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 rowblk
6 :
7 : import (
8 : "bytes"
9 : "context"
10 : "encoding/binary"
11 : "io"
12 : "slices"
13 : "sort"
14 : "unsafe"
15 :
16 : "github.com/cockroachdb/errors"
17 : "github.com/cockroachdb/pebble/internal/base"
18 : "github.com/cockroachdb/pebble/internal/invariants"
19 : "github.com/cockroachdb/pebble/internal/manual"
20 : "github.com/cockroachdb/pebble/internal/treeprinter"
21 : "github.com/cockroachdb/pebble/sstable/block"
22 : )
23 :
24 : // Iter is an iterator over a single block of data.
25 : //
26 : // An Iter provides an additional guarantee around key stability when a block
27 : // has a restart interval of 1 (i.e. when there is no prefix compression). Key
28 : // stability refers to whether the InternalKey.UserKey bytes returned by a
29 : // positioning call will remain stable after a subsequent positioning call. The
30 : // normal case is that a positioning call will invalidate any previously
31 : // returned InternalKey.UserKey. If a block has a restart interval of 1 (no
32 : // prefix compression), Iter guarantees that InternalKey.UserKey will point to
33 : // the key as stored in the block itself which will remain valid until the Iter
34 : // is closed. The key stability guarantee is used by the range tombstone and
35 : // range key code, which knows that the respective blocks are always encoded
36 : // with a restart interval of 1. This per-block key stability guarantee is
37 : // sufficient for range tombstones and range deletes as they are always encoded
38 : // in a single block. Note: this stability guarantee no longer holds for a block
39 : // iter with synthetic prefix/suffix replacement, but we don't use the synthetic
40 : // suffix/prefix functionality of Iter for range keys.
41 : //
42 : // An Iter also provides a value stability guarantee for range deletions and
43 : // range keys since there is only a single range deletion and range key block
44 : // per sstable and the Iter will not release the bytes for the block until it is
45 : // closed.
46 : //
47 : // Note on why Iter knows about lazyValueHandling:
48 : //
49 : // Iter's positioning functions (that return a LazyValue), are too
50 : // complex to inline even prior to lazyValueHandling. Iter.Next and
51 : // Iter.First were by far the cheapest and had costs 195 and 180
52 : // respectively, which exceeds the budget of 80. We initially tried to keep
53 : // the lazyValueHandling logic out of Iter by wrapping it with a
54 : // lazyValueDataBlockIter. singleLevelIter and twoLevelIter would use this
55 : // wrapped iter. The functions in lazyValueDataBlockIter were simple, in that
56 : // they called the corresponding Iter func and then decided whether the
57 : // value was in fact in-place (so return immediately) or needed further
58 : // handling. But these also turned out too costly for mid-stack inlining since
59 : // simple calls like the following have a high cost that is barely under the
60 : // budget of 80
61 : //
62 : // k, v := i.data.SeekGE(key, flags) // cost 74
63 : // k, v := i.data.Next() // cost 72
64 : //
65 : // We have 2 options for minimizing performance regressions:
66 : // - Include the lazyValueHandling logic in the already non-inlineable
67 : // Iter functions: Since most of the time is spent in data block iters,
68 : // it is acceptable to take the small hit of unnecessary branching (which
69 : // hopefully branch prediction will predict correctly) for other kinds of
70 : // blocks.
71 : // - Duplicate the logic of singleLevelIterator and twoLevelIterator for the
72 : // v3 sstable and only use the aforementioned lazyValueDataBlockIter for a
73 : // v3 sstable. We would want to manage these copies via code generation.
74 : //
75 : // We have picked the first option here.
76 : type Iter struct {
77 : cmp base.Compare
78 : split base.Split
79 :
80 : // Iterator transforms.
81 : //
82 : // SyntheticSuffix, if set, will replace the decoded ikey.UserKey suffix
83 : // before the key is returned to the user. A sequence of iter operations on a
84 : // block with a syntheticSuffix rule should return keys as if those operations
85 : // ran on a block with keys that all had the syntheticSuffix. As an example:
86 : // any sequence of block iter cmds should return the same keys for the
87 : // following two blocks:
88 : //
89 : // blockA: a@3,b@3,c@3
90 : // blockB: a@1,b@2,c@1 with syntheticSuffix=3
91 : //
92 : // To ensure this, Suffix replacement will not change the ordering of keys in
93 : // the block because the iter assumes that no two keys in the block share the
94 : // same prefix. Furthermore, during SeekGE and SeekLT operations, the block
95 : // iterator handles "off by one" errors (explained in more detail in those
96 : // functions) when, for a given key, originalSuffix < searchSuffix <
97 : // replacementSuffix, with integer comparison. To handle these cases, the
98 : // iterator assumes:
99 : //
100 : // pebble.Compare(keyPrefix{replacementSuffix},keyPrefix{originalSuffix}) < 0
101 : // for keys with a suffix.
102 : //
103 : // NB: it is possible for a block iter to add a synthetic suffix on a key
104 : // without a suffix, which implies
105 : // pebble.Compare(keyPrefix{replacementSuffix},keyPrefix{noSuffix}) > 0 ,
106 : // however, the iterator would never need to handle an off by one error in
107 : // this case since originalSuffix (empty) > searchSuffix (non empty), with
108 : // integer comparison.
109 : //
110 : //
111 : // In addition, we also assume that any block with rangekeys will not contain
112 : // a synthetic suffix.
113 : transforms block.IterTransforms
114 :
115 : // offset is the byte index that marks where the current key/value is
116 : // encoded in the block.
117 : offset int32
118 : // nextOffset is the byte index where the next key/value is encoded in the
119 : // block.
120 : nextOffset int32
121 : // A "restart point" in a block is a point where the full key is encoded,
122 : // instead of just having a suffix of the key encoded. See readEntry() for
123 : // how prefix compression of keys works. Keys in between two restart points
124 : // only have a suffix encoded in the block. When restart interval is 1, no
125 : // prefix compression of keys happens. This is the case with range tombstone
126 : // blocks.
127 : //
128 : // All restart offsets are listed in increasing order in
129 : // i.ptr[i.restarts:len(block)-4], while numRestarts is encoded in the last
130 : // 4 bytes of the block as a uint32 (i.ptr[len(block)-4:]). i.restarts can
131 : // therefore be seen as the point where data in the block ends, and a list
132 : // of offsets of all restart points begins.
133 : restarts int32
134 : // Number of restart points in this block. Encoded at the end of the block
135 : // as a uint32.
136 : numRestarts int32
137 : ptr unsafe.Pointer
138 : data []byte
139 : // key contains the raw key the iterator is currently pointed at. This may
140 : // point directly to data stored in the block (for a key which has no prefix
141 : // compression), to fullKey (for a prefix compressed key), or to a slice of
142 : // data stored in cachedBuf (during reverse iteration).
143 : //
144 : // NB: In general, key contains the same logical content as ikey
145 : // (i.e. ikey = decode(key)), but if the iterator contains a synthetic suffix
146 : // replacement rule, this will not be the case. Therefore, key should never
147 : // be used after ikey is set.
148 : key []byte
149 : // fullKey is a buffer used for key prefix decompression. Note that if
150 : // transforms.SyntheticPrifix is not nil, fullKey always starts with that
151 : // prefix.
152 : fullKey []byte
153 : // val contains the value the iterator is currently pointed at. If non-nil,
154 : // this points to a slice of the block data.
155 : val []byte
156 : // ikv contains the decoded internal KV the iterator is currently positioned
157 : // at.
158 : //
159 : // ikv.InternalKey contains the decoded InternalKey the iterator is
160 : // currently pointed at. Note that the memory backing ikv.UserKey is either
161 : // data stored directly in the block, fullKey, or cachedBuf. The key
162 : // stability guarantee for blocks built with a restart interval of 1 is
163 : // achieved by having ikv.UserKey always point to data stored directly in
164 : // the block.
165 : //
166 : // ikv.LazyValue is val turned into a LazyValue, whenever a positioning
167 : // method returns a non-nil key-value pair.
168 : ikv base.InternalKV
169 : // cached and cachedBuf are used during reverse iteration. They are needed
170 : // because we can't perform prefix decoding in reverse, only in the forward
171 : // direction. In order to iterate in reverse, we decode and cache the entries
172 : // between two restart points.
173 : //
174 : // Note that cached[len(cached)-1] contains the previous entry to the one the
175 : // blockIter is currently pointed at. As usual, nextOffset will contain the
176 : // offset of the next entry. During reverse iteration, nextOffset will be
177 : // updated to point to offset, and we'll set the blockIter to point at the
178 : // entry cached[len(cached)-1]. See Prev() for more details.
179 : //
180 : // For a block encoded with a restart interval of 1, cached and cachedBuf
181 : // will not be used as there are no prefix compressed entries between the
182 : // restart points.
183 : cached []blockEntry
184 : cachedBuf []byte
185 : handle block.BufferHandle
186 : // for block iteration for already loaded blocks.
187 : firstUserKey []byte
188 : lazyValueHandling struct {
189 : getValue block.GetLazyValueForPrefixAndValueHandler
190 : hasValuePrefix bool
191 : }
192 : synthSuffixBuf []byte
193 : firstUserKeyWithPrefixBuf []byte
194 : }
195 :
196 : type blockEntry struct {
197 : offset int32
198 : keyStart int32
199 : keyEnd int32
200 : valStart int32
201 : valSize int32
202 : }
203 :
204 : // *Iter implements the block.DataBlockIterator interface.
205 : var _ block.DataBlockIterator = (*Iter)(nil)
206 :
207 : // NewIter constructs a new row-oriented block iterator over the provided serialized block.
208 : func NewIter(
209 : cmp base.Compare, split base.Split, block []byte, transforms block.IterTransforms,
210 0 : ) (*Iter, error) {
211 0 : i := &Iter{}
212 0 : return i, i.Init(cmp, split, block, transforms)
213 0 : }
214 :
215 : // String implements fmt.Stringer.
