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 217299 : pub fn new(host: String, port: String) -> Self {
44 217299 : // port number is the same as NodeId
45 217299 : let port_num = port.parse::<u32>().unwrap();
46 217299 : Self {
47 217299 : host,
48 217299 : port,
49 217299 : node_id: port_num,
50 217299 : socket: None,
51 217299 : is_connecting: false,
52 217299 : is_start_wal_push: false,
53 217299 : raw_ptr: std::ptr::null_mut(),
54 217299 : }
55 217299 : }
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 72433 : pub fn new(os: NodeOs) -> Self {
72 72433 : let node_events = os.node_events();
73 72433 : Self {
74 72433 : os,
75 72433 : chans: vec![Box::new(node_events)],
76 72433 : sk_ptrs: vec![std::ptr::null_mut()],
77 72433 : masks: vec![WL_SOCKET_READABLE],
78 72433 : }
79 72433 : }
80 :
81 : /// Leaves all readable channels at the beginning of the array.
82 214333 : fn sort_readable(&mut self) -> usize {
83 214333 : let mut cnt = 1;
84 505243 : for i in 1..self.chans.len() {
85 505243 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
86 505243 : self.chans.swap(i, cnt);
87 505243 : self.sk_ptrs.swap(i, cnt);
88 505243 : self.masks.swap(i, cnt);
89 505243 : cnt += 1;
90 505243 : }
91 : }
92 214333 : cnt
93 214333 : }
94 :
95 528976 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
96 528976 : let index = self
97 528976 : .sk_ptrs
98 528976 : .iter()
99 1999469 : .position(|&ptr| ptr == conn.raw_ptr)
100 528976 : .expect("safekeeper should exist in event set");
101 528976 : self.masks[index] = event_mask;
102 528976 : }
103 :
104 481110 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
105 1164503 : for ptr in self.sk_ptrs.iter() {
106 1164503 : assert!(*ptr != sk.raw_ptr);
107 : }
108 :
109 481110 : self.chans.push(Box::new(
110 481110 : sk.socket
111 481110 : .as_ref()
112 481110 : .expect("socket should not be closed")
113 481110 : .recv_chan(),
114 481110 : ));
115 481110 : self.sk_ptrs.push(sk.raw_ptr);
116 481110 : self.masks.push(event_mask);
117 481110 : }
118 :
119 303353 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
120 644661 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
121 303353 : if index.is_none() {
122 36 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
123 36 : return;
124 303317 : }
125 303317 : let index = index.unwrap();
126 303317 :
127 303317 : self.chans.remove(index);
128 303317 : self.sk_ptrs.remove(index);
129 303317 : self.masks.remove(index);
130 303317 :
131 303317 : // to simulate the actual behaviour
132 303317 : self.refresh_event_set();
133 303353 : }
134 :
135 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
136 336766 : fn refresh_event_set(&mut self) {
137 943768 : for (i, mask) in self.masks.iter_mut().enumerate() {
138 943768 : if i == 0 {
139 336766 : continue;
140 607002 : }
141 607002 :
142 607002 : let mut mask_sk: u32 = 0;
143 607002 : let mut mask_nwr: u32 = 0;
144 607002 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
145 607002 :
146 607002 : 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 607002 : }
153 : }
154 336766 : }
155 :
156 : /// Wait for events on all channels.
