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::time::Duration;
13 : use std::time::Instant;
14 :
15 : use anyhow::{Context, Result};
16 : use chrono::{DateTime, Utc};
17 : use compute_api::spec::{Database, PgIdent, Role};
18 : use futures::future::join_all;
19 : use futures::stream::FuturesUnordered;
20 : use futures::StreamExt;
21 : use nix::unistd::Pid;
22 : use postgres;
23 : use postgres::error::SqlState;
24 : use postgres::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 :
38 : use crate::installed_extensions::get_installed_extensions;
39 : use crate::local_proxy;
40 : use crate::pg_helpers::*;
41 : use crate::spec::*;
42 : use crate::spec_apply::ApplySpecPhase::{
43 : CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreateSchemaNeon,
44 : CreateSuperUser, DropInvalidDatabases, DropRoles, FinalizeDropLogicalSubscriptions,
45 : HandleNeonExtension, HandleOtherExtensions, RenameAndDeleteDatabases, RenameRoles,
46 : RunInEachDatabase,
47 : };
48 : use crate::spec_apply::PerDatabasePhase;
49 : use crate::spec_apply::PerDatabasePhase::{
50 : ChangeSchemaPerms, DeleteDBRoleReferences, DropLogicalSubscriptions, HandleAnonExtension,
51 : };
52 : use crate::spec_apply::{apply_operations, MutableApplyContext, DB};
53 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
54 : use crate::{config, extension_server};
55 :
56 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
57 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
58 :
59 : /// Compute node info shared across several `compute_ctl` threads.
60 : pub struct ComputeNode {
61 : /// The ID of the compute
62 : pub compute_id: String,
63 : // Url type maintains proper escaping
64 : pub connstr: url::Url,
65 : // We connect to Postgres from many different places, so build configs once
66 : // and reuse them where needed.
67 : pub conn_conf: postgres::config::Config,
68 : pub tokio_conn_conf: tokio_postgres::config::Config,
69 : pub pgdata: String,
70 : pub pgbin: String,
71 : pub pgversion: String,
72 : /// We should only allow live re- / configuration of the compute node if
73 : /// it uses 'pull model', i.e. it can go to control-plane and fetch
74 : /// the latest configuration. Otherwise, there could be a case:
75 : /// - we start compute with some spec provided as argument
76 : /// - we push new spec and it does reconfiguration
77 : /// - but then something happens and compute pod / VM is destroyed,
78 : /// so k8s controller starts it again with the **old** spec
79 : ///
80 : /// and the same for empty computes:
81 : /// - we started compute without any spec
82 : /// - we push spec and it does configuration
83 : /// - but then it is restarted without any spec again
84 : pub live_config_allowed: bool,
85 : /// The port that the compute's external HTTP server listens on
86 : pub external_http_port: u16,
87 : /// The port that the compute's internal HTTP server listens on
88 : pub internal_http_port: u16,
89 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
90 : /// To allow HTTP API server to serving status requests, while configuration
91 : /// is in progress, lock should be held only for short periods of time to do
92 : /// read/write, not the whole configuration process.
93 : pub state: Mutex<ComputeState>,
94 : /// `Condvar` to allow notifying waiters about state changes.
95 : pub state_changed: Condvar,
96 : /// the address of extension storage proxy gateway
97 : pub ext_remote_storage: Option<String>,
98 : // key: ext_archive_name, value: started download time, download_completed?
99 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
100 : pub build_tag: String,
101 : }
102 :
103 : // store some metrics about download size that might impact startup time
104 : #[derive(Clone, Debug)]
105 : pub struct RemoteExtensionMetrics {
106 : num_ext_downloaded: u64,
107 : largest_ext_size: u64,
108 : total_ext_download_size: u64,
109 : }
110 :
111 : #[derive(Clone, Debug)]
112 : pub struct ComputeState {
113 : pub start_time: DateTime<Utc>,
114 : pub status: ComputeStatus,
115 : /// Timestamp of the last Postgres activity. It could be `None` if
116 : /// compute wasn't used since start.
117 : pub last_active: Option<DateTime<Utc>>,
118 : pub error: Option<String>,
119 : pub pspec: Option<ParsedSpec>,
120 : pub metrics: ComputeMetrics,
121 : }
122 :
123 : impl ComputeState {
124 0 : pub fn new() -> Self {
125 0 : Self {
126 0 : start_time: Utc::now(),
127 0 : status: ComputeStatus::Empty,
128 0 : last_active: None,
129 0 : error: None,
130 0 : pspec: None,
131 0 : metrics: ComputeMetrics::default(),
132 0 : }
133 0 : }
134 :
135 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
136 0 : let prev = self.status;
137 0 : info!("Changing compute status from {} to {}", prev, status);
138 0 : self.status = status;
139 0 : state_changed.notify_all();
140 0 : }
141 :
142 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
143 0 : self.error = Some(format!("{err:?}"));
144 0 : self.set_status(ComputeStatus::Failed, state_changed);
145 0 : }
146 : }
147 :
148 : impl Default for ComputeState {
149 0 : fn default() -> Self {
150 0 : Self::new()
151 0 : }
152 : }
153 :
154 : #[derive(Clone, Debug)]
155 : pub struct ParsedSpec {
156 : pub spec: ComputeSpec,
157 : pub tenant_id: TenantId,
158 : pub timeline_id: TimelineId,
159 : pub pageserver_connstr: String,
160 : pub safekeeper_connstrings: Vec<String>,
161 : pub storage_auth_token: Option<String>,
162 : }
163 :
164 : impl TryFrom<ComputeSpec> for ParsedSpec {
165 : type Error = String;
166 0 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
167 : // Extract the options from the spec file that are needed to connect to
168 : // the storage system.
169 : //
170 : // For backwards-compatibility, the top-level fields in the spec file
171 : // may be empty. In that case, we need to dig them from the GUCs in the
172 : // cluster.settings field.
173 0 : let pageserver_connstr = spec
174 0 : .pageserver_connstring
175 0 : .clone()
176 0 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
177 0 : .ok_or("pageserver connstr should be provided")?;
178 0 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
179 0 : if matches!(spec.mode, ComputeMode::Primary) {
180 0 : spec.cluster
181 0 : .settings
182 0 : .find("neon.safekeepers")
183 0 : .ok_or("safekeeper connstrings should be provided")?
184 0 : .split(',')
185 0 : .map(|str| str.to_string())
186 0 : .collect()
187 : } else {
188 0 : vec![]
189 : }
190 : } else {
191 0 : spec.safekeeper_connstrings.clone()
192 : };
193 0 : let storage_auth_token = spec.storage_auth_token.clone();
194 0 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
195 0 : tenant_id
196 : } else {
197 0 : spec.cluster
198 0 : .settings
199 0 : .find("neon.tenant_id")
200 0 : .ok_or("tenant id should be provided")
201 0 : .map(|s| TenantId::from_str(&s))?
202 0 : .or(Err("invalid tenant id"))?
203 : };
204 0 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
205 0 : timeline_id
206 : } else {
207 0 : spec.cluster
208 0 : .settings
209 0 : .find("neon.timeline_id")
210 0 : .ok_or("timeline id should be provided")
211 0 : .map(|s| TimelineId::from_str(&s))?
212 0 : .or(Err("invalid timeline id"))?
213 : };
214 :
215 0 : Ok(ParsedSpec {
216 0 : spec,
217 0 : pageserver_connstr,
218 0 : safekeeper_connstrings,
219 0 : storage_auth_token,
220 0 : tenant_id,
221 0 : timeline_id,
222 0 : })
223 0 : }
224 : }
225 :
226 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
227 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
228 : ///
229 : /// This function should be used to start postgres, as it will start it in the
230 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
231 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
232 : /// creates it during the sysinit phase of its inittab.
233 0 : fn maybe_cgexec(cmd: &str) -> Command {
234 0 : // The cplane sets this env var for autoscaling computes.
