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 : NeonWALReadResult, SafekeeperStateDesiredEvents, WL_SOCKET_READABLE, WL_SOCKET_WRITEABLE,
21 : },
22 : walproposer::{ApiImpl, Config},
23 : };
24 :
25 : use super::walproposer_disk::DiskWalProposer;
26 :
27 : /// Special state for each wp->sk connection.
28 : struct SafekeeperConn {
29 : host: String,
30 : port: String,
31 : node_id: NodeId,
32 : // socket is Some(..) equals to connection is established
33 : socket: Option<TCP>,
34 : // connection is in progress
35 : is_connecting: bool,
36 : // START_WAL_PUSH is in progress
37 : is_start_wal_push: bool,
38 : // pointer to Safekeeper in walproposer for callbacks
39 : raw_ptr: *mut walproposer::bindings::Safekeeper,
40 : }
41 :
42 : impl SafekeeperConn {
43 108891 : pub fn new(host: String, port: String) -> Self {
44 108891 : // port number is the same as NodeId
45 108891 : let port_num = port.parse::<u32>().unwrap();
46 108891 : Self {
47 108891 : host,
48 108891 : port,
49 108891 : node_id: port_num,
50 108891 : socket: None,
51 108891 : is_connecting: false,
52 108891 : is_start_wal_push: false,
53 108891 : raw_ptr: std::ptr::null_mut(),
54 108891 : }
55 108891 : }
56 : }
57 :
58 : /// Simulation version of a postgres WaitEventSet. At pos 0 there is always
59 : /// a special NodeEvents channel, which is used as a latch.
60 : struct EventSet {
61 : os: NodeOs,
62 : // all pollable channels, 0 is always NodeEvent channel
63 : chans: Vec<Box<dyn PollSome>>,
64 : // 0 is always nullptr
65 : sk_ptrs: Vec<*mut walproposer::bindings::Safekeeper>,
66 : // event mask for each channel
67 : masks: Vec<u32>,
68 : }
69 :
70 : impl EventSet {
71 36297 : pub fn new(os: NodeOs) -> Self {
72 36297 : let node_events = os.node_events();
73 36297 : Self {
74 36297 : os,
75 36297 : chans: vec![Box::new(node_events)],
76 36297 : sk_ptrs: vec![std::ptr::null_mut()],
77 36297 : masks: vec![WL_SOCKET_READABLE],
78 36297 : }
79 36297 : }
80 :
81 : /// Leaves all readable channels at the beginning of the array.
82 114066 : fn sort_readable(&mut self) -> usize {
83 114066 : let mut cnt = 1;
84 271171 : for i in 1..self.chans.len() {
85 271171 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
86 271171 : self.chans.swap(i, cnt);
87 271171 : self.sk_ptrs.swap(i, cnt);
88 271171 : self.masks.swap(i, cnt);
89 271171 : cnt += 1;
90 271171 : }
91 : }
92 114066 : cnt
93 114066 : }
94 :
95 268343 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
96 268343 : let index = self
97 268343 : .sk_ptrs
98 268343 : .iter()
99 1012093 : .position(|&ptr| ptr == conn.raw_ptr)
100 268343 : .expect("safekeeper should exist in event set");
101 268343 : self.masks[index] = event_mask;
102 268343 : }
103 :
104 241760 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
105 584922 : for ptr in self.sk_ptrs.iter() {
106 584922 : assert!(*ptr != sk.raw_ptr);
107 : }
108 :
109 241760 : self.chans.push(Box::new(
110 241760 : sk.socket
111 241760 : .as_ref()
112 241760 : .expect("socket should not be closed")
113 241760 : .recv_chan(),
114 241760 : ));
115 241760 : self.sk_ptrs.push(sk.raw_ptr);
116 241760 : self.masks.push(event_mask);
117 241760 : }
118 :
119 152910 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
120 325284 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
121 152910 : if index.is_none() {
122 21 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
123 21 : return;
124 152889 : }
125 152889 : let index = index.unwrap();
126 152889 :
127 152889 : self.chans.remove(index);
128 152889 : self.sk_ptrs.remove(index);
129 152889 : self.masks.remove(index);
130 152889 :
131 152889 : // to simulate the actual behaviour
132 152889 : self.refresh_event_set();
133 152910 : }
134 :
135 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
136 177276 : fn refresh_event_set(&mut self) {
137 504581 : for (i, mask) in self.masks.iter_mut().enumerate() {
138 504581 : if i == 0 {
139 177276 : continue;
140 327305 : }
141 327305 :
142 327305 : let mut mask_sk: u32 = 0;
143 327305 : let mut mask_nwr: u32 = 0;
144 327305 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
145 327305 :
146 327305 : if mask_sk != *mask {
147 0 : debug!(
148 0 : "refresh_event_set: sk={:?}, old_mask={:#b}, new_mask={:#b}",
149 0 : self.sk_ptrs[i], *mask, mask_sk
150 : );
151 0 : *mask = mask_sk;
152 327305 : }
153 : }
154 177276 : }
155 :
156 : /// Wait for events on all channels.
