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 9264 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 9264 : let mut timelines = HashMap::new();
47 :
48 9264 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 6871 : debug!("loading timeline {}", ttid);
50 6871 : let state = disk.state.lock().clone();
51 :
52 6871 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 6871 : }
55 :
56 6871 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 6871 : }
59 :
60 6871 : 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 6871 : }
67 :
68 6871 : let control_store = DiskStateStorage::new(disk.clone());
69 6871 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 6871 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
72 6871 : timelines.insert(
73 6871 : ttid,
74 6871 : SharedState {
75 6871 : sk,
76 6871 : disk: disk.clone(),
77 6871 : },
78 : );
79 : }
80 :
81 9264 : Ok(Self {
82 9264 : timelines,
83 9264 : conf,
84 9264 : disk,
85 9264 : })
86 9264 : }
87 :
88 1431 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 1431 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 1431 : }
92 :
93 1431 : debug!("creating new timeline {}", ttid);
94 :
95 1431 : let commit_lsn = Lsn::INVALID;
96 1431 : let local_start_lsn = Lsn::INVALID;
97 :
98 1431 : let state = TimelinePersistentState::new(
99 1431 : &ttid,
100 1431 : Configuration::empty(),
101 1431 : server_info,
102 1431 : commit_lsn,
103 1431 : local_start_lsn,
104 0 : )?;
105 :
106 1431 : let disk_timeline = self.disk.put_state(&ttid, state);
107 1431 : let control_store = DiskStateStorage::new(disk_timeline.clone());
108 1431 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
109 :
110 1431 : let sk = SafeKeeper::new(
111 1431 : TimelineState::new(control_store),
112 1431 : wal_store,
113 1431 : self.conf.my_id,
114 0 : )?;
115 :
116 1431 : self.timelines.insert(
117 1431 : ttid,
118 1431 : SharedState {
119 1431 : sk,
120 1431 : disk: disk_timeline,
121 1431 : },
122 : );
123 1431 : Ok(())
124 1431 : }
125 :
126 31109 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
127 31109 : self.timelines.get_mut(ttid).expect("timeline must exist")
128 31109 : }
129 :
130 18830 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
131 18830 : self.timelines.contains_key(ttid)
132 18830 : }
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 9264 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
147 9264 : let _enter = info_span!("safekeeper", id = os.id()).entered();
148 9264 : debug!("started server");
149 9264 : os.log_event("started;safekeeper".to_owned());
150 9264 : let conf = SafeKeeperConf {
151 9264 : workdir: Utf8PathBuf::from("."),
152 9264 : my_id: NodeId(os.id() as u64),
153 9264 : listen_pg_addr: String::new(),
154 9264 : listen_http_addr: String::new(),
155 9264 : listen_https_addr: None,
156 9264 : no_sync: false,
157 9264 : broker_endpoint: "/".parse::<Uri>().unwrap(),
158 9264 : broker_keepalive_interval: Duration::from_secs(0),
159 9264 : heartbeat_timeout: Duration::from_secs(0),
160 9264 : remote_storage: None,
161 9264 : max_offloader_lag_bytes: 0,
162 9264 : /* BEGIN_HADRON */
163 9264 : max_reelect_offloader_lag_bytes: 0,
164 9264 : max_timeline_disk_usage_bytes: 0,
165 9264 : /* END_HADRON */
166 9264 : wal_backup_enabled: false,
167 9264 : listen_pg_addr_tenant_only: None,
168 9264 : advertise_pg_addr: None,
169 9264 : availability_zone: None,
170 9264 : peer_recovery_enabled: false,
171 9264 : backup_parallel_jobs: 0,
172 9264 : pg_auth: None,
173 9264 : pg_tenant_only_auth: None,
174 9264 : http_auth: None,
175 9264 : sk_auth_token: None,
176 9264 : current_thread_runtime: false,
177 9264 : walsenders_keep_horizon: false,
178 9264 : partial_backup_timeout: Duration::from_secs(0),
179 9264 : disable_periodic_broker_push: false,
180 9264 : enable_offload: false,
181 9264 : delete_offloaded_wal: false,
182 9264 : control_file_save_interval: Duration::from_secs(1),
183 9264 : partial_backup_concurrency: 1,
184 9264 : eviction_min_resident: Duration::ZERO,
185 9264 : wal_reader_fanout: false,
186 9264 : max_delta_for_fanout: None,
187 9264 : ssl_key_file: Utf8PathBuf::from(""),
188 9264 : ssl_cert_file: Utf8PathBuf::from(""),
