Line data Source code
1 : // Copyright 2022 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 keyspan
6 :
7 : import (
8 : "bytes"
9 :
10 : "github.com/cockroachdb/pebble/internal/base"
11 : "github.com/cockroachdb/pebble/internal/bytealloc"
12 : "github.com/cockroachdb/pebble/internal/invariants"
13 : )
14 :
15 : // bufferReuseMaxCapacity is the maximum capacity of a DefragmentingIter buffer
16 : // that DefragmentingIter will reuse. Buffers larger than this will be
17 : // discarded and reallocated as necessary.
18 : const bufferReuseMaxCapacity = 10 << 10 // 10 KB
19 :
20 : // keysReuseMaxCapacity is the maximum capacity of a []keyspan.Key buffer that
21 : // DefragmentingIter will reuse. Buffers larger than this will be discarded and
22 : // reallocated as necessary.
23 : const keysReuseMaxCapacity = 100
24 :
25 : // DefragmentMethod configures the defragmentation performed by the
26 : // DefragmentingIter.
27 : type DefragmentMethod interface {
28 : // ShouldDefragment takes two abutting spans and returns whether the two
29 : // spans should be combined into a single, defragmented Span.
30 : ShouldDefragment(equal base.Equal, left, right *Span) bool
31 : }
32 :
33 : // The DefragmentMethodFunc type is an adapter to allow the use of ordinary
34 : // functions as DefragmentMethods. If f is a function with the appropriate
35 : // signature, DefragmentMethodFunc(f) is a DefragmentMethod that calls f.
36 : type DefragmentMethodFunc func(equal base.Equal, left, right *Span) bool
37 :
38 : // ShouldDefragment calls f(equal, left, right).
39 1 : func (f DefragmentMethodFunc) ShouldDefragment(equal base.Equal, left, right *Span) bool {
40 1 : return f(equal, left, right)
41 1 : }
42 :
43 : // DefragmentInternal configures a DefragmentingIter to defragment spans
44 : // only if they have identical keys. It requires spans' keys to be sorted in
45 : // trailer descending order.
46 : //
47 : // This defragmenting method is intended for use in compactions that may see
48 : // internal range keys fragments that may now be joined, because the state that
49 : // required their fragmentation has been dropped.
50 1 : var DefragmentInternal DefragmentMethod = DefragmentMethodFunc(func(equal base.Equal, a, b *Span) bool {
51 1 : if a.KeysOrder != ByTrailerDesc || b.KeysOrder != ByTrailerDesc {
52 0 : panic("pebble: span keys unexpectedly not in trailer descending order")
53 : }
54 1 : if len(a.Keys) != len(b.Keys) {
55 1 : return false
56 1 : }
57 1 : for i := range a.Keys {
58 1 : if a.Keys[i].Trailer != b.Keys[i].Trailer {
59 1 : return false
60 1 : }
61 1 : if !equal(a.Keys[i].Suffix, b.Keys[i].Suffix) {
62 1 : return false
63 1 : }
64 1 : if !bytes.Equal(a.Keys[i].Value, b.Keys[i].Value) {
65 1 : return false
66 1 : }
67 : }
68 1 : return true
69 : })
70 :
71 : // DefragmentReducer merges the current and next Key slices, returning a new Key
72 : // slice.
73 : //
74 : // Implementations should modify and return `cur` to save on allocations, or
75 : // consider allocating a new slice, as the `cur` slice may be retained by the
76 : // DefragmentingIter and mutated. The `next` slice must not be mutated.
77 : //
78 : // The incoming slices are sorted by (SeqNum, Kind) descending. The output slice
79 : // must also have this sort order.
80 : type DefragmentReducer func(cur, next []Key) []Key
81 :
82 : // StaticDefragmentReducer is a no-op DefragmentReducer that simply returns the
83 : // current key slice, effectively retaining the first set of keys encountered
84 : // for a defragmented span.
