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::ffi::OsString;
25 : use std::os::unix::fs::{PermissionsExt, symlink};
26 : use std::path::Path;
27 : use std::process::{Command, Stdio};
28 : use std::str::FromStr;
29 : use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
30 : use std::sync::{Arc, Condvar, Mutex, RwLock};
31 : use std::time::{Duration, Instant};
32 : use std::{env, fs};
33 : use tokio::{spawn, sync::watch, task::JoinHandle, time};
34 : use tracing::{Instrument, debug, error, info, instrument, warn};
35 : use url::Url;
36 : use utils::id::{TenantId, TimelineId};
37 : use utils::lsn::Lsn;
38 : use utils::measured_stream::MeasuredReader;
39 : use utils::pid_file;
40 : use utils::shard::{ShardCount, ShardIndex, ShardNumber};
41 :
42 : use crate::configurator::launch_configurator;
43 : use crate::disk_quota::set_disk_quota;
44 : use crate::installed_extensions::get_installed_extensions;
45 : use crate::logger::startup_context_from_env;
46 : use crate::lsn_lease::launch_lsn_lease_bg_task_for_static;
47 : use crate::metrics::COMPUTE_CTL_UP;
48 : use crate::monitor::launch_monitor;
49 : use crate::pg_helpers::*;
50 : use crate::pgbouncer::*;
51 : use crate::rsyslog::{
52 : PostgresLogsRsyslogConfig, configure_audit_rsyslog, configure_postgres_logs_export,
53 : launch_pgaudit_gc,
54 : };
55 : use crate::spec::*;
56 : use crate::swap::resize_swap;
57 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
58 : use crate::tls::watch_cert_for_changes;
59 : use crate::{config, extension_server, local_proxy};
60 :
61 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
62 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
63 : // This is an arbitrary build tag. Fine as a default / for testing purposes
64 : // in-case of not-set environment var
65 : const BUILD_TAG_DEFAULT: &str = "latest";
66 : /// Build tag/version of the compute node binaries/image. It's tricky and ugly
67 : /// to pass it everywhere as a part of `ComputeNodeParams`, so we use a
68 : /// global static variable.
69 0 : pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
70 0 : option_env!("BUILD_TAG")
71 0 : .unwrap_or(BUILD_TAG_DEFAULT)
72 0 : .to_string()
73 0 : });
74 : const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
75 :
76 : /// Static configuration params that don't change after startup. These mostly
77 : /// come from the CLI args, or are derived from them.
78 : #[derive(Clone, Debug)]
79 : pub struct ComputeNodeParams {
80 : /// The ID of the compute
81 : pub compute_id: String,
82 :
83 : /// Url type maintains proper escaping
84 : pub connstr: url::Url,
85 :
86 : /// The name of the 'weak' superuser role, which we give to the users.
87 : /// It follows the allow list approach, i.e., we take a standard role
88 : /// and grant it extra permissions with explicit GRANTs here and there,
89 : /// and core patches.
90 : pub privileged_role_name: String,
91 :
92 : pub resize_swap_on_bind: bool,
93 : pub set_disk_quota_for_fs: Option<String>,
94 :
95 : // VM monitor parameters
96 : #[cfg(target_os = "linux")]
97 : pub filecache_connstr: String,
98 : #[cfg(target_os = "linux")]
99 : pub cgroup: String,
100 : #[cfg(target_os = "linux")]
101 : pub vm_monitor_addr: String,
102 :
103 : pub pgdata: String,
104 : pub pgbin: String,
105 : pub pgversion: String,
106 :
107 : /// The port that the compute's external HTTP server listens on
108 : pub external_http_port: u16,
109 : /// The port that the compute's internal HTTP server listens on
110 : pub internal_http_port: u16,
111 :
112 : /// the address of extension storage proxy gateway
113 : pub remote_ext_base_url: Option<Url>,
114 :
115 : /// Interval for installed extensions collection
116 : pub installed_extensions_collection_interval: Arc<AtomicU64>,
117 : /// Hadron instance ID of the compute node.
118 : pub instance_id: Option<String>,
119 : /// Timeout of PG compute startup in the Init state.
120 : pub pg_init_timeout: Option<Duration>,
121 : // Path to the `pg_isready` binary.
122 : pub pg_isready_bin: String,
123 : pub lakebase_mode: bool,
124 :
125 : pub build_tag: String,
126 : pub control_plane_uri: Option<String>,
127 : pub config_path_test_only: Option<OsString>,
128 : }
129 :
130 : type TaskHandle = Mutex<Option<JoinHandle<()>>>;
131 :
132 : /// Compute node info shared across several `compute_ctl` threads.
133 : pub struct ComputeNode {
134 : pub params: ComputeNodeParams,
135 :
136 : // We connect to Postgres from many different places, so build configs once
137 : // and reuse them where needed. These are derived from 'params.connstr'
138 : pub conn_conf: postgres::config::Config,
139 : pub tokio_conn_conf: tokio_postgres::config::Config,
140 :
141 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
142 : /// To allow HTTP API server to serving status requests, while configuration
143 : /// is in progress, lock should be held only for short periods of time to do
144 : /// read/write, not the whole configuration process.
145 : pub state: Mutex<ComputeState>,
146 : /// `Condvar` to allow notifying waiters about state changes.
147 : pub state_changed: Condvar,
148 :
149 : // key: ext_archive_name, value: started download time, download_completed?
150 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
151 : pub compute_ctl_config: ComputeCtlConfig,
152 :
153 : /// Handle to the extension stats collection task
154 : extension_stats_task: TaskHandle,
155 : lfc_offload_task: TaskHandle,
156 : }
157 :
158 : // store some metrics about download size that might impact startup time
159 : #[derive(Clone, Debug)]
160 : pub struct RemoteExtensionMetrics {
161 : num_ext_downloaded: u64,
162 : largest_ext_size: u64,
163 : total_ext_download_size: u64,
164 : }
165 :
166 : #[derive(Clone, Debug)]
167 : pub struct ComputeState {
168 : pub start_time: DateTime<Utc>,
169 : pub pg_start_time: Option<DateTime<Utc>>,
170 : pub status: ComputeStatus,
171 : /// Timestamp of the last Postgres activity. It could be `None` if
172 : /// compute wasn't used since start.
173 : pub last_active: Option<DateTime<Utc>>,
174 : pub error: Option<String>,
175 :
176 : /// Compute spec. This can be received from the CLI or - more likely -
177 : /// passed by the control plane with a /configure HTTP request.
178 : pub pspec: Option<ParsedSpec>,
179 :
180 : /// If the spec is passed by a /configure request, 'startup_span' is the
181 : /// /configure request's tracing span. The main thread enters it when it
182 : /// processes the compute startup, so that the compute startup is considered
183 : /// to be part of the /configure request for tracing purposes.
184 : ///
185 : /// If the request handling thread/task called startup_compute() directly,
186 : /// it would automatically be a child of the request handling span, and we
187 : /// wouldn't need this. But because we use the main thread to perform the
188 : /// startup, and the /configure task just waits for it to finish, we need to
189 : /// set up the span relationship ourselves.
190 : pub startup_span: Option<tracing::span::Span>,
191 :
192 : pub lfc_prewarm_state: LfcPrewarmState,
193 : pub lfc_offload_state: LfcOffloadState,
194 :
195 : /// WAL flush LSN that is set after terminating Postgres and syncing safekeepers if
196 : /// mode == ComputeMode::Primary. None otherwise
197 : pub terminate_flush_lsn: Option<Lsn>,
198 : pub promote_state: Option<watch::Receiver<PromoteState>>,
199 :
200 : pub metrics: ComputeMetrics,
201 : }
202 :
203 : impl ComputeState {
204 0 : pub fn new() -> Self {
205 0 : Self {
206 0 : start_time: Utc::now(),
207 0 : pg_start_time: None,
208 0 : status: ComputeStatus::Empty,
209 0 : last_active: None,
210 0 : error: None,
211 0 : pspec: None,
212 0 : startup_span: None,
213 0 : metrics: ComputeMetrics::default(),
214 0 : lfc_prewarm_state: LfcPrewarmState::default(),
215 0 : lfc_offload_state: LfcOffloadState::default(),
216 0 : terminate_flush_lsn: None,
217 0 : promote_state: None,
218 0 : }
219 0 : }
220 :
221 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
222 0 : let prev = self.status;
223 0 : info!("Changing compute status from {} to {}", prev, status);
224 0 : self.status = status;
225 0 : state_changed.notify_all();
226 :
227 0 : COMPUTE_CTL_UP.reset();
228 0 : COMPUTE_CTL_UP
229 0 : .with_label_values(&[&BUILD_TAG, status.to_string().as_str()])
230 0 : .set(1);
231 0 : }
232 :
233 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
234 0 : self.error = Some(format!("{err:?}"));
235 0 : self.set_status(ComputeStatus::Failed, state_changed);
236 0 : }
237 : }
238 :
239 : impl Default for ComputeState {
240 0 : fn default() -> Self {
241 0 : Self::new()
242 0 : }
243 : }
244 :
245 : #[derive(Clone, Debug)]
246 : pub struct ParsedSpec {
247 : pub spec: ComputeSpec,
248 : pub tenant_id: TenantId,
249 : pub timeline_id: TimelineId,
250 : pub pageserver_connstr: String,
251 : pub safekeeper_connstrings: Vec<String>,
252 : pub storage_auth_token: Option<String>,
253 : /// k8s dns name and port
254 : pub endpoint_storage_addr: Option<String>,
255 : pub endpoint_storage_token: Option<String>,
256 : }
257 :
258 : impl ParsedSpec {
259 1 : pub fn validate(&self) -> Result<(), String> {
260 : // Only Primary nodes are using safekeeper_connstrings, and at the moment
261 : // this method only validates that part of the specs.
262 1 : if self.spec.mode != ComputeMode::Primary {
263 0 : return Ok(());
264 1 : }
265 :
266 : // While it seems like a good idea to check for an odd number of entries in
267 : // the safekeepers connection string, changes to the list of safekeepers might
268 : // incur appending a new server to a list of 3, in which case a list of 4
269 : // entries is okay in production.
270 : //
271 : // Still we want unique entries, and at least one entry in the vector
272 1 : if self.safekeeper_connstrings.is_empty() {
273 0 : return Err(String::from("safekeeper_connstrings is empty"));
274 1 : }
275 :
276 : // check for uniqueness of the connection strings in the set
277 1 : let mut connstrings = self.safekeeper_connstrings.clone();
278 :
279 1 : connstrings.sort();
280 1 : let mut previous = &connstrings[0];
281 :
282 2 : for current in connstrings.iter().skip(1) {
283 : // duplicate entry?
284 2 : if current == previous {
285 1 : return Err(format!(
286 1 : "duplicate entry in safekeeper_connstrings: {current}!",
287 1 : ));
288 1 : }
289 :
290 1 : previous = current;
291 : }
292 :
293 0 : Ok(())
294 1 : }
295 : }
296 :
297 : impl TryFrom<ComputeSpec> for ParsedSpec {
298 : type Error = String;
299 1 : fn try_from(spec: ComputeSpec) -> Result<Self, String> {
300 : // Extract the options from the spec file that are needed to connect to
301 : // the storage system.
302 : //
303 : // For backwards-compatibility, the top-level fields in the spec file
304 : // may be empty. In that case, we need to dig them from the GUCs in the
305 : // cluster.settings field.
306 1 : let pageserver_connstr = spec
307 1 : .pageserver_connstring
308 1 : .clone()
309 1 : .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
310 1 : .ok_or("pageserver connstr should be provided")?;
311 1 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
312 1 : if matches!(spec.mode, ComputeMode::Primary) {
313 1 : spec.cluster
314 1 : .settings
315 1 : .find("neon.safekeepers")
316 1 : .ok_or("safekeeper connstrings should be provided")?
317 1 : .split(',')
318 4 : .map(|str| str.to_string())
319 1 : .collect()
320 : } else {
321 0 : vec![]
322 : }
323 : } else {
324 0 : spec.safekeeper_connstrings.clone()
325 : };
326 :
327 1 : let storage_auth_token = spec.storage_auth_token.clone();
328 1 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
329 0 : tenant_id
330 : } else {
331 1 : spec.cluster
332 1 : .settings
333 1 : .find("neon.tenant_id")
334 1 : .ok_or("tenant id should be provided")
335 1 : .map(|s| TenantId::from_str(&s))?
336 1 : .or(Err("invalid tenant id"))?
