Line data Source code
1 : //! Code related to evicting WAL files to remote storage. The actual upload is done by the
2 : //! partial WAL backup code. This file has code to delete and re-download WAL files,
3 : //! cross-validate with partial WAL backup if local file is still present.
4 :
5 : use anyhow::Context;
6 : use camino::Utf8PathBuf;
7 : use remote_storage::RemotePath;
8 : use tokio::{
9 : fs::File,
10 : io::{AsyncRead, AsyncWriteExt},
11 : };
12 : use tracing::{debug, info, instrument, warn};
13 : use utils::crashsafe::durable_rename;
14 :
15 : use crate::{
16 : metrics::{EvictionEvent, EVICTION_EVENTS_COMPLETED, EVICTION_EVENTS_STARTED},
17 : rate_limit::rand_duration,
18 : timeline_manager::{Manager, StateSnapshot},
19 : wal_backup,
20 : wal_backup_partial::{self, PartialRemoteSegment},
21 : wal_storage::wal_file_paths,
22 : };
23 :
24 : impl Manager {
25 : /// Returns true if the timeline is ready for eviction.
26 : /// Current criteria:
27 : /// - no active tasks
28 : /// - control file is flushed (no next event scheduled)
29 : /// - no WAL residence guards
30 : /// - no pushes to the broker
31 : /// - partial WAL backup is uploaded
32 0 : pub(crate) fn ready_for_eviction(
33 0 : &self,
34 0 : next_event: &Option<tokio::time::Instant>,
35 0 : state: &StateSnapshot,
36 0 : ) -> bool {
37 0 : self.backup_task.is_none()
38 0 : && self.recovery_task.is_none()
39 0 : && self.wal_removal_task.is_none()
40 0 : && self.partial_backup_task.is_none()
41 0 : && self.partial_backup_uploaded.is_some()
42 0 : && next_event.is_none()
43 0 : && self.access_service.is_empty()
44 0 : && !self.tli_broker_active.get()
45 0 : && !wal_backup_partial::needs_uploading(state, &self.partial_backup_uploaded)
46 0 : && self
47 0 : .partial_backup_uploaded
48 0 : .as_ref()
49 0 : .unwrap()
50 0 : .flush_lsn
51 0 : .segment_number(self.wal_seg_size)
52 0 : == self.last_removed_segno + 1
53 0 : }
54 :
55 : /// Evict the timeline to remote storage.
56 0 : #[instrument(name = "evict_timeline", skip_all)]
57 : pub(crate) async fn evict_timeline(&mut self) {
58 : assert!(!self.is_offloaded);
59 : let partial_backup_uploaded = match &self.partial_backup_uploaded {
60 : Some(p) => p.clone(),
61 : None => {
62 : warn!("no partial backup uploaded, skipping eviction");
63 : return;
64 : }
65 : };
66 :
67 : info!("starting eviction, using {:?}", partial_backup_uploaded);
68 :
69 : EVICTION_EVENTS_STARTED
70 : .with_label_values(&[EvictionEvent::Evict.into()])
71 : .inc();
72 0 : let _guard = scopeguard::guard((), |_| {
73 0 : EVICTION_EVENTS_COMPLETED
74 0 : .with_label_values(&[EvictionEvent::Evict.into()])
75 0 : .inc();
76 0 : });
77 :
78 : if let Err(e) = do_eviction(self, &partial_backup_uploaded).await {
79 : warn!("failed to evict timeline: {:?}", e);
80 : return;
81 : }
82 :
83 : info!("successfully evicted timeline");
84 : }
85 :
86 : /// Restore evicted timeline from remote storage.