85 : //
86 : // This reducer can be used, for example, when the set of Keys for each Span
87 : // being reduced is not expected to change, and therefore the keys from the
88 : // first span encountered can be used without considering keys in subsequent
89 : // spans.
90 1 : var StaticDefragmentReducer DefragmentReducer = func(cur, _ []Key) []Key {
91 1 : return cur
92 1 : }
93 :
94 : // iterPos is an enum indicating the position of the defragmenting iter's
95 : // wrapped iter. The defragmenting iter must look ahead or behind when
96 : // defragmenting forward or backwards respectively, and this enum records that
97 : // current position.
98 : type iterPos int8
99 :
100 : const (
101 : iterPosPrev iterPos = -1
102 : iterPosCurr iterPos = 0
103 : iterPosNext iterPos = +1
104 : )
105 :
106 : // DefragmentingIter wraps a key span iterator, defragmenting physical
107 : // fragmentation during iteration.
108 : //
109 : // During flushes and compactions, keys applied over a span may be split at
110 : // sstable boundaries. This fragmentation can produce internal key bounds that
111 : // do not match any of the bounds ever supplied to a user operation. This
112 : // physical fragmentation is necessary to avoid excessively wide sstables.
113 : //
114 : // The defragmenting iterator undoes this physical fragmentation, joining spans
115 : // with abutting bounds and equal state. The defragmenting iterator takes a
116 : // DefragmentMethod to determine what is "equal state" for a span. The
117 : // DefragmentMethod is a function type, allowing arbitrary comparisons between
118 : // Span keys.
119 : //
120 : // Seeking (SeekGE, SeekLT) poses an obstacle to defragmentation. A seek may
121 : // land on a physical fragment in the middle of several fragments that must be
122 : // defragmented. A seek that lands in a fragment straddling the seek key must
123 : // first degfragment in the opposite direction of iteration to find the
124 : // beginning of the defragmented span, and then defragments in the iteration
125 : // direction, ensuring it's found a whole defragmented span.
126 : type DefragmentingIter struct {
127 : // DefragmentingBuffers holds buffers used for copying iterator state.
128 : *DefragmentingBuffers
129 : comparer *base.Comparer
130 : equal base.Equal
131 : iter FragmentIterator
132 : iterSpan *Span
133 : iterPos iterPos
134 :
135 : // curr holds the span at the current iterator position.
136 : curr Span
137 :
138 : // method is a comparison function for two spans. method is called when two
139 : // spans are abutting to determine whether they may be defragmented.
140 : // method does not itself check for adjacency for the two spans.
141 : method DefragmentMethod
142 :
143 : // reduce is the reducer function used to collect Keys across all spans that
144 : // constitute a defragmented span.
145 : reduce DefragmentReducer
146 : }
147 :
148 : // DefragmentingBuffers holds buffers used for copying iterator state.
149 : type DefragmentingBuffers struct {
150 : // currBuf is a buffer for use when copying user keys for curr. currBuf is
151 : // cleared between positioning methods.
152 : currBuf bytealloc.A
153 : // keysBuf is a buffer for use when copying Keys for DefragmentingIter.curr.
154 : keysBuf []Key
155 : // keyBuf is a buffer specifically for the defragmented start key when
156 : // defragmenting backwards or the defragmented end key when defragmenting
157 : // forwards. These bounds are overwritten repeatedly during defragmentation,
158 : // and the defragmentation routines overwrite keyBuf repeatedly to store
159 : // these extended bounds.
160 : keyBuf []byte
161 : }
162 :
163 : // PrepareForReuse discards any excessively large buffers.
164 1 : func (bufs *DefragmentingBuffers) PrepareForReuse() {
165 1 : if cap(bufs.currBuf) > bufferReuseMaxCapacity {
166 1 : bufs.currBuf = nil
167 1 : }
168 1 : if cap(bufs.keyBuf) > bufferReuseMaxCapacity {
169 0 : bufs.keyBuf = nil
170 0 : }
171 1 : if cap(bufs.keysBuf) > keysReuseMaxCapacity {
172 0 : bufs.keysBuf = nil
173 0 : }
174 : }
175 :
176 : // Assert that *DefragmentingIter implements the FragmentIterator interface.
