Line data Source code
1 : use std::collections::HashMap;
2 : use std::env;
3 : use std::fs;
4 : use std::io::BufRead;
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::{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;
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 :
36 : use remote_storage::{DownloadError, RemotePath};
37 :
38 : use crate::checker::create_availability_check_data;
39 : use crate::installed_extensions::get_installed_extensions_sync;
40 : use crate::local_proxy;
41 : use crate::logger::inlinify;
42 : use crate::pg_helpers::*;
43 : use crate::spec::*;
44 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
45 : use crate::{config, extension_server};
46 :
47 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
48 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
49 :
50 : /// Compute node info shared across several `compute_ctl` threads.
51 : pub struct ComputeNode {
52 : // Url type maintains proper escaping
53 : pub connstr: url::Url,
54 : pub pgdata: String,
55 : pub pgbin: String,
56 : pub pgversion: String,
57 : /// We should only allow live re- / configuration of the compute node if
58 : /// it uses 'pull model', i.e. it can go to control-plane and fetch
59 : /// the latest configuration. Otherwise, there could be a case:
60 : /// - we start compute with some spec provided as argument
61 : /// - we push new spec and it does reconfiguration
62 : /// - but then something happens and compute pod / VM is destroyed,
63 : /// so k8s controller starts it again with the **old** spec
64 : ///
65 : /// and the same for empty computes:
66 : /// - we started compute without any spec
67 : /// - we push spec and it does configuration
68 : /// - but then it is restarted without any spec again
69 : pub live_config_allowed: bool,
70 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
71 : /// To allow HTTP API server to serving status requests, while configuration
72 : /// is in progress, lock should be held only for short periods of time to do
73 : /// read/write, not the whole configuration process.
74 : pub state: Mutex<ComputeState>,
75 : /// `Condvar` to allow notifying waiters about state changes.
76 : pub state_changed: Condvar,
77 : /// the address of extension storage proxy gateway
78 : pub ext_remote_storage: Option<String>,
79 : // key: ext_archive_name, value: started download time, download_completed?
80 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
81 : pub build_tag: String,
82 : }
83 :
84 : // store some metrics about download size that might impact startup time
85 : #[derive(Clone, Debug)]
86 : pub struct RemoteExtensionMetrics {
87 : num_ext_downloaded: u64,
88 : largest_ext_size: u64,
89 : total_ext_download_size: u64,
90 : }
91 :
92 : #[derive(Clone, Debug)]
93 : pub struct ComputeState {
94 : pub start_time: DateTime<Utc>,
95 : pub status: ComputeStatus,
96 : /// Timestamp of the last Postgres activity. It could be `None` if
97 : /// compute wasn't used since start.
98 : pub last_active: Option<DateTime<Utc>>,
99 : pub error: Option<String>,
100 : pub pspec: Option<ParsedSpec>,
101 : pub metrics: ComputeMetrics,
102 : }
103 :
104 : impl ComputeState {
105 0 : pub fn new() -> Self {
106 0 : Self {
107 0 : start_time: Utc::now(),
108 0 : status: ComputeStatus::Empty,
109 0 : last_active: None,
110 0 : error: None,
111 0 : pspec: None,
112 0 : metrics: ComputeMetrics::default(),
113 0 : }
114 0 : }
115 :
116 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
117 0 : let prev = self.status;
118 0 : info!("Changing compute status from {} to {}", prev, status);
119 0 : self.status = status;
120 0 : state_changed.notify_all();
121 0 : }
122 :
123 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
124 0 : self.error = Some(format!("{err:?}"));
125 0 : self.set_status(ComputeStatus::Failed, state_changed);
126 0 : }
127 : }
128 :
129 : impl Default for ComputeState {
130 0 : fn default() -> Self {
131 0 : Self::new()
132 0 : }
133 : }
134 :
135 : #[derive(Clone, Debug)]
136 : pub struct ParsedSpec {
137 : pub spec: ComputeSpec,
138 : pub tenant_id: TenantId,
139 : pub timeline_id: TimelineId,
140 : pub pageserver_connstr: String,
141 : pub safekeeper_connstrings: Vec<String>,
142 : pub storage_auth_token: Option<String>,
143 : }
144 :
145 : impl TryFrom<ComputeSpec> for ParsedSpec {
146 : type Error = String;
147 0 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
148 : // Extract the options from the spec file that are needed to connect to
149 : // the storage system.
150 : //
151 : // For backwards-compatibility, the top-level fields in the spec file
152 : // may be empty. In that case, we need to dig them from the GUCs in the
153 : // cluster.settings field.
154 0 : let pageserver_connstr = spec
155 0 : .pageserver_connstring
156 0 : .clone()
157 0 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
158 0 : .ok_or("pageserver connstr should be provided")?;
159 0 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
160 0 : if matches!(spec.mode, ComputeMode::Primary) {
161 0 : spec.cluster
162 0 : .settings
163 0 : .find("neon.safekeepers")
164 0 : .ok_or("safekeeper connstrings should be provided")?
165 0 : .split(',')
166 0 : .map(|str| str.to_string())
167 0 : .collect()
168 : } else {
169 0 : vec![]
170 : }
171 : } else {
172 0 : spec.safekeeper_connstrings.clone()
173 : };
174 0 : let storage_auth_token = spec.storage_auth_token.clone();
175 0 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
176 0 : tenant_id
177 : } else {
178 0 : spec.cluster
179 0 : .settings
180 0 : .find("neon.tenant_id")
181 0 : .ok_or("tenant id should be provided")
182 0 : .map(|s| TenantId::from_str(&s))?
183 0 : .or(Err("invalid tenant id"))?
184 : };
185 0 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
186 0 : timeline_id
187 : } else {
188 0 : spec.cluster
189 0 : .settings
190 0 : .find("neon.timeline_id")
191 0 : .ok_or("timeline id should be provided")
192 0 : .map(|s| TimelineId::from_str(&s))?
193 0 : .or(Err("invalid timeline id"))?
194 : };
195 :
196 0 : Ok(ParsedSpec {
197 0 : spec,
198 0 : pageserver_connstr,
199 0 : safekeeper_connstrings,
200 0 : storage_auth_token,
201 0 : tenant_id,
202 0 : timeline_id,
203 0 : })
204 0 : }
205 : }
206 :
207 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
208 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
209 : ///
210 : /// This function should be used to start postgres, as it will start it in the
211 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
212 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
213 : /// creates it during the sysinit phase of its inittab.
214 0 : fn maybe_cgexec(cmd: &str) -> Command {
215 0 : // The cplane sets this env var for autoscaling computes.
