Line data Source code
1 : //! Safekeeper communication endpoint to WAL proposer (compute node).
2 : //! Gets messages from the network, passes them down to consensus module and
3 : //! sends replies back.
4 :
5 : use std::future;
6 : use std::net::SocketAddr;
7 : use std::sync::Arc;
8 :
9 : use anyhow::{Context, anyhow};
10 : use bytes::BytesMut;
11 : use parking_lot::{MappedMutexGuard, Mutex, MutexGuard};
12 : use postgres_backend::{CopyStreamHandlerEnd, PostgresBackend, PostgresBackendReader, QueryError};
13 : use pq_proto::BeMessage;
14 : use safekeeper_api::ServerInfo;
15 : use safekeeper_api::membership::Configuration;
16 : use safekeeper_api::models::{ConnectionId, WalReceiverState, WalReceiverStatus};
17 : use tokio::io::{AsyncRead, AsyncWrite};
18 : use tokio::sync::mpsc::error::SendTimeoutError;
19 : use tokio::sync::mpsc::{Receiver, Sender, channel};
20 : use tokio::task;
21 : use tokio::task::JoinHandle;
22 : use tokio::time::{Duration, Instant, MissedTickBehavior};
23 : use tracing::*;
24 : use utils::id::TenantTimelineId;
25 : use utils::lsn::Lsn;
26 : use utils::pageserver_feedback::PageserverFeedback;
27 :
28 : use crate::GlobalTimelines;
29 : use crate::handler::SafekeeperPostgresHandler;
30 : use crate::metrics::{
31 : WAL_RECEIVER_QUEUE_DEPTH, WAL_RECEIVER_QUEUE_DEPTH_TOTAL, WAL_RECEIVER_QUEUE_SIZE_TOTAL,
32 : WAL_RECEIVERS,
33 : };
34 : use crate::safekeeper::{AcceptorProposerMessage, ProposerAcceptorMessage};
35 : use crate::timeline::WalResidentTimeline;
36 :
37 : const DEFAULT_FEEDBACK_CAPACITY: usize = 8;
38 :
39 : /// Registry of WalReceivers (compute connections). Timeline holds it (wrapped
40 : /// in Arc).
41 : pub struct WalReceivers {
42 : mutex: Mutex<WalReceiversShared>,
43 : pageserver_feedback_tx: tokio::sync::broadcast::Sender<PageserverFeedback>,
44 :
45 : num_computes_tx: tokio::sync::watch::Sender<usize>,
46 : num_computes_rx: tokio::sync::watch::Receiver<usize>,
47 : }
48 :
49 : /// Id under which walreceiver is registered in shmem.
50 : type WalReceiverId = usize;
51 :
52 : impl WalReceivers {
53 5 : pub fn new() -> Arc<WalReceivers> {
54 5 : let (pageserver_feedback_tx, _) =
55 5 : tokio::sync::broadcast::channel(DEFAULT_FEEDBACK_CAPACITY);
56 5 :
57 5 : let (num_computes_tx, num_computes_rx) = tokio::sync::watch::channel(0usize);
58 5 :
59 5 : Arc::new(WalReceivers {
60 5 : mutex: Mutex::new(WalReceiversShared { slots: Vec::new() }),
61 5 : pageserver_feedback_tx,
62 5 : num_computes_tx,
63 5 : num_computes_rx,
64 5 : })
65 5 : }
66 :
67 : /// Register new walreceiver. Returned guard provides access to the slot and
68 : /// automatically deregisters in Drop.
69 5 : pub fn register(self: &Arc<WalReceivers>, conn_id: Option<ConnectionId>) -> WalReceiverGuard {
70 5 : let mut shared = self.mutex.lock();
71 5 : let slots = &mut shared.slots;
72 5 : let walreceiver = WalReceiverState {
73 5 : conn_id,
74 5 : status: WalReceiverStatus::Voting,
75 5 : };
76 : // find empty slot or create new one
77 5 : let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
78 0 : slots[pos] = Some(walreceiver);
79 0 : pos
80 : } else {
81 5 : let pos = slots.len();
82 5 : slots.push(Some(walreceiver));
83 5 : pos
84 : };
85 :
86 5 : self.update_num(&shared);
87 5 : WAL_RECEIVERS.inc();
88 5 :
89 5 : WalReceiverGuard {
90 5 : id: pos,
91 5 : walreceivers: self.clone(),
92 5 : }
93 5 : }
94 :
95 : /// Get reference to locked slot contents. Slot must exist (registered
96 : /// earlier).
