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::{
30 : FullTimelineInfo, MISC_OPERATION_SECONDS, WAL_STORAGE_LIMIT_ERRORS, WalStorageMetrics,
31 : };
32 :
33 : use crate::hadron::GLOBAL_DISK_LIMIT_EXCEEDED;
34 : use crate::rate_limit::RateLimiter;
35 : use crate::receive_wal::WalReceivers;
36 : use crate::safekeeper::{AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, TermLsn};
37 : use crate::send_wal::{WalSenders, WalSendersTimelineMetricValues};
38 : use crate::state::{EvictionState, TimelineMemState, TimelinePersistentState, TimelineState};
39 : use crate::timeline_guard::ResidenceGuard;
40 : use crate::timeline_manager::{AtomicStatus, ManagerCtl};
41 : use crate::timelines_set::TimelinesSet;
42 : use crate::wal_backup;
43 : use crate::wal_backup::{WalBackup, remote_timeline_path};
44 : use crate::wal_backup_partial::PartialRemoteSegment;
45 : use crate::wal_storage::{Storage as wal_storage_iface, WalReader};
46 : use crate::{SafeKeeperConf, control_file, debug_dump, timeline_manager, wal_storage};
47 :
48 0 : fn peer_info_from_sk_info(sk_info: &SafekeeperTimelineInfo, ts: Instant) -> PeerInfo {
49 0 : PeerInfo {
50 0 : sk_id: NodeId(sk_info.safekeeper_id),
51 0 : term: sk_info.term,
52 0 : last_log_term: sk_info.last_log_term,
53 0 : flush_lsn: Lsn(sk_info.flush_lsn),
54 0 : commit_lsn: Lsn(sk_info.commit_lsn),
55 0 : local_start_lsn: Lsn(sk_info.local_start_lsn),
56 0 : pg_connstr: sk_info.safekeeper_connstr.clone(),
57 0 : http_connstr: sk_info.http_connstr.clone(),
58 0 : https_connstr: sk_info.https_connstr.clone(),
59 0 : ts,
60 0 : }
61 0 : }
62 :
63 : // vector-based node id -> peer state map with very limited functionality we
64 : // need.
65 : #[derive(Debug, Clone, Default)]
66 : pub struct PeersInfo(pub Vec<PeerInfo>);
67 :
68 : impl PeersInfo {
69 0 : fn get(&mut self, id: NodeId) -> Option<&mut PeerInfo> {
70 0 : self.0.iter_mut().find(|p| p.sk_id == id)
71 0 : }
72 :
73 0 : fn upsert(&mut self, p: &PeerInfo) {
74 0 : match self.get(p.sk_id) {
75 0 : Some(rp) => *rp = p.clone(),
76 0 : None => self.0.push(p.clone()),
77 : }
78 0 : }
79 : }
80 :
81 : pub type ReadGuardSharedState<'a> = RwLockReadGuard<'a, SharedState>;
82 :
83 : /// WriteGuardSharedState is a wrapper around `RwLockWriteGuard<SharedState>` that
84 : /// automatically updates `watch::Sender` channels with state on drop.
85 : pub struct WriteGuardSharedState<'a> {
86 : tli: Arc<Timeline>,
87 : guard: RwLockWriteGuard<'a, SharedState>,
88 : }
89 :
90 : impl<'a> WriteGuardSharedState<'a> {
91 1249 : fn new(tli: Arc<Timeline>, guard: RwLockWriteGuard<'a, SharedState>) -> Self {
92 1249 : WriteGuardSharedState { tli, guard }
93 1249 : }
94 : }
95 :
96 : impl Deref for WriteGuardSharedState<'_> {
97 : type Target = SharedState;
98 :
99 0 : fn deref(&self) -> &Self::Target {
100 0 : &self.guard
101 0 : }
102 : }
103 :
104 : impl DerefMut for WriteGuardSharedState<'_> {
105 1244 : fn deref_mut(&mut self) -> &mut Self::Target {
106 1244 : &mut self.guard
107 1244 : }
108 : }
109 :
110 : impl Drop for WriteGuardSharedState<'_> {
111 1249 : fn drop(&mut self) {
112 1249 : let term_flush_lsn =
113 1249 : TermLsn::from((self.guard.sk.last_log_term(), self.guard.sk.flush_lsn()));
114 1249 : let commit_lsn = self.guard.sk.state().inmem.commit_lsn;
115 :
116 1249 : let _ = self.tli.term_flush_lsn_watch_tx.send_if_modified(|old| {
117 1249 : if *old != term_flush_lsn {
118 620 : *old = term_flush_lsn;
119 620 : true
120 : } else {
121 629 : false
122 : }
123 1249 : });
124 :
125 1249 : let _ = self.tli.commit_lsn_watch_tx.send_if_modified(|old| {
126 1249 : if *old != commit_lsn {
127 615 : *old = commit_lsn;
128 615 : true
129 : } else {
130 634 : false
131 : }
132 1249 : });
133 :
134 : // send notification about shared state update
135 1249 : self.tli.shared_state_version_tx.send_modify(|old| {
136 1249 : *old += 1;
137 1249 : });
138 1249 : }
139 : }
140 :
141 : /// This structure is stored in shared state and represents the state of the timeline.
142 : ///
143 : /// Usually it holds SafeKeeper, but it also supports offloaded timeline state. In this
144 : /// case, SafeKeeper is not available (because WAL is not present on disk) and all
145 : /// operations can be done only with control file.
146 : #[allow(clippy::large_enum_variant, reason = "TODO")]
147 : pub enum StateSK {
148 : Loaded(SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>),
149 : Offloaded(Box<TimelineState<control_file::FileStorage>>),
150 : // Not used, required for moving between states.
151 : Empty,
152 : }
153 :
154 : impl StateSK {
155 2576 : pub fn flush_lsn(&self) -> Lsn {
156 2576 : match self {
157 2576 : StateSK::Loaded(sk) => sk.wal_store.flush_lsn(),
158 0 : StateSK::Offloaded(state) => match state.eviction_state {
159 0 : EvictionState::Offloaded(flush_lsn) => flush_lsn,
160 0 : _ => panic!("StateSK::Offloaded mismatches with eviction_state from control_file"),
161 : },
162 0 : StateSK::Empty => unreachable!(),
163 : }
164 2576 : }
165 :
166 : /// Get a reference to the control file's timeline state.