235 0 : // use `var_os` so we don't have to worry about the variable being valid
236 0 : // unicode. Should never be an concern . . . but just in case
237 0 : if env::var_os("AUTOSCALING").is_some() {
238 0 : let mut command = Command::new("cgexec");
239 0 : command.args(["-g", "memory:neon-postgres"]);
240 0 : command.arg(cmd);
241 0 : command
242 : } else {
243 0 : Command::new(cmd)
244 : }
245 0 : }
246 :
247 0 : pub(crate) fn construct_superuser_query(spec: &ComputeSpec) -> String {
248 0 : let roles = spec
249 0 : .cluster
250 0 : .roles
251 0 : .iter()
252 0 : .map(|r| escape_literal(&r.name))
253 0 : .collect::<Vec<_>>();
254 0 :
255 0 : let dbs = spec
256 0 : .cluster
257 0 : .databases
258 0 : .iter()
259 0 : .map(|db| escape_literal(&db.name))
260 0 : .collect::<Vec<_>>();
261 :
262 0 : let roles_decl = if roles.is_empty() {
263 0 : String::from("roles text[] := NULL;")
264 : } else {
265 0 : format!(
266 0 : r#"
267 0 : roles text[] := ARRAY(SELECT rolname
268 0 : FROM pg_catalog.pg_roles
269 0 : WHERE rolname IN ({}));"#,
270 0 : roles.join(", ")
271 0 : )
272 : };
273 :
274 0 : let database_decl = if dbs.is_empty() {
275 0 : String::from("dbs text[] := NULL;")
276 : } else {
277 0 : format!(
278 0 : r#"
279 0 : dbs text[] := ARRAY(SELECT datname
280 0 : FROM pg_catalog.pg_database
281 0 : WHERE datname IN ({}));"#,
282 0 : dbs.join(", ")
283 0 : )
284 : };
285 :
286 : // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on all databases
287 : // (see https://www.postgresql.org/docs/current/ddl-priv.html)
288 0 : let query = format!(
289 0 : r#"
290 0 : DO $$
291 0 : DECLARE
292 0 : r text;
293 0 : {}
294 0 : {}
295 0 : BEGIN
296 0 : IF NOT EXISTS (
297 0 : SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
298 0 : THEN
299 0 : CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
300 0 : IF array_length(roles, 1) IS NOT NULL THEN
301 0 : EXECUTE format('GRANT neon_superuser TO %s',
302 0 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
303 0 : FOREACH r IN ARRAY roles LOOP
304 0 : EXECUTE format('ALTER ROLE %s CREATEROLE CREATEDB', quote_ident(r));
305 0 : END LOOP;
306 0 : END IF;
307 0 : IF array_length(dbs, 1) IS NOT NULL THEN
308 0 : EXECUTE format('GRANT ALL PRIVILEGES ON DATABASE %s TO neon_superuser',
309 0 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(dbs) as x), ', '));
310 0 : END IF;
311 0 : END IF;
312 0 : END
313 0 : $$;"#,
314 0 : roles_decl, database_decl,
315 0 : );
316 0 :
317 0 : query
318 0 : }
319 :
320 : impl ComputeNode {
321 : /// Check that compute node has corresponding feature enabled.
322 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
323 0 : let state = self.state.lock().unwrap();
324 :
325 0 : if let Some(s) = state.pspec.as_ref() {
326 0 : s.spec.features.contains(&feature)
327 : } else {
328 0 : false
329 : }
330 0 : }
331 :
332 0 : pub fn set_status(&self, status: ComputeStatus) {
333 0 : let mut state = self.state.lock().unwrap();
334 0 : state.set_status(status, &self.state_changed);
335 0 : }
336 :
337 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
338 0 : let mut state = self.state.lock().unwrap();
339 0 : state.set_failed_status(err, &self.state_changed);
340 0 : }
341 :
342 0 : pub fn get_status(&self) -> ComputeStatus {
343 0 : self.state.lock().unwrap().status
344 0 : }
345 :
346 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
347 0 : self.state
348 0 : .lock()
349 0 : .unwrap()
350 0 : .pspec
351 0 : .as_ref()
352 0 : .map(|s| s.timeline_id)
353 0 : }
354 :
355 : // Remove `pgdata` directory and create it again with right permissions.
356 0 : fn create_pgdata(&self) -> Result<()> {
357 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
358 0 : // If it is something different then create_dir() will error out anyway.
359 0 : let _ok = fs::remove_dir_all(&self.pgdata);
360 0 : fs::create_dir(&self.pgdata)?;
361 0 : fs::set_permissions(&self.pgdata, fs::Permissions::from_mode(0o700))?;
362 :
363 0 : Ok(())
364 0 : }
365 :
366 : // Get basebackup from the libpq connection to pageserver using `connstr` and
367 : // unarchive it to `pgdata` directory overriding all its previous content.
368 : #[instrument(skip_all, fields(%lsn))]
369 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
370 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
371 : let start_time = Instant::now();
372 :
373 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
374 : let mut config = postgres::Config::from_str(shard0_connstr)?;
375 :
376 : // Use the storage auth token from the config file, if given.
377 : // Note: this overrides any password set in the connection string.
378 : if let Some(storage_auth_token) = &spec.storage_auth_token {
379 : info!("Got storage auth token from spec file");
380 : config.password(storage_auth_token);
381 : } else {
382 : info!("Storage auth token not set");
383 : }
384 :
385 : // Connect to pageserver
386 : let mut client = config.connect(NoTls)?;
387 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
388 :
389 : let basebackup_cmd = match lsn {
390 : Lsn(0) => {
391 : if spec.spec.mode != ComputeMode::Primary {
392 : format!(
393 : "basebackup {} {} --gzip --replica",
394 : spec.tenant_id, spec.timeline_id
395 : )
396 : } else {
397 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
398 : }
399 : }
400 : _ => {
401 : if spec.spec.mode != ComputeMode::Primary {
402 : format!(
403 : "basebackup {} {} {} --gzip --replica",
404 : spec.tenant_id, spec.timeline_id, lsn
405 : )
406 : } else {
407 : format!(
408 : "basebackup {} {} {} --gzip",
409 : spec.tenant_id, spec.timeline_id, lsn
410 : )
411 : }
412 : }
413 : };
414 :
415 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
416 : let mut measured_reader = MeasuredReader::new(copyreader);
417 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
418 :
419 : // Read the archive directly from the `CopyOutReader`
420 : //
421 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
422 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
423 : // sends an Error after finishing the tarball, we will not notice it.
424 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
425 : ar.set_ignore_zeros(true);
426 : ar.unpack(&self.pgdata)?;
427 :
428 : // Report metrics
429 : let mut state = self.state.lock().unwrap();
430 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
431 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
432 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
433 : Ok(())
434 : }
435 :
436 : // Gets the basebackup in a retry loop
437 : #[instrument(skip_all, fields(%lsn))]
438 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
439 : let mut retry_period_ms = 500.0;
440 : let mut attempts = 0;
441 : const DEFAULT_ATTEMPTS: u16 = 10;
442 : #[cfg(feature = "testing")]
443 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
444 : u16::from_str(&v).unwrap()
445 : } else {
446 : DEFAULT_ATTEMPTS
447 : };
448 : #[cfg(not(feature = "testing"))]
449 : let max_attempts = DEFAULT_ATTEMPTS;
450 : loop {
451 : let result = self.try_get_basebackup(compute_state, lsn);
452 : match result {
453 : Ok(_) => {
454 : return result;
455 : }
456 : Err(ref e) if attempts < max_attempts => {
457 : warn!(
458 : "Failed to get basebackup: {} (attempt {}/{})",
459 : e, attempts, max_attempts
460 : );
461 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
462 : retry_period_ms *= 1.5;
463 : }
464 : Err(_) => {
465 : return result;
466 : }
467 : }
468 : attempts += 1;
469 : }
470 : }
471 :
472 0 : pub async fn check_safekeepers_synced_async(
473 0 : &self,
474 0 : compute_state: &ComputeState,
475 0 : ) -> Result<Option<Lsn>> {
476 0 : // Construct a connection config for each safekeeper
477 0 : let pspec: ParsedSpec = compute_state
478 0 : .pspec
479 0 : .as_ref()
480 0 : .expect("spec must be set")
481 0 : .clone();
482 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
483 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
484 0 : // Format connstr
485 0 : let id = connstr.clone();
486 0 : let connstr = format!("postgresql://no_user@{}", connstr);
487 0 : let options = format!(
488 0 : "-c timeline_id={} tenant_id={}",
489 0 : pspec.timeline_id, pspec.tenant_id
490 0 : );
491 0 :
492 0 : // Construct client
493 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
494 0 : config.options(&options);
495 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
496 0 : config.password(storage_auth_token);
497 0 : }
498 :
499 0 : (id, config)
500 0 : });
501 0 :
502 0 : // Create task set to query all safekeepers
503 0 : let mut tasks = FuturesUnordered::new();
504 0 : let quorum = sk_configs.len() / 2 + 1;
505 0 : for (id, config) in sk_configs {
506 0 : let timeout = tokio::time::Duration::from_millis(100);
507 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
508 0 : tasks.push(tokio::spawn(task));
509 0 : }
510 :
511 : // Get a quorum of responses or errors
512 0 : let mut responses = Vec::new();
513 0 : let mut join_errors = Vec::new();
514 0 : let mut task_errors = Vec::new();
515 0 : let mut timeout_errors = Vec::new();
516 0 : while let Some(response) = tasks.next().await {
517 0 : match response {
518 0 : Ok(Ok(Ok(r))) => responses.push(r),
519 0 : Ok(Ok(Err(e))) => task_errors.push(e),
520 0 : Ok(Err(e)) => timeout_errors.push(e),
521 0 : Err(e) => join_errors.push(e),
522 : };
523 0 : if responses.len() >= quorum {
524 0 : break;
525 0 : }
526 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
527 0 : break;
528 0 : }
529 : }
530 :
531 : // In case of error, log and fail the check, but don't crash.
532 : // We're playing it safe because these errors could be transient
533 : // and we don't yet retry. Also being careful here allows us to
534 : // be backwards compatible with safekeepers that don't have the
535 : // TIMELINE_STATUS API yet.
