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 std::collections::{BTreeMap, HashMap, HashSet, hash_map};
10 : use std::ops::{ControlFlow, Range};
11 :
12 : use anyhow::{Context, ensure};
13 : use bytes::{Buf, Bytes, BytesMut};
14 : use enum_map::Enum;
15 : use itertools::Itertools;
16 : use pageserver_api::key::{
17 : AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, CompactKey, DBDIR_KEY, Key, RelDirExists,
18 : TWOPHASEDIR_KEY, dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range,
19 : rel_size_to_key, rel_tag_sparse_key, rel_tag_sparse_key_range, relmap_file_key,
20 : repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
21 : slru_segment_key_range, slru_segment_size_to_key, twophase_file_key, twophase_key_range,
22 : };
23 : use pageserver_api::keyspace::SparseKeySpace;
24 : use pageserver_api::record::NeonWalRecord;
25 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
26 : use pageserver_api::shard::ShardIdentity;
27 : use pageserver_api::value::Value;
28 : use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
29 : use postgres_ffi::{BLCKSZ, Oid, RepOriginId, TimestampTz, TransactionId};
30 : use serde::{Deserialize, Serialize};
31 : use strum::IntoEnumIterator;
32 : use tokio_util::sync::CancellationToken;
33 : use tracing::{debug, info, trace, warn};
34 : use utils::bin_ser::{BeSer, DeserializeError};
35 : use utils::lsn::Lsn;
36 : use utils::pausable_failpoint;
37 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
38 :
39 : use super::tenant::{PageReconstructError, Timeline};
40 : use crate::aux_file;
41 : use crate::context::RequestContext;
42 : use crate::keyspace::{KeySpace, KeySpaceAccum};
43 : use crate::metrics::{
44 : RELSIZE_CACHE_ENTRIES, RELSIZE_CACHE_HITS, RELSIZE_CACHE_MISSES, RELSIZE_CACHE_MISSES_OLD,
45 : };
46 : use crate::span::{
47 : debug_assert_current_span_has_tenant_and_timeline_id,
48 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
49 : };
50 : use crate::tenant::storage_layer::IoConcurrency;
51 : use crate::tenant::timeline::GetVectoredError;
52 :
53 : /// 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.
54 : pub const MAX_AUX_FILE_DELTAS: usize = 1024;
55 :
56 : /// Max number of aux-file-related delta layers. The compaction will create a new image layer once this threshold is reached.
57 : pub const MAX_AUX_FILE_V2_DELTAS: usize = 16;
58 :
59 : #[derive(Debug)]
60 : pub enum LsnForTimestamp {
61 : /// Found commits both before and after the given timestamp
62 : Present(Lsn),
63 :
64 : /// Found no commits after the given timestamp, this means
65 : /// that the newest data in the branch is older than the given
66 : /// timestamp.
67 : ///
68 : /// All commits <= LSN happened before the given timestamp
69 : Future(Lsn),
70 :
71 : /// The queried timestamp is past our horizon we look back at (PITR)
72 : ///
73 : /// All commits > LSN happened after the given timestamp,
74 : /// but any commits < LSN might have happened before or after
75 : /// the given timestamp. We don't know because no data before
76 : /// the given lsn is available.
77 : Past(Lsn),
78 :
79 : /// We have found no commit with a timestamp,
80 : /// so we can't return anything meaningful.
81 : ///
82 : /// The associated LSN is the lower bound value we can safely
83 : /// create branches on, but no statement is made if it is
84 : /// older or newer than the timestamp.
85 : ///
86 : /// This variant can e.g. be returned right after a
87 : /// cluster import.
88 : NoData(Lsn),
89 : }
90 :
91 : #[derive(Debug, thiserror::Error)]
92 : pub(crate) enum CalculateLogicalSizeError {
93 : #[error("cancelled")]
94 : Cancelled,
95 :
96 : /// Something went wrong while reading the metadata we use to calculate logical size
97 : /// Note that cancellation variants of `PageReconstructError` are transformed to [`Self::Cancelled`]
98 : /// in the `From` implementation for this variant.
99 : #[error(transparent)]
100 : PageRead(PageReconstructError),
101 :
102 : /// Something went wrong deserializing metadata that we read to calculate logical size
103 : #[error("decode error: {0}")]
104 : Decode(#[from] DeserializeError),
105 : }
106 :
107 : #[derive(Debug, thiserror::Error)]
108 : pub(crate) enum CollectKeySpaceError {
109 : #[error(transparent)]
110 : Decode(#[from] DeserializeError),
111 : #[error(transparent)]
112 : PageRead(PageReconstructError),
113 : #[error("cancelled")]
114 : Cancelled,
115 : }
116 :
117 : impl From<PageReconstructError> for CollectKeySpaceError {
118 0 : fn from(err: PageReconstructError) -> Self {
119 0 : match err {
120 0 : PageReconstructError::Cancelled => Self::Cancelled,
121 0 : err => Self::PageRead(err),
122 : }
123 0 : }
124 : }
125 :
126 : impl From<PageReconstructError> for CalculateLogicalSizeError {
127 0 : fn from(pre: PageReconstructError) -> Self {
128 0 : match pre {
129 0 : PageReconstructError::Cancelled => Self::Cancelled,
130 0 : _ => Self::PageRead(pre),
131 : }
132 0 : }
133 : }
134 :
135 : #[derive(Debug, thiserror::Error)]
136 : pub enum RelationError {
137 : #[error("Relation Already Exists")]
138 : AlreadyExists,
139 : #[error("invalid relnode")]
140 : InvalidRelnode,
141 : #[error(transparent)]
142 : Other(#[from] anyhow::Error),
143 : }
144 :
145 : ///
146 : /// This impl provides all the functionality to store PostgreSQL relations, SLRUs,
147 : /// and other special kinds of files, in a versioned key-value store. The
148 : /// Timeline struct provides the key-value store.
149 : ///
150 : /// This is a separate impl, so that we can easily include all these functions in a Timeline
151 : /// implementation, and might be moved into a separate struct later.
152 : impl Timeline {
153 : /// Start ingesting a WAL record, or other atomic modification of
154 : /// the timeline.
155 : ///
156 : /// This provides a transaction-like interface to perform a bunch
157 : /// of modifications atomically.
158 : ///
159 : /// To ingest a WAL record, call begin_modification(lsn) to get a
160 : /// DatadirModification object. Use the functions in the object to
161 : /// modify the repository state, updating all the pages and metadata
162 : /// that the WAL record affects. When you're done, call commit() to
163 : /// commit the changes.
164 : ///
165 : /// Lsn stored in modification is advanced by `ingest_record` and
166 : /// is used by `commit()` to update `last_record_lsn`.
167 : ///
168 : /// Calling commit() will flush all the changes and reset the state,
169 : /// so the `DatadirModification` struct can be reused to perform the next modification.
170 : ///
171 : /// Note that any pending modifications you make through the
172 : /// modification object won't be visible to calls to the 'get' and list
173 : /// functions of the timeline until you finish! And if you update the
174 : /// same page twice, the last update wins.
175 : ///
176 536820 : pub fn begin_modification(&self, lsn: Lsn) -> DatadirModification
177 536820 : where
178 536820 : Self: Sized,
179 536820 : {
180 536820 : DatadirModification {
181 536820 : tline: self,
182 536820 : pending_lsns: Vec::new(),
183 536820 : pending_metadata_pages: HashMap::new(),
184 536820 : pending_data_batch: None,
185 536820 : pending_deletions: Vec::new(),
186 536820 : pending_nblocks: 0,
187 536820 : pending_directory_entries: Vec::new(),
188 536820 : pending_metadata_bytes: 0,
189 536820 : lsn,
190 536820 : }
191 536820 : }
192 :
193 : //------------------------------------------------------------------------------
194 : // Public GET functions
195 : //------------------------------------------------------------------------------
196 :
197 : /// Look up given page version.
198 36768 : pub(crate) async fn get_rel_page_at_lsn(
199 36768 : &self,
200 36768 : tag: RelTag,
201 36768 : blknum: BlockNumber,
202 36768 : version: Version<'_>,
203 36768 : ctx: &RequestContext,
204 36768 : io_concurrency: IoConcurrency,
205 36768 : ) -> Result<Bytes, PageReconstructError> {
206 36768 : match version {
207 36768 : Version::Lsn(effective_lsn) => {
208 36768 : let pages: smallvec::SmallVec<[_; 1]> = smallvec::smallvec![(tag, blknum)];
209 36768 : let res = self
210 36768 : .get_rel_page_at_lsn_batched(
211 36768 : pages.iter().map(|(tag, blknum)| (tag, blknum)),
212 36768 : effective_lsn,
213 36768 : io_concurrency.clone(),
214 36768 : ctx,
215 36768 : )
216 36768 : .await;
217 36768 : assert_eq!(res.len(), 1);
218 36768 : res.into_iter().next().unwrap()
219 : }
220 0 : Version::Modified(modification) => {
221 0 : if tag.relnode == 0 {
222 0 : return Err(PageReconstructError::Other(
223 0 : RelationError::InvalidRelnode.into(),
224 0 : ));
225 0 : }
226 :
227 0 : let nblocks = self.get_rel_size(tag, version, ctx).await?;
228 0 : if blknum >= nblocks {
229 0 : debug!(
230 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
231 0 : tag,
232 0 : blknum,
233 0 : version.get_lsn(),
234 : nblocks
235 : );
236 0 : return Ok(ZERO_PAGE.clone());
237 0 : }
238 0 :
239 0 : let key = rel_block_to_key(tag, blknum);
240 0 : modification.get(key, ctx).await
241 : }
242 : }
243 36768 : }
244 :
245 : /// Like [`Self::get_rel_page_at_lsn`], but returns a batch of pages.
246 : ///
247 : /// The ordering of the returned vec corresponds to the ordering of `pages`.
248 36768 : pub(crate) async fn get_rel_page_at_lsn_batched(
249 36768 : &self,
250 36768 : pages: impl ExactSizeIterator<Item = (&RelTag, &BlockNumber)>,
251 36768 : effective_lsn: Lsn,
252 36768 : io_concurrency: IoConcurrency,
253 36768 : ctx: &RequestContext,
254 36768 : ) -> Vec<Result<Bytes, PageReconstructError>> {
255 36768 : debug_assert_current_span_has_tenant_and_timeline_id();
256 36768 :
257 36768 : let mut slots_filled = 0;
258 36768 : let page_count = pages.len();
259 36768 :
260 36768 : // Would be nice to use smallvec here but it doesn't provide the spare_capacity_mut() API.
261 36768 : let mut result = Vec::with_capacity(pages.len());
262 36768 : let result_slots = result.spare_capacity_mut();
263 36768 :
264 36768 : let mut keys_slots: BTreeMap<Key, smallvec::SmallVec<[usize; 1]>> = BTreeMap::default();
265 36768 : for (response_slot_idx, (tag, blknum)) in pages.enumerate() {
266 36768 : if tag.relnode == 0 {
267 0 : result_slots[response_slot_idx].write(Err(PageReconstructError::Other(
268 0 : RelationError::InvalidRelnode.into(),
269 0 : )));
270 0 :
271 0 : slots_filled += 1;
272 0 : continue;
273 36768 : }
274 :
275 36768 : let nblocks = match self
276 36768 : .get_rel_size(*tag, Version::Lsn(effective_lsn), ctx)
277 36768 : .await
278 : {
279 36768 : Ok(nblocks) => nblocks,
280 0 : Err(err) => {
281 0 : result_slots[response_slot_idx].write(Err(err));
282 0 : slots_filled += 1;
283 0 : continue;
284 : }
285 : };
286 :
287 36768 : if *blknum >= nblocks {
288 0 : debug!(
289 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
290 : tag, blknum, effective_lsn, nblocks
291 : );
292 0 : result_slots[response_slot_idx].write(Ok(ZERO_PAGE.clone()));
293 0 : slots_filled += 1;
294 0 : continue;
295 36768 : }
296 36768 :
297 36768 : let key = rel_block_to_key(*tag, *blknum);
298 36768 :
299 36768 : let key_slots = keys_slots.entry(key).or_default();
300 36768 : key_slots.push(response_slot_idx);
301 : }
302 :
303 36768 : let keyspace = {
304 : // add_key requires monotonicity
305 36768 : let mut acc = KeySpaceAccum::new();
306 36768 : for key in keys_slots
307 36768 : .keys()
308 36768 : // in fact it requires strong monotonicity
309 36768 : .dedup()
310 36768 : {
311 36768 : acc.add_key(*key);
312 36768 : }
313 36768 : acc.to_keyspace()
314 36768 : };
315 36768 :
316 36768 : match self
317 36768 : .get_vectored(keyspace, effective_lsn, io_concurrency, ctx)
318 36768 : .await
319 : {
320 36768 : Ok(results) => {
321 73536 : for (key, res) in results {
322 36768 : let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
323 36768 : let first_slot = key_slots.next().unwrap();
324 :
325 36768 : for slot in key_slots {
326 0 : let clone = match &res {
327 0 : Ok(buf) => Ok(buf.clone()),
328 0 : Err(err) => Err(match err {
329 0 : PageReconstructError::Cancelled => PageReconstructError::Cancelled,
330 :
331 0 : x @ PageReconstructError::Other(_)
332 0 : | x @ PageReconstructError::AncestorLsnTimeout(_)
333 0 : | x @ PageReconstructError::WalRedo(_)
334 0 : | x @ PageReconstructError::MissingKey(_) => {
335 0 : PageReconstructError::Other(anyhow::anyhow!(
336 0 : "there was more than one request for this key in the batch, error logged once: {x:?}"
337 0 : ))
338 : }
339 : }),
340 : };
341 :
342 0 : result_slots[slot].write(clone);
343 0 : slots_filled += 1;
344 : }
345 :
346 36768 : result_slots[first_slot].write(res);
347 36768 : slots_filled += 1;
348 : }
349 : }
350 0 : Err(err) => {
351 : // this cannot really happen because get_vectored only errors globally on invalid LSN or too large batch size
352 : // (We enforce the max batch size outside of this function, in the code that constructs the batch request.)
353 0 : for slot in keys_slots.values().flatten() {
354 : // this whole `match` is a lot like `From<GetVectoredError> for PageReconstructError`
355 : // but without taking ownership of the GetVectoredError
356 0 : let err = match &err {
357 0 : GetVectoredError::Cancelled => Err(PageReconstructError::Cancelled),
358 : // TODO: restructure get_vectored API to make this error per-key
359 0 : GetVectoredError::MissingKey(err) => {
360 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
361 0 : "whole vectored get request failed because one or more of the requested keys were missing: {err:?}"
362 0 : )))
363 : }
364 : // TODO: restructure get_vectored API to make this error per-key
365 0 : GetVectoredError::GetReadyAncestorError(err) => {
366 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
367 0 : "whole vectored get request failed because one or more key required ancestor that wasn't ready: {err:?}"
368 0 : )))
369 : }
370 : // TODO: restructure get_vectored API to make this error per-key
371 0 : GetVectoredError::Other(err) => Err(PageReconstructError::Other(
372 0 : anyhow::anyhow!("whole vectored get request failed: {err:?}"),
373 0 : )),
374 : // TODO: we can prevent this error class by moving this check into the type system
375 0 : GetVectoredError::InvalidLsn(e) => {
376 0 : Err(anyhow::anyhow!("invalid LSN: {e:?}").into())
377 : }
378 : // NB: this should never happen in practice because we limit MAX_GET_VECTORED_KEYS
379 : // TODO: we can prevent this error class by moving this check into the type system
380 0 : GetVectoredError::Oversized(err) => {
381 0 : Err(anyhow::anyhow!("batching oversized: {err:?}").into())
382 : }
383 : };
384 :
385 0 : result_slots[*slot].write(err);
386 : }
387 :
388 0 : slots_filled += keys_slots.values().map(|slots| slots.len()).sum::<usize>();
389 0 : }
390 : };
391 :
392 36768 : assert_eq!(slots_filled, page_count);
393 : // SAFETY:
394 : // 1. `result` and any of its uninint members are not read from until this point
395 : // 2. The length below is tracked at run-time and matches the number of requested pages.