167 2609 : pub fn state(&self) -> &TimelineState<control_file::FileStorage> {
168 2609 : match self {
169 2609 : StateSK::Loaded(sk) => &sk.state,
170 0 : StateSK::Offloaded(s) => s,
171 0 : StateSK::Empty => unreachable!(),
172 : }
173 2609 : }
174 :
175 4 : pub fn state_mut(&mut self) -> &mut TimelineState<control_file::FileStorage> {
176 4 : match self {
177 4 : StateSK::Loaded(sk) => &mut sk.state,
178 0 : StateSK::Offloaded(s) => s,
179 0 : StateSK::Empty => unreachable!(),
180 : }
181 4 : }
182 :
183 1288 : pub fn last_log_term(&self) -> Term {
184 1288 : self.state()
185 1288 : .acceptor_state
186 1288 : .get_last_log_term(self.flush_lsn())
187 1288 : }
188 :
189 0 : pub async fn term_bump(&mut self, to: Option<Term>) -> Result<TimelineTermBumpResponse> {
190 0 : self.state_mut().term_bump(to).await
191 0 : }
192 :
193 0 : pub async fn membership_switch(
194 0 : &mut self,
195 0 : to: Configuration,
196 0 : ) -> Result<TimelineMembershipSwitchResponse> {
197 0 : let result = self.state_mut().membership_switch(to).await?;
198 :
199 0 : Ok(TimelineMembershipSwitchResponse {
200 0 : previous_conf: result.previous_conf,
201 0 : current_conf: result.current_conf,
202 0 : last_log_term: self.state().acceptor_state.term,
203 0 : flush_lsn: self.flush_lsn(),
204 0 : })
205 0 : }
206 :
207 : /// Close open WAL files to release FDs.
208 0 : fn close_wal_store(&mut self) {
209 0 : if let StateSK::Loaded(sk) = self {
210 0 : sk.wal_store.close();
211 0 : }
212 0 : }
213 :
214 : /// Update timeline state with peer safekeeper data.
215 0 : pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> {
216 : // update commit_lsn if safekeeper is loaded
217 0 : match self {
218 0 : StateSK::Loaded(sk) => sk.record_safekeeper_info(sk_info).await?,
219 0 : StateSK::Offloaded(_) => {}
220 0 : StateSK::Empty => unreachable!(),
221 : }
222 :
223 : // update everything else, including remote_consistent_lsn and backup_lsn
224 0 : let mut sync_control_file = false;
225 0 : let state = self.state_mut();
226 0 : let wal_seg_size = state.server.wal_seg_size as u64;
227 :
228 0 : state.inmem.backup_lsn = max(Lsn(sk_info.backup_lsn), state.inmem.backup_lsn);
229 0 : sync_control_file |= state.backup_lsn + wal_seg_size < state.inmem.backup_lsn;
230 :
231 0 : state.inmem.remote_consistent_lsn = max(
232 0 : Lsn(sk_info.remote_consistent_lsn),
233 0 : state.inmem.remote_consistent_lsn,
234 0 : );
235 0 : sync_control_file |=
236 0 : state.remote_consistent_lsn + wal_seg_size < state.inmem.remote_consistent_lsn;
237 :
238 0 : state.inmem.peer_horizon_lsn =
239 0 : max(Lsn(sk_info.peer_horizon_lsn), state.inmem.peer_horizon_lsn);
240 0 : sync_control_file |= state.peer_horizon_lsn + wal_seg_size < state.inmem.peer_horizon_lsn;
241 :
242 0 : if sync_control_file {
243 0 : state.flush().await?;
244 0 : }
245 0 : Ok(())
246 0 : }
247 :
248 : /// Previously known as epoch_start_lsn. Needed only for reference in some APIs.
249 0 : pub fn term_start_lsn(&self) -> Lsn {
250 0 : match self {
251 0 : StateSK::Loaded(sk) => sk.term_start_lsn,
252 0 : StateSK::Offloaded(_) => Lsn(0),
253 0 : StateSK::Empty => unreachable!(),
254 : }
255 0 : }
256 :
257 : /// Used for metrics only.
258 0 : pub fn wal_storage_metrics(&self) -> WalStorageMetrics {
259 0 : match self {
260 0 : StateSK::Loaded(sk) => sk.wal_store.get_metrics(),
261 0 : StateSK::Offloaded(_) => WalStorageMetrics::default(),
262 0 : StateSK::Empty => unreachable!(),
263 : }
264 0 : }
265 :
266 : /// Returns WAL storage internal LSNs for debug dump.
267 0 : pub fn wal_storage_internal_state(&self) -> (Lsn, Lsn, Lsn, bool) {
268 0 : match self {
269 0 : StateSK::Loaded(sk) => sk.wal_store.internal_state(),
270 : StateSK::Offloaded(_) => {
271 0 : let flush_lsn = self.flush_lsn();
272 0 : (flush_lsn, flush_lsn, flush_lsn, false)
273 : }
274 0 : StateSK::Empty => unreachable!(),
275 : }
276 0 : }
277 :
278 : /// Access to SafeKeeper object. Panics if offloaded, should be good to use from WalResidentTimeline.
279 1240 : pub fn safekeeper(
280 1240 : &mut self,
281 1240 : ) -> &mut SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage> {
282 1240 : match self {
283 1240 : StateSK::Loaded(sk) => sk,
284 : StateSK::Offloaded(_) => {
285 0 : panic!("safekeeper is offloaded, cannot be used")
286 : }
287 0 : StateSK::Empty => unreachable!(),
288 : }
289 1240 : }
290 :
291 : /// Moves control file's state structure out of the enum. Used to switch states.
292 0 : fn take_state(self) -> TimelineState<control_file::FileStorage> {
293 0 : match self {
294 0 : StateSK::Loaded(sk) => sk.state,
295 0 : StateSK::Offloaded(state) => *state,
296 0 : StateSK::Empty => unreachable!(),
297 : }
298 0 : }
299 : }
300 :
301 : /// Shared state associated with database instance
302 : pub struct SharedState {
303 : /// Safekeeper object
304 : pub(crate) sk: StateSK,
305 : /// In memory list containing state of peers sent in latest messages from them.
306 : pub(crate) peers_info: PeersInfo,
307 : // True value hinders old WAL removal; this is used by snapshotting. We
308 : // could make it a counter, but there is no need to.
309 : pub(crate) wal_removal_on_hold: bool,
310 : }
311 :
312 : impl SharedState {
313 : /// Creates a new SharedState.
314 5 : pub fn new(sk: StateSK) -> Self {
315 5 : Self {
316 5 : sk,
317 5 : peers_info: PeersInfo(vec![]),
318 5 : wal_removal_on_hold: false,
319 5 : }
320 5 : }
321 :
322 : /// Restore SharedState from control file. If file doesn't exist, bails out.
323 0 : pub fn restore(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Result<Self> {
324 0 : let timeline_dir = get_timeline_dir(conf, ttid);
325 0 : let control_store = control_file::FileStorage::restore_new(&timeline_dir, conf.no_sync)?;
326 0 : if control_store.server.wal_seg_size == 0 {
327 0 : bail!(TimelineError::UninitializedWalSegSize(*ttid));
328 0 : }
329 :
330 0 : let sk = match control_store.eviction_state {
331 : EvictionState::Present => {
332 0 : let wal_store = wal_storage::PhysicalStorage::new(
333 0 : ttid,
334 0 : &timeline_dir,
335 0 : &control_store,
336 0 : conf.no_sync,
337 0 : )?;
338 0 : StateSK::Loaded(SafeKeeper::new(
339 0 : TimelineState::new(control_store),
340 0 : wal_store,
341 0 : conf.my_id,
342 0 : )?)
