Line data Source code
1 : // Copyright 2023 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 pebble
6 :
7 : import (
8 : "context"
9 : "fmt"
10 : "slices"
11 :
12 : "github.com/cockroachdb/errors"
13 : "github.com/cockroachdb/pebble/internal/base"
14 : "github.com/cockroachdb/pebble/internal/invariants"
15 : "github.com/cockroachdb/pebble/internal/keyspan"
16 : "github.com/cockroachdb/pebble/internal/keyspan/keyspanimpl"
17 : "github.com/cockroachdb/pebble/internal/manifest"
18 : "github.com/cockroachdb/pebble/objstorage"
19 : "github.com/cockroachdb/pebble/objstorage/remote"
20 : "github.com/cockroachdb/pebble/sstable"
21 : )
22 :
23 : const (
24 : // In skip-shared iteration mode, keys in levels greater than
25 : // sharedLevelsStart (i.e. lower in the LSM) are skipped. Keys
26 : // in sharedLevelsStart are returned iff they are not in a
27 : // shared file.
28 : sharedLevelsStart = remote.SharedLevelsStart
29 :
30 : // In skip-external iteration mode, keys in levels greater
31 : // than externalSkipStart are skipped. Keys in
32 : // externalSkipStart are returned iff they are not in an
33 : // external file.
34 : externalSkipStart = 6
35 : )
36 :
37 : // ErrInvalidSkipSharedIteration is returned by ScanInternal if it was called
38 : // with a shared file visitor function, and a file in a shareable level (i.e.
39 : // level >= sharedLevelsStart) was found to not be in shared storage according
40 : // to objstorage.Provider, or not shareable for another reason such as for
41 : // containing keys newer than the snapshot sequence number.
42 : var ErrInvalidSkipSharedIteration = errors.New("pebble: cannot use skip-shared iteration due to non-shareable files in lower levels")
43 :
44 : // SharedSSTMeta represents an sstable on shared storage that can be ingested
45 : // by another pebble instance. This struct must contain all fields that are
46 : // required for a Pebble instance to ingest a foreign sstable on shared storage,
47 : // including constructing any relevant objstorage.Provider / remoteobjcat.Catalog
48 : // data structures, as well as creating virtual FileMetadatas.
49 : //
50 : // Note that the Pebble instance creating and returning a SharedSSTMeta might
51 : // not be the one that created the underlying sstable on shared storage to begin
52 : // with; it's possible for a Pebble instance to reshare an sstable that was
53 : // shared to it.
54 : type SharedSSTMeta struct {
55 : // Backing is the shared object underlying this SST. Can be attached to an
56 : // objstorage.Provider.
57 : Backing objstorage.RemoteObjectBackingHandle
58 :
59 : // Smallest and Largest internal keys for the overall bounds. The kind and
60 : // SeqNum of these will reflect what is physically present on the source Pebble
61 : // instance's view of the sstable; it's up to the ingesting instance to set the
62 : // sequence number in the trailer to match the read-time sequence numbers
63 : // reserved for the level this SST is being ingested into. The Kind is expected
64 : // to remain unchanged by the ingesting instance.
65 : //
66 : // Note that these bounds could be narrower than the bounds of the underlying
67 : // sstable; ScanInternal is expected to truncate sstable bounds to the user key
68 : // bounds passed into that method.
69 : Smallest, Largest InternalKey
70 :
71 : // SmallestRangeKey and LargestRangeKey are internal keys that denote the
72 : // range key bounds of this sstable. Must lie within [Smallest, Largest].
73 : SmallestRangeKey, LargestRangeKey InternalKey
74 :
75 : // SmallestPointKey and LargestPointKey are internal keys that denote the
76 : // point key bounds of this sstable. Must lie within [Smallest, Largest].
77 : SmallestPointKey, LargestPointKey InternalKey
78 :
79 : // Level denotes the level at which this file was present at read time.
80 : // For files visited by ScanInternal, this value will only be 5 or 6.
81 : Level uint8
82 :
83 : // Size contains an estimate of the size of this sstable.
84 : Size uint64
85 :
86 : // fileNum at time of creation in the creator instance. Only used for
87 : // debugging/tests.
88 : fileNum base.FileNum
89 : }
90 :
91 1 : func (s *SharedSSTMeta) cloneFromFileMeta(f *fileMetadata) {
92 1 : *s = SharedSSTMeta{
93 1 : Smallest: f.Smallest.Clone(),
94 1 : Largest: f.Largest.Clone(),
95 1 : SmallestRangeKey: f.SmallestRangeKey.Clone(),
96 1 : LargestRangeKey: f.LargestRangeKey.Clone(),
97 1 : SmallestPointKey: f.SmallestPointKey.Clone(),
98 1 : LargestPointKey: f.LargestPointKey.Clone(),
99 1 : Size: f.Size,
100 1 : fileNum: f.FileNum,
101 1 : }
102 1 : }
103 :
104 : type sharedByLevel []SharedSSTMeta
105 :
106 1 : func (s sharedByLevel) Len() int { return len(s) }
107 0 : func (s sharedByLevel) Swap(i, j int) { s[i], s[j] = s[j], s[i] }
108 1 : func (s sharedByLevel) Less(i, j int) bool { return s[i].Level < s[j].Level }
109 :
110 : type pcIterPos int
111 :
112 : const (
113 : pcIterPosCur pcIterPos = iota
114 : pcIterPosNext
115 : )
116 :
117 : // pointCollapsingIterator is an internalIterator that collapses point keys and
118 : // returns at most one point internal key for each user key. Merges and
119 : // SingleDels are not supported and result in a panic if encountered. Point keys
120 : // deleted by rangedels are considered shadowed and not exposed.
121 : //
122 : // Only used in ScanInternal to return at most one internal key per user key.
123 : type pointCollapsingIterator struct {
124 : iter keyspan.InterleavingIter
125 : pos pcIterPos
126 : comparer *base.Comparer
127 : merge base.Merge
128 : err error
129 : seqNum uint64
130 : // The current position of `iter`. Always owned by the underlying iter.
131 : iterKV *base.InternalKV
132 : // The last saved key. findNextEntry and similar methods are expected to save
133 : // the current value of iterKey to savedKey if they're iterating away from the
134 : // current key but still need to retain it. See comments in findNextEntry on
135 : // how this field is used.
136 : //
137 : // At the end of a positioning call:
138 : // - if pos == pcIterPosNext, iterKey is pointing to the next user key owned
139 : // by `iter` while savedKey is holding a copy to our current key.
140 : // - If pos == pcIterPosCur, iterKey is pointing to an `iter`-owned current
141 : // key, and savedKey is either undefined or pointing to a version of the
142 : // current key owned by this iterator (i.e. backed by savedKeyBuf).
143 : savedKey InternalKey
144 : savedKeyBuf []byte
145 : // If fixedSeqNum is non-zero, all emitted points are verified to have this
146 : // fixed sequence number.
147 : fixedSeqNum uint64
148 : }
149 :
150 1 : func (p *pointCollapsingIterator) Span() *keyspan.Span {
151 1 : return p.iter.Span()
152 1 : }
153 :
154 : // SeekPrefixGE implements the InternalIterator interface.
