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 bytes::{BufMut, Bytes, BytesMut};
18 : use fail::fail_point;
19 : use pageserver_api::key::{Key, rel_block_to_key};
20 : use pageserver_api::reltag::{RelTag, SlruKind};
21 : use postgres_ffi::pg_constants::{
22 : DEFAULTTABLESPACE_OID, GLOBALTABLESPACE_OID, PG_HBA, PGDATA_SPECIAL_FILES,
23 : };
24 : use postgres_ffi::relfile_utils::{INIT_FORKNUM, MAIN_FORKNUM};
25 : use postgres_ffi::{
26 : BLCKSZ, PG_TLI, RELSEG_SIZE, WAL_SEGMENT_SIZE, XLogFileName, dispatch_pgversion, pg_constants,
27 : };
28 : use tokio::io;
29 : use tokio::io::AsyncWrite;
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 0 : pub async fn send_basebackup_tarball<'a, W>(
101 0 : write: &'a mut W,
102 0 : timeline: &'a Timeline,
103 0 : req_lsn: Option<Lsn>,
104 0 : prev_lsn: Option<Lsn>,
105 0 : full_backup: bool,
106 0 : replica: bool,
107 0 : ctx: &'a RequestContext,
108 0 : ) -> Result<(), BasebackupError>
109 0 : where
110 0 : W: AsyncWrite + Send + Sync + Unpin,
111 0 : {
112 : // Compute postgres doesn't have any previous WAL files, but the first
113 : // record that it's going to write needs to include the LSN of the
114 : // previous record (xl_prev). We include prev_record_lsn in the
115 : // "zenith.signal" file, so that postgres can read it during startup.
116 : //
117 : // We don't keep full history of record boundaries in the page server,
118 : // however, only the predecessor of the latest record on each
119 : // timeline. So we can only provide prev_record_lsn when you take a
120 : // base backup at the end of the timeline, i.e. at last_record_lsn.
121 : // Even at the end of the timeline, we sometimes don't have a valid
122 : // prev_lsn value; that happens if the timeline was just branched from
123 : // an old LSN and it doesn't have any WAL of its own yet. We will set
124 : // prev_lsn to Lsn(0) if we cannot provide the correct value.
125 0 : let (backup_prev, backup_lsn) = if let Some(req_lsn) = req_lsn {
126 : // Backup was requested at a particular LSN. The caller should've
127 : // already checked that it's a valid LSN.
128 :
129 : // If the requested point is the end of the timeline, we can
130 : // provide prev_lsn. (get_last_record_rlsn() might return it as
131 : // zero, though, if no WAL has been generated on this timeline
132 : // yet.)
133 0 : let end_of_timeline = timeline.get_last_record_rlsn();
134 0 : if req_lsn == end_of_timeline.last {
135 0 : (end_of_timeline.prev, req_lsn)
136 : } else {
137 0 : (Lsn(0), req_lsn)
138 : }
139 : } else {
140 : // Backup was requested at end of the timeline.
141 0 : let end_of_timeline = timeline.get_last_record_rlsn();
142 0 : (end_of_timeline.prev, end_of_timeline.last)
143 : };
144 :
145 : // Consolidate the derived and the provided prev_lsn values
146 0 : let prev_lsn = if let Some(provided_prev_lsn) = prev_lsn {
147 0 : if backup_prev != Lsn(0) && backup_prev != provided_prev_lsn {
148 0 : return Err(BasebackupError::Server(anyhow!(
149 0 : "backup_prev {backup_prev} != provided_prev_lsn {provided_prev_lsn}"
150 0 : )));
151 0 : }
152 0 : provided_prev_lsn
153 : } else {
154 0 : backup_prev
155 : };
156 :
157 0 : info!(
158 0 : "taking basebackup lsn={}, prev_lsn={} (full_backup={}, replica={})",
159 : backup_lsn, prev_lsn, full_backup, replica
160 : );
161 :
162 0 : let basebackup = Basebackup {
163 0 : ar: Builder::new_non_terminated(write),
164 0 : timeline,
165 0 : lsn: backup_lsn,
166 0 : prev_record_lsn: prev_lsn,
167 0 : full_backup,
168 0 : replica,
169 0 : ctx,
170 0 : io_concurrency: IoConcurrency::spawn_from_conf(
171 0 : timeline.conf.get_vectored_concurrent_io,
172 0 : timeline
173 0 : .gate
174 0 : .enter()
175 0 : .map_err(|_| BasebackupError::Shutdown)?,
176 : ),
177 : };
178 0 : basebackup
179 0 : .send_tarball()
180 0 : .instrument(info_span!("send_tarball", backup_lsn=%backup_lsn))
181 0 : .await
182 0 : }
183 :
184 : /// This is short-living object only for the time of tarball creation,
185 : /// created mostly to avoid passing a lot of parameters between various functions
186 : /// used for constructing tarball.