216 0 : // use `var_os` so we don't have to worry about the variable being valid
217 0 : // unicode. Should never be an concern . . . but just in case
218 0 : if env::var_os("AUTOSCALING").is_some() {
219 0 : let mut command = Command::new("cgexec");
220 0 : command.args(["-g", "memory:neon-postgres"]);
221 0 : command.arg(cmd);
222 0 : command
223 : } else {
224 0 : Command::new(cmd)
225 : }
226 0 : }
227 :
228 : /// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
229 : /// that we give to customers
230 0 : #[instrument(skip_all)]
231 : fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
232 : let roles = spec
233 : .cluster
234 : .roles
235 : .iter()
236 0 : .map(|r| escape_literal(&r.name))
237 : .collect::<Vec<_>>();
238 :
239 : let dbs = spec
240 : .cluster
241 : .databases
242 : .iter()
243 0 : .map(|db| escape_literal(&db.name))
244 : .collect::<Vec<_>>();
245 :
246 : let roles_decl = if roles.is_empty() {
247 : String::from("roles text[] := NULL;")
248 : } else {
249 : format!(
250 : r#"
251 : roles text[] := ARRAY(SELECT rolname
252 : FROM pg_catalog.pg_roles
253 : WHERE rolname IN ({}));"#,
254 : roles.join(", ")
255 : )
256 : };
257 :
258 : let database_decl = if dbs.is_empty() {
259 : String::from("dbs text[] := NULL;")
260 : } else {
261 : format!(
262 : r#"
263 : dbs text[] := ARRAY(SELECT datname
264 : FROM pg_catalog.pg_database
265 : WHERE datname IN ({}));"#,
266 : dbs.join(", ")
267 : )
268 : };
269 :
270 : // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on all databases
271 : // (see https://www.postgresql.org/docs/current/ddl-priv.html)
272 : let query = format!(
273 : r#"
274 : DO $$
275 : DECLARE
276 : r text;
277 : {}
278 : {}
279 : BEGIN
280 : IF NOT EXISTS (
281 : SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
282 : THEN
283 : CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
284 : IF array_length(roles, 1) IS NOT NULL THEN
285 : EXECUTE format('GRANT neon_superuser TO %s',
286 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
287 : FOREACH r IN ARRAY roles LOOP
288 : EXECUTE format('ALTER ROLE %s CREATEROLE CREATEDB', quote_ident(r));
289 : END LOOP;
290 : END IF;
291 : IF array_length(dbs, 1) IS NOT NULL THEN
292 : EXECUTE format('GRANT ALL PRIVILEGES ON DATABASE %s TO neon_superuser',
293 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(dbs) as x), ', '));
294 : END IF;
295 : END IF;
296 : END
297 : $$;"#,
298 : roles_decl, database_decl,
299 : );
300 : info!("Neon superuser created: {}", inlinify(&query));
301 : client
302 : .simple_query(&query)
303 0 : .map_err(|e| anyhow::anyhow!(e).context(query))?;
304 : Ok(())
305 : }
306 :
307 : impl ComputeNode {
308 : /// Check that compute node has corresponding feature enabled.
309 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
310 0 : let state = self.state.lock().unwrap();
311 :
312 0 : if let Some(s) = state.pspec.as_ref() {
313 0 : s.spec.features.contains(&feature)
314 : } else {
315 0 : false
316 : }
317 0 : }
318 :
319 0 : pub fn set_status(&self, status: ComputeStatus) {
320 0 : let mut state = self.state.lock().unwrap();
321 0 : state.set_status(status, &self.state_changed);
322 0 : }
323 :
324 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
325 0 : let mut state = self.state.lock().unwrap();
326 0 : state.set_failed_status(err, &self.state_changed);
327 0 : }
328 :
329 0 : pub fn get_status(&self) -> ComputeStatus {
330 0 : self.state.lock().unwrap().status
331 0 : }
332 :
333 : // Remove `pgdata` directory and create it again with right permissions.
334 0 : fn create_pgdata(&self) -> Result<()> {
335 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
336 0 : // If it is something different then create_dir() will error out anyway.
337 0 : let _ok = fs::remove_dir_all(&self.pgdata);
338 0 : fs::create_dir(&self.pgdata)?;
339 0 : fs::set_permissions(&self.pgdata, fs::Permissions::from_mode(0o700))?;
340 :
341 0 : Ok(())
342 0 : }
343 :
344 : // Get basebackup from the libpq connection to pageserver using `connstr` and
345 : // unarchive it to `pgdata` directory overriding all its previous content.
346 0 : #[instrument(skip_all, fields(%lsn))]
347 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
348 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
349 : let start_time = Instant::now();
350 :
351 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
352 : let mut config = postgres::Config::from_str(shard0_connstr)?;
353 :
354 : // Use the storage auth token from the config file, if given.
355 : // Note: this overrides any password set in the connection string.
356 : if let Some(storage_auth_token) = &spec.storage_auth_token {
357 : info!("Got storage auth token from spec file");
358 : config.password(storage_auth_token);
359 : } else {
360 : info!("Storage auth token not set");
361 : }
362 :
363 : // Connect to pageserver
364 : let mut client = config.connect(NoTls)?;
365 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
366 :
367 : let basebackup_cmd = match lsn {
368 : // HACK We don't use compression on first start (Lsn(0)) because there's no API for it
369 : Lsn(0) => format!("basebackup {} {}", spec.tenant_id, spec.timeline_id),
370 : _ => format!(
371 : "basebackup {} {} {} --gzip",
372 : spec.tenant_id, spec.timeline_id, lsn
373 : ),
374 : };
375 :
376 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
377 : let mut measured_reader = MeasuredReader::new(copyreader);
378 :
379 : // Check the magic number to see if it's a gzip or not. Even though
380 : // we might explicitly ask for gzip, an old pageserver with no implementation
381 : // of gzip compression might send us uncompressed data. After some time
382 : // passes we can assume all pageservers know how to compress and we can
383 : // delete this check.
384 : //
385 : // If the data is not gzip, it will be tar. It will not be mistakenly
386 : // recognized as gzip because tar starts with an ascii encoding of a filename,
387 : // and 0x1f and 0x8b are unlikely first characters for any filename. Moreover,
388 : // we send the "global" directory first from the pageserver, so it definitely
389 : // won't be recognized as gzip.
390 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
391 : let gzip = {
392 : let peek = bufreader.fill_buf().unwrap();
393 : peek[0] == 0x1f && peek[1] == 0x8b
394 : };
395 :
396 : // Read the archive directly from the `CopyOutReader`
397 : //
398 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
399 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
400 : // sends an Error after finishing the tarball, we will not notice it.
