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 220062 : pub fn new(host: String, port: String) -> Self {
45 220062 : // port number is the same as NodeId
46 220062 : let port_num = port.parse::<u32>().unwrap();
47 220062 : Self {
48 220062 : host,
49 220062 : port,
50 220062 : node_id: port_num,
51 220062 : socket: None,
52 220062 : is_connecting: false,
53 220062 : is_start_wal_push: false,
54 220062 : raw_ptr: std::ptr::null_mut(),
55 220062 : }
56 220062 : }
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 73354 : pub fn new(os: NodeOs) -> Self {
73 73354 : let node_events = os.node_events();
74 73354 : Self {
75 73354 : os,
76 73354 : chans: vec![Box::new(node_events)],
77 73354 : sk_ptrs: vec![std::ptr::null_mut()],
78 73354 : masks: vec![WL_SOCKET_READABLE],
79 73354 : }
80 73354 : }
81 :
82 : /// Leaves all readable channels at the beginning of the array.
83 219809 : fn sort_readable(&mut self) -> usize {
84 219809 : let mut cnt = 1;
85 521783 : for i in 1..self.chans.len() {
86 521783 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
87 521783 : self.chans.swap(i, cnt);
88 521783 : self.sk_ptrs.swap(i, cnt);
89 521783 : self.masks.swap(i, cnt);
90 521783 : cnt += 1;
91 521783 : }
92 : }
93 219809 : cnt
94 219809 : }
95 :
96 536290 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
97 536290 : let index = self
98 536290 : .sk_ptrs
99 536290 : .iter()
100 2026279 : .position(|&ptr| ptr == conn.raw_ptr)
101 536290 : .expect("safekeeper should exist in event set");
102 536290 : self.masks[index] = event_mask;
103 536290 : }
104 :
105 486298 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
106 1177206 : for ptr in self.sk_ptrs.iter() {
107 1177206 : assert!(*ptr != sk.raw_ptr);
108 : }
109 :
110 486298 : self.chans.push(Box::new(
111 486298 : sk.socket
112 486298 : .as_ref()
113 486298 : .expect("socket should not be closed")
114 486298 : .recv_chan(),
115 486298 : ));
116 486298 : self.sk_ptrs.push(sk.raw_ptr);
117 486298 : self.masks.push(event_mask);
118 486298 : }
119 :
120 305807 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
121 650501 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
122 305807 : if index.is_none() {
123 47 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
124 47 : return;
125 305760 : }
126 305760 : let index = index.unwrap();
127 305760 :
128 305760 : self.chans.remove(index);
129 305760 : self.sk_ptrs.remove(index);
130 305760 : self.masks.remove(index);
131 305760 :
132 305760 : // to simulate the actual behaviour
133 305760 : self.refresh_event_set();
134 305807 : }
135 :
136 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
137 344576 : fn refresh_event_set(&mut self) {
138 971966 : for (i, mask) in self.masks.iter_mut().enumerate() {
139 971966 : if i == 0 {
140 344576 : continue;
141 627390 : }
142 627390 :
143 627390 : let mut mask_sk: u32 = 0;
144 627390 : let mut mask_nwr: u32 = 0;
145 627390 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
146 627390 :
147 627390 : 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 627390 : }
154 : }
155 344576 : }
156 :
157 : /// Wait for events on all channels.
