// internalGet looks up the first key/value pair whose (internal) key is >=
// ikey, according to the internal key ordering, and also returns whether or
// not that search was conclusive.
//
// If there is no such pair, or that pair's key and ikey do not share the same
// user key (according to ucmp), then conclusive will be false. Otherwise,
// conclusive will be true and:
// * if that pair's key's kind is set, that pair's value will be returned,
// * if that pair's key's kind is delete, db.ErrNotFound will be returned.
// If the returned error is non-nil then conclusive will be true.
func internalGet(t db.Iterator, ucmp db.Comparer, ukey []byte) (value []byte, conclusive bool, err error) {
if !t.Next() {
err = t.Close()
return nil, err != nil, err
}
ikey0 := internalKey(t.Key())
if !ikey0.valid() {
t.Close()
return nil, true, fmt.Errorf("leveldb: corrupt table: invalid internal key")
}
if ucmp.Compare(ukey, ikey0.ukey()) != 0 {
err = t.Close()
return nil, err != nil, err
}
if ikey0.kind() == internalKeyKindDelete {
t.Close()
return nil, true, db.ErrNotFound
}
return t.Value(), true, t.Close()
}
// internalGet looks up the first key/value pair whose (internal) key is >=
// ikey, according to the internal key ordering, and also returns whether or
// not that search was conclusive.
//
// If there is no such pair, or that pair's key and ikey do not share the same
// user key (according to ucmp), then conclusive will be false. Otherwise,
// conclusive will be true and:
// * if that pair's key's kind is set, that pair's value will be returned,
// * if that pair's key's kind is delete, db.ErrNotFound will be returned.
// If the returned error is non-nil then conclusive will be true.
func internalGet(t db.Iterator, ucmp db.Comparer, ukey []byte) (value []byte, conclusive bool, err error) {
if !t.Next() {
err = t.Close()
return nil, err != nil, err
}
ikey0 := internalKey(t.Key())
if !ikey0.valid() {
t.Close()
return nil, true, fmt.Errorf("leveldb: corrupt table: invalid internal key")
}
if ucmp.Compare(ukey, ikey0.ukey()) != 0 {
err = t.Close()
// 在memtable中没有找到，因为用户key不匹配
// 之后再从磁盘db文件中查找
return nil, err != nil, err
}
if ikey0.kind() == internalKeyKindDelete {
t.Close()
// 该key对应的数据项已经被删除了，所以也没找到
return nil, true, db.ErrNotFound
}
// 找到，返回value
return t.Value(), true, t.Close()
}
// 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
}
// 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
}