Line data Source code
1 : //!
2 : //! Generate a tarball with files needed to bootstrap ComputeNode.
3 : //!
4 : //! TODO: this module has nothing to do with PostgreSQL pg_basebackup.
5 : //! It could use a better name.
6 : //!
7 : //! Stateless Postgres compute node is launched by sending a tarball
8 : //! which contains non-relational data (multixacts, clog, filenodemaps, twophase files),
9 : //! generated pg_control and dummy segment of WAL.
10 : //! This module is responsible for creation of such tarball
11 : //! from data stored in object storage.
12 : //!
13 : use std::fmt::Write as FmtWrite;
14 : use std::time::{Instant, SystemTime};
15 :
16 : use anyhow::{Context, anyhow};
17 : use async_compression::tokio::write::GzipEncoder;
18 : use bytes::{BufMut, Bytes, BytesMut};
19 : use fail::fail_point;
20 : use pageserver_api::key::{Key, rel_block_to_key};
21 : use pageserver_api::reltag::{RelTag, SlruKind};
22 : use postgres_ffi::pg_constants::{PG_HBA, PGDATA_SPECIAL_FILES};
23 : use postgres_ffi::{
24 : BLCKSZ, PG_TLI, PgMajorVersion, RELSEG_SIZE, WAL_SEGMENT_SIZE, XLogFileName,
25 : dispatch_pgversion, pg_constants,
26 : };
27 : use postgres_ffi_types::constants::{DEFAULTTABLESPACE_OID, GLOBALTABLESPACE_OID};
28 : use postgres_ffi_types::forknum::{INIT_FORKNUM, MAIN_FORKNUM};
29 : use tokio::io::{self, AsyncWrite, AsyncWriteExt as _};
30 : use tokio_tar::{Builder, EntryType, Header};
31 : use tracing::*;
32 : use utils::lsn::Lsn;
33 :
34 : use crate::context::RequestContext;
35 : use crate::pgdatadir_mapping::Version;
36 : use crate::tenant::storage_layer::IoConcurrency;
37 : use crate::tenant::timeline::{GetVectoredError, VersionedKeySpaceQuery};
38 : use crate::tenant::{PageReconstructError, Timeline};
39 :
40 : #[derive(Debug, thiserror::Error)]
41 : pub enum BasebackupError {
42 : #[error("basebackup pageserver error {0:#}")]
43 : Server(#[from] anyhow::Error),
44 : #[error("basebackup client error {0:#} when {1}")]
45 : Client(#[source] io::Error, &'static str),
46 : #[error("basebackup during shutdown")]
47 : Shutdown,
48 : }
49 :
50 : impl From<PageReconstructError> for BasebackupError {
51 0 : fn from(value: PageReconstructError) -> Self {
52 0 : match value {
53 0 : PageReconstructError::Cancelled => BasebackupError::Shutdown,
54 0 : err => BasebackupError::Server(err.into()),
55 : }
56 0 : }
57 : }
58 :
59 : impl From<GetVectoredError> for BasebackupError {
60 0 : fn from(value: GetVectoredError) -> Self {
61 0 : match value {
62 0 : GetVectoredError::Cancelled => BasebackupError::Shutdown,
63 0 : err => BasebackupError::Server(err.into()),
64 : }
65 0 : }
66 : }
67 :
68 : impl From<BasebackupError> for postgres_backend::QueryError {
69 0 : fn from(err: BasebackupError) -> Self {
70 : use postgres_backend::QueryError;
71 : use pq_proto::framed::ConnectionError;
72 0 : match err {
73 0 : BasebackupError::Client(err, _) => QueryError::Disconnected(ConnectionError::Io(err)),
74 0 : BasebackupError::Server(err) => QueryError::Other(err),
75 0 : BasebackupError::Shutdown => QueryError::Shutdown,
76 : }
77 0 : }
78 : }
79 :
80 : impl From<BasebackupError> for tonic::Status {
81 0 : fn from(err: BasebackupError) -> Self {
82 : use tonic::Code;
83 0 : let code = match &err {
84 0 : BasebackupError::Client(_, _) => Code::Cancelled,
85 0 : BasebackupError::Server(_) => Code::Internal,
86 0 : BasebackupError::Shutdown => Code::Unavailable,
87 : };
88 0 : tonic::Status::new(code, err.to_string())
89 0 : }
90 : }
91 :
92 : /// Create basebackup with non-rel data in it.
93 : /// Only include relational data if 'full_backup' is true.
