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/manifest"
22 : "github.com/cockroachdb/pebble/internal/rangekey"
23 : "github.com/cockroachdb/pebble/sstable"
24 : "github.com/cockroachdb/redact"
25 : )
26 :
27 : // iterPos describes the state of the internal iterator, in terms of whether it
28 : // is at the position returned to the user (cur), one ahead of the position
29 : // returned (next for forward iteration and prev for reverse iteration). The cur
30 : // position is split into two states, for forward and reverse iteration, since
31 : // we need to differentiate for switching directions.
32 : //
33 : // There is subtlety in what is considered the current position of the Iterator.
34 : // The internal iterator exposes a sequence of internal keys. There is not
35 : // always a single internalIterator position corresponding to the position
36 : // returned to the user. Consider the example:
37 : //
38 : // a.MERGE.9 a.MERGE.8 a.MERGE.7 a.SET.6 b.DELETE.9 b.DELETE.5 b.SET.4
39 : // \ /
40 : // \ Iterator.Key() = 'a' /
41 : //
42 : // The Iterator exposes one valid position at user key 'a' and the two exhausted
43 : // positions at the beginning and end of iteration. The underlying
44 : // internalIterator contains 7 valid positions and 2 exhausted positions.
45 : //
46 : // Iterator positioning methods must set iterPos to iterPosCur{Foward,Backward}
47 : // iff the user key at the current internalIterator position equals the
48 : // Iterator.Key returned to the user. This guarantees that a call to nextUserKey
49 : // or prevUserKey will advance to the next or previous iterator position.
50 : // iterPosCur{Forward,Backward} does not make any guarantee about the internal
51 : // iterator position among internal keys with matching user keys, and it will
52 : // vary subtly depending on the particular key kinds encountered. In the above
53 : // example, the iterator returning 'a' may set iterPosCurForward if the internal
54 : // iterator is positioned at any of a.MERGE.9, a.MERGE.8, a.MERGE.7 or a.SET.6.
55 : //
56 : // When setting iterPos to iterPosNext or iterPosPrev, the internal iterator
57 : // must be advanced to the first internalIterator position at a user key greater
58 : // (iterPosNext) or less (iterPosPrev) than the key returned to the user. An
59 : // internalIterator position that's !Valid() must also be considered greater or
60 : // less—depending on the direction of iteration—than the last valid Iterator
61 : // position.
62 : type iterPos int8
63 :
64 : const (
65 : iterPosCurForward iterPos = 0
66 : iterPosNext iterPos = 1
67 : iterPosPrev iterPos = -1
68 : iterPosCurReverse iterPos = -2
69 :
70 : // For limited iteration. When the iterator is at iterPosCurForwardPaused
71 : // - Next*() call should behave as if the internal iterator is already
72 : // at next (akin to iterPosNext).
73 : // - Prev*() call should behave as if the internal iterator is at the
74 : // current key (akin to iterPosCurForward).
75 : //
76 : // Similar semantics apply to CurReversePaused.
77 : iterPosCurForwardPaused iterPos = 2
78 : iterPosCurReversePaused iterPos = -3
79 : )
80 :
81 : // Approximate gap in bytes between samples of data read during iteration.
82 : // This is multiplied with a default ReadSamplingMultiplier of 1 << 4 to yield
83 : // 1 << 20 (1MB). The 1MB factor comes from:
84 : // https://github.com/cockroachdb/pebble/issues/29#issuecomment-494477985
85 : const readBytesPeriod uint64 = 1 << 16
86 :
87 : var errReversePrefixIteration = errors.New("pebble: unsupported reverse prefix iteration")
88 :
89 : // IteratorMetrics holds per-iterator metrics. These do not change over the
90 : // lifetime of the iterator.
91 : type IteratorMetrics struct {
92 : // The read amplification experienced by this iterator. This is the sum of
93 : // the memtables, the L0 sublevels and the non-empty Ln levels. Higher read
94 : // amplification generally results in slower reads, though allowing higher
95 : // read amplification can also result in faster writes.
96 : ReadAmp int
97 : }
98 :
99 : // IteratorStatsKind describes the two kind of iterator stats.
100 : type IteratorStatsKind int8
101 :
102 : const (
103 : // InterfaceCall represents calls to Iterator.
104 : InterfaceCall IteratorStatsKind = iota
105 : // InternalIterCall represents calls by Iterator to its internalIterator.
106 : InternalIterCall
107 : // NumStatsKind is the number of kinds, and is used for array sizing.
108 : NumStatsKind
109 : )
110 :
111 : // IteratorStats contains iteration stats.
112 : type IteratorStats struct {
113 : // ForwardSeekCount includes SeekGE, SeekPrefixGE, First.
114 : ForwardSeekCount [NumStatsKind]int
115 : // ReverseSeek includes SeekLT, Last.
116 : ReverseSeekCount [NumStatsKind]int
117 : // ForwardStepCount includes Next.
118 : ForwardStepCount [NumStatsKind]int
119 : // ReverseStepCount includes Prev.
120 : ReverseStepCount [NumStatsKind]int
121 : InternalStats InternalIteratorStats
122 : RangeKeyStats RangeKeyIteratorStats
123 : }
124 :
125 : var _ redact.SafeFormatter = &IteratorStats{}
126 :
127 : // InternalIteratorStats contains miscellaneous stats produced by internal
128 : // iterators.
129 : type InternalIteratorStats = base.InternalIteratorStats
130 :
131 : // RangeKeyIteratorStats contains miscellaneous stats about range keys
132 : // encountered by the iterator.
133 : type RangeKeyIteratorStats struct {
134 : // Count records the number of range keys encountered during
135 : // iteration. Range keys may be counted multiple times if the iterator
136 : // leaves a range key's bounds and then returns.
137 : Count int
138 : // ContainedPoints records the number of point keys encountered within the
139 : // bounds of a range key. Note that this includes point keys with suffixes
140 : // that sort both above and below the covering range key's suffix.
141 : ContainedPoints int
142 : // SkippedPoints records the count of the subset of ContainedPoints point
143 : // keys that were skipped during iteration due to range-key masking. It does
144 : // not include point keys that were never loaded because a
145 : // RangeKeyMasking.Filter excluded the entire containing block.
146 : SkippedPoints int
147 : }
148 :
149 : // Merge adds all of the argument's statistics to the receiver. It may be used
150 : // to accumulate stats across multiple iterators.
151 1 : func (s *RangeKeyIteratorStats) Merge(o RangeKeyIteratorStats) {
152 1 : s.Count += o.Count
153 1 : s.ContainedPoints += o.ContainedPoints
154 1 : s.SkippedPoints += o.SkippedPoints
155 1 : }
156 :
157 : // LazyValue is a lazy value. See the long comment in base.LazyValue.
158 : type LazyValue = base.LazyValue
159 :
160 : // Iterator iterates over a DB's key/value pairs in key order.
161 : //
162 : // An iterator must be closed after use, but it is not necessary to read an
163 : // iterator until exhaustion.
164 : //
165 : // An iterator is not goroutine-safe, but it is safe to use multiple iterators
166 : // concurrently, with each in a dedicated goroutine.
167 : //
168 : // It is also safe to use an iterator concurrently with modifying its
169 : // underlying DB, if that DB permits modification. However, the resultant
170 : // key/value pairs are not guaranteed to be a consistent snapshot of that DB
171 : // at a particular point in time.
172 : //
173 : // If an iterator encounters an error during any operation, it is stored by
174 : // the Iterator and surfaced through the Error method. All absolute
175 : // positioning methods (eg, SeekLT, SeekGT, First, Last, etc) reset any
176 : // accumulated error before positioning. All relative positioning methods (eg,
177 : // Next, Prev) return without advancing if the iterator has an accumulated
178 : // error.
179 : type Iterator struct {
180 : // The context is stored here since (a) Iterators are expected to be
181 : // short-lived (since they pin memtables and sstables), (b) plumbing a
182 : // context into every method is very painful, (c) they do not (yet) respect
183 : // context cancellation and are only used for tracing.
184 : ctx context.Context
185 : opts IterOptions
186 : merge Merge
187 : comparer base.Comparer
188 : iter internalIterator
189 : pointIter internalIterator
190 : // Either readState or version is set, but not both.
191 : readState *readState
192 : version *version
193 : // rangeKey holds iteration state specific to iteration over range keys.
194 : // The range key field may be nil if the Iterator has never been configured
195 : // to iterate over range keys. Its non-nilness cannot be used to determine
196 : // if the Iterator is currently iterating over range keys: For that, consult
197 : // the IterOptions using opts.rangeKeys(). If non-nil, its rangeKeyIter
198 : // field is guaranteed to be non-nil too.
199 : rangeKey *iteratorRangeKeyState
200 : // rangeKeyMasking holds state for range-key masking of point keys.
201 : rangeKeyMasking rangeKeyMasking
202 : err error
203 : // When iterValidityState=IterValid, key represents the current key, which
204 : // is backed by keyBuf.
205 : key []byte
206 : keyBuf []byte
207 : value LazyValue
208 : // For use in LazyValue.Clone.
209 : valueBuf []byte
210 : fetcher base.LazyFetcher
211 : // For use in LazyValue.Value.
212 : lazyValueBuf []byte
213 : valueCloser io.Closer
214 : // boundsBuf holds two buffers used to store the lower and upper bounds.
215 : // Whenever the Iterator's bounds change, the new bounds are copied into
216 : // boundsBuf[boundsBufIdx]. The two bounds share a slice to reduce
217 : // allocations. opts.LowerBound and opts.UpperBound point into this slice.
218 : boundsBuf [2][]byte
219 : boundsBufIdx int
220 : // iterKey, iterValue reflect the latest position of iter, except when
221 : // SetBounds is called. In that case, these are explicitly set to nil.
222 : iterKey *InternalKey
223 : iterValue LazyValue
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 keyspan.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 : // Used for an optimization in external iterators to reduce the number of
303 : // merging levels.
304 : forwardOnly bool
305 : // batchOnlyIter is set to true for Batch.NewBatchOnlyIter.
306 : batchOnlyIter bool
307 : // closePointIterOnce is set to true if this point iter can only be Close()d
308 : // once, _and_ closing i.iter and then i.pointIter would close i.pointIter
309 : // twice. This is necessary to track if the point iter is an internal iterator
310 : // that could release its resources to a pool on Close(), making it harder for
311 : // that iterator to make its own closes idempotent.
312 : //
313 : // TODO(bilal): Update SetOptions to always close out point key iterators when
314 : // they won't be used, so that Close() doesn't need to default to closing
315 : // point iterators twice.
316 : closePointIterOnce bool
317 : // Used in some tests to disable the random disabling of seek optimizations.
318 : forceEnableSeekOpt bool
319 : // Set to true if NextPrefix is not currently permitted. Defaults to false
320 : // in case an iterator never had any bounds.
321 : nextPrefixNotPermittedByUpperBound bool
322 : }
323 :
324 : // cmp is a convenience shorthand for the i.comparer.Compare function.
325 1 : func (i *Iterator) cmp(a, b []byte) int {
326 1 : return i.comparer.Compare(a, b)
327 1 : }
328 :
329 : // split is a convenience shorthand for the i.comparer.Split function.
330 1 : func (i *Iterator) split(a []byte) int {
331 1 : return i.comparer.Split(a)
332 1 : }
333 :
334 : // equal is a convenience shorthand for the i.comparer.Equal function.
335 1 : func (i *Iterator) equal(a, b []byte) bool {
336 1 : return i.comparer.Equal(a, b)
337 1 : }
338 :
339 : // iteratorRangeKeyState holds an iterator's range key iteration state.
340 : type iteratorRangeKeyState struct {
341 : opts *IterOptions
342 : cmp base.Compare
343 : split base.Split
344 : // rangeKeyIter holds the range key iterator stack that iterates over the
345 : // merged spans across the entirety of the LSM.
346 : rangeKeyIter keyspan.FragmentIterator
347 : iiter keyspan.InterleavingIter
348 : // stale is set to true when the range key state recorded here (in start,
349 : // end and keys) may not be in sync with the current range key at the
350 : // interleaving iterator's current position.
351 : //
352 : // When the interelaving iterator passes over a new span, it invokes the
353 : // SpanChanged hook defined on the `rangeKeyMasking` type, which sets stale
354 : // to true if the span is non-nil.
355 : //
356 : // The parent iterator may not be positioned over the interleaving
357 : // iterator's current position (eg, i.iterPos = iterPos{Next,Prev}), so
358 : // {keys,start,end} are only updated to the new range key during a call to
359 : // Iterator.saveRangeKey.
360 : stale bool
361 : // updated is used to signal to the Iterator client whether the state of
362 : // range keys has changed since the previous iterator position through the
363 : // `RangeKeyChanged` method. It's set to true during an Iterator positioning
364 : // operation that changes the state of the current range key. Each Iterator
365 : // positioning operation sets it back to false before executing.
366 : //
367 : // TODO(jackson): The lifecycle of {stale,updated,prevPosHadRangeKey} is
368 : // intricate and confusing. Try to refactor to reduce complexity.
369 : updated bool
370 : // prevPosHadRangeKey records whether the previous Iterator position had a
371 : // range key (HasPointAndRage() = (_, true)). It's updated at the beginning
372 : // of each new Iterator positioning operation. It's required by saveRangeKey to
373 : // to set `updated` appropriately: Without this record of the previous iterator
374 : // state, it's ambiguous whether an iterator only temporarily stepped onto a
375 : // position without a range key.
376 : prevPosHadRangeKey bool
377 : // rangeKeyOnly is set to true if at the current iterator position there is
378 : // no point key, only a range key start boundary.
379 : rangeKeyOnly bool
380 : // hasRangeKey is true when the current iterator position has a covering
381 : // range key (eg, a range key with bounds [<lower>,<upper>) such that
382 : // <lower> ≤ Key() < <upper>).
383 : hasRangeKey bool
384 : // start and end are the [start, end) boundaries of the current range keys.
385 : start []byte
386 : end []byte
387 :
388 : rangeKeyBuffers
389 :
390 : // iterConfig holds fields that are used for the construction of the
391 : // iterator stack, but do not need to be directly accessed during iteration.
392 : // This struct is bundled within the iteratorRangeKeyState struct to reduce
393 : // allocations.
394 : iterConfig rangekey.UserIteratorConfig
395 : }
396 :
397 : type rangeKeyBuffers struct {
398 : // keys is sorted by Suffix ascending.
399 : keys []RangeKeyData
400 : // buf is used to save range-key data before moving the range-key iterator.
401 : // Start and end boundaries, suffixes and values are all copied into buf.
402 : buf bytealloc.A
403 : // internal holds buffers used by the range key internal iterators.
404 : internal rangekey.Buffers
405 : }
406 :
407 1 : func (b *rangeKeyBuffers) PrepareForReuse() {
408 1 : const maxKeysReuse = 100
409 1 : if len(b.keys) > maxKeysReuse {
410 0 : b.keys = nil
411 0 : }
412 : // Avoid caching the key buf if it is overly large. The constant is
413 : // fairly arbitrary.
414 1 : if cap(b.buf) >= maxKeyBufCacheSize {
415 1 : b.buf = nil
416 1 : } else {
417 1 : b.buf = b.buf[:0]
418 1 : }
419 1 : b.internal.PrepareForReuse()
420 : }
421 :
422 1 : func (i *iteratorRangeKeyState) init(cmp base.Compare, split base.Split, opts *IterOptions) {
423 1 : i.cmp = cmp
424 1 : i.split = split
425 1 : i.opts = opts
426 1 : }
427 :
428 : var iterRangeKeyStateAllocPool = sync.Pool{
429 1 : New: func() interface{} {
430 1 : return &iteratorRangeKeyState{}
431 1 : },
432 : }
433 :
434 : // isEphemeralPosition returns true iff the current iterator position is
435 : // ephemeral, and won't be visited during subsequent relative positioning
436 : // operations.
437 : //
438 : // The iterator position resulting from a SeekGE or SeekPrefixGE that lands on a
439 : // straddling range key without a coincident point key is such a position.
440 1 : func (i *Iterator) isEphemeralPosition() bool {
441 1 : return i.opts.rangeKeys() && i.rangeKey != nil && i.rangeKey.rangeKeyOnly &&
442 1 : !i.equal(i.rangeKey.start, i.key)
443 1 : }
444 :
445 : type lastPositioningOpKind int8
446 :
447 : const (
448 : unknownLastPositionOp lastPositioningOpKind = iota
449 : seekPrefixGELastPositioningOp
450 : seekGELastPositioningOp
451 : seekLTLastPositioningOp
452 : // invalidatedLastPositionOp is similar to unknownLastPositionOp and the
453 : // only reason to distinguish this is for the wider set of SeekGE
454 : // optimizations we permit for the external iterator Iterator.forwardOnly
455 : // case. Most code predicates should be doing equality comparisons with one
456 : // of the seek* enum values, so this duplication should not result in code
457 : // of the form:
458 : // if unknownLastPositionOp || invalidLastPositionOp
459 : invalidatedLastPositionOp
460 : )
461 :
462 : // Limited iteration mode. Not for use with prefix iteration.
