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 std::cmp::{max, min};
5 : use std::net::SocketAddr;
6 : use std::sync::Arc;
7 : use std::time::Duration;
8 :
9 : use anyhow::{Context as AnyhowContext, bail};
10 : use bytes::Bytes;
11 : use futures::FutureExt;
12 : use itertools::Itertools;
13 : use parking_lot::Mutex;
14 : use postgres_backend::{CopyStreamHandlerEnd, PostgresBackend, PostgresBackendReader, QueryError};
15 : use postgres_ffi::{MAX_SEND_SIZE, TimestampTz, get_current_timestamp};
16 : use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody};
17 : use safekeeper_api::Term;
18 : use safekeeper_api::models::{
19 : HotStandbyFeedback, INVALID_FULL_TRANSACTION_ID, ReplicationFeedback, StandbyFeedback,
20 : StandbyReply,
21 : };
22 : use tokio::io::{AsyncRead, AsyncWrite};
23 : use tokio::sync::watch::Receiver;
24 : use tokio::time::timeout;
25 : use tracing::*;
26 : use utils::bin_ser::BeSer;
27 : use utils::failpoint_support;
28 : use utils::lsn::Lsn;
29 : use utils::pageserver_feedback::PageserverFeedback;
30 : use utils::postgres_client::PostgresClientProtocol;
31 :
32 : use crate::handler::SafekeeperPostgresHandler;
33 : use crate::metrics::{RECEIVED_PS_FEEDBACKS, WAL_READERS};
34 : use crate::receive_wal::WalReceivers;
35 : use crate::safekeeper::TermLsn;
36 : use crate::send_interpreted_wal::{
37 : Batch, InterpretedWalReader, InterpretedWalReaderHandle, InterpretedWalSender,
38 : };
39 : use crate::timeline::WalResidentTimeline;
40 : use crate::wal_reader_stream::StreamingWalReader;
41 : use crate::wal_storage::WalReader;
42 :
43 : // See: https://www.postgresql.org/docs/13/protocol-replication.html
44 : const HOT_STANDBY_FEEDBACK_TAG_BYTE: u8 = b'h';
45 : const STANDBY_STATUS_UPDATE_TAG_BYTE: u8 = b'r';
46 : // neon extension of replication protocol
47 : const NEON_STATUS_UPDATE_TAG_BYTE: u8 = b'z';
48 :
49 : /// WalSenders registry. Timeline holds it (wrapped in Arc).
50 : pub struct WalSenders {
51 : mutex: Mutex<WalSendersShared>,
52 : walreceivers: Arc<WalReceivers>,
53 : }
54 :
55 : pub struct WalSendersTimelineMetricValues {
56 : pub ps_feedback_counter: u64,
57 : pub last_ps_feedback: PageserverFeedback,
58 : pub interpreted_wal_reader_tasks: usize,
59 : }
60 :
61 : impl WalSenders {
62 5 : pub fn new(walreceivers: Arc<WalReceivers>) -> Arc<WalSenders> {
63 5 : Arc::new(WalSenders {
64 5 : mutex: Mutex::new(WalSendersShared::new()),
65 5 : walreceivers,
66 5 : })
67 5 : }
68 :
69 : /// Register new walsender. Returned guard provides access to the slot and
70 : /// automatically deregisters in Drop.
71 0 : fn register(self: &Arc<WalSenders>, walsender_state: WalSenderState) -> WalSenderGuard {
72 0 : let slots = &mut self.mutex.lock().slots;
73 : // find empty slot or create new one
74 0 : let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
75 0 : slots[pos] = Some(walsender_state);
76 0 : pos
77 : } else {
78 0 : let pos = slots.len();
79 0 : slots.push(Some(walsender_state));
80 0 : pos
81 : };
82 0 : WalSenderGuard {
83 0 : id: pos,
84 0 : walsenders: self.clone(),
85 0 : }
86 0 : }
87 :
88 0 : fn create_or_update_interpreted_reader<
89 0 : FUp: FnOnce(&Arc<InterpretedWalReaderHandle>) -> anyhow::Result<()>,
90 0 : FNew: FnOnce() -> InterpretedWalReaderHandle,
91 0 : >(
92 0 : self: &Arc<WalSenders>,
93 0 : id: WalSenderId,
94 0 : start_pos: Lsn,
95 0 : max_delta_for_fanout: Option<u64>,
96 0 : update: FUp,
97 0 : create: FNew,
98 0 : ) -> anyhow::Result<()> {
99 0 : let state = &mut self.mutex.lock();
100 0 :
101 0 : let mut selected_interpreted_reader = None;
102 0 : for slot in state.slots.iter().flatten() {
103 0 : if let WalSenderState::Interpreted(slot_state) = slot {
104 0 : if let Some(ref interpreted_reader) = slot_state.interpreted_wal_reader {
105 0 : let select = match (interpreted_reader.current_position(), max_delta_for_fanout)
106 : {
107 0 : (Some(pos), Some(max_delta)) => {
108 0 : let delta = pos.0.abs_diff(start_pos.0);
109 0 : delta <= max_delta
110 : }
111 : // Reader is not active
112 0 : (None, _) => false,
113 : // Gating fanout by max delta is disabled.
114 : // Attach to any active reader.
115 0 : (_, None) => true,
116 : };
117 :
118 0 : if select {
119 0 : selected_interpreted_reader = Some(interpreted_reader.clone());
120 0 : break;
121 0 : }
122 0 : }
123 0 : }
124 : }
125 :
126 0 : let slot = state.get_slot_mut(id);
127 0 : let slot_state = match slot {
128 0 : WalSenderState::Interpreted(s) => s,
129 0 : WalSenderState::Vanilla(_) => unreachable!(),
130 : };
131 :
132 0 : let selected_or_new = match selected_interpreted_reader {
133 0 : Some(selected) => {
134 0 : update(&selected)?;
135 0 : selected
136 : }
137 0 : None => Arc::new(create()),
138 : };
139 :
140 0 : slot_state.interpreted_wal_reader = Some(selected_or_new);
141 0 :
142 0 : Ok(())
143 0 : }
144 :
145 : /// Get state of all walsenders.