155 : func (p *pointCollapsingIterator) SeekPrefixGE(
156 : prefix, key []byte, flags base.SeekGEFlags,
157 0 : ) *base.InternalKV {
158 0 : p.resetKey()
159 0 : p.iterKV = p.iter.SeekPrefixGE(prefix, key, flags)
160 0 : p.pos = pcIterPosCur
161 0 : if p.iterKV == nil {
162 0 : return nil
163 0 : }
164 0 : return p.findNextEntry()
165 : }
166 :
167 : // SeekGE implements the InternalIterator interface.
168 1 : func (p *pointCollapsingIterator) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
169 1 : p.resetKey()
170 1 : p.iterKV = p.iter.SeekGE(key, flags)
171 1 : p.pos = pcIterPosCur
172 1 : if p.iterKV == nil {
173 1 : return nil
174 1 : }
175 1 : return p.findNextEntry()
176 : }
177 :
178 : // SeekLT implements the InternalIterator interface.
179 0 : func (p *pointCollapsingIterator) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
180 0 : panic("unimplemented")
181 : }
182 :
183 1 : func (p *pointCollapsingIterator) resetKey() {
184 1 : p.savedKey.UserKey = p.savedKeyBuf[:0]
185 1 : p.savedKey.Trailer = 0
186 1 : p.iterKV = nil
187 1 : p.pos = pcIterPosCur
188 1 : }
189 :
190 1 : func (p *pointCollapsingIterator) verifySeqNum(kv *base.InternalKV) *base.InternalKV {
191 1 : if !invariants.Enabled {
192 0 : return kv
193 0 : }
194 1 : if p.fixedSeqNum == 0 || kv == nil || kv.Kind() == InternalKeyKindRangeDelete {
195 1 : return kv
196 1 : }
197 0 : if kv.SeqNum() != p.fixedSeqNum {
198 0 : panic(fmt.Sprintf("expected foreign point key to have seqnum %d, got %d", p.fixedSeqNum, kv.SeqNum()))
199 : }
200 0 : return kv
201 : }
202 :
203 : // findNextEntry is called to return the next key. p.iter must be positioned at the
204 : // start of the first user key we are interested in.
205 1 : func (p *pointCollapsingIterator) findNextEntry() *base.InternalKV {
206 1 : p.saveKey()
207 1 : // Saves a comparison in the fast path
208 1 : firstIteration := true
209 1 : for p.iterKV != nil {
210 1 : // NB: p.savedKey is either the current key (iff p.iterKV == firstKey),
211 1 : // or the previous key.
212 1 : if !firstIteration && !p.comparer.Equal(p.iterKV.K.UserKey, p.savedKey.UserKey) {
213 1 : p.saveKey()
214 1 : continue
215 : }
216 1 : firstIteration = false
217 1 : if s := p.iter.Span(); s != nil && s.CoversAt(p.seqNum, p.iterKV.SeqNum()) {
218 1 : // All future keys for this user key must be deleted.
219 1 : if p.savedKey.Kind() == InternalKeyKindSingleDelete {
220 0 : panic("cannot process singledel key in point collapsing iterator")
221 : }
222 : // Fast forward to the next user key.
223 1 : p.saveKey()
224 1 : p.iterKV = p.iter.Next()
225 1 : for p.iterKV != nil && p.savedKey.SeqNum() >= p.iterKV.SeqNum() && p.comparer.Equal(p.iterKV.K.UserKey, p.savedKey.UserKey) {
226 1 : p.iterKV = p.iter.Next()
227 1 : }
228 1 : continue
229 : }
230 1 : switch p.savedKey.Kind() {
231 1 : case InternalKeyKindSet, InternalKeyKindDelete, InternalKeyKindSetWithDelete, InternalKeyKindDeleteSized:
232 1 : // Note that we return SETs directly, even if they would otherwise get
233 1 : // compacted into a Del to turn into a SetWithDelete. This is a fast
234 1 : // path optimization that can break SINGLEDEL determinism. To lead to
235 1 : // consistent SINGLEDEL behaviour, this iterator should *not* be used for
236 1 : // a keyspace where SINGLEDELs could be in use. If this iterator observes
237 1 : // a SINGLEDEL as the first internal key for a user key, it will panic.
238 1 : //
239 1 : // As p.value is a lazy value owned by the child iterator, we can thread
240 1 : // it through without loading it into p.valueBuf.
241 1 : //
242 1 : // TODO(bilal): We can even avoid saving the key in this fast path if
243 1 : // we are in a block where setHasSamePrefix = false in a v3 sstable,
244 1 : // guaranteeing that there's only one internal key for each user key.
245 1 : // Thread this logic through the sstable iterators and/or consider
246 1 : // collapsing (ha) this logic into the sstable iterators that are aware
247 1 : // of blocks and can determine user key changes without doing key saves
248 1 : // or comparisons.
249 1 : p.pos = pcIterPosCur
250 1 : return p.verifySeqNum(p.iterKV)
251 0 : case InternalKeyKindSingleDelete:
252 0 : // Panic, as this iterator is not expected to observe single deletes.
253 0 : panic("cannot process singledel key in point collapsing iterator")
254 0 : case InternalKeyKindMerge:
255 0 : // Panic, as this iterator is not expected to observe merges.
256 0 : panic("cannot process merge key in point collapsing iterator")
257 1 : case InternalKeyKindRangeDelete:
258 1 : // These are interleaved by the interleaving iterator ahead of all points.
259 1 : // We should pass them as-is, but also account for any points ahead of
260 1 : // them.
261 1 : p.pos = pcIterPosCur
262 1 : return p.verifySeqNum(p.iterKV)
263 0 : default:
264 0 : panic(fmt.Sprintf("unexpected kind: %d", p.iterKV.Kind()))
265 : }
266 : }
267 0 : p.resetKey()
268 0 : return nil
269 : }
270 :
271 : // First implements the InternalIterator interface.
272 1 : func (p *pointCollapsingIterator) First() *base.InternalKV {
273 1 : p.resetKey()
274 1 : p.iterKV = p.iter.First()
275 1 : p.pos = pcIterPosCur
276 1 : if p.iterKV == nil {
277 0 : return nil
278 0 : }
279 1 : return p.findNextEntry()
280 : }
281 :
282 : // Last implements the InternalIterator interface.
283 0 : func (p *pointCollapsingIterator) Last() *base.InternalKV {
284 0 : panic("unimplemented")
285 : }
286 :
287 1 : func (p *pointCollapsingIterator) saveKey() {
288 1 : if p.iterKV == nil {
289 1 : p.savedKey = InternalKey{UserKey: p.savedKeyBuf[:0]}
290 1 : return
291 1 : }
292 1 : p.savedKeyBuf = append(p.savedKeyBuf[:0], p.iterKV.K.UserKey...)
293 1 : p.savedKey = InternalKey{UserKey: p.savedKeyBuf, Trailer: p.iterKV.K.Trailer}
294 : }
295 :
296 : // Next implements the InternalIterator interface.
