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 {
190 0 : bail!("timeline {} potential data loss: flush_lsn {} by find_end_of_wal is less than commit_lsn {} from control file", ttid.timeline_id, flush_lsn, state.commit_lsn);
191 0 : }
192 0 : if flush_lsn < state.peer_horizon_lsn {
193 0 : warn!(
194 0 : "timeline {}: flush_lsn {} is less than cfile peer_horizon_lsn {}",
195 : ttid.timeline_id, flush_lsn, state.peer_horizon_lsn
196 : );
197 0 : }
198 :
199 0 : Ok(PhysicalStorage {
200 0 : metrics: WalStorageMetrics::default(),
201 0 : timeline_dir,
202 0 : conf: conf.clone(),
203 0 : wal_seg_size,
204 0 : pg_version: state.server.pg_version,
205 0 : system_id: state.server.system_id,
206 0 : write_lsn,
207 0 : write_record_lsn: write_lsn,
208 0 : flush_record_lsn: flush_lsn,
209 0 : decoder: WalStreamDecoder::new(write_lsn, state.server.pg_version / 10000),
210 0 : file: None,
211 0 : is_truncated_after_restart: false,
212 0 : })
213 0 : }
214 :
215 : /// Get all known state of the storage.
216 0 : pub fn internal_state(&self) -> (Lsn, Lsn, Lsn, bool) {
217 0 : (
218 0 : self.write_lsn,
219 0 : self.write_record_lsn,
220 0 : self.flush_record_lsn,
221 0 : self.file.is_some(),
222 0 : )
223 0 : }
224 :
225 : /// Call fdatasync if config requires so.
226 0 : async fn fdatasync_file(&mut self, file: &File) -> Result<()> {
227 0 : if !self.conf.no_sync {
228 0 : self.metrics
229 0 : .observe_flush_seconds(time_io_closure(file.sync_data()).await?);
230 0 : }
231 0 : Ok(())
232 0 : }
233 :
234 : /// Open or create WAL segment file. Caller must call seek to the wanted position.
235 : /// Returns `file` and `is_partial`.
236 0 : async fn open_or_create(&mut self, segno: XLogSegNo) -> Result<(File, bool)> {
237 0 : let (wal_file_path, wal_file_partial_path) =
238 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
239 :
240 : // Try to open already completed segment
241 0 : if let Ok(file) = OpenOptions::new().write(true).open(&wal_file_path).await {
242 0 : Ok((file, false))
243 0 : } else if let Ok(file) = OpenOptions::new()
244 0 : .write(true)
245 0 : .open(&wal_file_partial_path)
246 0 : .await
247 : {
248 : // Try to open existing partial file
249 0 : Ok((file, true))
250 : } else {
251 : // Create and fill new partial file
252 : //
253 : // We're using fdatasync during WAL writing, so file size must not
254 : // change; to this end it is filled with zeros here. To avoid using
255 : // half initialized segment, first bake it under tmp filename and
256 : // then rename.
257 0 : let tmp_path = self.timeline_dir.join("waltmp");
258 0 : let mut file = File::create(&tmp_path)
259 0 : .await
260 0 : .with_context(|| format!("Failed to open tmp wal file {:?}", &tmp_path))?;
261 :
262 0 : write_zeroes(&mut file, self.wal_seg_size).await?;
263 :
264 : // Note: this doesn't get into observe_flush_seconds metric. But
265 : // segment init should be separate metric, if any.
266 0 : if let Err(e) =
267 0 : durable_rename(&tmp_path, &wal_file_partial_path, !self.conf.no_sync).await
268 : {
269 : // Probably rename succeeded, but fsync of it failed. Remove
270 : // the file then to avoid using it.
271 0 : remove_file(wal_file_partial_path)
272 0 : .await
273 0 : .or_else(utils::fs_ext::ignore_not_found)?;
274 0 : return Err(e.into());
275 0 : }
276 0 : Ok((file, true))
277 : }
278 0 : }
279 :
280 : /// Write WAL bytes, which are known to be located in a single WAL segment.