536 0 : if responses.len() < quorum {
537 0 : error!(
538 0 : "failed sync safekeepers check {:?} {:?} {:?}",
539 : join_errors, task_errors, timeout_errors
540 : );
541 0 : return Ok(None);
542 0 : }
543 0 :
544 0 : Ok(check_if_synced(responses))
545 0 : }
546 :
547 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
548 : // in one roundtrip. If not, we should do a full sync_safekeepers.
549 0 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
550 0 : let start_time = Utc::now();
551 0 :
552 0 : let rt = tokio::runtime::Handle::current();
553 0 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
554 0 :
555 0 : // Record runtime
556 0 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
557 0 : .signed_duration_since(start_time)
558 0 : .to_std()
559 0 : .unwrap()
560 0 : .as_millis() as u64;
561 0 : result
562 0 : }
563 :
564 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
565 : // and return the reported LSN back to the caller.
566 : #[instrument(skip_all)]
567 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
568 : let start_time = Utc::now();
569 :
570 : let mut sync_handle = maybe_cgexec(&self.pgbin)
571 : .args(["--sync-safekeepers"])
572 : .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
573 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
574 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
575 : } else {
576 : vec![]
577 : })
578 : .stdout(Stdio::piped())
579 : .stderr(Stdio::piped())
580 : .spawn()
581 : .expect("postgres --sync-safekeepers failed to start");
582 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
583 :
584 : // `postgres --sync-safekeepers` will print all log output to stderr and
585 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
586 : // will be collected in a child thread.
587 : let stderr = sync_handle
588 : .stderr
589 : .take()
590 : .expect("stderr should be captured");
591 : let logs_handle = handle_postgres_logs(stderr);
592 :
593 : let sync_output = sync_handle
594 : .wait_with_output()
595 : .expect("postgres --sync-safekeepers failed");
596 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
597 :
598 : // Process has exited, so we can join the logs thread.
599 : let _ = tokio::runtime::Handle::current()
600 : .block_on(logs_handle)
601 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
602 :
603 : if !sync_output.status.success() {
604 : anyhow::bail!(
605 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
606 : sync_output.status,
607 : String::from_utf8(sync_output.stdout)
608 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
609 : );
610 : }
611 :
612 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
613 : .signed_duration_since(start_time)
614 : .to_std()
615 : .unwrap()
616 : .as_millis() as u64;
617 :
618 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
619 :
620 : Ok(lsn)
621 : }
622 :
623 : /// Do all the preparations like PGDATA directory creation, configuration,
624 : /// safekeepers sync, basebackup, etc.
625 : #[instrument(skip_all)]
626 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
627 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
628 : let spec = &pspec.spec;
629 : let pgdata_path = Path::new(&self.pgdata);
630 :
631 : // Remove/create an empty pgdata directory and put configuration there.
632 : self.create_pgdata()?;
633 : config::write_postgres_conf(
634 : &pgdata_path.join("postgresql.conf"),
635 : &pspec.spec,
636 : self.internal_http_port,
637 : )?;
638 :
639 : // Syncing safekeepers is only safe with primary nodes: if a primary
640 : // is already connected it will be kicked out, so a secondary (standby)
641 : // cannot sync safekeepers.
642 : let lsn = match spec.mode {
643 : ComputeMode::Primary => {
644 : info!("checking if safekeepers are synced");
645 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
646 : lsn
647 : } else {
648 : info!("starting safekeepers syncing");
649 : self.sync_safekeepers(pspec.storage_auth_token.clone())
650 0 : .with_context(|| "failed to sync safekeepers")?
651 : };
652 : info!("safekeepers synced at LSN {}", lsn);
653 : lsn
654 : }
655 : ComputeMode::Static(lsn) => {
656 : info!("Starting read-only node at static LSN {}", lsn);
657 : lsn
658 : }
659 : ComputeMode::Replica => {
660 : info!("Initializing standby from latest Pageserver LSN");
661 : Lsn(0)
662 : }
663 : };
664 :
665 : info!(
666 : "getting basebackup@{} from pageserver {}",
667 : lsn, &pspec.pageserver_connstr
668 : );
669 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
670 0 : format!(
671 0 : "failed to get basebackup@{} from pageserver {}",
672 0 : lsn, &pspec.pageserver_connstr
673 0 : )
674 0 : })?;
675 :
676 : // Update pg_hba.conf received with basebackup.
677 : update_pg_hba(pgdata_path)?;
678 :
679 : // Place pg_dynshmem under /dev/shm. This allows us to use
680 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
681 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
682 : //
683 : // Why on earth don't we just stick to the 'posix' default, you might
684 : // ask. It turns out that making large allocations with 'posix' doesn't
685 : // work very well with autoscaling. The behavior we want is that:
686 : //
687 : // 1. You can make large DSM allocations, larger than the current RAM
688 : // size of the VM, without errors
689 : //
690 : // 2. If the allocated memory is really used, the VM is scaled up
691 : // automatically to accommodate that
692 : //
693 : // We try to make that possible by having swap in the VM. But with the
694 : // default 'posix' DSM implementation, we fail step 1, even when there's
695 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
696 : // the shmem segment, which is really just a file in /dev/shm in Linux,
697 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
698 : // than available RAM.
699 : //
700 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
701 : // the Postgres 'mmap' DSM implementation doesn't use
702 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
703 : // fill the file with zeros. It's weird that that differs between
704 : // 'posix' and 'mmap', but we take advantage of it. When the file is
705 : // filled slowly with write(2), the kernel allows it to grow larger, as
706 : // long as there's swap available.
707 : //
708 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
709 : // segment to be larger than currently available RAM. But because we
710 : // don't want to store it on a real file, which the kernel would try to
711 : // flush to disk, so symlink pg_dynshm to /dev/shm.
712 : //
713 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
714 : // control plane control that option. If 'mmap' is not used, this
715 : // symlink doesn't affect anything.
716 : //
717 : // See https://github.com/neondatabase/autoscaling/issues/800
718 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
719 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
720 :
721 : match spec.mode {
722 : ComputeMode::Primary => {}
723 : ComputeMode::Replica | ComputeMode::Static(..) => {
724 : add_standby_signal(pgdata_path)?;
725 : }
726 : }
727 :
728 : Ok(())
729 : }
730 :
731 : /// Start and stop a postgres process to warm up the VM for startup.
732 0 : pub fn prewarm_postgres(&self) -> Result<()> {
733 0 : info!("prewarming");
734 :
735 : // Create pgdata
736 0 : let pgdata = &format!("{}.warmup", self.pgdata);
737 0 : create_pgdata(pgdata)?;
738 :
739 : // Run initdb to completion
740 0 : info!("running initdb");
741 0 : let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
742 0 : Command::new(initdb_bin)
743 0 : .args(["--pgdata", pgdata])
744 0 : .output()
745 0 : .expect("cannot start initdb process");
746 :
747 : // Write conf
748 : use std::io::Write;
749 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
750 0 : let mut file = std::fs::File::create(conf_path)?;
751 0 : writeln!(file, "shared_buffers=65536")?;
752 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
753 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
754 :
755 : // Start postgres
756 0 : info!("starting postgres");
757 0 : let mut pg = maybe_cgexec(&self.pgbin)
758 0 : .args(["-D", pgdata])
759 0 : .spawn()
760 0 : .expect("cannot start postgres process");
761 0 :
762 0 : // Stop it when it's ready
763 0 : info!("waiting for postgres");
764 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
765 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
766 : // it to avoid orphaned processes prowling around while datadir is
767 : // wiped.
768 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
769 0 : kill(pm_pid, Signal::SIGQUIT)?;
770 0 : info!("sent SIGQUIT signal");
771 0 : pg.wait()?;
772 0 : info!("done prewarming");
773 :
774 : // clean up
775 0 : let _ok = fs::remove_dir_all(pgdata);
776 0 : Ok(())
777 0 : }
778 :
779 : /// Start Postgres as a child process and manage DBs/roles.
780 : /// After that this will hang waiting on the postmaster process to exit.
781 : /// Returns a handle to the child process and a handle to the logs thread.
782 : #[instrument(skip_all)]
783 : pub fn start_postgres(
784 : &self,
785 : storage_auth_token: Option<String>,
786 : ) -> Result<(std::process::Child, tokio::task::JoinHandle<Result<()>>)> {
787 : let pgdata_path = Path::new(&self.pgdata);
788 :
789 : // Run postgres as a child process.
790 : let mut pg = maybe_cgexec(&self.pgbin)
791 : .args(["-D", &self.pgdata])
792 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
793 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
794 : } else {
795 : vec![]
796 : })
797 : .stderr(Stdio::piped())
798 : .spawn()
799 : .expect("cannot start postgres process");
800 : PG_PID.store(pg.id(), Ordering::SeqCst);
801 :
802 : // Start a task to collect logs from stderr.
