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.SyntheticPrefix.IsSet() {
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 : // ResetForReuse resets the blockIter for reuse, retaining buffers to avoid
305 : // future allocations.
306 1 : func (i *Iter) ResetForReuse() {
307 1 : fullKey := i.fullKey[:0]
308 1 : cached := i.cached[:0]
309 1 : cachedBuf := i.cachedBuf[:0]
310 1 : firstUserKeyWithPrefixBuf := i.firstUserKeyWithPrefixBuf[:0]
311 1 : *i = Iter{
312 1 : fullKey: fullKey,
313 1 : cached: cached,
314 1 : cachedBuf: cachedBuf,
315 1 : firstUserKeyWithPrefixBuf: firstUserKeyWithPrefixBuf,
316 1 : }
317 1 : }
318 :
319 1 : func (i *Iter) readEntry() {
320 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
321 1 :
322 1 : // This is an ugly performance hack. Reading entries from blocks is one of
323 1 : // the inner-most routines and decoding the 3 varints per-entry takes
324 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
325 1 : // us, so we do it manually. This provides a 10-15% performance improvement
326 1 : // on blockIter benchmarks on both go1.11 and go1.12.
327 1 : //
328 1 : // TODO(peter): remove this hack if go:inline is ever supported.
329 1 :
330 1 : var shared uint32
331 1 : if a := *((*uint8)(ptr)); a < 128 {
332 1 : shared = uint32(a)
333 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
334 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
335 0 : shared = uint32(b)<<7 | uint32(a)
336 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
337 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
338 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
339 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
340 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
341 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
342 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
343 0 : } else {
344 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
345 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
346 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
347 0 : }
348 :
349 1 : var unshared uint32
350 1 : if a := *((*uint8)(ptr)); a < 128 {
351 1 : unshared = uint32(a)
352 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
353 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
354 0 : unshared = uint32(b)<<7 | uint32(a)
355 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
356 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
357 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
358 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
359 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
360 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
361 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
362 0 : } else {
363 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
364 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
365 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
366 0 : }
367 :
368 1 : var value uint32
369 1 : if a := *((*uint8)(ptr)); a < 128 {
370 1 : value = uint32(a)
371 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
372 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
373 1 : value = uint32(b)<<7 | uint32(a)
374 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
375 1 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
376 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
377 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
378 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
379 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
380 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
381 0 : } else {
382 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
383 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
384 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
385 0 : }
386 1 : shared += uint32(len(i.transforms.SyntheticPrefix))
387 1 : unsharedKey := getBytes(ptr, int(unshared))
388 1 : // TODO(sumeer): move this into the else block below.
389 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
390 1 : if shared == 0 {
391 1 : // Provide stability for the key across positioning calls if the key
392 1 : // doesn't share a prefix with the previous key. This removes requiring the
393 1 : // key to be copied if the caller knows the block has a restart interval of
394 1 : // 1. An important example of this is range-del blocks.
395 1 : i.key = unsharedKey
396 1 : } else {
397 1 : i.key = i.fullKey
398 1 : }
399 1 : ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
400 1 : i.val = getBytes(ptr, int(value))
401 1 : i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
402 : }
403 :
404 1 : func (i *Iter) readFirstKey() error {
405 1 : ptr := i.ptr
406 1 :
407 1 : // This is an ugly performance hack. Reading entries from blocks is one of
408 1 : // the inner-most routines and decoding the 3 varints per-entry takes
409 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
410 1 : // us, so we do it manually. This provides a 10-15% performance improvement
411 1 : // on blockIter benchmarks on both go1.11 and go1.12.
412 1 : //
413 1 : // TODO(peter): remove this hack if go:inline is ever supported.
414 1 :
415 1 : if shared := *((*uint8)(ptr)); shared == 0 {
416 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
417 1 : } else {
418 0 : // The shared length is != 0, which is invalid.
419 0 : panic("first key in block must have zero shared length")
420 : }
421 :
422 1 : var unshared uint32
423 1 : if a := *((*uint8)(ptr)); a < 128 {
424 1 : unshared = uint32(a)
425 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
426 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
427 0 : unshared = uint32(b)<<7 | uint32(a)
428 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
429 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
430 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
431 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
432 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
433 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
434 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
435 0 : } else {
436 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
437 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
438 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
439 0 : }
440 :
441 : // Skip the value length.
442 1 : if a := *((*uint8)(ptr)); a < 128 {
443 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
444 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); a < 128 {
445 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
446 1 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); a < 128 {
447 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
448 0 : } else if a := *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); a < 128 {
449 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
450 0 : } else {
451 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
452 0 : }
453 :
454 1 : firstKey := getBytes(ptr, int(unshared))
455 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
456 1 : // BlockIter benchmarks.
457 1 : if n := len(firstKey) - 8; n >= 0 {
458 1 : i.firstUserKey = firstKey[:n:n]
459 1 : } else {
460 0 : i.firstUserKey = nil
461 0 : return base.CorruptionErrorf("pebble/table: invalid firstKey in block")
462 0 : }
463 1 : if i.transforms.SyntheticPrefix != nil {
464 1 : i.firstUserKeyWithPrefixBuf = slices.Grow(i.firstUserKeyWithPrefixBuf[:0], len(i.transforms.SyntheticPrefix)+len(i.firstUserKey))
465 1 : i.firstUserKeyWithPrefixBuf = append(i.firstUserKeyWithPrefixBuf, i.transforms.SyntheticPrefix...)
466 1 : i.firstUserKeyWithPrefixBuf = append(i.firstUserKeyWithPrefixBuf, i.firstUserKey...)
467 1 : i.firstUserKey = i.firstUserKeyWithPrefixBuf
468 1 : }
469 1 : return nil
470 : }
471 :
472 1 : func (i *Iter) decodeInternalKey(key []byte) (hiddenPoint bool) {
473 1 : // Manually inlining base.DecodeInternalKey provides a 5-10% speedup on
474 1 : // BlockIter benchmarks.
475 1 : if n := len(key) - 8; n >= 0 {
476 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(key[n:]))
477 1 : hiddenPoint = i.transforms.HideObsoletePoints &&
478 1 : (trailer&TrailerObsoleteBit != 0)
479 1 : i.ikv.K.Trailer = trailer & TrailerObsoleteMask
480 1 : i.ikv.K.UserKey = key[:n:n]
481 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
482 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
483 1 : }
484 1 : } else {
485 1 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
486 1 : i.ikv.K.UserKey = nil
487 1 : }
488 1 : return hiddenPoint
489 : }
490 :
491 : // maybeReplaceSuffix replaces the suffix in i.ikey.UserKey with
492 : // i.transforms.syntheticSuffix.
493 1 : func (i *Iter) maybeReplaceSuffix() {
494 1 : if i.transforms.SyntheticSuffix.IsSet() && i.ikv.K.UserKey != nil {
495 1 : prefixLen := i.split(i.ikv.K.UserKey)
496 1 : // If ikey is cached or may get cached, we must copy
497 1 : // UserKey to a new buffer before suffix replacement.
498 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
499 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix...)
500 1 : i.ikv.K.UserKey = i.synthSuffixBuf
501 1 : }
502 : }
503 :
504 1 : func (i *Iter) clearCache() {
505 1 : i.cached = i.cached[:0]
506 1 : i.cachedBuf = i.cachedBuf[:0]
507 1 : }
508 :
509 1 : func (i *Iter) cacheEntry() {
510 1 : var valStart int32
511 1 : valSize := int32(len(i.val))
512 1 : if valSize > 0 {
513 1 : valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
514 1 : }
515 :
516 1 : i.cached = append(i.cached, blockEntry{
517 1 : offset: i.offset,
518 1 : keyStart: int32(len(i.cachedBuf)),
519 1 : keyEnd: int32(len(i.cachedBuf) + len(i.key)),
520 1 : valStart: valStart,
521 1 : valSize: valSize,
522 1 : })
523 1 : i.cachedBuf = append(i.cachedBuf, i.key...)
524 : }
525 :
526 : // IsLowerBound implements the block.DataBlockIterator interface.
527 1 : func (i *Iter) IsLowerBound(k []byte) bool {
528 1 : // Note: we ignore HideObsoletePoints, but false negatives are allowed.
529 1 : return i.cmp(i.firstUserKey, k) >= 0
530 1 : }
531 :
532 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
533 : // package.
534 1 : func (i *Iter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
535 1 : if invariants.Enabled && i.IsDataInvalidated() {
536 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
537 : }
538 1 : searchKey := key
539 1 : if i.transforms.SyntheticPrefix != nil {
540 1 : if !bytes.HasPrefix(key, i.transforms.SyntheticPrefix) {
541 0 : // The seek key is before or after the entire block of keys that start
542 0 : // with SyntheticPrefix. To determine which, we need to compare against a
543 0 : // valid key in the block. We use firstUserKey which has the synthetic
544 0 : // prefix.