157 114066 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
158 : // all channels are always writeable
159 385237 : for (i, mask) in self.masks.iter().enumerate() {
160 385237 : if *mask & WL_SOCKET_WRITEABLE != 0 {
161 0 : return walproposer::walproposer::WaitResult::Network(
162 0 : self.sk_ptrs[i],
163 0 : WL_SOCKET_WRITEABLE,
164 0 : );
165 385237 : }
166 : }
167 :
168 114066 : let cnt = self.sort_readable();
169 114066 :
170 114066 : let slice = &self.chans[0..cnt];
171 114066 : match executor::epoll_chans(slice, timeout_millis) {
172 45089 : None => walproposer::walproposer::WaitResult::Timeout,
173 : Some(0) => {
174 1599 : let msg = self.os.node_events().must_recv();
175 1599 : match msg {
176 1599 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
177 1599 : // got a notification about new WAL available
178 1599 : }
179 0 : NodeEvent::Internal(_) => unreachable!(),
180 0 : NodeEvent::Accept(_) => unreachable!(),
181 : }
182 1599 : walproposer::walproposer::WaitResult::Latch
183 : }
184 67378 : Some(index) => walproposer::walproposer::WaitResult::Network(
185 67378 : self.sk_ptrs[index],
186 67378 : WL_SOCKET_READABLE,
187 67378 : ),
188 : }
189 114066 : }
190 : }
191 :
192 : /// This struct handles all calls from walproposer into walproposer_api.
193 : pub struct SimulationApi {
194 : os: NodeOs,
195 : safekeepers: RefCell<Vec<SafekeeperConn>>,
196 : disk: Arc<DiskWalProposer>,
197 : redo_start_lsn: Option<Lsn>,
198 : last_logged_commit_lsn: u64,
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 36297 : pub fn new(args: Args) -> Self {
213 36297 : // initialize connection state for each safekeeper
214 36297 : let sk_conns = args
215 36297 : .config
216 36297 : .safekeepers_list
217 36297 : .iter()
218 108891 : .map(|s| {
219 108891 : SafekeeperConn::new(
220 108891 : s.split(':').next().unwrap().to_string(),
221 108891 : s.split(':').nth(1).unwrap().to_string(),
222 108891 : )
223 108891 : })
224 36297 : .collect::<Vec<_>>();
225 36297 :
226 36297 : Self {
227 36297 : os: args.os,
228 36297 : safekeepers: RefCell::new(sk_conns),
229 36297 : disk: args.disk,
230 36297 : redo_start_lsn: args.redo_start_lsn,
231 36297 : last_logged_commit_lsn: 0,
232 36297 : shmem: UnsafeCell::new(walproposer::api_bindings::empty_shmem()),
233 36297 : config: args.config,
234 36297 : event_set: RefCell::new(None),
235 36297 : }
236 36297 : }
237 :
238 : /// Get SafekeeperConn for the given Safekeeper.
