Line data Source code
1 : //! This module has everything to deal with WAL -- reading and writing to disk.
2 : //!
3 : //! Safekeeper WAL is stored in the timeline directory, in format similar to pg_wal.
4 : //! PG timeline is always 1, so WAL segments are usually have names like this:
5 : //! - 000000010000000000000001
6 : //! - 000000010000000000000002.partial
7 : //!
8 : //! Note that last file has `.partial` suffix, that's different from postgres.
9 :
10 : use anyhow::{bail, Context, Result};
11 : use bytes::Bytes;
12 : use camino::{Utf8Path, Utf8PathBuf};
13 : use futures::future::BoxFuture;
14 : use postgres_ffi::v14::xlog_utils::{IsPartialXLogFileName, IsXLogFileName, XLogFromFileName};
15 : use postgres_ffi::{dispatch_pgversion, XLogSegNo, PG_TLI};
16 : use remote_storage::RemotePath;
17 : use std::cmp::{max, min};
18 : use std::future::Future;
19 : use std::io::{self, SeekFrom};
20 : use std::pin::Pin;
21 : use tokio::fs::{self, remove_file, File, OpenOptions};
22 : use tokio::io::{AsyncRead, AsyncWriteExt};
23 : use tokio::io::{AsyncReadExt, AsyncSeekExt};
24 : use tracing::*;
25 : use utils::crashsafe::durable_rename;
26 :
27 : use crate::metrics::{
28 : time_io_closure, WalStorageMetrics, REMOVED_WAL_SEGMENTS, WAL_STORAGE_OPERATION_SECONDS,
29 : };
30 : use crate::state::TimelinePersistentState;
31 : use crate::wal_backup::{read_object, remote_timeline_path};
32 : use crate::SafeKeeperConf;
33 : use postgres_ffi::waldecoder::WalStreamDecoder;
34 : use postgres_ffi::XLogFileName;
35 : use postgres_ffi::XLOG_BLCKSZ;
36 : use pq_proto::SystemId;
37 : use utils::{id::TenantTimelineId, lsn::Lsn};
38 :
39 : pub trait Storage {
40 : // Last written LSN.
41 : fn write_lsn(&self) -> Lsn;
42 : /// LSN of last durably stored WAL record.
43 : fn flush_lsn(&self) -> Lsn;
44 :
45 : /// Initialize segment by creating proper long header at the beginning of
46 : /// the segment and short header at the page of given LSN. This is only used
47 : /// for timeline initialization because compute will stream data only since
48 : /// init_lsn. Other segment headers are included in compute stream.
49 : fn initialize_first_segment(
50 : &mut self,
51 : init_lsn: Lsn,
52 : ) -> impl Future<Output = Result<()>> + Send;
53 :
54 : /// Write piece of WAL from buf to disk, but not necessarily sync it.
55 : fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> impl Future<Output = Result<()>> + Send;
56 :
57 : /// Truncate WAL at specified LSN, which must be the end of WAL record.
58 : fn truncate_wal(&mut self, end_pos: Lsn) -> impl Future<Output = Result<()>> + Send;
59 :
60 : /// Durably store WAL on disk, up to the last written WAL record.
61 : fn flush_wal(&mut self) -> impl Future<Output = Result<()>> + Send;
62 :
63 : /// Remove all segments <= given segno. Returns function doing that as we
64 : /// want to perform it without timeline lock.
65 : fn remove_up_to(&self, segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>>;
66 :
67 : /// Release resources associated with the storage -- technically, close FDs.
68 : /// Currently we don't remove timelines until restart (#3146), so need to
69 : /// spare descriptors. This would be useful for temporary tli detach as
70 : /// well.
71 0 : fn close(&mut self) {}
72 :
73 : /// Get metrics for this timeline.
74 : fn get_metrics(&self) -> WalStorageMetrics;
75 : }
76 :
77 : /// PhysicalStorage is a storage that stores WAL on disk. Writes are separated from flushes
78 : /// for better performance. Storage is initialized in the constructor.
79 : ///
80 : /// WAL is stored in segments, each segment is a file. Last segment has ".partial" suffix in
81 : /// its filename and may be not fully flushed.
82 : ///
83 : /// Relationship of LSNs:
84 : /// `write_lsn` >= `write_record_lsn` >= `flush_record_lsn`
85 : ///
86 : /// When storage is created first time, all LSNs are zeroes and there are no segments on disk.
87 : pub struct PhysicalStorage {
88 : metrics: WalStorageMetrics,
89 : timeline_dir: Utf8PathBuf,
90 : conf: SafeKeeperConf,
91 :
92 : /// Size of WAL segment in bytes.
