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