Line data Source code
1 : //! The timeline manager task is responsible for managing the timeline's background tasks.
2 : //!
3 : //! It is spawned alongside each timeline and exits when the timeline is deleted.
4 : //! It watches for changes in the timeline state and decides when to spawn or kill background tasks.
5 : //! It also can manage some reactive state, like should the timeline be active for broker pushes or not.
6 : //!
7 : //! Be aware that you need to be extra careful with manager code, because it is not respawned on panic.
8 : //! Also, if it will stuck in some branch, it will prevent any further progress in the timeline.
9 :
10 : use std::{
11 : sync::{atomic::AtomicUsize, Arc},
12 : time::Duration,
13 : };
14 :
15 : use futures::channel::oneshot;
16 : use postgres_ffi::XLogSegNo;
17 : use serde::{Deserialize, Serialize};
18 : use tokio::{
19 : task::{JoinError, JoinHandle},
20 : time::Instant,
21 : };
22 : use tokio_util::sync::CancellationToken;
23 : use tracing::{debug, info, info_span, instrument, warn, Instrument};
24 : use utils::lsn::Lsn;
25 :
26 : use crate::{
27 : control_file::{FileStorage, Storage},
28 : metrics::{
29 : MANAGER_ACTIVE_CHANGES, MANAGER_ITERATIONS_TOTAL, MISC_OPERATION_SECONDS,
30 : NUM_EVICTED_TIMELINES,
31 : },
32 : rate_limit::{rand_duration, RateLimiter},
33 : recovery::recovery_main,
34 : remove_wal::calc_horizon_lsn,
35 : safekeeper::Term,
36 : send_wal::WalSenders,
37 : state::TimelineState,
38 : timeline::{ManagerTimeline, PeerInfo, ReadGuardSharedState, StateSK, WalResidentTimeline},
39 : timeline_guard::{AccessService, GuardId, ResidenceGuard},
40 : timelines_set::{TimelineSetGuard, TimelinesSet},
41 : wal_backup::{self, WalBackupTaskHandle},
42 : wal_backup_partial::{self, PartialBackup, PartialRemoteSegment},
43 : SafeKeeperConf,
44 : };
45 :
46 : pub(crate) struct StateSnapshot {
47 : // inmem values
48 : pub(crate) commit_lsn: Lsn,
49 : pub(crate) backup_lsn: Lsn,
50 : pub(crate) remote_consistent_lsn: Lsn,
51 :
52 : // persistent control file values
53 : pub(crate) cfile_commit_lsn: Lsn,
54 : pub(crate) cfile_remote_consistent_lsn: Lsn,
55 : pub(crate) cfile_backup_lsn: Lsn,
56 :
57 : // latest state
58 : pub(crate) flush_lsn: Lsn,
59 : pub(crate) last_log_term: Term,
60 :
61 : // misc
62 : pub(crate) cfile_last_persist_at: std::time::Instant,
63 : pub(crate) inmem_flush_pending: bool,
64 : pub(crate) wal_removal_on_hold: bool,
65 : pub(crate) peers: Vec<PeerInfo>,
66 : }
67 :
68 : impl StateSnapshot {
69 : /// Create a new snapshot of the timeline state.
70 0 : fn new(read_guard: ReadGuardSharedState, heartbeat_timeout: Duration) -> Self {
71 0 : let state = read_guard.sk.state();
72 0 : Self {
73 0 : commit_lsn: state.inmem.commit_lsn,
74 0 : backup_lsn: state.inmem.backup_lsn,
75 0 : remote_consistent_lsn: state.inmem.remote_consistent_lsn,
76 0 : cfile_commit_lsn: state.commit_lsn,
77 0 : cfile_remote_consistent_lsn: state.remote_consistent_lsn,
78 0 : cfile_backup_lsn: state.backup_lsn,
79 0 : flush_lsn: read_guard.sk.flush_lsn(),
80 0 : last_log_term: read_guard.sk.last_log_term(),
81 0 : cfile_last_persist_at: state.pers.last_persist_at(),
82 0 : inmem_flush_pending: Self::has_unflushed_inmem_state(state),
83 0 : wal_removal_on_hold: read_guard.wal_removal_on_hold,
84 0 : peers: read_guard.get_peers(heartbeat_timeout),
85 0 : }
86 0 : }
87 :
88 0 : fn has_unflushed_inmem_state(state: &TimelineState<FileStorage>) -> bool {
89 0 : state.inmem.commit_lsn > state.commit_lsn
90 0 : || state.inmem.backup_lsn > state.backup_lsn
91 0 : || state.inmem.peer_horizon_lsn > state.peer_horizon_lsn
92 0 : || state.inmem.remote_consistent_lsn > state.remote_consistent_lsn
93 0 : }
94 : }
95 :
96 : /// Control how often the manager task should wake up to check updates.
