499 lines
12 KiB
Go
499 lines
12 KiB
Go
package atree
|
||
|
||
import (
|
||
"errors"
|
||
"fmt"
|
||
"os"
|
||
"path/filepath"
|
||
"sync"
|
||
|
||
diploma "gordenko.dev/dima/qb"
|
||
"gordenko.dev/dima/qb/atree/redo"
|
||
"gordenko.dev/dima/qb/bin"
|
||
)
|
||
|
||
const (
|
||
filePerm = 0770
|
||
|
||
// index page
|
||
indexRecordsQtyIdx = IndexPageSize - 7
|
||
isDataPageNumbersIdx = IndexPageSize - 5
|
||
indexCRC32Idx = IndexPageSize - 4
|
||
|
||
// data page
|
||
timestampsSizeIdx = DataPageSize - 12
|
||
valuesSizeIdx = DataPageSize - 10
|
||
prevPageIdx = DataPageSize - 8
|
||
dataCRC32Idx = DataPageSize - 4
|
||
|
||
timestampSize = 4
|
||
pairSize = timestampSize + PageNoSize
|
||
indexFooterIdx = indexRecordsQtyIdx
|
||
dataFooterIdx = timestampsSizeIdx
|
||
|
||
DataPageSize = 8192
|
||
IndexPageSize = 1024
|
||
PageNoSize = 4
|
||
//
|
||
DataPagePayloadSize int = dataFooterIdx
|
||
)
|
||
|
||
type FreeList interface {
|
||
// використовується в allocPage
|
||
ReservePage() uint32
|
||
}
|
||
|
||
type _page struct {
|
||
PageNo uint32
|
||
Buf []byte
|
||
ReferenceCount int
|
||
}
|
||
|
||
type Atree struct {
|
||
redoDir string
|
||
indexFreelist FreeList
|
||
dataFreelist FreeList
|
||
dataFile *os.File
|
||
indexFile *os.File
|
||
mutex sync.Mutex
|
||
allocatedIndexPagesQty uint32
|
||
allocatedDataPagesQty uint32
|
||
indexPages map[uint32]*_page
|
||
dataPages map[uint32]*_page
|
||
indexWaits map[uint32][]chan readResult
|
||
dataWaits map[uint32][]chan readResult
|
||
indexPagesToRead []uint32
|
||
dataPagesToRead []uint32
|
||
readSignalCh chan struct{}
|
||
writeSignalCh chan struct{}
|
||
writeTasksQueue []WriteTask
|
||
}
|
||
|
||
type Options struct {
|
||
Dir string
|
||
RedoDir string
|
||
DatabaseName string
|
||
DataFreeList FreeList
|
||
IndexFreeList FreeList
|
||
}
|
||
|
||
func New(opt Options) (*Atree, error) {
|
||
if opt.Dir == "" {
|
||
return nil, errors.New("Dir option is required")
|
||
}
|
||
if opt.RedoDir == "" {
|
||
return nil, errors.New("RedoDir option is required")
|
||
}
|
||
if opt.DatabaseName == "" {
|
||
return nil, errors.New("DatabaseName option is required")
|
||
}
|
||
if opt.DataFreeList == nil {
|
||
return nil, errors.New("DataFreeList option is required")
|
||
}
|
||
if opt.IndexFreeList == nil {
|
||
return nil, errors.New("IndexFreeList option is required")
|
||
}
|
||
// открываю или создаю dbName.data и dbName.index файлы
|
||
var (
|
||
indexFileName = filepath.Join(opt.Dir, opt.DatabaseName+".index")
|
||
dataFileName = filepath.Join(opt.Dir, opt.DatabaseName+".data")
|
||
|
||
indexFile *os.File
|
||
dataFile *os.File
|
||
allocatedIndexPagesQty uint32
|
||
allocatedDataPagesQty uint32
|
||
)
|
||
|
||
// При создании data файла сразу создается индекс, поэтому корректное
|
||
// состояние БД: либо оба файла есть, либо ни одного файла нет.
