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;
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 : // TODO: make this more robust
629 0 : // now rsyslog starts once and there is no monitoring or restart if it fails
630 0 : configure_audit_rsyslog(
631 0 : log_directory_path.to_str().unwrap(),
632 0 : "hipaa",
633 0 : &remote_endpoint,
634 0 : )?;
635 0 : }
636 :
637 : // Launch remaining service threads
638 0 : let _monitor_handle = launch_monitor(self);
639 0 : let _configurator_handle = launch_configurator(self);
640 0 :
641 0 : // Wait for all the pre-tasks to finish before starting postgres
642 0 : let rt = tokio::runtime::Handle::current();
643 0 : while let Some(res) = rt.block_on(pre_tasks.join_next()) {
644 0 : res??;
645 : }
646 :
647 : ////// START POSTGRES
648 0 : let start_time = Utc::now();
649 0 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
650 0 : let postmaster_pid = pg_process.pid();
651 0 : *pg_handle = Some(pg_process);
652 0 :
653 0 : // If this is a primary endpoint, perform some post-startup configuration before
654 0 : // opening it up for the world.
655 0 : let config_time = Utc::now();
656 0 : if pspec.spec.mode == ComputeMode::Primary {
657 0 : self.configure_as_primary(&compute_state)?;
658 :
659 0 : let conf = self.get_tokio_conn_conf(None);
660 0 : tokio::task::spawn(async {
661 0 : let res = get_installed_extensions(conf).await;
662 0 : match res {
663 0 : Ok(extensions) => {
664 0 : info!(
665 0 : "[NEON_EXT_STAT] {}",
666 0 : serde_json::to_string(&extensions)
667 0 : .expect("failed to serialize extensions list")
668 : );
669 : }
670 0 : Err(err) => error!("could not get installed extensions: {err:?}"),
671 : }
672 0 : });
673 0 : }
674 :
675 : // All done!
676 0 : let startup_end_time = Utc::now();
677 0 : let metrics = {
678 0 : let mut state = self.state.lock().unwrap();
679 0 : state.metrics.start_postgres_ms = config_time
680 0 : .signed_duration_since(start_time)
681 0 : .to_std()
682 0 : .unwrap()
683 0 : .as_millis() as u64;
684 0 : state.metrics.config_ms = startup_end_time
685 0 : .signed_duration_since(config_time)
686 0 : .to_std()
687 0 : .unwrap()
688 0 : .as_millis() as u64;
689 0 : state.metrics.total_startup_ms = startup_end_time
690 0 : .signed_duration_since(compute_state.start_time)
691 0 : .to_std()
692 0 : .unwrap()
693 0 : .as_millis() as u64;
694 0 : state.metrics.clone()
695 0 : };
696 0 : self.set_status(ComputeStatus::Running);
697 0 :
698 0 : // Log metrics so that we can search for slow operations in logs
699 0 : info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
700 :
701 0 : Ok(())
702 0 : }
703 :
704 : #[instrument(skip_all)]
705 : async fn download_preload_extensions(&self, spec: &ComputeSpec) -> Result<()> {
706 : let remote_extensions = if let Some(remote_extensions) = &spec.remote_extensions {
707 : remote_extensions
708 : } else {
709 : return Ok(());
710 : };
711 :
712 : // First, create control files for all available extensions
713 : extension_server::create_control_files(remote_extensions, &self.params.pgbin);
714 :
715 : let library_load_start_time = Utc::now();
716 : let remote_ext_metrics = self.prepare_preload_libraries(spec).await?;
717 :
718 : let library_load_time = Utc::now()
719 : .signed_duration_since(library_load_start_time)
720 : .to_std()
721 : .unwrap()
722 : .as_millis() as u64;
723 : let mut state = self.state.lock().unwrap();
724 : state.metrics.load_ext_ms = library_load_time;
725 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
726 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
727 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
728 : info!(
729 : "Loading shared_preload_libraries took {:?}ms",
730 : library_load_time
731 : );
732 : info!("{:?}", remote_ext_metrics);
733 :
734 : Ok(())
735 : }
736 :
737 : /// Start the vm-monitor if directed to. The vm-monitor only runs on linux
738 : /// because it requires cgroups.
739 0 : fn start_vm_monitor(&self, disable_lfc_resizing: bool) -> StartVmMonitorResult {
740 : cfg_if::cfg_if! {
741 : if #[cfg(target_os = "linux")] {
742 : use std::env;
743 : use tokio_util::sync::CancellationToken;
744 :
745 : // This token is used internally by the monitor to clean up all threads
746 0 : let token = CancellationToken::new();
747 :
748 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
749 0 : let pgconnstr = if disable_lfc_resizing {
750 0 : None
751 : } else {
752 0 : Some(self.params.filecache_connstr.clone())
753 : };
754 :
755 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
756 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
757 0 : Box::leak(Box::new(vm_monitor::Args {
758 0 : cgroup: Some(self.params.cgroup.clone()),
759 0 : pgconnstr,
760 0 : addr: self.params.vm_monitor_addr.clone(),
761 0 : })),
762 0 : token.clone(),
763 0 : ));
764 0 : Some(vm_monitor)
765 : } else {
766 0 : None
767 : };
768 0 : StartVmMonitorResult { token, vm_monitor }
769 0 : } else {
770 0 : _ = disable_lfc_resizing; // appease unused lint on macOS
771 0 : StartVmMonitorResult { }
772 0 : }
773 0 : }
774 0 : }
775 :
776 0 : fn cleanup_after_postgres_exit(&self) -> Result<bool> {
777 0 : // Maybe sync safekeepers again, to speed up next startup
778 0 : let compute_state = self.state.lock().unwrap().clone();
779 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
780 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
781 0 : info!("syncing safekeepers on shutdown");
782 0 : let storage_auth_token = pspec.storage_auth_token.clone();
783 0 : let lsn = self.sync_safekeepers(storage_auth_token)?;
784 0 : info!("synced safekeepers at lsn {lsn}");
785 0 : }
786 :
787 0 : let mut delay_exit = false;
788 0 : let mut state = self.state.lock().unwrap();
789 0 : if state.status == ComputeStatus::TerminationPending {
790 0 : state.status = ComputeStatus::Terminated;
791 0 : self.state_changed.notify_all();
792 0 : // we were asked to terminate gracefully, don't exit to avoid restart
793 0 : delay_exit = true
794 0 : }
795 0 : drop(state);
796 :
797 0 : if let Err(err) = self.check_for_core_dumps() {
798 0 : error!("error while checking for core dumps: {err:?}");
799 0 : }
800 :
801 0 : Ok(delay_exit)
802 0 : }
803 :
804 : /// Check that compute node has corresponding feature enabled.
