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 : "sync"
12 : "unsafe"
13 :
14 : "github.com/cockroachdb/errors"
15 : "github.com/cockroachdb/pebble/internal/base"
16 : "github.com/cockroachdb/pebble/internal/bytealloc"
17 : "github.com/cockroachdb/pebble/internal/fastrand"
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/sstable"
25 : "github.com/cockroachdb/redact"
26 : )
27 :
28 : // iterPos describes the state of the internal iterator, in terms of whether it
29 : // is at the position returned to the user (cur), one ahead of the position
30 : // returned (next for forward iteration and prev for reverse iteration). The cur
31 : // position is split into two states, for forward and reverse iteration, since
32 : // we need to differentiate for switching directions.
33 : //
34 : // There is subtlety in what is considered the current position of the Iterator.
35 : // The internal iterator exposes a sequence of internal keys. There is not
36 : // always a single internalIterator position corresponding to the position
37 : // returned to the user. Consider the example:
38 : //
39 : // a.MERGE.9 a.MERGE.8 a.MERGE.7 a.SET.6 b.DELETE.9 b.DELETE.5 b.SET.4
40 : // \ /
41 : // \ Iterator.Key() = 'a' /
42 : //
43 : // The Iterator exposes one valid position at user key 'a' and the two exhausted
44 : // positions at the beginning and end of iteration. The underlying
45 : // internalIterator contains 7 valid positions and 2 exhausted positions.
46 : //
47 : // Iterator positioning methods must set iterPos to iterPosCur{Foward,Backward}
48 : // iff the user key at the current internalIterator position equals the
49 : // Iterator.Key returned to the user. This guarantees that a call to nextUserKey
50 : // or prevUserKey will advance to the next or previous iterator position.
51 : // iterPosCur{Forward,Backward} does not make any guarantee about the internal
52 : // iterator position among internal keys with matching user keys, and it will
53 : // vary subtly depending on the particular key kinds encountered. In the above
54 : // example, the iterator returning 'a' may set iterPosCurForward if the internal
55 : // iterator is positioned at any of a.MERGE.9, a.MERGE.8, a.MERGE.7 or a.SET.6.
56 : //
57 : // When setting iterPos to iterPosNext or iterPosPrev, the internal iterator
58 : // must be advanced to the first internalIterator position at a user key greater
59 : // (iterPosNext) or less (iterPosPrev) than the key returned to the user. An
60 : // internalIterator position that's !Valid() must also be considered greater or
61 : // less—depending on the direction of iteration—than the last valid Iterator
62 : // position.
63 : type iterPos int8
64 :
65 : const (
66 : iterPosCurForward iterPos = 0
67 : iterPosNext iterPos = 1
68 : iterPosPrev iterPos = -1
69 : iterPosCurReverse iterPos = -2
70 :
71 : // For limited iteration. When the iterator is at iterPosCurForwardPaused
72 : // - Next*() call should behave as if the internal iterator is already
73 : // at next (akin to iterPosNext).
74 : // - Prev*() call should behave as if the internal iterator is at the
75 : // current key (akin to iterPosCurForward).
76 : //
77 : // Similar semantics apply to CurReversePaused.
78 : iterPosCurForwardPaused iterPos = 2
79 : iterPosCurReversePaused iterPos = -3
80 : )
81 :
82 : // Approximate gap in bytes between samples of data read during iteration.
83 : // This is multiplied with a default ReadSamplingMultiplier of 1 << 4 to yield
84 : // 1 << 20 (1MB). The 1MB factor comes from:
85 : // https://github.com/cockroachdb/pebble/issues/29#issuecomment-494477985
86 : const readBytesPeriod uint64 = 1 << 16
87 :
88 : var errReversePrefixIteration = errors.New("pebble: unsupported reverse prefix iteration")
89 :
90 : // IteratorMetrics holds per-iterator metrics. These do not change over the
91 : // lifetime of the iterator.
92 : type IteratorMetrics struct {
93 : // The read amplification experienced by this iterator. This is the sum of
94 : // the memtables, the L0 sublevels and the non-empty Ln levels. Higher read
95 : // amplification generally results in slower reads, though allowing higher
96 : // read amplification can also result in faster writes.
97 : ReadAmp int
98 : }
99 :
100 : // IteratorStatsKind describes the two kind of iterator stats.
101 : type IteratorStatsKind int8
102 :
103 : const (
104 : // InterfaceCall represents calls to Iterator.
105 : InterfaceCall IteratorStatsKind = iota
106 : // InternalIterCall represents calls by Iterator to its internalIterator.
107 : InternalIterCall
108 : // NumStatsKind is the number of kinds, and is used for array sizing.
109 : NumStatsKind
110 : )
111 :
112 : // IteratorStats contains iteration stats.
113 : type IteratorStats struct {
114 : // ForwardSeekCount includes SeekGE, SeekPrefixGE, First.
115 : ForwardSeekCount [NumStatsKind]int
116 : // ReverseSeek includes SeekLT, Last.
117 : ReverseSeekCount [NumStatsKind]int
118 : // ForwardStepCount includes Next.
119 : ForwardStepCount [NumStatsKind]int
120 : // ReverseStepCount includes Prev.
121 : ReverseStepCount [NumStatsKind]int
122 : InternalStats InternalIteratorStats
123 : RangeKeyStats RangeKeyIteratorStats
124 : }
125 :
126 : var _ redact.SafeFormatter = &IteratorStats{}
127 :
128 : // InternalIteratorStats contains miscellaneous stats produced by internal
129 : // iterators.
130 : type InternalIteratorStats = base.InternalIteratorStats
131 :
132 : // RangeKeyIteratorStats contains miscellaneous stats about range keys
133 : // encountered by the iterator.
134 : type RangeKeyIteratorStats struct {
135 : // Count records the number of range keys encountered during
136 : // iteration. Range keys may be counted multiple times if the iterator
137 : // leaves a range key's bounds and then returns.
138 : Count int
139 : // ContainedPoints records the number of point keys encountered within the
140 : // bounds of a range key. Note that this includes point keys with suffixes
141 : // that sort both above and below the covering range key's suffix.
142 : ContainedPoints int
143 : // SkippedPoints records the count of the subset of ContainedPoints point
144 : // keys that were skipped during iteration due to range-key masking. It does
145 : // not include point keys that were never loaded because a
146 : // RangeKeyMasking.Filter excluded the entire containing block.
147 : SkippedPoints int
148 : }
149 :
150 : // Merge adds all of the argument's statistics to the receiver. It may be used
151 : // to accumulate stats across multiple iterators.
152 0 : func (s *RangeKeyIteratorStats) Merge(o RangeKeyIteratorStats) {
153 0 : s.Count += o.Count
154 0 : s.ContainedPoints += o.ContainedPoints
155 0 : s.SkippedPoints += o.SkippedPoints
156 0 : }
157 :
158 : // LazyValue is a lazy value. See the long comment in base.LazyValue.
159 : type LazyValue = base.LazyValue
160 :
161 : // Iterator iterates over a DB's key/value pairs in key order.
162 : //
163 : // An iterator must be closed after use, but it is not necessary to read an
164 : // iterator until exhaustion.
165 : //
166 : // An iterator is not goroutine-safe, but it is safe to use multiple iterators
167 : // concurrently, with each in a dedicated goroutine.
168 : //
169 : // It is also safe to use an iterator concurrently with modifying its
170 : // underlying DB, if that DB permits modification. However, the resultant
171 : // key/value pairs are not guaranteed to be a consistent snapshot of that DB
172 : // at a particular point in time.
173 : //
174 : // If an iterator encounters an error during any operation, it is stored by
175 : // the Iterator and surfaced through the Error method. All absolute
176 : // positioning methods (eg, SeekLT, SeekGT, First, Last, etc) reset any
177 : // accumulated error before positioning. All relative positioning methods (eg,
178 : // Next, Prev) return without advancing if the iterator has an accumulated
179 : // error.
180 : type Iterator struct {
181 : // The context is stored here since (a) Iterators are expected to be
182 : // short-lived (since they pin memtables and sstables), (b) plumbing a
183 : // context into every method is very painful, (c) they do not (yet) respect
184 : // context cancellation and are only used for tracing.
185 : ctx context.Context
186 : opts IterOptions
187 : merge Merge
188 : comparer base.Comparer
189 : iter internalIterator
190 : pointIter topLevelIterator
191 : // Either readState or version is set, but not both.
192 : readState *readState
193 : version *version
194 : // rangeKey holds iteration state specific to iteration over range keys.
195 : // The range key field may be nil if the Iterator has never been configured
196 : // to iterate over range keys. Its non-nilness cannot be used to determine
197 : // if the Iterator is currently iterating over range keys: For that, consult
198 : // the IterOptions using opts.rangeKeys(). If non-nil, its rangeKeyIter
199 : // field is guaranteed to be non-nil too.
200 : rangeKey *iteratorRangeKeyState
201 : // rangeKeyMasking holds state for range-key masking of point keys.
202 : rangeKeyMasking rangeKeyMasking
203 : err error
204 : // When iterValidityState=IterValid, key represents the current key, which
205 : // is backed by keyBuf.
206 : key []byte
207 : keyBuf []byte
208 : value LazyValue
209 : // For use in LazyValue.Clone.
210 : valueBuf []byte
211 : fetcher base.LazyFetcher
212 : // For use in LazyValue.Value.
213 : lazyValueBuf []byte
214 : valueCloser io.Closer
215 : // boundsBuf holds two buffers used to store the lower and upper bounds.
216 : // Whenever the Iterator's bounds change, the new bounds are copied into
217 : // boundsBuf[boundsBufIdx]. The two bounds share a slice to reduce
218 : // allocations. opts.LowerBound and opts.UpperBound point into this slice.
219 : boundsBuf [2][]byte
220 : boundsBufIdx int
221 : // iterKV reflects the latest position of iter, except when SetBounds is
222 : // called. In that case, it is explicitly set to nil.
223 : iterKV *base.InternalKV
224 : alloc *iterAlloc
225 : getIterAlloc *getIterAlloc
226 : prefixOrFullSeekKey []byte
227 : readSampling readSampling
228 : stats IteratorStats
229 : externalReaders [][]*sstable.Reader
230 :
231 : // Following fields used when constructing an iterator stack, eg, in Clone
232 : // and SetOptions or when re-fragmenting a batch's range keys/range dels.
233 : // Non-nil if this Iterator includes a Batch.
234 : batch *Batch
235 : newIters tableNewIters
236 : newIterRangeKey keyspanimpl.TableNewSpanIter
237 : lazyCombinedIter lazyCombinedIter
238 : seqNum uint64
239 : // batchSeqNum is used by Iterators over indexed batches to detect when the
240 : // underlying batch has been mutated. The batch beneath an indexed batch may
241 : // be mutated while the Iterator is open, but new keys are not surfaced
242 : // until the next call to SetOptions.
243 : batchSeqNum uint64
244 : // batch{PointIter,RangeDelIter,RangeKeyIter} are used when the Iterator is
245 : // configured to read through an indexed batch. If a batch is set, these
246 : // iterators will be included within the iterator stack regardless of
247 : // whether the batch currently contains any keys of their kind. These
248 : // pointers are used during a call to SetOptions to refresh the Iterator's
249 : // view of its indexed batch.
250 : batchPointIter batchIter
251 : batchRangeDelIter keyspan.Iter
252 : batchRangeKeyIter keyspan.Iter
253 : // merging is a pointer to this iterator's point merging iterator. It
254 : // appears here because key visibility is handled by the merging iterator.
255 : // During SetOptions on an iterator over an indexed batch, this field is
256 : // used to update the merging iterator's batch snapshot.
257 : merging *mergingIter
258 :
259 : // Keeping the bools here after all the 8 byte aligned fields shrinks the
260 : // sizeof this struct by 24 bytes.
261 :
262 : // INVARIANT:
263 : // iterValidityState==IterAtLimit <=>
264 : // pos==iterPosCurForwardPaused || pos==iterPosCurReversePaused
265 : iterValidityState IterValidityState
266 : // Set to true by SetBounds, SetOptions. Causes the Iterator to appear
267 : // exhausted externally, while preserving the correct iterValidityState for
268 : // the iterator's internal state. Preserving the correct internal validity
269 : // is used for SeekPrefixGE(..., trySeekUsingNext), and SeekGE/SeekLT
270 : // optimizations after "no-op" calls to SetBounds and SetOptions.
271 : requiresReposition bool
272 : // The position of iter. When this is iterPos{Prev,Next} the iter has been
273 : // moved past the current key-value, which can only happen if
274 : // iterValidityState=IterValid, i.e., there is something to return to the
275 : // client for the current position.
276 : pos iterPos
277 : // Relates to the prefixOrFullSeekKey field above.
278 : hasPrefix bool
279 : // Used for deriving the value of SeekPrefixGE(..., trySeekUsingNext),
280 : // and SeekGE/SeekLT optimizations
281 : lastPositioningOp lastPositioningOpKind
282 : // Used for determining when it's safe to perform SeekGE optimizations that
283 : // reuse the iterator state to avoid the cost of a full seek if the iterator
284 : // is already positioned in the correct place. If the iterator's view of its
285 : // indexed batch was just refreshed, some optimizations cannot be applied on
286 : // the first seek after the refresh:
287 : // - SeekGE has a no-op optimization that does not seek on the internal
288 : // iterator at all if the iterator is already in the correct place.
289 : // This optimization cannot be performed if the internal iterator was
290 : // last positioned when the iterator had a different view of an
291 : // underlying batch.
292 : // - Seek[Prefix]GE set flags.TrySeekUsingNext()=true when the seek key is
293 : // greater than the previous operation's seek key, under the expectation
294 : // that the various internal iterators can use their current position to
295 : // avoid a full expensive re-seek. This applies to the batchIter as well.
296 : // However, if the view of the batch was just refreshed, the batchIter's
297 : // position is not useful because it may already be beyond new keys less
298 : // than the seek key. To prevent the use of this optimization in
299 : // batchIter, Seek[Prefix]GE set flags.BatchJustRefreshed()=true if this
300 : // bit is enabled.
301 : batchJustRefreshed bool
302 : // batchOnlyIter is set to true for Batch.NewBatchOnlyIter.
303 : batchOnlyIter bool
304 : // Used in some tests to disable the random disabling of seek optimizations.
305 : forceEnableSeekOpt bool
306 : // Set to true if NextPrefix is not currently permitted. Defaults to false
307 : // in case an iterator never had any bounds.
308 : nextPrefixNotPermittedByUpperBound bool
309 : }
310 :
311 : // cmp is a convenience shorthand for the i.comparer.Compare function.
312 1 : func (i *Iterator) cmp(a, b []byte) int {
313 1 : return i.comparer.Compare(a, b)
314 1 : }
315 :
316 : // split is a convenience shorthand for the i.comparer.Split function.
317 1 : func (i *Iterator) split(a []byte) int {
318 1 : return i.comparer.Split(a)
319 1 : }
320 :
321 : // equal is a convenience shorthand for the i.comparer.Equal function.
322 1 : func (i *Iterator) equal(a, b []byte) bool {
323 1 : return i.comparer.Equal(a, b)
324 1 : }
325 :
326 : // iteratorRangeKeyState holds an iterator's range key iteration state.
327 : type iteratorRangeKeyState struct {
328 : opts *IterOptions
329 : cmp base.Compare
330 : split base.Split
331 : // rangeKeyIter holds the range key iterator stack that iterates over the
332 : // merged spans across the entirety of the LSM.
333 : rangeKeyIter keyspan.FragmentIterator
334 : iiter keyspan.InterleavingIter
335 : // stale is set to true when the range key state recorded here (in start,
336 : // end and keys) may not be in sync with the current range key at the
337 : // interleaving iterator's current position.
338 : //
339 : // When the interelaving iterator passes over a new span, it invokes the
340 : // SpanChanged hook defined on the `rangeKeyMasking` type, which sets stale
341 : // to true if the span is non-nil.
342 : //
343 : // The parent iterator may not be positioned over the interleaving
344 : // iterator's current position (eg, i.iterPos = iterPos{Next,Prev}), so
345 : // {keys,start,end} are only updated to the new range key during a call to
346 : // Iterator.saveRangeKey.
347 : stale bool
348 : // updated is used to signal to the Iterator client whether the state of
349 : // range keys has changed since the previous iterator position through the
350 : // `RangeKeyChanged` method. It's set to true during an Iterator positioning
351 : // operation that changes the state of the current range key. Each Iterator
352 : // positioning operation sets it back to false before executing.
353 : //
354 : // TODO(jackson): The lifecycle of {stale,updated,prevPosHadRangeKey} is
355 : // intricate and confusing. Try to refactor to reduce complexity.
356 : updated bool
357 : // prevPosHadRangeKey records whether the previous Iterator position had a
358 : // range key (HasPointAndRage() = (_, true)). It's updated at the beginning
359 : // of each new Iterator positioning operation. It's required by saveRangeKey to
360 : // to set `updated` appropriately: Without this record of the previous iterator
361 : // state, it's ambiguous whether an iterator only temporarily stepped onto a
362 : // position without a range key.
363 : prevPosHadRangeKey bool
364 : // rangeKeyOnly is set to true if at the current iterator position there is
365 : // no point key, only a range key start boundary.
366 : rangeKeyOnly bool
367 : // hasRangeKey is true when the current iterator position has a covering
368 : // range key (eg, a range key with bounds [<lower>,<upper>) such that
369 : // <lower> ≤ Key() < <upper>).
370 : hasRangeKey bool
371 : // start and end are the [start, end) boundaries of the current range keys.
372 : start []byte
373 : end []byte
374 :
375 : rangeKeyBuffers
376 :
377 : // iterConfig holds fields that are used for the construction of the
378 : // iterator stack, but do not need to be directly accessed during iteration.
379 : // This struct is bundled within the iteratorRangeKeyState struct to reduce
380 : // allocations.
