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 219990 : pub fn new(host: String, port: String) -> Self {
44 219990 : // port number is the same as NodeId
45 219990 : let port_num = port.parse::<u32>().unwrap();
46 219990 : Self {
47 219990 : host,
48 219990 : port,
49 219990 : node_id: port_num,
50 219990 : socket: None,
51 219990 : is_connecting: false,
52 219990 : is_start_wal_push: false,
53 219990 : raw_ptr: std::ptr::null_mut(),
54 219990 : }
55 219990 : }
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 73330 : pub fn new(os: NodeOs) -> Self {
72 73330 : let node_events = os.node_events();
73 73330 : Self {
74 73330 : os,
75 73330 : chans: vec![Box::new(node_events)],
76 73330 : sk_ptrs: vec![std::ptr::null_mut()],
77 73330 : masks: vec![WL_SOCKET_READABLE],
78 73330 : }
79 73330 : }
80 :
81 : /// Leaves all readable channels at the beginning of the array.
82 224647 : fn sort_readable(&mut self) -> usize {
83 224647 : let mut cnt = 1;
84 533892 : for i in 1..self.chans.len() {
85 533892 : if self.masks[i] & WL_SOCKET_READABLE != 0 {
86 533892 : self.chans.swap(i, cnt);
87 533892 : self.sk_ptrs.swap(i, cnt);
88 533892 : self.masks.swap(i, cnt);
89 533892 : cnt += 1;
90 533892 : }
91 : }
92 224647 : cnt
93 224647 : }
94 :
95 539041 : fn update_event_set(&mut self, conn: &SafekeeperConn, event_mask: u32) {
96 539041 : let index = self
97 539041 : .sk_ptrs
98 539041 : .iter()
99 2033623 : .position(|&ptr| ptr == conn.raw_ptr)
100 539041 : .expect("safekeeper should exist in event set");
101 539041 : self.masks[index] = event_mask;
102 539041 : }
103 :
104 486278 : fn add_safekeeper(&mut self, sk: &SafekeeperConn, event_mask: u32) {
105 1176718 : for ptr in self.sk_ptrs.iter() {
106 1176718 : assert!(*ptr != sk.raw_ptr);
107 : }
108 :
109 486278 : self.chans.push(Box::new(
110 486278 : sk.socket
111 486278 : .as_ref()
112 486278 : .expect("socket should not be closed")
113 486278 : .recv_chan(),
114 486278 : ));
115 486278 : self.sk_ptrs.push(sk.raw_ptr);
116 486278 : self.masks.push(event_mask);
117 486278 : }
118 :
119 306592 : fn remove_safekeeper(&mut self, sk: &SafekeeperConn) {
120 652186 : let index = self.sk_ptrs.iter().position(|&ptr| ptr == sk.raw_ptr);
121 306592 : if index.is_none() {
122 46 : debug!("remove_safekeeper: sk={:?} not found", sk.raw_ptr);
123 46 : return;
124 306546 : }
125 306546 : let index = index.unwrap();
126 306546 :
127 306546 : self.chans.remove(index);
128 306546 : self.sk_ptrs.remove(index);
129 306546 : self.masks.remove(index);
130 306546 :
131 306546 : // to simulate the actual behaviour
132 306546 : self.refresh_event_set();
133 306592 : }
134 :
135 : /// Updates all masks to match the result of a SafekeeperStateDesiredEvents.
136 348996 : fn refresh_event_set(&mut self) {
137 987449 : for (i, mask) in self.masks.iter_mut().enumerate() {
138 987449 : if i == 0 {
139 348996 : continue;
140 638453 : }
141 638453 :
142 638453 : let mut mask_sk: u32 = 0;
143 638453 : let mut mask_nwr: u32 = 0;
144 638453 : unsafe { SafekeeperStateDesiredEvents(self.sk_ptrs[i], &mut mask_sk, &mut mask_nwr) };
145 638453 :
146 638453 : 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 638453 : }
153 : }
154 348996 : }
155 :
156 : /// Wait for events on all channels.
