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::metrics::RECEIVED_PS_FEEDBACKS;
6 : use crate::receive_wal::WalReceivers;
7 : use crate::safekeeper::{Term, TermLsn};
8 : use crate::send_interpreted_wal::InterpretedWalSender;
9 : use crate::timeline::WalResidentTimeline;
10 : use crate::wal_reader_stream::WalReaderStreamBuilder;
11 : use crate::wal_service::ConnectionId;
12 : use crate::wal_storage::WalReader;
13 : use crate::GlobalTimelines;
14 : use anyhow::{bail, Context as AnyhowContext};
15 : use bytes::Bytes;
16 : use futures::future::Either;
17 : use parking_lot::Mutex;
18 : use postgres_backend::PostgresBackend;
19 : use postgres_backend::{CopyStreamHandlerEnd, PostgresBackendReader, QueryError};
20 : use postgres_ffi::get_current_timestamp;
21 : use postgres_ffi::{TimestampTz, MAX_SEND_SIZE};
22 : use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody};
23 : use serde::{Deserialize, Serialize};
24 : use tokio::io::{AsyncRead, AsyncWrite};
25 : use utils::failpoint_support;
26 : use utils::id::TenantTimelineId;
27 : use utils::pageserver_feedback::PageserverFeedback;
28 : use utils::postgres_client::PostgresClientProtocol;
29 :
30 : use std::cmp::{max, min};
31 : use std::net::SocketAddr;
32 : use std::str;
33 : use std::sync::Arc;
34 : use std::time::Duration;
35 : use tokio::sync::watch::Receiver;
36 : use tokio::time::timeout;
37 : use tracing::*;
38 : use utils::{bin_ser::BeSer, lsn::Lsn};
39 :
40 : // See: https://www.postgresql.org/docs/13/protocol-replication.html
41 : const HOT_STANDBY_FEEDBACK_TAG_BYTE: u8 = b'h';
42 : const STANDBY_STATUS_UPDATE_TAG_BYTE: u8 = b'r';
43 : // neon extension of replication protocol
44 : const NEON_STATUS_UPDATE_TAG_BYTE: u8 = b'z';
45 :
46 : type FullTransactionId = u64;
47 :
48 : /// Hot standby feedback received from replica
49 0 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
50 : pub struct HotStandbyFeedback {
51 : pub ts: TimestampTz,
52 : pub xmin: FullTransactionId,
53 : pub catalog_xmin: FullTransactionId,
54 : }
55 :
56 : const INVALID_FULL_TRANSACTION_ID: FullTransactionId = 0;
57 :
58 : impl HotStandbyFeedback {
59 3266 : pub fn empty() -> HotStandbyFeedback {
60 3266 : HotStandbyFeedback {
61 3266 : ts: 0,
62 3266 : xmin: 0,
63 3266 : catalog_xmin: 0,
64 3266 : }
65 3266 : }
66 : }
67 :
68 : /// Standby status update
69 0 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
70 : pub struct StandbyReply {
71 : pub write_lsn: Lsn, // The location of the last WAL byte + 1 received and written to disk in the standby.
72 : pub flush_lsn: Lsn, // The location of the last WAL byte + 1 flushed to disk in the standby.
73 : pub apply_lsn: Lsn, // The location of the last WAL byte + 1 applied in the standby.
74 : pub reply_ts: TimestampTz, // The client's system clock at the time of transmission, as microseconds since midnight on 2000-01-01.
75 : pub reply_requested: bool,
76 : }
77 :
78 : impl StandbyReply {
79 8 : fn empty() -> Self {
80 8 : StandbyReply {
81 8 : write_lsn: Lsn::INVALID,
82 8 : flush_lsn: Lsn::INVALID,
83 8 : apply_lsn: Lsn::INVALID,
84 8 : reply_ts: 0,
85 8 : reply_requested: false,
86 8 : }
87 8 : }
88 : }
89 :
90 0 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
91 : pub struct StandbyFeedback {
92 : pub reply: StandbyReply,
93 : pub hs_feedback: HotStandbyFeedback,
94 : }
95 :
96 : impl StandbyFeedback {
97 2 : pub fn empty() -> Self {
98 2 : StandbyFeedback {
99 2 : reply: StandbyReply::empty(),
100 2 : hs_feedback: HotStandbyFeedback::empty(),
101 2 : }
102 2 : }
103 : }
104 :
105 : /// WalSenders registry. Timeline holds it (wrapped in Arc).
106 : pub struct WalSenders {
107 : mutex: Mutex<WalSendersShared>,
108 : walreceivers: Arc<WalReceivers>,
109 : }
110 :
111 : impl WalSenders {
112 0 : pub fn new(walreceivers: Arc<WalReceivers>) -> Arc<WalSenders> {
113 0 : Arc::new(WalSenders {
114 0 : mutex: Mutex::new(WalSendersShared::new()),
115 0 : walreceivers,
116 0 : })
117 0 : }
118 :
119 : /// Register new walsender. Returned guard provides access to the slot and
120 : /// automatically deregisters in Drop.
