Line data Source code
1 : // Copyright 2012 The LevelDB-Go and Pebble Authors. All rights reserved. Use
2 : // of this source code is governed by a BSD-style license that can be found in
3 : // the LICENSE file.
4 :
5 : package pebble
6 :
7 : import (
8 : "bytes"
9 : "context"
10 : "encoding/binary"
11 : "fmt"
12 : "io"
13 : "math"
14 : "sort"
15 : "sync"
16 : "sync/atomic"
17 : "time"
18 : "unsafe"
19 :
20 : "github.com/cockroachdb/errors"
21 : "github.com/cockroachdb/pebble/batchrepr"
22 : "github.com/cockroachdb/pebble/internal/base"
23 : "github.com/cockroachdb/pebble/internal/batchskl"
24 : "github.com/cockroachdb/pebble/internal/humanize"
25 : "github.com/cockroachdb/pebble/internal/invariants"
26 : "github.com/cockroachdb/pebble/internal/keyspan"
27 : "github.com/cockroachdb/pebble/internal/private"
28 : "github.com/cockroachdb/pebble/internal/rangedel"
29 : "github.com/cockroachdb/pebble/internal/rangekey"
30 : "github.com/cockroachdb/pebble/internal/rawalloc"
31 : "github.com/cockroachdb/pebble/internal/treeprinter"
32 : )
33 :
34 : const (
35 : invalidBatchCount = 1<<32 - 1
36 : maxVarintLen32 = 5
37 :
38 : defaultBatchInitialSize = 1 << 10 // 1 KB
39 : defaultBatchMaxRetainedSize = 1 << 20 // 1 MB
40 : )
41 :
42 : // ErrNotIndexed means that a read operation on a batch failed because the
43 : // batch is not indexed and thus doesn't support reads.
44 : var ErrNotIndexed = errors.New("pebble: batch not indexed")
45 :
46 : // ErrInvalidBatch indicates that a batch is invalid or otherwise corrupted.
47 : var ErrInvalidBatch = batchrepr.ErrInvalidBatch
48 :
49 : // ErrBatchTooLarge indicates that a batch is invalid or otherwise corrupted.
50 : var ErrBatchTooLarge = base.MarkCorruptionError(errors.Newf("pebble: batch too large: >= %s", humanize.Bytes.Uint64(maxBatchSize)))
51 :
52 : // DeferredBatchOp represents a batch operation (eg. set, merge, delete) that is
53 : // being inserted into the batch. Indexing is not performed on the specified key
54 : // until Finish is called, hence the name deferred. This struct lets the caller
55 : // copy or encode keys/values directly into the batch representation instead of
56 : // copying into an intermediary buffer then having pebble.Batch copy off of it.
57 : type DeferredBatchOp struct {
58 : index *batchskl.Skiplist
59 :
60 : // Key and Value point to parts of the binary batch representation where
61 : // keys and values should be encoded/copied into. len(Key) and len(Value)
62 : // bytes must be copied into these slices respectively before calling
63 : // Finish(). Changing where these slices point to is not allowed.
64 : Key, Value []byte
65 : offset uint32
66 : }
67 :
68 : // Finish completes the addition of this batch operation, and adds it to the
69 : // index if necessary. Must be called once (and exactly once) keys/values
70 : // have been filled into Key and Value. Not calling Finish or not
71 : // copying/encoding keys will result in an incomplete index, and calling Finish
72 : // twice may result in a panic.
73 0 : func (d DeferredBatchOp) Finish() error {
74 0 : if d.index != nil {
75 0 : if err := d.index.Add(d.offset); err != nil {
76 0 : return err
77 0 : }
78 : }
79 0 : return nil
80 : }
81 :
82 : // A Batch is a sequence of Sets, Merges, Deletes, DeleteRanges, RangeKeySets,
83 : // RangeKeyUnsets, and/or RangeKeyDeletes that are applied atomically. Batch
84 : // implements the Reader interface, but only an indexed batch supports reading
85 : // (without error) via Get or NewIter. A non-indexed batch will return
86 : // ErrNotIndexed when read from. A batch is not safe for concurrent use, and
87 : // consumers should use a batch per goroutine or provide their own
88 : // synchronization.
89 : //
90 : // # Indexing
91 : //
92 : // Batches can be optionally indexed (see DB.NewIndexedBatch). An indexed batch
93 : // allows iteration via an Iterator (see Batch.NewIter). The iterator provides
94 : // a merged view of the operations in the batch and the underlying
95 : // database. This is implemented by treating the batch as an additional layer
96 : // in the LSM where every entry in the batch is considered newer than any entry
97 : // in the underlying database (batch entries have the InternalKeySeqNumBatch
98 : // bit set). By treating the batch as an additional layer in the LSM, iteration
99 : // supports all batch operations (i.e. Set, Merge, Delete, DeleteRange,
100 : // RangeKeySet, RangeKeyUnset, RangeKeyDelete) with minimal effort.
101 : //
102 : // The same key can be operated on multiple times in a batch, though only the
103 : // latest operation will be visible. For example, Put("a", "b"), Delete("a")
104 : // will cause the key "a" to not be visible in the batch. Put("a", "b"),
105 : // Put("a", "c") will cause a read of "a" to return the value "c".
106 : //
107 : // The batch index is implemented via an skiplist (internal/batchskl). While
108 : // the skiplist implementation is very fast, inserting into an indexed batch is
109 : // significantly slower than inserting into a non-indexed batch. Only use an
110 : // indexed batch if you require reading from it.
111 : //
112 : // # Atomic commit
113 : //
114 : // The operations in a batch are persisted by calling Batch.Commit which is
115 : // equivalent to calling DB.Apply(batch). A batch is committed atomically by
116 : // writing the internal batch representation to the WAL, adding all of the
117 : // batch operations to the memtable associated with the WAL, and then
118 : // incrementing the visible sequence number so that subsequent reads can see
119 : // the effects of the batch operations. If WriteOptions.Sync is true, a call to
120 : // Batch.Commit will guarantee that the batch is persisted to disk before
121 : // returning. See commitPipeline for more on the implementation details.
122 : //
123 : // # Large batches
124 : //
125 : // The size of a batch is limited only by available memory (be aware that
126 : // indexed batches require considerably additional memory for the skiplist
127 : // structure). A given WAL file has a single memtable associated with it (this
128 : // restriction could be removed, but doing so is onerous and complex). And a
129 : // memtable has a fixed size due to the underlying fixed size arena. Note that
130 : // this differs from RocksDB where a memtable can grow arbitrarily large using
131 : // a list of arena chunks. In RocksDB this is accomplished by storing pointers
132 : // in the arena memory, but that isn't possible in Go.
133 : //
134 : // During Batch.Commit, a batch which is larger than a threshold (>
135 : // MemTableSize/2) is wrapped in a flushableBatch and inserted into the queue
136 : // of memtables. A flushableBatch forces WAL to be rotated, but that happens
137 : // anyways when the memtable becomes full so this does not cause significant
138 : // WAL churn. Because the flushableBatch is readable as another layer in the
139 : // LSM, Batch.Commit returns as soon as the flushableBatch has been added to
140 : // the queue of memtables.
141 : //
142 : // Internally, a flushableBatch provides Iterator support by sorting the batch
143 : // contents (the batch is sorted once, when it is added to the memtable
144 : // queue). Sorting the batch contents and insertion of the contents into a
145 : // memtable have the same big-O time, but the constant factor dominates
146 : // here. Sorting is significantly faster and uses significantly less memory.
147 : //
148 : // # Internal representation
149 : //
150 : // The internal batch representation is a contiguous byte buffer with a fixed
151 : // 12-byte header, followed by a series of records.
152 : //
153 : // +-------------+------------+--- ... ---+
154 : // | SeqNum (8B) | Count (4B) | Entries |
155 : // +-------------+------------+--- ... ---+
156 : //
157 : // Each record has a 1-byte kind tag prefix, followed by 1 or 2 length prefixed
158 : // strings (varstring):
159 : //
160 : // +-----------+-----------------+-------------------+
161 : // | Kind (1B) | Key (varstring) | Value (varstring) |
162 : // +-----------+-----------------+-------------------+
163 : //
164 : // A varstring is a varint32 followed by N bytes of data. The Kind tags are
165 : // exactly those specified by InternalKeyKind. The following table shows the
166 : // format for records of each kind:
167 : //
168 : // InternalKeyKindDelete varstring
169 : // InternalKeyKindLogData varstring
170 : // InternalKeyKindIngestSST varstring
171 : // InternalKeyKindSet varstring varstring
172 : // InternalKeyKindMerge varstring varstring
173 : // InternalKeyKindRangeDelete varstring varstring
174 : // InternalKeyKindRangeKeySet varstring varstring
175 : // InternalKeyKindRangeKeyUnset varstring varstring
176 : // InternalKeyKindRangeKeyDelete varstring varstring
177 : //
178 : // The intuitive understanding here are that the arguments to Delete, Set,
179 : // Merge, DeleteRange and RangeKeyDelete are encoded into the batch. The
180 : // RangeKeySet and RangeKeyUnset operations are slightly more complicated,
181 : // encoding their end key, suffix and value [in the case of RangeKeySet] within
182 : // the Value varstring. For more information on the value encoding for
183 : // RangeKeySet and RangeKeyUnset, see the internal/rangekey package.
184 : //
185 : // The internal batch representation is the on disk format for a batch in the
186 : // WAL, and thus stable. New record kinds may be added, but the existing ones
187 : // will not be modified.
188 : type Batch struct {
189 : batchInternal
190 : applied atomic.Bool
191 : // lifecycle is used to negotiate the lifecycle of a Batch. A Batch and its
192 : // underlying batchInternal.data byte slice may be reused. There are two
193 : // mechanisms for reuse:
194 : //
195 : // 1. The caller may explicitly call [Batch.Reset] to reset the batch to be
196 : // empty (while retaining the underlying repr's buffer).
197 : // 2. The caller may call [Batch.Close], passing ownership off to Pebble,
198 : // which may reuse the batch's memory to service new callers to
199 : // [DB.NewBatch].
200 : //
201 : // There's a complication to reuse: When WAL failover is configured, the
202 : // Pebble commit pipeline may retain a pointer to the batch.data beyond the
203 : // return of [Batch.Commit]. The user of the Batch may commit their batch
204 : // and call Close or Reset before the commit pipeline is finished reading
205 : // the data slice. Recycling immediately would cause a data race.
206 : //
207 : // To resolve this data race, this [lifecycle] atomic is used to determine
208 : // safety and responsibility of reusing a batch. The low bits of the atomic
209 : // are used as a reference count (really just the lowest bit—in practice
210 : // there's only 1 code path that references). The [Batch] is passed into
211 : // [wal.Writer]'s WriteRecord method as a [RefCount] implementation. The
212 : // wal.Writer guarantees that if it will read [Batch.data] after the call to
213 : // WriteRecord returns, it will increment the reference count. When it's
214 : // complete, it'll unreference through invoking [Batch.Unref].
215 : //
216 : // When the committer of a batch indicates intent to recycle a Batch through
217 : // calling [Batch.Reset] or [Batch.Close], the lifecycle atomic is read. If
218 : // an outstanding reference remains, it's unsafe to reuse Batch.data yet. In
219 : // [Batch.Reset] the caller wants to reuse the [Batch] immediately, so we
220 : // discard b.data to recycle the struct but not the underlying byte slice.
221 : // In [Batch.Close], we set a special high bit [batchClosedBit] on lifecycle
222 : // that indicates that the user will not use [Batch] again and we're free to
223 : // recycle it when safe. When the commit pipeline eventually calls
224 : // [Batch.Unref], the [batchClosedBit] is noticed and the batch is
225 : // recycled.
226 : lifecycle atomic.Int32
227 : }
228 :
229 : // batchClosedBit is a bit stored on Batch.lifecycle to indicate that the user
230 : // called [Batch.Close] to release a Batch, but an open reference count
231 : // prevented immediate recycling.
232 : const batchClosedBit = 1 << 30
233 :
234 : // TODO(jackson): Hide the wal.RefCount implementation from the public Batch interface.
235 :
236 : // Ref implements wal.RefCount. If the WAL writer may need to read b.data after
237 : // it returns, it invokes Ref to increment the lifecycle's reference count. When
238 : // it's finished, it invokes Unref.
239 1 : func (b *Batch) Ref() {
240 1 : b.lifecycle.Add(+1)
241 1 : }
242 :
243 : // Unref implemets wal.RefCount.
244 1 : func (b *Batch) Unref() {
245 1 : if v := b.lifecycle.Add(-1); (v ^ batchClosedBit) == 0 {
246 1 : // The [batchClosedBit] high bit is set, and there are no outstanding
247 1 : // references. The user of the Batch called [Batch.Close], expecting the
248 1 : // batch to be recycled. However, our outstanding reference count
249 1 : // prevented recycling. As the last to dereference, we're now
250 1 : // responsible for releasing the batch.
251 1 : b.lifecycle.Store(0)
252 1 : b.release()
253 1 : }
254 : }
255 :
256 : // batchInternal contains the set of fields within Batch that are non-atomic and
257 : // capable of being reset using a *b = batchInternal{} struct copy.
258 : type batchInternal struct {
259 : // Data is the wire format of a batch's log entry:
260 : // - 8 bytes for a sequence number of the first batch element,
261 : // or zeroes if the batch has not yet been applied,
262 : // - 4 bytes for the count: the number of elements in the batch,
263 : // or "\xff\xff\xff\xff" if the batch is invalid,
264 : // - count elements, being:
265 : // - one byte for the kind
266 : // - the varint-string user key,
267 : // - the varint-string value (if kind != delete).
268 : // The sequence number and count are stored in little-endian order.
269 : //
270 : // The data field can be (but is not guaranteed to be) nil for new
271 : // batches. Large batches will set the data field to nil when committed as
272 : // the data has been moved to a flushableBatch and inserted into the queue of
273 : // memtables.
274 : data []byte
275 : comparer *base.Comparer
276 : opts batchOptions
277 :
278 : // An upper bound on required space to add this batch to a memtable.
279 : // Note that although batches are limited to 4 GiB in size, that limit
280 : // applies to len(data), not the memtable size. The upper bound on the
281 : // size of a memtable node is larger than the overhead of the batch's log
282 : // encoding, so memTableSize is larger than len(data) and may overflow a
283 : // uint32.
284 : memTableSize uint64
285 :
286 : // The db to which the batch will be committed. Do not change this field
287 : // after the batch has been created as it might invalidate internal state.
288 : // Batch.memTableSize is only refreshed if Batch.db is set. Setting db to
289 : // nil once it has been set implies that the Batch has encountered an error.
