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 sstable
6 :
7 : import (
8 : "context"
9 : "encoding/binary"
10 : "unsafe"
11 :
12 : "github.com/cockroachdb/errors"
13 : "github.com/cockroachdb/pebble/internal/base"
14 : "github.com/cockroachdb/pebble/internal/invariants"
15 : "github.com/cockroachdb/pebble/internal/keyspan"
16 : "github.com/cockroachdb/pebble/internal/manual"
17 : "github.com/cockroachdb/pebble/internal/rangedel"
18 : "github.com/cockroachdb/pebble/internal/rangekey"
19 : )
20 :
21 1 : func uvarintLen(v uint32) int {
22 1 : i := 0
23 1 : for v >= 0x80 {
24 0 : v >>= 7
25 0 : i++
26 0 : }
27 1 : return i + 1
28 : }
29 :
30 : type blockWriter struct {
31 : restartInterval int
32 : nEntries int
33 : nextRestart int
34 : buf []byte
35 : // For datablocks in TableFormatPebblev3, we steal the most significant bit
36 : // in restarts for encoding setHasSameKeyPrefixSinceLastRestart. This leaves
37 : // us with 31 bits, which is more than enough (no one needs > 2GB blocks).
38 : // Typically, restarts occur every 16 keys, and by storing this bit with the
39 : // restart, we can optimize for the case where a user wants to skip to the
40 : // next prefix which happens to be in the same data block, but is > 16 keys
41 : // away. We have seen production situations with 100+ versions per MVCC key
42 : // (which share the same prefix). Additionally, for such writers, the prefix
43 : // compression of the key, that shares the key with the preceding key, is
44 : // limited to the prefix part of the preceding key -- this ensures that when
45 : // doing NPrefix (see blockIter) we don't need to assemble the full key
46 : // for each step since by limiting the length of the shared key we are
47 : // ensuring that any of the keys with the same prefix can be used to
48 : // assemble the full key when the prefix does change.
49 : restarts []uint32
50 : // Do not read curKey directly from outside blockWriter since it can have
51 : // the InternalKeyKindSSTableInternalObsoleteBit set. Use getCurKey() or
52 : // getCurUserKey() instead.
53 : curKey []byte
54 : // curValue excludes the optional prefix provided to
55 : // storeWithOptionalValuePrefix.
56 : curValue []byte
57 : prevKey []byte
58 : tmp [4]byte
59 : // We don't know the state of the sets that were at the end of the previous
60 : // block, so this is initially 0. It may be true for the second and later
61 : // restarts in a block. Not having inter-block information is fine since we
62 : // will optimize by stepping through restarts only within the same block.
63 : // Note that the first restart is the first key in the block.
64 : setHasSameKeyPrefixSinceLastRestart bool
65 : }
66 :
67 1 : func (w *blockWriter) clear() {
68 1 : *w = blockWriter{
69 1 : buf: w.buf[:0],
70 1 : restarts: w.restarts[:0],
71 1 : curKey: w.curKey[:0],
72 1 : curValue: w.curValue[:0],
73 1 : prevKey: w.prevKey[:0],
74 1 : }
75 1 : }
76 :
77 : // MaximumBlockSize is an extremely generous maximum block size of 256MiB. We
78 : // explicitly place this limit to reserve a few bits in the restart for
79 : // internal use.
80 : const MaximumBlockSize = 1 << 28
81 : const setHasSameKeyPrefixRestartMask uint32 = 1 << 31
82 : const restartMaskLittleEndianHighByteWithoutSetHasSamePrefix byte = 0b0111_1111
83 : const restartMaskLittleEndianHighByteOnlySetHasSamePrefix byte = 0b1000_0000
84 :
85 1 : func (w *blockWriter) getCurKey() InternalKey {
86 1 : k := base.DecodeInternalKey(w.curKey)
87 1 : k.Trailer = k.Trailer & trailerObsoleteMask
88 1 : return k
89 1 : }
90 :
91 1 : func (w *blockWriter) getCurUserKey() []byte {
92 1 : n := len(w.curKey) - base.InternalTrailerLen
93 1 : if n < 0 {
94 0 : panic(errors.AssertionFailedf("corrupt key in blockWriter buffer"))
95 : }
96 1 : return w.curKey[:n:n]
97 : }
98 :
99 : // If !addValuePrefix, the valuePrefix is ignored.
100 : func (w *blockWriter) storeWithOptionalValuePrefix(
101 : keySize int,
102 : value []byte,
103 : maxSharedKeyLen int,
104 : addValuePrefix bool,
105 : valuePrefix valuePrefix,
106 : setHasSameKeyPrefix bool,
107 1 : ) {
108 1 : shared := 0
109 1 : if !setHasSameKeyPrefix {
110 1 : w.setHasSameKeyPrefixSinceLastRestart = false
111 1 : }
112 1 : if w.nEntries == w.nextRestart {
113 1 : w.nextRestart = w.nEntries + w.restartInterval
114 1 : restart := uint32(len(w.buf))
115 1 : if w.setHasSameKeyPrefixSinceLastRestart {
116 1 : restart = restart | setHasSameKeyPrefixRestartMask
117 1 : }
118 1 : w.setHasSameKeyPrefixSinceLastRestart = true
119 1 : w.restarts = append(w.restarts, restart)
120 1 : } else {
121 1 : // TODO(peter): Manually inlined version of base.SharedPrefixLen(). This
122 1 : // is 3% faster on BenchmarkWriter on go1.16. Remove if future versions
123 1 : // show this to not be a performance win. For now, functions that use of
124 1 : // unsafe cannot be inlined.
125 1 : n := maxSharedKeyLen
126 1 : if n > len(w.prevKey) {
127 1 : n = len(w.prevKey)
128 1 : }
129 1 : asUint64 := func(b []byte, i int) uint64 {
130 1 : return binary.LittleEndian.Uint64(b[i:])
131 1 : }
132 1 : for shared < n-7 && asUint64(w.curKey, shared) == asUint64(w.prevKey, shared) {
133 1 : shared += 8
134 1 : }
135 1 : for shared < n && w.curKey[shared] == w.prevKey[shared] {
136 1 : shared++
137 1 : }
138 : }
139 :
140 1 : lenValuePlusOptionalPrefix := len(value)
141 1 : if addValuePrefix {
142 1 : lenValuePlusOptionalPrefix++
143 1 : }
144 1 : needed := 3*binary.MaxVarintLen32 + len(w.curKey[shared:]) + lenValuePlusOptionalPrefix
145 1 : n := len(w.buf)
146 1 : if cap(w.buf) < n+needed {
147 1 : newCap := 2 * cap(w.buf)
148 1 : if newCap == 0 {
149 1 : newCap = 1024
150 1 : }
151 1 : for newCap < n+needed {
152 1 : newCap *= 2
153 1 : }
154 1 : newBuf := make([]byte, n, newCap)
155 1 : copy(newBuf, w.buf)
156 1 : w.buf = newBuf
157 : }
158 1 : w.buf = w.buf[:n+needed]
159 1 :
160 1 : // TODO(peter): Manually inlined versions of binary.PutUvarint(). This is 15%
161 1 : // faster on BenchmarkWriter on go1.13. Remove if go1.14 or future versions
162 1 : // show this to not be a performance win.
163 1 : {
164 1 : x := uint32(shared)
165 1 : for x >= 0x80 {
166 0 : w.buf[n] = byte(x) | 0x80
167 0 : x >>= 7
168 0 : n++
169 0 : }
170 1 : w.buf[n] = byte(x)
171 1 : n++
172 : }
173 :
174 1 : {
175 1 : x := uint32(keySize - shared)
176 1 : for x >= 0x80 {
177 0 : w.buf[n] = byte(x) | 0x80
178 0 : x >>= 7
179 0 : n++
180 0 : }
181 1 : w.buf[n] = byte(x)
182 1 : n++
183 : }
184 :
185 1 : {
186 1 : x := uint32(lenValuePlusOptionalPrefix)
187 1 : for x >= 0x80 {
188 1 : w.buf[n] = byte(x) | 0x80
189 1 : x >>= 7
190 1 : n++
191 1 : }
192 1 : w.buf[n] = byte(x)
193 1 : n++
194 : }
195 :
196 1 : n += copy(w.buf[n:], w.curKey[shared:])
197 1 : if addValuePrefix {
198 1 : w.buf[n : n+1][0] = byte(valuePrefix)
199 1 : n++
200 1 : }
201 1 : n += copy(w.buf[n:], value)
202 1 : w.buf = w.buf[:n]
203 1 :
204 1 : w.curValue = w.buf[n-len(value):]
205 1 :
206 1 : w.nEntries++
207 : }
208 :
209 1 : func (w *blockWriter) add(key InternalKey, value []byte) {
210 1 : w.addWithOptionalValuePrefix(
211 1 : key, false, value, len(key.UserKey), false, 0, false)
212 1 : }
213 :
214 : // Callers that always set addValuePrefix to false should use add() instead.
215 : //
216 : // isObsolete indicates whether this key-value pair is obsolete in this
217 : // sstable (only applicable when writing data blocks) -- see the comment in
218 : // table.go and the longer one in format.go. addValuePrefix adds a 1 byte
219 : // prefix to the value, specified in valuePrefix -- this is used for data
220 : // blocks in TableFormatPebblev3 onwards for SETs (see the comment in
221 : // format.go, with more details in value_block.go). setHasSameKeyPrefix is
222 : // also used in TableFormatPebblev3 onwards for SETs.
