// makeRoomForWrite ensures that there is room in d.mem for the next write.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) makeRoomForWrite(force bool) error {
allowDelay := !force
for {
// TODO: check any previous sticky error, if the paranoid option is set.
if allowDelay && len(d.versions.currentVersion().files[0]) > l0SlowdownWritesTrigger {
// We are getting close to hitting a hard limit on the number of
// L0 files. Rather than delaying a single write by several
// seconds when we hit the hard limit, start delaying each
// individual write by 1ms to reduce latency variance.
d.mu.Unlock()
time.Sleep(1 * time.Millisecond)
d.mu.Lock()
allowDelay = false
// TODO: how do we ensure we are still 'at the front of the writer queue'?
continue
}
if !force && d.mem.ApproximateMemoryUsage() <= d.opts.GetWriteBufferSize() {
// There is room in the current memtable.
break
}
if d.imm != nil {
// We have filled up the current memtable, but the previous
// one is still being compacted, so we wait.
d.compactionCond.Wait()
continue
}
if len(d.versions.currentVersion().files[0]) > l0StopWritesTrigger {
// There are too many level-0 files.
d.compactionCond.Wait()
continue
}
// Attempt to switch to a new memtable and trigger compaction of old
// TODO: drop and re-acquire d.mu around the I/O.
newLogNumber := d.versions.nextFileNum()
newLogFile, err := d.opts.GetFileSystem().Create(dbFilename(d.dirname, fileTypeLog, newLogNumber))
if err != nil {
return err
}
newLog := record.NewWriter(newLogFile)
if err := d.log.Close(); err != nil {
newLogFile.Close()
return err
}
if err := d.logFile.Close(); err != nil {
newLog.Close()
newLogFile.Close()
return err
}
d.logNumber, d.logFile, d.log = newLogNumber, newLogFile, newLog
d.imm, d.mem = d.mem, memdb.New(&d.icmpOpts)
force = false
d.maybeScheduleCompaction()
}
return nil
}
func createDB(dirname string, opts *db.Options) (retErr error) {
const manifestFileNum = 1
ve := versionEdit{
comparatorName: opts.GetComparer().Name(),
nextFileNumber: manifestFileNum + 1,
}
manifestFilename := dbFilename(dirname, fileTypeManifest, manifestFileNum)
f, err := opts.GetFileSystem().Create(manifestFilename)
if err != nil {
return fmt.Errorf("leveldb: could not create %q: %v", manifestFilename, err)
}
defer func() {
if retErr != nil {
opts.GetFileSystem().Remove(manifestFilename)
}
}()
defer f.Close()
recWriter := record.NewWriter(f)
w, err := recWriter.Next()
if err != nil {
return err
}
err = ve.encode(w)
if err != nil {
return err
}
err = recWriter.Close()
if err != nil {
return err
}
return setCurrentFile(dirname, opts.GetFileSystem(), manifestFileNum)
}
// createManifest creates a manifest file that contains a snapshot of vs.
func (vs *versionSet) createManifest(dirname string) (err error) {
var (
filename = dbFilename(dirname, fileTypeManifest, vs.manifestFileNumber)
manifestFile db.File
manifest *record.Writer
)
defer func() {
if manifest != nil {
manifest.Close()
}
if manifestFile != nil {
manifestFile.Close()
}
if err != nil {
vs.fs.Remove(filename)
}
}()
manifestFile, err = vs.fs.Create(filename)
if err != nil {
return err
}
manifest = record.NewWriter(manifestFile)
snapshot := versionEdit{
comparatorName: vs.ucmp.Name(),
}
// TODO: save compaction pointers.
for level, fileMetadata := range vs.currentVersion().files {
for _, meta := range fileMetadata {
snapshot.newFiles = append(snapshot.newFiles, newFileEntry{
level: level,
meta: meta,
})
}
}
w, err1 := manifest.Next()
if err1 != nil {
return err1
}
err1 = snapshot.encode(w)
if err1 != nil {
return err1
}
vs.manifest, manifest = manifest, nil
vs.manifestFile, manifestFile = manifestFile, nil
return nil
}
// Open opens a LevelDB whose files live in the given directory.
func Open(dirname string, opts *db.Options) (*DB, error) {
d := &DB{
dirname: dirname,
opts: opts,
icmp: internalKeyComparer{opts.GetComparer()},
pendingOutputs: make(map[uint64]struct{}),
}
if opts != nil {
d.icmpOpts = *opts
}
d.icmpOpts.Comparer = d.icmp
tableCacheSize := opts.GetMaxOpenFiles() - numNonTableCacheFiles
if tableCacheSize < minTableCacheSize {
tableCacheSize = minTableCacheSize
}
d.tableCache.init(dirname, opts.GetFileSystem(), &d.icmpOpts, tableCacheSize)
d.mem = memdb.New(&d.icmpOpts)
d.compactionCond = sync.Cond{L: &d.mu}
fs := opts.GetFileSystem()
d.mu.Lock()
defer d.mu.Unlock()
// Lock the database directory.
err := fs.MkdirAll(dirname, 0755)
if err != nil {
return nil, err
}
fileLock, err := fs.Lock(dbFilename(dirname, fileTypeLock, 0))
if err != nil {
return nil, err
}
defer func() {
if fileLock != nil {
fileLock.Close()
}
}()
if _, err := fs.Stat(dbFilename(dirname, fileTypeCurrent, 0)); os.IsNotExist(err) {
// Create the DB if it did not already exist.
if err := createDB(dirname, opts); err != nil {
return nil, err
}
} else if err != nil {
return nil, fmt.Errorf("leveldb: database %q: %v", dirname, err)
} else if opts.GetErrorIfDBExists() {
return nil, fmt.Errorf("leveldb: database %q already exists", dirname)
}
// Load the version set.
err = d.versions.load(dirname, opts)
if err != nil {
return nil, err
}
// Replay any newer log files than the ones named in the manifest.
var ve versionEdit
ls, err := fs.List(dirname)
if err != nil {
return nil, err
}
var logFiles fileNumAndNameSlice
for _, filename := range ls {
ft, fn, ok := parseDBFilename(filename)
if ok && ft == fileTypeLog && (fn >= d.versions.logNumber || fn == d.versions.prevLogNumber) {
logFiles = append(logFiles, fileNumAndName{fn, filename})
}
}
sort.Sort(logFiles)
for _, lf := range logFiles {
maxSeqNum, err := d.replayLogFile(&ve, fs, filepath.Join(dirname, lf.name))
if err != nil {
return nil, err
}
d.versions.markFileNumUsed(lf.num)
if d.versions.lastSequence < maxSeqNum {
d.versions.lastSequence = maxSeqNum
}
}
// Create an empty .log file.
ve.logNumber = d.versions.nextFileNum()
d.logNumber = ve.logNumber
logFile, err := fs.Create(dbFilename(dirname, fileTypeLog, ve.logNumber))
if err != nil {
return nil, err
}
defer func() {
if logFile != nil {
logFile.Close()
}
}()
d.log = record.NewWriter(logFile)
// Write a new manifest to disk.
if err := d.versions.logAndApply(dirname, &ve); err != nil {
return nil, err
}
d.deleteObsoleteFiles()
d.maybeScheduleCompaction()
d.logFile, logFile = logFile, nil
d.fileLock, fileLock = fileLock, nil
return d, nil
}