343 : }
344 : EvictionState::Offloaded(_) => {
345 0 : StateSK::Offloaded(Box::new(TimelineState::new(control_store)))
346 : }
347 : };
348 :
349 0 : Ok(Self::new(sk))
350 0 : }
351 :
352 5 : pub(crate) fn get_wal_seg_size(&self) -> usize {
353 5 : self.sk.state().server.wal_seg_size as usize
354 5 : }
355 :
356 0 : fn get_safekeeper_info(
357 0 : &self,
358 0 : ttid: &TenantTimelineId,
359 0 : conf: &SafeKeeperConf,
360 0 : standby_apply_lsn: Lsn,
361 0 : ) -> SafekeeperTimelineInfo {
362 0 : SafekeeperTimelineInfo {
363 0 : safekeeper_id: conf.my_id.0,
364 0 : tenant_timeline_id: Some(ProtoTenantTimelineId {
365 0 : tenant_id: ttid.tenant_id.as_ref().to_owned(),
366 0 : timeline_id: ttid.timeline_id.as_ref().to_owned(),
367 0 : }),
368 0 : term: self.sk.state().acceptor_state.term,
369 0 : last_log_term: self.sk.last_log_term(),
370 0 : flush_lsn: self.sk.flush_lsn().0,
371 0 : // note: this value is not flushed to control file yet and can be lost
372 0 : commit_lsn: self.sk.state().inmem.commit_lsn.0,
373 0 : remote_consistent_lsn: self.sk.state().inmem.remote_consistent_lsn.0,
374 0 : peer_horizon_lsn: self.sk.state().inmem.peer_horizon_lsn.0,
375 0 : safekeeper_connstr: conf
376 0 : .advertise_pg_addr
377 0 : .to_owned()
378 0 : .unwrap_or(conf.listen_pg_addr.clone()),
379 0 : http_connstr: conf.listen_http_addr.to_owned(),
380 0 : https_connstr: conf.listen_https_addr.to_owned(),
381 0 : backup_lsn: self.sk.state().inmem.backup_lsn.0,
382 0 : local_start_lsn: self.sk.state().local_start_lsn.0,
383 0 : availability_zone: conf.availability_zone.clone(),
384 0 : standby_horizon: standby_apply_lsn.0,
385 0 : }
386 0 : }
387 :
388 : /// Get our latest view of alive peers status on the timeline.
389 : /// We pass our own info through the broker as well, so when we don't have connection
390 : /// to the broker returned vec is empty.
391 34 : pub(crate) fn get_peers(&self, heartbeat_timeout: Duration) -> Vec<PeerInfo> {
392 34 : let now = Instant::now();
393 34 : self.peers_info
394 34 : .0
395 34 : .iter()
396 : // Regard peer as absent if we haven't heard from it within heartbeat_timeout.
397 34 : .filter(|p| now.duration_since(p.ts) <= heartbeat_timeout)
398 34 : .cloned()
399 34 : .collect()
400 34 : }
401 : }
402 :
403 : #[derive(Debug, thiserror::Error)]
404 : pub enum TimelineError {
405 : #[error("Timeline {0} was cancelled and cannot be used anymore")]
406 : Cancelled(TenantTimelineId),
407 : #[error("Timeline {0} was not found in global map")]
408 : NotFound(TenantTimelineId),
409 : #[error("Timeline {0} has been deleted")]
410 : Deleted(TenantTimelineId),
411 : #[error("Timeline {0} creation is in progress")]
412 : CreationInProgress(TenantTimelineId),
413 : #[error("Timeline {0} exists on disk, but wasn't loaded on startup")]
414 : Invalid(TenantTimelineId),
415 : #[error("Timeline {0} is already exists")]
416 : AlreadyExists(TenantTimelineId),
417 : #[error("Timeline {0} is not initialized, wal_seg_size is zero")]
418 : UninitializedWalSegSize(TenantTimelineId),
419 : #[error("Timeline {0} is not initialized, pg_version is unknown")]
420 : UninitialinzedPgVersion(TenantTimelineId),
421 : }
422 :
423 : // Convert to HTTP API error.
424 : impl From<TimelineError> for ApiError {
425 0 : fn from(te: TimelineError) -> ApiError {
426 0 : match te {
427 0 : TimelineError::NotFound(ttid) => {
428 0 : ApiError::NotFound(anyhow!("timeline {} not found", ttid).into())
429 : }
430 0 : TimelineError::Deleted(ttid) => {
431 0 : ApiError::NotFound(anyhow!("timeline {} deleted", ttid).into())
432 : }
433 0 : _ => ApiError::InternalServerError(anyhow!("{}", te)),
434 : }
435 0 : }
436 : }
437 :
438 : /// We run remote deletion in a background task, this is how it sends its results back.
439 : type RemoteDeletionReceiver = tokio::sync::watch::Receiver<Option<anyhow::Result<()>>>;
440 :
441 : /// Timeline struct manages lifecycle (creation, deletion, restore) of a safekeeper timeline.
442 : /// It also holds SharedState and provides mutually exclusive access to it.
443 : pub struct Timeline {
444 : pub ttid: TenantTimelineId,
445 : pub remote_path: RemotePath,
446 :
447 : /// Used to broadcast commit_lsn updates to all background jobs.
448 : commit_lsn_watch_tx: watch::Sender<Lsn>,
449 : commit_lsn_watch_rx: watch::Receiver<Lsn>,
450 :
451 : /// Broadcasts (current term, flush_lsn) updates, walsender is interested in
452 : /// them when sending in recovery mode (to walproposer or peers). Note: this
453 : /// is just a notification, WAL reading should always done with lock held as
454 : /// term can change otherwise.
455 : term_flush_lsn_watch_tx: watch::Sender<TermLsn>,
456 : term_flush_lsn_watch_rx: watch::Receiver<TermLsn>,
457 :
458 : /// Broadcasts shared state updates.
459 : shared_state_version_tx: watch::Sender<usize>,
460 : shared_state_version_rx: watch::Receiver<usize>,
461 :
462 : /// Safekeeper and other state, that should remain consistent and
463 : /// synchronized with the disk. This is tokio mutex as we write WAL to disk
464 : /// while holding it, ensuring that consensus checks are in order.
465 : mutex: RwLock<SharedState>,
466 : walsenders: Arc<WalSenders>,
467 : walreceivers: Arc<WalReceivers>,
468 : timeline_dir: Utf8PathBuf,
469 : manager_ctl: ManagerCtl,
470 : conf: Arc<SafeKeeperConf>,
471 :
472 : pub(crate) wal_backup: Arc<WalBackup>,
473 :
474 : remote_deletion: std::sync::Mutex<Option<RemoteDeletionReceiver>>,
475 :
476 : /// Hold this gate from code that depends on the Timeline's non-shut-down state. While holding
477 : /// this gate, you must respect [`Timeline::cancel`]
478 : pub(crate) gate: Gate,
479 :
480 : /// Delete/cancel will trigger this, background tasks should drop out as soon as it fires
481 : pub(crate) cancel: CancellationToken,
482 :
483 : // timeline_manager controlled state
484 : pub(crate) broker_active: AtomicBool,
485 : pub(crate) wal_backup_active: AtomicBool,
486 : pub(crate) last_removed_segno: AtomicU64,
487 : pub(crate) mgr_status: AtomicStatus,
488 : }
489 :
490 : impl Timeline {
491 : /// Constructs a new timeline.
