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::collections::HashMap;
25 : use std::sync::Arc;
26 : use std::sync::OnceLock;
27 : use std::time::Duration;
28 : use std::time::Instant;
29 : use std::time::SystemTime;
30 :
31 : use pageserver_api::shard::ShardIdentity;
32 : use postgres_ffi::fsm_logical_to_physical;
33 : use postgres_ffi::walrecord::*;
34 : use postgres_ffi::{dispatch_pgversion, enum_pgversion, enum_pgversion_dispatch, TimestampTz};
35 : use wal_decoder::models::*;
36 :
37 : use anyhow::{bail, Result};
38 : use bytes::{Buf, Bytes};
39 : use tracing::*;
40 : use utils::failpoint_support;
41 : use utils::rate_limit::RateLimit;
42 :
43 : use crate::context::RequestContext;
44 : use crate::metrics::WAL_INGEST;
45 : use crate::pgdatadir_mapping::{DatadirModification, Version};
46 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
47 : use crate::tenant::PageReconstructError;
48 : use crate::tenant::Timeline;
49 : use crate::ZERO_PAGE;
50 : use pageserver_api::key::rel_block_to_key;
51 : use pageserver_api::record::NeonWalRecord;
52 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
53 : use postgres_ffi::pg_constants;
54 : use postgres_ffi::relfile_utils::{FSM_FORKNUM, INIT_FORKNUM, MAIN_FORKNUM, VISIBILITYMAP_FORKNUM};
55 : use postgres_ffi::TransactionId;
56 : use utils::bin_ser::SerializeError;
57 : use utils::lsn::Lsn;
58 :
59 : enum_pgversion! {CheckPoint, pgv::CheckPoint}
60 :
61 : impl CheckPoint {
62 6 : fn encode(&self) -> Result<Bytes, SerializeError> {
63 6 : enum_pgversion_dispatch!(self, CheckPoint, cp, { cp.encode() })
64 6 : }
65 :
66 145834 : fn update_next_xid(&mut self, xid: u32) -> bool {
67 145834 : enum_pgversion_dispatch!(self, CheckPoint, cp, { cp.update_next_xid(xid) })
68 145834 : }
69 :
70 0 : pub fn update_next_multixid(&mut self, multi_xid: u32, multi_offset: u32) -> bool {
71 0 : enum_pgversion_dispatch!(self, CheckPoint, cp, {
72 0 : cp.update_next_multixid(multi_xid, multi_offset)
73 : })
74 0 : }
75 : }
76 :
77 : /// Temporary limitation of WAL lag warnings after attach
78 : ///
79 : /// After tenant attach, we want to limit WAL lag warnings because
80 : /// we don't look at the WAL until the attach is complete, which
81 : /// might take a while.
82 : pub struct WalLagCooldown {
83 : /// Until when should this limitation apply at all
84 : active_until: std::time::Instant,
85 : /// The maximum lag to suppress. Lags above this limit get reported anyways.
86 : max_lag: Duration,
87 : }
88 :
89 : impl WalLagCooldown {
90 0 : pub fn new(attach_start: Instant, attach_duration: Duration) -> Self {
91 0 : Self {
92 0 : active_until: attach_start + attach_duration * 3 + Duration::from_secs(120),
93 0 : max_lag: attach_duration * 2 + Duration::from_secs(60),
94 0 : }
95 0 : }
96 : }
97 :
98 : pub struct WalIngest {
99 : attach_wal_lag_cooldown: Arc<OnceLock<WalLagCooldown>>,
100 : shard: ShardIdentity,
101 : checkpoint: CheckPoint,
102 : checkpoint_modified: bool,
103 : warn_ingest_lag: WarnIngestLag,
104 : }
105 :
106 : struct WarnIngestLag {
107 : lag_msg_ratelimit: RateLimit,
108 : future_lsn_msg_ratelimit: RateLimit,
109 : timestamp_invalid_msg_ratelimit: RateLimit,
110 : }
111 :
112 : impl WalIngest {
113 12 : pub async fn new(
114 12 : timeline: &Timeline,
115 12 : startpoint: Lsn,
116 12 : ctx: &RequestContext,
117 12 : ) -> anyhow::Result<WalIngest> {
118 : // Fetch the latest checkpoint into memory, so that we can compare with it
119 : // quickly in `ingest_record` and update it when it changes.
120 12 : let checkpoint_bytes = timeline.get_checkpoint(startpoint, ctx).await?;
121 12 : let pgversion = timeline.pg_version;
122 :
123 12 : let checkpoint = dispatch_pgversion!(pgversion, {
124 0 : let checkpoint = pgv::CheckPoint::decode(&checkpoint_bytes)?;
125 0 : trace!("CheckPoint.nextXid = {}", checkpoint.nextXid.value);
126 0 : <pgv::CheckPoint as Into<CheckPoint>>::into(checkpoint)
127 : });
128 :
129 12 : Ok(WalIngest {
130 12 : shard: *timeline.get_shard_identity(),
131 12 : checkpoint,
132 12 : checkpoint_modified: false,
133 12 : attach_wal_lag_cooldown: timeline.attach_wal_lag_cooldown.clone(),
134 12 : warn_ingest_lag: WarnIngestLag {
135 12 : lag_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
136 12 : future_lsn_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
137 12 : timestamp_invalid_msg_ratelimit: RateLimit::new(std::time::Duration::from_secs(10)),
138 12 : },
139 12 : })
140 12 : }
141 :
142 : /// Ingest an interpreted PostgreSQL WAL record by doing writes to the underlying key value
143 : /// storage of a given timeline.
144 : ///
145 : /// This function updates `lsn` field of `DatadirModification`
146 : ///
147 : /// This function returns `true` if the record was ingested, and `false` if it was filtered out
148 145852 : pub async fn ingest_record(
149 145852 : &mut self,
150 145852 : interpreted: InterpretedWalRecord,
151 145852 : modification: &mut DatadirModification<'_>,
152 145852 : ctx: &RequestContext,
153 145852 : ) -> anyhow::Result<bool> {
154 145852 : WAL_INGEST.records_received.inc();
155 145852 : let prev_len = modification.len();
156 145852 :
157 145852 : modification.set_lsn(interpreted.next_record_lsn)?;
158 :
159 145852 : if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes) {
160 : // Records of this type should always be preceded by a commit(), as they
161 : // rely on reading data pages back from the Timeline.
162 0 : assert!(!modification.has_dirty_data());
163 145852 : }
164 :
165 145852 : assert!(!self.checkpoint_modified);
166 145852 : if interpreted.xid != pg_constants::INVALID_TRANSACTION_ID
167 145834 : && self.checkpoint.update_next_xid(interpreted.xid)
168 2 : {
169 2 : self.checkpoint_modified = true;
170 145850 : }
171 :
172 145852 : failpoint_support::sleep_millis_async!("wal-ingest-record-sleep");
173 :
174 66 : match interpreted.metadata_record {
175 12 : Some(MetadataRecord::Heapam(rec)) => match rec {
176 12 : HeapamRecord::ClearVmBits(clear_vm_bits) => {
177 12 : self.ingest_clear_vm_bits(clear_vm_bits, modification, ctx)
178 12 : .await?;
179 : }
180 : },
181 0 : Some(MetadataRecord::Neonrmgr(rec)) => match rec {
182 0 : NeonrmgrRecord::ClearVmBits(clear_vm_bits) => {
183 0 : self.ingest_clear_vm_bits(clear_vm_bits, modification, ctx)
184 0 : .await?;
185 : }
186 : },
187 16 : Some(MetadataRecord::Smgr(rec)) => match rec {
188 16 : SmgrRecord::Create(create) => {
189 16 : self.ingest_xlog_smgr_create(create, modification, ctx)
190 16 : .await?;
191 : }
192 0 : SmgrRecord::Truncate(truncate) => {
193 0 : self.ingest_xlog_smgr_truncate(truncate, modification, ctx)
194 0 : .await?;
195 : }
196 : },
197 0 : Some(MetadataRecord::Dbase(rec)) => match rec {
198 0 : DbaseRecord::Create(create) => {
199 0 : self.ingest_xlog_dbase_create(create, modification, ctx)
200 0 : .await?;
201 : }
202 0 : DbaseRecord::Drop(drop) => {
203 0 : self.ingest_xlog_dbase_drop(drop, modification, ctx).await?;
204 : }
205 : },
206 0 : Some(MetadataRecord::Clog(rec)) => match rec {
207 0 : ClogRecord::ZeroPage(zero_page) => {
208 0 : self.ingest_clog_zero_page(zero_page, modification, ctx)
209 0 : .await?;
210 : }
211 0 : ClogRecord::Truncate(truncate) => {
212 0 : self.ingest_clog_truncate(truncate, modification, ctx)
213 0 : .await?;
214 : }
215 : },
216 8 : Some(MetadataRecord::Xact(rec)) => {
217 8 : self.ingest_xact_record(rec, modification, ctx).await?;
218 : }
219 0 : Some(MetadataRecord::MultiXact(rec)) => match rec {
220 0 : MultiXactRecord::ZeroPage(zero_page) => {
221 0 : self.ingest_multixact_zero_page(zero_page, modification, ctx)
222 0 : .await?;
223 : }
224 0 : MultiXactRecord::Create(create) => {
225 0 : self.ingest_multixact_create(modification, &create)?;
226 : }
227 0 : MultiXactRecord::Truncate(truncate) => {
228 0 : self.ingest_multixact_truncate(modification, &truncate, ctx)
229 0 : .await?;
230 : }
231 : },
232 0 : Some(MetadataRecord::Relmap(rec)) => match rec {
233 0 : RelmapRecord::Update(update) => {
234 0 : self.ingest_relmap_update(update, modification, ctx).await?;
235 : }
236 : },
237 30 : Some(MetadataRecord::Xlog(rec)) => match rec {
238 30 : XlogRecord::Raw(raw) => {
239 30 : self.ingest_raw_xlog_record(raw, modification, ctx).await?;
240 : }
241 : },
242 0 : Some(MetadataRecord::LogicalMessage(rec)) => match rec {
243 0 : LogicalMessageRecord::Put(put) => {
244 0 : self.ingest_logical_message_put(put, modification, ctx)
245 0 : .await?;
246 : }
247 : #[cfg(feature = "testing")]
248 : LogicalMessageRecord::Failpoint => {
249 : // This is a convenient way to make the WAL ingestion pause at
250 : // particular point in the WAL. For more fine-grained control,
251 : // we could peek into the message and only pause if it contains
252 : // a particular string, for example, but this is enough for now.
253 0 : failpoint_support::sleep_millis_async!(
254 0 : "pageserver-wal-ingest-logical-message-sleep"
255 0 : );
256 : }
257 : },
258 0 : Some(MetadataRecord::Standby(rec)) => {
259 0 : self.ingest_standby_record(rec).unwrap();
260 0 : }
261 0 : Some(MetadataRecord::Replorigin(rec)) => {
262 0 : self.ingest_replorigin_record(rec, modification).await?;
263 : }
264 145786 : None => {
265 145786 : // There are two cases through which we end up here:
266 145786 : // 1. The resource manager for the original PG WAL record
267 145786 : // is [`pg_constants::RM_TBLSPC_ID`]. This is not a supported
268 145786 : // record type within Neon.
269 145786 : // 2. The resource manager id was unknown to
270 145786 : // [`wal_decoder::decoder::MetadataRecord::from_decoded`].