396 36768 : unsafe {
397 36768 : result.set_len(page_count);
398 36768 : }
399 36768 :
400 36768 : result
401 36768 : }
402 :
403 : /// Get size of a database in blocks. This is only accurate on shard 0. It will undercount on
404 : /// other shards, by only accounting for relations the shard has pages for, and only accounting
405 : /// for pages up to the highest page number it has stored.
406 0 : pub(crate) async fn get_db_size(
407 0 : &self,
408 0 : spcnode: Oid,
409 0 : dbnode: Oid,
410 0 : version: Version<'_>,
411 0 : ctx: &RequestContext,
412 0 : ) -> Result<usize, PageReconstructError> {
413 0 : let mut total_blocks = 0;
414 :
415 0 : let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
416 :
417 0 : for rel in rels {
418 0 : let n_blocks = self.get_rel_size(rel, version, ctx).await?;
419 0 : total_blocks += n_blocks as usize;
420 : }
421 0 : Ok(total_blocks)
422 0 : }
423 :
424 : /// Get size of a relation file. The relation must exist, otherwise an error is returned.
425 : ///
426 : /// This is only accurate on shard 0. On other shards, it will return the size up to the highest
427 : /// page number stored in the shard.
428 48868 : pub(crate) async fn get_rel_size(
429 48868 : &self,
430 48868 : tag: RelTag,
431 48868 : version: Version<'_>,
432 48868 : ctx: &RequestContext,
433 48868 : ) -> Result<BlockNumber, PageReconstructError> {
434 48868 : if tag.relnode == 0 {
435 0 : return Err(PageReconstructError::Other(
436 0 : RelationError::InvalidRelnode.into(),
437 0 : ));
438 48868 : }
439 :
440 48868 : if let Some(nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
441 38588 : return Ok(nblocks);
442 10280 : }
443 10280 :
444 10280 : if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
445 0 : && !self.get_rel_exists(tag, version, ctx).await?
446 : {
447 : // FIXME: Postgres sometimes calls smgrcreate() to create
448 : // FSM, and smgrnblocks() on it immediately afterwards,
449 : // without extending it. Tolerate that by claiming that
450 : // any non-existent FSM fork has size 0.
451 0 : return Ok(0);
452 10280 : }
453 10280 :
454 10280 : let key = rel_size_to_key(tag);
455 10280 : let mut buf = version.get(self, key, ctx).await?;
456 10272 : let nblocks = buf.get_u32_le();
457 10272 :
458 10272 : self.update_cached_rel_size(tag, version.get_lsn(), nblocks);
459 10272 :
460 10272 : Ok(nblocks)
461 48868 : }
462 :
463 : /// Does the relation exist?
464 : ///
465 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
466 : /// the shard stores pages for.
467 12100 : pub(crate) async fn get_rel_exists(
468 12100 : &self,
469 12100 : tag: RelTag,
470 12100 : version: Version<'_>,
471 12100 : ctx: &RequestContext,
472 12100 : ) -> Result<bool, PageReconstructError> {
473 12100 : if tag.relnode == 0 {
474 0 : return Err(PageReconstructError::Other(
475 0 : RelationError::InvalidRelnode.into(),
476 0 : ));
477 12100 : }
478 :
479 : // first try to lookup relation in cache
480 12100 : if let Some(_nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
481 12064 : return Ok(true);
482 36 : }
483 : // then check if the database was already initialized.
484 : // get_rel_exists can be called before dbdir is created.
485 36 : let buf = version.get(self, DBDIR_KEY, ctx).await?;
486 36 : let dbdirs = DbDirectory::des(&buf)?.dbdirs;
487 36 : if !dbdirs.contains_key(&(tag.spcnode, tag.dbnode)) {
488 0 : return Ok(false);
489 36 : }
490 36 :
491 36 : // Read path: first read the new reldir keyspace. Early return if the relation exists.
492 36 : // Otherwise, read the old reldir keyspace.
493 36 : // TODO: if IndexPart::rel_size_migration is `Migrated`, we only need to read from v2.
494 36 :
495 36 : if self.get_rel_size_v2_enabled() {
496 : // fetch directory listing (new)
497 0 : let key = rel_tag_sparse_key(tag.spcnode, tag.dbnode, tag.relnode, tag.forknum);
498 0 : let buf = RelDirExists::decode_option(version.sparse_get(self, key, ctx).await?)
499 0 : .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
500 0 : let exists_v2 = buf == RelDirExists::Exists;
501 0 : // Fast path: if the relation exists in the new format, return true.
502 0 : // TODO: we should have a verification mode that checks both keyspaces
503 0 : // to ensure the relation only exists in one of them.
504 0 : if exists_v2 {
505 0 : return Ok(true);
506 0 : }
507 36 : }
508 :
509 : // fetch directory listing (old)
510 :
511 36 : let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
512 36 : let buf = version.get(self, key, ctx).await?;
513 :
514 36 : let dir = RelDirectory::des(&buf)?;
515 36 : let exists_v1 = dir.rels.contains(&(tag.relnode, tag.forknum));
516 36 : Ok(exists_v1)
517 12100 : }
518 :
519 : /// Get a list of all existing relations in given tablespace and database.
520 : ///
521 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
522 : /// the shard stores pages for.
523 : ///
524 : /// # Cancel-Safety
525 : ///
526 : /// This method is cancellation-safe.
527 0 : pub(crate) async fn list_rels(
528 0 : &self,
529 0 : spcnode: Oid,
530 0 : dbnode: Oid,
531 0 : version: Version<'_>,
532 0 : ctx: &RequestContext,
533 0 : ) -> Result<HashSet<RelTag>, PageReconstructError> {
534 0 : // fetch directory listing (old)
535 0 : let key = rel_dir_to_key(spcnode, dbnode);
536 0 : let buf = version.get(self, key, ctx).await?;
537 :
538 0 : let dir = RelDirectory::des(&buf)?;
539 0 : let rels_v1: HashSet<RelTag> =
540 0 : HashSet::from_iter(dir.rels.iter().map(|(relnode, forknum)| RelTag {
541 0 : spcnode,
542 0 : dbnode,
543 0 : relnode: *relnode,
544 0 : forknum: *forknum,
545 0 : }));
546 0 :
547 0 : if !self.get_rel_size_v2_enabled() {
548 0 : return Ok(rels_v1);
549 0 : }
550 0 :
551 0 : // scan directory listing (new), merge with the old results
552 0 : let key_range = rel_tag_sparse_key_range(spcnode, dbnode);
553 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
554 0 : self.conf,
555 0 : self.gate
556 0 : .enter()
557 0 : .map_err(|_| PageReconstructError::Cancelled)?,
558 : );
559 0 : let results = self
560 0 : .scan(
561 0 : KeySpace::single(key_range),
562 0 : version.get_lsn(),
563 0 : ctx,
564 0 : io_concurrency,
565 0 : )
566 0 : .await?;
567 0 : let mut rels = rels_v1;
568 0 : for (key, val) in results {
569 0 : let val = RelDirExists::decode(&val?)
570 0 : .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
571 0 : assert_eq!(key.field6, 1);
572 0 : assert_eq!(key.field2, spcnode);
573 0 : assert_eq!(key.field3, dbnode);
574 0 : let tag = RelTag {
575 0 : spcnode,
576 0 : dbnode,
577 0 : relnode: key.field4,
578 0 : forknum: key.field5,
579 0 : };
580 0 : if val == RelDirExists::Removed {
581 0 : debug_assert!(!rels.contains(&tag), "removed reltag in v2");
582 0 : continue;
583 0 : }
584 0 : let did_not_contain = rels.insert(tag);
585 0 : debug_assert!(did_not_contain, "duplicate reltag in v2");
586 : }
587 0 : Ok(rels)
588 0 : }
589 :
590 : /// Get the whole SLRU segment
591 0 : pub(crate) async fn get_slru_segment(
592 0 : &self,
593 0 : kind: SlruKind,
594 0 : segno: u32,
595 0 : lsn: Lsn,
596 0 : ctx: &RequestContext,
597 0 : ) -> Result<Bytes, PageReconstructError> {
598 0 : assert!(self.tenant_shard_id.is_shard_zero());
599 0 : let n_blocks = self
600 0 : .get_slru_segment_size(kind, segno, Version::Lsn(lsn), ctx)
601 0 : .await?;
602 0 : let mut segment = BytesMut::with_capacity(n_blocks as usize * BLCKSZ as usize);
603 0 : for blkno in 0..n_blocks {
604 0 : let block = self
605 0 : .get_slru_page_at_lsn(kind, segno, blkno, lsn, ctx)
606 0 : .await?;
607 0 : segment.extend_from_slice(&block[..BLCKSZ as usize]);
608 : }
609 0 : Ok(segment.freeze())
610 0 : }
611 :
612 : /// Look up given SLRU page version.
613 0 : pub(crate) async fn get_slru_page_at_lsn(
614 0 : &self,
615 0 : kind: SlruKind,
616 0 : segno: u32,
617 0 : blknum: BlockNumber,
618 0 : lsn: Lsn,
619 0 : ctx: &RequestContext,
620 0 : ) -> Result<Bytes, PageReconstructError> {
621 0 : assert!(self.tenant_shard_id.is_shard_zero());
622 0 : let key = slru_block_to_key(kind, segno, blknum);
623 0 : self.get(key, lsn, ctx).await
624 0 : }
625 :
626 : /// Get size of an SLRU segment
627 0 : pub(crate) async fn get_slru_segment_size(
628 0 : &self,
629 0 : kind: SlruKind,
630 0 : segno: u32,
631 0 : version: Version<'_>,
632 0 : ctx: &RequestContext,
633 0 : ) -> Result<BlockNumber, PageReconstructError> {
634 0 : assert!(self.tenant_shard_id.is_shard_zero());
635 0 : let key = slru_segment_size_to_key(kind, segno);
636 0 : let mut buf = version.get(self, key, ctx).await?;
637 0 : Ok(buf.get_u32_le())
638 0 : }
639 :
640 : /// Get size of an SLRU segment
641 0 : pub(crate) async fn get_slru_segment_exists(
642 0 : &self,
643 0 : kind: SlruKind,
644 0 : segno: u32,
645 0 : version: Version<'_>,
646 0 : ctx: &RequestContext,
647 0 : ) -> Result<bool, PageReconstructError> {
648 0 : assert!(self.tenant_shard_id.is_shard_zero());
649 : // fetch directory listing
650 0 : let key = slru_dir_to_key(kind);
651 0 : let buf = version.get(self, key, ctx).await?;
652 :
653 0 : let dir = SlruSegmentDirectory::des(&buf)?;
654 0 : Ok(dir.segments.contains(&segno))
655 0 : }
656 :
657 : /// Locate LSN, such that all transactions that committed before
658 : /// 'search_timestamp' are visible, but nothing newer is.
659 : ///
660 : /// This is not exact. Commit timestamps are not guaranteed to be ordered,
661 : /// so it's not well defined which LSN you get if there were multiple commits
662 : /// "in flight" at that point in time.
663 : ///
664 0 : pub(crate) async fn find_lsn_for_timestamp(
665 0 : &self,
666 0 : search_timestamp: TimestampTz,
667 0 : cancel: &CancellationToken,
668 0 : ctx: &RequestContext,
669 0 : ) -> Result<LsnForTimestamp, PageReconstructError> {
670 0 : pausable_failpoint!("find-lsn-for-timestamp-pausable");
671 :
672 0 : let gc_cutoff_lsn_guard = self.get_applied_gc_cutoff_lsn();
673 0 : let gc_cutoff_planned = {
674 0 : let gc_info = self.gc_info.read().unwrap();
675 0 : gc_info.min_cutoff()
676 0 : };
677 0 : // Usually the planned cutoff is newer than the cutoff of the last gc run,
678 0 : // but let's be defensive.
679 0 : let gc_cutoff = gc_cutoff_planned.max(*gc_cutoff_lsn_guard);
680 0 : // We use this method to figure out the branching LSN for the new branch, but the
681 0 : // GC cutoff could be before the branching point and we cannot create a new branch
682 0 : // with LSN < `ancestor_lsn`. Thus, pick the maximum of these two to be
683 0 : // on the safe side.
684 0 : let min_lsn = std::cmp::max(gc_cutoff, self.get_ancestor_lsn());
685 0 : let max_lsn = self.get_last_record_lsn();
686 0 :
687 0 : // LSNs are always 8-byte aligned. low/mid/high represent the
688 0 : // LSN divided by 8.
689 0 : let mut low = min_lsn.0 / 8;
690 0 : let mut high = max_lsn.0 / 8 + 1;
691 0 :
692 0 : let mut found_smaller = false;
693 0 : let mut found_larger = false;
694 :
695 0 : while low < high {
696 0 : if cancel.is_cancelled() {
697 0 : return Err(PageReconstructError::Cancelled);
698 0 : }
699 0 : // cannot overflow, high and low are both smaller than u64::MAX / 2
700 0 : let mid = (high + low) / 2;
701 :
702 0 : let cmp = match self
703 0 : .is_latest_commit_timestamp_ge_than(
704 0 : search_timestamp,
705 0 : Lsn(mid * 8),
706 0 : &mut found_smaller,
707 0 : &mut found_larger,
708 0 : ctx,
709 0 : )
710 0 : .await
711 : {
712 0 : Ok(res) => res,
713 0 : Err(PageReconstructError::MissingKey(e)) => {
714 0 : warn!(
715 0 : "Missing key while find_lsn_for_timestamp. Either we might have already garbage-collected that data or the key is really missing. Last error: {:#}",
716 : e
717 : );
718 : // Return that we didn't find any requests smaller than the LSN, and logging the error.
719 0 : return Ok(LsnForTimestamp::Past(min_lsn));
720 : }
721 0 : Err(e) => return Err(e),
722 : };
723 :
724 0 : if cmp {
725 0 : high = mid;
726 0 : } else {
727 0 : low = mid + 1;
728 0 : }
729 : }
730 :
731 : // If `found_smaller == true`, `low = t + 1` where `t` is the target LSN,
732 : // so the LSN of the last commit record before or at `search_timestamp`.
733 : // Remove one from `low` to get `t`.
734 : //
735 : // FIXME: it would be better to get the LSN of the previous commit.
736 : // Otherwise, if you restore to the returned LSN, the database will
737 : // include physical changes from later commits that will be marked
738 : // as aborted, and will need to be vacuumed away.
739 0 : let commit_lsn = Lsn((low - 1) * 8);
740 0 : match (found_smaller, found_larger) {
741 : (false, false) => {
742 : // This can happen if no commit records have been processed yet, e.g.
743 : // just after importing a cluster.
744 0 : Ok(LsnForTimestamp::NoData(min_lsn))
745 : }
746 : (false, true) => {
747 : // Didn't find any commit timestamps smaller than the request
748 0 : Ok(LsnForTimestamp::Past(min_lsn))
749 : }
750 0 : (true, _) if commit_lsn < min_lsn => {
751 0 : // the search above did set found_smaller to true but it never increased the lsn.
752 0 : // Then, low is still the old min_lsn, and the subtraction above gave a value
753 0 : // below the min_lsn. We should never do that.
754 0 : Ok(LsnForTimestamp::Past(min_lsn))
755 : }
756 : (true, false) => {
757 : // Only found commits with timestamps smaller than the request.