545 0 : if i.cmp(i.firstUserKey, key) >= 0 {
546 0 : return i.First()
547 0 : }
548 : // Set the offset to the end of the block to mimic the offset of an
549 : // invalid iterator. This ensures a subsequent i.Prev() returns a valid
550 : // result.
551 0 : i.offset = i.restarts
552 0 : i.nextOffset = i.restarts
553 0 : return nil
554 : }
555 1 : searchKey = key[len(i.transforms.SyntheticPrefix):]
556 : }
557 :
558 1 : i.clearCache()
559 1 : // Find the index of the smallest restart point whose key is > the key
560 1 : // sought; index will be numRestarts if there is no such restart point.
561 1 : i.offset = 0
562 1 : var index int32
563 1 :
564 1 : {
565 1 : // NB: manually inlined sort.Seach is ~5% faster.
566 1 : //
567 1 : // Define f(-1) == false and f(n) == true.
568 1 : // Invariant: f(index-1) == false, f(upper) == true.
569 1 : upper := i.numRestarts
570 1 : for index < upper {
571 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
572 1 : // index ≤ h < upper
573 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
574 1 : // For a restart point, there are 0 bytes shared with the previous key.
575 1 : // The varint encoding of 0 occupies 1 byte.
576 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
577 1 :
578 1 : // Decode the key at that restart point, and compare it to the key
579 1 : // sought. See the comment in readEntry for why we manually inline the
580 1 : // varint decoding.
581 1 : var v1 uint32
582 1 : if a := *((*uint8)(ptr)); a < 128 {
583 1 : v1 = uint32(a)
584 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
585 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
586 0 : v1 = uint32(b)<<7 | uint32(a)
587 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
588 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
589 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
590 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
591 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
592 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
593 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
594 0 : } else {
595 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
596 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
597 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
598 0 : }
599 :
600 1 : if *((*uint8)(ptr)) < 128 {
601 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
602 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
603 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
604 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
605 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
606 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
607 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
608 0 : } else {
609 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
610 0 : }
611 :
612 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
613 : // speedup on BlockIter benchmarks.
614 1 : s := getBytes(ptr, int(v1))
615 1 : var k []byte
616 1 : if n := len(s) - 8; n >= 0 {
617 1 : k = s[:n:n]
618 1 : }
619 : // Else k is invalid, and left as nil
620 :
621 1 : if i.cmp(searchKey, k) > 0 {
622 1 : // The search key is greater than the user key at this restart point.
623 1 : // Search beyond this restart point, since we are trying to find the
624 1 : // first restart point with a user key >= the search key.
625 1 : index = h + 1 // preserves f(i-1) == false
626 1 : } else {
627 1 : // k >= search key, so prune everything after index (since index
628 1 : // satisfies the property we are looking for).
629 1 : upper = h // preserves f(j) == true
630 1 : }
631 : }
632 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
633 : // => answer is index.
634 : }
635 :
636 : // index is the first restart point with key >= search key. Define the keys
637 : // between a restart point and the next restart point as belonging to that
638 : // restart point.
639 : //
640 : // Since keys are strictly increasing, if index > 0 then the restart point
641 : // at index-1 will be the first one that has some keys belonging to it that
642 : // could be equal to the search key. If index == 0, then all keys in this
643 : // block are larger than the key sought, and offset remains at zero.
644 1 : if index > 0 {
645 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
646 1 : }
647 1 : i.readEntry()
648 1 : hiddenPoint := i.decodeInternalKey(i.key)
649 1 :
650 1 : // Iterate from that restart point to somewhere >= the key sought.
651 1 : if !i.Valid() {
652 0 : return nil
653 0 : }
654 :
655 : // A note on seeking in a block with a suffix replacement rule: even though
656 : // the binary search above was conducted on keys without suffix replacement,
657 : // Seek will still return the correct suffix replaced key. A binary
658 : // search without suffix replacement will land on a key that is _less_ than
659 : // the key the search would have landed on if all keys were already suffix
660 : // replaced. Since Seek then conducts forward iteration to the first suffix
661 : // replaced user key that is greater than or equal to the search key, the
662 : // correct key is still returned.
663 : //
664 : // As an example, consider the following block with a restart interval of 1,
665 : // with a replacement suffix of "4":
666 : // - Pre-suffix replacement: apple@1, banana@3
667 : // - Post-suffix replacement: apple@4, banana@4
668 : //
669 : // Suppose the client seeks with apple@3. Assuming suffixes sort in reverse
670 : // chronological order (i.e. apple@1>apple@3), the binary search without
671 : // suffix replacement would return apple@1. A binary search with suffix
672 : // replacement would return banana@4. After beginning forward iteration from
673 : // either returned restart point, forward iteration would
674 : // always return the correct key, banana@4.
675 : //
676 : // Further, if the user searched with apple@0 (i.e. a suffix less than the
677 : // pre replacement suffix) or with apple@5 (a suffix larger than the post
678 : // replacement suffix), the binary search with or without suffix replacement
679 : // would land on the same key, as we assume the following:
680 : // (1) no two keys in the sst share the same prefix.
681 : // (2) pebble.Compare(replacementSuffix,originalSuffix) > 0
682 :
683 1 : i.maybeReplaceSuffix()
684 1 :
685 1 : if !hiddenPoint && i.cmp(i.ikv.K.UserKey, key) >= 0 {
686 1 : // Initialize i.lazyValue
687 1 : if !i.lazyValueHandling.hasValuePrefix ||
688 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
689 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
690 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
691 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
692 1 : } else {
693 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
694 1 : }
695 1 : return &i.ikv
696 : }
697 1 : for i.Next(); i.Valid(); i.Next() {
698 1 : if i.cmp(i.ikv.K.UserKey, key) >= 0 {
699 1 : // i.Next() has already initialized i.ikv.LazyValue.
700 1 : return &i.ikv
701 1 : }
702 : }
703 1 : return nil
704 : }
705 :
706 : // SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
707 : // pebble package.
708 0 : func (i *Iter) SeekPrefixGE(prefix, key []byte, flags base.SeekGEFlags) *base.InternalKV {
709 0 : // This should never be called as prefix iteration is handled by sstable.Iterator.
710 0 : panic("pebble: SeekPrefixGE unimplemented")
711 : }
712 :
713 : // SeekLT implements internalIterator.SeekLT, as documented in the pebble
714 : // package.
715 1 : func (i *Iter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
716 1 : if invariants.Enabled && i.IsDataInvalidated() {
717 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
718 : }
719 1 : searchKey := key
720 1 : if i.transforms.SyntheticPrefix != nil {
721 1 : if !bytes.HasPrefix(key, i.transforms.SyntheticPrefix) {
722 1 : // The seek key is before or after the entire block of keys that start
723 1 : // with SyntheticPrefix. To determine which, we need to compare against a
724 1 : // valid key in the block. We use firstUserKey which has the synthetic
725 1 : // prefix.
726 1 : if i.cmp(i.firstUserKey, key) < 0 {
727 1 : return i.Last()
728 1 : }
729 : // Set the offset to the beginning of the block to mimic an exhausted
730 : // iterator that has conducted backward interation. This ensures a
731 : // subsequent Next() call returns the first key in the block.
732 1 : i.offset = -1
733 1 : i.nextOffset = 0
734 1 : return nil
735 : }
736 1 : searchKey = key[len(i.transforms.SyntheticPrefix):]
737 : }
738 :
739 1 : i.clearCache()
740 1 : // Find the index of the smallest restart point whose key is >= the key
741 1 : // sought; index will be numRestarts if there is no such restart point.
742 1 : i.offset = 0
743 1 : var index int32
744 1 :
745 1 : {
746 1 : // NB: manually inlined sort.Search is ~5% faster.
747 1 : //
748 1 : // Define f(-1) == false and f(n) == true.
749 1 : // Invariant: f(index-1) == false, f(upper) == true.
750 1 : upper := i.numRestarts
751 1 : for index < upper {
752 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
753 1 : // index ≤ h < upper
754 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
755 1 : // For a restart point, there are 0 bytes shared with the previous key.
756 1 : // The varint encoding of 0 occupies 1 byte.
757 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
758 1 :
759 1 : // Decode the key at that restart point, and compare it to the key
760 1 : // sought. See the comment in readEntry for why we manually inline the
761 1 : // varint decoding.