146 0 : pub fn get_all_public(self: &Arc<WalSenders>) -> Vec<safekeeper_api::models::WalSenderState> {
147 0 : self.mutex
148 0 : .lock()
149 0 : .slots
150 0 : .iter()
151 0 : .flatten()
152 0 : .map(|state| match state {
153 0 : WalSenderState::Vanilla(s) => {
154 0 : safekeeper_api::models::WalSenderState::Vanilla(s.clone())
155 : }
156 0 : WalSenderState::Interpreted(s) => {
157 0 : safekeeper_api::models::WalSenderState::Interpreted(s.public_state.clone())
158 : }
159 0 : })
160 0 : .collect()
161 0 : }
162 :
163 : /// Get LSN of the most lagging pageserver receiver. Return None if there are no
164 : /// active walsenders.
165 0 : pub fn laggard_lsn(self: &Arc<WalSenders>) -> Option<Lsn> {
166 0 : self.mutex
167 0 : .lock()
168 0 : .slots
169 0 : .iter()
170 0 : .flatten()
171 0 : .filter_map(|s| match s.get_feedback() {
172 0 : ReplicationFeedback::Pageserver(feedback) => Some(feedback.last_received_lsn),
173 0 : ReplicationFeedback::Standby(_) => None,
174 0 : })
175 0 : .min()
176 0 : }
177 :
178 : /// Returns total counter of pageserver feedbacks received and last feedback.
179 0 : pub fn info_for_metrics(self: &Arc<WalSenders>) -> WalSendersTimelineMetricValues {
180 0 : let shared = self.mutex.lock();
181 0 :
182 0 : let interpreted_wal_reader_tasks = shared
183 0 : .slots
184 0 : .iter()
185 0 : .filter_map(|ss| match ss {
186 0 : Some(WalSenderState::Interpreted(int)) => int.interpreted_wal_reader.as_ref(),
187 0 : Some(WalSenderState::Vanilla(_)) => None,
188 0 : None => None,
189 0 : })
190 0 : .unique_by(|reader| Arc::as_ptr(reader))
191 0 : .count();
192 0 :
193 0 : WalSendersTimelineMetricValues {
194 0 : ps_feedback_counter: shared.ps_feedback_counter,
195 0 : last_ps_feedback: shared.last_ps_feedback,
196 0 : interpreted_wal_reader_tasks,
197 0 : }
198 0 : }
199 :
200 : /// Get aggregated hot standby feedback (we send it to compute).
201 620 : pub fn get_hotstandby(self: &Arc<WalSenders>) -> StandbyFeedback {
202 620 : self.mutex.lock().agg_standby_feedback
203 620 : }
204 :
205 : /// Record new pageserver feedback, update aggregated values.
206 0 : fn record_ps_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &PageserverFeedback) {
207 0 : let mut shared = self.mutex.lock();
208 0 : *shared.get_slot_mut(id).get_mut_feedback() = ReplicationFeedback::Pageserver(*feedback);
209 0 : shared.last_ps_feedback = *feedback;
210 0 : shared.ps_feedback_counter += 1;
211 0 : drop(shared);
212 0 :
213 0 : RECEIVED_PS_FEEDBACKS.inc();
214 0 :
215 0 : // send feedback to connected walproposers
216 0 : self.walreceivers.broadcast_pageserver_feedback(*feedback);
217 0 : }
218 :
219 : /// Record standby reply.
220 0 : fn record_standby_reply(self: &Arc<WalSenders>, id: WalSenderId, reply: &StandbyReply) {
221 0 : let mut shared = self.mutex.lock();
222 0 : let slot = shared.get_slot_mut(id);
223 0 : debug!(
224 0 : "Record standby reply: ts={} apply_lsn={}",
225 : reply.reply_ts, reply.apply_lsn
226 : );
227 0 : match &mut slot.get_mut_feedback() {
228 0 : ReplicationFeedback::Standby(sf) => sf.reply = *reply,
229 : ReplicationFeedback::Pageserver(_) => {
230 0 : *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback {
231 0 : reply: *reply,
232 0 : hs_feedback: HotStandbyFeedback::empty(),
233 0 : })
234 : }
235 : }
236 0 : }
237 :
238 : /// Record hot standby feedback, update aggregated value.
239 0 : fn record_hs_feedback(self: &Arc<WalSenders>, id: WalSenderId, feedback: &HotStandbyFeedback) {
240 0 : let mut shared = self.mutex.lock();
241 0 : let slot = shared.get_slot_mut(id);
242 0 : match &mut slot.get_mut_feedback() {
243 0 : ReplicationFeedback::Standby(sf) => sf.hs_feedback = *feedback,
244 : ReplicationFeedback::Pageserver(_) => {
245 0 : *slot.get_mut_feedback() = ReplicationFeedback::Standby(StandbyFeedback {
246 0 : reply: StandbyReply::empty(),
247 0 : hs_feedback: *feedback,
248 0 : })
249 : }
250 : }
251 0 : shared.update_reply_feedback();
252 0 : }
253 :
254 : /// Get remote_consistent_lsn reported by the pageserver. Returns None if
255 : /// client is not pageserver.
256 0 : pub fn get_ws_remote_consistent_lsn(self: &Arc<WalSenders>, id: WalSenderId) -> Option<Lsn> {
257 0 : let shared = self.mutex.lock();
258 0 : let slot = shared.get_slot(id);
259 0 : match slot.get_feedback() {
260 0 : ReplicationFeedback::Pageserver(feedback) => Some(feedback.remote_consistent_lsn),
261 0 : _ => None,
262 : }
263 0 : }
264 :
265 : /// Unregister walsender.