93 : wal_seg_size: usize,
94 : pg_version: u32,
95 : system_id: u64,
96 :
97 : /// Written to disk, but possibly still in the cache and not fully persisted.
98 : /// Also can be ahead of record_lsn, if happen to be in the middle of a WAL record.
99 : write_lsn: Lsn,
100 :
101 : /// The LSN of the last WAL record written to disk. Still can be not fully
102 : /// flushed.
103 : ///
104 : /// Note: Normally it (and flush_record_lsn) is <= write_lsn, but after xlog
105 : /// switch ingest the reverse is true because we don't bump write_lsn up to
106 : /// the next segment: WAL stream from the compute doesn't have the gap and
107 : /// for simplicity / as a sanity check we disallow any non-sequential
108 : /// writes, so write zeros as is.
109 : ///
110 : /// Similar effect is in theory possible due to LSN alignment: if record
111 : /// ends at *2, decoder will report end lsn as *8 even though we haven't
112 : /// written these zeros yet. In practice compute likely never sends
113 : /// non-aligned chunks of data.
114 : write_record_lsn: Lsn,
115 :
116 : /// The LSN of the last WAL record flushed to disk.
117 : flush_record_lsn: Lsn,
118 :
119 : /// Decoder is required for detecting boundaries of WAL records.
120 : decoder: WalStreamDecoder,
121 :
122 : /// Cached open file for the last segment.
123 : ///
124 : /// If Some(file) is open, then it always:
125 : /// - has ".partial" suffix
126 : /// - points to write_lsn, so no seek is needed for writing
127 : /// - doesn't point to the end of the segment
128 : file: Option<File>,
129 :
130 : /// When false, we have just initialized storage using the LSN from find_end_of_wal().
131 : /// In this case, [`write_lsn`] can be less than actually written WAL on disk. In particular,
132 : /// there can be a case with unexpected .partial file.
133 : ///
134 : /// Imagine the following:
135 : /// - 000000010000000000000001
136 : /// - it was fully written, but the last record is split between 2 segments
137 : /// - after restart, `find_end_of_wal()` returned 0/1FFFFF0, which is in the end of this segment
138 : /// - `write_lsn`, `write_record_lsn` and `flush_record_lsn` were initialized to 0/1FFFFF0
139 : /// - 000000010000000000000002.partial
140 : /// - it has only 1 byte written, which is not enough to make a full WAL record
141 : ///
142 : /// Partial segment 002 has no WAL records, and it will be removed by the next truncate_wal().
143 : /// This flag will be set to true after the first truncate_wal() call.
144 : ///
145 : /// [`write_lsn`]: Self::write_lsn
146 : is_truncated_after_restart: bool,
147 : }
148 :
149 : impl PhysicalStorage {
150 : /// Create new storage. If commit_lsn is not zero, flush_lsn is tried to be restored from
151 : /// the disk. Otherwise, all LSNs are set to zero.
152 0 : pub fn new(
153 0 : ttid: &TenantTimelineId,
154 0 : timeline_dir: Utf8PathBuf,
155 0 : conf: &SafeKeeperConf,
156 0 : state: &TimelinePersistentState,
157 0 : ) -> Result<PhysicalStorage> {
158 0 : let wal_seg_size = state.server.wal_seg_size as usize;
159 :
160 : // Find out where stored WAL ends, starting at commit_lsn which is a
161 : // known recent record boundary (unless we don't have WAL at all).
162 : //
163 : // NB: find_end_of_wal MUST be backwards compatible with the previously
164 : // written WAL. If find_end_of_wal fails to read any WAL written by an
165 : // older version of the code, we could lose data forever.
166 0 : let write_lsn = if state.commit_lsn == Lsn(0) {
167 0 : Lsn(0)
168 : } else {
169 0 : let version = state.server.pg_version / 10000;
170 0 :
171 0 : dispatch_pgversion!(
172 0 : version,
173 0 : pgv::xlog_utils::find_end_of_wal(
174 0 : timeline_dir.as_std_path(),
175 0 : wal_seg_size,
176 0 : state.commit_lsn,
177 0 : )?,
178 0 : bail!("unsupported postgres version: {}", version)
179 : )
180 : };
181 :
182 : // note: this assumes we fsync'ed whole datadir on start.
