0 : .postgres
1354 0 : .wait()
1355 0 : .expect("failed to start waiting on Postgres process");
1356 0 : PG_PID.store(0, Ordering::SeqCst);
1357 0 :
1358 0 : // Process has exited. Wait for the log collecting task to finish.
1359 0 : let _ = tokio::runtime::Handle::current()
1360 0 : .block_on(pg_handle.log_collector)
1361 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1362 0 :
1363 0 : ecode
1364 0 : }
1365 :
1366 : /// Do post configuration of the already started Postgres. This function spawns a background task to
1367 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
1368 : /// version. In the future, it may upgrade all 3rd-party extensions.
1369 : #[instrument(skip_all)]
1370 : pub fn post_apply_config(&self) -> Result<()> {
1371 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
1372 0 : tokio::spawn(async move {
1373 0 : let res = async {
1374 0 : let (mut client, connection) = conf.connect(NoTls).await?;
1375 0 : tokio::spawn(async move {
1376 0 : if let Err(e) = connection.await {
1377 0 : eprintln!("connection error: {}", e);
1378 0 : }
1379 0 : });
1380 0 :
1381 0 : handle_neon_extension_upgrade(&mut client)
1382 0 : .await
1383 0 : .context("handle_neon_extension_upgrade")?;
1384 0 : Ok::<_, anyhow::Error>(())
1385 0 : }
1386 0 : .await;
1387 0 : if let Err(err) = res {
1388 0 : error!("error while post_apply_config: {err:#}");
1389 0 : }
1390 0 : });
1391 : Ok(())
1392 : }
1393 :
1394 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
1395 0 : let mut conf = self.conn_conf.clone();
1396 0 : if let Some(application_name) = application_name {
1397 0 : conf.application_name(application_name);
1398 0 : }
1399 0 : conf
1400 0 : }
1401 :
1402 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
1403 0 : let mut conf = self.tokio_conn_conf.clone();
1404 0 : if let Some(application_name) = application_name {
1405 0 : conf.application_name(application_name);
1406 0 : }
1407 0 : conf
1408 0 : }
1409 :
1410 0 : pub async fn get_maintenance_client(
1411 0 : conf: &tokio_postgres::Config,
1412 0 : ) -> Result<tokio_postgres::Client> {
1413 0 : let mut conf = conf.clone();
1414 0 : conf.application_name("compute_ctl:apply_config");
1415 :
1416 0 : let (client, conn) = match conf.connect(NoTls).await {
1417 : // If connection fails, it may be the old node with `zenith_admin` superuser.
1418 : //
1419 : // In this case we need to connect with old `zenith_admin` name
1420 : // and create new user. We cannot simply rename connected user,
1421 : // but we can create a new one and grant it all privileges.
1422 0 : Err(e) => match e.code() {
1423 : Some(&SqlState::INVALID_PASSWORD)
1424 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
1425 : // Connect with zenith_admin if cloud_admin could not authenticate
1426 0 : info!(
1427 0 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
1428 : e
1429 : );
1430 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
1431 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
1432 0 : zenith_admin_conf.user("zenith_admin");
1433 :
1434 0 : let mut client =
1435 0 : zenith_admin_conf.connect(NoTls)
1436 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
1437 :
1438 : // Disable forwarding so that users don't get a cloud_admin role
1439 0 : let mut func = || {
1440 0 : client.simple_query("SET neon.forward_ddl = false")?;
1441 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
1442 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
1443 0 : Ok::<_, anyhow::Error>(())
1444 0 : };
1445 0 : func().context("apply_config setup cloud_admin")?;
1446 :
1447 0 : drop(client);
1448 0 :
1449 0 : // Reconnect with connstring with expected name
1450 0 : conf.connect(NoTls).await?
1451 : }
1452 0 : _ => return Err(e.into()),
1453 : },
1454 0 : Ok((client, conn)) => (client, conn),
1455 : };
1456 :
1457 0 : spawn(async move {
1458 0 : if let Err(e) = conn.await {
1459 0 : error!("maintenance client connection error: {}", e);
1460 0 : }
1461 0 : });
1462 0 :
1463 0 : // Disable DDL forwarding because control plane already knows about the roles/databases
1464 0 : // we're about to modify.
