Line data Source code
1 : //!
2 : //! This provides an abstraction to store PostgreSQL relations and other files
3 : //! in the key-value store that implements the Repository interface.
4 : //!
5 : //! (TODO: The line between PUT-functions here and walingest.rs is a bit blurry, as
6 : //! walingest.rs handles a few things like implicit relation creation and extension.
7 : //! Clarify that)
8 : //!
9 : use std::collections::{HashMap, HashSet, hash_map};
10 : use std::ops::{ControlFlow, Range};
11 :
12 : use crate::walingest::{WalIngestError, WalIngestErrorKind};
13 : use crate::{PERF_TRACE_TARGET, ensure_walingest};
14 : use anyhow::Context;
15 : use bytes::{Buf, Bytes, BytesMut};
16 : use enum_map::Enum;
17 : use pageserver_api::key::{
18 : AUX_FILES_KEY, CHECKPOINT_KEY, CONTROLFILE_KEY, CompactKey, DBDIR_KEY, Key, RelDirExists,
19 : TWOPHASEDIR_KEY, dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range,
20 : rel_size_to_key, rel_tag_sparse_key, rel_tag_sparse_key_range, relmap_file_key,
21 : repl_origin_key, repl_origin_key_range, slru_block_to_key, slru_dir_to_key,
22 : slru_segment_key_range, slru_segment_size_to_key, twophase_file_key, twophase_key_range,
23 : };
24 : use pageserver_api::keyspace::{KeySpaceRandomAccum, SparseKeySpace};
25 : use pageserver_api::models::RelSizeMigration;
26 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
27 : use pageserver_api::shard::ShardIdentity;
28 : use postgres_ffi::{BLCKSZ, PgMajorVersion, TransactionId};
29 : use postgres_ffi_types::forknum::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
30 : use postgres_ffi_types::{Oid, RepOriginId, TimestampTz};
31 : use serde::{Deserialize, Serialize};
32 : use strum::IntoEnumIterator;
33 : use tokio_util::sync::CancellationToken;
34 : use tracing::{debug, info, info_span, trace, warn};
35 : use utils::bin_ser::{BeSer, DeserializeError};
36 : use utils::lsn::Lsn;
37 : use utils::pausable_failpoint;
38 : use wal_decoder::models::record::NeonWalRecord;
39 : use wal_decoder::models::value::Value;
40 : use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
41 :
42 : use super::tenant::{PageReconstructError, Timeline};
43 : use crate::aux_file;
44 : use crate::context::{PerfInstrumentFutureExt, RequestContext, RequestContextBuilder};
45 : use crate::keyspace::{KeySpace, KeySpaceAccum};
46 : use crate::metrics::{
47 : RELSIZE_CACHE_MISSES_OLD, RELSIZE_LATEST_CACHE_ENTRIES, RELSIZE_LATEST_CACHE_HITS,
48 : RELSIZE_LATEST_CACHE_MISSES, RELSIZE_SNAPSHOT_CACHE_ENTRIES, RELSIZE_SNAPSHOT_CACHE_HITS,
49 : RELSIZE_SNAPSHOT_CACHE_MISSES,
50 : };
51 : use crate::span::{
52 : debug_assert_current_span_has_tenant_and_timeline_id,
53 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
54 : };
55 : use crate::tenant::storage_layer::IoConcurrency;
56 : use crate::tenant::timeline::{GetVectoredError, VersionedKeySpaceQuery};
57 :
58 : /// 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.
59 : pub const MAX_AUX_FILE_DELTAS: usize = 1024;
60 :
61 : /// Max number of aux-file-related delta layers. The compaction will create a new image layer once this threshold is reached.
62 : pub const MAX_AUX_FILE_V2_DELTAS: usize = 16;
63 :
64 : #[derive(Debug)]
65 : pub enum LsnForTimestamp {
66 : /// Found commits both before and after the given timestamp
67 : Present(Lsn),
68 :
69 : /// Found no commits after the given timestamp, this means
70 : /// that the newest data in the branch is older than the given
71 : /// timestamp.
72 : ///
73 : /// All commits <= LSN happened before the given timestamp
74 : Future(Lsn),
75 :
76 : /// The queried timestamp is past our horizon we look back at (PITR)
77 : ///
78 : /// All commits > LSN happened after the given timestamp,
79 : /// but any commits < LSN might have happened before or after
80 : /// the given timestamp. We don't know because no data before
81 : /// the given lsn is available.
82 : Past(Lsn),
83 :
84 : /// We have found no commit with a timestamp,
85 : /// so we can't return anything meaningful.
86 : ///
87 : /// The associated LSN is the lower bound value we can safely
88 : /// create branches on, but no statement is made if it is
89 : /// older or newer than the timestamp.
90 : ///
91 : /// This variant can e.g. be returned right after a
92 : /// cluster import.
93 : NoData(Lsn),
94 : }
95 :
96 : /// Each request to page server contains LSN range: `not_modified_since..request_lsn`.
97 : /// See comments libs/pageserver_api/src/models.rs.
98 : /// Based on this range and `last_record_lsn` PS calculates `effective_lsn`.
99 : /// But to distinguish requests from primary and replicas we need also to pass `request_lsn`.
100 : #[derive(Debug, Clone, Copy, Default)]
101 : pub struct LsnRange {
102 : pub effective_lsn: Lsn,
103 : pub request_lsn: Lsn,
104 : }
105 :
106 : impl LsnRange {
107 0 : pub fn at(lsn: Lsn) -> LsnRange {
108 0 : LsnRange {
109 0 : effective_lsn: lsn,
110 0 : request_lsn: lsn,
111 0 : }
112 0 : }
113 16 : pub fn is_latest(&self) -> bool {
114 16 : self.request_lsn == Lsn::MAX
115 16 : }
116 : }
117 :
118 : #[derive(Debug, thiserror::Error)]
119 : pub(crate) enum CalculateLogicalSizeError {
120 : #[error("cancelled")]
121 : Cancelled,
122 :
123 : /// Something went wrong while reading the metadata we use to calculate logical size
124 : /// Note that cancellation variants of `PageReconstructError` are transformed to [`Self::Cancelled`]
125 : /// in the `From` implementation for this variant.
126 : #[error(transparent)]
127 : PageRead(PageReconstructError),
128 :
129 : /// Something went wrong deserializing metadata that we read to calculate logical size
130 : #[error("decode error: {0}")]
131 : Decode(#[from] DeserializeError),
132 : }
133 :
134 : #[derive(Debug, thiserror::Error)]
135 : pub(crate) enum CollectKeySpaceError {
136 : #[error(transparent)]
137 : Decode(#[from] DeserializeError),
138 : #[error(transparent)]
139 : PageRead(PageReconstructError),
140 : #[error("cancelled")]
141 : Cancelled,
142 : }
143 :
144 : impl CollectKeySpaceError {
145 0 : pub(crate) fn is_cancel(&self) -> bool {
146 0 : match self {
147 0 : CollectKeySpaceError::Decode(_) => false,
148 0 : CollectKeySpaceError::PageRead(e) => e.is_cancel(),
149 0 : CollectKeySpaceError::Cancelled => true,
150 : }
151 0 : }
152 0 : pub(crate) fn into_anyhow(self) -> anyhow::Error {
153 0 : match self {
154 0 : CollectKeySpaceError::Decode(e) => anyhow::Error::new(e),
155 0 : CollectKeySpaceError::PageRead(e) => anyhow::Error::new(e),
156 0 : CollectKeySpaceError::Cancelled => anyhow::Error::new(self),
157 : }
158 0 : }
159 : }
160 :
161 : impl From<PageReconstructError> for CollectKeySpaceError {
162 0 : fn from(err: PageReconstructError) -> Self {
163 0 : match err {
164 0 : PageReconstructError::Cancelled => Self::Cancelled,
165 0 : err => Self::PageRead(err),
166 : }
167 0 : }
168 : }
169 :
170 : impl From<PageReconstructError> for CalculateLogicalSizeError {
171 0 : fn from(pre: PageReconstructError) -> Self {
172 0 : match pre {
173 0 : PageReconstructError::Cancelled => Self::Cancelled,
174 0 : _ => Self::PageRead(pre),
175 : }
176 0 : }
177 : }
178 :
179 : #[derive(Debug, thiserror::Error)]
180 : pub enum RelationError {
181 : #[error("invalid relnode")]
182 : InvalidRelnode,
183 : }
184 :
185 : ///
186 : /// This impl provides all the functionality to store PostgreSQL relations, SLRUs,
187 : /// and other special kinds of files, in a versioned key-value store. The
188 : /// Timeline struct provides the key-value store.
189 : ///
190 : /// This is a separate impl, so that we can easily include all these functions in a Timeline
191 : /// implementation, and might be moved into a separate struct later.
192 : impl Timeline {
193 : /// Start ingesting a WAL record, or other atomic modification of
194 : /// the timeline.
195 : ///
196 : /// This provides a transaction-like interface to perform a bunch
197 : /// of modifications atomically.
198 : ///
199 : /// To ingest a WAL record, call begin_modification(lsn) to get a
200 : /// DatadirModification object. Use the functions in the object to
201 : /// modify the repository state, updating all the pages and metadata
202 : /// that the WAL record affects. When you're done, call commit() to
203 : /// commit the changes.
204 : ///
205 : /// Lsn stored in modification is advanced by `ingest_record` and
206 : /// is used by `commit()` to update `last_record_lsn`.
207 : ///
208 : /// Calling commit() will flush all the changes and reset the state,
209 : /// so the `DatadirModification` struct can be reused to perform the next modification.
210 : ///
211 : /// Note that any pending modifications you make through the
212 : /// modification object won't be visible to calls to the 'get' and list
213 : /// functions of the timeline until you finish! And if you update the
214 : /// same page twice, the last update wins.
215 : ///
216 134215 : pub fn begin_modification(&self, lsn: Lsn) -> DatadirModification
217 134215 : where
218 134215 : Self: Sized,
219 : {
220 134215 : DatadirModification {
221 134215 : tline: self,
222 134215 : pending_lsns: Vec::new(),
223 134215 : pending_metadata_pages: HashMap::new(),
224 134215 : pending_data_batch: None,
225 134215 : pending_deletions: Vec::new(),
226 134215 : pending_nblocks: 0,
227 134215 : pending_directory_entries: Vec::new(),
228 134215 : pending_metadata_bytes: 0,
229 134215 : lsn,
230 134215 : }
231 134215 : }
232 :
233 : //------------------------------------------------------------------------------
234 : // Public GET functions
235 : //------------------------------------------------------------------------------
236 :
237 : /// Look up given page version.
238 9192 : pub(crate) async fn get_rel_page_at_lsn(
239 9192 : &self,
240 9192 : tag: RelTag,
241 9192 : blknum: BlockNumber,
242 9192 : version: Version<'_>,
243 9192 : ctx: &RequestContext,
244 9192 : io_concurrency: IoConcurrency,
245 9192 : ) -> Result<Bytes, PageReconstructError> {
246 9192 : match version {
247 9192 : Version::LsnRange(lsns) => {
248 9192 : let pages: smallvec::SmallVec<[_; 1]> = smallvec::smallvec![(tag, blknum)];
249 9192 : let res = self
250 9192 : .get_rel_page_at_lsn_batched(
251 9192 : pages
252 9192 : .iter()
253 9192 : .map(|(tag, blknum)| (tag, blknum, lsns, ctx.attached_child())),
254 9192 : io_concurrency.clone(),
255 9192 : ctx,
256 : )
257 9192 : .await;
258 9192 : assert_eq!(res.len(), 1);
259 9192 : res.into_iter().next().unwrap()
260 : }
261 0 : Version::Modified(modification) => {
262 0 : if tag.relnode == 0 {
263 0 : return Err(PageReconstructError::Other(
264 0 : RelationError::InvalidRelnode.into(),
265 0 : ));
266 0 : }
267 :
268 0 : let nblocks = self.get_rel_size(tag, version, ctx).await?;
269 0 : if blknum >= nblocks {
270 0 : debug!(
271 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
272 : tag,
273 : blknum,
274 0 : version.get_lsn(),
275 : nblocks
276 : );
277 0 : return Ok(ZERO_PAGE.clone());
278 0 : }
279 :
280 0 : let key = rel_block_to_key(tag, blknum);
281 0 : modification.get(key, ctx).await
282 : }
283 : }
284 9192 : }
285 :
286 : /// Like [`Self::get_rel_page_at_lsn`], but returns a batch of pages.
287 : ///
288 : /// The ordering of the returned vec corresponds to the ordering of `pages`.
289 9192 : pub(crate) async fn get_rel_page_at_lsn_batched(
290 9192 : &self,
291 9192 : pages: impl ExactSizeIterator<Item = (&RelTag, &BlockNumber, LsnRange, RequestContext)>,
292 9192 : io_concurrency: IoConcurrency,
293 9192 : ctx: &RequestContext,
294 9192 : ) -> Vec<Result<Bytes, PageReconstructError>> {
295 9192 : debug_assert_current_span_has_tenant_and_timeline_id();
296 :
297 9192 : let mut slots_filled = 0;
298 9192 : let page_count = pages.len();
299 :
300 : // Would be nice to use smallvec here but it doesn't provide the spare_capacity_mut() API.
301 9192 : let mut result = Vec::with_capacity(pages.len());
302 9192 : let result_slots = result.spare_capacity_mut();
303 :
304 9192 : let mut keys_slots: HashMap<Key, smallvec::SmallVec<[(usize, RequestContext); 1]>> =
305 9192 : HashMap::with_capacity(pages.len());
306 :
307 9192 : let mut req_keyspaces: HashMap<Lsn, KeySpaceRandomAccum> =
308 9192 : HashMap::with_capacity(pages.len());
309 :
310 9192 : for (response_slot_idx, (tag, blknum, lsns, ctx)) in pages.enumerate() {
311 9192 : if tag.relnode == 0 {
312 0 : result_slots[response_slot_idx].write(Err(PageReconstructError::Other(
313 0 : RelationError::InvalidRelnode.into(),
314 0 : )));
315 :
316 0 : slots_filled += 1;
317 0 : continue;
318 9192 : }
319 9192 : let lsn = lsns.effective_lsn;
320 9192 : let nblocks = {
321 9192 : let ctx = RequestContextBuilder::from(&ctx)
322 9192 : .perf_span(|crnt_perf_span| {
323 0 : info_span!(
324 : target: PERF_TRACE_TARGET,
325 0 : parent: crnt_perf_span,
326 : "GET_REL_SIZE",
327 : reltag=%tag,
328 : lsn=%lsn,
329 : )
330 0 : })
331 9192 : .attached_child();
332 :
333 9192 : match self
334 9192 : .get_rel_size(*tag, Version::LsnRange(lsns), &ctx)
335 9192 : .maybe_perf_instrument(&ctx, |crnt_perf_span| crnt_perf_span.clone())
336 9192 : .await
337 : {
338 9192 : Ok(nblocks) => nblocks,
339 0 : Err(err) => {
340 0 : result_slots[response_slot_idx].write(Err(err));
341 0 : slots_filled += 1;
342 0 : continue;
343 : }
344 : }
345 : };
346 :
347 9192 : if *blknum >= nblocks {
348 0 : debug!(
349 0 : "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
350 : tag, blknum, lsn, nblocks
351 : );
352 0 : result_slots[response_slot_idx].write(Ok(ZERO_PAGE.clone()));
353 0 : slots_filled += 1;
354 0 : continue;
355 9192 : }
356 :
357 9192 : let key = rel_block_to_key(*tag, *blknum);
358 :
359 9192 : let ctx = RequestContextBuilder::from(&ctx)
360 9192 : .perf_span(|crnt_perf_span| {
361 0 : info_span!(
362 : target: PERF_TRACE_TARGET,
363 0 : parent: crnt_perf_span,
364 : "GET_BATCH",
365 : batch_size = %page_count,
366 : )
367 0 : })
368 9192 : .attached_child();
369 :
370 9192 : let key_slots = keys_slots.entry(key).or_default();
371 9192 : key_slots.push((response_slot_idx, ctx));
372 :
373 9192 : let acc = req_keyspaces.entry(lsn).or_default();
374 9192 : acc.add_key(key);
375 : }
376 :
377 9192 : let query: Vec<(Lsn, KeySpace)> = req_keyspaces
378 9192 : .into_iter()
379 9192 : .map(|(lsn, acc)| (lsn, acc.to_keyspace()))
380 9192 : .collect();
381 :
382 9192 : let query = VersionedKeySpaceQuery::scattered(query);
383 9192 : let res = self
384 9192 : .get_vectored(query, io_concurrency, ctx)
385 9192 : .maybe_perf_instrument(ctx, |current_perf_span| current_perf_span.clone())
386 9192 : .await;
387 :
388 9192 : match res {
389 9192 : Ok(results) => {
390 18384 : for (key, res) in results {
391 9192 : let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
392 9192 : let (first_slot, first_req_ctx) = key_slots.next().unwrap();
393 :
394 9192 : for (slot, req_ctx) in key_slots {
395 0 : let clone = match &res {
396 0 : Ok(buf) => Ok(buf.clone()),
397 0 : Err(err) => Err(match err {
398 0 : PageReconstructError::Cancelled => PageReconstructError::Cancelled,
399 :
400 0 : x @ PageReconstructError::Other(_)
401 0 : | x @ PageReconstructError::AncestorLsnTimeout(_)
402 0 : | x @ PageReconstructError::WalRedo(_)
403 0 : | x @ PageReconstructError::MissingKey(_) => {
404 0 : PageReconstructError::Other(anyhow::anyhow!(
405 0 : "there was more than one request for this key in the batch, error logged once: {x:?}"
406 0 : ))
407 : }
408 : }),
409 : };
410 :
411 0 : result_slots[slot].write(clone);
412 : // There is no standardized way to express that the batched span followed from N request spans.
