Line data Source code
1 : use std::{
2 : cell::{RefCell, RefMut, UnsafeCell},
3 : ffi::CStr,
4 : sync::Arc,
5 : };
6 :
7 : use bytes::Bytes;
8 : use desim::{
9 : executor::{self, PollSome},
10 : network::TCP,
11 : node_os::NodeOs,
12 : proto::{AnyMessage, NetEvent, NodeEvent},
13 : world::NodeId,
14 : };
15 : use tracing::debug;
16 : use utils::lsn::Lsn;
17 : use walproposer::{
18 : api_bindings::Level,
19 : bindings::{
20 : pg_atomic_uint64, NeonWALReadResult, PageserverFeedback, SafekeeperStateDesiredEvents,
21 : WL_SOCKET_READABLE, WL_SOCKET_WRITEABLE,
22 : },
23 : walproposer::{ApiImpl, Config},
24 : };
25 :
26 : use super::walproposer_disk::DiskWalProposer;
27 :
28 : /// Special state for each wp->sk connection.
29 : struct SafekeeperConn {
30 : host: String,
31 : port: String,
32 : node_id: NodeId,
33 : // socket is Some(..) equals to connection is established
34 : socket: Option<TCP>,
35 : // connection is in progress
36 : is_connecting: bool,
37 : // START_WAL_PUSH is in progress
38 : is_start_wal_push: bool,
39 : // pointer to Safekeeper in walproposer for callbacks
40 : raw_ptr: *mut walproposer::bindings::Safekeeper,
41 : }
42 :
43 : impl SafekeeperConn {
44 214971 : pub fn new(host: String, port: String) -> Self {
45 214971 : // port number is the same as NodeId
46 214971 : let port_num = port.parse::<u32>().unwrap();
47 214971 : Self {
48 214971 : host,
49 214971 : port,
50 214971 : node_id: port_num,
51 214971 : socket: None,
52 214971 : is_connecting: false,
53 214971 : is_start_wal_push: false,
54 214971 : raw_ptr: std::ptr::null_mut(),
55 214971 : }
56 214971 : }
57 : }
58 :
59 : /// Simulation version of a postgres WaitEventSet. At pos 0 there is always
60 : /// a special NodeEvents channel, which is used as a latch.
61 : struct EventSet {
62 : os: NodeOs,
63 : // all pollable channels, 0 is always NodeEvent channel
64 : chans: Vec<Box<dyn PollSome>>,
65 : // 0 is always nullptr
66 : sk_ptrs: Vec<*mut walproposer::bindings::Safekeeper>,
67 : // event mask for each channel
68 : masks: Vec<u32>,
69 : }
70 :
71 : impl EventSet {
72 71657 : pub fn new(os: NodeOs) -> Self {
73 71657 : let node_events = os.node_events();
74 71657 : Self {
75 71657 : os,
76 71657 : chans: vec![Box::new(node_events)],
77 71657 : sk_ptrs: vec![std::ptr::null_mut()],
78 71657 : masks: vec![WL_SOCKET_READABLE],
79 71657 : }
80 71657 : }
81 :
82 : /// Leaves all readable channels at the beginning of the array.
