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;
6 : use std::sync::Arc;
7 : use std::time::Duration;
8 :
9 : use anyhow::{Result, bail};
10 : use bytes::{Bytes, BytesMut};
11 : use camino::Utf8PathBuf;
12 : use desim::executor::{self, PollSome};
13 : use desim::network::TCP;
14 : use desim::node_os::NodeOs;
15 : use desim::proto::{AnyMessage, NetEvent, NodeEvent};
16 : use http::Uri;
17 : use postgres_backend::AuthType;
18 : use safekeeper::SafeKeeperConf;
19 : use safekeeper::safekeeper::{
20 : ProposerAcceptorMessage, SK_PROTO_VERSION_3, SafeKeeper, UNKNOWN_SERVER_VERSION,
21 : };
22 : use safekeeper::state::{TimelinePersistentState, TimelineState};
23 : use safekeeper::timeline::TimelineError;
24 : use safekeeper::wal_storage::Storage;
25 : use safekeeper_api::ServerInfo;
26 : use safekeeper_api::membership::Configuration;
27 : use tracing::{debug, info_span, warn};
28 : use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
29 : use utils::lsn::Lsn;
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 9623 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
47 9623 : let mut timelines = HashMap::new();
48 :
49 9623 : for (&ttid, disk) in disk.timelines.lock().iter() {
50 7223 : debug!("loading timeline {}", ttid);
51 7223 : let state = disk.state.lock().clone();
52 :
53 7223 : if state.server.wal_seg_size == 0 {
54 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
55 7223 : }
56 :
57 7223 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
58 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
59 7223 : }
60 :
61 7223 : if state.commit_lsn < state.local_start_lsn {
62 0 : bail!(
63 0 : "commit_lsn {} is smaller than local_start_lsn {}",
64 : state.commit_lsn,
65 : state.local_start_lsn
66 : );
67 7223 : }
68 :
69 7223 : let control_store = DiskStateStorage::new(disk.clone());
70 7223 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
71 :
72 7223 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
73 7223 : timelines.insert(
74 7223 : ttid,
75 7223 : SharedState {
76 7223 : sk,
77 7223 : disk: disk.clone(),
78 7223 : },
79 : );
80 : }
81 :
82 9623 : Ok(Self {
83 9623 : timelines,
84 9623 : conf,
85 9623 : disk,
86 9623 : })
87 9623 : }
88 :
89 1434 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
90 1434 : if self.timelines.contains_key(&ttid) {
91 0 : bail!("timeline {} already exists", ttid);
92 1434 : }
93 :
94 1434 : debug!("creating new timeline {}", ttid);
95 :
96 1434 : let commit_lsn = Lsn::INVALID;
97 1434 : let local_start_lsn = Lsn::INVALID;
98 :
99 1434 : let state = TimelinePersistentState::new(
100 1434 : &ttid,
101 1434 : Configuration::empty(),
102 1434 : server_info,
103 1434 : commit_lsn,
104 1434 : local_start_lsn,
105 0 : )?;
106 :
107 1434 : let disk_timeline = self.disk.put_state(&ttid, state);
108 1434 : let control_store = DiskStateStorage::new(disk_timeline.clone());
109 1434 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
110 :
111 1434 : let sk = SafeKeeper::new(
112 1434 : TimelineState::new(control_store),
113 1434 : wal_store,
114 1434 : self.conf.my_id,
115 0 : )?;
116 :
117 1434 : self.timelines.insert(
118 1434 : ttid,
119 1434 : SharedState {
120 1434 : sk,
121 1434 : disk: disk_timeline,
122 1434 : },
123 : );
124 1434 : Ok(())
125 1434 : }
126 :
127 33283 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
128 33283 : self.timelines.get_mut(ttid).expect("timeline must exist")
129 33283 : }
130 :
131 20450 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
132 20450 : self.timelines.contains_key(ttid)
133 20450 : }
134 : }
135 :
136 : /// State of a single connection to walproposer.