97 : /// There is no need to check for updates more often than this.
98 : const REFRESH_INTERVAL: Duration = Duration::from_millis(300);
99 :
100 : pub enum ManagerCtlMessage {
101 : /// Request to get a guard for WalResidentTimeline, with WAL files available locally.
102 : GuardRequest(tokio::sync::oneshot::Sender<anyhow::Result<ResidenceGuard>>),
103 : /// Get a guard for WalResidentTimeline if the timeline is not currently offloaded, else None
104 : TryGuardRequest(tokio::sync::oneshot::Sender<Option<ResidenceGuard>>),
105 : /// Request to drop the guard.
106 : GuardDrop(GuardId),
107 : /// Request to reset uploaded partial backup state.
108 : BackupPartialReset(oneshot::Sender<anyhow::Result<Vec<String>>>),
109 : }
110 :
111 : impl std::fmt::Debug for ManagerCtlMessage {
112 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
113 0 : match self {
114 0 : ManagerCtlMessage::GuardRequest(_) => write!(f, "GuardRequest"),
115 0 : ManagerCtlMessage::TryGuardRequest(_) => write!(f, "TryGuardRequest"),
116 0 : ManagerCtlMessage::GuardDrop(id) => write!(f, "GuardDrop({:?})", id),
117 0 : ManagerCtlMessage::BackupPartialReset(_) => write!(f, "BackupPartialReset"),
118 : }
119 0 : }
120 : }
121 :
122 : pub struct ManagerCtl {
123 : manager_tx: tokio::sync::mpsc::UnboundedSender<ManagerCtlMessage>,
124 :
125 : // this is used to initialize manager, it will be moved out in bootstrap().
126 : init_manager_rx:
127 : std::sync::Mutex<Option<tokio::sync::mpsc::UnboundedReceiver<ManagerCtlMessage>>>,
128 : }
129 :
130 : impl Default for ManagerCtl {
131 0 : fn default() -> Self {
132 0 : Self::new()
133 0 : }
134 : }
135 :
136 : impl ManagerCtl {
137 0 : pub fn new() -> Self {
138 0 : let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
139 0 : Self {
140 0 : manager_tx: tx,
141 0 : init_manager_rx: std::sync::Mutex::new(Some(rx)),
142 0 : }
143 0 : }
144 :
145 : /// Issue a new guard and wait for manager to prepare the timeline.
146 : /// Sends a message to the manager and waits for the response.
147 : /// Can be blocked indefinitely if the manager is stuck.
148 0 : pub async fn wal_residence_guard(&self) -> anyhow::Result<ResidenceGuard> {
149 0 : let (tx, rx) = tokio::sync::oneshot::channel();
150 0 : self.manager_tx.send(ManagerCtlMessage::GuardRequest(tx))?;
151 :
152 : // wait for the manager to respond with the guard
153 0 : rx.await
154 0 : .map_err(|e| anyhow::anyhow!("response read fail: {:?}", e))
155 0 : .and_then(std::convert::identity)
156 0 : }
157 :
158 : /// Issue a new guard if the timeline is currently not offloaded, else return None
159 : /// Sends a message to the manager and waits for the response.
160 : /// Can be blocked indefinitely if the manager is stuck.
161 0 : pub async fn try_wal_residence_guard(&self) -> anyhow::Result<Option<ResidenceGuard>> {
162 0 : let (tx, rx) = tokio::sync::oneshot::channel();
163 0 : self.manager_tx
164 0 : .send(ManagerCtlMessage::TryGuardRequest(tx))?;
165 :
166 : // wait for the manager to respond with the guard
167 0 : rx.await
168 0 : .map_err(|e| anyhow::anyhow!("response read fail: {:?}", e))
169 0 : }
170 :
171 : /// Request timeline manager to reset uploaded partial segment state and
172 : /// wait for the result.
173 0 : pub async fn backup_partial_reset(&self) -> anyhow::Result<Vec<String>> {
174 0 : let (tx, rx) = oneshot::channel();
175 0 : self.manager_tx
176 0 : .send(ManagerCtlMessage::BackupPartialReset(tx))
177 0 : .expect("manager task is not running");
178 0 : match rx.await {
179 0 : Ok(res) => res,
180 0 : Err(_) => anyhow::bail!("timeline manager is gone"),
181 : }
182 0 : }
183 :
184 : /// Must be called exactly once to bootstrap the manager.