266 0 : fn unregister(self: &Arc<WalSenders>, id: WalSenderId) {
267 0 : let mut shared = self.mutex.lock();
268 0 : shared.slots[id] = None;
269 0 : shared.update_reply_feedback();
270 0 : }
271 : }
272 :
273 : struct WalSendersShared {
274 : // aggregated over all walsenders value
275 : agg_standby_feedback: StandbyFeedback,
276 : // last feedback ever received from any pageserver, empty if none
277 : last_ps_feedback: PageserverFeedback,
278 : // total counter of pageserver feedbacks received
279 : ps_feedback_counter: u64,
280 : slots: Vec<Option<WalSenderState>>,
281 : }
282 :
283 : /// Safekeeper internal definitions of wal sender state
284 : ///
285 : /// As opposed to [`safekeeper_api::models::WalSenderState`] these struct may
286 : /// include state that we don not wish to expose to the public api.
287 : #[derive(Debug, Clone)]
288 : pub(crate) enum WalSenderState {
289 : Vanilla(VanillaWalSenderInternalState),
290 : Interpreted(InterpretedWalSenderInternalState),
291 : }
292 :
293 : type VanillaWalSenderInternalState = safekeeper_api::models::VanillaWalSenderState;
294 :
295 : #[derive(Debug, Clone)]
296 : pub(crate) struct InterpretedWalSenderInternalState {
297 : public_state: safekeeper_api::models::InterpretedWalSenderState,
298 : interpreted_wal_reader: Option<Arc<InterpretedWalReaderHandle>>,
299 : }
300 :
301 : impl WalSenderState {
302 0 : fn get_addr(&self) -> &SocketAddr {
303 0 : match self {
304 0 : WalSenderState::Vanilla(state) => &state.addr,
305 0 : WalSenderState::Interpreted(state) => &state.public_state.addr,
306 : }
307 0 : }
308 :
309 4 : fn get_feedback(&self) -> &ReplicationFeedback {
310 4 : match self {
311 4 : WalSenderState::Vanilla(state) => &state.feedback,
312 0 : WalSenderState::Interpreted(state) => &state.public_state.feedback,
313 : }
314 4 : }
315 :
316 0 : fn get_mut_feedback(&mut self) -> &mut ReplicationFeedback {
317 0 : match self {
318 0 : WalSenderState::Vanilla(state) => &mut state.feedback,
319 0 : WalSenderState::Interpreted(state) => &mut state.public_state.feedback,
320 : }
321 0 : }
322 : }
323 :
324 : impl WalSendersShared {
325 7 : fn new() -> Self {
326 7 : WalSendersShared {
327 7 : agg_standby_feedback: StandbyFeedback::empty(),
328 7 : last_ps_feedback: PageserverFeedback::empty(),
329 7 : ps_feedback_counter: 0,
330 7 : slots: Vec::new(),
331 7 : }
332 7 : }
333 :
334 : /// Get content of provided id slot, it must exist.
335 0 : fn get_slot(&self, id: WalSenderId) -> &WalSenderState {
336 0 : self.slots[id].as_ref().expect("walsender doesn't exist")
337 0 : }
338 :
339 : /// Get mut content of provided id slot, it must exist.
340 0 : fn get_slot_mut(&mut self, id: WalSenderId) -> &mut WalSenderState {
341 0 : self.slots[id].as_mut().expect("walsender doesn't exist")
342 0 : }
343 :
344 : /// Update aggregated hot standy and normal reply feedbacks. We just take min of valid xmins
345 : /// and ts.
346 2 : fn update_reply_feedback(&mut self) {
347 2 : let mut agg = HotStandbyFeedback::empty();
348 2 : let mut reply_agg = StandbyReply::empty();
349 4 : for ws_state in self.slots.iter().flatten() {
350 4 : if let ReplicationFeedback::Standby(standby_feedback) = ws_state.get_feedback() {
351 4 : let hs_feedback = standby_feedback.hs_feedback;
352 4 : // doing Option math like op1.iter().chain(op2.iter()).min()
353 4 : // would be nicer, but we serialize/deserialize this struct
354 4 : // directly, so leave as is for now
355 4 : if hs_feedback.xmin != INVALID_FULL_TRANSACTION_ID {
356 2 : if agg.xmin != INVALID_FULL_TRANSACTION_ID {
357 1 : agg.xmin = min(agg.xmin, hs_feedback.xmin);
358 1 : } else {
359 1 : agg.xmin = hs_feedback.xmin;
360 1 : }
361 2 : agg.ts = max(agg.ts, hs_feedback.ts);
362 2 : }
363 4 : if hs_feedback.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
364 0 : if agg.catalog_xmin != INVALID_FULL_TRANSACTION_ID {
365 0 : agg.catalog_xmin = min(agg.catalog_xmin, hs_feedback.catalog_xmin);
366 0 : } else {
367 0 : agg.catalog_xmin = hs_feedback.catalog_xmin;
368 0 : }
369 0 : agg.ts = max(agg.ts, hs_feedback.ts);
370 4 : }
371 4 : let reply = standby_feedback.reply;
372 4 : if reply.write_lsn != Lsn::INVALID {
373 0 : if reply_agg.write_lsn != Lsn::INVALID {
374 0 : reply_agg.write_lsn = Lsn::min(reply_agg.write_lsn, reply.write_lsn);
375 0 : } else {
376 0 : reply_agg.write_lsn = reply.write_lsn;
377 0 : }
378 4 : }
379 4 : if reply.flush_lsn != Lsn::INVALID {
380 0 : if reply_agg.flush_lsn != Lsn::INVALID {
381 0 : reply_agg.flush_lsn = Lsn::min(reply_agg.flush_lsn, reply.flush_lsn);
382 0 : } else {
383 0 : reply_agg.flush_lsn = reply.flush_lsn;
384 0 : }
385 4 : }
386 4 : if reply.apply_lsn != Lsn::INVALID {
387 0 : if reply_agg.apply_lsn != Lsn::INVALID {
388 0 : reply_agg.apply_lsn = Lsn::min(reply_agg.apply_lsn, reply.apply_lsn);
389 0 : } else {
390 0 : reply_agg.apply_lsn = reply.apply_lsn;
391 0 : }
392 4 : }
393 4 : if reply.reply_ts != 0 {
394 0 : if reply_agg.reply_ts != 0 {
395 0 : reply_agg.reply_ts = TimestampTz::min(reply_agg.reply_ts, reply.reply_ts);
396 0 : } else {
397 0 : reply_agg.reply_ts = reply.reply_ts;
398 0 : }
399 4 : }
400 0 : }
401 : }
402 2 : self.agg_standby_feedback = StandbyFeedback {
403 2 : reply: reply_agg,
404 2 : hs_feedback: agg,
405 2 : };
406 2 : }
407 : }
408 :
409 : // id of the occupied slot in WalSenders to access it (and save in the
410 : // WalSenderGuard). We could give Arc directly to the slot, but there is not
411 : // much sense in that as values aggregation which is performed on each feedback
412 : // receival iterates over all walsenders.
