// Creates new initialized table ops instance.
func newTableOps(s *session) *tOps {
var (
cacher cache.Cacher
bcache *cache.Cache
bpool *util.BufferPool
)
if s.o.GetOpenFilesCacheCapacity() > 0 {
cacher = cache.NewLRU(s.o.GetOpenFilesCacheCapacity())
}
if !s.o.GetDisableBlockCache() {
var bcacher cache.Cacher
if s.o.GetBlockCacheCapacity() > 0 {
bcacher = cache.NewLRU(s.o.GetBlockCacheCapacity())
}
bcache = cache.NewCache(bcacher)
}
if !s.o.GetDisableBufferPool() {
bpool = util.NewBufferPool(s.o.GetBlockSize() + 5)
}
return &tOps{
s: s,
noSync: s.o.GetNoSync(),
cache: cache.NewCache(cacher),
bcache: bcache,
bpool: bpool,
}
}
// Creates new initialized table ops instance.
func newTableOps(s *session, cacheCap int) *tOps {
c := cache.NewLRUCache(cacheCap)
return &tOps{
s: s,
cache: c,
cacheNS: c.GetNamespace(0),
bpool: util.NewBufferPool(s.o.GetBlockSize() + 5),
}
}
// Creates new initialized table ops instance.
func newTableOps(s *session) *tOps {
var (
cacher cache.Cacher
bcache *cache.Cache
)
if s.o.GetOpenFilesCacheCapacity() > 0 {
cacher = cache.NewLRU(s.o.GetOpenFilesCacheCapacity())
}
if !s.o.DisableBlockCache {
var bcacher cache.Cacher
if s.o.GetBlockCacheCapacity() > 0 {
bcacher = cache.NewLRU(s.o.GetBlockCacheCapacity())
}
bcache = cache.NewCache(bcacher)
}
return &tOps{
s: s,
cache: cache.NewCache(cacher),
bcache: bcache,
bpool: util.NewBufferPool(s.o.GetBlockSize() + 5),
}
}
// 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.
indexBH, n := decodeBlockHandle(footer[n:])
if n == 0 {
r.err = errors.New("leveldb/table: Reader: invalid table (bad index block handle)")
return r
}
// Read index block.
r.indexBlock, r.err = r.readBlock(indexBH, true)
if r.err != nil {
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:]
var filter filter.Filter
if f0 := o.GetFilter(); f0 != nil && f0.Name() == fn {
filter = f0
} else {
for _, f0 := range o.GetAltFilters() {
if f0.Name() == fn {
filter = f0
break
}
}
}
if filter != nil {
filterBH, n := decodeBlockHandle(metaIter.Value())
if n == 0 {
continue
}
// Update data end.
r.dataEnd = int64(filterBH.offset)
filterBlock, err := r.readFilterBlock(filterBH, filter)
if err != nil {
continue
}
r.filterBlock = filterBlock
break
}
}
metaIter.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)
}
func main() {
flag.Parse()
if enableBufferPool {
bpool = util.NewBufferPool(opt.DefaultBlockSize + 128)
}
log.Printf("Test DB stored at %q", dbPath)
if httpProf != "" {
log.Printf("HTTP pprof listening at %q", httpProf)
runtime.SetBlockProfileRate(1)
go func() {
if err := http.ListenAndServe(httpProf, nil); err != nil {
log.Fatalf("HTTPPROF: %v", err)
}
}()
}
runtime.GOMAXPROCS(runtime.NumCPU())
os.RemoveAll(dbPath)
stor, err := storage.OpenFile(dbPath, false)
if err != nil {
log.Fatal(err)
}
tstor := &testingStorage{stor}
defer tstor.Close()
fatalf := func(err error, format string, v ...interface{}) {
atomic.StoreUint32(&fail, 1)
atomic.StoreUint32(&done, 1)
log.Printf("FATAL: "+format, v...)
if err != nil && errors.IsCorrupted(err) {
cerr := err.(*errors.ErrCorrupted)
if !cerr.Fd.Zero() && cerr.Fd.Type == storage.TypeTable {
log.Print("FATAL: corruption detected, scanning...")
