Line data Source code
1 : //! This module implements the streaming side of replication protocol, starting
2 : //! with the "START_REPLICATION" message, and registry of walsenders.
3 :
4 : use crate::handler::SafekeeperPostgresHandler;
5 : use crate::safekeeper::{Term, TermLsn};
6 : use crate::timeline::Timeline;
7 : use crate::wal_service::ConnectionId;
8 : use crate::wal_storage::WalReader;
9 : use crate::GlobalTimelines;
10 : use anyhow::{bail, Context as AnyhowContext};
11 : use bytes::Bytes;
12 : use parking_lot::Mutex;
13 : use postgres_backend::PostgresBackend;
14 : use postgres_backend::{CopyStreamHandlerEnd, PostgresBackendReader, QueryError};
15 : use postgres_ffi::get_current_timestamp;
16 : use postgres_ffi::{TimestampTz, MAX_SEND_SIZE};
17 : use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody};
18 : use serde::{Deserialize, Serialize};
19 : use serde_with::{serde_as, DisplayFromStr};
20 : use tokio::io::{AsyncRead, AsyncWrite};
21 : use utils::id::TenantTimelineId;
22 : use utils::lsn::AtomicLsn;
23 : use utils::pageserver_feedback::PageserverFeedback;
24 :
25 : use std::cmp::{max, min};
26 : use std::net::SocketAddr;
27 : use std::str;
28 : use std::sync::Arc;
29 : use std::time::Duration;
30 : use tokio::sync::watch::Receiver;
31 : use tokio::time::timeout;
32 : use tracing::*;
33 : use utils::{bin_ser::BeSer, lsn::Lsn};
34 :
35 : // See: https://www.postgresql.org/docs/13/protocol-replication.html
36 : const HOT_STANDBY_FEEDBACK_TAG_BYTE: u8 = b'h';
37 : const STANDBY_STATUS_UPDATE_TAG_BYTE: u8 = b'r';
38 : // neon extension of replication protocol
39 : const NEON_STATUS_UPDATE_TAG_BYTE: u8 = b'z';
40 :
41 : type FullTransactionId = u64;
42 :
43 : /// Hot standby feedback received from replica
44 3 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
45 : pub struct HotStandbyFeedback {
46 : pub ts: TimestampTz,
47 : pub xmin: FullTransactionId,
48 : pub catalog_xmin: FullTransactionId,
49 : }
50 :
51 : const INVALID_FULL_TRANSACTION_ID: FullTransactionId = 0;
52 :
53 : impl HotStandbyFeedback {
54 2201429 : pub fn empty() -> HotStandbyFeedback {
55 2201429 : HotStandbyFeedback {
56 2201429 : ts: 0,
57 2201429 : xmin: 0,
58 2201429 : catalog_xmin: 0,
59 2201429 : }
60 2201429 : }
61 : }
62 :
63 : /// Standby status update
64 12 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
65 : pub struct StandbyReply {
66 : pub write_lsn: Lsn, // The location of the last WAL byte + 1 received and written to disk in the standby.
67 : pub flush_lsn: Lsn, // The location of the last WAL byte + 1 flushed to disk in the standby.
68 : pub apply_lsn: Lsn, // The location of the last WAL byte + 1 applied in the standby.
69 : pub reply_ts: TimestampTz, // The client's system clock at the time of transmission, as microseconds since midnight on 2000-01-01.
70 : pub reply_requested: bool,
71 : }
72 :
73 : impl StandbyReply {
74 4 : fn empty() -> Self {
75 4 : StandbyReply {
76 4 : write_lsn: Lsn::INVALID,
77 4 : flush_lsn: Lsn::INVALID,
78 4 : apply_lsn: Lsn::INVALID,
79 4 : reply_ts: 0,
80 4 : reply_requested: false,
81 4 : }
82 4 : }
83 : }
84 :
85 0 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
86 : pub struct StandbyFeedback {
87 : reply: StandbyReply,
88 : hs_feedback: HotStandbyFeedback,
89 : }
90 :
91 : /// WalSenders registry. Timeline holds it (wrapped in Arc).
92 : pub struct WalSenders {
93 : /// Lsn maximized over all walsenders *and* peer data, so might be higher
94 : /// than what we receive from replicas.