185 0 : pub fn bootstrap_manager(
186 0 : &self,
187 0 : ) -> (
188 0 : tokio::sync::mpsc::UnboundedSender<ManagerCtlMessage>,
189 0 : tokio::sync::mpsc::UnboundedReceiver<ManagerCtlMessage>,
190 0 : ) {
191 0 : let rx = self
192 0 : .init_manager_rx
193 0 : .lock()
194 0 : .expect("mutex init_manager_rx poisoned")
195 0 : .take()
196 0 : .expect("manager already bootstrapped");
197 0 :
198 0 : (self.manager_tx.clone(), rx)
199 0 : }
200 : }
201 :
202 : pub(crate) struct Manager {
203 : // configuration & dependencies
204 : pub(crate) tli: ManagerTimeline,
205 : pub(crate) conf: SafeKeeperConf,
206 : pub(crate) wal_seg_size: usize,
207 : pub(crate) walsenders: Arc<WalSenders>,
208 :
209 : // current state
210 : pub(crate) state_version_rx: tokio::sync::watch::Receiver<usize>,
211 : pub(crate) num_computes_rx: tokio::sync::watch::Receiver<usize>,
212 : pub(crate) tli_broker_active: TimelineSetGuard,
213 : pub(crate) last_removed_segno: XLogSegNo,
214 : pub(crate) is_offloaded: bool,
215 :
216 : // background tasks
217 : pub(crate) backup_task: Option<WalBackupTaskHandle>,
218 : pub(crate) recovery_task: Option<JoinHandle<()>>,
219 : pub(crate) wal_removal_task: Option<JoinHandle<anyhow::Result<u64>>>,
220 :
221 : // partial backup
222 : pub(crate) partial_backup_task:
223 : Option<(JoinHandle<Option<PartialRemoteSegment>>, CancellationToken)>,
224 : pub(crate) partial_backup_uploaded: Option<PartialRemoteSegment>,
225 :
226 : // misc
227 : pub(crate) access_service: AccessService,
228 : pub(crate) global_rate_limiter: RateLimiter,
229 :
230 : // Anti-flapping state: we evict timelines eagerly if they are inactive, but should not
231 : // evict them if they go inactive very soon after being restored.
232 : pub(crate) evict_not_before: Instant,
233 : }
234 :
235 : /// This task gets spawned alongside each timeline and is responsible for managing the timeline's
236 : /// background tasks.
237 : /// Be careful, this task is not respawned on panic, so it should not panic.
238 0 : #[instrument(name = "manager", skip_all, fields(ttid = %tli.ttid))]
239 : pub async fn main_task(
240 : tli: ManagerTimeline,
241 : conf: SafeKeeperConf,
242 : broker_active_set: Arc<TimelinesSet>,
243 : manager_tx: tokio::sync::mpsc::UnboundedSender<ManagerCtlMessage>,
244 : mut manager_rx: tokio::sync::mpsc::UnboundedReceiver<ManagerCtlMessage>,
245 : global_rate_limiter: RateLimiter,
246 : ) {
247 : tli.set_status(Status::Started);
248 :
249 : let defer_tli = tli.tli.clone();
250 : scopeguard::defer! {
251 : if defer_tli.is_cancelled() {
252 : info!("manager task finished");
253 : } else {
254 : warn!("manager task finished prematurely");
255 : }
256 : };
257 :
258 : let mut mgr = Manager::new(
259 : tli,
260 : conf,
261 : broker_active_set,
262 : manager_tx,
263 : global_rate_limiter,
264 : )
265 : .await;
266 :
267 : // Start recovery task which always runs on the timeline.
268 : if !mgr.is_offloaded && mgr.conf.peer_recovery_enabled {
269 : // Recovery task is only spawned if we can get a residence guard (i.e. timeline is not already shutting down)
270 : if let Ok(tli) = mgr.wal_resident_timeline() {
271 : mgr.recovery_task = Some(tokio::spawn(recovery_main(tli, mgr.conf.clone())));
272 : }
273 : }
274 :
275 : // If timeline is evicted, reflect that in the metric.
