Line data Source code
1 : use anyhow::{Context, Result};
2 : use chrono::{DateTime, Utc};
3 : use compute_api::privilege::Privilege;
4 : use compute_api::responses::{
5 : ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus, LfcOffloadState,
6 : LfcPrewarmState, PromoteState, TlsConfig,
7 : };
8 : use compute_api::spec::{
9 : ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
10 : };
11 : use futures::StreamExt;
12 : use futures::future::join_all;
13 : use futures::stream::FuturesUnordered;
14 : use itertools::Itertools;
15 : use nix::sys::signal::{Signal, kill};
16 : use nix::unistd::Pid;
17 : use once_cell::sync::Lazy;
18 : use pageserver_page_api::{self as page_api, BaseBackupCompression};
19 : use postgres;
20 : use postgres::NoTls;
21 : use postgres::error::SqlState;
22 : use remote_storage::{DownloadError, RemotePath};
23 : use std::collections::{HashMap, HashSet};
24 : use std::os::unix::fs::{PermissionsExt, symlink};
25 : use std::path::Path;
26 : use std::process::{Command, Stdio};
27 : use std::str::FromStr;
28 : use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
29 : use std::sync::{Arc, Condvar, Mutex, RwLock};
30 : use std::time::{Duration, Instant};
31 : use std::{env, fs};
32 : use tokio::{spawn, sync::watch, task::JoinHandle, time};
33 : use tracing::{Instrument, debug, error, info, instrument, warn};
34 : use url::Url;
35 : use utils::id::{TenantId, TimelineId};
36 : use utils::lsn::Lsn;
37 : use utils::measured_stream::MeasuredReader;
38 : use utils::pid_file;
39 : use utils::shard::{ShardCount, ShardIndex, ShardNumber};
40 :
41 : use crate::configurator::launch_configurator;
42 : use crate::disk_quota::set_disk_quota;
43 : use crate::installed_extensions::get_installed_extensions;
44 : use crate::logger::startup_context_from_env;
45 : use crate::lsn_lease::launch_lsn_lease_bg_task_for_static;
46 : use crate::metrics::COMPUTE_CTL_UP;
47 : use crate::monitor::launch_monitor;
48 : use crate::pg_helpers::*;
49 : use crate::pgbouncer::*;
50 : use crate::rsyslog::{
51 : PostgresLogsRsyslogConfig, configure_audit_rsyslog, configure_postgres_logs_export,
52 : launch_pgaudit_gc,
53 : };
54 : use crate::spec::*;
55 : use crate::swap::resize_swap;
56 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
57 : use crate::tls::watch_cert_for_changes;
58 : use crate::{config, extension_server, local_proxy};
59 :
60 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
61 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
62 : // This is an arbitrary build tag. Fine as a default / for testing purposes
63 : // in-case of not-set environment var
64 : const BUILD_TAG_DEFAULT: &str = "latest";
65 : /// Build tag/version of the compute node binaries/image. It's tricky and ugly
66 : /// to pass it everywhere as a part of `ComputeNodeParams`, so we use a
67 : /// global static variable.
68 0 : pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
69 0 : option_env!("BUILD_TAG")
70 0 : .unwrap_or(BUILD_TAG_DEFAULT)
71 0 : .to_string()
72 0 : });
73 : const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
74 :
75 : /// Static configuration params that don't change after startup. These mostly
76 : /// come from the CLI args, or are derived from them.
77 : pub struct ComputeNodeParams {
78 : /// The ID of the compute
79 : pub compute_id: String,
80 : // Url type maintains proper escaping
81 : pub connstr: url::Url,
82 :
83 : pub resize_swap_on_bind: bool,
84 : pub set_disk_quota_for_fs: Option<String>,
85 :
86 : // VM monitor parameters
87 : #[cfg(target_os = "linux")]
88 : pub filecache_connstr: String,
89 : #[cfg(target_os = "linux")]
90 : pub cgroup: String,
91 : #[cfg(target_os = "linux")]
92 : pub vm_monitor_addr: String,
93 :
94 : pub pgdata: String,
95 : pub pgbin: String,
96 : pub pgversion: String,
97 :
98 : /// The port that the compute's external HTTP server listens on
99 : pub external_http_port: u16,
100 : /// The port that the compute's internal HTTP server listens on
101 : pub internal_http_port: u16,
102 :
103 : /// the address of extension storage proxy gateway
104 : pub remote_ext_base_url: Option<Url>,
105 :
106 : /// Interval for installed extensions collection
107 : pub installed_extensions_collection_interval: Arc<AtomicU64>,
108 : }
109 :
110 : type TaskHandle = Mutex<Option<JoinHandle<()>>>;
111 :
112 : /// Compute node info shared across several `compute_ctl` threads.
113 : pub struct ComputeNode {
114 : pub params: ComputeNodeParams,
115 :
116 : // We connect to Postgres from many different places, so build configs once
117 : // and reuse them where needed. These are derived from 'params.connstr'
118 : pub conn_conf: postgres::config::Config,
119 : pub tokio_conn_conf: tokio_postgres::config::Config,
120 :
121 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
122 : /// To allow HTTP API server to serving status requests, while configuration
123 : /// is in progress, lock should be held only for short periods of time to do
124 : /// read/write, not the whole configuration process.
125 : pub state: Mutex<ComputeState>,
126 : /// `Condvar` to allow notifying waiters about state changes.
127 : pub state_changed: Condvar,
128 :
129 : // key: ext_archive_name, value: started download time, download_completed?
130 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
131 : pub compute_ctl_config: ComputeCtlConfig,
132 :
133 : /// Handle to the extension stats collection task
134 : extension_stats_task: TaskHandle,
135 : lfc_offload_task: TaskHandle,
136 : }
137 :
138 : // store some metrics about download size that might impact startup time
139 : #[derive(Clone, Debug)]
140 : pub struct RemoteExtensionMetrics {
141 : num_ext_downloaded: u64,
142 : largest_ext_size: u64,
143 : total_ext_download_size: u64,
144 : }
145 :
146 : #[derive(Clone, Debug)]
147 : pub struct ComputeState {
148 : pub start_time: DateTime<Utc>,
149 : pub status: ComputeStatus,
150 : /// Timestamp of the last Postgres activity. It could be `None` if
151 : /// compute wasn't used since start.
152 : pub last_active: Option<DateTime<Utc>>,
153 : pub error: Option<String>,
154 :
155 : /// Compute spec. This can be received from the CLI or - more likely -
156 : /// passed by the control plane with a /configure HTTP request.
157 : pub pspec: Option<ParsedSpec>,
158 :
159 : /// If the spec is passed by a /configure request, 'startup_span' is the
160 : /// /configure request's tracing span. The main thread enters it when it
161 : /// processes the compute startup, so that the compute startup is considered
162 : /// to be part of the /configure request for tracing purposes.
163 : ///
164 : /// If the request handling thread/task called startup_compute() directly,
165 : /// it would automatically be a child of the request handling span, and we
166 : /// wouldn't need this. But because we use the main thread to perform the
167 : /// startup, and the /configure task just waits for it to finish, we need to
168 : /// set up the span relationship ourselves.
169 : pub startup_span: Option<tracing::span::Span>,
170 :
171 : pub lfc_prewarm_state: LfcPrewarmState,
172 : pub lfc_offload_state: LfcOffloadState,
173 :
174 : /// WAL flush LSN that is set after terminating Postgres and syncing safekeepers if
175 : /// mode == ComputeMode::Primary. None otherwise
176 : pub terminate_flush_lsn: Option<Lsn>,
177 : pub promote_state: Option<watch::Receiver<PromoteState>>,
178 :
179 : pub metrics: ComputeMetrics,
180 : }
181 :
182 : impl ComputeState {
183 0 : pub fn new() -> Self {
184 0 : Self {
185 0 : start_time: Utc::now(),
186 0 : status: ComputeStatus::Empty,
187 0 : last_active: None,
188 0 : error: None,
189 0 : pspec: None,
190 0 : startup_span: None,
191 0 : metrics: ComputeMetrics::default(),
192 0 : lfc_prewarm_state: LfcPrewarmState::default(),
193 0 : lfc_offload_state: LfcOffloadState::default(),
194 0 : terminate_flush_lsn: None,
195 0 : promote_state: None,
196 0 : }
197 0 : }
198 :
199 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
200 0 : let prev = self.status;
201 0 : info!("Changing compute status from {} to {}", prev, status);
202 0 : self.status = status;
203 0 : state_changed.notify_all();
204 :
205 0 : COMPUTE_CTL_UP.reset();
206 0 : COMPUTE_CTL_UP
207 0 : .with_label_values(&[&BUILD_TAG, status.to_string().as_str()])
208 0 : .set(1);
209 0 : }
210 :
211 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
212 0 : self.error = Some(format!("{err:?}"));
213 0 : self.set_status(ComputeStatus::Failed, state_changed);
214 0 : }
215 : }
216 :
217 : impl Default for ComputeState {
218 0 : fn default() -> Self {
219 0 : Self::new()
220 0 : }
221 : }
222 :
223 : #[derive(Clone, Debug)]
224 : pub struct ParsedSpec {
225 : pub spec: ComputeSpec,
226 : pub tenant_id: TenantId,
227 : pub timeline_id: TimelineId,
228 : pub pageserver_connstr: String,
229 : pub safekeeper_connstrings: Vec<String>,
230 : pub storage_auth_token: Option<String>,
231 : /// k8s dns name and port
232 : pub endpoint_storage_addr: Option<String>,
233 : pub endpoint_storage_token: Option<String>,
234 : }
235 :
236 : impl ParsedSpec {
237 1 : pub fn validate(&self) -> Result<(), String> {
238 : // Only Primary nodes are using safekeeper_connstrings, and at the moment
239 : // this method only validates that part of the specs.
240 1 : if self.spec.mode != ComputeMode::Primary {
241 0 : return Ok(());
242 1 : }
243 :
244 : // While it seems like a good idea to check for an odd number of entries in
245 : // the safekeepers connection string, changes to the list of safekeepers might
246 : // incur appending a new server to a list of 3, in which case a list of 4
247 : // entries is okay in production.
248 : //
249 : // Still we want unique entries, and at least one entry in the vector
250 1 : if self.safekeeper_connstrings.is_empty() {
251 0 : return Err(String::from("safekeeper_connstrings is empty"));
252 1 : }
253 :
254 : // check for uniqueness of the connection strings in the set
255 1 : let mut connstrings = self.safekeeper_connstrings.clone();
256 :
257 1 : connstrings.sort();
258 1 : let mut previous = &connstrings[0];
259 :
260 2 : for current in connstrings.iter().skip(1) {
261 : // duplicate entry?
262 2 : if current == previous {
263 1 : return Err(format!(
264 1 : "duplicate entry in safekeeper_connstrings: {current}!",
265 1 : ));
266 1 : }
267 :
268 1 : previous = current;
269 : }
270 :
271 0 : Ok(())
272 1 : }
273 : }
274 :
275 : impl TryFrom<ComputeSpec> for ParsedSpec {
276 : type Error = String;
277 1 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
278 : // Extract the options from the spec file that are needed to connect to
279 : // the storage system.
280 : //
281 : // For backwards-compatibility, the top-level fields in the spec file
282 : // may be empty. In that case, we need to dig them from the GUCs in the
283 : // cluster.settings field.
284 1 : let pageserver_connstr = spec
285 1 : .pageserver_connstring
286 1 : .clone()
287 1 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
288 1 : .ok_or("pageserver connstr should be provided")?;
289 1 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
290 1 : if matches!(spec.mode, ComputeMode::Primary) {
291 1 : spec.cluster
292 1 : .settings
293 1 : .find("neon.safekeepers")
294 1 : .ok_or("safekeeper connstrings should be provided")?
295 1 : .split(',')
296 4 : .map(|str| str.to_string())
297 1 : .collect()
298 : } else {
299 0 : vec![]
300 : }
301 : } else {
302 0 : spec.safekeeper_connstrings.clone()
303 : };
304 :
305 1 : let storage_auth_token = spec.storage_auth_token.clone();
306 1 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
307 0 : tenant_id
308 : } else {
309 1 : spec.cluster
310 1 : .settings
311 1 : .find("neon.tenant_id")
312 1 : .ok_or("tenant id should be provided")
313 1 : .map(|s| TenantId::from_str(&s))?