157 224647 : fn wait(&mut self, timeout_millis: i64) -> walproposer::walproposer::WaitResult {
158 : // all channels are always writeable
159 758539 : for (i, mask) in self.masks.iter().enumerate() {
160 758539 : 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 758539 : }
166 : }
167 :
168 224647 : let cnt = self.sort_readable();
169 224647 :
170 224647 : let slice = &self.chans[0..cnt];
171 224647 : match executor::epoll_chans(slice, timeout_millis) {
172 91234 : None => walproposer::walproposer::WaitResult::Timeout,
173 : Some(0) => {
174 3128 : let msg = self.os.node_events().must_recv();
175 3128 : match msg {
176 3128 : NodeEvent::Internal(AnyMessage::Just32(0)) => {
177 3128 : // got a notification about new WAL available
178 3128 : }
179 0 : NodeEvent::Internal(_) => unreachable!(),
180 0 : NodeEvent::Accept(_) => unreachable!(),
181 : }
182 3128 : walproposer::walproposer::WaitResult::Latch
183 : }
184 130285 : Some(index) => walproposer::walproposer::WaitResult::Network(
185 130285 : self.sk_ptrs[index],
186 130285 : WL_SOCKET_READABLE,
187 130285 : ),
188 : }
189 224647 : }
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 73330 : pub fn new(args: Args) -> Self {
213 73330 : // initialize connection state for each safekeeper
214 73330 : let sk_conns = args
215 73330 : .config
216 73330 : .safekeepers_list
217 73330 : .iter()
218 219990 : .map(|s| {
219 219990 : SafekeeperConn::new(
220 219990 : s.split(':').next().unwrap().to_string(),
221 219990 : s.split(':').nth(1).unwrap().to_string(),
222 219990 : )
223 219990 : })
224 73330 : .collect::<Vec<_>>();
225 73330 :
226 73330 : Self {
227 73330 : os: args.os,
228 73330 : safekeepers: RefCell::new(sk_conns),
229 73330 : disk: args.disk,
230 73330 : redo_start_lsn: args.redo_start_lsn,
231 73330 : last_logged_commit_lsn: 0,
232 73330 : shmem: UnsafeCell::new(walproposer::api_bindings::empty_shmem()),
233 73330 : config: args.config,
234 73330 : event_set: RefCell::new(None),
235 73330 : }
236 73330 : }
237 :
238 : /// Get SafekeeperConn for the given Safekeeper.
239 2390911 : fn get_conn(&self, sk: &mut walproposer::bindings::Safekeeper) -> RefMut<'_, SafekeeperConn> {
240 2390911 : let sk_port = unsafe { CStr::from_ptr(sk.port).to_str().unwrap() };
241 2390911 : let state = self.safekeepers.borrow_mut();
242 2390911 : RefMut::map(state, |v| {
243 2390911 : v.iter_mut()
244 4777988 : .find(|conn| conn.port == sk_port)
245 2390911 : .expect("safekeeper conn not found by port")
246 2390911 : })
247 2390911 : }
248 : }
249 :
250 : impl ApiImpl for SimulationApi {
251 2552312 : fn get_current_timestamp(&self) -> i64 {
252 2552312 : debug!("get_current_timestamp");
253 : // PG TimestampTZ is microseconds, but simulation unit is assumed to be
254 : // milliseconds, so add 10^3
255 2552312 : self.os.now() as i64 * 1000
256 2552312 : }
257 :
258 7982 : fn update_donor(&self, donor: &mut walproposer::bindings::Safekeeper, donor_lsn: u64) {
259 7982 : let mut shmem = unsafe { *self.get_shmem_state() };
260 7982 : shmem.propEpochStartLsn.value = donor_lsn;
261 7982 : shmem.donor_conninfo = donor.conninfo;
262 7982 : }
263 :
264 269971 : fn conn_status(
265 269971 : &self,
266 269971 : _: &mut walproposer::bindings::Safekeeper,
267 269971 : ) -> walproposer::bindings::WalProposerConnStatusType {
268 269971 : debug!("conn_status");
269 : // break the connection with a 10% chance
270 269971 : if self.os.random(100) < 10 {
271 26832 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_BAD
272 : } else {
273 243139 : walproposer::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
274 : }
