389 lines
8.9 KiB
Go
389 lines
8.9 KiB
Go
package database
|
|
|
|
import (
|
|
"errors"
|
|
"fmt"
|
|
"log"
|
|
"net"
|
|
"os"
|
|
"path/filepath"
|
|
"sync"
|
|
"time"
|
|
|
|
"gordenko.dev/dima/qb/atree"
|
|
"gordenko.dev/dima/qb/freelist"
|
|
"gordenko.dev/dima/qb/storage"
|
|
)
|
|
|
|
func JoinSnapshotFileName(dir string, snapshotNumber int) string {
|
|
return filepath.Join(dir, fmt.Sprintf("%d.snapshot", snapshotNumber))
|
|
}
|
|
|
|
func JoinWALFileName(dir string, snapshotNumber int) string {
|
|
return filepath.Join(dir, fmt.Sprintf("%d.wal", snapshotNumber))
|
|
}
|
|
|
|
// type metricLockEntry struct {
|
|
// XLock bool
|
|
// RLocks int
|
|
// WaitQueue []any
|
|
// }
|
|
|
|
type Database struct {
|
|
mutex sync.Mutex
|
|
workerSignalCh chan struct{}
|
|
workerQueue []any
|
|
rLocksToRelease []uint32
|
|
metrics map[uint32]*_metric
|
|
//metricLockEntries map[uint32]*metricLockEntry
|
|
dataFreeList *freelist.FreeList
|
|
indexFreeList *freelist.FreeList
|
|
dir string
|
|
databaseName string
|
|
snapshotNumber int
|
|
storage *storage.Writer
|
|
atree *atree.Atree
|
|
tcpPort int
|
|
logfile *os.File
|
|
logger *log.Logger
|
|
exitCh chan struct{}
|
|
waitGroup *sync.WaitGroup
|
|
}
|
|
|
|
type Options struct {
|
|
TCPPort int
|
|
Dir string
|
|
DatabaseName string
|
|
RedoDir string
|
|
Logfile *os.File
|
|
ExitCh chan struct{}
|
|
WaitGroup *sync.WaitGroup
|
|
}
|
|
|
|
func New(opt Options) (_ *Database, err error) {
|
|
if opt.TCPPort <= 0 {
|
|
return nil, errors.New("TCPPort option is required")
|
|
}
|
|
if opt.Dir == "" {
|
|
return nil, errors.New("Dir option is required")
|
|
}
|
|
if opt.DatabaseName == "" {
|
|
return nil, errors.New("DatabaseName option is required")
|
|
}
|
|
if opt.RedoDir == "" {
|
|
return nil, errors.New("RedoDir option is required")
|
|
}
|
|
if opt.Logfile == nil {
|
|
return nil, errors.New("Logfile option is required")
|
|
}
|
|
if opt.ExitCh == nil {
|
|
return nil, errors.New("ExitCh option is required")
|
|
}
|
|
if opt.WaitGroup == nil {
|
|
return nil, errors.New("WaitGroup option is required")
|
|
}
|
|
|
|
// deltaFreeListFile, err := os.OpenFile(
|
|
// filepath.Join(opt.Dir, opt.DatabaseName+".deltafree"),
|
|
// os.O_RDWR|os.O_CREATE,
|
|
// 0666,
|
|
// )
|
|
// if err != nil {
|
|
// return nil, err
|
|
// }
|
|
|
|
dataFreeList, err := freelist.New(freelist.Options{
|
|
PageSize: 2048,
|
|
BaseFilePath: filepath.Join(opt.Dir, opt.DatabaseName+".free_data_base"),
|
|
DeltaFilePath: filepath.Join(opt.Dir, opt.DatabaseName+".free_data_delta"),
|
|
})
|
|
|
|
indexFreeList, err := freelist.New(freelist.Options{
|
|
PageSize: 2048,
|
|
BaseFilePath: filepath.Join(opt.Dir, opt.DatabaseName+".free_index_base"),
|
|
DeltaFilePath: filepath.Join(opt.Dir, opt.DatabaseName+".free_index_delta"),
|
|
})
|
|
|
|
s := &Database{
|
|
workerSignalCh: make(chan struct{}, 1),
|
|
dir: opt.Dir,
|
|
databaseName: opt.DatabaseName,
|
|
metrics: make(map[uint32]*_metric),
|
|
//metricLockEntries: make(map[uint32]*metricLockEntry),
|
|
dataFreeList: dataFreeList,
|
|
indexFreeList: indexFreeList,
|
|
tcpPort: opt.TCPPort,
|
|
logfile: opt.Logfile,
|
|
logger: log.New(opt.Logfile, "", log.LstdFlags),
|
|
exitCh: opt.ExitCh,
|
|
waitGroup: opt.WaitGroup,
|
|
}
|
|
return s, nil
|
|
}
|
|
|
|
func (s *Database) ListenAndServe() (err error) {
|
|
listener, err := net.Listen("tcp", fmt.Sprintf(":%d", s.tcpPort))
|
|
if err != nil {
|
|
return fmt.Errorf("net.Listen: %s; port=%d", err, s.tcpPort)
|
|
}
|
|
|
|
s.atree, err = atree.New(atree.Options{
|
|
Dir: s.dir,
|
|
DatabaseName: s.databaseName,
|
|
})
|
|
if err != nil {
|
|
return fmt.Errorf("atree.New: %s", err)
|
|
}
|
|
s.atree.Run()
|
|
|
|
go s.worker()
|
|
|
|
s.recovery()
|
|
|
|
s.logger.Println("database started")
|
|
for {
|
|
// Listen for an incoming connection.