271 145786 : // TODO(vlad): Tighten this up more once we build confidence
272 145786 : // that case (2) does not happen in the field.
273 145786 : }
274 : }
275 :
276 145852 : modification
277 145852 : .ingest_batch(interpreted.batch, &self.shard, ctx)
278 145852 : .await?;
279 :
280 : // If checkpoint data was updated, store the new version in the repository
281 145852 : if self.checkpoint_modified {
282 6 : let new_checkpoint_bytes = self.checkpoint.encode()?;
283 :
284 6 : modification.put_checkpoint(new_checkpoint_bytes)?;
285 6 : self.checkpoint_modified = false;
286 145846 : }
287 :
288 : // Note that at this point this record is only cached in the modification
289 : // until commit() is called to flush the data into the repository and update
290 : // the latest LSN.
291 :
292 145852 : Ok(modification.len() > prev_len)
293 145852 : }
294 :
295 : /// This is the same as AdjustToFullTransactionId(xid) in PostgreSQL
296 0 : fn adjust_to_full_transaction_id(&self, xid: TransactionId) -> Result<u64> {
297 0 : let next_full_xid =
298 0 : enum_pgversion_dispatch!(&self.checkpoint, CheckPoint, cp, { cp.nextXid.value });
299 :
300 0 : let next_xid = (next_full_xid) as u32;
301 0 : let mut epoch = (next_full_xid >> 32) as u32;
302 0 :
303 0 : if xid > next_xid {
304 : // Wraparound occurred, must be from a prev epoch.
305 0 : if epoch == 0 {
306 0 : bail!("apparent XID wraparound with prepared transaction XID {xid}, nextXid is {next_full_xid}");
307 0 : }
308 0 : epoch -= 1;
309 0 : }
310 :
311 0 : Ok((epoch as u64) << 32 | xid as u64)
312 0 : }
313 :
314 12 : async fn ingest_clear_vm_bits(
315 12 : &mut self,
316 12 : clear_vm_bits: ClearVmBits,
317 12 : modification: &mut DatadirModification<'_>,
318 12 : ctx: &RequestContext,
319 12 : ) -> anyhow::Result<()> {
320 12 : let ClearVmBits {
321 12 : new_heap_blkno,
322 12 : old_heap_blkno,
323 12 : flags,
324 12 : vm_rel,
325 12 : } = clear_vm_bits;
326 12 : // Clear the VM bits if required.
327 12 : let mut new_vm_blk = new_heap_blkno.map(pg_constants::HEAPBLK_TO_MAPBLOCK);
328 12 : let mut old_vm_blk = old_heap_blkno.map(pg_constants::HEAPBLK_TO_MAPBLOCK);
329 :
330 : // Sometimes, Postgres seems to create heap WAL records with the
331 : // ALL_VISIBLE_CLEARED flag set, even though the bit in the VM page is
332 : // not set. In fact, it's possible that the VM page does not exist at all.
333 : // In that case, we don't want to store a record to clear the VM bit;
334 : // replaying it would fail to find the previous image of the page, because
335 : // it doesn't exist. So check if the VM page(s) exist, and skip the WAL
336 : // record if it doesn't.
337 : //
338 : // TODO: analyze the metrics and tighten this up accordingly. This logic
339 : // implicitly assumes that VM pages see explicit WAL writes before
340 : // implicit ClearVmBits, and will otherwise silently drop updates.
341 12 : let Some(vm_size) = get_relsize(modification, vm_rel, ctx).await? else {
342 0 : WAL_INGEST
343 0 : .clear_vm_bits_unknown
344 0 : .with_label_values(&["relation"])
345 0 : .inc();
346 0 : return Ok(());
347 : };
348 12 : if let Some(blknum) = new_vm_blk {
349 12 : if blknum >= vm_size {
350 0 : WAL_INGEST
351 0 : .clear_vm_bits_unknown
352 0 : .with_label_values(&["new_page"])
353 0 : .inc();
354 0 : new_vm_blk = None;
355 12 : }
356 0 : }
357 12 : if let Some(blknum) = old_vm_blk {
358 0 : if blknum >= vm_size {
359 0 : WAL_INGEST
360 0 : .clear_vm_bits_unknown
361 0 : .with_label_values(&["old_page"])
362 0 : .inc();
363 0 : old_vm_blk = None;
364 0 : }
365 12 : }
366 :
367 12 : if new_vm_blk.is_some() || old_vm_blk.is_some() {
368 12 : if new_vm_blk == old_vm_blk {
369 : // An UPDATE record that needs to clear the bits for both old and the
370 : // new page, both of which reside on the same VM page.
371 0 : self.put_rel_wal_record(
372 0 : modification,
373 0 : vm_rel,
374 0 : new_vm_blk.unwrap(),
375 0 : NeonWalRecord::ClearVisibilityMapFlags {
376 0 : new_heap_blkno,
377 0 : old_heap_blkno,
378 0 : flags,
379 0 : },
380 0 : ctx,
381 0 : )
382 0 : .await?;
383 : } else {
384 : // Clear VM bits for one heap page, or for two pages that reside on
385 : // different VM pages.
386 12 : if let Some(new_vm_blk) = new_vm_blk {
387 12 : self.put_rel_wal_record(
388 12 : modification,
389 12 : vm_rel,
390 12 : new_vm_blk,
391 12 : NeonWalRecord::ClearVisibilityMapFlags {
392 12 : new_heap_blkno,
393 12 : old_heap_blkno: None,
394 12 : flags,
395 12 : },
396 12 : ctx,
397 12 : )
398 12 : .await?;
399 0 : }
400 12 : if let Some(old_vm_blk) = old_vm_blk {
401 0 : self.put_rel_wal_record(
402 0 : modification,
403 0 : vm_rel,
404 0 : old_vm_blk,
405 0 : NeonWalRecord::ClearVisibilityMapFlags {
406 0 : new_heap_blkno: None,
407 0 : old_heap_blkno,
408 0 : flags,
409 0 : },
410 0 : ctx,
411 0 : )
412 0 : .await?;
413 12 : }
414 : }
415 0 : }
416 :
417 12 : Ok(())
418 12 : }
419 :
420 : /// Subroutine of ingest_record(), to handle an XLOG_DBASE_CREATE record.
421 0 : async fn ingest_xlog_dbase_create(
422 0 : &mut self,
423 0 : create: DbaseCreate,
424 0 : modification: &mut DatadirModification<'_>,
425 0 : ctx: &RequestContext,
426 0 : ) -> anyhow::Result<()> {
427 0 : let DbaseCreate {
428 0 : db_id,
429 0 : tablespace_id,
430 0 : src_db_id,
431 0 : src_tablespace_id,
432 0 : } = create;
433 :
434 0 : let rels = modification
435 0 : .tline
436 0 : .list_rels(
437 0 : src_tablespace_id,
438 0 : src_db_id,
439 0 : Version::Modified(modification),
440 0 : ctx,
441 0 : )
442 0 : .await?;
443 :
444 0 : debug!("ingest_xlog_dbase_create: {} rels", rels.len());
445 :
446 : // Copy relfilemap
447 0 : let filemap = modification
448 0 : .tline
449 0 : .get_relmap_file(
450 0 : src_tablespace_id,
451 0 : src_db_id,
452 0 : Version::Modified(modification),
453 0 : ctx,
454 0 : )
455 0 : .await?;
456 0 : modification
457 0 : .put_relmap_file(tablespace_id, db_id, filemap, ctx)
458 0 : .await?;
459 :
460 0 : let mut num_rels_copied = 0;
461 0 : let mut num_blocks_copied = 0;
462 0 : for src_rel in rels {
463 0 : assert_eq!(src_rel.spcnode, src_tablespace_id);
464 0 : assert_eq!(src_rel.dbnode, src_db_id);
465 :
466 0 : let nblocks = modification
467 0 : .tline
468 0 : .get_rel_size(src_rel, Version::Modified(modification), ctx)
469 0 : .await?;
470 0 : let dst_rel = RelTag {
471 0 : spcnode: tablespace_id,
472 0 : dbnode: db_id,
473 0 : relnode: src_rel.relnode,
474 0 : forknum: src_rel.forknum,
475 0 : };
476 0 :
477 0 : modification.put_rel_creation(dst_rel, nblocks, ctx).await?;
478 :
479 : // Copy content
480 0 : debug!("copying rel {} to {}, {} blocks", src_rel, dst_rel, nblocks);
481 0 : for blknum in 0..nblocks {
482 : // Sharding:
483 : // - src and dst are always on the same shard, because they differ only by dbNode, and
484 : // dbNode is not included in the hash inputs for sharding.
485 : // - This WAL command is replayed on all shards, but each shard only copies the blocks
486 : // that belong to it.
487 0 : let src_key = rel_block_to_key(src_rel, blknum);
488 0 : if !self.shard.is_key_local(&src_key) {
489 0 : debug!(
490 0 : "Skipping non-local key {} during XLOG_DBASE_CREATE",
491 : src_key
492 : );
493 0 : continue;
494 0 : }
495 0 : debug!(
496 0 : "copying block {} from {} ({}) to {}",
497 : blknum, src_rel, src_key, dst_rel
498 : );
499 :
500 0 : let content = modification
501 0 : .tline
502 0 : .get_rel_page_at_lsn(src_rel, blknum, Version::Modified(modification), ctx)
503 0 : .await?;
504 0 : modification.put_rel_page_image(dst_rel, blknum, content)?;
505 0 : num_blocks_copied += 1;
506 : }
507 :
508 0 : num_rels_copied += 1;
509 : }
510 :
511 0 : info!(
512 0 : "Created database {}/{}, copied {} blocks in {} rels",
513 : tablespace_id, db_id, num_blocks_copied, num_rels_copied
514 : );
515 0 : Ok(())
516 0 : }
517 :
518 0 : async fn ingest_xlog_dbase_drop(
519 0 : &mut self,
520 0 : dbase_drop: DbaseDrop,
521 0 : modification: &mut DatadirModification<'_>,
522 0 : ctx: &RequestContext,
523 0 : ) -> anyhow::Result<()> {
524 0 : let DbaseDrop {
525 0 : db_id,
526 0 : tablespace_ids,
527 0 : } = dbase_drop;
528 0 : for tablespace_id in tablespace_ids {
529 0 : trace!("Drop db {}, {}", tablespace_id, db_id);
530 0 : modification.drop_dbdir(tablespace_id, db_id, ctx).await?;
531 : }
532 :
533 0 : Ok(())
534 0 : }
535 :
536 16 : async fn ingest_xlog_smgr_create(
537 16 : &mut self,
538 16 : create: SmgrCreate,
539 16 : modification: &mut DatadirModification<'_>,
540 16 : ctx: &RequestContext,
541 16 : ) -> anyhow::Result<()> {
542 16 : let SmgrCreate { rel } = create;
543 16 : self.put_rel_creation(modification, rel, ctx).await?;
544 16 : Ok(())
545 16 : }
546 :
547 : /// Subroutine of ingest_record(), to handle an XLOG_SMGR_TRUNCATE record.
548 : ///
549 : /// This is the same logic as in PostgreSQL's smgr_redo() function.
