Line data Source code
1 : //!
2 : //! Parse PostgreSQL WAL records and store them in a neon Timeline.
3 : //!
4 : //! The pipeline for ingesting WAL looks like this:
5 : //!
6 : //! WAL receiver -> [`wal_decoder`] -> WalIngest -> Repository
7 : //!
8 : //! The WAL receiver receives a stream of WAL from the WAL safekeepers.
9 : //! Records get decoded and interpreted in the [`wal_decoder`] module
10 : //! and then stored to the Repository by WalIngest.
11 : //!
12 : //! The neon Repository can store page versions in two formats: as
13 : //! page images, or a WAL records. [`wal_decoder::models::InterpretedWalRecord::from_bytes_filtered`]
14 : //! extracts page images out of some WAL records, but mostly it's WAL
15 : //! records. If a WAL record modifies multiple pages, WalIngest
16 : //! will call Repository::put_rel_wal_record or put_rel_page_image functions
17 : //! separately for each modified page.
18 : //!
19 : //! To reconstruct a page using a WAL record, the Repository calls the
20 : //! code in walredo.rs. walredo.rs passes most WAL records to the WAL
21 : //! redo Postgres process, but some records it can handle directly with
22 : //! bespoken Rust code.
23 :
24 : use std::backtrace::Backtrace;
25 : use std::collections::HashMap;
26 : use std::sync::{Arc, OnceLock};
27 : use std::time::{Duration, Instant, SystemTime};
28 :
29 : use bytes::{Buf, Bytes};
30 : use pageserver_api::key::{Key, rel_block_to_key};
31 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
32 : use pageserver_api::shard::ShardIdentity;
33 : use postgres_ffi::walrecord::*;
34 : use postgres_ffi::{
35 : PgMajorVersion, TransactionId, dispatch_pgversion, enum_pgversion, enum_pgversion_dispatch,
36 : fsm_logical_to_physical, pg_constants,
37 : };
38 : use postgres_ffi_types::TimestampTz;
39 : use postgres_ffi_types::forknum::{FSM_FORKNUM, INIT_FORKNUM, MAIN_FORKNUM, VISIBILITYMAP_FORKNUM};
40 : use tracing::*;
41 : use utils::bin_ser::{DeserializeError, SerializeError};
42 : use utils::lsn::Lsn;
43 : use utils::rate_limit::RateLimit;
44 : use utils::{critical_timeline, failpoint_support};
45 : use wal_decoder::models::record::NeonWalRecord;
46 : use wal_decoder::models::*;
47 :
48 : use crate::ZERO_PAGE;
49 : use crate::context::RequestContext;
50 : use crate::metrics::WAL_INGEST;
51 : use crate::pgdatadir_mapping::{DatadirModification, Version};
52 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
53 : use crate::tenant::{PageReconstructError, Timeline};
54 :
55 : enum_pgversion! {CheckPoint, pgv::CheckPoint}
56 :
57 : impl CheckPoint {
58 3 : fn encode(&self) -> Result<Bytes, SerializeError> {
59 3 : enum_pgversion_dispatch!(self, CheckPoint, cp, { cp.encode() })
60 3 : }
61 :
62 72917 : fn update_next_xid(&mut self, xid: u32) -> bool {
63 72917 : enum_pgversion_dispatch!(self, CheckPoint, cp, { cp.update_next_xid(xid) })
64 72917 : }
65 :
66 0 : pub fn update_next_multixid(&mut self, multi_xid: u32, multi_offset: u32) -> bool {
67 0 : enum_pgversion_dispatch!(self, CheckPoint, cp, {
68 0 : cp.update_next_multixid(multi_xid, multi_offset)
69 : })
70 0 : }
71 : }
72 :
73 : /// Temporary limitation of WAL lag warnings after attach
74 : ///
75 : /// After tenant attach, we want to limit WAL lag warnings because
76 : /// we don't look at the WAL until the attach is complete, which
77 : /// might take a while.
78 : pub struct WalLagCooldown {
79 : /// Until when should this limitation apply at all
80 : active_until: std::time::Instant,
81 : /// The maximum lag to suppress. Lags above this limit get reported anyways.
82 : max_lag: Duration,
83 : }
84 :
85 : impl WalLagCooldown {
86 0 : pub fn new(attach_start: Instant, attach_duration: Duration) -> Self {
87 0 : Self {
88 0 : active_until: attach_start + attach_duration * 3 + Duration::from_secs(120),
89 0 : max_lag: attach_duration * 2 + Duration::from_secs(60),
90 0 : }
91 0 : }
92 : }
93 :
94 : pub struct WalIngest {
95 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
96 : shard: ShardIdentity,
97 : checkpoint: CheckPoint,
98 : checkpoint_modified: bool,
99 : warn_ingest_lag: WarnIngestLag,
100 : }
101 :
102 : struct WarnIngestLag {
103 : lag_msg_ratelimit: RateLimit,
104 : future_lsn_msg_ratelimit: RateLimit,
105 : timestamp_invalid_msg_ratelimit: RateLimit,
106 : }
107 :
108 : pub struct WalIngestError {
109 : pub backtrace: std::backtrace::Backtrace,
110 : pub kind: WalIngestErrorKind,
111 : }
112 :
113 : #[derive(thiserror::Error, Debug)]
114 : pub enum WalIngestErrorKind {
115 : #[error(transparent)]
116 : #[allow(private_interfaces)]
117 : PageReconstructError(#[from] PageReconstructError),
118 : #[error(transparent)]
119 : DeserializationFailure(#[from] DeserializeError),
120 : #[error(transparent)]
121 : SerializationFailure(#[from] SerializeError),
122 : #[error("the request contains data not supported by pageserver: {0} @ {1}")]
123 : InvalidKey(Key, Lsn),
124 : #[error("twophase file for xid {0} already exists")]
125 : FileAlreadyExists(u64),
126 : #[error("slru segment {0:?}/{1} already exists")]
127 : SlruAlreadyExists(SlruKind, u32),
128 : #[error("relation already exists")]
129 : RelationAlreadyExists(RelTag),
130 : #[error("invalid reldir key {0}")]
131 : InvalidRelDirKey(Key),
132 :
133 : #[error(transparent)]
134 : LogicalError(anyhow::Error),
135 : #[error(transparent)]
136 : EncodeAuxFileError(anyhow::Error),
137 : #[error(transparent)]
138 : MaybeRelSizeV2Error(anyhow::Error),
139 :
140 : #[error("timeline shutting down")]
141 : Cancelled,
142 : }
143 :
144 : impl<T> From<T> for WalIngestError
145 : where
146 : WalIngestErrorKind: From<T>,
147 : {
148 0 : fn from(value: T) -> Self {
149 0 : WalIngestError {
150 0 : backtrace: Backtrace::capture(),
151 0 : kind: WalIngestErrorKind::from(value),
152 0 : }
153 0 : }
154 : }
155 :
156 : impl std::error::Error for WalIngestError {
157 0 : fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
158 0 : self.kind.source()
159 0 : }
160 : }
161 :
162 : impl core::fmt::Display for WalIngestError {
163 0 : fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
164 0 : self.kind.fmt(f)
165 0 : }
166 : }
167 :
168 : impl core::fmt::Debug for WalIngestError {
169 0 : fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
170 0 : if f.alternate() {
171 0 : f.debug_map()
172 0 : .key(&"backtrace")
173 0 : .value(&self.backtrace)
174 0 : .key(&"kind")
175 0 : .value(&self.kind)
176 0 : .finish()
177 : } else {
178 0 : writeln!(f, "Error: {:?}", self.kind)?;
179 0 : if self.backtrace.status() == std::backtrace::BacktraceStatus::Captured {
180 0 : writeln!(f, "Stack backtrace: {:?}", self.backtrace)?;
181 0 : }
182 0 : Ok(())
183 : }
184 0 : }
185 : }
186 :
187 : #[macro_export]
188 : macro_rules! ensure_walingest {
189 : ($($t:tt)*) => {
190 354701 : _ = || -> Result<(), anyhow::Error> {
191 354701 : anyhow::ensure!($($t)*);
192 354701 : Ok(())
193 354701 : }().map_err(WalIngestErrorKind::LogicalError)?;
194 : };
195 : }
196 :
197 : impl WalIngest {
198 6 : pub async fn new(
199 6 : timeline: &Timeline,
200 6 : startpoint: Lsn,
201 6 : ctx: &RequestContext,
202 6 : ) -> Result<WalIngest, WalIngestError> {
203 : // Fetch the latest checkpoint into memory, so that we can compare with it
204 : // quickly in `ingest_record` and update it when it changes.
205 6 : let checkpoint_bytes = timeline.get_checkpoint(startpoint, ctx).await?;
206 6 : let pgversion = timeline.pg_version;
207 :
208 6 : let checkpoint = dispatch_pgversion!(pgversion, {
209 0 : let checkpoint = pgv::CheckPoint::decode(&checkpoint_bytes)?;
210 4 : trace!("CheckPoint.nextXid = {}", checkpoint.nextXid.value);
211 0 : <pgv::CheckPoint as Into<CheckPoint>>::into(checkpoint)
212 : });
213 :
214 6 : Ok(WalIngest {
215 6 : shard: *timeline.get_shard_identity(),
216 6 : checkpoint,
217 6 : checkpoint_modified: false,
218 6 : attach_wal_lag_cooldown: timeline.attach_wal_lag_cooldown.clone(),
219 6 : warn_ingest_lag: WarnIngestLag {
220 6 : lag_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
221 6 : future_lsn_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
222 6 : timestamp_invalid_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
223 6 : },
224 6 : })
225 6 : }
226 :
227 : /// Ingest an interpreted PostgreSQL WAL record by doing writes to the underlying key value
228 : /// storage of a given timeline.
229 : ///
230 : /// This function updates `lsn` field of `DatadirModification`
231 : ///
232 : /// This function returns `true` if the record was ingested, and `false` if it was filtered out
233 72926 : pub async fn ingest_record(
234 72926 : &mut self,
235 72926 : interpreted: InterpretedWalRecord,
236 72926 : modification: &mut DatadirModification<'_>,
237 72926 : ctx: &RequestContext,
238 72926 : ) -> Result<bool, WalIngestError> {
239 72926 : WAL_INGEST.records_received.inc();
240 72926 : let prev_len = modification.len();
241 :
242 72926 : modification.set_lsn(interpreted.next_record_lsn)?;
243 :
244 72926 : if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes) {
245 : // Records of this type should always be preceded by a commit(), as they
246 : // rely on reading data pages back from the Timeline.
247 0 : assert!(!modification.has_dirty_data());
248 72926 : }
249 :
250 72926 : assert!(!self.checkpoint_modified);
251 72926 : if interpreted.xid != pg_constants::INVALID_TRANSACTION_ID
252 72917 : && self.checkpoint.update_next_xid(interpreted.xid)
253 1 : {
254 1 : self.checkpoint_modified = true;
255 72925 : }
256 :
257 72926 : failpoint_support::sleep_millis_async!("wal-ingest-record-sleep");
258 :
259 33 : match interpreted.metadata_record {
260 6 : Some(MetadataRecord::Heapam(rec)) => match rec {
261 6 : HeapamRecord::ClearVmBits(clear_vm_bits) => {
262 6 : self.ingest_clear_vm_bits(clear_vm_bits, modification, ctx)
263 6 : .await?;
264 : }
265 : },
266 0 : Some(MetadataRecord::Neonrmgr(rec)) => match rec {
267 0 : NeonrmgrRecord::ClearVmBits(clear_vm_bits) => {
268 0 : self.ingest_clear_vm_bits(clear_vm_bits, modification, ctx)
269 0 : .await?;
270 : }
271 : },
272 8 : Some(MetadataRecord::Smgr(rec)) => match rec {
273 8 : SmgrRecord::Create(create) => {
274 8 : self.ingest_xlog_smgr_create(create, modification, ctx)
275 8 : .await?;
276 : }
277 0 : SmgrRecord::Truncate(truncate) => {
278 0 : self.ingest_xlog_smgr_truncate(truncate, modification, ctx)
279 0 : .await?;
280 : }
281 : },
282 0 : Some(MetadataRecord::Dbase(rec)) => match rec {
283 0 : DbaseRecord::Create(create) => {
284 0 : self.ingest_xlog_dbase_create(create, modification, ctx)
285 0 : .await?;
286 : }
287 0 : DbaseRecord::Drop(drop) => {
288 0 : self.ingest_xlog_dbase_drop(drop, modification, ctx).await?;
289 : }
290 : },
291 0 : Some(MetadataRecord::Clog(rec)) => match rec {
292 0 : ClogRecord::ZeroPage(zero_page) => {
293 0 : self.ingest_clog_zero_page(zero_page, modification, ctx)
294 0 : .await?;
295 : }
296 0 : ClogRecord::Truncate(truncate) => {
297 0 : self.ingest_clog_truncate(truncate, modification, ctx)
298 0 : .await?;
299 : }
300 : },
301 4 : Some(MetadataRecord::Xact(rec)) => {
302 4 : self.ingest_xact_record(rec, modification, ctx).await?;
303 : }
304 0 : Some(MetadataRecord::MultiXact(rec)) => match rec {
305 0 : MultiXactRecord::ZeroPage(zero_page) => {
306 0 : self.ingest_multixact_zero_page(zero_page, modification, ctx)
307 0 : .await?;
308 : }
309 0 : MultiXactRecord::Create(create) => {
310 0 : self.ingest_multixact_create(modification, &create)?;
311 : }
312 0 : MultiXactRecord::Truncate(truncate) => {
313 0 : self.ingest_multixact_truncate(modification, &truncate, ctx)
314 0 : .await?;
315 : }
316 : },
317 0 : Some(MetadataRecord::Relmap(rec)) => match rec {
318 0 : RelmapRecord::Update(update) => {
319 0 : self.ingest_relmap_update(update, modification, ctx).await?;
320 : }
321 : },
322 15 : Some(MetadataRecord::Xlog(rec)) => match rec {
323 15 : XlogRecord::Raw(raw) => {
324 15 : self.ingest_raw_xlog_record(raw, modification, ctx).await?;
325 : }
326 : },
327 0 : Some(MetadataRecord::LogicalMessage(rec)) => match rec {
328 0 : LogicalMessageRecord::Put(put) => {
329 0 : self.ingest_logical_message_put(put, modification, ctx)
330 0 : .await?;
331 : }
332 : #[cfg(feature = "testing")]
333 : LogicalMessageRecord::Failpoint => {
334 : // This is a convenient way to make the WAL ingestion pause at
335 : // particular point in the WAL. For more fine-grained control,
336 : // we could peek into the message and only pause if it contains
337 : // a particular string, for example, but this is enough for now.
