Line data Source code
1 : use std::collections::{HashMap, HashSet};
2 : use std::env;
3 : use std::fs;
4 : use std::iter::once;
5 : use std::os::unix::fs::{symlink, PermissionsExt};
6 : use std::path::Path;
7 : use std::process::{Command, Stdio};
8 : use std::str::FromStr;
9 : use std::sync::atomic::AtomicU32;
10 : use std::sync::atomic::Ordering;
11 : use std::sync::{Arc, Condvar, Mutex, RwLock};
12 : use std::thread;
13 : use std::time::Duration;
14 : use std::time::Instant;
15 :
16 : use anyhow::{Context, Result};
17 : use chrono::{DateTime, Utc};
18 : use compute_api::spec::{PgIdent, Role};
19 : use futures::future::join_all;
20 : use futures::stream::FuturesUnordered;
21 : use futures::StreamExt;
22 : use nix::unistd::Pid;
23 : use postgres::error::SqlState;
24 : use postgres::{Client, NoTls};
25 : use tracing::{debug, error, info, instrument, warn};
26 : use utils::id::{TenantId, TimelineId};
27 : use utils::lsn::Lsn;
28 :
29 : use compute_api::privilege::Privilege;
30 : use compute_api::responses::{ComputeMetrics, ComputeStatus};
31 : use compute_api::spec::{ComputeFeature, ComputeMode, ComputeSpec, ExtVersion};
32 : use utils::measured_stream::MeasuredReader;
33 :
34 : use nix::sys::signal::{kill, Signal};
35 : use remote_storage::{DownloadError, RemotePath};
36 : use tokio::spawn;
37 : use url::Url;
38 :
39 : use crate::installed_extensions::get_installed_extensions_sync;
40 : use crate::local_proxy;
41 : use crate::pg_helpers::*;
42 : use crate::spec::*;
43 : use crate::spec_apply::ApplySpecPhase::{
44 : CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreateSuperUser,
45 : DropInvalidDatabases, DropRoles, HandleNeonExtension, HandleOtherExtensions,
46 : RenameAndDeleteDatabases, RenameRoles, RunInEachDatabase,
47 : };
48 : use crate::spec_apply::PerDatabasePhase::{
49 : ChangeSchemaPerms, DeleteDBRoleReferences, HandleAnonExtension,
50 : };
51 : use crate::spec_apply::{apply_operations, MutableApplyContext, DB};
52 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
53 : use crate::{config, extension_server};
54 :
55 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
56 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
57 :
58 : /// Compute node info shared across several `compute_ctl` threads.
59 : pub struct ComputeNode {
60 : // Url type maintains proper escaping
61 : pub connstr: url::Url,
62 : pub pgdata: String,
63 : pub pgbin: String,
64 : pub pgversion: String,
65 : /// We should only allow live re- / configuration of the compute node if
66 : /// it uses 'pull model', i.e. it can go to control-plane and fetch
67 : /// the latest configuration. Otherwise, there could be a case:
68 : /// - we start compute with some spec provided as argument
69 : /// - we push new spec and it does reconfiguration
70 : /// - but then something happens and compute pod / VM is destroyed,
71 : /// so k8s controller starts it again with the **old** spec
72 : ///
73 : /// and the same for empty computes:
74 : /// - we started compute without any spec
75 : /// - we push spec and it does configuration
76 : /// - but then it is restarted without any spec again
77 : pub live_config_allowed: bool,
78 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
79 : /// To allow HTTP API server to serving status requests, while configuration
80 : /// is in progress, lock should be held only for short periods of time to do
81 : /// read/write, not the whole configuration process.
82 : pub state: Mutex<ComputeState>,
83 : /// `Condvar` to allow notifying waiters about state changes.
84 : pub state_changed: Condvar,
85 : /// the address of extension storage proxy gateway
86 : pub ext_remote_storage: Option<String>,
87 : // key: ext_archive_name, value: started download time, download_completed?
88 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
89 : pub build_tag: String,
90 : }
91 :
92 : // store some metrics about download size that might impact startup time
93 : #[derive(Clone, Debug)]
94 : pub struct RemoteExtensionMetrics {
95 : num_ext_downloaded: u64,
96 : largest_ext_size: u64,
97 : total_ext_download_size: u64,
98 : }
99 :
100 : #[derive(Clone, Debug)]
101 : pub struct ComputeState {
102 : pub start_time: DateTime<Utc>,
103 : pub status: ComputeStatus,
104 : /// Timestamp of the last Postgres activity. It could be `None` if
105 : /// compute wasn't used since start.
106 : pub last_active: Option<DateTime<Utc>>,
107 : pub error: Option<String>,
108 : pub pspec: Option<ParsedSpec>,
109 : pub metrics: ComputeMetrics,
110 : }
111 :
112 : impl ComputeState {
113 0 : pub fn new() -> Self {
114 0 : Self {
115 0 : start_time: Utc::now(),
116 0 : status: ComputeStatus::Empty,
117 0 : last_active: None,
118 0 : error: None,
119 0 : pspec: None,
120 0 : metrics: ComputeMetrics::default(),
121 0 : }
122 0 : }
123 :
124 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
125 0 : let prev = self.status;
126 0 : info!("Changing compute status from {} to {}", prev, status);
127 0 : self.status = status;
128 0 : state_changed.notify_all();
129 0 : }
130 :
131 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
132 0 : self.error = Some(format!("{err:?}"));
133 0 : self.set_status(ComputeStatus::Failed, state_changed);
134 0 : }
135 : }
136 :
137 : impl Default for ComputeState {
138 0 : fn default() -> Self {
139 0 : Self::new()
140 0 : }
141 : }
142 :
143 : #[derive(Clone, Debug)]
144 : pub struct ParsedSpec {
145 : pub spec: ComputeSpec,
146 : pub tenant_id: TenantId,
147 : pub timeline_id: TimelineId,
148 : pub pageserver_connstr: String,
149 : pub safekeeper_connstrings: Vec<String>,
150 : pub storage_auth_token: Option<String>,
151 : }
152 :
153 : impl TryFrom<ComputeSpec> for ParsedSpec {
154 : type Error = String;
155 0 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
156 : // Extract the options from the spec file that are needed to connect to
157 : // the storage system.
158 : //
159 : // For backwards-compatibility, the top-level fields in the spec file
160 : // may be empty. In that case, we need to dig them from the GUCs in the
161 : // cluster.settings field.
162 0 : let pageserver_connstr = spec
163 0 : .pageserver_connstring
164 0 : .clone()
165 0 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
166 0 : .ok_or("pageserver connstr should be provided")?;
167 0 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
168 0 : if matches!(spec.mode, ComputeMode::Primary) {
169 0 : spec.cluster
170 0 : .settings
171 0 : .find("neon.safekeepers")
172 0 : .ok_or("safekeeper connstrings should be provided")?
173 0 : .split(',')
174 0 : .map(|str| str.to_string())
175 0 : .collect()
176 : } else {
177 0 : vec![]
178 : }
179 : } else {
180 0 : spec.safekeeper_connstrings.clone()
181 : };
182 0 : let storage_auth_token = spec.storage_auth_token.clone();
183 0 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
184 0 : tenant_id
185 : } else {
186 0 : spec.cluster
187 0 : .settings
188 0 : .find("neon.tenant_id")
189 0 : .ok_or("tenant id should be provided")
190 0 : .map(|s| TenantId::from_str(&s))?
191 0 : .or(Err("invalid tenant id"))?
192 : };
193 0 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
194 0 : timeline_id
195 : } else {
196 0 : spec.cluster
197 0 : .settings
198 0 : .find("neon.timeline_id")
199 0 : .ok_or("timeline id should be provided")
200 0 : .map(|s| TimelineId::from_str(&s))?
201 0 : .or(Err("invalid timeline id"))?
202 : };
203 :
204 0 : Ok(ParsedSpec {
205 0 : spec,
206 0 : pageserver_connstr,
207 0 : safekeeper_connstrings,
208 0 : storage_auth_token,
209 0 : tenant_id,
210 0 : timeline_id,
211 0 : })
212 0 : }
213 : }
214 :
215 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
216 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
217 : ///
218 : /// This function should be used to start postgres, as it will start it in the
219 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
220 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
221 : /// creates it during the sysinit phase of its inittab.