758 : // It's still a valid case for branch creation, return it.
759 : // And `update_gc_info()` ignores LSN for a `LsnForTimestamp::Future`
760 : // case, anyway.
761 0 : Ok(LsnForTimestamp::Future(commit_lsn))
762 : }
763 0 : (true, true) => Ok(LsnForTimestamp::Present(commit_lsn)),
764 : }
765 0 : }
766 :
767 : /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
768 : /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
769 : ///
770 : /// Additionally, sets 'found_smaller'/'found_Larger, if encounters any commits
771 : /// with a smaller/larger timestamp.
772 : ///
773 0 : pub(crate) async fn is_latest_commit_timestamp_ge_than(
774 0 : &self,
775 0 : search_timestamp: TimestampTz,
776 0 : probe_lsn: Lsn,
777 0 : found_smaller: &mut bool,
778 0 : found_larger: &mut bool,
779 0 : ctx: &RequestContext,
780 0 : ) -> Result<bool, PageReconstructError> {
781 0 : self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
782 0 : if timestamp >= search_timestamp {
783 0 : *found_larger = true;
784 0 : return ControlFlow::Break(true);
785 0 : } else {
786 0 : *found_smaller = true;
787 0 : }
788 0 : ControlFlow::Continue(())
789 0 : })
790 0 : .await
791 0 : }
792 :
793 : /// Obtain the possible timestamp range for the given lsn.
794 : ///
795 : /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
796 0 : pub(crate) async fn get_timestamp_for_lsn(
797 0 : &self,
798 0 : probe_lsn: Lsn,
799 0 : ctx: &RequestContext,
800 0 : ) -> Result<Option<TimestampTz>, PageReconstructError> {
801 0 : let mut max: Option<TimestampTz> = None;
802 0 : self.map_all_timestamps::<()>(probe_lsn, ctx, |timestamp| {
803 0 : if let Some(max_prev) = max {
804 0 : max = Some(max_prev.max(timestamp));
805 0 : } else {
806 0 : max = Some(timestamp);
807 0 : }
808 0 : ControlFlow::Continue(())
809 0 : })
810 0 : .await?;
811 :
812 0 : Ok(max)
813 0 : }
814 :
815 : /// Runs the given function on all the timestamps for a given lsn
816 : ///
817 : /// The return value is either given by the closure, or set to the `Default`
818 : /// impl's output.
819 0 : async fn map_all_timestamps<T: Default>(
820 0 : &self,
821 0 : probe_lsn: Lsn,
822 0 : ctx: &RequestContext,
823 0 : mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
824 0 : ) -> Result<T, PageReconstructError> {
825 0 : for segno in self
826 0 : .list_slru_segments(SlruKind::Clog, Version::Lsn(probe_lsn), ctx)
827 0 : .await?
828 : {
829 0 : let nblocks = self
830 0 : .get_slru_segment_size(SlruKind::Clog, segno, Version::Lsn(probe_lsn), ctx)
831 0 : .await?;
832 0 : for blknum in (0..nblocks).rev() {
833 0 : let clog_page = self
834 0 : .get_slru_page_at_lsn(SlruKind::Clog, segno, blknum, probe_lsn, ctx)
835 0 : .await?;
836 :
837 0 : if clog_page.len() == BLCKSZ as usize + 8 {
838 0 : let mut timestamp_bytes = [0u8; 8];
839 0 : timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
840 0 : let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
841 0 :
842 0 : match f(timestamp) {
843 0 : ControlFlow::Break(b) => return Ok(b),
844 0 : ControlFlow::Continue(()) => (),
845 : }
846 0 : }
847 : }
848 : }
849 0 : Ok(Default::default())
850 0 : }
851 :
852 0 : pub(crate) async fn get_slru_keyspace(
853 0 : &self,
854 0 : version: Version<'_>,
855 0 : ctx: &RequestContext,
856 0 : ) -> Result<KeySpace, PageReconstructError> {
857 0 : let mut accum = KeySpaceAccum::new();
858 :
859 0 : for kind in SlruKind::iter() {
860 0 : let mut segments: Vec<u32> = self
861 0 : .list_slru_segments(kind, version, ctx)
862 0 : .await?
863 0 : .into_iter()
864 0 : .collect();
865 0 : segments.sort_unstable();
866 :
867 0 : for seg in segments {
868 0 : let block_count = self.get_slru_segment_size(kind, seg, version, ctx).await?;
869 :
870 0 : accum.add_range(
871 0 : slru_block_to_key(kind, seg, 0)..slru_block_to_key(kind, seg, block_count),
872 0 : );
873 : }
874 : }
875 :
876 0 : Ok(accum.to_keyspace())
877 0 : }
878 :
879 : /// Get a list of SLRU segments
880 0 : pub(crate) async fn list_slru_segments(
881 0 : &self,
882 0 : kind: SlruKind,
883 0 : version: Version<'_>,
884 0 : ctx: &RequestContext,
885 0 : ) -> Result<HashSet<u32>, PageReconstructError> {
886 0 : // fetch directory entry
887 0 : let key = slru_dir_to_key(kind);
888 :
889 0 : let buf = version.get(self, key, ctx).await?;
890 0 : Ok(SlruSegmentDirectory::des(&buf)?.segments)
891 0 : }
892 :
893 0 : pub(crate) async fn get_relmap_file(
894 0 : &self,
895 0 : spcnode: Oid,
896 0 : dbnode: Oid,
897 0 : version: Version<'_>,
898 0 : ctx: &RequestContext,
899 0 : ) -> Result<Bytes, PageReconstructError> {
900 0 : let key = relmap_file_key(spcnode, dbnode);
901 :
902 0 : let buf = version.get(self, key, ctx).await?;
903 0 : Ok(buf)
904 0 : }
905 :
906 644 : pub(crate) async fn list_dbdirs(
907 644 : &self,
908 644 : lsn: Lsn,
909 644 : ctx: &RequestContext,
910 644 : ) -> Result<HashMap<(Oid, Oid), bool>, PageReconstructError> {
911 : // fetch directory entry
912 644 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
913 :
914 644 : Ok(DbDirectory::des(&buf)?.dbdirs)
915 644 : }
916 :
917 0 : pub(crate) async fn get_twophase_file(
918 0 : &self,
919 0 : xid: u64,
920 0 : lsn: Lsn,
921 0 : ctx: &RequestContext,
922 0 : ) -> Result<Bytes, PageReconstructError> {
923 0 : let key = twophase_file_key(xid);
924 0 : let buf = self.get(key, lsn, ctx).await?;
925 0 : Ok(buf)
926 0 : }
927 :
928 648 : pub(crate) async fn list_twophase_files(
929 648 : &self,
930 648 : lsn: Lsn,
931 648 : ctx: &RequestContext,
932 648 : ) -> Result<HashSet<u64>, PageReconstructError> {
933 : // fetch directory entry
934 648 : let buf = self.get(TWOPHASEDIR_KEY, lsn, ctx).await?;
935 :
936 648 : if self.pg_version >= 17 {
937 0 : Ok(TwoPhaseDirectoryV17::des(&buf)?.xids)
938 : } else {
939 648 : Ok(TwoPhaseDirectory::des(&buf)?
940 : .xids
941 648 : .iter()
942 648 : .map(|x| u64::from(*x))
943 648 : .collect())
944 : }
945 648 : }
946 :
947 0 : pub(crate) async fn get_control_file(
948 0 : &self,
949 0 : lsn: Lsn,
950 0 : ctx: &RequestContext,
951 0 : ) -> Result<Bytes, PageReconstructError> {
952 0 : self.get(CONTROLFILE_KEY, lsn, ctx).await
953 0 : }
954 :
955 24 : pub(crate) async fn get_checkpoint(
956 24 : &self,
957 24 : lsn: Lsn,
958 24 : ctx: &RequestContext,
959 24 : ) -> Result<Bytes, PageReconstructError> {
960 24 : self.get(CHECKPOINT_KEY, lsn, ctx).await
961 24 : }
962 :
963 24 : async fn list_aux_files_v2(
964 24 : &self,
965 24 : lsn: Lsn,
966 24 : ctx: &RequestContext,
967 24 : io_concurrency: IoConcurrency,
968 24 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
969 24 : let kv = self
970 24 : .scan(
971 24 : KeySpace::single(Key::metadata_aux_key_range()),
972 24 : lsn,
973 24 : ctx,
974 24 : io_concurrency,
975 24 : )
976 24 : .await?;
977 24 : let mut result = HashMap::new();
978 24 : let mut sz = 0;
979 60 : for (_, v) in kv {
980 36 : let v = v?;
981 36 : let v = aux_file::decode_file_value_bytes(&v)
982 36 : .context("value decode")
983 36 : .map_err(PageReconstructError::Other)?;
984 68 : for (fname, content) in v {
985 32 : sz += fname.len();
986 32 : sz += content.len();
987 32 : result.insert(fname, content);
988 32 : }
989 : }
990 24 : self.aux_file_size_estimator.on_initial(sz);
991 24 : Ok(result)
992 24 : }
993 :
994 0 : pub(crate) async fn trigger_aux_file_size_computation(
995 0 : &self,
996 0 : lsn: Lsn,
997 0 : ctx: &RequestContext,
998 0 : io_concurrency: IoConcurrency,
999 0 : ) -> Result<(), PageReconstructError> {
1000 0 : self.list_aux_files_v2(lsn, ctx, io_concurrency).await?;
1001 0 : Ok(())
1002 0 : }
1003 :
1004 24 : pub(crate) async fn list_aux_files(
1005 24 : &self,
1006 24 : lsn: Lsn,
1007 24 : ctx: &RequestContext,
1008 24 : io_concurrency: IoConcurrency,
1009 24 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
1010 24 : self.list_aux_files_v2(lsn, ctx, io_concurrency).await
1011 24 : }
1012 :
1013 0 : pub(crate) async fn get_replorigins(
1014 0 : &self,
1015 0 : lsn: Lsn,
1016 0 : ctx: &RequestContext,
1017 0 : io_concurrency: IoConcurrency,
1018 0 : ) -> Result<HashMap<RepOriginId, Lsn>, PageReconstructError> {
1019 0 : let kv = self
1020 0 : .scan(
1021 0 : KeySpace::single(repl_origin_key_range()),
1022 0 : lsn,
1023 0 : ctx,
1024 0 : io_concurrency,
1025 0 : )
1026 0 : .await?;
1027 0 : let mut result = HashMap::new();
1028 0 : for (k, v) in kv {
1029 0 : let v = v?;
1030 0 : let origin_id = k.field6 as RepOriginId;
1031 0 : let origin_lsn = Lsn::des(&v).unwrap();
1032 0 : if origin_lsn != Lsn::INVALID {
1033 0 : result.insert(origin_id, origin_lsn);
1034 0 : }
1035 : }
1036 0 : Ok(result)
1037 0 : }
1038 :
1039 : /// Does the same as get_current_logical_size but counted on demand.
1040 : /// Used to initialize the logical size tracking on startup.
1041 : ///
1042 : /// Only relation blocks are counted currently. That excludes metadata,
1043 : /// SLRUs, twophase files etc.
1044 : ///
1045 : /// # Cancel-Safety
1046 : ///
1047 : /// This method is cancellation-safe.
1048 0 : pub(crate) async fn get_current_logical_size_non_incremental(
1049 0 : &self,
1050 0 : lsn: Lsn,
1051 0 : ctx: &RequestContext,
1052 0 : ) -> Result<u64, CalculateLogicalSizeError> {
1053 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1054 0 :
1055 0 : fail::fail_point!("skip-logical-size-calculation", |_| { Ok(0) });
1056 :
1057 : // Fetch list of database dirs and iterate them
1058 0 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
1059 0 : let dbdir = DbDirectory::des(&buf)?;
1060 :
1061 0 : let mut total_size: u64 = 0;
1062 0 : for (spcnode, dbnode) in dbdir.dbdirs.keys() {
1063 0 : for rel in self
1064 0 : .list_rels(*spcnode, *dbnode, Version::Lsn(lsn), ctx)
1065 0 : .await?
1066 : {
1067 0 : if self.cancel.is_cancelled() {
1068 0 : return Err(CalculateLogicalSizeError::Cancelled);
1069 0 : }
1070 0 : let relsize_key = rel_size_to_key(rel);
1071 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
1072 0 : let relsize = buf.get_u32_le();
1073 0 :
1074 0 : total_size += relsize as u64;
1075 : }
1076 : }
1077 0 : Ok(total_size * BLCKSZ as u64)
1078 0 : }
1079 :
1080 : /// Get a KeySpace that covers all the Keys that are in use at AND below the given LSN. This is only used
1081 : /// for gc-compaction.
1082 : ///
1083 : /// gc-compaction cannot use the same `collect_keyspace` function as the legacy compaction because it
1084 : /// processes data at multiple LSNs and needs to be aware of the fact that some key ranges might need to
1085 : /// be kept only for a specific range of LSN.
1086 : ///
1087 : /// Consider the case that the user created branches at LSN 10 and 20, where the user created a table A at
1088 : /// LSN 10 and dropped that table at LSN 20. `collect_keyspace` at LSN 10 will return the key range
1089 : /// corresponding to that table, while LSN 20 won't. The keyspace info at a single LSN is not enough to
1090 : /// determine which keys to retain/drop for gc-compaction.
1091 : ///
1092 : /// For now, it only drops AUX-v1 keys. But in the future, the function will be extended to return the keyspace
1093 : /// to be retained for each of the branch LSN.
1094 : ///
1095 : /// The return value is (dense keyspace, sparse keyspace).
1096 104 : pub(crate) async fn collect_gc_compaction_keyspace(
1097 104 : &self,
1098 104 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
1099 104 : let metadata_key_begin = Key::metadata_key_range().start;
1100 104 : let aux_v1_key = AUX_FILES_KEY;
1101 104 : let dense_keyspace = KeySpace {
1102 104 : ranges: vec![Key::MIN..aux_v1_key, aux_v1_key.next()..metadata_key_begin],
1103 104 : };
1104 104 : Ok((
1105 104 : dense_keyspace,
1106 104 : SparseKeySpace(KeySpace::single(Key::metadata_key_range())),
1107 104 : ))
1108 104 : }
1109 :
1110 : ///
1111 : /// Get a KeySpace that covers all the Keys that are in use at the given LSN.
1112 : /// Anything that's not listed maybe removed from the underlying storage (from
1113 : /// that LSN forwards).
1114 : ///
1115 : /// The return value is (dense keyspace, sparse keyspace).
1116 644 : pub(crate) async fn collect_keyspace(
1117 644 : &self,
1118 644 : lsn: Lsn,
1119 644 : ctx: &RequestContext,
1120 644 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
1121 644 : // Iterate through key ranges, greedily packing them into partitions
1122 644 : let mut result = KeySpaceAccum::new();
1123 644 :
1124 644 : // The dbdir metadata always exists
1125 644 : result.add_key(DBDIR_KEY);
1126 :
1127 : // Fetch list of database dirs and iterate them
1128 644 : let dbdir = self.list_dbdirs(lsn, ctx).await?;
1129 644 : let mut dbs: Vec<((Oid, Oid), bool)> = dbdir.into_iter().collect();
1130 644 :
1131 644 : dbs.sort_unstable_by(|(k_a, _), (k_b, _)| k_a.cmp(k_b));
1132 644 : for ((spcnode, dbnode), has_relmap_file) in dbs {
1133 0 : if has_relmap_file {
1134 0 : result.add_key(relmap_file_key(spcnode, dbnode));
1135 0 : }
1136 0 : result.add_key(rel_dir_to_key(spcnode, dbnode));
1137 :
1138 0 : let mut rels: Vec<RelTag> = self
1139 0 : .list_rels(spcnode, dbnode, Version::Lsn(lsn), ctx)
1140 0 : .await?
