Line data Source code
1 : //!
2 : //! This provides an abstraction to store PostgreSQL relations and other files
3 : //! in the key-value store that implements the Repository interface.
4 : //!
5 : //! (TODO: The line between PUT-functions here and walingest.rs is a bit blurry, as
6 : //! walingest.rs handles a few things like implicit relation creation and extension.
7 : //! Clarify that)
8 : //!
9 : use super::tenant::{PageReconstructError, Timeline};
10 : use crate::aux_file;
11 : use crate::context::RequestContext;
12 : use crate::keyspace::{KeySpace, KeySpaceAccum};
13 : use crate::metrics::{
14 : RELSIZE_CACHE_ENTRIES, RELSIZE_CACHE_HITS, RELSIZE_CACHE_MISSES, RELSIZE_CACHE_MISSES_OLD,
15 : };
16 : use crate::span::{
17 : debug_assert_current_span_has_tenant_and_timeline_id,
18 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
19 : };
20 : use crate::tenant::timeline::GetVectoredError;
21 : use anyhow::{ensure, Context};
22 : use bytes::{Buf, Bytes, BytesMut};
23 : use enum_map::Enum;
24 : use itertools::Itertools;
25 : use pageserver_api::key::Key;
26 : use pageserver_api::key::{
27 : dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range, rel_size_to_key,
28 : relmap_file_key, repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
29 : slru_segment_key_range, slru_segment_size_to_key, twophase_file_key, twophase_key_range,
30 : CompactKey, AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, DBDIR_KEY, TWOPHASEDIR_KEY,
31 : };
32 : use pageserver_api::keyspace::SparseKeySpace;
33 : use pageserver_api::record::NeonWalRecord;
34 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
35 : use pageserver_api::shard::ShardIdentity;
36 : use pageserver_api::value::Value;
37 : use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
38 : use postgres_ffi::BLCKSZ;
39 : use postgres_ffi::{Oid, RepOriginId, TimestampTz, TransactionId};
40 : use serde::{Deserialize, Serialize};
41 : use std::collections::{hash_map, BTreeMap, HashMap, HashSet};
42 : use std::ops::ControlFlow;
43 : use std::ops::Range;
44 : use strum::IntoEnumIterator;
45 : use tokio_util::sync::CancellationToken;
46 : use tracing::{debug, trace, warn};
47 : use utils::bin_ser::DeserializeError;
48 : use utils::pausable_failpoint;
49 : use utils::{bin_ser::BeSer, lsn::Lsn};
50 : use wal_decoder::serialized_batch::SerializedValueBatch;
51 :
52 : /// Max delta records appended to the AUX_FILES_KEY (for aux v1). The write path will write a full image once this threshold is reached.
53 : pub const MAX_AUX_FILE_DELTAS: usize = 1024;
54 :
55 : /// Max number of aux-file-related delta layers. The compaction will create a new image layer once this threshold is reached.
56 : pub const MAX_AUX_FILE_V2_DELTAS: usize = 16;
57 :
58 : #[derive(Debug)]
59 : pub enum LsnForTimestamp {
60 : /// Found commits both before and after the given timestamp
61 : Present(Lsn),
62 :
63 : /// Found no commits after the given timestamp, this means
64 : /// that the newest data in the branch is older than the given
65 : /// timestamp.
66 : ///
67 : /// All commits <= LSN happened before the given timestamp
68 : Future(Lsn),
69 :
70 : /// The queried timestamp is past our horizon we look back at (PITR)
71 : ///
72 : /// All commits > LSN happened after the given timestamp,
73 : /// but any commits < LSN might have happened before or after
74 : /// the given timestamp. We don't know because no data before
75 : /// the given lsn is available.
76 : Past(Lsn),
77 :
78 : /// We have found no commit with a timestamp,
79 : /// so we can't return anything meaningful.
80 : ///
81 : /// The associated LSN is the lower bound value we can safely
82 : /// create branches on, but no statement is made if it is
83 : /// older or newer than the timestamp.
84 : ///
85 : /// This variant can e.g. be returned right after a
86 : /// cluster import.
87 : NoData(Lsn),
88 : }
89 :
90 : #[derive(Debug, thiserror::Error)]
91 : pub(crate) enum CalculateLogicalSizeError {
92 : #[error("cancelled")]
93 : Cancelled,
94 :
95 : /// Something went wrong while reading the metadata we use to calculate logical size
96 : /// Note that cancellation variants of `PageReconstructError` are transformed to [`Self::Cancelled`]
97 : /// in the `From` implementation for this variant.
98 : #[error(transparent)]
99 : PageRead(PageReconstructError),
100 :
101 : /// Something went wrong deserializing metadata that we read to calculate logical size
102 : #[error("decode error: {0}")]
103 : Decode(#[from] DeserializeError),
104 : }
105 :
106 : #[derive(Debug, thiserror::Error)]
107 : pub(crate) enum CollectKeySpaceError {
108 : #[error(transparent)]
109 : Decode(#[from] DeserializeError),
110 : #[error(transparent)]
111 : PageRead(PageReconstructError),
112 : #[error("cancelled")]
113 : Cancelled,
114 : }
115 :
116 : impl From<PageReconstructError> for CollectKeySpaceError {
117 0 : fn from(err: PageReconstructError) -> Self {
118 0 : match err {
119 0 : PageReconstructError::Cancelled => Self::Cancelled,
120 0 : err => Self::PageRead(err),
121 : }
122 0 : }
123 : }
124 :
125 : impl From<PageReconstructError> for CalculateLogicalSizeError {
126 0 : fn from(pre: PageReconstructError) -> Self {
127 0 : match pre {
128 0 : PageReconstructError::Cancelled => Self::Cancelled,
129 0 : _ => Self::PageRead(pre),
130 : }
131 0 : }
132 : }
133 :
134 : #[derive(Debug, thiserror::Error)]
135 : pub enum RelationError {
136 : #[error("Relation Already Exists")]
137 : AlreadyExists,
138 : #[error("invalid relnode")]
139 : InvalidRelnode,
140 : #[error(transparent)]
141 : Other(#[from] anyhow::Error),
142 : }
143 :
144 : ///
145 : /// This impl provides all the functionality to store PostgreSQL relations, SLRUs,
146 : /// and other special kinds of files, in a versioned key-value store. The
147 : /// Timeline struct provides the key-value store.
148 : ///
149 : /// This is a separate impl, so that we can easily include all these functions in a Timeline
150 : /// implementation, and might be moved into a separate struct later.
151 : impl Timeline {
152 : /// Start ingesting a WAL record, or other atomic modification of
153 : /// the timeline.
154 : ///
155 : /// This provides a transaction-like interface to perform a bunch
156 : /// of modifications atomically.
157 : ///
158 : /// To ingest a WAL record, call begin_modification(lsn) to get a
159 : /// DatadirModification object. Use the functions in the object to
160 : /// modify the repository state, updating all the pages and metadata
161 : /// that the WAL record affects. When you're done, call commit() to
162 : /// commit the changes.
163 : ///
164 : /// Lsn stored in modification is advanced by `ingest_record` and
165 : /// is used by `commit()` to update `last_record_lsn`.
166 : ///
167 : /// Calling commit() will flush all the changes and reset the state,
168 : /// so the `DatadirModification` struct can be reused to perform the next modification.
169 : ///
170 : /// Note that any pending modifications you make through the
171 : /// modification object won't be visible to calls to the 'get' and list
172 : /// functions of the timeline until you finish! And if you update the
173 : /// same page twice, the last update wins.
174 : ///
175 268380 : pub fn begin_modification(&self, lsn: Lsn) -> DatadirModification
176 268380 : where
177 268380 : Self: Sized,
178 268380 : {
179 268380 : DatadirModification {
180 268380 : tline: self,
181 268380 : pending_lsns: Vec::new(),
182 268380 : pending_metadata_pages: HashMap::new(),
183 268380 : pending_data_batch: None,
184 268380 : pending_deletions: Vec::new(),
185 268380 : pending_nblocks: 0,
186 268380 : pending_directory_entries: Vec::new(),
187 268380 : pending_metadata_bytes: 0,
188 268380 : lsn,
189 268380 : }
190 268380 : }
191 :
192 : //------------------------------------------------------------------------------
193 : // Public GET functions
194 : //------------------------------------------------------------------------------
195 :
196 : /// Look up given page version.
197 18384 : pub(crate) async fn get_rel_page_at_lsn(
198 18384 : &self,
199 18384 : tag: RelTag,
200 18384 : blknum: BlockNumber,
201 18384 : version: Version<'_>,
202 18384 : ctx: &RequestContext,
203 18384 : ) -> Result<Bytes, PageReconstructError> {
204 18384 : match version {
205 18384 : Version::Lsn(effective_lsn) => {
206 18384 : let pages: smallvec::SmallVec<[_; 1]> = smallvec::smallvec![(tag, blknum)];
207 18384 : let res = self
208 18384 : .get_rel_page_at_lsn_batched(
209 18384 : pages.iter().map(|(tag, blknum)| (tag, blknum)),
210 18384 : effective_lsn,
211 18384 : ctx,
212 18384 : )
213 18384 : .await;
214 18384 : assert_eq!(res.len(), 1);
215 18384 : res.into_iter().next().unwrap()
216 : }
217 0 : Version::Modified(modification) => {
218 0 : if tag.relnode == 0 {
219 0 : return Err(PageReconstructError::Other(
220 0 : RelationError::InvalidRelnode.into(),
221 0 : ));
222 0 : }
223 :
224 0 : let nblocks = self.get_rel_size(tag, version, ctx).await?;
225 0 : if blknum >= nblocks {
226 0 : debug!(
227 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
228 0 : tag,
229 0 : blknum,
230 0 : version.get_lsn(),
231 : nblocks
232 : );
233 0 : return Ok(ZERO_PAGE.clone());
234 0 : }
235 0 :
236 0 : let key = rel_block_to_key(tag, blknum);
237 0 : modification.get(key, ctx).await
238 : }
239 : }
240 18384 : }
241 :
242 : /// Like [`Self::get_rel_page_at_lsn`], but returns a batch of pages.
243 : ///
244 : /// The ordering of the returned vec corresponds to the ordering of `pages`.
245 18384 : pub(crate) async fn get_rel_page_at_lsn_batched(
246 18384 : &self,
247 18384 : pages: impl ExactSizeIterator<Item = (&RelTag, &BlockNumber)>,
248 18384 : effective_lsn: Lsn,
249 18384 : ctx: &RequestContext,
250 18384 : ) -> Vec<Result<Bytes, PageReconstructError>> {
251 18384 : debug_assert_current_span_has_tenant_and_timeline_id();
252 18384 :
253 18384 : let mut slots_filled = 0;
254 18384 : let page_count = pages.len();
255 18384 :
256 18384 : // Would be nice to use smallvec here but it doesn't provide the spare_capacity_mut() API.
257 18384 : let mut result = Vec::with_capacity(pages.len());
258 18384 : let result_slots = result.spare_capacity_mut();
259 18384 :
260 18384 : let mut keys_slots: BTreeMap<Key, smallvec::SmallVec<[usize; 1]>> = BTreeMap::default();
261 18384 : for (response_slot_idx, (tag, blknum)) in pages.enumerate() {
262 18384 : if tag.relnode == 0 {
263 0 : result_slots[response_slot_idx].write(Err(PageReconstructError::Other(
264 0 : RelationError::InvalidRelnode.into(),
265 0 : )));
266 0 :
267 0 : slots_filled += 1;
268 0 : continue;
269 18384 : }
270 :
271 18384 : let nblocks = match self
272 18384 : .get_rel_size(*tag, Version::Lsn(effective_lsn), ctx)
273 18384 : .await
274 : {
275 18384 : Ok(nblocks) => nblocks,
276 0 : Err(err) => {
277 0 : result_slots[response_slot_idx].write(Err(err));
278 0 : slots_filled += 1;
279 0 : continue;
280 : }
281 : };
282 :
283 18384 : if *blknum >= nblocks {
284 0 : debug!(
285 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
286 : tag, blknum, effective_lsn, nblocks
287 : );
288 0 : result_slots[response_slot_idx].write(Ok(ZERO_PAGE.clone()));
289 0 : slots_filled += 1;
290 0 : continue;
291 18384 : }
292 18384 :
293 18384 : let key = rel_block_to_key(*tag, *blknum);
294 18384 :
295 18384 : let key_slots = keys_slots.entry(key).or_default();
296 18384 : key_slots.push(response_slot_idx);
297 : }
298 :
299 18384 : let keyspace = {
300 : // add_key requires monotonicity
301 18384 : let mut acc = KeySpaceAccum::new();
302 18384 : for key in keys_slots
303 18384 : .keys()
304 18384 : // in fact it requires strong monotonicity
305 18384 : .dedup()
306 18384 : {
307 18384 : acc.add_key(*key);
308 18384 : }
309 18384 : acc.to_keyspace()
310 18384 : };
311 18384 :
312 18384 : match self.get_vectored(keyspace, effective_lsn, ctx).await {
313 18384 : Ok(results) => {
314 36768 : for (key, res) in results {
315 18384 : let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
316 18384 : let first_slot = key_slots.next().unwrap();
317 :
318 18384 : for slot in key_slots {
319 0 : let clone = match &res {
320 0 : Ok(buf) => Ok(buf.clone()),
321 0 : Err(err) => Err(match err {
322 : PageReconstructError::Cancelled => {
323 0 : PageReconstructError::Cancelled
324 : }
325 :
326 0 : x @ PageReconstructError::Other(_) |
327 0 : x @ PageReconstructError::AncestorLsnTimeout(_) |
328 0 : x @ PageReconstructError::WalRedo(_) |
329 0 : x @ PageReconstructError::MissingKey(_) => {
330 0 : PageReconstructError::Other(anyhow::anyhow!("there was more than one request for this key in the batch, error logged once: {x:?}"))
331 : },
332 : }),
333 : };
334 :
335 0 : result_slots[slot].write(clone);
336 0 : slots_filled += 1;
337 : }
338 :
339 18384 : result_slots[first_slot].write(res);
340 18384 : slots_filled += 1;
341 : }
342 : }
343 0 : Err(err) => {
344 : // this cannot really happen because get_vectored only errors globally on invalid LSN or too large batch size
345 : // (We enforce the max batch size outside of this function, in the code that constructs the batch request.)
