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 9931 : fn new(disk: Arc<SafekeeperDisk>, conf: SafeKeeperConf) -> Result<Self> {
46 9931 : let mut timelines = HashMap::new();
47 :
48 9931 : for (&ttid, disk) in disk.timelines.lock().iter() {
49 7553 : debug!("loading timeline {}", ttid);
50 7553 : let state = disk.state.lock().clone();
51 7553 :
52 7553 : if state.server.wal_seg_size == 0 {
53 0 : bail!(TimelineError::UninitializedWalSegSize(ttid));
54 7553 : }
55 7553 :
56 7553 : if state.server.pg_version == UNKNOWN_SERVER_VERSION {
57 0 : bail!(TimelineError::UninitialinzedPgVersion(ttid));
58 7553 : }
59 7553 :
60 7553 : if state.commit_lsn < state.local_start_lsn {
61 0 : bail!(
62 0 : "commit_lsn {} is smaller than local_start_lsn {}",
63 0 : state.commit_lsn,
64 0 : state.local_start_lsn
65 0 : );
66 7553 : }
67 7553 :
68 7553 : let control_store = DiskStateStorage::new(disk.clone());
69 7553 : let wal_store = DiskWALStorage::new(disk.clone(), &control_store)?;
70 :
71 7553 : let sk = SafeKeeper::new(TimelineState::new(control_store), wal_store, conf.my_id)?;
72 7553 : timelines.insert(
73 7553 : ttid,
74 7553 : SharedState {
75 7553 : sk,
76 7553 : disk: disk.clone(),
77 7553 : },
78 7553 : );
79 : }
80 :
81 9931 : Ok(Self {
82 9931 : timelines,
83 9931 : conf,
84 9931 : disk,
85 9931 : })
86 9931 : }
87 :
88 1460 : fn create(&mut self, ttid: TenantTimelineId, server_info: ServerInfo) -> Result<()> {
89 1460 : if self.timelines.contains_key(&ttid) {
90 0 : bail!("timeline {} already exists", ttid);
91 1460 : }
92 1460 :
93 1460 : debug!("creating new timeline {}", ttid);
94 :
95 1460 : let commit_lsn = Lsn::INVALID;
96 1460 : let local_start_lsn = Lsn::INVALID;
97 :
98 1460 : let state = TimelinePersistentState::new(
99 1460 : &ttid,
100 1460 : Configuration::empty(),
101 1460 : server_info,
102 1460 : commit_lsn,
103 1460 : local_start_lsn,
104 1460 : )?;
105 :
106 1460 : let disk_timeline = self.disk.put_state(&ttid, state);
107 1460 : let control_store = DiskStateStorage::new(disk_timeline.clone());
108 1460 : let wal_store = DiskWALStorage::new(disk_timeline.clone(), &control_store)?;
109 :
110 1460 : let sk = SafeKeeper::new(
111 1460 : TimelineState::new(control_store),
112 1460 : wal_store,
113 1460 : self.conf.my_id,
114 1460 : )?;
115 :
116 1460 : self.timelines.insert(
117 1460 : ttid,
118 1460 : SharedState {
119 1460 : sk,
120 1460 : disk: disk_timeline,
121 1460 : },
122 1460 : );
123 1460 : Ok(())
124 1460 : }
125 :
126 33683 : fn get(&mut self, ttid: &TenantTimelineId) -> &mut SharedState {
127 33683 : self.timelines.get_mut(ttid).expect("timeline must exist")
128 33683 : }
129 :
130 20761 : fn has_tli(&self, ttid: &TenantTimelineId) -> bool {
131 20761 : self.timelines.contains_key(ttid)
132 20761 : }
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 9931 : pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
147 9931 : let _enter = info_span!("safekeeper", id = os.id()).entered();
148 9931 : debug!("started server");
149 9931 : os.log_event("started;safekeeper".to_owned());
150 9931 : let conf = SafeKeeperConf {
151 9931 : workdir: Utf8PathBuf::from("."),
152 9931 : my_id: NodeId(os.id() as u64),
153 9931 : listen_pg_addr: String::new(),
154 9931 : listen_http_addr: String::new(),
155 9931 : no_sync: false,
156 9931 : broker_endpoint: "/".parse::<Uri>().unwrap(),
157 9931 : broker_keepalive_interval: Duration::from_secs(0),
158 9931 : heartbeat_timeout: Duration::from_secs(0),
159 9931 : remote_storage: None,
160 9931 : max_offloader_lag_bytes: 0,
