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