550 0 : async fn ingest_xlog_smgr_truncate(
551 0 : &mut self,
552 0 : truncate: XlSmgrTruncate,
553 0 : modification: &mut DatadirModification<'_>,
554 0 : ctx: &RequestContext,
555 0 : ) -> anyhow::Result<()> {
556 0 : let XlSmgrTruncate {
557 0 : blkno,
558 0 : rnode,
559 0 : flags,
560 0 : } = truncate;
561 0 :
562 0 : let spcnode = rnode.spcnode;
563 0 : let dbnode = rnode.dbnode;
564 0 : let relnode = rnode.relnode;
565 0 :
566 0 : if flags & pg_constants::SMGR_TRUNCATE_HEAP != 0 {
567 0 : let rel = RelTag {
568 0 : spcnode,
569 0 : dbnode,
570 0 : relnode,
571 0 : forknum: MAIN_FORKNUM,
572 0 : };
573 0 :
574 0 : self.put_rel_truncation(modification, rel, blkno, ctx)
575 0 : .await?;
576 0 : }
577 0 : if flags & pg_constants::SMGR_TRUNCATE_FSM != 0 {
578 0 : let rel = RelTag {
579 0 : spcnode,
580 0 : dbnode,
581 0 : relnode,
582 0 : forknum: FSM_FORKNUM,
583 0 : };
584 0 :
585 0 : // Zero out the last remaining FSM page, if this shard owns it. We are not precise here,
586 0 : // and instead of digging in the FSM bitmap format we just clear the whole page.
587 0 : let fsm_logical_page_no = blkno / pg_constants::SLOTS_PER_FSM_PAGE;
588 0 : let mut fsm_physical_page_no = fsm_logical_to_physical(fsm_logical_page_no);
589 0 : if blkno % pg_constants::SLOTS_PER_FSM_PAGE != 0
590 0 : && self
591 0 : .shard
592 0 : .is_key_local(&rel_block_to_key(rel, fsm_physical_page_no))
593 : {
594 0 : modification.put_rel_page_image_zero(rel, fsm_physical_page_no)?;
595 0 : fsm_physical_page_no += 1;
596 0 : }
597 : // Truncate this shard's view of the FSM relation size, if it even has one.
598 0 : let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
599 0 : if nblocks > fsm_physical_page_no {
600 0 : self.put_rel_truncation(modification, rel, fsm_physical_page_no, ctx)
601 0 : .await?;
602 0 : }
603 0 : }
604 0 : if flags & pg_constants::SMGR_TRUNCATE_VM != 0 {
605 0 : let rel = RelTag {
606 0 : spcnode,
607 0 : dbnode,
608 0 : relnode,
609 0 : forknum: VISIBILITYMAP_FORKNUM,
610 0 : };
611 0 :
612 0 : // last remaining block, byte, and bit
613 0 : let mut vm_page_no = blkno / (pg_constants::VM_HEAPBLOCKS_PER_PAGE as u32);
614 0 : let trunc_byte = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_PAGE
615 0 : / pg_constants::VM_HEAPBLOCKS_PER_BYTE;
616 0 : let trunc_offs = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_BYTE
617 0 : * pg_constants::VM_BITS_PER_HEAPBLOCK;
618 0 :
619 0 : // Unless the new size is exactly at a visibility map page boundary, the
620 0 : // tail bits in the last remaining map page, representing truncated heap
621 0 : // blocks, need to be cleared. This is not only tidy, but also necessary
622 0 : // because we don't get a chance to clear the bits if the heap is extended
623 0 : // again. Only do this on the shard that owns the page.
624 0 : if (trunc_byte != 0 || trunc_offs != 0)
625 0 : && self.shard.is_key_local(&rel_block_to_key(rel, vm_page_no))
626 : {
627 0 : modification.put_rel_wal_record(
628 0 : rel,
629 0 : vm_page_no,
630 0 : NeonWalRecord::TruncateVisibilityMap {
631 0 : trunc_byte,
632 0 : trunc_offs,
633 0 : },
634 0 : )?;
635 0 : vm_page_no += 1;
636 0 : }
637 : // Truncate this shard's view of the VM relation size, if it even has one.
638 0 : let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
639 0 : if nblocks > vm_page_no {
640 0 : self.put_rel_truncation(modification, rel, vm_page_no, ctx)
641 0 : .await?;
642 0 : }
643 0 : }
644 0 : Ok(())
645 0 : }
646 :
647 8 : fn warn_on_ingest_lag(
648 8 : &mut self,
649 8 : conf: &crate::config::PageServerConf,
650 8 : wal_timestamp: TimestampTz,
651 8 : ) {
652 8 : debug_assert_current_span_has_tenant_and_timeline_id();
653 8 : let now = SystemTime::now();
654 8 : let rate_limits = &mut self.warn_ingest_lag;
655 :
656 8 : let ts = enum_pgversion_dispatch!(&self.checkpoint, CheckPoint, _cp, {
657 0 : pgv::xlog_utils::try_from_pg_timestamp(wal_timestamp)
658 : });
659 :
660 8 : match ts {
661 8 : Ok(ts) => {
662 8 : match now.duration_since(ts) {
663 8 : Ok(lag) => {
664 8 : if lag > conf.wait_lsn_timeout {
665 8 : rate_limits.lag_msg_ratelimit.call2(|rate_limit_stats| {
666 2 : if let Some(cooldown) = self.attach_wal_lag_cooldown.get() {
667 0 : if std::time::Instant::now() < cooldown.active_until && lag <= cooldown.max_lag {
668 0 : return;
669 0 : }
670 2 : } else {
671 2 : // Still loading? We shouldn't be here
672 2 : }
673 2 : let lag = humantime::format_duration(lag);
674 2 : warn!(%rate_limit_stats, %lag, "ingesting record with timestamp lagging more than wait_lsn_timeout");
675 8 : })
676 0 : }
677 : }
678 0 : Err(e) => {
679 0 : let delta_t = e.duration();
680 : // determined by prod victoriametrics query: 1000 * (timestamp(node_time_seconds{neon_service="pageserver"}) - node_time_seconds)
681 : // => https://www.robustperception.io/time-metric-from-the-node-exporter/
682 : const IGNORED_DRIFT: Duration = Duration::from_millis(100);
683 0 : if delta_t > IGNORED_DRIFT {
684 0 : let delta_t = humantime::format_duration(delta_t);
685 0 : rate_limits.future_lsn_msg_ratelimit.call2(|rate_limit_stats| {
686 0 : warn!(%rate_limit_stats, %delta_t, "ingesting record with timestamp from future");
687 0 : })
688 0 : }
689 : }
690 : };
691 : }
692 0 : Err(error) => {
693 0 : rate_limits.timestamp_invalid_msg_ratelimit.call2(|rate_limit_stats| {
694 0 : warn!(%rate_limit_stats, %error, "ingesting record with invalid timestamp, cannot calculate lag and will fail find-lsn-for-timestamp type queries");
695 0 : })
696 : }
697 : }
698 8 : }
699 :
700 : /// Subroutine of ingest_record(), to handle an XLOG_XACT_* records.
701 : ///
702 8 : async fn ingest_xact_record(
703 8 : &mut self,
704 8 : record: XactRecord,
705 8 : modification: &mut DatadirModification<'_>,
706 8 : ctx: &RequestContext,
707 8 : ) -> anyhow::Result<()> {
708 8 : let (xact_common, is_commit, is_prepared) = match record {
709 0 : XactRecord::Prepare(XactPrepare { xl_xid, data }) => {
710 0 : let xid: u64 = if modification.tline.pg_version >= 17 {
711 0 : self.adjust_to_full_transaction_id(xl_xid)?
712 : } else {
713 0 : xl_xid as u64
714 : };
715 0 : return modification.put_twophase_file(xid, data, ctx).await;
716 : }
717 8 : XactRecord::Commit(common) => (common, true, false),
718 0 : XactRecord::Abort(common) => (common, false, false),
719 0 : XactRecord::CommitPrepared(common) => (common, true, true),
720 0 : XactRecord::AbortPrepared(common) => (common, false, true),
721 : };
722 :
723 : let XactCommon {
724 8 : parsed,
725 8 : origin_id,
726 8 : xl_xid,
727 8 : lsn,
728 8 : } = xact_common;
729 8 :
730 8 : // Record update of CLOG pages
731 8 : let mut pageno = parsed.xid / pg_constants::CLOG_XACTS_PER_PAGE;
732 8 : let mut segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
733 8 : let mut rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
734 8 : let mut page_xids: Vec<TransactionId> = vec![parsed.xid];
735 8 :
736 8 : self.warn_on_ingest_lag(modification.tline.conf, parsed.xact_time);
737 :
738 8 : for subxact in &parsed.subxacts {
739 0 : let subxact_pageno = subxact / pg_constants::CLOG_XACTS_PER_PAGE;
740 0 : if subxact_pageno != pageno {
741 : // This subxact goes to different page. Write the record
742 : // for all the XIDs on the previous page, and continue
743 : // accumulating XIDs on this new page.
744 0 : modification.put_slru_wal_record(
745 0 : SlruKind::Clog,
746 0 : segno,
747 0 : rpageno,
748 0 : if is_commit {
749 0 : NeonWalRecord::ClogSetCommitted {
750 0 : xids: page_xids,
751 0 : timestamp: parsed.xact_time,
752 0 : }
753 : } else {
754 0 : NeonWalRecord::ClogSetAborted { xids: page_xids }
755 : },
756 0 : )?;
757 0 : page_xids = Vec::new();
758 0 : }
759 0 : pageno = subxact_pageno;
760 0 : segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
761 0 : rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
762 0 : page_xids.push(*subxact);
763 : }
764 8 : modification.put_slru_wal_record(
765 8 : SlruKind::Clog,
766 8 : segno,
767 8 : rpageno,
768 8 : if is_commit {
769 8 : NeonWalRecord::ClogSetCommitted {
770 8 : xids: page_xids,
771 8 : timestamp: parsed.xact_time,
772 8 : }
773 : } else {
774 0 : NeonWalRecord::ClogSetAborted { xids: page_xids }
775 : },
776 0 : )?;
777 :
778 : // Group relations to drop by dbNode. This map will contain all relations that _might_
779 : // exist, we will reduce it to which ones really exist later. This map can be huge if
780 : // the transaction touches a huge number of relations (there is no bound on this in
781 : // postgres).
782 8 : let mut drop_relations: HashMap<(u32, u32), Vec<RelTag>> = HashMap::new();
783 :
784 8 : for xnode in &parsed.xnodes {
785 0 : for forknum in MAIN_FORKNUM..=INIT_FORKNUM {
786 0 : let rel = RelTag {
787 0 : forknum,
788 0 : spcnode: xnode.spcnode,
789 0 : dbnode: xnode.dbnode,
790 0 : relnode: xnode.relnode,
791 0 : };
792 0 : drop_relations
793 0 : .entry((xnode.spcnode, xnode.dbnode))
794 0 : .or_default()
795 0 : .push(rel);
796 0 : }
797 : }
798 :
799 : // Execute relation drops in a batch: the number may be huge, so deleting individually is prohibitively expensive
800 8 : modification.put_rel_drops(drop_relations, ctx).await?;
801 :
802 8 : if origin_id != 0 {
803 0 : modification
804 0 : .set_replorigin(origin_id, parsed.origin_lsn)
805 0 : .await?;
806 8 : }
807 :
808 8 : if is_prepared {
809 : // Remove twophase file. see RemoveTwoPhaseFile() in postgres code
810 0 : trace!(
811 0 : "Drop twophaseFile for xid {} parsed_xact.xid {} here at {}",
812 : xl_xid,
813 : parsed.xid,
814 : lsn,
815 : );
816 :
817 0 : let xid: u64 = if modification.tline.pg_version >= 17 {
818 0 : self.adjust_to_full_transaction_id(parsed.xid)?