297 1 : func (p *pointCollapsingIterator) Next() *base.InternalKV {
298 1 : switch p.pos {
299 1 : case pcIterPosCur:
300 1 : p.saveKey()
301 1 : if p.iterKV != nil && p.iterKV.Kind() == InternalKeyKindRangeDelete {
302 1 : // Step over the interleaved range delete and process the very next
303 1 : // internal key, even if it's at the same user key. This is because a
304 1 : // point for that user key has not been returned yet.
305 1 : p.iterKV = p.iter.Next()
306 1 : break
307 : }
308 : // Fast forward to the next user key.
309 1 : kv := p.iter.Next()
310 1 : // p.iterKV.SeqNum() >= key.SeqNum() is an optimization that allows us to
311 1 : // use p.iterKV.SeqNum() < key.SeqNum() as a sign that the user key has
312 1 : // changed, without needing to do the full key comparison.
313 1 : for kv != nil && p.savedKey.SeqNum() >= kv.SeqNum() &&
314 1 : p.comparer.Equal(p.savedKey.UserKey, kv.K.UserKey) {
315 1 : kv = p.iter.Next()
316 1 : }
317 1 : if kv == nil {
318 1 : // There are no keys to return.
319 1 : p.resetKey()
320 1 : return nil
321 1 : }
322 1 : p.iterKV = kv
323 0 : case pcIterPosNext:
324 0 : p.pos = pcIterPosCur
325 : }
326 1 : if p.iterKV == nil {
327 1 : p.resetKey()
328 1 : return nil
329 1 : }
330 1 : return p.findNextEntry()
331 : }
332 :
333 : // NextPrefix implements the InternalIterator interface.
334 0 : func (p *pointCollapsingIterator) NextPrefix(succKey []byte) *base.InternalKV {
335 0 : panic("unimplemented")
336 : }
337 :
338 : // Prev implements the InternalIterator interface.
339 0 : func (p *pointCollapsingIterator) Prev() *base.InternalKV {
340 0 : panic("unimplemented")
341 : }
342 :
343 : // Error implements the InternalIterator interface.
344 1 : func (p *pointCollapsingIterator) Error() error {
345 1 : if p.err != nil {
346 0 : return p.err
347 0 : }
348 1 : return p.iter.Error()
349 : }
350 :
351 : // Close implements the InternalIterator interface.
352 1 : func (p *pointCollapsingIterator) Close() error {
353 1 : return p.iter.Close()
354 1 : }
355 :
356 : // SetBounds implements the InternalIterator interface.
357 0 : func (p *pointCollapsingIterator) SetBounds(lower, upper []byte) {
358 0 : p.resetKey()
359 0 : p.iter.SetBounds(lower, upper)
360 0 : }
361 :
362 0 : func (p *pointCollapsingIterator) SetContext(ctx context.Context) {
363 0 : p.iter.SetContext(ctx)
364 0 : }
365 :
366 : // String implements the InternalIterator interface.
367 0 : func (p *pointCollapsingIterator) String() string {
368 0 : return p.iter.String()
369 0 : }
370 :
371 : var _ internalIterator = &pointCollapsingIterator{}
372 :
373 : // IteratorLevelKind is used to denote whether the current ScanInternal iterator
374 : // is unknown, belongs to a flushable, or belongs to an LSM level type.
375 : type IteratorLevelKind int8
376 :
377 : const (
378 : // IteratorLevelUnknown indicates an unknown LSM level.
379 : IteratorLevelUnknown IteratorLevelKind = iota
380 : // IteratorLevelLSM indicates an LSM level.
381 : IteratorLevelLSM
382 : // IteratorLevelFlushable indicates a flushable (i.e. memtable).
383 : IteratorLevelFlushable
384 : )
385 :
386 : // IteratorLevel is used with scanInternalIterator to surface additional iterator-specific info where possible.
387 : // Note: this is struct is only provided for point keys.
388 : type IteratorLevel struct {
389 : Kind IteratorLevelKind
390 : // FlushableIndex indicates the position within the flushable queue of this level.
391 : // Only valid if kind == IteratorLevelFlushable.
392 : FlushableIndex int
393 : // The level within the LSM. Only valid if Kind == IteratorLevelLSM.
394 : Level int
395 : // Sublevel is only valid if Kind == IteratorLevelLSM and Level == 0.
396 : Sublevel int
397 : }
398 :
399 : // scanInternalIterator is an iterator that returns all internal keys, including
400 : // tombstones. For instance, an InternalKeyKindDelete would be returned as an
401 : // InternalKeyKindDelete instead of the iterator skipping over to the next key.
402 : // Internal keys within a user key are collapsed, eg. if there are two SETs, the
403 : // one with the higher sequence is returned. Useful if an external user of Pebble
404 : // needs to observe and rebuild Pebble's history of internal keys, such as in
405 : // node-to-node replication. For use with {db,snapshot}.ScanInternal().
406 : //
407 : // scanInternalIterator is expected to ignore point keys deleted by range
408 : // deletions, and range keys shadowed by a range key unset or delete, however it
409 : // *must* return the range delete as well as the range key unset/delete that did
410 : // the shadowing.
411 : type scanInternalIterator struct {
412 : ctx context.Context
413 : db *DB
414 : opts scanInternalOptions
415 : comparer *base.Comparer
416 : merge Merge
417 : iter internalIterator
418 : readState *readState
419 : version *version
420 : rangeKey *iteratorRangeKeyState
421 : pointKeyIter internalIterator
422 : iterKV *base.InternalKV
423 : alloc *iterAlloc
424 : newIters tableNewIters
425 : newIterRangeKey keyspanimpl.TableNewSpanIter
426 : seqNum uint64
427 : iterLevels []IteratorLevel
428 : mergingIter *mergingIter
429 :
430 : // boundsBuf holds two buffers used to store the lower and upper bounds.
431 : // Whenever the InternalIterator's bounds change, the new bounds are copied
432 : // into boundsBuf[boundsBufIdx]. The two bounds share a slice to reduce
433 : // allocations. opts.LowerBound and opts.UpperBound point into this slice.
434 : boundsBuf [2][]byte
435 : boundsBufIdx int
436 : }
437 :
438 : // truncateExternalFile truncates an External file's [SmallestUserKey,
439 : // LargestUserKey] fields to [lower, upper). A ExternalFile is
440 : // produced that is suitable for external consumption by other Pebble
441 : // instances.
442 : //
443 : // truncateSharedFile reads the file to try to create the smallest
444 : // possible bounds. Here, we blindly truncate them. This may mean we
445 : // include this SST in iterations it isn't really needed in. Since we
446 : // don't expect External files to be long-lived in the pebble
447 : // instance, We think this is OK.
448 : //
449 : // TODO(ssd) 2024-01-26: Potentially de-duplicate with
450 : // truncateSharedFile.