762 1 : var v1 uint32
763 1 : if a := *((*uint8)(ptr)); a < 128 {
764 1 : v1 = uint32(a)
765 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
766 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
767 0 : v1 = uint32(b)<<7 | uint32(a)
768 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
769 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
770 0 : v1 = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
771 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
772 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
773 0 : v1 = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
774 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
775 0 : } else {
776 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
777 0 : v1 = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
778 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
779 0 : }
780 :
781 1 : if *((*uint8)(ptr)) < 128 {
782 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
783 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))) < 128 {
784 1 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
785 1 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))) < 128 {
786 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
787 0 : } else if *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))) < 128 {
788 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
789 0 : } else {
790 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
791 0 : }
792 :
793 : // Manually inlining part of base.DecodeInternalKey provides a 5-10%
794 : // speedup on BlockIter benchmarks.
795 1 : s := getBytes(ptr, int(v1))
796 1 : var k []byte
797 1 : if n := len(s) - 8; n >= 0 {
798 1 : k = s[:n:n]
799 1 : }
800 : // Else k is invalid, and left as nil
801 :
802 1 : if i.cmp(searchKey, k) > 0 {
803 1 : // The search key is greater than the user key at this restart point.
804 1 : // Search beyond this restart point, since we are trying to find the
805 1 : // first restart point with a user key >= the search key.
806 1 : index = h + 1 // preserves f(i-1) == false
807 1 : } else {
808 1 : // k >= search key, so prune everything after index (since index
809 1 : // satisfies the property we are looking for).
810 1 : upper = h // preserves f(j) == true
811 1 : }
812 : }
813 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
814 : // => answer is index.
815 : }
816 :
817 1 : if index == 0 {
818 1 : if i.transforms.SyntheticSuffix.IsSet() {
819 1 : // The binary search was conducted on keys without suffix replacement,
820 1 : // implying the first key in the block may be less than the search key. To
821 1 : // double check, get the first key in the block with suffix replacement
822 1 : // and compare to the search key. Consider the following example: suppose
823 1 : // the user searches with a@3, the first key in the block is a@2 and the
824 1 : // block contains a suffix replacement rule of 4. Since a@3 sorts before
825 1 : // a@2, the binary search would return index==0. Without conducting the
826 1 : // suffix replacement, the SeekLT would incorrectly return nil. With
827 1 : // suffix replacement though, a@4 should be returned as a@4 sorts before
828 1 : // a@3.
829 1 : ikv := i.First()
830 1 : if i.cmp(ikv.K.UserKey, key) < 0 {
831 1 : return ikv
832 1 : }
833 : }
834 : // If index == 0 then all keys in this block are larger than the key
835 : // sought, so there is no match.
836 1 : i.offset = -1
837 1 : i.nextOffset = 0
838 1 : return nil
839 : }
840 :
841 : // INVARIANT: index > 0
842 :
843 : // Ignoring suffix replacement, index is the first restart point with key >=
844 : // search key. Define the keys between a restart point and the next restart
845 : // point as belonging to that restart point. Note that index could be equal to
846 : // i.numRestarts, i.e., we are past the last restart. Since keys are strictly
847 : // increasing, then the restart point at index-1 will be the first one that
848 : // has some keys belonging to it that are less than the search key.
849 : //
850 : // Next, we will search between the restart at index-1 and the restart point
851 : // at index, for the first key >= key, and then on finding it, return
852 : // i.Prev(). We need to know when we have hit the offset for index, since then
853 : // we can stop searching. targetOffset encodes that offset for index.
854 1 : targetOffset := i.restarts
855 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
856 1 : if index < i.numRestarts {
857 1 : targetOffset = decodeRestart(i.data[i.restarts+4*(index):])
858 1 :
859 1 : if i.transforms.SyntheticSuffix.IsSet() {
860 0 : // The binary search was conducted on keys without suffix replacement,
861 0 : // implying the returned restart point (index) may be less than the search
862 0 : // key, breaking the assumption described above.
863 0 : //
864 0 : // For example: consider this block with a replacement ts of 4, and
865 0 : // restart interval of 1: - pre replacement: a@3,b@2,c@3 - post
866 0 : // replacement: a@4,b@4,c@4
867 0 : //
868 0 : // Suppose the client calls SeekLT(b@3), SeekLT must return b@4.
869 0 : //
870 0 : // If the client calls SeekLT(b@3), the binary search would return b@2,
871 0 : // the lowest key geq to b@3, pre-suffix replacement. Then, SeekLT will
872 0 : // begin forward iteration from a@3, the previous restart point, to
873 0 : // b{suffix}. The iteration stops when it encounters a key geq to the
874 0 : // search key or if it reaches the upper bound. Without suffix
875 0 : // replacement, we can assume that the upper bound of this forward
876 0 : // iteration, b{suffix}, is greater than the search key, as implied by the
877 0 : // binary search.
878 0 : //
879 0 : // If we naively hold this assumption with suffix replacement, the
880 0 : // iteration would terminate at the upper bound, b@4, call i.Prev, and
881 0 : // incorrectly return a@4. To correct for this, if the original returned
882 0 : // index is less than the search key, shift our forward iteration to begin
883 0 : // at index instead of index -1. With suffix replacement the key at index
884 0 : // is guaranteed to be the highest restart point less than the seach key
885 0 : // (i.e. the same property of index-1 for a block without suffix
886 0 : // replacement). This property holds because of the invariant that a block
887 0 : // with suffix replacement will not have two keys that share the same
888 0 : // prefix. To consider the above example, binary searching with b@3 landed
889 0 : // naively at a@3, but since b@4<b@3, we shift our forward iteration to
890 0 : // begin at b@4. We never need to shift by more than one restart point
891 0 : // (i.e. to c@4) because it's impossible for the search key to be greater
892 0 : // than the key at the next restart point in the block because that
893 0 : // key will always have a different prefix. Put another way, because no
894 0 : // key in the block shares the same prefix, naive binary search should
895 0 : // always land at most 1 restart point off the correct one.
896 0 :
897 0 : naiveOffset := i.offset
898 0 : // Shift up to the original binary search result and decode the key.
899 0 : i.offset = targetOffset
900 0 : i.readEntry()
901 0 : i.decodeInternalKey(i.key)
902 0 : i.maybeReplaceSuffix()
903 0 :
904 0 : // If the binary search point is actually less than the search key, post
905 0 : // replacement, bump the target offset.
906 0 : if i.cmp(i.ikv.K.UserKey, key) < 0 {
907 0 : i.offset = targetOffset
908 0 : if index+1 < i.numRestarts {
909 0 : // if index+1 is within the i.data bounds, use it to find the target
910 0 : // offset.
911 0 : targetOffset = decodeRestart(i.data[i.restarts+4*(index+1):])
912 0 : } else {
913 0 : targetOffset = i.restarts
914 0 : }
915 0 : } else {
916 0 : i.offset = naiveOffset
917 0 : }
918 : }
919 : }
920 :
921 : // Init nextOffset for the forward iteration below.
922 1 : i.nextOffset = i.offset
923 1 :
924 1 : for {
925 1 : i.offset = i.nextOffset
926 1 : i.readEntry()
927 1 : // When hidden keys are common, there is additional optimization possible
928 1 : // by not caching entries that are hidden (note that some calls to
929 1 : // cacheEntry don't decode the internal key before caching, but checking
930 1 : // whether a key is hidden does not require full decoding). However, we do
931 1 : // need to use the blockEntry.offset in the cache for the first entry at
932 1 : // the reset point to do the binary search when the cache is empty -- so
933 1 : // we would need to cache that first entry (though not the key) even if
934 1 : // was hidden. Our current assumption is that if there are large numbers
935 1 : // of hidden keys we will be able to skip whole blocks (using block
936 1 : // property filters) so we don't bother optimizing.
937 1 : hiddenPoint := i.decodeInternalKey(i.key)
938 1 : i.maybeReplaceSuffix()
939 1 :
940 1 : // NB: we don't use the hiddenPoint return value of decodeInternalKey
941 1 : // since we want to stop as soon as we reach a key >= ikey.UserKey, so
942 1 : // that we can reverse.
943 1 : if i.cmp(i.ikv.K.UserKey, key) >= 0 {
944 1 : // The current key is greater than or equal to our search key. Back up to
945 1 : // the previous key which was less than our search key. Note that this for
946 1 : // loop will execute at least once with this if-block not being true, so
947 1 : // the key we are backing up to is the last one this loop cached.
948 1 : return i.Prev()
949 1 : }
950 :
951 1 : if i.nextOffset >= targetOffset {
952 1 : // We've reached the end of the current restart block. Return the
953 1 : // current key if not hidden, else call Prev().
954 1 : //
955 1 : // When the restart interval is 1, the first iteration of the for loop
956 1 : // will bring us here. In that case ikey is backed by the block so we
957 1 : // get the desired key stability guarantee for the lifetime of the
958 1 : // blockIter. That is, we never cache anything and therefore never
959 1 : // return a key backed by cachedBuf.