1465 0 : client
1466 0 : .simple_query("SET neon.forward_ddl = false")
1467 0 : .await
1468 0 : .context("apply_config SET neon.forward_ddl = false")?;
1469 :
1470 0 : Ok(client)
1471 0 : }
1472 :
1473 : /// Do initial configuration of the already started Postgres.
1474 : #[instrument(skip_all)]
1475 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1476 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1477 :
1478 : let conf = Arc::new(conf);
1479 : let spec = Arc::new(
1480 : compute_state
1481 : .pspec
1482 : .as_ref()
1483 : .expect("spec must be set")
1484 : .spec
1485 : .clone(),
1486 : );
1487 :
1488 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1489 :
1490 : // Merge-apply spec & changes to PostgreSQL state.
1491 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1492 :
1493 : if let Some(local_proxy) = &spec.clone().local_proxy_config {
1494 : info!("configuring local_proxy");
1495 : local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
1496 : }
1497 :
1498 : // Run migrations separately to not hold up cold starts
1499 0 : tokio::spawn(async move {
1500 0 : let mut conf = conf.as_ref().clone();
1501 0 : conf.application_name("compute_ctl:migrations");
1502 0 :
1503 0 : match conf.connect(NoTls).await {
1504 0 : Ok((mut client, connection)) => {
1505 0 : tokio::spawn(async move {
1506 0 : if let Err(e) = connection.await {
1507 0 : eprintln!("connection error: {}", e);
1508 0 : }
1509 0 : });
1510 0 : if let Err(e) = handle_migrations(&mut client).await {
1511 0 : error!("Failed to run migrations: {}", e);
1512 0 : }
1513 : }
1514 0 : Err(e) => {
1515 0 : error!(
1516 0 : "Failed to connect to the compute for running migrations: {}",
1517 : e
1518 : );
1519 : }
1520 : };
1521 0 : });
1522 :
1523 : Ok::<(), anyhow::Error>(())
1524 : }
1525 :
1526 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1527 : // `pg_ctl` for start / stop.
1528 : #[instrument(skip_all)]
1529 : fn pg_reload_conf(&self) -> Result<()> {
1530 : let pgctl_bin = Path::new(&self.params.pgbin)
1531 : .parent()
1532 : .unwrap()
1533 : .join("pg_ctl");
1534 : Command::new(pgctl_bin)
1535 : .args(["reload", "-D", &self.params.pgdata])
1536 : .output()
1537 : .expect("cannot run pg_ctl process");
1538 : Ok(())
1539 : }
1540 :
1541 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1542 : /// as it's used to reconfigure a previously started and configured Postgres node.
1543 : #[instrument(skip_all)]
1544 : pub fn reconfigure(&self) -> Result<()> {
1545 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1546 :
1547 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1548 : info!("tuning pgbouncer");
1549 :
1550 : let pgbouncer_settings = pgbouncer_settings.clone();
1551 : let tls_config = self.compute_ctl_config.tls.clone();
1552 :
1553 : // Spawn a background task to do the tuning,
1554 : // so that we don't block the main thread that starts Postgres.
1555 0 : tokio::spawn(async move {
1556 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
1557 0 : if let Err(err) = res {
1558 0 : error!("error while tuning pgbouncer: {err:?}");
1559 0 : }
1560 0 : });
1561 : }
1562 :
1563 : if let Some(ref local_proxy) = spec.local_proxy_config {
1564 : info!("configuring local_proxy");
1565 :
1566 : // Spawn a background task to do the configuration,
1567 : // so that we don't block the main thread that starts Postgres.
