// NewWriter creates a new initialized table writer for the file.
//
// Table writer is not goroutine-safe.
func NewWriter(f io.Writer, o *opt.Options) *Writer {
w := &Writer{
writer: f,
cmp: o.GetComparer(),
filter: o.GetFilter(),
compression: o.GetCompression(),
blockSize: o.GetBlockSize(),
comparerScratch: make([]byte, 0),
}
// data block
w.dataBlock.restartInterval = o.GetBlockRestartInterval()
// The first 20-bytes are used for encoding block handle.
w.dataBlock.scratch = w.scratch[20:]
// index block
w.indexBlock.restartInterval = 1
w.indexBlock.scratch = w.scratch[20:]
// filter block
if w.filter != nil {
w.filterBlock.generator = w.filter.NewGenerator()
w.filterBlock.flush(0)
}
return w
}
// NewReader creates a new initialized table reader for the file.
// The cache and bpool is optional and can be nil.
//
// The returned table reader instance is goroutine-safe.
func NewReader(f io.ReaderAt, size int64, cache cache.Namespace, bpool *util.BufferPool, o *opt.Options) *Reader {
if bpool == nil {
bpool = util.NewBufferPool(o.GetBlockSize() + blockTrailerLen)
}
r := &Reader{
reader: f,
cache: cache,
bpool: bpool,
cmp: o.GetComparer(),
checksum: o.GetStrict(opt.StrictBlockChecksum),
strictIter: o.GetStrict(opt.StrictIterator),
}
if f == nil {
r.err = errors.New("leveldb/table: Reader: nil file")
return r
}
if size < footerLen {
r.err = errors.New("leveldb/table: Reader: invalid table (file size is too small)")
return r
}
var footer [footerLen]byte
if _, err := r.reader.ReadAt(footer[:], size-footerLen); err != nil && err != io.EOF {
r.err = fmt.Errorf("leveldb/table: Reader: invalid table (could not read footer): %v", err)
}
if string(footer[footerLen-len(magic):footerLen]) != magic {
r.err = errors.New("leveldb/table: Reader: invalid table (bad magic number)")
return r
}
// Decode the metaindex block handle.
metaBH, n := decodeBlockHandle(footer[:])
if n == 0 {
r.err = errors.New("leveldb/table: Reader: invalid table (bad metaindex block handle)")
return r
}
// Decode the index block handle.
r.indexBH, n = decodeBlockHandle(footer[n:])
if n == 0 {
r.err = errors.New("leveldb/table: Reader: invalid table (bad index block handle)")
return r
}
// Read metaindex block.
metaBlock, err := r.readBlock(metaBH, true)
if err != nil {
r.err = err
return r
}
// Set data end.
r.dataEnd = int64(metaBH.offset)
metaIter := metaBlock.newIterator(nil, false, nil)
for metaIter.Next() {
key := string(metaIter.Key())
if !strings.HasPrefix(key, "filter.") {
continue
}
fn := key[7:]
if f0 := o.GetFilter(); f0 != nil && f0.Name() == fn {
r.filter = f0
} else {
for _, f0 := range o.GetAltFilters() {
if f0.Name() == fn {
r.filter = f0
break
}
}
}
if r.filter != nil {
filterBH, n := decodeBlockHandle(metaIter.Value())
if n == 0 {
continue
}
r.filterBH = filterBH
// Update data end.
r.dataEnd = int64(filterBH.offset)
break
}
}
metaIter.Release()
metaBlock.Release()
return r
}
func recoverTable(s *session, o *opt.Options) error {
// Get all tables and sort it by file number.
tableFiles_, err := s.getFiles(storage.TypeTable)
if err != nil {
return err
}
tableFiles := files(tableFiles_)
tableFiles.sort()
var mSeq uint64
var good, corrupted int
rec := new(sessionRecord)
bpool := util.NewBufferPool(o.GetBlockSize() + 5)
buildTable := func(iter iterator.Iterator) (tmp storage.File, size int64, err error) {
tmp = s.newTemp()
writer, err := tmp.Create()
if err != nil {
return
}
defer func() {
writer.Close()
if err != nil {
tmp.Remove()
tmp = nil
}
}()
// Copy entries.
tw := table.NewWriter(writer, o)
for iter.Next() {
key := iter.Key()
if validIkey(key) {
err = tw.Append(key, iter.Value())
if err != nil {
return
}
}
}
err = iter.Error()
if err != nil {
return
}
err = tw.Close()
if err != nil {
return
}
err = writer.Sync()
if err != nil {
return
}
size = int64(tw.BytesLen())
return
}
recoverTable := func(file storage.File) error {
s.logf("table@recovery recovering @%d", file.Num())
reader, err := file.Open()
if err != nil {
return err
}
defer reader.Close()
// Get file size.
size, err := reader.Seek(0, 2)
if err != nil {
return err
}
var tSeq uint64
var tgood, tcorrupted, blockerr int
var imin, imax []byte
tr := table.NewReader(reader, size, nil, bpool, o)
iter := tr.NewIterator(nil, nil)
iter.(iterator.ErrorCallbackSetter).SetErrorCallback(func(err error) {
s.logf("table@recovery found error @%d %q", file.Num(), err)
blockerr++
})
// Scan the table.
for iter.Next() {
key := iter.Key()
_, seq, _, ok := parseIkey(key)
if !ok {
tcorrupted++
continue
}
tgood++
if seq > tSeq {
tSeq = seq
}
if imin == nil {
imin = append([]byte{}, key...)
}
imax = append(imax[:0], key...)
}
if err := iter.Error(); err != nil {
iter.Release()
return err
}
iter.Release()
if tgood > 0 {
if tcorrupted > 0 || blockerr > 0 {
// Rebuild the table.
s.logf("table@recovery rebuilding @%d", file.Num())
iter := tr.NewIterator(nil, nil)
tmp, newSize, err := buildTable(iter)
iter.Release()
if err != nil {
return err
}
reader.Close()
if err := file.Replace(tmp); err != nil {
return err
}
size = newSize
}
if tSeq > mSeq {
mSeq = tSeq
}
// Add table to level 0.
rec.addTable(0, file.Num(), uint64(size), imin, imax)
s.logf("table@recovery recovered @%d N·%d C·%d B·%d S·%d Q·%d", file.Num(), tgood, tcorrupted, blockerr, size, tSeq)
} else {
s.logf("table@recovery unrecoverable @%d C·%d B·%d S·%d", file.Num(), tcorrupted, blockerr, size)
}
good += tgood
corrupted += tcorrupted
return nil
}
// Recover all tables.
if len(tableFiles) > 0 {
s.logf("table@recovery F·%d", len(tableFiles))
// Mark file number as used.
s.markFileNum(tableFiles[len(tableFiles)-1].Num())
for _, file := range tableFiles {
if err := recoverTable(file); err != nil {
return err
}
}
s.logf("table@recovery recovered F·%d N·%d C·%d Q·%d", len(tableFiles), good, corrupted, mSeq)
}
// Set sequence number.
rec.setSeq(mSeq + 1)
// Create new manifest.
if err := s.create(); err != nil {
return err
}
// Commit.
return s.commit(rec)
}