187 : struct Basebackup<'a, W>
188 : where
189 : W: AsyncWrite + Send + Sync + Unpin,
190 : {
191 : ar: Builder<&'a mut W>,
192 : timeline: &'a Timeline,
193 : lsn: Lsn,
194 : prev_record_lsn: Lsn,
195 : full_backup: bool,
196 : replica: bool,
197 : ctx: &'a RequestContext,
198 : io_concurrency: IoConcurrency,
199 : }
200 :
201 : /// A sink that accepts SLRU blocks ordered by key and forwards
202 : /// full segments to the archive.
203 : struct SlruSegmentsBuilder<'a, 'b, W>
204 : where
205 : W: AsyncWrite + Send + Sync + Unpin,
206 : {
207 : ar: &'a mut Builder<&'b mut W>,
208 : buf: Vec<u8>,
209 : current_segment: Option<(SlruKind, u32)>,
210 : total_blocks: usize,
211 : }
212 :
213 : impl<'a, 'b, W> SlruSegmentsBuilder<'a, 'b, W>
214 : where
215 : W: AsyncWrite + Send + Sync + Unpin,
216 : {
217 0 : fn new(ar: &'a mut Builder<&'b mut W>) -> Self {
218 0 : Self {
219 0 : ar,
220 0 : buf: Vec::new(),
221 0 : current_segment: None,
222 0 : total_blocks: 0,
223 0 : }
224 0 : }
225 :
226 0 : async fn add_block(&mut self, key: &Key, block: Bytes) -> Result<(), BasebackupError> {
227 0 : let (kind, segno, _) = key.to_slru_block()?;
228 :
229 0 : match kind {
230 : SlruKind::Clog => {
231 0 : if !(block.len() == BLCKSZ as usize || block.len() == BLCKSZ as usize + 8) {
232 0 : return Err(BasebackupError::Server(anyhow!(
233 0 : "invalid SlruKind::Clog record: block.len()={}",
234 0 : block.len()
235 0 : )));
236 0 : }
237 : }
238 : SlruKind::MultiXactMembers | SlruKind::MultiXactOffsets => {
239 0 : if block.len() != BLCKSZ as usize {
240 0 : return Err(BasebackupError::Server(anyhow!(
241 0 : "invalid {:?} record: block.len()={}",
242 0 : kind,
243 0 : block.len()
244 0 : )));
245 0 : }
246 : }
247 : }
248 :
249 0 : let segment = (kind, segno);
250 0 : match self.current_segment {
251 0 : None => {
252 0 : self.current_segment = Some(segment);
253 0 : self.buf
254 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
255 0 : }
256 0 : Some(current_seg) if current_seg == segment => {
257 0 : self.buf
258 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
259 0 : }
260 : Some(_) => {
261 0 : self.flush().await?;
262 :
263 0 : self.current_segment = Some(segment);
264 0 : self.buf
265 0 : .extend_from_slice(block.slice(..BLCKSZ as usize).as_ref());
266 : }
267 : }
268 :
269 0 : Ok(())
270 0 : }
271 :
272 0 : async fn flush(&mut self) -> Result<(), BasebackupError> {
273 0 : let nblocks = self.buf.len() / BLCKSZ as usize;
274 0 : let (kind, segno) = self.current_segment.take().unwrap();
275 0 : let segname = format!("{kind}/{segno:>04X}");
276 0 : let header = new_tar_header(&segname, self.buf.len() as u64)?;
277 0 : self.ar
278 0 : .append(&header, self.buf.as_slice())
279 0 : .await
280 0 : .map_err(|e| BasebackupError::Client(e, "flush"))?;
281 :
282 0 : self.total_blocks += nblocks;
283 0 : debug!("Added to basebackup slru {} relsize {}", segname, nblocks);
284 :
