// writeLevel0Table writes a memtable to a level-0 on-disk table.
//
// If no error is returned, it adds the file number of that on-disk table to
// d.pendingOutputs. It is the caller's responsibility to remove that fileNum
// from that set when it has been applied to d.versions.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) writeLevel0Table(fs db.FileSystem, mem *memdb.MemDB) (meta fileMetadata, err error) {
// meta用于记录新创建的level0 db文件的元信息
meta.fileNum = d.versions.nextFileNum()
// filename：新db文件的文件名
filename := dbFilename(d.dirname, fileTypeTable, meta.fileNum)
d.pendingOutputs[meta.fileNum] = struct{}{}
defer func(fileNum uint64) {
// 如果异常退出(err不为nil)，则从d.pendingOutputs中删除新db文件的记录
// 否则d.pendingOutputs是用来干什么的呢？
if err != nil {
delete(d.pendingOutputs, fileNum)
}
}(meta.fileNum)
// Release the d.mu lock while doing I/O.
// Note the unusual order: Unlock and then Lock.
d.mu.Unlock()
defer d.mu.Lock()
var (
file db.File
tw *table.Writer
iter db.Iterator
)
defer func() {
if iter != nil {
err = firstError(err, iter.Close())
}
if tw != nil {
err = firstError(err, tw.Close())
}
if file != nil {
err = firstError(err, file.Close())
}
if err != nil {
fs.Remove(filename)
meta = fileMetadata{}
}
}()
file, err = fs.Create(filename)
if err != nil {
return fileMetadata{}, err
}
// table为磁盘db文件封装写入方式
tw = table.NewWriter(file, &db.Options{
Comparer: d.icmp,
})
// Find返回一个迭代器，用于遍历mem（这里即是d.imm）中的数据
// memtable是以skiplist来组织数据的，有序
// 所以取到的第一个数据的key即为当前imm中最小的key
iter = mem.Find(nil, nil)
iter.Next()
// meta.smallest记录新db文件中最小的内部key，在写入memtable时就已经将用户key封装成了内部key，那么为什么还要使用internalKey来做类型转换？
// 下面的meta.largest就没有进行类型转换就调用clone()方法了。
meta.smallest = internalKey(iter.Key()).clone()
for {
// 最后一次循环中的key即为最大的key，但为什么不封装成内部key呢？
meta.largest = iter.Key()
// 将key、value写到新db文件中
if err1 := tw.Set(meta.largest, iter.Value(), nil); err1 != nil {
return fileMetadata{}, err1
}
// 如果imm中的数据已经遍历完，全部存入新db文件，则break
if !iter.Next() {
break
}
}
meta.largest = meta.largest.clone()
if err1 := iter.Close(); err1 != nil {
iter = nil
return fileMetadata{}, err1
}
iter = nil
if err1 := tw.Close(); err1 != nil {
tw = nil
return fileMetadata{}, err1
}
tw = nil
// TODO: currently, closing a table.Writer closes its underlying file.
// We have to re-open the file to Sync or Stat it, which seems stupid.
file, err = fs.Open(filename)
if err != nil {
return fileMetadata{}, err
}
if err1 := file.Sync(); err1 != nil {
return fileMetadata{}, err1
}
if stat, err1 := file.Stat(); err1 != nil {
return fileMetadata{}, err1
} else {
size := stat.Size()
if size < 0 {
return fileMetadata{}, fmt.Errorf("leveldb: table file %q has negative size %d", filename, size)
}
// 将文件的大小值存入meta.size
meta.size = uint64(size)
}
// TODO: compaction stats.
/* 此时，meta的四个成员：
- filenum
- smallest
- largest
- size
已经全部填上了
*/
return meta, nil
}
// writeLevel0Table writes a memtable to a level-0 on-disk table.
//
// If no error is returned, it adds the file number of that on-disk table to
// d.pendingOutputs. It is the caller's responsibility to remove that fileNum
// from that set when it has been applied to d.versions.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) writeLevel0Table(fs db.FileSystem, mem *memdb.MemDB) (meta fileMetadata, err error) {
meta.fileNum = d.versions.nextFileNum()
filename := dbFilename(d.dirname, fileTypeTable, meta.fileNum)
d.pendingOutputs[meta.fileNum] = struct{}{}
defer func(fileNum uint64) {
if err != nil {
delete(d.pendingOutputs, fileNum)
}
}(meta.fileNum)
// Release the d.mu lock while doing I/O.
