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::Instant;
14 :
15 : use anyhow::{Context, Result};
16 : use chrono::{DateTime, Utc};
17 : use futures::future::join_all;
18 : use futures::stream::FuturesUnordered;
19 : use futures::StreamExt;
20 : use nix::unistd::Pid;
21 : use postgres::error::SqlState;
22 : use postgres::{Client, NoTls};
23 : use tracing::{debug, error, info, instrument, warn};
24 : use utils::id::{TenantId, TimelineId};
25 : use utils::lsn::Lsn;
26 :
27 : use compute_api::responses::{ComputeMetrics, ComputeStatus};
28 : use compute_api::spec::{ComputeFeature, ComputeMode, ComputeSpec};
29 : use utils::measured_stream::MeasuredReader;
30 :
31 : use nix::sys::signal::{kill, Signal};
32 :
33 : use remote_storage::{DownloadError, RemotePath};
34 :
35 : use crate::checker::create_availability_check_data;
36 : use crate::logger::inlinify;
37 : use crate::pg_helpers::*;
38 : use crate::spec::*;
39 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
40 : use crate::{config, extension_server};
41 :
42 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
43 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
44 :
45 : /// Compute node info shared across several `compute_ctl` threads.
46 : pub struct ComputeNode {
47 : // Url type maintains proper escaping
48 : pub connstr: url::Url,
49 : pub pgdata: String,
50 : pub pgbin: String,
51 : pub pgversion: String,
52 : /// We should only allow live re- / configuration of the compute node if
53 : /// it uses 'pull model', i.e. it can go to control-plane and fetch
54 : /// the latest configuration. Otherwise, there could be a case:
55 : /// - we start compute with some spec provided as argument
56 : /// - we push new spec and it does reconfiguration
57 : /// - but then something happens and compute pod / VM is destroyed,
58 : /// so k8s controller starts it again with the **old** spec
59 : ///
60 : /// and the same for empty computes:
61 : /// - we started compute without any spec
62 : /// - we push spec and it does configuration
63 : /// - but then it is restarted without any spec again
64 : pub live_config_allowed: bool,
65 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
66 : /// To allow HTTP API server to serving status requests, while configuration
67 : /// is in progress, lock should be held only for short periods of time to do
68 : /// read/write, not the whole configuration process.
69 : pub state: Mutex<ComputeState>,
70 : /// `Condvar` to allow notifying waiters about state changes.
71 : pub state_changed: Condvar,
72 : /// the address of extension storage proxy gateway
73 : pub ext_remote_storage: Option<String>,
74 : // key: ext_archive_name, value: started download time, download_completed?
75 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
76 : pub build_tag: String,
77 : }
78 :
79 : // store some metrics about download size that might impact startup time
80 : #[derive(Clone, Debug)]
81 : pub struct RemoteExtensionMetrics {
82 : num_ext_downloaded: u64,
83 : largest_ext_size: u64,
84 : total_ext_download_size: u64,
85 : }
86 :
87 : #[derive(Clone, Debug)]
88 : pub struct ComputeState {
89 : pub start_time: DateTime<Utc>,
90 : pub status: ComputeStatus,
91 : /// Timestamp of the last Postgres activity. It could be `None` if
92 : /// compute wasn't used since start.
93 : pub last_active: Option<DateTime<Utc>>,
94 : pub error: Option<String>,
95 : pub pspec: Option<ParsedSpec>,
96 : pub metrics: ComputeMetrics,
97 : }
98 :
99 : impl ComputeState {
100 0 : pub fn new() -> Self {
101 0 : Self {
102 0 : start_time: Utc::now(),
103 0 : status: ComputeStatus::Empty,
104 0 : last_active: None,
105 0 : error: None,
106 0 : pspec: None,
107 0 : metrics: ComputeMetrics::default(),
108 0 : }
109 0 : }
110 : }
111 :
112 : impl Default for ComputeState {
113 0 : fn default() -> Self {
114 0 : Self::new()
115 0 : }
116 : }
117 :
118 : #[derive(Clone, Debug)]
119 : pub struct ParsedSpec {
120 : pub spec: ComputeSpec,
121 : pub tenant_id: TenantId,
122 : pub timeline_id: TimelineId,
123 : pub pageserver_connstr: String,
124 : pub safekeeper_connstrings: Vec<String>,
125 : pub storage_auth_token: Option<String>,
126 : }
127 :
128 : impl TryFrom<ComputeSpec> for ParsedSpec {
129 : type Error = String;
130 0 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
131 : // Extract the options from the spec file that are needed to connect to
132 : // the storage system.
133 : //
134 : // For backwards-compatibility, the top-level fields in the spec file
135 : // may be empty. In that case, we need to dig them from the GUCs in the
136 : // cluster.settings field.
137 0 : let pageserver_connstr = spec
138 0 : .pageserver_connstring
139 0 : .clone()
140 0 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
141 0 : .ok_or("pageserver connstr should be provided")?;
142 0 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
143 0 : if matches!(spec.mode, ComputeMode::Primary) {
144 0 : spec.cluster
145 0 : .settings
146 0 : .find("neon.safekeepers")
147 0 : .ok_or("safekeeper connstrings should be provided")?
148 0 : .split(',')
149 0 : .map(|str| str.to_string())
150 0 : .collect()
151 : } else {
152 0 : vec![]
153 : }
154 : } else {
155 0 : spec.safekeeper_connstrings.clone()
156 : };
157 0 : let storage_auth_token = spec.storage_auth_token.clone();
158 0 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
159 0 : tenant_id
160 : } else {
161 0 : spec.cluster
162 0 : .settings
163 0 : .find("neon.tenant_id")
164 0 : .ok_or("tenant id should be provided")
165 0 : .map(|s| TenantId::from_str(&s))?
166 0 : .or(Err("invalid tenant id"))?
167 : };
168 0 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
169 0 : timeline_id
170 : } else {
171 0 : spec.cluster
172 0 : .settings
173 0 : .find("neon.timeline_id")
174 0 : .ok_or("timeline id should be provided")
175 0 : .map(|s| TimelineId::from_str(&s))?
176 0 : .or(Err("invalid timeline id"))?