463 : //
464 : // SeekGE, SeekLT, Prev, Next have WithLimit variants, that pause the iterator
465 : // at the limit in a best-effort manner. The client should behave correctly
466 : // even if the limits are ignored. These limits are not "deep", in that they
467 : // are not passed down to the underlying collection of internalIterators. This
468 : // is because the limits are transient, and apply only until the next
469 : // iteration call. They serve mainly as a way to bound the amount of work when
470 : // two (or more) Iterators are being coordinated at a higher level.
471 : //
472 : // In limited iteration mode:
473 : // - Avoid using Iterator.Valid if the last call was to a *WithLimit() method.
474 : // The return value from the *WithLimit() method provides a more precise
475 : // disposition.
476 : // - The limit is exclusive for forward and inclusive for reverse.
477 : //
478 : //
479 : // Limited iteration mode & range keys
480 : //
481 : // Limited iteration interacts with range-key iteration. When range key
482 : // iteration is enabled, range keys are interleaved at their start boundaries.
483 : // Limited iteration must ensure that if a range key exists within the limit,
484 : // the iterator visits the range key.
485 : //
486 : // During forward limited iteration, this is trivial: An overlapping range key
487 : // must have a start boundary less than the limit, and the range key's start
488 : // boundary will be interleaved and found to be within the limit.
489 : //
490 : // During reverse limited iteration, the tail of the range key may fall within
491 : // the limit. The range key must be surfaced even if the range key's start
492 : // boundary is less than the limit, and if there are no point keys between the
493 : // current iterator position and the limit. To provide this guarantee, reverse
494 : // limited iteration ignores the limit as long as there is a range key
495 : // overlapping the iteration position.
496 :
497 : // IterValidityState captures the state of the Iterator.
498 : type IterValidityState int8
499 :
500 : const (
501 : // IterExhausted represents an Iterator that is exhausted.
502 : IterExhausted IterValidityState = iota
503 : // IterValid represents an Iterator that is valid.
504 : IterValid
505 : // IterAtLimit represents an Iterator that has a non-exhausted
506 : // internalIterator, but has reached a limit without any key for the
507 : // caller.
508 : IterAtLimit
509 : )
510 :
511 : // readSampling stores variables used to sample a read to trigger a read
512 : // compaction
513 : type readSampling struct {
514 : bytesUntilReadSampling uint64
515 : initialSamplePassed bool
516 : pendingCompactions readCompactionQueue
517 : // forceReadSampling is used for testing purposes to force a read sample on every
518 : // call to Iterator.maybeSampleRead()
519 : forceReadSampling bool
520 : }
521 :
522 1 : func (i *Iterator) findNextEntry(limit []byte) {
523 1 : i.iterValidityState = IterExhausted
524 1 : i.pos = iterPosCurForward
525 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
526 1 : i.rangeKey.rangeKeyOnly = false
527 1 : }
528 :
529 : // Close the closer for the current value if one was open.
530 1 : if i.closeValueCloser() != nil {
531 0 : return
532 0 : }
533 :
534 1 : for i.iterKey != nil {
535 1 : key := *i.iterKey
536 1 :
537 1 : if i.hasPrefix {
538 1 : if n := i.split(key.UserKey); !i.equal(i.prefixOrFullSeekKey, key.UserKey[:n]) {
539 1 : return
540 1 : }
541 : }
542 : // Compare with limit every time we start at a different user key.
543 : // Note that given the best-effort contract of limit, we could avoid a
544 : // comparison in the common case by doing this only after
545 : // i.nextUserKey is called for the deletes below. However that makes
546 : // the behavior non-deterministic (since the behavior will vary based
547 : // on what has been compacted), which makes it hard to test with the
548 : // metamorphic test. So we forego that performance optimization.
549 1 : if limit != nil && i.cmp(limit, i.iterKey.UserKey) <= 0 {
550 1 : i.iterValidityState = IterAtLimit
551 1 : i.pos = iterPosCurForwardPaused
552 1 : return
553 1 : }
554 :
555 : // If the user has configured a SkipPoint function, invoke it to see
556 : // whether we should skip over the current user key.
557 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(i.iterKey.UserKey) {
558 1 : // NB: We could call nextUserKey, but in some cases the SkipPoint
559 1 : // predicate function might be cheaper than nextUserKey's key copy
560 1 : // and key comparison. This should be the case for MVCC suffix
561 1 : // comparisons, for example. In the future, we could expand the
562 1 : // SkipPoint interface to give the implementor more control over
563 1 : // whether we skip over just the internal key, the user key, or even
564 1 : // the key prefix.
565 1 : i.stats.ForwardStepCount[InternalIterCall]++
566 1 : i.iterKey, i.iterValue = i.iter.Next()
567 1 : continue
568 : }
569 :
570 1 : switch key.Kind() {
571 1 : case InternalKeyKindRangeKeySet:
572 1 : // Save the current key.
573 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
574 1 : i.key = i.keyBuf
575 1 : i.value = LazyValue{}
576 1 : // There may also be a live point key at this userkey that we have
577 1 : // not yet read. We need to find the next entry with this user key
578 1 : // to find it. Save the range key so we don't lose it when we Next
579 1 : // the underlying iterator.
580 1 : i.saveRangeKey()
581 1 : pointKeyExists := i.nextPointCurrentUserKey()
582 1 : if i.err != nil {
583 0 : i.iterValidityState = IterExhausted
584 0 : return
585 0 : }
586 1 : i.rangeKey.rangeKeyOnly = !pointKeyExists
587 1 : i.iterValidityState = IterValid
588 1 : return
589 :
590 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
591 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
592 1 : // only simpler, but is also necessary for correctness due to
593 1 : // InternalKeyKindSSTableInternalObsoleteBit.
594 1 : i.nextUserKey()
595 1 : continue
596 :
597 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
598 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
599 1 : i.key = i.keyBuf
600 1 : i.value = i.iterValue
601 1 : i.iterValidityState = IterValid
602 1 : i.saveRangeKey()
603 1 : return
604 :
605 1 : case InternalKeyKindMerge:
606 1 : // Resolving the merge may advance us to the next point key, which
607 1 : // may be covered by a different set of range keys. Save the range
608 1 : // key state so we don't lose it.
609 1 : i.saveRangeKey()
610 1 : if i.mergeForward(key) {
611 1 : i.iterValidityState = IterValid
612 1 : return
613 1 : }
614 :
615 : // The merge didn't yield a valid key, either because the value
616 : // merger indicated it should be deleted, or because an error was
617 : // encountered.
618 1 : i.iterValidityState = IterExhausted
619 1 : if i.err != nil {
620 0 : return
621 0 : }
622 1 : if i.pos != iterPosNext {
623 0 : i.nextUserKey()
624 0 : }
625 1 : if i.closeValueCloser() != nil {
626 0 : return
627 0 : }
628 1 : i.pos = iterPosCurForward
629 :
630 0 : default:
631 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
632 0 : i.iterValidityState = IterExhausted
633 0 : return
634 : }
635 : }
636 : }
637 :
638 1 : func (i *Iterator) nextPointCurrentUserKey() bool {
639 1 : // If the user has configured a SkipPoint function and the current user key
640 1 : // would be skipped by it, there's no need to step forward looking for a
641 1 : // point key. If we were to find one, it should be skipped anyways.
642 1 : if i.opts.SkipPoint != nil && i.opts.SkipPoint(i.key) {
643 1 : return false
644 1 : }
645 :
646 1 : i.pos = iterPosCurForward
647 1 :
648 1 : i.iterKey, i.iterValue = i.iter.Next()
649 1 : i.stats.ForwardStepCount[InternalIterCall]++
650 1 : if i.iterKey == nil || !i.equal(i.key, i.iterKey.UserKey) {
651 1 : i.pos = iterPosNext
652 1 : return false
653 1 : }
654 :
655 1 : key := *i.iterKey
656 1 : switch key.Kind() {
657 0 : case InternalKeyKindRangeKeySet:
658 0 : // RangeKeySets must always be interleaved as the first internal key
659 0 : // for a user key.
660 0 : i.err = base.CorruptionErrorf("pebble: unexpected range key set mid-user key")
661 0 : return false
662 :
663 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
664 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
665 1 : // only simpler, but is also necessary for correctness due to
666 1 : // InternalKeyKindSSTableInternalObsoleteBit.
667 1 : return false
668 :
669 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
670 1 : i.value = i.iterValue
671 1 : return true
672 :
673 1 : case InternalKeyKindMerge:
674 1 : return i.mergeForward(key)
675 :
676 0 : default:
677 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
678 0 : return false
679 : }
680 : }
681 :
682 : // mergeForward resolves a MERGE key, advancing the underlying iterator forward
683 : // to merge with subsequent keys with the same userkey. mergeForward returns a
684 : // boolean indicating whether or not the merge yielded a valid key. A merge may
685 : // not yield a valid key if an error occurred, in which case i.err is non-nil,
686 : // or the user's value merger specified the key to be deleted.
687 : //
688 : // mergeForward does not update iterValidityState.
689 1 : func (i *Iterator) mergeForward(key base.InternalKey) (valid bool) {
690 1 : var iterValue []byte
691 1 : iterValue, _, i.err = i.iterValue.Value(nil)
692 1 : if i.err != nil {
693 0 : return false
694 0 : }
695 1 : var valueMerger ValueMerger
696 1 : valueMerger, i.err = i.merge(key.UserKey, iterValue)
697 1 : if i.err != nil {
698 0 : return false
699 0 : }
700 :
701 1 : i.mergeNext(key, valueMerger)
702 1 : if i.err != nil {
703 0 : return false
704 0 : }
705 :
706 1 : var needDelete bool
707 1 : var value []byte
708 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(
709 1 : valueMerger, true /* includesBase */)
710 1 : i.value = base.MakeInPlaceValue(value)
711 1 : if i.err != nil {
712 0 : return false
713 0 : }
714 1 : if needDelete {
715 1 : _ = i.closeValueCloser()
716 1 : return false
717 1 : }
718 1 : return true
719 : }
720 :
721 1 : func (i *Iterator) closeValueCloser() error {
722 1 : if i.valueCloser != nil {
723 0 : i.err = i.valueCloser.Close()
724 0 : i.valueCloser = nil
725 0 : }
726 1 : return i.err
727 : }
728 :
729 1 : func (i *Iterator) nextUserKey() {
730 1 : if i.iterKey == nil {
731 1 : return
732 1 : }
733 1 : trailer := i.iterKey.Trailer
734 1 : done := i.iterKey.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
735 1 : if i.iterValidityState != IterValid {
736 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKey.UserKey...)
737 1 : i.key = i.keyBuf
738 1 : }
739 1 : for {
740 1 : i.iterKey, i.iterValue = i.iter.Next()
741 1 : i.stats.ForwardStepCount[InternalIterCall]++
742 1 : // NB: We're guaranteed to be on the next user key if the previous key
743 1 : // had a zero sequence number (`done`), or the new key has a trailer
744 1 : // greater or equal to the previous key's trailer. This is true because
745 1 : // internal keys with the same user key are sorted by Trailer in
746 1 : // strictly monotonically descending order. We expect the trailer
747 1 : // optimization to trigger around 50% of the time with randomly
748 1 : // distributed writes. We expect it to trigger very frequently when
749 1 : // iterating through ingested sstables, which contain keys that all have
750 1 : // the same sequence number.
751 1 : if done || i.iterKey == nil || i.iterKey.Trailer >= trailer {
752 1 : break
753 : }
754 1 : if !i.equal(i.key, i.iterKey.UserKey) {
755 1 : break
756 : }
757 1 : done = i.iterKey.Trailer <= base.InternalKeyZeroSeqnumMaxTrailer
758 1 : trailer = i.iterKey.Trailer
759 : }
760 : }
761 :
762 1 : func (i *Iterator) maybeSampleRead() {
763 1 : // This method is only called when a public method of Iterator is
764 1 : // returning, and below we exclude the case were the iterator is paused at
765 1 : // a limit. The effect of these choices is that keys that are deleted, but
766 1 : // are encountered during iteration, are not accounted for in the read
767 1 : // sampling and will not cause read driven compactions, even though we are
768 1 : // incurring cost in iterating over them. And this issue is not limited to
769 1 : // Iterator, which does not see the effect of range deletes, which may be
770 1 : // causing iteration work in mergingIter. It is not clear at this time
771 1 : // whether this is a deficiency worth addressing.
772 1 : if i.iterValidityState != IterValid {
773 1 : return
774 1 : }
775 1 : if i.readState == nil {
776 1 : return
777 1 : }
778 1 : if i.readSampling.forceReadSampling {
779 1 : i.sampleRead()
780 1 : return
781 1 : }
782 1 : samplingPeriod := int32(int64(readBytesPeriod) * i.readState.db.opts.Experimental.ReadSamplingMultiplier)
783 1 : if samplingPeriod <= 0 {
784 0 : return
785 0 : }
786 1 : bytesRead := uint64(len(i.key) + i.value.Len())
787 1 : for i.readSampling.bytesUntilReadSampling < bytesRead {
788 1 : i.readSampling.bytesUntilReadSampling += uint64(fastrand.Uint32n(2 * uint32(samplingPeriod)))
789 1 : // The block below tries to adjust for the case where this is the
790 1 : // first read in a newly-opened iterator. As bytesUntilReadSampling
791 1 : // starts off at zero, we don't want to sample the first read of
792 1 : // every newly-opened iterator, but we do want to sample some of them.
793 1 : if !i.readSampling.initialSamplePassed {
794 1 : i.readSampling.initialSamplePassed = true
795 1 : if fastrand.Uint32n(uint32(i.readSampling.bytesUntilReadSampling)) > uint32(bytesRead) {
796 1 : continue
797 : }
798 : }
799 1 : i.sampleRead()
800 : }
801 1 : i.readSampling.bytesUntilReadSampling -= bytesRead
802 : }
803 :
804 1 : func (i *Iterator) sampleRead() {
805 1 : var topFile *manifest.FileMetadata
806 1 : topLevel, numOverlappingLevels := numLevels, 0
807 1 : mi := i.merging
808 1 : if mi == nil {
809 1 : return
810 1 : }
811 1 : if len(mi.levels) > 1 {
812 1 : mi.ForEachLevelIter(func(li *levelIter) bool {
813 1 : l := manifest.LevelToInt(li.level)
814 1 : if f := li.iterFile; f != nil {
815 1 : var containsKey bool
816 1 : if i.pos == iterPosNext || i.pos == iterPosCurForward ||
817 1 : i.pos == iterPosCurForwardPaused {
818 1 : containsKey = i.cmp(f.SmallestPointKey.UserKey, i.key) <= 0
819 1 : } else if i.pos == iterPosPrev || i.pos == iterPosCurReverse ||
820 1 : i.pos == iterPosCurReversePaused {
821 1 : containsKey = i.cmp(f.LargestPointKey.UserKey, i.key) >= 0
822 1 : }
823 : // Do nothing if the current key is not contained in f's
824 : // bounds. We could seek the LevelIterator at this level
825 : // to find the right file, but the performance impacts of
826 : // doing that are significant enough to negate the benefits
827 : // of read sampling in the first place. See the discussion
828 : // at:
829 : // https://github.com/cockroachdb/pebble/pull/1041#issuecomment-763226492
830 1 : if containsKey {
831 1 : numOverlappingLevels++
832 1 : if numOverlappingLevels >= 2 {
833 1 : // Terminate the loop early if at least 2 overlapping levels are found.
834 1 : return true
835 1 : }
836 1 : topLevel = l
837 1 : topFile = f
838 : }
839 : }
840 1 : return false
841 : })
842 : }
843 1 : if topFile == nil || topLevel >= numLevels {
844 1 : return
845 1 : }
846 1 : if numOverlappingLevels >= 2 {
847 1 : allowedSeeks := topFile.AllowedSeeks.Add(-1)
848 1 : if allowedSeeks == 0 {
849 1 :
850 1 : // Since the compaction queue can handle duplicates, we can keep
851 1 : // adding to the queue even once allowedSeeks hits 0.
852 1 : // In fact, we NEED to keep adding to the queue, because the queue
853 1 : // is small and evicts older and possibly useful compactions.