|
||
isIndexExist, err := isFileExist(indexFileName)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("check index file is exist: %s", err)
|
||
}
|
||
|
||
isDataExist, err := isFileExist(dataFileName)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("check data file is exist: %s", err)
|
||
}
|
||
|
||
if isIndexExist {
|
||
if isDataExist {
|
||
// открываю оба файла
|
||
indexFile, allocatedIndexPagesQty, err = openFile(indexFileName, IndexPageSize)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("open index file: %s", err)
|
||
}
|
||
|
||
dataFile, allocatedDataPagesQty, err = openFile(dataFileName, DataPageSize)
|
||
if err != nil {
|
||
return nil, fmt.Errorf("open data file: %s", err)
|
||
}
|
||
} else {
|
||
// нет data файла
|
||
return nil, errors.New("not found data file")
|
||
}
|
||
} else {
|
||
if isDataExist {
|
||
// index файла нет
|
||
return nil, errors.New("not found index file")
|
||
} else {
|
||
// нет обоих файлов
|
||
indexFile, err = os.OpenFile(indexFileName, os.O_CREATE|os.O_RDWR, filePerm)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
dataFile, err = os.OpenFile(dataFileName, os.O_CREATE|os.O_RDWR, filePerm)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
}
|
||
}
|
||
|
||
tree := &Atree{
|
||
redoDir: opt.RedoDir,
|
||
indexFreelist: opt.IndexFreeList,
|
||
dataFreelist: opt.DataFreeList,
|
||
indexFile: indexFile,
|
||
dataFile: dataFile,
|
||
allocatedIndexPagesQty: allocatedIndexPagesQty,
|
||
allocatedDataPagesQty: allocatedDataPagesQty,
|
||
indexPages: make(map[uint32]*_page),
|
||
dataPages: make(map[uint32]*_page),
|
||
indexWaits: make(map[uint32][]chan readResult),
|
||
dataWaits: make(map[uint32][]chan readResult),
|
||
readSignalCh: make(chan struct{}, 1),
|
||
writeSignalCh: make(chan struct{}, 1),
|
||
}
|
||
|
||
return tree, nil
|
||
}
|
||
|
||
func (s *Atree) Run() {
|
||
go s.pageWriter()
|
||
go s.pageReader()
|
||
}
|
||
|
||
// FIND
|
||
|
||
func (s *Atree) findDataPage(rootPageNo uint32, timestamp uint32) (uint32, []byte, error) {
|
||
indexPageNo := rootPageNo
|
||
for {
|
||
buf, err := s.fetchIndexPage(indexPageNo)
|
||
if err != nil {
|
||
return 0, nil, fmt.Errorf("fetchIndexPage(%d): %s", indexPageNo, err)
|
||
}
|
||
|
||
foundPageNo := findPageNo(buf, timestamp)
|
||
s.releaseIndexPage(indexPageNo)
|
||
|
||
if buf[isDataPageNumbersIdx] == 1 {
|
||
buf, err := s.fetchDataPage(foundPageNo)
|
||
if err != nil {
|
||
return 0, nil, fmt.Errorf("fetchDataPage(%d): %s", foundPageNo, err)
|
||
}
|
||
return foundPageNo, buf, nil
|
||
}
|
||
// вглубь
|
||
indexPageNo = foundPageNo
|
||
}
|
||
}
|
||
|
||
type pathLeg struct {
|
||
PageNo uint32
|
||
Data []byte
|
||
}
|
||
|
||
type pathToDataPage struct {
|
||
Legs []pathLeg
|
||
LastPageNo uint32
|
||
}
|
||
|
||
func (s *Atree) findPathToLastPage(rootPageNo uint32) (_ pathToDataPage, err error) {
|
||
var (
|
||
pageNo = rootPageNo
|
||
legs []pathLeg
|
||
)
|
||
|
||
for {
|
||
var buf []byte
|
||
buf, err = s.fetchIndexPage(pageNo)
|
||
if err != nil {
|
||
err = fmt.Errorf("FetchIndexPage(%d): %s", pageNo, err)
|
||
return
|
||
}
|
||
|
||
legs = append(legs, pathLeg{
|
||
PageNo: pageNo,
|
||
Data: buf,
|
||
// childIdx не нужен
|
||
})
|
||
|
||
foundPageNo := getLastPageNo(buf)
|
||
|
||
if buf[isDataPageNumbersIdx] == 1 {
|
||
return pathToDataPage{
|
||
Legs: legs,
|
||
LastPageNo: foundPageNo,
|
||
}, nil
|
||
}
|
||
// вглубь
|
||
pageNo = foundPageNo
|
||
}
|
||
}
|
||
|
||
// APPEND DATA PAGE
|
||
|
||
type AppendDataPageReq struct {
|
||
MetricID uint32
|
||
Timestamp uint32
|
||
Value float64
|
||
Since uint32
|
||
RootPageNo uint32
