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 std::cmp::max;
5 : use std::ops::{Deref, DerefMut};
6 : use std::sync::Arc;
7 : use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
8 : use std::time::Duration;
9 :
10 : use anyhow::{Result, anyhow, bail};
11 : use camino::{Utf8Path, Utf8PathBuf};
12 : use http_utils::error::ApiError;
13 : use remote_storage::RemotePath;
14 : use safekeeper_api::Term;
15 : use safekeeper_api::membership::Configuration;
16 : use safekeeper_api::models::{
17 : PeerInfo, TimelineMembershipSwitchResponse, TimelineTermBumpResponse,
18 : };
19 : use storage_broker::proto::{SafekeeperTimelineInfo, TenantTimelineId as ProtoTenantTimelineId};
20 : use tokio::fs::{self};
21 : use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard, watch};
22 : use tokio::time::Instant;
23 : use tokio_util::sync::CancellationToken;
24 : use tracing::*;
25 : use utils::id::{NodeId, TenantId, TenantTimelineId};
26 : use utils::lsn::Lsn;
27 : use utils::sync::gate::Gate;
28 :
29 : use crate::metrics::{FullTimelineInfo, MISC_OPERATION_SECONDS, WalStorageMetrics};
30 : use crate::rate_limit::RateLimiter;
31 : use crate::receive_wal::WalReceivers;
32 : use crate::safekeeper::{AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, TermLsn};
33 : use crate::send_wal::{WalSenders, WalSendersTimelineMetricValues};
34 : use crate::state::{EvictionState, TimelineMemState, TimelinePersistentState, TimelineState};
35 : use crate::timeline_guard::ResidenceGuard;
36 : use crate::timeline_manager::{AtomicStatus, ManagerCtl};
37 : use crate::timelines_set::TimelinesSet;
38 : use crate::wal_backup;
39 : use crate::wal_backup::{WalBackup, remote_timeline_path};
40 : use crate::wal_backup_partial::PartialRemoteSegment;
41 : use crate::wal_storage::{Storage as wal_storage_iface, WalReader};
42 : use crate::{SafeKeeperConf, control_file, debug_dump, timeline_manager, wal_storage};
43 :
44 0 : fn peer_info_from_sk_info(sk_info: &SafekeeperTimelineInfo, ts: Instant) -> PeerInfo {
45 0 : PeerInfo {
46 0 : sk_id: NodeId(sk_info.safekeeper_id),
47 0 : term: sk_info.term,
48 0 : last_log_term: sk_info.last_log_term,
49 0 : flush_lsn: Lsn(sk_info.flush_lsn),
50 0 : commit_lsn: Lsn(sk_info.commit_lsn),
51 0 : local_start_lsn: Lsn(sk_info.local_start_lsn),
52 0 : pg_connstr: sk_info.safekeeper_connstr.clone(),
53 0 : http_connstr: sk_info.http_connstr.clone(),
54 0 : https_connstr: sk_info.https_connstr.clone(),
55 0 : ts,
56 0 : }
57 0 : }
58 :
59 : // vector-based node id -> peer state map with very limited functionality we
60 : // need.
61 : #[derive(Debug, Clone, Default)]
62 : pub struct PeersInfo(pub Vec<PeerInfo>);
63 :
64 : impl PeersInfo {
65 0 : fn get(&mut self, id: NodeId) -> Option<&mut PeerInfo> {
66 0 : self.0.iter_mut().find(|p| p.sk_id == id)
67 0 : }
68 :
69 0 : fn upsert(&mut self, p: &PeerInfo) {
70 0 : match self.get(p.sk_id) {
71 0 : Some(rp) => *rp = p.clone(),
72 0 : None => self.0.push(p.clone()),
73 : }
74 0 : }
75 : }
76 :
77 : pub type ReadGuardSharedState<'a> = RwLockReadGuard<'a, SharedState>;
78 :
79 : /// WriteGuardSharedState is a wrapper around `RwLockWriteGuard<SharedState>` that
80 : /// automatically updates `watch::Sender` channels with state on drop.
81 : pub struct WriteGuardSharedState<'a> {
82 : tli: Arc<Timeline>,
83 : guard: RwLockWriteGuard<'a, SharedState>,
84 : }
85 :
86 : impl<'a> WriteGuardSharedState<'a> {
87 1245 : fn new(tli: Arc<Timeline>, guard: RwLockWriteGuard<'a, SharedState>) -> Self {
88 1245 : WriteGuardSharedState { tli, guard }
89 1245 : }
90 : }
91 :
92 : impl Deref for WriteGuardSharedState<'_> {
93 : type Target = SharedState;
94 :
95 0 : fn deref(&self) -> &Self::Target {
96 0 : &self.guard
97 0 : }
98 : }
99 :
100 : impl DerefMut for WriteGuardSharedState<'_> {
101 1240 : fn deref_mut(&mut self) -> &mut Self::Target {
102 1240 : &mut self.guard
103 1240 : }
104 : }
105 :
106 : impl Drop for WriteGuardSharedState<'_> {
107 1245 : fn drop(&mut self) {
108 1245 : let term_flush_lsn =
109 1245 : TermLsn::from((self.guard.sk.last_log_term(), self.guard.sk.flush_lsn()));
110 1245 : let commit_lsn = self.guard.sk.state().inmem.commit_lsn;
111 1245 :
112 1245 : let _ = self.tli.term_flush_lsn_watch_tx.send_if_modified(|old| {
113 1245 : if *old != term_flush_lsn {
114 620 : *old = term_flush_lsn;
115 620 : true
116 : } else {
117 625 : false
118 : }
119 1245 : });
120 1245 :
121 1245 : let _ = self.tli.commit_lsn_watch_tx.send_if_modified(|old| {
122 1245 : if *old != commit_lsn {
123 615 : *old = commit_lsn;
124 615 : true
125 : } else {
126 630 : false
127 : }
128 1245 : });
129 1245 :
130 1245 : // send notification about shared state update
131 1245 : self.tli.shared_state_version_tx.send_modify(|old| {
132 1245 : *old += 1;
133 1245 : });
134 1245 : }
135 : }
136 :
137 : /// This structure is stored in shared state and represents the state of the timeline.
138 : ///
139 : /// Usually it holds SafeKeeper, but it also supports offloaded timeline state. In this
140 : /// case, SafeKeeper is not available (because WAL is not present on disk) and all
141 : /// operations can be done only with control file.
142 : #[allow(clippy::large_enum_variant, reason = "TODO")]
143 : pub enum StateSK {
144 : Loaded(SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>),
145 : Offloaded(Box<TimelineState<control_file::FileStorage>>),
146 : // Not used, required for moving between states.