346 0 : for slot in keys_slots.values().flatten() {
347 : // this whole `match` is a lot like `From<GetVectoredError> for PageReconstructError`
348 : // but without taking ownership of the GetVectoredError
349 0 : let err = match &err {
350 : GetVectoredError::Cancelled => {
351 0 : Err(PageReconstructError::Cancelled)
352 : }
353 : // TODO: restructure get_vectored API to make this error per-key
354 0 : GetVectoredError::MissingKey(err) => {
355 0 : Err(PageReconstructError::Other(anyhow::anyhow!("whole vectored get request failed because one or more of the requested keys were missing: {err:?}")))
356 : }
357 : // TODO: restructure get_vectored API to make this error per-key
358 0 : GetVectoredError::GetReadyAncestorError(err) => {
359 0 : Err(PageReconstructError::Other(anyhow::anyhow!("whole vectored get request failed because one or more key required ancestor that wasn't ready: {err:?}")))
360 : }
361 : // TODO: restructure get_vectored API to make this error per-key
362 0 : GetVectoredError::Other(err) => {
363 0 : Err(PageReconstructError::Other(
364 0 : anyhow::anyhow!("whole vectored get request failed: {err:?}"),
365 0 : ))
366 : }
367 : // TODO: we can prevent this error class by moving this check into the type system
368 0 : GetVectoredError::InvalidLsn(e) => {
369 0 : Err(anyhow::anyhow!("invalid LSN: {e:?}").into())
370 : }
371 : // NB: this should never happen in practice because we limit MAX_GET_VECTORED_KEYS
372 : // TODO: we can prevent this error class by moving this check into the type system
373 0 : GetVectoredError::Oversized(err) => {
374 0 : Err(anyhow::anyhow!(
375 0 : "batching oversized: {err:?}"
376 0 : )
377 0 : .into())
378 : }
379 : };
380 :
381 0 : result_slots[*slot].write(err);
382 : }
383 :
384 0 : slots_filled += keys_slots.values().map(|slots| slots.len()).sum::<usize>();
385 0 : }
386 : };
387 :
388 18384 : assert_eq!(slots_filled, page_count);
389 : // SAFETY:
390 : // 1. `result` and any of its uninint members are not read from until this point
391 : // 2. The length below is tracked at run-time and matches the number of requested pages.
392 18384 : unsafe {
393 18384 : result.set_len(page_count);
394 18384 : }
395 18384 :
396 18384 : result
397 18384 : }
398 :
399 : /// Get size of a database in blocks. This is only accurate on shard 0. It will undercount on
400 : /// other shards, by only accounting for relations the shard has pages for, and only accounting
401 : /// for pages up to the highest page number it has stored.
402 0 : pub(crate) async fn get_db_size(
403 0 : &self,
404 0 : spcnode: Oid,
405 0 : dbnode: Oid,
406 0 : version: Version<'_>,
407 0 : ctx: &RequestContext,
408 0 : ) -> Result<usize, PageReconstructError> {
409 0 : let mut total_blocks = 0;
410 :
411 0 : let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
412 :
413 0 : for rel in rels {
414 0 : let n_blocks = self.get_rel_size(rel, version, ctx).await?;
415 0 : total_blocks += n_blocks as usize;
416 : }
417 0 : Ok(total_blocks)
418 0 : }
419 :
420 : /// Get size of a relation file. The relation must exist, otherwise an error is returned.
421 : ///
422 : /// This is only accurate on shard 0. On other shards, it will return the size up to the highest
423 : /// page number stored in the shard.
424 24434 : pub(crate) async fn get_rel_size(
425 24434 : &self,
426 24434 : tag: RelTag,
427 24434 : version: Version<'_>,
428 24434 : ctx: &RequestContext,
429 24434 : ) -> Result<BlockNumber, PageReconstructError> {
430 24434 : if tag.relnode == 0 {
431 0 : return Err(PageReconstructError::Other(
432 0 : RelationError::InvalidRelnode.into(),
433 0 : ));
434 24434 : }
435 :
436 24434 : if let Some(nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
437 19294 : return Ok(nblocks);
438 5140 : }
439 5140 :
440 5140 : if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
441 0 : && !self.get_rel_exists(tag, version, ctx).await?
442 : {
443 : // FIXME: Postgres sometimes calls smgrcreate() to create
444 : // FSM, and smgrnblocks() on it immediately afterwards,
445 : // without extending it. Tolerate that by claiming that
446 : // any non-existent FSM fork has size 0.
447 0 : return Ok(0);
448 5140 : }
449 5140 :
450 5140 : let key = rel_size_to_key(tag);
451 5140 : let mut buf = version.get(self, key, ctx).await?;
452 5136 : let nblocks = buf.get_u32_le();
453 5136 :
454 5136 : self.update_cached_rel_size(tag, version.get_lsn(), nblocks);
455 5136 :
456 5136 : Ok(nblocks)
457 24434 : }
458 :
459 : /// Does the relation exist?
460 : ///
461 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
462 : /// the shard stores pages for.
463 6050 : pub(crate) async fn get_rel_exists(
464 6050 : &self,
465 6050 : tag: RelTag,
466 6050 : version: Version<'_>,
467 6050 : ctx: &RequestContext,
468 6050 : ) -> Result<bool, PageReconstructError> {
469 6050 : if tag.relnode == 0 {
470 0 : return Err(PageReconstructError::Other(
471 0 : RelationError::InvalidRelnode.into(),
472 0 : ));
473 6050 : }
474 :
475 : // first try to lookup relation in cache
476 6050 : if let Some(_nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
477 6032 : return Ok(true);
478 18 : }
479 : // then check if the database was already initialized.
480 : // get_rel_exists can be called before dbdir is created.
481 18 : let buf = version.get(self, DBDIR_KEY, ctx).await?;
482 18 : let dbdirs = DbDirectory::des(&buf)?.dbdirs;
483 18 : if !dbdirs.contains_key(&(tag.spcnode, tag.dbnode)) {
484 0 : return Ok(false);
485 18 : }
486 18 : // fetch directory listing
487 18 : let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
488 18 : let buf = version.get(self, key, ctx).await?;
489 :
490 18 : let dir = RelDirectory::des(&buf)?;
491 18 : Ok(dir.rels.contains(&(tag.relnode, tag.forknum)))
492 6050 : }
493 :
494 : /// Get a list of all existing relations in given tablespace and database.
495 : ///
496 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
497 : /// the shard stores pages for.
498 : ///
499 : /// # Cancel-Safety
500 : ///
501 : /// This method is cancellation-safe.
502 0 : pub(crate) async fn list_rels(
503 0 : &self,
504 0 : spcnode: Oid,
505 0 : dbnode: Oid,
506 0 : version: Version<'_>,
507 0 : ctx: &RequestContext,
508 0 : ) -> Result<HashSet<RelTag>, PageReconstructError> {
509 0 : // fetch directory listing
510 0 : let key = rel_dir_to_key(spcnode, dbnode);
511 0 : let buf = version.get(self, key, ctx).await?;
512 :
513 0 : let dir = RelDirectory::des(&buf)?;
514 0 : let rels: HashSet<RelTag> =
515 0 : HashSet::from_iter(dir.rels.iter().map(|(relnode, forknum)| RelTag {
516 0 : spcnode,
517 0 : dbnode,
518 0 : relnode: *relnode,
519 0 : forknum: *forknum,
520 0 : }));
521 0 :
522 0 : Ok(rels)
523 0 : }
524 :
525 : /// Get the whole SLRU segment
526 0 : pub(crate) async fn get_slru_segment(
527 0 : &self,
528 0 : kind: SlruKind,
529 0 : segno: u32,
530 0 : lsn: Lsn,
531 0 : ctx: &RequestContext,
532 0 : ) -> Result<Bytes, PageReconstructError> {
533 0 : assert!(self.tenant_shard_id.is_shard_zero());
534 0 : let n_blocks = self
535 0 : .get_slru_segment_size(kind, segno, Version::Lsn(lsn), ctx)
536 0 : .await?;
537 0 : let mut segment = BytesMut::with_capacity(n_blocks as usize * BLCKSZ as usize);
538 0 : for blkno in 0..n_blocks {
539 0 : let block = self
540 0 : .get_slru_page_at_lsn(kind, segno, blkno, lsn, ctx)
541 0 : .await?;
542 0 : segment.extend_from_slice(&block[..BLCKSZ as usize]);
543 : }
544 0 : Ok(segment.freeze())
545 0 : }
546 :
547 : /// Look up given SLRU page version.
548 0 : pub(crate) async fn get_slru_page_at_lsn(
549 0 : &self,
550 0 : kind: SlruKind,
551 0 : segno: u32,
552 0 : blknum: BlockNumber,
553 0 : lsn: Lsn,
554 0 : ctx: &RequestContext,
555 0 : ) -> Result<Bytes, PageReconstructError> {
556 0 : assert!(self.tenant_shard_id.is_shard_zero());
557 0 : let key = slru_block_to_key(kind, segno, blknum);
558 0 : self.get(key, lsn, ctx).await
559 0 : }
560 :
561 : /// Get size of an SLRU segment
562 0 : pub(crate) async fn get_slru_segment_size(
563 0 : &self,
564 0 : kind: SlruKind,
565 0 : segno: u32,
566 0 : version: Version<'_>,
567 0 : ctx: &RequestContext,
568 0 : ) -> Result<BlockNumber, PageReconstructError> {
569 0 : assert!(self.tenant_shard_id.is_shard_zero());
570 0 : let key = slru_segment_size_to_key(kind, segno);
571 0 : let mut buf = version.get(self, key, ctx).await?;
572 0 : Ok(buf.get_u32_le())
573 0 : }
574 :
575 : /// Get size of an SLRU segment
576 0 : pub(crate) async fn get_slru_segment_exists(
577 0 : &self,
578 0 : kind: SlruKind,
579 0 : segno: u32,
580 0 : version: Version<'_>,
581 0 : ctx: &RequestContext,
582 0 : ) -> Result<bool, PageReconstructError> {
583 0 : assert!(self.tenant_shard_id.is_shard_zero());
584 : // fetch directory listing
585 0 : let key = slru_dir_to_key(kind);
586 0 : let buf = version.get(self, key, ctx).await?;
587 :
588 0 : let dir = SlruSegmentDirectory::des(&buf)?;
589 0 : Ok(dir.segments.contains(&segno))
590 0 : }
591 :
592 : /// Locate LSN, such that all transactions that committed before
593 : /// 'search_timestamp' are visible, but nothing newer is.
594 : ///
595 : /// This is not exact. Commit timestamps are not guaranteed to be ordered,
596 : /// so it's not well defined which LSN you get if there were multiple commits
597 : /// "in flight" at that point in time.
598 : ///
599 0 : pub(crate) async fn find_lsn_for_timestamp(
600 0 : &self,
601 0 : search_timestamp: TimestampTz,
602 0 : cancel: &CancellationToken,
603 0 : ctx: &RequestContext,
604 0 : ) -> Result<LsnForTimestamp, PageReconstructError> {
605 0 : pausable_failpoint!("find-lsn-for-timestamp-pausable");
606 :
607 0 : let gc_cutoff_lsn_guard = self.get_latest_gc_cutoff_lsn();
608 0 : // We use this method to figure out the branching LSN for the new branch, but the
609 0 : // GC cutoff could be before the branching point and we cannot create a new branch
610 0 : // with LSN < `ancestor_lsn`. Thus, pick the maximum of these two to be
611 0 : // on the safe side.
612 0 : let min_lsn = std::cmp::max(*gc_cutoff_lsn_guard, self.get_ancestor_lsn());
613 0 : let max_lsn = self.get_last_record_lsn();
614 0 :
615 0 : // LSNs are always 8-byte aligned. low/mid/high represent the
616 0 : // LSN divided by 8.
617 0 : let mut low = min_lsn.0 / 8;
618 0 : let mut high = max_lsn.0 / 8 + 1;
619 0 :
620 0 : let mut found_smaller = false;
621 0 : let mut found_larger = false;
622 :
623 0 : while low < high {
624 0 : if cancel.is_cancelled() {
625 0 : return Err(PageReconstructError::Cancelled);
626 0 : }
627 0 : // cannot overflow, high and low are both smaller than u64::MAX / 2
628 0 : let mid = (high + low) / 2;
629 :
630 0 : let cmp = self
631 0 : .is_latest_commit_timestamp_ge_than(
632 0 : search_timestamp,
633 0 : Lsn(mid * 8),
634 0 : &mut found_smaller,
635 0 : &mut found_larger,
636 0 : ctx,
637 0 : )
638 0 : .await?;
639 :
640 0 : if cmp {
641 0 : high = mid;
642 0 : } else {
643 0 : low = mid + 1;
644 0 : }
645 : }
646 : // If `found_smaller == true`, `low = t + 1` where `t` is the target LSN,
647 : // so the LSN of the last commit record before or at `search_timestamp`.
648 : // Remove one from `low` to get `t`.
649 : //
650 : // FIXME: it would be better to get the LSN of the previous commit.
651 : // Otherwise, if you restore to the returned LSN, the database will
652 : // include physical changes from later commits that will be marked
653 : // as aborted, and will need to be vacuumed away.
654 0 : let commit_lsn = Lsn((low - 1) * 8);
655 0 : match (found_smaller, found_larger) {
656 : (false, false) => {
657 : // This can happen if no commit records have been processed yet, e.g.
658 : // just after importing a cluster.
659 0 : Ok(LsnForTimestamp::NoData(min_lsn))
660 : }
661 : (false, true) => {
662 : // Didn't find any commit timestamps smaller than the request
663 0 : Ok(LsnForTimestamp::Past(min_lsn))
664 : }
665 0 : (true, _) if commit_lsn < min_lsn => {
666 0 : // the search above did set found_smaller to true but it never increased the lsn.
667 0 : // Then, low is still the old min_lsn, and the subtraction above gave a value
668 0 : // below the min_lsn. We should never do that.