222 0 : fn maybe_cgexec(cmd: &str) -> Command {
223 0 : // The cplane sets this env var for autoscaling computes.
224 0 : // use `var_os` so we don't have to worry about the variable being valid
225 0 : // unicode. Should never be an concern . . . but just in case
226 0 : if env::var_os("AUTOSCALING").is_some() {
227 0 : let mut command = Command::new("cgexec");
228 0 : command.args(["-g", "memory:neon-postgres"]);
229 0 : command.arg(cmd);
230 0 : command
231 : } else {
232 0 : Command::new(cmd)
233 : }
234 0 : }
235 :
236 0 : pub(crate) fn construct_superuser_query(spec: &ComputeSpec) -> String {
237 0 : let roles = spec
238 0 : .cluster
239 0 : .roles
240 0 : .iter()
241 0 : .map(|r| escape_literal(&r.name))
242 0 : .collect::<Vec<_>>();
243 0 :
244 0 : let dbs = spec
245 0 : .cluster
246 0 : .databases
247 0 : .iter()
248 0 : .map(|db| escape_literal(&db.name))
249 0 : .collect::<Vec<_>>();
250 :
251 0 : let roles_decl = if roles.is_empty() {
252 0 : String::from("roles text[] := NULL;")
253 : } else {
254 0 : format!(
255 0 : r#"
256 0 : roles text[] := ARRAY(SELECT rolname
257 0 : FROM pg_catalog.pg_roles
258 0 : WHERE rolname IN ({}));"#,
259 0 : roles.join(", ")
260 0 : )
261 : };
262 :
263 0 : let database_decl = if dbs.is_empty() {
264 0 : String::from("dbs text[] := NULL;")
265 : } else {
266 0 : format!(
267 0 : r#"
268 0 : dbs text[] := ARRAY(SELECT datname
269 0 : FROM pg_catalog.pg_database
270 0 : WHERE datname IN ({}));"#,
271 0 : dbs.join(", ")
272 0 : )
273 : };
274 :
275 : // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on all databases
276 : // (see https://www.postgresql.org/docs/current/ddl-priv.html)
277 0 : let query = format!(
278 0 : r#"
279 0 : DO $$
280 0 : DECLARE
281 0 : r text;
282 0 : {}
283 0 : {}
284 0 : BEGIN
285 0 : IF NOT EXISTS (
286 0 : SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
287 0 : THEN
288 0 : CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
289 0 : IF array_length(roles, 1) IS NOT NULL THEN
290 0 : EXECUTE format('GRANT neon_superuser TO %s',
291 0 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
292 0 : FOREACH r IN ARRAY roles LOOP
293 0 : EXECUTE format('ALTER ROLE %s CREATEROLE CREATEDB', quote_ident(r));
294 0 : END LOOP;
295 0 : END IF;
296 0 : IF array_length(dbs, 1) IS NOT NULL THEN
297 0 : EXECUTE format('GRANT ALL PRIVILEGES ON DATABASE %s TO neon_superuser',
298 0 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(dbs) as x), ', '));
299 0 : END IF;
300 0 : END IF;
301 0 : END
302 0 : $$;"#,
303 0 : roles_decl, database_decl,
304 0 : );
305 0 :
306 0 : query
307 0 : }
308 :
309 : impl ComputeNode {
310 : /// Check that compute node has corresponding feature enabled.
311 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
312 0 : let state = self.state.lock().unwrap();
313 :
314 0 : if let Some(s) = state.pspec.as_ref() {
315 0 : s.spec.features.contains(&feature)
316 : } else {
317 0 : false
318 : }
319 0 : }
320 :
321 0 : pub fn set_status(&self, status: ComputeStatus) {
322 0 : let mut state = self.state.lock().unwrap();
323 0 : state.set_status(status, &self.state_changed);
324 0 : }
325 :
326 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
327 0 : let mut state = self.state.lock().unwrap();
328 0 : state.set_failed_status(err, &self.state_changed);
329 0 : }
330 :
331 0 : pub fn get_status(&self) -> ComputeStatus {
332 0 : self.state.lock().unwrap().status
333 0 : }
334 :
335 : // Remove `pgdata` directory and create it again with right permissions.
336 0 : fn create_pgdata(&self) -> Result<()> {
337 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
338 0 : // If it is something different then create_dir() will error out anyway.
339 0 : let _ok = fs::remove_dir_all(&self.pgdata);
340 0 : fs::create_dir(&self.pgdata)?;
341 0 : fs::set_permissions(&self.pgdata, fs::Permissions::from_mode(0o700))?;
342 :
343 0 : Ok(())
344 0 : }
345 :
346 : // Get basebackup from the libpq connection to pageserver using `connstr` and
347 : // unarchive it to `pgdata` directory overriding all its previous content.
348 0 : #[instrument(skip_all, fields(%lsn))]
349 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
350 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
351 : let start_time = Instant::now();
352 :
353 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
354 : let mut config = postgres::Config::from_str(shard0_connstr)?;
355 :
356 : // Use the storage auth token from the config file, if given.
357 : // Note: this overrides any password set in the connection string.
358 : if let Some(storage_auth_token) = &spec.storage_auth_token {
359 : info!("Got storage auth token from spec file");
360 : config.password(storage_auth_token);
361 : } else {
362 : info!("Storage auth token not set");
363 : }
364 :
365 : // Connect to pageserver
366 : let mut client = config.connect(NoTls)?;
367 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
368 :
369 : let basebackup_cmd = match lsn {
370 : Lsn(0) => {
371 : if spec.spec.mode != ComputeMode::Primary {
372 : format!(
373 : "basebackup {} {} --gzip --replica",
374 : spec.tenant_id, spec.timeline_id
375 : )
376 : } else {
377 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
378 : }
379 : }
380 : _ => {
381 : if spec.spec.mode != ComputeMode::Primary {
382 : format!(
383 : "basebackup {} {} {} --gzip --replica",
384 : spec.tenant_id, spec.timeline_id, lsn
385 : )
386 : } else {
387 : format!(
388 : "basebackup {} {} {} --gzip",
389 : spec.tenant_id, spec.timeline_id, lsn
390 : )
391 : }
392 : }
393 : };
394 :
395 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
396 : let mut measured_reader = MeasuredReader::new(copyreader);
397 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
398 :
399 : // Read the archive directly from the `CopyOutReader`
400 : //
401 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
402 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
403 : // sends an Error after finishing the tarball, we will not notice it.