147 : Empty,
148 : }
149 :
150 : impl StateSK {
151 2578 : pub fn flush_lsn(&self) -> Lsn {
152 2578 : match self {
153 2578 : StateSK::Loaded(sk) => sk.wal_store.flush_lsn(),
154 0 : StateSK::Offloaded(state) => match state.eviction_state {
155 0 : EvictionState::Offloaded(flush_lsn) => flush_lsn,
156 0 : _ => panic!("StateSK::Offloaded mismatches with eviction_state from control_file"),
157 : },
158 0 : StateSK::Empty => unreachable!(),
159 : }
160 2578 : }
161 :
162 : /// Get a reference to the control file's timeline state.
163 2611 : pub fn state(&self) -> &TimelineState<control_file::FileStorage> {
164 2611 : match self {
165 2611 : StateSK::Loaded(sk) => &sk.state,
166 0 : StateSK::Offloaded(s) => s,
167 0 : StateSK::Empty => unreachable!(),
168 : }
169 2611 : }
170 :
171 0 : pub fn state_mut(&mut self) -> &mut TimelineState<control_file::FileStorage> {
172 0 : match self {
173 0 : StateSK::Loaded(sk) => &mut sk.state,
174 0 : StateSK::Offloaded(s) => s,
175 0 : StateSK::Empty => unreachable!(),
176 : }
177 0 : }
178 :
179 1289 : pub fn last_log_term(&self) -> Term {
180 1289 : self.state()
181 1289 : .acceptor_state
182 1289 : .get_last_log_term(self.flush_lsn())
183 1289 : }
184 :
185 0 : pub async fn term_bump(&mut self, to: Option<Term>) -> Result<TimelineTermBumpResponse> {
186 0 : self.state_mut().term_bump(to).await
187 0 : }
188 :
189 0 : pub async fn membership_switch(
190 0 : &mut self,
191 0 : to: Configuration,
192 0 : ) -> Result<TimelineMembershipSwitchResponse> {
193 0 : self.state_mut().membership_switch(to).await
194 0 : }
195 :
196 : /// Close open WAL files to release FDs.
197 0 : fn close_wal_store(&mut self) {
198 0 : if let StateSK::Loaded(sk) = self {
199 0 : sk.wal_store.close();
200 0 : }
201 0 : }
202 :
203 : /// Update timeline state with peer safekeeper data.
204 0 : pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> {
205 0 : // update commit_lsn if safekeeper is loaded
206 0 : match self {
207 0 : StateSK::Loaded(sk) => sk.record_safekeeper_info(sk_info).await?,
208 0 : StateSK::Offloaded(_) => {}
209 0 : StateSK::Empty => unreachable!(),
210 : }
211 :
212 : // update everything else, including remote_consistent_lsn and backup_lsn
213 0 : let mut sync_control_file = false;
214 0 : let state = self.state_mut();
215 0 : let wal_seg_size = state.server.wal_seg_size as u64;
216 0 :
217 0 : state.inmem.backup_lsn = max(Lsn(sk_info.backup_lsn), state.inmem.backup_lsn);
218 0 : sync_control_file |= state.backup_lsn + wal_seg_size < state.inmem.backup_lsn;
219 0 :
220 0 : state.inmem.remote_consistent_lsn = max(
221 0 : Lsn(sk_info.remote_consistent_lsn),
222 0 : state.inmem.remote_consistent_lsn,
223 0 : );
224 0 : sync_control_file |=
225 0 : state.remote_consistent_lsn + wal_seg_size < state.inmem.remote_consistent_lsn;
226 0 :
227 0 : state.inmem.peer_horizon_lsn =
228 0 : max(Lsn(sk_info.peer_horizon_lsn), state.inmem.peer_horizon_lsn);
229 0 : sync_control_file |= state.peer_horizon_lsn + wal_seg_size < state.inmem.peer_horizon_lsn;
230 0 :
231 0 : if sync_control_file {
232 0 : state.flush().await?;
233 0 : }
234 0 : Ok(())
235 0 : }
236 :
237 : /// Previously known as epoch_start_lsn. Needed only for reference in some APIs.
238 0 : pub fn term_start_lsn(&self) -> Lsn {
239 0 : match self {
240 0 : StateSK::Loaded(sk) => sk.term_start_lsn,
241 0 : StateSK::Offloaded(_) => Lsn(0),
242 0 : StateSK::Empty => unreachable!(),
243 : }
244 0 : }
245 :
246 : /// Used for metrics only.
247 0 : pub fn wal_storage_metrics(&self) -> WalStorageMetrics {
248 0 : match self {
249 0 : StateSK::Loaded(sk) => sk.wal_store.get_metrics(),
250 0 : StateSK::Offloaded(_) => WalStorageMetrics::default(),
251 0 : StateSK::Empty => unreachable!(),
252 : }
253 0 : }
254 :
255 : /// Returns WAL storage internal LSNs for debug dump.
256 0 : pub fn wal_storage_internal_state(&self) -> (Lsn, Lsn, Lsn, bool) {
257 0 : match self {
258 0 : StateSK::Loaded(sk) => sk.wal_store.internal_state(),
259 : StateSK::Offloaded(_) => {
260 0 : let flush_lsn = self.flush_lsn();
261 0 : (flush_lsn, flush_lsn, flush_lsn, false)
262 : }
263 0 : StateSK::Empty => unreachable!(),
264 : }
265 0 : }
266 :
267 : /// Access to SafeKeeper object. Panics if offloaded, should be good to use from WalResidentTimeline.
268 1240 : pub fn safekeeper(
269 1240 : &mut self,
270 1240 : ) -> &mut SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage> {
271 1240 : match self {
272 1240 : StateSK::Loaded(sk) => sk,
273 : StateSK::Offloaded(_) => {
274 0 : panic!("safekeeper is offloaded, cannot be used")
275 : }
276 0 : StateSK::Empty => unreachable!(),
277 : }
278 1240 : }
279 :
280 : /// Moves control file's state structure out of the enum. Used to switch states.
281 0 : fn take_state(self) -> TimelineState<control_file::FileStorage> {
282 0 : match self {
283 0 : StateSK::Loaded(sk) => sk.state,
284 0 : StateSK::Offloaded(state) => *state,
285 0 : StateSK::Empty => unreachable!(),
286 : }
287 0 : }
288 : }
289 :
290 : /// Shared state associated with database instance
291 : pub struct SharedState {
292 : /// Safekeeper object
293 : pub(crate) sk: StateSK,
294 : /// In memory list containing state of peers sent in latest messages from them.
295 : pub(crate) peers_info: PeersInfo,
296 : // True value hinders old WAL removal; this is used by snapshotting. We
297 : // could make it a counter, but there is no need to.
298 : pub(crate) wal_removal_on_hold: bool,
299 : }
300 :
301 : impl SharedState {
302 : /// Creates a new SharedState.
303 5 : pub fn new(sk: StateSK) -> Self {
304 5 : Self {
305 5 : sk,
306 5 : peers_info: PeersInfo(vec![]),
307 5 : wal_removal_on_hold: false,
308 5 : }
309 5 : }
310 :
311 : /// Restore SharedState from control file. If file doesn't exist, bails out.