819 : } else {
820 0 : parsed.xid as u64
821 : };
822 0 : modification.drop_twophase_file(xid, ctx).await?;
823 8 : }
824 :
825 8 : Ok(())
826 8 : }
827 :
828 0 : async fn ingest_clog_truncate(
829 0 : &mut self,
830 0 : truncate: ClogTruncate,
831 0 : modification: &mut DatadirModification<'_>,
832 0 : ctx: &RequestContext,
833 0 : ) -> anyhow::Result<()> {
834 0 : let ClogTruncate {
835 0 : pageno,
836 0 : oldest_xid,
837 0 : oldest_xid_db,
838 0 : } = truncate;
839 0 :
840 0 : info!(
841 0 : "RM_CLOG_ID truncate pageno {} oldestXid {} oldestXidDB {}",
842 : pageno, oldest_xid, oldest_xid_db
843 : );
844 :
845 : // In Postgres, oldestXid and oldestXidDB are updated in memory when the CLOG is
846 : // truncated, but a checkpoint record with the updated values isn't written until
847 : // later. In Neon, a server can start at any LSN, not just on a checkpoint record,
848 : // so we keep the oldestXid and oldestXidDB up-to-date.
849 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
850 0 : cp.oldestXid = oldest_xid;
851 0 : cp.oldestXidDB = oldest_xid_db;
852 0 : });
853 0 : self.checkpoint_modified = true;
854 :
855 : // TODO Treat AdvanceOldestClogXid() or write a comment why we don't need it
856 :
857 0 : let latest_page_number =
858 0 : enum_pgversion_dispatch!(self.checkpoint, CheckPoint, cp, { cp.nextXid.value }) as u32
859 : / pg_constants::CLOG_XACTS_PER_PAGE;
860 :
861 : // Now delete all segments containing pages between xlrec.pageno
862 : // and latest_page_number.
863 :
864 : // First, make an important safety check:
865 : // the current endpoint page must not be eligible for removal.
866 : // See SimpleLruTruncate() in slru.c
867 0 : if dispatch_pgversion!(modification.tline.pg_version, {
868 0 : pgv::nonrelfile_utils::clogpage_precedes(latest_page_number, pageno)
869 : }) {
870 0 : info!("could not truncate directory pg_xact apparent wraparound");
871 0 : return Ok(());
872 0 : }
873 0 :
874 0 : // Iterate via SLRU CLOG segments and drop segments that we're ready to truncate
875 0 : //
876 0 : // We cannot pass 'lsn' to the Timeline.list_nonrels(), or it
877 0 : // will block waiting for the last valid LSN to advance up to
878 0 : // it. So we use the previous record's LSN in the get calls
879 0 : // instead.
880 0 : if modification.tline.get_shard_identity().is_shard_zero() {
881 0 : for segno in modification
882 0 : .tline
883 0 : .list_slru_segments(SlruKind::Clog, Version::Modified(modification), ctx)
884 0 : .await?
885 : {
886 0 : let segpage = segno * pg_constants::SLRU_PAGES_PER_SEGMENT;
887 :
888 0 : let may_delete = dispatch_pgversion!(modification.tline.pg_version, {
889 0 : pgv::nonrelfile_utils::slru_may_delete_clogsegment(segpage, pageno)
890 : });
891 :
892 0 : if may_delete {
893 0 : modification
894 0 : .drop_slru_segment(SlruKind::Clog, segno, ctx)
895 0 : .await?;
896 0 : trace!("Drop CLOG segment {:>04X}", segno);
897 0 : }
898 : }
899 0 : }
900 :
901 0 : Ok(())
902 0 : }
903 :
904 0 : async fn ingest_clog_zero_page(
905 0 : &mut self,
906 0 : zero_page: ClogZeroPage,
907 0 : modification: &mut DatadirModification<'_>,
908 0 : ctx: &RequestContext,
909 0 : ) -> anyhow::Result<()> {
910 0 : let ClogZeroPage { segno, rpageno } = zero_page;
911 0 :
912 0 : self.put_slru_page_image(
913 0 : modification,
914 0 : SlruKind::Clog,
915 0 : segno,
916 0 : rpageno,
917 0 : ZERO_PAGE.clone(),
918 0 : ctx,
919 0 : )
920 0 : .await
921 0 : }
922 :
923 0 : fn ingest_multixact_create(
924 0 : &mut self,
925 0 : modification: &mut DatadirModification,
926 0 : xlrec: &XlMultiXactCreate,
927 0 : ) -> Result<()> {
928 0 : // Create WAL record for updating the multixact-offsets page
929 0 : let pageno = xlrec.mid / pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32;
930 0 : let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
931 0 : let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
932 0 :
933 0 : modification.put_slru_wal_record(
934 0 : SlruKind::MultiXactOffsets,
935 0 : segno,
936 0 : rpageno,
937 0 : NeonWalRecord::MultixactOffsetCreate {
938 0 : mid: xlrec.mid,
939 0 : moff: xlrec.moff,
940 0 : },
941 0 : )?;
942 :
943 : // Create WAL records for the update of each affected multixact-members page
944 0 : let mut members = xlrec.members.iter();
945 0 : let mut offset = xlrec.moff;
946 : loop {
947 0 : let pageno = offset / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
948 0 :
949 0 : // How many members fit on this page?
950 0 : let page_remain = pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32
951 0 : - offset % pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
952 0 :
953 0 : let mut this_page_members: Vec<MultiXactMember> = Vec::new();
954 0 : for _ in 0..page_remain {
955 0 : if let Some(m) = members.next() {
956 0 : this_page_members.push(m.clone());
957 0 : } else {
958 0 : break;
959 : }
960 : }
961 0 : if this_page_members.is_empty() {
962 : // all done
963 0 : break;
964 0 : }
965 0 : let n_this_page = this_page_members.len();
966 0 :
967 0 : modification.put_slru_wal_record(
968 0 : SlruKind::MultiXactMembers,
969 0 : pageno / pg_constants::SLRU_PAGES_PER_SEGMENT,
970 0 : pageno % pg_constants::SLRU_PAGES_PER_SEGMENT,
971 0 : NeonWalRecord::MultixactMembersCreate {
972 0 : moff: offset,
973 0 : members: this_page_members,
974 0 : },
975 0 : )?;
976 :
977 : // Note: The multixact members can wrap around, even within one WAL record.
978 0 : offset = offset.wrapping_add(n_this_page as u32);
979 : }
980 0 : let next_offset = offset;
981 0 : assert!(xlrec.moff.wrapping_add(xlrec.nmembers) == next_offset);
982 :
983 : // Update next-multi-xid and next-offset
984 : //
985 : // NB: In PostgreSQL, the next-multi-xid stored in the control file is allowed to
986 : // go to 0, and it's fixed up by skipping to FirstMultiXactId in functions that
987 : // read it, like GetNewMultiXactId(). This is different from how nextXid is
988 : // incremented! nextXid skips over < FirstNormalTransactionId when the the value
989 : // is stored, so it's never 0 in a checkpoint.
990 : //
991 : // I don't know why it's done that way, it seems less error-prone to skip over 0
992 : // when the value is stored rather than when it's read. But let's do it the same
993 : // way here.
994 0 : let next_multi_xid = xlrec.mid.wrapping_add(1);
995 0 :
996 0 : if self
997 0 : .checkpoint
998 0 : .update_next_multixid(next_multi_xid, next_offset)
999 0 : {
1000 0 : self.checkpoint_modified = true;
1001 0 : }
1002 :
1003 : // Also update the next-xid with the highest member. According to the comments in
1004 : // multixact_redo(), this shouldn't be necessary, but let's do the same here.
1005 0 : let max_mbr_xid = xlrec.members.iter().fold(None, |acc, mbr| {
1006 0 : if let Some(max_xid) = acc {
1007 0 : if mbr.xid.wrapping_sub(max_xid) as i32 > 0 {
1008 0 : Some(mbr.xid)
1009 : } else {
1010 0 : acc
1011 : }
1012 : } else {
1013 0 : Some(mbr.xid)
1014 : }
1015 0 : });
1016 :
1017 0 : if let Some(max_xid) = max_mbr_xid {
1018 0 : if self.checkpoint.update_next_xid(max_xid) {
1019 0 : self.checkpoint_modified = true;
1020 0 : }
1021 0 : }
1022 0 : Ok(())
1023 0 : }
1024 :
1025 0 : async fn ingest_multixact_truncate(
1026 0 : &mut self,
1027 0 : modification: &mut DatadirModification<'_>,
1028 0 : xlrec: &XlMultiXactTruncate,
1029 0 : ctx: &RequestContext,
1030 0 : ) -> Result<()> {
1031 0 : let (maxsegment, startsegment, endsegment) =
1032 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1033 0 : cp.oldestMulti = xlrec.end_trunc_off;
1034 0 : cp.oldestMultiDB = xlrec.oldest_multi_db;
1035 0 : let maxsegment: i32 = pgv::nonrelfile_utils::mx_offset_to_member_segment(
1036 0 : pg_constants::MAX_MULTIXACT_OFFSET,
1037 0 : );
1038 0 : let startsegment: i32 =
1039 0 : pgv::nonrelfile_utils::mx_offset_to_member_segment(xlrec.start_trunc_memb);
1040 0 : let endsegment: i32 =
1041 0 : pgv::nonrelfile_utils::mx_offset_to_member_segment(xlrec.end_trunc_memb);
1042 0 : (maxsegment, startsegment, endsegment)
1043 : });
1044 :
1045 0 : self.checkpoint_modified = true;
1046 0 :
1047 0 : // PerformMembersTruncation
1048 0 : let mut segment: i32 = startsegment;
1049 0 :
1050 0 : // Delete all the segments except the last one. The last segment can still
1051 0 : // contain, possibly partially, valid data.