492 5 : pub fn new(
493 5 : ttid: TenantTimelineId,
494 5 : timeline_dir: &Utf8Path,
495 5 : remote_path: &RemotePath,
496 5 : shared_state: SharedState,
497 5 : conf: Arc<SafeKeeperConf>,
498 5 : wal_backup: Arc<WalBackup>,
499 5 : ) -> Arc<Self> {
500 5 : let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
501 5 : watch::channel(shared_state.sk.state().commit_lsn);
502 5 : let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
503 5 : shared_state.sk.last_log_term(),
504 5 : shared_state.sk.flush_lsn(),
505 5 : )));
506 5 : let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
507 :
508 5 : let walreceivers = WalReceivers::new();
509 :
510 5 : Arc::new(Self {
511 5 : ttid,
512 5 : remote_path: remote_path.to_owned(),
513 5 : timeline_dir: timeline_dir.to_owned(),
514 5 : commit_lsn_watch_tx,
515 5 : commit_lsn_watch_rx,
516 5 : term_flush_lsn_watch_tx,
517 5 : term_flush_lsn_watch_rx,
518 5 : shared_state_version_tx,
519 5 : shared_state_version_rx,
520 5 : mutex: RwLock::new(shared_state),
521 5 : walsenders: WalSenders::new(walreceivers.clone()),
522 5 : walreceivers,
523 5 : gate: Default::default(),
524 5 : cancel: CancellationToken::default(),
525 5 : remote_deletion: std::sync::Mutex::new(None),
526 5 : manager_ctl: ManagerCtl::new(),
527 5 : conf,
528 5 : broker_active: AtomicBool::new(false),
529 5 : wal_backup_active: AtomicBool::new(false),
530 5 : last_removed_segno: AtomicU64::new(0),
531 5 : mgr_status: AtomicStatus::new(),
532 5 : wal_backup,
533 5 : })
534 5 : }
535 :
536 : /// Load existing timeline from disk.
537 0 : pub fn load_timeline(
538 0 : conf: Arc<SafeKeeperConf>,
539 0 : ttid: TenantTimelineId,
540 0 : wal_backup: Arc<WalBackup>,
541 0 : ) -> Result<Arc<Timeline>> {
542 0 : let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
543 :
544 0 : let shared_state = SharedState::restore(conf.as_ref(), &ttid)?;
545 0 : let timeline_dir = get_timeline_dir(conf.as_ref(), &ttid);
546 0 : let remote_path = remote_timeline_path(&ttid)?;
547 :
548 0 : Ok(Timeline::new(
549 0 : ttid,
550 0 : &timeline_dir,
551 0 : &remote_path,
552 0 : shared_state,
553 0 : conf,
554 0 : wal_backup,
555 0 : ))
556 0 : }
557 :
558 : /// Bootstrap new or existing timeline starting background tasks.
559 5 : pub fn bootstrap(
560 5 : self: &Arc<Timeline>,
561 5 : _shared_state: &mut WriteGuardSharedState<'_>,
562 5 : conf: &SafeKeeperConf,
563 5 : broker_active_set: Arc<TimelinesSet>,
564 5 : partial_backup_rate_limiter: RateLimiter,
565 5 : wal_backup: Arc<WalBackup>,
566 5 : ) {
567 5 : let (tx, rx) = self.manager_ctl.bootstrap_manager();
568 :
569 5 : let Ok(gate_guard) = self.gate.enter() else {
570 : // Init raced with shutdown
571 0 : return;
572 : };
573 :
574 : // Start manager task which will monitor timeline state and update
575 : // background tasks.
576 5 : tokio::spawn({
577 5 : let this = self.clone();
578 5 : let conf = conf.clone();
579 5 : async move {
580 5 : let _gate_guard = gate_guard;
581 5 : timeline_manager::main_task(
582 5 : ManagerTimeline { tli: this },
583 5 : conf,
584 5 : broker_active_set,
585 5 : tx,
586 5 : rx,
587 5 : partial_backup_rate_limiter,
588 5 : wal_backup,
589 5 : )
590 5 : .await
591 0 : }
592 : });
593 5 : }
594 :
595 : /// Cancel the timeline, requesting background activity to stop. Closing
596 : /// the `self.gate` waits for that.
597 0 : pub async fn cancel(&self) {
598 0 : info!("timeline {} shutting down", self.ttid);
599 0 : self.cancel.cancel();
600 0 : }
601 :
602 : /// Background timeline activities (which hold Timeline::gate) will no
603 : /// longer run once this function completes. `Self::cancel` must have been
604 : /// already called.
605 0 : pub async fn close(&self) {
606 0 : assert!(self.cancel.is_cancelled());
607 :
608 : // Wait for any concurrent tasks to stop using this timeline, to avoid e.g. attempts
609 : // to read deleted files.
610 0 : self.gate.close().await;
611 0 : }
612 :
613 : /// Delete timeline from disk completely, by removing timeline directory.
614 : ///
615 : /// Also deletes WAL in s3. Might fail if e.g. s3 is unavailable, but
616 : /// deletion API endpoint is retriable.
617 : ///
618 : /// Timeline must be in shut-down state (i.e. call [`Self::close`] first)
619 0 : pub async fn delete(
620 0 : &self,
621 0 : shared_state: &mut WriteGuardSharedState<'_>,
622 0 : only_local: bool,
623 0 : ) -> Result<bool> {
624 : // Assert that [`Self::close`] was already called
625 0 : assert!(self.cancel.is_cancelled());
626 0 : assert!(self.gate.close_complete());
627 :
628 0 : info!("deleting timeline {} from disk", self.ttid);
629 :
630 : // Close associated FDs. Nobody will be able to touch timeline data once
631 : // it is cancelled, so WAL storage won't be opened again.
632 0 : shared_state.sk.close_wal_store();
633 :
634 0 : if !only_local {
635 0 : self.remote_delete().await?;
636 0 : }
637 :
638 0 : let dir_existed = delete_dir(&self.timeline_dir).await?;
639 0 : Ok(dir_existed)
640 0 : }
641 :
642 : /// Delete timeline content from remote storage. If the returned future is dropped,
643 : /// deletion will continue in the background.
644 : ///
645 : /// This function ordinarily spawns a task and stashes a result receiver into [`Self::remote_deletion`]. If
646 : /// deletion is already happening, it may simply wait for an existing task's result.
647 : ///
648 : /// Note: we concurrently delete remote storage data from multiple
649 : /// safekeepers. That's ok, s3 replies 200 if object doesn't exist and we
650 : /// do some retries anyway.
651 0 : async fn remote_delete(&self) -> Result<()> {
652 : // We will start a background task to do the deletion, so that it proceeds even if our
653 : // API request is dropped. Future requests will see the existing deletion task and wait
654 : // for it to complete.