158 219809 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
159 : // all channels are always writeable
160 741592 : for (i, mask) in self.masks.iter().enumerate() {
161 741592 : 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 741592 : }
167 : }
168 :
169 219809 : let cnt = self.sort_readable();
170 219809 :
171 219809 : let slice = &self.chans[0..cnt];
172 219809 : match executor::epoll_chans(slice, timeout_millis) {
173 91681 : None => walproposer::walproposer::WaitResult::Timeout,
174 : Some(0) => {
175 2981 : let msg = self.os.node_events().must_recv();
176 2981 : match msg {
177 2981 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
178 2981 : // got a notification about new WAL available
179 2981 : }
180 0 : NodeEvent::Internal(_) => unreachable!(),
181 0 : NodeEvent::Accept(_) => unreachable!(),
182 : }
183 2981 : walproposer::walproposer::WaitResult::Latch
184 : }
185 125147 : Some(index) => walproposer::walproposer::WaitResult::Network(
186 125147 : self.sk_ptrs[index],
187 125147 : WL_SOCKET_READABLE,
188 125147 : ),
189 : }
190 219809 : }
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 73354 : pub fn new(args: Args) -> Self {
213 73354 : // initialize connection state for each safekeeper
214 73354 : let sk_conns = args
215 73354 : .config
216 73354 : .safekeepers_list
217 73354 : .iter()
218 220062 : .map(|s| {
219 220062 : SafekeeperConn::new(
220 220062 : s.split(':').next().unwrap().to_string(),
221 220062 : s.split(':').nth(1).unwrap().to_string(),
222 220062 : )
223 220062 : })
224 73354 : .collect::<Vec<_>>();
225 73354 :
226 73354 : Self {
227 73354 : os: args.os,
228 73354 : safekeepers: RefCell::new(sk_conns),
229 73354 : disk: args.disk,
230 73354 : redo_start_lsn: args.redo_start_lsn,
231 73354 : shmem: UnsafeCell::new(walproposer::bindings::WalproposerShmemState {
232 73354 : mutex: 0,
233 73354 : feedback: PageserverFeedback {
234 73354 : currentClusterSize: 0,
235 73354 : last_received_lsn: 0,
236 73354 : disk_consistent_lsn: 0,
237 73354 : remote_consistent_lsn: 0,
238 73354 : replytime: 0,
239 73354 : },
240 73354 : mineLastElectedTerm: 0,
241 73354 : backpressureThrottlingTime: pg_atomic_uint64 { value: 0 },
242 73354 : }),
243 73354 : config: args.config,
244 73354 : event_set: RefCell::new(None),
245 73354 : }
246 73354 : }
247 :
248 : /// Get SafekeeperConn for the given Safekeeper.
249 2375912 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
250 2375912 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
251 2375912 : let state = self.safekeepers.borrow_mut();
252 2375912 : RefMut::map(state, |v| {
253 2375912 : v.iter_mut()
254 4747414 : .find(|conn| conn.port == sk_port)
255 2375912 : .expect("safekeeper conn not found by port")
256 2375912 : })
257 2375912 : }
258 : }
259 :
260 : impl ApiImpl for SimulationApi {
261 2534073 : fn get_current_timestamp(&self) -> i64 {
262 2534073 : debug!("get_current_timestamp");
263 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
264 : // milliseconds, so add 10^3
265 2534073 : self.os.now() as i64 * 1000
266 2534073 : }
267 :
268 270013 : fn conn_status(
269 270013 : &self,
270 270013 : _: &mut walproposer::bindings::Safekeeper,
271 270013 : ) -> walproposer::bindings::WalProposerConnStatusType {
272 270013 : debug!("conn_status");
273 : // break the connection with a 10% chance
274 270013 : if self.os.random(100) < 10 {
275 26864 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
276 : } else {
277 243149 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
278 : }
279 270013 : }
280 :
281 270013 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
282 270013 : debug!("conn_connect_start");
283 270013 : let mut conn = self.get_conn(sk);
284 270013 :
285 270013 : assert!(conn.socket.is_none());
286 270013 : let socket = self.os.open_tcp(conn.node_id);
287 270013 : conn.socket = Some(socket);
288 270013 : conn.raw_ptr = sk;
289 270013 : conn.is_connecting = true;
290 270013 : }
291 :
292 243149 : fn conn_connect_poll(
293 243149 : &self,
294 243149 : _: &mut walproposer::bindings::Safekeeper,
295 243149 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
296 243149 : debug!("conn_connect_poll");
297 : // TODO: break the connection here
298 243149 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