314 1 : .or(Err("invalid tenant id"))?
315 : };
316 1 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
317 0 : timeline_id
318 : } else {
319 1 : spec.cluster
320 1 : .settings
321 1 : .find("neon.timeline_id")
322 1 : .ok_or("timeline id should be provided")
323 1 : .map(|s| TimelineId::from_str(&s))?
324 1 : .or(Err("invalid timeline id"))?
325 : };
326 :
327 1 : let endpoint_storage_addr: Option<String> = spec
328 1 : .endpoint_storage_addr
329 1 : .clone()
330 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"));
331 1 : let endpoint_storage_token = spec
332 1 : .endpoint_storage_token
333 1 : .clone()
334 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_token"));
335 :
336 1 : let res = ParsedSpec {
337 1 : spec,
338 1 : pageserver_connstr,
339 1 : safekeeper_connstrings,
340 1 : storage_auth_token,
341 1 : tenant_id,
342 1 : timeline_id,
343 1 : endpoint_storage_addr,
344 1 : endpoint_storage_token,
345 1 : };
346 :
347 : // Now check validity of the parsed specification
348 1 : res.validate()?;
349 0 : Ok(res)
350 1 : }
351 : }
352 :
353 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
354 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
355 : ///
356 : /// This function should be used to start postgres, as it will start it in the
357 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
358 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
359 : /// creates it during the sysinit phase of its inittab.
360 0 : fn maybe_cgexec(cmd: &str) -> Command {
361 : // The cplane sets this env var for autoscaling computes.
362 : // use `var_os` so we don't have to worry about the variable being valid
363 : // unicode. Should never be an concern . . . but just in case
364 0 : if env::var_os("AUTOSCALING").is_some() {
365 0 : let mut command = Command::new("cgexec");
366 0 : command.args(["-g", "memory:neon-postgres"]);
367 0 : command.arg(cmd);
368 0 : command
369 : } else {
370 0 : Command::new(cmd)
371 : }
372 0 : }
373 :
374 : struct PostgresHandle {
375 : postgres: std::process::Child,
376 : log_collector: JoinHandle<Result<()>>,
377 : }
378 :
379 : impl PostgresHandle {
380 : /// Return PID of the postgres (postmaster) process
381 0 : fn pid(&self) -> Pid {
382 0 : Pid::from_raw(self.postgres.id() as i32)
383 0 : }
384 : }
385 :
386 : struct StartVmMonitorResult {
387 : #[cfg(target_os = "linux")]
388 : token: tokio_util::sync::CancellationToken,
389 : #[cfg(target_os = "linux")]
390 : vm_monitor: Option<JoinHandle<Result<()>>>,
391 : }
392 :
393 : impl ComputeNode {
394 0 : pub fn new(params: ComputeNodeParams, config: ComputeConfig) -> Result<Self> {
395 0 : let connstr = params.connstr.as_str();
396 0 : let mut conn_conf = postgres::config::Config::from_str(connstr)
397 0 : .context("cannot build postgres config from connstr")?;
398 0 : let mut tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr)
399 0 : .context("cannot build tokio postgres config from connstr")?;
400 :
401 : // Users can set some configuration parameters per database with
402 : // ALTER DATABASE ... SET ...
403 : //
404 : // There are at least these parameters:
405 : //
406 : // - role=some_other_role
407 : // - default_transaction_read_only=on
408 : // - statement_timeout=1, i.e., 1ms, which will cause most of the queries to fail
409 : // - search_path=non_public_schema, this should be actually safe because
410 : // we don't call any functions in user databases, but better to always reset
411 : // it to public.
412 : //
413 : // that can affect `compute_ctl` and prevent it from properly configuring the database schema.
414 : // Unset them via connection string options before connecting to the database.
415 : // N.B. keep it in sync with `ZENITH_OPTIONS` in `get_maintenance_client()`.
416 : const EXTRA_OPTIONS: &str = "-c role=cloud_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0 -c pgaudit.log=none";
417 0 : let options = match conn_conf.get_options() {
418 : // Allow the control plane to override any options set by the
419 : // compute
420 0 : Some(options) => format!("{EXTRA_OPTIONS} {options}"),
421 0 : None => EXTRA_OPTIONS.to_string(),
422 : };
423 0 : conn_conf.options(&options);
424 0 : tokio_conn_conf.options(&options);
425 :
426 0 : let mut new_state = ComputeState::new();
427 0 : if let Some(spec) = config.spec {
428 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
429 0 : new_state.pspec = Some(pspec);
430 0 : }
431 :
432 0 : Ok(ComputeNode {
433 0 : params,
434 0 : conn_conf,
435 0 : tokio_conn_conf,
436 0 : state: Mutex::new(new_state),
437 0 : state_changed: Condvar::new(),
438 0 : ext_download_progress: RwLock::new(HashMap::new()),
439 0 : compute_ctl_config: config.compute_ctl_config,
440 0 : extension_stats_task: Mutex::new(None),
441 0 : lfc_offload_task: Mutex::new(None),
442 0 : })
443 0 : }
444 :
445 : /// Top-level control flow of compute_ctl. Returns a process exit code we should
446 : /// exit with.
447 0 : pub fn run(self) -> Result<Option<i32>> {
448 0 : let this = Arc::new(self);
449 :
450 0 : let cli_spec = this.state.lock().unwrap().pspec.clone();
451 :
452 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
453 : // available for binding. Prewarming helps Postgres start quicker later,
454 : // because QEMU will already have its memory allocated from the host, and
455 : // the necessary binaries will already be cached.
456 0 : if cli_spec.is_none() {
457 0 : this.prewarm_postgres_vm_memory()?;
458 0 : }
459 :
460 : // Set the up metric with Empty status before starting the HTTP server.
461 : // That way on the first metric scrape, an external observer will see us
462 : // as 'up' and 'empty' (unless the compute was started with a spec or
463 : // already configured by control plane).
464 0 : COMPUTE_CTL_UP
465 0 : .with_label_values(&[&BUILD_TAG, ComputeStatus::Empty.to_string().as_str()])
466 0 : .set(1);
467 :
468 : // Launch the external HTTP server first, so that we can serve control plane
469 : // requests while configuration is still in progress.
470 0 : crate::http::server::Server::External {
471 0 : port: this.params.external_http_port,
472 0 : config: this.compute_ctl_config.clone(),
473 0 : compute_id: this.params.compute_id.clone(),
474 0 : }
475 0 : .launch(&this);
476 :
477 : // The internal HTTP server could be launched later, but there isn't much
478 : // sense in waiting.
479 0 : crate::http::server::Server::Internal {
480 0 : port: this.params.internal_http_port,
481 0 : }
482 0 : .launch(&this);
483 :
484 : // If we got a spec from the CLI already, use that. Otherwise wait for the
485 : // control plane to pass it to us with a /configure HTTP request
486 0 : let pspec = if let Some(cli_spec) = cli_spec {
487 0 : cli_spec
488 : } else {
489 0 : this.wait_spec()?
490 : };
491 :
492 0 : launch_lsn_lease_bg_task_for_static(&this);
493 :
494 : // We have a spec, start the compute
495 0 : let mut delay_exit = false;
496 0 : let mut vm_monitor = None;
497 0 : let mut pg_process: Option<PostgresHandle> = None;
498 :
499 0 : match this.start_compute(&mut pg_process) {
500 0 : Ok(()) => {
501 0 : // Success! Launch remaining services (just vm-monitor currently)
502 0 : vm_monitor =
503 0 : Some(this.start_vm_monitor(pspec.spec.disable_lfc_resizing.unwrap_or(false)));
504 0 : }
505 0 : Err(err) => {
506 : // Something went wrong with the startup. Log it and expose the error to
507 : // HTTP status requests.
508 0 : error!("could not start the compute node: {:#}", err);
509 0 : this.set_failed_status(err);
510 0 : delay_exit = true;
511 :
512 : // If the error happened after starting PostgreSQL, kill it
513 0 : if let Some(ref pg_process) = pg_process {
514 0 : kill(pg_process.pid(), Signal::SIGQUIT).ok();
515 0 : }
516 : }
517 : }
518 :
519 : // If startup was successful, or it failed in the late stages,
520 : // PostgreSQL is now running. Wait until it exits.
521 0 : let exit_code = if let Some(pg_handle) = pg_process {
522 0 : let exit_status = this.wait_postgres(pg_handle);
523 0 : info!("Postgres exited with code {}, shutting down", exit_status);
524 0 : exit_status.code()
525 : } else {
526 0 : None
527 : };
528 :
529 0 : this.terminate_extension_stats_task();
530 0 : this.terminate_lfc_offload_task();
531 :
532 : // Terminate the vm_monitor so it releases the file watcher on
533 : // /sys/fs/cgroup/neon-postgres.
534 : // Note: the vm-monitor only runs on linux because it requires cgroups.
535 0 : if let Some(vm_monitor) = vm_monitor {
536 : cfg_if::cfg_if! {
537 : if #[cfg(target_os = "linux")] {
538 : // Kills all threads spawned by the monitor
539 0 : vm_monitor.token.cancel();
540 0 : if let Some(handle) = vm_monitor.vm_monitor {
541 0 : // Kills the actual task running the monitor
542 0 : handle.abort();
543 0 : }
544 : } else {
545 : _ = vm_monitor; // appease unused lint on macOS
546 : }
547 : }
548 0 : }
549 :
550 : // Reap the postgres process
551 0 : delay_exit |= this.cleanup_after_postgres_exit()?;
552 :
553 : // /terminate returns LSN. If we don't sleep at all, connection will break and we
554 : // won't get result. If we sleep too much, tests will take significantly longer
555 : // and Github Action run will error out
556 0 : let sleep_duration = if delay_exit {
557 0 : Duration::from_secs(30)
558 : } else {
559 0 : Duration::from_millis(300)
560 : };
561 :
562 : // If launch failed, keep serving HTTP requests for a while, so the cloud
563 : // control plane can get the actual error.
564 0 : if delay_exit {
565 0 : info!("giving control plane 30s to collect the error before shutdown");
566 0 : }
567 0 : std::thread::sleep(sleep_duration);
568 0 : Ok(exit_code)
569 0 : }
570 :
571 0 : pub fn wait_spec(&self) -> Result<ParsedSpec> {
572 0 : info!("no compute spec provided, waiting");
573 0 : let mut state = self.state.lock().unwrap();
574 0 : while state.status != ComputeStatus::ConfigurationPending {
575 0 : state = self.state_changed.wait(state).unwrap();
576 0 : }
577 :
578 0 : info!("got spec, continue configuration");
579 0 : let spec = state.pspec.as_ref().unwrap().clone();
580 :
581 : // Record for how long we slept waiting for the spec.
582 0 : let now = Utc::now();
583 0 : state.metrics.wait_for_spec_ms = now
584 0 : .signed_duration_since(state.start_time)
585 0 : .to_std()
586 0 : .unwrap()
587 0 : .as_millis() as u64;
588 :
589 : // Reset start time, so that the total startup time that is calculated later will
590 : // not include the time that we waited for the spec.
591 0 : state.start_time = now;
592 :
593 0 : Ok(spec)
594 0 : }
595 :
596 : /// Start compute.
597 : ///
598 : /// Prerequisites:
599 : /// - the compute spec has been placed in self.state.pspec
600 : ///
601 : /// On success:
602 : /// - status is set to ComputeStatus::Running
603 : /// - self.running_postgres is set
604 : ///
605 : /// On error:
606 : /// - status is left in ComputeStatus::Init. The caller is responsible for setting it to Failed
607 : /// - if Postgres was started before the fatal error happened, self.running_postgres is
608 : /// set. The caller is responsible for killing it.
609 : ///
610 : /// Note that this is in the critical path of a compute cold start. Keep this fast.
611 : /// Try to do things concurrently, to hide the latencies.