137 : struct ConnState {
138 : tcp: TCP,
139 :
140 : greeting: bool,
141 : ttid: TenantTimelineId,
142 : flush_pending: bool,
143 :
144 : runtime: tokio::runtime::Runtime,
145 : }
146 :
147 9623 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
148 9623 : let _enter = info_span!("safekeeper", id = os.id()).entered();
149 9623 : debug!("started server");
150 9623 : os.log_event("started;safekeeper".to_owned());
151 9623 : let conf = SafeKeeperConf {
152 9623 : workdir: Utf8PathBuf::from("."),
153 9623 : my_id: NodeId(os.id() as u64),
154 9623 : listen_pg_addr: String::new(),
155 9623 : listen_http_addr: String::new(),
156 9623 : listen_https_addr: None,
157 9623 : no_sync: false,
158 9623 : broker_endpoint: "/".parse::<Uri>().unwrap(),
159 9623 : broker_keepalive_interval: Duration::from_secs(0),
160 9623 : heartbeat_timeout: Duration::from_secs(0),
161 9623 : remote_storage: None,
162 9623 : max_offloader_lag_bytes: 0,
163 9623 : /* BEGIN_HADRON */
164 9623 : max_reelect_offloader_lag_bytes: 0,
165 9623 : max_timeline_disk_usage_bytes: 0,
166 9623 : /* END_HADRON */
167 9623 : wal_backup_enabled: false,
168 9623 : listen_pg_addr_tenant_only: None,
169 9623 : advertise_pg_addr: None,
170 9623 : availability_zone: None,
171 9623 : peer_recovery_enabled: false,
172 9623 : backup_parallel_jobs: 0,
173 9623 : auth_type: AuthType::NeonJWT,
174 9623 : pg_auth: None,
175 9623 : pg_tenant_only_auth: None,
176 9623 : http_auth: None,
177 9623 : sk_auth_token: None,
178 9623 : current_thread_runtime: false,
179 9623 : walsenders_keep_horizon: false,
180 9623 : partial_backup_timeout: Duration::from_secs(0),
181 9623 : disable_periodic_broker_push: false,
182 9623 : enable_offload: false,
183 9623 : delete_offloaded_wal: false,
184 9623 : control_file_save_interval: Duration::from_secs(1),
185 9623 : partial_backup_concurrency: 1,
186 9623 : eviction_min_resident: Duration::ZERO,
187 9623 : wal_reader_fanout: false,
188 9623 : max_delta_for_fanout: None,
189 9623 : ssl_key_file: Utf8PathBuf::from(""),
190 9623 : ssl_cert_file: Utf8PathBuf::from(""),
191 9623 : ssl_cert_reload_period: Duration::ZERO,
192 9623 : ssl_ca_certs: Vec::new(),
193 9623 : use_https_safekeeper_api: false,
194 9623 : enable_tls_wal_service_api: false,
195 9623 : force_metric_collection_on_scrape: true,
196 9623 : /* BEGIN_HADRON */
197 9623 : enable_pull_timeline_on_startup: false,
198 9623 : advertise_pg_addr_tenant_only: None,
199 9623 : hcc_base_url: None,
200 9623 : global_disk_check_interval: Duration::from_secs(10),
201 9623 : max_global_disk_usage_ratio: 0.0,
202 9623 : /* END_HADRON */
203 9623 : };
204 :
205 9623 : let mut global = GlobalMap::new(disk, conf.clone())?;
206 9623 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
207 :
208 9623 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
209 7223 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
210 7223 : let commit_lsn = shared_state.sk.state.commit_lsn;
211 7223 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
212 7223 : }
213 :
214 9623 : let node_events = os.node_events();
215 9623 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
216 9623 : let mut epoll_idx: Vec<usize> = vec![];
217 :
218 : // TODO: batch events processing (multiple events per tick)
219 : loop {
220 79015 : epoll_vec.clear();
221 79015 : epoll_idx.clear();
222 :
223 : // node events channel
224 79015 : epoll_vec.push(Box::new(node_events.clone()));
225 79015 : epoll_idx.push(0);
226 :
227 : // tcp connections
228 304461 : for conn in conns.values() {
229 304461 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
230 304461 : epoll_idx.push(conn.tcp.connection_id());
231 304461 : }
232 :
233 : // waiting for the next message
234 79015 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
235 :
236 79015 : if index == 0 {
237 : // got a new connection
238 25803 : match node_events.must_recv() {
239 25803 : NodeEvent::Accept(tcp) => {
240 25803 : conns.insert(
241 25803 : tcp.connection_id(),
242 : ConnState {
243 25803 : tcp,
244 : greeting: false,
245 25803 : ttid: TenantTimelineId::empty(),
246 : flush_pending: false,
247 25803 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
248 : },
249 : );
250 : }
251 0 : NodeEvent::Internal(_) => unreachable!(),
252 : }
253 25803 : continue;
254 53212 : }
255 :
256 53212 : let connection_id = epoll_idx[index];
257 53212 : let conn = conns.get_mut(&connection_id).unwrap();
258 53212 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
259 :
260 : loop {
261 88244 : let event = match next_event {
262 44925 : Some(event) => event,
263 43319 : None => break,
264 : };
265 :
266 44925 : match event {
267 29522 : NetEvent::Message(msg) => {
268 29522 : let res = conn.process_any(msg, &mut global);
269 29522 : if res.is_err() {
270 9893 : let e = res.unwrap_err();
271 9893 : let estr = e.to_string();
272 9893 : if !estr.contains("finished processing START_REPLICATION") {
273 9623 : warn!("conn {:?} error: {:?}", connection_id, e);
274 9623 : panic!("unexpected error at safekeeper: {e:#}");
275 270 : }
276 270 : conns.remove(&connection_id);
277 270 : break;
278 19629 : }
279 : }
280 15403 : NetEvent::Closed => {
281 15403 : // TODO: remove from conns?
