Line data Source code
1 : //! This module implements Timeline lifecycle management and has all necessary code
2 : //! to glue together SafeKeeper and all other background services.
3 :
4 : use anyhow::{anyhow, bail, Result};
5 : use camino::Utf8PathBuf;
6 : use serde::{Deserialize, Serialize};
7 : use tokio::fs::{self};
8 : use tokio_util::sync::CancellationToken;
9 : use utils::id::TenantId;
10 :
11 : use std::cmp::max;
12 : use std::ops::{Deref, DerefMut};
13 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
14 : use std::sync::Arc;
15 : use std::time::Duration;
16 : use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
17 : use tokio::{sync::watch, time::Instant};
18 : use tracing::*;
19 : use utils::http::error::ApiError;
20 : use utils::{
21 : id::{NodeId, TenantTimelineId},
22 : lsn::Lsn,
23 : };
24 :
25 : use storage_broker::proto::SafekeeperTimelineInfo;
26 : use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
27 :
28 : use crate::rate_limit::RateLimiter;
29 : use crate::receive_wal::WalReceivers;
30 : use crate::safekeeper::{
31 : AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, ServerInfo, Term, TermLsn,
32 : INVALID_TERM,
33 : };
34 : use crate::send_wal::WalSenders;
35 : use crate::state::{EvictionState, TimelineMemState, TimelinePersistentState, TimelineState};
36 : use crate::timeline_guard::ResidenceGuard;
37 : use crate::timeline_manager::{AtomicStatus, ManagerCtl};
38 : use crate::timelines_set::TimelinesSet;
39 : use crate::wal_backup::{self};
40 : use crate::wal_backup_partial::PartialRemoteSegment;
41 : use crate::{control_file, safekeeper::UNKNOWN_SERVER_VERSION};
42 :
43 : use crate::metrics::{FullTimelineInfo, WalStorageMetrics, MISC_OPERATION_SECONDS};
44 : use crate::wal_storage::{Storage as wal_storage_iface, WalReader};
45 : use crate::{debug_dump, timeline_manager, wal_storage};
46 : use crate::{GlobalTimelines, SafeKeeperConf};
47 :
48 : /// Things safekeeper should know about timeline state on peers.
49 0 : #[derive(Debug, Clone, Serialize, Deserialize)]
50 : pub struct PeerInfo {
51 : pub sk_id: NodeId,
52 : pub term: Term,
53 : /// Term of the last entry.
54 : pub last_log_term: Term,
55 : /// LSN of the last record.
56 : pub flush_lsn: Lsn,
57 : pub commit_lsn: Lsn,
58 : /// Since which LSN safekeeper has WAL.
59 : pub local_start_lsn: Lsn,
60 : /// When info was received. Serde annotations are not very useful but make
61 : /// the code compile -- we don't rely on this field externally.
62 : #[serde(skip)]
63 : #[serde(default = "Instant::now")]
64 : ts: Instant,
65 : pub pg_connstr: String,
66 : pub http_connstr: String,
67 : }
68 :
69 : impl PeerInfo {
70 0 : fn from_sk_info(sk_info: &SafekeeperTimelineInfo, ts: Instant) -> PeerInfo {
71 0 : PeerInfo {
72 0 : sk_id: NodeId(sk_info.safekeeper_id),
73 0 : term: sk_info.term,
74 0 : last_log_term: sk_info.last_log_term,
75 0 : flush_lsn: Lsn(sk_info.flush_lsn),
76 0 : commit_lsn: Lsn(sk_info.commit_lsn),
77 0 : local_start_lsn: Lsn(sk_info.local_start_lsn),
78 0 : pg_connstr: sk_info.safekeeper_connstr.clone(),
79 0 : http_connstr: sk_info.http_connstr.clone(),
80 0 : ts,
81 0 : }
82 0 : }
83 : }
84 :
85 : // vector-based node id -> peer state map with very limited functionality we
86 : // need.
87 : #[derive(Debug, Clone, Default)]
88 : pub struct PeersInfo(pub Vec<PeerInfo>);
89 :
90 : impl PeersInfo {
91 0 : fn get(&mut self, id: NodeId) -> Option<&mut PeerInfo> {
92 0 : self.0.iter_mut().find(|p| p.sk_id == id)
93 0 : }
94 :
95 0 : fn upsert(&mut self, p: &PeerInfo) {
96 0 : match self.get(p.sk_id) {
97 0 : Some(rp) => *rp = p.clone(),
98 0 : None => self.0.push(p.clone()),
99 : }
100 0 : }
101 : }
102 :
103 : pub type ReadGuardSharedState<'a> = RwLockReadGuard<'a, SharedState>;
104 :
105 : /// WriteGuardSharedState is a wrapper around `RwLockWriteGuard<SharedState>` that
106 : /// automatically updates `watch::Sender` channels with state on drop.
107 : pub struct WriteGuardSharedState<'a> {
108 : tli: Arc<Timeline>,
109 : guard: RwLockWriteGuard<'a, SharedState>,
110 : skip_update: bool,
111 : }
112 :
113 : impl<'a> WriteGuardSharedState<'a> {
114 0 : fn new(tli: Arc<Timeline>, guard: RwLockWriteGuard<'a, SharedState>) -> Self {
115 0 : WriteGuardSharedState {
116 0 : tli,
117 0 : guard,
118 0 : skip_update: false,
119 0 : }
120 0 : }
121 : }
122 :
123 : impl<'a> Deref for WriteGuardSharedState<'a> {
124 : type Target = SharedState;
125 :
126 0 : fn deref(&self) -> &Self::Target {
127 0 : &self.guard
128 0 : }
129 : }
130 :
131 : impl<'a> DerefMut for WriteGuardSharedState<'a> {
132 0 : fn deref_mut(&mut self) -> &mut Self::Target {
133 0 : &mut self.guard
134 0 : }
135 : }
136 :
137 : impl<'a> Drop for WriteGuardSharedState<'a> {
138 0 : fn drop(&mut self) {
139 0 : let term_flush_lsn =
140 0 : TermLsn::from((self.guard.sk.last_log_term(), self.guard.sk.flush_lsn()));
141 0 : let commit_lsn = self.guard.sk.state().inmem.commit_lsn;
142 0 :
143 0 : let _ = self.tli.term_flush_lsn_watch_tx.send_if_modified(|old| {
144 0 : if *old != term_flush_lsn {
145 0 : *old = term_flush_lsn;
146 0 : true
147 : } else {
148 0 : false
149 : }
150 0 : });
151 0 :
152 0 : let _ = self.tli.commit_lsn_watch_tx.send_if_modified(|old| {
153 0 : if *old != commit_lsn {
154 0 : *old = commit_lsn;
155 0 : true
156 : } else {
157 0 : false
158 : }
159 0 : });
160 0 :
161 0 : if !self.skip_update {
162 0 : // send notification about shared state update
163 0 : self.tli.shared_state_version_tx.send_modify(|old| {
164 0 : *old += 1;
165 0 : });
166 0 : }
167 0 : }
168 : }
169 :
170 : /// This structure is stored in shared state and represents the state of the timeline.
