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};
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 76407 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 76407 : let mut timelines = HashMap::new();
47 :
48 76407 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 57312 : debug!("loading timeline {}", ttid);
50 57312 : let state = disk.state.lock().clone();
51 57312 :
52 57312 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 57312 : }
55 57312 :
56 57312 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 57312 : }
59 57312 :
60 57312 : 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 57312 : }
67 57312 :
68 57312 : let control_store = DiskStateStorage::new(disk.clone());
69 57312 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 57312 : let sk = SafeKeeper::new(control_store, wal_store, conf.my_id)?;
72 57312 : timelines.insert(
73 57312 : ttid,
74 57312 : SharedState {
75 57312 : sk,
76 57312 : disk: disk.clone(),
77 57312 : },
78 57312 : );
79 : }
80 :
81 76407 : Ok(Self {
82 76407 : timelines,
83 76407 : conf,
84 76407 : disk,
85 76407 : })
86 76407 : }
87 :
88 11447 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 11447 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 11447 : }
92 11447 :
93 11447 : debug!("creating new timeline {}", ttid);
94 :
95 11447 : let commit_lsn = Lsn::INVALID;
96 11447 : let local_start_lsn = Lsn::INVALID;
97 11447 :
98 11447 : let state =
99 11447 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
100 11447 :
101 11447 : if state.server.wal_seg_size == 0 {
102 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
103 11447 : }
104 11447 :
105 11447 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
106 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
107 11447 : }
108 11447 :
109 11447 : 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 11447 : }
116 11447 :
117 11447 : let disk_timeline = self.disk.put_state(&ttid, state);
118 11447 : let control_store = DiskStateStorage::new(disk_timeline.clone());
119 11447 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
120 :
121 11447 : let sk = SafeKeeper::new(control_store, wal_store, self.conf.my_id)?;
122 :
123 11447 : self.timelines.insert(
124 11447 : ttid,
125 11447 : SharedState {
126 11447 : sk,
127 11447 : disk: disk_timeline,
128 11447 : },
129 11447 : );
130 11447 : Ok(())
131 11447 : }
132 :
133 223156 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
134 223156 : self.timelines.get_mut(ttid).expect("timeline must exist")
135 223156 : }
136 :
137 153540 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
138 153540 : self.timelines.contains_key(ttid)
139 153540 : }
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 76407 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
154 76407 : let _enter = info_span!("safekeeper", id = os.id()).entered();
155 76407 : debug!("started server");
156 76407 : os.log_event("started;safekeeper".to_owned());
157 76407 : let conf = SafeKeeperConf {
158 76407 : workdir: Utf8PathBuf::from("."),
159 76407 : my_id: NodeId(os.id() as u64),
160 76407 : listen_pg_addr: String::new(),
161 76407 : listen_http_addr: String::new(),
162 76407 : no_sync: false,
163 76407 : broker_endpoint: "/".parse::<Uri>().unwrap(),
164 76407 : broker_keepalive_interval: Duration::from_secs(0),
165 76407 : heartbeat_timeout: Duration::from_secs(0),
166 76407 : remote_storage: None,
167 76407 : max_offloader_lag_bytes: 0,
168 76407 : wal_backup_enabled: false,
169 76407 : listen_pg_addr_tenant_only: None,
170 76407 : advertise_pg_addr: None,
171 76407 : availability_zone: None,
172 76407 : peer_recovery_enabled: false,
173 76407 : backup_parallel_jobs: 0,
174 76407 : pg_auth: None,
175 76407 : pg_tenant_only_auth: None,
176 76407 : http_auth: None,
177 76407 : current_thread_runtime: false,
178 76407 : walsenders_keep_horizon: false,
179 76407 : };
180 :
181 76407 : let mut global = GlobalMap::new(disk, conf.clone())?;
182 76407 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
183 :
184 76407 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
185 57312 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
186 57312 : let commit_lsn = shared_state.sk.state.commit_lsn;
187 57312 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
188 57312 : }
189 :
190 76407 : let node_events = os.node_events();
191 76407 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
192 76407 : let mut epoll_idx: Vec<usize> = vec![];
193 :
194 : // TODO: batch events processing (multiple events per tick)
195 : loop {
196 583168 : epoll_vec.clear();
197 583168 : epoll_idx.clear();
198 583168 :
199 583168 : // node events channel
200 583168 : epoll_vec.push(Box::new(node_events.clone()));
201 583168 : epoll_idx.push(0);
202 :
203 : // tcp connections
204 2100219 : for conn in conns.values() {
205 2100219 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
206 2100219 : epoll_idx.push(conn.tcp.connection_id());
207 2100219 : }
208 :
209 : // waiting for the next message
210 583168 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
211 583168 :
212 583168 : if index == 0 {
213 : // got a new connection
214 196988 : match node_events.must_recv() {
215 196988 : NodeEvent::Accept(tcp) => {
216 196988 : conns.insert(
217 196988 : tcp.connection_id(),
218 196988 : ConnState {
219 196988 : tcp,
220 196988 : greeting: false,
221 196988 : ttid: TenantTimelineId::empty(),
222 196988 : flush_pending: false,
223 196988 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
224 : },
225 : );
226 : }
227 0 : NodeEvent::Internal(_) => unreachable!(),
228 : }
229 196988 : continue;
230 386180 : }
231 386180 :
232 386180 : let connection_id = epoll_idx[index];
233 386180 : let conn = conns.get_mut(&connection_id).unwrap();
234 386180 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
235 :
236 : loop {
237 701382 : let event = match next_event {
238 393405 : Some(event) => event,
239 307977 : None => break,
240 : };
241 :
242 393405 : match event {
243 281736 : NetEvent::Message(msg) => {
244 281736 : let res = conn.process_any(msg, &mut global);
245 281736 : if res.is_err() {
246 78203 : debug!("conn {:?} error: {:#}", connection_id, res.unwrap_err());
247 1796 : conns.remove(&connection_id);
248 1796 : break;
249 203533 : }
250 : }
251 111669 : NetEvent::Closed => {
252 111669 : // TODO: remove from conns?