1568 : let mut local_proxy = local_proxy.clone();
1569 : local_proxy.tls = self.compute_ctl_config.tls.clone();
1570 0 : tokio::spawn(async move {
1571 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1572 0 : error!("error while configuring local_proxy: {err:?}");
1573 0 : }
1574 0 : });
1575 : }
1576 :
1577 : // Reconfigure rsyslog for Postgres logs export
1578 : let conf = PostgresLogsRsyslogConfig::new(spec.logs_export_host.as_deref());
1579 : configure_postgres_logs_export(conf)?;
1580 :
1581 : // Write new config
1582 : let pgdata_path = Path::new(&self.params.pgdata);
1583 : config::write_postgres_conf(
1584 : pgdata_path,
1585 : &spec,
1586 : self.params.internal_http_port,
1587 : &self.compute_ctl_config.tls,
1588 : )?;
1589 :
1590 : if !spec.skip_pg_catalog_updates {
1591 : let max_concurrent_connections = spec.reconfigure_concurrency;
1592 : // Temporarily reset max_cluster_size in config
1593 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1594 : // creating new extensions, roles, etc.
1595 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1596 0 : self.pg_reload_conf()?;
1597 :
1598 0 : if spec.mode == ComputeMode::Primary {
1599 0 : let mut conf =
1600 0 : tokio_postgres::Config::from_str(self.params.connstr.as_str()).unwrap();
1601 0 : conf.application_name("apply_config");
1602 0 : let conf = Arc::new(conf);
1603 0 :
1604 0 : let spec = Arc::new(spec.clone());
1605 0 :
1606 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
1607 0 : }
1608 :
1609 0 : Ok(())
1610 0 : })?;
1611 : }
1612 :
1613 : self.pg_reload_conf()?;
1614 :
1615 : let unknown_op = "unknown".to_string();
1616 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1617 : info!(
1618 : "finished reconfiguration of compute node for operation {}",
1619 : op_id
1620 : );
1621 :
1622 : Ok(())
1623 : }
1624 :
1625 : #[instrument(skip_all)]
1626 : pub fn configure_as_primary(&self, compute_state: &ComputeState) -> Result<()> {
1627 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1628 :
1629 : assert!(pspec.spec.mode == ComputeMode::Primary);
1630 : if !pspec.spec.skip_pg_catalog_updates {
1631 : let pgdata_path = Path::new(&self.params.pgdata);
1632 : // temporarily reset max_cluster_size in config
1633 : // to avoid the possibility of hitting the limit, while we are applying config:
1634 : // creating new extensions, roles, etc...
1635 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1636 0 : self.pg_reload_conf()?;
1637 :
1638 0 : self.apply_config(compute_state)?;
1639 :
1640 0 : Ok(())
1641 0 : })?;
1642 :
1643 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1644 : if config::line_in_file(
1645 : &postgresql_conf_path,
1646 : "neon.disable_logical_replication_subscribers=false",
1647 : )? {
1648 : info!(
1649 : "updated postgresql.conf to set neon.disable_logical_replication_subscribers=false"
1650 : );
1651 : }
1652 : self.pg_reload_conf()?;
1653 : }
1654 : self.post_apply_config()?;
1655 :
1656 : Ok(())
1657 : }
1658 :
1659 0 : pub async fn watch_cert_for_changes(self: Arc<Self>) {
1660 : // update status on cert renewal
1661 0 : if let Some(tls_config) = &self.compute_ctl_config.tls {
1662 0 : let tls_config = tls_config.clone();
1663 :
1664 : // wait until the cert exists.
1665 0 : let mut cert_watch = watch_cert_for_changes(tls_config.cert_path.clone()).await;
1666 :
1667 0 : tokio::task::spawn_blocking(move || {
1668 0 : let handle = tokio::runtime::Handle::current();
1669 : 'cert_update: loop {
1670 : // let postgres/pgbouncer/local_proxy know the new cert/key exists.
1671 : // we need to wait until it's configurable first.