451 : func (d *DB) truncateExternalFile(
452 : ctx context.Context,
453 : lower, upper []byte,
454 : level int,
455 : file *fileMetadata,
456 : objMeta objstorage.ObjectMetadata,
457 1 : ) (*ExternalFile, error) {
458 1 : cmp := d.cmp
459 1 : sst := &ExternalFile{
460 1 : Level: uint8(level),
461 1 : ObjName: objMeta.Remote.CustomObjectName,
462 1 : Locator: objMeta.Remote.Locator,
463 1 : HasPointKey: file.HasPointKeys,
464 1 : HasRangeKey: file.HasRangeKeys,
465 1 : Size: file.Size,
466 1 : SyntheticSuffix: slices.Clone(file.SyntheticSuffix),
467 1 : }
468 1 : if file.SyntheticPrefix.IsSet() {
469 0 : sst.SyntheticPrefix = slices.Clone(sst.SyntheticPrefix)
470 0 : }
471 :
472 1 : needsLowerTruncate := cmp(lower, file.Smallest.UserKey) > 0
473 1 : if needsLowerTruncate {
474 1 : sst.StartKey = slices.Clone(lower)
475 1 : } else {
476 1 : sst.StartKey = slices.Clone(file.Smallest.UserKey)
477 1 : }
478 :
479 1 : cmpUpper := cmp(upper, file.Largest.UserKey)
480 1 : needsUpperTruncate := cmpUpper < 0
481 1 : if needsUpperTruncate {
482 0 : sst.EndKey = slices.Clone(upper)
483 0 : sst.EndKeyIsInclusive = false
484 1 : } else {
485 1 : sst.EndKey = slices.Clone(file.Largest.UserKey)
486 1 : sst.EndKeyIsInclusive = !file.Largest.IsExclusiveSentinel()
487 1 : }
488 :
489 1 : if cmp(sst.StartKey, sst.EndKey) > 0 {
490 0 : return nil, base.AssertionFailedf("pebble: invalid external file bounds after truncation [%q, %q)", sst.StartKey, sst.EndKey)
491 0 : }
492 :
493 1 : if cmp(sst.StartKey, sst.EndKey) == 0 && !sst.EndKeyIsInclusive {
494 0 : return nil, base.AssertionFailedf("pebble: invalid external file bounds after truncation [%q, %q)", sst.StartKey, sst.EndKey)
495 0 : }
496 :
497 1 : return sst, nil
498 : }
499 :
500 : // truncateSharedFile truncates a shared file's [Smallest, Largest] fields to
501 : // [lower, upper), potentially opening iterators on the file to find keys within
502 : // the requested bounds. A SharedSSTMeta is produced that is suitable for
503 : // external consumption by other Pebble instances. If shouldSkip is true, this
504 : // file does not contain any keys in [lower, upper) and can be skipped.
505 : //
506 : // TODO(bilal): If opening iterators and doing reads in this method is too
507 : // inefficient, consider producing non-tight file bounds instead.
508 : func (d *DB) truncateSharedFile(
509 : ctx context.Context,
510 : lower, upper []byte,
511 : level int,
512 : file *fileMetadata,
513 : objMeta objstorage.ObjectMetadata,
514 1 : ) (sst *SharedSSTMeta, shouldSkip bool, err error) {
515 1 : cmp := d.cmp
516 1 : sst = &SharedSSTMeta{}
517 1 : sst.cloneFromFileMeta(file)
518 1 : sst.Level = uint8(level)
519 1 : sst.Backing, err = d.objProvider.RemoteObjectBacking(&objMeta)
520 1 : if err != nil {
521 0 : return nil, false, err
522 0 : }
523 1 : needsLowerTruncate := cmp(lower, file.Smallest.UserKey) > 0
524 1 : needsUpperTruncate := cmp(upper, file.Largest.UserKey) < 0 || (cmp(upper, file.Largest.UserKey) == 0 && !file.Largest.IsExclusiveSentinel())
525 1 : // Fast path: file is entirely within [lower, upper).
526 1 : if !needsLowerTruncate && !needsUpperTruncate {
527 1 : return sst, false, nil
528 1 : }
529 :
530 : // We will need to truncate file bounds in at least one direction. Open all
531 : // relevant iterators.
532 1 : iters, err := d.newIters(ctx, file, &IterOptions{
533 1 : LowerBound: lower,
534 1 : UpperBound: upper,
535 1 : level: manifest.Level(level),
536 1 : }, internalIterOpts{}, iterPointKeys|iterRangeDeletions|iterRangeKeys)
537 1 : if err != nil {
538 0 : return nil, false, err
539 0 : }
540 1 : defer iters.CloseAll()
541 1 : iter := iters.point
542 1 : rangeDelIter := iters.rangeDeletion
543 1 : rangeKeyIter := iters.rangeKey
544 1 : if rangeDelIter != nil {
545 1 : rangeDelIter = keyspan.Truncate(cmp, rangeDelIter, base.UserKeyBoundsEndExclusive(lower, upper))
546 1 : }
547 1 : if rangeKeyIter != nil {
548 1 : rangeKeyIter = keyspan.Truncate(cmp, rangeKeyIter, base.UserKeyBoundsEndExclusive(lower, upper))
549 1 : }
550 : // Check if we need to truncate on the left side. This means finding a new
551 : // LargestPointKey and LargestRangeKey that is >= lower.
552 1 : if needsLowerTruncate {
553 1 : sst.SmallestPointKey.UserKey = sst.SmallestPointKey.UserKey[:0]
554 1 : sst.SmallestPointKey.Trailer = 0
555 1 : kv := iter.SeekGE(lower, base.SeekGEFlagsNone)
556 1 : foundPointKey := kv != nil
557 1 : if kv != nil {
558 1 : sst.SmallestPointKey.CopyFrom(kv.K)
559 1 : }
560 1 : if rangeDelIter != nil {
561 1 : if span, err := rangeDelIter.SeekGE(lower); err != nil {
562 0 : return nil, false, err
563 1 : } else if span != nil && (len(sst.SmallestPointKey.UserKey) == 0 || base.InternalCompare(cmp, span.SmallestKey(), sst.SmallestPointKey) < 0) {
564 1 : sst.SmallestPointKey.CopyFrom(span.SmallestKey())
565 1 : foundPointKey = true
566 1 : }
567 : }
568 1 : if !foundPointKey {
569 1 : // There are no point keys in the span we're interested in.
570 1 : sst.SmallestPointKey = InternalKey{}
571 1 : sst.LargestPointKey = InternalKey{}
572 1 : }
573 1 : sst.SmallestRangeKey.UserKey = sst.SmallestRangeKey.UserKey[:0]
574 1 : sst.SmallestRangeKey.Trailer = 0
575 1 : if rangeKeyIter != nil {
576 1 : span, err := rangeKeyIter.SeekGE(lower)
577 1 : switch {
578 0 : case err != nil:
579 0 : return nil, false, err
580 1 : case span != nil:
581 1 : sst.SmallestRangeKey.CopyFrom(span.SmallestKey())
582 1 : default:
583 1 : // There are no range keys in the span we're interested in.
584 1 : sst.SmallestRangeKey = InternalKey{}
585 1 : sst.LargestRangeKey = InternalKey{}
586 : }
587 : }
588 : }
589 : // Check if we need to truncate on the right side. This means finding a new
590 : // LargestPointKey and LargestRangeKey that is < upper.
