Line data Source code
1 : //! Safekeeper communication endpoint to WAL proposer (compute node).
2 : //! Gets messages from the network, passes them down to consensus module and
3 : //! sends replies back.
4 :
5 : use std::{collections::HashMap, sync::Arc, time::Duration};
6 :
7 : use anyhow::{bail, Result};
8 : use bytes::{Bytes, BytesMut};
9 : use camino::Utf8PathBuf;
10 : use desim::{
11 : executor::{self, PollSome},
12 : network::TCP,
13 : node_os::NodeOs,
14 : proto::{AnyMessage, NetEvent, NodeEvent},
15 : };
16 : use hyper::Uri;
17 : use safekeeper::{
18 : safekeeper::{ProposerAcceptorMessage, SafeKeeper, ServerInfo, UNKNOWN_SERVER_VERSION},
19 : state::TimelinePersistentState,
20 : timeline::TimelineError,
21 : wal_storage::Storage,
22 : SafeKeeperConf,
23 : };
24 : use tracing::{debug, info_span, warn};
25 : use utils::{
26 : id::{NodeId, TenantId, TenantTimelineId, TimelineId},
27 : lsn::Lsn,
28 : };
29 :
30 : use super::safekeeper_disk::{DiskStateStorage, DiskWALStorage, SafekeeperDisk, TimelineDisk};
31 :
32 : struct SharedState {
33 : sk: SafeKeeper<DiskStateStorage, DiskWALStorage>,
34 : disk: Arc<TimelineDisk>,
35 : }
36 :
37 : struct GlobalMap {
38 : timelines: HashMap<TenantTimelineId, SharedState>,
39 : conf: SafeKeeperConf,
40 : disk: Arc<SafekeeperDisk>,
41 : }
42 :
43 : impl GlobalMap {
44 : /// Restores global state from disk.
45 76469 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 76469 : let mut timelines = HashMap::new();
47 :
48 76469 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 57348 : debug!("loading timeline {}", ttid);
50 57348 : let state = disk.state.lock().clone();
51 57348 :
52 57348 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 57348 : }
55 57348 :
56 57348 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 57348 : }
59 57348 :
60 57348 : if state.commit_lsn < state.local_start_lsn {
61 0 : bail!(
62 0 : "commit_lsn {} is higher than local_start_lsn {}",
63 0 : state.commit_lsn,
64 0 : state.local_start_lsn
65 0 : );
66 57348 : }
67 57348 :
68 57348 : let control_store = DiskStateStorage::new(disk.clone());
69 57348 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 57348 : let sk = SafeKeeper::new(control_store, wal_store, conf.my_id)?;
72 57348 : timelines.insert(
73 57348 : ttid,
74 57348 : SharedState {
75 57348 : sk,
76 57348 : disk: disk.clone(),
77 57348 : },
78 57348 : );
79 : }
80 :
81 76469 : Ok(Self {
82 76469 : timelines,
83 76469 : conf,
84 76469 : disk,
85 76469 : })
86 76469 : }
87 :
88 11462 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 11462 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 11462 : }
92 11462 :
93 11462 : debug!("creating new timeline {}", ttid);
94 :
95 11462 : let commit_lsn = Lsn::INVALID;
96 11462 : let local_start_lsn = Lsn::INVALID;
97 11462 :
98 11462 : let state =
99 11462 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
100 11462 :
101 11462 : if state.server.wal_seg_size == 0 {
102 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
103 11462 : }
104 11462 :
105 11462 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
106 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
107 11462 : }
108 11462 :
109 11462 : if state.commit_lsn < state.local_start_lsn {
110 0 : bail!(
111 0 : "commit_lsn {} is higher than local_start_lsn {}",
112 0 : state.commit_lsn,
113 0 : state.local_start_lsn
114 0 : );
115 11462 : }
116 11462 :
117 11462 : let disk_timeline = self.disk.put_state(&ttid, state);
118 11462 : let control_store = DiskStateStorage::new(disk_timeline.clone());
119 11462 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
120 :
121 11462 : let sk = SafeKeeper::new(control_store, wal_store, self.conf.my_id)?;
122 :
123 11462 : self.timelines.insert(
124 11462 : ttid,
125 11462 : SharedState {
126 11462 : sk,
127 11462 : disk: disk_timeline,
128 11462 : },
129 11462 : );
130 11462 : Ok(())
131 11462 : }
132 :
133 221443 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
134 221443 : self.timelines.get_mut(ttid).expect("timeline must exist")
135 221443 : }
136 :
137 154044 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
138 154044 : self.timelines.contains_key(ttid)
139 154044 : }
140 : }
141 :
142 : /// State of a single connection to walproposer.
