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 77188 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 77188 : let mut timelines = HashMap::new();
47 :
48 77188 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 58132 : debug!("loading timeline {}", ttid);
50 58132 : let state = disk.state.lock().clone();
51 58132 :
52 58132 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 58132 : }
55 58132 :
56 58132 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 58132 : }
59 58132 :
60 58132 : 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 58132 : }
67 58132 :
68 58132 : let control_store = DiskStateStorage::new(disk.clone());
69 58132 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 58132 : let sk = SafeKeeper::new(control_store, wal_store, conf.my_id)?;
72 58132 : timelines.insert(
73 58132 : ttid,
74 58132 : SharedState {
75 58132 : sk,
76 58132 : disk: disk.clone(),
77 58132 : },
78 58132 : );
79 : }
80 :
81 77188 : Ok(Self {
82 77188 : timelines,
83 77188 : conf,
84 77188 : disk,
85 77188 : })
86 77188 : }
87 :
88 11438 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 11438 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 11438 : }
92 11438 :
93 11438 : debug!("creating new timeline {}", ttid);
94 :
95 11438 : let commit_lsn = Lsn::INVALID;
96 11438 : let local_start_lsn = Lsn::INVALID;
97 11438 :
98 11438 : let state =
99 11438 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
100 11438 :
101 11438 : if state.server.wal_seg_size == 0 {
102 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
103 11438 : }
104 11438 :
105 11438 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
106 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
107 11438 : }
108 11438 :
109 11438 : 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 11438 : }
116 11438 :
117 11438 : let disk_timeline = self.disk.put_state(&ttid, state);
118 11438 : let control_store = DiskStateStorage::new(disk_timeline.clone());
119 11438 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
120 :
121 11438 : let sk = SafeKeeper::new(control_store, wal_store, self.conf.my_id)?;
122 :
123 11438 : self.timelines.insert(
124 11438 : ttid,
125 11438 : SharedState {
126 11438 : sk,
127 11438 : disk: disk_timeline,
128 11438 : },
129 11438 : );
130 11438 : Ok(())
131 11438 : }
132 :
133 234863 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
134 234863 : self.timelines.get_mut(ttid).expect("timeline must exist")
135 234863 : }
136 :
137 155592 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
138 155592 : self.timelines.contains_key(ttid)
139 155592 : }
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 77188 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
154 77188 : let _enter = info_span!("safekeeper", id = os.id()).entered();
155 77188 : debug!("started server");
156 77188 : os.log_event("started;safekeeper".to_owned());
157 77188 : let conf = SafeKeeperConf {
158 77188 : workdir: Utf8PathBuf::from("."),
159 77188 : my_id: NodeId(os.id() as u64),
160 77188 : listen_pg_addr: String::new(),
161 77188 : listen_http_addr: String::new(),
162 77188 : no_sync: false,
163 77188 : broker_endpoint: "/".parse::<Uri>().unwrap(),
164 77188 : broker_keepalive_interval: Duration::from_secs(0),
165 77188 : heartbeat_timeout: Duration::from_secs(0),
166 77188 : remote_storage: None,
167 77188 : max_offloader_lag_bytes: 0,
168 77188 : wal_backup_enabled: false,
169 77188 : listen_pg_addr_tenant_only: None,
170 77188 : advertise_pg_addr: None,
171 77188 : availability_zone: None,
172 77188 : peer_recovery_enabled: false,
173 77188 : backup_parallel_jobs: 0,
174 77188 : pg_auth: None,
175 77188 : pg_tenant_only_auth: None,
176 77188 : http_auth: None,
177 77188 : sk_auth_token: None,
178 77188 : current_thread_runtime: false,
179 77188 : walsenders_keep_horizon: false,
180 77188 : partial_backup_enabled: false,
181 77188 : partial_backup_timeout: Duration::from_secs(0),
182 77188 : disable_periodic_broker_push: false,
183 77188 : };
184 :
185 77188 : let mut global = GlobalMap::new(disk, conf.clone())?;
186 77188 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
187 :
188 77188 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
189 58132 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
190 58132 : let commit_lsn = shared_state.sk.state.commit_lsn;
191 58132 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
192 58132 : }
193 :
194 77188 : let node_events = os.node_events();
195 77188 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
196 77188 : let mut epoll_idx: Vec<usize> = vec![];
197 :
198 : // TODO: batch events processing (multiple events per tick)
199 : loop {
200 595821 : epoll_vec.clear();
201 595821 : epoll_idx.clear();
202 595821 :
203 595821 : // node events channel
204 595821 : epoll_vec.push(Box::new(node_events.clone()));
205 595821 : epoll_idx.push(0);
206 :
207 : // tcp connections
208 2149273 : for conn in conns.values() {
209 2149273 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
210 2149273 : epoll_idx.push(conn.tcp.connection_id());
211 2149273 : }
212 :
213 : // waiting for the next message
214 595821 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
215 595821 :
216 595821 : if index == 0 {
217 : // got a new connection
218 199345 : match node_events.must_recv() {
219 199345 : NodeEvent::Accept(tcp) => {
220 199345 : conns.insert(
221 199345 : tcp.connection_id(),
222 199345 : ConnState {
223 199345 : tcp,
224 199345 : greeting: false,
225 199345 : ttid: TenantTimelineId::empty(),
226 199345 : flush_pending: false,
227 199345 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
228 : },
229 : );
230 : }
231 0 : NodeEvent::Internal(_) => unreachable!(),
232 : }
233 199345 : continue;
234 396476 : }
235 396476 :
236 396476 : let connection_id = epoll_idx[index];
237 396476 : let conn = conns.get_mut(&connection_id).unwrap();
238 396476 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
239 :
240 : loop {
241 721561 : let event = match next_event {
242 404469 : Some(event) => event,
243 317092 : None => break,
244 : };
245 :
246 404469 : match event {
247 292163 : NetEvent::Message(msg) => {
248 292163 : let res = conn.process_any(msg, &mut global);
249 292163 : if res.is_err() {
250 79384 : debug!("conn {:?} error: {:#}", connection_id, res.unwrap_err());
251 2196 : conns.remove(&connection_id);
252 2196 : break;
253 212779 : }
254 : }
255 112306 : NetEvent::Closed => {
256 112306 : // TODO: remove from conns?