121 0 : fn register(
122 0 : self: &Arc<WalSenders>,
123 0 : ttid: TenantTimelineId,
124 0 : addr: SocketAddr,
125 0 : conn_id: ConnectionId,
126 0 : appname: Option<String>,
127 0 : ) -> WalSenderGuard {
128 0 : let slots = &mut self.mutex.lock().slots;
129 0 : let walsender_state = WalSenderState {
130 0 : ttid,
131 0 : addr,
132 0 : conn_id,
133 0 : appname,
134 0 : feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
135 0 : };
136 : // find empty slot or create new one
137 0 : let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
138 0 : slots[pos] = Some(walsender_state);
139 0 : pos
140 : } else {
141 0 : let pos = slots.len();
142 0 : slots.push(Some(walsender_state));
143 0 : pos
144 : };
145 0 : WalSenderGuard {
146 0 : id: pos,
147 0 : walsenders: self.clone(),
148 0 : }
149 0 : }
150 :
151 : /// Get state of all walsenders.
152 0 : pub fn get_all(self: &Arc<WalSenders>) -> Vec<WalSenderState> {
153 0 : self.mutex.lock().slots.iter().flatten().cloned().collect()
154 0 : }
155 :
156 : /// Get LSN of the most lagging pageserver receiver. Return None if there are no
157 : /// active walsenders.
158 0 : pub fn laggard_lsn(self: &Arc<WalSenders>) -> Option<Lsn> {
159 0 : self.mutex
160 0 : .lock()
161 0 : .slots
162 0 : .iter()
163 0 : .flatten()
164 0 : .filter_map(|s| match s.feedback {
165 0 : ReplicationFeedback::Pageserver(feedback) => Some(feedback.last_received_lsn),
166 0 : ReplicationFeedback::Standby(_) => None,
167 0 : })
168 0 : .min()
169 0 : }
170 :
171 : /// Returns total counter of pageserver feedbacks received and last feedback.
172 0 : pub fn get_ps_feedback_stats(self: &Arc<WalSenders>) -> (u64, PageserverFeedback) {
173 0 : let shared = self.mutex.lock();
174 0 : (shared.ps_feedback_counter, shared.last_ps_feedback)
175 0 : }
176 :
177 : /// Get aggregated hot standby feedback (we send it to compute).
178 0 : pub fn get_hotstandby(self: &Arc<WalSenders>) -> StandbyFeedback {
179 0 : self.mutex.lock().agg_standby_feedback
180 0 : }
181 :
182 : /// Record new pageserver feedback, update aggregated values.
183 0 : fn record_ps_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &PageserverFeedback) {
184 0 : let mut shared = self.mutex.lock();
185 0 : shared.get_slot_mut(id).feedback = ReplicationFeedback::Pageserver(*feedback);
186 0 : shared.last_ps_feedback = *feedback;
187 0 : shared.ps_feedback_counter += 1;
188 0 : drop(shared);
189 0 :
190 0 : RECEIVED_PS_FEEDBACKS.inc();
191 0 :
192 0 : // send feedback to connected walproposers
193 0 : self.walreceivers.broadcast_pageserver_feedback(*feedback);
194 0 : }
195 :
196 : /// Record standby reply.
197 0 : fn record_standby_reply(self: &Arc<WalSenders>, id: WalSenderId, reply: &StandbyReply) {
198 0 : let mut shared = self.mutex.lock();
199 0 : let slot = shared.get_slot_mut(id);
200 0 : debug!(
201 0 : "Record standby reply: ts={} apply_lsn={}",
202 : reply.reply_ts, reply.apply_lsn
203 : );
204 0 : match &mut slot.feedback {
205 0 : ReplicationFeedback::Standby(sf) => sf.reply = *reply,
206 : ReplicationFeedback::Pageserver(_) => {
207 0 : slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
208 0 : reply: *reply,
209 0 : hs_feedback: HotStandbyFeedback::empty(),
210 0 : })
211 : }
212 : }
213 0 : }
214 :
215 : /// Record hot standby feedback, update aggregated value.
216 0 : fn record_hs_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &HotStandbyFeedback) {
217 0 : let mut shared = self.mutex.lock();
218 0 : let slot = shared.get_slot_mut(id);
219 0 : match &mut slot.feedback {
220 0 : ReplicationFeedback::Standby(sf) => sf.hs_feedback = *feedback,
221 : ReplicationFeedback::Pageserver(_) => {
222 0 : slot.feedback = ReplicationFeedback::Standby(StandbyFeedback {
223 0 : reply: StandbyReply::empty(),
224 0 : hs_feedback: *feedback,
225 0 : })
226 : }
227 : }
228 0 : shared.update_reply_feedback();
229 0 : }
230 :
231 : /// Get remote_consistent_lsn reported by the pageserver. Returns None if
232 : /// client is not pageserver.