171 : /// Usually it holds SafeKeeper, but it also supports offloaded timeline state. In this
172 : /// case, SafeKeeper is not available (because WAL is not present on disk) and all
173 : /// operations can be done only with control file.
174 : pub enum StateSK {
175 : Loaded(SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>),
176 : Offloaded(Box<TimelineState<control_file::FileStorage>>),
177 : // Not used, required for moving between states.
178 : Empty,
179 : }
180 :
181 : impl StateSK {
182 0 : pub fn flush_lsn(&self) -> Lsn {
183 0 : match self {
184 0 : StateSK::Loaded(sk) => sk.wal_store.flush_lsn(),
185 0 : StateSK::Offloaded(state) => match state.eviction_state {
186 0 : EvictionState::Offloaded(flush_lsn) => flush_lsn,
187 0 : _ => panic!("StateSK::Offloaded mismatches with eviction_state from control_file"),
188 : },
189 0 : StateSK::Empty => unreachable!(),
190 : }
191 0 : }
192 :
193 : /// Get a reference to the control file's timeline state.
194 0 : pub fn state(&self) -> &TimelineState<control_file::FileStorage> {
195 0 : match self {
196 0 : StateSK::Loaded(sk) => &sk.state,
197 0 : StateSK::Offloaded(ref s) => s,
198 0 : StateSK::Empty => unreachable!(),
199 : }
200 0 : }
201 :
202 0 : pub fn state_mut(&mut self) -> &mut TimelineState<control_file::FileStorage> {
203 0 : match self {
204 0 : StateSK::Loaded(sk) => &mut sk.state,
205 0 : StateSK::Offloaded(ref mut s) => s,
206 0 : StateSK::Empty => unreachable!(),
207 : }
208 0 : }
209 :
210 0 : pub fn last_log_term(&self) -> Term {
211 0 : self.state()
212 0 : .acceptor_state
213 0 : .get_last_log_term(self.flush_lsn())
214 0 : }
215 :
216 : /// Close open WAL files to release FDs.
217 0 : fn close_wal_store(&mut self) {
218 0 : if let StateSK::Loaded(sk) = self {
219 0 : sk.wal_store.close();
220 0 : }
221 0 : }
222 :
223 : /// Update timeline state with peer safekeeper data.
224 0 : pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> {
225 0 : // update commit_lsn if safekeeper is loaded
226 0 : match self {
227 0 : StateSK::Loaded(sk) => sk.record_safekeeper_info(sk_info).await?,
228 0 : StateSK::Offloaded(_) => {}
229 0 : StateSK::Empty => unreachable!(),
230 : }
231 :
232 : // update everything else, including remote_consistent_lsn and backup_lsn
233 0 : let mut sync_control_file = false;
234 0 : let state = self.state_mut();
235 0 : let wal_seg_size = state.server.wal_seg_size as u64;
236 0 :
237 0 : state.inmem.backup_lsn = max(Lsn(sk_info.backup_lsn), state.inmem.backup_lsn);
238 0 : sync_control_file |= state.backup_lsn + wal_seg_size < state.inmem.backup_lsn;
239 0 :
240 0 : state.inmem.remote_consistent_lsn = max(
241 0 : Lsn(sk_info.remote_consistent_lsn),
242 0 : state.inmem.remote_consistent_lsn,
243 0 : );
244 0 : sync_control_file |=
245 0 : state.remote_consistent_lsn + wal_seg_size < state.inmem.remote_consistent_lsn;
246 0 :
247 0 : state.inmem.peer_horizon_lsn =
248 0 : max(Lsn(sk_info.peer_horizon_lsn), state.inmem.peer_horizon_lsn);
249 0 : sync_control_file |= state.peer_horizon_lsn + wal_seg_size < state.inmem.peer_horizon_lsn;
250 0 :
251 0 : if sync_control_file {
252 0 : state.flush().await?;
253 0 : }
254 0 : Ok(())
255 0 : }
256 :
257 : /// Previously known as epoch_start_lsn. Needed only for reference in some APIs.
258 0 : pub fn term_start_lsn(&self) -> Lsn {
259 0 : match self {
260 0 : StateSK::Loaded(sk) => sk.term_start_lsn,
261 0 : StateSK::Offloaded(_) => Lsn(0),
262 0 : StateSK::Empty => unreachable!(),
263 : }
264 0 : }
265 :
266 : /// Used for metrics only.
267 0 : pub fn wal_storage_metrics(&self) -> WalStorageMetrics {
268 0 : match self {
269 0 : StateSK::Loaded(sk) => sk.wal_store.get_metrics(),
270 0 : StateSK::Offloaded(_) => WalStorageMetrics::default(),
271 0 : StateSK::Empty => unreachable!(),
272 : }
273 0 : }
274 :
275 : /// Returns WAL storage internal LSNs for debug dump.
276 0 : pub fn wal_storage_internal_state(&self) -> (Lsn, Lsn, Lsn, bool) {
277 0 : match self {
278 0 : StateSK::Loaded(sk) => sk.wal_store.internal_state(),
279 : StateSK::Offloaded(_) => {
280 0 : let flush_lsn = self.flush_lsn();
281 0 : (flush_lsn, flush_lsn, flush_lsn, false)
282 : }
283 0 : StateSK::Empty => unreachable!(),
284 : }
285 0 : }
286 :
287 : /// Access to SafeKeeper object. Panics if offloaded, should be good to use from WalResidentTimeline.
288 0 : pub fn safekeeper(
289 0 : &mut self,
290 0 : ) -> &mut SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage> {
291 0 : match self {
292 0 : StateSK::Loaded(sk) => sk,
293 : StateSK::Offloaded(_) => {
294 0 : panic!("safekeeper is offloaded, cannot be used")
295 : }
296 0 : StateSK::Empty => unreachable!(),
297 : }
298 0 : }
299 :
300 : /// Moves control file's state structure out of the enum. Used to switch states.
