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