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 2580 : pub fn flush_lsn(&self) -> Lsn {
156 2580 : match self {
157 2580 : 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 2580 : }
165 :
166 : /// Get a reference to the control file's timeline state.
167 2613 : pub fn state(&self) -> &TimelineState<control_file::FileStorage> {
168 2613 : match self {
169 2613 : StateSK::Loaded(sk) => &sk.state,
170 0 : StateSK::Offloaded(s) => s,
171 0 : StateSK::Empty => unreachable!(),
172 : }
173 2613 : }
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 1290 : pub fn last_log_term(&self) -> Term {
184 1290 : self.state()
185 1290 : .acceptor_state
186 1290 : .get_last_log_term(self.flush_lsn())
187 1290 : }
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 36 : pub(crate) fn get_peers(&self, heartbeat_timeout: Duration) -> Vec<PeerInfo> {
392 36 : let now = Instant::now();
393 36 : self.peers_info
394 36 : .0
395 36 : .iter()
396 : // Regard peer as absent if we haven't heard from it within heartbeat_timeout.
397 36 : .filter(|p| now.duration_since(p.ts) <= heartbeat_timeout)
398 36 : .cloned()
399 36 : .collect()
400 36 : }
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 : _ => ApiError::InternalServerError(anyhow!("{}", te)),
431 : }
432 0 : }
433 : }
434 :
435 : /// We run remote deletion in a background task, this is how it sends its results back.
436 : type RemoteDeletionReceiver = tokio::sync::watch::Receiver<Option<anyhow::Result<()>>>;
437 :
438 : /// Timeline struct manages lifecycle (creation, deletion, restore) of a safekeeper timeline.
439 : /// It also holds SharedState and provides mutually exclusive access to it.
440 : pub struct Timeline {
441 : pub ttid: TenantTimelineId,
442 : pub remote_path: RemotePath,
443 :
444 : /// Used to broadcast commit_lsn updates to all background jobs.
445 : commit_lsn_watch_tx: watch::Sender<Lsn>,
446 : commit_lsn_watch_rx: watch::Receiver<Lsn>,
447 :
448 : /// Broadcasts (current term, flush_lsn) updates, walsender is interested in
449 : /// them when sending in recovery mode (to walproposer or peers). Note: this
450 : /// is just a notification, WAL reading should always done with lock held as
451 : /// term can change otherwise.
452 : term_flush_lsn_watch_tx: watch::Sender<TermLsn>,
453 : term_flush_lsn_watch_rx: watch::Receiver<TermLsn>,
454 :
455 : /// Broadcasts shared state updates.
456 : shared_state_version_tx: watch::Sender<usize>,
457 : shared_state_version_rx: watch::Receiver<usize>,
458 :
459 : /// Safekeeper and other state, that should remain consistent and
460 : /// synchronized with the disk. This is tokio mutex as we write WAL to disk
461 : /// while holding it, ensuring that consensus checks are in order.
462 : mutex: RwLock<SharedState>,
463 : walsenders: Arc<WalSenders>,
464 : walreceivers: Arc<WalReceivers>,
465 : timeline_dir: Utf8PathBuf,
466 : manager_ctl: ManagerCtl,
467 : conf: Arc<SafeKeeperConf>,
468 :
469 : pub(crate) wal_backup: Arc<WalBackup>,
470 :
471 : remote_deletion: std::sync::Mutex<Option<RemoteDeletionReceiver>>,
472 :
473 : /// Hold this gate from code that depends on the Timeline's non-shut-down state. While holding
474 : /// this gate, you must respect [`Timeline::cancel`]
475 : pub(crate) gate: Gate,
476 :
477 : /// Delete/cancel will trigger this, background tasks should drop out as soon as it fires
478 : pub(crate) cancel: CancellationToken,
479 :
480 : // timeline_manager controlled state
481 : pub(crate) broker_active: AtomicBool,
482 : pub(crate) wal_backup_active: AtomicBool,
483 : pub(crate) last_removed_segno: AtomicU64,
484 : pub(crate) mgr_status: AtomicStatus,
485 : }
486 :
487 : impl Timeline {
488 : /// Constructs a new timeline.