413 : pub type WalSenderId = usize;
414 :
415 : /// Scope guard to access slot in WalSenders registry and unregister from it in
416 : /// Drop.
417 : pub struct WalSenderGuard {
418 : id: WalSenderId,
419 : walsenders: Arc<WalSenders>,
420 : }
421 :
422 : impl WalSenderGuard {
423 0 : pub fn id(&self) -> WalSenderId {
424 0 : self.id
425 0 : }
426 :
427 0 : pub fn walsenders(&self) -> &Arc<WalSenders> {
428 0 : &self.walsenders
429 0 : }
430 : }
431 :
432 : impl Drop for WalSenderGuard {
433 0 : fn drop(&mut self) {
434 0 : self.walsenders.unregister(self.id);
435 0 : }
436 : }
437 :
438 : impl SafekeeperPostgresHandler {
439 : /// Wrapper around handle_start_replication_guts handling result. Error is
440 : /// handled here while we're still in walsender ttid span; with API
441 : /// extension, this can probably be moved into postgres_backend.
442 0 : pub async fn handle_start_replication<IO: AsyncRead + AsyncWrite + Unpin + Send>(
443 0 : &mut self,
444 0 : pgb: &mut PostgresBackend<IO>,
445 0 : start_pos: Lsn,
446 0 : term: Option<Term>,
447 0 : ) -> Result<(), QueryError> {
448 0 : let tli = self
449 0 : .global_timelines
450 0 : .get(self.ttid)
451 0 : .map_err(|e| QueryError::Other(e.into()))?;
452 0 : let residence_guard = tli.wal_residence_guard().await?;
453 :
454 0 : if let Err(end) = self
455 0 : .handle_start_replication_guts(pgb, start_pos, term, residence_guard)
456 0 : .await
457 : {
458 0 : let info = tli.get_safekeeper_info(&self.conf).await;
459 : // Log the result and probably send it to the client, closing the stream.
460 0 : pgb.handle_copy_stream_end(end)
461 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)))
462 0 : .await;
463 0 : }
464 0 : Ok(())
465 0 : }
466 :
467 0 : pub async fn handle_start_replication_guts<IO: AsyncRead + AsyncWrite + Unpin + Send>(
468 0 : &mut self,
469 0 : pgb: &mut PostgresBackend<IO>,
470 0 : start_pos: Lsn,
471 0 : term: Option<Term>,
472 0 : tli: WalResidentTimeline,
473 0 : ) -> Result<(), CopyStreamHandlerEnd> {
474 0 : let appname = self.appname.clone();
475 :
476 : // Use a guard object to remove our entry from the timeline when we are done.
477 0 : let ws_guard = match self.protocol() {
478 0 : PostgresClientProtocol::Vanilla => Arc::new(tli.get_walsenders().register(
479 0 : WalSenderState::Vanilla(VanillaWalSenderInternalState {
480 0 : ttid: self.ttid,
481 0 : addr: *pgb.get_peer_addr(),
482 0 : conn_id: self.conn_id,
483 0 : appname: self.appname.clone(),
484 0 : feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
485 0 : }),
486 0 : )),
487 0 : PostgresClientProtocol::Interpreted { .. } => Arc::new(tli.get_walsenders().register(
488 0 : WalSenderState::Interpreted(InterpretedWalSenderInternalState {
489 0 : public_state: safekeeper_api::models::InterpretedWalSenderState {
490 0 : ttid: self.ttid,
491 0 : shard: self.shard.unwrap(),
492 0 : addr: *pgb.get_peer_addr(),
493 0 : conn_id: self.conn_id,
494 0 : appname: self.appname.clone(),
495 0 : feedback: ReplicationFeedback::Pageserver(PageserverFeedback::empty()),
496 0 : },
497 0 : interpreted_wal_reader: None,
498 0 : }),
499 0 : )),
500 : };
501 :
502 : // Walsender can operate in one of two modes which we select by
503 : // application_name: give only committed WAL (used by pageserver) or all
504 : // existing WAL (up to flush_lsn, used by walproposer or peer recovery).
505 : // The second case is always driven by a consensus leader which term
506 : // must be supplied.
507 0 : let end_watch = if term.is_some() {
508 0 : EndWatch::Flush(tli.get_term_flush_lsn_watch_rx())
509 : } else {
510 0 : EndWatch::Commit(tli.get_commit_lsn_watch_rx())
511 : };
512 : // we don't check term here; it will be checked on first waiting/WAL reading anyway.