94 : ///
95 : /// Currently we use empty 'req_lsn' in two cases:
96 : /// * During the basebackup right after timeline creation
97 : /// * When working without safekeepers. In this situation it is important to match the lsn
98 : /// we are taking basebackup on with the lsn that is used in pageserver's walreceiver
99 : /// to start the replication.
100 : #[allow(clippy::too_many_arguments)]
101 0 : pub async fn send_basebackup_tarball<'a, W>(
102 0 : write: &'a mut W,
103 0 : timeline: &'a Timeline,
104 0 : req_lsn: Option<Lsn>,
105 0 : prev_lsn: Option<Lsn>,
106 0 : full_backup: bool,
107 0 : replica: bool,
108 0 : gzip_level: Option<async_compression::Level>,
109 0 : ctx: &'a RequestContext,
110 0 : ) -> Result<(), BasebackupError>
111 0 : where
112 0 : W: AsyncWrite + Send + Sync + Unpin,
113 0 : {
114 : // Compute postgres doesn't have any previous WAL files, but the first
115 : // record that it's going to write needs to include the LSN of the
116 : // previous record (xl_prev). We include prev_record_lsn in the
117 : // "neon.signal" file, so that postgres can read it during startup.
118 : //
119 : // We don't keep full history of record boundaries in the page server,
120 : // however, only the predecessor of the latest record on each
121 : // timeline. So we can only provide prev_record_lsn when you take a
122 : // base backup at the end of the timeline, i.e. at last_record_lsn.
123 : // Even at the end of the timeline, we sometimes don't have a valid
124 : // prev_lsn value; that happens if the timeline was just branched from
125 : // an old LSN and it doesn't have any WAL of its own yet. We will set
126 : // prev_lsn to Lsn(0) if we cannot provide the correct value.
127 0 : let (backup_prev, lsn) = if let Some(req_lsn) = req_lsn {
128 : // Backup was requested at a particular LSN. The caller should've
129 : // already checked that it's a valid LSN.
130 :
131 : // If the requested point is the end of the timeline, we can
132 : // provide prev_lsn. (get_last_record_rlsn() might return it as
133 : // zero, though, if no WAL has been generated on this timeline
134 : // yet.)
135 0 : let end_of_timeline = timeline.get_last_record_rlsn();
136 0 : if req_lsn == end_of_timeline.last {
137 0 : (end_of_timeline.prev, req_lsn)
138 : } else {
139 0 : (Lsn(0), req_lsn)
140 : }
141 : } else {
142 : // Backup was requested at end of the timeline.
143 0 : let end_of_timeline = timeline.get_last_record_rlsn();
144 0 : (end_of_timeline.prev, end_of_timeline.last)
145 : };
146 :
147 : // Consolidate the derived and the provided prev_lsn values
148 0 : let prev_record_lsn = if let Some(provided_prev_lsn) = prev_lsn {
149 0 : if backup_prev != Lsn(0) && backup_prev != provided_prev_lsn {
150 0 : return Err(BasebackupError::Server(anyhow!(
151 0 : "backup_prev {backup_prev} != provided_prev_lsn {provided_prev_lsn}"
152 0 : )));
153 0 : }
154 0 : provided_prev_lsn
155 : } else {
156 0 : backup_prev
157 : };
158 :
159 0 : info!(
160 0 : "taking basebackup lsn={lsn}, prev_lsn={prev_record_lsn} \
161 0 : (full_backup={full_backup}, replica={replica}, gzip={gzip_level:?})",
162 : );
163 0 : let span = info_span!("send_tarball", backup_lsn=%lsn);
164 :
165 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
166 0 : timeline.conf.get_vectored_concurrent_io,
167 0 : timeline
168 0 : .gate
169 0 : .enter()
170 0 : .map_err(|_| BasebackupError::Shutdown)?,
171 : );
172 :
173 0 : if let Some(gzip_level) = gzip_level {
174 0 : let mut encoder = GzipEncoder::with_quality(write, gzip_level);
175 0 : Basebackup {
176 0 : ar: Builder::new_non_terminated(&mut encoder),
177 0 : timeline,
178 0 : lsn,
179 0 : prev_record_lsn,
180 0 : full_backup,
181 0 : replica,
182 0 : ctx,
183 0 : io_concurrency,
184 0 : }
185 0 : .send_tarball()
186 0 : .instrument(span)
187 0 : .await?;
188 0 : encoder
189 0 : .shutdown()
190 0 : .await
191 0 : .map_err(|err| BasebackupError::Client(err, "gzip"))?;
192 : } else {
193 0 : Basebackup {
194 0 : ar: Builder::new_non_terminated(write),
195 0 : timeline,
196 0 : lsn,
197 0 : prev_record_lsn,
198 0 : full_backup,
199 0 : replica,
200 0 : ctx,
201 0 : io_concurrency,
202 0 : }
203 0 : .send_tarball()
204 0 : .instrument(span)
205 0 : .await?;
206 : }
207 :
208 0 : Ok(())
209 0 : }
210 :
211 : /// This is short-living object only for the time of tarball creation,
212 : /// created mostly to avoid passing a lot of parameters between various functions
213 : /// used for constructing tarball.