83 213531 : fn sort_readable(&mut self) -> usize {
84 213531 : let mut cnt = 1;
85 504831 : for i in 1..self.chans.len() {
86 504831 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
87 504831 : self.chans.swap(i, cnt);
88 504831 : self.sk_ptrs.swap(i, cnt);
89 504831 : self.masks.swap(i, cnt);
90 504831 : cnt += 1;
91 504831 : }
92 : }
93 213531 : cnt
94 213531 : }
95 :
96 523932 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
97 523932 : let index = self
98 523932 : .sk_ptrs
99 523932 : .iter()
100 1978704 : .position(|&ptr| ptr == conn.raw_ptr)
101 523932 : .expect("safekeeper should exist in event set");
102 523932 : self.masks[index] = event_mask;
103 523932 : }
104 :
105 475400 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
106 1150206 : for ptr in self.sk_ptrs.iter() {
107 1150206 : assert!(*ptr != sk.raw_ptr);
108 : }
109 :
110 475400 : self.chans.push(Box::new(
111 475400 : sk.socket
112 475400 : .as_ref()
113 475400 : .expect("socket should not be closed")
114 475400 : .recv_chan(),
115 475400 : ));
116 475400 : self.sk_ptrs.push(sk.raw_ptr);
117 475400 : self.masks.push(event_mask);
118 475400 : }
119 :
120 299350 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
121 637357 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
122 299350 : if index.is_none() {
123 44 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
124 44 : return;
125 299306 : }
126 299306 : let index = index.unwrap();
127 299306 :
128 299306 : self.chans.remove(index);
129 299306 : self.sk_ptrs.remove(index);
130 299306 : self.masks.remove(index);
131 299306 :
132 299306 : // to simulate the actual behaviour
133 299306 : self.refresh_event_set();
134 299350 : }
135 :
136 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
137 333830 : fn refresh_event_set(&mut self) {
138 936517 : for (i, mask) in self.masks.iter_mut().enumerate() {
139 936517 : if i == 0 {
140 333830 : continue;
141 602687 : }
142 602687 :
143 602687 : let mut mask_sk: u32 = 0;
144 602687 : let mut mask_nwr: u32 = 0;
145 602687 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
146 602687 :
147 602687 : if mask_sk != *mask {
148 0 : debug!(
149 0 : "refresh_event_set: sk={:?}, old_mask={:#b}, new_mask={:#b}",
150 0 : self.sk_ptrs[i], *mask, mask_sk
151 0 : );
152 0 : *mask = mask_sk;
153 602687 : }
154 : }
155 333830 : }
156 :
157 : /// Wait for events on all channels.
158 213531 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
159 : // all channels are always writeable
160 718362 : for (i, mask) in self.masks.iter().enumerate() {
161 718362 : if *mask & WL_SOCKET_WRITEABLE != 0 {
162 0 : return walproposer::walproposer::WaitResult::Network(
163 0 : self.sk_ptrs[i],
164 0 : WL_SOCKET_WRITEABLE,
165 0 : );
166 718362 : }
167 : }
168 :
169 213531 : let cnt = self.sort_readable();
170 213531 :
171 213531 : let slice = &self.chans[0..cnt];
172 213531 : match executor::epoll_chans(slice, timeout_millis) {
173 89837 : None => walproposer::walproposer::WaitResult::Timeout,
174 : Some(0) => {
175 2717 : let msg = self.os.node_events().must_recv();
176 2717 : match msg {
177 2717 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
178 2717 : // got a notification about new WAL available
179 2717 : }
180 0 : NodeEvent::Internal(_) => unreachable!(),
181 0 : NodeEvent::Accept(_) => unreachable!(),
182 : }
183 2717 : walproposer::walproposer::WaitResult::Latch
184 : }
185 120977 : Some(index) => walproposer::walproposer::WaitResult::Network(
186 120977 : self.sk_ptrs[index],
187 120977 : WL_SOCKET_READABLE,
188 120977 : ),
189 : }
190 213531 : }
191 : }
192 :
193 : /// This struct handles all calls from walproposer into walproposer_api.
194 : pub struct SimulationApi {
195 : os: NodeOs,
196 : safekeepers: RefCell<Vec<SafekeeperConn>>,
197 : disk: Arc<DiskWalProposer>,
198 : redo_start_lsn: Option<Lsn>,
199 : shmem: UnsafeCell<walproposer::bindings::WalproposerShmemState>,
200 : config: Config,
201 : event_set: RefCell<Option<EventSet>>,
202 : }
203 :
204 : pub struct Args {
205 : pub os: NodeOs,
206 : pub config: Config,
207 : pub disk: Arc<DiskWalProposer>,
208 : pub redo_start_lsn: Option<Lsn>,
209 : }
210 :
211 : impl SimulationApi {
212 71657 : pub fn new(args: Args) -> Self {
213 71657 : // initialize connection state for each safekeeper
214 71657 : let sk_conns = args
215 71657 : .config
216 71657 : .safekeepers_list
217 71657 : .iter()
218 214971 : .map(|s| {
219 214971 : SafekeeperConn::new(
220 214971 : s.split(':').next().unwrap().to_string(),
221 214971 : s.split(':').nth(1).unwrap().to_string(),
222 214971 : )
223 214971 : })
224 71657 : .collect::<Vec<_>>();
225 71657 :
226 71657 : Self {
227 71657 : os: args.os,
228 71657 : safekeepers: RefCell::new(sk_conns),
229 71657 : disk: args.disk,
230 71657 : redo_start_lsn: args.redo_start_lsn,
231 71657 : shmem: UnsafeCell::new(walproposer::bindings::WalproposerShmemState {
232 71657 : mutex: 0,
233 71657 : feedback: PageserverFeedback {
234 71657 : currentClusterSize: 0,
235 71657 : last_received_lsn: 0,
236 71657 : disk_consistent_lsn: 0,
237 71657 : remote_consistent_lsn: 0,
238 71657 : replytime: 0,
239 71657 : },
240 71657 : mineLastElectedTerm: 0,
241 71657 : backpressureThrottlingTime: pg_atomic_uint64 { value: 0 },
242 71657 : }),
243 71657 : config: args.config,
244 71657 : event_set: RefCell::new(None),
245 71657 : }
246 71657 : }
247 :
248 : /// Get SafekeeperConn for the given Safekeeper.
