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 pebble
6 :
7 : import (
8 : "bytes"
9 : "context"
10 : "io"
11 : "math/rand/v2"
12 : "sync"
13 : "unsafe"
14 :
15 : "github.com/cockroachdb/errors"
16 : "github.com/cockroachdb/pebble/internal/base"
17 : "github.com/cockroachdb/pebble/internal/bytealloc"
18 : "github.com/cockroachdb/pebble/internal/humanize"
19 : "github.com/cockroachdb/pebble/internal/invariants"
20 : "github.com/cockroachdb/pebble/internal/keyspan"
21 : "github.com/cockroachdb/pebble/internal/keyspan/keyspanimpl"
22 : "github.com/cockroachdb/pebble/internal/manifest"
23 : "github.com/cockroachdb/pebble/internal/rangekeystack"
24 : "github.com/cockroachdb/pebble/internal/treeprinter"
25 : "github.com/cockroachdb/pebble/sstable/blob"
26 : "github.com/cockroachdb/redact"
27 : )
28 :
29 : // iterPos describes the state of the internal iterator, in terms of whether it
30 : // is at the position returned to the user (cur), one ahead of the position
31 : // returned (next for forward iteration and prev for reverse iteration). The cur
32 : // position is split into two states, for forward and reverse iteration, since
33 : // we need to differentiate for switching directions.
34 : //
35 : // There is subtlety in what is considered the current position of the Iterator.
36 : // The internal iterator exposes a sequence of internal keys. There is not
37 : // always a single internalIterator position corresponding to the position
38 : // returned to the user. Consider the example:
39 : //
40 : // a.MERGE.9 a.MERGE.8 a.MERGE.7 a.SET.6 b.DELETE.9 b.DELETE.5 b.SET.4
41 : // \ /
42 : // \ Iterator.Key() = 'a' /
43 : //
44 : // The Iterator exposes one valid position at user key 'a' and the two exhausted
45 : // positions at the beginning and end of iteration. The underlying
46 : // internalIterator contains 7 valid positions and 2 exhausted positions.
47 : //
48 : // Iterator positioning methods must set iterPos to iterPosCur{Foward,Backward}
49 : // iff the user key at the current internalIterator position equals the
50 : // Iterator.Key returned to the user. This guarantees that a call to nextUserKey
51 : // or prevUserKey will advance to the next or previous iterator position.
52 : // iterPosCur{Forward,Backward} does not make any guarantee about the internal
53 : // iterator position among internal keys with matching user keys, and it will
54 : // vary subtly depending on the particular key kinds encountered. In the above
55 : // example, the iterator returning 'a' may set iterPosCurForward if the internal
56 : // iterator is positioned at any of a.MERGE.9, a.MERGE.8, a.MERGE.7 or a.SET.6.
57 : //
58 : // When setting iterPos to iterPosNext or iterPosPrev, the internal iterator
59 : // must be advanced to the first internalIterator position at a user key greater
60 : // (iterPosNext) or less (iterPosPrev) than the key returned to the user. An
61 : // internalIterator position that's !Valid() must also be considered greater or
62 : // less—depending on the direction of iteration—than the last valid Iterator
63 : // position.
64 : type iterPos int8
65 :
66 : const (
67 : iterPosCurForward iterPos = 0
68 : iterPosNext iterPos = 1
69 : iterPosPrev iterPos = -1
70 : iterPosCurReverse iterPos = -2
71 :
72 : // For limited iteration. When the iterator is at iterPosCurForwardPaused
73 : // - Next*() call should behave as if the internal iterator is already
74 : // at next (akin to iterPosNext).
75 : // - Prev*() call should behave as if the internal iterator is at the
76 : // current key (akin to iterPosCurForward).
77 : //
78 : // Similar semantics apply to CurReversePaused.
79 : iterPosCurForwardPaused iterPos = 2
80 : iterPosCurReversePaused iterPos = -3
81 : )
82 :
83 : // Approximate gap in bytes between samples of data read during iteration.
84 : // This is multiplied with a default ReadSamplingMultiplier of 1 << 4 to yield
85 : // 1 << 20 (1MB). The 1MB factor comes from:
86 : // https://github.com/cockroachdb/pebble/issues/29#issuecomment-494477985
87 : const readBytesPeriod uint64 = 1 << 16
88 :
89 : var errReversePrefixIteration = errors.New("pebble: unsupported reverse prefix iteration")
90 :
91 : // IteratorMetrics holds per-iterator metrics. These do not change over the
92 : // lifetime of the iterator.
93 : type IteratorMetrics struct {
94 : // The read amplification experienced by this iterator. This is the sum of
95 : // the memtables, the L0 sublevels and the non-empty Ln levels. Higher read
96 : // amplification generally results in slower reads, though allowing higher
97 : // read amplification can also result in faster writes.
98 : ReadAmp int
99 : }
100 :
101 : // IteratorStatsKind describes the two kind of iterator stats.
102 : type IteratorStatsKind int8
103 :
104 : const (
105 : // InterfaceCall represents calls to Iterator.
106 : InterfaceCall IteratorStatsKind = iota
107 : // InternalIterCall represents calls by Iterator to its internalIterator.
108 : InternalIterCall
109 : // NumStatsKind is the number of kinds, and is used for array sizing.
110 : NumStatsKind
111 : )
112 :
113 : // IteratorStats contains iteration stats.
114 : type IteratorStats struct {
115 : // ForwardSeekCount includes SeekGE, SeekPrefixGE, First.
116 : ForwardSeekCount [NumStatsKind]int
117 : // ReverseSeek includes SeekLT, Last.
118 : ReverseSeekCount [NumStatsKind]int
119 : // ForwardStepCount includes Next.
120 : ForwardStepCount [NumStatsKind]int
121 : // ReverseStepCount includes Prev.
122 : ReverseStepCount [NumStatsKind]int
123 : InternalStats InternalIteratorStats
124 : RangeKeyStats RangeKeyIteratorStats
125 : }
126 :
127 : var _ redact.SafeFormatter = &IteratorStats{}
128 :
129 : // InternalIteratorStats contains miscellaneous stats produced by internal
130 : // iterators.
131 : type InternalIteratorStats = base.InternalIteratorStats
132 :
133 : // BlockReadStats contains iterators stats about block reads.
134 : type BlockReadStats = base.BlockReadStats
135 :
136 : // RangeKeyIteratorStats contains miscellaneous stats about range keys
137 : // encountered by the iterator.
138 : type RangeKeyIteratorStats struct {
139 : // Count records the number of range keys encountered during
140 : // iteration. Range keys may be counted multiple times if the iterator
141 : // leaves a range key's bounds and then returns.
142 : Count int
143 : // ContainedPoints records the number of point keys encountered within the
144 : // bounds of a range key. Note that this includes point keys with suffixes
145 : // that sort both above and below the covering range key's suffix.
146 : ContainedPoints int
147 : // SkippedPoints records the count of the subset of ContainedPoints point
148 : // keys that were skipped during iteration due to range-key masking. It does
149 : // not include point keys that were never loaded because a
150 : // RangeKeyMasking.Filter excluded the entire containing block.
151 : SkippedPoints int
152 : }
153 :
154 : // Merge adds all of the argument's statistics to the receiver. It may be used
155 : // to accumulate stats across multiple iterators.
156 0 : func (s *RangeKeyIteratorStats) Merge(o RangeKeyIteratorStats) {
157 0 : s.Count += o.Count
158 0 : s.ContainedPoints += o.ContainedPoints
159 0 : s.SkippedPoints += o.SkippedPoints
160 0 : }
161 :
162 0 : func (s *RangeKeyIteratorStats) String() string {
163 0 : return redact.StringWithoutMarkers(s)
164 0 : }
165 :
166 : // SafeFormat implements the redact.SafeFormatter interface.
167 0 : func (s *RangeKeyIteratorStats) SafeFormat(p redact.SafePrinter, verb rune) {
168 0 : p.Printf("range keys: %s, contained points: %s (%s skipped)",
169 0 : humanize.Count.Uint64(uint64(s.Count)),
170 0 : humanize.Count.Uint64(uint64(s.ContainedPoints)),
171 0 : humanize.Count.Uint64(uint64(s.SkippedPoints)))
172 0 : }
173 :
174 : // LazyValue is a lazy value. See the long comment in base.LazyValue.
175 : type LazyValue = base.LazyValue
176 :
177 : // Iterator iterates over a DB's key/value pairs in key order.
178 : //
179 : // An iterator must be closed after use, but it is not necessary to read an
180 : // iterator until exhaustion.
181 : //
182 : // An iterator is not goroutine-safe, but it is safe to use multiple iterators
183 : // concurrently, with each in a dedicated goroutine.
184 : //
185 : // It is also safe to use an iterator concurrently with modifying its
186 : // underlying DB, if that DB permits modification. However, the resultant
187 : // key/value pairs are not guaranteed to be a consistent snapshot of that DB
188 : // at a particular point in time.
189 : //
190 : // If an iterator encounters an error during any operation, it is stored by
191 : // the Iterator and surfaced through the Error method. All absolute
192 : // positioning methods (eg, SeekLT, SeekGT, First, Last, etc) reset any
193 : // accumulated error before positioning. All relative positioning methods (eg,
194 : // Next, Prev) return without advancing if the iterator has an accumulated
195 : // error.
196 : type Iterator struct {
197 : // The context is stored here since (a) Iterators are expected to be
198 : // short-lived (since they pin memtables and sstables), (b) plumbing a
199 : // context into every method is very painful, (c) they do not (yet) respect
200 : // context cancellation and are only used for tracing.
201 : ctx context.Context
202 : opts IterOptions
203 : merge Merge
204 : comparer base.Comparer
205 : iter internalIterator
206 : pointIter topLevelIterator
207 : // Either readState or version is set, but not both.
208 : readState *readState
209 : version *manifest.Version
210 : // rangeKey holds iteration state specific to iteration over range keys.
211 : // The range key field may be nil if the Iterator has never been configured
212 : // to iterate over range keys. Its non-nilness cannot be used to determine
213 : // if the Iterator is currently iterating over range keys: For that, consult
214 : // the IterOptions using opts.rangeKeys(). If non-nil, its rangeKeyIter
215 : // field is guaranteed to be non-nil too.
216 : rangeKey *iteratorRangeKeyState
217 : // rangeKeyMasking holds state for range-key masking of point keys.
218 : rangeKeyMasking rangeKeyMasking
219 : err error
220 : // When iterValidityState=IterValid, key represents the current key, which
221 : // is backed by keyBuf.
222 : key []byte
223 : keyBuf []byte
224 : value base.InternalValue
225 : // For use in LazyValue.Clone.
226 : valueBuf []byte
227 : fetcher base.LazyFetcher
228 : // For use in LazyValue.Value.
229 : lazyValueBuf []byte
230 : valueCloser io.Closer
231 : // blobValueFetcher is the ValueFetcher to use when retrieving values stored
232 : // externally in blob files.
233 : blobValueFetcher blob.ValueFetcher
234 : // boundsBuf holds two buffers used to store the lower and upper bounds.
235 : // Whenever the Iterator's bounds change, the new bounds are copied into
236 : // boundsBuf[boundsBufIdx]. The two bounds share a slice to reduce
237 : // allocations. opts.LowerBound and opts.UpperBound point into this slice.
238 : boundsBuf [2][]byte
239 : boundsBufIdx int
240 : // iterKV reflects the latest position of iter, except when SetBounds is
241 : // called. In that case, it is explicitly set to nil.
242 : iterKV *base.InternalKV
243 : alloc *iterAlloc
244 : getIterAlloc *getIterAlloc
245 : prefixOrFullSeekKey []byte
246 : readSampling readSampling
247 : stats IteratorStats
248 : externalIter *externalIterState
249 : // Following fields used when constructing an iterator stack, eg, in Clone
250 : // and SetOptions or when re-fragmenting a batch's range keys/range dels.
251 : // Non-nil if this Iterator includes a Batch.
252 : batch *Batch
253 : fc *fileCacheHandle
254 : newIters tableNewIters
255 : newIterRangeKey keyspanimpl.TableNewSpanIter
256 : lazyCombinedIter lazyCombinedIter
257 : seqNum base.SeqNum
258 : // batchSeqNum is used by Iterators over indexed batches to detect when the
259 : // underlying batch has been mutated. The batch beneath an indexed batch may
260 : // be mutated while the Iterator is open, but new keys are not surfaced
261 : // until the next call to SetOptions.
262 : batchSeqNum base.SeqNum
263 : // batch{PointIter,RangeDelIter,RangeKeyIter} are used when the Iterator is
264 : // configured to read through an indexed batch. If a batch is set, these
265 : // iterators will be included within the iterator stack regardless of
266 : // whether the batch currently contains any keys of their kind. These
267 : // pointers are used during a call to SetOptions to refresh the Iterator's
268 : // view of its indexed batch.
269 : batchPointIter batchIter
270 : batchRangeDelIter keyspan.Iter
271 : batchRangeKeyIter keyspan.Iter
272 : // merging is a pointer to this iterator's point merging iterator. It
273 : // appears here because key visibility is handled by the merging iterator.
274 : // During SetOptions on an iterator over an indexed batch, this field is
275 : // used to update the merging iterator's batch snapshot.
276 : merging *mergingIter
277 :
278 : // Keeping the bools here after all the 8 byte aligned fields shrinks the
279 : // sizeof this struct by 24 bytes.
280 :
281 : // INVARIANT:
282 : // iterValidityState==IterAtLimit <=>
283 : // pos==iterPosCurForwardPaused || pos==iterPosCurReversePaused
284 : iterValidityState IterValidityState
285 : // Set to true by SetBounds, SetOptions. Causes the Iterator to appear
286 : // exhausted externally, while preserving the correct iterValidityState for
287 : // the iterator's internal state. Preserving the correct internal validity
288 : // is used for SeekPrefixGE(..., trySeekUsingNext), and SeekGE/SeekLT
289 : // optimizations after "no-op" calls to SetBounds and SetOptions.
290 : requiresReposition bool
291 : // The position of iter. When this is iterPos{Prev,Next} the iter has been
292 : // moved past the current key-value, which can only happen if
293 : // iterValidityState=IterValid, i.e., there is something to return to the
294 : // client for the current position.
295 : pos iterPos
296 : // Relates to the prefixOrFullSeekKey field above.
297 : hasPrefix bool
298 : // Used for deriving the value of SeekPrefixGE(..., trySeekUsingNext),
299 : // and SeekGE/SeekLT optimizations
300 : lastPositioningOp lastPositioningOpKind
301 : // Used for determining when it's safe to perform SeekGE optimizations that
302 : // reuse the iterator state to avoid the cost of a full seek if the iterator
303 : // is already positioned in the correct place. If the iterator's view of its
304 : // indexed batch was just refreshed, some optimizations cannot be applied on
305 : // the first seek after the refresh:
306 : // - SeekGE has a no-op optimization that does not seek on the internal
307 : // iterator at all if the iterator is already in the correct place.
308 : // This optimization cannot be performed if the internal iterator was
309 : // last positioned when the iterator had a different view of an
310 : // underlying batch.
311 : // - Seek[Prefix]GE set flags.TrySeekUsingNext()=true when the seek key is
312 : // greater than the previous operation's seek key, under the expectation
313 : // that the various internal iterators can use their current position to
314 : // avoid a full expensive re-seek. This applies to the batchIter as well.
315 : // However, if the view of the batch was just refreshed, the batchIter's
316 : // position is not useful because it may already be beyond new keys less
317 : // than the seek key. To prevent the use of this optimization in
318 : // batchIter, Seek[Prefix]GE set flags.BatchJustRefreshed()=true if this
319 : // bit is enabled.
320 : batchJustRefreshed bool
321 : // batchOnlyIter is set to true for Batch.NewBatchOnlyIter.
322 : batchOnlyIter bool
323 : // Used in some tests to disable the random disabling of seek optimizations.
324 : forceEnableSeekOpt bool
325 : // Set to true if NextPrefix is not currently permitted. Defaults to false
326 : // in case an iterator never had any bounds.
327 : nextPrefixNotPermittedByUpperBound bool
328 : }
329 :
330 : // cmp is a convenience shorthand for the i.comparer.Compare function.
331 1 : func (i *Iterator) cmp(a, b []byte) int {
332 1 : return i.comparer.Compare(a, b)
333 1 : }
334 :
335 : // equal is a convenience shorthand for the i.comparer.Equal function.
336 1 : func (i *Iterator) equal(a, b []byte) bool {
337 1 : return i.comparer.Equal(a, b)
338 1 : }
339 :
340 : // iteratorRangeKeyState holds an iterator's range key iteration state.
341 : type iteratorRangeKeyState struct {
342 : // rangeKeyIter holds the range key iterator stack that iterates over the
343 : // merged spans across the entirety of the LSM.
344 : rangeKeyIter keyspan.FragmentIterator
345 : iiter keyspan.InterleavingIter
346 : // stale is set to true when the range key state recorded here (in start,
347 : // end and keys) may not be in sync with the current range key at the
348 : // interleaving iterator's current position.
349 : //
350 : // When the interelaving iterator passes over a new span, it invokes the
351 : // SpanChanged hook defined on the `rangeKeyMasking` type, which sets stale
352 : // to true if the span is non-nil.
353 : //
354 : // The parent iterator may not be positioned over the interleaving
355 : // iterator's current position (eg, i.iterPos = iterPos{Next,Prev}), so
356 : // {keys,start,end} are only updated to the new range key during a call to
357 : // Iterator.saveRangeKey.
358 : stale bool
359 : // updated is used to signal to the Iterator client whether the state of
360 : // range keys has changed since the previous iterator position through the
361 : // `RangeKeyChanged` method. It's set to true during an Iterator positioning
362 : // operation that changes the state of the current range key. Each Iterator
363 : // positioning operation sets it back to false before executing.
364 : //
365 : // TODO(jackson): The lifecycle of {stale,updated,prevPosHadRangeKey} is
366 : // intricate and confusing. Try to refactor to reduce complexity.
367 : updated bool
368 : // prevPosHadRangeKey records whether the previous Iterator position had a
369 : // range key (HasPointAndRage() = (_, true)). It's updated at the beginning
370 : // of each new Iterator positioning operation. It's required by saveRangeKey to
371 : // to set `updated` appropriately: Without this record of the previous iterator
372 : // state, it's ambiguous whether an iterator only temporarily stepped onto a
373 : // position without a range key.
374 : prevPosHadRangeKey bool
375 : // rangeKeyOnly is set to true if at the current iterator position there is
376 : // no point key, only a range key start boundary.