489 5 : pub fn new(
490 5 : ttid: TenantTimelineId,
491 5 : timeline_dir: &Utf8Path,
492 5 : remote_path: &RemotePath,
493 5 : shared_state: SharedState,
494 5 : conf: Arc<SafeKeeperConf>,
495 5 : wal_backup: Arc<WalBackup>,
496 5 : ) -> Arc<Self> {
497 5 : let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
498 5 : watch::channel(shared_state.sk.state().commit_lsn);
499 5 : let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
500 5 : shared_state.sk.last_log_term(),
501 5 : shared_state.sk.flush_lsn(),
502 5 : )));
503 5 : let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
504 :
505 5 : let walreceivers = WalReceivers::new();
506 :
507 5 : Arc::new(Self {
508 5 : ttid,
509 5 : remote_path: remote_path.to_owned(),
510 5 : timeline_dir: timeline_dir.to_owned(),
511 5 : commit_lsn_watch_tx,
512 5 : commit_lsn_watch_rx,
513 5 : term_flush_lsn_watch_tx,
514 5 : term_flush_lsn_watch_rx,
515 5 : shared_state_version_tx,
516 5 : shared_state_version_rx,
517 5 : mutex: RwLock::new(shared_state),
518 5 : walsenders: WalSenders::new(walreceivers.clone()),
519 5 : walreceivers,
520 5 : gate: Default::default(),
521 5 : cancel: CancellationToken::default(),
522 5 : remote_deletion: std::sync::Mutex::new(None),
523 5 : manager_ctl: ManagerCtl::new(),
524 5 : conf,
525 5 : broker_active: AtomicBool::new(false),
526 5 : wal_backup_active: AtomicBool::new(false),
527 5 : last_removed_segno: AtomicU64::new(0),
528 5 : mgr_status: AtomicStatus::new(),
529 5 : wal_backup,
530 5 : })
531 5 : }
532 :
533 : /// Load existing timeline from disk.
534 0 : pub fn load_timeline(
535 0 : conf: Arc<SafeKeeperConf>,
536 0 : ttid: TenantTimelineId,
537 0 : wal_backup: Arc<WalBackup>,
538 0 : ) -> Result<Arc<Timeline>> {
539 0 : let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
540 :
541 0 : let shared_state = SharedState::restore(conf.as_ref(), &ttid)?;
542 0 : let timeline_dir = get_timeline_dir(conf.as_ref(), &ttid);
543 0 : let remote_path = remote_timeline_path(&ttid)?;
544 :
545 0 : Ok(Timeline::new(
546 0 : ttid,
547 0 : &timeline_dir,
548 0 : &remote_path,
549 0 : shared_state,
550 0 : conf,
551 0 : wal_backup,
552 0 : ))
553 0 : }
554 :
555 : /// Bootstrap new or existing timeline starting background tasks.
556 5 : pub fn bootstrap(
557 5 : self: &Arc<Timeline>,
558 5 : _shared_state: &mut WriteGuardSharedState<'_>,
559 5 : conf: &SafeKeeperConf,
560 5 : broker_active_set: Arc<TimelinesSet>,
561 5 : partial_backup_rate_limiter: RateLimiter,
562 5 : wal_backup: Arc<WalBackup>,
563 5 : ) {
564 5 : let (tx, rx) = self.manager_ctl.bootstrap_manager();
565 :
566 5 : let Ok(gate_guard) = self.gate.enter() else {
567 : // Init raced with shutdown
568 0 : return;
569 : };
570 :
571 : // Start manager task which will monitor timeline state and update
572 : // background tasks.
573 5 : tokio::spawn({
574 5 : let this = self.clone();
575 5 : let conf = conf.clone();
576 5 : async move {
577 5 : let _gate_guard = gate_guard;
578 5 : timeline_manager::main_task(
579 5 : ManagerTimeline { tli: this },
580 5 : conf,
581 5 : broker_active_set,
582 5 : tx,
583 5 : rx,
584 5 : partial_backup_rate_limiter,
585 5 : wal_backup,
586 5 : )
587 5 : .await
588 0 : }
589 : });
590 5 : }
591 :
592 : /// Cancel the timeline, requesting background activity to stop. Closing
593 : /// the `self.gate` waits for that.
594 0 : pub async fn cancel(&self) {
595 0 : info!("timeline {} shutting down", self.ttid);
596 0 : self.cancel.cancel();
597 0 : }
598 :
599 : /// Background timeline activities (which hold Timeline::gate) will no
600 : /// longer run once this function completes. `Self::cancel` must have been
601 : /// already called.
602 0 : pub async fn close(&self) {
603 0 : assert!(self.cancel.is_cancelled());
604 :
605 : // Wait for any concurrent tasks to stop using this timeline, to avoid e.g. attempts
606 : // to read deleted files.
607 0 : self.gate.close().await;
608 0 : }
609 :
610 : /// Delete timeline from disk completely, by removing timeline directory.
611 : ///
612 : /// Also deletes WAL in s3. Might fail if e.g. s3 is unavailable, but
613 : /// deletion API endpoint is retriable.
614 : ///
615 : /// Timeline must be in shut-down state (i.e. call [`Self::close`] first)
616 0 : pub async fn delete(
617 0 : &self,
618 0 : shared_state: &mut WriteGuardSharedState<'_>,
619 0 : only_local: bool,
620 0 : ) -> Result<bool> {
621 : // Assert that [`Self::close`] was already called
622 0 : assert!(self.cancel.is_cancelled());
623 0 : assert!(self.gate.close_complete());
624 :
625 0 : info!("deleting timeline {} from disk", self.ttid);
626 :
627 : // Close associated FDs. Nobody will be able to touch timeline data once
628 : // it is cancelled, so WAL storage won't be opened again.