381 : iterConfig rangekeystack.UserIteratorConfig
382 : }
383 :
384 : type rangeKeyBuffers struct {
385 : // keys is sorted by Suffix ascending.
386 : keys []RangeKeyData
387 : // buf is used to save range-key data before moving the range-key iterator.
388 : // Start and end boundaries, suffixes and values are all copied into buf.
389 : buf bytealloc.A
390 : // internal holds buffers used by the range key internal iterators.
391 : internal rangekeystack.Buffers
392 : }
393 :
394 1 : func (b *rangeKeyBuffers) PrepareForReuse() {
395 1 : const maxKeysReuse = 100
396 1 : if len(b.keys) > maxKeysReuse {
397 0 : b.keys = nil
398 0 : }
399 : // Avoid caching the key buf if it is overly large. The constant is
400 : // fairly arbitrary.
401 1 : if cap(b.buf) >= maxKeyBufCacheSize {
402 1 : b.buf = nil
403 1 : } else {
404 1 : b.buf = b.buf[:0]
405 1 : }
406 1 : b.internal.PrepareForReuse()
407 : }
408 :
409 1 : func (i *iteratorRangeKeyState) init(cmp base.Compare, split base.Split, opts *IterOptions) {
410 1 : i.cmp = cmp
411 1 : i.split = split
412 1 : i.opts = opts
413 1 : }
414 :
415 : var iterRangeKeyStateAllocPool = sync.Pool{
416 1 : New: func() interface{} {
417 1 : return &iteratorRangeKeyState{}
418 1 : },
419 : }
420 :
421 : // isEphemeralPosition returns true iff the current iterator position is
422 : // ephemeral, and won't be visited during subsequent relative positioning
423 : // operations.
424 : //
425 : // The iterator position resulting from a SeekGE or SeekPrefixGE that lands on a
426 : // straddling range key without a coincident point key is such a position.
427 1 : func (i *Iterator) isEphemeralPosition() bool {
428 1 : return i.opts.rangeKeys() && i.rangeKey != nil && i.rangeKey.rangeKeyOnly &&
429 1 : !i.equal(i.rangeKey.start, i.key)
430 1 : }
431 :
432 : type lastPositioningOpKind int8
433 :
434 : const (
435 : unknownLastPositionOp lastPositioningOpKind = iota
436 : seekPrefixGELastPositioningOp
437 : seekGELastPositioningOp
438 : seekLTLastPositioningOp
439 : // internalNextOp is a special internal iterator positioning operation used
440 : // by CanDeterministicallySingleDelete. It exists for enforcing requirements
441 : // around calling CanDeterministicallySingleDelete at most once per external
442 : // iterator position.
443 : internalNextOp
444 : )
445 :
446 : // Limited iteration mode. Not for use with prefix iteration.
447 : //
448 : // SeekGE, SeekLT, Prev, Next have WithLimit variants, that pause the iterator
449 : // at the limit in a best-effort manner. The client should behave correctly
450 : // even if the limits are ignored. These limits are not "deep", in that they
451 : // are not passed down to the underlying collection of internalIterators. This
452 : // is because the limits are transient, and apply only until the next
453 : // iteration call. They serve mainly as a way to bound the amount of work when
454 : // two (or more) Iterators are being coordinated at a higher level.
455 : //
456 : // In limited iteration mode:
457 : // - Avoid using Iterator.Valid if the last call was to a *WithLimit() method.
458 : // The return value from the *WithLimit() method provides a more precise
459 : // disposition.
460 : // - The limit is exclusive for forward and inclusive for reverse.
461 : //
462 : //
463 : // Limited iteration mode & range keys
464 : //
465 : // Limited iteration interacts with range-key iteration. When range key
466 : // iteration is enabled, range keys are interleaved at their start boundaries.
467 : // Limited iteration must ensure that if a range key exists within the limit,
468 : // the iterator visits the range key.
469 : //
470 : // During forward limited iteration, this is trivial: An overlapping range key
471 : // must have a start boundary less than the limit, and the range key's start
472 : // boundary will be interleaved and found to be within the limit.
473 : //
474 : // During reverse limited iteration, the tail of the range key may fall within
475 : // the limit. The range key must be surfaced even if the range key's start
476 : // boundary is less than the limit, and if there are no point keys between the
477 : // current iterator position and the limit. To provide this guarantee, reverse
478 : // limited iteration ignores the limit as long as there is a range key
479 : // overlapping the iteration position.
480 :
481 : // IterValidityState captures the state of the Iterator.
482 : type IterValidityState int8
483 :
484 : const (
485 : // IterExhausted represents an Iterator that is exhausted.
486 : IterExhausted IterValidityState = iota
487 : // IterValid represents an Iterator that is valid.
488 : IterValid
489 : // IterAtLimit represents an Iterator that has a non-exhausted
490 : // internalIterator, but has reached a limit without any key for the
491 : // caller.
492 : IterAtLimit
493 : )
494 :
495 : // readSampling stores variables used to sample a read to trigger a read
496 : // compaction
497 : type readSampling struct {
498 : bytesUntilReadSampling uint64
499 : initialSamplePassed bool
500 : pendingCompactions readCompactionQueue
501 : // forceReadSampling is used for testing purposes to force a read sample on every
502 : // call to Iterator.maybeSampleRead()
503 : forceReadSampling bool
504 : }
505 :
506 1 : func (i *Iterator) findNextEntry(limit []byte) {
507 1 : i.iterValidityState = IterExhausted
508 1 : i.pos = iterPosCurForward
509 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
510 1 : i.rangeKey.rangeKeyOnly = false
511 1 : }
512 :
513 : // Close the closer for the current value if one was open.
514 1 : if i.closeValueCloser() != nil {
515 1 : return
516 1 : }
517 :
518 1 : for i.iterKV != nil {
519 1 : key := i.iterKV.K
520 1 :
521 1 : // The topLevelIterator.StrictSeekPrefixGE contract requires that in
522 1 : // prefix mode [i.hasPrefix=t], every point key returned by the internal
523 1 : // iterator must have the current iteration prefix.
524 1 : if invariants.Enabled && i.hasPrefix {
525 1 : // Range keys are an exception to the contract and may return a different
526 1 : // prefix. This case is explicitly handled in the switch statement below.
527 1 : if key.Kind() != base.InternalKeyKindRangeKeySet {
528 1 : if n := i.split(key.UserKey); !i.equal(i.prefixOrFullSeekKey, key.UserKey[:n]) {
529 0 : i.opts.logger.Fatalf("pebble: prefix violation: key %q does not have prefix %q\n", key.UserKey, i.prefixOrFullSeekKey)
530 0 : }
531 : }
532 : }
533 :
534 : // Compare with limit every time we start at a different user key.
535 : // Note that given the best-effort contract of limit, we could avoid a
536 : // comparison in the common case by doing this only after
537 : // i.nextUserKey is called for the deletes below. However that makes
538 : // the behavior non-deterministic (since the behavior will vary based
539 : // on what has been compacted), which makes it hard to test with the
540 : // metamorphic test. So we forego that performance optimization.
541 1 : if limit != nil && i.cmp(limit, i.iterKV.K.UserKey) <= 0 {
542 1 : i.iterValidityState = IterAtLimit
543 1 : i.pos = iterPosCurForwardPaused
544 1 : return
545 1 : }
546 :
547 : // If the user has configured a SkipPoint function, invoke it to see
548 : // whether we should skip over the current user key.
549 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(i.iterKV.K.UserKey) {
550 1 : // NB: We could call nextUserKey, but in some cases the SkipPoint
551 1 : // predicate function might be cheaper than nextUserKey's key copy
552 1 : // and key comparison. This should be the case for MVCC suffix
553 1 : // comparisons, for example. In the future, we could expand the
554 1 : // SkipPoint interface to give the implementor more control over
555 1 : // whether we skip over just the internal key, the user key, or even
556 1 : // the key prefix.
557 1 : i.stats.ForwardStepCount[InternalIterCall]++
558 1 : i.iterKV = i.iter.Next()
559 1 : continue
560 : }
561 :
562 1 : switch key.Kind() {
563 1 : case InternalKeyKindRangeKeySet:
564 1 : if i.hasPrefix {
565 1 : if n := i.split(key.UserKey); !i.equal(i.prefixOrFullSeekKey, key.UserKey[:n]) {
566 1 : return
567 1 : }
568 : }
569 : // Save the current key.
570 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
571 1 : i.key = i.keyBuf
572 1 : i.value = LazyValue{}
573 1 : // There may also be a live point key at this userkey that we have
574 1 : // not yet read. We need to find the next entry with this user key
575 1 : // to find it. Save the range key so we don't lose it when we Next
576 1 : // the underlying iterator.
577 1 : i.saveRangeKey()
578 1 : pointKeyExists := i.nextPointCurrentUserKey()
579 1 : if i.err != nil {
580 0 : i.iterValidityState = IterExhausted
581 0 : return
582 0 : }
583 1 : i.rangeKey.rangeKeyOnly = !pointKeyExists
584 1 : i.iterValidityState = IterValid
585 1 : return
586 :
587 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
588 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
589 1 : // only simpler, but is also necessary for correctness due to
590 1 : // InternalKeyKindSSTableInternalObsoleteBit.
591 1 : i.nextUserKey()
592 1 : continue
593 :
594 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
595 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
596 1 : i.key = i.keyBuf
597 1 : i.value = i.iterKV.V
598 1 : i.iterValidityState = IterValid
599 1 : i.saveRangeKey()
600 1 : return
601 :
602 1 : case InternalKeyKindMerge:
603 1 : // Resolving the merge may advance us to the next point key, which
604 1 : // may be covered by a different set of range keys. Save the range
605 1 : // key state so we don't lose it.
606 1 : i.saveRangeKey()
607 1 : if i.mergeForward(key) {
608 1 : i.iterValidityState = IterValid
609 1 : return
610 1 : }
611 :
612 : // The merge didn't yield a valid key, either because the value
613 : // merger indicated it should be deleted, or because an error was
614 : // encountered.
615 0 : i.iterValidityState = IterExhausted
616 0 : if i.err != nil {
617 0 : return
618 0 : }
619 0 : if i.pos != iterPosNext {
620 0 : i.nextUserKey()
621 0 : }
622 0 : if i.closeValueCloser() != nil {
623 0 : return
624 0 : }
625 0 : i.pos = iterPosCurForward
626 :
627 0 : default:
628 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
629 0 : i.iterValidityState = IterExhausted
630 0 : return
631 : }
632 : }
633 :
634 : // Is iterKey nil due to an error?
635 1 : if err := i.iter.Error(); err != nil {
636 0 : i.err = err
637 0 : i.iterValidityState = IterExhausted
638 0 : }
639 : }
640 :
641 1 : func (i *Iterator) nextPointCurrentUserKey() bool {
642 1 : // If the user has configured a SkipPoint function and the current user key
643 1 : // would be skipped by it, there's no need to step forward looking for a
644 1 : // point key. If we were to find one, it should be skipped anyways.
645 1 : if i.opts.SkipPoint != nil && i.opts.SkipPoint(i.key) {
646 1 : return false
647 1 : }
648 :
649 1 : i.pos = iterPosCurForward
650 1 :
651 1 : i.iterKV = i.iter.Next()
652 1 : i.stats.ForwardStepCount[InternalIterCall]++
653 1 : if i.iterKV == nil {
654 1 : if err := i.iter.Error(); err != nil {
655 0 : i.err = err
656 1 : } else {
657 1 : i.pos = iterPosNext
658 1 : }
659 1 : return false
660 : }
661 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
662 1 : i.pos = iterPosNext
663 1 : return false
664 1 : }
665 :
666 1 : key := i.iterKV.K
667 1 : switch key.Kind() {
668 0 : case InternalKeyKindRangeKeySet:
669 0 : // RangeKeySets must always be interleaved as the first internal key
670 0 : // for a user key.
671 0 : i.err = base.CorruptionErrorf("pebble: unexpected range key set mid-user key")
672 0 : return false
673 :
674 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
675 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
676 1 : // only simpler, but is also necessary for correctness due to
677 1 : // InternalKeyKindSSTableInternalObsoleteBit.
678 1 : return false
679 :
680 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
681 1 : i.value = i.iterKV.V
682 1 : return true
683 :
684 1 : case InternalKeyKindMerge:
685 1 : return i.mergeForward(key)
686 :
687 0 : default:
688 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
689 0 : return false
690 : }
691 : }
692 :
693 : // mergeForward resolves a MERGE key, advancing the underlying iterator forward
694 : // to merge with subsequent keys with the same userkey. mergeForward returns a
695 : // boolean indicating whether or not the merge yielded a valid key. A merge may
696 : // not yield a valid key if an error occurred, in which case i.err is non-nil,
697 : // or the user's value merger specified the key to be deleted.
698 : //
699 : // mergeForward does not update iterValidityState.
700 1 : func (i *Iterator) mergeForward(key base.InternalKey) (valid bool) {
701 1 : var iterValue []byte
702 1 : iterValue, _, i.err = i.iterKV.Value(nil)
703 1 : if i.err != nil {
704 0 : return false
705 0 : }
706 1 : var valueMerger ValueMerger
707 1 : valueMerger, i.err = i.merge(key.UserKey, iterValue)
708 1 : if i.err != nil {
709 0 : return false
710 0 : }
711 :
712 1 : i.mergeNext(key, valueMerger)
713 1 : if i.err != nil {
714 0 : return false
715 0 : }
716 :
717 1 : var needDelete bool
718 1 : var value []byte
719 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(
720 1 : valueMerger, true /* includesBase */)
721 1 : i.value = base.MakeInPlaceValue(value)
722 1 : if i.err != nil {
723 0 : return false
724 0 : }
725 1 : if needDelete {
726 0 : _ = i.closeValueCloser()
727 0 : return false
728 0 : }
729 1 : return true
730 : }
731 :
732 1 : func (i *Iterator) closeValueCloser() error {
733 1 : if i.valueCloser != nil {
734 0 : i.err = i.valueCloser.Close()
735 0 : i.valueCloser = nil
736 0 : }
737 1 : return i.err
738 : }
739 :
740 1 : func (i *Iterator) nextUserKey() {
741 1 : if i.iterKV == nil {
742 1 : return
743 1 : }
744 1 : trailer := i.iterKV.K.Trailer
745 1 : done := i.iterKV.K.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
746 1 : if i.iterValidityState != IterValid {
747 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKV.K.UserKey...)
748 1 : i.key = i.keyBuf
749 1 : }
750 1 : for {
751 1 : i.stats.ForwardStepCount[InternalIterCall]++
752 1 : i.iterKV = i.iter.Next()
753 1 : if i.iterKV == nil {
754 1 : if err := i.iter.Error(); err != nil {
755 0 : i.err = err
756 0 : return
757 0 : }
758 : }
759 : // NB: We're guaranteed to be on the next user key if the previous key
760 : // had a zero sequence number (`done`), or the new key has a trailer
761 : // greater or equal to the previous key's trailer. This is true because
762 : // internal keys with the same user key are sorted by Trailer in
763 : // strictly monotonically descending order. We expect the trailer
764 : // optimization to trigger around 50% of the time with randomly
765 : // distributed writes. We expect it to trigger very frequently when
766 : // iterating through ingested sstables, which contain keys that all have
767 : // the same sequence number.
768 1 : if done || i.iterKV == nil || i.iterKV.K.Trailer >= trailer {
769 1 : break
770 : }
771 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
772 1 : break
773 : }
774 1 : done = i.iterKV.K.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
775 1 : trailer = i.iterKV.K.Trailer
776 : }
777 : }
778 :
779 1 : func (i *Iterator) maybeSampleRead() {
780 1 : // This method is only called when a public method of Iterator is
781 1 : // returning, and below we exclude the case were the iterator is paused at
782 1 : // a limit. The effect of these choices is that keys that are deleted, but
783 1 : // are encountered during iteration, are not accounted for in the read
784 1 : // sampling and will not cause read driven compactions, even though we are
785 1 : // incurring cost in iterating over them. And this issue is not limited to
786 1 : // Iterator, which does not see the effect of range deletes, which may be
787 1 : // causing iteration work in mergingIter. It is not clear at this time
788 1 : // whether this is a deficiency worth addressing.
789 1 : if i.iterValidityState != IterValid {
790 1 : return
791 1 : }
792 1 : if i.readState == nil {
793 1 : return
794 1 : }
795 1 : if i.readSampling.forceReadSampling {
796 0 : i.sampleRead()
797 0 : return
798 0 : }
799 1 : samplingPeriod := int32(int64(readBytesPeriod) * i.readState.db.opts.Experimental.ReadSamplingMultiplier)
800 1 : if samplingPeriod <= 0 {
801 0 : return
802 0 : }
803 1 : bytesRead := uint64(len(i.key) + i.value.Len())
804 1 : for i.readSampling.bytesUntilReadSampling < bytesRead {
805 1 : i.readSampling.bytesUntilReadSampling += uint64(fastrand.Uint32n(2 * uint32(samplingPeriod)))
806 1 : // The block below tries to adjust for the case where this is the
807 1 : // first read in a newly-opened iterator. As bytesUntilReadSampling
808 1 : // starts off at zero, we don't want to sample the first read of
809 1 : // every newly-opened iterator, but we do want to sample some of them.