161 9931 : wal_backup_enabled: false,
162 9931 : listen_pg_addr_tenant_only: None,
163 9931 : advertise_pg_addr: None,
164 9931 : availability_zone: None,
165 9931 : peer_recovery_enabled: false,
166 9931 : backup_parallel_jobs: 0,
167 9931 : pg_auth: None,
168 9931 : pg_tenant_only_auth: None,
169 9931 : http_auth: None,
170 9931 : sk_auth_token: None,
171 9931 : current_thread_runtime: false,
172 9931 : walsenders_keep_horizon: false,
173 9931 : partial_backup_timeout: Duration::from_secs(0),
174 9931 : disable_periodic_broker_push: false,
175 9931 : enable_offload: false,
176 9931 : delete_offloaded_wal: false,
177 9931 : control_file_save_interval: Duration::from_secs(1),
178 9931 : partial_backup_concurrency: 1,
179 9931 : eviction_min_resident: Duration::ZERO,
180 9931 : wal_reader_fanout: false,
181 9931 : max_delta_for_fanout: None,
182 9931 : };
183 :
184 9931 : let mut global = GlobalMap::new(disk, conf.clone())?;
185 9931 : let mut conns: HashMap<usize, ConnState> = HashMap::new();
186 :
187 9931 : for (&_ttid, shared_state) in global.timelines.iter_mut() {
188 7553 : let flush_lsn = shared_state.sk.wal_store.flush_lsn();
189 7553 : let commit_lsn = shared_state.sk.state.commit_lsn;
190 7553 : os.log_event(format!("tli_loaded;{};{}", flush_lsn.0, commit_lsn.0));
191 7553 : }
192 :
193 9931 : let node_events = os.node_events();
194 9931 : let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![];
195 9931 : let mut epoll_idx: Vec<usize> = vec![];
196 :
197 : // TODO: batch events processing (multiple events per tick)
198 : loop {
199 80062 : epoll_vec.clear();
200 80062 : epoll_idx.clear();
201 80062 :
202 80062 : // node events channel
203 80062 : epoll_vec.push(Box::new(node_events.clone()));
204 80062 : epoll_idx.push(0);
205 :
206 : // tcp connections
207 301050 : for conn in conns.values() {
208 301050 : epoll_vec.push(Box::new(conn.tcp.recv_chan()));
209 301050 : epoll_idx.push(conn.tcp.connection_id());
210 301050 : }
211 :
212 : // waiting for the next message
213 80062 : let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
214 80062 :
215 80062 : if index == 0 {
216 : // got a new connection
217 26434 : match node_events.must_recv() {
218 26434 : NodeEvent::Accept(tcp) => {
219 26434 : conns.insert(
220 26434 : tcp.connection_id(),
221 26434 : ConnState {
222 26434 : tcp,
223 26434 : greeting: false,
224 26434 : ttid: TenantTimelineId::empty(),
225 26434 : flush_pending: false,
226 26434 : runtime: tokio::runtime::Builder::new_current_thread().build()?,
227 : },
228 : );
229 : }
230 0 : NodeEvent::Internal(_) => unreachable!(),
231 : }
232 26434 : continue;
233 53628 : }
234 53628 :
235 53628 : let connection_id = epoll_idx[index];
236 53628 : let conn = conns.get_mut(&connection_id).unwrap();
237 53628 : let mut next_event = Some(conn.tcp.recv_chan().must_recv());
238 :
239 : loop {
240 98419 : let event = match next_event {
241 55051 : Some(event) => event,
242 43368 : None => break,
243 : };
244 :
245 55051 : match event {
246 40055 : NetEvent::Message(msg) => {
247 40055 : let res = conn.process_any(msg, &mut global);
248 40055 : if res.is_err() {
249 10260 : let e = res.unwrap_err();
250 10260 : let estr = e.to_string();
251 10260 : if !estr.contains("finished processing START_REPLICATION") {
252 9931 : warn!("conn {:?} error: {:?}", connection_id, e);
253 0 : panic!("unexpected error at safekeeper: {:#}", e);
254 329 : }
255 329 : conns.remove(&connection_id);
256 329 : break;
257 29795 : }
258 : }
259 14996 : NetEvent::Closed => {
260 14996 : // TODO: remove from conns?