629 0 : shared_state.sk.close_wal_store();
630 :
631 0 : if !only_local {
632 0 : self.remote_delete().await?;
633 0 : }
634 :
635 0 : let dir_existed = delete_dir(&self.timeline_dir).await?;
636 0 : Ok(dir_existed)
637 0 : }
638 :
639 : /// Delete timeline content from remote storage. If the returned future is dropped,
640 : /// deletion will continue in the background.
641 : ///
642 : /// This function ordinarily spawns a task and stashes a result receiver into [`Self::remote_deletion`]. If
643 : /// deletion is already happening, it may simply wait for an existing task's result.
644 : ///
645 : /// Note: we concurrently delete remote storage data from multiple
646 : /// safekeepers. That's ok, s3 replies 200 if object doesn't exist and we
647 : /// do some retries anyway.
648 0 : async fn remote_delete(&self) -> Result<()> {
649 : // We will start a background task to do the deletion, so that it proceeds even if our
650 : // API request is dropped. Future requests will see the existing deletion task and wait
651 : // for it to complete.
652 0 : let mut result_rx = {
653 0 : let mut remote_deletion_state = self.remote_deletion.lock().unwrap();
654 0 : let result_rx = if let Some(result_rx) = remote_deletion_state.as_ref() {
655 0 : if let Some(result) = result_rx.borrow().as_ref() {
656 0 : if let Err(e) = result {
657 : // A previous remote deletion failed: we will start a new one
658 0 : tracing::error!("remote deletion failed, will retry ({e})");
659 0 : None
660 : } else {
661 : // A previous remote deletion call already succeeded
662 0 : return Ok(());
663 : }
664 : } else {
665 : // Remote deletion is still in flight
666 0 : Some(result_rx.clone())
667 : }
668 : } else {
669 : // Remote deletion was not attempted yet, start it now.
670 0 : None
671 : };
672 :
673 0 : match result_rx {
674 0 : Some(result_rx) => result_rx,
675 0 : None => self.start_remote_delete(&mut remote_deletion_state),
676 : }
677 : };
678 :
679 : // Wait for a result
680 0 : let Ok(result) = result_rx.wait_for(|v| v.is_some()).await else {
681 : // Unexpected: sender should always send a result before dropping the channel, even if it has an error
682 0 : return Err(anyhow::anyhow!(
683 0 : "remote deletion task future was dropped without sending a result"
684 0 : ));
685 : };
686 :
687 0 : result
688 0 : .as_ref()
689 0 : .expect("We did a wait_for on this being Some above")
690 0 : .as_ref()
691 0 : .map(|_| ())
692 0 : .map_err(|e| anyhow::anyhow!("remote deletion failed: {e}"))
693 0 : }
694 :
695 : /// Spawn background task to do remote deletion, return a receiver for its outcome
696 0 : fn start_remote_delete(
697 0 : &self,
698 0 : guard: &mut std::sync::MutexGuard<Option<RemoteDeletionReceiver>>,
699 0 : ) -> RemoteDeletionReceiver {
700 0 : tracing::info!("starting remote deletion");
701 0 : let storage = self.wal_backup.get_storage().clone();
702 0 : let (result_tx, result_rx) = tokio::sync::watch::channel(None);
703 0 : let ttid = self.ttid;
704 0 : tokio::task::spawn(
705 0 : async move {
706 0 : let r = if let Some(storage) = storage {
707 0 : wal_backup::delete_timeline(&storage, &ttid).await
708 : } else {
709 0 : tracing::info!(
710 0 : "skipping remote deletion because no remote storage is configured; this effectively leaks the objects in remote storage"
711 : );
712 0 : Ok(())
713 : };
714 :
715 0 : if let Err(e) = &r {
716 : // Log error here in case nobody ever listens for our result (e.g. dropped API request)
717 0 : tracing::error!("remote deletion failed: {e}");
718 0 : }
719 :
720 : // Ignore send results: it's legal for the Timeline to give up waiting for us.
721 0 : let _ = result_tx.send(Some(r));
722 0 : }
723 0 : .instrument(info_span!("remote_delete", timeline = %self.ttid)),
724 : );
725 :
726 0 : **guard = Some(result_rx.clone());
727 :
728 0 : result_rx
729 0 : }
730 :
731 : /// Returns if timeline is cancelled.
732 2505 : pub fn is_cancelled(&self) -> bool {
733 2505 : self.cancel.is_cancelled()
734 2505 : }
735 :
736 : /// Take a writing mutual exclusive lock on timeline shared_state.
737 1249 : pub async fn write_shared_state(self: &Arc<Self>) -> WriteGuardSharedState<'_> {
738 1249 : WriteGuardSharedState::new(self.clone(), self.mutex.write().await)
739 1249 : }
740 :
741 55 : pub async fn read_shared_state(&self) -> ReadGuardSharedState {
742 55 : self.mutex.read().await
743 55 : }
744 :
745 : /// Returns commit_lsn watch channel.
746 5 : pub fn get_commit_lsn_watch_rx(&self) -> watch::Receiver<Lsn> {
747 5 : self.commit_lsn_watch_rx.clone()
748 5 : }
749 :
750 : /// Returns term_flush_lsn watch channel.