805 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
806 0 : let state = self.state.lock().unwrap();
807 :
808 0 : if let Some(s) = state.pspec.as_ref() {
809 0 : s.spec.features.contains(&feature)
810 : } else {
811 0 : false
812 : }
813 0 : }
814 :
815 0 : pub fn set_status(&self, status: ComputeStatus) {
816 0 : let mut state = self.state.lock().unwrap();
817 0 : state.set_status(status, &self.state_changed);
818 0 : }
819 :
820 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
821 0 : let mut state = self.state.lock().unwrap();
822 0 : state.set_failed_status(err, &self.state_changed);
823 0 : }
824 :
825 0 : pub fn get_status(&self) -> ComputeStatus {
826 0 : self.state.lock().unwrap().status
827 0 : }
828 :
829 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
830 0 : self.state
831 0 : .lock()
832 0 : .unwrap()
833 0 : .pspec
834 0 : .as_ref()
835 0 : .map(|s| s.timeline_id)
836 0 : }
837 :
838 : // Remove `pgdata` directory and create it again with right permissions.
839 0 : fn create_pgdata(&self) -> Result<()> {
840 0 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
841 0 : // If it is something different then create_dir() will error out anyway.
842 0 : let pgdata = &self.params.pgdata;
843 0 : let _ok = fs::remove_dir_all(pgdata);
844 0 : fs::create_dir(pgdata)?;
845 0 : fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
846 :
847 0 : Ok(())
848 0 : }
849 :
850 : // Get basebackup from the libpq connection to pageserver using `connstr` and
851 : // unarchive it to `pgdata` directory overriding all its previous content.
852 : #[instrument(skip_all, fields(%lsn))]
853 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
854 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
855 : let start_time = Instant::now();
856 :
857 : let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
858 : let mut config = postgres::Config::from_str(shard0_connstr)?;
859 :
860 : // Use the storage auth token from the config file, if given.
861 : // Note: this overrides any password set in the connection string.
862 : if let Some(storage_auth_token) = &spec.storage_auth_token {
863 : info!("Got storage auth token from spec file");
864 : config.password(storage_auth_token);
865 : } else {
866 : info!("Storage auth token not set");
867 : }
868 :
869 : // Connect to pageserver
870 : let mut client = config.connect(NoTls)?;
871 : let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
872 :
873 : let basebackup_cmd = match lsn {
874 : Lsn(0) => {
875 : if spec.spec.mode != ComputeMode::Primary {
876 : format!(
877 : "basebackup {} {} --gzip --replica",
878 : spec.tenant_id, spec.timeline_id
879 : )
880 : } else {
881 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
882 : }
883 : }
884 : _ => {
885 : if spec.spec.mode != ComputeMode::Primary {
886 : format!(
887 : "basebackup {} {} {} --gzip --replica",
888 : spec.tenant_id, spec.timeline_id, lsn
889 : )
890 : } else {
891 : format!(
892 : "basebackup {} {} {} --gzip",
893 : spec.tenant_id, spec.timeline_id, lsn
894 : )
895 : }
896 : }
897 : };
898 :
899 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
900 : let mut measured_reader = MeasuredReader::new(copyreader);
901 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
902 :
903 : // Read the archive directly from the `CopyOutReader`
904 : //
905 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
906 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
907 : // sends an Error after finishing the tarball, we will not notice it.
908 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
909 : ar.set_ignore_zeros(true);
910 : ar.unpack(&self.params.pgdata)?;
911 :
912 : // Report metrics
913 : let mut state = self.state.lock().unwrap();
914 : state.metrics.pageserver_connect_micros = pageserver_connect_micros;
915 : state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
916 : state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
917 : Ok(())
918 : }
919 :
920 : // Gets the basebackup in a retry loop
921 : #[instrument(skip_all, fields(%lsn))]
922 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
923 : let mut retry_period_ms = 500.0;
924 : let mut attempts = 0;
925 : const DEFAULT_ATTEMPTS: u16 = 10;
926 : #[cfg(feature = "testing")]
927 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
928 : u16::from_str(&v).unwrap()
929 : } else {
930 : DEFAULT_ATTEMPTS
931 : };
932 : #[cfg(not(feature = "testing"))]
933 : let max_attempts = DEFAULT_ATTEMPTS;
934 : loop {
935 : let result = self.try_get_basebackup(compute_state, lsn);
936 : match result {
937 : Ok(_) => {
938 : return result;
939 : }
940 : Err(ref e) if attempts < max_attempts => {
941 : warn!(
942 : "Failed to get basebackup: {} (attempt {}/{})",
943 : e, attempts, max_attempts
944 : );
945 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
946 : retry_period_ms *= 1.5;
947 : }
948 : Err(_) => {
949 : return result;
950 : }
951 : }
952 : attempts += 1;
953 : }
954 : }
955 :
956 0 : pub async fn check_safekeepers_synced_async(
957 0 : &self,
958 0 : compute_state: &ComputeState,
959 0 : ) -> Result<Option<Lsn>> {
960 0 : // Construct a connection config for each safekeeper
961 0 : let pspec: ParsedSpec = compute_state
962 0 : .pspec
963 0 : .as_ref()
964 0 : .expect("spec must be set")
965 0 : .clone();
966 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
967 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
968 0 : // Format connstr
969 0 : let id = connstr.clone();
970 0 : let connstr = format!("postgresql://no_user@{}", connstr);
971 0 : let options = format!(
972 0 : "-c timeline_id={} tenant_id={}",
973 0 : pspec.timeline_id, pspec.tenant_id
974 0 : );
975 0 :
976 0 : // Construct client
977 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
978 0 : config.options(&options);
979 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
980 0 : config.password(storage_auth_token);
981 0 : }
982 :
983 0 : (id, config)
984 0 : });
985 0 :
986 0 : // Create task set to query all safekeepers
987 0 : let mut tasks = FuturesUnordered::new();
988 0 : let quorum = sk_configs.len() / 2 + 1;
989 0 : for (id, config) in sk_configs {
990 0 : let timeout = tokio::time::Duration::from_millis(100);
991 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
992 0 : tasks.push(tokio::spawn(task));
993 0 : }
994 :
995 : // Get a quorum of responses or errors
996 0 : let mut responses = Vec::new();
997 0 : let mut join_errors = Vec::new();
998 0 : let mut task_errors = Vec::new();
999 0 : let mut timeout_errors = Vec::new();
1000 0 : while let Some(response) = tasks.next().await {
1001 0 : match response {
1002 0 : Ok(Ok(Ok(r))) => responses.push(r),
1003 0 : Ok(Ok(Err(e))) => task_errors.push(e),
1004 0 : Ok(Err(e)) => timeout_errors.push(e),
1005 0 : Err(e) => join_errors.push(e),
1006 : };
1007 0 : if responses.len() >= quorum {
1008 0 : break;
1009 0 : }
1010 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
1011 0 : break;
1012 0 : }
1013 : }
1014 :
1015 : // In case of error, log and fail the check, but don't crash.
