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"
2014-09-20 19:38:59 +00:00
"time"
2014-10-26 18:34:55 +00:00
2016-06-03 01:09:14 +00:00
"github.com/chrislusf/seaweedfs/weed/glog"
2016-09-29 05:57:23 +00:00
"github.com/chrislusf/seaweedfs/weed/util"
2014-05-20 02:24:35 +00:00
)
func ( v * Volume ) garbageLevel ( ) float64 {
return float64 ( v . nm . DeletedSize ( ) ) / float64 ( v . ContentSize ( ) )
}
func ( v * Volume ) Compact ( ) error {
2014-05-27 00:34:54 +00:00
glog . V ( 3 ) . Infof ( "Compacting ..." )
2014-09-20 19:38:59 +00:00
//no need to lock for copy on write
//v.accessLock.Lock()
//defer v.accessLock.Unlock()
//glog.V(3).Infof("Got Compaction lock...")
2014-05-20 02:24:35 +00:00
filePath := v . FileName ( )
2016-10-07 08:22:24 +00:00
v . lastCompactIndexOffset = v . nm . IndexFileSize ( )
v . lastCompactRevision = v . SuperBlock . CompactRevision
glog . V ( 3 ) . Infof ( "creating copies for volume %d ,last offset %d..." , v . Id , v . lastCompactIndexOffset )
2017-01-20 15:31:11 +00:00
return v . copyDataAndGenerateIndexFile ( filePath + ".cpd" , filePath + ".cpx" , 0 )
2014-05-20 02:24:35 +00:00
}
2016-09-23 03:31:17 +00:00
func ( v * Volume ) Compact2 ( ) error {
glog . V ( 3 ) . Infof ( "Compact2 ..." )
filePath := v . FileName ( )
glog . V ( 3 ) . Infof ( "creating copies for volume %d ..." , v . Id )
return v . copyDataBasedOnIndexFile ( filePath + ".cpd" , filePath + ".cpx" )
}
2014-05-20 02:24:35 +00:00
func ( v * Volume ) commitCompact ( ) error {
2017-01-06 18:22:20 +00:00
glog . V ( 0 ) . Infof ( "Committing vacuuming..." )
2015-05-23 17:16:01 +00:00
v . dataFileAccessLock . Lock ( )
defer v . dataFileAccessLock . Unlock ( )
2014-05-27 00:34:54 +00:00
glog . V ( 3 ) . Infof ( "Got Committing lock..." )
2016-05-09 17:04:21 +00:00
v . nm . Close ( )
2014-05-20 02:24:35 +00:00
_ = v . dataFile . Close ( )
2016-09-29 05:57:23 +00:00
2014-05-20 02:24:35 +00:00
var e error
2016-09-29 05:57:23 +00:00
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 failed %v" , e )
e = os . Remove ( v . FileName ( ) + ".cpd" )
if e != nil {
return e
}
e = os . Remove ( v . FileName ( ) + ".cpx" )
if e != nil {
return e
}
} else {
var e error
if e = os . Rename ( v . FileName ( ) + ".cpd" , v . FileName ( ) + ".dat" ) ; e != nil {
return e
}
if e = os . Rename ( v . FileName ( ) + ".cpx" , v . FileName ( ) + ".idx" ) ; e != nil {
return e
}
2014-05-20 02:24:35 +00:00
}
2016-09-29 05:57:23 +00:00
2014-05-27 00:34:54 +00:00
//glog.V(3).Infof("Pretending to be vacuuming...")