337 : };
338 1 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
339 0 : timeline_id
340 : } else {
341 1 : spec.cluster
342 1 : .settings
343 1 : .find("neon.timeline_id")
344 1 : .ok_or("timeline id should be provided")
345 1 : .map(|s| TimelineId::from_str(&s))?
346 1 : .or(Err("invalid timeline id"))?
347 : };
348 :
349 1 : let endpoint_storage_addr: Option<String> = spec
350 1 : .endpoint_storage_addr
351 1 : .clone()
352 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"));
353 1 : let endpoint_storage_token = spec
354 1 : .endpoint_storage_token
355 1 : .clone()
356 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_token"));
357 :
358 1 : let res = ParsedSpec {
359 1 : spec,
360 1 : pageserver_connstr,
361 1 : safekeeper_connstrings,
362 1 : storage_auth_token,
363 1 : tenant_id,
364 1 : timeline_id,
365 1 : endpoint_storage_addr,
366 1 : endpoint_storage_token,
367 1 : };
368 :
369 : // Now check validity of the parsed specification
370 1 : res.validate()?;
371 0 : Ok(res)
372 1 : }
373 : }
374 :
375 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
376 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
377 : ///
378 : /// This function should be used to start postgres, as it will start it in the
379 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
380 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
381 : /// creates it during the sysinit phase of its inittab.
382 0 : fn maybe_cgexec(cmd: &str) -> Command {
383 : // The cplane sets this env var for autoscaling computes.
384 : // use `var_os` so we don't have to worry about the variable being valid
385 : // unicode. Should never be an concern . . . but just in case
386 0 : if env::var_os("AUTOSCALING").is_some() {
387 0 : let mut command = Command::new("cgexec");
388 0 : command.args(["-g", "memory:neon-postgres"]);
389 0 : command.arg(cmd);
390 0 : command
391 : } else {
392 0 : Command::new(cmd)
393 : }
394 0 : }
395 :
396 : struct PostgresHandle {
397 : postgres: std::process::Child,
398 : log_collector: JoinHandle<Result<()>>,
399 : }
400 :
401 : impl PostgresHandle {
402 : /// Return PID of the postgres (postmaster) process
403 0 : fn pid(&self) -> Pid {
404 0 : Pid::from_raw(self.postgres.id() as i32)
405 0 : }
406 : }
407 :
408 : struct StartVmMonitorResult {
409 : #[cfg(target_os = "linux")]
410 : token: tokio_util::sync::CancellationToken,
411 : #[cfg(target_os = "linux")]
412 : vm_monitor: Option<JoinHandle<Result<()>>>,
413 : }
414 :
415 : impl ComputeNode {
416 0 : pub fn new(params: ComputeNodeParams, config: ComputeConfig) -> Result<Self> {
417 0 : let connstr = params.connstr.as_str();
418 0 : let mut conn_conf = postgres::config::Config::from_str(connstr)
419 0 : .context("cannot build postgres config from connstr")?;
420 0 : let mut tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr)
421 0 : .context("cannot build tokio postgres config from connstr")?;
422 :
423 : // Users can set some configuration parameters per database with
424 : // ALTER DATABASE ... SET ...
425 : //
426 : // There are at least these parameters:
427 : //
428 : // - role=some_other_role
429 : // - default_transaction_read_only=on
430 : // - statement_timeout=1, i.e., 1ms, which will cause most of the queries to fail
431 : // - search_path=non_public_schema, this should be actually safe because
432 : // we don't call any functions in user databases, but better to always reset
433 : // it to public.
434 : //
435 : // that can affect `compute_ctl` and prevent it from properly configuring the database schema.
436 : // Unset them via connection string options before connecting to the database.
437 : // N.B. keep it in sync with `ZENITH_OPTIONS` in `get_maintenance_client()`.
438 : 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";
439 0 : let options = match conn_conf.get_options() {
440 : // Allow the control plane to override any options set by the
441 : // compute
442 0 : Some(options) => format!("{EXTRA_OPTIONS} {options}"),
443 0 : None => EXTRA_OPTIONS.to_string(),
444 : };
445 0 : conn_conf.options(&options);
446 0 : tokio_conn_conf.options(&options);
447 :
448 0 : let mut new_state = ComputeState::new();
449 0 : if let Some(spec) = config.spec {
450 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
451 0 : new_state.pspec = Some(pspec);
452 0 : }
453 :
454 0 : Ok(ComputeNode {
455 0 : params,
456 0 : conn_conf,
457 0 : tokio_conn_conf,
458 0 : state: Mutex::new(new_state),
459 0 : state_changed: Condvar::new(),
460 0 : ext_download_progress: RwLock::new(HashMap::new()),
461 0 : compute_ctl_config: config.compute_ctl_config,
462 0 : extension_stats_task: Mutex::new(None),
463 0 : lfc_offload_task: Mutex::new(None),
464 0 : })
465 0 : }
466 :
467 : /// Top-level control flow of compute_ctl. Returns a process exit code we should
468 : /// exit with.
469 0 : pub fn run(self) -> Result<Option<i32>> {
470 0 : let this = Arc::new(self);
471 :
472 0 : let cli_spec = this.state.lock().unwrap().pspec.clone();
473 :
474 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
475 : // available for binding. Prewarming helps Postgres start quicker later,
476 : // because QEMU will already have its memory allocated from the host, and
477 : // the necessary binaries will already be cached.
478 0 : if cli_spec.is_none() {
479 0 : this.prewarm_postgres_vm_memory()?;
480 0 : }
481 :
482 : // Set the up metric with Empty status before starting the HTTP server.
483 : // That way on the first metric scrape, an external observer will see us
484 : // as 'up' and 'empty' (unless the compute was started with a spec or
485 : // already configured by control plane).
486 0 : COMPUTE_CTL_UP
487 0 : .with_label_values(&[&BUILD_TAG, ComputeStatus::Empty.to_string().as_str()])
488 0 : .set(1);
489 :
490 : // Launch the external HTTP server first, so that we can serve control plane
491 : // requests while configuration is still in progress.
492 0 : crate::http::server::Server::External {
493 0 : port: this.params.external_http_port,
494 0 : config: this.compute_ctl_config.clone(),
495 0 : compute_id: this.params.compute_id.clone(),
496 0 : instance_id: this.params.instance_id.clone(),
497 0 : }
498 0 : .launch(&this);
499 :
500 : // The internal HTTP server could be launched later, but there isn't much
501 : // sense in waiting.
502 0 : crate::http::server::Server::Internal {
503 0 : port: this.params.internal_http_port,
504 0 : }
505 0 : .launch(&this);
506 :
507 : // If we got a spec from the CLI already, use that. Otherwise wait for the
508 : // control plane to pass it to us with a /configure HTTP request
509 0 : let pspec = if let Some(cli_spec) = cli_spec {
510 0 : cli_spec
511 : } else {
512 0 : this.wait_spec()?
513 : };
514 :
515 0 : launch_lsn_lease_bg_task_for_static(&this);
516 :
517 : // We have a spec, start the compute
518 0 : let mut delay_exit = false;
519 0 : let mut vm_monitor = None;
520 0 : let mut pg_process: Option<PostgresHandle> = None;
521 :
522 0 : match this.start_compute(&mut pg_process) {
523 0 : Ok(()) => {
524 0 : // Success! Launch remaining services (just vm-monitor currently)
525 0 : vm_monitor =
526 0 : Some(this.start_vm_monitor(pspec.spec.disable_lfc_resizing.unwrap_or(false)));
527 0 : }
528 0 : Err(err) => {
529 : // Something went wrong with the startup. Log it and expose the error to
530 : // HTTP status requests.
531 0 : error!("could not start the compute node: {:#}", err);
532 0 : this.set_failed_status(err);
533 0 : delay_exit = true;
534 :
535 : // If the error happened after starting PostgreSQL, kill it
536 0 : if let Some(ref pg_process) = pg_process {
537 0 : kill(pg_process.pid(), Signal::SIGQUIT).ok();
538 0 : }
539 : }
540 : }
541 :
542 : // If startup was successful, or it failed in the late stages,
543 : // PostgreSQL is now running. Wait until it exits.
544 0 : let exit_code = if let Some(pg_handle) = pg_process {
545 0 : let exit_status = this.wait_postgres(pg_handle);
546 0 : info!("Postgres exited with code {}, shutting down", exit_status);
547 0 : exit_status.code()
548 : } else {
549 0 : None
550 : };
551 :
552 0 : this.terminate_extension_stats_task();
553 0 : this.terminate_lfc_offload_task();
554 :
555 : // Terminate the vm_monitor so it releases the file watcher on
556 : // /sys/fs/cgroup/neon-postgres.
557 : // Note: the vm-monitor only runs on linux because it requires cgroups.
558 0 : if let Some(vm_monitor) = vm_monitor {
559 : cfg_if::cfg_if! {
560 : if #[cfg(target_os = "linux")] {
561 : // Kills all threads spawned by the monitor
562 0 : vm_monitor.token.cancel();
563 0 : if let Some(handle) = vm_monitor.vm_monitor {
564 0 : // Kills the actual task running the monitor
565 0 : handle.abort();
566 0 : }
567 : } else {
568 : _ = vm_monitor; // appease unused lint on macOS
569 : }
570 : }
571 0 : }
572 :
573 : // Reap the postgres process
574 0 : delay_exit |= this.cleanup_after_postgres_exit()?;
575 :
576 : // /terminate returns LSN. If we don't sleep at all, connection will break and we
577 : // won't get result. If we sleep too much, tests will take significantly longer
578 : // and Github Action run will error out
579 0 : let sleep_duration = if delay_exit {
580 0 : Duration::from_secs(30)
581 : } else {
582 0 : Duration::from_millis(300)
583 : };
584 :
585 : // If launch failed, keep serving HTTP requests for a while, so the cloud
586 : // control plane can get the actual error.
587 0 : if delay_exit {
588 0 : info!("giving control plane 30s to collect the error before shutdown");
589 0 : }
590 0 : std::thread::sleep(sleep_duration);
591 0 : Ok(exit_code)
592 0 : }
593 :
594 0 : pub fn wait_spec(&self) -> Result<ParsedSpec> {
595 0 : info!("no compute spec provided, waiting");
596 0 : let mut state = self.state.lock().unwrap();
597 0 : while state.status != ComputeStatus::ConfigurationPending {
598 0 : state = self.state_changed.wait(state).unwrap();
599 0 : }
600 :
601 0 : info!("got spec, continue configuration");
602 0 : let spec = state.pspec.as_ref().unwrap().clone();
603 :
604 : // Record for how long we slept waiting for the spec.
605 0 : let now = Utc::now();
606 0 : state.metrics.wait_for_spec_ms = now
607 0 : .signed_duration_since(state.start_time)
608 0 : .to_std()
609 0 : .unwrap()
610 0 : .as_millis() as u64;
611 :
612 : // Reset start time, so that the total startup time that is calculated later will
613 : // not include the time that we waited for the spec.
614 0 : state.start_time = now;
615 :
616 0 : Ok(spec)
617 0 : }
618 :
619 : /// Start compute.
620 : ///
621 : /// Prerequisites:
622 : /// - the compute spec has been placed in self.state.pspec
623 : ///
624 : /// On success:
625 : /// - status is set to ComputeStatus::Running
626 : /// - self.running_postgres is set
627 : ///
628 : /// On error:
629 : /// - status is left in ComputeStatus::Init. The caller is responsible for setting it to Failed
630 : /// - if Postgres was started before the fatal error happened, self.running_postgres is
631 : /// set. The caller is responsible for killing it.
632 : ///
633 : /// Note that this is in the critical path of a compute cold start. Keep this fast.
634 : /// Try to do things concurrently, to hide the latencies.
635 0 : fn start_compute(self: &Arc<Self>, pg_handle: &mut Option<PostgresHandle>) -> Result<()> {
636 : let compute_state: ComputeState;
637 :
638 : let start_compute_span;
639 : let _this_entered;
640 : {
641 0 : let mut state_guard = self.state.lock().unwrap();
642 :
643 : // Create a tracing span for the startup operation.
644 : //
645 : // We could otherwise just annotate the function with #[instrument], but if
646 : // we're being configured from a /configure HTTP request, we want the
647 : // startup to be considered part of the /configure request.
648 : //
649 : // Similarly, if a trace ID was passed in env variables, attach it to the span.
650 0 : start_compute_span = {
651 : // Temporarily enter the parent span, so that the new span becomes its child.