312 0 : pub fn restore(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Result<Self> {
313 0 : let timeline_dir = get_timeline_dir(conf, ttid);
314 0 : let control_store = control_file::FileStorage::restore_new(&timeline_dir, conf.no_sync)?;
315 0 : if control_store.server.wal_seg_size == 0 {
316 0 : bail!(TimelineError::UninitializedWalSegSize(*ttid));
317 0 : }
318 :
319 0 : let sk = match control_store.eviction_state {
320 : EvictionState::Present => {
321 0 : let wal_store = wal_storage::PhysicalStorage::new(
322 0 : ttid,
323 0 : &timeline_dir,
324 0 : &control_store,
325 0 : conf.no_sync,
326 0 : )?;
327 0 : StateSK::Loaded(SafeKeeper::new(
328 0 : TimelineState::new(control_store),
329 0 : wal_store,
330 0 : conf.my_id,
331 0 : )?)
332 : }
333 : EvictionState::Offloaded(_) => {
334 0 : StateSK::Offloaded(Box::new(TimelineState::new(control_store)))
335 : }
336 : };
337 :
338 0 : Ok(Self::new(sk))
339 0 : }
340 :
341 5 : pub(crate) fn get_wal_seg_size(&self) -> usize {
342 5 : self.sk.state().server.wal_seg_size as usize
343 5 : }
344 :
345 0 : fn get_safekeeper_info(
346 0 : &self,
347 0 : ttid: &TenantTimelineId,
348 0 : conf: &SafeKeeperConf,
349 0 : standby_apply_lsn: Lsn,
350 0 : ) -> SafekeeperTimelineInfo {
351 0 : SafekeeperTimelineInfo {
352 0 : safekeeper_id: conf.my_id.0,
353 0 : tenant_timeline_id: Some(ProtoTenantTimelineId {
354 0 : tenant_id: ttid.tenant_id.as_ref().to_owned(),
355 0 : timeline_id: ttid.timeline_id.as_ref().to_owned(),
356 0 : }),
357 0 : term: self.sk.state().acceptor_state.term,
358 0 : last_log_term: self.sk.last_log_term(),
359 0 : flush_lsn: self.sk.flush_lsn().0,
360 0 : // note: this value is not flushed to control file yet and can be lost
361 0 : commit_lsn: self.sk.state().inmem.commit_lsn.0,
362 0 : remote_consistent_lsn: self.sk.state().inmem.remote_consistent_lsn.0,
363 0 : peer_horizon_lsn: self.sk.state().inmem.peer_horizon_lsn.0,
364 0 : safekeeper_connstr: conf
365 0 : .advertise_pg_addr
366 0 : .to_owned()
367 0 : .unwrap_or(conf.listen_pg_addr.clone()),
368 0 : http_connstr: conf.listen_http_addr.to_owned(),
369 0 : https_connstr: conf.listen_https_addr.to_owned(),
370 0 : backup_lsn: self.sk.state().inmem.backup_lsn.0,
371 0 : local_start_lsn: self.sk.state().local_start_lsn.0,
372 0 : availability_zone: conf.availability_zone.clone(),
373 0 : standby_horizon: standby_apply_lsn.0,
374 0 : }
375 0 : }
376 :
377 : /// Get our latest view of alive peers status on the timeline.
378 : /// We pass our own info through the broker as well, so when we don't have connection
379 : /// to the broker returned vec is empty.
380 39 : pub(crate) fn get_peers(&self, heartbeat_timeout: Duration) -> Vec<PeerInfo> {
381 39 : let now = Instant::now();
382 39 : self.peers_info
383 39 : .0
384 39 : .iter()
385 39 : // Regard peer as absent if we haven't heard from it within heartbeat_timeout.
386 39 : .filter(|p| now.duration_since(p.ts) <= heartbeat_timeout)
387 39 : .cloned()
388 39 : .collect()
389 39 : }
390 : }
391 :
392 : #[derive(Debug, thiserror::Error)]
393 : pub enum TimelineError {
394 : #[error("Timeline {0} was cancelled and cannot be used anymore")]
395 : Cancelled(TenantTimelineId),
396 : #[error("Timeline {0} was not found in global map")]
397 : NotFound(TenantTimelineId),
398 : #[error("Timeline {0} creation is in progress")]
399 : CreationInProgress(TenantTimelineId),
400 : #[error("Timeline {0} exists on disk, but wasn't loaded on startup")]
401 : Invalid(TenantTimelineId),
402 : #[error("Timeline {0} is already exists")]
403 : AlreadyExists(TenantTimelineId),
404 : #[error("Timeline {0} is not initialized, wal_seg_size is zero")]
405 : UninitializedWalSegSize(TenantTimelineId),
406 : #[error("Timeline {0} is not initialized, pg_version is unknown")]
407 : UninitialinzedPgVersion(TenantTimelineId),
408 : }
409 :
410 : // Convert to HTTP API error.
411 : impl From<TimelineError> for ApiError {
412 0 : fn from(te: TimelineError) -> ApiError {
413 0 : match te {
414 0 : TimelineError::NotFound(ttid) => {
415 0 : ApiError::NotFound(anyhow!("timeline {} not found", ttid).into())
416 : }
417 0 : _ => ApiError::InternalServerError(anyhow!("{}", te)),
418 : }
419 0 : }
420 : }
421 :
422 : /// We run remote deletion in a background task, this is how it sends its results back.
423 : type RemoteDeletionReceiver = tokio::sync::watch::Receiver<Option<anyhow::Result<()>>>;
424 :
425 : /// Timeline struct manages lifecycle (creation, deletion, restore) of a safekeeper timeline.
426 : /// It also holds SharedState and provides mutually exclusive access to it.
427 : pub struct Timeline {
428 : pub ttid: TenantTimelineId,
429 : pub remote_path: RemotePath,
430 :
431 : /// Used to broadcast commit_lsn updates to all background jobs.
432 : commit_lsn_watch_tx: watch::Sender<Lsn>,
433 : commit_lsn_watch_rx: watch::Receiver<Lsn>,
434 :
435 : /// Broadcasts (current term, flush_lsn) updates, walsender is interested in
436 : /// them when sending in recovery mode (to walproposer or peers). Note: this
437 : /// is just a notification, WAL reading should always done with lock held as
438 : /// term can change otherwise.
439 : term_flush_lsn_watch_tx: watch::Sender<TermLsn>,
440 : term_flush_lsn_watch_rx: watch::Receiver<TermLsn>,
441 :
442 : /// Broadcasts shared state updates.
443 : shared_state_version_tx: watch::Sender<usize>,
444 : shared_state_version_rx: watch::Receiver<usize>,
445 :
446 : /// Safekeeper and other state, that should remain consistent and
447 : /// synchronized with the disk. This is tokio mutex as we write WAL to disk
448 : /// while holding it, ensuring that consensus checks are in order.