290 : db *DB
291 :
292 : // The count of records in the batch. This count will be stored in the batch
293 : // data whenever Repr() is called.
294 : count uint64
295 :
296 : // The count of range deletions in the batch. Updated every time a range
297 : // deletion is added.
298 : countRangeDels uint64
299 :
300 : // The count of range key sets, unsets and deletes in the batch. Updated
301 : // every time a RANGEKEYSET, RANGEKEYUNSET or RANGEKEYDEL key is added.
302 : countRangeKeys uint64
303 :
304 : // A deferredOp struct, stored in the Batch so that a pointer can be returned
305 : // from the *Deferred() methods rather than a value.
306 : deferredOp DeferredBatchOp
307 :
308 : // An optional skiplist keyed by offset into data of the entry.
309 : index *batchskl.Skiplist
310 : rangeDelIndex *batchskl.Skiplist
311 : rangeKeyIndex *batchskl.Skiplist
312 :
313 : // Fragmented range deletion tombstones. Cached the first time a range
314 : // deletion iterator is requested. The cache is invalidated whenever a new
315 : // range deletion is added to the batch. This cache can only be used when
316 : // opening an iterator to read at a batch sequence number >=
317 : // tombstonesSeqNum. This is the case for all new iterators created over a
318 : // batch but it's not the case for all cloned iterators.
319 : tombstones []keyspan.Span
320 : tombstonesSeqNum base.SeqNum
321 :
322 : // Fragmented range key spans. Cached the first time a range key iterator is
323 : // requested. The cache is invalidated whenever a new range key
324 : // (RangeKey{Set,Unset,Del}) is added to the batch. This cache can only be
325 : // used when opening an iterator to read at a batch sequence number >=
326 : // tombstonesSeqNum. This is the case for all new iterators created over a
327 : // batch but it's not the case for all cloned iterators.
328 : rangeKeys []keyspan.Span
329 : rangeKeysSeqNum base.SeqNum
330 :
331 : // The flushableBatch wrapper if the batch is too large to fit in the
332 : // memtable.
333 : flushable *flushableBatch
334 :
335 : // minimumFormatMajorVersion indicates the format major version required in
336 : // order to commit this batch. If an operation requires a particular format
337 : // major version, it ratchets the batch's minimumFormatMajorVersion. When
338 : // the batch is committed, this is validated against the database's current
339 : // format major version.
340 : minimumFormatMajorVersion FormatMajorVersion
341 :
342 : // Synchronous Apply uses the commit WaitGroup for both publishing the
343 : // seqnum and waiting for the WAL fsync (if needed). Asynchronous
344 : // ApplyNoSyncWait, which implies WriteOptions.Sync is true, uses the commit
345 : // WaitGroup for publishing the seqnum and the fsyncWait WaitGroup for
346 : // waiting for the WAL fsync.
347 : //
348 : // TODO(sumeer): if we find that ApplyNoSyncWait in conjunction with
349 : // SyncWait is causing higher memory usage because of the time duration
350 : // between when the sync is already done, and a goroutine calls SyncWait
351 : // (followed by Batch.Close), we could separate out {fsyncWait, commitErr}
352 : // into a separate struct that is allocated separately (using another
353 : // sync.Pool), and only that struct needs to outlive Batch.Close (which
354 : // could then be called immediately after ApplyNoSyncWait). commitStats
355 : // will also need to be in this separate struct.
356 : commit sync.WaitGroup
357 : fsyncWait sync.WaitGroup
358 :
359 : commitStats BatchCommitStats
360 :
361 : commitErr error
362 :
363 : // Position bools together to reduce the sizeof the struct.
364 :
365 : // ingestedSSTBatch indicates that the batch contains one or more key kinds
366 : // of InternalKeyKindIngestSST. If the batch contains key kinds of IngestSST
367 : // then it will only contain key kinds of IngestSST.
368 : ingestedSSTBatch bool
369 :
370 : // committing is set to true when a batch begins to commit. It's used to
371 : // ensure the batch is not mutated concurrently. It is not an atomic
372 : // deliberately, so as to avoid the overhead on batch mutations. This is
373 : // okay, because under correct usage this field will never be accessed
374 : // concurrently. It's only under incorrect usage the memory accesses of this
375 : // variable may violate memory safety. Since we don't use atomics here,
376 : // false negatives are possible.
377 : committing bool
378 : }
379 :
380 : // BatchCommitStats exposes stats related to committing a batch.
381 : //
382 : // NB: there is no Pebble internal tracing (using LoggerAndTracer) of slow
383 : // batch commits. The caller can use these stats to do their own tracing as
384 : // needed.
385 : type BatchCommitStats struct {
386 : // TotalDuration is the time spent in DB.{Apply,ApplyNoSyncWait} or
387 : // Batch.Commit, plus the time waiting in Batch.SyncWait. If there is a gap
388 : // between calling ApplyNoSyncWait and calling SyncWait, that gap could
389 : // include some duration in which real work was being done for the commit
390 : // and will not be included here. This missing time is considered acceptable
391 : // since the goal of these stats is to understand user-facing latency.
392 : //
393 : // TotalDuration includes time spent in various queues both inside Pebble
394 : // and outside Pebble (I/O queues, goroutine scheduler queue, mutex wait
395 : // etc.). For some of these queues (which we consider important) the wait
396 : // times are included below -- these expose low-level implementation detail
397 : // and are meant for expert diagnosis and subject to change. There may be
398 : // unaccounted time after subtracting those values from TotalDuration.
399 : TotalDuration time.Duration
400 : // SemaphoreWaitDuration is the wait time for semaphores in
401 : // commitPipeline.Commit.
402 : SemaphoreWaitDuration time.Duration
403 : // WALQueueWaitDuration is the wait time for allocating memory blocks in the
404 : // LogWriter (due to the LogWriter not writing fast enough). At the moment
405 : // this is duration is always zero because a single WAL will allow
406 : // allocating memory blocks up to the entire memtable size. In the future,
407 : // we may pipeline WALs and bound the WAL queued blocks separately, so this
408 : // field is preserved for that possibility.
409 : WALQueueWaitDuration time.Duration
410 : // MemTableWriteStallDuration is the wait caused by a write stall due to too
411 : // many memtables (due to not flushing fast enough).
412 : MemTableWriteStallDuration time.Duration
413 : // L0ReadAmpWriteStallDuration is the wait caused by a write stall due to
414 : // high read amplification in L0 (due to not compacting fast enough out of
415 : // L0).
416 : L0ReadAmpWriteStallDuration time.Duration
417 : // WALRotationDuration is the wait time for WAL rotation, which includes
418 : // syncing and closing the old WAL and creating (or reusing) a new one.
419 : WALRotationDuration time.Duration
420 : // CommitWaitDuration is the wait for publishing the seqnum plus the
421 : // duration for the WAL sync (if requested). The former should be tiny and
422 : // one can assume that this is all due to the WAL sync.
423 : CommitWaitDuration time.Duration
424 : }
425 :
426 : var _ Reader = (*Batch)(nil)
427 : var _ Writer = (*Batch)(nil)
428 :
429 : var batchPool = sync.Pool{
430 1 : New: func() interface{} {
431 1 : return &Batch{}
432 1 : },
433 : }
434 :
435 : type indexedBatch struct {
436 : batch Batch
437 : index batchskl.Skiplist
438 : }
439 :
440 : var indexedBatchPool = sync.Pool{
441 1 : New: func() interface{} {
442 1 : return &indexedBatch{}
443 1 : },
444 : }
445 :
446 1 : func newBatch(db *DB, opts ...BatchOption) *Batch {
447 1 : b := batchPool.Get().(*Batch)
448 1 : b.db = db
449 1 : b.opts.ensureDefaults()
450 1 : for _, opt := range opts {
451 0 : opt(&b.opts)
452 0 : }
453 1 : return b
454 : }
455 :
456 0 : func newBatchWithSize(db *DB, size int, opts ...BatchOption) *Batch {
457 0 : b := newBatch(db, opts...)
458 0 : if cap(b.data) < size {
459 0 : b.data = rawalloc.New(0, size)
460 0 : }
461 0 : return b
462 : }
463 :
464 1 : func newIndexedBatch(db *DB, comparer *Comparer) *Batch {
465 1 : i := indexedBatchPool.Get().(*indexedBatch)
466 1 : i.batch.comparer = comparer
467 1 : i.batch.db = db
468 1 : i.batch.index = &i.index
469 1 : i.batch.index.Init(&i.batch.data, comparer.Compare, comparer.AbbreviatedKey)
470 1 : i.batch.opts.ensureDefaults()
471 1 : return &i.batch
472 1 : }
473 :
474 0 : func newIndexedBatchWithSize(db *DB, comparer *Comparer, size int) *Batch {
475 0 : b := newIndexedBatch(db, comparer)
476 0 : if cap(b.data) < size {
477 0 : b.data = rawalloc.New(0, size)
478 0 : }
479 0 : return b
480 : }
481 :
482 : // nextSeqNum returns the batch "sequence number" that will be given to the next
483 : // key written to the batch. During iteration keys within an indexed batch are
484 : // given a sequence number consisting of their offset within the batch combined
485 : // with the base.SeqNumBatchBit bit. These sequence numbers are only
486 : // used during iteration, and the keys are assigned ordinary sequence numbers
487 : // when the batch is committed.
488 1 : func (b *Batch) nextSeqNum() base.SeqNum {
489 1 : return base.SeqNum(len(b.data)) | base.SeqNumBatchBit
490 1 : }
491 :
492 1 : func (b *Batch) release() {
493 1 : if b.db == nil {
494 1 : // The batch was not created using newBatch or newIndexedBatch, or an error
495 1 : // was encountered. We don't try to reuse batches that encountered an error
496 1 : // because they might be stuck somewhere in the system and attempting to
497 1 : // reuse such batches is a recipe for onerous debugging sessions. Instead,
498 1 : // let the GC do its job.
499 1 : return
500 1 : }
501 1 : b.db = nil
502 1 :
503 1 : // NB: This is ugly (it would be cleaner if we could just assign a Batch{}),
504 1 : // but necessary so that we can use atomic.StoreUint32 for the Batch.applied
505 1 : // field. Without using an atomic to clear that field the Go race detector
506 1 : // complains.
507 1 : b.reset()
508 1 : b.comparer = nil
509 1 :
510 1 : if b.index == nil {
511 1 : batchPool.Put(b)
512 1 : } else {
513 1 : b.index, b.rangeDelIndex, b.rangeKeyIndex = nil, nil, nil
514 1 : indexedBatchPool.Put((*indexedBatch)(unsafe.Pointer(b)))
515 1 : }
516 : }
517 :
518 1 : func (b *Batch) refreshMemTableSize() error {
519 1 : b.memTableSize = 0
520 1 : if len(b.data) < batchrepr.HeaderLen {
521 0 : return nil
522 0 : }
523 :
524 1 : b.countRangeDels = 0
525 1 : b.countRangeKeys = 0
526 1 : b.minimumFormatMajorVersion = 0
527 1 : for r := b.Reader(); ; {
528 1 : kind, key, value, ok, err := r.Next()
529 1 : if !ok {
530 1 : if err != nil {
531 0 : return err
532 0 : }
533 1 : break
534 : }
535 1 : switch kind {
536 1 : case InternalKeyKindRangeDelete:
537 1 : b.countRangeDels++
538 1 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
539 1 : b.countRangeKeys++
540 1 : case InternalKeyKindSet, InternalKeyKindDelete, InternalKeyKindMerge, InternalKeyKindSingleDelete, InternalKeyKindSetWithDelete:
541 : // fallthrough
542 1 : case InternalKeyKindDeleteSized:
543 1 : if b.minimumFormatMajorVersion < FormatDeleteSizedAndObsolete {
544 1 : b.minimumFormatMajorVersion = FormatDeleteSizedAndObsolete
545 1 : }
546 0 : case InternalKeyKindLogData:
547 0 : // LogData does not contribute to memtable size.
548 0 : continue
549 1 : case InternalKeyKindIngestSST:
550 1 : if b.minimumFormatMajorVersion < FormatFlushableIngest {
551 1 : b.minimumFormatMajorVersion = FormatFlushableIngest
552 1 : }
553 : // This key kind doesn't contribute to the memtable size.
554 1 : continue
555 1 : case InternalKeyKindExcise:
556 1 : if b.minimumFormatMajorVersion < FormatFlushableIngestExcises {
557 1 : b.minimumFormatMajorVersion = FormatFlushableIngestExcises
558 1 : }
559 : // This key kind doesn't contribute to the memtable size.
560 1 : continue
561 0 : default:
562 0 : // Note In some circumstances this might be temporary memory
563 0 : // corruption that can be recovered by discarding the batch and
564 0 : // trying again. In other cases, the batch repr might've been
565 0 : // already persisted elsewhere, and we'll loop continuously trying
566 0 : // to commit the same corrupted batch. The caller is responsible for
567 0 : // distinguishing.
568 0 : return errors.Wrapf(ErrInvalidBatch, "unrecognized kind %v", kind)
569 : }
570 1 : b.memTableSize += memTableEntrySize(len(key), len(value))
571 : }
572 1 : return nil
573 : }
574 :
575 : // Apply the operations contained in the batch to the receiver batch.
576 : //
577 : // It is safe to modify the contents of the arguments after Apply returns.
578 : //
579 : // Apply returns ErrInvalidBatch if the provided batch is invalid in any way.
580 1 : func (b *Batch) Apply(batch *Batch, _ *WriteOptions) error {
581 1 : if b.ingestedSSTBatch {
582 0 : panic("pebble: invalid batch application")
583 : }
584 1 : if len(batch.data) == 0 {
585 1 : return nil
586 1 : }
587 1 : if len(batch.data) < batchrepr.HeaderLen {
588 0 : return ErrInvalidBatch
589 0 : }
590 :
591 1 : offset := len(b.data)
592 1 : if offset == 0 {
593 1 : b.init(offset)
594 1 : offset = batchrepr.HeaderLen
595 1 : }
596 1 : b.data = append(b.data, batch.data[batchrepr.HeaderLen:]...)
597 1 :
598 1 : b.setCount(b.Count() + batch.Count())
599 1 :
600 1 : if b.db != nil || b.index != nil {
601 1 : // Only iterate over the new entries if we need to track memTableSize or in
602 1 : // order to update the index.