655 0 : let mut result_rx = {
656 0 : let mut remote_deletion_state = self.remote_deletion.lock().unwrap();
657 0 : let result_rx = if let Some(result_rx) = remote_deletion_state.as_ref() {
658 0 : if let Some(result) = result_rx.borrow().as_ref() {
659 0 : if let Err(e) = result {
660 : // A previous remote deletion failed: we will start a new one
661 0 : tracing::error!("remote deletion failed, will retry ({e})");
662 0 : None
663 : } else {
664 : // A previous remote deletion call already succeeded
665 0 : return Ok(());
666 : }
667 : } else {
668 : // Remote deletion is still in flight
669 0 : Some(result_rx.clone())
670 : }
671 : } else {
672 : // Remote deletion was not attempted yet, start it now.
673 0 : None
674 : };
675 :
676 0 : match result_rx {
677 0 : Some(result_rx) => result_rx,
678 0 : None => self.start_remote_delete(&mut remote_deletion_state),
679 : }
680 : };
681 :
682 : // Wait for a result
683 0 : let Ok(result) = result_rx.wait_for(|v| v.is_some()).await else {
684 : // Unexpected: sender should always send a result before dropping the channel, even if it has an error
685 0 : return Err(anyhow::anyhow!(
686 0 : "remote deletion task future was dropped without sending a result"
687 0 : ));
688 : };
689 :
690 0 : result
691 0 : .as_ref()
692 0 : .expect("We did a wait_for on this being Some above")
693 0 : .as_ref()
694 0 : .map(|_| ())
695 0 : .map_err(|e| anyhow::anyhow!("remote deletion failed: {e}"))
696 0 : }
697 :
698 : /// Spawn background task to do remote deletion, return a receiver for its outcome
699 0 : fn start_remote_delete(
700 0 : &self,
701 0 : guard: &mut std::sync::MutexGuard<Option<RemoteDeletionReceiver>>,
702 0 : ) -> RemoteDeletionReceiver {
703 0 : tracing::info!("starting remote deletion");
704 0 : let storage = self.wal_backup.get_storage().clone();
705 0 : let (result_tx, result_rx) = tokio::sync::watch::channel(None);
706 0 : let ttid = self.ttid;
707 0 : tokio::task::spawn(
708 0 : async move {
709 0 : let r = if let Some(storage) = storage {
710 0 : wal_backup::delete_timeline(&storage, &ttid).await
711 : } else {
712 0 : tracing::info!(
713 0 : "skipping remote deletion because no remote storage is configured; this effectively leaks the objects in remote storage"
714 : );
715 0 : Ok(())
716 : };
717 :
718 0 : if let Err(e) = &r {
719 : // Log error here in case nobody ever listens for our result (e.g. dropped API request)
720 0 : tracing::error!("remote deletion failed: {e}");
721 0 : }
722 :
723 : // Ignore send results: it's legal for the Timeline to give up waiting for us.
724 0 : let _ = result_tx.send(Some(r));
725 0 : }
726 0 : .instrument(info_span!("remote_delete", timeline = %self.ttid)),
727 : );
728 :
729 0 : **guard = Some(result_rx.clone());
730 :
731 0 : result_rx
732 0 : }
733 :
734 : /// Returns if timeline is cancelled.
735 2505 : pub fn is_cancelled(&self) -> bool {
736 2505 : self.cancel.is_cancelled()
737 2505 : }
738 :
739 : /// Take a writing mutual exclusive lock on timeline shared_state.
740 1249 : pub async fn write_shared_state(self: &Arc<Self>) -> WriteGuardSharedState<'_> {
741 1249 : WriteGuardSharedState::new(self.clone(), self.mutex.write().await)
742 1249 : }
743 :
744 53 : pub async fn read_shared_state(&self) -> ReadGuardSharedState {
745 53 : self.mutex.read().await
746 53 : }
747 :
748 : /// Returns commit_lsn watch channel.
749 5 : pub fn get_commit_lsn_watch_rx(&self) -> watch::Receiver<Lsn> {
750 5 : self.commit_lsn_watch_rx.clone()
751 5 : }
752 :
753 : /// Returns term_flush_lsn watch channel.
754 0 : pub fn get_term_flush_lsn_watch_rx(&self) -> watch::Receiver<TermLsn> {
755 0 : self.term_flush_lsn_watch_rx.clone()
756 0 : }
757 :
758 : /// Returns watch channel for SharedState update version.
759 5 : pub fn get_state_version_rx(&self) -> watch::Receiver<usize> {
760 5 : self.shared_state_version_rx.clone()
761 5 : }
762 :
763 : /// Returns wal_seg_size.
764 5 : pub async fn get_wal_seg_size(&self) -> usize {
765 5 : self.read_shared_state().await.get_wal_seg_size()
766 5 : }
767 :
768 : /// Returns state of the timeline.
769 9 : pub async fn get_state(&self) -> (TimelineMemState, TimelinePersistentState) {
770 9 : let state = self.read_shared_state().await;
771 9 : (
772 9 : state.sk.state().inmem.clone(),
773 9 : TimelinePersistentState::clone(state.sk.state()),
774 9 : )
775 9 : }
776 :
777 : /// Returns latest backup_lsn.
778 0 : pub async fn get_wal_backup_lsn(&self) -> Lsn {
779 0 : self.read_shared_state().await.sk.state().inmem.backup_lsn
780 0 : }
781 :
782 : /// Sets backup_lsn to the given value.
783 0 : pub async fn set_wal_backup_lsn(self: &Arc<Self>, backup_lsn: Lsn) -> Result<()> {
784 0 : if self.is_cancelled() {
785 0 : bail!(TimelineError::Cancelled(self.ttid));
786 0 : }
787 :
788 0 : let mut state = self.write_shared_state().await;
789 0 : state.sk.state_mut().inmem.backup_lsn = max(state.sk.state().inmem.backup_lsn, backup_lsn);
790 : // we should check whether to shut down offloader, but this will be done
791 : // soon by peer communication anyway.
792 0 : Ok(())
793 0 : }
794 :
795 : /// Get safekeeper info for broadcasting to broker and other peers.
796 0 : pub async fn get_safekeeper_info(&self, conf: &SafeKeeperConf) -> SafekeeperTimelineInfo {
797 0 : let standby_apply_lsn = self.walsenders.get_hotstandby().reply.apply_lsn;
798 0 : let shared_state = self.read_shared_state().await;
799 0 : shared_state.get_safekeeper_info(&self.ttid, conf, standby_apply_lsn)
800 0 : }
801 :
802 : /// Update timeline state with peer safekeeper data.
803 0 : pub async fn record_safekeeper_info(
804 0 : self: &Arc<Self>,
805 0 : sk_info: SafekeeperTimelineInfo,
806 0 : ) -> Result<()> {
807 : {
808 0 : let mut shared_state = self.write_shared_state().await;
809 0 : shared_state.sk.record_safekeeper_info(&sk_info).await?;
810 0 : let peer_info = peer_info_from_sk_info(&sk_info, Instant::now());
811 0 : shared_state.peers_info.upsert(&peer_info);
812 : }
813 0 : Ok(())
814 0 : }
815 :
816 0 : pub async fn get_peers(&self, conf: &SafeKeeperConf) -> Vec<PeerInfo> {
817 0 : let shared_state = self.read_shared_state().await;
818 0 : shared_state.get_peers(conf.heartbeat_timeout)
819 0 : }
820 :
821 5 : pub fn get_walsenders(&self) -> &Arc<WalSenders> {
822 5 : &self.walsenders
823 5 : }
824 :
825 15 : pub fn get_walreceivers(&self) -> &Arc<WalReceivers> {
826 15 : &self.walreceivers
827 15 : }
828 :
829 : /// Returns flush_lsn.
