Line data Source code
1 : // Copyright 2018 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 : "fmt"
11 : "runtime/debug"
12 : "unsafe"
13 :
14 : "github.com/cockroachdb/errors"
15 : "github.com/cockroachdb/pebble/internal/base"
16 : "github.com/cockroachdb/pebble/internal/invariants"
17 : "github.com/cockroachdb/pebble/internal/keyspan"
18 : )
19 :
20 : type mergingIterLevel struct {
21 : index int
22 : iter internalIterator
23 : // rangeDelIter is set to the range-deletion iterator for the level. When
24 : // configured with a levelIter, this pointer changes as sstable boundaries
25 : // are crossed. See levelIter.initRangeDel and the Range Deletions comment
26 : // below.
27 : rangeDelIter keyspan.FragmentIterator
28 : // iterKV caches the current key-value pair iter points to.
29 : iterKV *base.InternalKV
30 : // levelIter is non-nil if this level's iter is ultimately backed by a
31 : // *levelIter. The handle in iter may have wrapped the levelIter with
32 : // intermediary internalIterator implementations.
33 : levelIter *levelIter
34 :
35 : // tombstone caches the tombstone rangeDelIter is currently pointed at. If
36 : // tombstone is nil, there are no further tombstones within the
37 : // current sstable in the current iterator direction. The cached tombstone is
38 : // only valid for the levels in the range [0,heap[0].index]. This avoids
39 : // positioning tombstones at lower levels which cannot possibly shadow the
40 : // current key.
41 : tombstone *keyspan.Span
42 : }
43 :
44 : // mergingIter provides a merged view of multiple iterators from different
45 : // levels of the LSM.
46 : //
47 : // The core of a mergingIter is a heap of internalIterators (see
48 : // mergingIterHeap). The heap can operate as either a min-heap, used during
49 : // forward iteration (First, SeekGE, Next) or a max-heap, used during reverse
50 : // iteration (Last, SeekLT, Prev). The heap is initialized in calls to First,
51 : // Last, SeekGE, and SeekLT. A call to Next or Prev takes the current top
52 : // element on the heap, advances its iterator, and then "fixes" the heap
53 : // property. When one of the child iterators is exhausted during Next/Prev
54 : // iteration, it is removed from the heap.
55 : //
56 : // # Range Deletions
57 : //
58 : // A mergingIter can optionally be configured with a slice of range deletion
59 : // iterators. The range deletion iterator slice must exactly parallel the point
60 : // iterators and the range deletion iterator must correspond to the same level
61 : // in the LSM as the point iterator. Note that each memtable and each table in
62 : // L0 is a different "level" from the mergingIter perspective. So level 0 below
63 : // does not correspond to L0 in the LSM.
64 : //
65 : // A range deletion iterator iterates over fragmented range tombstones. Range
66 : // tombstones are fragmented by splitting them at any overlapping points. This
67 : // fragmentation guarantees that within an sstable tombstones will either be
68 : // distinct or will have identical start and end user keys. While range
69 : // tombstones are fragmented within an sstable, the start and end keys are not truncated
70 : // to sstable boundaries. This is necessary because the tombstone end key is
71 : // exclusive and does not have a sequence number. Consider an sstable
72 : // containing the range tombstone [a,c)#9 and the key "b#8". The tombstone must
73 : // delete "b#8", yet older versions of "b" might spill over to the next
74 : // sstable. So the boundary key for this sstable must be "b#8". Adjusting the
75 : // end key of tombstones to be optionally inclusive or contain a sequence
76 : // number would be possible solutions (such solutions have potentially serious
77 : // issues: tombstones have exclusive end keys since an inclusive deletion end can
78 : // be converted to an exclusive one while the reverse transformation is not possible;
79 : // the semantics of a sequence number for the end key of a range tombstone are murky).
80 : //
81 : // The approach taken here performs an
82 : // implicit truncation of the tombstone to the sstable boundaries.
83 : //
84 : // During initialization of a mergingIter, the range deletion iterators for
85 : // batches, memtables, and L0 tables are populated up front. Note that Batches
86 : // and memtables index unfragmented tombstones. Batch.newRangeDelIter() and
87 : // memTable.newRangeDelIter() fragment and cache the tombstones on demand. The
88 : // L1-L6 range deletion iterators are populated by levelIter. When configured
89 : // to load range deletion iterators, whenever a levelIter loads a table it
90 : // loads both the point iterator and the range deletion
91 : // iterator. levelIter.rangeDelIter is configured to point to the right entry
92 : // in mergingIter.levels. The effect of this setup is that
93 : // mergingIter.levels[i].rangeDelIter always contains the fragmented range
94 : // tombstone for the current table in level i that the levelIter has open.
95 : //
96 : // Another crucial mechanism of levelIter is that it materializes fake point
97 : // entries for the table boundaries if the boundary is range deletion
98 : // key. Consider a table that contains only a range tombstone [a-e)#10. The
99 : // sstable boundaries for this table will be a#10,15 and
100 : // e#72057594037927935,15. During forward iteration levelIter will return
101 : // e#72057594037927935,15 as a key. During reverse iteration levelIter will
102 : // return a#10,15 as a key. These sentinel keys act as bookends to point
103 : // iteration and allow mergingIter to keep a table and its associated range
104 : // tombstones loaded as long as there are keys at lower levels that are within
105 : // the bounds of the table.
106 : //
107 : // The final piece to the range deletion puzzle is the LSM invariant that for a
108 : // given key K newer versions of K can only exist earlier in the level, or at
109 : // higher levels of the tree. For example, if K#4 exists in L3, k#5 can only
110 : // exist earlier in the L3 or in L0, L1, L2 or a memtable. Get very explicitly
111 : // uses this invariant to find the value for a key by walking the LSM level by
112 : // level. For range deletions, this invariant means that a range deletion at
113 : // level N will necessarily shadow any keys within its bounds in level Y where
114 : // Y > N. One wrinkle to this statement is that it only applies to keys that
115 : // lie within the sstable bounds as well, but we get that guarantee due to the
116 : // way the range deletion iterator and point iterator are bound together by a
117 : // levelIter.
118 : //
119 : // Tying the above all together, we get a picture where each level (index in
120 : // mergingIter.levels) is composed of both point operations (pX) and range
121 : // deletions (rX). The range deletions for level X shadow both the point
122 : // operations and range deletions for level Y where Y > X allowing mergingIter
123 : // to skip processing entries in that shadow. For example, consider the
124 : // scenario:
125 : //
126 : // r0: a---e
127 : // r1: d---h
128 : // r2: g---k
129 : // r3: j---n
130 : // r4: m---q
131 : //
132 : // This is showing 5 levels of range deletions. Consider what happens upon
133 : // SeekGE("b"). We first seek the point iterator for level 0 (the point values
134 : // are not shown above) and we then seek the range deletion iterator. That
135 : // returns the tombstone [a,e). This tombstone tells us that all keys in the
136 : // range [a,e) in lower levels are deleted so we can skip them. So we can
137 : // adjust the seek key to "e", the tombstone end key. For level 1 we seek to
138 : // "e" and find the range tombstone [d,h) and similar logic holds. By the time
139 : // we get to level 4 we're seeking to "n".
140 : //
141 : // One consequence of not truncating tombstone end keys to sstable boundaries
142 : // is the seeking process described above cannot always seek to the tombstone
143 : // end key in the older level. For example, imagine in the above example r3 is
144 : // a partitioned level (i.e., L1+ in our LSM), and the sstable containing [j,
145 : // n) has "k" as its upper boundary. In this situation, compactions involving
146 : // keys at or after "k" can output those keys to r4+, even if they're newer
147 : // than our tombstone [j, n). So instead of seeking to "n" in r4 we can only
148 : // seek to "k". To achieve this, the instance variable `largestUserKey.`
149 : // maintains the upper bounds of the current sstables in the partitioned
150 : // levels. In this example, `levels[3].largestUserKey` holds "k", telling us to
151 : // limit the seek triggered by a tombstone in r3 to "k".