669 0 : Ok(LsnForTimestamp::Past(min_lsn))
670 : }
671 : (true, false) => {
672 : // Only found commits with timestamps smaller than the request.
673 : // It's still a valid case for branch creation, return it.
674 : // And `update_gc_info()` ignores LSN for a `LsnForTimestamp::Future`
675 : // case, anyway.
676 0 : Ok(LsnForTimestamp::Future(commit_lsn))
677 : }
678 0 : (true, true) => Ok(LsnForTimestamp::Present(commit_lsn)),
679 : }
680 0 : }
681 :
682 : /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
683 : /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
684 : ///
685 : /// Additionally, sets 'found_smaller'/'found_Larger, if encounters any commits
686 : /// with a smaller/larger timestamp.
687 : ///
688 0 : pub(crate) async fn is_latest_commit_timestamp_ge_than(
689 0 : &self,
690 0 : search_timestamp: TimestampTz,
691 0 : probe_lsn: Lsn,
692 0 : found_smaller: &mut bool,
693 0 : found_larger: &mut bool,
694 0 : ctx: &RequestContext,
695 0 : ) -> Result<bool, PageReconstructError> {
696 0 : self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
697 0 : if timestamp >= search_timestamp {
698 0 : *found_larger = true;
699 0 : return ControlFlow::Break(true);
700 0 : } else {
701 0 : *found_smaller = true;
702 0 : }
703 0 : ControlFlow::Continue(())
704 0 : })
705 0 : .await
706 0 : }
707 :
708 : /// Obtain the possible timestamp range for the given lsn.
709 : ///
710 : /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
711 0 : pub(crate) async fn get_timestamp_for_lsn(
712 0 : &self,
713 0 : probe_lsn: Lsn,
714 0 : ctx: &RequestContext,
715 0 : ) -> Result<Option<TimestampTz>, PageReconstructError> {
716 0 : let mut max: Option<TimestampTz> = None;
717 0 : self.map_all_timestamps::<()>(probe_lsn, ctx, |timestamp| {
718 0 : if let Some(max_prev) = max {
719 0 : max = Some(max_prev.max(timestamp));
720 0 : } else {
721 0 : max = Some(timestamp);
722 0 : }
723 0 : ControlFlow::Continue(())
724 0 : })
725 0 : .await?;
726 :
727 0 : Ok(max)
728 0 : }
729 :
730 : /// Runs the given function on all the timestamps for a given lsn
731 : ///
732 : /// The return value is either given by the closure, or set to the `Default`
733 : /// impl's output.
734 0 : async fn map_all_timestamps<T: Default>(
735 0 : &self,
736 0 : probe_lsn: Lsn,
737 0 : ctx: &RequestContext,
738 0 : mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
739 0 : ) -> Result<T, PageReconstructError> {
740 0 : for segno in self
741 0 : .list_slru_segments(SlruKind::Clog, Version::Lsn(probe_lsn), ctx)
742 0 : .await?
743 : {
744 0 : let nblocks = self
745 0 : .get_slru_segment_size(SlruKind::Clog, segno, Version::Lsn(probe_lsn), ctx)
746 0 : .await?;
747 0 : for blknum in (0..nblocks).rev() {
748 0 : let clog_page = self
749 0 : .get_slru_page_at_lsn(SlruKind::Clog, segno, blknum, probe_lsn, ctx)
750 0 : .await?;
751 :
752 0 : if clog_page.len() == BLCKSZ as usize + 8 {
753 0 : let mut timestamp_bytes = [0u8; 8];
754 0 : timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
755 0 : let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
756 0 :
757 0 : match f(timestamp) {
758 0 : ControlFlow::Break(b) => return Ok(b),
759 0 : ControlFlow::Continue(()) => (),
760 : }
761 0 : }
762 : }
763 : }
764 0 : Ok(Default::default())
765 0 : }
766 :
767 0 : pub(crate) async fn get_slru_keyspace(
768 0 : &self,
769 0 : version: Version<'_>,
770 0 : ctx: &RequestContext,
771 0 : ) -> Result<KeySpace, PageReconstructError> {
772 0 : let mut accum = KeySpaceAccum::new();
773 :
774 0 : for kind in SlruKind::iter() {
775 0 : let mut segments: Vec<u32> = self
776 0 : .list_slru_segments(kind, version, ctx)
777 0 : .await?
778 0 : .into_iter()
779 0 : .collect();
780 0 : segments.sort_unstable();
781 :
782 0 : for seg in segments {
783 0 : let block_count = self.get_slru_segment_size(kind, seg, version, ctx).await?;
784 :
785 0 : accum.add_range(
786 0 : slru_block_to_key(kind, seg, 0)..slru_block_to_key(kind, seg, block_count),
787 0 : );
788 : }
789 : }
790 :
791 0 : Ok(accum.to_keyspace())
792 0 : }
793 :
794 : /// Get a list of SLRU segments
795 0 : pub(crate) async fn list_slru_segments(
796 0 : &self,
797 0 : kind: SlruKind,
798 0 : version: Version<'_>,
799 0 : ctx: &RequestContext,
800 0 : ) -> Result<HashSet<u32>, PageReconstructError> {
801 0 : // fetch directory entry
802 0 : let key = slru_dir_to_key(kind);
803 :
804 0 : let buf = version.get(self, key, ctx).await?;
805 0 : Ok(SlruSegmentDirectory::des(&buf)?.segments)
806 0 : }
807 :
808 0 : pub(crate) async fn get_relmap_file(
809 0 : &self,
810 0 : spcnode: Oid,
811 0 : dbnode: Oid,
812 0 : version: Version<'_>,
813 0 : ctx: &RequestContext,
814 0 : ) -> Result<Bytes, PageReconstructError> {
815 0 : let key = relmap_file_key(spcnode, dbnode);
816 :
817 0 : let buf = version.get(self, key, ctx).await?;
818 0 : Ok(buf)
819 0 : }
820 :
821 292 : pub(crate) async fn list_dbdirs(
822 292 : &self,
823 292 : lsn: Lsn,
824 292 : ctx: &RequestContext,
825 292 : ) -> Result<HashMap<(Oid, Oid), bool>, PageReconstructError> {
826 : // fetch directory entry
827 292 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
828 :
829 292 : Ok(DbDirectory::des(&buf)?.dbdirs)
830 292 : }
831 :
832 0 : pub(crate) async fn get_twophase_file(
833 0 : &self,
834 0 : xid: u64,
835 0 : lsn: Lsn,
836 0 : ctx: &RequestContext,
837 0 : ) -> Result<Bytes, PageReconstructError> {
838 0 : let key = twophase_file_key(xid);
839 0 : let buf = self.get(key, lsn, ctx).await?;
840 0 : Ok(buf)
841 0 : }
842 :
843 294 : pub(crate) async fn list_twophase_files(
844 294 : &self,
845 294 : lsn: Lsn,
846 294 : ctx: &RequestContext,
847 294 : ) -> Result<HashSet<u64>, PageReconstructError> {
848 : // fetch directory entry
849 294 : let buf = self.get(TWOPHASEDIR_KEY, lsn, ctx).await?;
850 :
851 294 : if self.pg_version >= 17 {
852 0 : Ok(TwoPhaseDirectoryV17::des(&buf)?.xids)
853 : } else {
854 294 : Ok(TwoPhaseDirectory::des(&buf)?
855 : .xids
856 294 : .iter()
857 294 : .map(|x| u64::from(*x))
858 294 : .collect())
859 : }
860 294 : }
861 :
862 0 : pub(crate) async fn get_control_file(
863 0 : &self,
864 0 : lsn: Lsn,
865 0 : ctx: &RequestContext,
866 0 : ) -> Result<Bytes, PageReconstructError> {
867 0 : self.get(CONTROLFILE_KEY, lsn, ctx).await
868 0 : }
869 :
870 12 : pub(crate) async fn get_checkpoint(
871 12 : &self,
872 12 : lsn: Lsn,
873 12 : ctx: &RequestContext,
874 12 : ) -> Result<Bytes, PageReconstructError> {
875 12 : self.get(CHECKPOINT_KEY, lsn, ctx).await
876 12 : }
877 :
878 12 : async fn list_aux_files_v2(
879 12 : &self,
880 12 : lsn: Lsn,
881 12 : ctx: &RequestContext,
882 12 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
883 12 : let kv = self
884 12 : .scan(KeySpace::single(Key::metadata_aux_key_range()), lsn, ctx)
885 12 : .await?;
886 12 : let mut result = HashMap::new();
887 12 : let mut sz = 0;
888 30 : for (_, v) in kv {
889 18 : let v = v?;
890 18 : let v = aux_file::decode_file_value_bytes(&v)
891 18 : .context("value decode")
892 18 : .map_err(PageReconstructError::Other)?;
893 34 : for (fname, content) in v {
894 16 : sz += fname.len();
895 16 : sz += content.len();
896 16 : result.insert(fname, content);
897 16 : }
898 : }
899 12 : self.aux_file_size_estimator.on_initial(sz);
900 12 : Ok(result)
901 12 : }
902 :
903 0 : pub(crate) async fn trigger_aux_file_size_computation(
904 0 : &self,
905 0 : lsn: Lsn,
906 0 : ctx: &RequestContext,
907 0 : ) -> Result<(), PageReconstructError> {
908 0 : self.list_aux_files_v2(lsn, ctx).await?;
909 0 : Ok(())
910 0 : }
911 :
912 12 : pub(crate) async fn list_aux_files(
913 12 : &self,
914 12 : lsn: Lsn,
915 12 : ctx: &RequestContext,
916 12 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
917 12 : self.list_aux_files_v2(lsn, ctx).await
918 12 : }
919 :
920 0 : pub(crate) async fn get_replorigins(
921 0 : &self,
922 0 : lsn: Lsn,
923 0 : ctx: &RequestContext,
924 0 : ) -> Result<HashMap<RepOriginId, Lsn>, PageReconstructError> {
925 0 : let kv = self
926 0 : .scan(KeySpace::single(repl_origin_key_range()), lsn, ctx)
927 0 : .await?;
928 0 : let mut result = HashMap::new();
929 0 : for (k, v) in kv {
930 0 : let v = v?;
931 0 : let origin_id = k.field6 as RepOriginId;
932 0 : let origin_lsn = Lsn::des(&v).unwrap();
933 0 : if origin_lsn != Lsn::INVALID {
934 0 : result.insert(origin_id, origin_lsn);
935 0 : }
936 : }
937 0 : Ok(result)
938 0 : }
939 :
940 : /// Does the same as get_current_logical_size but counted on demand.
941 : /// Used to initialize the logical size tracking on startup.
942 : ///
943 : /// Only relation blocks are counted currently. That excludes metadata,
944 : /// SLRUs, twophase files etc.
945 : ///
946 : /// # Cancel-Safety
947 : ///
948 : /// This method is cancellation-safe.
949 0 : pub(crate) async fn get_current_logical_size_non_incremental(
950 0 : &self,
951 0 : lsn: Lsn,
952 0 : ctx: &RequestContext,
953 0 : ) -> Result<u64, CalculateLogicalSizeError> {
954 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
955 :
956 : // Fetch list of database dirs and iterate them
957 0 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
958 0 : let dbdir = DbDirectory::des(&buf)?;
959 :
960 0 : let mut total_size: u64 = 0;
961 0 : for (spcnode, dbnode) in dbdir.dbdirs.keys() {
962 0 : for rel in self
963 0 : .list_rels(*spcnode, *dbnode, Version::Lsn(lsn), ctx)
964 0 : .await?
965 : {
966 0 : if self.cancel.is_cancelled() {
967 0 : return Err(CalculateLogicalSizeError::Cancelled);
968 0 : }
969 0 : let relsize_key = rel_size_to_key(rel);
970 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
971 0 : let relsize = buf.get_u32_le();
972 0 :
973 0 : total_size += relsize as u64;
974 : }
975 : }
976 0 : Ok(total_size * BLCKSZ as u64)
977 0 : }
978 :
979 : /// Get a KeySpace that covers all the Keys that are in use at AND below the given LSN. This is only used
980 : /// for gc-compaction.
981 : ///
982 : /// gc-compaction cannot use the same `collect_keyspace` function as the legacy compaction because it
983 : /// processes data at multiple LSNs and needs to be aware of the fact that some key ranges might need to
984 : /// be kept only for a specific range of LSN.
985 : ///
986 : /// Consider the case that the user created branches at LSN 10 and 20, where the user created a table A at
987 : /// LSN 10 and dropped that table at LSN 20. `collect_keyspace` at LSN 10 will return the key range
988 : /// corresponding to that table, while LSN 20 won't. The keyspace info at a single LSN is not enough to
989 : /// determine which keys to retain/drop for gc-compaction.
990 : ///
991 : /// For now, it only drops AUX-v1 keys. But in the future, the function will be extended to return the keyspace
992 : /// to be retained for each of the branch LSN.
993 : ///
994 : /// The return value is (dense keyspace, sparse keyspace).
995 40 : pub(crate) async fn collect_gc_compaction_keyspace(
996 40 : &self,
997 40 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
998 40 : let metadata_key_begin = Key::metadata_key_range().start;
999 40 : let aux_v1_key = AUX_FILES_KEY;
1000 40 : let dense_keyspace = KeySpace {
1001 40 : ranges: vec![Key::MIN..aux_v1_key, aux_v1_key.next()..metadata_key_begin],
1002 40 : };
1003 40 : Ok((
1004 40 : dense_keyspace,
1005 40 : SparseKeySpace(KeySpace::single(Key::metadata_key_range())),
1006 40 : ))
1007 40 : }
1008 :
1009 : ///
1010 : /// Get a KeySpace that covers all the Keys that are in use at the given LSN.
1011 : /// Anything that's not listed maybe removed from the underlying storage (from
1012 : /// that LSN forwards).
1013 : ///
1014 : /// The return value is (dense keyspace, sparse keyspace).