377 : rangeKeyOnly bool
378 : // hasRangeKey is true when the current iterator position has a covering
379 : // range key (eg, a range key with bounds [<lower>,<upper>) such that
380 : // <lower> ≤ Key() < <upper>).
381 : hasRangeKey bool
382 : // start and end are the [start, end) boundaries of the current range keys.
383 : start []byte
384 : end []byte
385 :
386 : rangeKeyBuffers
387 :
388 : // iterConfig holds fields that are used for the construction of the
389 : // iterator stack, but do not need to be directly accessed during iteration.
390 : // This struct is bundled within the iteratorRangeKeyState struct to reduce
391 : // allocations.
392 : iterConfig rangekeystack.UserIteratorConfig
393 : }
394 :
395 : type rangeKeyBuffers struct {
396 : // keys is sorted by Suffix ascending.
397 : keys []RangeKeyData
398 : // buf is used to save range-key data before moving the range-key iterator.
399 : // Start and end boundaries, suffixes and values are all copied into buf.
400 : buf bytealloc.A
401 : // internal holds buffers used by the range key internal iterators.
402 : internal rangekeystack.Buffers
403 : }
404 :
405 1 : func (b *rangeKeyBuffers) PrepareForReuse() {
406 1 : const maxKeysReuse = 100
407 1 : if len(b.keys) > maxKeysReuse {
408 0 : b.keys = nil
409 0 : }
410 : // Avoid caching the key buf if it is overly large. The constant is
411 : // fairly arbitrary.
412 1 : if cap(b.buf) >= maxKeyBufCacheSize {
413 1 : b.buf = nil
414 1 : } else {
415 1 : b.buf = b.buf[:0]
416 1 : }
417 1 : b.internal.PrepareForReuse()
418 : }
419 :
420 : var iterRangeKeyStateAllocPool = sync.Pool{
421 1 : New: func() interface{} {
422 1 : return &iteratorRangeKeyState{}
423 1 : },
424 : }
425 :
426 : // isEphemeralPosition returns true iff the current iterator position is
427 : // ephemeral, and won't be visited during subsequent relative positioning
428 : // operations.
429 : //
430 : // The iterator position resulting from a SeekGE or SeekPrefixGE that lands on a
431 : // straddling range key without a coincident point key is such a position.
432 1 : func (i *Iterator) isEphemeralPosition() bool {
433 1 : return i.opts.rangeKeys() && i.rangeKey != nil && i.rangeKey.rangeKeyOnly &&
434 1 : !i.equal(i.rangeKey.start, i.key)
435 1 : }
436 :
437 : type lastPositioningOpKind int8
438 :
439 : const (
440 : unknownLastPositionOp lastPositioningOpKind = iota
441 : seekPrefixGELastPositioningOp
442 : seekGELastPositioningOp
443 : seekLTLastPositioningOp
444 : // internalNextOp is a special internal iterator positioning operation used
445 : // by CanDeterministicallySingleDelete. It exists for enforcing requirements
446 : // around calling CanDeterministicallySingleDelete at most once per external
447 : // iterator position.
448 : internalNextOp
449 : )
450 :
451 : // Limited iteration mode. Not for use with prefix iteration.
452 : //
453 : // SeekGE, SeekLT, Prev, Next have WithLimit variants, that pause the iterator
454 : // at the limit in a best-effort manner. The client should behave correctly
455 : // even if the limits are ignored. These limits are not "deep", in that they
456 : // are not passed down to the underlying collection of internalIterators. This
457 : // is because the limits are transient, and apply only until the next
458 : // iteration call. They serve mainly as a way to bound the amount of work when
459 : // two (or more) Iterators are being coordinated at a higher level.
460 : //
461 : // In limited iteration mode:
462 : // - Avoid using Iterator.Valid if the last call was to a *WithLimit() method.
463 : // The return value from the *WithLimit() method provides a more precise
464 : // disposition.
465 : // - The limit is exclusive for forward and inclusive for reverse.
466 : //
467 : //
468 : // Limited iteration mode & range keys
469 : //
470 : // Limited iteration interacts with range-key iteration. When range key
471 : // iteration is enabled, range keys are interleaved at their start boundaries.
472 : // Limited iteration must ensure that if a range key exists within the limit,
473 : // the iterator visits the range key.
474 : //
475 : // During forward limited iteration, this is trivial: An overlapping range key
476 : // must have a start boundary less than the limit, and the range key's start
477 : // boundary will be interleaved and found to be within the limit.
478 : //
479 : // During reverse limited iteration, the tail of the range key may fall within
480 : // the limit. The range key must be surfaced even if the range key's start
481 : // boundary is less than the limit, and if there are no point keys between the
482 : // current iterator position and the limit. To provide this guarantee, reverse
483 : // limited iteration ignores the limit as long as there is a range key
484 : // overlapping the iteration position.
485 :
486 : // IterValidityState captures the state of the Iterator.
487 : type IterValidityState int8
488 :
489 : const (
490 : // IterExhausted represents an Iterator that is exhausted.
491 : IterExhausted IterValidityState = iota
492 : // IterValid represents an Iterator that is valid.
493 : IterValid
494 : // IterAtLimit represents an Iterator that has a non-exhausted
495 : // internalIterator, but has reached a limit without any key for the
496 : // caller.
497 : IterAtLimit
498 : )
499 :
500 : // readSampling stores variables used to sample a read to trigger a read
501 : // compaction
502 : type readSampling struct {
503 : bytesUntilReadSampling uint64
504 : initialSamplePassed bool
505 : pendingCompactions readCompactionQueue
506 : // forceReadSampling is used for testing purposes to force a read sample on every
507 : // call to Iterator.maybeSampleRead()
508 : forceReadSampling bool
509 : }
510 :
511 1 : func (i *Iterator) findNextEntry(limit []byte) {
512 1 : i.iterValidityState = IterExhausted
513 1 : i.pos = iterPosCurForward
514 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
515 1 : i.rangeKey.rangeKeyOnly = false
516 1 : }
517 :
518 : // Close the closer for the current value if one was open.
519 1 : if i.closeValueCloser() != nil {
520 0 : return
521 0 : }
522 :
523 1 : for i.iterKV != nil {
524 1 : key := i.iterKV.K
525 1 :
526 1 : // The topLevelIterator.StrictSeekPrefixGE contract requires that in
527 1 : // prefix mode [i.hasPrefix=t], every point key returned by the internal
528 1 : // iterator must have the current iteration prefix.
529 1 : if invariants.Enabled && i.hasPrefix {
530 1 : // Range keys are an exception to the contract and may return a different
531 1 : // prefix. This case is explicitly handled in the switch statement below.
532 1 : if key.Kind() != base.InternalKeyKindRangeKeySet {
533 1 : if p := i.comparer.Split.Prefix(key.UserKey); !i.equal(i.prefixOrFullSeekKey, p) {
534 0 : i.opts.logger.Fatalf("pebble: prefix violation: key %q does not have prefix %q\n", key.UserKey, i.prefixOrFullSeekKey)
535 0 : }
536 : }
537 : }
538 :
539 : // Compare with limit every time we start at a different user key.
540 : // Note that given the best-effort contract of limit, we could avoid a
541 : // comparison in the common case by doing this only after
542 : // i.nextUserKey is called for the deletes below. However that makes
543 : // the behavior non-deterministic (since the behavior will vary based
544 : // on what has been compacted), which makes it hard to test with the
545 : // metamorphic test. So we forego that performance optimization.
546 1 : if limit != nil && i.cmp(limit, i.iterKV.K.UserKey) <= 0 {
547 1 : i.iterValidityState = IterAtLimit
548 1 : i.pos = iterPosCurForwardPaused
549 1 : return
550 1 : }
551 :
552 : // If the user has configured a SkipPoint function, invoke it to see
553 : // whether we should skip over the current user key.
554 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(i.iterKV.K.UserKey) {
555 1 : // NB: We could call nextUserKey, but in some cases the SkipPoint
556 1 : // predicate function might be cheaper than nextUserKey's key copy
557 1 : // and key comparison. This should be the case for MVCC suffix
558 1 : // comparisons, for example. In the future, we could expand the
559 1 : // SkipPoint interface to give the implementor more control over
560 1 : // whether we skip over just the internal key, the user key, or even
561 1 : // the key prefix.
562 1 : i.stats.ForwardStepCount[InternalIterCall]++
563 1 : i.iterKV = i.iter.Next()
564 1 : continue
565 : }
566 :
567 1 : switch key.Kind() {
568 1 : case InternalKeyKindRangeKeySet:
569 1 : if i.hasPrefix {
570 1 : if p := i.comparer.Split.Prefix(key.UserKey); !i.equal(i.prefixOrFullSeekKey, p) {
571 1 : return
572 1 : }
573 : }
574 : // Save the current key.
575 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
576 1 : i.key = i.keyBuf
577 1 : i.value = base.InternalValue{}
578 1 : // There may also be a live point key at this userkey that we have
579 1 : // not yet read. We need to find the next entry with this user key
580 1 : // to find it. Save the range key so we don't lose it when we Next
581 1 : // the underlying iterator.
582 1 : i.saveRangeKey()
583 1 : pointKeyExists := i.nextPointCurrentUserKey()
584 1 : if i.err != nil {
585 0 : i.iterValidityState = IterExhausted
586 0 : return
587 0 : }
588 1 : i.rangeKey.rangeKeyOnly = !pointKeyExists
589 1 : i.iterValidityState = IterValid
590 1 : return
591 :
592 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
593 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
594 1 : // only simpler, but is also necessary for correctness due to
595 1 : // InternalKeyKindSSTableInternalObsoleteBit.
596 1 : i.nextUserKey()
597 1 : continue
598 :
599 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
600 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
601 1 : i.key = i.keyBuf
602 1 : i.value = i.iterKV.V
603 1 : i.iterValidityState = IterValid
604 1 : i.saveRangeKey()
605 1 : return
606 :
607 1 : case InternalKeyKindMerge:
608 1 : // Resolving the merge may advance us to the next point key, which
609 1 : // may be covered by a different set of range keys. Save the range
610 1 : // key state so we don't lose it.
611 1 : i.saveRangeKey()
612 1 : if i.mergeForward(key) {
613 1 : i.iterValidityState = IterValid
614 1 : return
615 1 : }
616 :
617 : // The merge didn't yield a valid key, either because the value
618 : // merger indicated it should be deleted, or because an error was
619 : // encountered.
620 0 : i.iterValidityState = IterExhausted
621 0 : if i.err != nil {
622 0 : return
623 0 : }
624 0 : if i.pos != iterPosNext {
625 0 : i.nextUserKey()
626 0 : }
627 0 : if i.closeValueCloser() != nil {
628 0 : return
629 0 : }
630 0 : i.pos = iterPosCurForward
631 :
632 0 : default:
633 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
634 0 : i.iterValidityState = IterExhausted
635 0 : return
636 : }
637 : }
638 :
639 : // Is iterKey nil due to an error?
640 1 : if err := i.iter.Error(); err != nil {
641 0 : i.err = err
642 0 : i.iterValidityState = IterExhausted
643 0 : }
644 : }
645 :
646 1 : func (i *Iterator) nextPointCurrentUserKey() bool {
647 1 : // If the user has configured a SkipPoint function and the current user key
648 1 : // would be skipped by it, there's no need to step forward looking for a
649 1 : // point key. If we were to find one, it should be skipped anyways.
650 1 : if i.opts.SkipPoint != nil && i.opts.SkipPoint(i.key) {
651 1 : return false
652 1 : }
653 :
654 1 : i.pos = iterPosCurForward
655 1 :
656 1 : i.iterKV = i.iter.Next()
657 1 : i.stats.ForwardStepCount[InternalIterCall]++
658 1 : if i.iterKV == nil {
659 1 : if err := i.iter.Error(); err != nil {
660 0 : i.err = err
661 1 : } else {
662 1 : i.pos = iterPosNext
663 1 : }
664 1 : return false
665 : }
666 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
667 1 : i.pos = iterPosNext
668 1 : return false
669 1 : }
670 :
671 1 : key := i.iterKV.K
672 1 : switch key.Kind() {
673 0 : case InternalKeyKindRangeKeySet:
674 0 : // RangeKeySets must always be interleaved as the first internal key
675 0 : // for a user key.
676 0 : i.err = base.CorruptionErrorf("pebble: unexpected range key set mid-user key")
677 0 : return false
678 :
679 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
680 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
681 1 : // only simpler, but is also necessary for correctness due to
682 1 : // InternalKeyKindSSTableInternalObsoleteBit.
683 1 : return false
684 :
685 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
686 1 : i.value = i.iterKV.V
687 1 : return true
688 :
689 1 : case InternalKeyKindMerge:
690 1 : return i.mergeForward(key)
691 :
692 0 : default:
693 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
694 0 : return false
695 : }
696 : }
697 :
698 : // mergeForward resolves a MERGE key, advancing the underlying iterator forward
699 : // to merge with subsequent keys with the same userkey. mergeForward returns a
700 : // boolean indicating whether or not the merge yielded a valid key. A merge may
701 : // not yield a valid key if an error occurred, in which case i.err is non-nil,
702 : // or the user's value merger specified the key to be deleted.
703 : //
704 : // mergeForward does not update iterValidityState.
705 1 : func (i *Iterator) mergeForward(key base.InternalKey) (valid bool) {
706 1 : var iterValue []byte
707 1 : iterValue, _, i.err = i.iterKV.Value(nil)
708 1 : if i.err != nil {
709 0 : return false
710 0 : }
711 1 : var valueMerger ValueMerger
712 1 : valueMerger, i.err = i.merge(key.UserKey, iterValue)
713 1 : if i.err != nil {
714 0 : return false
715 0 : }
716 :
717 1 : i.mergeNext(key, valueMerger)
718 1 : if i.err != nil {
719 0 : return false
720 0 : }
721 :
722 1 : var needDelete bool
723 1 : var value []byte
724 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(
725 1 : valueMerger, true /* includesBase */)
726 1 : i.value = base.MakeInPlaceValue(value)
727 1 : if i.err != nil {
728 0 : return false
729 0 : }
730 1 : if needDelete {
731 0 : _ = i.closeValueCloser()
732 0 : return false
733 0 : }
734 1 : return true
735 : }
736 :
737 1 : func (i *Iterator) closeValueCloser() error {
738 1 : if i.valueCloser != nil {
739 0 : i.err = i.valueCloser.Close()
740 0 : i.valueCloser = nil
741 0 : }
742 1 : return i.err
743 : }
744 :
745 1 : func (i *Iterator) nextUserKey() {
746 1 : if i.iterKV == nil {
747 1 : return
748 1 : }
749 1 : trailer := i.iterKV.K.Trailer
750 1 : done := i.iterKV.K.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
751 1 : if i.iterValidityState != IterValid {
752 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKV.K.UserKey...)
753 1 : i.key = i.keyBuf
754 1 : }
755 1 : for {
756 1 : i.stats.ForwardStepCount[InternalIterCall]++
757 1 : i.iterKV = i.iter.Next()
758 1 : if i.iterKV == nil {
759 1 : if err := i.iter.Error(); err != nil {
760 0 : i.err = err
761 0 : return
762 0 : }
763 : }
764 : // NB: We're guaranteed to be on the next user key if the previous key
765 : // had a zero sequence number (`done`), or the new key has a trailer
766 : // greater or equal to the previous key's trailer. This is true because
767 : // internal keys with the same user key are sorted by InternalKeyTrailer in
768 : // strictly monotonically descending order. We expect the trailer
769 : // optimization to trigger around 50% of the time with randomly
770 : // distributed writes. We expect it to trigger very frequently when
771 : // iterating through ingested sstables, which contain keys that all have
772 : // the same sequence number.
773 1 : if done || i.iterKV == nil || i.iterKV.K.Trailer >= trailer {
774 1 : break
775 : }
776 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
777 1 : break
778 : }
779 1 : done = i.iterKV.K.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
780 1 : trailer = i.iterKV.K.Trailer
781 : }
782 : }
783 :
784 1 : func (i *Iterator) maybeSampleRead() {
785 1 : // This method is only called when a public method of Iterator is
786 1 : // returning, and below we exclude the case were the iterator is paused at
787 1 : // a limit. The effect of these choices is that keys that are deleted, but
788 1 : // are encountered during iteration, are not accounted for in the read
789 1 : // sampling and will not cause read driven compactions, even though we are
790 1 : // incurring cost in iterating over them. And this issue is not limited to
791 1 : // Iterator, which does not see the effect of range deletes, which may be
792 1 : // causing iteration work in mergingIter. It is not clear at this time
793 1 : // whether this is a deficiency worth addressing.
794 1 : if i.iterValidityState != IterValid {
795 1 : return
796 1 : }
797 1 : if i.readState == nil {
798 1 : return
799 1 : }
800 1 : if i.readSampling.forceReadSampling {
801 0 : i.sampleRead()
802 0 : return
803 0 : }
804 1 : samplingPeriod := int32(int64(readBytesPeriod) * i.readState.db.opts.Experimental.ReadSamplingMultiplier)
805 1 : if samplingPeriod <= 0 {
806 0 : return
807 0 : }
808 1 : bytesRead := uint64(len(i.key) + i.value.Len())
809 1 : for i.readSampling.bytesUntilReadSampling < bytesRead {
810 1 : i.readSampling.bytesUntilReadSampling += uint64(rand.Uint32N(2 * uint32(samplingPeriod)))
811 1 : // The block below tries to adjust for the case where this is the
812 1 : // first read in a newly-opened iterator. As bytesUntilReadSampling
813 1 : // starts off at zero, we don't want to sample the first read of
814 1 : // every newly-opened iterator, but we do want to sample some of them.