223 : func (w *blockWriter) addWithOptionalValuePrefix(
224 : key InternalKey,
225 : isObsolete bool,
226 : value []byte,
227 : maxSharedKeyLen int,
228 : addValuePrefix bool,
229 : valuePrefix valuePrefix,
230 : setHasSameKeyPrefix bool,
231 1 : ) {
232 1 : w.curKey, w.prevKey = w.prevKey, w.curKey
233 1 :
234 1 : size := key.Size()
235 1 : if cap(w.curKey) < size {
236 1 : w.curKey = make([]byte, 0, size*2)
237 1 : }
238 1 : w.curKey = w.curKey[:size]
239 1 : if isObsolete {
240 1 : key.Trailer = key.Trailer | trailerObsoleteBit
241 1 : }
242 1 : key.Encode(w.curKey)
243 1 :
244 1 : w.storeWithOptionalValuePrefix(
245 1 : size, value, maxSharedKeyLen, addValuePrefix, valuePrefix, setHasSameKeyPrefix)
246 : }
247 :
248 1 : func (w *blockWriter) finish() []byte {
249 1 : // Write the restart points to the buffer.
250 1 : if w.nEntries == 0 {
251 1 : // Every block must have at least one restart point.
252 1 : if cap(w.restarts) > 0 {
253 1 : w.restarts = w.restarts[:1]
254 1 : w.restarts[0] = 0
255 1 : } else {
256 1 : w.restarts = append(w.restarts, 0)
257 1 : }
258 : }
259 1 : tmp4 := w.tmp[:4]
260 1 : for _, x := range w.restarts {
261 1 : binary.LittleEndian.PutUint32(tmp4, x)
262 1 : w.buf = append(w.buf, tmp4...)
263 1 : }
264 1 : binary.LittleEndian.PutUint32(tmp4, uint32(len(w.restarts)))
265 1 : w.buf = append(w.buf, tmp4...)
266 1 : result := w.buf
267 1 :
268 1 : // Reset the block state.
269 1 : w.nEntries = 0
270 1 : w.nextRestart = 0
271 1 : w.buf = w.buf[:0]
272 1 : w.restarts = w.restarts[:0]
273 1 : return result
274 : }
275 :
276 : // emptyBlockSize holds the size of an empty block. Every block ends
277 : // in a uint32 trailer encoding the number of restart points within the
278 : // block.
279 : const emptyBlockSize = 4
280 :
281 1 : func (w *blockWriter) estimatedSize() int {
282 1 : return len(w.buf) + 4*len(w.restarts) + emptyBlockSize
283 1 : }
284 :
285 : type blockEntry struct {
286 : offset int32
287 : keyStart int32
288 : keyEnd int32
289 : valStart int32
290 : valSize int32
291 : }
292 :
293 : // blockIter is an iterator over a single block of data.
294 : //
295 : // A blockIter provides an additional guarantee around key stability when a
296 : // block has a restart interval of 1 (i.e. when there is no prefix
297 : // compression). Key stability refers to whether the InternalKey.UserKey bytes
298 : // returned by a positioning call will remain stable after a subsequent
299 : // positioning call. The normal case is that a positioning call will invalidate
300 : // any previously returned InternalKey.UserKey. If a block has a restart
301 : // interval of 1 (no prefix compression), blockIter guarantees that
302 : // InternalKey.UserKey will point to the key as stored in the block itself
303 : // which will remain valid until the blockIter is closed. The key stability
304 : // guarantee is used by the range tombstone and range key code, which knows that
305 : // the respective blocks are always encoded with a restart interval of 1. This
306 : // per-block key stability guarantee is sufficient for range tombstones and
307 : // range deletes as they are always encoded in a single block.
308 : //
309 : // A blockIter also provides a value stability guarantee for range deletions and
310 : // range keys since there is only a single range deletion and range key block
311 : // per sstable and the blockIter will not release the bytes for the block until
312 : // it is closed.
313 : //
314 : // Note on why blockIter knows about lazyValueHandling:
315 : //
316 : // blockIter's positioning functions (that return a LazyValue), are too
317 : // complex to inline even prior to lazyValueHandling. blockIter.Next and
318 : // blockIter.First were by far the cheapest and had costs 195 and 180
319 : // respectively, which exceeds the budget of 80. We initially tried to keep
320 : // the lazyValueHandling logic out of blockIter by wrapping it with a
321 : // lazyValueDataBlockIter. singleLevelIter and twoLevelIter would use this
322 : // wrapped iter. The functions in lazyValueDataBlockIter were simple, in that
323 : // they called the corresponding blockIter func and then decided whether the
324 : // value was in fact in-place (so return immediately) or needed further
325 : // handling. But these also turned out too costly for mid-stack inlining since
326 : // simple calls like the following have a high cost that is barely under the
327 : // budget of 80
328 : //
329 : // k, v := i.data.SeekGE(key, flags) // cost 74
330 : // k, v := i.data.Next() // cost 72
331 : //
332 : // We have 2 options for minimizing performance regressions:
333 : // - Include the lazyValueHandling logic in the already non-inlineable
334 : // blockIter functions: Since most of the time is spent in data block iters,
335 : // it is acceptable to take the small hit of unnecessary branching (which
336 : // hopefully branch prediction will predict correctly) for other kinds of
337 : // blocks.
338 : // - Duplicate the logic of singleLevelIterator and twoLevelIterator for the
339 : // v3 sstable and only use the aforementioned lazyValueDataBlockIter for a
340 : // v3 sstable. We would want to manage these copies via code generation.
341 : //
342 : // We have picked the first option here.
343 : type blockIter struct {
344 : cmp Compare
345 : // offset is the byte index that marks where the current key/value is
346 : // encoded in the block.
347 : offset int32
348 : // nextOffset is the byte index where the next key/value is encoded in the
349 : // block.
350 : nextOffset int32
351 : // A "restart point" in a block is a point where the full key is encoded,
352 : // instead of just having a suffix of the key encoded. See readEntry() for
353 : // how prefix compression of keys works. Keys in between two restart points
354 : // only have a suffix encoded in the block. When restart interval is 1, no
355 : // prefix compression of keys happens. This is the case with range tombstone
356 : // blocks.
357 : //
358 : // All restart offsets are listed in increasing order in
359 : // i.ptr[i.restarts:len(block)-4], while numRestarts is encoded in the last
360 : // 4 bytes of the block as a uint32 (i.ptr[len(block)-4:]). i.restarts can
361 : // therefore be seen as the point where data in the block ends, and a list
362 : // of offsets of all restart points begins.
363 : restarts int32
364 : // Number of restart points in this block. Encoded at the end of the block
365 : // as a uint32.
366 : numRestarts int32
367 : globalSeqNum uint64
368 : ptr unsafe.Pointer
369 : data []byte
370 : // key contains the raw key the iterator is currently pointed at. This may
371 : // point directly to data stored in the block (for a key which has no prefix
372 : // compression), to fullKey (for a prefix compressed key), or to a slice of
373 : // data stored in cachedBuf (during reverse iteration).
374 : key []byte
375 : // fullKey is a buffer used for key prefix decompression.
376 : fullKey []byte
377 : // val contains the value the iterator is currently pointed at. If non-nil,
378 : // this points to a slice of the block data.
379 : val []byte
380 : // lazyValue is val turned into a LazyValue, whenever a positioning method
381 : // returns a non-nil key-value pair.
382 : lazyValue base.LazyValue
383 : // ikey contains the decoded InternalKey the iterator is currently pointed
384 : // at. Note that the memory backing ikey.UserKey is either data stored
385 : // directly in the block, fullKey, or cachedBuf. The key stability guarantee
386 : // for blocks built with a restart interval of 1 is achieved by having
387 : // ikey.UserKey always point to data stored directly in the block.
388 : ikey InternalKey
389 : // cached and cachedBuf are used during reverse iteration. They are needed
390 : // because we can't perform prefix decoding in reverse, only in the forward
391 : // direction. In order to iterate in reverse, we decode and cache the entries
392 : // between two restart points.
393 : //
394 : // Note that cached[len(cached)-1] contains the previous entry to the one the
395 : // blockIter is currently pointed at. As usual, nextOffset will contain the
396 : // offset of the next entry. During reverse iteration, nextOffset will be
397 : // updated to point to offset, and we'll set the blockIter to point at the
398 : // entry cached[len(cached)-1]. See Prev() for more details.
399 : //
400 : // For a block encoded with a restart interval of 1, cached and cachedBuf
401 : // will not be used as there are no prefix compressed entries between the
402 : // restart points.
403 : cached []blockEntry
404 : cachedBuf []byte
405 : handle bufferHandle
406 : // for block iteration for already loaded blocks.
407 : firstUserKey []byte
408 : lazyValueHandling struct {
409 : vbr *valueBlockReader
410 : hasValuePrefix bool
411 : }
412 : hideObsoletePoints bool
413 : }
414 :
415 : // blockIter implements the base.InternalIterator interface.
416 : var _ base.InternalIterator = (*blockIter)(nil)
417 :
418 1 : func newBlockIter(cmp Compare, block block) (*blockIter, error) {
419 1 : i := &blockIter{}
420 1 : return i, i.init(cmp, block, 0, false)
421 1 : }
422 :
423 0 : func (i *blockIter) String() string {
424 0 : return "block"
425 0 : }
426 :
427 : func (i *blockIter) init(
428 : cmp Compare, block block, globalSeqNum uint64, hideObsoletePoints bool,
429 1 : ) error {
430 1 : numRestarts := int32(binary.LittleEndian.Uint32(block[len(block)-4:]))
431 1 : if numRestarts == 0 {
432 0 : return base.CorruptionErrorf("pebble/table: invalid table (block has no restart points)")
433 0 : }
434 1 : i.cmp = cmp
435 1 : i.restarts = int32(len(block)) - 4*(1+numRestarts)
436 1 : i.numRestarts = numRestarts
437 1 : i.globalSeqNum = globalSeqNum
438 1 : i.ptr = unsafe.Pointer(&block[0])
439 1 : i.data = block
440 1 : i.fullKey = i.fullKey[:0]
441 1 : i.val = nil
442 1 : i.hideObsoletePoints = hideObsoletePoints
443 1 : i.clearCache()
444 1 : if i.restarts > 0 {
445 1 : if err := i.readFirstKey(); err != nil {
446 0 : return err
447 0 : }
448 1 : } else {
449 1 : // Block is empty.