189 9264 : ssl_cert_reload_period: Duration::ZERO,
190 9264 : ssl_ca_certs: Vec::new(),
191 9264 : use_https_safekeeper_api: false,
192 9264 : enable_tls_wal_service_api: false,
193 9264 : force_metric_collection_on_scrape: true,
194 9264 : /* BEGIN_HADRON */
195 9264 : enable_pull_timeline_on_startup: false,
196 9264 : advertise_pg_addr_tenant_only: None,
197 9264 : hcc_base_url: None,
198 9264 : global_disk_check_interval: Duration::from_secs(10),
199 9264 : max_global_disk_usage_ratio: 0.0,
200 9264 : /* END_HADRON */
201 9264 : };
202 :
203 9264 : let mut global = GlobalMap::new(disk, conf.clone())?;
204 9264 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
205 :
206 9264 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
207 6871 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
208 6871 : let commit_lsn = shared_state.sk.state.commit_lsn;
209 6871 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
210 6871 : }
211 :
212 9264 : let node_events = os.node_events();
213 9264 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
214 9264 : let mut epoll_idx: Vec<usize> = vec![];
215 :
216 : // TODO: batch events processing (multiple events per tick)
217 : loop {
218 73547 : epoll_vec.clear();
219 73547 : epoll_idx.clear();
220 :
221 : // node events channel
222 73547 : epoll_vec.push(Box::new(node_events.clone()));
223 73547 : epoll_idx.push(0);
224 :
225 : // tcp connections
226 274837 : for conn in conns.values() {
227 274837 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
228 274837 : epoll_idx.push(conn.tcp.connection_id());
229 274837 : }
230 :
231 : // waiting for the next message
232 73547 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
233 :
234 73547 : if index == 0 {
235 : // got a new connection
236 24042 : match node_events.must_recv() {
237 24042 : NodeEvent::Accept(tcp) => {
238 24042 : conns.insert(
239 24042 : tcp.connection_id(),
240 : ConnState {
241 24042 : tcp,
242 : greeting: false,
243 24042 : ttid: TenantTimelineId::empty(),
244 : flush_pending: false,
245 24042 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
246 : },
247 : );
248 : }
249 0 : NodeEvent::Internal(_) => unreachable!(),
250 : }
251 24042 : continue;
252 49505 : }
253 :
254 49505 : let connection_id = epoll_idx[index];
255 49505 : let conn = conns.get_mut(&connection_id).unwrap();
256 49505 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
257 :
258 : loop {
259 81528 : let event = match next_event {
260 41567 : Some(event) => event,
261 39961 : None => break,
262 : };
263 :
264 41567 : match event {
265 27707 : NetEvent::Message(msg) => {
266 27707 : let res = conn.process_any(msg, &mut global);
267 27707 : if res.is_err() {
268 9544 : let e = res.unwrap_err();
269 9544 : let estr = e.to_string();
270 9544 : if !estr.contains("finished processing START_REPLICATION") {
271 9264 : warn!("conn {:?} error: {:?}", connection_id, e);
272 9264 : panic!("unexpected error at safekeeper: {e:#}");
273 280 : }
274 280 : conns.remove(&connection_id);
275 280 : break;
276 18163 : }
277 : }
278 13860 : NetEvent::Closed => {
279 13860 : // TODO: remove from conns?
280 13860 : }
281 : }
282 :
283 32023 : next_event = conn.tcp.recv_chan().try_recv();
284 : }
285 :
286 182760 : conns.retain(|_, conn| {
287 182760 : let res = conn.flush(&mut global);
288 182760 : if res.is_err() {
289 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
290 182760 : }
291 182760 : res.is_ok()
292 182760 : });
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 27707 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
299 27707 : if let AnyMessage::Bytes(copy_data) = any {
300 27707 : let repl_prefix = b"START_REPLICATION ";
301 27707 : if !self.greeting && copy_data.starts_with(repl_prefix) {
302 280 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
303 280 : bail!("finished processing START_REPLICATION")
304 27427 : }
305 :
306 27427 : let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTO_VERSION_3)?;
307 27427 : debug!("got msg: {:?}", msg);
308 27427 : self.process(msg, global)
309 : } else {
310 0 : bail!("unexpected message, expected AnyMessage::Bytes");
311 : }
312 27707 : }
313 :
314 : /// Process START_REPLICATION request.