1141 0 : .into_iter()
1142 0 : .collect();
1143 0 : rels.sort_unstable();
1144 0 : for rel in rels {
1145 0 : let relsize_key = rel_size_to_key(rel);
1146 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
1147 0 : let relsize = buf.get_u32_le();
1148 0 :
1149 0 : result.add_range(rel_block_to_key(rel, 0)..rel_block_to_key(rel, relsize));
1150 0 : result.add_key(relsize_key);
1151 : }
1152 : }
1153 :
1154 : // Iterate SLRUs next
1155 644 : if self.tenant_shard_id.is_shard_zero() {
1156 1896 : for kind in [
1157 632 : SlruKind::Clog,
1158 632 : SlruKind::MultiXactMembers,
1159 632 : SlruKind::MultiXactOffsets,
1160 : ] {
1161 1896 : let slrudir_key = slru_dir_to_key(kind);
1162 1896 : result.add_key(slrudir_key);
1163 1896 : let buf = self.get(slrudir_key, lsn, ctx).await?;
1164 1896 : let dir = SlruSegmentDirectory::des(&buf)?;
1165 1896 : let mut segments: Vec<u32> = dir.segments.iter().cloned().collect();
1166 1896 : segments.sort_unstable();
1167 1896 : for segno in segments {
1168 0 : let segsize_key = slru_segment_size_to_key(kind, segno);
1169 0 : let mut buf = self.get(segsize_key, lsn, ctx).await?;
1170 0 : let segsize = buf.get_u32_le();
1171 0 :
1172 0 : result.add_range(
1173 0 : slru_block_to_key(kind, segno, 0)..slru_block_to_key(kind, segno, segsize),
1174 0 : );
1175 0 : result.add_key(segsize_key);
1176 : }
1177 : }
1178 12 : }
1179 :
1180 : // Then pg_twophase
1181 644 : result.add_key(TWOPHASEDIR_KEY);
1182 :
1183 644 : let mut xids: Vec<u64> = self
1184 644 : .list_twophase_files(lsn, ctx)
1185 644 : .await?
1186 644 : .iter()
1187 644 : .cloned()
1188 644 : .collect();
1189 644 : xids.sort_unstable();
1190 644 : for xid in xids {
1191 0 : result.add_key(twophase_file_key(xid));
1192 0 : }
1193 :
1194 644 : result.add_key(CONTROLFILE_KEY);
1195 644 : result.add_key(CHECKPOINT_KEY);
1196 644 :
1197 644 : // Add extra keyspaces in the test cases. Some test cases write keys into the storage without
1198 644 : // creating directory keys. These test cases will add such keyspaces into `extra_test_dense_keyspace`
1199 644 : // and the keys will not be garbage-colllected.
1200 644 : #[cfg(test)]
1201 644 : {
1202 644 : let guard = self.extra_test_dense_keyspace.load();
1203 644 : for kr in &guard.ranges {
1204 0 : result.add_range(kr.clone());
1205 0 : }
1206 0 : }
1207 0 :
1208 644 : let dense_keyspace = result.to_keyspace();
1209 644 : let sparse_keyspace = SparseKeySpace(KeySpace {
1210 644 : ranges: vec![
1211 644 : Key::metadata_aux_key_range(),
1212 644 : repl_origin_key_range(),
1213 644 : Key::rel_dir_sparse_key_range(),
1214 644 : ],
1215 644 : });
1216 644 :
1217 644 : if cfg!(debug_assertions) {
1218 : // Verify if the sparse keyspaces are ordered and non-overlapping.
1219 :
1220 : // We do not use KeySpaceAccum for sparse_keyspace because we want to ensure each
1221 : // category of sparse keys are split into their own image/delta files. If there
1222 : // are overlapping keyspaces, they will be automatically merged by keyspace accum,
1223 : // and we want the developer to keep the keyspaces separated.
1224 :
1225 644 : let ranges = &sparse_keyspace.0.ranges;
1226 :
1227 : // TODO: use a single overlaps_with across the codebase
1228 1932 : fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
1229 1932 : !(a.end <= b.start || b.end <= a.start)
1230 1932 : }
1231 1932 : for i in 0..ranges.len() {
1232 1932 : for j in 0..i {
1233 1932 : if overlaps_with(&ranges[i], &ranges[j]) {
1234 0 : panic!(
1235 0 : "overlapping sparse keyspace: {}..{} and {}..{}",
1236 0 : ranges[i].start, ranges[i].end, ranges[j].start, ranges[j].end
1237 0 : );
1238 1932 : }
1239 : }
1240 : }
1241 1288 : for i in 1..ranges.len() {
1242 1288 : assert!(
1243 1288 : ranges[i - 1].end <= ranges[i].start,
1244 0 : "unordered sparse keyspace: {}..{} and {}..{}",
1245 0 : ranges[i - 1].start,
1246 0 : ranges[i - 1].end,
1247 0 : ranges[i].start,
1248 0 : ranges[i].end
1249 : );
1250 : }
1251 0 : }
1252 :
1253 644 : Ok((dense_keyspace, sparse_keyspace))
1254 644 : }
1255 :
1256 : /// Get cached size of relation if it not updated after specified LSN
1257 897080 : pub fn get_cached_rel_size(&self, tag: &RelTag, lsn: Lsn) -> Option<BlockNumber> {
1258 897080 : let rel_size_cache = self.rel_size_cache.read().unwrap();
1259 897080 : if let Some((cached_lsn, nblocks)) = rel_size_cache.map.get(tag) {
1260 897036 : if lsn >= *cached_lsn {
1261 886744 : RELSIZE_CACHE_HITS.inc();
1262 886744 : return Some(*nblocks);
1263 10292 : }
1264 10292 : RELSIZE_CACHE_MISSES_OLD.inc();
1265 44 : }
1266 10336 : RELSIZE_CACHE_MISSES.inc();
1267 10336 : None
1268 897080 : }
1269 :
1270 : /// Update cached relation size if there is no more recent update
1271 10272 : pub fn update_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
1272 10272 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1273 10272 :
1274 10272 : if lsn < rel_size_cache.complete_as_of {
1275 : // Do not cache old values. It's safe to cache the size on read, as long as
1276 : // the read was at an LSN since we started the WAL ingestion. Reasoning: we
1277 : // never evict values from the cache, so if the relation size changed after
1278 : // 'lsn', the new value is already in the cache.
1279 0 : return;
1280 10272 : }
1281 10272 :
1282 10272 : match rel_size_cache.map.entry(tag) {
1283 10272 : hash_map::Entry::Occupied(mut entry) => {
1284 10272 : let cached_lsn = entry.get_mut();
1285 10272 : if lsn >= cached_lsn.0 {
1286 0 : *cached_lsn = (lsn, nblocks);
1287 10272 : }
1288 : }
1289 0 : hash_map::Entry::Vacant(entry) => {
1290 0 : entry.insert((lsn, nblocks));
1291 0 : RELSIZE_CACHE_ENTRIES.inc();
1292 0 : }
1293 : }
1294 10272 : }
1295 :
1296 : /// Store cached relation size
1297 565440 : pub fn set_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
1298 565440 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1299 565440 : if rel_size_cache.map.insert(tag, (lsn, nblocks)).is_none() {
1300 3840 : RELSIZE_CACHE_ENTRIES.inc();
1301 561600 : }
1302 565440 : }
1303 :
1304 : /// Remove cached relation size
1305 4 : pub fn remove_cached_rel_size(&self, tag: &RelTag) {
1306 4 : let mut rel_size_cache = self.rel_size_cache.write().unwrap();
1307 4 : if rel_size_cache.map.remove(tag).is_some() {
1308 4 : RELSIZE_CACHE_ENTRIES.dec();
1309 4 : }
1310 4 : }
1311 : }
1312 :
1313 : /// DatadirModification represents an operation to ingest an atomic set of
1314 : /// updates to the repository.
1315 : ///
1316 : /// It is created by the 'begin_record' function. It is called for each WAL
1317 : /// record, so that all the modifications by a one WAL record appear atomic.
1318 : pub struct DatadirModification<'a> {
1319 : /// The timeline this modification applies to. You can access this to
1320 : /// read the state, but note that any pending updates are *not* reflected
1321 : /// in the state in 'tline' yet.
1322 : pub tline: &'a Timeline,
1323 :
1324 : /// Current LSN of the modification
1325 : lsn: Lsn,
1326 :
1327 : // The modifications are not applied directly to the underlying key-value store.
1328 : // The put-functions add the modifications here, and they are flushed to the
1329 : // underlying key-value store by the 'finish' function.
1330 : pending_lsns: Vec<Lsn>,
1331 : pending_deletions: Vec<(Range<Key>, Lsn)>,
1332 : pending_nblocks: i64,
1333 :
1334 : /// Metadata writes, indexed by key so that they can be read from not-yet-committed modifications
1335 : /// while ingesting subsequent records. See [`Self::is_data_key`] for the definition of 'metadata'.
1336 : pending_metadata_pages: HashMap<CompactKey, Vec<(Lsn, usize, Value)>>,
1337 :
1338 : /// Data writes, ready to be flushed into an ephemeral layer. See [`Self::is_data_key`] for
1339 : /// which keys are stored here.
1340 : pending_data_batch: Option<SerializedValueBatch>,
1341 :
1342 : /// For special "directory" keys that store key-value maps, track the size of the map
1343 : /// if it was updated in this modification.
1344 : pending_directory_entries: Vec<(DirectoryKind, MetricsUpdate)>,
1345 :
1346 : /// An **approximation** of how many metadata bytes will be written to the EphemeralFile.
1347 : pending_metadata_bytes: usize,
1348 : }
1349 :
1350 : #[derive(Debug, Clone, Copy, PartialEq, Eq)]
1351 : pub enum MetricsUpdate {
1352 : /// Set the metrics to this value
1353 : Set(u64),
1354 : /// Increment the metrics by this value
1355 : Add(u64),
1356 : /// Decrement the metrics by this value
1357 : Sub(u64),
1358 : }
1359 :
1360 : impl DatadirModification<'_> {
1361 : // When a DatadirModification is committed, we do a monolithic serialization of all its contents. WAL records can
1362 : // contain multiple pages, so the pageserver's record-based batch size isn't sufficient to bound this allocation: we
1363 : // additionally specify a limit on how much payload a DatadirModification may contain before it should be committed.
1364 : pub(crate) const MAX_PENDING_BYTES: usize = 8 * 1024 * 1024;
1365 :
1366 : /// Get the current lsn
1367 836116 : pub(crate) fn get_lsn(&self) -> Lsn {
1368 836116 : self.lsn
1369 836116 : }
1370 :
1371 0 : pub(crate) fn approx_pending_bytes(&self) -> usize {
1372 0 : self.pending_data_batch
1373 0 : .as_ref()
1374 0 : .map_or(0, |b| b.buffer_size())
1375 0 : + self.pending_metadata_bytes
1376 0 : }
1377 :
1378 0 : pub(crate) fn has_dirty_data(&self) -> bool {
1379 0 : self.pending_data_batch
1380 0 : .as_ref()
1381 0 : .is_some_and(|b| b.has_data())
1382 0 : }
1383 :
1384 : /// Returns statistics about the currently pending modifications.
1385 0 : pub(crate) fn stats(&self) -> DatadirModificationStats {
1386 0 : let mut stats = DatadirModificationStats::default();
1387 0 : for (_, _, value) in self.pending_metadata_pages.values().flatten() {
1388 0 : match value {
1389 0 : Value::Image(_) => stats.metadata_images += 1,
1390 0 : Value::WalRecord(r) if r.will_init() => stats.metadata_images += 1,
1391 0 : Value::WalRecord(_) => stats.metadata_deltas += 1,
1392 : }
1393 : }
1394 0 : for valuemeta in self.pending_data_batch.iter().flat_map(|b| &b.metadata) {
1395 0 : match valuemeta {
1396 0 : ValueMeta::Serialized(s) if s.will_init => stats.data_images += 1,
1397 0 : ValueMeta::Serialized(_) => stats.data_deltas += 1,
1398 0 : ValueMeta::Observed(_) => {}
1399 : }
1400 : }
1401 0 : stats
1402 0 : }
1403 :
1404 : /// Set the current lsn
1405 291716 : pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> anyhow::Result<()> {
1406 291716 : ensure!(
1407 291716 : lsn >= self.lsn,
1408 0 : "setting an older lsn {} than {} is not allowed",
1409 : lsn,
1410 : self.lsn
1411 : );
1412 :
1413 291716 : if lsn > self.lsn {
1414 291716 : self.pending_lsns.push(self.lsn);
1415 291716 : self.lsn = lsn;
1416 291716 : }
1417 291716 : Ok(())
1418 291716 : }
1419 :
1420 : /// In this context, 'metadata' means keys that are only read by the pageserver internally, and 'data' means
1421 : /// keys that represent literal blocks that postgres can read. So data includes relation blocks and
1422 : /// SLRU blocks, which are read directly by postgres, and everything else is considered metadata.
1423 : ///
1424 : /// The distinction is important because data keys are handled on a fast path where dirty writes are
1425 : /// not readable until this modification is committed, whereas metadata keys are visible for read
1426 : /// via [`Self::get`] as soon as their record has been ingested.
1427 1701248 : fn is_data_key(key: &Key) -> bool {
1428 1701248 : key.is_rel_block_key() || key.is_slru_block_key()
1429 1701248 : }
1430 :
1431 : /// Initialize a completely new repository.
1432 : ///
1433 : /// This inserts the directory metadata entries that are assumed to
1434 : /// always exist.
1435 416 : pub fn init_empty(&mut self) -> anyhow::Result<()> {
1436 416 : let buf = DbDirectory::ser(&DbDirectory {
1437 416 : dbdirs: HashMap::new(),
1438 416 : })?;
1439 416 : self.pending_directory_entries
1440 416 : .push((DirectoryKind::Db, MetricsUpdate::Set(0)));
1441 416 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1442 :
1443 416 : let buf = if self.tline.pg_version >= 17 {
1444 0 : TwoPhaseDirectoryV17::ser(&TwoPhaseDirectoryV17 {
1445 0 : xids: HashSet::new(),
1446 0 : })
1447 : } else {
1448 416 : TwoPhaseDirectory::ser(&TwoPhaseDirectory {
1449 416 : xids: HashSet::new(),
1450 416 : })
1451 0 : }?;
1452 416 : self.pending_directory_entries
1453 416 : .push((DirectoryKind::TwoPhase, MetricsUpdate::Set(0)));
1454 416 : self.put(TWOPHASEDIR_KEY, Value::Image(buf.into()));
1455 :
1456 416 : let buf: Bytes = SlruSegmentDirectory::ser(&SlruSegmentDirectory::default())?.into();
1457 416 : let empty_dir = Value::Image(buf);
1458 416 :
1459 416 : // Initialize SLRUs on shard 0 only: creating these on other shards would be
1460 416 : // harmless but they'd just be dropped on later compaction.