152 : //
153 : // During actual iteration levels can contain both point operations and range
154 : // deletions. Within a level, when a range deletion contains a point operation
155 : // the sequence numbers must be checked to determine if the point operation is
156 : // newer or older than the range deletion tombstone. The mergingIter maintains
157 : // the invariant that the range deletion iterators for all levels newer that
158 : // the current iteration key (L < m.heap.items[0].index) are positioned at the
159 : // next (or previous during reverse iteration) range deletion tombstone. We
160 : // know those levels don't contain a range deletion tombstone that covers the
161 : // current key because if they did the current key would be deleted. The range
162 : // deletion iterator for the current key's level is positioned at a range
163 : // tombstone covering or past the current key. The position of all of other
164 : // range deletion iterators is unspecified. Whenever a key from those levels
165 : // becomes the current key, their range deletion iterators need to be
166 : // positioned. This lazy positioning avoids seeking the range deletion
167 : // iterators for keys that are never considered. (A similar bit of lazy
168 : // evaluation can be done for the point iterators, but is still TBD).
169 : //
170 : // For a full example, consider the following setup:
171 : //
172 : // p0: o
173 : // r0: m---q
174 : //
175 : // p1: n p
176 : // r1: g---k
177 : //
178 : // p2: b d i
179 : // r2: a---e q----v
180 : //
181 : // p3: e
182 : // r3:
183 : //
184 : // If we start iterating from the beginning, the first key we encounter is "b"
185 : // in p2. When the mergingIter is pointing at a valid entry, the range deletion
186 : // iterators for all of the levels < m.heap.items[0].index are positioned at
187 : // the next range tombstone past the current key. So r0 will point at [m,q) and
188 : // r1 at [g,k). When the key "b" is encountered, we check to see if the current
189 : // tombstone for r0 or r1 contains it, and whether the tombstone for r2, [a,e),
190 : // contains and is newer than "b".
191 : //
192 : // Advancing the iterator finds the next key at "d". This is in the same level
193 : // as the previous key "b" so we don't have to reposition any of the range
194 : // deletion iterators, but merely check whether "d" is now contained by any of
195 : // the range tombstones at higher levels or has stepped past the range
196 : // tombstone in its own level or higher levels. In this case, there is nothing to be done.
197 : //
198 : // Advancing the iterator again finds "e". Since "e" comes from p3, we have to
199 : // position the r3 range deletion iterator, which is empty. "e" is past the r2
200 : // tombstone of [a,e) so we need to advance the r2 range deletion iterator to
201 : // [q,v).
202 : //
203 : // The next key is "i". Because this key is in p2, a level above "e", we don't
204 : // have to reposition any range deletion iterators and instead see that "i" is
205 : // covered by the range tombstone [g,k). The iterator is immediately advanced
206 : // to "n" which is covered by the range tombstone [m,q) causing the iterator to
207 : // advance to "o" which is visible.
208 : //
209 : // # Error handling
210 : //
211 : // Any iterator operation may fail. The InternalIterator contract dictates that
212 : // an iterator must return a nil internal key when an error occurs, and a
213 : // subsequent call to Error() should return the error value. The exported
214 : // merging iterator positioning methods must adhere to this contract by setting
215 : // m.err to hold any error encountered by the individual level iterators and
216 : // returning a nil internal key. Some internal helpers (eg,
217 : // find[Next|Prev]Entry) also adhere to this contract, setting m.err directly).
218 : // Other internal functions return an explicit error return value and DO NOT set
219 : // m.err, relying on the caller to set m.err appropriately.
220 : //
221 : // TODO(jackson): Update the InternalIterator interface to return explicit error
222 : // return values (and an *InternalKV pointer).
223 : //
224 : // TODO(peter,rangedel): For testing, advance the iterator through various
225 : // scenarios and have each step display the current state (i.e. the current
226 : // heap and range-del iterator positioning).
227 : type mergingIter struct {
228 : logger Logger
229 : split Split
230 : dir int
231 : snapshot uint64
232 : batchSnapshot uint64
233 : levels []mergingIterLevel
234 : heap mergingIterHeap
235 : err error
236 : prefix []byte
237 : lower []byte
238 : upper []byte
239 : stats *InternalIteratorStats
240 :
241 : // levelsPositioned, if non-nil, is a slice of the same length as levels.
242 : // It's used by NextPrefix to record which levels have already been
243 : // repositioned. It's created lazily by the first call to NextPrefix.
244 : levelsPositioned []bool
245 :
246 : combinedIterState *combinedIterState
247 :
248 : // Used in some tests to disable the random disabling of seek optimizations.
249 : forceEnableSeekOpt bool
250 : }
251 :
252 : // mergingIter implements the base.InternalIterator interface.
253 : var _ base.InternalIterator = (*mergingIter)(nil)
254 :
255 : // newMergingIter returns an iterator that merges its input. Walking the
256 : // resultant iterator will return all key/value pairs of all input iterators
257 : // in strictly increasing key order, as defined by cmp. It is permissible to
258 : // pass a nil split parameter if the caller is never going to call
259 : // SeekPrefixGE.
260 : //
261 : // The input's key ranges may overlap, but there are assumed to be no duplicate
262 : // keys: if iters[i] contains a key k then iters[j] will not contain that key k.
263 : //
264 : // None of the iters may be nil.
265 : func newMergingIter(
266 : logger Logger,
267 : stats *base.InternalIteratorStats,
268 : cmp Compare,
269 : split Split,
270 : iters ...internalIterator,
271 1 : ) *mergingIter {
272 1 : m := &mergingIter{}
273 1 : levels := make([]mergingIterLevel, len(iters))
274 1 : for i := range levels {
275 1 : levels[i].iter = iters[i]
276 1 : }
277 1 : m.init(&IterOptions{logger: logger}, stats, cmp, split, levels...)
278 1 : return m
279 : }
280 :
281 : func (m *mergingIter) init(
282 : opts *IterOptions,
283 : stats *base.InternalIteratorStats,
284 : cmp Compare,
285 : split Split,
286 : levels ...mergingIterLevel,
287 1 : ) {
288 1 : m.err = nil // clear cached iteration error
289 1 : m.logger = opts.getLogger()
290 1 : if opts != nil {
291 1 : m.lower = opts.LowerBound
292 1 : m.upper = opts.UpperBound
293 1 : }
294 1 : m.snapshot = InternalKeySeqNumMax
295 1 : m.batchSnapshot = InternalKeySeqNumMax
296 1 : m.levels = levels
297 1 : m.heap.cmp = cmp
298 1 : m.split = split
299 1 : m.stats = stats
300 1 : if cap(m.heap.items) < len(levels) {
301 1 : m.heap.items = make([]*mergingIterLevel, 0, len(levels))
302 1 : } else {
303 1 : m.heap.items = m.heap.items[:0]
304 1 : }
305 1 : for l := range m.levels {
306 1 : m.levels[l].index = l
307 1 : }
308 : }
309 :
310 1 : func (m *mergingIter) initHeap() {
311 1 : m.heap.items = m.heap.items[:0]
312 1 : for i := range m.levels {
313 1 : if l := &m.levels[i]; l.iterKV != nil {
314 1 : m.heap.items = append(m.heap.items, l)
315 1 : }
316 : }
317 1 : m.heap.init()
318 : }
319 :
320 1 : func (m *mergingIter) initMinHeap() error {
321 1 : m.dir = 1
322 1 : m.heap.reverse = false
323 1 : m.initHeap()
324 1 : return m.initMinRangeDelIters(-1)
325 1 : }
326 :
327 : // The level of the previous top element was oldTopLevel. Note that all range delete
328 : // iterators < oldTopLevel are positioned past the key of the previous top element and
329 : // the range delete iterator == oldTopLevel is positioned at or past the key of the
330 : // previous top element. We need to position the range delete iterators from oldTopLevel + 1
331 : // to the level of the current top element.
332 1 : func (m *mergingIter) initMinRangeDelIters(oldTopLevel int) error {
333 1 : if m.heap.len() == 0 {
334 1 : return nil
335 1 : }
336 :
337 : // Position the range-del iterators at levels <= m.heap.items[0].index.