810 1 : if !i.readSampling.initialSamplePassed {
811 1 : i.readSampling.initialSamplePassed = true
812 1 : if fastrand.Uint32n(uint32(i.readSampling.bytesUntilReadSampling)) > uint32(bytesRead) {
813 1 : continue
814 : }
815 : }
816 1 : i.sampleRead()
817 : }
818 1 : i.readSampling.bytesUntilReadSampling -= bytesRead
819 : }
820 :
821 1 : func (i *Iterator) sampleRead() {
822 1 : var topFile *manifest.FileMetadata
823 1 : topLevel, numOverlappingLevels := numLevels, 0
824 1 : mi := i.merging
825 1 : if mi == nil {
826 1 : return
827 1 : }
828 1 : if len(mi.levels) > 1 {
829 1 : mi.ForEachLevelIter(func(li *levelIter) bool {
830 1 : l := manifest.LevelToInt(li.level)
831 1 : if f := li.iterFile; f != nil {
832 1 : var containsKey bool
833 1 : if i.pos == iterPosNext || i.pos == iterPosCurForward ||
834 1 : i.pos == iterPosCurForwardPaused {
835 1 : containsKey = i.cmp(f.SmallestPointKey.UserKey, i.key) <= 0
836 1 : } else if i.pos == iterPosPrev || i.pos == iterPosCurReverse ||
837 1 : i.pos == iterPosCurReversePaused {
838 1 : containsKey = i.cmp(f.LargestPointKey.UserKey, i.key) >= 0
839 1 : }
840 : // Do nothing if the current key is not contained in f's
841 : // bounds. We could seek the LevelIterator at this level
842 : // to find the right file, but the performance impacts of
843 : // doing that are significant enough to negate the benefits
844 : // of read sampling in the first place. See the discussion
845 : // at:
846 : // https://github.com/cockroachdb/pebble/pull/1041#issuecomment-763226492
847 1 : if containsKey {
848 1 : numOverlappingLevels++
849 1 : if numOverlappingLevels >= 2 {
850 1 : // Terminate the loop early if at least 2 overlapping levels are found.
851 1 : return true
852 1 : }
853 1 : topLevel = l
854 1 : topFile = f
855 : }
856 : }
857 1 : return false
858 : })
859 : }
860 1 : if topFile == nil || topLevel >= numLevels {
861 1 : return
862 1 : }
863 1 : if numOverlappingLevels >= 2 {
864 1 : allowedSeeks := topFile.AllowedSeeks.Add(-1)
865 1 : if allowedSeeks == 0 {
866 0 :
867 0 : // Since the compaction queue can handle duplicates, we can keep
868 0 : // adding to the queue even once allowedSeeks hits 0.
869 0 : // In fact, we NEED to keep adding to the queue, because the queue
870 0 : // is small and evicts older and possibly useful compactions.
871 0 : topFile.AllowedSeeks.Add(topFile.InitAllowedSeeks)
872 0 :
873 0 : read := readCompaction{
874 0 : start: topFile.SmallestPointKey.UserKey,
875 0 : end: topFile.LargestPointKey.UserKey,
876 0 : level: topLevel,
877 0 : fileNum: topFile.FileNum,
878 0 : }
879 0 : i.readSampling.pendingCompactions.add(&read, i.cmp)
880 0 : }
881 : }
882 : }
883 :
884 1 : func (i *Iterator) findPrevEntry(limit []byte) {
885 1 : i.iterValidityState = IterExhausted
886 1 : i.pos = iterPosCurReverse
887 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
888 1 : i.rangeKey.rangeKeyOnly = false
889 1 : }
890 :
891 : // Close the closer for the current value if one was open.
892 1 : if i.valueCloser != nil {
893 0 : i.err = i.valueCloser.Close()
894 0 : i.valueCloser = nil
895 0 : if i.err != nil {
896 0 : i.iterValidityState = IterExhausted
897 0 : return
898 0 : }
899 : }
900 :
901 1 : var valueMerger ValueMerger
902 1 : firstLoopIter := true
903 1 : rangeKeyBoundary := false
904 1 : // The code below compares with limit in multiple places. As documented in
905 1 : // findNextEntry, this is being done to make the behavior of limit
906 1 : // deterministic to allow for metamorphic testing. It is not required by
907 1 : // the best-effort contract of limit.
908 1 : for i.iterKV != nil {
909 1 : key := i.iterKV.K
910 1 :
911 1 : // NB: We cannot pause if the current key is covered by a range key.
912 1 : // Otherwise, the user might not ever learn of a range key that covers
913 1 : // the key space being iterated over in which there are no point keys.
914 1 : // Since limits are best effort, ignoring the limit in this case is
915 1 : // allowed by the contract of limit.
916 1 : if firstLoopIter && limit != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
917 1 : i.iterValidityState = IterAtLimit
918 1 : i.pos = iterPosCurReversePaused
919 1 : return
920 1 : }
921 1 : firstLoopIter = false
922 1 :
923 1 : if i.iterValidityState == IterValid {
924 1 : if !i.equal(key.UserKey, i.key) {
925 1 : // We've iterated to the previous user key.
926 1 : i.pos = iterPosPrev
927 1 : if valueMerger != nil {
928 1 : var needDelete bool
929 1 : var value []byte
930 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
931 1 : i.value = base.MakeInPlaceValue(value)
932 1 : if i.err == nil && needDelete {
933 0 : // The point key at this key is deleted. If we also have
934 0 : // a range key boundary at this key, we still want to
935 0 : // return. Otherwise, we need to continue looking for
936 0 : // a live key.
937 0 : i.value = LazyValue{}
938 0 : if rangeKeyBoundary {
939 0 : i.rangeKey.rangeKeyOnly = true
940 0 : } else {
941 0 : i.iterValidityState = IterExhausted
942 0 : if i.closeValueCloser() == nil {
943 0 : continue
944 : }
945 : }
946 : }
947 : }
948 1 : if i.err != nil {
949 0 : i.iterValidityState = IterExhausted
950 0 : }
951 1 : return
952 : }
953 : }
954 :
955 : // If the user has configured a SkipPoint function, invoke it to see
956 : // whether we should skip over the current user key.
957 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(key.UserKey) {
958 1 : // NB: We could call prevUserKey, but in some cases the SkipPoint
959 1 : // predicate function might be cheaper than prevUserKey's key copy
960 1 : // and key comparison. This should be the case for MVCC suffix
961 1 : // comparisons, for example. In the future, we could expand the
962 1 : // SkipPoint interface to give the implementor more control over
963 1 : // whether we skip over just the internal key, the user key, or even
964 1 : // the key prefix.
965 1 : i.stats.ReverseStepCount[InternalIterCall]++
966 1 : i.iterKV = i.iter.Prev()
967 1 : if i.iterKV == nil {
968 1 : if err := i.iter.Error(); err != nil {
969 0 : i.err = err
970 0 : i.iterValidityState = IterExhausted
971 0 : return
972 0 : }
973 : }
974 1 : if limit != nil && i.iterKV != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
975 1 : i.iterValidityState = IterAtLimit
976 1 : i.pos = iterPosCurReversePaused
977 1 : return
978 1 : }
979 1 : continue
980 : }
981 :
982 1 : switch key.Kind() {
983 1 : case InternalKeyKindRangeKeySet:
984 1 : // Range key start boundary markers are interleaved with the maximum
985 1 : // sequence number, so if there's a point key also at this key, we
986 1 : // must've already iterated over it.
987 1 : // This is the final entry at this user key, so we may return
988 1 : i.rangeKey.rangeKeyOnly = i.iterValidityState != IterValid
989 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
990 1 : i.key = i.keyBuf
991 1 : i.iterValidityState = IterValid
992 1 : i.saveRangeKey()
993 1 : // In all other cases, previous iteration requires advancing to
994 1 : // iterPosPrev in order to determine if the key is live and
995 1 : // unshadowed by another key at the same user key. In this case,
996 1 : // because range key start boundary markers are always interleaved
997 1 : // at the maximum sequence number, we know that there aren't any
998 1 : // additional keys with the same user key in the backward direction.
999 1 : //
1000 1 : // We Prev the underlying iterator once anyways for consistency, so
1001 1 : // that we can maintain the invariant during backward iteration that
1002 1 : // i.iterPos = iterPosPrev.
1003 1 : i.stats.ReverseStepCount[InternalIterCall]++
1004 1 : i.iterKV = i.iter.Prev()
1005 1 :
1006 1 : // Set rangeKeyBoundary so that on the next iteration, we know to
1007 1 : // return the key even if the MERGE point key is deleted.
1008 1 : rangeKeyBoundary = true
1009 :
1010 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
1011 1 : i.value = LazyValue{}
1012 1 : i.iterValidityState = IterExhausted
1013 1 : valueMerger = nil
1014 1 : i.stats.ReverseStepCount[InternalIterCall]++
1015 1 : i.iterKV = i.iter.Prev()
1016 1 : // Compare with the limit. We could optimize by only checking when
1017 1 : // we step to the previous user key, but detecting that requires a
1018 1 : // comparison too. Note that this position may already passed a
1019 1 : // number of versions of this user key, but they are all deleted, so
1020 1 : // the fact that a subsequent Prev*() call will not see them is
1021 1 : // harmless. Also note that this is the only place in the loop,
1022 1 : // other than the firstLoopIter and SkipPoint cases above, where we
1023 1 : // could step to a different user key and start processing it for
1024 1 : // returning to the caller.
1025 1 : if limit != nil && i.iterKV != nil && i.cmp(limit, i.iterKV.K.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
1026 1 : i.iterValidityState = IterAtLimit
1027 1 : i.pos = iterPosCurReversePaused
1028 1 : return
1029 1 : }
1030 1 : continue
1031 :
1032 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1033 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1034 1 : i.key = i.keyBuf
1035 1 : // iterValue is owned by i.iter and could change after the Prev()
1036 1 : // call, so use valueBuf instead. Note that valueBuf is only used
1037 1 : // in this one instance; everywhere else (eg. in findNextEntry),
1038 1 : // we just point i.value to the unsafe i.iter-owned value buffer.
1039 1 : i.value, i.valueBuf = i.iterKV.V.Clone(i.valueBuf[:0], &i.fetcher)
1040 1 : i.saveRangeKey()
1041 1 : i.iterValidityState = IterValid
1042 1 : i.iterKV = i.iter.Prev()
1043 1 : i.stats.ReverseStepCount[InternalIterCall]++
1044 1 : valueMerger = nil
1045 1 : continue
1046 :
1047 1 : case InternalKeyKindMerge:
1048 1 : if i.iterValidityState == IterExhausted {
1049 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1050 1 : i.key = i.keyBuf
1051 1 : i.saveRangeKey()
1052 1 : var iterValue []byte
1053 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1054 1 : if i.err != nil {
1055 0 : return
1056 0 : }
1057 1 : valueMerger, i.err = i.merge(i.key, iterValue)
1058 1 : if i.err != nil {
1059 0 : return
1060 0 : }
1061 1 : i.iterValidityState = IterValid
1062 1 : } else if valueMerger == nil {
1063 1 : // Extract value before iterValue since we use value before iterValue
1064 1 : // and the underlying iterator is not required to provide backing
1065 1 : // memory for both simultaneously.
1066 1 : var value []byte
1067 1 : var callerOwned bool
1068 1 : value, callerOwned, i.err = i.value.Value(i.lazyValueBuf)
1069 1 : if callerOwned {
1070 0 : i.lazyValueBuf = value[:0]
1071 0 : }
1072 1 : if i.err != nil {
1073 0 : i.iterValidityState = IterExhausted
1074 0 : return
1075 0 : }
1076 1 : valueMerger, i.err = i.merge(i.key, value)
1077 1 : var iterValue []byte
1078 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1079 1 : if i.err != nil {
1080 0 : i.iterValidityState = IterExhausted
1081 0 : return
1082 0 : }
1083 1 : if i.err == nil {
1084 1 : i.err = valueMerger.MergeNewer(iterValue)
1085 1 : }
1086 1 : if i.err != nil {
1087 0 : i.iterValidityState = IterExhausted
1088 0 : return
1089 0 : }
1090 1 : } else {
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 : i.err = valueMerger.MergeNewer(iterValue)
1098 1 : if i.err != nil {
1099 0 : i.iterValidityState = IterExhausted
1100 0 : return
1101 0 : }
1102 : }
1103 1 : i.iterKV = i.iter.Prev()
1104 1 : i.stats.ReverseStepCount[InternalIterCall]++
1105 1 : continue
1106 :
1107 0 : default:
1108 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1109 0 : i.iterValidityState = IterExhausted
1110 0 : return
1111 : }
1112 : }
1113 : // i.iterKV == nil, so broke out of the preceding loop.
1114 :
1115 : // Is iterKey nil due to an error?
1116 1 : if i.err = i.iter.Error(); i.err != nil {
1117 0 : i.iterValidityState = IterExhausted
1118 0 : return
1119 0 : }
1120 :
1121 1 : if i.iterValidityState == IterValid {
1122 1 : i.pos = iterPosPrev
1123 1 : if valueMerger != nil {
1124 1 : var needDelete bool
1125 1 : var value []byte
1126 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
1127 1 : i.value = base.MakeInPlaceValue(value)
1128 1 : if i.err == nil && needDelete {
1129 0 : i.key = nil
1130 0 : i.value = LazyValue{}
1131 0 : i.iterValidityState = IterExhausted
1132 0 : }
1133 : }
1134 1 : if i.err != nil {
1135 0 : i.iterValidityState = IterExhausted
1136 0 : }
1137 : }
1138 : }
1139 :
1140 1 : func (i *Iterator) prevUserKey() {
1141 1 : if i.iterKV == nil {
1142 1 : return
1143 1 : }
1144 1 : if i.iterValidityState != IterValid {
1145 1 : // If we're going to compare against the prev key, we need to save the
1146 1 : // current key.
1147 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKV.K.UserKey...)
1148 1 : i.key = i.keyBuf
1149 1 : }
1150 1 : for {
1151 1 : i.iterKV = i.iter.Prev()
1152 1 : i.stats.ReverseStepCount[InternalIterCall]++
1153 1 : if i.iterKV == nil {
1154 1 : if err := i.iter.Error(); err != nil {
1155 0 : i.err = err
1156 0 : i.iterValidityState = IterExhausted
1157 0 : }
1158 1 : break
1159 : }
1160 1 : if !i.equal(i.key, i.iterKV.K.UserKey) {
1161 1 : break
1162 : }
1163 : }
1164 : }
1165 :
1166 1 : func (i *Iterator) mergeNext(key InternalKey, valueMerger ValueMerger) {
1167 1 : // Save the current key.
1168 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1169 1 : i.key = i.keyBuf
1170 1 :
1171 1 : // Loop looking for older values for this key and merging them.
1172 1 : for {
1173 1 : i.iterKV = i.iter.Next()
1174 1 : i.stats.ForwardStepCount[InternalIterCall]++
1175 1 : if i.iterKV == nil {
1176 1 : if i.err = i.iter.Error(); i.err != nil {
1177 0 : return
1178 0 : }
1179 1 : i.pos = iterPosNext
1180 1 : return
1181 : }
1182 1 : key = i.iterKV.K
1183 1 : if !i.equal(i.key, key.UserKey) {
1184 1 : // We've advanced to the next key.
1185 1 : i.pos = iterPosNext
1186 1 : return
1187 1 : }
1188 1 : switch key.Kind() {
1189 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
1190 1 : // We've hit a deletion tombstone. Return everything up to this
1191 1 : // point.
1192 1 : //
1193 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
1194 1 : // only simpler, but is also necessary for correctness due to
1195 1 : // InternalKeyKindSSTableInternalObsoleteBit.
1196 1 : return
1197 :
1198 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1199 1 : // We've hit a Set value. Merge with the existing value and return.
1200 1 : var iterValue []byte
1201 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1202 1 : if i.err != nil {
1203 0 : return
1204 0 : }
1205 1 : i.err = valueMerger.MergeOlder(iterValue)
1206 1 : return
1207 :
1208 1 : case InternalKeyKindMerge:
1209 1 : // We've hit another Merge value. Merge with the existing value and
1210 1 : // continue looping.
1211 1 : var iterValue []byte
1212 1 : iterValue, _, i.err = i.iterKV.Value(nil)
1213 1 : if i.err != nil {
1214 0 : return
1215 0 : }
1216 1 : i.err = valueMerger.MergeOlder(iterValue)
1217 1 : if i.err != nil {
1218 0 : return
1219 0 : }
1220 1 : continue
1221 :
1222 0 : case InternalKeyKindRangeKeySet:
1223 0 : // The RANGEKEYSET marker must sort before a MERGE at the same user key.
1224 0 : i.err = base.CorruptionErrorf("pebble: out of order range key marker")
1225 0 : return
1226 :
1227 0 : default:
1228 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1229 0 : return
1230 : }
1231 : }
1232 : }
1233 :
1234 : // SeekGE moves the iterator to the first key/value pair whose key is greater
1235 : // than or equal to the given key. Returns true if the iterator is pointing at
1236 : // a valid entry and false otherwise.
1237 1 : func (i *Iterator) SeekGE(key []byte) bool {
1238 1 : return i.SeekGEWithLimit(key, nil) == IterValid
1239 1 : }
1240 :
1241 : // SeekGEWithLimit moves the iterator to the first key/value pair whose key is
1242 : // greater than or equal to the given key.
1243 : //
1244 : // If limit is provided, it serves as a best-effort exclusive limit. If the
1245 : // first key greater than or equal to the given search key is also greater than
1246 : // or equal to limit, the Iterator may pause and return IterAtLimit. Because
1247 : // limits are best-effort, SeekGEWithLimit may return a key beyond limit.
1248 : //
1249 : // If the Iterator is configured to iterate over range keys, SeekGEWithLimit
1250 : // guarantees it will surface any range keys with bounds overlapping the
1251 : // keyspace [key, limit).
1252 1 : func (i *Iterator) SeekGEWithLimit(key []byte, limit []byte) IterValidityState {
1253 1 : if i.rangeKey != nil {
1254 1 : // NB: Check Valid() before clearing requiresReposition.
1255 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1256 1 : // If we have a range key but did not expose it at the previous iterator
1257 1 : // position (because the iterator was not at a valid position), updated
1258 1 : // must be true. This ensures that after an iterator op sequence like:
1259 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1260 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1261 1 : // - SeekGE(...) → (IterValid, RangeBounds() = [a,b))
1262 1 : // the iterator returns RangeKeyChanged()=true.
1263 1 : //
1264 1 : // The remainder of this function will only update i.rangeKey.updated if
1265 1 : // the iterator moves into a new range key, or out of the current range
1266 1 : // key.