177 : var _ FragmentIterator = (*DefragmentingIter)(nil)
178 :
179 : // Init initializes the defragmenting iter using the provided defragment
180 : // method.
181 : func (i *DefragmentingIter) Init(
182 : comparer *base.Comparer,
183 : iter FragmentIterator,
184 : equal DefragmentMethod,
185 : reducer DefragmentReducer,
186 : bufs *DefragmentingBuffers,
187 1 : ) {
188 1 : *i = DefragmentingIter{
189 1 : DefragmentingBuffers: bufs,
190 1 : comparer: comparer,
191 1 : equal: comparer.Equal,
192 1 : iter: iter,
193 1 : method: equal,
194 1 : reduce: reducer,
195 1 : }
196 1 : }
197 :
198 : // Error returns any accumulated error.
199 1 : func (i *DefragmentingIter) Error() error {
200 1 : return i.iter.Error()
201 1 : }
202 :
203 : // Close closes the underlying iterators.
204 1 : func (i *DefragmentingIter) Close() error {
205 1 : return i.iter.Close()
206 1 : }
207 :
208 : // SeekGE moves the iterator to the first span covering a key greater than or
209 : // equal to the given key. This is equivalent to seeking to the first span with
210 : // an end key greater than the given key.
211 1 : func (i *DefragmentingIter) SeekGE(key []byte) *Span {
212 1 : i.iterSpan = i.iter.SeekGE(key)
213 1 : if i.iterSpan == nil {
214 1 : i.iterPos = iterPosCurr
215 1 : return nil
216 1 : } else if i.iterSpan.Empty() {
217 1 : i.iterPos = iterPosCurr
218 1 : return i.iterSpan
219 1 : }
220 : // If the span starts strictly after key, we know there mustn't be an
221 : // earlier span that ends at i.iterSpan.Start, otherwise i.iter would've
222 : // returned that span instead.
223 1 : if i.comparer.Compare(i.iterSpan.Start, key) > 0 {
224 1 : return i.defragmentForward()
225 1 : }
226 :
227 : // The span we landed on has a Start bound ≤ key. There may be additional
228 : // fragments before this span. Defragment backward to find the start of the
229 : // defragmented span.
230 1 : i.defragmentBackward()
231 1 :
232 1 : // Defragmenting backward may have stopped because it encountered an error.
233 1 : // If so, we must not continue so that i.iter.Error() (and thus i.Error())
234 1 : // yields the error.
235 1 : if i.iterSpan == nil && i.iter.Error() != nil {
236 0 : return nil
237 0 : }
238 :
239 1 : if i.iterPos == iterPosPrev {
240 1 : // Next once back onto the span.
241 1 : i.iterSpan = i.iter.Next()
242 1 : }
243 : // Defragment the full span from its start.
244 1 : return i.defragmentForward()
245 : }
246 :
247 : // SeekLT moves the iterator to the last span covering a key less than the
248 : // given key. This is equivalent to seeking to the last span with a start
249 : // key less than the given key.
250 1 : func (i *DefragmentingIter) SeekLT(key []byte) *Span {
251 1 : i.iterSpan = i.iter.SeekLT(key)
252 1 : if i.iterSpan == nil {
253 1 : i.iterPos = iterPosCurr
254 1 : return nil
255 1 : } else if i.iterSpan.Empty() {
256 1 : i.iterPos = iterPosCurr
257 1 : return i.iterSpan
258 1 : }
259 : // If the span ends strictly before key, we know there mustn't be a later
260 : // span that starts at i.iterSpan.End, otherwise i.iter would've returned
261 : // that span instead.
262 1 : if i.comparer.Compare(i.iterSpan.End, key) < 0 {
263 1 : return i.defragmentBackward()
264 1 : }
265 :
266 : // The span we landed on has a End bound ≥ key. There may be additional
267 : // fragments after this span. Defragment forward to find the end of the
268 : // defragmented span.