815 1 : if !i.readSampling.initialSamplePassed {
816 1 : i.readSampling.initialSamplePassed = true
817 1 : if i.readSampling.bytesUntilReadSampling > bytesRead {
818 1 : if rand.Uint64N(i.readSampling.bytesUntilReadSampling) > bytesRead {
819 1 : continue
820 : }
821 : }
822 : }
823 1 : i.sampleRead()
824 : }
825 1 : i.readSampling.bytesUntilReadSampling -= bytesRead
826 : }
827 :
828 1 : func (i *Iterator) sampleRead() {
829 1 : var topFile *manifest.TableMetadata
830 1 : topLevel, numOverlappingLevels := numLevels, 0
831 1 : mi := i.merging
832 1 : if mi == nil {
833 1 : return
834 1 : }
835 1 : if len(mi.levels) > 1 {
836 1 : mi.ForEachLevelIter(func(li *levelIter) (done bool) {
837 1 : if li.layer.IsFlushableIngests() {
838 0 : return false
839 0 : }
840 1 : l := li.layer.Level()
841 1 : if f := li.iterFile; f != nil {
842 1 : var containsKey bool
843 1 : if i.pos == iterPosNext || i.pos == iterPosCurForward ||
844 1 : i.pos == iterPosCurForwardPaused {
845 1 : containsKey = i.cmp(f.PointKeyBounds.SmallestUserKey(), i.key) <= 0
846 1 : } else if i.pos == iterPosPrev || i.pos == iterPosCurReverse ||
847 1 : i.pos == iterPosCurReversePaused {
848 1 : containsKey = i.cmp(f.PointKeyBounds.LargestUserKey(), i.key) >= 0
849 1 : }
850 : // Do nothing if the current key is not contained in f's
851 : // bounds. We could seek the LevelIterator at this level
852 : // to find the right file, but the performance impacts of
853 : // doing that are significant enough to negate the benefits
854 : // of read sampling in the first place. See the discussion
855 : // at:
856 : // https://github.com/cockroachdb/pebble/pull/1041#issuecomment-763226492
857 1 : if containsKey {
858 1 : numOverlappingLevels++
859 1 : if numOverlappingLevels >= 2 {
860 1 : // Terminate the loop early if at least 2 overlapping levels are found.
861 1 : return true
862 1 : }
863 1 : topLevel = l
864 1 : topFile = f
865 : }
866 : }
867 1 : return false
868 : })
869 : }
870 1 : if topFile == nil || topLevel >= numLevels {
871 1 : return
872 1 : }
873 1 : if numOverlappingLevels >= 2 {
874 1 : allowedSeeks := topFile.AllowedSeeks.Add(-1)
875 1 : if allowedSeeks == 0 {
876 0 :
877 0 : // Since the compaction queue can handle duplicates, we can keep
878 0 : // adding to the queue even once allowedSeeks hits 0.
879 0 : // In fact, we NEED to keep adding to the queue, because the queue
880 0 : // is small and evicts older and possibly useful compactions.
881 0 : topFile.AllowedSeeks.Add(topFile.InitAllowedSeeks)
882 0 :
883 0 : read := readCompaction{
884 0 : start: topFile.PointKeyBounds.SmallestUserKey(),
885 0 : end: topFile.PointKeyBounds.LargestUserKey(),
886 0 : level: topLevel,
887 0 : tableNum: topFile.TableNum,
888 0 : }
889 0 : i.readSampling.pendingCompactions.add(&read, i.cmp)
890 0 : }
891 : }
892 : }
893 :
894 1 : func (i *Iterator) findPrevEntry(limit []byte) {
895 1 : i.iterValidityState = IterExhausted
896 1 : i.pos = iterPosCurReverse
897 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
898 1 : i.rangeKey.rangeKeyOnly = false
899 1 : }
900 :
901 : // Close the closer for the current value if one was open.
902 1 : if i.valueCloser != nil {
903 0 : i.err = i.valueCloser.Close()
904 0 : i.valueCloser = nil
905 0 : if i.err != nil {
906 0 : i.iterValidityState = IterExhausted
907 0 : return
908 0 : }
909 : }
910 :
911 1 : var valueMerger ValueMerger
912 1 : firstLoopIter := true
913 1 : rangeKeyBoundary := false
914 1 : // The code below compares with limit in multiple places. As documented in
915 1 : // findNextEntry, this is being done to make the behavior of limit
916 1 : // deterministic to allow for metamorphic testing. It is not required by
917 1 : // the best-effort contract of limit.
918 1 : for i.iterKV != nil {
919 1 : key := i.iterKV.K
920 1 :
921 1 : // NB: We cannot pause if the current key is covered by a range key.
922 1 : // Otherwise, the user might not ever learn of a range key that covers
923 1 : // the key space being iterated over in which there are no point keys.
924 1 : // Since limits are best effort, ignoring the limit in this case is
925 1 : // allowed by the contract of limit.
926 1 : if firstLoopIter && limit != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
927 1 : i.iterValidityState = IterAtLimit
928 1 : i.pos = iterPosCurReversePaused
929 1 : return
930 1 : }
931 1 : firstLoopIter = false
932 1 :
933 1 : if i.iterValidityState == IterValid {
934 1 : if !i.equal(key.UserKey, i.key) {
935 1 : // We've iterated to the previous user key.
936 1 : i.pos = iterPosPrev
937 1 : if valueMerger != nil {
938 1 : var needDelete bool
939 1 : var value []byte
940 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
941 1 : i.value = base.MakeInPlaceValue(value)
942 1 : if i.err == nil && needDelete {
943 0 : // The point key at this key is deleted. If we also have
944 0 : // a range key boundary at this key, we still want to
945 0 : // return. Otherwise, we need to continue looking for
946 0 : // a live key.
947 0 : i.value = base.InternalValue{}
948 0 : if rangeKeyBoundary {
949 0 : i.rangeKey.rangeKeyOnly = true
950 0 : } else {
951 0 : i.iterValidityState = IterExhausted
952 0 : if i.closeValueCloser() == nil {
953 0 : continue
954 : }
955 : }
956 : }
957 : }
958 1 : if i.err != nil {
959 0 : i.iterValidityState = IterExhausted
960 0 : }
961 1 : return
962 : }
963 : }
964 :
965 : // If the user has configured a SkipPoint function, invoke it to see
966 : // whether we should skip over the current user key.
967 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(key.UserKey) {
968 1 : // NB: We could call prevUserKey, but in some cases the SkipPoint
969 1 : // predicate function might be cheaper than prevUserKey's key copy
970 1 : // and key comparison. This should be the case for MVCC suffix
971 1 : // comparisons, for example. In the future, we could expand the
972 1 : // SkipPoint interface to give the implementor more control over
973 1 : // whether we skip over just the internal key, the user key, or even
974 1 : // the key prefix.
975 1 : i.stats.ReverseStepCount[InternalIterCall]++
976 1 : i.iterKV = i.iter.Prev()
977 1 : if i.iterKV == nil {
978 1 : if err := i.iter.Error(); err != nil {
979 0 : i.err = err
980 0 : i.iterValidityState = IterExhausted
981 0 : return
982 0 : }
983 : }
984 1 : if limit != nil && i.iterKV != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
985 1 : i.iterValidityState = IterAtLimit
986 1 : i.pos = iterPosCurReversePaused
987 1 : return
988 1 : }
989 1 : continue
990 : }
991 :
992 1 : switch key.Kind() {
993 1 : case InternalKeyKindRangeKeySet:
994 1 : // Range key start boundary markers are interleaved with the maximum
995 1 : // sequence number, so if there's a point key also at this key, we
996 1 : // must've already iterated over it.
997 1 : // This is the final entry at this user key, so we may return
998 1 : i.rangeKey.rangeKeyOnly = i.iterValidityState != IterValid
999 1 : if i.rangeKey.rangeKeyOnly {
1000 1 : // The point iterator is now invalid, so clear the point value.
1001 1 : i.value = base.InternalValue{}
1002 1 : }
1003 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1004 1 : i.key = i.keyBuf
1005 1 : i.iterValidityState = IterValid
1006 1 : i.saveRangeKey()
1007 1 : // In all other cases, previous iteration requires advancing to
1008 1 : // iterPosPrev in order to determine if the key is live and
1009 1 : // unshadowed by another key at the same user key. In this case,
1010 1 : // because range key start boundary markers are always interleaved
1011 1 : // at the maximum sequence number, we know that there aren't any
1012 1 : // additional keys with the same user key in the backward direction.
1013 1 : //
1014 1 : // We Prev the underlying iterator once anyways for consistency, so
1015 1 : // that we can maintain the invariant during backward iteration that
1016 1 : // i.iterPos = iterPosPrev.
1017 1 : i.stats.ReverseStepCount[InternalIterCall]++
1018 1 : i.iterKV = i.iter.Prev()
1019 1 :
1020 1 : // Set rangeKeyBoundary so that on the next iteration, we know to
1021 1 : // return the key even if the MERGE point key is deleted.
1022 1 : rangeKeyBoundary = true
1023 :
1024 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
1025 1 : i.value = base.InternalValue{}
1026 1 : i.iterValidityState = IterExhausted
1027 1 : valueMerger = nil
1028 1 : i.stats.ReverseStepCount[InternalIterCall]++
1029 1 : i.iterKV = i.iter.Prev()
1030 1 : // Compare with the limit. We could optimize by only checking when
1031 1 : // we step to the previous user key, but detecting that requires a
1032 1 : // comparison too. Note that this position may already passed a
1033 1 : // number of versions of this user key, but they are all deleted, so
1034 1 : // the fact that a subsequent Prev*() call will not see them is
1035 1 : // harmless. Also note that this is the only place in the loop,
1036 1 : // other than the firstLoopIter and SkipPoint cases above, where we
1037 1 : // could step to a different user key and start processing it for
1038 1 : // returning to the caller.
1039 1 : if limit != nil && i.iterKV != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
1040 1 : i.iterValidityState = IterAtLimit
1041 1 : i.pos = iterPosCurReversePaused
1042 1 : return
1043 1 : }
1044 1 : continue
1045 :
1046 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1047 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1048 1 : i.key = i.keyBuf
1049 1 : // iterValue is owned by i.iter and could change after the Prev()
1050 1 : // call, so use valueBuf instead. Note that valueBuf is only used
1051 1 : // in this one instance; everywhere else (eg. in findNextEntry),
1052 1 : // we just point i.value to the unsafe i.iter-owned value buffer.
1053 1 : i.value, i.valueBuf = i.iterKV.V.Clone(i.valueBuf[:0], &i.fetcher)
1054 1 : i.saveRangeKey()
1055 1 : i.iterValidityState = IterValid
1056 1 : i.iterKV = i.iter.Prev()
1057 1 : i.stats.ReverseStepCount[InternalIterCall]++
1058 1 : valueMerger = nil
1059 1 : continue
1060 :
1061 1 : case InternalKeyKindMerge:
1062 1 : if i.iterValidityState == IterExhausted {
1063 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1064 1 : i.key = i.keyBuf
1065 1 : i.saveRangeKey()
1066 1 : var iterValue []byte
1067 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1068 1 : if i.err != nil {
1069 0 : return
1070 0 : }
1071 1 : valueMerger, i.err = i.merge(i.key, iterValue)
1072 1 : if i.err != nil {
1073 0 : return
1074 0 : }
1075 1 : i.iterValidityState = IterValid
1076 1 : } else if valueMerger == nil {
1077 1 : // Extract value before iterValue since we use value before iterValue
1078 1 : // and the underlying iterator is not required to provide backing
1079 1 : // memory for both simultaneously.
1080 1 : var value []byte
1081 1 : var callerOwned bool
1082 1 : value, callerOwned, i.err = i.value.Value(i.lazyValueBuf)
1083 1 : if callerOwned {
1084 0 : i.lazyValueBuf = value[:0]
1085 0 : }
1086 1 : if i.err != nil {
1087 0 : i.iterValidityState = IterExhausted
1088 0 : return
1089 0 : }
1090 1 : valueMerger, i.err = i.merge(i.key, value)
1091 1 : var iterValue []byte
1092 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1093 1 : if i.err != nil {
1094 0 : i.iterValidityState = IterExhausted
1095 0 : return
1096 0 : }
1097 1 : if i.err == nil {
1098 1 : i.err = valueMerger.MergeNewer(iterValue)
1099 1 : }
1100 1 : if i.err != nil {
1101 0 : i.iterValidityState = IterExhausted
1102 0 : return
1103 0 : }
1104 1 : } else {
1105 1 : var iterValue []byte
1106 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1107 1 : if i.err != nil {
1108 0 : i.iterValidityState = IterExhausted
1109 0 : return
1110 0 : }
1111 1 : i.err = valueMerger.MergeNewer(iterValue)
1112 1 : if i.err != nil {
1113 0 : i.iterValidityState = IterExhausted
1114 0 : return
1115 0 : }
1116 : }
1117 1 : i.iterKV = i.iter.Prev()
1118 1 : i.stats.ReverseStepCount[InternalIterCall]++
1119 1 : continue
1120 :
1121 0 : default:
1122 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1123 0 : i.iterValidityState = IterExhausted
1124 0 : return
1125 : }
1126 : }
1127 : // i.iterKV == nil, so broke out of the preceding loop.
1128 :
1129 : // Is iterKey nil due to an error?
1130 1 : if i.err = i.iter.Error(); i.err != nil {
1131 0 : i.iterValidityState = IterExhausted
1132 0 : return
1133 0 : }
1134 :
1135 1 : if i.iterValidityState == IterValid {
1136 1 : i.pos = iterPosPrev
1137 1 : if valueMerger != nil {
1138 1 : var needDelete bool
1139 1 : var value []byte
1140 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
1141 1 : i.value = base.MakeInPlaceValue(value)
1142 1 : if i.err == nil && needDelete {
1143 0 : i.key = nil
1144 0 : i.value = base.InternalValue{}
1145 0 : i.iterValidityState = IterExhausted
1146 0 : }
1147 : }
1148 1 : if i.err != nil {
1149 0 : i.iterValidityState = IterExhausted
1150 0 : }
1151 : }
1152 : }
1153 :
1154 1 : func (i *Iterator) prevUserKey() {
1155 1 : if i.iterKV == nil {
1156 1 : return
1157 1 : }
1158 1 : if i.iterValidityState != IterValid {
1159 1 : // If we're going to compare against the prev key, we need to save the
1160 1 : // current key.
1161 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKV.K.UserKey...)
1162 1 : i.key = i.keyBuf
1163 1 : }
1164 1 : for {
1165 1 : i.iterKV = i.iter.Prev()
1166 1 : i.stats.ReverseStepCount[InternalIterCall]++
1167 1 : if i.iterKV == nil {
1168 1 : if err := i.iter.Error(); err != nil {
1169 0 : i.err = err
1170 0 : i.iterValidityState = IterExhausted
1171 0 : }
1172 1 : break
1173 : }
1174 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
1175 1 : break
1176 : }
1177 : }
1178 : }
1179 :
1180 1 : func (i *Iterator) mergeNext(key InternalKey, valueMerger ValueMerger) {
1181 1 : // Save the current key.
1182 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1183 1 : i.key = i.keyBuf
1184 1 :
1185 1 : // Loop looking for older values for this key and merging them.
1186 1 : for {
1187 1 : i.iterKV = i.iter.Next()
1188 1 : i.stats.ForwardStepCount[InternalIterCall]++
1189 1 : if i.iterKV == nil {
1190 1 : if i.err = i.iter.Error(); i.err != nil {
1191 0 : return
1192 0 : }
1193 1 : i.pos = iterPosNext
1194 1 : return
1195 : }
1196 1 : key = i.iterKV.K
1197 1 : if !i.equal(i.key, key.UserKey) {
1198 1 : // We've advanced to the next key.
1199 1 : i.pos = iterPosNext
1200 1 : return
1201 1 : }
1202 1 : switch key.Kind() {
1203 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
1204 1 : // We've hit a deletion tombstone. Return everything up to this
1205 1 : // point.
1206 1 : //
1207 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
1208 1 : // only simpler, but is also necessary for correctness due to
1209 1 : // InternalKeyKindSSTableInternalObsoleteBit.
1210 1 : return
1211 :
1212 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1213 1 : // We've hit a Set value. Merge with the existing value and return.
1214 1 : var iterValue []byte
1215 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1216 1 : if i.err != nil {
1217 0 : return
1218 0 : }
1219 1 : i.err = valueMerger.MergeOlder(iterValue)
1220 1 : return
1221 :
1222 1 : case InternalKeyKindMerge:
1223 1 : // We've hit another Merge value. Merge with the existing value and
1224 1 : // continue looping.
1225 1 : var iterValue []byte
1226 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1227 1 : if i.err != nil {
1228 0 : return
1229 0 : }
1230 1 : i.err = valueMerger.MergeOlder(iterValue)
1231 1 : if i.err != nil {
1232 0 : return
1233 0 : }
1234 1 : continue
1235 :
1236 0 : case InternalKeyKindRangeKeySet:
1237 0 : // The RANGEKEYSET marker must sort before a MERGE at the same user key.
1238 0 : i.err = base.CorruptionErrorf("pebble: out of order range key marker")
1239 0 : return
1240 :
1241 0 : default:
1242 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1243 0 : return
1244 : }
1245 : }
1246 : }
1247 :
1248 : // SeekGE moves the iterator to the first key/value pair whose key is greater
1249 : // than or equal to the given key. Returns true if the iterator is pointing at
1250 : // a valid entry and false otherwise.
1251 1 : func (i *Iterator) SeekGE(key []byte) bool {
1252 1 : return i.SeekGEWithLimit(key, nil) == IterValid
1253 1 : }
1254 :
1255 : // SeekGEWithLimit moves the iterator to the first key/value pair whose key is
1256 : // greater than or equal to the given key.
1257 : //
1258 : // If limit is provided, it serves as a best-effort exclusive limit. If the
1259 : // first key greater than or equal to the given search key is also greater than
1260 : // or equal to limit, the Iterator may pause and return IterAtLimit. Because
1261 : // limits are best-effort, SeekGEWithLimit may return a key beyond limit.
1262 : //
1263 : // If the Iterator is configured to iterate over range keys, SeekGEWithLimit
1264 : // guarantees it will surface any range keys with bounds overlapping the
1265 : // keyspace [key, limit).
1266 1 : func (i *Iterator) SeekGEWithLimit(key []byte, limit []byte) IterValidityState {
1267 1 : if i.rangeKey != nil {
1268 1 : // NB: Check Valid() before clearing requiresReposition.
1269 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1270 1 : // If we have a range key but did not expose it at the previous iterator
1271 1 : // position (because the iterator was not at a valid position), updated
1272 1 : // must be true. This ensures that after an iterator op sequence like:
1273 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1274 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1275 1 : // - SeekGE(...) → (IterValid, RangeBounds() = [a,b))
1276 1 : // the iterator returns RangeKeyChanged()=true.
1277 1 : //
1278 1 : // The remainder of this function will only update i.rangeKey.updated if
1279 1 : // the iterator moves into a new range key, or out of the current range
1280 1 : // key.
1281 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1282 1 : }
1283 1 : lastPositioningOp := i.lastPositioningOp
1284 1 : // Set it to unknown, since this operation may not succeed, in which case
1285 1 : // the SeekGE following this should not make any assumption about iterator
1286 1 : // position.