1015 292 : pub(crate) async fn collect_keyspace(
1016 292 : &self,
1017 292 : lsn: Lsn,
1018 292 : ctx: &RequestContext,
1019 292 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
1020 292 : // Iterate through key ranges, greedily packing them into partitions
1021 292 : let mut result = KeySpaceAccum::new();
1022 292 :
1023 292 : // The dbdir metadata always exists
1024 292 : result.add_key(DBDIR_KEY);
1025 :
1026 : // Fetch list of database dirs and iterate them
1027 292 : let dbdir = self.list_dbdirs(lsn, ctx).await?;
1028 292 : let mut dbs: Vec<((Oid, Oid), bool)> = dbdir.into_iter().collect();
1029 292 :
1030 292 : dbs.sort_unstable_by(|(k_a, _), (k_b, _)| k_a.cmp(k_b));
1031 292 : for ((spcnode, dbnode), has_relmap_file) in dbs {
1032 0 : if has_relmap_file {
1033 0 : result.add_key(relmap_file_key(spcnode, dbnode));
1034 0 : }
1035 0 : result.add_key(rel_dir_to_key(spcnode, dbnode));
1036 :
1037 0 : let mut rels: Vec<RelTag> = self
1038 0 : .list_rels(spcnode, dbnode, Version::Lsn(lsn), ctx)
1039 0 : .await?
1040 0 : .into_iter()
1041 0 : .collect();
1042 0 : rels.sort_unstable();
1043 0 : for rel in rels {
1044 0 : let relsize_key = rel_size_to_key(rel);
1045 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
1046 0 : let relsize = buf.get_u32_le();
1047 0 :
1048 0 : result.add_range(rel_block_to_key(rel, 0)..rel_block_to_key(rel, relsize));
1049 0 : result.add_key(relsize_key);
1050 : }
1051 : }
1052 :
1053 : // Iterate SLRUs next
1054 292 : if self.tenant_shard_id.is_shard_zero() {
1055 858 : for kind in [
1056 286 : SlruKind::Clog,
1057 286 : SlruKind::MultiXactMembers,
1058 286 : SlruKind::MultiXactOffsets,
1059 : ] {
1060 858 : let slrudir_key = slru_dir_to_key(kind);
1061 858 : result.add_key(slrudir_key);
1062 858 : let buf = self.get(slrudir_key, lsn, ctx).await?;
1063 858 : let dir = SlruSegmentDirectory::des(&buf)?;
1064 858 : let mut segments: Vec<u32> = dir.segments.iter().cloned().collect();
1065 858 : segments.sort_unstable();
1066 858 : for segno in segments {
1067 0 : let segsize_key = slru_segment_size_to_key(kind, segno);
1068 0 : let mut buf = self.get(segsize_key, lsn, ctx).await?;
1069 0 : let segsize = buf.get_u32_le();
1070 0 :
1071 0 : result.add_range(
1072 0 : slru_block_to_key(kind, segno, 0)..slru_block_to_key(kind, segno, segsize),
1073 0 : );
1074 0 : result.add_key(segsize_key);
1075 : }
1076 : }
1077 6 : }
1078 :
1079 : // Then pg_twophase
1080 292 : result.add_key(TWOPHASEDIR_KEY);
1081 :
1082 292 : let mut xids: Vec<u64> = self
1083 292 : .list_twophase_files(lsn, ctx)
1084 292 : .await?
1085 292 : .iter()
1086 292 : .cloned()
1087 292 : .collect();
1088 292 : xids.sort_unstable();
1089 292 : for xid in xids {
1090 0 : result.add_key(twophase_file_key(xid));
1091 0 : }
1092 :
1093 292 : result.add_key(CONTROLFILE_KEY);
1094 292 : result.add_key(CHECKPOINT_KEY);
1095 292 :
1096 292 : // Add extra keyspaces in the test cases. Some test cases write keys into the storage without
1097 292 : // creating directory keys. These test cases will add such keyspaces into `extra_test_dense_keyspace`
1098 292 : // and the keys will not be garbage-colllected.
1099 292 : #[cfg(test)]
1100 292 : {
1101 292 : let guard = self.extra_test_dense_keyspace.load();
1102 292 : for kr in &guard.ranges {
1103 0 : result.add_range(kr.clone());
1104 0 : }
1105 0 : }
1106 0 :
1107 292 : let dense_keyspace = result.to_keyspace();
1108 292 : let sparse_keyspace = SparseKeySpace(KeySpace {
1109 292 : ranges: vec![Key::metadata_aux_key_range(), repl_origin_key_range()],
1110 292 : });
1111 292 :
1112 292 : if cfg!(debug_assertions) {
1113 : // Verify if the sparse keyspaces are ordered and non-overlapping.
1114 :
1115 : // We do not use KeySpaceAccum for sparse_keyspace because we want to ensure each
1116 : // category of sparse keys are split into their own image/delta files. If there
1117 : // are overlapping keyspaces, they will be automatically merged by keyspace accum,
1118 : // and we want the developer to keep the keyspaces separated.
1119 :
1120 292 : let ranges = &sparse_keyspace.0.ranges;
1121 :
1122 : // TODO: use a single overlaps_with across the codebase
1123 292 : fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
1124 292 : !(a.end <= b.start || b.end <= a.start)
1125 292 : }
1126 584 : for i in 0..ranges.len() {
1127 584 : for j in 0..i {
1128 292 : if overlaps_with(&ranges[i], &ranges[j]) {
1129 0 : panic!(
1130 0 : "overlapping sparse keyspace: {}..{} and {}..{}",
1131 0 : ranges[i].start, ranges[i].end, ranges[j].start, ranges[j].end
1132 0 : );
1133 292 : }
1134 : }
1135 : }
1136 292 : for i in 1..ranges.len() {
1137 292 : assert!(
1138 292 : ranges[i - 1].end <= ranges[i].start,
1139 0 : "unordered sparse keyspace: {}..{} and {}..{}",
1140 0 : ranges[i - 1].start,
1141 0 : ranges[i - 1].end,
1142 0 : ranges[i].start,
1143 0 : ranges[i].end
1144 : );
1145 : }
1146 0 : }
1147 :
1148 292 : Ok((dense_keyspace, sparse_keyspace))
1149 292 : }
1150 :
1151 : /// Get cached size of relation if it not updated after specified LSN
1152 448540 : pub fn get_cached_rel_size(&self, tag: &RelTag, lsn: Lsn) -> Option<BlockNumber> {
1153 448540 : let rel_size_cache = self.rel_size_cache.read().unwrap();
1154 448540 : if let Some((cached_lsn, nblocks)) = rel_size_cache.map.get(tag) {
1155 448518 : if lsn >= *cached_lsn {
1156 443372 : RELSIZE_CACHE_HITS.inc();
1157 443372 : return Some(*nblocks);
1158 5146 : }
1159 5146 : RELSIZE_CACHE_MISSES_OLD.inc();
1160 22 : }
1161 5168 : RELSIZE_CACHE_MISSES.inc();
1162 5168 : None
1163 448540 : }
1164 :
1165 : /// Update cached relation size if there is no more recent update
1166 5136 : pub fn update_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
1167 5136 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1168 5136 :
1169 5136 : if lsn < rel_size_cache.complete_as_of {
1170 : // Do not cache old values. It's safe to cache the size on read, as long as
1171 : // the read was at an LSN since we started the WAL ingestion. Reasoning: we
1172 : // never evict values from the cache, so if the relation size changed after
1173 : // 'lsn', the new value is already in the cache.
1174 0 : return;
1175 5136 : }
1176 5136 :
1177 5136 : match rel_size_cache.map.entry(tag) {
1178 5136 : hash_map::Entry::Occupied(mut entry) => {
1179 5136 : let cached_lsn = entry.get_mut();
1180 5136 : if lsn >= cached_lsn.0 {
1181 0 : *cached_lsn = (lsn, nblocks);
1182 5136 : }
1183 : }
1184 0 : hash_map::Entry::Vacant(entry) => {
1185 0 : entry.insert((lsn, nblocks));
1186 0 : RELSIZE_CACHE_ENTRIES.inc();
1187 0 : }
1188 : }
1189 5136 : }
1190 :
1191 : /// Store cached relation size
1192 282720 : pub fn set_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
1193 282720 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1194 282720 : if rel_size_cache.map.insert(tag, (lsn, nblocks)).is_none() {
1195 1920 : RELSIZE_CACHE_ENTRIES.inc();
1196 280800 : }
1197 282720 : }
1198 :
1199 : /// Remove cached relation size
1200 2 : pub fn remove_cached_rel_size(&self, tag: &RelTag) {
1201 2 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1202 2 : if rel_size_cache.map.remove(tag).is_some() {
1203 2 : RELSIZE_CACHE_ENTRIES.dec();
1204 2 : }
1205 2 : }
1206 : }
1207 :
1208 : /// DatadirModification represents an operation to ingest an atomic set of
1209 : /// updates to the repository.
1210 : ///
1211 : /// It is created by the 'begin_record' function. It is called for each WAL
1212 : /// record, so that all the modifications by a one WAL record appear atomic.
1213 : pub struct DatadirModification<'a> {
1214 : /// The timeline this modification applies to. You can access this to
1215 : /// read the state, but note that any pending updates are *not* reflected
1216 : /// in the state in 'tline' yet.
1217 : pub tline: &'a Timeline,
1218 :
1219 : /// Current LSN of the modification
1220 : lsn: Lsn,
1221 :
1222 : // The modifications are not applied directly to the underlying key-value store.
1223 : // The put-functions add the modifications here, and they are flushed to the
1224 : // underlying key-value store by the 'finish' function.
1225 : pending_lsns: Vec<Lsn>,
1226 : pending_deletions: Vec<(Range<Key>, Lsn)>,
1227 : pending_nblocks: i64,
1228 :
1229 : /// Metadata writes, indexed by key so that they can be read from not-yet-committed modifications
1230 : /// while ingesting subsequent records. See [`Self::is_data_key`] for the definition of 'metadata'.
1231 : pending_metadata_pages: HashMap<CompactKey, Vec<(Lsn, usize, Value)>>,
1232 :
1233 : /// Data writes, ready to be flushed into an ephemeral layer. See [`Self::is_data_key`] for
1234 : /// which keys are stored here.
1235 : pending_data_batch: Option<SerializedValueBatch>,
1236 :
1237 : /// For special "directory" keys that store key-value maps, track the size of the map
1238 : /// if it was updated in this modification.
1239 : pending_directory_entries: Vec<(DirectoryKind, usize)>,
1240 :
1241 : /// An **approximation** of how many metadata bytes will be written to the EphemeralFile.
1242 : pending_metadata_bytes: usize,
1243 : }
1244 :
1245 : impl<'a> DatadirModification<'a> {
1246 : // When a DatadirModification is committed, we do a monolithic serialization of all its contents. WAL records can
1247 : // contain multiple pages, so the pageserver's record-based batch size isn't sufficient to bound this allocation: we
1248 : // additionally specify a limit on how much payload a DatadirModification may contain before it should be committed.
1249 : pub(crate) const MAX_PENDING_BYTES: usize = 8 * 1024 * 1024;
1250 :
1251 : /// Get the current lsn
1252 418058 : pub(crate) fn get_lsn(&self) -> Lsn {
1253 418058 : self.lsn
1254 418058 : }
1255 :
1256 0 : pub(crate) fn approx_pending_bytes(&self) -> usize {
1257 0 : self.pending_data_batch
1258 0 : .as_ref()
1259 0 : .map_or(0, |b| b.buffer_size())
1260 0 : + self.pending_metadata_bytes
1261 0 : }
1262 :
1263 0 : pub(crate) fn has_dirty_data(&self) -> bool {
1264 0 : self.pending_data_batch
1265 0 : .as_ref()
1266 0 : .map_or(false, |b| b.has_data())
1267 0 : }
1268 :
1269 : /// Set the current lsn
1270 145858 : pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> anyhow::Result<()> {
1271 145858 : ensure!(
1272 145858 : lsn >= self.lsn,
1273 0 : "setting an older lsn {} than {} is not allowed",
1274 : lsn,
1275 : self.lsn
1276 : );
1277 :
1278 145858 : if lsn > self.lsn {
1279 145858 : self.pending_lsns.push(self.lsn);
1280 145858 : self.lsn = lsn;
1281 145858 : }
1282 145858 : Ok(())
1283 145858 : }
1284 :
1285 : /// In this context, 'metadata' means keys that are only read by the pageserver internally, and 'data' means
1286 : /// keys that represent literal blocks that postgres can read. So data includes relation blocks and
1287 : /// SLRU blocks, which are read directly by postgres, and everything else is considered metadata.
1288 : ///
1289 : /// The distinction is important because data keys are handled on a fast path where dirty writes are
1290 : /// not readable until this modification is committed, whereas metadata keys are visible for read
1291 : /// via [`Self::get`] as soon as their record has been ingested.
1292 850432 : fn is_data_key(key: &Key) -> bool {
1293 850432 : key.is_rel_block_key() || key.is_slru_block_key()
1294 850432 : }
1295 :
1296 : /// Initialize a completely new repository.
1297 : ///
1298 : /// This inserts the directory metadata entries that are assumed to
1299 : /// always exist.