269 1 : i.defragmentForward()
270 1 :
271 1 : // Defragmenting forward may have stopped because it encountered an error.
272 1 : // If so, we must not continue so that i.iter.Error() (and thus i.Error())
273 1 : // yields the error.
274 1 : if i.iterSpan == nil && i.iter.Error() != nil {
275 0 : return nil
276 0 : }
277 :
278 1 : if i.iterPos == iterPosNext {
279 1 : // Prev once back onto the span.
280 1 : i.iterSpan = i.iter.Prev()
281 1 : }
282 : // Defragment the full span from its end.
283 1 : return i.defragmentBackward()
284 : }
285 :
286 : // First seeks the iterator to the first span and returns it.
287 1 : func (i *DefragmentingIter) First() *Span {
288 1 : i.iterSpan = i.iter.First()
289 1 : if i.iterSpan == nil {
290 1 : i.iterPos = iterPosCurr
291 1 : return nil
292 1 : }
293 1 : return i.defragmentForward()
294 : }
295 :
296 : // Last seeks the iterator to the last span and returns it.
297 1 : func (i *DefragmentingIter) Last() *Span {
298 1 : i.iterSpan = i.iter.Last()
299 1 : if i.iterSpan == nil {
300 1 : i.iterPos = iterPosCurr
301 1 : return nil
302 1 : }
303 1 : return i.defragmentBackward()
304 : }
305 :
306 : // Next advances to the next span and returns it.
307 1 : func (i *DefragmentingIter) Next() *Span {
308 1 : switch i.iterPos {
309 1 : case iterPosPrev:
310 1 : // Switching directions; The iterator is currently positioned over the
311 1 : // last span of the previous set of fragments. In the below diagram,
312 1 : // the iterator is positioned over the last span that contributes to
313 1 : // the defragmented x position. We want to be positioned over the first
314 1 : // span that contributes to the z position.
315 1 : //
316 1 : // x x x y y y z z z
317 1 : // ^ ^
318 1 : // old new
319 1 : //
320 1 : // Next once to move onto y, defragment forward to land on the first z
321 1 : // position.
322 1 : i.iterSpan = i.iter.Next()
323 1 : if invariants.Enabled && i.iterSpan == nil && i.iter.Error() == nil {
324 0 : panic("pebble: invariant violation: no next span while switching directions")
325 : }
326 : // We're now positioned on the first span that was defragmented into the
327 : // current iterator position. Skip over the rest of the current iterator
328 : // position's constitutent fragments. In the above example, this would
329 : // land on the first 'z'.
330 1 : i.defragmentForward()
331 1 : if i.iterSpan == nil {
332 1 : i.iterPos = iterPosCurr
333 1 : return nil
334 1 : }
335 :
336 : // Now that we're positioned over the first of the next set of
337 : // fragments, defragment forward.
338 1 : return i.defragmentForward()
339 1 : case iterPosCurr:
340 1 : // iterPosCurr is only used when the iter is exhausted or when the iterator
341 1 : // is at an empty span.
342 1 : if invariants.Enabled && i.iterSpan != nil && !i.iterSpan.Empty() {
343 0 : panic("pebble: invariant violation: iterPosCurr with valid iterSpan")
344 : }
345 :
346 1 : i.iterSpan = i.iter.Next()
347 1 : if i.iterSpan == nil {
348 1 : return nil
349 1 : }
350 1 : return i.defragmentForward()
351 1 : case iterPosNext:
352 1 : // Already at the next span.
353 1 : if i.iterSpan == nil {
354 1 : i.iterPos = iterPosCurr
355 1 : return nil
356 1 : }
357 1 : return i.defragmentForward()
358 0 : default:
359 0 : panic("unreachable")
360 : }
361 : }
362 :
363 : // Prev steps back to the previous span and returns it.
364 1 : func (i *DefragmentingIter) Prev() *Span {
365 1 : switch i.iterPos {
366 1 : case iterPosPrev:
367 1 : // Already at the previous span.