157 214333 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
158 : // all channels are always writeable
159 719576 : for (i, mask) in self.masks.iter().enumerate() {
160 719576 : 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 719576 : }
166 : }
167 :
168 214333 : let cnt = self.sort_readable();
169 214333 :
170 214333 : let slice = &self.chans[0..cnt];
171 214333 : match executor::epoll_chans(slice, timeout_millis) {
172 90656 : None => walproposer::walproposer::WaitResult::Timeout,
173 : Some(0) => {
174 2685 : let msg = self.os.node_events().must_recv();
175 2685 : match msg {
176 2685 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
177 2685 : // got a notification about new WAL available
178 2685 : }
179 0 : NodeEvent::Internal(_) => unreachable!(),
180 0 : NodeEvent::Accept(_) => unreachable!(),
181 : }
182 2685 : walproposer::walproposer::WaitResult::Latch
183 : }
184 120992 : Some(index) => walproposer::walproposer::WaitResult::Network(
185 120992 : self.sk_ptrs[index],
186 120992 : WL_SOCKET_READABLE,
187 120992 : ),
188 : }
189 214333 : }
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 72433 : pub fn new(args: Args) -> Self {
213 72433 : // initialize connection state for each safekeeper
214 72433 : let sk_conns = args
215 72433 : .config
216 72433 : .safekeepers_list
217 72433 : .iter()
218 217299 : .map(|s| {
219 217299 : SafekeeperConn::new(
220 217299 : s.split(':').next().unwrap().to_string(),
221 217299 : s.split(':').nth(1).unwrap().to_string(),
222 217299 : )
223 217299 : })
224 72433 : .collect::<Vec<_>>();
225 72433 :
226 72433 : Self {
227 72433 : os: args.os,
228 72433 : safekeepers: RefCell::new(sk_conns),
229 72433 : disk: args.disk,
230 72433 : redo_start_lsn: args.redo_start_lsn,
231 72433 : last_logged_commit_lsn: 0,
232 72433 : shmem: UnsafeCell::new(walproposer::api_bindings::empty_shmem()),
233 72433 : config: args.config,
234 72433 : event_set: RefCell::new(None),
235 72433 : }
236 72433 : }
237 :
238 : /// Get SafekeeperConn for the given Safekeeper.
239 2340401 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
240 2340401 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
241 2340401 : let state = self.safekeepers.borrow_mut();
242 2340401 : RefMut::map(state, |v| {
243 2340401 : v.iter_mut()
244 4676110 : .find(|conn| conn.port == sk_port)
245 2340401 : .expect("safekeeper conn not found by port")
246 2340401 : })
247 2340401 : }
248 : }
249 :
250 : impl ApiImpl for SimulationApi {
251 2494844 : fn get_current_timestamp(&self) -> i64 {
252 2494844 : debug!("get_current_timestamp");
253 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
254 : // milliseconds, so add 10^3
255 2494844 : self.os.now() as i64 * 1000
256 2494844 : }
257 :
258 6396 : fn update_donor(&self, donor: &mut walproposer::bindings::Safekeeper, donor_lsn: u64) {
259 6396 : let mut shmem = unsafe { *self.get_shmem_state() };
260 6396 : shmem.propEpochStartLsn.value = donor_lsn;
261 6396 : shmem.donor_conninfo = donor.conninfo;
262 6396 : }
263 :
264 267111 : fn conn_status(
265 267111 : &self,
266 267111 : _: &mut walproposer::bindings::Safekeeper,
267 267111 : ) -> walproposer::bindings::WalProposerConnStatusType {
268 267111 : debug!("conn_status");
269 : // break the connection with a 10% chance
270 267111 : if self.os.random(100) < 10 {
271 26556 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
272 : } else {
273 240555 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
274 : }
275 267111 : }
276 :
277 267111 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
278 267111 : debug!("conn_connect_start");
279 267111 : let mut conn = self.get_conn(sk);
280 267111 :
281 267111 : assert!(conn.socket.is_none());
282 267111 : let socket = self.os.open_tcp(conn.node_id);
283 267111 : conn.socket = Some(socket);
284 267111 : conn.raw_ptr = sk;
285 267111 : conn.is_connecting = true;
286 267111 : }
287 :
288 240555 : fn conn_connect_poll(
289 240555 : &self,
290 240555 : _: &mut walproposer::bindings::Safekeeper,
291 240555 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
292 240555 : debug!("conn_connect_poll");
293 : // TODO: break the connection here
294 240555 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
295 240555 : }
296 :