751 0 : pub fn get_term_flush_lsn_watch_rx(&self) -> watch::Receiver<TermLsn> {
752 0 : self.term_flush_lsn_watch_rx.clone()
753 0 : }
754 :
755 : /// Returns watch channel for SharedState update version.
756 5 : pub fn get_state_version_rx(&self) -> watch::Receiver<usize> {
757 5 : self.shared_state_version_rx.clone()
758 5 : }
759 :
760 : /// Returns wal_seg_size.
761 5 : pub async fn get_wal_seg_size(&self) -> usize {
762 5 : self.read_shared_state().await.get_wal_seg_size()
763 5 : }
764 :
765 : /// Returns state of the timeline.
766 9 : pub async fn get_state(&self) -> (TimelineMemState, TimelinePersistentState) {
767 9 : let state = self.read_shared_state().await;
768 9 : (
769 9 : state.sk.state().inmem.clone(),
770 9 : TimelinePersistentState::clone(state.sk.state()),
771 9 : )
772 9 : }
773 :
774 : /// Returns latest backup_lsn.
775 0 : pub async fn get_wal_backup_lsn(&self) -> Lsn {
776 0 : self.read_shared_state().await.sk.state().inmem.backup_lsn
777 0 : }
778 :
779 : /// Sets backup_lsn to the given value.
780 0 : pub async fn set_wal_backup_lsn(self: &Arc<Self>, backup_lsn: Lsn) -> Result<()> {
781 0 : if self.is_cancelled() {
782 0 : bail!(TimelineError::Cancelled(self.ttid));
783 0 : }
784 :
785 0 : let mut state = self.write_shared_state().await;
786 0 : state.sk.state_mut().inmem.backup_lsn = max(state.sk.state().inmem.backup_lsn, backup_lsn);
787 : // we should check whether to shut down offloader, but this will be done
788 : // soon by peer communication anyway.
789 0 : Ok(())
790 0 : }
791 :
792 : /// Get safekeeper info for broadcasting to broker and other peers.
793 0 : pub async fn get_safekeeper_info(&self, conf: &SafeKeeperConf) -> SafekeeperTimelineInfo {
794 0 : let standby_apply_lsn = self.walsenders.get_hotstandby().reply.apply_lsn;
795 0 : let shared_state = self.read_shared_state().await;
796 0 : shared_state.get_safekeeper_info(&self.ttid, conf, standby_apply_lsn)
797 0 : }
798 :
799 : /// Update timeline state with peer safekeeper data.
800 0 : pub async fn record_safekeeper_info(
801 0 : self: &Arc<Self>,
802 0 : sk_info: SafekeeperTimelineInfo,
803 0 : ) -> Result<()> {
804 : {
805 0 : let mut shared_state = self.write_shared_state().await;
806 0 : shared_state.sk.record_safekeeper_info(&sk_info).await?;
807 0 : let peer_info = peer_info_from_sk_info(&sk_info, Instant::now());
808 0 : shared_state.peers_info.upsert(&peer_info);
809 : }
810 0 : Ok(())
811 0 : }
812 :
813 0 : pub async fn get_peers(&self, conf: &SafeKeeperConf) -> Vec<PeerInfo> {
814 0 : let shared_state = self.read_shared_state().await;
815 0 : shared_state.get_peers(conf.heartbeat_timeout)
816 0 : }
817 :
818 5 : pub fn get_walsenders(&self) -> &Arc<WalSenders> {
819 5 : &self.walsenders
820 5 : }
821 :
822 15 : pub fn get_walreceivers(&self) -> &Arc<WalReceivers> {
823 15 : &self.walreceivers
824 15 : }
825 :
826 : /// Returns flush_lsn.
827 0 : pub async fn get_flush_lsn(&self) -> Lsn {
828 0 : self.read_shared_state().await.sk.flush_lsn()
829 0 : }
830 :
831 : /// Gather timeline data for metrics.
832 0 : pub async fn info_for_metrics(&self) -> Option<FullTimelineInfo> {
833 0 : if self.is_cancelled() {
834 0 : return None;
835 0 : }
836 :
837 : let WalSendersTimelineMetricValues {
838 0 : ps_feedback_counter,
839 0 : last_ps_feedback,
840 0 : interpreted_wal_reader_tasks,
841 0 : } = self.walsenders.info_for_metrics();
842 :
843 0 : let state = self.read_shared_state().await;
844 0 : Some(FullTimelineInfo {
845 0 : ttid: self.ttid,
846 0 : ps_feedback_count: ps_feedback_counter,
847 0 : last_ps_feedback,
848 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
849 0 : timeline_is_active: self.broker_active.load(Ordering::Relaxed),
850 0 : num_computes: self.walreceivers.get_num() as u32,
851 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
852 0 : interpreted_wal_reader_tasks,
853 0 : epoch_start_lsn: state.sk.term_start_lsn(),
854 0 : mem_state: state.sk.state().inmem.clone(),
855 0 : persisted_state: TimelinePersistentState::clone(state.sk.state()),
856 0 : flush_lsn: state.sk.flush_lsn(),
857 0 : wal_storage: state.sk.wal_storage_metrics(),
858 0 : })
859 0 : }
860 :
861 : /// Returns in-memory timeline state to build a full debug dump.