97 0 : fn get_slot<'a>(
98 0 : self: &'a Arc<WalReceivers>,
99 0 : id: WalReceiverId,
100 0 : ) -> MappedMutexGuard<'a, WalReceiverState> {
101 0 : MutexGuard::map(self.mutex.lock(), |locked| {
102 0 : locked.slots[id]
103 0 : .as_mut()
104 0 : .expect("walreceiver doesn't exist")
105 0 : })
106 0 : }
107 :
108 : /// Get number of walreceivers (compute connections).
109 0 : pub fn get_num(self: &Arc<WalReceivers>) -> usize {
110 0 : self.mutex.lock().get_num()
111 0 : }
112 :
113 : /// Get channel for number of walreceivers.
114 5 : pub fn get_num_rx(self: &Arc<WalReceivers>) -> tokio::sync::watch::Receiver<usize> {
115 5 : self.num_computes_rx.clone()
116 5 : }
117 :
118 : /// Should get called after every update of slots.
119 10 : fn update_num(self: &Arc<WalReceivers>, shared: &MutexGuard<WalReceiversShared>) {
120 10 : let num = shared.get_num();
121 10 : self.num_computes_tx.send_replace(num);
122 10 : }
123 :
124 : /// Get state of all walreceivers.
125 0 : pub fn get_all(self: &Arc<WalReceivers>) -> Vec<WalReceiverState> {
126 0 : self.mutex.lock().slots.iter().flatten().cloned().collect()
127 0 : }
128 :
129 : /// Get number of streaming walreceivers (normally 0 or 1) from compute.
130 5 : pub fn get_num_streaming(self: &Arc<WalReceivers>) -> usize {
131 5 : self.mutex
132 5 : .lock()
133 5 : .slots
134 5 : .iter()
135 5 : .flatten()
136 5 : // conn_id.is_none skips recovery which also registers here
137 5 : .filter(|s| s.conn_id.is_some() && matches!(s.status, WalReceiverStatus::Streaming))
138 5 : .count()
139 5 : }
140 :
141 : /// Unregister walreceiver.
142 5 : fn unregister(self: &Arc<WalReceivers>, id: WalReceiverId) {
143 5 : let mut shared = self.mutex.lock();
144 5 : shared.slots[id] = None;
145 5 : self.update_num(&shared);
146 5 : WAL_RECEIVERS.dec();
147 5 : }
148 :
149 : /// Broadcast pageserver feedback to connected walproposers.
150 0 : pub fn broadcast_pageserver_feedback(&self, feedback: PageserverFeedback) {
151 0 : // Err means there is no subscribers, it is fine.
152 0 : let _ = self.pageserver_feedback_tx.send(feedback);
153 0 : }
154 : }
155 :
156 : /// Only a few connections are expected (normally one), so store in Vec.
157 : struct WalReceiversShared {
158 : slots: Vec<Option<WalReceiverState>>,
159 : }
160 :
161 : impl WalReceiversShared {
162 : /// Get number of walreceivers (compute connections).
163 10 : fn get_num(&self) -> usize {
164 10 : self.slots.iter().flatten().count()
165 10 : }
166 : }
167 :
168 : /// Scope guard to access slot in WalReceivers registry and unregister from
169 : /// it in Drop.
170 : pub struct WalReceiverGuard {
171 : id: WalReceiverId,
172 : walreceivers: Arc<WalReceivers>,
173 : }
174 :
175 : impl WalReceiverGuard {
176 : /// Get reference to locked shared state contents.
177 0 : fn get(&self) -> MappedMutexGuard<WalReceiverState> {
178 0 : self.walreceivers.get_slot(self.id)
179 0 : }
180 : }
181 :
182 : impl Drop for WalReceiverGuard {
183 5 : fn drop(&mut self) {
184 5 : self.walreceivers.unregister(self.id);
185 5 : }
186 : }
187 :
188 : pub const MSG_QUEUE_SIZE: usize = 256;
189 : pub const REPLY_QUEUE_SIZE: usize = 16;
190 :
191 : impl SafekeeperPostgresHandler {
192 : /// Wrapper around handle_start_wal_push_guts handling result. Error is
193 : /// handled here while we're still in walreceiver ttid span; with API
194 : /// extension, this can probably be moved into postgres_backend.