233 0 : pub fn get_ws_remote_consistent_lsn(self: &Arc<WalSenders>, id: WalSenderId) -> Option<Lsn> {
234 0 : let shared = self.mutex.lock();
235 0 : let slot = shared.get_slot(id);
236 0 : match slot.feedback {
237 0 : ReplicationFeedback::Pageserver(feedback) => Some(feedback.remote_consistent_lsn),
238 0 : _ => None,
239 : }
240 0 : }
241 :
242 : /// Unregister walsender.
243 0 : fn unregister(self: &Arc<WalSenders>, id: WalSenderId) {
244 0 : let mut shared = self.mutex.lock();
245 0 : shared.slots[id] = None;
246 0 : shared.update_reply_feedback();
247 0 : }
248 : }
249 :
250 : struct WalSendersShared {
251 : // aggregated over all walsenders value
252 : agg_standby_feedback: StandbyFeedback,
253 : // last feedback ever received from any pageserver, empty if none
254 : last_ps_feedback: PageserverFeedback,
255 : // total counter of pageserver feedbacks received
256 : ps_feedback_counter: u64,
257 : slots: Vec<Option<WalSenderState>>,
258 : }
259 :
260 : impl WalSendersShared {
261 2 : fn new() -> Self {
262 2 : WalSendersShared {
263 2 : agg_standby_feedback: StandbyFeedback::empty(),
264 2 : last_ps_feedback: PageserverFeedback::empty(),
265 2 : ps_feedback_counter: 0,
266 2 : slots: Vec::new(),
267 2 : }
268 2 : }
269 :
270 : /// Get content of provided id slot, it must exist.
271 0 : fn get_slot(&self, id: WalSenderId) -> &WalSenderState {
272 0 : self.slots[id].as_ref().expect("walsender doesn't exist")
273 0 : }
274 :
275 : /// Get mut content of provided id slot, it must exist.
276 0 : fn get_slot_mut(&mut self, id: WalSenderId) -> &mut WalSenderState {
277 0 : self.slots[id].as_mut().expect("walsender doesn't exist")
278 0 : }
279 :
280 : /// Update aggregated hot standy and normal reply feedbacks. We just take min of valid xmins
281 : /// and ts.
282 2 : fn update_reply_feedback(&mut self) {
283 2 : let mut agg = HotStandbyFeedback::empty();
284 2 : let mut reply_agg = StandbyReply::empty();
285 4 : for ws_state in self.slots.iter().flatten() {
286 4 : if let ReplicationFeedback::Standby(standby_feedback) = ws_state.feedback {
287 4 : let hs_feedback = standby_feedback.hs_feedback;
288 4 : // doing Option math like op1.iter().chain(op2.iter()).min()
289 4 : // would be nicer, but we serialize/deserialize this struct
290 4 : // directly, so leave as is for now
291 4 : if hs_feedback.xmin != INVALID_FULL_TRANSACTION_ID {
292 2 : if agg.xmin != INVALID_FULL_TRANSACTION_ID {
293 1 : agg.xmin = min(agg.xmin, hs_feedback.xmin);
294 1 : } else {
295 1 : agg.xmin = hs_feedback.xmin;
296 1 : }
297 2 : agg.ts = max(agg.ts, hs_feedback.ts);
298 2 : }
299 4 : if hs_feedback.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
300 0 : if agg.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
301 0 : agg.catalog_xmin = min(agg.catalog_xmin, hs_feedback.catalog_xmin);
302 0 : } else {
303 0 : agg.catalog_xmin = hs_feedback.catalog_xmin;
304 0 : }
305 0 : agg.ts = max(agg.ts, hs_feedback.ts);
306 4 : }
307 4 : let reply = standby_feedback.reply;
308 4 : if reply.write_lsn != Lsn::INVALID {
309 0 : if reply_agg.write_lsn != Lsn::INVALID {
310 0 : reply_agg.write_lsn = Lsn::min(reply_agg.write_lsn, reply.write_lsn);
311 0 : } else {
312 0 : reply_agg.write_lsn = reply.write_lsn;
313 0 : }
314 4 : }
315 4 : if reply.flush_lsn != Lsn::INVALID {
316 0 : if reply_agg.flush_lsn != Lsn::INVALID {
317 0 : reply_agg.flush_lsn = Lsn::min(reply_agg.flush_lsn, reply.flush_lsn);
318 0 : } else {
319 0 : reply_agg.flush_lsn = reply.flush_lsn;
320 0 : }
321 4 : }
322 4 : if reply.apply_lsn != Lsn::INVALID {
323 0 : if reply_agg.apply_lsn != Lsn::INVALID {
324 0 : reply_agg.apply_lsn = Lsn::min(reply_agg.apply_lsn, reply.apply_lsn);
325 0 : } else {
326 0 : reply_agg.apply_lsn = reply.apply_lsn;
327 0 : }
328 4 : }
329 4 : if reply.reply_ts != 0 {
330 0 : if reply_agg.reply_ts != 0 {
331 0 : reply_agg.reply_ts = TimestampTz::min(reply_agg.reply_ts, reply.reply_ts);
332 0 : } else {
333 0 : reply_agg.reply_ts = reply.reply_ts;
334 0 : }
335 4 : }
336 0 : }
337 : }
338 2 : self.agg_standby_feedback = StandbyFeedback {
339 2 : reply: reply_agg,
340 2 : hs_feedback: agg,
341 2 : };
342 2 : }
343 : }
344 :
345 : // Serialized is used only for pretty printing in json.