95 : remote_consistent_lsn: AtomicLsn,
96 : mutex: Mutex<WalSendersShared>,
97 : }
98 :
99 : impl WalSenders {
100 604 : pub fn new(remote_consistent_lsn: Lsn) -> Arc<WalSenders> {
101 604 : Arc::new(WalSenders {
102 604 : remote_consistent_lsn: AtomicLsn::from(remote_consistent_lsn),
103 604 : mutex: Mutex::new(WalSendersShared::new()),
104 604 : })
105 604 : }
106 :
107 : /// Register new walsender. Returned guard provides access to the slot and
108 : /// automatically deregisters in Drop.
109 830 : fn register(
110 830 : self: &Arc<WalSenders>,
111 830 : ttid: TenantTimelineId,
112 830 : addr: SocketAddr,
113 830 : conn_id: ConnectionId,
114 830 : appname: Option<String>,
115 830 : ) -> WalSenderGuard {
116 830 : let slots = &mut self.mutex.lock().slots;
117 830 : let walsender_state = WalSenderState {
118 830 : ttid,
119 830 : addr,
120 830 : conn_id,
121 830 : appname,
122 830 : feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
123 830 : };
124 : // find empty slot or create new one
125 830 : let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
126 279 : slots[pos] = Some(walsender_state);
127 279 : pos
128 : } else {
129 551 : let pos = slots.len();
130 551 : slots.push(Some(walsender_state));
131 551 : pos
132 : };
133 830 : WalSenderGuard {
134 830 : id: pos,
135 830 : walsenders: self.clone(),
136 830 : }
137 830 : }
138 :
139 : /// Get state of all walsenders.
140 204 : pub fn get_all(self: &Arc<WalSenders>) -> Vec<WalSenderState> {
141 204 : self.mutex.lock().slots.iter().flatten().cloned().collect()
142 204 : }
143 :
144 : /// Get aggregated pageserver feedback.
145 51 : pub fn get_ps_feedback(self: &Arc<WalSenders>) -> PageserverFeedback {
146 51 : self.mutex.lock().agg_ps_feedback
147 51 : }
148 :
149 : /// Get aggregated pageserver and hot standby feedback (we send them to compute).
150 2200283 : pub fn get_feedbacks(self: &Arc<WalSenders>) -> (PageserverFeedback, HotStandbyFeedback) {
151 2200283 : let shared = self.mutex.lock();
152 2200283 : (shared.agg_ps_feedback, shared.agg_hs_feedback)
153 2200283 : }
154 :
155 : /// Record new pageserver feedback, update aggregated values.
156 736446 : fn record_ps_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &PageserverFeedback) {
157 736446 : let mut shared = self.mutex.lock();
158 736446 : shared.get_slot_mut(id).feedback = ReplicationFeedback::Pageserver(*feedback);
159 736446 : shared.update_ps_feedback();
160 736446 : self.update_remote_consistent_lsn(shared.agg_ps_feedback.remote_consistent_lsn);
161 736446 : }
162 :
163 : /// Record standby reply.
164 12 : fn record_standby_reply(self: &Arc<WalSenders>, id: WalSenderId, reply: &StandbyReply) {
165 12 : let mut shared = self.mutex.lock();
166 12 : let slot = shared.get_slot_mut(id);
167 12 : match &mut slot.feedback {
168 11 : ReplicationFeedback::Standby(sf) => sf.reply = *reply,
169 : ReplicationFeedback::Pageserver(_) => {
170 1 : slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
171 1 : reply: *reply,
172 1 : hs_feedback: HotStandbyFeedback::empty(),
173 1 : })
174 : }
175 : }
176 12 : }
177 :
178 : /// Record hot standby feedback, update aggregated value.
179 1 : fn record_hs_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &HotStandbyFeedback) {
180 1 : let mut shared = self.mutex.lock();
181 1 : let slot = shared.get_slot_mut(id);
182 1 : match &mut slot.feedback {
183 1 : ReplicationFeedback::Standby(sf) => sf.hs_feedback = *feedback,
184 : ReplicationFeedback::Pageserver(_) => {
185 0 : slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
186 0 : reply: StandbyReply::empty(),
187 0 : hs_feedback: *feedback,
188 0 : })
189 : }
190 : }
191 1 : shared.update_hs_feedback();
192 1 : }
193 :
194 : /// Get remote_consistent_lsn reported by the pageserver. Returns None if
195 : /// client is not pageserver.
