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 216276 : pub fn new(host: String, port: String) -> Self {
45 216276 : // port number is the same as NodeId
46 216276 : let port_num = port.parse::<u32>().unwrap();
47 216276 : Self {
48 216276 : host,
49 216276 : port,
50 216276 : node_id: port_num,
51 216276 : socket: None,
52 216276 : is_connecting: false,
53 216276 : is_start_wal_push: false,
54 216276 : raw_ptr: std::ptr::null_mut(),
55 216276 : }
56 216276 : }
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 72092 : pub fn new(os: NodeOs) -> Self {
73 72092 : let node_events = os.node_events();
74 72092 : Self {
75 72092 : os,
76 72092 : chans: vec![Box::new(node_events)],
77 72092 : sk_ptrs: vec![std::ptr::null_mut()],
78 72092 : masks: vec![WL_SOCKET_READABLE],
79 72092 : }
80 72092 : }
81 :
82 : /// Leaves all readable channels at the beginning of the array.
83 214063 : fn sort_readable(&mut self) -> usize {
84 214063 : let mut cnt = 1;
85 507183 : for i in 1..self.chans.len() {
86 507183 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
87 507183 : self.chans.swap(i, cnt);
88 507183 : self.sk_ptrs.swap(i, cnt);
89 507183 : self.masks.swap(i, cnt);
90 507183 : cnt += 1;
91 507183 : }
92 : }
93 214063 : cnt
94 214063 : }
95 :
96 523833 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
97 523833 : let index = self
98 523833 : .sk_ptrs
99 523833 : .iter()
100 1980683 : .position(|&ptr| ptr == conn.raw_ptr)
101 523833 : .expect("safekeeper should exist in event set");
102 523833 : self.masks[index] = event_mask;
103 523833 : }
104 :
105 477114 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
106 1154740 : for ptr in self.sk_ptrs.iter() {
107 1154740 : assert!(*ptr != sk.raw_ptr);
108 : }
109 :
110 477114 : self.chans.push(Box::new(
111 477114 : sk.socket
112 477114 : .as_ref()
113 477114 : .expect("socket should not be closed")
114 477114 : .recv_chan(),
115 477114 : ));
116 477114 : self.sk_ptrs.push(sk.raw_ptr);
117 477114 : self.masks.push(event_mask);
118 477114 : }
119 :
120 299774 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
121 637238 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
122 299774 : if index.is_none() {
123 54 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
124 54 : return;
125 299720 : }
126 299720 : let index = index.unwrap();
127 299720 :
128 299720 : self.chans.remove(index);
129 299720 : self.sk_ptrs.remove(index);
130 299720 : self.masks.remove(index);
131 299720 :
132 299720 : // to simulate the actual behaviour
133 299720 : self.refresh_event_set();
134 299774 : }
135 :
136 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
137 338560 : fn refresh_event_set(&mut self) {
138 955824 : for (i, mask) in self.masks.iter_mut().enumerate() {
139 955824 : if i == 0 {
140 338560 : continue;
141 617264 : }
142 617264 :
143 617264 : let mut mask_sk: u32 = 0;
144 617264 : let mut mask_nwr: u32 = 0;
145 617264 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
146 617264 :
147 617264 : 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 617264 : }
154 : }
155 338560 : }
156 :
157 : /// Wait for events on all channels.
