seaweedfs/weed/storage/volume_vacuum.go

448 lines
14 KiB
Go
Raw Normal View History

2014-05-20 02:24:35 +00:00
package storage
import (
2014-05-20 03:54:39 +00:00
"fmt"
2014-05-20 02:24:35 +00:00
"os"
"runtime"
"time"
"github.com/chrislusf/seaweedfs/weed/glog"
"github.com/chrislusf/seaweedfs/weed/stats"
"github.com/chrislusf/seaweedfs/weed/storage/backend"
idx2 "github.com/chrislusf/seaweedfs/weed/storage/idx"
"github.com/chrislusf/seaweedfs/weed/storage/needle"
"github.com/chrislusf/seaweedfs/weed/storage/needle_map"
2019-12-23 20:48:20 +00:00
"github.com/chrislusf/seaweedfs/weed/storage/super_block"
. "github.com/chrislusf/seaweedfs/weed/storage/types"
"github.com/chrislusf/seaweedfs/weed/util"
2014-05-20 02:24:35 +00:00
)
func (v *Volume) garbageLevel() float64 {
if v.ContentSize() == 0 {
return 0
}
deletedSize := v.DeletedSize()
fileSize := v.ContentSize()
if v.DeletedCount() > 0 && v.DeletedSize() == 0 {
// this happens for .sdx converted back to normal .idx
// where deleted entry size is missing
datFileSize, _, _ := v.FileStat()
deletedSize = datFileSize - fileSize - super_block.SuperBlockSize
fileSize = datFileSize
}
return float64(deletedSize) / float64(fileSize)
2014-05-20 02:24:35 +00:00
}
// compact a volume based on deletions in .dat files
func (v *Volume) Compact(preallocate int64, compactionBytePerSecond int64) error {
2019-09-03 14:05:43 +00:00
2019-12-24 16:13:01 +00:00
if v.MemoryMapMaxSizeMb != 0 { //it makes no sense to compact in memory
2019-09-03 14:05:43 +00:00
return nil
}
2019-12-24 16:13:01 +00:00
glog.V(3).Infof("Compacting volume %d ...", v.Id)
//no need to lock for copy on write
//v.accessLock.Lock()
//defer v.accessLock.Unlock()
//glog.V(3).Infof("Got Compaction lock...")
v.isCompacting = true
defer func() {
v.isCompacting = false
}()
filePath := v.FileName()
v.lastCompactIndexOffset = v.IndexFileSize()
v.lastCompactRevision = v.SuperBlock.CompactionRevision
glog.V(3).Infof("creating copies for volume %d ,last offset %d...", v.Id, v.lastCompactIndexOffset)
if err := v.DataBackend.Sync(); err != nil {
glog.V(0).Infof("compact fail to sync volume %d", v.Id)
}
2020-03-21 06:38:46 +00:00
if err := v.nm.Sync(); err != nil {
glog.V(0).Infof("compact fail to sync volume idx %d", v.Id)
}
2019-12-24 16:13:01 +00:00
return v.copyDataAndGenerateIndexFile(filePath+".cpd", filePath+".cpx", preallocate, compactionBytePerSecond)
2014-05-20 02:24:35 +00:00
}
// compact a volume based on deletions in .idx files
func (v *Volume) Compact2(preallocate int64, compactionBytePerSecond int64) error {
2019-12-24 16:13:01 +00:00
if v.MemoryMapMaxSizeMb != 0 { //it makes no sense to compact in memory
2019-09-03 14:05:43 +00:00
return nil
}
2019-12-24 16:13:01 +00:00
glog.V(3).Infof("Compact2 volume %d ...", v.Id)
v.isCompacting = true
defer func() {
v.isCompacting = false
}()
filePath := v.FileName()
v.lastCompactIndexOffset = v.IndexFileSize()
v.lastCompactRevision = v.SuperBlock.CompactionRevision
2019-12-24 16:13:01 +00:00
glog.V(3).Infof("creating copies for volume %d ...", v.Id)
if err := v.DataBackend.Sync(); err != nil {
2020-04-27 19:41:31 +00:00
glog.V(0).Infof("compact2 fail to sync volume dat %d: %v", v.Id, err)
2020-03-21 06:38:46 +00:00
}
if err := v.nm.Sync(); err != nil {
2020-04-27 19:41:31 +00:00