413 : // So, abuse the system and mark the request contexts as follows_from the batch span, so we get
414 : // some linkage in our trace viewer. It allows us to answer: which GET_VECTORED did this GET_PAGE wait for.
415 0 : req_ctx.perf_follows_from(ctx);
416 0 : slots_filled += 1;
417 : }
418 :
419 9192 : result_slots[first_slot].write(res);
420 9192 : first_req_ctx.perf_follows_from(ctx);
421 9192 : slots_filled += 1;
422 : }
423 : }
424 0 : Err(err) => {
425 : // this cannot really happen because get_vectored only errors globally on invalid LSN or too large batch size
426 : // (We enforce the max batch size outside of this function, in the code that constructs the batch request.)
427 0 : for (slot, req_ctx) in keys_slots.values().flatten() {
428 : // this whole `match` is a lot like `From<GetVectoredError> for PageReconstructError`
429 : // but without taking ownership of the GetVectoredError
430 0 : let err = match &err {
431 0 : GetVectoredError::Cancelled => Err(PageReconstructError::Cancelled),
432 : // TODO: restructure get_vectored API to make this error per-key
433 0 : GetVectoredError::MissingKey(err) => {
434 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
435 0 : "whole vectored get request failed because one or more of the requested keys were missing: {err:?}"
436 0 : )))
437 : }
438 : // TODO: restructure get_vectored API to make this error per-key
439 0 : GetVectoredError::GetReadyAncestorError(err) => {
440 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
441 0 : "whole vectored get request failed because one or more key required ancestor that wasn't ready: {err:?}"
442 0 : )))
443 : }
444 : // TODO: restructure get_vectored API to make this error per-key
445 0 : GetVectoredError::Other(err) => Err(PageReconstructError::Other(
446 0 : anyhow::anyhow!("whole vectored get request failed: {err:?}"),
447 0 : )),
448 : // TODO: we can prevent this error class by moving this check into the type system
449 0 : GetVectoredError::InvalidLsn(e) => {
450 0 : Err(anyhow::anyhow!("invalid LSN: {e:?}").into())
451 : }
452 : // NB: this should never happen in practice because we limit batch size to be smaller than max_get_vectored_keys
453 : // TODO: we can prevent this error class by moving this check into the type system
454 0 : GetVectoredError::Oversized(err, max) => {
455 0 : Err(anyhow::anyhow!("batching oversized: {err} > {max}").into())
456 : }
457 : };
458 :
459 0 : req_ctx.perf_follows_from(ctx);
460 0 : result_slots[*slot].write(err);
461 : }
462 :
463 0 : slots_filled += keys_slots.values().map(|slots| slots.len()).sum::<usize>();
464 : }
465 : };
466 :
467 9192 : assert_eq!(slots_filled, page_count);
468 : // SAFETY:
469 : // 1. `result` and any of its uninint members are not read from until this point
470 : // 2. The length below is tracked at run-time and matches the number of requested pages.
471 9192 : unsafe {
472 9192 : result.set_len(page_count);
473 9192 : }
474 :
475 9192 : result
476 9192 : }
477 :
478 : /// Get size of a database in blocks. This is only accurate on shard 0. It will undercount on
479 : /// other shards, by only accounting for relations the shard has pages for, and only accounting
480 : /// for pages up to the highest page number it has stored.
481 0 : pub(crate) async fn get_db_size(
482 0 : &self,
483 0 : spcnode: Oid,
484 0 : dbnode: Oid,
485 0 : version: Version<'_>,
486 0 : ctx: &RequestContext,
487 0 : ) -> Result<usize, PageReconstructError> {
488 0 : let mut total_blocks = 0;
489 :
490 0 : let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
491 :
492 0 : if rels.is_empty() {
493 0 : return Ok(0);
494 0 : }
495 :
496 : // Pre-deserialize the rel directory to avoid duplicated work in `get_relsize_cached`.
497 0 : let reldir_key = rel_dir_to_key(spcnode, dbnode);
498 0 : let buf = version.get(self, reldir_key, ctx).await?;
499 0 : let reldir = RelDirectory::des(&buf)?;
500 :
501 0 : for rel in rels {
502 0 : let n_blocks = self
503 0 : .get_rel_size_in_reldir(rel, version, Some((reldir_key, &reldir)), ctx)
504 0 : .await?;
505 0 : total_blocks += n_blocks as usize;
506 : }
507 0 : Ok(total_blocks)
508 0 : }
509 :
510 : /// Get size of a relation file. The relation must exist, otherwise an error is returned.
511 : ///
512 : /// This is only accurate on shard 0. On other shards, it will return the size up to the highest
513 : /// page number stored in the shard.
514 12217 : pub(crate) async fn get_rel_size(
515 12217 : &self,
516 12217 : tag: RelTag,
517 12217 : version: Version<'_>,
518 12217 : ctx: &RequestContext,
519 12217 : ) -> Result<BlockNumber, PageReconstructError> {
520 12217 : self.get_rel_size_in_reldir(tag, version, None, ctx).await
521 12217 : }
522 :
523 : /// Get size of a relation file. The relation must exist, otherwise an error is returned.
524 : ///
525 : /// See [`Self::get_rel_exists_in_reldir`] on why we need `deserialized_reldir_v1`.
526 12217 : pub(crate) async fn get_rel_size_in_reldir(
527 12217 : &self,
528 12217 : tag: RelTag,
529 12217 : version: Version<'_>,
530 12217 : deserialized_reldir_v1: Option<(Key, &RelDirectory)>,
531 12217 : ctx: &RequestContext,
532 12217 : ) -> Result<BlockNumber, PageReconstructError> {
533 12217 : if tag.relnode == 0 {
534 0 : return Err(PageReconstructError::Other(
535 0 : RelationError::InvalidRelnode.into(),
536 0 : ));
537 12217 : }
538 :
539 12217 : if let Some(nblocks) = self.get_cached_rel_size(&tag, version) {
540 12210 : return Ok(nblocks);
541 7 : }
542 :
543 7 : if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
544 0 : && !self
545 0 : .get_rel_exists_in_reldir(tag, version, deserialized_reldir_v1, ctx)
546 0 : .await?
547 : {
548 : // FIXME: Postgres sometimes calls smgrcreate() to create
549 : // FSM, and smgrnblocks() on it immediately afterwards,
550 : // without extending it. Tolerate that by claiming that
551 : // any non-existent FSM fork has size 0.
552 0 : return Ok(0);
553 7 : }
554 :
555 7 : let key = rel_size_to_key(tag);
556 7 : let mut buf = version.get(self, key, ctx).await?;
557 5 : let nblocks = buf.get_u32_le();
558 :
559 5 : self.update_cached_rel_size(tag, version, nblocks);
560 :
561 5 : Ok(nblocks)
562 12217 : }
563 :
564 : /// Does the relation exist?
565 : ///
566 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
567 : /// the shard stores pages for.
568 : ///
569 3025 : pub(crate) async fn get_rel_exists(
570 3025 : &self,
571 3025 : tag: RelTag,
572 3025 : version: Version<'_>,
573 3025 : ctx: &RequestContext,
574 3025 : ) -> Result<bool, PageReconstructError> {
575 3025 : self.get_rel_exists_in_reldir(tag, version, None, ctx).await
576 3025 : }
577 :
578 : /// Does the relation exist? With a cached deserialized `RelDirectory`.
579 : ///
580 : /// There are some cases where the caller loops across all relations. In that specific case,
581 : /// the caller should obtain the deserialized `RelDirectory` first and then call this function
582 : /// to avoid duplicated work of deserliazation. This is a hack and should be removed by introducing
583 : /// a new API (e.g., `get_rel_exists_batched`).
584 3025 : pub(crate) async fn get_rel_exists_in_reldir(
585 3025 : &self,
586 3025 : tag: RelTag,
587 3025 : version: Version<'_>,
588 3025 : deserialized_reldir_v1: Option<(Key, &RelDirectory)>,
589 3025 : ctx: &RequestContext,
590 3025 : ) -> Result<bool, PageReconstructError> {
591 3025 : if tag.relnode == 0 {
592 0 : return Err(PageReconstructError::Other(
593 0 : RelationError::InvalidRelnode.into(),
594 0 : ));
595 3025 : }
596 :
597 : // first try to lookup relation in cache
598 3025 : if let Some(_nblocks) = self.get_cached_rel_size(&tag, version) {
599 3016 : return Ok(true);
600 9 : }
601 : // then check if the database was already initialized.
602 : // get_rel_exists can be called before dbdir is created.
603 9 : let buf = version.get(self, DBDIR_KEY, ctx).await?;
604 9 : let dbdirs = DbDirectory::des(&buf)?.dbdirs;
605 9 : if !dbdirs.contains_key(&(tag.spcnode, tag.dbnode)) {
606 0 : return Ok(false);
607 9 : }
608 :
609 : // Read path: first read the new reldir keyspace. Early return if the relation exists.
610 : // Otherwise, read the old reldir keyspace.
611 : // TODO: if IndexPart::rel_size_migration is `Migrated`, we only need to read from v2.
612 :
613 : if let RelSizeMigration::Migrated | RelSizeMigration::Migrating =
614 9 : self.get_rel_size_v2_status()
615 : {
616 : // fetch directory listing (new)
617 0 : let key = rel_tag_sparse_key(tag.spcnode, tag.dbnode, tag.relnode, tag.forknum);
618 0 : let buf = RelDirExists::decode_option(version.sparse_get(self, key, ctx).await?)
619 0 : .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
620 0 : let exists_v2 = buf == RelDirExists::Exists;
621 : // Fast path: if the relation exists in the new format, return true.
622 : // TODO: we should have a verification mode that checks both keyspaces
623 : // to ensure the relation only exists in one of them.
624 0 : if exists_v2 {
625 0 : return Ok(true);
626 0 : }
627 9 : }
628 :
629 : // fetch directory listing (old)
630 :
631 9 : let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
632 :
633 9 : if let Some((cached_key, dir)) = deserialized_reldir_v1 {
634 0 : if cached_key == key {
635 0 : return Ok(dir.rels.contains(&(tag.relnode, tag.forknum)));
636 0 : } else if cfg!(test) || cfg!(feature = "testing") {
637 0 : panic!("cached reldir key mismatch: {cached_key} != {key}");
638 : } else {
639 0 : warn!("cached reldir key mismatch: {cached_key} != {key}");
640 : }
641 : // Fallback to reading the directory from the datadir.
642 9 : }
643 9 : let buf = version.get(self, key, ctx).await?;
644 :
645 9 : let dir = RelDirectory::des(&buf)?;
646 9 : let exists_v1 = dir.rels.contains(&(tag.relnode, tag.forknum));
647 9 : Ok(exists_v1)
648 3025 : }
649 :
650 : /// Get a list of all existing relations in given tablespace and database.
651 : ///
652 : /// Only shard 0 has a full view of the relations. Other shards only know about relations that
653 : /// the shard stores pages for.
654 : ///
655 : /// # Cancel-Safety
656 : ///
657 : /// This method is cancellation-safe.
658 0 : pub(crate) async fn list_rels(
659 0 : &self,
660 0 : spcnode: Oid,
661 0 : dbnode: Oid,
662 0 : version: Version<'_>,
663 0 : ctx: &RequestContext,
664 0 : ) -> Result<HashSet<RelTag>, PageReconstructError> {
665 : // fetch directory listing (old)
666 0 : let key = rel_dir_to_key(spcnode, dbnode);
667 0 : let buf = version.get(self, key, ctx).await?;
668 :
669 0 : let dir = RelDirectory::des(&buf)?;
670 0 : let rels_v1: HashSet<RelTag> =
671 0 : HashSet::from_iter(dir.rels.iter().map(|(relnode, forknum)| RelTag {
672 0 : spcnode,
673 0 : dbnode,
674 0 : relnode: *relnode,
675 0 : forknum: *forknum,
676 0 : }));
677 :
678 0 : if let RelSizeMigration::Legacy = self.get_rel_size_v2_status() {
679 0 : return Ok(rels_v1);
680 0 : }
681 :
682 : // scan directory listing (new), merge with the old results
683 0 : let key_range = rel_tag_sparse_key_range(spcnode, dbnode);
684 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
685 0 : self.conf.get_vectored_concurrent_io,
686 0 : self.gate
687 0 : .enter()
688 0 : .map_err(|_| PageReconstructError::Cancelled)?,
689 : );
690 0 : let results = self
691 0 : .scan(
692 0 : KeySpace::single(key_range),
693 0 : version.get_lsn(),
694 0 : ctx,
695 0 : io_concurrency,
696 0 : )
697 0 : .await?;
698 0 : let mut rels = rels_v1;
699 0 : for (key, val) in results {
700 0 : let val = RelDirExists::decode(&val?)
701 0 : .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
702 0 : assert_eq!(key.field6, 1);
703 0 : assert_eq!(key.field2, spcnode);
704 0 : assert_eq!(key.field3, dbnode);
705 0 : let tag = RelTag {
706 0 : spcnode,
707 0 : dbnode,
708 0 : relnode: key.field4,
709 0 : forknum: key.field5,
710 0 : };
711 0 : if val == RelDirExists::Removed {
712 0 : debug_assert!(!rels.contains(&tag), "removed reltag in v2");
713 0 : continue;
714 0 : }
715 0 : let did_not_contain = rels.insert(tag);
716 0 : debug_assert!(did_not_contain, "duplicate reltag in v2");
717 : }
718 0 : Ok(rels)
719 0 : }
720 :
721 : /// Get the whole SLRU segment
722 0 : pub(crate) async fn get_slru_segment(
723 0 : &self,
724 0 : kind: SlruKind,
725 0 : segno: u32,
726 0 : lsn: Lsn,
727 0 : ctx: &RequestContext,
728 0 : ) -> Result<Bytes, PageReconstructError> {
729 0 : assert!(self.tenant_shard_id.is_shard_zero());
730 0 : let n_blocks = self
731 0 : .get_slru_segment_size(kind, segno, Version::at(lsn), ctx)
732 0 : .await?;
733 :
734 0 : let keyspace = KeySpace::single(
735 0 : slru_block_to_key(kind, segno, 0)..slru_block_to_key(kind, segno, n_blocks),
736 : );
737 :
738 0 : let batches = keyspace.partition(
739 0 : self.get_shard_identity(),
740 0 : self.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
741 0 : BLCKSZ as u64,
742 : );
743 :
744 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
745 0 : self.conf.get_vectored_concurrent_io,
746 0 : self.gate
747 0 : .enter()
748 0 : .map_err(|_| PageReconstructError::Cancelled)?,
749 : );
750 :
751 0 : let mut segment = BytesMut::with_capacity(n_blocks as usize * BLCKSZ as usize);
752 0 : for batch in batches.parts {
753 0 : let query = VersionedKeySpaceQuery::uniform(batch, lsn);
754 0 : let blocks = self
755 0 : .get_vectored(query, io_concurrency.clone(), ctx)
756 0 : .await?;
757 :
758 0 : for (_key, block) in blocks {
759 0 : let block = block?;
760 0 : segment.extend_from_slice(&block[..BLCKSZ as usize]);
761 : }
762 : }
763 :
764 0 : Ok(segment.freeze())
765 0 : }
766 :
767 : /// Get size of an SLRU segment
768 0 : pub(crate) async fn get_slru_segment_size(
769 0 : &self,
770 0 : kind: SlruKind,
771 0 : segno: u32,
772 0 : version: Version<'_>,
773 0 : ctx: &RequestContext,
774 0 : ) -> Result<BlockNumber, PageReconstructError> {
775 0 : assert!(self.tenant_shard_id.is_shard_zero());
776 0 : let key = slru_segment_size_to_key(kind, segno);
777 0 : let mut buf = version.get(self, key, ctx).await?;
778 0 : Ok(buf.get_u32_le())
779 0 : }
780 :
781 : /// Does the slru segment exist?
782 0 : pub(crate) async fn get_slru_segment_exists(
783 0 : &self,
784 0 : kind: SlruKind,
785 0 : segno: u32,
786 0 : version: Version<'_>,
787 0 : ctx: &RequestContext,
788 0 : ) -> Result<bool, PageReconstructError> {
789 0 : assert!(self.tenant_shard_id.is_shard_zero());
790 : // fetch directory listing
791 0 : let key = slru_dir_to_key(kind);
792 0 : let buf = version.get(self, key, ctx).await?;
793 :
794 0 : let dir = SlruSegmentDirectory::des(&buf)?;
795 0 : Ok(dir.segments.contains(&segno))
796 0 : }
797 :
798 : /// Locate LSN, such that all transactions that committed before
799 : /// 'search_timestamp' are visible, but nothing newer is.
800 : ///
801 : /// This is not exact. Commit timestamps are not guaranteed to be ordered,
802 : /// so it's not well defined which LSN you get if there were multiple commits
803 : /// "in flight" at that point in time.
