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 : let tli = mgr.wal_resident_timeline();
270 : mgr.recovery_task = Some(tokio::spawn(recovery_main(tli, mgr.conf.clone())));
271 : }
272 :
273 : // If timeline is evicted, reflect that in the metric.
274 : if mgr.is_offloaded {
275 : NUM_EVICTED_TIMELINES.inc();
276 : }
277 :
278 : let last_state = 'outer: loop {
279 : MANAGER_ITERATIONS_TOTAL.inc();
280 :
281 : mgr.set_status(Status::StateSnapshot);
282 : let state_snapshot = mgr.state_snapshot().await;
283 :
284 : let mut next_event: Option<Instant> = None;
285 : if !mgr.is_offloaded {
286 : let num_computes = *mgr.num_computes_rx.borrow();
287 :
288 : mgr.set_status(Status::UpdateBackup);
289 : let is_wal_backup_required = mgr.update_backup(num_computes, &state_snapshot).await;
290 : mgr.update_is_active(is_wal_backup_required, num_computes, &state_snapshot);
291 :
292 : mgr.set_status(Status::UpdateControlFile);
293 : mgr.update_control_file_save(&state_snapshot, &mut next_event)
294 : .await;
295 :
296 : mgr.set_status(Status::UpdateWalRemoval);
297 : mgr.update_wal_removal(&state_snapshot).await;
298 :
299 : mgr.set_status(Status::UpdatePartialBackup);
300 : mgr.update_partial_backup(&state_snapshot).await;
301 :
302 : let now = Instant::now();
303 : if mgr.evict_not_before > now {
304 : // we should wait until evict_not_before
305 : update_next_event(&mut next_event, mgr.evict_not_before);
306 : }
307 :
308 : if mgr.conf.enable_offload
309 : && mgr.evict_not_before <= now
310 : && mgr.ready_for_eviction(&next_event, &state_snapshot)
311 : {
312 : // check rate limiter and evict timeline if possible
313 : match mgr.global_rate_limiter.try_acquire_eviction() {
314 : Some(_permit) => {
315 : mgr.set_status(Status::EvictTimeline);
316 : if !mgr.evict_timeline().await {
317 : // eviction failed, try again later
318 : mgr.evict_not_before =
319 : Instant::now() + rand_duration(&mgr.conf.eviction_min_resident);
320 : update_next_event(&mut next_event, mgr.evict_not_before);
321 : }
322 : }
323 : None => {
324 : // we can't evict timeline now, will try again later
325 : mgr.evict_not_before =
326 : Instant::now() + rand_duration(&mgr.conf.eviction_min_resident);
327 : update_next_event(&mut next_event, mgr.evict_not_before);
328 : }
329 : }
330 : }
331 : }
332 :
333 : mgr.set_status(Status::Wait);
334 : // wait until something changes. tx channels are stored under Arc, so they will not be
335 : // dropped until the manager task is finished.
336 : tokio::select! {
337 : _ = mgr.tli.cancel.cancelled() => {
338 : // timeline was deleted
339 : break 'outer state_snapshot;
340 : }
341 0 : _ = async {
342 0 : // don't wake up on every state change, but at most every REFRESH_INTERVAL
343 0 : tokio::time::sleep(REFRESH_INTERVAL).await;
344 0 : let _ = mgr.state_version_rx.changed().await;
345 0 : } => {
346 : // state was updated
347 : }
348 : _ = mgr.num_computes_rx.changed() => {
349 : // number of connected computes was updated
350 : }
351 : _ = sleep_until(&next_event) => {
352 : // we were waiting for some event (e.g. cfile save)
353 : }
354 : res = await_task_finish(mgr.wal_removal_task.as_mut()) => {
355 : // WAL removal task finished
356 : mgr.wal_removal_task = None;
357 : mgr.update_wal_removal_end(res);
358 : }
359 0 : res = await_task_finish(mgr.partial_backup_task.as_mut().map(|(handle, _)| handle)) => {
360 : // partial backup task finished
361 : mgr.partial_backup_task = None;
362 : mgr.update_partial_backup_end(res);
363 : }
364 :
365 : msg = manager_rx.recv() => {
366 : mgr.set_status(Status::HandleMessage);
367 : mgr.handle_message(msg).await;
368 : }
369 : }
370 : };
371 : mgr.set_status(Status::Exiting);
372 :
373 : // remove timeline from the broker active set sooner, before waiting for background tasks
374 : mgr.tli_broker_active.set(false);
375 :
376 : // shutdown background tasks
377 : if mgr.conf.is_wal_backup_enabled() {
378 : wal_backup::update_task(&mut mgr, false, &last_state).await;
379 : }
380 :
381 : if let Some(recovery_task) = &mut mgr.recovery_task {
382 : if let Err(e) = recovery_task.await {
383 : warn!("recovery task failed: {:?}", e);
384 : }
385 : }
386 :
387 : if let Some((handle, cancel)) = &mut mgr.partial_backup_task {
388 : cancel.cancel();
389 : if let Err(e) = handle.await {
390 : warn!("partial backup task failed: {:?}", e);
391 : }
392 : }
393 :
394 : if let Some(wal_removal_task) = &mut mgr.wal_removal_task {
395 : let res = wal_removal_task.await;
396 : mgr.update_wal_removal_end(res);
397 : }
398 :
399 : // If timeline is deleted while evicted decrement the gauge.