177 : };
178 :
179 0 : Ok(ParsedSpec {
180 0 : spec,
181 0 : pageserver_connstr,
182 0 : safekeeper_connstrings,
183 0 : storage_auth_token,
184 0 : tenant_id,
185 0 : timeline_id,
186 0 : })
187 0 : }
188 : }
189 :
190 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
191 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
192 : ///
193 : /// This function should be used to start postgres, as it will start it in the
194 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
195 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
196 : /// creates it during the sysinit phase of its inittab.
197 0 : fn maybe_cgexec(cmd: &str) -> Command {
198 0 : // The cplane sets this env var for autoscaling computes.
199 0 : // use `var_os` so we don't have to worry about the variable being valid
200 0 : // unicode. Should never be an concern . . . but just in case
201 0 : if env::var_os("AUTOSCALING").is_some() {
202 0 : let mut command = Command::new("cgexec");
203 0 : command.args(["-g", "memory:neon-postgres"]);
204 0 : command.arg(cmd);
205 0 : command
206 : } else {
207 0 : Command::new(cmd)
208 : }
209 0 : }
210 :
211 : /// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
212 : /// that we give to customers
213 0 : #[instrument(skip_all)]
214 : fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
215 : let roles = spec
216 : .cluster
217 : .roles
218 : .iter()
219 0 : .map(|r| escape_literal(&r.name))
220 : .collect::<Vec<_>>();
221 :
222 : let dbs = spec
223 : .cluster
224 : .databases
225 : .iter()
226 0 : .map(|db| escape_literal(&db.name))
227 : .collect::<Vec<_>>();
228 :
229 : let roles_decl = if roles.is_empty() {
230 : String::from("roles text[] := NULL;")
231 : } else {
232 : format!(
233 : r#"
234 : roles text[] := ARRAY(SELECT rolname
235 : FROM pg_catalog.pg_roles
236 : WHERE rolname IN ({}));"#,
237 : roles.join(", ")
238 : )
239 : };
240 :
241 : let database_decl = if dbs.is_empty() {
242 : String::from("dbs text[] := NULL;")
243 : } else {
244 : format!(
245 : r#"
246 : dbs text[] := ARRAY(SELECT datname
247 : FROM pg_catalog.pg_database
248 : WHERE datname IN ({}));"#,
249 : dbs.join(", ")
250 : )
251 : };
252 :
253 : // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on all databases
254 : // (see https://www.postgresql.org/docs/current/ddl-priv.html)
255 : let query = format!(
256 : r#"
257 : DO $$
258 : DECLARE
259 : r text;
260 : {}
261 : {}
262 : BEGIN
263 : IF NOT EXISTS (
264 : SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
265 : THEN
266 : CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
267 : IF array_length(roles, 1) IS NOT NULL THEN
268 : EXECUTE format('GRANT neon_superuser TO %s',
269 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
270 : FOREACH r IN ARRAY roles LOOP
271 : EXECUTE format('ALTER ROLE %s CREATEROLE CREATEDB', quote_ident(r));
272 : END LOOP;
273 : END IF;
274 : IF array_length(dbs, 1) IS NOT NULL THEN
275 : EXECUTE format('GRANT ALL PRIVILEGES ON DATABASE %s TO neon_superuser',
276 : array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(dbs) as x), ', '));
277 : END IF;
278 : END IF;
279 : END
280 : $$;"#,
281 : roles_decl, database_decl,
282 : );
283 : info!("Neon superuser created: {}", inlinify(&query));
284 : client
285 : .simple_query(&query)
286 0 : .map_err(|e| anyhow::anyhow!(e).context(query))?;
287 : Ok(())
288 : }
289 :
290 : impl ComputeNode {
291 : /// Check that compute node has corresponding feature enabled.
292 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
293 0 : let state = self.state.lock().unwrap();
294 :
295 0 : if let Some(s) = state.pspec.as_ref() {
296 0 : s.spec.features.contains(&feature)
297 : } else {
298 0 : false
299 : }
300 0 : }
301 :
302 0 : pub fn set_status(&self, status: ComputeStatus) {
303 0 : let mut state = self.state.lock().unwrap();
304 0 : state.status = status;
305 0 : self.state_changed.notify_all();
306 0 : }
307 :
308 0 : pub fn get_status(&self) -> ComputeStatus {
309 0 : self.state.lock().unwrap().status
310 0 : }
311 :
312 : // Remove `pgdata` directory and create it again with right permissions.
313 0 : fn create_pgdata(&self) -> Result<()> {
314 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
315 0 : // If it is something different then create_dir() will error out anyway.
316 0 : let _ok = fs::remove_dir_all(&self.pgdata);
317 0 : fs::create_dir(&self.pgdata)?;
318 0 : fs::set_permissions(&self.pgdata, fs::Permissions::from_mode(0o700))?;
319 :
320 0 : Ok(())
321 0 : }
322 :
323 : // Get basebackup from the libpq connection to pageserver using `connstr` and
324 : // unarchive it to `pgdata` directory overriding all its previous content.
325 0 : #[instrument(skip_all, fields(%lsn))]
326 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
327 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
328 : let start_time = Instant::now();
329 :
330 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
331 : let mut config = postgres::Config::from_str(shard0_connstr)?;
332 :
333 : // Use the storage auth token from the config file, if given.
334 : // Note: this overrides any password set in the connection string.
335 : if let Some(storage_auth_token) = &spec.storage_auth_token {
336 : info!("Got storage auth token from spec file");
337 : config.password(storage_auth_token);
338 : } else {
339 : info!("Storage auth token not set");
340 : }
341 :
342 : // Connect to pageserver
343 : let mut client = config.connect(NoTls)?;
344 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
345 :
346 : let basebackup_cmd = match lsn {
347 : // HACK We don't use compression on first start (Lsn(0)) because there's no API for it
348 : Lsn(0) => format!("basebackup {} {}", spec.tenant_id, spec.timeline_id),
349 : _ => format!(
350 : "basebackup {} {} {} --gzip",
351 : spec.tenant_id, spec.timeline_id, lsn
352 : ),
353 : };
354 :
355 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
356 : let mut measured_reader = MeasuredReader::new(copyreader);
357 :
358 : // Check the magic number to see if it's a gzip or not. Even though
359 : // we might explicitly ask for gzip, an old pageserver with no implementation
360 : // of gzip compression might send us uncompressed data. After some time
361 : // passes we can assume all pageservers know how to compress and we can
362 : // delete this check.