216 0 : func (i *Iter) String() string {
217 0 : return "block"
218 0 : }
219 :
220 : // Init initializes the block iterator from the provided block.
221 : func (i *Iter) Init(
222 : cmp base.Compare, split base.Split, blk []byte, transforms block.IterTransforms,
223 1 : ) error {
224 1 : numRestarts := int32(binary.LittleEndian.Uint32(blk[len(blk)-4:]))
225 1 : if numRestarts == 0 {
226 0 : return base.CorruptionErrorf("pebble/table: invalid table (block has no restart points)")
227 0 : }
228 1 : i.transforms = transforms
229 1 : i.synthSuffixBuf = i.synthSuffixBuf[:0]
230 1 : i.split = split
231 1 : i.cmp = cmp
232 1 : i.restarts = int32(len(blk)) - 4*(1+numRestarts)
233 1 : i.numRestarts = numRestarts
234 1 : i.ptr = unsafe.Pointer(&blk[0])
235 1 : i.data = blk
236 1 : if i.transforms.HasSyntheticPrefix() {
237 1 : i.fullKey = append(i.fullKey[:0], i.transforms.SyntheticPrefix()...)
238 1 : } else {
239 1 : i.fullKey = i.fullKey[:0]
240 1 : }
241 1 : i.val = nil
242 1 : i.clearCache()
243 1 : if i.restarts > 0 {
244 1 : if err := i.readFirstKey(); err != nil {
245 0 : return err
246 0 : }
247 1 : } else {
248 1 : // Block is empty.
249 1 : i.firstUserKey = nil
250 1 : }
251 1 : return nil
252 : }
253 :
254 : // InitHandle initializes an iterator from the provided block handle.
255 : // NB: two cases of hideObsoletePoints:
256 : // - Local sstable iteration: syntheticSeqNum will be set iff the sstable was
257 : // ingested.
258 : // - Foreign sstable iteration: syntheticSeqNum is always set.
259 : func (i *Iter) InitHandle(
260 : cmp base.Compare, split base.Split, block block.BufferHandle, transforms block.IterTransforms,
261 1 : ) error {
262 1 : i.handle.Release()
263 1 : i.handle = block
264 1 : return i.Init(cmp, split, block.BlockData(), transforms)
265 1 : }
266 :
267 : // SetHasValuePrefix sets whether or not the block iterator should expect values
268 : // corresponding to Set keys to have a prefix byte.
269 1 : func (i *Iter) SetHasValuePrefix(hasValuePrefix bool) {
270 1 : i.lazyValueHandling.hasValuePrefix = hasValuePrefix
271 1 : }
272 :
273 : // SetGetLazyValuer sets the value block reader the iterator should use to get
274 : // lazy values when the value encodes a value prefix.
275 1 : func (i *Iter) SetGetLazyValuer(g block.GetLazyValueForPrefixAndValueHandler) {
276 1 : i.lazyValueHandling.getValue = g
277 1 :
278 1 : }
279 :
280 : // Handle returns the underlying block buffer handle, if the iterator was
281 : // initialized with one.
282 1 : func (i *Iter) Handle() block.BufferHandle {
283 1 : return i.handle
284 1 : }
285 :
286 : // Invalidate invalidates the block iterator, removing references to the block
287 : // it was initialized with.
288 1 : func (i *Iter) Invalidate() {
289 1 : i.clearCache()
290 1 : i.offset = 0
291 1 : i.nextOffset = 0
292 1 : i.restarts = 0
293 1 : i.numRestarts = 0
294 1 : i.data = nil
295 1 : }
296 :
297 : // IsDataInvalidated returns true when the blockIter has been invalidated
298 : // using an invalidate call. NB: this is different from blockIter.Valid
299 : // which is part of the InternalIterator implementation.
300 1 : func (i *Iter) IsDataInvalidated() bool {
301 1 : return i.data == nil
302 1 : }
303 :
304 1 : func (i *Iter) readEntry() {
305 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
306 1 :
307 1 : // This is an ugly performance hack. Reading entries from blocks is one of
308 1 : // the inner-most routines and decoding the 3 varints per-entry takes
309 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
310 1 : // us, so we do it manually. This provides a 10-15% performance improvement
311 1 : // on blockIter benchmarks on both go1.11 and go1.12.
312 1 : //
313 1 : // TODO(peter): remove this hack if go:inline is ever supported.
314 1 :
315 1 : var shared uint32
316 1 : if a := *((*uint8)(ptr)); a < 128 {
317 1 : shared = uint32(a)
318 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
319 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
320 0 : shared = uint32(b)<<7 | uint32(a)
321 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
322 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
323 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
324 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
325 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
326 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
327 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
328 0 : } else {
329 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
330 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
331 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
332 0 : }
333 :
334 1 : var unshared uint32
335 1 : if a := *((*uint8)(ptr)); a < 128 {
336 1 : unshared = uint32(a)
337 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
338 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
339 0 : unshared = uint32(b)<<7 | uint32(a)
340 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
341 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
342 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
343 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
344 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
345 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
346 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
347 0 : } else {
348 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
349 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
350 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
351 0 : }
352 :
353 1 : var value uint32
354 1 : if a := *((*uint8)(ptr)); a < 128 {
355 1 : value = uint32(a)
356 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
357 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
358 1 : value = uint32(b)<<7 | uint32(a)
359 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
360 1 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
361 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
362 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
363 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
364 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
365 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
366 0 : } else {
367 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
368 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
369 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
370 0 : }
371 1 : shared += i.transforms.SyntheticPrefixAndSuffix.PrefixLen()
372 1 : unsharedKey := getBytes(ptr, int(unshared))
373 1 : // TODO(sumeer): move this into the else block below.
374 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
375 1 : if shared == 0 {
376 1 : // Provide stability for the key across positioning calls if the key
377 1 : // doesn't share a prefix with the previous key. This removes requiring the
378 1 : // key to be copied if the caller knows the block has a restart interval of
379 1 : // 1. An important example of this is range-del blocks.
380 1 : i.key = unsharedKey
381 1 : } else {
382 1 : i.key = i.fullKey
383 1 : }
384 1 : ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
385 1 : i.val = getBytes(ptr, int(value))
386 1 : i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
387 : }
388 :
389 1 : func (i *Iter) readFirstKey() error {
390 1 : ptr := i.ptr
391 1 :
392 1 : // This is an ugly performance hack. Reading entries from blocks is one of
393 1 : // the inner-most routines and decoding the 3 varints per-entry takes
394 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
395 1 : // us, so we do it manually. This provides a 10-15% performance improvement
396 1 : // on blockIter benchmarks on both go1.11 and go1.12.
397 1 : //
398 1 : // TODO(peter): remove this hack if go:inline is ever supported.
399 1 :
400 1 : if shared := *((*uint8)(ptr)); shared == 0 {
401 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
402 1 : } else {
403 0 : // The shared length is != 0, which is invalid.
404 0 : panic("first key in block must have zero shared length")
405 : }
406 :
407 1 : var unshared uint32
408 1 : if a := *((*uint8)(ptr)); a < 128 {
409 1 : unshared = uint32(a)
410 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
411 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
412 0 : unshared = uint32(b)<<7 | uint32(a)
413 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
414 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
415 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
416 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
417 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
418 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
419 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
420 0 : } else {
421 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
422 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
423 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
424 0 : }
425 :
426 : // Skip the value length.
427 1 : if a := *((*uint8)(ptr)); a < 128 {
428 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
429 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); a < 128 {
430 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
431 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); a < 128 {
432 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
433 0 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); a < 128 {
434 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
435 0 : } else {
436 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
437 0 : }
438 :
439 1 : firstKey := getBytes(ptr, int(unshared))
440 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
441 1 : // BlockIter benchmarks.
442 1 : if n := len(firstKey) - 8; n >= 0 {
443 1 : i.firstUserKey = firstKey[:n:n]
444 1 : } else {
445 0 : i.firstUserKey = nil
446 0 : return base.CorruptionErrorf("pebble/table: invalid firstKey in block")
447 0 : }
448 1 : if i.transforms.HasSyntheticPrefix() {
449 1 : syntheticPrefix := i.transforms.SyntheticPrefix()
450 1 : i.firstUserKeyWithPrefixBuf = slices.Grow(i.firstUserKeyWithPrefixBuf[:0], len(syntheticPrefix)+len(i.firstUserKey))
451 1 : i.firstUserKeyWithPrefixBuf = append(i.firstUserKeyWithPrefixBuf, syntheticPrefix...)
452 1 : i.firstUserKeyWithPrefixBuf = append(i.firstUserKeyWithPrefixBuf, i.firstUserKey...)
453 1 : i.firstUserKey = i.firstUserKeyWithPrefixBuf
454 1 : }
455 1 : return nil
456 : }
457 :
458 1 : func (i *Iter) decodeInternalKey(key []byte) (hiddenPoint bool) {
459 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
460 1 : // BlockIter benchmarks.
461 1 : if n := len(key) - 8; n >= 0 {
462 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(key[n:]))
463 1 : hiddenPoint = i.transforms.HideObsoletePoints &&
464 1 : (trailer&TrailerObsoleteBit != 0)
465 1 : i.ikv.K.Trailer = trailer & TrailerObsoleteMask
466 1 : i.ikv.K.UserKey = key[:n:n]
467 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
468 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
469 1 : }
470 1 : } else {
471 1 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
472 1 : i.ikv.K.UserKey = nil
473 1 : }
474 1 : return hiddenPoint
475 : }
476 :
477 : // maybeReplaceSuffix replaces the suffix in i.ikey.UserKey with
478 : // i.transforms.syntheticSuffix.
479 1 : func (i *Iter) maybeReplaceSuffix() {
480 1 : if i.transforms.HasSyntheticSuffix() && i.ikv.K.UserKey != nil {
481 1 : prefixLen := i.split(i.ikv.K.UserKey)
482 1 : // If ikey is cached or may get cached, we must copy
483 1 : // UserKey to a new buffer before suffix replacement.
484 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
485 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix()...)