338 0 : failpoint_support::sleep_millis_async!(
339 : "pageserver-wal-ingest-logical-message-sleep"
340 : );
341 : }
342 : },
343 0 : Some(MetadataRecord::Standby(rec)) => {
344 0 : self.ingest_standby_record(rec).unwrap();
345 0 : }
346 0 : Some(MetadataRecord::Replorigin(rec)) => {
347 0 : self.ingest_replorigin_record(rec, modification).await?;
348 : }
349 72893 : None => {
350 72893 : // There are two cases through which we end up here:
351 72893 : // 1. The resource manager for the original PG WAL record
352 72893 : // is [`pg_constants::RM_TBLSPC_ID`]. This is not a supported
353 72893 : // record type within Neon.
354 72893 : // 2. The resource manager id was unknown to
355 72893 : // [`wal_decoder::decoder::MetadataRecord::from_decoded`].
356 72893 : // TODO(vlad): Tighten this up more once we build confidence
357 72893 : // that case (2) does not happen in the field.
358 72893 : }
359 : }
360 :
361 72926 : modification
362 72926 : .ingest_batch(interpreted.batch, &self.shard, ctx)
363 72926 : .await?;
364 :
365 : // If checkpoint data was updated, store the new version in the repository
366 72926 : if self.checkpoint_modified {
367 3 : let new_checkpoint_bytes = self.checkpoint.encode()?;
368 :
369 3 : modification.put_checkpoint(new_checkpoint_bytes)?;
370 3 : self.checkpoint_modified = false;
371 72923 : }
372 :
373 : // Note that at this point this record is only cached in the modification
374 : // until commit() is called to flush the data into the repository and update
375 : // the latest LSN.
376 :
377 72926 : Ok(modification.len() > prev_len)
378 72926 : }
379 :
380 : /// This is the same as AdjustToFullTransactionId(xid) in PostgreSQL
381 0 : fn adjust_to_full_transaction_id(&self, xid: TransactionId) -> Result<u64, WalIngestError> {
382 0 : let next_full_xid =
383 0 : enum_pgversion_dispatch!(&self.checkpoint, CheckPoint, cp, { cp.nextXid.value });
384 :
385 0 : let next_xid = (next_full_xid) as u32;
386 0 : let mut epoch = (next_full_xid >> 32) as u32;
387 :
388 0 : if xid > next_xid {
389 : // Wraparound occurred, must be from a prev epoch.
390 0 : if epoch == 0 {
391 0 : Err(WalIngestErrorKind::LogicalError(anyhow::anyhow!(
392 0 : "apparent XID wraparound with prepared transaction XID {xid}, nextXid is {next_full_xid}"
393 0 : )))?;
394 0 : }
395 0 : epoch -= 1;
396 0 : }
397 :
398 0 : Ok(((epoch as u64) << 32) | xid as u64)
399 0 : }
400 :
401 6 : async fn ingest_clear_vm_bits(
402 6 : &mut self,
403 6 : clear_vm_bits: ClearVmBits,
404 6 : modification: &mut DatadirModification<'_>,
405 6 : ctx: &RequestContext,
406 6 : ) -> Result<(), WalIngestError> {
407 : let ClearVmBits {
408 6 : new_heap_blkno,
409 6 : old_heap_blkno,
410 6 : flags,
411 6 : vm_rel,
412 6 : } = clear_vm_bits;
413 : // Clear the VM bits if required.
414 6 : let mut new_vm_blk = new_heap_blkno.map(pg_constants::HEAPBLK_TO_MAPBLOCK);
415 6 : let mut old_vm_blk = old_heap_blkno.map(pg_constants::HEAPBLK_TO_MAPBLOCK);
416 :
417 : // VM bits can only be cleared on the shard(s) owning the VM relation, and must be within
418 : // its view of the VM relation size. Out of caution, error instead of failing WAL ingestion,
419 : // as there has historically been cases where PostgreSQL has cleared spurious VM pages. See:
420 : // https://github.com/neondatabase/neon/pull/10634.
421 6 : let Some(vm_size) = get_relsize(modification, vm_rel, ctx).await? else {
422 0 : critical_timeline!(
423 0 : modification.tline.tenant_shard_id,
424 0 : modification.tline.timeline_id,
425 0 : "clear_vm_bits for unknown VM relation {vm_rel}"
426 : );
427 0 : return Ok(());
428 : };
429 6 : if let Some(blknum) = new_vm_blk {
430 6 : if blknum >= vm_size {
431 0 : critical_timeline!(
432 0 : modification.tline.tenant_shard_id,
433 0 : modification.tline.timeline_id,
434 0 : "new_vm_blk {blknum} not in {vm_rel} of size {vm_size}"
435 : );
436 0 : new_vm_blk = None;
437 6 : }
438 0 : }
439 6 : if let Some(blknum) = old_vm_blk {
440 0 : if blknum >= vm_size {
441 0 : critical_timeline!(
442 0 : modification.tline.tenant_shard_id,
443 0 : modification.tline.timeline_id,
444 0 : "old_vm_blk {blknum} not in {vm_rel} of size {vm_size}"
445 : );
446 0 : old_vm_blk = None;
447 0 : }
448 6 : }
449 :
450 6 : if new_vm_blk.is_none() && old_vm_blk.is_none() {
451 0 : return Ok(());
452 6 : } else if new_vm_blk == old_vm_blk {
453 : // An UPDATE record that needs to clear the bits for both old and the new page, both of
454 : // which reside on the same VM page.
455 0 : self.put_rel_wal_record(
456 0 : modification,
457 0 : vm_rel,
458 0 : new_vm_blk.unwrap(),
459 0 : NeonWalRecord::ClearVisibilityMapFlags {
460 0 : new_heap_blkno,
461 0 : old_heap_blkno,
462 0 : flags,
463 0 : },
464 0 : ctx,
465 0 : )
466 0 : .await?;
467 : } else {
468 : // Clear VM bits for one heap page, or for two pages that reside on different VM pages.
469 6 : if let Some(new_vm_blk) = new_vm_blk {
470 6 : self.put_rel_wal_record(
471 6 : modification,
472 6 : vm_rel,
473 6 : new_vm_blk,
474 6 : NeonWalRecord::ClearVisibilityMapFlags {
475 6 : new_heap_blkno,
476 6 : old_heap_blkno: None,
477 6 : flags,
478 6 : },
479 6 : ctx,
480 6 : )
481 6 : .await?;
482 0 : }
483 6 : if let Some(old_vm_blk) = old_vm_blk {
484 0 : self.put_rel_wal_record(
485 0 : modification,
486 0 : vm_rel,
487 0 : old_vm_blk,
488 0 : NeonWalRecord::ClearVisibilityMapFlags {
489 0 : new_heap_blkno: None,
490 0 : old_heap_blkno,
491 0 : flags,
492 0 : },
493 0 : ctx,
494 0 : )
495 0 : .await?;
496 6 : }
497 : }
498 6 : Ok(())
499 6 : }
500 :
501 : /// Subroutine of ingest_record(), to handle an XLOG_DBASE_CREATE record.
502 0 : async fn ingest_xlog_dbase_create(
503 0 : &mut self,
504 0 : create: DbaseCreate,
505 0 : modification: &mut DatadirModification<'_>,
506 0 : ctx: &RequestContext,
507 0 : ) -> Result<(), WalIngestError> {
508 : let DbaseCreate {
509 0 : db_id,
510 0 : tablespace_id,
511 0 : src_db_id,
512 0 : src_tablespace_id,
513 0 : } = create;
514 :
515 0 : let rels = modification
516 0 : .tline
517 0 : .list_rels(
518 0 : src_tablespace_id,
519 0 : src_db_id,
520 0 : Version::Modified(modification),
521 0 : ctx,
522 0 : )
523 0 : .await?;
524 :
525 0 : debug!("ingest_xlog_dbase_create: {} rels", rels.len());
526 :
527 : // Copy relfilemap
528 0 : let filemap = modification
529 0 : .tline
530 0 : .get_relmap_file(
531 0 : src_tablespace_id,
532 0 : src_db_id,
533 0 : Version::Modified(modification),
534 0 : ctx,
535 0 : )
536 0 : .await?;
537 0 : modification
538 0 : .put_relmap_file(tablespace_id, db_id, filemap, ctx)
539 0 : .await?;
540 :
541 0 : let mut num_rels_copied = 0;
542 0 : let mut num_blocks_copied = 0;
543 0 : for src_rel in rels {
544 0 : assert_eq!(src_rel.spcnode, src_tablespace_id);
545 0 : assert_eq!(src_rel.dbnode, src_db_id);
546 :
547 0 : let nblocks = modification
548 0 : .tline
549 0 : .get_rel_size(src_rel, Version::Modified(modification), ctx)
550 0 : .await?;
551 0 : let dst_rel = RelTag {
552 0 : spcnode: tablespace_id,
553 0 : dbnode: db_id,
554 0 : relnode: src_rel.relnode,
555 0 : forknum: src_rel.forknum,
556 0 : };
557 :
558 0 : modification.put_rel_creation(dst_rel, nblocks, ctx).await?;
559 :
560 : // Copy content
561 0 : debug!("copying rel {} to {}, {} blocks", src_rel, dst_rel, nblocks);
562 0 : for blknum in 0..nblocks {
563 : // Sharding:
564 : // - src and dst are always on the same shard, because they differ only by dbNode, and
565 : // dbNode is not included in the hash inputs for sharding.
566 : // - This WAL command is replayed on all shards, but each shard only copies the blocks
567 : // that belong to it.