449 : mutex: RwLock<SharedState>,
450 : walsenders: Arc<WalSenders>,
451 : walreceivers: Arc<WalReceivers>,
452 : timeline_dir: Utf8PathBuf,
453 : manager_ctl: ManagerCtl,
454 : conf: Arc<SafeKeeperConf>,
455 :
456 : pub(crate) wal_backup: Arc<WalBackup>,
457 :
458 : remote_deletion: std::sync::Mutex<Option<RemoteDeletionReceiver>>,
459 :
460 : /// Hold this gate from code that depends on the Timeline's non-shut-down state. While holding
461 : /// this gate, you must respect [`Timeline::cancel`]
462 : pub(crate) gate: Gate,
463 :
464 : /// Delete/cancel will trigger this, background tasks should drop out as soon as it fires
465 : pub(crate) cancel: CancellationToken,
466 :
467 : // timeline_manager controlled state
468 : pub(crate) broker_active: AtomicBool,
469 : pub(crate) wal_backup_active: AtomicBool,
470 : pub(crate) last_removed_segno: AtomicU64,
471 : pub(crate) mgr_status: AtomicStatus,
472 : }
473 :
474 : impl Timeline {
475 : /// Constructs a new timeline.
476 5 : pub fn new(
477 5 : ttid: TenantTimelineId,
478 5 : timeline_dir: &Utf8Path,
479 5 : remote_path: &RemotePath,
480 5 : shared_state: SharedState,
481 5 : conf: Arc<SafeKeeperConf>,
482 5 : wal_backup: Arc<WalBackup>,
483 5 : ) -> Arc<Self> {
484 5 : let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
485 5 : watch::channel(shared_state.sk.state().commit_lsn);
486 5 : let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
487 5 : shared_state.sk.last_log_term(),
488 5 : shared_state.sk.flush_lsn(),
489 5 : )));
490 5 : let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
491 5 :
492 5 : let walreceivers = WalReceivers::new();
493 5 :
494 5 : Arc::new(Self {
495 5 : ttid,
496 5 : remote_path: remote_path.to_owned(),
497 5 : timeline_dir: timeline_dir.to_owned(),
498 5 : commit_lsn_watch_tx,
499 5 : commit_lsn_watch_rx,
500 5 : term_flush_lsn_watch_tx,
501 5 : term_flush_lsn_watch_rx,
502 5 : shared_state_version_tx,
503 5 : shared_state_version_rx,
504 5 : mutex: RwLock::new(shared_state),
505 5 : walsenders: WalSenders::new(walreceivers.clone()),
506 5 : walreceivers,
507 5 : gate: Default::default(),
508 5 : cancel: CancellationToken::default(),
509 5 : remote_deletion: std::sync::Mutex::new(None),
510 5 : manager_ctl: ManagerCtl::new(),
511 5 : conf,
512 5 : broker_active: AtomicBool::new(false),
513 5 : wal_backup_active: AtomicBool::new(false),
514 5 : last_removed_segno: AtomicU64::new(0),
515 5 : mgr_status: AtomicStatus::new(),
516 5 : wal_backup,
517 5 : })
518 5 : }
519 :
520 : /// Load existing timeline from disk.
521 0 : pub fn load_timeline(
522 0 : conf: Arc<SafeKeeperConf>,
523 0 : ttid: TenantTimelineId,
524 0 : wal_backup: Arc<WalBackup>,
525 0 : ) -> Result<Arc<Timeline>> {
526 0 : let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
527 :
528 0 : let shared_state = SharedState::restore(conf.as_ref(), &ttid)?;
529 0 : let timeline_dir = get_timeline_dir(conf.as_ref(), &ttid);
530 0 : let remote_path = remote_timeline_path(&ttid)?;
531 :
532 0 : Ok(Timeline::new(
533 0 : ttid,
534 0 : &timeline_dir,
535 0 : &remote_path,
536 0 : shared_state,
537 0 : conf,
538 0 : wal_backup,
539 0 : ))
540 0 : }
541 :
542 : /// Bootstrap new or existing timeline starting background tasks.
543 5 : pub fn bootstrap(
544 5 : self: &Arc<Timeline>,
545 5 : _shared_state: &mut WriteGuardSharedState<'_>,
546 5 : conf: &SafeKeeperConf,
547 5 : broker_active_set: Arc<TimelinesSet>,
548 5 : partial_backup_rate_limiter: RateLimiter,
549 5 : wal_backup: Arc<WalBackup>,
550 5 : ) {
551 5 : let (tx, rx) = self.manager_ctl.bootstrap_manager();
552 :
553 5 : let Ok(gate_guard) = self.gate.enter() else {
554 : // Init raced with shutdown
555 0 : return;
556 : };
557 :
558 : // Start manager task which will monitor timeline state and update
559 : // background tasks.
560 5 : tokio::spawn({
561 5 : let this = self.clone();
562 5 : let conf = conf.clone();
563 5 : async move {
564 5 : let _gate_guard = gate_guard;
565 5 : timeline_manager::main_task(
566 5 : ManagerTimeline { tli: this },
567 5 : conf,
568 5 : broker_active_set,
569 5 : tx,
570 5 : rx,
571 5 : partial_backup_rate_limiter,
572 5 : wal_backup,
573 5 : )
574 5 : .await
575 5 : }
576 5 : });
577 5 : }
578 :
579 : /// Cancel the timeline, requesting background activity to stop. Closing
580 : /// the `self.gate` waits for that.
581 0 : pub async fn cancel(&self) {
582 0 : info!("timeline {} shutting down", self.ttid);
583 0 : self.cancel.cancel();
584 0 : }
585 :
586 : /// Background timeline activities (which hold Timeline::gate) will no
587 : /// longer run once this function completes. `Self::cancel` must have been
588 : /// already called.
589 0 : pub async fn close(&self) {
590 0 : assert!(self.cancel.is_cancelled());
591 :
592 : // Wait for any concurrent tasks to stop using this timeline, to avoid e.g. attempts
593 : // to read deleted files.
594 0 : self.gate.close().await;
595 0 : }
596 :
597 : /// Delete timeline from disk completely, by removing timeline directory.
598 : ///
599 : /// Also deletes WAL in s3. Might fail if e.g. s3 is unavailable, but
600 : /// deletion API endpoint is retriable.
601 : ///
602 : /// Timeline must be in shut-down state (i.e. call [`Self::close`] first)
603 0 : pub async fn delete(
604 0 : &self,
605 0 : shared_state: &mut WriteGuardSharedState<'_>,
606 0 : only_local: bool,
607 0 : ) -> Result<bool> {
608 0 : // Assert that [`Self::close`] was already called
609 0 : assert!(self.cancel.is_cancelled());
610 0 : assert!(self.gate.close_complete());
611 :
612 0 : info!("deleting timeline {} from disk", self.ttid);
613 :
614 : // Close associated FDs. Nobody will be able to touch timeline data once
615 : // it is cancelled, so WAL storage won't be opened again.