158 214063 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
159 : // all channels are always writeable
160 721246 : for (i, mask) in self.masks.iter().enumerate() {
161 721246 : 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 721246 : }
167 : }
168 :
169 214063 : let cnt = self.sort_readable();
170 214063 :
171 214063 : let slice = &self.chans[0..cnt];
172 214063 : match executor::epoll_chans(slice, timeout_millis) {
173 90196 : None => walproposer::walproposer::WaitResult::Timeout,
174 : Some(0) => {
175 2802 : let msg = self.os.node_events().must_recv();
176 2802 : match msg {
177 2802 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
178 2802 : // got a notification about new WAL available
179 2802 : }
180 0 : NodeEvent::Internal(_) => unreachable!(),
181 0 : NodeEvent::Accept(_) => unreachable!(),
182 : }
183 2802 : walproposer::walproposer::WaitResult::Latch
184 : }
185 121065 : Some(index) => walproposer::walproposer::WaitResult::Network(
186 121065 : self.sk_ptrs[index],
187 121065 : WL_SOCKET_READABLE,
188 121065 : ),
189 : }
190 214063 : }
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 72092 : pub fn new(args: Args) -> Self {
213 72092 : // initialize connection state for each safekeeper
214 72092 : let sk_conns = args
215 72092 : .config
216 72092 : .safekeepers_list
217 72092 : .iter()
218 216276 : .map(|s| {
219 216276 : SafekeeperConn::new(
220 216276 : s.split(':').next().unwrap().to_string(),
221 216276 : s.split(':').nth(1).unwrap().to_string(),
222 216276 : )
223 216276 : })
224 72092 : .collect::<Vec<_>>();
225 72092 :
226 72092 : Self {
227 72092 : os: args.os,
228 72092 : safekeepers: RefCell::new(sk_conns),
229 72092 : disk: args.disk,
230 72092 : redo_start_lsn: args.redo_start_lsn,
231 72092 : shmem: UnsafeCell::new(walproposer::bindings::WalproposerShmemState {
232 72092 : mutex: 0,
233 72092 : feedback: PageserverFeedback {
234 72092 : currentClusterSize: 0,
235 72092 : last_received_lsn: 0,
236 72092 : disk_consistent_lsn: 0,
237 72092 : remote_consistent_lsn: 0,
238 72092 : replytime: 0,
239 72092 : },
240 72092 : mineLastElectedTerm: 0,
241 72092 : backpressureThrottlingTime: pg_atomic_uint64 { value: 0 },
242 72092 : }),
243 72092 : config: args.config,
244 72092 : event_set: RefCell::new(None),
245 72092 : }
246 72092 : }
247 :
248 : /// Get SafekeeperConn for the given Safekeeper.
249 2325798 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
250 2325798 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
251 2325798 : let state = self.safekeepers.borrow_mut();
252 2325798 : RefMut::map(state, |v| {
253 2325798 : v.iter_mut()
254 4647336 : .find(|conn| conn.port == sk_port)
255 2325798 : .expect("safekeeper conn not found by port")
256 2325798 : })
257 2325798 : }
258 : }
259 :
260 : impl ApiImpl for SimulationApi {
261 2479853 : fn get_current_timestamp(&self) -> i64 {
262 2479853 : debug!("get_current_timestamp");
263 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
264 : // milliseconds, so add 10^3
265 2479853 : self.os.now() as i64 * 1000
266 2479853 : }
267 :
268 264955 : fn conn_status(
269 264955 : &self,
270 264955 : _: &mut walproposer::bindings::Safekeeper,
271 264955 : ) -> walproposer::bindings::WalProposerConnStatusType {
272 264955 : debug!("conn_status");
273 : // break the connection with a 10% chance
274 264955 : if self.os.random(100) < 10 {
275 26398 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
276 : } else {
277 238557 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
278 : }
279 264955 : }
280 :
281 264955 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
282 264955 : debug!("conn_connect_start");
283 264955 : let mut conn = self.get_conn(sk);
284 264955 :
285 264955 : assert!(conn.socket.is_none());
286 264955 : let socket = self.os.open_tcp(conn.node_id);
287 264955 : conn.socket = Some(socket);
288 264955 : conn.raw_ptr = sk;
289 264955 : conn.is_connecting = true;
290 264955 : }
291 :
292 238557 : fn conn_connect_poll(
293 238557 : &self,
294 238557 : _: &mut walproposer::bindings::Safekeeper,
295 238557 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
296 238557 : debug!("conn_connect_poll");
297 : // TODO: break the connection here