// Note the unusual order: Unlock and then Lock.
d.mu.Unlock()
defer d.mu.Lock()
var (
file db.File
tw *table.Writer
iter db.Iterator
)
defer func() {
if iter != nil {
err = firstError(err, iter.Close())
}
if tw != nil {
err = firstError(err, tw.Close())
}
if file != nil {
err = firstError(err, file.Close())
}
if err != nil {
fs.Remove(filename)
meta = fileMetadata{}
}
}()
file, err = fs.Create(filename)
if err != nil {
return fileMetadata{}, err
}
tw = table.NewWriter(file, &db.Options{
Comparer: d.icmp,
})
iter = mem.Find(nil, nil)
iter.Next()
meta.smallest = internalKey(iter.Key()).clone()
for {
meta.largest = iter.Key()
if err1 := tw.Set(meta.largest, iter.Value(), nil); err1 != nil {
return fileMetadata{}, err1
}
if !iter.Next() {
break
}
}
meta.largest = meta.largest.clone()
if err1 := iter.Close(); err1 != nil {
iter = nil
return fileMetadata{}, err1
}
iter = nil
if err1 := tw.Close(); err1 != nil {
tw = nil
return fileMetadata{}, err1
}
tw = nil
// TODO: currently, closing a table.Writer closes its underlying file.
// We have to re-open the file to Sync or Stat it, which seems stupid.
file, err = fs.Open(filename)
if err != nil {
return fileMetadata{}, err
}
if err1 := file.Sync(); err1 != nil {
return fileMetadata{}, err1
}
if stat, err1 := file.Stat(); err1 != nil {
return fileMetadata{}, err1
} else {
size := stat.Size()
if size < 0 {
return fileMetadata{}, fmt.Errorf("leveldb: table file %q has negative size %d", filename, size)
}
meta.size = uint64(size)
}
// TODO: compaction stats.
return meta, nil
}
func (d *DB) writeLevel0Table(fs db.FileSystem, mem *memdb.MemDB) (meta fileMetadata, err error) {
meta.fileNum = d.versions.nextFileNum()
filename := dbFilename(d.dirname, fileTypeTable, meta.fileNum)
// TODO: add meta.fileNum to a set of 'pending outputs' so that a
// concurrent sweep of obsolete db files won't delete the fileNum file.
// It is the caller's responsibility to remove that fileNum from the
// set of pending outputs.
var (
file db.File
tw *table.Writer
iter db.Iterator
)
defer func() {
if iter != nil {
err = firstError(err, iter.Close())
}
if tw != nil {
err = firstError(err, tw.Close())
}
if file != nil {
err = firstError(err, file.Close())
}
if err != nil {
fs.Remove(filename)
meta = fileMetadata{}
}
}()
file, err = fs.Create(filename)
if err != nil {
return fileMetadata{}, err
}
tw = table.NewWriter(file, &db.Options{
Comparer: d.icmp,
})
iter = mem.Find(nil, nil)
iter.Next()
meta.smallest = internalKey(iter.Key()).clone()
for {
meta.largest = iter.Key()
if err1 := tw.Set(meta.largest, iter.Value(), nil); err1 != nil {
return fileMetadata{}, err1
}
if !iter.Next() {
break
}
}
meta.largest = meta.largest.clone()
if err1 := iter.Close(); err1 != nil {
iter = nil
return fileMetadata{}, err1
}
iter = nil
if err1 := tw.Close(); err1 != nil {
tw = nil
return fileMetadata{}, err1
}
tw = nil
// TODO: currently, closing a table.Writer closes its underlying file.