804 : ///
805 0 : pub(crate) async fn find_lsn_for_timestamp(
806 0 : &self,
807 0 : search_timestamp: TimestampTz,
808 0 : cancel: &CancellationToken,
809 0 : ctx: &RequestContext,
810 0 : ) -> Result<LsnForTimestamp, PageReconstructError> {
811 0 : pausable_failpoint!("find-lsn-for-timestamp-pausable");
812 :
813 0 : let gc_cutoff_lsn_guard = self.get_applied_gc_cutoff_lsn();
814 0 : let gc_cutoff_planned = {
815 0 : let gc_info = self.gc_info.read().unwrap();
816 0 : gc_info.min_cutoff()
817 : };
818 : // Usually the planned cutoff is newer than the cutoff of the last gc run,
819 : // but let's be defensive.
820 0 : let gc_cutoff = gc_cutoff_planned.max(*gc_cutoff_lsn_guard);
821 : // We use this method to figure out the branching LSN for the new branch, but the
822 : // GC cutoff could be before the branching point and we cannot create a new branch
823 : // with LSN < `ancestor_lsn`. Thus, pick the maximum of these two to be
824 : // on the safe side.
825 0 : let min_lsn = std::cmp::max(gc_cutoff, self.get_ancestor_lsn());
826 0 : let max_lsn = self.get_last_record_lsn();
827 :
828 : // LSNs are always 8-byte aligned. low/mid/high represent the
829 : // LSN divided by 8.
830 0 : let mut low = min_lsn.0 / 8;
831 0 : let mut high = max_lsn.0 / 8 + 1;
832 :
833 0 : let mut found_smaller = false;
834 0 : let mut found_larger = false;
835 :
836 0 : while low < high {
837 0 : if cancel.is_cancelled() {
838 0 : return Err(PageReconstructError::Cancelled);
839 0 : }
840 : // cannot overflow, high and low are both smaller than u64::MAX / 2
841 0 : let mid = (high + low) / 2;
842 :
843 0 : let cmp = match self
844 0 : .is_latest_commit_timestamp_ge_than(
845 0 : search_timestamp,
846 0 : Lsn(mid * 8),
847 0 : &mut found_smaller,
848 0 : &mut found_larger,
849 0 : ctx,
850 : )
851 0 : .await
852 : {
853 0 : Ok(res) => res,
854 0 : Err(PageReconstructError::MissingKey(e)) => {
855 0 : warn!(
856 0 : "Missing key while find_lsn_for_timestamp. Either we might have already garbage-collected that data or the key is really missing. Last error: {:#}",
857 : e
858 : );
859 : // Return that we didn't find any requests smaller than the LSN, and logging the error.
860 0 : return Ok(LsnForTimestamp::Past(min_lsn));
861 : }
862 0 : Err(e) => return Err(e),
863 : };
864 :
865 0 : if cmp {
866 0 : high = mid;
867 0 : } else {
868 0 : low = mid + 1;
869 0 : }
870 : }
871 :
872 : // If `found_smaller == true`, `low = t + 1` where `t` is the target LSN,
873 : // so the LSN of the last commit record before or at `search_timestamp`.
874 : // Remove one from `low` to get `t`.
875 : //
876 : // FIXME: it would be better to get the LSN of the previous commit.
877 : // Otherwise, if you restore to the returned LSN, the database will
878 : // include physical changes from later commits that will be marked
879 : // as aborted, and will need to be vacuumed away.
880 0 : let commit_lsn = Lsn((low - 1) * 8);
881 0 : match (found_smaller, found_larger) {
882 : (false, false) => {
883 : // This can happen if no commit records have been processed yet, e.g.
884 : // just after importing a cluster.
885 0 : Ok(LsnForTimestamp::NoData(min_lsn))
886 : }
887 : (false, true) => {
888 : // Didn't find any commit timestamps smaller than the request
889 0 : Ok(LsnForTimestamp::Past(min_lsn))
890 : }
891 0 : (true, _) if commit_lsn < min_lsn => {
892 : // the search above did set found_smaller to true but it never increased the lsn.
893 : // Then, low is still the old min_lsn, and the subtraction above gave a value
894 : // below the min_lsn. We should never do that.
895 0 : Ok(LsnForTimestamp::Past(min_lsn))
896 : }
897 : (true, false) => {
898 : // Only found commits with timestamps smaller than the request.
899 : // It's still a valid case for branch creation, return it.
900 : // And `update_gc_info()` ignores LSN for a `LsnForTimestamp::Future`
901 : // case, anyway.
902 0 : Ok(LsnForTimestamp::Future(commit_lsn))
903 : }
904 0 : (true, true) => Ok(LsnForTimestamp::Present(commit_lsn)),
905 : }
906 0 : }
907 :
908 : /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
909 : /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
910 : ///
911 : /// Additionally, sets 'found_smaller'/'found_Larger, if encounters any commits
912 : /// with a smaller/larger timestamp.
913 : ///
914 0 : pub(crate) async fn is_latest_commit_timestamp_ge_than(
915 0 : &self,
916 0 : search_timestamp: TimestampTz,
917 0 : probe_lsn: Lsn,
918 0 : found_smaller: &mut bool,
919 0 : found_larger: &mut bool,
920 0 : ctx: &RequestContext,
921 0 : ) -> Result<bool, PageReconstructError> {
922 0 : self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
923 0 : if timestamp >= search_timestamp {
924 0 : *found_larger = true;
925 0 : return ControlFlow::Break(true);
926 0 : } else {
927 0 : *found_smaller = true;
928 0 : }
929 0 : ControlFlow::Continue(())
930 0 : })
931 0 : .await
932 0 : }
933 :
934 : /// Obtain the timestamp for the given lsn.
935 : ///
936 : /// If the lsn has no timestamps (e.g. no commits), returns None.
937 0 : pub(crate) async fn get_timestamp_for_lsn(
938 0 : &self,
939 0 : probe_lsn: Lsn,
940 0 : ctx: &RequestContext,
941 0 : ) -> Result<Option<TimestampTz>, PageReconstructError> {
942 0 : let mut max: Option<TimestampTz> = None;
943 0 : self.map_all_timestamps::<()>(probe_lsn, ctx, |timestamp| {
944 0 : if let Some(max_prev) = max {
945 0 : max = Some(max_prev.max(timestamp));
946 0 : } else {
947 0 : max = Some(timestamp);
948 0 : }
949 0 : ControlFlow::Continue(())
950 0 : })
951 0 : .await?;
952 :
953 0 : Ok(max)
954 0 : }
955 :
956 : /// Runs the given function on all the timestamps for a given lsn
957 : ///
958 : /// The return value is either given by the closure, or set to the `Default`
959 : /// impl's output.
960 0 : async fn map_all_timestamps<T: Default>(
961 0 : &self,
962 0 : probe_lsn: Lsn,
963 0 : ctx: &RequestContext,
964 0 : mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
965 0 : ) -> Result<T, PageReconstructError> {
966 0 : for segno in self
967 0 : .list_slru_segments(SlruKind::Clog, Version::at(probe_lsn), ctx)
968 0 : .await?
969 : {
970 0 : let nblocks = self
971 0 : .get_slru_segment_size(SlruKind::Clog, segno, Version::at(probe_lsn), ctx)
972 0 : .await?;
973 :
974 0 : let keyspace = KeySpace::single(
975 0 : slru_block_to_key(SlruKind::Clog, segno, 0)
976 0 : ..slru_block_to_key(SlruKind::Clog, segno, nblocks),
977 : );
978 :
979 0 : let batches = keyspace.partition(
980 0 : self.get_shard_identity(),
981 0 : self.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
982 0 : BLCKSZ as u64,
983 : );
984 :
985 0 : let io_concurrency = IoConcurrency::spawn_from_conf(
986 0 : self.conf.get_vectored_concurrent_io,
987 0 : self.gate
988 0 : .enter()
989 0 : .map_err(|_| PageReconstructError::Cancelled)?,
990 : );
991 :
992 0 : for batch in batches.parts.into_iter().rev() {
993 0 : let query = VersionedKeySpaceQuery::uniform(batch, probe_lsn);
994 0 : let blocks = self
995 0 : .get_vectored(query, io_concurrency.clone(), ctx)
996 0 : .await?;
997 :
998 0 : for (_key, clog_page) in blocks.into_iter().rev() {
999 0 : let clog_page = clog_page?;
1000 :
1001 0 : if clog_page.len() == BLCKSZ as usize + 8 {
1002 0 : let mut timestamp_bytes = [0u8; 8];
1003 0 : timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
1004 0 : let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
1005 :
1006 0 : match f(timestamp) {
1007 0 : ControlFlow::Break(b) => return Ok(b),
1008 0 : ControlFlow::Continue(()) => (),
1009 : }
1010 0 : }
1011 : }
1012 : }
1013 : }
1014 0 : Ok(Default::default())
1015 0 : }
1016 :
1017 0 : pub(crate) async fn get_slru_keyspace(
1018 0 : &self,
1019 0 : version: Version<'_>,
1020 0 : ctx: &RequestContext,
1021 0 : ) -> Result<KeySpace, PageReconstructError> {
1022 0 : let mut accum = KeySpaceAccum::new();
1023 :
1024 0 : for kind in SlruKind::iter() {
1025 0 : let mut segments: Vec<u32> = self
1026 0 : .list_slru_segments(kind, version, ctx)
1027 0 : .await?
1028 0 : .into_iter()
1029 0 : .collect();
1030 0 : segments.sort_unstable();
1031 :
1032 0 : for seg in segments {
1033 0 : let block_count = self.get_slru_segment_size(kind, seg, version, ctx).await?;
1034 :
1035 0 : accum.add_range(
1036 0 : slru_block_to_key(kind, seg, 0)..slru_block_to_key(kind, seg, block_count),
1037 : );
1038 : }
1039 : }
1040 :
1041 0 : Ok(accum.to_keyspace())
1042 0 : }
1043 :
1044 : /// Get a list of SLRU segments
1045 0 : pub(crate) async fn list_slru_segments(
1046 0 : &self,
1047 0 : kind: SlruKind,
1048 0 : version: Version<'_>,
1049 0 : ctx: &RequestContext,
1050 0 : ) -> Result<HashSet<u32>, PageReconstructError> {
1051 : // fetch directory entry
1052 0 : let key = slru_dir_to_key(kind);
1053 :
1054 0 : let buf = version.get(self, key, ctx).await?;
1055 0 : Ok(SlruSegmentDirectory::des(&buf)?.segments)
1056 0 : }
1057 :
1058 0 : pub(crate) async fn get_relmap_file(
1059 0 : &self,
1060 0 : spcnode: Oid,
1061 0 : dbnode: Oid,
1062 0 : version: Version<'_>,
1063 0 : ctx: &RequestContext,
1064 0 : ) -> Result<Bytes, PageReconstructError> {
1065 0 : let key = relmap_file_key(spcnode, dbnode);
1066 :
1067 0 : let buf = version.get(self, key, ctx).await?;
1068 0 : Ok(buf)
1069 0 : }
1070 :
1071 169 : pub(crate) async fn list_dbdirs(
1072 169 : &self,
1073 169 : lsn: Lsn,
1074 169 : ctx: &RequestContext,
1075 169 : ) -> Result<HashMap<(Oid, Oid), bool>, PageReconstructError> {
1076 : // fetch directory entry
1077 169 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
1078 :
1079 169 : Ok(DbDirectory::des(&buf)?.dbdirs)
1080 169 : }
1081 :
1082 0 : pub(crate) async fn get_twophase_file(
1083 0 : &self,
1084 0 : xid: u64,
1085 0 : lsn: Lsn,
1086 0 : ctx: &RequestContext,
1087 0 : ) -> Result<Bytes, PageReconstructError> {
1088 0 : let key = twophase_file_key(xid);
1089 0 : let buf = self.get(key, lsn, ctx).await?;
1090 0 : Ok(buf)
1091 0 : }
1092 :
1093 170 : pub(crate) async fn list_twophase_files(
1094 170 : &self,
1095 170 : lsn: Lsn,
1096 170 : ctx: &RequestContext,
1097 170 : ) -> Result<HashSet<u64>, PageReconstructError> {
1098 : // fetch directory entry
1099 170 : let buf = self.get(TWOPHASEDIR_KEY, lsn, ctx).await?;
1100 :
1101 170 : if self.pg_version >= PgMajorVersion::PG17 {
1102 157 : Ok(TwoPhaseDirectoryV17::des(&buf)?.xids)
1103 : } else {
1104 13 : Ok(TwoPhaseDirectory::des(&buf)?
1105 : .xids
1106 13 : .iter()
1107 13 : .map(|x| u64::from(*x))
1108 13 : .collect())
1109 : }
1110 170 : }
1111 :
1112 0 : pub(crate) async fn get_control_file(
1113 0 : &self,
1114 0 : lsn: Lsn,
1115 0 : ctx: &RequestContext,
1116 0 : ) -> Result<Bytes, PageReconstructError> {
1117 0 : self.get(CONTROLFILE_KEY, lsn, ctx).await
1118 0 : }
1119 :
1120 6 : pub(crate) async fn get_checkpoint(
1121 6 : &self,
1122 6 : lsn: Lsn,
1123 6 : ctx: &RequestContext,
1124 6 : ) -> Result<Bytes, PageReconstructError> {
1125 6 : self.get(CHECKPOINT_KEY, lsn, ctx).await
1126 6 : }
1127 :
1128 6 : async fn list_aux_files_v2(
1129 6 : &self,
1130 6 : lsn: Lsn,
1131 6 : ctx: &RequestContext,
1132 6 : io_concurrency: IoConcurrency,
1133 6 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
1134 6 : let kv = self
1135 6 : .scan(
1136 6 : KeySpace::single(Key::metadata_aux_key_range()),
1137 6 : lsn,
1138 6 : ctx,
1139 6 : io_concurrency,
1140 6 : )
1141 6 : .await?;
1142 6 : let mut result = HashMap::new();
1143 6 : let mut sz = 0;
1144 15 : for (_, v) in kv {
1145 9 : let v = v?;
1146 9 : let v = aux_file::decode_file_value_bytes(&v)
1147 9 : .context("value decode")
1148 9 : .map_err(PageReconstructError::Other)?;
1149 17 : for (fname, content) in v {
1150 8 : sz += fname.len();
1151 8 : sz += content.len();
1152 8 : result.insert(fname, content);
1153 8 : }
1154 : }
1155 6 : self.aux_file_size_estimator.on_initial(sz);
1156 6 : Ok(result)
1157 6 : }
1158 :
1159 0 : pub(crate) async fn trigger_aux_file_size_computation(
1160 0 : &self,
1161 0 : lsn: Lsn,
1162 0 : ctx: &RequestContext,
1163 0 : io_concurrency: IoConcurrency,
1164 0 : ) -> Result<(), PageReconstructError> {
1165 0 : self.list_aux_files_v2(lsn, ctx, io_concurrency).await?;
1166 0 : Ok(())
1167 0 : }
1168 :
1169 6 : pub(crate) async fn list_aux_files(
1170 6 : &self,
1171 6 : lsn: Lsn,
1172 6 : ctx: &RequestContext,
1173 6 : io_concurrency: IoConcurrency,
1174 6 : ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
1175 6 : self.list_aux_files_v2(lsn, ctx, io_concurrency).await
1176 6 : }
1177 :
1178 2 : pub(crate) async fn get_replorigins(
1179 2 : &self,
1180 2 : lsn: Lsn,
1181 2 : ctx: &RequestContext,
1182 2 : io_concurrency: IoConcurrency,
1183 2 : ) -> Result<HashMap<RepOriginId, Lsn>, PageReconstructError> {
1184 2 : let kv = self
1185 2 : .scan(
1186 2 : KeySpace::single(repl_origin_key_range()),
1187 2 : lsn,
1188 2 : ctx,
1189 2 : io_concurrency,
1190 2 : )
1191 2 : .await?;
1192 2 : let mut result = HashMap::new();
1193 6 : for (k, v) in kv {
1194 5 : let v = v?;
1195 5 : if v.is_empty() {
1196 : // This is a tombstone -- we can skip it.
1197 : // Originally, the replorigin code uses `Lsn::INVALID` to represent a tombstone. However, as it part of
1198 : // the sparse keyspace and the sparse keyspace uses an empty image to universally represent a tombstone,
1199 : // we also need to consider that. Such tombstones might be written on the detach ancestor code path to
1200 : // avoid the value going into the child branch. (See [`crate::tenant::timeline::detach_ancestor::generate_tombstone_image_layer`] for more details.)
1201 2 : continue;
1202 3 : }
1203 3 : let origin_id = k.field6 as RepOriginId;
1204 3 : let origin_lsn = Lsn::des(&v)
1205 3 : .with_context(|| format!("decode replorigin value for {origin_id}: {v:?}"))?;
1206 2 : if origin_lsn != Lsn::INVALID {
1207 2 : result.insert(origin_id, origin_lsn);
1208 2 : }
1209 : }
1210 1 : Ok(result)
1211 2 : }
1212 :
1213 : /// Does the same as get_current_logical_size but counted on demand.
1214 : /// Used to initialize the logical size tracking on startup.
1215 : ///
1216 : /// Only relation blocks are counted currently. That excludes metadata,
1217 : /// SLRUs, twophase files etc.
1218 : ///
1219 : /// # Cancel-Safety
1220 : ///
1221 : /// This method is cancellation-safe.