400 : if mgr.tli.is_cancelled() && mgr.is_offloaded {
401 : NUM_EVICTED_TIMELINES.dec();
402 : }
403 :
404 : mgr.set_status(Status::Finished);
405 : }
406 :
407 : impl Manager {
408 0 : async fn new(
409 0 : tli: ManagerTimeline,
410 0 : conf: SafeKeeperConf,
411 0 : broker_active_set: Arc<TimelinesSet>,
412 0 : manager_tx: tokio::sync::mpsc::UnboundedSender<ManagerCtlMessage>,
413 0 : global_rate_limiter: RateLimiter,
414 0 : ) -> Manager {
415 0 : let (is_offloaded, partial_backup_uploaded) = tli.bootstrap_mgr().await;
416 : Manager {
417 0 : wal_seg_size: tli.get_wal_seg_size().await,
418 0 : walsenders: tli.get_walsenders().clone(),
419 0 : state_version_rx: tli.get_state_version_rx(),
420 0 : num_computes_rx: tli.get_walreceivers().get_num_rx(),
421 0 : tli_broker_active: broker_active_set.guard(tli.clone()),
422 0 : last_removed_segno: 0,
423 0 : is_offloaded,
424 0 : backup_task: None,
425 0 : recovery_task: None,
426 0 : wal_removal_task: None,
427 0 : partial_backup_task: None,
428 0 : partial_backup_uploaded,
429 0 : access_service: AccessService::new(manager_tx),
430 0 : tli,
431 0 : global_rate_limiter,
432 0 : // to smooth out evictions spike after restart
433 0 : evict_not_before: Instant::now() + rand_duration(&conf.eviction_min_resident),
434 0 : conf,
435 0 : }
436 0 : }
437 :
438 0 : fn set_status(&self, status: Status) {
439 0 : self.tli.set_status(status);
440 0 : }
441 :
442 : /// Get a WalResidentTimeline.
443 : /// Manager code must use this function instead of one from `Timeline`
444 : /// directly, because it will deadlock.
445 0 : pub(crate) fn wal_resident_timeline(&mut self) -> WalResidentTimeline {
446 0 : assert!(!self.is_offloaded);
447 0 : let guard = self.access_service.create_guard();
448 0 : WalResidentTimeline::new(self.tli.clone(), guard)
449 0 : }
450 :
451 : /// Get a snapshot of the timeline state.
452 0 : async fn state_snapshot(&self) -> StateSnapshot {
453 0 : let _timer = MISC_OPERATION_SECONDS
454 0 : .with_label_values(&["state_snapshot"])
455 0 : .start_timer();
456 0 :
457 0 : StateSnapshot::new(
458 0 : self.tli.read_shared_state().await,
459 0 : self.conf.heartbeat_timeout,
460 0 : )
461 0 : }
462 :
463 : /// Spawns/kills backup task and returns true if backup is required.
464 0 : async fn update_backup(&mut self, num_computes: usize, state: &StateSnapshot) -> bool {
465 0 : let is_wal_backup_required =
466 0 : wal_backup::is_wal_backup_required(self.wal_seg_size, num_computes, state);
467 0 :
468 0 : if self.conf.is_wal_backup_enabled() {
469 0 : wal_backup::update_task(self, is_wal_backup_required, state).await;
470 0 : }
471 :
472 : // update the state in Arc<Timeline>
473 0 : self.tli.wal_backup_active.store(
474 0 : self.backup_task.is_some(),
475 0 : std::sync::atomic::Ordering::Relaxed,
476 0 : );
477 0 : is_wal_backup_required
478 0 : }
479 :
480 : /// Update is_active flag and returns its value.