1287 1 : i.lastPositioningOp = unknownLastPositionOp
1288 1 : i.requiresReposition = false
1289 1 : i.err = nil // clear cached iteration error
1290 1 : i.hasPrefix = false
1291 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1292 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1293 1 : key = lowerBound
1294 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1295 1 : key = upperBound
1296 1 : }
1297 1 : seekInternalIter := true
1298 1 :
1299 1 : var flags base.SeekGEFlags
1300 1 : if i.batchJustRefreshed {
1301 0 : i.batchJustRefreshed = false
1302 0 : flags = flags.EnableBatchJustRefreshed()
1303 0 : }
1304 1 : if lastPositioningOp == seekGELastPositioningOp {
1305 1 : cmp := i.cmp(i.prefixOrFullSeekKey, key)
1306 1 : // If this seek is to the same or later key, and the iterator is
1307 1 : // already positioned there, this is a noop. This can be helpful for
1308 1 : // sparse key spaces that have many deleted keys, where one can avoid
1309 1 : // the overhead of iterating past them again and again.
1310 1 : if cmp <= 0 {
1311 1 : if !flags.BatchJustRefreshed() &&
1312 1 : (i.iterValidityState == IterExhausted ||
1313 1 : (i.iterValidityState == IterValid && i.cmp(key, i.key) <= 0 &&
1314 1 : (limit == nil || i.cmp(i.key, limit) < 0))) {
1315 1 : // Noop
1316 1 : if i.forceEnableSeekOpt || !testingDisableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1317 1 : i.lastPositioningOp = seekGELastPositioningOp
1318 1 : return i.iterValidityState
1319 1 : }
1320 : }
1321 : // cmp == 0 is not safe to optimize since
1322 : // - i.pos could be at iterPosNext, due to a merge.
1323 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1324 : // SET pair for a key, and the iterator would have moved past DELETE
1325 : // but stayed at iterPosCurForward. A similar situation occurs for a
1326 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1327 : // at the SET.
1328 : // We also leverage the IterAtLimit <=> i.pos invariant defined in the
1329 : // comment on iterValidityState, to exclude any cases where i.pos
1330 : // is iterPosCur{Forward,Reverse}Paused. This avoids the need to
1331 : // special-case those iterator positions and their interactions with
1332 : // TrySeekUsingNext, as the main uses for TrySeekUsingNext in CockroachDB
1333 : // do not use limited Seeks in the first place.
1334 1 : if cmp < 0 && i.iterValidityState != IterAtLimit && limit == nil {
1335 1 : flags = flags.EnableTrySeekUsingNext()
1336 1 : }
1337 1 : if testingDisableSeekOpt(key, uintptr(unsafe.Pointer(i))) && !i.forceEnableSeekOpt {
1338 1 : flags = flags.DisableTrySeekUsingNext()
1339 1 : }
1340 1 : if !flags.BatchJustRefreshed() && i.pos == iterPosCurForwardPaused && i.cmp(key, i.iterKV.K.UserKey) <= 0 {
1341 1 : // Have some work to do, but don't need to seek, and we can
1342 1 : // start doing findNextEntry from i.iterKey.
1343 1 : seekInternalIter = false
1344 1 : }
1345 : }
1346 : }
1347 1 : if seekInternalIter {
1348 1 : i.iterKV = i.iter.SeekGE(key, flags)
1349 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1350 1 : if err := i.iter.Error(); err != nil {
1351 0 : i.err = err
1352 0 : i.iterValidityState = IterExhausted
1353 0 : return i.iterValidityState
1354 0 : }
1355 : }
1356 1 : i.findNextEntry(limit)
1357 1 : i.maybeSampleRead()
1358 1 : if i.Error() == nil {
1359 1 : // Prepare state for a future noop optimization.
1360 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1361 1 : i.lastPositioningOp = seekGELastPositioningOp
1362 1 : }
1363 1 : return i.iterValidityState
1364 : }
1365 :
1366 : // SeekPrefixGE moves the iterator to the first key/value pair whose key is
1367 : // greater than or equal to the given key and which has the same "prefix" as
1368 : // the given key. The prefix for a key is determined by the user-defined
1369 : // Comparer.Split function. The iterator will not observe keys not matching the
1370 : // "prefix" of the search key. Calling SeekPrefixGE puts the iterator in prefix
1371 : // iteration mode. The iterator remains in prefix iteration until a subsequent
1372 : // call to another absolute positioning method (SeekGE, SeekLT, First,
1373 : // Last). Reverse iteration (Prev) is not supported when an iterator is in
1374 : // prefix iteration mode. Returns true if the iterator is pointing at a valid
1375 : // entry and false otherwise.
1376 : //
1377 : // The semantics of SeekPrefixGE are slightly unusual and designed for
1378 : // iteration to be able to take advantage of bloom filters that have been
1379 : // created on the "prefix". If you're not using bloom filters, there is no
1380 : // reason to use SeekPrefixGE.
1381 : //
1382 : // An example Split function may separate a timestamp suffix from the prefix of
1383 : // the key.
1384 : //
1385 : // Split(<key>@<timestamp>) -> <key>
1386 : //
1387 : // Consider the keys "a@1", "a@2", "aa@3", "aa@4". The prefixes for these keys
1388 : // are "a", and "aa". Note that despite "a" and "aa" sharing a prefix by the
1389 : // usual definition, those prefixes differ by the definition of the Split
1390 : // function. To see how this works, consider the following set of calls on this
1391 : // data set:
1392 : //
1393 : // SeekPrefixGE("a@0") -> "a@1"
1394 : // Next() -> "a@2"
1395 : // Next() -> EOF
1396 : //
1397 : // If you're just looking to iterate over keys with a shared prefix, as
1398 : // defined by the configured comparer, set iterator bounds instead:
1399 : //
1400 : // iter := db.NewIter(&pebble.IterOptions{
1401 : // LowerBound: []byte("prefix"),
1402 : // UpperBound: []byte("prefiy"),
1403 : // })
1404 : // for iter.First(); iter.Valid(); iter.Next() {
1405 : // // Only keys beginning with "prefix" will be visited.
1406 : // }
1407 : //
1408 : // See ExampleIterator_SeekPrefixGE for a working example.
1409 : //
1410 : // When iterating with range keys enabled, all range keys encountered are
1411 : // truncated to the seek key's prefix's bounds. The truncation of the upper
1412 : // bound requires that the database's Comparer is configured with a
1413 : // ImmediateSuccessor method. For example, a SeekPrefixGE("a@9") call with the
1414 : // prefix "a" will truncate range key bounds to [a,ImmediateSuccessor(a)].
1415 1 : func (i *Iterator) SeekPrefixGE(key []byte) bool {
1416 1 : if i.rangeKey != nil {
1417 1 : // NB: Check Valid() before clearing requiresReposition.
1418 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1419 1 : // If we have a range key but did not expose it at the previous iterator
1420 1 : // position (because the iterator was not at a valid position), updated
1421 1 : // must be true. This ensures that after an iterator op sequence like:
1422 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1423 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1424 1 : // - SeekPrefixGE(...) → (IterValid, RangeBounds() = [a,b))
1425 1 : // the iterator returns RangeKeyChanged()=true.
1426 1 : //
1427 1 : // The remainder of this function will only update i.rangeKey.updated if
1428 1 : // the iterator moves into a new range key, or out of the current range
1429 1 : // key.
1430 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1431 1 : }
1432 1 : lastPositioningOp := i.lastPositioningOp
1433 1 : // Set it to unknown, since this operation may not succeed, in which case
1434 1 : // the SeekPrefixGE following this should not make any assumption about
1435 1 : // iterator position.
1436 1 : i.lastPositioningOp = unknownLastPositionOp
1437 1 : i.requiresReposition = false
1438 1 : i.err = nil // clear cached iteration error
1439 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1440 1 : if i.comparer.ImmediateSuccessor == nil && i.opts.KeyTypes != IterKeyTypePointsOnly {
1441 0 : panic("pebble: ImmediateSuccessor must be provided for SeekPrefixGE with range keys")
1442 : }
1443 1 : prefixLen := i.comparer.Split(key)
1444 1 : keyPrefix := key[:prefixLen]
1445 1 : var flags base.SeekGEFlags
1446 1 : if i.batchJustRefreshed {
1447 0 : flags = flags.EnableBatchJustRefreshed()
1448 0 : i.batchJustRefreshed = false
1449 0 : }
1450 1 : if lastPositioningOp == seekPrefixGELastPositioningOp {
1451 1 : if !i.hasPrefix {
1452 0 : panic("lastPositioningOpsIsSeekPrefixGE is true, but hasPrefix is false")
1453 : }
1454 : // The iterator has not been repositioned after the last SeekPrefixGE.
1455 : // See if we are seeking to a larger key, since then we can optimize
1456 : // the seek by using next. Note that we could also optimize if Next
1457 : // has been called, if the iterator is not exhausted and the current
1458 : // position is <= the seek key. We are keeping this limited for now
1459 : // since such optimizations require care for correctness, and to not
1460 : // become de-optimizations (if one usually has to do all the next
1461 : // calls and then the seek). This SeekPrefixGE optimization
1462 : // specifically benefits CockroachDB.
1463 1 : cmp := i.cmp(i.prefixOrFullSeekKey, keyPrefix)
1464 1 : // cmp == 0 is not safe to optimize since
1465 1 : // - i.pos could be at iterPosNext, due to a merge.
1466 1 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1467 1 : // SET pair for a key, and the iterator would have moved past DELETE
1468 1 : // but stayed at iterPosCurForward. A similar situation occurs for a
1469 1 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1470 1 : // at the SET.
1471 1 : // In general some versions of i.prefix could have been consumed by
1472 1 : // the iterator, so we only optimize for cmp < 0.
1473 1 : if cmp < 0 {
1474 1 : flags = flags.EnableTrySeekUsingNext()
1475 1 : }
1476 1 : if testingDisableSeekOpt(key, uintptr(unsafe.Pointer(i))) && !i.forceEnableSeekOpt {
1477 1 : flags = flags.DisableTrySeekUsingNext()
1478 1 : }
1479 : }
1480 : // Make a copy of the prefix so that modifications to the key after
1481 : // SeekPrefixGE returns does not affect the stored prefix.
1482 1 : if cap(i.prefixOrFullSeekKey) < prefixLen {
1483 1 : i.prefixOrFullSeekKey = make([]byte, prefixLen)
1484 1 : } else {
1485 1 : i.prefixOrFullSeekKey = i.prefixOrFullSeekKey[:prefixLen]
1486 1 : }
1487 1 : i.hasPrefix = true
1488 1 : copy(i.prefixOrFullSeekKey, keyPrefix)
1489 1 :
1490 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1491 1 : if p := i.comparer.Split.Prefix(lowerBound); !bytes.Equal(i.prefixOrFullSeekKey, p) {
1492 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of lower bound")
1493 1 : i.iterValidityState = IterExhausted
1494 1 : return false
1495 1 : }
1496 1 : key = lowerBound
1497 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1498 1 : if p := i.comparer.Split.Prefix(upperBound); !bytes.Equal(i.prefixOrFullSeekKey, p) {
1499 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of upper bound")
1500 1 : i.iterValidityState = IterExhausted
1501 1 : return false
1502 1 : }
1503 1 : key = upperBound
1504 : }
1505 1 : i.iterKV = i.iter.SeekPrefixGE(i.prefixOrFullSeekKey, key, flags)
1506 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1507 1 : i.findNextEntry(nil)
1508 1 : i.maybeSampleRead()
1509 1 : if i.Error() == nil {
1510 1 : i.lastPositioningOp = seekPrefixGELastPositioningOp
1511 1 : }
1512 1 : return i.iterValidityState == IterValid
1513 : }
1514 :
1515 : // Deterministic disabling (in testing mode) of the seek optimizations. It uses
1516 : // the iterator pointer, since we want diversity in iterator behavior for the
1517 : // same key. Used for tests.
1518 1 : func testingDisableSeekOpt(key []byte, ptr uintptr) bool {
1519 1 : if !invariants.Enabled {
1520 0 : return false
1521 0 : }
1522 : // Fibonacci hash https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
1523 1 : simpleHash := (11400714819323198485 * uint64(ptr)) >> 63
1524 1 : return key != nil && key[0]&byte(1) == 0 && simpleHash == 0
1525 : }
1526 :
1527 : // SeekLT moves the iterator to the last key/value pair whose key is less than
1528 : // the given key. Returns true if the iterator is pointing at a valid entry and
1529 : // false otherwise.
1530 1 : func (i *Iterator) SeekLT(key []byte) bool {
1531 1 : return i.SeekLTWithLimit(key, nil) == IterValid
1532 1 : }
1533 :
1534 : // SeekLTWithLimit moves the iterator to the last key/value pair whose key is
1535 : // less than the given key.
1536 : //
1537 : // If limit is provided, it serves as a best-effort inclusive limit. If the last
1538 : // key less than the given search key is also less than limit, the Iterator may
1539 : // pause and return IterAtLimit. Because limits are best-effort, SeekLTWithLimit
1540 : // may return a key beyond limit.
1541 : //
1542 : // If the Iterator is configured to iterate over range keys, SeekLTWithLimit
1543 : // guarantees it will surface any range keys with bounds overlapping the
1544 : // keyspace up to limit.
1545 1 : func (i *Iterator) SeekLTWithLimit(key []byte, limit []byte) IterValidityState {
1546 1 : if i.rangeKey != nil {
1547 1 : // NB: Check Valid() before clearing requiresReposition.
1548 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1549 1 : // If we have a range key but did not expose it at the previous iterator
1550 1 : // position (because the iterator was not at a valid position), updated
1551 1 : // must be true. This ensures that after an iterator op sequence like:
1552 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1553 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1554 1 : // - SeekLTWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1555 1 : // the iterator returns RangeKeyChanged()=true.
1556 1 : //
1557 1 : // The remainder of this function will only update i.rangeKey.updated if
1558 1 : // the iterator moves into a new range key, or out of the current range
1559 1 : // key.
1560 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1561 1 : }
1562 1 : lastPositioningOp := i.lastPositioningOp
1563 1 : // Set it to unknown, since this operation may not succeed, in which case
1564 1 : // the SeekLT following this should not make any assumption about iterator
1565 1 : // position.
1566 1 : i.lastPositioningOp = unknownLastPositionOp
1567 1 : i.batchJustRefreshed = false
1568 1 : i.requiresReposition = false
1569 1 : i.err = nil // clear cached iteration error
1570 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1571 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1572 1 : key = upperBound
1573 1 : } else if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1574 1 : key = lowerBound
1575 1 : }
1576 1 : i.hasPrefix = false
1577 1 : seekInternalIter := true
1578 1 : // The following noop optimization only applies when i.batch == nil, since
1579 1 : // an iterator over a batch is iterating over mutable data, that may have
1580 1 : // changed since the last seek.
1581 1 : if lastPositioningOp == seekLTLastPositioningOp && i.batch == nil {
1582 1 : cmp := i.cmp(key, i.prefixOrFullSeekKey)
1583 1 : // If this seek is to the same or earlier key, and the iterator is
1584 1 : // already positioned there, this is a noop. This can be helpful for
1585 1 : // sparse key spaces that have many deleted keys, where one can avoid
1586 1 : // the overhead of iterating past them again and again.
1587 1 : if cmp <= 0 {
1588 1 : // NB: when pos != iterPosCurReversePaused, the invariant
1589 1 : // documented earlier implies that iterValidityState !=
1590 1 : // IterAtLimit.
1591 1 : if i.iterValidityState == IterExhausted ||
1592 1 : (i.iterValidityState == IterValid && i.cmp(i.key, key) < 0 &&
1593 1 : (limit == nil || i.cmp(limit, i.key) <= 0)) {
1594 1 : if !testingDisableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1595 1 : i.lastPositioningOp = seekLTLastPositioningOp
1596 1 : return i.iterValidityState
1597 1 : }
1598 : }
1599 1 : if i.pos == iterPosCurReversePaused && i.cmp(i.iterKV.K.UserKey, key) < 0 {
1600 1 : // Have some work to do, but don't need to seek, and we can
1601 1 : // start doing findPrevEntry from i.iterKey.
1602 1 : seekInternalIter = false
1603 1 : }
1604 : }
1605 : }
1606 1 : if seekInternalIter {
1607 1 : i.iterKV = i.iter.SeekLT(key, base.SeekLTFlagsNone)
1608 1 : i.stats.ReverseSeekCount[InternalIterCall]++
1609 1 : if err := i.iter.Error(); err != nil {
1610 0 : i.err = err
1611 0 : i.iterValidityState = IterExhausted
1612 0 : return i.iterValidityState
1613 0 : }
1614 : }
1615 1 : i.findPrevEntry(limit)
1616 1 : i.maybeSampleRead()
1617 1 : if i.Error() == nil && i.batch == nil {
1618 1 : // Prepare state for a future noop optimization.
1619 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1620 1 : i.lastPositioningOp = seekLTLastPositioningOp
1621 1 : }
1622 1 : return i.iterValidityState
1623 : }
1624 :
1625 : // First moves the iterator the first key/value pair. Returns true if the
1626 : // iterator is pointing at a valid entry and false otherwise.
1627 1 : func (i *Iterator) First() bool {
1628 1 : if i.rangeKey != nil {
1629 1 : // NB: Check Valid() before clearing requiresReposition.
1630 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1631 1 : // If we have a range key but did not expose it at the previous iterator
1632 1 : // position (because the iterator was not at a valid position), updated
1633 1 : // must be true. This ensures that after an iterator op sequence like:
1634 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1635 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1636 1 : // - First(...) → (IterValid, RangeBounds() = [a,b))
1637 1 : // the iterator returns RangeKeyChanged()=true.
1638 1 : //
1639 1 : // The remainder of this function will only update i.rangeKey.updated if
1640 1 : // the iterator moves into a new range key, or out of the current range
1641 1 : // key.
1642 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1643 1 : }
1644 1 : i.err = nil // clear cached iteration error
1645 1 : i.hasPrefix = false
1646 1 : i.batchJustRefreshed = false
1647 1 : i.lastPositioningOp = unknownLastPositionOp
1648 1 : i.requiresReposition = false
1649 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1650 1 :
1651 1 : i.err = i.iterFirstWithinBounds()
1652 1 : if i.err != nil {
1653 0 : i.iterValidityState = IterExhausted
1654 0 : return false
1655 0 : }
1656 1 : i.findNextEntry(nil)
1657 1 : i.maybeSampleRead()
1658 1 : return i.iterValidityState == IterValid
1659 : }
1660 :
1661 : // Last moves the iterator the last key/value pair. Returns true if the
1662 : // iterator is pointing at a valid entry and false otherwise.