276 : if mgr.is_offloaded {
277 : NUM_EVICTED_TIMELINES.inc();
278 : }
279 :
280 : let last_state = 'outer: loop {
281 : MANAGER_ITERATIONS_TOTAL.inc();
282 :
283 : mgr.set_status(Status::StateSnapshot);
284 : let state_snapshot = mgr.state_snapshot().await;
285 :
286 : let mut next_event: Option<Instant> = None;
287 : if !mgr.is_offloaded {
288 : let num_computes = *mgr.num_computes_rx.borrow();
289 :
290 : mgr.set_status(Status::UpdateBackup);
291 : let is_wal_backup_required = mgr.update_backup(num_computes, &state_snapshot).await;
292 : mgr.update_is_active(is_wal_backup_required, num_computes, &state_snapshot);
293 :
294 : mgr.set_status(Status::UpdateControlFile);
295 : mgr.update_control_file_save(&state_snapshot, &mut next_event)
296 : .await;
297 :
298 : mgr.set_status(Status::UpdateWalRemoval);
299 : mgr.update_wal_removal(&state_snapshot).await;
300 :
301 : mgr.set_status(Status::UpdatePartialBackup);
302 : mgr.update_partial_backup(&state_snapshot).await;
303 :
304 : let now = Instant::now();
305 : if mgr.evict_not_before > now {
306 : // we should wait until evict_not_before
307 : update_next_event(&mut next_event, mgr.evict_not_before);
308 : }
309 :
310 : if mgr.conf.enable_offload
311 : && mgr.evict_not_before <= now
312 : && mgr.ready_for_eviction(&next_event, &state_snapshot)
313 : {
314 : // check rate limiter and evict timeline if possible
315 : match mgr.global_rate_limiter.try_acquire_eviction() {
316 : Some(_permit) => {
317 : mgr.set_status(Status::EvictTimeline);
318 : if !mgr.evict_timeline().await {
319 : // eviction failed, try again later
320 : mgr.evict_not_before =
321 : Instant::now() + rand_duration(&mgr.conf.eviction_min_resident);
322 : update_next_event(&mut next_event, mgr.evict_not_before);
323 : }
324 : }
325 : None => {
326 : // we can't evict timeline now, will try again later
327 : mgr.evict_not_before =
328 : Instant::now() + rand_duration(&mgr.conf.eviction_min_resident);
329 : update_next_event(&mut next_event, mgr.evict_not_before);
330 : }
331 : }
332 : }
333 : }
334 :
335 : mgr.set_status(Status::Wait);
336 : // wait until something changes. tx channels are stored under Arc, so they will not be
337 : // dropped until the manager task is finished.
338 : tokio::select! {
339 : _ = mgr.tli.cancel.cancelled() => {
340 : // timeline was deleted
341 : break 'outer state_snapshot;
342 : }
343 0 : _ = async {
344 0 : // don't wake up on every state change, but at most every REFRESH_INTERVAL
345 0 : tokio::time::sleep(REFRESH_INTERVAL).await;
346 0 : let _ = mgr.state_version_rx.changed().await;
347 0 : } => {
348 : // state was updated
349 : }
350 : _ = mgr.num_computes_rx.changed() => {
351 : // number of connected computes was updated
352 : }
353 : _ = sleep_until(&next_event) => {
354 : // we were waiting for some event (e.g. cfile save)
355 : }
356 : res = await_task_finish(mgr.wal_removal_task.as_mut()) => {
357 : // WAL removal task finished
358 : mgr.wal_removal_task = None;
359 : mgr.update_wal_removal_end(res);
360 : }
361 0 : res = await_task_finish(mgr.partial_backup_task.as_mut().map(|(handle, _)| handle)) => {
362 : // partial backup task finished
363 : mgr.partial_backup_task = None;
364 : mgr.update_partial_backup_end(res);
365 : }
366 :
367 : msg = manager_rx.recv() => {
368 : mgr.set_status(Status::HandleMessage);
369 : mgr.handle_message(msg).await;
370 : }
371 : }
372 : };
373 : mgr.set_status(Status::Exiting);
374 :
375 : // remove timeline from the broker active set sooner, before waiting for background tasks
376 : mgr.tli_broker_active.set(false);
377 :
378 : // shutdown background tasks
379 : if mgr.conf.is_wal_backup_enabled() {
380 : if let Some(backup_task) = mgr.backup_task.take() {
381 : // If we fell through here, then the timeline is shutting down. This is important
382 : // because otherwise joining on the wal_backup handle might hang.
383 : assert!(mgr.tli.cancel.is_cancelled());
384 :
385 : backup_task.join().await;
386 : }
387 : wal_backup::update_task(&mut mgr, false, &last_state).await;
388 : }
389 :
390 : if let Some(recovery_task) = &mut mgr.recovery_task {
391 : if let Err(e) = recovery_task.await {
392 : warn!("recovery task failed: {:?}", e);
393 : }
394 : }
395 :
396 : if let Some((handle, cancel)) = &mut mgr.partial_backup_task {
397 : cancel.cancel();
398 : if let Err(e) = handle.await {
399 : warn!("partial backup task failed: {:?}", e);
400 : }
401 : }
402 :
403 : if let Some(wal_removal_task) = &mut mgr.wal_removal_task {
404 : let res = wal_removal_task.await;
405 : mgr.update_wal_removal_end(res);
406 : }
407 :
408 : // If timeline is deleted while evicted decrement the gauge.