612 0 : fn start_compute(self: &Arc<Self>, pg_handle: &mut Option<PostgresHandle>) -> Result<()> {
613 : let compute_state: ComputeState;
614 :
615 : let start_compute_span;
616 : let _this_entered;
617 : {
618 0 : let mut state_guard = self.state.lock().unwrap();
619 :
620 : // Create a tracing span for the startup operation.
621 : //
622 : // We could otherwise just annotate the function with #[instrument], but if
623 : // we're being configured from a /configure HTTP request, we want the
624 : // startup to be considered part of the /configure request.
625 : //
626 : // Similarly, if a trace ID was passed in env variables, attach it to the span.
627 0 : start_compute_span = {
628 : // Temporarily enter the parent span, so that the new span becomes its child.
629 0 : if let Some(p) = state_guard.startup_span.take() {
630 0 : let _parent_entered = p.entered();
631 0 : tracing::info_span!("start_compute")
632 0 : } else if let Some(otel_context) = startup_context_from_env() {
633 : use tracing_opentelemetry::OpenTelemetrySpanExt;
634 0 : let span = tracing::info_span!("start_compute");
635 0 : span.set_parent(otel_context);
636 0 : span
637 : } else {
638 0 : tracing::info_span!("start_compute")
639 : }
640 : };
641 0 : _this_entered = start_compute_span.enter();
642 :
643 0 : state_guard.set_status(ComputeStatus::Init, &self.state_changed);
644 0 : compute_state = state_guard.clone()
645 : }
646 :
647 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
648 0 : info!(
649 0 : "starting compute for project {}, operation {}, tenant {}, timeline {}, project {}, branch {}, endpoint {}, features {:?}, spec.remote_extensions {:?}",
650 0 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
651 0 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
652 : pspec.tenant_id,
653 : pspec.timeline_id,
654 0 : pspec.spec.project_id.as_deref().unwrap_or("None"),
655 0 : pspec.spec.branch_id.as_deref().unwrap_or("None"),
656 0 : pspec.spec.endpoint_id.as_deref().unwrap_or("None"),
657 : pspec.spec.features,
658 : pspec.spec.remote_extensions,
659 : );
660 :
661 : ////// PRE-STARTUP PHASE: things that need to be finished before we start the Postgres process
662 :
663 : // Collect all the tasks that must finish here
664 0 : let mut pre_tasks = tokio::task::JoinSet::new();
665 :
666 : // Make sure TLS certificates are properly loaded and in the right place.
667 0 : if self.compute_ctl_config.tls.is_some() {
668 0 : let this = self.clone();
669 0 : pre_tasks.spawn(async move {
670 0 : this.watch_cert_for_changes().await;
671 :
672 0 : Ok::<(), anyhow::Error>(())
673 0 : });
674 0 : }
675 :
676 0 : let tls_config = self.tls_config(&pspec.spec);
677 :
678 : // If there are any remote extensions in shared_preload_libraries, start downloading them
679 0 : if pspec.spec.remote_extensions.is_some() {
680 0 : let (this, spec) = (self.clone(), pspec.spec.clone());
681 0 : pre_tasks.spawn(async move {
682 0 : this.download_preload_extensions(&spec)
683 0 : .in_current_span()
684 0 : .await
685 0 : });
686 0 : }
687 :
688 : // Prepare pgdata directory. This downloads the basebackup, among other things.
689 : {
690 0 : let (this, cs) = (self.clone(), compute_state.clone());
691 0 : pre_tasks.spawn_blocking_child(move || this.prepare_pgdata(&cs));
692 : }
693 :
694 : // Resize swap to the desired size if the compute spec says so
695 0 : if let (Some(size_bytes), true) =
696 0 : (pspec.spec.swap_size_bytes, self.params.resize_swap_on_bind)
697 : {
698 0 : pre_tasks.spawn_blocking_child(move || {
699 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
700 : // *before* starting postgres.
701 : //
702 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
703 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
704 : // OOM-killed during startup because swap wasn't available yet.
705 0 : resize_swap(size_bytes).context("failed to resize swap")?;
706 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
707 0 : info!(%size_bytes, %size_mib, "resized swap");
708 :
709 0 : Ok::<(), anyhow::Error>(())
710 0 : });
711 0 : }
712 :
713 : // Set disk quota if the compute spec says so
714 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) = (
715 0 : pspec.spec.disk_quota_bytes,
716 0 : self.params.set_disk_quota_for_fs.as_ref(),
717 : ) {
718 0 : let disk_quota_fs_mountpoint = disk_quota_fs_mountpoint.clone();
719 0 : pre_tasks.spawn_blocking_child(move || {
720 0 : set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint)
721 0 : .context("failed to set disk quota")?;
722 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
723 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
724 :
725 0 : Ok::<(), anyhow::Error>(())
726 0 : });
727 0 : }
728 :
729 : // tune pgbouncer
730 0 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
731 0 : info!("tuning pgbouncer");
732 :
733 0 : let pgbouncer_settings = pgbouncer_settings.clone();
734 0 : let tls_config = tls_config.clone();
735 :
736 : // Spawn a background task to do the tuning,
737 : // so that we don't block the main thread that starts Postgres.
738 0 : let _handle = tokio::spawn(async move {
739 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
740 0 : if let Err(err) = res {
741 0 : error!("error while tuning pgbouncer: {err:?}");
742 : // Continue with the startup anyway
743 0 : }
744 0 : });
745 0 : }
746 :
747 : // configure local_proxy
748 0 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
749 0 : info!("configuring local_proxy");
750 :
751 : // Spawn a background task to do the configuration,
752 : // so that we don't block the main thread that starts Postgres.
753 :
754 0 : let mut local_proxy = local_proxy.clone();
755 0 : local_proxy.tls = tls_config.clone();
756 :
757 0 : let _handle = tokio::spawn(async move {
758 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
759 0 : error!("error while configuring local_proxy: {err:?}");
760 : // Continue with the startup anyway
761 0 : }
762 0 : });
763 0 : }
764 :
765 : // Configure and start rsyslog for compliance audit logging
766 0 : match pspec.spec.audit_log_level {
767 : ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
768 0 : let remote_tls_endpoint =
769 0 : std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
770 0 : let remote_plain_endpoint =
771 0 : std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
772 :
773 0 : if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
774 0 : anyhow::bail!(
775 0 : "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
776 : );
777 0 : }
778 :
779 0 : let log_directory_path = Path::new(&self.params.pgdata).join("log");
780 0 : let log_directory_path = log_directory_path.to_string_lossy().to_string();
781 :
782 : // Add project_id,endpoint_id to identify the logs.
783 : //
784 : // These ids are passed from cplane,
785 0 : let endpoint_id = pspec.spec.endpoint_id.as_deref().unwrap_or("");
786 0 : let project_id = pspec.spec.project_id.as_deref().unwrap_or("");
787 :
788 0 : configure_audit_rsyslog(
789 0 : log_directory_path.clone(),
790 0 : endpoint_id,
791 0 : project_id,
792 0 : &remote_plain_endpoint,
793 0 : &remote_tls_endpoint,
794 0 : )?;
795 :
796 : // Launch a background task to clean up the audit logs
797 0 : launch_pgaudit_gc(log_directory_path);
798 : }
799 0 : _ => {}
800 : }
801 :
802 : // Configure and start rsyslog for Postgres logs export
803 0 : let conf = PostgresLogsRsyslogConfig::new(pspec.spec.logs_export_host.as_deref());
804 0 : configure_postgres_logs_export(conf)?;
805 :
806 : // Launch remaining service threads
807 0 : let _monitor_handle = launch_monitor(self);
808 0 : let _configurator_handle = launch_configurator(self);
809 :
810 : // Wait for all the pre-tasks to finish before starting postgres
811 0 : let rt = tokio::runtime::Handle::current();
812 0 : while let Some(res) = rt.block_on(pre_tasks.join_next()) {
813 0 : res??;
814 : }
815 :
816 : ////// START POSTGRES
817 0 : let start_time = Utc::now();
818 0 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
819 0 : let postmaster_pid = pg_process.pid();
820 0 : *pg_handle = Some(pg_process);
821 :
822 : // If this is a primary endpoint, perform some post-startup configuration before
823 : // opening it up for the world.
824 0 : let config_time = Utc::now();
825 0 : if pspec.spec.mode == ComputeMode::Primary {
826 0 : self.configure_as_primary(&compute_state)?;
827 :
828 0 : let conf = self.get_tokio_conn_conf(None);
829 0 : tokio::task::spawn(async {
830 0 : let _ = installed_extensions(conf).await;
831 0 : });
832 0 : }
833 :
834 : // All done!
835 0 : let startup_end_time = Utc::now();
836 0 : let metrics = {
837 0 : let mut state = self.state.lock().unwrap();
838 0 : state.metrics.start_postgres_ms = config_time
839 0 : .signed_duration_since(start_time)
840 0 : .to_std()
841 0 : .unwrap()
842 0 : .as_millis() as u64;
843 0 : state.metrics.config_ms = startup_end_time
844 0 : .signed_duration_since(config_time)
845 0 : .to_std()
846 0 : .unwrap()
847 0 : .as_millis() as u64;
848 0 : state.metrics.total_startup_ms = startup_end_time
849 0 : .signed_duration_since(compute_state.start_time)
850 0 : .to_std()
851 0 : .unwrap()
852 0 : .as_millis() as u64;
853 0 : state.metrics.clone()
854 : };
855 0 : self.set_status(ComputeStatus::Running);
856 :
857 : // Log metrics so that we can search for slow operations in logs
858 0 : info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
859 :
860 0 : self.spawn_extension_stats_task();
861 :
862 0 : if pspec.spec.autoprewarm {
863 0 : info!("autoprewarming on startup as requested");
864 0 : self.prewarm_lfc(None);
865 0 : }
866 0 : if let Some(seconds) = pspec.spec.offload_lfc_interval_seconds {
867 0 : self.spawn_lfc_offload_task(Duration::from_secs(seconds.into()));
868 0 : };
869 0 : Ok(())
870 0 : }
871 :
872 : #[instrument(skip_all)]
873 : async fn download_preload_extensions(&self, spec: &ComputeSpec) -> Result<()> {
874 : let remote_extensions = if let Some(remote_extensions) = &spec.remote_extensions {
875 : remote_extensions
876 : } else {
877 : return Ok(());
878 : };
879 :
880 : // First, create control files for all available extensions
881 : extension_server::create_control_files(remote_extensions, &self.params.pgbin);
882 :
883 : let library_load_start_time = Utc::now();
884 : let remote_ext_metrics = self.prepare_preload_libraries(spec).await?;
885 :
886 : let library_load_time = Utc::now()
887 : .signed_duration_since(library_load_start_time)
888 : .to_std()
889 : .unwrap()
890 : .as_millis() as u64;
891 : let mut state = self.state.lock().unwrap();
892 : state.metrics.load_ext_ms = library_load_time;
893 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
894 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
895 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
896 : info!(
897 : "Loading shared_preload_libraries took {:?}ms",
898 : library_load_time
899 : );
900 : info!("{:?}", remote_ext_metrics);
901 :
902 : Ok(())
903 : }
904 :
905 : /// Start the vm-monitor if directed to. The vm-monitor only runs on linux
906 : /// because it requires cgroups.