195 0 : pub async fn handle_start_wal_push<IO: AsyncRead + AsyncWrite + Unpin>(
196 0 : &mut self,
197 0 : pgb: &mut PostgresBackend<IO>,
198 0 : proto_version: u32,
199 0 : allow_timeline_creation: bool,
200 0 : ) -> Result<(), QueryError> {
201 0 : let mut tli: Option<WalResidentTimeline> = None;
202 0 : if let Err(end) = self
203 0 : .handle_start_wal_push_guts(pgb, &mut tli, proto_version, allow_timeline_creation)
204 0 : .await
205 : {
206 : // Log the result and probably send it to the client, closing the stream.
207 0 : let handle_end_fut = pgb.handle_copy_stream_end(end);
208 : // If we managed to create the timeline, augment logging with current LSNs etc.
209 0 : if let Some(tli) = tli {
210 0 : let info = tli.get_safekeeper_info(&self.conf).await;
211 0 : handle_end_fut
212 0 : .instrument(info_span!("", term=%info.term, last_log_term=%info.last_log_term, flush_lsn=%Lsn(info.flush_lsn), commit_lsn=%Lsn(info.commit_lsn)))
213 0 : .await;
214 : } else {
215 0 : handle_end_fut.await;
216 : }
217 0 : }
218 0 : Ok(())
219 0 : }
220 :
221 0 : pub async fn handle_start_wal_push_guts<IO: AsyncRead + AsyncWrite + Unpin>(
222 0 : &mut self,
223 0 : pgb: &mut PostgresBackend<IO>,
224 0 : tli: &mut Option<WalResidentTimeline>,
225 0 : proto_version: u32,
226 0 : allow_timeline_creation: bool,
227 0 : ) -> Result<(), CopyStreamHandlerEnd> {
228 0 : // The `tli` parameter is only used for passing _out_ a timeline, one should
229 0 : // not have been passed in.
230 0 : assert!(tli.is_none());
231 :
232 : // Notify the libpq client that it's allowed to send `CopyData` messages
233 0 : pgb.write_message(&BeMessage::CopyBothResponse).await?;
234 :
235 : // Experiments [1] confirm that doing network IO in one (this) thread and
236 : // processing with disc IO in another significantly improves
237 : // performance; we spawn off WalAcceptor thread for message processing
238 : // to this end.
239 : //
240 : // [1] https://github.com/neondatabase/neon/pull/1318
241 0 : let (msg_tx, msg_rx) = channel(MSG_QUEUE_SIZE);
242 0 : let (reply_tx, reply_rx) = channel(REPLY_QUEUE_SIZE);
243 0 : let mut acceptor_handle: Option<JoinHandle<anyhow::Result<()>>> = None;
244 :
245 : // Concurrently receive and send data; replies are not synchronized with
246 : // sends, so this avoids deadlocks.
247 0 : let mut pgb_reader = pgb.split().context("START_WAL_PUSH split")?;
248 0 : let peer_addr = *pgb.get_peer_addr();
249 0 :
250 0 : let mut network_reader = NetworkReader {
251 0 : ttid: self.ttid,
252 0 : conn_id: self.conn_id,
253 0 : pgb_reader: &mut pgb_reader,
254 0 : peer_addr,
255 0 : proto_version,
256 0 : acceptor_handle: &mut acceptor_handle,
257 0 : global_timelines: self.global_timelines.clone(),
258 0 : };
259 :
260 : // Read first message and create timeline if needed and allowed. This
261 : // won't be when timelines will be always created by storcon and
262 : // allow_timeline_creation becomes false.
263 0 : let res = network_reader
264 0 : .read_first_message(allow_timeline_creation)
265 0 : .await;
266 :
267 0 : let network_res = if let Ok((timeline, next_msg)) = res {
268 0 : let pageserver_feedback_rx: tokio::sync::broadcast::Receiver<PageserverFeedback> =
269 0 : timeline
270 0 : .get_walreceivers()
271 0 : .pageserver_feedback_tx
272 0 : .subscribe();
273 0 : *tli = Some(timeline.wal_residence_guard().await?);
274 :
275 0 : let timeline_cancel = timeline.cancel.clone();
276 0 : tokio::select! {
277 : // todo: add read|write .context to these errors
278 0 : r = network_reader.run(msg_tx, msg_rx, reply_tx, timeline, next_msg) => r,
279 0 : r = network_write(pgb, reply_rx, pageserver_feedback_rx, proto_version) => r,
280 0 : _ = timeline_cancel.cancelled() => {
281 0 : return Err(CopyStreamHandlerEnd::Cancelled);
282 : }
283 : }
284 : } else {
285 0 : res.map(|_| ())
286 : };
287 :
288 : // Join pg backend back.