960 1 : if hiddenPoint {
961 1 : return i.Prev()
962 1 : }
963 1 : break
964 : }
965 1 : i.cacheEntry()
966 : }
967 :
968 1 : if !i.Valid() {
969 1 : return nil
970 1 : }
971 1 : if !i.lazyValueHandling.hasValuePrefix ||
972 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
973 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
974 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
975 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
976 1 : } else {
977 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
978 1 : }
979 1 : return &i.ikv
980 : }
981 :
982 : // First implements internalIterator.First, as documented in the pebble
983 : // package.
984 1 : func (i *Iter) First() *base.InternalKV {
985 1 : if invariants.Enabled && i.IsDataInvalidated() {
986 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
987 : }
988 :
989 1 : i.offset = 0
990 1 : if !i.Valid() {
991 1 : return nil
992 1 : }
993 1 : i.clearCache()
994 1 : i.readEntry()
995 1 : hiddenPoint := i.decodeInternalKey(i.key)
996 1 : if hiddenPoint {
997 1 : return i.Next()
998 1 : }
999 1 : i.maybeReplaceSuffix()
1000 1 : if !i.lazyValueHandling.hasValuePrefix ||
1001 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1002 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1003 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1004 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1005 1 : } else {
1006 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1007 1 : }
1008 1 : return &i.ikv
1009 : }
1010 :
1011 : const restartMaskLittleEndianHighByteWithoutSetHasSamePrefix byte = 0b0111_1111
1012 : const restartMaskLittleEndianHighByteOnlySetHasSamePrefix byte = 0b1000_0000
1013 :
1014 1 : func decodeRestart(b []byte) int32 {
1015 1 : _ = b[3] // bounds check hint to compiler; see golang.org/issue/14808
1016 1 : return int32(uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 |
1017 1 : uint32(b[3]&restartMaskLittleEndianHighByteWithoutSetHasSamePrefix)<<24)
1018 1 : }
1019 :
1020 : // Last implements internalIterator.Last, as documented in the pebble package.
1021 1 : func (i *Iter) Last() *base.InternalKV {
1022 1 : if invariants.Enabled && i.IsDataInvalidated() {
1023 0 : panic(errors.AssertionFailedf("invalidated blockIter used"))
1024 : }
1025 :
1026 : // Seek forward from the last restart point.
1027 1 : i.offset = decodeRestart(i.data[i.restarts+4*(i.numRestarts-1):])
1028 1 : if !i.Valid() {
1029 1 : return nil
1030 1 : }
1031 :
1032 1 : i.readEntry()
1033 1 : i.clearCache()
1034 1 :
1035 1 : for i.nextOffset < i.restarts {
1036 1 : i.cacheEntry()
1037 1 : i.offset = i.nextOffset
1038 1 : i.readEntry()
1039 1 : }
1040 :
1041 1 : hiddenPoint := i.decodeInternalKey(i.key)
1042 1 : if hiddenPoint {
1043 1 : return i.Prev()
1044 1 : }
1045 1 : i.maybeReplaceSuffix()
1046 1 : if !i.lazyValueHandling.hasValuePrefix ||
1047 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1048 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1049 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1050 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1051 1 : } else {
1052 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1053 1 : }
1054 1 : return &i.ikv
1055 : }
1056 :
1057 : // Next implements internalIterator.Next, as documented in the pebble
1058 : // package.
1059 1 : func (i *Iter) Next() *base.InternalKV {
1060 1 : if len(i.cachedBuf) > 0 {
1061 1 : // We're switching from reverse iteration to forward iteration. We need to
1062 1 : // populate i.fullKey with the current key we're positioned at so that
1063 1 : // readEntry() can use i.fullKey for key prefix decompression. Note that we
1064 1 : // don't know whether i.key is backed by i.cachedBuf or i.fullKey (if
1065 1 : // SeekLT was the previous call, i.key may be backed by i.fullKey), but
1066 1 : // copying into i.fullKey works for both cases.
1067 1 : //
1068 1 : // TODO(peter): Rather than clearing the cache, we could instead use the
1069 1 : // cache until it is exhausted. This would likely be faster than falling
1070 1 : // through to the normal forward iteration code below.
1071 1 : i.fullKey = append(i.fullKey[:0], i.key...)
1072 1 : i.clearCache()
1073 1 : }
1074 :
1075 : start:
1076 1 : i.offset = i.nextOffset
1077 1 : if !i.Valid() {
1078 1 : return nil
1079 1 : }
1080 1 : i.readEntry()
1081 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1082 1 : if n := len(i.key) - 8; n >= 0 {
1083 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1084 1 : hiddenPoint := i.transforms.HideObsoletePoints &&
1085 1 : (trailer&TrailerObsoleteBit != 0)
1086 1 : i.ikv.K.Trailer = trailer & TrailerObsoleteMask
1087 1 : i.ikv.K.UserKey = i.key[:n:n]
1088 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1089 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1090 1 : }
1091 1 : if hiddenPoint {
1092 1 : goto start
1093 : }
1094 1 : if i.transforms.SyntheticSuffix.IsSet() {
1095 1 : // Inlined version of i.maybeReplaceSuffix()
1096 1 : prefixLen := i.split(i.ikv.K.UserKey)
1097 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1098 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix...)
1099 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1100 1 : }
1101 0 : } else {
1102 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1103 0 : i.ikv.K.UserKey = nil
1104 0 : }
1105 1 : if !i.lazyValueHandling.hasValuePrefix ||
1106 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1107 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1108 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1109 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1110 1 : } else {
1111 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1112 1 : }
1113 1 : return &i.ikv
1114 : }
1115 :
1116 : // NextPrefix implements (base.InternalIterator).NextPrefix.
1117 1 : func (i *Iter) NextPrefix(succKey []byte) *base.InternalKV {
1118 1 : if i.lazyValueHandling.hasValuePrefix {
1119 1 : return i.nextPrefixV3(succKey)
1120 1 : }
1121 1 : const nextsBeforeSeek = 3
1122 1 : kv := i.Next()
1123 1 : for j := 1; kv != nil && i.cmp(kv.K.UserKey, succKey) < 0; j++ {
1124 1 : if j >= nextsBeforeSeek {
1125 1 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1126 1 : }
1127 1 : kv = i.Next()
1128 : }
1129 1 : return kv
1130 : }
1131 :
1132 1 : func (i *Iter) nextPrefixV3(succKey []byte) *base.InternalKV {
1133 1 : // Doing nexts that involve a key comparison can be expensive (and the cost
1134 1 : // depends on the key length), so we use the same threshold of 3 that we use
1135 1 : // for TableFormatPebblev2 in blockIter.nextPrefix above. The next fast path
1136 1 : // that looks at setHasSamePrefix takes ~5ns per key, which is ~150x faster
1137 1 : // than doing a SeekGE within the block, so we do this 16 times
1138 1 : // (~5ns*16=80ns), and then switch to looking at restarts. Doing the binary
1139 1 : // search for the restart consumes > 100ns. If the number of versions is >
1140 1 : // 17, we will increment nextFastCount to 17, then do a binary search, and
1141 1 : // on average need to find a key between two restarts, so another 8 steps
1142 1 : // corresponding to nextFastCount, for a mean total of 17 + 8 = 25 such
1143 1 : // steps.
1144 1 : //
1145 1 : // TODO(sumeer): use the configured restartInterval for the sstable when it
1146 1 : // was written (which we don't currently store) instead of the default value
1147 1 : // of 16.
1148 1 : const nextCmpThresholdBeforeSeek = 3
1149 1 : const nextFastThresholdBeforeRestarts = 16
1150 1 : nextCmpCount := 0
1151 1 : nextFastCount := 0
1152 1 : usedRestarts := false
1153 1 : // INVARIANT: blockIter is valid.
1154 1 : if invariants.Enabled && !i.Valid() {
1155 0 : panic(errors.AssertionFailedf("nextPrefixV3 called on invalid blockIter"))
1156 : }
1157 1 : prevKeyIsSet := i.ikv.Kind() == base.InternalKeyKindSet
1158 1 : for {
1159 1 : i.offset = i.nextOffset
1160 1 : if !i.Valid() {
1161 1 : return nil
1162 1 : }
1163 : // Need to decode the length integers, so we can compute nextOffset.
1164 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1165 1 : // This is an ugly performance hack. Reading entries from blocks is one of
1166 1 : // the inner-most routines and decoding the 3 varints per-entry takes
1167 1 : // significant time. Neither go1.11 or go1.12 will inline decodeVarint for
1168 1 : // us, so we do it manually. This provides a 10-15% performance improvement
1169 1 : // on blockIter benchmarks on both go1.11 and go1.12.