|
|
conn, err := listener.Accept()
|
|
if err != nil {
|
|
s.logger.Printf("listener.Accept: %s\n", err)
|
|
time.Sleep(time.Second)
|
|
} else {
|
|
go s.handleTCPConn(conn)
|
|
}
|
|
}
|
|
}
|
|
|
|
func (s *Database) recovery() {
|
|
// FIX
|
|
// advisor, err := recovery.NewRecoveryAdvisor(recovery.RecoveryAdvisorOptions{
|
|
// Dir: s.dir,
|
|
// VerifySnapshot: s.verifySnapshot,
|
|
// })
|
|
// if err != nil {
|
|
// panic(err)
|
|
// }
|
|
|
|
// recipe, err := advisor.GetRecipe()
|
|
// if err != nil {
|
|
// qb.Abort(qb.GetRecoveryRecipeFailed, err)
|
|
// }
|
|
|
|
// var logNumber int
|
|
|
|
// if recipe != nil {
|
|
// if recipe.Snapshot != "" {
|
|
// err = s.loadSnapshot(recipe.Snapshot)
|
|
// if err != nil {
|
|
// qb.Abort(qb.LoadSnapshotFailed, err)
|
|
// }
|
|
// }
|
|
// for _, changesFileName := range recipe.Changes {
|
|
// err = s.replayChanges(changesFileName)
|
|
// if err != nil {
|
|
// qb.Abort(qb.ReplayChangesFailed, err)
|
|
// }
|
|
// }
|
|
// logNumber = recipe.LogNumber
|
|
// }
|
|
|
|
// s.storage, err = storage.NewWriter(storage.WriterOptions{
|
|
// Dir: s.dir,
|
|
// LogNumber: logNumber,
|
|
// AppendToWorkerQueue: s.appendJobToWorkerQueue,
|
|
// FreeList: s.freeList,
|
|
// Atree: s.atree,
|
|
// ExitCh: s.exitCh,
|
|
// WaitGroup: s.waitGroup,
|
|
// })
|
|
// if err != nil {
|
|
// qb.Abort(qb.CreateChangesWriterFailed, err)
|
|
|
|
// }
|
|
// go s.storage.Run()
|
|
|
|
// // fileNames, err := s.searchREDOFiles()
|
|
// // if err != nil {
|
|
// // qb.Abort(qb.SearchREDOFilesFailed, err)
|
|
// // }
|
|
|
|
// // if len(fileNames) > 0 {
|
|
// // for _, fileName := range fileNames {
|
|
// // err = s.replayREDOFile(fileName)
|
|
// // if err != nil {
|
|
// // qb.Abort(qb.ReplayREDOFileFailed, err)
|
|
// // }
|
|
// // }
|
|
|
|
// // for _, fileName := range fileNames {
|
|
// // err = os.Remove(fileName)
|
|
// // if err != nil {
|
|
// // qb.Abort(qb.RemoveREDOFileFailed, err)
|
|
// // }
|
|
// // }
|
|
// // }
|
|
|
|
// if recipe != nil {
|
|
// if recipe.CompleteSnapshot {
|
|
// err = s.dumpSnapshot(logNumber)
|
|
// if err != nil {
|
|
// qb.Abort(qb.DumpSnapshotFailed, err)
|
|
// }
|
|
// }
|
|
|
|
// for _, fileName := range recipe.ToDelete {
|
|
// err = os.Remove(fileName)
|
|
// if err != nil {
|
|
// qb.Abort(qb.RemoveRecipeFileFailed, err)
|
|
// }
|
|
// }
|
|
// }
|
|
}
|
|
|
|
//
|
|
|
|
// зробити object?