854 1 : topFile.AllowedSeeks.Add(topFile.InitAllowedSeeks)
855 1 :
856 1 : read := readCompaction{
857 1 : start: topFile.SmallestPointKey.UserKey,
858 1 : end: topFile.LargestPointKey.UserKey,
859 1 : level: topLevel,
860 1 : fileNum: topFile.FileNum,
861 1 : }
862 1 : i.readSampling.pendingCompactions.add(&read, i.cmp)
863 1 : }
864 : }
865 : }
866 :
867 1 : func (i *Iterator) findPrevEntry(limit []byte) {
868 1 : i.iterValidityState = IterExhausted
869 1 : i.pos = iterPosCurReverse
870 1 : if i.opts.rangeKeys() && i.rangeKey != nil {
871 1 : i.rangeKey.rangeKeyOnly = false
872 1 : }
873 :
874 : // Close the closer for the current value if one was open.
875 1 : if i.valueCloser != nil {
876 0 : i.err = i.valueCloser.Close()
877 0 : i.valueCloser = nil
878 0 : if i.err != nil {
879 0 : i.iterValidityState = IterExhausted
880 0 : return
881 0 : }
882 : }
883 :
884 1 : var valueMerger ValueMerger
885 1 : firstLoopIter := true
886 1 : rangeKeyBoundary := false
887 1 : // The code below compares with limit in multiple places. As documented in
888 1 : // findNextEntry, this is being done to make the behavior of limit
889 1 : // deterministic to allow for metamorphic testing. It is not required by
890 1 : // the best-effort contract of limit.
891 1 : for i.iterKey != nil {
892 1 : key := *i.iterKey
893 1 :
894 1 : // NB: We cannot pause if the current key is covered by a range key.
895 1 : // Otherwise, the user might not ever learn of a range key that covers
896 1 : // the key space being iterated over in which there are no point keys.
897 1 : // Since limits are best effort, ignoring the limit in this case is
898 1 : // allowed by the contract of limit.
899 1 : if firstLoopIter && limit != nil && i.cmp(limit, i.iterKey.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
900 1 : i.iterValidityState = IterAtLimit
901 1 : i.pos = iterPosCurReversePaused
902 1 : return
903 1 : }
904 1 : firstLoopIter = false
905 1 :
906 1 : if i.iterValidityState == IterValid {
907 1 : if !i.equal(key.UserKey, i.key) {
908 1 : // We've iterated to the previous user key.
909 1 : i.pos = iterPosPrev
910 1 : if valueMerger != nil {
911 1 : var needDelete bool
912 1 : var value []byte
913 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
914 1 : i.value = base.MakeInPlaceValue(value)
915 1 : if i.err == nil && needDelete {
916 1 : // The point key at this key is deleted. If we also have
917 1 : // a range key boundary at this key, we still want to
918 1 : // return. Otherwise, we need to continue looking for
919 1 : // a live key.
920 1 : i.value = LazyValue{}
921 1 : if rangeKeyBoundary {
922 0 : i.rangeKey.rangeKeyOnly = true
923 1 : } else {
924 1 : i.iterValidityState = IterExhausted
925 1 : if i.closeValueCloser() == nil {
926 1 : continue
927 : }
928 : }
929 : }
930 : }
931 1 : if i.err != nil {
932 0 : i.iterValidityState = IterExhausted
933 0 : }
934 1 : return
935 : }
936 : }
937 :
938 : // If the user has configured a SkipPoint function, invoke it to see
939 : // whether we should skip over the current user key.
940 1 : if i.opts.SkipPoint != nil && key.Kind() != InternalKeyKindRangeKeySet && i.opts.SkipPoint(key.UserKey) {
941 1 : // NB: We could call prevUserKey, but in some cases the SkipPoint
942 1 : // predicate function might be cheaper than prevUserKey's key copy
943 1 : // and key comparison. This should be the case for MVCC suffix
944 1 : // comparisons, for example. In the future, we could expand the
945 1 : // SkipPoint interface to give the implementor more control over
946 1 : // whether we skip over just the internal key, the user key, or even
947 1 : // the key prefix.
948 1 : i.stats.ReverseStepCount[InternalIterCall]++
949 1 : i.iterKey, i.iterValue = i.iter.Prev()
950 1 : if limit != nil && i.iterKey != nil && i.cmp(limit, i.iterKey.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
951 1 : i.iterValidityState = IterAtLimit
952 1 : i.pos = iterPosCurReversePaused
953 1 : return
954 1 : }
955 1 : continue
956 : }
957 :
958 1 : switch key.Kind() {
959 1 : case InternalKeyKindRangeKeySet:
960 1 : // Range key start boundary markers are interleaved with the maximum
961 1 : // sequence number, so if there's a point key also at this key, we
962 1 : // must've already iterated over it.
963 1 : // This is the final entry at this user key, so we may return
964 1 : i.rangeKey.rangeKeyOnly = i.iterValidityState != IterValid
965 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
966 1 : i.key = i.keyBuf
967 1 : i.iterValidityState = IterValid
968 1 : i.saveRangeKey()
969 1 : // In all other cases, previous iteration requires advancing to
970 1 : // iterPosPrev in order to determine if the key is live and
971 1 : // unshadowed by another key at the same user key. In this case,
972 1 : // because range key start boundary markers are always interleaved
973 1 : // at the maximum sequence number, we know that there aren't any
974 1 : // additional keys with the same user key in the backward direction.
975 1 : //
976 1 : // We Prev the underlying iterator once anyways for consistency, so
977 1 : // that we can maintain the invariant during backward iteration that
978 1 : // i.iterPos = iterPosPrev.
979 1 : i.stats.ReverseStepCount[InternalIterCall]++
980 1 : i.iterKey, i.iterValue = i.iter.Prev()
981 1 :
982 1 : // Set rangeKeyBoundary so that on the next iteration, we know to
983 1 : // return the key even if the MERGE point key is deleted.
984 1 : rangeKeyBoundary = true
985 :
986 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
987 1 : i.value = LazyValue{}
988 1 : i.iterValidityState = IterExhausted
989 1 : valueMerger = nil
990 1 : i.iterKey, i.iterValue = i.iter.Prev()
991 1 : i.stats.ReverseStepCount[InternalIterCall]++
992 1 : // Compare with the limit. We could optimize by only checking when
993 1 : // we step to the previous user key, but detecting that requires a
994 1 : // comparison too. Note that this position may already passed a
995 1 : // number of versions of this user key, but they are all deleted, so
996 1 : // the fact that a subsequent Prev*() call will not see them is
997 1 : // harmless. Also note that this is the only place in the loop,
998 1 : // other than the firstLoopIter and SkipPoint cases above, where we
999 1 : // could step to a different user key and start processing it for
1000 1 : // returning to the caller.
1001 1 : if limit != nil && i.iterKey != nil && i.cmp(limit, i.iterKey.UserKey) > 0 && !i.rangeKeyWithinLimit(limit) {
1002 1 : i.iterValidityState = IterAtLimit
1003 1 : i.pos = iterPosCurReversePaused
1004 1 : return
1005 1 : }
1006 1 : continue
1007 :
1008 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1009 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1010 1 : i.key = i.keyBuf
1011 1 : // iterValue is owned by i.iter and could change after the Prev()
1012 1 : // call, so use valueBuf instead. Note that valueBuf is only used
1013 1 : // in this one instance; everywhere else (eg. in findNextEntry),
1014 1 : // we just point i.value to the unsafe i.iter-owned value buffer.
1015 1 : i.value, i.valueBuf = i.iterValue.Clone(i.valueBuf[:0], &i.fetcher)
1016 1 : i.saveRangeKey()
1017 1 : i.iterValidityState = IterValid
1018 1 : i.iterKey, i.iterValue = i.iter.Prev()
1019 1 : i.stats.ReverseStepCount[InternalIterCall]++
1020 1 : valueMerger = nil
1021 1 : continue
1022 :
1023 1 : case InternalKeyKindMerge:
1024 1 : if i.iterValidityState == IterExhausted {
1025 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1026 1 : i.key = i.keyBuf
1027 1 : i.saveRangeKey()
1028 1 : var iterValue []byte
1029 1 : iterValue, _, i.err = i.iterValue.Value(nil)
1030 1 : if i.err != nil {
1031 0 : return
1032 0 : }
1033 1 : valueMerger, i.err = i.merge(i.key, iterValue)
1034 1 : if i.err != nil {
1035 0 : return
1036 0 : }
1037 1 : i.iterValidityState = IterValid
1038 1 : } else if valueMerger == nil {
1039 1 : // Extract value before iterValue since we use value before iterValue
1040 1 : // and the underlying iterator is not required to provide backing
1041 1 : // memory for both simultaneously.
1042 1 : var value []byte
1043 1 : var callerOwned bool
1044 1 : value, callerOwned, i.err = i.value.Value(i.lazyValueBuf)
1045 1 : if callerOwned {
1046 0 : i.lazyValueBuf = value[:0]
1047 0 : }
1048 1 : if i.err != nil {
1049 0 : return
1050 0 : }
1051 1 : valueMerger, i.err = i.merge(i.key, value)
1052 1 : var iterValue []byte
1053 1 : iterValue, _, i.err = i.iterValue.Value(nil)
1054 1 : if i.err != nil {
1055 0 : return
1056 0 : }
1057 1 : if i.err == nil {
1058 1 : i.err = valueMerger.MergeNewer(iterValue)
1059 1 : }
1060 1 : if i.err != nil {
1061 0 : i.iterValidityState = IterExhausted
1062 0 : return
1063 0 : }
1064 1 : } else {
1065 1 : var iterValue []byte
1066 1 : iterValue, _, i.err = i.iterValue.Value(nil)
1067 1 : if i.err != nil {
1068 0 : return
1069 0 : }
1070 1 : i.err = valueMerger.MergeNewer(iterValue)
1071 1 : if i.err != nil {
1072 0 : i.iterValidityState = IterExhausted
1073 0 : return
1074 0 : }
1075 : }
1076 1 : i.iterKey, i.iterValue = i.iter.Prev()
1077 1 : i.stats.ReverseStepCount[InternalIterCall]++
1078 1 : continue
1079 :
1080 0 : default:
1081 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1082 0 : i.iterValidityState = IterExhausted
1083 0 : return
1084 : }
1085 : }
1086 :
1087 : // i.iterKey == nil, so broke out of the preceding loop.
1088 1 : if i.iterValidityState == IterValid {
1089 1 : i.pos = iterPosPrev
1090 1 : if valueMerger != nil {
1091 1 : var needDelete bool
1092 1 : var value []byte
1093 1 : value, needDelete, i.valueCloser, i.err = finishValueMerger(valueMerger, true /* includesBase */)
1094 1 : i.value = base.MakeInPlaceValue(value)
1095 1 : if i.err == nil && needDelete {
1096 1 : i.key = nil
1097 1 : i.value = LazyValue{}
1098 1 : i.iterValidityState = IterExhausted
1099 1 : }
1100 : }
1101 1 : if i.err != nil {
1102 0 : i.iterValidityState = IterExhausted
1103 0 : }
1104 : }
1105 : }
1106 :
1107 1 : func (i *Iterator) prevUserKey() {
1108 1 : if i.iterKey == nil {
1109 1 : return
1110 1 : }
1111 1 : if i.iterValidityState != IterValid {
1112 1 : // If we're going to compare against the prev key, we need to save the
1113 1 : // current key.
1114 1 : i.keyBuf = append(i.keyBuf[:0], i.iterKey.UserKey...)
1115 1 : i.key = i.keyBuf
1116 1 : }
1117 1 : for {
1118 1 : i.iterKey, i.iterValue = i.iter.Prev()
1119 1 : i.stats.ReverseStepCount[InternalIterCall]++
1120 1 : if i.iterKey == nil {
1121 1 : break
1122 : }
1123 1 : if !i.equal(i.key, i.iterKey.UserKey) {
1124 1 : break
1125 : }
1126 : }
1127 : }
1128 :
1129 1 : func (i *Iterator) mergeNext(key InternalKey, valueMerger ValueMerger) {
1130 1 : // Save the current key.
1131 1 : i.keyBuf = append(i.keyBuf[:0], key.UserKey...)
1132 1 : i.key = i.keyBuf
1133 1 :
1134 1 : // Loop looking for older values for this key and merging them.
1135 1 : for {
1136 1 : i.iterKey, i.iterValue = i.iter.Next()
1137 1 : i.stats.ForwardStepCount[InternalIterCall]++
1138 1 : if i.iterKey == nil {
1139 1 : i.pos = iterPosNext
1140 1 : return
1141 1 : }
1142 1 : key = *i.iterKey
1143 1 : if !i.equal(i.key, key.UserKey) {
1144 1 : // We've advanced to the next key.
1145 1 : i.pos = iterPosNext
1146 1 : return
1147 1 : }
1148 1 : switch key.Kind() {
1149 1 : case InternalKeyKindDelete, InternalKeyKindSingleDelete, InternalKeyKindDeleteSized:
1150 1 : // We've hit a deletion tombstone. Return everything up to this
1151 1 : // point.
1152 1 : //
1153 1 : // NB: treating InternalKeyKindSingleDelete as equivalent to DEL is not
1154 1 : // only simpler, but is also necessary for correctness due to
1155 1 : // InternalKeyKindSSTableInternalObsoleteBit.
1156 1 : return
1157 :
1158 1 : case InternalKeyKindSet, InternalKeyKindSetWithDelete:
1159 1 : // We've hit a Set value. Merge with the existing value and return.
1160 1 : var iterValue []byte
1161 1 : iterValue, _, i.err = i.iterValue.Value(nil)
1162 1 : if i.err != nil {
1163 0 : return
1164 0 : }
1165 1 : i.err = valueMerger.MergeOlder(iterValue)
1166 1 : return
1167 :
1168 1 : case InternalKeyKindMerge:
1169 1 : // We've hit another Merge value. Merge with the existing value and
1170 1 : // continue looping.
1171 1 : var iterValue []byte
1172 1 : iterValue, _, i.err = i.iterValue.Value(nil)
1173 1 : if i.err != nil {
1174 0 : return
1175 0 : }
1176 1 : i.err = valueMerger.MergeOlder(iterValue)
1177 1 : if i.err != nil {
1178 0 : return
1179 0 : }
1180 1 : continue
1181 :
1182 0 : case InternalKeyKindRangeKeySet:
1183 0 : // The RANGEKEYSET marker must sort before a MERGE at the same user key.
1184 0 : i.err = base.CorruptionErrorf("pebble: out of order range key marker")
1185 0 : return
1186 :
1187 0 : default:
1188 0 : i.err = base.CorruptionErrorf("pebble: invalid internal key kind: %d", errors.Safe(key.Kind()))
1189 0 : return
1190 : }
1191 : }
1192 : }
1193 :
1194 : // SeekGE moves the iterator to the first key/value pair whose key is greater
1195 : // than or equal to the given key. Returns true if the iterator is pointing at
1196 : // a valid entry and false otherwise.
1197 1 : func (i *Iterator) SeekGE(key []byte) bool {
1198 1 : return i.SeekGEWithLimit(key, nil) == IterValid
1199 1 : }
1200 :
1201 : // SeekGEWithLimit moves the iterator to the first key/value pair whose key is
1202 : // greater than or equal to the given key.
1203 : //
1204 : // If limit is provided, it serves as a best-effort exclusive limit. If the
1205 : // first key greater than or equal to the given search key is also greater than
1206 : // or equal to limit, the Iterator may pause and return IterAtLimit. Because
1207 : // limits are best-effort, SeekGEWithLimit may return a key beyond limit.
1208 : //
1209 : // If the Iterator is configured to iterate over range keys, SeekGEWithLimit
1210 : // guarantees it will surface any range keys with bounds overlapping the
1211 : // keyspace [key, limit).
1212 1 : func (i *Iterator) SeekGEWithLimit(key []byte, limit []byte) IterValidityState {
1213 1 : if i.rangeKey != nil {
1214 1 : // NB: Check Valid() before clearing requiresReposition.
1215 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1216 1 : // If we have a range key but did not expose it at the previous iterator
1217 1 : // position (because the iterator was not at a valid position), updated
1218 1 : // must be true. This ensures that after an iterator op sequence like:
1219 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1220 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1221 1 : // - SeekGE(...) → (IterValid, RangeBounds() = [a,b))
1222 1 : // the iterator returns RangeKeyChanged()=true.
1223 1 : //
1224 1 : // The remainder of this function will only update i.rangeKey.updated if
1225 1 : // the iterator moves into a new range key, or out of the current range
1226 1 : // key.
1227 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1228 1 : }
1229 1 : lastPositioningOp := i.lastPositioningOp
1230 1 : hasPrefix := i.hasPrefix
1231 1 : // Set it to unknown, since this operation may not succeed, in which case
1232 1 : // the SeekGE following this should not make any assumption about iterator
1233 1 : // position.