1672 :
1673 0 : let mut state = self.state.lock().unwrap();
1674 : 'status_update: loop {
1675 0 : match state.status {
1676 : // let's update the state to config pending
1677 : ComputeStatus::ConfigurationPending | ComputeStatus::Running => {
1678 0 : state.set_status(
1679 0 : ComputeStatus::ConfigurationPending,
1680 0 : &self.state_changed,
1681 0 : );
1682 0 : break 'status_update;
1683 : }
1684 :
1685 : // exit loop
1686 : ComputeStatus::Failed
1687 : | ComputeStatus::TerminationPending
1688 0 : | ComputeStatus::Terminated => break 'cert_update,
1689 :
1690 : // wait
1691 : ComputeStatus::Init
1692 : | ComputeStatus::Configuration
1693 0 : | ComputeStatus::Empty => {
1694 0 : state = self.state_changed.wait(state).unwrap();
1695 0 : }
1696 : }
1697 : }
1698 0 : drop(state);
1699 0 :
1700 0 : // wait for a new certificate update
1701 0 : if handle.block_on(cert_watch.changed()).is_err() {
1702 0 : break;
1703 0 : }
1704 : }
1705 0 : });
1706 0 : }
1707 0 : }
1708 :
1709 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1710 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1711 0 : let mut state = self.state.lock().unwrap();
1712 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1713 0 : if last_active > state.last_active {
1714 0 : state.last_active = last_active;
1715 0 : debug!("set the last compute activity time to: {:?}", last_active);
1716 0 : }
1717 0 : }
1718 :
1719 : // Look for core dumps and collect backtraces.
1720 : //
1721 : // EKS worker nodes have following core dump settings:
1722 : // /proc/sys/kernel/core_pattern -> core
1723 : // /proc/sys/kernel/core_uses_pid -> 1
1724 : // ulimit -c -> unlimited
1725 : // which results in core dumps being written to postgres data directory as core.<pid>.
1726 : //
1727 : // Use that as a default location and pattern, except macos where core dumps are written
1728 : // to /cores/ directory by default.
1729 : //
1730 : // With default Linux settings, the core dump file is called just "core", so check for
1731 : // that too.
1732 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1733 0 : let core_dump_dir = match std::env::consts::OS {
1734 0 : "macos" => Path::new("/cores/"),
1735 0 : _ => Path::new(&self.params.pgdata),
1736 : };
1737 :
1738 : // Collect core dump paths if any
1739 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1740 0 : let files = fs::read_dir(core_dump_dir)?;
1741 0 : let cores = files.filter_map(|entry| {
1742 0 : let entry = entry.ok()?;
1743 :
1744 0 : let is_core_dump = match entry.file_name().to_str()? {
1745 0 : n if n.starts_with("core.") => true,
1746 0 : "core" => true,
1747 0 : _ => false,
1748 : };
1749 0 : if is_core_dump {
1750 0 : Some(entry.path())
1751 : } else {
1752 0 : None
1753 : }
1754 0 : });
1755 :
1756 : // Print backtrace for each core dump
1757 0 : for core_path in cores {
1758 0 : warn!(
1759 0 : "core dump found: {}, collecting backtrace",
1760 0 : core_path.display()
1761 : );
1762 :
1763 : // Try first with gdb
1764 0 : let backtrace = Command::new("gdb")
1765 0 : .args(["--batch", "-q", "-ex", "bt", &self.params.pgbin])
1766 0 : .arg(&core_path)
1767 0 : .output();
1768 :
1769 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1770 0 : let backtrace = match backtrace {
1771 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1772 0 : warn!("cannot find gdb, trying lldb");
1773 0 : Command::new("lldb")
1774 0 : .arg("-c")
1775 0 : .arg(&core_path)
1776 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1777 0 : .output()
1778 : }
1779 0 : _ => backtrace,
1780 0 : }?;
1781 :
1782 0 : warn!(
1783 0 : "core dump backtrace: {}",
1784 0 : String::from_utf8_lossy(&backtrace.stdout)
1785 : );
1786 0 : warn!(
1787 0 : "debugger stderr: {}",
1788 0 : String::from_utf8_lossy(&backtrace.stderr)
1789 : );
1790 : }
1791 :
1792 0 : Ok(())
1793 0 : }
1794 :
1795 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1796 0 : pub async fn collect_insights(&self) -> String {
1797 0 : let mut result_rows: Vec<String> = Vec::new();
1798 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
1799 0 : let connect_result = conf.connect(NoTls).await;
1800 0 : let (client, connection) = connect_result.unwrap();
1801 0 : tokio::spawn(async move {
1802 0 : if let Err(e) = connection.await {