281 0 : async fn write_in_segment(&mut self, segno: u64, xlogoff: usize, buf: &[u8]) -> Result<()> {
282 0 : let mut file = if let Some(file) = self.file.take() {
283 0 : file
284 : } else {
285 0 : let (mut file, is_partial) = self.open_or_create(segno).await?;
286 0 : assert!(is_partial, "unexpected write into non-partial segment file");
287 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
288 0 : file
289 : };
290 :
291 0 : file.write_all(buf).await?;
292 : // Note: flush just ensures write above reaches the OS (this is not
293 : // needed in case of sync IO as Write::write there calls directly write
294 : // syscall, but needed in case of async). It does *not* fsyncs the file.
295 0 : file.flush().await?;
296 :
297 0 : if xlogoff + buf.len() == self.wal_seg_size {
298 : // If we reached the end of a WAL segment, flush and close it.
299 0 : self.fdatasync_file(&file).await?;
300 :
301 : // Rename partial file to completed file
302 0 : let (wal_file_path, wal_file_partial_path) =
303 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
304 0 : fs::rename(wal_file_partial_path, wal_file_path).await?;
305 0 : } else {
306 0 : // otherwise, file can be reused later
307 0 : self.file = Some(file);
308 0 : }
309 :
310 0 : Ok(())
311 0 : }
312 :
313 : /// Writes WAL to the segment files, until everything is writed. If some segments
314 : /// are fully written, they are flushed to disk. The last (partial) segment can
315 : /// be flushed separately later.
316 : ///
317 : /// Updates `write_lsn`.
318 0 : async fn write_exact(&mut self, pos: Lsn, mut buf: &[u8]) -> Result<()> {
319 0 : if self.write_lsn != pos {
320 : // need to flush the file before discarding it
321 0 : if let Some(file) = self.file.take() {
322 0 : self.fdatasync_file(&file).await?;
323 0 : }
324 :
325 0 : self.write_lsn = pos;
326 0 : }
327 :
328 0 : while !buf.is_empty() {
329 : // Extract WAL location for this block
330 0 : let xlogoff = self.write_lsn.segment_offset(self.wal_seg_size);
331 0 : let segno = self.write_lsn.segment_number(self.wal_seg_size);
332 :
333 : // If crossing a WAL boundary, only write up until we reach wal segment size.
334 0 : let bytes_write = if xlogoff + buf.len() > self.wal_seg_size {
335 0 : self.wal_seg_size - xlogoff
336 : } else {
337 0 : buf.len()
338 : };
339 :
340 0 : self.write_in_segment(segno, xlogoff, &buf[..bytes_write])
341 0 : .await?;
342 0 : self.write_lsn += bytes_write as u64;
343 0 : buf = &buf[bytes_write..];
344 : }
345 :
346 0 : Ok(())
347 0 : }
348 : }
349 :
350 : impl Storage for PhysicalStorage {
351 : // Last written LSN.
352 0 : fn write_lsn(&self) -> Lsn {
353 0 : self.write_lsn
354 0 : }
355 : /// flush_lsn returns LSN of last durably stored WAL record.
356 0 : fn flush_lsn(&self) -> Lsn {
357 0 : self.flush_record_lsn
358 0 : }
359 :
360 0 : async fn initialize_first_segment(&mut self, init_lsn: Lsn) -> Result<()> {
361 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
362 0 : .with_label_values(&["initialize_first_segment"])
363 0 : .start_timer();
364 0 :
365 0 : let segno = init_lsn.segment_number(self.wal_seg_size);
366 0 : let (mut file, _) = self.open_or_create(segno).await?;
367 0 : let major_pg_version = self.pg_version / 10000;
368 0 : let wal_seg =
369 0 : postgres_ffi::generate_wal_segment(segno, self.system_id, major_pg_version, init_lsn)?;
370 0 : file.seek(SeekFrom::Start(0)).await?;
371 0 : file.write_all(&wal_seg).await?;
372 0 : file.flush().await?;
373 0 : info!("initialized segno {} at lsn {}", segno, init_lsn);
374 : // note: file is *not* fsynced
375 0 : Ok(())
376 0 : }
377 :
378 : /// Write WAL to disk.