346 0 : #[derive(Debug, Clone, Serialize, Deserialize)]
347 : pub struct WalSenderState {
348 : ttid: TenantTimelineId,
349 : addr: SocketAddr,
350 : conn_id: ConnectionId,
351 : // postgres application_name
352 : appname: Option<String>,
353 : feedback: ReplicationFeedback,
354 : }
355 :
356 : // Receiver is either pageserver or regular standby, which have different
357 : // feedbacks.
358 0 : #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
359 : enum ReplicationFeedback {
360 : Pageserver(PageserverFeedback),
361 : Standby(StandbyFeedback),
362 : }
363 :
364 : // id of the occupied slot in WalSenders to access it (and save in the
365 : // WalSenderGuard). We could give Arc directly to the slot, but there is not
366 : // much sense in that as values aggregation which is performed on each feedback
367 : // receival iterates over all walsenders.
368 : pub type WalSenderId = usize;
369 :
370 : /// Scope guard to access slot in WalSenders registry and unregister from it in
371 : /// Drop.
372 : pub struct WalSenderGuard {
373 : id: WalSenderId,
374 : walsenders: Arc<WalSenders>,
375 : }
376 :
377 : impl WalSenderGuard {
378 0 : pub fn id(&self) -> WalSenderId {
379 0 : self.id
380 0 : }
381 :
382 0 : pub fn walsenders(&self) -> &Arc<WalSenders> {
383 0 : &self.walsenders
384 0 : }
385 : }
386 :
387 : impl Drop for WalSenderGuard {
388 0 : fn drop(&mut self) {
389 0 : self.walsenders.unregister(self.id);
390 0 : }
391 : }
392 :
393 : impl SafekeeperPostgresHandler {
394 : /// Wrapper around handle_start_replication_guts handling result. Error is
395 : /// handled here while we're still in walsender ttid span; with API
396 : /// extension, this can probably be moved into postgres_backend.
397 0 : pub async fn handle_start_replication<IO: AsyncRead + AsyncWrite + Unpin>(
398 0 : &mut self,
399 0 : pgb: &mut PostgresBackend<IO>,
400 0 : start_pos: Lsn,
401 0 : term: Option<Term>,
402 0 : ) -> Result<(), QueryError> {
403 0 : let tli = GlobalTimelines::get(self.ttid).map_err(|e| QueryError::Other(e.into()))?;
404 0 : let residence_guard = tli.wal_residence_guard().await?;
405 :
406 0 : if let Err(end) = self
407 0 : .handle_start_replication_guts(pgb, start_pos, term, residence_guard)
408 0 : .await
409 : {
410 0 : let info = tli.get_safekeeper_info(&self.conf).await;
411 : // Log the result and probably send it to the client, closing the stream.
412 0 : pgb.handle_copy_stream_end(end)
413 0 : .instrument(info_span!("", term=%info.term, last_log_term=%info.last_log_term, flush_lsn=%Lsn(info.flush_lsn), commit_lsn=%Lsn(info.flush_lsn)))
414 0 : .await;
415 0 : }
416 0 : Ok(())
417 0 : }
418 :
419 0 : pub async fn handle_start_replication_guts<IO: AsyncRead + AsyncWrite + Unpin>(
420 0 : &mut self,
421 0 : pgb: &mut PostgresBackend<IO>,
422 0 : start_pos: Lsn,
423 0 : term: Option<Term>,
424 0 : tli: WalResidentTimeline,
425 0 : ) -> Result<(), CopyStreamHandlerEnd> {
426 0 : let appname = self.appname.clone();
427 0 :
428 0 : // Use a guard object to remove our entry from the timeline when we are done.
429 0 : let ws_guard = Arc::new(tli.get_walsenders().register(
430 0 : self.ttid,
431 0 : *pgb.get_peer_addr(),
432 0 : self.conn_id,
433 0 : self.appname.clone(),
434 0 : ));
435 :
436 : // Walsender can operate in one of two modes which we select by
437 : // application_name: give only committed WAL (used by pageserver) or all
438 : // existing WAL (up to flush_lsn, used by walproposer or peer recovery).
439 : // The second case is always driven by a consensus leader which term
440 : // must be supplied.
441 0 : let end_watch = if term.is_some() {
442 0 : EndWatch::Flush(tli.get_term_flush_lsn_watch_rx())
443 : } else {
444 0 : EndWatch::Commit(tli.get_commit_lsn_watch_rx())
445 : };
446 : // we don't check term here; it will be checked on first waiting/WAL reading anyway.