450 1 : i.firstUserKey = nil
451 1 : }
452 1 : return nil
453 : }
454 :
455 : // NB: two cases of hideObsoletePoints:
456 : // - Local sstable iteration: globalSeqNum will be set iff the sstable was
457 : // ingested.
458 : // - Foreign sstable iteration: globalSeqNum is always set.
459 : func (i *blockIter) initHandle(
460 : cmp Compare, block bufferHandle, globalSeqNum uint64, hideObsoletePoints bool,
461 1 : ) error {
462 1 : i.handle.Release()
463 1 : i.handle = block
464 1 : return i.init(cmp, block.Get(), globalSeqNum, hideObsoletePoints)
465 1 : }
466 :
467 1 : func (i *blockIter) invalidate() {
468 1 : i.clearCache()
469 1 : i.offset = 0
470 1 : i.nextOffset = 0
471 1 : i.restarts = 0
472 1 : i.numRestarts = 0
473 1 : i.data = nil
474 1 : }
475 :
476 : // isDataInvalidated returns true when the blockIter has been invalidated
477 : // using an invalidate call. NB: this is different from blockIter.Valid
478 : // which is part of the InternalIterator implementation.
479 1 : func (i *blockIter) isDataInvalidated() bool {
480 1 : return i.data == nil
481 1 : }
482 :
483 1 : func (i *blockIter) resetForReuse() blockIter {
484 1 : return blockIter{
485 1 : fullKey: i.fullKey[:0],
486 1 : cached: i.cached[:0],
487 1 : cachedBuf: i.cachedBuf[:0],
488 1 : data: nil,
489 1 : }
490 1 : }
491 :
492 1 : func (i *blockIter) readEntry() {
493 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
494 1 :
495 1 : // This is an ugly performance hack. Reading entries from blocks is one of
496 1 : // the inner-most routines and decoding the 3 varints per-entry takes
497 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
498 1 : // us, so we do it manually. This provides a 10-15% performance improvement
499 1 : // on blockIter benchmarks on both go1.11 and go1.12.
500 1 : //
501 1 : // TODO(peter): remove this hack if go:inline is ever supported.
502 1 :
503 1 : var shared uint32
504 1 : if a := *((*uint8)(ptr)); a < 128 {
505 1 : shared = uint32(a)
506 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
507 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
508 0 : shared = uint32(b)<<7 | uint32(a)
509 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
510 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
511 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
512 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
513 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
514 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
515 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
516 0 : } else {
517 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
518 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
519 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
520 0 : }
521 :
522 1 : var unshared uint32
523 1 : if a := *((*uint8)(ptr)); a < 128 {
524 1 : unshared = uint32(a)
525 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
526 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
527 0 : unshared = uint32(b)<<7 | uint32(a)
528 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
529 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
530 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
531 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
532 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
533 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
534 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
535 0 : } else {
536 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
537 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
538 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
539 0 : }
540 :
541 1 : var value uint32
542 1 : if a := *((*uint8)(ptr)); a < 128 {
543 1 : value = uint32(a)
544 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
545 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
546 1 : value = uint32(b)<<7 | uint32(a)
547 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
548 1 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
549 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
550 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
551 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
552 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
553 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
554 0 : } else {
555 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
556 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
557 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
558 0 : }
559 :
560 1 : unsharedKey := getBytes(ptr, int(unshared))
561 1 : // TODO(sumeer): move this into the else block below.
562 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
563 1 : if shared == 0 {
564 1 : // Provide stability for the key across positioning calls if the key
565 1 : // doesn't share a prefix with the previous key. This removes requiring the
566 1 : // key to be copied if the caller knows the block has a restart interval of
567 1 : // 1. An important example of this is range-del blocks.
568 1 : i.key = unsharedKey
569 1 : } else {
570 1 : i.key = i.fullKey
571 1 : }
572 1 : ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
573 1 : i.val = getBytes(ptr, int(value))
574 1 : i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
575 : }
576 :
577 1 : func (i *blockIter) readFirstKey() error {
578 1 : ptr := i.ptr
579 1 :
580 1 : // This is an ugly performance hack. Reading entries from blocks is one of
581 1 : // the inner-most routines and decoding the 3 varints per-entry takes
582 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
583 1 : // us, so we do it manually. This provides a 10-15% performance improvement
584 1 : // on blockIter benchmarks on both go1.11 and go1.12.
585 1 : //
586 1 : // TODO(peter): remove this hack if go:inline is ever supported.
587 1 :
588 1 : if shared := *((*uint8)(ptr)); shared == 0 {
589 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
590 1 : } else {
591 0 : // The shared length is != 0, which is invalid.
592 0 : panic("first key in block must have zero shared length")
593 : }
594 :
595 1 : var unshared uint32
596 1 : if a := *((*uint8)(ptr)); a < 128 {
597 1 : unshared = uint32(a)
598 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
599 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
600 0 : unshared = uint32(b)<<7 | uint32(a)
601 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
602 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
603 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
604 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
605 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
606 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
607 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
608 0 : } else {
609 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
610 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
611 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
612 0 : }
613 :
614 : // Skip the value length.
615 1 : if a := *((*uint8)(ptr)); a < 128 {
616 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
617 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); a < 128 {
618 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
619 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); a < 128 {
620 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
621 0 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); a < 128 {
622 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
623 0 : } else {
624 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
625 0 : }
626 :
627 1 : firstKey := getBytes(ptr, int(unshared))
628 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
629 1 : // BlockIter benchmarks.
630 1 : if n := len(firstKey) - 8; n >= 0 {
631 1 : i.firstUserKey = firstKey[:n:n]
632 1 : } else {
633 0 : i.firstUserKey = nil
634 0 : return base.CorruptionErrorf("pebble/table: invalid firstKey in block")
635 0 : }
636 1 : return nil
637 : }
638 :
639 : // The sstable internal obsolete bit is set when writing a block and unset by
640 : // blockIter, so no code outside block writing/reading code ever sees it.
641 : const trailerObsoleteBit = uint64(base.InternalKeyKindSSTableInternalObsoleteBit)
642 : const trailerObsoleteMask = (InternalKeySeqNumMax << 8) | uint64(base.InternalKeyKindSSTableInternalObsoleteMask)
643 :
644 1 : func (i *blockIter) decodeInternalKey(key []byte) (hiddenPoint bool) {
645 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
646 1 : // BlockIter benchmarks.
647 1 : if n := len(key) - 8; n >= 0 {
648 1 : trailer := binary.LittleEndian.Uint64(key[n:])
649 1 : hiddenPoint = i.hideObsoletePoints &&
650 1 : (trailer&trailerObsoleteBit != 0)
651 1 : i.ikey.Trailer = trailer & trailerObsoleteMask
652 1 : i.ikey.UserKey = key[:n:n]
653 1 : if i.globalSeqNum != 0 {
654 1 : i.ikey.SetSeqNum(i.globalSeqNum)
655 1 : }
656 1 : } else {
657 1 : i.ikey.Trailer = uint64(InternalKeyKindInvalid)
658 1 : i.ikey.UserKey = nil
659 1 : }
660 1 : return hiddenPoint
661 : }
662 :
663 1 : func (i *blockIter) clearCache() {
664 1 : i.cached = i.cached[:0]
665 1 : i.cachedBuf = i.cachedBuf[:0]
666 1 : }
667 :
668 1 : func (i *blockIter) cacheEntry() {
669 1 : var valStart int32
670 1 : valSize := int32(len(i.val))
671 1 : if valSize > 0 {
672 1 : valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
673 1 : }
674 :
675 1 : i.cached = append(i.cached, blockEntry{
676 1 : offset: i.offset,
677 1 : keyStart: int32(len(i.cachedBuf)),
678 1 : keyEnd: int32(len(i.cachedBuf) + len(i.key)),
679 1 : valStart: valStart,
680 1 : valSize: valSize,
681 1 : })
682 1 : i.cachedBuf = append(i.cachedBuf, i.key...)
683 : }
684 :
685 1 : func (i *blockIter) getFirstUserKey() []byte {
686 1 : return i.firstUserKey
687 1 : }
688 :
689 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
690 : // package.
691 1 : func (i *blockIter) SeekGE(key []byte, flags base.SeekGEFlags) (*InternalKey, base.LazyValue) {
692 1 : if invariants.Enabled && i.isDataInvalidated() {
693 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
694 : }
695 :
696 1 : i.clearCache()
697 1 : // Find the index of the smallest restart point whose key is > the key
698 1 : // sought; index will be numRestarts if there is no such restart point.
699 1 : i.offset = 0
700 1 : var index int32
701 1 :
702 1 : {
703 1 : // NB: manually inlined sort.Seach is ~5% faster.
704 1 : //
705 1 : // Define f(-1) == false and f(n) == true.
706 1 : // Invariant: f(index-1) == false, f(upper) == true.
707 1 : upper := i.numRestarts
708 1 : for index < upper {
709 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
710 1 : // index ≤ h < upper
711 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
712 1 : // For a restart point, there are 0 bytes shared with the previous key.
713 1 : // The varint encoding of 0 occupies 1 byte.
714 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
715 1 :
716 1 : // Decode the key at that restart point, and compare it to the key
717 1 : // sought. See the comment in readEntry for why we manually inline the
718 1 : // varint decoding.