830 0 : pub async fn get_flush_lsn(&self) -> Lsn {
831 0 : self.read_shared_state().await.sk.flush_lsn()
832 0 : }
833 :
834 : /// Gather timeline data for metrics.
835 0 : pub async fn info_for_metrics(&self) -> Option<FullTimelineInfo> {
836 0 : if self.is_cancelled() {
837 0 : return None;
838 0 : }
839 :
840 : let WalSendersTimelineMetricValues {
841 0 : ps_feedback_counter,
842 0 : last_ps_feedback,
843 0 : interpreted_wal_reader_tasks,
844 0 : } = self.walsenders.info_for_metrics();
845 :
846 0 : let state = self.read_shared_state().await;
847 0 : Some(FullTimelineInfo {
848 0 : ttid: self.ttid,
849 0 : ps_feedback_count: ps_feedback_counter,
850 0 : last_ps_feedback,
851 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
852 0 : timeline_is_active: self.broker_active.load(Ordering::Relaxed),
853 0 : num_computes: self.walreceivers.get_num() as u32,
854 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
855 0 : interpreted_wal_reader_tasks,
856 0 : epoch_start_lsn: state.sk.term_start_lsn(),
857 0 : mem_state: state.sk.state().inmem.clone(),
858 0 : persisted_state: TimelinePersistentState::clone(state.sk.state()),
859 0 : flush_lsn: state.sk.flush_lsn(),
860 0 : wal_storage: state.sk.wal_storage_metrics(),
861 0 : })
862 0 : }
863 :
864 : /// Returns in-memory timeline state to build a full debug dump.
865 0 : pub async fn memory_dump(&self) -> debug_dump::Memory {
866 0 : let state = self.read_shared_state().await;
867 :
868 0 : let (write_lsn, write_record_lsn, flush_lsn, file_open) =
869 0 : state.sk.wal_storage_internal_state();
870 :
871 0 : debug_dump::Memory {
872 0 : is_cancelled: self.is_cancelled(),
873 0 : peers_info_len: state.peers_info.0.len(),
874 0 : walsenders: self.walsenders.get_all_public(),
875 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
876 0 : active: self.broker_active.load(Ordering::Relaxed),
877 0 : num_computes: self.walreceivers.get_num() as u32,
878 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
879 0 : epoch_start_lsn: state.sk.term_start_lsn(),
880 0 : mem_state: state.sk.state().inmem.clone(),
881 0 : mgr_status: self.mgr_status.get(),
882 0 : write_lsn,
883 0 : write_record_lsn,
884 0 : flush_lsn,
885 0 : file_open,
886 0 : }
887 0 : }
888 :
889 : /// Apply a function to the control file state and persist it.
890 0 : pub async fn map_control_file<T>(
891 0 : self: &Arc<Self>,
892 0 : f: impl FnOnce(&mut TimelinePersistentState) -> Result<T>,
893 0 : ) -> Result<T> {
894 0 : let mut state = self.write_shared_state().await;
895 0 : let mut persistent_state = state.sk.state_mut().start_change();
896 : // If f returns error, we abort the change and don't persist anything.
897 0 : let res = f(&mut persistent_state)?;
898 : // If persisting fails, we abort the change and return error.
899 0 : state
900 0 : .sk
901 0 : .state_mut()
902 0 : .finish_change(&persistent_state)
903 0 : .await?;
904 0 : Ok(res)
905 0 : }
906 :
907 0 : pub async fn term_bump(self: &Arc<Self>, to: Option<Term>) -> Result<TimelineTermBumpResponse> {
908 0 : let mut state = self.write_shared_state().await;
909 0 : state.sk.term_bump(to).await
910 0 : }
911 :
912 0 : pub async fn membership_switch(
913 0 : self: &Arc<Self>,
914 0 : to: Configuration,
915 0 : ) -> Result<TimelineMembershipSwitchResponse> {
916 0 : let mut state = self.write_shared_state().await;
917 0 : state.sk.membership_switch(to).await
918 0 : }
919 :
920 : /// Guts of [`Self::wal_residence_guard`] and [`Self::try_wal_residence_guard`]
921 10 : async fn do_wal_residence_guard(
922 10 : self: &Arc<Self>,
923 10 : block: bool,
924 10 : ) -> Result<Option<WalResidentTimeline>> {
925 10 : let op_label = if block {
926 10 : "wal_residence_guard"
927 : } else {
928 0 : "try_wal_residence_guard"
929 : };
930 :
931 10 : if self.is_cancelled() {
932 0 : bail!(TimelineError::Cancelled(self.ttid));
933 10 : }
934 :
935 10 : debug!("requesting WalResidentTimeline guard");
936 10 : let started_at = Instant::now();
937 10 : let status_before = self.mgr_status.get();
938 :
939 : // Wait 30 seconds for the guard to be acquired. It can time out if someone is
940 : // holding the lock (e.g. during `SafeKeeper::process_msg()`) or manager task
941 : // is stuck.
942 10 : let res = tokio::time::timeout_at(started_at + Duration::from_secs(30), async {
943 10 : if block {
944 10 : self.manager_ctl.wal_residence_guard().await.map(Some)
945 : } else {
946 0 : self.manager_ctl.try_wal_residence_guard().await
947 : }
948 10 : })
949 10 : .await;
950 :
951 10 : let guard = match res {
952 10 : Ok(Ok(guard)) => {
953 10 : let finished_at = Instant::now();
954 10 : let elapsed = finished_at - started_at;
955 10 : MISC_OPERATION_SECONDS
956 10 : .with_label_values(&[op_label])
957 10 : .observe(elapsed.as_secs_f64());
958 :
959 10 : guard
960 : }
961 0 : Ok(Err(e)) => {
962 0 : warn!(
963 0 : "error acquiring in {op_label}, statuses {:?} => {:?}",
964 : status_before,
965 0 : self.mgr_status.get()
966 : );
967 0 : return Err(e);
968 : }
969 : Err(_) => {
970 0 : warn!(
971 0 : "timeout acquiring in {op_label} guard, statuses {:?} => {:?}",
972 : status_before,
973 0 : self.mgr_status.get()
974 : );
975 0 : anyhow::bail!("timeout while acquiring WalResidentTimeline guard");
976 : }
977 : };
978 :
979 10 : Ok(guard.map(|g| WalResidentTimeline::new(self.clone(), g)))
980 10 : }
981 :
982 : /// Get the timeline guard for reading/writing WAL files.
983 : /// If WAL files are not present on disk (evicted), they will be automatically
984 : /// downloaded from remote storage. This is done in the manager task, which is
985 : /// responsible for issuing all guards.
986 : ///
987 : /// NB: don't use this function from timeline_manager, it will deadlock.
988 : /// NB: don't use this function while holding shared_state lock.