447 0 : let end_pos = end_watch.get();
448 0 :
449 0 : if end_pos < start_pos {
450 0 : warn!(
451 0 : "requested start_pos {} is ahead of available WAL end_pos {}",
452 : start_pos, end_pos
453 : );
454 0 : }
455 :
456 0 : info!(
457 0 : "starting streaming from {:?}, available WAL ends at {}, recovery={}, appname={:?}, protocol={:?}",
458 : start_pos,
459 : end_pos,
460 0 : matches!(end_watch, EndWatch::Flush(_)),
461 : appname,
462 0 : self.protocol(),
463 : );
464 :
465 : // switch to copy
466 0 : pgb.write_message(&BeMessage::CopyBothResponse).await?;
467 :
468 0 : let wal_reader = tli.get_walreader(start_pos).await?;
469 :
470 : // Split to concurrently receive and send data; replies are generally
471 : // not synchronized with sends, so this avoids deadlocks.
472 0 : let reader = pgb.split().context("START_REPLICATION split")?;
473 :
474 0 : let send_fut = match self.protocol() {
475 : PostgresClientProtocol::Vanilla => {
476 0 : let sender = WalSender {
477 0 : pgb,
478 0 : // should succeed since we're already holding another guard
479 0 : tli: tli.wal_residence_guard().await?,
480 0 : appname,
481 0 : start_pos,
482 0 : end_pos,
483 0 : term,
484 0 : end_watch,
485 0 : ws_guard: ws_guard.clone(),
486 0 : wal_reader,
487 0 : send_buf: vec![0u8; MAX_SEND_SIZE],
488 0 : };
489 0 :
490 0 : Either::Left(sender.run())
491 : }
492 : PostgresClientProtocol::Interpreted {
493 0 : format,
494 0 : compression,
495 : } => {
496 0 : let pg_version = tli.tli.get_state().await.1.server.pg_version / 10000;
497 0 : let end_watch_view = end_watch.view();
498 0 : let wal_stream_builder = WalReaderStreamBuilder {
499 0 : tli: tli.wal_residence_guard().await?,
500 0 : start_pos,
501 0 : end_pos,
502 0 : term,
503 0 : end_watch,
504 0 : wal_sender_guard: ws_guard.clone(),
505 0 : };
506 0 :
507 0 : let sender = InterpretedWalSender {
508 0 : format,
509 0 : compression,
510 0 : pgb,
511 0 : wal_stream_builder,
512 0 : end_watch_view,
513 0 : shard: self.shard.unwrap(),
514 0 : pg_version,
515 0 : appname,
516 0 : };
517 0 :
518 0 : Either::Right(sender.run())
519 : }
520 : };
521 :
522 0 : let tli_cancel = tli.cancel.clone();
523 0 :
524 0 : let mut reply_reader = ReplyReader {
525 0 : reader,
526 0 : ws_guard: ws_guard.clone(),
527 0 : tli,
528 0 : };
529 :
530 0 : let res = tokio::select! {
531 : // todo: add read|write .context to these errors
532 0 : r = send_fut => r,
533 0 : r = reply_reader.run() => r,
534 0 : _ = tli_cancel.cancelled() => {
535 0 : return Err(CopyStreamHandlerEnd::Cancelled);
536 : }
537 : };
538 :
539 0 : let ws_state = ws_guard
540 0 : .walsenders
541 0 : .mutex
542 0 : .lock()
543 0 : .get_slot(ws_guard.id)
544 0 : .clone();
545 0 : info!(
546 0 : "finished streaming to {}, feedback={:?}",
547 : ws_state.addr, ws_state.feedback,
548 : );
549 :
550 : // Join pg backend back.
551 0 : pgb.unsplit(reply_reader.reader)?;
552 :
553 0 : res
554 0 : }
555 : }
556 :
557 : /// TODO(vlad): maybe lift this instead
558 : /// Walsender streams either up to commit_lsn (normally) or flush_lsn in the
559 : /// given term (recovery by walproposer or peer safekeeper).
560 : #[derive(Clone)]
561 : pub(crate) enum EndWatch {
562 : Commit(Receiver<Lsn>),
563 : Flush(Receiver<TermLsn>),
564 : }
565 :
566 : impl EndWatch {
567 0 : pub(crate) fn view(&self) -> EndWatchView {
568 0 : EndWatchView(self.clone())
569 0 : }
570 :
571 : /// Get current end of WAL.
572 0 : pub(crate) fn get(&self) -> Lsn {
573 0 : match self {
574 0 : EndWatch::Commit(r) => *r.borrow(),
575 0 : EndWatch::Flush(r) => r.borrow().lsn,
576 : }
577 0 : }
578 :
579 : /// Wait for the update.