603 1 : for iter := batchrepr.Reader(b.data[offset:]); len(iter) > 0; {
604 1 : offset := uintptr(unsafe.Pointer(&iter[0])) - uintptr(unsafe.Pointer(&b.data[0]))
605 1 : kind, key, value, ok, err := iter.Next()
606 1 : if !ok {
607 0 : if err != nil {
608 0 : return err
609 0 : }
610 0 : break
611 : }
612 1 : switch kind {
613 0 : case InternalKeyKindRangeDelete:
614 0 : b.countRangeDels++
615 0 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
616 0 : b.countRangeKeys++
617 0 : case InternalKeyKindIngestSST, InternalKeyKindExcise:
618 0 : panic("pebble: invalid key kind for batch")
619 0 : case InternalKeyKindLogData:
620 0 : // LogData does not contribute to memtable size.
621 0 : continue
622 : case InternalKeyKindSet, InternalKeyKindDelete, InternalKeyKindMerge,
623 1 : InternalKeyKindSingleDelete, InternalKeyKindSetWithDelete, InternalKeyKindDeleteSized:
624 : // fallthrough
625 0 : default:
626 0 : // Note In some circumstances this might be temporary memory
627 0 : // corruption that can be recovered by discarding the batch and
628 0 : // trying again. In other cases, the batch repr might've been
629 0 : // already persisted elsewhere, and we'll loop continuously
630 0 : // trying to commit the same corrupted batch. The caller is
631 0 : // responsible for distinguishing.
632 0 : return errors.Wrapf(ErrInvalidBatch, "unrecognized kind %v", kind)
633 : }
634 1 : if b.index != nil {
635 1 : var err error
636 1 : switch kind {
637 0 : case InternalKeyKindRangeDelete:
638 0 : b.tombstones = nil
639 0 : b.tombstonesSeqNum = 0
640 0 : if b.rangeDelIndex == nil {
641 0 : b.rangeDelIndex = batchskl.NewSkiplist(&b.data, b.comparer.Compare, b.comparer.AbbreviatedKey)
642 0 : }
643 0 : err = b.rangeDelIndex.Add(uint32(offset))
644 0 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
645 0 : b.rangeKeys = nil
646 0 : b.rangeKeysSeqNum = 0
647 0 : if b.rangeKeyIndex == nil {
648 0 : b.rangeKeyIndex = batchskl.NewSkiplist(&b.data, b.comparer.Compare, b.comparer.AbbreviatedKey)
649 0 : }
650 0 : err = b.rangeKeyIndex.Add(uint32(offset))
651 1 : default:
652 1 : err = b.index.Add(uint32(offset))
653 : }
654 1 : if err != nil {
655 0 : return err
656 0 : }
657 : }
658 1 : b.memTableSize += memTableEntrySize(len(key), len(value))
659 : }
660 : }
661 1 : return nil
662 : }
663 :
664 : // Get gets the value for the given key. It returns ErrNotFound if the Batch
665 : // does not contain the key.
666 : //
667 : // The caller should not modify the contents of the returned slice, but it is
668 : // safe to modify the contents of the argument after Get returns. The returned
669 : // slice will remain valid until the returned Closer is closed. On success, the
670 : // caller MUST call closer.Close() or a memory leak will occur.
671 1 : func (b *Batch) Get(key []byte) ([]byte, io.Closer, error) {
672 1 : if b.index == nil {
673 0 : return nil, nil, ErrNotIndexed
674 0 : }
675 1 : return b.db.getInternal(key, b, nil /* snapshot */)
676 : }
677 :
678 1 : func (b *Batch) prepareDeferredKeyValueRecord(keyLen, valueLen int, kind InternalKeyKind) {
679 1 : if b.committing {
680 0 : panic("pebble: batch already committing")
681 : }
682 1 : if len(b.data) == 0 {
683 1 : b.init(keyLen + valueLen + 2*binary.MaxVarintLen64 + batchrepr.HeaderLen)
684 1 : }
685 1 : b.count++
686 1 : b.memTableSize += memTableEntrySize(keyLen, valueLen)
687 1 :
688 1 : pos := len(b.data)
689 1 : b.deferredOp.offset = uint32(pos)
690 1 : b.grow(1 + 2*maxVarintLen32 + keyLen + valueLen)
691 1 : b.data[pos] = byte(kind)
692 1 : pos++
693 1 :
694 1 : {
695 1 : // TODO(peter): Manually inlined version binary.PutUvarint(). This is 20%
696 1 : // faster on BenchmarkBatchSet on go1.13. Remove if go1.14 or future
697 1 : // versions show this to not be a performance win.
698 1 : x := uint32(keyLen)
699 1 : for x >= 0x80 {
700 0 : b.data[pos] = byte(x) | 0x80
701 0 : x >>= 7
702 0 : pos++
703 0 : }
704 1 : b.data[pos] = byte(x)
705 1 : pos++
706 : }
707 :
708 1 : b.deferredOp.Key = b.data[pos : pos+keyLen]
709 1 : pos += keyLen
710 1 :
711 1 : {
712 1 : // TODO(peter): Manually inlined version binary.PutUvarint(). This is 20%
713 1 : // faster on BenchmarkBatchSet on go1.13. Remove if go1.14 or future
714 1 : // versions show this to not be a performance win.
715 1 : x := uint32(valueLen)
716 1 : for x >= 0x80 {
717 0 : b.data[pos] = byte(x) | 0x80
718 0 : x >>= 7
719 0 : pos++
720 0 : }
721 1 : b.data[pos] = byte(x)
722 1 : pos++
723 : }
724 :
725 1 : b.deferredOp.Value = b.data[pos : pos+valueLen]
726 1 : // Shrink data since varints may be shorter than the upper bound.
727 1 : b.data = b.data[:pos+valueLen]
728 : }
729 :
730 1 : func (b *Batch) prepareDeferredKeyRecord(keyLen int, kind InternalKeyKind) {
731 1 : if b.committing {
732 0 : panic("pebble: batch already committing")
733 : }
734 1 : if len(b.data) == 0 {
735 1 : b.init(keyLen + binary.MaxVarintLen64 + batchrepr.HeaderLen)
736 1 : }
737 1 : b.count++
738 1 : b.memTableSize += memTableEntrySize(keyLen, 0)
739 1 :
740 1 : pos := len(b.data)
741 1 : b.deferredOp.offset = uint32(pos)
742 1 : b.grow(1 + maxVarintLen32 + keyLen)
743 1 : b.data[pos] = byte(kind)
744 1 : pos++
745 1 :
746 1 : {
747 1 : // TODO(peter): Manually inlined version binary.PutUvarint(). Remove if
748 1 : // go1.13 or future versions show this to not be a performance win. See
749 1 : // BenchmarkBatchSet.
750 1 : x := uint32(keyLen)
751 1 : for x >= 0x80 {
752 0 : b.data[pos] = byte(x) | 0x80
753 0 : x >>= 7
754 0 : pos++
755 0 : }
756 1 : b.data[pos] = byte(x)
757 1 : pos++
758 : }
759 :
760 1 : b.deferredOp.Key = b.data[pos : pos+keyLen]
761 1 : b.deferredOp.Value = nil
762 1 :
763 1 : // Shrink data since varint may be shorter than the upper bound.
764 1 : b.data = b.data[:pos+keyLen]
765 : }
766 :
767 : // AddInternalKey allows the caller to add an internal key of point key or range
768 : // key kinds (but not RangeDelete) to a batch. Passing in an internal key of
769 : // kind RangeDelete will result in a panic. Note that the seqnum in the internal
770 : // key is effectively ignored, even though the Kind is preserved. This is
771 : // because the batch format does not allow for a per-key seqnum to be specified,
772 : // only a batch-wide one.
773 : //
774 : // Note that non-indexed keys (IngestKeyKind{LogData,IngestSST}) are not
775 : // supported with this method as they require specialized logic.
776 1 : func (b *Batch) AddInternalKey(key *base.InternalKey, value []byte, _ *WriteOptions) error {
777 1 : keyLen := len(key.UserKey)
778 1 : hasValue := false
779 1 : switch kind := key.Kind(); kind {
780 0 : case InternalKeyKindRangeDelete:
781 0 : panic("unexpected range delete in AddInternalKey")
782 0 : case InternalKeyKindSingleDelete, InternalKeyKindDelete:
783 0 : b.prepareDeferredKeyRecord(keyLen, kind)
784 0 : b.deferredOp.index = b.index
785 1 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
786 1 : b.prepareDeferredKeyValueRecord(keyLen, len(value), kind)
787 1 : hasValue = true
788 1 : b.incrementRangeKeysCount()
789 0 : default:
790 0 : b.prepareDeferredKeyValueRecord(keyLen, len(value), kind)
791 0 : hasValue = true
792 0 : b.deferredOp.index = b.index
793 : }
794 1 : copy(b.deferredOp.Key, key.UserKey)
795 1 : if hasValue {
796 1 : copy(b.deferredOp.Value, value)
797 1 : }
798 :
799 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
800 : // in go1.13 will remove the need for this.
801 1 : if b.index != nil {
802 0 : if err := b.index.Add(b.deferredOp.offset); err != nil {
803 0 : return err
804 0 : }
805 : }
806 1 : return nil
807 : }
808 :
809 : // Set adds an action to the batch that sets the key to map to the value.
810 : //
811 : // It is safe to modify the contents of the arguments after Set returns.
812 1 : func (b *Batch) Set(key, value []byte, _ *WriteOptions) error {
813 1 : deferredOp := b.SetDeferred(len(key), len(value))
814 1 : copy(deferredOp.Key, key)
815 1 : copy(deferredOp.Value, value)
816 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
817 1 : // in go1.13 will remove the need for this.
818 1 : if b.index != nil {
819 1 : if err := b.index.Add(deferredOp.offset); err != nil {
820 0 : return err
821 0 : }
822 : }
823 1 : return nil
824 : }
825 :
826 : // SetDeferred is similar to Set in that it adds a set operation to the batch,
827 : // except it only takes in key/value lengths instead of complete slices,
828 : // letting the caller encode into those objects and then call Finish() on the
829 : // returned object.
830 1 : func (b *Batch) SetDeferred(keyLen, valueLen int) *DeferredBatchOp {
831 1 : b.prepareDeferredKeyValueRecord(keyLen, valueLen, InternalKeyKindSet)
832 1 : b.deferredOp.index = b.index
833 1 : return &b.deferredOp
834 1 : }
835 :
836 : // Merge adds an action to the batch that merges the value at key with the new
837 : // value. The details of the merge are dependent upon the configured merge
838 : // operator.
839 : //
840 : // It is safe to modify the contents of the arguments after Merge returns.
841 1 : func (b *Batch) Merge(key, value []byte, _ *WriteOptions) error {
842 1 : deferredOp := b.MergeDeferred(len(key), len(value))
843 1 : copy(deferredOp.Key, key)
844 1 : copy(deferredOp.Value, value)
845 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
846 1 : // in go1.13 will remove the need for this.
847 1 : if b.index != nil {
848 1 : if err := b.index.Add(deferredOp.offset); err != nil {
849 0 : return err
850 0 : }
851 : }
852 1 : return nil
853 : }
854 :
855 : // MergeDeferred is similar to Merge in that it adds a merge operation to the
856 : // batch, except it only takes in key/value lengths instead of complete slices,
857 : // letting the caller encode into those objects and then call Finish() on the
858 : // returned object.
859 1 : func (b *Batch) MergeDeferred(keyLen, valueLen int) *DeferredBatchOp {
860 1 : b.prepareDeferredKeyValueRecord(keyLen, valueLen, InternalKeyKindMerge)
861 1 : b.deferredOp.index = b.index
862 1 : return &b.deferredOp
863 1 : }
864 :
865 : // Delete adds an action to the batch that deletes the entry for key.
866 : //
867 : // It is safe to modify the contents of the arguments after Delete returns.
868 1 : func (b *Batch) Delete(key []byte, _ *WriteOptions) error {
869 1 : deferredOp := b.DeleteDeferred(len(key))
870 1 : copy(deferredOp.Key, key)
871 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
872 1 : // in go1.13 will remove the need for this.
873 1 : if b.index != nil {
874 1 : if err := b.index.Add(deferredOp.offset); err != nil {
875 0 : return err
876 0 : }
877 : }
878 1 : return nil
879 : }
880 :
881 : // DeleteDeferred is similar to Delete in that it adds a delete operation to
882 : // the batch, except it only takes in key/value lengths instead of complete
883 : // slices, letting the caller encode into those objects and then call Finish()
884 : // on the returned object.
885 1 : func (b *Batch) DeleteDeferred(keyLen int) *DeferredBatchOp {
886 1 : b.prepareDeferredKeyRecord(keyLen, InternalKeyKindDelete)
887 1 : b.deferredOp.index = b.index
888 1 : return &b.deferredOp
889 1 : }
890 :
891 : // DeleteSized behaves identically to Delete, but takes an additional
892 : // argument indicating the size of the value being deleted. DeleteSized
893 : // should be preferred when the caller has the expectation that there exists
894 : // a single internal KV pair for the key (eg, the key has not been
895 : // overwritten recently), and the caller knows the size of its value.
896 : //
897 : // DeleteSized will record the value size within the tombstone and use it to
898 : // inform compaction-picking heuristics which strive to reduce space
899 : // amplification in the LSM. This "calling your shot" mechanic allows the
900 : // storage engine to more accurately estimate and reduce space amplification.
901 : //
902 : // It is safe to modify the contents of the arguments after DeleteSized
903 : // returns.
904 1 : func (b *Batch) DeleteSized(key []byte, deletedValueSize uint32, _ *WriteOptions) error {
905 1 : deferredOp := b.DeleteSizedDeferred(len(key), deletedValueSize)
906 1 : copy(b.deferredOp.Key, key)
907 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Check if in a
908 1 : // later Go release this is unnecessary.
909 1 : if b.index != nil {
910 1 : if err := b.index.Add(deferredOp.offset); err != nil {
911 0 : return err
912 0 : }
913 : }
914 1 : return nil
915 : }
916 :
917 : // DeleteSizedDeferred is similar to DeleteSized in that it adds a sized delete
918 : // operation to the batch, except it only takes in key length instead of a
919 : // complete key slice, letting the caller encode into the DeferredBatchOp.Key
920 : // slice and then call Finish() on the returned object.
921 1 : func (b *Batch) DeleteSizedDeferred(keyLen int, deletedValueSize uint32) *DeferredBatchOp {
922 1 : if b.minimumFormatMajorVersion < FormatDeleteSizedAndObsolete {
923 1 : b.minimumFormatMajorVersion = FormatDeleteSizedAndObsolete
924 1 : }
925 :
926 : // Encode the sum of the key length and the value in the value.
927 1 : v := uint64(deletedValueSize) + uint64(keyLen)
928 1 :
929 1 : // Encode `v` as a varint.
