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 crate::handler::SafekeeperPostgresHandler;
6 : use crate::safekeeper::AcceptorProposerMessage;
7 : use crate::safekeeper::ProposerAcceptorMessage;
8 : use crate::safekeeper::ServerInfo;
9 : use crate::timeline::Timeline;
10 : use crate::wal_service::ConnectionId;
11 : use crate::GlobalTimelines;
12 : use anyhow::{anyhow, Context};
13 : use bytes::BytesMut;
14 : use parking_lot::MappedMutexGuard;
15 : use parking_lot::Mutex;
16 : use parking_lot::MutexGuard;
17 : use postgres_backend::CopyStreamHandlerEnd;
18 : use postgres_backend::PostgresBackend;
19 : use postgres_backend::PostgresBackendReader;
20 : use postgres_backend::QueryError;
21 : use pq_proto::BeMessage;
22 : use serde::Deserialize;
23 : use serde::Serialize;
24 : use std::net::SocketAddr;
25 : use std::sync::Arc;
26 : use tokio::io::AsyncRead;
27 : use tokio::io::AsyncWrite;
28 : use tokio::sync::mpsc::channel;
29 : use tokio::sync::mpsc::error::TryRecvError;
30 : use tokio::sync::mpsc::Receiver;
31 : use tokio::sync::mpsc::Sender;
32 : use tokio::task;
33 : use tokio::task::JoinHandle;
34 : use tokio::time::Duration;
35 : use tokio::time::Instant;
36 : use tracing::*;
37 : use utils::id::TenantTimelineId;
38 : use utils::lsn::Lsn;
39 : use utils::pageserver_feedback::PageserverFeedback;
40 :
41 : const DEFAULT_FEEDBACK_CAPACITY: usize = 8;
42 :
43 : /// Registry of WalReceivers (compute connections). Timeline holds it (wrapped
44 : /// in Arc).
45 : pub struct WalReceivers {
46 : mutex: Mutex<WalReceiversShared>,
47 : pageserver_feedback_tx: tokio::sync::broadcast::Sender<PageserverFeedback>,
48 : }
49 :
50 : /// Id under which walreceiver is registered in shmem.
51 : type WalReceiverId = usize;
52 :
53 : impl WalReceivers {
54 0 : pub fn new() -> Arc<WalReceivers> {
55 0 : let (pageserver_feedback_tx, _) =
56 0 : tokio::sync::broadcast::channel(DEFAULT_FEEDBACK_CAPACITY);
57 0 :
58 0 : Arc::new(WalReceivers {
59 0 : mutex: Mutex::new(WalReceiversShared { slots: Vec::new() }),
60 0 : pageserver_feedback_tx,
61 0 : })
62 0 : }
63 :
64 : /// Register new walreceiver. Returned guard provides access to the slot and
65 : /// automatically deregisters in Drop.
66 0 : pub fn register(self: &Arc<WalReceivers>, conn_id: Option<ConnectionId>) -> WalReceiverGuard {
67 0 : let slots = &mut self.mutex.lock().slots;
68 0 : let walreceiver = WalReceiverState {
69 0 : conn_id,
70 0 : status: WalReceiverStatus::Voting,
71 0 : };
72 : // find empty slot or create new one
73 0 : let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
74 0 : slots[pos] = Some(walreceiver);
75 0 : pos
76 : } else {
77 0 : let pos = slots.len();
78 0 : slots.push(Some(walreceiver));
79 0 : pos
80 : };
81 0 : WalReceiverGuard {
82 0 : id: pos,
83 0 : walreceivers: self.clone(),
84 0 : }
85 0 : }
86 :
87 : /// Get reference to locked slot contents. Slot must exist (registered
88 : /// earlier).
89 0 : fn get_slot<'a>(
90 0 : self: &'a Arc<WalReceivers>,
91 0 : id: WalReceiverId,
92 0 : ) -> MappedMutexGuard<'a, WalReceiverState> {
93 0 : MutexGuard::map(self.mutex.lock(), |locked| {
94 0 : locked.slots[id]
95 0 : .as_mut()
96 0 : .expect("walreceiver doesn't exist")
97 0 : })
98 0 : }
99 :
100 : /// Get number of walreceivers (compute connections).