591 1 : if needsUpperTruncate {
592 1 : sst.LargestPointKey.UserKey = sst.LargestPointKey.UserKey[:0]
593 1 : sst.LargestPointKey.Trailer = 0
594 1 : kv := iter.SeekLT(upper, base.SeekLTFlagsNone)
595 1 : foundPointKey := kv != nil
596 1 : if kv != nil {
597 1 : sst.LargestPointKey.CopyFrom(kv.K)
598 1 : }
599 1 : if rangeDelIter != nil {
600 1 : if span, err := rangeDelIter.SeekLT(upper); err != nil {
601 0 : return nil, false, err
602 1 : } else if span != nil && (len(sst.LargestPointKey.UserKey) == 0 || base.InternalCompare(cmp, span.LargestKey(), sst.LargestPointKey) > 0) {
603 1 : sst.LargestPointKey.CopyFrom(span.LargestKey())
604 1 : foundPointKey = true
605 1 : }
606 : }
607 1 : if !foundPointKey {
608 1 : // There are no point keys in the span we're interested in.
609 1 : sst.SmallestPointKey = InternalKey{}
610 1 : sst.LargestPointKey = InternalKey{}
611 1 : }
612 1 : sst.LargestRangeKey.UserKey = sst.LargestRangeKey.UserKey[:0]
613 1 : sst.LargestRangeKey.Trailer = 0
614 1 : if rangeKeyIter != nil {
615 1 : span, err := rangeKeyIter.SeekLT(upper)
616 1 : switch {
617 0 : case err != nil:
618 0 : return nil, false, err
619 1 : case span != nil:
620 1 : sst.LargestRangeKey.CopyFrom(span.LargestKey())
621 1 : default:
622 1 : // There are no range keys in the span we're interested in.
623 1 : sst.SmallestRangeKey = InternalKey{}
624 1 : sst.LargestRangeKey = InternalKey{}
625 : }
626 : }
627 : }
628 : // Set overall bounds based on {Smallest,Largest}{Point,Range}Key.
629 1 : switch {
630 1 : case len(sst.SmallestRangeKey.UserKey) == 0:
631 1 : sst.Smallest = sst.SmallestPointKey
632 1 : case len(sst.SmallestPointKey.UserKey) == 0:
633 1 : sst.Smallest = sst.SmallestRangeKey
634 1 : default:
635 1 : sst.Smallest = sst.SmallestPointKey
636 1 : if base.InternalCompare(cmp, sst.SmallestRangeKey, sst.SmallestPointKey) < 0 {
637 1 : sst.Smallest = sst.SmallestRangeKey
638 1 : }
639 : }
640 1 : switch {
641 1 : case len(sst.LargestRangeKey.UserKey) == 0:
642 1 : sst.Largest = sst.LargestPointKey
643 1 : case len(sst.LargestPointKey.UserKey) == 0:
644 1 : sst.Largest = sst.LargestRangeKey
645 1 : default:
646 1 : sst.Largest = sst.LargestPointKey
647 1 : if base.InternalCompare(cmp, sst.LargestRangeKey, sst.LargestPointKey) > 0 {
648 1 : sst.Largest = sst.LargestRangeKey
649 1 : }
650 : }
651 : // On rare occasion, a file might overlap with [lower, upper) but not actually
652 : // have any keys within those bounds. Skip such files.
653 1 : if len(sst.Smallest.UserKey) == 0 {
654 1 : return nil, true, nil
655 1 : }
656 1 : sst.Size, err = d.tableCache.estimateSize(file, sst.Smallest.UserKey, sst.Largest.UserKey)
657 1 : if err != nil {
658 0 : return nil, false, err
659 0 : }
660 : // On occasion, estimateSize gives us a low estimate, i.e. a 0 file size. This
661 : // can cause panics in places where we divide by file sizes. Correct for it
662 : // here.
663 1 : if sst.Size == 0 {
664 1 : sst.Size = 1
665 1 : }
666 1 : return sst, false, nil
667 : }
668 :
669 : func scanInternalImpl(
670 : ctx context.Context, lower, upper []byte, iter *scanInternalIterator, opts *scanInternalOptions,
671 1 : ) error {
672 1 : if opts.visitSharedFile != nil && (lower == nil || upper == nil) {
673 0 : panic("lower and upper bounds must be specified in skip-shared iteration mode")
674 : }
675 1 : if opts.visitSharedFile != nil && opts.visitExternalFile != nil {
676 0 : return base.AssertionFailedf("cannot provide both a shared-file and external-file visitor")
677 0 : }
678 :
679 : // Before starting iteration, check if any files in levels sharedLevelsStart
680 : // and below are *not* shared. Error out if that is the case, as skip-shared
681 : // iteration will not produce a consistent point-in-time view of this range
682 : // of keys. For files that are shared, call visitSharedFile with a truncated
683 : // version of that file.
684 1 : cmp := iter.comparer.Compare
685 1 : provider := iter.db.ObjProvider()
686 1 : seqNum := iter.seqNum
687 1 : current := iter.version
688 1 : if current == nil {
689 1 : current = iter.readState.current
690 1 : }
691 :
692 1 : if opts.visitSharedFile != nil || opts.visitExternalFile != nil {
693 1 : if provider == nil {
694 0 : panic("expected non-nil Provider in skip-shared iteration mode")
695 : }
696 :
697 1 : firstLevelWithRemote := opts.skipLevelForOpts()
698 1 : for level := firstLevelWithRemote; level < numLevels; level++ {
699 1 : files := current.Levels[level].Iter()
700 1 : for f := files.SeekGE(cmp, lower); f != nil && cmp(f.Smallest.UserKey, upper) < 0; f = files.Next() {
701 1 : if cmp(lower, f.Largest.UserKey) == 0 && f.Largest.IsExclusiveSentinel() {
702 0 : continue
703 : }
704 :
705 1 : var objMeta objstorage.ObjectMetadata
706 1 : var err error
707 1 : objMeta, err = provider.Lookup(fileTypeTable, f.FileBacking.DiskFileNum)
708 1 : if err != nil {
709 0 : return err
710 0 : }
711 :
712 : // We allow a mix of files at the first level.