930 1 : var buf [binary.MaxVarintLen64]byte
931 1 : n := 0
932 1 : {
933 1 : x := v
934 1 : for x >= 0x80 {
935 0 : buf[n] = byte(x) | 0x80
936 0 : x >>= 7
937 0 : n++
938 0 : }
939 1 : buf[n] = byte(x)
940 1 : n++
941 : }
942 :
943 : // NB: In batch entries and sstable entries, values are stored as
944 : // varstrings. Here, the value is itself a simple varint. This results in an
945 : // unnecessary double layer of encoding:
946 : // varint(n) varint(deletedValueSize)
947 : // The first varint will always be 1-byte, since a varint-encoded uint64
948 : // will never exceed 128 bytes. This unnecessary extra byte and wrapping is
949 : // preserved to avoid special casing across the database, and in particular
950 : // in sstable block decoding which is performance sensitive.
951 1 : b.prepareDeferredKeyValueRecord(keyLen, n, InternalKeyKindDeleteSized)
952 1 : b.deferredOp.index = b.index
953 1 : copy(b.deferredOp.Value, buf[:n])
954 1 : return &b.deferredOp
955 : }
956 :
957 : // SingleDelete adds an action to the batch that single deletes the entry for key.
958 : // See Writer.SingleDelete for more details on the semantics of SingleDelete.
959 : //
960 : // It is safe to modify the contents of the arguments after SingleDelete returns.
961 1 : func (b *Batch) SingleDelete(key []byte, _ *WriteOptions) error {
962 1 : deferredOp := b.SingleDeleteDeferred(len(key))
963 1 : copy(deferredOp.Key, key)
964 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
965 1 : // in go1.13 will remove the need for this.
966 1 : if b.index != nil {
967 1 : if err := b.index.Add(deferredOp.offset); err != nil {
968 0 : return err
969 0 : }
970 : }
971 1 : return nil
972 : }
973 :
974 : // SingleDeleteDeferred is similar to SingleDelete in that it adds a single delete
975 : // operation to the batch, except it only takes in key/value lengths instead of
976 : // complete slices, letting the caller encode into those objects and then call
977 : // Finish() on the returned object.
978 1 : func (b *Batch) SingleDeleteDeferred(keyLen int) *DeferredBatchOp {
979 1 : b.prepareDeferredKeyRecord(keyLen, InternalKeyKindSingleDelete)
980 1 : b.deferredOp.index = b.index
981 1 : return &b.deferredOp
982 1 : }
983 :
984 : // DeleteRange deletes all of the point keys (and values) in the range
985 : // [start,end) (inclusive on start, exclusive on end). DeleteRange does NOT
986 : // delete overlapping range keys (eg, keys set via RangeKeySet).
987 : //
988 : // It is safe to modify the contents of the arguments after DeleteRange
989 : // returns.
990 1 : func (b *Batch) DeleteRange(start, end []byte, _ *WriteOptions) error {
991 1 : deferredOp := b.DeleteRangeDeferred(len(start), len(end))
992 1 : copy(deferredOp.Key, start)
993 1 : copy(deferredOp.Value, end)
994 1 : // TODO(peter): Manually inline DeferredBatchOp.Finish(). Mid-stack inlining
995 1 : // in go1.13 will remove the need for this.
996 1 : if deferredOp.index != nil {
997 1 : if err := deferredOp.index.Add(deferredOp.offset); err != nil {
998 0 : return err
999 0 : }
1000 : }
1001 1 : return nil
1002 : }
1003 :
1004 : // DeleteRangeDeferred is similar to DeleteRange in that it adds a delete range
1005 : // operation to the batch, except it only takes in key lengths instead of
1006 : // complete slices, letting the caller encode into those objects and then call
1007 : // Finish() on the returned object. Note that DeferredBatchOp.Key should be
1008 : // populated with the start key, and DeferredBatchOp.Value should be populated
1009 : // with the end key.
1010 1 : func (b *Batch) DeleteRangeDeferred(startLen, endLen int) *DeferredBatchOp {
1011 1 : b.prepareDeferredKeyValueRecord(startLen, endLen, InternalKeyKindRangeDelete)
1012 1 : b.countRangeDels++
1013 1 : if b.index != nil {
1014 1 : b.tombstones = nil
1015 1 : b.tombstonesSeqNum = 0
1016 1 : // Range deletions are rare, so we lazily allocate the index for them.
1017 1 : if b.rangeDelIndex == nil {
1018 1 : b.rangeDelIndex = batchskl.NewSkiplist(&b.data, b.comparer.Compare, b.comparer.AbbreviatedKey)
1019 1 : }
1020 1 : b.deferredOp.index = b.rangeDelIndex
1021 : }
1022 1 : return &b.deferredOp
1023 : }
1024 :
1025 : // RangeKeySet sets a range key mapping the key range [start, end) at the MVCC
1026 : // timestamp suffix to value. The suffix is optional. If any portion of the key
1027 : // range [start, end) is already set by a range key with the same suffix value,
1028 : // RangeKeySet overrides it.
1029 : //
1030 : // It is safe to modify the contents of the arguments after RangeKeySet returns.
1031 1 : func (b *Batch) RangeKeySet(start, end, suffix, value []byte, _ *WriteOptions) error {
1032 1 : if invariants.Enabled && b.db != nil {
1033 1 : // RangeKeySet is only supported on prefix keys.
1034 1 : if b.db.opts.Comparer.Split(start) != len(start) {
1035 0 : panic("RangeKeySet called with suffixed start key")
1036 : }
1037 1 : if b.db.opts.Comparer.Split(end) != len(end) {
1038 0 : panic("RangeKeySet called with suffixed end key")
1039 : }
1040 : }
1041 1 : suffixValues := [1]rangekey.SuffixValue{{Suffix: suffix, Value: value}}
1042 1 : internalValueLen := rangekey.EncodedSetValueLen(end, suffixValues[:])
1043 1 :
1044 1 : deferredOp := b.rangeKeySetDeferred(len(start), internalValueLen)
1045 1 : copy(deferredOp.Key, start)
1046 1 : n := rangekey.EncodeSetValue(deferredOp.Value, end, suffixValues[:])
1047 1 : if n != internalValueLen {
1048 0 : panic("unexpected internal value length mismatch")
1049 : }
1050 :
1051 : // Manually inline DeferredBatchOp.Finish().
1052 1 : if deferredOp.index != nil {
1053 1 : if err := deferredOp.index.Add(deferredOp.offset); err != nil {
1054 0 : return err
1055 0 : }
1056 : }
1057 1 : return nil
1058 : }
1059 :
1060 1 : func (b *Batch) rangeKeySetDeferred(startLen, internalValueLen int) *DeferredBatchOp {
1061 1 : b.prepareDeferredKeyValueRecord(startLen, internalValueLen, InternalKeyKindRangeKeySet)
1062 1 : b.incrementRangeKeysCount()
1063 1 : return &b.deferredOp
1064 1 : }
1065 :
1066 1 : func (b *Batch) incrementRangeKeysCount() {
1067 1 : b.countRangeKeys++
1068 1 : if b.index != nil {
1069 1 : b.rangeKeys = nil
1070 1 : b.rangeKeysSeqNum = 0
1071 1 : // Range keys are rare, so we lazily allocate the index for them.
1072 1 : if b.rangeKeyIndex == nil {
1073 1 : b.rangeKeyIndex = batchskl.NewSkiplist(&b.data, b.comparer.Compare, b.comparer.AbbreviatedKey)
1074 1 : }
1075 1 : b.deferredOp.index = b.rangeKeyIndex
1076 : }
1077 : }
1078 :
1079 : // RangeKeyUnset removes a range key mapping the key range [start, end) at the
1080 : // MVCC timestamp suffix. The suffix may be omitted to remove an unsuffixed
1081 : // range key. RangeKeyUnset only removes portions of range keys that fall within
1082 : // the [start, end) key span, and only range keys with suffixes that exactly
1083 : // match the unset suffix.
1084 : //
1085 : // It is safe to modify the contents of the arguments after RangeKeyUnset
1086 : // returns.
1087 1 : func (b *Batch) RangeKeyUnset(start, end, suffix []byte, _ *WriteOptions) error {
1088 1 : if invariants.Enabled && b.db != nil {
1089 1 : // RangeKeyUnset is only supported on prefix keys.
1090 1 : if b.db.opts.Comparer.Split(start) != len(start) {
1091 0 : panic("RangeKeyUnset called with suffixed start key")
1092 : }
1093 1 : if b.db.opts.Comparer.Split(end) != len(end) {
1094 0 : panic("RangeKeyUnset called with suffixed end key")
1095 : }
1096 : }
1097 1 : suffixes := [1][]byte{suffix}
1098 1 : internalValueLen := rangekey.EncodedUnsetValueLen(end, suffixes[:])
1099 1 :
1100 1 : deferredOp := b.rangeKeyUnsetDeferred(len(start), internalValueLen)
1101 1 : copy(deferredOp.Key, start)
1102 1 : n := rangekey.EncodeUnsetValue(deferredOp.Value, end, suffixes[:])
1103 1 : if n != internalValueLen {
1104 0 : panic("unexpected internal value length mismatch")
1105 : }
1106 :
1107 : // Manually inline DeferredBatchOp.Finish()
1108 1 : if deferredOp.index != nil {
1109 1 : if err := deferredOp.index.Add(deferredOp.offset); err != nil {
1110 0 : return err
1111 0 : }
1112 : }
1113 1 : return nil
1114 : }
1115 :
1116 1 : func (b *Batch) rangeKeyUnsetDeferred(startLen, internalValueLen int) *DeferredBatchOp {
1117 1 : b.prepareDeferredKeyValueRecord(startLen, internalValueLen, InternalKeyKindRangeKeyUnset)
1118 1 : b.incrementRangeKeysCount()
1119 1 : return &b.deferredOp
1120 1 : }
1121 :
1122 : // RangeKeyDelete deletes all of the range keys in the range [start,end)
1123 : // (inclusive on start, exclusive on end). It does not delete point keys (for
1124 : // that use DeleteRange). RangeKeyDelete removes all range keys within the
1125 : // bounds, including those with or without suffixes.
1126 : //
1127 : // It is safe to modify the contents of the arguments after RangeKeyDelete
1128 : // returns.
1129 1 : func (b *Batch) RangeKeyDelete(start, end []byte, _ *WriteOptions) error {
1130 1 : if invariants.Enabled && b.db != nil {
1131 1 : // RangeKeyDelete is only supported on prefix keys.
1132 1 : if b.db.opts.Comparer.Split(start) != len(start) {
1133 0 : panic("RangeKeyDelete called with suffixed start key")
1134 : }
1135 1 : if b.db.opts.Comparer.Split(end) != len(end) {
1136 0 : panic("RangeKeyDelete called with suffixed end key")
1137 : }
1138 : }
1139 1 : deferredOp := b.RangeKeyDeleteDeferred(len(start), len(end))
1140 1 : copy(deferredOp.Key, start)
1141 1 : copy(deferredOp.Value, end)
1142 1 : // Manually inline DeferredBatchOp.Finish().
1143 1 : if deferredOp.index != nil {
1144 1 : if err := deferredOp.index.Add(deferredOp.offset); err != nil {
1145 0 : return err
1146 0 : }
1147 : }
1148 1 : return nil
1149 : }
1150 :
1151 : // RangeKeyDeleteDeferred is similar to RangeKeyDelete in that it adds an
1152 : // operation to delete range keys to the batch, except it only takes in key
1153 : // lengths instead of complete slices, letting the caller encode into those
1154 : // objects and then call Finish() on the returned object. Note that
1155 : // DeferredBatchOp.Key should be populated with the start key, and
1156 : // DeferredBatchOp.Value should be populated with the end key.
1157 1 : func (b *Batch) RangeKeyDeleteDeferred(startLen, endLen int) *DeferredBatchOp {
1158 1 : b.prepareDeferredKeyValueRecord(startLen, endLen, InternalKeyKindRangeKeyDelete)
1159 1 : b.incrementRangeKeysCount()
1160 1 : return &b.deferredOp
1161 1 : }
1162 :
1163 : // LogData adds the specified to the batch. The data will be written to the
1164 : // WAL, but not added to memtables or sstables. Log data is never indexed,
1165 : // which makes it useful for testing WAL performance.
1166 : //
1167 : // It is safe to modify the contents of the argument after LogData returns.
1168 1 : func (b *Batch) LogData(data []byte, _ *WriteOptions) error {
1169 1 : origCount, origMemTableSize := b.count, b.memTableSize
1170 1 : b.prepareDeferredKeyRecord(len(data), InternalKeyKindLogData)
1171 1 : copy(b.deferredOp.Key, data)
1172 1 : // Since LogData only writes to the WAL and does not affect the memtable, we
1173 1 : // restore b.count and b.memTableSize to their origin values. Note that
1174 1 : // Batch.count only refers to records that are added to the memtable.
1175 1 : b.count, b.memTableSize = origCount, origMemTableSize
1176 1 : return nil
1177 1 : }
1178 :
1179 : // IngestSST adds the FileNum for an sstable to the batch. The data will only be
1180 : // written to the WAL (not added to memtables or sstables).
1181 1 : func (b *Batch) ingestSST(fileNum base.FileNum) {
1182 1 : if b.Empty() {
1183 1 : b.ingestedSSTBatch = true
1184 1 : } else if !b.ingestedSSTBatch {
1185 0 : // Batch contains other key kinds.
1186 0 : panic("pebble: invalid call to ingestSST")
1187 : }
1188 :
1189 1 : origMemTableSize := b.memTableSize
1190 1 : var buf [binary.MaxVarintLen64]byte
1191 1 : length := binary.PutUvarint(buf[:], uint64(fileNum))
1192 1 : b.prepareDeferredKeyRecord(length, InternalKeyKindIngestSST)
1193 1 : copy(b.deferredOp.Key, buf[:length])
1194 1 : // Since IngestSST writes only to the WAL and does not affect the memtable,
1195 1 : // we restore b.memTableSize to its original value. Note that Batch.count
1196 1 : // is not reset because for the InternalKeyKindIngestSST the count is the
1197 1 : // number of sstable paths which have been added to the batch.
1198 1 : b.memTableSize = origMemTableSize
1199 1 : b.minimumFormatMajorVersion = FormatFlushableIngest
1200 : }
1201 :
1202 : // Excise adds the excise span for a flushable ingest containing an excise. The data
1203 : // will only be written to the WAL (not added to memtables or sstables).
1204 1 : func (b *Batch) excise(start, end []byte) {
1205 1 : if b.Empty() {
1206 1 : b.ingestedSSTBatch = true
1207 1 : } else if !b.ingestedSSTBatch {
1208 0 : // Batch contains other key kinds.