glog.V(0).Infof("compact2 fail to sync volume idx %d: %v", v.Id, err)
}
return copyDataBasedOnIndexFile(filePath+".dat", filePath+".idx", filePath+".cpd", filePath+".cpx", v.SuperBlock, v.Version(), preallocate, compactionBytePerSecond)
}
2019-03-25 16:16:12 +00:00
func (v *Volume) CommitCompact() error {
2019-12-24 16:13:01 +00:00
if v.MemoryMapMaxSizeMb != 0 { //it makes no sense to compact in memory
return nil
}
glog.V(0).Infof("Committing volume %d vacuuming...", v.Id)
2019-09-03 14:05:43 +00:00
2019-12-24 16:13:01 +00:00
v.isCompacting = true
defer func() {
v.isCompacting = false
}()
2019-09-03 14:05:43 +00:00
2019-12-24 16:13:01 +00:00
v.dataFileAccessLock.Lock()
defer v.dataFileAccessLock.Unlock()
2019-09-03 14:05:43 +00:00
2019-12-24 16:13:01 +00:00
glog.V(3).Infof("Got volume %d committing lock...", v.Id)
v.nm.Close()
2020-02-27 08:07:25 +00:00
if v.DataBackend != nil {
if err := v.DataBackend.Close(); err != nil {
glog.V(0).Infof("fail to close volume %d", v.Id)
}
2019-12-24 16:13:01 +00:00
}
v.DataBackend = nil
stats.VolumeServerVolumeCounter.WithLabelValues(v.Collection, "volume").Dec()
var e error
if e = v.makeupDiff(v.FileName()+".cpd", v.FileName()+".cpx", v.FileName()+".dat", v.FileName()+".idx"); e != nil {
glog.V(0).Infof("makeupDiff in CommitCompact volume %d failed %v", v.Id, e)
e = os.Remove(v.FileName() + ".cpd")
if e != nil {
return e
}
2019-12-24 16:13:01 +00:00
e = os.Remove(v.FileName() + ".cpx")
if e != nil {
return e
}
} else {
if runtime.GOOS == "windows" {
2020-03-10 20:33:26 +00:00
e = os.RemoveAll(v.FileName() + ".dat")
if e != nil {
return e
}
e = os.RemoveAll(v.FileName() + ".idx")
if e != nil {
return e
}
}
var e error
2019-12-24 16:13:01 +00:00
if e = os.Rename(v.FileName()+".cpd", v.FileName()+".dat"); e != nil {
return fmt.Errorf("rename %s: %v", v.FileName()+".cpd", e)
}
2019-12-24 16:13:01 +00:00
if e = os.Rename(v.FileName()+".cpx", v.FileName()+".idx"); e != nil {
return fmt.Errorf("rename %s: %v", v.FileName()+".cpx", e)
}
}
2019-12-24 16:13:01 +00:00
//glog.V(3).Infof("Pretending to be vacuuming...")
//time.Sleep(20 * time.Second)
2019-12-24 16:13:01 +00:00
os.RemoveAll(v.FileName() + ".ldb")
2019-12-24 16:13:01 +00:00
glog.V(3).Infof("Loading volume %d commit file...", v.Id)
if e = v.load(true, false, v.needleMapKind, 0); e != nil {
return e
2014-05-20 02:24:35 +00:00
}
2019-12-24 18:18:56 +00:00
return nil
2014-05-20 02:24:35 +00:00
}
2014-05-20 03:54:39 +00:00
func (v *Volume) cleanupCompact() error {
2018-10-19 03:47:30 +00:00
glog.V(0).Infof("Cleaning up volume %d vacuuming...", v.Id)
e1 := os.Remove(v.FileName() + ".cpd")
e2 := os.Remove(v.FileName() + ".cpx")
if e1 != nil {
return e1
}
if e2 != nil {
return e2
}
return nil
}
func fetchCompactRevisionFromDatFile(datBackend backend.BackendStorageFile) (compactRevision uint16, err error) {
2019-12-23 20:48:20 +00:00
superBlock, err := super_block.ReadSuperBlock(datBackend)
2016-10-07 08:22:24 +00:00
if err != nil {
return 0, err
}
return superBlock.CompactionRevision, nil
2016-10-07 08:22:24 +00:00
}
// if old .dat and .idx files are updated, this func tries to apply the same changes to new files accordingly
func (v *Volume) makeupDiff(newDatFileName, newIdxFileName, oldDatFileName, oldIdxFileName string) (err error) {