401 : if gzip {
402 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
403 : ar.set_ignore_zeros(true);
404 : ar.unpack(&self.pgdata)?;
405 : } else {
406 : let mut ar = tar::Archive::new(&mut bufreader);
407 : ar.set_ignore_zeros(true);
408 : ar.unpack(&self.pgdata)?;
409 : };
410 :
411 : // Report metrics
412 : let mut state = self.state.lock().unwrap();
413 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
414 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
415 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
416 : Ok(())
417 : }
418 :
419 : // Gets the basebackup in a retry loop
420 0 : #[instrument(skip_all, fields(%lsn))]
421 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
422 : let mut retry_period_ms = 500.0;
423 : let mut attempts = 0;
424 : const DEFAULT_ATTEMPTS: u16 = 10;
425 : #[cfg(feature = "testing")]
426 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
427 : u16::from_str(&v).unwrap()
428 : } else {
429 : DEFAULT_ATTEMPTS
430 : };
431 : #[cfg(not(feature = "testing"))]
432 : let max_attempts = DEFAULT_ATTEMPTS;
433 : loop {
434 : let result = self.try_get_basebackup(compute_state, lsn);
435 : match result {
436 : Ok(_) => {
437 : return result;
438 : }
439 : Err(ref e) if attempts < max_attempts => {
440 : warn!(
441 : "Failed to get basebackup: {} (attempt {}/{})",
442 : e, attempts, max_attempts
443 : );
444 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
445 : retry_period_ms *= 1.5;
446 : }
447 : Err(_) => {
448 : return result;
449 : }
450 : }
451 : attempts += 1;
452 : }
453 : }
454 :
455 0 : pub async fn check_safekeepers_synced_async(
456 0 : &self,
457 0 : compute_state: &ComputeState,
458 0 : ) -> Result<Option<Lsn>> {
459 0 : // Construct a connection config for each safekeeper
460 0 : let pspec: ParsedSpec = compute_state
461 0 : .pspec
462 0 : .as_ref()
463 0 : .expect("spec must be set")
464 0 : .clone();
465 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
466 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
467 0 : // Format connstr
468 0 : let id = connstr.clone();
469 0 : let connstr = format!("postgresql://no_user@{}", connstr);
470 0 : let options = format!(
471 0 : "-c timeline_id={} tenant_id={}",
472 0 : pspec.timeline_id, pspec.tenant_id
473 0 : );
474 0 :
475 0 : // Construct client
476 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
477 0 : config.options(&options);
478 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
479 0 : config.password(storage_auth_token);
480 0 : }
481 :
482 0 : (id, config)
483 0 : });
484 0 :
485 0 : // Create task set to query all safekeepers
486 0 : let mut tasks = FuturesUnordered::new();
487 0 : let quorum = sk_configs.len() / 2 + 1;
488 0 : for (id, config) in sk_configs {
489 0 : let timeout = tokio::time::Duration::from_millis(100);
490 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
491 0 : tasks.push(tokio::spawn(task));
492 0 : }
493 :
494 : // Get a quorum of responses or errors
495 0 : let mut responses = Vec::new();
496 0 : let mut join_errors = Vec::new();
497 0 : let mut task_errors = Vec::new();
498 0 : let mut timeout_errors = Vec::new();
499 0 : while let Some(response) = tasks.next().await {
500 0 : match response {
501 0 : Ok(Ok(Ok(r))) => responses.push(r),
502 0 : Ok(Ok(Err(e))) => task_errors.push(e),
503 0 : Ok(Err(e)) => timeout_errors.push(e),
504 0 : Err(e) => join_errors.push(e),
505 : };
506 0 : if responses.len() >= quorum {
507 0 : break;
508 0 : }
509 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
510 0 : break;
511 0 : }
512 : }
513 :
514 : // In case of error, log and fail the check, but don't crash.
515 : // We're playing it safe because these errors could be transient
516 : // and we don't yet retry. Also being careful here allows us to
517 : // be backwards compatible with safekeepers that don't have the
518 : // TIMELINE_STATUS API yet.
519 0 : if responses.len() < quorum {
520 0 : error!(
521 0 : "failed sync safekeepers check {:?} {:?} {:?}",
522 : join_errors, task_errors, timeout_errors
523 : );
524 0 : return Ok(None);
525 0 : }
526 0 :
527 0 : Ok(check_if_synced(responses))
528 0 : }
529 :
530 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
531 : // in one roundtrip. If not, we should do a full sync_safekeepers.
532 0 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
533 0 : let start_time = Utc::now();
534 0 :
535 0 : // Run actual work with new tokio runtime
536 0 : let rt = tokio::runtime::Builder::new_current_thread()
537 0 : .enable_all()
538 0 : .build()
539 0 : .expect("failed to create rt");
540 0 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
541 0 :
542 0 : // Record runtime
543 0 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
544 0 : .signed_duration_since(start_time)
545 0 : .to_std()
546 0 : .unwrap()
547 0 : .as_millis() as u64;
548 0 : result
549 0 : }
550 :
551 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
552 : // and return the reported LSN back to the caller.
553 0 : #[instrument(skip_all)]
554 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
555 : let start_time = Utc::now();
556 :
557 : let mut sync_handle = maybe_cgexec(&self.pgbin)
558 : .args(["--sync-safekeepers"])
559 : .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
560 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
561 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
562 : } else {
563 : vec![]
564 : })
565 : .stdout(Stdio::piped())
566 : .stderr(Stdio::piped())
567 : .spawn()
568 : .expect("postgres --sync-safekeepers failed to start");
569 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
570 :
571 : // `postgres --sync-safekeepers` will print all log output to stderr and
572 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
573 : // will be collected in a child thread.
574 : let stderr = sync_handle
575 : .stderr
576 : .take()
577 : .expect("stderr should be captured");
578 : let logs_handle = handle_postgres_logs(stderr);
579 :
580 : let sync_output = sync_handle
581 : .wait_with_output()
582 : .expect("postgres --sync-safekeepers failed");
583 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
584 :
585 : // Process has exited, so we can join the logs thread.
586 : let _ = logs_handle
587 : .join()
588 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
589 :
590 : if !sync_output.status.success() {
591 : anyhow::bail!(
592 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
593 : sync_output.status,
594 : String::from_utf8(sync_output.stdout)
595 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
596 : );
597 : }
598 :
599 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
600 : .signed_duration_since(start_time)
601 : .to_std()
602 : .unwrap()
603 : .as_millis() as u64;
604 :
605 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
606 :
607 : Ok(lsn)
608 : }
609 :
610 : /// Do all the preparations like PGDATA directory creation, configuration,
611 : /// safekeepers sync, basebackup, etc.
612 0 : #[instrument(skip_all)]
613 : pub fn prepare_pgdata(
614 : &self,
615 : compute_state: &ComputeState,
616 : extension_server_port: u16,
617 : ) -> Result<()> {
618 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
619 : let spec = &pspec.spec;
620 : let pgdata_path = Path::new(&self.pgdata);
621 :
622 : // Remove/create an empty pgdata directory and put configuration there.
623 : self.create_pgdata()?;
624 : config::write_postgres_conf(
625 : &pgdata_path.join("postgresql.conf"),
626 : &pspec.spec,
627 : Some(extension_server_port),
628 : )?;
629 :
630 : // Syncing safekeepers is only safe with primary nodes: if a primary
631 : // is already connected it will be kicked out, so a secondary (standby)
632 : // cannot sync safekeepers.