379 0 : async fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> Result<()> {
380 0 : // Disallow any non-sequential writes, which can result in gaps or overwrites.
381 0 : // If we need to move the pointer, use truncate_wal() instead.
382 0 : if self.write_lsn > startpos {
383 0 : bail!(
384 0 : "write_wal rewrites WAL written before, write_lsn={}, startpos={}",
385 0 : self.write_lsn,
386 0 : startpos
387 0 : );
388 0 : }
389 0 : if self.write_lsn < startpos && self.write_lsn != Lsn(0) {
390 0 : bail!(
391 0 : "write_wal creates gap in written WAL, write_lsn={}, startpos={}",
392 0 : self.write_lsn,
393 0 : startpos
394 0 : );
395 0 : }
396 :
397 0 : let write_seconds = time_io_closure(self.write_exact(startpos, buf)).await?;
398 : // WAL is written, updating write metrics
399 0 : self.metrics.observe_write_seconds(write_seconds);
400 0 : self.metrics.observe_write_bytes(buf.len());
401 0 :
402 0 : // figure out last record's end lsn for reporting (if we got the
403 0 : // whole record)
404 0 : if self.decoder.available() != startpos {
405 0 : info!(
406 0 : "restart decoder from {} to {}",
407 0 : self.decoder.available(),
408 : startpos,
409 : );
410 0 : let pg_version = self.decoder.pg_version;
411 0 : self.decoder = WalStreamDecoder::new(startpos, pg_version);
412 0 : }
413 0 : self.decoder.feed_bytes(buf);
414 : loop {
415 0 : match self.decoder.poll_decode()? {
416 0 : None => break, // no full record yet
417 0 : Some((lsn, _rec)) => {
418 0 : self.write_record_lsn = lsn;
419 0 : }
420 : }
421 : }
422 :
423 0 : Ok(())
424 0 : }
425 :
426 0 : async fn flush_wal(&mut self) -> Result<()> {
427 0 : if self.flush_record_lsn == self.write_record_lsn {
428 : // no need to do extra flush
429 0 : return Ok(());
430 0 : }
431 :
432 0 : if let Some(unflushed_file) = self.file.take() {
433 0 : self.fdatasync_file(&unflushed_file).await?;
434 0 : self.file = Some(unflushed_file);
435 : } else {
436 : // We have unflushed data (write_lsn != flush_lsn), but no file.
437 : // This should only happen if last file was fully written and flushed,
438 : // but haven't updated flush_lsn yet.
439 0 : if self.write_lsn.segment_offset(self.wal_seg_size) != 0 {
440 0 : bail!(
441 0 : "unexpected unflushed data with no open file, write_lsn={}, flush_lsn={}",
442 0 : self.write_lsn,
443 0 : self.flush_record_lsn
444 0 : );
445 0 : }
446 : }
447 :
448 : // everything is flushed now, let's update flush_lsn
449 0 : self.flush_record_lsn = self.write_record_lsn;
450 0 : Ok(())
451 0 : }
452 :
453 : /// Truncate written WAL by removing all WAL segments after the given LSN.
454 : /// end_pos must point to the end of the WAL record.
455 0 : async fn truncate_wal(&mut self, end_pos: Lsn) -> Result<()> {
456 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
457 0 : .with_label_values(&["truncate_wal"])
458 0 : .start_timer();
459 0 :
460 0 : // Streaming must not create a hole, so truncate cannot be called on
461 0 : // non-written lsn.