301 0 : fn take_state(self) -> TimelineState<control_file::FileStorage> {
302 0 : match self {
303 0 : StateSK::Loaded(sk) => sk.state,
304 0 : StateSK::Offloaded(state) => *state,
305 0 : StateSK::Empty => unreachable!(),
306 : }
307 0 : }
308 : }
309 :
310 : /// Shared state associated with database instance
311 : pub struct SharedState {
312 : /// Safekeeper object
313 : pub(crate) sk: StateSK,
314 : /// In memory list containing state of peers sent in latest messages from them.
315 : pub(crate) peers_info: PeersInfo,
316 : // True value hinders old WAL removal; this is used by snapshotting. We
317 : // could make it a counter, but there is no need to.
318 : pub(crate) wal_removal_on_hold: bool,
319 : }
320 :
321 : impl SharedState {
322 : /// Initialize fresh timeline state without persisting anything to disk.
323 0 : fn create_new(
324 0 : conf: &SafeKeeperConf,
325 0 : ttid: &TenantTimelineId,
326 0 : state: TimelinePersistentState,
327 0 : ) -> Result<Self> {
328 0 : if state.server.wal_seg_size == 0 {
329 0 : bail!(TimelineError::UninitializedWalSegSize(*ttid));
330 0 : }
331 0 :
332 0 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
333 0 : bail!(TimelineError::UninitialinzedPgVersion(*ttid));
334 0 : }
335 0 :
336 0 : if state.commit_lsn < state.local_start_lsn {
337 0 : bail!(
338 0 : "commit_lsn {} is higher than local_start_lsn {}",
339 0 : state.commit_lsn,
340 0 : state.local_start_lsn
341 0 : );
342 0 : }
343 0 :
344 0 : // We don't want to write anything to disk, because we may have existing timeline there.
345 0 : // These functions should not change anything on disk.
346 0 : let timeline_dir = get_timeline_dir(conf, ttid);
347 0 : let control_store =
348 0 : control_file::FileStorage::create_new(timeline_dir.clone(), conf, state)?;
349 0 : let wal_store =
350 0 : wal_storage::PhysicalStorage::new(ttid, timeline_dir, conf, &control_store)?;
351 0 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
352 :
353 0 : Ok(Self {
354 0 : sk: StateSK::Loaded(sk),
355 0 : peers_info: PeersInfo(vec![]),
356 0 : wal_removal_on_hold: false,
357 0 : })
358 0 : }
359 :
360 : /// Restore SharedState from control file. If file doesn't exist, bails out.
361 0 : fn restore(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Result<Self> {
362 0 : let timeline_dir = get_timeline_dir(conf, ttid);
363 0 : let control_store = control_file::FileStorage::restore_new(ttid, conf)?;
364 0 : if control_store.server.wal_seg_size == 0 {
365 0 : bail!(TimelineError::UninitializedWalSegSize(*ttid));
366 0 : }
367 :
368 0 : let sk = match control_store.eviction_state {
369 : EvictionState::Present => {
370 0 : let wal_store =
371 0 : wal_storage::PhysicalStorage::new(ttid, timeline_dir, conf, &control_store)?;
372 0 : StateSK::Loaded(SafeKeeper::new(
373 0 : TimelineState::new(control_store),
374 0 : wal_store,
375 0 : conf.my_id,
376 0 : )?)
377 : }
378 : EvictionState::Offloaded(_) => {
379 0 : StateSK::Offloaded(Box::new(TimelineState::new(control_store)))
380 : }
381 : };
382 :
383 0 : Ok(Self {
384 0 : sk,
385 0 : peers_info: PeersInfo(vec![]),
386 0 : wal_removal_on_hold: false,
387 0 : })
388 0 : }
389 :
390 0 : pub(crate) fn get_wal_seg_size(&self) -> usize {
391 0 : self.sk.state().server.wal_seg_size as usize
392 0 : }
393 :
394 0 : fn get_safekeeper_info(
395 0 : &self,
396 0 : ttid: &TenantTimelineId,
397 0 : conf: &SafeKeeperConf,
398 0 : standby_apply_lsn: Lsn,
399 0 : ) -> SafekeeperTimelineInfo {
400 0 : SafekeeperTimelineInfo {
401 0 : safekeeper_id: conf.my_id.0,
402 0 : tenant_timeline_id: Some(ProtoTenantTimelineId {
403 0 : tenant_id: ttid.tenant_id.as_ref().to_owned(),
404 0 : timeline_id: ttid.timeline_id.as_ref().to_owned(),
405 0 : }),
406 0 : term: self.sk.state().acceptor_state.term,
407 0 : last_log_term: self.sk.last_log_term(),
408 0 : flush_lsn: self.sk.flush_lsn().0,
409 0 : // note: this value is not flushed to control file yet and can be lost
410 0 : commit_lsn: self.sk.state().inmem.commit_lsn.0,
411 0 : remote_consistent_lsn: self.sk.state().inmem.remote_consistent_lsn.0,
412 0 : peer_horizon_lsn: self.sk.state().inmem.peer_horizon_lsn.0,
413 0 : safekeeper_connstr: conf
414 0 : .advertise_pg_addr
415 0 : .to_owned()
416 0 : .unwrap_or(conf.listen_pg_addr.clone()),
417 0 : http_connstr: conf.listen_http_addr.to_owned(),
418 0 : backup_lsn: self.sk.state().inmem.backup_lsn.0,
419 0 : local_start_lsn: self.sk.state().local_start_lsn.0,
420 0 : availability_zone: conf.availability_zone.clone(),
421 0 : standby_horizon: standby_apply_lsn.0,
422 0 : }
423 0 : }
424 :
425 : /// Get our latest view of alive peers status on the timeline.
426 : /// We pass our own info through the broker as well, so when we don't have connection
427 : /// to the broker returned vec is empty.
428 0 : pub(crate) fn get_peers(&self, heartbeat_timeout: Duration) -> Vec<PeerInfo> {
429 0 : let now = Instant::now();
430 0 : self.peers_info
431 0 : .0
432 0 : .iter()
433 0 : // Regard peer as absent if we haven't heard from it within heartbeat_timeout.