580 0 : pub(crate) async fn changed(&mut self) -> anyhow::Result<()> {
581 0 : match self {
582 0 : EndWatch::Commit(r) => r.changed().await?,
583 0 : EndWatch::Flush(r) => r.changed().await?,
584 : }
585 0 : Ok(())
586 0 : }
587 :
588 0 : pub(crate) async fn wait_for_lsn(
589 0 : &mut self,
590 0 : lsn: Lsn,
591 0 : client_term: Option<Term>,
592 0 : ) -> anyhow::Result<Lsn> {
593 : loop {
594 0 : let end_pos = self.get();
595 0 : if end_pos > lsn {
596 0 : return Ok(end_pos);
597 0 : }
598 0 : if let EndWatch::Flush(rx) = &self {
599 0 : let curr_term = rx.borrow().term;
600 0 : if let Some(client_term) = client_term {
601 0 : if curr_term != client_term {
602 0 : bail!("term changed: requested {}, now {}", client_term, curr_term);
603 0 : }
604 0 : }
605 0 : }
606 0 : self.changed().await?;
607 : }
608 0 : }
609 : }
610 :
611 : pub(crate) struct EndWatchView(EndWatch);
612 :
613 : impl EndWatchView {
614 0 : pub(crate) fn get(&self) -> Lsn {
615 0 : self.0.get()
616 0 : }
617 : }
618 : /// A half driving sending WAL.
619 : struct WalSender<'a, IO> {
620 : pgb: &'a mut PostgresBackend<IO>,
621 : tli: WalResidentTimeline,
622 : appname: Option<String>,
623 : // Position since which we are sending next chunk.
624 : start_pos: Lsn,
625 : // WAL up to this position is known to be locally available.
626 : // Usually this is the same as the latest commit_lsn, but in case of
627 : // walproposer recovery, this is flush_lsn.
628 : //
629 : // We send this LSN to the receiver as wal_end, so that it knows how much
630 : // WAL this safekeeper has. This LSN should be as fresh as possible.
631 : end_pos: Lsn,
632 : /// When streaming uncommitted part, the term the client acts as the leader
633 : /// in. Streaming is stopped if local term changes to a different (higher)
634 : /// value.
635 : term: Option<Term>,
636 : /// Watch channel receiver to learn end of available WAL (and wait for its advancement).
637 : end_watch: EndWatch,
638 : ws_guard: Arc<WalSenderGuard>,
639 : wal_reader: WalReader,
640 : // buffer for readling WAL into to send it
641 : send_buf: Vec<u8>,
642 : }
643 :
644 : const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1);
645 :
646 : impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
647 : /// Send WAL until
648 : /// - an error occurs
649 : /// - receiver is caughtup and there is no computes (if streaming up to commit_lsn)
650 : /// - timeline's cancellation token fires
651 : ///
652 : /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
653 : /// convenience.
654 0 : async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> {
655 : loop {
656 : // Wait for the next portion if it is not there yet, or just
657 : // update our end of WAL available for sending value, we
658 : // communicate it to the receiver.
659 0 : self.wait_wal().await?;
660 0 : assert!(
661 0 : self.end_pos > self.start_pos,
662 0 : "nothing to send after waiting for WAL"
663 : );
664 :
665 : // try to send as much as available, capped by MAX_SEND_SIZE
666 0 : let mut chunk_end_pos = self.start_pos + MAX_SEND_SIZE as u64;
667 0 : // if we went behind available WAL, back off
668 0 : if chunk_end_pos >= self.end_pos {
669 0 : chunk_end_pos = self.end_pos;
670 0 : } else {
671 0 : // If sending not up to end pos, round down to page boundary to
672 0 : // avoid breaking WAL record not at page boundary, as protocol
673 0 : // demands. See walsender.c (XLogSendPhysical).
674 0 : chunk_end_pos = chunk_end_pos
675 0 : .checked_sub(chunk_end_pos.block_offset())
676 0 : .unwrap();
677 0 : }
678 0 : let send_size = (chunk_end_pos.0 - self.start_pos.0) as usize;
679 0 : let send_buf = &mut self.send_buf[..send_size];
680 : let send_size: usize;
681 : {
682 : // If uncommitted part is being pulled, check that the term is
683 : // still the expected one.
684 0 : let _term_guard = if let Some(t) = self.term {
685 0 : Some(self.tli.acquire_term(t).await?)
686 : } else {
687 0 : None
688 : };
689 : // Read WAL into buffer. send_size can be additionally capped to
690 : // segment boundary here.
691 0 : send_size = self.wal_reader.read(send_buf).await?
