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 1524 : pub fn new() -> Self {
33 1524 : SafekeeperDisk {
34 1524 : timelines: Mutex::new(HashMap::new()),
35 1524 : }
36 1524 : }
37 :
38 1462 : pub fn put_state(
39 1462 : &self,
40 1462 : ttid: &TenantTimelineId,
41 1462 : state: TimelinePersistentState,
42 1462 : ) -> Arc<TimelineDisk> {
43 1462 : self.timelines
44 1462 : .lock()
45 1462 : .entry(*ttid)
46 1462 : .and_modify(|e| {
47 0 : let mut mu = e.state.lock();
48 0 : *mu = state.clone();
49 1462 : })
50 1462 : .or_insert_with(|| {
51 1462 : Arc::new(TimelineDisk {
52 1462 : state: Mutex::new(state),
53 1462 : wal: Mutex::new(BlockStorage::new()),
54 1462 : })
55 1462 : })
56 1462 : .clone()
57 1462 : }
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 8989 : pub fn new(disk: Arc<TimelineDisk>) -> Self {
75 8989 : let guard = disk.state.lock();
76 8989 : let state = guard.clone();
77 8989 : drop(guard);
78 8989 : DiskStateStorage {
79 8989 : persisted_state: state,
80 8989 : disk,
81 8989 : last_persist_at: Instant::now(),
82 8989 : }
83 8989 : }
84 : }
85 :
86 : impl control_file::Storage for DiskStateStorage {
87 : /// Persist safekeeper state on disk and update internal state.
88 3692 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()> {
89 3692 : self.persisted_state = s.clone();
90 3692 : *self.disk.state.lock() = s.clone();
91 3692 : Ok(())
92 3692 : }
93 :
94 : /// Timestamp of last persist.
95 0 : fn last_persist_at(&self) -> Instant {
96 0 : // TODO: don't rely on it in tests
97 0 : self.last_persist_at
98 0 : }
99 : }
100 :
101 : impl Deref for DiskStateStorage {
102 : type Target = TimelinePersistentState;
103 :
104 250550 : fn deref(&self) -> &Self::Target {
105 250550 : &self.persisted_state
106 250550 : }
107 : }
108 :
109 : /// Implementation of `wal_storage::Storage` trait.
110 : pub struct DiskWALStorage {
111 : /// Written to disk, but possibly still in the cache and not fully persisted.
112 : /// Also can be ahead of record_lsn, if happen to be in the middle of a WAL record.
113 : write_lsn: Lsn,
114 :
115 : /// The LSN of the last WAL record written to disk. Still can be not fully flushed.
116 : write_record_lsn: Lsn,
117 :
118 : /// The LSN of the last WAL record flushed to disk.
119 : flush_record_lsn: Lsn,
120 :
121 : /// Decoder is required for detecting boundaries of WAL records.
122 : decoder: WalStreamDecoder,
123 :
124 : /// Bytes of WAL records that are not yet written to disk.
125 : unflushed_bytes: BytesMut,
126 :
127 : /// Contains BlockStorage for WAL.
128 : disk: Arc<TimelineDisk>,
129 : }
130 :
131 : impl DiskWALStorage {
132 8989 : pub fn new(disk: Arc<TimelineDisk>, state: &TimelinePersistentState) -> Result<Self> {
133 8989 : let write_lsn = if state.commit_lsn == Lsn(0) {
134 8365 : Lsn(0)
135 : } else {
136 624 : Self::find_end_of_wal(disk.clone(), state.commit_lsn)?
137 : };
138 :
139 8989 : let flush_lsn = write_lsn;
140 8989 : Ok(DiskWALStorage {
141 8989 : write_lsn,
142 8989 : write_record_lsn: flush_lsn,
143 8989 : flush_record_lsn: flush_lsn,
144 8989 : decoder: WalStreamDecoder::new(flush_lsn, 16),
145 8989 : unflushed_bytes: BytesMut::new(),
146 8989 : disk,
147 8989 : })
148 8989 : }
149 :
150 624 : fn find_end_of_wal(disk: Arc<TimelineDisk>, start_lsn: Lsn) -> Result<Lsn> {
151 624 : let mut buf = [0; 8192];
152 624 : let mut pos = start_lsn.0;
153 624 : let mut decoder = WalStreamDecoder::new(start_lsn, 16);
154 624 : let mut result = start_lsn;
155 : loop {
156 624 : disk.wal.lock().read(pos, &mut buf);
157 624 : pos += buf.len() as u64;
158 624 : decoder.feed_bytes(&buf);
159 :
160 : loop {
161 3189 : match decoder.poll_decode() {
162 2565 : Ok(Some(record)) => result = record.0,
163 624 : Err(e) => {
164 624 : debug!(
165 0 : "find_end_of_wal reached end at {:?}, decode error: {:?}",
166 : result, e
167 : );
168 624 : return Ok(result);
169 : }
170 0 : Ok(None) => break, // need more data
171 : }
172 : }
173 : }
174 624 : }
175 : }
176 :
177 : impl wal_storage::Storage for DiskWALStorage {
178 : // Last written LSN.
