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 safekeeper::SafeKeeperConf;
18 : use safekeeper::safekeeper::{
19 : ProposerAcceptorMessage, SK_PROTO_VERSION_3, SafeKeeper, UNKNOWN_SERVER_VERSION,
20 : };
21 : use safekeeper::state::{TimelinePersistentState, TimelineState};
22 : use safekeeper::timeline::TimelineError;
23 : use safekeeper::wal_storage::Storage;
24 : use safekeeper_api::ServerInfo;
25 : use safekeeper_api::membership::Configuration;
26 : use tracing::{debug, info_span, warn};
27 : use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
28 : use utils::lsn::Lsn;
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 9672 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 9672 : let mut timelines = HashMap::new();
47 :
48 9672 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 7293 : debug!("loading timeline {}", ttid);
50 7293 : let state = disk.state.lock().clone();
51 :
52 7293 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 7293 : }
55 :
56 7293 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 7293 : }
59 :
60 7293 : if state.commit_lsn < state.local_start_lsn {
61 0 : bail!(
62 0 : "commit_lsn {} is smaller than local_start_lsn {}",
63 : state.commit_lsn,
64 : state.local_start_lsn
65 : );
66 7293 : }
67 :
68 7293 : let control_store = DiskStateStorage::new(disk.clone());
69 7293 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 7293 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
72 7293 : timelines.insert(
73 7293 : ttid,
74 7293 : SharedState {
75 7293 : sk,
76 7293 : disk: disk.clone(),
77 7293 : },
78 : );
79 : }
80 :
81 9672 : Ok(Self {
82 9672 : timelines,
83 9672 : conf,
84 9672 : disk,
85 9672 : })
86 9672 : }
87 :
88 1443 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 1443 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 1443 : }
92 :
93 1443 : debug!("creating new timeline {}", ttid);
94 :
95 1443 : let commit_lsn = Lsn::INVALID;
96 1443 : let local_start_lsn = Lsn::INVALID;
97 :
98 1443 : let state = TimelinePersistentState::new(
99 1443 : &ttid,
100 1443 : Configuration::empty(),
101 1443 : server_info,
102 1443 : commit_lsn,
103 1443 : local_start_lsn,
104 0 : )?;
105 :
106 1443 : let disk_timeline = self.disk.put_state(&ttid, state);
107 1443 : let control_store = DiskStateStorage::new(disk_timeline.clone());
108 1443 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
109 :
110 1443 : let sk = SafeKeeper::new(
111 1443 : TimelineState::new(control_store),
112 1443 : wal_store,
113 1443 : self.conf.my_id,
114 0 : )?;
115 :
116 1443 : self.timelines.insert(
117 1443 : ttid,
118 1443 : SharedState {
119 1443 : sk,
120 1443 : disk: disk_timeline,
121 1443 : },
122 : );
123 1443 : Ok(())
124 1443 : }
125 :
126 31498 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
127 31498 : self.timelines.get_mut(ttid).expect("timeline must exist")
128 31498 : }
129 :
130 19415 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
131 19415 : self.timelines.contains_key(ttid)
132 19415 : }
133 : }
134 :
135 : /// State of a single connection to walproposer.
136 : struct ConnState {
137 : tcp: TCP,
138 :
139 : greeting: bool,
140 : ttid: TenantTimelineId,
141 : flush_pending: bool,
142 :
143 : runtime: tokio::runtime::Runtime,
144 : }
145 :
146 9672 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
147 9672 : let _enter = info_span!("safekeeper", id = os.id()).entered();
148 9672 : debug!("started server");
149 9672 : os.log_event("started;safekeeper".to_owned());
150 9672 : let conf = SafeKeeperConf {
151 9672 : workdir: Utf8PathBuf::from("."),
152 9672 : my_id: NodeId(os.id() as u64),
153 9672 : listen_pg_addr: String::new(),
154 9672 : listen_http_addr: String::new(),
155 9672 : listen_https_addr: None,
156 9672 : no_sync: false,
157 9672 : broker_endpoint: "/".parse::<Uri>().unwrap(),
158 9672 : broker_keepalive_interval: Duration::from_secs(0),
159 9672 : heartbeat_timeout: Duration::from_secs(0),
160 9672 : remote_storage: None,
161 9672 : max_offloader_lag_bytes: 0,
162 9672 : /* BEGIN_HADRON */
163 9672 : max_reelect_offloader_lag_bytes: 0,
164 9672 : max_timeline_disk_usage_bytes: 0,