803 : let stderr = pg.stderr.take().expect("stderr should be captured");
804 : let logs_handle = handle_postgres_logs(stderr);
805 :
806 : wait_for_postgres(&mut pg, pgdata_path)?;
807 :
808 : Ok((pg, logs_handle))
809 : }
810 :
811 : /// Do post configuration of the already started Postgres. This function spawns a background task to
812 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
813 : /// version. In the future, it may upgrade all 3rd-party extensions.
814 : #[instrument(skip_all)]
815 : pub fn post_apply_config(&self) -> Result<()> {
816 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
817 0 : tokio::spawn(async move {
818 0 : let res = async {
819 0 : let (mut client, connection) = conf.connect(NoTls).await?;
820 0 : tokio::spawn(async move {
821 0 : if let Err(e) = connection.await {
822 0 : eprintln!("connection error: {}", e);
823 0 : }
824 0 : });
825 0 :
826 0 : handle_neon_extension_upgrade(&mut client)
827 0 : .await
828 0 : .context("handle_neon_extension_upgrade")?;
829 0 : Ok::<_, anyhow::Error>(())
830 0 : }
831 0 : .await;
832 0 : if let Err(err) = res {
833 0 : error!("error while post_apply_config: {err:#}");
834 0 : }
835 0 : });
836 : Ok(())
837 : }
838 :
839 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
840 0 : let mut conf = self.conn_conf.clone();
841 0 : if let Some(application_name) = application_name {
842 0 : conf.application_name(application_name);
843 0 : }
844 0 : conf
845 0 : }
846 :
847 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
848 0 : let mut conf = self.tokio_conn_conf.clone();
849 0 : if let Some(application_name) = application_name {
850 0 : conf.application_name(application_name);
851 0 : }
852 0 : conf
853 0 : }
854 :
855 0 : pub async fn get_maintenance_client(
856 0 : conf: &tokio_postgres::Config,
857 0 : ) -> Result<tokio_postgres::Client> {
858 0 : let mut conf = conf.clone();
859 0 : conf.application_name("compute_ctl:apply_config");
860 :
861 0 : let (client, conn) = match conf.connect(NoTls).await {
862 : // If connection fails, it may be the old node with `zenith_admin` superuser.
863 : //
864 : // In this case we need to connect with old `zenith_admin` name
865 : // and create new user. We cannot simply rename connected user,
866 : // but we can create a new one and grant it all privileges.
867 0 : Err(e) => match e.code() {
868 : Some(&SqlState::INVALID_PASSWORD)
869 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
870 : // Connect with zenith_admin if cloud_admin could not authenticate
871 0 : info!(
872 0 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
873 : e
874 : );
875 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
876 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
877 0 : zenith_admin_conf.user("zenith_admin");
878 :
879 0 : let mut client =
880 0 : zenith_admin_conf.connect(NoTls)
881 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
882 :
883 : // Disable forwarding so that users don't get a cloud_admin role
884 0 : let mut func = || {
885 0 : client.simple_query("SET neon.forward_ddl = false")?;
886 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
887 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
888 0 : Ok::<_, anyhow::Error>(())
889 0 : };
890 0 : func().context("apply_config setup cloud_admin")?;
891 :
892 0 : drop(client);
893 0 :
894 0 : // Reconnect with connstring with expected name
895 0 : conf.connect(NoTls).await?
896 : }
897 0 : _ => return Err(e.into()),
898 : },
899 0 : Ok((client, conn)) => (client, conn),
900 : };
901 :
902 0 : spawn(async move {
903 0 : if let Err(e) = conn.await {
904 0 : error!("maintenance client connection error: {}", e);
905 0 : }
906 0 : });
907 0 :
908 0 : // Disable DDL forwarding because control plane already knows about the roles/databases
909 0 : // we're about to modify.
910 0 : client
911 0 : .simple_query("SET neon.forward_ddl = false")
912 0 : .await
913 0 : .context("apply_config SET neon.forward_ddl = false")?;
914 :
915 0 : Ok(client)
916 0 : }
917 :
918 : /// Apply the spec to the running PostgreSQL instance.
919 : /// The caller can decide to run with multiple clients in parallel, or
920 : /// single mode. Either way, the commands executed will be the same, and
921 : /// only commands run in different databases are parallelized.
922 : #[instrument(skip_all)]
923 : pub fn apply_spec_sql(
924 : &self,
925 : spec: Arc<ComputeSpec>,
926 : conf: Arc<tokio_postgres::Config>,
927 : concurrency: usize,
928 : ) -> Result<()> {
929 : info!("Applying config with max {} concurrency", concurrency);
930 : debug!("Config: {:?}", spec);
931 :
932 : let rt = tokio::runtime::Handle::current();
933 0 : rt.block_on(async {
934 : // Proceed with post-startup configuration. Note, that order of operations is important.
935 0 : let client = Self::get_maintenance_client(&conf).await?;
936 0 : let spec = spec.clone();
937 :
938 0 : let databases = get_existing_dbs_async(&client).await?;
939 0 : let roles = get_existing_roles_async(&client)
940 0 : .await?
941 0 : .into_iter()
942 0 : .map(|role| (role.name.clone(), role))
943 0 : .collect::<HashMap<String, Role>>();
944 0 :
945 0 : // Check if we need to drop subscriptions before starting the endpoint.
946 0 : //
947 0 : // It is important to do this operation exactly once when endpoint starts on a new branch.
948 0 : // Otherwise, we may drop not inherited, but newly created subscriptions.
949 0 : //
950 0 : // We cannot rely only on spec.drop_subscriptions_before_start flag,
951 0 : // because if for some reason compute restarts inside VM,
952 0 : // it will start again with the same spec and flag value.
953 0 : //
954 0 : // To handle this, we save the fact of the operation in the database
955 0 : // in the neon.drop_subscriptions_done table.
956 0 : // If the table does not exist, we assume that the operation was never performed, so we must do it.
957 0 : // If table exists, we check if the operation was performed on the current timelilne.
958 0 : //
959 0 : let mut drop_subscriptions_done = false;
960 0 :
961 0 : if spec.drop_subscriptions_before_start {
962 0 : let timeline_id = self.get_timeline_id().context("timeline_id must be set")?;
963 0 : let query = format!("select 1 from neon.drop_subscriptions_done where timeline_id = '{}'", timeline_id);
964 0 :
965 0 : info!("Checking if drop subscription operation was already performed for timeline_id: {}", timeline_id);
966 :
967 : drop_subscriptions_done = match
968 0 : client.simple_query(&query).await {
969 0 : Ok(result) => {
970 0 : matches!(&result[0], postgres::SimpleQueryMessage::Row(_))
971 : },
972 0 : Err(e) =>
973 0 : {
974 0 : match e.code() {
975 0 : Some(&SqlState::UNDEFINED_TABLE) => false,
976 : _ => {
977 : // We don't expect any other error here, except for the schema/table not existing
978 0 : error!("Error checking if drop subscription operation was already performed: {}", e);
979 0 : return Err(e.into());
980 : }
981 : }
982 : }
983 : }
984 0 : };
985 :
986 :
987 0 : let jwks_roles = Arc::new(
988 0 : spec.as_ref()
989 0 : .local_proxy_config
990 0 : .iter()
991 0 : .flat_map(|it| &it.jwks)
992 0 : .flatten()
993 0 : .flat_map(|setting| &setting.role_names)
994 0 : .cloned()
995 0 : .collect::<HashSet<_>>(),
996 0 : );
997 0 :
998 0 : let ctx = Arc::new(tokio::sync::RwLock::new(MutableApplyContext {
999 0 : roles,
1000 0 : dbs: databases,
1001 0 : }));
1002 0 :
1003 0 : // Apply special pre drop database phase.
1004 0 : // NOTE: we use the code of RunInEachDatabase phase for parallelism
1005 0 : // and connection management, but we don't really run it in *each* database,
1006 0 : // only in databases, we're about to drop.
1007 0 : info!("Applying PerDatabase (pre-dropdb) phase");
1008 0 : let concurrency_token = Arc::new(tokio::sync::Semaphore::new(concurrency));
1009 0 :
1010 0 : // Run the phase for each database that we're about to drop.
