Line data Source code
1 : // Copyright 2011 The LevelDB-Go and Pebble Authors. All rights reserved. Use
2 : // of this source code is governed by a BSD-style license that can be found in
3 : // the LICENSE file.
4 :
5 : package sstable
6 :
7 : import (
8 : "bytes"
9 : "cmp"
10 : "context"
11 : "encoding/binary"
12 : "io"
13 : "os"
14 : "slices"
15 : "time"
16 :
17 : "github.com/cespare/xxhash/v2"
18 : "github.com/cockroachdb/errors"
19 : "github.com/cockroachdb/pebble/internal/base"
20 : "github.com/cockroachdb/pebble/internal/bytealloc"
21 : "github.com/cockroachdb/pebble/internal/cache"
22 : "github.com/cockroachdb/pebble/internal/crc"
23 : "github.com/cockroachdb/pebble/internal/invariants"
24 : "github.com/cockroachdb/pebble/internal/keyspan"
25 : "github.com/cockroachdb/pebble/internal/private"
26 : "github.com/cockroachdb/pebble/objstorage"
27 : "github.com/cockroachdb/pebble/objstorage/objstorageprovider/objiotracing"
28 : )
29 :
30 : var errCorruptIndexEntry = base.CorruptionErrorf("pebble/table: corrupt index entry")
31 : var errReaderClosed = errors.New("pebble/table: reader is closed")
32 :
33 : // decodeBlockHandle returns the block handle encoded at the start of src, as
34 : // well as the number of bytes it occupies. It returns zero if given invalid
35 : // input. A block handle for a data block or a first/lower level index block
36 : // should not be decoded using decodeBlockHandle since the caller may validate
37 : // that the number of bytes decoded is equal to the length of src, which will
38 : // be false if the properties are not decoded. In those cases the caller
39 : // should use decodeBlockHandleWithProperties.
40 1 : func decodeBlockHandle(src []byte) (BlockHandle, int) {
41 1 : offset, n := binary.Uvarint(src)
42 1 : length, m := binary.Uvarint(src[n:])
43 1 : if n == 0 || m == 0 {
44 0 : return BlockHandle{}, 0
45 0 : }
46 1 : return BlockHandle{offset, length}, n + m
47 : }
48 :
49 : // decodeBlockHandleWithProperties returns the block handle and properties
50 : // encoded in src. src needs to be exactly the length that was encoded. This
51 : // method must be used for data block and first/lower level index blocks. The
52 : // properties in the block handle point to the bytes in src.
53 1 : func decodeBlockHandleWithProperties(src []byte) (BlockHandleWithProperties, error) {
54 1 : bh, n := decodeBlockHandle(src)
55 1 : if n == 0 {
56 0 : return BlockHandleWithProperties{}, errors.Errorf("invalid BlockHandle")
57 0 : }
58 1 : return BlockHandleWithProperties{
59 1 : BlockHandle: bh,
60 1 : Props: src[n:],
61 1 : }, nil
62 : }
63 :
64 1 : func encodeBlockHandle(dst []byte, b BlockHandle) int {
65 1 : n := binary.PutUvarint(dst, b.Offset)
66 1 : m := binary.PutUvarint(dst[n:], b.Length)
67 1 : return n + m
68 1 : }
69 :
70 1 : func encodeBlockHandleWithProperties(dst []byte, b BlockHandleWithProperties) []byte {
71 1 : n := encodeBlockHandle(dst, b.BlockHandle)
72 1 : dst = append(dst[:n], b.Props...)
73 1 : return dst
74 1 : }
75 :
76 : // block is a []byte that holds a sequence of key/value pairs plus an index
77 : // over those pairs.
78 : type block []byte
79 :
80 : type loadBlockResult int8
81 :
82 : const (
83 : loadBlockOK loadBlockResult = iota
84 : // Could be due to error or because no block left to load.
85 : loadBlockFailed
86 : loadBlockIrrelevant
87 : )
88 :
89 : type blockTransform func([]byte) ([]byte, error)
90 :
91 : // ReaderOption provide an interface to do work on Reader while it is being
92 : // opened.
93 : type ReaderOption interface {
94 : // readerApply is called on the reader during opening in order to set internal
95 : // parameters.
96 : readerApply(*Reader)
97 : }
98 :
99 : // Comparers is a map from comparer name to comparer. It is used for debugging
100 : // tools which may be used on multiple databases configured with different
101 : // comparers. Comparers implements the OpenOption interface and can be passed
102 : // as a parameter to NewReader.
103 : type Comparers map[string]*Comparer
104 :
105 1 : func (c Comparers) readerApply(r *Reader) {
106 1 : if r.Compare != nil || r.Properties.ComparerName == "" {
107 1 : return
108 1 : }
109 1 : if comparer, ok := c[r.Properties.ComparerName]; ok {
110 1 : r.Compare = comparer.Compare
111 1 : r.Equal = comparer.Equal
112 1 : r.FormatKey = comparer.FormatKey
113 1 : r.Split = comparer.Split
114 1 : }
115 : }
116 :
117 : // Mergers is a map from merger name to merger. It is used for debugging tools
118 : // which may be used on multiple databases configured with different
119 : // mergers. Mergers implements the OpenOption interface and can be passed as
120 : // a parameter to NewReader.
121 : type Mergers map[string]*Merger
122 :
123 1 : func (m Mergers) readerApply(r *Reader) {
124 1 : if r.mergerOK || r.Properties.MergerName == "" {
125 1 : return
126 1 : }
127 1 : _, r.mergerOK = m[r.Properties.MergerName]
128 : }
129 :
130 : // cacheOpts is a Reader open option for specifying the cache ID and sstable file
131 : // number. If not specified, a unique cache ID will be used.
132 : type cacheOpts struct {
133 : cacheID uint64
134 : fileNum base.DiskFileNum
135 : }
136 :
137 : // Marker function to indicate the option should be applied before reading the
138 : // sstable properties and, in the write path, before writing the default
139 : // sstable properties.
140 0 : func (c *cacheOpts) preApply() {}
141 :
142 1 : func (c *cacheOpts) readerApply(r *Reader) {
143 1 : if r.cacheID == 0 {
144 1 : r.cacheID = c.cacheID
145 1 : }
146 1 : if r.fileNum == 0 {
147 1 : r.fileNum = c.fileNum
148 1 : }
149 : }
150 :
151 1 : func (c *cacheOpts) writerApply(w *Writer) {
152 1 : if w.cacheID == 0 {
153 1 : w.cacheID = c.cacheID
154 1 : }
155 1 : if w.fileNum == 0 {
156 1 : w.fileNum = c.fileNum
157 1 : }
158 : }
159 :
160 : // rawTombstonesOpt is a Reader open option for specifying that range
161 : // tombstones returned by Reader.NewRangeDelIter() should not be
162 : // fragmented. Used by debug tools to get a raw view of the tombstones
163 : // contained in an sstable.
164 : type rawTombstonesOpt struct{}
165 :
166 0 : func (rawTombstonesOpt) preApply() {}
167 :
168 1 : func (rawTombstonesOpt) readerApply(r *Reader) {
169 1 : r.rawTombstones = true
170 1 : }
171 :
172 1 : func init() {
173 1 : private.SSTableCacheOpts = func(cacheID uint64, fileNum base.DiskFileNum) interface{} {
174 1 : return &cacheOpts{cacheID, fileNum}
175 1 : }
176 1 : private.SSTableRawTombstonesOpt = rawTombstonesOpt{}
177 : }
178 :
179 : // CommonReader abstracts functionality over a Reader or a VirtualReader. This
180 : // can be used by code which doesn't care to distinguish between a reader and a
181 : // virtual reader.