1300 178 : pub fn init_empty(&mut self) -> anyhow::Result<()> {
1301 178 : let buf = DbDirectory::ser(&DbDirectory {
1302 178 : dbdirs: HashMap::new(),
1303 178 : })?;
1304 178 : self.pending_directory_entries.push((DirectoryKind::Db, 0));
1305 178 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1306 :
1307 178 : let buf = if self.tline.pg_version >= 17 {
1308 0 : TwoPhaseDirectoryV17::ser(&TwoPhaseDirectoryV17 {
1309 0 : xids: HashSet::new(),
1310 0 : })
1311 : } else {
1312 178 : TwoPhaseDirectory::ser(&TwoPhaseDirectory {
1313 178 : xids: HashSet::new(),
1314 178 : })
1315 0 : }?;
1316 178 : self.pending_directory_entries
1317 178 : .push((DirectoryKind::TwoPhase, 0));
1318 178 : self.put(TWOPHASEDIR_KEY, Value::Image(buf.into()));
1319 :
1320 178 : let buf: Bytes = SlruSegmentDirectory::ser(&SlruSegmentDirectory::default())?.into();
1321 178 : let empty_dir = Value::Image(buf);
1322 178 : self.put(slru_dir_to_key(SlruKind::Clog), empty_dir.clone());
1323 178 : self.pending_directory_entries
1324 178 : .push((DirectoryKind::SlruSegment(SlruKind::Clog), 0));
1325 178 : self.put(
1326 178 : slru_dir_to_key(SlruKind::MultiXactMembers),
1327 178 : empty_dir.clone(),
1328 178 : );
1329 178 : self.pending_directory_entries
1330 178 : .push((DirectoryKind::SlruSegment(SlruKind::Clog), 0));
1331 178 : self.put(slru_dir_to_key(SlruKind::MultiXactOffsets), empty_dir);
1332 178 : self.pending_directory_entries
1333 178 : .push((DirectoryKind::SlruSegment(SlruKind::MultiXactOffsets), 0));
1334 178 :
1335 178 : Ok(())
1336 178 : }
1337 :
1338 : #[cfg(test)]
1339 176 : pub fn init_empty_test_timeline(&mut self) -> anyhow::Result<()> {
1340 176 : self.init_empty()?;
1341 176 : self.put_control_file(bytes::Bytes::from_static(
1342 176 : b"control_file contents do not matter",
1343 176 : ))
1344 176 : .context("put_control_file")?;
1345 176 : self.put_checkpoint(bytes::Bytes::from_static(
1346 176 : b"checkpoint_file contents do not matter",
1347 176 : ))
1348 176 : .context("put_checkpoint_file")?;
1349 176 : Ok(())
1350 176 : }
1351 :
1352 : /// Creates a relation if it is not already present.
1353 : /// Returns the current size of the relation
1354 418056 : pub(crate) async fn create_relation_if_required(
1355 418056 : &mut self,
1356 418056 : rel: RelTag,
1357 418056 : ctx: &RequestContext,
1358 418056 : ) -> Result<u32, PageReconstructError> {
1359 : // Get current size and put rel creation if rel doesn't exist
1360 : //
1361 : // NOTE: we check the cache first even though get_rel_exists and get_rel_size would
1362 : // check the cache too. This is because eagerly checking the cache results in
1363 : // less work overall and 10% better performance. It's more work on cache miss
1364 : // but cache miss is rare.
1365 418056 : if let Some(nblocks) = self.tline.get_cached_rel_size(&rel, self.get_lsn()) {
1366 418046 : Ok(nblocks)
1367 10 : } else if !self
1368 10 : .tline
1369 10 : .get_rel_exists(rel, Version::Modified(self), ctx)
1370 10 : .await?
1371 : {
1372 : // create it with 0 size initially, the logic below will extend it
1373 10 : self.put_rel_creation(rel, 0, ctx)
1374 10 : .await
1375 10 : .context("Relation Error")?;
1376 10 : Ok(0)
1377 : } else {
1378 0 : self.tline
1379 0 : .get_rel_size(rel, Version::Modified(self), ctx)
1380 0 : .await
1381 : }
1382 418056 : }
1383 :
1384 : /// Given a block number for a relation (which represents a newly written block),
1385 : /// the previous block count of the relation, and the shard info, find the gaps
1386 : /// that were created by the newly written block if any.
1387 145670 : fn find_gaps(
1388 145670 : rel: RelTag,
1389 145670 : blkno: u32,
1390 145670 : previous_nblocks: u32,
1391 145670 : shard: &ShardIdentity,
1392 145670 : ) -> Option<KeySpace> {
1393 145670 : let mut key = rel_block_to_key(rel, blkno);
1394 145670 : let mut gap_accum = None;
1395 :
1396 145670 : for gap_blkno in previous_nblocks..blkno {
1397 32 : key.field6 = gap_blkno;
1398 32 :
1399 32 : if shard.get_shard_number(&key) != shard.number {
1400 8 : continue;
1401 24 : }
1402 24 :
1403 24 : gap_accum
1404 24 : .get_or_insert_with(KeySpaceAccum::new)
1405 24 : .add_key(key);
1406 : }
1407 :
1408 145670 : gap_accum.map(|accum| accum.to_keyspace())
1409 145670 : }
1410 :
1411 145852 : pub async fn ingest_batch(
1412 145852 : &mut self,
1413 145852 : mut batch: SerializedValueBatch,
1414 145852 : // TODO(vlad): remove this argument and replace the shard check with is_key_local
1415 145852 : shard: &ShardIdentity,
1416 145852 : ctx: &RequestContext,
1417 145852 : ) -> anyhow::Result<()> {
1418 145852 : let mut gaps_at_lsns = Vec::default();
1419 :
1420 145852 : for meta in batch.metadata.iter() {
1421 145642 : let (rel, blkno) = Key::from_compact(meta.key()).to_rel_block()?;
1422 145642 : let new_nblocks = blkno + 1;
1423 :
1424 145642 : let old_nblocks = self.create_relation_if_required(rel, ctx).await?;
1425 145642 : if new_nblocks > old_nblocks {
1426 2390 : self.put_rel_extend(rel, new_nblocks, ctx).await?;
1427 143252 : }
1428 :
1429 145642 : if let Some(gaps) = Self::find_gaps(rel, blkno, old_nblocks, shard) {
1430 0 : gaps_at_lsns.push((gaps, meta.lsn()));
1431 145642 : }
1432 : }
1433 :
1434 145852 : if !gaps_at_lsns.is_empty() {
1435 0 : batch.zero_gaps(gaps_at_lsns);
1436 145852 : }
1437 :
1438 145852 : match self.pending_data_batch.as_mut() {
1439 20 : Some(pending_batch) => {
1440 20 : pending_batch.extend(batch);
1441 20 : }
1442 145832 : None if batch.has_data() => {
1443 145630 : self.pending_data_batch = Some(batch);
1444 145630 : }
1445 202 : None => {
1446 202 : // Nothing to initialize the batch with
1447 202 : }
1448 : }
1449 :
1450 145852 : Ok(())
1451 145852 : }
1452 :
1453 : /// Put a new page version that can be constructed from a WAL record
1454 : ///
1455 : /// NOTE: this will *not* implicitly extend the relation, if the page is beyond the
1456 : /// current end-of-file. It's up to the caller to check that the relation size
1457 : /// matches the blocks inserted!
1458 12 : pub fn put_rel_wal_record(
1459 12 : &mut self,
1460 12 : rel: RelTag,
1461 12 : blknum: BlockNumber,
1462 12 : rec: NeonWalRecord,
1463 12 : ) -> anyhow::Result<()> {
1464 12 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1465 12 : self.put(rel_block_to_key(rel, blknum), Value::WalRecord(rec));
1466 12 : Ok(())
1467 12 : }
1468 :
1469 : // Same, but for an SLRU.
1470 8 : pub fn put_slru_wal_record(
1471 8 : &mut self,
1472 8 : kind: SlruKind,
1473 8 : segno: u32,
1474 8 : blknum: BlockNumber,
1475 8 : rec: NeonWalRecord,
1476 8 : ) -> anyhow::Result<()> {
1477 8 : if !self.tline.tenant_shard_id.is_shard_zero() {
1478 0 : return Ok(());
1479 8 : }
1480 8 :
1481 8 : self.put(
1482 8 : slru_block_to_key(kind, segno, blknum),
1483 8 : Value::WalRecord(rec),
1484 8 : );
1485 8 : Ok(())
1486 8 : }
1487 :
1488 : /// Like put_wal_record, but with ready-made image of the page.
1489 277842 : pub fn put_rel_page_image(
1490 277842 : &mut self,
1491 277842 : rel: RelTag,
1492 277842 : blknum: BlockNumber,
1493 277842 : img: Bytes,
1494 277842 : ) -> anyhow::Result<()> {
1495 277842 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1496 277842 : let key = rel_block_to_key(rel, blknum);
1497 277842 : if !key.is_valid_key_on_write_path() {
1498 0 : anyhow::bail!(
1499 0 : "the request contains data not supported by pageserver at {}",
1500 0 : key
1501 0 : );
1502 277842 : }
1503 277842 : self.put(rel_block_to_key(rel, blknum), Value::Image(img));
1504 277842 : Ok(())
1505 277842 : }
1506 :
1507 6 : pub fn put_slru_page_image(
1508 6 : &mut self,
1509 6 : kind: SlruKind,
1510 6 : segno: u32,
1511 6 : blknum: BlockNumber,
1512 6 : img: Bytes,
1513 6 : ) -> anyhow::Result<()> {
1514 6 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1515 :
1516 6 : let key = slru_block_to_key(kind, segno, blknum);
1517 6 : if !key.is_valid_key_on_write_path() {
1518 0 : anyhow::bail!(
1519 0 : "the request contains data not supported by pageserver at {}",
1520 0 : key
1521 0 : );
1522 6 : }
1523 6 : self.put(key, Value::Image(img));
1524 6 : Ok(())
1525 6 : }
1526 :
1527 2998 : pub(crate) fn put_rel_page_image_zero(
1528 2998 : &mut self,
1529 2998 : rel: RelTag,
1530 2998 : blknum: BlockNumber,
1531 2998 : ) -> anyhow::Result<()> {
1532 2998 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1533 2998 : let key = rel_block_to_key(rel, blknum);
1534 2998 : if !key.is_valid_key_on_write_path() {
1535 0 : anyhow::bail!(
1536 0 : "the request contains data not supported by pageserver: {} @ {}",
1537 0 : key,
1538 0 : self.lsn
1539 0 : );
1540 2998 : }
1541 2998 :
1542 2998 : let batch = self
1543 2998 : .pending_data_batch
1544 2998 : .get_or_insert_with(SerializedValueBatch::default);
1545 2998 :
1546 2998 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1547 2998 :
1548 2998 : Ok(())
1549 2998 : }
1550 :
1551 0 : pub(crate) fn put_slru_page_image_zero(
1552 0 : &mut self,
1553 0 : kind: SlruKind,
1554 0 : segno: u32,
1555 0 : blknum: BlockNumber,
1556 0 : ) -> anyhow::Result<()> {
1557 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1558 0 : let key = slru_block_to_key(kind, segno, blknum);
1559 0 : if !key.is_valid_key_on_write_path() {
1560 0 : anyhow::bail!(
1561 0 : "the request contains data not supported by pageserver: {} @ {}",
1562 0 : key,
1563 0 : self.lsn
1564 0 : );
1565 0 : }
1566 0 :
1567 0 : let batch = self
1568 0 : .pending_data_batch
1569 0 : .get_or_insert_with(SerializedValueBatch::default);
1570 0 :
1571 0 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1572 0 :
1573 0 : Ok(())
1574 0 : }
1575 :
1576 : /// Store a relmapper file (pg_filenode.map) in the repository
1577 16 : pub async fn put_relmap_file(
1578 16 : &mut self,
1579 16 : spcnode: Oid,
1580 16 : dbnode: Oid,
1581 16 : img: Bytes,
1582 16 : ctx: &RequestContext,
1583 16 : ) -> anyhow::Result<()> {
1584 : // Add it to the directory (if it doesn't exist already)
1585 16 : let buf = self.get(DBDIR_KEY, ctx).await?;
1586 16 : let mut dbdir = DbDirectory::des(&buf)?;
1587 :
1588 16 : let r = dbdir.dbdirs.insert((spcnode, dbnode), true);
1589 16 : if r.is_none() || r == Some(false) {
1590 : // The dbdir entry didn't exist, or it contained a
1591 : // 'false'. The 'insert' call already updated it with
1592 : // 'true', now write the updated 'dbdirs' map back.
1593 16 : let buf = DbDirectory::ser(&dbdir)?;
1594 16 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1595 0 : }
1596 16 : if r.is_none() {
1597 : // Create RelDirectory
1598 8 : let buf = RelDirectory::ser(&RelDirectory {
1599 8 : rels: HashSet::new(),
1600 8 : })?;
1601 8 : self.pending_directory_entries.push((DirectoryKind::Rel, 0));
1602 8 : self.put(
1603 8 : rel_dir_to_key(spcnode, dbnode),
1604 8 : Value::Image(Bytes::from(buf)),
1605 8 : );
1606 8 : }
1607 :
1608 16 : self.put(relmap_file_key(spcnode, dbnode), Value::Image(img));
1609 16 : Ok(())
1610 16 : }
1611 :
1612 0 : pub async fn put_twophase_file(
1613 0 : &mut self,
1614 0 : xid: u64,
1615 0 : img: Bytes,
1616 0 : ctx: &RequestContext,
1617 0 : ) -> anyhow::Result<()> {
1618 : // Add it to the directory entry
1619 0 : let dirbuf = self.get(TWOPHASEDIR_KEY, ctx).await?;
1620 0 : let newdirbuf = if self.tline.pg_version >= 17 {
1621 0 : let mut dir = TwoPhaseDirectoryV17::des(&dirbuf)?;
1622 0 : if !dir.xids.insert(xid) {
1623 0 : anyhow::bail!("twophase file for xid {} already exists", xid);
1624 0 : }
1625 0 : self.pending_directory_entries
1626 0 : .push((DirectoryKind::TwoPhase, dir.xids.len()));
1627 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
1628 : } else {
1629 0 : let xid = xid as u32;
1630 0 : let mut dir = TwoPhaseDirectory::des(&dirbuf)?;
1631 0 : if !dir.xids.insert(xid) {
1632 0 : anyhow::bail!("twophase file for xid {} already exists", xid);
1633 0 : }
1634 0 : self.pending_directory_entries
1635 0 : .push((DirectoryKind::TwoPhase, dir.xids.len()));
1636 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
1637 : };
1638 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
1639 0 :
1640 0 : self.put(twophase_file_key(xid), Value::Image(img));
1641 0 : Ok(())
1642 0 : }
1643 :
1644 0 : pub async fn set_replorigin(
1645 0 : &mut self,
1646 0 : origin_id: RepOriginId,
1647 0 : origin_lsn: Lsn,
1648 0 : ) -> anyhow::Result<()> {
1649 0 : let key = repl_origin_key(origin_id);
1650 0 : self.put(key, Value::Image(origin_lsn.ser().unwrap().into()));
1651 0 : Ok(())
1652 0 : }
1653 :
1654 0 : pub async fn drop_replorigin(&mut self, origin_id: RepOriginId) -> anyhow::Result<()> {
1655 0 : self.set_replorigin(origin_id, Lsn::INVALID).await
1656 0 : }
1657 :
1658 178 : pub fn put_control_file(&mut self, img: Bytes) -> anyhow::Result<()> {
1659 178 : self.put(CONTROLFILE_KEY, Value::Image(img));
1660 178 : Ok(())
1661 178 : }
1662 :
1663 192 : pub fn put_checkpoint(&mut self, img: Bytes) -> anyhow::Result<()> {
1664 192 : self.put(CHECKPOINT_KEY, Value::Image(img));
1665 192 : Ok(())
1666 192 : }
1667 :
1668 0 : pub async fn drop_dbdir(
1669 0 : &mut self,
1670 0 : spcnode: Oid,
1671 0 : dbnode: Oid,
1672 0 : ctx: &RequestContext,
1673 0 : ) -> anyhow::Result<()> {
1674 0 : let total_blocks = self
1675 0 : .tline
1676 0 : .get_db_size(spcnode, dbnode, Version::Modified(self), ctx)
1677 0 : .await?;
1678 :
1679 : // Remove entry from dbdir
1680 0 : let buf = self.get(DBDIR_KEY, ctx).await?;
1681 0 : let mut dir = DbDirectory::des(&buf)?;
1682 0 : if dir.dbdirs.remove(&(spcnode, dbnode)).is_some() {
1683 0 : let buf = DbDirectory::ser(&dir)?;
1684 0 : self.pending_directory_entries
1685 0 : .push((DirectoryKind::Db, dir.dbdirs.len()));
1686 0 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1687 : } else {
1688 0 : warn!(
1689 0 : "dropped dbdir for spcnode {} dbnode {} did not exist in db directory",
1690 : spcnode, dbnode
1691 : );
1692 : }
1693 :
1694 : // Update logical database size.