1170 1 : //
1171 1 : // TODO(peter): remove this hack if go:inline is ever supported.
1172 1 :
1173 1 : // Decode the shared key length integer.
1174 1 : var shared uint32
1175 1 : if a := *((*uint8)(ptr)); a < 128 {
1176 1 : shared = uint32(a)
1177 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1178 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1179 0 : shared = uint32(b)<<7 | uint32(a)
1180 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1181 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1182 0 : shared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1183 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1184 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1185 0 : shared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1186 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1187 0 : } else {
1188 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1189 0 : shared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1190 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1191 0 : }
1192 : // Decode the unshared key length integer.
1193 1 : var unshared uint32
1194 1 : if a := *((*uint8)(ptr)); a < 128 {
1195 1 : unshared = uint32(a)
1196 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1197 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1198 0 : unshared = uint32(b)<<7 | uint32(a)
1199 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1200 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1201 0 : unshared = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1202 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1203 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1204 0 : unshared = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1205 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1206 0 : } else {
1207 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1208 0 : unshared = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1209 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1210 0 : }
1211 : // Decode the value length integer.
1212 1 : var value uint32
1213 1 : if a := *((*uint8)(ptr)); a < 128 {
1214 1 : value = uint32(a)
1215 1 : ptr = unsafe.Pointer(uintptr(ptr) + 1)
1216 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1217 0 : value = uint32(b)<<7 | uint32(a)
1218 0 : ptr = unsafe.Pointer(uintptr(ptr) + 2)
1219 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1220 0 : value = uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1221 0 : ptr = unsafe.Pointer(uintptr(ptr) + 3)
1222 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1223 0 : value = uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1224 0 : ptr = unsafe.Pointer(uintptr(ptr) + 4)
1225 0 : } else {
1226 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1227 0 : value = uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a)
1228 0 : ptr = unsafe.Pointer(uintptr(ptr) + 5)
1229 0 : }
1230 1 : if i.transforms.SyntheticPrefix != nil {
1231 0 : shared += uint32(len(i.transforms.SyntheticPrefix))
1232 0 : }
1233 : // The starting position of the value.
1234 1 : valuePtr := unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
1235 1 : i.nextOffset = int32(uintptr(valuePtr)-uintptr(i.ptr)) + int32(value)
1236 1 : if invariants.Enabled && unshared < 8 {
1237 0 : // This should not happen since only the key prefix is shared, so even
1238 0 : // if the prefix length is the same as the user key length, the unshared
1239 0 : // will include the trailer.
1240 0 : panic(errors.AssertionFailedf("unshared %d is too small", unshared))
1241 : }
1242 : // The trailer is written in little endian, so the key kind is the first
1243 : // byte in the trailer that is encoded in the slice [unshared-8:unshared].
1244 1 : keyKind := base.InternalKeyKind((*[manual.MaxArrayLen]byte)(ptr)[unshared-8])
1245 1 : keyKind = keyKind & base.InternalKeyKindSSTableInternalObsoleteMask
1246 1 : prefixChanged := false
1247 1 : if keyKind == base.InternalKeyKindSet {
1248 1 : if invariants.Enabled && value == 0 {
1249 0 : panic(errors.AssertionFailedf("value is of length 0, but we expect a valuePrefix"))
1250 : }
1251 1 : valPrefix := *((*block.ValuePrefix)(valuePtr))
1252 1 : if valPrefix.SetHasSamePrefix() {
1253 1 : // Fast-path. No need to assemble i.fullKey, or update i.key. We know
1254 1 : // that subsequent keys will not have a shared length that is greater
1255 1 : // than the prefix of the current key, which is also the prefix of
1256 1 : // i.key. Since we are continuing to iterate, we don't need to
1257 1 : // initialize i.ikey and i.lazyValue (these are initialized before
1258 1 : // returning).
1259 1 : nextFastCount++
1260 1 : if nextFastCount > nextFastThresholdBeforeRestarts {
1261 0 : if usedRestarts {
1262 0 : // Exhausted iteration budget. This will never happen unless
1263 0 : // someone is using a restart interval > 16. It is just to guard
1264 0 : // against long restart intervals causing too much iteration.
1265 0 : break
1266 : }
1267 : // Haven't used restarts yet, so find the first restart at or beyond
1268 : // the current offset.
1269 0 : targetOffset := i.offset
1270 0 : var index int32
1271 0 : {
1272 0 : // NB: manually inlined sort.Sort is ~5% faster.
1273 0 : //
1274 0 : // f defined for a restart point is true iff the offset >=
1275 0 : // targetOffset.
1276 0 : // Define f(-1) == false and f(i.numRestarts) == true.
1277 0 : // Invariant: f(index-1) == false, f(upper) == true.
1278 0 : upper := i.numRestarts
1279 0 : for index < upper {
1280 0 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1281 0 : // index ≤ h < upper
1282 0 : offset := decodeRestart(i.data[i.restarts+4*h:])
1283 0 : if offset < targetOffset {
1284 0 : index = h + 1 // preserves f(index-1) == false
1285 0 : } else {
1286 0 : upper = h // preserves f(upper) == true
1287 0 : }
1288 : }
1289 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1290 : // => answer is index.
1291 : }
1292 0 : usedRestarts = true
1293 0 : nextFastCount = 0
1294 0 : if index == i.numRestarts {
1295 0 : // Already past the last real restart, so iterate a bit more until
1296 0 : // we are done with the block.
1297 0 : continue
1298 : }
1299 : // Have some real restarts after index. NB: index is the first
1300 : // restart at or beyond the current offset.
1301 0 : startingIndex := index
1302 0 : for index != i.numRestarts &&
1303 0 : // The restart at index is 4 bytes written in little endian format
1304 0 : // starting at i.restart+4*index. The 0th byte is the least
1305 0 : // significant and the 3rd byte is the most significant. Since the
1306 0 : // most significant bit of the 3rd byte is what we use for
1307 0 : // encoding the set-has-same-prefix information, the indexing
1308 0 : // below has +3.
1309 0 : i.data[i.restarts+4*index+3]&restartMaskLittleEndianHighByteOnlySetHasSamePrefix != 0 {
1310 0 : // We still have the same prefix, so move to the next restart.
1311 0 : index++
1312 0 : }
1313 : // index is the first restart that did not have the same prefix.
1314 0 : if index != startingIndex {
1315 0 : // Managed to skip past at least one restart. Resume iteration
1316 0 : // from index-1. Since nextFastCount has been reset to 0, we
1317 0 : // should be able to iterate to the next prefix.
1318 0 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1319 0 : i.readEntry()
1320 0 : }
1321 : // Else, unable to skip past any restart. Resume iteration. Since
1322 : // nextFastCount has been reset to 0, we should be able to iterate
1323 : // to the next prefix.
1324 0 : continue
1325 : }
1326 1 : continue
1327 1 : } else if prevKeyIsSet {
1328 1 : prefixChanged = true
1329 1 : }
1330 1 : } else {
1331 1 : prevKeyIsSet = false
1332 1 : }
1333 : // Slow-path cases:
1334 : // - (Likely) The prefix has changed.
1335 : // - (Unlikely) The prefix has not changed.
1336 : // We assemble the key etc. under the assumption that it is the likely
1337 : // case.
1338 1 : unsharedKey := getBytes(ptr, int(unshared))
1339 1 : // TODO(sumeer): move this into the else block below. This is a bit tricky
1340 1 : // since the current logic assumes we have always copied the latest key
1341 1 : // into fullKey, which is why when we get to the next key we can (a)
1342 1 : // access i.fullKey[:shared], (b) append only the unsharedKey to
1343 1 : // i.fullKey. For (a), we can access i.key[:shared] since that memory is
1344 1 : // valid (even if unshared). For (b), we will need to remember whether
1345 1 : // i.key refers to i.fullKey or not, and can append the unsharedKey only
1346 1 : // in the former case and for the latter case need to copy the shared part
1347 1 : // too. This same comment applies to the other place where we can do this
1348 1 : // optimization, in readEntry().
1349 1 : i.fullKey = append(i.fullKey[:shared], unsharedKey...)
1350 1 : i.val = getBytes(valuePtr, int(value))
1351 1 : if shared == 0 {
1352 1 : // Provide stability for the key across positioning calls if the key
1353 1 : // doesn't share a prefix with the previous key. This removes requiring the
1354 1 : // key to be copied if the caller knows the block has a restart interval of
1355 1 : // 1. An important example of this is range-del blocks.
1356 1 : i.key = unsharedKey
1357 1 : } else {
1358 1 : i.key = i.fullKey
1359 1 : }
1360 : // Manually inlined version of i.decodeInternalKey(i.key).