481 0 : fn update_is_active(
482 0 : &mut self,
483 0 : is_wal_backup_required: bool,
484 0 : num_computes: usize,
485 0 : state: &StateSnapshot,
486 0 : ) {
487 0 : let is_active = is_wal_backup_required
488 0 : || num_computes > 0
489 0 : || state.remote_consistent_lsn < state.commit_lsn;
490 :
491 : // update the broker timeline set
492 0 : if self.tli_broker_active.set(is_active) {
493 : // write log if state has changed
494 0 : info!(
495 0 : "timeline active={} now, remote_consistent_lsn={}, commit_lsn={}",
496 : is_active, state.remote_consistent_lsn, state.commit_lsn,
497 : );
498 :
499 0 : MANAGER_ACTIVE_CHANGES.inc();
500 0 : }
501 :
502 : // update the state in Arc<Timeline>
503 0 : self.tli
504 0 : .broker_active
505 0 : .store(is_active, std::sync::atomic::Ordering::Relaxed);
506 0 : }
507 :
508 : /// Save control file if needed. Returns Instant if we should persist the control file in the future.
509 0 : async fn update_control_file_save(
510 0 : &self,
511 0 : state: &StateSnapshot,
512 0 : next_event: &mut Option<Instant>,
513 0 : ) {
514 0 : if !state.inmem_flush_pending {
515 0 : return;
516 0 : }
517 0 :
518 0 : if state.cfile_last_persist_at.elapsed() > self.conf.control_file_save_interval
519 : // If the control file's commit_lsn lags more than one segment behind the current
520 : // commit_lsn, flush immediately to limit recovery time in case of a crash. We don't do
521 : // this on the WAL ingest hot path since it incurs fsync latency.
522 0 : || state.commit_lsn.saturating_sub(state.cfile_commit_lsn).0 >= self.wal_seg_size as u64
523 : {
524 0 : let mut write_guard = self.tli.write_shared_state().await;
525 : // it should be done in the background because it blocks manager task, but flush() should
526 : // be fast enough not to be a problem now
527 0 : if let Err(e) = write_guard.sk.state_mut().flush().await {
528 0 : warn!("failed to save control file: {:?}", e);
529 0 : }
530 0 : } else {
531 0 : // we should wait until some time passed until the next save
532 0 : update_next_event(
533 0 : next_event,
534 0 : (state.cfile_last_persist_at + self.conf.control_file_save_interval).into(),
535 0 : );
536 0 : }
537 0 : }
538 :
539 : /// Spawns WAL removal task if needed.
540 0 : async fn update_wal_removal(&mut self, state: &StateSnapshot) {
541 0 : if self.wal_removal_task.is_some() || state.wal_removal_on_hold {
542 : // WAL removal is already in progress or hold off
543 0 : return;
544 0 : }
545 :
546 : // If enabled, we use LSN of the most lagging walsender as a WAL removal horizon.
547 : // This allows to get better read speed for pageservers that are lagging behind,
548 : // at the cost of keeping more WAL on disk.
549 0 : let replication_horizon_lsn = if self.conf.walsenders_keep_horizon {
550 0 : self.walsenders.laggard_lsn()
551 : } else {
552 0 : None
553 : };
554 :
555 0 : let removal_horizon_lsn = calc_horizon_lsn(state, replication_horizon_lsn);
556 0 : let removal_horizon_segno = removal_horizon_lsn
557 0 : .segment_number(self.wal_seg_size)
558 0 : .saturating_sub(1);
559 0 :
560 0 : if removal_horizon_segno > self.last_removed_segno {
561 : // we need to remove WAL
562 0 : let remover = match self.tli.read_shared_state().await.sk {
563 0 : StateSK::Loaded(ref sk) => {
564 0 : crate::wal_storage::Storage::remove_up_to(&sk.wal_store, removal_horizon_segno)
565 : }
566 : StateSK::Offloaded(_) => {
567 : // we can't remove WAL if it's not loaded
568 0 : warn!("unexpectedly trying to run WAL removal on offloaded timeline");
569 0 : return;
570 : }
571 0 : StateSK::Empty => unreachable!(),
572 : };
573 :
574 0 : self.wal_removal_task = Some(tokio::spawn(
575 0 : async move {
576 0 : remover.await?;
577 0 : Ok(removal_horizon_segno)
578 0 : }
579 0 : .instrument(info_span!("WAL removal", ttid=%self.tli.ttid)),
580 : ));
581 0 : }
582 0 : }
583 :
584 : /// Update the state after WAL removal task finished.