568 0 : let src_key = rel_block_to_key(src_rel, blknum);
569 0 : if !self.shard.is_key_local(&src_key) {
570 0 : debug!(
571 0 : "Skipping non-local key {} during XLOG_DBASE_CREATE",
572 : src_key
573 : );
574 0 : continue;
575 0 : }
576 0 : debug!(
577 0 : "copying block {} from {} ({}) to {}",
578 : blknum, src_rel, src_key, dst_rel
579 : );
580 :
581 0 : let content = modification
582 0 : .tline
583 0 : .get_rel_page_at_lsn(
584 0 : src_rel,
585 0 : blknum,
586 0 : Version::Modified(modification),
587 0 : ctx,
588 0 : crate::tenant::storage_layer::IoConcurrency::sequential(),
589 0 : )
590 0 : .await?;
591 0 : modification.put_rel_page_image(dst_rel, blknum, content)?;
592 0 : num_blocks_copied += 1;
593 : }
594 :
595 0 : num_rels_copied += 1;
596 : }
597 :
598 0 : info!(
599 0 : "Created database {}/{}, copied {} blocks in {} rels",
600 : tablespace_id, db_id, num_blocks_copied, num_rels_copied
601 : );
602 0 : Ok(())
603 0 : }
604 :
605 0 : async fn ingest_xlog_dbase_drop(
606 0 : &mut self,
607 0 : dbase_drop: DbaseDrop,
608 0 : modification: &mut DatadirModification<'_>,
609 0 : ctx: &RequestContext,
610 0 : ) -> Result<(), WalIngestError> {
611 : let DbaseDrop {
612 0 : db_id,
613 0 : tablespace_ids,
614 0 : } = dbase_drop;
615 0 : for tablespace_id in tablespace_ids {
616 0 : trace!("Drop db {}, {}", tablespace_id, db_id);
617 0 : modification.drop_dbdir(tablespace_id, db_id, ctx).await?;
618 : }
619 :
620 0 : Ok(())
621 0 : }
622 :
623 8 : async fn ingest_xlog_smgr_create(
624 8 : &mut self,
625 8 : create: SmgrCreate,
626 8 : modification: &mut DatadirModification<'_>,
627 8 : ctx: &RequestContext,
628 8 : ) -> Result<(), WalIngestError> {
629 8 : let SmgrCreate { rel } = create;
630 8 : self.put_rel_creation(modification, rel, ctx).await?;
631 8 : Ok(())
632 8 : }
633 :
634 : /// Subroutine of ingest_record(), to handle an XLOG_SMGR_TRUNCATE record.
635 : ///
636 : /// This is the same logic as in PostgreSQL's smgr_redo() function.
637 0 : async fn ingest_xlog_smgr_truncate(
638 0 : &mut self,
639 0 : truncate: XlSmgrTruncate,
640 0 : modification: &mut DatadirModification<'_>,
641 0 : ctx: &RequestContext,
642 0 : ) -> Result<(), WalIngestError> {
643 : let XlSmgrTruncate {
644 0 : blkno,
645 0 : rnode,
646 0 : flags,
647 0 : } = truncate;
648 :
649 0 : let spcnode = rnode.spcnode;
650 0 : let dbnode = rnode.dbnode;
651 0 : let relnode = rnode.relnode;
652 :
653 0 : if flags & pg_constants::SMGR_TRUNCATE_HEAP != 0 {
654 0 : let rel = RelTag {
655 0 : spcnode,
656 0 : dbnode,
657 0 : relnode,
658 0 : forknum: MAIN_FORKNUM,
659 0 : };
660 :
661 0 : self.put_rel_truncation(modification, rel, blkno, ctx)
662 0 : .await?;
663 0 : }
664 0 : if flags & pg_constants::SMGR_TRUNCATE_FSM != 0 {
665 0 : let rel = RelTag {
666 0 : spcnode,
667 0 : dbnode,
668 0 : relnode,
669 0 : forknum: FSM_FORKNUM,
670 0 : };
671 :
672 : // Zero out the last remaining FSM page, if this shard owns it. We are not precise here,
673 : // and instead of digging in the FSM bitmap format we just clear the whole page.
674 0 : let fsm_logical_page_no = blkno / pg_constants::SLOTS_PER_FSM_PAGE;
675 0 : let mut fsm_physical_page_no = fsm_logical_to_physical(fsm_logical_page_no);
676 0 : if blkno % pg_constants::SLOTS_PER_FSM_PAGE != 0
677 0 : && self
678 0 : .shard
679 0 : .is_key_local(&rel_block_to_key(rel, fsm_physical_page_no))
680 : {
681 0 : modification.put_rel_page_image_zero(rel, fsm_physical_page_no)?;
682 0 : fsm_physical_page_no += 1;
683 0 : }
684 : // Truncate this shard's view of the FSM relation size, if it even has one.
685 0 : let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
686 0 : if nblocks > fsm_physical_page_no {
687 0 : self.put_rel_truncation(modification, rel, fsm_physical_page_no, ctx)
688 0 : .await?;
689 0 : }
690 0 : }
691 0 : if flags & pg_constants::SMGR_TRUNCATE_VM != 0 {
692 0 : let rel = RelTag {
693 0 : spcnode,
694 0 : dbnode,
695 0 : relnode,
696 0 : forknum: VISIBILITYMAP_FORKNUM,
697 0 : };
698 :
699 : // last remaining block, byte, and bit
700 0 : let mut vm_page_no = blkno / (pg_constants::VM_HEAPBLOCKS_PER_PAGE as u32);
701 0 : let trunc_byte = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_PAGE
702 0 : / pg_constants::VM_HEAPBLOCKS_PER_BYTE;
703 0 : let trunc_offs = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_BYTE
704 0 : * pg_constants::VM_BITS_PER_HEAPBLOCK;
705 :
706 : // Unless the new size is exactly at a visibility map page boundary, the
707 : // tail bits in the last remaining map page, representing truncated heap
708 : // blocks, need to be cleared. This is not only tidy, but also necessary
709 : // because we don't get a chance to clear the bits if the heap is extended
710 : // again. Only do this on the shard that owns the page.
711 0 : if (trunc_byte != 0 || trunc_offs != 0)
712 0 : && self.shard.is_key_local(&rel_block_to_key(rel, vm_page_no))
713 : {
714 0 : modification.put_rel_wal_record(
715 0 : rel,
716 0 : vm_page_no,
717 0 : NeonWalRecord::TruncateVisibilityMap {
718 0 : trunc_byte,
719 0 : trunc_offs,
720 0 : },
721 0 : )?;
722 0 : vm_page_no += 1;
723 0 : }
724 : // Truncate this shard's view of the VM relation size, if it even has one.
725 0 : let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
726 0 : if nblocks > vm_page_no {
727 0 : self.put_rel_truncation(modification, rel, vm_page_no, ctx)
728 0 : .await?;
729 0 : }
730 0 : }
731 0 : Ok(())
732 0 : }
733 :
734 4 : fn warn_on_ingest_lag(
735 4 : &mut self,
736 4 : conf: &crate::config::PageServerConf,
737 4 : wal_timestamp: TimestampTz,
738 4 : ) {
739 4 : debug_assert_current_span_has_tenant_and_timeline_id();
740 4 : let now = SystemTime::now();
741 4 : let rate_limits = &mut self.warn_ingest_lag;
742 :
743 4 : let ts = enum_pgversion_dispatch!(&self.checkpoint, CheckPoint, _cp, {
744 0 : pgv::xlog_utils::try_from_pg_timestamp(wal_timestamp)
745 : });
746 :
747 4 : match ts {
748 4 : Ok(ts) => {
749 4 : match now.duration_since(ts) {
750 4 : Ok(lag) => {
751 4 : if lag > conf.wait_lsn_timeout {
752 4 : rate_limits.lag_msg_ratelimit.call2(|rate_limit_stats| {
753 1 : if let Some(cooldown) = self.attach_wal_lag_cooldown.get() {
754 0 : if std::time::Instant::now() < cooldown.active_until && lag <= cooldown.max_lag {
755 0 : return;
756 0 : }
757 1 : } else {
758 1 : // Still loading? We shouldn't be here
759 1 : }
760 1 : let lag = humantime::format_duration(lag);
761 1 : warn!(%rate_limit_stats, %lag, "ingesting record with timestamp lagging more than wait_lsn_timeout");
762 1 : })
763 0 : }
764 : }
765 0 : Err(e) => {
766 0 : let delta_t = e.duration();
767 : // determined by prod victoriametrics query: 1000 * (timestamp(node_time_seconds{neon_service="pageserver"}) - node_time_seconds)
768 : // => https://www.robustperception.io/time-metric-from-the-node-exporter/
769 : const IGNORED_DRIFT: Duration = Duration::from_millis(100);
770 0 : if delta_t > IGNORED_DRIFT {
771 0 : let delta_t = humantime::format_duration(delta_t);
772 0 : rate_limits.future_lsn_msg_ratelimit.call2(|rate_limit_stats| {
773 0 : warn!(%rate_limit_stats, %delta_t, "ingesting record with timestamp from future");
774 0 : })
775 0 : }
776 : }
777 : };
778 : }
779 0 : Err(error) => {
780 0 : rate_limits.timestamp_invalid_msg_ratelimit.call2(|rate_limit_stats| {
781 0 : warn!(%rate_limit_stats, %error, "ingesting record with invalid timestamp, cannot calculate lag and will fail find-lsn-for-timestamp type queries");
782 0 : })
783 : }
784 : }
785 4 : }
786 :
787 : /// Subroutine of ingest_record(), to handle an XLOG_XACT_* records.
788 : ///
789 4 : async fn ingest_xact_record(
790 4 : &mut self,
791 4 : record: XactRecord,
792 4 : modification: &mut DatadirModification<'_>,
793 4 : ctx: &RequestContext,
794 4 : ) -> Result<(), WalIngestError> {
795 4 : let (xact_common, is_commit, is_prepared) = match record {
796 0 : XactRecord::Prepare(XactPrepare { xl_xid, data }) => {
797 0 : let xid: u64 = if modification.tline.pg_version >= PgMajorVersion::PG17 {
798 0 : self.adjust_to_full_transaction_id(xl_xid)?
799 : } else {
800 0 : xl_xid as u64
801 : };
802 0 : return modification.put_twophase_file(xid, data, ctx).await;
803 : }
804 4 : XactRecord::Commit(common) => (common, true, false),
805 0 : XactRecord::Abort(common) => (common, false, false),
806 0 : XactRecord::CommitPrepared(common) => (common, true, true),
807 0 : XactRecord::AbortPrepared(common) => (common, false, true),
808 : };
809 :
810 : let XactCommon {
811 4 : parsed,
812 4 : origin_id,
813 4 : xl_xid,
814 4 : lsn,
815 4 : } = xact_common;
816 :
817 : // Record update of CLOG pages
818 4 : let mut pageno = parsed.xid / pg_constants::CLOG_XACTS_PER_PAGE;
819 4 : let mut segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
820 4 : let mut rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
821 4 : let mut page_xids: Vec<TransactionId> = vec![parsed.xid];
822 :
823 4 : self.warn_on_ingest_lag(modification.tline.conf, parsed.xact_time);
824 :
825 4 : for subxact in &parsed.subxacts {
826 0 : let subxact_pageno = subxact / pg_constants::CLOG_XACTS_PER_PAGE;
827 0 : if subxact_pageno != pageno {
828 : // This subxact goes to different page. Write the record
829 : // for all the XIDs on the previous page, and continue
830 : // accumulating XIDs on this new page.
831 0 : modification.put_slru_wal_record(
832 0 : SlruKind::Clog,
833 0 : segno,
834 0 : rpageno,
835 0 : if is_commit {
836 0 : NeonWalRecord::ClogSetCommitted {
837 0 : xids: page_xids,
838 0 : timestamp: parsed.xact_time,
839 0 : }
840 : } else {
841 0 : NeonWalRecord::ClogSetAborted { xids: page_xids }
842 : },
843 0 : )?;
844 0 : page_xids = Vec::new();
845 0 : }
846 0 : pageno = subxact_pageno;
847 0 : segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
848 0 : rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
849 0 : page_xids.push(*subxact);
850 : }
851 4 : modification.put_slru_wal_record(
852 4 : SlruKind::Clog,
853 4 : segno,
854 4 : rpageno,
855 4 : if is_commit {
856 4 : NeonWalRecord::ClogSetCommitted {
857 4 : xids: page_xids,
858 4 : timestamp: parsed.xact_time,
859 4 : }
860 : } else {
861 0 : NeonWalRecord::ClogSetAborted { xids: page_xids }
862 : },
863 0 : )?;
864 :
865 : // Group relations to drop by dbNode. This map will contain all relations that _might_
866 : // exist, we will reduce it to which ones really exist later. This map can be huge if
867 : // the transaction touches a huge number of relations (there is no bound on this in
868 : // postgres).
869 4 : let mut drop_relations: HashMap<(u32, u32), Vec<RelTag>> = HashMap::new();
870 :
871 4 : for xnode in &parsed.xnodes {
872 0 : for forknum in MAIN_FORKNUM..=INIT_FORKNUM {
873 0 : let rel = RelTag {
874 0 : forknum,
875 0 : spcnode: xnode.spcnode,
876 0 : dbnode: xnode.dbnode,
877 0 : relnode: xnode.relnode,
878 0 : };
879 0 : drop_relations
880 0 : .entry((xnode.spcnode, xnode.dbnode))
881 0 : .or_default()
882 0 : .push(rel);
883 0 : }
884 : }
885 :
886 : // Execute relation drops in a batch: the number may be huge, so deleting individually is prohibitively expensive
887 4 : modification.put_rel_drops(drop_relations, ctx).await?;
888 :
889 4 : if origin_id != 0 {
890 0 : modification
891 0 : .set_replorigin(origin_id, parsed.origin_lsn)
892 0 : .await?;
893 4 : }
894 :
895 4 : if is_prepared {
896 : // Remove twophase file. see RemoveTwoPhaseFile() in postgres code
897 0 : trace!(
898 0 : "Drop twophaseFile for xid {} parsed_xact.xid {} here at {}",
899 : xl_xid, parsed.xid, lsn,
900 : );
901 :
902 0 : let xid: u64 = if modification.tline.pg_version >= PgMajorVersion::PG17 {
903 0 : self.adjust_to_full_transaction_id(parsed.xid)?