275 269971 : }
276 :
277 269971 : fn conn_connect_start(&self, sk: &mut walproposer::bindings::Safekeeper) {
278 269971 : debug!("conn_connect_start");
279 269971 : let mut conn = self.get_conn(sk);
280 269971 :
281 269971 : assert!(conn.socket.is_none());
282 269971 : let socket = self.os.open_tcp(conn.node_id);
283 269971 : conn.socket = Some(socket);
284 269971 : conn.raw_ptr = sk;
285 269971 : conn.is_connecting = true;
286 269971 : }
287 :
288 243139 : fn conn_connect_poll(
289 243139 : &self,
290 243139 : _: &mut walproposer::bindings::Safekeeper,
291 243139 : ) -> walproposer::bindings::WalProposerConnectPollStatusType {
292 243139 : debug!("conn_connect_poll");
293 : // TODO: break the connection here
294 243139 : walproposer::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
295 243139 : }
296 :
297 243139 : fn conn_send_query(&self, sk: &mut walproposer::bindings::Safekeeper, query: &str) -> bool {
298 243139 : debug!("conn_send_query: {}", query);
299 243139 : self.get_conn(sk).is_start_wal_push = true;
300 243139 : true
301 243139 : }
302 :
303 243139 : fn conn_get_query_result(
304 243139 : &self,
305 243139 : _: &mut walproposer::bindings::Safekeeper,
306 243139 : ) -> walproposer::bindings::WalProposerExecStatusType {
307 243139 : debug!("conn_get_query_result");
308 : // TODO: break the connection here
309 243139 : walproposer::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
310 243139 : }
311 :
312 142399 : fn conn_async_read(
313 142399 : &self,
314 142399 : sk: &mut walproposer::bindings::Safekeeper,
315 142399 : vec: &mut Vec<u8>,
316 142399 : ) -> walproposer::bindings::PGAsyncReadResult {
317 142399 : debug!("conn_async_read");
318 142399 : let mut conn = self.get_conn(sk);
319 :
320 142399 : let socket = if let Some(socket) = conn.socket.as_mut() {
321 142399 : socket
322 : } else {
323 : // socket is already closed
324 0 : return walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL;
325 : };
326 :
327 142399 : let msg = socket.recv_chan().try_recv();
328 :
329 128625 : match msg {
330 : None => {
331 : // no message is ready
332 13774 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_TRY_AGAIN
333 : }
334 : Some(NetEvent::Closed) => {
335 : // connection is closed
336 61246 : debug!("conn_async_read: connection is closed");
337 61246 : conn.socket = None;
338 61246 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_FAIL
339 : }
340 67379 : Some(NetEvent::Message(msg)) => {
341 : // got a message
342 67379 : let b = match msg {
343 67379 : desim::proto::AnyMessage::Bytes(b) => b,
344 0 : _ => unreachable!(),
345 : };
346 67379 : vec.extend_from_slice(&b);
347 67379 : walproposer::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
348 : }
349 : }
350 142399 : }
351 :
352 273734 : fn conn_blocking_write(&self, sk: &mut walproposer::bindings::Safekeeper, buf: &[u8]) -> bool {
353 273734 : let mut conn = self.get_conn(sk);
354 273734 : debug!("conn_blocking_write to {}: {:?}", conn.node_id, buf);
355 273734 : let socket = conn.socket.as_mut().unwrap();
356 273734 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
357 273734 : true
358 273734 : }
359 :
360 37015 : fn conn_async_write(
361 37015 : &self,
362 37015 : sk: &mut walproposer::bindings::Safekeeper,
363 37015 : buf: &[u8],
364 37015 : ) -> walproposer::bindings::PGAsyncWriteResult {
365 37015 : let mut conn = self.get_conn(sk);
366 37015 : debug!("conn_async_write to {}: {:?}", conn.node_id, buf);
367 37015 : if let Some(socket) = conn.socket.as_mut() {
368 37015 : socket.send(desim::proto::AnyMessage::Bytes(Bytes::copy_from_slice(buf)));