633 : let lsn = match spec.mode {
634 : ComputeMode::Primary => {
635 : info!("checking if safekeepers are synced");
636 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
637 : lsn
638 : } else {
639 : info!("starting safekeepers syncing");
640 : self.sync_safekeepers(pspec.storage_auth_token.clone())
641 0 : .with_context(|| "failed to sync safekeepers")?
642 : };
643 : info!("safekeepers synced at LSN {}", lsn);
644 : lsn
645 : }
646 : ComputeMode::Static(lsn) => {
647 : info!("Starting read-only node at static LSN {}", lsn);
648 : lsn
649 : }
650 : ComputeMode::Replica => {
651 : info!("Initializing standby from latest Pageserver LSN");
652 : Lsn(0)
653 : }
654 : };
655 :
656 : info!(
657 : "getting basebackup@{} from pageserver {}",
658 : lsn, &pspec.pageserver_connstr
659 : );
660 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
661 0 : format!(
662 0 : "failed to get basebackup@{} from pageserver {}",
663 0 : lsn, &pspec.pageserver_connstr
664 0 : )
665 0 : })?;
666 :
667 : // Update pg_hba.conf received with basebackup.
668 : update_pg_hba(pgdata_path)?;
669 :
670 : // Place pg_dynshmem under /dev/shm. This allows us to use
671 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
672 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
673 : //
674 : // Why on earth don't we just stick to the 'posix' default, you might
675 : // ask. It turns out that making large allocations with 'posix' doesn't
676 : // work very well with autoscaling. The behavior we want is that:
677 : //
678 : // 1. You can make large DSM allocations, larger than the current RAM
679 : // size of the VM, without errors
680 : //
681 : // 2. If the allocated memory is really used, the VM is scaled up
682 : // automatically to accommodate that
683 : //
684 : // We try to make that possible by having swap in the VM. But with the
685 : // default 'posix' DSM implementation, we fail step 1, even when there's
686 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
687 : // the shmem segment, which is really just a file in /dev/shm in Linux,
688 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
689 : // than available RAM.
690 : //
691 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
692 : // the Postgres 'mmap' DSM implementation doesn't use
693 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
694 : // fill the file with zeros. It's weird that that differs between
695 : // 'posix' and 'mmap', but we take advantage of it. When the file is
696 : // filled slowly with write(2), the kernel allows it to grow larger, as
697 : // long as there's swap available.
698 : //
699 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
700 : // segment to be larger than currently available RAM. But because we
701 : // don't want to store it on a real file, which the kernel would try to
702 : // flush to disk, so symlink pg_dynshm to /dev/shm.
703 : //
704 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
705 : // control plane control that option. If 'mmap' is not used, this
706 : // symlink doesn't affect anything.
707 : //
708 : // See https://github.com/neondatabase/autoscaling/issues/800
709 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
710 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
711 :
712 : match spec.mode {
713 : ComputeMode::Primary => {}
714 : ComputeMode::Replica | ComputeMode::Static(..) => {
715 : add_standby_signal(pgdata_path)?;
716 : }
717 : }
718 :
719 : Ok(())
720 : }
721 :
722 : /// Start and stop a postgres process to warm up the VM for startup.
723 0 : pub fn prewarm_postgres(&self) -> Result<()> {
724 0 : info!("prewarming");
725 :
726 : // Create pgdata
727 0 : let pgdata = &format!("{}.warmup", self.pgdata);
728 0 : create_pgdata(pgdata)?;
729 :
730 : // Run initdb to completion
731 0 : info!("running initdb");
732 0 : let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
733 0 : Command::new(initdb_bin)
734 0 : .args(["--pgdata", pgdata])
735 0 : .output()
736 0 : .expect("cannot start initdb process");
737 :
738 : // Write conf
739 : use std::io::Write;
740 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
741 0 : let mut file = std::fs::File::create(conf_path)?;
742 0 : writeln!(file, "shared_buffers=65536")?;
743 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
744 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
745 :
746 : // Start postgres
747 0 : info!("starting postgres");
748 0 : let mut pg = maybe_cgexec(&self.pgbin)
749 0 : .args(["-D", pgdata])
750 0 : .spawn()
751 0 : .expect("cannot start postgres process");
752 0 :
753 0 : // Stop it when it's ready
754 0 : info!("waiting for postgres");
755 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
756 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
757 : // it to avoid orphaned processes prowling around while datadir is
758 : // wiped.
759 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
760 0 : kill(pm_pid, Signal::SIGQUIT)?;
761 0 : info!("sent SIGQUIT signal");
762 0 : pg.wait()?;
763 0 : info!("done prewarming");
764 :
765 : // clean up
766 0 : let _ok = fs::remove_dir_all(pgdata);
767 0 : Ok(())
768 0 : }
769 :
770 : /// Start Postgres as a child process and manage DBs/roles.
771 : /// After that this will hang waiting on the postmaster process to exit.
772 : /// Returns a handle to the child process and a handle to the logs thread.
773 0 : #[instrument(skip_all)]
774 : pub fn start_postgres(
775 : &self,
776 : storage_auth_token: Option<String>,
777 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
778 : let pgdata_path = Path::new(&self.pgdata);
779 :
780 : // Run postgres as a child process.
781 : let mut pg = maybe_cgexec(&self.pgbin)
782 : .args(["-D", &self.pgdata])
783 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
784 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
785 : } else {
786 : vec![]
787 : })
788 : .stderr(Stdio::piped())
789 : .spawn()
790 : .expect("cannot start postgres process");
791 : PG_PID.store(pg.id(), Ordering::SeqCst);
792 :
793 : // Start a thread to collect logs from stderr.
794 : let stderr = pg.stderr.take().expect("stderr should be captured");
795 : let logs_handle = handle_postgres_logs(stderr);
796 :
797 : wait_for_postgres(&mut pg, pgdata_path)?;
798 :
799 : Ok((pg, logs_handle))
800 : }
801 :
802 : /// Do post configuration of the already started Postgres. This function spawns a background thread to
803 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
804 : /// version. In the future, it may upgrade all 3rd-party extensions.
805 0 : #[instrument(skip_all)]
806 : pub fn post_apply_config(&self) -> Result<()> {
807 : let connstr = self.connstr.clone();
808 0 : thread::spawn(move || {
809 0 : let func = || {
810 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
811 0 : handle_neon_extension_upgrade(&mut client)
812 0 : .context("handle_neon_extension_upgrade")?;
813 0 : Ok::<_, anyhow::Error>(())
814 0 : };
815 0 : if let Err(err) = func() {
816 0 : error!("error while post_apply_config: {err:#}");
817 0 : }
818 0 : });
819 : Ok(())
820 : }
821 :
822 : /// Do initial configuration of the already started Postgres.
823 0 : #[instrument(skip_all)]
824 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
825 : // If connection fails,
826 : // it may be the old node with `zenith_admin` superuser.