183 0 : let flush_lsn = write_lsn;
184 0 :
185 0 : debug!(
186 0 : "initialized storage for timeline {}, flush_lsn={}, commit_lsn={}, peer_horizon_lsn={}",
187 : ttid.timeline_id, flush_lsn, state.commit_lsn, state.peer_horizon_lsn,
188 : );
189 0 : if flush_lsn < state.commit_lsn || flush_lsn < state.peer_horizon_lsn {
190 0 : warn!("timeline {} potential data loss: flush_lsn by find_end_of_wal is less than either commit_lsn or peer_horizon_lsn from control file", ttid.timeline_id);
191 0 : }
192 :
193 0 : Ok(PhysicalStorage {
194 0 : metrics: WalStorageMetrics::default(),
195 0 : timeline_dir,
196 0 : conf: conf.clone(),
197 0 : wal_seg_size,
198 0 : pg_version: state.server.pg_version,
199 0 : system_id: state.server.system_id,
200 0 : write_lsn,
201 0 : write_record_lsn: write_lsn,
202 0 : flush_record_lsn: flush_lsn,
203 0 : decoder: WalStreamDecoder::new(write_lsn, state.server.pg_version / 10000),
204 0 : file: None,
205 0 : is_truncated_after_restart: false,
206 0 : })
207 0 : }
208 :
209 : /// Get all known state of the storage.
210 0 : pub fn internal_state(&self) -> (Lsn, Lsn, Lsn, bool) {
211 0 : (
212 0 : self.write_lsn,
213 0 : self.write_record_lsn,
214 0 : self.flush_record_lsn,
215 0 : self.file.is_some(),
216 0 : )
217 0 : }
218 :
219 : /// Call fdatasync if config requires so.
220 0 : async fn fdatasync_file(&mut self, file: &File) -> Result<()> {
221 0 : if !self.conf.no_sync {
222 0 : self.metrics
223 0 : .observe_flush_seconds(time_io_closure(file.sync_data()).await?);
224 0 : }
225 0 : Ok(())
226 0 : }
227 :
228 : /// Open or create WAL segment file. Caller must call seek to the wanted position.
229 : /// Returns `file` and `is_partial`.
230 0 : async fn open_or_create(&mut self, segno: XLogSegNo) -> Result<(File, bool)> {
231 0 : let (wal_file_path, wal_file_partial_path) =
232 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
233 :
234 : // Try to open already completed segment
235 0 : if let Ok(file) = OpenOptions::new().write(true).open(&wal_file_path).await {
236 0 : Ok((file, false))
237 0 : } else if let Ok(file) = OpenOptions::new()
238 0 : .write(true)
239 0 : .open(&wal_file_partial_path)
240 0 : .await
241 : {
242 : // Try to open existing partial file
243 0 : Ok((file, true))
244 : } else {
245 : // Create and fill new partial file
246 : //
247 : // We're using fdatasync during WAL writing, so file size must not
248 : // change; to this end it is filled with zeros here. To avoid using
249 : // half initialized segment, first bake it under tmp filename and
250 : // then rename.
251 0 : let tmp_path = self.timeline_dir.join("waltmp");
252 0 : let mut file = File::create(&tmp_path)
253 0 : .await
254 0 : .with_context(|| format!("Failed to open tmp wal file {:?}", &tmp_path))?;
255 :
256 0 : write_zeroes(&mut file, self.wal_seg_size).await?;
257 :
258 : // Note: this doesn't get into observe_flush_seconds metric. But
259 : // segment init should be separate metric, if any.
260 0 : if let Err(e) =
261 0 : durable_rename(&tmp_path, &wal_file_partial_path, !self.conf.no_sync).await
262 : {
263 : // Probably rename succeeded, but fsync of it failed. Remove
264 : // the file then to avoid using it.
265 0 : remove_file(wal_file_partial_path)
266 0 : .await
267 0 : .or_else(utils::fs_ext::ignore_not_found)?;
268 0 : return Err(e.into());
269 0 : }
270 0 : Ok((file, true))
271 : }
272 0 : }
273 :
274 : /// Write WAL bytes, which are known to be located in a single WAL segment.
275 0 : async fn write_in_segment(&mut self, segno: u64, xlogoff: usize, buf: &[u8]) -> Result<()> {
276 0 : let mut file = if let Some(file) = self.file.take() {
277 0 : file
278 : } else {
279 0 : let (mut file, is_partial) = self.open_or_create(segno).await?;
280 0 : assert!(is_partial, "unexpected write into non-partial segment file");
281 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
282 0 : file
283 : };
284 :
285 0 : file.write_all(buf).await?;
286 : // Note: flush just ensures write above reaches the OS (this is not
287 : // needed in case of sync IO as Write::write there calls directly write
288 : // syscall, but needed in case of async). It does *not* fsyncs the file.