652 0 : if let Some(p) = state_guard.startup_span.take() {
653 0 : let _parent_entered = p.entered();
654 0 : tracing::info_span!("start_compute")
655 0 : } else if let Some(otel_context) = startup_context_from_env() {
656 : use tracing_opentelemetry::OpenTelemetrySpanExt;
657 0 : let span = tracing::info_span!("start_compute");
658 0 : span.set_parent(otel_context);
659 0 : span
660 : } else {
661 0 : tracing::info_span!("start_compute")
662 : }
663 : };
664 0 : _this_entered = start_compute_span.enter();
665 :
666 : // Hadron: Record postgres start time (used to enforce pg_init_timeout).
667 0 : state_guard.pg_start_time.replace(Utc::now());
668 :
669 0 : state_guard.set_status(ComputeStatus::Init, &self.state_changed);
670 0 : compute_state = state_guard.clone()
671 : }
672 :
673 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
674 0 : info!(
675 0 : "starting compute for project {}, operation {}, tenant {}, timeline {}, project {}, branch {}, endpoint {}, features {:?}, spec.remote_extensions {:?}",
676 0 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
677 0 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
678 : pspec.tenant_id,
679 : pspec.timeline_id,
680 0 : pspec.spec.project_id.as_deref().unwrap_or("None"),
681 0 : pspec.spec.branch_id.as_deref().unwrap_or("None"),
682 0 : pspec.spec.endpoint_id.as_deref().unwrap_or("None"),
683 : pspec.spec.features,
684 : pspec.spec.remote_extensions,
685 : );
686 :
687 : ////// PRE-STARTUP PHASE: things that need to be finished before we start the Postgres process
688 :
689 : // Collect all the tasks that must finish here
690 0 : let mut pre_tasks = tokio::task::JoinSet::new();
691 :
692 : // Make sure TLS certificates are properly loaded and in the right place.
693 0 : if self.compute_ctl_config.tls.is_some() {
694 0 : let this = self.clone();
695 0 : pre_tasks.spawn(async move {
696 0 : this.watch_cert_for_changes().await;
697 :
698 0 : Ok::<(), anyhow::Error>(())
699 0 : });
700 0 : }
701 :
702 0 : let tls_config = self.tls_config(&pspec.spec);
703 :
704 : // If there are any remote extensions in shared_preload_libraries, start downloading them
705 0 : if pspec.spec.remote_extensions.is_some() {
706 0 : let (this, spec) = (self.clone(), pspec.spec.clone());
707 0 : pre_tasks.spawn(async move {
708 0 : this.download_preload_extensions(&spec)
709 0 : .in_current_span()
710 0 : .await
711 0 : });
712 0 : }
713 :
714 : // Prepare pgdata directory. This downloads the basebackup, among other things.
715 : {
716 0 : let (this, cs) = (self.clone(), compute_state.clone());
717 0 : pre_tasks.spawn_blocking_child(move || this.prepare_pgdata(&cs));
718 : }
719 :
720 : // Resize swap to the desired size if the compute spec says so
721 0 : if let (Some(size_bytes), true) =
722 0 : (pspec.spec.swap_size_bytes, self.params.resize_swap_on_bind)
723 : {
724 0 : pre_tasks.spawn_blocking_child(move || {
725 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
726 : // *before* starting postgres.
727 : //
728 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
729 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
730 : // OOM-killed during startup because swap wasn't available yet.
731 0 : resize_swap(size_bytes).context("failed to resize swap")?;
732 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
733 0 : info!(%size_bytes, %size_mib, "resized swap");
734 :
735 0 : Ok::<(), anyhow::Error>(())
736 0 : });
737 0 : }
738 :
739 : // Set disk quota if the compute spec says so
740 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) = (
741 0 : pspec.spec.disk_quota_bytes,
742 0 : self.params.set_disk_quota_for_fs.as_ref(),
743 : ) {
744 0 : let disk_quota_fs_mountpoint = disk_quota_fs_mountpoint.clone();
745 0 : pre_tasks.spawn_blocking_child(move || {
746 0 : set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint)
747 0 : .context("failed to set disk quota")?;
748 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
749 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
750 :
751 0 : Ok::<(), anyhow::Error>(())
752 0 : });
753 0 : }
754 :
755 : // tune pgbouncer
756 0 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
757 0 : info!("tuning pgbouncer");
758 :
759 0 : let pgbouncer_settings = pgbouncer_settings.clone();
760 0 : let tls_config = tls_config.clone();
761 :
762 : // Spawn a background task to do the tuning,
763 : // so that we don't block the main thread that starts Postgres.
764 0 : let _handle = tokio::spawn(async move {
765 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
766 0 : if let Err(err) = res {
767 0 : error!("error while tuning pgbouncer: {err:?}");
768 : // Continue with the startup anyway
769 0 : }
770 0 : });
771 0 : }
772 :
773 : // configure local_proxy
774 0 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
775 0 : info!("configuring local_proxy");
776 :
777 : // Spawn a background task to do the configuration,
778 : // so that we don't block the main thread that starts Postgres.
779 :
780 0 : let mut local_proxy = local_proxy.clone();
781 0 : local_proxy.tls = tls_config.clone();
782 :
783 0 : let _handle = tokio::spawn(async move {
784 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
785 0 : error!("error while configuring local_proxy: {err:?}");
786 : // Continue with the startup anyway
787 0 : }
788 0 : });
789 0 : }
790 :
791 : // Configure and start rsyslog for compliance audit logging
792 0 : match pspec.spec.audit_log_level {
793 : ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
794 0 : let remote_tls_endpoint =
795 0 : std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
796 0 : let remote_plain_endpoint =
797 0 : std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
798 :
799 0 : if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
800 0 : anyhow::bail!(
801 0 : "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
802 : );
803 0 : }
804 :
805 0 : let log_directory_path = Path::new(&self.params.pgdata).join("log");
806 0 : let log_directory_path = log_directory_path.to_string_lossy().to_string();
807 :
808 : // Add project_id,endpoint_id to identify the logs.
809 : //
810 : // These ids are passed from cplane,
811 0 : let endpoint_id = pspec.spec.endpoint_id.as_deref().unwrap_or("");
812 0 : let project_id = pspec.spec.project_id.as_deref().unwrap_or("");
813 :
814 0 : configure_audit_rsyslog(
815 0 : log_directory_path.clone(),
816 0 : endpoint_id,
817 0 : project_id,
818 0 : &remote_plain_endpoint,
819 0 : &remote_tls_endpoint,
820 0 : )?;
821 :
822 : // Launch a background task to clean up the audit logs
823 0 : launch_pgaudit_gc(log_directory_path);
824 : }
825 0 : _ => {}
826 : }
827 :
828 : // Configure and start rsyslog for Postgres logs export
829 0 : let conf = PostgresLogsRsyslogConfig::new(pspec.spec.logs_export_host.as_deref());
830 0 : configure_postgres_logs_export(conf)?;
831 :
832 : // Launch remaining service threads
833 0 : let _monitor_handle = launch_monitor(self);
834 0 : let _configurator_handle = launch_configurator(self);
835 :
836 : // Wait for all the pre-tasks to finish before starting postgres
837 0 : let rt = tokio::runtime::Handle::current();
838 0 : while let Some(res) = rt.block_on(pre_tasks.join_next()) {
839 0 : res??;
840 : }
841 :
842 : ////// START POSTGRES
843 0 : let start_time = Utc::now();
844 0 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
845 0 : let postmaster_pid = pg_process.pid();
846 0 : *pg_handle = Some(pg_process);
847 :
848 : // If this is a primary endpoint, perform some post-startup configuration before
849 : // opening it up for the world.
850 0 : let config_time = Utc::now();
851 0 : if pspec.spec.mode == ComputeMode::Primary {
852 0 : self.configure_as_primary(&compute_state)?;
853 :
854 0 : let conf = self.get_tokio_conn_conf(None);
855 0 : tokio::task::spawn(async {
856 0 : let _ = installed_extensions(conf).await;
857 0 : });
858 0 : }
859 :
860 : // All done!
861 0 : let startup_end_time = Utc::now();
862 0 : let metrics = {
863 0 : let mut state = self.state.lock().unwrap();
864 0 : state.metrics.start_postgres_ms = config_time
865 0 : .signed_duration_since(start_time)
866 0 : .to_std()
867 0 : .unwrap()
868 0 : .as_millis() as u64;
869 0 : state.metrics.config_ms = startup_end_time
870 0 : .signed_duration_since(config_time)
871 0 : .to_std()
872 0 : .unwrap()
873 0 : .as_millis() as u64;
874 0 : state.metrics.total_startup_ms = startup_end_time
875 0 : .signed_duration_since(compute_state.start_time)
876 0 : .to_std()
877 0 : .unwrap()
878 0 : .as_millis() as u64;
879 0 : state.metrics.clone()
880 : };
881 0 : self.set_status(ComputeStatus::Running);
882 :
883 : // Log metrics so that we can search for slow operations in logs
884 0 : info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
885 :
886 0 : self.spawn_extension_stats_task();
887 :
888 0 : if pspec.spec.autoprewarm {
889 0 : info!("autoprewarming on startup as requested");
890 0 : self.prewarm_lfc(None);
891 0 : }
892 0 : if let Some(seconds) = pspec.spec.offload_lfc_interval_seconds {
893 0 : self.spawn_lfc_offload_task(Duration::from_secs(seconds.into()));
894 0 : };
895 0 : Ok(())
896 0 : }
897 :
898 : #[instrument(skip_all)]
899 : async fn download_preload_extensions(&self, spec: &ComputeSpec) -> Result<()> {
900 : let remote_extensions = if let Some(remote_extensions) = &spec.remote_extensions {
901 : remote_extensions
902 : } else {
903 : return Ok(());
904 : };
905 :
906 : // First, create control files for all available extensions
907 : extension_server::create_control_files(remote_extensions, &self.params.pgbin);
908 :
909 : let library_load_start_time = Utc::now();
910 : let remote_ext_metrics = self.prepare_preload_libraries(spec).await?;
911 :
912 : let library_load_time = Utc::now()
913 : .signed_duration_since(library_load_start_time)
914 : .to_std()
915 : .unwrap()
916 : .as_millis() as u64;
917 : let mut state = self.state.lock().unwrap();
918 : state.metrics.load_ext_ms = library_load_time;
919 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
920 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
921 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
922 : info!(
923 : "Loading shared_preload_libraries took {:?}ms",
924 : library_load_time
925 : );
926 : info!("{:?}", remote_ext_metrics);
927 :
928 : Ok(())
929 : }
930 :
931 : /// Start the vm-monitor if directed to. The vm-monitor only runs on linux
932 : /// because it requires cgroups.
933 0 : fn start_vm_monitor(&self, disable_lfc_resizing: bool) -> StartVmMonitorResult {
934 : cfg_if::cfg_if! {
935 : if #[cfg(target_os = "linux")] {
936 : use std::env;
937 : use tokio_util::sync::CancellationToken;
938 :
939 : // This token is used internally by the monitor to clean up all threads
940 0 : let token = CancellationToken::new();
941 :
942 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
943 0 : let pgconnstr = if disable_lfc_resizing {
944 0 : None
945 : } else {
946 0 : Some(self.params.filecache_connstr.clone())
947 : };
948 :
949 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
950 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
951 0 : Box::leak(Box::new(vm_monitor::Args {
952 0 : cgroup: Some(self.params.cgroup.clone()),
953 0 : pgconnstr,
954 0 : addr: self.params.vm_monitor_addr.clone(),
955 0 : })),
956 0 : token.clone(),
957 : ));
958 0 : Some(vm_monitor)
959 : } else {
960 0 : None
961 : };
962 0 : StartVmMonitorResult { token, vm_monitor }
963 : } else {
964 : _ = disable_lfc_resizing; // appease unused lint on macOS
965 : StartVmMonitorResult { }
966 : }
967 : }
968 0 : }
969 :
970 0 : fn cleanup_after_postgres_exit(&self) -> Result<bool> {
971 : // Maybe sync safekeepers again, to speed up next startup
972 0 : let compute_state = self.state.lock().unwrap().clone();
973 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
974 0 : let lsn = if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
975 0 : info!("syncing safekeepers on shutdown");
976 0 : let storage_auth_token = pspec.storage_auth_token.clone();
977 0 : let lsn = self.sync_safekeepers(storage_auth_token)?;
978 0 : info!(%lsn, "synced safekeepers");
979 0 : Some(lsn)
980 : } else {
981 0 : info!("not primary, not syncing safekeepers");
982 0 : None
983 : };
984 :
985 0 : let mut state = self.state.lock().unwrap();
986 0 : state.terminate_flush_lsn = lsn;
987 :
988 0 : let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
989 0 : if state.status == ComputeStatus::TerminationPendingFast
990 0 : || state.status == ComputeStatus::TerminationPendingImmediate
991 : {
992 0 : info!(
993 0 : "Changing compute status from {} to {}",
994 0 : state.status,
995 : ComputeStatus::Terminated
996 : );
997 0 : state.status = ComputeStatus::Terminated;
998 0 : self.state_changed.notify_all();
999 0 : }
1000 0 : drop(state);
1001 :
1002 0 : if let Err(err) = self.check_for_core_dumps() {
1003 0 : error!("error while checking for core dumps: {err:?}");
1004 0 : }
1005 :
1006 0 : Ok(delay_exit)
1007 0 : }
1008 :
1009 : /// Check that compute node has corresponding feature enabled.