285 0 : self.buf.clear();
286 0 :
287 0 : Ok(())
288 0 : }
289 :
290 0 : async fn finish(mut self) -> Result<(), BasebackupError> {
291 0 : let res = if self.current_segment.is_none() || self.buf.is_empty() {
292 0 : Ok(())
293 : } else {
294 0 : self.flush().await
295 : };
296 :
297 0 : info!("Collected {} SLRU blocks", self.total_blocks);
298 :
299 0 : res
300 0 : }
301 : }
302 :
303 : impl<W> Basebackup<'_, W>
304 : where
305 : W: AsyncWrite + Send + Sync + Unpin,
306 : {
307 0 : async fn send_tarball(mut self) -> Result<(), BasebackupError> {
308 : // TODO include checksum
309 :
310 : // Construct the pg_control file from the persisted checkpoint and pg_control
311 : // information. But we only add this to the tarball at the end, so that if the
312 : // writing is interrupted half-way through, the resulting incomplete tarball will
313 : // be missing the pg_control file, which prevents PostgreSQL from starting up on
314 : // it. With proper error handling, you should never try to start up from an
315 : // incomplete basebackup in the first place, of course, but this is a nice little
316 : // extra safety measure.
317 0 : let checkpoint_bytes = self
318 0 : .timeline
319 0 : .get_checkpoint(self.lsn, self.ctx)
320 0 : .await
321 0 : .context("failed to get checkpoint bytes")?;
322 0 : let pg_control_bytes = self
323 0 : .timeline
324 0 : .get_control_file(self.lsn, self.ctx)
325 0 : .await
326 0 : .context("failed to get control bytes")?;
327 0 : let (pg_control_bytes, system_identifier, was_shutdown) =
328 0 : postgres_ffi::generate_pg_control(
329 0 : &pg_control_bytes,
330 0 : &checkpoint_bytes,
331 0 : self.lsn,
332 0 : self.timeline.pg_version,
333 0 : )?;
334 :
335 0 : let lazy_slru_download = self.timeline.get_lazy_slru_download() && !self.full_backup;
336 :
337 0 : let pgversion = self.timeline.pg_version;
338 0 : let subdirs = dispatch_pgversion!(pgversion, &pgv::bindings::PGDATA_SUBDIRS[..]);
339 :
340 : // Create pgdata subdirs structure
341 0 : for dir in subdirs.iter() {
342 0 : let header = new_tar_header_dir(dir)?;
343 0 : self.ar
344 0 : .append(&header, io::empty())
345 0 : .await
346 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball"))?;
347 : }
348 :
349 : // Send config files.
350 0 : for filepath in PGDATA_SPECIAL_FILES.iter() {
351 0 : if *filepath == "pg_hba.conf" {
352 0 : let data = PG_HBA.as_bytes();
353 0 : let header = new_tar_header(filepath, data.len() as u64)?;
354 0 : self.ar
355 0 : .append(&header, data)
356 0 : .await
357 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,pg_hba.conf"))?;
358 : } else {
359 0 : let header = new_tar_header(filepath, 0)?;
360 0 : self.ar
361 0 : .append(&header, io::empty())
362 0 : .await
363 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,add_config_file"))?;
364 : }
365 : }
366 0 : if !lazy_slru_download {
367 : // Gather non-relational files from object storage pages.