907 0 : fn start_vm_monitor(&self, disable_lfc_resizing: bool) -> StartVmMonitorResult {
908 : cfg_if::cfg_if! {
909 : if #[cfg(target_os = "linux")] {
910 : use std::env;
911 : use tokio_util::sync::CancellationToken;
912 :
913 : // This token is used internally by the monitor to clean up all threads
914 0 : let token = CancellationToken::new();
915 :
916 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
917 0 : let pgconnstr = if disable_lfc_resizing {
918 0 : None
919 : } else {
920 0 : Some(self.params.filecache_connstr.clone())
921 : };
922 :
923 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
924 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
925 0 : Box::leak(Box::new(vm_monitor::Args {
926 0 : cgroup: Some(self.params.cgroup.clone()),
927 0 : pgconnstr,
928 0 : addr: self.params.vm_monitor_addr.clone(),
929 0 : })),
930 0 : token.clone(),
931 : ));
932 0 : Some(vm_monitor)
933 : } else {
934 0 : None
935 : };
936 0 : StartVmMonitorResult { token, vm_monitor }
937 : } else {
938 : _ = disable_lfc_resizing; // appease unused lint on macOS
939 : StartVmMonitorResult { }
940 : }
941 : }
942 0 : }
943 :
944 0 : fn cleanup_after_postgres_exit(&self) -> Result<bool> {
945 : // Maybe sync safekeepers again, to speed up next startup
946 0 : let compute_state = self.state.lock().unwrap().clone();
947 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
948 0 : let lsn = if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
949 0 : info!("syncing safekeepers on shutdown");
950 0 : let storage_auth_token = pspec.storage_auth_token.clone();
951 0 : let lsn = self.sync_safekeepers(storage_auth_token)?;
952 0 : info!(%lsn, "synced safekeepers");
953 0 : Some(lsn)
954 : } else {
955 0 : info!("not primary, not syncing safekeepers");
956 0 : None
957 : };
958 :
959 0 : let mut state = self.state.lock().unwrap();
960 0 : state.terminate_flush_lsn = lsn;
961 :
962 0 : let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
963 0 : if state.status == ComputeStatus::TerminationPendingFast
964 0 : || state.status == ComputeStatus::TerminationPendingImmediate
965 : {
966 0 : info!(
967 0 : "Changing compute status from {} to {}",
968 0 : state.status,
969 : ComputeStatus::Terminated
970 : );
971 0 : state.status = ComputeStatus::Terminated;
972 0 : self.state_changed.notify_all();
973 0 : }
974 0 : drop(state);
975 :
976 0 : if let Err(err) = self.check_for_core_dumps() {
977 0 : error!("error while checking for core dumps: {err:?}");
978 0 : }
979 :
980 0 : Ok(delay_exit)
981 0 : }
982 :
983 : /// Check that compute node has corresponding feature enabled.
984 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
985 0 : let state = self.state.lock().unwrap();
986 :
987 0 : if let Some(s) = state.pspec.as_ref() {
988 0 : s.spec.features.contains(&feature)
989 : } else {
990 0 : false
991 : }
992 0 : }
993 :
994 0 : pub fn set_status(&self, status: ComputeStatus) {
995 0 : let mut state = self.state.lock().unwrap();
996 0 : state.set_status(status, &self.state_changed);
997 0 : }
998 :
999 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
1000 0 : let mut state = self.state.lock().unwrap();
1001 0 : state.set_failed_status(err, &self.state_changed);
1002 0 : }
1003 :
1004 0 : pub fn get_status(&self) -> ComputeStatus {
1005 0 : self.state.lock().unwrap().status
1006 0 : }
1007 :
1008 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
1009 0 : self.state
1010 0 : .lock()
1011 0 : .unwrap()
1012 0 : .pspec
1013 0 : .as_ref()
1014 0 : .map(|s| s.timeline_id)
1015 0 : }
1016 :
1017 : // Remove `pgdata` directory and create it again with right permissions.
1018 0 : fn create_pgdata(&self) -> Result<()> {
1019 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
1020 : // If it is something different then create_dir() will error out anyway.
1021 0 : let pgdata = &self.params.pgdata;
1022 0 : let _ok = fs::remove_dir_all(pgdata);
1023 0 : fs::create_dir(pgdata)?;
1024 0 : fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
1025 :
1026 0 : Ok(())
1027 0 : }
1028 :
1029 : /// Fetches a basebackup from the Pageserver using the compute state's Pageserver connstring and
1030 : /// unarchives it to `pgdata` directory, replacing any existing contents.
1031 : #[instrument(skip_all, fields(%lsn))]
1032 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1033 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
1034 :
1035 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
1036 : let started = Instant::now();
1037 :
1038 : let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
1039 : PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
1040 : PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
1041 : };
1042 :
1043 : self.fix_zenith_signal_neon_signal()?;
1044 :
1045 : let mut state = self.state.lock().unwrap();
1046 : state.metrics.pageserver_connect_micros =
1047 : connected.duration_since(started).as_micros() as u64;
1048 : state.metrics.basebackup_bytes = size as u64;
1049 : state.metrics.basebackup_ms = started.elapsed().as_millis() as u64;
1050 :
1051 : Ok(())
1052 : }
1053 :
1054 : /// Move the Zenith signal file to Neon signal file location.
1055 : /// This makes Compute compatible with older PageServers that don't yet
1056 : /// know about the Zenith->Neon rename.
1057 0 : fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
1058 0 : let datadir = Path::new(&self.params.pgdata);
1059 :
1060 0 : let neonsig = datadir.join("neon.signal");
1061 :
1062 0 : if neonsig.is_file() {
1063 0 : return Ok(());
1064 0 : }
1065 :
1066 0 : let zenithsig = datadir.join("zenith.signal");
1067 :
1068 0 : if zenithsig.is_file() {
1069 0 : fs::copy(zenithsig, neonsig)?;
1070 0 : }
1071 :
1072 0 : Ok(())
1073 0 : }
1074 :
1075 : /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
1076 : /// the connection was established, and the (compressed) size of the basebackup.
1077 0 : fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1078 0 : let shard0_connstr = spec
1079 0 : .pageserver_connstr
1080 0 : .split(',')
1081 0 : .next()
1082 0 : .unwrap()
1083 0 : .to_string();
1084 0 : let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
1085 0 : 0 | 1 => ShardIndex::unsharded(),
1086 0 : count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
1087 : };
1088 :
1089 0 : let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
1090 0 : let mut client = page_api::Client::connect(
1091 0 : shard0_connstr,
1092 0 : spec.tenant_id,
1093 0 : spec.timeline_id,
1094 0 : shard_index,
1095 0 : spec.storage_auth_token.clone(),
1096 0 : None, // NB: base backups use payload compression
1097 0 : )
1098 0 : .await?;
1099 0 : let connected = Instant::now();
1100 0 : let reader = client
1101 0 : .get_base_backup(page_api::GetBaseBackupRequest {
1102 0 : lsn: (lsn != Lsn(0)).then_some(lsn),
1103 0 : compression: BaseBackupCompression::Gzip,
1104 0 : replica: spec.spec.mode != ComputeMode::Primary,
1105 0 : full: false,
1106 0 : })
1107 0 : .await?;
1108 0 : anyhow::Ok((reader, connected))
1109 0 : })?;
1110 :
1111 0 : let mut reader = MeasuredReader::new(tokio_util::io::SyncIoBridge::new(reader));
1112 :
1113 : // Set `ignore_zeros` so that unpack() reads the entire stream and doesn't just stop at the
1114 : // end-of-archive marker. If the server errors, the tar::Builder drop handler will write an
1115 : // end-of-archive marker before the error is emitted, and we would not see the error.
1116 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut reader));
1117 0 : ar.set_ignore_zeros(true);
1118 0 : ar.unpack(&self.params.pgdata)?;
1119 :
1120 0 : Ok((connected, reader.get_byte_count()))
1121 0 : }
1122 :
1123 : /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
1124 : /// when the connection was established, and the (compressed) size of the basebackup.
1125 0 : fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1126 0 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
1127 0 : let mut config = postgres::Config::from_str(shard0_connstr)?;
1128 :
1129 : // Use the storage auth token from the config file, if given.
1130 : // Note: this overrides any password set in the connection string.
1131 0 : if let Some(storage_auth_token) = &spec.storage_auth_token {
1132 0 : info!("Got storage auth token from spec file");
1133 0 : config.password(storage_auth_token);
1134 : } else {
1135 0 : info!("Storage auth token not set");
1136 : }
1137 :
1138 0 : config.application_name("compute_ctl");
1139 0 : config.options(&format!(
1140 0 : "-c neon.compute_mode={}",
1141 0 : spec.spec.mode.to_type_str()
1142 0 : ));
1143 :
1144 : // Connect to pageserver
1145 0 : let mut client = config.connect(NoTls)?;
1146 0 : let connected = Instant::now();
1147 :
1148 0 : let basebackup_cmd = match lsn {
1149 : Lsn(0) => {
1150 0 : if spec.spec.mode != ComputeMode::Primary {
1151 0 : format!(
1152 0 : "basebackup {} {} --gzip --replica",
1153 : spec.tenant_id, spec.timeline_id
1154 : )
1155 : } else {
1156 0 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
1157 : }
1158 : }
1159 : _ => {
1160 0 : if spec.spec.mode != ComputeMode::Primary {
1161 0 : format!(
1162 0 : "basebackup {} {} {} --gzip --replica",
1163 : spec.tenant_id, spec.timeline_id, lsn
1164 : )
1165 : } else {
1166 0 : format!(
1167 0 : "basebackup {} {} {} --gzip",
1168 : spec.tenant_id, spec.timeline_id, lsn
1169 : )
1170 : }
1171 : }
1172 : };
1173 :
1174 0 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
1175 0 : let mut measured_reader = MeasuredReader::new(copyreader);
1176 0 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
1177 :
1178 : // Read the archive directly from the `CopyOutReader`
1179 : //
1180 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
1181 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
1182 : // sends an Error after finishing the tarball, we will not notice it.
1183 : // The tar::Builder drop handler will write an end-of-archive marker
1184 : // before emitting the error, and we would not see it otherwise.
1185 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
1186 0 : ar.set_ignore_zeros(true);
1187 0 : ar.unpack(&self.params.pgdata)?;
1188 :
1189 0 : Ok((connected, measured_reader.get_byte_count()))
1190 0 : }
1191 :
1192 : // Gets the basebackup in a retry loop
1193 : #[instrument(skip_all, fields(%lsn))]
1194 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1195 : let mut retry_period_ms = 500.0;
1196 : let mut attempts = 0;
1197 : const DEFAULT_ATTEMPTS: u16 = 10;
1198 : #[cfg(feature = "testing")]
1199 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
1200 : u16::from_str(&v).unwrap()
1201 : } else {
1202 : DEFAULT_ATTEMPTS
1203 : };
1204 : #[cfg(not(feature = "testing"))]
1205 : let max_attempts = DEFAULT_ATTEMPTS;
1206 : loop {
1207 : let result = self.try_get_basebackup(compute_state, lsn);
1208 : match result {
1209 : Ok(_) => {
1210 : return result;
1211 : }
1212 : Err(ref e) if attempts < max_attempts => {
1213 : warn!(
1214 : "Failed to get basebackup: {} (attempt {}/{})",
1215 : e, attempts, max_attempts
1216 : );
1217 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
1218 : retry_period_ms *= 1.5;
1219 : }
1220 : Err(_) => {
1221 : return result;
1222 : }
1223 : }
1224 : attempts += 1;
1225 : }
1226 : }
1227 :
1228 0 : pub async fn check_safekeepers_synced_async(
1229 0 : &self,
1230 0 : compute_state: &ComputeState,
1231 0 : ) -> Result<Option<Lsn>> {
1232 : // Construct a connection config for each safekeeper
1233 0 : let pspec: ParsedSpec = compute_state
1234 0 : .pspec
1235 0 : .as_ref()
1236 0 : .expect("spec must be set")
1237 0 : .clone();
1238 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
1239 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
1240 : // Format connstr
1241 0 : let id = connstr.clone();
1242 0 : let connstr = format!("postgresql://no_user@{connstr}");
1243 0 : let options = format!(
1244 0 : "-c timeline_id={} tenant_id={}",
1245 : pspec.timeline_id, pspec.tenant_id
1246 : );
1247 :
1248 : // Construct client
1249 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
1250 0 : config.options(&options);
1251 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
1252 0 : config.password(storage_auth_token);
1253 0 : }
1254 :
1255 0 : (id, config)
1256 0 : });
1257 :
1258 : // Create task set to query all safekeepers
1259 0 : let mut tasks = FuturesUnordered::new();
1260 0 : let quorum = sk_configs.len() / 2 + 1;
1261 0 : for (id, config) in sk_configs {
1262 0 : let timeout = tokio::time::Duration::from_millis(100);
1263 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
1264 0 : tasks.push(tokio::spawn(task));
1265 0 : }
1266 :
1267 : // Get a quorum of responses or errors
1268 0 : let mut responses = Vec::new();
1269 0 : let mut join_errors = Vec::new();
1270 0 : let mut task_errors = Vec::new();
1271 0 : let mut timeout_errors = Vec::new();
1272 0 : while let Some(response) = tasks.next().await {
1273 0 : match response {
1274 0 : Ok(Ok(Ok(r))) => responses.push(r),
1275 0 : Ok(Ok(Err(e))) => task_errors.push(e),
1276 0 : Ok(Err(e)) => timeout_errors.push(e),
1277 0 : Err(e) => join_errors.push(e),
1278 : };
1279 0 : if responses.len() >= quorum {
1280 0 : break;
1281 0 : }
1282 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
1283 0 : break;
1284 0 : }
1285 : }
1286 :
1287 : // In case of error, log and fail the check, but don't crash.