182 : type CommonReader interface {
183 : NewRawRangeKeyIter() (keyspan.FragmentIterator, error)
184 : NewRawRangeDelIter() (keyspan.FragmentIterator, error)
185 : NewIterWithBlockPropertyFiltersAndContextEtc(
186 : ctx context.Context, lower, upper []byte,
187 : filterer *BlockPropertiesFilterer,
188 : hideObsoletePoints, useFilterBlock bool,
189 : stats *base.InternalIteratorStats,
190 : categoryAndQoS CategoryAndQoS,
191 : statsCollector *CategoryStatsCollector,
192 : rp ReaderProvider,
193 : ) (Iterator, error)
194 : NewCompactionIter(
195 : bytesIterated *uint64,
196 : categoryAndQoS CategoryAndQoS,
197 : statsCollector *CategoryStatsCollector,
198 : rp ReaderProvider,
199 : bufferPool *BufferPool,
200 : ) (Iterator, error)
201 : EstimateDiskUsage(start, end []byte) (uint64, error)
202 : CommonProperties() *CommonProperties
203 : }
204 :
205 : // Reader is a table reader.
206 : type Reader struct {
207 : readable objstorage.Readable
208 : cacheID uint64
209 : fileNum base.DiskFileNum
210 : err error
211 : indexBH BlockHandle
212 : filterBH BlockHandle
213 : rangeDelBH BlockHandle
214 : rangeKeyBH BlockHandle
215 : rangeDelTransform blockTransform
216 : valueBIH valueBlocksIndexHandle
217 : propertiesBH BlockHandle
218 : metaIndexBH BlockHandle
219 : footerBH BlockHandle
220 : opts ReaderOptions
221 : Compare Compare
222 : Equal Equal
223 : FormatKey base.FormatKey
224 : Split Split
225 : tableFilter *tableFilterReader
226 : // Keep types that are not multiples of 8 bytes at the end and with
227 : // decreasing size.
228 : Properties Properties
229 : tableFormat TableFormat
230 : rawTombstones bool
231 : mergerOK bool
232 : checksumType ChecksumType
233 : // metaBufferPool is a buffer pool used exclusively when opening a table and
234 : // loading its meta blocks. metaBufferPoolAlloc is used to batch-allocate
235 : // the BufferPool.pool slice as a part of the Reader allocation. It's
236 : // capacity 3 to accommodate the meta block (1), and both the compressed
237 : // properties block (1) and decompressed properties block (1)
238 : // simultaneously.
239 : metaBufferPool BufferPool
240 : metaBufferPoolAlloc [3]allocedBuffer
241 : }
242 :
243 : // Close implements DB.Close, as documented in the pebble package.
244 1 : func (r *Reader) Close() error {
245 1 : r.opts.Cache.Unref()
246 1 :
247 1 : if r.readable != nil {
248 1 : r.err = firstError(r.err, r.readable.Close())
249 1 : r.readable = nil
250 1 : }
251 :
252 1 : if r.err != nil {
253 1 : return r.err
254 1 : }
255 : // Make any future calls to Get, NewIter or Close return an error.
256 1 : r.err = errReaderClosed
257 1 : return nil
258 : }
259 :
260 : // NewIterWithBlockPropertyFilters returns an iterator for the contents of the
261 : // table. If an error occurs, NewIterWithBlockPropertyFilters cleans up after
262 : // itself and returns a nil iterator.
263 : func (r *Reader) NewIterWithBlockPropertyFilters(
264 : lower, upper []byte,
265 : filterer *BlockPropertiesFilterer,
266 : useFilterBlock bool,
267 : stats *base.InternalIteratorStats,
268 : categoryAndQoS CategoryAndQoS,
269 : statsCollector *CategoryStatsCollector,
270 : rp ReaderProvider,
271 1 : ) (Iterator, error) {
272 1 : return r.newIterWithBlockPropertyFiltersAndContext(
273 1 : context.Background(), lower, upper, filterer, false, useFilterBlock, stats,
274 1 : categoryAndQoS, statsCollector, rp, nil)
275 1 : }
276 :
277 : // NewIterWithBlockPropertyFiltersAndContextEtc is similar to
278 : // NewIterWithBlockPropertyFilters and additionally accepts a context for
279 : // tracing.
280 : //
281 : // If hideObsoletePoints, the callee assumes that filterer already includes
282 : // obsoleteKeyBlockPropertyFilter. The caller can satisfy this contract by
283 : // first calling TryAddBlockPropertyFilterForHideObsoletePoints.
284 : func (r *Reader) NewIterWithBlockPropertyFiltersAndContextEtc(
285 : ctx context.Context,
286 : lower, upper []byte,
287 : filterer *BlockPropertiesFilterer,
288 : hideObsoletePoints, useFilterBlock bool,
289 : stats *base.InternalIteratorStats,
290 : categoryAndQoS CategoryAndQoS,
291 : statsCollector *CategoryStatsCollector,
292 : rp ReaderProvider,
293 1 : ) (Iterator, error) {
294 1 : return r.newIterWithBlockPropertyFiltersAndContext(
295 1 : ctx, lower, upper, filterer, hideObsoletePoints, useFilterBlock, stats, categoryAndQoS,
296 1 : statsCollector, rp, nil)
297 1 : }
298 :
299 : // TryAddBlockPropertyFilterForHideObsoletePoints is expected to be called
300 : // before the call to NewIterWithBlockPropertyFiltersAndContextEtc, to get the
301 : // value of hideObsoletePoints and potentially add a block property filter.
302 : func (r *Reader) TryAddBlockPropertyFilterForHideObsoletePoints(
303 : snapshotForHideObsoletePoints uint64,
304 : fileLargestSeqNum uint64,
305 : pointKeyFilters []BlockPropertyFilter,
306 1 : ) (hideObsoletePoints bool, filters []BlockPropertyFilter) {
307 1 : hideObsoletePoints = r.tableFormat >= TableFormatPebblev4 &&
308 1 : snapshotForHideObsoletePoints > fileLargestSeqNum
309 1 : if hideObsoletePoints {
310 1 : pointKeyFilters = append(pointKeyFilters, obsoleteKeyBlockPropertyFilter{})
311 1 : }
312 1 : return hideObsoletePoints, pointKeyFilters
313 : }
314 :
315 : func (r *Reader) newIterWithBlockPropertyFiltersAndContext(
316 : ctx context.Context,
317 : lower, upper []byte,
318 : filterer *BlockPropertiesFilterer,
319 : hideObsoletePoints bool,
320 : useFilterBlock bool,
321 : stats *base.InternalIteratorStats,
322 : categoryAndQoS CategoryAndQoS,
323 : statsCollector *CategoryStatsCollector,
324 : rp ReaderProvider,
325 : v *virtualState,
326 1 : ) (Iterator, error) {
327 1 : // NB: pebble.tableCache wraps the returned iterator with one which performs
328 1 : // reference counting on the Reader, preventing the Reader from being closed
329 1 : // until the final iterator closes.