1267 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1268 1 : }
1269 1 : lastPositioningOp := i.lastPositioningOp
1270 1 : // Set it to unknown, since this operation may not succeed, in which case
1271 1 : // the SeekGE following this should not make any assumption about iterator
1272 1 : // position.
1273 1 : i.lastPositioningOp = unknownLastPositionOp
1274 1 : i.requiresReposition = false
1275 1 : i.err = nil // clear cached iteration error
1276 1 : i.hasPrefix = false
1277 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1278 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1279 1 : key = lowerBound
1280 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1281 1 : key = upperBound
1282 1 : }
1283 1 : seekInternalIter := true
1284 1 :
1285 1 : var flags base.SeekGEFlags
1286 1 : if i.batchJustRefreshed {
1287 0 : i.batchJustRefreshed = false
1288 0 : flags = flags.EnableBatchJustRefreshed()
1289 0 : }
1290 1 : if lastPositioningOp == seekGELastPositioningOp {
1291 1 : cmp := i.cmp(i.prefixOrFullSeekKey, key)
1292 1 : // If this seek is to the same or later key, and the iterator is
1293 1 : // already positioned there, this is a noop. This can be helpful for
1294 1 : // sparse key spaces that have many deleted keys, where one can avoid
1295 1 : // the overhead of iterating past them again and again.
1296 1 : if cmp <= 0 {
1297 1 : if !flags.BatchJustRefreshed() &&
1298 1 : (i.iterValidityState == IterExhausted ||
1299 1 : (i.iterValidityState == IterValid && i.cmp(key, i.key) <= 0 &&
1300 1 : (limit == nil || i.cmp(i.key, limit) < 0))) {
1301 1 : // Noop
1302 1 : if !invariants.Enabled || !disableSeekOpt(key, uintptr(unsafe.Pointer(i))) || i.forceEnableSeekOpt {
1303 1 : i.lastPositioningOp = seekGELastPositioningOp
1304 1 : return i.iterValidityState
1305 1 : }
1306 : }
1307 : // cmp == 0 is not safe to optimize since
1308 : // - i.pos could be at iterPosNext, due to a merge.
1309 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1310 : // SET pair for a key, and the iterator would have moved past DELETE
1311 : // but stayed at iterPosCurForward. A similar situation occurs for a
1312 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1313 : // at the SET.
1314 : // We also leverage the IterAtLimit <=> i.pos invariant defined in the
1315 : // comment on iterValidityState, to exclude any cases where i.pos
1316 : // is iterPosCur{Forward,Reverse}Paused. This avoids the need to
1317 : // special-case those iterator positions and their interactions with
1318 : // TrySeekUsingNext, as the main uses for TrySeekUsingNext in CockroachDB
1319 : // do not use limited Seeks in the first place.
1320 1 : if cmp < 0 && i.iterValidityState != IterAtLimit && limit == nil {
1321 1 : flags = flags.EnableTrySeekUsingNext()
1322 1 : }
1323 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !i.forceEnableSeekOpt && disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1324 1 : flags = flags.DisableTrySeekUsingNext()
1325 1 : }
1326 1 : if !flags.BatchJustRefreshed() && i.pos == iterPosCurForwardPaused && i.cmp(key, i.iterKV.K.UserKey) <= 0 {
1327 1 : // Have some work to do, but don't need to seek, and we can
1328 1 : // start doing findNextEntry from i.iterKey.
1329 1 : seekInternalIter = false
1330 1 : }
1331 : }
1332 : }
1333 1 : if seekInternalIter {
1334 1 : i.iterKV = i.iter.SeekGE(key, flags)
1335 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1336 1 : if err := i.iter.Error(); err != nil {
1337 0 : i.err = err
1338 0 : i.iterValidityState = IterExhausted
1339 0 : return i.iterValidityState
1340 0 : }
1341 : }
1342 1 : i.findNextEntry(limit)
1343 1 : i.maybeSampleRead()
1344 1 : if i.Error() == nil {
1345 1 : // Prepare state for a future noop optimization.
1346 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1347 1 : i.lastPositioningOp = seekGELastPositioningOp
1348 1 : }
1349 1 : return i.iterValidityState
1350 : }
1351 :
1352 : // SeekPrefixGE moves the iterator to the first key/value pair whose key is
1353 : // greater than or equal to the given key and which has the same "prefix" as
1354 : // the given key. The prefix for a key is determined by the user-defined
1355 : // Comparer.Split function. The iterator will not observe keys not matching the
1356 : // "prefix" of the search key. Calling SeekPrefixGE puts the iterator in prefix
1357 : // iteration mode. The iterator remains in prefix iteration until a subsequent
1358 : // call to another absolute positioning method (SeekGE, SeekLT, First,
1359 : // Last). Reverse iteration (Prev) is not supported when an iterator is in
1360 : // prefix iteration mode. Returns true if the iterator is pointing at a valid
1361 : // entry and false otherwise.
1362 : //
1363 : // The semantics of SeekPrefixGE are slightly unusual and designed for
1364 : // iteration to be able to take advantage of bloom filters that have been
1365 : // created on the "prefix". If you're not using bloom filters, there is no
1366 : // reason to use SeekPrefixGE.
1367 : //
1368 : // An example Split function may separate a timestamp suffix from the prefix of
1369 : // the key.
1370 : //
1371 : // Split(<key>@<timestamp>) -> <key>
1372 : //
1373 : // Consider the keys "a@1", "a@2", "aa@3", "aa@4". The prefixes for these keys
1374 : // are "a", and "aa". Note that despite "a" and "aa" sharing a prefix by the
1375 : // usual definition, those prefixes differ by the definition of the Split
1376 : // function. To see how this works, consider the following set of calls on this
1377 : // data set:
1378 : //
1379 : // SeekPrefixGE("a@0") -> "a@1"
1380 : // Next() -> "a@2"
1381 : // Next() -> EOF
1382 : //
1383 : // If you're just looking to iterate over keys with a shared prefix, as
1384 : // defined by the configured comparer, set iterator bounds instead:
1385 : //
1386 : // iter := db.NewIter(&pebble.IterOptions{
1387 : // LowerBound: []byte("prefix"),
1388 : // UpperBound: []byte("prefiy"),
1389 : // })
1390 : // for iter.First(); iter.Valid(); iter.Next() {
1391 : // // Only keys beginning with "prefix" will be visited.
1392 : // }
1393 : //
1394 : // See ExampleIterator_SeekPrefixGE for a working example.
1395 : //
1396 : // When iterating with range keys enabled, all range keys encountered are
1397 : // truncated to the seek key's prefix's bounds. The truncation of the upper
1398 : // bound requires that the database's Comparer is configured with a
1399 : // ImmediateSuccessor method. For example, a SeekPrefixGE("a@9") call with the
1400 : // prefix "a" will truncate range key bounds to [a,ImmediateSuccessor(a)].
1401 1 : func (i *Iterator) SeekPrefixGE(key []byte) bool {
1402 1 : if i.rangeKey != nil {
1403 1 : // NB: Check Valid() before clearing requiresReposition.
1404 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1405 1 : // If we have a range key but did not expose it at the previous iterator
1406 1 : // position (because the iterator was not at a valid position), updated
1407 1 : // must be true. This ensures that after an iterator op sequence like:
1408 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1409 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1410 1 : // - SeekPrefixGE(...) → (IterValid, RangeBounds() = [a,b))
1411 1 : // the iterator returns RangeKeyChanged()=true.
1412 1 : //
1413 1 : // The remainder of this function will only update i.rangeKey.updated if
1414 1 : // the iterator moves into a new range key, or out of the current range
1415 1 : // key.
1416 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1417 1 : }
1418 1 : lastPositioningOp := i.lastPositioningOp
1419 1 : // Set it to unknown, since this operation may not succeed, in which case
1420 1 : // the SeekPrefixGE following this should not make any assumption about
1421 1 : // iterator position.
1422 1 : i.lastPositioningOp = unknownLastPositionOp
1423 1 : i.requiresReposition = false
1424 1 : i.err = nil // clear cached iteration error
1425 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1426 1 : if i.comparer.ImmediateSuccessor == nil && i.opts.KeyTypes != IterKeyTypePointsOnly {
1427 0 : panic("pebble: ImmediateSuccessor must be provided for SeekPrefixGE with range keys")
1428 : }
1429 1 : prefixLen := i.split(key)
1430 1 : keyPrefix := key[:prefixLen]
1431 1 : var flags base.SeekGEFlags
1432 1 : if i.batchJustRefreshed {
1433 0 : flags = flags.EnableBatchJustRefreshed()
1434 0 : i.batchJustRefreshed = false
1435 0 : }
1436 1 : if lastPositioningOp == seekPrefixGELastPositioningOp {
1437 1 : if !i.hasPrefix {
1438 0 : panic("lastPositioningOpsIsSeekPrefixGE is true, but hasPrefix is false")
1439 : }
1440 : // The iterator has not been repositioned after the last SeekPrefixGE.
1441 : // See if we are seeking to a larger key, since then we can optimize
1442 : // the seek by using next. Note that we could also optimize if Next
1443 : // has been called, if the iterator is not exhausted and the current
1444 : // position is <= the seek key. We are keeping this limited for now
1445 : // since such optimizations require care for correctness, and to not
1446 : // become de-optimizations (if one usually has to do all the next
1447 : // calls and then the seek). This SeekPrefixGE optimization
1448 : // specifically benefits CockroachDB.
1449 1 : cmp := i.cmp(i.prefixOrFullSeekKey, keyPrefix)
1450 1 : // cmp == 0 is not safe to optimize since
1451 1 : // - i.pos could be at iterPosNext, due to a merge.
1452 1 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1453 1 : // SET pair for a key, and the iterator would have moved past DELETE
1454 1 : // but stayed at iterPosCurForward. A similar situation occurs for a
1455 1 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1456 1 : // at the SET.
1457 1 : // In general some versions of i.prefix could have been consumed by
1458 1 : // the iterator, so we only optimize for cmp < 0.
1459 1 : if cmp < 0 {
1460 1 : flags = flags.EnableTrySeekUsingNext()
1461 1 : }
1462 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !i.forceEnableSeekOpt && disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1463 1 : flags = flags.DisableTrySeekUsingNext()
1464 1 : }
1465 : }
1466 : // Make a copy of the prefix so that modifications to the key after
1467 : // SeekPrefixGE returns does not affect the stored prefix.
1468 1 : if cap(i.prefixOrFullSeekKey) < prefixLen {
1469 1 : i.prefixOrFullSeekKey = make([]byte, prefixLen)
1470 1 : } else {
1471 1 : i.prefixOrFullSeekKey = i.prefixOrFullSeekKey[:prefixLen]
1472 1 : }
1473 1 : i.hasPrefix = true
1474 1 : copy(i.prefixOrFullSeekKey, keyPrefix)
1475 1 :
1476 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1477 1 : if n := i.split(lowerBound); !bytes.Equal(i.prefixOrFullSeekKey, lowerBound[:n]) {
1478 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of lower bound")
1479 1 : i.iterValidityState = IterExhausted
1480 1 : return false
1481 1 : }
1482 1 : key = lowerBound
1483 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1484 1 : if n := i.split(upperBound); !bytes.Equal(i.prefixOrFullSeekKey, upperBound[:n]) {
1485 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of upper bound")
1486 1 : i.iterValidityState = IterExhausted
1487 1 : return false
1488 1 : }
1489 1 : key = upperBound
1490 : }
1491 1 : i.iterKV = i.iter.SeekPrefixGE(i.prefixOrFullSeekKey, key, flags)
1492 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1493 1 : i.findNextEntry(nil)
1494 1 : i.maybeSampleRead()
1495 1 : if i.Error() == nil {
1496 1 : i.lastPositioningOp = seekPrefixGELastPositioningOp
1497 1 : }
1498 1 : return i.iterValidityState == IterValid
1499 : }
1500 :
1501 : // Deterministic disabling of the seek optimizations. It uses the iterator
1502 : // pointer, since we want diversity in iterator behavior for the same key. Used
1503 : // for tests.
1504 1 : func disableSeekOpt(key []byte, ptr uintptr) bool {
1505 1 : // Fibonacci hash https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
1506 1 : simpleHash := (11400714819323198485 * uint64(ptr)) >> 63
1507 1 : return key != nil && key[0]&byte(1) == 0 && simpleHash == 0
1508 1 : }
1509 :
1510 : // SeekLT moves the iterator to the last key/value pair whose key is less than
1511 : // the given key. Returns true if the iterator is pointing at a valid entry and
1512 : // false otherwise.
1513 1 : func (i *Iterator) SeekLT(key []byte) bool {
1514 1 : return i.SeekLTWithLimit(key, nil) == IterValid
1515 1 : }
1516 :
1517 : // SeekLTWithLimit moves the iterator to the last key/value pair whose key is
1518 : // less than the given key.
1519 : //
1520 : // If limit is provided, it serves as a best-effort inclusive limit. If the last
1521 : // key less than the given search key is also less than limit, the Iterator may
1522 : // pause and return IterAtLimit. Because limits are best-effort, SeekLTWithLimit
1523 : // may return a key beyond limit.
1524 : //
1525 : // If the Iterator is configured to iterate over range keys, SeekLTWithLimit
1526 : // guarantees it will surface any range keys with bounds overlapping the
1527 : // keyspace up to limit.
1528 1 : func (i *Iterator) SeekLTWithLimit(key []byte, limit []byte) IterValidityState {
1529 1 : if i.rangeKey != nil {
1530 1 : // NB: Check Valid() before clearing requiresReposition.
1531 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1532 1 : // If we have a range key but did not expose it at the previous iterator
1533 1 : // position (because the iterator was not at a valid position), updated
1534 1 : // must be true. This ensures that after an iterator op sequence like:
1535 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1536 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1537 1 : // - SeekLTWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1538 1 : // the iterator returns RangeKeyChanged()=true.
1539 1 : //
1540 1 : // The remainder of this function will only update i.rangeKey.updated if
1541 1 : // the iterator moves into a new range key, or out of the current range
1542 1 : // key.
1543 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1544 1 : }
1545 1 : lastPositioningOp := i.lastPositioningOp
1546 1 : // Set it to unknown, since this operation may not succeed, in which case
1547 1 : // the SeekLT following this should not make any assumption about iterator
1548 1 : // position.
1549 1 : i.lastPositioningOp = unknownLastPositionOp
1550 1 : i.batchJustRefreshed = false
1551 1 : i.requiresReposition = false
1552 1 : i.err = nil // clear cached iteration error
1553 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1554 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1555 1 : key = upperBound
1556 1 : } else if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1557 1 : key = lowerBound
1558 1 : }
1559 1 : i.hasPrefix = false
1560 1 : seekInternalIter := true
1561 1 : // The following noop optimization only applies when i.batch == nil, since
1562 1 : // an iterator over a batch is iterating over mutable data, that may have
1563 1 : // changed since the last seek.
1564 1 : if lastPositioningOp == seekLTLastPositioningOp && i.batch == nil {
1565 1 : cmp := i.cmp(key, i.prefixOrFullSeekKey)
1566 1 : // If this seek is to the same or earlier key, and the iterator is
1567 1 : // already positioned there, this is a noop. This can be helpful for
1568 1 : // sparse key spaces that have many deleted keys, where one can avoid
1569 1 : // the overhead of iterating past them again and again.
1570 1 : if cmp <= 0 {
1571 1 : // NB: when pos != iterPosCurReversePaused, the invariant
1572 1 : // documented earlier implies that iterValidityState !=
1573 1 : // IterAtLimit.
1574 1 : if i.iterValidityState == IterExhausted ||
1575 1 : (i.iterValidityState == IterValid && i.cmp(i.key, key) < 0 &&
1576 1 : (limit == nil || i.cmp(limit, i.key) <= 0)) {
1577 1 : if !invariants.Enabled || !disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1578 1 : i.lastPositioningOp = seekLTLastPositioningOp
1579 1 : return i.iterValidityState
1580 1 : }
1581 : }
1582 1 : if i.pos == iterPosCurReversePaused && i.cmp(i.iterKV.K.UserKey, key) < 0 {
1583 1 : // Have some work to do, but don't need to seek, and we can
1584 1 : // start doing findPrevEntry from i.iterKey.
1585 1 : seekInternalIter = false
1586 1 : }
1587 : }
1588 : }
1589 1 : if seekInternalIter {
1590 1 : i.iterKV = i.iter.SeekLT(key, base.SeekLTFlagsNone)
1591 1 : i.stats.ReverseSeekCount[InternalIterCall]++
1592 1 : if err := i.iter.Error(); err != nil {
1593 0 : i.err = err
1594 0 : i.iterValidityState = IterExhausted
1595 0 : return i.iterValidityState
1596 0 : }
1597 : }
1598 1 : i.findPrevEntry(limit)
1599 1 : i.maybeSampleRead()
1600 1 : if i.Error() == nil && i.batch == nil {
1601 1 : // Prepare state for a future noop optimization.
1602 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1603 1 : i.lastPositioningOp = seekLTLastPositioningOp
1604 1 : }
1605 1 : return i.iterValidityState
1606 : }
1607 :
1608 : // First moves the iterator the first key/value pair. Returns true if the
1609 : // iterator is pointing at a valid entry and false otherwise.
1610 1 : func (i *Iterator) First() bool {
1611 1 : if i.rangeKey != nil {
1612 1 : // NB: Check Valid() before clearing requiresReposition.
1613 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1614 1 : // If we have a range key but did not expose it at the previous iterator
1615 1 : // position (because the iterator was not at a valid position), updated
1616 1 : // must be true. This ensures that after an iterator op sequence like:
1617 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1618 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1619 1 : // - First(...) → (IterValid, RangeBounds() = [a,b))
1620 1 : // the iterator returns RangeKeyChanged()=true.
1621 1 : //
1622 1 : // The remainder of this function will only update i.rangeKey.updated if
1623 1 : // the iterator moves into a new range key, or out of the current range
1624 1 : // key.