1011 0 : let db_processes = spec
1012 0 : .delta_operations
1013 0 : .iter()
1014 0 : .flatten()
1015 0 : .filter_map(move |op| {
1016 0 : if op.action.as_str() == "delete_db" {
1017 0 : Some(op.name.clone())
1018 : } else {
1019 0 : None
1020 : }
1021 0 : })
1022 0 : .map(|dbname| {
1023 0 : let spec = spec.clone();
1024 0 : let ctx = ctx.clone();
1025 0 : let jwks_roles = jwks_roles.clone();
1026 0 : let mut conf = conf.as_ref().clone();
1027 0 : let concurrency_token = concurrency_token.clone();
1028 0 : // We only need dbname field for this phase, so set other fields to dummy values
1029 0 : let db = DB::UserDB(Database {
1030 0 : name: dbname.clone(),
1031 0 : owner: "cloud_admin".to_string(),
1032 0 : options: None,
1033 0 : restrict_conn: false,
1034 0 : invalid: false,
1035 0 : });
1036 0 :
1037 0 : debug!("Applying per-database phases for Database {:?}", &db);
1038 :
1039 0 : match &db {
1040 0 : DB::SystemDB => {}
1041 0 : DB::UserDB(db) => {
1042 0 : conf.dbname(db.name.as_str());
1043 0 : }
1044 : }
1045 :
1046 0 : let conf = Arc::new(conf);
1047 0 : let fut = Self::apply_spec_sql_db(
1048 0 : spec.clone(),
1049 0 : conf,
1050 0 : ctx.clone(),
1051 0 : jwks_roles.clone(),
1052 0 : concurrency_token.clone(),
1053 0 : db,
1054 0 : [DropLogicalSubscriptions].to_vec(),
1055 0 : );
1056 0 :
1057 0 : Ok(spawn(fut))
1058 0 : })
1059 0 : .collect::<Vec<Result<_, anyhow::Error>>>();
1060 :
1061 0 : for process in db_processes.into_iter() {
1062 0 : let handle = process?;
1063 0 : if let Err(e) = handle.await? {
1064 : // Handle the error case where the database does not exist
1065 : // We do not check whether the DB exists or not in the deletion phase,
1066 : // so we shouldn't be strict about it in pre-deletion cleanup as well.
1067 0 : if e.to_string().contains("does not exist") {
1068 0 : warn!("Error dropping subscription: {}", e);
1069 : } else {
1070 0 : return Err(e);
1071 : }
1072 0 : };
1073 : }
1074 :
1075 0 : for phase in [
1076 0 : CreateSuperUser,
1077 0 : DropInvalidDatabases,
1078 0 : RenameRoles,
1079 0 : CreateAndAlterRoles,
1080 0 : RenameAndDeleteDatabases,
1081 0 : CreateAndAlterDatabases,
1082 0 : CreateSchemaNeon,
1083 : ] {
1084 0 : info!("Applying phase {:?}", &phase);
1085 0 : apply_operations(
1086 0 : spec.clone(),
1087 0 : ctx.clone(),
1088 0 : jwks_roles.clone(),
1089 0 : phase,
1090 0 : || async { Ok(&client) },
1091 0 : )
1092 0 : .await?;
1093 : }
1094 :
1095 0 : info!("Applying RunInEachDatabase2 phase");
1096 0 : let concurrency_token = Arc::new(tokio::sync::Semaphore::new(concurrency));
1097 0 :
1098 0 : let db_processes = spec
1099 0 : .cluster
1100 0 : .databases
1101 0 : .iter()
1102 0 : .map(|db| DB::new(db.clone()))
1103 0 : // include
1104 0 : .chain(once(DB::SystemDB))
1105 0 : .map(|db| {
1106 0 : let spec = spec.clone();
1107 0 : let ctx = ctx.clone();
1108 0 : let jwks_roles = jwks_roles.clone();
1109 0 : let mut conf = conf.as_ref().clone();
1110 0 : let concurrency_token = concurrency_token.clone();
1111 0 : let db = db.clone();
1112 0 :
1113 0 : debug!("Applying per-database phases for Database {:?}", &db);
1114 :
1115 0 : match &db {
1116 0 : DB::SystemDB => {}
1117 0 : DB::UserDB(db) => {
1118 0 : conf.dbname(db.name.as_str());
1119 0 : }
1120 : }
1121 :
1122 0 : let conf = Arc::new(conf);
1123 0 : let mut phases = vec![
1124 0 : DeleteDBRoleReferences,
1125 0 : ChangeSchemaPerms,
1126 0 : HandleAnonExtension,
1127 0 : ];
1128 0 :
1129 0 : if spec.drop_subscriptions_before_start && !drop_subscriptions_done {
1130 0 : info!("Adding DropLogicalSubscriptions phase because drop_subscriptions_before_start is set");
1131 0 : phases.push(DropLogicalSubscriptions);
1132 0 : }
1133 :
1134 0 : let fut = Self::apply_spec_sql_db(
1135 0 : spec.clone(),
1136 0 : conf,
1137 0 : ctx.clone(),
1138 0 : jwks_roles.clone(),
1139 0 : concurrency_token.clone(),
1140 0 : db,
1141 0 : phases,
1142 0 : );
1143 0 :
1144 0 : Ok(spawn(fut))
1145 0 : })
1146 0 : .collect::<Vec<Result<_, anyhow::Error>>>();
1147 :
1148 0 : for process in db_processes.into_iter() {
1149 0 : let handle = process?;
1150 0 : handle.await??;
1151 : }
1152 :
1153 0 : let mut phases = vec![
1154 0 : HandleOtherExtensions,
1155 0 : HandleNeonExtension, // This step depends on CreateSchemaNeon
1156 0 : CreateAvailabilityCheck,
1157 0 : DropRoles,
1158 0 : ];
1159 0 :
1160 0 : // This step depends on CreateSchemaNeon
1161 0 : if spec.drop_subscriptions_before_start && !drop_subscriptions_done {
1162 0 : info!("Adding FinalizeDropLogicalSubscriptions phase because drop_subscriptions_before_start is set");
1163 0 : phases.push(FinalizeDropLogicalSubscriptions);
1164 0 : }
1165 :
1166 0 : for phase in phases {
1167 0 : debug!("Applying phase {:?}", &phase);
1168 0 : apply_operations(
1169 0 : spec.clone(),
1170 0 : ctx.clone(),
1171 0 : jwks_roles.clone(),
1172 0 : phase,
1173 0 : || async { Ok(&client) },
1174 0 : )
1175 0 : .await?;
1176 : }
1177 :
1178 0 : Ok::<(), anyhow::Error>(())
1179 0 : })?;
1180 :
1181 : Ok(())
1182 : }
1183 :
1184 : /// Apply SQL migrations of the RunInEachDatabase phase.
1185 : ///
1186 : /// May opt to not connect to databases that don't have any scheduled
1187 : /// operations. The function is concurrency-controlled with the provided
1188 : /// semaphore. The caller has to make sure the semaphore isn't exhausted.
1189 0 : async fn apply_spec_sql_db(
1190 0 : spec: Arc<ComputeSpec>,
1191 0 : conf: Arc<tokio_postgres::Config>,
1192 0 : ctx: Arc<tokio::sync::RwLock<MutableApplyContext>>,
1193 0 : jwks_roles: Arc<HashSet<String>>,
1194 0 : concurrency_token: Arc<tokio::sync::Semaphore>,
1195 0 : db: DB,
1196 0 : subphases: Vec<PerDatabasePhase>,
1197 0 : ) -> Result<()> {
1198 0 : let _permit = concurrency_token.acquire().await?;
1199 :
1200 0 : let mut client_conn = None;
1201 :
1202 0 : for subphase in subphases {
1203 0 : apply_operations(
1204 0 : spec.clone(),
1205 0 : ctx.clone(),
1206 0 : jwks_roles.clone(),
1207 0 : RunInEachDatabase {
1208 0 : db: db.clone(),
1209 0 : subphase,
1210 0 : },
1211 0 : // Only connect if apply_operation actually wants a connection.
1212 0 : // It's quite possible this database doesn't need any queries,
1213 0 : // so by not connecting we save time and effort connecting to
1214 0 : // that database.
1215 0 : || async {
1216 0 : if client_conn.is_none() {
1217 0 : let db_client = Self::get_maintenance_client(&conf).await?;
1218 0 : client_conn.replace(db_client);
1219 0 : }
1220 0 : let client = client_conn.as_ref().unwrap();
1221 0 : Ok(client)
1222 0 : },
1223 0 : )
1224 0 : .await?;
1225 : }
1226 :
1227 0 : drop(client_conn);
1228 0 :
1229 0 : Ok::<(), anyhow::Error>(())
1230 0 : }
1231 :
1232 : /// Choose how many concurrent connections to use for applying the spec changes.
1233 0 : pub fn max_service_connections(
1234 0 : &self,
1235 0 : compute_state: &ComputeState,
1236 0 : spec: &ComputeSpec,
1237 0 : ) -> usize {
1238 0 : // If the cluster is in Init state we don't have to deal with user connections,
1239 0 : // and can thus use all `max_connections` connection slots. However, that's generally not
1240 0 : // very efficient, so we generally still limit it to a smaller number.
1241 0 : if compute_state.status == ComputeStatus::Init {
1242 : // If the settings contain 'max_connections', use that as template
1243 0 : if let Some(config) = spec.cluster.settings.find("max_connections") {
1244 0 : config.parse::<usize>().ok()
1245 : } else {
1246 : // Otherwise, try to find the setting in the postgresql_conf string
1247 0 : spec.cluster
1248 0 : .postgresql_conf
1249 0 : .iter()
1250 0 : .flat_map(|conf| conf.split("\n"))
1251 0 : .filter_map(|line| {
1252 0 : if !line.contains("max_connections") {
1253 0 : return None;
1254 0 : }
1255 :
1256 0 : let (key, value) = line.split_once("=")?;
1257 0 : let key = key
1258 0 : .trim_start_matches(char::is_whitespace)
1259 0 : .trim_end_matches(char::is_whitespace);
1260 0 :
1261 0 : let value = value
1262 0 : .trim_start_matches(char::is_whitespace)
1263 0 : .trim_end_matches(char::is_whitespace);
1264 0 :
1265 0 : if key != "max_connections" {
1266 0 : return None;
1267 0 : }
1268 0 :
1269 0 : value.parse::<usize>().ok()
1270 0 : })
1271 0 : .next()
1272 : }
1273 : // If max_connections is present, use at most 1/3rd of that.