1288 : // We're playing it safe because these errors could be transient
1289 : // and we don't yet retry.
1290 0 : if responses.len() < quorum {
1291 0 : error!(
1292 0 : "failed sync safekeepers check {:?} {:?} {:?}",
1293 : join_errors, task_errors, timeout_errors
1294 : );
1295 0 : return Ok(None);
1296 0 : }
1297 :
1298 0 : Ok(check_if_synced(responses))
1299 0 : }
1300 :
1301 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
1302 : // in one roundtrip. If not, we should do a full sync_safekeepers.
1303 : #[instrument(skip_all)]
1304 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
1305 : let start_time = Utc::now();
1306 :
1307 : let rt = tokio::runtime::Handle::current();
1308 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
1309 :
1310 : // Record runtime
1311 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
1312 : .signed_duration_since(start_time)
1313 : .to_std()
1314 : .unwrap()
1315 : .as_millis() as u64;
1316 : result
1317 : }
1318 :
1319 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
1320 : // and return the reported LSN back to the caller.
1321 : #[instrument(skip_all)]
1322 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
1323 : let start_time = Utc::now();
1324 :
1325 : let mut sync_handle = maybe_cgexec(&self.params.pgbin)
1326 : .args(["--sync-safekeepers"])
1327 : .env("PGDATA", &self.params.pgdata) // we cannot use -D in this mode
1328 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1329 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1330 : } else {
1331 : vec![]
1332 : })
1333 : .stdout(Stdio::piped())
1334 : .stderr(Stdio::piped())
1335 : .spawn()
1336 : .expect("postgres --sync-safekeepers failed to start");
1337 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
1338 :
1339 : // `postgres --sync-safekeepers` will print all log output to stderr and
1340 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
1341 : // will be collected in a child thread.
1342 : let stderr = sync_handle
1343 : .stderr
1344 : .take()
1345 : .expect("stderr should be captured");
1346 : let logs_handle = handle_postgres_logs(stderr);
1347 :
1348 : let sync_output = sync_handle
1349 : .wait_with_output()
1350 : .expect("postgres --sync-safekeepers failed");
1351 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
1352 :
1353 : // Process has exited, so we can join the logs thread.
1354 : let _ = tokio::runtime::Handle::current()
1355 : .block_on(logs_handle)
1356 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1357 :
1358 : if !sync_output.status.success() {
1359 : anyhow::bail!(
1360 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
1361 : sync_output.status,
1362 : String::from_utf8(sync_output.stdout)
1363 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
1364 : );
1365 : }
1366 :
1367 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
1368 : .signed_duration_since(start_time)
1369 : .to_std()
1370 : .unwrap()
1371 : .as_millis() as u64;
1372 :
1373 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
1374 :
1375 : Ok(lsn)
1376 : }
1377 :
1378 : /// Do all the preparations like PGDATA directory creation, configuration,
1379 : /// safekeepers sync, basebackup, etc.
1380 : #[instrument(skip_all)]
1381 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
1382 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1383 : let spec = &pspec.spec;
1384 : let pgdata_path = Path::new(&self.params.pgdata);
1385 :
1386 : let tls_config = self.tls_config(&pspec.spec);
1387 :
1388 : // Remove/create an empty pgdata directory and put configuration there.
1389 : self.create_pgdata()?;
1390 : config::write_postgres_conf(
1391 : pgdata_path,
1392 : &pspec.spec,
1393 : self.params.internal_http_port,
1394 : tls_config,
1395 : )?;
1396 :
1397 : // Syncing safekeepers is only safe with primary nodes: if a primary
1398 : // is already connected it will be kicked out, so a secondary (standby)
1399 : // cannot sync safekeepers.
1400 : let lsn = match spec.mode {
1401 : ComputeMode::Primary => {
1402 : info!("checking if safekeepers are synced");
1403 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
1404 : lsn
1405 : } else {
1406 : info!("starting safekeepers syncing");
1407 : self.sync_safekeepers(pspec.storage_auth_token.clone())
1408 : .with_context(|| "failed to sync safekeepers")?
1409 : };
1410 : info!("safekeepers synced at LSN {}", lsn);
1411 : lsn
1412 : }
1413 : ComputeMode::Static(lsn) => {
1414 : info!("Starting read-only node at static LSN {}", lsn);
1415 : lsn
1416 : }
1417 : ComputeMode::Replica => {
1418 : info!("Initializing standby from latest Pageserver LSN");
1419 : Lsn(0)
1420 : }
1421 : };
1422 :
1423 : info!(
1424 : "getting basebackup@{} from pageserver {}",
1425 : lsn, &pspec.pageserver_connstr
1426 : );
1427 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
1428 0 : format!(
1429 0 : "failed to get basebackup@{} from pageserver {}",
1430 0 : lsn, &pspec.pageserver_connstr
1431 : )
1432 0 : })?;
1433 :
1434 : // Update pg_hba.conf received with basebackup.
1435 : update_pg_hba(pgdata_path)?;
1436 :
1437 : // Place pg_dynshmem under /dev/shm. This allows us to use
1438 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
1439 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
1440 : //
1441 : // Why on earth don't we just stick to the 'posix' default, you might
1442 : // ask. It turns out that making large allocations with 'posix' doesn't
1443 : // work very well with autoscaling. The behavior we want is that:
1444 : //
1445 : // 1. You can make large DSM allocations, larger than the current RAM
1446 : // size of the VM, without errors
1447 : //
1448 : // 2. If the allocated memory is really used, the VM is scaled up
1449 : // automatically to accommodate that
1450 : //
1451 : // We try to make that possible by having swap in the VM. But with the
1452 : // default 'posix' DSM implementation, we fail step 1, even when there's
1453 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
1454 : // the shmem segment, which is really just a file in /dev/shm in Linux,
1455 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
1456 : // than available RAM.
1457 : //
1458 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
1459 : // the Postgres 'mmap' DSM implementation doesn't use
1460 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
1461 : // fill the file with zeros. It's weird that that differs between
1462 : // 'posix' and 'mmap', but we take advantage of it. When the file is
1463 : // filled slowly with write(2), the kernel allows it to grow larger, as
1464 : // long as there's swap available.
1465 : //
1466 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
1467 : // segment to be larger than currently available RAM. But because we
1468 : // don't want to store it on a real file, which the kernel would try to
1469 : // flush to disk, so symlink pg_dynshm to /dev/shm.
1470 : //
1471 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
1472 : // control plane control that option. If 'mmap' is not used, this
1473 : // symlink doesn't affect anything.
1474 : //
1475 : // See https://github.com/neondatabase/autoscaling/issues/800
1476 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
1477 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
1478 :
1479 : match spec.mode {
1480 : ComputeMode::Primary => {}
1481 : ComputeMode::Replica | ComputeMode::Static(..) => {
1482 : add_standby_signal(pgdata_path)?;
1483 : }
1484 : }
1485 :
1486 : Ok(())
1487 : }
1488 :
1489 : /// Start and stop a postgres process to warm up the VM for startup.
1490 0 : pub fn prewarm_postgres_vm_memory(&self) -> Result<()> {
1491 0 : info!("prewarming VM memory");
1492 :
1493 : // Create pgdata
1494 0 : let pgdata = &format!("{}.warmup", self.params.pgdata);
1495 0 : create_pgdata(pgdata)?;
1496 :
1497 : // Run initdb to completion
1498 0 : info!("running initdb");
1499 0 : let initdb_bin = Path::new(&self.params.pgbin)
1500 0 : .parent()
1501 0 : .unwrap()
1502 0 : .join("initdb");
1503 0 : Command::new(initdb_bin)
1504 0 : .args(["--pgdata", pgdata])
1505 0 : .output()
1506 0 : .expect("cannot start initdb process");
1507 :
1508 : // Write conf
1509 : use std::io::Write;
1510 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
1511 0 : let mut file = std::fs::File::create(conf_path)?;
1512 0 : writeln!(file, "shared_buffers=65536")?;
1513 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
1514 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
1515 :
1516 : // Start postgres
1517 0 : info!("starting postgres");
1518 0 : let mut pg = maybe_cgexec(&self.params.pgbin)
1519 0 : .args(["-D", pgdata])
1520 0 : .spawn()
1521 0 : .expect("cannot start postgres process");
1522 :
1523 : // Stop it when it's ready
1524 0 : info!("waiting for postgres");
1525 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
1526 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
1527 : // it to avoid orphaned processes prowling around while datadir is
1528 : // wiped.
1529 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
1530 0 : kill(pm_pid, Signal::SIGQUIT)?;
1531 0 : info!("sent SIGQUIT signal");
1532 0 : pg.wait()?;
1533 0 : info!("done prewarming vm memory");
1534 :
1535 : // clean up
1536 0 : let _ok = fs::remove_dir_all(pgdata);
1537 0 : Ok(())
1538 0 : }
1539 :
1540 : /// Start Postgres as a child process and wait for it to start accepting
1541 : /// connections.
1542 : ///
1543 : /// Returns a handle to the child process and a handle to the logs thread.
1544 : #[instrument(skip_all)]
1545 : pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
1546 : let pgdata_path = Path::new(&self.params.pgdata);
1547 :
1548 : // Run postgres as a child process.
1549 : let mut pg = maybe_cgexec(&self.params.pgbin)
1550 : .args(["-D", &self.params.pgdata])
1551 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1552 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1553 : } else {
1554 : vec![]
1555 : })
1556 : .stderr(Stdio::piped())
1557 : .spawn()
1558 : .expect("cannot start postgres process");
1559 : PG_PID.store(pg.id(), Ordering::SeqCst);
1560 :
1561 : // Start a task to collect logs from stderr.
1562 : let stderr = pg.stderr.take().expect("stderr should be captured");
1563 : let logs_handle = handle_postgres_logs(stderr);
1564 :
1565 : wait_for_postgres(&mut pg, pgdata_path)?;
1566 :
1567 : Ok(PostgresHandle {
1568 : postgres: pg,
1569 : log_collector: logs_handle,
1570 : })
1571 : }
1572 :
1573 : /// Wait for the child Postgres process forever. In this state Ctrl+C will
1574 : /// propagate to Postgres and it will be shut down as well.
1575 0 : fn wait_postgres(&self, mut pg_handle: PostgresHandle) -> std::process::ExitStatus {
1576 0 : info!(postmaster_pid = %pg_handle.postgres.id(), "Waiting for Postgres to exit");
1577 :
1578 0 : let ecode = pg_handle
1579 0 : .postgres
1580 0 : .wait()
1581 0 : .expect("failed to start waiting on Postgres process");
1582 0 : PG_PID.store(0, Ordering::SeqCst);
1583 :
1584 : // Process has exited. Wait for the log collecting task to finish.
1585 0 : let _ = tokio::runtime::Handle::current()
1586 0 : .block_on(pg_handle.log_collector)
1587 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1588 :
1589 0 : ecode
1590 0 : }
1591 :
1592 : /// Do post configuration of the already started Postgres. This function spawns a background task to
1593 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
1594 : /// version. In the future, it may upgrade all 3rd-party extensions.