616 0 : shared_state.sk.close_wal_store();
617 0 :
618 0 : if !only_local {
619 0 : self.remote_delete().await?;
620 0 : }
621 :
622 0 : let dir_existed = delete_dir(&self.timeline_dir).await?;
623 0 : Ok(dir_existed)
624 0 : }
625 :
626 : /// Delete timeline content from remote storage. If the returned future is dropped,
627 : /// deletion will continue in the background.
628 : ///
629 : /// This function ordinarily spawns a task and stashes a result receiver into [`Self::remote_deletion`]. If
630 : /// deletion is already happening, it may simply wait for an existing task's result.
631 : ///
632 : /// Note: we concurrently delete remote storage data from multiple
633 : /// safekeepers. That's ok, s3 replies 200 if object doesn't exist and we
634 : /// do some retries anyway.
635 0 : async fn remote_delete(&self) -> Result<()> {
636 : // We will start a background task to do the deletion, so that it proceeds even if our
637 : // API request is dropped. Future requests will see the existing deletion task and wait
638 : // for it to complete.
639 0 : let mut result_rx = {
640 0 : let mut remote_deletion_state = self.remote_deletion.lock().unwrap();
641 0 : let result_rx = if let Some(result_rx) = remote_deletion_state.as_ref() {
642 0 : if let Some(result) = result_rx.borrow().as_ref() {
643 0 : if let Err(e) = result {
644 : // A previous remote deletion failed: we will start a new one
645 0 : tracing::error!("remote deletion failed, will retry ({e})");
646 0 : None
647 : } else {
648 : // A previous remote deletion call already succeeded
649 0 : return Ok(());
650 : }
651 : } else {
652 : // Remote deletion is still in flight
653 0 : Some(result_rx.clone())
654 : }
655 : } else {
656 : // Remote deletion was not attempted yet, start it now.
657 0 : None
658 : };
659 :
660 0 : match result_rx {
661 0 : Some(result_rx) => result_rx,
662 0 : None => self.start_remote_delete(&mut remote_deletion_state),
663 : }
664 : };
665 :
666 : // Wait for a result
667 0 : let Ok(result) = result_rx.wait_for(|v| v.is_some()).await else {
668 : // Unexpected: sender should always send a result before dropping the channel, even if it has an error
669 0 : return Err(anyhow::anyhow!(
670 0 : "remote deletion task future was dropped without sending a result"
671 0 : ));
672 : };
673 :
674 0 : result
675 0 : .as_ref()
676 0 : .expect("We did a wait_for on this being Some above")
677 0 : .as_ref()
678 0 : .map(|_| ())
679 0 : .map_err(|e| anyhow::anyhow!("remote deletion failed: {e}"))
680 0 : }
681 :
682 : /// Spawn background task to do remote deletion, return a receiver for its outcome
683 0 : fn start_remote_delete(
684 0 : &self,
685 0 : guard: &mut std::sync::MutexGuard<Option<RemoteDeletionReceiver>>,
686 0 : ) -> RemoteDeletionReceiver {
687 0 : tracing::info!("starting remote deletion");
688 0 : let storage = self.wal_backup.get_storage().clone();
689 0 : let (result_tx, result_rx) = tokio::sync::watch::channel(None);
690 0 : let ttid = self.ttid;
691 0 : tokio::task::spawn(
692 0 : async move {
693 0 : let r = if let Some(storage) = storage {
694 0 : wal_backup::delete_timeline(&storage, &ttid).await
695 : } else {
696 0 : tracing::info!(
697 0 : "skipping remote deletion because no remote storage is configured; this effectively leaks the objects in remote storage"
698 : );
699 0 : Ok(())
700 : };
701 :
702 0 : if let Err(e) = &r {
703 : // Log error here in case nobody ever listens for our result (e.g. dropped API request)
704 0 : tracing::error!("remote deletion failed: {e}");
705 0 : }
706 :
707 : // Ignore send results: it's legal for the Timeline to give up waiting for us.
708 0 : let _ = result_tx.send(Some(r));
709 0 : }
710 0 : .instrument(info_span!("remote_delete", timeline = %self.ttid)),
711 : );
712 :
713 0 : **guard = Some(result_rx.clone());
714 0 :
715 0 : result_rx
716 0 : }
717 :
718 : /// Returns if timeline is cancelled.
719 2505 : pub fn is_cancelled(&self) -> bool {
720 2505 : self.cancel.is_cancelled()
721 2505 : }
722 :
723 : /// Take a writing mutual exclusive lock on timeline shared_state.
724 1245 : pub async fn write_shared_state(self: &Arc<Self>) -> WriteGuardSharedState<'_> {
725 1245 : WriteGuardSharedState::new(self.clone(), self.mutex.write().await)
726 1245 : }
727 :
728 58 : pub async fn read_shared_state(&self) -> ReadGuardSharedState {
729 58 : self.mutex.read().await
730 58 : }
731 :
732 : /// Returns commit_lsn watch channel.
733 5 : pub fn get_commit_lsn_watch_rx(&self) -> watch::Receiver<Lsn> {
734 5 : self.commit_lsn_watch_rx.clone()
735 5 : }
736 :
737 : /// Returns term_flush_lsn watch channel.
738 0 : pub fn get_term_flush_lsn_watch_rx(&self) -> watch::Receiver<TermLsn> {
739 0 : self.term_flush_lsn_watch_rx.clone()
740 0 : }
741 :
742 : /// Returns watch channel for SharedState update version.
743 5 : pub fn get_state_version_rx(&self) -> watch::Receiver<usize> {
744 5 : self.shared_state_version_rx.clone()
745 5 : }
746 :
747 : /// Returns wal_seg_size.
748 5 : pub async fn get_wal_seg_size(&self) -> usize {
749 5 : self.read_shared_state().await.get_wal_seg_size()
750 5 : }
751 :
752 : /// Returns state of the timeline.
753 9 : pub async fn get_state(&self) -> (TimelineMemState, TimelinePersistentState) {
754 9 : let state = self.read_shared_state().await;
755 9 : (
756 9 : state.sk.state().inmem.clone(),
757 9 : TimelinePersistentState::clone(state.sk.state()),
758 9 : )
759 9 : }
760 :
761 : /// Returns latest backup_lsn.
762 0 : pub async fn get_wal_backup_lsn(&self) -> Lsn {
763 0 : self.read_shared_state().await.sk.state().inmem.backup_lsn
764 0 : }
765 :
766 : /// Sets backup_lsn to the given value.
767 0 : pub async fn set_wal_backup_lsn(self: &Arc<Self>, backup_lsn: Lsn) -> Result<()> {
768 0 : if self.is_cancelled() {
769 0 : bail!(TimelineError::Cancelled(self.ttid));
770 0 : }
771 :
772 0 : let mut state = self.write_shared_state().await;
773 0 : state.sk.state_mut().inmem.backup_lsn = max(state.sk.state().inmem.backup_lsn, backup_lsn);
774 0 : // we should check whether to shut down offloader, but this will be done
775 0 : // soon by peer communication anyway.