1461 416 : if self.tline.tenant_shard_id.is_shard_zero() {
1462 404 : self.put(slru_dir_to_key(SlruKind::Clog), empty_dir.clone());
1463 404 : self.pending_directory_entries.push((
1464 404 : DirectoryKind::SlruSegment(SlruKind::Clog),
1465 404 : MetricsUpdate::Set(0),
1466 404 : ));
1467 404 : self.put(
1468 404 : slru_dir_to_key(SlruKind::MultiXactMembers),
1469 404 : empty_dir.clone(),
1470 404 : );
1471 404 : self.pending_directory_entries.push((
1472 404 : DirectoryKind::SlruSegment(SlruKind::Clog),
1473 404 : MetricsUpdate::Set(0),
1474 404 : ));
1475 404 : self.put(slru_dir_to_key(SlruKind::MultiXactOffsets), empty_dir);
1476 404 : self.pending_directory_entries.push((
1477 404 : DirectoryKind::SlruSegment(SlruKind::MultiXactOffsets),
1478 404 : MetricsUpdate::Set(0),
1479 404 : ));
1480 404 : }
1481 :
1482 416 : Ok(())
1483 416 : }
1484 :
1485 : #[cfg(test)]
1486 412 : pub fn init_empty_test_timeline(&mut self) -> anyhow::Result<()> {
1487 412 : self.init_empty()?;
1488 412 : self.put_control_file(bytes::Bytes::from_static(
1489 412 : b"control_file contents do not matter",
1490 412 : ))
1491 412 : .context("put_control_file")?;
1492 412 : self.put_checkpoint(bytes::Bytes::from_static(
1493 412 : b"checkpoint_file contents do not matter",
1494 412 : ))
1495 412 : .context("put_checkpoint_file")?;
1496 412 : Ok(())
1497 412 : }
1498 :
1499 : /// Creates a relation if it is not already present.
1500 : /// Returns the current size of the relation
1501 836112 : pub(crate) async fn create_relation_if_required(
1502 836112 : &mut self,
1503 836112 : rel: RelTag,
1504 836112 : ctx: &RequestContext,
1505 836112 : ) -> Result<u32, PageReconstructError> {
1506 : // Get current size and put rel creation if rel doesn't exist
1507 : //
1508 : // NOTE: we check the cache first even though get_rel_exists and get_rel_size would
1509 : // check the cache too. This is because eagerly checking the cache results in
1510 : // less work overall and 10% better performance. It's more work on cache miss
1511 : // but cache miss is rare.
1512 836112 : if let Some(nblocks) = self.tline.get_cached_rel_size(&rel, self.get_lsn()) {
1513 836092 : Ok(nblocks)
1514 20 : } else if !self
1515 20 : .tline
1516 20 : .get_rel_exists(rel, Version::Modified(self), ctx)
1517 20 : .await?
1518 : {
1519 : // create it with 0 size initially, the logic below will extend it
1520 20 : self.put_rel_creation(rel, 0, ctx)
1521 20 : .await
1522 20 : .context("Relation Error")?;
1523 20 : Ok(0)
1524 : } else {
1525 0 : self.tline
1526 0 : .get_rel_size(rel, Version::Modified(self), ctx)
1527 0 : .await
1528 : }
1529 836112 : }
1530 :
1531 : /// Given a block number for a relation (which represents a newly written block),
1532 : /// the previous block count of the relation, and the shard info, find the gaps
1533 : /// that were created by the newly written block if any.
1534 291340 : fn find_gaps(
1535 291340 : rel: RelTag,
1536 291340 : blkno: u32,
1537 291340 : previous_nblocks: u32,
1538 291340 : shard: &ShardIdentity,
1539 291340 : ) -> Option<KeySpace> {
1540 291340 : let mut key = rel_block_to_key(rel, blkno);
1541 291340 : let mut gap_accum = None;
1542 :
1543 291340 : for gap_blkno in previous_nblocks..blkno {
1544 64 : key.field6 = gap_blkno;
1545 64 :
1546 64 : if shard.get_shard_number(&key) != shard.number {
1547 16 : continue;
1548 48 : }
1549 48 :
1550 48 : gap_accum
1551 48 : .get_or_insert_with(KeySpaceAccum::new)
1552 48 : .add_key(key);
1553 : }
1554 :
1555 291340 : gap_accum.map(|accum| accum.to_keyspace())
1556 291340 : }
1557 :
1558 291704 : pub async fn ingest_batch(
1559 291704 : &mut self,
1560 291704 : mut batch: SerializedValueBatch,
1561 291704 : // TODO(vlad): remove this argument and replace the shard check with is_key_local
1562 291704 : shard: &ShardIdentity,
1563 291704 : ctx: &RequestContext,
1564 291704 : ) -> anyhow::Result<()> {
1565 291704 : let mut gaps_at_lsns = Vec::default();
1566 :
1567 291704 : for meta in batch.metadata.iter() {
1568 291284 : let (rel, blkno) = Key::from_compact(meta.key()).to_rel_block()?;
1569 291284 : let new_nblocks = blkno + 1;
1570 :
1571 291284 : let old_nblocks = self.create_relation_if_required(rel, ctx).await?;
1572 291284 : if new_nblocks > old_nblocks {
1573 4780 : self.put_rel_extend(rel, new_nblocks, ctx).await?;
1574 286504 : }
1575 :
1576 291284 : if let Some(gaps) = Self::find_gaps(rel, blkno, old_nblocks, shard) {
1577 0 : gaps_at_lsns.push((gaps, meta.lsn()));
1578 291284 : }
1579 : }
1580 :
1581 291704 : if !gaps_at_lsns.is_empty() {
1582 0 : batch.zero_gaps(gaps_at_lsns);
1583 291704 : }
1584 :
1585 291704 : match self.pending_data_batch.as_mut() {
1586 40 : Some(pending_batch) => {
1587 40 : pending_batch.extend(batch);
1588 40 : }
1589 291664 : None if batch.has_data() => {
1590 291260 : self.pending_data_batch = Some(batch);
1591 291260 : }
1592 404 : None => {
1593 404 : // Nothing to initialize the batch with
1594 404 : }
1595 : }
1596 :
1597 291704 : Ok(())
1598 291704 : }
1599 :
1600 : /// Put a new page version that can be constructed from a WAL record
1601 : ///
1602 : /// NOTE: this will *not* implicitly extend the relation, if the page is beyond the
1603 : /// current end-of-file. It's up to the caller to check that the relation size
1604 : /// matches the blocks inserted!
1605 24 : pub fn put_rel_wal_record(
1606 24 : &mut self,
1607 24 : rel: RelTag,
1608 24 : blknum: BlockNumber,
1609 24 : rec: NeonWalRecord,
1610 24 : ) -> anyhow::Result<()> {
1611 24 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1612 24 : self.put(rel_block_to_key(rel, blknum), Value::WalRecord(rec));
1613 24 : Ok(())
1614 24 : }
1615 :
1616 : // Same, but for an SLRU.
1617 16 : pub fn put_slru_wal_record(
1618 16 : &mut self,
1619 16 : kind: SlruKind,
1620 16 : segno: u32,
1621 16 : blknum: BlockNumber,
1622 16 : rec: NeonWalRecord,
1623 16 : ) -> anyhow::Result<()> {
1624 16 : if !self.tline.tenant_shard_id.is_shard_zero() {
1625 0 : return Ok(());
1626 16 : }
1627 16 :
1628 16 : self.put(
1629 16 : slru_block_to_key(kind, segno, blknum),
1630 16 : Value::WalRecord(rec),
1631 16 : );
1632 16 : Ok(())
1633 16 : }
1634 :
1635 : /// Like put_wal_record, but with ready-made image of the page.
1636 555684 : pub fn put_rel_page_image(
1637 555684 : &mut self,
1638 555684 : rel: RelTag,
1639 555684 : blknum: BlockNumber,
1640 555684 : img: Bytes,
1641 555684 : ) -> anyhow::Result<()> {
1642 555684 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1643 555684 : let key = rel_block_to_key(rel, blknum);
1644 555684 : if !key.is_valid_key_on_write_path() {
1645 0 : anyhow::bail!(
1646 0 : "the request contains data not supported by pageserver at {}",
1647 0 : key
1648 0 : );
1649 555684 : }
1650 555684 : self.put(rel_block_to_key(rel, blknum), Value::Image(img));
1651 555684 : Ok(())
1652 555684 : }
1653 :
1654 12 : pub fn put_slru_page_image(
1655 12 : &mut self,
1656 12 : kind: SlruKind,
1657 12 : segno: u32,
1658 12 : blknum: BlockNumber,
1659 12 : img: Bytes,
1660 12 : ) -> anyhow::Result<()> {
1661 12 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1662 :
1663 12 : let key = slru_block_to_key(kind, segno, blknum);
1664 12 : if !key.is_valid_key_on_write_path() {
1665 0 : anyhow::bail!(
1666 0 : "the request contains data not supported by pageserver at {}",
1667 0 : key
1668 0 : );
1669 12 : }
1670 12 : self.put(key, Value::Image(img));
1671 12 : Ok(())
1672 12 : }
1673 :
1674 5996 : pub(crate) fn put_rel_page_image_zero(
1675 5996 : &mut self,
1676 5996 : rel: RelTag,
1677 5996 : blknum: BlockNumber,
1678 5996 : ) -> anyhow::Result<()> {
1679 5996 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1680 5996 : let key = rel_block_to_key(rel, blknum);
1681 5996 : if !key.is_valid_key_on_write_path() {
1682 0 : anyhow::bail!(
1683 0 : "the request contains data not supported by pageserver: {} @ {}",
1684 0 : key,
1685 0 : self.lsn
1686 0 : );
1687 5996 : }
1688 5996 :
1689 5996 : let batch = self
1690 5996 : .pending_data_batch
1691 5996 : .get_or_insert_with(SerializedValueBatch::default);
1692 5996 :
1693 5996 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1694 5996 :
1695 5996 : Ok(())
1696 5996 : }
1697 :
1698 0 : pub(crate) fn put_slru_page_image_zero(
1699 0 : &mut self,
1700 0 : kind: SlruKind,
1701 0 : segno: u32,
1702 0 : blknum: BlockNumber,
1703 0 : ) -> anyhow::Result<()> {
1704 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1705 0 : let key = slru_block_to_key(kind, segno, blknum);
1706 0 : if !key.is_valid_key_on_write_path() {
1707 0 : anyhow::bail!(
1708 0 : "the request contains data not supported by pageserver: {} @ {}",
1709 0 : key,
1710 0 : self.lsn
1711 0 : );
1712 0 : }
1713 0 :
1714 0 : let batch = self
1715 0 : .pending_data_batch
1716 0 : .get_or_insert_with(SerializedValueBatch::default);
1717 0 :
1718 0 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1719 0 :
1720 0 : Ok(())
1721 0 : }
1722 :
1723 : /// Store a relmapper file (pg_filenode.map) in the repository
1724 32 : pub async fn put_relmap_file(
1725 32 : &mut self,
1726 32 : spcnode: Oid,
1727 32 : dbnode: Oid,
1728 32 : img: Bytes,
1729 32 : ctx: &RequestContext,
1730 32 : ) -> anyhow::Result<()> {
1731 : // Add it to the directory (if it doesn't exist already)
1732 32 : let buf = self.get(DBDIR_KEY, ctx).await?;
1733 32 : let mut dbdir = DbDirectory::des(&buf)?;
1734 :
1735 32 : let r = dbdir.dbdirs.insert((spcnode, dbnode), true);
1736 32 : if r.is_none() || r == Some(false) {
1737 : // The dbdir entry didn't exist, or it contained a
1738 : // 'false'. The 'insert' call already updated it with
1739 : // 'true', now write the updated 'dbdirs' map back.
1740 32 : let buf = DbDirectory::ser(&dbdir)?;
1741 32 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1742 0 : }
1743 32 : if r.is_none() {
1744 : // Create RelDirectory
1745 : // TODO: if we have fully migrated to v2, no need to create this directory
1746 16 : let buf = RelDirectory::ser(&RelDirectory {
1747 16 : rels: HashSet::new(),
1748 16 : })?;
1749 16 : self.pending_directory_entries
1750 16 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
1751 16 : if self.tline.get_rel_size_v2_enabled() {
1752 0 : self.pending_directory_entries
1753 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
1754 16 : }
1755 16 : self.put(
1756 16 : rel_dir_to_key(spcnode, dbnode),
1757 16 : Value::Image(Bytes::from(buf)),
1758 16 : );
1759 16 : }
1760 :
1761 32 : self.put(relmap_file_key(spcnode, dbnode), Value::Image(img));
1762 32 : Ok(())
1763 32 : }
1764 :
1765 0 : pub async fn put_twophase_file(
1766 0 : &mut self,
1767 0 : xid: u64,
1768 0 : img: Bytes,
1769 0 : ctx: &RequestContext,
1770 0 : ) -> anyhow::Result<()> {
1771 : // Add it to the directory entry
1772 0 : let dirbuf = self.get(TWOPHASEDIR_KEY, ctx).await?;
1773 0 : let newdirbuf = if self.tline.pg_version >= 17 {
1774 0 : let mut dir = TwoPhaseDirectoryV17::des(&dirbuf)?;
1775 0 : if !dir.xids.insert(xid) {
1776 0 : anyhow::bail!("twophase file for xid {} already exists", xid);
1777 0 : }
1778 0 : self.pending_directory_entries.push((
1779 0 : DirectoryKind::TwoPhase,
1780 0 : MetricsUpdate::Set(dir.xids.len() as u64),
1781 0 : ));
1782 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
1783 : } else {
1784 0 : let xid = xid as u32;
1785 0 : let mut dir = TwoPhaseDirectory::des(&dirbuf)?;
1786 0 : if !dir.xids.insert(xid) {
1787 0 : anyhow::bail!("twophase file for xid {} already exists", xid);
1788 0 : }
1789 0 : self.pending_directory_entries.push((
1790 0 : DirectoryKind::TwoPhase,
1791 0 : MetricsUpdate::Set(dir.xids.len() as u64),
1792 0 : ));
1793 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
1794 : };
1795 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
1796 0 :
1797 0 : self.put(twophase_file_key(xid), Value::Image(img));
1798 0 : Ok(())
1799 0 : }
1800 :
1801 0 : pub async fn set_replorigin(
1802 0 : &mut self,
1803 0 : origin_id: RepOriginId,
1804 0 : origin_lsn: Lsn,
1805 0 : ) -> anyhow::Result<()> {
1806 0 : let key = repl_origin_key(origin_id);
1807 0 : self.put(key, Value::Image(origin_lsn.ser().unwrap().into()));
1808 0 : Ok(())
1809 0 : }
1810 :
1811 0 : pub async fn drop_replorigin(&mut self, origin_id: RepOriginId) -> anyhow::Result<()> {
1812 0 : self.set_replorigin(origin_id, Lsn::INVALID).await
1813 0 : }
1814 :
1815 416 : pub fn put_control_file(&mut self, img: Bytes) -> anyhow::Result<()> {
1816 416 : self.put(CONTROLFILE_KEY, Value::Image(img));
1817 416 : Ok(())
1818 416 : }
1819 :
1820 444 : pub fn put_checkpoint(&mut self, img: Bytes) -> anyhow::Result<()> {
1821 444 : self.put(CHECKPOINT_KEY, Value::Image(img));
1822 444 : Ok(())
1823 444 : }
1824 :
1825 0 : pub async fn drop_dbdir(
1826 0 : &mut self,
1827 0 : spcnode: Oid,
1828 0 : dbnode: Oid,
1829 0 : ctx: &RequestContext,
1830 0 : ) -> anyhow::Result<()> {
1831 0 : let total_blocks = self
1832 0 : .tline
1833 0 : .get_db_size(spcnode, dbnode, Version::Modified(self), ctx)
1834 0 : .await?;
1835 :
1836 : // Remove entry from dbdir
1837 0 : let buf = self.get(DBDIR_KEY, ctx).await?;
1838 0 : let mut dir = DbDirectory::des(&buf)?;
1839 0 : if dir.dbdirs.remove(&(spcnode, dbnode)).is_some() {
1840 0 : let buf = DbDirectory::ser(&dir)?;
1841 0 : self.pending_directory_entries.push((
1842 0 : DirectoryKind::Db,
1843 0 : MetricsUpdate::Set(dir.dbdirs.len() as u64),
1844 0 : ));
1845 0 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1846 : } else {
1847 0 : warn!(
1848 0 : "dropped dbdir for spcnode {} dbnode {} did not exist in db directory",
1849 : spcnode, dbnode
1850 : );
1851 : }
1852 :
1853 : // Update logical database size.