904 : } else {
905 0 : parsed.xid as u64
906 : };
907 0 : modification.drop_twophase_file(xid, ctx).await?;
908 4 : }
909 :
910 4 : Ok(())
911 4 : }
912 :
913 0 : async fn ingest_clog_truncate(
914 0 : &mut self,
915 0 : truncate: ClogTruncate,
916 0 : modification: &mut DatadirModification<'_>,
917 0 : ctx: &RequestContext,
918 0 : ) -> Result<(), WalIngestError> {
919 : let ClogTruncate {
920 0 : pageno,
921 0 : oldest_xid,
922 0 : oldest_xid_db,
923 0 : } = truncate;
924 :
925 0 : info!(
926 0 : "RM_CLOG_ID truncate pageno {} oldestXid {} oldestXidDB {}",
927 : pageno, oldest_xid, oldest_xid_db
928 : );
929 :
930 : // In Postgres, oldestXid and oldestXidDB are updated in memory when the CLOG is
931 : // truncated, but a checkpoint record with the updated values isn't written until
932 : // later. In Neon, a server can start at any LSN, not just on a checkpoint record,
933 : // so we keep the oldestXid and oldestXidDB up-to-date.
934 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
935 0 : cp.oldestXid = oldest_xid;
936 0 : cp.oldestXidDB = oldest_xid_db;
937 0 : });
938 0 : self.checkpoint_modified = true;
939 :
940 : // TODO Treat AdvanceOldestClogXid() or write a comment why we don't need it
941 :
942 0 : let latest_page_number =
943 0 : enum_pgversion_dispatch!(self.checkpoint, CheckPoint, cp, { cp.nextXid.value }) as u32
944 : / pg_constants::CLOG_XACTS_PER_PAGE;
945 :
946 : // Now delete all segments containing pages between xlrec.pageno
947 : // and latest_page_number.
948 :
949 : // First, make an important safety check:
950 : // the current endpoint page must not be eligible for removal.
951 : // See SimpleLruTruncate() in slru.c
952 0 : if dispatch_pgversion!(modification.tline.pg_version, {
953 0 : pgv::nonrelfile_utils::clogpage_precedes(latest_page_number, pageno)
954 : }) {
955 0 : info!("could not truncate directory pg_xact apparent wraparound");
956 0 : return Ok(());
957 0 : }
958 :
959 : // Iterate via SLRU CLOG segments and drop segments that we're ready to truncate
960 : //
961 : // We cannot pass 'lsn' to the Timeline.list_nonrels(), or it
962 : // will block waiting for the last valid LSN to advance up to
963 : // it. So we use the previous record's LSN in the get calls
964 : // instead.
965 0 : if modification.tline.get_shard_identity().is_shard_zero() {
966 0 : for segno in modification
967 0 : .tline
968 0 : .list_slru_segments(SlruKind::Clog, Version::Modified(modification), ctx)
969 0 : .await?
970 : {
971 0 : let segpage = segno * pg_constants::SLRU_PAGES_PER_SEGMENT;
972 :
973 0 : let may_delete = dispatch_pgversion!(modification.tline.pg_version, {
974 0 : pgv::nonrelfile_utils::slru_may_delete_clogsegment(segpage, pageno)
975 : });
976 :
977 0 : if may_delete {
978 0 : modification
979 0 : .drop_slru_segment(SlruKind::Clog, segno, ctx)
980 0 : .await?;
981 0 : trace!("Drop CLOG segment {:>04X}", segno);
982 0 : }
983 : }
984 0 : }
985 :
986 0 : Ok(())
987 0 : }
988 :
989 0 : async fn ingest_clog_zero_page(
990 0 : &mut self,
991 0 : zero_page: ClogZeroPage,
992 0 : modification: &mut DatadirModification<'_>,
993 0 : ctx: &RequestContext,
994 0 : ) -> Result<(), WalIngestError> {
995 0 : let ClogZeroPage { segno, rpageno } = zero_page;
996 :
997 0 : self.put_slru_page_image(
998 0 : modification,
999 0 : SlruKind::Clog,
1000 0 : segno,
1001 0 : rpageno,
1002 0 : ZERO_PAGE.clone(),
1003 0 : ctx,
1004 0 : )
1005 0 : .await
1006 0 : }
1007 :
1008 0 : fn ingest_multixact_create(
1009 0 : &mut self,
1010 0 : modification: &mut DatadirModification,
1011 0 : xlrec: &XlMultiXactCreate,
1012 0 : ) -> Result<(), WalIngestError> {
1013 : // Create WAL record for updating the multixact-offsets page
1014 0 : let pageno = xlrec.mid / pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32;
1015 0 : let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
1016 0 : let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
1017 :
1018 0 : modification.put_slru_wal_record(
1019 0 : SlruKind::MultiXactOffsets,
1020 0 : segno,
1021 0 : rpageno,
1022 0 : NeonWalRecord::MultixactOffsetCreate {
1023 0 : mid: xlrec.mid,
1024 0 : moff: xlrec.moff,
1025 0 : },
1026 0 : )?;
1027 :
1028 : // Create WAL records for the update of each affected multixact-members page
1029 0 : let mut members = xlrec.members.iter();
1030 0 : let mut offset = xlrec.moff;
1031 : loop {
1032 0 : let pageno = offset / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
1033 :
1034 : // How many members fit on this page?
1035 0 : let page_remain = pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32
1036 0 : - offset % pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
1037 :
1038 0 : let mut this_page_members: Vec<MultiXactMember> = Vec::new();
1039 0 : for _ in 0..page_remain {
1040 0 : if let Some(m) = members.next() {
1041 0 : this_page_members.push(m.clone());
1042 0 : } else {
1043 0 : break;
1044 : }
1045 : }
1046 0 : if this_page_members.is_empty() {
1047 : // all done
1048 0 : break;
1049 0 : }
1050 0 : let n_this_page = this_page_members.len();
1051 :
1052 0 : modification.put_slru_wal_record(
1053 0 : SlruKind::MultiXactMembers,
1054 0 : pageno / pg_constants::SLRU_PAGES_PER_SEGMENT,
1055 0 : pageno % pg_constants::SLRU_PAGES_PER_SEGMENT,
1056 0 : NeonWalRecord::MultixactMembersCreate {
1057 0 : moff: offset,
1058 0 : members: this_page_members,
1059 0 : },
1060 0 : )?;
1061 :
1062 : // Note: The multixact members can wrap around, even within one WAL record.
1063 0 : offset = offset.wrapping_add(n_this_page as u32);
1064 : }
1065 0 : let next_offset = offset;
1066 0 : assert!(xlrec.moff.wrapping_add(xlrec.nmembers) == next_offset);
1067 :
1068 : // Update next-multi-xid and next-offset
1069 : //
1070 : // NB: In PostgreSQL, the next-multi-xid stored in the control file is allowed to
1071 : // go to 0, and it's fixed up by skipping to FirstMultiXactId in functions that
1072 : // read it, like GetNewMultiXactId(). This is different from how nextXid is
1073 : // incremented! nextXid skips over < FirstNormalTransactionId when the value
1074 : // is stored, so it's never 0 in a checkpoint.
1075 : //
1076 : // I don't know why it's done that way, it seems less error-prone to skip over 0
1077 : // when the value is stored rather than when it's read. But let's do it the same
1078 : // way here.
1079 0 : let next_multi_xid = xlrec.mid.wrapping_add(1);
1080 :
1081 0 : if self
1082 0 : .checkpoint
1083 0 : .update_next_multixid(next_multi_xid, next_offset)
1084 0 : {
1085 0 : self.checkpoint_modified = true;
1086 0 : }
1087 :
1088 : // Also update the next-xid with the highest member. According to the comments in
1089 : // multixact_redo(), this shouldn't be necessary, but let's do the same here.
1090 0 : let max_mbr_xid = xlrec.members.iter().fold(None, |acc, mbr| {
1091 0 : if let Some(max_xid) = acc {
1092 0 : if mbr.xid.wrapping_sub(max_xid) as i32 > 0 {
1093 0 : Some(mbr.xid)
1094 : } else {
1095 0 : acc
1096 : }
1097 : } else {
1098 0 : Some(mbr.xid)
1099 : }
1100 0 : });
1101 :
1102 0 : if let Some(max_xid) = max_mbr_xid {
1103 0 : if self.checkpoint.update_next_xid(max_xid) {
1104 0 : self.checkpoint_modified = true;
1105 0 : }
1106 0 : }
1107 0 : Ok(())
1108 0 : }
1109 :
1110 0 : async fn ingest_multixact_truncate(
1111 0 : &mut self,
1112 0 : modification: &mut DatadirModification<'_>,
1113 0 : xlrec: &XlMultiXactTruncate,
1114 0 : ctx: &RequestContext,
1115 0 : ) -> Result<(), WalIngestError> {
1116 0 : let (maxsegment, startsegment, endsegment) =
1117 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1118 0 : cp.oldestMulti = xlrec.end_trunc_off;
1119 0 : cp.oldestMultiDB = xlrec.oldest_multi_db;
1120 0 : let maxsegment: i32 = pgv::nonrelfile_utils::mx_offset_to_member_segment(
1121 : pg_constants::MAX_MULTIXACT_OFFSET,
1122 : );
1123 0 : let startsegment: i32 =
1124 0 : pgv::nonrelfile_utils::mx_offset_to_member_segment(xlrec.start_trunc_memb);
1125 0 : let endsegment: i32 =
1126 0 : pgv::nonrelfile_utils::mx_offset_to_member_segment(xlrec.end_trunc_memb);
1127 0 : (maxsegment, startsegment, endsegment)
1128 : });
1129 :
1130 0 : self.checkpoint_modified = true;
1131 :
1132 : // PerformMembersTruncation
1133 0 : let mut segment: i32 = startsegment;
1134 :
1135 : // Delete all the segments except the last one. The last segment can still
1136 : // contain, possibly partially, valid data.