1803 0 : eprintln!("connection error: {}", e);
1804 0 : }
1805 0 : });
1806 0 : let result = client
1807 0 : .simple_query(
1808 0 : "SELECT
1809 0 : row_to_json(pg_stat_statements)
1810 0 : FROM
1811 0 : pg_stat_statements
1812 0 : WHERE
1813 0 : userid != 'cloud_admin'::regrole::oid
1814 0 : ORDER BY
1815 0 : (mean_exec_time + mean_plan_time) DESC
1816 0 : LIMIT 100",
1817 0 : )
1818 0 : .await;
1819 :
1820 0 : if let Ok(raw_rows) = result {
1821 0 : for message in raw_rows.iter() {
1822 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1823 0 : if let Some(json) = row.get(0) {
1824 0 : result_rows.push(json.to_string());
1825 0 : }
1826 0 : }
1827 : }
1828 :
1829 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1830 : } else {
1831 0 : "{{\"pg_stat_statements\": []}}".to_string()
1832 : }
1833 0 : }
1834 :
1835 : // download an archive, unzip and place files in correct locations
1836 0 : pub async fn download_extension(
1837 0 : &self,
1838 0 : real_ext_name: String,
1839 0 : ext_path: RemotePath,
1840 0 : ) -> Result<u64, DownloadError> {
1841 0 : let ext_remote_storage =
1842 0 : self.params
1843 0 : .ext_remote_storage
1844 0 : .as_ref()
1845 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1846 0 : "Remote extensions storage is not configured",
1847 0 : )))?;
1848 :
1849 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1850 0 :
1851 0 : let mut first_try = false;
1852 0 : if !self
1853 0 : .ext_download_progress
1854 0 : .read()
1855 0 : .expect("lock err")
1856 0 : .contains_key(ext_archive_name)
1857 0 : {
1858 0 : self.ext_download_progress
1859 0 : .write()
1860 0 : .expect("lock err")
1861 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1862 0 : first_try = true;
1863 0 : }
1864 0 : let (download_start, download_completed) =
1865 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1866 0 : let start_time_delta = Utc::now()
1867 0 : .signed_duration_since(download_start)
1868 0 : .to_std()
1869 0 : .unwrap()
1870 0 : .as_millis() as u64;
1871 :
1872 : // how long to wait for extension download if it was started by another process
1873 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1874 :
1875 0 : if download_completed {
1876 0 : info!("extension already downloaded, skipping re-download");
1877 0 : return Ok(0);
1878 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1879 0 : info!(
1880 0 : "download {ext_archive_name} already started by another process, hanging untill completion or timeout"
1881 : );
1882 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1883 : loop {
1884 0 : info!("waiting for download");
1885 0 : interval.tick().await;
1886 0 : let (_, download_completed_now) =
1887 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1888 0 : if download_completed_now {
1889 0 : info!("download finished by whoever else downloaded it");
1890 0 : return Ok(0);
1891 0 : }
1892 : }
1893 : // NOTE: the above loop will get terminated
1894 : // based on the timeout of the download function
1895 0 : }
1896 0 :
1897 0 : // if extension hasn't been downloaded before or the previous
1898 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1899 0 : // then we try to download it here
1900 0 : info!("downloading new extension {ext_archive_name}");
1901 :
1902 0 : let download_size = extension_server::download_extension(
1903 0 : &real_ext_name,
1904 0 : &ext_path,
1905 0 : ext_remote_storage,
1906 0 : &self.params.pgbin,
1907 0 : )
1908 0 : .await
1909 0 : .map_err(DownloadError::Other);
1910 0 :
1911 0 : if download_size.is_ok() {
1912 0 : self.ext_download_progress
1913 0 : .write()
1914 0 : .expect("bad lock")
1915 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1916 0 : }
1917 :
1918 0 : download_size
1919 0 : }
1920 :
1921 0 : pub async fn set_role_grants(
1922 0 : &self,
1923 0 : db_name: &PgIdent,
1924 0 : schema_name: &PgIdent,
1925 0 : privileges: &[Privilege],
1926 0 : role_name: &PgIdent,
1927 0 : ) -> Result<()> {
1928 : use tokio_postgres::NoTls;
1929 :
1930 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
1931 0 : conf.dbname(db_name);
1932 :
1933 0 : let (db_client, conn) = conf
1934 0 : .connect(NoTls)
1935 0 : .await
1936 0 : .context("Failed to connect to the database")?;
1937 0 : tokio::spawn(conn);
1938 0 :
1939 0 : // TODO: support other types of grants apart from schemas?