404 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
405 : ar.set_ignore_zeros(true);
406 : ar.unpack(&self.pgdata)?;
407 :
408 : // Report metrics
409 : let mut state = self.state.lock().unwrap();
410 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
411 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
412 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
413 : Ok(())
414 : }
415 :
416 : // Gets the basebackup in a retry loop
417 0 : #[instrument(skip_all, fields(%lsn))]
418 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
419 : let mut retry_period_ms = 500.0;
420 : let mut attempts = 0;
421 : const DEFAULT_ATTEMPTS: u16 = 10;
422 : #[cfg(feature = "testing")]
423 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
424 : u16::from_str(&v).unwrap()
425 : } else {
426 : DEFAULT_ATTEMPTS
427 : };
428 : #[cfg(not(feature = "testing"))]
429 : let max_attempts = DEFAULT_ATTEMPTS;
430 : loop {
431 : let result = self.try_get_basebackup(compute_state, lsn);
432 : match result {
433 : Ok(_) => {
434 : return result;
435 : }
436 : Err(ref e) if attempts < max_attempts => {
437 : warn!(
438 : "Failed to get basebackup: {} (attempt {}/{})",
439 : e, attempts, max_attempts
440 : );
441 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
442 : retry_period_ms *= 1.5;
443 : }
444 : Err(_) => {
445 : return result;
446 : }
447 : }
448 : attempts += 1;
449 : }
450 : }
451 :
452 0 : pub async fn check_safekeepers_synced_async(
453 0 : &self,
454 0 : compute_state: &ComputeState,
455 0 : ) -> Result<Option<Lsn>> {
456 0 : // Construct a connection config for each safekeeper
457 0 : let pspec: ParsedSpec = compute_state
458 0 : .pspec
459 0 : .as_ref()
460 0 : .expect("spec must be set")
461 0 : .clone();
462 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
463 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
464 0 : // Format connstr
465 0 : let id = connstr.clone();
466 0 : let connstr = format!("postgresql://no_user@{}", connstr);
467 0 : let options = format!(
468 0 : "-c timeline_id={} tenant_id={}",
469 0 : pspec.timeline_id, pspec.tenant_id
470 0 : );
471 0 :
472 0 : // Construct client
473 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
474 0 : config.options(&options);
475 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
476 0 : config.password(storage_auth_token);
477 0 : }
478 :
479 0 : (id, config)
480 0 : });
481 0 :
482 0 : // Create task set to query all safekeepers
483 0 : let mut tasks = FuturesUnordered::new();
484 0 : let quorum = sk_configs.len() / 2 + 1;
485 0 : for (id, config) in sk_configs {
486 0 : let timeout = tokio::time::Duration::from_millis(100);
487 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
488 0 : tasks.push(tokio::spawn(task));
489 0 : }
490 :
491 : // Get a quorum of responses or errors
492 0 : let mut responses = Vec::new();
493 0 : let mut join_errors = Vec::new();
494 0 : let mut task_errors = Vec::new();
495 0 : let mut timeout_errors = Vec::new();
496 0 : while let Some(response) = tasks.next().await {
497 0 : match response {
498 0 : Ok(Ok(Ok(r))) => responses.push(r),
499 0 : Ok(Ok(Err(e))) => task_errors.push(e),
500 0 : Ok(Err(e)) => timeout_errors.push(e),
501 0 : Err(e) => join_errors.push(e),
502 : };
503 0 : if responses.len() >= quorum {
504 0 : break;
505 0 : }
506 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
507 0 : break;
508 0 : }
509 : }
510 :
511 : // In case of error, log and fail the check, but don't crash.
512 : // We're playing it safe because these errors could be transient
513 : // and we don't yet retry. Also being careful here allows us to
514 : // be backwards compatible with safekeepers that don't have the
515 : // TIMELINE_STATUS API yet.
516 0 : if responses.len() < quorum {
517 0 : error!(
518 0 : "failed sync safekeepers check {:?} {:?} {:?}",
519 : join_errors, task_errors, timeout_errors
520 : );
521 0 : return Ok(None);
522 0 : }
523 0 :
524 0 : Ok(check_if_synced(responses))
525 0 : }
526 :
527 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
528 : // in one roundtrip. If not, we should do a full sync_safekeepers.
529 0 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
530 0 : let start_time = Utc::now();
531 0 :
532 0 : // Run actual work with new tokio runtime
533 0 : let rt = tokio::runtime::Builder::new_current_thread()
534 0 : .enable_all()
535 0 : .build()
536 0 : .expect("failed to create rt");
537 0 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
538 0 :
539 0 : // Record runtime
540 0 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
541 0 : .signed_duration_since(start_time)
542 0 : .to_std()
543 0 : .unwrap()
544 0 : .as_millis() as u64;
545 0 : result
546 0 : }
547 :
548 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
549 : // and return the reported LSN back to the caller.
550 0 : #[instrument(skip_all)]
551 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
552 : let start_time = Utc::now();
553 :
554 : let mut sync_handle = maybe_cgexec(&self.pgbin)
555 : .args(["--sync-safekeepers"])
556 : .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
557 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
558 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
559 : } else {
560 : vec![]
561 : })
562 : .stdout(Stdio::piped())
563 : .stderr(Stdio::piped())
564 : .spawn()
565 : .expect("postgres --sync-safekeepers failed to start");
566 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
567 :
568 : // `postgres --sync-safekeepers` will print all log output to stderr and
569 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
570 : // will be collected in a child thread.
571 : let stderr = sync_handle
572 : .stderr
573 : .take()
574 : .expect("stderr should be captured");
575 : let logs_handle = handle_postgres_logs(stderr);
576 :
577 : let sync_output = sync_handle
578 : .wait_with_output()
579 : .expect("postgres --sync-safekeepers failed");
580 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
581 :
582 : // Process has exited, so we can join the logs thread.
583 : let _ = logs_handle
584 : .join()
585 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
586 :
587 : if !sync_output.status.success() {
588 : anyhow::bail!(
589 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
590 : sync_output.status,
591 : String::from_utf8(sync_output.stdout)
592 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
593 : );
594 : }
595 :
596 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
597 : .signed_duration_since(start_time)
598 : .to_std()
599 : .unwrap()
600 : .as_millis() as u64;
601 :
602 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
603 :
604 : Ok(lsn)
605 : }
606 :
607 : /// Do all the preparations like PGDATA directory creation, configuration,
608 : /// safekeepers sync, basebackup, etc.
609 0 : #[instrument(skip_all)]
610 : pub fn prepare_pgdata(
611 : &self,
612 : compute_state: &ComputeState,
613 : extension_server_port: u16,
614 : ) -> Result<()> {
615 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
616 : let spec = &pspec.spec;
617 : let pgdata_path = Path::new(&self.pgdata);
618 :
619 : // Remove/create an empty pgdata directory and put configuration there.
620 : self.create_pgdata()?;
621 : config::write_postgres_conf(
622 : &pgdata_path.join("postgresql.conf"),
623 : &pspec.spec,
624 : Some(extension_server_port),
625 : )?;
626 :
627 : // Syncing safekeepers is only safe with primary nodes: if a primary
628 : // is already connected it will be kicked out, so a secondary (standby)
629 : // cannot sync safekeepers.
630 : let lsn = match spec.mode {
631 : ComputeMode::Primary => {
632 : info!("checking if safekeepers are synced");
633 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
634 : lsn
635 : } else {
636 : info!("starting safekeepers syncing");
637 : self.sync_safekeepers(pspec.storage_auth_token.clone())
638 0 : .with_context(|| "failed to sync safekeepers")?
639 : };
640 : info!("safekeepers synced at LSN {}", lsn);
641 : lsn
642 : }
643 : ComputeMode::Static(lsn) => {
644 : info!("Starting read-only node at static LSN {}", lsn);
645 : lsn
646 : }
647 : ComputeMode::Replica => {
648 : info!("Initializing standby from latest Pageserver LSN");
649 : Lsn(0)
650 : }
651 : };
652 :
653 : info!(
654 : "getting basebackup@{} from pageserver {}",
655 : lsn, &pspec.pageserver_connstr
656 : );
657 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
658 0 : format!(
659 0 : "failed to get basebackup@{} from pageserver {}",
660 0 : lsn, &pspec.pageserver_connstr
661 0 : )
662 0 : })?;
663 :
664 : // Update pg_hba.conf received with basebackup.
665 : update_pg_hba(pgdata_path)?;
666 :
667 : // Place pg_dynshmem under /dev/shm. This allows us to use
668 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
669 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
670 : //
671 : // Why on earth don't we just stick to the 'posix' default, you might
672 : // ask. It turns out that making large allocations with 'posix' doesn't
673 : // work very well with autoscaling. The behavior we want is that:
674 : //
675 : // 1. You can make large DSM allocations, larger than the current RAM
676 : // size of the VM, without errors
677 : //
678 : // 2. If the allocated memory is really used, the VM is scaled up
679 : // automatically to accommodate that
680 : //
681 : // We try to make that possible by having swap in the VM. But with the
682 : // default 'posix' DSM implementation, we fail step 1, even when there's
683 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
684 : // the shmem segment, which is really just a file in /dev/shm in Linux,
685 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
686 : // than available RAM.
687 : //
688 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
689 : // the Postgres 'mmap' DSM implementation doesn't use
690 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
691 : // fill the file with zeros. It's weird that that differs between
692 : // 'posix' and 'mmap', but we take advantage of it. When the file is
693 : // filled slowly with write(2), the kernel allows it to grow larger, as
694 : // long as there's swap available.
695 : //
696 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
697 : // segment to be larger than currently available RAM. But because we
698 : // don't want to store it on a real file, which the kernel would try to
699 : // flush to disk, so symlink pg_dynshm to /dev/shm.