434 0 : .filter(|p| now.duration_since(p.ts) <= heartbeat_timeout)
435 0 : .cloned()
436 0 : .collect()
437 0 : }
438 : }
439 :
440 0 : #[derive(Debug, thiserror::Error)]
441 : pub enum TimelineError {
442 : #[error("Timeline {0} was cancelled and cannot be used anymore")]
443 : Cancelled(TenantTimelineId),
444 : #[error("Timeline {0} was not found in global map")]
445 : NotFound(TenantTimelineId),
446 : #[error("Timeline {0} exists on disk, but wasn't loaded on startup")]
447 : Invalid(TenantTimelineId),
448 : #[error("Timeline {0} is already exists")]
449 : AlreadyExists(TenantTimelineId),
450 : #[error("Timeline {0} is not initialized, wal_seg_size is zero")]
451 : UninitializedWalSegSize(TenantTimelineId),
452 : #[error("Timeline {0} is not initialized, pg_version is unknown")]
453 : UninitialinzedPgVersion(TenantTimelineId),
454 : }
455 :
456 : // Convert to HTTP API error.
457 : impl From<TimelineError> for ApiError {
458 0 : fn from(te: TimelineError) -> ApiError {
459 0 : match te {
460 0 : TimelineError::NotFound(ttid) => {
461 0 : ApiError::NotFound(anyhow!("timeline {} not found", ttid).into())
462 : }
463 0 : _ => ApiError::InternalServerError(anyhow!("{}", te)),
464 : }
465 0 : }
466 : }
467 :
468 : /// Timeline struct manages lifecycle (creation, deletion, restore) of a safekeeper timeline.
469 : /// It also holds SharedState and provides mutually exclusive access to it.
470 : pub struct Timeline {
471 : pub ttid: TenantTimelineId,
472 :
473 : /// Used to broadcast commit_lsn updates to all background jobs.
474 : commit_lsn_watch_tx: watch::Sender<Lsn>,
475 : commit_lsn_watch_rx: watch::Receiver<Lsn>,
476 :
477 : /// Broadcasts (current term, flush_lsn) updates, walsender is interested in
478 : /// them when sending in recovery mode (to walproposer or peers). Note: this
479 : /// is just a notification, WAL reading should always done with lock held as
480 : /// term can change otherwise.
481 : term_flush_lsn_watch_tx: watch::Sender<TermLsn>,
482 : term_flush_lsn_watch_rx: watch::Receiver<TermLsn>,
483 :
484 : /// Broadcasts shared state updates.
485 : shared_state_version_tx: watch::Sender<usize>,
486 : shared_state_version_rx: watch::Receiver<usize>,
487 :
488 : /// Safekeeper and other state, that should remain consistent and
489 : /// synchronized with the disk. This is tokio mutex as we write WAL to disk
490 : /// while holding it, ensuring that consensus checks are in order.
491 : mutex: RwLock<SharedState>,
492 : walsenders: Arc<WalSenders>,
493 : walreceivers: Arc<WalReceivers>,
494 : timeline_dir: Utf8PathBuf,
495 : manager_ctl: ManagerCtl,
496 :
497 : /// Delete/cancel will trigger this, background tasks should drop out as soon as it fires
498 : pub(crate) cancel: CancellationToken,
499 :
500 : // timeline_manager controlled state
501 : pub(crate) broker_active: AtomicBool,
502 : pub(crate) wal_backup_active: AtomicBool,
503 : pub(crate) last_removed_segno: AtomicU64,
504 : pub(crate) mgr_status: AtomicStatus,
505 : }
506 :
507 : impl Timeline {
508 : /// Load existing timeline from disk.
509 0 : pub fn load_timeline(conf: &SafeKeeperConf, ttid: TenantTimelineId) -> Result<Timeline> {
510 0 : let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
511 :
512 0 : let shared_state = SharedState::restore(conf, &ttid)?;
513 0 : let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
514 0 : watch::channel(shared_state.sk.state().commit_lsn);
515 0 : let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
516 0 : shared_state.sk.last_log_term(),
517 0 : shared_state.sk.flush_lsn(),
518 0 : )));
519 0 : let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
520 0 :
521 0 : let walreceivers = WalReceivers::new();
522 0 : Ok(Timeline {
523 0 : ttid,
524 0 : commit_lsn_watch_tx,
525 0 : commit_lsn_watch_rx,
526 0 : term_flush_lsn_watch_tx,
527 0 : term_flush_lsn_watch_rx,
528 0 : shared_state_version_tx,
529 0 : shared_state_version_rx,
530 0 : mutex: RwLock::new(shared_state),
531 0 : walsenders: WalSenders::new(walreceivers.clone()),
532 0 : walreceivers,
533 0 : cancel: CancellationToken::default(),
534 0 : timeline_dir: get_timeline_dir(conf, &ttid),
535 0 : manager_ctl: ManagerCtl::new(),
536 0 : broker_active: AtomicBool::new(false),
537 0 : wal_backup_active: AtomicBool::new(false),
538 0 : last_removed_segno: AtomicU64::new(0),
539 0 : mgr_status: AtomicStatus::new(),
540 0 : })
541 0 : }
542 :
543 : /// Create a new timeline, which is not yet persisted to disk.
544 0 : pub fn create_empty(
545 0 : conf: &SafeKeeperConf,
546 0 : ttid: TenantTimelineId,
547 0 : server_info: ServerInfo,
548 0 : commit_lsn: Lsn,
549 0 : local_start_lsn: Lsn,
550 0 : ) -> Result<Timeline> {
551 0 : let (commit_lsn_watch_tx, commit_lsn_watch_rx) = watch::channel(Lsn::INVALID);
552 0 : let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) =
553 0 : watch::channel(TermLsn::from((INVALID_TERM, Lsn::INVALID)));
554 0 : let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
555 0 :
556 0 : let state =
557 0 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
558 0 :
559 0 : let walreceivers = WalReceivers::new();
560 0 : Ok(Timeline {
561 0 : ttid,
562 0 : commit_lsn_watch_tx,
563 0 : commit_lsn_watch_rx,
564 0 : term_flush_lsn_watch_tx,
565 0 : term_flush_lsn_watch_rx,
566 0 : shared_state_version_tx,
567 0 : shared_state_version_rx,
568 0 : mutex: RwLock::new(SharedState::create_new(conf, &ttid, state)?),
569 0 : walsenders: WalSenders::new(walreceivers.clone()),
570 0 : walreceivers,
571 0 : cancel: CancellationToken::default(),
572 0 : timeline_dir: get_timeline_dir(conf, &ttid),
573 0 : manager_ctl: ManagerCtl::new(),
574 0 : broker_active: AtomicBool::new(false),
575 0 : wal_backup_active: AtomicBool::new(false),
576 0 : last_removed_segno: AtomicU64::new(0),
577 0 : mgr_status: AtomicStatus::new(),
578 : })
579 0 : }
580 :
581 : /// Initialize fresh timeline on disk and start background tasks. If init
582 : /// fails, timeline is cancelled and cannot be used anymore.