214 : struct Basebackup<'a, W>
215 : where
216 : W: AsyncWrite + Send + Sync + Unpin,
217 : {
218 : ar: Builder<&'a mut W>,
219 : timeline: &'a Timeline,
220 : lsn: Lsn,
221 : prev_record_lsn: Lsn,
222 : full_backup: bool,
223 : replica: bool,
224 : ctx: &'a RequestContext,
225 : io_concurrency: IoConcurrency,
226 : }
227 :
228 : /// A sink that accepts SLRU blocks ordered by key and forwards
229 : /// full segments to the archive.
230 : struct SlruSegmentsBuilder<'a, 'b, W>
231 : where
232 : W: AsyncWrite + Send + Sync + Unpin,
233 : {
234 : ar: &'a mut Builder<&'b mut W>,
235 : buf: Vec<u8>,
236 : current_segment: Option<(SlruKind, u32)>,
237 : total_blocks: usize,
238 : }
239 :
240 : impl<'a, 'b, W> SlruSegmentsBuilder<'a, 'b, W>
241 : where
242 : W: AsyncWrite + Send + Sync + Unpin,
243 : {
244 0 : fn new(ar: &'a mut Builder<&'b mut W>) -> Self {
245 0 : Self {
246 0 : ar,
247 0 : buf: Vec::new(),
248 0 : current_segment: None,
249 0 : total_blocks: 0,
250 0 : }
251 0 : }
252 :
253 0 : async fn add_block(&mut self, key: &Key, block: Bytes) -> Result<(), BasebackupError> {
254 0 : let (kind, segno, _) = key.to_slru_block()?;
255 :
256 0 : match kind {
257 : SlruKind::Clog => {
258 0 : if !(block.len() == BLCKSZ as usize || block.len() == BLCKSZ as usize + 8) {
259 0 : return Err(BasebackupError::Server(anyhow!(
260 0 : "invalid SlruKind::Clog record: block.len()={}",
261 0 : block.len()
262 0 : )));
263 0 : }
264 : }
265 : SlruKind::MultiXactMembers | SlruKind::MultiXactOffsets => {
266 0 : if block.len() != BLCKSZ as usize {
267 0 : return Err(BasebackupError::Server(anyhow!(
268 0 : "invalid {:?} record: block.len()={}",
269 0 : kind,
270 0 : block.len()
271 0 : )));
272 0 : }
273 : }
274 : }
275 :
276 0 : let segment = (kind, segno);
277 0 : match self.current_segment {
278 0 : None => {
279 0 : self.current_segment = Some(segment);
280 0 : self.buf
281 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
282 0 : }
283 0 : Some(current_seg) if current_seg == segment => {
284 0 : self.buf
285 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
286 0 : }
287 : Some(_) => {
288 0 : self.flush().await?;
289 :
290 0 : self.current_segment = Some(segment);
291 0 : self.buf
292 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
293 : }
294 : }
295 :
296 0 : Ok(())
297 0 : }
298 :
299 0 : async fn flush(&mut self) -> Result<(), BasebackupError> {
300 0 : let nblocks = self.buf.len() / BLCKSZ as usize;
301 0 : let (kind, segno) = self.current_segment.take().unwrap();
302 0 : let segname = format!("{kind}/{segno:>04X}");
303 0 : let header = new_tar_header(&segname, self.buf.len() as u64)?;
304 0 : self.ar
305 0 : .append(&header, self.buf.as_slice())
306 0 : .await
307 0 : .map_err(|e| BasebackupError::Client(e, "flush"))?;
308 :
309 0 : self.total_blocks += nblocks;
310 0 : debug!("Added to basebackup slru {} relsize {}", segname, nblocks);
311 :
312 0 : self.buf.clear();
313 :
314 0 : Ok(())
315 0 : }
316 :
317 0 : async fn finish(mut self) -> Result<(), BasebackupError> {
318 0 : let res = if self.current_segment.is_none() || self.buf.is_empty() {
319 0 : Ok(())
320 : } else {
321 0 : self.flush().await
322 : };
323 :
324 0 : info!("Collected {} SLRU blocks", self.total_blocks);
325 :
326 0 : res
327 0 : }
328 : }
329 :
330 : impl<W> Basebackup<'_, W>
331 : where
332 : W: AsyncWrite + Send + Sync + Unpin,
333 : {
334 0 : async fn send_tarball(mut self) -> Result<(), BasebackupError> {
335 : // TODO include checksum
336 :
337 : // Construct the pg_control file from the persisted checkpoint and pg_control
338 : // information. But we only add this to the tarball at the end, so that if the
339 : // writing is interrupted half-way through, the resulting incomplete tarball will
340 : // be missing the pg_control file, which prevents PostgreSQL from starting up on
341 : // it. With proper error handling, you should never try to start up from an
342 : // incomplete basebackup in the first place, of course, but this is a nice little
343 : // extra safety measure.