1209 0 : panic("pebble: invalid call to excise")
1210 : }
1211 :
1212 1 : origMemTableSize := b.memTableSize
1213 1 : b.prepareDeferredKeyValueRecord(len(start), len(end), InternalKeyKindExcise)
1214 1 : copy(b.deferredOp.Key, start)
1215 1 : copy(b.deferredOp.Value, end)
1216 1 : // Since excise writes only to the WAL and does not affect the memtable,
1217 1 : // we restore b.memTableSize to its original value. Note that Batch.count
1218 1 : // is not reset because for the InternalKeyKindIngestSST/Excise the count
1219 1 : // is the number of sstable paths which have been added to the batch.
1220 1 : b.memTableSize = origMemTableSize
1221 1 : b.minimumFormatMajorVersion = FormatFlushableIngestExcises
1222 : }
1223 :
1224 : // Empty returns true if the batch is empty, and false otherwise.
1225 1 : func (b *Batch) Empty() bool {
1226 1 : return batchrepr.IsEmpty(b.data)
1227 1 : }
1228 :
1229 : // Len returns the current size of the batch in bytes.
1230 0 : func (b *Batch) Len() int {
1231 0 : return max(batchrepr.HeaderLen, len(b.data))
1232 0 : }
1233 :
1234 : // Repr returns the underlying batch representation. It is not safe to modify
1235 : // the contents. Reset() will not change the contents of the returned value,
1236 : // though any other mutation operation may do so.
1237 1 : func (b *Batch) Repr() []byte {
1238 1 : if len(b.data) == 0 {
1239 0 : b.init(batchrepr.HeaderLen)
1240 0 : }
1241 1 : batchrepr.SetCount(b.data, b.Count())
1242 1 : return b.data
1243 : }
1244 :
1245 : // SetRepr sets the underlying batch representation. The batch takes ownership
1246 : // of the supplied slice. It is not safe to modify it afterwards until the
1247 : // Batch is no longer in use.
1248 : //
1249 : // SetRepr may return ErrInvalidBatch if the supplied slice fails to decode in
1250 : // any way. It will not return an error in any other circumstance.
1251 1 : func (b *Batch) SetRepr(data []byte) error {
1252 1 : h, ok := batchrepr.ReadHeader(data)
1253 1 : if !ok {
1254 0 : return ErrInvalidBatch
1255 0 : }
1256 1 : b.data = data
1257 1 : b.count = uint64(h.Count)
1258 1 : var err error
1259 1 : if b.db != nil {
1260 1 : // Only track memTableSize for batches that will be committed to the DB.
1261 1 : err = b.refreshMemTableSize()
1262 1 : }
1263 1 : return err
1264 : }
1265 :
1266 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
1267 : // return false). The iterator can be positioned via a call to SeekGE,
1268 : // SeekPrefixGE, SeekLT, First or Last. Only indexed batches support iterators.
1269 : //
1270 : // The returned Iterator observes all of the Batch's existing mutations, but no
1271 : // later mutations. Its view can be refreshed via RefreshBatchSnapshot or
1272 : // SetOptions().
1273 1 : func (b *Batch) NewIter(o *IterOptions) (*Iterator, error) {
1274 1 : return b.NewIterWithContext(context.Background(), o)
1275 1 : }
1276 :
1277 : // NewIterWithContext is like NewIter, and additionally accepts a context for
1278 : // tracing.
1279 1 : func (b *Batch) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
1280 1 : if b.index == nil {
1281 0 : return nil, ErrNotIndexed
1282 0 : }
1283 1 : return b.db.newIter(ctx, b, newIterOpts{}, o), nil
1284 : }
1285 :
1286 : // NewBatchOnlyIter constructs an iterator that only reads the contents of the
1287 : // batch, and does not overlay the batch mutations on top of the DB state.
1288 : //
1289 : // The returned Iterator observes all of the Batch's existing mutations, but
1290 : // no later mutations. Its view can be refreshed via RefreshBatchSnapshot or
1291 : // SetOptions().
1292 0 : func (b *Batch) NewBatchOnlyIter(ctx context.Context, o *IterOptions) (*Iterator, error) {
1293 0 : if b.index == nil {
1294 0 : return nil, ErrNotIndexed
1295 0 : }
1296 0 : return b.db.newIter(ctx, b, newIterOpts{batch: batchIterOpts{batchOnly: true}}, o), nil
1297 : }
1298 :
1299 : // newInternalIter creates a new internalIterator that iterates over the
1300 : // contents of the batch.
1301 1 : func (b *Batch) newInternalIter(o *IterOptions) *batchIter {
1302 1 : iter := &batchIter{}
1303 1 : b.initInternalIter(o, iter)
1304 1 : return iter
1305 1 : }
1306 :
1307 1 : func (b *Batch) initInternalIter(o *IterOptions, iter *batchIter) {
1308 1 : *iter = batchIter{
1309 1 : batch: b,
1310 1 : iter: b.index.NewIter(o.GetLowerBound(), o.GetUpperBound()),
1311 1 : // NB: We explicitly do not propagate the batch snapshot to the point
1312 1 : // key iterator. Filtering point keys within the batch iterator can
1313 1 : // cause pathological behavior where a batch iterator advances
1314 1 : // significantly farther than necessary filtering many batch keys that
1315 1 : // are not visible at the batch sequence number. Instead, the merging
1316 1 : // iterator enforces bounds.
1317 1 : //
1318 1 : // For example, consider an engine that contains the committed keys
1319 1 : // 'bar' and 'bax', with no keys between them. Consider a batch
1320 1 : // containing keys 1,000 keys within the range [a,z]. All of the
1321 1 : // batch keys were added to the batch after the iterator was
1322 1 : // constructed, so they are not visible to the iterator. A call to
1323 1 : // SeekGE('bax') would seek the LSM iterators and discover the key
1324 1 : // 'bax'. It would also seek the batch iterator, landing on the key
1325 1 : // 'baz' but discover it that it's not visible. The batch iterator would
1326 1 : // next through the rest of the batch's keys, only to discover there are
1327 1 : // no visible keys greater than or equal to 'bax'.
1328 1 : //
1329 1 : // Filtering these batch points within the merging iterator ensures that
1330 1 : // the batch iterator never needs to iterate beyond 'baz', because it
1331 1 : // already found a smaller, visible key 'bax'.
1332 1 : snapshot: base.SeqNumMax,
1333 1 : }
1334 1 : }
1335 :
1336 1 : func (b *Batch) newRangeDelIter(o *IterOptions, batchSnapshot base.SeqNum) *keyspan.Iter {
1337 1 : // Construct an iterator even if rangeDelIndex is nil, because it is allowed
1338 1 : // to refresh later, so we need the container to exist.
1339 1 : iter := new(keyspan.Iter)
1340 1 : b.initRangeDelIter(o, iter, batchSnapshot)
1341 1 : return iter
1342 1 : }
1343 :
1344 1 : func (b *Batch) initRangeDelIter(_ *IterOptions, iter *keyspan.Iter, batchSnapshot base.SeqNum) {
1345 1 : if b.rangeDelIndex == nil {
1346 1 : iter.Init(b.comparer.Compare, nil)
1347 1 : return
1348 1 : }
1349 :
1350 : // Fragment the range tombstones the first time a range deletion iterator is
1351 : // requested. The cached tombstones are invalidated if another range
1352 : // deletion tombstone is added to the batch. This cache is only guaranteed
1353 : // to be correct if we're opening an iterator to read at a batch sequence
1354 : // number at least as high as tombstonesSeqNum. The cache is guaranteed to
1355 : // include all tombstones up to tombstonesSeqNum, and if any additional
1356 : // tombstones were added after that sequence number the cache would've been
1357 : // cleared.
1358 1 : nextSeqNum := b.nextSeqNum()
1359 1 : if b.tombstones != nil && b.tombstonesSeqNum <= batchSnapshot {
1360 1 : iter.Init(b.comparer.Compare, b.tombstones)
1361 1 : return
1362 1 : }
1363 :
1364 1 : tombstones := make([]keyspan.Span, 0, b.countRangeDels)
1365 1 : frag := &keyspan.Fragmenter{
1366 1 : Cmp: b.comparer.Compare,
1367 1 : Format: b.comparer.FormatKey,
1368 1 : Emit: func(s keyspan.Span) {
1369 1 : tombstones = append(tombstones, s)
1370 1 : },
1371 : }
1372 1 : it := &batchIter{
1373 1 : batch: b,
1374 1 : iter: b.rangeDelIndex.NewIter(nil, nil),
1375 1 : snapshot: batchSnapshot,
1376 1 : }
1377 1 : fragmentRangeDels(frag, it, int(b.countRangeDels))
1378 1 : iter.Init(b.comparer.Compare, tombstones)
1379 1 :
1380 1 : // If we just read all the tombstones in the batch (eg, batchSnapshot was
1381 1 : // set to b.nextSeqNum()), then cache the tombstones so that a subsequent
1382 1 : // call to initRangeDelIter may use them without refragmenting.
1383 1 : if nextSeqNum == batchSnapshot {
1384 1 : b.tombstones = tombstones
1385 1 : b.tombstonesSeqNum = nextSeqNum
1386 1 : }
1387 : }
1388 :
1389 1 : func fragmentRangeDels(frag *keyspan.Fragmenter, it internalIterator, count int) {
1390 1 : // The memory management here is a bit subtle. The keys and values returned
1391 1 : // by the iterator are slices in Batch.data. Thus the fragmented tombstones
1392 1 : // are slices within Batch.data. If additional entries are added to the
1393 1 : // Batch, Batch.data may be reallocated. The references in the fragmented
1394 1 : // tombstones will remain valid, pointing into the old Batch.data. GC for
1395 1 : // the win.
1396 1 :
1397 1 : // Use a single []keyspan.Key buffer to avoid allocating many
1398 1 : // individual []keyspan.Key slices with a single element each.
1399 1 : keyBuf := make([]keyspan.Key, 0, count)
1400 1 : for kv := it.First(); kv != nil; kv = it.Next() {
1401 1 : s := rangedel.Decode(kv.K, kv.InPlaceValue(), keyBuf)
1402 1 : keyBuf = s.Keys[len(s.Keys):]
1403 1 :
1404 1 : // Set a fixed capacity to avoid accidental overwriting.
1405 1 : s.Keys = s.Keys[:len(s.Keys):len(s.Keys)]
1406 1 : frag.Add(s)
1407 1 : }
1408 1 : frag.Finish()
1409 : }
1410 :
1411 1 : func (b *Batch) newRangeKeyIter(o *IterOptions, batchSnapshot base.SeqNum) *keyspan.Iter {
1412 1 : // Construct an iterator even if rangeKeyIndex is nil, because it is allowed
1413 1 : // to refresh later, so we need the container to exist.
1414 1 : iter := new(keyspan.Iter)
1415 1 : b.initRangeKeyIter(o, iter, batchSnapshot)
1416 1 : return iter
1417 1 : }
1418 :
1419 1 : func (b *Batch) initRangeKeyIter(_ *IterOptions, iter *keyspan.Iter, batchSnapshot base.SeqNum) {
1420 1 : if b.rangeKeyIndex == nil {
1421 1 : iter.Init(b.comparer.Compare, nil)
1422 1 : return
1423 1 : }
1424 :
1425 : // Fragment the range keys the first time a range key iterator is requested.
1426 : // The cached spans are invalidated if another range key is added to the
1427 : // batch. This cache is only guaranteed to be correct if we're opening an
1428 : // iterator to read at a batch sequence number at least as high as
1429 : // rangeKeysSeqNum. The cache is guaranteed to include all range keys up to
1430 : // rangeKeysSeqNum, and if any additional range keys were added after that
1431 : // sequence number the cache would've been cleared.
1432 1 : nextSeqNum := b.nextSeqNum()
1433 1 : if b.rangeKeys != nil && b.rangeKeysSeqNum <= batchSnapshot {
1434 1 : iter.Init(b.comparer.Compare, b.rangeKeys)
1435 1 : return
1436 1 : }
1437 :
1438 1 : rangeKeys := make([]keyspan.Span, 0, b.countRangeKeys)
1439 1 : frag := &keyspan.Fragmenter{
1440 1 : Cmp: b.comparer.Compare,
1441 1 : Format: b.comparer.FormatKey,
1442 1 : Emit: func(s keyspan.Span) {
1443 1 : rangeKeys = append(rangeKeys, s)
1444 1 : },
1445 : }
1446 1 : it := &batchIter{
1447 1 : batch: b,
1448 1 : iter: b.rangeKeyIndex.NewIter(nil, nil),
1449 1 : snapshot: batchSnapshot,
1450 1 : }
1451 1 : fragmentRangeKeys(frag, it, int(b.countRangeKeys))
1452 1 : iter.Init(b.comparer.Compare, rangeKeys)
1453 1 :
1454 1 : // If we just read all the range keys in the batch (eg, batchSnapshot was
1455 1 : // set to b.nextSeqNum()), then cache the range keys so that a subsequent
1456 1 : // call to initRangeKeyIter may use them without refragmenting.
1457 1 : if nextSeqNum == batchSnapshot {
1458 1 : b.rangeKeys = rangeKeys
1459 1 : b.rangeKeysSeqNum = nextSeqNum
1460 1 : }
1461 : }
1462 :
1463 1 : func fragmentRangeKeys(frag *keyspan.Fragmenter, it internalIterator, count int) error {
1464 1 : // The memory management here is a bit subtle. The keys and values
1465 1 : // returned by the iterator are slices in Batch.data. Thus the
1466 1 : // fragmented key spans are slices within Batch.data. If additional
1467 1 : // entries are added to the Batch, Batch.data may be reallocated. The
1468 1 : // references in the fragmented keys will remain valid, pointing into
1469 1 : // the old Batch.data. GC for the win.
1470 1 :
1471 1 : // Use a single []keyspan.Key buffer to avoid allocating many
1472 1 : // individual []keyspan.Key slices with a single element each.
1473 1 : keyBuf := make([]keyspan.Key, 0, count)
1474 1 : for kv := it.First(); kv != nil; kv = it.Next() {
1475 1 : s, err := rangekey.Decode(kv.K, kv.InPlaceValue(), keyBuf)
1476 1 : if err != nil {
1477 0 : return err
1478 0 : }
1479 1 : keyBuf = s.Keys[len(s.Keys):]
1480 1 :
1481 1 : // Set a fixed capacity to avoid accidental overwriting.
