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 http::Uri;
17 : use safekeeper::{
18 : safekeeper::{ProposerAcceptorMessage, SafeKeeper, UNKNOWN_SERVER_VERSION},
19 : state::{TimelinePersistentState, TimelineState},
20 : timeline::TimelineError,
21 : wal_storage::Storage,
22 : SafeKeeperConf,
23 : };
24 : use safekeeper_api::ServerInfo;
25 : use tracing::{debug, info_span, warn};
26 : use utils::{
27 : id::{NodeId, TenantId, TenantTimelineId, TimelineId},
28 : lsn::Lsn,
29 : };
30 :
31 : use super::safekeeper_disk::{DiskStateStorage, DiskWALStorage, SafekeeperDisk, TimelineDisk};
32 :
33 : struct SharedState {
34 : sk: SafeKeeper<DiskStateStorage, DiskWALStorage>,
35 : disk: Arc<TimelineDisk>,
36 : }
37 :
38 : struct GlobalMap {
39 : timelines: HashMap<TenantTimelineId, SharedState>,
40 : conf: SafeKeeperConf,
41 : disk: Arc<SafekeeperDisk>,
42 : }
43 :
44 : impl GlobalMap {
45 : /// Restores global state from disk.
46 10070 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
47 10070 : let mut timelines = HashMap::new();
48 :
49 10070 : for (&ttid, disk) in disk.timelines.lock().iter() {
50 7669 : debug!("loading timeline {}", ttid);
51 7669 : let state = disk.state.lock().clone();
52 7669 :
53 7669 : if state.server.wal_seg_size == 0 {
54 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
55 7669 : }
56 7669 :
57 7669 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
58 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
59 7669 : }
60 7669 :
61 7669 : if state.commit_lsn < state.local_start_lsn {
62 0 : bail!(
63 0 : "commit_lsn {} is smaller than local_start_lsn {}",
64 0 : state.commit_lsn,
65 0 : state.local_start_lsn
66 0 : );
67 7669 : }
68 7669 :
69 7669 : let control_store = DiskStateStorage::new(disk.clone());
70 7669 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
71 :
72 7669 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
73 7669 : timelines.insert(
74 7669 : ttid,
75 7669 : SharedState {
76 7669 : sk,
77 7669 : disk: disk.clone(),
78 7669 : },
79 7669 : );
80 : }
81 :
82 10070 : Ok(Self {
83 10070 : timelines,
84 10070 : conf,
85 10070 : disk,
86 10070 : })
87 10070 : }
88 :
89 1445 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
90 1445 : if self.timelines.contains_key(&ttid) {
91 0 : bail!("timeline {} already exists", ttid);
92 1445 : }
93 1445 :
94 1445 : debug!("creating new timeline {}", ttid);
95 :
96 1445 : let commit_lsn = Lsn::INVALID;
97 1445 : let local_start_lsn = Lsn::INVALID;
98 :
99 1445 : let state =
100 1445 : TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?;
101 :
102 1445 : let disk_timeline = self.disk.put_state(&ttid, state);
103 1445 : let control_store = DiskStateStorage::new(disk_timeline.clone());
104 1445 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
105 :
106 1445 : let sk = SafeKeeper::new(
107 1445 : TimelineState::new(control_store),
108 1445 : wal_store,
109 1445 : self.conf.my_id,
110 1445 : )?;
111 :
112 1445 : self.timelines.insert(
113 1445 : ttid,
114 1445 : SharedState {
115 1445 : sk,
116 1445 : disk: disk_timeline,
117 1445 : },
118 1445 : );
119 1445 : Ok(())
120 1445 : }
121 :
122 29246 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
123 29246 : self.timelines.get_mut(ttid).expect("timeline must exist")
124 29246 : }
125 :
126 19878 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
127 19878 : self.timelines.contains_key(ttid)
128 19878 : }
129 : }
130 :
131 : /// State of a single connection to walproposer.