338 1 : item := m.heap.items[0]
339 1 : for level := oldTopLevel + 1; level <= item.index; level++ {
340 1 : l := &m.levels[level]
341 1 : if l.rangeDelIter == nil {
342 1 : continue
343 : }
344 1 : var err error
345 1 : l.tombstone, err = l.rangeDelIter.SeekGE(item.iterKV.K.UserKey)
346 1 : if err != nil {
347 0 : return err
348 0 : }
349 : }
350 1 : return nil
351 : }
352 :
353 1 : func (m *mergingIter) initMaxHeap() error {
354 1 : m.dir = -1
355 1 : m.heap.reverse = true
356 1 : m.initHeap()
357 1 : return m.initMaxRangeDelIters(-1)
358 1 : }
359 :
360 : // The level of the previous top element was oldTopLevel. Note that all range delete
361 : // iterators < oldTopLevel are positioned before the key of the previous top element and
362 : // the range delete iterator == oldTopLevel is positioned at or before the key of the
363 : // previous top element. We need to position the range delete iterators from oldTopLevel + 1
364 : // to the level of the current top element.
365 1 : func (m *mergingIter) initMaxRangeDelIters(oldTopLevel int) error {
366 1 : if m.heap.len() == 0 {
367 1 : return nil
368 1 : }
369 : // Position the range-del iterators at levels <= m.heap.items[0].index.
370 1 : item := m.heap.items[0]
371 1 : for level := oldTopLevel + 1; level <= item.index; level++ {
372 1 : l := &m.levels[level]
373 1 : if l.rangeDelIter == nil {
374 1 : continue
375 : }
376 1 : tomb, err := keyspan.SeekLE(m.heap.cmp, l.rangeDelIter, item.iterKV.K.UserKey)
377 1 : if err != nil {
378 0 : return err
379 0 : }
380 1 : l.tombstone = tomb
381 : }
382 1 : return nil
383 : }
384 :
385 1 : func (m *mergingIter) switchToMinHeap() error {
386 1 : if m.heap.len() == 0 {
387 1 : if m.lower != nil {
388 1 : m.SeekGE(m.lower, base.SeekGEFlagsNone)
389 1 : } else {
390 1 : m.First()
391 1 : }
392 1 : return m.err
393 : }
394 :
395 : // We're switching from using a max heap to a min heap. We need to advance
396 : // any iterator that is less than or equal to the current key. Consider the
397 : // scenario where we have 2 iterators being merged (user-key:seq-num):
398 : //
399 : // i1: *a:2 b:2
400 : // i2: a:1 b:1
401 : //
402 : // The current key is a:2 and i2 is pointed at a:1. When we switch to forward
403 : // iteration, we want to return a key that is greater than a:2.
404 :
405 1 : key := m.heap.items[0].iterKV.K
406 1 : cur := m.heap.items[0]
407 1 :
408 1 : for i := range m.levels {
409 1 : l := &m.levels[i]
410 1 : if l == cur {
411 1 : continue
412 : }
413 1 : for l.iterKV = l.iter.Next(); l.iterKV != nil; l.iterKV = l.iter.Next() {
414 1 : if base.InternalCompare(m.heap.cmp, key, l.iterKV.K) < 0 {
415 1 : // key < iter-key
416 1 : break
417 : }
418 : // key >= iter-key
419 : }
420 1 : if l.iterKV == nil {
421 1 : if err := l.iter.Error(); err != nil {
422 1 : return err
423 1 : }
424 : }
425 : }
426 :
427 : // Special handling for the current iterator because we were using its key
428 : // above.
429 1 : cur.iterKV = cur.iter.Next()
430 1 : if cur.iterKV == nil {
431 1 : if err := cur.iter.Error(); err != nil {
432 1 : return err
433 1 : }
434 : }
435 1 : return m.initMinHeap()
436 : }
437 :
438 1 : func (m *mergingIter) switchToMaxHeap() error {
439 1 : if m.heap.len() == 0 {
440 1 : if m.upper != nil {
441 1 : m.SeekLT(m.upper, base.SeekLTFlagsNone)
442 1 : } else {
443 1 : m.Last()
444 1 : }
445 1 : return m.err
446 : }
447 :
448 : // We're switching from using a min heap to a max heap. We need to backup any
449 : // iterator that is greater than or equal to the current key. Consider the
450 : // scenario where we have 2 iterators being merged (user-key:seq-num):
451 : //
452 : // i1: a:2 *b:2
453 : // i2: a:1 b:1
454 : //
455 : // The current key is b:2 and i2 is pointing at b:1. When we switch to
456 : // reverse iteration, we want to return a key that is less than b:2.
457 1 : key := m.heap.items[0].iterKV.K
458 1 : cur := m.heap.items[0]
459 1 :
460 1 : for i := range m.levels {
461 1 : l := &m.levels[i]
462 1 : if l == cur {
463 1 : continue
464 : }
465 :
466 1 : for l.iterKV = l.iter.Prev(); l.iterKV != nil; l.iterKV = l.iter.Prev() {
467 1 : if base.InternalCompare(m.heap.cmp, key, l.iterKV.K) > 0 {
468 1 : // key > iter-key
469 1 : break
470 : }
471 : // key <= iter-key
472 : }
473 1 : if l.iterKV == nil {
474 1 : if err := l.iter.Error(); err != nil {
475 1 : return err
476 1 : }
477 : }
478 : }
479 :
480 : // Special handling for the current iterator because we were using its key
481 : // above.
482 1 : cur.iterKV = cur.iter.Prev()
483 1 : if cur.iterKV == nil {
484 1 : if err := cur.iter.Error(); err != nil {
485 1 : return err
486 1 : }
487 : }
488 1 : return m.initMaxHeap()
489 : }
490 :
491 : // nextEntry unconditionally steps to the next entry. item is the current top
492 : // item in the heap.
493 1 : func (m *mergingIter) nextEntry(l *mergingIterLevel, succKey []byte) error {
494 1 : // INVARIANT: If in prefix iteration mode, item.iterKey must have a prefix equal
495 1 : // to m.prefix. This invariant is important for ensuring TrySeekUsingNext
496 1 : // optimizations behave correctly.
497 1 : //
498 1 : // During prefix iteration, the iterator does not have a full view of the
499 1 : // LSM. Some level iterators may omit keys that are known to fall outside
500 1 : // the seek prefix (eg, due to sstable bloom filter exclusion). It's
501 1 : // important that in such cases we don't position any iterators beyond
502 1 : // m.prefix, because doing so may interfere with future seeks.
503 1 : //
504 1 : // Let prefixes P1 < P2 < P3. Imagine a SeekPrefixGE to prefix P1, followed
505 1 : // by a SeekPrefixGE to prefix P2. Imagine there exist live keys at prefix
506 1 : // P2, but they're not visible to the SeekPrefixGE(P1) (because of
507 1 : // bloom-filter exclusion or a range tombstone that deletes prefix P1 but
508 1 : // not P2). If the SeekPrefixGE(P1) is allowed to move any level iterators
509 1 : // to P3, the SeekPrefixGE(P2, TrySeekUsingNext=true) may mistakenly think
510 1 : // the level contains no point keys or range tombstones within the prefix
511 1 : // P2. Care is taken to avoid ever advancing the iterator beyond the current
512 1 : // prefix. If nextEntry is ever invoked while we're already beyond the
513 1 : // current prefix, we're violating the invariant.