239 1196512 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
240 1196512 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
241 1196512 : let state = self.safekeepers.borrow_mut();
242 1196512 : RefMut::map(state, |v| {
243 1196512 : v.iter_mut()
244 2391353 : .find(|conn| conn.port == sk_port)
245 1196512 : .expect("safekeeper conn not found by port")
246 1196512 : })
247 1196512 : }
248 : }
249 :
250 : impl ApiImpl for SimulationApi {
251 1278122 : fn get_current_timestamp(&self) -> i64 {
252 1278122 : debug!("get_current_timestamp");
253 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
254 : // milliseconds, so add 10^3
255 1278122 : self.os.now() as i64 * 1000
256 1278122 : }
257 :
258 4386 : fn update_donor(&self, donor: &mut walproposer::bindings::Safekeeper, donor_lsn: u64) {
259 4386 : let mut shmem = unsafe { *self.get_shmem_state() };
260 4386 : shmem.propEpochStartLsn.value = donor_lsn;
261 4386 : shmem.donor_conninfo = donor.conninfo;
262 4386 : }
263 :
264 134216 : fn conn_status(
265 134216 : &self,
266 134216 : _: &mut walproposer::bindings::Safekeeper,
267 134216 : ) -> walproposer::bindings::WalProposerConnStatusType {
268 134216 : debug!("conn_status");
269 : // break the connection with a 10% chance
270 134216 : if self.os.random(100) < 10 {
271 13336 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
272 : } else {
273 120880 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
274 : }
275 134216 : }
276 :
277 134216 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
278 134216 : debug!("conn_connect_start");
279 134216 : let mut conn = self.get_conn(sk);
280 134216 :
281 134216 : assert!(conn.socket.is_none());
282 134216 : let socket = self.os.open_tcp(conn.node_id);
283 134216 : conn.socket = Some(socket);
284 134216 : conn.raw_ptr = sk;
285 134216 : conn.is_connecting = true;
286 134216 : }
287 :
288 120880 : fn conn_connect_poll(
289 120880 : &self,
290 120880 : _: &mut walproposer::bindings::Safekeeper,
291 120880 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
292 120880 : debug!("conn_connect_poll");
293 : // TODO: break the connection here
294 120880 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
295 120880 : }
296 :
297 120880 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
298 120880 : debug!("conn_send_query: {}", query);
299 120880 : self.get_conn(sk).is_start_wal_push = true;
300 120880 : true
301 120880 : }
302 :
303 120880 : fn conn_get_query_result(
304 120880 : &self,
305 120880 : _: &mut walproposer::bindings::Safekeeper,
306 120880 : ) -> walproposer::bindings::WalProposerExecStatusType {
307 120880 : debug!("conn_get_query_result");
308 : // TODO: break the connection here
309 120880 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
310 120880 : }
311 :
312 74853 : fn conn_async_read(
313 74853 : &self,
314 74853 : sk: &mut walproposer::bindings::Safekeeper,
315 74853 : vec: &mut Vec<u8>,
316 74853 : ) -> walproposer::bindings::PGAsyncReadResult {
317 74853 : debug!("conn_async_read");
318 74853 : let mut conn = self.get_conn(sk);
319 :
320 74853 : let socket = if let Some(socket) = conn.socket.as_mut() {
321 74853 : socket
322 : } else {
323 : // socket is already closed
324 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
325 : };
326 :
327 74853 : let msg = socket.recv_chan().try_recv();
328 :
329 66559 : match msg {
330 : None => {
331 : // no message is ready
332 8294 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
333 : }
334 : Some(NetEvent::Closed) => {
335 : // connection is closed
336 30890 : debug!("conn_async_read: connection is closed");
337 30890 : conn.socket = None;
338 30890 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
339 : }
340 35669 : Some(NetEvent::Message(msg)) => {
341 : // got a message
342 35669 : let b = match msg {
343 35669 : desim::proto::AnyMessage::Bytes(b) => b,
344 0 : _ => unreachable!(),
345 : };
346 35669 : vec.extend_from_slice(&b);
347 35669 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
348 : }
349 : }
350 74853 : }
351 :
352 136305 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
353 136305 : let mut conn = self.get_conn(sk);
354 136305 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
355 136305 : let socket = conn.socket.as_mut().unwrap();
356 136305 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
357 136305 : true
358 136305 : }
359 :
360 20684 : fn conn_async_write(
361 20684 : &self,
362 20684 : sk: &mut walproposer::bindings::Safekeeper,
363 20684 : buf: &[u8],
364 20684 : ) -> walproposer::bindings::PGAsyncWriteResult {
365 20684 : let mut conn = self.get_conn(sk);
366 20684 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
367 20684 : if let Some(socket) = conn.socket.as_mut() {
368 20684 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
369 20684 : } else {
370 : // connection is already closed
371 0 : debug!("conn_async_write: writing to a closed socket!");
372 : // TODO: maybe we should return error here?