//time.Sleep(20 * time.Second)
2015-07-11 19:20:09 +00:00
glog . V ( 3 ) . Infof ( "Loading Commit file..." )
2017-01-08 19:01:46 +00:00
if e = v . load ( true , false , v . needleMapKind , 0 ) ; e != nil {
2014-05-20 02:24:35 +00:00
return e
}
return nil
}
2014-05-20 03:54:39 +00:00
2016-10-07 08:22:24 +00:00
func fetchCompactRevisionFromDatFile ( file * os . File ) ( compactRevision uint16 , err error ) {
if _ , err = file . Seek ( 0 , 0 ) ; err != nil {
return 0 , fmt . Errorf ( "cannot seek to the beginning of %s: %v" , file . Name ( ) , err )
}
header := make ( [ ] byte , SuperBlockSize )
if _ , e := file . Read ( header ) ; e != nil {
return 0 , fmt . Errorf ( "cannot read file %s 's super block: %v" , file . Name ( ) , e )
}
superBlock , err := ParseSuperBlock ( header )
if err != nil {
return 0 , err
}
return superBlock . CompactRevision , nil
}
2016-09-29 05:57:23 +00:00
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 )
defer oldDatFile . Close ( )
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 {
2016-09-29 05:57:23 +00:00
return nil
}
2016-10-07 08:22:24 +00:00
oldDatCompactRevision , err := fetchCompactRevisionFromDatFile ( oldDatFile )
if err != nil {
return
}
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 uint32
size uint32
}
incrementedHasUpdatedIndexEntry := make ( map [ uint64 ] keyField )
for idx_offset := indexSize - NeedleIndexSize ; uint64 ( idx_offset ) >= v . lastCompactIndexOffset ; idx_offset -= NeedleIndexSize {
2016-09-29 05:57:23 +00:00
var IdxEntry [ ] byte
if IdxEntry , err = readIndexEntryAtOffset ( oldIdxFile , idx_offset ) ; err != nil {
return fmt . Errorf ( "readIndexEntry %s at offset %d failed: %v" , oldIdxFileName , idx_offset , err )
}
key , offset , size := idxFileEntry ( IdxEntry )
2016-10-07 08:22:24 +00:00
if _ , found := incrementedHasUpdatedIndexEntry [ key ] ; ! found {
incrementedHasUpdatedIndexEntry [ key ] = keyField {
2016-09-29 05:57:23 +00:00
offset : offset ,
size : size ,
}
}
}
2016-10-07 08:22:24 +00:00
if len ( incrementedHasUpdatedIndexEntry ) > 0 {
2016-09-29 05:57:23 +00:00
var (
dst , idx * os . File
)
2016-10-07 08:22:24 +00:00
if dst , err = os . OpenFile ( newDatFileName , os . O_RDWR , 0644 ) ; err != nil {
2016-09-29 05:57:23 +00:00
return
}
defer dst . Close ( )
2016-10-07 08:22:24 +00:00
if idx , err = os . OpenFile ( newIdxFileName , os . O_RDWR , 0644 ) ; err != nil {
2016-09-29 05:57:23 +00:00
return
}
defer idx . Close ( )
2016-10-07 08:22:24 +00:00
var newDatCompactRevision uint16
newDatCompactRevision , err = fetchCompactRevisionFromDatFile ( dst )
if err != nil {
return
}
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 )
}
2016-09-29 05:57:23 +00:00
idx_entry_bytes := make ( [ ] byte , 16 )
2016-10-07 08:22:24 +00:00
for key , incre_idx_entry := range incrementedHasUpdatedIndexEntry {
2016-09-29 05:57:23 +00:00
util . Uint64toBytes ( idx_entry_bytes [ 0 : 8 ] , key )
util . Uint32toBytes ( idx_entry_bytes [ 8 : 12 ] , incre_idx_entry . offset )
util . Uint32toBytes ( idx_entry_bytes [ 12 : 16 ] , incre_idx_entry . 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 )
if offset , err = v . dataFile . Seek ( offset , 0 ) ; err != nil {
glog . V ( 0 ) . Infof ( "failed to align in datafile %s: %v" , v . dataFile . Name ( ) , err )
return
}
}
2016-10-07 08:22:24 +00:00
//updated needle
if incre_idx_entry . offset != 0 && incre_idx_entry . size != 0 {
//even the needle cache in memory is hit, the need_bytes is correct
var needle_bytes [ ] byte
2017-06-03 18:44:24 +00:00
needle_bytes , err = ReadNeedleBlob ( oldDatFile , int64 ( incre_idx_entry . offset ) * NeedlePaddingSize , incre_idx_entry . size )
2016-10-07 08:22:24 +00:00
if err != nil {
return
}
dst . Write ( needle_bytes )
util . Uint32toBytes ( idx_entry_bytes [ 8 : 12 ] , uint32 ( offset / NeedlePaddingSize ) )
} else { //deleted needle
//fakeDelNeedle 's default Data field is nil
fakeDelNeedle := new ( Needle )
fakeDelNeedle . Id = key
fakeDelNeedle . Cookie = 0x12345678
2017-01-06 18:22:20 +00:00
_ , _ , err = fakeDelNeedle . Append ( dst , v . Version ( ) )
2016-10-07 08:22:24 +00:00
if err != nil {
return
}
util . Uint32toBytes ( idx_entry_bytes [ 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 ( idx_entry_bytes )
2016-09-29 05:57:23 +00:00
}
}
2016-09-27 05:30:44 +00:00
return nil
2016-09-27 05:26:39 +00:00
}
2017-01-08 19:01:46 +00:00
func ( v * Volume ) copyDataAndGenerateIndexFile ( dstName , idxName string , preallocate int64 ) ( err error ) {
2014-05-20 03:54:39 +00:00
var (
dst , idx * os . File
)
2017-01-08 19:01:46 +00:00
if dst , err = createVolumeFile ( dstName , preallocate ) ; err != nil {
2014-05-20 03:54:39 +00:00
return
}
defer dst . Close ( )
if idx , err = os . OpenFile ( idxName , os . O_WRONLY | os . O_CREATE | os . O_TRUNC , 0644 ) ; err != nil {
return
}
defer idx . Close ( )
2017-05-27 05:51:25 +00:00
nm := NewBtreeNeedleMap ( idx )
2014-05-20 03:54:39 +00:00
new_offset := int64 ( SuperBlockSize )
2014-09-20 19:38:59 +00:00
now := uint64 ( time . Now ( ) . Unix ( ) )
Add boltdb for volume needle map
boltdb is fairly slow to write, about 6 minutes for recreating index
for 1553934 files. Boltdb loads 1,553,934 x 16 = 24,862,944bytes from
disk, and generate the boltdb as large as 134,217,728 bytes in 6
minutes.