583 : ///
584 : /// Init is transactional, so if it fails, created files will be deleted,
585 : /// and state on disk should remain unchanged.
586 0 : pub async fn init_new(
587 0 : self: &Arc<Timeline>,
588 0 : shared_state: &mut WriteGuardSharedState<'_>,
589 0 : conf: &SafeKeeperConf,
590 0 : broker_active_set: Arc<TimelinesSet>,
591 0 : partial_backup_rate_limiter: RateLimiter,
592 0 : ) -> Result<()> {
593 0 : match fs::metadata(&self.timeline_dir).await {
594 : Ok(_) => {
595 : // Timeline directory exists on disk, we should leave state unchanged
596 : // and return error.
597 0 : bail!(TimelineError::Invalid(self.ttid));
598 : }
599 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
600 0 : Err(e) => {
601 0 : return Err(e.into());
602 : }
603 : }
604 :
605 : // Create timeline directory.
606 0 : fs::create_dir_all(&self.timeline_dir).await?;
607 :
608 : // Write timeline to disk and start background tasks.
609 0 : if let Err(e) = shared_state.sk.state_mut().flush().await {
610 : // Bootstrap failed, cancel timeline and remove timeline directory.
611 0 : self.cancel(shared_state);
612 :
613 0 : if let Err(fs_err) = fs::remove_dir_all(&self.timeline_dir).await {
614 0 : warn!(
615 0 : "failed to remove timeline {} directory after bootstrap failure: {}",
616 0 : self.ttid, fs_err
617 : );
618 0 : }
619 :
620 0 : return Err(e);
621 0 : }
622 0 : self.bootstrap(conf, broker_active_set, partial_backup_rate_limiter);
623 0 : Ok(())
624 0 : }
625 :
626 : /// Bootstrap new or existing timeline starting background tasks.
627 0 : pub fn bootstrap(
628 0 : self: &Arc<Timeline>,
629 0 : conf: &SafeKeeperConf,
630 0 : broker_active_set: Arc<TimelinesSet>,
631 0 : partial_backup_rate_limiter: RateLimiter,
632 0 : ) {
633 0 : let (tx, rx) = self.manager_ctl.bootstrap_manager();
634 0 :
635 0 : // Start manager task which will monitor timeline state and update
636 0 : // background tasks.
637 0 : tokio::spawn(timeline_manager::main_task(
638 0 : ManagerTimeline { tli: self.clone() },
639 0 : conf.clone(),
640 0 : broker_active_set,
641 0 : tx,
642 0 : rx,
643 0 : partial_backup_rate_limiter,
644 0 : ));
645 0 : }
646 :
647 : /// Delete timeline from disk completely, by removing timeline directory.
648 : /// Background timeline activities will stop eventually.
649 : ///
650 : /// Also deletes WAL in s3. Might fail if e.g. s3 is unavailable, but
651 : /// deletion API endpoint is retriable.
652 0 : pub async fn delete(
653 0 : &self,
654 0 : shared_state: &mut WriteGuardSharedState<'_>,
655 0 : only_local: bool,
656 0 : ) -> Result<bool> {
657 0 : self.cancel(shared_state);
658 0 :
659 0 : // TODO: It's better to wait for s3 offloader termination before
660 0 : // removing data from s3. Though since s3 doesn't have transactions it
661 0 : // still wouldn't guarantee absense of data after removal.
662 0 : let conf = GlobalTimelines::get_global_config();
663 0 : if !only_local && conf.is_wal_backup_enabled() {
664 : // Note: we concurrently delete remote storage data from multiple
665 : // safekeepers. That's ok, s3 replies 200 if object doesn't exist and we
666 : // do some retries anyway.
667 0 : wal_backup::delete_timeline(&self.ttid).await?;
668 0 : }
669 0 : let dir_existed = delete_dir(&self.timeline_dir).await?;
670 0 : Ok(dir_existed)
671 0 : }
672 :
673 : /// Cancel timeline to prevent further usage. Background tasks will stop
674 : /// eventually after receiving cancellation signal.
675 0 : fn cancel(&self, shared_state: &mut WriteGuardSharedState<'_>) {
676 0 : info!("timeline {} is cancelled", self.ttid);
677 0 : self.cancel.cancel();
678 0 : // Close associated FDs. Nobody will be able to touch timeline data once
679 0 : // it is cancelled, so WAL storage won't be opened again.
680 0 : shared_state.sk.close_wal_store();
681 0 : }
682 :
683 : /// Returns if timeline is cancelled.
684 0 : pub fn is_cancelled(&self) -> bool {
685 0 : self.cancel.is_cancelled()
686 0 : }
687 :
688 : /// Take a writing mutual exclusive lock on timeline shared_state.
689 0 : pub async fn write_shared_state<'a>(self: &'a Arc<Self>) -> WriteGuardSharedState<'a> {
690 0 : WriteGuardSharedState::new(self.clone(), self.mutex.write().await)
691 0 : }
692 :
693 0 : pub async fn read_shared_state(&self) -> ReadGuardSharedState {
694 0 : self.mutex.read().await
695 0 : }
696 :
697 : /// Returns commit_lsn watch channel.
698 0 : pub fn get_commit_lsn_watch_rx(&self) -> watch::Receiver<Lsn> {
699 0 : self.commit_lsn_watch_rx.clone()
700 0 : }
701 :
702 : /// Returns term_flush_lsn watch channel.
703 0 : pub fn get_term_flush_lsn_watch_rx(&self) -> watch::Receiver<TermLsn> {
704 0 : self.term_flush_lsn_watch_rx.clone()
705 0 : }
706 :
707 : /// Returns watch channel for SharedState update version.
708 0 : pub fn get_state_version_rx(&self) -> watch::Receiver<usize> {
709 0 : self.shared_state_version_rx.clone()
710 0 : }
711 :
712 : /// Returns wal_seg_size.
713 0 : pub async fn get_wal_seg_size(&self) -> usize {
714 0 : self.read_shared_state().await.get_wal_seg_size()
715 0 : }
716 :
717 : /// Returns state of the timeline.