330 1 : if r.Properties.IndexType == twoLevelIndex {
331 1 : i := twoLevelIterPool.Get().(*twoLevelIterator)
332 1 : err := i.init(ctx, r, v, lower, upper, filterer, useFilterBlock, hideObsoletePoints, stats,
333 1 : categoryAndQoS, statsCollector, rp, nil /* bufferPool */)
334 1 : if err != nil {
335 1 : return nil, err
336 1 : }
337 1 : return i, nil
338 : }
339 :
340 1 : i := singleLevelIterPool.Get().(*singleLevelIterator)
341 1 : err := i.init(ctx, r, v, lower, upper, filterer, useFilterBlock, hideObsoletePoints, stats,
342 1 : categoryAndQoS, statsCollector, rp, nil /* bufferPool */)
343 1 : if err != nil {
344 1 : return nil, err
345 1 : }
346 1 : return i, nil
347 : }
348 :
349 : // NewIter returns an iterator for the contents of the table. If an error
350 : // occurs, NewIter cleans up after itself and returns a nil iterator. NewIter
351 : // must only be used when the Reader is guaranteed to outlive any LazyValues
352 : // returned from the iter.
353 1 : func (r *Reader) NewIter(lower, upper []byte) (Iterator, error) {
354 1 : return r.NewIterWithBlockPropertyFilters(
355 1 : lower, upper, nil, true /* useFilterBlock */, nil, /* stats */
356 1 : CategoryAndQoS{}, nil /*statsCollector */, TrivialReaderProvider{Reader: r})
357 1 : }
358 :
359 : // NewCompactionIter returns an iterator similar to NewIter but it also increments
360 : // the number of bytes iterated. If an error occurs, NewCompactionIter cleans up
361 : // after itself and returns a nil iterator.
362 : func (r *Reader) NewCompactionIter(
363 : bytesIterated *uint64,
364 : categoryAndQoS CategoryAndQoS,
365 : statsCollector *CategoryStatsCollector,
366 : rp ReaderProvider,
367 : bufferPool *BufferPool,
368 1 : ) (Iterator, error) {
369 1 : return r.newCompactionIter(bytesIterated, categoryAndQoS, statsCollector, rp, nil, bufferPool)
370 1 : }
371 :
372 : func (r *Reader) newCompactionIter(
373 : bytesIterated *uint64,
374 : categoryAndQoS CategoryAndQoS,
375 : statsCollector *CategoryStatsCollector,
376 : rp ReaderProvider,
377 : v *virtualState,
378 : bufferPool *BufferPool,
379 1 : ) (Iterator, error) {
380 1 : if r.Properties.IndexType == twoLevelIndex {
381 1 : i := twoLevelIterPool.Get().(*twoLevelIterator)
382 1 : err := i.init(
383 1 : context.Background(),
384 1 : r, v, nil /* lower */, nil /* upper */, nil,
385 1 : false /* useFilter */, v != nil && v.isSharedIngested, /* hideObsoletePoints */
386 1 : nil /* stats */, categoryAndQoS, statsCollector, rp, bufferPool,
387 1 : )
388 1 : if err != nil {
389 0 : return nil, err
390 0 : }
391 1 : i.setupForCompaction()
392 1 : return &twoLevelCompactionIterator{
393 1 : twoLevelIterator: i,
394 1 : bytesIterated: bytesIterated,
395 1 : }, nil
396 : }
397 1 : i := singleLevelIterPool.Get().(*singleLevelIterator)
398 1 : err := i.init(
399 1 : context.Background(), r, v, nil /* lower */, nil, /* upper */
400 1 : nil, false /* useFilter */, v != nil && v.isSharedIngested, /* hideObsoletePoints */
401 1 : nil /* stats */, categoryAndQoS, statsCollector, rp, bufferPool,
402 1 : )
403 1 : if err != nil {
404 0 : return nil, err
405 0 : }
406 1 : i.setupForCompaction()
407 1 : return &compactionIterator{
408 1 : singleLevelIterator: i,
409 1 : bytesIterated: bytesIterated,
410 1 : }, nil
411 : }
412 :
413 : // NewRawRangeDelIter returns an internal iterator for the contents of the
414 : // range-del block for the table. Returns nil if the table does not contain
415 : // any range deletions.
416 : //
417 : // TODO(sumeer): plumb context.Context since this path is relevant in the user-facing
418 : // iterator. Add WithContext methods since the existing ones are public.
419 1 : func (r *Reader) NewRawRangeDelIter() (keyspan.FragmentIterator, error) {
420 1 : if r.rangeDelBH.Length == 0 {
421 1 : return nil, nil
422 1 : }
423 1 : h, err := r.readRangeDel(nil /* stats */, nil /* iterStats */)
424 1 : if err != nil {
425 1 : return nil, err
426 1 : }
427 1 : i := &fragmentBlockIter{elideSameSeqnum: true}
428 1 : // It's okay for hideObsoletePoints to be false here, even for shared ingested
429 1 : // sstables. This is because rangedels do not apply to points in the same
430 1 : // sstable at the same sequence number anyway, so exposing obsolete rangedels
431 1 : // is harmless.
432 1 : if err := i.blockIter.initHandle(r.Compare, h, r.Properties.GlobalSeqNum, false); err != nil {
433 0 : return nil, err
434 0 : }
435 1 : return i, nil
436 : }
437 :
438 1 : func (r *Reader) newRawRangeKeyIter(vState *virtualState) (keyspan.FragmentIterator, error) {
439 1 : if r.rangeKeyBH.Length == 0 {
440 1 : return nil, nil
441 1 : }
442 1 : h, err := r.readRangeKey(nil /* stats */, nil /* iterStats */)
443 1 : if err != nil {
444 1 : return nil, err
445 1 : }
446 1 : i := rangeKeyFragmentBlockIterPool.Get().(*rangeKeyFragmentBlockIter)
447 1 : var globalSeqNum uint64
448 1 : // Don't pass a global sequence number for shared ingested sstables. The
449 1 : // virtual reader needs to know the materialized sequence numbers, and will
450 1 : // do the appropriate sequence number substitution.
451 1 : if vState == nil || !vState.isSharedIngested {
452 1 : globalSeqNum = r.Properties.GlobalSeqNum
453 1 : }
454 1 : if err := i.blockIter.initHandle(r.Compare, h, globalSeqNum, false /* hideObsoletePoints */); err != nil {
455 0 : return nil, err
456 0 : }
457 1 : return i, nil
458 : }
459 :
460 : // NewRawRangeKeyIter returns an internal iterator for the contents of the
461 : // range-key block for the table. Returns nil if the table does not contain any
462 : // range keys.
463 : //
464 : // TODO(sumeer): plumb context.Context since this path is relevant in the user-facing
465 : // iterator. Add WithContext methods since the existing ones are public.