1663 1 : func (i *Iterator) Last() bool {
1664 1 : if i.rangeKey != nil {
1665 1 : // NB: Check Valid() before clearing requiresReposition.
1666 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1667 1 : // If we have a range key but did not expose it at the previous iterator
1668 1 : // position (because the iterator was not at a valid position), updated
1669 1 : // must be true. This ensures that after an iterator op sequence like:
1670 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1671 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1672 1 : // - Last(...) → (IterValid, RangeBounds() = [a,b))
1673 1 : // the iterator returns RangeKeyChanged()=true.
1674 1 : //
1675 1 : // The remainder of this function will only update i.rangeKey.updated if
1676 1 : // the iterator moves into a new range key, or out of the current range
1677 1 : // key.
1678 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1679 1 : }
1680 1 : i.err = nil // clear cached iteration error
1681 1 : i.hasPrefix = false
1682 1 : i.batchJustRefreshed = false
1683 1 : i.lastPositioningOp = unknownLastPositionOp
1684 1 : i.requiresReposition = false
1685 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1686 1 :
1687 1 : if i.err = i.iterLastWithinBounds(); i.err != nil {
1688 0 : i.iterValidityState = IterExhausted
1689 0 : return false
1690 0 : }
1691 1 : i.findPrevEntry(nil)
1692 1 : i.maybeSampleRead()
1693 1 : return i.iterValidityState == IterValid
1694 : }
1695 :
1696 : // Next moves the iterator to the next key/value pair. Returns true if the
1697 : // iterator is pointing at a valid entry and false otherwise.
1698 1 : func (i *Iterator) Next() bool {
1699 1 : return i.nextWithLimit(nil) == IterValid
1700 1 : }
1701 :
1702 : // NextWithLimit moves the iterator to the next key/value pair.
1703 : //
1704 : // If limit is provided, it serves as a best-effort exclusive limit. If the next
1705 : // key is greater than or equal to limit, the Iterator may pause and return
1706 : // IterAtLimit. Because limits are best-effort, NextWithLimit may return a key
1707 : // beyond limit.
1708 : //
1709 : // If the Iterator is configured to iterate over range keys, NextWithLimit
1710 : // guarantees it will surface any range keys with bounds overlapping the
1711 : // keyspace up to limit.
1712 1 : func (i *Iterator) NextWithLimit(limit []byte) IterValidityState {
1713 1 : return i.nextWithLimit(limit)
1714 1 : }
1715 :
1716 : // NextPrefix moves the iterator to the next key/value pair with a key
1717 : // containing a different prefix than the current key. Prefixes are determined
1718 : // by Comparer.Split. Exhausts the iterator if invoked while in prefix-iteration
1719 : // mode.
1720 : //
1721 : // It is not permitted to invoke NextPrefix while at a IterAtLimit position.
1722 : // When called in this condition, NextPrefix has non-deterministic behavior.
1723 : //
1724 : // It is not permitted to invoke NextPrefix when the Iterator has an
1725 : // upper-bound that is a versioned MVCC key (see the comment for
1726 : // Comparer.Split). It returns an error in this case.
1727 1 : func (i *Iterator) NextPrefix() bool {
1728 1 : if i.nextPrefixNotPermittedByUpperBound {
1729 1 : i.lastPositioningOp = unknownLastPositionOp
1730 1 : i.requiresReposition = false
1731 1 : i.err = errors.Errorf("NextPrefix not permitted with upper bound %s",
1732 1 : i.comparer.FormatKey(i.opts.UpperBound))
1733 1 : i.iterValidityState = IterExhausted
1734 1 : return false
1735 1 : }
1736 1 : if i.hasPrefix {
1737 1 : i.iterValidityState = IterExhausted
1738 1 : return false
1739 1 : }
1740 1 : if i.Error() != nil {
1741 1 : return false
1742 1 : }
1743 1 : return i.nextPrefix() == IterValid
1744 : }
1745 :
1746 1 : func (i *Iterator) nextPrefix() IterValidityState {
1747 1 : if i.rangeKey != nil {
1748 1 : // NB: Check Valid() before clearing requiresReposition.
1749 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1750 1 : // If we have a range key but did not expose it at the previous iterator
1751 1 : // position (because the iterator was not at a valid position), updated
1752 1 : // must be true. This ensures that after an iterator op sequence like:
1753 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1754 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1755 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1756 1 : // the iterator returns RangeKeyChanged()=true.
1757 1 : //
1758 1 : // The remainder of this function will only update i.rangeKey.updated if
1759 1 : // the iterator moves into a new range key, or out of the current range
1760 1 : // key.
1761 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1762 1 : }
1763 :
1764 : // Although NextPrefix documents that behavior at IterAtLimit is undefined,
1765 : // this function handles these cases as a simple prefix-agnostic Next. This
1766 : // is done for deterministic behavior in the metamorphic tests.
1767 : //
1768 : // TODO(jackson): If the metamorphic test operation generator is adjusted to
1769 : // make generation of some operations conditional on the previous
1770 : // operations, then we can remove this behavior and explicitly error.
1771 :
1772 1 : i.lastPositioningOp = unknownLastPositionOp
1773 1 : i.requiresReposition = false
1774 1 : switch i.pos {
1775 1 : case iterPosCurForward:
1776 1 : // Positioned on the current key. Advance to the next prefix.
1777 1 : i.internalNextPrefix(i.comparer.Split(i.key))
1778 1 : case iterPosCurForwardPaused:
1779 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1780 : // up above. The iterator is already positioned at the next key.
1781 1 : case iterPosCurReverse:
1782 1 : // Switching directions.
1783 1 : // Unless the iterator was exhausted, reverse iteration needs to
1784 1 : // position the iterator at iterPosPrev.
1785 1 : if i.iterKV != nil {
1786 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1787 0 : i.iterValidityState = IterExhausted
1788 0 : return i.iterValidityState
1789 0 : }
1790 : // The Iterator is exhausted and i.iter is positioned before the first
1791 : // key. Reposition to point to the first internal key.
1792 1 : if i.err = i.iterFirstWithinBounds(); i.err != nil {
1793 0 : i.iterValidityState = IterExhausted
1794 0 : return i.iterValidityState
1795 0 : }
1796 1 : case iterPosCurReversePaused:
1797 1 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1798 1 : // up above.
1799 1 : //
1800 1 : // Switching directions; The iterator must not be exhausted since it
1801 1 : // paused.
1802 1 : if i.iterKV == nil {
1803 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1804 0 : i.iterValidityState = IterExhausted
1805 0 : return i.iterValidityState
1806 0 : }
1807 1 : i.nextUserKey()
1808 1 : case iterPosPrev:
1809 1 : // The underlying iterator is pointed to the previous key (this can
1810 1 : // only happen when switching iteration directions).
1811 1 : if i.iterKV == nil {
1812 1 : // We're positioned before the first key. Need to reposition to point to
1813 1 : // the first key.
1814 1 : i.err = i.iterFirstWithinBounds()
1815 1 : if i.iterKV == nil {
1816 0 : i.iterValidityState = IterExhausted
1817 0 : return i.iterValidityState
1818 0 : }
1819 1 : if invariants.Enabled && !i.equal(i.iterKV.K.UserKey, i.key) {
1820 0 : i.opts.getLogger().Fatalf("pebble: invariant violation: First internal iterator from iterPosPrev landed on %q, not %q",
1821 0 : i.iterKV.K.UserKey, i.key)
1822 0 : }
1823 1 : } else {
1824 1 : // Move the internal iterator back onto the user key stored in
1825 1 : // i.key. iterPosPrev guarantees that it's positioned at the last
1826 1 : // key with the user key less than i.key, so we're guaranteed to
1827 1 : // land on the correct key with a single Next.
1828 1 : i.iterKV = i.iter.Next()
1829 1 : if i.iterKV == nil {
1830 0 : // This should only be possible if i.iter.Next() encountered an
1831 0 : // error.
1832 0 : if i.iter.Error() == nil {
1833 0 : i.opts.getLogger().Fatalf("pebble: invariant violation: Nexting internal iterator from iterPosPrev found nothing")
1834 0 : }
1835 : // NB: Iterator.Error() will return i.iter.Error().
1836 0 : i.iterValidityState = IterExhausted
1837 0 : return i.iterValidityState
1838 : }
1839 1 : if invariants.Enabled && !i.equal(i.iterKV.K.UserKey, i.key) {
1840 0 : i.opts.getLogger().Fatalf("pebble: invariant violation: Nexting internal iterator from iterPosPrev landed on %q, not %q",
1841 0 : i.iterKV.K.UserKey, i.key)
1842 0 : }
1843 : }
1844 : // The internal iterator is now positioned at i.key. Advance to the next
1845 : // prefix.
1846 1 : i.internalNextPrefix(i.comparer.Split(i.key))
1847 1 : case iterPosNext:
1848 1 : // Already positioned on the next key. Only call nextPrefixKey if the
1849 1 : // next key shares the same prefix.
1850 1 : if i.iterKV != nil {
1851 1 : currKeyPrefixLen := i.comparer.Split(i.key)
1852 1 : if bytes.Equal(i.comparer.Split.Prefix(i.iterKV.K.UserKey), i.key[:currKeyPrefixLen]) {
1853 1 : i.internalNextPrefix(currKeyPrefixLen)
1854 1 : }
1855 : }
1856 : }
1857 :
1858 1 : i.stats.ForwardStepCount[InterfaceCall]++
1859 1 : i.findNextEntry(nil /* limit */)
1860 1 : i.maybeSampleRead()
1861 1 : return i.iterValidityState
1862 : }
1863 :
1864 1 : func (i *Iterator) internalNextPrefix(currKeyPrefixLen int) {
1865 1 : if i.iterKV == nil {
1866 1 : return
1867 1 : }
1868 : // The Next "fast-path" is not really a fast-path when there is more than
1869 : // one version. However, even with TableFormatPebblev3, there is a small
1870 : // slowdown (~10%) for one version if we remove it and only call NextPrefix.
1871 : // When there are two versions, only calling NextPrefix is ~30% faster.
1872 1 : i.stats.ForwardStepCount[InternalIterCall]++
1873 1 : if i.iterKV = i.iter.Next(); i.iterKV == nil {
1874 1 : return
1875 1 : }
1876 1 : if !bytes.Equal(i.comparer.Split.Prefix(i.iterKV.K.UserKey), i.key[:currKeyPrefixLen]) {
1877 1 : return
1878 1 : }
1879 1 : i.stats.ForwardStepCount[InternalIterCall]++
1880 1 : i.prefixOrFullSeekKey = i.comparer.ImmediateSuccessor(i.prefixOrFullSeekKey[:0], i.key[:currKeyPrefixLen])
1881 1 : if i.iterKV.K.IsExclusiveSentinel() {
1882 0 : panic(errors.AssertionFailedf("pebble: unexpected exclusive sentinel key: %q", i.iterKV.K))
1883 : }
1884 :
1885 1 : i.iterKV = i.iter.NextPrefix(i.prefixOrFullSeekKey)
1886 1 : if invariants.Enabled && i.iterKV != nil {
1887 1 : if p := i.comparer.Split.Prefix(i.iterKV.K.UserKey); i.cmp(p, i.prefixOrFullSeekKey) < 0 {
1888 0 : panic(errors.AssertionFailedf("pebble: iter.NextPrefix did not advance beyond the current prefix: now at %q; expected to be geq %q",
1889 0 : i.iterKV.K, i.prefixOrFullSeekKey))
1890 : }
1891 : }
1892 : }
1893 :
1894 1 : func (i *Iterator) nextWithLimit(limit []byte) IterValidityState {
1895 1 : i.stats.ForwardStepCount[InterfaceCall]++
1896 1 : if i.hasPrefix {
1897 1 : if limit != nil {
1898 1 : i.err = errors.New("cannot use limit with prefix iteration")
1899 1 : i.iterValidityState = IterExhausted
1900 1 : return i.iterValidityState
1901 1 : } else if i.iterValidityState == IterExhausted {
1902 1 : // No-op, already exhausted. We avoid executing the Next because it
1903 1 : // can break invariants: Specifically, a file that fails the bloom
1904 1 : // filter test may result in its level being removed from the
1905 1 : // merging iterator. The level's removal can cause a lazy combined
1906 1 : // iterator to miss range keys and trigger a switch to combined
1907 1 : // iteration at a larger key, breaking keyspan invariants.
1908 1 : return i.iterValidityState
1909 1 : }
1910 : }
1911 1 : if i.err != nil {
1912 1 : return i.iterValidityState
1913 1 : }
1914 1 : if i.rangeKey != nil {
1915 1 : // NB: Check Valid() before clearing requiresReposition.
1916 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1917 1 : // If we have a range key but did not expose it at the previous iterator
1918 1 : // position (because the iterator was not at a valid position), updated
1919 1 : // must be true. This ensures that after an iterator op sequence like:
1920 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1921 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1922 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1923 1 : // the iterator returns RangeKeyChanged()=true.
1924 1 : //
1925 1 : // The remainder of this function will only update i.rangeKey.updated if
1926 1 : // the iterator moves into a new range key, or out of the current range
1927 1 : // key.
1928 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1929 1 : }
1930 1 : i.lastPositioningOp = unknownLastPositionOp
1931 1 : i.requiresReposition = false
1932 1 : switch i.pos {
1933 1 : case iterPosCurForward:
1934 1 : i.nextUserKey()
1935 1 : case iterPosCurForwardPaused:
1936 : // Already at the right place.
1937 1 : case iterPosCurReverse:
1938 1 : // Switching directions.
1939 1 : // Unless the iterator was exhausted, reverse iteration needs to
1940 1 : // position the iterator at iterPosPrev.
1941 1 : if i.iterKV != nil {
1942 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1943 0 : i.iterValidityState = IterExhausted
1944 0 : return i.iterValidityState
1945 0 : }
1946 : // We're positioned before the first key. Need to reposition to point to
1947 : // the first key.
1948 1 : if i.err = i.iterFirstWithinBounds(); i.err != nil {
1949 0 : i.iterValidityState = IterExhausted
1950 0 : return i.iterValidityState
1951 0 : }
1952 1 : case iterPosCurReversePaused:
1953 1 : // Switching directions.
1954 1 : // The iterator must not be exhausted since it paused.
1955 1 : if i.iterKV == nil {
1956 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1957 0 : i.iterValidityState = IterExhausted
1958 0 : return i.iterValidityState
1959 0 : }
1960 1 : i.nextUserKey()
1961 1 : case iterPosPrev:
1962 1 : // The underlying iterator is pointed to the previous key (this can
1963 1 : // only happen when switching iteration directions). We set
1964 1 : // i.iterValidityState to IterExhausted here to force the calls to
1965 1 : // nextUserKey to save the current key i.iter is pointing at in order
1966 1 : // to determine when the next user-key is reached.
1967 1 : i.iterValidityState = IterExhausted
1968 1 : if i.iterKV == nil {
1969 1 : // We're positioned before the first key. Need to reposition to point to
1970 1 : // the first key.
1971 1 : i.err = i.iterFirstWithinBounds()
1972 1 : } else {
1973 1 : i.nextUserKey()
1974 1 : }
1975 1 : if i.err != nil {
1976 0 : i.iterValidityState = IterExhausted
1977 0 : return i.iterValidityState
1978 0 : }
1979 1 : i.nextUserKey()
1980 1 : case iterPosNext:
1981 : // Already at the right place.
1982 : }
1983 1 : i.findNextEntry(limit)
1984 1 : i.maybeSampleRead()
1985 1 : return i.iterValidityState
1986 : }
1987 :
1988 : // Prev moves the iterator to the previous key/value pair. Returns true if the
1989 : // iterator is pointing at a valid entry and false otherwise.
1990 1 : func (i *Iterator) Prev() bool {
1991 1 : return i.PrevWithLimit(nil) == IterValid
1992 1 : }
1993 :
1994 : // PrevWithLimit moves the iterator to the previous key/value pair.
1995 : //
1996 : // If limit is provided, it serves as a best-effort inclusive limit. If the
1997 : // previous key is less than limit, the Iterator may pause and return
1998 : // IterAtLimit. Because limits are best-effort, PrevWithLimit may return a key
1999 : // beyond limit.
2000 : //
2001 : // If the Iterator is configured to iterate over range keys, PrevWithLimit
2002 : // guarantees it will surface any range keys with bounds overlapping the
2003 : // keyspace up to limit.
2004 1 : func (i *Iterator) PrevWithLimit(limit []byte) IterValidityState {
2005 1 : i.stats.ReverseStepCount[InterfaceCall]++
2006 1 : if i.err != nil {
2007 1 : return i.iterValidityState
2008 1 : }
2009 1 : if i.rangeKey != nil {
2010 1 : // NB: Check Valid() before clearing requiresReposition.
2011 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
2012 1 : // If we have a range key but did not expose it at the previous iterator
2013 1 : // position (because the iterator was not at a valid position), updated
2014 1 : // must be true. This ensures that after an iterator op sequence like:
2015 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
2016 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
2017 1 : // - PrevWithLimit(...) → (IterValid, RangeBounds() = [a,b))
2018 1 : // the iterator returns RangeKeyChanged()=true.
2019 1 : //
2020 1 : // The remainder of this function will only update i.rangeKey.updated if
2021 1 : // the iterator moves into a new range key, or out of the current range
2022 1 : // key.
2023 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
2024 1 : }
2025 1 : i.lastPositioningOp = unknownLastPositionOp
2026 1 : i.requiresReposition = false
2027 1 : if i.hasPrefix {
2028 1 : i.err = errReversePrefixIteration
2029 1 : i.iterValidityState = IterExhausted
2030 1 : return i.iterValidityState
2031 1 : }
2032 1 : switch i.pos {
2033 1 : case iterPosCurForward:
2034 : // Switching directions, and will handle this below.
2035 1 : case iterPosCurForwardPaused:
2036 : // Switching directions, and will handle this below.
2037 1 : case iterPosCurReverse:
2038 1 : i.prevUserKey()
2039 1 : case iterPosCurReversePaused:
2040 : // Already at the right place.
2041 1 : case iterPosNext:
2042 : // The underlying iterator is pointed to the next key (this can only happen
2043 : // when switching iteration directions). We will handle this below.
2044 1 : case iterPosPrev:
2045 : // Already at the right place.
2046 : }
2047 1 : if i.pos == iterPosCurForward || i.pos == iterPosNext || i.pos == iterPosCurForwardPaused {
2048 1 : // Switching direction.