1625 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1626 1 : }
1627 1 : i.err = nil // clear cached iteration error
1628 1 : i.hasPrefix = false
1629 1 : i.batchJustRefreshed = false
1630 1 : i.lastPositioningOp = unknownLastPositionOp
1631 1 : i.requiresReposition = false
1632 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1633 1 :
1634 1 : i.iterFirstWithinBounds()
1635 1 : if i.err != nil {
1636 0 : i.iterValidityState = IterExhausted
1637 0 : return false
1638 0 : }
1639 1 : i.findNextEntry(nil)
1640 1 : i.maybeSampleRead()
1641 1 : return i.iterValidityState == IterValid
1642 : }
1643 :
1644 : // Last moves the iterator the last key/value pair. Returns true if the
1645 : // iterator is pointing at a valid entry and false otherwise.
1646 1 : func (i *Iterator) Last() bool {
1647 1 : if i.rangeKey != nil {
1648 1 : // NB: Check Valid() before clearing requiresReposition.
1649 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1650 1 : // If we have a range key but did not expose it at the previous iterator
1651 1 : // position (because the iterator was not at a valid position), updated
1652 1 : // must be true. This ensures that after an iterator op sequence like:
1653 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1654 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1655 1 : // - Last(...) → (IterValid, RangeBounds() = [a,b))
1656 1 : // the iterator returns RangeKeyChanged()=true.
1657 1 : //
1658 1 : // The remainder of this function will only update i.rangeKey.updated if
1659 1 : // the iterator moves into a new range key, or out of the current range
1660 1 : // key.
1661 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1662 1 : }
1663 1 : i.err = nil // clear cached iteration error
1664 1 : i.hasPrefix = false
1665 1 : i.batchJustRefreshed = false
1666 1 : i.lastPositioningOp = unknownLastPositionOp
1667 1 : i.requiresReposition = false
1668 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1669 1 :
1670 1 : i.iterLastWithinBounds()
1671 1 : if i.err != nil {
1672 0 : i.iterValidityState = IterExhausted
1673 0 : return false
1674 0 : }
1675 1 : i.findPrevEntry(nil)
1676 1 : i.maybeSampleRead()
1677 1 : return i.iterValidityState == IterValid
1678 : }
1679 :
1680 : // Next moves the iterator to the next key/value pair. Returns true if the
1681 : // iterator is pointing at a valid entry and false otherwise.
1682 1 : func (i *Iterator) Next() bool {
1683 1 : return i.nextWithLimit(nil) == IterValid
1684 1 : }
1685 :
1686 : // NextWithLimit moves the iterator to the next key/value pair.
1687 : //
1688 : // If limit is provided, it serves as a best-effort exclusive limit. If the next
1689 : // key is greater than or equal to limit, the Iterator may pause and return
1690 : // IterAtLimit. Because limits are best-effort, NextWithLimit may return a key
1691 : // beyond limit.
1692 : //
1693 : // If the Iterator is configured to iterate over range keys, NextWithLimit
1694 : // guarantees it will surface any range keys with bounds overlapping the
1695 : // keyspace up to limit.
1696 1 : func (i *Iterator) NextWithLimit(limit []byte) IterValidityState {
1697 1 : return i.nextWithLimit(limit)
1698 1 : }
1699 :
1700 : // NextPrefix moves the iterator to the next key/value pair with a key
1701 : // containing a different prefix than the current key. Prefixes are determined
1702 : // by Comparer.Split. Exhausts the iterator if invoked while in prefix-iteration
1703 : // mode.
1704 : //
1705 : // It is not permitted to invoke NextPrefix while at a IterAtLimit position.
1706 : // When called in this condition, NextPrefix has non-deterministic behavior.
1707 : //
1708 : // It is not permitted to invoke NextPrefix when the Iterator has an
1709 : // upper-bound that is a versioned MVCC key (see the comment for
1710 : // Comparer.Split). It returns an error in this case.
1711 1 : func (i *Iterator) NextPrefix() bool {
1712 1 : if i.nextPrefixNotPermittedByUpperBound {
1713 1 : i.lastPositioningOp = unknownLastPositionOp
1714 1 : i.requiresReposition = false
1715 1 : i.err = errors.Errorf("NextPrefix not permitted with upper bound %s",
1716 1 : i.comparer.FormatKey(i.opts.UpperBound))
1717 1 : i.iterValidityState = IterExhausted
1718 1 : return false
1719 1 : }
1720 1 : if i.hasPrefix {
1721 1 : i.iterValidityState = IterExhausted
1722 1 : return false
1723 1 : }
1724 1 : return i.nextPrefix() == IterValid
1725 : }
1726 :
1727 1 : func (i *Iterator) nextPrefix() IterValidityState {
1728 1 : if i.rangeKey != nil {
1729 1 : // NB: Check Valid() before clearing requiresReposition.
1730 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1731 1 : // If we have a range key but did not expose it at the previous iterator
1732 1 : // position (because the iterator was not at a valid position), updated
1733 1 : // must be true. This ensures that after an iterator op sequence like:
1734 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1735 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1736 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1737 1 : // the iterator returns RangeKeyChanged()=true.
1738 1 : //
1739 1 : // The remainder of this function will only update i.rangeKey.updated if
1740 1 : // the iterator moves into a new range key, or out of the current range
1741 1 : // key.
1742 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1743 1 : }
1744 :
1745 : // Although NextPrefix documents that behavior at IterAtLimit is undefined,
1746 : // this function handles these cases as a simple prefix-agnostic Next. This
1747 : // is done for deterministic behavior in the metamorphic tests.
1748 : //
1749 : // TODO(jackson): If the metamorphic test operation generator is adjusted to
1750 : // make generation of some operations conditional on the previous
1751 : // operations, then we can remove this behavior and explicitly error.
1752 :
1753 1 : i.lastPositioningOp = unknownLastPositionOp
1754 1 : i.requiresReposition = false
1755 1 : switch i.pos {
1756 1 : case iterPosCurForward:
1757 1 : // Positioned on the current key. Advance to the next prefix.
1758 1 : i.internalNextPrefix(i.split(i.key))
1759 1 : case iterPosCurForwardPaused:
1760 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1761 : // up above. The iterator is already positioned at the next key.
1762 1 : case iterPosCurReverse:
1763 1 : // Switching directions.
1764 1 : // Unless the iterator was exhausted, reverse iteration needs to
1765 1 : // position the iterator at iterPosPrev.
1766 1 : if i.iterKV != nil {
1767 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1768 0 : i.iterValidityState = IterExhausted
1769 0 : return i.iterValidityState
1770 0 : }
1771 : // The Iterator is exhausted and i.iter is positioned before the first
1772 : // key. Reposition to point to the first internal key.
1773 1 : i.iterFirstWithinBounds()
1774 1 : case iterPosCurReversePaused:
1775 1 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1776 1 : // up above.
1777 1 : //
1778 1 : // Switching directions; The iterator must not be exhausted since it
1779 1 : // paused.
1780 1 : if i.iterKV == nil {
1781 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1782 0 : i.iterValidityState = IterExhausted
1783 0 : return i.iterValidityState
1784 0 : }
1785 1 : i.nextUserKey()
1786 1 : case iterPosPrev:
1787 1 : // The underlying iterator is pointed to the previous key (this can
1788 1 : // only happen when switching iteration directions).
1789 1 : if i.iterKV == nil {
1790 1 : // We're positioned before the first key. Need to reposition to point to
1791 1 : // the first key.
1792 1 : i.iterFirstWithinBounds()
1793 1 : } else {
1794 1 : // Move the internal iterator back onto the user key stored in
1795 1 : // i.key. iterPosPrev guarantees that it's positioned at the last
1796 1 : // key with the user key less than i.key, so we're guaranteed to
1797 1 : // land on the correct key with a single Next.
1798 1 : i.iterKV = i.iter.Next()
1799 1 : if i.iterKV == nil {
1800 0 : // This should only be possible if i.iter.Next() encountered an
1801 0 : // error.
1802 0 : if i.iter.Error() == nil {
1803 0 : i.opts.logger.Fatalf("pebble: invariant violation: Nexting internal iterator from iterPosPrev found nothing")
1804 0 : }
1805 : // NB: Iterator.Error() will return i.iter.Error().
1806 0 : i.iterValidityState = IterExhausted
1807 0 : return i.iterValidityState
1808 : }
1809 1 : if invariants.Enabled && !i.equal(i.iterKV.K.UserKey, i.key) {
1810 0 : i.opts.logger.Fatalf("pebble: invariant violation: Nexting internal iterator from iterPosPrev landed on %q, not %q",
1811 0 : i.iterKV.K.UserKey, i.key)
1812 0 : }
1813 : }
1814 : // The internal iterator is now positioned at i.key. Advance to the next
1815 : // prefix.
1816 1 : i.internalNextPrefix(i.split(i.key))
1817 1 : case iterPosNext:
1818 1 : // Already positioned on the next key. Only call nextPrefixKey if the
1819 1 : // next key shares the same prefix.
1820 1 : if i.iterKV != nil {
1821 1 : currKeyPrefixLen := i.split(i.key)
1822 1 : iterKeyPrefixLen := i.split(i.iterKV.K.UserKey)
1823 1 : if bytes.Equal(i.iterKV.K.UserKey[:iterKeyPrefixLen], i.key[:currKeyPrefixLen]) {
1824 1 : i.internalNextPrefix(currKeyPrefixLen)
1825 1 : }
1826 : }
1827 : }
1828 :
1829 1 : i.stats.ForwardStepCount[InterfaceCall]++
1830 1 : i.findNextEntry(nil /* limit */)
1831 1 : i.maybeSampleRead()
1832 1 : return i.iterValidityState
1833 : }
1834 :
1835 1 : func (i *Iterator) internalNextPrefix(currKeyPrefixLen int) {
1836 1 : if i.iterKV == nil {
1837 1 : return
1838 1 : }
1839 : // The Next "fast-path" is not really a fast-path when there is more than
1840 : // one version. However, even with TableFormatPebblev3, there is a small
1841 : // slowdown (~10%) for one version if we remove it and only call NextPrefix.
1842 : // When there are two versions, only calling NextPrefix is ~30% faster.
1843 1 : i.stats.ForwardStepCount[InternalIterCall]++
1844 1 : if i.iterKV = i.iter.Next(); i.iterKV == nil {
1845 1 : return
1846 1 : }
1847 1 : iterKeyPrefixLen := i.split(i.iterKV.K.UserKey)
1848 1 : if !bytes.Equal(i.iterKV.K.UserKey[:iterKeyPrefixLen], i.key[:currKeyPrefixLen]) {
1849 1 : return
1850 1 : }
1851 1 : i.stats.ForwardStepCount[InternalIterCall]++
1852 1 : i.prefixOrFullSeekKey = i.comparer.ImmediateSuccessor(i.prefixOrFullSeekKey[:0], i.key[:currKeyPrefixLen])
1853 1 : i.iterKV = i.iter.NextPrefix(i.prefixOrFullSeekKey)
1854 1 : if invariants.Enabled && i.iterKV != nil {
1855 1 : if iterKeyPrefixLen := i.split(i.iterKV.K.UserKey); i.cmp(i.iterKV.K.UserKey[:iterKeyPrefixLen], i.prefixOrFullSeekKey) < 0 {
1856 0 : panic(errors.AssertionFailedf("pebble: iter.NextPrefix did not advance beyond the current prefix: now at %q; expected to be geq %q",
1857 0 : i.iterKV.K, i.prefixOrFullSeekKey))
1858 : }
1859 : }
1860 : }
1861 :
1862 1 : func (i *Iterator) nextWithLimit(limit []byte) IterValidityState {
1863 1 : i.stats.ForwardStepCount[InterfaceCall]++
1864 1 : if i.hasPrefix {
1865 1 : if limit != nil {
1866 1 : i.err = errors.New("cannot use limit with prefix iteration")
1867 1 : i.iterValidityState = IterExhausted
1868 1 : return i.iterValidityState
1869 1 : } else if i.iterValidityState == IterExhausted {
1870 1 : // No-op, already exhausted. We avoid executing the Next because it
1871 1 : // can break invariants: Specifically, a file that fails the bloom
1872 1 : // filter test may result in its level being removed from the
1873 1 : // merging iterator. The level's removal can cause a lazy combined
1874 1 : // iterator to miss range keys and trigger a switch to combined
1875 1 : // iteration at a larger key, breaking keyspan invariants.
1876 1 : return i.iterValidityState
1877 1 : }
1878 : }
1879 1 : if i.err != nil {
1880 1 : return i.iterValidityState
1881 1 : }
1882 1 : if i.rangeKey != nil {
1883 1 : // NB: Check Valid() before clearing requiresReposition.
1884 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1885 1 : // If we have a range key but did not expose it at the previous iterator
1886 1 : // position (because the iterator was not at a valid position), updated
1887 1 : // must be true. This ensures that after an iterator op sequence like:
1888 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1889 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1890 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1891 1 : // the iterator returns RangeKeyChanged()=true.
1892 1 : //
1893 1 : // The remainder of this function will only update i.rangeKey.updated if
1894 1 : // the iterator moves into a new range key, or out of the current range
1895 1 : // key.
1896 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1897 1 : }
1898 1 : i.lastPositioningOp = unknownLastPositionOp
1899 1 : i.requiresReposition = false
1900 1 : switch i.pos {
1901 1 : case iterPosCurForward:
1902 1 : i.nextUserKey()
1903 1 : case iterPosCurForwardPaused:
1904 : // Already at the right place.
1905 1 : case iterPosCurReverse:
1906 1 : // Switching directions.
1907 1 : // Unless the iterator was exhausted, reverse iteration needs to
1908 1 : // position the iterator at iterPosPrev.
1909 1 : if i.iterKV != nil {
1910 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1911 0 : i.iterValidityState = IterExhausted
1912 0 : return i.iterValidityState
1913 0 : }
1914 : // We're positioned before the first key. Need to reposition to point to
1915 : // the first key.
1916 1 : i.iterFirstWithinBounds()
1917 1 : case iterPosCurReversePaused:
1918 1 : // Switching directions.
1919 1 : // The iterator must not be exhausted since it paused.
1920 1 : if i.iterKV == nil {
1921 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1922 0 : i.iterValidityState = IterExhausted
1923 0 : return i.iterValidityState
1924 0 : }
1925 1 : i.nextUserKey()
1926 1 : case iterPosPrev:
1927 1 : // The underlying iterator is pointed to the previous key (this can
1928 1 : // only happen when switching iteration directions). We set
1929 1 : // i.iterValidityState to IterExhausted here to force the calls to
1930 1 : // nextUserKey to save the current key i.iter is pointing at in order
1931 1 : // to determine when the next user-key is reached.
1932 1 : i.iterValidityState = IterExhausted
1933 1 : if i.iterKV == nil {
1934 1 : // We're positioned before the first key. Need to reposition to point to
1935 1 : // the first key.
1936 1 : i.iterFirstWithinBounds()
1937 1 : } else {
1938 1 : i.nextUserKey()
1939 1 : }
1940 1 : i.nextUserKey()
1941 1 : case iterPosNext:
1942 : // Already at the right place.
1943 : }
1944 1 : i.findNextEntry(limit)
1945 1 : i.maybeSampleRead()
1946 1 : return i.iterValidityState
1947 : }
1948 :
1949 : // Prev moves the iterator to the previous key/value pair. Returns true if the
1950 : // iterator is pointing at a valid entry and false otherwise.
1951 1 : func (i *Iterator) Prev() bool {
1952 1 : return i.PrevWithLimit(nil) == IterValid
1953 1 : }
1954 :
1955 : // PrevWithLimit moves the iterator to the previous key/value pair.
1956 : //
1957 : // If limit is provided, it serves as a best-effort inclusive limit. If the
1958 : // previous key is less than limit, the Iterator may pause and return
1959 : // IterAtLimit. Because limits are best-effort, PrevWithLimit may return a key
1960 : // beyond limit.
1961 : //
1962 : // If the Iterator is configured to iterate over range keys, PrevWithLimit
1963 : // guarantees it will surface any range keys with bounds overlapping the
1964 : // keyspace up to limit.
1965 1 : func (i *Iterator) PrevWithLimit(limit []byte) IterValidityState {
1966 1 : i.stats.ReverseStepCount[InterfaceCall]++
1967 1 : if i.err != nil {
1968 1 : return i.iterValidityState
1969 1 : }
1970 1 : if i.rangeKey != nil {
1971 1 : // NB: Check Valid() before clearing requiresReposition.
1972 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1973 1 : // If we have a range key but did not expose it at the previous iterator
1974 1 : // position (because the iterator was not at a valid position), updated
1975 1 : // must be true. This ensures that after an iterator op sequence like:
1976 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1977 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1978 1 : // - PrevWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1979 1 : // the iterator returns RangeKeyChanged()=true.
1980 1 : //
1981 1 : // The remainder of this function will only update i.rangeKey.updated if
1982 1 : // the iterator moves into a new range key, or out of the current range
1983 1 : // key.
1984 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1985 1 : }
1986 1 : i.lastPositioningOp = unknownLastPositionOp
1987 1 : i.requiresReposition = false
1988 1 : if i.hasPrefix {
1989 1 : i.err = errReversePrefixIteration
1990 1 : i.iterValidityState = IterExhausted
1991 1 : return i.iterValidityState
1992 1 : }
1993 1 : switch i.pos {
1994 1 : case iterPosCurForward:
1995 : // Switching directions, and will handle this below.
1996 1 : case iterPosCurForwardPaused:
1997 : // Switching directions, and will handle this below.
1998 1 : case iterPosCurReverse:
1999 1 : i.prevUserKey()
2000 1 : case iterPosCurReversePaused:
2001 : // Already at the right place.
2002 1 : case iterPosNext:
2003 : // The underlying iterator is pointed to the next key (this can only happen
2004 : // when switching iteration directions). We will handle this below.
2005 1 : case iterPosPrev:
2006 : // Already at the right place.