196 1941 : fn get_ws_remote_consistent_lsn(self: &Arc<WalSenders>, id: WalSenderId) -> Option<Lsn> {
197 1941 : let shared = self.mutex.lock();
198 1941 : let slot = shared.get_slot(id);
199 1941 : match slot.feedback {
200 1941 : ReplicationFeedback::Pageserver(feedback) => Some(feedback.remote_consistent_lsn),
201 0 : _ => None,
202 : }
203 1941 : }
204 :
205 : /// Get remote_consistent_lsn maximized across all walsenders and peers.
206 21682 : pub fn get_remote_consistent_lsn(self: &Arc<WalSenders>) -> Lsn {
207 21682 : self.remote_consistent_lsn.load()
208 21682 : }
209 :
210 : /// Update maximized remote_consistent_lsn, return new (potentially) value.
211 745523 : pub fn update_remote_consistent_lsn(self: &Arc<WalSenders>, candidate: Lsn) -> Lsn {
212 745523 : self.remote_consistent_lsn
213 745523 : .fetch_max(candidate)
214 745523 : .max(candidate)
215 745523 : }
216 :
217 : /// Unregister walsender.
218 533 : fn unregister(self: &Arc<WalSenders>, id: WalSenderId) {
219 533 : let mut shared = self.mutex.lock();
220 533 : shared.slots[id] = None;
221 533 : shared.update_hs_feedback();
222 533 : }
223 : }
224 :
225 : struct WalSendersShared {
226 : // aggregated over all walsenders value
227 : agg_hs_feedback: HotStandbyFeedback,
228 : // aggregated over all walsenders value
229 : agg_ps_feedback: PageserverFeedback,
230 : slots: Vec<Option<WalSenderState>>,
231 : }
232 :
233 : impl WalSendersShared {
234 607 : fn new() -> Self {
235 607 : WalSendersShared {
236 607 : agg_hs_feedback: HotStandbyFeedback::empty(),
237 607 : agg_ps_feedback: PageserverFeedback::empty(),
238 607 : slots: Vec::new(),
239 607 : }
240 607 : }
241 :
242 : /// Get content of provided id slot, it must exist.
243 1941 : fn get_slot(&self, id: WalSenderId) -> &WalSenderState {
244 1941 : self.slots[id].as_ref().expect("walsender doesn't exist")
245 1941 : }
246 :
247 : /// Get mut content of provided id slot, it must exist.
248 736459 : fn get_slot_mut(&mut self, id: WalSenderId) -> &mut WalSenderState {
249 736459 : self.slots[id].as_mut().expect("walsender doesn't exist")
250 736459 : }
251 :
252 : /// Update aggregated hot standy feedback. We just take min of valid xmins
253 : /// and ts.
254 536 : fn update_hs_feedback(&mut self) {
255 536 : let mut agg = HotStandbyFeedback::empty();
256 536 : for ws_state in self.slots.iter().flatten() {
257 61 : if let ReplicationFeedback::Standby(standby_feedback) = ws_state.feedback {
258 5 : let hs_feedback = standby_feedback.hs_feedback;
259 5 : // doing Option math like op1.iter().chain(op2.iter()).min()
260 5 : // would be nicer, but we serialize/deserialize this struct
261 5 : // directly, so leave as is for now
262 5 : if hs_feedback.xmin != INVALID_FULL_TRANSACTION_ID {
263 2 : if agg.xmin != INVALID_FULL_TRANSACTION_ID {
264 1 : agg.xmin = min(agg.xmin, hs_feedback.xmin);
265 1 : } else {
266 1 : agg.xmin = hs_feedback.xmin;
267 1 : }
268 2 : agg.ts = min(agg.ts, hs_feedback.ts);
269 3 : }
270 5 : if hs_feedback.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
271 0 : if agg.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
272 0 : agg.catalog_xmin = min(agg.catalog_xmin, hs_feedback.catalog_xmin);
273 0 : } else {
274 0 : agg.catalog_xmin = hs_feedback.catalog_xmin;
275 0 : }
276 0 : agg.ts = min(agg.ts, hs_feedback.ts);
277 5 : }
278 56 : }
279 : }
280 536 : self.agg_hs_feedback = agg;
281 536 : }
282 :
283 : /// Update aggregated pageserver feedback. LSNs (last_received,
284 : /// disk_consistent, remote_consistent) and reply timestamp are just
285 : /// maximized; timeline_size if taken from feedback with highest
286 : /// last_received lsn. This is generally reasonable, but we might want to
287 : /// implement other policies once multiple pageservers start to be actively
288 : /// used.