513 0 : let end_pos = end_watch.get();
514 0 :
515 0 : if end_pos < start_pos {
516 0 : warn!(
517 0 : "requested start_pos {} is ahead of available WAL end_pos {}",
518 : start_pos, end_pos
519 : );
520 0 : }
521 :
522 0 : info!(
523 0 : "starting streaming from {:?}, available WAL ends at {}, recovery={}, appname={:?}, protocol={:?}",
524 : start_pos,
525 : end_pos,
526 0 : matches!(end_watch, EndWatch::Flush(_)),
527 : appname,
528 0 : self.protocol(),
529 : );
530 :
531 : // switch to copy
532 0 : pgb.write_message(&BeMessage::CopyBothResponse).await?;
533 :
534 0 : let wal_reader = tli.get_walreader(start_pos).await?;
535 :
536 : // Split to concurrently receive and send data; replies are generally
537 : // not synchronized with sends, so this avoids deadlocks.
538 0 : let reader = pgb.split().context("START_REPLICATION split")?;
539 :
540 0 : let send_fut = match self.protocol() {
541 : PostgresClientProtocol::Vanilla => {
542 0 : let sender = WalSender {
543 0 : pgb,
544 0 : // should succeed since we're already holding another guard
545 0 : tli: tli.wal_residence_guard().await?,
546 0 : appname: appname.clone(),
547 0 : start_pos,
548 0 : end_pos,
549 0 : term,
550 0 : end_watch,
551 0 : ws_guard: ws_guard.clone(),
552 0 : wal_reader,
553 0 : send_buf: vec![0u8; MAX_SEND_SIZE],
554 0 : };
555 0 :
556 0 : FutureExt::boxed(sender.run())
557 : }
558 : PostgresClientProtocol::Interpreted {
559 0 : format,
560 0 : compression,
561 : } => {
562 0 : let pg_version = tli.tli.get_state().await.1.server.pg_version / 10000;
563 0 : let end_watch_view = end_watch.view();
564 0 : let wal_residence_guard = tli.wal_residence_guard().await?;
565 0 : let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(2);
566 0 : let shard = self.shard.unwrap();
567 0 :
568 0 : if self.conf.wal_reader_fanout && !shard.is_unsharded() {
569 0 : let ws_id = ws_guard.id();
570 0 : ws_guard.walsenders().create_or_update_interpreted_reader(
571 0 : ws_id,
572 0 : start_pos,
573 0 : self.conf.max_delta_for_fanout,
574 0 : {
575 0 : let tx = tx.clone();
576 0 : |reader| {
577 0 : tracing::info!(
578 0 : "Fanning out interpreted wal reader at {}",
579 : start_pos
580 : );
581 0 : reader
582 0 : .fanout(shard, tx, start_pos)
583 0 : .with_context(|| "Failed to fan out reader")
584 0 : }
585 0 : },
586 0 : || {
587 0 : tracing::info!("Spawning interpreted wal reader at {}", start_pos);
588 :
589 0 : let wal_stream = StreamingWalReader::new(
590 0 : wal_residence_guard,
591 0 : term,
592 0 : start_pos,
593 0 : end_pos,
594 0 : end_watch,
595 0 : MAX_SEND_SIZE,
596 0 : );
597 0 :
598 0 : InterpretedWalReader::spawn(
599 0 : wal_stream, start_pos, tx, shard, pg_version, &appname,
600 0 : )
601 0 : },
602 0 : )?;
603 :
604 0 : let sender = InterpretedWalSender {
605 0 : format,
606 0 : compression,
607 0 : appname,
608 0 : tli: tli.wal_residence_guard().await?,
609 0 : start_lsn: start_pos,
610 0 : pgb,
611 0 : end_watch_view,
612 0 : wal_sender_guard: ws_guard.clone(),
613 0 : rx,
614 0 : };
615 0 :
616 0 : FutureExt::boxed(sender.run())
617 : } else {
618 0 : let wal_reader = StreamingWalReader::new(
619 0 : wal_residence_guard,
620 0 : term,
621 0 : start_pos,
622 0 : end_pos,
623 0 : end_watch,
624 0 : MAX_SEND_SIZE,
625 0 : );
626 0 :
627 0 : let reader = InterpretedWalReader::new(
628 0 : wal_reader, start_pos, tx, shard, pg_version, None,
629 0 : );
630 :
631 0 : let sender = InterpretedWalSender {
632 0 : format,
633 0 : compression,
634 0 : appname: appname.clone(),
635 0 : tli: tli.wal_residence_guard().await?,
636 0 : start_lsn: start_pos,
637 0 : pgb,
638 0 : end_watch_view,
639 0 : wal_sender_guard: ws_guard.clone(),
640 0 : rx,
641 0 : };
642 0 :
643 0 : FutureExt::boxed(async move {
644 : // Sender returns an Err on all code paths.
645 : // If the sender finishes first, we will drop the reader future.
646 : // If the reader finishes first, the sender will finish too since
647 : // the wal sender has dropped.
648 0 : let res = tokio::try_join!(sender.run(), reader.run(start_pos, &appname));
649 0 : match res.map(|_| ()) {
650 0 : Ok(_) => unreachable!("sender finishes with Err by convention"),
651 0 : err_res => err_res,
652 0 : }
653 0 : })
654 : }
655 : }
656 : };
657 :
658 0 : let tli_cancel = tli.cancel.clone();
659 0 :
660 0 : let mut reply_reader = ReplyReader {
661 0 : reader,
662 0 : ws_guard: ws_guard.clone(),
663 0 : tli,
664 0 : };
665 :
666 0 : let res = tokio::select! {
667 : // todo: add read|write .context to these errors
668 0 : r = send_fut => r,
669 0 : r = reply_reader.run() => r,
670 0 : _ = tli_cancel.cancelled() => {
671 0 : return Err(CopyStreamHandlerEnd::Cancelled);
672 : }
673 : };
674 :
675 0 : let ws_state = ws_guard
676 0 : .walsenders
677 0 : .mutex
678 0 : .lock()
679 0 : .get_slot(ws_guard.id)
680 0 : .clone();
681 0 : info!(
682 0 : "finished streaming to {}, feedback={:?}",
683 0 : ws_state.get_addr(),
684 0 : ws_state.get_feedback(),
685 : );
686 :
687 : // Join pg backend back.