1234 1 : i.lastPositioningOp = unknownLastPositionOp
1235 1 : i.requiresReposition = false
1236 1 : i.err = nil // clear cached iteration error
1237 1 : i.hasPrefix = false
1238 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1239 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1240 1 : key = lowerBound
1241 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1242 1 : key = upperBound
1243 1 : }
1244 1 : seekInternalIter := true
1245 1 :
1246 1 : var flags base.SeekGEFlags
1247 1 : if i.batchJustRefreshed {
1248 1 : i.batchJustRefreshed = false
1249 1 : flags = flags.EnableBatchJustRefreshed()
1250 1 : }
1251 1 : if lastPositioningOp == seekGELastPositioningOp {
1252 1 : cmp := i.cmp(i.prefixOrFullSeekKey, key)
1253 1 : // If this seek is to the same or later key, and the iterator is
1254 1 : // already positioned there, this is a noop. This can be helpful for
1255 1 : // sparse key spaces that have many deleted keys, where one can avoid
1256 1 : // the overhead of iterating past them again and again.
1257 1 : if cmp <= 0 {
1258 1 : if !flags.BatchJustRefreshed() &&
1259 1 : (i.iterValidityState == IterExhausted ||
1260 1 : (i.iterValidityState == IterValid && i.cmp(key, i.key) <= 0 &&
1261 1 : (limit == nil || i.cmp(i.key, limit) < 0))) {
1262 1 : // Noop
1263 1 : if !invariants.Enabled || !disableSeekOpt(key, uintptr(unsafe.Pointer(i))) || i.forceEnableSeekOpt {
1264 1 : i.lastPositioningOp = seekGELastPositioningOp
1265 1 : return i.iterValidityState
1266 1 : }
1267 : }
1268 : // cmp == 0 is not safe to optimize since
1269 : // - i.pos could be at iterPosNext, due to a merge.
1270 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1271 : // SET pair for a key, and the iterator would have moved past DELETE
1272 : // but stayed at iterPosCurForward. A similar situation occurs for a
1273 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1274 : // at the SET.
1275 : // We also leverage the IterAtLimit <=> i.pos invariant defined in the
1276 : // comment on iterValidityState, to exclude any cases where i.pos
1277 : // is iterPosCur{Forward,Reverse}Paused. This avoids the need to
1278 : // special-case those iterator positions and their interactions with
1279 : // TrySeekUsingNext, as the main uses for TrySeekUsingNext in CockroachDB
1280 : // do not use limited Seeks in the first place.
1281 1 : if cmp < 0 && i.iterValidityState != IterAtLimit && limit == nil {
1282 1 : flags = flags.EnableTrySeekUsingNext()
1283 1 : }
1284 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !i.forceEnableSeekOpt && disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1285 1 : flags = flags.DisableTrySeekUsingNext()
1286 1 : }
1287 1 : if !flags.BatchJustRefreshed() && i.pos == iterPosCurForwardPaused && i.cmp(key, i.iterKey.UserKey) <= 0 {
1288 1 : // Have some work to do, but don't need to seek, and we can
1289 1 : // start doing findNextEntry from i.iterKey.
1290 1 : seekInternalIter = false
1291 1 : }
1292 : }
1293 : }
1294 : // Check for another TrySeekUsingNext optimization opportunity, currently
1295 : // specifically tailored to external iterators. This case is intended to
1296 : // trigger in instances of Seek-ing with monotonically increasing keys with
1297 : // Nexts interspersed. At the time of writing, this is the case for
1298 : // CockroachDB scans. This optimization is important for external iterators
1299 : // to avoid re-seeking within an already-exhausted sstable. It is not always
1300 : // a performance win more generally, so we restrict it to external iterators
1301 : // that are configured to only use forward positioning operations.
1302 : //
1303 : // TODO(jackson): This optimization should be obsolete once we introduce and
1304 : // use the NextPrefix iterator positioning operation.
1305 1 : if seekInternalIter && i.forwardOnly && lastPositioningOp != invalidatedLastPositionOp &&
1306 1 : i.pos == iterPosCurForward && !hasPrefix && i.iterValidityState == IterValid &&
1307 1 : i.cmp(key, i.iterKey.UserKey) > 0 {
1308 1 : flags = flags.EnableTrySeekUsingNext()
1309 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !i.forceEnableSeekOpt && disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1310 0 : flags = flags.DisableTrySeekUsingNext()
1311 0 : }
1312 : }
1313 1 : if seekInternalIter {
1314 1 : i.iterKey, i.iterValue = i.iter.SeekGE(key, flags)
1315 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1316 1 : }
1317 1 : i.findNextEntry(limit)
1318 1 : i.maybeSampleRead()
1319 1 : if i.Error() == nil {
1320 1 : // Prepare state for a future noop optimization.
1321 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1322 1 : i.lastPositioningOp = seekGELastPositioningOp
1323 1 : }
1324 1 : return i.iterValidityState
1325 : }
1326 :
1327 : // SeekPrefixGE moves the iterator to the first key/value pair whose key is
1328 : // greater than or equal to the given key and which has the same "prefix" as
1329 : // the given key. The prefix for a key is determined by the user-defined
1330 : // Comparer.Split function. The iterator will not observe keys not matching the
1331 : // "prefix" of the search key. Calling SeekPrefixGE puts the iterator in prefix
1332 : // iteration mode. The iterator remains in prefix iteration until a subsequent
1333 : // call to another absolute positioning method (SeekGE, SeekLT, First,
1334 : // Last). Reverse iteration (Prev) is not supported when an iterator is in
1335 : // prefix iteration mode. Returns true if the iterator is pointing at a valid
1336 : // entry and false otherwise.
1337 : //
1338 : // The semantics of SeekPrefixGE are slightly unusual and designed for
1339 : // iteration to be able to take advantage of bloom filters that have been
1340 : // created on the "prefix". If you're not using bloom filters, there is no
1341 : // reason to use SeekPrefixGE.
1342 : //
1343 : // An example Split function may separate a timestamp suffix from the prefix of
1344 : // the key.
1345 : //
1346 : // Split(<key>@<timestamp>) -> <key>
1347 : //
1348 : // Consider the keys "a@1", "a@2", "aa@3", "aa@4". The prefixes for these keys
1349 : // are "a", and "aa". Note that despite "a" and "aa" sharing a prefix by the
1350 : // usual definition, those prefixes differ by the definition of the Split
1351 : // function. To see how this works, consider the following set of calls on this
1352 : // data set:
1353 : //
1354 : // SeekPrefixGE("a@0") -> "a@1"
1355 : // Next() -> "a@2"
1356 : // Next() -> EOF
1357 : //
1358 : // If you're just looking to iterate over keys with a shared prefix, as
1359 : // defined by the configured comparer, set iterator bounds instead:
1360 : //
1361 : // iter := db.NewIter(&pebble.IterOptions{
1362 : // LowerBound: []byte("prefix"),
1363 : // UpperBound: []byte("prefiy"),
1364 : // })
1365 : // for iter.First(); iter.Valid(); iter.Next() {
1366 : // // Only keys beginning with "prefix" will be visited.
1367 : // }
1368 : //
1369 : // See ExampleIterator_SeekPrefixGE for a working example.
1370 : //
1371 : // When iterating with range keys enabled, all range keys encountered are
1372 : // truncated to the seek key's prefix's bounds. The truncation of the upper
1373 : // bound requires that the database's Comparer is configured with a
1374 : // ImmediateSuccessor method. For example, a SeekPrefixGE("a@9") call with the
1375 : // prefix "a" will truncate range key bounds to [a,ImmediateSuccessor(a)].
1376 1 : func (i *Iterator) SeekPrefixGE(key []byte) bool {
1377 1 : if i.rangeKey != nil {
1378 1 : // NB: Check Valid() before clearing requiresReposition.
1379 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1380 1 : // If we have a range key but did not expose it at the previous iterator
1381 1 : // position (because the iterator was not at a valid position), updated
1382 1 : // must be true. This ensures that after an iterator op sequence like:
1383 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1384 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1385 1 : // - SeekPrefixGE(...) → (IterValid, RangeBounds() = [a,b))
1386 1 : // the iterator returns RangeKeyChanged()=true.
1387 1 : //
1388 1 : // The remainder of this function will only update i.rangeKey.updated if
1389 1 : // the iterator moves into a new range key, or out of the current range
1390 1 : // key.
1391 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1392 1 : }
1393 1 : lastPositioningOp := i.lastPositioningOp
1394 1 : // Set it to unknown, since this operation may not succeed, in which case
1395 1 : // the SeekPrefixGE following this should not make any assumption about
1396 1 : // iterator position.
1397 1 : i.lastPositioningOp = unknownLastPositionOp
1398 1 : i.requiresReposition = false
1399 1 : i.err = nil // clear cached iteration error
1400 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1401 1 : if i.comparer.Split == nil {
1402 0 : panic("pebble: split must be provided for SeekPrefixGE")
1403 : }
1404 1 : if i.comparer.ImmediateSuccessor == nil && i.opts.KeyTypes != IterKeyTypePointsOnly {
1405 0 : panic("pebble: ImmediateSuccessor must be provided for SeekPrefixGE with range keys")
1406 : }
1407 1 : prefixLen := i.split(key)
1408 1 : keyPrefix := key[:prefixLen]
1409 1 : var flags base.SeekGEFlags
1410 1 : if i.batchJustRefreshed {
1411 1 : flags = flags.EnableBatchJustRefreshed()
1412 1 : i.batchJustRefreshed = false
1413 1 : }
1414 1 : if lastPositioningOp == seekPrefixGELastPositioningOp {
1415 1 : if !i.hasPrefix {
1416 0 : panic("lastPositioningOpsIsSeekPrefixGE is true, but hasPrefix is false")
1417 : }
1418 : // The iterator has not been repositioned after the last SeekPrefixGE.
1419 : // See if we are seeking to a larger key, since then we can optimize
1420 : // the seek by using next. Note that we could also optimize if Next
1421 : // has been called, if the iterator is not exhausted and the current
1422 : // position is <= the seek key. We are keeping this limited for now
1423 : // since such optimizations require care for correctness, and to not
1424 : // become de-optimizations (if one usually has to do all the next
1425 : // calls and then the seek). This SeekPrefixGE optimization
1426 : // specifically benefits CockroachDB.
1427 1 : cmp := i.cmp(i.prefixOrFullSeekKey, keyPrefix)
1428 1 : // cmp == 0 is not safe to optimize since
1429 1 : // - i.pos could be at iterPosNext, due to a merge.
1430 1 : // - Even if i.pos were at iterPosCurForward, we could have a DELETE,
1431 1 : // SET pair for a key, and the iterator would have moved past DELETE
1432 1 : // but stayed at iterPosCurForward. A similar situation occurs for a
1433 1 : // MERGE, SET pair where the MERGE is consumed and the iterator is
1434 1 : // at the SET.
1435 1 : // In general some versions of i.prefix could have been consumed by
1436 1 : // the iterator, so we only optimize for cmp < 0.
1437 1 : if cmp < 0 {
1438 1 : flags = flags.EnableTrySeekUsingNext()
1439 1 : }
1440 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !i.forceEnableSeekOpt && disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1441 1 : flags = flags.DisableTrySeekUsingNext()
1442 1 : }
1443 : }
1444 : // Make a copy of the prefix so that modifications to the key after
1445 : // SeekPrefixGE returns does not affect the stored prefix.
1446 1 : if cap(i.prefixOrFullSeekKey) < prefixLen {
1447 1 : i.prefixOrFullSeekKey = make([]byte, prefixLen)
1448 1 : } else {
1449 1 : i.prefixOrFullSeekKey = i.prefixOrFullSeekKey[:prefixLen]
1450 1 : }
1451 1 : i.hasPrefix = true
1452 1 : copy(i.prefixOrFullSeekKey, keyPrefix)
1453 1 :
1454 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1455 1 : if n := i.split(lowerBound); !bytes.Equal(i.prefixOrFullSeekKey, lowerBound[:n]) {
1456 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of lower bound")
1457 1 : i.iterValidityState = IterExhausted
1458 1 : return false
1459 1 : }
1460 1 : key = lowerBound
1461 1 : } else if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1462 1 : if n := i.split(upperBound); !bytes.Equal(i.prefixOrFullSeekKey, upperBound[:n]) {
1463 1 : i.err = errors.New("pebble: SeekPrefixGE supplied with key outside of upper bound")
1464 1 : i.iterValidityState = IterExhausted
1465 1 : return false
1466 1 : }
1467 1 : key = upperBound
1468 : }
1469 1 : i.iterKey, i.iterValue = i.iter.SeekPrefixGE(i.prefixOrFullSeekKey, key, flags)
1470 1 : i.stats.ForwardSeekCount[InternalIterCall]++
1471 1 : i.findNextEntry(nil)
1472 1 : i.maybeSampleRead()
1473 1 : if i.Error() == nil {
1474 1 : i.lastPositioningOp = seekPrefixGELastPositioningOp
1475 1 : }
1476 1 : return i.iterValidityState == IterValid
1477 : }
1478 :
1479 : // Deterministic disabling of the seek optimizations. It uses the iterator
1480 : // pointer, since we want diversity in iterator behavior for the same key. Used
1481 : // for tests.
1482 1 : func disableSeekOpt(key []byte, ptr uintptr) bool {
1483 1 : // Fibonacci hash https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
1484 1 : simpleHash := (11400714819323198485 * uint64(ptr)) >> 63
1485 1 : return key != nil && key[0]&byte(1) == 0 && simpleHash == 0
1486 1 : }
1487 :
1488 : // SeekLT moves the iterator to the last key/value pair whose key is less than
1489 : // the given key. Returns true if the iterator is pointing at a valid entry and
1490 : // false otherwise.
1491 1 : func (i *Iterator) SeekLT(key []byte) bool {
1492 1 : return i.SeekLTWithLimit(key, nil) == IterValid
1493 1 : }
1494 :
1495 : // SeekLTWithLimit moves the iterator to the last key/value pair whose key is
1496 : // less than the given key.
1497 : //
1498 : // If limit is provided, it serves as a best-effort inclusive limit. If the last
1499 : // key less than the given search key is also less than limit, the Iterator may
1500 : // pause and return IterAtLimit. Because limits are best-effort, SeekLTWithLimit
1501 : // may return a key beyond limit.
1502 : //
1503 : // If the Iterator is configured to iterate over range keys, SeekLTWithLimit
1504 : // guarantees it will surface any range keys with bounds overlapping the
1505 : // keyspace up to limit.
1506 1 : func (i *Iterator) SeekLTWithLimit(key []byte, limit []byte) IterValidityState {
1507 1 : if i.rangeKey != nil {
1508 1 : // NB: Check Valid() before clearing requiresReposition.
1509 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1510 1 : // If we have a range key but did not expose it at the previous iterator
1511 1 : // position (because the iterator was not at a valid position), updated
1512 1 : // must be true. This ensures that after an iterator op sequence like:
1513 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1514 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1515 1 : // - SeekLTWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1516 1 : // the iterator returns RangeKeyChanged()=true.
1517 1 : //
1518 1 : // The remainder of this function will only update i.rangeKey.updated if
1519 1 : // the iterator moves into a new range key, or out of the current range
1520 1 : // key.
1521 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1522 1 : }
1523 1 : lastPositioningOp := i.lastPositioningOp
1524 1 : // Set it to unknown, since this operation may not succeed, in which case
1525 1 : // the SeekLT following this should not make any assumption about iterator
1526 1 : // position.
1527 1 : i.lastPositioningOp = unknownLastPositionOp
1528 1 : i.batchJustRefreshed = false
1529 1 : i.requiresReposition = false
1530 1 : i.err = nil // clear cached iteration error
1531 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1532 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil && i.cmp(key, upperBound) > 0 {
1533 1 : key = upperBound
1534 1 : } else if lowerBound := i.opts.GetLowerBound(); lowerBound != nil && i.cmp(key, lowerBound) < 0 {
1535 1 : key = lowerBound
1536 1 : }
1537 1 : i.hasPrefix = false
1538 1 : seekInternalIter := true
1539 1 : // The following noop optimization only applies when i.batch == nil, since
1540 1 : // an iterator over a batch is iterating over mutable data, that may have
1541 1 : // changed since the last seek.
1542 1 : if lastPositioningOp == seekLTLastPositioningOp && i.batch == nil {
1543 1 : cmp := i.cmp(key, i.prefixOrFullSeekKey)
1544 1 : // If this seek is to the same or earlier key, and the iterator is
1545 1 : // already positioned there, this is a noop. This can be helpful for
1546 1 : // sparse key spaces that have many deleted keys, where one can avoid
1547 1 : // the overhead of iterating past them again and again.