1361 1 : hiddenPoint := false
1362 1 : if n := len(i.key) - 8; n >= 0 {
1363 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1364 1 : hiddenPoint = i.transforms.HideObsoletePoints &&
1365 1 : (trailer&TrailerObsoleteBit != 0)
1366 1 : i.ikv.K = base.InternalKey{
1367 1 : Trailer: trailer & TrailerObsoleteMask,
1368 1 : UserKey: i.key[:n:n],
1369 1 : }
1370 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1371 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1372 1 : }
1373 1 : if i.transforms.SyntheticSuffix.IsSet() {
1374 0 : // Inlined version of i.maybeReplaceSuffix()
1375 0 : prefixLen := i.split(i.ikv.K.UserKey)
1376 0 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1377 0 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix...)
1378 0 : i.ikv.K.UserKey = i.synthSuffixBuf
1379 0 : }
1380 0 : } else {
1381 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1382 0 : i.ikv.K.UserKey = nil
1383 0 : }
1384 1 : nextCmpCount++
1385 1 : if invariants.Enabled && prefixChanged && i.cmp(i.ikv.K.UserKey, succKey) < 0 {
1386 0 : panic(errors.AssertionFailedf("prefix should have changed but %x < %x",
1387 0 : i.ikv.K.UserKey, succKey))
1388 : }
1389 1 : if prefixChanged || i.cmp(i.ikv.K.UserKey, succKey) >= 0 {
1390 1 : // Prefix has changed.
1391 1 : if hiddenPoint {
1392 1 : return i.Next()
1393 1 : }
1394 1 : if invariants.Enabled && !i.lazyValueHandling.hasValuePrefix {
1395 0 : panic(errors.AssertionFailedf("nextPrefixV3 being run for non-v3 sstable"))
1396 : }
1397 1 : if i.ikv.K.Kind() != base.InternalKeyKindSet {
1398 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1399 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1400 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1401 1 : } else {
1402 0 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1403 0 : }
1404 1 : return &i.ikv
1405 : }
1406 : // Else prefix has not changed.
1407 :
1408 1 : if nextCmpCount >= nextCmpThresholdBeforeSeek {
1409 0 : break
1410 : }
1411 : }
1412 0 : return i.SeekGE(succKey, base.SeekGEFlagsNone)
1413 : }
1414 :
1415 : // Prev implements internalIterator.Prev, as documented in the pebble
1416 : // package.
1417 1 : func (i *Iter) Prev() *base.InternalKV {
1418 1 : start:
1419 1 : for n := len(i.cached) - 1; n >= 0; n-- {
1420 1 : i.nextOffset = i.offset
1421 1 : e := &i.cached[n]
1422 1 : i.offset = e.offset
1423 1 : i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
1424 1 : // Manually inlined version of i.decodeInternalKey(i.key).
1425 1 : i.key = i.cachedBuf[e.keyStart:e.keyEnd]
1426 1 : if n := len(i.key) - 8; n >= 0 {
1427 1 : trailer := base.InternalKeyTrailer(binary.LittleEndian.Uint64(i.key[n:]))
1428 1 : hiddenPoint := i.transforms.HideObsoletePoints &&
1429 1 : (trailer&TrailerObsoleteBit != 0)
1430 1 : if hiddenPoint {
1431 1 : continue
1432 : }
1433 1 : i.ikv.K = base.InternalKey{
1434 1 : Trailer: trailer & TrailerObsoleteMask,
1435 1 : UserKey: i.key[:n:n],
1436 1 : }
1437 1 : if n := i.transforms.SyntheticSeqNum; n != 0 {
1438 1 : i.ikv.K.SetSeqNum(base.SeqNum(n))
1439 1 : }
1440 1 : if i.transforms.SyntheticSuffix.IsSet() {
1441 1 : // Inlined version of i.maybeReplaceSuffix()
1442 1 : prefixLen := i.split(i.ikv.K.UserKey)
1443 1 : // If ikey is cached or may get cached, we must de-reference
1444 1 : // UserKey before suffix replacement.
1445 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1446 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix...)
1447 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1448 1 : }
1449 0 : } else {
1450 0 : i.ikv.K.Trailer = base.InternalKeyTrailer(base.InternalKeyKindInvalid)
1451 0 : i.ikv.K.UserKey = nil
1452 0 : }
1453 1 : i.cached = i.cached[:n]
1454 1 : if !i.lazyValueHandling.hasValuePrefix ||
1455 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1456 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1457 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1458 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1459 1 : } else {
1460 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1461 1 : }
1462 1 : return &i.ikv
1463 : }
1464 :
1465 1 : i.clearCache()
1466 1 : if i.offset <= 0 {
1467 1 : i.offset = -1
1468 1 : i.nextOffset = 0
1469 1 : return nil
1470 1 : }
1471 :
1472 1 : targetOffset := i.offset
1473 1 : var index int32
1474 1 :
1475 1 : {
1476 1 : // NB: manually inlined sort.Sort is ~5% faster.
1477 1 : //
1478 1 : // Define f(-1) == false and f(n) == true.
1479 1 : // Invariant: f(index-1) == false, f(upper) == true.
1480 1 : upper := i.numRestarts
1481 1 : for index < upper {
1482 1 : h := int32(uint(index+upper) >> 1) // avoid overflow when computing h
1483 1 : // index ≤ h < upper
1484 1 : offset := decodeRestart(i.data[i.restarts+4*h:])
1485 1 : if offset < targetOffset {
1486 1 : // Looking for the first restart that has offset >= targetOffset, so
1487 1 : // ignore h and earlier.
1488 1 : index = h + 1 // preserves f(i-1) == false
1489 1 : } else {
1490 1 : upper = h // preserves f(j) == true
1491 1 : }
1492 : }
1493 : // index == upper, f(index-1) == false, and f(upper) (= f(index)) == true
1494 : // => answer is index.
1495 : }
1496 :
1497 : // index is first restart with offset >= targetOffset. Note that
1498 : // targetOffset may not be at a restart point since one can call Prev()
1499 : // after Next() (so the cache was not populated) and targetOffset refers to
1500 : // the current entry. index-1 must have an offset < targetOffset (it can't
1501 : // be equal to targetOffset since the binary search would have selected that
1502 : // as the index).
1503 1 : i.offset = 0
1504 1 : if index > 0 {
1505 1 : i.offset = decodeRestart(i.data[i.restarts+4*(index-1):])
1506 1 : }
1507 : // TODO(sumeer): why is the else case not an error given targetOffset is a
1508 : // valid offset.
1509 :
1510 1 : i.readEntry()
1511 1 :
1512 1 : // We stop when i.nextOffset == targetOffset since the targetOffset is the
1513 1 : // entry we are stepping back from, and we don't need to cache the entry
1514 1 : // before it, since it is the candidate to return.
1515 1 : for i.nextOffset < targetOffset {
1516 1 : i.cacheEntry()
1517 1 : i.offset = i.nextOffset
1518 1 : i.readEntry()
1519 1 : }
1520 :
1521 1 : hiddenPoint := i.decodeInternalKey(i.key)
1522 1 : if hiddenPoint {
1523 1 : // Use the cache.
1524 1 : goto start
1525 : }
1526 1 : if i.transforms.SyntheticSuffix.IsSet() {
1527 1 : // Inlined version of i.maybeReplaceSuffix()
1528 1 : prefixLen := i.split(i.ikv.K.UserKey)
1529 1 : // If ikey is cached or may get cached, we must de-reference
1530 1 : // UserKey before suffix replacement.
1531 1 : i.synthSuffixBuf = append(i.synthSuffixBuf[:0], i.ikv.K.UserKey[:prefixLen]...)
1532 1 : i.synthSuffixBuf = append(i.synthSuffixBuf, i.transforms.SyntheticSuffix...)
1533 1 : i.ikv.K.UserKey = i.synthSuffixBuf
1534 1 : }
1535 1 : if !i.lazyValueHandling.hasValuePrefix ||
1536 1 : i.ikv.K.Kind() != base.InternalKeyKindSet {
1537 1 : i.ikv.V = base.MakeInPlaceValue(i.val)
1538 1 : } else if i.lazyValueHandling.getValue == nil || !block.ValuePrefix(i.val[0]).IsValueHandle() {
1539 1 : i.ikv.V = base.MakeInPlaceValue(i.val[1:])
1540 1 : } else {
1541 1 : i.ikv.V = i.lazyValueHandling.getValue.GetLazyValueForPrefixAndValueHandle(i.val)
1542 1 : }
1543 1 : return &i.ikv
1544 : }
1545 :
1546 : // Key returns the internal key at the current iterator position.