514 1 : if invariants.Enabled && m.prefix != nil {
515 1 : if p := m.split.Prefix(l.iterKV.K.UserKey); !bytes.Equal(m.prefix, p) {
516 0 : m.logger.Fatalf("mergingIter: prefix violation: nexting beyond prefix %q; existing heap root %q\n%s",
517 0 : m.prefix, l.iterKV, debug.Stack())
518 0 : }
519 : }
520 :
521 1 : oldTopLevel := l.index
522 1 : oldRangeDelIter := l.rangeDelIter
523 1 :
524 1 : if succKey == nil {
525 1 : l.iterKV = l.iter.Next()
526 1 : } else {
527 1 : l.iterKV = l.iter.NextPrefix(succKey)
528 1 : }
529 :
530 1 : if l.iterKV == nil {
531 1 : if err := l.iter.Error(); err != nil {
532 1 : return err
533 1 : }
534 1 : m.heap.pop()
535 1 : } else {
536 1 : if m.prefix != nil && !bytes.Equal(m.prefix, m.split.Prefix(l.iterKV.K.UserKey)) {
537 1 : // Set keys without a matching prefix to their zero values when in prefix
538 1 : // iteration mode and remove iterated level from heap.
539 1 : l.iterKV = nil
540 1 : m.heap.pop()
541 1 : } else if m.heap.len() > 1 {
542 1 : m.heap.fix(0)
543 1 : }
544 1 : if l.rangeDelIter != oldRangeDelIter {
545 1 : // The rangeDelIter changed which indicates that the l.iter moved to the
546 1 : // next sstable. We have to update the tombstone for oldTopLevel as well.
547 1 : oldTopLevel--
548 1 : }
549 : }
550 :
551 : // The cached tombstones are only valid for the levels
552 : // [0,oldTopLevel]. Updated the cached tombstones for any levels in the range
553 : // [oldTopLevel+1,heap[0].index].
554 1 : return m.initMinRangeDelIters(oldTopLevel)
555 : }
556 :
557 : // isNextEntryDeleted starts from the current entry (as the next entry) and if
558 : // it is deleted, moves the iterators forward as needed and returns true, else
559 : // it returns false. item is the top item in the heap. If any of the required
560 : // iterator operations error, the error is returned without updating m.err.
561 : //
562 : // During prefix iteration mode, isNextEntryDeleted will exhaust the iterator by
563 : // clearing the heap if the deleted key(s) extend beyond the iteration prefix
564 : // during prefix-iteration mode.
565 1 : func (m *mergingIter) isNextEntryDeleted(item *mergingIterLevel) (bool, error) {
566 1 : // Look for a range deletion tombstone containing item.iterKV at higher
567 1 : // levels (level < item.index). If we find such a range tombstone we know
568 1 : // it deletes the key in the current level. Also look for a range
569 1 : // deletion at the current level (level == item.index). If we find such a
570 1 : // range deletion we need to check whether it is newer than the current
571 1 : // entry.
572 1 : for level := 0; level <= item.index; level++ {
573 1 : l := &m.levels[level]
574 1 : if l.rangeDelIter == nil || l.tombstone == nil {
575 1 : // If l.tombstone is nil, there are no further tombstones
576 1 : // in the current sstable in the current (forward) iteration
577 1 : // direction.
578 1 : continue
579 : }
580 1 : if m.heap.cmp(l.tombstone.End, item.iterKV.K.UserKey) <= 0 {
581 1 : // The current key is at or past the tombstone end key.
582 1 : //
583 1 : // NB: for the case that this l.rangeDelIter is provided by a levelIter we know that
584 1 : // the levelIter must be positioned at a key >= item.iterKV. So it is sufficient to seek the
585 1 : // current l.rangeDelIter (since any range del iterators that will be provided by the
586 1 : // levelIter in the future cannot contain item.iterKV). Also, it is possible that we
587 1 : // will encounter parts of the range delete that should be ignored -- we handle that
588 1 : // below.
589 1 : var err error
590 1 : l.tombstone, err = l.rangeDelIter.SeekGE(item.iterKV.K.UserKey)
591 1 : if err != nil {
592 1 : return false, err
593 1 : }
594 : }
595 1 : if l.tombstone == nil {
596 1 : continue
597 : }
598 :
599 1 : if l.tombstone.VisibleAt(m.snapshot) && m.heap.cmp(l.tombstone.Start, item.iterKV.K.UserKey) <= 0 {
600 1 : if level < item.index {
601 1 : // We could also do m.seekGE(..., level + 1). The levels from
602 1 : // [level + 1, item.index) are already after item.iterKV so seeking them may be
603 1 : // wasteful.
604 1 :
605 1 : // We can seek up to tombstone.End.
606 1 : //
607 1 : // Progress argument: Since this file is at a higher level than item.iterKV we know
608 1 : // that the iterator in this file must be positioned within its bounds and at a key
609 1 : // X > item.iterKV (otherwise it would be the min of the heap). It is not
610 1 : // possible for X.UserKey == item.iterKV.UserKey, since it is incompatible with
611 1 : // X > item.iterKV (a lower version cannot be in a higher sstable), so it must be that
612 1 : // X.UserKey > item.iterKV.UserKey. Which means l.largestUserKey > item.key.UserKey.
613 1 : // We also know that l.tombstone.End > item.iterKV.UserKey. So the min of these,
614 1 : // seekKey, computed below, is > item.iterKV.UserKey, so the call to seekGE() will
615 1 : // make forward progress.
616 1 : seekKey := l.tombstone.End
617 1 : // This seek is not directly due to a SeekGE call, so we don't know
618 1 : // enough about the underlying iterator positions, and so we keep the
619 1 : // try-seek-using-next optimization disabled. Additionally, if we're in
620 1 : // prefix-seek mode and a re-seek would have moved us past the original
621 1 : // prefix, we can remove all merging iter levels below the rangedel
622 1 : // tombstone's level and return immediately instead of re-seeking. This
623 1 : // is correct since those levels cannot provide a key that matches the
624 1 : // prefix, and is also visible. Additionally, this is important to make
625 1 : // subsequent `TrySeekUsingNext` work correctly, as a re-seek on a
626 1 : // different prefix could have resulted in this iterator skipping visible
627 1 : // keys at prefixes in between m.prefix and seekKey, that are currently
628 1 : // not in the heap due to a bloom filter mismatch.
629 1 : //
630 1 : // Additionally, we set the relative-seek flag. This is
631 1 : // important when iterating with lazy combined iteration. If
632 1 : // there's a range key between this level's current file and the
633 1 : // file the seek will land on, we need to detect it in order to
634 1 : // trigger construction of the combined iterator.
635 1 : if m.prefix != nil {
636 1 : if !bytes.Equal(m.prefix, m.split.Prefix(seekKey)) {
637 1 : for i := item.index; i < len(m.levels); i++ {
638 1 : // Remove this level from the heap. Setting iterKV
639 1 : // to nil should be sufficient for initMinHeap to
640 1 : // not re-initialize the heap with them in it. Other
641 1 : // fields in mergingIterLevel can remain as-is; the
642 1 : // iter/rangeDelIter needs to stay intact for future
643 1 : // trySeekUsingNexts to work, the level iter
644 1 : // boundary context is owned by the levelIter which
645 1 : // is not being repositioned, and any tombstones in
646 1 : // these levels will be irrelevant for us anyway.
647 1 : m.levels[i].iterKV = nil
648 1 : }
649 : // TODO(bilal): Consider a more efficient way of removing levels from
650 : // the heap without reinitializing all of it. This would likely
651 : // necessitate tracking the heap positions of each mergingIterHeap
652 : // item in the mergingIterLevel, and then swapping that item in the
653 : // heap with the last-positioned heap item, and shrinking the heap by
654 : // one.
655 1 : if err := m.initMinHeap(); err != nil {
656 0 : return false, err
657 0 : }
658 1 : return true, nil
659 : }
660 : }
661 1 : if err := m.seekGE(seekKey, item.index, base.SeekGEFlagsNone.EnableRelativeSeek()); err != nil {
662 1 : return false, err
663 1 : }
664 1 : return true, nil
665 : }
666 1 : if l.tombstone.CoversAt(m.snapshot, item.iterKV.SeqNum()) {
667 1 : if err := m.nextEntry(item, nil /* succKey */); err != nil {
668 0 : return false, err
669 0 : }
670 1 : return true, nil
671 : }
672 : }
673 : }
674 1 : return false, nil
675 : }
676 :
677 : // Starting from the current entry, finds the first (next) entry that can be returned.