1010 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
1011 0 : let state = self.state.lock().unwrap();
1012 :
1013 0 : if let Some(s) = state.pspec.as_ref() {
1014 0 : s.spec.features.contains(&feature)
1015 : } else {
1016 0 : false
1017 : }
1018 0 : }
1019 :
1020 0 : pub fn set_status(&self, status: ComputeStatus) {
1021 0 : let mut state = self.state.lock().unwrap();
1022 0 : state.set_status(status, &self.state_changed);
1023 0 : }
1024 :
1025 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
1026 0 : let mut state = self.state.lock().unwrap();
1027 0 : state.set_failed_status(err, &self.state_changed);
1028 0 : }
1029 :
1030 0 : pub fn get_status(&self) -> ComputeStatus {
1031 0 : self.state.lock().unwrap().status
1032 0 : }
1033 :
1034 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
1035 0 : self.state
1036 0 : .lock()
1037 0 : .unwrap()
1038 0 : .pspec
1039 0 : .as_ref()
1040 0 : .map(|s| s.timeline_id)
1041 0 : }
1042 :
1043 : // Remove `pgdata` directory and create it again with right permissions.
1044 0 : fn create_pgdata(&self) -> Result<()> {
1045 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
1046 : // If it is something different then create_dir() will error out anyway.
1047 0 : let pgdata = &self.params.pgdata;
1048 0 : let _ok = fs::remove_dir_all(pgdata);
1049 0 : fs::create_dir(pgdata)?;
1050 0 : fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
1051 :
1052 0 : Ok(())
1053 0 : }
1054 :
1055 : /// Fetches a basebackup from the Pageserver using the compute state's Pageserver connstring and
1056 : /// unarchives it to `pgdata` directory, replacing any existing contents.
1057 : #[instrument(skip_all, fields(%lsn))]
1058 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1059 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
1060 :
1061 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
1062 : let started = Instant::now();
1063 :
1064 : let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
1065 : PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
1066 : PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
1067 : };
1068 :
1069 : self.fix_zenith_signal_neon_signal()?;
1070 :
1071 : let mut state = self.state.lock().unwrap();
1072 : state.metrics.pageserver_connect_micros =
1073 : connected.duration_since(started).as_micros() as u64;
1074 : state.metrics.basebackup_bytes = size as u64;
1075 : state.metrics.basebackup_ms = started.elapsed().as_millis() as u64;
1076 :
1077 : Ok(())
1078 : }
1079 :
1080 : /// Move the Zenith signal file to Neon signal file location.
1081 : /// This makes Compute compatible with older PageServers that don't yet
1082 : /// know about the Zenith->Neon rename.
1083 0 : fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
1084 0 : let datadir = Path::new(&self.params.pgdata);
1085 :
1086 0 : let neonsig = datadir.join("neon.signal");
1087 :
1088 0 : if neonsig.is_file() {
1089 0 : return Ok(());
1090 0 : }
1091 :
1092 0 : let zenithsig = datadir.join("zenith.signal");
1093 :
1094 0 : if zenithsig.is_file() {
1095 0 : fs::copy(zenithsig, neonsig)?;
1096 0 : }
1097 :
1098 0 : Ok(())
1099 0 : }
1100 :
1101 : /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
1102 : /// the connection was established, and the (compressed) size of the basebackup.
1103 0 : fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1104 0 : let shard0_connstr = spec
1105 0 : .pageserver_connstr
1106 0 : .split(',')
1107 0 : .next()
1108 0 : .unwrap()
1109 0 : .to_string();
1110 0 : let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
1111 0 : 0 | 1 => ShardIndex::unsharded(),
1112 0 : count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
1113 : };
1114 :
1115 0 : let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
1116 0 : let mut client = page_api::Client::connect(
1117 0 : shard0_connstr,
1118 0 : spec.tenant_id,
1119 0 : spec.timeline_id,
1120 0 : shard_index,
1121 0 : spec.storage_auth_token.clone(),
1122 0 : None, // NB: base backups use payload compression
1123 0 : )
1124 0 : .await?;
1125 0 : let connected = Instant::now();
1126 0 : let reader = client
1127 0 : .get_base_backup(page_api::GetBaseBackupRequest {
1128 0 : lsn: (lsn != Lsn(0)).then_some(lsn),
1129 0 : compression: BaseBackupCompression::Gzip,
1130 0 : replica: spec.spec.mode != ComputeMode::Primary,
1131 0 : full: false,
1132 0 : })
1133 0 : .await?;
1134 0 : anyhow::Ok((reader, connected))
1135 0 : })?;
1136 :
1137 0 : let mut reader = MeasuredReader::new(tokio_util::io::SyncIoBridge::new(reader));
1138 :
1139 : // Set `ignore_zeros` so that unpack() reads the entire stream and doesn't just stop at the
1140 : // end-of-archive marker. If the server errors, the tar::Builder drop handler will write an
1141 : // end-of-archive marker before the error is emitted, and we would not see the error.
1142 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut reader));
1143 0 : ar.set_ignore_zeros(true);
1144 0 : ar.unpack(&self.params.pgdata)?;
1145 :
1146 0 : Ok((connected, reader.get_byte_count()))
1147 0 : }
1148 :
1149 : /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
1150 : /// when the connection was established, and the (compressed) size of the basebackup.
1151 0 : fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1152 0 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
1153 0 : let mut config = postgres::Config::from_str(shard0_connstr)?;
1154 :
1155 : // Use the storage auth token from the config file, if given.
1156 : // Note: this overrides any password set in the connection string.
1157 0 : if let Some(storage_auth_token) = &spec.storage_auth_token {
1158 0 : info!("Got storage auth token from spec file");
1159 0 : config.password(storage_auth_token);
1160 : } else {
1161 0 : info!("Storage auth token not set");
1162 : }
1163 :
1164 0 : config.application_name("compute_ctl");
1165 0 : config.options(&format!(
1166 0 : "-c neon.compute_mode={}",
1167 0 : spec.spec.mode.to_type_str()
1168 0 : ));
1169 :
1170 : // Connect to pageserver
1171 0 : let mut client = config.connect(NoTls)?;
1172 0 : let connected = Instant::now();
1173 :
1174 0 : let basebackup_cmd = match lsn {
1175 : Lsn(0) => {
1176 0 : if spec.spec.mode != ComputeMode::Primary {
1177 0 : format!(
1178 0 : "basebackup {} {} --gzip --replica",
1179 : spec.tenant_id, spec.timeline_id
1180 : )
1181 : } else {
1182 0 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
1183 : }
1184 : }
1185 : _ => {
1186 0 : if spec.spec.mode != ComputeMode::Primary {
1187 0 : format!(
1188 0 : "basebackup {} {} {} --gzip --replica",
1189 : spec.tenant_id, spec.timeline_id, lsn
1190 : )
1191 : } else {
1192 0 : format!(
1193 0 : "basebackup {} {} {} --gzip",
1194 : spec.tenant_id, spec.timeline_id, lsn
1195 : )
1196 : }
1197 : }
1198 : };
1199 :
1200 0 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
1201 0 : let mut measured_reader = MeasuredReader::new(copyreader);
1202 0 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
1203 :
1204 : // Read the archive directly from the `CopyOutReader`
1205 : //
1206 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
1207 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
1208 : // sends an Error after finishing the tarball, we will not notice it.
1209 : // The tar::Builder drop handler will write an end-of-archive marker
1210 : // before emitting the error, and we would not see it otherwise.
1211 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
1212 0 : ar.set_ignore_zeros(true);
1213 0 : ar.unpack(&self.params.pgdata)?;
1214 :
1215 0 : Ok((connected, measured_reader.get_byte_count()))
1216 0 : }
1217 :
1218 : // Gets the basebackup in a retry loop
1219 : #[instrument(skip_all, fields(%lsn))]
1220 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1221 : let mut retry_period_ms = 500.0;
1222 : let mut attempts = 0;
1223 : const DEFAULT_ATTEMPTS: u16 = 10;
1224 : #[cfg(feature = "testing")]
1225 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
1226 : u16::from_str(&v).unwrap()
1227 : } else {
1228 : DEFAULT_ATTEMPTS
1229 : };
1230 : #[cfg(not(feature = "testing"))]
1231 : let max_attempts = DEFAULT_ATTEMPTS;
1232 : loop {
1233 : let result = self.try_get_basebackup(compute_state, lsn);
1234 : match result {
1235 : Ok(_) => {
1236 : return result;
1237 : }
1238 : Err(ref e) if attempts < max_attempts => {
1239 : warn!(
1240 : "Failed to get basebackup: {} (attempt {}/{})",
1241 : e, attempts, max_attempts
1242 : );
1243 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
1244 : retry_period_ms *= 1.5;
1245 : }
1246 : Err(_) => {
1247 : return result;
1248 : }
1249 : }
1250 : attempts += 1;
1251 : }
1252 : }
1253 :
1254 0 : pub async fn check_safekeepers_synced_async(
1255 0 : &self,
1256 0 : compute_state: &ComputeState,
1257 0 : ) -> Result<Option<Lsn>> {
1258 : // Construct a connection config for each safekeeper
1259 0 : let pspec: ParsedSpec = compute_state
1260 0 : .pspec
1261 0 : .as_ref()
1262 0 : .expect("spec must be set")
1263 0 : .clone();
1264 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
1265 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
1266 : // Format connstr
1267 0 : let id = connstr.clone();
1268 0 : let connstr = format!("postgresql://no_user@{connstr}");
1269 0 : let options = format!(
1270 0 : "-c timeline_id={} tenant_id={}",
1271 : pspec.timeline_id, pspec.tenant_id
1272 : );
1273 :
1274 : // Construct client
1275 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
1276 0 : config.options(&options);
1277 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
1278 0 : config.password(storage_auth_token);
1279 0 : }
1280 :
1281 0 : (id, config)
1282 0 : });
1283 :
1284 : // Create task set to query all safekeepers
1285 0 : let mut tasks = FuturesUnordered::new();
1286 0 : let quorum = sk_configs.len() / 2 + 1;
1287 0 : for (id, config) in sk_configs {
1288 0 : let timeout = tokio::time::Duration::from_millis(100);
1289 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
1290 0 : tasks.push(tokio::spawn(task));
1291 0 : }
1292 :
1293 : // Get a quorum of responses or errors
1294 0 : let mut responses = Vec::new();
1295 0 : let mut join_errors = Vec::new();
1296 0 : let mut task_errors = Vec::new();
1297 0 : let mut timeout_errors = Vec::new();
1298 0 : while let Some(response) = tasks.next().await {
1299 0 : match response {
1300 0 : Ok(Ok(Ok(r))) => responses.push(r),
1301 0 : Ok(Ok(Err(e))) => task_errors.push(e),
1302 0 : Ok(Err(e)) => timeout_errors.push(e),
1303 0 : Err(e) => join_errors.push(e),
1304 : };
1305 0 : if responses.len() >= quorum {
1306 0 : break;
1307 0 : }
1308 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
1309 0 : break;
1310 0 : }
1311 : }
1312 :
1313 : // In case of error, log and fail the check, but don't crash.
1314 : // We're playing it safe because these errors could be transient
1315 : // and we don't yet retry.
1316 0 : if responses.len() < quorum {
1317 0 : error!(
1318 0 : "failed sync safekeepers check {:?} {:?} {:?}",
1319 : join_errors, task_errors, timeout_errors
1320 : );
1321 0 : return Ok(None);
1322 0 : }
1323 :
1324 0 : Ok(check_if_synced(responses))
1325 0 : }
1326 :
1327 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
1328 : // in one roundtrip. If not, we should do a full sync_safekeepers.