369 37015 : } 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 37015 : walproposer::bindings::PGAsyncWriteResult_PG_ASYNC_WRITE_SUCCESS
375 37015 : }
376 :
377 8339 : fn wal_reader_allocate(&self, _: &mut walproposer::bindings::Safekeeper) -> NeonWALReadResult {
378 8339 : debug!("wal_reader_allocate");
379 8339 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
380 8339 : }
381 :
382 28676 : fn wal_read(
383 28676 : &self,
384 28676 : _sk: &mut walproposer::bindings::Safekeeper,
385 28676 : buf: &mut [u8],
386 28676 : startpos: u64,
387 28676 : ) -> NeonWALReadResult {
388 28676 : self.disk.lock().read(startpos, buf);
389 28676 : walproposer::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
390 28676 : }
391 :
392 73330 : fn init_event_set(&self, _: &mut walproposer::bindings::WalProposer) {
393 73330 : debug!("init_event_set");
394 73330 : let new_event_set = EventSet::new(self.os.clone());
395 73330 : let old_event_set = self.event_set.replace(Some(new_event_set));
396 73330 : assert!(old_event_set.is_none());
397 73330 : }
398 :
399 539041 : fn update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper, event_mask: u32) {
400 539041 : debug!(
401 0 : "update_event_set, sk={:?}, events_mask={:#b}",
402 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
403 : );
404 539041 : let conn = self.get_conn(sk);
405 539041 :
406 539041 : self.event_set
407 539041 : .borrow_mut()
408 539041 : .as_mut()
409 539041 : .unwrap()
410 539041 : .update_event_set(&conn, event_mask);
411 539041 : }
412 :
413 486278 : fn add_safekeeper_event_set(
414 486278 : &self,
415 486278 : sk: &mut walproposer::bindings::Safekeeper,
416 486278 : event_mask: u32,
417 486278 : ) {
418 486278 : debug!(
419 0 : "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
420 0 : sk as *mut walproposer::bindings::Safekeeper, event_mask
421 : );
422 :
423 486278 : self.event_set
424 486278 : .borrow_mut()
425 486278 : .as_mut()
426 486278 : .unwrap()
427 486278 : .add_safekeeper(&self.get_conn(sk), event_mask);
428 486278 : }
429 :
430 306592 : fn rm_safekeeper_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
431 306592 : debug!(
432 0 : "rm_safekeeper_event_set, sk={:?}",
433 0 : sk as *mut walproposer::bindings::Safekeeper,
434 : );
435 :
436 306592 : self.event_set
437 306592 : .borrow_mut()
438 306592 : .as_mut()
439 306592 : .unwrap()
440 306592 : .remove_safekeeper(&self.get_conn(sk));
441 306592 : }
442 :
443 42450 : fn active_state_update_event_set(&self, sk: &mut walproposer::bindings::Safekeeper) {
444 42450 : debug!("active_state_update_event_set");
445 :
446 42450 : assert!(sk.state == walproposer::bindings::SafekeeperState_SS_ACTIVE);
447 42450 : self.event_set
448 42450 : .borrow_mut()
449 42450 : .as_mut()
450 42450 : .unwrap()
451 42450 : .refresh_event_set();
452 42450 : }
453 :
454 120634 : fn wal_reader_events(&self, _sk: &mut walproposer::bindings::Safekeeper) -> u32 {
455 120634 : 0
456 120634 : }
457 :
458 710925 : fn wait_event_set(
459 710925 : &self,
460 710925 : _: &mut walproposer::bindings::WalProposer,
461 710925 : timeout_millis: i64,
462 710925 : ) -> walproposer::walproposer::WaitResult {
463 710925 : // TODO: handle multiple stages as part of the simulation (e.g. connect, start_wal_push, etc)
464 710925 : let mut conns = self.safekeepers.borrow_mut();
465 1646469 : for conn in conns.iter_mut() {
466 1646469 : if conn.socket.is_some() && conn.is_connecting {
467 243139 : conn.is_connecting = false;
468 243139 : debug!("wait_event_set, connecting to {}:{}", conn.host, conn.port);
469 243139 : return walproposer::walproposer::WaitResult::Network(