344 0 : let checkpoint_bytes = self
345 0 : .timeline
346 0 : .get_checkpoint(self.lsn, self.ctx)
347 0 : .await
348 0 : .context("failed to get checkpoint bytes")?;
349 0 : let pg_control_bytes = self
350 0 : .timeline
351 0 : .get_control_file(self.lsn, self.ctx)
352 0 : .await
353 0 : .context("failed to get control bytes")?;
354 0 : let (pg_control_bytes, system_identifier, was_shutdown) =
355 0 : postgres_ffi::generate_pg_control(
356 0 : &pg_control_bytes,
357 0 : &checkpoint_bytes,
358 0 : self.lsn,
359 0 : self.timeline.pg_version,
360 0 : )?;
361 :
362 0 : let lazy_slru_download = self.timeline.get_lazy_slru_download() && !self.full_backup;
363 :
364 0 : let pgversion = self.timeline.pg_version;
365 0 : let subdirs = dispatch_pgversion!(pgversion, &pgv::bindings::PGDATA_SUBDIRS[..]);
366 :
367 : // Create pgdata subdirs structure
368 0 : for dir in subdirs.iter() {
369 0 : let header = new_tar_header_dir(dir)?;
370 0 : self.ar
371 0 : .append(&header, io::empty())
372 0 : .await
373 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball"))?;
374 : }
375 :
376 : // Send config files.
377 0 : for filepath in PGDATA_SPECIAL_FILES.iter() {
378 0 : if *filepath == "pg_hba.conf" {
379 0 : let data = PG_HBA.as_bytes();
380 0 : let header = new_tar_header(filepath, data.len() as u64)?;
381 0 : self.ar
382 0 : .append(&header, data)
383 0 : .await
384 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,pg_hba.conf"))?;
385 : } else {
386 0 : let header = new_tar_header(filepath, 0)?;
387 0 : self.ar
388 0 : .append(&header, io::empty())
389 0 : .await
390 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,add_config_file"))?;
391 : }
392 : }
393 0 : if !lazy_slru_download {
394 : // Gather non-relational files from object storage pages.
395 0 : let slru_partitions = self
396 0 : .timeline
397 0 : .get_slru_keyspace(Version::at(self.lsn), self.ctx)
398 0 : .await?
399 0 : .partition(
400 0 : self.timeline.get_shard_identity(),
401 0 : self.timeline.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
402 0 : BLCKSZ as u64,
403 : );
404 :
405 0 : let mut slru_builder = SlruSegmentsBuilder::new(&mut self.ar);
406 :
407 0 : for part in slru_partitions.parts {
408 0 : let query = VersionedKeySpaceQuery::uniform(part, self.lsn);
409 0 : let blocks = self
410 0 : .timeline
411 0 : .get_vectored(query, self.io_concurrency.clone(), self.ctx)
412 0 : .await?;
413 :
414 0 : for (key, block) in blocks {
415 0 : let block = block?;
416 0 : slru_builder.add_block(&key, block).await?;
417 : }
418 : }
419 0 : slru_builder.finish().await?;
420 0 : }
421 :
422 0 : let mut min_restart_lsn: Lsn = Lsn::MAX;
423 : // Create tablespace directories
424 0 : for ((spcnode, dbnode), has_relmap_file) in
425 0 : self.timeline.list_dbdirs(self.lsn, self.ctx).await?
426 : {
427 0 : self.add_dbdir(spcnode, dbnode, has_relmap_file).await?;
428 :
429 : // If full backup is requested, include all relation files.
430 : // Otherwise only include init forks of unlogged relations.
