func decodeBatchToMem(data []byte, expectSeq uint64, mdb *memdb.DB) (seq uint64, batchLen int, err error) {
seq, batchLen, err = decodeBatchHeader(data)
if err != nil {
return 0, 0, err
}
if seq < expectSeq {
return 0, 0, newErrBatchCorrupted("invalid sequence number")
}
data = data[batchHeaderLen:]
var ik []byte
var decodedLen int
err = decodeBatch(data, func(i int, index batchIndex) error {
if i >= batchLen {
return newErrBatchCorrupted("invalid records length")
}
ik = makeInternalKey(ik, index.k(data), seq+uint64(i), index.keyType)
if err := mdb.Put(ik, index.v(data)); err != nil {
return err
}
decodedLen++
return nil
})
if err == nil && decodedLen != batchLen {
err = newErrBatchCorrupted(fmt.Sprintf("invalid records length: %d vs %d", batchLen, decodedLen))
}
return
}
func (b *Batch) revertMemReplay(to *memdb.DB) error {
var ikScratch []byte
return b.decodeRec(func(i int, kt keyType, key, value []byte) error {
ikScratch := makeInternalKey(ikScratch, key, b.seq+uint64(i), kt)
return to.Delete(ikScratch)
})
}
func (b *Batch) memReplay(to *memdb.DB) error {
var ikScratch []byte
return b.decodeRec(func(i int, kt kType, key, value []byte) error {
ikScratch = makeIkey(ikScratch, key, b.seq+uint64(i), kt)
return to.Put(ikScratch, value)
})
}
func (d *DB) memCompaction(mem *memdb.DB) {
s := d.s
c := newCMem(s)
stats := new(cStatsStaging)
s.printf("MemCompaction: started, size=%d entries=%d", mem.Size(), mem.Len())
d.transact(func() (err error) {
stats.startTimer()
defer stats.stopTimer()
return c.flush(mem, -1)
})
d.transact(func() (err error) {
stats.startTimer()
defer stats.stopTimer()
return c.commit(d.journal.file.Num(), d.fseq)
})
stats.write = c.t.size
d.cstats[c.level].add(stats)
// drop frozen mem
d.dropFrozenMem()
c = nil
}
func (b *Batch) revertMem(seq uint64, mdb *memdb.DB) error {
var ik []byte
for i, index := range b.index {
ik = makeInternalKey(ik, index.k(b.data), seq+uint64(i), index.keyType)
if err := mdb.Delete(ik); err != nil {
return err
}
}
return nil
}
func (db *DB) mpoolGet(n int) *memDB {
var mdb *memdb.DB
select {
case mdb = <-db.memPool:
default:
}
if mdb == nil || mdb.Capacity() < n {
mdb = memdb.New(db.s.icmp, maxInt(db.s.o.GetWriteBuffer(), n))
}
return &memDB{
db: db,
DB: mdb,
}
}
func (s *session) flushMemdb(rec *sessionRecord, mdb *memdb.DB, level int) (level_ int, err error) {
// Create sorted table.
iter := mdb.NewIterator(nil)
defer iter.Release()
t, n, err := s.tops.createFrom(iter)
if err != nil {
return level, err
}
// Pick level and add to record.
if level < 0 {
level = s.pickMemdbLevel(t.imin.ukey(), t.imax.ukey())
}
rec.addTableFile(level, t)
s.logf("memdb@flush created L%d@%d N·%d S·%s %q:%q", level, t.file.Num(), n, shortenb(int(t.size)), t.imin, t.imax)
return level, nil
}
func (c *cMem) flush(mem *memdb.DB, level int) error {
s := c.s
// Write memdb to table
t, n, err := s.tops.createFrom(mem.NewIterator(nil))
if err != nil {
return err
}
if level < 0 {
level = s.version_NB().pickLevel(t.min.ukey(), t.max.ukey())
}
c.rec.addTableFile(level, t)
s.logf("mem@flush created L%d@%d N·%d S·%s %q:%q", level, t.file.Num(), n, shortenb(int(t.size)), t.min, t.max)
c.level = level
return nil
}
func memGet(mdb *memdb.DB, ikey internalKey, icmp *iComparer) (ok bool, mv []byte, err error) {
mk, mv, err := mdb.Find(ikey)
if err == nil {
ukey, _, kt, kerr := parseInternalKey(mk)
if kerr != nil {
// Shouldn't have had happen.