101 0 : pub fn get_num(self: &Arc<WalReceivers>) -> usize {
102 0 : self.mutex.lock().slots.iter().flatten().count()
103 0 : }
104 :
105 : /// Get state of all walreceivers.
106 0 : pub fn get_all(self: &Arc<WalReceivers>) -> Vec<WalReceiverState> {
107 0 : self.mutex.lock().slots.iter().flatten().cloned().collect()
108 0 : }
109 :
110 : /// Get number of streaming walreceivers (normally 0 or 1) from compute.
111 0 : pub fn get_num_streaming(self: &Arc<WalReceivers>) -> usize {
112 0 : self.mutex
113 0 : .lock()
114 0 : .slots
115 0 : .iter()
116 0 : .flatten()
117 0 : // conn_id.is_none skips recovery which also registers here
118 0 : .filter(|s| s.conn_id.is_some() && matches!(s.status, WalReceiverStatus::Streaming))
119 0 : .count()
120 0 : }
121 :
122 : /// Unregister walreceiver.
123 0 : fn unregister(self: &Arc<WalReceivers>, id: WalReceiverId) {
124 0 : let mut shared = self.mutex.lock();
125 0 : shared.slots[id] = None;
126 0 : }
127 :
128 : /// Broadcast pageserver feedback to connected walproposers.
129 0 : pub fn broadcast_pageserver_feedback(&self, feedback: PageserverFeedback) {
130 0 : // Err means there is no subscribers, it is fine.
131 0 : let _ = self.pageserver_feedback_tx.send(feedback);
132 0 : }
133 : }
134 :
135 : /// Only a few connections are expected (normally one), so store in Vec.
136 : struct WalReceiversShared {
137 : slots: Vec<Option<WalReceiverState>>,
138 : }
139 :
140 0 : #[derive(Debug, Clone, Serialize, Deserialize)]
141 : pub struct WalReceiverState {
142 : /// None means it is recovery initiated by us (this safekeeper).
143 : pub conn_id: Option<ConnectionId>,
144 : pub status: WalReceiverStatus,
145 : }
146 :
147 : /// Walreceiver status. Currently only whether it passed voting stage and
148 : /// started receiving the stream, but it is easy to add more if needed.
149 0 : #[derive(Debug, Clone, Serialize, Deserialize)]
150 : pub enum WalReceiverStatus {
151 : Voting,
152 : Streaming,
153 : }
154 :
155 : /// Scope guard to access slot in WalReceivers registry and unregister from
156 : /// it in Drop.
157 : pub struct WalReceiverGuard {
158 : id: WalReceiverId,
159 : walreceivers: Arc<WalReceivers>,
160 : }
161 :
162 : impl WalReceiverGuard {
163 : /// Get reference to locked shared state contents.
164 0 : fn get(&self) -> MappedMutexGuard<WalReceiverState> {
165 0 : self.walreceivers.get_slot(self.id)
166 0 : }
167 : }
168 :
169 : impl Drop for WalReceiverGuard {
170 0 : fn drop(&mut self) {
171 0 : self.walreceivers.unregister(self.id);
172 0 : }
173 : }
174 :
175 : pub const MSG_QUEUE_SIZE: usize = 256;
176 : pub const REPLY_QUEUE_SIZE: usize = 16;
177 :
178 : impl SafekeeperPostgresHandler {
179 : /// Wrapper around handle_start_wal_push_guts handling result. Error is
180 : /// handled here while we're still in walreceiver ttid span; with API
181 : /// extension, this can probably be moved into postgres_backend.
182 0 : pub async fn handle_start_wal_push<IO: AsyncRead + AsyncWrite + Unpin>(
183 0 : &mut self,
184 0 : pgb: &mut PostgresBackend<IO>,
185 0 : ) -> Result<(), QueryError> {
186 0 : if let Err(end) = self.handle_start_wal_push_guts(pgb).await {
187 : // Log the result and probably send it to the client, closing the stream.