132 : struct ConnState {
133 : tcp: TCP,
134 :
135 : greeting: bool,
136 : ttid: TenantTimelineId,
137 : flush_pending: bool,
138 :
139 : runtime: tokio::runtime::Runtime,
140 : }
141 :
142 10070 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
143 10070 : let _enter = info_span!("safekeeper", id = os.id()).entered();
144 10070 : debug!("started server");
145 10070 : os.log_event("started;safekeeper".to_owned());
146 10070 : let conf = SafeKeeperConf {
147 10070 : workdir: Utf8PathBuf::from("."),
148 10070 : my_id: NodeId(os.id() as u64),
149 10070 : listen_pg_addr: String::new(),
150 10070 : listen_http_addr: String::new(),
151 10070 : no_sync: false,
152 10070 : broker_endpoint: "/".parse::<Uri>().unwrap(),
153 10070 : broker_keepalive_interval: Duration::from_secs(0),
154 10070 : heartbeat_timeout: Duration::from_secs(0),
155 10070 : remote_storage: None,
156 10070 : max_offloader_lag_bytes: 0,
157 10070 : wal_backup_enabled: false,
158 10070 : listen_pg_addr_tenant_only: None,
159 10070 : advertise_pg_addr: None,
160 10070 : availability_zone: None,
161 10070 : peer_recovery_enabled: false,
162 10070 : backup_parallel_jobs: 0,
163 10070 : pg_auth: None,
164 10070 : pg_tenant_only_auth: None,
165 10070 : http_auth: None,
166 10070 : sk_auth_token: None,
167 10070 : current_thread_runtime: false,
168 10070 : walsenders_keep_horizon: false,
169 10070 : partial_backup_timeout: Duration::from_secs(0),
170 10070 : disable_periodic_broker_push: false,
171 10070 : enable_offload: false,
172 10070 : delete_offloaded_wal: false,
173 10070 : control_file_save_interval: Duration::from_secs(1),
174 10070 : partial_backup_concurrency: 1,
175 10070 : eviction_min_resident: Duration::ZERO,
176 10070 : };
177 :
178 10070 : let mut global = GlobalMap::new(disk, conf.clone())?;
179 10070 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
180 :
181 10070 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
182 7669 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
183 7669 : let commit_lsn = shared_state.sk.state.commit_lsn;
184 7669 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
185 7669 : }
186 :
187 10070 : let node_events = os.node_events();
188 10070 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
189 10070 : let mut epoll_idx: Vec<usize> = vec![];
190 :
191 : // TODO: batch events processing (multiple events per tick)
192 : loop {
193 75770 : epoll_vec.clear();
194 75770 : epoll_idx.clear();
195 75770 :
196 75770 : // node events channel
197 75770 : epoll_vec.push(Box::new(node_events.clone()));
198 75770 : epoll_idx.push(0);
199 :
200 : // tcp connections
201 266724 : for conn in conns.values() {
202 266724 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
203 266724 : epoll_idx.push(conn.tcp.connection_id());
204 266724 : }
205 :
206 : // waiting for the next message
207 75770 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
208 75770 :
209 75770 : if index == 0 {
210 : // got a new connection
211 25237 : match node_events.must_recv() {
212 25237 : NodeEvent::Accept(tcp) => {
213 25237 : conns.insert(
214 25237 : tcp.connection_id(),
215 25237 : ConnState {
216 25237 : tcp,
217 25237 : greeting: false,
218 25237 : ttid: TenantTimelineId::empty(),
219 25237 : flush_pending: false,
220 25237 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
221 : },
222 : );
223 : }
224 0 : NodeEvent::Internal(_) => unreachable!(),
225 : }
226 25237 : continue;
227 50533 : }
228 50533 :
229 50533 : let connection_id = epoll_idx[index];
230 50533 : let conn = conns.get_mut(&connection_id).unwrap();
231 50533 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
232 :
233 : loop {
234 91720 : let event = match next_event {
235 51475 : Some(event) => event,
236 40245 : None => break,
237 : };
238 :
239 51475 : match event {
240 36814 : NetEvent::Message(msg) => {
241 36814 : let res = conn.process_any(msg, &mut global);
242 36814 : if res.is_err() {
243 10288 : let e = res.unwrap_err();
244 10288 : let estr = e.to_string();
245 10288 : if !estr.contains("finished processing START_REPLICATION") {
246 10070 : warn!("conn {:?} error: {:?}", connection_id, e);
247 0 : panic!("unexpected error at safekeeper: {:#}", e);
248 218 : }
249 218 : conns.remove(&connection_id);
250 218 : break;
251 26526 : }
252 : }
253 14661 : NetEvent::Closed => {
254 14661 : // TODO: remove from conns?
255 14661 : }
256 : }
257 :
258 41187 : next_event = conn.tcp.recv_chan().try_recv();
259 : }
260 :
261 174467 : conns.retain(|_, conn| {
262 174467 : let res = conn.flush(&mut global);
263 174467 : if res.is_err() {
264 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
265 174467 : }
266 174467 : res.is_ok()
267 174467 : });
268 : }
269 0 : }
270 :
271 : impl ConnState {
272 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
273 26744 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
274 26744 : if let AnyMessage::Bytes(copy_data) = any {
275 26744 : let repl_prefix = b"START_REPLICATION ";
276 26744 : if !self.greeting && copy_data.starts_with(repl_prefix) {
277 218 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
278 218 : bail!("finished processing START_REPLICATION")
279 26526 : }
280 :
281 26526 : let msg = ProposerAcceptorMessage::parse(copy_data)?;
282 26526 : debug!("got msg: {:?}", msg);
283 26526 : self.process(msg, global)
284 : } else {
285 0 : bail!("unexpected message, expected AnyMessage::Bytes");
286 : }
287 26744 : }
288 :
289 : /// Process START_REPLICATION request.