719 1 : var v1 uint32
720 1 : if a := *((*uint8)(ptr)); a < 128 {
721 1 : v1 = uint32(a)
722 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
723 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
724 0 : v1 = uint32(b)<<7 | uint32(a)
725 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
726 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
727 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
728 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
729 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
730 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
731 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
732 0 : } else {
733 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
734 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
735 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
736 0 : }
737 :
738 1 : if *((*uint8)(ptr)) < 128 {
739 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
740 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
741 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
742 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
743 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
744 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
745 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
746 0 : } else {
747 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
748 0 : }
749 :
750 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
751 : // speedup on BlockIter benchmarks.
752 1 : s := getBytes(ptr, int(v1))
753 1 : var k []byte
754 1 : if n := len(s) - 8; n >= 0 {
755 1 : k = s[:n:n]
756 1 : }
757 : // Else k is invalid, and left as nil
758 :
759 1 : if i.cmp(key, k) > 0 {
760 1 : // The search key is greater than the user key at this restart point.
761 1 : // Search beyond this restart point, since we are trying to find the
762 1 : // first restart point with a user key >= the search key.
763 1 : index = h + 1 // preserves f(i-1) == false
764 1 : } else {
765 1 : // k >= search key, so prune everything after index (since index
766 1 : // satisfies the property we are looking for).
767 1 : upper = h // preserves f(j) == true
768 1 : }
769 : }
770 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
771 : // => answer is index.
772 : }
773 :
774 : // index is the first restart point with key >= search key. Define the keys
775 : // between a restart point and the next restart point as belonging to that
776 : // restart point.
777 : //
778 : // Since keys are strictly increasing, if index > 0 then the restart point
779 : // at index-1 will be the first one that has some keys belonging to it that
780 : // could be equal to the search key. If index == 0, then all keys in this
781 : // block are larger than the key sought, and offset remains at zero.
782 1 : if index > 0 {
783 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
784 1 : }
785 1 : i.readEntry()
786 1 : hiddenPoint := i.decodeInternalKey(i.key)
787 1 :
788 1 : // Iterate from that restart point to somewhere >= the key sought.
789 1 : if !i.valid() {
790 0 : return nil, base.LazyValue{}
791 0 : }
792 1 : if !hiddenPoint && i.cmp(i.ikey.UserKey, key) >= 0 {
793 1 : // Initialize i.lazyValue
794 1 : if !i.lazyValueHandling.hasValuePrefix ||
795 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
796 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
797 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
798 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
799 1 : } else {
800 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
801 1 : }
802 1 : return &i.ikey, i.lazyValue
803 : }
804 1 : for i.Next(); i.valid(); i.Next() {
805 1 : if i.cmp(i.ikey.UserKey, key) >= 0 {
806 1 : // i.Next() has already initialized i.lazyValue.
807 1 : return &i.ikey, i.lazyValue
808 1 : }
809 : }
810 1 : return nil, base.LazyValue{}
811 : }
812 :
813 : // SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
814 : // pebble package.
815 : func (i *blockIter) SeekPrefixGE(
816 : prefix, key []byte, flags base.SeekGEFlags,
817 0 : ) (*base.InternalKey, base.LazyValue) {
818 0 : // This should never be called as prefix iteration is handled by sstable.Iterator.
819 0 : panic("pebble: SeekPrefixGE unimplemented")
820 : }
821 :
822 : // SeekLT implements internalIterator.SeekLT, as documented in the pebble
823 : // package.
824 1 : func (i *blockIter) SeekLT(key []byte, flags base.SeekLTFlags) (*InternalKey, base.LazyValue) {
825 1 : if invariants.Enabled && i.isDataInvalidated() {
826 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
827 : }
828 :
829 1 : i.clearCache()
830 1 : // Find the index of the smallest restart point whose key is >= the key
831 1 : // sought; index will be numRestarts if there is no such restart point.
832 1 : i.offset = 0
833 1 : var index int32
834 1 :
835 1 : {
836 1 : // NB: manually inlined sort.Search is ~5% faster.
837 1 : //
838 1 : // Define f(-1) == false and f(n) == true.
839 1 : // Invariant: f(index-1) == false, f(upper) == true.
840 1 : upper := i.numRestarts
841 1 : for index < upper {
842 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
843 1 : // index ≤ h < upper
844 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
845 1 : // For a restart point, there are 0 bytes shared with the previous key.
846 1 : // The varint encoding of 0 occupies 1 byte.
847 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
848 1 :
849 1 : // Decode the key at that restart point, and compare it to the key
850 1 : // sought. See the comment in readEntry for why we manually inline the
851 1 : // varint decoding.
852 1 : var v1 uint32
853 1 : if a := *((*uint8)(ptr)); a < 128 {
854 1 : v1 = uint32(a)
855 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
856 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
857 0 : v1 = uint32(b)<<7 | uint32(a)
858 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
859 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
860 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
861 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
862 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
863 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
864 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
865 0 : } else {
866 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
867 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
868 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
869 0 : }
870 :
871 1 : if *((*uint8)(ptr)) < 128 {
872 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
873 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
874 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
875 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
876 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
877 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
878 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
879 0 : } else {
880 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
881 0 : }
882 :
883 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
884 : // speedup on BlockIter benchmarks.
885 1 : s := getBytes(ptr, int(v1))
886 1 : var k []byte
887 1 : if n := len(s) - 8; n >= 0 {
888 1 : k = s[:n:n]
889 1 : }
890 : // Else k is invalid, and left as nil
891 :
892 1 : if i.cmp(key, k) > 0 {
893 1 : // The search key is greater than the user key at this restart point.
894 1 : // Search beyond this restart point, since we are trying to find the
895 1 : // first restart point with a user key >= the search key.
896 1 : index = h + 1 // preserves f(i-1) == false
897 1 : } else {
898 1 : // k >= search key, so prune everything after index (since index
899 1 : // satisfies the property we are looking for).
900 1 : upper = h // preserves f(j) == true
901 1 : }
902 : }
903 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
904 : // => answer is index.
905 : }
906 :
907 : // index is the first restart point with key >= search key. Define the keys
908 : // between a restart point and the next restart point as belonging to that
909 : // restart point. Note that index could be equal to i.numRestarts, i.e., we
910 : // are past the last restart.
911 : //
912 : // Since keys are strictly increasing, if index > 0 then the restart point
913 : // at index-1 will be the first one that has some keys belonging to it that
914 : // are less than the search key. If index == 0, then all keys in this block
915 : // are larger than the search key, so there is no match.
916 1 : targetOffset := i.restarts
917 1 : if index > 0 {
918 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
919 1 : if index < i.numRestarts {
920 1 : targetOffset = decodeRestart(i.data[i.restarts+4*(index):])
921 1 : }
922 1 : } else if index == 0 {
923 1 : // If index == 0 then all keys in this block are larger than the key
924 1 : // sought.
925 1 : i.offset = -1
926 1 : i.nextOffset = 0
927 1 : return nil, base.LazyValue{}
928 1 : }
929 :
930 : // Iterate from that restart point to somewhere >= the key sought, then back
931 : // up to the previous entry. The expectation is that we'll be performing
932 : // reverse iteration, so we cache the entries as we advance forward.
933 1 : i.nextOffset = i.offset
934 1 :
935 1 : for {
936 1 : i.offset = i.nextOffset
937 1 : i.readEntry()
938 1 : // When hidden keys are common, there is additional optimization possible
939 1 : // by not caching entries that are hidden (note that some calls to
940 1 : // cacheEntry don't decode the internal key before caching, but checking
941 1 : // whether a key is hidden does not require full decoding). However, we do
942 1 : // need to use the blockEntry.offset in the cache for the first entry at
943 1 : // the reset point to do the binary search when the cache is empty -- so
944 1 : // we would need to cache that first entry (though not the key) even if
945 1 : // was hidden. Our current assumption is that if there are large numbers
946 1 : // of hidden keys we will be able to skip whole blocks (using block
947 1 : // property filters) so we don't bother optimizing.
948 1 : hiddenPoint := i.decodeInternalKey(i.key)
949 1 :
950 1 : // NB: we don't use the hiddenPoint return value of decodeInternalKey
951 1 : // since we want to stop as soon as we reach a key >= ikey.UserKey, so
952 1 : // that we can reverse.
953 1 : if i.cmp(i.ikey.UserKey, key) >= 0 {
954 1 : // The current key is greater than or equal to our search key. Back up to
955 1 : // the previous key which was less than our search key. Note that this for
956 1 : // loop will execute at least once with this if-block not being true, so
957 1 : // the key we are backing up to is the last one this loop cached.
958 1 : return i.Prev()
959 1 : }
960 :
961 1 : if i.nextOffset >= targetOffset {
962 1 : // We've reached the end of the current restart block. Return the
963 1 : // current key if not hidden, else call Prev().
964 1 : //
965 1 : // When the restart interval is 1, the first iteration of the for loop
966 1 : // will bring us here. In that case ikey is backed by the block so we
967 1 : // get the desired key stability guarantee for the lifetime of the
968 1 : // blockIter. That is, we never cache anything and therefore never
969 1 : // return a key backed by cachedBuf.
970 1 : if hiddenPoint {
971 1 : return i.Prev()
972 1 : }
973 1 : break
974 : }
975 :
976 1 : i.cacheEntry()
977 : }
978 :
979 1 : if !i.valid() {
980 1 : return nil, base.LazyValue{}
981 1 : }
982 1 : if !i.lazyValueHandling.hasValuePrefix ||
983 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
984 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
985 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
986 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
987 1 : } else {
988 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
989 1 : }
990 1 : return &i.ikey, i.lazyValue
991 : }
992 :
993 : // First implements internalIterator.First, as documented in the pebble
994 : // package.