486 1 : i.ikv.K.UserKey = i.synthSuffixBuf
487 1 : }
488 : }
489 :
490 1 : func (i *Iter) clearCache() {
491 1 : i.cached = i.cached[:0]
492 1 : i.cachedBuf = i.cachedBuf[:0]
493 1 : }
494 :
495 1 : func (i *Iter) cacheEntry() {
496 1 : var valStart int32
497 1 : valSize := int32(len(i.val))
498 1 : if valSize > 0 {
499 1 : valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
500 1 : }
501 :
502 1 : i.cached = append(i.cached, blockEntry{
503 1 : offset: i.offset,
504 1 : keyStart: int32(len(i.cachedBuf)),
505 1 : keyEnd: int32(len(i.cachedBuf) + len(i.key)),
506 1 : valStart: valStart,
507 1 : valSize: valSize,
508 1 : })
509 1 : i.cachedBuf = append(i.cachedBuf, i.key...)
510 : }
511 :
512 : // IsLowerBound implements the block.DataBlockIterator interface.
513 1 : func (i *Iter) IsLowerBound(k []byte) bool {
514 1 : // Note: we ignore HideObsoletePoints, but false negatives are allowed.
515 1 : return i.cmp(i.firstUserKey, k) >= 0
516 1 : }
517 :
518 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
519 : // package.
520 1 : func (i *Iter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
521 1 : if invariants.Enabled && i.IsDataInvalidated() {
522 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
523 : }
524 1 : searchKey := key
525 1 : if i.transforms.HasSyntheticPrefix() {
526 1 : syntheticPrefix := i.transforms.SyntheticPrefix()
527 1 : if !bytes.HasPrefix(key, syntheticPrefix) {
528 0 : // The seek key is before or after the entire block of keys that start
529 0 : // with SyntheticPrefix. To determine which, we need to compare against a
530 0 : // valid key in the block. We use firstUserKey which has the synthetic
531 0 : // prefix.
532 0 : if i.cmp(i.firstUserKey, key) >= 0 {
533 0 : return i.First()
534 0 : }
535 : // Set the offset to the end of the block to mimic the offset of an
536 : // invalid iterator. This ensures a subsequent i.Prev() returns a valid
537 : // result.
538 0 : i.offset = i.restarts
539 0 : i.nextOffset = i.restarts
540 0 : return nil
541 : }
542 1 : searchKey = key[len(syntheticPrefix):]
543 : }
544 :
545 1 : i.clearCache()
546 1 : // Find the index of the smallest restart point whose key is > the key
547 1 : // sought; index will be numRestarts if there is no such restart point.
548 1 : i.offset = 0
549 1 : var index int32
550 1 :
551 1 : {
552 1 : // NB: manually inlined sort.Seach is ~5% faster.
553 1 : //
554 1 : // Define f(-1) == false and f(n) == true.
555 1 : // Invariant: f(index-1) == false, f(upper) == true.
556 1 : upper := i.numRestarts
557 1 : for index < upper {
558 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
559 1 : // index ≤ h < upper
560 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
561 1 : // For a restart point, there are 0 bytes shared with the previous key.
562 1 : // The varint encoding of 0 occupies 1 byte.
563 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
564 1 :
565 1 : // Decode the key at that restart point, and compare it to the key
566 1 : // sought. See the comment in readEntry for why we manually inline the
567 1 : // varint decoding.
568 1 : var v1 uint32
569 1 : if a := *((*uint8)(ptr)); a < 128 {
570 1 : v1 = uint32(a)
571 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
572 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
573 0 : v1 = uint32(b)<<7 | uint32(a)
574 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
575 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
576 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
577 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
578 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
579 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
580 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
581 0 : } else {
582 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
583 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
584 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
585 0 : }
586 :
587 1 : if *((*uint8)(ptr)) < 128 {
588 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
589 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
590 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
591 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
592 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
593 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
594 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
595 0 : } else {
596 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
597 0 : }
598 :
599 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
600 : // speedup on BlockIter benchmarks.
601 1 : s := getBytes(ptr, int(v1))
602 1 : var k []byte
603 1 : if n := len(s) - 8; n >= 0 {
604 1 : k = s[:n:n]
605 1 : }
606 : // Else k is invalid, and left as nil
607 :
608 1 : if i.cmp(searchKey, k) > 0 {
609 1 : // The search key is greater than the user key at this restart point.
610 1 : // Search beyond this restart point, since we are trying to find the
611 1 : // first restart point with a user key >= the search key.
612 1 : index = h + 1 // preserves f(i-1) == false
613 1 : } else {
614 1 : // k >= search key, so prune everything after index (since index
615 1 : // satisfies the property we are looking for).
616 1 : upper = h // preserves f(j) == true
617 1 : }
618 : }
619 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
620 : // => answer is index.
621 : }
622 :
623 : // index is the first restart point with key >= search key. Define the keys
624 : // between a restart point and the next restart point as belonging to that
625 : // restart point.
626 : //
627 : // Since keys are strictly increasing, if index > 0 then the restart point
628 : // at index-1 will be the first one that has some keys belonging to it that
629 : // could be equal to the search key. If index == 0, then all keys in this
630 : // block are larger than the key sought, and offset remains at zero.
631 1 : if index > 0 {
632 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
633 1 : }
634 1 : i.readEntry()
635 1 : hiddenPoint := i.decodeInternalKey(i.key)
636 1 :
637 1 : // Iterate from that restart point to somewhere >= the key sought.
638 1 : if !i.Valid() {
639 0 : return nil
640 0 : }
641 :
642 : // A note on seeking in a block with a suffix replacement rule: even though
643 : // the binary search above was conducted on keys without suffix replacement,
644 : // Seek will still return the correct suffix replaced key. A binary
645 : // search without suffix replacement will land on a key that is _less_ than
646 : // the key the search would have landed on if all keys were already suffix
647 : // replaced. Since Seek then conducts forward iteration to the first suffix
648 : // replaced user key that is greater than or equal to the search key, the
649 : // correct key is still returned.
650 : //
651 : // As an example, consider the following block with a restart interval of 1,
652 : // with a replacement suffix of "4":
653 : // - Pre-suffix replacement: apple@1, banana@3
654 : // - Post-suffix replacement: apple@4, banana@4
655 : //
656 : // Suppose the client seeks with apple@3. Assuming suffixes sort in reverse
657 : // chronological order (i.e. apple@1>apple@3), the binary search without
658 : // suffix replacement would return apple@1. A binary search with suffix
659 : // replacement would return banana@4. After beginning forward iteration from
660 : // either returned restart point, forward iteration would
661 : // always return the correct key, banana@4.
662 : //
663 : // Further, if the user searched with apple@0 (i.e. a suffix less than the
664 : // pre replacement suffix) or with apple@5 (a suffix larger than the post
665 : // replacement suffix), the binary search with or without suffix replacement
666 : // would land on the same key, as we assume the following:
667 : // (1) no two keys in the sst share the same prefix.
668 : // (2) pebble.Compare(replacementSuffix,originalSuffix) > 0
669 :
670 1 : i.maybeReplaceSuffix()
671 1 :
672 1 : if !hiddenPoint && i.cmp(i.ikv.K.UserKey, key) >= 0 {
673 1 : // Initialize i.lazyValue
674 1 : if !i.lazyValueHandling.hasValuePrefix ||
675 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
676 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
677 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
678 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
679 1 : } else {
680 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
681 1 : }
682 1 : return &i.ikv
683 : }
684 1 : for i.Next(); i.Valid(); i.Next() {
685 1 : if i.cmp(i.ikv.K.UserKey, key) >= 0 {
686 1 : // i.Next() has already initialized i.ikv.LazyValue.
687 1 : return &i.ikv
688 1 : }
689 : }
690 1 : return nil
691 : }
692 :
693 : // SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
694 : // pebble package.
695 0 : func (i *Iter) SeekPrefixGE(prefix, key []byte, flags base.SeekGEFlags) *base.InternalKV {
696 0 : // This should never be called as prefix iteration is handled by sstable.Iterator.
697 0 : panic("pebble: SeekPrefixGE unimplemented")
698 : }
699 :
700 : // SeekLT implements internalIterator.SeekLT, as documented in the pebble
701 : // package.
702 1 : func (i *Iter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
703 1 : if invariants.Enabled && i.IsDataInvalidated() {
704 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
705 : }
706 1 : searchKey := key
707 1 : if i.transforms.HasSyntheticPrefix() {
708 1 : syntheticPrefix := i.transforms.SyntheticPrefix()
709 1 : if !bytes.HasPrefix(key, syntheticPrefix) {
710 1 : // The seek key is before or after the entire block of keys that start
711 1 : // with SyntheticPrefix. To determine which, we need to compare against a
712 1 : // valid key in the block. We use firstUserKey which has the synthetic
713 1 : // prefix.
714 1 : if i.cmp(i.firstUserKey, key) < 0 {
715 1 : return i.Last()
716 1 : }
717 : // Set the offset to the beginning of the block to mimic an exhausted
718 : // iterator that has conducted backward interation. This ensures a
719 : // subsequent Next() call returns the first key in the block.
720 1 : i.offset = -1
721 1 : i.nextOffset = 0
722 1 : return nil
723 : }
724 1 : searchKey = key[len(syntheticPrefix):]
725 : }
726 :
727 1 : i.clearCache()
728 1 : // Find the index of the smallest restart point whose key is >= the key
729 1 : // sought; index will be numRestarts if there is no such restart point.
730 1 : i.offset = 0
731 1 : var index int32
732 1 :
733 1 : {
734 1 : // NB: manually inlined sort.Search is ~5% faster.
735 1 : //
736 1 : // Define f(-1) == false and f(n) == true.
737 1 : // Invariant: f(index-1) == false, f(upper) == true.
738 1 : upper := i.numRestarts
739 1 : for index < upper {
740 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
741 1 : // index ≤ h < upper
742 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
743 1 : // For a restart point, there are 0 bytes shared with the previous key.
744 1 : // The varint encoding of 0 occupies 1 byte.