688 0 : pgb.unsplit(reply_reader.reader)?;
689 :
690 0 : res
691 0 : }
692 : }
693 :
694 : /// TODO(vlad): maybe lift this instead
695 : /// Walsender streams either up to commit_lsn (normally) or flush_lsn in the
696 : /// given term (recovery by walproposer or peer safekeeper).
697 : #[derive(Clone)]
698 : pub(crate) enum EndWatch {
699 : Commit(Receiver<Lsn>),
700 : Flush(Receiver<TermLsn>),
701 : }
702 :
703 : impl EndWatch {
704 0 : pub(crate) fn view(&self) -> EndWatchView {
705 0 : EndWatchView(self.clone())
706 0 : }
707 :
708 : /// Get current end of WAL.
709 12 : pub(crate) fn get(&self) -> Lsn {
710 12 : match self {
711 12 : EndWatch::Commit(r) => *r.borrow(),
712 0 : EndWatch::Flush(r) => r.borrow().lsn,
713 : }
714 12 : }
715 :
716 : /// Wait for the update.
717 7 : pub(crate) async fn changed(&mut self) -> anyhow::Result<()> {
718 7 : match self {
719 7 : EndWatch::Commit(r) => r.changed().await?,
720 0 : EndWatch::Flush(r) => r.changed().await?,
721 : }
722 3 : Ok(())
723 3 : }
724 :
725 4 : pub(crate) async fn wait_for_lsn(
726 4 : &mut self,
727 4 : lsn: Lsn,
728 4 : client_term: Option<Term>,
729 4 : ) -> anyhow::Result<Lsn> {
730 : loop {
731 7 : let end_pos = self.get();
732 7 : if end_pos > lsn {
733 0 : return Ok(end_pos);
734 7 : }
735 7 : if let EndWatch::Flush(rx) = &self {
736 0 : let curr_term = rx.borrow().term;
737 0 : if let Some(client_term) = client_term {
738 0 : if curr_term != client_term {
739 0 : bail!("term changed: requested {}, now {}", client_term, curr_term);
740 0 : }
741 0 : }
742 7 : }
743 7 : self.changed().await?;
744 : }
745 0 : }
746 : }
747 :
748 : pub(crate) struct EndWatchView(EndWatch);
749 :
750 : impl EndWatchView {
751 0 : pub(crate) fn get(&self) -> Lsn {
752 0 : self.0.get()
753 0 : }
754 : }
755 : /// A half driving sending WAL.
756 : struct WalSender<'a, IO> {
757 : pgb: &'a mut PostgresBackend<IO>,
758 : tli: WalResidentTimeline,
759 : appname: Option<String>,
760 : // Position since which we are sending next chunk.
761 : start_pos: Lsn,
762 : // WAL up to this position is known to be locally available.
763 : // Usually this is the same as the latest commit_lsn, but in case of
764 : // walproposer recovery, this is flush_lsn.
765 : //
766 : // We send this LSN to the receiver as wal_end, so that it knows how much
767 : // WAL this safekeeper has. This LSN should be as fresh as possible.
768 : end_pos: Lsn,
769 : /// When streaming uncommitted part, the term the client acts as the leader
770 : /// in. Streaming is stopped if local term changes to a different (higher)
771 : /// value.
772 : term: Option<Term>,
773 : /// Watch channel receiver to learn end of available WAL (and wait for its advancement).
774 : end_watch: EndWatch,
775 : ws_guard: Arc<WalSenderGuard>,
776 : wal_reader: WalReader,
777 : // buffer for readling WAL into to send it
778 : send_buf: Vec<u8>,
779 : }
780 :
781 : const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1);
782 :
783 : impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
784 : /// Send WAL until
785 : /// - an error occurs
786 : /// - receiver is caughtup and there is no computes (if streaming up to commit_lsn)
787 : /// - timeline's cancellation token fires
788 : ///
789 : /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
790 : /// convenience.
791 0 : async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> {
792 0 : let metric = WAL_READERS
793 0 : .get_metric_with_label_values(&[
794 0 : "future",
795 0 : self.appname.as_deref().unwrap_or("safekeeper"),
796 0 : ])
797 0 : .unwrap();
798 0 :
799 0 : metric.inc();
800 0 : scopeguard::defer! {
801 0 : metric.dec();
802 0 : }
803 :
804 : loop {
805 : // Wait for the next portion if it is not there yet, or just
806 : // update our end of WAL available for sending value, we
807 : // communicate it to the receiver.
808 0 : self.wait_wal().await?;
809 0 : assert!(
810 0 : self.end_pos > self.start_pos,
811 0 : "nothing to send after waiting for WAL"
812 : );
813 :
814 : // try to send as much as available, capped by MAX_SEND_SIZE
815 0 : let mut chunk_end_pos = self.start_pos + MAX_SEND_SIZE as u64;
816 0 : // if we went behind available WAL, back off
817 0 : if chunk_end_pos >= self.end_pos {
818 0 : chunk_end_pos = self.end_pos;
819 0 : } else {
820 0 : // If sending not up to end pos, round down to page boundary to
821 0 : // avoid breaking WAL record not at page boundary, as protocol
822 0 : // demands. See walsender.c (XLogSendPhysical).
823 0 : chunk_end_pos = chunk_end_pos
824 0 : .checked_sub(chunk_end_pos.block_offset())
825 0 : .unwrap();
826 0 : }
827 0 : let send_size = (chunk_end_pos.0 - self.start_pos.0) as usize;
828 0 : let send_buf = &mut self.send_buf[..send_size];
829 : let send_size: usize;
830 : {
831 : // If uncommitted part is being pulled, check that the term is
832 : // still the expected one.