995 1 : func (i *blockIter) First() (*InternalKey, base.LazyValue) {
996 1 : if invariants.Enabled && i.isDataInvalidated() {
997 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
998 : }
999 :
1000 1 : i.offset = 0
1001 1 : if !i.valid() {
1002 1 : return nil, base.LazyValue{}
1003 1 : }
1004 1 : i.clearCache()
1005 1 : i.readEntry()
1006 1 : hiddenPoint := i.decodeInternalKey(i.key)
1007 1 : if hiddenPoint {
1008 1 : return i.Next()
1009 1 : }
1010 1 : if !i.lazyValueHandling.hasValuePrefix ||
1011 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1012 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1013 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1014 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1015 1 : } else {
1016 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1017 1 : }
1018 1 : return &i.ikey, i.lazyValue
1019 : }
1020 :
1021 1 : func decodeRestart(b []byte) int32 {
1022 1 : _ = b[3] // bounds check hint to compiler; see golang.org/issue/14808
1023 1 : return int32(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 |
1024 1 : uint32(b[3]&restartMaskLittleEndianHighByteWithoutSetHasSamePrefix)<<24)
1025 1 : }
1026 :
1027 : // Last implements internalIterator.Last, as documented in the pebble package.
1028 1 : func (i *blockIter) Last() (*InternalKey, base.LazyValue) {
1029 1 : if invariants.Enabled && i.isDataInvalidated() {
1030 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
1031 : }
1032 :
1033 : // Seek forward from the last restart point.
1034 1 : i.offset = decodeRestart(i.data[i.restarts+4*(i.numRestarts-1):])
1035 1 : if !i.valid() {
1036 1 : return nil, base.LazyValue{}
1037 1 : }
1038 :
1039 1 : i.readEntry()
1040 1 : i.clearCache()
1041 1 :
1042 1 : for i.nextOffset < i.restarts {
1043 1 : i.cacheEntry()
1044 1 : i.offset = i.nextOffset
1045 1 : i.readEntry()
1046 1 : }
1047 :
1048 1 : hiddenPoint := i.decodeInternalKey(i.key)
1049 1 : if hiddenPoint {
1050 1 : return i.Prev()
1051 1 : }
1052 1 : if !i.lazyValueHandling.hasValuePrefix ||
1053 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1054 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1055 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1056 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1057 1 : } else {
1058 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1059 1 : }
1060 1 : return &i.ikey, i.lazyValue
1061 : }
1062 :
1063 : // Next implements internalIterator.Next, as documented in the pebble
1064 : // package.
1065 1 : func (i *blockIter) Next() (*InternalKey, base.LazyValue) {
1066 1 : if len(i.cachedBuf) > 0 {
1067 1 : // We're switching from reverse iteration to forward iteration. We need to
1068 1 : // populate i.fullKey with the current key we're positioned at so that
1069 1 : // readEntry() can use i.fullKey for key prefix decompression. Note that we
1070 1 : // don't know whether i.key is backed by i.cachedBuf or i.fullKey (if
1071 1 : // SeekLT was the previous call, i.key may be backed by i.fullKey), but
1072 1 : // copying into i.fullKey works for both cases.
1073 1 : //
1074 1 : // TODO(peter): Rather than clearing the cache, we could instead use the
1075 1 : // cache until it is exhausted. This would likely be faster than falling
1076 1 : // through to the normal forward iteration code below.
1077 1 : i.fullKey = append(i.fullKey[:0], i.key...)
1078 1 : i.clearCache()
1079 1 : }
1080 :
1081 : start:
1082 1 : i.offset = i.nextOffset
1083 1 : if !i.valid() {
1084 1 : return nil, base.LazyValue{}
1085 1 : }
1086 1 : i.readEntry()
1087 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1088 1 : if n := len(i.key) - 8; n >= 0 {
1089 1 : trailer := binary.LittleEndian.Uint64(i.key[n:])
1090 1 : hiddenPoint := i.hideObsoletePoints &&
1091 1 : (trailer&trailerObsoleteBit != 0)
1092 1 : i.ikey.Trailer = trailer & trailerObsoleteMask
1093 1 : i.ikey.UserKey = i.key[:n:n]
1094 1 : if i.globalSeqNum != 0 {
1095 1 : i.ikey.SetSeqNum(i.globalSeqNum)
1096 1 : }
1097 1 : if hiddenPoint {
1098 1 : goto start
1099 : }
1100 0 : } else {
1101 0 : i.ikey.Trailer = uint64(InternalKeyKindInvalid)
1102 0 : i.ikey.UserKey = nil
1103 0 : }
1104 1 : if !i.lazyValueHandling.hasValuePrefix ||
1105 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1106 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1107 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1108 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1109 1 : } else {
1110 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1111 1 : }
1112 1 : return &i.ikey, i.lazyValue
1113 : }
1114 :
1115 : // NextPrefix implements (base.InternalIterator).NextPrefix.
1116 1 : func (i *blockIter) NextPrefix(succKey []byte) (*InternalKey, base.LazyValue) {
1117 1 : if i.lazyValueHandling.hasValuePrefix {
1118 1 : return i.nextPrefixV3(succKey)
1119 1 : }
1120 1 : const nextsBeforeSeek = 3
1121 1 : k, v := i.Next()
1122 1 : for j := 1; k != nil && i.cmp(k.UserKey, succKey) < 0; j++ {
1123 1 : if j >= nextsBeforeSeek {
1124 1 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1125 1 : }
1126 1 : k, v = i.Next()
1127 : }
1128 1 : return k, v
1129 : }
1130 :
1131 1 : func (i *blockIter) nextPrefixV3(succKey []byte) (*InternalKey, base.LazyValue) {
1132 1 : // Doing nexts that involve a key comparison can be expensive (and the cost
1133 1 : // depends on the key length), so we use the same threshold of 3 that we use
1134 1 : // for TableFormatPebblev2 in blockIter.nextPrefix above. The next fast path
1135 1 : // that looks at setHasSamePrefix takes ~5ns per key, which is ~150x faster
1136 1 : // than doing a SeekGE within the block, so we do this 16 times
1137 1 : // (~5ns*16=80ns), and then switch to looking at restarts. Doing the binary
1138 1 : // search for the restart consumes > 100ns. If the number of versions is >
1139 1 : // 17, we will increment nextFastCount to 17, then do a binary search, and
1140 1 : // on average need to find a key between two restarts, so another 8 steps
1141 1 : // corresponding to nextFastCount, for a mean total of 17 + 8 = 25 such
1142 1 : // steps.
1143 1 : //
1144 1 : // TODO(sumeer): use the configured restartInterval for the sstable when it
1145 1 : // was written (which we don't currently store) instead of the default value
1146 1 : // of 16.
1147 1 : const nextCmpThresholdBeforeSeek = 3
1148 1 : const nextFastThresholdBeforeRestarts = 16
1149 1 : nextCmpCount := 0
1150 1 : nextFastCount := 0
1151 1 : usedRestarts := false
1152 1 : // INVARIANT: blockIter is valid.
1153 1 : if invariants.Enabled && !i.valid() {
1154 0 : panic(errors.AssertionFailedf("nextPrefixV3 called on invalid blockIter"))
1155 : }
1156 1 : prevKeyIsSet := i.ikey.Kind() == InternalKeyKindSet
1157 1 : for {
1158 1 : i.offset = i.nextOffset
1159 1 : if !i.valid() {
1160 1 : return nil, base.LazyValue{}
1161 1 : }
1162 : // Need to decode the length integers, so we can compute nextOffset.
1163 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1164 1 : // This is an ugly performance hack. Reading entries from blocks is one of
1165 1 : // the inner-most routines and decoding the 3 varints per-entry takes
1166 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
1167 1 : // us, so we do it manually. This provides a 10-15% performance improvement
1168 1 : // on blockIter benchmarks on both go1.11 and go1.12.
1169 1 : //
1170 1 : // TODO(peter): remove this hack if go:inline is ever supported.
1171 1 :
1172 1 : // Decode the shared key length integer.
1173 1 : var shared uint32
1174 1 : if a := *((*uint8)(ptr)); a < 128 {
1175 1 : shared = uint32(a)
1176 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1177 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1178 0 : shared = uint32(b)<<7 | uint32(a)
1179 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1180 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1181 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1182 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1183 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1184 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1185 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1186 0 : } else {
1187 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1188 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1189 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1190 0 : }
1191 : // Decode the unshared key length integer.
1192 1 : var unshared uint32
1193 1 : if a := *((*uint8)(ptr)); a < 128 {
1194 1 : unshared = uint32(a)
1195 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1196 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1197 0 : unshared = uint32(b)<<7 | uint32(a)
1198 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1199 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1200 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1201 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1202 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1203 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1204 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1205 0 : } else {
1206 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1207 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1208 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1209 0 : }
1210 : // Decode the value length integer.
1211 1 : var value uint32
1212 1 : if a := *((*uint8)(ptr)); a < 128 {
1213 1 : value = uint32(a)
1214 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1215 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1216 0 : value = uint32(b)<<7 | uint32(a)
1217 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1218 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1219 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1220 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1221 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1222 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1223 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1224 0 : } else {
1225 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1226 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1227 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1228 0 : }
1229 : // The starting position of the value.
1230 1 : valuePtr := unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
1231 1 : i.nextOffset = int32(uintptr(valuePtr)-uintptr(i.ptr)) + int32(value)
1232 1 : if invariants.Enabled && unshared < 8 {
1233 0 : // This should not happen since only the key prefix is shared, so even
1234 0 : // if the prefix length is the same as the user key length, the unshared
1235 0 : // will include the trailer.
1236 0 : panic(errors.AssertionFailedf("unshared %d is too small", unshared))
1237 : }
1238 : // The trailer is written in little endian, so the key kind is the first
1239 : // byte in the trailer that is encoded in the slice [unshared-8:unshared].