1548 1 : if cmp <= 0 {
1549 1 : // NB: when pos != iterPosCurReversePaused, the invariant
1550 1 : // documented earlier implies that iterValidityState !=
1551 1 : // IterAtLimit.
1552 1 : if i.iterValidityState == IterExhausted ||
1553 1 : (i.iterValidityState == IterValid && i.cmp(i.key, key) < 0 &&
1554 1 : (limit == nil || i.cmp(limit, i.key) <= 0)) {
1555 1 : if !invariants.Enabled || !disableSeekOpt(key, uintptr(unsafe.Pointer(i))) {
1556 1 : i.lastPositioningOp = seekLTLastPositioningOp
1557 1 : return i.iterValidityState
1558 1 : }
1559 : }
1560 1 : if i.pos == iterPosCurReversePaused && i.cmp(i.iterKey.UserKey, key) < 0 {
1561 1 : // Have some work to do, but don't need to seek, and we can
1562 1 : // start doing findPrevEntry from i.iterKey.
1563 1 : seekInternalIter = false
1564 1 : }
1565 : }
1566 : }
1567 1 : if seekInternalIter {
1568 1 : i.iterKey, i.iterValue = i.iter.SeekLT(key, base.SeekLTFlagsNone)
1569 1 : i.stats.ReverseSeekCount[InternalIterCall]++
1570 1 : }
1571 1 : i.findPrevEntry(limit)
1572 1 : i.maybeSampleRead()
1573 1 : if i.Error() == nil && i.batch == nil {
1574 1 : // Prepare state for a future noop optimization.
1575 1 : i.prefixOrFullSeekKey = append(i.prefixOrFullSeekKey[:0], key...)
1576 1 : i.lastPositioningOp = seekLTLastPositioningOp
1577 1 : }
1578 1 : return i.iterValidityState
1579 : }
1580 :
1581 : // First moves the iterator the the first key/value pair. Returns true if the
1582 : // iterator is pointing at a valid entry and false otherwise.
1583 1 : func (i *Iterator) First() bool {
1584 1 : if i.rangeKey != nil {
1585 1 : // NB: Check Valid() before clearing requiresReposition.
1586 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1587 1 : // If we have a range key but did not expose it at the previous iterator
1588 1 : // position (because the iterator was not at a valid position), updated
1589 1 : // must be true. This ensures that after an iterator op sequence like:
1590 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1591 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1592 1 : // - First(...) → (IterValid, RangeBounds() = [a,b))
1593 1 : // the iterator returns RangeKeyChanged()=true.
1594 1 : //
1595 1 : // The remainder of this function will only update i.rangeKey.updated if
1596 1 : // the iterator moves into a new range key, or out of the current range
1597 1 : // key.
1598 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1599 1 : }
1600 1 : i.err = nil // clear cached iteration error
1601 1 : i.hasPrefix = false
1602 1 : i.batchJustRefreshed = false
1603 1 : i.lastPositioningOp = unknownLastPositionOp
1604 1 : i.requiresReposition = false
1605 1 : i.stats.ForwardSeekCount[InterfaceCall]++
1606 1 :
1607 1 : i.iterFirstWithinBounds()
1608 1 : i.findNextEntry(nil)
1609 1 : i.maybeSampleRead()
1610 1 : return i.iterValidityState == IterValid
1611 : }
1612 :
1613 : // Last moves the iterator the the last key/value pair. Returns true if the
1614 : // iterator is pointing at a valid entry and false otherwise.
1615 1 : func (i *Iterator) Last() bool {
1616 1 : if i.rangeKey != nil {
1617 1 : // NB: Check Valid() before clearing requiresReposition.
1618 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1619 1 : // If we have a range key but did not expose it at the previous iterator
1620 1 : // position (because the iterator was not at a valid position), updated
1621 1 : // must be true. This ensures that after an iterator op sequence like:
1622 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1623 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1624 1 : // - Last(...) → (IterValid, RangeBounds() = [a,b))
1625 1 : // the iterator returns RangeKeyChanged()=true.
1626 1 : //
1627 1 : // The remainder of this function will only update i.rangeKey.updated if
1628 1 : // the iterator moves into a new range key, or out of the current range
1629 1 : // key.
1630 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1631 1 : }
1632 1 : i.err = nil // clear cached iteration error
1633 1 : i.hasPrefix = false
1634 1 : i.batchJustRefreshed = false
1635 1 : i.lastPositioningOp = unknownLastPositionOp
1636 1 : i.requiresReposition = false
1637 1 : i.stats.ReverseSeekCount[InterfaceCall]++
1638 1 :
1639 1 : i.iterLastWithinBounds()
1640 1 : i.findPrevEntry(nil)
1641 1 : i.maybeSampleRead()
1642 1 : return i.iterValidityState == IterValid
1643 : }
1644 :
1645 : // Next moves the iterator to the next key/value pair. Returns true if the
1646 : // iterator is pointing at a valid entry and false otherwise.
1647 1 : func (i *Iterator) Next() bool {
1648 1 : return i.nextWithLimit(nil) == IterValid
1649 1 : }
1650 :
1651 : // NextWithLimit moves the iterator to the next key/value pair.
1652 : //
1653 : // If limit is provided, it serves as a best-effort exclusive limit. If the next
1654 : // key is greater than or equal to limit, the Iterator may pause and return
1655 : // IterAtLimit. Because limits are best-effort, NextWithLimit may return a key
1656 : // beyond limit.
1657 : //
1658 : // If the Iterator is configured to iterate over range keys, NextWithLimit
1659 : // guarantees it will surface any range keys with bounds overlapping the
1660 : // keyspace up to limit.
1661 1 : func (i *Iterator) NextWithLimit(limit []byte) IterValidityState {
1662 1 : return i.nextWithLimit(limit)
1663 1 : }
1664 :
1665 : // NextPrefix moves the iterator to the next key/value pair with a key
1666 : // containing a different prefix than the current key. Prefixes are determined
1667 : // by Comparer.Split. Exhausts the iterator if invoked while in prefix-iteration
1668 : // mode.
1669 : //
1670 : // It is not permitted to invoke NextPrefix while at a IterAtLimit position.
1671 : // When called in this condition, NextPrefix has non-deterministic behavior.
1672 : //
1673 : // It is not permitted to invoke NextPrefix when the Iterator has an
1674 : // upper-bound that is a versioned MVCC key (see the comment for
1675 : // Comparer.Split). It returns an error in this case.
1676 1 : func (i *Iterator) NextPrefix() bool {
1677 1 : if i.nextPrefixNotPermittedByUpperBound {
1678 1 : i.lastPositioningOp = unknownLastPositionOp
1679 1 : i.requiresReposition = false
1680 1 : i.err = errors.Errorf("NextPrefix not permitted with upper bound %s",
1681 1 : i.comparer.FormatKey(i.opts.UpperBound))
1682 1 : i.iterValidityState = IterExhausted
1683 1 : return false
1684 1 : }
1685 1 : if i.hasPrefix {
1686 1 : i.iterValidityState = IterExhausted
1687 1 : return false
1688 1 : }
1689 1 : return i.nextPrefix() == IterValid
1690 : }
1691 :
1692 1 : func (i *Iterator) nextPrefix() IterValidityState {
1693 1 : if i.rangeKey != nil {
1694 1 : // NB: Check Valid() before clearing requiresReposition.
1695 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1696 1 : // If we have a range key but did not expose it at the previous iterator
1697 1 : // position (because the iterator was not at a valid position), updated
1698 1 : // must be true. This ensures that after an iterator op sequence like:
1699 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1700 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1701 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1702 1 : // the iterator returns RangeKeyChanged()=true.
1703 1 : //
1704 1 : // The remainder of this function will only update i.rangeKey.updated if
1705 1 : // the iterator moves into a new range key, or out of the current range
1706 1 : // key.
1707 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1708 1 : }
1709 :
1710 : // Although NextPrefix documents that behavior at IterAtLimit is undefined,
1711 : // this function handles these cases as a simple prefix-agnostic Next. This
1712 : // is done for deterministic behavior in the metamorphic tests.
1713 : //
1714 : // TODO(jackson): If the metamorphic test operation generator is adjusted to
1715 : // make generation of some operations conditional on the previous
1716 : // operations, then we can remove this behavior and explicitly error.
1717 :
1718 1 : i.lastPositioningOp = unknownLastPositionOp
1719 1 : i.requiresReposition = false
1720 1 : switch i.pos {
1721 1 : case iterPosCurForward:
1722 1 : // Positioned on the current key. Advance to the next prefix.
1723 1 : i.internalNextPrefix(i.split(i.key))
1724 0 : case iterPosCurForwardPaused:
1725 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1726 : // up above. The iterator is already positioned at the next key.
1727 1 : case iterPosCurReverse:
1728 1 : // Switching directions.
1729 1 : // Unless the iterator was exhausted, reverse iteration needs to
1730 1 : // position the iterator at iterPosPrev.
1731 1 : if i.iterKey != nil {
1732 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1733 0 : i.iterValidityState = IterExhausted
1734 0 : return i.iterValidityState
1735 0 : }
1736 : // The Iterator is exhausted and i.iter is positioned before the first
1737 : // key. Reposition to point to the first internal key.
1738 1 : i.iterFirstWithinBounds()
1739 0 : case iterPosCurReversePaused:
1740 0 : // Positioned at a limit. Implement as a prefix-agnostic Next. See TODO
1741 0 : // up above.
1742 0 : //
1743 0 : // Switching directions; The iterator must not be exhausted since it
1744 0 : // paused.
1745 0 : if i.iterKey == nil {
1746 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1747 0 : i.iterValidityState = IterExhausted
1748 0 : return i.iterValidityState
1749 0 : }
1750 0 : i.nextUserKey()
1751 1 : case iterPosPrev:
1752 1 : // The underlying iterator is pointed to the previous key (this can
1753 1 : // only happen when switching iteration directions).
1754 1 : if i.iterKey == nil {
1755 1 : // We're positioned before the first key. Need to reposition to point to
1756 1 : // the first key.
1757 1 : i.iterFirstWithinBounds()
1758 1 : } else {
1759 1 : // Move the internal iterator back onto the user key stored in
1760 1 : // i.key. iterPosPrev guarantees that it's positioned at the last
1761 1 : // key with the user key less than i.key, so we're guaranteed to
1762 1 : // land on the correct key with a single Next.
1763 1 : i.iterKey, i.iterValue = i.iter.Next()
1764 1 : if invariants.Enabled && !i.equal(i.iterKey.UserKey, i.key) {
1765 0 : i.opts.logger.Fatalf("pebble: invariant violation: Nexting internal iterator from iterPosPrev landed on %q, not %q",
1766 0 : i.iterKey.UserKey, i.key)
1767 0 : }
1768 : }
1769 : // The internal iterator is now positioned at i.key. Advance to the next
1770 : // prefix.
1771 1 : i.internalNextPrefix(i.split(i.key))
1772 1 : case iterPosNext:
1773 1 : // Already positioned on the next key. Only call nextPrefixKey if the
1774 1 : // next key shares the same prefix.
1775 1 : if i.iterKey != nil {
1776 1 : currKeyPrefixLen := i.split(i.key)
1777 1 : iterKeyPrefixLen := i.split(i.iterKey.UserKey)
1778 1 : if bytes.Equal(i.iterKey.UserKey[:iterKeyPrefixLen], i.key[:currKeyPrefixLen]) {
1779 1 : i.internalNextPrefix(currKeyPrefixLen)
1780 1 : }
1781 : }
1782 : }
1783 :
1784 1 : i.stats.ForwardStepCount[InterfaceCall]++
1785 1 : i.findNextEntry(nil /* limit */)
1786 1 : i.maybeSampleRead()
1787 1 : return i.iterValidityState
1788 : }
1789 :
1790 1 : func (i *Iterator) internalNextPrefix(currKeyPrefixLen int) {
1791 1 : if i.iterKey == nil {
1792 1 : return
1793 1 : }
1794 : // The Next "fast-path" is not really a fast-path when there is more than
1795 : // one version. However, even with TableFormatPebblev3, there is a small
1796 : // slowdown (~10%) for one version if we remove it and only call NextPrefix.
1797 : // When there are two versions, only calling NextPrefix is ~30% faster.
1798 1 : i.stats.ForwardStepCount[InternalIterCall]++
1799 1 : if i.iterKey, i.iterValue = i.iter.Next(); i.iterKey == nil {
1800 1 : return
1801 1 : }
1802 1 : iterKeyPrefixLen := i.split(i.iterKey.UserKey)
1803 1 : if !bytes.Equal(i.iterKey.UserKey[:iterKeyPrefixLen], i.key[:currKeyPrefixLen]) {
1804 1 : return
1805 1 : }
1806 1 : i.stats.ForwardStepCount[InternalIterCall]++
1807 1 : i.prefixOrFullSeekKey = i.comparer.ImmediateSuccessor(i.prefixOrFullSeekKey[:0], i.key[:currKeyPrefixLen])
1808 1 : i.iterKey, i.iterValue = i.iter.NextPrefix(i.prefixOrFullSeekKey)
1809 1 : if invariants.Enabled && i.iterKey != nil {
1810 1 : if iterKeyPrefixLen := i.split(i.iterKey.UserKey); i.cmp(i.iterKey.UserKey[:iterKeyPrefixLen], i.prefixOrFullSeekKey) < 0 {
1811 0 : panic(errors.AssertionFailedf("pebble: iter.NextPrefix did not advance beyond the current prefix: now at %q; expected to be geq %q",
1812 0 : i.iterKey, i.prefixOrFullSeekKey))
1813 : }
1814 : }
1815 : }
1816 :
1817 1 : func (i *Iterator) nextWithLimit(limit []byte) IterValidityState {
1818 1 : i.stats.ForwardStepCount[InterfaceCall]++
1819 1 : if i.hasPrefix {
1820 1 : if limit != nil {
1821 1 : i.err = errors.New("cannot use limit with prefix iteration")
1822 1 : i.iterValidityState = IterExhausted
1823 1 : return i.iterValidityState
1824 1 : } else if i.iterValidityState == IterExhausted {
1825 1 : // No-op, already exhasuted. We avoid executing the Next because it
1826 1 : // can break invariants: Specifically, a file that fails the bloom
1827 1 : // filter test may result in its level being removed from the
1828 1 : // merging iterator. The level's removal can cause a lazy combined
1829 1 : // iterator to miss range keys and trigger a switch to combined
1830 1 : // iteration at a larger key, breaking keyspan invariants.
1831 1 : return i.iterValidityState
1832 1 : }
1833 : }
1834 1 : if i.err != nil {
1835 1 : return i.iterValidityState
1836 1 : }
1837 1 : if i.rangeKey != nil {
1838 1 : // NB: Check Valid() before clearing requiresReposition.
1839 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1840 1 : // If we have a range key but did not expose it at the previous iterator
1841 1 : // position (because the iterator was not at a valid position), updated
1842 1 : // must be true. This ensures that after an iterator op sequence like:
1843 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1844 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1845 1 : // - NextWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1846 1 : // the iterator returns RangeKeyChanged()=true.
1847 1 : //
1848 1 : // The remainder of this function will only update i.rangeKey.updated if
1849 1 : // the iterator moves into a new range key, or out of the current range
1850 1 : // key.
1851 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1852 1 : }
1853 1 : i.lastPositioningOp = unknownLastPositionOp
1854 1 : i.requiresReposition = false
1855 1 : switch i.pos {
1856 1 : case iterPosCurForward:
1857 1 : i.nextUserKey()
1858 1 : case iterPosCurForwardPaused:
1859 : // Already at the right place.
1860 1 : case iterPosCurReverse:
1861 1 : // Switching directions.
1862 1 : // Unless the iterator was exhausted, reverse iteration needs to
1863 1 : // position the iterator at iterPosPrev.
1864 1 : if i.iterKey != nil {
1865 0 : i.err = errors.New("switching from reverse to forward but iter is not at prev")
1866 0 : i.iterValidityState = IterExhausted
1867 0 : return i.iterValidityState
1868 0 : }
1869 : // We're positioned before the first key. Need to reposition to point to
1870 : // the first key.
1871 1 : i.iterFirstWithinBounds()
1872 1 : case iterPosCurReversePaused:
1873 1 : // Switching directions.
1874 1 : // The iterator must not be exhausted since it paused.