87 0 : #[instrument(name = "unevict_timeline", skip_all)]
88 : pub(crate) async fn unevict_timeline(&mut self) {
89 : assert!(self.is_offloaded);
90 : let partial_backup_uploaded = match &self.partial_backup_uploaded {
91 : Some(p) => p.clone(),
92 : None => {
93 : warn!("no partial backup uploaded, cannot unevict");
94 : return;
95 : }
96 : };
97 :
98 : info!("starting uneviction, using {:?}", partial_backup_uploaded);
99 :
100 : EVICTION_EVENTS_STARTED
101 : .with_label_values(&[EvictionEvent::Restore.into()])
102 : .inc();
103 0 : let _guard = scopeguard::guard((), |_| {
104 0 : EVICTION_EVENTS_COMPLETED
105 0 : .with_label_values(&[EvictionEvent::Restore.into()])
106 0 : .inc();
107 0 : });
108 :
109 : if let Err(e) = do_uneviction(self, &partial_backup_uploaded).await {
110 : warn!("failed to unevict timeline: {:?}", e);
111 : return;
112 : }
113 :
114 : self.evict_not_before =
115 : tokio::time::Instant::now() + rand_duration(&self.conf.eviction_min_resident);
116 :
117 : info!("successfully restored evicted timeline");
118 : }
119 : }
120 :
121 : /// Ensure that content matches the remote partial backup, if local segment exists.
122 : /// Then change state in control file and in-memory. If `delete_offloaded_wal` is set,
123 : /// delete the local segment.
124 0 : async fn do_eviction(mgr: &mut Manager, partial: &PartialRemoteSegment) -> anyhow::Result<()> {
125 0 : compare_local_segment_with_remote(mgr, partial).await?;
126 :
127 0 : mgr.tli.switch_to_offloaded(partial).await?;
128 : // switch manager state as soon as possible
129 0 : mgr.is_offloaded = true;
130 0 :
131 0 : if mgr.conf.delete_offloaded_wal {
132 0 : delete_local_segment(mgr, partial).await?;
133 0 : }
134 :
135 0 : Ok(())
136 0 : }
137 :
138 : /// Ensure that content matches the remote partial backup, if local segment exists.
139 : /// Then download segment to local disk and change state in control file and in-memory.
140 0 : async fn do_uneviction(mgr: &mut Manager, partial: &PartialRemoteSegment) -> anyhow::Result<()> {
141 0 : // if the local segment is present, validate it
142 0 : compare_local_segment_with_remote(mgr, partial).await?;
143 :
144 : // atomically download the partial segment
145 0 : redownload_partial_segment(mgr, partial).await?;
146 :
147 0 : mgr.tli.switch_to_present().await?;
148 : // switch manager state as soon as possible
149 0 : mgr.is_offloaded = false;
150 0 :
151 0 : Ok(())
152 0 : }
153 :
154 : /// Delete local WAL segment.
155 0 : async fn delete_local_segment(mgr: &Manager, partial: &PartialRemoteSegment) -> anyhow::Result<()> {
156 0 : let local_path = local_segment_path(mgr, partial);
157 0 :
158 0 : info!("deleting WAL file to evict: {}", local_path);
159 0 : tokio::fs::remove_file(&local_path).await?;
160 0 : Ok(())
161 0 : }
162 :
163 : /// Redownload partial segment from remote storage.
164 : /// The segment is downloaded to a temporary file and then renamed to the final path.
165 0 : async fn redownload_partial_segment(
166 0 : mgr: &Manager,
167 0 : partial: &PartialRemoteSegment,
168 0 : ) -> anyhow::Result<()> {
169 0 : let tmp_file = mgr.tli.timeline_dir().join("remote_partial.tmp");
170 0 : let remote_segfile = remote_segment_path(mgr, partial)?;
171 :
172 0 : debug!(
173 0 : "redownloading partial segment: {} -> {}",
174 : remote_segfile, tmp_file
175 : );
176 :
177 0 : let mut reader = wal_backup::read_object(&remote_segfile, 0).await?;
178 0 : let mut file = File::create(&tmp_file).await?;
179 :
180 0 : let actual_len = tokio::io::copy(&mut reader, &mut file).await?;
181 0 : let expected_len = partial.flush_lsn.segment_offset(mgr.wal_seg_size);
182 0 :
183 0 : if actual_len != expected_len as u64 {
184 0 : anyhow::bail!(
185 0 : "partial downloaded {} bytes, expected {}",
186 0 : actual_len,
187 0 : expected_len
188 0 : );
189 0 : }
190 0 :
191 0 : if actual_len > mgr.wal_seg_size as u64 {
192 0 : anyhow::bail!(
193 0 : "remote segment is too long: {} bytes, expected {}",
194 0 : actual_len,
195 0 : mgr.wal_seg_size
196 0 : );
197 0 : }
198 0 : file.set_len(mgr.wal_seg_size as u64).await?;
199 0 : file.flush().await?;
200 :
201 0 : let final_path = local_segment_path(mgr, partial);
202 0 : info!("downloaded {actual_len} bytes, renaming to {final_path}");
203 0 : if let Err(e) = durable_rename(&tmp_file, &final_path, !mgr.conf.no_sync).await {
204 : // Probably rename succeeded, but fsync of it failed. Remove
205 : // the file then to avoid using it.