718 0 : pub async fn get_state(&self) -> (TimelineMemState, TimelinePersistentState) {
719 0 : let state = self.read_shared_state().await;
720 0 : (
721 0 : state.sk.state().inmem.clone(),
722 0 : TimelinePersistentState::clone(state.sk.state()),
723 0 : )
724 0 : }
725 :
726 : /// Returns latest backup_lsn.
727 0 : pub async fn get_wal_backup_lsn(&self) -> Lsn {
728 0 : self.read_shared_state().await.sk.state().inmem.backup_lsn
729 0 : }
730 :
731 : /// Sets backup_lsn to the given value.
732 0 : pub async fn set_wal_backup_lsn(self: &Arc<Self>, backup_lsn: Lsn) -> Result<()> {
733 0 : if self.is_cancelled() {
734 0 : bail!(TimelineError::Cancelled(self.ttid));
735 0 : }
736 :
737 0 : let mut state = self.write_shared_state().await;
738 0 : state.sk.state_mut().inmem.backup_lsn = max(state.sk.state().inmem.backup_lsn, backup_lsn);
739 0 : // we should check whether to shut down offloader, but this will be done
740 0 : // soon by peer communication anyway.
741 0 : Ok(())
742 0 : }
743 :
744 : /// Get safekeeper info for broadcasting to broker and other peers.
745 0 : pub async fn get_safekeeper_info(&self, conf: &SafeKeeperConf) -> SafekeeperTimelineInfo {
746 0 : let standby_apply_lsn = self.walsenders.get_hotstandby().reply.apply_lsn;
747 0 : let shared_state = self.read_shared_state().await;
748 0 : shared_state.get_safekeeper_info(&self.ttid, conf, standby_apply_lsn)
749 0 : }
750 :
751 : /// Update timeline state with peer safekeeper data.
752 0 : pub async fn record_safekeeper_info(
753 0 : self: &Arc<Self>,
754 0 : sk_info: SafekeeperTimelineInfo,
755 0 : ) -> Result<()> {
756 : {
757 0 : let mut shared_state = self.write_shared_state().await;
758 0 : shared_state.sk.record_safekeeper_info(&sk_info).await?;
759 0 : let peer_info = PeerInfo::from_sk_info(&sk_info, Instant::now());
760 0 : shared_state.peers_info.upsert(&peer_info);
761 0 : }
762 0 : Ok(())
763 0 : }
764 :
765 0 : pub async fn get_peers(&self, conf: &SafeKeeperConf) -> Vec<PeerInfo> {
766 0 : let shared_state = self.read_shared_state().await;
767 0 : shared_state.get_peers(conf.heartbeat_timeout)
768 0 : }
769 :
770 0 : pub fn get_walsenders(&self) -> &Arc<WalSenders> {
771 0 : &self.walsenders
772 0 : }
773 :
774 0 : pub fn get_walreceivers(&self) -> &Arc<WalReceivers> {
775 0 : &self.walreceivers
776 0 : }
777 :
778 : /// Returns flush_lsn.
779 0 : pub async fn get_flush_lsn(&self) -> Lsn {
780 0 : self.read_shared_state().await.sk.flush_lsn()
781 0 : }
782 :
783 : /// Gather timeline data for metrics.
784 0 : pub async fn info_for_metrics(&self) -> Option<FullTimelineInfo> {
785 0 : if self.is_cancelled() {
786 0 : return None;
787 0 : }
788 0 :
789 0 : let (ps_feedback_count, last_ps_feedback) = self.walsenders.get_ps_feedback_stats();
790 0 : let state = self.read_shared_state().await;
791 0 : Some(FullTimelineInfo {
792 0 : ttid: self.ttid,
793 0 : ps_feedback_count,
794 0 : last_ps_feedback,
795 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
796 0 : timeline_is_active: self.broker_active.load(Ordering::Relaxed),
797 0 : num_computes: self.walreceivers.get_num() as u32,
798 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
799 0 : epoch_start_lsn: state.sk.term_start_lsn(),
800 0 : mem_state: state.sk.state().inmem.clone(),
801 0 : persisted_state: TimelinePersistentState::clone(state.sk.state()),
802 0 : flush_lsn: state.sk.flush_lsn(),
803 0 : wal_storage: state.sk.wal_storage_metrics(),
804 0 : })
805 0 : }
806 :
807 : /// Returns in-memory timeline state to build a full debug dump.
808 0 : pub async fn memory_dump(&self) -> debug_dump::Memory {
809 0 : let state = self.read_shared_state().await;
810 :
811 0 : let (write_lsn, write_record_lsn, flush_lsn, file_open) =
812 0 : state.sk.wal_storage_internal_state();
813 0 :
814 0 : debug_dump::Memory {
815 0 : is_cancelled: self.is_cancelled(),
816 0 : peers_info_len: state.peers_info.0.len(),
817 0 : walsenders: self.walsenders.get_all(),
818 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
819 0 : active: self.broker_active.load(Ordering::Relaxed),
820 0 : num_computes: self.walreceivers.get_num() as u32,
821 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
822 0 : epoch_start_lsn: state.sk.term_start_lsn(),
823 0 : mem_state: state.sk.state().inmem.clone(),
824 0 : mgr_status: self.mgr_status.get(),
825 0 : write_lsn,
826 0 : write_record_lsn,
827 0 : flush_lsn,
828 0 : file_open,
829 0 : }
830 0 : }
831 :
832 : /// Apply a function to the control file state and persist it.
833 0 : pub async fn map_control_file<T>(
834 0 : self: &Arc<Self>,
835 0 : f: impl FnOnce(&mut TimelinePersistentState) -> Result<T>,
836 0 : ) -> Result<T> {
837 0 : let mut state = self.write_shared_state().await;
838 0 : let mut persistent_state = state.sk.state_mut().start_change();
839 : // If f returns error, we abort the change and don't persist anything.
840 0 : let res = f(&mut persistent_state)?;
841 : // If persisting fails, we abort the change and return error.
842 0 : state
843 0 : .sk
844 0 : .state_mut()
845 0 : .finish_change(&persistent_state)
846 0 : .await?;
847 0 : Ok(res)
848 0 : }
849 :
850 : /// Get the timeline guard for reading/writing WAL files.
851 : /// If WAL files are not present on disk (evicted), they will be automatically
852 : /// downloaded from remote storage. This is done in the manager task, which is
853 : /// responsible for issuing all guards.
854 : ///
855 : /// NB: don't use this function from timeline_manager, it will deadlock.
856 : /// NB: don't use this function while holding shared_state lock.