if !tstor.scanTable(storage.FileDesc{Type: storage.TypeTable, Num: cerr.Fd.Num}, false) {
log.Printf("FATAL: unable to find corrupted key/value pair in table %v", cerr.Fd)
}
}
}
runtime.Goexit()
}
if openFilesCacheCapacity == 0 {
openFilesCacheCapacity = -1
}
o := &opt.Options{
OpenFilesCacheCapacity: openFilesCacheCapacity,
DisableBufferPool: !enableBufferPool,
DisableBlockCache: !enableBlockCache,
ErrorIfExist: true,
Compression: opt.NoCompression,
}
if enableCompression {
o.Compression = opt.DefaultCompression
}
db, err := leveldb.Open(tstor, o)
if err != nil {
log.Fatal(err)
}
defer db.Close()
var (
mu = &sync.Mutex{}
gGetStat = &latencyStats{}
gIterStat = &latencyStats{}
gWriteStat = &latencyStats{}
gTrasactionStat = &latencyStats{}
startTime = time.Now()
writeReq = make(chan *leveldb.Batch)
writeAck = make(chan error)
writeAckAck = make(chan struct{})
)
go func() {
for b := range writeReq {
var err error
if mrand.Float64() < transactionProb {
log.Print("> Write using transaction")
gTrasactionStat.start()
var tr *leveldb.Transaction
if tr, err = db.OpenTransaction(); err == nil {
if err = tr.Write(b, nil); err == nil {
if err = tr.Commit(); err == nil {
gTrasactionStat.record(b.Len())
}
} else {
tr.Discard()
}
}
} else {
gWriteStat.start()
if err = db.Write(b, nil); err == nil {
gWriteStat.record(b.Len())
}
}
writeAck <- err
<-writeAckAck
}
}()
go func() {
for {
time.Sleep(3 * time.Second)
log.Print("------------------------")
log.Printf("> Elapsed=%v", time.Now().Sub(startTime))
mu.Lock()
log.Printf("> GetLatencyMin=%v GetLatencyMax=%v GetLatencyAvg=%v GetRatePerSec=%d",
gGetStat.min, gGetStat.max, gGetStat.avg(), gGetStat.ratePerSec())
log.Printf("> IterLatencyMin=%v IterLatencyMax=%v IterLatencyAvg=%v IterRatePerSec=%d",
gIterStat.min, gIterStat.max, gIterStat.avg(), gIterStat.ratePerSec())
log.Printf("> WriteLatencyMin=%v WriteLatencyMax=%v WriteLatencyAvg=%v WriteRatePerSec=%d",
gWriteStat.min, gWriteStat.max, gWriteStat.avg(), gWriteStat.ratePerSec())
log.Printf("> TransactionLatencyMin=%v TransactionLatencyMax=%v TransactionLatencyAvg=%v TransactionRatePerSec=%d",
gTrasactionStat.min, gTrasactionStat.max, gTrasactionStat.avg(), gTrasactionStat.ratePerSec())
mu.Unlock()
cachedblock, _ := db.GetProperty("leveldb.cachedblock")
openedtables, _ := db.GetProperty("leveldb.openedtables")
alivesnaps, _ := db.GetProperty("leveldb.alivesnaps")
aliveiters, _ := db.GetProperty("leveldb.aliveiters")
blockpool, _ := db.GetProperty("leveldb.blockpool")
log.Printf("> BlockCache=%s OpenedTables=%s AliveSnaps=%s AliveIter=%s BlockPool=%q",
cachedblock, openedtables, alivesnaps, aliveiters, blockpool)
log.Print("------------------------")
}
}()
for ns, numKey := range numKeys {
func(ns, numKey int) {
log.Printf("[%02d] STARTING: numKey=%d", ns, numKey)
keys := make([][]byte, numKey)
for i := range keys {
keys[i] = randomData(nil, byte(ns), 1, uint32(i), keyLen)
}
wg.Add(1)
go func() {
var wi uint32
defer func() {