// We have to re-open the file to Sync or Stat it, which seems stupid.
file, err = fs.Open(filename)
if err != nil {
return fileMetadata{}, err
}
if err1 := file.Sync(); err1 != nil {
return fileMetadata{}, err1
}
if stat, err1 := file.Stat(); err1 != nil {
return fileMetadata{}, err1
} else {
size := stat.Size()
if size < 0 {
return fileMetadata{}, fmt.Errorf("leveldb: table file %q has negative size %d", filename, size)
}
meta.size = uint64(size)
}
// TODO: compaction stats.
return meta, nil
}
// compactDiskTables runs a compaction that produces new on-disk tables from
// old on-disk tables.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) compactDiskTables(c *compaction) (ve *versionEdit, pendingOutputs []uint64, retErr error) {
defer func() {
if retErr != nil {
for _, fileNum := range pendingOutputs {
delete(d.pendingOutputs, fileNum)
}
pendingOutputs = nil
}
}()
// TODO: track snapshots.
smallestSnapshot := d.versions.lastSequence
// Release the d.mu lock while doing I/O.
// Note the unusual order: Unlock and then Lock.
d.mu.Unlock()
defer d.mu.Lock()
iter, err := compactionIterator(&d.tableCache, d.icmp, c)
if err != nil {
return nil, pendingOutputs, err
}
// TODO: output to more than one table, if it would otherwise be too large.
var (
fileNum uint64
filename string
tw *table.Writer
)
defer func() {
if iter != nil {
retErr = firstError(retErr, iter.Close())
}
if tw != nil {
retErr = firstError(retErr, tw.Close())
}
if retErr != nil {
d.opts.GetFileSystem().Remove(filename)
}
}()
currentUkey := make([]byte, 0, 4096)
hasCurrentUkey := false
lastSeqNumForKey := internalKeySeqNumMax
smallest, largest := internalKey(nil), internalKey(nil)
for iter.Next() {
// TODO: prioritize compacting d.imm.
// TODO: support c.shouldStopBefore.
ikey := internalKey(iter.Key())
if !ikey.valid() {
// Do not hide invalid keys.
currentUkey = currentUkey[:0]
hasCurrentUkey = false
lastSeqNumForKey = internalKeySeqNumMax
} else {
ukey := ikey.ukey()
if !hasCurrentUkey || d.icmp.userCmp.Compare(currentUkey, ukey) != 0 {
// This is the first occurrence of this user key.
currentUkey = append(currentUkey[:0], ukey...)
hasCurrentUkey = true
lastSeqNumForKey = internalKeySeqNumMax
}
drop, ikeySeqNum := false, ikey.seqNum()
if lastSeqNumForKey <= smallestSnapshot {
drop = true // Rule (A) referenced below.
} else if ikey.kind() == internalKeyKindDelete &&
ikeySeqNum <= smallestSnapshot &&
c.isBaseLevelForUkey(d.icmp.userCmp, ukey) {
// For this user key:
// (1) there is no data in higher levels
// (2) data in lower levels will have larger sequence numbers
// (3) data in layers that are being compacted here and have
// smaller sequence numbers will be dropped in the next
// few iterations of this loop (by rule (A) above).
// Therefore this deletion marker is obsolete and can be dropped.
drop = true
}
lastSeqNumForKey = ikeySeqNum
if drop {
continue
}
}
if tw == nil {
d.mu.Lock()
fileNum = d.versions.nextFileNum()
d.pendingOutputs[fileNum] = struct{}{}
pendingOutputs = append(pendingOutputs, fileNum)
d.mu.Unlock()
filename = dbFilename(d.dirname, fileTypeTable, fileNum)
file, err := d.opts.GetFileSystem().Create(filename)
if err != nil {
return nil, pendingOutputs, err
}
tw = table.NewWriter(file, &d.icmpOpts)
smallest = make(internalKey, len(ikey))
copy(smallest, ikey)
largest = make(internalKey, 0, 2*len(ikey))
}
largest = append(largest[:0], ikey...)
if err := tw.Set(ikey, iter.Value(), nil); err != nil {
return nil, pendingOutputs, err
}
}
ve = &versionEdit{
deletedFiles: map[deletedFileEntry]bool{},
newFiles: []newFileEntry{
{
level: c.level + 1,
meta: fileMetadata{
fileNum: fileNum,
size: 1,
smallest: smallest,
largest: largest,
},
},
},
}
for i := 0; i < 2; i++ {
for _, f := range c.inputs[i] {
ve.deletedFiles[deletedFileEntry{
level: c.level + i,
fileNum: f.fileNum,
}] = true
}
}
return ve, pendingOutputs, nil
}