289 736447 : fn update_ps_feedback(&mut self) {
290 736447 : let init = PageserverFeedback::empty();
291 736447 : let acc =
292 736447 : self.slots
293 736447 : .iter()
294 736447 : .flatten()
295 741121 : .fold(init, |mut acc, ws_state| match ws_state.feedback {
296 741117 : ReplicationFeedback::Pageserver(feedback) => {
297 741117 : if feedback.last_received_lsn > acc.last_received_lsn {
298 737649 : acc.current_timeline_size = feedback.current_timeline_size;
299 737649 : }
300 741117 : acc.last_received_lsn =
301 741117 : max(feedback.last_received_lsn, acc.last_received_lsn);
302 741117 : acc.disk_consistent_lsn =
303 741117 : max(feedback.disk_consistent_lsn, acc.disk_consistent_lsn);
304 741117 : acc.remote_consistent_lsn =
305 741117 : max(feedback.remote_consistent_lsn, acc.remote_consistent_lsn);
306 741117 : acc.replytime = max(feedback.replytime, acc.replytime);
307 741117 : acc
308 : }
309 4 : ReplicationFeedback::Standby(_) => acc,
310 741121 : });
311 736447 : self.agg_ps_feedback = acc;
312 736447 : }
313 : }
314 :
315 : // Serialized is used only for pretty printing in json.
316 : #[serde_as]
317 86 : #[derive(Debug, Clone, Serialize, Deserialize)]
318 : pub struct WalSenderState {
319 : #[serde_as(as = "DisplayFromStr")]
320 : ttid: TenantTimelineId,
321 : addr: SocketAddr,
322 : conn_id: ConnectionId,
323 : // postgres application_name
324 : appname: Option<String>,
325 : feedback: ReplicationFeedback,
326 : }
327 :
328 : // Receiver is either pageserver or regular standby, which have different
329 : // feedbacks.
330 86 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
331 : enum ReplicationFeedback {
332 : Pageserver(PageserverFeedback),
333 : Standby(StandbyFeedback),
334 : }
335 :
336 : // id of the occupied slot in WalSenders to access it (and save in the
337 : // WalSenderGuard). We could give Arc directly to the slot, but there is not
338 : // much sense in that as values aggregation which is performed on each feedback
339 : // receival iterates over all walsenders.
340 : pub type WalSenderId = usize;
341 :
342 : /// Scope guard to access slot in WalSenders registry and unregister from it in
343 : /// Drop.
344 : pub struct WalSenderGuard {
345 : id: WalSenderId,
346 : walsenders: Arc<WalSenders>,
347 : }
348 :
349 : impl Drop for WalSenderGuard {
350 533 : fn drop(&mut self) {
351 533 : self.walsenders.unregister(self.id);
352 533 : }
353 : }
354 :
355 : impl SafekeeperPostgresHandler {
356 : /// Wrapper around handle_start_replication_guts handling result. Error is
357 : /// handled here while we're still in walsender ttid span; with API
358 : /// extension, this can probably be moved into postgres_backend.
359 830 : pub async fn handle_start_replication<IO: AsyncRead + AsyncWrite + Unpin>(
360 830 : &mut self,
361 830 : pgb: &mut PostgresBackend<IO>,
362 830 : start_pos: Lsn,
363 830 : term: Option<Term>,
364 830 : ) -> Result<(), QueryError> {
365 830 : if let Err(end) = self
366 830 : .handle_start_replication_guts(pgb, start_pos, term)
367 2769161 : .await
368 : {
369 : // Log the result and probably send it to the client, closing the stream.
370 533 : pgb.handle_copy_stream_end(end).await;
371 0 : }
372 533 : Ok(())
373 533 : }
374 :
375 830 : pub async fn handle_start_replication_guts<IO: AsyncRead + AsyncWrite + Unpin>(
376 830 : &mut self,
377 830 : pgb: &mut PostgresBackend<IO>,
378 830 : start_pos: Lsn,
379 830 : term: Option<Term>,
380 830 : ) -> Result<(), CopyStreamHandlerEnd> {
381 830 : let appname = self.appname.clone();
382 830 : let tli =
383 830 : GlobalTimelines::get(self.ttid).map_err(|e| CopyStreamHandlerEnd::Other(e.into()))?;
384 :
385 : // Use a guard object to remove our entry from the timeline when we are done.