1137 0 : if modification.tline.get_shard_identity().is_shard_zero() {
1138 0 : while segment != endsegment {
1139 0 : modification
1140 0 : .drop_slru_segment(SlruKind::MultiXactMembers, segment as u32, ctx)
1141 0 : .await?;
1142 :
1143 : /* move to next segment, handling wraparound correctly */
1144 0 : if segment == maxsegment {
1145 0 : segment = 0;
1146 0 : } else {
1147 0 : segment += 1;
1148 0 : }
1149 : }
1150 0 : }
1151 :
1152 : // Truncate offsets
1153 : // FIXME: this did not handle wraparound correctly
1154 :
1155 0 : Ok(())
1156 0 : }
1157 :
1158 0 : async fn ingest_multixact_zero_page(
1159 0 : &mut self,
1160 0 : zero_page: MultiXactZeroPage,
1161 0 : modification: &mut DatadirModification<'_>,
1162 0 : ctx: &RequestContext,
1163 0 : ) -> Result<(), WalIngestError> {
1164 : let MultiXactZeroPage {
1165 0 : slru_kind,
1166 0 : segno,
1167 0 : rpageno,
1168 0 : } = zero_page;
1169 0 : self.put_slru_page_image(
1170 0 : modification,
1171 0 : slru_kind,
1172 0 : segno,
1173 0 : rpageno,
1174 0 : ZERO_PAGE.clone(),
1175 0 : ctx,
1176 0 : )
1177 0 : .await
1178 0 : }
1179 :
1180 0 : async fn ingest_relmap_update(
1181 0 : &mut self,
1182 0 : update: RelmapUpdate,
1183 0 : modification: &mut DatadirModification<'_>,
1184 0 : ctx: &RequestContext,
1185 0 : ) -> Result<(), WalIngestError> {
1186 0 : let RelmapUpdate { update, buf } = update;
1187 :
1188 0 : modification
1189 0 : .put_relmap_file(update.tsid, update.dbid, buf, ctx)
1190 0 : .await
1191 0 : }
1192 :
1193 15 : async fn ingest_raw_xlog_record(
1194 15 : &mut self,
1195 15 : raw_record: RawXlogRecord,
1196 15 : modification: &mut DatadirModification<'_>,
1197 15 : ctx: &RequestContext,
1198 15 : ) -> Result<(), WalIngestError> {
1199 15 : let RawXlogRecord { info, lsn, mut buf } = raw_record;
1200 15 : let pg_version = modification.tline.pg_version;
1201 :
1202 15 : if info == pg_constants::XLOG_PARAMETER_CHANGE {
1203 1 : if let CheckPoint::V17(cp) = &mut self.checkpoint {
1204 0 : let rec = v17::XlParameterChange::decode(&mut buf);
1205 0 : cp.wal_level = rec.wal_level;
1206 0 : self.checkpoint_modified = true;
1207 1 : }
1208 14 : } else if info == pg_constants::XLOG_END_OF_RECOVERY {
1209 0 : if let CheckPoint::V17(cp) = &mut self.checkpoint {
1210 0 : let rec = v17::XlEndOfRecovery::decode(&mut buf);
1211 0 : cp.wal_level = rec.wal_level;
1212 0 : self.checkpoint_modified = true;
1213 0 : }
1214 14 : }
1215 :
1216 15 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1217 0 : if info == pg_constants::XLOG_NEXTOID {
1218 0 : let next_oid = buf.get_u32_le();
1219 0 : if cp.nextOid != next_oid {
1220 0 : cp.nextOid = next_oid;
1221 0 : self.checkpoint_modified = true;
1222 0 : }
1223 0 : } else if info == pg_constants::XLOG_CHECKPOINT_ONLINE
1224 0 : || info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
1225 : {
1226 0 : let mut checkpoint_bytes = [0u8; pgv::xlog_utils::SIZEOF_CHECKPOINT];
1227 0 : buf.copy_to_slice(&mut checkpoint_bytes);
1228 0 : let xlog_checkpoint = pgv::CheckPoint::decode(&checkpoint_bytes)?;
1229 0 : trace!(
1230 0 : "xlog_checkpoint.oldestXid={}, checkpoint.oldestXid={}",
1231 : xlog_checkpoint.oldestXid, cp.oldestXid
1232 : );
1233 0 : if (cp.oldestXid.wrapping_sub(xlog_checkpoint.oldestXid) as i32) < 0 {
1234 0 : cp.oldestXid = xlog_checkpoint.oldestXid;
1235 0 : }
1236 0 : trace!(
1237 0 : "xlog_checkpoint.oldestActiveXid={}, checkpoint.oldestActiveXid={}",
1238 : xlog_checkpoint.oldestActiveXid, cp.oldestActiveXid
1239 : );
1240 :
1241 : // A shutdown checkpoint has `oldestActiveXid == InvalidTransactionid`,
1242 : // because at shutdown, all in-progress transactions will implicitly
1243 : // end. Postgres startup code knows that, and allows hot standby to start
1244 : // immediately from a shutdown checkpoint.
1245 : //
1246 : // In Neon, Postgres hot standby startup always behaves as if starting from
1247 : // an online checkpoint. It needs a valid `oldestActiveXid` value, so
1248 : // instead of overwriting self.checkpoint.oldestActiveXid with
1249 : // InvalidTransactionid from the checkpoint WAL record, update it to a
1250 : // proper value, knowing that there are no in-progress transactions at this
1251 : // point, except for prepared transactions.
1252 : //
1253 : // See also the neon code changes in the InitWalRecovery() function.
1254 0 : if xlog_checkpoint.oldestActiveXid == pg_constants::INVALID_TRANSACTION_ID
1255 0 : && info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
1256 : {
1257 0 : let oldest_active_xid = if pg_version >= PgMajorVersion::PG17 {
1258 0 : let mut oldest_active_full_xid = cp.nextXid.value;
1259 0 : for xid in modification.tline.list_twophase_files(lsn, ctx).await? {
1260 0 : if xid < oldest_active_full_xid {
1261 0 : oldest_active_full_xid = xid;
1262 0 : }
1263 : }
1264 0 : oldest_active_full_xid as u32
1265 : } else {
1266 0 : let mut oldest_active_xid = cp.nextXid.value as u32;
1267 0 : for xid in modification.tline.list_twophase_files(lsn, ctx).await? {
1268 0 : let narrow_xid = xid as u32;
1269 0 : if (narrow_xid.wrapping_sub(oldest_active_xid) as i32) < 0 {
1270 0 : oldest_active_xid = narrow_xid;
1271 0 : }
1272 : }
1273 0 : oldest_active_xid
1274 : };
1275 0 : cp.oldestActiveXid = oldest_active_xid;
1276 0 : } else {
1277 0 : cp.oldestActiveXid = xlog_checkpoint.oldestActiveXid;
1278 0 : }
1279 : // NB: We abuse the Checkpoint.redo field:
1280 : //
1281 : // - In PostgreSQL, the Checkpoint struct doesn't store the information
1282 : // of whether this is an online checkpoint or a shutdown checkpoint. It's
1283 : // stored in the XLOG info field of the WAL record, shutdown checkpoints
1284 : // use record type XLOG_CHECKPOINT_SHUTDOWN and online checkpoints use
1285 : // XLOG_CHECKPOINT_ONLINE. We don't store the original WAL record headers
1286 : // in the pageserver, however.
1287 : //
1288 : // - In PostgreSQL, the Checkpoint.redo field stores the *start* of the
1289 : // checkpoint record, if it's a shutdown checkpoint. But when we are
1290 : // starting from a shutdown checkpoint, the basebackup LSN is the *end*
1291 : // of the shutdown checkpoint WAL record. That makes it difficult to
1292 : // correctly detect whether we're starting from a shutdown record or
1293 : // not.
1294 : //
1295 : // To address both of those issues, we store 0 in the redo field if it's
1296 : // an online checkpoint record, and the record's *end* LSN if it's a
1297 : // shutdown checkpoint. We don't need the original redo pointer in neon,
1298 : // because we don't perform WAL replay at startup anyway, so we can get
1299 : // away with abusing the redo field like this.
1300 : //
1301 : // XXX: Ideally, we would persist the extra information in a more
1302 : // explicit format, rather than repurpose the fields of the Postgres
1303 : // struct like this. However, we already have persisted data like this,
1304 : // so we need to maintain backwards compatibility.
1305 : //
1306 : // NB: We didn't originally have this convention, so there are still old
1307 : // persisted records that didn't do this. Before, we didn't update the
1308 : // persisted redo field at all. That means that old records have a bogus
1309 : // redo pointer that points to some old value, from the checkpoint record
1310 : // that was originally imported from the data directory. If it was a
1311 : // project created in Neon, that means it points to the first checkpoint
1312 : // after initdb. That's OK for our purposes: all such old checkpoints are
1313 : // treated as old online checkpoints when the basebackup is created.
1314 0 : cp.redo = if info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN {
1315 : // Store the *end* LSN of the checkpoint record. Or to be precise,
1316 : // the start LSN of the *next* record, i.e. if the record ends
1317 : // exactly at page boundary, the redo LSN points to just after the
1318 : // page header on the next page.
1319 0 : lsn.into()
1320 : } else {
1321 0 : Lsn::INVALID.into()
1322 : };
1323 :
1324 : // Write a new checkpoint key-value pair on every checkpoint record, even
1325 : // if nothing really changed. Not strictly required, but it seems nice to
1326 : // have some trace of the checkpoint records in the layer files at the same
1327 : // LSNs.
1328 0 : self.checkpoint_modified = true;
1329 0 : }
1330 : });
1331 :
1332 15 : if info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN {
1333 1 : modification.tline.prepare_basebackup(lsn);
1334 14 : }
1335 :
1336 15 : Ok(())
1337 15 : }
1338 :
1339 0 : async fn ingest_logical_message_put(
1340 0 : &mut self,
1341 0 : put: PutLogicalMessage,
1342 0 : modification: &mut DatadirModification<'_>,
1343 0 : ctx: &RequestContext,
1344 0 : ) -> Result<(), WalIngestError> {
1345 0 : let PutLogicalMessage { path, buf } = put;
1346 0 : modification.put_file(path.as_str(), &buf, ctx).await
1347 0 : }
1348 :
1349 0 : fn ingest_standby_record(&mut self, record: StandbyRecord) -> Result<(), WalIngestError> {
1350 0 : match record {
1351 0 : StandbyRecord::RunningXacts(running_xacts) => {
1352 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1353 0 : cp.oldestActiveXid = running_xacts.oldest_running_xid;
1354 0 : });
1355 :
1356 0 : self.checkpoint_modified = true;
1357 : }
1358 : }
1359 :
1360 0 : Ok(())
1361 0 : }
1362 :
1363 0 : async fn ingest_replorigin_record(
1364 0 : &mut self,
1365 0 : record: ReploriginRecord,
1366 0 : modification: &mut DatadirModification<'_>,
1367 0 : ) -> Result<(), WalIngestError> {
1368 0 : match record {
1369 0 : ReploriginRecord::Set(set) => {
1370 0 : modification
1371 0 : .set_replorigin(set.node_id, set.remote_lsn)
1372 0 : .await?;
1373 : }
1374 0 : ReploriginRecord::Drop(drop) => {
1375 0 : modification.drop_replorigin(drop.node_id).await?;
1376 : }
1377 : }
1378 :
1379 0 : Ok(())
1380 0 : }
1381 :
1382 9 : async fn put_rel_creation(
1383 9 : &mut self,
1384 9 : modification: &mut DatadirModification<'_>,
1385 9 : rel: RelTag,
1386 9 : ctx: &RequestContext,
1387 9 : ) -> Result<(), WalIngestError> {
1388 9 : modification.put_rel_creation(rel, 0, ctx).await?;
1389 9 : Ok(())
1390 9 : }
1391 :
1392 : #[cfg(test)]
1393 136201 : async fn put_rel_page_image(
1394 136201 : &mut self,
1395 136201 : modification: &mut DatadirModification<'_>,
1396 136201 : rel: RelTag,
1397 136201 : blknum: BlockNumber,
1398 136201 : img: Bytes,
1399 136201 : ctx: &RequestContext,
1400 136201 : ) -> Result<(), WalIngestError> {
1401 136201 : self.handle_rel_extend(modification, rel, blknum, ctx)
1402 136201 : .await?;
1403 136201 : modification.put_rel_page_image(rel, blknum, img)?;
1404 136201 : Ok(())
1405 136201 : }
1406 :
1407 6 : async fn put_rel_wal_record(
1408 6 : &mut self,
1409 6 : modification: &mut DatadirModification<'_>,
1410 6 : rel: RelTag,
1411 6 : blknum: BlockNumber,
1412 6 : rec: NeonWalRecord,
1413 6 : ctx: &RequestContext,
1414 6 : ) -> Result<(), WalIngestError> {
1415 6 : self.handle_rel_extend(modification, rel, blknum, ctx)
1416 6 : .await?;
1417 6 : modification.put_rel_wal_record(rel, blknum, rec)?;
1418 6 : Ok(())
1419 6 : }
1420 :
1421 3006 : async fn put_rel_truncation(
1422 3006 : &mut self,
1423 3006 : modification: &mut DatadirModification<'_>,
1424 3006 : rel: RelTag,
1425 3006 : nblocks: BlockNumber,
1426 3006 : ctx: &RequestContext,
1427 3006 : ) -> Result<(), WalIngestError> {
1428 3006 : modification.put_rel_truncation(rel, nblocks, ctx).await?;
1429 3006 : Ok(())
1430 3006 : }
1431 :
1432 136207 : async fn handle_rel_extend(
1433 136207 : &mut self,
1434 136207 : modification: &mut DatadirModification<'_>,
1435 136207 : rel: RelTag,
1436 136207 : blknum: BlockNumber,
1437 136207 : ctx: &RequestContext,
1438 136207 : ) -> Result<(), WalIngestError> {
1439 136207 : let new_nblocks = blknum + 1;
1440 : // Check if the relation exists. We implicitly create relations on first
1441 : // record.
1442 136207 : let old_nblocks = modification.create_relation_if_required(rel, ctx).await?;
1443 :
1444 136207 : if new_nblocks > old_nblocks {
1445 : //info!("extending {} {} to {}", rel, old_nblocks, new_nblocks);
1446 136199 : modification.put_rel_extend(rel, new_nblocks, ctx).await?;
1447 :
1448 136199 : let mut key = rel_block_to_key(rel, blknum);
1449 :
1450 : // fill the gap with zeros
1451 136199 : let mut gap_blocks_filled: u64 = 0;
1452 136199 : for gap_blknum in old_nblocks..blknum {
1453 1499 : key.field6 = gap_blknum;
1454 :
1455 1499 : if self.shard.get_shard_number(&key) != self.shard.number {
1456 0 : continue;
1457 1499 : }
1458 :
1459 1499 : modification.put_rel_page_image_zero(rel, gap_blknum)?;
1460 1499 : gap_blocks_filled += 1;
1461 : }
1462 :
1463 136199 : WAL_INGEST
1464 136199 : .gap_blocks_zeroed_on_rel_extend
1465 136199 : .inc_by(gap_blocks_filled);
1466 :
1467 : // Log something when relation extends cause use to fill gaps
1468 : // with zero pages. Logging is rate limited per pg version to
1469 : // avoid skewing.