1854 0 : self.pending_nblocks -= total_blocks as i64;
1855 0 :
1856 0 : // Delete all relations and metadata files for the spcnode/dnode
1857 0 : self.delete(dbdir_key_range(spcnode, dbnode));
1858 0 : Ok(())
1859 0 : }
1860 :
1861 : /// Create a relation fork.
1862 : ///
1863 : /// 'nblocks' is the initial size.
1864 3840 : pub async fn put_rel_creation(
1865 3840 : &mut self,
1866 3840 : rel: RelTag,
1867 3840 : nblocks: BlockNumber,
1868 3840 : ctx: &RequestContext,
1869 3840 : ) -> Result<(), RelationError> {
1870 3840 : if rel.relnode == 0 {
1871 0 : return Err(RelationError::InvalidRelnode);
1872 3840 : }
1873 : // It's possible that this is the first rel for this db in this
1874 : // tablespace. Create the reldir entry for it if so.
1875 3840 : let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await.context("read db")?)
1876 3840 : .context("deserialize db")?;
1877 :
1878 3840 : let dbdir_exists =
1879 3840 : if let hash_map::Entry::Vacant(e) = dbdir.dbdirs.entry((rel.spcnode, rel.dbnode)) {
1880 : // Didn't exist. Update dbdir
1881 16 : e.insert(false);
1882 16 : let buf = DbDirectory::ser(&dbdir).context("serialize db")?;
1883 16 : self.pending_directory_entries.push((
1884 16 : DirectoryKind::Db,
1885 16 : MetricsUpdate::Set(dbdir.dbdirs.len() as u64),
1886 16 : ));
1887 16 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1888 16 : false
1889 : } else {
1890 3824 : true
1891 : };
1892 :
1893 3840 : let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
1894 3840 : let mut rel_dir = if !dbdir_exists {
1895 : // Create the RelDirectory
1896 16 : RelDirectory::default()
1897 : } else {
1898 : // reldir already exists, fetch it
1899 3824 : RelDirectory::des(&self.get(rel_dir_key, ctx).await.context("read db")?)
1900 3824 : .context("deserialize db")?
1901 : };
1902 :
1903 : // Add the new relation to the rel directory entry, and write it back
1904 3840 : if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
1905 0 : return Err(RelationError::AlreadyExists);
1906 3840 : }
1907 3840 :
1908 3840 : if self.tline.get_rel_size_v2_enabled() {
1909 0 : let sparse_rel_dir_key =
1910 0 : rel_tag_sparse_key(rel.spcnode, rel.dbnode, rel.relnode, rel.forknum);
1911 : // check if the rel_dir_key exists in v2
1912 0 : let val = self
1913 0 : .sparse_get(sparse_rel_dir_key, ctx)
1914 0 : .await
1915 0 : .map_err(|e| RelationError::Other(e.into()))?;
1916 0 : let val = RelDirExists::decode_option(val)
1917 0 : .map_err(|_| RelationError::Other(anyhow::anyhow!("invalid reldir key")))?;
1918 0 : if val == RelDirExists::Exists {
1919 0 : return Err(RelationError::AlreadyExists);
1920 0 : }
1921 0 : self.put(
1922 0 : sparse_rel_dir_key,
1923 0 : Value::Image(RelDirExists::Exists.encode()),
1924 0 : );
1925 0 : if !dbdir_exists {
1926 0 : self.pending_directory_entries
1927 0 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
1928 0 : self.pending_directory_entries
1929 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
1930 0 : // We don't write `rel_dir_key -> rel_dir.rels` back to the storage in the v2 path unless it's the initial creation.
1931 0 : // TODO: if we have fully migrated to v2, no need to create this directory. Otherwise, there
1932 0 : // will be key not found errors if we don't create an empty one for rel_size_v2.
1933 0 : self.put(
1934 0 : rel_dir_key,
1935 0 : Value::Image(Bytes::from(
1936 0 : RelDirectory::ser(&RelDirectory::default()).context("serialize")?,
1937 : )),
1938 : );
1939 0 : }
1940 0 : self.pending_directory_entries
1941 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Add(1)));
1942 : } else {
1943 3840 : if !dbdir_exists {
1944 16 : self.pending_directory_entries
1945 16 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)))
1946 3824 : }
1947 3840 : self.pending_directory_entries
1948 3840 : .push((DirectoryKind::Rel, MetricsUpdate::Add(1)));
1949 3840 : self.put(
1950 3840 : rel_dir_key,
1951 3840 : Value::Image(Bytes::from(
1952 3840 : RelDirectory::ser(&rel_dir).context("serialize")?,
1953 : )),
1954 : );
1955 : }
1956 : // Put size
1957 3840 : let size_key = rel_size_to_key(rel);
1958 3840 : let buf = nblocks.to_le_bytes();
1959 3840 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1960 3840 :
1961 3840 : self.pending_nblocks += nblocks as i64;
1962 3840 :
1963 3840 : // Update relation size cache
1964 3840 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
1965 3840 :
1966 3840 : // Even if nblocks > 0, we don't insert any actual blocks here. That's up to the
1967 3840 : // caller.
1968 3840 : Ok(())
1969 3840 : }
1970 :
1971 : /// Truncate relation
1972 12024 : pub async fn put_rel_truncation(
1973 12024 : &mut self,
1974 12024 : rel: RelTag,
1975 12024 : nblocks: BlockNumber,
1976 12024 : ctx: &RequestContext,
1977 12024 : ) -> anyhow::Result<()> {
1978 12024 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
1979 12024 : if self
1980 12024 : .tline
1981 12024 : .get_rel_exists(rel, Version::Modified(self), ctx)
1982 12024 : .await?
1983 : {
1984 12024 : let size_key = rel_size_to_key(rel);
1985 : // Fetch the old size first
1986 12024 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
1987 12024 :
1988 12024 : // Update the entry with the new size.
1989 12024 : let buf = nblocks.to_le_bytes();
1990 12024 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
1991 12024 :
1992 12024 : // Update relation size cache
1993 12024 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
1994 12024 :
1995 12024 : // Update logical database size.
1996 12024 : self.pending_nblocks -= old_size as i64 - nblocks as i64;
1997 0 : }
1998 12024 : Ok(())
1999 12024 : }
2000 :
2001 : /// Extend relation
2002 : /// If new size is smaller, do nothing.
2003 553360 : pub async fn put_rel_extend(
2004 553360 : &mut self,
2005 553360 : rel: RelTag,
2006 553360 : nblocks: BlockNumber,
2007 553360 : ctx: &RequestContext,
2008 553360 : ) -> anyhow::Result<()> {
2009 553360 : anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
2010 :
2011 : // Put size
2012 553360 : let size_key = rel_size_to_key(rel);
2013 553360 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
2014 553360 :
2015 553360 : // only extend relation here. never decrease the size
2016 553360 : if nblocks > old_size {
2017 549576 : let buf = nblocks.to_le_bytes();
2018 549576 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2019 549576 :
2020 549576 : // Update relation size cache
2021 549576 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
2022 549576 :
2023 549576 : self.pending_nblocks += nblocks as i64 - old_size as i64;
2024 549576 : }
2025 553360 : Ok(())
2026 553360 : }
2027 :
2028 : /// Drop some relations
2029 20 : pub(crate) async fn put_rel_drops(
2030 20 : &mut self,
2031 20 : drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
2032 20 : ctx: &RequestContext,
2033 20 : ) -> anyhow::Result<()> {
2034 24 : for ((spc_node, db_node), rel_tags) in drop_relations {
2035 4 : let dir_key = rel_dir_to_key(spc_node, db_node);
2036 4 : let buf = self.get(dir_key, ctx).await?;
2037 4 : let mut dir = RelDirectory::des(&buf)?;
2038 :
2039 4 : let mut dirty = false;
2040 8 : for rel_tag in rel_tags {
2041 4 : let found = if dir.rels.remove(&(rel_tag.relnode, rel_tag.forknum)) {
2042 4 : self.pending_directory_entries
2043 4 : .push((DirectoryKind::Rel, MetricsUpdate::Sub(1)));
2044 4 : dirty = true;
2045 4 : true
2046 0 : } else if self.tline.get_rel_size_v2_enabled() {
2047 : // The rel is not found in the old reldir key, so we need to check the new sparse keyspace.
2048 : // Note that a relation can only exist in one of the two keyspaces (guaranteed by the ingestion
2049 : // logic).
2050 0 : let key =
2051 0 : rel_tag_sparse_key(spc_node, db_node, rel_tag.relnode, rel_tag.forknum);
2052 0 : let val = RelDirExists::decode_option(self.sparse_get(key, ctx).await?)
2053 0 : .map_err(|_| RelationError::Other(anyhow::anyhow!("invalid reldir key")))?;
2054 0 : if val == RelDirExists::Exists {
2055 0 : self.pending_directory_entries
2056 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Sub(1)));
2057 0 : // put tombstone
2058 0 : self.put(key, Value::Image(RelDirExists::Removed.encode()));
2059 0 : // no need to set dirty to true
2060 0 : true
2061 : } else {
2062 0 : false
2063 : }
2064 : } else {
2065 0 : false
2066 : };
2067 :
2068 4 : if found {
2069 : // update logical size
2070 4 : let size_key = rel_size_to_key(rel_tag);
2071 4 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
2072 4 : self.pending_nblocks -= old_size as i64;
2073 4 :
2074 4 : // Remove entry from relation size cache
2075 4 : self.tline.remove_cached_rel_size(&rel_tag);
2076 4 :
2077 4 : // Delete size entry, as well as all blocks
2078 4 : self.delete(rel_key_range(rel_tag));
2079 0 : }
2080 : }
2081 :
2082 4 : if dirty {
2083 4 : self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
2084 0 : }
2085 : }
2086 :
2087 20 : Ok(())
2088 20 : }
2089 :
2090 12 : pub async fn put_slru_segment_creation(
2091 12 : &mut self,
2092 12 : kind: SlruKind,
2093 12 : segno: u32,
2094 12 : nblocks: BlockNumber,
2095 12 : ctx: &RequestContext,
2096 12 : ) -> anyhow::Result<()> {
2097 12 : assert!(self.tline.tenant_shard_id.is_shard_zero());
2098 :
2099 : // Add it to the directory entry
2100 12 : let dir_key = slru_dir_to_key(kind);
2101 12 : let buf = self.get(dir_key, ctx).await?;
2102 12 : let mut dir = SlruSegmentDirectory::des(&buf)?;
2103 :
2104 12 : if !dir.segments.insert(segno) {
2105 0 : anyhow::bail!("slru segment {kind:?}/{segno} already exists");
2106 12 : }
2107 12 : self.pending_directory_entries.push((
2108 12 : DirectoryKind::SlruSegment(kind),
2109 12 : MetricsUpdate::Set(dir.segments.len() as u64),
2110 12 : ));
2111 12 : self.put(
2112 12 : dir_key,
2113 12 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
2114 : );
2115 :
2116 : // Put size
2117 12 : let size_key = slru_segment_size_to_key(kind, segno);
2118 12 : let buf = nblocks.to_le_bytes();
2119 12 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2120 12 :
2121 12 : // even if nblocks > 0, we don't insert any actual blocks here
2122 12 :
2123 12 : Ok(())
2124 12 : }
2125 :
2126 : /// Extend SLRU segment
2127 0 : pub fn put_slru_extend(
2128 0 : &mut self,
2129 0 : kind: SlruKind,
2130 0 : segno: u32,
2131 0 : nblocks: BlockNumber,
2132 0 : ) -> anyhow::Result<()> {
2133 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
2134 :
2135 : // Put size
2136 0 : let size_key = slru_segment_size_to_key(kind, segno);
2137 0 : let buf = nblocks.to_le_bytes();
2138 0 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2139 0 : Ok(())
2140 0 : }
2141 :
2142 : /// This method is used for marking truncated SLRU files
2143 0 : pub async fn drop_slru_segment(
2144 0 : &mut self,
2145 0 : kind: SlruKind,
2146 0 : segno: u32,
2147 0 : ctx: &RequestContext,
2148 0 : ) -> anyhow::Result<()> {
2149 0 : // Remove it from the directory entry
2150 0 : let dir_key = slru_dir_to_key(kind);
2151 0 : let buf = self.get(dir_key, ctx).await?;
2152 0 : let mut dir = SlruSegmentDirectory::des(&buf)?;
2153 :
2154 0 : if !dir.segments.remove(&segno) {
2155 0 : warn!("slru segment {:?}/{} does not exist", kind, segno);
2156 0 : }
2157 0 : self.pending_directory_entries.push((
2158 0 : DirectoryKind::SlruSegment(kind),
2159 0 : MetricsUpdate::Set(dir.segments.len() as u64),
2160 0 : ));
2161 0 : self.put(
2162 0 : dir_key,
2163 0 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
2164 : );
2165 :
2166 : // Delete size entry, as well as all blocks
2167 0 : self.delete(slru_segment_key_range(kind, segno));
2168 0 :
2169 0 : Ok(())
2170 0 : }
2171 :
2172 : /// Drop a relmapper file (pg_filenode.map)
2173 0 : pub fn drop_relmap_file(&mut self, _spcnode: Oid, _dbnode: Oid) -> anyhow::Result<()> {
2174 0 : // TODO
2175 0 : Ok(())
2176 0 : }
2177 :
2178 : /// This method is used for marking truncated SLRU files
2179 0 : pub async fn drop_twophase_file(
2180 0 : &mut self,
2181 0 : xid: u64,
2182 0 : ctx: &RequestContext,
2183 0 : ) -> anyhow::Result<()> {
2184 : // Remove it from the directory entry
2185 0 : let buf = self.get(TWOPHASEDIR_KEY, ctx).await?;
2186 0 : let newdirbuf = if self.tline.pg_version >= 17 {
2187 0 : let mut dir = TwoPhaseDirectoryV17::des(&buf)?;
2188 :
2189 0 : if !dir.xids.remove(&xid) {
2190 0 : warn!("twophase file for xid {} does not exist", xid);
2191 0 : }
2192 0 : self.pending_directory_entries.push((
2193 0 : DirectoryKind::TwoPhase,
2194 0 : MetricsUpdate::Set(dir.xids.len() as u64),
2195 0 : ));
2196 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
2197 : } else {
2198 0 : let xid: u32 = u32::try_from(xid)?;
2199 0 : let mut dir = TwoPhaseDirectory::des(&buf)?;
2200 :
2201 0 : if !dir.xids.remove(&xid) {
2202 0 : warn!("twophase file for xid {} does not exist", xid);
2203 0 : }
2204 0 : self.pending_directory_entries.push((
2205 0 : DirectoryKind::TwoPhase,
2206 0 : MetricsUpdate::Set(dir.xids.len() as u64),
2207 0 : ));
2208 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
2209 : };
2210 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
2211 0 :
2212 0 : // Delete it
2213 0 : self.delete(twophase_key_range(xid));
2214 0 :
2215 0 : Ok(())
2216 0 : }
2217 :
2218 32 : pub async fn put_file(
2219 32 : &mut self,
2220 32 : path: &str,
2221 32 : content: &[u8],
2222 32 : ctx: &RequestContext,
2223 32 : ) -> anyhow::Result<()> {
2224 32 : let key = aux_file::encode_aux_file_key(path);
2225 : // retrieve the key from the engine
2226 32 : let old_val = match self.get(key, ctx).await {
2227 8 : Ok(val) => Some(val),
2228 24 : Err(PageReconstructError::MissingKey(_)) => None,
2229 0 : Err(e) => return Err(e.into()),
2230 : };
2231 32 : let files: Vec<(&str, &[u8])> = if let Some(ref old_val) = old_val {
2232 8 : aux_file::decode_file_value(old_val)?