289 0 : file.flush().await?;
290 :
291 0 : if xlogoff + buf.len() == self.wal_seg_size {
292 : // If we reached the end of a WAL segment, flush and close it.
293 0 : self.fdatasync_file(&file).await?;
294 :
295 : // Rename partial file to completed file
296 0 : let (wal_file_path, wal_file_partial_path) =
297 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
298 0 : fs::rename(wal_file_partial_path, wal_file_path).await?;
299 0 : } else {
300 0 : // otherwise, file can be reused later
301 0 : self.file = Some(file);
302 0 : }
303 :
304 0 : Ok(())
305 0 : }
306 :
307 : /// Writes WAL to the segment files, until everything is writed. If some segments
308 : /// are fully written, they are flushed to disk. The last (partial) segment can
309 : /// be flushed separately later.
310 : ///
311 : /// Updates `write_lsn`.
312 0 : async fn write_exact(&mut self, pos: Lsn, mut buf: &[u8]) -> Result<()> {
313 0 : if self.write_lsn != pos {
314 : // need to flush the file before discarding it
315 0 : if let Some(file) = self.file.take() {
316 0 : self.fdatasync_file(&file).await?;
317 0 : }
318 :
319 0 : self.write_lsn = pos;
320 0 : }
321 :
322 0 : while !buf.is_empty() {
323 : // Extract WAL location for this block
324 0 : let xlogoff = self.write_lsn.segment_offset(self.wal_seg_size);
325 0 : let segno = self.write_lsn.segment_number(self.wal_seg_size);
326 :
327 : // If crossing a WAL boundary, only write up until we reach wal segment size.
328 0 : let bytes_write = if xlogoff + buf.len() > self.wal_seg_size {
329 0 : self.wal_seg_size - xlogoff
330 : } else {
331 0 : buf.len()
332 : };
333 :
334 0 : self.write_in_segment(segno, xlogoff, &buf[..bytes_write])
335 0 : .await?;
336 0 : self.write_lsn += bytes_write as u64;
337 0 : buf = &buf[bytes_write..];
338 : }
339 :
340 0 : Ok(())
341 0 : }
342 : }
343 :
344 : impl Storage for PhysicalStorage {
345 : // Last written LSN.
346 0 : fn write_lsn(&self) -> Lsn {
347 0 : self.write_lsn
348 0 : }
349 : /// flush_lsn returns LSN of last durably stored WAL record.
350 0 : fn flush_lsn(&self) -> Lsn {
351 0 : self.flush_record_lsn
352 0 : }
353 :
354 0 : async fn initialize_first_segment(&mut self, init_lsn: Lsn) -> Result<()> {
355 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
356 0 : .with_label_values(&["initialize_first_segment"])
357 0 : .start_timer();
358 0 :
359 0 : let segno = init_lsn.segment_number(self.wal_seg_size);
360 0 : let (mut file, _) = self.open_or_create(segno).await?;
361 0 : let major_pg_version = self.pg_version / 10000;
362 0 : let wal_seg =
363 0 : postgres_ffi::generate_wal_segment(segno, self.system_id, major_pg_version, init_lsn)?;
364 0 : file.seek(SeekFrom::Start(0)).await?;
365 0 : file.write_all(&wal_seg).await?;
366 0 : file.flush().await?;
367 0 : info!("initialized segno {} at lsn {}", segno, init_lsn);
368 : // note: file is *not* fsynced
369 0 : Ok(())
370 0 : }
371 :
372 : /// Write WAL to disk.
373 0 : async fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> Result<()> {
374 0 : // Disallow any non-sequential writes, which can result in gaps or overwrites.
375 0 : // If we need to move the pointer, use truncate_wal() instead.