585 0 : fn update_wal_removal_end(&mut self, res: Result<anyhow::Result<u64>, JoinError>) {
586 0 : let new_last_removed_segno = match res {
587 0 : Ok(Ok(segno)) => segno,
588 0 : Err(e) => {
589 0 : warn!("WAL removal task failed: {:?}", e);
590 0 : return;
591 : }
592 0 : Ok(Err(e)) => {
593 0 : warn!("WAL removal task failed: {:?}", e);
594 0 : return;
595 : }
596 : };
597 :
598 0 : self.last_removed_segno = new_last_removed_segno;
599 0 : // update the state in Arc<Timeline>
600 0 : self.tli
601 0 : .last_removed_segno
602 0 : .store(new_last_removed_segno, std::sync::atomic::Ordering::Relaxed);
603 0 : }
604 :
605 : /// Spawns partial WAL backup task if needed.
606 0 : async fn update_partial_backup(&mut self, state: &StateSnapshot) {
607 0 : // check if WAL backup is enabled and should be started
608 0 : if !self.conf.is_wal_backup_enabled() {
609 0 : return;
610 0 : }
611 0 :
612 0 : if self.partial_backup_task.is_some() {
613 : // partial backup is already running
614 0 : return;
615 0 : }
616 0 :
617 0 : if !wal_backup_partial::needs_uploading(state, &self.partial_backup_uploaded) {
618 : // nothing to upload
619 0 : return;
620 0 : }
621 0 :
622 0 : // Get WalResidentTimeline and start partial backup task.
623 0 : let cancel = CancellationToken::new();
624 0 : let handle = tokio::spawn(wal_backup_partial::main_task(
625 0 : self.wal_resident_timeline(),
626 0 : self.conf.clone(),
627 0 : self.global_rate_limiter.clone(),
628 0 : cancel.clone(),
629 0 : ));
630 0 : self.partial_backup_task = Some((handle, cancel));
631 0 : }
632 :
633 : /// Update the state after partial WAL backup task finished.
634 0 : fn update_partial_backup_end(&mut self, res: Result<Option<PartialRemoteSegment>, JoinError>) {
635 0 : match res {
636 0 : Ok(new_upload_state) => {
637 0 : self.partial_backup_uploaded = new_upload_state;
638 0 : }
639 0 : Err(e) => {
640 0 : warn!("partial backup task panicked: {:?}", e);
641 : }
642 : }
643 0 : }
644 :
645 : /// Reset partial backup state and remove its remote storage data. Since it
646 : /// might concurrently uploading something, cancel the task first.
647 0 : async fn backup_partial_reset(&mut self) -> anyhow::Result<Vec<String>> {
648 0 : info!("resetting partial backup state");
649 : // Force unevict timeline if it is evicted before erasing partial backup
650 : // state. The intended use of this function is to drop corrupted remote
651 : // state; we haven't enabled local files deletion yet anywhere,
652 : // so direct switch is safe.
653 0 : if self.is_offloaded {
654 0 : self.tli.switch_to_present().await?;
655 : // switch manager state as soon as possible
656 0 : self.is_offloaded = false;
657 0 : }
658 :
659 0 : if let Some((handle, cancel)) = &mut self.partial_backup_task {
660 0 : cancel.cancel();
661 0 : info!("cancelled partial backup task, awaiting it");
662 : // we're going to reset .partial_backup_uploaded to None anyway, so ignore the result
663 0 : handle.await.ok();
664 0 : self.partial_backup_task = None;
665 0 : }
666 :
667 0 : let tli = self.wal_resident_timeline();
668 0 : let mut partial_backup = PartialBackup::new(tli, self.conf.clone()).await;
669 : // Reset might fail e.g. when cfile is already reset but s3 removal
670 : // failed, so set manager state to None beforehand. In any case caller
671 : // is expected to retry until success.
672 0 : self.partial_backup_uploaded = None;
673 0 : let res = partial_backup.reset().await?;
674 0 : info!("reset is done");
675 0 : Ok(res)
676 0 : }
677 :
678 : /// Handle message arrived from ManagerCtl.