var indexSize int64
oldIdxFile, err := os.Open(oldIdxFileName)
defer oldIdxFile.Close()
oldDatFile, err := os.Open(oldDatFileName)
2019-10-30 05:37:36 +00:00
oldDatBackend := backend.NewDiskFile(oldDatFile)
defer oldDatBackend.Close()
// skip if the old .idx file has not changed
if indexSize, err = verifyIndexFileIntegrity(oldIdxFile); err != nil {
return fmt.Errorf("verifyIndexFileIntegrity %s failed: %v", oldIdxFileName, err)
}
2016-10-07 08:22:24 +00:00
if indexSize == 0 || uint64(indexSize) <= v.lastCompactIndexOffset {
return nil
}
// fail if the old .dat file has changed to a new revision
oldDatCompactRevision, err := fetchCompactRevisionFromDatFile(oldDatBackend)
2016-10-07 08:22:24 +00:00
if err != nil {
return fmt.Errorf("fetchCompactRevisionFromDatFile src %s failed: %v", oldDatFile.Name(), err)
2016-10-07 08:22:24 +00:00
}
if oldDatCompactRevision != v.lastCompactRevision {
return fmt.Errorf("current old dat file's compact revision %d is not the expected one %d", oldDatCompactRevision, v.lastCompactRevision)
}
type keyField struct {
offset Offset
2016-10-07 08:22:24 +00:00
size uint32
}
incrementedHasUpdatedIndexEntry := make(map[NeedleId]keyField)
2016-10-07 08:22:24 +00:00
2019-04-19 07:39:34 +00:00
for idxOffset := indexSize - NeedleMapEntrySize; uint64(idxOffset) >= v.lastCompactIndexOffset; idxOffset -= NeedleMapEntrySize {
var IdxEntry []byte
2019-01-17 01:17:19 +00:00
if IdxEntry, err = readIndexEntryAtOffset(oldIdxFile, idxOffset); err != nil {
return fmt.Errorf("readIndexEntry %s at offset %d failed: %v", oldIdxFileName, idxOffset, err)
}
key, offset, size := idx2.IdxFileEntry(IdxEntry)
2018-07-07 07:48:58 +00:00
glog.V(4).Infof("key %d offset %d size %d", key, offset, size)
2016-10-07 08:22:24 +00:00
if _, found := incrementedHasUpdatedIndexEntry[key]; !found {
incrementedHasUpdatedIndexEntry[key] = keyField{
offset: offset,
size: size,
}
}
}
// no updates during commit step
if len(incrementedHasUpdatedIndexEntry) == 0 {
return nil
}
// deal with updates during commit step
var (
dst, idx *os.File
)
if dst, err = os.OpenFile(newDatFileName, os.O_RDWR, 0644); err != nil {
return fmt.Errorf("open dat file %s failed: %v", newDatFileName, err)
}
2019-10-30 05:37:36 +00:00
dstDatBackend := backend.NewDiskFile(dst)
defer dstDatBackend.Close()
if idx, err = os.OpenFile(newIdxFileName, os.O_RDWR, 0644); err != nil {
return fmt.Errorf("open idx file %s failed: %v", newIdxFileName, err)
}
defer idx.Close()
var newDatCompactRevision uint16
newDatCompactRevision, err = fetchCompactRevisionFromDatFile(dstDatBackend)
if err != nil {
return fmt.Errorf("fetchCompactRevisionFromDatFile dst %s failed: %v", dst.Name(), err)
}
if oldDatCompactRevision+1 != newDatCompactRevision {
return fmt.Errorf("oldDatFile %s 's compact revision is %d while newDatFile %s 's compact revision is %d", oldDatFileName, oldDatCompactRevision, newDatFileName, newDatCompactRevision)
}
2016-10-07 08:22:24 +00:00
for key, increIdxEntry := range incrementedHasUpdatedIndexEntry {
2019-07-21 20:50:24 +00:00
idxEntryBytes := needle_map.ToBytes(key, increIdxEntry.offset, increIdxEntry.size)
var offset int64
if offset, err = dst.Seek(0, 2); err != nil {
glog.V(0).Infof("failed to seek the end of file: %v", err)
return
}
//ensure file writing starting from aligned positions