862 0 : pub async fn memory_dump(&self) -> debug_dump::Memory {
863 0 : let state = self.read_shared_state().await;
864 :
865 0 : let (write_lsn, write_record_lsn, flush_lsn, file_open) =
866 0 : state.sk.wal_storage_internal_state();
867 :
868 0 : debug_dump::Memory {
869 0 : is_cancelled: self.is_cancelled(),
870 0 : peers_info_len: state.peers_info.0.len(),
871 0 : walsenders: self.walsenders.get_all_public(),
872 0 : wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
873 0 : active: self.broker_active.load(Ordering::Relaxed),
874 0 : num_computes: self.walreceivers.get_num() as u32,
875 0 : last_removed_segno: self.last_removed_segno.load(Ordering::Relaxed),
876 0 : epoch_start_lsn: state.sk.term_start_lsn(),
877 0 : mem_state: state.sk.state().inmem.clone(),
878 0 : mgr_status: self.mgr_status.get(),
879 0 : write_lsn,
880 0 : write_record_lsn,
881 0 : flush_lsn,
882 0 : file_open,
883 0 : }
884 0 : }
885 :
886 : /// Apply a function to the control file state and persist it.
887 0 : pub async fn map_control_file<T>(
888 0 : self: &Arc<Self>,
889 0 : f: impl FnOnce(&mut TimelinePersistentState) -> Result<T>,
890 0 : ) -> Result<T> {
891 0 : let mut state = self.write_shared_state().await;
892 0 : let mut persistent_state = state.sk.state_mut().start_change();
893 : // If f returns error, we abort the change and don't persist anything.
894 0 : let res = f(&mut persistent_state)?;
895 : // If persisting fails, we abort the change and return error.
896 0 : state
897 0 : .sk
898 0 : .state_mut()
899 0 : .finish_change(&persistent_state)
900 0 : .await?;
901 0 : Ok(res)
902 0 : }
903 :
904 0 : pub async fn term_bump(self: &Arc<Self>, to: Option<Term>) -> Result<TimelineTermBumpResponse> {
905 0 : let mut state = self.write_shared_state().await;
906 0 : state.sk.term_bump(to).await
907 0 : }
908 :
909 0 : pub async fn membership_switch(
910 0 : self: &Arc<Self>,
911 0 : to: Configuration,
912 0 : ) -> Result<TimelineMembershipSwitchResponse> {
913 0 : let mut state = self.write_shared_state().await;
914 0 : state.sk.membership_switch(to).await
915 0 : }
916 :
917 : /// Guts of [`Self::wal_residence_guard`] and [`Self::try_wal_residence_guard`]
918 10 : async fn do_wal_residence_guard(
919 10 : self: &Arc<Self>,
920 10 : block: bool,
921 10 : ) -> Result<Option<WalResidentTimeline>> {
922 10 : let op_label = if block {
923 10 : "wal_residence_guard"
924 : } else {
925 0 : "try_wal_residence_guard"
926 : };
927 :
928 10 : if self.is_cancelled() {
929 0 : bail!(TimelineError::Cancelled(self.ttid));
930 10 : }
931 :
932 10 : debug!("requesting WalResidentTimeline guard");
933 10 : let started_at = Instant::now();
934 10 : let status_before = self.mgr_status.get();
935 :
936 : // Wait 30 seconds for the guard to be acquired. It can time out if someone is
937 : // holding the lock (e.g. during `SafeKeeper::process_msg()`) or manager task
938 : // is stuck.
939 10 : let res = tokio::time::timeout_at(started_at + Duration::from_secs(30), async {
940 10 : if block {
941 10 : self.manager_ctl.wal_residence_guard().await.map(Some)
942 : } else {
943 0 : self.manager_ctl.try_wal_residence_guard().await
944 : }
945 10 : })
946 10 : .await;
947 :
948 10 : let guard = match res {
949 10 : Ok(Ok(guard)) => {
950 10 : let finished_at = Instant::now();
951 10 : let elapsed = finished_at - started_at;
952 10 : MISC_OPERATION_SECONDS
953 10 : .with_label_values(&[op_label])
954 10 : .observe(elapsed.as_secs_f64());
955 :
956 10 : guard
957 : }
958 0 : Ok(Err(e)) => {
959 0 : warn!(
960 0 : "error acquiring in {op_label}, statuses {:?} => {:?}",
961 : status_before,
962 0 : self.mgr_status.get()
963 : );
964 0 : return Err(e);
965 : }
966 : Err(_) => {
967 0 : warn!(
968 0 : "timeout acquiring in {op_label} guard, statuses {:?} => {:?}",
969 : status_before,
970 0 : self.mgr_status.get()
971 : );
972 0 : anyhow::bail!("timeout while acquiring WalResidentTimeline guard");
973 : }
974 : };
975 :
976 10 : Ok(guard.map(|g| WalResidentTimeline::new(self.clone(), g)))
977 10 : }
978 :
979 : /// Get the timeline guard for reading/writing WAL files.