833 0 : let _term_guard = if let Some(t) = self.term {
834 0 : Some(self.tli.acquire_term(t).await?)
835 : } else {
836 0 : None
837 : };
838 : // Read WAL into buffer. send_size can be additionally capped to
839 : // segment boundary here.
840 0 : send_size = self.wal_reader.read(send_buf).await?
841 : };
842 0 : let send_buf = &send_buf[..send_size];
843 0 :
844 0 : // and send it, while respecting Timeline::cancel
845 0 : let msg = BeMessage::XLogData(XLogDataBody {
846 0 : wal_start: self.start_pos.0,
847 0 : wal_end: self.end_pos.0,
848 0 : timestamp: get_current_timestamp(),
849 0 : data: send_buf,
850 0 : });
851 0 : self.pgb.write_message(&msg).await?;
852 :
853 0 : if let Some(appname) = &self.appname {
854 0 : if appname == "replica" {
855 0 : failpoint_support::sleep_millis_async!("sk-send-wal-replica-sleep");
856 0 : }
857 0 : }
858 0 : trace!(
859 0 : "sent {} bytes of WAL {}-{}",
860 0 : send_size,
861 0 : self.start_pos,
862 0 : self.start_pos + send_size as u64
863 : );
864 0 : self.start_pos += send_size as u64;
865 : }
866 0 : }
867 :
868 : /// wait until we have WAL to stream, sending keepalives and checking for
869 : /// exit in the meanwhile
870 0 : async fn wait_wal(&mut self) -> Result<(), CopyStreamHandlerEnd> {
871 : loop {
872 0 : self.end_pos = self.end_watch.get();
873 0 : let have_something_to_send = (|| {
874 0 : fail::fail_point!(
875 0 : "sk-pause-send",
876 0 : self.appname.as_deref() != Some("pageserver"),
877 0 : |_| { false }
878 0 : );
879 0 : self.end_pos > self.start_pos
880 0 : })();
881 0 :
882 0 : if have_something_to_send {
883 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
884 0 : return Ok(());
885 0 : }
886 :
887 : // Wait for WAL to appear, now self.end_pos == self.start_pos.
888 0 : if let Some(lsn) = self.wait_for_lsn().await? {
889 0 : self.end_pos = lsn;
890 0 : trace!("got end_pos {:?}, streaming", self.end_pos);
891 0 : return Ok(());
892 0 : }
893 0 :
894 0 : // Timed out waiting for WAL, check for termination and send KA.
895 0 : // Check for termination only if we are streaming up to commit_lsn
896 0 : // (to pageserver).
897 0 : if let EndWatch::Commit(_) = self.end_watch {
898 0 : if let Some(remote_consistent_lsn) = self
899 0 : .ws_guard
900 0 : .walsenders
901 0 : .get_ws_remote_consistent_lsn(self.ws_guard.id)
902 : {
903 0 : if self.tli.should_walsender_stop(remote_consistent_lsn).await {
904 : // Terminate if there is nothing more to send.
905 : // Note that "ending streaming" part of the string is used by
906 : // pageserver to identify WalReceiverError::SuccessfulCompletion,
907 : // do not change this string without updating pageserver.
908 0 : return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
909 0 : "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
910 0 : self.appname, self.start_pos,
911 0 : )));
912 0 : }
913 0 : }
914 0 : }
915 :
916 0 : let msg = BeMessage::KeepAlive(WalSndKeepAlive {
917 0 : wal_end: self.end_pos.0,
918 0 : timestamp: get_current_timestamp(),
919 0 : request_reply: true,
920 0 : });
921 0 :
922 0 : self.pgb.write_message(&msg).await?;
923 : }
924 0 : }
925 :
926 : /// Wait until we have available WAL > start_pos or timeout expires. Returns
927 : /// - Ok(Some(end_pos)) if needed lsn is successfully observed;
928 : /// - Ok(None) if timeout expired;
929 : /// - Err in case of error -- only if 1) term changed while fetching in recovery
930 : /// mode 2) watch channel closed, which must never happen.
931 0 : async fn wait_for_lsn(&mut self) -> anyhow::Result<Option<Lsn>> {
932 0 : let fp = (|| {
933 0 : fail::fail_point!(
934 0 : "sk-pause-send",
935 0 : self.appname.as_deref() != Some("pageserver"),
936 0 : |_| { true }
937 0 : );
938 0 : false
939 0 : })();
940 0 : if fp {
941 0 : tokio::time::sleep(POLL_STATE_TIMEOUT).await;
942 0 : return Ok(None);
943 0 : }
944 :
945 0 : let res = timeout(POLL_STATE_TIMEOUT, async move {
946 : loop {
947 0 : let end_pos = self.end_watch.get();
948 0 : if end_pos > self.start_pos {
949 0 : return Ok(end_pos);
950 0 : }
951 0 : if let EndWatch::Flush(rx) = &self.end_watch {
952 0 : let curr_term = rx.borrow().term;
953 0 : if let Some(client_term) = self.term {
954 0 : if curr_term != client_term {
955 0 : bail!("term changed: requested {}, now {}", client_term, curr_term);
956 0 : }
957 0 : }
958 0 : }
959 0 : self.end_watch.changed().await?;
960 : }
961 0 : })
962 0 : .await;
963 :
964 0 : match res {
965 : // success
966 0 : Ok(Ok(commit_lsn)) => Ok(Some(commit_lsn)),
967 : // error inside closure
968 0 : Ok(Err(err)) => Err(err),
969 : // timeout
970 0 : Err(_) => Ok(None),
971 : }
972 0 : }
973 : }
974 :
975 : /// A half driving receiving replies.