249 2317699 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
250 2317699 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
251 2317699 : let state = self.safekeepers.borrow_mut();
252 2317699 : RefMut::map(state, |v| {
253 2317699 : v.iter_mut()
254 4630905 : .find(|conn| conn.port == sk_port)
255 2317699 : .expect("safekeeper conn not found by port")
256 2317699 : })
257 2317699 : }
258 : }
259 :
260 : impl ApiImpl for SimulationApi {
261 2472414 : fn get_current_timestamp(&self) -> i64 {
262 2472414 : debug!("get_current_timestamp");
263 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
264 : // milliseconds, so add 10^3
265 2472414 : self.os.now() as i64 * 1000
266 2472414 : }
267 :
268 264050 : fn conn_status(
269 264050 : &self,
270 264050 : _: &mut walproposer::bindings::Safekeeper,
271 264050 : ) -> walproposer::bindings::WalProposerConnStatusType {
272 264050 : debug!("conn_status");
273 : // break the connection with a 10% chance
274 264050 : if self.os.random(100) < 10 {
275 26350 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
276 : } else {
277 237700 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
278 : }
279 264050 : }
280 :
281 264050 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
282 264050 : debug!("conn_connect_start");
283 264050 : let mut conn = self.get_conn(sk);
284 264050 :
285 264050 : assert!(conn.socket.is_none());
286 264050 : let socket = self.os.open_tcp(conn.node_id);
287 264050 : conn.socket = Some(socket);
288 264050 : conn.raw_ptr = sk;
289 264050 : conn.is_connecting = true;
290 264050 : }
291 :
292 237700 : fn conn_connect_poll(
293 237700 : &self,
294 237700 : _: &mut walproposer::bindings::Safekeeper,
295 237700 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
296 237700 : debug!("conn_connect_poll");
297 : // TODO: break the connection here
298 237700 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
299 237700 : }
300 :
301 237700 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
302 237700 : debug!("conn_send_query: {}", query);
303 237700 : self.get_conn(sk).is_start_wal_push = true;
304 237700 : true
305 237700 : }
306 :
307 237700 : fn conn_get_query_result(
308 237700 : &self,
309 237700 : _: &mut walproposer::bindings::Safekeeper,
310 237700 : ) -> walproposer::bindings::WalProposerExecStatusType {
311 237700 : debug!("conn_get_query_result");
312 : // TODO: break the connection here
313 237700 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
314 237700 : }
315 :
316 130670 : fn conn_async_read(
317 130670 : &self,
318 130670 : sk: &mut walproposer::bindings::Safekeeper,
319 130670 : vec: &mut Vec<u8>,
320 130670 : ) -> walproposer::bindings::PGAsyncReadResult {
321 130670 : debug!("conn_async_read");
322 130670 : let mut conn = self.get_conn(sk);
323 :
324 130670 : let socket = if let Some(socket) = conn.socket.as_mut() {
325 130670 : socket
326 : } else {
327 : // socket is already closed
328 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
329 : };
330 :
331 130670 : let msg = socket.recv_chan().try_recv();
332 :
333 119273 : match msg {
334 : None => {
335 : // no message is ready
336 11397 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
337 : }
338 : Some(NetEvent::Closed) => {
339 : // connection is closed
340 59411 : debug!("conn_async_read: connection is closed");
341 59411 : conn.socket = None;
342 59411 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
343 : }
344 59862 : Some(NetEvent::Message(msg)) => {
345 : // got a message
346 59862 : let b = match msg {
347 59862 : desim::proto::AnyMessage::Bytes(b) => b,
348 0 : _ => unreachable!(),
349 : };
350 59862 : vec.extend_from_slice(&b);
351 59862 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