678 : //
679 : // If an error occurs, m.err is updated to hold the error and findNextentry
680 : // returns a nil internal key.
681 1 : func (m *mergingIter) findNextEntry() *base.InternalKV {
682 1 : for m.heap.len() > 0 && m.err == nil {
683 1 : item := m.heap.items[0]
684 1 :
685 1 : // The levelIter internal iterator will interleave exclusive sentinel
686 1 : // keys to keep files open until their range deletions are no longer
687 1 : // necessary. Sometimes these are interleaved with the user key of a
688 1 : // file's largest key, in which case they may simply be stepped over to
689 1 : // move to the next file in the forward direction. Other times they're
690 1 : // interleaved at the user key of the user-iteration boundary, if that
691 1 : // falls within the bounds of a file. In the latter case, there are no
692 1 : // more keys < m.upper, and we can stop iterating.
693 1 : //
694 1 : // We perform a key comparison to differentiate between these two cases.
695 1 : // This key comparison is considered okay because it only happens for
696 1 : // sentinel keys. It may be eliminated after #2863.
697 1 : if m.levels[item.index].iterKV.K.IsExclusiveSentinel() {
698 1 : if m.upper != nil && m.heap.cmp(m.levels[item.index].iterKV.K.UserKey, m.upper) >= 0 {
699 1 : break
700 : }
701 : // This key is the largest boundary of a file and can be skipped now
702 : // that the file's range deletions are no longer relevant.
703 1 : m.err = m.nextEntry(item, nil /* succKey */)
704 1 : if m.err != nil {
705 1 : return nil
706 1 : }
707 1 : continue
708 : }
709 :
710 1 : m.addItemStats(item)
711 1 :
712 1 : // Check if the heap root key is deleted by a range tombstone in a
713 1 : // higher level. If it is, isNextEntryDeleted will advance the iterator
714 1 : // to a later key (through seeking or nexting).
715 1 : isDeleted, err := m.isNextEntryDeleted(item)
716 1 : if err != nil {
717 1 : m.err = err
718 1 : return nil
719 1 : } else if isDeleted {
720 1 : m.stats.PointsCoveredByRangeTombstones++
721 1 : continue
722 : }
723 :
724 : // Check if the key is visible at the iterator sequence numbers.
725 1 : if !item.iterKV.Visible(m.snapshot, m.batchSnapshot) {
726 1 : m.err = m.nextEntry(item, nil /* succKey */)
727 1 : if m.err != nil {
728 0 : return nil
729 0 : }
730 1 : continue
731 : }
732 :
733 : // The heap root is visible and not deleted by any range tombstones.
734 : // Return it.
735 1 : return item.iterKV
736 : }
737 1 : return nil
738 : }
739 :
740 : // Steps to the prev entry. item is the current top item in the heap.
741 1 : func (m *mergingIter) prevEntry(l *mergingIterLevel) error {
742 1 : oldTopLevel := l.index
743 1 : oldRangeDelIter := l.rangeDelIter
744 1 : if l.iterKV = l.iter.Prev(); l.iterKV != nil {
745 1 : if m.heap.len() > 1 {
746 1 : m.heap.fix(0)
747 1 : }
748 1 : if l.rangeDelIter != oldRangeDelIter && l.rangeDelIter != nil {
749 1 : // The rangeDelIter changed which indicates that the l.iter moved to the
750 1 : // previous sstable. We have to update the tombstone for oldTopLevel as
751 1 : // well.
752 1 : oldTopLevel--
753 1 : }
754 1 : } else {
755 1 : if err := l.iter.Error(); err != nil {
756 1 : return err
757 1 : }
758 1 : m.heap.pop()
759 : }
760 :
761 : // The cached tombstones are only valid for the levels
762 : // [0,oldTopLevel]. Updated the cached tombstones for any levels in the range
763 : // [oldTopLevel+1,heap[0].index].
764 1 : return m.initMaxRangeDelIters(oldTopLevel)
765 : }
766 :
767 : // isPrevEntryDeleted() starts from the current entry (as the prev entry) and if it is deleted,
768 : // moves the iterators backward as needed and returns true, else it returns false. item is the top
769 : // item in the heap.
770 1 : func (m *mergingIter) isPrevEntryDeleted(item *mergingIterLevel) (bool, error) {
771 1 : // Look for a range deletion tombstone containing item.iterKV at higher
772 1 : // levels (level < item.index). If we find such a range tombstone we know
773 1 : // it deletes the key in the current level. Also look for a range
774 1 : // deletion at the current level (level == item.index). If we find such a
775 1 : // range deletion we need to check whether it is newer than the current
776 1 : // entry.
777 1 : for level := 0; level <= item.index; level++ {
778 1 : l := &m.levels[level]
779 1 : if l.rangeDelIter == nil || l.tombstone == nil {
780 1 : // If l.tombstone is nil, there are no further tombstones
781 1 : // in the current sstable in the current (reverse) iteration
782 1 : // direction.
783 1 : continue
784 : }
785 1 : if m.heap.cmp(item.iterKV.K.UserKey, l.tombstone.Start) < 0 {
786 1 : // The current key is before the tombstone start key.
787 1 : //
788 1 : // NB: for the case that this l.rangeDelIter is provided by a levelIter we know that
789 1 : // the levelIter must be positioned at a key < item.iterKV. So it is sufficient to seek the
790 1 : // current l.rangeDelIter (since any range del iterators that will be provided by the
791 1 : // levelIter in the future cannot contain item.iterKV). Also, it is it is possible that we
792 1 : // will encounter parts of the range delete that should be ignored -- we handle that
793 1 : // below.
794 1 :
795 1 : tomb, err := keyspan.SeekLE(m.heap.cmp, l.rangeDelIter, item.iterKV.K.UserKey)
796 1 : if err != nil {
797 0 : return false, err
798 0 : }
799 1 : l.tombstone = tomb
800 : }
801 1 : if l.tombstone == nil {
802 1 : continue
803 : }
804 1 : if l.tombstone.VisibleAt(m.snapshot) && m.heap.cmp(l.tombstone.End, item.iterKV.K.UserKey) > 0 {
805 1 : if level < item.index {
806 1 : // We could also do m.seekLT(..., level + 1). The levels from
807 1 : // [level + 1, item.index) are already before item.iterKV so seeking them may be
808 1 : // wasteful.
809 1 :
810 1 : // We can seek up to tombstone.Start.UserKey.
811 1 : //
812 1 : // Progress argument: We know that the iterator in this file is positioned within
813 1 : // its bounds and at a key X < item.iterKV (otherwise it would be the max of the heap).
814 1 : // So smallestUserKey <= item.iterKV.UserKey and we already know that
815 1 : // l.tombstone.Start.UserKey <= item.iterKV.UserKey. So the seekKey computed below
816 1 : // is <= item.iterKV.UserKey, and since we do a seekLT() we will make backwards
817 1 : // progress.
818 1 : seekKey := l.tombstone.Start
819 1 : // We set the relative-seek flag. This is important when
820 1 : // iterating with lazy combined iteration. If there's a range
821 1 : // key between this level's current file and the file the seek
822 1 : // will land on, we need to detect it in order to trigger
823 1 : // construction of the combined iterator.
824 1 : if err := m.seekLT(seekKey, item.index, base.SeekLTFlagsNone.EnableRelativeSeek()); err != nil {
825 1 : return false, err
826 1 : }
827 1 : return true, nil
828 : }
829 1 : if l.tombstone.CoversAt(m.snapshot, item.iterKV.SeqNum()) {
830 1 : if err := m.prevEntry(item); err != nil {
831 0 : return false, err
832 0 : }
833 1 : return true, nil
834 : }
835 : }
836 : }
837 1 : return false, nil
838 : }
839 :
840 : // Starting from the current entry, finds the first (prev) entry that can be returned.
841 : //
842 : // If an error occurs, m.err is updated to hold the error and findNextentry
843 : // returns a nil internal key.
