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::{
19 : ProposerAcceptorMessage, SafeKeeper, SK_PROTOCOL_VERSION, UNKNOWN_SERVER_VERSION,
20 : },
21 : state::{TimelinePersistentState, TimelineState},
22 : timeline::TimelineError,
23 : wal_storage::Storage,
24 : SafeKeeperConf,
25 : };
26 : use safekeeper_api::{membership::Configuration, ServerInfo};
27 : use tracing::{debug, info_span, warn};
28 : use utils::{
29 : id::{NodeId, TenantId, TenantTimelineId, TimelineId},
30 : lsn::Lsn,
31 : };
32 :
33 : use super::safekeeper_disk::{DiskStateStorage, DiskWALStorage, SafekeeperDisk, TimelineDisk};
34 :
35 : struct SharedState {
36 : sk: SafeKeeper<DiskStateStorage, DiskWALStorage>,
37 : disk: Arc<TimelineDisk>,
38 : }
39 :
40 : struct GlobalMap {
41 : timelines: HashMap<TenantTimelineId, SharedState>,
42 : conf: SafeKeeperConf,
43 : disk: Arc<SafekeeperDisk>,
44 : }
45 :
46 : impl GlobalMap {
47 : /// Restores global state from disk.
48 9376 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
49 9376 : let mut timelines = HashMap::new();
50 :
51 9376 : for (&ttid, disk) in disk.timelines.lock().iter() {
52 7000 : debug!("loading timeline {}", ttid);
53 7000 : let state = disk.state.lock().clone();
54 7000 :
55 7000 : if state.server.wal_seg_size == 0 {
56 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
57 7000 : }
58 7000 :
59 7000 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
60 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
61 7000 : }
62 7000 :
63 7000 : if state.commit_lsn < state.local_start_lsn {
64 0 : bail!(
65 0 : "commit_lsn {} is smaller than local_start_lsn {}",
66 0 : state.commit_lsn,
67 0 : state.local_start_lsn
68 0 : );
69 7000 : }
70 7000 :
71 7000 : let control_store = DiskStateStorage::new(disk.clone());
72 7000 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
73 :
74 7000 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
75 7000 : timelines.insert(
76 7000 : ttid,
77 7000 : SharedState {
78 7000 : sk,
79 7000 : disk: disk.clone(),
80 7000 : },
81 7000 : );
82 : }
83 :
84 9376 : Ok(Self {
85 9376 : timelines,
86 9376 : conf,
87 9376 : disk,
88 9376 : })
89 9376 : }
90 :
91 1432 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
92 1432 : if self.timelines.contains_key(&ttid) {
93 0 : bail!("timeline {} already exists", ttid);
94 1432 : }
95 1432 :
96 1432 : debug!("creating new timeline {}", ttid);
97 :
98 1432 : let commit_lsn = Lsn::INVALID;
99 1432 : let local_start_lsn = Lsn::INVALID;
100 :
101 1432 : let state = TimelinePersistentState::new(
102 1432 : &ttid,
103 1432 : Configuration::empty(),
104 1432 : server_info,
105 1432 : commit_lsn,
106 1432 : local_start_lsn,
107 1432 : )?;
108 :
109 1432 : let disk_timeline = self.disk.put_state(&ttid, state);
110 1432 : let control_store = DiskStateStorage::new(disk_timeline.clone());
111 1432 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
112 :
113 1432 : let sk = SafeKeeper::new(
114 1432 : TimelineState::new(control_store),
115 1432 : wal_store,
116 1432 : self.conf.my_id,
117 1432 : )?;
118 :
119 1432 : self.timelines.insert(
120 1432 : ttid,
121 1432 : SharedState {
122 1432 : sk,
123 1432 : disk: disk_timeline,
124 1432 : },
125 1432 : );
126 1432 : Ok(())
127 1432 : }
128 :
129 29604 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
130 29604 : self.timelines.get_mut(ttid).expect("timeline must exist")
131 29604 : }
132 :
133 19090 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
134 19090 : self.timelines.contains_key(ttid)
135 19090 : }
136 : }
137 :
138 : /// State of a single connection to walproposer.