299 243149 : }
300 :
301 243149 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
302 243149 : debug!("conn_send_query: {}", query);
303 243149 : self.get_conn(sk).is_start_wal_push = true;
304 243149 : true
305 243149 : }
306 :
307 243149 : fn conn_get_query_result(
308 243149 : &self,
309 243149 : _: &mut walproposer::bindings::Safekeeper,
310 243149 : ) -> walproposer::bindings::WalProposerExecStatusType {
311 243149 : debug!("conn_get_query_result");
312 : // TODO: break the connection here
313 243149 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
314 243149 : }
315 :
316 136613 : fn conn_async_read(
317 136613 : &self,
318 136613 : sk: &mut walproposer::bindings::Safekeeper,
319 136613 : vec: &mut Vec<u8>,
320 136613 : ) -> walproposer::bindings::PGAsyncReadResult {
321 136613 : debug!("conn_async_read");
322 136613 : let mut conn = self.get_conn(sk);
323 :
324 136613 : let socket = if let Some(socket) = conn.socket.as_mut() {
325 136613 : socket
326 : } else {
327 : // socket is already closed
328 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
329 : };
330 :
331 136613 : let msg = socket.recv_chan().try_recv();
332 :
333 123581 : match msg {
334 : None => {
335 : // no message is ready
336 13032 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
337 : }
338 : Some(NetEvent::Closed) => {
339 : // connection is closed
340 60526 : debug!("conn_async_read: connection is closed");
341 60526 : conn.socket = None;
342 60526 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
343 : }
344 63055 : Some(NetEvent::Message(msg)) => {
345 : // got a message
346 63055 : let b = match msg {
347 63055 : desim::proto::AnyMessage::Bytes(b) => b,
348 0 : _ => unreachable!(),
349 : };
350 63055 : vec.extend_from_slice(&b);
351 63055 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
352 : }
353 : }
354 136613 : }
355 :
356 272086 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
357 272086 : let mut conn = self.get_conn(sk);
358 272086 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
359 272086 : let socket = conn.socket.as_mut().unwrap();
360 272086 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
361 272086 : true
362 272086 : }
363 :
364 33923 : fn conn_async_write(
365 33923 : &self,
366 33923 : sk: &mut walproposer::bindings::Safekeeper,
367 33923 : buf: &[u8],
368 33923 : ) -> walproposer::bindings::PGAsyncWriteResult {
369 33923 : let mut conn = self.get_conn(sk);
370 33923 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
371 33923 : if let Some(socket) = conn.socket.as_mut() {
372 33923 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
373 33923 : } 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 33923 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
379 33923 : }
380 :
381 7643 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
382 7643 : debug!("wal_reader_allocate");
383 7643 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
384 7643 : }
385 :
386 26280 : fn wal_read(
387 26280 : &self,
388 26280 : _sk: &mut walproposer::bindings::Safekeeper,
389 26280 : buf: &mut [u8],
390 26280 : startpos: u64,
391 26280 : ) -> NeonWALReadResult {
392 26280 : self.disk.lock().read(startpos, buf);
393 26280 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
394 26280 : }
395 :
396 73354 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
397 73354 : debug!("init_event_set");
398 73354 : let new_event_set = EventSet::new(self.os.clone());
399 73354 : let old_event_set = self.event_set.replace(Some(new_event_set));
400 73354 : assert!(old_event_set.is_none());
401 73354 : }
402 :
403 536290 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
404 536290 : debug!(
405 1166 : "update_event_set, sk={:?}, events_mask={:#b}",
406 1166 : sk as *mut walproposer::bindings::Safekeeper, event_mask
407 1166 : );
408 536290 : let conn = self.get_conn(sk);
409 536290 :
410 536290 : self.event_set
411 536290 : .borrow_mut()
412 536290 : .as_mut()
413 536290 : .unwrap()
414 536290 : .update_event_set(&conn, event_mask);
415 536290 : }
416 :
417 486298 : fn add_safekeeper_event_set(
418 486298 : &self,
419 486298 : sk: &mut walproposer::bindings::Safekeeper,
420 486298 : event_mask: u32,
421 486298 : ) {