panic(kerr)
}
if icmp.uCompare(ukey, ikey.ukey()) == 0 {
if kt == keyTypeDel {
return true, nil, ErrNotFound
}
return true, mv, nil
}
} else if err != ErrNotFound {
return true, nil, err
}
return
}
func (c *cMem) flush(mem *memdb.DB, level int) error {
s := c.s
// Write memdb to table
t, n, err := s.tops.createFrom(mem.NewIterator())
if err != nil {
return err
}
if level < 0 {
level = s.version().pickLevel(t.min.ukey(), t.max.ukey())
}
c.rec.addTableFile(level, t)
s.printf("Compaction: table created, source=mem level=%d num=%d size=%d entries=%d min=%q max=%q",
level, t.file.Num(), t.size, n, t.min, t.max)
c.level = level
c.t = t
return nil
}
func (s *session) flushMemdb(rec *sessionRecord, mdb *memdb.DB, maxLevel int) (int, error) {
// Create sorted table.
iter := mdb.NewIterator(nil)
defer iter.Release()
t, n, err := s.tops.createFrom(iter)
if err != nil {
return 0, err
}
// Pick level other than zero can cause compaction issue with large
// bulk insert and delete on strictly incrementing key-space. The
// problem is that the small deletion markers trapped at lower level,
// while key/value entries keep growing at higher level. Since the
// key-space is strictly incrementing it will not overlaps with
// higher level, thus maximum possible level is always picked, while
// overlapping deletion marker pushed into lower level.
// See: https://github.com/syndtr/goleveldb/issues/127.
flushLevel := s.pickMemdbLevel(t.imin.ukey(), t.imax.ukey(), maxLevel)
rec.addTableFile(flushLevel, t)
s.logf("memdb@flush created L%d@%d N·%d S·%s %q:%q", flushLevel, t.fd.Num, n, shortenb(int(t.size)), t.imin, t.imax)
return flushLevel, nil
}
func (c *cMem) flush(mem *memdb.DB, level int) error {
s := c.s
// Write memdb to table.
iter := mem.NewIterator(nil)
defer iter.Release()
t, n, err := s.tops.createFrom(iter)
if err != nil {
return err
}
// Pick level.
if level < 0 {
v := s.version()
level = v.pickLevel(t.imin.ukey(), t.imax.ukey())
v.release()
}
c.rec.addTableFile(level, t)
s.logf("mem@flush created L%d@%d N·%d S·%s %q:%q", level, t.file.Num(), n, shortenb(int(t.size)), t.imin, t.imax)
c.level = level
return nil
}
func isMemOverlaps(icmp *iComparer, mem *memdb.DB, min, max []byte) bool {
iter := mem.NewIterator(nil)
defer iter.Release()
return (max == nil || (iter.First() && icmp.uCompare(max, internalKey(iter.Key()).ukey()) >= 0)) &&
(min == nil || (iter.Last() && icmp.uCompare(min, internalKey(iter.Key()).ukey()) <= 0))
}
func (db *DB) recoverJournal() error {
// Get all tables and sort it by file number.
journalFiles_, err := db.s.getFiles(storage.TypeJournal)
if err != nil {
return err
}
journalFiles := files(journalFiles_)
journalFiles.sort()
// Discard older journal.
prev := -1
for i, file := range journalFiles {
if file.Num() >= db.s.stJournalNum {
if prev >= 0 {
i--
journalFiles[i] = journalFiles[prev]
}
journalFiles = journalFiles[i:]
break
} else if file.Num() == db.s.stPrevJournalNum {
prev = i
}
}
var jr *journal.Reader
var of storage.File
var mem *memdb.DB
batch := new(Batch)
cm := newCMem(db.s)
buf := new(util.Buffer)
// Options.
strict := db.s.o.GetStrict(opt.StrictJournal)
checksum := db.s.o.GetStrict(opt.StrictJournalChecksum)
writeBuffer := db.s.o.GetWriteBuffer()
recoverJournal := func(file storage.File) error {
db.logf("journal@recovery recovering @%d", file.Num())
reader, err := file.Open()
if err != nil {
return err
}
defer reader.Close()
// Create/reset journal reader instance.
if jr == nil {
jr = journal.NewReader(reader, dropper{db.s, file}, strict, checksum)
} else {
jr.Reset(reader, dropper{db.s, file}, strict, checksum)
}
// Flush memdb and remove obsolete journal file.