431 0 : let rels = self
432 0 : .timeline
433 0 : .list_rels(spcnode, dbnode, Version::at(self.lsn), self.ctx)
434 0 : .await?;
435 0 : for &rel in rels.iter() {
436 : // Send init fork as main fork to provide well formed empty
437 : // contents of UNLOGGED relations. Postgres copies it in
438 : // `reinit.c` during recovery.
439 0 : if rel.forknum == INIT_FORKNUM {
440 : // I doubt we need _init fork itself, but having it at least
441 : // serves as a marker relation is unlogged.
442 0 : self.add_rel(rel, rel).await?;
443 0 : self.add_rel(rel, rel.with_forknum(MAIN_FORKNUM)).await?;
444 0 : continue;
445 0 : }
446 :
447 0 : if self.full_backup {
448 0 : if rel.forknum == MAIN_FORKNUM && rels.contains(&rel.with_forknum(INIT_FORKNUM))
449 : {
450 : // skip this, will include it when we reach the init fork
451 0 : continue;
452 0 : }
453 0 : self.add_rel(rel, rel).await?;
454 0 : }
455 : }
456 : }
457 :
458 0 : let start_time = Instant::now();
459 0 : let aux_files = self
460 0 : .timeline
461 0 : .list_aux_files(self.lsn, self.ctx, self.io_concurrency.clone())
462 0 : .await?;
463 0 : let aux_scan_time = start_time.elapsed();
464 0 : let aux_estimated_size = aux_files
465 0 : .values()
466 0 : .map(|content| content.len())
467 0 : .sum::<usize>();
468 0 : info!(
469 0 : "Scanned {} aux files in {}ms, aux file content size = {}",
470 0 : aux_files.len(),
471 0 : aux_scan_time.as_millis(),
472 : aux_estimated_size
473 : );
474 :
475 0 : for (path, content) in aux_files {
476 0 : if path.starts_with("pg_replslot") {
477 : // Do not create LR slots at standby because they are not used but prevent WAL truncation
478 0 : if self.replica {
479 0 : continue;
480 0 : }
481 0 : let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
482 0 : let restart_lsn = Lsn(u64::from_le_bytes(
483 0 : content[offs..offs + 8].try_into().unwrap(),
484 0 : ));
485 0 : info!("Replication slot {} restart LSN={}", path, restart_lsn);
486 0 : min_restart_lsn = Lsn::min(min_restart_lsn, restart_lsn);
487 0 : } else if path == "pg_logical/replorigin_checkpoint" {
488 : // replorigin_checkoint is written only on compute shutdown, so it contains
489 : // deteriorated values. So we generate our own version of this file for the particular LSN
490 : // based on information about replorigins extracted from transaction commit records.
491 : // In future we will not generate AUX record for "pg_logical/replorigin_checkpoint" at all,
492 : // but now we should handle (skip) it for backward compatibility.
493 0 : continue;
494 0 : } else if path == "pg_stat/pgstat.stat" && !was_shutdown {
495 : // Drop statistic in case of abnormal termination, i.e. if we're not starting from the exact LSN
496 : // of a shutdown checkpoint.
497 0 : continue;
498 0 : }
499 0 : let header = new_tar_header(&path, content.len() as u64)?;
500 0 : self.ar
501 0 : .append(&header, &*content)
502 0 : .await
503 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,add_aux_file"))?;
504 : }
505 :
506 0 : if min_restart_lsn != Lsn::MAX {
507 0 : info!(
508 0 : "Min restart LSN for logical replication is {}",
509 : min_restart_lsn
510 : );
511 0 : let data = min_restart_lsn.0.to_le_bytes();
512 0 : let header = new_tar_header("restart.lsn", data.len() as u64)?;
513 0 : self.ar
514 0 : .append(&header, &data[..])
515 0 : .await
516 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,restart.lsn"))?;
517 0 : }
518 0 : for xid in self
519 0 : .timeline
520 0 : .list_twophase_files(self.lsn, self.ctx)
521 0 : .await?
522 : {
523 0 : self.add_twophase_file(xid).await?;
524 : }
525 0 : let repl_origins = self
526 0 : .timeline
527 0 : .get_replorigins(self.lsn, self.ctx, self.io_concurrency.clone())
528 0 : .await?;
529 0 : let n_origins = repl_origins.len();
530 0 : if n_origins != 0 {
531 : //
532 : // Construct "pg_logical/replorigin_checkpoint" file based on information about replication origins
533 : // extracted from transaction commit record. We are using this file to pass information about replication
534 : // origins to compute to allow logical replication to restart from proper point.