827 : //
828 : // In this case we need to connect with old `zenith_admin` name
829 : // and create new user. We cannot simply rename connected user,
830 : // but we can create a new one and grant it all privileges.
831 : let mut connstr = self.connstr.clone();
832 : connstr
833 : .query_pairs_mut()
834 : .append_pair("application_name", "apply_config");
835 :
836 : let mut client = match Client::connect(connstr.as_str(), NoTls) {
837 : Err(e) => match e.code() {
838 : Some(&SqlState::INVALID_PASSWORD)
839 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
840 : // connect with zenith_admin if cloud_admin could not authenticate
841 : info!(
842 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
843 : e
844 : );
845 : let mut zenith_admin_connstr = connstr.clone();
846 :
847 : zenith_admin_connstr
848 : .set_username("zenith_admin")
849 0 : .map_err(|_| anyhow::anyhow!("invalid connstr"))?;
850 :
851 : let mut client =
852 : Client::connect(zenith_admin_connstr.as_str(), NoTls)
853 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
854 : // Disable forwarding so that users don't get a cloud_admin role
855 :
856 0 : let mut func = || {
857 0 : client.simple_query("SET neon.forward_ddl = false")?;
858 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
859 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
860 0 : Ok::<_, anyhow::Error>(())
861 0 : };
862 : func().context("apply_config setup cloud_admin")?;
863 :
864 : drop(client);
865 :
866 : // reconnect with connstring with expected name
867 : Client::connect(connstr.as_str(), NoTls)?
868 : }
869 : _ => return Err(e.into()),
870 : },
871 : Ok(client) => client,
872 : };
873 :
874 : // Disable DDL forwarding because control plane already knows about these roles/databases.
875 : client
876 : .simple_query("SET neon.forward_ddl = false")
877 : .context("apply_config SET neon.forward_ddl = false")?;
878 :
879 : // Proceed with post-startup configuration. Note, that order of operations is important.
880 : let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
881 : create_neon_superuser(spec, &mut client).context("apply_config create_neon_superuser")?;
882 : cleanup_instance(&mut client).context("apply_config cleanup_instance")?;
883 : handle_roles(spec, &mut client).context("apply_config handle_roles")?;
884 : handle_databases(spec, &mut client).context("apply_config handle_databases")?;
885 : handle_role_deletions(spec, connstr.as_str(), &mut client)
886 : .context("apply_config handle_role_deletions")?;
887 : handle_grants(
888 : spec,
889 : &mut client,
890 : connstr.as_str(),
891 : self.has_feature(ComputeFeature::AnonExtension),
892 : )
893 : .context("apply_config handle_grants")?;
894 : handle_extensions(spec, &mut client).context("apply_config handle_extensions")?;
895 : handle_extension_neon(&mut client).context("apply_config handle_extension_neon")?;
896 : create_availability_check_data(&mut client)
897 : .context("apply_config create_availability_check_data")?;
898 :
899 : // 'Close' connection
900 : drop(client);
901 :
902 : if let Some(ref local_proxy) = spec.local_proxy_config {
903 : info!("configuring local_proxy");
904 : local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
905 : }
906 :
907 : // Run migrations separately to not hold up cold starts
908 0 : thread::spawn(move || {
909 0 : let mut connstr = connstr.clone();
910 0 : connstr
911 0 : .query_pairs_mut()
912 0 : .append_pair("application_name", "migrations");
913 :
914 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
915 0 : handle_migrations(&mut client).context("apply_config handle_migrations")
916 0 : });
917 : Ok(())
918 : }
919 :
920 : // Wrapped this around `pg_ctl reload`, but right now we don't use
921 : // `pg_ctl` for start / stop.
922 0 : #[instrument(skip_all)]
923 : fn pg_reload_conf(&self) -> Result<()> {
924 : let pgctl_bin = Path::new(&self.pgbin).parent().unwrap().join("pg_ctl");
925 : Command::new(pgctl_bin)
926 : .args(["reload", "-D", &self.pgdata])
927 : .output()
928 : .expect("cannot run pg_ctl process");
929 : Ok(())
930 : }
931 :
932 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
933 : /// as it's used to reconfigure a previously started and configured Postgres node.
934 0 : #[instrument(skip_all)]
935 : pub fn reconfigure(&self) -> Result<()> {
936 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
937 :
938 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
939 : info!("tuning pgbouncer");
940 :
941 : let rt = tokio::runtime::Builder::new_current_thread()
942 : .enable_all()
943 : .build()
944 : .expect("failed to create rt");
945 :
946 : // Spawn a thread to do the tuning,
947 : // so that we don't block the main thread that starts Postgres.
948 : let pgbouncer_settings = pgbouncer_settings.clone();
949 0 : let _handle = thread::spawn(move || {
950 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
951 0 : if let Err(err) = res {
952 0 : error!("error while tuning pgbouncer: {err:?}");
953 0 : }
954 0 : });
955 : }
956 :
957 : if let Some(ref local_proxy) = spec.local_proxy_config {
958 : info!("configuring local_proxy");
959 :
960 : // Spawn a thread to do the configuration,
961 : // so that we don't block the main thread that starts Postgres.
962 : let local_proxy = local_proxy.clone();
963 0 : let _handle = Some(thread::spawn(move || {
964 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
965 0 : error!("error while configuring local_proxy: {err:?}");
966 0 : }
967 0 : }));
968 : }
969 :
970 : // Write new config
971 : let pgdata_path = Path::new(&self.pgdata);
972 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
973 : config::write_postgres_conf(&postgresql_conf_path, &spec, None)?;
974 : // temporarily reset max_cluster_size in config
975 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
976 : // creating new extensions, roles, etc...
977 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
978 0 : self.pg_reload_conf()?;
979 :
980 0 : let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
981 :
982 : // Proceed with post-startup configuration. Note, that order of operations is important.
983 : // Disable DDL forwarding because control plane already knows about these roles/databases.
984 0 : if spec.mode == ComputeMode::Primary {
985 0 : client.simple_query("SET neon.forward_ddl = false")?;
986 0 : cleanup_instance(&mut client)?;
987 0 : handle_roles(&spec, &mut client)?;
988 0 : handle_databases(&spec, &mut client)?;
989 0 : handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
990 0 : handle_grants(
991 0 : &spec,
992 0 : &mut client,
993 0 : self.connstr.as_str(),
994 0 : self.has_feature(ComputeFeature::AnonExtension),
995 0 : )?;
996 0 : handle_extensions(&spec, &mut client)?;
997 0 : handle_extension_neon(&mut client)?;
998 : // We can skip handle_migrations here because a new migration can only appear
999 : // if we have a new version of the compute_ctl binary, which can only happen
1000 : // if compute got restarted, in which case we'll end up inside of apply_config
1001 : // instead of reconfigure.