745 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
746 1 :
747 1 : // Decode the key at that restart point, and compare it to the key
748 1 : // sought. See the comment in readEntry for why we manually inline the
749 1 : // varint decoding.
750 1 : var v1 uint32
751 1 : if a := *((*uint8)(ptr)); a < 128 {
752 1 : v1 = uint32(a)
753 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
754 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
755 0 : v1 = uint32(b)<<7 | uint32(a)
756 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
757 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
758 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
759 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
760 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
761 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
762 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
763 0 : } else {
764 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
765 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
766 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
767 0 : }
768 :
769 1 : if *((*uint8)(ptr)) < 128 {
770 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
771 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
772 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
773 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
774 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
775 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
776 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
777 0 : } else {
778 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
779 0 : }
780 :
781 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
782 : // speedup on BlockIter benchmarks.
783 1 : s := getBytes(ptr, int(v1))
784 1 : var k []byte
785 1 : if n := len(s) - 8; n >= 0 {
786 1 : k = s[:n:n]
787 1 : }
788 : // Else k is invalid, and left as nil
789 :
790 1 : if i.cmp(searchKey, k) > 0 {
791 1 : // The search key is greater than the user key at this restart point.
792 1 : // Search beyond this restart point, since we are trying to find the
793 1 : // first restart point with a user key >= the search key.
794 1 : index = h + 1 // preserves f(i-1) == false
795 1 : } else {
796 1 : // k >= search key, so prune everything after index (since index
797 1 : // satisfies the property we are looking for).
798 1 : upper = h // preserves f(j) == true
799 1 : }
800 : }
801 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
802 : // => answer is index.
803 : }
804 :
805 1 : if index == 0 {
806 1 : if i.transforms.HasSyntheticSuffix() {
807 1 : // The binary search was conducted on keys without suffix replacement,
808 1 : // implying the first key in the block may be less than the search key. To
809 1 : // double check, get the first key in the block with suffix replacement
810 1 : // and compare to the search key. Consider the following example: suppose
811 1 : // the user searches with a@3, the first key in the block is a@2 and the
812 1 : // block contains a suffix replacement rule of 4. Since a@3 sorts before
813 1 : // a@2, the binary search would return index==0. Without conducting the
814 1 : // suffix replacement, the SeekLT would incorrectly return nil. With
815 1 : // suffix replacement though, a@4 should be returned as a@4 sorts before
816 1 : // a@3.
817 1 : ikv := i.First()
818 1 : if i.cmp(ikv.K.UserKey, key) < 0 {
819 1 : return ikv
820 1 : }
821 : }
822 : // If index == 0 then all keys in this block are larger than the key
823 : // sought, so there is no match.
824 1 : i.offset = -1
825 1 : i.nextOffset = 0
826 1 : return nil
827 : }
828 :
829 : // INVARIANT: index > 0
830 :
831 : // Ignoring suffix replacement, index is the first restart point with key >=
832 : // search key. Define the keys between a restart point and the next restart
833 : // point as belonging to that restart point. Note that index could be equal to
834 : // i.numRestarts, i.e., we are past the last restart. Since keys are strictly
835 : // increasing, then the restart point at index-1 will be the first one that
836 : // has some keys belonging to it that are less than the search key.
837 : //
838 : // Next, we will search between the restart at index-1 and the restart point
839 : // at index, for the first key >= key, and then on finding it, return
840 : // i.Prev(). We need to know when we have hit the offset for index, since then
841 : // we can stop searching. targetOffset encodes that offset for index.
842 1 : targetOffset := i.restarts
843 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
844 1 : if index < i.numRestarts {
845 1 : targetOffset = decodeRestart(i.data[i.restarts+4*(index):])
846 1 :
847 1 : if i.transforms.HasSyntheticSuffix() {
848 1 : // The binary search was conducted on keys without suffix replacement,
849 1 : // implying the returned restart point (index) may be less than the search
850 1 : // key, breaking the assumption described above.
851 1 : //
852 1 : // For example: consider this block with a replacement ts of 4, and
853 1 : // restart interval of 1: - pre replacement: a@3,b@2,c@3 - post
854 1 : // replacement: a@4,b@4,c@4
855 1 : //
856 1 : // Suppose the client calls SeekLT(b@3), SeekLT must return b@4.
857 1 : //
858 1 : // If the client calls SeekLT(b@3), the binary search would return b@2,
859 1 : // the lowest key geq to b@3, pre-suffix replacement. Then, SeekLT will
860 1 : // begin forward iteration from a@3, the previous restart point, to
861 1 : // b{suffix}. The iteration stops when it encounters a key geq to the
862 1 : // search key or if it reaches the upper bound. Without suffix
863 1 : // replacement, we can assume that the upper bound of this forward
864 1 : // iteration, b{suffix}, is greater than the search key, as implied by the
865 1 : // binary search.
866 1 : //
867 1 : // If we naively hold this assumption with suffix replacement, the
868 1 : // iteration would terminate at the upper bound, b@4, call i.Prev, and
869 1 : // incorrectly return a@4. To correct for this, if the original returned
870 1 : // index is less than the search key, shift our forward iteration to begin
871 1 : // at index instead of index -1. With suffix replacement the key at index
872 1 : // is guaranteed to be the highest restart point less than the seach key
873 1 : // (i.e. the same property of index-1 for a block without suffix
874 1 : // replacement). This property holds because of the invariant that a block
875 1 : // with suffix replacement will not have two keys that share the same
876 1 : // prefix. To consider the above example, binary searching with b@3 landed
877 1 : // naively at a@3, but since b@4<b@3, we shift our forward iteration to
878 1 : // begin at b@4. We never need to shift by more than one restart point
879 1 : // (i.e. to c@4) because it's impossible for the search key to be greater
880 1 : // than the key at the next restart point in the block because that
881 1 : // key will always have a different prefix. Put another way, because no
882 1 : // key in the block shares the same prefix, naive binary search should
883 1 : // always land at most 1 restart point off the correct one.
884 1 :
885 1 : naiveOffset := i.offset
886 1 : // Shift up to the original binary search result and decode the key.
887 1 : i.offset = targetOffset
888 1 : i.readEntry()
889 1 : i.decodeInternalKey(i.key)
890 1 : i.maybeReplaceSuffix()
891 1 :
892 1 : // If the binary search point is actually less than the search key, post
893 1 : // replacement, bump the target offset.
894 1 : if i.cmp(i.ikv.K.UserKey, key) < 0 {
895 0 : i.offset = targetOffset
896 0 : if index+1 < i.numRestarts {
897 0 : // if index+1 is within the i.data bounds, use it to find the target
898 0 : // offset.
899 0 : targetOffset = decodeRestart(i.data[i.restarts+4*(index+1):])
900 0 : } else {
901 0 : targetOffset = i.restarts
902 0 : }
903 1 : } else {
904 1 : i.offset = naiveOffset
905 1 : }
906 : }
907 : }
908 :
909 : // Init nextOffset for the forward iteration below.
910 1 : i.nextOffset = i.offset
911 1 :
912 1 : for {
913 1 : i.offset = i.nextOffset
914 1 : i.readEntry()
915 1 : // When hidden keys are common, there is additional optimization possible
916 1 : // by not caching entries that are hidden (note that some calls to
917 1 : // cacheEntry don't decode the internal key before caching, but checking
918 1 : // whether a key is hidden does not require full decoding). However, we do
919 1 : // need to use the blockEntry.offset in the cache for the first entry at
920 1 : // the reset point to do the binary search when the cache is empty -- so
921 1 : // we would need to cache that first entry (though not the key) even if
922 1 : // was hidden. Our current assumption is that if there are large numbers
923 1 : // of hidden keys we will be able to skip whole blocks (using block
924 1 : // property filters) so we don't bother optimizing.
925 1 : hiddenPoint := i.decodeInternalKey(i.key)
926 1 : i.maybeReplaceSuffix()
927 1 :
928 1 : // NB: we don't use the hiddenPoint return value of decodeInternalKey
929 1 : // since we want to stop as soon as we reach a key >= ikey.UserKey, so
930 1 : // that we can reverse.
931 1 : if i.cmp(i.ikv.K.UserKey, key) >= 0 {
932 1 : // The current key is greater than or equal to our search key. Back up to
933 1 : // the previous key which was less than our search key. Note that this for
934 1 : // loop will execute at least once with this if-block not being true, so
935 1 : // the key we are backing up to is the last one this loop cached.
936 1 : return i.Prev()
937 1 : }
938 :
939 1 : if i.nextOffset >= targetOffset {
940 1 : // We've reached the end of the current restart block. Return the
941 1 : // current key if not hidden, else call Prev().
942 1 : //
943 1 : // When the restart interval is 1, the first iteration of the for loop
944 1 : // will bring us here. In that case ikey is backed by the block so we
945 1 : // get the desired key stability guarantee for the lifetime of the
946 1 : // blockIter. That is, we never cache anything and therefore never
947 1 : // return a key backed by cachedBuf.
948 1 : if hiddenPoint {
949 1 : return i.Prev()
950 1 : }
951 1 : break
952 : }
953 1 : i.cacheEntry()
954 : }
955 :
956 1 : if !i.Valid() {
957 1 : return nil
958 1 : }
959 1 : if !i.lazyValueHandling.hasValuePrefix ||
960 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
961 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
962 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
963 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
964 1 : } else {
965 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
966 1 : }
967 1 : return &i.ikv
968 : }
969 :
970 : // First implements internalIterator.First, as documented in the pebble
971 : // package.