297 240555 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
298 240555 : debug!("conn_send_query: {}", query);
299 240555 : self.get_conn(sk).is_start_wal_push = true;
300 240555 : true
301 240555 : }
302 :
303 240555 : fn conn_get_query_result(
304 240555 : &self,
305 240555 : _: &mut walproposer::bindings::Safekeeper,
306 240555 : ) -> walproposer::bindings::WalProposerExecStatusType {
307 240555 : debug!("conn_get_query_result");
308 : // TODO: break the connection here
309 240555 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
310 240555 : }
311 :
312 130173 : fn conn_async_read(
313 130173 : &self,
314 130173 : sk: &mut walproposer::bindings::Safekeeper,
315 130173 : vec: &mut Vec<u8>,
316 130173 : ) -> walproposer::bindings::PGAsyncReadResult {
317 130173 : debug!("conn_async_read");
318 130173 : let mut conn = self.get_conn(sk);
319 :
320 130173 : let socket = if let Some(socket) = conn.socket.as_mut() {
321 130173 : socket
322 : } else {
323 : // socket is already closed
324 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
325 : };
326 :
327 130173 : let msg = socket.recv_chan().try_recv();
328 :
329 119462 : match msg {
330 : None => {
331 : // no message is ready
332 10711 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
333 : }
334 : Some(NetEvent::Closed) => {
335 : // connection is closed
336 60789 : debug!("conn_async_read: connection is closed");
337 60789 : conn.socket = None;
338 60789 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
339 : }
340 58673 : Some(NetEvent::Message(msg)) => {
341 : // got a message
342 58673 : let b = match msg {
343 58673 : desim::proto::AnyMessage::Bytes(b) => b,
344 0 : _ => unreachable!(),
345 : };
346 58673 : vec.extend_from_slice(&b);
347 58673 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
348 : }
349 : }
350 130173 : }
351 :
352 268297 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
353 268297 : let mut conn = self.get_conn(sk);
354 268297 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
355 268297 : let socket = conn.socket.as_mut().unwrap();
356 268297 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
357 268297 : true
358 268297 : }
359 :
360 29503 : fn conn_async_write(
361 29503 : &self,
362 29503 : sk: &mut walproposer::bindings::Safekeeper,
363 29503 : buf: &[u8],
364 29503 : ) -> walproposer::bindings::PGAsyncWriteResult {
365 29503 : let mut conn = self.get_conn(sk);
366 29503 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
367 29503 : if let Some(socket) = conn.socket.as_mut() {
368 29503 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
369 29503 : } 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 29503 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
375 29503 : }
376 :
377 6765 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
378 6765 : debug!("wal_reader_allocate");
379 6765 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
380 6765 : }
381 :
382 22738 : fn wal_read(
383 22738 : &self,
384 22738 : _sk: &mut walproposer::bindings::Safekeeper,
385 22738 : buf: &mut [u8],
386 22738 : startpos: u64,
387 22738 : ) -> NeonWALReadResult {
388 22738 : self.disk.lock().read(startpos, buf);
389 22738 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
390 22738 : }
391 :
392 72433 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
393 72433 : debug!("init_event_set");
394 72433 : let new_event_set = EventSet::new(self.os.clone());
395 72433 : let old_event_set = self.event_set.replace(Some(new_event_set));
396 72433 : assert!(old_event_set.is_none());
397 72433 : }
398 :
399 528976 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
400 528976 : debug!(
401 0 : "update_event_set, sk={:?}, events_mask={:#b}",
402 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
403 : );
404 528976 : let conn = self.get_conn(sk);
405 528976 :
406 528976 : self.event_set
407 528976 : .borrow_mut()
408 528976 : .as_mut()
409 528976 : .unwrap()
410 528976 : .update_event_set(&conn, event_mask);
411 528976 : }
412 :
413 481110 : fn add_safekeeper_event_set(
414 481110 : &self,
415 481110 : sk: &mut walproposer::bindings::Safekeeper,
416 481110 : event_mask: u32,
417 481110 : ) {
418 481110 : debug!(
419 0 : "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
420 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
