Line data Source code
1 : use std::collections::HashMap;
2 : use std::sync::Arc;
3 :
4 : use parking_lot::Mutex;
5 : use safekeeper::state::TimelinePersistentState;
6 : use utils::id::TenantTimelineId;
7 :
8 : use super::block_storage::BlockStorage;
9 :
10 : use std::{ops::Deref, time::Instant};
11 :
12 : use anyhow::Result;
13 : use bytes::{Buf, BytesMut};
14 : use futures::future::BoxFuture;
15 : use postgres_ffi::{waldecoder::WalStreamDecoder, XLogSegNo};
16 : use safekeeper::{control_file, metrics::WalStorageMetrics, wal_storage};
17 : use tracing::{debug, info};
18 : use utils::lsn::Lsn;
19 :
20 : /// All safekeeper state that is usually saved to disk.
21 : pub struct SafekeeperDisk {
22 : pub timelines: Mutex<HashMap<TenantTimelineId, Arc<TimelineDisk>>>,
23 : }
24 :
25 : impl Default for SafekeeperDisk {
26 0 : fn default() -> Self {
27 0 : Self::new()
28 0 : }
29 : }
30 :
31 : impl SafekeeperDisk {
32 12048 : pub fn new() -> Self {
33 12048 : SafekeeperDisk {
34 12048 : timelines: Mutex::new(HashMap::new()),
35 12048 : }
36 12048 : }
37 :
38 11438 : pub fn put_state(
39 11438 : &self,
40 11438 : ttid: &TenantTimelineId,
41 11438 : state: TimelinePersistentState,
42 11438 : ) -> Arc<TimelineDisk> {
43 11438 : self.timelines
44 11438 : .lock()
45 11438 : .entry(*ttid)
46 11438 : .and_modify(|e| {
47 0 : let mut mu = e.state.lock();
48 0 : *mu = state.clone();
49 11438 : })
50 11438 : .or_insert_with(|| {
51 11438 : Arc::new(TimelineDisk {
52 11438 : state: Mutex::new(state),
53 11438 : wal: Mutex::new(BlockStorage::new()),
54 11438 : })
55 11438 : })
56 11438 : .clone()
57 11438 : }
58 : }
59 :
60 : /// Control file state and WAL storage.
61 : pub struct TimelineDisk {
62 : pub state: Mutex<TimelinePersistentState>,
63 : pub wal: Mutex<BlockStorage>,
64 : }
65 :
66 : /// Implementation of `control_file::Storage` trait.
67 : pub struct DiskStateStorage {
68 : persisted_state: TimelinePersistentState,
69 : disk: Arc<TimelineDisk>,
70 : last_persist_at: Instant,
71 : }
72 :
73 : impl DiskStateStorage {
74 69570 : pub fn new(disk: Arc<TimelineDisk>) -> Self {
75 69570 : let guard = disk.state.lock();
76 69570 : let state = guard.clone();
77 69570 : drop(guard);
78 69570 : DiskStateStorage {
79 69570 : persisted_state: state,
80 69570 : disk,
81 69570 : last_persist_at: Instant::now(),
82 69570 : }
83 69570 : }
84 : }
85 :
86 : #[async_trait::async_trait]
87 : impl control_file::Storage for DiskStateStorage {
88 : /// Persist safekeeper state on disk and update internal state.
89 32953 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()> {
90 32953 : self.persisted_state = s.clone();
91 32953 : *self.disk.state.lock() = s.clone();
92 32953 : Ok(())
93 32953 : }
94 :
95 : /// Timestamp of last persist.
96 0 : fn last_persist_at(&self) -> Instant {
97 0 : // TODO: don't rely on it in tests
98 0 : self.last_persist_at
99 0 : }
100 : }
101 :
102 : impl Deref for DiskStateStorage {
103 : type Target = TimelinePersistentState;
104 :
105 1970349 : fn deref(&self) -> &Self::Target {
106 1970349 : &self.persisted_state
107 1970349 : }
108 : }
109 :
110 : /// Implementation of `wal_storage::Storage` trait.
111 : pub struct DiskWALStorage {
112 : /// Written to disk, but possibly still in the cache and not fully persisted.
113 : /// Also can be ahead of record_lsn, if happen to be in the middle of a WAL record.
114 : write_lsn: Lsn,
115 :
116 : /// The LSN of the last WAL record written to disk. Still can be not fully flushed.
117 : write_record_lsn: Lsn,
118 :
119 : /// The LSN of the last WAL record flushed to disk.
120 : flush_record_lsn: Lsn,
121 :
122 : /// Decoder is required for detecting boundaries of WAL records.
123 : decoder: WalStreamDecoder,
124 :
125 : /// Bytes of WAL records that are not yet written to disk.
126 : unflushed_bytes: BytesMut,
127 :
128 : /// Contains BlockStorage for WAL.
129 : disk: Arc<TimelineDisk>,
130 : }
131 :
132 : impl DiskWALStorage {
133 69570 : pub fn new(disk: Arc<TimelineDisk>, state: &TimelinePersistentState) -> Result<Self> {
134 69570 : let write_lsn = if state.commit_lsn == Lsn(0) {
135 64684 : Lsn(0)
136 : } else {
137 4886 : Self::find_end_of_wal(disk.clone(), state.commit_lsn)?