972 1 : func (i *Iter) First() *base.InternalKV {
973 1 : if invariants.Enabled && i.IsDataInvalidated() {
974 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
975 : }
976 :
977 1 : i.offset = 0
978 1 : if !i.Valid() {
979 1 : return nil
980 1 : }
981 1 : i.clearCache()
982 1 : i.readEntry()
983 1 : hiddenPoint := i.decodeInternalKey(i.key)
984 1 : if hiddenPoint {
985 1 : return i.Next()
986 1 : }
987 1 : i.maybeReplaceSuffix()
988 1 : if !i.lazyValueHandling.hasValuePrefix ||
989 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
990 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
991 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
992 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
993 1 : } else {
994 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
995 1 : }
996 1 : return &i.ikv
997 : }
998 :
999 : const restartMaskLittleEndianHighByteWithoutSetHasSamePrefix byte = 0b0111_1111
1000 : const restartMaskLittleEndianHighByteOnlySetHasSamePrefix byte = 0b1000_0000
1001 :
1002 1 : func decodeRestart(b []byte) int32 {
1003 1 : _ = b[3] // bounds check hint to compiler; see golang.org/issue/14808
1004 1 : return int32(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 |
1005 1 : uint32(b[3]&restartMaskLittleEndianHighByteWithoutSetHasSamePrefix)<<24)
1006 1 : }
1007 :
1008 : // Last implements internalIterator.Last, as documented in the pebble package.
1009 1 : func (i *Iter) Last() *base.InternalKV {
1010 1 : if invariants.Enabled && i.IsDataInvalidated() {
1011 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
1012 : }
1013 :
1014 : // Seek forward from the last restart point.
1015 1 : i.offset = decodeRestart(i.data[i.restarts+4*(i.numRestarts-1):])
1016 1 : if !i.Valid() {
1017 1 : return nil
1018 1 : }
1019 :
1020 1 : i.readEntry()
1021 1 : i.clearCache()
1022 1 :
1023 1 : for i.nextOffset < i.restarts {
1024 1 : i.cacheEntry()
1025 1 : i.offset = i.nextOffset
1026 1 : i.readEntry()
1027 1 : }
1028 :
1029 1 : hiddenPoint := i.decodeInternalKey(i.key)
1030 1 : if hiddenPoint {
1031 1 : return i.Prev()
1032 1 : }
1033 1 : i.maybeReplaceSuffix()
1034 1 : if !i.lazyValueHandling.hasValuePrefix ||
1035 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1036 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1037 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1038 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1039 1 : } else {
1040 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1041 1 : }
1042 1 : return &i.ikv
1043 : }
1044 :
1045 : // Next implements internalIterator.Next, as documented in the pebble
1046 : // package.
1047 1 : func (i *Iter) Next() *base.InternalKV {
1048 1 : if len(i.cachedBuf) > 0 {
1049 1 : // We're switching from reverse iteration to forward iteration. We need to
1050 1 : // populate i.fullKey with the current key we're positioned at so that
1051 1 : // readEntry() can use i.fullKey for key prefix decompression. Note that we
1052 1 : // don't know whether i.key is backed by i.cachedBuf or i.fullKey (if
1053 1 : // SeekLT was the previous call, i.key may be backed by i.fullKey), but
1054 1 : // copying into i.fullKey works for both cases.
1055 1 : //
1056 1 : // TODO(peter): Rather than clearing the cache, we could instead use the
1057 1 : // cache until it is exhausted. This would likely be faster than falling
1058 1 : // through to the normal forward iteration code below.
1059 1 : i.fullKey = append(i.fullKey[:0], i.key...)
1060 1 : i.clearCache()
1061 1 : }
1062 :
1063 : start:
1064 1 : i.offset = i.nextOffset
1065 1 : if !i.Valid() {
1066 1 : return nil
1067 1 : }
1068 1 : i.readEntry()
1069 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1070 1 : if n := len(i.key) - 8; n >= 0 {
1071 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1072 1 : hiddenPoint := i.transforms.HideObsoletePoints &&
1073 1 : (trailer&TrailerObsoleteBit != 0)
1074 1 : i.ikv.K.Trailer = trailer & TrailerObsoleteMask
1075 1 : i.ikv.K.UserKey = i.key[:n:n]
1076 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1077 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1078 1 : }
1079 1 : if hiddenPoint {
1080 1 : goto start
1081 : }
1082 1 : if i.transforms.HasSyntheticSuffix() {
1083 1 : // Inlined version of i.maybeReplaceSuffix()
1084 1 : prefixLen := i.split(i.ikv.K.UserKey)
1085 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1086 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix()...)
1087 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1088 1 : }
1089 0 : } else {
1090 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1091 0 : i.ikv.K.UserKey = nil
1092 0 : }
1093 1 : if !i.lazyValueHandling.hasValuePrefix ||
1094 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1095 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1096 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1097 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1098 1 : } else {
1099 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1100 1 : }
1101 1 : return &i.ikv
1102 : }
1103 :
1104 : // NextPrefix implements (base.InternalIterator).NextPrefix.
1105 1 : func (i *Iter) NextPrefix(succKey []byte) *base.InternalKV {
1106 1 : if i.lazyValueHandling.hasValuePrefix {
1107 1 : return i.nextPrefixV3(succKey)
1108 1 : }
1109 1 : const nextsBeforeSeek = 3
1110 1 : kv := i.Next()
1111 1 : for j := 1; kv != nil && i.cmp(kv.K.UserKey, succKey) < 0; j++ {
1112 0 : if j >= nextsBeforeSeek {
1113 0 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1114 0 : }
1115 0 : kv = i.Next()
1116 : }
1117 1 : return kv
1118 : }
1119 :
1120 1 : func (i *Iter) nextPrefixV3(succKey []byte) *base.InternalKV {
1121 1 : // Doing nexts that involve a key comparison can be expensive (and the cost
1122 1 : // depends on the key length), so we use the same threshold of 3 that we use
1123 1 : // for TableFormatPebblev2 in blockIter.nextPrefix above. The next fast path
1124 1 : // that looks at setHasSamePrefix takes ~5ns per key, which is ~150x faster
1125 1 : // than doing a SeekGE within the block, so we do this 16 times
1126 1 : // (~5ns*16=80ns), and then switch to looking at restarts. Doing the binary
1127 1 : // search for the restart consumes > 100ns. If the number of versions is >
1128 1 : // 17, we will increment nextFastCount to 17, then do a binary search, and
1129 1 : // on average need to find a key between two restarts, so another 8 steps
1130 1 : // corresponding to nextFastCount, for a mean total of 17 + 8 = 25 such
1131 1 : // steps.
1132 1 : //
1133 1 : // TODO(sumeer): use the configured restartInterval for the sstable when it
1134 1 : // was written (which we don't currently store) instead of the default value
1135 1 : // of 16.
1136 1 : const nextCmpThresholdBeforeSeek = 3
1137 1 : const nextFastThresholdBeforeRestarts = 16
1138 1 : nextCmpCount := 0
1139 1 : nextFastCount := 0
1140 1 : usedRestarts := false
1141 1 : // INVARIANT: blockIter is valid.
1142 1 : if invariants.Enabled && !i.Valid() {
1143 0 : panic(errors.AssertionFailedf("nextPrefixV3 called on invalid blockIter"))
1144 : }
1145 1 : prevKeyIsSet := i.ikv.Kind() == base.InternalKeyKindSet
1146 1 : for {
1147 1 : i.offset = i.nextOffset
1148 1 : if !i.Valid() {
1149 1 : return nil
1150 1 : }
1151 : // Need to decode the length integers, so we can compute nextOffset.
1152 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1153 1 : // This is an ugly performance hack. Reading entries from blocks is one of
1154 1 : // the inner-most routines and decoding the 3 varints per-entry takes
1155 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
1156 1 : // us, so we do it manually. This provides a 10-15% performance improvement
1157 1 : // on blockIter benchmarks on both go1.11 and go1.12.
1158 1 : //
1159 1 : // TODO(peter): remove this hack if go:inline is ever supported.
1160 1 :
1161 1 : // Decode the shared key length integer.
1162 1 : var shared uint32
1163 1 : if a := *((*uint8)(ptr)); a < 128 {
1164 1 : shared = uint32(a)
1165 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1166 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1167 0 : shared = uint32(b)<<7 | uint32(a)
1168 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1169 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1170 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1171 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1172 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1173 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1174 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1175 0 : } else {
1176 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1177 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1178 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1179 0 : }
1180 : // Decode the unshared key length integer.
1181 1 : var unshared uint32
1182 1 : if a := *((*uint8)(ptr)); a < 128 {
1183 1 : unshared = uint32(a)
1184 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1185 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1186 0 : unshared = uint32(b)<<7 | uint32(a)
1187 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1188 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1189 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1190 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1191 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1192 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1193 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1194 0 : } else {
1195 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1196 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1197 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1198 0 : }
1199 : // Decode the value length integer.
1200 1 : var value uint32
1201 1 : if a := *((*uint8)(ptr)); a < 128 {
1202 1 : value = uint32(a)
1203 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1204 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1205 0 : value = uint32(b)<<7 | uint32(a)
1206 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1207 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1208 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1209 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1210 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1211 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1212 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1213 0 : } else {
1214 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1215 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1216 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1217 0 : }
1218 1 : shared += i.transforms.SyntheticPrefixAndSuffix.PrefixLen()
1219 1 : // The starting position of the value.
1220 1 : valuePtr := unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
1221 1 : i.nextOffset = int32(uintptr(valuePtr)-uintptr(i.ptr)) + int32(value)
1222 1 : if invariants.Enabled && unshared < 8 {
1223 0 : // This should not happen since only the key prefix is shared, so even
1224 0 : // if the prefix length is the same as the user key length, the unshared
1225 0 : // will include the trailer.
1226 0 : panic(errors.AssertionFailedf("unshared %d is too small", unshared))
1227 : }
1228 : // The trailer is written in little endian, so the key kind is the first
1229 : // byte in the trailer that is encoded in the slice [unshared-8:unshared].
1230 1 : keyKind := base.InternalKeyKind((*[manual.MaxArrayLen]byte)(ptr)[unshared-8])
1231 1 : keyKind = keyKind & base.InternalKeyKindSSTableInternalObsoleteMask
1232 1 : prefixChanged := false
1233 1 : if keyKind == base.InternalKeyKindSet {
1234 1 : if invariants.Enabled && value == 0 {
1235 0 : panic(errors.AssertionFailedf("value is of length 0, but we expect a valuePrefix"))
1236 : }
1237 1 : valPrefix := *((*block.ValuePrefix)(valuePtr))
1238 1 : if valPrefix.SetHasSamePrefix() {
1239 1 : // Fast-path. No need to assemble i.fullKey, or update i.key. We know
1240 1 : // that subsequent keys will not have a shared length that is greater
1241 1 : // than the prefix of the current key, which is also the prefix of
1242 1 : // i.key. Since we are continuing to iterate, we don't need to
1243 1 : // initialize i.ikey and i.lazyValue (these are initialized before
1244 1 : // returning).