1329 : #[instrument(skip_all)]
1330 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
1331 : let start_time = Utc::now();
1332 :
1333 : let rt = tokio::runtime::Handle::current();
1334 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
1335 :
1336 : // Record runtime
1337 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
1338 : .signed_duration_since(start_time)
1339 : .to_std()
1340 : .unwrap()
1341 : .as_millis() as u64;
1342 : result
1343 : }
1344 :
1345 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
1346 : // and return the reported LSN back to the caller.
1347 : #[instrument(skip_all)]
1348 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
1349 : let start_time = Utc::now();
1350 :
1351 : let mut sync_handle = maybe_cgexec(&self.params.pgbin)
1352 : .args(["--sync-safekeepers"])
1353 : .env("PGDATA", &self.params.pgdata) // we cannot use -D in this mode
1354 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1355 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1356 : } else {
1357 : vec![]
1358 : })
1359 : .stdout(Stdio::piped())
1360 : .stderr(Stdio::piped())
1361 : .spawn()
1362 : .expect("postgres --sync-safekeepers failed to start");
1363 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
1364 :
1365 : // `postgres --sync-safekeepers` will print all log output to stderr and
1366 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
1367 : // will be collected in a child thread.
1368 : let stderr = sync_handle
1369 : .stderr
1370 : .take()
1371 : .expect("stderr should be captured");
1372 : let logs_handle = handle_postgres_logs(stderr);
1373 :
1374 : let sync_output = sync_handle
1375 : .wait_with_output()
1376 : .expect("postgres --sync-safekeepers failed");
1377 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
1378 :
1379 : // Process has exited, so we can join the logs thread.
1380 : let _ = tokio::runtime::Handle::current()
1381 : .block_on(logs_handle)
1382 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1383 :
1384 : if !sync_output.status.success() {
1385 : anyhow::bail!(
1386 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
1387 : sync_output.status,
1388 : String::from_utf8(sync_output.stdout)
1389 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
1390 : );
1391 : }
1392 :
1393 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
1394 : .signed_duration_since(start_time)
1395 : .to_std()
1396 : .unwrap()
1397 : .as_millis() as u64;
1398 :
1399 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
1400 :
1401 : Ok(lsn)
1402 : }
1403 :
1404 : /// Do all the preparations like PGDATA directory creation, configuration,
1405 : /// safekeepers sync, basebackup, etc.
1406 : #[instrument(skip_all)]
1407 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
1408 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1409 : let spec = &pspec.spec;
1410 : let pgdata_path = Path::new(&self.params.pgdata);
1411 :
1412 : let tls_config = self.tls_config(&pspec.spec);
1413 :
1414 : // Remove/create an empty pgdata directory and put configuration there.
1415 : self.create_pgdata()?;
1416 : config::write_postgres_conf(
1417 : pgdata_path,
1418 : &self.params,
1419 : &pspec.spec,
1420 : self.params.internal_http_port,
1421 : tls_config,
1422 : )?;
1423 :
1424 : // Syncing safekeepers is only safe with primary nodes: if a primary
1425 : // is already connected it will be kicked out, so a secondary (standby)
1426 : // cannot sync safekeepers.
1427 : let lsn = match spec.mode {
1428 : ComputeMode::Primary => {
1429 : info!("checking if safekeepers are synced");
1430 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
1431 : lsn
1432 : } else {
1433 : info!("starting safekeepers syncing");
1434 : self.sync_safekeepers(pspec.storage_auth_token.clone())
1435 : .with_context(|| "failed to sync safekeepers")?
1436 : };
1437 : info!("safekeepers synced at LSN {}", lsn);
1438 : lsn
1439 : }
1440 : ComputeMode::Static(lsn) => {
1441 : info!("Starting read-only node at static LSN {}", lsn);
1442 : lsn
1443 : }
1444 : ComputeMode::Replica => {
1445 : info!("Initializing standby from latest Pageserver LSN");
1446 : Lsn(0)
1447 : }
1448 : };
1449 :
1450 : info!(
1451 : "getting basebackup@{} from pageserver {}",
1452 : lsn, &pspec.pageserver_connstr
1453 : );
1454 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
1455 0 : format!(
1456 0 : "failed to get basebackup@{} from pageserver {}",
1457 0 : lsn, &pspec.pageserver_connstr
1458 : )
1459 0 : })?;
1460 :
1461 : // Update pg_hba.conf received with basebackup.
1462 : update_pg_hba(pgdata_path, None)?;
1463 :
1464 : // Place pg_dynshmem under /dev/shm. This allows us to use
1465 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
1466 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
1467 : //
1468 : // Why on earth don't we just stick to the 'posix' default, you might
1469 : // ask. It turns out that making large allocations with 'posix' doesn't
1470 : // work very well with autoscaling. The behavior we want is that:
1471 : //
1472 : // 1. You can make large DSM allocations, larger than the current RAM
1473 : // size of the VM, without errors
1474 : //
1475 : // 2. If the allocated memory is really used, the VM is scaled up
1476 : // automatically to accommodate that
1477 : //
1478 : // We try to make that possible by having swap in the VM. But with the
1479 : // default 'posix' DSM implementation, we fail step 1, even when there's
1480 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
1481 : // the shmem segment, which is really just a file in /dev/shm in Linux,
1482 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
1483 : // than available RAM.
1484 : //
1485 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
1486 : // the Postgres 'mmap' DSM implementation doesn't use
1487 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
1488 : // fill the file with zeros. It's weird that that differs between
1489 : // 'posix' and 'mmap', but we take advantage of it. When the file is
1490 : // filled slowly with write(2), the kernel allows it to grow larger, as
1491 : // long as there's swap available.
1492 : //
1493 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
1494 : // segment to be larger than currently available RAM. But because we
1495 : // don't want to store it on a real file, which the kernel would try to
1496 : // flush to disk, so symlink pg_dynshm to /dev/shm.
1497 : //
1498 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
1499 : // control plane control that option. If 'mmap' is not used, this
1500 : // symlink doesn't affect anything.
1501 : //
1502 : // See https://github.com/neondatabase/autoscaling/issues/800
1503 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
1504 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
1505 :
1506 : match spec.mode {
1507 : ComputeMode::Primary => {}
1508 : ComputeMode::Replica | ComputeMode::Static(..) => {
1509 : add_standby_signal(pgdata_path)?;
1510 : }
1511 : }
1512 :
1513 : Ok(())
1514 : }
1515 :
1516 : /// Start and stop a postgres process to warm up the VM for startup.
1517 0 : pub fn prewarm_postgres_vm_memory(&self) -> Result<()> {
1518 0 : info!("prewarming VM memory");
1519 :
1520 : // Create pgdata
1521 0 : let pgdata = &format!("{}.warmup", self.params.pgdata);
1522 0 : create_pgdata(pgdata)?;
1523 :
1524 : // Run initdb to completion
1525 0 : info!("running initdb");
1526 0 : let initdb_bin = Path::new(&self.params.pgbin)
1527 0 : .parent()
1528 0 : .unwrap()
1529 0 : .join("initdb");
1530 0 : Command::new(initdb_bin)
1531 0 : .args(["--pgdata", pgdata])
1532 0 : .output()
1533 0 : .expect("cannot start initdb process");
1534 :
1535 : // Write conf
1536 : use std::io::Write;
1537 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
1538 0 : let mut file = std::fs::File::create(conf_path)?;
1539 0 : writeln!(file, "shared_buffers=65536")?;
1540 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
1541 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
1542 :
1543 : // Start postgres
1544 0 : info!("starting postgres");
1545 0 : let mut pg = maybe_cgexec(&self.params.pgbin)
1546 0 : .args(["-D", pgdata])
1547 0 : .spawn()
1548 0 : .expect("cannot start postgres process");
1549 :
1550 : // Stop it when it's ready
1551 0 : info!("waiting for postgres");
1552 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
1553 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
1554 : // it to avoid orphaned processes prowling around while datadir is
1555 : // wiped.
1556 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
1557 0 : kill(pm_pid, Signal::SIGQUIT)?;
1558 0 : info!("sent SIGQUIT signal");
1559 0 : pg.wait()?;
1560 0 : info!("done prewarming vm memory");
1561 :
1562 : // clean up
1563 0 : let _ok = fs::remove_dir_all(pgdata);
1564 0 : Ok(())
1565 0 : }
1566 :
1567 : /// Start Postgres as a child process and wait for it to start accepting
1568 : /// connections.
1569 : ///
1570 : /// Returns a handle to the child process and a handle to the logs thread.
1571 : #[instrument(skip_all)]
1572 : pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
1573 : let pgdata_path = Path::new(&self.params.pgdata);
1574 :
1575 : // Run postgres as a child process.
1576 : let mut pg = maybe_cgexec(&self.params.pgbin)
1577 : .args(["-D", &self.params.pgdata])
1578 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1579 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1580 : } else {
1581 : vec![]
1582 : })
1583 : .stderr(Stdio::piped())
1584 : .spawn()
1585 : .expect("cannot start postgres process");
1586 : PG_PID.store(pg.id(), Ordering::SeqCst);
1587 :
1588 : // Start a task to collect logs from stderr.
1589 : let stderr = pg.stderr.take().expect("stderr should be captured");
1590 : let logs_handle = handle_postgres_logs(stderr);
1591 :
1592 : wait_for_postgres(&mut pg, pgdata_path)?;
1593 :
1594 : Ok(PostgresHandle {
1595 : postgres: pg,
1596 : log_collector: logs_handle,
1597 : })
1598 : }
1599 :
1600 : /// Wait for the child Postgres process forever. In this state Ctrl+C will
1601 : /// propagate to Postgres and it will be shut down as well.
1602 0 : fn wait_postgres(&self, mut pg_handle: PostgresHandle) -> std::process::ExitStatus {
1603 0 : info!(postmaster_pid = %pg_handle.postgres.id(), "Waiting for Postgres to exit");
1604 :
1605 0 : let ecode = pg_handle
1606 0 : .postgres
1607 0 : .wait()
1608 0 : .expect("failed to start waiting on Postgres process");
1609 0 : PG_PID.store(0, Ordering::SeqCst);
1610 :
1611 : // Process has exited. Wait for the log collecting task to finish.
1612 0 : let _ = tokio::runtime::Handle::current()
1613 0 : .block_on(pg_handle.log_collector)
1614 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1615 :
1616 0 : ecode
1617 0 : }
1618 :
1619 : /// Do post configuration of the already started Postgres. This function spawns a background task to
1620 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
1621 : /// version. In the future, it may upgrade all 3rd-party extensions.
1622 : #[instrument(skip_all)]
1623 : pub fn post_apply_config(&self) -> Result<()> {
1624 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
1625 0 : tokio::spawn(async move {
1626 0 : let res = async {
1627 0 : let (mut client, connection) = conf.connect(NoTls).await?;
1628 0 : tokio::spawn(async move {
1629 0 : if let Err(e) = connection.await {
1630 0 : eprintln!("connection error: {e}");
1631 0 : }
1632 0 : });
1633 :
1634 0 : handle_neon_extension_upgrade(&mut client)
1635 0 : .await
1636 0 : .context("handle_neon_extension_upgrade")?;
1637 0 : Ok::<_, anyhow::Error>(())
1638 0 : }
1639 0 : .await;
1640 0 : if let Err(err) = res {
1641 0 : error!("error while post_apply_config: {err:#}");
1642 0 : }
1643 0 : });
1644 : Ok(())
1645 : }
1646 :
1647 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
1648 0 : let mut conf = self.conn_conf.clone();
1649 0 : if let Some(application_name) = application_name {
1650 0 : conf.application_name(application_name);
1651 0 : }
1652 0 : conf
1653 0 : }
1654 :
1655 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
1656 0 : let mut conf = self.tokio_conn_conf.clone();
1657 0 : if let Some(application_name) = application_name {
1658 0 : conf.application_name(application_name);
1659 0 : }
1660 0 : conf
1661 0 : }
1662 :
1663 0 : pub async fn get_maintenance_client(
1664 0 : conf: &tokio_postgres::Config,
1665 0 : ) -> Result<tokio_postgres::Client> {
1666 0 : let mut conf = conf.clone();
1667 0 : conf.application_name("compute_ctl:apply_config");
1668 :
1669 0 : let (client, conn) = match conf.connect(NoTls).await {
1670 : // If connection fails, it may be the old node with `zenith_admin` superuser.
1671 : //
1672 : // In this case we need to connect with old `zenith_admin` name
1673 : // and create new user. We cannot simply rename connected user,
1674 : // but we can create a new one and grant it all privileges.