713 1 : if level != firstLevelWithRemote {
714 1 : if !objMeta.IsShared() && !objMeta.IsExternal() {
715 0 : return errors.Wrapf(ErrInvalidSkipSharedIteration, "file %s is not shared or external", objMeta.DiskFileNum)
716 0 : }
717 : }
718 :
719 1 : if objMeta.IsShared() && opts.visitSharedFile == nil {
720 0 : return errors.Wrapf(ErrInvalidSkipSharedIteration, "shared file is present but no shared file visitor is defined")
721 0 : }
722 :
723 1 : if objMeta.IsExternal() && opts.visitExternalFile == nil {
724 1 : return errors.Wrapf(ErrInvalidSkipSharedIteration, "external file is present but no external file visitor is defined")
725 1 : }
726 :
727 1 : if !base.Visible(f.LargestSeqNum, seqNum, base.InternalKeySeqNumMax) {
728 1 : return errors.Wrapf(ErrInvalidSkipSharedIteration, "file %s contains keys newer than snapshot", objMeta.DiskFileNum)
729 1 : }
730 :
731 1 : if level != firstLevelWithRemote && (!objMeta.IsShared() && !objMeta.IsExternal()) {
732 0 : return errors.Wrapf(ErrInvalidSkipSharedIteration, "file %s is not shared or external", objMeta.DiskFileNum)
733 0 : }
734 :
735 1 : if objMeta.IsShared() {
736 1 : var sst *SharedSSTMeta
737 1 : var skip bool
738 1 : sst, skip, err = iter.db.truncateSharedFile(ctx, lower, upper, level, f, objMeta)
739 1 : if err != nil {
740 0 : return err
741 0 : }
742 1 : if skip {
743 1 : continue
744 : }
745 1 : if err = opts.visitSharedFile(sst); err != nil {
746 0 : return err
747 0 : }
748 1 : } else if objMeta.IsExternal() {
749 1 : sst, err := iter.db.truncateExternalFile(ctx, lower, upper, level, f, objMeta)
750 1 : if err != nil {
751 0 : return err
752 0 : }
753 1 : if err := opts.visitExternalFile(sst); err != nil {
754 0 : return err
755 0 : }
756 : }
757 :
758 : }
759 : }
760 : }
761 :
762 1 : for valid := iter.seekGE(lower); valid && iter.error() == nil; valid = iter.next() {
763 1 : key := iter.unsafeKey()
764 1 :
765 1 : if opts.rateLimitFunc != nil {
766 0 : if err := opts.rateLimitFunc(key, iter.lazyValue()); err != nil {
767 0 : return err
768 0 : }
769 : }
770 :
771 1 : switch key.Kind() {
772 1 : case InternalKeyKindRangeKeyDelete, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeySet:
773 1 : if opts.visitRangeKey != nil {
774 1 : span := iter.unsafeSpan()
775 1 : // NB: The caller isn't interested in the sequence numbers of these
776 1 : // range keys. Rather, the caller wants them to be in trailer order
777 1 : // _after_ zeroing of sequence numbers. Copy span.Keys, sort it, and then
778 1 : // call visitRangeKey.
779 1 : keysCopy := make([]keyspan.Key, len(span.Keys))
780 1 : for i := range span.Keys {
781 1 : keysCopy[i] = span.Keys[i]
782 1 : keysCopy[i].Trailer = base.MakeTrailer(0, span.Keys[i].Kind())
783 1 : }
784 1 : keyspan.SortKeysByTrailer(&keysCopy)
785 1 : if err := opts.visitRangeKey(span.Start, span.End, keysCopy); err != nil {
786 0 : return err
787 0 : }
788 : }
789 1 : case InternalKeyKindRangeDelete:
790 1 : if opts.visitRangeDel != nil {
791 1 : rangeDel := iter.unsafeRangeDel()
792 1 : if err := opts.visitRangeDel(rangeDel.Start, rangeDel.End, rangeDel.LargestSeqNum()); err != nil {
793 0 : return err
794 0 : }
795 : }
796 1 : default:
797 1 : if opts.visitPointKey != nil {
798 1 : var info IteratorLevel
799 1 : if len(iter.mergingIter.heap.items) > 0 {
800 1 : mergingIterIdx := iter.mergingIter.heap.items[0].index
801 1 : info = iter.iterLevels[mergingIterIdx]
802 1 : } else {
803 0 : info = IteratorLevel{Kind: IteratorLevelUnknown}
804 0 : }
805 1 : val := iter.lazyValue()
806 1 : if err := opts.visitPointKey(key, val, info); err != nil {
807 0 : return err
808 0 : }
809 : }
810 : }
811 : }
812 :
813 1 : return nil
814 : }
815 :
816 1 : func (opts *scanInternalOptions) skipLevelForOpts() int {
817 1 : if opts.visitSharedFile != nil {
818 1 : return sharedLevelsStart
819 1 : }
820 1 : if opts.visitExternalFile != nil {
821 1 : return externalSkipStart
822 1 : }
823 1 : return numLevels
824 : }
825 :
826 : // constructPointIter constructs a merging iterator and sets i.iter to it.
827 : func (i *scanInternalIterator) constructPointIter(
828 : categoryAndQoS sstable.CategoryAndQoS, memtables flushableList, buf *iterAlloc,
829 1 : ) error {
830 1 : // Merging levels and levels from iterAlloc.
831 1 : mlevels := buf.mlevels[:0]
832 1 : levels := buf.levels[:0]
833 1 :
834 1 : // We compute the number of levels needed ahead of time and reallocate a slice if
835 1 : // the array from the iterAlloc isn't large enough. Doing this allocation once
836 1 : // should improve the performance.
837 1 : numMergingLevels := len(memtables)
838 1 : numLevelIters := 0
839 1 :
840 1 : current := i.version
841 1 : if current == nil {
842 1 : current = i.readState.current
843 1 : }
844 1 : numMergingLevels += len(current.L0SublevelFiles)
845 1 : numLevelIters += len(current.L0SublevelFiles)
846 1 :
847 1 : skipStart := i.opts.skipLevelForOpts()
848 1 : for level := 1; level < len(current.Levels); level++ {
849 1 : if current.Levels[level].Empty() {
850 1 : continue
851 : }
852 1 : if level > skipStart {
853 1 : continue
854 : }
855 1 : numMergingLevels++
856 1 : numLevelIters++
857 : }
858 :
859 1 : if numMergingLevels > cap(mlevels) {
860 0 : mlevels = make([]mergingIterLevel, 0, numMergingLevels)
861 0 : }
862 1 : if numLevelIters > cap(levels) {
863 0 : levels = make([]levelIter, 0, numLevelIters)
864 0 : }
865 : // TODO(bilal): Push these into the iterAlloc buf.
866 1 : var rangeDelMiter keyspanimpl.MergingIter
867 1 : rangeDelIters := make([]keyspan.FragmentIterator, 0, numMergingLevels)
868 1 : rangeDelLevels := make([]keyspanimpl.LevelIter, 0, numLevelIters)
869 1 :
870 1 : i.iterLevels = make([]IteratorLevel, numMergingLevels)
871 1 : mlevelsIndex := 0
872 1 :
873 1 : // Next are the memtables.
874 1 : for j := len(memtables) - 1; j >= 0; j-- {
875 1 : mem := memtables[j]
876 1 : mlevels = append(mlevels, mergingIterLevel{
877 1 : iter: mem.newIter(&i.opts.IterOptions),
878 1 : })
879 1 : i.iterLevels[mlevelsIndex] = IteratorLevel{
880 1 : Kind: IteratorLevelFlushable,
881 1 : FlushableIndex: j,
882 1 : }
883 1 : mlevelsIndex++
884 1 : if rdi := mem.newRangeDelIter(&i.opts.IterOptions); rdi != nil {
885 1 : rangeDelIters = append(rangeDelIters, rdi)
886 1 : }
887 : }
888 :
889 : // Next are the file levels: L0 sub-levels followed by lower levels.