373 : }
374 20684 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
375 20684 : }
376 :
377 4591 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
378 4591 : debug!("wal_reader_allocate");
379 4591 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
380 4591 : }
381 :
382 16093 : fn wal_read(
383 16093 : &self,
384 16093 : _sk: &mut walproposer::bindings::Safekeeper,
385 16093 : buf: &mut [u8],
386 16093 : startpos: u64,
387 16093 : ) -> NeonWALReadResult {
388 16093 : self.disk.lock().read(startpos, buf);
389 16093 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
390 16093 : }
391 :
392 36297 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
393 36297 : debug!("init_event_set");
394 36297 : let new_event_set = EventSet::new(self.os.clone());
395 36297 : let old_event_set = self.event_set.replace(Some(new_event_set));
396 36297 : assert!(old_event_set.is_none());
397 36297 : }
398 :
399 268343 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
400 268343 : debug!(
401 0 : "update_event_set, sk={:?}, events_mask={:#b}",
402 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
403 : );
404 268343 : let conn = self.get_conn(sk);
405 268343 :
406 268343 : self.event_set
407 268343 : .borrow_mut()
408 268343 : .as_mut()
409 268343 : .unwrap()
410 268343 : .update_event_set(&conn, event_mask);
411 268343 : }
412 :
413 241760 : fn add_safekeeper_event_set(
414 241760 : &self,
415 241760 : sk: &mut walproposer::bindings::Safekeeper,
416 241760 : event_mask: u32,
417 241760 : ) {
418 241760 : debug!(
419 0 : "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
420 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
421 : );
422 :
423 241760 : self.event_set
424 241760 : .borrow_mut()
425 241760 : .as_mut()
426 241760 : .unwrap()
427 241760 : .add_safekeeper(&self.get_conn(sk), event_mask);
428 241760 : }
429 :
430 152910 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
431 152910 : debug!(
432 0 : "rm_safekeeper_event_set, sk={:?}",
433 0 : sk as *mut walproposer::bindings::Safekeeper,
434 : );
435 :
436 152910 : self.event_set
437 152910 : .borrow_mut()
438 152910 : .as_mut()
439 152910 : .unwrap()
440 152910 : .remove_safekeeper(&self.get_conn(sk));
441 152910 : }
442 :
443 24387 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
444 24387 : debug!("active_state_update_event_set");
445 :
446 24387 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
447 24387 : self.event_set
448 24387 : .borrow_mut()
449 24387 : .as_mut()
450 24387 : .unwrap()
451 24387 : .refresh_event_set();
452 24387 : }
453 :
454 70505 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
455 70505 : 0
456 70505 : }
457 :
458 355826 : fn wait_event_set(
459 355826 : &self,
460 355826 : _: &mut walproposer::bindings::WalProposer,
461 355826 : timeout_millis: i64,
462 355826 : ) -> walproposer::walproposer::WaitResult {
463 355826 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
464 355826 : let mut conns = self.safekeepers.borrow_mut();
465 825766 : for conn in conns.iter_mut() {
466 825766 : if conn.socket.is_some() && conn.is_connecting {
467 120880 : conn.is_connecting = false;
468 120880 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
469 120880 : return walproposer::walproposer::WaitResult::Network(
470 120880 : conn.raw_ptr,
471 120880 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
472 120880 : );
473 704886 : }