466 1 : func (r *Reader) NewRawRangeKeyIter() (keyspan.FragmentIterator, error) {
467 1 : return r.newRawRangeKeyIter(nil /* vState */)
468 1 : }
469 :
470 : type rangeKeyFragmentBlockIter struct {
471 : fragmentBlockIter
472 : }
473 :
474 1 : func (i *rangeKeyFragmentBlockIter) Close() error {
475 1 : err := i.fragmentBlockIter.Close()
476 1 : i.fragmentBlockIter = i.fragmentBlockIter.resetForReuse()
477 1 : rangeKeyFragmentBlockIterPool.Put(i)
478 1 : return err
479 1 : }
480 :
481 : func (r *Reader) readIndex(
482 : ctx context.Context, stats *base.InternalIteratorStats, iterStats *iterStatsAccumulator,
483 1 : ) (bufferHandle, error) {
484 1 : ctx = objiotracing.WithBlockType(ctx, objiotracing.MetadataBlock)
485 1 : return r.readBlock(ctx, r.indexBH, nil, nil, stats, iterStats, nil /* buffer pool */)
486 1 : }
487 :
488 : func (r *Reader) readFilter(
489 : ctx context.Context, stats *base.InternalIteratorStats, iterStats *iterStatsAccumulator,
490 1 : ) (bufferHandle, error) {
491 1 : ctx = objiotracing.WithBlockType(ctx, objiotracing.FilterBlock)
492 1 : return r.readBlock(ctx, r.filterBH, nil /* transform */, nil /* readHandle */, stats, iterStats, nil /* buffer pool */)
493 1 : }
494 :
495 : func (r *Reader) readRangeDel(
496 : stats *base.InternalIteratorStats, iterStats *iterStatsAccumulator,
497 1 : ) (bufferHandle, error) {
498 1 : ctx := objiotracing.WithBlockType(context.Background(), objiotracing.MetadataBlock)
499 1 : return r.readBlock(ctx, r.rangeDelBH, r.rangeDelTransform, nil /* readHandle */, stats, iterStats, nil /* buffer pool */)
500 1 : }
501 :
502 : func (r *Reader) readRangeKey(
503 : stats *base.InternalIteratorStats, iterStats *iterStatsAccumulator,
504 1 : ) (bufferHandle, error) {
505 1 : ctx := objiotracing.WithBlockType(context.Background(), objiotracing.MetadataBlock)
506 1 : return r.readBlock(ctx, r.rangeKeyBH, nil /* transform */, nil /* readHandle */, stats, iterStats, nil /* buffer pool */)
507 1 : }
508 :
509 : func checkChecksum(
510 : checksumType ChecksumType, b []byte, bh BlockHandle, fileNum base.DiskFileNum,
511 1 : ) error {
512 1 : expectedChecksum := binary.LittleEndian.Uint32(b[bh.Length+1:])
513 1 : var computedChecksum uint32
514 1 : switch checksumType {
515 1 : case ChecksumTypeCRC32c:
516 1 : computedChecksum = crc.New(b[:bh.Length+1]).Value()
517 1 : case ChecksumTypeXXHash64:
518 1 : computedChecksum = uint32(xxhash.Sum64(b[:bh.Length+1]))
519 0 : default:
520 0 : return errors.Errorf("unsupported checksum type: %d", checksumType)
521 : }
522 :
523 1 : if expectedChecksum != computedChecksum {
524 1 : return base.CorruptionErrorf(
525 1 : "pebble/table: invalid table %s (checksum mismatch at %d/%d)",
526 1 : fileNum, errors.Safe(bh.Offset), errors.Safe(bh.Length))
527 1 : }
528 1 : return nil
529 : }
530 :
531 : type cacheValueOrBuf struct {
532 : // buf.Valid() returns true if backed by a BufferPool.
533 : buf Buf
534 : // v is non-nil if backed by the block cache.
535 : v *cache.Value
536 : }
537 :
538 1 : func (b cacheValueOrBuf) get() []byte {
539 1 : if b.buf.Valid() {
540 1 : return b.buf.p.pool[b.buf.i].b
541 1 : }
542 1 : return b.v.Buf()
543 : }
544 :
545 1 : func (b cacheValueOrBuf) release() {
546 1 : if b.buf.Valid() {
547 1 : b.buf.Release()
548 1 : } else {
549 1 : cache.Free(b.v)
550 1 : }
551 : }
552 :
553 1 : func (b cacheValueOrBuf) truncate(n int) {
554 1 : if b.buf.Valid() {
555 1 : b.buf.p.pool[b.buf.i].b = b.buf.p.pool[b.buf.i].b[:n]
556 1 : } else {
557 1 : b.v.Truncate(n)
558 1 : }
559 : }
560 :
561 : // DeterministicReadBlockDurationForTesting is for tests that want a
562 : // deterministic value of the time to read a block (that is not in the cache).
563 : // The return value is a function that must be called before the test exits.
564 1 : func DeterministicReadBlockDurationForTesting() func() {
565 1 : drbdForTesting := deterministicReadBlockDurationForTesting
566 1 : deterministicReadBlockDurationForTesting = true
567 1 : return func() {
568 1 : deterministicReadBlockDurationForTesting = drbdForTesting
569 1 : }
570 : }
571 :
572 : var deterministicReadBlockDurationForTesting = false
573 :
574 : func (r *Reader) readBlock(
575 : ctx context.Context,
576 : bh BlockHandle,
577 : transform blockTransform,
578 : readHandle objstorage.ReadHandle,
579 : stats *base.InternalIteratorStats,
580 : iterStats *iterStatsAccumulator,
581 : bufferPool *BufferPool,
582 1 : ) (handle bufferHandle, _ error) {
583 1 : if h := r.opts.Cache.Get(r.cacheID, r.fileNum, bh.Offset); h.Get() != nil {
584 1 : // Cache hit.
585 1 : if readHandle != nil {
586 1 : readHandle.RecordCacheHit(ctx, int64(bh.Offset), int64(bh.Length+blockTrailerLen))
587 1 : }
588 1 : if stats != nil {
589 1 : stats.BlockBytes += bh.Length
590 1 : stats.BlockBytesInCache += bh.Length
591 1 : }
592 1 : if iterStats != nil {
593 1 : iterStats.reportStats(bh.Length, bh.Length, 0)
594 1 : }
595 : // This block is already in the cache; return a handle to existing vlaue
596 : // in the cache.
597 1 : return bufferHandle{h: h}, nil
598 : }
599 :
600 : // Cache miss.
601 1 : var compressed cacheValueOrBuf
602 1 : if bufferPool != nil {
603 1 : compressed = cacheValueOrBuf{
604 1 : buf: bufferPool.Alloc(int(bh.Length + blockTrailerLen)),
605 1 : }
606 1 : } else {
607 1 : compressed = cacheValueOrBuf{
608 1 : v: cache.Alloc(int(bh.Length + blockTrailerLen)),
609 1 : }
610 1 : }
611 :
612 1 : readStartTime := time.Now()
613 1 : var err error
614 1 : if readHandle != nil {
615 1 : err = readHandle.ReadAt(ctx, compressed.get(), int64(bh.Offset))
616 1 : } else {
617 1 : err = r.readable.ReadAt(ctx, compressed.get(), int64(bh.Offset))
618 1 : }
619 1 : readDuration := time.Since(readStartTime)
620 1 : // TODO(sumeer): should the threshold be configurable.
621 1 : const slowReadTracingThreshold = 5 * time.Millisecond
622 1 : // For deterministic testing.
623 1 : if deterministicReadBlockDurationForTesting {
624 1 : readDuration = slowReadTracingThreshold
625 1 : }
626 : // Call IsTracingEnabled to avoid the allocations of boxing integers into an
627 : // interface{}, unless necessary.
628 1 : if readDuration >= slowReadTracingThreshold && r.opts.LoggerAndTracer.IsTracingEnabled(ctx) {
629 1 : r.opts.LoggerAndTracer.Eventf(ctx, "reading %d bytes took %s",
630 1 : int(bh.Length+blockTrailerLen), readDuration.String())
631 1 : }
632 1 : if stats != nil {
633 1 : stats.BlockReadDuration += readDuration
634 1 : }
635 1 : if err != nil {
636 1 : compressed.release()
637 1 : return bufferHandle{}, err
638 1 : }
639 1 : if err := checkChecksum(r.checksumType, compressed.get(), bh, r.fileNum); err != nil {
640 1 : compressed.release()
641 1 : return bufferHandle{}, err
642 1 : }
643 :
644 1 : typ := blockType(compressed.get()[bh.Length])
645 1 : compressed.truncate(int(bh.Length))
646 1 :
647 1 : var decompressed cacheValueOrBuf
648 1 : if typ == noCompressionBlockType {
649 1 : decompressed = compressed
650 1 : } else {
651 1 : // Decode the length of the decompressed value.