368 1 : if i.iterSpan == nil {
369 1 : i.iterPos = iterPosCurr
370 1 : return nil
371 1 : }
372 1 : return i.defragmentBackward()
373 1 : case iterPosCurr:
374 1 : // iterPosCurr is only used when the iter is exhausted or when the iterator
375 1 : // is at an empty span.
376 1 : if invariants.Enabled && i.iterSpan != nil && !i.iterSpan.Empty() {
377 0 : panic("pebble: invariant violation: iterPosCurr with valid iterSpan")
378 : }
379 :
380 1 : i.iterSpan = i.iter.Prev()
381 1 : if i.iterSpan == nil {
382 1 : return nil
383 1 : }
384 1 : return i.defragmentBackward()
385 1 : case iterPosNext:
386 1 : // Switching directions; The iterator is currently positioned over the
387 1 : // first fragment of the next set of fragments. In the below diagram,
388 1 : // the iterator is positioned over the first span that contributes to
389 1 : // the defragmented z position. We want to be positioned over the last
390 1 : // span that contributes to the x position.
391 1 : //
392 1 : // x x x y y y z z z
393 1 : // ^ ^
394 1 : // new old
395 1 : //
396 1 : // Prev once to move onto y, defragment backward to land on the last x
397 1 : // position.
398 1 : i.iterSpan = i.iter.Prev()
399 1 : if invariants.Enabled && i.iterSpan == nil && i.iter.Error() == nil {
400 0 : panic("pebble: invariant violation: no previous span while switching directions")
401 : }
402 : // We're now positioned on the last span that was defragmented into the
403 : // current iterator position. Skip over the rest of the current iterator
404 : // position's constitutent fragments. In the above example, this would
405 : // land on the last 'x'.
406 1 : i.defragmentBackward()
407 1 :
408 1 : // Now that we're positioned over the last of the prev set of
409 1 : // fragments, defragment backward.
410 1 : if i.iterSpan == nil {
411 1 : i.iterPos = iterPosCurr
412 1 : return nil
413 1 : }
414 1 : return i.defragmentBackward()
415 0 : default:
416 0 : panic("unreachable")
417 : }
418 : }
419 :
420 : // checkEqual checks the two spans for logical equivalence. It uses the passed-in
421 : // DefragmentMethod and ensures both spans are NOT empty; not defragmenting empty
422 : // spans is an optimization that lets us load fewer sstable blocks.
423 1 : func (i *DefragmentingIter) checkEqual(left, right *Span) bool {
424 1 : return (!left.Empty() && !right.Empty()) && i.method.ShouldDefragment(i.equal, i.iterSpan, &i.curr)
425 1 : }
426 :
427 : // defragmentForward defragments spans in the forward direction, starting from
428 : // i.iter's current position. The span at the current position must be non-nil,
429 : // but may be Empty().
430 1 : func (i *DefragmentingIter) defragmentForward() *Span {
431 1 : if i.iterSpan.Empty() {
432 1 : // An empty span will never be equal to another span; see checkEqual for
433 1 : // why. To avoid loading non-empty range keys further ahead by calling Next,
434 1 : // return early.
435 1 : i.iterPos = iterPosCurr
436 1 : return i.iterSpan
437 1 : }
438 1 : i.saveCurrent()
439 1 :
440 1 : i.iterPos = iterPosNext
441 1 : i.iterSpan = i.iter.Next()
442 1 : for i.iterSpan != nil {
443 1 : if !i.equal(i.curr.End, i.iterSpan.Start) {
444 1 : // Not a continuation.
445 1 : break
446 : }
447 1 : if !i.checkEqual(i.iterSpan, &i.curr) {
448 1 : // Not a continuation.
449 1 : break
450 : }
451 1 : i.keyBuf = append(i.keyBuf[:0], i.iterSpan.End...)