1052 0 : if modification.tline.get_shard_identity().is_shard_zero() {
1053 0 : while segment != endsegment {
1054 0 : modification
1055 0 : .drop_slru_segment(SlruKind::MultiXactMembers, segment as u32, ctx)
1056 0 : .await?;
1057 :
1058 : /* move to next segment, handling wraparound correctly */
1059 0 : if segment == maxsegment {
1060 0 : segment = 0;
1061 0 : } else {
1062 0 : segment += 1;
1063 0 : }
1064 : }
1065 0 : }
1066 :
1067 : // Truncate offsets
1068 : // FIXME: this did not handle wraparound correctly
1069 :
1070 0 : Ok(())
1071 0 : }
1072 :
1073 0 : async fn ingest_multixact_zero_page(
1074 0 : &mut self,
1075 0 : zero_page: MultiXactZeroPage,
1076 0 : modification: &mut DatadirModification<'_>,
1077 0 : ctx: &RequestContext,
1078 0 : ) -> Result<()> {
1079 0 : let MultiXactZeroPage {
1080 0 : slru_kind,
1081 0 : segno,
1082 0 : rpageno,
1083 0 : } = zero_page;
1084 0 : self.put_slru_page_image(
1085 0 : modification,
1086 0 : slru_kind,
1087 0 : segno,
1088 0 : rpageno,
1089 0 : ZERO_PAGE.clone(),
1090 0 : ctx,
1091 0 : )
1092 0 : .await
1093 0 : }
1094 :
1095 0 : async fn ingest_relmap_update(
1096 0 : &mut self,
1097 0 : update: RelmapUpdate,
1098 0 : modification: &mut DatadirModification<'_>,
1099 0 : ctx: &RequestContext,
1100 0 : ) -> Result<()> {
1101 0 : let RelmapUpdate { update, buf } = update;
1102 0 :
1103 0 : modification
1104 0 : .put_relmap_file(update.tsid, update.dbid, buf, ctx)
1105 0 : .await
1106 0 : }
1107 :
1108 30 : async fn ingest_raw_xlog_record(
1109 30 : &mut self,
1110 30 : raw_record: RawXlogRecord,
1111 30 : modification: &mut DatadirModification<'_>,
1112 30 : ctx: &RequestContext,
1113 30 : ) -> Result<()> {
1114 30 : let RawXlogRecord { info, lsn, mut buf } = raw_record;
1115 30 : let pg_version = modification.tline.pg_version;
1116 30 :
1117 30 : if info == pg_constants::XLOG_PARAMETER_CHANGE {
1118 2 : if let CheckPoint::V17(cp) = &mut self.checkpoint {
1119 0 : let rec = v17::XlParameterChange::decode(&mut buf);
1120 0 : cp.wal_level = rec.wal_level;
1121 0 : self.checkpoint_modified = true;
1122 2 : }
1123 28 : } else if info == pg_constants::XLOG_END_OF_RECOVERY {
1124 0 : if let CheckPoint::V17(cp) = &mut self.checkpoint {
1125 0 : let rec = v17::XlEndOfRecovery::decode(&mut buf);
1126 0 : cp.wal_level = rec.wal_level;
1127 0 : self.checkpoint_modified = true;
1128 0 : }
1129 28 : }
1130 :
1131 30 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1132 0 : if info == pg_constants::XLOG_NEXTOID {
1133 0 : let next_oid = buf.get_u32_le();
1134 0 : if cp.nextOid != next_oid {
1135 0 : cp.nextOid = next_oid;
1136 0 : self.checkpoint_modified = true;
1137 0 : }
1138 0 : } else if info == pg_constants::XLOG_CHECKPOINT_ONLINE
1139 0 : || info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
1140 : {
1141 0 : let mut checkpoint_bytes = [0u8; pgv::xlog_utils::SIZEOF_CHECKPOINT];
1142 0 : buf.copy_to_slice(&mut checkpoint_bytes);
1143 0 : let xlog_checkpoint = pgv::CheckPoint::decode(&checkpoint_bytes)?;
1144 0 : trace!(
1145 0 : "xlog_checkpoint.oldestXid={}, checkpoint.oldestXid={}",
1146 : xlog_checkpoint.oldestXid,
1147 : cp.oldestXid
1148 : );
1149 0 : if (cp.oldestXid.wrapping_sub(xlog_checkpoint.oldestXid) as i32) < 0 {
1150 0 : cp.oldestXid = xlog_checkpoint.oldestXid;
1151 0 : }
1152 0 : trace!(
1153 0 : "xlog_checkpoint.oldestActiveXid={}, checkpoint.oldestActiveXid={}",
1154 : xlog_checkpoint.oldestActiveXid,
1155 : cp.oldestActiveXid
1156 : );
1157 :
1158 : // A shutdown checkpoint has `oldestActiveXid == InvalidTransactionid`,
1159 : // because at shutdown, all in-progress transactions will implicitly
1160 : // end. Postgres startup code knows that, and allows hot standby to start
1161 : // immediately from a shutdown checkpoint.
1162 : //
1163 : // In Neon, Postgres hot standby startup always behaves as if starting from
1164 : // an online checkpoint. It needs a valid `oldestActiveXid` value, so
1165 : // instead of overwriting self.checkpoint.oldestActiveXid with
1166 : // InvalidTransactionid from the checkpoint WAL record, update it to a
1167 : // proper value, knowing that there are no in-progress transactions at this
1168 : // point, except for prepared transactions.
1169 : //
1170 : // See also the neon code changes in the InitWalRecovery() function.
1171 0 : if xlog_checkpoint.oldestActiveXid == pg_constants::INVALID_TRANSACTION_ID
1172 0 : && info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
1173 : {
1174 0 : let oldest_active_xid = if pg_version >= 17 {
1175 0 : let mut oldest_active_full_xid = cp.nextXid.value;
1176 0 : for xid in modification.tline.list_twophase_files(lsn, ctx).await? {
1177 0 : if xid < oldest_active_full_xid {
1178 0 : oldest_active_full_xid = xid;
1179 0 : }
1180 : }
1181 0 : oldest_active_full_xid as u32
1182 : } else {
1183 0 : let mut oldest_active_xid = cp.nextXid.value as u32;
1184 0 : for xid in modification.tline.list_twophase_files(lsn, ctx).await? {
1185 0 : let narrow_xid = xid as u32;
1186 0 : if (narrow_xid.wrapping_sub(oldest_active_xid) as i32) < 0 {
1187 0 : oldest_active_xid = narrow_xid;
1188 0 : }
1189 : }
1190 0 : oldest_active_xid
1191 : };
1192 0 : cp.oldestActiveXid = oldest_active_xid;
1193 0 : } else {
1194 0 : cp.oldestActiveXid = xlog_checkpoint.oldestActiveXid;
1195 0 : }
1196 :
1197 : // Write a new checkpoint key-value pair on every checkpoint record, even
1198 : // if nothing really changed. Not strictly required, but it seems nice to
1199 : // have some trace of the checkpoint records in the layer files at the same
1200 : // LSNs.
1201 0 : self.checkpoint_modified = true;
1202 0 : }
1203 : });
1204 :
1205 30 : Ok(())
1206 30 : }
1207 :
1208 0 : async fn ingest_logical_message_put(
1209 0 : &mut self,
1210 0 : put: PutLogicalMessage,
1211 0 : modification: &mut DatadirModification<'_>,
1212 0 : ctx: &RequestContext,
1213 0 : ) -> Result<()> {
1214 0 : let PutLogicalMessage { path, buf } = put;
1215 0 : modification.put_file(path.as_str(), &buf, ctx).await
1216 0 : }
1217 :
1218 0 : fn ingest_standby_record(&mut self, record: StandbyRecord) -> Result<()> {
1219 0 : match record {
1220 0 : StandbyRecord::RunningXacts(running_xacts) => {
1221 0 : enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
1222 0 : cp.oldestActiveXid = running_xacts.oldest_running_xid;
1223 0 : });
1224 :
1225 0 : self.checkpoint_modified = true;
1226 0 : }
1227 0 : }
1228 0 :
1229 0 : Ok(())
1230 0 : }
1231 :
1232 0 : async fn ingest_replorigin_record(
1233 0 : &mut self,
1234 0 : record: ReploriginRecord,
1235 0 : modification: &mut DatadirModification<'_>,
1236 0 : ) -> Result<()> {
1237 0 : match record {
1238 0 : ReploriginRecord::Set(set) => {
1239 0 : modification
1240 0 : .set_replorigin(set.node_id, set.remote_lsn)
1241 0 : .await?;
1242 : }
1243 0 : ReploriginRecord::Drop(drop) => {
1244 0 : modification.drop_replorigin(drop.node_id).await?;
1245 : }
1246 : }
1247 :
1248 0 : Ok(())
1249 0 : }
1250 :
1251 18 : async fn put_rel_creation(
1252 18 : &mut self,
1253 18 : modification: &mut DatadirModification<'_>,
1254 18 : rel: RelTag,
1255 18 : ctx: &RequestContext,
1256 18 : ) -> Result<()> {
1257 18 : modification.put_rel_creation(rel, 0, ctx).await?;
1258 18 : Ok(())
1259 18 : }
1260 :
1261 : #[cfg(test)]
1262 272402 : async fn put_rel_page_image(
1263 272402 : &mut self,
1264 272402 : modification: &mut DatadirModification<'_>,
1265 272402 : rel: RelTag,
1266 272402 : blknum: BlockNumber,
1267 272402 : img: Bytes,
1268 272402 : ctx: &RequestContext,
1269 272402 : ) -> Result<(), PageReconstructError> {
1270 272402 : self.handle_rel_extend(modification, rel, blknum, ctx)
1271 272402 : .await?;
1272 272402 : modification.put_rel_page_image(rel, blknum, img)?;
1273 272402 : Ok(())
1274 272402 : }
1275 :
1276 12 : async fn put_rel_wal_record(
1277 12 : &mut self,
1278 12 : modification: &mut DatadirModification<'_>,
1279 12 : rel: RelTag,
1280 12 : blknum: BlockNumber,
1281 12 : rec: NeonWalRecord,
1282 12 : ctx: &RequestContext,
1283 12 : ) -> Result<()> {
1284 12 : self.handle_rel_extend(modification, rel, blknum, ctx)
1285 12 : .await?;
1286 12 : modification.put_rel_wal_record(rel, blknum, rec)?;
1287 12 : Ok(())
1288 12 : }
1289 :
1290 6012 : async fn put_rel_truncation(
1291 6012 : &mut self,
1292 6012 : modification: &mut DatadirModification<'_>,
1293 6012 : rel: RelTag,
1294 6012 : nblocks: BlockNumber,
1295 6012 : ctx: &RequestContext,
1296 6012 : ) -> anyhow::Result<()> {
1297 6012 : modification.put_rel_truncation(rel, nblocks, ctx).await?;
1298 6012 : Ok(())
1299 6012 : }
1300 :
1301 272414 : async fn handle_rel_extend(
1302 272414 : &mut self,
1303 272414 : modification: &mut DatadirModification<'_>,
1304 272414 : rel: RelTag,
1305 272414 : blknum: BlockNumber,
1306 272414 : ctx: &RequestContext,
1307 272414 : ) -> Result<(), PageReconstructError> {
1308 272414 : let new_nblocks = blknum + 1;
1309 : // Check if the relation exists. We implicitly create relations on first
1310 : // record.
1311 272414 : let old_nblocks = modification.create_relation_if_required(rel, ctx).await?;
1312 :
1313 272414 : if new_nblocks > old_nblocks {
1314 : //info!("extending {} {} to {}", rel, old_nblocks, new_nblocks);
1315 272398 : modification.put_rel_extend(rel, new_nblocks, ctx).await?;
1316 :
1317 272398 : let mut key = rel_block_to_key(rel, blknum);
1318 272398 :
1319 272398 : // fill the gap with zeros
1320 272398 : let mut gap_blocks_filled: u64 = 0;
1321 272398 : for gap_blknum in old_nblocks..blknum {
1322 2998 : key.field6 = gap_blknum;
1323 2998 :
1324 2998 : if self.shard.get_shard_number(&key) != self.shard.number {
1325 0 : continue;
1326 2998 : }
1327 2998 :
1328 2998 : modification.put_rel_page_image_zero(rel, gap_blknum)?;
1329 2998 : gap_blocks_filled += 1;
1330 : }
1331 :
1332 272398 : WAL_INGEST
1333 272398 : .gap_blocks_zeroed_on_rel_extend
1334 272398 : .inc_by(gap_blocks_filled);
1335 272398 :
1336 272398 : // Log something when relation extends cause use to fill gaps
1337 272398 : // with zero pages. Logging is rate limited per pg version to
1338 272398 : // avoid skewing.