To compare, for leveldb, it recreates index in leveldb as large as
27,188,148 bytes in 8 seconds.
For in memory version, it loads the index in
To test the memory consumption, the leveldb or boltdb index are
created. And the server is restarted. Using the benchmark tool to read
lots of files. There are 7 volumes in benchmark collection, each with
about 1553K files.
For leveldb, the memory starts at 142,884KB, and stays at 179,340KB.
For boltdb, the memory starts at 73,756KB, and stays at 144,564KB.
For in-memory, the memory starts at 368,152KB, and stays at 448,032KB.
2015-03-29 18:04:32 +00:00
err = ScanVolumeFile ( v . dir , v . Collection , v . Id , v . needleMapKind ,
func ( superBlock SuperBlock ) error {
2015-05-09 06:43:59 +00:00
superBlock . CompactRevision ++
Add boltdb for volume needle map
boltdb is fairly slow to write, about 6 minutes for recreating index
for 1553934 files. Boltdb loads 1,553,934 x 16 = 24,862,944bytes from
disk, and generate the boltdb as large as 134,217,728 bytes in 6
minutes.
To compare, for leveldb, it recreates index in leveldb as large as
27,188,148 bytes in 8 seconds.
For in memory version, it loads the index in
To test the memory consumption, the leveldb or boltdb index are
created. And the server is restarted. Using the benchmark tool to read
lots of files. There are 7 volumes in benchmark collection, each with
about 1553K files.
For leveldb, the memory starts at 142,884KB, and stays at 179,340KB.
For boltdb, the memory starts at 73,756KB, and stays at 144,564KB.
For in-memory, the memory starts at 368,152KB, and stays at 448,032KB.
2015-03-29 18:04:32 +00:00
_ , err = dst . Write ( superBlock . Bytes ( ) )
return err
} , true , func ( n * Needle , offset int64 ) error {
if n . HasTtl ( ) && now >= n . LastModified + uint64 ( v . Ttl . Minutes ( ) * 60 ) {
return nil
2014-05-20 03:54:39 +00:00
}
Add boltdb for volume needle map
boltdb is fairly slow to write, about 6 minutes for recreating index
for 1553934 files. Boltdb loads 1,553,934 x 16 = 24,862,944bytes from
disk, and generate the boltdb as large as 134,217,728 bytes in 6
minutes.
To compare, for leveldb, it recreates index in leveldb as large as
27,188,148 bytes in 8 seconds.
For in memory version, it loads the index in
To test the memory consumption, the leveldb or boltdb index are
created. And the server is restarted. Using the benchmark tool to read
lots of files. There are 7 volumes in benchmark collection, each with
about 1553K files.
For leveldb, the memory starts at 142,884KB, and stays at 179,340KB.
For boltdb, the memory starts at 73,756KB, and stays at 144,564KB.
For in-memory, the memory starts at 368,152KB, and stays at 448,032KB.
2015-03-29 18:04:32 +00:00
nv , ok := v . nm . Get ( n . Id )
glog . V ( 4 ) . Infoln ( "needle expected offset " , offset , "ok" , ok , "nv" , nv )
if ok && int64 ( nv . Offset ) * NeedlePaddingSize == offset && nv . Size > 0 {
if err = nm . Put ( n . Id , uint32 ( new_offset / NeedlePaddingSize ) , n . Size ) ; err != nil {
return fmt . Errorf ( "cannot put needle: %s" , err )
}
2017-01-06 18:22:20 +00:00
if _ , _ , err := n . Append ( dst , v . Version ( ) ) ; err != nil {
Add boltdb for volume needle map
boltdb is fairly slow to write, about 6 minutes for recreating index
for 1553934 files. Boltdb loads 1,553,934 x 16 = 24,862,944bytes from
disk, and generate the boltdb as large as 134,217,728 bytes in 6
minutes.