298 238557 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
299 238557 : }
300 :
301 238557 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
302 238557 : debug!("conn_send_query: {}", query);
303 238557 : self.get_conn(sk).is_start_wal_push = true;
304 238557 : true
305 238557 : }
306 :
307 238557 : fn conn_get_query_result(
308 238557 : &self,
309 238557 : _: &mut walproposer::bindings::Safekeeper,
310 238557 : ) -> walproposer::bindings::WalProposerExecStatusType {
311 238557 : debug!("conn_get_query_result");
312 : // TODO: break the connection here
313 238557 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
314 238557 : }
315 :
316 132975 : fn conn_async_read(
317 132975 : &self,
318 132975 : sk: &mut walproposer::bindings::Safekeeper,
319 132975 : vec: &mut Vec<u8>,
320 132975 : ) -> walproposer::bindings::PGAsyncReadResult {
321 132975 : debug!("conn_async_read");
322 132975 : let mut conn = self.get_conn(sk);
323 :
324 132975 : let socket = if let Some(socket) = conn.socket.as_mut() {
325 132975 : socket
326 : } else {
327 : // socket is already closed
328 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
329 : };
330 :
331 132975 : let msg = socket.recv_chan().try_recv();
332 :
333 119598 : match msg {
334 : None => {
335 : // no message is ready
336 13377 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
337 : }
338 : Some(NetEvent::Closed) => {
339 : // connection is closed
340 59169 : debug!("conn_async_read: connection is closed");
341 59169 : conn.socket = None;
342 59169 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
343 : }
344 60429 : Some(NetEvent::Message(msg)) => {
345 : // got a message
346 60429 : let b = match msg {
347 60429 : desim::proto::AnyMessage::Bytes(b) => b,
348 0 : _ => unreachable!(),
349 : };
350 60429 : vec.extend_from_slice(&b);
351 60429 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
352 : }
353 : }
354 132975 : }
355 :
356 265479 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
357 265479 : let mut conn = self.get_conn(sk);
358 265479 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
359 265479 : let socket = conn.socket.as_mut().unwrap();
360 265479 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
361 265479 : true
362 265479 : }
363 :
364 33463 : fn conn_async_write(
365 33463 : &self,
366 33463 : sk: &mut walproposer::bindings::Safekeeper,
367 33463 : buf: &[u8],
368 33463 : ) -> walproposer::bindings::PGAsyncWriteResult {
369 33463 : let mut conn = self.get_conn(sk);
370 33463 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
371 33463 : if let Some(socket) = conn.socket.as_mut() {
372 33463 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
373 33463 : } 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 33463 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
379 33463 : }
380 :
381 7504 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
382 7504 : debug!("wal_reader_allocate");
383 7504 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
384 7504 : }
385 :
386 25959 : fn wal_read(
387 25959 : &self,
388 25959 : _sk: &mut walproposer::bindings::Safekeeper,
389 25959 : buf: &mut [u8],
390 25959 : startpos: u64,
391 25959 : ) -> NeonWALReadResult {
392 25959 : self.disk.lock().read(startpos, buf);
393 25959 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
394 25959 : }
395 :
396 72092 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
397 72092 : debug!("init_event_set");
398 72092 : let new_event_set = EventSet::new(self.os.clone());
399 72092 : let old_event_set = self.event_set.replace(Some(new_event_set));
400 72092 : assert!(old_event_set.is_none());
401 72092 : }
402 :
403 523833 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
404 523833 : debug!(
405 1166 : "update_event_set, sk={:?}, events_mask={:#b}",
406 1166 : sk as *mut walproposer::bindings::Safekeeper, event_mask
407 1166 : );
408 523833 : let conn = self.get_conn(sk);
409 523833 :
410 523833 : self.event_set
411 523833 : .borrow_mut()
412 523833 : .as_mut()
413 523833 : .unwrap()
414 523833 : .update_event_set(&conn, event_mask);
415 523833 : }
416 :
417 477114 : fn add_safekeeper_event_set(
418 477114 : &self,