if of != nil {
if mem.Len() > 0 {
if err := cm.flush(mem, 0); err != nil {
return err
}
}
if err := cm.commit(file.Num(), db.seq); err != nil {
return err
}
cm.reset()
of.Remove()
of = nil
}
// Replay journal to memdb.
mem.Reset()
for {
r, err := jr.Next()
if err != nil {
if err == io.EOF {
break
}
return err
}
buf.Reset()
if _, err := buf.ReadFrom(r); err != nil {
if err == io.ErrUnexpectedEOF {
continue
} else {
return err
}
}
if err := batch.decode(buf.Bytes()); err != nil {
return err
}
if err := batch.memReplay(mem); err != nil {
return err
}
// Save sequence number.
db.seq = batch.seq + uint64(batch.len())
// Flush it if large enough.
if mem.Size() >= writeBuffer {
if err := cm.flush(mem, 0); err != nil {
return err
}
mem.Reset()
}
}
of = file
return nil
}
// Recover all journals.
if len(journalFiles) > 0 {
db.logf("journal@recovery F·%d", len(journalFiles))
// Mark file number as used.
db.s.markFileNum(journalFiles[len(journalFiles)-1].Num())
mem = memdb.New(db.s.icmp, writeBuffer)
for _, file := range journalFiles {
if err := recoverJournal(file); err != nil {
return err
}
}
// Flush the last journal.
if mem.Len() > 0 {
if err := cm.flush(mem, 0); err != nil {
return err
}
}
}
// Create a new journal.
if _, err := db.newMem(0); err != nil {
return err
}
// Commit.
if err := cm.commit(db.journalFile.Num(), db.seq); err != nil {
// Close journal.
if db.journal != nil {
db.journal.Close()
db.journalWriter.Close()
}
return err
}
// Remove the last obsolete journal file.
if of != nil {
of.Remove()
}
return nil
}
func (d *DB) recoverJournal() (err error) {
s := d.s
icmp := s.cmp
s.printf("JournalRecovery: started, min=%d", s.stJournalNum)
var mem *memdb.DB
batch := new(Batch)
cm := newCMem(s)
journals := files(s.getFiles(storage.TypeJournal))
journals.sort()
rJournals := make([]storage.File, 0, len(journals))
for _, journal := range journals {
if journal.Num() >= s.stJournalNum || journal.Num() == s.stPrevJournalNum {
s.markFileNum(journal.Num())
rJournals = append(rJournals, journal)
}
}
var r, fr *journalReader
for _, journal := range rJournals {
s.printf("JournalRecovery: recovering, num=%d", journal.Num())
r, err = newJournalReader(journal, true, s.journalDropFunc("journal", journal.Num()))
if err != nil {
return
}
if mem != nil {
if mem.Len() > 0 {
err = cm.flush(mem, 0)
if err != nil {
return
}
}
err = cm.commit(r.file.Num(), d.seq)
if err != nil {
return
}
cm.reset()
fr.remove()
fr = nil
}
mem = memdb.New(icmp)
for r.journal.Next() {
err = batch.decode(r.journal.Record())
if err != nil {
return
}
err = batch.memReplay(mem)
if err != nil {
return
}
d.seq = batch.seq + uint64(batch.len())
if mem.Size() > s.o.GetWriteBuffer() {
// flush to table
err = cm.flush(mem, 0)
if err != nil {
return
}
// create new memdb
mem = memdb.New(icmp)
}
}
err = r.journal.Error()
if err != nil {
return
}
r.close()
fr = r
}
// create new journal
_, err = d.newMem()
if err != nil {
return
}
if mem != nil && mem.Len() > 0 {
err = cm.flush(mem, 0)
if err != nil {
return
}
}
err = cm.commit(d.journal.file.Num(), d.seq)
if err != nil {
return
}
if fr != nil {
fr.remove()
}
return
}
func (b *Batch) revertMemReplay(to *memdb.DB) error {
return b.decodeRec(func(i int, kt kType, key, value []byte) {
ikey := newIkey(key, b.seq+uint64(i), kt)
to.Delete(ikey)
})
}
func (d *DB) recoverJournal() error {
s := d.s
ff0, err := s.getFiles(storage.TypeJournal)
if err != nil {
return err
}
ff1 := files(ff0)
ff1.sort()
ff2 := make([]storage.File, 0, len(ff1))
for _, file := range ff1 {
if file.Num() >= s.stJournalNum || file.Num() == s.stPrevJournalNum {
s.markFileNum(file.Num())
ff2 = append(ff2, file)
}
}
var jr *journal.Reader
var of storage.File
var mem *memdb.DB
batch := new(Batch)
cm := newCMem(s)
buf := new(util.Buffer)
// Options.