315 280 : fn process_start_replication(
316 280 : &mut self,
317 280 : copy_data: Bytes,
318 280 : global: &mut GlobalMap,
319 280 : ) -> Result<()> {
320 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
321 280 : let str = String::from_utf8(copy_data.to_vec())?;
322 :
323 280 : let mut parts = str.split(' ');
324 280 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
325 280 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
326 280 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
327 280 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
328 :
329 280 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
330 280 : let shared_state = global.get(&ttid);
331 :
332 : // read bytes from start_lsn to end_lsn
333 280 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
334 280 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
335 :
336 : // send bytes to the client
337 280 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
338 280 : Ok(())
339 280 : }
340 :
341 : /// Get or create a timeline.
342 18830 : fn init_timeline(
343 18830 : &mut self,
344 18830 : ttid: TenantTimelineId,
345 18830 : server_info: ServerInfo,
346 18830 : global: &mut GlobalMap,
347 18830 : ) -> Result<()> {
348 18830 : self.ttid = ttid;
349 18830 : if global.has_tli(&ttid) {
350 17399 : return Ok(());
351 1431 : }
352 :
353 1431 : global.create(ttid, server_info)
354 18830 : }
355 :
356 : /// Process a ProposerAcceptorMessage.
357 27427 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
358 27427 : if !self.greeting {
359 18830 : self.greeting = true;
360 :
361 18830 : match msg {
362 18830 : ProposerAcceptorMessage::Greeting(ref greeting) => {
363 18830 : tracing::info!(
364 0 : "start handshake with walproposer {:?} {:?}",
365 : self.tcp,
366 : greeting
367 : );
368 18830 : let server_info = ServerInfo {
369 18830 : pg_version: greeting.pg_version,
370 18830 : system_id: greeting.system_id,
371 18830 : wal_seg_size: greeting.wal_seg_size,
372 18830 : };
373 18830 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
374 18830 : self.init_timeline(ttid, server_info, global)?
375 : }
376 : _ => {
377 0 : bail!("unexpected message {msg:?} instead of greeting");
378 : }
379 : }
380 8597 : }
381 :
382 27427 : let tli = global.get(&self.ttid);
383 :
384 27427 : match msg {
385 4124 : ProposerAcceptorMessage::AppendRequest(append_request) => {
386 4124 : self.flush_pending = true;
387 4124 : self.process_sk_msg(
388 4124 : tli,
389 4124 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
390 0 : )?;
391 : }
392 23303 : other => {
393 23303 : self.process_sk_msg(tli, &other)?;
394 : }
395 : }
396 :
397 27427 : Ok(())
398 27427 : }
399 :
400 : /// Process FlushWAL if needed.
401 182760 : 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 182760 : if !self.flush_pending {
404 179358 : return Ok(());
405 3402 : }
406 3402 : self.flush_pending = false;
407 3402 : let shared_state = global.get(&self.ttid);
408 3402 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
409 182760 : }
410 :
411 : /// Make safekeeper process a message and send a reply to the TCP
412 30829 : fn process_sk_msg(
413 30829 : &mut self,
414 30829 : shared_state: &mut SharedState,
415 30829 : msg: &ProposerAcceptorMessage,
416 30829 : ) -> Result<()> {
417 30829 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
418 30829 : if let Some(reply) = &mut reply {
419 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
420 :
421 25653 : let mut buf = BytesMut::with_capacity(128);
422 25653 : reply.serialize(&mut buf, SK_PROTO_VERSION_3)?;
423 :
424 25653 : self.tcp.send(AnyMessage::Bytes(buf.into()));
425 5176 : }
426 30829 : Ok(())
427 30829 : }
428 : }
429 :
430 : impl Drop for ConnState {
431 24014 : fn drop(&mut self) {
432 24014 : debug!("dropping conn: {:?}", self.tcp);
433 24014 : if !std::thread::panicking() {
434 280 : self.tcp.close();
435 23734 : }
436 : // TODO: clean up non-fsynced WAL
437 24014 : }
438 : }
|