261 14996 : }
262 : }
263 :
264 44791 : next_event = conn.tcp.recv_chan().try_recv();
265 : }
266 :
267 199313 : conns.retain(|_, conn| {
268 199313 : let res = conn.flush(&mut global);
269 199313 : if res.is_err() {
270 0 : debug!("conn {:?} error: {:?}", conn.tcp, res);
271 199313 : }
272 199313 : res.is_ok()
273 199313 : });
274 : }
275 0 : }
276 :
277 : impl ConnState {
278 : /// Process a message from the network. It can be START_REPLICATION request or a valid ProposerAcceptorMessage message.
279 30124 : fn process_any(&mut self, any: AnyMessage, global: &mut GlobalMap) -> Result<()> {
280 30124 : if let AnyMessage::Bytes(copy_data) = any {
281 30124 : let repl_prefix = b"START_REPLICATION ";
282 30124 : if !self.greeting && copy_data.starts_with(repl_prefix) {
283 329 : self.process_start_replication(copy_data.slice(repl_prefix.len()..), global)?;
284 329 : bail!("finished processing START_REPLICATION")
285 29795 : }
286 :
287 29795 : let msg = ProposerAcceptorMessage::parse(copy_data, SK_PROTO_VERSION_3)?;
288 29795 : debug!("got msg: {:?}", msg);
289 29795 : self.process(msg, global)
290 : } else {
291 0 : bail!("unexpected message, expected AnyMessage::Bytes");
292 : }
293 30124 : }
294 :
295 : /// Process START_REPLICATION request.
296 329 : fn process_start_replication(
297 329 : &mut self,
298 329 : copy_data: Bytes,
299 329 : global: &mut GlobalMap,
300 329 : ) -> Result<()> {
301 : // format is "<tenant_id> <timeline_id> <start_lsn> <end_lsn>"
302 329 : let str = String::from_utf8(copy_data.to_vec())?;
303 :
304 329 : let mut parts = str.split(' ');
305 329 : let tenant_id = parts.next().unwrap().parse::<TenantId>()?;
306 329 : let timeline_id = parts.next().unwrap().parse::<TimelineId>()?;
307 329 : let start_lsn = parts.next().unwrap().parse::<u64>()?;
308 329 : let end_lsn = parts.next().unwrap().parse::<u64>()?;
309 :
310 329 : let ttid = TenantTimelineId::new(tenant_id, timeline_id);
311 329 : let shared_state = global.get(&ttid);
312 329 :
313 329 : // read bytes from start_lsn to end_lsn
314 329 : let mut buf = vec![0; (end_lsn - start_lsn) as usize];
315 329 : shared_state.disk.wal.lock().read(start_lsn, &mut buf);
316 329 :
317 329 : // send bytes to the client
318 329 : self.tcp.send(AnyMessage::Bytes(Bytes::from(buf)));
319 329 : Ok(())
320 329 : }
321 :
322 : /// Get or create a timeline.