179 3040 : fn write_lsn(&self) -> Lsn {
180 3040 : self.write_lsn
181 3040 : }
182 : /// LSN of last durably stored WAL record.
183 20333 : fn flush_lsn(&self) -> Lsn {
184 20333 : self.flush_record_lsn
185 20333 : }
186 :
187 148 : async fn initialize_first_segment(&mut self, _init_lsn: Lsn) -> Result<()> {
188 148 : Ok(())
189 148 : }
190 :
191 : /// Write piece of WAL from buf to disk, but not necessarily sync it.
192 730 : async fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> Result<()> {
193 730 : if self.write_lsn != startpos {
194 0 : panic!("write_wal called with wrong startpos");
195 730 : }
196 730 :
197 730 : self.unflushed_bytes.extend_from_slice(buf);
198 730 : self.write_lsn += buf.len() as u64;
199 730 :
200 730 : if self.decoder.available() != startpos {
201 0 : info!(
202 0 : "restart decoder from {} to {}",
203 0 : self.decoder.available(),
204 : startpos,
205 : );
206 0 : self.decoder = WalStreamDecoder::new(startpos, 16);
207 730 : }
208 730 : self.decoder.feed_bytes(buf);
209 : loop {
210 22390 : match self.decoder.poll_decode()? {
211 730 : None => break, // no full record yet
212 21660 : Some((lsn, _rec)) => {
213 21660 : self.write_record_lsn = lsn;
214 21660 : }
215 : }
216 : }
217 :
218 730 : Ok(())
219 730 : }
220 :
221 : /// Truncate WAL at specified LSN, which must be the end of WAL record.
222 758 : async fn truncate_wal(&mut self, end_pos: Lsn) -> Result<()> {
223 758 : if self.write_lsn != Lsn(0) && end_pos > self.write_lsn {
224 0 : panic!(
225 0 : "truncate_wal called on non-written WAL, write_lsn={}, end_pos={}",
226 0 : self.write_lsn, end_pos
227 0 : );
228 758 : }
229 758 :
230 758 : self.flush_wal().await?;
231 :
232 : // write zeroes to disk from end_pos until self.write_lsn
233 758 : let buf = [0; 8192];
234 758 : let mut pos = end_pos.0;
235 762 : while pos < self.write_lsn.0 {
236 4 : self.disk.wal.lock().write(pos, &buf);
237 4 : pos += buf.len() as u64;
238 4 : }
239 :
240 758 : self.write_lsn = end_pos;
241 758 : self.write_record_lsn = end_pos;
242 758 : self.flush_record_lsn = end_pos;
243 758 : self.unflushed_bytes.clear();
244 758 : self.decoder = WalStreamDecoder::new(end_pos, 16);
245 758 :
246 758 : Ok(())
247 758 : }
248 :
249 : /// Durably store WAL on disk, up to the last written WAL record.
250 6345 : async fn flush_wal(&mut self) -> Result<()> {
251 6345 : if self.flush_record_lsn == self.write_record_lsn {
252 : // no need to do extra flush
253 5616 : return Ok(());
254 729 : }
255 729 :
256 729 : let num_bytes = self.write_record_lsn.0 - self.flush_record_lsn.0;
257 729 :
258 729 : self.disk.wal.lock().write(
259 729 : self.flush_record_lsn.0,
260 729 : &self.unflushed_bytes[..num_bytes as usize],
261 729 : );
262 729 : self.unflushed_bytes.advance(num_bytes as usize);
263 729 : self.flush_record_lsn = self.write_record_lsn;
264 729 :
265 729 : Ok(())
266 6345 : }
267 :
268 : /// Remove all segments <= given segno. Returns function doing that as we
269 : /// want to perform it without timeline lock.
270 0 : fn remove_up_to(&self, _segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>> {
271 0 : Box::pin(async move { Ok(()) })
272 0 : }
273 :
274 : /// Release resources associated with the storage -- technically, close FDs.
275 : /// Currently we don't remove timelines until restart (#3146), so need to
276 : /// spare descriptors. This would be useful for temporary tli detach as
277 : /// well.
278 0 : fn close(&mut self) {}
279 :
280 : /// Get metrics for this timeline.
281 0 : fn get_metrics(&self) -> WalStorageMetrics {
282 0 : WalStorageMetrics::default()
283 0 : }
284 : }
|
