// NewReader creates a new initialized table reader for the file.
// The fi, cache and bpool is optional and can be nil.
//
// The returned table reader instance is goroutine-safe.
func NewReader(f io.ReaderAt, size int64, fd storage.FileDesc, cache *cache.NamespaceGetter, bpool *util.BufferPool, o *opt.Options) (*Reader, error) {
if f == nil {
return nil, errors.New("leveldb/table: nil file")
}
r := &Reader{
fd: fd,
reader: f,
cache: cache,
bpool: bpool,
o: o,
cmp: o.GetComparer(),
verifyChecksum: o.GetStrict(opt.StrictBlockChecksum),
}
if size < footerLen {
r.err = r.newErrCorrupted(0, size, "table", "too small")
return r, nil
}
footerPos := size - footerLen
var footer [footerLen]byte
if _, err := r.reader.ReadAt(footer[:], footerPos); err != nil && err != io.EOF {
return nil, err
}
if string(footer[footerLen-len(magic):footerLen]) != magic {
r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad magic number")
return r, nil
}
var n int
// Decode the metaindex block handle.
r.metaBH, n = decodeBlockHandle(footer[:])
if n == 0 {
r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad metaindex block handle")
return r, nil
}
// Decode the index block handle.
r.indexBH, n = decodeBlockHandle(footer[n:])
if n == 0 {
r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad index block handle")
return r, nil
}
// Read metaindex block.
metaBlock, err := r.readBlock(r.metaBH, true)
if err != nil {
if errors.IsCorrupted(err) {
r.err = err
return r, nil
} else {
return nil, err
}
}
// Set data end.
r.dataEnd = int64(r.metaBH.offset)
// Read metaindex.
metaIter := r.newBlockIter(metaBlock, nil, nil, true)
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()
// Cache index and filter block locally, since we don't have global cache.
if cache == nil {
r.indexBlock, err = r.readBlock(r.indexBH, true)
if err != nil {
if errors.IsCorrupted(err) {
r.err = err
return r, nil
} else {
return nil, err
}
}
if r.filter != nil {
r.filterBlock, err = r.readFilterBlock(r.filterBH)
if err != nil {
if !errors.IsCorrupted(err) {
return nil, err
}
// Don't use filter then.
r.filter = nil
}
}
}
return r, nil
}
func recoverTable(s *session, o *opt.Options) error {
o = dupOptions(o)
// Mask StrictReader, lets StrictRecovery doing its job.
o.Strict &= ^opt.StrictReader
// Get all tables and sort it by file number.
fds, err := s.stor.List(storage.TypeTable)
if err != nil {
return err
}
sortFds(fds)
var (
maxSeq uint64
recoveredKey, goodKey, corruptedKey, corruptedBlock, droppedTable int
// We will drop corrupted table.
strict = o.GetStrict(opt.StrictRecovery)
noSync = o.GetNoSync()
rec = &sessionRecord{}
bpool = util.NewBufferPool(o.GetBlockSize() + 5)
)
buildTable := func(iter iterator.Iterator) (tmpFd storage.FileDesc, size int64, err error) {
tmpFd = s.newTemp()
writer, err := s.stor.Create(tmpFd)
if err != nil {
return
}
defer func() {
writer.Close()
if err != nil {
s.stor.Remove(tmpFd)
tmpFd = storage.FileDesc{}
}
}()
// Copy entries.
tw := table.NewWriter(writer, o)
for iter.Next() {
key := iter.Key()
if validInternalKey(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
}
if !noSync {
err = writer.Sync()
if err != nil {
return
}
}
size = int64(tw.BytesLen())
return
}
recoverTable := func(fd storage.FileDesc) error {
s.logf("table@recovery recovering @%d", fd.Num)
reader, err := s.stor.Open(fd)
if err != nil {
return err
}
var closed bool
defer func() {
if !closed {
reader.Close()
}
}()
// Get file size.
size, err := reader.Seek(0, 2)
if err != nil {
return err
}
var (
tSeq uint64
tgoodKey, tcorruptedKey, tcorruptedBlock int
imin, imax []byte
)
tr, err := table.NewReader(reader, size, fd, nil, bpool, o)
if err != nil {
return err
}
iter := tr.NewIterator(nil, nil)
if itererr, ok := iter.(iterator.ErrorCallbackSetter); ok {
itererr.SetErrorCallback(func(err error) {
if errors.IsCorrupted(err) {
s.logf("table@recovery block corruption @%d %q", fd.Num, err)
tcorruptedBlock++
}
})
}
// Scan the table.
for iter.Next() {
key := iter.Key()
_, seq, _, kerr := parseInternalKey(key)
if kerr != nil {
tcorruptedKey++
continue
}
tgoodKey++
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()
goodKey += tgoodKey
corruptedKey += tcorruptedKey
corruptedBlock += tcorruptedBlock
if strict && (tcorruptedKey > 0 || tcorruptedBlock > 0) {
droppedTable++
s.logf("table@recovery dropped @%d Gk·%d Ck·%d Cb·%d S·%d Q·%d", fd.Num, tgoodKey, tcorruptedKey, tcorruptedBlock, size, tSeq)
return nil
}
if tgoodKey > 0 {
if tcorruptedKey > 0 || tcorruptedBlock > 0 {
// Rebuild the table.
s.logf("table@recovery rebuilding @%d", fd.Num)
iter := tr.NewIterator(nil, nil)
tmpFd, newSize, err := buildTable(iter)
iter.Release()
if err != nil {
return err
}
closed = true
reader.Close()
if err := s.stor.Rename(tmpFd, fd); err != nil {
return err
}
size = newSize
}
if tSeq > maxSeq {
maxSeq = tSeq
}
recoveredKey += tgoodKey
// Add table to level 0.
rec.addTable(0, fd.Num, size, imin, imax)
s.logf("table@recovery recovered @%d Gk·%d Ck·%d Cb·%d S·%d Q·%d", fd.Num, tgoodKey, tcorruptedKey, tcorruptedBlock, size, tSeq)
} else {
droppedTable++
s.logf("table@recovery unrecoverable @%d Ck·%d Cb·%d S·%d", fd.Num, tcorruptedKey, tcorruptedBlock, size)
}
return nil
}
// Recover all tables.
if len(fds) > 0 {
s.logf("table@recovery F·%d", len(fds))
// Mark file number as used.
s.markFileNum(fds[len(fds)-1].Num)
for _, fd := range fds {
if err := recoverTable(fd); err != nil {
return err
}
}
s.logf("table@recovery recovered F·%d N·%d Gk·%d Ck·%d Q·%d", len(fds), recoveredKey, goodKey, corruptedKey, maxSeq)
}
// Set sequence number.
rec.setSeqNum(maxSeq)
// Create new manifest.
if err := s.create(); err != nil {
return err
}
// Commit.
return s.commit(rec)
}