143 : struct ConnState {
144 : tcp: TCP,
145 :
146 : greeting: bool,
147 : ttid: TenantTimelineId,
148 : flush_pending: bool,
149 :
150 : runtime: tokio::runtime::Runtime,
151 : }
152 :
153 76469 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
154 76469 : let _enter = info_span!("safekeeper", id = os.id()).entered();
155 76469 : debug!("started server");
156 76469 : os.log_event("started;safekeeper".to_owned());
157 76469 : let conf = SafeKeeperConf {
158 76469 : workdir: Utf8PathBuf::from("."),
159 76469 : my_id: NodeId(os.id() as u64),
160 76469 : listen_pg_addr: String::new(),
161 76469 : listen_http_addr: String::new(),
162 76469 : no_sync: false,
163 76469 : broker_endpoint: "/".parse::<Uri>().unwrap(),
164 76469 : broker_keepalive_interval: Duration::from_secs(0),
165 76469 : heartbeat_timeout: Duration::from_secs(0),
166 76469 : remote_storage: None,
167 76469 : max_offloader_lag_bytes: 0,
168 76469 : wal_backup_enabled: false,
169 76469 : listen_pg_addr_tenant_only: None,
170 76469 : advertise_pg_addr: None,
171 76469 : availability_zone: None,
172 76469 : peer_recovery_enabled: false,
173 76469 : backup_parallel_jobs: 0,
174 76469 : pg_auth: None,
175 76469 : pg_tenant_only_auth: None,
176 76469 : http_auth: None,
177 76469 : sk_auth_token: None,
178 76469 : current_thread_runtime: false,
179 76469 : walsenders_keep_horizon: false,
180 76469 : partial_backup_enabled: false,
181 76469 : partial_backup_timeout: Duration::from_secs(0),
182 76469 : disable_periodic_broker_push: false,
183 76469 : };
184 :
185 76469 : let mut global = GlobalMap::new(disk, conf.clone())?;
186 76469 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
187 :
188 76469 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
189 57348 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
190 57348 : let commit_lsn = shared_state.sk.state.commit_lsn;
191 57348 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
192 57348 : }
193 :
194 76469 : let node_events = os.node_events();
195 76469 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
196 76469 : let mut epoll_idx: Vec<usize> = vec![];
197 :
198 : // TODO: batch events processing (multiple events per tick)
199 : loop {
200 584375 : epoll_vec.clear();
201 584375 : epoll_idx.clear();
202 584375 :
203 584375 : // node events channel
204 584375 : epoll_vec.push(Box::new(node_events.clone()));
205 584375 : epoll_idx.push(0);
206 :
207 : // tcp connections
208 2102675 : for conn in conns.values() {
209 2102675 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
210 2102675 : epoll_idx.push(conn.tcp.connection_id());
211 2102675 : }
212 :
213 : // waiting for the next message
214 584375 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
215 584375 :
216 584375 : if index == 0 {
217 : // got a new connection
218 197363 : match node_events.must_recv() {
219 197363 : NodeEvent::Accept(tcp) => {
220 197363 : conns.insert(
221 197363 : tcp.connection_id(),
222 197363 : ConnState {
223 197363 : tcp,
224 197363 : greeting: false,
225 197363 : ttid: TenantTimelineId::empty(),
226 197363 : flush_pending: false,
227 197363 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
228 : },
229 : );
230 : }
231 0 : NodeEvent::Internal(_) => unreachable!(),
232 : }
233 197363 : continue;
234 387012 : }
235 387012 :
236 387012 : let connection_id = epoll_idx[index];
237 387012 : let conn = conns.get_mut(&connection_id).unwrap();
238 387012 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
239 :
240 : loop {
241 702462 : let event = match next_event {
242 393806 : Some(event) => event,
243 308656 : None => break,
244 : };
245 :
246 393806 : match event {
247 281603 : NetEvent::Message(msg) => {
248 281603 : let res = conn.process_any(msg, &mut global);
249 281603 : if res.is_err() {
250 78356 : let e = res.unwrap_err();
251 78356 : let estr = e.to_string();
252 78356 : if !estr.contains("finished processing START_REPLICATION") {
253 76469 : warn!("conn {:?} error: {:?}", connection_id, e);
254 0 : panic!("unexpected error at safekeeper: {:#}", e);
255 1887 : }
256 1887 : conns.remove(&connection_id);
257 1887 : break;
258 203247 : }
259 : }
260 112203 : NetEvent::Closed => {
261 112203 : // TODO: remove from conns?