To compare, for leveldb, it recreates index in leveldb as large as
27,188,148 bytes in 8 seconds.
For in memory version, it loads the index in
To test the memory consumption, the leveldb or boltdb index are
created. And the server is restarted. Using the benchmark tool to read
lots of files. There are 7 volumes in benchmark collection, each with
about 1553K files.
For leveldb, the memory starts at 142,884KB, and stays at 179,340KB.
For boltdb, the memory starts at 73,756KB, and stays at 144,564KB.
For in-memory, the memory starts at 368,152KB, and stays at 448,032KB.
2015-03-29 18:04:32 +00:00
return fmt . Errorf ( "cannot append needle: %s" , err )
}
new_offset += n . DiskSize ( )
glog . V ( 3 ) . Infoln ( "saving key" , n . Id , "volume offset" , offset , "=>" , new_offset , "data_size" , n . Size )
2014-05-20 03:54:39 +00:00
}
Add boltdb for volume needle map
boltdb is fairly slow to write, about 6 minutes for recreating index
for 1553934 files. Boltdb loads 1,553,934 x 16 = 24,862,944bytes from
disk, and generate the boltdb as large as 134,217,728 bytes in 6
minutes.
To compare, for leveldb, it recreates index in leveldb as large as
27,188,148 bytes in 8 seconds.
For in memory version, it loads the index in
To test the memory consumption, the leveldb or boltdb index are
created. And the server is restarted. Using the benchmark tool to read
lots of files. There are 7 volumes in benchmark collection, each with
about 1553K files.
For leveldb, the memory starts at 142,884KB, and stays at 179,340KB.
For boltdb, the memory starts at 73,756KB, and stays at 144,564KB.
For in-memory, the memory starts at 368,152KB, and stays at 448,032KB.
2015-03-29 18:04:32 +00:00
return nil
} )
2014-05-20 03:54:39 +00:00
return
}
2016-09-23 03:31:17 +00:00
func ( v * Volume ) copyDataBasedOnIndexFile ( dstName , idxName string ) ( err error ) {
var (
dst , idx , oldIndexFile * os . File
)
if dst , err = os . OpenFile ( dstName , os . O_WRONLY | os . O_CREATE | os . O_TRUNC , 0644 ) ; err != nil {
return
}
defer dst . Close ( )
if idx , err = os . OpenFile ( idxName , os . O_WRONLY | os . O_CREATE | os . O_TRUNC , 0644 ) ; err != nil {
return
}
defer idx . Close ( )
if oldIndexFile , err = os . OpenFile ( v . FileName ( ) + ".idx" , os . O_RDONLY , 0644 ) ; err != nil {
return
}
defer oldIndexFile . Close ( )
2017-05-27 05:51:25 +00:00
nm := NewBtreeNeedleMap ( idx )
2016-09-23 03:31:17 +00:00
now := uint64 ( time . Now ( ) . Unix ( ) )
v . SuperBlock . CompactRevision ++
dst . Write ( v . SuperBlock . Bytes ( ) )
new_offset := int64 ( SuperBlockSize )
WalkIndexFile ( oldIndexFile , func ( key uint64 , offset , size uint32 ) error {
2017-01-06 18:22:20 +00:00
if offset == 0 || size == TombstoneFileSize {
2016-09-23 03:31:17 +00:00
return nil
}
nv , ok := v . nm . Get ( key )
if ! ok {
return nil
}
n := new ( Needle )
n . ReadData ( v . dataFile , int64 ( offset ) * NeedlePaddingSize , size , v . Version ( ) )
if n . HasTtl ( ) && now >= n . LastModified + uint64 ( v . Ttl . Minutes ( ) * 60 ) {
return nil
}
glog . V ( 4 ) . Infoln ( "needle expected offset " , offset , "ok" , ok , "nv" , nv )
if nv . Offset == offset && nv . Size > 0 {
if err = nm . Put ( n . Id , uint32 ( new_offset / NeedlePaddingSize ) , n . Size ) ; err != nil {
return fmt . Errorf ( "cannot put needle: %s" , err )
}
2017-01-06 18:22:20 +00:00
if _ , _ , err = n . Append ( dst , v . Version ( ) ) ; err != nil {
2016-09-23 03:31:17 +00:00
return fmt . Errorf ( "cannot append needle: %s" , err )
}
new_offset += n . DiskSize ( )
glog . V ( 3 ) . Infoln ( "saving key" , n . Id , "volume offset" , offset , "=>" , new_offset , "data_size" , n . Size )
}
return nil
} )
return
}