386 830 : let ws_guard = Arc::new(tli.get_walsenders().register(
387 830 : self.ttid,
388 830 : *pgb.get_peer_addr(),
389 830 : self.conn_id,
390 830 : self.appname.clone(),
391 830 : ));
392 :
393 : // Walsender can operate in one of two modes which we select by
394 : // application_name: give only committed WAL (used by pageserver) or all
395 : // existing WAL (up to flush_lsn, used by walproposer or peer recovery).
396 : // The second case is always driven by a consensus leader which term
397 : // must generally be also supplied. However we're sloppy to do this in
398 : // walproposer recovery which will be removed soon. So TODO is to make
399 : // it not Option'al then.
400 : //
401 : // Fetching WAL without term in recovery creates a small risk of this
402 : // WAL getting concurrently garbaged if another compute rises which
403 : // collects majority and starts fixing log on this safekeeper itself.
404 : // That's ok as (old) proposer will never be able to commit such WAL.
405 830 : let end_watch = if self.is_walproposer_recovery() {
406 64 : EndWatch::Flush(tli.get_term_flush_lsn_watch_rx())
407 : } else {
408 766 : EndWatch::Commit(tli.get_commit_lsn_watch_rx())
409 : };
410 : // we don't check term here; it will be checked on first waiting/WAL reading anyway.
411 830 : let end_pos = end_watch.get();
412 830 :
413 830 : if end_pos < start_pos {
414 16 : warn!(
415 16 : "requested start_pos {} is ahead of available WAL end_pos {}",
416 16 : start_pos, end_pos
417 16 : );
418 814 : }
419 :
420 830 : info!(
421 830 : "starting streaming from {:?}, available WAL ends at {}, recovery={}",
422 830 : start_pos,
423 830 : end_pos,
424 830 : matches!(end_watch, EndWatch::Flush(_))
425 830 : );
426 :
427 : // switch to copy
428 830 : pgb.write_message(&BeMessage::CopyBothResponse).await?;
429 :
430 830 : let (_, persisted_state) = tli.get_state().await;
431 830 : let wal_reader = WalReader::new(
432 830 : self.conf.workdir.clone(),
433 830 : self.conf.timeline_dir(&tli.ttid),
434 830 : &persisted_state,
435 830 : start_pos,
436 830 : self.conf.wal_backup_enabled,
437 830 : )?;
438 :
439 : // Split to concurrently receive and send data; replies are generally
440 : // not synchronized with sends, so this avoids deadlocks.
441 830 : let reader = pgb.split().context("START_REPLICATION split")?;
442 :
443 830 : let mut sender = WalSender {
444 830 : pgb,
445 830 : tli: tli.clone(),
446 830 : appname,
447 830 : start_pos,
448 830 : end_pos,
449 830 : term,
450 830 : end_watch,
451 830 : ws_guard: ws_guard.clone(),
452 830 : wal_reader,
453 830 : send_buf: [0; MAX_SEND_SIZE],
454 830 : };
455 830 : let mut reply_reader = ReplyReader { reader, ws_guard };
456 :
457 830 : let res = tokio::select! {
458 : // todo: add read|write .context to these errors
459 140 : r = sender.run() => r,
460 393 : r = reply_reader.run() => r,
461 : };
462 : // Join pg backend back.
463 533 : pgb.unsplit(reply_reader.reader)?;
464 :
465 533 : res
466 533 : }
467 : }
468 :
469 : /// Walsender streams either up to commit_lsn (normally) or flush_lsn in the
470 : /// given term (recovery by walproposer or peer safekeeper).
471 : enum EndWatch {
472 : Commit(Receiver<Lsn>),
473 : Flush(Receiver<TermLsn>),
474 : }
475 :
476 : impl EndWatch {
477 : /// Get current end of WAL.
478 2813205 : fn get(&self) -> Lsn {
479 2813205 : match self {
480 2809241 : EndWatch::Commit(r) => *r.borrow(),
481 3964 : EndWatch::Flush(r) => r.borrow().lsn,
482 : }
483 2813205 : }
484 :
485 : /// Wait for the update.
486 1427721 : async fn changed(&mut self) -> anyhow::Result<()> {
487 1427721 : match self {
488 2007177 : EndWatch::Commit(r) => r.changed().await?,
489 128 : EndWatch::Flush(r) => r.changed().await?,
490 : }
491 1425106 : Ok(())
492 1425106 : }
493 : }
494 :
495 : /// A half driving sending WAL.
496 : struct WalSender<'a, IO> {
497 : pgb: &'a mut PostgresBackend<IO>,
498 : tli: Arc<Timeline>,
499 : appname: Option<String>,
500 : // Position since which we are sending next chunk.