1482 1 : s.Keys = s.Keys[:len(s.Keys):len(s.Keys)]
1483 1 : frag.Add(s)
1484 : }
1485 1 : frag.Finish()
1486 1 : return nil
1487 : }
1488 :
1489 : // Commit applies the batch to its parent writer.
1490 1 : func (b *Batch) Commit(o *WriteOptions) error {
1491 1 : return b.db.Apply(b, o)
1492 1 : }
1493 :
1494 : // Close closes the batch without committing it.
1495 1 : func (b *Batch) Close() error {
1496 1 : // The storage engine commit pipeline may retain a pointer to b.data beyond
1497 1 : // when Commit() returns. This is possible when configured for WAL failover;
1498 1 : // we don't know if we might need to read the batch data again until the
1499 1 : // batch has been durably synced [even if the committer doesn't care to wait
1500 1 : // for the sync and Sync()=false].
1501 1 : //
1502 1 : // We still want to recycle these batches. The b.lifecycle atomic negotiates
1503 1 : // the batch's lifecycle. If the commit pipeline still might read b.data,
1504 1 : // b.lifecycle will be nonzeroed [the low bits hold a ref count].
1505 1 : for {
1506 1 : v := b.lifecycle.Load()
1507 1 : switch {
1508 1 : case v == 0:
1509 1 : // A zero value indicates that the commit pipeline has no
1510 1 : // outstanding references to the batch. The commit pipeline is
1511 1 : // required to acquire a ref synchronously, so there is no risk that
1512 1 : // the commit pipeline will grab a ref after the call to release. We
1513 1 : // can simply release the batch.
1514 1 : b.release()
1515 1 : return nil
1516 0 : case (v & batchClosedBit) != 0:
1517 0 : // The batch has a batchClosedBit: This batch has already been closed.
1518 0 : return ErrClosed
1519 1 : default:
1520 1 : // There's an outstanding reference. Set the batch released bit so
1521 1 : // that the commit pipeline knows it should release the batch when
1522 1 : // it unrefs.
1523 1 : if b.lifecycle.CompareAndSwap(v, v|batchClosedBit) {
1524 1 : return nil
1525 1 : }
1526 : // CAS Failed—this indicates the outstanding reference just
1527 : // decremented (or the caller illegally closed the batch twice).
1528 : // Loop to reload.
1529 : }
1530 : }
1531 : }
1532 :
1533 : // Indexed returns true if the batch is indexed (i.e. supports read
1534 : // operations).
1535 1 : func (b *Batch) Indexed() bool {
1536 1 : return b.index != nil
1537 1 : }
1538 :
1539 : // init ensures that the batch data slice is initialized to meet the
1540 : // minimum required size and allocates space for the batch header.
1541 1 : func (b *Batch) init(size int) {
1542 1 : b.opts.ensureDefaults()
1543 1 : n := b.opts.initialSizeBytes
1544 1 : for n < size {
1545 0 : n *= 2
1546 0 : }
1547 1 : if cap(b.data) < n {
1548 1 : b.data = rawalloc.New(batchrepr.HeaderLen, n)
1549 1 : }
1550 1 : b.data = b.data[:batchrepr.HeaderLen]
1551 1 : clear(b.data) // Zero the sequence number in the header
1552 : }
1553 :
1554 : // Reset resets the batch for reuse. The underlying byte slice (that is
1555 : // returned by Repr()) may not be modified. It is only necessary to call this
1556 : // method if a batch is explicitly being reused. Close automatically takes are
1557 : // of releasing resources when appropriate for batches that are internally
1558 : // being reused.
1559 0 : func (b *Batch) Reset() {
1560 0 : // In some configurations (WAL failover) the commit pipeline may retain
1561 0 : // b.data beyond a call to commit the batch. When this happens, b.lifecycle
1562 0 : // is nonzero (see the comment above b.lifecycle). In this case it's unsafe
1563 0 : // to mutate b.data, so we discard it. Note that Reset must not be called on
1564 0 : // a closed batch, so v > 0 implies a non-zero ref count and not
1565 0 : // batchClosedBit being set.
1566 0 : if v := b.lifecycle.Load(); v > 0 {
1567 0 : b.data = nil
1568 0 : }
1569 0 : b.reset()
1570 : }
1571 :
1572 1 : func (b *Batch) reset() {
1573 1 : // Zero out the struct, retaining only the fields necessary for manual
1574 1 : // reuse.
1575 1 : b.batchInternal = batchInternal{
1576 1 : data: b.data,
1577 1 : comparer: b.comparer,
1578 1 : opts: b.opts,
1579 1 : index: b.index,
1580 1 : db: b.db,
1581 1 : }
1582 1 : b.applied.Store(false)
1583 1 : if b.data != nil {
1584 1 : if cap(b.data) > b.opts.maxRetainedSizeBytes {
1585 0 : // If the capacity of the buffer is larger than our maximum
1586 0 : // retention size, don't re-use it. Let it be GC-ed instead.
1587 0 : // This prevents the memory from an unusually large batch from
1588 0 : // being held on to indefinitely.
1589 0 : b.data = nil
1590 1 : } else {
1591 1 : // Otherwise, reset the buffer for re-use.
1592 1 : b.data = b.data[:batchrepr.HeaderLen]
1593 1 : clear(b.data)
1594 1 : }
1595 : }
1596 1 : if b.index != nil {
1597 1 : b.index.Init(&b.data, b.comparer.Compare, b.comparer.AbbreviatedKey)
1598 1 : }
1599 : }
1600 :
1601 1 : func (b *Batch) grow(n int) {
1602 1 : newSize := len(b.data) + n
1603 1 : if uint64(newSize) >= maxBatchSize {
1604 0 : panic(ErrBatchTooLarge)
1605 : }
1606 1 : if newSize > cap(b.data) {
1607 0 : newCap := 2 * cap(b.data)
1608 0 : for newCap < newSize {
1609 0 : newCap *= 2
1610 0 : }
1611 0 : newData := rawalloc.New(len(b.data), newCap)
1612 0 : copy(newData, b.data)
1613 0 : b.data = newData
1614 : }
1615 1 : b.data = b.data[:newSize]
1616 : }
1617 :
1618 1 : func (b *Batch) setSeqNum(seqNum base.SeqNum) {
1619 1 : batchrepr.SetSeqNum(b.data, seqNum)
1620 1 : }
1621 :
1622 : // SeqNum returns the batch sequence number which is applied to the first
1623 : // record in the batch. The sequence number is incremented for each subsequent
1624 : // record. It returns zero if the batch is empty.
1625 1 : func (b *Batch) SeqNum() base.SeqNum {
1626 1 : if len(b.data) == 0 {
1627 0 : b.init(batchrepr.HeaderLen)
1628 0 : }
1629 1 : return batchrepr.ReadSeqNum(b.data)
1630 : }
1631 :
1632 1 : func (b *Batch) setCount(v uint32) {
1633 1 : b.count = uint64(v)
1634 1 : }
1635 :
1636 : // Count returns the count of memtable-modifying operations in this batch. All
1637 : // operations with the except of LogData increment this count. For IngestSSTs,
1638 : // count is only used to indicate the number of SSTs ingested in the record, the
1639 : // batch isn't applied to the memtable.
1640 1 : func (b *Batch) Count() uint32 {
1641 1 : if b.count > math.MaxUint32 {
1642 0 : panic(batchrepr.ErrInvalidBatch)
1643 : }
1644 1 : return uint32(b.count)
1645 : }
1646 :
1647 : // Reader returns a batchrepr.Reader for the current batch contents. If the
1648 : // batch is mutated, the new entries will not be visible to the reader.
1649 1 : func (b *Batch) Reader() batchrepr.Reader {
1650 1 : if len(b.data) == 0 {
1651 1 : b.init(batchrepr.HeaderLen)
1652 1 : }
1653 1 : return batchrepr.Read(b.data)
1654 : }
1655 :
1656 : // SyncWait is to be used in conjunction with DB.ApplyNoSyncWait.
1657 1 : func (b *Batch) SyncWait() error {
1658 1 : now := time.Now()
1659 1 : b.fsyncWait.Wait()
1660 1 : if b.commitErr != nil {
1661 0 : b.db = nil // prevent batch reuse on error
1662 0 : }
1663 1 : waitDuration := time.Since(now)
1664 1 : b.commitStats.CommitWaitDuration += waitDuration
1665 1 : b.commitStats.TotalDuration += waitDuration
1666 1 : return b.commitErr
1667 : }
1668 :
1669 : // CommitStats returns stats related to committing the batch. Should be called
1670 : // after Batch.Commit, DB.Apply. If DB.ApplyNoSyncWait is used, should be
1671 : // called after Batch.SyncWait.
1672 0 : func (b *Batch) CommitStats() BatchCommitStats {
1673 0 : return b.commitStats
1674 0 : }
1675 :
1676 : // Note: batchIter mirrors the implementation of flushableBatchIter. Keep the
1677 : // two in sync.
1678 : type batchIter struct {
1679 : batch *Batch
1680 : iter batchskl.Iterator
1681 : kv base.InternalKV
1682 : err error
1683 : // snapshot holds a batch "sequence number" at which the batch is being
1684 : // read. This sequence number has the InternalKeySeqNumBatch bit set, so it
1685 : // encodes an offset within the batch. Only batch entries earlier than the
1686 : // offset are visible during iteration.
1687 : snapshot base.SeqNum
1688 : }
1689 :
1690 : // batchIter implements the base.InternalIterator interface.
1691 : var _ base.InternalIterator = (*batchIter)(nil)
1692 :
1693 0 : func (i *batchIter) String() string {
1694 0 : return "batch"
1695 0 : }
1696 :
1697 1 : func (i *batchIter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
1698 1 : // Ignore TrySeekUsingNext if the view of the batch changed.
1699 1 : if flags.TrySeekUsingNext() && flags.BatchJustRefreshed() {
1700 0 : flags = flags.DisableTrySeekUsingNext()
1701 0 : }
1702 :
1703 1 : i.err = nil // clear cached iteration error
1704 1 : ikey := i.iter.SeekGE(key, flags)
1705 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1706 0 : ikey = i.iter.Next()
1707 0 : }
1708 1 : if ikey == nil {
1709 1 : i.kv = base.InternalKV{}
1710 1 : return nil
1711 1 : }
1712 1 : i.kv.K = *ikey
1713 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1714 1 : return &i.kv
1715 : }
1716 :
1717 1 : func (i *batchIter) SeekPrefixGE(prefix, key []byte, flags base.SeekGEFlags) *base.InternalKV {
1718 1 : kv := i.SeekGE(key, flags)
1719 1 : if kv == nil {
1720 1 : return nil
1721 1 : }
1722 : // If the key doesn't have the sought prefix, return nil.
1723 1 : if !bytes.Equal(i.batch.comparer.Split.Prefix(kv.K.UserKey), prefix) {
1724 1 : i.kv = base.InternalKV{}
1725 1 : return nil
1726 1 : }
1727 1 : return kv
1728 : }
1729 :
1730 1 : func (i *batchIter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
1731 1 : i.err = nil // clear cached iteration error
1732 1 : ikey := i.iter.SeekLT(key)
1733 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1734 0 : ikey = i.iter.Prev()
1735 0 : }
1736 1 : if ikey == nil {
1737 1 : i.kv = base.InternalKV{}
1738 1 : return nil
1739 1 : }
1740 1 : i.kv.K = *ikey
1741 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1742 1 : return &i.kv
1743 : }
1744 :
1745 1 : func (i *batchIter) First() *base.InternalKV {
1746 1 : i.err = nil // clear cached iteration error
1747 1 : ikey := i.iter.First()
1748 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1749 0 : ikey = i.iter.Next()
1750 0 : }
1751 1 : if ikey == nil {
1752 1 : i.kv = base.InternalKV{}
1753 1 : return nil
1754 1 : }
1755 1 : i.kv.K = *ikey
1756 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1757 1 : return &i.kv
1758 : }
1759 :
1760 1 : func (i *batchIter) Last() *base.InternalKV {
1761 1 : i.err = nil // clear cached iteration error
1762 1 : ikey := i.iter.Last()
1763 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1764 0 : ikey = i.iter.Prev()
1765 0 : }
1766 1 : if ikey == nil {
1767 1 : i.kv = base.InternalKV{}
1768 1 : return nil
1769 1 : }
1770 1 : i.kv.K = *ikey
1771 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1772 1 : return &i.kv
1773 : }
1774 :
1775 1 : func (i *batchIter) Next() *base.InternalKV {
1776 1 : ikey := i.iter.Next()
1777 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1778 0 : ikey = i.iter.Next()
1779 0 : }
1780 1 : if ikey == nil {
1781 1 : i.kv = base.InternalKV{}
1782 1 : return nil
1783 1 : }
1784 1 : i.kv.K = *ikey
1785 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1786 1 : return &i.kv
1787 : }
1788 :
1789 0 : func (i *batchIter) NextPrefix(succKey []byte) *base.InternalKV {
1790 0 : // Because NextPrefix was invoked `succKey` must be ≥ the key at i's current
1791 0 : // position. Seek the arena iterator using TrySeekUsingNext.
1792 0 : ikey := i.iter.SeekGE(succKey, base.SeekGEFlagsNone.EnableTrySeekUsingNext())
1793 0 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1794 0 : ikey = i.iter.Next()
1795 0 : }
1796 0 : if ikey == nil {
1797 0 : i.kv = base.InternalKV{}
1798 0 : return nil
1799 0 : }
1800 0 : i.kv.K = *ikey
1801 0 : i.kv.V = base.MakeInPlaceValue(i.value())
1802 0 : return &i.kv
1803 : }
1804 :
1805 1 : func (i *batchIter) Prev() *base.InternalKV {
1806 1 : ikey := i.iter.Prev()
1807 1 : for ikey != nil && ikey.SeqNum() >= i.snapshot {
1808 0 : ikey = i.iter.Prev()
1809 0 : }
1810 1 : if ikey == nil {
1811 1 : i.kv = base.InternalKV{}
1812 1 : return nil
1813 1 : }
1814 1 : i.kv.K = *ikey
1815 1 : i.kv.V = base.MakeInPlaceValue(i.value())
1816 1 : return &i.kv
1817 : }
1818 :
1819 1 : func (i *batchIter) value() []byte {
1820 1 : offset, _, keyEnd := i.iter.KeyInfo()
1821 1 : data := i.batch.data
1822 1 : if len(data[offset:]) == 0 {
1823 0 : i.err = base.CorruptionErrorf("corrupted batch")
1824 0 : return nil
1825 0 : }
1826 :
1827 1 : switch InternalKeyKind(data[offset]) {
1828 : case InternalKeyKindSet, InternalKeyKindMerge, InternalKeyKindRangeDelete,
1829 : InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete,
1830 1 : InternalKeyKindDeleteSized:
1831 1 : _, value, ok := batchrepr.DecodeStr(data[keyEnd:])
1832 1 : if !ok {
1833 0 : return nil
1834 0 : }
1835 1 : return value
1836 1 : default:
1837 1 : return nil
1838 : }
1839 : }
1840 :
1841 1 : func (i *batchIter) Error() error {
1842 1 : return i.err
1843 1 : }
1844 :
1845 1 : func (i *batchIter) Close() error {
1846 1 : _ = i.iter.Close()
1847 1 : return i.err
1848 1 : }
1849 :
1850 1 : func (i *batchIter) SetBounds(lower, upper []byte) {
1851 1 : i.iter.SetBounds(lower, upper)
1852 1 : }
1853 :
1854 0 : func (i *batchIter) SetContext(_ context.Context) {}
1855 :
1856 : // DebugTree is part of the InternalIterator interface.