188 0 : pgb.handle_copy_stream_end(end).await;
189 0 : }
190 0 : Ok(())
191 0 : }
192 :
193 0 : pub async fn handle_start_wal_push_guts<IO: AsyncRead + AsyncWrite + Unpin>(
194 0 : &mut self,
195 0 : pgb: &mut PostgresBackend<IO>,
196 0 : ) -> Result<(), CopyStreamHandlerEnd> {
197 0 : // Notify the libpq client that it's allowed to send `CopyData` messages
198 0 : pgb.write_message(&BeMessage::CopyBothResponse).await?;
199 :
200 : // Experiments [1] confirm that doing network IO in one (this) thread and
201 : // processing with disc IO in another significantly improves
202 : // performance; we spawn off WalAcceptor thread for message processing
203 : // to this end.
204 : //
205 : // [1] https://github.com/neondatabase/neon/pull/1318
206 0 : let (msg_tx, msg_rx) = channel(MSG_QUEUE_SIZE);
207 0 : let (reply_tx, reply_rx) = channel(REPLY_QUEUE_SIZE);
208 0 : let mut acceptor_handle: Option<JoinHandle<anyhow::Result<()>>> = None;
209 :
210 : // Concurrently receive and send data; replies are not synchronized with
211 : // sends, so this avoids deadlocks.
212 0 : let mut pgb_reader = pgb.split().context("START_WAL_PUSH split")?;
213 0 : let peer_addr = *pgb.get_peer_addr();
214 0 : let mut network_reader = NetworkReader {
215 0 : ttid: self.ttid,
216 0 : conn_id: self.conn_id,
217 0 : pgb_reader: &mut pgb_reader,
218 0 : peer_addr,
219 0 : acceptor_handle: &mut acceptor_handle,
220 0 : };
221 :
222 : // Read first message and create timeline if needed.
223 0 : let res = network_reader.read_first_message().await;
224 :
225 0 : let res = if let Ok((tli, next_msg)) = res {
226 0 : let pageserver_feedback_rx: tokio::sync::broadcast::Receiver<PageserverFeedback> =
227 0 : tli.get_walreceivers().pageserver_feedback_tx.subscribe();
228 :
229 : tokio::select! {
230 : // todo: add read|write .context to these errors
231 : r = network_reader.run(msg_tx, msg_rx, reply_tx, tli.clone(), next_msg) => r,
232 : r = network_write(pgb, reply_rx, pageserver_feedback_rx) => r,
233 : }
234 : } else {
235 0 : res.map(|_| ())
236 : };
237 :
238 : // Join pg backend back.
239 0 : pgb.unsplit(pgb_reader)?;
240 :
241 : // Join the spawned WalAcceptor. At this point chans to/from it passed
242 : // to network routines are dropped, so it will exit as soon as it
243 : // touches them.
244 0 : match acceptor_handle {
245 : None => {
246 : // failed even before spawning; read_network should have error
247 0 : Err(res.expect_err("no error with WalAcceptor not spawn"))
248 : }
249 0 : Some(handle) => {
250 0 : let wal_acceptor_res = handle.await;
251 :
252 : // If there was any network error, return it.
253 0 : res?;
254 :
255 : // Otherwise, WalAcceptor thread must have errored.
256 0 : match wal_acceptor_res {
257 0 : Ok(Ok(_)) => Ok(()), // can't happen currently; would be if we add graceful termination
258 0 : Ok(Err(e)) => Err(CopyStreamHandlerEnd::Other(e.context("WAL acceptor"))),
259 0 : Err(_) => Err(CopyStreamHandlerEnd::Other(anyhow!(
260 0 : "WalAcceptor task panicked",
261 0 : ))),
262 : }
263 : }
264 : }
265 0 : }
266 : }
267 :
268 : struct NetworkReader<'a, IO> {
269 : ttid: TenantTimelineId,
270 : conn_id: ConnectionId,
271 : pgb_reader: &'a mut PostgresBackendReader<IO>,
272 : peer_addr: SocketAddr,
273 : // WalAcceptor is spawned when we learn server info from walproposer and
274 : // create timeline; handle is put here.