1274 : // When max_connections is lower than 30, try to use at least 10 connections, but
1275 : // never more than max_connections.
1276 0 : .map(|limit| match limit {
1277 0 : 0..10 => limit,
1278 0 : 10..30 => 10,
1279 0 : 30.. => limit / 3,
1280 0 : })
1281 0 : // If we didn't find max_connections, default to 10 concurrent connections.
1282 0 : .unwrap_or(10)
1283 : } else {
1284 : // state == Running
1285 : // Because the cluster is already in the Running state, we should assume users are
1286 : // already connected to the cluster, and high concurrency could negatively
1287 : // impact user connectivity. Therefore, we can limit concurrency to the number of
1288 : // reserved superuser connections, which users wouldn't be able to use anyway.
1289 0 : spec.cluster
1290 0 : .settings
1291 0 : .find("superuser_reserved_connections")
1292 0 : .iter()
1293 0 : .filter_map(|val| val.parse::<usize>().ok())
1294 0 : .map(|val| if val > 1 { val - 1 } else { 1 })
1295 0 : .last()
1296 0 : .unwrap_or(3)
1297 : }
1298 0 : }
1299 :
1300 : /// Do initial configuration of the already started Postgres.
1301 : #[instrument(skip_all)]
1302 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1303 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1304 :
1305 : let conf = Arc::new(conf);
1306 : let spec = Arc::new(
1307 : compute_state
1308 : .pspec
1309 : .as_ref()
1310 : .expect("spec must be set")
1311 : .spec
1312 : .clone(),
1313 : );
1314 :
1315 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1316 :
1317 : // Merge-apply spec & changes to PostgreSQL state.
1318 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1319 :
1320 : if let Some(ref local_proxy) = &spec.clone().local_proxy_config {
1321 : info!("configuring local_proxy");
1322 : local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
1323 : }
1324 :
1325 : // Run migrations separately to not hold up cold starts
1326 0 : tokio::spawn(async move {
1327 0 : let mut conf = conf.as_ref().clone();
1328 0 : conf.application_name("compute_ctl:migrations");
1329 0 :
1330 0 : match conf.connect(NoTls).await {
1331 0 : Ok((mut client, connection)) => {
1332 0 : tokio::spawn(async move {
1333 0 : if let Err(e) = connection.await {
1334 0 : eprintln!("connection error: {}", e);
1335 0 : }
1336 0 : });
1337 0 : if let Err(e) = handle_migrations(&mut client).await {
1338 0 : error!("Failed to run migrations: {}", e);
1339 0 : }
1340 : }
1341 0 : Err(e) => {
1342 0 : error!(
1343 0 : "Failed to connect to the compute for running migrations: {}",
1344 : e
1345 : );
1346 : }
1347 : };
1348 0 : });
1349 :
1350 : Ok::<(), anyhow::Error>(())
1351 : }
1352 :
1353 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1354 : // `pg_ctl` for start / stop.
1355 : #[instrument(skip_all)]
1356 : fn pg_reload_conf(&self) -> Result<()> {
1357 : let pgctl_bin = Path::new(&self.pgbin).parent().unwrap().join("pg_ctl");
1358 : Command::new(pgctl_bin)
1359 : .args(["reload", "-D", &self.pgdata])
1360 : .output()
1361 : .expect("cannot run pg_ctl process");
1362 : Ok(())
1363 : }
1364 :
1365 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1366 : /// as it's used to reconfigure a previously started and configured Postgres node.
1367 : #[instrument(skip_all)]
1368 : pub fn reconfigure(&self) -> Result<()> {
1369 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1370 :
1371 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1372 : info!("tuning pgbouncer");
1373 :
1374 : // Spawn a background task to do the tuning,
1375 : // so that we don't block the main thread that starts Postgres.
1376 : let pgbouncer_settings = pgbouncer_settings.clone();
1377 0 : tokio::spawn(async move {
1378 0 : let res = tune_pgbouncer(pgbouncer_settings).await;
1379 0 : if let Err(err) = res {
1380 0 : error!("error while tuning pgbouncer: {err:?}");
1381 0 : }
1382 0 : });
1383 : }
1384 :
1385 : if let Some(ref local_proxy) = spec.local_proxy_config {
1386 : info!("configuring local_proxy");
1387 :
1388 : // Spawn a background task to do the configuration,
1389 : // so that we don't block the main thread that starts Postgres.
1390 : let local_proxy = local_proxy.clone();
1391 0 : tokio::spawn(async move {
1392 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1393 0 : error!("error while configuring local_proxy: {err:?}");
1394 0 : }
1395 0 : });
1396 : }
1397 :
1398 : // Write new config
1399 : let pgdata_path = Path::new(&self.pgdata);
1400 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1401 : config::write_postgres_conf(&postgresql_conf_path, &spec, self.internal_http_port)?;
1402 :
1403 : if !spec.skip_pg_catalog_updates {
1404 : let max_concurrent_connections = spec.reconfigure_concurrency;
1405 : // Temporarily reset max_cluster_size in config
1406 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1407 : // creating new extensions, roles, etc.
1408 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1409 0 : self.pg_reload_conf()?;
1410 :
1411 0 : if spec.mode == ComputeMode::Primary {
1412 0 : let mut conf = tokio_postgres::Config::from_str(self.connstr.as_str()).unwrap();
1413 0 : conf.application_name("apply_config");
1414 0 : let conf = Arc::new(conf);
1415 0 :
1416 0 : let spec = Arc::new(spec.clone());
1417 0 :
1418 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
1419 0 : }
1420 :
1421 0 : Ok(())
1422 0 : })?;
1423 : }
1424 :
1425 : self.pg_reload_conf()?;
1426 :
1427 : let unknown_op = "unknown".to_string();
1428 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1429 : info!(
1430 : "finished reconfiguration of compute node for operation {}",
1431 : op_id
1432 : );
1433 :
1434 : Ok(())
1435 : }
1436 :
1437 : #[instrument(skip_all)]
1438 : pub fn start_compute(
1439 : &self,
1440 : ) -> Result<(std::process::Child, tokio::task::JoinHandle<Result<()>>)> {
1441 : let compute_state = self.state.lock().unwrap().clone();
1442 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1443 : info!(
1444 : "starting compute for project {}, operation {}, tenant {}, timeline {}",
1445 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
1446 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
1447 : pspec.tenant_id,
1448 : pspec.timeline_id,
1449 : );
1450 :
1451 : // tune pgbouncer
1452 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
1453 : info!("tuning pgbouncer");
1454 :
1455 : // Spawn a background task to do the tuning,
1456 : // so that we don't block the main thread that starts Postgres.
1457 : let pgbouncer_settings = pgbouncer_settings.clone();
1458 0 : let _handle = tokio::spawn(async move {
1459 0 : let res = tune_pgbouncer(pgbouncer_settings).await;
1460 0 : if let Err(err) = res {
1461 0 : error!("error while tuning pgbouncer: {err:?}");
1462 0 : }
1463 0 : });
1464 : }
1465 :
1466 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
1467 : info!("configuring local_proxy");
1468 :
1469 : // Spawn a background task to do the configuration,
1470 : // so that we don't block the main thread that starts Postgres.
1471 : let local_proxy = local_proxy.clone();
1472 0 : let _handle = tokio::spawn(async move {
1473 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1474 0 : error!("error while configuring local_proxy: {err:?}");
1475 0 : }
1476 0 : });
1477 : }
1478 :
1479 : info!(
1480 : "start_compute spec.remote_extensions {:?}",
1481 : pspec.spec.remote_extensions
1482 : );
1483 :
1484 : // This part is sync, because we need to download
1485 : // remote shared_preload_libraries before postgres start (if any)
1486 : if let Some(remote_extensions) = &pspec.spec.remote_extensions {
1487 : // First, create control files for all availale extensions
1488 : extension_server::create_control_files(remote_extensions, &self.pgbin);
1489 :
1490 : let library_load_start_time = Utc::now();
1491 : let rt = tokio::runtime::Handle::current();
1492 : let remote_ext_metrics = rt.block_on(self.prepare_preload_libraries(&pspec.spec))?;
1493 :
1494 : let library_load_time = Utc::now()
1495 : .signed_duration_since(library_load_start_time)
1496 : .to_std()
1497 : .unwrap()
1498 : .as_millis() as u64;
1499 : let mut state = self.state.lock().unwrap();
1500 : state.metrics.load_ext_ms = library_load_time;
1501 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
1502 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
1503 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
1504 : info!(
1505 : "Loading shared_preload_libraries took {:?}ms",
1506 : library_load_time
1507 : );
1508 : info!("{:?}", remote_ext_metrics);
1509 : }
1510 :
1511 : self.prepare_pgdata(&compute_state)?;
1512 :
1513 : let start_time = Utc::now();
1514 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
1515 :
1516 : let config_time = Utc::now();
1517 : if pspec.spec.mode == ComputeMode::Primary {
1518 : if !pspec.spec.skip_pg_catalog_updates {
1519 : let pgdata_path = Path::new(&self.pgdata);
1520 : // temporarily reset max_cluster_size in config
1521 : // to avoid the possibility of hitting the limit, while we are applying config:
1522 : // creating new extensions, roles, etc...