652 1 : decodedLen, prefixLen, err := decompressedLen(typ, compressed.get())
653 1 : if err != nil {
654 0 : compressed.release()
655 0 : return bufferHandle{}, err
656 0 : }
657 :
658 1 : if bufferPool != nil {
659 1 : decompressed = cacheValueOrBuf{buf: bufferPool.Alloc(decodedLen)}
660 1 : } else {
661 1 : decompressed = cacheValueOrBuf{v: cache.Alloc(decodedLen)}
662 1 : }
663 1 : if _, err := decompressInto(typ, compressed.get()[prefixLen:], decompressed.get()); err != nil {
664 0 : compressed.release()
665 0 : return bufferHandle{}, err
666 0 : }
667 1 : compressed.release()
668 : }
669 :
670 1 : if transform != nil {
671 1 : // Transforming blocks is very rare, so the extra copy of the
672 1 : // transformed data is not problematic.
673 1 : tmpTransformed, err := transform(decompressed.get())
674 1 : if err != nil {
675 0 : decompressed.release()
676 0 : return bufferHandle{}, err
677 0 : }
678 :
679 1 : var transformed cacheValueOrBuf
680 1 : if bufferPool != nil {
681 0 : transformed = cacheValueOrBuf{buf: bufferPool.Alloc(len(tmpTransformed))}
682 1 : } else {
683 1 : transformed = cacheValueOrBuf{v: cache.Alloc(len(tmpTransformed))}
684 1 : }
685 1 : copy(transformed.get(), tmpTransformed)
686 1 : decompressed.release()
687 1 : decompressed = transformed
688 : }
689 :
690 1 : if stats != nil {
691 1 : stats.BlockBytes += bh.Length
692 1 : }
693 1 : if iterStats != nil {
694 1 : iterStats.reportStats(bh.Length, 0, readDuration)
695 1 : }
696 1 : if decompressed.buf.Valid() {
697 1 : return bufferHandle{b: decompressed.buf}, nil
698 1 : }
699 1 : h := r.opts.Cache.Set(r.cacheID, r.fileNum, bh.Offset, decompressed.v)
700 1 : return bufferHandle{h: h}, nil
701 : }
702 :
703 1 : func (r *Reader) transformRangeDelV1(b []byte) ([]byte, error) {
704 1 : // Convert v1 (RocksDB format) range-del blocks to v2 blocks on the fly. The
705 1 : // v1 format range-del blocks have unfragmented and unsorted range
706 1 : // tombstones. We need properly fragmented and sorted range tombstones in
707 1 : // order to serve from them directly.
708 1 : iter := &blockIter{}
709 1 : if err := iter.init(r.Compare, b, r.Properties.GlobalSeqNum, false); err != nil {
710 0 : return nil, err
711 0 : }
712 1 : var tombstones []keyspan.Span
713 1 : for key, value := iter.First(); key != nil; key, value = iter.Next() {
714 1 : t := keyspan.Span{
715 1 : Start: key.UserKey,
716 1 : End: value.InPlaceValue(),
717 1 : Keys: []keyspan.Key{{Trailer: key.Trailer}},
718 1 : }
719 1 : tombstones = append(tombstones, t)
720 1 : }
721 1 : keyspan.Sort(r.Compare, tombstones)
722 1 :
723 1 : // Fragment the tombstones, outputting them directly to a block writer.
724 1 : rangeDelBlock := blockWriter{
725 1 : restartInterval: 1,
726 1 : }
727 1 : frag := keyspan.Fragmenter{
728 1 : Cmp: r.Compare,
729 1 : Format: r.FormatKey,
730 1 : Emit: func(s keyspan.Span) {
731 1 : for _, k := range s.Keys {
732 1 : startIK := InternalKey{UserKey: s.Start, Trailer: k.Trailer}
733 1 : rangeDelBlock.add(startIK, s.End)
734 1 : }
735 : },
736 : }
737 1 : for i := range tombstones {
738 1 : frag.Add(tombstones[i])
739 1 : }
740 1 : frag.Finish()
741 1 :
742 1 : // Return the contents of the constructed v2 format range-del block.
743 1 : return rangeDelBlock.finish(), nil
744 : }
745 :
746 1 : func (r *Reader) readMetaindex(metaindexBH BlockHandle) error {
747 1 : // We use a BufferPool when reading metaindex blocks in order to avoid
748 1 : // populating the block cache with these blocks. In heavy-write workloads,
749 1 : // especially with high compaction concurrency, new tables may be created
750 1 : // frequently. Populating the block cache with these metaindex blocks adds
751 1 : // additional contention on the block cache mutexes (see #1997).
752 1 : // Additionally, these blocks are exceedingly unlikely to be read again
753 1 : // while they're still in the block cache except in misconfigurations with
754 1 : // excessive sstables counts or a table cache that's far too small.
755 1 : r.metaBufferPool.initPreallocated(r.metaBufferPoolAlloc[:0])
756 1 : // When we're finished, release the buffers we've allocated back to memory
757 1 : // allocator. We don't expect to use metaBufferPool again.