422 486298 : 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 486298 : self.event_set
428 486298 : .borrow_mut()
429 486298 : .as_mut()
430 486298 : .unwrap()
431 486298 : .add_safekeeper(&self.get_conn(sk), event_mask);
432 486298 : }
433 :
434 305807 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
435 305807 : debug!(
436 400 : "rm_safekeeper_event_set, sk={:?}",
437 400 : sk as *mut walproposer::bindings::Safekeeper,
438 400 : );
439 :
440 305807 : self.event_set
441 305807 : .borrow_mut()
442 305807 : .as_mut()
443 305807 : .unwrap()
444 305807 : .remove_safekeeper(&self.get_conn(sk));
445 305807 : }
446 :
447 38816 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
448 38816 : debug!("active_state_update_event_set");
449 :
450 38816 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
451 38816 : self.event_set
452 38816 : .borrow_mut()
453 38816 : .as_mut()
454 38816 : .unwrap()
455 38816 : .refresh_event_set();
456 38816 : }
457 :
458 110114 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
459 110114 : 0
460 110114 : }
461 :
462 706107 : fn wait_event_set(
463 706107 : &self,
464 706107 : _: &mut walproposer::bindings::WalProposer,
465 706107 : timeout_millis: i64,
466 706107 : ) -> walproposer::walproposer::WaitResult {
467 706107 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
468 706107 : let mut conns = self.safekeepers.borrow_mut();
469 1631987 : for conn in conns.iter_mut() {
470 1631987 : if conn.socket.is_some() && conn.is_connecting {
471 243149 : conn.is_connecting = false;
472 243149 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
473 243149 : return walproposer::walproposer::WaitResult::Network(
474 243149 : conn.raw_ptr,
475 243149 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
476 243149 : );
477 1388838 : }
478 1388838 : if conn.socket.is_some() && conn.is_start_wal_push {
479 243149 : conn.is_start_wal_push = false;
480 243149 : debug!(
481 364 : "wait_event_set, start wal push to {}:{}",
482 364 : conn.host, conn.port
483 364 : );
484 243149 : return walproposer::walproposer::WaitResult::Network(
485 243149 : conn.raw_ptr,
486 243149 : WL_SOCKET_READABLE,
487 243149 : );
488 1145689 : }
489 : }
490 219809 : drop(conns);
491 219809 :
492 219809 : let res = self
493 219809 : .event_set
494 219809 : .borrow_mut()
495 219809 : .as_mut()
496 219809 : .unwrap()
497 219809 : .wait(timeout_millis);
498 219809 :
499 219809 : debug!(
500 690 : "wait_event_set, timeout_millis={}, res={:?}",
501 690 : timeout_millis, res,
502 690 : );
503 150241 : res
504 636539 : }
505 :
506 73354 : fn strong_random(&self, buf: &mut [u8]) -> bool {
507 73354 : debug!("strong_random");
508 73354 : buf.fill(0);
509 73354 : true
510 73354 : }
511 :
512 2757 : fn finish_sync_safekeepers(&self, lsn: u64) {
513 2757 : debug!("finish_sync_safekeepers, lsn={}", lsn);
514 2757 : executor::exit(0, Lsn(lsn).to_string());
515 2757 : }
516 :
517 788290 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
518 788290 : debug!("wp_log[{}] {}", level, msg);
519 788290 : if level == Level::Fatal || level == Level::Panic {
520 497 : if msg.contains("rejects our connection request with term") {
521 282 : // collected quorum with lower term, then got rejected by next connected safekeeper
522 282 : executor::exit(1, msg.to_owned());
523 282 : }
524 497 : if msg.contains("collected propEpochStartLsn") && msg.contains(", but basebackup LSN ")
525 22 : {
526 22 : // sync-safekeepers collected wrong quorum, walproposer collected another quorum
527 22 : executor::exit(1, msg.to_owned());
528 475 : }
529 497 : if msg.contains("failed to download WAL for logical replicaiton") {
530 107 : // Recovery connection broken and recovery was failed
531 107 : executor::exit(1, msg.to_owned());
532 390 : }
533 497 : if msg.contains("missing majority of votes, collected") {
534 86 : // Voting bug when safekeeper disconnects after voting
535 86 : executor::exit(1, msg.to_owned());
536 411 : }
537 497 : panic!("unknown FATAL error from walproposer: {}", msg);
538 787793 : }
539 787793 : }
540 :
541 5633 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
542 5633 : let prop_lsn = wp.propEpochStartLsn;
543 5633 : let prop_term = wp.propTerm;
544 5633 :
545 5633 : let mut prev_lsn: u64 = 0;