1675 0 : Err(e) => match e.code() {
1676 : Some(&SqlState::INVALID_PASSWORD)
1677 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
1678 : // Connect with `zenith_admin` if `cloud_admin` could not authenticate
1679 0 : info!(
1680 0 : "cannot connect to Postgres: {}, retrying with 'zenith_admin' username",
1681 : e
1682 : );
1683 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
1684 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
1685 0 : zenith_admin_conf.user("zenith_admin");
1686 :
1687 : // It doesn't matter what were the options before, here we just want
1688 : // to connect and create a new superuser role.
1689 : const ZENITH_OPTIONS: &str = "-c role=zenith_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0";
1690 0 : zenith_admin_conf.options(ZENITH_OPTIONS);
1691 :
1692 0 : let mut client =
1693 0 : zenith_admin_conf.connect(NoTls)
1694 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
1695 :
1696 : // Disable forwarding so that users don't get a cloud_admin role
1697 0 : let mut func = || {
1698 0 : client.simple_query("SET neon.forward_ddl = false")?;
1699 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
1700 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
1701 0 : Ok::<_, anyhow::Error>(())
1702 0 : };
1703 0 : func().context("apply_config setup cloud_admin")?;
1704 :
1705 0 : drop(client);
1706 :
1707 : // Reconnect with connstring with expected name
1708 0 : conf.connect(NoTls).await?
1709 : }
1710 0 : _ => return Err(e.into()),
1711 : },
1712 0 : Ok((client, conn)) => (client, conn),
1713 : };
1714 :
1715 0 : spawn(async move {
1716 0 : if let Err(e) = conn.await {
1717 0 : error!("maintenance client connection error: {}", e);
1718 0 : }
1719 0 : });
1720 :
1721 : // Disable DDL forwarding because control plane already knows about the roles/databases
1722 : // we're about to modify.
1723 0 : client
1724 0 : .simple_query("SET neon.forward_ddl = false")
1725 0 : .await
1726 0 : .context("apply_config SET neon.forward_ddl = false")?;
1727 :
1728 0 : Ok(client)
1729 0 : }
1730 :
1731 : /// Do initial configuration of the already started Postgres.
1732 : #[instrument(skip_all)]
1733 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1734 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1735 :
1736 : let conf = Arc::new(conf);
1737 : let spec = Arc::new(
1738 : compute_state
1739 : .pspec
1740 : .as_ref()
1741 : .expect("spec must be set")
1742 : .spec
1743 : .clone(),
1744 : );
1745 :
1746 : let mut tls_config = None::<TlsConfig>;
1747 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
1748 : tls_config = self.compute_ctl_config.tls.clone();
1749 : }
1750 :
1751 : self.update_installed_extensions_collection_interval(&spec);
1752 :
1753 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1754 :
1755 : // Merge-apply spec & changes to PostgreSQL state.
1756 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1757 :
1758 : if let Some(local_proxy) = &spec.clone().local_proxy_config {
1759 : let mut local_proxy = local_proxy.clone();
1760 : local_proxy.tls = tls_config.clone();
1761 :
1762 : info!("configuring local_proxy");
1763 : local_proxy::configure(&local_proxy).context("apply_config local_proxy")?;
1764 : }
1765 :
1766 : // Run migrations separately to not hold up cold starts
1767 : let lakebase_mode = self.params.lakebase_mode;
1768 : let params = self.params.clone();
1769 0 : tokio::spawn(async move {
1770 0 : let mut conf = conf.as_ref().clone();
1771 0 : conf.application_name("compute_ctl:migrations");
1772 :
1773 0 : match conf.connect(NoTls).await {
1774 0 : Ok((mut client, connection)) => {
1775 0 : tokio::spawn(async move {
1776 0 : if let Err(e) = connection.await {
1777 0 : eprintln!("connection error: {e}");
1778 0 : }
1779 0 : });
1780 0 : if let Err(e) = handle_migrations(params, &mut client, lakebase_mode).await {
1781 0 : error!("Failed to run migrations: {}", e);
1782 0 : }
1783 : }
1784 0 : Err(e) => {
1785 0 : error!(
1786 0 : "Failed to connect to the compute for running migrations: {}",
1787 : e
1788 : );
1789 : }
1790 : };
1791 0 : });
1792 :
1793 : Ok::<(), anyhow::Error>(())
1794 : }
1795 :
1796 : // Signal to the configurator to refresh the configuration by pulling a new spec from the HCC.
1797 : // Note that this merely triggers a notification on a condition variable the configurator thread
1798 : // waits on. The configurator thread (in configurator.rs) pulls the new spec from the HCC and
1799 : // applies it.
1800 0 : pub async fn signal_refresh_configuration(&self) -> Result<()> {
1801 0 : let states_allowing_configuration_refresh = [
1802 0 : ComputeStatus::Running,
1803 0 : ComputeStatus::Failed,
1804 0 : ComputeStatus::RefreshConfigurationPending,
1805 0 : ];
1806 :
1807 0 : let mut state = self.state.lock().expect("state lock poisoned");
1808 0 : if states_allowing_configuration_refresh.contains(&state.status) {
1809 0 : state.status = ComputeStatus::RefreshConfigurationPending;
1810 0 : self.state_changed.notify_all();
1811 0 : Ok(())
1812 0 : } else if state.status == ComputeStatus::Init {
1813 : // If the compute is in Init state, we can't refresh the configuration immediately,
1814 : // but we should be able to do that soon.
1815 0 : Ok(())
1816 : } else {
1817 0 : Err(anyhow::anyhow!(
1818 0 : "Cannot refresh compute configuration in state {:?}",
1819 0 : state.status
1820 0 : ))
1821 : }
1822 0 : }
1823 :
1824 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1825 : // `pg_ctl` for start / stop.
1826 : #[instrument(skip_all)]
1827 : fn pg_reload_conf(&self) -> Result<()> {
1828 : let pgctl_bin = Path::new(&self.params.pgbin)
1829 : .parent()
1830 : .unwrap()
1831 : .join("pg_ctl");
1832 : Command::new(pgctl_bin)
1833 : .args(["reload", "-D", &self.params.pgdata])
1834 : .output()
1835 : .expect("cannot run pg_ctl process");
1836 : Ok(())
1837 : }
1838 :
1839 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1840 : /// as it's used to reconfigure a previously started and configured Postgres node.
1841 : #[instrument(skip_all)]
1842 : pub fn reconfigure(&self) -> Result<()> {
1843 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1844 :
1845 : let tls_config = self.tls_config(&spec);
1846 :
1847 : self.update_installed_extensions_collection_interval(&spec);
1848 :
1849 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1850 : info!("tuning pgbouncer");
1851 :
1852 : let pgbouncer_settings = pgbouncer_settings.clone();
1853 : let tls_config = tls_config.clone();
1854 :
1855 : // Spawn a background task to do the tuning,
1856 : // so that we don't block the main thread that starts Postgres.
1857 0 : tokio::spawn(async move {
1858 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
1859 0 : if let Err(err) = res {
1860 0 : error!("error while tuning pgbouncer: {err:?}");
1861 0 : }
1862 0 : });
1863 : }
1864 :
1865 : if let Some(ref local_proxy) = spec.local_proxy_config {
1866 : info!("configuring local_proxy");
1867 :
1868 : // Spawn a background task to do the configuration,
1869 : // so that we don't block the main thread that starts Postgres.
1870 : let mut local_proxy = local_proxy.clone();
1871 : local_proxy.tls = tls_config.clone();
1872 0 : tokio::spawn(async move {
1873 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1874 0 : error!("error while configuring local_proxy: {err:?}");
1875 0 : }
1876 0 : });
1877 : }
1878 :
1879 : // Reconfigure rsyslog for Postgres logs export
1880 : let conf = PostgresLogsRsyslogConfig::new(spec.logs_export_host.as_deref());
1881 : configure_postgres_logs_export(conf)?;
1882 :
1883 : // Write new config
1884 : let pgdata_path = Path::new(&self.params.pgdata);
1885 : config::write_postgres_conf(
1886 : pgdata_path,
1887 : &self.params,
1888 : &spec,
1889 : self.params.internal_http_port,
1890 : tls_config,
1891 : )?;
1892 :
1893 : self.pg_reload_conf()?;
1894 :
1895 : if !spec.skip_pg_catalog_updates {
1896 : let max_concurrent_connections = spec.reconfigure_concurrency;
1897 : // Temporarily reset max_cluster_size in config
1898 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1899 : // creating new extensions, roles, etc.
1900 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1901 0 : self.pg_reload_conf()?;
1902 :
1903 0 : if spec.mode == ComputeMode::Primary {
1904 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:reconfigure"));
1905 0 : let conf = Arc::new(conf);
1906 :
1907 0 : let spec = Arc::new(spec.clone());
1908 :
1909 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
1910 0 : }
1911 :
1912 0 : Ok(())
1913 0 : })?;
1914 : self.pg_reload_conf()?;
1915 : }
1916 :
1917 : let unknown_op = "unknown".to_string();
1918 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1919 : info!(
1920 : "finished reconfiguration of compute node for operation {}",
1921 : op_id
1922 : );
1923 :
1924 : Ok(())
1925 : }
1926 :
1927 : #[instrument(skip_all)]
1928 : pub fn configure_as_primary(&self, compute_state: &ComputeState) -> Result<()> {
1929 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1930 :
1931 : assert!(pspec.spec.mode == ComputeMode::Primary);
1932 : if !pspec.spec.skip_pg_catalog_updates {
1933 : let pgdata_path = Path::new(&self.params.pgdata);
1934 : // temporarily reset max_cluster_size in config
1935 : // to avoid the possibility of hitting the limit, while we are applying config:
1936 : // creating new extensions, roles, etc...
1937 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1938 0 : self.pg_reload_conf()?;
1939 :
1940 0 : self.apply_config(compute_state)?;
1941 :
1942 0 : Ok(())
1943 0 : })?;
1944 :
1945 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1946 : if config::line_in_file(
1947 : &postgresql_conf_path,
1948 : "neon.disable_logical_replication_subscribers=false",
1949 : )? {
1950 : info!(
1951 : "updated postgresql.conf to set neon.disable_logical_replication_subscribers=false"
1952 : );
1953 : }
1954 : self.pg_reload_conf()?;
1955 : }
1956 : self.post_apply_config()?;
1957 :
1958 : Ok(())
1959 : }
1960 :
1961 0 : pub async fn watch_cert_for_changes(self: Arc<Self>) {
1962 : // update status on cert renewal
1963 0 : if let Some(tls_config) = &self.compute_ctl_config.tls {
1964 0 : let tls_config = tls_config.clone();
1965 :
1966 : // wait until the cert exists.
1967 0 : let mut cert_watch = watch_cert_for_changes(tls_config.cert_path.clone()).await;
1968 :
1969 0 : tokio::task::spawn_blocking(move || {
1970 0 : let handle = tokio::runtime::Handle::current();
1971 : 'cert_update: loop {
1972 : // let postgres/pgbouncer/local_proxy know the new cert/key exists.
1973 : // we need to wait until it's configurable first.
1974 :
1975 0 : let mut state = self.state.lock().unwrap();
1976 : 'status_update: loop {
1977 0 : match state.status {
1978 : // let's update the state to config pending
1979 : ComputeStatus::ConfigurationPending | ComputeStatus::Running => {
1980 0 : state.set_status(
1981 0 : ComputeStatus::ConfigurationPending,
1982 0 : &self.state_changed,
1983 0 : );
1984 0 : break 'status_update;
1985 : }
1986 :
1987 : // exit loop
1988 : ComputeStatus::Failed
1989 : | ComputeStatus::TerminationPendingFast
1990 : | ComputeStatus::TerminationPendingImmediate
1991 0 : | ComputeStatus::Terminated => break 'cert_update,
1992 :
1993 : // wait
1994 : ComputeStatus::Init
1995 : | ComputeStatus::Configuration
1996 : | ComputeStatus::RefreshConfigurationPending
1997 0 : | ComputeStatus::Empty => {
1998 0 : state = self.state_changed.wait(state).unwrap();
1999 0 : }
2000 : }
2001 : }
2002 0 : drop(state);
2003 :
2004 : // wait for a new certificate update
2005 0 : if handle.block_on(cert_watch.changed()).is_err() {
2006 0 : break;
2007 0 : }
2008 : }
2009 0 : });
2010 0 : }
2011 0 : }
2012 :
2013 0 : pub fn tls_config(&self, spec: &ComputeSpec) -> &Option<TlsConfig> {
2014 0 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
2015 0 : &self.compute_ctl_config.tls
2016 : } else {
2017 0 : &None::<TlsConfig>
2018 : }
2019 0 : }
2020 :
2021 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
2022 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
2023 0 : let mut state = self.state.lock().unwrap();
2024 : // NB: `Some(<DateTime>)` is always greater than `None`.