758 1 : defer r.metaBufferPool.Release()
759 1 :
760 1 : b, err := r.readBlock(
761 1 : context.Background(), metaindexBH, nil /* transform */, nil /* readHandle */, nil, /* stats */
762 1 : nil /* iterStats */, &r.metaBufferPool)
763 1 : if err != nil {
764 1 : return err
765 1 : }
766 1 : data := b.Get()
767 1 : defer b.Release()
768 1 :
769 1 : if uint64(len(data)) != metaindexBH.Length {
770 0 : return base.CorruptionErrorf("pebble/table: unexpected metaindex block size: %d vs %d",
771 0 : errors.Safe(len(data)), errors.Safe(metaindexBH.Length))
772 0 : }
773 :
774 1 : i, err := newRawBlockIter(bytes.Compare, data)
775 1 : if err != nil {
776 0 : return err
777 0 : }
778 :
779 1 : meta := map[string]BlockHandle{}
780 1 : for valid := i.First(); valid; valid = i.Next() {
781 1 : value := i.Value()
782 1 : if bytes.Equal(i.Key().UserKey, []byte(metaValueIndexName)) {
783 1 : vbih, n, err := decodeValueBlocksIndexHandle(i.Value())
784 1 : if err != nil {
785 0 : return err
786 0 : }
787 1 : if n == 0 || n != len(value) {
788 0 : return base.CorruptionErrorf("pebble/table: invalid table (bad value blocks index handle)")
789 0 : }
790 1 : r.valueBIH = vbih
791 1 : } else {
792 1 : bh, n := decodeBlockHandle(value)
793 1 : if n == 0 || n != len(value) {
794 0 : return base.CorruptionErrorf("pebble/table: invalid table (bad block handle)")
795 0 : }
796 1 : meta[string(i.Key().UserKey)] = bh
797 : }
798 : }
799 1 : if err := i.Close(); err != nil {
800 0 : return err
801 0 : }
802 :
803 1 : if bh, ok := meta[metaPropertiesName]; ok {
804 1 : b, err = r.readBlock(
805 1 : context.Background(), bh, nil /* transform */, nil /* readHandle */, nil, /* stats */
806 1 : nil /* iterStats */, nil /* buffer pool */)
807 1 : if err != nil {
808 1 : return err
809 1 : }
810 1 : r.propertiesBH = bh
811 1 : err := r.Properties.load(b.Get(), bh.Offset, r.opts.DeniedUserProperties)
812 1 : b.Release()
813 1 : if err != nil {
814 0 : return err
815 0 : }
816 : }
817 :
818 1 : if bh, ok := meta[metaRangeDelV2Name]; ok {
819 1 : r.rangeDelBH = bh
820 1 : } else if bh, ok := meta[metaRangeDelName]; ok {
821 1 : r.rangeDelBH = bh
822 1 : if !r.rawTombstones {
823 1 : r.rangeDelTransform = r.transformRangeDelV1
824 1 : }
825 : }
826 :
827 1 : if bh, ok := meta[metaRangeKeyName]; ok {
828 1 : r.rangeKeyBH = bh
829 1 : }
830 :
831 1 : for name, fp := range r.opts.Filters {
832 1 : types := []struct {
833 1 : ftype FilterType
834 1 : prefix string
835 1 : }{
836 1 : {TableFilter, "fullfilter."},
837 1 : }
838 1 : var done bool
839 1 : for _, t := range types {
840 1 : if bh, ok := meta[t.prefix+name]; ok {
841 1 : r.filterBH = bh
842 1 :
843 1 : switch t.ftype {
844 1 : case TableFilter:
845 1 : r.tableFilter = newTableFilterReader(fp)
846 0 : default:
847 0 : return base.CorruptionErrorf("unknown filter type: %v", errors.Safe(t.ftype))
848 : }
849 :
850 1 : done = true
851 1 : break
852 : }
853 : }
854 1 : if done {
855 1 : break
856 : }
857 : }
858 1 : return nil
859 : }
860 :
861 : // Layout returns the layout (block organization) for an sstable.
862 1 : func (r *Reader) Layout() (*Layout, error) {
863 1 : if r.err != nil {
864 0 : return nil, r.err
865 0 : }
866 :
867 1 : l := &Layout{
868 1 : Data: make([]BlockHandleWithProperties, 0, r.Properties.NumDataBlocks),
869 1 : Filter: r.filterBH,
870 1 : RangeDel: r.rangeDelBH,
871 1 : RangeKey: r.rangeKeyBH,
872 1 : ValueIndex: r.valueBIH.h,
873 1 : Properties: r.propertiesBH,
874 1 : MetaIndex: r.metaIndexBH,
875 1 : Footer: r.footerBH,
876 1 : Format: r.tableFormat,
877 1 : }
878 1 :
879 1 : indexH, err := r.readIndex(context.Background(), nil, nil)
880 1 : if err != nil {
881 1 : return nil, err
882 1 : }
883 1 : defer indexH.Release()
884 1 :
885 1 : var alloc bytealloc.A
886 1 :
887 1 : if r.Properties.IndexPartitions == 0 {
888 1 : l.Index = append(l.Index, r.indexBH)
889 1 : iter, _ := newBlockIter(r.Compare, indexH.Get())
890 1 : for key, value := iter.First(); key != nil; key, value = iter.Next() {
891 1 : dataBH, err := decodeBlockHandleWithProperties(value.InPlaceValue())
892 1 : if err != nil {
893 0 : return nil, errCorruptIndexEntry
894 0 : }
895 1 : if len(dataBH.Props) > 0 {
896 1 : alloc, dataBH.Props = alloc.Copy(dataBH.Props)
897 1 : }
898 1 : l.Data = append(l.Data, dataBH)
899 : }
900 1 : } else {
901 1 : l.TopIndex = r.indexBH
902 1 : topIter, _ := newBlockIter(r.Compare, indexH.Get())
903 1 : iter := &blockIter{}
904 1 : for key, value := topIter.First(); key != nil; key, value = topIter.Next() {
905 1 : indexBH, err := decodeBlockHandleWithProperties(value.InPlaceValue())
906 1 : if err != nil {
907 0 : return nil, errCorruptIndexEntry
908 0 : }
909 1 : l.Index = append(l.Index, indexBH.BlockHandle)
910 1 :
911 1 : subIndex, err := r.readBlock(context.Background(), indexBH.BlockHandle,
912 1 : nil /* transform */, nil /* readHandle */, nil /* stats */, nil /* iterStats */, nil /* buffer pool */)
913 1 : if err != nil {
914 1 : return nil, err
915 1 : }
916 1 : if err := iter.init(r.Compare, subIndex.Get(), 0, /* globalSeqNum */
917 1 : false /* hideObsoletePoints */); err != nil {
918 0 : return nil, err
919 0 : }
920 1 : for key, value := iter.First(); key != nil; key, value = iter.Next() {
921 1 : dataBH, err := decodeBlockHandleWithProperties(value.InPlaceValue())
922 1 : if len(dataBH.Props) > 0 {
923 1 : alloc, dataBH.Props = alloc.Copy(dataBH.Props)
924 1 : }
925 1 : if err != nil {
926 0 : return nil, errCorruptIndexEntry
927 0 : }
928 1 : l.Data = append(l.Data, dataBH)
929 : }
930 1 : subIndex.Release()
931 1 : *iter = iter.resetForReuse()
932 : }
933 : }
934 1 : if r.valueBIH.h.Length != 0 {
935 1 : vbiH, err := r.readBlock(context.Background(), r.valueBIH.h, nil, nil, nil, nil, nil /* buffer pool */)
936 1 : if err != nil {
937 0 : return nil, err
938 0 : }
939 1 : defer vbiH.Release()
940 1 : vbiBlock := vbiH.Get()
941 1 : indexEntryLen := int(r.valueBIH.blockNumByteLength + r.valueBIH.blockOffsetByteLength +
942 1 : r.valueBIH.blockLengthByteLength)
943 1 : i := 0
944 1 : for len(vbiBlock) != 0 {
945 1 : if len(vbiBlock) < indexEntryLen {
946 0 : return nil, errors.Errorf(
947 0 : "remaining value index block %d does not contain a full entry of length %d",
948 0 : len(vbiBlock), indexEntryLen)
949 0 : }
950 1 : n := int(r.valueBIH.blockNumByteLength)
951 1 : bn := int(littleEndianGet(vbiBlock, n))
952 1 : if bn != i {
953 0 : return nil, errors.Errorf("unexpected block num %d, expected %d",
954 0 : bn, i)
955 0 : }
956 1 : i++
957 1 : vbiBlock = vbiBlock[n:]
958 1 : n = int(r.valueBIH.blockOffsetByteLength)
959 1 : blockOffset := littleEndianGet(vbiBlock, n)
960 1 : vbiBlock = vbiBlock[n:]
961 1 : n = int(r.valueBIH.blockLengthByteLength)
962 1 : blockLen := littleEndianGet(vbiBlock, n)
963 1 : vbiBlock = vbiBlock[n:]
964 1 : l.ValueBlock = append(l.ValueBlock, BlockHandle{Offset: blockOffset, Length: blockLen})
965 : }
966 : }
967 :
968 1 : return l, nil
969 : }
970 :
971 : // ValidateBlockChecksums validates the checksums for each block in the SSTable.