282 15403 : }
283 : }
284 :
285 35032 : next_event = conn.tcp.recv_chan().try_recv();
286 : }
287 :
288 202137 : conns.retain(|_, conn| {
289 202137 : let res = conn.flush(&mut global);
290 202137 : if res.is_err() {
291 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
292 202137 : }
293 202137 : res.is_ok()
294 202137 : });
295 : }
296 0 : }
297 :
298 : impl ConnState {
299 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
300 29522 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
301 29522 : if let AnyMessage::Bytes(copy_data) = any {
302 29522 : let repl_prefix = b"START_REPLICATION ";
303 29522 : if !self.greeting && copy_data.starts_with(repl_prefix) {
304 270 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
305 270 : bail!("finished processing START_REPLICATION")
306 29252 : }
307 :
308 29252 : let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTO_VERSION_3)?;
309 29252 : debug!("got msg: {:?}", msg);
310 29252 : self.process(msg, global)
311 : } else {
312 0 : bail!("unexpected message, expected AnyMessage::Bytes");
313 : }
314 29522 : }
315 :
316 : /// Process START_REPLICATION request.
317 270 : fn process_start_replication(
318 270 : &mut self,
319 270 : copy_data: Bytes,
320 270 : global: &mut GlobalMap,
321 270 : ) -> Result<()> {
322 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
323 270 : let str = String::from_utf8(copy_data.to_vec())?;
324 :
325 270 : let mut parts = str.split(' ');
326 270 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
327 270 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
328 270 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
329 270 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
330 :
331 270 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
332 270 : let shared_state = global.get(&ttid);
333 :
334 : // read bytes from start_lsn to end_lsn
335 270 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
336 270 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
337 :
338 : // send bytes to the client
339 270 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
340 270 : Ok(())
341 270 : }
342 :
343 : /// Get or create a timeline.
344 20450 : fn init_timeline(
345 20450 : &mut self,
346 20450 : ttid: TenantTimelineId,
347 20450 : server_info: ServerInfo,
348 20450 : global: &mut GlobalMap,
349 20450 : ) -> Result<()> {
350 20450 : self.ttid = ttid;
351 20450 : if global.has_tli(&ttid) {
352 19016 : return Ok(());
353 1434 : }
354 :
355 1434 : global.create(ttid, server_info)
356 20450 : }
357 :
358 : /// Process a ProposerAcceptorMessage.
359 29252 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
360 29252 : if !self.greeting {
361 20450 : self.greeting = true;
362 :
363 20450 : match msg {
364 20450 : ProposerAcceptorMessage::Greeting(ref greeting) => {
365 20450 : tracing::info!(
366 0 : "start handshake with walproposer {:?} {:?}",
367 : self.tcp,
368 : greeting
369 : );
370 20450 : let server_info = ServerInfo {
371 20450 : pg_version: greeting.pg_version,
372 20450 : system_id: greeting.system_id,
373 20450 : wal_seg_size: greeting.wal_seg_size,
374 20450 : };
375 20450 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
376 20450 : self.init_timeline(ttid, server_info, global)?
377 : }
378 : _ => {
379 0 : bail!("unexpected message {msg:?} instead of greeting");
380 : }
381 : }
382 8802 : }
383 :
384 29252 : let tli = global.get(&self.ttid);
385 :
386 29252 : match msg {
387 4493 : ProposerAcceptorMessage::AppendRequest(append_request) => {
388 4493 : self.flush_pending = true;
389 4493 : self.process_sk_msg(
390 4493 : tli,
391 4493 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
392 0 : )?;
393 : }
394 24759 : other => {
395 24759 : self.process_sk_msg(tli, &other)?;
396 : }
397 : }
398 :
399 29252 : Ok(())
400 29252 : }
401 :
402 : /// Process FlushWAL if needed.
403 202137 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
404 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
405 202137 : if !self.flush_pending {
406 198376 : return Ok(());
407 3761 : }
408 3761 : self.flush_pending = false;
409 3761 : let shared_state = global.get(&self.ttid);
410 3761 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
411 202137 : }
412 :
413 : /// Make safekeeper process a message and send a reply to the TCP
414 33013 : fn process_sk_msg(
415 33013 : &mut self,
416 33013 : shared_state: &mut SharedState,
417 33013 : msg: &ProposerAcceptorMessage,
418 33013 : ) -> Result<()> {
419 33013 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
420 33013 : if let Some(reply) = &mut reply {
421 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
422 :
423 27484 : let mut buf = BytesMut::with_capacity(128);
424 27484 : reply.serialize(&mut buf, SK_PROTO_VERSION_3)?;
425 :
426 27484 : self.tcp.send(AnyMessage::Bytes(buf.into()));
427 5529 : }
428 33013 : Ok(())
429 33013 : }
430 : }
431 :
432 : impl Drop for ConnState {
433 25777 : fn drop(&mut self) {
434 25777 : debug!("dropping conn: {:?}", self.tcp);
435 25777 : if !std::thread::panicking() {
436 270 : self.tcp.close();
437 25507 : }
438 : // TODO: clean up non-fsynced WAL
439 25777 : }
440 : }
|