1547 0 : func (i *Iter) Key() *base.InternalKey {
1548 0 : return &i.ikv.K
1549 0 : }
1550 :
1551 : // KV returns the internal KV at the current iterator position.
1552 1 : func (i *Iter) KV() *base.InternalKV {
1553 1 : return &i.ikv
1554 1 : }
1555 :
1556 : // Value returns the value at the current iterator position.
1557 0 : func (i *Iter) Value() base.LazyValue {
1558 0 : return i.ikv.V
1559 0 : }
1560 :
1561 : // Error implements internalIterator.Error, as documented in the pebble
1562 : // package.
1563 1 : func (i *Iter) Error() error {
1564 1 : return nil // infallible
1565 1 : }
1566 :
1567 : // Close implements internalIterator.Close, as documented in the pebble
1568 : // package.
1569 1 : func (i *Iter) Close() error {
1570 1 : i.handle.Release()
1571 1 : i.handle = block.BufferHandle{}
1572 1 : i.val = nil
1573 1 : i.ikv = base.InternalKV{}
1574 1 : i.lazyValueHandling.getValue = nil
1575 1 : return nil
1576 1 : }
1577 :
1578 : // SetBounds implements base.InternalIterator. It panics, as bounds should
1579 : // always be handled the by the parent sstable iterator.
1580 0 : func (i *Iter) SetBounds(lower, upper []byte) {
1581 0 : // This should never be called as bounds are handled by sstable.Iterator.
1582 0 : panic("pebble: SetBounds unimplemented")
1583 : }
1584 :
1585 : // SetContext implements base.InternalIterator.
1586 0 : func (i *Iter) SetContext(_ context.Context) {}
1587 :
1588 : // Valid returns true if the iterator is currently positioned at a valid KV.
1589 1 : func (i *Iter) Valid() bool {
1590 1 : return i.offset >= 0 && i.offset < i.restarts
1591 1 : }
1592 :
1593 : // DebugTree is part of the InternalIterator interface.
1594 0 : func (i *Iter) DebugTree(tp treeprinter.Node) {
1595 0 : tp.Childf("%T(%p)", i, i)
1596 0 : }
1597 :
1598 0 : func (i *Iter) getRestart(idx int) int32 {
1599 0 : return int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*int32(idx):]))
1600 0 : }
1601 :
1602 0 : func (i *Iter) isRestartPoint() bool {
1603 0 : j := sort.Search(int(i.numRestarts), func(j int) bool {
1604 0 : return i.getRestart(j) >= i.offset
1605 0 : })
1606 0 : return j < int(i.numRestarts) && i.getRestart(j) == i.offset
1607 : }
1608 :
1609 : // DescribeKV is a function that formats a key-value pair, writing the
1610 : // description to w.
1611 : type DescribeKV func(w io.Writer, key *base.InternalKey, val []byte, enc KVEncoding)
1612 :
1613 : // KVEncoding describes the encoding of a key-value pair within the block.
1614 : type KVEncoding struct {
1615 : // IsRestart is true if the key is a restart point.
1616 : IsRestart bool
1617 : // Offset is the position within the block at which the key-value pair is
1618 : // encoded.
1619 : Offset int32
1620 : // Length is the total length of the KV pair as it is encoded in the block
1621 : // format.
1622 : Length int32
1623 : // KeyShared is the number of bytes this KV's user key shared with its predecessor.
1624 : KeyShared uint32
1625 : // KeyUnshared is the number of bytes this KV's user key did not share with
1626 : // its predecessor.
1627 : KeyUnshared uint32
1628 : // ValueLen is the length of the internal value.
1629 : ValueLen uint32
1630 : }
1631 :
1632 : // Describe describes the contents of a block, writing the description to w.
1633 : // It invokes fmtKV to describe each key-value pair.
1634 0 : func (i *Iter) Describe(tp treeprinter.Node, fmtKV DescribeKV) {
1635 0 : var buf bytes.Buffer
1636 0 : for kv := i.First(); kv != nil; kv = i.Next() {
1637 0 : enc := KVEncoding{
1638 0 : IsRestart: i.isRestartPoint(),
1639 0 : Offset: i.offset,
1640 0 : Length: int32(i.nextOffset - i.offset),
1641 0 : }
1642 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1643 0 : enc.KeyShared, ptr = decodeVarint(ptr)
1644 0 : enc.KeyUnshared, ptr = decodeVarint(ptr)
1645 0 : enc.ValueLen, _ = decodeVarint(ptr)
1646 0 : buf.Reset()
1647 0 : fmtKV(&buf, &kv.K, kv.V.ValueOrHandle, enc)
1648 0 : tp.Child(buf.String())
1649 0 : }
1650 : // Format the restart points.
1651 0 : n := tp.Child("restart points")
1652 0 : // Format the restart points.
1653 0 : for j := 0; j < int(i.numRestarts); j++ {
1654 0 : offset := i.getRestart(j)
1655 0 : n.Childf("%05d [restart %d]", uint64(i.restarts+4*int32(j)), offset)
1656 0 : }
1657 : }
1658 :
1659 : // RawIter is an iterator over a single block of data. Unlike blockIter,
1660 : // keys are stored in "raw" format (i.e. not as internal keys). Note that there
1661 : // is significant similarity between this code and the code in blockIter. Yet
1662 : // reducing duplication is difficult due to the blockIter being performance
1663 : // critical. RawIter must only be used for blocks where the value is
1664 : // stored together with the key.
1665 : type RawIter struct {
1666 : cmp base.Compare
1667 : offset int32
1668 : nextOffset int32
1669 : restarts int32
1670 : numRestarts int32
1671 : ptr unsafe.Pointer
1672 : data []byte
1673 : key, val []byte
1674 : ikey base.InternalKey
1675 : cached []blockEntry
1676 : cachedBuf []byte
1677 : }
1678 :
1679 : // NewRawIter constructs a new raw block iterator.
1680 1 : func NewRawIter(cmp base.Compare, block []byte) (*RawIter, error) {
1681 1 : i := &RawIter{}
1682 1 : return i, i.Init(cmp, block)
1683 1 : }
1684 :
1685 : // Init initializes the raw block iterator.
1686 1 : func (i *RawIter) Init(cmp base.Compare, blk []byte) error {
1687 1 : numRestarts := int32(binary.LittleEndian.Uint32(blk[len(blk)-4:]))
1688 1 : if numRestarts == 0 {
1689 0 : return base.CorruptionErrorf("pebble/table: invalid table (block has no restart points)")
1690 0 : }
1691 1 : i.cmp = cmp
1692 1 : i.restarts = int32(len(blk)) - 4*(1+numRestarts)
1693 1 : i.numRestarts = numRestarts
1694 1 : i.ptr = unsafe.Pointer(&blk[0])
1695 1 : i.data = blk
1696 1 : if i.key == nil {
1697 1 : i.key = make([]byte, 0, 256)
1698 1 : } else {
1699 0 : i.key = i.key[:0]
1700 0 : }
1701 1 : i.val = nil
1702 1 : i.clearCache()
1703 1 : return nil
1704 : }
1705 :
1706 1 : func (i *RawIter) readEntry() {
1707 1 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1708 1 : shared, ptr := decodeVarint(ptr)
1709 1 : unshared, ptr := decodeVarint(ptr)
1710 1 : value, ptr := decodeVarint(ptr)
1711 1 : i.key = append(i.key[:shared], getBytes(ptr, int(unshared))...)
1712 1 : i.key = i.key[:len(i.key):len(i.key)]
1713 1 : ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
1714 1 : i.val = getBytes(ptr, int(value))
1715 1 : i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
1716 1 : }
1717 :
1718 1 : func (i *RawIter) loadEntry() {
1719 1 : i.readEntry()
1720 1 : i.ikey.UserKey = i.key
1721 1 : }
1722 :
1723 1 : func (i *RawIter) clearCache() {
1724 1 : i.cached = i.cached[:0]
1725 1 : i.cachedBuf = i.cachedBuf[:0]
1726 1 : }
1727 :
1728 0 : func (i *RawIter) cacheEntry() {
1729 0 : var valStart int32
1730 0 : valSize := int32(len(i.val))
1731 0 : if valSize > 0 {
1732 0 : valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
1733 0 : }
1734 :
1735 0 : i.cached = append(i.cached, blockEntry{
1736 0 : offset: i.offset,
1737 0 : keyStart: int32(len(i.cachedBuf)),
1738 0 : keyEnd: int32(len(i.cachedBuf) + len(i.key)),
1739 0 : valStart: valStart,
1740 0 : valSize: valSize,
1741 0 : })
1742 0 : i.cachedBuf = append(i.cachedBuf, i.key...)
1743 : }
1744 :
1745 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
1746 : // package.