139 : struct ConnState {
140 : tcp: TCP,
141 :
142 : greeting: bool,
143 : ttid: TenantTimelineId,
144 : flush_pending: bool,
145 :
146 : runtime: tokio::runtime::Runtime,
147 : }
148 :
149 9376 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
150 9376 : let _enter = info_span!("safekeeper", id = os.id()).entered();
151 9376 : debug!("started server");
152 9376 : os.log_event("started;safekeeper".to_owned());
153 9376 : let conf = SafeKeeperConf {
154 9376 : workdir: Utf8PathBuf::from("."),
155 9376 : my_id: NodeId(os.id() as u64),
156 9376 : listen_pg_addr: String::new(),
157 9376 : listen_http_addr: String::new(),
158 9376 : no_sync: false,
159 9376 : broker_endpoint: "/".parse::<Uri>().unwrap(),
160 9376 : broker_keepalive_interval: Duration::from_secs(0),
161 9376 : heartbeat_timeout: Duration::from_secs(0),
162 9376 : remote_storage: None,
163 9376 : max_offloader_lag_bytes: 0,
164 9376 : wal_backup_enabled: false,
165 9376 : listen_pg_addr_tenant_only: None,
166 9376 : advertise_pg_addr: None,
167 9376 : availability_zone: None,
168 9376 : peer_recovery_enabled: false,
169 9376 : backup_parallel_jobs: 0,
170 9376 : pg_auth: None,
171 9376 : pg_tenant_only_auth: None,
172 9376 : http_auth: None,
173 9376 : sk_auth_token: None,
174 9376 : current_thread_runtime: false,
175 9376 : walsenders_keep_horizon: false,
176 9376 : partial_backup_timeout: Duration::from_secs(0),
177 9376 : disable_periodic_broker_push: false,
178 9376 : enable_offload: false,
179 9376 : delete_offloaded_wal: false,
180 9376 : control_file_save_interval: Duration::from_secs(1),
181 9376 : partial_backup_concurrency: 1,
182 9376 : eviction_min_resident: Duration::ZERO,
183 9376 : wal_reader_fanout: false,
184 9376 : max_delta_for_fanout: None,
185 9376 : };
186 :
187 9376 : let mut global = GlobalMap::new(disk, conf.clone())?;
188 9376 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
189 :
190 9376 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
191 7000 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
192 7000 : let commit_lsn = shared_state.sk.state.commit_lsn;
193 7000 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
194 7000 : }
195 :
196 9376 : let node_events = os.node_events();
197 9376 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
198 9376 : let mut epoll_idx: Vec<usize> = vec![];
199 :
200 : // TODO: batch events processing (multiple events per tick)
201 : loop {
202 73603 : epoll_vec.clear();
203 73603 : epoll_idx.clear();
204 73603 :
205 73603 : // node events channel
206 73603 : epoll_vec.push(Box::new(node_events.clone()));
207 73603 : epoll_idx.push(0);
208 :
209 : // tcp connections
210 270268 : for conn in conns.values() {
211 270268 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
212 270268 : epoll_idx.push(conn.tcp.connection_id());
213 270268 : }
214 :
215 : // waiting for the next message
216 73603 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
217 73603 :
218 73603 : if index == 0 {
219 : // got a new connection
220 24410 : match node_events.must_recv() {
221 24410 : NodeEvent::Accept(tcp) => {
222 24410 : conns.insert(
223 24410 : tcp.connection_id(),
224 24410 : ConnState {
225 24410 : tcp,
226 24410 : greeting: false,
227 24410 : ttid: TenantTimelineId::empty(),
228 24410 : flush_pending: false,
229 24410 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
230 : },
231 : );
232 : }
233 0 : NodeEvent::Internal(_) => unreachable!(),
234 : }
235 24410 : continue;
236 49193 : }
237 49193 :
238 49193 : let connection_id = epoll_idx[index];
239 49193 : let conn = conns.get_mut(&connection_id).unwrap();
240 49193 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
241 :
242 : loop {
243 89791 : let event = match next_event {
244 50202 : Some(event) => event,
245 39589 : None => break,
246 : };
247 :
248 50202 : match event {
249 36184 : NetEvent::Message(msg) => {
250 36184 : let res = conn.process_any(msg, &mut global);
251 36184 : if res.is_err() {
252 9604 : let e = res.unwrap_err();
253 9604 : let estr = e.to_string();
254 9604 : if !estr.contains("finished processing START_REPLICATION") {
255 9376 : warn!("conn {:?} error: {:?}", connection_id, e);
256 0 : panic!("unexpected error at safekeeper: {:#}", e);
257 228 : }
258 228 : conns.remove(&connection_id);
259 228 : break;
260 26580 : }
261 : }
262 14018 : NetEvent::Closed => {
263 14018 : // TODO: remove from conns?