776 0 : Ok(())
777 0 : }
778 :
779 : /// Get safekeeper info for broadcasting to broker and other peers.
780 0 : pub async fn get_safekeeper_info(&self, conf: &SafeKeeperConf) -> SafekeeperTimelineInfo {
781 0 : let standby_apply_lsn = self.walsenders.get_hotstandby().reply.apply_lsn;
782 0 : let shared_state = self.read_shared_state().await;
783 0 : shared_state.get_safekeeper_info(&self.ttid, conf, standby_apply_lsn)
784 0 : }
785 :
786 : /// Update timeline state with peer safekeeper data.
787 0 : pub async fn record_safekeeper_info(
788 0 : self: &Arc<Self>,
789 0 : sk_info: SafekeeperTimelineInfo,
790 0 : ) -> Result<()> {
791 : {
792 0 : let mut shared_state = self.write_shared_state().await;
793 0 : shared_state.sk.record_safekeeper_info(&sk_info).await?;
794 0 : let peer_info = peer_info_from_sk_info(&sk_info, Instant::now());
795 0 : shared_state.peers_info.upsert(&peer_info);
796 0 : }
797 0 : Ok(())
798 0 : }
799 :
800 0 : pub async fn get_peers(&self, conf: &SafeKeeperConf) -> Vec<PeerInfo> {
801 0 : let shared_state = self.read_shared_state().await;
802 0 : shared_state.get_peers(conf.heartbeat_timeout)
803 0 : }
804 :
805 5 : pub fn get_walsenders(&self) -> &Arc<WalSenders> {
806 5 : &self.walsenders
807 5 : }
808 :
809 15 : pub fn get_walreceivers(&self) -> &Arc<WalReceivers> {
810 15 : &self.walreceivers
811 15 : }
812 :
813 : /// Returns flush_lsn.
814 0 : pub async fn get_flush_lsn(&self) -> Lsn {
815 0 : self.read_shared_state().await.sk.flush_lsn()
816 0 : }
817 :
818 : /// Gather timeline data for metrics.
819 0 : pub async fn info_for_metrics(&self) -> Option<FullTimelineInfo> {
820 0 : if self.is_cancelled() {
821 0 : return None;
822 0 : }
823 0 :
824 0 : let WalSendersTimelineMetricValues {
825 0 : ps_feedback_counter,
826 0 : last_ps_feedback,
827 0 : interpreted_wal_reader_tasks,
828 0 : } = self.walsenders.info_for_metrics();
829 :
830 0 : let state = self.read_shared_state().await;
831 0 : Some(FullTimelineInfo {
832 0 : ttid: self.ttid,
833 0 : ps_feedback_count: ps_feedback_counter,
834 0 : last_ps_feedback,
835 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
836 0 : timeline_is_active: self.broker_active.load(Ordering::Relaxed),
837 0 : num_computes: self.walreceivers.get_num() as u32,
838 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
839 0 : interpreted_wal_reader_tasks,
840 0 : epoch_start_lsn: state.sk.term_start_lsn(),
841 0 : mem_state: state.sk.state().inmem.clone(),
842 0 : persisted_state: TimelinePersistentState::clone(state.sk.state()),
843 0 : flush_lsn: state.sk.flush_lsn(),
844 0 : wal_storage: state.sk.wal_storage_metrics(),
845 0 : })
846 0 : }
847 :
848 : /// Returns in-memory timeline state to build a full debug dump.
849 0 : pub async fn memory_dump(&self) -> debug_dump::Memory {
850 0 : let state = self.read_shared_state().await;
851 :
852 0 : let (write_lsn, write_record_lsn, flush_lsn, file_open) =
853 0 : state.sk.wal_storage_internal_state();
854 0 :
855 0 : debug_dump::Memory {
856 0 : is_cancelled: self.is_cancelled(),
857 0 : peers_info_len: state.peers_info.0.len(),
858 0 : walsenders: self.walsenders.get_all_public(),
859 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
860 0 : active: self.broker_active.load(Ordering::Relaxed),
861 0 : num_computes: self.walreceivers.get_num() as u32,
862 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
863 0 : epoch_start_lsn: state.sk.term_start_lsn(),
864 0 : mem_state: state.sk.state().inmem.clone(),
865 0 : mgr_status: self.mgr_status.get(),
866 0 : write_lsn,
867 0 : write_record_lsn,
868 0 : flush_lsn,
869 0 : file_open,
870 0 : }
871 0 : }
872 :
873 : /// Apply a function to the control file state and persist it.
874 0 : pub async fn map_control_file<T>(
875 0 : self: &Arc<Self>,
876 0 : f: impl FnOnce(&mut TimelinePersistentState) -> Result<T>,
877 0 : ) -> Result<T> {
878 0 : let mut state = self.write_shared_state().await;
879 0 : let mut persistent_state = state.sk.state_mut().start_change();
880 : // If f returns error, we abort the change and don't persist anything.
881 0 : let res = f(&mut persistent_state)?;
882 : // If persisting fails, we abort the change and return error.
883 0 : state
884 0 : .sk
885 0 : .state_mut()
886 0 : .finish_change(&persistent_state)
887 0 : .await?;
888 0 : Ok(res)
889 0 : }
890 :
891 0 : pub async fn term_bump(self: &Arc<Self>, to: Option<Term>) -> Result<TimelineTermBumpResponse> {
892 0 : let mut state = self.write_shared_state().await;
893 0 : state.sk.term_bump(to).await
894 0 : }
895 :
896 0 : pub async fn membership_switch(
897 0 : self: &Arc<Self>,
898 0 : to: Configuration,
899 0 : ) -> Result<TimelineMembershipSwitchResponse> {
900 0 : let mut state = self.write_shared_state().await;
901 0 : state.sk.membership_switch(to).await
902 0 : }
903 :
904 : /// Guts of [`Self::wal_residence_guard`] and [`Self::try_wal_residence_guard`]
905 10 : async fn do_wal_residence_guard(
906 10 : self: &Arc<Self>,
907 10 : block: bool,
908 10 : ) -> Result<Option<WalResidentTimeline>> {
909 10 : let op_label = if block {
910 10 : "wal_residence_guard"
911 : } else {
912 0 : "try_wal_residence_guard"
913 : };
914 :
915 10 : if self.is_cancelled() {
916 0 : bail!(TimelineError::Cancelled(self.ttid));
917 10 : }
918 10 :
919 10 : debug!("requesting WalResidentTimeline guard");
920 10 : let started_at = Instant::now();
921 10 : let status_before = self.mgr_status.get();
922 :
923 : // Wait 30 seconds for the guard to be acquired. It can time out if someone is
924 : // holding the lock (e.g. during `SafeKeeper::process_msg()`) or manager task
925 : // is stuck.