1002 0 : }
1003 :
1004 : // 'Close' connection
1005 0 : drop(client);
1006 0 :
1007 0 : Ok(())
1008 0 : })?;
1009 :
1010 : self.pg_reload_conf()?;
1011 :
1012 : let unknown_op = "unknown".to_string();
1013 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1014 : info!(
1015 : "finished reconfiguration of compute node for operation {}",
1016 : op_id
1017 : );
1018 :
1019 : Ok(())
1020 : }
1021 :
1022 0 : #[instrument(skip_all)]
1023 : pub fn start_compute(
1024 : &self,
1025 : extension_server_port: u16,
1026 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
1027 : let compute_state = self.state.lock().unwrap().clone();
1028 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1029 : info!(
1030 : "starting compute for project {}, operation {}, tenant {}, timeline {}",
1031 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
1032 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
1033 : pspec.tenant_id,
1034 : pspec.timeline_id,
1035 : );
1036 :
1037 : // tune pgbouncer
1038 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
1039 : info!("tuning pgbouncer");
1040 :
1041 : let rt = tokio::runtime::Builder::new_current_thread()
1042 : .enable_all()
1043 : .build()
1044 : .expect("failed to create rt");
1045 :
1046 : // Spawn a thread to do the tuning,
1047 : // so that we don't block the main thread that starts Postgres.
1048 : let pgbouncer_settings = pgbouncer_settings.clone();
1049 0 : let _handle = thread::spawn(move || {
1050 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
1051 0 : if let Err(err) = res {
1052 0 : error!("error while tuning pgbouncer: {err:?}");
1053 0 : }
1054 0 : });
1055 : }
1056 :
1057 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
1058 : info!("configuring local_proxy");
1059 :
1060 : // Spawn a thread to do the configuration,
1061 : // so that we don't block the main thread that starts Postgres.
1062 : let local_proxy = local_proxy.clone();
1063 0 : let _handle = thread::spawn(move || {
1064 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1065 0 : error!("error while configuring local_proxy: {err:?}");
1066 0 : }
1067 0 : });
1068 : }
1069 :
1070 : info!(
1071 : "start_compute spec.remote_extensions {:?}",
1072 : pspec.spec.remote_extensions
1073 : );
1074 :
1075 : // This part is sync, because we need to download
1076 : // remote shared_preload_libraries before postgres start (if any)
1077 : if let Some(remote_extensions) = &pspec.spec.remote_extensions {
1078 : // First, create control files for all availale extensions
1079 : extension_server::create_control_files(remote_extensions, &self.pgbin);
1080 :
1081 : let library_load_start_time = Utc::now();
1082 : let remote_ext_metrics = self.prepare_preload_libraries(&pspec.spec)?;
1083 :
1084 : let library_load_time = Utc::now()
1085 : .signed_duration_since(library_load_start_time)
1086 : .to_std()
1087 : .unwrap()
1088 : .as_millis() as u64;
1089 : let mut state = self.state.lock().unwrap();
1090 : state.metrics.load_ext_ms = library_load_time;
1091 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
1092 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
1093 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
1094 : info!(
1095 : "Loading shared_preload_libraries took {:?}ms",
1096 : library_load_time
1097 : );
1098 : info!("{:?}", remote_ext_metrics);
1099 : }
1100 :
1101 : self.prepare_pgdata(&compute_state, extension_server_port)?;
1102 :
1103 : let start_time = Utc::now();
1104 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
1105 :
1106 : let config_time = Utc::now();
1107 : if pspec.spec.mode == ComputeMode::Primary {
1108 : if !pspec.spec.skip_pg_catalog_updates {
1109 : let pgdata_path = Path::new(&self.pgdata);
1110 : // temporarily reset max_cluster_size in config
1111 : // to avoid the possibility of hitting the limit, while we are applying config:
1112 : // creating new extensions, roles, etc...
1113 : config::with_compute_ctl_tmp_override(
1114 : pgdata_path,
1115 : "neon.max_cluster_size=-1",
1116 0 : || {
1117 0 : self.pg_reload_conf()?;
1118 :
1119 0 : self.apply_config(&compute_state)?;
1120 :
1121 0 : Ok(())
1122 0 : },
1123 : )?;
1124 : self.pg_reload_conf()?;
1125 : }
1126 : self.post_apply_config()?;
1127 :
1128 : let connstr = self.connstr.clone();
1129 0 : thread::spawn(move || {
1130 0 : get_installed_extensions_sync(connstr).context("get_installed_extensions")
1131 0 : });
1132 : }
1133 :
1134 : let startup_end_time = Utc::now();
1135 : {
1136 : let mut state = self.state.lock().unwrap();
1137 : state.metrics.start_postgres_ms = config_time
1138 : .signed_duration_since(start_time)
1139 : .to_std()
1140 : .unwrap()
1141 : .as_millis() as u64;
1142 : state.metrics.config_ms = startup_end_time
1143 : .signed_duration_since(config_time)
1144 : .to_std()
1145 : .unwrap()
1146 : .as_millis() as u64;
1147 : state.metrics.total_startup_ms = startup_end_time
1148 : .signed_duration_since(compute_state.start_time)
1149 : .to_std()
1150 : .unwrap()
1151 : .as_millis() as u64;
1152 : }
1153 : self.set_status(ComputeStatus::Running);
1154 :
1155 : info!(
1156 : "finished configuration of compute for project {}",
1157 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
1158 : );
1159 :
1160 : // Log metrics so that we can search for slow operations in logs
1161 : let metrics = {
1162 : let state = self.state.lock().unwrap();
1163 : state.metrics.clone()
1164 : };
1165 : info!(?metrics, "compute start finished");
1166 :
1167 : Ok(pg_process)
1168 : }
1169 :
1170 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1171 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1172 0 : let mut state = self.state.lock().unwrap();
1173 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1174 0 : if last_active > state.last_active {
1175 0 : state.last_active = last_active;
1176 0 : debug!("set the last compute activity time to: {:?}", last_active);
1177 0 : }
1178 0 : }
1179 :
1180 : // Look for core dumps and collect backtraces.
1181 : //
1182 : // EKS worker nodes have following core dump settings:
1183 : // /proc/sys/kernel/core_pattern -> core
1184 : // /proc/sys/kernel/core_uses_pid -> 1
1185 : // ulimit -c -> unlimited
1186 : // which results in core dumps being written to postgres data directory as core.<pid>.
1187 : //
1188 : // Use that as a default location and pattern, except macos where core dumps are written
1189 : // to /cores/ directory by default.
1190 : //
1191 : // With default Linux settings, the core dump file is called just "core", so check for
1192 : // that too.