363 : //
364 : // If the data is not gzip, it will be tar. It will not be mistakenly
365 : // recognized as gzip because tar starts with an ascii encoding of a filename,
366 : // and 0x1f and 0x8b are unlikely first characters for any filename. Moreover,
367 : // we send the "global" directory first from the pageserver, so it definitely
368 : // won't be recognized as gzip.
369 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
370 : let gzip = {
371 : let peek = bufreader.fill_buf().unwrap();
372 : peek[0] == 0x1f && peek[1] == 0x8b
373 : };
374 :
375 : // Read the archive directly from the `CopyOutReader`
376 : //
377 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
378 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
379 : // sends an Error after finishing the tarball, we will not notice it.
380 : if gzip {
381 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
382 : ar.set_ignore_zeros(true);
383 : ar.unpack(&self.pgdata)?;
384 : } else {
385 : let mut ar = tar::Archive::new(&mut bufreader);
386 : ar.set_ignore_zeros(true);
387 : ar.unpack(&self.pgdata)?;
388 : };
389 :
390 : // Report metrics
391 : let mut state = self.state.lock().unwrap();
392 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
393 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
394 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
395 : Ok(())
396 : }
397 :
398 : // Gets the basebackup in a retry loop
399 0 : #[instrument(skip_all, fields(%lsn))]
400 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
401 : let mut retry_period_ms = 500.0;
402 : let mut attempts = 0;
403 : const DEFAULT_ATTEMPTS: u16 = 10;
404 : #[cfg(feature = "testing")]
405 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
406 : u16::from_str(&v).unwrap()
407 : } else {
408 : DEFAULT_ATTEMPTS
409 : };
410 : #[cfg(not(feature = "testing"))]
411 : let max_attempts = DEFAULT_ATTEMPTS;
412 : loop {
413 : let result = self.try_get_basebackup(compute_state, lsn);
414 : match result {
415 : Ok(_) => {
416 : return result;
417 : }
418 : Err(ref e) if attempts < max_attempts => {
419 : warn!(
420 : "Failed to get basebackup: {} (attempt {}/{})",
421 : e, attempts, max_attempts
422 : );
423 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
424 : retry_period_ms *= 1.5;
425 : }
426 : Err(_) => {
427 : return result;
428 : }
429 : }
430 : attempts += 1;
431 : }
432 : }
433 :
434 0 : pub async fn check_safekeepers_synced_async(
435 0 : &self,
436 0 : compute_state: &ComputeState,
437 0 : ) -> Result<Option<Lsn>> {
438 0 : // Construct a connection config for each safekeeper
439 0 : let pspec: ParsedSpec = compute_state
440 0 : .pspec
441 0 : .as_ref()
442 0 : .expect("spec must be set")
443 0 : .clone();
444 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
445 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
446 0 : // Format connstr
447 0 : let id = connstr.clone();
448 0 : let connstr = format!("postgresql://no_user@{}", connstr);
449 0 : let options = format!(
450 0 : "-c timeline_id={} tenant_id={}",
451 0 : pspec.timeline_id, pspec.tenant_id
452 0 : );
453 0 :
454 0 : // Construct client
455 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
456 0 : config.options(&options);
457 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
458 0 : config.password(storage_auth_token);
459 0 : }
460 :
461 0 : (id, config)
462 0 : });
463 0 :
464 0 : // Create task set to query all safekeepers
465 0 : let mut tasks = FuturesUnordered::new();
466 0 : let quorum = sk_configs.len() / 2 + 1;
467 0 : for (id, config) in sk_configs {
468 0 : let timeout = tokio::time::Duration::from_millis(100);
469 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
470 0 : tasks.push(tokio::spawn(task));
471 0 : }
472 :
473 : // Get a quorum of responses or errors
474 0 : let mut responses = Vec::new();
475 0 : let mut join_errors = Vec::new();
476 0 : let mut task_errors = Vec::new();
477 0 : let mut timeout_errors = Vec::new();
478 0 : while let Some(response) = tasks.next().await {
479 0 : match response {
480 0 : Ok(Ok(Ok(r))) => responses.push(r),
481 0 : Ok(Ok(Err(e))) => task_errors.push(e),
482 0 : Ok(Err(e)) => timeout_errors.push(e),
483 0 : Err(e) => join_errors.push(e),
484 : };
485 0 : if responses.len() >= quorum {
486 0 : break;
487 0 : }
488 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
489 0 : break;
490 0 : }
491 : }
492 :
493 : // In case of error, log and fail the check, but don't crash.
494 : // We're playing it safe because these errors could be transient
495 : // and we don't yet retry. Also being careful here allows us to
496 : // be backwards compatible with safekeepers that don't have the
497 : // TIMELINE_STATUS API yet.
498 0 : if responses.len() < quorum {
499 0 : error!(
500 0 : "failed sync safekeepers check {:?} {:?} {:?}",
501 : join_errors, task_errors, timeout_errors
502 : );
503 0 : return Ok(None);
504 0 : }
505 0 :
506 0 : Ok(check_if_synced(responses))
507 0 : }
508 :
509 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
510 : // in one roundtrip. If not, we should do a full sync_safekeepers.
511 0 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
512 0 : let start_time = Utc::now();
513 0 :
514 0 : // Run actual work with new tokio runtime
515 0 : let rt = tokio::runtime::Builder::new_current_thread()
516 0 : .enable_all()
517 0 : .build()
518 0 : .expect("failed to create rt");
519 0 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
520 0 :
521 0 : // Record runtime
522 0 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
523 0 : .signed_duration_since(start_time)
524 0 : .to_std()
525 0 : .unwrap()
526 0 : .as_millis() as u64;
527 0 : result
528 0 : }
529 :
530 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
531 : // and return the reported LSN back to the caller.