1222 7 : pub(crate) async fn get_current_logical_size_non_incremental(
1223 7 : &self,
1224 7 : lsn: Lsn,
1225 7 : ctx: &RequestContext,
1226 7 : ) -> Result<u64, CalculateLogicalSizeError> {
1227 7 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1228 :
1229 7 : fail::fail_point!("skip-logical-size-calculation", |_| { Ok(0) });
1230 :
1231 : // Fetch list of database dirs and iterate them
1232 7 : let buf = self.get(DBDIR_KEY, lsn, ctx).await?;
1233 7 : let dbdir = DbDirectory::des(&buf)?;
1234 :
1235 7 : let mut total_size: u64 = 0;
1236 7 : for (spcnode, dbnode) in dbdir.dbdirs.keys() {
1237 0 : for rel in self
1238 0 : .list_rels(*spcnode, *dbnode, Version::at(lsn), ctx)
1239 0 : .await?
1240 : {
1241 0 : if self.cancel.is_cancelled() {
1242 0 : return Err(CalculateLogicalSizeError::Cancelled);
1243 0 : }
1244 0 : let relsize_key = rel_size_to_key(rel);
1245 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
1246 0 : let relsize = buf.get_u32_le();
1247 :
1248 0 : total_size += relsize as u64;
1249 : }
1250 : }
1251 7 : Ok(total_size * BLCKSZ as u64)
1252 7 : }
1253 :
1254 : /// Get a KeySpace that covers all the Keys that are in use at AND below the given LSN. This is only used
1255 : /// for gc-compaction.
1256 : ///
1257 : /// gc-compaction cannot use the same `collect_keyspace` function as the legacy compaction because it
1258 : /// processes data at multiple LSNs and needs to be aware of the fact that some key ranges might need to
1259 : /// be kept only for a specific range of LSN.
1260 : ///
1261 : /// Consider the case that the user created branches at LSN 10 and 20, where the user created a table A at
1262 : /// LSN 10 and dropped that table at LSN 20. `collect_keyspace` at LSN 10 will return the key range
1263 : /// corresponding to that table, while LSN 20 won't. The keyspace info at a single LSN is not enough to
1264 : /// determine which keys to retain/drop for gc-compaction.
1265 : ///
1266 : /// For now, it only drops AUX-v1 keys. But in the future, the function will be extended to return the keyspace
1267 : /// to be retained for each of the branch LSN.
1268 : ///
1269 : /// The return value is (dense keyspace, sparse keyspace).
1270 27 : pub(crate) async fn collect_gc_compaction_keyspace(
1271 27 : &self,
1272 27 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
1273 27 : let metadata_key_begin = Key::metadata_key_range().start;
1274 27 : let aux_v1_key = AUX_FILES_KEY;
1275 27 : let dense_keyspace = KeySpace {
1276 27 : ranges: vec![Key::MIN..aux_v1_key, aux_v1_key.next()..metadata_key_begin],
1277 27 : };
1278 27 : Ok((
1279 27 : dense_keyspace,
1280 27 : SparseKeySpace(KeySpace::single(Key::metadata_key_range())),
1281 27 : ))
1282 27 : }
1283 :
1284 : ///
1285 : /// Get a KeySpace that covers all the Keys that are in use at the given LSN.
1286 : /// Anything that's not listed maybe removed from the underlying storage (from
1287 : /// that LSN forwards).
1288 : ///
1289 : /// The return value is (dense keyspace, sparse keyspace).
1290 169 : pub(crate) async fn collect_keyspace(
1291 169 : &self,
1292 169 : lsn: Lsn,
1293 169 : ctx: &RequestContext,
1294 169 : ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
1295 : // Iterate through key ranges, greedily packing them into partitions
1296 169 : let mut result = KeySpaceAccum::new();
1297 :
1298 : // The dbdir metadata always exists
1299 169 : result.add_key(DBDIR_KEY);
1300 :
1301 : // Fetch list of database dirs and iterate them
1302 169 : let dbdir = self.list_dbdirs(lsn, ctx).await?;
1303 169 : let mut dbs: Vec<((Oid, Oid), bool)> = dbdir.into_iter().collect();
1304 :
1305 169 : dbs.sort_unstable_by(|(k_a, _), (k_b, _)| k_a.cmp(k_b));
1306 169 : for ((spcnode, dbnode), has_relmap_file) in dbs {
1307 0 : if has_relmap_file {
1308 0 : result.add_key(relmap_file_key(spcnode, dbnode));
1309 0 : }
1310 0 : result.add_key(rel_dir_to_key(spcnode, dbnode));
1311 :
1312 0 : let mut rels: Vec<RelTag> = self
1313 0 : .list_rels(spcnode, dbnode, Version::at(lsn), ctx)
1314 0 : .await?
1315 0 : .into_iter()
1316 0 : .collect();
1317 0 : rels.sort_unstable();
1318 0 : for rel in rels {
1319 0 : let relsize_key = rel_size_to_key(rel);
1320 0 : let mut buf = self.get(relsize_key, lsn, ctx).await?;
1321 0 : let relsize = buf.get_u32_le();
1322 :
1323 0 : result.add_range(rel_block_to_key(rel, 0)..rel_block_to_key(rel, relsize));
1324 0 : result.add_key(relsize_key);
1325 : }
1326 : }
1327 :
1328 : // Iterate SLRUs next
1329 169 : if self.tenant_shard_id.is_shard_zero() {
1330 498 : for kind in [
1331 166 : SlruKind::Clog,
1332 166 : SlruKind::MultiXactMembers,
1333 166 : SlruKind::MultiXactOffsets,
1334 : ] {
1335 498 : let slrudir_key = slru_dir_to_key(kind);
1336 498 : result.add_key(slrudir_key);
1337 498 : let buf = self.get(slrudir_key, lsn, ctx).await?;
1338 498 : let dir = SlruSegmentDirectory::des(&buf)?;
1339 498 : let mut segments: Vec<u32> = dir.segments.iter().cloned().collect();
1340 498 : segments.sort_unstable();
1341 498 : for segno in segments {
1342 0 : let segsize_key = slru_segment_size_to_key(kind, segno);
1343 0 : let mut buf = self.get(segsize_key, lsn, ctx).await?;
1344 0 : let segsize = buf.get_u32_le();
1345 :
1346 0 : result.add_range(
1347 0 : slru_block_to_key(kind, segno, 0)..slru_block_to_key(kind, segno, segsize),
1348 : );
1349 0 : result.add_key(segsize_key);
1350 : }
1351 : }
1352 3 : }
1353 :
1354 : // Then pg_twophase
1355 169 : result.add_key(TWOPHASEDIR_KEY);
1356 :
1357 169 : let mut xids: Vec<u64> = self
1358 169 : .list_twophase_files(lsn, ctx)
1359 169 : .await?
1360 169 : .iter()
1361 169 : .cloned()
1362 169 : .collect();
1363 169 : xids.sort_unstable();
1364 169 : for xid in xids {
1365 0 : result.add_key(twophase_file_key(xid));
1366 0 : }
1367 :
1368 169 : result.add_key(CONTROLFILE_KEY);
1369 169 : result.add_key(CHECKPOINT_KEY);
1370 :
1371 : // Add extra keyspaces in the test cases. Some test cases write keys into the storage without
1372 : // creating directory keys. These test cases will add such keyspaces into `extra_test_dense_keyspace`
1373 : // and the keys will not be garbage-colllected.
1374 : #[cfg(test)]
1375 : {
1376 169 : let guard = self.extra_test_dense_keyspace.load();
1377 169 : for kr in &guard.ranges {
1378 0 : result.add_range(kr.clone());
1379 0 : }
1380 : }
1381 :
1382 169 : let dense_keyspace = result.to_keyspace();
1383 169 : let sparse_keyspace = SparseKeySpace(KeySpace {
1384 169 : ranges: vec![
1385 169 : Key::metadata_aux_key_range(),
1386 169 : repl_origin_key_range(),
1387 169 : Key::rel_dir_sparse_key_range(),
1388 169 : ],
1389 169 : });
1390 :
1391 169 : if cfg!(debug_assertions) {
1392 : // Verify if the sparse keyspaces are ordered and non-overlapping.
1393 :
1394 : // We do not use KeySpaceAccum for sparse_keyspace because we want to ensure each
1395 : // category of sparse keys are split into their own image/delta files. If there
1396 : // are overlapping keyspaces, they will be automatically merged by keyspace accum,
1397 : // and we want the developer to keep the keyspaces separated.
1398 :
1399 169 : let ranges = &sparse_keyspace.0.ranges;
1400 :
1401 : // TODO: use a single overlaps_with across the codebase
1402 507 : fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
1403 507 : !(a.end <= b.start || b.end <= a.start)
1404 507 : }
1405 507 : for i in 0..ranges.len() {
1406 507 : for j in 0..i {
1407 507 : if overlaps_with(&ranges[i], &ranges[j]) {
1408 0 : panic!(
1409 0 : "overlapping sparse keyspace: {}..{} and {}..{}",
1410 0 : ranges[i].start, ranges[i].end, ranges[j].start, ranges[j].end
1411 : );
1412 507 : }
1413 : }
1414 : }
1415 338 : for i in 1..ranges.len() {
1416 338 : assert!(
1417 338 : ranges[i - 1].end <= ranges[i].start,
1418 0 : "unordered sparse keyspace: {}..{} and {}..{}",
1419 0 : ranges[i - 1].start,
1420 0 : ranges[i - 1].end,
1421 0 : ranges[i].start,
1422 0 : ranges[i].end
1423 : );
1424 : }
1425 0 : }
1426 :
1427 169 : Ok((dense_keyspace, sparse_keyspace))
1428 169 : }
1429 :
1430 : /// Get cached size of relation. There are two caches: one for primary updates, it captures the latest state of
1431 : /// of the timeline and snapshot cache, which key includes LSN and so can be used by replicas to get relation size
1432 : /// at the particular LSN (snapshot).
1433 224270 : pub fn get_cached_rel_size(&self, tag: &RelTag, version: Version<'_>) -> Option<BlockNumber> {
1434 224270 : let lsn = version.get_lsn();
1435 : {
1436 224270 : let rel_size_cache = self.rel_size_latest_cache.read().unwrap();
1437 224270 : if let Some((cached_lsn, nblocks)) = rel_size_cache.get(tag) {
1438 224259 : if lsn >= *cached_lsn {
1439 221686 : RELSIZE_LATEST_CACHE_HITS.inc();
1440 221686 : return Some(*nblocks);
1441 2573 : }
1442 2573 : RELSIZE_CACHE_MISSES_OLD.inc();
1443 11 : }
1444 : }
1445 : {
1446 2584 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
1447 2584 : if let Some(nblock) = rel_size_cache.get(&(lsn, *tag)) {
1448 2563 : RELSIZE_SNAPSHOT_CACHE_HITS.inc();
1449 2563 : return Some(*nblock);
1450 21 : }
1451 : }
1452 21 : if version.is_latest() {
1453 10 : RELSIZE_LATEST_CACHE_MISSES.inc();
1454 11 : } else {
1455 11 : RELSIZE_SNAPSHOT_CACHE_MISSES.inc();
1456 11 : }
1457 21 : None
1458 224270 : }
1459 :
1460 : /// Update cached relation size if there is no more recent update
1461 5 : pub fn update_cached_rel_size(&self, tag: RelTag, version: Version<'_>, nblocks: BlockNumber) {
1462 5 : let lsn = version.get_lsn();
1463 5 : if version.is_latest() {
1464 0 : let mut rel_size_cache = self.rel_size_latest_cache.write().unwrap();
1465 0 : match rel_size_cache.entry(tag) {
1466 0 : hash_map::Entry::Occupied(mut entry) => {
1467 0 : let cached_lsn = entry.get_mut();
1468 0 : if lsn >= cached_lsn.0 {
1469 0 : *cached_lsn = (lsn, nblocks);
1470 0 : }
1471 : }
1472 0 : hash_map::Entry::Vacant(entry) => {
1473 0 : entry.insert((lsn, nblocks));
1474 0 : RELSIZE_LATEST_CACHE_ENTRIES.inc();
1475 0 : }
1476 : }
1477 : } else {
1478 5 : let mut rel_size_cache = self.rel_size_snapshot_cache.lock().unwrap();
1479 5 : if rel_size_cache.capacity() != 0 {
1480 5 : rel_size_cache.insert((lsn, tag), nblocks);
1481 5 : RELSIZE_SNAPSHOT_CACHE_ENTRIES.set(rel_size_cache.len() as u64);
1482 5 : }
1483 : }
1484 5 : }
1485 :
1486 : /// Store cached relation size
1487 141360 : pub fn set_cached_rel_size(&self, tag: RelTag, lsn: Lsn, nblocks: BlockNumber) {
1488 141360 : let mut rel_size_cache = self.rel_size_latest_cache.write().unwrap();
1489 141360 : if rel_size_cache.insert(tag, (lsn, nblocks)).is_none() {
1490 960 : RELSIZE_LATEST_CACHE_ENTRIES.inc();
1491 140400 : }
1492 141360 : }
1493 :
1494 : /// Remove cached relation size
1495 1 : pub fn remove_cached_rel_size(&self, tag: &RelTag) {
1496 1 : let mut rel_size_cache = self.rel_size_latest_cache.write().unwrap();
1497 1 : if rel_size_cache.remove(tag).is_some() {
1498 1 : RELSIZE_LATEST_CACHE_ENTRIES.dec();
1499 1 : }
1500 1 : }
1501 : }
1502 :
1503 : /// DatadirModification represents an operation to ingest an atomic set of
1504 : /// updates to the repository.
1505 : ///
1506 : /// It is created by the 'begin_record' function. It is called for each WAL
1507 : /// record, so that all the modifications by a one WAL record appear atomic.
1508 : pub struct DatadirModification<'a> {
1509 : /// The timeline this modification applies to. You can access this to
1510 : /// read the state, but note that any pending updates are *not* reflected
1511 : /// in the state in 'tline' yet.
1512 : pub tline: &'a Timeline,
1513 :
1514 : /// Current LSN of the modification
1515 : lsn: Lsn,
1516 :
1517 : // The modifications are not applied directly to the underlying key-value store.
1518 : // The put-functions add the modifications here, and they are flushed to the
1519 : // underlying key-value store by the 'finish' function.
1520 : pending_lsns: Vec<Lsn>,
1521 : pending_deletions: Vec<(Range<Key>, Lsn)>,
1522 : pending_nblocks: i64,
1523 :
1524 : /// Metadata writes, indexed by key so that they can be read from not-yet-committed modifications
1525 : /// while ingesting subsequent records. See [`Self::is_data_key`] for the definition of 'metadata'.
1526 : pending_metadata_pages: HashMap<CompactKey, Vec<(Lsn, usize, Value)>>,
1527 :
1528 : /// Data writes, ready to be flushed into an ephemeral layer. See [`Self::is_data_key`] for
1529 : /// which keys are stored here.
1530 : pending_data_batch: Option<SerializedValueBatch>,
1531 :
1532 : /// For special "directory" keys that store key-value maps, track the size of the map
1533 : /// if it was updated in this modification.
1534 : pending_directory_entries: Vec<(DirectoryKind, MetricsUpdate)>,
1535 :
1536 : /// An **approximation** of how many metadata bytes will be written to the EphemeralFile.
1537 : pending_metadata_bytes: usize,
1538 : }
1539 :
1540 : #[derive(Debug, Clone, Copy, PartialEq, Eq)]
1541 : pub enum MetricsUpdate {
1542 : /// Set the metrics to this value
1543 : Set(u64),
1544 : /// Increment the metrics by this value
1545 : Add(u64),
1546 : /// Decrement the metrics by this value
1547 : Sub(u64),
1548 : }
1549 :
1550 : impl DatadirModification<'_> {
1551 : // When a DatadirModification is committed, we do a monolithic serialization of all its contents. WAL records can
1552 : // contain multiple pages, so the pageserver's record-based batch size isn't sufficient to bound this allocation: we
1553 : // additionally specify a limit on how much payload a DatadirModification may contain before it should be committed.
1554 : pub(crate) const MAX_PENDING_BYTES: usize = 8 * 1024 * 1024;
1555 :
1556 : /// Get the current lsn
1557 1 : pub(crate) fn get_lsn(&self) -> Lsn {
1558 1 : self.lsn
1559 1 : }
1560 :
1561 0 : pub(crate) fn approx_pending_bytes(&self) -> usize {
1562 0 : self.pending_data_batch
1563 0 : .as_ref()
1564 0 : .map_or(0, |b| b.buffer_size())
1565 0 : + self.pending_metadata_bytes
1566 0 : }
1567 :
1568 0 : pub(crate) fn has_dirty_data(&self) -> bool {
1569 0 : self.pending_data_batch
1570 0 : .as_ref()
1571 0 : .is_some_and(|b| b.has_data())
1572 0 : }
1573 :
1574 : /// Returns statistics about the currently pending modifications.