1339 272398 : if gap_blocks_filled > 0 {
1340 : use once_cell::sync::Lazy;
1341 : use std::sync::Mutex;
1342 : use utils::rate_limit::RateLimit;
1343 :
1344 : struct RateLimitPerPgVersion {
1345 : rate_limiters: [Lazy<Mutex<RateLimit>>; 4],
1346 : }
1347 :
1348 : impl RateLimitPerPgVersion {
1349 0 : const fn new() -> Self {
1350 0 : Self {
1351 0 : rate_limiters: [const {
1352 2 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(30))))
1353 0 : }; 4],
1354 0 : }
1355 0 : }
1356 :
1357 4 : const fn rate_limiter(
1358 4 : &self,
1359 4 : pg_version: u32,
1360 4 : ) -> Option<&Lazy<Mutex<RateLimit>>> {
1361 : const MIN_PG_VERSION: u32 = 14;
1362 : const MAX_PG_VERSION: u32 = 17;
1363 :
1364 4 : if pg_version < MIN_PG_VERSION || pg_version > MAX_PG_VERSION {
1365 0 : return None;
1366 4 : }
1367 4 :
1368 4 : Some(&self.rate_limiters[(pg_version - MIN_PG_VERSION) as usize])
1369 4 : }
1370 : }
1371 :
1372 : static LOGGED: RateLimitPerPgVersion = RateLimitPerPgVersion::new();
1373 4 : if let Some(rate_limiter) = LOGGED.rate_limiter(modification.tline.pg_version) {
1374 4 : if let Ok(mut locked) = rate_limiter.try_lock() {
1375 4 : locked.call(|| {
1376 2 : info!(
1377 0 : lsn=%modification.get_lsn(),
1378 0 : pg_version=%modification.tline.pg_version,
1379 0 : rel=%rel,
1380 0 : "Filled {} gap blocks on rel extend to {} from {}",
1381 : gap_blocks_filled,
1382 : new_nblocks,
1383 : old_nblocks);
1384 4 : });
1385 4 : }
1386 0 : }
1387 272394 : }
1388 16 : }
1389 272414 : Ok(())
1390 272414 : }
1391 :
1392 0 : async fn put_slru_page_image(
1393 0 : &mut self,
1394 0 : modification: &mut DatadirModification<'_>,
1395 0 : kind: SlruKind,
1396 0 : segno: u32,
1397 0 : blknum: BlockNumber,
1398 0 : img: Bytes,
1399 0 : ctx: &RequestContext,
1400 0 : ) -> Result<()> {
1401 0 : if !self.shard.is_shard_zero() {
1402 0 : return Ok(());
1403 0 : }
1404 0 :
1405 0 : self.handle_slru_extend(modification, kind, segno, blknum, ctx)
1406 0 : .await?;
1407 0 : modification.put_slru_page_image(kind, segno, blknum, img)?;
1408 0 : Ok(())
1409 0 : }
1410 :
1411 0 : async fn handle_slru_extend(
1412 0 : &mut self,
1413 0 : modification: &mut DatadirModification<'_>,
1414 0 : kind: SlruKind,
1415 0 : segno: u32,
1416 0 : blknum: BlockNumber,
1417 0 : ctx: &RequestContext,
1418 0 : ) -> anyhow::Result<()> {
1419 0 : // we don't use a cache for this like we do for relations. SLRUS are explcitly
1420 0 : // extended with ZEROPAGE records, not with commit records, so it happens
1421 0 : // a lot less frequently.
1422 0 :
1423 0 : let new_nblocks = blknum + 1;
1424 : // Check if the relation exists. We implicitly create relations on first
1425 : // record.
1426 : // TODO: would be nice if to be more explicit about it
1427 0 : let old_nblocks = if !modification
1428 0 : .tline
1429 0 : .get_slru_segment_exists(kind, segno, Version::Modified(modification), ctx)
1430 0 : .await?
1431 : {
1432 : // create it with 0 size initially, the logic below will extend it
1433 0 : modification
1434 0 : .put_slru_segment_creation(kind, segno, 0, ctx)
1435 0 : .await?;
1436 0 : 0
1437 : } else {
1438 0 : modification
1439 0 : .tline
1440 0 : .get_slru_segment_size(kind, segno, Version::Modified(modification), ctx)
1441 0 : .await?
1442 : };
1443 :
1444 0 : if new_nblocks > old_nblocks {
1445 0 : trace!(
1446 0 : "extending SLRU {:?} seg {} from {} to {} blocks",
1447 : kind,
1448 : segno,
1449 : old_nblocks,
1450 : new_nblocks
1451 : );
1452 0 : modification.put_slru_extend(kind, segno, new_nblocks)?;
1453 :
1454 : // fill the gap with zeros
1455 0 : for gap_blknum in old_nblocks..blknum {
1456 0 : modification.put_slru_page_image_zero(kind, segno, gap_blknum)?;
1457 : }
1458 0 : }
1459 0 : Ok(())
1460 0 : }
1461 : }
1462 :
1463 : /// Returns the size of the relation as of this modification, or None if the relation doesn't exist.
1464 : ///
1465 : /// This is only accurate on shard 0. On other shards, it will return the size up to the highest
1466 : /// page number stored in the shard, or None if the shard does not have any pages for it.
1467 12 : async fn get_relsize(
1468 12 : modification: &DatadirModification<'_>,
1469 12 : rel: RelTag,
1470 12 : ctx: &RequestContext,
1471 12 : ) -> Result<Option<BlockNumber>, PageReconstructError> {
1472 12 : if !modification
1473 12 : .tline
1474 12 : .get_rel_exists(rel, Version::Modified(modification), ctx)
1475 12 : .await?
1476 : {
1477 0 : return Ok(None);
1478 12 : }
1479 12 : modification
1480 12 : .tline
1481 12 : .get_rel_size(rel, Version::Modified(modification), ctx)
1482 12 : .await
1483 12 : .map(Some)
1484 12 : }
1485 :
1486 : #[allow(clippy::bool_assert_comparison)]
1487 : #[cfg(test)]
1488 : mod tests {
1489 : use super::*;
1490 : use crate::tenant::harness::*;
1491 : use crate::tenant::remote_timeline_client::{remote_initdb_archive_path, INITDB_PATH};
1492 : use postgres_ffi::RELSEG_SIZE;
1493 :
1494 : use crate::DEFAULT_PG_VERSION;
1495 :
1496 : /// Arbitrary relation tag, for testing.
1497 : const TESTREL_A: RelTag = RelTag {
1498 : spcnode: 0,
1499 : dbnode: 111,
1500 : relnode: 1000,
1501 : forknum: 0,
1502 : };
1503 :
1504 12 : fn assert_current_logical_size(_timeline: &Timeline, _lsn: Lsn) {
1505 12 : // TODO
1506 12 : }
1507 :
1508 : #[tokio::test]
1509 2 : async fn test_zeroed_checkpoint_decodes_correctly() -> Result<()> {
1510 8 : for i in 14..=16 {
1511 6 : dispatch_pgversion!(i, {
1512 2 : pgv::CheckPoint::decode(&pgv::ZERO_CHECKPOINT)?;
1513 2 : });
1514 2 : }
1515 2 :
1516 2 : Ok(())
1517 2 : }
1518 :
1519 8 : async fn init_walingest_test(tline: &Timeline, ctx: &RequestContext) -> Result<WalIngest> {
1520 8 : let mut m = tline.begin_modification(Lsn(0x10));
1521 8 : m.put_checkpoint(dispatch_pgversion!(
1522 8 : tline.pg_version,
1523 0 : pgv::ZERO_CHECKPOINT.clone()
1524 0 : ))?;
1525 8 : m.put_relmap_file(0, 111, Bytes::from(""), ctx).await?; // dummy relmapper file
1526 8 : m.commit(ctx).await?;
1527 8 : let walingest = WalIngest::new(tline, Lsn(0x10), ctx).await?;
1528 :
1529 8 : Ok(walingest)
1530 8 : }
1531 :
1532 : #[tokio::test]
1533 2 : async fn test_relsize() -> Result<()> {
1534 2 : let (tenant, ctx) = TenantHarness::create("test_relsize").await?.load().await;
1535 2 : let tline = tenant
1536 2 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
1537 2 : .await?;
1538 2 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
1539 2 :
1540 2 : let mut m = tline.begin_modification(Lsn(0x20));
1541 2 : walingest.put_rel_creation(&mut m, TESTREL_A, &ctx).await?;
1542 2 : walingest
1543 2 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 2"), &ctx)
1544 2 : .await?;
1545 2 : m.commit(&ctx).await?;
1546 2 : let mut m = tline.begin_modification(Lsn(0x30));
1547 2 : walingest
1548 2 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 3"), &ctx)
1549 2 : .await?;
1550 2 : m.commit(&ctx).await?;
1551 2 : let mut m = tline.begin_modification(Lsn(0x40));
1552 2 : walingest
1553 2 : .put_rel_page_image(&mut m, TESTREL_A, 1, test_img("foo blk 1 at 4"), &ctx)
1554 2 : .await?;
1555 2 : m.commit(&ctx).await?;
1556 2 : let mut m = tline.begin_modification(Lsn(0x50));
1557 2 : walingest
1558 2 : .put_rel_page_image(&mut m, TESTREL_A, 2, test_img("foo blk 2 at 5"), &ctx)
1559 2 : .await?;
1560 2 : m.commit(&ctx).await?;
1561 2 :
1562 2 : assert_current_logical_size(&tline, Lsn(0x50));
1563 2 :
1564 2 : let test_span = tracing::info_span!(parent: None, "test",
1565 2 : tenant_id=%tline.tenant_shard_id.tenant_id,
1566 0 : shard_id=%tline.tenant_shard_id.shard_slug(),
1567 0 : timeline_id=%tline.timeline_id);
1568 2 :
1569 2 : // The relation was created at LSN 2, not visible at LSN 1 yet.