2007 : }
2008 1 : if i.pos == iterPosCurForward || i.pos == iterPosNext || i.pos == iterPosCurForwardPaused {
2009 1 : // Switching direction.
2010 1 : stepAgain := i.pos == iterPosNext
2011 1 :
2012 1 : // Synthetic range key markers are a special case. Consider SeekGE(b)
2013 1 : // which finds a range key [a, c). To ensure the user observes the range
2014 1 : // key, the Iterator pauses at Key() = b. The iterator must advance the
2015 1 : // internal iterator to see if there's also a coincident point key at
2016 1 : // 'b', leaving the iterator at iterPosNext if there's not.
2017 1 : //
2018 1 : // This is a problem: Synthetic range key markers are only interleaved
2019 1 : // during the original seek. A subsequent Prev() of i.iter will not move
2020 1 : // back onto the synthetic range key marker. In this case where the
2021 1 : // previous iterator position was a synthetic range key start boundary,
2022 1 : // we must not step a second time.
2023 1 : if i.isEphemeralPosition() {
2024 1 : stepAgain = false
2025 1 : }
2026 :
2027 : // We set i.iterValidityState to IterExhausted here to force the calls
2028 : // to prevUserKey to save the current key i.iter is pointing at in
2029 : // order to determine when the prev user-key is reached.
2030 1 : i.iterValidityState = IterExhausted
2031 1 : if i.iterKV == nil {
2032 1 : // We're positioned after the last key. Need to reposition to point to
2033 1 : // the last key.
2034 1 : i.iterLastWithinBounds()
2035 1 : } else {
2036 1 : i.prevUserKey()
2037 1 : if i.err != nil {
2038 0 : return i.iterValidityState
2039 0 : }
2040 : }
2041 1 : if stepAgain {
2042 1 : i.prevUserKey()
2043 1 : if i.err != nil {
2044 0 : return i.iterValidityState
2045 0 : }
2046 : }
2047 : }
2048 1 : i.findPrevEntry(limit)
2049 1 : i.maybeSampleRead()
2050 1 : return i.iterValidityState
2051 : }
2052 :
2053 : // iterFirstWithinBounds moves the internal iterator to the first key,
2054 : // respecting bounds.
2055 1 : func (i *Iterator) iterFirstWithinBounds() {
2056 1 : i.stats.ForwardSeekCount[InternalIterCall]++
2057 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil {
2058 1 : i.iterKV = i.iter.SeekGE(lowerBound, base.SeekGEFlagsNone)
2059 1 : } else {
2060 1 : i.iterKV = i.iter.First()
2061 1 : }
2062 : }
2063 :
2064 : // iterLastWithinBounds moves the internal iterator to the last key, respecting
2065 : // bounds.
2066 1 : func (i *Iterator) iterLastWithinBounds() {
2067 1 : i.stats.ReverseSeekCount[InternalIterCall]++
2068 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil {
2069 1 : i.iterKV = i.iter.SeekLT(upperBound, base.SeekLTFlagsNone)
2070 1 : } else {
2071 1 : i.iterKV = i.iter.Last()
2072 1 : }
2073 : }
2074 :
2075 : // RangeKeyData describes a range key's data, set through RangeKeySet. The key
2076 : // boundaries of the range key is provided by Iterator.RangeBounds.
2077 : type RangeKeyData struct {
2078 : Suffix []byte
2079 : Value []byte
2080 : }
2081 :
2082 : // rangeKeyWithinLimit is called during limited reverse iteration when
2083 : // positioned over a key beyond the limit. If there exists a range key that lies
2084 : // within the limit, the iterator must not pause in order to ensure the user has
2085 : // an opportunity to observe the range key within limit.
2086 : //
2087 : // It would be valid to ignore the limit whenever there's a range key covering
2088 : // the key, but that would introduce nondeterminism. To preserve determinism for
2089 : // testing, the iterator ignores the limit only if the covering range key does
2090 : // cover the keyspace within the limit.
2091 : //
2092 : // This awkwardness exists because range keys are interleaved at their inclusive
2093 : // start positions. Note that limit is inclusive.
2094 1 : func (i *Iterator) rangeKeyWithinLimit(limit []byte) bool {
2095 1 : if i.rangeKey == nil || !i.opts.rangeKeys() {
2096 1 : return false
2097 1 : }
2098 1 : s := i.rangeKey.iiter.Span()
2099 1 : // If the range key ends beyond the limit, then the range key does not cover
2100 1 : // any portion of the keyspace within the limit and it is safe to pause.
2101 1 : return s != nil && i.cmp(s.End, limit) > 0
2102 : }
2103 :
2104 : // saveRangeKey saves the current range key to the underlying iterator's current
2105 : // range key state. If the range key has not changed, saveRangeKey is a no-op.
2106 : // If there is a new range key, saveRangeKey copies all of the key, value and
2107 : // suffixes into Iterator-managed buffers.
2108 1 : func (i *Iterator) saveRangeKey() {
2109 1 : if i.rangeKey == nil || i.opts.KeyTypes == IterKeyTypePointsOnly {
2110 1 : return
2111 1 : }
2112 :
2113 1 : s := i.rangeKey.iiter.Span()
2114 1 : if s == nil {
2115 1 : i.rangeKey.hasRangeKey = false
2116 1 : i.rangeKey.updated = i.rangeKey.prevPosHadRangeKey
2117 1 : return
2118 1 : } else if !i.rangeKey.stale {
2119 1 : // The range key `s` is identical to the one currently saved. No-op.
2120 1 : return
2121 1 : }
2122 :
2123 1 : if s.KeysOrder != keyspan.BySuffixAsc {
2124 0 : panic("pebble: range key span's keys unexpectedly not in ascending suffix order")
2125 : }
2126 :
2127 : // Although `i.rangeKey.stale` is true, the span s may still be identical
2128 : // to the currently saved span. This is possible when seeking the iterator,
2129 : // which may land back on the same range key. If we previously had a range
2130 : // key and the new one has an identical start key, then it must be the same
2131 : // range key and we can avoid copying and keep `i.rangeKey.updated=false`.
2132 : //
2133 : // TODO(jackson): These key comparisons could be avoidable during relative
2134 : // positioning operations continuing in the same direction, because these
2135 : // ops will never encounter the previous position's range key while
2136 : // stale=true. However, threading whether the current op is a seek or step
2137 : // maybe isn't worth it. This key comparison is only necessary once when we
2138 : // step onto a new range key, which should be relatively rare.
2139 1 : if i.rangeKey.prevPosHadRangeKey && i.equal(i.rangeKey.start, s.Start) &&
2140 1 : i.equal(i.rangeKey.end, s.End) {
2141 1 : i.rangeKey.updated = false
2142 1 : i.rangeKey.stale = false
2143 1 : i.rangeKey.hasRangeKey = true
2144 1 : return
2145 1 : }
2146 1 : i.stats.RangeKeyStats.Count += len(s.Keys)
2147 1 : i.rangeKey.buf.Reset()
2148 1 : i.rangeKey.hasRangeKey = true
2149 1 : i.rangeKey.updated = true
2150 1 : i.rangeKey.stale = false
2151 1 : i.rangeKey.buf, i.rangeKey.start = i.rangeKey.buf.Copy(s.Start)
2152 1 : i.rangeKey.buf, i.rangeKey.end = i.rangeKey.buf.Copy(s.End)
2153 1 : i.rangeKey.keys = i.rangeKey.keys[:0]
2154 1 : for j := 0; j < len(s.Keys); j++ {
2155 1 : if invariants.Enabled {
2156 1 : if s.Keys[j].Kind() != base.InternalKeyKindRangeKeySet {
2157 0 : panic("pebble: user iteration encountered non-RangeKeySet key kind")
2158 1 : } else if j > 0 && i.cmp(s.Keys[j].Suffix, s.Keys[j-1].Suffix) < 0 {
2159 0 : panic("pebble: user iteration encountered range keys not in suffix order")
2160 : }
2161 : }
2162 1 : var rkd RangeKeyData
2163 1 : i.rangeKey.buf, rkd.Suffix = i.rangeKey.buf.Copy(s.Keys[j].Suffix)
2164 1 : i.rangeKey.buf, rkd.Value = i.rangeKey.buf.Copy(s.Keys[j].Value)
2165 1 : i.rangeKey.keys = append(i.rangeKey.keys, rkd)
2166 : }
2167 : }
2168 :
2169 : // RangeKeyChanged indicates whether the most recent iterator positioning
2170 : // operation resulted in the iterator stepping into or out of a new range key.
2171 : // If true, previously returned range key bounds and data has been invalidated.
2172 : // If false, previously obtained range key bounds, suffix and value slices are
2173 : // still valid and may continue to be read.
2174 : //
2175 : // Invalid iterator positions are considered to not hold range keys, meaning
2176 : // that if an iterator steps from an IterExhausted or IterAtLimit position onto
2177 : // a position with a range key, RangeKeyChanged will yield true.
2178 1 : func (i *Iterator) RangeKeyChanged() bool {
2179 1 : return i.iterValidityState == IterValid && i.rangeKey != nil && i.rangeKey.updated
2180 1 : }
2181 :
2182 : // HasPointAndRange indicates whether there exists a point key, a range key or
2183 : // both at the current iterator position.
2184 1 : func (i *Iterator) HasPointAndRange() (hasPoint, hasRange bool) {
2185 1 : if i.iterValidityState != IterValid || i.requiresReposition {
2186 0 : return false, false
2187 0 : }
2188 1 : if i.opts.KeyTypes == IterKeyTypePointsOnly {
2189 1 : return true, false
2190 1 : }
2191 1 : return i.rangeKey == nil || !i.rangeKey.rangeKeyOnly, i.rangeKey != nil && i.rangeKey.hasRangeKey
2192 : }
2193 :
2194 : // RangeBounds returns the start (inclusive) and end (exclusive) bounds of the
2195 : // range key covering the current iterator position. RangeBounds returns nil
2196 : // bounds if there is no range key covering the current iterator position, or
2197 : // the iterator is not configured to surface range keys.
2198 : //
2199 : // If valid, the returned start bound is less than or equal to Key() and the
2200 : // returned end bound is greater than Key().
2201 1 : func (i *Iterator) RangeBounds() (start, end []byte) {
2202 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2203 0 : return nil, nil
2204 0 : }
2205 1 : return i.rangeKey.start, i.rangeKey.end
2206 : }
2207 :
2208 : // Key returns the key of the current key/value pair, or nil if done. The
2209 : // caller should not modify the contents of the returned slice, and its
2210 : // contents may change on the next call to Next.
2211 : //
2212 : // If positioned at an iterator position that only holds a range key, Key()
2213 : // always returns the start bound of the range key. Otherwise, it returns the
2214 : // point key's key.
2215 1 : func (i *Iterator) Key() []byte {
2216 1 : return i.key
2217 1 : }
2218 :
2219 : // Value returns the value of the current key/value pair, or nil if done. The
2220 : // caller should not modify the contents of the returned slice, and its
2221 : // contents may change on the next call to Next.
2222 : //
2223 : // Only valid if HasPointAndRange() returns true for hasPoint.
2224 : // Deprecated: use ValueAndErr instead.
2225 1 : func (i *Iterator) Value() []byte {
2226 1 : val, _ := i.ValueAndErr()
2227 1 : return val
2228 1 : }
2229 :
2230 : // ValueAndErr returns the value, and any error encountered in extracting the value.
2231 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2232 : //
2233 : // The caller should not modify the contents of the returned slice, and its
2234 : // contents may change on the next call to Next.
2235 1 : func (i *Iterator) ValueAndErr() ([]byte, error) {
2236 1 : val, callerOwned, err := i.value.Value(i.lazyValueBuf)
2237 1 : if err != nil {
2238 0 : i.err = err
2239 0 : i.iterValidityState = IterExhausted
2240 0 : }
2241 1 : if callerOwned {
2242 1 : i.lazyValueBuf = val[:0]
2243 1 : }
2244 1 : return val, err
2245 : }
2246 :
2247 : // LazyValue returns the LazyValue. Only for advanced use cases.
2248 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2249 0 : func (i *Iterator) LazyValue() LazyValue {
2250 0 : return i.value
2251 0 : }
2252 :
2253 : // RangeKeys returns the range key values and their suffixes covering the
2254 : // current iterator position. The range bounds may be retrieved separately
2255 : // through Iterator.RangeBounds().
2256 1 : func (i *Iterator) RangeKeys() []RangeKeyData {
2257 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2258 0 : return nil
2259 0 : }
2260 1 : return i.rangeKey.keys
2261 : }
2262 :
2263 : // Valid returns true if the iterator is positioned at a valid key/value pair
2264 : // and false otherwise.
2265 1 : func (i *Iterator) Valid() bool {
2266 1 : valid := i.iterValidityState == IterValid && !i.requiresReposition
2267 1 : if invariants.Enabled {
2268 1 : if err := i.Error(); valid && err != nil {
2269 0 : panic(errors.AssertionFailedf("pebble: iterator is valid with non-nil Error: %+v", err))
2270 : }
2271 : }
2272 1 : return valid
2273 : }
2274 :
2275 : // Error returns any accumulated error.
2276 1 : func (i *Iterator) Error() error {
2277 1 : if i.err != nil {
2278 1 : return i.err
2279 1 : }
2280 1 : if i.iter != nil {
2281 1 : return i.iter.Error()
2282 1 : }
2283 0 : return nil
2284 : }
2285 :
2286 : const maxKeyBufCacheSize = 4 << 10 // 4 KB
2287 :
2288 : // Close closes the iterator and returns any accumulated error. Exhausting
2289 : // all the key/value pairs in a table is not considered to be an error.
2290 : // It is not valid to call any method, including Close, after the iterator
2291 : // has been closed.
2292 1 : func (i *Iterator) Close() error {
2293 1 : // Close the child iterator before releasing the readState because when the
2294 1 : // readState is released sstables referenced by the readState may be deleted
2295 1 : // which will fail on Windows if the sstables are still open by the child
2296 1 : // iterator.
2297 1 : if i.iter != nil {
2298 1 : i.err = firstError(i.err, i.iter.Close())
2299 1 :
2300 1 : // Closing i.iter did not necessarily close the point and range key
2301 1 : // iterators. Calls to SetOptions may have 'disconnected' either one
2302 1 : // from i.iter if iteration key types were changed. Both point and range
2303 1 : // key iterators are preserved in case the iterator needs to switch key
2304 1 : // types again. We explicitly close both of these iterators here.
2305 1 : //
2306 1 : // NB: If the iterators were still connected to i.iter, they may be
2307 1 : // closed, but calling Close on a closed internal iterator or fragment
2308 1 : // iterator is allowed.
2309 1 : if i.pointIter != nil {
2310 1 : i.err = firstError(i.err, i.pointIter.Close())
2311 1 : }
2312 1 : if i.rangeKey != nil && i.rangeKey.rangeKeyIter != nil {
2313 1 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2314 1 : }
2315 : }
2316 1 : err := i.err
2317 1 :
2318 1 : if i.readState != nil {
2319 1 : if i.readSampling.pendingCompactions.size > 0 {
2320 0 : // Copy pending read compactions using db.mu.Lock()
2321 0 : i.readState.db.mu.Lock()
2322 0 : i.readState.db.mu.compact.readCompactions.combine(&i.readSampling.pendingCompactions, i.cmp)
2323 0 : reschedule := i.readState.db.mu.compact.rescheduleReadCompaction
2324 0 : i.readState.db.mu.compact.rescheduleReadCompaction = false
2325 0 : concurrentCompactions := i.readState.db.mu.compact.compactingCount
2326 0 : i.readState.db.mu.Unlock()
2327 0 :
2328 0 : if reschedule && concurrentCompactions == 0 {
2329 0 : // In a read heavy workload, flushes may not happen frequently enough to
2330 0 : // schedule compactions.
2331 0 : i.readState.db.compactionSchedulers.Add(1)
2332 0 : go i.readState.db.maybeScheduleCompactionAsync()
2333 0 : }
2334 : }
2335 :
2336 1 : i.readState.unref()
2337 1 : i.readState = nil
2338 : }
2339 :
2340 1 : if i.version != nil {
2341 1 : i.version.Unref()
2342 1 : }
2343 :
2344 1 : for _, readers := range i.externalReaders {
2345 0 : for _, r := range readers {
2346 0 : err = firstError(err, r.Close())
2347 0 : }
2348 : }
2349 :
2350 : // Close the closer for the current value if one was open.
2351 1 : if i.valueCloser != nil {
2352 0 : err = firstError(err, i.valueCloser.Close())
2353 0 : i.valueCloser = nil
2354 0 : }
2355 :
2356 1 : if i.rangeKey != nil {
2357 1 :
2358 1 : i.rangeKey.rangeKeyBuffers.PrepareForReuse()
2359 1 : *i.rangeKey = iteratorRangeKeyState{
2360 1 : rangeKeyBuffers: i.rangeKey.rangeKeyBuffers,
2361 1 : }
2362 1 : iterRangeKeyStateAllocPool.Put(i.rangeKey)
2363 1 : i.rangeKey = nil
2364 1 : }
2365 1 : if alloc := i.alloc; alloc != nil {
2366 1 : // Avoid caching the key buf if it is overly large. The constant is fairly
2367 1 : // arbitrary.