1240 1 : keyKind := InternalKeyKind((*[manual.MaxArrayLen]byte)(ptr)[unshared-8])
1241 1 : keyKind = keyKind & base.InternalKeyKindSSTableInternalObsoleteMask
1242 1 : prefixChanged := false
1243 1 : if keyKind == InternalKeyKindSet {
1244 1 : if invariants.Enabled && value == 0 {
1245 0 : panic(errors.AssertionFailedf("value is of length 0, but we expect a valuePrefix"))
1246 : }
1247 1 : valPrefix := *((*valuePrefix)(valuePtr))
1248 1 : if setHasSamePrefix(valPrefix) {
1249 1 : // Fast-path. No need to assemble i.fullKey, or update i.key. We know
1250 1 : // that subsequent keys will not have a shared length that is greater
1251 1 : // than the prefix of the current key, which is also the prefix of
1252 1 : // i.key. Since we are continuing to iterate, we don't need to
1253 1 : // initialize i.ikey and i.lazyValue (these are initialized before
1254 1 : // returning).
1255 1 : nextFastCount++
1256 1 : if nextFastCount > nextFastThresholdBeforeRestarts {
1257 1 : if usedRestarts {
1258 1 : // Exhausted iteration budget. This will never happen unless
1259 1 : // someone is using a restart interval > 16. It is just to guard
1260 1 : // against long restart intervals causing too much iteration.
1261 1 : break
1262 : }
1263 : // Haven't used restarts yet, so find the first restart at or beyond
1264 : // the current offset.
1265 1 : targetOffset := i.offset
1266 1 : var index int32
1267 1 : {
1268 1 : // NB: manually inlined sort.Sort is ~5% faster.
1269 1 : //
1270 1 : // f defined for a restart point is true iff the offset >=
1271 1 : // targetOffset.
1272 1 : // Define f(-1) == false and f(i.numRestarts) == true.
1273 1 : // Invariant: f(index-1) == false, f(upper) == true.
1274 1 : upper := i.numRestarts
1275 1 : for index < upper {
1276 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1277 1 : // index ≤ h < upper
1278 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
1279 1 : if offset < targetOffset {
1280 1 : index = h + 1 // preserves f(index-1) == false
1281 1 : } else {
1282 1 : upper = h // preserves f(upper) == true
1283 1 : }
1284 : }
1285 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1286 : // => answer is index.
1287 : }
1288 1 : usedRestarts = true
1289 1 : nextFastCount = 0
1290 1 : if index == i.numRestarts {
1291 1 : // Already past the last real restart, so iterate a bit more until
1292 1 : // we are done with the block.
1293 1 : continue
1294 : }
1295 : // Have some real restarts after index. NB: index is the first
1296 : // restart at or beyond the current offset.
1297 1 : startingIndex := index
1298 1 : for index != i.numRestarts &&
1299 1 : // The restart at index is 4 bytes written in little endian format
1300 1 : // starting at i.restart+4*index. The 0th byte is the least
1301 1 : // significant and the 3rd byte is the most significant. Since the
1302 1 : // most significant bit of the 3rd byte is what we use for
1303 1 : // encoding the set-has-same-prefix information, the indexing
1304 1 : // below has +3.
1305 1 : i.data[i.restarts+4*index+3]&restartMaskLittleEndianHighByteOnlySetHasSamePrefix != 0 {
1306 1 : // We still have the same prefix, so move to the next restart.
1307 1 : index++
1308 1 : }
1309 : // index is the first restart that did not have the same prefix.
1310 1 : if index != startingIndex {
1311 1 : // Managed to skip past at least one restart. Resume iteration
1312 1 : // from index-1. Since nextFastCount has been reset to 0, we
1313 1 : // should be able to iterate to the next prefix.
1314 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1315 1 : i.readEntry()
1316 1 : }
1317 : // Else, unable to skip past any restart. Resume iteration. Since
1318 : // nextFastCount has been reset to 0, we should be able to iterate
1319 : // to the next prefix.
1320 1 : continue
1321 : }
1322 1 : continue
1323 1 : } else if prevKeyIsSet {
1324 1 : prefixChanged = true
1325 1 : }
1326 1 : } else {
1327 1 : prevKeyIsSet = false
1328 1 : }
1329 : // Slow-path cases:
1330 : // - (Likely) The prefix has changed.
1331 : // - (Unlikely) The prefix has not changed.
1332 : // We assemble the key etc. under the assumption that it is the likely
1333 : // case.
1334 1 : unsharedKey := getBytes(ptr, int(unshared))
1335 1 : // TODO(sumeer): move this into the else block below. This is a bit tricky
1336 1 : // since the current logic assumes we have always copied the latest key
1337 1 : // into fullKey, which is why when we get to the next key we can (a)
1338 1 : // access i.fullKey[:shared], (b) append only the unsharedKey to
1339 1 : // i.fullKey. For (a), we can access i.key[:shared] since that memory is
1340 1 : // valid (even if unshared). For (b), we will need to remember whether
1341 1 : // i.key refers to i.fullKey or not, and can append the unsharedKey only
1342 1 : // in the former case and for the latter case need to copy the shared part
1343 1 : // too. This same comment applies to the other place where we can do this
1344 1 : // optimization, in readEntry().
1345 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
1346 1 : i.val = getBytes(valuePtr, int(value))
1347 1 : if shared == 0 {
1348 1 : // Provide stability for the key across positioning calls if the key
1349 1 : // doesn't share a prefix with the previous key. This removes requiring the
1350 1 : // key to be copied if the caller knows the block has a restart interval of
1351 1 : // 1. An important example of this is range-del blocks.
1352 1 : i.key = unsharedKey
1353 1 : } else {
1354 1 : i.key = i.fullKey
1355 1 : }
1356 : // Manually inlined version of i.decodeInternalKey(i.key).
1357 1 : hiddenPoint := false
1358 1 : if n := len(i.key) - 8; n >= 0 {
1359 1 : trailer := binary.LittleEndian.Uint64(i.key[n:])
1360 1 : hiddenPoint = i.hideObsoletePoints &&
1361 1 : (trailer&trailerObsoleteBit != 0)
1362 1 : i.ikey.Trailer = trailer & trailerObsoleteMask
1363 1 : i.ikey.UserKey = i.key[:n:n]
1364 1 : if i.globalSeqNum != 0 {
1365 1 : i.ikey.SetSeqNum(i.globalSeqNum)
1366 1 : }
1367 0 : } else {
1368 0 : i.ikey.Trailer = uint64(InternalKeyKindInvalid)
1369 0 : i.ikey.UserKey = nil
1370 0 : }
1371 1 : nextCmpCount++
1372 1 : if invariants.Enabled && prefixChanged && i.cmp(i.ikey.UserKey, succKey) < 0 {
1373 0 : panic(errors.AssertionFailedf("prefix should have changed but %x < %x",
1374 0 : i.ikey.UserKey, succKey))
1375 : }
1376 1 : if prefixChanged || i.cmp(i.ikey.UserKey, succKey) >= 0 {
1377 1 : // Prefix has changed.
1378 1 : if hiddenPoint {
1379 1 : return i.Next()
1380 1 : }
1381 1 : if invariants.Enabled && !i.lazyValueHandling.hasValuePrefix {
1382 0 : panic(errors.AssertionFailedf("nextPrefixV3 being run for non-v3 sstable"))
1383 : }
1384 1 : if base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1385 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1386 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1387 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1388 1 : } else {
1389 0 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1390 0 : }
1391 1 : return &i.ikey, i.lazyValue
1392 : }
1393 : // Else prefix has not changed.
1394 :
1395 1 : if nextCmpCount >= nextCmpThresholdBeforeSeek {
1396 0 : break
1397 : }
1398 : }
1399 1 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1400 : }
1401 :
1402 : // Prev implements internalIterator.Prev, as documented in the pebble
1403 : // package.
1404 1 : func (i *blockIter) Prev() (*InternalKey, base.LazyValue) {
1405 1 : start:
1406 1 : for n := len(i.cached) - 1; n >= 0; n-- {
1407 1 : i.nextOffset = i.offset
1408 1 : e := &i.cached[n]
1409 1 : i.offset = e.offset
1410 1 : i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
1411 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1412 1 : i.key = i.cachedBuf[e.keyStart:e.keyEnd]
1413 1 : if n := len(i.key) - 8; n >= 0 {
1414 1 : trailer := binary.LittleEndian.Uint64(i.key[n:])
1415 1 : hiddenPoint := i.hideObsoletePoints &&
1416 1 : (trailer&trailerObsoleteBit != 0)
1417 1 : if hiddenPoint {
1418 1 : continue
1419 : }
1420 1 : i.ikey.Trailer = trailer & trailerObsoleteMask
1421 1 : i.ikey.UserKey = i.key[:n:n]
1422 1 : if i.globalSeqNum != 0 {
1423 1 : i.ikey.SetSeqNum(i.globalSeqNum)
1424 1 : }
1425 0 : } else {
1426 0 : i.ikey.Trailer = uint64(InternalKeyKindInvalid)
1427 0 : i.ikey.UserKey = nil
1428 0 : }
1429 1 : i.cached = i.cached[:n]
1430 1 : if !i.lazyValueHandling.hasValuePrefix ||
1431 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1432 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1433 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1434 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1435 1 : } else {
1436 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1437 1 : }
1438 1 : return &i.ikey, i.lazyValue
1439 : }
1440 :
1441 1 : i.clearCache()
1442 1 : if i.offset <= 0 {
1443 1 : i.offset = -1
1444 1 : i.nextOffset = 0
1445 1 : return nil, base.LazyValue{}
1446 1 : }
1447 :
1448 1 : targetOffset := i.offset
1449 1 : var index int32
1450 1 :
1451 1 : {
1452 1 : // NB: manually inlined sort.Sort is ~5% faster.
1453 1 : //
1454 1 : // Define f(-1) == false and f(n) == true.
1455 1 : // Invariant: f(index-1) == false, f(upper) == true.
1456 1 : upper := i.numRestarts
1457 1 : for index < upper {
1458 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1459 1 : // index ≤ h < upper
1460 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
1461 1 : if offset < targetOffset {
1462 1 : // Looking for the first restart that has offset >= targetOffset, so
1463 1 : // ignore h and earlier.