log.Printf("[%02d] WRITER DONE #%d", ns, wi)
wg.Done()
}()
var (
b = new(leveldb.Batch)
k2, v2 []byte
nReader int32
)
for atomic.LoadUint32(&done) == 0 {
log.Printf("[%02d] WRITER #%d", ns, wi)
b.Reset()
for _, k1 := range keys {
k2 = randomData(k2, byte(ns), 2, wi, keyLen)
v2 = randomData(v2, byte(ns), 3, wi, valueLen)
b.Put(k2, v2)
b.Put(k1, k2)
}
writeReq <- b
if err := <-writeAck; err != nil {
writeAckAck <- struct{}{}
fatalf(err, "[%02d] WRITER #%d db.Write: %v", ns, wi, err)
}
snap, err := db.GetSnapshot()
if err != nil {
writeAckAck <- struct{}{}
fatalf(err, "[%02d] WRITER #%d db.GetSnapshot: %v", ns, wi, err)
}
writeAckAck <- struct{}{}
wg.Add(1)
atomic.AddInt32(&nReader, 1)
go func(snapwi uint32, snap *leveldb.Snapshot) {
var (
ri int
iterStat = &latencyStats{}
getStat = &latencyStats{}
)
defer func() {
mu.Lock()
gGetStat.add(getStat)
gIterStat.add(iterStat)
mu.Unlock()
atomic.AddInt32(&nReader, -1)
log.Printf("[%02d] READER #%d.%d DONE Snap=%v Alive=%d IterLatency=%v GetLatency=%v", ns, snapwi, ri, snap, atomic.LoadInt32(&nReader), iterStat.avg(), getStat.avg())
snap.Release()
wg.Done()
}()
stopi := snapwi + 3
for (ri < 3 || atomic.LoadUint32(&wi) < stopi) && atomic.LoadUint32(&done) == 0 {
var n int
iter := snap.NewIterator(dataPrefixSlice(byte(ns), 1), nil)
iterStat.start()
for iter.Next() {
k1 := iter.Key()
k2 := iter.Value()
iterStat.record(1)
if dataNS(k2) != byte(ns) {
fatalf(nil, "[%02d] READER #%d.%d K%d invalid in-key NS: want=%d got=%d", ns, snapwi, ri, n, ns, dataNS(k2))
}
kwritei := dataI(k2)
if kwritei != snapwi {
fatalf(nil, "[%02d] READER #%d.%d K%d invalid in-key iter num: %d", ns, snapwi, ri, n, kwritei)
}
getStat.start()
v2, err := snap.Get(k2, nil)
if err != nil {
fatalf(err, "[%02d] READER #%d.%d K%d snap.Get: %v\nk1: %x\n -> k2: %x", ns, snapwi, ri, n, err, k1, k2)
}
getStat.record(1)
if checksum0, checksum1 := dataChecksum(v2); checksum0 != checksum1 {
err := &errors.ErrCorrupted{Fd: storage.FileDesc{0xff, 0}, Err: fmt.Errorf("v2: %x: checksum mismatch: %v vs %v", v2, checksum0, checksum1)}
fatalf(err, "[%02d] READER #%d.%d K%d snap.Get: %v\nk1: %x\n -> k2: %x", ns, snapwi, ri, n, err, k1, k2)
}
n++
iterStat.start()
}
iter.Release()
if err := iter.Error(); err != nil {
fatalf(err, "[%02d] READER #%d.%d K%d iter.Error: %v", ns, snapwi, ri, numKey, err)
}
if n != numKey {
fatalf(nil, "[%02d] READER #%d.%d missing keys: want=%d got=%d", ns, snapwi, ri, numKey, n)
}
ri++
}
}(wi, snap)
atomic.AddUint32(&wi, 1)
}
}()
delB := new(leveldb.Batch)
wg.Add(1)
go func() {
var (
i int
iterStat = &latencyStats{}
)
defer func() {
log.Printf("[%02d] SCANNER DONE #%d", ns, i)
wg.Done()
}()
time.Sleep(2 * time.Second)
for atomic.LoadUint32(&done) == 0 {
var n int
delB.Reset()
iter := db.NewIterator(dataNsSlice(byte(ns)), nil)
iterStat.start()
for iter.Next() && atomic.LoadUint32(&done) == 0 {
k := iter.Key()
v := iter.Value()
iterStat.record(1)
for ci, x := range [...][]byte{k, v} {