1875 1 : if i.iterKey == nil {
1876 0 : i.err = errors.New("switching paused from reverse to forward but iter is exhausted")
1877 0 : i.iterValidityState = IterExhausted
1878 0 : return i.iterValidityState
1879 0 : }
1880 1 : i.nextUserKey()
1881 1 : case iterPosPrev:
1882 1 : // The underlying iterator is pointed to the previous key (this can
1883 1 : // only happen when switching iteration directions). We set
1884 1 : // i.iterValidityState to IterExhausted here to force the calls to
1885 1 : // nextUserKey to save the current key i.iter is pointing at in order
1886 1 : // to determine when the next user-key is reached.
1887 1 : i.iterValidityState = IterExhausted
1888 1 : if i.iterKey == nil {
1889 1 : // We're positioned before the first key. Need to reposition to point to
1890 1 : // the first key.
1891 1 : i.iterFirstWithinBounds()
1892 1 : } else {
1893 1 : i.nextUserKey()
1894 1 : }
1895 1 : i.nextUserKey()
1896 1 : case iterPosNext:
1897 : // Already at the right place.
1898 : }
1899 1 : i.findNextEntry(limit)
1900 1 : i.maybeSampleRead()
1901 1 : return i.iterValidityState
1902 : }
1903 :
1904 : // Prev moves the iterator to the previous key/value pair. Returns true if the
1905 : // iterator is pointing at a valid entry and false otherwise.
1906 1 : func (i *Iterator) Prev() bool {
1907 1 : return i.PrevWithLimit(nil) == IterValid
1908 1 : }
1909 :
1910 : // PrevWithLimit moves the iterator to the previous key/value pair.
1911 : //
1912 : // If limit is provided, it serves as a best-effort inclusive limit. If the
1913 : // previous key is less than limit, the Iterator may pause and return
1914 : // IterAtLimit. Because limits are best-effort, PrevWithLimit may return a key
1915 : // beyond limit.
1916 : //
1917 : // If the Iterator is configured to iterate over range keys, PrevWithLimit
1918 : // guarantees it will surface any range keys with bounds overlapping the
1919 : // keyspace up to limit.
1920 1 : func (i *Iterator) PrevWithLimit(limit []byte) IterValidityState {
1921 1 : i.stats.ReverseStepCount[InterfaceCall]++
1922 1 : if i.err != nil {
1923 1 : return i.iterValidityState
1924 1 : }
1925 1 : if i.rangeKey != nil {
1926 1 : // NB: Check Valid() before clearing requiresReposition.
1927 1 : i.rangeKey.prevPosHadRangeKey = i.rangeKey.hasRangeKey && i.Valid()
1928 1 : // If we have a range key but did not expose it at the previous iterator
1929 1 : // position (because the iterator was not at a valid position), updated
1930 1 : // must be true. This ensures that after an iterator op sequence like:
1931 1 : // - Next() → (IterValid, RangeBounds() = [a,b))
1932 1 : // - NextWithLimit(...) → (IterAtLimit, RangeBounds() = -)
1933 1 : // - PrevWithLimit(...) → (IterValid, RangeBounds() = [a,b))
1934 1 : // the iterator returns RangeKeyChanged()=true.
1935 1 : //
1936 1 : // The remainder of this function will only update i.rangeKey.updated if
1937 1 : // the iterator moves into a new range key, or out of the current range
1938 1 : // key.
1939 1 : i.rangeKey.updated = i.rangeKey.hasRangeKey && !i.Valid() && i.opts.rangeKeys()
1940 1 : }
1941 1 : i.lastPositioningOp = unknownLastPositionOp
1942 1 : i.requiresReposition = false
1943 1 : if i.hasPrefix {
1944 1 : i.err = errReversePrefixIteration
1945 1 : i.iterValidityState = IterExhausted
1946 1 : return i.iterValidityState
1947 1 : }
1948 1 : switch i.pos {
1949 1 : case iterPosCurForward:
1950 : // Switching directions, and will handle this below.
1951 1 : case iterPosCurForwardPaused:
1952 : // Switching directions, and will handle this below.
1953 1 : case iterPosCurReverse:
1954 1 : i.prevUserKey()
1955 1 : case iterPosCurReversePaused:
1956 : // Already at the right place.
1957 1 : case iterPosNext:
1958 : // The underlying iterator is pointed to the next key (this can only happen
1959 : // when switching iteration directions). We will handle this below.
1960 1 : case iterPosPrev:
1961 : // Already at the right place.
1962 : }
1963 1 : if i.pos == iterPosCurForward || i.pos == iterPosNext || i.pos == iterPosCurForwardPaused {
1964 1 : // Switching direction.
1965 1 : stepAgain := i.pos == iterPosNext
1966 1 :
1967 1 : // Synthetic range key markers are a special case. Consider SeekGE(b)
1968 1 : // which finds a range key [a, c). To ensure the user observes the range
1969 1 : // key, the Iterator pauses at Key() = b. The iterator must advance the
1970 1 : // internal iterator to see if there's also a coincident point key at
1971 1 : // 'b', leaving the iterator at iterPosNext if there's not.
1972 1 : //
1973 1 : // This is a problem: Synthetic range key markers are only interleaved
1974 1 : // during the original seek. A subsequent Prev() of i.iter will not move
1975 1 : // back onto the synthetic range key marker. In this case where the
1976 1 : // previous iterator position was a synthetic range key start boundary,
1977 1 : // we must not step a second time.
1978 1 : if i.isEphemeralPosition() {
1979 1 : stepAgain = false
1980 1 : }
1981 :
1982 : // We set i.iterValidityState to IterExhausted here to force the calls
1983 : // to prevUserKey to save the current key i.iter is pointing at in
1984 : // order to determine when the prev user-key is reached.
1985 1 : i.iterValidityState = IterExhausted
1986 1 : if i.iterKey == nil {
1987 1 : // We're positioned after the last key. Need to reposition to point to
1988 1 : // the last key.
1989 1 : i.iterLastWithinBounds()
1990 1 : } else {
1991 1 : i.prevUserKey()
1992 1 : }
1993 1 : if stepAgain {
1994 1 : i.prevUserKey()
1995 1 : }
1996 : }
1997 1 : i.findPrevEntry(limit)
1998 1 : i.maybeSampleRead()
1999 1 : return i.iterValidityState
2000 : }
2001 :
2002 : // iterFirstWithinBounds moves the internal iterator to the first key,
2003 : // respecting bounds.
2004 1 : func (i *Iterator) iterFirstWithinBounds() {
2005 1 : i.stats.ForwardSeekCount[InternalIterCall]++
2006 1 : if lowerBound := i.opts.GetLowerBound(); lowerBound != nil {
2007 1 : i.iterKey, i.iterValue = i.iter.SeekGE(lowerBound, base.SeekGEFlagsNone)
2008 1 : } else {
2009 1 : i.iterKey, i.iterValue = i.iter.First()
2010 1 : }
2011 : }
2012 :
2013 : // iterLastWithinBounds moves the internal iterator to the last key, respecting
2014 : // bounds.
2015 1 : func (i *Iterator) iterLastWithinBounds() {
2016 1 : i.stats.ReverseSeekCount[InternalIterCall]++
2017 1 : if upperBound := i.opts.GetUpperBound(); upperBound != nil {
2018 1 : i.iterKey, i.iterValue = i.iter.SeekLT(upperBound, base.SeekLTFlagsNone)
2019 1 : } else {
2020 1 : i.iterKey, i.iterValue = i.iter.Last()
2021 1 : }
2022 : }
2023 :
2024 : // RangeKeyData describes a range key's data, set through RangeKeySet. The key
2025 : // boundaries of the range key is provided by Iterator.RangeBounds.
2026 : type RangeKeyData struct {
2027 : Suffix []byte
2028 : Value []byte
2029 : }
2030 :
2031 : // rangeKeyWithinLimit is called during limited reverse iteration when
2032 : // positioned over a key beyond the limit. If there exists a range key that lies
2033 : // within the limit, the iterator must not pause in order to ensure the user has
2034 : // an opportunity to observe the range key within limit.
2035 : //
2036 : // It would be valid to ignore the limit whenever there's a range key covering
2037 : // the key, but that would introduce nondeterminism. To preserve determinism for
2038 : // testing, the iterator ignores the limit only if the covering range key does
2039 : // cover the keyspace within the limit.
2040 : //
2041 : // This awkwardness exists because range keys are interleaved at their inclusive
2042 : // start positions. Note that limit is inclusive.
2043 1 : func (i *Iterator) rangeKeyWithinLimit(limit []byte) bool {
2044 1 : if i.rangeKey == nil || !i.opts.rangeKeys() {
2045 1 : return false
2046 1 : }
2047 1 : s := i.rangeKey.iiter.Span()
2048 1 : // If the range key ends beyond the limit, then the range key does not cover
2049 1 : // any portion of the keyspace within the limit and it is safe to pause.
2050 1 : return s != nil && i.cmp(s.End, limit) > 0
2051 : }
2052 :
2053 : // saveRangeKey saves the current range key to the underlying iterator's current
2054 : // range key state. If the range key has not changed, saveRangeKey is a no-op.
2055 : // If there is a new range key, saveRangeKey copies all of the key, value and
2056 : // suffixes into Iterator-managed buffers.
2057 1 : func (i *Iterator) saveRangeKey() {
2058 1 : if i.rangeKey == nil || i.opts.KeyTypes == IterKeyTypePointsOnly {
2059 1 : return
2060 1 : }
2061 :
2062 1 : s := i.rangeKey.iiter.Span()
2063 1 : if s == nil {
2064 1 : i.rangeKey.hasRangeKey = false
2065 1 : i.rangeKey.updated = i.rangeKey.prevPosHadRangeKey
2066 1 : return
2067 1 : } else if !i.rangeKey.stale {
2068 1 : // The range key `s` is identical to the one currently saved. No-op.
2069 1 : return
2070 1 : }
2071 :
2072 1 : if s.KeysOrder != keyspan.BySuffixAsc {
2073 0 : panic("pebble: range key span's keys unexpectedly not in ascending suffix order")
2074 : }
2075 :
2076 : // Although `i.rangeKey.stale` is true, the span s may still be identical
2077 : // to the currently saved span. This is possible when seeking the iterator,
2078 : // which may land back on the same range key. If we previously had a range
2079 : // key and the new one has an identical start key, then it must be the same
2080 : // range key and we can avoid copying and keep `i.rangeKey.updated=false`.
2081 : //
2082 : // TODO(jackson): These key comparisons could be avoidable during relative
2083 : // positioning operations continuing in the same direction, because these
2084 : // ops will never encounter the previous position's range key while
2085 : // stale=true. However, threading whether the current op is a seek or step
2086 : // maybe isn't worth it. This key comparison is only necessary once when we
2087 : // step onto a new range key, which should be relatively rare.
2088 1 : if i.rangeKey.prevPosHadRangeKey && i.equal(i.rangeKey.start, s.Start) &&
2089 1 : i.equal(i.rangeKey.end, s.End) {
2090 1 : i.rangeKey.updated = false
2091 1 : i.rangeKey.stale = false
2092 1 : i.rangeKey.hasRangeKey = true
2093 1 : return
2094 1 : }
2095 1 : i.stats.RangeKeyStats.Count += len(s.Keys)
2096 1 : i.rangeKey.buf.Reset()
2097 1 : i.rangeKey.hasRangeKey = true
2098 1 : i.rangeKey.updated = true
2099 1 : i.rangeKey.stale = false
2100 1 : i.rangeKey.buf, i.rangeKey.start = i.rangeKey.buf.Copy(s.Start)
2101 1 : i.rangeKey.buf, i.rangeKey.end = i.rangeKey.buf.Copy(s.End)
2102 1 : i.rangeKey.keys = i.rangeKey.keys[:0]
2103 1 : for j := 0; j < len(s.Keys); j++ {
2104 1 : if invariants.Enabled {
2105 1 : if s.Keys[j].Kind() != base.InternalKeyKindRangeKeySet {
2106 0 : panic("pebble: user iteration encountered non-RangeKeySet key kind")
2107 1 : } else if j > 0 && i.cmp(s.Keys[j].Suffix, s.Keys[j-1].Suffix) < 0 {
2108 0 : panic("pebble: user iteration encountered range keys not in suffix order")
2109 : }
2110 : }
2111 1 : var rkd RangeKeyData
2112 1 : i.rangeKey.buf, rkd.Suffix = i.rangeKey.buf.Copy(s.Keys[j].Suffix)
2113 1 : i.rangeKey.buf, rkd.Value = i.rangeKey.buf.Copy(s.Keys[j].Value)
2114 1 : i.rangeKey.keys = append(i.rangeKey.keys, rkd)
2115 : }
2116 : }
2117 :
2118 : // RangeKeyChanged indicates whether the most recent iterator positioning
2119 : // operation resulted in the iterator stepping into or out of a new range key.
2120 : // If true, previously returned range key bounds and data has been invalidated.
2121 : // If false, previously obtained range key bounds, suffix and value slices are
2122 : // still valid and may continue to be read.
2123 : //
2124 : // Invalid iterator positions are considered to not hold range keys, meaning
2125 : // that if an iterator steps from an IterExhausted or IterAtLimit position onto
2126 : // a position with a range key, RangeKeyChanged will yield true.
2127 1 : func (i *Iterator) RangeKeyChanged() bool {
2128 1 : return i.iterValidityState == IterValid && i.rangeKey != nil && i.rangeKey.updated
2129 1 : }
2130 :
2131 : // HasPointAndRange indicates whether there exists a point key, a range key or
2132 : // both at the current iterator position.
2133 1 : func (i *Iterator) HasPointAndRange() (hasPoint, hasRange bool) {
2134 1 : if i.iterValidityState != IterValid || i.requiresReposition {
2135 1 : return false, false
2136 1 : }
2137 1 : if i.opts.KeyTypes == IterKeyTypePointsOnly {
2138 1 : return true, false
2139 1 : }
2140 1 : return i.rangeKey == nil || !i.rangeKey.rangeKeyOnly, i.rangeKey != nil && i.rangeKey.hasRangeKey
2141 : }
2142 :
2143 : // RangeBounds returns the start (inclusive) and end (exclusive) bounds of the
2144 : // range key covering the current iterator position. RangeBounds returns nil
2145 : // bounds if there is no range key covering the current iterator position, or
2146 : // the iterator is not configured to surface range keys.
2147 : //
2148 : // If valid, the returned start bound is less than or equal to Key() and the
2149 : // returned end bound is greater than Key().
2150 1 : func (i *Iterator) RangeBounds() (start, end []byte) {
2151 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2152 0 : return nil, nil
2153 0 : }
2154 1 : return i.rangeKey.start, i.rangeKey.end
2155 : }
2156 :
2157 : // Key returns the key of the current key/value pair, or nil if done. The
2158 : // caller should not modify the contents of the returned slice, and its
2159 : // contents may change on the next call to Next.
2160 : //
2161 : // If positioned at an iterator position that only holds a range key, Key()
2162 : // always returns the start bound of the range key. Otherwise, it returns the
2163 : // point key's key.
2164 1 : func (i *Iterator) Key() []byte {
2165 1 : return i.key
2166 1 : }
2167 :
2168 : // Value returns the value of the current key/value pair, or nil if done. The
2169 : // caller should not modify the contents of the returned slice, and its
2170 : // contents may change on the next call to Next.
2171 : //
2172 : // Only valid if HasPointAndRange() returns true for hasPoint.
2173 : // Deprecated: use ValueAndErr instead.
2174 1 : func (i *Iterator) Value() []byte {
2175 1 : val, _ := i.ValueAndErr()
2176 1 : return val
2177 1 : }
2178 :
2179 : // ValueAndErr returns the value, and any error encountered in extracting the value.
2180 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2181 : //
2182 : // The caller should not modify the contents of the returned slice, and its
2183 : // contents may change on the next call to Next.
2184 1 : func (i *Iterator) ValueAndErr() ([]byte, error) {
2185 1 : val, callerOwned, err := i.value.Value(i.lazyValueBuf)
2186 1 : if err != nil {
2187 0 : i.err = err
2188 0 : }
2189 1 : if callerOwned {
2190 1 : i.lazyValueBuf = val[:0]
2191 1 : }
2192 1 : return val, err
2193 : }
2194 :
2195 : // LazyValue returns the LazyValue. Only for advanced use cases.
2196 : // REQUIRES: i.Error()==nil and HasPointAndRange() returns true for hasPoint.
2197 0 : func (i *Iterator) LazyValue() LazyValue {
2198 0 : return i.value
2199 0 : }
2200 :
2201 : // RangeKeys returns the range key values and their suffixes covering the
2202 : // current iterator position. The range bounds may be retrieved separately
2203 : // through Iterator.RangeBounds().