1016 : // We're playing it safe because these errors could be transient
1017 : // and we don't yet retry. Also being careful here allows us to
1018 : // be backwards compatible with safekeepers that don't have the
1019 : // TIMELINE_STATUS API yet.
1020 0 : if responses.len() < quorum {
1021 0 : error!(
1022 0 : "failed sync safekeepers check {:?} {:?} {:?}",
1023 : join_errors, task_errors, timeout_errors
1024 : );
1025 0 : return Ok(None);
1026 0 : }
1027 0 :
1028 0 : Ok(check_if_synced(responses))
1029 0 : }
1030 :
1031 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
1032 : // in one roundtrip. If not, we should do a full sync_safekeepers.
1033 : #[instrument(skip_all)]
1034 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
1035 : let start_time = Utc::now();
1036 :
1037 : let rt = tokio::runtime::Handle::current();
1038 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
1039 :
1040 : // Record runtime
1041 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
1042 : .signed_duration_since(start_time)
1043 : .to_std()
1044 : .unwrap()
1045 : .as_millis() as u64;
1046 : result
1047 : }
1048 :
1049 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
1050 : // and return the reported LSN back to the caller.
1051 : #[instrument(skip_all)]
1052 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
1053 : let start_time = Utc::now();
1054 :
1055 : let mut sync_handle = maybe_cgexec(&self.params.pgbin)
1056 : .args(["--sync-safekeepers"])
1057 : .env("PGDATA", &self.params.pgdata) // we cannot use -D in this mode
1058 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1059 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1060 : } else {
1061 : vec![]
1062 : })
1063 : .stdout(Stdio::piped())
1064 : .stderr(Stdio::piped())
1065 : .spawn()
1066 : .expect("postgres --sync-safekeepers failed to start");
1067 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
1068 :
1069 : // `postgres --sync-safekeepers` will print all log output to stderr and
1070 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
1071 : // will be collected in a child thread.
1072 : let stderr = sync_handle
1073 : .stderr
1074 : .take()
1075 : .expect("stderr should be captured");
1076 : let logs_handle = handle_postgres_logs(stderr);
1077 :
1078 : let sync_output = sync_handle
1079 : .wait_with_output()
1080 : .expect("postgres --sync-safekeepers failed");
1081 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
1082 :
1083 : // Process has exited, so we can join the logs thread.
1084 : let _ = tokio::runtime::Handle::current()
1085 : .block_on(logs_handle)
1086 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1087 :
1088 : if !sync_output.status.success() {
1089 : anyhow::bail!(
1090 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
1091 : sync_output.status,
1092 : String::from_utf8(sync_output.stdout)
1093 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
1094 : );
1095 : }
1096 :
1097 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
1098 : .signed_duration_since(start_time)
1099 : .to_std()
1100 : .unwrap()
1101 : .as_millis() as u64;
1102 :
1103 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
1104 :
1105 : Ok(lsn)
1106 : }
1107 :
1108 : /// Do all the preparations like PGDATA directory creation, configuration,
1109 : /// safekeepers sync, basebackup, etc.
1110 : #[instrument(skip_all)]
1111 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
1112 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1113 : let spec = &pspec.spec;
1114 : let pgdata_path = Path::new(&self.params.pgdata);
1115 :
1116 : // Remove/create an empty pgdata directory and put configuration there.
1117 : self.create_pgdata()?;
1118 : config::write_postgres_conf(
1119 : pgdata_path,
1120 : &pspec.spec,
1121 : self.params.internal_http_port,
1122 : &self.compute_ctl_config.tls,
1123 : )?;
1124 :
1125 : // Syncing safekeepers is only safe with primary nodes: if a primary
1126 : // is already connected it will be kicked out, so a secondary (standby)
1127 : // cannot sync safekeepers.
1128 : let lsn = match spec.mode {
1129 : ComputeMode::Primary => {
1130 : info!("checking if safekeepers are synced");
1131 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
1132 : lsn
1133 : } else {
1134 : info!("starting safekeepers syncing");
1135 : self.sync_safekeepers(pspec.storage_auth_token.clone())
1136 0 : .with_context(|| "failed to sync safekeepers")?
1137 : };
1138 : info!("safekeepers synced at LSN {}", lsn);
1139 : lsn
1140 : }
1141 : ComputeMode::Static(lsn) => {
1142 : info!("Starting read-only node at static LSN {}", lsn);
1143 : lsn
1144 : }
1145 : ComputeMode::Replica => {
1146 : info!("Initializing standby from latest Pageserver LSN");
1147 : Lsn(0)
1148 : }
1149 : };
1150 :
1151 : info!(
1152 : "getting basebackup@{} from pageserver {}",
1153 : lsn, &pspec.pageserver_connstr
1154 : );
1155 0 : self.get_basebackup(compute_state, lsn).with_context(|| {
1156 0 : format!(
1157 0 : "failed to get basebackup@{} from pageserver {}",
1158 0 : lsn, &pspec.pageserver_connstr
1159 0 : )
1160 0 : })?;
1161 :
1162 : // Update pg_hba.conf received with basebackup.
1163 : update_pg_hba(pgdata_path)?;
1164 :
1165 : // Place pg_dynshmem under /dev/shm. This allows us to use
1166 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
1167 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
1168 : //
1169 : // Why on earth don't we just stick to the 'posix' default, you might
1170 : // ask. It turns out that making large allocations with 'posix' doesn't
1171 : // work very well with autoscaling. The behavior we want is that:
1172 : //
1173 : // 1. You can make large DSM allocations, larger than the current RAM
1174 : // size of the VM, without errors
1175 : //
1176 : // 2. If the allocated memory is really used, the VM is scaled up
1177 : // automatically to accommodate that
1178 : //
1179 : // We try to make that possible by having swap in the VM. But with the
1180 : // default 'posix' DSM implementation, we fail step 1, even when there's
1181 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
1182 : // the shmem segment, which is really just a file in /dev/shm in Linux,
1183 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
1184 : // than available RAM.
1185 : //
1186 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
1187 : // the Postgres 'mmap' DSM implementation doesn't use
1188 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
1189 : // fill the file with zeros. It's weird that that differs between
1190 : // 'posix' and 'mmap', but we take advantage of it. When the file is
1191 : // filled slowly with write(2), the kernel allows it to grow larger, as
1192 : // long as there's swap available.