368 0 : let slru_partitions = self
369 0 : .timeline
370 0 : .get_slru_keyspace(Version::at(self.lsn), self.ctx)
371 0 : .await?
372 0 : .partition(
373 0 : self.timeline.get_shard_identity(),
374 0 : self.timeline.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
375 0 : );
376 0 :
377 0 : let mut slru_builder = SlruSegmentsBuilder::new(&mut self.ar);
378 :
379 0 : for part in slru_partitions.parts {
380 0 : let query = VersionedKeySpaceQuery::uniform(part, self.lsn);
381 0 : let blocks = self
382 0 : .timeline
383 0 : .get_vectored(query, self.io_concurrency.clone(), self.ctx)
384 0 : .await?;
385 :
386 0 : for (key, block) in blocks {
387 0 : let block = block?;
388 0 : slru_builder.add_block(&key, block).await?;
389 : }
390 : }
391 0 : slru_builder.finish().await?;
392 0 : }
393 :
394 0 : let mut min_restart_lsn: Lsn = Lsn::MAX;
395 : // Create tablespace directories
396 0 : for ((spcnode, dbnode), has_relmap_file) in
397 0 : self.timeline.list_dbdirs(self.lsn, self.ctx).await?
398 : {
399 0 : self.add_dbdir(spcnode, dbnode, has_relmap_file).await?;
400 :
401 : // If full backup is requested, include all relation files.
402 : // Otherwise only include init forks of unlogged relations.
403 0 : let rels = self
404 0 : .timeline
405 0 : .list_rels(spcnode, dbnode, Version::at(self.lsn), self.ctx)
406 0 : .await?;
407 0 : for &rel in rels.iter() {
408 : // Send init fork as main fork to provide well formed empty
409 : // contents of UNLOGGED relations. Postgres copies it in
410 : // `reinit.c` during recovery.
411 0 : if rel.forknum == INIT_FORKNUM {
412 : // I doubt we need _init fork itself, but having it at least
413 : // serves as a marker relation is unlogged.
414 0 : self.add_rel(rel, rel).await?;
415 0 : self.add_rel(rel, rel.with_forknum(MAIN_FORKNUM)).await?;
416 0 : continue;
417 0 : }
418 0 :
419 0 : if self.full_backup {
420 0 : if rel.forknum == MAIN_FORKNUM && rels.contains(&rel.with_forknum(INIT_FORKNUM))
421 : {
422 : // skip this, will include it when we reach the init fork
423 0 : continue;
424 0 : }
425 0 : self.add_rel(rel, rel).await?;
426 0 : }
427 : }
428 : }
429 :
430 0 : let start_time = Instant::now();
431 0 : let aux_files = self
432 0 : .timeline
433 0 : .list_aux_files(self.lsn, self.ctx, self.io_concurrency.clone())
434 0 : .await?;
435 0 : let aux_scan_time = start_time.elapsed();
436 0 : let aux_estimated_size = aux_files
437 0 : .values()
438 0 : .map(|content| content.len())
439 0 : .sum::<usize>();
440 0 : info!(
441 0 : "Scanned {} aux files in {}ms, aux file content size = {}",
442 0 : aux_files.len(),
443 0 : aux_scan_time.as_millis(),
444 : aux_estimated_size
445 : );
446 :
447 0 : for (path, content) in aux_files {
448 0 : if path.starts_with("pg_replslot") {
449 : // Do not create LR slots at standby because they are not used but prevent WAL truncation
450 0 : if self.replica {
451 0 : continue;
452 0 : }
453 0 : let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
454 0 : let restart_lsn = Lsn(u64::from_le_bytes(
455 0 : content[offs..offs + 8].try_into().unwrap(),
456 0 : ));
457 0 : info!("Replication slot {} restart LSN={}", path, restart_lsn);
458 0 : min_restart_lsn = Lsn::min(min_restart_lsn, restart_lsn);
459 0 : } else if path == "pg_logical/replorigin_checkpoint" {
460 : // replorigin_checkoint is written only on compute shutdown, so it contains
461 : // deteriorated values. So we generate our own version of this file for the particular LSN
462 : // based on information about replorigins extracted from transaction commit records.