289 0 : pgb.unsplit(pgb_reader)?;
290 :
291 : // Join the spawned WalAcceptor. At this point chans to/from it passed
292 : // to network routines are dropped, so it will exit as soon as it
293 : // touches them.
294 0 : match acceptor_handle {
295 : None => {
296 : // failed even before spawning; read_network should have error
297 0 : Err(network_res.expect_err("no error with WalAcceptor not spawn"))
298 : }
299 0 : Some(handle) => {
300 0 : let wal_acceptor_res = handle.await;
301 :
302 : // If there was any network error, return it.
303 0 : network_res?;
304 :
305 : // Otherwise, WalAcceptor thread must have errored.
306 0 : match wal_acceptor_res {
307 0 : Ok(Ok(_)) => Ok(()), // Clean shutdown
308 0 : Ok(Err(e)) => Err(CopyStreamHandlerEnd::Other(e.context("WAL acceptor"))),
309 0 : Err(_) => Err(CopyStreamHandlerEnd::Other(anyhow!(
310 0 : "WalAcceptor task panicked",
311 0 : ))),
312 : }
313 : }
314 : }
315 0 : }
316 : }
317 :
318 : struct NetworkReader<'a, IO> {
319 : ttid: TenantTimelineId,
320 : conn_id: ConnectionId,
321 : pgb_reader: &'a mut PostgresBackendReader<IO>,
322 : peer_addr: SocketAddr,
323 : proto_version: u32,
324 : // WalAcceptor is spawned when we learn server info from walproposer and
325 : // create timeline; handle is put here.
326 : acceptor_handle: &'a mut Option<JoinHandle<anyhow::Result<()>>>,
327 : global_timelines: Arc<GlobalTimelines>,
328 : }
329 :
330 : impl<IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'_, IO> {
331 0 : async fn read_first_message(
332 0 : &mut self,
333 0 : allow_timeline_creation: bool,
334 0 : ) -> Result<(WalResidentTimeline, ProposerAcceptorMessage), CopyStreamHandlerEnd> {
335 : // Receive information about server to create timeline, if not yet.
336 0 : let next_msg = read_message(self.pgb_reader, self.proto_version).await?;
337 0 : let tli = match next_msg {
338 0 : ProposerAcceptorMessage::Greeting(ref greeting) => {
339 0 : info!(
340 0 : "start handshake with walproposer {} sysid {}",
341 : self.peer_addr, greeting.system_id,
342 : );
343 0 : let server_info = ServerInfo {
344 0 : pg_version: greeting.pg_version,
345 0 : system_id: greeting.system_id,
346 0 : wal_seg_size: greeting.wal_seg_size,
347 0 : };
348 0 : let tli = if allow_timeline_creation {
349 0 : self.global_timelines
350 0 : .create(
351 0 : self.ttid,
352 0 : Configuration::empty(),
353 0 : server_info,
354 0 : Lsn::INVALID,
355 0 : Lsn::INVALID,
356 0 : )
357 0 : .await
358 0 : .context("create timeline")?
359 : } else {
360 0 : self.global_timelines
361 0 : .get(self.ttid)
362 0 : .context("get timeline")?
363 : };
364 0 : tli.wal_residence_guard().await?
365 : }
366 : _ => {
367 0 : return Err(CopyStreamHandlerEnd::Other(anyhow::anyhow!(
368 0 : "unexpected message {next_msg:?} instead of greeting"
369 0 : )));
370 : }
371 : };
372 0 : Ok((tli, next_msg))
373 0 : }
374 :
375 : /// This function is cancellation-safe (only does network I/O and channel read/writes).