138 : };
139 :
140 69570 : let flush_lsn = write_lsn;
141 69570 : Ok(DiskWALStorage {
142 69570 : write_lsn,
143 69570 : write_record_lsn: flush_lsn,
144 69570 : flush_record_lsn: flush_lsn,
145 69570 : decoder: WalStreamDecoder::new(flush_lsn, 16),
146 69570 : unflushed_bytes: BytesMut::new(),
147 69570 : disk,
148 69570 : })
149 69570 : }
150 :
151 4886 : fn find_end_of_wal(disk: Arc<TimelineDisk>, start_lsn: Lsn) -> Result<Lsn> {
152 4886 : let mut buf = [0; 8192];
153 4886 : let mut pos = start_lsn.0;
154 4886 : let mut decoder = WalStreamDecoder::new(start_lsn, 16);
155 4886 : let mut result = start_lsn;
156 4886 : loop {
157 4886 : disk.wal.lock().read(pos, &mut buf);
158 4886 : pos += buf.len() as u64;
159 4886 : decoder.feed_bytes(&buf);
160 :
161 : loop {
162 24920 : match decoder.poll_decode() {
163 20034 : Ok(Some(record)) => result = record.0,
164 4886 : Err(e) => {
165 4886 : debug!(
166 0 : "find_end_of_wal reached end at {:?}, decode error: {:?}",
167 : result, e
168 : );
169 4886 : return Ok(result);
170 : }
171 0 : Ok(None) => break, // need more data
172 : }
173 : }
174 : }
175 4886 : }
176 : }
177 :
178 : #[async_trait::async_trait]
179 : impl wal_storage::Storage for DiskWALStorage {
180 : /// LSN of last durably stored WAL record.
181 156731 : fn flush_lsn(&self) -> Lsn {
182 156731 : self.flush_record_lsn
183 156731 : }
184 :
185 1125 : async fn initialize_first_segment(&mut self, _init_lsn: Lsn) -> Result<()> {
186 1125 : Ok(())
187 1125 : }
188 :
189 : /// Write piece of WAL from buf to disk, but not necessarily sync it.
190 4275 : async fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> Result<()> {
191 4275 : if self.write_lsn != startpos {
192 4275 : panic!("write_wal called with wrong startpos");
193 4275 : }
194 4275 :
195 4275 : self.unflushed_bytes.extend_from_slice(buf);
196 4275 : self.write_lsn += buf.len() as u64;
197 4275 :
198 4275 : if self.decoder.available() != startpos {
199 4275 : info!(
200 4275 : "restart decoder from {} to {}",
201 0 : self.decoder.available(),
202 4275 : startpos,
203 4275 : );
204 4275 : self.decoder = WalStreamDecoder::new(startpos, 16);
205 4275 : }
206 4275 : self.decoder.feed_bytes(buf);
207 4275 : loop {
208 65638 : match self.decoder.poll_decode()? {
209 4275 : None => break, // no full record yet
210 61363 : Some((lsn, _rec)) => {
211 61363 : self.write_record_lsn = lsn;
212 61363 : }
213 4275 : }
214 4275 : }
215 4275 :
216 4275 : Ok(())
217 4275 : }
218 :
219 : /// Truncate WAL at specified LSN, which must be the end of WAL record.
220 7352 : async fn truncate_wal(&mut self, end_pos: Lsn) -> Result<()> {
221 7352 : if self.write_lsn != Lsn(0) && end_pos > self.write_lsn {
222 7352 : panic!(
223 0 : "truncate_wal called on non-written WAL, write_lsn={}, end_pos={}",
224 0 : self.write_lsn, end_pos
225 0 : );
226 7352 : }
227 7352 :
228 7352 : self.flush_wal().await?;
229 7352 :
230 7352 : // write zeroes to disk from end_pos until self.write_lsn
231 7352 : let buf = [0; 8192];
232 7352 : let mut pos = end_pos.0;
233 7383 : while pos < self.write_lsn.0 {
234 31 : self.disk.wal.lock().write(pos, &buf);
235 31 : pos += buf.len() as u64;
236 31 : }
237 7352 :
238 7352 : self.write_lsn = end_pos;
239 7352 : self.write_record_lsn = end_pos;
240 7352 : self.flush_record_lsn = end_pos;
241 7352 : self.unflushed_bytes.clear();
242 7352 : self.decoder = WalStreamDecoder::new(end_pos, 16);
243 7352 :
244 7352 : Ok(())
245 7352 : }
246 :
247 : /// Durably store WAL on disk, up to the last written WAL record.
248 53104 : async fn flush_wal(&mut self) -> Result<()> {
249 53104 : if self.flush_record_lsn == self.write_record_lsn {
250 53104 : // no need to do extra flush
251 53104 : return Ok(());
252 53104 : }
253 4248 :
254 4248 : let num_bytes = self.write_record_lsn.0 - self.flush_record_lsn.0;
255 4248 :
256 4248 : self.disk.wal.lock().write(
257 4248 : self.flush_record_lsn.0,
258 4248 : &self.unflushed_bytes[..num_bytes as usize],
259 4248 : );
260 4248 : self.unflushed_bytes.advance(num_bytes as usize);
261 4248 : self.flush_record_lsn = self.write_record_lsn;
262 4248 :
263 4248 : Ok(())
264 53104 : }
265 :
266 : /// Remove all segments <= given segno. Returns function doing that as we
267 : /// want to perform it without timeline lock.
268 0 : fn remove_up_to(&self, _segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>> {
269 0 : Box::pin(async move { Ok(()) })
270 0 : }
271 :
272 : /// Release resources associated with the storage -- technically, close FDs.
273 : /// Currently we don't remove timelines until restart (#3146), so need to
274 : /// spare descriptors. This would be useful for temporary tli detach as
275 : /// well.
276 0 : fn close(&mut self) {}
277 :
278 : /// Get metrics for this timeline.
279 0 : fn get_metrics(&self) -> WalStorageMetrics {
280 0 : WalStorageMetrics::default()
281 0 : }
282 : }
|