532 0 : #[instrument(skip_all)]
533 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
534 : let start_time = Utc::now();
535 :
536 : let mut sync_handle = maybe_cgexec(&self.pgbin)
537 : .args(["--sync-safekeepers"])
538 : .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
539 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
540 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
541 : } else {
542 : vec![]
543 : })
544 : .stdout(Stdio::piped())
545 : .stderr(Stdio::piped())
546 : .spawn()
547 : .expect("postgres --sync-safekeepers failed to start");
548 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
549 :
550 : // `postgres --sync-safekeepers` will print all log output to stderr and
551 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
552 : // will be collected in a child thread.
553 : let stderr = sync_handle
554 : .stderr
555 : .take()
556 : .expect("stderr should be captured");
557 : let logs_handle = handle_postgres_logs(stderr);
558 :
559 : let sync_output = sync_handle
560 : .wait_with_output()
561 : .expect("postgres --sync-safekeepers failed");
562 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
563 :
564 : // Process has exited, so we can join the logs thread.
565 : let _ = logs_handle
566 : .join()
567 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
568 :
569 : if !sync_output.status.success() {
570 : anyhow::bail!(
571 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
572 : sync_output.status,
573 : String::from_utf8(sync_output.stdout)
574 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
575 : );
576 : }
577 :
578 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
579 : .signed_duration_since(start_time)
580 : .to_std()
581 : .unwrap()
582 : .as_millis() as u64;
583 :
584 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
585 :
586 : Ok(lsn)
587 : }
588 :
589 : /// Do all the preparations like PGDATA directory creation, configuration,
590 : /// safekeepers sync, basebackup, etc.
591 0 : #[instrument(skip_all)]
592 : pub fn prepare_pgdata(
593 : &self,
594 : compute_state: &ComputeState,
595 : extension_server_port: u16,
596 : ) -> Result<()> {
597 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
598 : let spec = &pspec.spec;
599 : let pgdata_path = Path::new(&self.pgdata);
600 :
601 : // Remove/create an empty pgdata directory and put configuration there.
602 : self.create_pgdata()?;
603 : config::write_postgres_conf(
604 : &pgdata_path.join("postgresql.conf"),
605 : &pspec.spec,
606 : Some(extension_server_port),
607 : )?;
608 :
609 : // Syncing safekeepers is only safe with primary nodes: if a primary
610 : // is already connected it will be kicked out, so a secondary (standby)
611 : // cannot sync safekeepers.
612 : let lsn = match spec.mode {
613 : ComputeMode::Primary => {
614 : info!("checking if safekeepers are synced");
615 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
616 : lsn
617 : } else {
618 : info!("starting safekeepers syncing");
619 : self.sync_safekeepers(pspec.storage_auth_token.clone())
620 0 : .with_context(|| "failed to sync safekeepers")?
621 : };
622 : info!("safekeepers synced at LSN {}", lsn);
623 : lsn
624 : }
625 : ComputeMode::Static(lsn) => {
626 : info!("Starting read-only node at static LSN {}", lsn);
627 : lsn
628 : }
629 : ComputeMode::Replica => {
630 : info!("Initializing standby from latest Pageserver LSN");
631 : Lsn(0)
632 : }
633 : };
634 :
635 : info!(
636 : "getting basebackup@{} from pageserver {}",
637 : lsn, &pspec.pageserver_connstr
638 : );
639 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
640 0 : format!(
641 0 : "failed to get basebackup@{} from pageserver {}",
642 0 : lsn, &pspec.pageserver_connstr
643 0 : )
644 0 : })?;
645 :
646 : // Update pg_hba.conf received with basebackup.
647 : update_pg_hba(pgdata_path)?;
648 :
649 : // Place pg_dynshmem under /dev/shm. This allows us to use
650 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
651 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
652 : //
653 : // Why on earth don't we just stick to the 'posix' default, you might
654 : // ask. It turns out that making large allocations with 'posix' doesn't
655 : // work very well with autoscaling. The behavior we want is that:
656 : //
657 : // 1. You can make large DSM allocations, larger than the current RAM
658 : // size of the VM, without errors
659 : //
660 : // 2. If the allocated memory is really used, the VM is scaled up
661 : // automatically to accommodate that
662 : //
663 : // We try to make that possible by having swap in the VM. But with the
664 : // default 'posix' DSM implementation, we fail step 1, even when there's
665 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
666 : // the shmem segment, which is really just a file in /dev/shm in Linux,
667 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
668 : // than available RAM.
669 : //
670 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
671 : // the Postgres 'mmap' DSM implementation doesn't use
672 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
673 : // fill the file with zeros. It's weird that that differs between
674 : // 'posix' and 'mmap', but we take advantage of it. When the file is
675 : // filled slowly with write(2), the kernel allows it to grow larger, as
676 : // long as there's swap available.
677 : //
678 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
679 : // segment to be larger than currently available RAM. But because we
680 : // don't want to store it on a real file, which the kernel would try to
681 : // flush to disk, so symlink pg_dynshm to /dev/shm.
682 : //
683 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
684 : // control plane control that option. If 'mmap' is not used, this
685 : // symlink doesn't affect anything.
686 : //
687 : // See https://github.com/neondatabase/autoscaling/issues/800
688 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
689 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
690 :
691 : match spec.mode {
692 : ComputeMode::Primary => {}
693 : ComputeMode::Replica | ComputeMode::Static(..) => {
694 : add_standby_signal(pgdata_path)?;
695 : }
696 : }
697 :
698 : Ok(())
699 : }
700 :
701 : /// Start and stop a postgres process to warm up the VM for startup.
702 0 : pub fn prewarm_postgres(&self) -> Result<()> {
703 0 : info!("prewarming");
704 :
705 : // Create pgdata
706 0 : let pgdata = &format!("{}.warmup", self.pgdata);
707 0 : create_pgdata(pgdata)?;
708 :
709 : // Run initdb to completion
710 0 : info!("running initdb");
711 0 : let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
712 0 : Command::new(initdb_bin)
713 0 : .args(["-D", pgdata])
714 0 : .output()
715 0 : .expect("cannot start initdb process");
716 :
717 : // Write conf
718 : use std::io::Write;
719 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
720 0 : let mut file = std::fs::File::create(conf_path)?;
721 0 : writeln!(file, "shared_buffers=65536")?;
722 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
723 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
724 :
725 : // Start postgres
726 0 : info!("starting postgres");
727 0 : let mut pg = maybe_cgexec(&self.pgbin)
728 0 : .args(["-D", pgdata])
729 0 : .spawn()
730 0 : .expect("cannot start postgres process");
731 0 :
732 0 : // Stop it when it's ready
733 0 : info!("waiting for postgres");
734 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
735 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
736 : // it to avoid orphaned processes prowling around while datadir is
737 : // wiped.