1747 0 : func (i *RawIter) SeekGE(key []byte) bool {
1748 0 : // Find the index of the smallest restart point whose key is > the key
1749 0 : // sought; index will be numRestarts if there is no such restart point.
1750 0 : i.offset = 0
1751 0 : index := sort.Search(int(i.numRestarts), func(j int) bool {
1752 0 : offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
1753 0 : // For a restart point, there are 0 bytes shared with the previous key.
1754 0 : // The varint encoding of 0 occupies 1 byte.
1755 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
1756 0 : // Decode the key at that restart point, and compare it to the key sought.
1757 0 : v1, ptr := decodeVarint(ptr)
1758 0 : _, ptr = decodeVarint(ptr)
1759 0 : s := getBytes(ptr, int(v1))
1760 0 : return i.cmp(key, s) < 0
1761 0 : })
1762 :
1763 : // Since keys are strictly increasing, if index > 0 then the restart point at
1764 : // index-1 will be the largest whose key is <= the key sought. If index ==
1765 : // 0, then all keys in this block are larger than the key sought, and offset
1766 : // remains at zero.
1767 0 : if index > 0 {
1768 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
1769 0 : }
1770 0 : i.loadEntry()
1771 0 :
1772 0 : // Iterate from that restart point to somewhere >= the key sought.
1773 0 : for valid := i.Valid(); valid; valid = i.Next() {
1774 0 : if i.cmp(key, i.key) <= 0 {
1775 0 : break
1776 : }
1777 : }
1778 0 : return i.Valid()
1779 : }
1780 :
1781 : // First implements internalIterator.First, as documented in the pebble
1782 : // package.
1783 1 : func (i *RawIter) First() bool {
1784 1 : i.offset = 0
1785 1 : i.loadEntry()
1786 1 : return i.Valid()
1787 1 : }
1788 :
1789 : // Last implements internalIterator.Last, as documented in the pebble package.
1790 0 : func (i *RawIter) Last() bool {
1791 0 : // Seek forward from the last restart point.
1792 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*(i.numRestarts-1):]))
1793 0 :
1794 0 : i.readEntry()
1795 0 : i.clearCache()
1796 0 : i.cacheEntry()
1797 0 :
1798 0 : for i.nextOffset < i.restarts {
1799 0 : i.offset = i.nextOffset
1800 0 : i.readEntry()
1801 0 : i.cacheEntry()
1802 0 : }
1803 :
1804 0 : i.ikey.UserKey = i.key
1805 0 : return i.Valid()
1806 : }
1807 :
1808 : // Next implements internalIterator.Next, as documented in the pebble
1809 : // package.
1810 1 : func (i *RawIter) Next() bool {
1811 1 : i.offset = i.nextOffset
1812 1 : if !i.Valid() {
1813 1 : return false
1814 1 : }
1815 1 : i.loadEntry()
1816 1 : return true
1817 : }
1818 :
1819 : // Prev implements internalIterator.Prev, as documented in the pebble
1820 : // package.
1821 0 : func (i *RawIter) Prev() bool {
1822 0 : if n := len(i.cached) - 1; n > 0 && i.cached[n].offset == i.offset {
1823 0 : i.nextOffset = i.offset
1824 0 : e := &i.cached[n-1]
1825 0 : i.offset = e.offset
1826 0 : i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
1827 0 : i.ikey.UserKey = i.cachedBuf[e.keyStart:e.keyEnd]
1828 0 : i.cached = i.cached[:n]
1829 0 : return true
1830 0 : }
1831 :
1832 0 : if i.offset == 0 {
1833 0 : i.offset = -1
1834 0 : i.nextOffset = 0
1835 0 : return false
1836 0 : }
1837 :
1838 0 : targetOffset := i.offset
1839 0 : index := sort.Search(int(i.numRestarts), func(j int) bool {
1840 0 : offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
1841 0 : return offset >= targetOffset
1842 0 : })
1843 0 : i.offset = 0
1844 0 : if index > 0 {
1845 0 : i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
1846 0 : }
1847 :
1848 0 : i.readEntry()
1849 0 : i.clearCache()
1850 0 : i.cacheEntry()
1851 0 :
1852 0 : for i.nextOffset < targetOffset {
1853 0 : i.offset = i.nextOffset
1854 0 : i.readEntry()
1855 0 : i.cacheEntry()
1856 0 : }
1857 :
1858 0 : i.ikey.UserKey = i.key
1859 0 : return true
1860 : }
1861 :
1862 : // Key implements internalIterator.Key, as documented in the pebble package.
1863 1 : func (i *RawIter) Key() base.InternalKey {
1864 1 : return i.ikey
1865 1 : }
1866 :
1867 : // Value implements internalIterator.Value, as documented in the pebble
1868 : // package.
1869 1 : func (i *RawIter) Value() []byte {
1870 1 : return i.val
1871 1 : }
1872 :
1873 : // Valid implements internalIterator.Valid, as documented in the pebble
1874 : // package.
1875 1 : func (i *RawIter) Valid() bool {
1876 1 : return i.offset >= 0 && i.offset < i.restarts
1877 1 : }
1878 :
1879 : // Error implements internalIterator.Error, as documented in the pebble
1880 : // package.
1881 0 : func (i *RawIter) Error() error {
1882 0 : return nil
1883 0 : }
1884 :
1885 : // Close implements internalIterator.Close, as documented in the pebble
1886 : // package.
1887 1 : func (i *RawIter) Close() error {
1888 1 : i.val = nil
1889 1 : return nil
1890 1 : }
1891 :
1892 : // DebugTree is part of the InternalIterator interface.
1893 0 : func (i *RawIter) DebugTree(tp treeprinter.Node) {
1894 0 : tp.Childf("%T(%p)", i, i)
1895 0 : }
1896 :
1897 0 : func (i *RawIter) getRestart(idx int) int32 {
1898 0 : return int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*int32(idx):]))
1899 0 : }
1900 :
1901 0 : func (i *RawIter) isRestartPoint() bool {
1902 0 : j := sort.Search(int(i.numRestarts), func(j int) bool {
1903 0 : return i.getRestart(j) >= i.offset
1904 0 : })
1905 0 : return j < int(i.numRestarts) && i.getRestart(j) == i.offset
1906 : }
1907 :
1908 : // Describe describes the contents of a block, writing the description to w.
1909 : // It invokes fmtKV to describe each key-value pair.
1910 0 : func (i *RawIter) Describe(tp treeprinter.Node, fmtKV DescribeKV) {
1911 0 : var buf bytes.Buffer
1912 0 : for valid := i.First(); valid; valid = i.Next() {
1913 0 : enc := KVEncoding{
1914 0 : IsRestart: i.isRestartPoint(),
1915 0 : Offset: i.offset,
1916 0 : Length: int32(i.nextOffset - i.offset),
1917 0 : }
1918 0 : ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
1919 0 : enc.KeyShared, ptr = decodeVarint(ptr)
1920 0 : enc.KeyUnshared, ptr = decodeVarint(ptr)
1921 0 : enc.ValueLen, _ = decodeVarint(ptr)
1922 0 : buf.Reset()
1923 0 : fmtKV(&buf, &i.ikey, i.val, enc)
1924 0 : if i.isRestartPoint() {
1925 0 : buf.WriteString(" [restart]")
1926 0 : }
1927 0 : tp.Child(buf.String())
1928 : }
1929 0 : n := tp.Child("restart points")
1930 0 : // Format the restart points.
1931 0 : for j := 0; j < int(i.numRestarts); j++ {
1932 0 : offset := i.getRestart(j)
1933 0 : n.Childf("%05d [restart %d]", uint64(i.restarts+4*int32(j)), offset)
1934 0 : }
1935 : }
1936 :
1937 1 : func getBytes(ptr unsafe.Pointer, length int) []byte {
1938 1 : return (*[manual.MaxArrayLen]byte)(ptr)[:length:length]
1939 1 : }
1940 :
1941 1 : func decodeVarint(ptr unsafe.Pointer) (uint32, unsafe.Pointer) {
1942 1 : if a := *((*uint8)(ptr)); a < 128 {
1943 1 : return uint32(a),
1944 1 : unsafe.Pointer(uintptr(ptr) + 1)
1945 1 : } else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
1946 0 : return uint32(b)<<7 | uint32(a),
1947 0 : unsafe.Pointer(uintptr(ptr) + 2)
1948 0 : } else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
1949 0 : return uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1950 0 : unsafe.Pointer(uintptr(ptr) + 3)
1951 0 : } else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
1952 0 : return uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1953 0 : unsafe.Pointer(uintptr(ptr) + 4)
1954 0 : } else {
1955 0 : d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
1956 0 : return uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
1957 0 : unsafe.Pointer(uintptr(ptr) + 5)
1958 0 : }
1959 : }
|