679 0 : async fn handle_message(&mut self, msg: Option<ManagerCtlMessage>) {
680 0 : debug!("received manager message: {:?}", msg);
681 0 : match msg {
682 0 : Some(ManagerCtlMessage::GuardRequest(tx)) => {
683 0 : if self.is_offloaded {
684 : // trying to unevict timeline, but without gurarantee that it will be successful
685 0 : self.unevict_timeline().await;
686 0 : }
687 :
688 0 : let guard = if self.is_offloaded {
689 0 : Err(anyhow::anyhow!("timeline is offloaded, can't get a guard"))
690 : } else {
691 0 : Ok(self.access_service.create_guard())
692 : };
693 :
694 0 : if tx.send(guard).is_err() {
695 0 : warn!("failed to reply with a guard, receiver dropped");
696 0 : }
697 : }
698 0 : Some(ManagerCtlMessage::TryGuardRequest(tx)) => {
699 0 : let result = if self.is_offloaded {
700 0 : None
701 : } else {
702 0 : Some(self.access_service.create_guard())
703 : };
704 :
705 0 : if tx.send(result).is_err() {
706 0 : warn!("failed to reply with a guard, receiver dropped");
707 0 : }
708 : }
709 0 : Some(ManagerCtlMessage::GuardDrop(guard_id)) => {
710 0 : self.access_service.drop_guard(guard_id);
711 0 : }
712 0 : Some(ManagerCtlMessage::BackupPartialReset(tx)) => {
713 0 : info!("resetting uploaded partial backup state");
714 0 : let res = self.backup_partial_reset().await;
715 0 : if let Err(ref e) = res {
716 0 : warn!("failed to reset partial backup state: {:?}", e);
717 0 : }
718 0 : if tx.send(res).is_err() {
719 0 : warn!("failed to send partial backup reset result, receiver dropped");
720 0 : }
721 : }
722 : None => {
723 : // can't happen, we're holding the sender
724 0 : unreachable!();
725 : }
726 : }
727 0 : }
728 : }
729 :
730 : // utility functions
731 0 : async fn sleep_until(option: &Option<tokio::time::Instant>) {
732 0 : if let Some(timeout) = option {
733 0 : tokio::time::sleep_until(*timeout).await;
734 : } else {
735 0 : futures::future::pending::<()>().await;
736 : }
737 0 : }
738 :
739 : /// Future that resolves when the task is finished or never if the task is None.
740 : ///
741 : /// Note: it accepts Option<&mut> instead of &mut Option<> because mapping the
742 : /// option to get the latter is hard.
743 0 : async fn await_task_finish<T>(option: Option<&mut JoinHandle<T>>) -> Result<T, JoinError> {
744 0 : if let Some(task) = option {
745 0 : task.await
746 : } else {
747 0 : futures::future::pending().await
748 : }
749 0 : }
750 :
751 : /// Update next_event if candidate is earlier.
752 0 : fn update_next_event(next_event: &mut Option<Instant>, candidate: Instant) {
753 0 : if let Some(next) = next_event {
754 0 : if candidate < *next {
755 0 : *next = candidate;
756 0 : }
757 0 : } else {
758 0 : *next_event = Some(candidate);
759 0 : }
760 0 : }
761 :
762 : #[repr(usize)]
763 0 : #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
764 : pub enum Status {
765 : NotStarted,
766 : Started,
767 : StateSnapshot,
768 : UpdateBackup,
769 : UpdateControlFile,
770 : UpdateWalRemoval,
771 : UpdatePartialBackup,
772 : EvictTimeline,
773 : Wait,
774 : HandleMessage,
775 : Exiting,
776 : Finished,
777 : }
778 :
779 : /// AtomicStatus is a wrapper around AtomicUsize adapted for the Status enum.
780 : pub struct AtomicStatus {
781 : inner: AtomicUsize,
782 : }
783 :
784 : impl Default for AtomicStatus {
785 0 : fn default() -> Self {
786 0 : Self::new()
787 0 : }
788 : }
789 :
790 : impl AtomicStatus {
791 0 : pub fn new() -> Self {
792 0 : AtomicStatus {
793 0 : inner: AtomicUsize::new(Status::NotStarted as usize),
794 0 : }
795 0 : }
796 :
797 0 : pub fn load(&self, order: std::sync::atomic::Ordering) -> Status {
798 0 : // Safety: This line of code uses `std::mem::transmute` to reinterpret the loaded value as `Status`.
799 0 : // It is safe to use `transmute` in this context because `Status` is a repr(usize) enum,
800 0 : // which means it has the same memory layout as usize.
801 0 : // However, it is important to ensure that the loaded value is a valid variant of `Status`,
802 0 : // otherwise, the behavior will be undefined.
803 0 : unsafe { std::mem::transmute(self.inner.load(order)) }
804 0 : }
805 :
806 0 : pub fn get(&self) -> Status {
807 0 : self.load(std::sync::atomic::Ordering::Relaxed)
808 0 : }
809 :
810 0 : pub fn store(&self, val: Status, order: std::sync::atomic::Ordering) {
811 0 : self.inner.store(val as usize, order);
812 0 : }
813 : }
|