738 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
739 0 : kill(pm_pid, Signal::SIGQUIT)?;
740 0 : info!("sent SIGQUIT signal");
741 0 : pg.wait()?;
742 0 : info!("done prewarming");
743 :
744 : // clean up
745 0 : let _ok = fs::remove_dir_all(pgdata);
746 0 : Ok(())
747 0 : }
748 :
749 : /// Start Postgres as a child process and manage DBs/roles.
750 : /// After that this will hang waiting on the postmaster process to exit.
751 : /// Returns a handle to the child process and a handle to the logs thread.
752 0 : #[instrument(skip_all)]
753 : pub fn start_postgres(
754 : &self,
755 : storage_auth_token: Option<String>,
756 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
757 : let pgdata_path = Path::new(&self.pgdata);
758 :
759 : // Run postgres as a child process.
760 : let mut pg = maybe_cgexec(&self.pgbin)
761 : .args(["-D", &self.pgdata])
762 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
763 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
764 : } else {
765 : vec![]
766 : })
767 : .stderr(Stdio::piped())
768 : .spawn()
769 : .expect("cannot start postgres process");
770 : PG_PID.store(pg.id(), Ordering::SeqCst);
771 :
772 : // Start a thread to collect logs from stderr.
773 : let stderr = pg.stderr.take().expect("stderr should be captured");
774 : let logs_handle = handle_postgres_logs(stderr);
775 :
776 : wait_for_postgres(&mut pg, pgdata_path)?;
777 :
778 : Ok((pg, logs_handle))
779 : }
780 :
781 : /// Do post configuration of the already started Postgres. This function spawns a background thread to
782 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
783 : /// version. In the future, it may upgrade all 3rd-party extensions.
784 0 : #[instrument(skip_all)]
785 : pub fn post_apply_config(&self) -> Result<()> {
786 : let connstr = self.connstr.clone();
787 0 : thread::spawn(move || {
788 0 : let func = || {
789 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
790 0 : handle_neon_extension_upgrade(&mut client)
791 0 : .context("handle_neon_extension_upgrade")?;
792 0 : Ok::<_, anyhow::Error>(())
793 0 : };
794 0 : if let Err(err) = func() {
795 0 : error!("error while post_apply_config: {err:#}");
796 0 : }
797 0 : });
798 : Ok(())
799 : }
800 :
801 : /// Do initial configuration of the already started Postgres.
802 0 : #[instrument(skip_all)]
803 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
804 : // If connection fails,
805 : // it may be the old node with `zenith_admin` superuser.
806 : //
807 : // In this case we need to connect with old `zenith_admin` name
808 : // and create new user. We cannot simply rename connected user,
809 : // but we can create a new one and grant it all privileges.
810 : let mut connstr = self.connstr.clone();
811 : connstr
812 : .query_pairs_mut()
813 : .append_pair("application_name", "apply_config");
814 :
815 : let mut client = match Client::connect(connstr.as_str(), NoTls) {
816 : Err(e) => match e.code() {
817 : Some(&SqlState::INVALID_PASSWORD)
818 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
819 : // connect with zenith_admin if cloud_admin could not authenticate
820 : info!(
821 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
822 : e
823 : );
824 : let mut zenith_admin_connstr = connstr.clone();
825 :
826 : zenith_admin_connstr
827 : .set_username("zenith_admin")
828 0 : .map_err(|_| anyhow::anyhow!("invalid connstr"))?;
829 :
830 : let mut client =
831 : Client::connect(zenith_admin_connstr.as_str(), NoTls)
832 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
833 : // Disable forwarding so that users don't get a cloud_admin role
834 :
835 0 : let mut func = || {
836 0 : client.simple_query("SET neon.forward_ddl = false")?;
837 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
838 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
839 0 : Ok::<_, anyhow::Error>(())
840 0 : };
841 : func().context("apply_config setup cloud_admin")?;
842 :
843 : drop(client);
844 :
845 : // reconnect with connstring with expected name
846 : Client::connect(connstr.as_str(), NoTls)?
847 : }
848 : _ => return Err(e.into()),
849 : },
850 : Ok(client) => client,
851 : };
852 :
853 : // Disable DDL forwarding because control plane already knows about these roles/databases.
854 : client
855 : .simple_query("SET neon.forward_ddl = false")
856 : .context("apply_config SET neon.forward_ddl = false")?;
857 :
858 : // Proceed with post-startup configuration. Note, that order of operations is important.
859 : let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
860 : create_neon_superuser(spec, &mut client).context("apply_config create_neon_superuser")?;
861 : cleanup_instance(&mut client).context("apply_config cleanup_instance")?;
862 : handle_roles(spec, &mut client).context("apply_config handle_roles")?;
863 : handle_databases(spec, &mut client).context("apply_config handle_databases")?;
864 : handle_role_deletions(spec, connstr.as_str(), &mut client)
865 : .context("apply_config handle_role_deletions")?;
866 : handle_grants(
867 : spec,
868 : &mut client,
869 : connstr.as_str(),
870 : self.has_feature(ComputeFeature::AnonExtension),
871 : )
872 : .context("apply_config handle_grants")?;
873 : handle_extensions(spec, &mut client).context("apply_config handle_extensions")?;
874 : handle_extension_neon(&mut client).context("apply_config handle_extension_neon")?;
875 : create_availability_check_data(&mut client)
876 : .context("apply_config create_availability_check_data")?;
877 :
878 : // 'Close' connection
879 : drop(client);
880 :
881 : // Run migrations separately to not hold up cold starts
882 0 : thread::spawn(move || {
883 0 : let mut connstr = connstr.clone();
884 0 : connstr
885 0 : .query_pairs_mut()
886 0 : .append_pair("application_name", "migrations");
887 :
888 0 : let mut client = Client::connect(connstr.as_str(), NoTls)?;
889 0 : handle_migrations(&mut client).context("apply_config handle_migrations")
890 0 : });
891 : Ok(())
892 : }
893 :
894 : // Wrapped this around `pg_ctl reload`, but right now we don't use
895 : // `pg_ctl` for start / stop.