1695 0 : self.pending_nblocks -= total_blocks as i64;
1696 0 :
1697 0 : // Delete all relations and metadata files for the spcnode/dnode
1698 0 : self.delete(dbdir_key_range(spcnode, dbnode));
1699 0 : Ok(())
1700 0 : }
1701 :
1702 : /// Create a relation fork.
1703 : ///
1704 : /// 'nblocks' is the initial size.
1705 1920 : pub async fn put_rel_creation(
1706 1920 : &mut self,
1707 1920 : rel: RelTag,
1708 1920 : nblocks: BlockNumber,
1709 1920 : ctx: &RequestContext,
1710 1920 : ) -> Result<(), RelationError> {
1711 1920 : if rel.relnode == 0 {
1712 0 : return Err(RelationError::InvalidRelnode);
1713 1920 : }
1714 : // It's possible that this is the first rel for this db in this
1715 : // tablespace. Create the reldir entry for it if so.
1716 1920 : let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await.context("read db")?)
1717 1920 : .context("deserialize db")?;
1718 1920 : let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
1719 1920 : let mut rel_dir =
1720 1920 : if let hash_map::Entry::Vacant(e) = dbdir.dbdirs.entry((rel.spcnode, rel.dbnode)) {
1721 : // Didn't exist. Update dbdir
1722 8 : e.insert(false);
1723 8 : let buf = DbDirectory::ser(&dbdir).context("serialize db")?;
1724 8 : self.pending_directory_entries
1725 8 : .push((DirectoryKind::Db, dbdir.dbdirs.len()));
1726 8 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1727 8 :
1728 8 : // and create the RelDirectory
1729 8 : RelDirectory::default()
1730 : } else {
1731 : // reldir already exists, fetch it
1732 1912 : RelDirectory::des(&self.get(rel_dir_key, ctx).await.context("read db")?)
1733 1912 : .context("deserialize db")?
1734 : };
1735 :
1736 : // Add the new relation to the rel directory entry, and write it back
1737 1920 : if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
1738 0 : return Err(RelationError::AlreadyExists);
1739 1920 : }
1740 1920 :
1741 1920 : self.pending_directory_entries
1742 1920 : .push((DirectoryKind::Rel, rel_dir.rels.len()));
1743 1920 :
1744 1920 : self.put(
1745 1920 : rel_dir_key,
1746 1920 : Value::Image(Bytes::from(
1747 1920 : RelDirectory::ser(&rel_dir).context("serialize")?,
1748 : )),
1749 : );
1750 :
1751 : // Put size
1752 1920 : let size_key = rel_size_to_key(rel);
1753 1920 : let buf = nblocks.to_le_bytes();
1754 1920 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1755 1920 :
1756 1920 : self.pending_nblocks += nblocks as i64;
1757 1920 :
1758 1920 : // Update relation size cache
1759 1920 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
1760 1920 :
1761 1920 : // Even if nblocks > 0, we don't insert any actual blocks here. That's up to the
1762 1920 : // caller.
1763 1920 : Ok(())
1764 1920 : }
1765 :
1766 : /// Truncate relation
1767 6012 : pub async fn put_rel_truncation(
1768 6012 : &mut self,
1769 6012 : rel: RelTag,
1770 6012 : nblocks: BlockNumber,
1771 6012 : ctx: &RequestContext,
1772 6012 : ) -> anyhow::Result<()> {
1773 6012 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1774 6012 : if self
1775 6012 : .tline
1776 6012 : .get_rel_exists(rel, Version::Modified(self), ctx)
1777 6012 : .await?
1778 : {
1779 6012 : let size_key = rel_size_to_key(rel);
1780 : // Fetch the old size first
1781 6012 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
1782 6012 :
1783 6012 : // Update the entry with the new size.
1784 6012 : let buf = nblocks.to_le_bytes();
1785 6012 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1786 6012 :
1787 6012 : // Update relation size cache
1788 6012 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
1789 6012 :
1790 6012 : // Update logical database size.
1791 6012 : self.pending_nblocks -= old_size as i64 - nblocks as i64;
1792 0 : }
1793 6012 : Ok(())
1794 6012 : }
1795 :
1796 : /// Extend relation
1797 : /// If new size is smaller, do nothing.
1798 276680 : pub async fn put_rel_extend(
1799 276680 : &mut self,
1800 276680 : rel: RelTag,
1801 276680 : nblocks: BlockNumber,
1802 276680 : ctx: &RequestContext,
1803 276680 : ) -> anyhow::Result<()> {
1804 276680 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1805 :
1806 : // Put size
1807 276680 : let size_key = rel_size_to_key(rel);
1808 276680 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
1809 276680 :
1810 276680 : // only extend relation here. never decrease the size
1811 276680 : if nblocks > old_size {
1812 274788 : let buf = nblocks.to_le_bytes();
1813 274788 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1814 274788 :
1815 274788 : // Update relation size cache
1816 274788 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
1817 274788 :
1818 274788 : self.pending_nblocks += nblocks as i64 - old_size as i64;
1819 274788 : }
1820 276680 : Ok(())
1821 276680 : }
1822 :
1823 : /// Drop some relations
1824 10 : pub(crate) async fn put_rel_drops(
1825 10 : &mut self,
1826 10 : drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
1827 10 : ctx: &RequestContext,
1828 10 : ) -> anyhow::Result<()> {
1829 12 : for ((spc_node, db_node), rel_tags) in drop_relations {
1830 2 : let dir_key = rel_dir_to_key(spc_node, db_node);
1831 2 : let buf = self.get(dir_key, ctx).await?;
1832 2 : let mut dir = RelDirectory::des(&buf)?;
1833 :
1834 2 : let mut dirty = false;
1835 4 : for rel_tag in rel_tags {
1836 2 : if dir.rels.remove(&(rel_tag.relnode, rel_tag.forknum)) {
1837 2 : dirty = true;
1838 2 :
1839 2 : // update logical size
1840 2 : let size_key = rel_size_to_key(rel_tag);
1841 2 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
1842 2 : self.pending_nblocks -= old_size as i64;
1843 2 :
1844 2 : // Remove entry from relation size cache
1845 2 : self.tline.remove_cached_rel_size(&rel_tag);
1846 2 :
1847 2 : // Delete size entry, as well as all blocks
1848 2 : self.delete(rel_key_range(rel_tag));
1849 0 : }
1850 : }
1851 :
1852 2 : if dirty {
1853 2 : self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
1854 2 : self.pending_directory_entries
1855 2 : .push((DirectoryKind::Rel, dir.rels.len()));
1856 0 : }
1857 : }
1858 :
1859 10 : Ok(())
1860 10 : }
1861 :
1862 6 : pub async fn put_slru_segment_creation(
1863 6 : &mut self,
1864 6 : kind: SlruKind,
1865 6 : segno: u32,
1866 6 : nblocks: BlockNumber,
1867 6 : ctx: &RequestContext,
1868 6 : ) -> anyhow::Result<()> {
1869 6 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1870 :
1871 : // Add it to the directory entry
1872 6 : let dir_key = slru_dir_to_key(kind);
1873 6 : let buf = self.get(dir_key, ctx).await?;
1874 6 : let mut dir = SlruSegmentDirectory::des(&buf)?;
1875 :
1876 6 : if !dir.segments.insert(segno) {
1877 0 : anyhow::bail!("slru segment {kind:?}/{segno} already exists");
1878 6 : }
1879 6 : self.pending_directory_entries
1880 6 : .push((DirectoryKind::SlruSegment(kind), dir.segments.len()));
1881 6 : self.put(
1882 6 : dir_key,
1883 6 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
1884 : );
1885 :
1886 : // Put size
1887 6 : let size_key = slru_segment_size_to_key(kind, segno);
1888 6 : let buf = nblocks.to_le_bytes();
1889 6 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1890 6 :
1891 6 : // even if nblocks > 0, we don't insert any actual blocks here
1892 6 :
1893 6 : Ok(())
1894 6 : }
1895 :
1896 : /// Extend SLRU segment
1897 0 : pub fn put_slru_extend(
1898 0 : &mut self,
1899 0 : kind: SlruKind,
1900 0 : segno: u32,
1901 0 : nblocks: BlockNumber,
1902 0 : ) -> anyhow::Result<()> {
1903 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1904 :
1905 : // Put size
1906 0 : let size_key = slru_segment_size_to_key(kind, segno);
1907 0 : let buf = nblocks.to_le_bytes();
1908 0 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1909 0 : Ok(())
1910 0 : }
1911 :
1912 : /// This method is used for marking truncated SLRU files
1913 0 : pub async fn drop_slru_segment(
1914 0 : &mut self,
1915 0 : kind: SlruKind,
1916 0 : segno: u32,
1917 0 : ctx: &RequestContext,
1918 0 : ) -> anyhow::Result<()> {
1919 0 : // Remove it from the directory entry
1920 0 : let dir_key = slru_dir_to_key(kind);
1921 0 : let buf = self.get(dir_key, ctx).await?;
1922 0 : let mut dir = SlruSegmentDirectory::des(&buf)?;
1923 :
1924 0 : if !dir.segments.remove(&segno) {
1925 0 : warn!("slru segment {:?}/{} does not exist", kind, segno);
1926 0 : }
1927 0 : self.pending_directory_entries
1928 0 : .push((DirectoryKind::SlruSegment(kind), dir.segments.len()));
1929 0 : self.put(
1930 0 : dir_key,
1931 0 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
1932 : );
1933 :
1934 : // Delete size entry, as well as all blocks
1935 0 : self.delete(slru_segment_key_range(kind, segno));
1936 0 :
1937 0 : Ok(())
1938 0 : }
1939 :
1940 : /// Drop a relmapper file (pg_filenode.map)
1941 0 : pub fn drop_relmap_file(&mut self, _spcnode: Oid, _dbnode: Oid) -> anyhow::Result<()> {
1942 0 : // TODO
1943 0 : Ok(())
1944 0 : }
1945 :
1946 : /// This method is used for marking truncated SLRU files
1947 0 : pub async fn drop_twophase_file(
1948 0 : &mut self,
1949 0 : xid: u64,
1950 0 : ctx: &RequestContext,
1951 0 : ) -> anyhow::Result<()> {
1952 : // Remove it from the directory entry
1953 0 : let buf = self.get(TWOPHASEDIR_KEY, ctx).await?;
1954 0 : let newdirbuf = if self.tline.pg_version >= 17 {
1955 0 : let mut dir = TwoPhaseDirectoryV17::des(&buf)?;
1956 :
1957 0 : if !dir.xids.remove(&xid) {
1958 0 : warn!("twophase file for xid {} does not exist", xid);
1959 0 : }
1960 0 : self.pending_directory_entries
1961 0 : .push((DirectoryKind::TwoPhase, dir.xids.len()));
1962 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
1963 : } else {
1964 0 : let xid: u32 = u32::try_from(xid)?;
1965 0 : let mut dir = TwoPhaseDirectory::des(&buf)?;
1966 :
1967 0 : if !dir.xids.remove(&xid) {
1968 0 : warn!("twophase file for xid {} does not exist", xid);
1969 0 : }
1970 0 : self.pending_directory_entries
1971 0 : .push((DirectoryKind::TwoPhase, dir.xids.len()));
1972 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
1973 : };
1974 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
1975 0 :
1976 0 : // Delete it
1977 0 : self.delete(twophase_key_range(xid));
1978 0 :
1979 0 : Ok(())
1980 0 : }
1981 :
1982 16 : pub async fn put_file(
1983 16 : &mut self,
1984 16 : path: &str,
1985 16 : content: &[u8],
1986 16 : ctx: &RequestContext,
1987 16 : ) -> anyhow::Result<()> {
1988 16 : let key = aux_file::encode_aux_file_key(path);
1989 : // retrieve the key from the engine
1990 16 : let old_val = match self.get(key, ctx).await {
1991 4 : Ok(val) => Some(val),
1992 12 : Err(PageReconstructError::MissingKey(_)) => None,
1993 0 : Err(e) => return Err(e.into()),
1994 : };
1995 16 : let files: Vec<(&str, &[u8])> = if let Some(ref old_val) = old_val {
1996 4 : aux_file::decode_file_value(old_val)?