275 : acceptor_handle: &'a mut Option<JoinHandle<anyhow::Result<()>>>,
276 : }
277 :
278 : impl<'a, IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'a, IO> {
279 0 : async fn read_first_message(
280 0 : &mut self,
281 0 : ) -> Result<(Arc<Timeline>, ProposerAcceptorMessage), CopyStreamHandlerEnd> {
282 : // Receive information about server to create timeline, if not yet.
283 0 : let next_msg = read_message(self.pgb_reader).await?;
284 0 : let tli = match next_msg {
285 0 : ProposerAcceptorMessage::Greeting(ref greeting) => {
286 0 : info!(
287 0 : "start handshake with walproposer {} sysid {} timeline {}",
288 : self.peer_addr, greeting.system_id, greeting.tli,
289 : );
290 0 : let server_info = ServerInfo {
291 0 : pg_version: greeting.pg_version,
292 0 : system_id: greeting.system_id,
293 0 : wal_seg_size: greeting.wal_seg_size,
294 0 : };
295 0 : GlobalTimelines::create(self.ttid, server_info, Lsn::INVALID, Lsn::INVALID).await?
296 : }
297 : _ => {
298 0 : return Err(CopyStreamHandlerEnd::Other(anyhow::anyhow!(
299 0 : "unexpected message {next_msg:?} instead of greeting"
300 0 : )))
301 : }
302 : };
303 0 : Ok((tli, next_msg))
304 0 : }
305 :
306 0 : async fn run(
307 0 : self,
308 0 : msg_tx: Sender<ProposerAcceptorMessage>,
309 0 : msg_rx: Receiver<ProposerAcceptorMessage>,
310 0 : reply_tx: Sender<AcceptorProposerMessage>,
311 0 : tli: Arc<Timeline>,
312 0 : next_msg: ProposerAcceptorMessage,
313 0 : ) -> Result<(), CopyStreamHandlerEnd> {
314 0 : *self.acceptor_handle = Some(WalAcceptor::spawn(
315 0 : tli,
316 0 : msg_rx,
317 0 : reply_tx,
318 0 : Some(self.conn_id),
319 0 : ));
320 0 :
321 0 : // Forward all messages to WalAcceptor
322 0 : read_network_loop(self.pgb_reader, msg_tx, next_msg).await
323 0 : }
324 : }
325 :
326 : /// Read next message from walproposer.
327 : /// TODO: Return Ok(None) on graceful termination.
328 0 : async fn read_message<IO: AsyncRead + AsyncWrite + Unpin>(
329 0 : pgb_reader: &mut PostgresBackendReader<IO>,
330 0 : ) -> Result<ProposerAcceptorMessage, CopyStreamHandlerEnd> {
331 0 : let copy_data = pgb_reader.read_copy_message().await?;
332 0 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
333 0 : Ok(msg)
334 0 : }
335 :
336 0 : async fn read_network_loop<IO: AsyncRead + AsyncWrite + Unpin>(
337 0 : pgb_reader: &mut PostgresBackendReader<IO>,
338 0 : msg_tx: Sender<ProposerAcceptorMessage>,
339 0 : mut next_msg: ProposerAcceptorMessage,
340 0 : ) -> Result<(), CopyStreamHandlerEnd> {
341 : loop {
342 0 : if msg_tx.send(next_msg).await.is_err() {
343 0 : return Ok(()); // chan closed, WalAcceptor terminated
344 0 : }
345 0 : next_msg = read_message(pgb_reader).await?;
346 : }
347 0 : }
348 :
349 : /// Read replies from WalAcceptor and pass them back to socket. Returns Ok(())
350 : /// if reply_rx closed; it must mean WalAcceptor terminated, joining it should
351 : /// tell the error.