352 : }
353 : }
354 130670 : }
355 :
356 265726 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
357 265726 : let mut conn = self.get_conn(sk);
358 265726 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
359 265726 : let socket = conn.socket.as_mut().unwrap();
360 265726 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
361 265726 : true
362 265726 : }
363 :
364 30658 : fn conn_async_write(
365 30658 : &self,
366 30658 : sk: &mut walproposer::bindings::Safekeeper,
367 30658 : buf: &[u8],
368 30658 : ) -> walproposer::bindings::PGAsyncWriteResult {
369 30658 : let mut conn = self.get_conn(sk);
370 30658 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
371 30658 : if let Some(socket) = conn.socket.as_mut() {
372 30658 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
373 30658 : } else {
374 : // connection is already closed
375 0 : debug!("conn_async_write: writing to a closed socket!");
376 : // TODO: maybe we should return error here?
377 : }
378 30658 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
379 30658 : }
380 :
381 7078 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
382 7078 : debug!("wal_reader_allocate");
383 7078 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
384 7078 : }
385 :
386 23580 : fn wal_read(
387 23580 : &self,
388 23580 : _sk: &mut walproposer::bindings::Safekeeper,
389 23580 : buf: &mut [u8],
390 23580 : startpos: u64,
391 23580 : ) -> NeonWALReadResult {
392 23580 : self.disk.lock().read(startpos, buf);
393 23580 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
394 23580 : }
395 :
396 71657 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
397 71657 : debug!("init_event_set");
398 71657 : let new_event_set = EventSet::new(self.os.clone());
399 71657 : let old_event_set = self.event_set.replace(Some(new_event_set));
400 71657 : assert!(old_event_set.is_none());
401 71657 : }
402 :
403 523932 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
404 523932 : debug!(
405 1166 : "update_event_set, sk={:?}, events_mask={:#b}",
406 1166 : sk as *mut walproposer::bindings::Safekeeper, event_mask
407 1166 : );
408 523932 : let conn = self.get_conn(sk);
409 523932 :
410 523932 : self.event_set
411 523932 : .borrow_mut()
412 523932 : .as_mut()
413 523932 : .unwrap()
414 523932 : .update_event_set(&conn, event_mask);
415 523932 : }
416 :
417 475400 : fn add_safekeeper_event_set(
418 475400 : &self,
419 475400 : sk: &mut walproposer::bindings::Safekeeper,
420 475400 : event_mask: u32,
421 475400 : ) {
422 475400 : debug!(
423 728 : "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
424 728 : sk as *mut walproposer::bindings::Safekeeper, event_mask
425 728 : );
426 :
427 475400 : self.event_set
428 475400 : .borrow_mut()
429 475400 : .as_mut()
430 475400 : .unwrap()
431 475400 : .add_safekeeper(&self.get_conn(sk), event_mask);
432 475400 : }
433 :
434 299350 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
435 299350 : debug!(
436 400 : "rm_safekeeper_event_set, sk={:?}",
437 400 : sk as *mut walproposer::bindings::Safekeeper,
438 400 : );
439 :
440 299350 : self.event_set
441 299350 : .borrow_mut()
442 299350 : .as_mut()
443 299350 : .unwrap()
444 299350 : .remove_safekeeper(&self.get_conn(sk));
445 299350 : }
446 :
447 34524 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
448 34524 : debug!("active_state_update_event_set");
449 :
450 34524 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
451 34524 : self.event_set
452 34524 : .borrow_mut()
453 34524 : .as_mut()
454 34524 : .unwrap()
455 34524 : .refresh_event_set();
456 34524 : }
457 :
458 97378 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
459 97378 : 0
460 97378 : }
461 :
462 688931 : fn wait_event_set(