1193 : //
1194 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
1195 : // segment to be larger than currently available RAM. But because we
1196 : // don't want to store it on a real file, which the kernel would try to
1197 : // flush to disk, so symlink pg_dynshm to /dev/shm.
1198 : //
1199 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
1200 : // control plane control that option. If 'mmap' is not used, this
1201 : // symlink doesn't affect anything.
1202 : //
1203 : // See https://github.com/neondatabase/autoscaling/issues/800
1204 : std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
1205 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
1206 :
1207 : match spec.mode {
1208 : ComputeMode::Primary => {}
1209 : ComputeMode::Replica | ComputeMode::Static(..) => {
1210 : add_standby_signal(pgdata_path)?;
1211 : }
1212 : }
1213 :
1214 : Ok(())
1215 : }
1216 :
1217 : /// Start and stop a postgres process to warm up the VM for startup.
1218 0 : pub fn prewarm_postgres(&self) -> Result<()> {
1219 0 : info!("prewarming");
1220 :
1221 : // Create pgdata
1222 0 : let pgdata = &format!("{}.warmup", self.params.pgdata);
1223 0 : create_pgdata(pgdata)?;
1224 :
1225 : // Run initdb to completion
1226 0 : info!("running initdb");
1227 0 : let initdb_bin = Path::new(&self.params.pgbin)
1228 0 : .parent()
1229 0 : .unwrap()
1230 0 : .join("initdb");
1231 0 : Command::new(initdb_bin)
1232 0 : .args(["--pgdata", pgdata])
1233 0 : .output()
1234 0 : .expect("cannot start initdb process");
1235 :
1236 : // Write conf
1237 : use std::io::Write;
1238 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
1239 0 : let mut file = std::fs::File::create(conf_path)?;
1240 0 : writeln!(file, "shared_buffers=65536")?;
1241 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
1242 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
1243 :
1244 : // Start postgres
1245 0 : info!("starting postgres");
1246 0 : let mut pg = maybe_cgexec(&self.params.pgbin)
1247 0 : .args(["-D", pgdata])
1248 0 : .spawn()
1249 0 : .expect("cannot start postgres process");
1250 0 :
1251 0 : // Stop it when it's ready
1252 0 : info!("waiting for postgres");
1253 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
1254 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
1255 : // it to avoid orphaned processes prowling around while datadir is
1256 : // wiped.
1257 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
1258 0 : kill(pm_pid, Signal::SIGQUIT)?;
1259 0 : info!("sent SIGQUIT signal");
1260 0 : pg.wait()?;
1261 0 : info!("done prewarming");
1262 :
1263 : // clean up
1264 0 : let _ok = fs::remove_dir_all(pgdata);
1265 0 : Ok(())
1266 0 : }
1267 :
1268 : /// Start Postgres as a child process and wait for it to start accepting
1269 : /// connections.
1270 : ///
1271 : /// Returns a handle to the child process and a handle to the logs thread.
1272 : #[instrument(skip_all)]
1273 : pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
1274 : let pgdata_path = Path::new(&self.params.pgdata);
1275 :
1276 : // Run postgres as a child process.
1277 : let mut pg = maybe_cgexec(&self.params.pgbin)
1278 : .args(["-D", &self.params.pgdata])
1279 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1280 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1281 : } else {
1282 : vec![]
1283 : })
1284 : .stderr(Stdio::piped())
1285 : .spawn()
1286 : .expect("cannot start postgres process");
1287 : PG_PID.store(pg.id(), Ordering::SeqCst);
1288 :
1289 : // Start a task to collect logs from stderr.
1290 : let stderr = pg.stderr.take().expect("stderr should be captured");
1291 : let logs_handle = handle_postgres_logs(stderr);
1292 :
1293 : wait_for_postgres(&mut pg, pgdata_path)?;
1294 :
1295 : Ok(PostgresHandle {
1296 : postgres: pg,
1297 : log_collector: logs_handle,
1298 : })
1299 : }
1300 :
1301 : /// Wait for the child Postgres process forever. In this state Ctrl+C will
1302 : /// propagate to Postgres and it will be shut down as well.
1303 0 : fn wait_postgres(&self, mut pg_handle: PostgresHandle) -> std::process::ExitStatus {
1304 0 : info!(postmaster_pid = %pg_handle.postgres.id(), "Waiting for Postgres to exit");
1305 :
1306 0 : let ecode = pg_handle
1307 0 : .postgres
1308 0 : .wait()
1309 0 : .expect("failed to start waiting on Postgres process");
1310 0 : PG_PID.store(0, Ordering::SeqCst);
1311 0 :
1312 0 : // Process has exited. Wait for the log collecting task to finish.
1313 0 : let _ = tokio::runtime::Handle::current()
1314 0 : .block_on(pg_handle.log_collector)
1315 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1316 0 :
1317 0 : ecode
1318 0 : }
1319 :
1320 : /// Do post configuration of the already started Postgres. This function spawns a background task to
1321 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
1322 : /// version. In the future, it may upgrade all 3rd-party extensions.
1323 : #[instrument(skip_all)]
1324 : pub fn post_apply_config(&self) -> Result<()> {
1325 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
1326 0 : tokio::spawn(async move {
1327 0 : let res = async {
1328 0 : let (mut client, connection) = conf.connect(NoTls).await?;
1329 0 : tokio::spawn(async move {
1330 0 : if let Err(e) = connection.await {
1331 0 : eprintln!("connection error: {}", e);
1332 0 : }
1333 0 : });
1334 0 :
1335 0 : handle_neon_extension_upgrade(&mut client)
1336 0 : .await
1337 0 : .context("handle_neon_extension_upgrade")?;
1338 0 : Ok::<_, anyhow::Error>(())
1339 0 : }
1340 0 : .await;
1341 0 : if let Err(err) = res {
1342 0 : error!("error while post_apply_config: {err:#}");
1343 0 : }
1344 0 : });
1345 : Ok(())
1346 : }
1347 :
1348 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
1349 0 : let mut conf = self.conn_conf.clone();
1350 0 : if let Some(application_name) = application_name {
1351 0 : conf.application_name(application_name);
1352 0 : }
1353 0 : conf
1354 0 : }
1355 :
1356 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
1357 0 : let mut conf = self.tokio_conn_conf.clone();
1358 0 : if let Some(application_name) = application_name {
1359 0 : conf.application_name(application_name);
1360 0 : }
1361 0 : conf
1362 0 : }
1363 :
1364 0 : pub async fn get_maintenance_client(
1365 0 : conf: &tokio_postgres::Config,
1366 0 : ) -> Result<tokio_postgres::Client> {
1367 0 : let mut conf = conf.clone();
1368 0 : conf.application_name("compute_ctl:apply_config");
1369 :
1370 0 : let (client, conn) = match conf.connect(NoTls).await {
1371 : // If connection fails, it may be the old node with `zenith_admin` superuser.