700 : //
701 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
702 : // control plane control that option. If 'mmap' is not used, this
703 : // symlink doesn't affect anything.
704 : //
705 : // See https://github.com/neondatabase/autoscaling/issues/800
706 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
707 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
708 :
709 : match spec.mode {
710 : ComputeMode::Primary => {}
711 : ComputeMode::Replica | ComputeMode::Static(..) => {
712 : add_standby_signal(pgdata_path)?;
713 : }
714 : }
715 :
716 : Ok(())
717 : }
718 :
719 : /// Start and stop a postgres process to warm up the VM for startup.
720 0 : pub fn prewarm_postgres(&self) -> Result<()> {
721 0 : info!("prewarming");
722 :
723 : // Create pgdata
724 0 : let pgdata = &format!("{}.warmup", self.pgdata);
725 0 : create_pgdata(pgdata)?;
726 :
727 : // Run initdb to completion
728 0 : info!("running initdb");
729 0 : let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
730 0 : Command::new(initdb_bin)
731 0 : .args(["--pgdata", pgdata])
732 0 : .output()
733 0 : .expect("cannot start initdb process");
734 :
735 : // Write conf
736 : use std::io::Write;
737 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
738 0 : let mut file = std::fs::File::create(conf_path)?;
739 0 : writeln!(file, "shared_buffers=65536")?;
740 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
741 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
742 :
743 : // Start postgres
744 0 : info!("starting postgres");
745 0 : let mut pg = maybe_cgexec(&self.pgbin)
746 0 : .args(["-D", pgdata])
747 0 : .spawn()
748 0 : .expect("cannot start postgres process");
749 0 :
750 0 : // Stop it when it's ready
751 0 : info!("waiting for postgres");
752 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
753 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
754 : // it to avoid orphaned processes prowling around while datadir is
755 : // wiped.
756 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
757 0 : kill(pm_pid, Signal::SIGQUIT)?;
758 0 : info!("sent SIGQUIT signal");
759 0 : pg.wait()?;
760 0 : info!("done prewarming");
761 :
762 : // clean up
763 0 : let _ok = fs::remove_dir_all(pgdata);
764 0 : Ok(())
765 0 : }
766 :
767 : /// Start Postgres as a child process and manage DBs/roles.
768 : /// After that this will hang waiting on the postmaster process to exit.
769 : /// Returns a handle to the child process and a handle to the logs thread.
770 0 : #[instrument(skip_all)]
771 : pub fn start_postgres(
772 : &self,
773 : storage_auth_token: Option<String>,
774 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
775 : let pgdata_path = Path::new(&self.pgdata);
776 :
777 : // Run postgres as a child process.
778 : let mut pg = maybe_cgexec(&self.pgbin)
779 : .args(["-D", &self.pgdata])
780 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
781 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
782 : } else {
783 : vec![]
784 : })
785 : .stderr(Stdio::piped())
786 : .spawn()
787 : .expect("cannot start postgres process");
788 : PG_PID.store(pg.id(), Ordering::SeqCst);
789 :
790 : // Start a thread to collect logs from stderr.
791 : let stderr = pg.stderr.take().expect("stderr should be captured");
792 : let logs_handle = handle_postgres_logs(stderr);
793 :
794 : wait_for_postgres(&mut pg, pgdata_path)?;
795 :
796 : Ok((pg, logs_handle))
797 : }
798 :
799 : /// Do post configuration of the already started Postgres. This function spawns a background thread to
800 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
801 : /// version. In the future, it may upgrade all 3rd-party extensions.
802 0 : #[instrument(skip_all)]
803 : pub fn post_apply_config(&self) -> Result<()> {
804 : let connstr = self.connstr.clone();
805 0 : thread::spawn(move || {
806 0 : let func = || {
807 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
808 0 : handle_neon_extension_upgrade(&mut client)
809 0 : .context("handle_neon_extension_upgrade")?;
810 0 : Ok::<_, anyhow::Error>(())
811 0 : };
812 0 : if let Err(err) = func() {
813 0 : error!("error while post_apply_config: {err:#}");
814 0 : }
815 0 : });
816 : Ok(())
817 : }
818 :
819 0 : async fn get_maintenance_client(url: &Url) -> Result<tokio_postgres::Client> {
820 0 : let mut connstr = url.clone();
821 0 :
822 0 : connstr
823 0 : .query_pairs_mut()
824 0 : .append_pair("application_name", "apply_config");
825 :
826 0 : let (client, conn) = match tokio_postgres::connect(connstr.as_str(), NoTls).await {
827 0 : Err(e) => match e.code() {
828 : Some(&SqlState::INVALID_PASSWORD)
829 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
830 : // connect with zenith_admin if cloud_admin could not authenticate
831 0 : info!(
832 0 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
833 : e
834 : );
835 0 : let mut zenith_admin_connstr = connstr.clone();
836 0 :
837 0 : zenith_admin_connstr
838 0 : .set_username("zenith_admin")
839 0 : .map_err(|_| anyhow::anyhow!("invalid connstr"))?;
840 :
841 0 : let mut client =
842 0 : Client::connect(zenith_admin_connstr.as_str(), NoTls)
843 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
844 :
845 : // Disable forwarding so that users don't get a cloud_admin role
846 0 : let mut func = || {
847 0 : client.simple_query("SET neon.forward_ddl = false")?;
848 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
849 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
850 0 : Ok::<_, anyhow::Error>(())
851 0 : };
852 0 : func().context("apply_config setup cloud_admin")?;
853 :
854 0 : drop(client);
855 0 :
856 0 : // reconnect with connstring with expected name
857 0 : tokio_postgres::connect(connstr.as_str(), NoTls).await?
858 : }
859 0 : _ => return Err(e.into()),
860 : },
861 0 : Ok((client, conn)) => (client, conn),
862 : };
863 :
864 0 : spawn(async move {
865 0 : if let Err(e) = conn.await {
866 0 : error!("maintenance client connection error: {}", e);
867 0 : }
868 0 : });
869 0 :
870 0 : // Disable DDL forwarding because control plane already knows about the roles/databases
871 0 : // we're about to modify.
872 0 : client
873 0 : .simple_query("SET neon.forward_ddl = false")
874 0 : .await
875 0 : .context("apply_config SET neon.forward_ddl = false")?;
876 :
877 0 : Ok(client)
878 0 : }
879 :
880 : /// Apply the spec to the running PostgreSQL instance.
881 : /// The caller can decide to run with multiple clients in parallel, or
882 : /// single mode. Either way, the commands executed will be the same, and
883 : /// only commands run in different databases are parallelized.
884 0 : #[instrument(skip_all)]
885 : pub fn apply_spec_sql(
886 : &self,
887 : spec: Arc<ComputeSpec>,
888 : url: Arc<Url>,
889 : concurrency: usize,
890 : ) -> Result<()> {
891 : let rt = tokio::runtime::Builder::new_multi_thread()
892 : .enable_all()
893 : .build()?;
894 :
895 : info!("Applying config with max {} concurrency", concurrency);
896 : debug!("Config: {:?}", spec);
897 :
898 0 : rt.block_on(async {
899 : // Proceed with post-startup configuration. Note, that order of operations is important.
900 0 : let client = Self::get_maintenance_client(&url).await?;
901 0 : let spec = spec.clone();
902 :
903 0 : let databases = get_existing_dbs_async(&client).await?;
904 0 : let roles = get_existing_roles_async(&client)
905 0 : .await?