376 0 : if self.write_lsn > startpos {
377 0 : bail!(
378 0 : "write_wal rewrites WAL written before, write_lsn={}, startpos={}",
379 0 : self.write_lsn,
380 0 : startpos
381 0 : );
382 0 : }
383 0 : if self.write_lsn < startpos && self.write_lsn != Lsn(0) {
384 0 : bail!(
385 0 : "write_wal creates gap in written WAL, write_lsn={}, startpos={}",
386 0 : self.write_lsn,
387 0 : startpos
388 0 : );
389 0 : }
390 :
391 0 : let write_seconds = time_io_closure(self.write_exact(startpos, buf)).await?;
392 : // WAL is written, updating write metrics
393 0 : self.metrics.observe_write_seconds(write_seconds);
394 0 : self.metrics.observe_write_bytes(buf.len());
395 0 :
396 0 : // figure out last record's end lsn for reporting (if we got the
397 0 : // whole record)
398 0 : if self.decoder.available() != startpos {
399 0 : info!(
400 0 : "restart decoder from {} to {}",
401 0 : self.decoder.available(),
402 : startpos,
403 : );
404 0 : let pg_version = self.decoder.pg_version;
405 0 : self.decoder = WalStreamDecoder::new(startpos, pg_version);
406 0 : }
407 0 : self.decoder.feed_bytes(buf);
408 : loop {
409 0 : match self.decoder.poll_decode()? {
410 0 : None => break, // no full record yet
411 0 : Some((lsn, _rec)) => {
412 0 : self.write_record_lsn = lsn;
413 0 : }
414 : }
415 : }
416 :
417 0 : Ok(())
418 0 : }
419 :
420 0 : async fn flush_wal(&mut self) -> Result<()> {
421 0 : if self.flush_record_lsn == self.write_record_lsn {
422 : // no need to do extra flush
423 0 : return Ok(());
424 0 : }
425 :
426 0 : if let Some(unflushed_file) = self.file.take() {
427 0 : self.fdatasync_file(&unflushed_file).await?;
428 0 : self.file = Some(unflushed_file);
429 : } else {
430 : // We have unflushed data (write_lsn != flush_lsn), but no file.
431 : // This should only happen if last file was fully written and flushed,
432 : // but haven't updated flush_lsn yet.
433 0 : if self.write_lsn.segment_offset(self.wal_seg_size) != 0 {
434 0 : bail!(
435 0 : "unexpected unflushed data with no open file, write_lsn={}, flush_lsn={}",
436 0 : self.write_lsn,
437 0 : self.flush_record_lsn
438 0 : );
439 0 : }
440 : }
441 :
442 : // everything is flushed now, let's update flush_lsn
443 0 : self.flush_record_lsn = self.write_record_lsn;
444 0 : Ok(())
445 0 : }
446 :
447 : /// Truncate written WAL by removing all WAL segments after the given LSN.
448 : /// end_pos must point to the end of the WAL record.
449 0 : async fn truncate_wal(&mut self, end_pos: Lsn) -> Result<()> {
450 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
451 0 : .with_label_values(&["truncate_wal"])
452 0 : .start_timer();
453 0 :
454 0 : // Streaming must not create a hole, so truncate cannot be called on
455 0 : // non-written lsn.
456 0 : if self.write_record_lsn != Lsn(0) && end_pos > self.write_record_lsn {
457 0 : bail!(
458 0 : "truncate_wal called on non-written WAL, write_record_lsn={}, end_pos={}",
459 0 : self.write_record_lsn,
460 0 : end_pos
461 0 : );
462 0 : }
463 0 :
464 0 : // Quick exit if nothing to do to avoid writing up to 16 MiB of zeros on
465 0 : // disk (this happens on each connect).
466 0 : if self.is_truncated_after_restart
467 0 : && end_pos == self.write_lsn
468 0 : && end_pos == self.flush_record_lsn
469 : {
470 0 : return Ok(());
471 0 : }
472 :
473 : // Close previously opened file, if any
474 0 : if let Some(unflushed_file) = self.file.take() {
475 0 : self.fdatasync_file(&unflushed_file).await?;
476 0 : }
477 :
478 0 : let xlogoff = end_pos.segment_offset(self.wal_seg_size);
479 0 : let segno = end_pos.segment_number(self.wal_seg_size);
480 0 :
481 0 : // Remove all segments after the given LSN.
482 0 : remove_segments_from_disk(&self.timeline_dir, self.wal_seg_size, |x| x > segno).await?;
483 :
484 0 : let (mut file, is_partial) = self.open_or_create(segno).await?;
485 :
486 : // Fill end with zeroes
487 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
488 0 : write_zeroes(&mut file, self.wal_seg_size - xlogoff).await?;
489 0 : self.fdatasync_file(&file).await?;
490 :
491 0 : if !is_partial {
492 : // Make segment partial once again
493 0 : let (wal_file_path, wal_file_partial_path) =
494 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
495 0 : fs::rename(wal_file_path, wal_file_partial_path).await?;
496 0 : }
497 :
498 : // Update LSNs
499 0 : self.write_lsn = end_pos;
500 0 : self.write_record_lsn = end_pos;
501 0 : self.flush_record_lsn = end_pos;
502 0 : self.is_truncated_after_restart = true;
503 0 : Ok(())
504 0 : }
505 :
506 0 : fn remove_up_to(&self, segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>> {
507 0 : let timeline_dir = self.timeline_dir.clone();
508 0 : let wal_seg_size = self.wal_seg_size;
509 0 : Box::pin(async move {
510 0 : remove_segments_from_disk(&timeline_dir, wal_seg_size, |x| x <= segno_up_to).await
511 0 : })
512 0 : }
513 :
514 0 : fn close(&mut self) {
515 0 : // close happens in destructor
516 0 : let _open_file = self.file.take();
517 0 : }
518 :
519 0 : fn get_metrics(&self) -> WalStorageMetrics {
520 0 : self.metrics.clone()
521 0 : }
522 : }
523 :
524 : /// Remove all WAL segments in timeline_dir that match the given predicate.