352 0 : async fn network_write<IO: AsyncRead + AsyncWrite + Unpin>(
353 0 : pgb_writer: &mut PostgresBackend<IO>,
354 0 : mut reply_rx: Receiver<AcceptorProposerMessage>,
355 0 : mut pageserver_feedback_rx: tokio::sync::broadcast::Receiver<PageserverFeedback>,
356 0 : ) -> Result<(), CopyStreamHandlerEnd> {
357 0 : let mut buf = BytesMut::with_capacity(128);
358 0 :
359 0 : // storing append_response to inject PageserverFeedback into it
360 0 : let mut last_append_response = None;
361 :
362 : loop {
363 : // trying to read either AcceptorProposerMessage or PageserverFeedback
364 0 : let msg = tokio::select! {
365 : reply = reply_rx.recv() => {
366 : if let Some(msg) = reply {
367 : if let AcceptorProposerMessage::AppendResponse(append_response) = &msg {
368 : last_append_response = Some(append_response.clone());
369 : }
370 : Some(msg)
371 : } else {
372 : return Ok(()); // chan closed, WalAcceptor terminated
373 : }
374 : }
375 :
376 : feedback = pageserver_feedback_rx.recv() =>
377 : match (feedback, &last_append_response) {
378 : (Ok(feedback), Some(append_response)) => {
379 : // clone AppendResponse and inject PageserverFeedback into it
380 : let mut append_response = append_response.clone();
381 : append_response.pageserver_feedback = Some(feedback);
382 : Some(AcceptorProposerMessage::AppendResponse(append_response))
383 : }
384 : _ => None,
385 : }
386 : };
387 :
388 0 : let Some(msg) = msg else {
389 0 : continue;
390 : };
391 :
392 0 : buf.clear();
393 0 : msg.serialize(&mut buf)?;
394 0 : pgb_writer.write_message(&BeMessage::CopyData(&buf)).await?;
395 : }
396 0 : }
397 :
398 : // Send keepalive messages to walproposer, to make sure it receives updates
399 : // even when it writes a steady stream of messages.
400 : const KEEPALIVE_INTERVAL: Duration = Duration::from_secs(1);
401 :
402 : /// Encapsulates a task which takes messages from msg_rx, processes and pushes
403 : /// replies to reply_tx; reading from socket and writing to disk in parallel is
404 : /// beneficial for performance, this struct provides writing to disk part.
405 : pub struct WalAcceptor {
406 : tli: Arc<Timeline>,
407 : msg_rx: Receiver<ProposerAcceptorMessage>,
408 : reply_tx: Sender<AcceptorProposerMessage>,
409 : conn_id: Option<ConnectionId>,
410 : }
411 :
412 : impl WalAcceptor {
413 : /// Spawn task with WalAcceptor running, return handle to it. Task returns
414 : /// Ok(()) if either of channels has closed, and Err if any error during
415 : /// message processing is encountered.
416 : ///
417 : /// conn_id None means WalAcceptor is used by recovery initiated at this safekeeper.
418 0 : pub fn spawn(
419 0 : tli: Arc<Timeline>,
420 0 : msg_rx: Receiver<ProposerAcceptorMessage>,
421 0 : reply_tx: Sender<AcceptorProposerMessage>,
422 0 : conn_id: Option<ConnectionId>,
423 0 : ) -> JoinHandle<anyhow::Result<()>> {
424 0 : task::spawn(async move {
425 0 : let mut wa = WalAcceptor {
426 0 : tli,
427 0 : msg_rx,
428 0 : reply_tx,
429 0 : conn_id,
430 0 : };
431 0 :
432 0 : let span_ttid = wa.tli.ttid; // satisfy borrow checker
433 0 : wa.run()
434 0 : .instrument(
435 0 : info_span!("WAL acceptor", cid = %conn_id.unwrap_or(0), ttid = %span_ttid),
436 : )
437 0 : .await
438 0 : })
439 0 : }
440 :
441 : /// The main loop. Returns Ok(()) if either msg_rx or reply_tx got closed;
442 : /// it must mean that network thread terminated.