1245 1 : nextFastCount++
1246 1 : if nextFastCount > nextFastThresholdBeforeRestarts {
1247 0 : if usedRestarts {
1248 0 : // Exhausted iteration budget. This will never happen unless
1249 0 : // someone is using a restart interval > 16. It is just to guard
1250 0 : // against long restart intervals causing too much iteration.
1251 0 : break
1252 : }
1253 : // Haven't used restarts yet, so find the first restart at or beyond
1254 : // the current offset.
1255 0 : targetOffset := i.offset
1256 0 : var index int32
1257 0 : {
1258 0 : // NB: manually inlined sort.Sort is ~5% faster.
1259 0 : //
1260 0 : // f defined for a restart point is true iff the offset >=
1261 0 : // targetOffset.
1262 0 : // Define f(-1) == false and f(i.numRestarts) == true.
1263 0 : // Invariant: f(index-1) == false, f(upper) == true.
1264 0 : upper := i.numRestarts
1265 0 : for index < upper {
1266 0 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1267 0 : // index ≤ h < upper
1268 0 : offset := decodeRestart(i.data[i.restarts+4*h:])
1269 0 : if offset < targetOffset {
1270 0 : index = h + 1 // preserves f(index-1) == false
1271 0 : } else {
1272 0 : upper = h // preserves f(upper) == true
1273 0 : }
1274 : }
1275 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1276 : // => answer is index.
1277 : }
1278 0 : usedRestarts = true
1279 0 : nextFastCount = 0
1280 0 : if index == i.numRestarts {
1281 0 : // Already past the last real restart, so iterate a bit more until
1282 0 : // we are done with the block.
1283 0 : continue
1284 : }
1285 : // Have some real restarts after index. NB: index is the first
1286 : // restart at or beyond the current offset.
1287 0 : startingIndex := index
1288 0 : for index != i.numRestarts &&
1289 0 : // The restart at index is 4 bytes written in little endian format
1290 0 : // starting at i.restart+4*index. The 0th byte is the least
1291 0 : // significant and the 3rd byte is the most significant. Since the
1292 0 : // most significant bit of the 3rd byte is what we use for
1293 0 : // encoding the set-has-same-prefix information, the indexing
1294 0 : // below has +3.
1295 0 : i.data[i.restarts+4*index+3]&restartMaskLittleEndianHighByteOnlySetHasSamePrefix != 0 {
1296 0 : // We still have the same prefix, so move to the next restart.
1297 0 : index++
1298 0 : }
1299 : // index is the first restart that did not have the same prefix.
1300 0 : if index != startingIndex {
1301 0 : // Managed to skip past at least one restart. Resume iteration
1302 0 : // from index-1. Since nextFastCount has been reset to 0, we
1303 0 : // should be able to iterate to the next prefix.
1304 0 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1305 0 : i.readEntry()
1306 0 : }
1307 : // Else, unable to skip past any restart. Resume iteration. Since
1308 : // nextFastCount has been reset to 0, we should be able to iterate
1309 : // to the next prefix.
1310 0 : continue
1311 : }
1312 1 : continue
1313 1 : } else if prevKeyIsSet {
1314 1 : prefixChanged = true
1315 1 : }
1316 1 : } else {
1317 1 : prevKeyIsSet = false
1318 1 : }
1319 : // Slow-path cases:
1320 : // - (Likely) The prefix has changed.
1321 : // - (Unlikely) The prefix has not changed.
1322 : // We assemble the key etc. under the assumption that it is the likely
1323 : // case.
1324 1 : unsharedKey := getBytes(ptr, int(unshared))
1325 1 : // TODO(sumeer): move this into the else block below. This is a bit tricky
1326 1 : // since the current logic assumes we have always copied the latest key
1327 1 : // into fullKey, which is why when we get to the next key we can (a)
1328 1 : // access i.fullKey[:shared], (b) append only the unsharedKey to
1329 1 : // i.fullKey. For (a), we can access i.key[:shared] since that memory is
1330 1 : // valid (even if unshared). For (b), we will need to remember whether
1331 1 : // i.key refers to i.fullKey or not, and can append the unsharedKey only
1332 1 : // in the former case and for the latter case need to copy the shared part
1333 1 : // too. This same comment applies to the other place where we can do this
1334 1 : // optimization, in readEntry().
1335 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
1336 1 : i.val = getBytes(valuePtr, int(value))
1337 1 : if shared == 0 {
1338 1 : // Provide stability for the key across positioning calls if the key
1339 1 : // doesn't share a prefix with the previous key. This removes requiring the
1340 1 : // key to be copied if the caller knows the block has a restart interval of
1341 1 : // 1. An important example of this is range-del blocks.
1342 1 : i.key = unsharedKey
1343 1 : } else {
1344 1 : i.key = i.fullKey
1345 1 : }
1346 : // Manually inlined version of i.decodeInternalKey(i.key).
1347 1 : hiddenPoint := false
1348 1 : if n := len(i.key) - 8; n >= 0 {
1349 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1350 1 : hiddenPoint = i.transforms.HideObsoletePoints &&
1351 1 : (trailer&TrailerObsoleteBit != 0)
1352 1 : i.ikv.K = base.InternalKey{
1353 1 : Trailer: trailer & TrailerObsoleteMask,
1354 1 : UserKey: i.key[:n:n],
1355 1 : }
1356 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1357 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1358 1 : }
1359 1 : if i.transforms.HasSyntheticSuffix() {
1360 0 : // Inlined version of i.maybeReplaceSuffix()
1361 0 : prefixLen := i.split(i.ikv.K.UserKey)
1362 0 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1363 0 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix()...)
1364 0 : i.ikv.K.UserKey = i.synthSuffixBuf
1365 0 : }
1366 0 : } else {
1367 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1368 0 : i.ikv.K.UserKey = nil
1369 0 : }
1370 1 : nextCmpCount++
1371 1 : if invariants.Enabled && prefixChanged && i.cmp(i.ikv.K.UserKey, succKey) < 0 {
1372 0 : panic(errors.AssertionFailedf("prefix should have changed but %x < %x",
1373 0 : i.ikv.K.UserKey, succKey))
1374 : }
1375 1 : if prefixChanged || i.cmp(i.ikv.K.UserKey, succKey) >= 0 {
1376 1 : // Prefix has changed.
1377 1 : if hiddenPoint {
1378 1 : return i.Next()
1379 1 : }
1380 1 : if invariants.Enabled && !i.lazyValueHandling.hasValuePrefix {
1381 0 : panic(errors.AssertionFailedf("nextPrefixV3 being run for non-v3 sstable"))
1382 : }
1383 1 : if i.ikv.K.Kind() != base.InternalKeyKindSet {
1384 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1385 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1386 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1387 1 : } else {
1388 0 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1389 0 : }
1390 1 : return &i.ikv
1391 : }
1392 : // Else prefix has not changed.
1393 :
1394 1 : if nextCmpCount >= nextCmpThresholdBeforeSeek {
1395 0 : break
1396 : }
1397 : }
1398 0 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1399 : }
1400 :
1401 : // Prev implements internalIterator.Prev, as documented in the pebble
1402 : // package.
1403 1 : func (i *Iter) Prev() *base.InternalKV {
1404 1 : start:
1405 1 : for n := len(i.cached) - 1; n >= 0; n-- {
1406 1 : i.nextOffset = i.offset
1407 1 : e := &i.cached[n]
1408 1 : i.offset = e.offset
1409 1 : i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
1410 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1411 1 : i.key = i.cachedBuf[e.keyStart:e.keyEnd]
1412 1 : if n := len(i.key) - 8; n >= 0 {
1413 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1414 1 : hiddenPoint := i.transforms.HideObsoletePoints &&
1415 1 : (trailer&TrailerObsoleteBit != 0)
1416 1 : if hiddenPoint {
1417 1 : continue
1418 : }
1419 1 : i.ikv.K = base.InternalKey{
1420 1 : Trailer: trailer & TrailerObsoleteMask,
1421 1 : UserKey: i.key[:n:n],
1422 1 : }
1423 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1424 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1425 1 : }
1426 1 : if i.transforms.HasSyntheticSuffix() {
1427 1 : // Inlined version of i.maybeReplaceSuffix()
1428 1 : prefixLen := i.split(i.ikv.K.UserKey)
1429 1 : // If ikey is cached or may get cached, we must de-reference
1430 1 : // UserKey before suffix replacement.
1431 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1432 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix()...)
1433 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1434 1 : }
1435 0 : } else {
1436 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1437 0 : i.ikv.K.UserKey = nil
1438 0 : }
1439 1 : i.cached = i.cached[:n]
1440 1 : if !i.lazyValueHandling.hasValuePrefix ||
1441 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1442 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1443 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1444 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1445 1 : } else {
1446 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1447 1 : }
1448 1 : return &i.ikv
1449 : }
1450 :
1451 1 : i.clearCache()
1452 1 : if i.offset <= 0 {
1453 1 : i.offset = -1
1454 1 : i.nextOffset = 0
1455 1 : return nil
1456 1 : }
1457 :
1458 1 : targetOffset := i.offset
1459 1 : var index int32
1460 1 :
1461 1 : {
1462 1 : // NB: manually inlined sort.Sort is ~5% faster.
1463 1 : //
1464 1 : // Define f(-1) == false and f(n) == true.
1465 1 : // Invariant: f(index-1) == false, f(upper) == true.