1464 1 : index = h + 1 // preserves f(i-1) == false
1465 1 : } else {
1466 1 : upper = h // preserves f(j) == true
1467 1 : }
1468 : }
1469 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1470 : // => answer is index.
1471 : }
1472 :
1473 : // index is first restart with offset >= targetOffset. Note that
1474 : // targetOffset may not be at a restart point since one can call Prev()
1475 : // after Next() (so the cache was not populated) and targetOffset refers to
1476 : // the current entry. index-1 must have an offset < targetOffset (it can't
1477 : // be equal to targetOffset since the binary search would have selected that
1478 : // as the index).
1479 1 : i.offset = 0
1480 1 : if index > 0 {
1481 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1482 1 : }
1483 : // TODO(sumeer): why is the else case not an error given targetOffset is a
1484 : // valid offset.
1485 :
1486 1 : i.readEntry()
1487 1 :
1488 1 : // We stop when i.nextOffset == targetOffset since the targetOffset is the
1489 1 : // entry we are stepping back from, and we don't need to cache the entry
1490 1 : // before it, since it is the candidate to return.
1491 1 : for i.nextOffset < targetOffset {
1492 1 : i.cacheEntry()
1493 1 : i.offset = i.nextOffset
1494 1 : i.readEntry()
1495 1 : }
1496 :
1497 1 : hiddenPoint := i.decodeInternalKey(i.key)
1498 1 : if hiddenPoint {
1499 1 : // Use the cache.
1500 1 : goto start
1501 : }
1502 1 : if !i.lazyValueHandling.hasValuePrefix ||
1503 1 : base.TrailerKind(i.ikey.Trailer) != InternalKeyKindSet {
1504 1 : i.lazyValue = base.MakeInPlaceValue(i.val)
1505 1 : } else if i.lazyValueHandling.vbr == nil || !isValueHandle(valuePrefix(i.val[0])) {
1506 1 : i.lazyValue = base.MakeInPlaceValue(i.val[1:])
1507 1 : } else {
1508 1 : i.lazyValue = i.lazyValueHandling.vbr.getLazyValueForPrefixAndValueHandle(i.val)
1509 1 : }
1510 1 : return &i.ikey, i.lazyValue
1511 : }
1512 :
1513 : // Key implements internalIterator.Key, as documented in the pebble package.
1514 1 : func (i *blockIter) Key() *InternalKey {
1515 1 : return &i.ikey
1516 1 : }
1517 :
1518 1 : func (i *blockIter) value() base.LazyValue {
1519 1 : return i.lazyValue
1520 1 : }
1521 :
1522 : // Error implements internalIterator.Error, as documented in the pebble
1523 : // package.
1524 1 : func (i *blockIter) Error() error {
1525 1 : return nil // infallible
1526 1 : }
1527 :
1528 : // Close implements internalIterator.Close, as documented in the pebble
1529 : // package.
1530 1 : func (i *blockIter) Close() error {
1531 1 : i.handle.Release()
1532 1 : i.handle = bufferHandle{}
1533 1 : i.val = nil
1534 1 : i.lazyValue = base.LazyValue{}
1535 1 : i.lazyValueHandling.vbr = nil
1536 1 : return nil
1537 1 : }
1538 :
1539 0 : func (i *blockIter) SetBounds(lower, upper []byte) {
1540 0 : // This should never be called as bounds are handled by sstable.Iterator.
1541 0 : panic("pebble: SetBounds unimplemented")
1542 : }
1543 :
1544 0 : func (i *blockIter) SetContext(_ context.Context) {}
1545 :
1546 1 : func (i *blockIter) valid() bool {
1547 1 : return i.offset >= 0 && i.offset < i.restarts
1548 1 : }
1549 :
1550 : // fragmentBlockIter wraps a blockIter, implementing the
1551 : // keyspan.FragmentIterator interface. It's used for reading range deletion and
1552 : // range key blocks.
1553 : //
1554 : // Range deletions and range keys are fragmented before they're persisted to the
1555 : // block. Overlapping fragments have identical bounds. The fragmentBlockIter
1556 : // gathers all the fragments with identical bounds within a block and returns a
1557 : // single keyspan.Span describing all the keys defined over the span.
1558 : //
1559 : // # Memory lifetime
1560 : //
1561 : // A Span returned by fragmentBlockIter is only guaranteed to be stable until
1562 : // the next fragmentBlockIter iteration positioning method. A Span's Keys slice
1563 : // may be reused, so the user must not assume it's stable.
1564 : //
1565 : // Blocks holding range deletions and range keys are configured to use a restart
1566 : // interval of 1. This provides key stability. The caller may treat the various
1567 : // byte slices (start, end, suffix, value) as stable for the lifetime of the
1568 : // iterator.
1569 : type fragmentBlockIter struct {
1570 : blockIter blockIter
1571 : keyBuf [2]keyspan.Key
1572 : span keyspan.Span
1573 : err error
1574 : dir int8
1575 : closeHook func(i keyspan.FragmentIterator) error
1576 :
1577 : // elideSameSeqnum, if true, returns only the first-occurring (in forward
1578 : // order) Key for each sequence number.
1579 : elideSameSeqnum bool
1580 : }
1581 :
1582 1 : func (i *fragmentBlockIter) resetForReuse() fragmentBlockIter {
1583 1 : return fragmentBlockIter{blockIter: i.blockIter.resetForReuse()}
1584 1 : }
1585 :
1586 1 : func (i *fragmentBlockIter) decodeSpanKeys(k *InternalKey, internalValue []byte) {
1587 1 : // TODO(jackson): The use of i.span.Keys to accumulate keys across multiple
1588 1 : // calls to Decode is too confusing and subtle. Refactor to make it
1589 1 : // explicit.
1590 1 :
1591 1 : // decode the contents of the fragment's value. This always includes at
1592 1 : // least the end key: RANGEDELs store the end key directly as the value,
1593 1 : // whereas the various range key kinds store are more complicated. The
1594 1 : // details of the range key internal value format are documented within the
1595 1 : // internal/rangekey package.
1596 1 : switch k.Kind() {
1597 1 : case base.InternalKeyKindRangeDelete:
1598 1 : i.span = rangedel.Decode(*k, internalValue, i.span.Keys)
1599 1 : i.err = nil
1600 1 : case base.InternalKeyKindRangeKeySet, base.InternalKeyKindRangeKeyUnset, base.InternalKeyKindRangeKeyDelete:
1601 1 : i.span, i.err = rangekey.Decode(*k, internalValue, i.span.Keys)
1602 0 : default:
1603 0 : i.span = keyspan.Span{}
1604 0 : i.err = base.CorruptionErrorf("pebble: corrupt keyspan fragment of kind %d", k.Kind())
1605 : }
1606 : }
1607 :
1608 1 : func (i *fragmentBlockIter) elideKeysOfSameSeqNum() {
1609 1 : if invariants.Enabled {
1610 1 : if !i.elideSameSeqnum || len(i.span.Keys) == 0 {
1611 0 : panic("elideKeysOfSameSeqNum called when it should not be")
1612 : }
1613 : }
1614 1 : lastSeqNum := i.span.Keys[0].SeqNum()
1615 1 : k := 1
1616 1 : for j := 1; j < len(i.span.Keys); j++ {
1617 1 : if lastSeqNum != i.span.Keys[j].SeqNum() {
1618 1 : lastSeqNum = i.span.Keys[j].SeqNum()
1619 1 : i.span.Keys[k] = i.span.Keys[j]
1620 1 : k++
1621 1 : }
1622 : }
1623 1 : i.span.Keys = i.span.Keys[:k]
1624 : }
1625 :
1626 : // gatherForward gathers internal keys with identical bounds. Keys defined over
1627 : // spans of the keyspace are fragmented such that any overlapping key spans have
1628 : // identical bounds. When these spans are persisted to a range deletion or range
1629 : // key block, they may be persisted as multiple internal keys in order to encode
1630 : // multiple sequence numbers or key kinds.
1631 : //
1632 : // gatherForward iterates forward, re-combining the fragmented internal keys to
1633 : // reconstruct a keyspan.Span that holds all the keys defined over the span.
1634 1 : func (i *fragmentBlockIter) gatherForward(k *InternalKey, lazyValue base.LazyValue) *keyspan.Span {
1635 1 : i.span = keyspan.Span{}
1636 1 : if k == nil || !i.blockIter.valid() {
1637 1 : return nil
1638 1 : }
1639 1 : i.err = nil
1640 1 : // Use the i.keyBuf array to back the Keys slice to prevent an allocation
1641 1 : // when a span contains few keys.
1642 1 : i.span.Keys = i.keyBuf[:0]
1643 1 :
1644 1 : // Decode the span's end key and individual keys from the value.
1645 1 : internalValue := lazyValue.InPlaceValue()
1646 1 : i.decodeSpanKeys(k, internalValue)
1647 1 : if i.err != nil {
1648 0 : return nil
1649 0 : }
1650 1 : prevEnd := i.span.End
1651 1 :
1652 1 : // There might exist additional internal keys with identical bounds encoded
1653 1 : // within the block. Iterate forward, accumulating all the keys with
1654 1 : // identical bounds to s.
1655 1 : k, lazyValue = i.blockIter.Next()
1656 1 : internalValue = lazyValue.InPlaceValue()
1657 1 : for k != nil && i.blockIter.cmp(k.UserKey, i.span.Start) == 0 {
1658 1 : i.decodeSpanKeys(k, internalValue)
1659 1 : if i.err != nil {
1660 0 : return nil
1661 0 : }
1662 :
1663 : // Since k indicates an equal start key, the encoded end key must
1664 : // exactly equal the original end key from the first internal key.
1665 : // Overlapping fragments are required to have exactly equal start and
1666 : // end bounds.