2368 1 : if cap(i.keyBuf) >= maxKeyBufCacheSize {
2369 0 : alloc.keyBuf = nil
2370 1 : } else {
2371 1 : alloc.keyBuf = i.keyBuf
2372 1 : }
2373 1 : if cap(i.prefixOrFullSeekKey) >= maxKeyBufCacheSize {
2374 0 : alloc.prefixOrFullSeekKey = nil
2375 1 : } else {
2376 1 : alloc.prefixOrFullSeekKey = i.prefixOrFullSeekKey
2377 1 : }
2378 1 : for j := range i.boundsBuf {
2379 1 : if cap(i.boundsBuf[j]) >= maxKeyBufCacheSize {
2380 0 : alloc.boundsBuf[j] = nil
2381 1 : } else {
2382 1 : alloc.boundsBuf[j] = i.boundsBuf[j]
2383 1 : }
2384 : }
2385 1 : *alloc = iterAlloc{
2386 1 : keyBuf: alloc.keyBuf,
2387 1 : boundsBuf: alloc.boundsBuf,
2388 1 : prefixOrFullSeekKey: alloc.prefixOrFullSeekKey,
2389 1 : }
2390 1 : iterAllocPool.Put(alloc)
2391 1 : } else if alloc := i.getIterAlloc; alloc != nil {
2392 1 : if cap(i.keyBuf) >= maxKeyBufCacheSize {
2393 0 : alloc.keyBuf = nil
2394 1 : } else {
2395 1 : alloc.keyBuf = i.keyBuf
2396 1 : }
2397 1 : *alloc = getIterAlloc{
2398 1 : keyBuf: alloc.keyBuf,
2399 1 : }
2400 1 : getIterAllocPool.Put(alloc)
2401 : }
2402 1 : return err
2403 : }
2404 :
2405 : // SetBounds sets the lower and upper bounds for the iterator. Once SetBounds
2406 : // returns, the caller is free to mutate the provided slices.
2407 : //
2408 : // The iterator will always be invalidated and must be repositioned with a call
2409 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2410 1 : func (i *Iterator) SetBounds(lower, upper []byte) {
2411 1 : // Ensure that the Iterator appears exhausted, regardless of whether we
2412 1 : // actually have to invalidate the internal iterator. Optimizations that
2413 1 : // avoid exhaustion are an internal implementation detail that shouldn't
2414 1 : // leak through the interface. The caller should still call an absolute
2415 1 : // positioning method to reposition the iterator.
2416 1 : i.requiresReposition = true
2417 1 :
2418 1 : if ((i.opts.LowerBound == nil) == (lower == nil)) &&
2419 1 : ((i.opts.UpperBound == nil) == (upper == nil)) &&
2420 1 : i.equal(i.opts.LowerBound, lower) &&
2421 1 : i.equal(i.opts.UpperBound, upper) {
2422 1 : // Unchanged, noop.
2423 1 : return
2424 1 : }
2425 :
2426 : // Copy the user-provided bounds into an Iterator-owned buffer, and set them
2427 : // on i.opts.{Lower,Upper}Bound.
2428 1 : i.processBounds(lower, upper)
2429 1 :
2430 1 : i.iter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2431 1 : // If the iterator has an open point iterator that's not currently being
2432 1 : // used, propagate the new bounds to it.
2433 1 : if i.pointIter != nil && !i.opts.pointKeys() {
2434 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2435 1 : }
2436 : // If the iterator has a range key iterator, propagate bounds to it. The
2437 : // top-level SetBounds on the interleaving iterator (i.iter) won't propagate
2438 : // bounds to the range key iterator stack, because the FragmentIterator
2439 : // interface doesn't define a SetBounds method. We need to directly inform
2440 : // the iterConfig stack.
2441 1 : if i.rangeKey != nil {
2442 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2443 1 : }
2444 :
2445 : // Even though this is not a positioning operation, the alteration of the
2446 : // bounds means we cannot optimize Seeks by using Next.
2447 1 : i.invalidate()
2448 : }
2449 :
2450 : // SetContext replaces the context provided at iterator creation, or the last
2451 : // one provided by SetContext. Even though iterators are expected to be
2452 : // short-lived, there are some cases where either (a) iterators are used far
2453 : // from the code that created them, (b) iterators are reused (while being
2454 : // short-lived) for processing different requests. For such scenarios, we
2455 : // allow the caller to replace the context.
2456 0 : func (i *Iterator) SetContext(ctx context.Context) {
2457 0 : i.ctx = ctx
2458 0 : i.iter.SetContext(ctx)
2459 0 : // If the iterator has an open point iterator that's not currently being
2460 0 : // used, propagate the new context to it.
2461 0 : if i.pointIter != nil && !i.opts.pointKeys() {
2462 0 : i.pointIter.SetContext(i.ctx)
2463 0 : }
2464 : }
2465 :
2466 : // Initialization and changing of the bounds must call processBounds.
2467 : // processBounds saves the bounds and computes derived state from those
2468 : // bounds.
2469 1 : func (i *Iterator) processBounds(lower, upper []byte) {
2470 1 : // Copy the user-provided bounds into an Iterator-owned buffer. We can't
2471 1 : // overwrite the current bounds, because some internal iterators compare old
2472 1 : // and new bounds for optimizations.
2473 1 :
2474 1 : buf := i.boundsBuf[i.boundsBufIdx][:0]
2475 1 : if lower != nil {
2476 1 : buf = append(buf, lower...)
2477 1 : i.opts.LowerBound = buf
2478 1 : } else {
2479 1 : i.opts.LowerBound = nil
2480 1 : }
2481 1 : i.nextPrefixNotPermittedByUpperBound = false
2482 1 : if upper != nil {
2483 1 : buf = append(buf, upper...)
2484 1 : i.opts.UpperBound = buf[len(buf)-len(upper):]
2485 1 : if i.comparer.Split(i.opts.UpperBound) != len(i.opts.UpperBound) {
2486 1 : // Setting an upper bound that is a versioned MVCC key. This means
2487 1 : // that a key can have some MVCC versions before the upper bound and
2488 1 : // some after. This causes significant complications for NextPrefix,
2489 1 : // so we bar the user of NextPrefix.
2490 1 : i.nextPrefixNotPermittedByUpperBound = true
2491 1 : }
2492 1 : } else {
2493 1 : i.opts.UpperBound = nil
2494 1 : }
2495 1 : i.boundsBuf[i.boundsBufIdx] = buf
2496 1 : i.boundsBufIdx = 1 - i.boundsBufIdx
2497 : }
2498 :
2499 : // SetOptions sets new iterator options for the iterator. Note that the lower
2500 : // and upper bounds applied here will supersede any bounds set by previous calls
2501 : // to SetBounds.
2502 : //
2503 : // Note that the slices provided in this SetOptions must not be changed by the
2504 : // caller until the iterator is closed, or a subsequent SetBounds or SetOptions
2505 : // has returned. This is because comparisons between the existing and new bounds
2506 : // are sometimes used to optimize seeking. See the extended commentary on
2507 : // SetBounds.
2508 : //
2509 : // If the iterator was created over an indexed mutable batch, the iterator's
2510 : // view of the mutable batch is refreshed.
2511 : //
2512 : // The iterator will always be invalidated and must be repositioned with a call
2513 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2514 : //
2515 : // If only lower and upper bounds need to be modified, prefer SetBounds.
2516 1 : func (i *Iterator) SetOptions(o *IterOptions) {
2517 1 : if i.externalReaders != nil {
2518 0 : if err := validateExternalIterOpts(o); err != nil {
2519 0 : panic(err)
2520 : }
2521 : }
2522 :
2523 : // Ensure that the Iterator appears exhausted, regardless of whether we
2524 : // actually have to invalidate the internal iterator. Optimizations that
2525 : // avoid exhaustion are an internal implementation detail that shouldn't
2526 : // leak through the interface. The caller should still call an absolute
2527 : // positioning method to reposition the iterator.
2528 1 : i.requiresReposition = true
2529 1 :
2530 1 : // Check if global state requires we close all internal iterators.
2531 1 : //
2532 1 : // If the Iterator is in an error state, invalidate the existing iterators
2533 1 : // so that we reconstruct an iterator state from scratch.
2534 1 : //
2535 1 : // If OnlyReadGuaranteedDurable changed, the iterator stacks are incorrect,
2536 1 : // improperly including or excluding memtables. Invalidate them so that
2537 1 : // finishInitializingIter will reconstruct them.
2538 1 : //
2539 1 : // If either the original options or the new options specify a table filter,
2540 1 : // we need to reconstruct the iterator stacks. If they both supply a table
2541 1 : // filter, we can't be certain that it's the same filter since we have no
2542 1 : // mechanism to compare the filter closures.
2543 1 : closeBoth := i.err != nil ||
2544 1 : o.OnlyReadGuaranteedDurable != i.opts.OnlyReadGuaranteedDurable ||
2545 1 : o.TableFilter != nil || i.opts.TableFilter != nil
2546 1 :
2547 1 : // If either options specify block property filters for an iterator stack,
2548 1 : // reconstruct it.
2549 1 : if i.pointIter != nil && (closeBoth || len(o.PointKeyFilters) > 0 || len(i.opts.PointKeyFilters) > 0 ||
2550 1 : o.RangeKeyMasking.Filter != nil || i.opts.RangeKeyMasking.Filter != nil || o.SkipPoint != nil ||
2551 1 : i.opts.SkipPoint != nil) {
2552 1 : i.err = firstError(i.err, i.pointIter.Close())
2553 1 : i.pointIter = nil
2554 1 : }
2555 1 : if i.rangeKey != nil {
2556 1 : if closeBoth || len(o.RangeKeyFilters) > 0 || len(i.opts.RangeKeyFilters) > 0 {
2557 1 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2558 1 : i.rangeKey = nil
2559 1 : } else {
2560 1 : // If there's still a range key iterator stack, invalidate the
2561 1 : // iterator. This ensures RangeKeyChanged() returns true if a
2562 1 : // subsequent positioning operation discovers a range key. It also
2563 1 : // prevents seek no-op optimizations.
2564 1 : i.invalidate()
2565 1 : }
2566 : }
2567 :
2568 : // If the iterator is backed by a batch that's been mutated, refresh its
2569 : // existing point and range-key iterators, and invalidate the iterator to
2570 : // prevent seek-using-next optimizations. If we don't yet have a point-key
2571 : // iterator or range-key iterator but we require one, it'll be created in
2572 : // the slow path that reconstructs the iterator in finishInitializingIter.
2573 1 : if i.batch != nil {
2574 1 : nextBatchSeqNum := (uint64(len(i.batch.data)) | base.InternalKeySeqNumBatch)
2575 1 : if nextBatchSeqNum != i.batchSeqNum {
2576 0 : i.batchSeqNum = nextBatchSeqNum
2577 0 : if i.merging != nil {
2578 0 : i.merging.batchSnapshot = nextBatchSeqNum
2579 0 : }
2580 : // Prevent a no-op seek optimization on the next seek. We won't be
2581 : // able to reuse the top-level Iterator state, because it may be
2582 : // incorrect after the inclusion of new batch mutations.
2583 0 : i.batchJustRefreshed = true
2584 0 : if i.pointIter != nil && i.batch.countRangeDels > 0 {
2585 0 : if i.batchRangeDelIter.Count() == 0 {
2586 0 : // When we constructed this iterator, there were no
2587 0 : // rangedels in the batch. Iterator construction will
2588 0 : // have excluded the batch rangedel iterator from the
2589 0 : // point iterator stack. We need to reconstruct the
2590 0 : // point iterator to add i.batchRangeDelIter into the
2591 0 : // iterator stack.
2592 0 : i.err = firstError(i.err, i.pointIter.Close())
2593 0 : i.pointIter = nil
2594 0 : } else {
2595 0 : // There are range deletions in the batch and we already
2596 0 : // have a batch rangedel iterator. We can update the
2597 0 : // batch rangedel iterator in place.
2598 0 : //
2599 0 : // NB: There may or may not be new range deletions. We
2600 0 : // can't tell based on i.batchRangeDelIter.Count(),
2601 0 : // which is the count of fragmented range deletions, NOT
2602 0 : // the number of range deletions written to the batch
2603 0 : // [i.batch.countRangeDels].
2604 0 : i.batch.initRangeDelIter(&i.opts, &i.batchRangeDelIter, nextBatchSeqNum)
2605 0 : }
2606 : }
2607 0 : if i.rangeKey != nil && i.batch.countRangeKeys > 0 {
2608 0 : if i.batchRangeKeyIter.Count() == 0 {
2609 0 : // When we constructed this iterator, there were no range
2610 0 : // keys in the batch. Iterator construction will have
2611 0 : // excluded the batch rangekey iterator from the range key
2612 0 : // iterator stack. We need to reconstruct the range key
2613 0 : // iterator to add i.batchRangeKeyIter into the iterator
2614 0 : // stack.
2615 0 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2616 0 : i.rangeKey = nil
2617 0 : } else {
2618 0 : // There are range keys in the batch and we already
2619 0 : // have a batch rangekey iterator. We can update the batch
2620 0 : // rangekey iterator in place.
2621 0 : //
2622 0 : // NB: There may or may not be new range keys. We can't
2623 0 : // tell based on i.batchRangeKeyIter.Count(), which is the
2624 0 : // count of fragmented range keys, NOT the number of
2625 0 : // range keys written to the batch [i.batch.countRangeKeys].
2626 0 : i.batch.initRangeKeyIter(&i.opts, &i.batchRangeKeyIter, nextBatchSeqNum)
2627 0 : i.invalidate()
2628 0 : }
2629 : }
2630 : }
2631 : }
2632 :
2633 : // Reset combinedIterState.initialized in case the iterator key types
2634 : // changed. If there's already a range key iterator stack, the combined
2635 : // iterator is already initialized. Additionally, if the iterator is not
2636 : // configured to include range keys, mark it as initialized to signal that
2637 : // lower level iterators should not trigger a switch to combined iteration.
2638 1 : i.lazyCombinedIter.combinedIterState = combinedIterState{
2639 1 : initialized: i.rangeKey != nil || !i.opts.rangeKeys(),
2640 1 : }
2641 1 :
2642 1 : boundsEqual := ((i.opts.LowerBound == nil) == (o.LowerBound == nil)) &&
2643 1 : ((i.opts.UpperBound == nil) == (o.UpperBound == nil)) &&
2644 1 : i.equal(i.opts.LowerBound, o.LowerBound) &&
2645 1 : i.equal(i.opts.UpperBound, o.UpperBound)
2646 1 :
2647 1 : if boundsEqual && o.KeyTypes == i.opts.KeyTypes &&
2648 1 : (i.pointIter != nil || !i.opts.pointKeys()) &&
2649 1 : (i.rangeKey != nil || !i.opts.rangeKeys() || i.opts.KeyTypes == IterKeyTypePointsAndRanges) &&
2650 1 : i.equal(o.RangeKeyMasking.Suffix, i.opts.RangeKeyMasking.Suffix) &&
2651 1 : o.UseL6Filters == i.opts.UseL6Filters {
2652 1 : // The options are identical, so we can likely use the fast path. In
2653 1 : // addition to all the above constraints, we cannot use the fast path if
2654 1 : // configured to perform lazy combined iteration but an indexed batch
2655 1 : // used by the iterator now contains range keys. Lazy combined iteration
2656 1 : // is not compatible with batch range keys because we always need to
2657 1 : // merge the batch's range keys into iteration.
2658 1 : if i.rangeKey != nil || !i.opts.rangeKeys() || i.batch == nil || i.batch.countRangeKeys == 0 {
2659 1 : // Fast path. This preserves the Seek-using-Next optimizations as
2660 1 : // long as the iterator wasn't already invalidated up above.
2661 1 : return
2662 1 : }
2663 : }
2664 : // Slow path.
2665 :
2666 : // The options changed. Save the new ones to i.opts.
2667 1 : if boundsEqual {
2668 1 : // Copying the options into i.opts will overwrite LowerBound and
2669 1 : // UpperBound fields with the user-provided slices. We need to hold on
2670 1 : // to the Pebble-owned slices, so save them and re-set them after the
2671 1 : // copy.
2672 1 : lower, upper := i.opts.LowerBound, i.opts.UpperBound
2673 1 : i.opts = *o
2674 1 : i.opts.LowerBound, i.opts.UpperBound = lower, upper
2675 1 : } else {
2676 1 : i.opts = *o
2677 1 : i.processBounds(o.LowerBound, o.UpperBound)
2678 1 : // Propagate the changed bounds to the existing point iterator.
2679 1 : // NB: We propagate i.opts.{Lower,Upper}Bound, not o.{Lower,Upper}Bound
2680 1 : // because i.opts now point to buffers owned by Pebble.
2681 1 : if i.pointIter != nil {
2682 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2683 1 : }
2684 1 : if i.rangeKey != nil {
2685 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2686 1 : }
2687 : }
2688 :
2689 : // Even though this is not a positioning operation, the invalidation of the
2690 : // iterator stack means we cannot optimize Seeks by using Next.
2691 1 : i.invalidate()
2692 1 :
2693 1 : // Iterators created through NewExternalIter have a different iterator
2694 1 : // initialization process.
2695 1 : if i.externalReaders != nil {
2696 0 : finishInitializingExternal(i.ctx, i)
2697 0 : return
2698 0 : }
2699 1 : finishInitializingIter(i.ctx, i.alloc)
2700 : }
2701 :
2702 1 : func (i *Iterator) invalidate() {
2703 1 : i.lastPositioningOp = unknownLastPositionOp
2704 1 : i.hasPrefix = false
2705 1 : i.iterKV = nil
2706 1 : i.err = nil
2707 1 : // This switch statement isn't necessary for correctness since callers
2708 1 : // should call a repositioning method. We could have arbitrarily set i.pos
2709 1 : // to one of the values. But it results in more intuitive behavior in
2710 1 : // tests, which do not always reposition.
2711 1 : switch i.pos {
2712 1 : case iterPosCurForward, iterPosNext, iterPosCurForwardPaused:
2713 1 : i.pos = iterPosCurForward
2714 1 : case iterPosCurReverse, iterPosPrev, iterPosCurReversePaused:
2715 1 : i.pos = iterPosCurReverse
2716 : }
2717 1 : i.iterValidityState = IterExhausted
2718 1 : if i.rangeKey != nil {
2719 1 : i.rangeKey.iiter.Invalidate()
2720 1 : i.rangeKey.prevPosHadRangeKey = false
2721 1 : }
2722 : }
2723 :
2724 : // Metrics returns per-iterator metrics.