419 477114 : sk: &mut walproposer::bindings::Safekeeper,
420 477114 : event_mask: u32,
421 477114 : ) {
422 477114 : 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 477114 : self.event_set
428 477114 : .borrow_mut()
429 477114 : .as_mut()
430 477114 : .unwrap()
431 477114 : .add_safekeeper(&self.get_conn(sk), event_mask);
432 477114 : }
433 :
434 299774 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
435 299774 : debug!(
436 400 : "rm_safekeeper_event_set, sk={:?}",
437 400 : sk as *mut walproposer::bindings::Safekeeper,
438 400 : );
439 :
440 299774 : self.event_set
441 299774 : .borrow_mut()
442 299774 : .as_mut()
443 299774 : .unwrap()
444 299774 : .remove_safekeeper(&self.get_conn(sk));
445 299774 : }
446 :
447 38840 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
448 38840 : debug!("active_state_update_event_set");
449 :
450 38840 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
451 38840 : self.event_set
452 38840 : .borrow_mut()
453 38840 : .as_mut()
454 38840 : .unwrap()
455 38840 : .refresh_event_set();
456 38840 : }
457 :
458 110467 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
459 110467 : 0
460 110467 : }
461 :
462 691177 : fn wait_event_set(
463 691177 : &self,
464 691177 : _: &mut walproposer::bindings::WalProposer,
465 691177 : timeout_millis: i64,
466 691177 : ) -> walproposer::walproposer::WaitResult {
467 691177 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
468 691177 : let mut conns = self.safekeepers.borrow_mut();
469 1596405 : for conn in conns.iter_mut() {
470 1596405 : if conn.socket.is_some() && conn.is_connecting {
471 238557 : conn.is_connecting = false;
472 238557 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
473 238557 : return walproposer::walproposer::WaitResult::Network(
474 238557 : conn.raw_ptr,
475 238557 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
476 238557 : );
477 1357848 : }
478 1357848 : if conn.socket.is_some() && conn.is_start_wal_push {
479 238557 : conn.is_start_wal_push = false;
480 238557 : debug!(
481 364 : "wait_event_set, start wal push to {}:{}",
482 364 : conn.host, conn.port
483 364 : );
484 238557 : return walproposer::walproposer::WaitResult::Network(
485 238557 : conn.raw_ptr,
486 238557 : WL_SOCKET_READABLE,
487 238557 : );
488 1119291 : }
489 : }
490 214063 : drop(conns);
491 214063 :
492 214063 : let res = self
493 214063 : .event_set
494 214063 : .borrow_mut()
495 214063 : .as_mut()
496 214063 : .unwrap()
497 214063 : .wait(timeout_millis);
498 214063 :
499 214063 : debug!(
500 690 : "wait_event_set, timeout_millis={}, res={:?}",
501 690 : timeout_millis, res,
502 690 : );
503 145530 : res
504 622644 : }
505 :
506 72092 : fn strong_random(&self, buf: &mut [u8]) -> bool {
507 72092 : debug!("strong_random");
508 72092 : buf.fill(0);
509 72092 : true
510 72092 : }
511 :
512 2594 : fn finish_sync_safekeepers(&self, lsn: u64) {
513 2594 : debug!("finish_sync_safekeepers, lsn={}", lsn);
514 2594 : executor::exit(0, Lsn(lsn).to_string());
515 2594 : }
516 :
517 772743 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
518 772743 : debug!("wp_log[{}] {}", level, msg);
519 772743 : if level == Level::Fatal || level == Level::Panic {
520 504 : if msg.contains("rejects our connection request with term") {
521 310 : // collected quorum with lower term, then got rejected by next connected safekeeper
522 310 : executor::exit(1, msg.to_owned());
523 310 : }
524 504 : 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());
528 482 : }
529 504 : if msg.contains("failed to download WAL for logical replicaiton") {
530 99 : // Recovery connection broken and recovery was failed
531 99 : executor::exit(1, msg.to_owned());
532 405 : }
533 504 : if msg.contains("missing majority of votes, collected") {
534 73 : // Voting bug when safekeeper disconnects after voting
535 73 : executor::exit(1, msg.to_owned());
536 431 : }
537 504 : panic!("unknown FATAL error from walproposer: {}", msg);
538 772239 : }
539 772239 : }
540 :
541 5334 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
542 5334 : let prop_lsn = wp.propEpochStartLsn;
543 5334 : let prop_term = wp.propTerm;
544 5334 :
545 5334 : let mut prev_lsn: u64 = 0;
546 5334 : let mut prev_term: u64 = 0;
547 5334 :