989 10 : pub async fn wal_residence_guard(self: &Arc<Self>) -> Result<WalResidentTimeline> {
990 10 : self.do_wal_residence_guard(true)
991 10 : .await
992 10 : .map(|m| m.expect("Always get Some in block=true mode"))
993 10 : }
994 :
995 : /// Get the timeline guard for reading/writing WAL files if the timeline is resident,
996 : /// else return None
997 0 : pub(crate) async fn try_wal_residence_guard(
998 0 : self: &Arc<Self>,
999 0 : ) -> Result<Option<WalResidentTimeline>> {
1000 0 : self.do_wal_residence_guard(false).await
1001 0 : }
1002 :
1003 0 : pub async fn backup_partial_reset(self: &Arc<Self>) -> Result<Vec<String>> {
1004 0 : self.manager_ctl.backup_partial_reset().await
1005 0 : }
1006 : }
1007 :
1008 : /// This is a guard that allows to read/write disk timeline state.
1009 : /// All tasks that are trying to read/write WAL from disk should use this guard.
1010 : pub struct WalResidentTimeline {
1011 : pub tli: Arc<Timeline>,
1012 : _guard: ResidenceGuard,
1013 : }
1014 :
1015 : impl WalResidentTimeline {
1016 15 : pub fn new(tli: Arc<Timeline>, _guard: ResidenceGuard) -> Self {
1017 15 : WalResidentTimeline { tli, _guard }
1018 15 : }
1019 : }
1020 :
1021 : impl Deref for WalResidentTimeline {
1022 : type Target = Arc<Timeline>;
1023 :
1024 6911 : fn deref(&self) -> &Self::Target {
1025 6911 : &self.tli
1026 6911 : }
1027 : }
1028 :
1029 : impl WalResidentTimeline {
1030 : /// Returns true if walsender should stop sending WAL to pageserver. We
1031 : /// terminate it if remote_consistent_lsn reached commit_lsn and there is no
1032 : /// computes. While there might be nothing to stream already, we learn about
1033 : /// remote_consistent_lsn update through replication feedback, and we want
1034 : /// to stop pushing to the broker if pageserver is fully caughtup.
1035 0 : pub async fn should_walsender_stop(&self, reported_remote_consistent_lsn: Lsn) -> bool {
1036 0 : if self.is_cancelled() {
1037 0 : return true;
1038 0 : }
1039 0 : let shared_state = self.read_shared_state().await;
1040 0 : if self.walreceivers.get_num() == 0 {
1041 0 : return shared_state.sk.state().inmem.commit_lsn == Lsn(0) || // no data at all yet
1042 0 : reported_remote_consistent_lsn >= shared_state.sk.state().inmem.commit_lsn;
1043 0 : }
1044 0 : false
1045 0 : }
1046 :
1047 : /// Ensure that current term is t, erroring otherwise, and lock the state.
1048 0 : pub async fn acquire_term(&self, t: Term) -> Result<ReadGuardSharedState> {
1049 0 : let ss = self.read_shared_state().await;
1050 0 : if ss.sk.state().acceptor_state.term != t {
1051 0 : bail!(
1052 0 : "failed to acquire term {}, current term {}",
1053 : t,
1054 0 : ss.sk.state().acceptor_state.term
1055 : );
1056 0 : }
1057 0 : Ok(ss)
1058 0 : }
1059 :
1060 : // BEGIN HADRON
1061 : // Check if disk usage by WAL segment files for this timeline exceeds the configured limit.
1062 1240 : fn hadron_check_disk_usage(
1063 1240 : &self,
1064 1240 : shared_state_locked: &mut WriteGuardSharedState<'_>,
1065 1240 : ) -> Result<()> {
1066 : // The disk usage is calculated based on the number of segments between `last_removed_segno`
1067 : // and the current flush LSN segment number. `last_removed_segno` is advanced after
1068 : // unneeded WAL files are physically removed from disk (see `update_wal_removal_end()`
1069 : // in `timeline_manager.rs`).
1070 1240 : let max_timeline_disk_usage_bytes = self.conf.max_timeline_disk_usage_bytes;
1071 1240 : if max_timeline_disk_usage_bytes > 0 {
1072 0 : let last_removed_segno = self.last_removed_segno.load(Ordering::Relaxed);
1073 0 : let flush_lsn = shared_state_locked.sk.flush_lsn();
1074 0 : let wal_seg_size = shared_state_locked.sk.state().server.wal_seg_size as u64;
1075 0 : let current_segno = flush_lsn.segment_number(wal_seg_size as usize);
1076 :
1077 0 : let segno_count = current_segno - last_removed_segno;
1078 0 : let disk_usage_bytes = segno_count * wal_seg_size;
1079 :
1080 0 : if disk_usage_bytes > max_timeline_disk_usage_bytes {
1081 0 : WAL_STORAGE_LIMIT_ERRORS.inc();
1082 0 : bail!(
1083 0 : "WAL storage utilization exceeds configured limit of {} bytes: current disk usage: {} bytes",
1084 : max_timeline_disk_usage_bytes,
1085 : disk_usage_bytes
1086 : );
1087 0 : }
1088 1240 : }
1089 :
1090 1240 : if GLOBAL_DISK_LIMIT_EXCEEDED.load(Ordering::Relaxed) {
1091 0 : bail!("Global disk usage exceeded limit");
1092 1240 : }
1093 :
1094 1240 : Ok(())
1095 1240 : }
1096 : // END HADRON
1097 :
1098 : /// Pass arrived message to the safekeeper.
1099 1240 : pub async fn process_msg(
1100 1240 : &self,
1101 1240 : msg: &ProposerAcceptorMessage,
1102 1240 : ) -> Result<Option<AcceptorProposerMessage>> {
1103 1240 : if self.is_cancelled() {
1104 0 : bail!(TimelineError::Cancelled(self.ttid));
1105 1240 : }
1106 :
1107 : let mut rmsg: Option<AcceptorProposerMessage>;
1108 : {
1109 1240 : let mut shared_state = self.write_shared_state().await;
1110 : // BEGIN HADRON
1111 : // Errors from the `hadron_check_disk_usage()` function fail the process_msg() function, which
1112 : // gets propagated upward and terminates the entire WalAcceptor. This will cause postgres to
1113 : // disconnect from the safekeeper and reestablish another connection. Postgres will keep retrying
1114 : // safekeeper connections every second until it can successfully propose WAL to the SK again.
1115 1240 : self.hadron_check_disk_usage(&mut shared_state)?;
1116 : // END HADRON
1117 1240 : rmsg = shared_state.sk.safekeeper().process_msg(msg).await?;
1118 :
1119 : // if this is AppendResponse, fill in proper hot standby feedback.
1120 620 : if let Some(AcceptorProposerMessage::AppendResponse(ref mut resp)) = rmsg {
1121 620 : resp.hs_feedback = self.walsenders.get_hotstandby().hs_feedback;
1122 620 : }
1123 : }
1124 1240 : Ok(rmsg)
1125 1240 : }
1126 :
1127 9 : pub async fn get_walreader(&self, start_lsn: Lsn) -> Result<WalReader> {
1128 9 : let (_, persisted_state) = self.get_state().await;
1129 :
1130 9 : WalReader::new(
1131 9 : &self.ttid,
1132 9 : self.timeline_dir.clone(),
1133 9 : &persisted_state,
1134 9 : start_lsn,
1135 9 : self.wal_backup.clone(),
1136 : )
1137 9 : }
1138 :
1139 0 : pub fn get_timeline_dir(&self) -> Utf8PathBuf {
1140 0 : self.timeline_dir.clone()
1141 0 : }
1142 :
1143 : /// Update in memory remote consistent lsn.