1570 2 : assert_eq!(
1571 2 : tline
1572 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x10)), &ctx)
1573 2 : .await?,
1574 2 : false
1575 2 : );
1576 2 : assert!(tline
1577 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x10)), &ctx)
1578 2 : .await
1579 2 : .is_err());
1580 2 : assert_eq!(
1581 2 : tline
1582 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1583 2 : .await?,
1584 2 : true
1585 2 : );
1586 2 : assert_eq!(
1587 2 : tline
1588 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1589 2 : .await?,
1590 2 : 1
1591 2 : );
1592 2 : assert_eq!(
1593 2 : tline
1594 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x50)), &ctx)
1595 2 : .await?,
1596 2 : 3
1597 2 : );
1598 2 :
1599 2 : // Check page contents at each LSN
1600 2 : assert_eq!(
1601 2 : tline
1602 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x20)), &ctx)
1603 2 : .instrument(test_span.clone())
1604 2 : .await?,
1605 2 : test_img("foo blk 0 at 2")
1606 2 : );
1607 2 :
1608 2 : assert_eq!(
1609 2 : tline
1610 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x30)), &ctx)
1611 2 : .instrument(test_span.clone())
1612 2 : .await?,
1613 2 : test_img("foo blk 0 at 3")
1614 2 : );
1615 2 :
1616 2 : assert_eq!(
1617 2 : tline
1618 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x40)), &ctx)
1619 2 : .instrument(test_span.clone())
1620 2 : .await?,
1621 2 : test_img("foo blk 0 at 3")
1622 2 : );
1623 2 : assert_eq!(
1624 2 : tline
1625 2 : .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x40)), &ctx)
1626 2 : .instrument(test_span.clone())
1627 2 : .await?,
1628 2 : test_img("foo blk 1 at 4")
1629 2 : );
1630 2 :
1631 2 : assert_eq!(
1632 2 : tline
1633 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x50)), &ctx)
1634 2 : .instrument(test_span.clone())
1635 2 : .await?,
1636 2 : test_img("foo blk 0 at 3")
1637 2 : );
1638 2 : assert_eq!(
1639 2 : tline
1640 2 : .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x50)), &ctx)
1641 2 : .instrument(test_span.clone())
1642 2 : .await?,
1643 2 : test_img("foo blk 1 at 4")
1644 2 : );
1645 2 : assert_eq!(
1646 2 : tline
1647 2 : .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
1648 2 : .instrument(test_span.clone())
1649 2 : .await?,
1650 2 : test_img("foo blk 2 at 5")
1651 2 : );
1652 2 :
1653 2 : // Truncate last block
1654 2 : let mut m = tline.begin_modification(Lsn(0x60));
1655 2 : walingest
1656 2 : .put_rel_truncation(&mut m, TESTREL_A, 2, &ctx)
1657 2 : .await?;
1658 2 : m.commit(&ctx).await?;
1659 2 : assert_current_logical_size(&tline, Lsn(0x60));
1660 2 :
1661 2 : // Check reported size and contents after truncation
1662 2 : assert_eq!(
1663 2 : tline
1664 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x60)), &ctx)
1665 2 : .await?,
1666 2 : 2
1667 2 : );
1668 2 : assert_eq!(
1669 2 : tline
1670 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x60)), &ctx)
1671 2 : .instrument(test_span.clone())
1672 2 : .await?,
1673 2 : test_img("foo blk 0 at 3")
1674 2 : );
1675 2 : assert_eq!(
1676 2 : tline
1677 2 : .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x60)), &ctx)
1678 2 : .instrument(test_span.clone())
1679 2 : .await?,
1680 2 : test_img("foo blk 1 at 4")
1681 2 : );
1682 2 :
1683 2 : // should still see the truncated block with older LSN
1684 2 : assert_eq!(
1685 2 : tline
1686 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x50)), &ctx)
1687 2 : .await?,
1688 2 : 3
1689 2 : );
1690 2 : assert_eq!(
1691 2 : tline
1692 2 : .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
1693 2 : .instrument(test_span.clone())
1694 2 : .await?,
1695 2 : test_img("foo blk 2 at 5")
1696 2 : );
1697 2 :
1698 2 : // Truncate to zero length
1699 2 : let mut m = tline.begin_modification(Lsn(0x68));
1700 2 : walingest
1701 2 : .put_rel_truncation(&mut m, TESTREL_A, 0, &ctx)
1702 2 : .await?;
1703 2 : m.commit(&ctx).await?;
1704 2 : assert_eq!(
1705 2 : tline
1706 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x68)), &ctx)
1707 2 : .await?,
1708 2 : 0
1709 2 : );
1710 2 :
1711 2 : // Extend from 0 to 2 blocks, leaving a gap
1712 2 : let mut m = tline.begin_modification(Lsn(0x70));
1713 2 : walingest
1714 2 : .put_rel_page_image(&mut m, TESTREL_A, 1, test_img("foo blk 1"), &ctx)
1715 2 : .await?;
1716 2 : m.commit(&ctx).await?;
1717 2 : assert_eq!(
1718 2 : tline
1719 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x70)), &ctx)
1720 2 : .await?,
1721 2 : 2
1722 2 : );
1723 2 : assert_eq!(
1724 2 : tline
1725 2 : .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x70)), &ctx)
1726 2 : .instrument(test_span.clone())
1727 2 : .await?,
1728 2 : ZERO_PAGE
1729 2 : );
1730 2 : assert_eq!(
1731 2 : tline
1732 2 : .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x70)), &ctx)
1733 2 : .instrument(test_span.clone())
1734 2 : .await?,
1735 2 : test_img("foo blk 1")
1736 2 : );
1737 2 :
1738 2 : // Extend a lot more, leaving a big gap that spans across segments
1739 2 : let mut m = tline.begin_modification(Lsn(0x80));
1740 2 : walingest
1741 2 : .put_rel_page_image(&mut m, TESTREL_A, 1500, test_img("foo blk 1500"), &ctx)
1742 2 : .await?;
1743 2 : m.commit(&ctx).await?;
1744 2 : assert_eq!(
1745 2 : tline
1746 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x80)), &ctx)
1747 2 : .await?,
1748 2 : 1501
1749 2 : );
1750 2998 : for blk in 2..1500 {
1751 2996 : assert_eq!(
1752 2996 : tline
1753 2996 : .get_rel_page_at_lsn(TESTREL_A, blk, Version::Lsn(Lsn(0x80)), &ctx)
1754 2996 : .instrument(test_span.clone())
1755 2996 : .await?,
1756 2996 : ZERO_PAGE
1757 2 : );
1758 2 : }
1759 2 : assert_eq!(
1760 2 : tline
1761 2 : .get_rel_page_at_lsn(TESTREL_A, 1500, Version::Lsn(Lsn(0x80)), &ctx)
1762 2 : .instrument(test_span.clone())
1763 2 : .await?,
1764 2 : test_img("foo blk 1500")
1765 2 : );
1766 2 :
1767 2 : Ok(())
1768 2 : }
1769 :
1770 : // Test what happens if we dropped a relation
1771 : // and then created it again within the same layer.
1772 : #[tokio::test]
1773 2 : async fn test_drop_extend() -> Result<()> {
1774 2 : let (tenant, ctx) = TenantHarness::create("test_drop_extend")
1775 2 : .await?
1776 2 : .load()
1777 2 : .await;
1778 2 : let tline = tenant
1779 2 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
1780 2 : .await?;
1781 2 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
1782 2 :
1783 2 : let mut m = tline.begin_modification(Lsn(0x20));
1784 2 : walingest
1785 2 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 2"), &ctx)
1786 2 : .await?;
1787 2 : m.commit(&ctx).await?;
1788 2 :
1789 2 : // Check that rel exists and size is correct
1790 2 : assert_eq!(
1791 2 : tline
1792 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1793 2 : .await?,
1794 2 : true
1795 2 : );
1796 2 : assert_eq!(
1797 2 : tline
1798 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1799 2 : .await?,
1800 2 : 1
1801 2 : );
1802 2 :
1803 2 : // Drop rel
1804 2 : let mut m = tline.begin_modification(Lsn(0x30));
1805 2 : let mut rel_drops = HashMap::new();
1806 2 : rel_drops.insert((TESTREL_A.spcnode, TESTREL_A.dbnode), vec![TESTREL_A]);
1807 2 : m.put_rel_drops(rel_drops, &ctx).await?;
1808 2 : m.commit(&ctx).await?;
1809 2 :
1810 2 : // Check that rel is not visible anymore
1811 2 : assert_eq!(
1812 2 : tline
1813 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x30)), &ctx)
1814 2 : .await?,
1815 2 : false
1816 2 : );
1817 2 :
1818 2 : // FIXME: should fail
1819 2 : //assert!(tline.get_rel_size(TESTREL_A, Lsn(0x30), false)?.is_none());
1820 2 :
1821 2 : // Re-create it
1822 2 : let mut m = tline.begin_modification(Lsn(0x40));
1823 2 : walingest
1824 2 : .put_rel_page_image(&mut m, TESTREL_A, 0, test_img("foo blk 0 at 4"), &ctx)
1825 2 : .await?;
1826 2 : m.commit(&ctx).await?;
1827 2 :
1828 2 : // Check that rel exists and size is correct
1829 2 : assert_eq!(
1830 2 : tline
1831 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x40)), &ctx)
1832 2 : .await?,
1833 2 : true
1834 2 : );
1835 2 : assert_eq!(
1836 2 : tline
1837 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x40)), &ctx)
1838 2 : .await?,
1839 2 : 1
1840 2 : );
1841 2 :
1842 2 : Ok(())
1843 2 : }
1844 :
1845 : // Test what happens if we truncated a relation
1846 : // so that one of its segments was dropped
1847 : // and then extended it again within the same layer.
1848 : #[tokio::test]
1849 2 : async fn test_truncate_extend() -> Result<()> {
1850 2 : let (tenant, ctx) = TenantHarness::create("test_truncate_extend")
1851 2 : .await?
1852 2 : .load()
1853 2 : .await;
1854 2 : let tline = tenant
1855 2 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
1856 2 : .await?;
1857 2 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
1858 2 :
1859 2 : // Create a 20 MB relation (the size is arbitrary)
1860 2 : let relsize = 20 * 1024 * 1024 / 8192;
1861 2 : let mut m = tline.begin_modification(Lsn(0x20));
1862 5120 : for blkno in 0..relsize {
1863 5120 : let data = format!("foo blk {} at {}", blkno, Lsn(0x20));
1864 5120 : walingest
1865 5120 : .put_rel_page_image(&mut m, TESTREL_A, blkno, test_img(&data), &ctx)
1866 5120 : .await?;
1867 2 : }
1868 2 : m.commit(&ctx).await?;
1869 2 :
1870 2 : let test_span = tracing::info_span!(parent: None, "test",
1871 2 : tenant_id=%tline.tenant_shard_id.tenant_id,
1872 0 : shard_id=%tline.tenant_shard_id.shard_slug(),
1873 0 : timeline_id=%tline.timeline_id);
1874 2 :
1875 2 : // The relation was created at LSN 20, not visible at LSN 1 yet.
1876 2 : assert_eq!(
1877 2 : tline
1878 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x10)), &ctx)
1879 2 : .await?,
1880 2 : false
1881 2 : );
1882 2 : assert!(tline
1883 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x10)), &ctx)
1884 2 : .await
1885 2 : .is_err());
1886 2 :
1887 2 : assert_eq!(
1888 2 : tline
1889 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1890 2 : .await?,
1891 2 : true
1892 2 : );
1893 2 : assert_eq!(
1894 2 : tline
1895 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x20)), &ctx)
1896 2 : .await?,
1897 2 : relsize
1898 2 : );
1899 2 :
1900 2 : // Check relation content
1901 5120 : for blkno in 0..relsize {
1902 5120 : let lsn = Lsn(0x20);
1903 5120 : let data = format!("foo blk {} at {}", blkno, lsn);
1904 5120 : assert_eq!(
1905 5120 : tline
1906 5120 : .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(lsn), &ctx)
1907 5120 : .instrument(test_span.clone())
1908 5120 : .await?,
1909 5120 : test_img(&data)
1910 2 : );
1911 2 : }
1912 2 :
1913 2 : // Truncate relation so that second segment was dropped
1914 2 : // - only leave one page
1915 2 : let mut m = tline.begin_modification(Lsn(0x60));
1916 2 : walingest
1917 2 : .put_rel_truncation(&mut m, TESTREL_A, 1, &ctx)
1918 2 : .await?;
1919 2 : m.commit(&ctx).await?;
1920 2 :
1921 2 : // Check reported size and contents after truncation
1922 2 : assert_eq!(
1923 2 : tline
1924 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x60)), &ctx)
1925 2 : .await?,
1926 2 : 1
1927 2 : );
1928 2 :
1929 4 : for blkno in 0..1 {
1930 2 : let lsn = Lsn(0x20);
1931 2 : let data = format!("foo blk {} at {}", blkno, lsn);
1932 2 : assert_eq!(
1933 2 : tline
1934 2 : .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x60)), &ctx)
1935 2 : .instrument(test_span.clone())
1936 2 : .await?,
1937 2 : test_img(&data)
1938 2 : );
1939 2 : }
1940 2 :
1941 2 : // should still see all blocks with older LSN
1942 2 : assert_eq!(
1943 2 : tline
1944 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x50)), &ctx)
1945 2 : .await?,
1946 2 : relsize
1947 2 : );
1948 5120 : for blkno in 0..relsize {
1949 5120 : let lsn = Lsn(0x20);
1950 5120 : let data = format!("foo blk {} at {}", blkno, lsn);
1951 5120 : assert_eq!(
1952 5120 : tline
1953 5120 : .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x50)), &ctx)
1954 5120 : .instrument(test_span.clone())
1955 5120 : .await?,
1956 5120 : test_img(&data)
1957 2 : );
1958 2 : }
1959 2 :
1960 2 : // Extend relation again.