906 0 : .into_iter()
907 0 : .map(|role| (role.name.clone(), role))
908 0 : .collect::<HashMap<String, Role>>();
909 0 :
910 0 : let jwks_roles = Arc::new(
911 0 : spec.as_ref()
912 0 : .local_proxy_config
913 0 : .iter()
914 0 : .flat_map(|it| &it.jwks)
915 0 : .flatten()
916 0 : .flat_map(|setting| &setting.role_names)
917 0 : .cloned()
918 0 : .collect::<HashSet<_>>(),
919 0 : );
920 0 :
921 0 : let ctx = Arc::new(tokio::sync::RwLock::new(MutableApplyContext {
922 0 : roles,
923 0 : dbs: databases,
924 0 : }));
925 :
926 0 : for phase in [
927 0 : CreateSuperUser,
928 0 : DropInvalidDatabases,
929 0 : RenameRoles,
930 0 : CreateAndAlterRoles,
931 0 : RenameAndDeleteDatabases,
932 0 : CreateAndAlterDatabases,
933 : ] {
934 0 : debug!("Applying phase {:?}", &phase);
935 0 : apply_operations(
936 0 : spec.clone(),
937 0 : ctx.clone(),
938 0 : jwks_roles.clone(),
939 0 : phase,
940 0 : || async { Ok(&client) },
941 0 : )
942 0 : .await?;
943 : }
944 :
945 0 : let concurrency_token = Arc::new(tokio::sync::Semaphore::new(concurrency));
946 0 :
947 0 : let db_processes = spec
948 0 : .cluster
949 0 : .databases
950 0 : .iter()
951 0 : .map(|db| DB::new(db.clone()))
952 0 : // include
953 0 : .chain(once(DB::SystemDB))
954 0 : .map(|db| {
955 0 : let spec = spec.clone();
956 0 : let ctx = ctx.clone();
957 0 : let jwks_roles = jwks_roles.clone();
958 0 : let mut url = url.as_ref().clone();
959 0 : let concurrency_token = concurrency_token.clone();
960 0 : let db = db.clone();
961 0 :
962 0 : debug!("Applying per-database phases for Database {:?}", &db);
963 :
964 0 : match &db {
965 0 : DB::SystemDB => {}
966 0 : DB::UserDB(db) => {
967 0 : url.set_path(db.name.as_str());
968 0 : }
969 : }
970 :
971 0 : let url = Arc::new(url);
972 0 : let fut = Self::apply_spec_sql_db(
973 0 : spec.clone(),
974 0 : url,
975 0 : ctx.clone(),
976 0 : jwks_roles.clone(),
977 0 : concurrency_token.clone(),
978 0 : db,
979 0 : );
980 0 :
981 0 : Ok(spawn(fut))
982 0 : })
983 0 : .collect::<Vec<Result<_, anyhow::Error>>>();
984 :
985 0 : for process in db_processes.into_iter() {
986 0 : let handle = process?;
987 0 : handle.await??;
988 : }
989 :
990 0 : for phase in vec![
991 0 : HandleOtherExtensions,
992 0 : HandleNeonExtension,
993 0 : CreateAvailabilityCheck,
994 0 : DropRoles,
995 0 : ] {
996 0 : debug!("Applying phase {:?}", &phase);
997 0 : apply_operations(
998 0 : spec.clone(),
999 0 : ctx.clone(),
1000 0 : jwks_roles.clone(),
1001 0 : phase,
1002 0 : || async { Ok(&client) },
1003 0 : )
1004 0 : .await?;
1005 : }
1006 :
1007 0 : Ok::<(), anyhow::Error>(())
1008 0 : })?;
1009 :
1010 : Ok(())
1011 : }
1012 :
1013 : /// Apply SQL migrations of the RunInEachDatabase phase.
1014 : ///
1015 : /// May opt to not connect to databases that don't have any scheduled
1016 : /// operations. The function is concurrency-controlled with the provided
1017 : /// semaphore. The caller has to make sure the semaphore isn't exhausted.
1018 0 : async fn apply_spec_sql_db(
1019 0 : spec: Arc<ComputeSpec>,
1020 0 : url: Arc<Url>,
1021 0 : ctx: Arc<tokio::sync::RwLock<MutableApplyContext>>,
1022 0 : jwks_roles: Arc<HashSet<String>>,
1023 0 : concurrency_token: Arc<tokio::sync::Semaphore>,
1024 0 : db: DB,
1025 0 : ) -> Result<()> {
1026 0 : let _permit = concurrency_token.acquire().await?;
1027 :
1028 0 : let mut client_conn = None;
1029 :
1030 0 : for subphase in [
1031 0 : DeleteDBRoleReferences,
1032 0 : ChangeSchemaPerms,
1033 0 : HandleAnonExtension,
1034 : ] {
1035 0 : apply_operations(
1036 0 : spec.clone(),
1037 0 : ctx.clone(),
1038 0 : jwks_roles.clone(),
1039 0 : RunInEachDatabase {
1040 0 : db: db.clone(),
1041 0 : subphase,
1042 0 : },
1043 0 : // Only connect if apply_operation actually wants a connection.
1044 0 : // It's quite possible this database doesn't need any queries,
1045 0 : // so by not connecting we save time and effort connecting to
1046 0 : // that database.
1047 0 : || async {
1048 0 : if client_conn.is_none() {
1049 0 : let db_client = Self::get_maintenance_client(&url).await?;
1050 0 : client_conn.replace(db_client);
1051 0 : }
1052 0 : let client = client_conn.as_ref().unwrap();
1053 0 : Ok(client)
1054 0 : },
1055 0 : )
1056 0 : .await?;
1057 : }
1058 :
1059 0 : drop(client_conn);
1060 0 :
1061 0 : Ok::<(), anyhow::Error>(())
1062 0 : }
1063 :
1064 : /// Do initial configuration of the already started Postgres.
1065 0 : #[instrument(skip_all)]
1066 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1067 : // If connection fails,
1068 : // it may be the old node with `zenith_admin` superuser.
1069 : //
1070 : // In this case we need to connect with old `zenith_admin` name
1071 : // and create new user. We cannot simply rename connected user,
1072 : // but we can create a new one and grant it all privileges.
1073 : let mut url = self.connstr.clone();
1074 : url.query_pairs_mut()
1075 : .append_pair("application_name", "apply_config");
1076 :
1077 : let url = Arc::new(url);
1078 : let spec = Arc::new(
1079 : compute_state
1080 : .pspec
1081 : .as_ref()
1082 : .expect("spec must be set")
1083 : .spec
1084 : .clone(),
1085 : );
1086 :
1087 : // Choose how many concurrent connections to use for applying the spec changes.
1088 : // If the cluster is not currently Running we don't have to deal with user connections,
1089 : // and can thus use all `max_connections` connection slots. However, that's generally not
1090 : // very efficient, so we generally still limit it to a smaller number.
1091 : let max_concurrent_connections = if compute_state.status != ComputeStatus::Running {
1092 : // If the settings contain 'max_connections', use that as template
1093 : if let Some(config) = spec.cluster.settings.find("max_connections") {
1094 : config.parse::<usize>().ok()
1095 : } else {
1096 : // Otherwise, try to find the setting in the postgresql_conf string
1097 : spec.cluster
1098 : .postgresql_conf
1099 : .iter()
1100 0 : .flat_map(|conf| conf.split("\n"))
1101 0 : .filter_map(|line| {
1102 0 : if !line.contains("max_connections") {
1103 0 : return None;
1104 0 : }
1105 :
1106 0 : let (key, value) = line.split_once("=")?;
1107 0 : let key = key
1108 0 : .trim_start_matches(char::is_whitespace)
1109 0 : .trim_end_matches(char::is_whitespace);
1110 0 :
1111 0 : let value = value
1112 0 : .trim_start_matches(char::is_whitespace)
1113 0 : .trim_end_matches(char::is_whitespace);
1114 0 :
1115 0 : if key != "max_connections" {
1116 0 : return None;
1117 0 : }
1118 0 :
1119 0 : value.parse::<usize>().ok()
1120 0 : })
1121 : .next()
1122 : }
1123 : // If max_connections is present, use at most 1/3rd of that.
1124 : // When max_connections is lower than 30, try to use at least 10 connections, but
1125 : // never more than max_connections.
1126 0 : .map(|limit| match limit {
1127 0 : 0..10 => limit,
1128 0 : 10..30 => 10,
1129 0 : 30.. => limit / 3,
1130 0 : })
1131 : // If we didn't find max_connections, default to 10 concurrent connections.
1132 : .unwrap_or(10)
1133 : } else {
1134 : // state == Running
1135 : // Because the cluster is already in the Running state, we should assume users are
1136 : // already connected to the cluster, and high concurrency could negatively
1137 : // impact user connectivity. Therefore, we can limit concurrency to the number of
1138 : // reserved superuser connections, which users wouldn't be able to use anyway.