463 688931 : &self,
464 688931 : _: &mut walproposer::bindings::WalProposer,
465 688931 : timeout_millis: i64,
466 688931 : ) -> walproposer::walproposer::WaitResult {
467 688931 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
468 688931 : let mut conns = self.safekeepers.borrow_mut();
469 1591265 : for conn in conns.iter_mut() {
470 1591265 : if conn.socket.is_some() && conn.is_connecting {
471 237700 : conn.is_connecting = false;
472 237700 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
473 237700 : return walproposer::walproposer::WaitResult::Network(
474 237700 : conn.raw_ptr,
475 237700 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
476 237700 : );
477 1353565 : }
478 1353565 : if conn.socket.is_some() && conn.is_start_wal_push {
479 237700 : conn.is_start_wal_push = false;
480 237700 : debug!(
481 364 : "wait_event_set, start wal push to {}:{}",
482 364 : conn.host, conn.port
483 364 : );
484 237700 : return walproposer::walproposer::WaitResult::Network(
485 237700 : conn.raw_ptr,
486 237700 : WL_SOCKET_READABLE,
487 237700 : );
488 1115865 : }
489 : }
490 213531 : drop(conns);
491 213531 :
492 213531 : let res = self
493 213531 : .event_set
494 213531 : .borrow_mut()
495 213531 : .as_mut()
496 213531 : .unwrap()
497 213531 : .wait(timeout_millis);
498 213531 :
499 213531 : debug!(
500 690 : "wait_event_set, timeout_millis={}, res={:?}",
501 690 : timeout_millis, res,
502 690 : );
503 145537 : res
504 620937 : }
505 :
506 71657 : fn strong_random(&self, buf: &mut [u8]) -> bool {
507 71657 : debug!("strong_random");
508 71657 : buf.fill(0);
509 71657 : true
510 71657 : }
511 :
512 2553 : fn finish_sync_safekeepers(&self, lsn: u64) {
513 2553 : debug!("finish_sync_safekeepers, lsn={}", lsn);
514 2553 : executor::exit(0, Lsn(lsn).to_string());
515 2553 : }
516 :
517 759620 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
518 759620 : debug!("wp_log[{}] {}", level, msg);
519 759620 : if level == Level::Fatal || level == Level::Panic {
520 549 : if msg.contains("rejects our connection request with term") {
521 301 : // collected quorum with lower term, then got rejected by next connected safekeeper
522 301 : executor::exit(1, msg.to_owned());
523 301 : }
524 549 : if msg.contains("collected propEpochStartLsn") && msg.contains(", but basebackup LSN ")
525 20 : {
526 20 : // sync-safekeepers collected wrong quorum, walproposer collected another quorum
527 20 : executor::exit(1, msg.to_owned());
528 529 : }
529 549 : if msg.contains("failed to download WAL for logical replicaiton") {
530 153 : // Recovery connection broken and recovery was failed
531 153 : executor::exit(1, msg.to_owned());
532 396 : }
533 549 : if msg.contains("missing majority of votes, collected") {
534 75 : // Voting bug when safekeeper disconnects after voting
535 75 : executor::exit(1, msg.to_owned());
536 474 : }
537 549 : panic!("unknown FATAL error from walproposer: {}", msg);
538 759071 : }
539 759071 : }
540 :
541 5335 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
542 5335 : let prop_lsn = wp.propEpochStartLsn;
543 5335 : let prop_term = wp.propTerm;
544 5335 :
545 5335 : let mut prev_lsn: u64 = 0;
546 5335 : let mut prev_term: u64 = 0;
547 5335 :
548 5335 : unsafe {
549 5335 : let history = wp.propTermHistory.entries;
550 5335 : let len = wp.propTermHistory.n_entries as usize;
551 5335 : if len > 1 {
552 3508 : let entry = *history.wrapping_add(len - 2);
553 3508 : prev_lsn = entry.lsn;
554 3508 : prev_term = entry.term;
555 3508 : }
556 : }
557 :
558 5335 : let msg = format!(
559 5335 : "prop_elected;{};{};{};{}",
560 5335 : prop_lsn, prop_term, prev_lsn, prev_term
561 5335 : );
562 5335 :
563 5335 : debug!(msg);