257 112306 : }
258 : }
259 :
260 325085 : next_event = conn.tcp.recv_chan().try_recv();
261 : }
262 :
263 1408387 : conns.retain(|_, conn| {
264 1408387 : let res = conn.flush(&mut global);
265 1408387 : if res.is_err() {
266 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
267 1408387 : }
268 1408387 : res.is_ok()
269 1408387 : });
270 : }
271 0 : }
272 :
273 : impl ConnState {
274 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
275 214975 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
276 214975 : if let AnyMessage::Bytes(copy_data) = any {
277 214975 : let repl_prefix = b"START_REPLICATION ";
278 214975 : if !self.greeting && copy_data.starts_with(repl_prefix) {
279 2196 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
280 2196 : bail!("finished processing START_REPLICATION")
281 212779 : }
282 :
283 212779 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
284 212779 : debug!("got msg: {:?}", msg);
285 212779 : self.process(msg, global)
286 : } else {
287 0 : bail!("unexpected message, expected AnyMessage::Bytes");
288 : }
289 214975 : }
290 :
291 : /// Process START_REPLICATION request.
292 2196 : fn process_start_replication(
293 2196 : &mut self,
294 2196 : copy_data: Bytes,
295 2196 : global: &mut GlobalMap,
296 2196 : ) -> Result<()> {
297 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
298 2196 : let str = String::from_utf8(copy_data.to_vec())?;
299 :
300 2196 : let mut parts = str.split(' ');
301 2196 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
302 2196 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
303 2196 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
304 2196 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
305 :
306 2196 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
307 2196 : let shared_state = global.get(&ttid);
308 2196 :
309 2196 : // read bytes from start_lsn to end_lsn
310 2196 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
311 2196 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
312 2196 :
313 2196 : // send bytes to the client
314 2196 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
315 2196 : Ok(())
316 2196 : }
317 :
318 : /// Get or create a timeline.
319 155592 : fn init_timeline(
320 155592 : &mut self,
321 155592 : ttid: TenantTimelineId,
322 155592 : server_info: ServerInfo,
323 155592 : global: &mut GlobalMap,
324 155592 : ) -> Result<()> {
325 155592 : self.ttid = ttid;
326 155592 : if global.has_tli(&ttid) {
327 144154 : return Ok(());
328 11438 : }
329 11438 :
330 11438 : global.create(ttid, server_info)
331 155592 : }
332 :
333 : /// Process a ProposerAcceptorMessage.
334 212779 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
335 212779 : if !self.greeting {
336 155592 : self.greeting = true;
337 155592 :
338 155592 : match msg {
339 155592 : ProposerAcceptorMessage::Greeting(ref greeting) => {
340 155592 : tracing::info!(
341 0 : "start handshake with walproposer {:?} {:?}",
342 : self.tcp,
343 : greeting
344 : );
345 155592 : let server_info = ServerInfo {
346 155592 : pg_version: greeting.pg_version,
347 155592 : system_id: greeting.system_id,
348 155592 : wal_seg_size: greeting.wal_seg_size,
349 155592 : };
350 155592 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
351 155592 : self.init_timeline(ttid, server_info, global)?
352 : }
353 : _ => {
354 0 : bail!("unexpected message {msg:?} instead of greeting");
355 : }
356 : }
357 57187 : }
358 :
359 212779 : let tli = global.get(&self.ttid);
360 212779 :
361 212779 : match msg {
362 23971 : ProposerAcceptorMessage::AppendRequest(append_request) => {
363 23971 : self.flush_pending = true;
364 23971 : self.process_sk_msg(
365 23971 : tli,
366 23971 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
367 23971 : )?;
368 : }
369 188808 : other => {
370 188808 : self.process_sk_msg(tli, &other)?;
371 : }
372 : }
373 :
374 212779 : Ok(())
375 212779 : }
376 :
377 : /// Process FlushWAL if needed.
378 1408387 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
379 1408387 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
380 1408387 : if !self.flush_pending {
381 1388499 : return Ok(());
382 19888 : }
383 19888 : self.flush_pending = false;
384 19888 : let shared_state = global.get(&self.ttid);
385 19888 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
386 1408387 : }
387 :
388 : /// Make safekeeper process a message and send a reply to the TCP
389 232667 : fn process_sk_msg(
390 232667 : &mut self,
391 232667 : shared_state: &mut SharedState,
392 232667 : msg: &ProposerAcceptorMessage,
393 232667 : ) -> Result<()> {
394 232667 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
395 232667 : if let Some(reply) = &mut reply {
396 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
397 :
398 201344 : let mut buf = BytesMut::with_capacity(128);
399 201344 : reply.serialize(&mut buf)?;
400 :
401 201344 : self.tcp.send(AnyMessage::Bytes(buf.into()));
402 31323 : }
403 232667 : Ok(())
404 232667 : }
405 : }
406 :
407 : impl Drop for ConnState {
408 199289 : fn drop(&mut self) {
409 199289 : debug!("dropping conn: {:?}", self.tcp);
410 199289 : if !std::thread::panicking() {
411 2196 : self.tcp.close();
412 197093 : }
413 : // TODO: clean up non-fsynced WAL
414 199289 : }
415 : }
|