376 0 : async fn run(
377 0 : self,
378 0 : msg_tx: Sender<ProposerAcceptorMessage>,
379 0 : msg_rx: Receiver<ProposerAcceptorMessage>,
380 0 : reply_tx: Sender<AcceptorProposerMessage>,
381 0 : tli: WalResidentTimeline,
382 0 : next_msg: ProposerAcceptorMessage,
383 0 : ) -> Result<(), CopyStreamHandlerEnd> {
384 0 : *self.acceptor_handle = Some(WalAcceptor::spawn(
385 0 : tli,
386 0 : msg_rx,
387 0 : reply_tx,
388 0 : Some(self.conn_id),
389 0 : ));
390 0 :
391 0 : // Forward all messages to WalAcceptor
392 0 : read_network_loop(self.pgb_reader, msg_tx, next_msg, self.proto_version).await
393 0 : }
394 : }
395 :
396 : /// Read next message from walproposer.
397 : /// TODO: Return Ok(None) on graceful termination.
398 0 : async fn read_message<IO: AsyncRead + AsyncWrite + Unpin>(
399 0 : pgb_reader: &mut PostgresBackendReader<IO>,
400 0 : proto_version: u32,
401 0 : ) -> Result<ProposerAcceptorMessage, CopyStreamHandlerEnd> {
402 0 : let copy_data = pgb_reader.read_copy_message().await?;
403 0 : let msg = ProposerAcceptorMessage::parse(copy_data, proto_version)?;
404 0 : Ok(msg)
405 0 : }
406 :
407 0 : async fn read_network_loop<IO: AsyncRead + AsyncWrite + Unpin>(
408 0 : pgb_reader: &mut PostgresBackendReader<IO>,
409 0 : msg_tx: Sender<ProposerAcceptorMessage>,
410 0 : mut next_msg: ProposerAcceptorMessage,
411 0 : proto_version: u32,
412 0 : ) -> Result<(), CopyStreamHandlerEnd> {
413 : /// Threshold for logging slow WalAcceptor sends.
414 : const SLOW_THRESHOLD: Duration = Duration::from_secs(5);
415 :
416 : loop {
417 0 : let started = Instant::now();
418 0 : let size = next_msg.size();
419 0 :
420 0 : match msg_tx.send_timeout(next_msg, SLOW_THRESHOLD).await {
421 0 : Ok(()) => {}
422 : // Slow send, log a message and keep trying. Log context has timeline ID.
423 0 : Err(SendTimeoutError::Timeout(next_msg)) => {
424 0 : warn!(
425 0 : "slow WalAcceptor send blocked for {:.3}s",
426 0 : Instant::now().duration_since(started).as_secs_f64()
427 : );
428 0 : if msg_tx.send(next_msg).await.is_err() {
429 0 : return Ok(()); // WalAcceptor terminated
430 0 : }
431 0 : warn!(
432 0 : "slow WalAcceptor send completed after {:.3}s",
433 0 : Instant::now().duration_since(started).as_secs_f64()
434 : )
435 : }
436 : // WalAcceptor terminated.
437 0 : Err(SendTimeoutError::Closed(_)) => return Ok(()),
438 : }
439 :
440 : // Update metrics. Will be decremented in WalAcceptor.
441 0 : WAL_RECEIVER_QUEUE_DEPTH_TOTAL.inc();
442 0 : WAL_RECEIVER_QUEUE_SIZE_TOTAL.add(size as i64);
443 :
444 0 : next_msg = read_message(pgb_reader, proto_version).await?;
445 : }
446 0 : }
447 :
448 : /// Read replies from WalAcceptor and pass them back to socket. Returns Ok(())
449 : /// if reply_rx closed; it must mean WalAcceptor terminated, joining it should
450 : /// tell the error.
451 : ///
452 : /// This function is cancellation-safe (only does network I/O and channel read/writes).