463 : // In future we will not generate AUX record for "pg_logical/replorigin_checkpoint" at all,
464 : // but now we should handle (skip) it for backward compatibility.
465 0 : continue;
466 0 : } else if path == "pg_stat/pgstat.stat" && !was_shutdown {
467 : // Drop statistic in case of abnormal termination, i.e. if we're not starting from the exact LSN
468 : // of a shutdown checkpoint.
469 0 : continue;
470 0 : }
471 0 : let header = new_tar_header(&path, content.len() as u64)?;
472 0 : self.ar
473 0 : .append(&header, &*content)
474 0 : .await
475 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,add_aux_file"))?;
476 : }
477 :
478 0 : if min_restart_lsn != Lsn::MAX {
479 0 : info!(
480 0 : "Min restart LSN for logical replication is {}",
481 : min_restart_lsn
482 : );
483 0 : let data = min_restart_lsn.0.to_le_bytes();
484 0 : let header = new_tar_header("restart.lsn", data.len() as u64)?;
485 0 : self.ar
486 0 : .append(&header, &data[..])
487 0 : .await
488 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,restart.lsn"))?;
489 0 : }
490 0 : for xid in self
491 0 : .timeline
492 0 : .list_twophase_files(self.lsn, self.ctx)
493 0 : .await?
494 : {
495 0 : self.add_twophase_file(xid).await?;
496 : }
497 0 : let repl_origins = self
498 0 : .timeline
499 0 : .get_replorigins(self.lsn, self.ctx, self.io_concurrency.clone())
500 0 : .await?;
501 0 : let n_origins = repl_origins.len();
502 0 : if n_origins != 0 {
503 : //
504 : // Construct "pg_logical/replorigin_checkpoint" file based on information about replication origins
505 : // extracted from transaction commit record. We are using this file to pass information about replication
506 : // origins to compute to allow logical replication to restart from proper point.
507 : //
508 0 : let mut content = Vec::with_capacity(n_origins * 16 + 8);
509 0 : content.extend_from_slice(&pg_constants::REPLICATION_STATE_MAGIC.to_le_bytes());
510 0 : for (origin_id, origin_lsn) in repl_origins {
511 0 : content.extend_from_slice(&origin_id.to_le_bytes());
512 0 : content.extend_from_slice(&[0u8; 6]); // align to 8 bytes
513 0 : content.extend_from_slice(&origin_lsn.0.to_le_bytes());
514 0 : }
515 0 : let crc32 = crc32c::crc32c(&content);
516 0 : content.extend_from_slice(&crc32.to_le_bytes());
517 0 : let header = new_tar_header("pg_logical/replorigin_checkpoint", content.len() as u64)?;
518 0 : self.ar.append(&header, &*content).await.map_err(|e| {
519 0 : BasebackupError::Client(e, "send_tarball,pg_logical/replorigin_checkpoint")
520 0 : })?;
521 0 : }
522 :
523 0 : fail_point!("basebackup-before-control-file", |_| {
524 0 : Err(BasebackupError::Server(anyhow!(
525 0 : "failpoint basebackup-before-control-file"
526 0 : )))
527 0 : });
528 :
529 : // Last, add the pg_control file and bootstrap WAL segment.
530 0 : self.add_pgcontrol_file(pg_control_bytes, system_identifier)
531 0 : .await?;
532 0 : self.ar
533 0 : .finish()
534 0 : .await
535 0 : .map_err(|e| BasebackupError::Client(e, "send_tarball,finish"))?;
536 0 : debug!("all tarred up!");
537 0 : Ok(())
538 0 : }
539 :
540 : /// Add contents of relfilenode `src`, naming it as `dst`.