1372 : //
1373 : // In this case we need to connect with old `zenith_admin` name
1374 : // and create new user. We cannot simply rename connected user,
1375 : // but we can create a new one and grant it all privileges.
1376 0 : Err(e) => match e.code() {
1377 : Some(&SqlState::INVALID_PASSWORD)
1378 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
1379 : // Connect with zenith_admin if cloud_admin could not authenticate
1380 0 : info!(
1381 0 : "cannot connect to postgres: {}, retrying with `zenith_admin` username",
1382 : e
1383 : );
1384 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
1385 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
1386 0 : zenith_admin_conf.user("zenith_admin");
1387 :
1388 0 : let mut client =
1389 0 : zenith_admin_conf.connect(NoTls)
1390 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
1391 :
1392 : // Disable forwarding so that users don't get a cloud_admin role
1393 0 : let mut func = || {
1394 0 : client.simple_query("SET neon.forward_ddl = false")?;
1395 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
1396 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
1397 0 : Ok::<_, anyhow::Error>(())
1398 0 : };
1399 0 : func().context("apply_config setup cloud_admin")?;
1400 :
1401 0 : drop(client);
1402 0 :
1403 0 : // Reconnect with connstring with expected name
1404 0 : conf.connect(NoTls).await?
1405 : }
1406 0 : _ => return Err(e.into()),
1407 : },
1408 0 : Ok((client, conn)) => (client, conn),
1409 : };
1410 :
1411 0 : spawn(async move {
1412 0 : if let Err(e) = conn.await {
1413 0 : error!("maintenance client connection error: {}", e);
1414 0 : }
1415 0 : });
1416 0 :
1417 0 : // Disable DDL forwarding because control plane already knows about the roles/databases
1418 0 : // we're about to modify.
1419 0 : client
1420 0 : .simple_query("SET neon.forward_ddl = false")
1421 0 : .await
1422 0 : .context("apply_config SET neon.forward_ddl = false")?;
1423 :
1424 0 : Ok(client)
1425 0 : }
1426 :
1427 : /// Do initial configuration of the already started Postgres.
1428 : #[instrument(skip_all)]
1429 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1430 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1431 :
1432 : let conf = Arc::new(conf);
1433 : let spec = Arc::new(
1434 : compute_state
1435 : .pspec
1436 : .as_ref()
1437 : .expect("spec must be set")
1438 : .spec
1439 : .clone(),
1440 : );
1441 :
1442 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1443 :
1444 : // Merge-apply spec & changes to PostgreSQL state.
1445 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1446 :
1447 : if let Some(local_proxy) = &spec.clone().local_proxy_config {
1448 : info!("configuring local_proxy");
1449 : local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
1450 : }
1451 :
1452 : // Run migrations separately to not hold up cold starts
1453 0 : tokio::spawn(async move {
1454 0 : let mut conf = conf.as_ref().clone();
1455 0 : conf.application_name("compute_ctl:migrations");
1456 0 :
1457 0 : match conf.connect(NoTls).await {
1458 0 : Ok((mut client, connection)) => {
1459 0 : tokio::spawn(async move {
1460 0 : if let Err(e) = connection.await {
1461 0 : eprintln!("connection error: {}", e);
1462 0 : }
1463 0 : });
1464 0 : if let Err(e) = handle_migrations(&mut client).await {
1465 0 : error!("Failed to run migrations: {}", e);
1466 0 : }
1467 : }
1468 0 : Err(e) => {
1469 0 : error!(
1470 0 : "Failed to connect to the compute for running migrations: {}",
1471 : e
1472 : );
1473 : }
1474 : };
1475 0 : });
1476 :
1477 : Ok::<(), anyhow::Error>(())
1478 : }
1479 :
1480 : // Wrapped this around `pg_ctl reload`, but right now we don't use
1481 : // `pg_ctl` for start / stop.
1482 : #[instrument(skip_all)]
1483 : fn pg_reload_conf(&self) -> Result<()> {
1484 : let pgctl_bin = Path::new(&self.params.pgbin)
1485 : .parent()
1486 : .unwrap()
1487 : .join("pg_ctl");
1488 : Command::new(pgctl_bin)
1489 : .args(["reload", "-D", &self.params.pgdata])
1490 : .output()
1491 : .expect("cannot run pg_ctl process");
1492 : Ok(())
1493 : }
1494 :
1495 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
1496 : /// as it's used to reconfigure a previously started and configured Postgres node.
1497 : #[instrument(skip_all)]
1498 : pub fn reconfigure(&self) -> Result<()> {
1499 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
1500 :
1501 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
1502 : info!("tuning pgbouncer");
1503 :
1504 : let pgbouncer_settings = pgbouncer_settings.clone();
1505 : let tls_config = self.compute_ctl_config.tls.clone();
1506 :
1507 : // Spawn a background task to do the tuning,
1508 : // so that we don't block the main thread that starts Postgres.
1509 0 : tokio::spawn(async move {
1510 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
1511 0 : if let Err(err) = res {
1512 0 : error!("error while tuning pgbouncer: {err:?}");
1513 0 : }
1514 0 : });
1515 : }
1516 :
1517 : if let Some(ref local_proxy) = spec.local_proxy_config {
1518 : info!("configuring local_proxy");
1519 :
1520 : // Spawn a background task to do the configuration,
1521 : // so that we don't block the main thread that starts Postgres.
1522 : let mut local_proxy = local_proxy.clone();
1523 : local_proxy.tls = self.compute_ctl_config.tls.clone();
1524 0 : tokio::spawn(async move {
1525 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
1526 0 : error!("error while configuring local_proxy: {err:?}");
1527 0 : }
1528 0 : });
1529 : }
1530 :
1531 : // Write new config
1532 : let pgdata_path = Path::new(&self.params.pgdata);
1533 : config::write_postgres_conf(
1534 : pgdata_path,
1535 : &spec,
1536 : self.params.internal_http_port,
1537 : &self.compute_ctl_config.tls,
1538 : )?;
1539 :
1540 : if !spec.skip_pg_catalog_updates {
1541 : let max_concurrent_connections = spec.reconfigure_concurrency;
1542 : // Temporarily reset max_cluster_size in config
1543 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
1544 : // creating new extensions, roles, etc.