564 5335 : self.os.log_event(msg);
565 5335 : }
566 :
567 1961 : fn get_redo_start_lsn(&self) -> u64 {
568 1961 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
569 1961 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
570 1961 : }
571 :
572 1135 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
573 1135 : self.shmem.get()
574 1135 : }
575 :
576 1095 : fn start_streaming(
577 1095 : &self,
578 1095 : startpos: u64,
579 1095 : callback: &walproposer::walproposer::StreamingCallback,
580 1095 : ) {
581 1095 : let disk = &self.disk;
582 1095 : let disk_lsn = disk.lock().flush_rec_ptr().0;
583 1095 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
584 1095 : if startpos < disk_lsn {
585 338 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
586 757 : }
587 1095 : assert!(startpos <= disk_lsn);
588 1095 : let mut broadcasted = Lsn(startpos);
589 :
590 : loop {
591 3620 : let available = disk.lock().flush_rec_ptr();
592 3620 : assert!(available >= broadcasted);
593 2525 : callback.broadcast(broadcasted, available);
594 2525 : broadcasted = available;
595 2525 : callback.poll();
596 : }
597 : }
598 :
599 11397 : fn process_safekeeper_feedback(
600 11397 : &self,
601 11397 : wp: &mut walproposer::bindings::WalProposer,
602 11397 : commit_lsn: u64,
603 11397 : ) {
604 11397 : debug!("process_safekeeper_feedback, commit_lsn={}", commit_lsn);
605 11397 : if commit_lsn > wp.lastSentCommitLsn {
606 2539 : self.os.log_event(format!("commit_lsn;{}", commit_lsn));
607 8858 : }
608 11397 : }
609 :
610 726 : fn get_flush_rec_ptr(&self) -> u64 {
611 726 : let lsn = self.disk.lock().flush_rec_ptr();
612 726 : debug!("get_flush_rec_ptr: {}", lsn);
613 726 : lsn.0
614 726 : }
615 :
616 5335 : fn recovery_download(
617 5335 : &self,
618 5335 : wp: &mut walproposer::bindings::WalProposer,
619 5335 : sk: &mut walproposer::bindings::Safekeeper,
620 5335 : ) -> bool {
621 5335 : let mut startpos = wp.truncateLsn;
622 5335 : let endpos = wp.propEpochStartLsn;
623 5335 :
624 5335 : if startpos == endpos {
625 3122 : debug!("recovery_download: nothing to download");
626 3122 : return true;
627 2213 : }
628 2213 :
629 2213 : debug!("recovery_download from {} to {}", startpos, endpos,);
630 :
631 2213 : let replication_prompt = format!(
632 2213 : "START_REPLICATION {} {} {} {}",
633 2213 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
634 2213 : );
635 2213 : let async_conn = self.get_conn(sk);
636 2213 :
637 2213 : let conn = self.os.open_tcp(async_conn.node_id);
638 2213 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
639 2213 :
640 2213 : let chan = conn.recv_chan();
641 3712 : while startpos < endpos {
642 2213 : let event = chan.recv();
643 2060 : match event {
644 : NetEvent::Closed => {
645 153 : debug!("connection closed in recovery");
646 153 : break;
647 : }
648 2060 : NetEvent::Message(AnyMessage::Bytes(b)) => {
649 2060 : debug!("got recovery bytes from safekeeper");
650 1499 : self.disk.lock().write(startpos, &b);
651 1499 : startpos += b.len() as u64;
652 : }
653 0 : NetEvent::Message(_) => unreachable!(),
654 : }
655 : }
656 :
657 1652 : debug!("recovery finished at {}", startpos);
658 :
659 1652 : startpos == endpos
660 4774 : }
661 :
662 88000 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
663 88000 : let mut conn = self.get_conn(sk);
664 88000 : debug!("conn_finish to {}", conn.node_id);
665 88000 : if let Some(socket) = conn.socket.as_mut() {
666 28545 : socket.close();
667 59455 : } else {
668 59455 : // connection is already closed
669 59455 : }
670 88000 : conn.socket = None;
671 88000 : }
672 :
673 85761 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
674 85761 : "connection is closed, probably".into()
675 85761 : }
676 : }
|