2725 0 : func (i *Iterator) Metrics() IteratorMetrics {
2726 0 : m := IteratorMetrics{
2727 0 : ReadAmp: 1,
2728 0 : }
2729 0 : if mi, ok := i.iter.(*mergingIter); ok {
2730 0 : m.ReadAmp = len(mi.levels)
2731 0 : }
2732 0 : return m
2733 : }
2734 :
2735 : // ResetStats resets the stats to 0.
2736 0 : func (i *Iterator) ResetStats() {
2737 0 : i.stats = IteratorStats{}
2738 0 : }
2739 :
2740 : // Stats returns the current stats.
2741 0 : func (i *Iterator) Stats() IteratorStats {
2742 0 : return i.stats
2743 0 : }
2744 :
2745 : // CloneOptions configures an iterator constructed through Iterator.Clone.
2746 : type CloneOptions struct {
2747 : // IterOptions, if non-nil, define the iterator options to configure a
2748 : // cloned iterator. If nil, the clone adopts the same IterOptions as the
2749 : // iterator being cloned.
2750 : IterOptions *IterOptions
2751 : // RefreshBatchView may be set to true when cloning an Iterator over an
2752 : // indexed batch. When false, the clone adopts the same (possibly stale)
2753 : // view of the indexed batch as the cloned Iterator. When true, the clone is
2754 : // constructed with a refreshed view of the batch, observing all of the
2755 : // batch's mutations at the time of the Clone. If the cloned iterator was
2756 : // not constructed to read over an indexed batch, RefreshVatchView has no
2757 : // effect.
2758 : RefreshBatchView bool
2759 : }
2760 :
2761 : // Clone creates a new Iterator over the same underlying data, i.e., over the
2762 : // same {batch, memtables, sstables}). The resulting iterator is not positioned.
2763 : // It starts with the same IterOptions, unless opts.IterOptions is set.
2764 : //
2765 : // When called on an Iterator over an indexed batch, the clone's visibility of
2766 : // the indexed batch is determined by CloneOptions.RefreshBatchView. If false,
2767 : // the clone inherits the iterator's current (possibly stale) view of the batch,
2768 : // and callers may call SetOptions to subsequently refresh the clone's view to
2769 : // include all batch mutations. If true, the clone is constructed with a
2770 : // complete view of the indexed batch's mutations at the time of the Clone.
2771 : //
2772 : // Callers can use Clone if they need multiple iterators that need to see
2773 : // exactly the same underlying state of the DB. This should not be used to
2774 : // extend the lifetime of the data backing the original Iterator since that
2775 : // will cause an increase in memory and disk usage (use NewSnapshot for that
2776 : // purpose).
2777 1 : func (i *Iterator) Clone(opts CloneOptions) (*Iterator, error) {
2778 1 : return i.CloneWithContext(context.Background(), opts)
2779 1 : }
2780 :
2781 : // CloneWithContext is like Clone, and additionally accepts a context for
2782 : // tracing.
2783 1 : func (i *Iterator) CloneWithContext(ctx context.Context, opts CloneOptions) (*Iterator, error) {
2784 1 : if opts.IterOptions == nil {
2785 1 : opts.IterOptions = &i.opts
2786 1 : }
2787 1 : if i.batchOnlyIter {
2788 0 : return nil, errors.Errorf("cannot Clone a batch-only Iterator")
2789 0 : }
2790 1 : readState := i.readState
2791 1 : vers := i.version
2792 1 : if readState == nil && vers == nil {
2793 0 : return nil, errors.Errorf("cannot Clone a closed Iterator")
2794 0 : }
2795 : // i is already holding a ref, so there is no race with unref here.
2796 : //
2797 : // TODO(bilal): If the underlying iterator was created on a snapshot, we could
2798 : // grab a reference to the current readState instead of reffing the original
2799 : // readState. This allows us to release references to some zombie sstables
2800 : // and memtables.
2801 1 : if readState != nil {
2802 1 : readState.ref()
2803 1 : }
2804 1 : if vers != nil {
2805 1 : vers.Ref()
2806 1 : }
2807 : // Bundle various structures under a single umbrella in order to allocate
2808 : // them together.
2809 1 : buf := iterAllocPool.Get().(*iterAlloc)
2810 1 : dbi := &buf.dbi
2811 1 : *dbi = Iterator{
2812 1 : ctx: ctx,
2813 1 : opts: *opts.IterOptions,
2814 1 : alloc: buf,
2815 1 : merge: i.merge,
2816 1 : comparer: i.comparer,
2817 1 : readState: readState,
2818 1 : version: vers,
2819 1 : keyBuf: buf.keyBuf,
2820 1 : prefixOrFullSeekKey: buf.prefixOrFullSeekKey,
2821 1 : boundsBuf: buf.boundsBuf,
2822 1 : batch: i.batch,
2823 1 : batchSeqNum: i.batchSeqNum,
2824 1 : newIters: i.newIters,
2825 1 : newIterRangeKey: i.newIterRangeKey,
2826 1 : seqNum: i.seqNum,
2827 1 : }
2828 1 : dbi.processBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
2829 1 :
2830 1 : // If the caller requested the clone have a current view of the indexed
2831 1 : // batch, set the clone's batch sequence number appropriately.
2832 1 : if i.batch != nil && opts.RefreshBatchView {
2833 0 : dbi.batchSeqNum = (uint64(len(i.batch.data)) | base.InternalKeySeqNumBatch)
2834 0 : }
2835 :
2836 1 : return finishInitializingIter(ctx, buf), nil
2837 : }
2838 :
2839 : // Merge adds all of the argument's statistics to the receiver. It may be used
2840 : // to accumulate stats across multiple iterators.
2841 0 : func (stats *IteratorStats) Merge(o IteratorStats) {
2842 0 : for i := InterfaceCall; i < NumStatsKind; i++ {
2843 0 : stats.ForwardSeekCount[i] += o.ForwardSeekCount[i]
2844 0 : stats.ReverseSeekCount[i] += o.ReverseSeekCount[i]
2845 0 : stats.ForwardStepCount[i] += o.ForwardStepCount[i]
2846 0 : stats.ReverseStepCount[i] += o.ReverseStepCount[i]
2847 0 : }
2848 0 : stats.InternalStats.Merge(o.InternalStats)
2849 0 : stats.RangeKeyStats.Merge(o.RangeKeyStats)
2850 : }
2851 :
2852 0 : func (stats *IteratorStats) String() string {
2853 0 : return redact.StringWithoutMarkers(stats)
2854 0 : }
2855 :
2856 : // SafeFormat implements the redact.SafeFormatter interface.
2857 0 : func (stats *IteratorStats) SafeFormat(s redact.SafePrinter, verb rune) {
2858 0 : for i := range stats.ForwardStepCount {
2859 0 : switch IteratorStatsKind(i) {
2860 0 : case InterfaceCall:
2861 0 : s.SafeString("(interface (dir, seek, step): ")
2862 0 : case InternalIterCall:
2863 0 : s.SafeString(", (internal (dir, seek, step): ")
2864 : }
2865 0 : s.Printf("(fwd, %d, %d), (rev, %d, %d))",
2866 0 : redact.Safe(stats.ForwardSeekCount[i]), redact.Safe(stats.ForwardStepCount[i]),
2867 0 : redact.Safe(stats.ReverseSeekCount[i]), redact.Safe(stats.ReverseStepCount[i]))
2868 : }
2869 0 : if stats.InternalStats != (InternalIteratorStats{}) {
2870 0 : s.SafeString(",\n(internal-stats: ")
2871 0 : s.Printf("(block-bytes: (total %s, cached %s, read-time %s)), "+
2872 0 : "(points: (count %s, key-bytes %s, value-bytes %s, tombstoned %s))",
2873 0 : humanize.Bytes.Uint64(stats.InternalStats.BlockBytes),
2874 0 : humanize.Bytes.Uint64(stats.InternalStats.BlockBytesInCache),
2875 0 : humanize.FormattedString(stats.InternalStats.BlockReadDuration.String()),
2876 0 : humanize.Count.Uint64(stats.InternalStats.PointCount),
2877 0 : humanize.Bytes.Uint64(stats.InternalStats.KeyBytes),
2878 0 : humanize.Bytes.Uint64(stats.InternalStats.ValueBytes),
2879 0 : humanize.Count.Uint64(stats.InternalStats.PointsCoveredByRangeTombstones),
2880 0 : )
2881 0 : if stats.InternalStats.SeparatedPointValue.Count != 0 {
2882 0 : s.Printf(", (separated: (count %s, bytes %s, fetched %s)))",
2883 0 : humanize.Count.Uint64(stats.InternalStats.SeparatedPointValue.Count),
2884 0 : humanize.Bytes.Uint64(stats.InternalStats.SeparatedPointValue.ValueBytes),
2885 0 : humanize.Bytes.Uint64(stats.InternalStats.SeparatedPointValue.ValueBytesFetched))
2886 0 : } else {
2887 0 : s.Printf(")")
2888 0 : }
2889 : }
2890 0 : if stats.RangeKeyStats != (RangeKeyIteratorStats{}) {
2891 0 : s.SafeString(",\n(range-key-stats: ")
2892 0 : s.Printf("(count %d), (contained points: (count %d, skipped %d)))",
2893 0 : stats.RangeKeyStats.Count,
2894 0 : stats.RangeKeyStats.ContainedPoints,
2895 0 : stats.RangeKeyStats.SkippedPoints)
2896 0 : }
2897 : }
2898 :
2899 : // CanDeterministicallySingleDelete takes a valid iterator and examines internal
2900 : // state to determine if a SingleDelete deleting Iterator.Key() would
2901 : // deterministically delete the key. CanDeterministicallySingleDelete requires
2902 : // the iterator to be oriented in the forward direction (eg, the last
2903 : // positioning operation must've been a First, a Seek[Prefix]GE, or a
2904 : // Next[Prefix][WithLimit]).
2905 : //
2906 : // This function does not change the external position of the iterator, and all
2907 : // positioning methods should behave the same as if it was never called. This
2908 : // function will only return a meaningful result the first time it's invoked at
2909 : // an iterator position. This function invalidates the iterator Value's memory,
2910 : // and the caller must not rely on the memory safety of the previous Iterator
2911 : // position.
2912 : //
2913 : // If CanDeterministicallySingleDelete returns true AND the key at the iterator
2914 : // position is not modified between the creation of the Iterator and the commit
2915 : // of a batch containing a SingleDelete over the key, then the caller can be
2916 : // assured that SingleDelete is equivalent to Delete on the local engine, but it
2917 : // may not be true on another engine that received the same writes and with
2918 : // logically equivalent state since this engine may have collapsed multiple SETs
2919 : // into one.
2920 1 : func CanDeterministicallySingleDelete(it *Iterator) (bool, error) {
2921 1 : // This function may only be called once per external iterator position. We
2922 1 : // can validate this by checking the last positioning operation.
2923 1 : if it.lastPositioningOp == internalNextOp {
2924 1 : return false, errors.New("pebble: CanDeterministicallySingleDelete called twice")
2925 1 : }
2926 1 : validity, kind := it.internalNext()
2927 1 : var shadowedBySingleDelete bool
2928 1 : for validity == internalNextValid {
2929 1 : switch kind {
2930 1 : case InternalKeyKindDelete, InternalKeyKindDeleteSized:
2931 1 : // A DEL or DELSIZED tombstone is okay. An internal key
2932 1 : // sequence like SINGLEDEL; SET; DEL; SET can be handled
2933 1 : // deterministically. If there are SETs further down, we
2934 1 : // don't care about them.
2935 1 : return true, nil
2936 1 : case InternalKeyKindSingleDelete:
2937 1 : // A SingleDelete is okay as long as when that SingleDelete was
2938 1 : // written, it was written deterministically (eg, with its own
2939 1 : // CanDeterministicallySingleDelete check). Validate that it was
2940 1 : // written deterministically. We'll allow one set to appear after
2941 1 : // the SingleDelete.
2942 1 : shadowedBySingleDelete = true
2943 1 : validity, kind = it.internalNext()
2944 1 : continue
2945 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete, InternalKeyKindMerge:
2946 1 : // If we observed a single delete, it's allowed to delete 1 key.
2947 1 : // We'll keep looping to validate that the internal keys beneath the
2948 1 : // already-written single delete are copacetic.
2949 1 : if shadowedBySingleDelete {
2950 1 : shadowedBySingleDelete = false
2951 1 : validity, kind = it.internalNext()
2952 1 : continue
2953 : }
2954 : // We encountered a shadowed SET, SETWITHDEL, MERGE. A SINGLEDEL
2955 : // that deleted the KV at the original iterator position could
2956 : // result in this key becoming visible.
2957 1 : return false, nil
2958 0 : case InternalKeyKindRangeDelete:
2959 0 : // RangeDeletes are handled by the merging iterator and should never
2960 0 : // be observed by the top-level Iterator.
2961 0 : panic(errors.AssertionFailedf("pebble: unexpected range delete"))
2962 0 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
2963 0 : // Range keys are interleaved at the maximal sequence number and
2964 0 : // should never be observed within a user key.
2965 0 : panic(errors.AssertionFailedf("pebble: unexpected range key"))
2966 0 : default:
2967 0 : panic(errors.AssertionFailedf("pebble: unexpected key kind: %s", errors.Safe(kind)))
2968 : }
2969 : }
2970 1 : if validity == internalNextError {
2971 1 : return false, it.Error()
2972 1 : }
2973 1 : return true, nil
2974 : }
2975 :
2976 : // internalNextValidity enumerates the potential outcomes of a call to
2977 : // internalNext.
2978 : type internalNextValidity int8
2979 :
2980 : const (
2981 : // internalNextError is returned by internalNext when an error occurred and
2982 : // the caller is responsible for checking iter.Error().
2983 : internalNextError internalNextValidity = iota
2984 : // internalNextExhausted is returned by internalNext when the next internal
2985 : // key is an internal key with a different user key than Iterator.Key().
2986 : internalNextExhausted
2987 : // internalNextValid is returned by internalNext when the internal next
2988 : // found a shadowed internal key with a user key equal to Iterator.Key().
2989 : internalNextValid
2990 : )
2991 :
2992 : // internalNext advances internal Iterator state forward to expose the
2993 : // InternalKeyKind of the next internal key with a user key equal to Key().
2994 : //
2995 : // internalNext is a highly specialized operation and is unlikely to be
2996 : // generally useful. See Iterator.Next for how to reposition the iterator to the
2997 : // next key. internalNext requires the Iterator to be at a valid position in the
2998 : // forward direction (the last positioning operation must've been a First, a
2999 : // Seek[Prefix]GE, or a Next[Prefix][WithLimit] and Valid() must return true).
3000 : //
3001 : // internalNext, unlike all other Iterator methods, exposes internal LSM state.
3002 : // internalNext advances the Iterator's internal iterator to the next shadowed
3003 : // key with a user key equal to Key(). When a key is overwritten or deleted, its
3004 : // removal from the LSM occurs lazily as a part of compactions. internalNext
3005 : // allows the caller to see whether an obsolete internal key exists with the
3006 : // current Key(), and what it's key kind is. Note that the existence of an
3007 : // internal key is nondeterministic and dependent on internal LSM state. These
3008 : // semantics are unlikely to be applicable to almost all use cases.
3009 : //
3010 : // If internalNext finds a key that shares the same user key as Key(), it
3011 : // returns internalNextValid and the internal key's kind. If internalNext
3012 : // encounters an error, it returns internalNextError and the caller is expected
3013 : // to call Iterator.Error() to retrieve it. In all other circumstances,
3014 : // internalNext returns internalNextExhausted, indicating that there are no more
3015 : // additional internal keys with the user key Key().
3016 : //
3017 : // internalNext does not change the external position of the iterator, and a
3018 : // Next operation should behave the same as if internalNext was never called.
3019 : // internalNext does invalidate the iterator Value's memory, and the caller must
3020 : // not rely on the memory safety of the previous Iterator position.
3021 1 : func (i *Iterator) internalNext() (internalNextValidity, base.InternalKeyKind) {
3022 1 : i.stats.ForwardStepCount[InterfaceCall]++
3023 1 : if i.err != nil {
3024 1 : return internalNextError, base.InternalKeyKindInvalid
3025 1 : } else if i.iterValidityState != IterValid {
3026 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3027 1 : }
3028 1 : i.lastPositioningOp = internalNextOp
3029 1 :
3030 1 : switch i.pos {
3031 1 : case iterPosCurForward:
3032 1 : i.iterKV = i.iter.Next()
3033 1 : if i.iterKV == nil {
3034 1 : // We check i.iter.Error() here and return an internalNextError enum
3035 1 : // variant so that the caller does not need to check i.iter.Error()
3036 1 : // in the common case that the next internal key has a new user key.
3037 1 : if i.err = i.iter.Error(); i.err != nil {
3038 0 : return internalNextError, base.InternalKeyKindInvalid
3039 0 : }
3040 1 : i.pos = iterPosNext
3041 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3042 1 : } else if i.comparer.Equal(i.iterKV.K.UserKey, i.key) {
3043 1 : return internalNextValid, i.iterKV.Kind()
3044 1 : }
3045 1 : i.pos = iterPosNext
3046 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3047 1 : case iterPosCurReverse, iterPosCurReversePaused, iterPosPrev:
3048 1 : i.err = errors.New("switching from reverse to forward via internalNext is prohibited")
3049 1 : i.iterValidityState = IterExhausted
3050 1 : return internalNextError, base.InternalKeyKindInvalid
3051 1 : case iterPosNext, iterPosCurForwardPaused:
3052 1 : // The previous method already moved onto the next user key. This is
3053 1 : // only possible if
3054 1 : // - the last positioning method was a call to internalNext, and we
3055 1 : // advanced to a new user key.
3056 1 : // - the previous non-internalNext iterator operation encountered a
3057 1 : // range key or merge, forcing an internal Next that found a new
3058 1 : // user key that's not equal to i.Iterator.Key().
3059 1 : return internalNextExhausted, base.InternalKeyKindInvalid
3060 0 : default:
3061 0 : panic("unreachable")
3062 : }
3063 : }
|