165 9672 : /* END_HADRON */
166 9672 : wal_backup_enabled: false,
167 9672 : listen_pg_addr_tenant_only: None,
168 9672 : advertise_pg_addr: None,
169 9672 : availability_zone: None,
170 9672 : peer_recovery_enabled: false,
171 9672 : backup_parallel_jobs: 0,
172 9672 : pg_auth: None,
173 9672 : pg_tenant_only_auth: None,
174 9672 : http_auth: None,
175 9672 : sk_auth_token: None,
176 9672 : current_thread_runtime: false,
177 9672 : walsenders_keep_horizon: false,
178 9672 : partial_backup_timeout: Duration::from_secs(0),
179 9672 : disable_periodic_broker_push: false,
180 9672 : enable_offload: false,
181 9672 : delete_offloaded_wal: false,
182 9672 : control_file_save_interval: Duration::from_secs(1),
183 9672 : partial_backup_concurrency: 1,
184 9672 : eviction_min_resident: Duration::ZERO,
185 9672 : wal_reader_fanout: false,
186 9672 : max_delta_for_fanout: None,
187 9672 : ssl_key_file: Utf8PathBuf::from(""),
188 9672 : ssl_cert_file: Utf8PathBuf::from(""),
189 9672 : ssl_cert_reload_period: Duration::ZERO,
190 9672 : ssl_ca_certs: Vec::new(),
191 9672 : use_https_safekeeper_api: false,
192 9672 : enable_tls_wal_service_api: false,
193 9672 : force_metric_collection_on_scrape: true,
194 9672 : /* BEGIN_HADRON */
195 9672 : enable_pull_timeline_on_startup: false,
196 9672 : advertise_pg_addr_tenant_only: None,
197 9672 : hcc_base_url: None,
198 9672 : global_disk_check_interval: Duration::from_secs(10),
199 9672 : max_global_disk_usage_ratio: 0.0,
200 9672 : /* END_HADRON */
201 9672 : };
202 :
203 9672 : let mut global = GlobalMap::new(disk, conf.clone())?;
204 9672 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
205 :
206 9672 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
207 7293 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
208 7293 : let commit_lsn = shared_state.sk.state.commit_lsn;
209 7293 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
210 7293 : }
211 :
212 9672 : let node_events = os.node_events();
213 9672 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
214 9672 : let mut epoll_idx: Vec<usize> = vec![];
215 :
216 : // TODO: batch events processing (multiple events per tick)
217 : loop {
218 76046 : epoll_vec.clear();
219 76046 : epoll_idx.clear();
220 :
221 : // node events channel
222 76046 : epoll_vec.push(Box::new(node_events.clone()));
223 76046 : epoll_idx.push(0);
224 :
225 : // tcp connections
226 264647 : for conn in conns.values() {
227 264647 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
228 264647 : epoll_idx.push(conn.tcp.connection_id());
229 264647 : }
230 :
231 : // waiting for the next message
232 76046 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
233 :
234 76046 : if index == 0 {
235 : // got a new connection
236 24885 : match node_events.must_recv() {
237 24885 : NodeEvent::Accept(tcp) => {
238 24885 : conns.insert(
239 24885 : tcp.connection_id(),
240 : ConnState {
241 24885 : tcp,
242 : greeting: false,
243 24885 : ttid: TenantTimelineId::empty(),
244 : flush_pending: false,
245 24885 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
246 : },
247 : );
248 : }
249 0 : NodeEvent::Internal(_) => unreachable!(),
250 : }
251 24885 : continue;
252 51161 : }
253 :
254 51161 : let connection_id = epoll_idx[index];
255 51161 : let conn = conns.get_mut(&connection_id).unwrap();
256 51161 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
257 :
258 : loop {
259 83879 : let event = match next_event {
260 42701 : Some(event) => event,
261 41178 : None => break,
262 : };
263 :
264 42701 : match event {
265 28225 : NetEvent::Message(msg) => {
266 28225 : let res = conn.process_any(msg, &mut global);
267 28225 : if res.is_err() {
268 9983 : let e = res.unwrap_err();
269 9983 : let estr = e.to_string();
270 9983 : if !estr.contains("finished processing START_REPLICATION") {
271 9672 : warn!("conn {:?} error: {:?}", connection_id, e);
272 9672 : panic!("unexpected error at safekeeper: {e:#}");
273 311 : }
274 311 : conns.remove(&connection_id);
275 311 : break;
276 18242 : }
277 : }
278 14476 : NetEvent::Closed => {
279 14476 : // TODO: remove from conns?