1470 136199 : if gap_blocks_filled > 0 {
1471 : use std::sync::Mutex;
1472 :
1473 : use once_cell::sync::Lazy;
1474 : use utils::rate_limit::RateLimit;
1475 :
1476 : struct RateLimitPerPgVersion {
1477 : rate_limiters: [Lazy<Mutex<RateLimit>>; 4],
1478 : }
1479 :
1480 : impl RateLimitPerPgVersion {
1481 0 : const fn new() -> Self {
1482 : Self {
1483 : rate_limiters: [const {
1484 1 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(30))))
1485 : }; 4],
1486 : }
1487 0 : }
1488 :
1489 2 : const fn rate_limiter(
1490 2 : &self,
1491 2 : pg_version: PgMajorVersion,
1492 2 : ) -> Option<&Lazy<Mutex<RateLimit>>> {
1493 : const MIN_PG_VERSION: u32 = PgMajorVersion::PG14.major_version_num();
1494 : const MAX_PG_VERSION: u32 = PgMajorVersion::PG17.major_version_num();
1495 2 : let pg_version = pg_version.major_version_num();
1496 :
1497 2 : if pg_version < MIN_PG_VERSION || pg_version > MAX_PG_VERSION {
1498 0 : return None;
1499 2 : }
1500 :
1501 2 : Some(&self.rate_limiters[(pg_version - MIN_PG_VERSION) as usize])
1502 2 : }
1503 : }
1504 :
1505 : static LOGGED: RateLimitPerPgVersion = RateLimitPerPgVersion::new();
1506 2 : if let Some(rate_limiter) = LOGGED.rate_limiter(modification.tline.pg_version) {
1507 2 : if let Ok(mut locked) = rate_limiter.try_lock() {
1508 2 : locked.call(|| {
1509 1 : info!(
1510 0 : lsn=%modification.get_lsn(),
1511 : pg_version=%modification.tline.pg_version,
1512 : rel=%rel,
1513 0 : "Filled {} gap blocks on rel extend to {} from {}",
1514 : gap_blocks_filled,
1515 : new_nblocks,
1516 : old_nblocks);
1517 1 : });
1518 0 : }
1519 0 : }
1520 136197 : }
1521 8 : }
1522 136207 : Ok(())
1523 136207 : }
1524 :
1525 0 : async fn put_slru_page_image(
1526 0 : &mut self,
1527 0 : modification: &mut DatadirModification<'_>,
1528 0 : kind: SlruKind,
1529 0 : segno: u32,
1530 0 : blknum: BlockNumber,
1531 0 : img: Bytes,
1532 0 : ctx: &RequestContext,
1533 0 : ) -> Result<(), WalIngestError> {
1534 0 : if !self.shard.is_shard_zero() {
1535 0 : return Ok(());
1536 0 : }
1537 :
1538 0 : self.handle_slru_extend(modification, kind, segno, blknum, ctx)
1539 0 : .await?;
1540 0 : modification.put_slru_page_image(kind, segno, blknum, img)?;
1541 0 : Ok(())
1542 0 : }
1543 :
1544 0 : async fn handle_slru_extend(
1545 0 : &mut self,
1546 0 : modification: &mut DatadirModification<'_>,
1547 0 : kind: SlruKind,
1548 0 : segno: u32,
1549 0 : blknum: BlockNumber,
1550 0 : ctx: &RequestContext,
1551 0 : ) -> Result<(), WalIngestError> {
1552 : // we don't use a cache for this like we do for relations. SLRUS are explcitly
1553 : // extended with ZEROPAGE records, not with commit records, so it happens
1554 : // a lot less frequently.
1555 :
1556 0 : let new_nblocks = blknum + 1;
1557 : // Check if the relation exists. We implicitly create relations on first
1558 : // record.
1559 : // TODO: would be nice if to be more explicit about it
1560 0 : let old_nblocks = if !modification
1561 0 : .tline
1562 0 : .get_slru_segment_exists(kind, segno, Version::Modified(modification), ctx)
1563 0 : .await?
1564 : {
1565 : // create it with 0 size initially, the logic below will extend it
1566 0 : modification
1567 0 : .put_slru_segment_creation(kind, segno, 0, ctx)
1568 0 : .await?;
1569 0 : 0
1570 : } else {
1571 0 : modification
1572 0 : .tline
1573 0 : .get_slru_segment_size(kind, segno, Version::Modified(modification), ctx)
1574 0 : .await?
1575 : };
1576 :
1577 0 : if new_nblocks > old_nblocks {
1578 0 : trace!(
1579 0 : "extending SLRU {:?} seg {} from {} to {} blocks",
1580 : kind, segno, old_nblocks, new_nblocks
1581 : );
1582 0 : modification.put_slru_extend(kind, segno, new_nblocks)?;
1583 :
1584 : // fill the gap with zeros
1585 0 : for gap_blknum in old_nblocks..blknum {
1586 0 : modification.put_slru_page_image_zero(kind, segno, gap_blknum)?;
1587 : }
1588 0 : }
1589 0 : Ok(())
1590 0 : }
1591 : }
1592 :
1593 : /// Returns the size of the relation as of this modification, or None if the relation doesn't exist.
1594 : ///
1595 : /// This is only accurate on shard 0. On other shards, it will return the size up to the highest
1596 : /// page number stored in the shard, or None if the shard does not have any pages for it.
1597 6 : async fn get_relsize(
1598 6 : modification: &DatadirModification<'_>,
1599 6 : rel: RelTag,
1600 6 : ctx: &RequestContext,
1601 6 : ) -> Result<Option<BlockNumber>, PageReconstructError> {
1602 6 : if !modification
1603 6 : .tline
1604 6 : .get_rel_exists(rel, Version::Modified(modification), ctx)
1605 6 : .await?
1606 : {
1607 0 : return Ok(None);
1608 6 : }
1609 6 : modification
1610 6 : .tline
1611 6 : .get_rel_size(rel, Version::Modified(modification), ctx)
1612 6 : .await
1613 6 : .map(Some)
1614 6 : }
1615 :
1616 : #[allow(clippy::bool_assert_comparison)]
1617 : #[cfg(test)]
1618 : mod tests {
1619 : use anyhow::Result;
1620 : use postgres_ffi::PgMajorVersion;
1621 : use postgres_ffi::RELSEG_SIZE;
1622 :
1623 : use super::*;
1624 : use crate::DEFAULT_PG_VERSION;
1625 : use crate::tenant::harness::*;
1626 : use crate::tenant::remote_timeline_client::{INITDB_PATH, remote_initdb_archive_path};
1627 : use crate::tenant::storage_layer::IoConcurrency;
1628 :
1629 : /// Arbitrary relation tag, for testing.
1630 : const TESTREL_A: RelTag = RelTag {
1631 : spcnode: 0,
1632 : dbnode: 111,
1633 : relnode: 1000,
1634 : forknum: 0,
1635 : };
1636 :
1637 6 : fn assert_current_logical_size(_timeline: &Timeline, _lsn: Lsn) {
1638 : // TODO
1639 6 : }
1640 :
1641 : #[tokio::test]
1642 1 : async fn test_zeroed_checkpoint_decodes_correctly() -> Result<(), anyhow::Error> {
1643 5 : for i in PgMajorVersion::ALL {
1644 4 : dispatch_pgversion!(i, {
1645 1 : pgv::CheckPoint::decode(&pgv::ZERO_CHECKPOINT)?;
1646 1 : });
1647 1 : }
1648 1 :
1649 1 : Ok(())
1650 1 : }
1651 :
1652 4 : async fn init_walingest_test(tline: &Timeline, ctx: &RequestContext) -> Result<WalIngest> {
1653 4 : let mut m = tline.begin_modification(Lsn(0x10));
1654 4 : m.put_checkpoint(dispatch_pgversion!(
1655 4 : tline.pg_version,
1656 0 : pgv::ZERO_CHECKPOINT.clone()
1657 0 : ))?;
1658 4 : m.put_relmap_file(0, 111, Bytes::from(""), ctx).await?; // dummy relmapper file
1659 4 : m.commit(ctx).await?;
1660 4 : let walingest = WalIngest::new(tline, Lsn(0x10), ctx).await?;
1661 :
1662 4 : Ok(walingest)
1663 4 : }
1664 :
1665 : #[tokio::test]
1666 1 : async fn test_relsize() -> Result<()> {
1667 1 : let (tenant, ctx) = TenantHarness::create("test_relsize").await?.load().await;
1668 1 : let io_concurrency = IoConcurrency::spawn_for_test();
1669 1 : let tline = tenant
1670 1 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
1671 1 : .await?;
1672 1 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
1673 :
1674 1 : let mut m = tline.begin_modification(Lsn(0x20));
1675 1 : walingest.put_rel_creation(&mut m, TESTREL_A, &ctx).await?;
1676 1 : walingest
1677 1 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 2"), &ctx)
1678 1 : .await?;
1679 1 : m.commit(&ctx).await?;
1680 1 : let mut m = tline.begin_modification(Lsn(0x30));
1681 1 : walingest
1682 1 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 3"), &ctx)
1683 1 : .await?;
1684 1 : m.commit(&ctx).await?;
1685 1 : let mut m = tline.begin_modification(Lsn(0x40));
1686 1 : walingest
1687 1 : .put_rel_page_image(&mut m, TESTREL_A, 1, test_img("foo blk 1 at 4"), &ctx)
1688 1 : .await?;
1689 1 : m.commit(&ctx).await?;
1690 1 : let mut m = tline.begin_modification(Lsn(0x50));
1691 1 : walingest
1692 1 : .put_rel_page_image(&mut m, TESTREL_A, 2, test_img("foo blk 2 at 5"), &ctx)
1693 1 : .await?;
1694 1 : m.commit(&ctx).await?;
1695 :
1696 1 : assert_current_logical_size(&tline, Lsn(0x50));
1697 :
1698 1 : let test_span = tracing::info_span!(parent: None, "test",
1699 0 : tenant_id=%tline.tenant_shard_id.tenant_id,
1700 0 : shard_id=%tline.tenant_shard_id.shard_slug(),
1701 0 : timeline_id=%tline.timeline_id);
1702 :
1703 : // The relation was created at LSN 2, not visible at LSN 1 yet.