1139 : spec.cluster
1140 : .settings
1141 : .find("superuser_reserved_connections")
1142 : .iter()
1143 0 : .filter_map(|val| val.parse::<usize>().ok())
1144 0 : .map(|val| if val > 1 { val - 1 } else { 1 })
1145 : .last()
1146 : .unwrap_or(3)
1147 : };
1148 :
1149 : // Merge-apply spec & changes to PostgreSQL state.
1150 : self.apply_spec_sql(spec.clone(), url.clone(), max_concurrent_connections)?;
1151 :
1152 : if let Some(ref local_proxy) = &spec.clone().local_proxy_config {
1153 : info!("configuring local_proxy");
1154 : local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
1155 : }
1156 :
1157 : // Run migrations separately to not hold up cold starts
1158 0 : thread::spawn(move || {
1159 0 : let mut connstr = url.as_ref().clone();
1160 0 : connstr
1161 0 : .query_pairs_mut()
1162 0 : .append_pair("application_name", "migrations");
1163 :
1164 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
1165 0 : handle_migrations(&mut client).context("apply_config handle_migrations")
1166 0 : });
1167 :
1168 : Ok::<(), anyhow::Error>(())
1169 : }
1170 :
1171 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1172 : // `pg_ctl` for start / stop.
1173 0 : #[instrument(skip_all)]
1174 : fn pg_reload_conf(&self) -> Result<()> {
1175 : let pgctl_bin = Path::new(&self.pgbin).parent().unwrap().join("pg_ctl");
1176 : Command::new(pgctl_bin)
1177 : .args(["reload", "-D", &self.pgdata])
1178 : .output()
1179 : .expect("cannot run pg_ctl process");
1180 : Ok(())
1181 : }
1182 :
1183 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1184 : /// as it's used to reconfigure a previously started and configured Postgres node.
1185 0 : #[instrument(skip_all)]
1186 : pub fn reconfigure(&self) -> Result<()> {
1187 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1188 :
1189 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1190 : info!("tuning pgbouncer");
1191 :
1192 : let rt = tokio::runtime::Builder::new_current_thread()
1193 : .enable_all()
1194 : .build()
1195 : .expect("failed to create rt");
1196 :
1197 : // Spawn a thread to do the tuning,
1198 : // so that we don't block the main thread that starts Postgres.
1199 : let pgbouncer_settings = pgbouncer_settings.clone();
1200 0 : let _handle = thread::spawn(move || {
1201 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
1202 0 : if let Err(err) = res {
1203 0 : error!("error while tuning pgbouncer: {err:?}");
1204 0 : }
1205 0 : });
1206 : }
1207 :
1208 : if let Some(ref local_proxy) = spec.local_proxy_config {
1209 : info!("configuring local_proxy");
1210 :
1211 : // Spawn a thread to do the configuration,
1212 : // so that we don't block the main thread that starts Postgres.
1213 : let local_proxy = local_proxy.clone();
1214 0 : let _handle = Some(thread::spawn(move || {
1215 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1216 0 : error!("error while configuring local_proxy: {err:?}");
1217 0 : }
1218 0 : }));
1219 : }
1220 :
1221 : // Write new config
1222 : let pgdata_path = Path::new(&self.pgdata);
1223 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1224 : config::write_postgres_conf(&postgresql_conf_path, &spec, None)?;
1225 : // temporarily reset max_cluster_size in config
1226 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1227 : // creating new extensions, roles, etc...
1228 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1229 0 : self.pg_reload_conf()?;
1230 :
1231 0 : if spec.mode == ComputeMode::Primary {
1232 0 : let mut url = self.connstr.clone();
1233 0 : url.query_pairs_mut()
1234 0 : .append_pair("application_name", "apply_config");
1235 0 : let url = Arc::new(url);
1236 0 :
1237 0 : let spec = Arc::new(spec.clone());
1238 0 :
1239 0 : self.apply_spec_sql(spec, url, 1)?;
1240 0 : }
1241 :
1242 0 : Ok(())
1243 0 : })?;
1244 :
1245 : self.pg_reload_conf()?;
1246 :
1247 : let unknown_op = "unknown".to_string();
1248 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1249 : info!(
1250 : "finished reconfiguration of compute node for operation {}",
1251 : op_id
1252 : );
1253 :
1254 : Ok(())
1255 : }
1256 :
1257 0 : #[instrument(skip_all)]
1258 : pub fn start_compute(
1259 : &self,
1260 : extension_server_port: u16,
1261 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
1262 : let compute_state = self.state.lock().unwrap().clone();
1263 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1264 : info!(
1265 : "starting compute for project {}, operation {}, tenant {}, timeline {}",
1266 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
1267 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
1268 : pspec.tenant_id,
1269 : pspec.timeline_id,
1270 : );
1271 :
1272 : // tune pgbouncer
1273 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
1274 : info!("tuning pgbouncer");
1275 :
1276 : let rt = tokio::runtime::Builder::new_current_thread()
1277 : .enable_all()
1278 : .build()
1279 : .expect("failed to create rt");
1280 :
1281 : // Spawn a thread to do the tuning,
1282 : // so that we don't block the main thread that starts Postgres.
1283 : let pgbouncer_settings = pgbouncer_settings.clone();
1284 0 : let _handle = thread::spawn(move || {
1285 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
1286 0 : if let Err(err) = res {
1287 0 : error!("error while tuning pgbouncer: {err:?}");
1288 0 : }
1289 0 : });
1290 : }
1291 :
1292 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
1293 : info!("configuring local_proxy");
1294 :
1295 : // Spawn a thread to do the configuration,
1296 : // so that we don't block the main thread that starts Postgres.
1297 : let local_proxy = local_proxy.clone();
1298 0 : let _handle = thread::spawn(move || {
1299 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1300 0 : error!("error while configuring local_proxy: {err:?}");
1301 0 : }
1302 0 : });
1303 : }
1304 :
1305 : info!(
1306 : "start_compute spec.remote_extensions {:?}",
1307 : pspec.spec.remote_extensions
1308 : );
1309 :
1310 : // This part is sync, because we need to download
1311 : // remote shared_preload_libraries before postgres start (if any)
1312 : if let Some(remote_extensions) = &pspec.spec.remote_extensions {
1313 : // First, create control files for all availale extensions
1314 : extension_server::create_control_files(remote_extensions, &self.pgbin);
1315 :
1316 : let library_load_start_time = Utc::now();
1317 : let remote_ext_metrics = self.prepare_preload_libraries(&pspec.spec)?;
1318 :
1319 : let library_load_time = Utc::now()
1320 : .signed_duration_since(library_load_start_time)
1321 : .to_std()
1322 : .unwrap()
1323 : .as_millis() as u64;
1324 : let mut state = self.state.lock().unwrap();
1325 : state.metrics.load_ext_ms = library_load_time;
1326 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
1327 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
1328 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
1329 : info!(
1330 : "Loading shared_preload_libraries took {:?}ms",
1331 : library_load_time
1332 : );
1333 : info!("{:?}", remote_ext_metrics);
1334 : }
1335 :
1336 : self.prepare_pgdata(&compute_state, extension_server_port)?;
1337 :
1338 : let start_time = Utc::now();
1339 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
1340 :
1341 : let config_time = Utc::now();
1342 : if pspec.spec.mode == ComputeMode::Primary {
1343 : if !pspec.spec.skip_pg_catalog_updates {
1344 : let pgdata_path = Path::new(&self.pgdata);
1345 : // temporarily reset max_cluster_size in config
1346 : // to avoid the possibility of hitting the limit, while we are applying config:
1347 : // creating new extensions, roles, etc...