976 : struct ReplyReader<IO> {
977 : reader: PostgresBackendReader<IO>,
978 : ws_guard: Arc<WalSenderGuard>,
979 : tli: WalResidentTimeline,
980 : }
981 :
982 : impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
983 0 : async fn run(&mut self) -> Result<(), CopyStreamHandlerEnd> {
984 : loop {
985 0 : let msg = self.reader.read_copy_message().await?;
986 0 : self.handle_feedback(&msg).await?
987 : }
988 0 : }
989 :
990 0 : async fn handle_feedback(&mut self, msg: &Bytes) -> anyhow::Result<()> {
991 0 : match msg.first().cloned() {
992 : Some(HOT_STANDBY_FEEDBACK_TAG_BYTE) => {
993 : // Note: deserializing is on m[1..] because we skip the tag byte.
994 0 : let mut hs_feedback = HotStandbyFeedback::des(&msg[1..])
995 0 : .context("failed to deserialize HotStandbyFeedback")?;
996 : // TODO: xmin/catalog_xmin are serialized by walreceiver.c in this way:
997 : // pq_sendint32(&reply_message, xmin);
998 : // pq_sendint32(&reply_message, xmin_epoch);
999 : // So it is two big endian 32-bit words in low endian order!
1000 0 : hs_feedback.xmin = hs_feedback.xmin.rotate_left(32);
1001 0 : hs_feedback.catalog_xmin = hs_feedback.catalog_xmin.rotate_left(32);
1002 0 : self.ws_guard
1003 0 : .walsenders
1004 0 : .record_hs_feedback(self.ws_guard.id, &hs_feedback);
1005 : }
1006 : Some(STANDBY_STATUS_UPDATE_TAG_BYTE) => {
1007 0 : let reply =
1008 0 : StandbyReply::des(&msg[1..]).context("failed to deserialize StandbyReply")?;
1009 0 : self.ws_guard
1010 0 : .walsenders
1011 0 : .record_standby_reply(self.ws_guard.id, &reply);
1012 : }
1013 : Some(NEON_STATUS_UPDATE_TAG_BYTE) => {
1014 : // pageserver sends this.
1015 : // Note: deserializing is on m[9..] because we skip the tag byte and len bytes.
1016 0 : let buf = Bytes::copy_from_slice(&msg[9..]);
1017 0 : let ps_feedback = PageserverFeedback::parse(buf);
1018 0 :
1019 0 : trace!("PageserverFeedback is {:?}", ps_feedback);
1020 0 : self.ws_guard
1021 0 : .walsenders
1022 0 : .record_ps_feedback(self.ws_guard.id, &ps_feedback);
1023 0 : self.tli
1024 0 : .update_remote_consistent_lsn(ps_feedback.remote_consistent_lsn)
1025 0 : .await;
1026 : // in principle new remote_consistent_lsn could allow to
1027 : // deactivate the timeline, but we check that regularly through
1028 : // broker updated, not need to do it here
1029 : }
1030 0 : _ => warn!("unexpected message {:?}", msg),
1031 : }
1032 0 : Ok(())
1033 0 : }
1034 : }
1035 :
1036 : #[cfg(test)]
1037 : mod tests {
1038 : use safekeeper_api::models::FullTransactionId;
1039 : use utils::id::{TenantId, TenantTimelineId, TimelineId};
1040 :
1041 : use super::*;
1042 :
1043 4 : fn mock_ttid() -> TenantTimelineId {
1044 4 : TenantTimelineId {
1045 4 : tenant_id: TenantId::from_slice(&[0x00; 16]).unwrap(),
1046 4 : timeline_id: TimelineId::from_slice(&[0x00; 16]).unwrap(),
1047 4 : }
1048 4 : }
1049 :
1050 4 : fn mock_addr() -> SocketAddr {
1051 4 : "127.0.0.1:8080".parse().unwrap()
1052 4 : }
1053 :
1054 : // add to wss specified feedback setting other fields to dummy values
1055 4 : fn push_feedback(wss: &mut WalSendersShared, feedback: ReplicationFeedback) {
1056 4 : let walsender_state = WalSenderState::Vanilla(VanillaWalSenderInternalState {
1057 4 : ttid: mock_ttid(),
1058 4 : addr: mock_addr(),
1059 4 : conn_id: 1,
1060 4 : appname: None,
1061 4 : feedback,
1062 4 : });
1063 4 : wss.slots.push(Some(walsender_state))
1064 4 : }
1065 :
1066 : // form standby feedback with given hot standby feedback ts/xmin and the
1067 : // rest set to dummy values.
1068 4 : fn hs_feedback(ts: TimestampTz, xmin: FullTransactionId) -> ReplicationFeedback {
1069 4 : ReplicationFeedback::Standby(StandbyFeedback {
1070 4 : reply: StandbyReply::empty(),
1071 4 : hs_feedback: HotStandbyFeedback {
1072 4 : ts,
1073 4 : xmin,
1074 4 : catalog_xmin: 0,
1075 4 : },
1076 4 : })
1077 4 : }
1078 :
1079 : // test that hs aggregation works as expected
1080 : #[test]
1081 1 : fn test_hs_feedback_no_valid() {
1082 1 : let mut wss = WalSendersShared::new();
1083 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
1084 1 : wss.update_reply_feedback();
1085 1 : assert_eq!(
1086 1 : wss.agg_standby_feedback.hs_feedback.xmin,
1087 1 : INVALID_FULL_TRANSACTION_ID
1088 1 : );
1089 1 : }
1090 :
1091 : #[test]
1092 1 : fn test_hs_feedback() {
1093 1 : let mut wss = WalSendersShared::new();
1094 1 : push_feedback(&mut wss, hs_feedback(1, INVALID_FULL_TRANSACTION_ID));
1095 1 : push_feedback(&mut wss, hs_feedback(1, 42));
1096 1 : push_feedback(&mut wss, hs_feedback(1, 64));
1097 1 : wss.update_reply_feedback();
1098 1 : assert_eq!(wss.agg_standby_feedback.hs_feedback.xmin, 42);
1099 1 : }
1100 : }
|