470 243139 : conn.raw_ptr,
471 243139 : WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE,
472 243139 : );
473 1403330 : }
474 1403330 : if conn.socket.is_some() && conn.is_start_wal_push {
475 243139 : conn.is_start_wal_push = false;
476 243139 : debug!(
477 0 : "wait_event_set, start wal push to {}:{}",
478 : conn.host, conn.port
479 : );
480 243139 : return walproposer::walproposer::WaitResult::Network(
481 243139 : conn.raw_ptr,
482 243139 : WL_SOCKET_READABLE,
483 243139 : );
484 1160191 : }
485 : }
486 224647 : drop(conns);
487 224647 :
488 224647 : let res = self
489 224647 : .event_set
490 224647 : .borrow_mut()
491 224647 : .as_mut()
492 224647 : .unwrap()
493 224647 : .wait(timeout_millis);
494 224647 :
495 224647 : debug!(
496 0 : "wait_event_set, timeout_millis={}, res={:?}",
497 : timeout_millis, res,
498 : );
499 155459 : res
500 641737 : }
501 :
502 73330 : fn strong_random(&self, buf: &mut [u8]) -> bool {
503 73330 : debug!("strong_random");
504 73330 : buf.fill(0);
505 73330 : true
506 73330 : }
507 :
508 2971 : fn finish_sync_safekeepers(&self, lsn: u64) {
509 2971 : debug!("finish_sync_safekeepers, lsn={}", lsn);
510 2971 : executor::exit(0, Lsn(lsn).to_string());
511 2971 : }
512 :
513 732192 : fn log_internal(&self, _wp: &mut walproposer::bindings::WalProposer, level: Level, msg: &str) {
514 732192 : debug!("wp_log[{}] {}", level, msg);
515 732192 : if level == Level::Fatal || level == Level::Panic {
516 605 : if msg.contains("rejects our connection request with term") {
517 313 : // collected quorum with lower term, then got rejected by next connected safekeeper
518 313 : executor::exit(1, msg.to_owned());
519 313 : }
520 605 : if msg.contains("collected propEpochStartLsn") && msg.contains(", but basebackup LSN ")
521 23 : {
522 23 : // sync-safekeepers collected wrong quorum, walproposer collected another quorum
523 23 : executor::exit(1, msg.to_owned());
524 582 : }
525 605 : if msg.contains("failed to download WAL for logical replicaiton") {
526 171 : // Recovery connection broken and recovery was failed
527 171 : executor::exit(1, msg.to_owned());
528 434 : }
529 605 : if msg.contains("missing majority of votes, collected") {
530 98 : // Voting bug when safekeeper disconnects after voting
531 98 : executor::exit(1, msg.to_owned());
532 507 : }
533 605 : panic!("unknown FATAL error from walproposer: {}", msg);
534 731587 : }
535 731587 : }
536 :
537 6135 : fn after_election(&self, wp: &mut walproposer::bindings::WalProposer) {
538 6135 : let prop_lsn = wp.propEpochStartLsn;
539 6135 : let prop_term = wp.propTerm;
540 6135 :
541 6135 : let mut prev_lsn: u64 = 0;
542 6135 : let mut prev_term: u64 = 0;
543 6135 :
544 6135 : unsafe {
545 6135 : let history = wp.propTermHistory.entries;
546 6135 : let len = wp.propTermHistory.n_entries as usize;
547 6135 : if len > 1 {
548 4134 : let entry = *history.wrapping_add(len - 2);
549 4134 : prev_lsn = entry.lsn;
550 4134 : prev_term = entry.term;
551 4134 : }
552 : }
553 :
554 6135 : let msg = format!(
555 6135 : "prop_elected;{};{};{};{}",
556 6135 : prop_lsn, prop_term, prev_lsn, prev_term
557 6135 : );
558 6135 :
559 6135 : debug!(msg);
560 6135 : self.os.log_event(msg);
561 6135 : }
562 :
563 2318 : fn get_redo_start_lsn(&self) -> u64 {
564 2318 : debug!("get_redo_start_lsn -> {:?}", self.redo_start_lsn);
565 2318 : self.redo_start_lsn.expect("redo_start_lsn is not set").0
566 2318 : }
567 :
568 20298 : fn get_shmem_state(&self) -> *mut walproposer::bindings::WalproposerShmemState {
569 20298 : self.shmem.get()
570 20298 : }
571 :
572 1347 : fn start_streaming(