474 704886 : if conn.socket.is_some() && conn.is_start_wal_push {
475 120880 : conn.is_start_wal_push = false;
476 120880 : debug!(
477 0 : "wait_event_set, start wal push to {}:{}",
478 : conn.host, conn.port
479 : );
480 120880 : return walproposer::walproposer::WaitResult::Network(
481 120880 : conn.raw_ptr,
482 120880 : WL_SOCKET_READABLE,
483 120880 : );
484 584006 : }
485 : }
486 114066 : drop(conns);
487 114066 :
488 114066 : let res = self
489 114066 : .event_set
490 114066 : .borrow_mut()
491 114066 : .as_mut()
492 114066 : .unwrap()
493 114066 : .wait(timeout_millis);
494 114066 :
495 114066 : debug!(
496 0 : "wait_event_set, timeout_millis={}, res={:?}",
497 : timeout_millis, res,
498 : );
499 79852 : res
500 321612 : }
501 :
502 36297 : fn strong_random(&self, buf: &mut [u8]) -> bool {
503 36297 : debug!("strong_random");
504 36297 : buf.fill(0);
505 36297 : true
506 36297 : }
507 :
508 1514 : fn finish_sync_safekeepers(&self, lsn: u64) {
509 1514 : debug!("finish_sync_safekeepers, lsn={}", lsn);
510 1514 : executor::exit(0, Lsn(lsn).to_string());
511 1514 : }
512 :
513 369231 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
514 369231 : debug!("wp_log[{}] {}", level, msg);
515 369231 : if level == Level::Fatal || level == Level::Panic {
516 298 : if msg.contains("rejects our connection request with term") {
517 151 : // collected quorum with lower term, then got rejected by next connected safekeeper
518 151 : executor::exit(1, msg.to_owned());
519 151 : }
520 298 : if msg.contains("collected propEpochStartLsn") && msg.contains(", but basebackup LSN ")
521 10 : {
522 10 : // sync-safekeepers collected wrong quorum, walproposer collected another quorum
523 10 : executor::exit(1, msg.to_owned());
524 288 : }
525 298 : if msg.contains("failed to download WAL for logical replicaiton") {
526 91 : // Recovery connection broken and recovery was failed
527 91 : executor::exit(1, msg.to_owned());
528 207 : }
529 298 : if msg.contains("missing majority of votes, collected") {
530 46 : // Voting bug when safekeeper disconnects after voting
531 46 : executor::exit(1, msg.to_owned());
532 252 : }
533 298 : panic!("unknown FATAL error from walproposer: {}", msg);
534 368933 : }
535 368933 : }
536 :
537 3150 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
538 3150 : let prop_lsn = wp.propEpochStartLsn;
539 3150 : let prop_term = wp.propTerm;
540 3150 :
541 3150 : let mut prev_lsn: u64 = 0;
542 3150 : let mut prev_term: u64 = 0;
543 3150 :
544 3150 : unsafe {
545 3150 : let history = wp.propTermHistory.entries;
546 3150 : let len = wp.propTermHistory.n_entries as usize;
547 3150 : if len > 1 {
548 2231 : let entry = *history.wrapping_add(len - 2);
549 2231 : prev_lsn = entry.lsn;
550 2231 : prev_term = entry.term;
551 2231 : }
552 : }
553 :
554 3150 : let msg = format!(
555 3150 : "prop_elected;{};{};{};{}",
556 3150 : prop_lsn, prop_term, prev_lsn, prev_term
557 3150 : );
558 3150 :
559 3150 : debug!(msg);
560 3150 : self.os.log_event(msg);
561 3150 : }
562 :
563 1258 : fn get_redo_start_lsn(&self) -> u64 {
564 1258 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
565 1258 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
566 1258 : }
567 :
568 10706 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
569 10706 : self.shmem.get()
570 10706 : }
571 :
572 799 : fn start_streaming(
573 799 : &self,