1940 0 : let query = format!(
1941 0 : "GRANT {} ON SCHEMA {} TO {}",
1942 0 : privileges
1943 0 : .iter()
1944 0 : // should not be quoted as it's part of the command.
1945 0 : // is already sanitized so it's ok
1946 0 : .map(|p| p.as_str())
1947 0 : .collect::<Vec<&'static str>>()
1948 0 : .join(", "),
1949 0 : // quote the schema and role name as identifiers to sanitize them.
1950 0 : schema_name.pg_quote(),
1951 0 : role_name.pg_quote(),
1952 0 : );
1953 0 : db_client
1954 0 : .simple_query(&query)
1955 0 : .await
1956 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1957 :
1958 0 : Ok(())
1959 0 : }
1960 :
1961 0 : pub async fn install_extension(
1962 0 : &self,
1963 0 : ext_name: &PgIdent,
1964 0 : db_name: &PgIdent,
1965 0 : ext_version: ExtVersion,
1966 0 : ) -> Result<ExtVersion> {
1967 : use tokio_postgres::NoTls;
1968 :
1969 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
1970 0 : conf.dbname(db_name);
1971 :
1972 0 : let (db_client, conn) = conf
1973 0 : .connect(NoTls)
1974 0 : .await
1975 0 : .context("Failed to connect to the database")?;
1976 0 : tokio::spawn(conn);
1977 0 :
1978 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
1979 0 : let version: Option<ExtVersion> = db_client
1980 0 : .query_opt(version_query, &[&ext_name])
1981 0 : .await
1982 0 : .with_context(|| format!("Failed to execute query: {}", version_query))?
1983 0 : .map(|row| row.get(0));
1984 0 :
1985 0 : // sanitize the inputs as postgres idents.
1986 0 : let ext_name: String = ext_name.pg_quote();
1987 0 : let quoted_version: String = ext_version.pg_quote();
1988 :
1989 0 : if let Some(installed_version) = version {
1990 0 : if installed_version == ext_version {
1991 0 : return Ok(installed_version);
1992 0 : }
1993 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
1994 0 : db_client
1995 0 : .simple_query(&query)
1996 0 : .await
1997 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1998 : } else {
1999 0 : let query =
2000 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
2001 0 : db_client
2002 0 : .simple_query(&query)
2003 0 : .await
2004 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
2005 : }
2006 :
2007 0 : Ok(ext_version)
2008 0 : }
2009 :
2010 0 : pub async fn prepare_preload_libraries(
2011 0 : &self,
2012 0 : spec: &ComputeSpec,
2013 0 : ) -> Result<RemoteExtensionMetrics> {
2014 0 : if self.params.ext_remote_storage.is_none() {
2015 0 : return Ok(RemoteExtensionMetrics {
2016 0 : num_ext_downloaded: 0,
2017 0 : largest_ext_size: 0,
2018 0 : total_ext_download_size: 0,
2019 0 : });
2020 0 : }
2021 0 : let remote_extensions = spec
2022 0 : .remote_extensions
2023 0 : .as_ref()
2024 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
2025 :
2026 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
2027 0 : let mut libs_vec = Vec::new();
2028 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
2029 0 : libs_vec = libs
2030 0 : .split(&[',', '\'', ' '])
2031 0 : .filter(|s| *s != "neon" && !s.is_empty())
2032 0 : .map(str::to_string)
2033 0 : .collect();
2034 0 : }
2035 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
2036 :
2037 : // that is used in neon_local and python tests
2038 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
2039 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
2040 0 : let mut shared_preload_libraries_line = "";
2041 0 : for line in conf_lines {
2042 0 : if line.starts_with("shared_preload_libraries") {
2043 0 : shared_preload_libraries_line = line;
2044 0 : }
2045 : }
2046 0 : let mut preload_libs_vec = Vec::new();
2047 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
2048 0 : preload_libs_vec = libs
2049 0 : .split(&[',', '\'', ' '])
2050 0 : .filter(|s| *s != "neon" && !s.is_empty())
2051 0 : .map(str::to_string)
2052 0 : .collect();
2053 0 : }
2054 0 : libs_vec.extend(preload_libs_vec);
2055 0 : }
2056 :
2057 : // Don't try to download libraries that are not in the index.