501 : start_pos: Lsn,
502 : // WAL up to this position is known to be locally available.
503 : // Usually this is the same as the latest commit_lsn, but in case of
504 : // walproposer recovery, this is flush_lsn.
505 : //
506 : // We send this LSN to the receiver as wal_end, so that it knows how much
507 : // WAL this safekeeper has. This LSN should be as fresh as possible.
508 : end_pos: Lsn,
509 : /// When streaming uncommitted part, the term the client acts as the leader
510 : /// in. Streaming is stopped if local term changes to a different (higher)
511 : /// value.
512 : term: Option<Term>,
513 : /// Watch channel receiver to learn end of available WAL (and wait for its advancement).
514 : end_watch: EndWatch,
515 : ws_guard: Arc<WalSenderGuard>,
516 : wal_reader: WalReader,
517 : // buffer for readling WAL into to send it
518 : send_buf: [u8; MAX_SEND_SIZE],
519 : }
520 :
521 : impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
522 : /// Send WAL until
523 : /// - an error occurs
524 : /// - receiver is caughtup and there is no computes (if streaming up to commit_lsn)
525 : ///
526 : /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
527 : /// convenience.
528 830 : async fn run(&mut self) -> Result<(), CopyStreamHandlerEnd> {
529 : loop {
530 : // Wait for the next portion if it is not there yet, or just
531 : // update our end of WAL available for sending value, we
532 : // communicate it to the receiver.
533 2007222 : self.wait_wal().await?;
534 740439 : assert!(
535 740439 : self.end_pos > self.start_pos,
536 0 : "nothing to send after waiting for WAL"
537 : );
538 :
539 : // try to send as much as available, capped by MAX_SEND_SIZE
540 740439 : let mut send_size = self
541 740439 : .end_pos
542 740439 : .checked_sub(self.start_pos)
543 740439 : .context("reading wal without waiting for it first")?
544 : .0 as usize;
545 740439 : send_size = min(send_size, self.send_buf.len());
546 740439 : let send_buf = &mut self.send_buf[..send_size];
547 : let send_size: usize;
548 : {
549 : // If uncommitted part is being pulled, check that the term is
550 : // still the expected one.
551 740439 : let _term_guard = if let Some(t) = self.term {
552 1 : Some(self.tli.acquire_term(t).await?)
553 : } else {
554 740438 : None
555 : };
556 : // read wal into buffer
557 740588 : send_size = self.wal_reader.read(send_buf).await?
558 : };
559 740437 : let send_buf = &send_buf[..send_size];
560 740437 :
561 740437 : // and send it
562 740437 : self.pgb
563 740437 : .write_message(&BeMessage::XLogData(XLogDataBody {
564 740437 : wal_start: self.start_pos.0,
565 740437 : wal_end: self.end_pos.0,
566 740437 : timestamp: get_current_timestamp(),
567 740437 : data: send_buf,
568 740437 : }))
569 21121 : .await?;
570 :
571 0 : trace!(
572 0 : "sent {} bytes of WAL {}-{}",
573 0 : send_size,
574 0 : self.start_pos,
575 0 : self.start_pos + send_size as u64
576 0 : );
577 740417 : self.start_pos += send_size as u64;
578 : }
579 140 : }
580 :
581 : /// wait until we have WAL to stream, sending keepalives and checking for
582 : /// exit in the meanwhile
583 741246 : async fn wait_wal(&mut self) -> Result<(), CopyStreamHandlerEnd> {
584 : loop {
585 743054 : self.end_pos = self.end_watch.get();
586 743054 : if self.end_pos > self.start_pos {
587 : // We have something to send.
588 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
589 98839 : return Ok(());
590 644215 : }
591 :
592 : // Wait for WAL to appear, now self.end_pos == self.start_pos.
593 2007216 : if let Some(lsn) = wait_for_lsn(&mut self.end_watch, self.term, self.start_pos).await? {
594 641600 : self.end_pos = lsn;
595 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
596 641600 : return Ok(());
597 1941 : }
598 1941 :
599 1941 : // Timed out waiting for WAL, check for termination and send KA.
600 1941 : // Check for termination only if we are streaming up to commit_lsn
601 1941 : // (to pageserver).
602 1941 : if let EndWatch::Commit(_) = self.end_watch {
603 1941 : if let Some(remote_consistent_lsn) = self
604 1941 : .ws_guard
605 1941 : .walsenders
606 1941 : .get_ws_remote_consistent_lsn(self.ws_guard.id)
607 : {
608 1941 : if self.tli.should_walsender_stop(remote_consistent_lsn).await {
609 : // Terminate if there is nothing more to send.