253 111669 : }
254 : }
255 :
256 315202 : next_event = conn.tcp.recv_chan().try_recv();
257 : }
258 :
259 1367649 : conns.retain(|_, conn| {
260 1367649 : let res = conn.flush(&mut global);
261 1367649 : if res.is_err() {
262 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
263 1367649 : }
264 1367649 : res.is_ok()
265 1367649 : });
266 : }
267 0 : }
268 :
269 : impl ConnState {
270 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
271 205329 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
272 205329 : if let AnyMessage::Bytes(copy_data) = any {
273 205329 : let repl_prefix = b"START_REPLICATION ";
274 205329 : if !self.greeting && copy_data.starts_with(repl_prefix) {
275 1796 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
276 1796 : bail!("finished processing START_REPLICATION")
277 203533 : }
278 :
279 203533 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
280 203533 : debug!("got msg: {:?}", msg);
281 203533 : self.process(msg, global)
282 : } else {
283 0 : bail!("unexpected message, expected AnyMessage::Bytes");
284 : }
285 205329 : }
286 :
287 : /// Process START_REPLICATION request.
288 1796 : fn process_start_replication(
289 1796 : &mut self,
290 1796 : copy_data: Bytes,
291 1796 : global: &mut GlobalMap,
292 1796 : ) -> Result<()> {
293 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
294 1796 : let str = String::from_utf8(copy_data.to_vec())?;
295 :
296 1796 : let mut parts = str.split(' ');
297 1796 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
298 1796 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
299 1796 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
300 1796 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
301 :
302 1796 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
303 1796 : let shared_state = global.get(&ttid);
304 1796 :
305 1796 : // read bytes from start_lsn to end_lsn
306 1796 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
307 1796 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
308 1796 :
309 1796 : // send bytes to the client
310 1796 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
311 1796 : Ok(())
312 1796 : }
313 :
314 : /// Get or create a timeline.
315 153540 : fn init_timeline(
316 153540 : &mut self,
317 153540 : ttid: TenantTimelineId,
318 153540 : server_info: ServerInfo,
319 153540 : global: &mut GlobalMap,
320 153540 : ) -> Result<()> {
321 153540 : self.ttid = ttid;
322 153540 : if global.has_tli(&ttid) {
323 142093 : return Ok(());
324 11447 : }
325 11447 :
326 11447 : global.create(ttid, server_info)
327 153540 : }
328 :
329 : /// Process a ProposerAcceptorMessage.
330 203533 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
331 203533 : if !self.greeting {
332 153540 : self.greeting = true;
333 153540 :
334 153540 : match msg {
335 153540 : ProposerAcceptorMessage::Greeting(ref greeting) => {
336 153540 : tracing::info!(
337 302 : "start handshake with walproposer {:?} {:?}",
338 302 : self.tcp,
339 302 : greeting
340 302 : );
341 153540 : let server_info = ServerInfo {
342 153540 : pg_version: greeting.pg_version,
343 153540 : system_id: greeting.system_id,
344 153540 : wal_seg_size: greeting.wal_seg_size,
345 153540 : };
346 153540 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
347 153540 : self.init_timeline(ttid, server_info, global)?
348 : }
349 : _ => {
350 0 : bail!("unexpected message {msg:?} instead of greeting");
351 : }
352 : }
353 49993 : }
354 :
355 203533 : let tli = global.get(&self.ttid);
356 203533 :
357 203533 : match msg {
358 21452 : ProposerAcceptorMessage::AppendRequest(append_request) => {
359 21452 : self.flush_pending = true;
360 21452 : self.process_sk_msg(
361 21452 : tli,
362 21452 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
363 21452 : )?;
364 : }
365 182081 : other => {
366 182081 : self.process_sk_msg(tli, &other)?;
367 : }
368 : }
369 :
370 203533 : Ok(())
371 203533 : }
372 :
373 : /// Process FlushWAL if needed.
374 1367649 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
375 1367649 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
376 1367649 : if !self.flush_pending {
377 1349822 : return Ok(());
378 17827 : }
379 17827 : self.flush_pending = false;
380 17827 : let shared_state = global.get(&self.ttid);
381 17827 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
382 1367649 : }
383 :
384 : /// Make safekeeper process a message and send a reply to the TCP
385 221360 : fn process_sk_msg(
386 221360 : &mut self,
387 221360 : shared_state: &mut SharedState,
388 221360 : msg: &ProposerAcceptorMessage,
389 221360 : ) -> Result<()> {
390 221360 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
391 221360 : if let Some(reply) = &mut reply {
392 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
393 :
394 193490 : let mut buf = BytesMut::with_capacity(128);
395 193490 : reply.serialize(&mut buf)?;
396 :
397 193490 : self.tcp.send(AnyMessage::Bytes(buf.into()));
398 27870 : }
399 221360 : Ok(())
400 221360 : }
401 : }
402 :
403 : impl Drop for ConnState {
404 196932 : fn drop(&mut self) {
405 196932 : debug!("dropping conn: {:?}", self.tcp);
406 196932 : if !std::thread::panicking() {
407 1796 : self.tcp.close();
408 195136 : }
409 : // TODO: clean up non-fsynced WAL
410 196932 : }
411 : }
|