462 0 : if self.write_record_lsn != Lsn(0) && end_pos > self.write_record_lsn {
463 0 : bail!(
464 0 : "truncate_wal called on non-written WAL, write_record_lsn={}, end_pos={}",
465 0 : self.write_record_lsn,
466 0 : end_pos
467 0 : );
468 0 : }
469 0 :
470 0 : // Quick exit if nothing to do to avoid writing up to 16 MiB of zeros on
471 0 : // disk (this happens on each connect).
472 0 : if self.is_truncated_after_restart
473 0 : && end_pos == self.write_lsn
474 0 : && end_pos == self.flush_record_lsn
475 : {
476 0 : return Ok(());
477 0 : }
478 :
479 : // Close previously opened file, if any
480 0 : if let Some(unflushed_file) = self.file.take() {
481 0 : self.fdatasync_file(&unflushed_file).await?;
482 0 : }
483 :
484 0 : let xlogoff = end_pos.segment_offset(self.wal_seg_size);
485 0 : let segno = end_pos.segment_number(self.wal_seg_size);
486 0 :
487 0 : // Remove all segments after the given LSN.
488 0 : remove_segments_from_disk(&self.timeline_dir, self.wal_seg_size, |x| x > segno).await?;
489 :
490 0 : let (mut file, is_partial) = self.open_or_create(segno).await?;
491 :
492 : // Fill end with zeroes
493 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
494 0 : write_zeroes(&mut file, self.wal_seg_size - xlogoff).await?;
495 0 : self.fdatasync_file(&file).await?;
496 :
497 0 : if !is_partial {
498 : // Make segment partial once again
499 0 : let (wal_file_path, wal_file_partial_path) =
500 0 : wal_file_paths(&self.timeline_dir, segno, self.wal_seg_size);
501 0 : fs::rename(wal_file_path, wal_file_partial_path).await?;
502 0 : }
503 :
504 : // Update LSNs
505 0 : self.write_lsn = end_pos;
506 0 : self.write_record_lsn = end_pos;
507 0 : self.flush_record_lsn = end_pos;
508 0 : self.is_truncated_after_restart = true;
509 0 : Ok(())
510 0 : }
511 :
512 0 : fn remove_up_to(&self, segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>> {
513 0 : let timeline_dir = self.timeline_dir.clone();
514 0 : let wal_seg_size = self.wal_seg_size;
515 0 : Box::pin(async move {
516 0 : remove_segments_from_disk(&timeline_dir, wal_seg_size, |x| x <= segno_up_to).await
517 0 : })
518 0 : }
519 :
520 0 : fn close(&mut self) {
521 0 : // close happens in destructor
522 0 : let _open_file = self.file.take();
523 0 : }
524 :
525 0 : fn get_metrics(&self) -> WalStorageMetrics {
526 0 : self.metrics.clone()
527 0 : }
528 : }
529 :
530 : /// Remove all WAL segments in timeline_dir that match the given predicate.