844 1 : func (m *mergingIter) findPrevEntry() *base.InternalKV {
845 1 : for m.heap.len() > 0 && m.err == nil {
846 1 : item := m.heap.items[0]
847 1 :
848 1 : // The levelIter internal iterator will interleave exclusive sentinel
849 1 : // keys to keep files open until their range deletions are no longer
850 1 : // necessary. Sometimes these are interleaved with the user key of a
851 1 : // file's smallest key, in which case they may simply be stepped over to
852 1 : // move to the next file in the backward direction. Other times they're
853 1 : // interleaved at the user key of the user-iteration boundary, if that
854 1 : // falls within the bounds of a file. In the latter case, there are no
855 1 : // more keys ≥ m.lower, and we can stop iterating.
856 1 : //
857 1 : // We perform a key comparison to differentiate between these two cases.
858 1 : // This key comparison is considered okay because it only happens for
859 1 : // sentinel keys. It may be eliminated after #2863.
860 1 : if m.levels[item.index].iterKV.K.IsExclusiveSentinel() {
861 1 : if m.lower != nil && m.heap.cmp(m.levels[item.index].iterKV.K.UserKey, m.lower) <= 0 {
862 1 : break
863 : }
864 : // This key is the smallest boundary of a file and can be skipped
865 : // now that the file's range deletions are no longer relevant.
866 1 : m.err = m.prevEntry(item)
867 1 : if m.err != nil {
868 1 : return nil
869 1 : }
870 1 : continue
871 : }
872 :
873 1 : m.addItemStats(item)
874 1 : if isDeleted, err := m.isPrevEntryDeleted(item); err != nil {
875 1 : m.err = err
876 1 : return nil
877 1 : } else if isDeleted {
878 1 : m.stats.PointsCoveredByRangeTombstones++
879 1 : continue
880 : }
881 1 : if item.iterKV.Visible(m.snapshot, m.batchSnapshot) {
882 1 : return item.iterKV
883 1 : }
884 1 : m.err = m.prevEntry(item)
885 : }
886 1 : return nil
887 : }
888 :
889 : // Seeks levels >= level to >= key. Additionally uses range tombstones to extend the seeks.
890 : //
891 : // If an error occurs, seekGE returns the error without setting m.err.
892 1 : func (m *mergingIter) seekGE(key []byte, level int, flags base.SeekGEFlags) error {
893 1 : // When seeking, we can use tombstones to adjust the key we seek to on each
894 1 : // level. Consider the series of range tombstones:
895 1 : //
896 1 : // 1: a---e
897 1 : // 2: d---h
898 1 : // 3: g---k
899 1 : // 4: j---n
900 1 : // 5: m---q
901 1 : //
902 1 : // If we SeekGE("b") we also find the tombstone "b" resides within in the
903 1 : // first level which is [a,e). Regardless of whether this tombstone deletes
904 1 : // "b" in that level, we know it deletes "b" in all lower levels, so we
905 1 : // adjust the search key in the next level to the tombstone end key "e". We
906 1 : // then SeekGE("e") in the second level and find the corresponding tombstone
907 1 : // [d,h). This process continues and we end up seeking for "h" in the 3rd
908 1 : // level, "k" in the 4th level and "n" in the last level.
909 1 : //
910 1 : // TODO(peter,rangedel): In addition to the above we can delay seeking a
911 1 : // level (and any lower levels) when the current iterator position is
912 1 : // contained within a range tombstone at a higher level.
913 1 :
914 1 : // Deterministically disable the TrySeekUsingNext optimizations sometimes in
915 1 : // invariant builds to encourage the metamorphic tests to surface bugs. Note
916 1 : // that we cannot disable the optimization within individual levels. It must
917 1 : // be disabled for all levels or none. If one lower-level iterator performs
918 1 : // a fresh seek whereas another takes advantage of its current iterator
919 1 : // position, the heap can become inconsistent. Consider the following
920 1 : // example:
921 1 : //
922 1 : // L5: [ [b-c) ] [ d ]*
923 1 : // L6: [ b ] [e]*
924 1 : //
925 1 : // Imagine a SeekGE(a). The [b-c) range tombstone deletes the L6 point key
926 1 : // 'b', resulting in the iterator positioned at d with the heap:
927 1 : //
928 1 : // {L5: d, L6: e}
929 1 : //
930 1 : // A subsequent SeekGE(b) is seeking to a larger key, so the caller may set
931 1 : // TrySeekUsingNext()=true. If the L5 iterator used the TrySeekUsingNext
932 1 : // optimization but the L6 iterator did not, the iterator would have the
933 1 : // heap:
934 1 : //
935 1 : // {L6: b, L5: d}
936 1 : //
937 1 : // Because the L5 iterator has already advanced to the next sstable, the
938 1 : // merging iterator cannot observe the [b-c) range tombstone and will
939 1 : // mistakenly return L6's deleted point key 'b'.
940 1 : if invariants.Enabled && flags.TrySeekUsingNext() && !m.forceEnableSeekOpt &&
941 1 : disableSeekOpt(key, uintptr(unsafe.Pointer(m))) {
942 1 : flags = flags.DisableTrySeekUsingNext()
943 1 : }
944 :
945 1 : for ; level < len(m.levels); level++ {
946 1 : if invariants.Enabled && m.lower != nil && m.heap.cmp(key, m.lower) < 0 {
947 0 : m.logger.Fatalf("mergingIter: lower bound violation: %s < %s\n%s", key, m.lower, debug.Stack())
948 0 : }
949 :
950 1 : l := &m.levels[level]
951 1 : if m.prefix != nil {
952 1 : l.iterKV = l.iter.SeekPrefixGE(m.prefix, key, flags)
953 1 : if l.iterKV != nil {
954 1 : if !bytes.Equal(m.prefix, m.split.Prefix(l.iterKV.K.UserKey)) {
955 1 : // Prevent keys without a matching prefix from being added to the heap by setting
956 1 : // iterKey and iterValue to their zero values before calling initMinHeap.
957 1 : l.iterKV = nil
958 1 : }
959 : }
960 1 : } else {
961 1 : l.iterKV = l.iter.SeekGE(key, flags)
962 1 : }
963 1 : if l.iterKV == nil {
964 1 : if err := l.iter.Error(); err != nil {
965 1 : return err
966 1 : }
967 : }
968 :
969 : // If this level contains overlapping range tombstones, alter the seek
970 : // key accordingly. Caveat: If we're performing lazy-combined iteration,
971 : // we cannot alter the seek key: Range tombstones don't delete range
972 : // keys, and there might exist live range keys within the range
973 : // tombstone's span that need to be observed to trigger a switch to
974 : // combined iteration.
975 1 : if rangeDelIter := l.rangeDelIter; rangeDelIter != nil &&
976 1 : (m.combinedIterState == nil || m.combinedIterState.initialized) {
977 1 : // The level has a range-del iterator. Find the tombstone containing
978 1 : // the search key.
979 1 : var err error
980 1 : l.tombstone, err = rangeDelIter.SeekGE(key)
981 1 : if err != nil {
982 0 : return err
983 0 : }
984 1 : if l.tombstone != nil && l.tombstone.VisibleAt(m.snapshot) && m.heap.cmp(l.tombstone.Start, key) <= 0 {
985 1 : // Based on the containment condition tombstone.End > key, so
986 1 : // the assignment to key results in a monotonically
987 1 : // non-decreasing key across iterations of this loop.
988 1 : //
989 1 : // The adjustment of key here can only move it to a larger key.
990 1 : // Since the caller of seekGE guaranteed that the original key
991 1 : // was greater than or equal to m.lower, the new key will
992 1 : // continue to be greater than or equal to m.lower.
993 1 : key = l.tombstone.End
994 1 : }
995 : }
996 : }
997 1 : return m.initMinHeap()
998 : }
999 :
1000 0 : func (m *mergingIter) String() string {
1001 0 : return "merging"
1002 0 : }
1003 :
1004 : // SeekGE implements base.InternalIterator.SeekGE. Note that SeekGE only checks
1005 : // the upper bound. It is up to the caller to ensure that key is greater than
1006 : // or equal to the lower bound.