1595 : #[instrument(skip_all)]
1596 : pub fn post_apply_config(&self) -> Result<()> {
1597 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
1598 0 : tokio::spawn(async move {
1599 0 : let res = async {
1600 0 : let (mut client, connection) = conf.connect(NoTls).await?;
1601 0 : tokio::spawn(async move {
1602 0 : if let Err(e) = connection.await {
1603 0 : eprintln!("connection error: {e}");
1604 0 : }
1605 0 : });
1606 :
1607 0 : handle_neon_extension_upgrade(&mut client)
1608 0 : .await
1609 0 : .context("handle_neon_extension_upgrade")?;
1610 0 : Ok::<_, anyhow::Error>(())
1611 0 : }
1612 0 : .await;
1613 0 : if let Err(err) = res {
1614 0 : error!("error while post_apply_config: {err:#}");
1615 0 : }
1616 0 : });
1617 : Ok(())
1618 : }
1619 :
1620 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
1621 0 : let mut conf = self.conn_conf.clone();
1622 0 : if let Some(application_name) = application_name {
1623 0 : conf.application_name(application_name);
1624 0 : }
1625 0 : conf
1626 0 : }
1627 :
1628 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
1629 0 : let mut conf = self.tokio_conn_conf.clone();
1630 0 : if let Some(application_name) = application_name {
1631 0 : conf.application_name(application_name);
1632 0 : }
1633 0 : conf
1634 0 : }
1635 :
1636 0 : pub async fn get_maintenance_client(
1637 0 : conf: &tokio_postgres::Config,
1638 0 : ) -> Result<tokio_postgres::Client> {
1639 0 : let mut conf = conf.clone();
1640 0 : conf.application_name("compute_ctl:apply_config");
1641 :
1642 0 : let (client, conn) = match conf.connect(NoTls).await {
1643 : // If connection fails, it may be the old node with `zenith_admin` superuser.
1644 : //
1645 : // In this case we need to connect with old `zenith_admin` name
1646 : // and create new user. We cannot simply rename connected user,
1647 : // but we can create a new one and grant it all privileges.
1648 0 : Err(e) => match e.code() {
1649 : Some(&SqlState::INVALID_PASSWORD)
1650 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
1651 : // Connect with `zenith_admin` if `cloud_admin` could not authenticate
1652 0 : info!(
1653 0 : "cannot connect to Postgres: {}, retrying with 'zenith_admin' username",
1654 : e
1655 : );
1656 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
1657 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
1658 0 : zenith_admin_conf.user("zenith_admin");
1659 :
1660 : // It doesn't matter what were the options before, here we just want
1661 : // to connect and create a new superuser role.
1662 : const ZENITH_OPTIONS: &str = "-c role=zenith_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0";
1663 0 : zenith_admin_conf.options(ZENITH_OPTIONS);
1664 :
1665 0 : let mut client =
1666 0 : zenith_admin_conf.connect(NoTls)
1667 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
1668 :
1669 : // Disable forwarding so that users don't get a cloud_admin role
1670 0 : let mut func = || {
1671 0 : client.simple_query("SET neon.forward_ddl = false")?;
1672 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
1673 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
1674 0 : Ok::<_, anyhow::Error>(())
1675 0 : };
1676 0 : func().context("apply_config setup cloud_admin")?;
1677 :
1678 0 : drop(client);
1679 :
1680 : // Reconnect with connstring with expected name
1681 0 : conf.connect(NoTls).await?
1682 : }
1683 0 : _ => return Err(e.into()),
1684 : },
1685 0 : Ok((client, conn)) => (client, conn),
1686 : };
1687 :
1688 0 : spawn(async move {
1689 0 : if let Err(e) = conn.await {
1690 0 : error!("maintenance client connection error: {}", e);
1691 0 : }
1692 0 : });
1693 :
1694 : // Disable DDL forwarding because control plane already knows about the roles/databases
1695 : // we're about to modify.
1696 0 : client
1697 0 : .simple_query("SET neon.forward_ddl = false")
1698 0 : .await
1699 0 : .context("apply_config SET neon.forward_ddl = false")?;
1700 :
1701 0 : Ok(client)
1702 0 : }
1703 :
1704 : /// Do initial configuration of the already started Postgres.
1705 : #[instrument(skip_all)]
1706 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1707 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1708 :
1709 : let conf = Arc::new(conf);
1710 : let spec = Arc::new(
1711 : compute_state
1712 : .pspec
1713 : .as_ref()
1714 : .expect("spec must be set")
1715 : .spec
1716 : .clone(),
1717 : );
1718 :
1719 : let mut tls_config = None::<TlsConfig>;
1720 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
1721 : tls_config = self.compute_ctl_config.tls.clone();
1722 : }
1723 :
1724 : self.update_installed_extensions_collection_interval(&spec);
1725 :
1726 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1727 :
1728 : // Merge-apply spec & changes to PostgreSQL state.
1729 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1730 :
1731 : if let Some(local_proxy) = &spec.clone().local_proxy_config {
1732 : let mut local_proxy = local_proxy.clone();
1733 : local_proxy.tls = tls_config.clone();
1734 :
1735 : info!("configuring local_proxy");
1736 : local_proxy::configure(&local_proxy).context("apply_config local_proxy")?;
1737 : }
1738 :
1739 : // Run migrations separately to not hold up cold starts
1740 0 : tokio::spawn(async move {
1741 0 : let mut conf = conf.as_ref().clone();
1742 0 : conf.application_name("compute_ctl:migrations");
1743 :
1744 0 : match conf.connect(NoTls).await {
1745 0 : Ok((mut client, connection)) => {
1746 0 : tokio::spawn(async move {
1747 0 : if let Err(e) = connection.await {
1748 0 : eprintln!("connection error: {e}");
1749 0 : }
1750 0 : });
1751 0 : if let Err(e) = handle_migrations(&mut client).await {
1752 0 : error!("Failed to run migrations: {}", e);
1753 0 : }
1754 : }
1755 0 : Err(e) => {
1756 0 : error!(
1757 0 : "Failed to connect to the compute for running migrations: {}",
1758 : e
1759 : );
1760 : }
1761 : };
1762 0 : });
1763 :
1764 : Ok::<(), anyhow::Error>(())
1765 : }
1766 :
1767 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1768 : // `pg_ctl` for start / stop.
1769 : #[instrument(skip_all)]
1770 : fn pg_reload_conf(&self) -> Result<()> {
1771 : let pgctl_bin = Path::new(&self.params.pgbin)
1772 : .parent()
1773 : .unwrap()
1774 : .join("pg_ctl");
1775 : Command::new(pgctl_bin)
1776 : .args(["reload", "-D", &self.params.pgdata])
1777 : .output()
1778 : .expect("cannot run pg_ctl process");
1779 : Ok(())
1780 : }
1781 :
1782 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1783 : /// as it's used to reconfigure a previously started and configured Postgres node.
1784 : #[instrument(skip_all)]
1785 : pub fn reconfigure(&self) -> Result<()> {
1786 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1787 :
1788 : let tls_config = self.tls_config(&spec);
1789 :
1790 : self.update_installed_extensions_collection_interval(&spec);
1791 :
1792 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1793 : info!("tuning pgbouncer");
1794 :
1795 : let pgbouncer_settings = pgbouncer_settings.clone();
1796 : let tls_config = tls_config.clone();
1797 :
1798 : // Spawn a background task to do the tuning,
1799 : // so that we don't block the main thread that starts Postgres.
1800 0 : tokio::spawn(async move {
1801 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
1802 0 : if let Err(err) = res {
1803 0 : error!("error while tuning pgbouncer: {err:?}");
1804 0 : }
1805 0 : });
1806 : }
1807 :
1808 : if let Some(ref local_proxy) = spec.local_proxy_config {
1809 : info!("configuring local_proxy");
1810 :
1811 : // Spawn a background task to do the configuration,
1812 : // so that we don't block the main thread that starts Postgres.
1813 : let mut local_proxy = local_proxy.clone();
1814 : local_proxy.tls = tls_config.clone();
1815 0 : tokio::spawn(async move {
1816 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1817 0 : error!("error while configuring local_proxy: {err:?}");
1818 0 : }
1819 0 : });
1820 : }
1821 :
1822 : // Reconfigure rsyslog for Postgres logs export
1823 : let conf = PostgresLogsRsyslogConfig::new(spec.logs_export_host.as_deref());
1824 : configure_postgres_logs_export(conf)?;
1825 :
1826 : // Write new config
1827 : let pgdata_path = Path::new(&self.params.pgdata);
1828 : config::write_postgres_conf(
1829 : pgdata_path,
1830 : &spec,
1831 : self.params.internal_http_port,
1832 : tls_config,
1833 : )?;
1834 :
1835 : self.pg_reload_conf()?;
1836 :
1837 : if !spec.skip_pg_catalog_updates {
1838 : let max_concurrent_connections = spec.reconfigure_concurrency;
1839 : // Temporarily reset max_cluster_size in config
1840 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1841 : // creating new extensions, roles, etc.
1842 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1843 0 : self.pg_reload_conf()?;
1844 :
1845 0 : if spec.mode == ComputeMode::Primary {
1846 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:reconfigure"));
1847 0 : let conf = Arc::new(conf);
1848 :
1849 0 : let spec = Arc::new(spec.clone());
1850 :
1851 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
1852 0 : }
1853 :
1854 0 : Ok(())
1855 0 : })?;
1856 : self.pg_reload_conf()?;
1857 : }
1858 :
1859 : let unknown_op = "unknown".to_string();
1860 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1861 : info!(
1862 : "finished reconfiguration of compute node for operation {}",
1863 : op_id
1864 : );
1865 :
1866 : Ok(())
1867 : }
1868 :
1869 : #[instrument(skip_all)]
1870 : pub fn configure_as_primary(&self, compute_state: &ComputeState) -> Result<()> {
1871 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1872 :
1873 : assert!(pspec.spec.mode == ComputeMode::Primary);
1874 : if !pspec.spec.skip_pg_catalog_updates {
1875 : let pgdata_path = Path::new(&self.params.pgdata);
1876 : // temporarily reset max_cluster_size in config
1877 : // to avoid the possibility of hitting the limit, while we are applying config:
1878 : // creating new extensions, roles, etc...
1879 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1880 0 : self.pg_reload_conf()?;
1881 :
1882 0 : self.apply_config(compute_state)?;
1883 :
1884 0 : Ok(())
1885 0 : })?;
1886 :
1887 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1888 : if config::line_in_file(
1889 : &postgresql_conf_path,
1890 : "neon.disable_logical_replication_subscribers=false",
1891 : )? {
1892 : info!(
1893 : "updated postgresql.conf to set neon.disable_logical_replication_subscribers=false"
1894 : );
1895 : }
1896 : self.pg_reload_conf()?;
1897 : }
1898 : self.post_apply_config()?;
1899 :
1900 : Ok(())
1901 : }
1902 :
1903 0 : pub async fn watch_cert_for_changes(self: Arc<Self>) {
1904 : // update status on cert renewal
1905 0 : if let Some(tls_config) = &self.compute_ctl_config.tls {
1906 0 : let tls_config = tls_config.clone();
1907 :
1908 : // wait until the cert exists.
1909 0 : let mut cert_watch = watch_cert_for_changes(tls_config.cert_path.clone()).await;
1910 :
1911 0 : tokio::task::spawn_blocking(move || {
1912 0 : let handle = tokio::runtime::Handle::current();
1913 : 'cert_update: loop {
1914 : // let postgres/pgbouncer/local_proxy know the new cert/key exists.
1915 : // we need to wait until it's configurable first.
1916 :
1917 0 : let mut state = self.state.lock().unwrap();
1918 : 'status_update: loop {
1919 0 : match state.status {
1920 : // let's update the state to config pending
1921 : ComputeStatus::ConfigurationPending | ComputeStatus::Running => {
1922 0 : state.set_status(
1923 0 : ComputeStatus::ConfigurationPending,
1924 0 : &self.state_changed,
1925 0 : );
1926 0 : break 'status_update;
1927 : }
1928 :
1929 : // exit loop
1930 : ComputeStatus::Failed
1931 : | ComputeStatus::TerminationPendingFast
1932 : | ComputeStatus::TerminationPendingImmediate
1933 0 : | ComputeStatus::Terminated => break 'cert_update,
1934 :
1935 : // wait
1936 : ComputeStatus::Init
1937 : | ComputeStatus::Configuration
1938 0 : | ComputeStatus::Empty => {
1939 0 : state = self.state_changed.wait(state).unwrap();
1940 0 : }
1941 : }
1942 : }
1943 0 : drop(state);
1944 :
1945 : // wait for a new certificate update
1946 0 : if handle.block_on(cert_watch.changed()).is_err() {
1947 0 : break;
1948 0 : }
1949 : }
1950 0 : });
1951 0 : }
1952 0 : }
1953 :
1954 0 : pub fn tls_config(&self, spec: &ComputeSpec) -> &Option<TlsConfig> {
1955 0 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
1956 0 : &self.compute_ctl_config.tls
1957 : } else {
1958 0 : &None::<TlsConfig>
1959 : }
1960 0 : }
1961 :
1962 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1963 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1964 0 : let mut state = self.state.lock().unwrap();
1965 : // NB: `Some(<DateTime>)` is always greater than `None`.