535 : //
536 0 : let mut content = Vec::with_capacity(n_origins * 16 + 8);
537 0 : content.extend_from_slice(&pg_constants::REPLICATION_STATE_MAGIC.to_le_bytes());
538 0 : for (origin_id, origin_lsn) in repl_origins {
539 0 : content.extend_from_slice(&origin_id.to_le_bytes());
540 0 : content.extend_from_slice(&[0u8; 6]); // align to 8 bytes
541 0 : content.extend_from_slice(&origin_lsn.0.to_le_bytes());
542 0 : }
543 0 : let crc32 = crc32c::crc32c(&content);
544 0 : content.extend_from_slice(&crc32.to_le_bytes());
545 0 : let header = new_tar_header("pg_logical/replorigin_checkpoint", content.len() as u64)?;
546 0 : self.ar.append(&header, &*content).await.map_err(|e| {
547 0 : BasebackupError::Client(e, "send_tarball,pg_logical/replorigin_checkpoint")
548 0 : })?;
549 0 : }
550 :
551 0 : fail_point!("basebackup-before-control-file", |_| {
552 0 : Err(BasebackupError::Server(anyhow!(
553 0 : "failpoint basebackup-before-control-file"
554 0 : )))
555 0 : });
556 :
557 : // Last, add the pg_control file and bootstrap WAL segment.
558 0 : self.add_pgcontrol_file(pg_control_bytes, system_identifier)
559 0 : .await?;
560 0 : self.ar
561 0 : .finish()
562 0 : .await
563 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,finish"))?;
564 0 : debug!("all tarred up!");
565 0 : Ok(())
566 0 : }
567 :
568 : /// Add contents of relfilenode `src`, naming it as `dst`.
569 0 : async fn add_rel(&mut self, src: RelTag, dst: RelTag) -> Result<(), BasebackupError> {
570 0 : let nblocks = self
571 0 : .timeline
572 0 : .get_rel_size(src, Version::at(self.lsn), self.ctx)
573 0 : .await?;
574 :
575 : // If the relation is empty, create an empty file
576 0 : if nblocks == 0 {
577 0 : let file_name = dst.to_segfile_name(0);
578 0 : let header = new_tar_header(&file_name, 0)?;
579 0 : self.ar
580 0 : .append(&header, io::empty())
581 0 : .await
582 0 : .map_err(|e| BasebackupError::Client(e, "add_rel,empty"))?;
583 0 : return Ok(());
584 0 : }
585 :
586 : // Add a file for each chunk of blocks (aka segment)
587 0 : let mut startblk = 0;
588 0 : let mut seg = 0;
589 0 : while startblk < nblocks {
590 0 : let endblk = std::cmp::min(startblk + RELSEG_SIZE, nblocks);
591 :
592 0 : let mut segment_data: Vec<u8> = vec![];
593 0 : for blknum in startblk..endblk {
594 0 : let img = self
595 0 : .timeline
596 0 : // TODO: investigate using get_vectored for the entire startblk..endblk range.
597 0 : // But this code path is not on the critical path for most basebackups (?).
598 0 : .get(rel_block_to_key(src, blknum), self.lsn, self.ctx)
599 0 : .await?;
600 0 : segment_data.extend_from_slice(&img[..]);
601 : }
602 :
603 0 : let file_name = dst.to_segfile_name(seg as u32);
604 0 : let header = new_tar_header(&file_name, segment_data.len() as u64)?;
605 0 : self.ar
606 0 : .append(&header, segment_data.as_slice())
607 0 : .await
608 0 : .map_err(|e| BasebackupError::Client(e, "add_rel,segment"))?;
609 :
610 0 : seg += 1;
611 0 : startblk = endblk;
612 : }
613 :
614 0 : Ok(())
615 0 : }
616 :
617 : //
618 : // Include database/tablespace directories.
619 : //
620 : // Each directory contains a PG_VERSION file, and the default database
621 : // directories also contain pg_filenode.map files.