290 218 : fn process_start_replication(
291 218 : &mut self,
292 218 : copy_data: Bytes,
293 218 : global: &mut GlobalMap,
294 218 : ) -> Result<()> {
295 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
296 218 : let str = String::from_utf8(copy_data.to_vec())?;
297 :
298 218 : let mut parts = str.split(' ');
299 218 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
300 218 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
301 218 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
302 218 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
303 :
304 218 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
305 218 : let shared_state = global.get(&ttid);
306 218 :
307 218 : // read bytes from start_lsn to end_lsn
308 218 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
309 218 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
310 218 :
311 218 : // send bytes to the client
312 218 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
313 218 : Ok(())
314 218 : }
315 :
316 : /// Get or create a timeline.
317 19878 : fn init_timeline(
318 19878 : &mut self,
319 19878 : ttid: TenantTimelineId,
320 19878 : server_info: ServerInfo,
321 19878 : global: &mut GlobalMap,
322 19878 : ) -> Result<()> {
323 19878 : self.ttid = ttid;
324 19878 : if global.has_tli(&ttid) {
325 18433 : return Ok(());
326 1445 : }
327 1445 :
328 1445 : global.create(ttid, server_info)
329 19878 : }
330 :
331 : /// Process a ProposerAcceptorMessage.
332 26526 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
333 26526 : if !self.greeting {
334 19878 : self.greeting = true;
335 19878 :
336 19878 : match msg {
337 19878 : ProposerAcceptorMessage::Greeting(ref greeting) => {
338 19878 : tracing::info!(
339 0 : "start handshake with walproposer {:?} {:?}",
340 : self.tcp,
341 : greeting
342 : );
343 19878 : let server_info = ServerInfo {
344 19878 : pg_version: greeting.pg_version,
345 19878 : system_id: greeting.system_id,
346 19878 : wal_seg_size: greeting.wal_seg_size,
347 19878 : };
348 19878 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
349 19878 : self.init_timeline(ttid, server_info, global)?
350 : }
351 : _ => {
352 0 : bail!("unexpected message {msg:?} instead of greeting");
353 : }
354 : }
355 6648 : }
356 :
357 26526 : let tli = global.get(&self.ttid);
358 26526 :
359 26526 : match msg {
360 3052 : ProposerAcceptorMessage::AppendRequest(append_request) => {
361 3052 : self.flush_pending = true;
362 3052 : self.process_sk_msg(
363 3052 : tli,
364 3052 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
365 3052 : )?;
366 : }
367 23474 : other => {
368 23474 : self.process_sk_msg(tli, &other)?;
369 : }
370 : }
371 :
372 26526 : Ok(())
373 26526 : }
374 :
375 : /// Process FlushWAL if needed.
376 174467 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
377 174467 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
378 174467 : if !self.flush_pending {
379 171965 : return Ok(());
380 2502 : }
381 2502 : self.flush_pending = false;
382 2502 : let shared_state = global.get(&self.ttid);
383 2502 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
384 174467 : }
385 :
386 : /// Make safekeeper process a message and send a reply to the TCP
387 29028 : fn process_sk_msg(
388 29028 : &mut self,
389 29028 : shared_state: &mut SharedState,
390 29028 : msg: &ProposerAcceptorMessage,
391 29028 : ) -> Result<()> {
392 29028 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
393 29028 : if let Some(reply) = &mut reply {
394 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
395 :
396 25221 : let mut buf = BytesMut::with_capacity(128);
397 25221 : reply.serialize(&mut buf)?;
398 :
399 25221 : self.tcp.send(AnyMessage::Bytes(buf.into()));
400 3807 : }
401 29028 : Ok(())
402 29028 : }
403 : }
404 :
405 : impl Drop for ConnState {
406 25209 : fn drop(&mut self) {
407 25209 : debug!("dropping conn: {:?}", self.tcp);
408 25209 : if !std::thread::panicking() {
409 218 : self.tcp.close();
410 24991 : }
411 : // TODO: clean up non-fsynced WAL
412 25209 : }
413 : }
|