890 1 : levelsIndex := len(levels)
891 1 : mlevels = mlevels[:numMergingLevels]
892 1 : levels = levels[:numLevelIters]
893 1 : rangeDelLevels = rangeDelLevels[:numLevelIters]
894 1 : i.opts.IterOptions.snapshotForHideObsoletePoints = i.seqNum
895 1 : i.opts.IterOptions.CategoryAndQoS = categoryAndQoS
896 1 : addLevelIterForFiles := func(files manifest.LevelIterator, level manifest.Level) {
897 1 : li := &levels[levelsIndex]
898 1 : rli := &rangeDelLevels[levelsIndex]
899 1 :
900 1 : li.init(
901 1 : i.ctx, i.opts.IterOptions, i.comparer, i.newIters, files, level,
902 1 : internalIterOpts{})
903 1 : li.initBoundaryContext(&mlevels[mlevelsIndex].levelIterBoundaryContext)
904 1 : mlevels[mlevelsIndex].iter = li
905 1 : rli.Init(keyspan.SpanIterOptions{RangeKeyFilters: i.opts.RangeKeyFilters},
906 1 : i.comparer.Compare, tableNewRangeDelIter(i.ctx, i.newIters), files, level,
907 1 : manifest.KeyTypePoint)
908 1 : rangeDelIters = append(rangeDelIters, rli)
909 1 :
910 1 : levelsIndex++
911 1 : mlevelsIndex++
912 1 : }
913 :
914 1 : for j := len(current.L0SublevelFiles) - 1; j >= 0; j-- {
915 1 : i.iterLevels[mlevelsIndex] = IteratorLevel{
916 1 : Kind: IteratorLevelLSM,
917 1 : Level: 0,
918 1 : Sublevel: j,
919 1 : }
920 1 : addLevelIterForFiles(current.L0SublevelFiles[j].Iter(), manifest.L0Sublevel(j))
921 1 : }
922 : // Add level iterators for the non-empty non-L0 levels.
923 1 : for level := 1; level < numLevels; level++ {
924 1 : if current.Levels[level].Empty() {
925 1 : continue
926 : }
927 :
928 1 : if level > skipStart {
929 1 : continue
930 : }
931 1 : i.iterLevels[mlevelsIndex] = IteratorLevel{Kind: IteratorLevelLSM, Level: level}
932 1 : levIter := current.Levels[level].Iter()
933 1 : if level == skipStart {
934 1 : nonRemoteFiles := make([]*manifest.FileMetadata, 0)
935 1 : for f := levIter.First(); f != nil; f = levIter.Next() {
936 1 : meta, err := i.db.objProvider.Lookup(fileTypeTable, f.FileBacking.DiskFileNum)
937 1 : if err != nil {
938 0 : return err
939 0 : }
940 1 : if (meta.IsShared() && i.opts.visitSharedFile != nil) ||
941 1 : (meta.IsExternal() && i.opts.visitExternalFile != nil) {
942 1 : // Skip this file.
943 1 : continue
944 : }
945 1 : nonRemoteFiles = append(nonRemoteFiles, f)
946 : }
947 1 : levSlice := manifest.NewLevelSliceKeySorted(i.db.cmp, nonRemoteFiles)
948 1 : levIter = levSlice.Iter()
949 : }
950 :
951 1 : addLevelIterForFiles(levIter, manifest.Level(level))
952 : }
953 :
954 1 : buf.merging.init(&i.opts.IterOptions, &InternalIteratorStats{}, i.comparer.Compare, i.comparer.Split, mlevels...)
955 1 : buf.merging.snapshot = i.seqNum
956 1 : rangeDelMiter.Init(i.comparer.Compare, keyspan.VisibleTransform(i.seqNum), new(keyspanimpl.MergingBuffers), rangeDelIters...)
957 1 :
958 1 : if i.opts.includeObsoleteKeys {
959 1 : iiter := &keyspan.InterleavingIter{}
960 1 : iiter.Init(i.comparer, &buf.merging, &rangeDelMiter,
961 1 : keyspan.InterleavingIterOpts{
962 1 : LowerBound: i.opts.LowerBound,
963 1 : UpperBound: i.opts.UpperBound,
964 1 : })
965 1 : i.pointKeyIter = iiter
966 1 : } else {
967 1 : pcIter := &pointCollapsingIterator{
968 1 : comparer: i.comparer,
969 1 : merge: i.merge,
970 1 : seqNum: i.seqNum,
971 1 : }
972 1 : pcIter.iter.Init(i.comparer, &buf.merging, &rangeDelMiter, keyspan.InterleavingIterOpts{
973 1 : LowerBound: i.opts.LowerBound,
974 1 : UpperBound: i.opts.UpperBound,
975 1 : })
976 1 : i.pointKeyIter = pcIter
977 1 : }
978 1 : i.iter = i.pointKeyIter
979 1 : return nil
980 : }
981 :
982 : // constructRangeKeyIter constructs the range-key iterator stack, populating
983 : // i.rangeKey.rangeKeyIter with the resulting iterator. This is similar to
984 : // Iterator.constructRangeKeyIter, except it doesn't handle batches and ensures
985 : // iterConfig does *not* elide unsets/deletes.
986 1 : func (i *scanInternalIterator) constructRangeKeyIter() error {
987 1 : // We want the bounded iter from iterConfig, but not the collapsing of
988 1 : // RangeKeyUnsets and RangeKeyDels.
989 1 : i.rangeKey.rangeKeyIter = i.rangeKey.iterConfig.Init(
990 1 : i.comparer, i.seqNum, i.opts.LowerBound, i.opts.UpperBound,
991 1 : nil /* hasPrefix */, nil /* prefix */, true, /* internalKeys */
992 1 : &i.rangeKey.rangeKeyBuffers.internal)
993 1 :
994 1 : // Next are the flushables: memtables and large batches.
995 1 : if i.readState != nil {
996 1 : for j := len(i.readState.memtables) - 1; j >= 0; j-- {
997 1 : mem := i.readState.memtables[j]
998 1 : // We only need to read from memtables which contain sequence numbers older
999 1 : // than seqNum.
1000 1 : if logSeqNum := mem.logSeqNum; logSeqNum >= i.seqNum {
1001 1 : continue
1002 : }
1003 1 : if rki := mem.newRangeKeyIter(&i.opts.IterOptions); rki != nil {
1004 1 : i.rangeKey.iterConfig.AddLevel(rki)
1005 1 : }
1006 : }
1007 : }
1008 :
1009 1 : current := i.version
1010 1 : if current == nil {
1011 1 : current = i.readState.current
1012 1 : }
1013 : // Next are the file levels: L0 sub-levels followed by lower levels.
1014 : //
1015 : // Add file-specific iterators for L0 files containing range keys. This is less
1016 : // efficient than using levelIters for sublevels of L0 files containing
1017 : // range keys, but range keys are expected to be sparse anyway, reducing the
1018 : // cost benefit of maintaining a separate L0Sublevels instance for range key
1019 : // files and then using it here.
1020 : //
1021 : // NB: We iterate L0's files in reverse order. They're sorted by
1022 : // LargestSeqNum ascending, and we need to add them to the merging iterator
1023 : // in LargestSeqNum descending to preserve the merging iterator's invariants
1024 : // around Key Trailer order.