1966 0 : if last_active > state.last_active {
1967 0 : state.last_active = last_active;
1968 0 : debug!("set the last compute activity time to: {:?}", last_active);
1969 0 : }
1970 0 : }
1971 :
1972 : // Look for core dumps and collect backtraces.
1973 : //
1974 : // EKS worker nodes have following core dump settings:
1975 : // /proc/sys/kernel/core_pattern -> core
1976 : // /proc/sys/kernel/core_uses_pid -> 1
1977 : // ulimit -c -> unlimited
1978 : // which results in core dumps being written to postgres data directory as core.<pid>.
1979 : //
1980 : // Use that as a default location and pattern, except macos where core dumps are written
1981 : // to /cores/ directory by default.
1982 : //
1983 : // With default Linux settings, the core dump file is called just "core", so check for
1984 : // that too.
1985 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1986 0 : let core_dump_dir = match std::env::consts::OS {
1987 0 : "macos" => Path::new("/cores/"),
1988 0 : _ => Path::new(&self.params.pgdata),
1989 : };
1990 :
1991 : // Collect core dump paths if any
1992 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1993 0 : let files = fs::read_dir(core_dump_dir)?;
1994 0 : let cores = files.filter_map(|entry| {
1995 0 : let entry = entry.ok()?;
1996 :
1997 0 : let is_core_dump = match entry.file_name().to_str()? {
1998 0 : n if n.starts_with("core.") => true,
1999 0 : "core" => true,
2000 0 : _ => false,
2001 : };
2002 0 : if is_core_dump {
2003 0 : Some(entry.path())
2004 : } else {
2005 0 : None
2006 : }
2007 0 : });
2008 :
2009 : // Print backtrace for each core dump
2010 0 : for core_path in cores {
2011 0 : warn!(
2012 0 : "core dump found: {}, collecting backtrace",
2013 0 : core_path.display()
2014 : );
2015 :
2016 : // Try first with gdb
2017 0 : let backtrace = Command::new("gdb")
2018 0 : .args(["--batch", "-q", "-ex", "bt", &self.params.pgbin])
2019 0 : .arg(&core_path)
2020 0 : .output();
2021 :
2022 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
2023 0 : let backtrace = match backtrace {
2024 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
2025 0 : warn!("cannot find gdb, trying lldb");
2026 0 : Command::new("lldb")
2027 0 : .arg("-c")
2028 0 : .arg(&core_path)
2029 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
2030 0 : .output()
2031 : }
2032 0 : _ => backtrace,
2033 0 : }?;
2034 :
2035 0 : warn!(
2036 0 : "core dump backtrace: {}",
2037 0 : String::from_utf8_lossy(&backtrace.stdout)
2038 : );
2039 0 : warn!(
2040 0 : "debugger stderr: {}",
2041 0 : String::from_utf8_lossy(&backtrace.stderr)
2042 : );
2043 : }
2044 :
2045 0 : Ok(())
2046 0 : }
2047 :
2048 : /// Select `pg_stat_statements` data and return it as a stringified JSON
2049 0 : pub async fn collect_insights(&self) -> String {
2050 0 : let mut result_rows: Vec<String> = Vec::new();
2051 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
2052 0 : let connect_result = conf.connect(NoTls).await;
2053 0 : let (client, connection) = connect_result.unwrap();
2054 0 : tokio::spawn(async move {
2055 0 : if let Err(e) = connection.await {
2056 0 : eprintln!("connection error: {e}");
2057 0 : }
2058 0 : });
2059 0 : let result = client
2060 0 : .simple_query(
2061 0 : "SELECT
2062 0 : row_to_json(pg_stat_statements)
2063 0 : FROM
2064 0 : pg_stat_statements
2065 0 : WHERE
2066 0 : userid != 'cloud_admin'::regrole::oid
2067 0 : ORDER BY
2068 0 : (mean_exec_time + mean_plan_time) DESC
2069 0 : LIMIT 100",
2070 0 : )
2071 0 : .await;
2072 :
2073 0 : if let Ok(raw_rows) = result {
2074 0 : for message in raw_rows.iter() {
2075 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
2076 0 : if let Some(json) = row.get(0) {
2077 0 : result_rows.push(json.to_string());
2078 0 : }
2079 0 : }
2080 : }
2081 :
2082 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
2083 : } else {
2084 0 : "{{\"pg_stat_statements\": []}}".to_string()
2085 : }
2086 0 : }
2087 :
2088 : // download an archive, unzip and place files in correct locations
2089 0 : pub async fn download_extension(
2090 0 : &self,
2091 0 : real_ext_name: String,
2092 0 : ext_path: RemotePath,
2093 0 : ) -> Result<u64, DownloadError> {
2094 0 : let remote_ext_base_url =
2095 0 : self.params
2096 0 : .remote_ext_base_url
2097 0 : .as_ref()
2098 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
2099 0 : "Remote extensions storage is not configured",
2100 0 : )))?;
2101 :
2102 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
2103 :
2104 0 : let mut first_try = false;
2105 0 : if !self
2106 0 : .ext_download_progress
2107 0 : .read()
2108 0 : .expect("lock err")
2109 0 : .contains_key(ext_archive_name)
2110 0 : {
2111 0 : self.ext_download_progress
2112 0 : .write()
2113 0 : .expect("lock err")
2114 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
2115 0 : first_try = true;
2116 0 : }
2117 0 : let (download_start, download_completed) =
2118 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
2119 0 : let start_time_delta = Utc::now()
2120 0 : .signed_duration_since(download_start)
2121 0 : .to_std()
2122 0 : .unwrap()
2123 0 : .as_millis() as u64;
2124 :
2125 : // how long to wait for extension download if it was started by another process
2126 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
2127 :
2128 0 : if download_completed {
2129 0 : info!("extension already downloaded, skipping re-download");
2130 0 : return Ok(0);
2131 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
2132 0 : info!(
2133 0 : "download {ext_archive_name} already started by another process, hanging untill completion or timeout"
2134 : );
2135 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
2136 : loop {
2137 0 : info!("waiting for download");
2138 0 : interval.tick().await;
2139 0 : let (_, download_completed_now) =
2140 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
2141 0 : if download_completed_now {
2142 0 : info!("download finished by whoever else downloaded it");
2143 0 : return Ok(0);
2144 0 : }
2145 : }
2146 : // NOTE: the above loop will get terminated
2147 : // based on the timeout of the download function
2148 0 : }
2149 :
2150 : // if extension hasn't been downloaded before or the previous
2151 : // attempt to download was at least HANG_TIMEOUT ms ago
2152 : // then we try to download it here
2153 0 : info!("downloading new extension {ext_archive_name}");
2154 :
2155 0 : let download_size = extension_server::download_extension(
2156 0 : &real_ext_name,
2157 0 : &ext_path,
2158 0 : remote_ext_base_url,
2159 0 : &self.params.pgbin,
2160 0 : )
2161 0 : .await
2162 0 : .map_err(DownloadError::Other);
2163 :
2164 0 : if download_size.is_ok() {
2165 0 : self.ext_download_progress
2166 0 : .write()
2167 0 : .expect("bad lock")
2168 0 : .insert(ext_archive_name.to_string(), (download_start, true));
2169 0 : }
2170 :
2171 0 : download_size
2172 0 : }
2173 :
2174 0 : pub async fn set_role_grants(
2175 0 : &self,
2176 0 : db_name: &PgIdent,
2177 0 : schema_name: &PgIdent,
2178 0 : privileges: &[Privilege],
2179 0 : role_name: &PgIdent,
2180 0 : ) -> Result<()> {
2181 : use tokio_postgres::NoTls;
2182 :
2183 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
2184 0 : conf.dbname(db_name);
2185 :
2186 0 : let (db_client, conn) = conf
2187 0 : .connect(NoTls)
2188 0 : .await
2189 0 : .context("Failed to connect to the database")?;
2190 0 : tokio::spawn(conn);
2191 :
2192 : // TODO: support other types of grants apart from schemas?
2193 :
2194 : // check the role grants first - to gracefully handle read-replicas.
2195 0 : let select = "SELECT privilege_type
2196 0 : FROM pg_namespace
2197 0 : JOIN LATERAL (SELECT * FROM aclexplode(nspacl) AS x) acl ON true
2198 0 : JOIN pg_user users ON acl.grantee = users.usesysid
2199 0 : WHERE users.usename = $1
2200 0 : AND nspname = $2";
2201 0 : let rows = db_client
2202 0 : .query(select, &[role_name, schema_name])
2203 0 : .await
2204 0 : .with_context(|| format!("Failed to execute query: {select}"))?;
2205 :
2206 0 : let already_granted: HashSet<String> = rows.into_iter().map(|row| row.get(0)).collect();
2207 :
2208 0 : let grants = privileges
2209 0 : .iter()
2210 0 : .filter(|p| !already_granted.contains(p.as_str()))
2211 : // should not be quoted as it's part of the command.
2212 : // is already sanitized so it's ok
2213 0 : .map(|p| p.as_str())
2214 0 : .join(", ");
2215 :
2216 0 : if !grants.is_empty() {
2217 : // quote the schema and role name as identifiers to sanitize them.
2218 0 : let schema_name = schema_name.pg_quote();
2219 0 : let role_name = role_name.pg_quote();
2220 :
2221 0 : let query = format!("GRANT {grants} ON SCHEMA {schema_name} TO {role_name}",);
2222 0 : db_client
2223 0 : .simple_query(&query)
2224 0 : .await
2225 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2226 0 : }
2227 :
2228 0 : Ok(())
2229 0 : }
2230 :
2231 0 : pub async fn install_extension(
2232 0 : &self,
2233 0 : ext_name: &PgIdent,
2234 0 : db_name: &PgIdent,
2235 0 : ext_version: ExtVersion,
2236 0 : ) -> Result<ExtVersion> {
2237 : use tokio_postgres::NoTls;
2238 :
2239 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
2240 0 : conf.dbname(db_name);
2241 :
2242 0 : let (db_client, conn) = conf
2243 0 : .connect(NoTls)
2244 0 : .await
2245 0 : .context("Failed to connect to the database")?;
2246 0 : tokio::spawn(conn);
2247 :
2248 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
2249 0 : let version: Option<ExtVersion> = db_client
2250 0 : .query_opt(version_query, &[&ext_name])
2251 0 : .await
2252 0 : .with_context(|| format!("Failed to execute query: {version_query}"))?
2253 0 : .map(|row| row.get(0));
2254 :
2255 : // sanitize the inputs as postgres idents.
