// cloneFileSystem returns a new memory-backed file system whose root contains
// a copy of the directory dirname in the source file system srcFS. The copy
// is not recursive; directories under dirname are not copied.
//
// Changes to the resultant file system do not modify the source file system.
//
// For example, if srcFS contained:
// - /bar
// - /baz/0
// - /foo/x
// - /foo/y
// - /foo/z/A
// - /foo/z/B
// then calling cloneFileSystem(srcFS, "/foo") would result in a file system
// containing:
// - /x
// - /y
func cloneFileSystem(srcFS db.FileSystem, dirname string) (db.FileSystem, error) {
if len(dirname) == 0 || dirname[len(dirname)-1] != os.PathSeparator {
dirname += string(os.PathSeparator)
}
dstFS := memfs.New()
list, err := srcFS.List(dirname)
if err != nil {
return nil, err
}
for _, name := range list {
srcFile, err := srcFS.Open(dirname + name)
if err != nil {
return nil, err
}
stat, err := srcFile.Stat()
if err != nil {
return nil, err
}
if stat.IsDir() {
err = srcFile.Close()
if err != nil {
return nil, err
}
continue
}
data := make([]byte, stat.Size())
_, err = io.ReadFull(srcFile, data)
if err != nil {
return nil, err
}
err = srcFile.Close()
if err != nil {
return nil, err
}
dstFile, err := dstFS.Create(name)
if err != nil {
return nil, err
}
_, err = dstFile.Write(data)
if err != nil {
return nil, err
}
err = dstFile.Close()
if err != nil {
return nil, err
}
}
return dstFS, 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
}
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
}
// replayLogFile replays the edits in the named log file.
//
// 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) replayLogFile(ve *versionEdit, fs db.FileSystem, filename string) (maxSeqNum uint64, err error) {
file, err := fs.Open(filename)
if err != nil {
return 0, err
}
defer file.Close()
var (
mem *memdb.MemDB
batchBuf = new(bytes.Buffer)
ikey = make(internalKey, 512)
rr = record.NewReader(file)
)
for {
r, err := rr.Next()
if err == io.EOF {
break
}
if err != nil {
return 0, err
}
_, err = io.Copy(batchBuf, r)
if err != nil {
return 0, err
}
if batchBuf.Len() < batchHeaderLen {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
b := Batch{batchBuf.Bytes()}
seqNum := b.seqNum()
seqNum1 := seqNum + uint64(b.count())
if maxSeqNum < seqNum1 {
maxSeqNum = seqNum1
}
if mem == nil {
mem = memdb.New(&d.icmpOpts)
}
t := b.iter()
for ; seqNum != seqNum1; seqNum++ {
kind, ukey, value, ok := t.next()
if !ok {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
// Convert seqNum, kind and key into an internalKey, and add that ikey/value
// pair to mem.
//
// TODO: instead of copying to an intermediate buffer (ikey), is it worth
// adding a SetTwoPartKey(db.TwoPartKey{key0, key1}, value, opts) method to
// memdb.MemDB? What effect does that have on the db.Comparer interface?
//
// The C++ LevelDB code does not need an intermediate copy because its memdb
// implementation is a private implementation detail, and copies each internal
// key component from the Batch format straight to the skiplist buffer.
//
// Go's LevelDB considers the memdb functionality to be useful in its own
// right, and so leveldb/memdb is a separate package that is usable without
// having to import the top-level leveldb package. That extra abstraction
// means that we need to copy to an intermediate buffer here, to reconstruct
// the complete internal key to pass to the memdb.
ikey = makeInternalKey(ikey, ukey, kind, seqNum)
mem.Set(ikey, value, nil)
}
if len(t) != 0 {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
// TODO: if mem is large enough, write it to a level-0 table and set mem = nil.
batchBuf.Reset()
}
if mem != nil && !mem.Empty() {
// 如果初始化后memtable中有数据，则存入磁盘level0层级文件中
meta, err := d.writeLevel0Table(fs, mem)
if err != nil {
return 0, err
}
ve.newFiles = append(ve.newFiles, newFileEntry{level: 0, meta: meta})
// Strictly speaking, it's too early to delete meta.fileNum from d.pendingOutputs,
// but we are replaying the log file, which happens before Open returns, so there
// is no possibility of deleteObsoleteFiles being called concurrently here.
delete(d.pendingOutputs, meta.fileNum)
}
return maxSeqNum, 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
}