453 0 : async fn network_write<IO: AsyncRead + AsyncWrite + Unpin>(
454 0 : pgb_writer: &mut PostgresBackend<IO>,
455 0 : mut reply_rx: Receiver<AcceptorProposerMessage>,
456 0 : mut pageserver_feedback_rx: tokio::sync::broadcast::Receiver<PageserverFeedback>,
457 0 : proto_version: u32,
458 0 : ) -> Result<(), CopyStreamHandlerEnd> {
459 0 : let mut buf = BytesMut::with_capacity(128);
460 0 :
461 0 : // storing append_response to inject PageserverFeedback into it
462 0 : let mut last_append_response = None;
463 :
464 : loop {
465 : // trying to read either AcceptorProposerMessage or PageserverFeedback
466 0 : let msg = tokio::select! {
467 0 : reply = reply_rx.recv() => {
468 0 : if let Some(msg) = reply {
469 0 : if let AcceptorProposerMessage::AppendResponse(append_response) = &msg {
470 0 : last_append_response = Some(append_response.clone());
471 0 : }
472 0 : Some(msg)
473 : } else {
474 0 : return Ok(()); // chan closed, WalAcceptor terminated
475 : }
476 : }
477 :
478 0 : feedback = pageserver_feedback_rx.recv() =>
479 0 : match (feedback, &last_append_response) {
480 0 : (Ok(feedback), Some(append_response)) => {
481 0 : // clone AppendResponse and inject PageserverFeedback into it
482 0 : let mut append_response = append_response.clone();
483 0 : append_response.pageserver_feedback = Some(feedback);
484 0 : Some(AcceptorProposerMessage::AppendResponse(append_response))
485 : }
486 0 : _ => None,
487 : },
488 : };
489 :
490 0 : let Some(msg) = msg else {
491 0 : continue;
492 : };
493 :
494 0 : buf.clear();
495 0 : msg.serialize(&mut buf, proto_version)?;
496 0 : pgb_writer.write_message(&BeMessage::CopyData(&buf)).await?;
497 : }
498 0 : }
499 :
500 : /// The WAL flush interval. This ensures we periodically flush the WAL and send AppendResponses to
501 : /// walproposer, even when it's writing a steady stream of messages.
502 : const FLUSH_INTERVAL: Duration = Duration::from_secs(1);
503 :
504 : /// The metrics computation interval.
505 : ///
506 : /// The Prometheus poll interval is 60 seconds at the time of writing. We sample the queue depth
507 : /// every 5 seconds, for 12 samples per poll. This will give a count of up to 12x active timelines.
508 : const METRICS_INTERVAL: Duration = Duration::from_secs(5);
509 :
510 : /// Encapsulates a task which takes messages from msg_rx, processes and pushes
511 : /// replies to reply_tx.
512 : ///
513 : /// Reading from socket and writing to disk in parallel is beneficial for
514 : /// performance, this struct provides the writing to disk part.
515 : pub struct WalAcceptor {
516 : tli: WalResidentTimeline,
517 : msg_rx: Receiver<ProposerAcceptorMessage>,
518 : reply_tx: Sender<AcceptorProposerMessage>,
519 : conn_id: Option<ConnectionId>,
520 : }
521 :
522 : impl WalAcceptor {
523 : /// Spawn task with WalAcceptor running, return handle to it. Task returns
524 : /// Ok(()) if either of channels has closed, and Err if any error during
525 : /// message processing is encountered.
526 : ///
527 : /// conn_id None means WalAcceptor is used by recovery initiated at this safekeeper.
528 5 : pub fn spawn(
529 5 : tli: WalResidentTimeline,
530 5 : msg_rx: Receiver<ProposerAcceptorMessage>,
531 5 : reply_tx: Sender<AcceptorProposerMessage>,
532 5 : conn_id: Option<ConnectionId>,
533 5 : ) -> JoinHandle<anyhow::Result<()>> {
534 5 : task::spawn(async move {
535 5 : let mut wa = WalAcceptor {
536 5 : tli,
537 5 : msg_rx,
538 5 : reply_tx,
539 5 : conn_id,
540 5 : };
541 5 :
542 5 : let span_ttid = wa.tli.ttid; // satisfy borrow checker
543 5 : wa.run()
544 5 : .instrument(
545 5 : info_span!("WAL acceptor", cid = %conn_id.unwrap_or(0), ttid = %span_ttid),
546 : )
547 5 : .await
548 5 : })
549 5 : }
550 :
551 : /// The main loop. Returns Ok(()) if either msg_rx or reply_tx got closed;
552 : /// it must mean that network thread terminated.
553 : ///
554 : /// This function is *not* cancellation safe, it does local disk I/O: it should always
555 : /// be allowed to run to completion. It respects Timeline::cancel and shuts down cleanly
556 : /// when that gets triggered.
557 5 : async fn run(&mut self) -> anyhow::Result<()> {
558 5 : let walreceiver_guard = self.tli.get_walreceivers().register(self.conn_id);
559 5 :
560 5 : // Periodically flush the WAL and compute metrics.