692 : };
693 0 : let send_buf = &send_buf[..send_size];
694 0 :
695 0 : // and send it, while respecting Timeline::cancel
696 0 : let msg = BeMessage::XLogData(XLogDataBody {
697 0 : wal_start: self.start_pos.0,
698 0 : wal_end: self.end_pos.0,
699 0 : timestamp: get_current_timestamp(),
700 0 : data: send_buf,
701 0 : });
702 0 : self.pgb.write_message(&msg).await?;
703 :
704 0 : if let Some(appname) = &self.appname {
705 0 : if appname == "replica" {
706 0 : failpoint_support::sleep_millis_async!("sk-send-wal-replica-sleep");
707 0 : }
708 0 : }
709 0 : trace!(
710 0 : "sent {} bytes of WAL {}-{}",
711 0 : send_size,
712 0 : self.start_pos,
713 0 : self.start_pos + send_size as u64
714 : );
715 0 : self.start_pos += send_size as u64;
716 : }
717 0 : }
718 :
719 : /// wait until we have WAL to stream, sending keepalives and checking for
720 : /// exit in the meanwhile
721 0 : async fn wait_wal(&mut self) -> Result<(), CopyStreamHandlerEnd> {
722 : loop {
723 0 : self.end_pos = self.end_watch.get();
724 0 : let have_something_to_send = (|| {
725 0 : fail::fail_point!(
726 0 : "sk-pause-send",
727 0 : self.appname.as_deref() != Some("pageserver"),
728 0 : |_| { false }
729 0 : );
730 0 : self.end_pos > self.start_pos
731 0 : })();
732 0 :
733 0 : if have_something_to_send {
734 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
735 0 : return Ok(());
736 0 : }
737 :
738 : // Wait for WAL to appear, now self.end_pos == self.start_pos.
739 0 : if let Some(lsn) = self.wait_for_lsn().await? {
740 0 : self.end_pos = lsn;
741 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
742 0 : return Ok(());
743 0 : }
744 0 :
745 0 : // Timed out waiting for WAL, check for termination and send KA.
746 0 : // Check for termination only if we are streaming up to commit_lsn
747 0 : // (to pageserver).
748 0 : if let EndWatch::Commit(_) = self.end_watch {
749 0 : if let Some(remote_consistent_lsn) = self
750 0 : .ws_guard
751 0 : .walsenders
752 0 : .get_ws_remote_consistent_lsn(self.ws_guard.id)
753 : {
754 0 : if self.tli.should_walsender_stop(remote_consistent_lsn).await {
755 : // Terminate if there is nothing more to send.
756 : // Note that "ending streaming" part of the string is used by
757 : // pageserver to identify WalReceiverError::SuccessfulCompletion,
758 : // do not change this string without updating pageserver.
759 0 : return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
760 0 : "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
761 0 : self.appname, self.start_pos,
762 0 : )));
763 0 : }
764 0 : }
765 0 : }
766 :
767 0 : let msg = BeMessage::KeepAlive(WalSndKeepAlive {
768 0 : wal_end: self.end_pos.0,
769 0 : timestamp: get_current_timestamp(),
770 0 : request_reply: true,
771 0 : });
772 0 :
773 0 : self.pgb.write_message(&msg).await?;
774 : }
775 0 : }
776 :
777 : /// Wait until we have available WAL > start_pos or timeout expires. Returns
778 : /// - Ok(Some(end_pos)) if needed lsn is successfully observed;
779 : /// - Ok(None) if timeout expired;
780 : /// - Err in case of error -- only if 1) term changed while fetching in recovery
781 : /// mode 2) watch channel closed, which must never happen.
782 0 : async fn wait_for_lsn(&mut self) -> anyhow::Result<Option<Lsn>> {
783 0 : let fp = (|| {
784 0 : fail::fail_point!(
785 0 : "sk-pause-send",
786 0 : self.appname.as_deref() != Some("pageserver"),
787 0 : |_| { true }
788 0 : );
789 0 : false
790 0 : })();
791 0 : if fp {
792 0 : tokio::time::sleep(POLL_STATE_TIMEOUT).await;
793 0 : return Ok(None);
794 0 : }
795 :
796 0 : let res = timeout(POLL_STATE_TIMEOUT, async move {
797 : loop {
798 0 : let end_pos = self.end_watch.get();
799 0 : if end_pos > self.start_pos {
800 0 : return Ok(end_pos);
801 0 : }
802 0 : if let EndWatch::Flush(rx) = &self.end_watch {
803 0 : let curr_term = rx.borrow().term;
804 0 : if let Some(client_term) = self.term {
805 0 : if curr_term != client_term {
806 0 : bail!("term changed: requested {}, now {}", client_term, curr_term);
807 0 : }
808 0 : }
809 0 : }
810 0 : self.end_watch.changed().await?;
811 : }
812 0 : })
813 0 : .await;
814 :
815 0 : match res {
816 : // success
817 0 : Ok(Ok(commit_lsn)) => Ok(Some(commit_lsn)),
818 : // error inside closure
819 0 : Ok(Err(err)) => Err(err),
820 : // timeout
821 0 : Err(_) => Ok(None),
822 : }
823 0 : }
824 : }
825 :
826 : /// A half driving receiving replies.
827 : struct ReplyReader<IO> {
828 : reader: PostgresBackendReader<IO>,
829 : ws_guard: Arc<WalSenderGuard>,
830 : tli: WalResidentTimeline,
831 : }
832 :
833 : impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
834 0 : async fn run(&mut self) -> Result<(), CopyStreamHandlerEnd> {
835 : loop {
836 0 : let msg = self.reader.read_copy_message().await?;
837 0 : self.handle_feedback(&msg).await?