525 0 : async fn remove_segments_from_disk(
526 0 : timeline_dir: &Utf8Path,
527 0 : wal_seg_size: usize,
528 0 : remove_predicate: impl Fn(XLogSegNo) -> bool,
529 0 : ) -> Result<()> {
530 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
531 0 : .with_label_values(&["remove_segments_from_disk"])
532 0 : .start_timer();
533 0 :
534 0 : let mut n_removed = 0;
535 0 : let mut min_removed = u64::MAX;
536 0 : let mut max_removed = u64::MIN;
537 :
538 0 : let mut entries = fs::read_dir(timeline_dir).await?;
539 0 : while let Some(entry) = entries.next_entry().await? {
540 0 : let entry_path = entry.path();
541 0 : let fname = entry_path.file_name().unwrap();
542 :
543 0 : if let Some(fname_str) = fname.to_str() {
544 : /* Ignore files that are not XLOG segments */
545 0 : if !IsXLogFileName(fname_str) && !IsPartialXLogFileName(fname_str) {
546 0 : continue;
547 0 : }
548 0 : let (segno, _) = XLogFromFileName(fname_str, wal_seg_size);
549 0 : if remove_predicate(segno) {
550 0 : remove_file(entry_path).await?;
551 0 : n_removed += 1;
552 0 : min_removed = min(min_removed, segno);
553 0 : max_removed = max(max_removed, segno);
554 0 : REMOVED_WAL_SEGMENTS.inc();
555 0 : }
556 0 : }
557 : }
558 :
559 0 : if n_removed > 0 {
560 0 : info!(
561 0 : "removed {} WAL segments [{}; {}]",
562 : n_removed, min_removed, max_removed
563 : );
564 0 : }
565 0 : Ok(())
566 0 : }
567 :
568 : pub struct WalReader {
569 : remote_path: RemotePath,
570 : timeline_dir: Utf8PathBuf,
571 : wal_seg_size: usize,
572 : pos: Lsn,
573 : wal_segment: Option<Pin<Box<dyn AsyncRead + Send + Sync>>>,
574 :
575 : // S3 will be used to read WAL if LSN is not available locally
576 : enable_remote_read: bool,
577 :
578 : // We don't have WAL locally if LSN is less than local_start_lsn
579 : local_start_lsn: Lsn,
580 : // We will respond with zero-ed bytes before this Lsn as long as
581 : // pos is in the same segment as timeline_start_lsn.
582 : timeline_start_lsn: Lsn,
583 : // integer version number of PostgreSQL, e.g. 14; 15; 16
584 : pg_version: u32,
585 : system_id: SystemId,
586 : timeline_start_segment: Option<Bytes>,
587 : }
588 :
589 : impl WalReader {
590 0 : pub fn new(
591 0 : ttid: &TenantTimelineId,
592 0 : timeline_dir: Utf8PathBuf,
593 0 : state: &TimelinePersistentState,
594 0 : start_pos: Lsn,
595 0 : enable_remote_read: bool,
596 0 : ) -> Result<Self> {
597 0 : if state.server.wal_seg_size == 0 || state.local_start_lsn == Lsn(0) {
598 0 : bail!("state uninitialized, no data to read");
599 0 : }
600 0 :
601 0 : // TODO: Upgrade to bail!() once we know this couldn't possibly happen
602 0 : if state.timeline_start_lsn == Lsn(0) {
603 0 : warn!("timeline_start_lsn uninitialized before initializing wal reader");
604 0 : }
605 :
606 0 : if start_pos
607 0 : < state
608 0 : .timeline_start_lsn
609 0 : .segment_lsn(state.server.wal_seg_size as usize)
610 : {
611 0 : bail!(
612 0 : "Requested streaming from {}, which is before the start of the timeline {}, and also doesn't start at the first segment of that timeline",
613 0 : start_pos,
614 0 : state.timeline_start_lsn
615 0 : );
616 0 : }
617 0 :
618 0 : Ok(Self {
619 0 : remote_path: remote_timeline_path(ttid)?,
620 0 : timeline_dir,
621 0 : wal_seg_size: state.server.wal_seg_size as usize,
622 0 : pos: start_pos,
623 0 : wal_segment: None,
624 0 : enable_remote_read,
625 0 : local_start_lsn: state.local_start_lsn,
626 0 : timeline_start_lsn: state.timeline_start_lsn,
627 0 : pg_version: state.server.pg_version / 10000,
628 0 : system_id: state.server.system_id,
629 0 : timeline_start_segment: None,
630 : })
631 0 : }
632 :
633 : /// Read WAL at current position into provided buf, returns number of bytes
634 : /// read. It can be smaller than buf size only if segment boundary is
635 : /// reached.