if offset%NeedlePaddingSize != 0 {
offset = offset + (NeedlePaddingSize - offset%NeedlePaddingSize)
2019-07-21 20:50:42 +00:00
if offset, err = dst.Seek(offset, 0); err != nil {
glog.V(0).Infof("failed to align in datafile %s: %v", dst.Name(), err)
return
}
}
2016-10-07 08:22:24 +00:00
//updated needle
2019-04-09 02:40:56 +00:00
if !increIdxEntry.offset.IsZero() && increIdxEntry.size != 0 && increIdxEntry.size != TombstoneFileSize {
//even the needle cache in memory is hit, the need_bytes is correct
2019-04-09 02:40:56 +00:00
glog.V(4).Infof("file %d offset %d size %d", key, increIdxEntry.offset.ToAcutalOffset(), increIdxEntry.size)
var needleBytes []byte
needleBytes, err = needle.ReadNeedleBlob(oldDatBackend, increIdxEntry.offset.ToAcutalOffset(), increIdxEntry.size, v.Version())
if err != nil {
2019-04-09 02:40:56 +00:00
return fmt.Errorf("ReadNeedleBlob %s key %d offset %d size %d failed: %v", oldDatFile.Name(), key, increIdxEntry.offset.ToAcutalOffset(), increIdxEntry.size, err)
2016-10-07 08:22:24 +00:00
}
dst.Write(needleBytes)
util.Uint32toBytes(idxEntryBytes[8:12], uint32(offset/NeedlePaddingSize))
} else { //deleted needle
//fakeDelNeedle 's default Data field is nil
2019-04-19 04:43:36 +00:00
fakeDelNeedle := new(needle.Needle)
fakeDelNeedle.Id = key
fakeDelNeedle.Cookie = 0x12345678
fakeDelNeedle.AppendAtNs = uint64(time.Now().UnixNano())
_, _, _, err = fakeDelNeedle.Append(dstDatBackend, v.Version())
if err != nil {
return fmt.Errorf("append deleted %d failed: %v", key, err)
2016-10-07 08:22:24 +00:00
}
util.Uint32toBytes(idxEntryBytes[8:12], uint32(0))
}
if _, err := idx.Seek(0, 2); err != nil {
return fmt.Errorf("cannot seek end of indexfile %s: %v",
newIdxFileName, err)
}
_, err = idx.Write(idxEntryBytes)
}
2016-09-27 05:30:44 +00:00
return nil
2016-09-27 05:26:39 +00:00
}
type VolumeFileScanner4Vacuum struct {
version needle.Version
v *Volume
dstBackend backend.BackendStorageFile
2019-12-24 18:18:56 +00:00
nm *needle_map.MemDb
newOffset int64
now uint64
writeThrottler *util.WriteThrottler
}
2019-12-23 20:48:20 +00:00
func (scanner *VolumeFileScanner4Vacuum) VisitSuperBlock(superBlock super_block.SuperBlock) error {
scanner.version = superBlock.Version
superBlock.CompactionRevision++
_, err := scanner.dstBackend.WriteAt(superBlock.Bytes(), 0)
scanner.newOffset = int64(superBlock.BlockSize())
return err
}
func (scanner *VolumeFileScanner4Vacuum) ReadNeedleBody() bool {
return true
}
2019-10-22 07:50:30 +00:00
func (scanner *VolumeFileScanner4Vacuum) VisitNeedle(n *needle.Needle, offset int64, needleHeader, needleBody []byte) error {
if n.HasTtl() && scanner.now >= n.LastModified+uint64(scanner.v.Ttl.Minutes()*60) {
return nil
}
nv, ok := scanner.v.nm.Get(n.Id)
glog.V(4).Infoln("needle expected offset ", offset, "ok", ok, "nv", nv)
2019-04-09 02:40:56 +00:00
if ok && nv.Offset.ToAcutalOffset() == offset && nv.Size > 0 && nv.Size != TombstoneFileSize {
2019-12-24 18:18:56 +00:00
if err := scanner.nm.Set(n.Id, ToOffset(scanner.newOffset), n.Size); err != nil {
return fmt.Errorf("cannot put needle: %s", err)
}
if _, _, _, err := n.Append(scanner.dstBackend, scanner.v.Version()); err != nil {
return fmt.Errorf("cannot append needle: %s", err)
}
delta := n.DiskSize(scanner.version)