2049 1 : stepAgain := i.pos == iterPosNext
2050 1 :
2051 1 : // Synthetic range key markers are a special case. Consider SeekGE(b)
2052 1 : // which finds a range key [a, c). To ensure the user observes the range
2053 1 : // key, the Iterator pauses at Key() = b. The iterator must advance the
2054 1 : // internal iterator to see if there's also a coincident point key at
2055 1 : // 'b', leaving the iterator at iterPosNext if there's not.
2056 1 : //
2057 1 : // This is a problem: Synthetic range key markers are only interleaved
2058 1 : // during the original seek. A subsequent Prev() of i.iter will not move
2059 1 : // back onto the synthetic range key marker. In this case where the
2060 1 : // previous iterator position was a synthetic range key start boundary,
2061 1 : // we must not step a second time.
2062 1 : if i.isEphemeralPosition() {
2063 1 : stepAgain = false
2064 1 : }
2065 :
2066 : // We set i.iterValidityState to IterExhausted here to force the calls
2067 : // to prevUserKey to save the current key i.iter is pointing at in
2068 : // order to determine when the prev user-key is reached.
2069 1 : i.iterValidityState = IterExhausted
2070 1 : if i.iterKV == nil {
2071 1 : // We're positioned after the last key. Need to reposition to point to
2072 1 : // the last key.
2073 1 : i.err = i.iterLastWithinBounds()
2074 1 : } else {
2075 1 : i.prevUserKey()
2076 1 : }
2077 1 : if i.err != nil {
2078 0 : return i.iterValidityState
2079 0 : }
2080 1 : if stepAgain {
2081 1 : i.prevUserKey()
2082 1 : if i.err != nil {
2083 0 : return i.iterValidityState
2084 0 : }
2085 : }
2086 : }
2087 1 : i.findPrevEntry(limit)
2088 1 : i.maybeSampleRead()
2089 1 : return i.iterValidityState
2090 : }
2091 :
2092 : // iterFirstWithinBounds moves the internal iterator to the first key,
2093 : // respecting bounds.
2094 1 : func (i *Iterator) iterFirstWithinBounds() error {
2095 1 : i.stats.ForwardSeekCount[InternalIterCall]++
2096 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil {
2097 1 : i.iterKV = i.iter.SeekGE(lowerBound, base.SeekGEFlagsNone)
2098 1 : } else {
2099 1 : i.iterKV = i.iter.First()
2100 1 : }
2101 1 : if i.iterKV == nil {
2102 1 : return i.iter.Error()
2103 1 : }
2104 1 : return nil
2105 : }
2106 :
2107 : // iterLastWithinBounds moves the internal iterator to the last key, respecting
2108 : // bounds.
2109 1 : func (i *Iterator) iterLastWithinBounds() error {
2110 1 : i.stats.ReverseSeekCount[InternalIterCall]++
2111 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil {
2112 1 : i.iterKV = i.iter.SeekLT(upperBound, base.SeekLTFlagsNone)
2113 1 : } else {
2114 1 : i.iterKV = i.iter.Last()
2115 1 : }
2116 1 : if i.iterKV == nil {
2117 1 : return i.iter.Error()
2118 1 : }
2119 1 : return nil
2120 : }
2121 :
2122 : // RangeKeyData describes a range key's data, set through RangeKeySet. The key
2123 : // boundaries of the range key is provided by Iterator.RangeBounds.
2124 : type RangeKeyData struct {
2125 : Suffix []byte
2126 : Value []byte
2127 : }
2128 :
2129 : // rangeKeyWithinLimit is called during limited reverse iteration when
2130 : // positioned over a key beyond the limit. If there exists a range key that lies
2131 : // within the limit, the iterator must not pause in order to ensure the user has
2132 : // an opportunity to observe the range key within limit.
2133 : //
2134 : // It would be valid to ignore the limit whenever there's a range key covering
2135 : // the key, but that would introduce nondeterminism. To preserve determinism for
2136 : // testing, the iterator ignores the limit only if the covering range key does
2137 : // cover the keyspace within the limit.
2138 : //
2139 : // This awkwardness exists because range keys are interleaved at their inclusive
2140 : // start positions. Note that limit is inclusive.
2141 1 : func (i *Iterator) rangeKeyWithinLimit(limit []byte) bool {
2142 1 : if i.rangeKey == nil || !i.opts.rangeKeys() {
2143 1 : return false
2144 1 : }
2145 1 : s := i.rangeKey.iiter.Span()
2146 1 : // If the range key ends beyond the limit, then the range key does not cover
2147 1 : // any portion of the keyspace within the limit and it is safe to pause.
2148 1 : return s != nil && i.cmp(s.End, limit) > 0
2149 : }
2150 :
2151 : // saveRangeKey saves the current range key to the underlying iterator's current
2152 : // range key state. If the range key has not changed, saveRangeKey is a no-op.
2153 : // If there is a new range key, saveRangeKey copies all of the key, value and
2154 : // suffixes into Iterator-managed buffers.
2155 1 : func (i *Iterator) saveRangeKey() {
2156 1 : if i.rangeKey == nil || i.opts.KeyTypes == IterKeyTypePointsOnly {
2157 1 : return
2158 1 : }
2159 :
2160 1 : s := i.rangeKey.iiter.Span()
2161 1 : if s == nil {
2162 1 : i.rangeKey.hasRangeKey = false
2163 1 : i.rangeKey.updated = i.rangeKey.prevPosHadRangeKey
2164 1 : return
2165 1 : } else if !i.rangeKey.stale {
2166 1 : // The range key `s` is identical to the one currently saved. No-op.
2167 1 : return
2168 1 : }
2169 :
2170 1 : if s.KeysOrder != keyspan.BySuffixAsc {
2171 0 : panic("pebble: range key span's keys unexpectedly not in ascending suffix order")
2172 : }
2173 :
2174 : // Although `i.rangeKey.stale` is true, the span s may still be identical
2175 : // to the currently saved span. This is possible when seeking the iterator,
2176 : // which may land back on the same range key. If we previously had a range
2177 : // key and the new one has an identical start key, then it must be the same
2178 : // range key and we can avoid copying and keep `i.rangeKey.updated=false`.
2179 : //
2180 : // TODO(jackson): These key comparisons could be avoidable during relative
2181 : // positioning operations continuing in the same direction, because these
2182 : // ops will never encounter the previous position's range key while
2183 : // stale=true. However, threading whether the current op is a seek or step
2184 : // maybe isn't worth it. This key comparison is only necessary once when we
2185 : // step onto a new range key, which should be relatively rare.
2186 1 : if i.rangeKey.prevPosHadRangeKey && i.equal(i.rangeKey.start, s.Start) &&
2187 1 : i.equal(i.rangeKey.end, s.End) {
2188 1 : i.rangeKey.updated = false
2189 1 : i.rangeKey.stale = false
2190 1 : i.rangeKey.hasRangeKey = true
2191 1 : return
2192 1 : }
2193 1 : i.stats.RangeKeyStats.Count += len(s.Keys)
2194 1 : i.rangeKey.buf.Reset()
2195 1 : i.rangeKey.hasRangeKey = true
2196 1 : i.rangeKey.updated = true
2197 1 : i.rangeKey.stale = false
2198 1 : i.rangeKey.buf, i.rangeKey.start = i.rangeKey.buf.Copy(s.Start)
2199 1 : i.rangeKey.buf, i.rangeKey.end = i.rangeKey.buf.Copy(s.End)
2200 1 : i.rangeKey.keys = i.rangeKey.keys[:0]
2201 1 : for j := 0; j < len(s.Keys); j++ {
2202 1 : if invariants.Enabled {
2203 1 : if s.Keys[j].Kind() != base.InternalKeyKindRangeKeySet {
2204 0 : panic("pebble: user iteration encountered non-RangeKeySet key kind")
2205 1 : } else if j > 0 && i.comparer.CompareRangeSuffixes(s.Keys[j].Suffix, s.Keys[j-1].Suffix) < 0 {
2206 0 : panic("pebble: user iteration encountered range keys not in suffix order")
2207 : }
2208 : }
2209 1 : var rkd RangeKeyData
2210 1 : i.rangeKey.buf, rkd.Suffix = i.rangeKey.buf.Copy(s.Keys[j].Suffix)
2211 1 : i.rangeKey.buf, rkd.Value = i.rangeKey.buf.Copy(s.Keys[j].Value)
2212 1 : i.rangeKey.keys = append(i.rangeKey.keys, rkd)
2213 : }
2214 : }
2215 :
2216 : // RangeKeyChanged indicates whether the most recent iterator positioning
2217 : // operation resulted in the iterator stepping into or out of a new range key.
2218 : // If true, previously returned range key bounds and data has been invalidated.
2219 : // If false, previously obtained range key bounds, suffix and value slices are
2220 : // still valid and may continue to be read.
2221 : //
2222 : // Invalid iterator positions are considered to not hold range keys, meaning
2223 : // that if an iterator steps from an IterExhausted or IterAtLimit position onto
2224 : // a position with a range key, RangeKeyChanged will yield true.
2225 1 : func (i *Iterator) RangeKeyChanged() bool {
2226 1 : return i.iterValidityState == IterValid && i.rangeKey != nil && i.rangeKey.updated
2227 1 : }
2228 :
2229 : // HasPointAndRange indicates whether there exists a point key, a range key or
2230 : // both at the current iterator position.
2231 1 : func (i *Iterator) HasPointAndRange() (hasPoint, hasRange bool) {
2232 1 : if i.iterValidityState != IterValid || i.requiresReposition {
2233 0 : return false, false
2234 0 : }
2235 1 : if i.opts.KeyTypes == IterKeyTypePointsOnly {
2236 1 : return true, false
2237 1 : }
2238 1 : return i.rangeKey == nil || !i.rangeKey.rangeKeyOnly, i.rangeKey != nil && i.rangeKey.hasRangeKey
2239 : }
2240 :
2241 : // RangeBounds returns the start (inclusive) and end (exclusive) bounds of the
2242 : // range key covering the current iterator position. RangeBounds returns nil
2243 : // bounds if there is no range key covering the current iterator position, or
2244 : // the iterator is not configured to surface range keys.
2245 : //
2246 : // If valid, the returned start bound is less than or equal to Key() and the
2247 : // returned end bound is greater than Key().
2248 1 : func (i *Iterator) RangeBounds() (start, end []byte) {
2249 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2250 0 : return nil, nil
2251 0 : }
2252 1 : return i.rangeKey.start, i.rangeKey.end
2253 : }
2254 :
2255 : // Key returns the key of the current key/value pair, or nil if done. The
2256 : // caller should not modify the contents of the returned slice, and its
2257 : // contents may change on the next call to Next.
2258 : //
2259 : // If positioned at an iterator position that only holds a range key, Key()
2260 : // always returns the start bound of the range key. Otherwise, it returns the
2261 : // point key's key.
2262 1 : func (i *Iterator) Key() []byte {
2263 1 : return i.key
2264 1 : }
2265 :
2266 : // Value returns the value of the current key/value pair, or nil if done. The
2267 : // caller should not modify the contents of the returned slice, and its
2268 : // contents may change on the next call to Next.
2269 : //
2270 : // Only valid if HasPointAndRange() returns true for hasPoint.
2271 : // Deprecated: use ValueAndErr instead.
2272 1 : func (i *Iterator) Value() []byte {
2273 1 : val, _ := i.ValueAndErr()
2274 1 : return val
2275 1 : }
2276 :
2277 : // ValueAndErr returns the value, and any error encountered in extracting the value.
2278 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2279 : //
2280 : // The caller should not modify the contents of the returned slice, and its
2281 : // contents may change on the next call to Next.
2282 1 : func (i *Iterator) ValueAndErr() ([]byte, error) {
2283 1 : val, callerOwned, err := i.value.Value(i.lazyValueBuf)
2284 1 : if err != nil {
2285 0 : i.err = err
2286 0 : i.iterValidityState = IterExhausted
2287 0 : }
2288 1 : if callerOwned {
2289 1 : i.lazyValueBuf = val[:0]
2290 1 : }
2291 1 : return val, err
2292 : }
2293 :
2294 : // LazyValue returns the LazyValue. Only for advanced use cases.
2295 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2296 0 : func (i *Iterator) LazyValue() LazyValue {
2297 0 : return i.value.LazyValue()
2298 0 : }
2299 :
2300 : // RangeKeys returns the range key values and their suffixes covering the
2301 : // current iterator position. The range bounds may be retrieved separately
2302 : // through Iterator.RangeBounds().
2303 1 : func (i *Iterator) RangeKeys() []RangeKeyData {
2304 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2305 0 : return nil
2306 0 : }
2307 1 : return i.rangeKey.keys
2308 : }
2309 :
2310 : // Valid returns true if the iterator is positioned at a valid key/value pair
2311 : // and false otherwise.
2312 1 : func (i *Iterator) Valid() bool {
2313 1 : valid := i.iterValidityState == IterValid && !i.requiresReposition
2314 1 : if invariants.Enabled {
2315 1 : if err := i.Error(); valid && err != nil {
2316 0 : panic(errors.AssertionFailedf("pebble: iterator is valid with non-nil Error: %+v", err))
2317 : }
2318 : }
2319 1 : return valid
2320 : }
2321 :
2322 : // Error returns any accumulated error.
2323 1 : func (i *Iterator) Error() error {
2324 1 : if i.err != nil {
2325 1 : return i.err
2326 1 : }
2327 1 : if i.iter != nil {
2328 1 : return i.iter.Error()
2329 1 : }
2330 0 : return nil
2331 : }
2332 :
2333 : const maxKeyBufCacheSize = 4 << 10 // 4 KB
2334 :
2335 : // Close closes the iterator and returns any accumulated error. Exhausting
2336 : // all the key/value pairs in a table is not considered to be an error.
2337 : // It is not valid to call any method, including Close, after the iterator
2338 : // has been closed.
2339 1 : func (i *Iterator) Close() error {
2340 1 : // Close the child iterator before releasing the readState because when the
2341 1 : // readState is released sstables referenced by the readState may be deleted
2342 1 : // which will fail on Windows if the sstables are still open by the child
2343 1 : // iterator.
2344 1 : if i.iter != nil {
2345 1 : i.err = firstError(i.err, i.iter.Close())
2346 1 :
2347 1 : // Closing i.iter did not necessarily close the point and range key
2348 1 : // iterators. Calls to SetOptions may have 'disconnected' either one
2349 1 : // from i.iter if iteration key types were changed. Both point and range
2350 1 : // key iterators are preserved in case the iterator needs to switch key
2351 1 : // types again. We explicitly close both of these iterators here.
2352 1 : //
2353 1 : // NB: If the iterators were still connected to i.iter, they may be
2354 1 : // closed, but calling Close on a closed internal iterator or fragment
2355 1 : // iterator is allowed.
2356 1 : if i.pointIter != nil {
2357 1 : i.err = firstError(i.err, i.pointIter.Close())
2358 1 : }
2359 1 : if i.rangeKey != nil && i.rangeKey.rangeKeyIter != nil {
2360 1 : i.rangeKey.rangeKeyIter.Close()
2361 1 : }
2362 1 : i.err = firstError(i.err, i.blobValueFetcher.Close())
2363 : }
2364 1 : err := i.err
2365 1 :
2366 1 : if i.readState != nil {
2367 1 : if i.readSampling.pendingCompactions.size > 0 {
2368 0 : // Copy pending read compactions using db.mu.Lock()
2369 0 : i.readState.db.mu.Lock()
2370 0 : i.readState.db.mu.compact.readCompactions.combine(&i.readSampling.pendingCompactions, i.cmp)
2371 0 : reschedule := i.readState.db.mu.compact.rescheduleReadCompaction
2372 0 : i.readState.db.mu.compact.rescheduleReadCompaction = false
2373 0 : concurrentCompactions := i.readState.db.mu.compact.compactingCount
2374 0 : i.readState.db.mu.Unlock()
2375 0 :
2376 0 : if reschedule && concurrentCompactions == 0 {
2377 0 : // In a read heavy workload, flushes may not happen frequently enough to
2378 0 : // schedule compactions.
2379 0 : i.readState.db.compactionSchedulers.Add(1)
2380 0 : go i.readState.db.maybeScheduleCompactionAsync()
2381 0 : }
2382 : }
2383 :
2384 1 : i.readState.unref()
2385 1 : i.readState = nil
2386 : }
2387 :
2388 1 : if i.version != nil {
2389 1 : i.version.Unref()
2390 1 : }
2391 1 : if i.externalIter != nil {
2392 0 : err = firstError(err, i.externalIter.Close())
2393 0 : }
2394 :
2395 : // Close the closer for the current value if one was open.
2396 1 : if i.valueCloser != nil {
2397 0 : err = firstError(err, i.valueCloser.Close())
2398 0 : i.valueCloser = nil
2399 0 : }
2400 :
2401 1 : if i.rangeKey != nil {
2402 1 : i.rangeKey.rangeKeyBuffers.PrepareForReuse()
2403 1 : *i.rangeKey = iteratorRangeKeyState{
2404 1 : rangeKeyBuffers: i.rangeKey.rangeKeyBuffers,
2405 1 : }
2406 1 : iterRangeKeyStateAllocPool.Put(i.rangeKey)
2407 1 : i.rangeKey = nil
2408 1 : }
2409 1 : if alloc := i.alloc; alloc != nil {
2410 1 : var (
2411 1 : keyBuf []byte
2412 1 : boundsBuf [2][]byte
2413 1 : prefixOrFullSeekKey []byte
2414 1 : mergingIterHeapItems []mergingIterHeapItem
2415 1 : )
2416 1 :
2417 1 : // Avoid caching the key buf if it is overly large. The constant is fairly
2418 1 : // arbitrary.
2419 1 : if cap(i.keyBuf) < maxKeyBufCacheSize {
2420 1 : keyBuf = i.keyBuf
2421 1 : }
2422 1 : if cap(i.prefixOrFullSeekKey) < maxKeyBufCacheSize {
2423 1 : prefixOrFullSeekKey = i.prefixOrFullSeekKey
2424 1 : }
2425 1 : for j := range i.boundsBuf {
2426 1 : if cap(i.boundsBuf[j]) < maxKeyBufCacheSize {
2427 1 : boundsBuf[j] = i.boundsBuf[j]
2428 1 : }
2429 : }
2430 1 : mergingIterHeapItems = alloc.merging.heap.items
2431 1 :
2432 1 : // Reset the alloc struct, re-assign the fields that are being recycled, and
2433 1 : // then return it to the pool. Splitting the first two steps performs better
2434 1 : // than doing them in a single step (e.g. *alloc = iterAlloc{...}) because
2435 1 : // the compiler can avoid the use of a stack allocated autotmp iterAlloc
2436 1 : // variable (~12KB, as of Dec 2024), which must first be zeroed out, then
2437 1 : // assigned into, then copied over into the heap-allocated alloc. Instead,
2438 1 : // the two-step process allows the compiler to quickly zero out the heap
2439 1 : // allocated object and then assign the few fields we want to preserve.