1997 : } else {
1998 12 : Vec::new()
1999 : };
2000 16 : let mut other_files = Vec::with_capacity(files.len());
2001 16 : let mut modifying_file = None;
2002 20 : for file @ (p, content) in files {
2003 4 : if path == p {
2004 4 : assert!(
2005 4 : modifying_file.is_none(),
2006 0 : "duplicated entries found for {}",
2007 : path
2008 : );
2009 4 : modifying_file = Some(content);
2010 0 : } else {
2011 0 : other_files.push(file);
2012 0 : }
2013 : }
2014 16 : let mut new_files = other_files;
2015 16 : match (modifying_file, content.is_empty()) {
2016 2 : (Some(old_content), false) => {
2017 2 : self.tline
2018 2 : .aux_file_size_estimator
2019 2 : .on_update(old_content.len(), content.len());
2020 2 : new_files.push((path, content));
2021 2 : }
2022 2 : (Some(old_content), true) => {
2023 2 : self.tline
2024 2 : .aux_file_size_estimator
2025 2 : .on_remove(old_content.len());
2026 2 : // not adding the file key to the final `new_files` vec.
2027 2 : }
2028 12 : (None, false) => {
2029 12 : self.tline.aux_file_size_estimator.on_add(content.len());
2030 12 : new_files.push((path, content));
2031 12 : }
2032 0 : (None, true) => warn!("removing non-existing aux file: {}", path),
2033 : }
2034 16 : let new_val = aux_file::encode_file_value(&new_files)?;
2035 16 : self.put(key, Value::Image(new_val.into()));
2036 16 :
2037 16 : Ok(())
2038 16 : }
2039 :
2040 : ///
2041 : /// Flush changes accumulated so far to the underlying repository.
2042 : ///
2043 : /// Usually, changes made in DatadirModification are atomic, but this allows
2044 : /// you to flush them to the underlying repository before the final `commit`.
2045 : /// That allows to free up the memory used to hold the pending changes.
2046 : ///
2047 : /// Currently only used during bulk import of a data directory. In that
2048 : /// context, breaking the atomicity is OK. If the import is interrupted, the
2049 : /// whole import fails and the timeline will be deleted anyway.
2050 : /// (Or to be precise, it will be left behind for debugging purposes and
2051 : /// ignored, see <https://github.com/neondatabase/neon/pull/1809>)
2052 : ///
2053 : /// Note: A consequence of flushing the pending operations is that they
2054 : /// won't be visible to subsequent operations until `commit`. The function
2055 : /// retains all the metadata, but data pages are flushed. That's again OK
2056 : /// for bulk import, where you are just loading data pages and won't try to
2057 : /// modify the same pages twice.
2058 1930 : pub(crate) async fn flush(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2059 1930 : // Unless we have accumulated a decent amount of changes, it's not worth it
2060 1930 : // to scan through the pending_updates list.
2061 1930 : let pending_nblocks = self.pending_nblocks;
2062 1930 : if pending_nblocks < 10000 {
2063 1930 : return Ok(());
2064 0 : }
2065 :
2066 0 : let mut writer = self.tline.writer().await;
2067 :
2068 : // Flush relation and SLRU data blocks, keep metadata.
2069 0 : if let Some(batch) = self.pending_data_batch.take() {
2070 0 : tracing::debug!(
2071 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2072 0 : batch.max_lsn,
2073 0 : self.tline.get_last_record_lsn()
2074 : );
2075 :
2076 : // This bails out on first error without modifying pending_updates.
2077 : // That's Ok, cf this function's doc comment.
2078 0 : writer.put_batch(batch, ctx).await?;
2079 0 : }
2080 :
2081 0 : if pending_nblocks != 0 {
2082 0 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2083 0 : self.pending_nblocks = 0;
2084 0 : }
2085 :
2086 0 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2087 0 : writer.update_directory_entries_count(kind, count as u64);
2088 0 : }
2089 :
2090 0 : Ok(())
2091 1930 : }
2092 :
2093 : ///
2094 : /// Finish this atomic update, writing all the updated keys to the
2095 : /// underlying timeline.
2096 : /// All the modifications in this atomic update are stamped by the specified LSN.
2097 : ///
2098 743064 : pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2099 743064 : let mut writer = self.tline.writer().await;
2100 :
2101 743064 : let pending_nblocks = self.pending_nblocks;
2102 743064 : self.pending_nblocks = 0;
2103 :
2104 : // Ordering: the items in this batch do not need to be in any global order, but values for
2105 : // a particular Key must be in Lsn order relative to one another. InMemoryLayer relies on
2106 : // this to do efficient updates to its index. See [`wal_decoder::serialized_batch`] for
2107 : // more details.
2108 :
2109 743064 : let metadata_batch = {
2110 743064 : let pending_meta = self
2111 743064 : .pending_metadata_pages
2112 743064 : .drain()
2113 743064 : .flat_map(|(key, values)| {
2114 273826 : values
2115 273826 : .into_iter()
2116 273826 : .map(move |(lsn, value_size, value)| (key, lsn, value_size, value))
2117 743064 : })
2118 743064 : .collect::<Vec<_>>();
2119 743064 :
2120 743064 : if pending_meta.is_empty() {
2121 472278 : None
2122 : } else {
2123 270786 : Some(SerializedValueBatch::from_values(pending_meta))
2124 : }
2125 : };
2126 :
2127 743064 : let data_batch = self.pending_data_batch.take();
2128 :
2129 743064 : let maybe_batch = match (data_batch, metadata_batch) {
2130 264556 : (Some(mut data), Some(metadata)) => {
2131 264556 : data.extend(metadata);
2132 264556 : Some(data)
2133 : }
2134 143262 : (Some(data), None) => Some(data),
2135 6230 : (None, Some(metadata)) => Some(metadata),
2136 329016 : (None, None) => None,
2137 : };
2138 :
2139 743064 : if let Some(batch) = maybe_batch {
2140 414048 : tracing::debug!(
2141 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2142 0 : batch.max_lsn,
2143 0 : self.tline.get_last_record_lsn()
2144 : );
2145 :
2146 : // This bails out on first error without modifying pending_updates.
2147 : // That's Ok, cf this function's doc comment.
2148 414048 : writer.put_batch(batch, ctx).await?;
2149 329016 : }
2150 :
2151 743064 : if !self.pending_deletions.is_empty() {
2152 2 : writer.delete_batch(&self.pending_deletions, ctx).await?;
2153 2 : self.pending_deletions.clear();
2154 743062 : }
2155 :
2156 743064 : self.pending_lsns.push(self.lsn);
2157 888922 : for pending_lsn in self.pending_lsns.drain(..) {
2158 888922 : // TODO(vlad): pretty sure the comment below is not valid anymore
2159 888922 : // and we can call finish write with the latest LSN
2160 888922 : //
2161 888922 : // Ideally, we should be able to call writer.finish_write() only once
2162 888922 : // with the highest LSN. However, the last_record_lsn variable in the
2163 888922 : // timeline keeps track of the latest LSN and the immediate previous LSN
2164 888922 : // so we need to record every LSN to not leave a gap between them.
2165 888922 : writer.finish_write(pending_lsn);
2166 888922 : }
2167 :
2168 743064 : if pending_nblocks != 0 {
2169 270570 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2170 472494 : }
2171 :
2172 743064 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2173 2834 : writer.update_directory_entries_count(kind, count as u64);
2174 2834 : }
2175 :
2176 743064 : self.pending_metadata_bytes = 0;
2177 743064 :
2178 743064 : Ok(())
2179 743064 : }
2180 :
2181 291704 : pub(crate) fn len(&self) -> usize {
2182 291704 : self.pending_metadata_pages.len()
2183 291704 : + self.pending_data_batch.as_ref().map_or(0, |b| b.len())
2184 291704 : + self.pending_deletions.len()
2185 291704 : }
2186 :
2187 : /// Read a page from the Timeline we are writing to. For metadata pages, this passes through
2188 : /// a cache in Self, which makes writes earlier in this modification visible to WAL records later
2189 : /// in the modification.
2190 : ///
2191 : /// For data pages, reads pass directly to the owning Timeline: any ingest code which reads a data
2192 : /// page must ensure that the pages they read are already committed in Timeline, for example
2193 : /// DB create operations are always preceded by a call to commit(). This is special cased because
2194 : /// it's rare: all the 'normal' WAL operations will only read metadata pages such as relation sizes,
2195 : /// and not data pages.
2196 286586 : async fn get(&self, key: Key, ctx: &RequestContext) -> Result<Bytes, PageReconstructError> {
2197 286586 : if !Self::is_data_key(&key) {
2198 : // Have we already updated the same key? Read the latest pending updated
2199 : // version in that case.
2200 : //
2201 : // Note: we don't check pending_deletions. It is an error to request a
2202 : // value that has been removed, deletion only avoids leaking storage.
2203 286586 : if let Some(values) = self.pending_metadata_pages.get(&key.to_compact()) {
2204 15928 : if let Some((_, _, value)) = values.last() {
2205 15928 : return if let Value::Image(img) = value {
2206 15928 : Ok(img.clone())
2207 : } else {
2208 : // Currently, we never need to read back a WAL record that we
2209 : // inserted in the same "transaction". All the metadata updates
2210 : // work directly with Images, and we never need to read actual
2211 : // data pages. We could handle this if we had to, by calling
2212 : // the walredo manager, but let's keep it simple for now.
2213 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
2214 0 : "unexpected pending WAL record"
2215 0 : )))
2216 : };
2217 0 : }
2218 270658 : }
2219 : } else {
2220 : // This is an expensive check, so we only do it in debug mode. If reading a data key,
2221 : // this key should never be present in pending_data_pages. We ensure this by committing
2222 : // modifications before ingesting DB create operations, which are the only kind that reads
2223 : // data pages during ingest.
2224 0 : if cfg!(debug_assertions) {
2225 0 : assert!(!self
2226 0 : .pending_data_batch
2227 0 : .as_ref()
2228 0 : .map_or(false, |b| b.updates_key(&key)));
2229 0 : }
2230 : }
2231 :
2232 : // Metadata page cache miss, or we're reading a data page.
2233 270658 : let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
2234 270658 : self.tline.get(key, lsn, ctx).await
2235 286586 : }
2236 :
2237 563846 : fn put(&mut self, key: Key, val: Value) {
2238 563846 : if Self::is_data_key(&key) {
2239 277868 : self.put_data(key.to_compact(), val)
2240 : } else {
2241 285978 : self.put_metadata(key.to_compact(), val)
2242 : }
2243 563846 : }
2244 :
2245 277868 : fn put_data(&mut self, key: CompactKey, val: Value) {
2246 277868 : let batch = self
2247 277868 : .pending_data_batch
2248 277868 : .get_or_insert_with(SerializedValueBatch::default);
2249 277868 : batch.put(key, val, self.lsn);
2250 277868 : }
2251 :
2252 285978 : fn put_metadata(&mut self, key: CompactKey, val: Value) {
2253 285978 : let values = self.pending_metadata_pages.entry(key).or_default();
2254 : // Replace the previous value if it exists at the same lsn
2255 285978 : if let Some((last_lsn, last_value_ser_size, last_value)) = values.last_mut() {
2256 12152 : if *last_lsn == self.lsn {
2257 : // Update the pending_metadata_bytes contribution from this entry, and update the serialized size in place
2258 12152 : self.pending_metadata_bytes -= *last_value_ser_size;
2259 12152 : *last_value_ser_size = val.serialized_size().unwrap() as usize;
2260 12152 : self.pending_metadata_bytes += *last_value_ser_size;
2261 12152 :
2262 12152 : // Use the latest value, this replaces any earlier write to the same (key,lsn), such as much
2263 12152 : // have been generated by synthesized zero page writes prior to the first real write to a page.
2264 12152 : *last_value = val;
2265 12152 : return;
2266 0 : }
2267 273826 : }
2268 :
2269 273826 : let val_serialized_size = val.serialized_size().unwrap() as usize;
2270 273826 : self.pending_metadata_bytes += val_serialized_size;
2271 273826 : values.push((self.lsn, val_serialized_size, val));
2272 273826 :
2273 273826 : if key == CHECKPOINT_KEY.to_compact() {
2274 192 : tracing::debug!("Checkpoint key added to pending with size {val_serialized_size}");
2275 273634 : }
2276 285978 : }
2277 :
2278 2 : fn delete(&mut self, key_range: Range<Key>) {
2279 2 : trace!("DELETE {}-{}", key_range.start, key_range.end);
2280 2 : self.pending_deletions.push((key_range, self.lsn));
2281 2 : }
2282 : }
2283 :
2284 : /// This struct facilitates accessing either a committed key from the timeline at a
2285 : /// specific LSN, or the latest uncommitted key from a pending modification.
2286 : ///
2287 : /// During WAL ingestion, the records from multiple LSNs may be batched in the same
2288 : /// modification before being flushed to the timeline. Hence, the routines in WalIngest
2289 : /// need to look up the keys in the modification first before looking them up in the
2290 : /// timeline to not miss the latest updates.
2291 : #[derive(Clone, Copy)]
2292 : pub enum Version<'a> {
2293 : Lsn(Lsn),
2294 : Modified(&'a DatadirModification<'a>),
2295 : }
2296 :
2297 : impl<'a> Version<'a> {
2298 5176 : async fn get(
2299 5176 : &self,
2300 5176 : timeline: &Timeline,
2301 5176 : key: Key,
2302 5176 : ctx: &RequestContext,
2303 5176 : ) -> Result<Bytes, PageReconstructError> {
2304 5176 : match self {
2305 5156 : Version::Lsn(lsn) => timeline.get(key, *lsn, ctx).await,
2306 20 : Version::Modified(modification) => modification.get(key, ctx).await,
2307 : }
2308 5176 : }
2309 :
2310 35620 : fn get_lsn(&self) -> Lsn {
2311 35620 : match self {
2312 29574 : Version::Lsn(lsn) => *lsn,
2313 6046 : Version::Modified(modification) => modification.lsn,
2314 : }
2315 35620 : }
2316 : }
2317 :
2318 : //--- Metadata structs stored in key-value pairs in the repository.