980 : /// If WAL files are not present on disk (evicted), they will be automatically
981 : /// downloaded from remote storage. This is done in the manager task, which is
982 : /// responsible for issuing all guards.
983 : ///
984 : /// NB: don't use this function from timeline_manager, it will deadlock.
985 : /// NB: don't use this function while holding shared_state lock.
986 10 : pub async fn wal_residence_guard(self: &Arc<Self>) -> Result<WalResidentTimeline> {
987 10 : self.do_wal_residence_guard(true)
988 10 : .await
989 10 : .map(|m| m.expect("Always get Some in block=true mode"))
990 10 : }
991 :
992 : /// Get the timeline guard for reading/writing WAL files if the timeline is resident,
993 : /// else return None
994 0 : pub(crate) async fn try_wal_residence_guard(
995 0 : self: &Arc<Self>,
996 0 : ) -> Result<Option<WalResidentTimeline>> {
997 0 : self.do_wal_residence_guard(false).await
998 0 : }
999 :
1000 0 : pub async fn backup_partial_reset(self: &Arc<Self>) -> Result<Vec<String>> {
1001 0 : self.manager_ctl.backup_partial_reset().await
1002 0 : }
1003 : }
1004 :
1005 : /// This is a guard that allows to read/write disk timeline state.
1006 : /// All tasks that are trying to read/write WAL from disk should use this guard.
1007 : pub struct WalResidentTimeline {
1008 : pub tli: Arc<Timeline>,
1009 : _guard: ResidenceGuard,
1010 : }
1011 :
1012 : impl WalResidentTimeline {
1013 15 : pub fn new(tli: Arc<Timeline>, _guard: ResidenceGuard) -> Self {
1014 15 : WalResidentTimeline { tli, _guard }
1015 15 : }
1016 : }
1017 :
1018 : impl Deref for WalResidentTimeline {
1019 : type Target = Arc<Timeline>;
1020 :
1021 6911 : fn deref(&self) -> &Self::Target {
1022 6911 : &self.tli
1023 6911 : }
1024 : }
1025 :
1026 : impl WalResidentTimeline {
1027 : /// Returns true if walsender should stop sending WAL to pageserver. We
1028 : /// terminate it if remote_consistent_lsn reached commit_lsn and there is no
1029 : /// computes. While there might be nothing to stream already, we learn about
1030 : /// remote_consistent_lsn update through replication feedback, and we want
1031 : /// to stop pushing to the broker if pageserver is fully caughtup.
1032 0 : pub async fn should_walsender_stop(&self, reported_remote_consistent_lsn: Lsn) -> bool {
1033 0 : if self.is_cancelled() {
1034 0 : return true;
1035 0 : }
1036 0 : let shared_state = self.read_shared_state().await;
1037 0 : if self.walreceivers.get_num() == 0 {
1038 0 : return shared_state.sk.state().inmem.commit_lsn == Lsn(0) || // no data at all yet
1039 0 : reported_remote_consistent_lsn >= shared_state.sk.state().inmem.commit_lsn;
1040 0 : }
1041 0 : false
1042 0 : }
1043 :
1044 : /// Ensure that current term is t, erroring otherwise, and lock the state.
1045 0 : pub async fn acquire_term(&self, t: Term) -> Result<ReadGuardSharedState> {
1046 0 : let ss = self.read_shared_state().await;
1047 0 : if ss.sk.state().acceptor_state.term != t {
1048 0 : bail!(
1049 0 : "failed to acquire term {}, current term {}",
1050 : t,
1051 0 : ss.sk.state().acceptor_state.term
1052 : );
1053 0 : }
1054 0 : Ok(ss)
1055 0 : }
1056 :
1057 : // BEGIN HADRON
1058 : // Check if disk usage by WAL segment files for this timeline exceeds the configured limit.
1059 1240 : fn hadron_check_disk_usage(
1060 1240 : &self,
1061 1240 : shared_state_locked: &mut WriteGuardSharedState<'_>,
1062 1240 : ) -> Result<()> {
1063 : // The disk usage is calculated based on the number of segments between `last_removed_segno`
1064 : // and the current flush LSN segment number. `last_removed_segno` is advanced after
1065 : // unneeded WAL files are physically removed from disk (see `update_wal_removal_end()`
1066 : // in `timeline_manager.rs`).