2233 : } else {
2234 24 : Vec::new()
2235 : };
2236 32 : let mut other_files = Vec::with_capacity(files.len());
2237 32 : let mut modifying_file = None;
2238 40 : for file @ (p, content) in files {
2239 8 : if path == p {
2240 8 : assert!(
2241 8 : modifying_file.is_none(),
2242 0 : "duplicated entries found for {}",
2243 : path
2244 : );
2245 8 : modifying_file = Some(content);
2246 0 : } else {
2247 0 : other_files.push(file);
2248 0 : }
2249 : }
2250 32 : let mut new_files = other_files;
2251 32 : match (modifying_file, content.is_empty()) {
2252 4 : (Some(old_content), false) => {
2253 4 : self.tline
2254 4 : .aux_file_size_estimator
2255 4 : .on_update(old_content.len(), content.len());
2256 4 : new_files.push((path, content));
2257 4 : }
2258 4 : (Some(old_content), true) => {
2259 4 : self.tline
2260 4 : .aux_file_size_estimator
2261 4 : .on_remove(old_content.len());
2262 4 : // not adding the file key to the final `new_files` vec.
2263 4 : }
2264 24 : (None, false) => {
2265 24 : self.tline.aux_file_size_estimator.on_add(content.len());
2266 24 : new_files.push((path, content));
2267 24 : }
2268 : // Compute may request delete of old version of pgstat AUX file if new one exceeds size limit.
2269 : // Compute doesn't know if previous version of this file exists or not, so
2270 : // attempt to delete non-existing file can cause this message.
2271 : // To avoid false alarms, log it as info rather than warning.
2272 0 : (None, true) if path.starts_with("pg_stat/") => {
2273 0 : info!("removing non-existing pg_stat file: {}", path)
2274 : }
2275 0 : (None, true) => warn!("removing non-existing aux file: {}", path),
2276 : }
2277 32 : let new_val = aux_file::encode_file_value(&new_files)?;
2278 32 : self.put(key, Value::Image(new_val.into()));
2279 32 :
2280 32 : Ok(())
2281 32 : }
2282 :
2283 : ///
2284 : /// Flush changes accumulated so far to the underlying repository.
2285 : ///
2286 : /// Usually, changes made in DatadirModification are atomic, but this allows
2287 : /// you to flush them to the underlying repository before the final `commit`.
2288 : /// That allows to free up the memory used to hold the pending changes.
2289 : ///
2290 : /// Currently only used during bulk import of a data directory. In that
2291 : /// context, breaking the atomicity is OK. If the import is interrupted, the
2292 : /// whole import fails and the timeline will be deleted anyway.
2293 : /// (Or to be precise, it will be left behind for debugging purposes and
2294 : /// ignored, see <https://github.com/neondatabase/neon/pull/1809>)
2295 : ///
2296 : /// Note: A consequence of flushing the pending operations is that they
2297 : /// won't be visible to subsequent operations until `commit`. The function
2298 : /// retains all the metadata, but data pages are flushed. That's again OK
2299 : /// for bulk import, where you are just loading data pages and won't try to
2300 : /// modify the same pages twice.
2301 3860 : pub(crate) async fn flush(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2302 3860 : // Unless we have accumulated a decent amount of changes, it's not worth it
2303 3860 : // to scan through the pending_updates list.
2304 3860 : let pending_nblocks = self.pending_nblocks;
2305 3860 : if pending_nblocks < 10000 {
2306 3860 : return Ok(());
2307 0 : }
2308 :
2309 0 : let mut writer = self.tline.writer().await;
2310 :
2311 : // Flush relation and SLRU data blocks, keep metadata.
2312 0 : if let Some(batch) = self.pending_data_batch.take() {
2313 0 : tracing::debug!(
2314 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2315 0 : batch.max_lsn,
2316 0 : self.tline.get_last_record_lsn()
2317 : );
2318 :
2319 : // This bails out on first error without modifying pending_updates.
2320 : // That's Ok, cf this function's doc comment.
2321 0 : writer.put_batch(batch, ctx).await?;
2322 0 : }
2323 :
2324 0 : if pending_nblocks != 0 {
2325 0 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2326 0 : self.pending_nblocks = 0;
2327 0 : }
2328 :
2329 0 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2330 0 : writer.update_directory_entries_count(kind, count);
2331 0 : }
2332 :
2333 0 : Ok(())
2334 3860 : }
2335 :
2336 : ///
2337 : /// Finish this atomic update, writing all the updated keys to the
2338 : /// underlying timeline.
2339 : /// All the modifications in this atomic update are stamped by the specified LSN.
2340 : ///
2341 1486188 : pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2342 1486188 : let mut writer = self.tline.writer().await;
2343 :
2344 1486188 : let pending_nblocks = self.pending_nblocks;
2345 1486188 : self.pending_nblocks = 0;
2346 :
2347 : // Ordering: the items in this batch do not need to be in any global order, but values for
2348 : // a particular Key must be in Lsn order relative to one another. InMemoryLayer relies on
2349 : // this to do efficient updates to its index. See [`wal_decoder::serialized_batch`] for
2350 : // more details.
2351 :
2352 1486188 : let metadata_batch = {
2353 1486188 : let pending_meta = self
2354 1486188 : .pending_metadata_pages
2355 1486188 : .drain()
2356 1486188 : .flat_map(|(key, values)| {
2357 548036 : values
2358 548036 : .into_iter()
2359 548036 : .map(move |(lsn, value_size, value)| (key, lsn, value_size, value))
2360 1486188 : })
2361 1486188 : .collect::<Vec<_>>();
2362 1486188 :
2363 1486188 : if pending_meta.is_empty() {
2364 944556 : None
2365 : } else {
2366 541632 : Some(SerializedValueBatch::from_values(pending_meta))
2367 : }
2368 : };
2369 :
2370 1486188 : let data_batch = self.pending_data_batch.take();
2371 :
2372 1486188 : let maybe_batch = match (data_batch, metadata_batch) {
2373 529112 : (Some(mut data), Some(metadata)) => {
2374 529112 : data.extend(metadata);
2375 529112 : Some(data)
2376 : }
2377 286524 : (Some(data), None) => Some(data),
2378 12520 : (None, Some(metadata)) => Some(metadata),
2379 658032 : (None, None) => None,
2380 : };
2381 :
2382 1486188 : if let Some(batch) = maybe_batch {
2383 828156 : tracing::debug!(
2384 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2385 0 : batch.max_lsn,
2386 0 : self.tline.get_last_record_lsn()
2387 : );
2388 :
2389 : // This bails out on first error without modifying pending_updates.
2390 : // That's Ok, cf this function's doc comment.
2391 828156 : writer.put_batch(batch, ctx).await?;
2392 658032 : }
2393 :
2394 1486188 : if !self.pending_deletions.is_empty() {
2395 4 : writer.delete_batch(&self.pending_deletions, ctx).await?;
2396 4 : self.pending_deletions.clear();
2397 1486184 : }
2398 :
2399 1486188 : self.pending_lsns.push(self.lsn);
2400 1777904 : for pending_lsn in self.pending_lsns.drain(..) {
2401 1777904 : // TODO(vlad): pretty sure the comment below is not valid anymore
2402 1777904 : // and we can call finish write with the latest LSN
2403 1777904 : //
2404 1777904 : // Ideally, we should be able to call writer.finish_write() only once
2405 1777904 : // with the highest LSN. However, the last_record_lsn variable in the
2406 1777904 : // timeline keeps track of the latest LSN and the immediate previous LSN
2407 1777904 : // so we need to record every LSN to not leave a gap between them.
2408 1777904 : writer.finish_write(pending_lsn);
2409 1777904 : }
2410 :
2411 1486188 : if pending_nblocks != 0 {
2412 541140 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2413 945048 : }
2414 :
2415 1486188 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2416 5948 : writer.update_directory_entries_count(kind, count);
2417 5948 : }
2418 :
2419 1486188 : self.pending_metadata_bytes = 0;
2420 1486188 :
2421 1486188 : Ok(())
2422 1486188 : }
2423 :
2424 583408 : pub(crate) fn len(&self) -> usize {
2425 583408 : self.pending_metadata_pages.len()
2426 583408 : + self.pending_data_batch.as_ref().map_or(0, |b| b.len())
2427 583408 : + self.pending_deletions.len()
2428 583408 : }
2429 :
2430 : /// Read a page from the Timeline we are writing to. For metadata pages, this passes through
2431 : /// a cache in Self, which makes writes earlier in this modification visible to WAL records later
2432 : /// in the modification.
2433 : ///
2434 : /// For data pages, reads pass directly to the owning Timeline: any ingest code which reads a data
2435 : /// page must ensure that the pages they read are already committed in Timeline, for example
2436 : /// DB create operations are always preceded by a call to commit(). This is special cased because
2437 : /// it's rare: all the 'normal' WAL operations will only read metadata pages such as relation sizes,
2438 : /// and not data pages.
2439 573172 : async fn get(&self, key: Key, ctx: &RequestContext) -> Result<Bytes, PageReconstructError> {
2440 573172 : if !Self::is_data_key(&key) {
2441 : // Have we already updated the same key? Read the latest pending updated
2442 : // version in that case.
2443 : //
2444 : // Note: we don't check pending_deletions. It is an error to request a
2445 : // value that has been removed, deletion only avoids leaking storage.
2446 573172 : if let Some(values) = self.pending_metadata_pages.get(&key.to_compact()) {
2447 31856 : if let Some((_, _, value)) = values.last() {
2448 31856 : return if let Value::Image(img) = value {
2449 31856 : Ok(img.clone())
2450 : } else {
2451 : // Currently, we never need to read back a WAL record that we
2452 : // inserted in the same "transaction". All the metadata updates
2453 : // work directly with Images, and we never need to read actual
2454 : // data pages. We could handle this if we had to, by calling
2455 : // the walredo manager, but let's keep it simple for now.
2456 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
2457 0 : "unexpected pending WAL record"
2458 0 : )))
2459 : };
2460 0 : }
2461 541316 : }
2462 : } else {
2463 : // This is an expensive check, so we only do it in debug mode. If reading a data key,
2464 : // this key should never be present in pending_data_pages. We ensure this by committing
2465 : // modifications before ingesting DB create operations, which are the only kind that reads
2466 : // data pages during ingest.
2467 0 : if cfg!(debug_assertions) {
2468 0 : assert!(
2469 0 : !self
2470 0 : .pending_data_batch
2471 0 : .as_ref()
2472 0 : .is_some_and(|b| b.updates_key(&key))
2473 0 : );
2474 0 : }
2475 : }
2476 :
2477 : // Metadata page cache miss, or we're reading a data page.
2478 541316 : let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
2479 541316 : self.tline.get(key, lsn, ctx).await
2480 573172 : }
2481 :
2482 : /// Get a key from the sparse keyspace. Automatically converts the missing key error
2483 : /// and the empty value into None.
2484 0 : async fn sparse_get(
2485 0 : &self,
2486 0 : key: Key,
2487 0 : ctx: &RequestContext,
2488 0 : ) -> Result<Option<Bytes>, PageReconstructError> {
2489 0 : let val = self.get(key, ctx).await;
2490 0 : match val {
2491 0 : Ok(val) if val.is_empty() => Ok(None),
2492 0 : Ok(val) => Ok(Some(val)),
2493 0 : Err(PageReconstructError::MissingKey(_)) => Ok(None),
2494 0 : Err(e) => Err(e),
2495 : }
2496 0 : }
2497 :
2498 1128076 : fn put(&mut self, key: Key, val: Value) {
2499 1128076 : if Self::is_data_key(&key) {
2500 555736 : self.put_data(key.to_compact(), val)
2501 : } else {
2502 572340 : self.put_metadata(key.to_compact(), val)
2503 : }
2504 1128076 : }
2505 :
2506 555736 : fn put_data(&mut self, key: CompactKey, val: Value) {
2507 555736 : let batch = self
2508 555736 : .pending_data_batch
2509 555736 : .get_or_insert_with(SerializedValueBatch::default);
2510 555736 : batch.put(key, val, self.lsn);
2511 555736 : }
2512 :
2513 572340 : fn put_metadata(&mut self, key: CompactKey, val: Value) {
2514 572340 : let values = self.pending_metadata_pages.entry(key).or_default();
2515 : // Replace the previous value if it exists at the same lsn
2516 572340 : if let Some((last_lsn, last_value_ser_size, last_value)) = values.last_mut() {
2517 24304 : if *last_lsn == self.lsn {
2518 : // Update the pending_metadata_bytes contribution from this entry, and update the serialized size in place
2519 24304 : self.pending_metadata_bytes -= *last_value_ser_size;
2520 24304 : *last_value_ser_size = val.serialized_size().unwrap() as usize;
2521 24304 : self.pending_metadata_bytes += *last_value_ser_size;
2522 24304 :
2523 24304 : // Use the latest value, this replaces any earlier write to the same (key,lsn), such as much
2524 24304 : // have been generated by synthesized zero page writes prior to the first real write to a page.
2525 24304 : *last_value = val;
2526 24304 : return;
2527 0 : }
2528 548036 : }
2529 :
2530 548036 : let val_serialized_size = val.serialized_size().unwrap() as usize;
2531 548036 : self.pending_metadata_bytes += val_serialized_size;
2532 548036 : values.push((self.lsn, val_serialized_size, val));
2533 548036 :
2534 548036 : if key == CHECKPOINT_KEY.to_compact() {
2535 444 : tracing::debug!("Checkpoint key added to pending with size {val_serialized_size}");
2536 547592 : }
2537 572340 : }
2538 :
2539 4 : fn delete(&mut self, key_range: Range<Key>) {
2540 4 : trace!("DELETE {}-{}", key_range.start, key_range.end);
2541 4 : self.pending_deletions.push((key_range, self.lsn));
2542 4 : }
2543 : }
2544 :
2545 : /// Statistics for a DatadirModification.
2546 : #[derive(Default)]
2547 : pub struct DatadirModificationStats {
2548 : pub metadata_images: u64,
2549 : pub metadata_deltas: u64,
2550 : pub data_images: u64,
2551 : pub data_deltas: u64,
2552 : }
2553 :
2554 : /// This struct facilitates accessing either a committed key from the timeline at a
2555 : /// specific LSN, or the latest uncommitted key from a pending modification.
2556 : ///
2557 : /// During WAL ingestion, the records from multiple LSNs may be batched in the same
2558 : /// modification before being flushed to the timeline. Hence, the routines in WalIngest
2559 : /// need to look up the keys in the modification first before looking them up in the
2560 : /// timeline to not miss the latest updates.
2561 : #[derive(Clone, Copy)]
2562 : pub enum Version<'a> {
2563 : Lsn(Lsn),
2564 : Modified(&'a DatadirModification<'a>),
2565 : }
2566 :
2567 : impl Version<'_> {
2568 10352 : async fn get(
2569 10352 : &self,
2570 10352 : timeline: &Timeline,
2571 10352 : key: Key,
2572 10352 : ctx: &RequestContext,
2573 10352 : ) -> Result<Bytes, PageReconstructError> {
2574 10352 : match self {
2575 10312 : Version::Lsn(lsn) => timeline.get(key, *lsn, ctx).await,
2576 40 : Version::Modified(modification) => modification.get(key, ctx).await,
2577 : }
2578 10352 : }
2579 :
2580 : /// Get a key from the sparse keyspace. Automatically converts the missing key error
2581 : /// and the empty value into None.