206 0 : tokio::fs::remove_file(tmp_file)
207 0 : .await
208 0 : .or_else(utils::fs_ext::ignore_not_found)?;
209 0 : return Err(e.into());
210 0 : }
211 0 :
212 0 : Ok(())
213 0 : }
214 :
215 : /// Compare local WAL segment with partial WAL backup in remote storage.
216 : /// If the local segment is not present, the function does nothing.
217 : /// If the local segment is present, it compares the local segment with the remote one.
218 0 : async fn compare_local_segment_with_remote(
219 0 : mgr: &Manager,
220 0 : partial: &PartialRemoteSegment,
221 0 : ) -> anyhow::Result<()> {
222 0 : let local_path = local_segment_path(mgr, partial);
223 0 :
224 0 : match File::open(&local_path).await {
225 0 : Ok(mut local_file) => do_validation(mgr, &mut local_file, mgr.wal_seg_size, partial)
226 0 : .await
227 0 : .context("validation failed"),
228 : Err(_) => {
229 0 : info!(
230 0 : "local WAL file {} is not present, skipping validation",
231 : local_path
232 : );
233 0 : Ok(())
234 : }
235 : }
236 0 : }
237 :
238 : /// Compare opened local WAL segment with partial WAL backup in remote storage.
239 : /// Validate full content of both files.
240 0 : async fn do_validation(
241 0 : mgr: &Manager,
242 0 : file: &mut File,
243 0 : wal_seg_size: usize,
244 0 : partial: &PartialRemoteSegment,
245 0 : ) -> anyhow::Result<()> {
246 0 : let local_size = file.metadata().await?.len() as usize;
247 0 : if local_size != wal_seg_size {
248 0 : anyhow::bail!(
249 0 : "local segment size is invalid: found {}, expected {}",
250 0 : local_size,
251 0 : wal_seg_size
252 0 : );
253 0 : }
254 :
255 0 : let remote_segfile = remote_segment_path(mgr, partial)?;
256 0 : let mut remote_reader: std::pin::Pin<Box<dyn AsyncRead + Send + Sync>> =
257 0 : wal_backup::read_object(&remote_segfile, 0).await?;
258 :
259 : // remote segment should have bytes excatly up to `flush_lsn`
260 0 : let expected_remote_size = partial.flush_lsn.segment_offset(mgr.wal_seg_size);
261 0 : // let's compare the first `expected_remote_size` bytes
262 0 : compare_n_bytes(&mut remote_reader, file, expected_remote_size).await?;
263 : // and check that the remote segment ends here
264 0 : check_end(&mut remote_reader).await?;
265 :
266 : // if local segment is longer, the rest should be zeroes
267 0 : read_n_zeroes(file, mgr.wal_seg_size - expected_remote_size).await?;
268 : // and check that the local segment ends here
269 0 : check_end(file).await?;
270 :
271 0 : Ok(())
272 0 : }
273 :
274 0 : fn local_segment_path(mgr: &Manager, partial: &PartialRemoteSegment) -> Utf8PathBuf {
275 0 : let flush_lsn = partial.flush_lsn;
276 0 : let segno = flush_lsn.segment_number(mgr.wal_seg_size);
277 0 : let (_, local_partial_segfile) =
278 0 : wal_file_paths(mgr.tli.timeline_dir(), segno, mgr.wal_seg_size);
279 0 : local_partial_segfile
280 0 : }
281 :
282 0 : fn remote_segment_path(
283 0 : mgr: &Manager,
284 0 : partial: &PartialRemoteSegment,
285 0 : ) -> anyhow::Result<RemotePath> {
286 0 : let remote_timeline_path = wal_backup::remote_timeline_path(&mgr.tli.ttid)?;
287 0 : Ok(partial.remote_path(&remote_timeline_path))
288 0 : }
289 :
290 : /// Compare first `n` bytes of two readers. If the bytes differ, return an error.