531 0 : async fn remove_segments_from_disk(
532 0 : timeline_dir: &Utf8Path,
533 0 : wal_seg_size: usize,
534 0 : remove_predicate: impl Fn(XLogSegNo) -> bool,
535 0 : ) -> Result<()> {
536 0 : let _timer = WAL_STORAGE_OPERATION_SECONDS
537 0 : .with_label_values(&["remove_segments_from_disk"])
538 0 : .start_timer();
539 0 :
540 0 : let mut n_removed = 0;
541 0 : let mut min_removed = u64::MAX;
542 0 : let mut max_removed = u64::MIN;
543 :
544 0 : let mut entries = fs::read_dir(timeline_dir).await?;
545 0 : while let Some(entry) = entries.next_entry().await? {
546 0 : let entry_path = entry.path();
547 0 : let fname = entry_path.file_name().unwrap();
548 0 : /* Ignore files that are not XLOG segments */
549 0 : if !IsXLogFileName(fname) && !IsPartialXLogFileName(fname) {
550 0 : continue;
551 0 : }
552 0 : let (segno, _) = XLogFromFileName(fname, wal_seg_size)?;
553 0 : if remove_predicate(segno) {
554 0 : remove_file(entry_path).await?;
555 0 : n_removed += 1;
556 0 : min_removed = min(min_removed, segno);
557 0 : max_removed = max(max_removed, segno);
558 0 : REMOVED_WAL_SEGMENTS.inc();
559 0 : }
560 : }
561 :
562 0 : if n_removed > 0 {
563 0 : info!(
564 0 : "removed {} WAL segments [{}; {}]",
565 : n_removed, min_removed, max_removed
566 : );
567 0 : }
568 0 : Ok(())
569 0 : }
570 :
571 : pub struct WalReader {
572 : remote_path: RemotePath,
573 : timeline_dir: Utf8PathBuf,
574 : wal_seg_size: usize,
575 : pos: Lsn,
576 : wal_segment: Option<Pin<Box<dyn AsyncRead + Send + Sync>>>,
577 :
578 : // S3 will be used to read WAL if LSN is not available locally
579 : enable_remote_read: bool,
580 :
581 : // We don't have WAL locally if LSN is less than local_start_lsn
582 : local_start_lsn: Lsn,
583 : // We will respond with zero-ed bytes before this Lsn as long as
584 : // pos is in the same segment as timeline_start_lsn.
585 : timeline_start_lsn: Lsn,
586 : // integer version number of PostgreSQL, e.g. 14; 15; 16
587 : pg_version: u32,
588 : system_id: SystemId,
589 : timeline_start_segment: Option<Bytes>,
590 : }
591 :
592 : impl WalReader {
593 0 : pub fn new(
594 0 : ttid: &TenantTimelineId,
595 0 : timeline_dir: Utf8PathBuf,
596 0 : state: &TimelinePersistentState,
597 0 : start_pos: Lsn,
598 0 : enable_remote_read: bool,
599 0 : ) -> Result<Self> {
600 0 : if state.server.wal_seg_size == 0 || state.local_start_lsn == Lsn(0) {
601 0 : bail!("state uninitialized, no data to read");
602 0 : }
603 0 :
604 0 : // TODO: Upgrade to bail!() once we know this couldn't possibly happen
605 0 : if state.timeline_start_lsn == Lsn(0) {
606 0 : warn!("timeline_start_lsn uninitialized before initializing wal reader");
607 0 : }
608 :
609 0 : if start_pos
610 0 : < state
611 0 : .timeline_start_lsn
612 0 : .segment_lsn(state.server.wal_seg_size as usize)
613 : {
614 0 : bail!(
615 0 : "Requested streaming from {}, which is before the start of the timeline {}, and also doesn't start at the first segment of that timeline",
616 0 : start_pos,
617 0 : state.timeline_start_lsn
618 0 : );
619 0 : }
620 0 :
621 0 : Ok(Self {
622 0 : remote_path: remote_timeline_path(ttid)?,
623 0 : timeline_dir,
624 0 : wal_seg_size: state.server.wal_seg_size as usize,
625 0 : pos: start_pos,
626 0 : wal_segment: None,
627 0 : enable_remote_read,
628 0 : local_start_lsn: state.local_start_lsn,
629 0 : timeline_start_lsn: state.timeline_start_lsn,
630 0 : pg_version: state.server.pg_version / 10000,
631 0 : system_id: state.server.system_id,
632 0 : timeline_start_segment: None,
633 : })
634 0 : }
635 :
636 : /// Read WAL at current position into provided buf, returns number of bytes
637 : /// read. It can be smaller than buf size only if segment boundary is
638 : /// reached.
639 0 : pub async fn read(&mut self, buf: &mut [u8]) -> Result<usize> {
640 0 : // If this timeline is new, we may not have a full segment yet, so
641 0 : // we pad the first bytes of the timeline's first WAL segment with 0s
642 0 : if self.pos < self.timeline_start_lsn {
643 0 : debug_assert_eq!(
644 0 : self.pos.segment_number(self.wal_seg_size),
645 0 : self.timeline_start_lsn.segment_number(self.wal_seg_size)
646 : );
647 :
648 : // All bytes after timeline_start_lsn are in WAL, but those before
649 : // are not, so we manually construct an empty segment for the bytes
650 : // not available in this timeline.