2204 1 : func (i *Iterator) RangeKeys() []RangeKeyData {
2205 1 : if i.rangeKey == nil || !i.opts.rangeKeys() || !i.rangeKey.hasRangeKey {
2206 0 : return nil
2207 0 : }
2208 1 : return i.rangeKey.keys
2209 : }
2210 :
2211 : // Valid returns true if the iterator is positioned at a valid key/value pair
2212 : // and false otherwise.
2213 1 : func (i *Iterator) Valid() bool {
2214 1 : valid := i.iterValidityState == IterValid && !i.requiresReposition
2215 1 : if invariants.Enabled {
2216 1 : if err := i.Error(); valid && err != nil {
2217 0 : panic(errors.WithSecondaryError(errors.AssertionFailedf("pebble: iterator is valid with non-nil Error"), err))
2218 : }
2219 : }
2220 1 : return valid
2221 : }
2222 :
2223 : // Error returns any accumulated error.
2224 1 : func (i *Iterator) Error() error {
2225 1 : if i.iter != nil {
2226 1 : return firstError(i.err, i.iter.Error())
2227 1 : }
2228 0 : return i.err
2229 : }
2230 :
2231 : const maxKeyBufCacheSize = 4 << 10 // 4 KB
2232 :
2233 : // Close closes the iterator and returns any accumulated error. Exhausting
2234 : // all the key/value pairs in a table is not considered to be an error.
2235 : // It is not valid to call any method, including Close, after the iterator
2236 : // has been closed.
2237 1 : func (i *Iterator) Close() error {
2238 1 : // Close the child iterator before releasing the readState because when the
2239 1 : // readState is released sstables referenced by the readState may be deleted
2240 1 : // which will fail on Windows if the sstables are still open by the child
2241 1 : // iterator.
2242 1 : if i.iter != nil {
2243 1 : i.err = firstError(i.err, i.iter.Close())
2244 1 :
2245 1 : // Closing i.iter did not necessarily close the point and range key
2246 1 : // iterators. Calls to SetOptions may have 'disconnected' either one
2247 1 : // from i.iter if iteration key types were changed. Both point and range
2248 1 : // key iterators are preserved in case the iterator needs to switch key
2249 1 : // types again. We explicitly close both of these iterators here.
2250 1 : //
2251 1 : // NB: If the iterators were still connected to i.iter, they may be
2252 1 : // closed, but calling Close on a closed internal iterator or fragment
2253 1 : // iterator is allowed.
2254 1 : if i.pointIter != nil && !i.closePointIterOnce {
2255 1 : i.err = firstError(i.err, i.pointIter.Close())
2256 1 : }
2257 1 : if i.rangeKey != nil && i.rangeKey.rangeKeyIter != nil {
2258 1 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2259 1 : }
2260 : }
2261 1 : err := i.err
2262 1 :
2263 1 : if i.readState != nil {
2264 1 : if i.readSampling.pendingCompactions.size > 0 {
2265 1 : // Copy pending read compactions using db.mu.Lock()
2266 1 : i.readState.db.mu.Lock()
2267 1 : i.readState.db.mu.compact.readCompactions.combine(&i.readSampling.pendingCompactions, i.cmp)
2268 1 : reschedule := i.readState.db.mu.compact.rescheduleReadCompaction
2269 1 : i.readState.db.mu.compact.rescheduleReadCompaction = false
2270 1 : concurrentCompactions := i.readState.db.mu.compact.compactingCount
2271 1 : i.readState.db.mu.Unlock()
2272 1 :
2273 1 : if reschedule && concurrentCompactions == 0 {
2274 0 : // In a read heavy workload, flushes may not happen frequently enough to
2275 0 : // schedule compactions.
2276 0 : i.readState.db.compactionSchedulers.Add(1)
2277 0 : go i.readState.db.maybeScheduleCompactionAsync()
2278 0 : }
2279 : }
2280 :
2281 1 : i.readState.unref()
2282 1 : i.readState = nil
2283 : }
2284 :
2285 1 : if i.version != nil {
2286 1 : i.version.Unref()
2287 1 : }
2288 :
2289 1 : for _, readers := range i.externalReaders {
2290 1 : for _, r := range readers {
2291 1 : err = firstError(err, r.Close())
2292 1 : }
2293 : }
2294 :
2295 : // Close the closer for the current value if one was open.
2296 1 : if i.valueCloser != nil {
2297 1 : err = firstError(err, i.valueCloser.Close())
2298 1 : i.valueCloser = nil
2299 1 : }
2300 :
2301 1 : if i.rangeKey != nil {
2302 1 :
2303 1 : i.rangeKey.rangeKeyBuffers.PrepareForReuse()
2304 1 : *i.rangeKey = iteratorRangeKeyState{
2305 1 : rangeKeyBuffers: i.rangeKey.rangeKeyBuffers,
2306 1 : }
2307 1 : iterRangeKeyStateAllocPool.Put(i.rangeKey)
2308 1 : i.rangeKey = nil
2309 1 : }
2310 1 : if alloc := i.alloc; alloc != nil {
2311 1 : // Avoid caching the key buf if it is overly large. The constant is fairly
2312 1 : // arbitrary.
2313 1 : if cap(i.keyBuf) >= maxKeyBufCacheSize {
2314 0 : alloc.keyBuf = nil
2315 1 : } else {
2316 1 : alloc.keyBuf = i.keyBuf
2317 1 : }
2318 1 : if cap(i.prefixOrFullSeekKey) >= maxKeyBufCacheSize {
2319 0 : alloc.prefixOrFullSeekKey = nil
2320 1 : } else {
2321 1 : alloc.prefixOrFullSeekKey = i.prefixOrFullSeekKey
2322 1 : }
2323 1 : for j := range i.boundsBuf {
2324 1 : if cap(i.boundsBuf[j]) >= maxKeyBufCacheSize {
2325 0 : alloc.boundsBuf[j] = nil
2326 1 : } else {
2327 1 : alloc.boundsBuf[j] = i.boundsBuf[j]
2328 1 : }
2329 : }
2330 1 : *alloc = iterAlloc{
2331 1 : keyBuf: alloc.keyBuf,
2332 1 : boundsBuf: alloc.boundsBuf,
2333 1 : prefixOrFullSeekKey: alloc.prefixOrFullSeekKey,
2334 1 : }
2335 1 : iterAllocPool.Put(alloc)
2336 1 : } else if alloc := i.getIterAlloc; alloc != nil {
2337 1 : if cap(i.keyBuf) >= maxKeyBufCacheSize {
2338 0 : alloc.keyBuf = nil
2339 1 : } else {
2340 1 : alloc.keyBuf = i.keyBuf
2341 1 : }
2342 1 : *alloc = getIterAlloc{
2343 1 : keyBuf: alloc.keyBuf,
2344 1 : }
2345 1 : getIterAllocPool.Put(alloc)
2346 : }
2347 1 : return err
2348 : }
2349 :
2350 : // SetBounds sets the lower and upper bounds for the iterator. Once SetBounds
2351 : // returns, the caller is free to mutate the provided slices.
2352 : //
2353 : // The iterator will always be invalidated and must be repositioned with a call
2354 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2355 1 : func (i *Iterator) SetBounds(lower, upper []byte) {
2356 1 : // Ensure that the Iterator appears exhausted, regardless of whether we
2357 1 : // actually have to invalidate the internal iterator. Optimizations that
2358 1 : // avoid exhaustion are an internal implementation detail that shouldn't
2359 1 : // leak through the interface. The caller should still call an absolute
2360 1 : // positioning method to reposition the iterator.
2361 1 : i.requiresReposition = true
2362 1 :
2363 1 : if ((i.opts.LowerBound == nil) == (lower == nil)) &&
2364 1 : ((i.opts.UpperBound == nil) == (upper == nil)) &&
2365 1 : i.equal(i.opts.LowerBound, lower) &&
2366 1 : i.equal(i.opts.UpperBound, upper) {
2367 1 : // Unchanged, noop.
2368 1 : return
2369 1 : }
2370 :
2371 : // Copy the user-provided bounds into an Iterator-owned buffer, and set them
2372 : // on i.opts.{Lower,Upper}Bound.
2373 1 : i.processBounds(lower, upper)
2374 1 :
2375 1 : i.iter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2376 1 : // If the iterator has an open point iterator that's not currently being
2377 1 : // used, propagate the new bounds to it.
2378 1 : if i.pointIter != nil && !i.opts.pointKeys() {
2379 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2380 1 : }
2381 : // If the iterator has a range key iterator, propagate bounds to it. The
2382 : // top-level SetBounds on the interleaving iterator (i.iter) won't propagate
2383 : // bounds to the range key iterator stack, because the FragmentIterator
2384 : // interface doesn't define a SetBounds method. We need to directly inform
2385 : // the iterConfig stack.
2386 1 : if i.rangeKey != nil {
2387 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2388 1 : }
2389 :
2390 : // Even though this is not a positioning operation, the alteration of the
2391 : // bounds means we cannot optimize Seeks by using Next.
2392 1 : i.invalidate()
2393 : }
2394 :
2395 : // SetContext replaces the context provided at iterator creation, or the last
2396 : // one provided by SetContext. Even though iterators are expected to be
2397 : // short-lived, there are some cases where either (a) iterators are used far
2398 : // from the code that created them, (b) iterators are reused (while being
2399 : // short-lived) for processing different requests. For such scenarios, we
2400 : // allow the caller to replace the context.
2401 0 : func (i *Iterator) SetContext(ctx context.Context) {
2402 0 : i.ctx = ctx
2403 0 : i.iter.SetContext(ctx)
2404 0 : // If the iterator has an open point iterator that's not currently being
2405 0 : // used, propagate the new context to it.
2406 0 : if i.pointIter != nil && !i.opts.pointKeys() {
2407 0 : i.pointIter.SetContext(i.ctx)
2408 0 : }
2409 : }
2410 :
2411 : // Initialization and changing of the bounds must call processBounds.
2412 : // processBounds saves the bounds and computes derived state from those
2413 : // bounds.
2414 1 : func (i *Iterator) processBounds(lower, upper []byte) {
2415 1 : // Copy the user-provided bounds into an Iterator-owned buffer. We can't
2416 1 : // overwrite the current bounds, because some internal iterators compare old
2417 1 : // and new bounds for optimizations.
2418 1 :
2419 1 : buf := i.boundsBuf[i.boundsBufIdx][:0]
2420 1 : if lower != nil {
2421 1 : buf = append(buf, lower...)
2422 1 : i.opts.LowerBound = buf
2423 1 : } else {
2424 1 : i.opts.LowerBound = nil
2425 1 : }
2426 1 : i.nextPrefixNotPermittedByUpperBound = false
2427 1 : if upper != nil {
2428 1 : buf = append(buf, upper...)
2429 1 : i.opts.UpperBound = buf[len(buf)-len(upper):]
2430 1 : if i.comparer.Split != nil {
2431 1 : if i.comparer.Split(i.opts.UpperBound) != len(i.opts.UpperBound) {
2432 1 : // Setting an upper bound that is a versioned MVCC key. This means
2433 1 : // that a key can have some MVCC versions before the upper bound and
2434 1 : // some after. This causes significant complications for NextPrefix,
2435 1 : // so we bar the user of NextPrefix.
2436 1 : i.nextPrefixNotPermittedByUpperBound = true
2437 1 : }
2438 : }
2439 1 : } else {
2440 1 : i.opts.UpperBound = nil
2441 1 : }
2442 1 : i.boundsBuf[i.boundsBufIdx] = buf
2443 1 : i.boundsBufIdx = 1 - i.boundsBufIdx
2444 : }
2445 :
2446 : // SetOptions sets new iterator options for the iterator. Note that the lower
2447 : // and upper bounds applied here will supersede any bounds set by previous calls
2448 : // to SetBounds.
2449 : //
2450 : // Note that the slices provided in this SetOptions must not be changed by the
2451 : // caller until the iterator is closed, or a subsequent SetBounds or SetOptions
2452 : // has returned. This is because comparisons between the existing and new bounds
2453 : // are sometimes used to optimize seeking. See the extended commentary on
2454 : // SetBounds.
2455 : //
2456 : // If the iterator was created over an indexed mutable batch, the iterator's
2457 : // view of the mutable batch is refreshed.
2458 : //
2459 : // The iterator will always be invalidated and must be repositioned with a call
2460 : // to SeekGE, SeekPrefixGE, SeekLT, First, or Last.
2461 : //
2462 : // If only lower and upper bounds need to be modified, prefer SetBounds.
2463 1 : func (i *Iterator) SetOptions(o *IterOptions) {
2464 1 : if i.externalReaders != nil {
2465 1 : if err := validateExternalIterOpts(o); err != nil {
2466 0 : panic(err)
2467 : }
2468 : }
2469 :
2470 : // Ensure that the Iterator appears exhausted, regardless of whether we
2471 : // actually have to invalidate the internal iterator. Optimizations that
2472 : // avoid exhaustion are an internal implementation detail that shouldn't
2473 : // leak through the interface. The caller should still call an absolute
2474 : // positioning method to reposition the iterator.
2475 1 : i.requiresReposition = true
2476 1 :
2477 1 : // Check if global state requires we close all internal iterators.
2478 1 : //
2479 1 : // If the Iterator is in an error state, invalidate the existing iterators
2480 1 : // so that we reconstruct an iterator state from scratch.
2481 1 : //
2482 1 : // If OnlyReadGuaranteedDurable changed, the iterator stacks are incorrect,
2483 1 : // improperly including or excluding memtables. Invalidate them so that
2484 1 : // finishInitializingIter will reconstruct them.
2485 1 : //
2486 1 : // If either the original options or the new options specify a table filter,
2487 1 : // we need to reconstruct the iterator stacks. If they both supply a table
2488 1 : // filter, we can't be certain that it's the same filter since we have no
2489 1 : // mechanism to compare the filter closures.
2490 1 : closeBoth := i.err != nil ||
2491 1 : o.OnlyReadGuaranteedDurable != i.opts.OnlyReadGuaranteedDurable ||
2492 1 : o.TableFilter != nil || i.opts.TableFilter != nil
2493 1 :
2494 1 : // If either options specify block property filters for an iterator stack,
2495 1 : // reconstruct it.
2496 1 : if i.pointIter != nil && (closeBoth || len(o.PointKeyFilters) > 0 || len(i.opts.PointKeyFilters) > 0 ||
2497 1 : o.RangeKeyMasking.Filter != nil || i.opts.RangeKeyMasking.Filter != nil || o.SkipPoint != nil ||
2498 1 : i.opts.SkipPoint != nil) {
2499 1 : i.err = firstError(i.err, i.pointIter.Close())
2500 1 : i.pointIter = nil
2501 1 : }
2502 1 : if i.rangeKey != nil {
2503 1 : if closeBoth || len(o.RangeKeyFilters) > 0 || len(i.opts.RangeKeyFilters) > 0 {
2504 1 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2505 1 : i.rangeKey = nil
2506 1 : } else {
2507 1 : // If there's still a range key iterator stack, invalidate the
2508 1 : // iterator. This ensures RangeKeyChanged() returns true if a
2509 1 : // subsequent positioning operation discovers a range key. It also
2510 1 : // prevents seek no-op optimizations.
2511 1 : i.invalidate()
2512 1 : }
2513 : }
2514 :
2515 : // If the iterator is backed by a batch that's been mutated, refresh its
2516 : // existing point and range-key iterators, and invalidate the iterator to
2517 : // prevent seek-using-next optimizations. If we don't yet have a point-key
2518 : // iterator or range-key iterator but we require one, it'll be created in
2519 : // the slow path that reconstructs the iterator in finishInitializingIter.
2520 1 : if i.batch != nil {
2521 1 : nextBatchSeqNum := (uint64(len(i.batch.data)) | base.InternalKeySeqNumBatch)
2522 1 : if nextBatchSeqNum != i.batchSeqNum {
2523 1 : i.batchSeqNum = nextBatchSeqNum
2524 1 : if i.merging != nil {
2525 1 : i.merging.batchSnapshot = nextBatchSeqNum
2526 1 : }
2527 : // Prevent a no-op seek optimization on the next seek. We won't be
2528 : // able to reuse the top-level Iterator state, because it may be
2529 : // incorrect after the inclusion of new batch mutations.
2530 1 : i.batchJustRefreshed = true
2531 1 : if i.pointIter != nil && i.batch.countRangeDels > 0 {
2532 1 : if i.batchRangeDelIter.Count() == 0 {
2533 1 : // When we constructed this iterator, there were no
2534 1 : // rangedels in the batch. Iterator construction will
2535 1 : // have excluded the batch rangedel iterator from the
2536 1 : // point iterator stack. We need to reconstruct the
2537 1 : // point iterator to add i.batchRangeDelIter into the
2538 1 : // iterator stack.