610 : // Note that "ending streaming" part of the string is used by
611 : // pageserver to identify WalReceiverError::SuccessfulCompletion,
612 : // do not change this string without updating pageserver.
613 133 : return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
614 133 : "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
615 133 : self.appname, self.start_pos,
616 133 : )));
617 1808 : }
618 0 : }
619 0 : }
620 :
621 1808 : self.pgb
622 1808 : .write_message(&BeMessage::KeepAlive(WalSndKeepAlive {
623 1808 : wal_end: self.end_pos.0,
624 1808 : timestamp: get_current_timestamp(),
625 1808 : request_reply: true,
626 1808 : }))
627 0 : .await?;
628 : }
629 740572 : }
630 : }
631 :
632 : /// A half driving receiving replies.
633 : struct ReplyReader<IO> {
634 : reader: PostgresBackendReader<IO>,
635 : ws_guard: Arc<WalSenderGuard>,
636 : }
637 :
638 : impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
639 830 : async fn run(&mut self) -> Result<(), CopyStreamHandlerEnd> {
640 : loop {
641 2769038 : let msg = self.reader.read_copy_message().await?;
642 736458 : self.handle_feedback(&msg)?
643 : }
644 393 : }
645 :
646 736458 : fn handle_feedback(&mut self, msg: &Bytes) -> anyhow::Result<()> {
647 736458 : match msg.first().cloned() {
648 1 : Some(HOT_STANDBY_FEEDBACK_TAG_BYTE) => {
649 1 : // Note: deserializing is on m[1..] because we skip the tag byte.
650 1 : let hs_feedback = HotStandbyFeedback::des(&msg[1..])
651 1 : .context("failed to deserialize HotStandbyFeedback")?;
652 1 : self.ws_guard
653 1 : .walsenders
654 1 : .record_hs_feedback(self.ws_guard.id, &hs_feedback);
655 : }
656 12 : Some(STANDBY_STATUS_UPDATE_TAG_BYTE) => {
657 12 : let reply =
658 12 : StandbyReply::des(&msg[1..]).context("failed to deserialize StandbyReply")?;
659 12 : self.ws_guard
660 12 : .walsenders
661 12 : .record_standby_reply(self.ws_guard.id, &reply);
662 : }
663 : Some(NEON_STATUS_UPDATE_TAG_BYTE) => {
664 : // pageserver sends this.
665 : // Note: deserializing is on m[9..] because we skip the tag byte and len bytes.
666 736445 : let buf = Bytes::copy_from_slice(&msg[9..]);
667 736445 : let ps_feedback = PageserverFeedback::parse(buf);
668 736445 :
669 736445 : trace!("PageserverFeedback is {:?}", ps_feedback);
670 736445 : self.ws_guard
671 736445 : .walsenders
672 736445 : .record_ps_feedback(self.ws_guard.id, &ps_feedback);
673 : // in principle new remote_consistent_lsn could allow to
674 : // deactivate the timeline, but we check that regularly through
675 : // broker updated, not need to do it here
676 : }
677 0 : _ => warn!("unexpected message {:?}", msg),
678 : }
679 736458 : Ok(())
680 736458 : }
681 : }
682 :
683 : const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1);
684 :
685 : /// Wait until we have available WAL > start_pos or timeout expires. Returns
686 : /// - Ok(Some(end_pos)) if needed lsn is successfully observed;
687 : /// - Ok(None) if timeout expired;
688 : /// - Err in case of error -- only if 1) term changed while fetching in recovery
689 : /// mode 2) watch channel closed, which must never happen.