561 5 : let mut flush_ticker = tokio::time::interval(FLUSH_INTERVAL);
562 5 : flush_ticker.set_missed_tick_behavior(MissedTickBehavior::Delay);
563 5 : flush_ticker.tick().await; // skip the initial, immediate tick
564 :
565 5 : let mut metrics_ticker = tokio::time::interval(METRICS_INTERVAL);
566 5 : metrics_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
567 5 :
568 5 : // Tracks whether we have unflushed appends.
569 5 : let mut dirty = false;
570 :
571 1250 : while !self.tli.is_cancelled() {
572 1250 : let reply = tokio::select! {
573 : // Process inbound message.
574 1250 : msg = self.msg_rx.recv() => {
575 : // If disconnected, break to flush WAL and return.
576 624 : let Some(mut msg) = msg else {
577 4 : break;
578 : };
579 :
580 : // Update gauge metrics.
581 620 : WAL_RECEIVER_QUEUE_DEPTH_TOTAL.dec();
582 620 : WAL_RECEIVER_QUEUE_SIZE_TOTAL.sub(msg.size() as i64);
583 620 :
584 620 : // Update walreceiver state in shmem for reporting.
585 620 : if let ProposerAcceptorMessage::Elected(_) = &msg {
586 0 : walreceiver_guard.get().status = WalReceiverStatus::Streaming;
587 620 : }
588 :
589 : // Don't flush the WAL on every append, only periodically via flush_ticker.
590 : // This batches multiple appends per fsync. If the channel is empty after
591 : // sending the reply, we'll schedule an immediate flush.
592 : //
593 : // Note that a flush can still happen on segment bounds, which will result
594 : // in an AppendResponse.
595 620 : if let ProposerAcceptorMessage::AppendRequest(append_request) = msg {
596 620 : msg = ProposerAcceptorMessage::NoFlushAppendRequest(append_request);
597 620 : dirty = true;
598 620 : }
599 :
600 620 : self.tli.process_msg(&msg).await?
601 : }
602 :
603 : // While receiving AppendRequests, flush the WAL periodically and respond with an
604 : // AppendResponse to let walproposer know we're still alive.
605 1250 : _ = flush_ticker.tick(), if dirty => {
606 0 : dirty = false;
607 0 : self.tli
608 0 : .process_msg(&ProposerAcceptorMessage::FlushWAL)
609 0 : .await?
610 : }
611 :
612 : // If there are no pending messages, flush the WAL immediately.
613 : //
614 : // TODO: this should be done via flush_ticker.reset_immediately(), but that's always
615 : // delayed by 1ms due to this bug: https://github.com/tokio-rs/tokio/issues/6866.
616 1250 : _ = future::ready(()), if dirty && self.msg_rx.is_empty() => {
617 620 : dirty = false;
618 620 : flush_ticker.reset();
619 620 : self.tli
620 620 : .process_msg(&ProposerAcceptorMessage::FlushWAL)
621 620 : .await?
622 : }
623 :
624 : // Update histogram metrics periodically.
625 1250 : _ = metrics_ticker.tick() => {
626 5 : WAL_RECEIVER_QUEUE_DEPTH.observe(self.msg_rx.len() as f64);
627 5 : None // no reply
628 : }
629 :
630 1250 : _ = self.tli.cancel.cancelled() => {
631 0 : break;
632 : }
633 : };
634 :
635 : // Send reply, if any.
636 1245 : if let Some(reply) = reply {
637 620 : if self.reply_tx.send(reply).await.is_err() {
638 0 : break; // disconnected, break to flush WAL and return
639 620 : }
640 625 : }
641 : }
642 :
643 : // Flush WAL on disconnect, see https://github.com/neondatabase/neon/issues/9259.
644 4 : if dirty && !self.tli.cancel.is_cancelled() {
645 0 : self.tli
646 0 : .process_msg(&ProposerAcceptorMessage::FlushWAL)
647 0 : .await?;
648 4 : }
649 :
650 4 : Ok(())
651 4 : }
652 : }
653 :
654 : /// On drop, drain msg_rx and update metrics to avoid leaks.
655 : impl Drop for WalAcceptor {
656 5 : fn drop(&mut self) {
657 5 : self.msg_rx.close(); // prevent further sends
658 5 : while let Ok(msg) = self.msg_rx.try_recv() {
659 0 : WAL_RECEIVER_QUEUE_DEPTH_TOTAL.dec();
660 0 : WAL_RECEIVER_QUEUE_SIZE_TOTAL.sub(msg.size() as i64);
661 0 : }
662 5 : }
663 : }
|