2582 0 : async fn sparse_get(
2583 0 : &self,
2584 0 : timeline: &Timeline,
2585 0 : key: Key,
2586 0 : ctx: &RequestContext,
2587 0 : ) -> Result<Option<Bytes>, PageReconstructError> {
2588 0 : let val = self.get(timeline, key, ctx).await;
2589 0 : match val {
2590 0 : Ok(val) if val.is_empty() => Ok(None),
2591 0 : Ok(val) => Ok(Some(val)),
2592 0 : Err(PageReconstructError::MissingKey(_)) => Ok(None),
2593 0 : Err(e) => Err(e),
2594 : }
2595 0 : }
2596 :
2597 71240 : fn get_lsn(&self) -> Lsn {
2598 71240 : match self {
2599 59148 : Version::Lsn(lsn) => *lsn,
2600 12092 : Version::Modified(modification) => modification.lsn,
2601 : }
2602 71240 : }
2603 : }
2604 :
2605 : //--- Metadata structs stored in key-value pairs in the repository.
2606 :
2607 0 : #[derive(Debug, Serialize, Deserialize)]
2608 : pub(crate) struct DbDirectory {
2609 : // (spcnode, dbnode) -> (do relmapper and PG_VERSION files exist)
2610 : pub(crate) dbdirs: HashMap<(Oid, Oid), bool>,
2611 : }
2612 :
2613 : // The format of TwoPhaseDirectory changed in PostgreSQL v17, because the filenames of
2614 : // pg_twophase files was expanded from 32-bit XIDs to 64-bit XIDs. Previously, the files
2615 : // were named like "pg_twophase/000002E5", now they're like
2616 : // "pg_twophsae/0000000A000002E4".
2617 :
2618 0 : #[derive(Debug, Serialize, Deserialize)]
2619 : pub(crate) struct TwoPhaseDirectory {
2620 : pub(crate) xids: HashSet<TransactionId>,
2621 : }
2622 :
2623 0 : #[derive(Debug, Serialize, Deserialize)]
2624 : struct TwoPhaseDirectoryV17 {
2625 : xids: HashSet<u64>,
2626 : }
2627 :
2628 0 : #[derive(Debug, Serialize, Deserialize, Default)]
2629 : pub(crate) struct RelDirectory {
2630 : // Set of relations that exist. (relfilenode, forknum)
2631 : //
2632 : // TODO: Store it as a btree or radix tree or something else that spans multiple
2633 : // key-value pairs, if you have a lot of relations
2634 : pub(crate) rels: HashSet<(Oid, u8)>,
2635 : }
2636 :
2637 0 : #[derive(Debug, Serialize, Deserialize)]
2638 : struct RelSizeEntry {
2639 : nblocks: u32,
2640 : }
2641 :
2642 0 : #[derive(Debug, Serialize, Deserialize, Default)]
2643 : pub(crate) struct SlruSegmentDirectory {
2644 : // Set of SLRU segments that exist.
2645 : pub(crate) segments: HashSet<u32>,
2646 : }
2647 :
2648 : #[derive(Copy, Clone, PartialEq, Eq, Debug, enum_map::Enum)]
2649 : #[repr(u8)]
2650 : pub(crate) enum DirectoryKind {
2651 : Db,
2652 : TwoPhase,
2653 : Rel,
2654 : AuxFiles,
2655 : SlruSegment(SlruKind),
2656 : RelV2,
2657 : }
2658 :
2659 : impl DirectoryKind {
2660 : pub(crate) const KINDS_NUM: usize = <DirectoryKind as Enum>::LENGTH;
2661 17848 : pub(crate) fn offset(&self) -> usize {
2662 17848 : self.into_usize()
2663 17848 : }
2664 : }
2665 :
2666 : static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
2667 :
2668 : #[allow(clippy::bool_assert_comparison)]
2669 : #[cfg(test)]
2670 : mod tests {
2671 : use hex_literal::hex;
2672 : use pageserver_api::models::ShardParameters;
2673 : use pageserver_api::shard::ShardStripeSize;
2674 : use utils::id::TimelineId;
2675 : use utils::shard::{ShardCount, ShardNumber};
2676 :
2677 : use super::*;
2678 : use crate::DEFAULT_PG_VERSION;
2679 : use crate::tenant::harness::TenantHarness;
2680 :
2681 : /// Test a round trip of aux file updates, from DatadirModification to reading back from the Timeline
2682 : #[tokio::test]
2683 4 : async fn aux_files_round_trip() -> anyhow::Result<()> {
2684 4 : let name = "aux_files_round_trip";
2685 4 : let harness = TenantHarness::create(name).await?;
2686 4 :
2687 4 : pub const TIMELINE_ID: TimelineId =
2688 4 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
2689 4 :
2690 4 : let (tenant, ctx) = harness.load().await;
2691 4 : let tline = tenant
2692 4 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2693 4 : .await?;
2694 4 : let tline = tline.raw_timeline().unwrap();
2695 4 :
2696 4 : // First modification: insert two keys
2697 4 : let mut modification = tline.begin_modification(Lsn(0x1000));
2698 4 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
2699 4 : modification.set_lsn(Lsn(0x1008))?;
2700 4 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
2701 4 : modification.commit(&ctx).await?;
2702 4 : let expect_1008 = HashMap::from([
2703 4 : ("foo/bar1".to_string(), Bytes::from_static(b"content1")),
2704 4 : ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
2705 4 : ]);
2706 4 :
2707 4 : let io_concurrency = IoConcurrency::spawn_for_test();
2708 4 :
2709 4 : let readback = tline
2710 4 : .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
2711 4 : .await?;
2712 4 : assert_eq!(readback, expect_1008);
2713 4 :
2714 4 : // Second modification: update one key, remove the other
2715 4 : let mut modification = tline.begin_modification(Lsn(0x2000));
2716 4 : modification.put_file("foo/bar1", b"content3", &ctx).await?;
2717 4 : modification.set_lsn(Lsn(0x2008))?;
2718 4 : modification.put_file("foo/bar2", b"", &ctx).await?;
2719 4 : modification.commit(&ctx).await?;
2720 4 : let expect_2008 =
2721 4 : HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
2722 4 :
2723 4 : let readback = tline
2724 4 : .list_aux_files(Lsn(0x2008), &ctx, io_concurrency.clone())
2725 4 : .await?;
2726 4 : assert_eq!(readback, expect_2008);
2727 4 :
2728 4 : // Reading back in time works
2729 4 : let readback = tline
2730 4 : .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
2731 4 : .await?;
2732 4 : assert_eq!(readback, expect_1008);
2733 4 :
2734 4 : Ok(())
2735 4 : }
2736 :
2737 : #[test]
2738 4 : fn gap_finding() {
2739 4 : let rel = RelTag {
2740 4 : spcnode: 1663,
2741 4 : dbnode: 208101,
2742 4 : relnode: 2620,
2743 4 : forknum: 0,
2744 4 : };
2745 4 : let base_blkno = 1;
2746 4 :
2747 4 : let base_key = rel_block_to_key(rel, base_blkno);
2748 4 : let before_base_key = rel_block_to_key(rel, base_blkno - 1);
2749 4 :
2750 4 : let shard = ShardIdentity::unsharded();
2751 4 :
2752 4 : let mut previous_nblocks = 0;
2753 44 : for i in 0..10 {
2754 40 : let crnt_blkno = base_blkno + i;
2755 40 : let gaps = DatadirModification::find_gaps(rel, crnt_blkno, previous_nblocks, &shard);
2756 40 :
2757 40 : previous_nblocks = crnt_blkno + 1;
2758 40 :
2759 40 : if i == 0 {
2760 : // The first block we write is 1, so we should find the gap.
2761 4 : assert_eq!(gaps.unwrap(), KeySpace::single(before_base_key..base_key));
2762 : } else {
2763 36 : assert!(gaps.is_none());
2764 : }
2765 : }
2766 :
2767 : // This is an update to an already existing block. No gaps here.
2768 4 : let update_blkno = 5;
2769 4 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
2770 4 : assert!(gaps.is_none());
2771 :
2772 : // This is an update past the current end block.
2773 4 : let after_gap_blkno = 20;
2774 4 : let gaps = DatadirModification::find_gaps(rel, after_gap_blkno, previous_nblocks, &shard);
2775 4 :
2776 4 : let gap_start_key = rel_block_to_key(rel, previous_nblocks);
2777 4 : let after_gap_key = rel_block_to_key(rel, after_gap_blkno);
2778 4 : assert_eq!(
2779 4 : gaps.unwrap(),
2780 4 : KeySpace::single(gap_start_key..after_gap_key)
2781 4 : );
2782 4 : }
2783 :
2784 : #[test]
2785 4 : fn sharded_gap_finding() {
2786 4 : let rel = RelTag {
2787 4 : spcnode: 1663,
2788 4 : dbnode: 208101,
2789 4 : relnode: 2620,
2790 4 : forknum: 0,
2791 4 : };
2792 4 :
2793 4 : let first_blkno = 6;
2794 4 :
2795 4 : // This shard will get the even blocks
2796 4 : let shard = ShardIdentity::from_params(
2797 4 : ShardNumber(0),
2798 4 : &ShardParameters {
2799 4 : count: ShardCount(2),
2800 4 : stripe_size: ShardStripeSize(1),
2801 4 : },
2802 4 : );
2803 4 :
2804 4 : // Only keys belonging to this shard are considered as gaps.
2805 4 : let mut previous_nblocks = 0;
2806 4 : let gaps =
2807 4 : DatadirModification::find_gaps(rel, first_blkno, previous_nblocks, &shard).unwrap();
2808 4 : assert!(!gaps.ranges.is_empty());
2809 12 : for gap_range in gaps.ranges {
2810 8 : let mut k = gap_range.start;
2811 16 : while k != gap_range.end {
2812 8 : assert_eq!(shard.get_shard_number(&k), shard.number);
2813 8 : k = k.next();
2814 : }
2815 : }
2816 :
2817 4 : previous_nblocks = first_blkno;
2818 4 :
2819 4 : let update_blkno = 2;
2820 4 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
2821 4 : assert!(gaps.is_none());
2822 4 : }
2823 :
2824 : /*
2825 : fn assert_current_logical_size<R: Repository>(timeline: &DatadirTimeline<R>, lsn: Lsn) {
2826 : let incremental = timeline.get_current_logical_size();
2827 : let non_incremental = timeline
2828 : .get_current_logical_size_non_incremental(lsn)
2829 : .unwrap();
2830 : assert_eq!(incremental, non_incremental);
2831 : }
2832 : */
2833 :
2834 : /*
2835 : ///
2836 : /// Test list_rels() function, with branches and dropped relations
2837 : ///
2838 : #[test]
2839 : fn test_list_rels_drop() -> Result<()> {
2840 : let repo = RepoHarness::create("test_list_rels_drop")?.load();
2841 : let tline = create_empty_timeline(repo, TIMELINE_ID)?;
2842 : const TESTDB: u32 = 111;
2843 :
2844 : // Import initial dummy checkpoint record, otherwise the get_timeline() call
2845 : // after branching fails below
2846 : let mut writer = tline.begin_record(Lsn(0x10));
2847 : writer.put_checkpoint(ZERO_CHECKPOINT.clone())?;
2848 : writer.finish()?;
2849 :
2850 : // Create a relation on the timeline
2851 : let mut writer = tline.begin_record(Lsn(0x20));
2852 : writer.put_rel_page_image(TESTREL_A, 0, TEST_IMG("foo blk 0 at 2"))?;
2853 : writer.finish()?;
2854 :
2855 : let writer = tline.begin_record(Lsn(0x00));
2856 : writer.finish()?;
2857 :
2858 : // Check that list_rels() lists it after LSN 2, but no before it
2859 : assert!(!tline.list_rels(0, TESTDB, Lsn(0x10))?.contains(&TESTREL_A));
2860 : assert!(tline.list_rels(0, TESTDB, Lsn(0x20))?.contains(&TESTREL_A));
2861 : assert!(tline.list_rels(0, TESTDB, Lsn(0x30))?.contains(&TESTREL_A));
2862 :
2863 : // Create a branch, check that the relation is visible there
2864 : repo.branch_timeline(&tline, NEW_TIMELINE_ID, Lsn(0x30))?;
2865 : let newtline = match repo.get_timeline(NEW_TIMELINE_ID)?.local_timeline() {
2866 : Some(timeline) => timeline,
2867 : None => panic!("Should have a local timeline"),
2868 : };
2869 : let newtline = DatadirTimelineImpl::new(newtline);
2870 : assert!(newtline
2871 : .list_rels(0, TESTDB, Lsn(0x30))?
2872 : .contains(&TESTREL_A));
2873 :
2874 : // Drop it on the branch
2875 : let mut new_writer = newtline.begin_record(Lsn(0x40));
2876 : new_writer.drop_relation(TESTREL_A)?;
2877 : new_writer.finish()?;
2878 :
2879 : // Check that it's no longer listed on the branch after the point where it was dropped
2880 : assert!(newtline
2881 : .list_rels(0, TESTDB, Lsn(0x30))?
2882 : .contains(&TESTREL_A));
2883 : assert!(!newtline
2884 : .list_rels(0, TESTDB, Lsn(0x40))?
2885 : .contains(&TESTREL_A));
2886 :
2887 : // Run checkpoint and garbage collection and check that it's still not visible
2888 : newtline.checkpoint(CheckpointConfig::Forced)?;
2889 : repo.gc_iteration(Some(NEW_TIMELINE_ID), 0, true)?;
2890 :
2891 : assert!(!newtline
2892 : .list_rels(0, TESTDB, Lsn(0x40))?
2893 : .contains(&TESTREL_A));
2894 :
2895 : Ok(())
2896 : }
2897 : */
2898 :
2899 : /*
2900 : #[test]
2901 : fn test_read_beyond_eof() -> Result<()> {
2902 : let repo = RepoHarness::create("test_read_beyond_eof")?.load();
2903 : let tline = create_test_timeline(repo, TIMELINE_ID)?;
2904 :
2905 : make_some_layers(&tline, Lsn(0x20))?;
2906 : let mut writer = tline.begin_record(Lsn(0x60));
2907 : walingest.put_rel_page_image(
2908 : &mut writer,
2909 : TESTREL_A,
2910 : 0,
2911 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x60))),
2912 : )?;
2913 : writer.finish()?;
2914 :
2915 : // Test read before rel creation. Should error out.
2916 : assert!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x10), false).is_err());
2917 :
2918 : // Read block beyond end of relation at different points in time.
2919 : // These reads should fall into different delta, image, and in-memory layers.
2920 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x20), false)?, ZERO_PAGE);
2921 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x25), false)?, ZERO_PAGE);
2922 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x30), false)?, ZERO_PAGE);
2923 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x35), false)?, ZERO_PAGE);
2924 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x40), false)?, ZERO_PAGE);
2925 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x45), false)?, ZERO_PAGE);
2926 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x50), false)?, ZERO_PAGE);
2927 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x55), false)?, ZERO_PAGE);
2928 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x60), false)?, ZERO_PAGE);
2929 :
2930 : // Test on an in-memory layer with no preceding layer
2931 : let mut writer = tline.begin_record(Lsn(0x70));
2932 : walingest.put_rel_page_image(
2933 : &mut writer,
2934 : TESTREL_B,
2935 : 0,
2936 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x70))),
2937 : )?;
2938 : writer.finish()?;
2939 :
2940 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_B, 1, Lsn(0x70), false)?6, ZERO_PAGE);
2941 :
2942 : Ok(())
2943 : }
2944 : */
2945 : }
|