409 : if mgr.tli.is_cancelled() && mgr.is_offloaded {
410 : NUM_EVICTED_TIMELINES.dec();
411 : }
412 :
413 : mgr.set_status(Status::Finished);
414 : }
415 :
416 : impl Manager {
417 0 : async fn new(
418 0 : tli: ManagerTimeline,
419 0 : conf: SafeKeeperConf,
420 0 : broker_active_set: Arc<TimelinesSet>,
421 0 : manager_tx: tokio::sync::mpsc::UnboundedSender<ManagerCtlMessage>,
422 0 : global_rate_limiter: RateLimiter,
423 0 : ) -> Manager {
424 0 : let (is_offloaded, partial_backup_uploaded) = tli.bootstrap_mgr().await;
425 : Manager {
426 0 : wal_seg_size: tli.get_wal_seg_size().await,
427 0 : walsenders: tli.get_walsenders().clone(),
428 0 : state_version_rx: tli.get_state_version_rx(),
429 0 : num_computes_rx: tli.get_walreceivers().get_num_rx(),
430 0 : tli_broker_active: broker_active_set.guard(tli.clone()),
431 0 : last_removed_segno: 0,
432 0 : is_offloaded,
433 0 : backup_task: None,
434 0 : recovery_task: None,
435 0 : wal_removal_task: None,
436 0 : partial_backup_task: None,
437 0 : partial_backup_uploaded,
438 0 : access_service: AccessService::new(manager_tx),
439 0 : tli,
440 0 : global_rate_limiter,
441 0 : // to smooth out evictions spike after restart
442 0 : evict_not_before: Instant::now() + rand_duration(&conf.eviction_min_resident),
443 0 : conf,
444 0 : }
445 0 : }
446 :
447 0 : fn set_status(&self, status: Status) {
448 0 : self.tli.set_status(status);
449 0 : }
450 :
451 : /// Get a WalResidentTimeline.
452 : /// Manager code must use this function instead of one from `Timeline`
453 : /// directly, because it will deadlock.
454 : ///
455 : /// This function is fallible because the guard may not be created if the timeline is
456 : /// shutting down.
457 0 : pub(crate) fn wal_resident_timeline(&mut self) -> anyhow::Result<WalResidentTimeline> {
458 0 : assert!(!self.is_offloaded);
459 0 : let guard = self.access_service.create_guard(
460 0 : self.tli
461 0 : .gate
462 0 : .enter()
463 0 : .map_err(|_| anyhow::anyhow!("Timeline shutting down"))?,
464 : );
465 0 : Ok(WalResidentTimeline::new(self.tli.clone(), guard))
466 0 : }
467 :
468 : /// Get a snapshot of the timeline state.
469 0 : async fn state_snapshot(&self) -> StateSnapshot {
470 0 : let _timer = MISC_OPERATION_SECONDS
471 0 : .with_label_values(&["state_snapshot"])
472 0 : .start_timer();
473 0 :
474 0 : StateSnapshot::new(
475 0 : self.tli.read_shared_state().await,
476 0 : self.conf.heartbeat_timeout,
477 0 : )
478 0 : }
479 :
480 : /// Spawns/kills backup task and returns true if backup is required.
481 0 : async fn update_backup(&mut self, num_computes: usize, state: &StateSnapshot) -> bool {
482 0 : let is_wal_backup_required =
483 0 : wal_backup::is_wal_backup_required(self.wal_seg_size, num_computes, state);
484 0 :
485 0 : if self.conf.is_wal_backup_enabled() {
486 0 : wal_backup::update_task(self, is_wal_backup_required, state).await;
487 0 : }
488 :
489 : // update the state in Arc<Timeline>
490 0 : self.tli.wal_backup_active.store(
491 0 : self.backup_task.is_some(),
492 0 : std::sync::atomic::Ordering::Relaxed,
493 0 : );
494 0 : is_wal_backup_required
495 0 : }
496 :
497 : /// Update is_active flag and returns its value.
498 0 : fn update_is_active(
499 0 : &mut self,
500 0 : is_wal_backup_required: bool,
501 0 : num_computes: usize,
502 0 : state: &StateSnapshot,
503 0 : ) {
504 0 : let is_active = is_wal_backup_required
505 0 : || num_computes > 0
506 0 : || state.remote_consistent_lsn < state.commit_lsn;
507 :
508 : // update the broker timeline set
509 0 : if self.tli_broker_active.set(is_active) {
510 : // write log if state has changed
511 0 : info!(
512 0 : "timeline active={} now, remote_consistent_lsn={}, commit_lsn={}",
513 : is_active, state.remote_consistent_lsn, state.commit_lsn,
514 : );
515 :
516 0 : MANAGER_ACTIVE_CHANGES.inc();
517 0 : }
518 :
519 : // update the state in Arc<Timeline>
520 0 : self.tli
521 0 : .broker_active
522 0 : .store(is_active, std::sync::atomic::Ordering::Relaxed);
523 0 : }
524 :
525 : /// Save control file if needed. Returns Instant if we should persist the control file in the future.