strict := s.o.GetStrict(opt.StrictJournal)
checksum := s.o.GetStrict(opt.StrictJournalChecksum)
writeBuffer := s.o.GetWriteBuffer()
recoverJournal := func(file storage.File) error {
s.logf("journal@recovery recovering @%d", file.Num())
reader, err := file.Open()
if err != nil {
return err
}
defer reader.Close()
if jr == nil {
jr = journal.NewReader(reader, dropper{s, file}, strict, checksum)
} else {
jr.Reset(reader, dropper{s, file}, strict, checksum)
}
if of != nil {
if mem.Len() > 0 {
if err := cm.flush(mem, 0); err != nil {
return err
}
}
if err := cm.commit(file.Num(), d.seq); err != nil {
return err
}
cm.reset()
of.Remove()
of = nil
}
// Reset memdb.
mem.Reset()
for {
r, err := jr.Next()
if err != nil {
if err == io.EOF {
break
}
return err
}
buf.Reset()
if _, err := buf.ReadFrom(r); err != nil {
if strict {
return err
}
continue
}
if err := batch.decode(buf.Bytes()); err != nil {
return err
}
if err := batch.memReplay(mem); err != nil {
return err
}
d.seq = batch.seq + uint64(batch.len())
if mem.Size() >= writeBuffer {
// Large enough, flush it.
if err := cm.flush(mem, 0); err != nil {
return err
}
// Reset memdb.
mem.Reset()
}
}
of = file
return nil
}
// Recover all journals.
if len(ff2) > 0 {
s.logf("journal@recovery F·%d", len(ff2))
mem = memdb.New(s.icmp, writeBuffer)
for _, file := range ff2 {
if err := recoverJournal(file); err != nil {
return err
}
}
// Flush the last journal.
if mem.Len() > 0 {
if err := cm.flush(mem, 0); err != nil {
return err
}
}
}
// Create a new journal.
if _, err := d.newMem(0); err != nil {
return err
}
// Commit.
if err := cm.commit(d.journalFile.Num(), d.seq); err != nil {
return err
}
// Remove the last journal.
if of != nil {
of.Remove()
}
return nil
}
func (b *Batch) memReplay(to *memdb.DB) error {
return b.decodeRec(func(i int, t vType, key, value []byte) {
ikey := newIKey(key, b.seq+uint64(i), t)
to.Put(ikey, value)
})
}
func (d *DB) recoverLog() (err error) {
s := d.s
icmp := s.cmp
s.printf("LogRecovery: started, min=%d", s.stLogNum)
var mem *memdb.DB
batch := new(Batch)
cm := newCMem(s)
logs, skip := files(s.getFiles(desc.TypeLog)), 0
logs.sort()
for _, log := range logs {
if log.Num() < s.stLogNum {
skip++
continue
}
s.markFileNum(log.Num())
}
var r, fr *logReader
for _, log := range logs[skip:] {
s.printf("LogRecovery: recovering, num=%d", log.Num())
r, err = newLogReader(log, true, s.logDropFunc("log", log.Num()))
if err != nil {
return
}
if mem != nil {
if mem.Len() > 0 {
err = cm.flush(mem, 0)
if err != nil {
return
}
}
err = cm.commit(r.file.Num(), d.seq)
if err != nil {
return
}
cm.reset()
fr.remove()
fr = nil
}
mem = memdb.New(icmp)
for r.log.Next() {
err = batch.decode(r.log.Record())
if err != nil {
return
}
err = batch.memReplay(mem)
if err != nil {
return
}
d.seq = batch.seq + uint64(batch.len())
if mem.Size() > s.o.GetWriteBuffer() {
// flush to table
err = cm.flush(mem, 0)
if err != nil {
return
}
// create new memdb
mem = memdb.New(icmp)
}
}
err = r.log.Error()
if err != nil {
return
}
r.close()
fr = r
}
// create new log
_, err = d.newMem()
if err != nil {
return
}
if mem != nil && mem.Len() > 0 {
err = cm.flush(mem, 0)
if err != nil {
return
}
}
err = cm.commit(d.log.file.Num(), d.seq)
if err != nil {
return
}
if fr != nil {
fr.remove()
}
return
}