1704 1 : assert_eq!(
1705 1 : tline
1706 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x10)), &ctx)
1707 1 : .await?,
1708 : false
1709 : );
1710 1 : assert!(
1711 1 : tline
1712 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x10)), &ctx)
1713 1 : .await
1714 1 : .is_err()
1715 : );
1716 1 : assert_eq!(
1717 1 : tline
1718 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
1719 1 : .await?,
1720 : true
1721 : );
1722 1 : assert_eq!(
1723 1 : tline
1724 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
1725 1 : .await?,
1726 : 1
1727 : );
1728 1 : assert_eq!(
1729 1 : tline
1730 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x50)), &ctx)
1731 1 : .await?,
1732 : 3
1733 : );
1734 :
1735 : // Check page contents at each LSN
1736 1 : assert_eq!(
1737 1 : tline
1738 1 : .get_rel_page_at_lsn(
1739 1 : TESTREL_A,
1740 1 : 0,
1741 1 : Version::at(Lsn(0x20)),
1742 1 : &ctx,
1743 1 : io_concurrency.clone()
1744 1 : )
1745 1 : .instrument(test_span.clone())
1746 1 : .await?,
1747 1 : test_img("foo blk 0 at 2")
1748 : );
1749 :
1750 1 : assert_eq!(
1751 1 : tline
1752 1 : .get_rel_page_at_lsn(
1753 1 : TESTREL_A,
1754 1 : 0,
1755 1 : Version::at(Lsn(0x30)),
1756 1 : &ctx,
1757 1 : io_concurrency.clone()
1758 1 : )
1759 1 : .instrument(test_span.clone())
1760 1 : .await?,
1761 1 : test_img("foo blk 0 at 3")
1762 : );
1763 :
1764 1 : assert_eq!(
1765 1 : tline
1766 1 : .get_rel_page_at_lsn(
1767 1 : TESTREL_A,
1768 1 : 0,
1769 1 : Version::at(Lsn(0x40)),
1770 1 : &ctx,
1771 1 : io_concurrency.clone()
1772 1 : )
1773 1 : .instrument(test_span.clone())
1774 1 : .await?,
1775 1 : test_img("foo blk 0 at 3")
1776 : );
1777 1 : assert_eq!(
1778 1 : tline
1779 1 : .get_rel_page_at_lsn(
1780 1 : TESTREL_A,
1781 1 : 1,
1782 1 : Version::at(Lsn(0x40)),
1783 1 : &ctx,
1784 1 : io_concurrency.clone()
1785 1 : )
1786 1 : .instrument(test_span.clone())
1787 1 : .await?,
1788 1 : test_img("foo blk 1 at 4")
1789 : );
1790 :
1791 1 : assert_eq!(
1792 1 : tline
1793 1 : .get_rel_page_at_lsn(
1794 1 : TESTREL_A,
1795 1 : 0,
1796 1 : Version::at(Lsn(0x50)),
1797 1 : &ctx,
1798 1 : io_concurrency.clone()
1799 1 : )
1800 1 : .instrument(test_span.clone())
1801 1 : .await?,
1802 1 : test_img("foo blk 0 at 3")
1803 : );
1804 1 : assert_eq!(
1805 1 : tline
1806 1 : .get_rel_page_at_lsn(
1807 1 : TESTREL_A,
1808 1 : 1,
1809 1 : Version::at(Lsn(0x50)),
1810 1 : &ctx,
1811 1 : io_concurrency.clone()
1812 1 : )
1813 1 : .instrument(test_span.clone())
1814 1 : .await?,
1815 1 : test_img("foo blk 1 at 4")
1816 : );
1817 1 : assert_eq!(
1818 1 : tline
1819 1 : .get_rel_page_at_lsn(
1820 1 : TESTREL_A,
1821 1 : 2,
1822 1 : Version::at(Lsn(0x50)),
1823 1 : &ctx,
1824 1 : io_concurrency.clone()
1825 1 : )
1826 1 : .instrument(test_span.clone())
1827 1 : .await?,
1828 1 : test_img("foo blk 2 at 5")
1829 : );
1830 :
1831 : // Truncate last block
1832 1 : let mut m = tline.begin_modification(Lsn(0x60));
1833 1 : walingest
1834 1 : .put_rel_truncation(&mut m, TESTREL_A, 2, &ctx)
1835 1 : .await?;
1836 1 : m.commit(&ctx).await?;
1837 1 : assert_current_logical_size(&tline, Lsn(0x60));
1838 :
1839 : // Check reported size and contents after truncation
1840 1 : assert_eq!(
1841 1 : tline
1842 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x60)), &ctx)
1843 1 : .await?,
1844 : 2
1845 : );
1846 1 : assert_eq!(
1847 1 : tline
1848 1 : .get_rel_page_at_lsn(
1849 1 : TESTREL_A,
1850 1 : 0,
1851 1 : Version::at(Lsn(0x60)),
1852 1 : &ctx,
1853 1 : io_concurrency.clone()
1854 1 : )
1855 1 : .instrument(test_span.clone())
1856 1 : .await?,
1857 1 : test_img("foo blk 0 at 3")
1858 : );
1859 1 : assert_eq!(
1860 1 : tline
1861 1 : .get_rel_page_at_lsn(
1862 1 : TESTREL_A,
1863 1 : 1,
1864 1 : Version::at(Lsn(0x60)),
1865 1 : &ctx,
1866 1 : io_concurrency.clone()
1867 1 : )
1868 1 : .instrument(test_span.clone())
1869 1 : .await?,
1870 1 : test_img("foo blk 1 at 4")
1871 : );
1872 :
1873 : // should still see the truncated block with older LSN
1874 1 : assert_eq!(
1875 1 : tline
1876 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x50)), &ctx)
1877 1 : .await?,
1878 : 3
1879 : );
1880 1 : assert_eq!(
1881 1 : tline
1882 1 : .get_rel_page_at_lsn(
1883 1 : TESTREL_A,
1884 1 : 2,
1885 1 : Version::at(Lsn(0x50)),
1886 1 : &ctx,
1887 1 : io_concurrency.clone()
1888 1 : )
1889 1 : .instrument(test_span.clone())
1890 1 : .await?,
1891 1 : test_img("foo blk 2 at 5")
1892 : );
1893 :
1894 : // Truncate to zero length
1895 1 : let mut m = tline.begin_modification(Lsn(0x68));
1896 1 : walingest
1897 1 : .put_rel_truncation(&mut m, TESTREL_A, 0, &ctx)
1898 1 : .await?;
1899 1 : m.commit(&ctx).await?;
1900 1 : assert_eq!(
1901 1 : tline
1902 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x68)), &ctx)
1903 1 : .await?,
1904 : 0
1905 : );
1906 :
1907 : // Extend from 0 to 2 blocks, leaving a gap
1908 1 : let mut m = tline.begin_modification(Lsn(0x70));
1909 1 : walingest
1910 1 : .put_rel_page_image(&mut m, TESTREL_A, 1, test_img("foo blk 1"), &ctx)
1911 1 : .await?;
1912 1 : m.commit(&ctx).await?;
1913 1 : assert_eq!(
1914 1 : tline
1915 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x70)), &ctx)
1916 1 : .await?,
1917 : 2
1918 : );
1919 1 : assert_eq!(
1920 1 : tline
1921 1 : .get_rel_page_at_lsn(
1922 1 : TESTREL_A,
1923 1 : 0,
1924 1 : Version::at(Lsn(0x70)),
1925 1 : &ctx,
1926 1 : io_concurrency.clone()
1927 1 : )
1928 1 : .instrument(test_span.clone())
1929 1 : .await?,
1930 1 : ZERO_PAGE
1931 : );
1932 1 : assert_eq!(
1933 1 : tline
1934 1 : .get_rel_page_at_lsn(
1935 1 : TESTREL_A,
1936 1 : 1,
1937 1 : Version::at(Lsn(0x70)),
1938 1 : &ctx,
1939 1 : io_concurrency.clone()
1940 1 : )
1941 1 : .instrument(test_span.clone())
1942 1 : .await?,
1943 1 : test_img("foo blk 1")
1944 : );
1945 :
1946 : // Extend a lot more, leaving a big gap that spans across segments
1947 1 : let mut m = tline.begin_modification(Lsn(0x80));
1948 1 : walingest
1949 1 : .put_rel_page_image(&mut m, TESTREL_A, 1500, test_img("foo blk 1500"), &ctx)
1950 1 : .await?;
1951 1 : m.commit(&ctx).await?;
1952 1 : assert_eq!(
1953 1 : tline
1954 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x80)), &ctx)
1955 1 : .await?,
1956 : 1501
1957 : );
1958 1499 : for blk in 2..1500 {
1959 1498 : assert_eq!(
1960 1498 : tline
1961 1498 : .get_rel_page_at_lsn(
1962 1498 : TESTREL_A,
1963 1498 : blk,
1964 1498 : Version::at(Lsn(0x80)),
1965 1498 : &ctx,
1966 1498 : io_concurrency.clone()
1967 1498 : )
1968 1498 : .instrument(test_span.clone())
1969 1498 : .await?,
1970 1498 : ZERO_PAGE
1971 : );
1972 : }
1973 1 : assert_eq!(
1974 1 : tline
1975 1 : .get_rel_page_at_lsn(
1976 1 : TESTREL_A,
1977 1 : 1500,
1978 1 : Version::at(Lsn(0x80)),
1979 1 : &ctx,
1980 1 : io_concurrency.clone()
1981 1 : )
1982 1 : .instrument(test_span.clone())
1983 1 : .await?,
1984 1 : test_img("foo blk 1500")
1985 : );
1986 :
1987 2 : Ok(())
1988 1 : }
1989 :
1990 : // Test what happens if we dropped a relation
1991 : // and then created it again within the same layer.
1992 : #[tokio::test]
1993 1 : async fn test_drop_extend() -> Result<()> {
1994 1 : let (tenant, ctx) = TenantHarness::create("test_drop_extend")
1995 1 : .await?
1996 1 : .load()
1997 1 : .await;
1998 1 : let tline = tenant
1999 1 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2000 1 : .await?;
2001 1 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
2002 :
2003 1 : let mut m = tline.begin_modification(Lsn(0x20));
2004 1 : walingest
2005 1 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 2"), &ctx)
2006 1 : .await?;
2007 1 : m.commit(&ctx).await?;
2008 :
2009 : // Check that rel exists and size is correct
2010 1 : assert_eq!(
2011 1 : tline
2012 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
2013 1 : .await?,
2014 : true
2015 : );
2016 1 : assert_eq!(
2017 1 : tline
2018 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
2019 1 : .await?,
2020 : 1
2021 : );
2022 :
2023 : // Drop rel
2024 1 : let mut m = tline.begin_modification(Lsn(0x30));
2025 1 : let mut rel_drops = HashMap::new();
2026 1 : rel_drops.insert((TESTREL_A.spcnode, TESTREL_A.dbnode), vec![TESTREL_A]);
2027 1 : m.put_rel_drops(rel_drops, &ctx).await?;
2028 1 : m.commit(&ctx).await?;
2029 :
2030 : // Check that rel is not visible anymore
2031 1 : assert_eq!(
2032 1 : tline
2033 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x30)), &ctx)
2034 1 : .await?,
2035 : false
2036 : );
2037 :
2038 : // FIXME: should fail
2039 : //assert!(tline.get_rel_size(TESTREL_A, Lsn(0x30), false)?.is_none());
2040 :
2041 : // Re-create it
2042 1 : let mut m = tline.begin_modification(Lsn(0x40));
2043 1 : walingest
2044 1 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 4"), &ctx)
2045 1 : .await?;
2046 1 : m.commit(&ctx).await?;
2047 :
2048 : // Check that rel exists and size is correct
2049 1 : assert_eq!(
2050 1 : tline
2051 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x40)), &ctx)
2052 1 : .await?,
2053 : true
2054 : );
2055 1 : assert_eq!(
2056 1 : tline
2057 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x40)), &ctx)
2058 1 : .await?,
2059 : 1
2060 : );
2061 :
2062 2 : Ok(())
2063 1 : }
2064 :
2065 : // Test what happens if we truncated a relation
2066 : // so that one of its segments was dropped
2067 : // and then extended it again within the same layer.
2068 : #[tokio::test]
2069 1 : async fn test_truncate_extend() -> Result<()> {
2070 1 : let (tenant, ctx) = TenantHarness::create("test_truncate_extend")
2071 1 : .await?
2072 1 : .load()
2073 1 : .await;
2074 1 : let io_concurrency = IoConcurrency::spawn_for_test();
2075 1 : let tline = tenant
2076 1 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2077 1 : .await?;
2078 1 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
2079 :
2080 : // Create a 20 MB relation (the size is arbitrary)
2081 1 : let relsize = 20 * 1024 * 1024 / 8192;
2082 1 : let mut m = tline.begin_modification(Lsn(0x20));
2083 2560 : for blkno in 0..relsize {
2084 2560 : let data = format!("foo blk {} at {}", blkno, Lsn(0x20));
2085 2560 : walingest
2086 2560 : .put_rel_page_image(&mut m, TESTREL_A, blkno, test_img(&data), &ctx)
2087 2560 : .await?;
2088 : }
2089 1 : m.commit(&ctx).await?;
2090 :
2091 1 : let test_span = tracing::info_span!(parent: None, "test",
2092 0 : tenant_id=%tline.tenant_shard_id.tenant_id,
2093 0 : shard_id=%tline.tenant_shard_id.shard_slug(),
2094 0 : timeline_id=%tline.timeline_id);
2095 :
2096 : // The relation was created at LSN 20, not visible at LSN 1 yet.