526 0 : async fn update_control_file_save(
527 0 : &self,
528 0 : state: &StateSnapshot,
529 0 : next_event: &mut Option<Instant>,
530 0 : ) {
531 0 : if !state.inmem_flush_pending {
532 0 : return;
533 0 : }
534 0 :
535 0 : if state.cfile_last_persist_at.elapsed() > self.conf.control_file_save_interval
536 : // If the control file's commit_lsn lags more than one segment behind the current
537 : // commit_lsn, flush immediately to limit recovery time in case of a crash. We don't do
538 : // this on the WAL ingest hot path since it incurs fsync latency.
539 0 : || state.commit_lsn.saturating_sub(state.cfile_commit_lsn).0 >= self.wal_seg_size as u64
540 : {
541 0 : let mut write_guard = self.tli.write_shared_state().await;
542 : // it should be done in the background because it blocks manager task, but flush() should
543 : // be fast enough not to be a problem now
544 0 : if let Err(e) = write_guard.sk.state_mut().flush().await {
545 0 : warn!("failed to save control file: {:?}", e);
546 0 : }
547 0 : } else {
548 0 : // we should wait until some time passed until the next save
549 0 : update_next_event(
550 0 : next_event,
551 0 : (state.cfile_last_persist_at + self.conf.control_file_save_interval).into(),
552 0 : );
553 0 : }
554 0 : }
555 :
556 : /// Spawns WAL removal task if needed.
557 0 : async fn update_wal_removal(&mut self, state: &StateSnapshot) {
558 0 : if self.wal_removal_task.is_some() || state.wal_removal_on_hold {
559 : // WAL removal is already in progress or hold off
560 0 : return;
561 0 : }
562 :
563 : // If enabled, we use LSN of the most lagging walsender as a WAL removal horizon.
564 : // This allows to get better read speed for pageservers that are lagging behind,
565 : // at the cost of keeping more WAL on disk.
566 0 : let replication_horizon_lsn = if self.conf.walsenders_keep_horizon {
567 0 : self.walsenders.laggard_lsn()
568 : } else {
569 0 : None
570 : };
571 :
572 0 : let removal_horizon_lsn = calc_horizon_lsn(state, replication_horizon_lsn);
573 0 : let removal_horizon_segno = removal_horizon_lsn
574 0 : .segment_number(self.wal_seg_size)
575 0 : .saturating_sub(1);
576 0 :
577 0 : if removal_horizon_segno > self.last_removed_segno {
578 : // we need to remove WAL
579 0 : let Ok(timeline_gate_guard) = self.tli.gate.enter() else {
580 0 : tracing::info!("Timeline shutdown, not spawning WAL removal task");
581 0 : return;
582 : };
583 :
584 0 : let remover = match self.tli.read_shared_state().await.sk {
585 0 : StateSK::Loaded(ref sk) => {
586 0 : crate::wal_storage::Storage::remove_up_to(&sk.wal_store, removal_horizon_segno)
587 : }
588 : StateSK::Offloaded(_) => {
589 : // we can't remove WAL if it's not loaded
590 0 : warn!("unexpectedly trying to run WAL removal on offloaded timeline");
591 0 : return;
592 : }
593 0 : StateSK::Empty => unreachable!(),
594 : };
595 :
596 0 : self.wal_removal_task = Some(tokio::spawn(
597 0 : async move {
598 0 : let _timeline_gate_guard = timeline_gate_guard;
599 0 :
600 0 : remover.await?;
601 0 : Ok(removal_horizon_segno)
602 0 : }
603 0 : .instrument(info_span!("WAL removal", ttid=%self.tli.ttid)),
604 : ));
605 0 : }
606 0 : }
607 :
608 : /// Update the state after WAL removal task finished.