1575 0 : pub(crate) fn stats(&self) -> DatadirModificationStats {
1576 0 : let mut stats = DatadirModificationStats::default();
1577 0 : for (_, _, value) in self.pending_metadata_pages.values().flatten() {
1578 0 : match value {
1579 0 : Value::Image(_) => stats.metadata_images += 1,
1580 0 : Value::WalRecord(r) if r.will_init() => stats.metadata_images += 1,
1581 0 : Value::WalRecord(_) => stats.metadata_deltas += 1,
1582 : }
1583 : }
1584 0 : for valuemeta in self.pending_data_batch.iter().flat_map(|b| &b.metadata) {
1585 0 : match valuemeta {
1586 0 : ValueMeta::Serialized(s) if s.will_init => stats.data_images += 1,
1587 0 : ValueMeta::Serialized(_) => stats.data_deltas += 1,
1588 0 : ValueMeta::Observed(_) => {}
1589 : }
1590 : }
1591 0 : stats
1592 0 : }
1593 :
1594 : /// Set the current lsn
1595 72929 : pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> Result<(), WalIngestError> {
1596 72929 : ensure_walingest!(
1597 : lsn >= self.lsn,
1598 : "setting an older lsn {} than {} is not allowed",
1599 : lsn,
1600 : self.lsn
1601 : );
1602 :
1603 72929 : if lsn > self.lsn {
1604 72929 : self.pending_lsns.push(self.lsn);
1605 72929 : self.lsn = lsn;
1606 72929 : }
1607 72929 : Ok(())
1608 72929 : }
1609 :
1610 : /// In this context, 'metadata' means keys that are only read by the pageserver internally, and 'data' means
1611 : /// keys that represent literal blocks that postgres can read. So data includes relation blocks and
1612 : /// SLRU blocks, which are read directly by postgres, and everything else is considered metadata.
1613 : ///
1614 : /// The distinction is important because data keys are handled on a fast path where dirty writes are
1615 : /// not readable until this modification is committed, whereas metadata keys are visible for read
1616 : /// via [`Self::get`] as soon as their record has been ingested.
1617 425371 : fn is_data_key(key: &Key) -> bool {
1618 425371 : key.is_rel_block_key() || key.is_slru_block_key()
1619 425371 : }
1620 :
1621 : /// Initialize a completely new repository.
1622 : ///
1623 : /// This inserts the directory metadata entries that are assumed to
1624 : /// always exist.
1625 112 : pub fn init_empty(&mut self) -> anyhow::Result<()> {
1626 112 : let buf = DbDirectory::ser(&DbDirectory {
1627 112 : dbdirs: HashMap::new(),
1628 112 : })?;
1629 112 : self.pending_directory_entries
1630 112 : .push((DirectoryKind::Db, MetricsUpdate::Set(0)));
1631 112 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1632 :
1633 112 : let buf = if self.tline.pg_version >= PgMajorVersion::PG17 {
1634 99 : TwoPhaseDirectoryV17::ser(&TwoPhaseDirectoryV17 {
1635 99 : xids: HashSet::new(),
1636 99 : })
1637 : } else {
1638 13 : TwoPhaseDirectory::ser(&TwoPhaseDirectory {
1639 13 : xids: HashSet::new(),
1640 13 : })
1641 0 : }?;
1642 112 : self.pending_directory_entries
1643 112 : .push((DirectoryKind::TwoPhase, MetricsUpdate::Set(0)));
1644 112 : self.put(TWOPHASEDIR_KEY, Value::Image(buf.into()));
1645 :
1646 112 : let buf: Bytes = SlruSegmentDirectory::ser(&SlruSegmentDirectory::default())?.into();
1647 112 : let empty_dir = Value::Image(buf);
1648 :
1649 : // Initialize SLRUs on shard 0 only: creating these on other shards would be
1650 : // harmless but they'd just be dropped on later compaction.
1651 112 : if self.tline.tenant_shard_id.is_shard_zero() {
1652 109 : self.put(slru_dir_to_key(SlruKind::Clog), empty_dir.clone());
1653 109 : self.pending_directory_entries.push((
1654 109 : DirectoryKind::SlruSegment(SlruKind::Clog),
1655 109 : MetricsUpdate::Set(0),
1656 109 : ));
1657 109 : self.put(
1658 109 : slru_dir_to_key(SlruKind::MultiXactMembers),
1659 109 : empty_dir.clone(),
1660 109 : );
1661 109 : self.pending_directory_entries.push((
1662 109 : DirectoryKind::SlruSegment(SlruKind::Clog),
1663 109 : MetricsUpdate::Set(0),
1664 109 : ));
1665 109 : self.put(slru_dir_to_key(SlruKind::MultiXactOffsets), empty_dir);
1666 109 : self.pending_directory_entries.push((
1667 109 : DirectoryKind::SlruSegment(SlruKind::MultiXactOffsets),
1668 109 : MetricsUpdate::Set(0),
1669 109 : ));
1670 109 : }
1671 :
1672 112 : Ok(())
1673 112 : }
1674 :
1675 : #[cfg(test)]
1676 111 : pub fn init_empty_test_timeline(&mut self) -> anyhow::Result<()> {
1677 111 : self.init_empty()?;
1678 111 : self.put_control_file(bytes::Bytes::from_static(
1679 111 : b"control_file contents do not matter",
1680 : ))
1681 111 : .context("put_control_file")?;
1682 111 : self.put_checkpoint(bytes::Bytes::from_static(
1683 111 : b"checkpoint_file contents do not matter",
1684 : ))
1685 111 : .context("put_checkpoint_file")?;
1686 111 : Ok(())
1687 111 : }
1688 :
1689 : /// Creates a relation if it is not already present.
1690 : /// Returns the current size of the relation
1691 209028 : pub(crate) async fn create_relation_if_required(
1692 209028 : &mut self,
1693 209028 : rel: RelTag,
1694 209028 : ctx: &RequestContext,
1695 209028 : ) -> Result<u32, WalIngestError> {
1696 : // Get current size and put rel creation if rel doesn't exist
1697 : //
1698 : // NOTE: we check the cache first even though get_rel_exists and get_rel_size would
1699 : // check the cache too. This is because eagerly checking the cache results in
1700 : // less work overall and 10% better performance. It's more work on cache miss
1701 : // but cache miss is rare.
1702 209028 : if let Some(nblocks) = self
1703 209028 : .tline
1704 209028 : .get_cached_rel_size(&rel, Version::Modified(self))
1705 : {
1706 209023 : Ok(nblocks)
1707 5 : } else if !self
1708 5 : .tline
1709 5 : .get_rel_exists(rel, Version::Modified(self), ctx)
1710 5 : .await?
1711 : {
1712 : // create it with 0 size initially, the logic below will extend it
1713 5 : self.put_rel_creation(rel, 0, ctx).await?;
1714 5 : Ok(0)
1715 : } else {
1716 0 : Ok(self
1717 0 : .tline
1718 0 : .get_rel_size(rel, Version::Modified(self), ctx)
1719 0 : .await?)
1720 : }
1721 209028 : }
1722 :
1723 : /// Given a block number for a relation (which represents a newly written block),
1724 : /// the previous block count of the relation, and the shard info, find the gaps
1725 : /// that were created by the newly written block if any.
1726 72835 : fn find_gaps(
1727 72835 : rel: RelTag,
1728 72835 : blkno: u32,
1729 72835 : previous_nblocks: u32,
1730 72835 : shard: &ShardIdentity,
1731 72835 : ) -> Option<KeySpace> {
1732 72835 : let mut key = rel_block_to_key(rel, blkno);
1733 72835 : let mut gap_accum = None;
1734 :
1735 72835 : for gap_blkno in previous_nblocks..blkno {
1736 16 : key.field6 = gap_blkno;
1737 :
1738 16 : if shard.get_shard_number(&key) != shard.number {
1739 4 : continue;
1740 12 : }
1741 :
1742 12 : gap_accum
1743 12 : .get_or_insert_with(KeySpaceAccum::new)
1744 12 : .add_key(key);
1745 : }
1746 :
1747 72835 : gap_accum.map(|accum| accum.to_keyspace())
1748 72835 : }
1749 :
1750 72926 : pub async fn ingest_batch(
1751 72926 : &mut self,
1752 72926 : mut batch: SerializedValueBatch,
1753 72926 : // TODO(vlad): remove this argument and replace the shard check with is_key_local
1754 72926 : shard: &ShardIdentity,
1755 72926 : ctx: &RequestContext,
1756 72926 : ) -> Result<(), WalIngestError> {
1757 72926 : let mut gaps_at_lsns = Vec::default();
1758 :
1759 72926 : for meta in batch.metadata.iter() {
1760 72821 : let key = Key::from_compact(meta.key());
1761 72821 : let (rel, blkno) = key
1762 72821 : .to_rel_block()
1763 72821 : .map_err(|_| WalIngestErrorKind::InvalidKey(key, meta.lsn()))?;
1764 72821 : let new_nblocks = blkno + 1;
1765 :
1766 72821 : let old_nblocks = self.create_relation_if_required(rel, ctx).await?;
1767 72821 : if new_nblocks > old_nblocks {
1768 1195 : self.put_rel_extend(rel, new_nblocks, ctx).await?;
1769 71626 : }
1770 :
1771 72821 : if let Some(gaps) = Self::find_gaps(rel, blkno, old_nblocks, shard) {
1772 0 : gaps_at_lsns.push((gaps, meta.lsn()));
1773 72821 : }
1774 : }
1775 :
1776 72926 : if !gaps_at_lsns.is_empty() {
1777 0 : batch.zero_gaps(gaps_at_lsns);
1778 72926 : }
1779 :
1780 72926 : match self.pending_data_batch.as_mut() {
1781 10 : Some(pending_batch) => {
1782 10 : pending_batch.extend(batch);
1783 10 : }
1784 72916 : None if batch.has_data() => {
1785 72815 : self.pending_data_batch = Some(batch);
1786 72815 : }
1787 101 : None => {
1788 101 : // Nothing to initialize the batch with
1789 101 : }
1790 : }
1791 :
1792 72926 : Ok(())
1793 72926 : }
1794 :
1795 : /// Put a new page version that can be constructed from a WAL record
1796 : ///
1797 : /// NOTE: this will *not* implicitly extend the relation, if the page is beyond the
1798 : /// current end-of-file. It's up to the caller to check that the relation size
1799 : /// matches the blocks inserted!
1800 6 : pub fn put_rel_wal_record(
1801 6 : &mut self,
1802 6 : rel: RelTag,
1803 6 : blknum: BlockNumber,
1804 6 : rec: NeonWalRecord,
1805 6 : ) -> Result<(), WalIngestError> {
1806 6 : ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
1807 6 : self.put(rel_block_to_key(rel, blknum), Value::WalRecord(rec));
1808 6 : Ok(())
1809 6 : }
1810 :
1811 : // Same, but for an SLRU.
1812 4 : pub fn put_slru_wal_record(
1813 4 : &mut self,
1814 4 : kind: SlruKind,
1815 4 : segno: u32,
1816 4 : blknum: BlockNumber,
1817 4 : rec: NeonWalRecord,
1818 4 : ) -> Result<(), WalIngestError> {
1819 4 : if !self.tline.tenant_shard_id.is_shard_zero() {
1820 0 : return Ok(());
1821 4 : }
1822 :
1823 4 : self.put(
1824 4 : slru_block_to_key(kind, segno, blknum),
1825 4 : Value::WalRecord(rec),
1826 : );
1827 4 : Ok(())
1828 4 : }
1829 :
1830 : /// Like put_wal_record, but with ready-made image of the page.
1831 138921 : pub fn put_rel_page_image(
1832 138921 : &mut self,
1833 138921 : rel: RelTag,
1834 138921 : blknum: BlockNumber,
1835 138921 : img: Bytes,
1836 138921 : ) -> Result<(), WalIngestError> {
1837 138921 : ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
1838 138921 : let key = rel_block_to_key(rel, blknum);
1839 138921 : if !key.is_valid_key_on_write_path() {
1840 0 : Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
1841 138921 : }
1842 138921 : self.put(rel_block_to_key(rel, blknum), Value::Image(img));
1843 138921 : Ok(())
1844 138921 : }
1845 :
1846 3 : pub fn put_slru_page_image(
1847 3 : &mut self,
1848 3 : kind: SlruKind,
1849 3 : segno: u32,
1850 3 : blknum: BlockNumber,
1851 3 : img: Bytes,
1852 3 : ) -> Result<(), WalIngestError> {
1853 3 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1854 :
1855 3 : let key = slru_block_to_key(kind, segno, blknum);
1856 3 : if !key.is_valid_key_on_write_path() {
1857 0 : Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
1858 3 : }
1859 3 : self.put(key, Value::Image(img));
1860 3 : Ok(())
1861 3 : }
1862 :
1863 1499 : pub(crate) fn put_rel_page_image_zero(
1864 1499 : &mut self,
1865 1499 : rel: RelTag,
1866 1499 : blknum: BlockNumber,
1867 1499 : ) -> Result<(), WalIngestError> {
1868 1499 : ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
1869 1499 : let key = rel_block_to_key(rel, blknum);
1870 1499 : if !key.is_valid_key_on_write_path() {
1871 0 : Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
1872 1499 : }
1873 :
1874 1499 : let batch = self
1875 1499 : .pending_data_batch
1876 1499 : .get_or_insert_with(SerializedValueBatch::default);
1877 :
1878 1499 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1879 :
1880 1499 : Ok(())
1881 1499 : }
1882 :
1883 0 : pub(crate) fn put_slru_page_image_zero(
1884 0 : &mut self,
1885 0 : kind: SlruKind,
1886 0 : segno: u32,
1887 0 : blknum: BlockNumber,
1888 0 : ) -> Result<(), WalIngestError> {
1889 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
1890 0 : let key = slru_block_to_key(kind, segno, blknum);
1891 0 : if !key.is_valid_key_on_write_path() {
1892 0 : Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
1893 0 : }
1894 :
1895 0 : let batch = self
1896 0 : .pending_data_batch
1897 0 : .get_or_insert_with(SerializedValueBatch::default);
1898 :
1899 0 : batch.put(key.to_compact(), Value::Image(ZERO_PAGE.clone()), self.lsn);
1900 :
1901 0 : Ok(())
1902 0 : }
1903 :
1904 : /// Returns `true` if the rel_size_v2 write path is enabled. If it is the first time that
1905 : /// we enable it, we also need to persist it in `index_part.json`.
1906 973 : pub fn maybe_enable_rel_size_v2(&mut self) -> anyhow::Result<bool> {
1907 973 : let status = self.tline.get_rel_size_v2_status();
1908 973 : let config = self.tline.get_rel_size_v2_enabled();
1909 973 : match (config, status) {
1910 : (false, RelSizeMigration::Legacy) => {
1911 : // tenant config didn't enable it and we didn't write any reldir_v2 key yet
1912 973 : Ok(false)
1913 : }
1914 : (false, RelSizeMigration::Migrating | RelSizeMigration::Migrated) => {
1915 : // index_part already persisted that the timeline has enabled rel_size_v2
1916 0 : Ok(true)
1917 : }
1918 : (true, RelSizeMigration::Legacy) => {
1919 : // The first time we enable it, we need to persist it in `index_part.json`
1920 0 : self.tline
1921 0 : .update_rel_size_v2_status(RelSizeMigration::Migrating)?;
1922 0 : tracing::info!("enabled rel_size_v2");
1923 0 : Ok(true)
1924 : }
1925 : (true, RelSizeMigration::Migrating | RelSizeMigration::Migrated) => {
1926 : // index_part already persisted that the timeline has enabled rel_size_v2
1927 : // and we don't need to do anything
1928 0 : Ok(true)
1929 : }
1930 : }
1931 973 : }
1932 :
1933 : /// Store a relmapper file (pg_filenode.map) in the repository
1934 8 : pub async fn put_relmap_file(
1935 8 : &mut self,
1936 8 : spcnode: Oid,
1937 8 : dbnode: Oid,
1938 8 : img: Bytes,
1939 8 : ctx: &RequestContext,
1940 8 : ) -> Result<(), WalIngestError> {
1941 8 : let v2_enabled = self
1942 8 : .maybe_enable_rel_size_v2()
1943 8 : .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
1944 :
1945 : // Add it to the directory (if it doesn't exist already)
1946 8 : let buf = self.get(DBDIR_KEY, ctx).await?;
1947 8 : let mut dbdir = DbDirectory::des(&buf)?;
1948 :
1949 8 : let r = dbdir.dbdirs.insert((spcnode, dbnode), true);
1950 8 : if r.is_none() || r == Some(false) {
1951 : // The dbdir entry didn't exist, or it contained a
1952 : // 'false'. The 'insert' call already updated it with
1953 : // 'true', now write the updated 'dbdirs' map back.