636 0 : pub async fn read(&mut self, buf: &mut [u8]) -> Result<usize> {
637 0 : // If this timeline is new, we may not have a full segment yet, so
638 0 : // we pad the first bytes of the timeline's first WAL segment with 0s
639 0 : if self.pos < self.timeline_start_lsn {
640 0 : debug_assert_eq!(
641 0 : self.pos.segment_number(self.wal_seg_size),
642 0 : self.timeline_start_lsn.segment_number(self.wal_seg_size)
643 : );
644 :
645 : // All bytes after timeline_start_lsn are in WAL, but those before
646 : // are not, so we manually construct an empty segment for the bytes
647 : // not available in this timeline.
648 0 : if self.timeline_start_segment.is_none() {
649 0 : let it = postgres_ffi::generate_wal_segment(
650 0 : self.timeline_start_lsn.segment_number(self.wal_seg_size),
651 0 : self.system_id,
652 0 : self.pg_version,
653 0 : self.timeline_start_lsn,
654 0 : )?;
655 0 : self.timeline_start_segment = Some(it);
656 0 : }
657 :
658 0 : assert!(self.timeline_start_segment.is_some());
659 0 : let segment = self.timeline_start_segment.take().unwrap();
660 0 :
661 0 : let seg_bytes = &segment[..];
662 0 :
663 0 : // How much of the current segment have we already consumed?
664 0 : let pos_seg_offset = self.pos.segment_offset(self.wal_seg_size);
665 0 :
666 0 : // How many bytes may we consume in total?
667 0 : let tl_start_seg_offset = self.timeline_start_lsn.segment_offset(self.wal_seg_size);
668 0 :
669 0 : debug_assert!(seg_bytes.len() > pos_seg_offset);
670 0 : debug_assert!(seg_bytes.len() > tl_start_seg_offset);
671 :
672 : // Copy as many bytes as possible into the buffer
673 0 : let len = (tl_start_seg_offset - pos_seg_offset).min(buf.len());
674 0 : buf[0..len].copy_from_slice(&seg_bytes[pos_seg_offset..pos_seg_offset + len]);
675 0 :
676 0 : self.pos += len as u64;
677 0 :
678 0 : // If we're done with the segment, we can release it's memory.
679 0 : // However, if we're not yet done, store it so that we don't have to
680 0 : // construct the segment the next time this function is called.
681 0 : if self.pos < self.timeline_start_lsn {
682 0 : self.timeline_start_segment = Some(segment);
683 0 : }
684 :
685 0 : return Ok(len);
686 0 : }
687 :
688 0 : let mut wal_segment = match self.wal_segment.take() {
689 0 : Some(reader) => reader,
690 0 : None => self.open_segment().await?,
691 : };
692 :
693 : // How much to read and send in message? We cannot cross the WAL file
694 : // boundary, and we don't want send more than provided buffer.
695 0 : let xlogoff = self.pos.segment_offset(self.wal_seg_size);
696 0 : let send_size = min(buf.len(), self.wal_seg_size - xlogoff);
697 0 :
698 0 : // Read some data from the file.
699 0 : let buf = &mut buf[0..send_size];
700 0 : let send_size = wal_segment.read_exact(buf).await?;
701 0 : self.pos += send_size as u64;
702 0 :
703 0 : // Decide whether to reuse this file. If we don't set wal_segment here
704 0 : // a new reader will be opened next time.
705 0 : if self.pos.segment_offset(self.wal_seg_size) != 0 {
706 0 : self.wal_segment = Some(wal_segment);
707 0 : }
708 :
709 0 : Ok(send_size)
710 0 : }
711 :
712 : /// Open WAL segment at the current position of the reader.