609 0 : fn update_wal_removal_end(&mut self, res: Result<anyhow::Result<u64>, JoinError>) {
610 0 : let new_last_removed_segno = match res {
611 0 : Ok(Ok(segno)) => segno,
612 0 : Err(e) => {
613 0 : warn!("WAL removal task failed: {:?}", e);
614 0 : return;
615 : }
616 0 : Ok(Err(e)) => {
617 0 : warn!("WAL removal task failed: {:?}", e);
618 0 : return;
619 : }
620 : };
621 :
622 0 : self.last_removed_segno = new_last_removed_segno;
623 0 : // update the state in Arc<Timeline>
624 0 : self.tli
625 0 : .last_removed_segno
626 0 : .store(new_last_removed_segno, std::sync::atomic::Ordering::Relaxed);
627 0 : }
628 :
629 : /// Spawns partial WAL backup task if needed.
630 0 : async fn update_partial_backup(&mut self, state: &StateSnapshot) {
631 0 : // check if WAL backup is enabled and should be started
632 0 : if !self.conf.is_wal_backup_enabled() {
633 0 : return;
634 0 : }
635 0 :
636 0 : if self.partial_backup_task.is_some() {
637 : // partial backup is already running
638 0 : return;
639 0 : }
640 0 :
641 0 : if !wal_backup_partial::needs_uploading(state, &self.partial_backup_uploaded) {
642 : // nothing to upload
643 0 : return;
644 0 : }
645 :
646 0 : let Ok(resident) = self.wal_resident_timeline() else {
647 : // Shutting down
648 0 : return;
649 : };
650 :
651 : // Get WalResidentTimeline and start partial backup task.
652 0 : let cancel = CancellationToken::new();
653 0 : let handle = tokio::spawn(wal_backup_partial::main_task(
654 0 : resident,
655 0 : self.conf.clone(),
656 0 : self.global_rate_limiter.clone(),
657 0 : cancel.clone(),
658 0 : ));
659 0 : self.partial_backup_task = Some((handle, cancel));
660 0 : }
661 :
662 : /// Update the state after partial WAL backup task finished.
663 0 : fn update_partial_backup_end(&mut self, res: Result<Option<PartialRemoteSegment>, JoinError>) {
664 0 : match res {
665 0 : Ok(new_upload_state) => {
666 0 : self.partial_backup_uploaded = new_upload_state;
667 0 : }
668 0 : Err(e) => {
669 0 : warn!("partial backup task panicked: {:?}", e);
670 : }
671 : }
672 0 : }
673 :
674 : /// Reset partial backup state and remove its remote storage data. Since it
675 : /// might concurrently uploading something, cancel the task first.
676 0 : async fn backup_partial_reset(&mut self) -> anyhow::Result<Vec<String>> {
677 0 : info!("resetting partial backup state");
678 : // Force unevict timeline if it is evicted before erasing partial backup
679 : // state. The intended use of this function is to drop corrupted remote
680 : // state; we haven't enabled local files deletion yet anywhere,
681 : // so direct switch is safe.
682 0 : if self.is_offloaded {
683 0 : self.tli.switch_to_present().await?;
684 : // switch manager state as soon as possible
685 0 : self.is_offloaded = false;
686 0 : }
687 :
688 0 : if let Some((handle, cancel)) = &mut self.partial_backup_task {
689 0 : cancel.cancel();
690 0 : info!("cancelled partial backup task, awaiting it");
691 : // we're going to reset .partial_backup_uploaded to None anyway, so ignore the result
692 0 : handle.await.ok();
693 0 : self.partial_backup_task = None;
694 0 : }
695 :
696 0 : let tli = self.wal_resident_timeline()?;
697 0 : let mut partial_backup = PartialBackup::new(tli, self.conf.clone()).await;
698 : // Reset might fail e.g. when cfile is already reset but s3 removal
699 : // failed, so set manager state to None beforehand. In any case caller
700 : // is expected to retry until success.
701 0 : self.partial_backup_uploaded = None;
702 0 : let res = partial_backup.reset().await?;
703 0 : info!("reset is done");
704 0 : Ok(res)
705 0 : }
706 :
707 : /// Handle message arrived from ManagerCtl.
