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