1348 : config::with_compute_ctl_tmp_override(
1349 : pgdata_path,
1350 : "neon.max_cluster_size=-1",
1351 0 : || {
1352 0 : self.pg_reload_conf()?;
1353 :
1354 0 : self.apply_config(&compute_state)?;
1355 :
1356 0 : Ok(())
1357 0 : },
1358 : )?;
1359 : self.pg_reload_conf()?;
1360 : }
1361 : self.post_apply_config()?;
1362 :
1363 : let connstr = self.connstr.clone();
1364 0 : thread::spawn(move || {
1365 0 : get_installed_extensions_sync(connstr).context("get_installed_extensions")
1366 0 : });
1367 : }
1368 :
1369 : let startup_end_time = Utc::now();
1370 : {
1371 : let mut state = self.state.lock().unwrap();
1372 : state.metrics.start_postgres_ms = config_time
1373 : .signed_duration_since(start_time)
1374 : .to_std()
1375 : .unwrap()
1376 : .as_millis() as u64;
1377 : state.metrics.config_ms = startup_end_time
1378 : .signed_duration_since(config_time)
1379 : .to_std()
1380 : .unwrap()
1381 : .as_millis() as u64;
1382 : state.metrics.total_startup_ms = startup_end_time
1383 : .signed_duration_since(compute_state.start_time)
1384 : .to_std()
1385 : .unwrap()
1386 : .as_millis() as u64;
1387 : }
1388 : self.set_status(ComputeStatus::Running);
1389 :
1390 : info!(
1391 : "finished configuration of compute for project {}",
1392 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
1393 : );
1394 :
1395 : // Log metrics so that we can search for slow operations in logs
1396 : let metrics = {
1397 : let state = self.state.lock().unwrap();
1398 : state.metrics.clone()
1399 : };
1400 : info!(?metrics, "compute start finished");
1401 :
1402 : Ok(pg_process)
1403 : }
1404 :
1405 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1406 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1407 0 : let mut state = self.state.lock().unwrap();
1408 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1409 0 : if last_active > state.last_active {
1410 0 : state.last_active = last_active;
1411 0 : debug!("set the last compute activity time to: {:?}", last_active);
1412 0 : }
1413 0 : }
1414 :
1415 : // Look for core dumps and collect backtraces.
1416 : //
1417 : // EKS worker nodes have following core dump settings:
1418 : // /proc/sys/kernel/core_pattern -> core
1419 : // /proc/sys/kernel/core_uses_pid -> 1
1420 : // ulimit -c -> unlimited
1421 : // which results in core dumps being written to postgres data directory as core.<pid>.
1422 : //
1423 : // Use that as a default location and pattern, except macos where core dumps are written
1424 : // to /cores/ directory by default.
1425 : //
1426 : // With default Linux settings, the core dump file is called just "core", so check for
1427 : // that too.
1428 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1429 0 : let core_dump_dir = match std::env::consts::OS {
1430 0 : "macos" => Path::new("/cores/"),
1431 0 : _ => Path::new(&self.pgdata),
1432 : };
1433 :
1434 : // Collect core dump paths if any
1435 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1436 0 : let files = fs::read_dir(core_dump_dir)?;
1437 0 : let cores = files.filter_map(|entry| {
1438 0 : let entry = entry.ok()?;
1439 :
1440 0 : let is_core_dump = match entry.file_name().to_str()? {
1441 0 : n if n.starts_with("core.") => true,
1442 0 : "core" => true,
1443 0 : _ => false,
1444 : };
1445 0 : if is_core_dump {
1446 0 : Some(entry.path())
1447 : } else {
1448 0 : None
1449 : }
1450 0 : });
1451 :
1452 : // Print backtrace for each core dump
1453 0 : for core_path in cores {
1454 0 : warn!(
1455 0 : "core dump found: {}, collecting backtrace",
1456 0 : core_path.display()
1457 : );
1458 :
1459 : // Try first with gdb
1460 0 : let backtrace = Command::new("gdb")
1461 0 : .args(["--batch", "-q", "-ex", "bt", &self.pgbin])
1462 0 : .arg(&core_path)
1463 0 : .output();
1464 :
1465 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1466 0 : let backtrace = match backtrace {
1467 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1468 0 : warn!("cannot find gdb, trying lldb");
1469 0 : Command::new("lldb")
1470 0 : .arg("-c")
1471 0 : .arg(&core_path)
1472 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1473 0 : .output()
1474 : }
1475 0 : _ => backtrace,
1476 0 : }?;
1477 :
1478 0 : warn!(
1479 0 : "core dump backtrace: {}",
1480 0 : String::from_utf8_lossy(&backtrace.stdout)
1481 : );
1482 0 : warn!(
1483 0 : "debugger stderr: {}",
1484 0 : String::from_utf8_lossy(&backtrace.stderr)
1485 : );
1486 : }
1487 :
1488 0 : Ok(())
1489 0 : }
1490 :
1491 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1492 0 : pub async fn collect_insights(&self) -> String {
1493 0 : let mut result_rows: Vec<String> = Vec::new();
1494 0 : let connect_result = tokio_postgres::connect(self.connstr.as_str(), NoTls).await;
1495 0 : let (client, connection) = connect_result.unwrap();
1496 0 : tokio::spawn(async move {
1497 0 : if let Err(e) = connection.await {
1498 0 : eprintln!("connection error: {}", e);
1499 0 : }
1500 0 : });
1501 0 : let result = client
1502 0 : .simple_query(
1503 0 : "SELECT
1504 0 : row_to_json(pg_stat_statements)
1505 0 : FROM
1506 0 : pg_stat_statements
1507 0 : WHERE
1508 0 : userid != 'cloud_admin'::regrole::oid
1509 0 : ORDER BY
1510 0 : (mean_exec_time + mean_plan_time) DESC
1511 0 : LIMIT 100",
1512 0 : )
1513 0 : .await;
1514 :
1515 0 : if let Ok(raw_rows) = result {
1516 0 : for message in raw_rows.iter() {
1517 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1518 0 : if let Some(json) = row.get(0) {
1519 0 : result_rows.push(json.to_string());
1520 0 : }
1521 0 : }
1522 : }
1523 :
1524 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1525 : } else {
1526 0 : "{{\"pg_stat_statements\": []}}".to_string()
1527 : }
1528 0 : }
1529 :
1530 : // download an archive, unzip and place files in correct locations
1531 0 : pub async fn download_extension(
1532 0 : &self,
1533 0 : real_ext_name: String,
1534 0 : ext_path: RemotePath,
1535 0 : ) -> Result<u64, DownloadError> {
1536 0 : let ext_remote_storage =
1537 0 : self.ext_remote_storage
1538 0 : .as_ref()
1539 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1540 0 : "Remote extensions storage is not configured",
1541 0 : )))?;
1542 :
1543 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1544 0 :
1545 0 : let mut first_try = false;
1546 0 : if !self
1547 0 : .ext_download_progress
1548 0 : .read()
1549 0 : .expect("lock err")
1550 0 : .contains_key(ext_archive_name)
1551 0 : {
1552 0 : self.ext_download_progress
1553 0 : .write()
1554 0 : .expect("lock err")
1555 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1556 0 : first_try = true;
1557 0 : }
1558 0 : let (download_start, download_completed) =
1559 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1560 0 : let start_time_delta = Utc::now()
1561 0 : .signed_duration_since(download_start)
1562 0 : .to_std()
1563 0 : .unwrap()
1564 0 : .as_millis() as u64;
1565 :
1566 : // how long to wait for extension download if it was started by another process
1567 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1568 :
1569 0 : if download_completed {
1570 0 : info!("extension already downloaded, skipping re-download");
1571 0 : return Ok(0);
1572 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1573 0 : info!("download {ext_archive_name} already started by another process, hanging untill completion or timeout");
1574 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1575 : loop {
1576 0 : info!("waiting for download");
1577 0 : interval.tick().await;
1578 0 : let (_, download_completed_now) =
1579 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1580 0 : if download_completed_now {
1581 0 : info!("download finished by whoever else downloaded it");
1582 0 : return Ok(0);
1583 0 : }
1584 : }
1585 : // NOTE: the above loop will get terminated
1586 : // based on the timeout of the download function
1587 0 : }
1588 0 :
1589 0 : // if extension hasn't been downloaded before or the previous
1590 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1591 0 : // then we try to download it here
1592 0 : info!("downloading new extension {ext_archive_name}");
1593 :
1594 0 : let download_size = extension_server::download_extension(
1595 0 : &real_ext_name,
1596 0 : &ext_path,
1597 0 : ext_remote_storage,
1598 0 : &self.pgbin,
1599 0 : )
1600 0 : .await
1601 0 : .map_err(DownloadError::Other);
1602 0 :
1603 0 : if download_size.is_ok() {
1604 0 : self.ext_download_progress
1605 0 : .write()
1606 0 : .expect("bad lock")
1607 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1608 0 : }
1609 :
1610 0 : download_size
1611 0 : }
1612 :
1613 0 : pub async fn set_role_grants(
1614 0 : &self,
1615 0 : db_name: &PgIdent,
1616 0 : schema_name: &PgIdent,
1617 0 : privileges: &[Privilege],
1618 0 : role_name: &PgIdent,
1619 0 : ) -> Result<()> {
1620 : use tokio_postgres::config::Config;
1621 : use tokio_postgres::NoTls;
1622 :
1623 0 : let mut conf = Config::from_str(self.connstr.as_str()).unwrap();
1624 0 : conf.dbname(db_name);
1625 :
1626 0 : let (db_client, conn) = conf
1627 0 : .connect(NoTls)
1628 0 : .await
1629 0 : .context("Failed to connect to the database")?;
1630 0 : tokio::spawn(conn);
1631 0 :
1632 0 : // TODO: support other types of grants apart from schemas?