1193 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1194 0 : let core_dump_dir = match std::env::consts::OS {
1195 0 : "macos" => Path::new("/cores/"),
1196 0 : _ => Path::new(&self.pgdata),
1197 : };
1198 :
1199 : // Collect core dump paths if any
1200 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1201 0 : let files = fs::read_dir(core_dump_dir)?;
1202 0 : let cores = files.filter_map(|entry| {
1203 0 : let entry = entry.ok()?;
1204 :
1205 0 : let is_core_dump = match entry.file_name().to_str()? {
1206 0 : n if n.starts_with("core.") => true,
1207 0 : "core" => true,
1208 0 : _ => false,
1209 : };
1210 0 : if is_core_dump {
1211 0 : Some(entry.path())
1212 : } else {
1213 0 : None
1214 : }
1215 0 : });
1216 :
1217 : // Print backtrace for each core dump
1218 0 : for core_path in cores {
1219 0 : warn!(
1220 0 : "core dump found: {}, collecting backtrace",
1221 0 : core_path.display()
1222 : );
1223 :
1224 : // Try first with gdb
1225 0 : let backtrace = Command::new("gdb")
1226 0 : .args(["--batch", "-q", "-ex", "bt", &self.pgbin])
1227 0 : .arg(&core_path)
1228 0 : .output();
1229 :
1230 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1231 0 : let backtrace = match backtrace {
1232 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1233 0 : warn!("cannot find gdb, trying lldb");
1234 0 : Command::new("lldb")
1235 0 : .arg("-c")
1236 0 : .arg(&core_path)
1237 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1238 0 : .output()
1239 : }
1240 0 : _ => backtrace,
1241 0 : }?;
1242 :
1243 0 : warn!(
1244 0 : "core dump backtrace: {}",
1245 0 : String::from_utf8_lossy(&backtrace.stdout)
1246 : );
1247 0 : warn!(
1248 0 : "debugger stderr: {}",
1249 0 : String::from_utf8_lossy(&backtrace.stderr)
1250 : );
1251 : }
1252 :
1253 0 : Ok(())
1254 0 : }
1255 :
1256 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1257 0 : pub async fn collect_insights(&self) -> String {
1258 0 : let mut result_rows: Vec<String> = Vec::new();
1259 0 : let connect_result = tokio_postgres::connect(self.connstr.as_str(), NoTls).await;
1260 0 : let (client, connection) = connect_result.unwrap();
1261 0 : tokio::spawn(async move {
1262 0 : if let Err(e) = connection.await {
1263 0 : eprintln!("connection error: {}", e);
1264 0 : }
1265 0 : });
1266 0 : let result = client
1267 0 : .simple_query(
1268 0 : "SELECT
1269 0 : row_to_json(pg_stat_statements)
1270 0 : FROM
1271 0 : pg_stat_statements
1272 0 : WHERE
1273 0 : userid != 'cloud_admin'::regrole::oid
1274 0 : ORDER BY
1275 0 : (mean_exec_time + mean_plan_time) DESC
1276 0 : LIMIT 100",
1277 0 : )
1278 0 : .await;
1279 :
1280 0 : if let Ok(raw_rows) = result {
1281 0 : for message in raw_rows.iter() {
1282 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1283 0 : if let Some(json) = row.get(0) {
1284 0 : result_rows.push(json.to_string());
1285 0 : }
1286 0 : }
1287 : }
1288 :
1289 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1290 : } else {
1291 0 : "{{\"pg_stat_statements\": []}}".to_string()
1292 : }
1293 0 : }
1294 :
1295 : // download an archive, unzip and place files in correct locations
1296 0 : pub async fn download_extension(
1297 0 : &self,
1298 0 : real_ext_name: String,
1299 0 : ext_path: RemotePath,
1300 0 : ) -> Result<u64, DownloadError> {
1301 0 : let ext_remote_storage =
1302 0 : self.ext_remote_storage
1303 0 : .as_ref()
1304 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1305 0 : "Remote extensions storage is not configured",
1306 0 : )))?;
1307 :
1308 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1309 0 :
1310 0 : let mut first_try = false;
1311 0 : if !self
1312 0 : .ext_download_progress
1313 0 : .read()
1314 0 : .expect("lock err")
1315 0 : .contains_key(ext_archive_name)
1316 0 : {
1317 0 : self.ext_download_progress
1318 0 : .write()
1319 0 : .expect("lock err")
1320 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1321 0 : first_try = true;
1322 0 : }
1323 0 : let (download_start, download_completed) =
1324 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1325 0 : let start_time_delta = Utc::now()
1326 0 : .signed_duration_since(download_start)
1327 0 : .to_std()
1328 0 : .unwrap()
1329 0 : .as_millis() as u64;
1330 :
1331 : // how long to wait for extension download if it was started by another process
1332 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1333 :
1334 0 : if download_completed {
1335 0 : info!("extension already downloaded, skipping re-download");
1336 0 : return Ok(0);
1337 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1338 0 : info!("download {ext_archive_name} already started by another process, hanging untill completion or timeout");
1339 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1340 : loop {
1341 0 : info!("waiting for download");
1342 0 : interval.tick().await;
1343 0 : let (_, download_completed_now) =
1344 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1345 0 : if download_completed_now {
1346 0 : info!("download finished by whoever else downloaded it");
1347 0 : return Ok(0);
1348 0 : }
1349 : }
1350 : // NOTE: the above loop will get terminated
1351 : // based on the timeout of the download function
1352 0 : }
1353 0 :
1354 0 : // if extension hasn't been downloaded before or the previous
1355 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1356 0 : // then we try to download it here
1357 0 : info!("downloading new extension {ext_archive_name}");
1358 :
1359 0 : let download_size = extension_server::download_extension(
1360 0 : &real_ext_name,
1361 0 : &ext_path,
1362 0 : ext_remote_storage,
1363 0 : &self.pgbin,
1364 0 : )
1365 0 : .await
1366 0 : .map_err(DownloadError::Other);
1367 0 :
1368 0 : if download_size.is_ok() {
1369 0 : self.ext_download_progress
1370 0 : .write()
1371 0 : .expect("bad lock")
1372 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1373 0 : }
1374 :
1375 0 : download_size
1376 0 : }
1377 :
1378 0 : pub async fn set_role_grants(
1379 0 : &self,
1380 0 : db_name: &PgIdent,
1381 0 : schema_name: &PgIdent,
1382 0 : privileges: &[Privilege],
1383 0 : role_name: &PgIdent,
1384 0 : ) -> Result<()> {
1385 : use tokio_postgres::config::Config;
1386 : use tokio_postgres::NoTls;
1387 :
1388 0 : let mut conf = Config::from_str(self.connstr.as_str()).unwrap();
1389 0 : conf.dbname(db_name);
1390 :
1391 0 : let (db_client, conn) = conf
1392 0 : .connect(NoTls)
1393 0 : .await
1394 0 : .context("Failed to connect to the database")?;
1395 0 : tokio::spawn(conn);
1396 0 :
1397 0 : // TODO: support other types of grants apart from schemas?
1398 0 : let query = format!(
1399 0 : "GRANT {} ON SCHEMA {} TO {}",
1400 0 : privileges
1401 0 : .iter()
1402 0 : // should not be quoted as it's part of the command.