1466 1 : upper := i.numRestarts
1467 1 : for index < upper {
1468 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1469 1 : // index ≤ h < upper
1470 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
1471 1 : if offset < targetOffset {
1472 1 : // Looking for the first restart that has offset >= targetOffset, so
1473 1 : // ignore h and earlier.
1474 1 : index = h + 1 // preserves f(i-1) == false
1475 1 : } else {
1476 1 : upper = h // preserves f(j) == true
1477 1 : }
1478 : }
1479 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1480 : // => answer is index.
1481 : }
1482 :
1483 : // index is first restart with offset >= targetOffset. Note that
1484 : // targetOffset may not be at a restart point since one can call Prev()
1485 : // after Next() (so the cache was not populated) and targetOffset refers to
1486 : // the current entry. index-1 must have an offset < targetOffset (it can't
1487 : // be equal to targetOffset since the binary search would have selected that
1488 : // as the index).
1489 1 : i.offset = 0
1490 1 : if index > 0 {
1491 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1492 1 : }
1493 : // TODO(sumeer): why is the else case not an error given targetOffset is a
1494 : // valid offset.
1495 :
1496 1 : i.readEntry()
1497 1 :
1498 1 : // We stop when i.nextOffset == targetOffset since the targetOffset is the
1499 1 : // entry we are stepping back from, and we don't need to cache the entry
1500 1 : // before it, since it is the candidate to return.
1501 1 : for i.nextOffset < targetOffset {
1502 1 : i.cacheEntry()
1503 1 : i.offset = i.nextOffset
1504 1 : i.readEntry()
1505 1 : }
1506 :
1507 1 : hiddenPoint := i.decodeInternalKey(i.key)
1508 1 : if hiddenPoint {
1509 1 : // Use the cache.
1510 1 : goto start
1511 : }
1512 1 : if i.transforms.HasSyntheticSuffix() {
1513 1 : // Inlined version of i.maybeReplaceSuffix()
1514 1 : prefixLen := i.split(i.ikv.K.UserKey)
1515 1 : // If ikey is cached or may get cached, we must de-reference
1516 1 : // UserKey before suffix replacement.
1517 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1518 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix()...)
1519 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1520 1 : }
1521 1 : if !i.lazyValueHandling.hasValuePrefix ||
1522 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1523 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1524 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1525 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1526 1 : } else {
1527 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1528 1 : }
1529 1 : return &i.ikv
1530 : }
1531 :
1532 : // Key returns the internal key at the current iterator position.
1533 0 : func (i *Iter) Key() *base.InternalKey {
1534 0 : return &i.ikv.K
1535 0 : }
1536 :
1537 : // KV returns the internal KV at the current iterator position.
1538 1 : func (i *Iter) KV() *base.InternalKV {
1539 1 : return &i.ikv
1540 1 : }
1541 :
1542 : // Value returns the value at the current iterator position.
1543 0 : func (i *Iter) Value() base.LazyValue {
1544 0 : return i.ikv.V
1545 0 : }
1546 :
1547 : // Error implements internalIterator.Error, as documented in the pebble
1548 : // package.
1549 1 : func (i *Iter) Error() error {
1550 1 : return nil // infallible
1551 1 : }
1552 :
1553 : // Close implements internalIterator.Close, as documented in the pebble
1554 : // package.
1555 1 : func (i *Iter) Close() error {
1556 1 : i.handle.Release()
1557 1 : fullKey := i.fullKey[:0]
1558 1 : cached := i.cached[:0]
1559 1 : cachedBuf := i.cachedBuf[:0]
1560 1 : firstUserKeyWithPrefixBuf := i.firstUserKeyWithPrefixBuf[:0]
1561 1 : *i = Iter{
1562 1 : fullKey: fullKey,
1563 1 : cached: cached,
1564 1 : cachedBuf: cachedBuf,
1565 1 : firstUserKeyWithPrefixBuf: firstUserKeyWithPrefixBuf,
1566 1 : }
1567 1 : return nil
1568 1 : }
1569 :
1570 : // SetBounds implements base.InternalIterator. It panics, as bounds should
1571 : // always be handled the by the parent sstable iterator.
1572 0 : func (i *Iter) SetBounds(lower, upper []byte) {
1573 0 : // This should never be called as bounds are handled by sstable.Iterator.
1574 0 : panic("pebble: SetBounds unimplemented")
1575 : }
1576 :
1577 : // SetContext implements base.InternalIterator.
1578 0 : func (i *Iter) SetContext(_ context.Context) {}
1579 :
1580 : // Valid returns true if the iterator is currently positioned at a valid KV.
1581 1 : func (i *Iter) Valid() bool {
1582 1 : return i.offset >= 0 && i.offset < i.restarts
1583 1 : }
1584 :
1585 : // DebugTree is part of the InternalIterator interface.
1586 0 : func (i *Iter) DebugTree(tp treeprinter.Node) {
1587 0 : tp.Childf("%T(%p)", i, i)
1588 0 : }
1589 :
1590 0 : func (i *Iter) getRestart(idx int) int32 {
1591 0 : return int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*int32(idx):]))
1592 0 : }
1593 :
1594 0 : func (i *Iter) isRestartPoint() bool {
1595 0 : j := sort.Search(int(i.numRestarts), func(j int) bool {
1596 0 : return i.getRestart(j) >= i.offset
1597 0 : })
1598 0 : return j < int(i.numRestarts) && i.getRestart(j) == i.offset
1599 : }
1600 :
1601 : // DescribeKV is a function that formats a key-value pair, writing the
1602 : // description to w.
1603 : type DescribeKV func(w io.Writer, key *base.InternalKey, val []byte, enc KVEncoding)
1604 :
1605 : // KVEncoding describes the encoding of a key-value pair within the block.
1606 : type KVEncoding struct {
1607 : // IsRestart is true if the key is a restart point.
1608 : IsRestart bool
1609 : // Offset is the position within the block at which the key-value pair is
1610 : // encoded.
1611 : Offset int32
1612 : // Length is the total length of the KV pair as it is encoded in the block
1613 : // format.
1614 : Length int32
1615 : // KeyShared is the number of bytes this KV's user key shared with its predecessor.
1616 : KeyShared uint32
1617 : // KeyUnshared is the number of bytes this KV's user key did not share with
1618 : // its predecessor.
1619 : KeyUnshared uint32
1620 : // ValueLen is the length of the internal value.
1621 : ValueLen uint32
1622 : }
1623 :
1624 : // Describe describes the contents of a block, writing the description to w.
1625 : // It invokes fmtKV to describe each key-value pair.
1626 0 : func (i *Iter) Describe(tp treeprinter.Node, fmtKV DescribeKV) {
1627 0 : var buf bytes.Buffer
1628 0 : for kv := i.First(); kv != nil; kv = i.Next() {
1629 0 : enc := KVEncoding{
1630 0 : IsRestart: i.isRestartPoint(),
1631 0 : Offset: i.offset,
1632 0 : Length: int32(i.nextOffset - i.offset),
1633 0 : }
1634 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1635 0 : enc.KeyShared, ptr = decodeVarint(ptr)
1636 0 : enc.KeyUnshared, ptr = decodeVarint(ptr)
1637 0 : enc.ValueLen, _ = decodeVarint(ptr)
1638 0 : buf.Reset()
1639 0 : fmtKV(&buf, &kv.K, kv.V.ValueOrHandle, enc)
1640 0 : tp.Child(buf.String())
1641 0 : }
1642 : // Format the restart points.
1643 0 : n := tp.Child("restart points")
1644 0 : // Format the restart points.
1645 0 : for j := 0; j < int(i.numRestarts); j++ {
1646 0 : offset := i.getRestart(j)
1647 0 : n.Childf("%05d [restart %d]", uint64(i.restarts+4*int32(j)), offset)
1648 0 : }
1649 : }
1650 :
1651 : // RawIter is an iterator over a single block of data. Unlike blockIter,
1652 : // keys are stored in "raw" format (i.e. not as internal keys). Note that there
1653 : // is significant similarity between this code and the code in blockIter. Yet
1654 : // reducing duplication is difficult due to the blockIter being performance
1655 : // critical. RawIter must only be used for blocks where the value is
1656 : // stored together with the key.
1657 : type RawIter struct {
1658 : cmp base.Compare
1659 : offset int32
1660 : nextOffset int32
1661 : restarts int32
1662 : numRestarts int32
1663 : ptr unsafe.Pointer
1664 : data []byte
1665 : key, val []byte
1666 : ikey base.InternalKey
1667 : cached []blockEntry
1668 : cachedBuf []byte
1669 : }
1670 :
1671 : // NewRawIter constructs a new raw block iterator.
1672 1 : func NewRawIter(cmp base.Compare, block []byte) (*RawIter, error) {
1673 1 : i := &RawIter{}
1674 1 : return i, i.Init(cmp, block)
1675 1 : }
1676 :
1677 : // Init initializes the raw block iterator.
1678 1 : func (i *RawIter) Init(cmp base.Compare, blk []byte) error {
1679 1 : numRestarts := int32(binary.LittleEndian.Uint32(blk[len(blk)-4:]))
1680 1 : if numRestarts == 0 {
1681 0 : return base.CorruptionErrorf("pebble/table: invalid table (block has no restart points)")
1682 0 : }
1683 1 : i.cmp = cmp
1684 1 : i.restarts = int32(len(blk)) - 4*(1+numRestarts)
1685 1 : i.numRestarts = numRestarts
1686 1 : i.ptr = unsafe.Pointer(&blk[0])
1687 1 : i.data = blk
1688 1 : if i.key == nil {
1689 1 : i.key = make([]byte, 0, 256)
1690 1 : } else {
1691 0 : i.key = i.key[:0]
1692 0 : }
1693 1 : i.val = nil
1694 1 : i.clearCache()
1695 1 : return nil
1696 : }
1697 :
1698 1 : func (i *RawIter) readEntry() {
1699 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1700 1 : shared, ptr := decodeVarint(ptr)
1701 1 : unshared, ptr := decodeVarint(ptr)
1702 1 : value, ptr := decodeVarint(ptr)
1703 1 : i.key = append(i.key[:shared], getBytes(ptr, int(unshared))...)