1067 1240 : let max_timeline_disk_usage_bytes = self.conf.max_timeline_disk_usage_bytes;
1068 1240 : if max_timeline_disk_usage_bytes > 0 {
1069 0 : let last_removed_segno = self.last_removed_segno.load(Ordering::Relaxed);
1070 0 : let flush_lsn = shared_state_locked.sk.flush_lsn();
1071 0 : let wal_seg_size = shared_state_locked.sk.state().server.wal_seg_size as u64;
1072 0 : let current_segno = flush_lsn.segment_number(wal_seg_size as usize);
1073 :
1074 0 : let segno_count = current_segno - last_removed_segno;
1075 0 : let disk_usage_bytes = segno_count * wal_seg_size;
1076 :
1077 0 : if disk_usage_bytes > max_timeline_disk_usage_bytes {
1078 0 : WAL_STORAGE_LIMIT_ERRORS.inc();
1079 0 : bail!(
1080 0 : "WAL storage utilization exceeds configured limit of {} bytes: current disk usage: {} bytes",
1081 : max_timeline_disk_usage_bytes,
1082 : disk_usage_bytes
1083 : );
1084 0 : }
1085 1240 : }
1086 :
1087 1240 : if GLOBAL_DISK_LIMIT_EXCEEDED.load(Ordering::Relaxed) {
1088 0 : bail!("Global disk usage exceeded limit");
1089 1240 : }
1090 :
1091 1240 : Ok(())
1092 1240 : }
1093 : // END HADRON
1094 :
1095 : /// Pass arrived message to the safekeeper.
1096 1240 : pub async fn process_msg(
1097 1240 : &self,
1098 1240 : msg: &ProposerAcceptorMessage,
1099 1240 : ) -> Result<Option<AcceptorProposerMessage>> {
1100 1240 : if self.is_cancelled() {
1101 0 : bail!(TimelineError::Cancelled(self.ttid));
1102 1240 : }
1103 :
1104 : let mut rmsg: Option<AcceptorProposerMessage>;
1105 : {
1106 1240 : let mut shared_state = self.write_shared_state().await;
1107 : // BEGIN HADRON
1108 : // Errors from the `hadron_check_disk_usage()` function fail the process_msg() function, which
1109 : // gets propagated upward and terminates the entire WalAcceptor. This will cause postgres to
1110 : // disconnect from the safekeeper and reestablish another connection. Postgres will keep retrying
1111 : // safekeeper connections every second until it can successfully propose WAL to the SK again.
1112 1240 : self.hadron_check_disk_usage(&mut shared_state)?;
1113 : // END HADRON
1114 1240 : rmsg = shared_state.sk.safekeeper().process_msg(msg).await?;
1115 :
1116 : // if this is AppendResponse, fill in proper hot standby feedback.
1117 620 : if let Some(AcceptorProposerMessage::AppendResponse(ref mut resp)) = rmsg {
1118 620 : resp.hs_feedback = self.walsenders.get_hotstandby().hs_feedback;
1119 620 : }
1120 : }
1121 1240 : Ok(rmsg)
1122 1240 : }
1123 :
1124 9 : pub async fn get_walreader(&self, start_lsn: Lsn) -> Result<WalReader> {
1125 9 : let (_, persisted_state) = self.get_state().await;
1126 :
1127 9 : WalReader::new(
1128 9 : &self.ttid,
1129 9 : self.timeline_dir.clone(),
1130 9 : &persisted_state,
1131 9 : start_lsn,
1132 9 : self.wal_backup.clone(),
1133 : )
1134 9 : }
1135 :
1136 0 : pub fn get_timeline_dir(&self) -> Utf8PathBuf {
1137 0 : self.timeline_dir.clone()
1138 0 : }
1139 :
1140 : /// Update in memory remote consistent lsn.
1141 0 : pub async fn update_remote_consistent_lsn(&self, candidate: Lsn) {
1142 0 : let mut shared_state = self.write_shared_state().await;
1143 0 : shared_state.sk.state_mut().inmem.remote_consistent_lsn = max(
1144 0 : shared_state.sk.state().inmem.remote_consistent_lsn,
1145 0 : candidate,
1146 0 : );
1147 0 : }
1148 : }
1149 :
1150 : /// This struct contains methods that are used by timeline manager task.
1151 : pub(crate) struct ManagerTimeline {
1152 : pub(crate) tli: Arc<Timeline>,
1153 : }
1154 :
1155 : impl Deref for ManagerTimeline {
1156 : type Target = Arc<Timeline>;
1157 :
1158 391 : fn deref(&self) -> &Self::Target {
1159 391 : &self.tli
1160 391 : }
1161 : }
1162 :
1163 : impl ManagerTimeline {
1164 0 : pub(crate) fn timeline_dir(&self) -> &Utf8PathBuf {
1165 0 : &self.tli.timeline_dir
1166 0 : }
1167 :
1168 : /// Manager requests this state on startup.
1169 5 : pub(crate) async fn bootstrap_mgr(&self) -> (bool, Option<PartialRemoteSegment>) {
1170 5 : let shared_state = self.read_shared_state().await;
1171 5 : let is_offloaded = matches!(
1172 5 : shared_state.sk.state().eviction_state,
1173 : EvictionState::Offloaded(_)
1174 : );
1175 5 : let partial_backup_uploaded = shared_state.sk.state().partial_backup.uploaded_segment();
1176 :
1177 5 : (is_offloaded, partial_backup_uploaded)
1178 5 : }
1179 :
1180 : /// Try to switch state Present->Offloaded.