896 0 : #[instrument(skip_all)]
897 : fn pg_reload_conf(&self) -> Result<()> {
898 : let pgctl_bin = Path::new(&self.pgbin).parent().unwrap().join("pg_ctl");
899 : Command::new(pgctl_bin)
900 : .args(["reload", "-D", &self.pgdata])
901 : .output()
902 : .expect("cannot run pg_ctl process");
903 : Ok(())
904 : }
905 :
906 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
907 : /// as it's used to reconfigure a previously started and configured Postgres node.
908 0 : #[instrument(skip_all)]
909 : pub fn reconfigure(&self) -> Result<()> {
910 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
911 :
912 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
913 : info!("tuning pgbouncer");
914 :
915 : let rt = tokio::runtime::Builder::new_current_thread()
916 : .enable_all()
917 : .build()
918 : .expect("failed to create rt");
919 :
920 : // Spawn a thread to do the tuning,
921 : // so that we don't block the main thread that starts Postgres.
922 : let pgbouncer_settings = pgbouncer_settings.clone();
923 0 : let _handle = thread::spawn(move || {
924 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
925 0 : if let Err(err) = res {
926 0 : error!("error while tuning pgbouncer: {err:?}");
927 0 : }
928 0 : });
929 : }
930 :
931 : // Write new config
932 : let pgdata_path = Path::new(&self.pgdata);
933 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
934 : config::write_postgres_conf(&postgresql_conf_path, &spec, None)?;
935 : // temporarily reset max_cluster_size in config
936 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
937 : // creating new extensions, roles, etc...
938 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
939 0 : self.pg_reload_conf()?;
940 :
941 0 : let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
942 :
943 : // Proceed with post-startup configuration. Note, that order of operations is important.
944 : // Disable DDL forwarding because control plane already knows about these roles/databases.
945 0 : if spec.mode == ComputeMode::Primary {
946 0 : client.simple_query("SET neon.forward_ddl = false")?;
947 0 : cleanup_instance(&mut client)?;
948 0 : handle_roles(&spec, &mut client)?;
949 0 : handle_databases(&spec, &mut client)?;
950 0 : handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
951 0 : handle_grants(
952 0 : &spec,
953 0 : &mut client,
954 0 : self.connstr.as_str(),
955 0 : self.has_feature(ComputeFeature::AnonExtension),
956 0 : )?;
957 0 : handle_extensions(&spec, &mut client)?;
958 0 : handle_extension_neon(&mut client)?;
959 : // We can skip handle_migrations here because a new migration can only appear
960 : // if we have a new version of the compute_ctl binary, which can only happen
961 : // if compute got restarted, in which case we'll end up inside of apply_config
962 : // instead of reconfigure.
963 0 : }
964 :
965 : // 'Close' connection
966 0 : drop(client);
967 0 :
968 0 : Ok(())
969 0 : })?;
970 :
971 : self.pg_reload_conf()?;
972 :
973 : let unknown_op = "unknown".to_string();
974 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
975 : info!(
976 : "finished reconfiguration of compute node for operation {}",
977 : op_id
978 : );
979 :
980 : Ok(())
981 : }
982 :
983 0 : #[instrument(skip_all)]
984 : pub fn start_compute(
985 : &self,
986 : extension_server_port: u16,
987 : ) -> Result<(std::process::Child, std::thread::JoinHandle<()>)> {
988 : let compute_state = self.state.lock().unwrap().clone();
989 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
990 : info!(
991 : "starting compute for project {}, operation {}, tenant {}, timeline {}",
992 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
993 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
994 : pspec.tenant_id,
995 : pspec.timeline_id,
996 : );
997 :
998 : // tune pgbouncer
999 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
1000 : info!("tuning pgbouncer");
1001 :
1002 : let rt = tokio::runtime::Builder::new_current_thread()
1003 : .enable_all()
1004 : .build()
1005 : .expect("failed to create rt");
1006 :
1007 : // Spawn a thread to do the tuning,
1008 : // so that we don't block the main thread that starts Postgres.
1009 : let pgbouncer_settings = pgbouncer_settings.clone();
1010 0 : let _handle = thread::spawn(move || {
1011 0 : let res = rt.block_on(tune_pgbouncer(pgbouncer_settings));
1012 0 : if let Err(err) = res {
1013 0 : error!("error while tuning pgbouncer: {err:?}");
1014 0 : }
1015 0 : });
1016 : }
1017 :
1018 : info!(
1019 : "start_compute spec.remote_extensions {:?}",
1020 : pspec.spec.remote_extensions
1021 : );
1022 :
1023 : // This part is sync, because we need to download
1024 : // remote shared_preload_libraries before postgres start (if any)
1025 : if let Some(remote_extensions) = &pspec.spec.remote_extensions {
1026 : // First, create control files for all availale extensions
1027 : extension_server::create_control_files(remote_extensions, &self.pgbin);
1028 :
1029 : let library_load_start_time = Utc::now();
1030 : let remote_ext_metrics = self.prepare_preload_libraries(&pspec.spec)?;
1031 :
1032 : let library_load_time = Utc::now()
1033 : .signed_duration_since(library_load_start_time)
1034 : .to_std()
1035 : .unwrap()
1036 : .as_millis() as u64;
1037 : let mut state = self.state.lock().unwrap();
1038 : state.metrics.load_ext_ms = library_load_time;
1039 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
1040 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
1041 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
1042 : info!(
1043 : "Loading shared_preload_libraries took {:?}ms",
1044 : library_load_time
1045 : );
1046 : info!("{:?}", remote_ext_metrics);
1047 : }
1048 :
1049 : self.prepare_pgdata(&compute_state, extension_server_port)?;
1050 :
1051 : let start_time = Utc::now();
1052 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
1053 :
1054 : let config_time = Utc::now();
1055 : if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
1056 : let pgdata_path = Path::new(&self.pgdata);
1057 : // temporarily reset max_cluster_size in config
1058 : // to avoid the possibility of hitting the limit, while we are applying config:
1059 : // creating new extensions, roles, etc...