1007 1 : func (m *mergingIter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
1008 1 : m.prefix = nil
1009 1 : m.err = m.seekGE(key, 0 /* start level */, flags)
1010 1 : if m.err != nil {
1011 1 : return nil
1012 1 : }
1013 1 : return m.findNextEntry()
1014 : }
1015 :
1016 : // SeekPrefixGE implements base.InternalIterator.SeekPrefixGE.
1017 1 : func (m *mergingIter) SeekPrefixGE(prefix, key []byte, flags base.SeekGEFlags) *base.InternalKV {
1018 1 : return m.SeekPrefixGEStrict(prefix, key, flags)
1019 1 : }
1020 :
1021 : // SeekPrefixGEStrict implements topLevelIterator.SeekPrefixGEStrict. Note that
1022 : // SeekPrefixGEStrict explicitly checks that the key has a matching prefix.
1023 : func (m *mergingIter) SeekPrefixGEStrict(
1024 : prefix, key []byte, flags base.SeekGEFlags,
1025 1 : ) *base.InternalKV {
1026 1 : m.prefix = prefix
1027 1 : m.err = m.seekGE(key, 0 /* start level */, flags)
1028 1 : if m.err != nil {
1029 1 : return nil
1030 1 : }
1031 :
1032 1 : iterKV := m.findNextEntry()
1033 1 : if invariants.Enabled && iterKV != nil {
1034 1 : if !bytes.Equal(m.prefix, m.split.Prefix(iterKV.K.UserKey)) {
1035 0 : m.logger.Fatalf("mergingIter: prefix violation: returning key %q without prefix %q\n", iterKV, m.prefix)
1036 0 : }
1037 : }
1038 1 : return iterKV
1039 : }
1040 :
1041 : // Seeks levels >= level to < key. Additionally uses range tombstones to extend the seeks.
1042 1 : func (m *mergingIter) seekLT(key []byte, level int, flags base.SeekLTFlags) error {
1043 1 : // See the comment in seekGE regarding using tombstones to adjust the seek
1044 1 : // target per level.
1045 1 : m.prefix = nil
1046 1 : for ; level < len(m.levels); level++ {
1047 1 : if invariants.Enabled && m.upper != nil && m.heap.cmp(key, m.upper) > 0 {
1048 0 : m.logger.Fatalf("mergingIter: upper bound violation: %s > %s\n%s", key, m.upper, debug.Stack())
1049 0 : }
1050 :
1051 1 : l := &m.levels[level]
1052 1 : l.iterKV = l.iter.SeekLT(key, flags)
1053 1 : if l.iterKV == nil {
1054 1 : if err := l.iter.Error(); err != nil {
1055 1 : return err
1056 1 : }
1057 : }
1058 :
1059 : // If this level contains overlapping range tombstones, alter the seek
1060 : // key accordingly. Caveat: If we're performing lazy-combined iteration,
1061 : // we cannot alter the seek key: Range tombstones don't delete range
1062 : // keys, and there might exist live range keys within the range
1063 : // tombstone's span that need to be observed to trigger a switch to
1064 : // combined iteration.
1065 1 : if rangeDelIter := l.rangeDelIter; rangeDelIter != nil &&
1066 1 : (m.combinedIterState == nil || m.combinedIterState.initialized) {
1067 1 : // The level has a range-del iterator. Find the tombstone containing
1068 1 : // the search key.
1069 1 : tomb, err := keyspan.SeekLE(m.heap.cmp, rangeDelIter, key)
1070 1 : if err != nil {
1071 0 : return err
1072 0 : }
1073 1 : l.tombstone = tomb
1074 1 : // Since SeekLT is exclusive on `key` and a tombstone's end key is
1075 1 : // also exclusive, a seek key equal to a tombstone's end key still
1076 1 : // enables the seek optimization (Note this is different than the
1077 1 : // check performed by (*keyspan.Span).Contains).
1078 1 : if l.tombstone != nil && l.tombstone.VisibleAt(m.snapshot) &&
1079 1 : m.heap.cmp(key, l.tombstone.End) <= 0 {
1080 1 : // NB: Based on the containment condition
1081 1 : // tombstone.Start.UserKey <= key, so the assignment to key
1082 1 : // results in a monotonically non-increasing key across
1083 1 : // iterations of this loop.
1084 1 : //
1085 1 : // The adjustment of key here can only move it to a smaller key.
1086 1 : // Since the caller of seekLT guaranteed that the original key
1087 1 : // was less than or equal to m.upper, the new key will continue
1088 1 : // to be less than or equal to m.upper.
1089 1 : key = l.tombstone.Start
1090 1 : }
1091 : }
1092 : }
1093 :
1094 1 : return m.initMaxHeap()
1095 : }
1096 :
1097 : // SeekLT implements base.InternalIterator.SeekLT. Note that SeekLT only checks
1098 : // the lower bound. It is up to the caller to ensure that key is less than the
1099 : // upper bound.
1100 1 : func (m *mergingIter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
1101 1 : m.prefix = nil
1102 1 : m.err = m.seekLT(key, 0 /* start level */, flags)
1103 1 : if m.err != nil {
1104 1 : return nil
1105 1 : }
1106 1 : return m.findPrevEntry()
1107 : }
1108 :
1109 : // First implements base.InternalIterator.First. Note that First only checks
1110 : // the upper bound. It is up to the caller to ensure that key is greater than
1111 : // or equal to the lower bound (e.g. via a call to SeekGE(lower)).
1112 1 : func (m *mergingIter) First() *base.InternalKV {
1113 1 : m.err = nil // clear cached iteration error
1114 1 : m.prefix = nil
1115 1 : m.heap.items = m.heap.items[:0]
1116 1 : for i := range m.levels {
1117 1 : l := &m.levels[i]
1118 1 : l.iterKV = l.iter.First()
1119 1 : if l.iterKV == nil {
1120 1 : if m.err = l.iter.Error(); m.err != nil {
1121 1 : return nil
1122 1 : }
1123 : }
1124 : }
1125 1 : if m.err = m.initMinHeap(); m.err != nil {
1126 0 : return nil
1127 0 : }
1128 1 : return m.findNextEntry()
1129 : }
1130 :
1131 : // Last implements base.InternalIterator.Last. Note that Last only checks the
1132 : // lower bound. It is up to the caller to ensure that key is less than the
1133 : // upper bound (e.g. via a call to SeekLT(upper))
1134 1 : func (m *mergingIter) Last() *base.InternalKV {
1135 1 : m.err = nil // clear cached iteration error
1136 1 : m.prefix = nil
1137 1 : for i := range m.levels {
1138 1 : l := &m.levels[i]
1139 1 : l.iterKV = l.iter.Last()
1140 1 : if l.iterKV == nil {
1141 1 : if m.err = l.iter.Error(); m.err != nil {
1142 1 : return nil
1143 1 : }
1144 : }
1145 : }
1146 1 : if m.err = m.initMaxHeap(); m.err != nil {
1147 0 : return nil
1148 0 : }
1149 1 : return m.findPrevEntry()
1150 : }
1151 :
1152 1 : func (m *mergingIter) Next() *base.InternalKV {
1153 1 : if m.err != nil {
1154 1 : return nil
1155 1 : }
1156 :
1157 1 : if m.dir != 1 {
1158 1 : if m.err = m.switchToMinHeap(); m.err != nil {
1159 1 : return nil
1160 1 : }
1161 1 : return m.findNextEntry()
1162 : }
1163 :
1164 1 : if m.heap.len() == 0 {
1165 1 : return nil
1166 1 : }
1167 :
1168 : // NB: It's okay to call nextEntry directly even during prefix iteration
1169 : // mode. During prefix iteration mode, we rely on the caller to not call
1170 : // Next if the iterator has already advanced beyond the iteration prefix.
1171 : // See the comment above the base.InternalIterator interface.