1961 2 : // Add enough blocks to create second segment
1962 2 : let lsn = Lsn(0x80);
1963 2 : let mut m = tline.begin_modification(lsn);
1964 5120 : for blkno in 0..relsize {
1965 5120 : let data = format!("foo blk {} at {}", blkno, lsn);
1966 5120 : walingest
1967 5120 : .put_rel_page_image(&mut m, TESTREL_A, blkno, test_img(&data), &ctx)
1968 5120 : .await?;
1969 2 : }
1970 2 : m.commit(&ctx).await?;
1971 2 :
1972 2 : assert_eq!(
1973 2 : tline
1974 2 : .get_rel_exists(TESTREL_A, Version::Lsn(Lsn(0x80)), &ctx)
1975 2 : .await?,
1976 2 : true
1977 2 : );
1978 2 : assert_eq!(
1979 2 : tline
1980 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(0x80)), &ctx)
1981 2 : .await?,
1982 2 : relsize
1983 2 : );
1984 2 : // Check relation content
1985 5120 : for blkno in 0..relsize {
1986 5120 : let lsn = Lsn(0x80);
1987 5120 : let data = format!("foo blk {} at {}", blkno, lsn);
1988 5120 : assert_eq!(
1989 5120 : tline
1990 5120 : .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x80)), &ctx)
1991 5120 : .instrument(test_span.clone())
1992 5120 : .await?,
1993 5120 : test_img(&data)
1994 2 : );
1995 2 : }
1996 2 :
1997 2 : Ok(())
1998 2 : }
1999 :
2000 : /// Test get_relsize() and truncation with a file larger than 1 GB, so that it's
2001 : /// split into multiple 1 GB segments in Postgres.
2002 : #[tokio::test]
2003 2 : async fn test_large_rel() -> Result<()> {
2004 2 : let (tenant, ctx) = TenantHarness::create("test_large_rel").await?.load().await;
2005 2 : let tline = tenant
2006 2 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2007 2 : .await?;
2008 2 : let mut walingest = init_walingest_test(&tline, &ctx).await?;
2009 2 :
2010 2 : let mut lsn = 0x10;
2011 262146 : for blknum in 0..RELSEG_SIZE + 1 {
2012 262146 : lsn += 0x10;
2013 262146 : let mut m = tline.begin_modification(Lsn(lsn));
2014 262146 : let img = test_img(&format!("foo blk {} at {}", blknum, Lsn(lsn)));
2015 262146 : walingest
2016 262146 : .put_rel_page_image(&mut m, TESTREL_A, blknum as BlockNumber, img, &ctx)
2017 262146 : .await?;
2018 262146 : m.commit(&ctx).await?;
2019 2 : }
2020 2 :
2021 2 : assert_current_logical_size(&tline, Lsn(lsn));
2022 2 :
2023 2 : assert_eq!(
2024 2 : tline
2025 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(lsn)), &ctx)
2026 2 : .await?,
2027 2 : RELSEG_SIZE + 1
2028 2 : );
2029 2 :
2030 2 : // Truncate one block
2031 2 : lsn += 0x10;
2032 2 : let mut m = tline.begin_modification(Lsn(lsn));
2033 2 : walingest
2034 2 : .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE, &ctx)
2035 2 : .await?;
2036 2 : m.commit(&ctx).await?;
2037 2 : assert_eq!(
2038 2 : tline
2039 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(lsn)), &ctx)
2040 2 : .await?,
2041 2 : RELSEG_SIZE
2042 2 : );
2043 2 : assert_current_logical_size(&tline, Lsn(lsn));
2044 2 :
2045 2 : // Truncate another block
2046 2 : lsn += 0x10;
2047 2 : let mut m = tline.begin_modification(Lsn(lsn));
2048 2 : walingest
2049 2 : .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE - 1, &ctx)
2050 2 : .await?;
2051 2 : m.commit(&ctx).await?;
2052 2 : assert_eq!(
2053 2 : tline
2054 2 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(lsn)), &ctx)
2055 2 : .await?,
2056 2 : RELSEG_SIZE - 1
2057 2 : );
2058 2 : assert_current_logical_size(&tline, Lsn(lsn));
2059 2 :
2060 2 : // Truncate to 1500, and then truncate all the way down to 0, one block at a time
2061 2 : // This tests the behavior at segment boundaries
2062 2 : let mut size: i32 = 3000;
2063 6004 : while size >= 0 {
2064 6002 : lsn += 0x10;
2065 6002 : let mut m = tline.begin_modification(Lsn(lsn));
2066 6002 : walingest
2067 6002 : .put_rel_truncation(&mut m, TESTREL_A, size as BlockNumber, &ctx)
2068 6002 : .await?;
2069 6002 : m.commit(&ctx).await?;
2070 6002 : assert_eq!(
2071 6002 : tline
2072 6002 : .get_rel_size(TESTREL_A, Version::Lsn(Lsn(lsn)), &ctx)
2073 6002 : .await?,
2074 6002 : size as BlockNumber
2075 2 : );
2076 2 :
2077 6002 : size -= 1;
2078 2 : }
2079 2 : assert_current_logical_size(&tline, Lsn(lsn));
2080 2 :
2081 2 : Ok(())
2082 2 : }
2083 :
2084 : /// Replay a wal segment file taken directly from safekeepers.
2085 : ///
2086 : /// This test is useful for benchmarking since it allows us to profile only
2087 : /// the walingest code in a single-threaded executor, and iterate more quickly
2088 : /// without waiting for unrelated steps.
2089 : #[tokio::test]
2090 2 : async fn test_ingest_real_wal() {
2091 2 : use crate::tenant::harness::*;
2092 2 : use postgres_ffi::waldecoder::WalStreamDecoder;
2093 2 : use postgres_ffi::WAL_SEGMENT_SIZE;
2094 2 :
2095 2 : // Define test data path and constants.
2096 2 : //
2097 2 : // Steps to reconstruct the data, if needed:
2098 2 : // 1. Run the pgbench python test
2099 2 : // 2. Take the first wal segment file from safekeeper
2100 2 : // 3. Compress it using `zstd --long input_file`
2101 2 : // 4. Copy initdb.tar.zst from local_fs_remote_storage
2102 2 : // 5. Grep sk logs for "restart decoder" to get startpoint
2103 2 : // 6. Run just the decoder from this test to get the endpoint.
2104 2 : // It's the last LSN the decoder will output.
2105 2 : let pg_version = 15; // The test data was generated by pg15
2106 2 : let path = "test_data/sk_wal_segment_from_pgbench";
2107 2 : let wal_segment_path = format!("{path}/000000010000000000000001.zst");
2108 2 : let source_initdb_path = format!("{path}/{INITDB_PATH}");
2109 2 : let startpoint = Lsn::from_hex("14AEC08").unwrap();
2110 2 : let _endpoint = Lsn::from_hex("1FFFF98").unwrap();
2111 2 :
2112 2 : let harness = TenantHarness::create("test_ingest_real_wal").await.unwrap();
2113 2 : let span = harness
2114 2 : .span()
2115 2 : .in_scope(|| info_span!("timeline_span", timeline_id=%TIMELINE_ID));
2116 2 : let (tenant, ctx) = harness.load().await;
2117 2 :
2118 2 : let remote_initdb_path =
2119 2 : remote_initdb_archive_path(&tenant.tenant_shard_id().tenant_id, &TIMELINE_ID);
2120 2 : let initdb_path = harness.remote_fs_dir.join(remote_initdb_path.get_path());
2121 2 :
2122 2 : std::fs::create_dir_all(initdb_path.parent().unwrap())
2123 2 : .expect("creating test dir should work");
2124 2 : std::fs::copy(source_initdb_path, initdb_path).expect("copying the initdb.tar.zst works");
2125 2 :
2126 2 : // Bootstrap a real timeline. We can't use create_test_timeline because
2127 2 : // it doesn't create a real checkpoint, and Walingest::new tries to parse
2128 2 : // the garbage data.
2129 2 : let tline = tenant
2130 2 : .bootstrap_timeline_test(TIMELINE_ID, pg_version, Some(TIMELINE_ID), &ctx)
2131 2 : .await
2132 2 : .unwrap();
2133 2 :
2134 2 : // We fully read and decompress this into memory before decoding
2135 2 : // to get a more accurate perf profile of the decoder.
2136 2 : let bytes = {
2137 2 : use async_compression::tokio::bufread::ZstdDecoder;
2138 2 : let file = tokio::fs::File::open(wal_segment_path).await.unwrap();
2139 2 : let reader = tokio::io::BufReader::new(file);
2140 2 : let decoder = ZstdDecoder::new(reader);
2141 2 : let mut reader = tokio::io::BufReader::new(decoder);
2142 2 : let mut buffer = Vec::new();
2143 2 : tokio::io::copy_buf(&mut reader, &mut buffer).await.unwrap();
2144 2 : buffer
2145 2 : };
2146 2 :
2147 2 : // TODO start a profiler too
2148 2 : let started_at = std::time::Instant::now();
2149 2 :
2150 2 : // Initialize walingest
2151 2 : let xlogoff: usize = startpoint.segment_offset(WAL_SEGMENT_SIZE);
2152 2 : let mut decoder = WalStreamDecoder::new(startpoint, pg_version);
2153 2 : let mut walingest = WalIngest::new(tline.as_ref(), startpoint, &ctx)
2154 2 : .await
2155 2 : .unwrap();
2156 2 : let mut modification = tline.begin_modification(startpoint);
2157 2 : println!("decoding {} bytes", bytes.len() - xlogoff);
2158 2 :
2159 2 : // Decode and ingest wal. We process the wal in chunks because
2160 2 : // that's what happens when we get bytes from safekeepers.
2161 474686 : for chunk in bytes[xlogoff..].chunks(50) {
2162 474686 : decoder.feed_bytes(chunk);
2163 620536 : while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() {
2164 145850 : let interpreted = InterpretedWalRecord::from_bytes_filtered(
2165 145850 : recdata,
2166 145850 : modification.tline.get_shard_identity(),
2167 145850 : lsn,
2168 145850 : modification.tline.pg_version,
2169 145850 : )
2170 145850 : .unwrap();
2171 145850 :
2172 145850 : walingest
2173 145850 : .ingest_record(interpreted, &mut modification, &ctx)
2174 145850 : .instrument(span.clone())
2175 145850 : .await
2176 145850 : .unwrap();
2177 2 : }
2178 474686 : modification.commit(&ctx).await.unwrap();
2179 2 : }
2180 2 :
2181 2 : let duration = started_at.elapsed();
2182 2 : println!("done in {:?}", duration);
2183 2 : }
2184 : }
|