541 0 : async fn add_rel(&mut self, src: RelTag, dst: RelTag) -> Result<(), BasebackupError> {
542 0 : let nblocks = self
543 0 : .timeline
544 0 : .get_rel_size(src, Version::at(self.lsn), self.ctx)
545 0 : .await?;
546 :
547 : // If the relation is empty, create an empty file
548 0 : if nblocks == 0 {
549 0 : let file_name = dst.to_segfile_name(0);
550 0 : let header = new_tar_header(&file_name, 0)?;
551 0 : self.ar
552 0 : .append(&header, io::empty())
553 0 : .await
554 0 : .map_err(|e| BasebackupError::Client(e, "add_rel,empty"))?;
555 0 : return Ok(());
556 0 : }
557 0 :
558 0 : // Add a file for each chunk of blocks (aka segment)
559 0 : let mut startblk = 0;
560 0 : let mut seg = 0;
561 0 : while startblk < nblocks {
562 0 : let endblk = std::cmp::min(startblk + RELSEG_SIZE, nblocks);
563 0 :
564 0 : let mut segment_data: Vec<u8> = vec![];
565 0 : for blknum in startblk..endblk {
566 0 : let img = self
567 0 : .timeline
568 0 : // TODO: investigate using get_vectored for the entire startblk..endblk range.
569 0 : // But this code path is not on the critical path for most basebackups (?).
570 0 : .get(rel_block_to_key(src, blknum), self.lsn, self.ctx)
571 0 : .await?;
572 0 : segment_data.extend_from_slice(&img[..]);
573 : }
574 :
575 0 : let file_name = dst.to_segfile_name(seg as u32);
576 0 : let header = new_tar_header(&file_name, segment_data.len() as u64)?;
577 0 : self.ar
578 0 : .append(&header, segment_data.as_slice())
579 0 : .await
580 0 : .map_err(|e| BasebackupError::Client(e, "add_rel,segment"))?;
581 :
582 0 : seg += 1;
583 0 : startblk = endblk;
584 : }
585 :
586 0 : Ok(())
587 0 : }
588 :
589 : //
590 : // Include database/tablespace directories.
591 : //
592 : // Each directory contains a PG_VERSION file, and the default database
593 : // directories also contain pg_filenode.map files.
594 : //
595 0 : async fn add_dbdir(
596 0 : &mut self,
597 0 : spcnode: u32,
598 0 : dbnode: u32,
599 0 : has_relmap_file: bool,
600 0 : ) -> Result<(), BasebackupError> {
601 0 : let relmap_img = if has_relmap_file {
602 0 : let img = self
603 0 : .timeline
604 0 : .get_relmap_file(spcnode, dbnode, Version::at(self.lsn), self.ctx)
605 0 : .await?;
606 :
607 0 : if img.len()
608 0 : != dispatch_pgversion!(self.timeline.pg_version, pgv::bindings::SIZEOF_RELMAPFILE)
609 : {
610 0 : return Err(BasebackupError::Server(anyhow!(
611 0 : "img.len() != SIZE_OF_RELMAPFILE, img.len()={}",
612 0 : img.len(),
613 0 : )));
614 0 : }
615 0 :
616 0 : Some(img)
617 : } else {
618 0 : None
619 : };
620 :
621 0 : if spcnode == GLOBALTABLESPACE_OID {
622 0 : let pg_version_str = match self.timeline.pg_version {
623 0 : 14 | 15 => self.timeline.pg_version.to_string(),
624 0 : ver => format!("{ver}\x0A"),
625 : };
626 0 : let header = new_tar_header("PG_VERSION", pg_version_str.len() as u64)?;
627 0 : self.ar
628 0 : .append(&header, pg_version_str.as_bytes())
629 0 : .await
630 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,PG_VERSION"))?;
631 :
632 0 : info!("timeline.pg_version {}", self.timeline.pg_version);
633 :
634 0 : if let Some(img) = relmap_img {
635 : // filenode map for global tablespace
636 0 : let header = new_tar_header("global/pg_filenode.map", img.len() as u64)?;
637 0 : self.ar
638 0 : .append(&header, &img[..])