2058 : // Assume that they are already present locally.
2059 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
2060 0 :
2061 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
2062 :
2063 0 : let mut download_tasks = Vec::new();
2064 0 : for library in &libs_vec {
2065 0 : let (ext_name, ext_path) =
2066 0 : remote_extensions.get_ext(library, true, &BUILD_TAG, &self.params.pgversion)?;
2067 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
2068 : }
2069 0 : let results = join_all(download_tasks).await;
2070 :
2071 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
2072 0 : num_ext_downloaded: 0,
2073 0 : largest_ext_size: 0,
2074 0 : total_ext_download_size: 0,
2075 0 : };
2076 0 : for result in results {
2077 0 : let download_size = match result {
2078 0 : Ok(res) => {
2079 0 : remote_ext_metrics.num_ext_downloaded += 1;
2080 0 : res
2081 : }
2082 0 : Err(err) => {
2083 0 : // if we failed to download an extension, we don't want to fail the whole
2084 0 : // process, but we do want to log the error
2085 0 : error!("Failed to download extension: {}", err);
2086 0 : 0
2087 : }
2088 : };
2089 :
2090 0 : remote_ext_metrics.largest_ext_size =
2091 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
2092 0 : remote_ext_metrics.total_ext_download_size += download_size;
2093 : }
2094 0 : Ok(remote_ext_metrics)
2095 0 : }
2096 :
2097 : /// Waits until current thread receives a state changed notification and
2098 : /// the pageserver connection strings has changed.
2099 : ///
2100 : /// The operation will time out after a specified duration.
2101 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
2102 0 : let state = self.state.lock().unwrap();
2103 0 : let old_pageserver_connstr = state
2104 0 : .pspec
2105 0 : .as_ref()
2106 0 : .expect("spec must be set")
2107 0 : .pageserver_connstr
2108 0 : .clone();
2109 0 : let mut unchanged = true;
2110 0 : let _ = self
2111 0 : .state_changed
2112 0 : .wait_timeout_while(state, duration, |s| {
2113 0 : let pageserver_connstr = &s
2114 0 : .pspec
2115 0 : .as_ref()
2116 0 : .expect("spec must be set")
2117 0 : .pageserver_connstr;
2118 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
2119 0 : unchanged
2120 0 : })
2121 0 : .unwrap();
2122 0 : if !unchanged {
2123 0 : info!("Pageserver config changed");
2124 0 : }
2125 0 : }
2126 : }
2127 :
2128 0 : pub fn forward_termination_signal() {
2129 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2130 0 : if ss_pid != 0 {
2131 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2132 0 : kill(ss_pid, Signal::SIGTERM).ok();
2133 0 : }
2134 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2135 0 : if pg_pid != 0 {
2136 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2137 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2138 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2139 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2140 0 : kill(pg_pid, Signal::SIGINT).ok();
2141 0 : }
2142 0 : }
2143 :
2144 : // helper trait to call JoinSet::spawn_blocking(f), but propagates the current
2145 : // tracing span to the thread.
2146 : trait JoinSetExt<T> {
2147 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2148 : where
2149 : F: FnOnce() -> T + Send + 'static,
2150 : T: Send;
2151 : }
2152 :
2153 : impl<T: 'static> JoinSetExt<T> for tokio::task::JoinSet<T> {
2154 0 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2155 0 : where
2156 0 : F: FnOnce() -> T + Send + 'static,
2157 0 : T: Send,
2158 0 : {
2159 0 : let sp = tracing::Span::current();
2160 0 : self.spawn_blocking(move || {
2161 0 : let _e = sp.enter();
2162 0 : f()
2163 0 : })
2164 0 : }
2165 : }
|