1144 0 : pub async fn update_remote_consistent_lsn(&self, candidate: Lsn) {
1145 0 : let mut shared_state = self.write_shared_state().await;
1146 0 : shared_state.sk.state_mut().inmem.remote_consistent_lsn = max(
1147 0 : shared_state.sk.state().inmem.remote_consistent_lsn,
1148 0 : candidate,
1149 0 : );
1150 0 : }
1151 : }
1152 :
1153 : /// This struct contains methods that are used by timeline manager task.
1154 : pub(crate) struct ManagerTimeline {
1155 : pub(crate) tli: Arc<Timeline>,
1156 : }
1157 :
1158 : impl Deref for ManagerTimeline {
1159 : type Target = Arc<Timeline>;
1160 :
1161 366 : fn deref(&self) -> &Self::Target {
1162 366 : &self.tli
1163 366 : }
1164 : }
1165 :
1166 : impl ManagerTimeline {
1167 0 : pub(crate) fn timeline_dir(&self) -> &Utf8PathBuf {
1168 0 : &self.tli.timeline_dir
1169 0 : }
1170 :
1171 : /// Manager requests this state on startup.
1172 5 : pub(crate) async fn bootstrap_mgr(&self) -> (bool, Option<PartialRemoteSegment>) {
1173 5 : let shared_state = self.read_shared_state().await;
1174 5 : let is_offloaded = matches!(
1175 5 : shared_state.sk.state().eviction_state,
1176 : EvictionState::Offloaded(_)
1177 : );
1178 5 : let partial_backup_uploaded = shared_state.sk.state().partial_backup.uploaded_segment();
1179 :
1180 5 : (is_offloaded, partial_backup_uploaded)
1181 5 : }
1182 :
1183 : /// Try to switch state Present->Offloaded.
1184 0 : pub(crate) async fn switch_to_offloaded(
1185 0 : &self,
1186 0 : partial: &PartialRemoteSegment,
1187 0 : ) -> anyhow::Result<()> {
1188 0 : let mut shared = self.write_shared_state().await;
1189 :
1190 : // updating control file
1191 0 : let mut pstate = shared.sk.state_mut().start_change();
1192 :
1193 0 : if !matches!(pstate.eviction_state, EvictionState::Present) {
1194 0 : bail!(
1195 0 : "cannot switch to offloaded state, current state is {:?}",
1196 : pstate.eviction_state
1197 : );
1198 0 : }
1199 :
1200 0 : if partial.flush_lsn != shared.sk.flush_lsn() {
1201 0 : bail!(
1202 0 : "flush_lsn mismatch in partial backup, expected {}, got {}",
1203 0 : shared.sk.flush_lsn(),
1204 : partial.flush_lsn
1205 : );
1206 0 : }
1207 :
1208 0 : if partial.commit_lsn != pstate.commit_lsn {
1209 0 : bail!(
1210 0 : "commit_lsn mismatch in partial backup, expected {}, got {}",
1211 : pstate.commit_lsn,
1212 : partial.commit_lsn
1213 : );
1214 0 : }
1215 :
1216 0 : if partial.term != shared.sk.last_log_term() {
1217 0 : bail!(
1218 0 : "term mismatch in partial backup, expected {}, got {}",
1219 0 : shared.sk.last_log_term(),
1220 : partial.term
1221 : );
1222 0 : }
1223 :
1224 0 : pstate.eviction_state = EvictionState::Offloaded(shared.sk.flush_lsn());
1225 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1226 : // control file is now switched to Offloaded state
1227 :
1228 : // now we can switch shared.sk to Offloaded, shouldn't fail
1229 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1230 0 : let cfile_state = prev_sk.take_state();
1231 0 : shared.sk = StateSK::Offloaded(Box::new(cfile_state));
1232 :
1233 0 : Ok(())
1234 0 : }
1235 :
1236 : /// Try to switch state Offloaded->Present.
1237 0 : pub(crate) async fn switch_to_present(&self) -> anyhow::Result<()> {
1238 0 : let mut shared = self.write_shared_state().await;
1239 :
1240 : // trying to restore WAL storage
1241 0 : let wal_store = wal_storage::PhysicalStorage::new(
1242 0 : &self.ttid,
1243 0 : &self.timeline_dir,
1244 0 : shared.sk.state(),
1245 0 : self.conf.no_sync,
1246 0 : )?;
1247 :
1248 : // updating control file
1249 0 : let mut pstate = shared.sk.state_mut().start_change();
1250 :
1251 0 : if !matches!(pstate.eviction_state, EvictionState::Offloaded(_)) {
1252 0 : bail!(
1253 0 : "cannot switch to present state, current state is {:?}",
1254 : pstate.eviction_state
1255 : );
1256 0 : }
1257 :
1258 0 : if wal_store.flush_lsn() != shared.sk.flush_lsn() {
1259 0 : bail!(
1260 0 : "flush_lsn mismatch in restored WAL, expected {}, got {}",
1261 0 : shared.sk.flush_lsn(),
1262 0 : wal_store.flush_lsn()
1263 : );
1264 0 : }
1265 :
1266 0 : pstate.eviction_state = EvictionState::Present;
1267 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1268 :
1269 : // now we can switch shared.sk to Present, shouldn't fail
1270 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1271 0 : let cfile_state = prev_sk.take_state();
1272 0 : shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, self.conf.my_id)?);
1273 :
1274 0 : Ok(())
1275 0 : }
1276 :
1277 : /// Update current manager state, useful for debugging manager deadlocks.
1278 192 : pub(crate) fn set_status(&self, status: timeline_manager::Status) {
1279 192 : self.mgr_status.store(status, Ordering::Relaxed);
1280 192 : }
1281 : }
1282 :
1283 : /// Deletes directory and it's contents. Returns false if directory does not exist.
1284 0 : pub async fn delete_dir(path: &Utf8PathBuf) -> Result<bool> {
1285 0 : match fs::remove_dir_all(path).await {
1286 0 : Ok(_) => Ok(true),
1287 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(false),
1288 0 : Err(e) => Err(e.into()),
1289 : }
1290 0 : }
1291 :
1292 : /// Get a path to the tenant directory. If you just need to get a timeline directory,
1293 : /// use WalResidentTimeline::get_timeline_dir instead.
1294 10 : pub fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
1295 10 : conf.workdir.join(tenant_id.to_string())
1296 10 : }
1297 :
1298 : /// Get a path to the timeline directory. If you need to read WAL files from disk,
1299 : /// use WalResidentTimeline::get_timeline_dir instead. This function does not check
1300 : /// timeline eviction status and WAL files might not be present on disk.
1301 10 : pub fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
1302 10 : get_tenant_dir(conf, &ttid.tenant_id).join(ttid.timeline_id.to_string())
1303 10 : }
|