713 0 : async fn open_segment(&self) -> Result<Pin<Box<dyn AsyncRead + Send + Sync>>> {
714 0 : let xlogoff = self.pos.segment_offset(self.wal_seg_size);
715 0 : let segno = self.pos.segment_number(self.wal_seg_size);
716 0 : let wal_file_name = XLogFileName(PG_TLI, segno, self.wal_seg_size);
717 0 :
718 0 : // Try to open local file, if we may have WAL locally
719 0 : if self.pos >= self.local_start_lsn {
720 0 : let res = open_wal_file(&self.timeline_dir, segno, self.wal_seg_size).await;
721 0 : match res {
722 0 : Ok((mut file, _)) => {
723 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
724 0 : return Ok(Box::pin(file));
725 : }
726 0 : Err(e) => {
727 0 : let is_not_found = e.chain().any(|e| {
728 0 : if let Some(e) = e.downcast_ref::<io::Error>() {
729 0 : e.kind() == io::ErrorKind::NotFound
730 : } else {
731 0 : false
732 : }
733 0 : });
734 0 : if !is_not_found {
735 0 : return Err(e);
736 0 : }
737 : // NotFound is expected, fall through to remote read
738 : }
739 : };
740 0 : }
741 :
742 : // Try to open remote file, if remote reads are enabled
743 0 : if self.enable_remote_read {
744 0 : let remote_wal_file_path = self.remote_path.join(&wal_file_name);
745 0 : return read_object(&remote_wal_file_path, xlogoff as u64).await;
746 0 : }
747 0 :
748 0 : bail!("WAL segment is not found")
749 0 : }
750 : }
751 :
752 : /// Zero block for filling created WAL segments.
753 : const ZERO_BLOCK: &[u8] = &[0u8; XLOG_BLCKSZ];
754 :
755 : /// Helper for filling file with zeroes.
756 0 : async fn write_zeroes(file: &mut File, mut count: usize) -> Result<()> {
757 0 : fail::fail_point!("sk-write-zeroes", |_| {
758 0 : info!("write_zeroes hit failpoint");
759 0 : Err(anyhow::anyhow!("failpoint: sk-write-zeroes"))
760 0 : });
761 :
762 0 : while count >= XLOG_BLCKSZ {
763 0 : file.write_all(ZERO_BLOCK).await?;
764 0 : count -= XLOG_BLCKSZ;
765 : }
766 0 : file.write_all(&ZERO_BLOCK[0..count]).await?;
767 0 : file.flush().await?;
768 0 : Ok(())
769 0 : }
770 :
771 : /// Helper function for opening WAL segment `segno` in `dir`. Returns file and
772 : /// whether it is .partial.
773 0 : pub(crate) async fn open_wal_file(
774 0 : timeline_dir: &Utf8Path,
775 0 : segno: XLogSegNo,
776 0 : wal_seg_size: usize,
777 0 : ) -> Result<(tokio::fs::File, bool)> {
778 0 : let (wal_file_path, wal_file_partial_path) = wal_file_paths(timeline_dir, segno, wal_seg_size);
779 0 :
780 0 : // First try to open the .partial file.
781 0 : let mut partial_path = wal_file_path.to_owned();
782 0 : partial_path.set_extension("partial");
783 0 : if let Ok(opened_file) = tokio::fs::File::open(&wal_file_partial_path).await {
784 0 : return Ok((opened_file, true));
785 0 : }
786 :
787 : // If that failed, try it without the .partial extension.
788 0 : let pf = tokio::fs::File::open(&wal_file_path)
789 0 : .await
790 0 : .with_context(|| format!("failed to open WAL file {:#}", wal_file_path))
791 0 : .map_err(|e| {
792 0 : warn!("{}", e);
793 0 : e
794 0 : })?;
795 :
796 0 : Ok((pf, false))
797 0 : }
798 :
799 : /// Helper returning full path to WAL segment file and its .partial brother.
800 0 : pub fn wal_file_paths(
801 0 : timeline_dir: &Utf8Path,
802 0 : segno: XLogSegNo,
803 0 : wal_seg_size: usize,
804 0 : ) -> (Utf8PathBuf, Utf8PathBuf) {
805 0 : let wal_file_name = XLogFileName(PG_TLI, segno, wal_seg_size);
806 0 : let wal_file_path = timeline_dir.join(wal_file_name.clone());
807 0 : let wal_file_partial_path = timeline_dir.join(wal_file_name + ".partial");
808 0 : (wal_file_path, wal_file_partial_path)
809 0 : }
|