546 5633 : let mut prev_term: u64 = 0;
547 5633 :
548 5633 : unsafe {
549 5633 : let history = wp.propTermHistory.entries;
550 5633 : let len = wp.propTermHistory.n_entries as usize;
551 5633 : if len > 1 {
552 3757 : let entry = *history.wrapping_add(len - 2);
553 3757 : prev_lsn = entry.lsn;
554 3757 : prev_term = entry.term;
555 3757 : }
556 : }
557 :
558 5633 : let msg = format!(
559 5633 : "prop_elected;{};{};{};{}",
560 5633 : prop_lsn, prop_term, prev_lsn, prev_term
561 5633 : );
562 5633 :
563 5633 : debug!(msg);
564 5633 : self.os.log_event(msg);
565 5633 : }
566 :
567 2086 : fn get_redo_start_lsn(&self) -> u64 {
568 2086 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
569 2086 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
570 2086 : }
571 :
572 1264 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
573 1264 : self.shmem.get()
574 1264 : }
575 :
576 1214 : fn start_streaming(
577 1214 : &self,
578 1214 : startpos: u64,
579 1214 : callback: &walproposer::walproposer::StreamingCallback,
580 1214 : ) {
581 1214 : let disk = &self.disk;
582 1214 : let disk_lsn = disk.lock().flush_rec_ptr().0;
583 1214 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
584 1214 : if startpos < disk_lsn {
585 341 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
586 873 : }
587 1214 : assert!(startpos <= disk_lsn);
588 1214 : let mut broadcasted = Lsn(startpos);
589 :
590 : loop {
591 4046 : let available = disk.lock().flush_rec_ptr();
592 4046 : assert!(available >= broadcasted);
593 2832 : callback.broadcast(broadcasted, available);
594 2832 : broadcasted = available;
595 2832 : callback.poll();
596 : }
597 : }
598 :
599 13032 : fn process_safekeeper_feedback(
600 13032 : &self,
601 13032 : wp: &mut walproposer::bindings::WalProposer,
602 13032 : commit_lsn: u64,
603 13032 : ) {
604 13032 : debug!("process_safekeeper_feedback, commit_lsn={}", commit_lsn);
605 13032 : if commit_lsn > wp.lastSentCommitLsn {
606 2845 : self.os.log_event(format!("commit_lsn;{}", commit_lsn));
607 10187 : }
608 13032 : }
609 :
610 817 : fn get_flush_rec_ptr(&self) -> u64 {
611 817 : let lsn = self.disk.lock().flush_rec_ptr();
612 817 : debug!("get_flush_rec_ptr: {}", lsn);
613 817 : lsn.0
614 817 : }
615 :
616 5633 : fn recovery_download(
617 5633 : &self,
618 5633 : wp: &mut walproposer::bindings::WalProposer,
619 5633 : sk: &mut walproposer::bindings::Safekeeper,
620 5633 : ) -> bool {
621 5633 : let mut startpos = wp.truncateLsn;
622 5633 : let endpos = wp.propEpochStartLsn;
623 5633 :
624 5633 : if startpos == endpos {
625 3422 : debug!("recovery_download: nothing to download");
626 3422 : return true;
627 2211 : }
628 2211 :
629 2211 : debug!("recovery_download from {} to {}", startpos, endpos,);
630 :
631 2211 : let replication_prompt = format!(
632 2211 : "START_REPLICATION {} {} {} {}",
633 2211 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
634 2211 : );
635 2211 : let async_conn = self.get_conn(sk);
636 2211 :
637 2211 : let conn = self.os.open_tcp(async_conn.node_id);
638 2211 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
639 2211 :
640 2211 : let chan = conn.recv_chan();
641 3783 : while startpos < endpos {
642 2211 : let event = chan.recv();
643 2104 : match event {
644 : NetEvent::Closed => {
645 107 : debug!("connection closed in recovery");
646 107 : break;
647 : }
648 2104 : NetEvent::Message(AnyMessage::Bytes(b)) => {
649 2104 : debug!("got recovery bytes from safekeeper");
650 1572 : self.disk.lock().write(startpos, &b);
651 1572 : startpos += b.len() as u64;
652 : }
653 0 : NetEvent::Message(_) => unreachable!(),
654 : }
655 : }
656 :
657 1679 : debug!("recovery finished at {}", startpos);
658 :
659 1679 : startpos == endpos
660 5101 : }
661 :
662 89522 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
663 89522 : let mut conn = self.get_conn(sk);
664 89522 : debug!("conn_finish to {}", conn.node_id);
665 89522 : if let Some(socket) = conn.socket.as_mut() {
666 28949 : socket.close();
667 60573 : } else {
668 60573 : // connection is already closed
669 60573 : }
670 89522 : conn.socket = None;
671 89522 : }
672 :
673 87390 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
674 87390 : "connection is closed, probably".into()
675 87390 : }
676 : }
|