651 0 : if self.timeline_start_segment.is_none() {
652 0 : let it = postgres_ffi::generate_wal_segment(
653 0 : self.timeline_start_lsn.segment_number(self.wal_seg_size),
654 0 : self.system_id,
655 0 : self.pg_version,
656 0 : self.timeline_start_lsn,
657 0 : )?;
658 0 : self.timeline_start_segment = Some(it);
659 0 : }
660 :
661 0 : assert!(self.timeline_start_segment.is_some());
662 0 : let segment = self.timeline_start_segment.take().unwrap();
663 0 :
664 0 : let seg_bytes = &segment[..];
665 0 :
666 0 : // How much of the current segment have we already consumed?
667 0 : let pos_seg_offset = self.pos.segment_offset(self.wal_seg_size);
668 0 :
669 0 : // How many bytes may we consume in total?
670 0 : let tl_start_seg_offset = self.timeline_start_lsn.segment_offset(self.wal_seg_size);
671 0 :
672 0 : debug_assert!(seg_bytes.len() > pos_seg_offset);
673 0 : debug_assert!(seg_bytes.len() > tl_start_seg_offset);
674 :
675 : // Copy as many bytes as possible into the buffer
676 0 : let len = (tl_start_seg_offset - pos_seg_offset).min(buf.len());
677 0 : buf[0..len].copy_from_slice(&seg_bytes[pos_seg_offset..pos_seg_offset + len]);
678 0 :
679 0 : self.pos += len as u64;
680 0 :
681 0 : // If we're done with the segment, we can release it's memory.
682 0 : // However, if we're not yet done, store it so that we don't have to
683 0 : // construct the segment the next time this function is called.
684 0 : if self.pos < self.timeline_start_lsn {
685 0 : self.timeline_start_segment = Some(segment);
686 0 : }
687 :
688 0 : return Ok(len);
689 0 : }
690 :
691 0 : let mut wal_segment = match self.wal_segment.take() {
692 0 : Some(reader) => reader,
693 0 : None => self.open_segment().await?,
694 : };
695 :
696 : // How much to read and send in message? We cannot cross the WAL file
697 : // boundary, and we don't want send more than provided buffer.
698 0 : let xlogoff = self.pos.segment_offset(self.wal_seg_size);
699 0 : let send_size = min(buf.len(), self.wal_seg_size - xlogoff);
700 0 :
701 0 : // Read some data from the file.
702 0 : let buf = &mut buf[0..send_size];
703 0 : let send_size = wal_segment.read_exact(buf).await?;
704 0 : self.pos += send_size as u64;
705 0 :
706 0 : // Decide whether to reuse this file. If we don't set wal_segment here
707 0 : // a new reader will be opened next time.
708 0 : if self.pos.segment_offset(self.wal_seg_size) != 0 {
709 0 : self.wal_segment = Some(wal_segment);
710 0 : }
711 :
712 0 : Ok(send_size)
713 0 : }
714 :
715 : /// Open WAL segment at the current position of the reader.