1545 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1546 0 : self.pg_reload_conf()?;
1547 :
1548 0 : if spec.mode == ComputeMode::Primary {
1549 0 : let mut conf =
1550 0 : tokio_postgres::Config::from_str(self.params.connstr.as_str()).unwrap();
1551 0 : conf.application_name("apply_config");
1552 0 : let conf = Arc::new(conf);
1553 0 :
1554 0 : let spec = Arc::new(spec.clone());
1555 0 :
1556 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
1557 0 : }
1558 :
1559 0 : Ok(())
1560 0 : })?;
1561 : }
1562 :
1563 : self.pg_reload_conf()?;
1564 :
1565 : let unknown_op = "unknown".to_string();
1566 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
1567 : info!(
1568 : "finished reconfiguration of compute node for operation {}",
1569 : op_id
1570 : );
1571 :
1572 : Ok(())
1573 : }
1574 :
1575 : #[instrument(skip_all)]
1576 : pub fn configure_as_primary(&self, compute_state: &ComputeState) -> Result<()> {
1577 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1578 :
1579 : assert!(pspec.spec.mode == ComputeMode::Primary);
1580 : if !pspec.spec.skip_pg_catalog_updates {
1581 : let pgdata_path = Path::new(&self.params.pgdata);
1582 : // temporarily reset max_cluster_size in config
1583 : // to avoid the possibility of hitting the limit, while we are applying config:
1584 : // creating new extensions, roles, etc...
1585 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
1586 0 : self.pg_reload_conf()?;
1587 :
1588 0 : self.apply_config(compute_state)?;
1589 :
1590 0 : Ok(())
1591 0 : })?;
1592 :
1593 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
1594 : if config::line_in_file(
1595 : &postgresql_conf_path,
1596 : "neon.disable_logical_replication_subscribers=false",
1597 : )? {
1598 : info!(
1599 : "updated postgresql.conf to set neon.disable_logical_replication_subscribers=false"
1600 : );
1601 : }
1602 : self.pg_reload_conf()?;
1603 : }
1604 : self.post_apply_config()?;
1605 :
1606 : Ok(())
1607 : }
1608 :
1609 0 : pub async fn watch_cert_for_changes(self: Arc<Self>) {
1610 : // update status on cert renewal
1611 0 : if let Some(tls_config) = &self.compute_ctl_config.tls {
1612 0 : let tls_config = tls_config.clone();
1613 :
1614 : // wait until the cert exists.
1615 0 : let mut cert_watch = watch_cert_for_changes(tls_config.cert_path.clone()).await;
1616 :
1617 0 : tokio::task::spawn_blocking(move || {
1618 0 : let handle = tokio::runtime::Handle::current();
1619 : 'cert_update: loop {
1620 : // let postgres/pgbouncer/local_proxy know the new cert/key exists.
1621 : // we need to wait until it's configurable first.
1622 :
1623 0 : let mut state = self.state.lock().unwrap();
1624 : 'status_update: loop {
1625 0 : match state.status {
1626 : // let's update the state to config pending
1627 : ComputeStatus::ConfigurationPending | ComputeStatus::Running => {
1628 0 : state.set_status(
1629 0 : ComputeStatus::ConfigurationPending,
1630 0 : &self.state_changed,
1631 0 : );
1632 0 : break 'status_update;
1633 : }
1634 :
1635 : // exit loop
1636 : ComputeStatus::Failed
1637 : | ComputeStatus::TerminationPending
1638 0 : | ComputeStatus::Terminated => break 'cert_update,
1639 :
1640 : // wait
1641 : ComputeStatus::Init
1642 : | ComputeStatus::Configuration
1643 0 : | ComputeStatus::Empty => {
1644 0 : state = self.state_changed.wait(state).unwrap();
1645 0 : }
1646 : }
1647 : }
1648 0 : drop(state);
1649 0 :
1650 0 : // wait for a new certificate update
1651 0 : if handle.block_on(cert_watch.changed()).is_err() {
1652 0 : break;
1653 0 : }
1654 : }
1655 0 : });
1656 0 : }
1657 0 : }
1658 :
1659 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
1660 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
1661 0 : let mut state = self.state.lock().unwrap();
1662 0 : // NB: `Some(<DateTime>)` is always greater than `None`.
1663 0 : if last_active > state.last_active {
1664 0 : state.last_active = last_active;
1665 0 : debug!("set the last compute activity time to: {:?}", last_active);
1666 0 : }
1667 0 : }
1668 :
1669 : // Look for core dumps and collect backtraces.
1670 : //
1671 : // EKS worker nodes have following core dump settings:
1672 : // /proc/sys/kernel/core_pattern -> core
1673 : // /proc/sys/kernel/core_uses_pid -> 1
1674 : // ulimit -c -> unlimited
1675 : // which results in core dumps being written to postgres data directory as core.<pid>.
1676 : //
1677 : // Use that as a default location and pattern, except macos where core dumps are written
1678 : // to /cores/ directory by default.
1679 : //
1680 : // With default Linux settings, the core dump file is called just "core", so check for
1681 : // that too.