574 799 : startpos: u64,
575 799 : callback: &walproposer::walproposer::StreamingCallback,
576 799 : ) {
577 799 : let disk = &self.disk;
578 799 : let disk_lsn = disk.lock().flush_rec_ptr().0;
579 799 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
580 799 : if startpos < disk_lsn {
581 183 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
582 616 : }
583 799 : assert!(startpos <= disk_lsn);
584 799 : let mut broadcasted = Lsn(startpos);
585 :
586 : loop {
587 2567 : let available = disk.lock().flush_rec_ptr();
588 2567 : assert!(available >= broadcasted);
589 1768 : callback.broadcast(broadcasted, available);
590 1768 : broadcasted = available;
591 1768 : callback.poll();
592 : }
593 : }
594 :
595 8918 : fn process_safekeeper_feedback(
596 8918 : &mut self,
597 8918 : wp: &mut walproposer::bindings::WalProposer,
598 8918 : _sk: &mut walproposer::bindings::Safekeeper,
599 8918 : ) {
600 8918 : debug!("process_safekeeper_feedback, commit_lsn={}", wp.commitLsn);
601 8918 : if wp.commitLsn > self.last_logged_commit_lsn {
602 1832 : self.os.log_event(format!("commit_lsn;{}", wp.commitLsn));
603 1832 : self.last_logged_commit_lsn = wp.commitLsn;
604 7086 : }
605 8918 : }
606 :
607 457 : fn get_flush_rec_ptr(&self) -> u64 {
608 457 : let lsn = self.disk.lock().flush_rec_ptr();
609 457 : debug!("get_flush_rec_ptr: {}", lsn);
610 457 : lsn.0
611 457 : }
612 :
613 3150 : fn recovery_download(
614 3150 : &self,
615 3150 : wp: &mut walproposer::bindings::WalProposer,
616 3150 : sk: &mut walproposer::bindings::Safekeeper,
617 3150 : ) -> bool {
618 3150 : let mut startpos = wp.truncateLsn;
619 3150 : let endpos = wp.propEpochStartLsn;
620 3150 :
621 3150 : if startpos == endpos {
622 1955 : debug!("recovery_download: nothing to download");
623 1955 : return true;
624 1195 : }
625 1195 :
626 1195 : debug!("recovery_download from {} to {}", startpos, endpos,);
627 :
628 1195 : let replication_prompt = format!(
629 1195 : "START_REPLICATION {} {} {} {}",
630 1195 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
631 1195 : );
632 1195 : let async_conn = self.get_conn(sk);
633 1195 :
634 1195 : let conn = self.os.open_tcp(async_conn.node_id);
635 1195 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
636 1195 :
637 1195 : let chan = conn.recv_chan();
638 2028 : while startpos < endpos {
639 1195 : let event = chan.recv();
640 1104 : match event {
641 : NetEvent::Closed => {
642 91 : debug!("connection closed in recovery");
643 91 : break;
644 : }
645 1104 : NetEvent::Message(AnyMessage::Bytes(b)) => {
646 1104 : debug!("got recovery bytes from safekeeper");
647 833 : self.disk.lock().write(startpos, &b);
648 833 : startpos += b.len() as u64;
649 : }
650 0 : NetEvent::Message(_) => unreachable!(),
651 : }
652 : }
653 :
654 924 : debug!("recovery finished at {}", startpos);
655 :
656 924 : startpos == endpos
657 2879 : }
658 :
659 45366 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
660 45366 : let mut conn = self.get_conn(sk);
661 45366 : debug!("conn_finish to {}", conn.node_id);
662 45366 : if let Some(socket) = conn.socket.as_mut() {
663 14455 : socket.close();
664 30911 : } else {
665 30911 : // connection is already closed
666 30911 : }
667 45366 : conn.socket = None;
668 45366 : }
669 :
670 44226 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
671 44226 : "connection is closed, probably".into()
672 44226 : }
673 : }
|