291 : /// If the readers are shorter than `n`, return an error.
292 0 : async fn compare_n_bytes<R1, R2>(reader1: &mut R1, reader2: &mut R2, n: usize) -> anyhow::Result<()>
293 0 : where
294 0 : R1: AsyncRead + Unpin,
295 0 : R2: AsyncRead + Unpin,
296 0 : {
297 0 : use tokio::io::AsyncReadExt;
298 0 :
299 0 : const BUF_SIZE: usize = 32 * 1024;
300 0 :
301 0 : let mut buffer1 = vec![0u8; BUF_SIZE];
302 0 : let mut buffer2 = vec![0u8; BUF_SIZE];
303 0 :
304 0 : let mut offset = 0;
305 :
306 0 : while offset < n {
307 0 : let bytes_to_read = std::cmp::min(BUF_SIZE, n - offset);
308 :
309 0 : let bytes_read1 = reader1
310 0 : .read(&mut buffer1[..bytes_to_read])
311 0 : .await
312 0 : .with_context(|| format!("failed to read from reader1 at offset {}", offset))?;
313 0 : if bytes_read1 == 0 {
314 0 : anyhow::bail!("unexpected EOF from reader1 at offset {}", offset);
315 0 : }
316 :
317 0 : let bytes_read2 = reader2
318 0 : .read_exact(&mut buffer2[..bytes_read1])
319 0 : .await
320 0 : .with_context(|| {
321 0 : format!(
322 0 : "failed to read {} bytes from reader2 at offset {}",
323 0 : bytes_read1, offset
324 0 : )
325 0 : })?;
326 0 : assert!(bytes_read2 == bytes_read1);
327 :
328 0 : if buffer1[..bytes_read1] != buffer2[..bytes_read2] {
329 0 : let diff_offset = buffer1[..bytes_read1]
330 0 : .iter()
331 0 : .zip(buffer2[..bytes_read2].iter())
332 0 : .position(|(a, b)| a != b)
333 0 : .expect("mismatched buffers, but no difference found");
334 0 : anyhow::bail!("mismatch at offset {}", offset + diff_offset);
335 0 : }
336 0 :
337 0 : offset += bytes_read1;
338 : }
339 :
340 0 : Ok(())
341 0 : }
342 :
343 0 : async fn check_end<R>(mut reader: R) -> anyhow::Result<()>
344 0 : where
345 0 : R: AsyncRead + Unpin,
346 0 : {
347 0 : use tokio::io::AsyncReadExt;
348 0 :
349 0 : let mut buffer = [0u8; 1];
350 0 : let bytes_read = reader.read(&mut buffer).await?;
351 0 : if bytes_read != 0 {
352 0 : anyhow::bail!("expected EOF, found bytes");
353 0 : }
354 0 : Ok(())
355 0 : }
356 :
357 0 : async fn read_n_zeroes<R>(reader: &mut R, n: usize) -> anyhow::Result<()>
358 0 : where
359 0 : R: AsyncRead + Unpin,
360 0 : {
361 0 : use tokio::io::AsyncReadExt;
362 0 :
363 0 : const BUF_SIZE: usize = 32 * 1024;
364 0 : let mut buffer = vec![0u8; BUF_SIZE];
365 0 : let mut offset = 0;
366 :
367 0 : while offset < n {
368 0 : let bytes_to_read = std::cmp::min(BUF_SIZE, n - offset);
369 :
370 0 : let bytes_read = reader
371 0 : .read(&mut buffer[..bytes_to_read])
372 0 : .await
373 0 : .context("expected zeroes, got read error")?;
374 0 : if bytes_read == 0 {
375 0 : anyhow::bail!("expected zeroes, got EOF");
376 0 : }
377 0 :
378 0 : if buffer[..bytes_read].iter().all(|&b| b == 0) {
379 0 : offset += bytes_read;
380 0 : } else {
381 0 : anyhow::bail!("non-zero byte found");
382 : }
383 : }
384 :
385 0 : Ok(())
386 0 : }
|