1633 0 : let query = format!(
1634 0 : "GRANT {} ON SCHEMA {} TO {}",
1635 0 : privileges
1636 0 : .iter()
1637 0 : // should not be quoted as it's part of the command.
1638 0 : // is already sanitized so it's ok
1639 0 : .map(|p| p.as_str())
1640 0 : .collect::<Vec<&'static str>>()
1641 0 : .join(", "),
1642 0 : // quote the schema and role name as identifiers to sanitize them.
1643 0 : schema_name.pg_quote(),
1644 0 : role_name.pg_quote(),
1645 0 : );
1646 0 : db_client
1647 0 : .simple_query(&query)
1648 0 : .await
1649 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1650 :
1651 0 : Ok(())
1652 0 : }
1653 :
1654 0 : pub async fn install_extension(
1655 0 : &self,
1656 0 : ext_name: &PgIdent,
1657 0 : db_name: &PgIdent,
1658 0 : ext_version: ExtVersion,
1659 0 : ) -> Result<ExtVersion> {
1660 : use tokio_postgres::config::Config;
1661 : use tokio_postgres::NoTls;
1662 :
1663 0 : let mut conf = Config::from_str(self.connstr.as_str()).unwrap();
1664 0 : conf.dbname(db_name);
1665 :
1666 0 : let (db_client, conn) = conf
1667 0 : .connect(NoTls)
1668 0 : .await
1669 0 : .context("Failed to connect to the database")?;
1670 0 : tokio::spawn(conn);
1671 0 :
1672 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
1673 0 : let version: Option<ExtVersion> = db_client
1674 0 : .query_opt(version_query, &[&ext_name])
1675 0 : .await
1676 0 : .with_context(|| format!("Failed to execute query: {}", version_query))?
1677 0 : .map(|row| row.get(0));
1678 0 :
1679 0 : // sanitize the inputs as postgres idents.
1680 0 : let ext_name: String = ext_name.pg_quote();
1681 0 : let quoted_version: String = ext_version.pg_quote();
1682 :
1683 0 : if let Some(installed_version) = version {
1684 0 : if installed_version == ext_version {
1685 0 : return Ok(installed_version);
1686 0 : }
1687 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
1688 0 : db_client
1689 0 : .simple_query(&query)
1690 0 : .await
1691 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1692 : } else {
1693 0 : let query =
1694 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
1695 0 : db_client
1696 0 : .simple_query(&query)
1697 0 : .await
1698 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1699 : }
1700 :
1701 0 : Ok(ext_version)
1702 0 : }
1703 :
1704 : #[tokio::main]
1705 0 : pub async fn prepare_preload_libraries(
1706 0 : &self,
1707 0 : spec: &ComputeSpec,
1708 0 : ) -> Result<RemoteExtensionMetrics> {
1709 0 : if self.ext_remote_storage.is_none() {
1710 0 : return Ok(RemoteExtensionMetrics {
1711 0 : num_ext_downloaded: 0,
1712 0 : largest_ext_size: 0,
1713 0 : total_ext_download_size: 0,
1714 0 : });
1715 0 : }
1716 0 : let remote_extensions = spec
1717 0 : .remote_extensions
1718 0 : .as_ref()
1719 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
1720 0 :
1721 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
1722 0 : let mut libs_vec = Vec::new();
1723 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
1724 0 : libs_vec = libs
1725 0 : .split(&[',', '\'', ' '])
1726 0 : .filter(|s| *s != "neon" && !s.is_empty())
1727 0 : .map(str::to_string)
1728 0 : .collect();
1729 0 : }
1730 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
1731 0 :
1732 0 : // that is used in neon_local and python tests
1733 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
1734 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
1735 0 : let mut shared_preload_libraries_line = "";
1736 0 : for line in conf_lines {
1737 0 : if line.starts_with("shared_preload_libraries") {
1738 0 : shared_preload_libraries_line = line;
1739 0 : }
1740 0 : }
1741 0 : let mut preload_libs_vec = Vec::new();
1742 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
1743 0 : preload_libs_vec = libs
1744 0 : .split(&[',', '\'', ' '])
1745 0 : .filter(|s| *s != "neon" && !s.is_empty())
1746 0 : .map(str::to_string)
1747 0 : .collect();
1748 0 : }
1749 0 : libs_vec.extend(preload_libs_vec);
1750 0 : }
1751 0 :
1752 0 : // Don't try to download libraries that are not in the index.
1753 0 : // Assume that they are already present locally.
1754 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
1755 0 :
1756 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
1757 0 :
1758 0 : let mut download_tasks = Vec::new();
1759 0 : for library in &libs_vec {
1760 0 : let (ext_name, ext_path) =
1761 0 : remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
1762 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
1763 0 : }
1764 0 : let results = join_all(download_tasks).await;
1765 0 :
1766 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
1767 0 : num_ext_downloaded: 0,
1768 0 : largest_ext_size: 0,
1769 0 : total_ext_download_size: 0,
1770 0 : };
1771 0 : for result in results {
1772 0 : let download_size = match result {
1773 0 : Ok(res) => {
1774 0 : remote_ext_metrics.num_ext_downloaded += 1;
1775 0 : res
1776 0 : }
1777 0 : Err(err) => {
1778 0 : // if we failed to download an extension, we don't want to fail the whole
1779 0 : // process, but we do want to log the error
1780 0 : error!("Failed to download extension: {}", err);
1781 0 : 0
1782 0 : }
1783 0 : };
1784 0 :
1785 0 : remote_ext_metrics.largest_ext_size =
1786 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
1787 0 : remote_ext_metrics.total_ext_download_size += download_size;
1788 0 : }
1789 0 : Ok(remote_ext_metrics)
1790 0 : }
1791 :
1792 : /// Waits until current thread receives a state changed notification and
1793 : /// the pageserver connection strings has changed.
1794 : ///
1795 : /// The operation will time out after a specified duration.
1796 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
1797 0 : let state = self.state.lock().unwrap();
1798 0 : let old_pageserver_connstr = state
1799 0 : .pspec
1800 0 : .as_ref()
1801 0 : .expect("spec must be set")
1802 0 : .pageserver_connstr
1803 0 : .clone();
1804 0 : let mut unchanged = true;
1805 0 : let _ = self
1806 0 : .state_changed
1807 0 : .wait_timeout_while(state, duration, |s| {
1808 0 : let pageserver_connstr = &s
1809 0 : .pspec
1810 0 : .as_ref()
1811 0 : .expect("spec must be set")
1812 0 : .pageserver_connstr;
1813 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
1814 0 : unchanged
1815 0 : })
1816 0 : .unwrap();
1817 0 : if !unchanged {
1818 0 : info!("Pageserver config changed");
1819 0 : }
1820 0 : }
1821 : }
1822 :
1823 0 : pub fn forward_termination_signal() {
1824 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
1825 0 : if ss_pid != 0 {
1826 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
1827 0 : kill(ss_pid, Signal::SIGTERM).ok();
1828 0 : }
1829 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
1830 0 : if pg_pid != 0 {
1831 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
1832 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
1833 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
1834 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
1835 0 : kill(pg_pid, Signal::SIGINT).ok();
1836 0 : }
1837 0 : }
|