2097 1 : assert_eq!(
2098 1 : tline
2099 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x10)), &ctx)
2100 1 : .await?,
2101 : false
2102 : );
2103 1 : assert!(
2104 1 : tline
2105 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x10)), &ctx)
2106 1 : .await
2107 1 : .is_err()
2108 : );
2109 :
2110 1 : assert_eq!(
2111 1 : tline
2112 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
2113 1 : .await?,
2114 : true
2115 : );
2116 1 : assert_eq!(
2117 1 : tline
2118 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x20)), &ctx)
2119 1 : .await?,
2120 : relsize
2121 : );
2122 :
2123 : // Check relation content
2124 2560 : for blkno in 0..relsize {
2125 2560 : let lsn = Lsn(0x20);
2126 2560 : let data = format!("foo blk {blkno} at {lsn}");
2127 2560 : assert_eq!(
2128 2560 : tline
2129 2560 : .get_rel_page_at_lsn(
2130 2560 : TESTREL_A,
2131 2560 : blkno,
2132 2560 : Version::at(lsn),
2133 2560 : &ctx,
2134 2560 : io_concurrency.clone()
2135 2560 : )
2136 2560 : .instrument(test_span.clone())
2137 2560 : .await?,
2138 2560 : test_img(&data)
2139 : );
2140 : }
2141 :
2142 : // Truncate relation so that second segment was dropped
2143 : // - only leave one page
2144 1 : let mut m = tline.begin_modification(Lsn(0x60));
2145 1 : walingest
2146 1 : .put_rel_truncation(&mut m, TESTREL_A, 1, &ctx)
2147 1 : .await?;
2148 1 : m.commit(&ctx).await?;
2149 :
2150 : // Check reported size and contents after truncation
2151 1 : assert_eq!(
2152 1 : tline
2153 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x60)), &ctx)
2154 1 : .await?,
2155 : 1
2156 : );
2157 :
2158 2 : for blkno in 0..1 {
2159 1 : let lsn = Lsn(0x20);
2160 1 : let data = format!("foo blk {blkno} at {lsn}");
2161 1 : assert_eq!(
2162 1 : tline
2163 1 : .get_rel_page_at_lsn(
2164 1 : TESTREL_A,
2165 1 : blkno,
2166 1 : Version::at(Lsn(0x60)),
2167 1 : &ctx,
2168 1 : io_concurrency.clone()
2169 1 : )
2170 1 : .instrument(test_span.clone())
2171 1 : .await?,
2172 1 : test_img(&data)
2173 : );
2174 : }
2175 :
2176 : // should still see all blocks with older LSN
2177 1 : assert_eq!(
2178 1 : tline
2179 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x50)), &ctx)
2180 1 : .await?,
2181 : relsize
2182 : );
2183 2560 : for blkno in 0..relsize {
2184 2560 : let lsn = Lsn(0x20);
2185 2560 : let data = format!("foo blk {blkno} at {lsn}");
2186 2560 : assert_eq!(
2187 2560 : tline
2188 2560 : .get_rel_page_at_lsn(
2189 2560 : TESTREL_A,
2190 2560 : blkno,
2191 2560 : Version::at(Lsn(0x50)),
2192 2560 : &ctx,
2193 2560 : io_concurrency.clone()
2194 2560 : )
2195 2560 : .instrument(test_span.clone())
2196 2560 : .await?,
2197 2560 : test_img(&data)
2198 : );
2199 : }
2200 :
2201 : // Extend relation again.
2202 : // Add enough blocks to create second segment
2203 1 : let lsn = Lsn(0x80);
2204 1 : let mut m = tline.begin_modification(lsn);
2205 2560 : for blkno in 0..relsize {
2206 2560 : let data = format!("foo blk {blkno} at {lsn}");
2207 2560 : walingest
2208 2560 : .put_rel_page_image(&mut m, TESTREL_A, blkno, test_img(&data), &ctx)
2209 2560 : .await?;
2210 : }
2211 1 : m.commit(&ctx).await?;
2212 :
2213 1 : assert_eq!(
2214 1 : tline
2215 1 : .get_rel_exists(TESTREL_A, Version::at(Lsn(0x80)), &ctx)
2216 1 : .await?,
2217 : true
2218 : );
2219 1 : assert_eq!(
2220 1 : tline
2221 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(0x80)), &ctx)
2222 1 : .await?,
2223 : relsize
2224 : );
2225 : // Check relation content
2226 2560 : for blkno in 0..relsize {
2227 2560 : let lsn = Lsn(0x80);
2228 2560 : let data = format!("foo blk {blkno} at {lsn}");
2229 2560 : assert_eq!(
2230 2560 : tline
2231 2560 : .get_rel_page_at_lsn(
2232 2560 : TESTREL_A,
2233 2560 : blkno,
2234 2560 : Version::at(Lsn(0x80)),
2235 2560 : &ctx,
2236 2560 : io_concurrency.clone()
2237 2560 : )
2238 2560 : .instrument(test_span.clone())
2239 2560 : .await?,
2240 2560 : test_img(&data)
2241 1 : );
2242 1 : }
2243 1 :
2244 1 : Ok(())
2245 1 : }
2246 :
2247 : /// Test get_relsize() and truncation with a file larger than 1 GB, so that it's
2248 : /// split into multiple 1 GB segments in Postgres.
2249 : #[tokio::test]
2250 1 : async fn test_large_rel() -> Result<()> {
2251 1 : let (tenant, ctx) = TenantHarness::create("test_large_rel").await?.load().await;
2252 1 : let tline = tenant
2253 1 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2254 1 : .await?;
2255 1 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
2256 :
2257 1 : let mut lsn = 0x10;
2258 131073 : for blknum in 0..RELSEG_SIZE + 1 {
2259 131073 : lsn += 0x10;
2260 131073 : let mut m = tline.begin_modification(Lsn(lsn));
2261 131073 : let img = test_img(&format!("foo blk {} at {}", blknum, Lsn(lsn)));
2262 131073 : walingest
2263 131073 : .put_rel_page_image(&mut m, TESTREL_A, blknum as BlockNumber, img, &ctx)
2264 131073 : .await?;
2265 131073 : m.commit(&ctx).await?;
2266 : }
2267 :
2268 1 : assert_current_logical_size(&tline, Lsn(lsn));
2269 :
2270 1 : assert_eq!(
2271 1 : tline
2272 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(lsn)), &ctx)
2273 1 : .await?,
2274 1 : RELSEG_SIZE + 1
2275 : );
2276 :
2277 : // Truncate one block
2278 1 : lsn += 0x10;
2279 1 : let mut m = tline.begin_modification(Lsn(lsn));
2280 1 : walingest
2281 1 : .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE, &ctx)
2282 1 : .await?;
2283 1 : m.commit(&ctx).await?;
2284 1 : assert_eq!(
2285 1 : tline
2286 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(lsn)), &ctx)
2287 1 : .await?,
2288 : RELSEG_SIZE
2289 : );
2290 1 : assert_current_logical_size(&tline, Lsn(lsn));
2291 :
2292 : // Truncate another block
2293 1 : lsn += 0x10;
2294 1 : let mut m = tline.begin_modification(Lsn(lsn));
2295 1 : walingest
2296 1 : .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE - 1, &ctx)
2297 1 : .await?;
2298 1 : m.commit(&ctx).await?;
2299 1 : assert_eq!(
2300 1 : tline
2301 1 : .get_rel_size(TESTREL_A, Version::at(Lsn(lsn)), &ctx)
2302 1 : .await?,
2303 1 : RELSEG_SIZE - 1
2304 : );
2305 1 : assert_current_logical_size(&tline, Lsn(lsn));
2306 :
2307 : // Truncate to 1500, and then truncate all the way down to 0, one block at a time
2308 : // This tests the behavior at segment boundaries
2309 1 : let mut size: i32 = 3000;
2310 3002 : while size >= 0 {
2311 3001 : lsn += 0x10;
2312 3001 : let mut m = tline.begin_modification(Lsn(lsn));
2313 3001 : walingest
2314 3001 : .put_rel_truncation(&mut m, TESTREL_A, size as BlockNumber, &ctx)
2315 3001 : .await?;
2316 3001 : m.commit(&ctx).await?;
2317 3001 : assert_eq!(
2318 3001 : tline
2319 3001 : .get_rel_size(TESTREL_A, Version::at(Lsn(lsn)), &ctx)
2320 3001 : .await?,
2321 3001 : size as BlockNumber
2322 : );
2323 :
2324 3001 : size -= 1;
2325 : }
2326 1 : assert_current_logical_size(&tline, Lsn(lsn));
2327 :
2328 2 : Ok(())
2329 1 : }
2330 :
2331 : /// Replay a wal segment file taken directly from safekeepers.
2332 : ///
2333 : /// This test is useful for benchmarking since it allows us to profile only
2334 : /// the walingest code in a single-threaded executor, and iterate more quickly
2335 : /// without waiting for unrelated steps.
2336 : #[tokio::test]
2337 1 : async fn test_ingest_real_wal() {
2338 : use postgres_ffi::WAL_SEGMENT_SIZE;
2339 : use postgres_ffi::waldecoder::WalStreamDecoder;
2340 :
2341 : use crate::tenant::harness::*;
2342 :
2343 : // Define test data path and constants.
2344 : //
2345 : // Steps to reconstruct the data, if needed:
2346 : // 1. Run the pgbench python test
2347 : // 2. Take the first wal segment file from safekeeper
2348 : // 3. Compress it using `zstd --long input_file`
2349 : // 4. Copy initdb.tar.zst from local_fs_remote_storage
2350 : // 5. Grep sk logs for "restart decoder" to get startpoint
2351 : // 6. Run just the decoder from this test to get the endpoint.
2352 : // It's the last LSN the decoder will output.
2353 1 : let pg_version = PgMajorVersion::PG15; // The test data was generated by pg15
2354 1 : let path = "test_data/sk_wal_segment_from_pgbench";
2355 1 : let wal_segment_path = format!("{path}/000000010000000000000001.zst");
2356 1 : let source_initdb_path = format!("{path}/{INITDB_PATH}");
2357 1 : let startpoint = Lsn::from_hex("14AEC08").unwrap();
2358 1 : let _endpoint = Lsn::from_hex("1FFFF98").unwrap();
2359 :
2360 1 : let harness = TenantHarness::create("test_ingest_real_wal").await.unwrap();
2361 1 : let span = harness
2362 1 : .span()
2363 1 : .in_scope(|| info_span!("timeline_span", timeline_id=%TIMELINE_ID));
2364 1 : let (tenant, ctx) = harness.load().await;
2365 :
2366 1 : let remote_initdb_path =
2367 1 : remote_initdb_archive_path(&tenant.tenant_shard_id().tenant_id, &TIMELINE_ID);
2368 1 : let initdb_path = harness.remote_fs_dir.join(remote_initdb_path.get_path());
2369 :
2370 1 : std::fs::create_dir_all(initdb_path.parent().unwrap())
2371 1 : .expect("creating test dir should work");
2372 1 : std::fs::copy(source_initdb_path, initdb_path).expect("copying the initdb.tar.zst works");
2373 :
2374 : // Bootstrap a real timeline. We can't use create_test_timeline because
2375 : // it doesn't create a real checkpoint, and Walingest::new tries to parse
2376 : // the garbage data.
2377 1 : let tline = tenant
2378 1 : .bootstrap_timeline_test(TIMELINE_ID, pg_version, Some(TIMELINE_ID), &ctx)
2379 1 : .await
2380 1 : .unwrap();
2381 :
2382 : // We fully read and decompress this into memory before decoding
2383 : // to get a more accurate perf profile of the decoder.
2384 1 : let bytes = {
2385 : use async_compression::tokio::bufread::ZstdDecoder;
2386 1 : let file = tokio::fs::File::open(wal_segment_path).await.unwrap();
2387 1 : let reader = tokio::io::BufReader::new(file);
2388 1 : let decoder = ZstdDecoder::new(reader);
2389 1 : let mut reader = tokio::io::BufReader::new(decoder);
2390 1 : let mut buffer = Vec::new();
2391 1 : tokio::io::copy_buf(&mut reader, &mut buffer).await.unwrap();
2392 1 : buffer
2393 : };
2394 :
2395 : // TODO start a profiler too
2396 1 : let started_at = std::time::Instant::now();
2397 :
2398 : // Initialize walingest
2399 1 : let xlogoff: usize = startpoint.segment_offset(WAL_SEGMENT_SIZE);
2400 1 : let mut decoder = WalStreamDecoder::new(startpoint, pg_version);
2401 1 : let mut walingest = WalIngest::new(tline.as_ref(), startpoint, &ctx)
2402 1 : .await
2403 1 : .unwrap();
2404 1 : let mut modification = tline.begin_modification(startpoint);
2405 1 : println!("decoding {} bytes", bytes.len() - xlogoff);
2406 :
2407 : // Decode and ingest wal. We process the wal in chunks because
2408 : // that's what happens when we get bytes from safekeepers.
2409 237343 : for chunk in bytes[xlogoff..].chunks(50) {
2410 237343 : decoder.feed_bytes(chunk);
2411 310268 : while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() {
2412 72925 : let interpreted = InterpretedWalRecord::from_bytes_filtered(
2413 72925 : recdata,
2414 72925 : &[*modification.tline.get_shard_identity()],
2415 72925 : lsn,
2416 72925 : modification.tline.pg_version,
2417 72925 : )
2418 72925 : .unwrap()
2419 72925 : .remove(modification.tline.get_shard_identity())
2420 72925 : .unwrap();
2421 :
2422 72925 : walingest
2423 72925 : .ingest_record(interpreted, &mut modification, &ctx)
2424 72925 : .instrument(span.clone())
2425 72925 : .await
2426 72925 : .unwrap();
2427 : }
2428 237343 : modification.commit(&ctx).await.unwrap();
2429 : }
2430 :
2431 1 : let duration = started_at.elapsed();
2432 1 : println!("done in {duration:?}");
2433 1 : }
2434 : }
|