639 0 : .await
640 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,global/pg_filenode.map"))?;
641 : } else {
642 0 : warn!("global/pg_filenode.map is missing");
643 : }
644 : } else {
645 : // User defined tablespaces are not supported. However, as
646 : // a special case, if a tablespace/db directory is
647 : // completely empty, we can leave it out altogether. This
648 : // makes taking a base backup after the 'tablespace'
649 : // regression test pass, because the test drops the
650 : // created tablespaces after the tests.
651 : //
652 : // FIXME: this wouldn't be necessary, if we handled
653 : // XLOG_TBLSPC_DROP records. But we probably should just
654 : // throw an error on CREATE TABLESPACE in the first place.
655 0 : if !has_relmap_file
656 0 : && self
657 0 : .timeline
658 0 : .list_rels(spcnode, dbnode, Version::at(self.lsn), self.ctx)
659 0 : .await?
660 0 : .is_empty()
661 : {
662 0 : return Ok(());
663 0 : }
664 0 : // User defined tablespaces are not supported
665 0 : if spcnode != DEFAULTTABLESPACE_OID {
666 0 : return Err(BasebackupError::Server(anyhow!(
667 0 : "spcnode != DEFAULTTABLESPACE_OID, spcnode={spcnode}"
668 0 : )));
669 0 : }
670 0 :
671 0 : // Append dir path for each database
672 0 : let path = format!("base/{}", dbnode);
673 0 : let header = new_tar_header_dir(&path)?;
674 0 : self.ar
675 0 : .append(&header, io::empty())
676 0 : .await
677 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base"))?;
678 :
679 0 : if let Some(img) = relmap_img {
680 0 : let dst_path = format!("base/{}/PG_VERSION", dbnode);
681 :
682 0 : let pg_version_str = match self.timeline.pg_version {
683 0 : 14 | 15 => self.timeline.pg_version.to_string(),
684 0 : ver => format!("{ver}\x0A"),
685 : };
686 0 : let header = new_tar_header(&dst_path, pg_version_str.len() as u64)?;
687 0 : self.ar
688 0 : .append(&header, pg_version_str.as_bytes())
689 0 : .await
690 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base/PG_VERSION"))?;
691 :
692 0 : let relmap_path = format!("base/{}/pg_filenode.map", dbnode);
693 0 : let header = new_tar_header(&relmap_path, img.len() as u64)?;
694 0 : self.ar
695 0 : .append(&header, &img[..])
696 0 : .await
697 0 : .map_err(|e| BasebackupError::Client(e, "add_dbdir,base/pg_filenode.map"))?;
698 0 : }
699 : };
700 0 : Ok(())
701 0 : }
702 :
703 : //
704 : // Extract twophase state files
705 : //
706 0 : async fn add_twophase_file(&mut self, xid: u64) -> Result<(), BasebackupError> {
707 0 : let img = self
708 0 : .timeline
709 0 : .get_twophase_file(xid, self.lsn, self.ctx)
710 0 : .await?;
711 :
712 0 : let mut buf = BytesMut::new();
713 0 : buf.extend_from_slice(&img[..]);
714 0 : let crc = crc32c::crc32c(&img[..]);
715 0 : buf.put_u32_le(crc);
716 0 : let path = if self.timeline.pg_version < 17 {
717 0 : format!("pg_twophase/{:>08X}", xid)
718 : } else {
719 0 : format!("pg_twophase/{:>016X}", xid)
720 : };
721 0 : let header = new_tar_header(&path, buf.len() as u64)?;
722 0 : self.ar
723 0 : .append(&header, &buf[..])
724 0 : .await
725 0 : .map_err(|e| BasebackupError::Client(e, "add_twophase_file"))?;
726 :
727 0 : Ok(())
728 0 : }
729 :
730 : //
731 : // Add generated pg_control file and bootstrap WAL segment.
732 : // Also send zenith.signal file with extra bootstrap data.