1025 1 : iter := current.RangeKeyLevels[0].Iter()
1026 1 : for f := iter.Last(); f != nil; f = iter.Prev() {
1027 1 : spanIter, err := i.newIterRangeKey(f, i.opts.SpanIterOptions())
1028 1 : if err != nil {
1029 0 : return err
1030 0 : }
1031 1 : i.rangeKey.iterConfig.AddLevel(spanIter)
1032 : }
1033 : // Add level iterators for the non-empty non-L0 levels.
1034 1 : skipStart := i.opts.skipLevelForOpts()
1035 1 : for level := 1; level < len(current.RangeKeyLevels); level++ {
1036 1 : if current.RangeKeyLevels[level].Empty() {
1037 1 : continue
1038 : }
1039 1 : if level > skipStart {
1040 1 : continue
1041 : }
1042 1 : li := i.rangeKey.iterConfig.NewLevelIter()
1043 1 : spanIterOpts := i.opts.SpanIterOptions()
1044 1 : levIter := current.RangeKeyLevels[level].Iter()
1045 1 : if level == skipStart {
1046 1 : nonRemoteFiles := make([]*manifest.FileMetadata, 0)
1047 1 : for f := levIter.First(); f != nil; f = levIter.Next() {
1048 1 : meta, err := i.db.objProvider.Lookup(fileTypeTable, f.FileBacking.DiskFileNum)
1049 1 : if err != nil {
1050 0 : return err
1051 0 : }
1052 1 : if (meta.IsShared() && i.opts.visitSharedFile != nil) ||
1053 1 : (meta.IsExternal() && i.opts.visitExternalFile != nil) {
1054 1 : // Skip this file.
1055 1 : continue
1056 : }
1057 0 : nonRemoteFiles = append(nonRemoteFiles, f)
1058 : }
1059 1 : levSlice := manifest.NewLevelSliceKeySorted(i.db.cmp, nonRemoteFiles)
1060 1 : levIter = levSlice.Iter()
1061 : }
1062 1 : li.Init(spanIterOpts, i.comparer.Compare, i.newIterRangeKey, levIter,
1063 1 : manifest.Level(level), manifest.KeyTypeRange)
1064 1 : i.rangeKey.iterConfig.AddLevel(li)
1065 : }
1066 1 : return nil
1067 : }
1068 :
1069 : // seekGE seeks this iterator to the first key that's greater than or equal
1070 : // to the specified user key.
1071 1 : func (i *scanInternalIterator) seekGE(key []byte) bool {
1072 1 : i.iterKV = i.iter.SeekGE(key, base.SeekGEFlagsNone)
1073 1 : return i.iterKV != nil
1074 1 : }
1075 :
1076 : // unsafeKey returns the unsafe InternalKey at the current position. The value
1077 : // is nil if the iterator is invalid or exhausted.
1078 1 : func (i *scanInternalIterator) unsafeKey() *InternalKey {
1079 1 : return &i.iterKV.K
1080 1 : }
1081 :
1082 : // lazyValue returns a value pointer to the value at the current iterator
1083 : // position. Behaviour undefined if unsafeKey() returns a Range key or Rangedel
1084 : // kind key.
1085 1 : func (i *scanInternalIterator) lazyValue() LazyValue {
1086 1 : return i.iterKV.V
1087 1 : }
1088 :
1089 : // unsafeRangeDel returns a range key span. Behaviour undefined if UnsafeKey returns
1090 : // a non-rangedel kind.
1091 1 : func (i *scanInternalIterator) unsafeRangeDel() *keyspan.Span {
1092 1 : type spanInternalIterator interface {
1093 1 : Span() *keyspan.Span
1094 1 : }
1095 1 : return i.pointKeyIter.(spanInternalIterator).Span()
1096 1 : }
1097 :
1098 : // unsafeSpan returns a range key span. Behaviour undefined if UnsafeKey returns
1099 : // a non-rangekey type.
1100 1 : func (i *scanInternalIterator) unsafeSpan() *keyspan.Span {
1101 1 : return i.rangeKey.iiter.Span()
1102 1 : }
1103 :
1104 : // next advances the iterator in the forward direction, and returns the
1105 : // iterator's new validity state.
1106 1 : func (i *scanInternalIterator) next() bool {
1107 1 : i.iterKV = i.iter.Next()
1108 1 : return i.iterKV != nil
1109 1 : }
1110 :
1111 : // error returns an error from the internal iterator, if there's any.
1112 1 : func (i *scanInternalIterator) error() error {
1113 1 : return i.iter.Error()
1114 1 : }
1115 :
1116 : // close closes this iterator, and releases any pooled objects.
1117 1 : func (i *scanInternalIterator) close() error {
1118 1 : if err := i.iter.Close(); err != nil {
1119 0 : return err
1120 0 : }
1121 1 : if i.readState != nil {
1122 1 : i.readState.unref()
1123 1 : }
1124 1 : if i.version != nil {
1125 1 : i.version.Unref()
1126 1 : }
1127 1 : if i.rangeKey != nil {
1128 1 : i.rangeKey.PrepareForReuse()
1129 1 : *i.rangeKey = iteratorRangeKeyState{
1130 1 : rangeKeyBuffers: i.rangeKey.rangeKeyBuffers,
1131 1 : }
1132 1 : iterRangeKeyStateAllocPool.Put(i.rangeKey)
1133 1 : i.rangeKey = nil
1134 1 : }
1135 1 : if alloc := i.alloc; alloc != nil {
1136 1 : for j := range i.boundsBuf {
1137 1 : if cap(i.boundsBuf[j]) >= maxKeyBufCacheSize {
1138 0 : alloc.boundsBuf[j] = nil
1139 1 : } else {
1140 1 : alloc.boundsBuf[j] = i.boundsBuf[j]
1141 1 : }
1142 : }
1143 1 : *alloc = iterAlloc{
1144 1 : keyBuf: alloc.keyBuf[:0],
1145 1 : boundsBuf: alloc.boundsBuf,
1146 1 : prefixOrFullSeekKey: alloc.prefixOrFullSeekKey[:0],
1147 1 : }
1148 1 : iterAllocPool.Put(alloc)
1149 1 : i.alloc = nil
1150 : }
1151 1 : return nil
1152 : }
1153 :
1154 1 : func (i *scanInternalIterator) initializeBoundBufs(lower, upper []byte) {
1155 1 : buf := i.boundsBuf[i.boundsBufIdx][:0]
1156 1 : if lower != nil {
1157 1 : buf = append(buf, lower...)
1158 1 : i.opts.LowerBound = buf
1159 1 : } else {
1160 1 : i.opts.LowerBound = nil
1161 1 : }
1162 1 : if upper != nil {
1163 1 : buf = append(buf, upper...)
1164 1 : i.opts.UpperBound = buf[len(buf)-len(upper):]
1165 1 : } else {
1166 1 : i.opts.UpperBound = nil
1167 1 : }
1168 1 : i.boundsBuf[i.boundsBufIdx] = buf
1169 1 : i.boundsBufIdx = 1 - i.boundsBufIdx
1170 : }
|