548 5334 : unsafe {
549 5334 : let history = wp.propTermHistory.entries;
550 5334 : let len = wp.propTermHistory.n_entries as usize;
551 5334 : if len > 1 {
552 3615 : let entry = *history.wrapping_add(len - 2);
553 3615 : prev_lsn = entry.lsn;
554 3615 : prev_term = entry.term;
555 3615 : }
556 : }
557 :
558 5334 : let msg = format!(
559 5334 : "prop_elected;{};{};{};{}",
560 5334 : prop_lsn, prop_term, prev_lsn, prev_term
561 5334 : );
562 5334 :
563 5334 : debug!(msg);
564 5334 : self.os.log_event(msg);
565 5334 : }
566 :
567 2110 : fn get_redo_start_lsn(&self) -> u64 {
568 2110 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
569 2110 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
570 2110 : }
571 :
572 1294 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
573 1294 : self.shmem.get()
574 1294 : }
575 :
576 1259 : fn start_streaming(
577 1259 : &self,
578 1259 : startpos: u64,
579 1259 : callback: &walproposer::walproposer::StreamingCallback,
580 1259 : ) {
581 1259 : let disk = &self.disk;
582 1259 : let disk_lsn = disk.lock().flush_rec_ptr().0;
583 1259 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
584 1259 : if startpos < disk_lsn {
585 329 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
586 930 : }
587 1259 : assert!(startpos <= disk_lsn);
588 1259 : let mut broadcasted = Lsn(startpos);
589 :
590 : loop {
591 4106 : let available = disk.lock().flush_rec_ptr();
592 4106 : assert!(available >= broadcasted);
593 2847 : callback.broadcast(broadcasted, available);
594 2847 : broadcasted = available;
595 2847 : callback.poll();
596 : }
597 : }
598 :
599 13377 : fn process_safekeeper_feedback(
600 13377 : &self,
601 13377 : wp: &mut walproposer::bindings::WalProposer,
602 13377 : commit_lsn: u64,
603 13377 : ) {
604 13377 : debug!("process_safekeeper_feedback, commit_lsn={}", commit_lsn);
605 13377 : if commit_lsn > wp.lastSentCommitLsn {
606 2883 : self.os.log_event(format!("commit_lsn;{}", commit_lsn));
607 10494 : }
608 13377 : }
609 :
610 793 : fn get_flush_rec_ptr(&self) -> u64 {
611 793 : let lsn = self.disk.lock().flush_rec_ptr();
612 793 : debug!("get_flush_rec_ptr: {}", lsn);
613 793 : lsn.0
614 793 : }
615 :
616 5334 : fn recovery_download(
617 5334 : &self,
618 5334 : wp: &mut walproposer::bindings::WalProposer,
619 5334 : sk: &mut walproposer::bindings::Safekeeper,
620 5334 : ) -> bool {
621 5334 : let mut startpos = wp.truncateLsn;
622 5334 : let endpos = wp.propEpochStartLsn;
623 5334 :
624 5334 : if startpos == endpos {
625 3301 : debug!("recovery_download: nothing to download");
626 3301 : return true;
627 2033 : }
628 2033 :
629 2033 : debug!("recovery_download from {} to {}", startpos, endpos,);
630 :
631 2033 : let replication_prompt = format!(
632 2033 : "START_REPLICATION {} {} {} {}",
633 2033 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
634 2033 : );
635 2033 : let async_conn = self.get_conn(sk);
636 2033 :
637 2033 : let conn = self.os.open_tcp(async_conn.node_id);
638 2033 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
639 2033 :
640 2033 : let chan = conn.recv_chan();
641 3506 : while startpos < endpos {
642 2033 : let event = chan.recv();
643 1934 : match event {
644 : NetEvent::Closed => {
645 99 : debug!("connection closed in recovery");
646 99 : break;
647 : }
648 1934 : NetEvent::Message(AnyMessage::Bytes(b)) => {
649 1934 : debug!("got recovery bytes from safekeeper");
650 1473 : self.disk.lock().write(startpos, &b);
651 1473 : startpos += b.len() as u64;
652 : }
653 0 : NetEvent::Message(_) => unreachable!(),
654 : }
655 : }
656 :
657 1572 : debug!("recovery finished at {}", startpos);
658 :
659 1572 : startpos == endpos
660 4873 : }
661 :
662 87615 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
663 87615 : let mut conn = self.get_conn(sk);
664 87615 : debug!("conn_finish to {}", conn.node_id);
665 87615 : if let Some(socket) = conn.socket.as_mut() {
666 28392 : socket.close();
667 59223 : } else {
668 59223 : // connection is already closed
669 59223 : }
670 87615 : conn.socket = None;
671 87615 : }
672 :
673 85567 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
674 85567 : "connection is closed, probably".into()
675 85567 : }
676 : }
|