421 : );
422 :
423 481110 : self.event_set
424 481110 : .borrow_mut()
425 481110 : .as_mut()
426 481110 : .unwrap()
427 481110 : .add_safekeeper(&self.get_conn(sk), event_mask);
428 481110 : }
429 :
430 303353 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
431 303353 : debug!(
432 0 : "rm_safekeeper_event_set, sk={:?}",
433 0 : sk as *mut walproposer::bindings::Safekeeper,
434 : );
435 :
436 303353 : self.event_set
437 303353 : .borrow_mut()
438 303353 : .as_mut()
439 303353 : .unwrap()
440 303353 : .remove_safekeeper(&self.get_conn(sk));
441 303353 : }
442 :
443 33449 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
444 33449 : debug!("active_state_update_event_set");
445 :
446 33449 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
447 33449 : self.event_set
448 33449 : .borrow_mut()
449 33449 : .as_mut()
450 33449 : .unwrap()
451 33449 : .refresh_event_set();
452 33449 : }
453 :
454 94159 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
455 94159 : 0
456 94159 : }
457 :
458 695443 : fn wait_event_set(
459 695443 : &self,
460 695443 : _: &mut walproposer::bindings::WalProposer,
461 695443 : timeout_millis: i64,
462 695443 : ) -> walproposer::walproposer::WaitResult {
463 695443 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
464 695443 : let mut conns = self.safekeepers.borrow_mut();
465 1605003 : for conn in conns.iter_mut() {
466 1605003 : if conn.socket.is_some() && conn.is_connecting {
467 240555 : conn.is_connecting = false;
468 240555 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
469 240555 : return walproposer::walproposer::WaitResult::Network(
470 240555 : conn.raw_ptr,
471 240555 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
472 240555 : );
473 1364448 : }
474 1364448 : if conn.socket.is_some() && conn.is_start_wal_push {
475 240555 : conn.is_start_wal_push = false;
476 240555 : debug!(
477 0 : "wait_event_set, start wal push to {}:{}",
478 : conn.host, conn.port
479 : );
480 240555 : return walproposer::walproposer::WaitResult::Network(
481 240555 : conn.raw_ptr,
482 240555 : WL_SOCKET_READABLE,
483 240555 : );
484 1123893 : }
485 : }
486 214333 : drop(conns);
487 214333 :
488 214333 : let res = self
489 214333 : .event_set
490 214333 : .borrow_mut()
491 214333 : .as_mut()
492 214333 : .unwrap()
493 214333 : .wait(timeout_millis);
494 214333 :
495 214333 : debug!(
496 0 : "wait_event_set, timeout_millis={}, res={:?}",
497 : timeout_millis, res,
498 : );
499 145445 : res
500 626555 : }
501 :
502 72433 : fn strong_random(&self, buf: &mut [u8]) -> bool {
503 72433 : debug!("strong_random");
504 72433 : buf.fill(0);
505 72433 : true
506 72433 : }
507 :
508 2572 : fn finish_sync_safekeepers(&self, lsn: u64) {
509 2572 : debug!("finish_sync_safekeepers, lsn={}", lsn);
510 2572 : executor::exit(0, Lsn(lsn).to_string());
511 2572 : }
512 :
513 704645 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
514 704645 : debug!("wp_log[{}] {}", level, msg);
515 704645 : if level == Level::Fatal || level == Level::Panic {
516 514 : if msg.contains("rejects our connection request with term") {
517 291 : // collected quorum with lower term, then got rejected by next connected safekeeper
518 291 : executor::exit(1, msg.to_owned());
519 291 : }
520 514 : if msg.contains("collected propEpochStartLsn") && msg.contains(", but basebackup LSN ")
521 25 : {
522 25 : // sync-safekeepers collected wrong quorum, walproposer collected another quorum
523 25 : executor::exit(1, msg.to_owned());
524 489 : }
525 514 : if msg.contains("failed to download WAL for logical replicaiton") {
526 136 : // Recovery connection broken and recovery was failed
527 136 : executor::exit(1, msg.to_owned());
528 378 : }
529 514 : if msg.contains("missing majority of votes, collected") {
530 62 : // Voting bug when safekeeper disconnects after voting
531 62 : executor::exit(1, msg.to_owned());
532 452 : }
533 514 : panic!("unknown FATAL error from walproposer: {}", msg);
534 704131 : }
535 704131 : }
536 :
537 5174 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
538 5174 : let prop_lsn = wp.propEpochStartLsn;
539 5174 : let prop_term = wp.propTerm;
540 5174 :
541 5174 : let mut prev_lsn: u64 = 0;
542 5174 : let mut prev_term: u64 = 0;
543 5174 :
544 5174 : unsafe {