1181 0 : pub(crate) async fn switch_to_offloaded(
1182 0 : &self,
1183 0 : partial: &PartialRemoteSegment,
1184 0 : ) -> anyhow::Result<()> {
1185 0 : let mut shared = self.write_shared_state().await;
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::Present) {
1191 0 : bail!(
1192 0 : "cannot switch to offloaded state, current state is {:?}",
1193 : pstate.eviction_state
1194 : );
1195 0 : }
1196 :
1197 0 : if partial.flush_lsn != shared.sk.flush_lsn() {
1198 0 : bail!(
1199 0 : "flush_lsn mismatch in partial backup, expected {}, got {}",
1200 0 : shared.sk.flush_lsn(),
1201 : partial.flush_lsn
1202 : );
1203 0 : }
1204 :
1205 0 : if partial.commit_lsn != pstate.commit_lsn {
1206 0 : bail!(
1207 0 : "commit_lsn mismatch in partial backup, expected {}, got {}",
1208 : pstate.commit_lsn,
1209 : partial.commit_lsn
1210 : );
1211 0 : }
1212 :
1213 0 : if partial.term != shared.sk.last_log_term() {
1214 0 : bail!(
1215 0 : "term mismatch in partial backup, expected {}, got {}",
1216 0 : shared.sk.last_log_term(),
1217 : partial.term
1218 : );
1219 0 : }
1220 :
1221 0 : pstate.eviction_state = EvictionState::Offloaded(shared.sk.flush_lsn());
1222 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1223 : // control file is now switched to Offloaded state
1224 :
1225 : // now we can switch shared.sk to Offloaded, shouldn't fail
1226 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1227 0 : let cfile_state = prev_sk.take_state();
1228 0 : shared.sk = StateSK::Offloaded(Box::new(cfile_state));
1229 :
1230 0 : Ok(())
1231 0 : }
1232 :
1233 : /// Try to switch state Offloaded->Present.
1234 0 : pub(crate) async fn switch_to_present(&self) -> anyhow::Result<()> {
1235 0 : let mut shared = self.write_shared_state().await;
1236 :
1237 : // trying to restore WAL storage
1238 0 : let wal_store = wal_storage::PhysicalStorage::new(
1239 0 : &self.ttid,
1240 0 : &self.timeline_dir,
1241 0 : shared.sk.state(),
1242 0 : self.conf.no_sync,
1243 0 : )?;
1244 :
1245 : // updating control file
1246 0 : let mut pstate = shared.sk.state_mut().start_change();
1247 :
1248 0 : if !matches!(pstate.eviction_state, EvictionState::Offloaded(_)) {
1249 0 : bail!(
1250 0 : "cannot switch to present state, current state is {:?}",
1251 : pstate.eviction_state
1252 : );
1253 0 : }
1254 :
1255 0 : if wal_store.flush_lsn() != shared.sk.flush_lsn() {
1256 0 : bail!(
1257 0 : "flush_lsn mismatch in restored WAL, expected {}, got {}",
1258 0 : shared.sk.flush_lsn(),
1259 0 : wal_store.flush_lsn()
1260 : );
1261 0 : }
1262 :
1263 0 : pstate.eviction_state = EvictionState::Present;
1264 0 : shared.sk.state_mut().finish_change(&pstate).await?;
1265 :
1266 : // now we can switch shared.sk to Present, shouldn't fail
1267 0 : let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
1268 0 : let cfile_state = prev_sk.take_state();
1269 0 : shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, self.conf.my_id)?);
1270 :
1271 0 : Ok(())
1272 0 : }
1273 :
1274 : /// Update current manager state, useful for debugging manager deadlocks.
1275 208 : pub(crate) fn set_status(&self, status: timeline_manager::Status) {
1276 208 : self.mgr_status.store(status, Ordering::Relaxed);
1277 208 : }
1278 : }
1279 :
1280 : /// Deletes directory and it's contents. Returns false if directory does not exist.
1281 0 : pub async fn delete_dir(path: &Utf8PathBuf) -> Result<bool> {
1282 0 : match fs::remove_dir_all(path).await {
1283 0 : Ok(_) => Ok(true),
1284 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(false),
1285 0 : Err(e) => Err(e.into()),
1286 : }
1287 0 : }
1288 :
1289 : /// Get a path to the tenant directory. If you just need to get a timeline directory,
1290 : /// use WalResidentTimeline::get_timeline_dir instead.
1291 10 : pub fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
1292 10 : conf.workdir.join(tenant_id.to_string())
1293 10 : }
1294 :
1295 : /// Get a path to the timeline directory. If you need to read WAL files from disk,
1296 : /// use WalResidentTimeline::get_timeline_dir instead. This function does not check
1297 : /// timeline eviction status and WAL files might not be present on disk.
1298 10 : pub fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
1299 10 : get_tenant_dir(conf, &ttid.tenant_id).join(ttid.timeline_id.to_string())
1300 10 : }
|