2539 1 : i.err = firstError(i.err, i.pointIter.Close())
2540 1 : i.pointIter = nil
2541 1 : } else {
2542 1 : // There are range deletions in the batch and we already
2543 1 : // have a batch rangedel iterator. We can update the
2544 1 : // batch rangedel iterator in place.
2545 1 : //
2546 1 : // NB: There may or may not be new range deletions. We
2547 1 : // can't tell based on i.batchRangeDelIter.Count(),
2548 1 : // which is the count of fragmented range deletions, NOT
2549 1 : // the number of range deletions written to the batch
2550 1 : // [i.batch.countRangeDels].
2551 1 : i.batch.initRangeDelIter(&i.opts, &i.batchRangeDelIter, nextBatchSeqNum)
2552 1 : }
2553 : }
2554 1 : if i.rangeKey != nil && i.batch.countRangeKeys > 0 {
2555 1 : if i.batchRangeKeyIter.Count() == 0 {
2556 1 : // When we constructed this iterator, there were no range
2557 1 : // keys in the batch. Iterator construction will have
2558 1 : // excluded the batch rangekey iterator from the range key
2559 1 : // iterator stack. We need to reconstruct the range key
2560 1 : // iterator to add i.batchRangeKeyIter into the iterator
2561 1 : // stack.
2562 1 : i.err = firstError(i.err, i.rangeKey.rangeKeyIter.Close())
2563 1 : i.rangeKey = nil
2564 1 : } else {
2565 1 : // There are range keys in the batch and we already
2566 1 : // have a batch rangekey iterator. We can update the batch
2567 1 : // rangekey iterator in place.
2568 1 : //
2569 1 : // NB: There may or may not be new range keys. We can't
2570 1 : // tell based on i.batchRangeKeyIter.Count(), which is the
2571 1 : // count of fragmented range keys, NOT the number of
2572 1 : // range keys written to the batch [i.batch.countRangeKeys].
2573 1 : i.batch.initRangeKeyIter(&i.opts, &i.batchRangeKeyIter, nextBatchSeqNum)
2574 1 : i.invalidate()
2575 1 : }
2576 : }
2577 : }
2578 : }
2579 :
2580 : // Reset combinedIterState.initialized in case the iterator key types
2581 : // changed. If there's already a range key iterator stack, the combined
2582 : // iterator is already initialized. Additionally, if the iterator is not
2583 : // configured to include range keys, mark it as initialized to signal that
2584 : // lower level iterators should not trigger a switch to combined iteration.
2585 1 : i.lazyCombinedIter.combinedIterState = combinedIterState{
2586 1 : initialized: i.rangeKey != nil || !i.opts.rangeKeys(),
2587 1 : }
2588 1 :
2589 1 : boundsEqual := ((i.opts.LowerBound == nil) == (o.LowerBound == nil)) &&
2590 1 : ((i.opts.UpperBound == nil) == (o.UpperBound == nil)) &&
2591 1 : i.equal(i.opts.LowerBound, o.LowerBound) &&
2592 1 : i.equal(i.opts.UpperBound, o.UpperBound)
2593 1 :
2594 1 : if boundsEqual && o.KeyTypes == i.opts.KeyTypes &&
2595 1 : (i.pointIter != nil || !i.opts.pointKeys()) &&
2596 1 : (i.rangeKey != nil || !i.opts.rangeKeys() || i.opts.KeyTypes == IterKeyTypePointsAndRanges) &&
2597 1 : i.equal(o.RangeKeyMasking.Suffix, i.opts.RangeKeyMasking.Suffix) &&
2598 1 : o.UseL6Filters == i.opts.UseL6Filters {
2599 1 : // The options are identical, so we can likely use the fast path. In
2600 1 : // addition to all the above constraints, we cannot use the fast path if
2601 1 : // configured to perform lazy combined iteration but an indexed batch
2602 1 : // used by the iterator now contains range keys. Lazy combined iteration
2603 1 : // is not compatible with batch range keys because we always need to
2604 1 : // merge the batch's range keys into iteration.
2605 1 : if i.rangeKey != nil || !i.opts.rangeKeys() || i.batch == nil || i.batch.countRangeKeys == 0 {
2606 1 : // Fast path. This preserves the Seek-using-Next optimizations as
2607 1 : // long as the iterator wasn't already invalidated up above.
2608 1 : return
2609 1 : }
2610 : }
2611 : // Slow path.
2612 :
2613 : // The options changed. Save the new ones to i.opts.
2614 1 : if boundsEqual {
2615 1 : // Copying the options into i.opts will overwrite LowerBound and
2616 1 : // UpperBound fields with the user-provided slices. We need to hold on
2617 1 : // to the Pebble-owned slices, so save them and re-set them after the
2618 1 : // copy.
2619 1 : lower, upper := i.opts.LowerBound, i.opts.UpperBound
2620 1 : i.opts = *o
2621 1 : i.opts.LowerBound, i.opts.UpperBound = lower, upper
2622 1 : } else {
2623 1 : i.opts = *o
2624 1 : i.processBounds(o.LowerBound, o.UpperBound)
2625 1 : // Propagate the changed bounds to the existing point iterator.
2626 1 : // NB: We propagate i.opts.{Lower,Upper}Bound, not o.{Lower,Upper}Bound
2627 1 : // because i.opts now point to buffers owned by Pebble.
2628 1 : if i.pointIter != nil {
2629 1 : i.pointIter.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2630 1 : }
2631 1 : if i.rangeKey != nil {
2632 1 : i.rangeKey.iterConfig.SetBounds(i.opts.LowerBound, i.opts.UpperBound)
2633 1 : }
2634 : }
2635 :
2636 : // Even though this is not a positioning operation, the invalidation of the
2637 : // iterator stack means we cannot optimize Seeks by using Next.
2638 1 : i.invalidate()
2639 1 :
2640 1 : // Iterators created through NewExternalIter have a different iterator
2641 1 : // initialization process.
2642 1 : if i.externalReaders != nil {
2643 1 : finishInitializingExternal(i.ctx, i)
2644 1 : return
2645 1 : }
2646 1 : finishInitializingIter(i.ctx, i.alloc)
2647 : }
2648 :
2649 1 : func (i *Iterator) invalidate() {
2650 1 : i.lastPositioningOp = invalidatedLastPositionOp
2651 1 : i.hasPrefix = false
2652 1 : i.iterKey = nil
2653 1 : i.iterValue = LazyValue{}
2654 1 : i.err = nil
2655 1 : // This switch statement isn't necessary for correctness since callers
2656 1 : // should call a repositioning method. We could have arbitrarily set i.pos
2657 1 : // to one of the values. But it results in more intuitive behavior in
2658 1 : // tests, which do not always reposition.
2659 1 : switch i.pos {
2660 1 : case iterPosCurForward, iterPosNext, iterPosCurForwardPaused:
2661 1 : i.pos = iterPosCurForward
2662 1 : case iterPosCurReverse, iterPosPrev, iterPosCurReversePaused:
2663 1 : i.pos = iterPosCurReverse
2664 : }
2665 1 : i.iterValidityState = IterExhausted
2666 1 : if i.rangeKey != nil {
2667 1 : i.rangeKey.iiter.Invalidate()
2668 1 : i.rangeKey.prevPosHadRangeKey = false
2669 1 : }
2670 : }
2671 :
2672 : // Metrics returns per-iterator metrics.
2673 0 : func (i *Iterator) Metrics() IteratorMetrics {
2674 0 : m := IteratorMetrics{
2675 0 : ReadAmp: 1,
2676 0 : }
2677 0 : if mi, ok := i.iter.(*mergingIter); ok {
2678 0 : m.ReadAmp = len(mi.levels)
2679 0 : }
2680 0 : return m
2681 : }
2682 :
2683 : // ResetStats resets the stats to 0.
2684 0 : func (i *Iterator) ResetStats() {
2685 0 : i.stats = IteratorStats{}
2686 0 : }
2687 :
2688 : // Stats returns the current stats.
2689 1 : func (i *Iterator) Stats() IteratorStats {
2690 1 : return i.stats
2691 1 : }
2692 :
2693 : // CloneOptions configures an iterator constructed through Iterator.Clone.
2694 : type CloneOptions struct {
2695 : // IterOptions, if non-nil, define the iterator options to configure a
2696 : // cloned iterator. If nil, the clone adopts the same IterOptions as the
2697 : // iterator being cloned.
2698 : IterOptions *IterOptions
2699 : // RefreshBatchView may be set to true when cloning an Iterator over an
2700 : // indexed batch. When false, the clone adopts the same (possibly stale)
2701 : // view of the indexed batch as the cloned Iterator. When true, the clone is
2702 : // constructed with a refreshed view of the batch, observing all of the
2703 : // batch's mutations at the time of the Clone. If the cloned iterator was
2704 : // not constructed to read over an indexed batch, RefreshVatchView has no
2705 : // effect.
2706 : RefreshBatchView bool
2707 : }
2708 :
2709 : // Clone creates a new Iterator over the same underlying data, i.e., over the
2710 : // same {batch, memtables, sstables}). The resulting iterator is not positioned.
2711 : // It starts with the same IterOptions, unless opts.IterOptions is set.
2712 : //
2713 : // When called on an Iterator over an indexed batch, the clone's visibility of
2714 : // the indexed batch is determined by CloneOptions.RefreshBatchView. If false,
2715 : // the clone inherits the iterator's current (possibly stale) view of the batch,
2716 : // and callers may call SetOptions to subsequently refresh the clone's view to
2717 : // include all batch mutations. If true, the clone is constructed with a
2718 : // complete view of the indexed batch's mutations at the time of the Clone.
2719 : //
2720 : // Callers can use Clone if they need multiple iterators that need to see
2721 : // exactly the same underlying state of the DB. This should not be used to
2722 : // extend the lifetime of the data backing the original Iterator since that
2723 : // will cause an increase in memory and disk usage (use NewSnapshot for that
2724 : // purpose).
2725 1 : func (i *Iterator) Clone(opts CloneOptions) (*Iterator, error) {
2726 1 : return i.CloneWithContext(context.Background(), opts)
2727 1 : }
2728 :
2729 : // CloneWithContext is like Clone, and additionally accepts a context for
2730 : // tracing.
2731 1 : func (i *Iterator) CloneWithContext(ctx context.Context, opts CloneOptions) (*Iterator, error) {
2732 1 : if opts.IterOptions == nil {
2733 1 : opts.IterOptions = &i.opts
2734 1 : }
2735 1 : if i.batchOnlyIter {
2736 1 : return nil, errors.Errorf("cannot Clone a batch-only Iterator")
2737 1 : }
2738 1 : readState := i.readState
2739 1 : vers := i.version
2740 1 : if readState == nil && vers == nil {
2741 1 : return nil, errors.Errorf("cannot Clone a closed Iterator")
2742 1 : }
2743 : // i is already holding a ref, so there is no race with unref here.
2744 : //
2745 : // TODO(bilal): If the underlying iterator was created on a snapshot, we could
2746 : // grab a reference to the current readState instead of reffing the original
2747 : // readState. This allows us to release references to some zombie sstables
2748 : // and memtables.
2749 1 : if readState != nil {
2750 1 : readState.ref()
2751 1 : }
2752 1 : if vers != nil {
2753 1 : vers.Ref()
2754 1 : }
2755 : // Bundle various structures under a single umbrella in order to allocate
2756 : // them together.
2757 1 : buf := iterAllocPool.Get().(*iterAlloc)
2758 1 : dbi := &buf.dbi
2759 1 : *dbi = Iterator{
2760 1 : ctx: ctx,
2761 1 : opts: *opts.IterOptions,
2762 1 : alloc: buf,
2763 1 : merge: i.merge,
2764 1 : comparer: i.comparer,
2765 1 : readState: readState,
2766 1 : version: vers,
2767 1 : keyBuf: buf.keyBuf,
2768 1 : prefixOrFullSeekKey: buf.prefixOrFullSeekKey,
2769 1 : boundsBuf: buf.boundsBuf,
2770 1 : batch: i.batch,
2771 1 : batchSeqNum: i.batchSeqNum,
2772 1 : newIters: i.newIters,
2773 1 : newIterRangeKey: i.newIterRangeKey,
2774 1 : seqNum: i.seqNum,
2775 1 : }
2776 1 : dbi.processBounds(dbi.opts.LowerBound, dbi.opts.UpperBound)
2777 1 :
2778 1 : // If the caller requested the clone have a current view of the indexed
2779 1 : // batch, set the clone's batch sequence number appropriately.
2780 1 : if i.batch != nil && opts.RefreshBatchView {
2781 1 : dbi.batchSeqNum = (uint64(len(i.batch.data)) | base.InternalKeySeqNumBatch)
2782 1 : }
2783 :
2784 1 : return finishInitializingIter(ctx, buf), nil
2785 : }
2786 :
2787 : // Merge adds all of the argument's statistics to the receiver. It may be used
2788 : // to accumulate stats across multiple iterators.
2789 1 : func (stats *IteratorStats) Merge(o IteratorStats) {
2790 1 : for i := InterfaceCall; i < NumStatsKind; i++ {
2791 1 : stats.ForwardSeekCount[i] += o.ForwardSeekCount[i]
2792 1 : stats.ReverseSeekCount[i] += o.ReverseSeekCount[i]
2793 1 : stats.ForwardStepCount[i] += o.ForwardStepCount[i]
2794 1 : stats.ReverseStepCount[i] += o.ReverseStepCount[i]
2795 1 : }
2796 1 : stats.InternalStats.Merge(o.InternalStats)
2797 1 : stats.RangeKeyStats.Merge(o.RangeKeyStats)
2798 : }
2799 :
2800 1 : func (stats *IteratorStats) String() string {
2801 1 : return redact.StringWithoutMarkers(stats)
2802 1 : }
2803 :
2804 : // SafeFormat implements the redact.SafeFormatter interface.
2805 1 : func (stats *IteratorStats) SafeFormat(s redact.SafePrinter, verb rune) {
2806 1 : for i := range stats.ForwardStepCount {
2807 1 : switch IteratorStatsKind(i) {
2808 1 : case InterfaceCall:
2809 1 : s.SafeString("(interface (dir, seek, step): ")
2810 1 : case InternalIterCall:
2811 1 : s.SafeString(", (internal (dir, seek, step): ")
2812 : }
2813 1 : s.Printf("(fwd, %d, %d), (rev, %d, %d))",
2814 1 : redact.Safe(stats.ForwardSeekCount[i]), redact.Safe(stats.ForwardStepCount[i]),
2815 1 : redact.Safe(stats.ReverseSeekCount[i]), redact.Safe(stats.ReverseStepCount[i]))
2816 : }
2817 1 : if stats.InternalStats != (InternalIteratorStats{}) {
2818 1 : s.SafeString(",\n(internal-stats: ")
2819 1 : s.Printf("(block-bytes: (total %s, cached %s, read-time %s)), "+
2820 1 : "(points: (count %s, key-bytes %s, value-bytes %s, tombstoned %s))",
2821 1 : humanize.Bytes.Uint64(stats.InternalStats.BlockBytes),
2822 1 : humanize.Bytes.Uint64(stats.InternalStats.BlockBytesInCache),
2823 1 : humanize.FormattedString(stats.InternalStats.BlockReadDuration.String()),
2824 1 : humanize.Count.Uint64(stats.InternalStats.PointCount),
2825 1 : humanize.Bytes.Uint64(stats.InternalStats.KeyBytes),
2826 1 : humanize.Bytes.Uint64(stats.InternalStats.ValueBytes),
2827 1 : humanize.Count.Uint64(stats.InternalStats.PointsCoveredByRangeTombstones),
2828 1 : )
2829 1 : if stats.InternalStats.SeparatedPointValue.Count != 0 {
2830 1 : s.Printf(", (separated: (count %s, bytes %s, fetched %s)))",
2831 1 : humanize.Count.Uint64(stats.InternalStats.SeparatedPointValue.Count),
2832 1 : humanize.Bytes.Uint64(stats.InternalStats.SeparatedPointValue.ValueBytes),
2833 1 : humanize.Bytes.Uint64(stats.InternalStats.SeparatedPointValue.ValueBytesFetched))
2834 1 : } else {
2835 1 : s.Printf(")")
2836 1 : }
2837 : }
2838 1 : if stats.RangeKeyStats != (RangeKeyIteratorStats{}) {
2839 1 : s.SafeString(",\n(range-key-stats: ")
2840 1 : s.Printf("(count %d), (contained points: (count %d, skipped %d)))",
2841 1 : stats.RangeKeyStats.Count,
2842 1 : stats.RangeKeyStats.ContainedPoints,
2843 1 : stats.RangeKeyStats.SkippedPoints)
2844 1 : }
2845 : }
|