622 : //
623 0 : async fn add_dbdir(
624 0 : &mut self,
625 0 : spcnode: u32,
626 0 : dbnode: u32,
627 0 : has_relmap_file: bool,
628 0 : ) -> Result<(), BasebackupError> {
629 0 : let relmap_img = if has_relmap_file {
630 0 : let img = self
631 0 : .timeline
632 0 : .get_relmap_file(spcnode, dbnode, Version::at(self.lsn), self.ctx)
633 0 : .await?;
634 :
635 0 : if img.len()
636 0 : != dispatch_pgversion!(self.timeline.pg_version, pgv::bindings::SIZEOF_RELMAPFILE)
637 : {
638 0 : return Err(BasebackupError::Server(anyhow!(
639 0 : "img.len() != SIZE_OF_RELMAPFILE, img.len()={}",
640 0 : img.len(),
641 0 : )));
642 0 : }
643 :
644 0 : Some(img)
645 : } else {
646 0 : None
647 : };
648 :
649 0 : if spcnode == GLOBALTABLESPACE_OID {
650 0 : let pg_version_str = self.timeline.pg_version.versionfile_string();
651 0 : let header = new_tar_header("PG_VERSION", pg_version_str.len() as u64)?;
652 0 : self.ar
653 0 : .append(&header, pg_version_str.as_bytes())
654 0 : .await
655 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,PG_VERSION"))?;
656 :
657 0 : info!("timeline.pg_version {}", self.timeline.pg_version);
658 :
659 0 : if let Some(img) = relmap_img {
660 : // filenode map for global tablespace
661 0 : let header = new_tar_header("global/pg_filenode.map", img.len() as u64)?;
662 0 : self.ar
663 0 : .append(&header, &img[..])
664 0 : .await
665 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,global/pg_filenode.map"))?;
666 : } else {
667 0 : warn!("global/pg_filenode.map is missing");
668 : }
669 : } else {
670 : // User defined tablespaces are not supported. However, as
671 : // a special case, if a tablespace/db directory is
672 : // completely empty, we can leave it out altogether. This
673 : // makes taking a base backup after the 'tablespace'
674 : // regression test pass, because the test drops the
675 : // created tablespaces after the tests.
676 : //
677 : // FIXME: this wouldn't be necessary, if we handled
678 : // XLOG_TBLSPC_DROP records. But we probably should just
679 : // throw an error on CREATE TABLESPACE in the first place.
680 0 : if !has_relmap_file
681 0 : && self
682 0 : .timeline
683 0 : .list_rels(spcnode, dbnode, Version::at(self.lsn), self.ctx)
684 0 : .await?
685 0 : .is_empty()
686 : {
687 0 : return Ok(());
688 0 : }
689 : // User defined tablespaces are not supported
690 0 : if spcnode != DEFAULTTABLESPACE_OID {
691 0 : return Err(BasebackupError::Server(anyhow!(
692 0 : "spcnode != DEFAULTTABLESPACE_OID, spcnode={spcnode}"
693 0 : )));
694 0 : }
695 :
696 : // Append dir path for each database
697 0 : let path = format!("base/{dbnode}");
698 0 : let header = new_tar_header_dir(&path)?;
699 0 : self.ar
700 0 : .append(&header, io::empty())
701 0 : .await
702 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base"))?;
703 :
704 0 : if let Some(img) = relmap_img {
705 0 : let dst_path = format!("base/{dbnode}/PG_VERSION");
706 :
707 0 : let pg_version_str = self.timeline.pg_version.versionfile_string();
708 0 : let header = new_tar_header(&dst_path, pg_version_str.len() as u64)?;
709 0 : self.ar
710 0 : .append(&header, pg_version_str.as_bytes())
711 0 : .await
712 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base/PG_VERSION"))?;
713 :
714 0 : let relmap_path = format!("base/{dbnode}/pg_filenode.map");
715 0 : let header = new_tar_header(&relmap_path, img.len() as u64)?;
716 0 : self.ar
717 0 : .append(&header, &img[..])
718 0 : .await
719 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base/pg_filenode.map"))?;
720 0 : }
721 : };
722 0 : Ok(())
723 0 : }
724 :
725 : //
726 : // Extract twophase state files
727 : //
728 0 : async fn add_twophase_file(&mut self, xid: u64) -> Result<(), BasebackupError> {
729 0 : let img = self
730 0 : .timeline
731 0 : .get_twophase_file(xid, self.lsn, self.ctx)
732 0 : .await?;
733 :
734 0 : let mut buf = BytesMut::new();
735 0 : buf.extend_from_slice(&img[..]);
736 0 : let crc = crc32c::crc32c(&img[..]);
737 0 : buf.put_u32_le(crc);
738 0 : let path = if self.timeline.pg_version < PgMajorVersion::PG17 {
739 0 : format!("pg_twophase/{xid:>08X}")
740 : } else {
741 0 : format!("pg_twophase/{xid:>016X}")
742 : };
743 0 : let header = new_tar_header(&path, buf.len() as u64)?;
744 0 : self.ar
745 0 : .append(&header, &buf[..])