2025 0 : if last_active > state.last_active {
2026 0 : state.last_active = last_active;
2027 0 : debug!("set the last compute activity time to: {:?}", last_active);
2028 0 : }
2029 0 : }
2030 :
2031 : // Look for core dumps and collect backtraces.
2032 : //
2033 : // EKS worker nodes have following core dump settings:
2034 : // /proc/sys/kernel/core_pattern -> core
2035 : // /proc/sys/kernel/core_uses_pid -> 1
2036 : // ulimit -c -> unlimited
2037 : // which results in core dumps being written to postgres data directory as core.<pid>.
2038 : //
2039 : // Use that as a default location and pattern, except macos where core dumps are written
2040 : // to /cores/ directory by default.
2041 : //
2042 : // With default Linux settings, the core dump file is called just "core", so check for
2043 : // that too.
2044 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
2045 0 : let core_dump_dir = match std::env::consts::OS {
2046 0 : "macos" => Path::new("/cores/"),
2047 0 : _ => Path::new(&self.params.pgdata),
2048 : };
2049 :
2050 : // Collect core dump paths if any
2051 0 : info!("checking for core dumps in {}", core_dump_dir.display());
2052 0 : let files = fs::read_dir(core_dump_dir)?;
2053 0 : let cores = files.filter_map(|entry| {
2054 0 : let entry = entry.ok()?;
2055 :
2056 0 : let is_core_dump = match entry.file_name().to_str()? {
2057 0 : n if n.starts_with("core.") => true,
2058 0 : "core" => true,
2059 0 : _ => false,
2060 : };
2061 0 : if is_core_dump {
2062 0 : Some(entry.path())
2063 : } else {
2064 0 : None
2065 : }
2066 0 : });
2067 :
2068 : // Print backtrace for each core dump
2069 0 : for core_path in cores {
2070 0 : warn!(
2071 0 : "core dump found: {}, collecting backtrace",
2072 0 : core_path.display()
2073 : );
2074 :
2075 : // Try first with gdb
2076 0 : let backtrace = Command::new("gdb")
2077 0 : .args(["--batch", "-q", "-ex", "bt", &self.params.pgbin])
2078 0 : .arg(&core_path)
2079 0 : .output();
2080 :
2081 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
2082 0 : let backtrace = match backtrace {
2083 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
2084 0 : warn!("cannot find gdb, trying lldb");
2085 0 : Command::new("lldb")
2086 0 : .arg("-c")
2087 0 : .arg(&core_path)
2088 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
2089 0 : .output()
2090 : }
2091 0 : _ => backtrace,
2092 0 : }?;
2093 :
2094 0 : warn!(
2095 0 : "core dump backtrace: {}",
2096 0 : String::from_utf8_lossy(&backtrace.stdout)
2097 : );
2098 0 : warn!(
2099 0 : "debugger stderr: {}",
2100 0 : String::from_utf8_lossy(&backtrace.stderr)
2101 : );
2102 : }
2103 :
2104 0 : Ok(())
2105 0 : }
2106 :
2107 : /// Select `pg_stat_statements` data and return it as a stringified JSON
2108 0 : pub async fn collect_insights(&self) -> String {
2109 0 : let mut result_rows: Vec<String> = Vec::new();
2110 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
2111 0 : let connect_result = conf.connect(NoTls).await;
2112 0 : let (client, connection) = connect_result.unwrap();
2113 0 : tokio::spawn(async move {
2114 0 : if let Err(e) = connection.await {
2115 0 : eprintln!("connection error: {e}");
2116 0 : }
2117 0 : });
2118 0 : let result = client
2119 0 : .simple_query(
2120 0 : "SELECT
2121 0 : row_to_json(pg_stat_statements)
2122 0 : FROM
2123 0 : pg_stat_statements
2124 0 : WHERE
2125 0 : userid != 'cloud_admin'::regrole::oid
2126 0 : ORDER BY
2127 0 : (mean_exec_time + mean_plan_time) DESC
2128 0 : LIMIT 100",
2129 0 : )
2130 0 : .await;
2131 :
2132 0 : if let Ok(raw_rows) = result {
2133 0 : for message in raw_rows.iter() {
2134 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
2135 0 : if let Some(json) = row.get(0) {
2136 0 : result_rows.push(json.to_string());
2137 0 : }
2138 0 : }
2139 : }
2140 :
2141 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
2142 : } else {
2143 0 : "{{\"pg_stat_statements\": []}}".to_string()
2144 : }
2145 0 : }
2146 :
2147 : // download an archive, unzip and place files in correct locations
2148 0 : pub async fn download_extension(
2149 0 : &self,
2150 0 : real_ext_name: String,
2151 0 : ext_path: RemotePath,
2152 0 : ) -> Result<u64, DownloadError> {
2153 0 : let remote_ext_base_url =
2154 0 : self.params
2155 0 : .remote_ext_base_url
2156 0 : .as_ref()
2157 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
2158 0 : "Remote extensions storage is not configured",
2159 0 : )))?;
2160 :
2161 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
2162 :
2163 0 : let mut first_try = false;
2164 0 : if !self
2165 0 : .ext_download_progress
2166 0 : .read()
2167 0 : .expect("lock err")
2168 0 : .contains_key(ext_archive_name)
2169 0 : {
2170 0 : self.ext_download_progress
2171 0 : .write()
2172 0 : .expect("lock err")
2173 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
2174 0 : first_try = true;
2175 0 : }
2176 0 : let (download_start, download_completed) =
2177 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
2178 0 : let start_time_delta = Utc::now()
2179 0 : .signed_duration_since(download_start)
2180 0 : .to_std()
2181 0 : .unwrap()
2182 0 : .as_millis() as u64;
2183 :
2184 : // how long to wait for extension download if it was started by another process
2185 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
2186 :
2187 0 : if download_completed {
2188 0 : info!("extension already downloaded, skipping re-download");
2189 0 : return Ok(0);
2190 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
2191 0 : info!(
2192 0 : "download {ext_archive_name} already started by another process, hanging untill completion or timeout"
2193 : );
2194 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
2195 : loop {
2196 0 : info!("waiting for download");
2197 0 : interval.tick().await;
2198 0 : let (_, download_completed_now) =
2199 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
2200 0 : if download_completed_now {
2201 0 : info!("download finished by whoever else downloaded it");
2202 0 : return Ok(0);
2203 0 : }
2204 : }
2205 : // NOTE: the above loop will get terminated
2206 : // based on the timeout of the download function
2207 0 : }
2208 :
2209 : // if extension hasn't been downloaded before or the previous
2210 : // attempt to download was at least HANG_TIMEOUT ms ago
2211 : // then we try to download it here
2212 0 : info!("downloading new extension {ext_archive_name}");
2213 :
2214 0 : let download_size = extension_server::download_extension(
2215 0 : &real_ext_name,
2216 0 : &ext_path,
2217 0 : remote_ext_base_url,
2218 0 : &self.params.pgbin,
2219 0 : )
2220 0 : .await
2221 0 : .map_err(DownloadError::Other);
2222 :
2223 0 : if download_size.is_ok() {
2224 0 : self.ext_download_progress
2225 0 : .write()
2226 0 : .expect("bad lock")
2227 0 : .insert(ext_archive_name.to_string(), (download_start, true));
2228 0 : }
2229 :
2230 0 : download_size
2231 0 : }
2232 :
2233 0 : pub async fn set_role_grants(
2234 0 : &self,
2235 0 : db_name: &PgIdent,
2236 0 : schema_name: &PgIdent,
2237 0 : privileges: &[Privilege],
2238 0 : role_name: &PgIdent,
2239 0 : ) -> Result<()> {
2240 : use tokio_postgres::NoTls;
2241 :
2242 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
2243 0 : conf.dbname(db_name);
2244 :
2245 0 : let (db_client, conn) = conf
2246 0 : .connect(NoTls)
2247 0 : .await
2248 0 : .context("Failed to connect to the database")?;
2249 0 : tokio::spawn(conn);
2250 :
2251 : // TODO: support other types of grants apart from schemas?
2252 :
2253 : // check the role grants first - to gracefully handle read-replicas.
2254 0 : let select = "SELECT privilege_type
2255 0 : FROM pg_namespace
2256 0 : JOIN LATERAL (SELECT * FROM aclexplode(nspacl) AS x) acl ON true
2257 0 : JOIN pg_user users ON acl.grantee = users.usesysid
2258 0 : WHERE users.usename = $1
2259 0 : AND nspname = $2";
2260 0 : let rows = db_client
2261 0 : .query(select, &[role_name, schema_name])
2262 0 : .await
2263 0 : .with_context(|| format!("Failed to execute query: {select}"))?;
2264 :
2265 0 : let already_granted: HashSet<String> = rows.into_iter().map(|row| row.get(0)).collect();
2266 :
2267 0 : let grants = privileges
2268 0 : .iter()
2269 0 : .filter(|p| !already_granted.contains(p.as_str()))
2270 : // should not be quoted as it's part of the command.
2271 : // is already sanitized so it's ok
2272 0 : .map(|p| p.as_str())
2273 0 : .join(", ");
2274 :
2275 0 : if !grants.is_empty() {
2276 : // quote the schema and role name as identifiers to sanitize them.
2277 0 : let schema_name = schema_name.pg_quote();
2278 0 : let role_name = role_name.pg_quote();
2279 :
2280 0 : let query = format!("GRANT {grants} ON SCHEMA {schema_name} TO {role_name}",);
2281 0 : db_client
2282 0 : .simple_query(&query)
2283 0 : .await
2284 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2285 0 : }
2286 :
2287 0 : Ok(())
2288 0 : }
2289 :
2290 0 : pub async fn install_extension(
2291 0 : &self,
2292 0 : ext_name: &PgIdent,
2293 0 : db_name: &PgIdent,
2294 0 : ext_version: ExtVersion,
2295 0 : ) -> Result<ExtVersion> {
2296 : use tokio_postgres::NoTls;
2297 :
2298 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
2299 0 : conf.dbname(db_name);
2300 :
2301 0 : let (db_client, conn) = conf
2302 0 : .connect(NoTls)
2303 0 : .await
2304 0 : .context("Failed to connect to the database")?;
2305 0 : tokio::spawn(conn);
2306 :
2307 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
2308 0 : let version: Option<ExtVersion> = db_client
2309 0 : .query_opt(version_query, &[&ext_name])
2310 0 : .await
2311 0 : .with_context(|| format!("Failed to execute query: {version_query}"))?
2312 0 : .map(|row| row.get(0));
2313 :
2314 : // sanitize the inputs as postgres idents.
2315 0 : let ext_name: String = ext_name.pg_quote();
2316 0 : let quoted_version: String = ext_version.pg_quote();
2317 :
2318 0 : if let Some(installed_version) = version {
2319 0 : if installed_version == ext_version {
2320 0 : return Ok(installed_version);
2321 0 : }
2322 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
2323 0 : db_client
2324 0 : .simple_query(&query)
2325 0 : .await
2326 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2327 : } else {
2328 0 : let query =
2329 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
2330 0 : db_client
2331 0 : .simple_query(&query)
2332 0 : .await
2333 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2334 : }
2335 :
2336 0 : Ok(ext_version)
2337 0 : }
2338 :
2339 0 : pub async fn prepare_preload_libraries(
2340 0 : &self,
2341 0 : spec: &ComputeSpec,
2342 0 : ) -> Result<RemoteExtensionMetrics> {
2343 0 : if self.params.remote_ext_base_url.is_none() {
2344 0 : return Ok(RemoteExtensionMetrics {
2345 0 : num_ext_downloaded: 0,
2346 0 : largest_ext_size: 0,
2347 0 : total_ext_download_size: 0,
2348 0 : });
2349 0 : }
2350 0 : let remote_extensions = spec
2351 0 : .remote_extensions
2352 0 : .as_ref()
2353 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
2354 :
2355 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
2356 0 : let mut libs_vec = Vec::new();
2357 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
2358 0 : libs_vec = libs
2359 0 : .split(&[',', '\'', ' '])
2360 0 : .filter(|s| *s != "neon" && !s.is_empty())
2361 0 : .map(str::to_string)
2362 0 : .collect();
2363 0 : }
2364 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
2365 :
2366 : // that is used in neon_local and python tests
2367 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
2368 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
2369 0 : let mut shared_preload_libraries_line = "";
2370 0 : for line in conf_lines {
2371 0 : if line.starts_with("shared_preload_libraries") {
2372 0 : shared_preload_libraries_line = line;
2373 0 : }
2374 : }
2375 0 : let mut preload_libs_vec = Vec::new();
2376 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
2377 0 : preload_libs_vec = libs
2378 0 : .split(&[',', '\'', ' '])
2379 0 : .filter(|s| *s != "neon" && !s.is_empty())
2380 0 : .map(str::to_string)
2381 0 : .collect();
2382 0 : }
2383 0 : libs_vec.extend(preload_libs_vec);
2384 0 : }
2385 :
2386 : // Don't try to download libraries that are not in the index.