2440 1 : //
2441 1 : // TODO(nvanbenschoten): even with this optimization, zeroing out the alloc
2442 1 : // struct still shows up in profiles because it is such a large struct. Can
2443 1 : // we do something better here? We are hanging 22 separated iterators off of
2444 1 : // the alloc struct (or more, depending on how you count), many of which are
2445 1 : // only used in a few cases. Can those iterators be responsible for zeroing
2446 1 : // out their own memory on Close, allowing us to assume that most of the
2447 1 : // alloc struct is already zeroed out by this point?
2448 1 : *alloc = iterAlloc{}
2449 1 : alloc.keyBuf = keyBuf
2450 1 : alloc.boundsBuf = boundsBuf
2451 1 : alloc.prefixOrFullSeekKey = prefixOrFullSeekKey
2452 1 : alloc.merging.heap.items = mergingIterHeapItems
2453 1 :
2454 1 : iterAllocPool.Put(alloc)
2455 1 : } else if alloc := i.getIterAlloc; alloc != nil {
2456 1 : if cap(i.keyBuf) >= maxKeyBufCacheSize {
2457 0 : alloc.keyBuf = nil
2458 1 : } else {
2459 1 : alloc.keyBuf = i.keyBuf
2460 1 : }
2461 1 : *alloc = getIterAlloc{
2462 1 : keyBuf: alloc.keyBuf,
2463 1 : }
2464 1 : getIterAllocPool.Put(alloc)
2465 : }
2466 1 : return err
2467 : }
2468 :
2469 : // SetBounds sets the lower and upper bounds for the iterator. Once SetBounds
2470 : // returns, the caller is free to mutate the provided slices.
2471 : //
2472 : // The iterator will always be invalidated and must be repositioned with a call
2473 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2474 1 : func (i *Iterator) SetBounds(lower, upper []byte) {
2475 1 : // Ensure that the Iterator appears exhausted, regardless of whether we
2476 1 : // actually have to invalidate the internal iterator. Optimizations that
2477 1 : // avoid exhaustion are an internal implementation detail that shouldn't
2478 1 : // leak through the interface. The caller should still call an absolute
2479 1 : // positioning method to reposition the iterator.
2480 1 : i.requiresReposition = true
2481 1 :
2482 1 : if ((i.opts.LowerBound == nil) == (lower == nil)) &&
2483 1 : ((i.opts.UpperBound == nil) == (upper == nil)) &&
2484 1 : i.equal(i.opts.LowerBound, lower) &&
2485 1 : i.equal(i.opts.UpperBound, upper) {
2486 1 : // Unchanged, noop.
2487 1 : return
2488 1 : }
2489 :
2490 : // Copy the user-provided bounds into an Iterator-owned buffer, and set them
2491 : // on i.opts.{Lower,Upper}Bound.
2492 1 : i.processBounds(lower, upper)
2493 1 :
2494 1 : i.iter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2495 1 : // If the iterator has an open point iterator that's not currently being
2496 1 : // used, propagate the new bounds to it.
2497 1 : if i.pointIter != nil && !i.opts.pointKeys() {
2498 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2499 1 : }
2500 : // If the iterator has a range key iterator, propagate bounds to it. The
2501 : // top-level SetBounds on the interleaving iterator (i.iter) won't propagate
2502 : // bounds to the range key iterator stack, because the FragmentIterator
2503 : // interface doesn't define a SetBounds method. We need to directly inform
2504 : // the iterConfig stack.
2505 1 : if i.rangeKey != nil {
2506 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2507 1 : }
2508 :
2509 : // Even though this is not a positioning operation, the alteration of the
2510 : // bounds means we cannot optimize Seeks by using Next.
2511 1 : i.invalidate()
2512 : }
2513 :
2514 : // SetContext replaces the context provided at iterator creation, or the last
2515 : // one provided by SetContext. Even though iterators are expected to be
2516 : // short-lived, there are some cases where either (a) iterators are used far
2517 : // from the code that created them, (b) iterators are reused (while being
2518 : // short-lived) for processing different requests. For such scenarios, we
2519 : // allow the caller to replace the context.
2520 0 : func (i *Iterator) SetContext(ctx context.Context) {
2521 0 : i.ctx = ctx
2522 0 : i.iter.SetContext(ctx)
2523 0 : // If the iterator has an open point iterator that's not currently being
2524 0 : // used, propagate the new context to it.
2525 0 : if i.pointIter != nil && !i.opts.pointKeys() {
2526 0 : i.pointIter.SetContext(i.ctx)
2527 0 : }
2528 : }
2529 :
2530 : // Initialization and changing of the bounds must call processBounds.
2531 : // processBounds saves the bounds and computes derived state from those
2532 : // bounds.
2533 1 : func (i *Iterator) processBounds(lower, upper []byte) {
2534 1 : // Copy the user-provided bounds into an Iterator-owned buffer. We can't
2535 1 : // overwrite the current bounds, because some internal iterators compare old
2536 1 : // and new bounds for optimizations.
2537 1 :
2538 1 : buf := i.boundsBuf[i.boundsBufIdx][:0]
2539 1 : if lower != nil {
2540 1 : buf = append(buf, lower...)
2541 1 : i.opts.LowerBound = buf
2542 1 : } else {
2543 1 : i.opts.LowerBound = nil
2544 1 : }
2545 1 : i.nextPrefixNotPermittedByUpperBound = false
2546 1 : if upper != nil {
2547 1 : buf = append(buf, upper...)
2548 1 : i.opts.UpperBound = buf[len(buf)-len(upper):]
2549 1 : if i.comparer.Split(i.opts.UpperBound) != len(i.opts.UpperBound) {
2550 1 : // Setting an upper bound that is a versioned MVCC key. This means
2551 1 : // that a key can have some MVCC versions before the upper bound and
2552 1 : // some after. This causes significant complications for NextPrefix,
2553 1 : // so we bar the user of NextPrefix.
2554 1 : i.nextPrefixNotPermittedByUpperBound = true
2555 1 : }
2556 1 : } else {
2557 1 : i.opts.UpperBound = nil
2558 1 : }
2559 1 : i.boundsBuf[i.boundsBufIdx] = buf
2560 1 : i.boundsBufIdx = 1 - i.boundsBufIdx
2561 : }
2562 :
2563 : // SetOptions sets new iterator options for the iterator. Note that the lower
2564 : // and upper bounds applied here will supersede any bounds set by previous calls
2565 : // to SetBounds.
2566 : //
2567 : // Note that the slices provided in this SetOptions must not be changed by the
2568 : // caller until the iterator is closed, or a subsequent SetBounds or SetOptions
2569 : // has returned. This is because comparisons between the existing and new bounds
2570 : // are sometimes used to optimize seeking. See the extended commentary on
2571 : // SetBounds.
2572 : //
2573 : // If the iterator was created over an indexed mutable batch, the iterator's
2574 : // view of the mutable batch is refreshed.
2575 : //
2576 : // The iterator will always be invalidated and must be repositioned with a call
2577 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2578 : //
2579 : // If only lower and upper bounds need to be modified, prefer SetBounds.
2580 1 : func (i *Iterator) SetOptions(o *IterOptions) {
2581 1 : if i.externalIter != nil {
2582 0 : if err := validateExternalIterOpts(o); err != nil {
2583 0 : panic(err)
2584 : }
2585 : }
2586 :
2587 : // Ensure that the Iterator appears exhausted, regardless of whether we
2588 : // actually have to invalidate the internal iterator. Optimizations that
2589 : // avoid exhaustion are an internal implementation detail that shouldn't
2590 : // leak through the interface. The caller should still call an absolute
2591 : // positioning method to reposition the iterator.
2592 1 : i.requiresReposition = true
2593 1 :
2594 1 : // Check if global state requires we close all internal iterators.
2595 1 : //
2596 1 : // If the Iterator is in an error state, invalidate the existing iterators
2597 1 : // so that we reconstruct an iterator state from scratch.
2598 1 : //
2599 1 : // If OnlyReadGuaranteedDurable changed, the iterator stacks are incorrect,
2600 1 : // improperly including or excluding memtables. Invalidate them so that
2601 1 : // finishInitializingIter will reconstruct them.
2602 1 : closeBoth := i.err != nil ||
2603 1 : o.OnlyReadGuaranteedDurable != i.opts.OnlyReadGuaranteedDurable
2604 1 :
2605 1 : // If either options specify block property filters for an iterator stack,
2606 1 : // reconstruct it.
2607 1 : if i.pointIter != nil && (closeBoth || len(o.PointKeyFilters) > 0 || len(i.opts.PointKeyFilters) > 0 ||
2608 1 : o.RangeKeyMasking.Filter != nil || i.opts.RangeKeyMasking.Filter != nil || o.SkipPoint != nil ||
2609 1 : i.opts.SkipPoint != nil) {
2610 1 : i.err = firstError(i.err, i.pointIter.Close())
2611 1 : i.pointIter = nil
2612 1 : }
2613 1 : if i.rangeKey != nil {
2614 1 : if closeBoth || len(o.RangeKeyFilters) > 0 || len(i.opts.RangeKeyFilters) > 0 {
2615 1 : i.rangeKey.rangeKeyIter.Close()
2616 1 : i.rangeKey = nil
2617 1 : } else {
2618 1 : // If there's still a range key iterator stack, invalidate the
2619 1 : // iterator. This ensures RangeKeyChanged() returns true if a
2620 1 : // subsequent positioning operation discovers a range key. It also
2621 1 : // prevents seek no-op optimizations.
2622 1 : i.invalidate()
2623 1 : }
2624 : }
2625 :
2626 : // If the iterator is backed by a batch that's been mutated, refresh its
2627 : // existing point and range-key iterators, and invalidate the iterator to
2628 : // prevent seek-using-next optimizations. If we don't yet have a point-key
2629 : // iterator or range-key iterator but we require one, it'll be created in
2630 : // the slow path that reconstructs the iterator in finishInitializingIter.
2631 1 : if i.batch != nil {
2632 0 : nextBatchSeqNum := (base.SeqNum(len(i.batch.data)) | base.SeqNumBatchBit)
2633 0 : if nextBatchSeqNum != i.batchSeqNum {
2634 0 : i.batchSeqNum = nextBatchSeqNum
2635 0 : if i.merging != nil {
2636 0 : i.merging.batchSnapshot = nextBatchSeqNum
2637 0 : }
2638 : // Prevent a no-op seek optimization on the next seek. We won't be
2639 : // able to reuse the top-level Iterator state, because it may be
2640 : // incorrect after the inclusion of new batch mutations.
2641 0 : i.batchJustRefreshed = true
2642 0 : if i.pointIter != nil && i.batch.countRangeDels > 0 {
2643 0 : if i.batchRangeDelIter.Count() == 0 {
2644 0 : // When we constructed this iterator, there were no
2645 0 : // rangedels in the batch. Iterator construction will
2646 0 : // have excluded the batch rangedel iterator from the
2647 0 : // point iterator stack. We need to reconstruct the
2648 0 : // point iterator to add i.batchRangeDelIter into the
2649 0 : // iterator stack.
2650 0 : i.err = firstError(i.err, i.pointIter.Close())
2651 0 : i.pointIter = nil
2652 0 : } else {
2653 0 : // There are range deletions in the batch and we already
2654 0 : // have a batch rangedel iterator. We can update the
2655 0 : // batch rangedel iterator in place.
2656 0 : //
2657 0 : // NB: There may or may not be new range deletions. We
2658 0 : // can't tell based on i.batchRangeDelIter.Count(),
2659 0 : // which is the count of fragmented range deletions, NOT
2660 0 : // the number of range deletions written to the batch
2661 0 : // [i.batch.countRangeDels].
2662 0 : i.batch.initRangeDelIter(&i.opts, &i.batchRangeDelIter, nextBatchSeqNum)
2663 0 : }
2664 : }
2665 0 : if i.rangeKey != nil && i.batch.countRangeKeys > 0 {
2666 0 : if i.batchRangeKeyIter.Count() == 0 {
2667 0 : // When we constructed this iterator, there were no range
2668 0 : // keys in the batch. Iterator construction will have
2669 0 : // excluded the batch rangekey iterator from the range key
2670 0 : // iterator stack. We need to reconstruct the range key
2671 0 : // iterator to add i.batchRangeKeyIter into the iterator
2672 0 : // stack.
2673 0 : i.rangeKey.rangeKeyIter.Close()
2674 0 : i.rangeKey = nil
2675 0 : } else {
2676 0 : // There are range keys in the batch and we already
2677 0 : // have a batch rangekey iterator. We can update the batch
2678 0 : // rangekey iterator in place.
2679 0 : //
2680 0 : // NB: There may or may not be new range keys. We can't
2681 0 : // tell based on i.batchRangeKeyIter.Count(), which is the
2682 0 : // count of fragmented range keys, NOT the number of
2683 0 : // range keys written to the batch [i.batch.countRangeKeys].
2684 0 : i.batch.initRangeKeyIter(&i.opts, &i.batchRangeKeyIter, nextBatchSeqNum)
2685 0 : i.invalidate()
2686 0 : }
2687 : }
2688 : }
2689 : }
2690 :
2691 : // Reset combinedIterState.initialized in case the iterator key types
2692 : // changed. If there's already a range key iterator stack, the combined
2693 : // iterator is already initialized. Additionally, if the iterator is not
2694 : // configured to include range keys, mark it as initialized to signal that
2695 : // lower level iterators should not trigger a switch to combined iteration.
2696 1 : i.lazyCombinedIter.combinedIterState = combinedIterState{
2697 1 : initialized: i.rangeKey != nil || !i.opts.rangeKeys(),
2698 1 : }
2699 1 :
2700 1 : boundsEqual := ((i.opts.LowerBound == nil) == (o.LowerBound == nil)) &&
2701 1 : ((i.opts.UpperBound == nil) == (o.UpperBound == nil)) &&
2702 1 : i.equal(i.opts.LowerBound, o.LowerBound) &&
2703 1 : i.equal(i.opts.UpperBound, o.UpperBound)
2704 1 :
2705 1 : if boundsEqual && o.KeyTypes == i.opts.KeyTypes &&
2706 1 : (i.pointIter != nil || !i.opts.pointKeys()) &&
2707 1 : (i.rangeKey != nil || !i.opts.rangeKeys() || i.opts.KeyTypes == IterKeyTypePointsAndRanges) &&
2708 1 : i.comparer.CompareRangeSuffixes(o.RangeKeyMasking.Suffix, i.opts.RangeKeyMasking.Suffix) == 0 &&
2709 1 : o.UseL6Filters == i.opts.UseL6Filters {
2710 1 : // The options are identical, so we can likely use the fast path. In
2711 1 : // addition to all the above constraints, we cannot use the fast path if
2712 1 : // configured to perform lazy combined iteration but an indexed batch
2713 1 : // used by the iterator now contains range keys. Lazy combined iteration
2714 1 : // is not compatible with batch range keys because we always need to
2715 1 : // merge the batch's range keys into iteration.
2716 1 : if i.rangeKey != nil || !i.opts.rangeKeys() || i.batch == nil || i.batch.countRangeKeys == 0 {
2717 1 : // Fast path. This preserves the Seek-using-Next optimizations as
2718 1 : // long as the iterator wasn't already invalidated up above.
2719 1 : return
2720 1 : }
2721 : }
2722 : // Slow path.
2723 :
2724 : // The options changed. Save the new ones to i.opts.
2725 1 : if boundsEqual {
2726 1 : // Copying the options into i.opts will overwrite LowerBound and
2727 1 : // UpperBound fields with the user-provided slices. We need to hold on
2728 1 : // to the Pebble-owned slices, so save them and re-set them after the
2729 1 : // copy.
2730 1 : lower, upper := i.opts.LowerBound, i.opts.UpperBound
2731 1 : i.opts = *o
2732 1 : i.opts.LowerBound, i.opts.UpperBound = lower, upper
2733 1 : } else {
2734 1 : i.opts = *o
2735 1 : i.processBounds(o.LowerBound, o.UpperBound)
2736 1 : // Propagate the changed bounds to the existing point iterator.
2737 1 : // NB: We propagate i.opts.{Lower,Upper}Bound, not o.{Lower,Upper}Bound
2738 1 : // because i.opts now point to buffers owned by Pebble.
2739 1 : if i.pointIter != nil {
2740 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2741 1 : }
2742 1 : if i.rangeKey != nil {
2743 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2744 1 : }
2745 : }
2746 :
2747 : // Even though this is not a positioning operation, the invalidation of the
2748 : // iterator stack means we cannot optimize Seeks by using Next.
2749 1 : i.invalidate()
2750 1 :
2751 1 : // Iterators created through NewExternalIter have a different iterator
2752 1 : // initialization process.
2753 1 : if i.externalIter != nil {
2754 0 : _ = finishInitializingExternal(i.ctx, i)
2755 0 : return
2756 0 : }
2757 1 : finishInitializingIter(i.ctx, i.alloc)
2758 : }
2759 :
2760 1 : func (i *Iterator) invalidate() {
2761 1 : i.lastPositioningOp = unknownLastPositionOp
2762 1 : i.hasPrefix = false
2763 1 : i.iterKV = nil
2764 1 : i.err = nil
2765 1 : // This switch statement isn't necessary for correctness since callers
2766 1 : // should call a repositioning method. We could have arbitrarily set i.pos
2767 1 : // to one of the values. But it results in more intuitive behavior in
2768 1 : // tests, which do not always reposition.
2769 1 : switch i.pos {
2770 1 : case iterPosCurForward, iterPosNext, iterPosCurForwardPaused:
2771 1 : i.pos = iterPosCurForward
2772 1 : case iterPosCurReverse, iterPosPrev, iterPosCurReversePaused:
2773 1 : i.pos = iterPosCurReverse
2774 : }
2775 1 : i.iterValidityState = IterExhausted
2776 1 : if i.rangeKey != nil {
2777 1 : i.rangeKey.iiter.Invalidate()
2778 1 : i.rangeKey.prevPosHadRangeKey = false
2779 1 : }
2780 : }
2781 :
2782 : // Metrics returns per-iterator metrics.
2783 0 : func (i *Iterator) Metrics() IteratorMetrics {
2784 0 : m := IteratorMetrics{
2785 0 : ReadAmp: 1,
2786 0 : }
2787 0 : if mi, ok := i.iter.(*mergingIter); ok {
2788 0 : m.ReadAmp = len(mi.levels)
2789 0 : }
2790 0 : return m
2791 : }
2792 :
2793 : // ResetStats resets the stats to 0.