1403 0 : // is already sanitized so it's ok
1404 0 : .map(|p| p.as_str())
1405 0 : .collect::<Vec<&'static str>>()
1406 0 : .join(", "),
1407 0 : // quote the schema and role name as identifiers to sanitize them.
1408 0 : schema_name.pg_quote(),
1409 0 : role_name.pg_quote(),
1410 0 : );
1411 0 : db_client
1412 0 : .simple_query(&query)
1413 0 : .await
1414 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1415 :
1416 0 : Ok(())
1417 0 : }
1418 :
1419 0 : pub async fn install_extension(
1420 0 : &self,
1421 0 : ext_name: &PgIdent,
1422 0 : db_name: &PgIdent,
1423 0 : ext_version: ExtVersion,
1424 0 : ) -> Result<ExtVersion> {
1425 : use tokio_postgres::config::Config;
1426 : use tokio_postgres::NoTls;
1427 :
1428 0 : let mut conf = Config::from_str(self.connstr.as_str()).unwrap();
1429 0 : conf.dbname(db_name);
1430 :
1431 0 : let (db_client, conn) = conf
1432 0 : .connect(NoTls)
1433 0 : .await
1434 0 : .context("Failed to connect to the database")?;
1435 0 : tokio::spawn(conn);
1436 0 :
1437 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
1438 0 : let version: Option<ExtVersion> = db_client
1439 0 : .query_opt(version_query, &[&ext_name])
1440 0 : .await
1441 0 : .with_context(|| format!("Failed to execute query: {}", version_query))?
1442 0 : .map(|row| row.get(0));
1443 0 :
1444 0 : // sanitize the inputs as postgres idents.
1445 0 : let ext_name: String = ext_name.pg_quote();
1446 0 : let quoted_version: String = ext_version.pg_quote();
1447 :
1448 0 : if let Some(installed_version) = version {
1449 0 : if installed_version == ext_version {
1450 0 : return Ok(installed_version);
1451 0 : }
1452 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
1453 0 : db_client
1454 0 : .simple_query(&query)
1455 0 : .await
1456 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1457 : } else {
1458 0 : let query =
1459 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
1460 0 : db_client
1461 0 : .simple_query(&query)
1462 0 : .await
1463 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1464 : }
1465 :
1466 0 : Ok(ext_version)
1467 0 : }
1468 :
1469 : #[tokio::main]
1470 0 : pub async fn prepare_preload_libraries(
1471 0 : &self,
1472 0 : spec: &ComputeSpec,
1473 0 : ) -> Result<RemoteExtensionMetrics> {
1474 0 : if self.ext_remote_storage.is_none() {
1475 0 : return Ok(RemoteExtensionMetrics {
1476 0 : num_ext_downloaded: 0,
1477 0 : largest_ext_size: 0,
1478 0 : total_ext_download_size: 0,
1479 0 : });
1480 0 : }
1481 0 : let remote_extensions = spec
1482 0 : .remote_extensions
1483 0 : .as_ref()
1484 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
1485 0 :
1486 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
1487 0 : let mut libs_vec = Vec::new();
1488 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
1489 0 : libs_vec = libs
1490 0 : .split(&[',', '\'', ' '])
1491 0 : .filter(|s| *s != "neon" && !s.is_empty())
1492 0 : .map(str::to_string)
1493 0 : .collect();
1494 0 : }
1495 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
1496 0 :
1497 0 : // that is used in neon_local and python tests
1498 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
1499 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
1500 0 : let mut shared_preload_libraries_line = "";
1501 0 : for line in conf_lines {
1502 0 : if line.starts_with("shared_preload_libraries") {
1503 0 : shared_preload_libraries_line = line;
1504 0 : }
1505 0 : }
1506 0 : let mut preload_libs_vec = Vec::new();
1507 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
1508 0 : preload_libs_vec = libs
1509 0 : .split(&[',', '\'', ' '])
1510 0 : .filter(|s| *s != "neon" && !s.is_empty())
1511 0 : .map(str::to_string)
1512 0 : .collect();
1513 0 : }
1514 0 : libs_vec.extend(preload_libs_vec);
1515 0 : }
1516 0 :
1517 0 : // Don't try to download libraries that are not in the index.
1518 0 : // Assume that they are already present locally.
1519 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
1520 0 :
1521 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
1522 0 :
1523 0 : let mut download_tasks = Vec::new();
1524 0 : for library in &libs_vec {
1525 0 : let (ext_name, ext_path) =
1526 0 : remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
1527 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
1528 0 : }
1529 0 : let results = join_all(download_tasks).await;
1530 0 :
1531 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
1532 0 : num_ext_downloaded: 0,
1533 0 : largest_ext_size: 0,
1534 0 : total_ext_download_size: 0,
1535 0 : };
1536 0 : for result in results {
1537 0 : let download_size = match result {
1538 0 : Ok(res) => {
1539 0 : remote_ext_metrics.num_ext_downloaded += 1;
1540 0 : res
1541 0 : }
1542 0 : Err(err) => {
1543 0 : // if we failed to download an extension, we don't want to fail the whole
1544 0 : // process, but we do want to log the error
1545 0 : error!("Failed to download extension: {}", err);
1546 0 : 0
1547 0 : }
1548 0 : };
1549 0 :
1550 0 : remote_ext_metrics.largest_ext_size =
1551 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
1552 0 : remote_ext_metrics.total_ext_download_size += download_size;
1553 0 : }
1554 0 : Ok(remote_ext_metrics)
1555 0 : }
1556 :
1557 : /// Waits until current thread receives a state changed notification and
1558 : /// the pageserver connection strings has changed.
1559 : ///
1560 : /// The operation will time out after a specified duration.
1561 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
1562 0 : let state = self.state.lock().unwrap();
1563 0 : let old_pageserver_connstr = state
1564 0 : .pspec
1565 0 : .as_ref()
1566 0 : .expect("spec must be set")
1567 0 : .pageserver_connstr
1568 0 : .clone();
1569 0 : let mut unchanged = true;
1570 0 : let _ = self
1571 0 : .state_changed
1572 0 : .wait_timeout_while(state, duration, |s| {
1573 0 : let pageserver_connstr = &s
1574 0 : .pspec
1575 0 : .as_ref()
1576 0 : .expect("spec must be set")
1577 0 : .pageserver_connstr;
1578 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
1579 0 : unchanged
1580 0 : })
1581 0 : .unwrap();
1582 0 : if !unchanged {
1583 0 : info!("Pageserver config changed");
1584 0 : }
1585 0 : }
1586 : }
1587 :
1588 0 : pub fn forward_termination_signal() {
1589 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
1590 0 : if ss_pid != 0 {
1591 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
1592 0 : kill(ss_pid, Signal::SIGTERM).ok();
1593 0 : }
1594 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
1595 0 : if pg_pid != 0 {
1596 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
1597 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
1598 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
1599 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
1600 0 : kill(pg_pid, Signal::SIGINT).ok();
1601 0 : }
1602 0 : }
|