690 644215 : async fn wait_for_lsn(
691 644215 : rx: &mut EndWatch,
692 644215 : client_term: Option<Term>,
693 644215 : start_pos: Lsn,
694 644215 : ) -> anyhow::Result<Option<Lsn>> {
695 644215 : let res = timeout(POLL_STATE_TIMEOUT, async move {
696 : loop {
697 2069321 : let end_pos = rx.get();
698 2069321 : if end_pos > start_pos {
699 641600 : return Ok(end_pos);
700 1427721 : }
701 1427721 : if let EndWatch::Flush(rx) = rx {
702 128 : let curr_term = rx.borrow().term;
703 128 : if let Some(client_term) = client_term {
704 0 : if curr_term != client_term {
705 0 : bail!("term changed: requested {}, now {}", client_term, curr_term);
706 0 : }
707 128 : }
708 1427593 : }
709 2007216 : rx.changed().await?;
710 : }
711 644215 : })
712 2007216 : .await;
713 :
714 641600 : match res {
715 : // success
716 641600 : Ok(Ok(commit_lsn)) => Ok(Some(commit_lsn)),
717 : // error inside closure
718 0 : Ok(Err(err)) => Err(err),
719 : // timeout
720 1941 : Err(_) => Ok(None),
721 : }
722 643541 : }
723 :
724 : #[cfg(test)]
725 : mod tests {
726 : use postgres_protocol::PG_EPOCH;
727 : use utils::id::{TenantId, TimelineId};
728 :
729 : use super::*;
730 :
731 6 : fn mock_ttid() -> TenantTimelineId {
732 6 : TenantTimelineId {
733 6 : tenant_id: TenantId::from_slice(&[0x00; 16]).unwrap(),
734 6 : timeline_id: TimelineId::from_slice(&[0x00; 16]).unwrap(),
735 6 : }
736 6 : }
737 :
738 6 : fn mock_addr() -> SocketAddr {
739 6 : "127.0.0.1:8080".parse().unwrap()
740 6 : }
741 :
742 : // add to wss specified feedback setting other fields to dummy values
743 6 : fn push_feedback(wss: &mut WalSendersShared, feedback: ReplicationFeedback) {
744 6 : let walsender_state = WalSenderState {
745 6 : ttid: mock_ttid(),
746 6 : addr: mock_addr(),
747 6 : conn_id: 1,
748 6 : appname: None,
749 6 : feedback,
750 6 : };
751 6 : wss.slots.push(Some(walsender_state))
752 6 : }
753 :
754 : // form standby feedback with given hot standby feedback ts/xmin and the
755 : // rest set to dummy values.
756 4 : fn hs_feedback(ts: TimestampTz, xmin: FullTransactionId) -> ReplicationFeedback {
757 4 : ReplicationFeedback::Standby(StandbyFeedback {
758 4 : reply: StandbyReply::empty(),
759 4 : hs_feedback: HotStandbyFeedback {
760 4 : ts,
761 4 : xmin,
762 4 : catalog_xmin: 0,
763 4 : },
764 4 : })
765 4 : }
766 :
767 : // test that hs aggregation works as expected
768 1 : #[test]
769 1 : fn test_hs_feedback_no_valid() {
770 1 : let mut wss = WalSendersShared::new();
771 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
772 1 : wss.update_hs_feedback();
773 1 : assert_eq!(wss.agg_hs_feedback.xmin, INVALID_FULL_TRANSACTION_ID);
774 1 : }
775 :
776 1 : #[test]
777 1 : fn test_hs_feedback() {
778 1 : let mut wss = WalSendersShared::new();
779 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
780 1 : push_feedback(&mut wss, hs_feedback(1, 42));
781 1 : push_feedback(&mut wss, hs_feedback(1, 64));
782 1 : wss.update_hs_feedback();
783 1 : assert_eq!(wss.agg_hs_feedback.xmin, 42);
784 1 : }
785 :
786 : // form pageserver feedback with given last_record_lsn / tli size and the
787 : // rest set to dummy values.
788 2 : fn ps_feedback(current_timeline_size: u64, last_received_lsn: Lsn) -> ReplicationFeedback {
789 2 : ReplicationFeedback::Pageserver(PageserverFeedback {
790 2 : current_timeline_size,
791 2 : last_received_lsn,
792 2 : disk_consistent_lsn: Lsn::INVALID,
793 2 : remote_consistent_lsn: Lsn::INVALID,
794 2 : replytime: *PG_EPOCH,
795 2 : })
796 2 : }
797 :
798 : // test that ps aggregation works as expected
799 1 : #[test]
800 1 : fn test_ps_feedback() {
801 1 : let mut wss = WalSendersShared::new();
802 1 : push_feedback(&mut wss, ps_feedback(8, Lsn(42)));
803 1 : push_feedback(&mut wss, ps_feedback(4, Lsn(84)));
804 1 : wss.update_ps_feedback();
805 1 : assert_eq!(wss.agg_ps_feedback.current_timeline_size, 4);
806 1 : assert_eq!(wss.agg_ps_feedback.last_received_lsn, Lsn(84));
807 1 : }
808 : }
|