1682 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
1683 0 : let core_dump_dir = match std::env::consts::OS {
1684 0 : "macos" => Path::new("/cores/"),
1685 0 : _ => Path::new(&self.params.pgdata),
1686 : };
1687 :
1688 : // Collect core dump paths if any
1689 0 : info!("checking for core dumps in {}", core_dump_dir.display());
1690 0 : let files = fs::read_dir(core_dump_dir)?;
1691 0 : let cores = files.filter_map(|entry| {
1692 0 : let entry = entry.ok()?;
1693 :
1694 0 : let is_core_dump = match entry.file_name().to_str()? {
1695 0 : n if n.starts_with("core.") => true,
1696 0 : "core" => true,
1697 0 : _ => false,
1698 : };
1699 0 : if is_core_dump {
1700 0 : Some(entry.path())
1701 : } else {
1702 0 : None
1703 : }
1704 0 : });
1705 :
1706 : // Print backtrace for each core dump
1707 0 : for core_path in cores {
1708 0 : warn!(
1709 0 : "core dump found: {}, collecting backtrace",
1710 0 : core_path.display()
1711 : );
1712 :
1713 : // Try first with gdb
1714 0 : let backtrace = Command::new("gdb")
1715 0 : .args(["--batch", "-q", "-ex", "bt", &self.params.pgbin])
1716 0 : .arg(&core_path)
1717 0 : .output();
1718 :
1719 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
1720 0 : let backtrace = match backtrace {
1721 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
1722 0 : warn!("cannot find gdb, trying lldb");
1723 0 : Command::new("lldb")
1724 0 : .arg("-c")
1725 0 : .arg(&core_path)
1726 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
1727 0 : .output()
1728 : }
1729 0 : _ => backtrace,
1730 0 : }?;
1731 :
1732 0 : warn!(
1733 0 : "core dump backtrace: {}",
1734 0 : String::from_utf8_lossy(&backtrace.stdout)
1735 : );
1736 0 : warn!(
1737 0 : "debugger stderr: {}",
1738 0 : String::from_utf8_lossy(&backtrace.stderr)
1739 : );
1740 : }
1741 :
1742 0 : Ok(())
1743 0 : }
1744 :
1745 : /// Select `pg_stat_statements` data and return it as a stringified JSON
1746 0 : pub async fn collect_insights(&self) -> String {
1747 0 : let mut result_rows: Vec<String> = Vec::new();
1748 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
1749 0 : let connect_result = conf.connect(NoTls).await;
1750 0 : let (client, connection) = connect_result.unwrap();
1751 0 : tokio::spawn(async move {
1752 0 : if let Err(e) = connection.await {
1753 0 : eprintln!("connection error: {}", e);
1754 0 : }
1755 0 : });
1756 0 : let result = client
1757 0 : .simple_query(
1758 0 : "SELECT
1759 0 : row_to_json(pg_stat_statements)
1760 0 : FROM
1761 0 : pg_stat_statements
1762 0 : WHERE
1763 0 : userid != 'cloud_admin'::regrole::oid
1764 0 : ORDER BY
1765 0 : (mean_exec_time + mean_plan_time) DESC
1766 0 : LIMIT 100",
1767 0 : )
1768 0 : .await;
1769 :
1770 0 : if let Ok(raw_rows) = result {
1771 0 : for message in raw_rows.iter() {
1772 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
1773 0 : if let Some(json) = row.get(0) {
1774 0 : result_rows.push(json.to_string());
1775 0 : }
1776 0 : }
1777 : }
1778 :
1779 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
1780 : } else {
1781 0 : "{{\"pg_stat_statements\": []}}".to_string()
1782 : }
1783 0 : }
1784 :
1785 : // download an archive, unzip and place files in correct locations
1786 0 : pub async fn download_extension(
1787 0 : &self,
1788 0 : real_ext_name: String,
1789 0 : ext_path: RemotePath,
1790 0 : ) -> Result<u64, DownloadError> {
1791 0 : let ext_remote_storage =
1792 0 : self.params
1793 0 : .ext_remote_storage
1794 0 : .as_ref()
1795 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
1796 0 : "Remote extensions storage is not configured",
1797 0 : )))?;
1798 :
1799 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
1800 0 :
1801 0 : let mut first_try = false;
1802 0 : if !self
1803 0 : .ext_download_progress
1804 0 : .read()
1805 0 : .expect("lock err")
1806 0 : .contains_key(ext_archive_name)
1807 0 : {
1808 0 : self.ext_download_progress
1809 0 : .write()
1810 0 : .expect("lock err")
1811 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
1812 0 : first_try = true;
1813 0 : }
1814 0 : let (download_start, download_completed) =
1815 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
1816 0 : let start_time_delta = Utc::now()
1817 0 : .signed_duration_since(download_start)
1818 0 : .to_std()
1819 0 : .unwrap()
1820 0 : .as_millis() as u64;
1821 :
1822 : // how long to wait for extension download if it was started by another process
1823 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
1824 :
1825 0 : if download_completed {
1826 0 : info!("extension already downloaded, skipping re-download");
1827 0 : return Ok(0);
1828 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
1829 0 : info!(
1830 0 : "download {ext_archive_name} already started by another process, hanging untill completion or timeout"
1831 : );
1832 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
1833 : loop {
1834 0 : info!("waiting for download");
1835 0 : interval.tick().await;
1836 0 : let (_, download_completed_now) =
1837 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
1838 0 : if download_completed_now {
1839 0 : info!("download finished by whoever else downloaded it");
1840 0 : return Ok(0);
1841 0 : }
1842 : }
1843 : // NOTE: the above loop will get terminated
1844 : // based on the timeout of the download function
1845 0 : }
1846 0 :
1847 0 : // if extension hasn't been downloaded before or the previous
1848 0 : // attempt to download was at least HANG_TIMEOUT ms ago
1849 0 : // then we try to download it here
1850 0 : info!("downloading new extension {ext_archive_name}");
1851 :
1852 0 : let download_size = extension_server::download_extension(
1853 0 : &real_ext_name,
1854 0 : &ext_path,
1855 0 : ext_remote_storage,
1856 0 : &self.params.pgbin,
1857 0 : )
1858 0 : .await
1859 0 : .map_err(DownloadError::Other);
1860 0 :
1861 0 : if download_size.is_ok() {
1862 0 : self.ext_download_progress
1863 0 : .write()
1864 0 : .expect("bad lock")
1865 0 : .insert(ext_archive_name.to_string(), (download_start, true));
1866 0 : }
1867 :
1868 0 : download_size
1869 0 : }
1870 :
1871 0 : pub async fn set_role_grants(
1872 0 : &self,
1873 0 : db_name: &PgIdent,
1874 0 : schema_name: &PgIdent,
1875 0 : privileges: &[Privilege],
1876 0 : role_name: &PgIdent,
1877 0 : ) -> Result<()> {
1878 : use tokio_postgres::NoTls;
1879 :
1880 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
1881 0 : conf.dbname(db_name);
1882 :
1883 0 : let (db_client, conn) = conf
1884 0 : .connect(NoTls)
1885 0 : .await
1886 0 : .context("Failed to connect to the database")?;
1887 0 : tokio::spawn(conn);
1888 0 :
1889 0 : // TODO: support other types of grants apart from schemas?
1890 0 : let query = format!(
1891 0 : "GRANT {} ON SCHEMA {} TO {}",
1892 0 : privileges
1893 0 : .iter()
1894 0 : // should not be quoted as it's part of the command.
1895 0 : // is already sanitized so it's ok
1896 0 : .map(|p| p.as_str())
1897 0 : .collect::<Vec<&'static str>>()
1898 0 : .join(", "),
1899 0 : // quote the schema and role name as identifiers to sanitize them.
1900 0 : schema_name.pg_quote(),
1901 0 : role_name.pg_quote(),
1902 0 : );
1903 0 : db_client
1904 0 : .simple_query(&query)
1905 0 : .await
1906 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1907 :
1908 0 : Ok(())
1909 0 : }
1910 :
1911 0 : pub async fn install_extension(
1912 0 : &self,
1913 0 : ext_name: &PgIdent,
1914 0 : db_name: &PgIdent,
1915 0 : ext_version: ExtVersion,
1916 0 : ) -> Result<ExtVersion> {
1917 : use tokio_postgres::NoTls;
1918 :
1919 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
1920 0 : conf.dbname(db_name);
1921 :
1922 0 : let (db_client, conn) = conf
1923 0 : .connect(NoTls)
1924 0 : .await
1925 0 : .context("Failed to connect to the database")?;
1926 0 : tokio::spawn(conn);
1927 0 :
1928 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
1929 0 : let version: Option<ExtVersion> = db_client
1930 0 : .query_opt(version_query, &[&ext_name])
1931 0 : .await
1932 0 : .with_context(|| format!("Failed to execute query: {}", version_query))?