1060 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1061 0 : self.pg_reload_conf()?;
1062 :
1063 0 : self.apply_config(&compute_state)?;
1064 :
1065 0 : Ok(())
1066 0 : })?;
1067 : self.pg_reload_conf()?;
1068 : }
1069 :
1070 : let startup_end_time = Utc::now();
1071 : {
1072 : let mut state = self.state.lock().unwrap();
1073 : state.metrics.start_postgres_ms = config_time
1074 : .signed_duration_since(start_time)
1075 : .to_std()
1076 : .unwrap()
1077 : .as_millis() as u64;
1078 : state.metrics.config_ms = startup_end_time
1079 : .signed_duration_since(config_time)
1080 : .to_std()
1081 : .unwrap()
1082 : .as_millis() as u64;
1083 : state.metrics.total_startup_ms = startup_end_time
1084 : .signed_duration_since(compute_state.start_time)
1085 : .to_std()
1086 : .unwrap()
1087 : .as_millis() as u64;
1088 : }
1089 : self.set_status(ComputeStatus::Running);
1090 :
1091 : info!(
1092 : "finished configuration of compute for project {}",
1093 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
1094 : );
1095 :
1096 : // Log metrics so that we can search for slow operations in logs
1097 : let metrics = {
1098 : let state = self.state.lock().unwrap();
1099 : state.metrics.clone()
1100 : };
1101 : info!(?metrics, "compute start finished");
1102 :
1103 : Ok(pg_process)
1104 : }
1105 :
1106 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1107 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1108 0 : let mut state = self.state.lock().unwrap();
1109 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1110 0 : if last_active > state.last_active {
1111 0 : state.last_active = last_active;
1112 0 : debug!("set the last compute activity time to: {:?}", last_active);
1113 0 : }
1114 0 : }
1115 :
1116 : // Look for core dumps and collect backtraces.
1117 : //
1118 : // EKS worker nodes have following core dump settings:
1119 : // /proc/sys/kernel/core_pattern -> core
1120 : // /proc/sys/kernel/core_uses_pid -> 1
1121 : // ulimit -c -> unlimited
1122 : // which results in core dumps being written to postgres data directory as core.<pid>.
1123 : //
1124 : // Use that as a default location and pattern, except macos where core dumps are written
1125 : // to /cores/ directory by default.
1126 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1127 0 : let core_dump_dir = match std::env::consts::OS {
1128 0 : "macos" => Path::new("/cores/"),
1129 0 : _ => Path::new(&self.pgdata),
1130 : };
1131 :
1132 : // Collect core dump paths if any
1133 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1134 0 : let files = fs::read_dir(core_dump_dir)?;
1135 0 : let cores = files.filter_map(|entry| {
1136 0 : let entry = entry.ok()?;
1137 0 : let _ = entry.file_name().to_str()?.strip_prefix("core.")?;
1138 0 : Some(entry.path())
1139 0 : });
1140 :
1141 : // Print backtrace for each core dump
1142 0 : for core_path in cores {
1143 0 : warn!(
1144 0 : "core dump found: {}, collecting backtrace",
1145 0 : core_path.display()
1146 : );
1147 :
1148 : // Try first with gdb
1149 0 : let backtrace = Command::new("gdb")
1150 0 : .args(["--batch", "-q", "-ex", "bt", &self.pgbin])
1151 0 : .arg(&core_path)
1152 0 : .output();
1153 :
1154 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1155 0 : let backtrace = match backtrace {
1156 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1157 0 : warn!("cannot find gdb, trying lldb");
1158 0 : Command::new("lldb")
1159 0 : .arg("-c")
1160 0 : .arg(&core_path)
1161 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1162 0 : .output()
1163 : }
1164 0 : _ => backtrace,
1165 0 : }?;
1166 :
1167 0 : warn!(
1168 0 : "core dump backtrace: {}",
1169 0 : String::from_utf8_lossy(&backtrace.stdout)
1170 : );
1171 0 : warn!(
1172 0 : "debugger stderr: {}",
1173 0 : String::from_utf8_lossy(&backtrace.stderr)
1174 : );
1175 : }
1176 :
1177 0 : Ok(())
1178 0 : }
1179 :
1180 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1181 0 : pub async fn collect_insights(&self) -> String {
1182 0 : let mut result_rows: Vec<String> = Vec::new();
1183 0 : let connect_result = tokio_postgres::connect(self.connstr.as_str(), NoTls).await;
1184 0 : let (client, connection) = connect_result.unwrap();
1185 0 : tokio::spawn(async move {
1186 0 : if let Err(e) = connection.await {
1187 0 : eprintln!("connection error: {}", e);
1188 0 : }
1189 0 : });
1190 0 : let result = client
1191 0 : .simple_query(
1192 0 : "SELECT
1193 0 : row_to_json(pg_stat_statements)
1194 0 : FROM
1195 0 : pg_stat_statements
1196 0 : WHERE
1197 0 : userid != 'cloud_admin'::regrole::oid
1198 0 : ORDER BY
1199 0 : (mean_exec_time + mean_plan_time) DESC
1200 0 : LIMIT 100",
1201 0 : )
1202 0 : .await;
1203 :
1204 0 : if let Ok(raw_rows) = result {
1205 0 : for message in raw_rows.iter() {
1206 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1207 0 : if let Some(json) = row.get(0) {
1208 0 : result_rows.push(json.to_string());
1209 0 : }
1210 0 : }
1211 : }
1212 :
1213 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1214 : } else {
1215 0 : "{{\"pg_stat_statements\": []}}".to_string()
1216 : }
1217 0 : }
1218 :
1219 : // download an archive, unzip and place files in correct locations
1220 0 : pub async fn download_extension(
1221 0 : &self,
1222 0 : real_ext_name: String,
1223 0 : ext_path: RemotePath,
1224 0 : ) -> Result<u64, DownloadError> {
1225 0 : let ext_remote_storage =
1226 0 : self.ext_remote_storage
1227 0 : .as_ref()
1228 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1229 0 : "Remote extensions storage is not configured",
1230 0 : )))?;
1231 :
1232 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1233 0 :
1234 0 : let mut first_try = false;
1235 0 : if !self
1236 0 : .ext_download_progress
1237 0 : .read()
1238 0 : .expect("lock err")
1239 0 : .contains_key(ext_archive_name)
1240 0 : {
1241 0 : self.ext_download_progress
1242 0 : .write()