280 14476 : }
281 : }
282 :
283 32718 : next_event = conn.tcp.recv_chan().try_recv();
284 : }
285 :
286 174003 : conns.retain(|_, conn| {
287 174003 : let res = conn.flush(&mut global);
288 174003 : if res.is_err() {
289 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
290 174003 : }
291 174003 : res.is_ok()
292 174003 : });
293 : }
294 0 : }
295 :
296 : impl ConnState {
297 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
298 28225 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
299 28225 : if let AnyMessage::Bytes(copy_data) = any {
300 28225 : let repl_prefix = b"START_REPLICATION ";
301 28225 : if !self.greeting && copy_data.starts_with(repl_prefix) {
302 311 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
303 311 : bail!("finished processing START_REPLICATION")
304 27914 : }
305 :
306 27914 : let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTO_VERSION_3)?;
307 27914 : debug!("got msg: {:?}", msg);
308 27914 : self.process(msg, global)
309 : } else {
310 0 : bail!("unexpected message, expected AnyMessage::Bytes");
311 : }
312 28225 : }
313 :
314 : /// Process START_REPLICATION request.
315 311 : fn process_start_replication(
316 311 : &mut self,
317 311 : copy_data: Bytes,
318 311 : global: &mut GlobalMap,
319 311 : ) -> Result<()> {
320 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
321 311 : let str = String::from_utf8(copy_data.to_vec())?;
322 :
323 311 : let mut parts = str.split(' ');
324 311 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
325 311 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
326 311 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
327 311 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
328 :
329 311 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
330 311 : let shared_state = global.get(&ttid);
331 :
332 : // read bytes from start_lsn to end_lsn
333 311 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
334 311 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
335 :
336 : // send bytes to the client
337 311 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
338 311 : Ok(())
339 311 : }
340 :
341 : /// Get or create a timeline.
342 19415 : fn init_timeline(
343 19415 : &mut self,
344 19415 : ttid: TenantTimelineId,
345 19415 : server_info: ServerInfo,
346 19415 : global: &mut GlobalMap,
347 19415 : ) -> Result<()> {
348 19415 : self.ttid = ttid;
349 19415 : if global.has_tli(&ttid) {
350 17972 : return Ok(());
351 1443 : }
352 :
353 1443 : global.create(ttid, server_info)
354 19415 : }
355 :
356 : /// Process a ProposerAcceptorMessage.
357 27914 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
358 27914 : if !self.greeting {
359 19415 : self.greeting = true;
360 :
361 19415 : match msg {
362 19415 : ProposerAcceptorMessage::Greeting(ref greeting) => {
363 19415 : tracing::info!(
364 0 : "start handshake with walproposer {:?} {:?}",
365 : self.tcp,
366 : greeting
367 : );
368 19415 : let server_info = ServerInfo {
369 19415 : pg_version: greeting.pg_version,
370 19415 : system_id: greeting.system_id,
371 19415 : wal_seg_size: greeting.wal_seg_size,
372 19415 : };
373 19415 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
374 19415 : self.init_timeline(ttid, server_info, global)?
375 : }
376 : _ => {
377 0 : bail!("unexpected message {msg:?} instead of greeting");
378 : }
379 : }
380 8499 : }
381 :
382 27914 : let tli = global.get(&self.ttid);
383 :
384 27914 : match msg {
385 3961 : ProposerAcceptorMessage::AppendRequest(append_request) => {
386 3961 : self.flush_pending = true;
387 3961 : self.process_sk_msg(
388 3961 : tli,
389 3961 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
390 0 : )?;
391 : }
392 23953 : other => {
393 23953 : self.process_sk_msg(tli, &other)?;
394 : }
395 : }
396 :
397 27914 : Ok(())
398 27914 : }
399 :
400 : /// Process FlushWAL if needed.
401 174003 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
402 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
403 174003 : if !self.flush_pending {
404 170730 : return Ok(());
405 3273 : }
406 3273 : self.flush_pending = false;
407 3273 : let shared_state = global.get(&self.ttid);
408 3273 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
409 174003 : }
410 :
411 : /// Make safekeeper process a message and send a reply to the TCP
412 31187 : fn process_sk_msg(
413 31187 : &mut self,
414 31187 : shared_state: &mut SharedState,
415 31187 : msg: &ProposerAcceptorMessage,
416 31187 : ) -> Result<()> {
417 31187 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
418 31187 : if let Some(reply) = &mut reply {
419 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
420 :
421 26275 : let mut buf = BytesMut::with_capacity(128);
422 26275 : reply.serialize(&mut buf, SK_PROTO_VERSION_3)?;
423 :
424 26275 : self.tcp.send(AnyMessage::Bytes(buf.into()));
425 4912 : }
426 31187 : Ok(())
427 31187 : }
428 : }
429 :
430 : impl Drop for ConnState {
431 24859 : fn drop(&mut self) {
432 24859 : debug!("dropping conn: {:?}", self.tcp);
433 24859 : if !std::thread::panicking() {
434 311 : self.tcp.close();
435 24548 : }
436 : // TODO: clean up non-fsynced WAL
437 24859 : }
438 : }
|