972 1 : func (r *Reader) ValidateBlockChecksums() error {
973 1 : // Pre-compute the BlockHandles for the underlying file.
974 1 : l, err := r.Layout()
975 1 : if err != nil {
976 1 : return err
977 1 : }
978 :
979 : // Construct the set of blocks to check. Note that the footer is not checked
980 : // as it is not a block with a checksum.
981 1 : blocks := make([]BlockHandle, len(l.Data))
982 1 : for i := range l.Data {
983 1 : blocks[i] = l.Data[i].BlockHandle
984 1 : }
985 1 : blocks = append(blocks, l.Index...)
986 1 : blocks = append(blocks, l.TopIndex, l.Filter, l.RangeDel, l.RangeKey, l.Properties, l.MetaIndex)
987 1 :
988 1 : // Sorting by offset ensures we are performing a sequential scan of the
989 1 : // file.
990 1 : slices.SortFunc(blocks, func(a, b BlockHandle) int {
991 1 : return cmp.Compare(a.Offset, b.Offset)
992 1 : })
993 :
994 : // Check all blocks sequentially. Make use of read-ahead, given we are
995 : // scanning the entire file from start to end.
996 1 : rh := r.readable.NewReadHandle(context.TODO())
997 1 : defer rh.Close()
998 1 :
999 1 : for _, bh := range blocks {
1000 1 : // Certain blocks may not be present, in which case we skip them.
1001 1 : if bh.Length == 0 {
1002 1 : continue
1003 : }
1004 :
1005 : // Read the block, which validates the checksum.
1006 1 : h, err := r.readBlock(context.Background(), bh, nil, rh, nil, nil /* iterStats */, nil /* buffer pool */)
1007 1 : if err != nil {
1008 1 : return err
1009 1 : }
1010 1 : h.Release()
1011 : }
1012 :
1013 1 : return nil
1014 : }
1015 :
1016 : // CommonProperties implemented the CommonReader interface.
1017 1 : func (r *Reader) CommonProperties() *CommonProperties {
1018 1 : return &r.Properties.CommonProperties
1019 1 : }
1020 :
1021 : // EstimateDiskUsage returns the total size of data blocks overlapping the range
1022 : // `[start, end]`. Even if a data block partially overlaps, or we cannot
1023 : // determine overlap due to abbreviated index keys, the full data block size is
1024 : // included in the estimation.
1025 : //
1026 : // This function does not account for any metablock space usage. Assumes there
1027 : // is at least partial overlap, i.e., `[start, end]` falls neither completely
1028 : // before nor completely after the file's range.
1029 : //
1030 : // Only blocks containing point keys are considered. Range deletion and range
1031 : // key blocks are not considered.
1032 : //
1033 : // TODO(ajkr): account for metablock space usage. Perhaps look at the fraction of
1034 : // data blocks overlapped and add that same fraction of the metadata blocks to the
1035 : // estimate.
1036 1 : func (r *Reader) EstimateDiskUsage(start, end []byte) (uint64, error) {
1037 1 : if r.err != nil {
1038 0 : return 0, r.err
1039 0 : }
1040 :
1041 1 : indexH, err := r.readIndex(context.Background(), nil, nil)
1042 1 : if err != nil {
1043 1 : return 0, err
1044 1 : }
1045 1 : defer indexH.Release()
1046 1 :
1047 1 : // Iterators over the bottom-level index blocks containing start and end.
1048 1 : // These may be different in case of partitioned index but will both point
1049 1 : // to the same blockIter over the single index in the unpartitioned case.
1050 1 : var startIdxIter, endIdxIter *blockIter
1051 1 : if r.Properties.IndexPartitions == 0 {
1052 1 : iter, err := newBlockIter(r.Compare, indexH.Get())
1053 1 : if err != nil {
1054 0 : return 0, err
1055 0 : }
1056 1 : startIdxIter = iter
1057 1 : endIdxIter = iter
1058 1 : } else {
1059 1 : topIter, err := newBlockIter(r.Compare, indexH.Get())
1060 1 : if err != nil {
1061 0 : return 0, err
1062 0 : }
1063 :
1064 1 : key, val := topIter.SeekGE(start, base.SeekGEFlagsNone)
1065 1 : if key == nil {
1066 0 : // The range falls completely after this file, or an error occurred.
1067 0 : return 0, topIter.Error()
1068 0 : }
1069 1 : startIdxBH, err := decodeBlockHandleWithProperties(val.InPlaceValue())
1070 1 : if err != nil {
1071 0 : return 0, errCorruptIndexEntry
1072 0 : }
1073 1 : startIdxBlock, err := r.readBlock(context.Background(), startIdxBH.BlockHandle,
1074 1 : nil /* transform */, nil /* readHandle */, nil /* stats */, nil /* iterStats */, nil /* buffer pool */)
1075 1 : if err != nil {
1076 1 : return 0, err
1077 1 : }
1078 1 : defer startIdxBlock.Release()
1079 1 : startIdxIter, err = newBlockIter(r.Compare, startIdxBlock.Get())
1080 1 : if err != nil {
1081 0 : return 0, err
1082 0 : }
1083 :
1084 1 : key, val = topIter.SeekGE(end, base.SeekGEFlagsNone)
1085 1 : if key == nil {
1086 0 : if err := topIter.Error(); err != nil {
1087 0 : return 0, err
1088 0 : }
1089 1 : } else {
1090 1 : endIdxBH, err := decodeBlockHandleWithProperties(val.InPlaceValue())
1091 1 : if err != nil {
1092 0 : return 0, errCorruptIndexEntry
1093 0 : }
1094 1 : endIdxBlock, err := r.readBlock(context.Background(),
1095 1 : endIdxBH.BlockHandle, nil /* transform */, nil /* readHandle */, nil /* stats */, nil /* iterStats */, nil /* buffer pool */)
1096 1 : if err != nil {
1097 1 : return 0, err
1098 1 : }
1099 1 : defer endIdxBlock.Release()
1100 1 : endIdxIter, err = newBlockIter(r.Compare, endIdxBlock.Get())
1101 1 : if err != nil {
1102 0 : return 0, err
1103 0 : }
1104 : }
1105 : }
1106 : // startIdxIter should not be nil at this point, while endIdxIter can be if the
1107 : // range spans past the end of the file.
1108 :
1109 1 : key, val := startIdxIter.SeekGE(start, base.SeekGEFlagsNone)
1110 1 : if key == nil {
1111 1 : // The range falls completely after this file, or an error occurred.
1112 1 : return 0, startIdxIter.Error()
1113 1 : }
1114 1 : startBH, err := decodeBlockHandleWithProperties(val.InPlaceValue())
1115 1 : if err != nil {
1116 0 : return 0, errCorruptIndexEntry
1117 0 : }
1118 :
1119 1 : includeInterpolatedValueBlocksSize := func(dataBlockSize uint64) uint64 {
1120 1 : // INVARIANT: r.Properties.DataSize > 0 since startIdxIter is not nil.
1121 1 : // Linearly interpolate what is stored in value blocks.
1122 1 : //
1123 1 : // TODO(sumeer): if we need more accuracy, without loading any data blocks
1124 1 : // (which contain the value handles, and which may also be insufficient if
1125 1 : // the values are in separate files), we will need to accumulate the
1126 1 : // logical size of the key-value pairs and store the cumulative value for
1127 1 : // each data block in the index block entry. This increases the size of
1128 1 : // the BlockHandle, so wait until this becomes necessary.