2319 :
2320 2246 : #[derive(Debug, Serialize, Deserialize)]
2321 : pub(crate) struct DbDirectory {
2322 : // (spcnode, dbnode) -> (do relmapper and PG_VERSION files exist)
2323 : pub(crate) dbdirs: HashMap<(Oid, Oid), bool>,
2324 : }
2325 :
2326 : // The format of TwoPhaseDirectory changed in PostgreSQL v17, because the filenames of
2327 : // pg_twophase files was expanded from 32-bit XIDs to 64-bit XIDs. Previously, the files
2328 : // were named like "pg_twophase/000002E5", now they're like
2329 : // "pg_twophsae/0000000A000002E4".
2330 :
2331 294 : #[derive(Debug, Serialize, Deserialize)]
2332 : pub(crate) struct TwoPhaseDirectory {
2333 : pub(crate) xids: HashSet<TransactionId>,
2334 : }
2335 :
2336 0 : #[derive(Debug, Serialize, Deserialize)]
2337 : struct TwoPhaseDirectoryV17 {
2338 : xids: HashSet<u64>,
2339 : }
2340 :
2341 1932 : #[derive(Debug, Serialize, Deserialize, Default)]
2342 : pub(crate) struct RelDirectory {
2343 : // Set of relations that exist. (relfilenode, forknum)
2344 : //
2345 : // TODO: Store it as a btree or radix tree or something else that spans multiple
2346 : // key-value pairs, if you have a lot of relations
2347 : pub(crate) rels: HashSet<(Oid, u8)>,
2348 : }
2349 :
2350 0 : #[derive(Debug, Serialize, Deserialize)]
2351 : struct RelSizeEntry {
2352 : nblocks: u32,
2353 : }
2354 :
2355 864 : #[derive(Debug, Serialize, Deserialize, Default)]
2356 : pub(crate) struct SlruSegmentDirectory {
2357 : // Set of SLRU segments that exist.
2358 : pub(crate) segments: HashSet<u32>,
2359 : }
2360 :
2361 : #[derive(Copy, Clone, PartialEq, Eq, Debug, enum_map::Enum)]
2362 : #[repr(u8)]
2363 : pub(crate) enum DirectoryKind {
2364 : Db,
2365 : TwoPhase,
2366 : Rel,
2367 : AuxFiles,
2368 : SlruSegment(SlruKind),
2369 : }
2370 :
2371 : impl DirectoryKind {
2372 : pub(crate) const KINDS_NUM: usize = <DirectoryKind as Enum>::LENGTH;
2373 5668 : pub(crate) fn offset(&self) -> usize {
2374 5668 : self.into_usize()
2375 5668 : }
2376 : }
2377 :
2378 : static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
2379 :
2380 : #[allow(clippy::bool_assert_comparison)]
2381 : #[cfg(test)]
2382 : mod tests {
2383 : use hex_literal::hex;
2384 : use pageserver_api::{models::ShardParameters, shard::ShardStripeSize};
2385 : use utils::{
2386 : id::TimelineId,
2387 : shard::{ShardCount, ShardNumber},
2388 : };
2389 :
2390 : use super::*;
2391 :
2392 : use crate::{tenant::harness::TenantHarness, DEFAULT_PG_VERSION};
2393 :
2394 : /// Test a round trip of aux file updates, from DatadirModification to reading back from the Timeline
2395 : #[tokio::test]
2396 2 : async fn aux_files_round_trip() -> anyhow::Result<()> {
2397 2 : let name = "aux_files_round_trip";
2398 2 : let harness = TenantHarness::create(name).await?;
2399 2 :
2400 2 : pub const TIMELINE_ID: TimelineId =
2401 2 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
2402 2 :
2403 2 : let (tenant, ctx) = harness.load().await;
2404 2 : let tline = tenant
2405 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2406 2 : .await?;
2407 2 : let tline = tline.raw_timeline().unwrap();
2408 2 :
2409 2 : // First modification: insert two keys
2410 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
2411 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
2412 2 : modification.set_lsn(Lsn(0x1008))?;
2413 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
2414 2 : modification.commit(&ctx).await?;
2415 2 : let expect_1008 = HashMap::from([
2416 2 : ("foo/bar1".to_string(), Bytes::from_static(b"content1")),
2417 2 : ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
2418 2 : ]);
2419 2 :
2420 2 : let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
2421 2 : assert_eq!(readback, expect_1008);
2422 2 :
2423 2 : // Second modification: update one key, remove the other
2424 2 : let mut modification = tline.begin_modification(Lsn(0x2000));
2425 2 : modification.put_file("foo/bar1", b"content3", &ctx).await?;
2426 2 : modification.set_lsn(Lsn(0x2008))?;
2427 2 : modification.put_file("foo/bar2", b"", &ctx).await?;
2428 2 : modification.commit(&ctx).await?;
2429 2 : let expect_2008 =
2430 2 : HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
2431 2 :
2432 2 : let readback = tline.list_aux_files(Lsn(0x2008), &ctx).await?;
2433 2 : assert_eq!(readback, expect_2008);
2434 2 :
2435 2 : // Reading back in time works
2436 2 : let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
2437 2 : assert_eq!(readback, expect_1008);
2438 2 :
2439 2 : Ok(())
2440 2 : }
2441 :
2442 : #[test]
2443 2 : fn gap_finding() {
2444 2 : let rel = RelTag {
2445 2 : spcnode: 1663,
2446 2 : dbnode: 208101,
2447 2 : relnode: 2620,
2448 2 : forknum: 0,
2449 2 : };
2450 2 : let base_blkno = 1;
2451 2 :
2452 2 : let base_key = rel_block_to_key(rel, base_blkno);
2453 2 : let before_base_key = rel_block_to_key(rel, base_blkno - 1);
2454 2 :
2455 2 : let shard = ShardIdentity::unsharded();
2456 2 :
2457 2 : let mut previous_nblocks = 0;
2458 22 : for i in 0..10 {
2459 20 : let crnt_blkno = base_blkno + i;
2460 20 : let gaps = DatadirModification::find_gaps(rel, crnt_blkno, previous_nblocks, &shard);
2461 20 :
2462 20 : previous_nblocks = crnt_blkno + 1;
2463 20 :
2464 20 : if i == 0 {
2465 : // The first block we write is 1, so we should find the gap.
2466 2 : assert_eq!(gaps.unwrap(), KeySpace::single(before_base_key..base_key));
2467 : } else {
2468 18 : assert!(gaps.is_none());
2469 : }
2470 : }
2471 :
2472 : // This is an update to an already existing block. No gaps here.
2473 2 : let update_blkno = 5;
2474 2 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
2475 2 : assert!(gaps.is_none());
2476 :
2477 : // This is an update past the current end block.
2478 2 : let after_gap_blkno = 20;
2479 2 : let gaps = DatadirModification::find_gaps(rel, after_gap_blkno, previous_nblocks, &shard);
2480 2 :
2481 2 : let gap_start_key = rel_block_to_key(rel, previous_nblocks);
2482 2 : let after_gap_key = rel_block_to_key(rel, after_gap_blkno);
2483 2 : assert_eq!(
2484 2 : gaps.unwrap(),
2485 2 : KeySpace::single(gap_start_key..after_gap_key)
2486 2 : );
2487 2 : }
2488 :
2489 : #[test]
2490 2 : fn sharded_gap_finding() {
2491 2 : let rel = RelTag {
2492 2 : spcnode: 1663,
2493 2 : dbnode: 208101,
2494 2 : relnode: 2620,
2495 2 : forknum: 0,
2496 2 : };
2497 2 :
2498 2 : let first_blkno = 6;
2499 2 :
2500 2 : // This shard will get the even blocks
2501 2 : let shard = ShardIdentity::from_params(
2502 2 : ShardNumber(0),
2503 2 : &ShardParameters {
2504 2 : count: ShardCount(2),
2505 2 : stripe_size: ShardStripeSize(1),
2506 2 : },
2507 2 : );
2508 2 :
2509 2 : // Only keys belonging to this shard are considered as gaps.
2510 2 : let mut previous_nblocks = 0;
2511 2 : let gaps =
2512 2 : DatadirModification::find_gaps(rel, first_blkno, previous_nblocks, &shard).unwrap();
2513 2 : assert!(!gaps.ranges.is_empty());
2514 6 : for gap_range in gaps.ranges {
2515 4 : let mut k = gap_range.start;
2516 8 : while k != gap_range.end {
2517 4 : assert_eq!(shard.get_shard_number(&k), shard.number);
2518 4 : k = k.next();
2519 : }
2520 : }
2521 :
2522 2 : previous_nblocks = first_blkno;
2523 2 :
2524 2 : let update_blkno = 2;
2525 2 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
2526 2 : assert!(gaps.is_none());
2527 2 : }
2528 :
2529 : /*
2530 : fn assert_current_logical_size<R: Repository>(timeline: &DatadirTimeline<R>, lsn: Lsn) {
2531 : let incremental = timeline.get_current_logical_size();
2532 : let non_incremental = timeline
2533 : .get_current_logical_size_non_incremental(lsn)
2534 : .unwrap();
2535 : assert_eq!(incremental, non_incremental);
2536 : }
2537 : */
2538 :
2539 : /*
2540 : ///
2541 : /// Test list_rels() function, with branches and dropped relations
2542 : ///
2543 : #[test]
2544 : fn test_list_rels_drop() -> Result<()> {
2545 : let repo = RepoHarness::create("test_list_rels_drop")?.load();
2546 : let tline = create_empty_timeline(repo, TIMELINE_ID)?;
2547 : const TESTDB: u32 = 111;
2548 :
2549 : // Import initial dummy checkpoint record, otherwise the get_timeline() call
2550 : // after branching fails below
2551 : let mut writer = tline.begin_record(Lsn(0x10));
2552 : writer.put_checkpoint(ZERO_CHECKPOINT.clone())?;
2553 : writer.finish()?;
2554 :
2555 : // Create a relation on the timeline
2556 : let mut writer = tline.begin_record(Lsn(0x20));
2557 : writer.put_rel_page_image(TESTREL_A, 0, TEST_IMG("foo blk 0 at 2"))?;
2558 : writer.finish()?;
2559 :
2560 : let writer = tline.begin_record(Lsn(0x00));
2561 : writer.finish()?;
2562 :
2563 : // Check that list_rels() lists it after LSN 2, but no before it
2564 : assert!(!tline.list_rels(0, TESTDB, Lsn(0x10))?.contains(&TESTREL_A));
2565 : assert!(tline.list_rels(0, TESTDB, Lsn(0x20))?.contains(&TESTREL_A));
2566 : assert!(tline.list_rels(0, TESTDB, Lsn(0x30))?.contains(&TESTREL_A));
2567 :
2568 : // Create a branch, check that the relation is visible there
2569 : repo.branch_timeline(&tline, NEW_TIMELINE_ID, Lsn(0x30))?;
2570 : let newtline = match repo.get_timeline(NEW_TIMELINE_ID)?.local_timeline() {
2571 : Some(timeline) => timeline,
2572 : None => panic!("Should have a local timeline"),
2573 : };
2574 : let newtline = DatadirTimelineImpl::new(newtline);
2575 : assert!(newtline
2576 : .list_rels(0, TESTDB, Lsn(0x30))?
2577 : .contains(&TESTREL_A));
2578 :
2579 : // Drop it on the branch
2580 : let mut new_writer = newtline.begin_record(Lsn(0x40));
2581 : new_writer.drop_relation(TESTREL_A)?;
2582 : new_writer.finish()?;
2583 :
2584 : // Check that it's no longer listed on the branch after the point where it was dropped
2585 : assert!(newtline
2586 : .list_rels(0, TESTDB, Lsn(0x30))?
2587 : .contains(&TESTREL_A));
2588 : assert!(!newtline
2589 : .list_rels(0, TESTDB, Lsn(0x40))?
2590 : .contains(&TESTREL_A));
2591 :
2592 : // Run checkpoint and garbage collection and check that it's still not visible
2593 : newtline.checkpoint(CheckpointConfig::Forced)?;
2594 : repo.gc_iteration(Some(NEW_TIMELINE_ID), 0, true)?;
2595 :
2596 : assert!(!newtline
2597 : .list_rels(0, TESTDB, Lsn(0x40))?
2598 : .contains(&TESTREL_A));
2599 :
2600 : Ok(())
2601 : }
2602 : */
2603 :
2604 : /*
2605 : #[test]
2606 : fn test_read_beyond_eof() -> Result<()> {
2607 : let repo = RepoHarness::create("test_read_beyond_eof")?.load();
2608 : let tline = create_test_timeline(repo, TIMELINE_ID)?;
2609 :
2610 : make_some_layers(&tline, Lsn(0x20))?;
2611 : let mut writer = tline.begin_record(Lsn(0x60));
2612 : walingest.put_rel_page_image(
2613 : &mut writer,
2614 : TESTREL_A,
2615 : 0,
2616 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x60))),
2617 : )?;
2618 : writer.finish()?;
2619 :
2620 : // Test read before rel creation. Should error out.
2621 : assert!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x10), false).is_err());
2622 :
2623 : // Read block beyond end of relation at different points in time.
2624 : // These reads should fall into different delta, image, and in-memory layers.
2625 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x20), false)?, ZERO_PAGE);
2626 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x25), false)?, ZERO_PAGE);
2627 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x30), false)?, ZERO_PAGE);
2628 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x35), false)?, ZERO_PAGE);
2629 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x40), false)?, ZERO_PAGE);
2630 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x45), false)?, ZERO_PAGE);
2631 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x50), false)?, ZERO_PAGE);
2632 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x55), false)?, ZERO_PAGE);
2633 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x60), false)?, ZERO_PAGE);
2634 :
2635 : // Test on an in-memory layer with no preceding layer
2636 : let mut writer = tline.begin_record(Lsn(0x70));
2637 : walingest.put_rel_page_image(
2638 : &mut writer,
2639 : TESTREL_B,
2640 : 0,
2641 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x70))),
2642 : )?;
2643 : writer.finish()?;
2644 :
2645 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_B, 1, Lsn(0x70), false)?6, ZERO_PAGE);
2646 :
2647 : Ok(())
2648 : }
2649 : */
2650 : }
|