1954 8 : let buf = DbDirectory::ser(&dbdir)?;
1955 8 : self.put(DBDIR_KEY, Value::Image(buf.into()));
1956 0 : }
1957 8 : if r.is_none() {
1958 : // Create RelDirectory
1959 : // TODO: if we have fully migrated to v2, no need to create this directory
1960 4 : let buf = RelDirectory::ser(&RelDirectory {
1961 4 : rels: HashSet::new(),
1962 4 : })?;
1963 4 : self.pending_directory_entries
1964 4 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
1965 4 : if v2_enabled {
1966 0 : self.pending_directory_entries
1967 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
1968 4 : }
1969 4 : self.put(
1970 4 : rel_dir_to_key(spcnode, dbnode),
1971 4 : Value::Image(Bytes::from(buf)),
1972 : );
1973 4 : }
1974 :
1975 8 : self.put(relmap_file_key(spcnode, dbnode), Value::Image(img));
1976 8 : Ok(())
1977 8 : }
1978 :
1979 0 : pub async fn put_twophase_file(
1980 0 : &mut self,
1981 0 : xid: u64,
1982 0 : img: Bytes,
1983 0 : ctx: &RequestContext,
1984 0 : ) -> Result<(), WalIngestError> {
1985 : // Add it to the directory entry
1986 0 : let dirbuf = self.get(TWOPHASEDIR_KEY, ctx).await?;
1987 0 : let newdirbuf = if self.tline.pg_version >= PgMajorVersion::PG17 {
1988 0 : let mut dir = TwoPhaseDirectoryV17::des(&dirbuf)?;
1989 0 : if !dir.xids.insert(xid) {
1990 0 : Err(WalIngestErrorKind::FileAlreadyExists(xid))?;
1991 0 : }
1992 0 : self.pending_directory_entries.push((
1993 0 : DirectoryKind::TwoPhase,
1994 0 : MetricsUpdate::Set(dir.xids.len() as u64),
1995 0 : ));
1996 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
1997 : } else {
1998 0 : let xid = xid as u32;
1999 0 : let mut dir = TwoPhaseDirectory::des(&dirbuf)?;
2000 0 : if !dir.xids.insert(xid) {
2001 0 : Err(WalIngestErrorKind::FileAlreadyExists(xid.into()))?;
2002 0 : }
2003 0 : self.pending_directory_entries.push((
2004 0 : DirectoryKind::TwoPhase,
2005 0 : MetricsUpdate::Set(dir.xids.len() as u64),
2006 0 : ));
2007 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
2008 : };
2009 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
2010 :
2011 0 : self.put(twophase_file_key(xid), Value::Image(img));
2012 0 : Ok(())
2013 0 : }
2014 :
2015 1 : pub async fn set_replorigin(
2016 1 : &mut self,
2017 1 : origin_id: RepOriginId,
2018 1 : origin_lsn: Lsn,
2019 1 : ) -> Result<(), WalIngestError> {
2020 1 : let key = repl_origin_key(origin_id);
2021 1 : self.put(key, Value::Image(origin_lsn.ser().unwrap().into()));
2022 1 : Ok(())
2023 1 : }
2024 :
2025 0 : pub async fn drop_replorigin(&mut self, origin_id: RepOriginId) -> Result<(), WalIngestError> {
2026 0 : self.set_replorigin(origin_id, Lsn::INVALID).await
2027 0 : }
2028 :
2029 112 : pub fn put_control_file(&mut self, img: Bytes) -> Result<(), WalIngestError> {
2030 112 : self.put(CONTROLFILE_KEY, Value::Image(img));
2031 112 : Ok(())
2032 112 : }
2033 :
2034 119 : pub fn put_checkpoint(&mut self, img: Bytes) -> Result<(), WalIngestError> {
2035 119 : self.put(CHECKPOINT_KEY, Value::Image(img));
2036 119 : Ok(())
2037 119 : }
2038 :
2039 0 : pub async fn drop_dbdir(
2040 0 : &mut self,
2041 0 : spcnode: Oid,
2042 0 : dbnode: Oid,
2043 0 : ctx: &RequestContext,
2044 0 : ) -> Result<(), WalIngestError> {
2045 0 : let total_blocks = self
2046 0 : .tline
2047 0 : .get_db_size(spcnode, dbnode, Version::Modified(self), ctx)
2048 0 : .await?;
2049 :
2050 : // Remove entry from dbdir
2051 0 : let buf = self.get(DBDIR_KEY, ctx).await?;
2052 0 : let mut dir = DbDirectory::des(&buf)?;
2053 0 : if dir.dbdirs.remove(&(spcnode, dbnode)).is_some() {
2054 0 : let buf = DbDirectory::ser(&dir)?;
2055 0 : self.pending_directory_entries.push((
2056 0 : DirectoryKind::Db,
2057 0 : MetricsUpdate::Set(dir.dbdirs.len() as u64),
2058 0 : ));
2059 0 : self.put(DBDIR_KEY, Value::Image(buf.into()));
2060 : } else {
2061 0 : warn!(
2062 0 : "dropped dbdir for spcnode {} dbnode {} did not exist in db directory",
2063 : spcnode, dbnode
2064 : );
2065 : }
2066 :
2067 : // Update logical database size.
2068 0 : self.pending_nblocks -= total_blocks as i64;
2069 :
2070 : // Delete all relations and metadata files for the spcnode/dnode
2071 0 : self.delete(dbdir_key_range(spcnode, dbnode));
2072 0 : Ok(())
2073 0 : }
2074 :
2075 : /// Create a relation fork.
2076 : ///
2077 : /// 'nblocks' is the initial size.
2078 960 : pub async fn put_rel_creation(
2079 960 : &mut self,
2080 960 : rel: RelTag,
2081 960 : nblocks: BlockNumber,
2082 960 : ctx: &RequestContext,
2083 960 : ) -> Result<(), WalIngestError> {
2084 960 : if rel.relnode == 0 {
2085 0 : Err(WalIngestErrorKind::LogicalError(anyhow::anyhow!(
2086 0 : "invalid relnode"
2087 0 : )))?;
2088 960 : }
2089 : // It's possible that this is the first rel for this db in this
2090 : // tablespace. Create the reldir entry for it if so.
2091 960 : let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await?)?;
2092 :
2093 960 : let dbdir_exists =
2094 960 : if let hash_map::Entry::Vacant(e) = dbdir.dbdirs.entry((rel.spcnode, rel.dbnode)) {
2095 : // Didn't exist. Update dbdir
2096 4 : e.insert(false);
2097 4 : let buf = DbDirectory::ser(&dbdir)?;
2098 4 : self.pending_directory_entries.push((
2099 4 : DirectoryKind::Db,
2100 4 : MetricsUpdate::Set(dbdir.dbdirs.len() as u64),
2101 4 : ));
2102 4 : self.put(DBDIR_KEY, Value::Image(buf.into()));
2103 4 : false
2104 : } else {
2105 956 : true
2106 : };
2107 :
2108 960 : let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
2109 960 : let mut rel_dir = if !dbdir_exists {
2110 : // Create the RelDirectory
2111 4 : RelDirectory::default()
2112 : } else {
2113 : // reldir already exists, fetch it
2114 956 : RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?
2115 : };
2116 :
2117 960 : let v2_enabled = self
2118 960 : .maybe_enable_rel_size_v2()
2119 960 : .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
2120 :
2121 960 : if v2_enabled {
2122 0 : if rel_dir.rels.contains(&(rel.relnode, rel.forknum)) {
2123 0 : Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
2124 0 : }
2125 0 : let sparse_rel_dir_key =
2126 0 : rel_tag_sparse_key(rel.spcnode, rel.dbnode, rel.relnode, rel.forknum);
2127 : // check if the rel_dir_key exists in v2
2128 0 : let val = self.sparse_get(sparse_rel_dir_key, ctx).await?;
2129 0 : let val = RelDirExists::decode_option(val)
2130 0 : .map_err(|_| WalIngestErrorKind::InvalidRelDirKey(sparse_rel_dir_key))?;
2131 0 : if val == RelDirExists::Exists {
2132 0 : Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
2133 0 : }
2134 0 : self.put(
2135 0 : sparse_rel_dir_key,
2136 0 : Value::Image(RelDirExists::Exists.encode()),
2137 : );
2138 0 : if !dbdir_exists {
2139 0 : self.pending_directory_entries
2140 0 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
2141 0 : self.pending_directory_entries
2142 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
2143 : // We don't write `rel_dir_key -> rel_dir.rels` back to the storage in the v2 path unless it's the initial creation.
2144 : // TODO: if we have fully migrated to v2, no need to create this directory. Otherwise, there
2145 : // will be key not found errors if we don't create an empty one for rel_size_v2.
2146 0 : self.put(
2147 0 : rel_dir_key,
2148 0 : Value::Image(Bytes::from(RelDirectory::ser(&RelDirectory::default())?)),
2149 : );
2150 0 : }
2151 0 : self.pending_directory_entries
2152 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Add(1)));
2153 : } else {
2154 : // Add the new relation to the rel directory entry, and write it back
2155 960 : if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
2156 0 : Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
2157 960 : }
2158 960 : if !dbdir_exists {
2159 4 : self.pending_directory_entries
2160 4 : .push((DirectoryKind::Rel, MetricsUpdate::Set(0)))
2161 956 : }
2162 960 : self.pending_directory_entries
2163 960 : .push((DirectoryKind::Rel, MetricsUpdate::Add(1)));
2164 960 : self.put(
2165 960 : rel_dir_key,
2166 960 : Value::Image(Bytes::from(RelDirectory::ser(&rel_dir)?)),
2167 : );
2168 : }
2169 :
2170 : // Put size
2171 960 : let size_key = rel_size_to_key(rel);
2172 960 : let buf = nblocks.to_le_bytes();
2173 960 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2174 :
2175 960 : self.pending_nblocks += nblocks as i64;
2176 :
2177 : // Update relation size cache
2178 960 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
2179 :
2180 : // Even if nblocks > 0, we don't insert any actual blocks here. That's up to the
2181 : // caller.
2182 960 : Ok(())
2183 960 : }
2184 :
2185 : /// Truncate relation
2186 3006 : pub async fn put_rel_truncation(
2187 3006 : &mut self,
2188 3006 : rel: RelTag,
2189 3006 : nblocks: BlockNumber,
2190 3006 : ctx: &RequestContext,
2191 3006 : ) -> Result<(), WalIngestError> {
2192 3006 : ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
2193 3006 : if self
2194 3006 : .tline
2195 3006 : .get_rel_exists(rel, Version::Modified(self), ctx)
2196 3006 : .await?
2197 : {
2198 3006 : let size_key = rel_size_to_key(rel);
2199 : // Fetch the old size first
2200 3006 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
2201 :
2202 : // Update the entry with the new size.
2203 3006 : let buf = nblocks.to_le_bytes();
2204 3006 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2205 :
2206 : // Update relation size cache
2207 3006 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
2208 :
2209 : // Update logical database size.
2210 3006 : self.pending_nblocks -= old_size as i64 - nblocks as i64;
2211 0 : }
2212 3006 : Ok(())
2213 3006 : }
2214 :
2215 : /// Extend relation
2216 : /// If new size is smaller, do nothing.
2217 138340 : pub async fn put_rel_extend(
2218 138340 : &mut self,
2219 138340 : rel: RelTag,
2220 138340 : nblocks: BlockNumber,
2221 138340 : ctx: &RequestContext,
2222 138340 : ) -> Result<(), WalIngestError> {
2223 138340 : ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
2224 :
2225 : // Put size
2226 138340 : let size_key = rel_size_to_key(rel);
2227 138340 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
2228 :
2229 : // only extend relation here. never decrease the size
2230 138340 : if nblocks > old_size {
2231 137394 : let buf = nblocks.to_le_bytes();
2232 137394 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2233 137394 :
2234 137394 : // Update relation size cache
2235 137394 : self.tline.set_cached_rel_size(rel, self.lsn, nblocks);
2236 137394 :
2237 137394 : self.pending_nblocks += nblocks as i64 - old_size as i64;
2238 137394 : }
2239 138340 : Ok(())
2240 138340 : }
2241 :
2242 : /// Drop some relations
2243 5 : pub(crate) async fn put_rel_drops(
2244 5 : &mut self,
2245 5 : drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
2246 5 : ctx: &RequestContext,
2247 5 : ) -> Result<(), WalIngestError> {
2248 5 : let v2_enabled = self
2249 5 : .maybe_enable_rel_size_v2()
2250 5 : .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
2251 6 : for ((spc_node, db_node), rel_tags) in drop_relations {
2252 1 : let dir_key = rel_dir_to_key(spc_node, db_node);
2253 1 : let buf = self.get(dir_key, ctx).await?;
2254 1 : let mut dir = RelDirectory::des(&buf)?;
2255 :
2256 1 : let mut dirty = false;
2257 2 : for rel_tag in rel_tags {
2258 1 : let found = if dir.rels.remove(&(rel_tag.relnode, rel_tag.forknum)) {
2259 1 : self.pending_directory_entries
2260 1 : .push((DirectoryKind::Rel, MetricsUpdate::Sub(1)));
2261 1 : dirty = true;
2262 1 : true
2263 0 : } else if v2_enabled {
2264 : // The rel is not found in the old reldir key, so we need to check the new sparse keyspace.
2265 : // Note that a relation can only exist in one of the two keyspaces (guaranteed by the ingestion
2266 : // logic).
2267 0 : let key =
2268 0 : rel_tag_sparse_key(spc_node, db_node, rel_tag.relnode, rel_tag.forknum);
2269 0 : let val = RelDirExists::decode_option(self.sparse_get(key, ctx).await?)
2270 0 : .map_err(|_| WalIngestErrorKind::InvalidKey(key, self.lsn))?;
2271 0 : if val == RelDirExists::Exists {
2272 0 : self.pending_directory_entries
2273 0 : .push((DirectoryKind::RelV2, MetricsUpdate::Sub(1)));
2274 : // put tombstone
2275 0 : self.put(key, Value::Image(RelDirExists::Removed.encode()));
2276 : // no need to set dirty to true
2277 0 : true
2278 : } else {
2279 0 : false
2280 : }
2281 : } else {
2282 0 : false
2283 : };
2284 :
2285 1 : if found {
2286 : // update logical size
2287 1 : let size_key = rel_size_to_key(rel_tag);
2288 1 : let old_size = self.get(size_key, ctx).await?.get_u32_le();
2289 1 : self.pending_nblocks -= old_size as i64;
2290 :
2291 : // Remove entry from relation size cache
2292 1 : self.tline.remove_cached_rel_size(&rel_tag);
2293 :
2294 : // Delete size entry, as well as all blocks; this is currently a no-op because we haven't implemented tombstones in storage.
2295 1 : self.delete(rel_key_range(rel_tag));
2296 0 : }
2297 : }
2298 :
2299 1 : if dirty {
2300 1 : self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
2301 0 : }
2302 : }
2303 :
2304 5 : Ok(())
2305 5 : }
2306 :
2307 3 : pub async fn put_slru_segment_creation(
2308 3 : &mut self,
2309 3 : kind: SlruKind,
2310 3 : segno: u32,
2311 3 : nblocks: BlockNumber,
2312 3 : ctx: &RequestContext,
2313 3 : ) -> Result<(), WalIngestError> {
2314 3 : assert!(self.tline.tenant_shard_id.is_shard_zero());
2315 :
2316 : // Add it to the directory entry
2317 3 : let dir_key = slru_dir_to_key(kind);
2318 3 : let buf = self.get(dir_key, ctx).await?;
2319 3 : let mut dir = SlruSegmentDirectory::des(&buf)?;
2320 :
2321 3 : if !dir.segments.insert(segno) {
2322 0 : Err(WalIngestErrorKind::SlruAlreadyExists(kind, segno))?;
2323 3 : }
2324 3 : self.pending_directory_entries.push((
2325 3 : DirectoryKind::SlruSegment(kind),
2326 3 : MetricsUpdate::Set(dir.segments.len() as u64),
2327 3 : ));
2328 3 : self.put(
2329 3 : dir_key,
2330 3 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
2331 : );
2332 :
2333 : // Put size
2334 3 : let size_key = slru_segment_size_to_key(kind, segno);
2335 3 : let buf = nblocks.to_le_bytes();
2336 3 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2337 :
2338 : // even if nblocks > 0, we don't insert any actual blocks here
2339 :
2340 3 : Ok(())
2341 3 : }
2342 :
2343 : /// Extend SLRU segment
2344 0 : pub fn put_slru_extend(
2345 0 : &mut self,
2346 0 : kind: SlruKind,
2347 0 : segno: u32,
2348 0 : nblocks: BlockNumber,
2349 0 : ) -> Result<(), WalIngestError> {
2350 0 : assert!(self.tline.tenant_shard_id.is_shard_zero());
2351 :
2352 : // Put size
2353 0 : let size_key = slru_segment_size_to_key(kind, segno);
2354 0 : let buf = nblocks.to_le_bytes();
2355 0 : self.put(size_key, Value::Image(Bytes::from(buf.to_vec())));
2356 0 : Ok(())
2357 0 : }
2358 :
2359 : /// This method is used for marking truncated SLRU files
2360 0 : pub async fn drop_slru_segment(
2361 0 : &mut self,
2362 0 : kind: SlruKind,
2363 0 : segno: u32,
2364 0 : ctx: &RequestContext,
2365 0 : ) -> Result<(), WalIngestError> {
2366 : // Remove it from the directory entry
2367 0 : let dir_key = slru_dir_to_key(kind);
2368 0 : let buf = self.get(dir_key, ctx).await?;
2369 0 : let mut dir = SlruSegmentDirectory::des(&buf)?;
2370 :
2371 0 : if !dir.segments.remove(&segno) {
2372 0 : warn!("slru segment {:?}/{} does not exist", kind, segno);
2373 0 : }
2374 0 : self.pending_directory_entries.push((
2375 0 : DirectoryKind::SlruSegment(kind),
2376 0 : MetricsUpdate::Set(dir.segments.len() as u64),
2377 0 : ));
2378 0 : self.put(
2379 0 : dir_key,
2380 0 : Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
2381 : );
2382 :
2383 : // Delete size entry, as well as all blocks
2384 0 : self.delete(slru_segment_key_range(kind, segno));
2385 :
2386 0 : Ok(())
2387 0 : }
2388 :
2389 : /// Drop a relmapper file (pg_filenode.map)
2390 0 : pub fn drop_relmap_file(&mut self, _spcnode: Oid, _dbnode: Oid) -> Result<(), WalIngestError> {
2391 : // TODO
2392 0 : Ok(())
2393 0 : }
2394 :
2395 : /// This method is used for marking truncated SLRU files
2396 0 : pub async fn drop_twophase_file(
2397 0 : &mut self,
2398 0 : xid: u64,
2399 0 : ctx: &RequestContext,
2400 0 : ) -> Result<(), WalIngestError> {
2401 : // Remove it from the directory entry
2402 0 : let buf = self.get(TWOPHASEDIR_KEY, ctx).await?;
2403 0 : let newdirbuf = if self.tline.pg_version >= PgMajorVersion::PG17 {
2404 0 : let mut dir = TwoPhaseDirectoryV17::des(&buf)?;
2405 :
2406 0 : if !dir.xids.remove(&xid) {
2407 0 : warn!("twophase file for xid {} does not exist", xid);
2408 0 : }
2409 0 : self.pending_directory_entries.push((
2410 0 : DirectoryKind::TwoPhase,
2411 0 : MetricsUpdate::Set(dir.xids.len() as u64),
2412 0 : ));
2413 0 : Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
2414 : } else {
2415 0 : let xid: u32 = u32::try_from(xid)
2416 0 : .map_err(|e| WalIngestErrorKind::LogicalError(anyhow::Error::from(e)))?;
2417 0 : let mut dir = TwoPhaseDirectory::des(&buf)?;
2418 :
2419 0 : if !dir.xids.remove(&xid) {
2420 0 : warn!("twophase file for xid {} does not exist", xid);
2421 0 : }
2422 0 : self.pending_directory_entries.push((
2423 0 : DirectoryKind::TwoPhase,
2424 0 : MetricsUpdate::Set(dir.xids.len() as u64),
2425 0 : ));
2426 0 : Bytes::from(TwoPhaseDirectory::ser(&dir)?)