262 112203 : }
263 : }
264 :
265 315450 : next_event = conn.tcp.recv_chan().try_recv();
266 : }
267 :
268 1365862 : conns.retain(|_, conn| {
269 1365862 : let res = conn.flush(&mut global);
270 1365862 : if res.is_err() {
271 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
272 1365862 : }
273 1365862 : res.is_ok()
274 1365862 : });
275 : }
276 0 : }
277 :
278 : impl ConnState {
279 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
280 205134 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
281 205134 : if let AnyMessage::Bytes(copy_data) = any {
282 205134 : let repl_prefix = b"START_REPLICATION ";
283 205134 : if !self.greeting && copy_data.starts_with(repl_prefix) {
284 1887 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
285 1887 : bail!("finished processing START_REPLICATION")
286 203247 : }
287 :
288 203247 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
289 203247 : debug!("got msg: {:?}", msg);
290 203247 : self.process(msg, global)
291 : } else {
292 0 : bail!("unexpected message, expected AnyMessage::Bytes");
293 : }
294 205134 : }
295 :
296 : /// Process START_REPLICATION request.
297 1887 : fn process_start_replication(
298 1887 : &mut self,
299 1887 : copy_data: Bytes,
300 1887 : global: &mut GlobalMap,
301 1887 : ) -> Result<()> {
302 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
303 1887 : let str = String::from_utf8(copy_data.to_vec())?;
304 :
305 1887 : let mut parts = str.split(' ');
306 1887 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
307 1887 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
308 1887 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
309 1887 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
310 :
311 1887 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
312 1887 : let shared_state = global.get(&ttid);
313 1887 :
314 1887 : // read bytes from start_lsn to end_lsn
315 1887 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
316 1887 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
317 1887 :
318 1887 : // send bytes to the client
319 1887 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
320 1887 : Ok(())
321 1887 : }
322 :
323 : /// Get or create a timeline.
324 154044 : fn init_timeline(
325 154044 : &mut self,
326 154044 : ttid: TenantTimelineId,
327 154044 : server_info: ServerInfo,
328 154044 : global: &mut GlobalMap,
329 154044 : ) -> Result<()> {
330 154044 : self.ttid = ttid;
331 154044 : if global.has_tli(&ttid) {
332 142582 : return Ok(());
333 11462 : }
334 11462 :
335 11462 : global.create(ttid, server_info)
336 154044 : }
337 :
338 : /// Process a ProposerAcceptorMessage.
339 203247 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
340 203247 : if !self.greeting {
341 154044 : self.greeting = true;
342 154044 :
343 154044 : match msg {
344 154044 : ProposerAcceptorMessage::Greeting(ref greeting) => {
345 154044 : tracing::info!(
346 0 : "start handshake with walproposer {:?} {:?}",
347 : self.tcp,
348 : greeting
349 : );
350 154044 : let server_info = ServerInfo {
351 154044 : pg_version: greeting.pg_version,
352 154044 : system_id: greeting.system_id,
353 154044 : wal_seg_size: greeting.wal_seg_size,
354 154044 : };
355 154044 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
356 154044 : self.init_timeline(ttid, server_info, global)?
357 : }
358 : _ => {
359 0 : bail!("unexpected message {msg:?} instead of greeting");
360 : }
361 : }
362 49203 : }
363 :
364 203247 : let tli = global.get(&self.ttid);
365 203247 :
366 203247 : match msg {
367 19800 : ProposerAcceptorMessage::AppendRequest(append_request) => {
368 19800 : self.flush_pending = true;
369 19800 : self.process_sk_msg(
370 19800 : tli,
371 19800 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
372 19800 : )?;
373 : }
374 183447 : other => {
375 183447 : self.process_sk_msg(tli, &other)?;
376 : }
377 : }
378 :
379 203247 : Ok(())
380 203247 : }
381 :
382 : /// Process FlushWAL if needed.
383 1365862 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
384 1365862 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
385 1365862 : if !self.flush_pending {
386 1349553 : return Ok(());
387 16309 : }
388 16309 : self.flush_pending = false;
389 16309 : let shared_state = global.get(&self.ttid);
390 16309 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
391 1365862 : }
392 :
393 : /// Make safekeeper process a message and send a reply to the TCP
394 219556 : fn process_sk_msg(
395 219556 : &mut self,
396 219556 : shared_state: &mut SharedState,
397 219556 : msg: &ProposerAcceptorMessage,
398 219556 : ) -> Result<()> {
399 219556 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
400 219556 : if let Some(reply) = &mut reply {
401 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
402 :
403 193954 : let mut buf = BytesMut::with_capacity(128);
404 193954 : reply.serialize(&mut buf)?;
405 :
406 193954 : self.tcp.send(AnyMessage::Bytes(buf.into()));
407 25602 : }
408 219556 : Ok(())
409 219556 : }
410 : }
411 :
412 : impl Drop for ConnState {
413 197307 : fn drop(&mut self) {
414 197307 : debug!("dropping conn: {:?}", self.tcp);
415 197307 : if !std::thread::panicking() {
416 1887 : self.tcp.close();
417 195420 : }
418 : // TODO: clean up non-fsynced WAL
419 197307 : }
420 : }
|