452 1 : i.curr.End = i.keyBuf
453 1 : i.keysBuf = i.reduce(i.keysBuf, i.iterSpan.Keys)
454 1 : i.iterSpan = i.iter.Next()
455 : }
456 : // i.iterSpan == nil
457 : //
458 : // The inner iterator may return nil when it encounters an error. If there
459 : // was an error, we don't know whether there is another span we should
460 : // defragment or not. Return nil so that the caller knows they should check
461 : // Error().
462 1 : if i.iter.Error() != nil {
463 0 : return nil
464 0 : }
465 1 : i.curr.Keys = i.keysBuf
466 1 : return &i.curr
467 : }
468 :
469 : // defragmentBackward defragments spans in the backward direction, starting from
470 : // i.iter's current position. The span at the current position must be non-nil,
471 : // but may be Empty().
472 1 : func (i *DefragmentingIter) defragmentBackward() *Span {
473 1 : if i.iterSpan.Empty() {
474 1 : // An empty span will never be equal to another span; see checkEqual for
475 1 : // why. To avoid loading non-empty range keys further ahead by calling Next,
476 1 : // return early.
477 1 : i.iterPos = iterPosCurr
478 1 : return i.iterSpan
479 1 : }
480 1 : i.saveCurrent()
481 1 :
482 1 : i.iterPos = iterPosPrev
483 1 : i.iterSpan = i.iter.Prev()
484 1 : for i.iterSpan != nil {
485 1 : if !i.equal(i.curr.Start, i.iterSpan.End) {
486 1 : // Not a continuation.
487 1 : break
488 : }
489 1 : if !i.checkEqual(i.iterSpan, &i.curr) {
490 1 : // Not a continuation.
491 1 : break
492 : }
493 1 : i.keyBuf = append(i.keyBuf[:0], i.iterSpan.Start...)
494 1 : i.curr.Start = i.keyBuf
495 1 : i.keysBuf = i.reduce(i.keysBuf, i.iterSpan.Keys)
496 1 : i.iterSpan = i.iter.Prev()
497 : }
498 : // i.iterSpan == nil
499 : //
500 : // The inner iterator may return nil when it encounters an error. If there
501 : // was an error, we don't know whether there is another span we should
502 : // defragment or not. Return nil so that the caller knows they should check
503 : // Error().
504 1 : if i.iter.Error() != nil {
505 0 : return nil
506 0 : }
507 1 : i.curr.Keys = i.keysBuf
508 1 : return &i.curr
509 : }
510 :
511 1 : func (i *DefragmentingIter) saveCurrent() {
512 1 : i.currBuf.Reset()
513 1 : i.keysBuf = i.keysBuf[:0]
514 1 : i.keyBuf = i.keyBuf[:0]
515 1 : if i.iterSpan == nil {
516 0 : return
517 0 : }
518 1 : i.curr = Span{
519 1 : Start: i.saveBytes(i.iterSpan.Start),
520 1 : End: i.saveBytes(i.iterSpan.End),
521 1 : KeysOrder: i.iterSpan.KeysOrder,
522 1 : }
523 1 : for j := range i.iterSpan.Keys {
524 1 : i.keysBuf = append(i.keysBuf, Key{
525 1 : Trailer: i.iterSpan.Keys[j].Trailer,
526 1 : Suffix: i.saveBytes(i.iterSpan.Keys[j].Suffix),
527 1 : Value: i.saveBytes(i.iterSpan.Keys[j].Value),
528 1 : })
529 1 : }
530 1 : i.curr.Keys = i.keysBuf
531 : }
532 :
533 1 : func (i *DefragmentingIter) saveBytes(b []byte) []byte {
534 1 : if b == nil {
535 1 : return nil
536 1 : }
537 1 : i.currBuf, b = i.currBuf.Copy(b)
538 1 : return b
539 : }
540 :
541 : // WrapChildren implements FragmentIterator.
542 0 : func (i *DefragmentingIter) WrapChildren(wrap WrapFn) {
543 0 : i.iter = wrap(i.iter)
544 0 : }
|