2427 : };
2428 0 : self.put(TWOPHASEDIR_KEY, Value::Image(newdirbuf));
2429 :
2430 : // Delete it
2431 0 : self.delete(twophase_key_range(xid));
2432 :
2433 0 : Ok(())
2434 0 : }
2435 :
2436 8 : pub async fn put_file(
2437 8 : &mut self,
2438 8 : path: &str,
2439 8 : content: &[u8],
2440 8 : ctx: &RequestContext,
2441 8 : ) -> Result<(), WalIngestError> {
2442 8 : let key = aux_file::encode_aux_file_key(path);
2443 : // retrieve the key from the engine
2444 8 : let old_val = match self.get(key, ctx).await {
2445 2 : Ok(val) => Some(val),
2446 6 : Err(PageReconstructError::MissingKey(_)) => None,
2447 0 : Err(e) => return Err(e.into()),
2448 : };
2449 8 : let files: Vec<(&str, &[u8])> = if let Some(ref old_val) = old_val {
2450 2 : aux_file::decode_file_value(old_val).map_err(WalIngestErrorKind::EncodeAuxFileError)?
2451 : } else {
2452 6 : Vec::new()
2453 : };
2454 8 : let mut other_files = Vec::with_capacity(files.len());
2455 8 : let mut modifying_file = None;
2456 10 : for file @ (p, content) in files {
2457 2 : if path == p {
2458 2 : assert!(
2459 2 : modifying_file.is_none(),
2460 0 : "duplicated entries found for {path}"
2461 : );
2462 2 : modifying_file = Some(content);
2463 0 : } else {
2464 0 : other_files.push(file);
2465 0 : }
2466 : }
2467 8 : let mut new_files = other_files;
2468 8 : match (modifying_file, content.is_empty()) {
2469 1 : (Some(old_content), false) => {
2470 1 : self.tline
2471 1 : .aux_file_size_estimator
2472 1 : .on_update(old_content.len(), content.len());
2473 1 : new_files.push((path, content));
2474 1 : }
2475 1 : (Some(old_content), true) => {
2476 1 : self.tline
2477 1 : .aux_file_size_estimator
2478 1 : .on_remove(old_content.len());
2479 1 : // not adding the file key to the final `new_files` vec.
2480 1 : }
2481 6 : (None, false) => {
2482 6 : self.tline.aux_file_size_estimator.on_add(content.len());
2483 6 : new_files.push((path, content));
2484 6 : }
2485 : // Compute may request delete of old version of pgstat AUX file if new one exceeds size limit.
2486 : // Compute doesn't know if previous version of this file exists or not, so
2487 : // attempt to delete non-existing file can cause this message.
2488 : // To avoid false alarms, log it as info rather than warning.
2489 0 : (None, true) if path.starts_with("pg_stat/") => {
2490 0 : info!("removing non-existing pg_stat file: {}", path)
2491 : }
2492 0 : (None, true) => warn!("removing non-existing aux file: {}", path),
2493 : }
2494 8 : let new_val = aux_file::encode_file_value(&new_files)
2495 8 : .map_err(WalIngestErrorKind::EncodeAuxFileError)?;
2496 8 : self.put(key, Value::Image(new_val.into()));
2497 :
2498 8 : Ok(())
2499 8 : }
2500 :
2501 : ///
2502 : /// Flush changes accumulated so far to the underlying repository.
2503 : ///
2504 : /// Usually, changes made in DatadirModification are atomic, but this allows
2505 : /// you to flush them to the underlying repository before the final `commit`.
2506 : /// That allows to free up the memory used to hold the pending changes.
2507 : ///
2508 : /// Currently only used during bulk import of a data directory. In that
2509 : /// context, breaking the atomicity is OK. If the import is interrupted, the
2510 : /// whole import fails and the timeline will be deleted anyway.
2511 : /// (Or to be precise, it will be left behind for debugging purposes and
2512 : /// ignored, see <https://github.com/neondatabase/neon/pull/1809>)
2513 : ///
2514 : /// Note: A consequence of flushing the pending operations is that they
2515 : /// won't be visible to subsequent operations until `commit`. The function
2516 : /// retains all the metadata, but data pages are flushed. That's again OK
2517 : /// for bulk import, where you are just loading data pages and won't try to
2518 : /// modify the same pages twice.
2519 965 : pub(crate) async fn flush(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2520 : // Unless we have accumulated a decent amount of changes, it's not worth it
2521 : // to scan through the pending_updates list.
2522 965 : let pending_nblocks = self.pending_nblocks;
2523 965 : if pending_nblocks < 10000 {
2524 965 : return Ok(());
2525 0 : }
2526 :
2527 0 : let mut writer = self.tline.writer().await;
2528 :
2529 : // Flush relation and SLRU data blocks, keep metadata.
2530 0 : if let Some(batch) = self.pending_data_batch.take() {
2531 0 : tracing::debug!(
2532 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2533 : batch.max_lsn,
2534 0 : self.tline.get_last_record_lsn()
2535 : );
2536 :
2537 : // This bails out on first error without modifying pending_updates.
2538 : // That's Ok, cf this function's doc comment.
2539 0 : writer.put_batch(batch, ctx).await?;
2540 0 : }
2541 :
2542 0 : if pending_nblocks != 0 {
2543 0 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2544 0 : self.pending_nblocks = 0;
2545 0 : }
2546 :
2547 0 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2548 0 : writer.update_directory_entries_count(kind, count);
2549 0 : }
2550 :
2551 0 : Ok(())
2552 965 : }
2553 :
2554 : ///
2555 : /// Finish this atomic update, writing all the updated keys to the
2556 : /// underlying timeline.
2557 : /// All the modifications in this atomic update are stamped by the specified LSN.
2558 : ///
2559 371557 : pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
2560 371557 : let mut writer = self.tline.writer().await;
2561 :
2562 371557 : let pending_nblocks = self.pending_nblocks;
2563 371557 : self.pending_nblocks = 0;
2564 :
2565 : // Ordering: the items in this batch do not need to be in any global order, but values for
2566 : // a particular Key must be in Lsn order relative to one another. InMemoryLayer relies on
2567 : // this to do efficient updates to its index. See [`wal_decoder::serialized_batch`] for
2568 : // more details.
2569 :
2570 371557 : let metadata_batch = {
2571 371557 : let pending_meta = self
2572 371557 : .pending_metadata_pages
2573 371557 : .drain()
2574 371557 : .flat_map(|(key, values)| {
2575 137068 : values
2576 137068 : .into_iter()
2577 137068 : .map(move |(lsn, value_size, value)| (key, lsn, value_size, value))
2578 137068 : })
2579 371557 : .collect::<Vec<_>>();
2580 :
2581 371557 : if pending_meta.is_empty() {
2582 236139 : None
2583 : } else {
2584 135418 : Some(SerializedValueBatch::from_values(pending_meta))
2585 : }
2586 : };
2587 :
2588 371557 : let data_batch = self.pending_data_batch.take();
2589 :
2590 371557 : let maybe_batch = match (data_batch, metadata_batch) {
2591 132278 : (Some(mut data), Some(metadata)) => {
2592 132278 : data.extend(metadata);
2593 132278 : Some(data)
2594 : }
2595 71631 : (Some(data), None) => Some(data),
2596 3140 : (None, Some(metadata)) => Some(metadata),
2597 164508 : (None, None) => None,
2598 : };
2599 :
2600 371557 : if let Some(batch) = maybe_batch {
2601 207049 : tracing::debug!(
2602 0 : "Flushing batch with max_lsn={}. Last record LSN is {}",
2603 : batch.max_lsn,
2604 0 : self.tline.get_last_record_lsn()
2605 : );
2606 :
2607 : // This bails out on first error without modifying pending_updates.
2608 : // That's Ok, cf this function's doc comment.
2609 207049 : writer.put_batch(batch, ctx).await?;
2610 164508 : }
2611 :
2612 371557 : if !self.pending_deletions.is_empty() {
2613 1 : writer.delete_batch(&self.pending_deletions, ctx).await?;
2614 1 : self.pending_deletions.clear();
2615 371556 : }
2616 :
2617 371557 : self.pending_lsns.push(self.lsn);
2618 444486 : for pending_lsn in self.pending_lsns.drain(..) {
2619 444486 : // TODO(vlad): pretty sure the comment below is not valid anymore
2620 444486 : // and we can call finish write with the latest LSN
2621 444486 : //
2622 444486 : // Ideally, we should be able to call writer.finish_write() only once
2623 444486 : // with the highest LSN. However, the last_record_lsn variable in the
2624 444486 : // timeline keeps track of the latest LSN and the immediate previous LSN
2625 444486 : // so we need to record every LSN to not leave a gap between them.
2626 444486 : writer.finish_write(pending_lsn);
2627 444486 : }
2628 :
2629 371557 : if pending_nblocks != 0 {
2630 135285 : writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
2631 236272 : }
2632 :
2633 371557 : for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
2634 1527 : writer.update_directory_entries_count(kind, count);
2635 1527 : }
2636 :
2637 371557 : self.pending_metadata_bytes = 0;
2638 :
2639 371557 : Ok(())
2640 371557 : }
2641 :
2642 145852 : pub(crate) fn len(&self) -> usize {
2643 145852 : self.pending_metadata_pages.len()
2644 145852 : + self.pending_data_batch.as_ref().map_or(0, |b| b.len())
2645 145852 : + self.pending_deletions.len()
2646 145852 : }
2647 :
2648 : /// Read a page from the Timeline we are writing to. For metadata pages, this passes through
2649 : /// a cache in Self, which makes writes earlier in this modification visible to WAL records later
2650 : /// in the modification.
2651 : ///
2652 : /// For data pages, reads pass directly to the owning Timeline: any ingest code which reads a data
2653 : /// page must ensure that the pages they read are already committed in Timeline, for example
2654 : /// DB create operations are always preceded by a call to commit(). This is special cased because
2655 : /// it's rare: all the 'normal' WAL operations will only read metadata pages such as relation sizes,
2656 : /// and not data pages.
2657 143293 : async fn get(&self, key: Key, ctx: &RequestContext) -> Result<Bytes, PageReconstructError> {
2658 143293 : if !Self::is_data_key(&key) {
2659 : // Have we already updated the same key? Read the latest pending updated
2660 : // version in that case.
2661 : //
2662 : // Note: we don't check pending_deletions. It is an error to request a
2663 : // value that has been removed, deletion only avoids leaking storage.
2664 143293 : if let Some(values) = self.pending_metadata_pages.get(&key.to_compact()) {
2665 7964 : if let Some((_, _, value)) = values.last() {
2666 7964 : return if let Value::Image(img) = value {
2667 7964 : Ok(img.clone())
2668 : } else {
2669 : // Currently, we never need to read back a WAL record that we
2670 : // inserted in the same "transaction". All the metadata updates
2671 : // work directly with Images, and we never need to read actual
2672 : // data pages. We could handle this if we had to, by calling
2673 : // the walredo manager, but let's keep it simple for now.
2674 0 : Err(PageReconstructError::Other(anyhow::anyhow!(
2675 0 : "unexpected pending WAL record"
2676 0 : )))
2677 : };
2678 0 : }
2679 135329 : }
2680 : } else {
2681 : // This is an expensive check, so we only do it in debug mode. If reading a data key,
2682 : // this key should never be present in pending_data_pages. We ensure this by committing
2683 : // modifications before ingesting DB create operations, which are the only kind that reads
2684 : // data pages during ingest.
2685 0 : if cfg!(debug_assertions) {
2686 0 : assert!(
2687 0 : !self
2688 0 : .pending_data_batch
2689 0 : .as_ref()
2690 0 : .is_some_and(|b| b.updates_key(&key))
2691 : );
2692 0 : }
2693 : }
2694 :
2695 : // Metadata page cache miss, or we're reading a data page.
2696 135329 : let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
2697 135329 : self.tline.get(key, lsn, ctx).await
2698 143293 : }
2699 :
2700 : /// Get a key from the sparse keyspace. Automatically converts the missing key error
2701 : /// and the empty value into None.
2702 0 : async fn sparse_get(
2703 0 : &self,
2704 0 : key: Key,
2705 0 : ctx: &RequestContext,
2706 0 : ) -> Result<Option<Bytes>, PageReconstructError> {
2707 0 : let val = self.get(key, ctx).await;
2708 0 : match val {
2709 0 : Ok(val) if val.is_empty() => Ok(None),
2710 0 : Ok(val) => Ok(Some(val)),
2711 0 : Err(PageReconstructError::MissingKey(_)) => Ok(None),
2712 0 : Err(e) => Err(e),
2713 : }
2714 0 : }
2715 :
2716 : #[cfg(test)]
2717 2 : pub fn put_for_unit_test(&mut self, key: Key, val: Value) {
2718 2 : self.put(key, val);
2719 2 : }
2720 :
2721 282078 : fn put(&mut self, key: Key, val: Value) {
2722 282078 : if Self::is_data_key(&key) {
2723 138934 : self.put_data(key.to_compact(), val)
2724 : } else {
2725 143144 : self.put_metadata(key.to_compact(), val)
2726 : }
2727 282078 : }
2728 :
2729 138934 : fn put_data(&mut self, key: CompactKey, val: Value) {
2730 138934 : let batch = self
2731 138934 : .pending_data_batch
2732 138934 : .get_or_insert_with(SerializedValueBatch::default);
2733 138934 : batch.put(key, val, self.lsn);
2734 138934 : }
2735 :
2736 143144 : fn put_metadata(&mut self, key: CompactKey, val: Value) {
2737 143144 : let values = self.pending_metadata_pages.entry(key).or_default();
2738 : // Replace the previous value if it exists at the same lsn
2739 143144 : if let Some((last_lsn, last_value_ser_size, last_value)) = values.last_mut() {
2740 6076 : if *last_lsn == self.lsn {
2741 : // Update the pending_metadata_bytes contribution from this entry, and update the serialized size in place
2742 6076 : self.pending_metadata_bytes -= *last_value_ser_size;
2743 6076 : *last_value_ser_size = val.serialized_size().unwrap() as usize;
2744 6076 : self.pending_metadata_bytes += *last_value_ser_size;
2745 :
2746 : // Use the latest value, this replaces any earlier write to the same (key,lsn), such as much
2747 : // have been generated by synthesized zero page writes prior to the first real write to a page.
2748 6076 : *last_value = val;
2749 6076 : return;
2750 0 : }
2751 137068 : }
2752 :
2753 137068 : let val_serialized_size = val.serialized_size().unwrap() as usize;
2754 137068 : self.pending_metadata_bytes += val_serialized_size;
2755 137068 : values.push((self.lsn, val_serialized_size, val));
2756 :
2757 137068 : if key == CHECKPOINT_KEY.to_compact() {
2758 119 : tracing::debug!("Checkpoint key added to pending with size {val_serialized_size}");
2759 136949 : }
2760 143144 : }
2761 :
2762 1 : fn delete(&mut self, key_range: Range<Key>) {
2763 1 : trace!("DELETE {}-{}", key_range.start, key_range.end);
2764 1 : self.pending_deletions.push((key_range, self.lsn));
2765 1 : }
2766 : }
2767 :
2768 : /// Statistics for a DatadirModification.
2769 : #[derive(Default)]
2770 : pub struct DatadirModificationStats {
2771 : pub metadata_images: u64,
2772 : pub metadata_deltas: u64,
2773 : pub data_images: u64,
2774 : pub data_deltas: u64,
2775 : }
2776 :
2777 : /// This struct facilitates accessing either a committed key from the timeline at a
2778 : /// specific LSN, or the latest uncommitted key from a pending modification.