1523 : config::with_compute_ctl_tmp_override(
1524 : pgdata_path,
1525 : "neon.max_cluster_size=-1",
1526 0 : || {
1527 0 : self.pg_reload_conf()?;
1528 :
1529 0 : self.apply_config(&compute_state)?;
1530 :
1531 0 : Ok(())
1532 0 : },
1533 : )?;
1534 :
1535 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1536 : if config::line_in_file(
1537 : &postgresql_conf_path,
1538 : "neon.disable_logical_replication_subscribers=false",
1539 : )? {
1540 : info!("updated postgresql.conf to set neon.disable_logical_replication_subscribers=false");
1541 : }
1542 : self.pg_reload_conf()?;
1543 : }
1544 : self.post_apply_config()?;
1545 :
1546 : let conf = self.get_conn_conf(None);
1547 0 : tokio::task::spawn_blocking(|| {
1548 0 : let res = get_installed_extensions(conf);
1549 0 : match res {
1550 0 : Ok(extensions) => {
1551 0 : info!(
1552 0 : "[NEON_EXT_STAT] {}",
1553 0 : serde_json::to_string(&extensions)
1554 0 : .expect("failed to serialize extensions list")
1555 : );
1556 : }
1557 0 : Err(err) => error!("could not get installed extensions: {err:?}"),
1558 : }
1559 0 : });
1560 : }
1561 :
1562 : let startup_end_time = Utc::now();
1563 : {
1564 : let mut state = self.state.lock().unwrap();
1565 : state.metrics.start_postgres_ms = config_time
1566 : .signed_duration_since(start_time)
1567 : .to_std()
1568 : .unwrap()
1569 : .as_millis() as u64;
1570 : state.metrics.config_ms = startup_end_time
1571 : .signed_duration_since(config_time)
1572 : .to_std()
1573 : .unwrap()
1574 : .as_millis() as u64;
1575 : state.metrics.total_startup_ms = startup_end_time
1576 : .signed_duration_since(compute_state.start_time)
1577 : .to_std()
1578 : .unwrap()
1579 : .as_millis() as u64;
1580 : }
1581 : self.set_status(ComputeStatus::Running);
1582 :
1583 : info!(
1584 : "finished configuration of compute for project {}",
1585 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
1586 : );
1587 :
1588 : // Log metrics so that we can search for slow operations in logs
1589 : let metrics = {
1590 : let state = self.state.lock().unwrap();
1591 : state.metrics.clone()
1592 : };
1593 : info!(?metrics, "compute start finished");
1594 :
1595 : Ok(pg_process)
1596 : }
1597 :
1598 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1599 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1600 0 : let mut state = self.state.lock().unwrap();
1601 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1602 0 : if last_active > state.last_active {
1603 0 : state.last_active = last_active;
1604 0 : debug!("set the last compute activity time to: {:?}", last_active);
1605 0 : }
1606 0 : }
1607 :
1608 : // Look for core dumps and collect backtraces.
1609 : //
1610 : // EKS worker nodes have following core dump settings:
1611 : // /proc/sys/kernel/core_pattern -> core
1612 : // /proc/sys/kernel/core_uses_pid -> 1
1613 : // ulimit -c -> unlimited
1614 : // which results in core dumps being written to postgres data directory as core.<pid>.
1615 : //
1616 : // Use that as a default location and pattern, except macos where core dumps are written
1617 : // to /cores/ directory by default.
1618 : //
1619 : // With default Linux settings, the core dump file is called just "core", so check for
1620 : // that too.
1621 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1622 0 : let core_dump_dir = match std::env::consts::OS {
1623 0 : "macos" => Path::new("/cores/"),
1624 0 : _ => Path::new(&self.pgdata),
1625 : };
1626 :
1627 : // Collect core dump paths if any
1628 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1629 0 : let files = fs::read_dir(core_dump_dir)?;
1630 0 : let cores = files.filter_map(|entry| {
1631 0 : let entry = entry.ok()?;
1632 :
1633 0 : let is_core_dump = match entry.file_name().to_str()? {
1634 0 : n if n.starts_with("core.") => true,
1635 0 : "core" => true,
1636 0 : _ => false,
1637 : };
1638 0 : if is_core_dump {
1639 0 : Some(entry.path())
1640 : } else {
1641 0 : None
1642 : }
1643 0 : });
1644 :
1645 : // Print backtrace for each core dump
1646 0 : for core_path in cores {
1647 0 : warn!(
1648 0 : "core dump found: {}, collecting backtrace",
1649 0 : core_path.display()
1650 : );
1651 :
1652 : // Try first with gdb
1653 0 : let backtrace = Command::new("gdb")
1654 0 : .args(["--batch", "-q", "-ex", "bt", &self.pgbin])
1655 0 : .arg(&core_path)
1656 0 : .output();
1657 :
1658 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1659 0 : let backtrace = match backtrace {
1660 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1661 0 : warn!("cannot find gdb, trying lldb");
1662 0 : Command::new("lldb")
1663 0 : .arg("-c")
1664 0 : .arg(&core_path)
1665 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1666 0 : .output()
1667 : }
1668 0 : _ => backtrace,
1669 0 : }?;
1670 :
1671 0 : warn!(
1672 0 : "core dump backtrace: {}",
1673 0 : String::from_utf8_lossy(&backtrace.stdout)
1674 : );
1675 0 : warn!(
1676 0 : "debugger stderr: {}",
1677 0 : String::from_utf8_lossy(&backtrace.stderr)
1678 : );
1679 : }
1680 :
1681 0 : Ok(())
1682 0 : }
1683 :
1684 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1685 0 : pub async fn collect_insights(&self) -> String {
1686 0 : let mut result_rows: Vec<String> = Vec::new();
1687 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
1688 0 : let connect_result = conf.connect(NoTls).await;
1689 0 : let (client, connection) = connect_result.unwrap();
1690 0 : tokio::spawn(async move {
1691 0 : if let Err(e) = connection.await {
1692 0 : eprintln!("connection error: {}", e);
1693 0 : }
1694 0 : });
1695 0 : let result = client
1696 0 : .simple_query(
1697 0 : "SELECT
1698 0 : row_to_json(pg_stat_statements)
1699 0 : FROM
1700 0 : pg_stat_statements
1701 0 : WHERE
1702 0 : userid != 'cloud_admin'::regrole::oid
1703 0 : ORDER BY
1704 0 : (mean_exec_time + mean_plan_time) DESC
1705 0 : LIMIT 100",
1706 0 : )
1707 0 : .await;
1708 :
1709 0 : if let Ok(raw_rows) = result {
1710 0 : for message in raw_rows.iter() {
1711 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1712 0 : if let Some(json) = row.get(0) {
1713 0 : result_rows.push(json.to_string());
1714 0 : }
1715 0 : }
1716 : }
1717 :
1718 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1719 : } else {
1720 0 : "{{\"pg_stat_statements\": []}}".to_string()
1721 : }
1722 0 : }
1723 :
1724 : // download an archive, unzip and place files in correct locations
1725 0 : pub async fn download_extension(
1726 0 : &self,
1727 0 : real_ext_name: String,
1728 0 : ext_path: RemotePath,
1729 0 : ) -> Result<u64, DownloadError> {
1730 0 : let ext_remote_storage =
1731 0 : self.ext_remote_storage
1732 0 : .as_ref()
1733 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1734 0 : "Remote extensions storage is not configured",
1735 0 : )))?;
1736 :
1737 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1738 0 :
1739 0 : let mut first_try = false;
1740 0 : if !self
1741 0 : .ext_download_progress
1742 0 : .read()
1743 0 : .expect("lock err")
1744 0 : .contains_key(ext_archive_name)
1745 0 : {
1746 0 : self.ext_download_progress
1747 0 : .write()
1748 0 : .expect("lock err")
1749 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1750 0 : first_try = true;
1751 0 : }
1752 0 : let (download_start, download_completed) =
1753 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1754 0 : let start_time_delta = Utc::now()
1755 0 : .signed_duration_since(download_start)
1756 0 : .to_std()
1757 0 : .unwrap()
1758 0 : .as_millis() as u64;
1759 :
1760 : // how long to wait for extension download if it was started by another process
1761 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1762 :
1763 0 : if download_completed {
1764 0 : info!("extension already downloaded, skipping re-download");
1765 0 : return Ok(0);
1766 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1767 0 : info!("download {ext_archive_name} already started by another process, hanging untill completion or timeout");
1768 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1769 : loop {
1770 0 : info!("waiting for download");
1771 0 : interval.tick().await;
1772 0 : let (_, download_completed_now) =
1773 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1774 0 : if download_completed_now {
1775 0 : info!("download finished by whoever else downloaded it");
1776 0 : return Ok(0);
1777 0 : }
1778 : }
1779 : // NOTE: the above loop will get terminated
1780 : // based on the timeout of the download function
1781 0 : }
1782 0 :
1783 0 : // if extension hasn't been downloaded before or the previous
1784 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1785 0 : // then we try to download it here
1786 0 : info!("downloading new extension {ext_archive_name}");
1787 :
1788 0 : let download_size = extension_server::download_extension(
1789 0 : &real_ext_name,
1790 0 : &ext_path,
1791 0 : ext_remote_storage,
1792 0 : &self.pgbin,
1793 0 : )
1794 0 : .await
1795 0 : .map_err(DownloadError::Other);
1796 0 :
1797 0 : if download_size.is_ok() {
1798 0 : self.ext_download_progress
1799 0 : .write()
1800 0 : .expect("bad lock")
1801 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1802 0 : }
1803 :
1804 0 : download_size
1805 0 : }
1806 :
1807 0 : pub async fn set_role_grants(
1808 0 : &self,
1809 0 : db_name: &PgIdent,
1810 0 : schema_name: &PgIdent,
1811 0 : privileges: &[Privilege],
1812 0 : role_name: &PgIdent,
1813 0 : ) -> Result<()> {
1814 : use tokio_postgres::NoTls;
1815 :
1816 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
1817 0 : conf.dbname(db_name);
1818 :
1819 0 : let (db_client, conn) = conf
1820 0 : .connect(NoTls)
1821 0 : .await
1822 0 : .context("Failed to connect to the database")?;
1823 0 : tokio::spawn(conn);
1824 0 :
1825 0 : // TODO: support other types of grants apart from schemas?