1129 1 : return dataBlockSize +
1130 1 : uint64((float64(dataBlockSize)/float64(r.Properties.DataSize))*
1131 1 : float64(r.Properties.ValueBlocksSize))
1132 1 : }
1133 1 : if endIdxIter == nil {
1134 0 : // The range spans beyond this file. Include data blocks through the last.
1135 0 : return includeInterpolatedValueBlocksSize(r.Properties.DataSize - startBH.Offset), nil
1136 0 : }
1137 1 : key, val = endIdxIter.SeekGE(end, base.SeekGEFlagsNone)
1138 1 : if key == nil {
1139 1 : if err := endIdxIter.Error(); err != nil {
1140 0 : return 0, err
1141 0 : }
1142 : // The range spans beyond this file. Include data blocks through the last.
1143 1 : return includeInterpolatedValueBlocksSize(r.Properties.DataSize - startBH.Offset), nil
1144 : }
1145 1 : endBH, err := decodeBlockHandleWithProperties(val.InPlaceValue())
1146 1 : if err != nil {
1147 0 : return 0, errCorruptIndexEntry
1148 0 : }
1149 1 : return includeInterpolatedValueBlocksSize(
1150 1 : endBH.Offset + endBH.Length + blockTrailerLen - startBH.Offset), nil
1151 : }
1152 :
1153 : // TableFormat returns the format version for the table.
1154 1 : func (r *Reader) TableFormat() (TableFormat, error) {
1155 1 : if r.err != nil {
1156 0 : return TableFormatUnspecified, r.err
1157 0 : }
1158 1 : return r.tableFormat, nil
1159 : }
1160 :
1161 : // NewReader returns a new table reader for the file. Closing the reader will
1162 : // close the file.
1163 1 : func NewReader(f objstorage.Readable, o ReaderOptions, extraOpts ...ReaderOption) (*Reader, error) {
1164 1 : o = o.ensureDefaults()
1165 1 : r := &Reader{
1166 1 : readable: f,
1167 1 : opts: o,
1168 1 : }
1169 1 : if r.opts.Cache == nil {
1170 1 : r.opts.Cache = cache.New(0)
1171 1 : } else {
1172 1 : r.opts.Cache.Ref()
1173 1 : }
1174 :
1175 1 : if f == nil {
1176 1 : r.err = errors.New("pebble/table: nil file")
1177 1 : return nil, r.Close()
1178 1 : }
1179 :
1180 : // Note that the extra options are applied twice. First here for pre-apply
1181 : // options, and then below for post-apply options. Pre and post refer to
1182 : // before and after reading the metaindex and properties.
1183 1 : type preApply interface{ preApply() }
1184 1 : for _, opt := range extraOpts {
1185 1 : if _, ok := opt.(preApply); ok {
1186 1 : opt.readerApply(r)
1187 1 : }
1188 : }
1189 1 : if r.cacheID == 0 {
1190 1 : r.cacheID = r.opts.Cache.NewID()
1191 1 : }
1192 :
1193 1 : footer, err := readFooter(f)
1194 1 : if err != nil {
1195 1 : r.err = err
1196 1 : return nil, r.Close()
1197 1 : }
1198 1 : r.checksumType = footer.checksum
1199 1 : r.tableFormat = footer.format
1200 1 : // Read the metaindex.
1201 1 : if err := r.readMetaindex(footer.metaindexBH); err != nil {
1202 1 : r.err = err
1203 1 : return nil, r.Close()
1204 1 : }
1205 1 : r.indexBH = footer.indexBH
1206 1 : r.metaIndexBH = footer.metaindexBH
1207 1 : r.footerBH = footer.footerBH
1208 1 :
1209 1 : if r.Properties.ComparerName == "" || o.Comparer.Name == r.Properties.ComparerName {
1210 1 : r.Compare = o.Comparer.Compare
1211 1 : r.Equal = o.Comparer.Equal
1212 1 : r.FormatKey = o.Comparer.FormatKey
1213 1 : r.Split = o.Comparer.Split
1214 1 : }
1215 :
1216 1 : if o.MergerName == r.Properties.MergerName {
1217 1 : r.mergerOK = true
1218 1 : }
1219 :
1220 : // Apply the extra options again now that the comparer and merger names are
1221 : // known.
1222 1 : for _, opt := range extraOpts {
1223 1 : if _, ok := opt.(preApply); !ok {
1224 1 : opt.readerApply(r)
1225 1 : }
1226 : }
1227 :
1228 1 : if r.Compare == nil {
1229 1 : r.err = errors.Errorf("pebble/table: %d: unknown comparer %s",
1230 1 : errors.Safe(r.fileNum), errors.Safe(r.Properties.ComparerName))
1231 1 : }
1232 1 : if !r.mergerOK {
1233 1 : if name := r.Properties.MergerName; name != "" && name != "nullptr" {
1234 1 : r.err = errors.Errorf("pebble/table: %d: unknown merger %s",
1235 1 : errors.Safe(r.fileNum), errors.Safe(r.Properties.MergerName))
1236 1 : }
1237 : }
1238 1 : if r.err != nil {
1239 1 : return nil, r.Close()
1240 1 : }
1241 :
1242 1 : return r, nil
1243 : }
1244 :
1245 : // ReadableFile describes the smallest subset of vfs.File that is required for
1246 : // reading SSTs.
1247 : type ReadableFile interface {
1248 : io.ReaderAt
1249 : io.Closer
1250 : Stat() (os.FileInfo, error)
1251 : }
1252 :
1253 : // NewSimpleReadable wraps a ReadableFile in a objstorage.Readable
1254 : // implementation (which does not support read-ahead)
1255 1 : func NewSimpleReadable(r ReadableFile) (objstorage.Readable, error) {
1256 1 : info, err := r.Stat()
1257 1 : if err != nil {
1258 1 : return nil, err
1259 1 : }
1260 1 : res := &simpleReadable{
1261 1 : f: r,
1262 1 : size: info.Size(),
1263 1 : }
1264 1 : res.rh = objstorage.MakeNoopReadHandle(res)
1265 1 : return res, nil
1266 : }
1267 :
1268 : // simpleReadable wraps a ReadableFile to implement objstorage.Readable.
1269 : type simpleReadable struct {
1270 : f ReadableFile
1271 : size int64
1272 : rh objstorage.NoopReadHandle
1273 : }
1274 :
1275 : var _ objstorage.Readable = (*simpleReadable)(nil)
1276 :
1277 : // ReadAt is part of the objstorage.Readable interface.
1278 1 : func (s *simpleReadable) ReadAt(_ context.Context, p []byte, off int64) error {
1279 1 : n, err := s.f.ReadAt(p, off)
1280 1 : if invariants.Enabled && err == nil && n != len(p) {
1281 0 : panic("short read")
1282 : }
1283 1 : return err
1284 : }
1285 :
1286 : // Close is part of the objstorage.Readable interface.
1287 1 : func (s *simpleReadable) Close() error {
1288 1 : return s.f.Close()
1289 1 : }
1290 :
1291 : // Size is part of the objstorage.Readable interface.
1292 1 : func (s *simpleReadable) Size() int64 {
1293 1 : return s.size
1294 1 : }
1295 :
1296 : // NewReaddHandle is part of the objstorage.Readable interface.
1297 1 : func (s *simpleReadable) NewReadHandle(_ context.Context) objstorage.ReadHandle {
1298 1 : return &s.rh
1299 1 : }
|