573 1347 : &self,
574 1347 : startpos: u64,
575 1347 : callback: &walproposer::walproposer::StreamingCallback,
576 1347 : ) {
577 1347 : let disk = &self.disk;
578 1347 : let disk_lsn = disk.lock().flush_rec_ptr().0;
579 1347 : debug!("start_streaming at {} (disk_lsn={})", startpos, disk_lsn);
580 1347 : if startpos < disk_lsn {
581 389 : debug!("startpos < disk_lsn, it means we wrote some transaction even before streaming started");
582 958 : }
583 1347 : assert!(startpos <= disk_lsn);
584 1347 : let mut broadcasted = Lsn(startpos);
585 :
586 : loop {
587 4463 : let available = disk.lock().flush_rec_ptr();
588 4463 : assert!(available >= broadcasted);
589 3116 : callback.broadcast(broadcasted, available);
590 3116 : broadcasted = available;
591 3116 : callback.poll();
592 : }
593 : }
594 :
595 14844 : fn process_safekeeper_feedback(
596 14844 : &mut self,
597 14844 : wp: &mut walproposer::bindings::WalProposer,
598 14844 : _sk: &mut walproposer::bindings::Safekeeper,
599 14844 : ) {
600 14844 : debug!("process_safekeeper_feedback, commit_lsn={}", wp.commitLsn);
601 14844 : if wp.commitLsn > self.last_logged_commit_lsn {
602 3138 : self.os.log_event(format!("commit_lsn;{}", wp.commitLsn));
603 3138 : self.last_logged_commit_lsn = wp.commitLsn;
604 11706 : }
605 14844 : }
606 :
607 845 : fn get_flush_rec_ptr(&self) -> u64 {
608 845 : let lsn = self.disk.lock().flush_rec_ptr();
609 845 : debug!("get_flush_rec_ptr: {}", lsn);
610 845 : lsn.0
611 845 : }
612 :
613 6135 : fn recovery_download(
614 6135 : &self,
615 6135 : wp: &mut walproposer::bindings::WalProposer,
616 6135 : sk: &mut walproposer::bindings::Safekeeper,
617 6135 : ) -> bool {
618 6135 : let mut startpos = wp.truncateLsn;
619 6135 : let endpos = wp.propEpochStartLsn;
620 6135 :
621 6135 : if startpos == endpos {
622 3678 : debug!("recovery_download: nothing to download");
623 3678 : return true;
624 2457 : }
625 2457 :
626 2457 : debug!("recovery_download from {} to {}", startpos, endpos,);
627 :
628 2457 : let replication_prompt = format!(
629 2457 : "START_REPLICATION {} {} {} {}",
630 2457 : self.config.ttid.tenant_id, self.config.ttid.timeline_id, startpos, endpos,
631 2457 : );
632 2457 : let async_conn = self.get_conn(sk);
633 2457 :
634 2457 : let conn = self.os.open_tcp(async_conn.node_id);
635 2457 : conn.send(desim::proto::AnyMessage::Bytes(replication_prompt.into()));
636 2457 :
637 2457 : let chan = conn.recv_chan();
638 4177 : while startpos < endpos {
639 2457 : let event = chan.recv();
640 2286 : match event {
641 : NetEvent::Closed => {
642 171 : debug!("connection closed in recovery");
643 171 : break;
644 : }
645 2286 : NetEvent::Message(AnyMessage::Bytes(b)) => {
646 2286 : debug!("got recovery bytes from safekeeper");
647 1720 : self.disk.lock().write(startpos, &b);
648 1720 : startpos += b.len() as u64;
649 : }
650 0 : NetEvent::Message(_) => unreachable!(),
651 : }
652 : }
653 :
654 1891 : debug!("recovery finished at {}", startpos);
655 :
656 1891 : startpos == endpos
657 5569 : }
658 :
659 90285 : fn conn_finish(&self, sk: &mut walproposer::bindings::Safekeeper) {
660 90285 : let mut conn = self.get_conn(sk);
661 90285 : debug!("conn_finish to {}", conn.node_id);
662 90285 : if let Some(socket) = conn.socket.as_mut() {
663 28993 : socket.close();
664 61292 : } else {
665 61292 : // connection is already closed
666 61292 : }
667 90285 : conn.socket = None;
668 90285 : }
669 :
670 88078 : fn conn_error_message(&self, _sk: &mut walproposer::bindings::Safekeeper) -> String {
671 88078 : "connection is closed, probably".into()
672 88078 : }
673 : }
|