2779 : ///
2780 : /// During WAL ingestion, the records from multiple LSNs may be batched in the same
2781 : /// modification before being flushed to the timeline. Hence, the routines in WalIngest
2782 : /// need to look up the keys in the modification first before looking them up in the
2783 : /// timeline to not miss the latest updates.
2784 : #[derive(Clone, Copy)]
2785 : pub enum Version<'a> {
2786 : LsnRange(LsnRange),
2787 : Modified(&'a DatadirModification<'a>),
2788 : }
2789 :
2790 : impl Version<'_> {
2791 25 : async fn get(
2792 25 : &self,
2793 25 : timeline: &Timeline,
2794 25 : key: Key,
2795 25 : ctx: &RequestContext,
2796 25 : ) -> Result<Bytes, PageReconstructError> {
2797 25 : match self {
2798 15 : Version::LsnRange(lsns) => timeline.get(key, lsns.effective_lsn, ctx).await,
2799 10 : Version::Modified(modification) => modification.get(key, ctx).await,
2800 : }
2801 25 : }
2802 :
2803 : /// Get a key from the sparse keyspace. Automatically converts the missing key error
2804 : /// and the empty value into None.
2805 0 : async fn sparse_get(
2806 0 : &self,
2807 0 : timeline: &Timeline,
2808 0 : key: Key,
2809 0 : ctx: &RequestContext,
2810 0 : ) -> Result<Option<Bytes>, PageReconstructError> {
2811 0 : let val = self.get(timeline, key, ctx).await;
2812 0 : match val {
2813 0 : Ok(val) if val.is_empty() => Ok(None),
2814 0 : Ok(val) => Ok(Some(val)),
2815 0 : Err(PageReconstructError::MissingKey(_)) => Ok(None),
2816 0 : Err(e) => Err(e),
2817 : }
2818 0 : }
2819 :
2820 26 : pub fn is_latest(&self) -> bool {
2821 26 : match self {
2822 16 : Version::LsnRange(lsns) => lsns.is_latest(),
2823 10 : Version::Modified(_) => true,
2824 : }
2825 26 : }
2826 :
2827 224275 : pub fn get_lsn(&self) -> Lsn {
2828 224275 : match self {
2829 12224 : Version::LsnRange(lsns) => lsns.effective_lsn,
2830 212051 : Version::Modified(modification) => modification.lsn,
2831 : }
2832 224275 : }
2833 :
2834 12219 : pub fn at(lsn: Lsn) -> Self {
2835 12219 : Version::LsnRange(LsnRange {
2836 12219 : effective_lsn: lsn,
2837 12219 : request_lsn: lsn,
2838 12219 : })
2839 12219 : }
2840 : }
2841 :
2842 : //--- Metadata structs stored in key-value pairs in the repository.
2843 :
2844 0 : #[derive(Debug, Serialize, Deserialize)]
2845 : pub(crate) struct DbDirectory {
2846 : // (spcnode, dbnode) -> (do relmapper and PG_VERSION files exist)
2847 : pub(crate) dbdirs: HashMap<(Oid, Oid), bool>,
2848 : }
2849 :
2850 : // The format of TwoPhaseDirectory changed in PostgreSQL v17, because the filenames of
2851 : // pg_twophase files was expanded from 32-bit XIDs to 64-bit XIDs. Previously, the files
2852 : // were named like "pg_twophase/000002E5", now they're like
2853 : // "pg_twophsae/0000000A000002E4".
2854 :
2855 0 : #[derive(Debug, Serialize, Deserialize)]
2856 : pub(crate) struct TwoPhaseDirectory {
2857 : pub(crate) xids: HashSet<TransactionId>,
2858 : }
2859 :
2860 0 : #[derive(Debug, Serialize, Deserialize)]
2861 : struct TwoPhaseDirectoryV17 {
2862 : xids: HashSet<u64>,
2863 : }
2864 :
2865 0 : #[derive(Debug, Serialize, Deserialize, Default)]
2866 : pub(crate) struct RelDirectory {
2867 : // Set of relations that exist. (relfilenode, forknum)
2868 : //
2869 : // TODO: Store it as a btree or radix tree or something else that spans multiple
2870 : // key-value pairs, if you have a lot of relations
2871 : pub(crate) rels: HashSet<(Oid, u8)>,
2872 : }
2873 :
2874 0 : #[derive(Debug, Serialize, Deserialize)]
2875 : struct RelSizeEntry {
2876 : nblocks: u32,
2877 : }
2878 :
2879 0 : #[derive(Debug, Serialize, Deserialize, Default)]
2880 : pub(crate) struct SlruSegmentDirectory {
2881 : // Set of SLRU segments that exist.
2882 : pub(crate) segments: HashSet<u32>,
2883 : }
2884 :
2885 : #[derive(Copy, Clone, PartialEq, Eq, Debug, enum_map::Enum)]
2886 : #[repr(u8)]
2887 : pub(crate) enum DirectoryKind {
2888 : Db,
2889 : TwoPhase,
2890 : Rel,
2891 : AuxFiles,
2892 : SlruSegment(SlruKind),
2893 : RelV2,
2894 : }
2895 :
2896 : impl DirectoryKind {
2897 : pub(crate) const KINDS_NUM: usize = <DirectoryKind as Enum>::LENGTH;
2898 4582 : pub(crate) fn offset(&self) -> usize {
2899 4582 : self.into_usize()
2900 4582 : }
2901 : }
2902 :
2903 : static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
2904 :
2905 : #[allow(clippy::bool_assert_comparison)]
2906 : #[cfg(test)]
2907 : mod tests {
2908 : use hex_literal::hex;
2909 : use pageserver_api::models::ShardParameters;
2910 : use pageserver_api::shard::ShardStripeSize;
2911 : use utils::id::TimelineId;
2912 : use utils::shard::{ShardCount, ShardNumber};
2913 :
2914 : use super::*;
2915 : use crate::DEFAULT_PG_VERSION;
2916 : use crate::tenant::harness::TenantHarness;
2917 :
2918 : /// Test a round trip of aux file updates, from DatadirModification to reading back from the Timeline
2919 : #[tokio::test]
2920 1 : async fn aux_files_round_trip() -> anyhow::Result<()> {
2921 1 : let name = "aux_files_round_trip";
2922 1 : let harness = TenantHarness::create(name).await?;
2923 :
2924 : pub const TIMELINE_ID: TimelineId =
2925 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
2926 :
2927 1 : let (tenant, ctx) = harness.load().await;
2928 1 : let (tline, ctx) = tenant
2929 1 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2930 1 : .await?;
2931 1 : let tline = tline.raw_timeline().unwrap();
2932 :
2933 : // First modification: insert two keys
2934 1 : let mut modification = tline.begin_modification(Lsn(0x1000));
2935 1 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
2936 1 : modification.set_lsn(Lsn(0x1008))?;
2937 1 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
2938 1 : modification.commit(&ctx).await?;
2939 1 : let expect_1008 = HashMap::from([
2940 1 : ("foo/bar1".to_string(), Bytes::from_static(b"content1")),
2941 1 : ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
2942 1 : ]);
2943 :
2944 1 : let io_concurrency = IoConcurrency::spawn_for_test();
2945 :
2946 1 : let readback = tline
2947 1 : .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
2948 1 : .await?;
2949 1 : assert_eq!(readback, expect_1008);
2950 :
2951 : // Second modification: update one key, remove the other
2952 1 : let mut modification = tline.begin_modification(Lsn(0x2000));
2953 1 : modification.put_file("foo/bar1", b"content3", &ctx).await?;
2954 1 : modification.set_lsn(Lsn(0x2008))?;
2955 1 : modification.put_file("foo/bar2", b"", &ctx).await?;
2956 1 : modification.commit(&ctx).await?;
2957 1 : let expect_2008 =
2958 1 : HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
2959 :
2960 1 : let readback = tline
2961 1 : .list_aux_files(Lsn(0x2008), &ctx, io_concurrency.clone())
2962 1 : .await?;
2963 1 : assert_eq!(readback, expect_2008);
2964 :
2965 : // Reading back in time works
2966 1 : let readback = tline
2967 1 : .list_aux_files(Lsn(0x1008), &ctx, io_concurrency.clone())
2968 1 : .await?;
2969 1 : assert_eq!(readback, expect_1008);
2970 :
2971 2 : Ok(())
2972 1 : }
2973 :
2974 : #[test]
2975 1 : fn gap_finding() {
2976 1 : let rel = RelTag {
2977 1 : spcnode: 1663,
2978 1 : dbnode: 208101,
2979 1 : relnode: 2620,
2980 1 : forknum: 0,
2981 1 : };
2982 1 : let base_blkno = 1;
2983 :
2984 1 : let base_key = rel_block_to_key(rel, base_blkno);
2985 1 : let before_base_key = rel_block_to_key(rel, base_blkno - 1);
2986 :
2987 1 : let shard = ShardIdentity::unsharded();
2988 :
2989 1 : let mut previous_nblocks = 0;
2990 11 : for i in 0..10 {
2991 10 : let crnt_blkno = base_blkno + i;
2992 10 : let gaps = DatadirModification::find_gaps(rel, crnt_blkno, previous_nblocks, &shard);
2993 :
2994 10 : previous_nblocks = crnt_blkno + 1;
2995 :
2996 10 : if i == 0 {
2997 : // The first block we write is 1, so we should find the gap.
2998 1 : assert_eq!(gaps.unwrap(), KeySpace::single(before_base_key..base_key));
2999 : } else {
3000 9 : assert!(gaps.is_none());
3001 : }
3002 : }
3003 :
3004 : // This is an update to an already existing block. No gaps here.
3005 1 : let update_blkno = 5;
3006 1 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
3007 1 : assert!(gaps.is_none());
3008 :
3009 : // This is an update past the current end block.
3010 1 : let after_gap_blkno = 20;
3011 1 : let gaps = DatadirModification::find_gaps(rel, after_gap_blkno, previous_nblocks, &shard);
3012 :
3013 1 : let gap_start_key = rel_block_to_key(rel, previous_nblocks);
3014 1 : let after_gap_key = rel_block_to_key(rel, after_gap_blkno);
3015 1 : assert_eq!(
3016 1 : gaps.unwrap(),
3017 1 : KeySpace::single(gap_start_key..after_gap_key)
3018 : );
3019 1 : }
3020 :
3021 : #[test]
3022 1 : fn sharded_gap_finding() {
3023 1 : let rel = RelTag {
3024 1 : spcnode: 1663,
3025 1 : dbnode: 208101,
3026 1 : relnode: 2620,
3027 1 : forknum: 0,
3028 1 : };
3029 :
3030 1 : let first_blkno = 6;
3031 :
3032 : // This shard will get the even blocks
3033 1 : let shard = ShardIdentity::from_params(
3034 1 : ShardNumber(0),
3035 1 : ShardParameters {
3036 1 : count: ShardCount(2),
3037 1 : stripe_size: ShardStripeSize(1),
3038 1 : },
3039 : );
3040 :
3041 : // Only keys belonging to this shard are considered as gaps.
3042 1 : let mut previous_nblocks = 0;
3043 1 : let gaps =
3044 1 : DatadirModification::find_gaps(rel, first_blkno, previous_nblocks, &shard).unwrap();
3045 1 : assert!(!gaps.ranges.is_empty());
3046 3 : for gap_range in gaps.ranges {
3047 2 : let mut k = gap_range.start;
3048 4 : while k != gap_range.end {
3049 2 : assert_eq!(shard.get_shard_number(&k), shard.number);
3050 2 : k = k.next();
3051 : }
3052 : }
3053 :
3054 1 : previous_nblocks = first_blkno;
3055 :
3056 1 : let update_blkno = 2;
3057 1 : let gaps = DatadirModification::find_gaps(rel, update_blkno, previous_nblocks, &shard);
3058 1 : assert!(gaps.is_none());
3059 1 : }
3060 :
3061 : /*
3062 : fn assert_current_logical_size<R: Repository>(timeline: &DatadirTimeline<R>, lsn: Lsn) {
3063 : let incremental = timeline.get_current_logical_size();
3064 : let non_incremental = timeline
3065 : .get_current_logical_size_non_incremental(lsn)
3066 : .unwrap();
3067 : assert_eq!(incremental, non_incremental);
3068 : }
3069 : */
3070 :
3071 : /*
3072 : ///
3073 : /// Test list_rels() function, with branches and dropped relations
3074 : ///
3075 : #[test]
3076 : fn test_list_rels_drop() -> Result<()> {
3077 : let repo = RepoHarness::create("test_list_rels_drop")?.load();
3078 : let tline = create_empty_timeline(repo, TIMELINE_ID)?;
3079 : const TESTDB: u32 = 111;
3080 :
3081 : // Import initial dummy checkpoint record, otherwise the get_timeline() call
3082 : // after branching fails below
3083 : let mut writer = tline.begin_record(Lsn(0x10));
3084 : writer.put_checkpoint(ZERO_CHECKPOINT.clone())?;
3085 : writer.finish()?;
3086 :
3087 : // Create a relation on the timeline
3088 : let mut writer = tline.begin_record(Lsn(0x20));
3089 : writer.put_rel_page_image(TESTREL_A, 0, TEST_IMG("foo blk 0 at 2"))?;
3090 : writer.finish()?;
3091 :
3092 : let writer = tline.begin_record(Lsn(0x00));
3093 : writer.finish()?;
3094 :
3095 : // Check that list_rels() lists it after LSN 2, but no before it
3096 : assert!(!tline.list_rels(0, TESTDB, Lsn(0x10))?.contains(&TESTREL_A));
3097 : assert!(tline.list_rels(0, TESTDB, Lsn(0x20))?.contains(&TESTREL_A));
3098 : assert!(tline.list_rels(0, TESTDB, Lsn(0x30))?.contains(&TESTREL_A));
3099 :
3100 : // Create a branch, check that the relation is visible there
3101 : repo.branch_timeline(&tline, NEW_TIMELINE_ID, Lsn(0x30))?;
3102 : let newtline = match repo.get_timeline(NEW_TIMELINE_ID)?.local_timeline() {
3103 : Some(timeline) => timeline,
3104 : None => panic!("Should have a local timeline"),
3105 : };
3106 : let newtline = DatadirTimelineImpl::new(newtline);
3107 : assert!(newtline
3108 : .list_rels(0, TESTDB, Lsn(0x30))?
3109 : .contains(&TESTREL_A));
3110 :
3111 : // Drop it on the branch
3112 : let mut new_writer = newtline.begin_record(Lsn(0x40));
3113 : new_writer.drop_relation(TESTREL_A)?;
3114 : new_writer.finish()?;
3115 :
3116 : // Check that it's no longer listed on the branch after the point where it was dropped
3117 : assert!(newtline
3118 : .list_rels(0, TESTDB, Lsn(0x30))?
3119 : .contains(&TESTREL_A));
3120 : assert!(!newtline
3121 : .list_rels(0, TESTDB, Lsn(0x40))?
3122 : .contains(&TESTREL_A));
3123 :
3124 : // Run checkpoint and garbage collection and check that it's still not visible
3125 : newtline.checkpoint(CheckpointConfig::Forced)?;
3126 : repo.gc_iteration(Some(NEW_TIMELINE_ID), 0, true)?;
3127 :
3128 : assert!(!newtline
3129 : .list_rels(0, TESTDB, Lsn(0x40))?
3130 : .contains(&TESTREL_A));
3131 :
3132 : Ok(())
3133 : }
3134 : */
3135 :
3136 : /*
3137 : #[test]
3138 : fn test_read_beyond_eof() -> Result<()> {
3139 : let repo = RepoHarness::create("test_read_beyond_eof")?.load();
3140 : let tline = create_test_timeline(repo, TIMELINE_ID)?;
3141 :
3142 : make_some_layers(&tline, Lsn(0x20))?;
3143 : let mut writer = tline.begin_record(Lsn(0x60));
3144 : walingest.put_rel_page_image(
3145 : &mut writer,
3146 : TESTREL_A,
3147 : 0,
3148 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x60))),
3149 : )?;
3150 : writer.finish()?;
3151 :
3152 : // Test read before rel creation. Should error out.
3153 : assert!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x10), false).is_err());
3154 :
3155 : // Read block beyond end of relation at different points in time.
3156 : // These reads should fall into different delta, image, and in-memory layers.
3157 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x20), false)?, ZERO_PAGE);
3158 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x25), false)?, ZERO_PAGE);
3159 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x30), false)?, ZERO_PAGE);
3160 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x35), false)?, ZERO_PAGE);
3161 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x40), false)?, ZERO_PAGE);
3162 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x45), false)?, ZERO_PAGE);
3163 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x50), false)?, ZERO_PAGE);
3164 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x55), false)?, ZERO_PAGE);
3165 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_A, 1, Lsn(0x60), false)?, ZERO_PAGE);
3166 :
3167 : // Test on an in-memory layer with no preceding layer
3168 : let mut writer = tline.begin_record(Lsn(0x70));
3169 : walingest.put_rel_page_image(
3170 : &mut writer,
3171 : TESTREL_B,
3172 : 0,
3173 : TEST_IMG(&format!("foo blk 0 at {}", Lsn(0x70))),
3174 : )?;
3175 : writer.finish()?;
3176 :
3177 : assert_eq!(tline.get_rel_page_at_lsn(TESTREL_B, 1, Lsn(0x70), false)?6, ZERO_PAGE);
3178 :
3179 : Ok(())
3180 : }
3181 : */
3182 : }
|
