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 78103 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 78103 : let mut timelines = HashMap::new();
47 :
48 78103 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 59042 : debug!("loading timeline {}", ttid);
50 59042 : let state = disk.state.lock().clone();
51 59042 :
52 59042 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 59042 : }
55 59042 :
56 59042 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 59042 : }
59 59042 :
60 59042 : 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 59042 : }
67 59042 :
68 59042 : let control_store = DiskStateStorage::new(disk.clone());
69 59042 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 59042 : let sk = SafeKeeper::new(control_store, wal_store, conf.my_id)?;
72 59042 : timelines.insert(
73 59042 : ttid,
74 59042 : SharedState {
75 59042 : sk,
76 59042 : disk: disk.clone(),
77 59042 : },
78 59042 : );
79 : }
80 :
81 78103 : Ok(Self {
82 78103 : timelines,
83 78103 : conf,
84 78103 : disk,
85 78103 : })
86 78103 : }
87 :
88 11382 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 11382 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 11382 : }
92 11382 :
93 11382 : debug!("creating new timeline {}", ttid);
94 :
95 11382 : let commit_lsn = Lsn::INVALID;
96 11382 : let local_start_lsn = Lsn::INVALID;
97 11382 :
98 11382 : let state =
99 11382 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
100 11382 :
101 11382 : if state.server.wal_seg_size == 0 {
102 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
103 11382 : }
104 11382 :
105 11382 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
106 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
107 11382 : }
108 11382 :
109 11382 : 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 11382 : }
116 11382 :
117 11382 : let disk_timeline = self.disk.put_state(&ttid, state);
118 11382 : let control_store = DiskStateStorage::new(disk_timeline.clone());
119 11382 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
120 :
121 11382 : let sk = SafeKeeper::new(control_store, wal_store, self.conf.my_id)?;
122 :
123 11382 : self.timelines.insert(
124 11382 : ttid,
125 11382 : SharedState {
126 11382 : sk,
127 11382 : disk: disk_timeline,
128 11382 : },
129 11382 : );
130 11382 : Ok(())
131 11382 : }
132 :
133 229321 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
134 229321 : self.timelines.get_mut(ttid).expect("timeline must exist")
135 229321 : }
136 :
137 158218 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
138 158218 : self.timelines.contains_key(ttid)
139 158218 : }
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 78103 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
154 78103 : let _enter = info_span!("safekeeper", id = os.id()).entered();
155 78103 : debug!("started server");
156 78103 : os.log_event("started;safekeeper".to_owned());
157 78103 : let conf = SafeKeeperConf {
158 78103 : workdir: Utf8PathBuf::from("."),
159 78103 : my_id: NodeId(os.id() as u64),
160 78103 : listen_pg_addr: String::new(),
161 78103 : listen_http_addr: String::new(),
162 78103 : no_sync: false,
163 78103 : broker_endpoint: "/".parse::<Uri>().unwrap(),
164 78103 : broker_keepalive_interval: Duration::from_secs(0),
165 78103 : heartbeat_timeout: Duration::from_secs(0),
166 78103 : remote_storage: None,
167 78103 : max_offloader_lag_bytes: 0,
168 78103 : wal_backup_enabled: false,
169 78103 : listen_pg_addr_tenant_only: None,
170 78103 : advertise_pg_addr: None,
171 78103 : availability_zone: None,
172 78103 : peer_recovery_enabled: false,
173 78103 : backup_parallel_jobs: 0,
174 78103 : pg_auth: None,
175 78103 : pg_tenant_only_auth: None,
176 78103 : http_auth: None,
177 78103 : current_thread_runtime: false,
178 78103 : };
179 :
180 78103 : let mut global = GlobalMap::new(disk, conf.clone())?;
181 78103 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
182 :
183 78103 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
184 59042 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
185 59042 : let commit_lsn = shared_state.sk.state.commit_lsn;
186 59042 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
187 59042 : }
188 :
189 78103 : let node_events = os.node_events();
190 78103 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
191 78103 : let mut epoll_idx: Vec<usize> = vec![];
192 :
193 : // TODO: batch events processing (multiple events per tick)
194 : loop {
195 598216 : epoll_vec.clear();
196 598216 : epoll_idx.clear();
197 598216 :
198 598216 : // node events channel
199 598216 : epoll_vec.push(Box::new(node_events.clone()));
200 598216 : epoll_idx.push(0);
201 :
202 : // tcp connections
203 2164587 : for conn in conns.values() {
204 2164587 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
205 2164587 : epoll_idx.push(conn.tcp.connection_id());
206 2164587 : }
207 :
208 : // waiting for the next message
209 598216 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
210 598216 :
211 598216 : if index == 0 {
212 : // got a new connection
213 201477 : match node_events.must_recv() {
214 201477 : NodeEvent::Accept(tcp) => {
215 201477 : conns.insert(
216 201477 : tcp.connection_id(),
217 201477 : ConnState {
218 201477 : tcp,
219 201477 : greeting: false,
220 201477 : ttid: TenantTimelineId::empty(),
221 201477 : flush_pending: false,
222 201477 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
223 : },
224 : );
225 : }
226 0 : NodeEvent::Internal(_) => unreachable!(),
227 : }
228 201477 : continue;
229 396739 : }
230 396739 :
231 396739 : let connection_id = epoll_idx[index];
232 396739 : let conn = conns.get_mut(&connection_id).unwrap();
233 396739 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
234 :
235 : loop {
236 720595 : let event = match next_event {
237 403909 : Some(event) => event,
238 316686 : None => break,
239 : };
240 :
241 403909 : match event {
242 289779 : NetEvent::Message(msg) => {
243 289779 : let res = conn.process_any(msg, &mut global);
244 289779 : if res.is_err() {
245 80053 : debug!("conn {:?} error: {:#}", connection_id, res.unwrap_err());
246 1950 : conns.remove(&connection_id);
247 1950 : break;
248 209726 : }
249 : }
250 114130 : NetEvent::Closed => {
251 114130 : // TODO: remove from conns?