2256 0 : let ext_name: String = ext_name.pg_quote();
2257 0 : let quoted_version: String = ext_version.pg_quote();
2258 :
2259 0 : if let Some(installed_version) = version {
2260 0 : if installed_version == ext_version {
2261 0 : return Ok(installed_version);
2262 0 : }
2263 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
2264 0 : db_client
2265 0 : .simple_query(&query)
2266 0 : .await
2267 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2268 : } else {
2269 0 : let query =
2270 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
2271 0 : db_client
2272 0 : .simple_query(&query)
2273 0 : .await
2274 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2275 : }
2276 :
2277 0 : Ok(ext_version)
2278 0 : }
2279 :
2280 0 : pub async fn prepare_preload_libraries(
2281 0 : &self,
2282 0 : spec: &ComputeSpec,
2283 0 : ) -> Result<RemoteExtensionMetrics> {
2284 0 : if self.params.remote_ext_base_url.is_none() {
2285 0 : return Ok(RemoteExtensionMetrics {
2286 0 : num_ext_downloaded: 0,
2287 0 : largest_ext_size: 0,
2288 0 : total_ext_download_size: 0,
2289 0 : });
2290 0 : }
2291 0 : let remote_extensions = spec
2292 0 : .remote_extensions
2293 0 : .as_ref()
2294 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
2295 :
2296 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
2297 0 : let mut libs_vec = Vec::new();
2298 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
2299 0 : libs_vec = libs
2300 0 : .split(&[',', '\'', ' '])
2301 0 : .filter(|s| *s != "neon" && !s.is_empty())
2302 0 : .map(str::to_string)
2303 0 : .collect();
2304 0 : }
2305 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
2306 :
2307 : // that is used in neon_local and python tests
2308 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
2309 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
2310 0 : let mut shared_preload_libraries_line = "";
2311 0 : for line in conf_lines {
2312 0 : if line.starts_with("shared_preload_libraries") {
2313 0 : shared_preload_libraries_line = line;
2314 0 : }
2315 : }
2316 0 : let mut preload_libs_vec = Vec::new();
2317 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
2318 0 : preload_libs_vec = libs
2319 0 : .split(&[',', '\'', ' '])
2320 0 : .filter(|s| *s != "neon" && !s.is_empty())
2321 0 : .map(str::to_string)
2322 0 : .collect();
2323 0 : }
2324 0 : libs_vec.extend(preload_libs_vec);
2325 0 : }
2326 :
2327 : // Don't try to download libraries that are not in the index.
2328 : // Assume that they are already present locally.
2329 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
2330 :
2331 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
2332 :
2333 0 : let mut download_tasks = Vec::new();
2334 0 : for library in &libs_vec {
2335 0 : let (ext_name, ext_path) =
2336 0 : remote_extensions.get_ext(library, true, &BUILD_TAG, &self.params.pgversion)?;
2337 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
2338 : }
2339 0 : let results = join_all(download_tasks).await;
2340 :
2341 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
2342 0 : num_ext_downloaded: 0,
2343 0 : largest_ext_size: 0,
2344 0 : total_ext_download_size: 0,
2345 0 : };
2346 0 : for result in results {
2347 0 : let download_size = match result {
2348 0 : Ok(res) => {
2349 0 : remote_ext_metrics.num_ext_downloaded += 1;
2350 0 : res
2351 : }
2352 0 : Err(err) => {
2353 : // if we failed to download an extension, we don't want to fail the whole
2354 : // process, but we do want to log the error
2355 0 : error!("Failed to download extension: {}", err);
2356 0 : 0
2357 : }
2358 : };
2359 :
2360 0 : remote_ext_metrics.largest_ext_size =
2361 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
2362 0 : remote_ext_metrics.total_ext_download_size += download_size;
2363 : }
2364 0 : Ok(remote_ext_metrics)
2365 0 : }
2366 :
2367 : /// Waits until current thread receives a state changed notification and
2368 : /// the pageserver connection strings has changed.
2369 : ///
2370 : /// The operation will time out after a specified duration.
2371 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
2372 0 : let state = self.state.lock().unwrap();
2373 0 : let old_pageserver_connstr = state
2374 0 : .pspec
2375 0 : .as_ref()
2376 0 : .expect("spec must be set")
2377 0 : .pageserver_connstr
2378 0 : .clone();
2379 0 : let mut unchanged = true;
2380 0 : let _ = self
2381 0 : .state_changed
2382 0 : .wait_timeout_while(state, duration, |s| {
2383 0 : let pageserver_connstr = &s
2384 0 : .pspec
2385 0 : .as_ref()
2386 0 : .expect("spec must be set")
2387 0 : .pageserver_connstr;
2388 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
2389 0 : unchanged
2390 0 : })
2391 0 : .unwrap();
2392 0 : if !unchanged {
2393 0 : info!("Pageserver config changed");
2394 0 : }
2395 0 : }
2396 :
2397 0 : pub fn spawn_extension_stats_task(&self) {
2398 0 : self.terminate_extension_stats_task();
2399 :
2400 0 : let conf = self.tokio_conn_conf.clone();
2401 0 : let atomic_interval = self.params.installed_extensions_collection_interval.clone();
2402 0 : let mut installed_extensions_collection_interval =
2403 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst);
2404 0 : info!(
2405 0 : "[NEON_EXT_SPAWN] Spawning background installed extensions worker with Timeout: {}",
2406 : installed_extensions_collection_interval
2407 : );
2408 0 : let handle = tokio::spawn(async move {
2409 : loop {
2410 0 : info!(
2411 0 : "[NEON_EXT_INT_SLEEP]: Interval: {}",
2412 : installed_extensions_collection_interval
2413 : );
2414 : // Sleep at the start of the loop to ensure that two collections don't happen at the same time.
2415 : // The first collection happens during compute startup.
2416 0 : tokio::time::sleep(tokio::time::Duration::from_secs(
2417 0 : installed_extensions_collection_interval,
2418 0 : ))
2419 0 : .await;
2420 0 : let _ = installed_extensions(conf.clone()).await;
2421 : // Acquire a read lock on the compute spec and then update the interval if necessary
2422 0 : installed_extensions_collection_interval = std::cmp::max(
2423 0 : installed_extensions_collection_interval,
2424 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst),
2425 0 : );
2426 : }
2427 : });
2428 :
2429 : // Store the new task handle
2430 0 : *self.extension_stats_task.lock().unwrap() = Some(handle);
2431 0 : }
2432 :
2433 0 : fn terminate_extension_stats_task(&self) {
2434 0 : if let Some(h) = self.extension_stats_task.lock().unwrap().take() {
2435 0 : h.abort()
2436 0 : }
2437 0 : }
2438 :
2439 0 : pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
2440 0 : self.terminate_lfc_offload_task();
2441 0 : let secs = interval.as_secs();
2442 0 : info!("spawning lfc offload worker with {secs}s interval");
2443 0 : let this = self.clone();
2444 0 : let handle = spawn(async move {
2445 0 : let mut interval = time::interval(interval);
2446 0 : interval.tick().await; // returns immediately
2447 : loop {
2448 0 : interval.tick().await;
2449 0 : this.offload_lfc_async().await;
2450 : }
2451 : });
2452 0 : *self.lfc_offload_task.lock().unwrap() = Some(handle);
2453 0 : }
2454 :
2455 0 : fn terminate_lfc_offload_task(&self) {
2456 0 : if let Some(h) = self.lfc_offload_task.lock().unwrap().take() {
2457 0 : h.abort()
2458 0 : }
2459 0 : }
2460 :
2461 0 : fn update_installed_extensions_collection_interval(&self, spec: &ComputeSpec) {
2462 : // Update the interval for collecting installed extensions statistics
2463 : // If the value is -1, we never suspend so set the value to default collection.
2464 : // If the value is 0, it means default, we will just continue to use the default.
2465 0 : if spec.suspend_timeout_seconds == -1 || spec.suspend_timeout_seconds == 0 {
2466 0 : self.params.installed_extensions_collection_interval.store(
2467 0 : DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL,
2468 0 : std::sync::atomic::Ordering::SeqCst,
2469 0 : );
2470 0 : } else {
2471 0 : self.params.installed_extensions_collection_interval.store(
2472 0 : spec.suspend_timeout_seconds as u64,
2473 0 : std::sync::atomic::Ordering::SeqCst,
2474 0 : );
2475 0 : }
2476 0 : }
2477 : }
2478 :
2479 0 : pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
2480 0 : let res = get_installed_extensions(conf).await;
2481 0 : match res {
2482 0 : Ok(extensions) => {
2483 0 : info!(
2484 0 : "[NEON_EXT_STAT] {}",
2485 0 : serde_json::to_string(&extensions).expect("failed to serialize extensions list")
2486 : );
2487 : }
2488 0 : Err(err) => error!("could not get installed extensions: {err}"),
2489 : }
2490 0 : Ok(())
2491 0 : }
2492 :
2493 0 : pub fn forward_termination_signal(dev_mode: bool) {
2494 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2495 0 : if ss_pid != 0 {
2496 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2497 0 : kill(ss_pid, Signal::SIGTERM).ok();
2498 0 : }
2499 :
2500 0 : if !dev_mode {
2501 : // Terminate pgbouncer with SIGKILL
2502 0 : match pid_file::read(PGBOUNCER_PIDFILE.into()) {
2503 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2504 0 : info!("sending SIGKILL to pgbouncer process pid: {}", pid);
2505 0 : if let Err(e) = kill(pid, Signal::SIGKILL) {
2506 0 : error!("failed to terminate pgbouncer: {}", e);
2507 0 : }
2508 : }
2509 : // pgbouncer does not lock the pid file, so we read and kill the process directly
2510 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2511 0 : if let Ok(pid_str) = std::fs::read_to_string(PGBOUNCER_PIDFILE) {
2512 0 : if let Ok(pid) = pid_str.trim().parse::<i32>() {
2513 0 : info!(
2514 0 : "sending SIGKILL to pgbouncer process pid: {} (from unlocked pid file)",
2515 : pid
2516 : );
2517 0 : if let Err(e) = kill(Pid::from_raw(pid), Signal::SIGKILL) {
2518 0 : error!("failed to terminate pgbouncer: {}", e);
2519 0 : }
2520 0 : }
2521 : } else {
2522 0 : info!("pgbouncer pid file exists but process not running");
2523 : }
2524 : }
2525 : Ok(pid_file::PidFileRead::NotExist) => {
2526 0 : info!("pgbouncer pid file not found, process may not be running");
2527 : }
2528 0 : Err(e) => {
2529 0 : error!("error reading pgbouncer pid file: {}", e);
2530 : }
2531 : }
2532 :
2533 : // Terminate local_proxy
2534 0 : match pid_file::read("/etc/local_proxy/pid".into()) {
2535 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2536 0 : info!("sending SIGTERM to local_proxy process pid: {}", pid);
2537 0 : if let Err(e) = kill(pid, Signal::SIGTERM) {
2538 0 : error!("failed to terminate local_proxy: {}", e);
2539 0 : }
2540 : }
2541 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2542 0 : info!("local_proxy PID file exists but process not running");
2543 : }
2544 : Ok(pid_file::PidFileRead::NotExist) => {
2545 0 : info!("local_proxy PID file not found, process may not be running");
2546 : }
2547 0 : Err(e) => {
2548 0 : error!("error reading local_proxy PID file: {}", e);
2549 : }
2550 : }
2551 : } else {
2552 0 : info!("Skipping pgbouncer and local_proxy termination because in dev mode");
2553 : }
2554 :
2555 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2556 0 : if pg_pid != 0 {
2557 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2558 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2559 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2560 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2561 0 : kill(pg_pid, Signal::SIGINT).ok();
2562 0 : }
2563 0 : }
2564 :
2565 : // helper trait to call JoinSet::spawn_blocking(f), but propagates the current
2566 : // tracing span to the thread.
2567 : trait JoinSetExt<T> {
2568 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2569 : where
2570 : F: FnOnce() -> T + Send + 'static,
2571 : T: Send;
2572 : }
2573 :
2574 : impl<T: 'static> JoinSetExt<T> for tokio::task::JoinSet<T> {
2575 0 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2576 0 : where
2577 0 : F: FnOnce() -> T + Send + 'static,
2578 0 : T: Send,
2579 : {
2580 0 : let sp = tracing::Span::current();
2581 0 : self.spawn_blocking(move || {
2582 0 : let _e = sp.enter();
2583 0 : f()
2584 0 : })
2585 0 : }
2586 : }
2587 :
2588 : #[cfg(test)]
2589 : mod tests {
2590 : use std::fs::File;
2591 :
2592 : use super::*;
2593 :
2594 : #[test]
2595 1 : fn duplicate_safekeeper_connstring() {
2596 1 : let file = File::open("tests/cluster_spec.json").unwrap();
2597 1 : let spec: ComputeSpec = serde_json::from_reader(file).unwrap();
2598 :
2599 1 : match ParsedSpec::try_from(spec.clone()) {
2600 0 : Ok(_p) => panic!("Failed to detect duplicate entry"),
2601 1 : Err(e) => assert!(e.starts_with("duplicate entry in safekeeper_connstrings:")),
2602 : };
2603 1 : }
2604 : }
|