|
||
PrevPageNo uint32
|
||
TimestampsChunks [][]byte
|
||
TimestampsSize uint16
|
||
ValuesChunks [][]byte
|
||
ValuesSize uint16
|
||
}
|
||
|
||
func (s *Atree) AppendDataPage(req AppendDataPageReq) (_ redo.Report, err error) {
|
||
var (
|
||
flags byte
|
||
dataPagesToRelease []uint32
|
||
indexPagesToRelease []uint32
|
||
)
|
||
|
||
newDataPage := s.allocDataPage()
|
||
dataPagesToRelease = append(dataPagesToRelease, newDataPage.PageNo)
|
||
|
||
chunksToDataPage(newDataPage.Data, chunksToDataPageReq{
|
||
PrevPageNo: req.PrevPageNo,
|
||
TimestampsChunks: req.TimestampsChunks,
|
||
TimestampsSize: req.TimestampsSize,
|
||
ValuesChunks: req.ValuesChunks,
|
||
ValuesSize: req.ValuesSize,
|
||
})
|
||
|
||
redoWriter, err := redo.NewWriter(redo.WriterOptions{
|
||
Dir: s.redoDir,
|
||
MetricID: req.MetricID,
|
||
Timestamp: req.Timestamp,
|
||
Value: req.Value,
|
||
IsDataPageReused: newDataPage.IsReused,
|
||
DataPageNo: newDataPage.PageNo,
|
||
Page: newDataPage.Data,
|
||
})
|
||
if err != nil {
|
||
return
|
||
}
|
||
|
||
if req.RootPageNo > 0 {
|
||
var path pathToDataPage
|
||
path, err = s.findPathToLastPage(req.RootPageNo)
|
||
if err != nil {
|
||
return
|
||
}
|
||
for _, leg := range path.Legs {
|
||
indexPagesToRelease = append(indexPagesToRelease, leg.PageNo)
|
||
}
|
||
|
||
if path.LastPageNo != req.PrevPageNo {
|
||
diploma.Abort(
|
||
diploma.WrongPrevPageNo,
|
||
fmt.Errorf("bug: last pageNo %d in tree != prev pageNo %d in _metric",
|
||
path.LastPageNo, req.PrevPageNo),
|
||
)
|
||
}
|
||
|
||
newPageNo := newDataPage.PageNo
|
||
lastIdx := len(path.Legs) - 1
|
||
|
||
for legIdx := lastIdx; legIdx >= 0; legIdx-- {
|
||
leg := path.Legs[legIdx]
|
||
|
||
ok := appendPair(leg.Data, req.Since, newPageNo)
|
||
if ok {
|
||
err = redoWriter.AppendIndexPage(leg.PageNo, leg.Data, 0)
|
||
if err != nil {
|
||
return
|
||
}
|
||
break
|
||
}
|
||
|
||
newIndexPage := s.allocIndexPage()
|
||
indexPagesToRelease = append(indexPagesToRelease, newIndexPage.PageNo)
|
||
appendPair(newIndexPage.Data, req.Since, newPageNo)
|
||
// ставлю мітку що всі pageNo на сторінці - це data pageNo
|
||
if legIdx == lastIdx {
|
||
newIndexPage.Data[isDataPageNumbersIdx] = 1
|
||
}
|
||
|
||
flags = 0
|
||
if newIndexPage.IsReused {
|
||
flags |= redo.FlagReused
|
||
}
|
||
err = redoWriter.AppendIndexPage(newIndexPage.PageNo, newIndexPage.Data, flags)
|
||
if err != nil {
|
||
return
|
||
}
|
||
//
|
||
newPageNo = newIndexPage.PageNo
|
||
|
||
if legIdx == 0 {
|
||
newRoot := s.allocIndexPage()
|
||
indexPagesToRelease = append(indexPagesToRelease, newRoot.PageNo)
|
||
appendPair(newRoot.Data, getSince(leg.Data), leg.PageNo) // old rootPageNo
|
||
appendPair(newRoot.Data, req.Since, newIndexPage.PageNo)
|
||
|
||
// Фиксирую новый root в REDO логе
|
||
flags = redo.FlagNewRoot
|
||
if newRoot.IsReused {
|
||
flags |= redo.FlagReused
|
||
}
|
||
err = redoWriter.AppendIndexPage(newRoot.PageNo, newRoot.Data, flags)
|
||
if err != nil {
|
||
return
|
||
}
|
||
break
|
||
}
|
||
}
|
||
} else {
|
||
newRoot := s.allocIndexPage()
|
||
indexPagesToRelease = append(indexPagesToRelease, newRoot.PageNo)
|
||
newRoot.Data[isDataPageNumbersIdx] = 1
|
||
appendPair(newRoot.Data, req.Since, newDataPage.PageNo)
|
||
|
||
flags = redo.FlagNewRoot
|
||
if newRoot.IsReused {
|
||
flags |= redo.FlagReused
|
||
}
|
||
err = redoWriter.AppendIndexPage(newRoot.PageNo, newRoot.Data, flags)
|
||
if err != nil {
|
||
return
|
||
}
|
||
}
|
||
|
||
err = redoWriter.Close()
|
||
if err != nil {
|
||
return
|
||
}
|
||
|
||
// На данний момен схема - наступна. Всі сторінки - data та index - зафіксовані в кеші.
