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