1667 1 : if i.blockIter.cmp(prevEnd, i.span.End) != 0 {
1668 0 : i.err = base.CorruptionErrorf("pebble: corrupt keyspan fragmentation")
1669 0 : i.span = keyspan.Span{}
1670 0 : return nil
1671 0 : }
1672 1 : k, lazyValue = i.blockIter.Next()
1673 1 : internalValue = lazyValue.InPlaceValue()
1674 : }
1675 1 : if i.elideSameSeqnum && len(i.span.Keys) > 0 {
1676 1 : i.elideKeysOfSameSeqNum()
1677 1 : }
1678 : // i.blockIter is positioned over the first internal key for the next span.
1679 1 : return &i.span
1680 : }
1681 :
1682 : // gatherBackward gathers internal keys with identical bounds. Keys defined over
1683 : // spans of the keyspace are fragmented such that any overlapping key spans have
1684 : // identical bounds. When these spans are persisted to a range deletion or range
1685 : // key block, they may be persisted as multiple internal keys in order to encode
1686 : // multiple sequence numbers or key kinds.
1687 : //
1688 : // gatherBackward iterates backwards, re-combining the fragmented internal keys
1689 : // to reconstruct a keyspan.Span that holds all the keys defined over the span.
1690 1 : func (i *fragmentBlockIter) gatherBackward(k *InternalKey, lazyValue base.LazyValue) *keyspan.Span {
1691 1 : i.span = keyspan.Span{}
1692 1 : if k == nil || !i.blockIter.valid() {
1693 1 : return nil
1694 1 : }
1695 1 : i.err = nil
1696 1 : // Use the i.keyBuf array to back the Keys slice to prevent an allocation
1697 1 : // when a span contains few keys.
1698 1 : i.span.Keys = i.keyBuf[:0]
1699 1 :
1700 1 : // Decode the span's end key and individual keys from the value.
1701 1 : internalValue := lazyValue.InPlaceValue()
1702 1 : i.decodeSpanKeys(k, internalValue)
1703 1 : if i.err != nil {
1704 0 : return nil
1705 0 : }
1706 1 : prevEnd := i.span.End
1707 1 :
1708 1 : // There might exist additional internal keys with identical bounds encoded
1709 1 : // within the block. Iterate backward, accumulating all the keys with
1710 1 : // identical bounds to s.
1711 1 : k, lazyValue = i.blockIter.Prev()
1712 1 : internalValue = lazyValue.InPlaceValue()
1713 1 : for k != nil && i.blockIter.cmp(k.UserKey, i.span.Start) == 0 {
1714 1 : i.decodeSpanKeys(k, internalValue)
1715 1 : if i.err != nil {
1716 0 : return nil
1717 0 : }
1718 :
1719 : // Since k indicates an equal start key, the encoded end key must
1720 : // exactly equal the original end key from the first internal key.
1721 : // Overlapping fragments are required to have exactly equal start and
1722 : // end bounds.
1723 1 : if i.blockIter.cmp(prevEnd, i.span.End) != 0 {
1724 0 : i.err = base.CorruptionErrorf("pebble: corrupt keyspan fragmentation")
1725 0 : i.span = keyspan.Span{}
1726 0 : return nil
1727 0 : }
1728 1 : k, lazyValue = i.blockIter.Prev()
1729 1 : internalValue = lazyValue.InPlaceValue()
1730 : }
1731 : // i.blockIter is positioned over the last internal key for the previous
1732 : // span.
1733 :
1734 : // Backwards iteration encounters internal keys in the wrong order.
1735 1 : keyspan.SortKeysByTrailer(&i.span.Keys)
1736 1 :
1737 1 : if i.elideSameSeqnum && len(i.span.Keys) > 0 {
1738 1 : i.elideKeysOfSameSeqNum()
1739 1 : }
1740 1 : return &i.span
1741 : }
1742 :
1743 : // Error implements (keyspan.FragmentIterator).Error.
1744 1 : func (i *fragmentBlockIter) Error() error {
1745 1 : return i.err
1746 1 : }
1747 :
1748 : // Close implements (keyspan.FragmentIterator).Close.
1749 1 : func (i *fragmentBlockIter) Close() error {
1750 1 : var err error
1751 1 : if i.closeHook != nil {
1752 0 : err = i.closeHook(i)
1753 0 : }
1754 1 : err = firstError(err, i.blockIter.Close())
1755 1 : return err
1756 : }
1757 :
1758 : // First implements (keyspan.FragmentIterator).First
1759 1 : func (i *fragmentBlockIter) First() *keyspan.Span {
1760 1 : i.dir = +1
1761 1 : return i.gatherForward(i.blockIter.First())
1762 1 : }
1763 :
1764 : // Last implements (keyspan.FragmentIterator).Last.
1765 1 : func (i *fragmentBlockIter) Last() *keyspan.Span {
1766 1 : i.dir = -1
1767 1 : return i.gatherBackward(i.blockIter.Last())
1768 1 : }
1769 :
1770 : // Next implements (keyspan.FragmentIterator).Next.
1771 1 : func (i *fragmentBlockIter) Next() *keyspan.Span {
1772 1 : switch {
1773 1 : case i.dir == -1 && !i.span.Valid():
1774 1 : // Switching directions.
1775 1 : //
1776 1 : // i.blockIter is exhausted, before the first key. Move onto the first.
1777 1 : i.blockIter.First()
1778 1 : i.dir = +1
1779 1 : case i.dir == -1 && i.span.Valid():
1780 1 : // Switching directions.
1781 1 : //
1782 1 : // i.blockIter is currently positioned over the last internal key for
1783 1 : // the previous span. Next it once to move to the first internal key
1784 1 : // that makes up the current span, and gatherForwaad to land on the
1785 1 : // first internal key making up the next span.
1786 1 : //
1787 1 : // In the diagram below, if the last span returned to the user during
1788 1 : // reverse iteration was [b,c), i.blockIter is currently positioned at
1789 1 : // [a,b). The block iter must be positioned over [d,e) to gather the
1790 1 : // next span's fragments.
1791 1 : //
1792 1 : // ... [a,b) [b,c) [b,c) [b,c) [d,e) ...
1793 1 : // ^ ^
1794 1 : // i.blockIter want
1795 1 : if x := i.gatherForward(i.blockIter.Next()); invariants.Enabled && !x.Valid() {
1796 0 : panic("pebble: invariant violation: next entry unexpectedly invalid")
1797 : }
1798 1 : i.dir = +1
1799 : }
1800 : // We know that this blockIter has in-place values.
1801 1 : return i.gatherForward(&i.blockIter.ikey, base.MakeInPlaceValue(i.blockIter.val))
1802 : }
1803 :
1804 : // Prev implements (keyspan.FragmentIterator).Prev.
1805 1 : func (i *fragmentBlockIter) Prev() *keyspan.Span {
1806 1 : switch {
1807 1 : case i.dir == +1 && !i.span.Valid():
1808 1 : // Switching directions.
1809 1 : //
1810 1 : // i.blockIter is exhausted, after the last key. Move onto the last.
1811 1 : i.blockIter.Last()
1812 1 : i.dir = -1
1813 1 : case i.dir == +1 && i.span.Valid():
1814 1 : // Switching directions.
1815 1 : //
1816 1 : // i.blockIter is currently positioned over the first internal key for
1817 1 : // the next span. Prev it once to move to the last internal key that
1818 1 : // makes up the current span, and gatherBackward to land on the last
1819 1 : // internal key making up the previous span.
1820 1 : //
1821 1 : // In the diagram below, if the last span returned to the user during
1822 1 : // forward iteration was [b,c), i.blockIter is currently positioned at
1823 1 : // [d,e). The block iter must be positioned over [a,b) to gather the
1824 1 : // previous span's fragments.
1825 1 : //
1826 1 : // ... [a,b) [b,c) [b,c) [b,c) [d,e) ...
1827 1 : // ^ ^
1828 1 : // want i.blockIter
1829 1 : if x := i.gatherBackward(i.blockIter.Prev()); invariants.Enabled && !x.Valid() {
1830 0 : panic("pebble: invariant violation: previous entry unexpectedly invalid")
1831 : }
1832 1 : i.dir = -1
1833 : }
1834 : // We know that this blockIter has in-place values.
1835 1 : return i.gatherBackward(&i.blockIter.ikey, base.MakeInPlaceValue(i.blockIter.val))
1836 : }
1837 :
1838 : // SeekGE implements (keyspan.FragmentIterator).SeekGE.
1839 1 : func (i *fragmentBlockIter) SeekGE(k []byte) *keyspan.Span {
1840 1 : if s := i.SeekLT(k); s != nil && i.blockIter.cmp(k, s.End) < 0 {
1841 1 : return s
1842 1 : }
1843 : // TODO(jackson): If the above i.SeekLT(k) discovers a span but the span
1844 : // doesn't meet the k < s.End comparison, then there's no need for the
1845 : // SeekLT to gatherBackward.
1846 1 : return i.Next()
1847 : }
1848 :
1849 : // SeekLT implements (keyspan.FragmentIterator).SeekLT.
1850 1 : func (i *fragmentBlockIter) SeekLT(k []byte) *keyspan.Span {
1851 1 : i.dir = -1
1852 1 : return i.gatherBackward(i.blockIter.SeekLT(k, base.SeekLTFlagsNone))
1853 1 : }
1854 :
1855 : // String implements fmt.Stringer.
1856 0 : func (i *fragmentBlockIter) String() string {
1857 0 : return "fragment-block-iter"
1858 0 : }
1859 :
1860 : // SetCloseHook implements sstable.FragmentIterator.
1861 0 : func (i *fragmentBlockIter) SetCloseHook(fn func(i keyspan.FragmentIterator) error) {
1862 0 : i.closeHook = fn
1863 0 : }
|