857 0 : pub async fn wal_residence_guard(self: &Arc<Self>) -> Result<WalResidentTimeline> {
858 0 : if self.is_cancelled() {
859 0 : bail!(TimelineError::Cancelled(self.ttid));
860 0 : }
861 0 :
862 0 : debug!("requesting WalResidentTimeline guard");
863 0 : let started_at = Instant::now();
864 0 : let status_before = self.mgr_status.get();
865 :
866 : // Wait 30 seconds for the guard to be acquired. It can time out if someone is
867 : // holding the lock (e.g. during `SafeKeeper::process_msg()`) or manager task
868 : // is stuck.
869 0 : let res = tokio::time::timeout_at(
870 0 : started_at + Duration::from_secs(30),
871 0 : self.manager_ctl.wal_residence_guard(),
872 0 : )
873 0 : .await;
874 :
875 0 : let guard = match res {
876 0 : Ok(Ok(guard)) => {
877 0 : let finished_at = Instant::now();
878 0 : let elapsed = finished_at - started_at;
879 0 : MISC_OPERATION_SECONDS
880 0 : .with_label_values(&["wal_residence_guard"])
881 0 : .observe(elapsed.as_secs_f64());
882 0 :
883 0 : guard
884 : }
885 0 : Ok(Err(e)) => {
886 0 : warn!(
887 0 : "error while acquiring WalResidentTimeline guard, statuses {:?} => {:?}",
888 0 : status_before,
889 0 : self.mgr_status.get()
890 : );
891 0 : return Err(e);
892 : }
893 : Err(_) => {
894 0 : warn!(
895 0 : "timeout while acquiring WalResidentTimeline guard, statuses {:?} => {:?}",
896 0 : status_before,
897 0 : self.mgr_status.get()
898 : );
899 0 : anyhow::bail!("timeout while acquiring WalResidentTimeline guard");
900 : }
901 : };
902 :
903 0 : Ok(WalResidentTimeline::new(self.clone(), guard))
904 0 : }
905 : }
906 :
907 : /// This is a guard that allows to read/write disk timeline state.
908 : /// All tasks that are trying to read/write WAL from disk should use this guard.
909 : pub struct WalResidentTimeline {
910 : pub tli: Arc<Timeline>,
911 : _guard: ResidenceGuard,
912 : }
913 :
914 : impl WalResidentTimeline {
915 0 : pub fn new(tli: Arc<Timeline>, _guard: ResidenceGuard) -> Self {
916 0 : WalResidentTimeline { tli, _guard }
917 0 : }
918 : }
919 :
920 : impl Deref for WalResidentTimeline {
921 : type Target = Arc<Timeline>;
922 :
923 0 : fn deref(&self) -> &Self::Target {
924 0 : &self.tli
925 0 : }
926 : }
927 :
928 : impl WalResidentTimeline {
929 : /// Returns true if walsender should stop sending WAL to pageserver. We
930 : /// terminate it if remote_consistent_lsn reached commit_lsn and there is no
931 : /// computes. While there might be nothing to stream already, we learn about
932 : /// remote_consistent_lsn update through replication feedback, and we want
933 : /// to stop pushing to the broker if pageserver is fully caughtup.
934 0 : pub async fn should_walsender_stop(&self, reported_remote_consistent_lsn: Lsn) -> bool {
935 0 : if self.is_cancelled() {
936 0 : return true;
937 0 : }
938 0 : let shared_state = self.read_shared_state().await;
939 0 : if self.walreceivers.get_num() == 0 {
940 0 : return shared_state.sk.state().inmem.commit_lsn == Lsn(0) || // no data at all yet
941 0 : reported_remote_consistent_lsn >= shared_state.sk.state().inmem.commit_lsn;
942 0 : }
943 0 : false
944 0 : }
945 :
946 : /// Ensure that current term is t, erroring otherwise, and lock the state.
947 0 : pub async fn acquire_term(&self, t: Term) -> Result<ReadGuardSharedState> {
948 0 : let ss = self.read_shared_state().await;
949 0 : if ss.sk.state().acceptor_state.term != t {
950 0 : bail!(
951 0 : "failed to acquire term {}, current term {}",
952 0 : t,
953 0 : ss.sk.state().acceptor_state.term
954 0 : );
955 0 : }
956 0 : Ok(ss)
957 0 : }
958 :
959 : /// Pass arrived message to the safekeeper.
960 0 : pub async fn process_msg(
961 0 : &self,
962 0 : msg: &ProposerAcceptorMessage,
963 0 : ) -> Result<Option<AcceptorProposerMessage>> {
964 0 : if self.is_cancelled() {
965 0 : bail!(TimelineError::Cancelled(self.ttid));
966 0 : }
967 :
968 : let mut rmsg: Option<AcceptorProposerMessage>;
969 : {
970 0 : let mut shared_state = self.write_shared_state().await;
971 0 : rmsg = shared_state.sk.safekeeper().process_msg(msg).await?;
972 :
973 : // if this is AppendResponse, fill in proper hot standby feedback.
974 0 : if let Some(AcceptorProposerMessage::AppendResponse(ref mut resp)) = rmsg {
975 0 : resp.hs_feedback = self.walsenders.get_hotstandby().hs_feedback;
976 0 : }
977 : }
978 0 : Ok(rmsg)
979 0 : }
980 :
981 0 : pub async fn get_walreader(&self, start_lsn: Lsn) -> Result<WalReader> {
982 0 : let (_, persisted_state) = self.get_state().await;
983 0 : let enable_remote_read = GlobalTimelines::get_global_config().is_wal_backup_enabled();
984 0 :
985 0 : WalReader::new(
986 0 : &self.ttid,
987 0 : self.timeline_dir.clone(),
988 0 : &persisted_state,
989 0 : start_lsn,
990 0 : enable_remote_read,
991 0 : )
992 0 : }
993 :
994 0 : pub fn get_timeline_dir(&self) -> Utf8PathBuf {
995 0 : self.timeline_dir.clone()
996 0 : }
997 :
998 : /// Update in memory remote consistent lsn.
999 0 : pub async fn update_remote_consistent_lsn(&self, candidate: Lsn) {
1000 0 : let mut shared_state = self.write_shared_state().await;
1001 0 : shared_state.sk.state_mut().inmem.remote_consistent_lsn = max(
1002 0 : shared_state.sk.state().inmem.remote_consistent_lsn,
1003 0 : candidate,
1004 0 : );
1005 0 : }
1006 : }
1007 :
1008 : /// This struct contains methods that are used by timeline manager task.