1704 1 : i.key = i.key[:len(i.key):len(i.key)]
1705 1 : ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
1706 1 : i.val = getBytes(ptr, int(value))
1707 1 : i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
1708 1 : }
1709 :
1710 1 : func (i *RawIter) loadEntry() {
1711 1 : i.readEntry()
1712 1 : i.ikey.UserKey = i.key
1713 1 : }
1714 :
1715 1 : func (i *RawIter) clearCache() {
1716 1 : i.cached = i.cached[:0]
1717 1 : i.cachedBuf = i.cachedBuf[:0]
1718 1 : }
1719 :
1720 0 : func (i *RawIter) cacheEntry() {
1721 0 : var valStart int32
1722 0 : valSize := int32(len(i.val))
1723 0 : if valSize > 0 {
1724 0 : valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
1725 0 : }
1726 :
1727 0 : i.cached = append(i.cached, blockEntry{
1728 0 : offset: i.offset,
1729 0 : keyStart: int32(len(i.cachedBuf)),
1730 0 : keyEnd: int32(len(i.cachedBuf) + len(i.key)),
1731 0 : valStart: valStart,
1732 0 : valSize: valSize,
1733 0 : })
1734 0 : i.cachedBuf = append(i.cachedBuf, i.key...)
1735 : }
1736 :
1737 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
1738 : // package.
1739 0 : func (i *RawIter) SeekGE(key []byte) bool {
1740 0 : // Find the index of the smallest restart point whose key is > the key
1741 0 : // sought; index will be numRestarts if there is no such restart point.
1742 0 : i.offset = 0
1743 0 : index := sort.Search(int(i.numRestarts), func(j int) bool {
1744 0 : offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
1745 0 : // For a restart point, there are 0 bytes shared with the previous key.
1746 0 : // The varint encoding of 0 occupies 1 byte.
1747 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
1748 0 : // Decode the key at that restart point, and compare it to the key sought.
1749 0 : v1, ptr := decodeVarint(ptr)
1750 0 : _, ptr = decodeVarint(ptr)
1751 0 : s := getBytes(ptr, int(v1))
1752 0 : return i.cmp(key, s) < 0
1753 0 : })
1754 :
1755 : // Since keys are strictly increasing, if index > 0 then the restart point at
1756 : // index-1 will be the largest whose key is <= the key sought. If index ==
1757 : // 0, then all keys in this block are larger than the key sought, and offset
1758 : // remains at zero.
1759 0 : if index > 0 {
1760 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
1761 0 : }
1762 0 : i.loadEntry()
1763 0 :
1764 0 : // Iterate from that restart point to somewhere >= the key sought.
1765 0 : for valid := i.Valid(); valid; valid = i.Next() {
1766 0 : if i.cmp(key, i.key) <= 0 {
1767 0 : break
1768 : }
1769 : }
1770 0 : return i.Valid()
1771 : }
1772 :
1773 : // First implements internalIterator.First, as documented in the pebble
1774 : // package.
1775 1 : func (i *RawIter) First() bool {
1776 1 : i.offset = 0
1777 1 : i.loadEntry()
1778 1 : return i.Valid()
1779 1 : }
1780 :
1781 : // Last implements internalIterator.Last, as documented in the pebble package.
1782 0 : func (i *RawIter) Last() bool {
1783 0 : // Seek forward from the last restart point.
1784 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*(i.numRestarts-1):]))
1785 0 :
1786 0 : i.readEntry()
1787 0 : i.clearCache()
1788 0 : i.cacheEntry()
1789 0 :
1790 0 : for i.nextOffset < i.restarts {
1791 0 : i.offset = i.nextOffset
1792 0 : i.readEntry()
1793 0 : i.cacheEntry()
1794 0 : }
1795 :
1796 0 : i.ikey.UserKey = i.key
1797 0 : return i.Valid()
1798 : }
1799 :
1800 : // Next implements internalIterator.Next, as documented in the pebble
1801 : // package.
1802 1 : func (i *RawIter) Next() bool {
1803 1 : i.offset = i.nextOffset
1804 1 : if !i.Valid() {
1805 1 : return false
1806 1 : }
1807 1 : i.loadEntry()
1808 1 : return true
1809 : }
1810 :
1811 : // Prev implements internalIterator.Prev, as documented in the pebble
1812 : // package.
1813 0 : func (i *RawIter) Prev() bool {
1814 0 : if n := len(i.cached) - 1; n > 0 && i.cached[n].offset == i.offset {
1815 0 : i.nextOffset = i.offset
1816 0 : e := &i.cached[n-1]
1817 0 : i.offset = e.offset
1818 0 : i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
1819 0 : i.ikey.UserKey = i.cachedBuf[e.keyStart:e.keyEnd]
1820 0 : i.cached = i.cached[:n]
1821 0 : return true
1822 0 : }
1823 :
1824 0 : if i.offset == 0 {
1825 0 : i.offset = -1
1826 0 : i.nextOffset = 0
1827 0 : return false
1828 0 : }
1829 :
1830 0 : targetOffset := i.offset
1831 0 : index := sort.Search(int(i.numRestarts), func(j int) bool {
1832 0 : offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
1833 0 : return offset >= targetOffset
1834 0 : })
1835 0 : i.offset = 0
1836 0 : if index > 0 {
1837 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
1838 0 : }
1839 :
1840 0 : i.readEntry()
1841 0 : i.clearCache()
1842 0 : i.cacheEntry()
1843 0 :
1844 0 : for i.nextOffset < targetOffset {
1845 0 : i.offset = i.nextOffset
1846 0 : i.readEntry()
1847 0 : i.cacheEntry()
1848 0 : }
1849 :
1850 0 : i.ikey.UserKey = i.key
1851 0 : return true
1852 : }
1853 :
1854 : // Key implements internalIterator.Key, as documented in the pebble package.
1855 1 : func (i *RawIter) Key() base.InternalKey {
1856 1 : return i.ikey
1857 1 : }
1858 :
1859 : // Value implements internalIterator.Value, as documented in the pebble
1860 : // package.
1861 1 : func (i *RawIter) Value() []byte {
1862 1 : return i.val
1863 1 : }
1864 :
1865 : // Valid implements internalIterator.Valid, as documented in the pebble
1866 : // package.
1867 1 : func (i *RawIter) Valid() bool {
1868 1 : return i.offset >= 0 && i.offset < i.restarts
1869 1 : }
1870 :
1871 : // Error implements internalIterator.Error, as documented in the pebble
1872 : // package.
1873 0 : func (i *RawIter) Error() error {
1874 0 : return nil
1875 0 : }
1876 :
1877 : // Close implements internalIterator.Close, as documented in the pebble
1878 : // package.
1879 1 : func (i *RawIter) Close() error {
1880 1 : i.val = nil
1881 1 : return nil
1882 1 : }
1883 :
1884 : // DebugTree is part of the InternalIterator interface.
1885 0 : func (i *RawIter) DebugTree(tp treeprinter.Node) {
1886 0 : tp.Childf("%T(%p)", i, i)
1887 0 : }
1888 :
1889 0 : func (i *RawIter) getRestart(idx int) int32 {
1890 0 : return int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*int32(idx):]))
1891 0 : }
1892 :
1893 0 : func (i *RawIter) isRestartPoint() bool {
1894 0 : j := sort.Search(int(i.numRestarts), func(j int) bool {
1895 0 : return i.getRestart(j) >= i.offset
1896 0 : })
1897 0 : return j < int(i.numRestarts) && i.getRestart(j) == i.offset
1898 : }
1899 :
1900 : // Describe describes the contents of a block, writing the description to w.
1901 : // It invokes fmtKV to describe each key-value pair.
1902 0 : func (i *RawIter) Describe(tp treeprinter.Node, fmtKV DescribeKV) {
1903 0 : var buf bytes.Buffer
1904 0 : for valid := i.First(); valid; valid = i.Next() {
1905 0 : enc := KVEncoding{
1906 0 : IsRestart: i.isRestartPoint(),
1907 0 : Offset: i.offset,
1908 0 : Length: int32(i.nextOffset - i.offset),
1909 0 : }
1910 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1911 0 : enc.KeyShared, ptr = decodeVarint(ptr)
1912 0 : enc.KeyUnshared, ptr = decodeVarint(ptr)
1913 0 : enc.ValueLen, _ = decodeVarint(ptr)
1914 0 : buf.Reset()
1915 0 : fmtKV(&buf, &i.ikey, i.val, enc)
1916 0 : if i.isRestartPoint() {
1917 0 : buf.WriteString(" [restart]")
1918 0 : }
1919 0 : tp.Child(buf.String())
1920 : }
1921 0 : n := tp.Child("restart points")
1922 0 : // Format the restart points.
1923 0 : for j := 0; j < int(i.numRestarts); j++ {
1924 0 : offset := i.getRestart(j)
1925 0 : n.Childf("%05d [restart %d]", uint64(i.restarts+4*int32(j)), offset)
1926 0 : }
1927 : }
1928 :
1929 1 : func getBytes(ptr unsafe.Pointer, length int) []byte {
1930 1 : return (*[manual.MaxArrayLen]byte)(ptr)[:length:length]
1931 1 : }
1932 :
1933 1 : func decodeVarint(ptr unsafe.Pointer) (uint32, unsafe.Pointer) {
1934 1 : if a := *((*uint8)(ptr)); a < 128 {
1935 1 : return uint32(a),
1936 1 : unsafe.Pointer(uintptr(ptr) + 1)
1937 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1938 0 : return uint32(b)<<7 | uint32(a),
1939 0 : unsafe.Pointer(uintptr(ptr) + 2)
1940 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1941 0 : return uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1942 0 : unsafe.Pointer(uintptr(ptr) + 3)
1943 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1944 0 : return uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1945 0 : unsafe.Pointer(uintptr(ptr) + 4)
1946 0 : } else {
1947 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1948 0 : return uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1949 0 : unsafe.Pointer(uintptr(ptr) + 5)
1950 0 : }
1951 : }
|