323 20761 : fn init_timeline(
324 20761 : &mut self,
325 20761 : ttid: TenantTimelineId,
326 20761 : server_info: ServerInfo,
327 20761 : global: &mut GlobalMap,
328 20761 : ) -> Result<()> {
329 20761 : self.ttid = ttid;
330 20761 : if global.has_tli(&ttid) {
331 19301 : return Ok(());
332 1460 : }
333 1460 :
334 1460 : global.create(ttid, server_info)
335 20761 : }
336 :
337 : /// Process a ProposerAcceptorMessage.
338 29795 : fn process(&mut self, msg: ProposerAcceptorMessage, global: &mut GlobalMap) -> Result<()> {
339 29795 : if !self.greeting {
340 20761 : self.greeting = true;
341 20761 :
342 20761 : match msg {
343 20761 : ProposerAcceptorMessage::Greeting(ref greeting) => {
344 20761 : tracing::info!(
345 0 : "start handshake with walproposer {:?} {:?}",
346 : self.tcp,
347 : greeting
348 : );
349 20761 : let server_info = ServerInfo {
350 20761 : pg_version: greeting.pg_version,
351 20761 : system_id: greeting.system_id,
352 20761 : wal_seg_size: greeting.wal_seg_size,
353 20761 : };
354 20761 : let ttid = TenantTimelineId::new(greeting.tenant_id, greeting.timeline_id);
355 20761 : self.init_timeline(ttid, server_info, global)?
356 : }
357 : _ => {
358 0 : bail!("unexpected message {msg:?} instead of greeting");
359 : }
360 : }
361 9034 : }
362 :
363 29795 : let tli = global.get(&self.ttid);
364 29795 :
365 29795 : match msg {
366 4301 : ProposerAcceptorMessage::AppendRequest(append_request) => {
367 4301 : self.flush_pending = true;
368 4301 : self.process_sk_msg(
369 4301 : tli,
370 4301 : &ProposerAcceptorMessage::NoFlushAppendRequest(append_request),
371 4301 : )?;
372 : }
373 25494 : other => {
374 25494 : self.process_sk_msg(tli, &other)?;
375 : }
376 : }
377 :
378 29795 : Ok(())
379 29795 : }
380 :
381 : /// Process FlushWAL if needed.
382 199313 : fn flush(&mut self, global: &mut GlobalMap) -> Result<()> {
383 199313 : // TODO: try to add extra flushes in simulation, to verify that extra flushes don't break anything
384 199313 : if !self.flush_pending {
385 195754 : return Ok(());
386 3559 : }
387 3559 : self.flush_pending = false;
388 3559 : let shared_state = global.get(&self.ttid);
389 3559 : self.process_sk_msg(shared_state, &ProposerAcceptorMessage::FlushWAL)
390 199313 : }
391 :
392 : /// Make safekeeper process a message and send a reply to the TCP
393 33354 : fn process_sk_msg(
394 33354 : &mut self,
395 33354 : shared_state: &mut SharedState,
396 33354 : msg: &ProposerAcceptorMessage,
397 33354 : ) -> Result<()> {
398 33354 : let mut reply = self.runtime.block_on(shared_state.sk.process_msg(msg))?;
399 33354 : if let Some(reply) = &mut reply {
400 : // TODO: if this is AppendResponse, fill in proper hot standby feedback and disk consistent lsn
401 :
402 27952 : let mut buf = BytesMut::with_capacity(128);
403 27952 : reply.serialize(&mut buf, SK_PROTO_VERSION_3)?;
404 :
405 27952 : self.tcp.send(AnyMessage::Bytes(buf.into()));
406 5402 : }
407 33354 : Ok(())
408 33354 : }
409 : }
410 :
411 : impl Drop for ConnState {
412 26406 : fn drop(&mut self) {
413 26406 : debug!("dropping conn: {:?}", self.tcp);
414 26406 : if !std::thread::panicking() {
415 329 : self.tcp.close();
416 26077 : }
417 : // TODO: clean up non-fsynced WAL
418 26406 : }
419 : }
|