1009 : pub(crate) struct ManagerTimeline {
1010 : pub(crate) tli: Arc<Timeline>,
1011 : }
1012 :
1013 : impl Deref for ManagerTimeline {
1014 : type Target = Arc<Timeline>;
1015 :
1016 0 : fn deref(&self) -> &Self::Target {
1017 0 : &self.tli
1018 0 : }
1019 : }
1020 :
1021 : impl ManagerTimeline {
1022 0 : pub(crate) fn timeline_dir(&self) -> &Utf8PathBuf {
1023 0 : &self.tli.timeline_dir
1024 0 : }
1025 :
1026 : /// Manager requests this state on startup.
1027 0 : pub(crate) async fn bootstrap_mgr(&self) -> (bool, Option<PartialRemoteSegment>) {
1028 0 : let shared_state = self.read_shared_state().await;
1029 0 : let is_offloaded = matches!(
1030 0 : shared_state.sk.state().eviction_state,
1031 : EvictionState::Offloaded(_)
1032 : );
1033 0 : let partial_backup_uploaded = shared_state.sk.state().partial_backup.uploaded_segment();
1034 0 :
1035 0 : (is_offloaded, partial_backup_uploaded)
1036 0 : }
1037 :
1038 : /// Try to switch state Present->Offloaded.
1039 0 : pub(crate) async fn switch_to_offloaded(
1040 0 : &self,
1041 0 : partial: &PartialRemoteSegment,
1042 0 : ) -> anyhow::Result<()> {
1043 0 : let mut shared = self.write_shared_state().await;
1044 :
1045 : // updating control file
1046 0 : let mut pstate = shared.sk.state_mut().start_change();
1047 :
1048 0 : if !matches!(pstate.eviction_state, EvictionState::Present) {
1049 0 : bail!(
1050 0 : "cannot switch to offloaded state, current state is {:?}",
1051 0 : pstate.eviction_state
1052 0 : );
1053 0 : }
1054 0 :
1055 0 : if partial.flush_lsn != shared.sk.flush_lsn() {
1056 0 : bail!(
1057 0 : "flush_lsn mismatch in partial backup, expected {}, got {}",
1058 0 : shared.sk.flush_lsn(),
1059 0 : partial.flush_lsn
1060 0 : );
1061 0 : }
1062 0 :
1063 0 : if partial.commit_lsn != pstate.commit_lsn {
1064 0 : bail!(
1065 0 : "commit_lsn mismatch in partial backup, expected {}, got {}",
1066 0 : pstate.commit_lsn,
1067 0 : partial.commit_lsn
1068 0 : );
1069 0 : }
1070 0 :
1071 0 : if partial.term != shared.sk.last_log_term() {
1072 0 : bail!(
1073 0 : "term mismatch in partial backup, expected {}, got {}",
1074 0 : shared.sk.last_log_term(),
1075 0 : partial.term
1076 0 : );
1077 0 : }
1078 0 :
1079 0 : pstate.eviction_state = EvictionState::Offloaded(shared.sk.flush_lsn());
1080 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1081 : // control file is now switched to Offloaded state
1082 :
1083 : // now we can switch shared.sk to Offloaded, shouldn't fail
1084 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1085 0 : let cfile_state = prev_sk.take_state();
1086 0 : shared.sk = StateSK::Offloaded(Box::new(cfile_state));
1087 0 :
1088 0 : Ok(())
1089 0 : }
1090 :
1091 : /// Try to switch state Offloaded->Present.
1092 0 : pub(crate) async fn switch_to_present(&self) -> anyhow::Result<()> {
1093 0 : let conf = GlobalTimelines::get_global_config();
1094 0 : let mut shared = self.write_shared_state().await;
1095 :
1096 : // trying to restore WAL storage
1097 0 : let wal_store = wal_storage::PhysicalStorage::new(
1098 0 : &self.ttid,
1099 0 : self.timeline_dir.clone(),
1100 0 : &conf,
1101 0 : shared.sk.state(),
1102 0 : )?;
1103 :
1104 : // updating control file
1105 0 : let mut pstate = shared.sk.state_mut().start_change();
1106 :
1107 0 : if !matches!(pstate.eviction_state, EvictionState::Offloaded(_)) {
1108 0 : bail!(
1109 0 : "cannot switch to present state, current state is {:?}",
1110 0 : pstate.eviction_state
1111 0 : );
1112 0 : }
1113 0 :
1114 0 : if wal_store.flush_lsn() != shared.sk.flush_lsn() {
1115 0 : bail!(
1116 0 : "flush_lsn mismatch in restored WAL, expected {}, got {}",
1117 0 : shared.sk.flush_lsn(),
1118 0 : wal_store.flush_lsn()
1119 0 : );
1120 0 : }
1121 0 :
1122 0 : pstate.eviction_state = EvictionState::Present;
1123 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1124 :
1125 : // now we can switch shared.sk to Present, shouldn't fail
1126 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1127 0 : let cfile_state = prev_sk.take_state();
1128 0 : shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, conf.my_id)?);
1129 :
1130 0 : Ok(())
1131 0 : }
1132 :
1133 : /// Update current manager state, useful for debugging manager deadlocks.
1134 0 : pub(crate) fn set_status(&self, status: timeline_manager::Status) {
1135 0 : self.mgr_status.store(status, Ordering::Relaxed);
1136 0 : }
1137 : }
1138 :
1139 : /// Deletes directory and it's contents. Returns false if directory does not exist.
1140 0 : async fn delete_dir(path: &Utf8PathBuf) -> Result<bool> {
1141 0 : match fs::remove_dir_all(path).await {
1142 0 : Ok(_) => Ok(true),
1143 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(false),
1144 0 : Err(e) => Err(e.into()),
1145 : }
1146 0 : }
1147 :
1148 : /// Get a path to the tenant directory. If you just need to get a timeline directory,
1149 : /// use WalResidentTimeline::get_timeline_dir instead.
1150 14 : pub(crate) fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
1151 14 : conf.workdir.join(tenant_id.to_string())
1152 14 : }
1153 :
1154 : /// Get a path to the timeline directory. If you need to read WAL files from disk,
1155 : /// use WalResidentTimeline::get_timeline_dir instead. This function does not check
1156 : /// timeline eviction status and WAL files might not be present on disk.
1157 14 : pub(crate) fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
1158 14 : get_tenant_dir(conf, &ttid.tenant_id).join(ttid.timeline_id.to_string())
1159 14 : }
|