838 : }
839 0 : }
840 :
841 0 : async fn handle_feedback(&mut self, msg: &Bytes) -> anyhow::Result<()> {
842 0 : match msg.first().cloned() {
843 : Some(HOT_STANDBY_FEEDBACK_TAG_BYTE) => {
844 : // Note: deserializing is on m[1..] because we skip the tag byte.
845 0 : let mut hs_feedback = HotStandbyFeedback::des(&msg[1..])
846 0 : .context("failed to deserialize HotStandbyFeedback")?;
847 : // TODO: xmin/catalog_xmin are serialized by walreceiver.c in this way:
848 : // pq_sendint32(&reply_message, xmin);
849 : // pq_sendint32(&reply_message, xmin_epoch);
850 : // So it is two big endian 32-bit words in low endian order!
851 0 : hs_feedback.xmin = hs_feedback.xmin.rotate_left(32);
852 0 : hs_feedback.catalog_xmin = hs_feedback.catalog_xmin.rotate_left(32);
853 0 : self.ws_guard
854 0 : .walsenders
855 0 : .record_hs_feedback(self.ws_guard.id, &hs_feedback);
856 : }
857 : Some(STANDBY_STATUS_UPDATE_TAG_BYTE) => {
858 0 : let reply =
859 0 : StandbyReply::des(&msg[1..]).context("failed to deserialize StandbyReply")?;
860 0 : self.ws_guard
861 0 : .walsenders
862 0 : .record_standby_reply(self.ws_guard.id, &reply);
863 : }
864 : Some(NEON_STATUS_UPDATE_TAG_BYTE) => {
865 : // pageserver sends this.
866 : // Note: deserializing is on m[9..] because we skip the tag byte and len bytes.
867 0 : let buf = Bytes::copy_from_slice(&msg[9..]);
868 0 : let ps_feedback = PageserverFeedback::parse(buf);
869 0 :
870 0 : trace!("PageserverFeedback is {:?}", ps_feedback);
871 0 : self.ws_guard
872 0 : .walsenders
873 0 : .record_ps_feedback(self.ws_guard.id, &ps_feedback);
874 0 : self.tli
875 0 : .update_remote_consistent_lsn(ps_feedback.remote_consistent_lsn)
876 0 : .await;
877 : // in principle new remote_consistent_lsn could allow to
878 : // deactivate the timeline, but we check that regularly through
879 : // broker updated, not need to do it here
880 : }
881 0 : _ => warn!("unexpected message {:?}", msg),
882 : }
883 0 : Ok(())
884 0 : }
885 : }
886 :
887 : #[cfg(test)]
888 : mod tests {
889 : use utils::id::{TenantId, TimelineId};
890 :
891 : use super::*;
892 :
893 4 : fn mock_ttid() -> TenantTimelineId {
894 4 : TenantTimelineId {
895 4 : tenant_id: TenantId::from_slice(&[0x00; 16]).unwrap(),
896 4 : timeline_id: TimelineId::from_slice(&[0x00; 16]).unwrap(),
897 4 : }
898 4 : }
899 :
900 4 : fn mock_addr() -> SocketAddr {
901 4 : "127.0.0.1:8080".parse().unwrap()
902 4 : }
903 :
904 : // add to wss specified feedback setting other fields to dummy values
905 4 : fn push_feedback(wss: &mut WalSendersShared, feedback: ReplicationFeedback) {
906 4 : let walsender_state = WalSenderState {
907 4 : ttid: mock_ttid(),
908 4 : addr: mock_addr(),
909 4 : conn_id: 1,
910 4 : appname: None,
911 4 : feedback,
912 4 : };
913 4 : wss.slots.push(Some(walsender_state))
914 4 : }
915 :
916 : // form standby feedback with given hot standby feedback ts/xmin and the
917 : // rest set to dummy values.
918 4 : fn hs_feedback(ts: TimestampTz, xmin: FullTransactionId) -> ReplicationFeedback {
919 4 : ReplicationFeedback::Standby(StandbyFeedback {
920 4 : reply: StandbyReply::empty(),
921 4 : hs_feedback: HotStandbyFeedback {
922 4 : ts,
923 4 : xmin,
924 4 : catalog_xmin: 0,
925 4 : },
926 4 : })
927 4 : }
928 :
929 : // test that hs aggregation works as expected
930 : #[test]
931 1 : fn test_hs_feedback_no_valid() {
932 1 : let mut wss = WalSendersShared::new();
933 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
934 1 : wss.update_reply_feedback();
935 1 : assert_eq!(
936 1 : wss.agg_standby_feedback.hs_feedback.xmin,
937 1 : INVALID_FULL_TRANSACTION_ID
938 1 : );
939 1 : }
940 :
941 : #[test]
942 1 : fn test_hs_feedback() {
943 1 : let mut wss = WalSendersShared::new();
944 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
945 1 : push_feedback(&mut wss, hs_feedback(1, 42));
946 1 : push_feedback(&mut wss, hs_feedback(1, 64));
947 1 : wss.update_reply_feedback();
948 1 : assert_eq!(wss.agg_standby_feedback.hs_feedback.xmin, 42);
949 1 : }
950 : }
|