716 0 : async fn open_segment(&self) -> Result<Pin<Box<dyn AsyncRead + Send + Sync>>> {
717 0 : let xlogoff = self.pos.segment_offset(self.wal_seg_size);
718 0 : let segno = self.pos.segment_number(self.wal_seg_size);
719 0 : let wal_file_name = XLogFileName(PG_TLI, segno, self.wal_seg_size);
720 0 :
721 0 : // Try to open local file, if we may have WAL locally
722 0 : if self.pos >= self.local_start_lsn {
723 0 : let res = open_wal_file(&self.timeline_dir, segno, self.wal_seg_size).await;
724 0 : match res {
725 0 : Ok((mut file, _)) => {
726 0 : file.seek(SeekFrom::Start(xlogoff as u64)).await?;
727 0 : return Ok(Box::pin(file));
728 : }
729 0 : Err(e) => {
730 0 : let is_not_found = e.chain().any(|e| {
731 0 : if let Some(e) = e.downcast_ref::<io::Error>() {
732 0 : e.kind() == io::ErrorKind::NotFound
733 : } else {
734 0 : false
735 : }
736 0 : });
737 0 : if !is_not_found {
738 0 : return Err(e);
739 0 : }
740 : // NotFound is expected, fall through to remote read
741 : }
742 : };
743 0 : }
744 :
745 : // Try to open remote file, if remote reads are enabled
746 0 : if self.enable_remote_read {
747 0 : let remote_wal_file_path = self.remote_path.join(&wal_file_name);
748 0 : return read_object(&remote_wal_file_path, xlogoff as u64).await;
749 0 : }
750 0 :
751 0 : bail!("WAL segment is not found")
752 0 : }
753 : }
754 :
755 : /// Zero block for filling created WAL segments.
756 : const ZERO_BLOCK: &[u8] = &[0u8; XLOG_BLCKSZ];
757 :
758 : /// Helper for filling file with zeroes.
759 0 : async fn write_zeroes(file: &mut File, mut count: usize) -> Result<()> {
760 0 : fail::fail_point!("sk-write-zeroes", |_| {
761 0 : info!("write_zeroes hit failpoint");
762 0 : Err(anyhow::anyhow!("failpoint: sk-write-zeroes"))
763 0 : });
764 :
765 0 : while count >= XLOG_BLCKSZ {
766 0 : file.write_all(ZERO_BLOCK).await?;
767 0 : count -= XLOG_BLCKSZ;
768 : }
769 0 : file.write_all(&ZERO_BLOCK[0..count]).await?;
770 0 : file.flush().await?;
771 0 : Ok(())
772 0 : }
773 :
774 : /// Helper function for opening WAL segment `segno` in `dir`. Returns file and
775 : /// whether it is .partial.
776 0 : pub(crate) async fn open_wal_file(
777 0 : timeline_dir: &Utf8Path,
778 0 : segno: XLogSegNo,
779 0 : wal_seg_size: usize,
780 0 : ) -> Result<(tokio::fs::File, bool)> {
781 0 : let (wal_file_path, wal_file_partial_path) = wal_file_paths(timeline_dir, segno, wal_seg_size);
782 0 :
783 0 : // First try to open the .partial file.
784 0 : let mut partial_path = wal_file_path.to_owned();
785 0 : partial_path.set_extension("partial");
786 0 : if let Ok(opened_file) = tokio::fs::File::open(&wal_file_partial_path).await {
787 0 : return Ok((opened_file, true));
788 0 : }
789 :
790 : // If that failed, try it without the .partial extension.
791 0 : let pf = tokio::fs::File::open(&wal_file_path)
792 0 : .await
793 0 : .with_context(|| format!("failed to open WAL file {:#}", wal_file_path))
794 0 : .map_err(|e| {
795 0 : warn!("{}", e);
796 0 : e
797 0 : })?;
798 :
799 0 : Ok((pf, false))
800 0 : }
801 :
802 : /// Helper returning full path to WAL segment file and its .partial brother.
803 0 : pub fn wal_file_paths(
804 0 : timeline_dir: &Utf8Path,
805 0 : segno: XLogSegNo,
806 0 : wal_seg_size: usize,
807 0 : ) -> (Utf8PathBuf, Utf8PathBuf) {
808 0 : let wal_file_name = XLogFileName(PG_TLI, segno, wal_seg_size);
809 0 : let wal_file_path = timeline_dir.join(wal_file_name.clone());
810 0 : let wal_file_partial_path = timeline_dir.join(wal_file_name + ".partial");
811 0 : (wal_file_path, wal_file_partial_path)
812 0 : }
|