264 14018 : }
265 : }
266 :
267 40598 : next_event = conn.tcp.recv_chan().try_recv();
268 : }
269 :
270 177754 : conns.retain(|_, conn| {
271 177754 : let res = conn.flush(&mut global);
272 177754 : if res.is_err() {
273 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
274 177754 : }
275 177754 : res.is_ok()
276 177754 : });
277 : }
278 0 : }
279 :
280 : impl ConnState {
281 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
282 26808 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
283 26808 : if let AnyMessage::Bytes(copy_data) = any {
284 26808 : let repl_prefix = b"START_REPLICATION ";
285 26808 : if !self.greeting && copy_data.starts_with(repl_prefix) {
286 228 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
287 228 : bail!("finished processing START_REPLICATION")
288 26580 : }
289 :
290 26580 : let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTOCOL_VERSION)?;
291 26580 : debug!("got msg: {:?}", msg);
292 26580 : self.process(msg, global)
293 : } else {
294 0 : bail!("unexpected message, expected AnyMessage::Bytes");
295 : }
296 26808 : }
297 :
298 : /// Process START_REPLICATION request.
299 228 : fn process_start_replication(
300 228 : &mut self,
301 228 : copy_data: Bytes,
302 228 : global: &mut GlobalMap,
303 228 : ) -> Result<()> {
304 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
305 228 : let str = String::from_utf8(copy_data.to_vec())?;
306 :
307 228 : let mut parts = str.split(' ');
308 228 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
309 228 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
310 228 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
311 228 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
312 :
313 228 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
314 228 : let shared_state = global.get(&ttid);
315 228 :
316 228 : // read bytes from start_lsn to end_lsn
317 228 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
318 228 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
319 228 :
320 228 : // send bytes to the client
321 228 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
322 228 : Ok(())
323 228 : }
324 :
325 : /// Get or create a timeline.
326 19090 : fn init_timeline(
327 19090 : &mut self,
328 19090 : ttid: TenantTimelineId,
329 19090 : server_info: ServerInfo,
330 19090 : global: &mut GlobalMap,
331 19090 : ) -> Result<()> {
332 19090 : self.ttid = ttid;
333 19090 : if global.has_tli(&ttid) {
334 17658 : return Ok(());
335 1432 : }
336 1432 :
337 1432 : global.create(ttid, server_info)
338 19090 : }
339 :
340 : /// Process a ProposerAcceptorMessage.
341 26580 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
342 26580 : if !self.greeting {
343 19090 : self.greeting = true;
344 19090 :
345 19090 : match msg {
346 19090 : ProposerAcceptorMessage::Greeting(ref greeting) => {
347 19090 : tracing::info!(
348 0 : "start handshake with walproposer {:?} {:?}",
349 : self.tcp,
350 : greeting
351 : );
352 19090 : let server_info = ServerInfo {
353 19090 : pg_version: greeting.pg_version,
354 19090 : system_id: greeting.system_id,
355 19090 : wal_seg_size: greeting.wal_seg_size,
356 19090 : };
357 19090 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
358 19090 : self.init_timeline(ttid, server_info, global)?
359 : }
360 : _ => {
361 0 : bail!("unexpected message {msg:?} instead of greeting");
362 : }
363 : }
364 7490 : }
365 :
366 26580 : let tli = global.get(&self.ttid);
367 26580 :
368 26580 : match msg {
369 3352 : ProposerAcceptorMessage::AppendRequest(append_request) => {
370 3352 : self.flush_pending = true;
371 3352 : self.process_sk_msg(
372 3352 : tli,
373 3352 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
374 3352 : )?;
375 : }
376 23228 : other => {
377 23228 : self.process_sk_msg(tli, &other)?;
378 : }
379 : }
380 :
381 26580 : Ok(())
382 26580 : }
383 :
384 : /// Process FlushWAL if needed.
385 177754 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
386 177754 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
387 177754 : if !self.flush_pending {
388 174958 : return Ok(());
389 2796 : }
390 2796 : self.flush_pending = false;
391 2796 : let shared_state = global.get(&self.ttid);
392 2796 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
393 177754 : }
394 :
395 : /// Make safekeeper process a message and send a reply to the TCP
396 29376 : fn process_sk_msg(
397 29376 : &mut self,
398 29376 : shared_state: &mut SharedState,
399 29376 : msg: &ProposerAcceptorMessage,
400 29376 : ) -> Result<()> {
401 29376 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
402 29376 : if let Some(reply) = &mut reply {
403 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
404 :
405 25192 : let mut buf = BytesMut::with_capacity(128);
406 25192 : reply.serialize(&mut buf)?;
407 :
408 25192 : self.tcp.send(AnyMessage::Bytes(buf.into()));
409 4184 : }
410 29376 : Ok(())
411 29376 : }
412 : }
413 :
414 : impl Drop for ConnState {
415 24382 : fn drop(&mut self) {
416 24382 : debug!("dropping conn: {:?}", self.tcp);
417 24382 : if !std::thread::panicking() {
418 228 : self.tcp.close();
419 24154 : }
420 : // TODO: clean up non-fsynced WAL
421 24382 : }
422 : }
|