746 0 : .await
747 0 : .map_err(|e| BasebackupError::Client(e, "add_twophase_file"))?;
748 :
749 0 : Ok(())
750 0 : }
751 :
752 : //
753 : // Add generated pg_control file and bootstrap WAL segment.
754 : // Also send neon.signal and zenith.signal file with extra bootstrap data.
755 : //
756 0 : async fn add_pgcontrol_file(
757 0 : &mut self,
758 0 : pg_control_bytes: Bytes,
759 0 : system_identifier: u64,
760 0 : ) -> Result<(), BasebackupError> {
761 : // add neon.signal file
762 0 : let mut neon_signal = String::new();
763 0 : if self.prev_record_lsn == Lsn(0) {
764 0 : if self.timeline.is_ancestor_lsn(self.lsn) {
765 0 : write!(neon_signal, "PREV LSN: none")
766 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
767 : } else {
768 0 : write!(neon_signal, "PREV LSN: invalid")
769 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
770 : }
771 : } else {
772 0 : write!(neon_signal, "PREV LSN: {}", self.prev_record_lsn)
773 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
774 : }
775 :
776 : // TODO: Remove zenith.signal once all historical computes have been replaced
777 : // ... and thus support the neon.signal file.
778 0 : for signalfilename in ["neon.signal", "zenith.signal"] {
779 0 : self.ar
780 0 : .append(
781 0 : &new_tar_header(signalfilename, neon_signal.len() as u64)?,
782 0 : neon_signal.as_bytes(),
783 : )
784 0 : .await
785 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,neon.signal"))?;
786 : }
787 :
788 : //send pg_control
789 0 : let header = new_tar_header("global/pg_control", pg_control_bytes.len() as u64)?;
790 0 : self.ar
791 0 : .append(&header, &pg_control_bytes[..])
792 0 : .await
793 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,pg_control"))?;
794 :
795 : //send wal segment
796 0 : let segno = self.lsn.segment_number(WAL_SEGMENT_SIZE);
797 0 : let wal_file_name = XLogFileName(PG_TLI, segno, WAL_SEGMENT_SIZE);
798 0 : let wal_file_path = format!("pg_wal/{wal_file_name}");
799 0 : let header = new_tar_header(&wal_file_path, WAL_SEGMENT_SIZE as u64)?;
800 :
801 0 : let wal_seg = postgres_ffi::generate_wal_segment(
802 0 : segno,
803 0 : system_identifier,
804 0 : self.timeline.pg_version,
805 0 : self.lsn,
806 : )
807 0 : .map_err(|e| anyhow!(e).context("Failed generating wal segment"))?;
808 0 : if wal_seg.len() != WAL_SEGMENT_SIZE {
809 0 : return Err(BasebackupError::Server(anyhow!(
810 0 : "wal_seg.len() != WAL_SEGMENT_SIZE, wal_seg.len()={}",
811 0 : wal_seg.len()
812 0 : )));
813 0 : }
814 0 : self.ar
815 0 : .append(&header, &wal_seg[..])
816 0 : .await
817 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,wal_segment"))?;
818 0 : Ok(())
819 0 : }
820 : }
821 :
822 : //
823 : // Create new tarball entry header
824 : //
825 0 : fn new_tar_header(path: &str, size: u64) -> anyhow::Result<Header> {
826 0 : let mut header = Header::new_gnu();
827 0 : header.set_size(size);
828 0 : header.set_path(path)?;
829 0 : header.set_mode(0b110000000); // -rw-------
830 0 : header.set_mtime(
831 : // use currenttime as last modified time
832 0 : SystemTime::now()
833 0 : .duration_since(SystemTime::UNIX_EPOCH)
834 0 : .unwrap()
835 0 : .as_secs(),
836 : );
837 0 : header.set_cksum();
838 0 : Ok(header)
839 0 : }
840 :
841 0 : fn new_tar_header_dir(path: &str) -> anyhow::Result<Header> {
842 0 : let mut header = Header::new_gnu();
843 0 : header.set_size(0);
844 0 : header.set_path(path)?;
845 0 : header.set_mode(0o755); // -rw-------
846 0 : header.set_entry_type(EntryType::dir());
847 0 : header.set_mtime(
848 : // use currenttime as last modified time
849 0 : SystemTime::now()
850 0 : .duration_since(SystemTime::UNIX_EPOCH)
851 0 : .unwrap()
852 0 : .as_secs(),
853 : );
854 0 : header.set_cksum();
855 0 : Ok(header)
856 0 : }
|