545 5174 : let history = wp.propTermHistory.entries;
546 5174 : let len = wp.propTermHistory.n_entries as usize;
547 5174 : if len > 1 {
548 3324 : let entry = *history.wrapping_add(len - 2);
549 3324 : prev_lsn = entry.lsn;
550 3324 : prev_term = entry.term;
551 3324 : }
552 : }
553 :
554 5174 : let msg = format!(
555 5174 : "prop_elected;{};{};{};{}",
556 5174 : prop_lsn, prop_term, prev_lsn, prev_term
557 5174 : );
558 5174 :
559 5174 : debug!(msg);
560 5174 : self.os.log_event(msg);
561 5174 : }
562 :
563 1951 : fn get_redo_start_lsn(&self) -> u64 {
564 1951 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
565 1951 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
566 1951 : }
567 :
568 16794 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
569 16794 : self.shmem.get()
570 16794 : }
571 :
572 1089 : fn start_streaming(
573 1089 : &self,
574 1089 : startpos: u64,
575 1089 : callback: &walproposer::walproposer::StreamingCallback,
576 1089 : ) {
577 1089 : let disk = &self.disk;
578 1089 : let disk_lsn = disk.lock().flush_rec_ptr().0;
579 1089 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
580 1089 : if startpos < disk_lsn {
581 320 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
582 769 : }
583 1089 : assert!(startpos <= disk_lsn);
584 1089 : let mut broadcasted = Lsn(startpos);
585 :
586 : loop {
587 3561 : let available = disk.lock().flush_rec_ptr();
588 3561 : assert!(available >= broadcasted);
589 2472 : callback.broadcast(broadcasted, available);
590 2472 : broadcasted = available;
591 2472 : callback.poll();
592 : }
593 : }
594 :
595 11556 : fn process_safekeeper_feedback(
596 11556 : &mut self,
597 11556 : wp: &mut walproposer::bindings::WalProposer,
598 11556 : _sk: &mut walproposer::bindings::Safekeeper,
599 11556 : ) {
600 11556 : debug!("process_safekeeper_feedback, commit_lsn={}", wp.commitLsn);
601 11556 : if wp.commitLsn > self.last_logged_commit_lsn {
602 2498 : self.os.log_event(format!("commit_lsn;{}", wp.commitLsn));
603 2498 : self.last_logged_commit_lsn = wp.commitLsn;
604 9058 : }
605 11556 : }
606 :
607 721 : fn get_flush_rec_ptr(&self) -> u64 {
608 721 : let lsn = self.disk.lock().flush_rec_ptr();
609 721 : debug!("get_flush_rec_ptr: {}", lsn);
610 721 : lsn.0
611 721 : }
612 :
613 5174 : fn recovery_download(
614 5174 : &self,
615 5174 : wp: &mut walproposer::bindings::WalProposer,
616 5174 : sk: &mut walproposer::bindings::Safekeeper,
617 5174 : ) -> bool {
618 5174 : let mut startpos = wp.truncateLsn;
619 5174 : let endpos = wp.propEpochStartLsn;
620 5174 :
621 5174 : if startpos == endpos {
622 3205 : debug!("recovery_download: nothing to download");
623 3205 : return true;
624 1969 : }
625 1969 :
626 1969 : debug!("recovery_download from {} to {}", startpos, endpos,);
627 :
628 1969 : let replication_prompt = format!(
629 1969 : "START_REPLICATION {} {} {} {}",
630 1969 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
631 1969 : );
632 1969 : let async_conn = self.get_conn(sk);
633 1969 :
634 1969 : let conn = self.os.open_tcp(async_conn.node_id);
635 1969 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
636 1969 :
637 1969 : let chan = conn.recv_chan();
638 3343 : while startpos < endpos {
639 1969 : let event = chan.recv();
640 1833 : match event {
641 : NetEvent::Closed => {
642 136 : debug!("connection closed in recovery");
643 136 : break;
644 : }
645 1833 : NetEvent::Message(AnyMessage::Bytes(b)) => {
646 1833 : debug!("got recovery bytes from safekeeper");
647 1374 : self.disk.lock().write(startpos, &b);
648 1374 : startpos += b.len() as u64;
649 : }
650 0 : NetEvent::Message(_) => unreachable!(),
651 : }
652 : }
653 :
654 1510 : debug!("recovery finished at {}", startpos);
655 :
656 1510 : startpos == endpos
657 4715 : }
658 :
659 89354 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
660 89354 : let mut conn = self.get_conn(sk);
661 89354 : debug!("conn_finish to {}", conn.node_id);
662 89354 : if let Some(socket) = conn.socket.as_mut() {
663 28529 : socket.close();
664 60825 : } else {
665 60825 : // connection is already closed
666 60825 : }
667 89354 : conn.socket = None;
668 89354 : }
669 :
670 87345 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
671 87345 : "connection is closed, probably".into()
672 87345 : }
673 : }
|