1172 1 : if m.err = m.nextEntry(m.heap.items[0], nil /* succKey */); m.err != nil {
1173 1 : return nil
1174 1 : }
1175 :
1176 1 : iterKV := m.findNextEntry()
1177 1 : if invariants.Enabled && m.prefix != nil && iterKV != nil {
1178 1 : if !bytes.Equal(m.prefix, m.split.Prefix(iterKV.K.UserKey)) {
1179 0 : m.logger.Fatalf("mergingIter: prefix violation: returning key %q without prefix %q\n", iterKV, m.prefix)
1180 0 : }
1181 : }
1182 1 : return iterKV
1183 : }
1184 :
1185 1 : func (m *mergingIter) NextPrefix(succKey []byte) *base.InternalKV {
1186 1 : if m.dir != 1 {
1187 0 : panic("pebble: cannot switch directions with NextPrefix")
1188 : }
1189 1 : if m.err != nil || m.heap.len() == 0 {
1190 0 : return nil
1191 0 : }
1192 1 : if m.levelsPositioned == nil {
1193 1 : m.levelsPositioned = make([]bool, len(m.levels))
1194 1 : } else {
1195 1 : for i := range m.levelsPositioned {
1196 1 : m.levelsPositioned[i] = false
1197 1 : }
1198 : }
1199 :
1200 : // The heap root necessarily must be positioned at a key < succKey, because
1201 : // NextPrefix was invoked.
1202 1 : root := m.heap.items[0]
1203 1 : if invariants.Enabled && m.heap.cmp((*root).iterKV.K.UserKey, succKey) >= 0 {
1204 0 : m.logger.Fatalf("pebble: invariant violation: NextPrefix(%q) called on merging iterator already positioned at %q",
1205 0 : succKey, (*root).iterKV)
1206 0 : }
1207 : // NB: root is the heap root before we call nextEntry; nextEntry may change
1208 : // the heap root, so we must not `root` to still be the root of the heap, or
1209 : // even to be in the heap if the level's iterator becomes exhausted.
1210 1 : if m.err = m.nextEntry(root, succKey); m.err != nil {
1211 1 : return nil
1212 1 : }
1213 : // We only consider the level to be conclusively positioned at the next
1214 : // prefix if our call to nextEntry did not advance the level onto a range
1215 : // deletion's boundary. Range deletions may have bounds within the prefix
1216 : // that are still surfaced by NextPrefix.
1217 1 : m.levelsPositioned[root.index] = root.iterKV == nil || !root.iterKV.K.IsExclusiveSentinel()
1218 1 :
1219 1 : for m.heap.len() > 0 {
1220 1 : root := m.heap.items[0]
1221 1 : if m.levelsPositioned[root.index] {
1222 1 : // A level we've previously positioned is at the top of the heap, so
1223 1 : // there are no other levels positioned at keys < succKey. We've
1224 1 : // advanced as far as we need to.
1225 1 : break
1226 : }
1227 : // If the current heap root is a sentinel key, we need to skip it.
1228 : // Calling NextPrefix while positioned at a sentinel key is not
1229 : // supported.
1230 1 : if root.iterKV.K.IsExclusiveSentinel() {
1231 1 : if m.err = m.nextEntry(root, nil); m.err != nil {
1232 0 : return nil
1233 0 : }
1234 1 : continue
1235 : }
1236 :
1237 : // Since this level was not the original heap root when NextPrefix was
1238 : // called, we don't know whether this level's current key has the
1239 : // previous prefix or a new one.
1240 1 : if m.heap.cmp(root.iterKV.K.UserKey, succKey) >= 0 {
1241 1 : break
1242 : }
1243 1 : if m.err = m.nextEntry(root, succKey); m.err != nil {
1244 0 : return nil
1245 0 : }
1246 : // We only consider the level to be conclusively positioned at the next
1247 : // prefix if our call to nextEntry did not land onto a range deletion's
1248 : // boundary. Range deletions may have bounds within the prefix that are
1249 : // still surfaced by NextPrefix.
1250 1 : m.levelsPositioned[root.index] = root.iterKV == nil || !root.iterKV.K.IsExclusiveSentinel()
1251 : }
1252 1 : return m.findNextEntry()
1253 : }
1254 :
1255 1 : func (m *mergingIter) Prev() *base.InternalKV {
1256 1 : if m.err != nil {
1257 0 : return nil
1258 0 : }
1259 :
1260 1 : if m.dir != -1 {
1261 1 : if m.prefix != nil {
1262 1 : m.err = errors.New("pebble: unsupported reverse prefix iteration")
1263 1 : return nil
1264 1 : }
1265 1 : if m.err = m.switchToMaxHeap(); m.err != nil {
1266 1 : return nil
1267 1 : }
1268 1 : return m.findPrevEntry()
1269 : }
1270 :
1271 1 : if m.heap.len() == 0 {
1272 1 : return nil
1273 1 : }
1274 1 : if m.err = m.prevEntry(m.heap.items[0]); m.err != nil {
1275 1 : return nil
1276 1 : }
1277 1 : return m.findPrevEntry()
1278 : }
1279 :
1280 1 : func (m *mergingIter) Error() error {
1281 1 : if m.heap.len() == 0 || m.err != nil {
1282 1 : return m.err
1283 1 : }
1284 1 : return m.levels[m.heap.items[0].index].iter.Error()
1285 : }
1286 :
1287 1 : func (m *mergingIter) Close() error {
1288 1 : for i := range m.levels {
1289 1 : iter := m.levels[i].iter
1290 1 : if err := iter.Close(); err != nil && m.err == nil {
1291 0 : m.err = err
1292 0 : }
1293 1 : if rangeDelIter := m.levels[i].rangeDelIter; rangeDelIter != nil {
1294 1 : if err := rangeDelIter.Close(); err != nil && m.err == nil {
1295 0 : m.err = err
1296 0 : }
1297 : }
1298 : }
1299 1 : m.levels = nil
1300 1 : m.heap.items = m.heap.items[:0]
1301 1 : return m.err
1302 : }
1303 :
1304 1 : func (m *mergingIter) SetBounds(lower, upper []byte) {
1305 1 : m.prefix = nil
1306 1 : m.lower = lower
1307 1 : m.upper = upper
1308 1 : for i := range m.levels {
1309 1 : m.levels[i].iter.SetBounds(lower, upper)
1310 1 : }
1311 1 : m.heap.clear()
1312 : }
1313 :
1314 1 : func (m *mergingIter) SetContext(ctx context.Context) {
1315 1 : for i := range m.levels {
1316 1 : m.levels[i].iter.SetContext(ctx)
1317 1 : }
1318 : }
1319 :
1320 0 : func (m *mergingIter) DebugString() string {
1321 0 : var buf bytes.Buffer
1322 0 : sep := ""
1323 0 : for m.heap.len() > 0 {
1324 0 : item := m.heap.pop()
1325 0 : fmt.Fprintf(&buf, "%s%s", sep, item.iterKV.K)
1326 0 : sep = " "
1327 0 : }
1328 0 : var err error
1329 0 : if m.dir == 1 {
1330 0 : err = m.initMinHeap()
1331 0 : } else {
1332 0 : err = m.initMaxHeap()
1333 0 : }
1334 0 : if err != nil {
1335 0 : fmt.Fprintf(&buf, "err=<%s>", err)
1336 0 : }
1337 0 : return buf.String()
1338 : }
1339 :
1340 1 : func (m *mergingIter) ForEachLevelIter(fn func(li *levelIter) bool) {
1341 1 : for _, ml := range m.levels {
1342 1 : if ml.levelIter != nil {
1343 1 : if done := fn(ml.levelIter); done {
1344 1 : break
1345 : }
1346 : }
1347 : }
1348 : }
1349 :
1350 1 : func (m *mergingIter) addItemStats(l *mergingIterLevel) {
1351 1 : m.stats.PointCount++
1352 1 : m.stats.KeyBytes += uint64(len(l.iterKV.K.UserKey))
1353 1 : m.stats.ValueBytes += uint64(len(l.iterKV.V.ValueOrHandle))
1354 1 : }
1355 :
1356 : var _ internalIterator = &mergingIter{}
|