1826 0 : let query = format!(
1827 0 : "GRANT {} ON SCHEMA {} TO {}",
1828 0 : privileges
1829 0 : .iter()
1830 0 : // should not be quoted as it's part of the command.
1831 0 : // is already sanitized so it's ok
1832 0 : .map(|p| p.as_str())
1833 0 : .collect::<Vec<&'static str>>()
1834 0 : .join(", "),
1835 0 : // quote the schema and role name as identifiers to sanitize them.
1836 0 : schema_name.pg_quote(),
1837 0 : role_name.pg_quote(),
1838 0 : );
1839 0 : db_client
1840 0 : .simple_query(&query)
1841 0 : .await
1842 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1843 :
1844 0 : Ok(())
1845 0 : }
1846 :
1847 0 : pub async fn install_extension(
1848 0 : &self,
1849 0 : ext_name: &PgIdent,
1850 0 : db_name: &PgIdent,
1851 0 : ext_version: ExtVersion,
1852 0 : ) -> Result<ExtVersion> {
1853 : use tokio_postgres::NoTls;
1854 :
1855 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
1856 0 : conf.dbname(db_name);
1857 :
1858 0 : let (db_client, conn) = conf
1859 0 : .connect(NoTls)
1860 0 : .await
1861 0 : .context("Failed to connect to the database")?;
1862 0 : tokio::spawn(conn);
1863 0 :
1864 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
1865 0 : let version: Option<ExtVersion> = db_client
1866 0 : .query_opt(version_query, &[&ext_name])
1867 0 : .await
1868 0 : .with_context(|| format!("Failed to execute query: {}", version_query))?
1869 0 : .map(|row| row.get(0));
1870 0 :
1871 0 : // sanitize the inputs as postgres idents.
1872 0 : let ext_name: String = ext_name.pg_quote();
1873 0 : let quoted_version: String = ext_version.pg_quote();
1874 :
1875 0 : if let Some(installed_version) = version {
1876 0 : if installed_version == ext_version {
1877 0 : return Ok(installed_version);
1878 0 : }
1879 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
1880 0 : db_client
1881 0 : .simple_query(&query)
1882 0 : .await
1883 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1884 : } else {
1885 0 : let query =
1886 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
1887 0 : db_client
1888 0 : .simple_query(&query)
1889 0 : .await
1890 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1891 : }
1892 :
1893 0 : Ok(ext_version)
1894 0 : }
1895 :
1896 0 : pub async fn prepare_preload_libraries(
1897 0 : &self,
1898 0 : spec: &ComputeSpec,
1899 0 : ) -> Result<RemoteExtensionMetrics> {
1900 0 : if self.ext_remote_storage.is_none() {
1901 0 : return Ok(RemoteExtensionMetrics {
1902 0 : num_ext_downloaded: 0,
1903 0 : largest_ext_size: 0,
1904 0 : total_ext_download_size: 0,
1905 0 : });
1906 0 : }
1907 0 : let remote_extensions = spec
1908 0 : .remote_extensions
1909 0 : .as_ref()
1910 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
1911 :
1912 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
1913 0 : let mut libs_vec = Vec::new();
1914 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
1915 0 : libs_vec = libs
1916 0 : .split(&[',', '\'', ' '])
1917 0 : .filter(|s| *s != "neon" && !s.is_empty())
1918 0 : .map(str::to_string)
1919 0 : .collect();
1920 0 : }
1921 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
1922 :
1923 : // that is used in neon_local and python tests
1924 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
1925 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
1926 0 : let mut shared_preload_libraries_line = "";
1927 0 : for line in conf_lines {
1928 0 : if line.starts_with("shared_preload_libraries") {
1929 0 : shared_preload_libraries_line = line;
1930 0 : }
1931 : }
1932 0 : let mut preload_libs_vec = Vec::new();
1933 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
1934 0 : preload_libs_vec = libs
1935 0 : .split(&[',', '\'', ' '])
1936 0 : .filter(|s| *s != "neon" && !s.is_empty())
1937 0 : .map(str::to_string)
1938 0 : .collect();
1939 0 : }
1940 0 : libs_vec.extend(preload_libs_vec);
1941 0 : }
1942 :
1943 : // Don't try to download libraries that are not in the index.
1944 : // Assume that they are already present locally.
1945 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
1946 0 :
1947 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
1948 :
1949 0 : let mut download_tasks = Vec::new();
1950 0 : for library in &libs_vec {
1951 0 : let (ext_name, ext_path) =
1952 0 : remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
1953 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
1954 : }
1955 0 : let results = join_all(download_tasks).await;
1956 :
1957 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
1958 0 : num_ext_downloaded: 0,
1959 0 : largest_ext_size: 0,
1960 0 : total_ext_download_size: 0,
1961 0 : };
1962 0 : for result in results {
1963 0 : let download_size = match result {
1964 0 : Ok(res) => {
1965 0 : remote_ext_metrics.num_ext_downloaded += 1;
1966 0 : res
1967 : }
1968 0 : Err(err) => {
1969 0 : // if we failed to download an extension, we don't want to fail the whole
1970 0 : // process, but we do want to log the error
1971 0 : error!("Failed to download extension: {}", err);
1972 0 : 0
1973 : }
1974 : };
1975 :
1976 0 : remote_ext_metrics.largest_ext_size =
1977 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
1978 0 : remote_ext_metrics.total_ext_download_size += download_size;
1979 : }
1980 0 : Ok(remote_ext_metrics)
1981 0 : }
1982 :
1983 : /// Waits until current thread receives a state changed notification and
1984 : /// the pageserver connection strings has changed.
1985 : ///
1986 : /// The operation will time out after a specified duration.
1987 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
1988 0 : let state = self.state.lock().unwrap();
1989 0 : let old_pageserver_connstr = state
1990 0 : .pspec
1991 0 : .as_ref()
1992 0 : .expect("spec must be set")
1993 0 : .pageserver_connstr
1994 0 : .clone();
1995 0 : let mut unchanged = true;
1996 0 : let _ = self
1997 0 : .state_changed
1998 0 : .wait_timeout_while(state, duration, |s| {
1999 0 : let pageserver_connstr = &s
2000 0 : .pspec
2001 0 : .as_ref()
2002 0 : .expect("spec must be set")
2003 0 : .pageserver_connstr;
2004 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
2005 0 : unchanged
2006 0 : })
2007 0 : .unwrap();
2008 0 : if !unchanged {
2009 0 : info!("Pageserver config changed");
2010 0 : }
2011 0 : }
2012 : }
2013 :
2014 0 : pub fn forward_termination_signal() {
2015 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2016 0 : if ss_pid != 0 {
2017 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2018 0 : kill(ss_pid, Signal::SIGTERM).ok();
2019 0 : }
2020 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2021 0 : if pg_pid != 0 {
2022 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2023 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2024 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2025 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2026 0 : kill(pg_pid, Signal::SIGINT).ok();
2027 0 : }
2028 0 : }
|