926 10 : let res = tokio::time::timeout_at(started_at + Duration::from_secs(30), async {
927 10 : if block {
928 10 : self.manager_ctl.wal_residence_guard().await.map(Some)
929 : } else {
930 0 : self.manager_ctl.try_wal_residence_guard().await
931 : }
932 10 : })
933 10 : .await;
934 :
935 10 : let guard = match res {
936 10 : Ok(Ok(guard)) => {
937 10 : let finished_at = Instant::now();
938 10 : let elapsed = finished_at - started_at;
939 10 : MISC_OPERATION_SECONDS
940 10 : .with_label_values(&[op_label])
941 10 : .observe(elapsed.as_secs_f64());
942 10 :
943 10 : guard
944 : }
945 0 : Ok(Err(e)) => {
946 0 : warn!(
947 0 : "error acquiring in {op_label}, statuses {:?} => {:?}",
948 0 : status_before,
949 0 : self.mgr_status.get()
950 : );
951 0 : return Err(e);
952 : }
953 : Err(_) => {
954 0 : warn!(
955 0 : "timeout acquiring in {op_label} guard, statuses {:?} => {:?}",
956 0 : status_before,
957 0 : self.mgr_status.get()
958 : );
959 0 : anyhow::bail!("timeout while acquiring WalResidentTimeline guard");
960 : }
961 : };
962 :
963 10 : Ok(guard.map(|g| WalResidentTimeline::new(self.clone(), g)))
964 10 : }
965 :
966 : /// Get the timeline guard for reading/writing WAL files.
967 : /// If WAL files are not present on disk (evicted), they will be automatically
968 : /// downloaded from remote storage. This is done in the manager task, which is
969 : /// responsible for issuing all guards.
970 : ///
971 : /// NB: don't use this function from timeline_manager, it will deadlock.
972 : /// NB: don't use this function while holding shared_state lock.
973 10 : pub async fn wal_residence_guard(self: &Arc<Self>) -> Result<WalResidentTimeline> {
974 10 : self.do_wal_residence_guard(true)
975 10 : .await
976 10 : .map(|m| m.expect("Always get Some in block=true mode"))
977 10 : }
978 :
979 : /// Get the timeline guard for reading/writing WAL files if the timeline is resident,
980 : /// else return None
981 0 : pub(crate) async fn try_wal_residence_guard(
982 0 : self: &Arc<Self>,
983 0 : ) -> Result<Option<WalResidentTimeline>> {
984 0 : self.do_wal_residence_guard(false).await
985 0 : }
986 :
987 0 : pub async fn backup_partial_reset(self: &Arc<Self>) -> Result<Vec<String>> {
988 0 : self.manager_ctl.backup_partial_reset().await
989 0 : }
990 : }
991 :
992 : /// This is a guard that allows to read/write disk timeline state.
993 : /// All tasks that are trying to read/write WAL from disk should use this guard.
994 : pub struct WalResidentTimeline {
995 : pub tli: Arc<Timeline>,
996 : _guard: ResidenceGuard,
997 : }
998 :
999 : impl WalResidentTimeline {
1000 15 : pub fn new(tli: Arc<Timeline>, _guard: ResidenceGuard) -> Self {
1001 15 : WalResidentTimeline { tli, _guard }
1002 15 : }
1003 : }
1004 :
1005 : impl Deref for WalResidentTimeline {
1006 : type Target = Arc<Timeline>;
1007 :
1008 5666 : fn deref(&self) -> &Self::Target {
1009 5666 : &self.tli
1010 5666 : }
1011 : }
1012 :
1013 : impl WalResidentTimeline {
1014 : /// Returns true if walsender should stop sending WAL to pageserver. We
1015 : /// terminate it if remote_consistent_lsn reached commit_lsn and there is no
1016 : /// computes. While there might be nothing to stream already, we learn about
1017 : /// remote_consistent_lsn update through replication feedback, and we want
1018 : /// to stop pushing to the broker if pageserver is fully caughtup.
1019 0 : pub async fn should_walsender_stop(&self, reported_remote_consistent_lsn: Lsn) -> bool {
1020 0 : if self.is_cancelled() {
1021 0 : return true;
1022 0 : }
1023 0 : let shared_state = self.read_shared_state().await;
1024 0 : if self.walreceivers.get_num() == 0 {
1025 0 : return shared_state.sk.state().inmem.commit_lsn == Lsn(0) || // no data at all yet
1026 0 : reported_remote_consistent_lsn >= shared_state.sk.state().inmem.commit_lsn;
1027 0 : }
1028 0 : false
1029 0 : }
1030 :
1031 : /// Ensure that current term is t, erroring otherwise, and lock the state.
1032 0 : pub async fn acquire_term(&self, t: Term) -> Result<ReadGuardSharedState> {
1033 0 : let ss = self.read_shared_state().await;
1034 0 : if ss.sk.state().acceptor_state.term != t {
1035 0 : bail!(
1036 0 : "failed to acquire term {}, current term {}",
1037 0 : t,
1038 0 : ss.sk.state().acceptor_state.term
1039 0 : );
1040 0 : }
1041 0 : Ok(ss)
1042 0 : }
1043 :
1044 : /// Pass arrived message to the safekeeper.
1045 1240 : pub async fn process_msg(
1046 1240 : &self,
1047 1240 : msg: &ProposerAcceptorMessage,
1048 1240 : ) -> Result<Option<AcceptorProposerMessage>> {
1049 1240 : if self.is_cancelled() {
1050 0 : bail!(TimelineError::Cancelled(self.ttid));
1051 1240 : }
1052 :
1053 : let mut rmsg: Option<AcceptorProposerMessage>;
1054 : {
1055 1240 : let mut shared_state = self.write_shared_state().await;
1056 1240 : rmsg = shared_state.sk.safekeeper().process_msg(msg).await?;
1057 :
1058 : // if this is AppendResponse, fill in proper hot standby feedback.
1059 620 : if let Some(AcceptorProposerMessage::AppendResponse(ref mut resp)) = rmsg {
1060 620 : resp.hs_feedback = self.walsenders.get_hotstandby().hs_feedback;
1061 620 : }
1062 : }
1063 1240 : Ok(rmsg)
1064 1240 : }
1065 :
1066 9 : pub async fn get_walreader(&self, start_lsn: Lsn) -> Result<WalReader> {
1067 9 : let (_, persisted_state) = self.get_state().await;
1068 :
1069 9 : WalReader::new(
1070 9 : &self.ttid,
1071 9 : self.timeline_dir.clone(),
1072 9 : &persisted_state,
1073 9 : start_lsn,
1074 9 : self.wal_backup.clone(),
1075 9 : )
1076 9 : }
1077 :
1078 0 : pub fn get_timeline_dir(&self) -> Utf8PathBuf {
1079 0 : self.timeline_dir.clone()
1080 0 : }
1081 :
1082 : /// Update in memory remote consistent lsn.