|
|
|
|
func (s *Database) replayChanges(snapshotNumber int) (err error) {
|
|
snapshot, err := readSnapshot(JoinSnapshotFileName(s.dir, snapshotNumber))
|
|
if err != nil {
|
|
return
|
|
}
|
|
|
|
walReader, err := storage.NewWALReader(storage.WALReaderOptions{
|
|
FileName: JoinWALFileName(s.dir, snapshotNumber),
|
|
BufferSize: 8 * 1024 * 1024,
|
|
})
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
walReplayer, err := NewWALReplayer(WALReplayerOptions{
|
|
WALReader: walReader,
|
|
Metrics: snapshot.Metrics,
|
|
FreeIndexPages: snapshot.IndexPageNumbers,
|
|
FreeDataPages: snapshot.DataPageNumbers,
|
|
})
|
|
if err != nil {
|
|
return
|
|
}
|
|
|
|
err = walReplayer.Replay()
|
|
if err != nil {
|
|
return
|
|
}
|
|
|
|
err = s.storage.WritePagesToAtree(walReplayer.IndexPages(), walReplayer.DataPages())
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
// FIX free pages sync
|
|
|
|
return nil
|
|
}
|
|
|
|
func (s *Database) relayMetricsToMetrics(replayMetrics map[uint32]*ReplayMetric) {
|
|
for metricID, x := range replayMetrics {
|
|
s.metrics[metricID] = x.ToMetric()
|
|
}
|
|
}
|
|
|
|
// FIX
|
|
// УВАГА!
|
|
// Якщо буфер metric.buffer < розміру сторінки - для timestamps доступний весь буфер.
|
|
// Якщо буфер досяг розміру сторінки - в timestamps я передаю buffer[:DataPagePayloadSize]
|
|
|
|
var dataBufferSizes = []int{
|
|
1024,
|
|
2048,
|
|
4096,
|
|
8192,
|
|
}
|
|
|
|
func calculateDataBufferSize(payloadSize int) int {
|
|
for _, bufferSize := range dataBufferSizes {
|
|
if payloadSize <= bufferSize {
|
|
return bufferSize
|
|
}
|
|
}
|
|
return storage.DataPageSize
|
|
}
|
|
|
|
func allocateBuffers(requiredSpace int) (buf, databuf []byte) {
|
|
bufferSize := calculateDataBufferSize(requiredSpace)
|
|
buf = make([]byte, bufferSize)
|
|
if bufferSize == storage.DataPageSize {
|
|
databuf = buf[:storage.DataPagePayloadSize]
|
|
} else {
|
|
databuf = buf
|
|
}
|
|
return
|
|
}
|
|
|
|
// src - databuf
|
|
type growBuffersIn struct {
|
|
databuf []byte
|
|
tSize int
|
|
vSize int
|
|
requiredSpace int
|
|
}
|
|
|
|
func growBuffers(in growBuffersIn) (buf, databuf []byte) {
|
|
buf, databuf = allocateBuffers(in.requiredSpace)
|
|
copy(databuf[len(databuf)-in.tSize:], in.databuf[len(in.databuf)-in.tSize:])
|
|
copy(databuf, in.databuf[:in.vSize])
|
|
return
|
|
}
|
|
|
|
type shrinkBufferIn struct {
|
|
Buffer []byte
|
|
TimestampsSize int
|
|
ValuesSize int
|
|
NewBufferSize int
|
|
}
|
|
|
|
func shrinkBuffer(in shrinkBufferIn) []byte {
|
|
buf := make([]byte, in.NewBufferSize)
|
|
copy(buf, in.Buffer[:in.ValuesSize])
|
|
copy(buf[:len(buf)-in.TimestampsSize], in.Buffer[len(in.Buffer)-in.TimestampsSize:])
|
|
return buf
|
|
}
|
|
|
|
//func (s *Database) verifySnapshot(fileName string) (_ bool, err error) {
|
|
// file, err := os.Open(fileName)
|
|
// if err != nil {
|
|
// return
|
|
// }
|
|
// defer file.Close()
|
|
|
|
// stat, err := file.Stat()
|
|
// if err != nil {
|
|
// return
|
|
// }
|
|
|
|
// if stat.Size() <= 4 {
|
|
// return false, nil
|
|
// }
|
|
|
|
// var (
|
|
// payloadSize = stat.Size() - 4
|
|
// hash = crc32.NewIEEE()
|
|
// )
|
|
|
|
// _, err = io.CopyN(hash, file, payloadSize)
|
|
// if err != nil {
|
|
// return
|
|
// }
|
|
// calculatedCRC := hash.Sum32()
|
|
|
|
// storedCRC, err := bin.ReadUint32(file)
|
|
// if err != nil {
|
|
// return
|
|
// }
|
|
// if storedCRC != calculatedCRC {
|
|
// return false, fmt.Errorf("strored CRC %d not equal calculated CRC %d",
|
|
// storedCRC, calculatedCRC)
|
|
// }
|
|
// return true, nil
|
|
// }
|