checksum0, checksum1 := dataChecksum(x)
if checksum0 != checksum1 {
if ci == 0 {
fatalf(nil, "[%02d] SCANNER %d.%d invalid key checksum: want %d, got %d\n%x -> %x", ns, i, n, checksum0, checksum1, k, v)
} else {
fatalf(nil, "[%02d] SCANNER %d.%d invalid value checksum: want %d, got %d\n%x -> %x", ns, i, n, checksum0, checksum1, k, v)
}
}
}
if dataPrefix(k) == 2 || mrand.Int()%999 == 0 {
delB.Delete(k)
}
n++
iterStat.start()
}
iter.Release()
if err := iter.Error(); err != nil {
fatalf(err, "[%02d] SCANNER #%d.%d iter.Error: %v", ns, i, n, err)
}
if n > 0 {
log.Printf("[%02d] SCANNER #%d IterLatency=%v", ns, i, iterStat.avg())
}
if delB.Len() > 0 && atomic.LoadUint32(&done) == 0 {
t := time.Now()
writeReq <- delB
if err := <-writeAck; err != nil {
writeAckAck <- struct{}{}
fatalf(err, "[%02d] SCANNER #%d db.Write: %v", ns, i, err)
} else {
writeAckAck <- struct{}{}
}
log.Printf("[%02d] SCANNER #%d Deleted=%d Time=%v", ns, i, delB.Len(), time.Now().Sub(t))
}
i++
}
}()
}(ns, numKey)
}
go func() {
sig := make(chan os.Signal)
signal.Notify(sig, os.Interrupt, os.Kill)
log.Printf("Got signal: %v, exiting...", <-sig)
atomic.StoreUint32(&done, 1)
}()
wg.Wait()
}
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.
tableFiles_, err := s.getFiles(storage.TypeTable)
if err != nil {
return err
}
tableFiles := files(tableFiles_)
tableFiles.sort()
var (
maxSeq uint64
recoveredKey, goodKey, corruptedKey, corruptedBlock, droppedTable int
// We will drop corrupted table.
strict = o.GetStrict(opt.StrictRecovery)
rec = &sessionRecord{numLevel: o.GetNumLevel()}
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
}
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, storage.NewFileInfo(file), nil, bpool, o)
if err != nil {
return err
}
iter := tr.NewIterator(nil, nil)
iter.(iterator.ErrorCallbackSetter).SetErrorCallback(func(err error) {
if errors.IsCorrupted(err) {
s.logf("table@recovery block corruption @%d %q", file.Num(), err)
tcorruptedBlock++
}
})
// Scan the table.
for iter.Next() {
key := iter.Key()
_, seq, _, kerr := parseIkey(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", file.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", file.Num())
iter := tr.NewIterator(nil, nil)
tmp, newSize, err := buildTable(iter)
iter.Release()
if err != nil {
return err
}
closed = true
reader.Close()
if err := file.Replace(tmp); err != nil {
return err
}
size = newSize
}
if tSeq > maxSeq {
maxSeq = tSeq
}
recoveredKey += tgoodKey
// Add table to level 0.
rec.addTable(0, file.Num(), uint64(size), imin, imax)
s.logf("table@recovery recovered @%d Gk·%d Ck·%d Cb·%d S·%d Q·%d", file.Num(), tgoodKey, tcorruptedKey, tcorruptedBlock, size, tSeq)
} else {
droppedTable++
s.logf("table@recovery unrecoverable @%d Ck·%d Cb·%d S·%d", file.Num(), tcorruptedKey, tcorruptedBlock, size)
}
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 Gk·%d Ck·%d Q·%d", len(tableFiles), 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)
}