708 0 : async fn handle_message(&mut self, msg: Option<ManagerCtlMessage>) {
709 0 : debug!("received manager message: {:?}", msg);
710 0 : match msg {
711 0 : Some(ManagerCtlMessage::GuardRequest(tx)) => {
712 0 : if self.is_offloaded {
713 : // trying to unevict timeline, but without gurarantee that it will be successful
714 0 : self.unevict_timeline().await;
715 0 : }
716 :
717 0 : let guard = if self.is_offloaded {
718 0 : Err(anyhow::anyhow!("timeline is offloaded, can't get a guard"))
719 : } else {
720 0 : match self.tli.gate.enter() {
721 0 : Ok(gate_guard) => Ok(self.access_service.create_guard(gate_guard)),
722 0 : Err(_) => Err(anyhow::anyhow!(
723 0 : "timeline is shutting down, can't get a guard"
724 0 : )),
725 : }
726 : };
727 :
728 0 : if tx.send(guard).is_err() {
729 0 : warn!("failed to reply with a guard, receiver dropped");
730 0 : }
731 : }
732 0 : Some(ManagerCtlMessage::TryGuardRequest(tx)) => {
733 0 : let result = if self.is_offloaded {
734 0 : None
735 : } else {
736 0 : match self.tli.gate.enter() {
737 0 : Ok(gate_guard) => Some(self.access_service.create_guard(gate_guard)),
738 0 : Err(_) => None,
739 : }
740 : };
741 :
742 0 : if tx.send(result).is_err() {
743 0 : warn!("failed to reply with a guard, receiver dropped");
744 0 : }
745 : }
746 0 : Some(ManagerCtlMessage::GuardDrop(guard_id)) => {
747 0 : self.access_service.drop_guard(guard_id);
748 0 : }
749 0 : Some(ManagerCtlMessage::BackupPartialReset(tx)) => {
750 0 : info!("resetting uploaded partial backup state");
751 0 : let res = self.backup_partial_reset().await;
752 0 : if let Err(ref e) = res {
753 0 : warn!("failed to reset partial backup state: {:?}", e);
754 0 : }
755 0 : if tx.send(res).is_err() {
756 0 : warn!("failed to send partial backup reset result, receiver dropped");
757 0 : }
758 : }
759 : None => {
760 : // can't happen, we're holding the sender
761 0 : unreachable!();
762 : }
763 : }
764 0 : }
765 : }
766 :
767 : // utility functions
768 0 : async fn sleep_until(option: &Option<tokio::time::Instant>) {
769 0 : if let Some(timeout) = option {
770 0 : tokio::time::sleep_until(*timeout).await;
771 : } else {
772 0 : futures::future::pending::<()>().await;
773 : }
774 0 : }
775 :
776 : /// Future that resolves when the task is finished or never if the task is None.
777 : ///
778 : /// Note: it accepts Option<&mut> instead of &mut Option<> because mapping the
779 : /// option to get the latter is hard.
780 0 : async fn await_task_finish<T>(option: Option<&mut JoinHandle<T>>) -> Result<T, JoinError> {
781 0 : if let Some(task) = option {
782 0 : task.await
783 : } else {
784 0 : futures::future::pending().await
785 : }
786 0 : }
787 :
788 : /// Update next_event if candidate is earlier.
789 0 : fn update_next_event(next_event: &mut Option<Instant>, candidate: Instant) {
790 0 : if let Some(next) = next_event {
791 0 : if candidate < *next {
792 0 : *next = candidate;
793 0 : }
794 0 : } else {
795 0 : *next_event = Some(candidate);
796 0 : }
797 0 : }
798 :
799 : #[repr(usize)]
800 0 : #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
801 : pub enum Status {
802 : NotStarted,
803 : Started,
804 : StateSnapshot,
805 : UpdateBackup,
806 : UpdateControlFile,
807 : UpdateWalRemoval,
808 : UpdatePartialBackup,
809 : EvictTimeline,
810 : Wait,
811 : HandleMessage,
812 : Exiting,
813 : Finished,
814 : }
815 :
816 : /// AtomicStatus is a wrapper around AtomicUsize adapted for the Status enum.
817 : pub struct AtomicStatus {
818 : inner: AtomicUsize,
819 : }
820 :
821 : impl Default for AtomicStatus {
822 0 : fn default() -> Self {
823 0 : Self::new()
824 0 : }
825 : }
826 :
827 : impl AtomicStatus {
828 0 : pub fn new() -> Self {
829 0 : AtomicStatus {
830 0 : inner: AtomicUsize::new(Status::NotStarted as usize),
831 0 : }
832 0 : }
833 :
834 0 : pub fn load(&self, order: std::sync::atomic::Ordering) -> Status {
835 0 : // Safety: This line of code uses `std::mem::transmute` to reinterpret the loaded value as `Status`.
836 0 : // It is safe to use `transmute` in this context because `Status` is a repr(usize) enum,
837 0 : // which means it has the same memory layout as usize.
838 0 : // However, it is important to ensure that the loaded value is a valid variant of `Status`,
839 0 : // otherwise, the behavior will be undefined.
840 0 : unsafe { std::mem::transmute(self.inner.load(order)) }
841 0 : }
842 :
843 0 : pub fn get(&self) -> Status {
844 0 : self.load(std::sync::atomic::Ordering::Relaxed)
845 0 : }
846 :
847 0 : pub fn store(&self, val: Status, order: std::sync::atomic::Ordering) {
848 0 : self.inner.store(val as usize, order);
849 0 : }
850 : }
|