2387 : // Assume that they are already present locally.
2388 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
2389 :
2390 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
2391 :
2392 0 : let mut download_tasks = Vec::new();
2393 0 : for library in &libs_vec {
2394 0 : let (ext_name, ext_path) =
2395 0 : remote_extensions.get_ext(library, true, &BUILD_TAG, &self.params.pgversion)?;
2396 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
2397 : }
2398 0 : let results = join_all(download_tasks).await;
2399 :
2400 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
2401 0 : num_ext_downloaded: 0,
2402 0 : largest_ext_size: 0,
2403 0 : total_ext_download_size: 0,
2404 0 : };
2405 0 : for result in results {
2406 0 : let download_size = match result {
2407 0 : Ok(res) => {
2408 0 : remote_ext_metrics.num_ext_downloaded += 1;
2409 0 : res
2410 : }
2411 0 : Err(err) => {
2412 : // if we failed to download an extension, we don't want to fail the whole
2413 : // process, but we do want to log the error
2414 0 : error!("Failed to download extension: {}", err);
2415 0 : 0
2416 : }
2417 : };
2418 :
2419 0 : remote_ext_metrics.largest_ext_size =
2420 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
2421 0 : remote_ext_metrics.total_ext_download_size += download_size;
2422 : }
2423 0 : Ok(remote_ext_metrics)
2424 0 : }
2425 :
2426 : /// Waits until current thread receives a state changed notification and
2427 : /// the pageserver connection strings has changed.
2428 : ///
2429 : /// The operation will time out after a specified duration.
2430 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
2431 0 : let state = self.state.lock().unwrap();
2432 0 : let old_pageserver_connstr = state
2433 0 : .pspec
2434 0 : .as_ref()
2435 0 : .expect("spec must be set")
2436 0 : .pageserver_connstr
2437 0 : .clone();
2438 0 : let mut unchanged = true;
2439 0 : let _ = self
2440 0 : .state_changed
2441 0 : .wait_timeout_while(state, duration, |s| {
2442 0 : let pageserver_connstr = &s
2443 0 : .pspec
2444 0 : .as_ref()
2445 0 : .expect("spec must be set")
2446 0 : .pageserver_connstr;
2447 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
2448 0 : unchanged
2449 0 : })
2450 0 : .unwrap();
2451 0 : if !unchanged {
2452 0 : info!("Pageserver config changed");
2453 0 : }
2454 0 : }
2455 :
2456 0 : pub fn spawn_extension_stats_task(&self) {
2457 0 : self.terminate_extension_stats_task();
2458 :
2459 0 : let conf = self.tokio_conn_conf.clone();
2460 0 : let atomic_interval = self.params.installed_extensions_collection_interval.clone();
2461 0 : let mut installed_extensions_collection_interval =
2462 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst);
2463 0 : info!(
2464 0 : "[NEON_EXT_SPAWN] Spawning background installed extensions worker with Timeout: {}",
2465 : installed_extensions_collection_interval
2466 : );
2467 0 : let handle = tokio::spawn(async move {
2468 : loop {
2469 0 : info!(
2470 0 : "[NEON_EXT_INT_SLEEP]: Interval: {}",
2471 : installed_extensions_collection_interval
2472 : );
2473 : // Sleep at the start of the loop to ensure that two collections don't happen at the same time.
2474 : // The first collection happens during compute startup.
2475 0 : tokio::time::sleep(tokio::time::Duration::from_secs(
2476 0 : installed_extensions_collection_interval,
2477 0 : ))
2478 0 : .await;
2479 0 : let _ = installed_extensions(conf.clone()).await;
2480 : // Acquire a read lock on the compute spec and then update the interval if necessary
2481 0 : installed_extensions_collection_interval = std::cmp::max(
2482 0 : installed_extensions_collection_interval,
2483 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst),
2484 0 : );
2485 : }
2486 : });
2487 :
2488 : // Store the new task handle
2489 0 : *self.extension_stats_task.lock().unwrap() = Some(handle);
2490 0 : }
2491 :
2492 0 : fn terminate_extension_stats_task(&self) {
2493 0 : if let Some(h) = self.extension_stats_task.lock().unwrap().take() {
2494 0 : h.abort()
2495 0 : }
2496 0 : }
2497 :
2498 0 : pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
2499 0 : self.terminate_lfc_offload_task();
2500 0 : let secs = interval.as_secs();
2501 0 : let this = self.clone();
2502 :
2503 0 : info!("spawning LFC offload worker with {secs}s interval");
2504 0 : let handle = spawn(async move {
2505 0 : let mut interval = time::interval(interval);
2506 0 : interval.tick().await; // returns immediately
2507 : loop {
2508 0 : interval.tick().await;
2509 :
2510 0 : let prewarm_state = this.state.lock().unwrap().lfc_prewarm_state.clone();
2511 : // Do not offload LFC state if we are currently prewarming or any issue occurred.
2512 : // If we'd do that, we might override the LFC state in endpoint storage with some
2513 : // incomplete state. Imagine a situation:
2514 : // 1. Endpoint started with `autoprewarm: true`
2515 : // 2. While prewarming is not completed, we upload the new incomplete state
2516 : // 3. Compute gets interrupted and restarts
2517 : // 4. We start again and try to prewarm with the state from 2. instead of the previous complete state
2518 0 : if matches!(
2519 0 : prewarm_state,
2520 : LfcPrewarmState::Completed
2521 : | LfcPrewarmState::NotPrewarmed
2522 : | LfcPrewarmState::Skipped
2523 : ) {
2524 0 : this.offload_lfc_async().await;
2525 0 : }
2526 : }
2527 : });
2528 0 : *self.lfc_offload_task.lock().unwrap() = Some(handle);
2529 0 : }
2530 :
2531 0 : fn terminate_lfc_offload_task(&self) {
2532 0 : if let Some(h) = self.lfc_offload_task.lock().unwrap().take() {
2533 0 : h.abort()
2534 0 : }
2535 0 : }
2536 :
2537 0 : fn update_installed_extensions_collection_interval(&self, spec: &ComputeSpec) {
2538 : // Update the interval for collecting installed extensions statistics
2539 : // If the value is -1, we never suspend so set the value to default collection.
2540 : // If the value is 0, it means default, we will just continue to use the default.
2541 0 : if spec.suspend_timeout_seconds == -1 || spec.suspend_timeout_seconds == 0 {
2542 0 : self.params.installed_extensions_collection_interval.store(
2543 0 : DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL,
2544 0 : std::sync::atomic::Ordering::SeqCst,
2545 0 : );
2546 0 : } else {
2547 0 : self.params.installed_extensions_collection_interval.store(
2548 0 : spec.suspend_timeout_seconds as u64,
2549 0 : std::sync::atomic::Ordering::SeqCst,
2550 0 : );
2551 0 : }
2552 0 : }
2553 : }
2554 :
2555 0 : pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
2556 0 : let res = get_installed_extensions(conf).await;
2557 0 : match res {
2558 0 : Ok(extensions) => {
2559 0 : info!(
2560 0 : "[NEON_EXT_STAT] {}",
2561 0 : serde_json::to_string(&extensions).expect("failed to serialize extensions list")
2562 : );
2563 : }
2564 0 : Err(err) => error!("could not get installed extensions: {err}"),
2565 : }
2566 0 : Ok(())
2567 0 : }
2568 :
2569 0 : pub fn forward_termination_signal(dev_mode: bool) {
2570 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2571 0 : if ss_pid != 0 {
2572 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2573 0 : kill(ss_pid, Signal::SIGTERM).ok();
2574 0 : }
2575 :
2576 0 : if !dev_mode {
2577 : // Terminate pgbouncer with SIGKILL
2578 0 : match pid_file::read(PGBOUNCER_PIDFILE.into()) {
2579 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2580 0 : info!("sending SIGKILL to pgbouncer process pid: {}", pid);
2581 0 : if let Err(e) = kill(pid, Signal::SIGKILL) {
2582 0 : error!("failed to terminate pgbouncer: {}", e);
2583 0 : }
2584 : }
2585 : // pgbouncer does not lock the pid file, so we read and kill the process directly
2586 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2587 0 : if let Ok(pid_str) = std::fs::read_to_string(PGBOUNCER_PIDFILE) {
2588 0 : if let Ok(pid) = pid_str.trim().parse::<i32>() {
2589 0 : info!(
2590 0 : "sending SIGKILL to pgbouncer process pid: {} (from unlocked pid file)",
2591 : pid
2592 : );
2593 0 : if let Err(e) = kill(Pid::from_raw(pid), Signal::SIGKILL) {
2594 0 : error!("failed to terminate pgbouncer: {}", e);
2595 0 : }
2596 0 : }
2597 : } else {
2598 0 : info!("pgbouncer pid file exists but process not running");
2599 : }
2600 : }
2601 : Ok(pid_file::PidFileRead::NotExist) => {
2602 0 : info!("pgbouncer pid file not found, process may not be running");
2603 : }
2604 0 : Err(e) => {
2605 0 : error!("error reading pgbouncer pid file: {}", e);
2606 : }
2607 : }
2608 :
2609 : // Terminate local_proxy
2610 0 : match pid_file::read("/etc/local_proxy/pid".into()) {
2611 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2612 0 : info!("sending SIGTERM to local_proxy process pid: {}", pid);
2613 0 : if let Err(e) = kill(pid, Signal::SIGTERM) {
2614 0 : error!("failed to terminate local_proxy: {}", e);
2615 0 : }
2616 : }
2617 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2618 0 : info!("local_proxy PID file exists but process not running");
2619 : }
2620 : Ok(pid_file::PidFileRead::NotExist) => {
2621 0 : info!("local_proxy PID file not found, process may not be running");
2622 : }
2623 0 : Err(e) => {
2624 0 : error!("error reading local_proxy PID file: {}", e);
2625 : }
2626 : }
2627 : } else {
2628 0 : info!("Skipping pgbouncer and local_proxy termination because in dev mode");
2629 : }
2630 :
2631 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2632 0 : if pg_pid != 0 {
2633 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2634 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2635 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2636 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2637 0 : kill(pg_pid, Signal::SIGINT).ok();
2638 0 : }
2639 0 : }
2640 :
2641 : // helper trait to call JoinSet::spawn_blocking(f), but propagates the current
2642 : // tracing span to the thread.
2643 : trait JoinSetExt<T> {
2644 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2645 : where
2646 : F: FnOnce() -> T + Send + 'static,
2647 : T: Send;
2648 : }
2649 :
2650 : impl<T: 'static> JoinSetExt<T> for tokio::task::JoinSet<T> {
2651 0 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2652 0 : where
2653 0 : F: FnOnce() -> T + Send + 'static,
2654 0 : T: Send,
2655 : {
2656 0 : let sp = tracing::Span::current();
2657 0 : self.spawn_blocking(move || {
2658 0 : let _e = sp.enter();
2659 0 : f()
2660 0 : })
2661 0 : }
2662 : }
2663 :
2664 : #[cfg(test)]
2665 : mod tests {
2666 : use std::fs::File;
2667 :
2668 : use super::*;
2669 :
2670 : #[test]
2671 1 : fn duplicate_safekeeper_connstring() {
2672 1 : let file = File::open("tests/cluster_spec.json").unwrap();
2673 1 : let spec: ComputeSpec = serde_json::from_reader(file).unwrap();
2674 :
2675 1 : match ParsedSpec::try_from(spec.clone()) {
2676 0 : Ok(_p) => panic!("Failed to detect duplicate entry"),
2677 1 : Err(e) => assert!(e.starts_with("duplicate entry in safekeeper_connstrings:")),
2678 : };
2679 1 : }
2680 : }
|