1933 0 : .map(|row| row.get(0));
1934 0 :
1935 0 : // sanitize the inputs as postgres idents.
1936 0 : let ext_name: String = ext_name.pg_quote();
1937 0 : let quoted_version: String = ext_version.pg_quote();
1938 :
1939 0 : if let Some(installed_version) = version {
1940 0 : if installed_version == ext_version {
1941 0 : return Ok(installed_version);
1942 0 : }
1943 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
1944 0 : db_client
1945 0 : .simple_query(&query)
1946 0 : .await
1947 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1948 : } else {
1949 0 : let query =
1950 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
1951 0 : db_client
1952 0 : .simple_query(&query)
1953 0 : .await
1954 0 : .with_context(|| format!("Failed to execute query: {}", query))?;
1955 : }
1956 :
1957 0 : Ok(ext_version)
1958 0 : }
1959 :
1960 0 : pub async fn prepare_preload_libraries(
1961 0 : &self,
1962 0 : spec: &ComputeSpec,
1963 0 : ) -> Result<RemoteExtensionMetrics> {
1964 0 : if self.params.ext_remote_storage.is_none() {
1965 0 : return Ok(RemoteExtensionMetrics {
1966 0 : num_ext_downloaded: 0,
1967 0 : largest_ext_size: 0,
1968 0 : total_ext_download_size: 0,
1969 0 : });
1970 0 : }
1971 0 : let remote_extensions = spec
1972 0 : .remote_extensions
1973 0 : .as_ref()
1974 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
1975 :
1976 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
1977 0 : let mut libs_vec = Vec::new();
1978 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
1979 0 : libs_vec = libs
1980 0 : .split(&[',', '\'', ' '])
1981 0 : .filter(|s| *s != "neon" && !s.is_empty())
1982 0 : .map(str::to_string)
1983 0 : .collect();
1984 0 : }
1985 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
1986 :
1987 : // that is used in neon_local and python tests
1988 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
1989 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
1990 0 : let mut shared_preload_libraries_line = "";
1991 0 : for line in conf_lines {
1992 0 : if line.starts_with("shared_preload_libraries") {
1993 0 : shared_preload_libraries_line = line;
1994 0 : }
1995 : }
1996 0 : let mut preload_libs_vec = Vec::new();
1997 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
1998 0 : preload_libs_vec = libs
1999 0 : .split(&[',', '\'', ' '])
2000 0 : .filter(|s| *s != "neon" && !s.is_empty())
2001 0 : .map(str::to_string)
2002 0 : .collect();
2003 0 : }
2004 0 : libs_vec.extend(preload_libs_vec);
2005 0 : }
2006 :
2007 : // Don't try to download libraries that are not in the index.
2008 : // Assume that they are already present locally.
2009 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
2010 0 :
2011 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
2012 :
2013 0 : let mut download_tasks = Vec::new();
2014 0 : for library in &libs_vec {
2015 0 : let (ext_name, ext_path) = remote_extensions.get_ext(
2016 0 : library,
2017 0 : true,
2018 0 : &self.params.build_tag,
2019 0 : &self.params.pgversion,
2020 0 : )?;
2021 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
2022 : }
2023 0 : let results = join_all(download_tasks).await;
2024 :
2025 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
2026 0 : num_ext_downloaded: 0,
2027 0 : largest_ext_size: 0,
2028 0 : total_ext_download_size: 0,
2029 0 : };
2030 0 : for result in results {
2031 0 : let download_size = match result {
2032 0 : Ok(res) => {
2033 0 : remote_ext_metrics.num_ext_downloaded += 1;
2034 0 : res
2035 : }
2036 0 : Err(err) => {
2037 0 : // if we failed to download an extension, we don't want to fail the whole
2038 0 : // process, but we do want to log the error
2039 0 : error!("Failed to download extension: {}", err);
2040 0 : 0
2041 : }
2042 : };
2043 :
2044 0 : remote_ext_metrics.largest_ext_size =
2045 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
2046 0 : remote_ext_metrics.total_ext_download_size += download_size;
2047 : }
2048 0 : Ok(remote_ext_metrics)
2049 0 : }
2050 :
2051 : /// Waits until current thread receives a state changed notification and
2052 : /// the pageserver connection strings has changed.
2053 : ///
2054 : /// The operation will time out after a specified duration.
2055 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
2056 0 : let state = self.state.lock().unwrap();
2057 0 : let old_pageserver_connstr = state
2058 0 : .pspec
2059 0 : .as_ref()
2060 0 : .expect("spec must be set")
2061 0 : .pageserver_connstr
2062 0 : .clone();
2063 0 : let mut unchanged = true;
2064 0 : let _ = self
2065 0 : .state_changed
2066 0 : .wait_timeout_while(state, duration, |s| {
2067 0 : let pageserver_connstr = &s
2068 0 : .pspec
2069 0 : .as_ref()
2070 0 : .expect("spec must be set")
2071 0 : .pageserver_connstr;
2072 0 : unchanged = pageserver_connstr == &old_pageserver_connstr;
2073 0 : unchanged
2074 0 : })
2075 0 : .unwrap();
2076 0 : if !unchanged {
2077 0 : info!("Pageserver config changed");
2078 0 : }
2079 0 : }
2080 : }
2081 :
2082 0 : pub fn forward_termination_signal() {
2083 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2084 0 : if ss_pid != 0 {
2085 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2086 0 : kill(ss_pid, Signal::SIGTERM).ok();
2087 0 : }
2088 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2089 0 : if pg_pid != 0 {
2090 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2091 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2092 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2093 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2094 0 : kill(pg_pid, Signal::SIGINT).ok();
2095 0 : }
2096 0 : }
2097 :
2098 : // helper trait to call JoinSet::spawn_blocking(f), but propagates the current
2099 : // tracing span to the thread.
2100 : trait JoinSetExt<T> {
2101 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2102 : where
2103 : F: FnOnce() -> T + Send + 'static,
2104 : T: Send;
2105 : }
2106 :
2107 : impl<T: 'static> JoinSetExt<T> for tokio::task::JoinSet<T> {
2108 0 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2109 0 : where
2110 0 : F: FnOnce() -> T + Send + 'static,
2111 0 : T: Send,
2112 0 : {
2113 0 : let sp = tracing::Span::current();
2114 0 : self.spawn_blocking(move || {
2115 0 : let _e = sp.enter();
2116 0 : f()
2117 0 : })
2118 0 : }
2119 : }
|