|
||
// Отже запис на диск пройде максимально швидко. Після цього ReferenceCount кожної
|
||
// сторінки зменшиться на 1. Оскільки на метрику утримується XLock, сторінки мають
|
||
// ReferenceCount = 1 (немає інших читачів).
|
||
waitCh := make(chan struct{})
|
||
|
||
task := WriteTask{
|
||
WaitCh: waitCh,
|
||
DataPage: redo.PageToWrite{
|
||
PageNo: newDataPage.PageNo,
|
||
Data: newDataPage.Data,
|
||
},
|
||
IndexPages: redoWriter.IndexPagesToWrite(),
|
||
}
|
||
|
||
s.appendWriteTaskToQueue(task)
|
||
|
||
<-waitCh
|
||
|
||
for _, pageNo := range dataPagesToRelease {
|
||
s.releaseDataPage(pageNo)
|
||
}
|
||
for _, pageNo := range indexPagesToRelease {
|
||
s.releaseIndexPage(pageNo)
|
||
}
|
||
return redoWriter.GetReport(), nil
|
||
}
|
||
|
||
// DELETE
|
||
|
||
type PageLists struct {
|
||
DataPages []uint32
|
||
IndexPages []uint32
|
||
}
|
||
|
||
type Level struct {
|
||
PageNo uint32
|
||
PageData []byte
|
||
Idx int
|
||
ChildQty int
|
||
}
|
||
|
||
func (s *Atree) GetAllPages(rootPageNo uint32) (_ PageLists, err error) {
|
||
var (
|
||
dataPages []uint32
|
||
indexPages []uint32
|
||
levels []*Level
|
||
)
|
||
|
||
buf, err := s.fetchIndexPage(rootPageNo)
|
||
if err != nil {
|
||
err = fmt.Errorf("fetchIndexPage(%d): %s", rootPageNo, err)
|
||
return
|
||
}
|
||
indexPages = append(indexPages, rootPageNo)
|
||
|
||
if buf[isDataPageNumbersIdx] == 1 {
|
||
pageNumbers := listPageNumbers(buf)
|
||
dataPages = append(dataPages, pageNumbers...)
|
||
|
||
s.releaseIndexPage(rootPageNo)
|
||
|
||
return PageLists{
|
||
DataPages: dataPages,
|
||
IndexPages: indexPages,
|
||
}, nil
|
||
}
|
||
|
||
levels = append(levels, &Level{
|
||
PageNo: rootPageNo,
|
||
PageData: buf,
|
||
Idx: 0,
|
||
ChildQty: bin.GetUint16AsInt(buf[indexRecordsQtyIdx:]),
|
||
})
|
||
|
||
for {
|
||
if len(levels) == 0 {
|
||
return PageLists{
|
||
DataPages: dataPages,
|
||
IndexPages: indexPages,
|
||
}, nil
|
||
}
|
||
|
||
lastIdx := len(levels) - 1
|
||
level := levels[lastIdx]
|
||
|
||
if level.Idx < level.ChildQty {
|
||
pageNo := getPageNo(level.PageData, level.Idx)
|
||
level.Idx++
|
||
|
||
var buf []byte
|
||
buf, err = s.fetchIndexPage(pageNo)
|
||
if err != nil {
|
||
err = fmt.Errorf("fetchIndexPage(%d): %s", pageNo, err)
|
||
return
|
||
}
|
||
indexPages = append(indexPages, pageNo)
|
||
|
||
if buf[isDataPageNumbersIdx] == 1 {
|
||
pageNumbers := listPageNumbers(buf)
|
||
dataPages = append(dataPages, pageNumbers...)
|
||
|
||
s.releaseIndexPage(pageNo)
|
||
} else {
|
||
levels = append(levels, &Level{
|
||
PageNo: pageNo,
|
||
PageData: buf,
|
||
Idx: 0,
|
||
ChildQty: bin.GetUint16AsInt(buf[indexRecordsQtyIdx:]),
|
||
})
|
||
}
|
||
} else {
|
||
s.releaseIndexPage(level.PageNo)
|
||
levels = levels[:lastIdx]
|
||
}
|
||
}
|
||
}
|