733 : //
734 0 : async fn add_pgcontrol_file(
735 0 : &mut self,
736 0 : pg_control_bytes: Bytes,
737 0 : system_identifier: u64,
738 0 : ) -> Result<(), BasebackupError> {
739 0 : // add zenith.signal file
740 0 : let mut zenith_signal = String::new();
741 0 : if self.prev_record_lsn == Lsn(0) {
742 0 : if self.timeline.is_ancestor_lsn(self.lsn) {
743 0 : write!(zenith_signal, "PREV LSN: none")
744 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
745 : } else {
746 0 : write!(zenith_signal, "PREV LSN: invalid")
747 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
748 : }
749 : } else {
750 0 : write!(zenith_signal, "PREV LSN: {}", self.prev_record_lsn)
751 0 : .map_err(|e| BasebackupError::Server(e.into()))?;
752 : }
753 0 : self.ar
754 0 : .append(
755 0 : &new_tar_header("zenith.signal", zenith_signal.len() as u64)?,
756 0 : zenith_signal.as_bytes(),
757 0 : )
758 0 : .await
759 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,zenith.signal"))?;
760 :
761 : //send pg_control
762 0 : let header = new_tar_header("global/pg_control", pg_control_bytes.len() as u64)?;
763 0 : self.ar
764 0 : .append(&header, &pg_control_bytes[..])
765 0 : .await
766 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,pg_control"))?;
767 :
768 : //send wal segment
769 0 : let segno = self.lsn.segment_number(WAL_SEGMENT_SIZE);
770 0 : let wal_file_name = XLogFileName(PG_TLI, segno, WAL_SEGMENT_SIZE);
771 0 : let wal_file_path = format!("pg_wal/{}", wal_file_name);
772 0 : let header = new_tar_header(&wal_file_path, WAL_SEGMENT_SIZE as u64)?;
773 :
774 0 : let wal_seg = postgres_ffi::generate_wal_segment(
775 0 : segno,
776 0 : system_identifier,
777 0 : self.timeline.pg_version,
778 0 : self.lsn,
779 0 : )
780 0 : .map_err(|e| anyhow!(e).context("Failed generating wal segment"))?;
781 0 : if wal_seg.len() != WAL_SEGMENT_SIZE {
782 0 : return Err(BasebackupError::Server(anyhow!(
783 0 : "wal_seg.len() != WAL_SEGMENT_SIZE, wal_seg.len()={}",
784 0 : wal_seg.len()
785 0 : )));
786 0 : }
787 0 : self.ar
788 0 : .append(&header, &wal_seg[..])
789 0 : .await
790 0 : .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,wal_segment"))?;
791 0 : Ok(())
792 0 : }
793 : }
794 :
795 : //
796 : // Create new tarball entry header
797 : //
798 0 : fn new_tar_header(path: &str, size: u64) -> anyhow::Result<Header> {
799 0 : let mut header = Header::new_gnu();
800 0 : header.set_size(size);
801 0 : header.set_path(path)?;
802 0 : header.set_mode(0b110000000); // -rw-------
803 0 : header.set_mtime(
804 0 : // use currenttime as last modified time
805 0 : SystemTime::now()
806 0 : .duration_since(SystemTime::UNIX_EPOCH)
807 0 : .unwrap()
808 0 : .as_secs(),
809 0 : );
810 0 : header.set_cksum();
811 0 : Ok(header)
812 0 : }
813 :
814 0 : fn new_tar_header_dir(path: &str) -> anyhow::Result<Header> {
815 0 : let mut header = Header::new_gnu();
816 0 : header.set_size(0);
817 0 : header.set_path(path)?;
818 0 : header.set_mode(0o755); // -rw-------
819 0 : header.set_entry_type(EntryType::dir());
820 0 : header.set_mtime(
821 0 : // use currenttime as last modified time
822 0 : SystemTime::now()
823 0 : .duration_since(SystemTime::UNIX_EPOCH)
824 0 : .unwrap()
825 0 : .as_secs(),
826 0 : );
827 0 : header.set_cksum();
828 0 : Ok(header)
829 0 : }
|