1083 0 : pub async fn update_remote_consistent_lsn(&self, candidate: Lsn) {
1084 0 : let mut shared_state = self.write_shared_state().await;
1085 0 : shared_state.sk.state_mut().inmem.remote_consistent_lsn = max(
1086 0 : shared_state.sk.state().inmem.remote_consistent_lsn,
1087 0 : candidate,
1088 0 : );
1089 0 : }
1090 : }
1091 :
1092 : /// This struct contains methods that are used by timeline manager task.
1093 : pub(crate) struct ManagerTimeline {
1094 : pub(crate) tli: Arc<Timeline>,
1095 : }
1096 :
1097 : impl Deref for ManagerTimeline {
1098 : type Target = Arc<Timeline>;
1099 :
1100 417 : fn deref(&self) -> &Self::Target {
1101 417 : &self.tli
1102 417 : }
1103 : }
1104 :
1105 : impl ManagerTimeline {
1106 0 : pub(crate) fn timeline_dir(&self) -> &Utf8PathBuf {
1107 0 : &self.tli.timeline_dir
1108 0 : }
1109 :
1110 : /// Manager requests this state on startup.
1111 5 : pub(crate) async fn bootstrap_mgr(&self) -> (bool, Option<PartialRemoteSegment>) {
1112 5 : let shared_state = self.read_shared_state().await;
1113 5 : let is_offloaded = matches!(
1114 5 : shared_state.sk.state().eviction_state,
1115 : EvictionState::Offloaded(_)
1116 : );
1117 5 : let partial_backup_uploaded = shared_state.sk.state().partial_backup.uploaded_segment();
1118 5 :
1119 5 : (is_offloaded, partial_backup_uploaded)
1120 5 : }
1121 :
1122 : /// Try to switch state Present->Offloaded.
1123 0 : pub(crate) async fn switch_to_offloaded(
1124 0 : &self,
1125 0 : partial: &PartialRemoteSegment,
1126 0 : ) -> anyhow::Result<()> {
1127 0 : let mut shared = self.write_shared_state().await;
1128 :
1129 : // updating control file
1130 0 : let mut pstate = shared.sk.state_mut().start_change();
1131 :
1132 0 : if !matches!(pstate.eviction_state, EvictionState::Present) {
1133 0 : bail!(
1134 0 : "cannot switch to offloaded state, current state is {:?}",
1135 0 : pstate.eviction_state
1136 0 : );
1137 0 : }
1138 0 :
1139 0 : if partial.flush_lsn != shared.sk.flush_lsn() {
1140 0 : bail!(
1141 0 : "flush_lsn mismatch in partial backup, expected {}, got {}",
1142 0 : shared.sk.flush_lsn(),
1143 0 : partial.flush_lsn
1144 0 : );
1145 0 : }
1146 0 :
1147 0 : if partial.commit_lsn != pstate.commit_lsn {
1148 0 : bail!(
1149 0 : "commit_lsn mismatch in partial backup, expected {}, got {}",
1150 0 : pstate.commit_lsn,
1151 0 : partial.commit_lsn
1152 0 : );
1153 0 : }
1154 0 :
1155 0 : if partial.term != shared.sk.last_log_term() {
1156 0 : bail!(
1157 0 : "term mismatch in partial backup, expected {}, got {}",
1158 0 : shared.sk.last_log_term(),
1159 0 : partial.term
1160 0 : );
1161 0 : }
1162 0 :
1163 0 : pstate.eviction_state = EvictionState::Offloaded(shared.sk.flush_lsn());
1164 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1165 : // control file is now switched to Offloaded state
1166 :
1167 : // now we can switch shared.sk to Offloaded, shouldn't fail
1168 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1169 0 : let cfile_state = prev_sk.take_state();
1170 0 : shared.sk = StateSK::Offloaded(Box::new(cfile_state));
1171 0 :
1172 0 : Ok(())
1173 0 : }
1174 :
1175 : /// Try to switch state Offloaded->Present.
1176 0 : pub(crate) async fn switch_to_present(&self) -> anyhow::Result<()> {
1177 0 : let mut shared = self.write_shared_state().await;
1178 :
1179 : // trying to restore WAL storage
1180 0 : let wal_store = wal_storage::PhysicalStorage::new(
1181 0 : &self.ttid,
1182 0 : &self.timeline_dir,
1183 0 : shared.sk.state(),
1184 0 : self.conf.no_sync,
1185 0 : )?;
1186 :
1187 : // updating control file
1188 0 : let mut pstate = shared.sk.state_mut().start_change();
1189 :
1190 0 : if !matches!(pstate.eviction_state, EvictionState::Offloaded(_)) {
1191 0 : bail!(
1192 0 : "cannot switch to present state, current state is {:?}",
1193 0 : pstate.eviction_state
1194 0 : );
1195 0 : }
1196 0 :
1197 0 : if wal_store.flush_lsn() != shared.sk.flush_lsn() {
1198 0 : bail!(
1199 0 : "flush_lsn mismatch in restored WAL, expected {}, got {}",
1200 0 : shared.sk.flush_lsn(),
1201 0 : wal_store.flush_lsn()
1202 0 : );
1203 0 : }
1204 0 :
1205 0 : pstate.eviction_state = EvictionState::Present;
1206 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1207 :
1208 : // now we can switch shared.sk to Present, shouldn't fail
1209 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1210 0 : let cfile_state = prev_sk.take_state();
1211 0 : shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, self.conf.my_id)?);
1212 :
1213 0 : Ok(())
1214 0 : }
1215 :
1216 : /// Update current manager state, useful for debugging manager deadlocks.
1217 226 : pub(crate) fn set_status(&self, status: timeline_manager::Status) {
1218 226 : self.mgr_status.store(status, Ordering::Relaxed);
1219 226 : }
1220 : }
1221 :
1222 : /// Deletes directory and it's contents. Returns false if directory does not exist.
1223 0 : pub async fn delete_dir(path: &Utf8PathBuf) -> Result<bool> {
1224 0 : match fs::remove_dir_all(path).await {
1225 0 : Ok(_) => Ok(true),
1226 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(false),
1227 0 : Err(e) => Err(e.into()),
1228 : }
1229 0 : }
1230 :
1231 : /// Get a path to the tenant directory. If you just need to get a timeline directory,
1232 : /// use WalResidentTimeline::get_timeline_dir instead.
1233 10 : pub fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
1234 10 : conf.workdir.join(tenant_id.to_string())
1235 10 : }
1236 :
1237 : /// Get a path to the timeline directory. If you need to read WAL files from disk,
1238 : /// use WalResidentTimeline::get_timeline_dir instead. This function does not check
1239 : /// timeline eviction status and WAL files might not be present on disk.
1240 10 : pub fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
1241 10 : get_tenant_dir(conf, &ttid.tenant_id).join(ttid.timeline_id.to_string())
1242 10 : }
|