2794 0 : func (i *Iterator) ResetStats() {
2795 0 : i.stats = IteratorStats{}
2796 0 : }
2797 :
2798 : // Stats returns the current stats.
2799 0 : func (i *Iterator) Stats() IteratorStats {
2800 0 : return i.stats
2801 0 : }
2802 :
2803 : // CloneOptions configures an iterator constructed through Iterator.Clone.
2804 : type CloneOptions struct {
2805 : // IterOptions, if non-nil, define the iterator options to configure a
2806 : // cloned iterator. If nil, the clone adopts the same IterOptions as the
2807 : // iterator being cloned.
2808 : IterOptions *IterOptions
2809 : // RefreshBatchView may be set to true when cloning an Iterator over an
2810 : // indexed batch. When false, the clone adopts the same (possibly stale)
2811 : // view of the indexed batch as the cloned Iterator. When true, the clone is
2812 : // constructed with a refreshed view of the batch, observing all of the
2813 : // batch's mutations at the time of the Clone. If the cloned iterator was
2814 : // not constructed to read over an indexed batch, RefreshVatchView has no
2815 : // effect.
2816 : RefreshBatchView bool
2817 : }
2818 :
2819 : // Clone creates a new Iterator over the same underlying data, i.e., over the
2820 : // same {batch, memtables, sstables}). The resulting iterator is not positioned.
2821 : // It starts with the same IterOptions, unless opts.IterOptions is set.
2822 : //
2823 : // When called on an Iterator over an indexed batch, the clone's visibility of
2824 : // the indexed batch is determined by CloneOptions.RefreshBatchView. If false,
2825 : // the clone inherits the iterator's current (possibly stale) view of the batch,
2826 : // and callers may call SetOptions to subsequently refresh the clone's view to
2827 : // include all batch mutations. If true, the clone is constructed with a
2828 : // complete view of the indexed batch's mutations at the time of the Clone.
2829 : //
2830 : // Callers can use Clone if they need multiple iterators that need to see
2831 : // exactly the same underlying state of the DB. This should not be used to
2832 : // extend the lifetime of the data backing the original Iterator since that
2833 : // will cause an increase in memory and disk usage (use NewSnapshot for that
2834 : // purpose).
2835 1 : func (i *Iterator) Clone(opts CloneOptions) (*Iterator, error) {
2836 1 : return i.CloneWithContext(context.Background(), opts)
2837 1 : }
2838 :
2839 : // CloneWithContext is like Clone, and additionally accepts a context for
2840 : // tracing.
2841 1 : func (i *Iterator) CloneWithContext(ctx context.Context, opts CloneOptions) (*Iterator, error) {
2842 1 : if opts.IterOptions == nil {
2843 1 : opts.IterOptions = &i.opts
2844 1 : }
2845 1 : if i.batchOnlyIter {
2846 0 : return nil, errors.Errorf("cannot Clone a batch-only Iterator")
2847 0 : }
2848 1 : readState := i.readState
2849 1 : vers := i.version
2850 1 : if readState == nil && vers == nil {
2851 0 : return nil, errors.Errorf("cannot Clone a closed Iterator")
2852 0 : }
2853 : // i is already holding a ref, so there is no race with unref here.
2854 : //
2855 : // TODO(bilal): If the underlying iterator was created on a snapshot, we could
2856 : // grab a reference to the current readState instead of reffing the original
2857 : // readState. This allows us to release references to some zombie sstables
2858 : // and memtables.
2859 1 : if readState != nil {
2860 1 : readState.ref()
2861 1 : }
2862 1 : if vers != nil {
2863 1 : vers.Ref()
2864 1 : }
2865 : // Bundle various structures under a single umbrella in order to allocate
2866 : // them together.
2867 1 : buf := iterAllocPool.Get().(*iterAlloc)
2868 1 : dbi := &buf.dbi
2869 1 : *dbi = Iterator{
2870 1 : ctx: ctx,
2871 1 : opts: *opts.IterOptions,
2872 1 : alloc: buf,
2873 1 : merge: i.merge,
2874 1 : comparer: i.comparer,
2875 1 : readState: readState,
2876 1 : version: vers,
2877 1 : keyBuf: buf.keyBuf,
2878 1 : prefixOrFullSeekKey: buf.prefixOrFullSeekKey,
2879 1 : boundsBuf: buf.boundsBuf,
2880 1 : batch: i.batch,
2881 1 : batchSeqNum: i.batchSeqNum,
2882 1 : fc: i.fc,
2883 1 : newIters: i.newIters,
2884 1 : newIterRangeKey: i.newIterRangeKey,
2885 1 : seqNum: i.seqNum,
2886 1 : }
2887 1 : dbi.processBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
2888 1 :
2889 1 : // If the caller requested the clone have a current view of the indexed
2890 1 : // batch, set the clone's batch sequence number appropriately.
2891 1 : if i.batch != nil && opts.RefreshBatchView {
2892 0 : dbi.batchSeqNum = (base.SeqNum(len(i.batch.data)) | base.SeqNumBatchBit)
2893 0 : }
2894 :
2895 1 : return finishInitializingIter(ctx, buf), nil
2896 : }
2897 :
2898 : // Merge adds all of the argument's statistics to the receiver. It may be used
2899 : // to accumulate stats across multiple iterators.
2900 0 : func (stats *IteratorStats) Merge(o IteratorStats) {
2901 0 : for i := InterfaceCall; i < NumStatsKind; i++ {
2902 0 : stats.ForwardSeekCount[i] += o.ForwardSeekCount[i]
2903 0 : stats.ReverseSeekCount[i] += o.ReverseSeekCount[i]
2904 0 : stats.ForwardStepCount[i] += o.ForwardStepCount[i]
2905 0 : stats.ReverseStepCount[i] += o.ReverseStepCount[i]
2906 0 : }
2907 0 : stats.InternalStats.Merge(o.InternalStats)
2908 0 : stats.RangeKeyStats.Merge(o.RangeKeyStats)
2909 : }
2910 :
2911 0 : func (stats *IteratorStats) String() string {
2912 0 : return redact.StringWithoutMarkers(stats)
2913 0 : }
2914 :
2915 : // SafeFormat implements the redact.SafeFormatter interface.
2916 0 : func (stats *IteratorStats) SafeFormat(s redact.SafePrinter, verb rune) {
2917 0 : if stats.ReverseSeekCount[InterfaceCall] == 0 && stats.ReverseSeekCount[InternalIterCall] == 0 {
2918 0 : s.Printf("seeked %s times (%s internal)",
2919 0 : humanize.Count.Uint64(uint64(stats.ForwardSeekCount[InterfaceCall])),
2920 0 : humanize.Count.Uint64(uint64(stats.ForwardSeekCount[InternalIterCall])),
2921 0 : )
2922 0 : } else {
2923 0 : s.Printf("seeked %s times (%s fwd/%s rev, internal: %s fwd/%s rev)",
2924 0 : humanize.Count.Uint64(uint64(stats.ForwardSeekCount[InterfaceCall]+stats.ReverseSeekCount[InterfaceCall])),
2925 0 : humanize.Count.Uint64(uint64(stats.ForwardSeekCount[InterfaceCall])),
2926 0 : humanize.Count.Uint64(uint64(stats.ReverseSeekCount[InterfaceCall])),
2927 0 : humanize.Count.Uint64(uint64(stats.ForwardSeekCount[InternalIterCall])),
2928 0 : humanize.Count.Uint64(uint64(stats.ReverseSeekCount[InternalIterCall])),
2929 0 : )
2930 0 : }
2931 0 : s.SafeString("; ")
2932 0 :
2933 0 : if stats.ReverseStepCount[InterfaceCall] == 0 && stats.ReverseStepCount[InternalIterCall] == 0 {
2934 0 : s.Printf("stepped %s times (%s internal)",
2935 0 : humanize.Count.Uint64(uint64(stats.ForwardStepCount[InterfaceCall])),
2936 0 : humanize.Count.Uint64(uint64(stats.ForwardStepCount[InternalIterCall])),
2937 0 : )
2938 0 : } else {
2939 0 : s.Printf("stepped %s times (%s fwd/%s rev, internal: %s fwd/%s rev)",
2940 0 : humanize.Count.Uint64(uint64(stats.ForwardStepCount[InterfaceCall]+stats.ReverseStepCount[InterfaceCall])),
2941 0 : humanize.Count.Uint64(uint64(stats.ForwardStepCount[InterfaceCall])),
2942 0 : humanize.Count.Uint64(uint64(stats.ReverseStepCount[InterfaceCall])),
2943 0 : humanize.Count.Uint64(uint64(stats.ForwardStepCount[InternalIterCall])),
2944 0 : humanize.Count.Uint64(uint64(stats.ReverseStepCount[InternalIterCall])),
2945 0 : )
2946 0 : }
2947 :
2948 0 : if stats.InternalStats != (InternalIteratorStats{}) {
2949 0 : s.SafeString("; ")
2950 0 : stats.InternalStats.SafeFormat(s, verb)
2951 0 : }
2952 0 : if stats.RangeKeyStats != (RangeKeyIteratorStats{}) {
2953 0 : s.SafeString(", ")
2954 0 : stats.RangeKeyStats.SafeFormat(s, verb)
2955 0 : }
2956 : }
2957 :
2958 : // CanDeterministicallySingleDelete takes a valid iterator and examines internal
2959 : // state to determine if a SingleDelete deleting Iterator.Key() would
2960 : // deterministically delete the key. CanDeterministicallySingleDelete requires
2961 : // the iterator to be oriented in the forward direction (eg, the last
2962 : // positioning operation must've been a First, a Seek[Prefix]GE, or a
2963 : // Next[Prefix][WithLimit]).
2964 : //
2965 : // This function does not change the external position of the iterator, and all
2966 : // positioning methods should behave the same as if it was never called. This
2967 : // function will only return a meaningful result the first time it's invoked at
2968 : // an iterator position. This function invalidates the iterator Value's memory,
2969 : // and the caller must not rely on the memory safety of the previous Iterator
2970 : // position.
2971 : //
2972 : // If CanDeterministicallySingleDelete returns true AND the key at the iterator
2973 : // position is not modified between the creation of the Iterator and the commit
2974 : // of a batch containing a SingleDelete over the key, then the caller can be
2975 : // assured that SingleDelete is equivalent to Delete on the local engine, but it
2976 : // may not be true on another engine that received the same writes and with
2977 : // logically equivalent state since this engine may have collapsed multiple SETs
2978 : // into one.
2979 1 : func CanDeterministicallySingleDelete(it *Iterator) (bool, error) {
2980 1 : // This function may only be called once per external iterator position. We
2981 1 : // can validate this by checking the last positioning operation.
2982 1 : if it.lastPositioningOp == internalNextOp {
2983 1 : return false, errors.New("pebble: CanDeterministicallySingleDelete called twice")
2984 1 : }
2985 1 : validity, kind := it.internalNext()
2986 1 : var shadowedBySingleDelete bool
2987 1 : for validity == internalNextValid {
2988 1 : switch kind {
2989 1 : case InternalKeyKindDelete, InternalKeyKindDeleteSized:
2990 1 : // A DEL or DELSIZED tombstone is okay. An internal key
2991 1 : // sequence like SINGLEDEL; SET; DEL; SET can be handled
2992 1 : // deterministically. If there are SETs further down, we
2993 1 : // don't care about them.
2994 1 : return true, nil
2995 1 : case InternalKeyKindSingleDelete:
2996 1 : // A SingleDelete is okay as long as when that SingleDelete was
2997 1 : // written, it was written deterministically (eg, with its own
2998 1 : // CanDeterministicallySingleDelete check). Validate that it was
2999 1 : // written deterministically. We'll allow one set to appear after
3000 1 : // the SingleDelete.
3001 1 : shadowedBySingleDelete = true
3002 1 : validity, kind = it.internalNext()
3003 1 : continue
3004 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete, InternalKeyKindMerge:
3005 1 : // If we observed a single delete, it's allowed to delete 1 key.
3006 1 : // We'll keep looping to validate that the internal keys beneath the
3007 1 : // already-written single delete are copacetic.
3008 1 : if shadowedBySingleDelete {
3009 1 : shadowedBySingleDelete = false
3010 1 : validity, kind = it.internalNext()
3011 1 : continue
3012 : }
3013 : // We encountered a shadowed SET, SETWITHDEL, MERGE. A SINGLEDEL
3014 : // that deleted the KV at the original iterator position could
3015 : // result in this key becoming visible.
3016 1 : return false, nil
3017 0 : case InternalKeyKindRangeDelete:
3018 0 : // RangeDeletes are handled by the merging iterator and should never
3019 0 : // be observed by the top-level Iterator.
3020 0 : panic(errors.AssertionFailedf("pebble: unexpected range delete"))
3021 0 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
3022 0 : // Range keys are interleaved at the maximal sequence number and
3023 0 : // should never be observed within a user key.
3024 0 : panic(errors.AssertionFailedf("pebble: unexpected range key"))
3025 0 : default:
3026 0 : panic(errors.AssertionFailedf("pebble: unexpected key kind: %s", errors.Safe(kind)))
3027 : }
3028 : }
3029 1 : if validity == internalNextError {
3030 1 : return false, it.Error()
3031 1 : }
3032 1 : return true, nil
3033 : }
3034 :
3035 : // internalNextValidity enumerates the potential outcomes of a call to
3036 : // internalNext.
3037 : type internalNextValidity int8
3038 :
3039 : const (
3040 : // internalNextError is returned by internalNext when an error occurred and
3041 : // the caller is responsible for checking iter.Error().
3042 : internalNextError internalNextValidity = iota
3043 : // internalNextExhausted is returned by internalNext when the next internal
3044 : // key is an internal key with a different user key than Iterator.Key().
3045 : internalNextExhausted
3046 : // internalNextValid is returned by internalNext when the internal next
3047 : // found a shadowed internal key with a user key equal to Iterator.Key().
3048 : internalNextValid
3049 : )
3050 :
3051 : // internalNext advances internal Iterator state forward to expose the
3052 : // InternalKeyKind of the next internal key with a user key equal to Key().
3053 : //
3054 : // internalNext is a highly specialized operation and is unlikely to be
3055 : // generally useful. See Iterator.Next for how to reposition the iterator to the
3056 : // next key. internalNext requires the Iterator to be at a valid position in the
3057 : // forward direction (the last positioning operation must've been a First, a
3058 : // Seek[Prefix]GE, or a Next[Prefix][WithLimit] and Valid() must return true).
3059 : //
3060 : // internalNext, unlike all other Iterator methods, exposes internal LSM state.
3061 : // internalNext advances the Iterator's internal iterator to the next shadowed
3062 : // key with a user key equal to Key(). When a key is overwritten or deleted, its
3063 : // removal from the LSM occurs lazily as a part of compactions. internalNext
3064 : // allows the caller to see whether an obsolete internal key exists with the
3065 : // current Key(), and what it's key kind is. Note that the existence of an
3066 : // internal key is nondeterministic and dependent on internal LSM state. These
3067 : // semantics are unlikely to be applicable to almost all use cases.
3068 : //
3069 : // If internalNext finds a key that shares the same user key as Key(), it
3070 : // returns internalNextValid and the internal key's kind. If internalNext
3071 : // encounters an error, it returns internalNextError and the caller is expected
3072 : // to call Iterator.Error() to retrieve it. In all other circumstances,
3073 : // internalNext returns internalNextExhausted, indicating that there are no more
3074 : // additional internal keys with the user key Key().
3075 : //
3076 : // internalNext does not change the external position of the iterator, and a
3077 : // Next operation should behave the same as if internalNext was never called.
3078 : // internalNext does invalidate the iterator Value's memory, and the caller must
3079 : // not rely on the memory safety of the previous Iterator position.
3080 1 : func (i *Iterator) internalNext() (internalNextValidity, base.InternalKeyKind) {
3081 1 : i.stats.ForwardStepCount[InterfaceCall]++
3082 1 : if i.err != nil {
3083 1 : return internalNextError, base.InternalKeyKindInvalid
3084 1 : } else if i.iterValidityState != IterValid {
3085 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3086 1 : }
3087 1 : i.lastPositioningOp = internalNextOp
3088 1 :
3089 1 : switch i.pos {
3090 1 : case iterPosCurForward:
3091 1 : i.iterKV = i.iter.Next()
3092 1 : if i.iterKV == nil {
3093 1 : // We check i.iter.Error() here and return an internalNextError enum
3094 1 : // variant so that the caller does not need to check i.iter.Error()
3095 1 : // in the common case that the next internal key has a new user key.
3096 1 : if i.err = i.iter.Error(); i.err != nil {
3097 0 : return internalNextError, base.InternalKeyKindInvalid
3098 0 : }
3099 1 : i.pos = iterPosNext
3100 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3101 1 : } else if i.comparer.Equal(i.iterKV.K.UserKey, i.key) {
3102 1 : return internalNextValid, i.iterKV.Kind()
3103 1 : }
3104 1 : i.pos = iterPosNext
3105 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3106 1 : case iterPosCurReverse, iterPosCurReversePaused, iterPosPrev:
3107 1 : i.err = errors.New("switching from reverse to forward via internalNext is prohibited")
3108 1 : i.iterValidityState = IterExhausted
3109 1 : return internalNextError, base.InternalKeyKindInvalid
3110 1 : case iterPosNext, iterPosCurForwardPaused:
3111 1 : // The previous method already moved onto the next user key. This is
3112 1 : // only possible if
3113 1 : // - the last positioning method was a call to internalNext, and we
3114 1 : // advanced to a new user key.
3115 1 : // - the previous non-internalNext iterator operation encountered a
3116 1 : // range key or merge, forcing an internal Next that found a new
3117 1 : // user key that's not equal to i.Iterator.Key().
3118 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3119 0 : default:
3120 0 : panic("unreachable")
3121 : }
3122 : }
3123 :
3124 : var _ base.IteratorDebug = (*Iterator)(nil)
3125 :
3126 : // DebugTree implements the base.IteratorDebug interface.
3127 0 : func (i *Iterator) DebugTree(tp treeprinter.Node) {
3128 0 : n := tp.Childf("%T(%p)", i, i)
3129 0 : if i.iter != nil {
3130 0 : i.iter.DebugTree(n)
3131 0 : }
3132 0 : if i.pointIter != nil {
3133 0 : i.pointIter.DebugTree(n)
3134 0 : }
3135 : }
|