1857 0 : func (i *batchIter) DebugTree(tp treeprinter.Node) {
1858 0 : tp.Childf("%T(%p)", i, i)
1859 0 : }
1860 :
1861 : type flushableBatchEntry struct {
1862 : // offset is the byte offset of the record within the batch repr.
1863 : offset uint32
1864 : // index is the 0-based ordinal number of the record within the batch. Used
1865 : // to compute the seqnum for the record.
1866 : index uint32
1867 : // key{Start,End} are the start and end byte offsets of the key within the
1868 : // batch repr. Cached to avoid decoding the key length on every
1869 : // comparison. The value is stored starting at keyEnd.
1870 : keyStart uint32
1871 : keyEnd uint32
1872 : }
1873 :
1874 : // flushableBatch wraps an existing batch and provides the interfaces needed
1875 : // for making the batch flushable (i.e. able to mimic a memtable).
1876 : type flushableBatch struct {
1877 : cmp Compare
1878 : comparer *base.Comparer
1879 : data []byte
1880 :
1881 : // The base sequence number for the entries in the batch. This is the same
1882 : // value as Batch.seqNum() and is cached here for performance.
1883 : seqNum base.SeqNum
1884 :
1885 : // A slice of offsets and indices for the entries in the batch. Used to
1886 : // implement flushableBatchIter. Unlike the indexing on a normal batch, a
1887 : // flushable batch is indexed such that batch entry i will be given the
1888 : // sequence number flushableBatch.seqNum+i.
1889 : //
1890 : // Sorted in increasing order of key and decreasing order of offset (since
1891 : // higher offsets correspond to higher sequence numbers).
1892 : //
1893 : // Does not include range deletion entries or range key entries.
1894 : offsets []flushableBatchEntry
1895 :
1896 : // Fragmented range deletion tombstones.
1897 : tombstones []keyspan.Span
1898 :
1899 : // Fragmented range keys.
1900 : rangeKeys []keyspan.Span
1901 : }
1902 :
1903 : var _ flushable = (*flushableBatch)(nil)
1904 :
1905 : // newFlushableBatch creates a new batch that implements the flushable
1906 : // interface. This allows the batch to act like a memtable and be placed in the
1907 : // queue of flushable memtables. Note that the flushable batch takes ownership
1908 : // of the batch data.
1909 1 : func newFlushableBatch(batch *Batch, comparer *Comparer) (*flushableBatch, error) {
1910 1 : b := &flushableBatch{
1911 1 : data: batch.data,
1912 1 : cmp: comparer.Compare,
1913 1 : comparer: comparer,
1914 1 : offsets: make([]flushableBatchEntry, 0, batch.Count()),
1915 1 : }
1916 1 : if b.data != nil {
1917 1 : // Note that this sequence number is not correct when this batch has not
1918 1 : // been applied since the sequence number has not been assigned yet. The
1919 1 : // correct sequence number will be set later. But it is correct when the
1920 1 : // batch is being replayed from the WAL.
1921 1 : b.seqNum = batch.SeqNum()
1922 1 : }
1923 1 : var rangeDelOffsets []flushableBatchEntry
1924 1 : var rangeKeyOffsets []flushableBatchEntry
1925 1 : if len(b.data) > batchrepr.HeaderLen {
1926 1 : // Non-empty batch.
1927 1 : var index uint32
1928 1 : for iter := batchrepr.Read(b.data); len(iter) > 0; {
1929 1 : offset := uintptr(unsafe.Pointer(&iter[0])) - uintptr(unsafe.Pointer(&b.data[0]))
1930 1 : kind, key, _, ok, err := iter.Next()
1931 1 : if !ok {
1932 0 : if err != nil {
1933 0 : return nil, err
1934 0 : }
1935 0 : break
1936 : }
1937 1 : entry := flushableBatchEntry{
1938 1 : offset: uint32(offset),
1939 1 : index: uint32(index),
1940 1 : }
1941 1 : if keySize := uint32(len(key)); keySize == 0 {
1942 1 : // Must add 2 to the offset. One byte encodes `kind` and the next
1943 1 : // byte encodes `0`, which is the length of the key.
1944 1 : entry.keyStart = uint32(offset) + 2
1945 1 : entry.keyEnd = entry.keyStart
1946 1 : } else {
1947 1 : entry.keyStart = uint32(uintptr(unsafe.Pointer(&key[0])) -
1948 1 : uintptr(unsafe.Pointer(&b.data[0])))
1949 1 : entry.keyEnd = entry.keyStart + keySize
1950 1 : }
1951 1 : switch kind {
1952 1 : case InternalKeyKindRangeDelete:
1953 1 : rangeDelOffsets = append(rangeDelOffsets, entry)
1954 1 : case InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete:
1955 1 : rangeKeyOffsets = append(rangeKeyOffsets, entry)
1956 1 : case InternalKeyKindLogData:
1957 1 : // Skip it; we never want to iterate over LogDatas.
1958 1 : continue
1959 : case InternalKeyKindSet, InternalKeyKindDelete, InternalKeyKindMerge,
1960 1 : InternalKeyKindSingleDelete, InternalKeyKindSetWithDelete, InternalKeyKindDeleteSized:
1961 1 : b.offsets = append(b.offsets, entry)
1962 0 : default:
1963 0 : // Note In some circumstances this might be temporary memory
1964 0 : // corruption that can be recovered by discarding the batch and
1965 0 : // trying again. In other cases, the batch repr might've been
1966 0 : // already persisted elsewhere, and we'll loop continuously trying
1967 0 : // to commit the same corrupted batch. The caller is responsible for
1968 0 : // distinguishing.
1969 0 : return nil, errors.Wrapf(ErrInvalidBatch, "unrecognized kind %v", kind)
1970 : }
1971 : // NB: index (used for entry.offset above) must not reach the
1972 : // batch.count, because the offset is used in conjunction with the
1973 : // batch's sequence number to assign sequence numbers to keys within
1974 : // the batch. If we assign KV's indexes as high as batch.count,
1975 : // we'll begin assigning keys sequence numbers that weren't
1976 : // allocated.
1977 1 : if index >= uint32(batch.count) {
1978 0 : return nil, base.AssertionFailedf("pebble: batch entry index %d ≥ batch.count %d", index, batch.count)
1979 0 : }
1980 1 : index++
1981 : }
1982 : }
1983 :
1984 : // Sort all of offsets, rangeDelOffsets and rangeKeyOffsets, using *batch's
1985 : // sort.Interface implementation.
1986 1 : pointOffsets := b.offsets
1987 1 : sort.Sort(b)
1988 1 : b.offsets = rangeDelOffsets
1989 1 : sort.Sort(b)
1990 1 : b.offsets = rangeKeyOffsets
1991 1 : sort.Sort(b)
1992 1 : b.offsets = pointOffsets
1993 1 :
1994 1 : if len(rangeDelOffsets) > 0 {
1995 1 : frag := &keyspan.Fragmenter{
1996 1 : Cmp: b.cmp,
1997 1 : Format: b.comparer.FormatKey,
1998 1 : Emit: func(s keyspan.Span) {
1999 1 : b.tombstones = append(b.tombstones, s)
2000 1 : },
2001 : }
2002 1 : it := &flushableBatchIter{
2003 1 : batch: b,
2004 1 : data: b.data,
2005 1 : offsets: rangeDelOffsets,
2006 1 : cmp: b.cmp,
2007 1 : index: -1,
2008 1 : }
2009 1 : fragmentRangeDels(frag, it, len(rangeDelOffsets))
2010 : }
2011 1 : if len(rangeKeyOffsets) > 0 {
2012 1 : frag := &keyspan.Fragmenter{
2013 1 : Cmp: b.cmp,
2014 1 : Format: b.comparer.FormatKey,
2015 1 : Emit: func(s keyspan.Span) {
2016 1 : b.rangeKeys = append(b.rangeKeys, s)
2017 1 : },
2018 : }
2019 1 : it := &flushableBatchIter{
2020 1 : batch: b,
2021 1 : data: b.data,
2022 1 : offsets: rangeKeyOffsets,
2023 1 : cmp: b.cmp,
2024 1 : index: -1,
2025 1 : }
2026 1 : fragmentRangeKeys(frag, it, len(rangeKeyOffsets))
2027 : }
2028 1 : return b, nil
2029 : }
2030 :
2031 1 : func (b *flushableBatch) setSeqNum(seqNum base.SeqNum) {
2032 1 : if b.seqNum != 0 {
2033 0 : panic(fmt.Sprintf("pebble: flushableBatch.seqNum already set: %d", b.seqNum))
2034 : }
2035 1 : b.seqNum = seqNum
2036 1 : for i := range b.tombstones {
2037 1 : for j := range b.tombstones[i].Keys {
2038 1 : b.tombstones[i].Keys[j].Trailer = base.MakeTrailer(
2039 1 : b.tombstones[i].Keys[j].SeqNum()+seqNum,
2040 1 : b.tombstones[i].Keys[j].Kind(),
2041 1 : )
2042 1 : }
2043 : }
2044 1 : for i := range b.rangeKeys {
2045 1 : for j := range b.rangeKeys[i].Keys {
2046 1 : b.rangeKeys[i].Keys[j].Trailer = base.MakeTrailer(
2047 1 : b.rangeKeys[i].Keys[j].SeqNum()+seqNum,
2048 1 : b.rangeKeys[i].Keys[j].Kind(),
2049 1 : )
2050 1 : }
2051 : }
2052 : }
2053 :
2054 1 : func (b *flushableBatch) Len() int {
2055 1 : return len(b.offsets)
2056 1 : }
2057 :
2058 1 : func (b *flushableBatch) Less(i, j int) bool {
2059 1 : ei := &b.offsets[i]
2060 1 : ej := &b.offsets[j]
2061 1 : ki := b.data[ei.keyStart:ei.keyEnd]
2062 1 : kj := b.data[ej.keyStart:ej.keyEnd]
2063 1 : switch c := b.cmp(ki, kj); {
2064 1 : case c < 0:
2065 1 : return true
2066 1 : case c > 0:
2067 1 : return false
2068 1 : default:
2069 1 : return ei.offset > ej.offset
2070 : }
2071 : }
2072 :
2073 1 : func (b *flushableBatch) Swap(i, j int) {
2074 1 : b.offsets[i], b.offsets[j] = b.offsets[j], b.offsets[i]
2075 1 : }
2076 :
2077 : // newIter is part of the flushable interface.
2078 1 : func (b *flushableBatch) newIter(o *IterOptions) internalIterator {
2079 1 : return &flushableBatchIter{
2080 1 : batch: b,
2081 1 : data: b.data,
2082 1 : offsets: b.offsets,
2083 1 : cmp: b.cmp,
2084 1 : index: -1,
2085 1 : lower: o.GetLowerBound(),
2086 1 : upper: o.GetUpperBound(),
2087 1 : }
2088 1 : }
2089 :
2090 : // newFlushIter is part of the flushable interface.
2091 1 : func (b *flushableBatch) newFlushIter(o *IterOptions) internalIterator {
2092 1 : return &flushFlushableBatchIter{
2093 1 : flushableBatchIter: flushableBatchIter{
2094 1 : batch: b,
2095 1 : data: b.data,
2096 1 : offsets: b.offsets,
2097 1 : cmp: b.cmp,
2098 1 : index: -1,
2099 1 : },
2100 1 : }
2101 1 : }
2102 :
2103 : // newRangeDelIter is part of the flushable interface.
2104 1 : func (b *flushableBatch) newRangeDelIter(o *IterOptions) keyspan.FragmentIterator {
2105 1 : if len(b.tombstones) == 0 {
2106 1 : return nil
2107 1 : }
2108 1 : return keyspan.NewIter(b.cmp, b.tombstones)
2109 : }
2110 :
2111 : // newRangeKeyIter is part of the flushable interface.
2112 1 : func (b *flushableBatch) newRangeKeyIter(o *IterOptions) keyspan.FragmentIterator {
2113 1 : if len(b.rangeKeys) == 0 {
2114 1 : return nil
2115 1 : }
2116 1 : return keyspan.NewIter(b.cmp, b.rangeKeys)
2117 : }
2118 :
2119 : // containsRangeKeys is part of the flushable interface.
2120 1 : func (b *flushableBatch) containsRangeKeys() bool { return len(b.rangeKeys) > 0 }
2121 :
2122 : // inuseBytes is part of the flushable interface.
2123 1 : func (b *flushableBatch) inuseBytes() uint64 {
2124 1 : return uint64(len(b.data) - batchrepr.HeaderLen)
2125 1 : }
2126 :
2127 : // totalBytes is part of the flushable interface.
2128 1 : func (b *flushableBatch) totalBytes() uint64 {
2129 1 : return uint64(cap(b.data))
2130 1 : }
2131 :
2132 : // readyForFlush is part of the flushable interface.
2133 1 : func (b *flushableBatch) readyForFlush() bool {
2134 1 : // A flushable batch is always ready for flush; it must be flushed together
2135 1 : // with the previous memtable.
2136 1 : return true
2137 1 : }
2138 :
2139 : // computePossibleOverlaps is part of the flushable interface.
2140 : func (b *flushableBatch) computePossibleOverlaps(
2141 : fn func(bounded) shouldContinue, bounded ...bounded,
2142 1 : ) {
2143 1 : computePossibleOverlapsGenericImpl[*flushableBatch](b, b.cmp, fn, bounded)
2144 1 : }
2145 :
2146 : // Note: flushableBatchIter mirrors the implementation of batchIter. Keep the
2147 : // two in sync.
2148 : type flushableBatchIter struct {
2149 : // Members to be initialized by creator.