scanner.newOffset += delta
scanner.writeThrottler.MaybeSlowdown(delta)
glog.V(4).Infoln("saving key", n.Id, "volume offset", offset, "=>", scanner.newOffset, "data_size", n.Size)
}
return nil
}
func (v *Volume) copyDataAndGenerateIndexFile(dstName, idxName string, preallocate int64, compactionBytePerSecond int64) (err error) {
2014-05-20 03:54:39 +00:00
var (
dst backend.BackendStorageFile
2014-05-20 03:54:39 +00:00
)
2020-04-11 21:27:25 +00:00
if dst, err = backend.CreateVolumeFile(dstName, preallocate, 0); err != nil {
2014-05-20 03:54:39 +00:00
return
}
defer dst.Close()
2019-12-24 18:18:56 +00:00
nm := needle_map.NewMemDb()
2020-03-10 20:31:23 +00:00
defer nm.Close()
2014-05-20 03:54:39 +00:00
scanner := &VolumeFileScanner4Vacuum{
v: v,
now: uint64(time.Now().Unix()),
2019-12-24 18:18:56 +00:00
nm: nm,
dstBackend: dst,
writeThrottler: util.NewWriteThrottler(compactionBytePerSecond),
}
err = ScanVolumeFile(v.dir, v.Collection, v.Id, v.needleMapKind, scanner)
2019-12-24 18:18:56 +00:00
if err != nil {
return nil
}
err = nm.SaveToIdx(idxName)
2014-05-20 03:54:39 +00:00
return
}
func copyDataBasedOnIndexFile(srcDatName, srcIdxName, dstDatName, datIdxName string, sb super_block.SuperBlock, version needle.Version, preallocate int64, compactionBytePerSecond int64) (err error) {
var (
srcDatBackend, dstDatBackend backend.BackendStorageFile
dataFile *os.File
)
2020-04-11 21:27:25 +00:00
if dstDatBackend, err = backend.CreateVolumeFile(dstDatName, preallocate, 0); err != nil {
return
}
defer dstDatBackend.Close()
2020-01-08 17:45:03 +00:00
oldNm := needle_map.NewMemDb()
2020-03-10 20:31:23 +00:00
defer oldNm.Close()
2020-01-08 17:45:03 +00:00
newNm := needle_map.NewMemDb()
2020-03-10 20:31:23 +00:00
defer newNm.Close()
if err = oldNm.LoadFromIdx(srcIdxName); err != nil {
return
}
if dataFile, err = os.Open(srcDatName); err != nil {
return err
}
srcDatBackend = backend.NewDiskFile(dataFile)
2020-03-10 20:32:22 +00:00
defer srcDatBackend.Close()
2019-12-24 18:18:56 +00:00
now := uint64(time.Now().Unix())
sb.CompactionRevision++
dstDatBackend.WriteAt(sb.Bytes(), 0)
newOffset := int64(sb.BlockSize())
writeThrottler := util.NewWriteThrottler(compactionBytePerSecond)
2020-01-08 17:45:03 +00:00
oldNm.AscendingVisit(func(value needle_map.NeedleValue) error {
offset, size := value.Offset, value.Size
2020-01-08 17:45:03 +00:00
if offset.IsZero() || size == TombstoneFileSize {
return nil
}
2019-04-19 04:43:36 +00:00
n := new(needle.Needle)
err := n.ReadData(srcDatBackend, offset.ToAcutalOffset(), size, version)
2019-03-19 12:34:43 +00:00
if err != nil {
return nil
}
if n.HasTtl() && now >= n.LastModified+uint64(sb.Ttl.Minutes()*60) {
return nil
}
2020-01-08 17:45:03 +00:00
if err = newNm.Set(n.Id, ToOffset(newOffset), n.Size); err != nil {
return fmt.Errorf("cannot put needle: %s", err)
}
if _, _, _, err = n.Append(dstDatBackend, sb.Version); err != nil {
2020-01-08 17:45:03 +00:00
return fmt.Errorf("cannot append needle: %s", err)
}
delta := n.DiskSize(version)
newOffset += delta
writeThrottler.MaybeSlowdown(delta)
2020-03-10 20:33:26 +00:00
glog.V(4).Infoln("saving key", n.Id, "volume offset", offset, "=>", newOffset, "data_size", n.Size)
2020-01-08 17:45:03 +00:00
return nil
})
newNm.SaveToIdx(datIdxName)
2019-12-24 18:18:56 +00:00
return
}