252 114130 : }
253 : }
254 :
255 323856 : next_event = conn.tcp.recv_chan().try_recv();
256 : }
257 :
258 1412644 : conns.retain(|_, conn| {
259 1412644 : let res = conn.flush(&mut global);
260 1412644 : if res.is_err() {
261 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
262 1412644 : }
263 1412644 : res.is_ok()
264 1412644 : });
265 : }
266 0 : }
267 :
268 : impl ConnState {
269 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
270 211676 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
271 211676 : if let AnyMessage::Bytes(copy_data) = any {
272 211676 : let repl_prefix = b"START_REPLICATION ";
273 211676 : if !self.greeting && copy_data.starts_with(repl_prefix) {
274 1950 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
275 1950 : bail!("finished processing START_REPLICATION")
276 209726 : }
277 :
278 209726 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
279 209726 : debug!("got msg: {:?}", msg);
280 209726 : self.process(msg, global)
281 : } else {
282 0 : bail!("unexpected message, expected AnyMessage::Bytes");
283 : }
284 211676 : }
285 :
286 : /// Process START_REPLICATION request.
287 1950 : fn process_start_replication(
288 1950 : &mut self,
289 1950 : copy_data: Bytes,
290 1950 : global: &mut GlobalMap,
291 1950 : ) -> Result<()> {
292 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
293 1950 : let str = String::from_utf8(copy_data.to_vec())?;
294 :
295 1950 : let mut parts = str.split(' ');
296 1950 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
297 1950 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
298 1950 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
299 1950 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
300 :
301 1950 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
302 1950 : let shared_state = global.get(&ttid);
303 1950 :
304 1950 : // read bytes from start_lsn to end_lsn
305 1950 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
306 1950 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
307 1950 :
308 1950 : // send bytes to the client
309 1950 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
310 1950 : Ok(())
311 1950 : }
312 :
313 : /// Get or create a timeline.
314 158218 : fn init_timeline(
315 158218 : &mut self,
316 158218 : ttid: TenantTimelineId,
317 158218 : server_info: ServerInfo,
318 158218 : global: &mut GlobalMap,
319 158218 : ) -> Result<()> {
320 158218 : self.ttid = ttid;
321 158218 : if global.has_tli(&ttid) {
322 146836 : return Ok(());
323 11382 : }
324 11382 :
325 11382 : global.create(ttid, server_info)
326 158218 : }
327 :
328 : /// Process a ProposerAcceptorMessage.
329 209726 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
330 209726 : if !self.greeting {
331 158218 : self.greeting = true;
332 158218 :
333 158218 : match msg {
334 158218 : ProposerAcceptorMessage::Greeting(ref greeting) => {
335 158218 : tracing::info!(
336 302 : "start handshake with walproposer {:?} {:?}",
337 302 : self.tcp,
338 302 : greeting
339 302 : );
340 158218 : let server_info = ServerInfo {
341 158218 : pg_version: greeting.pg_version,
342 158218 : system_id: greeting.system_id,
343 158218 : wal_seg_size: greeting.wal_seg_size,
344 158218 : };
345 158218 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
346 158218 : self.init_timeline(ttid, server_info, global)?
347 : }
348 : _ => {
349 0 : bail!("unexpected message {msg:?} instead of greeting");
350 : }
351 : }
352 51508 : }
353 :
354 209726 : let tli = global.get(&self.ttid);
355 209726 :
356 209726 : match msg {
357 21303 : ProposerAcceptorMessage::AppendRequest(append_request) => {
358 21303 : self.flush_pending = true;
359 21303 : self.process_sk_msg(
360 21303 : tli,
361 21303 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
362 21303 : )?;
363 : }
364 188423 : other => {
365 188423 : self.process_sk_msg(tli, &other)?;
366 : }
367 : }
368 :
369 209726 : Ok(())
370 209726 : }
371 :
372 : /// Process FlushWAL if needed.
373 1412644 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
374 1412644 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
375 1412644 : if !self.flush_pending {
376 1394999 : return Ok(());
377 17645 : }
378 17645 : self.flush_pending = false;
379 17645 : let shared_state = global.get(&self.ttid);
380 17645 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
381 1412644 : }
382 :
383 : /// Make safekeeper process a message and send a reply to the TCP
384 227371 : fn process_sk_msg(
385 227371 : &mut self,
386 227371 : shared_state: &mut SharedState,
387 227371 : msg: &ProposerAcceptorMessage,
388 227371 : ) -> Result<()> {
389 227371 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
390 227371 : if let Some(reply) = &mut reply {
391 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
392 :
393 199686 : let mut buf = BytesMut::with_capacity(128);
394 199686 : reply.serialize(&mut buf)?;
395 :
396 199686 : self.tcp.send(AnyMessage::Bytes(buf.into()));
397 27685 : }
398 227371 : Ok(())
399 227371 : }
400 : }
401 :
402 : impl Drop for ConnState {
403 201421 : fn drop(&mut self) {
404 201421 : debug!("dropping conn: {:?}", self.tcp);
405 201421 : if !std::thread::panicking() {
406 1950 : self.tcp.close();
407 199471 : }
408 : // TODO: clean up non-fsynced WAL
409 201421 : }
410 : }
|