2150 : batch *flushableBatch
2151 : // The bytes backing the batch. Always the same as batch.data?
2152 : data []byte
2153 : // The sorted entries. This is not always equal to batch.offsets.
2154 : offsets []flushableBatchEntry
2155 : cmp Compare
2156 : // Must be initialized to -1. It is the index into offsets that represents
2157 : // the current iterator position.
2158 : index int
2159 :
2160 : // For internal use by the implementation.
2161 : kv base.InternalKV
2162 : err error
2163 :
2164 : // Optionally initialize to bounds of iteration, if any.
2165 : lower []byte
2166 : upper []byte
2167 : }
2168 :
2169 : // flushableBatchIter implements the base.InternalIterator interface.
2170 : var _ base.InternalIterator = (*flushableBatchIter)(nil)
2171 :
2172 1 : func (i *flushableBatchIter) String() string {
2173 1 : return "flushable-batch"
2174 1 : }
2175 :
2176 : // SeekGE implements internalIterator.SeekGE, as documented in the pebble
2177 : // package. Ignore flags.TrySeekUsingNext() since we don't expect this
2178 : // optimization to provide much benefit here at the moment.
2179 1 : func (i *flushableBatchIter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
2180 1 : i.err = nil // clear cached iteration error
2181 1 : ikey := base.MakeSearchKey(key)
2182 1 : i.index = sort.Search(len(i.offsets), func(j int) bool {
2183 1 : return base.InternalCompare(i.cmp, ikey, i.getKey(j)) <= 0
2184 1 : })
2185 1 : if i.index >= len(i.offsets) {
2186 1 : return nil
2187 1 : }
2188 1 : kv := i.getKV(i.index)
2189 1 : if i.upper != nil && i.cmp(kv.K.UserKey, i.upper) >= 0 {
2190 1 : i.index = len(i.offsets)
2191 1 : return nil
2192 1 : }
2193 1 : return kv
2194 : }
2195 :
2196 : // SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
2197 : // pebble package.
2198 : func (i *flushableBatchIter) SeekPrefixGE(
2199 : prefix, key []byte, flags base.SeekGEFlags,
2200 1 : ) *base.InternalKV {
2201 1 : kv := i.SeekGE(key, flags)
2202 1 : if kv == nil {
2203 1 : return nil
2204 1 : }
2205 : // If the key doesn't have the sought prefix, return nil.
2206 1 : if !bytes.Equal(i.batch.comparer.Split.Prefix(kv.K.UserKey), prefix) {
2207 1 : return nil
2208 1 : }
2209 1 : return kv
2210 : }
2211 :
2212 : // SeekLT implements internalIterator.SeekLT, as documented in the pebble
2213 : // package.
2214 1 : func (i *flushableBatchIter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
2215 1 : i.err = nil // clear cached iteration error
2216 1 : ikey := base.MakeSearchKey(key)
2217 1 : i.index = sort.Search(len(i.offsets), func(j int) bool {
2218 1 : return base.InternalCompare(i.cmp, ikey, i.getKey(j)) <= 0
2219 1 : })
2220 1 : i.index--
2221 1 : if i.index < 0 {
2222 1 : return nil
2223 1 : }
2224 1 : kv := i.getKV(i.index)
2225 1 : if i.lower != nil && i.cmp(kv.K.UserKey, i.lower) < 0 {
2226 1 : i.index = -1
2227 1 : return nil
2228 1 : }
2229 1 : return kv
2230 : }
2231 :
2232 : // First implements internalIterator.First, as documented in the pebble
2233 : // package.
2234 1 : func (i *flushableBatchIter) First() *base.InternalKV {
2235 1 : i.err = nil // clear cached iteration error
2236 1 : if len(i.offsets) == 0 {
2237 1 : return nil
2238 1 : }
2239 1 : i.index = 0
2240 1 : kv := i.getKV(i.index)
2241 1 : if i.upper != nil && i.cmp(kv.K.UserKey, i.upper) >= 0 {
2242 1 : i.index = len(i.offsets)
2243 1 : return nil
2244 1 : }
2245 1 : return kv
2246 : }
2247 :
2248 : // Last implements internalIterator.Last, as documented in the pebble
2249 : // package.
2250 1 : func (i *flushableBatchIter) Last() *base.InternalKV {
2251 1 : i.err = nil // clear cached iteration error
2252 1 : if len(i.offsets) == 0 {
2253 1 : return nil
2254 1 : }
2255 1 : i.index = len(i.offsets) - 1
2256 1 : kv := i.getKV(i.index)
2257 1 : if i.lower != nil && i.cmp(kv.K.UserKey, i.lower) < 0 {
2258 0 : i.index = -1
2259 0 : return nil
2260 0 : }
2261 1 : return kv
2262 : }
2263 :
2264 : // Note: flushFlushableBatchIter.Next mirrors the implementation of
2265 : // flushableBatchIter.Next due to performance. Keep the two in sync.
2266 1 : func (i *flushableBatchIter) Next() *base.InternalKV {
2267 1 : if i.index == len(i.offsets) {
2268 0 : return nil
2269 0 : }
2270 1 : i.index++
2271 1 : if i.index == len(i.offsets) {
2272 1 : return nil
2273 1 : }
2274 1 : kv := i.getKV(i.index)
2275 1 : if i.upper != nil && i.cmp(kv.K.UserKey, i.upper) >= 0 {
2276 1 : i.index = len(i.offsets)
2277 1 : return nil
2278 1 : }
2279 1 : return kv
2280 : }
2281 :
2282 1 : func (i *flushableBatchIter) Prev() *base.InternalKV {
2283 1 : if i.index < 0 {
2284 0 : return nil
2285 0 : }
2286 1 : i.index--
2287 1 : if i.index < 0 {
2288 1 : return nil
2289 1 : }
2290 1 : kv := i.getKV(i.index)
2291 1 : if i.lower != nil && i.cmp(kv.K.UserKey, i.lower) < 0 {
2292 1 : i.index = -1
2293 1 : return nil
2294 1 : }
2295 1 : return kv
2296 : }
2297 :
2298 : // Note: flushFlushableBatchIter.NextPrefix mirrors the implementation of
2299 : // flushableBatchIter.NextPrefix due to performance. Keep the two in sync.
2300 0 : func (i *flushableBatchIter) NextPrefix(succKey []byte) *base.InternalKV {
2301 0 : return i.SeekGE(succKey, base.SeekGEFlagsNone.EnableTrySeekUsingNext())
2302 0 : }
2303 :
2304 1 : func (i *flushableBatchIter) getKey(index int) InternalKey {
2305 1 : e := &i.offsets[index]
2306 1 : kind := InternalKeyKind(i.data[e.offset])
2307 1 : key := i.data[e.keyStart:e.keyEnd]
2308 1 : return base.MakeInternalKey(key, i.batch.seqNum+base.SeqNum(e.index), kind)
2309 1 : }
2310 :
2311 1 : func (i *flushableBatchIter) getKV(index int) *base.InternalKV {
2312 1 : i.kv = base.InternalKV{
2313 1 : K: i.getKey(index),
2314 1 : V: i.extractValue(),
2315 1 : }
2316 1 : return &i.kv
2317 1 : }
2318 :
2319 1 : func (i *flushableBatchIter) extractValue() base.LazyValue {
2320 1 : p := i.data[i.offsets[i.index].offset:]
2321 1 : if len(p) == 0 {
2322 0 : i.err = base.CorruptionErrorf("corrupted batch")
2323 0 : return base.LazyValue{}
2324 0 : }
2325 1 : kind := InternalKeyKind(p[0])
2326 1 : if kind > InternalKeyKindMax {
2327 0 : i.err = base.CorruptionErrorf("corrupted batch")
2328 0 : return base.LazyValue{}
2329 0 : }
2330 1 : var value []byte
2331 1 : var ok bool
2332 1 : switch kind {
2333 : case InternalKeyKindSet, InternalKeyKindMerge, InternalKeyKindRangeDelete,
2334 : InternalKeyKindRangeKeySet, InternalKeyKindRangeKeyUnset, InternalKeyKindRangeKeyDelete,
2335 1 : InternalKeyKindDeleteSized:
2336 1 : keyEnd := i.offsets[i.index].keyEnd
2337 1 : _, value, ok = batchrepr.DecodeStr(i.data[keyEnd:])
2338 1 : if !ok {
2339 0 : i.err = base.CorruptionErrorf("corrupted batch")
2340 0 : return base.LazyValue{}
2341 0 : }
2342 : }
2343 1 : return base.MakeInPlaceValue(value)
2344 : }
2345 :
2346 0 : func (i *flushableBatchIter) Valid() bool {
2347 0 : return i.index >= 0 && i.index < len(i.offsets)
2348 0 : }
2349 :
2350 1 : func (i *flushableBatchIter) Error() error {
2351 1 : return i.err
2352 1 : }
2353 :
2354 1 : func (i *flushableBatchIter) Close() error {
2355 1 : return i.err
2356 1 : }
2357 :
2358 1 : func (i *flushableBatchIter) SetBounds(lower, upper []byte) {
2359 1 : i.lower = lower
2360 1 : i.upper = upper
2361 1 : }
2362 :
2363 0 : func (i *flushableBatchIter) SetContext(_ context.Context) {}
2364 :
2365 : // DebugTree is part of the InternalIterator interface.
2366 0 : func (i *flushableBatchIter) DebugTree(tp treeprinter.Node) {
2367 0 : tp.Childf("%T(%p)", i, i)
2368 0 : }
2369 :
2370 : // flushFlushableBatchIter is similar to flushableBatchIter but it keeps track
2371 : // of number of bytes iterated.
2372 : type flushFlushableBatchIter struct {
2373 : flushableBatchIter
2374 : }
2375 :
2376 : // flushFlushableBatchIter implements the base.InternalIterator interface.
2377 : var _ base.InternalIterator = (*flushFlushableBatchIter)(nil)
2378 :
2379 0 : func (i *flushFlushableBatchIter) String() string {
2380 0 : return "flushable-batch"
2381 0 : }
2382 :
2383 0 : func (i *flushFlushableBatchIter) SeekGE(key []byte, flags base.SeekGEFlags) *base.InternalKV {
2384 0 : panic("pebble: SeekGE unimplemented")
2385 : }
2386 :
2387 : func (i *flushFlushableBatchIter) SeekPrefixGE(
2388 : prefix, key []byte, flags base.SeekGEFlags,
2389 0 : ) *base.InternalKV {
2390 0 : panic("pebble: SeekPrefixGE unimplemented")
2391 : }
2392 :
2393 0 : func (i *flushFlushableBatchIter) SeekLT(key []byte, flags base.SeekLTFlags) *base.InternalKV {
2394 0 : panic("pebble: SeekLT unimplemented")
2395 : }
2396 :
2397 1 : func (i *flushFlushableBatchIter) First() *base.InternalKV {
2398 1 : i.err = nil // clear cached iteration error
2399 1 : return i.flushableBatchIter.First()
2400 1 : }
2401 :
2402 0 : func (i *flushFlushableBatchIter) NextPrefix(succKey []byte) *base.InternalKV {
2403 0 : panic("pebble: Prev unimplemented")
2404 : }
2405 :
2406 : // Note: flushFlushableBatchIter.Next mirrors the implementation of
2407 : // flushableBatchIter.Next due to performance. Keep the two in sync.
2408 1 : func (i *flushFlushableBatchIter) Next() *base.InternalKV {
2409 1 : if i.index == len(i.offsets) {
2410 0 : return nil
2411 0 : }
2412 1 : i.index++
2413 1 : if i.index == len(i.offsets) {
2414 1 : return nil
2415 1 : }
2416 1 : return i.getKV(i.index)
2417 : }
2418 :
2419 0 : func (i flushFlushableBatchIter) Prev() *base.InternalKV {
2420 0 : panic("pebble: Prev unimplemented")
2421 : }
2422 :
2423 : // batchOptions holds the parameters to configure batch.
2424 : type batchOptions struct {
2425 : initialSizeBytes int
2426 : maxRetainedSizeBytes int
2427 : }
2428 :
2429 : // ensureDefaults creates batch options with default values.
2430 1 : func (o *batchOptions) ensureDefaults() {
2431 1 : if o.initialSizeBytes <= 0 {
2432 1 : o.initialSizeBytes = defaultBatchInitialSize
2433 1 : }
2434 1 : if o.maxRetainedSizeBytes <= 0 {
2435 1 : o.maxRetainedSizeBytes = defaultBatchMaxRetainedSize
2436 1 : }
2437 : }
2438 :
2439 : // BatchOption allows customizing the batch.
2440 : type BatchOption func(*batchOptions)
2441 :
2442 : // WithInitialSizeBytes sets a custom initial size for the batch. Defaults
2443 : // to 1KB.
2444 0 : func WithInitialSizeBytes(s int) BatchOption {
2445 0 : return func(opts *batchOptions) {
2446 0 : opts.initialSizeBytes = s
2447 0 : }
2448 : }
2449 :
2450 : // WithMaxRetainedSizeBytes sets a custom max size for the batch to be
2451 : // re-used. Any batch which exceeds the max retained size would be GC-ed.
2452 : // Defaults to 1MB.
2453 0 : func WithMaxRetainedSizeBytes(s int) BatchOption {
2454 0 : return func(opts *batchOptions) {
2455 0 : opts.maxRetainedSizeBytes = s
2456 0 : }
2457 : }
2458 :
2459 : // batchSort returns iterators for the sorted contents of the batch. It is
2460 : // intended for testing use only. The batch.Sort dance is done to prevent
2461 : // exposing this method in the public pebble interface.
2462 : func batchSort(
2463 : i interface{},
2464 : ) (
2465 : points internalIterator,
2466 : rangeDels keyspan.FragmentIterator,
2467 : rangeKeys keyspan.FragmentIterator,
2468 1 : ) {
2469 1 : b := i.(*Batch)
2470 1 : if b.Indexed() {
2471 1 : pointIter := b.newInternalIter(nil)
2472 1 : rangeDelIter := b.newRangeDelIter(nil, math.MaxUint64)
2473 1 : rangeKeyIter := b.newRangeKeyIter(nil, math.MaxUint64)
2474 1 : return pointIter, rangeDelIter, rangeKeyIter
2475 1 : }
2476 1 : f, err := newFlushableBatch(b, b.db.opts.Comparer)
2477 1 : if err != nil {
2478 0 : panic(err)
2479 : }
2480 1 : return f.newIter(nil), f.newRangeDelIter(nil), f.newRangeKeyIter(nil)
2481 : }
2482 :
2483 1 : func init() {
2484 1 : private.BatchSort = batchSort
2485 1 : }
|