443 0 : async fn run(&mut self) -> anyhow::Result<()> {
444 0 : // Register the connection and defer unregister.
445 0 : // Order of the next two lines is important: we want first to remove our entry and then
446 0 : // update status which depends on registered connections.
447 0 : let _compute_conn_guard = ComputeConnectionGuard {
448 0 : timeline: Arc::clone(&self.tli),
449 0 : };
450 0 : let walreceiver_guard = self.tli.get_walreceivers().register(self.conn_id);
451 0 : self.tli.update_status_notify().await?;
452 :
453 : // After this timestamp we will stop processing AppendRequests and send a response
454 : // to the walproposer. walproposer sends at least one AppendRequest per second,
455 : // we will send keepalives by replying to these requests once per second.
456 0 : let mut next_keepalive = Instant::now();
457 :
458 : loop {
459 0 : let opt_msg = self.msg_rx.recv().await;
460 0 : if opt_msg.is_none() {
461 0 : return Ok(()); // chan closed, streaming terminated
462 0 : }
463 0 : let mut next_msg = opt_msg.unwrap();
464 0 :
465 0 : // Update walreceiver state in shmem for reporting.
466 0 : if let ProposerAcceptorMessage::Elected(_) = &next_msg {
467 0 : walreceiver_guard.get().status = WalReceiverStatus::Streaming;
468 0 : }
469 :
470 0 : let reply_msg = if matches!(next_msg, ProposerAcceptorMessage::AppendRequest(_)) {
471 : // loop through AppendRequest's while it's readily available to
472 : // write as many WAL as possible without fsyncing
473 : //
474 : // Note: this will need to be rewritten if we want to read non-AppendRequest messages here.
475 : // Otherwise, we might end up in a situation where we read a message, but don't
476 : // process it.
477 0 : while let ProposerAcceptorMessage::AppendRequest(append_request) = next_msg {
478 0 : let noflush_msg = ProposerAcceptorMessage::NoFlushAppendRequest(append_request);
479 :
480 0 : if let Some(reply) = self.tli.process_msg(&noflush_msg).await? {
481 0 : if self.reply_tx.send(reply).await.is_err() {
482 0 : return Ok(()); // chan closed, streaming terminated
483 0 : }
484 0 : }
485 :
486 : // get out of this loop if keepalive time is reached
487 0 : if Instant::now() >= next_keepalive {
488 0 : break;
489 0 : }
490 0 :
491 0 : match self.msg_rx.try_recv() {
492 0 : Ok(msg) => next_msg = msg,
493 0 : Err(TryRecvError::Empty) => break,
494 0 : Err(TryRecvError::Disconnected) => return Ok(()), // chan closed, streaming terminated
495 : }
496 : }
497 :
498 : // flush all written WAL to the disk
499 0 : self.tli
500 0 : .process_msg(&ProposerAcceptorMessage::FlushWAL)
501 0 : .await?
502 : } else {
503 : // process message other than AppendRequest
504 0 : self.tli.process_msg(&next_msg).await?
505 : };
506 :
507 0 : if let Some(reply) = reply_msg {
508 0 : if self.reply_tx.send(reply).await.is_err() {
509 0 : return Ok(()); // chan closed, streaming terminated
510 0 : }
511 0 : // reset keepalive time
512 0 : next_keepalive = Instant::now() + KEEPALIVE_INTERVAL;
513 0 : }
514 : }
515 0 : }
516 : }
517 :
518 : /// Calls update_status_notify in drop to update timeline status.
519 : struct ComputeConnectionGuard {
520 : timeline: Arc<Timeline>,
521 : }
522 :
523 : impl Drop for ComputeConnectionGuard {
524 0 : fn drop(&mut self) {
525 0 : let tli = self.timeline.clone();
526 0 : tokio::spawn(async move {
527 0 : if let Err(e) = tli.update_status_notify().await {
528 0 : error!("failed to update timeline status: {}", e);
529 0 : }
530 0 : });
531 0 : }
532 : }
|