1243 0 : .expect("lock err")
1244 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1245 0 : first_try = true;
1246 0 : }
1247 0 : let (download_start, download_completed) =
1248 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1249 0 : let start_time_delta = Utc::now()
1250 0 : .signed_duration_since(download_start)
1251 0 : .to_std()
1252 0 : .unwrap()
1253 0 : .as_millis() as u64;
1254 :
1255 : // how long to wait for extension download if it was started by another process
1256 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1257 :
1258 0 : if download_completed {
1259 0 : info!("extension already downloaded, skipping re-download");
1260 0 : return Ok(0);
1261 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1262 0 : info!("download {ext_archive_name} already started by another process, hanging untill completion or timeout");
1263 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1264 : loop {
1265 0 : info!("waiting for download");
1266 0 : interval.tick().await;
1267 0 : let (_, download_completed_now) =
1268 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1269 0 : if download_completed_now {
1270 0 : info!("download finished by whoever else downloaded it");
1271 0 : return Ok(0);
1272 0 : }
1273 : }
1274 : // NOTE: the above loop will get terminated
1275 : // based on the timeout of the download function
1276 0 : }
1277 0 :
1278 0 : // if extension hasn't been downloaded before or the previous
1279 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1280 0 : // then we try to download it here
1281 0 : info!("downloading new extension {ext_archive_name}");
1282 :
1283 0 : let download_size = extension_server::download_extension(
1284 0 : &real_ext_name,
1285 0 : &ext_path,
1286 0 : ext_remote_storage,
1287 0 : &self.pgbin,
1288 0 : )
1289 0 : .await
1290 0 : .map_err(DownloadError::Other);
1291 0 :
1292 0 : if download_size.is_ok() {
1293 0 : self.ext_download_progress
1294 0 : .write()
1295 0 : .expect("bad lock")
1296 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1297 0 : }
1298 :
1299 0 : download_size
1300 0 : }
1301 :
1302 : #[tokio::main]
1303 0 : pub async fn prepare_preload_libraries(
1304 0 : &self,
1305 0 : spec: &ComputeSpec,
1306 0 : ) -> Result<RemoteExtensionMetrics> {
1307 0 : if self.ext_remote_storage.is_none() {
1308 0 : return Ok(RemoteExtensionMetrics {
1309 0 : num_ext_downloaded: 0,
1310 0 : largest_ext_size: 0,
1311 0 : total_ext_download_size: 0,
1312 0 : });
1313 0 : }
1314 0 : let remote_extensions = spec
1315 0 : .remote_extensions
1316 0 : .as_ref()
1317 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
1318 0 :
1319 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
1320 0 : let mut libs_vec = Vec::new();
1321 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
1322 0 : libs_vec = libs
1323 0 : .split(&[',', '\'', ' '])
1324 0 : .filter(|s| *s != "neon" && !s.is_empty())
1325 0 : .map(str::to_string)
1326 0 : .collect();
1327 0 : }
1328 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
1329 0 :
1330 0 : // that is used in neon_local and python tests
1331 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
1332 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
1333 0 : let mut shared_preload_libraries_line = "";
1334 0 : for line in conf_lines {
1335 0 : if line.starts_with("shared_preload_libraries") {
1336 0 : shared_preload_libraries_line = line;
1337 0 : }
1338 0 : }
1339 0 : let mut preload_libs_vec = Vec::new();
1340 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
1341 0 : preload_libs_vec = libs
1342 0 : .split(&[',', '\'', ' '])
1343 0 : .filter(|s| *s != "neon" && !s.is_empty())
1344 0 : .map(str::to_string)
1345 0 : .collect();
1346 0 : }
1347 0 : libs_vec.extend(preload_libs_vec);
1348 0 : }
1349 0 :
1350 0 : // Don't try to download libraries that are not in the index.
1351 0 : // Assume that they are already present locally.
1352 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
1353 0 :
1354 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
1355 0 :
1356 0 : let mut download_tasks = Vec::new();
1357 0 : for library in &libs_vec {
1358 0 : let (ext_name, ext_path) =
1359 0 : remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
1360 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
1361 0 : }
1362 0 : let results = join_all(download_tasks).await;
1363 0 :
1364 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
1365 0 : num_ext_downloaded: 0,
1366 0 : largest_ext_size: 0,
1367 0 : total_ext_download_size: 0,
1368 0 : };
1369 0 : for result in results {
1370 0 : let download_size = match result {
1371 0 : Ok(res) => {
1372 0 : remote_ext_metrics.num_ext_downloaded += 1;
1373 0 : res
1374 0 : }
1375 0 : Err(err) => {
1376 0 : // if we failed to download an extension, we don't want to fail the whole
1377 0 : // process, but we do want to log the error
1378 0 : error!("Failed to download extension: {}", err);
1379 0 : 0
1380 0 : }
1381 0 : };
1382 0 :
1383 0 : remote_ext_metrics.largest_ext_size =
1384 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
1385 0 : remote_ext_metrics.total_ext_download_size += download_size;
1386 0 : }
1387 0 : Ok(remote_ext_metrics)
1388 0 : }
1389 : }
1390 :
1391 0 : pub fn forward_termination_signal() {
1392 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
1393 0 : if ss_pid != 0 {
1394 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
1395 0 : kill(ss_pid, Signal::SIGTERM).ok();
1396 0 : }
1397 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
1398 0 : if pg_pid != 0 {
1399 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
1400 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
1401 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
1402 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
1403 0 : kill(pg_pid, Signal::SIGINT).ok();
1404 0 : }
1405 0 : }
|
