419 lines
9.2 KiB
Go
419 lines
9.2 KiB
Go
package storage
|
||
|
||
import (
|
||
"bytes"
|
||
"errors"
|
||
"fmt"
|
||
"os"
|
||
"path/filepath"
|
||
"sync"
|
||
|
||
"gordenko.dev/dima/qb"
|
||
"gordenko.dev/dima/qb/freelist"
|
||
"gordenko.dev/dima/qb/inbox"
|
||
)
|
||
|
||
const (
|
||
lsnSize = 4
|
||
packetPrefixSize = 12 // 4 lsn + 4 packet length + 4 crc32
|
||
|
||
lengthIdx = 0
|
||
checksumIdx = 4
|
||
lsnIdx = 8
|
||
|
||
filePerm = 0770
|
||
|
||
dumpSnapshotAfterNBytes = 1024 * 1024 * 1024 // 1 GB
|
||
|
||
writeBufferSize = 4 * 1024 * 1024
|
||
)
|
||
|
||
const (
|
||
CodeMetricAdd byte = 1
|
||
CodeMetricDelete byte = 2
|
||
CodeMeasuresAppendWithGrow byte = 5
|
||
CodeMeasuresAppend byte = 6
|
||
CodeMeasuresDelete byte = 7
|
||
)
|
||
|
||
type FreeList interface {
|
||
// використовується в allocPage
|
||
GetPageNumber() (uint32, error)
|
||
}
|
||
|
||
type MetricsState struct {
|
||
//Metrics []Metric
|
||
WaitCh chan struct{}
|
||
}
|
||
|
||
type Changes struct {
|
||
Commits []any
|
||
// for snapshot only
|
||
SnapshotNumberCh chan int // log number
|
||
FrozenIndexPagesCount int
|
||
IndexPageNumbers []uint32
|
||
FrozenDataPagesCount int
|
||
DataPageNumbers []uint32
|
||
}
|
||
|
||
type Writer struct {
|
||
mutex sync.Mutex
|
||
inbox *inbox.Inbox
|
||
workerInbox *inbox.Inbox
|
||
dataPagesCount uint32
|
||
indexPagesCount uint32
|
||
dataFreeList *freelist.FreeList
|
||
indexFreeList *freelist.FreeList
|
||
//atree *atree.Atree
|
||
dir string
|
||
databaseName string
|
||
w *bytes.Buffer // wal буфер для упаковки даних перед записом на диск
|
||
wal *os.File
|
||
dataFile *os.File
|
||
indexFile *os.File
|
||
input []any
|
||
writePreparer *WritePreparer
|
||
written int64
|
||
isExited bool
|
||
exitCh chan struct{}
|
||
waitGroup *sync.WaitGroup
|
||
}
|
||
|
||
type WriterOptions struct {
|
||
Inbox *inbox.Inbox
|
||
WorkerInbox *inbox.Inbox
|
||
Dir string
|
||
DatabaseName string
|
||
WAL string
|
||
DataFreeList *freelist.FreeList
|
||
IndexFreeList *freelist.FreeList
|
||
DataFile *os.File
|
||
IndexFile *os.File
|
||
ExitCh chan struct{}
|
||
WaitGroup *sync.WaitGroup
|
||
}
|
||
|
||
func NewWriter(opt WriterOptions) (*Writer, error) {
|
||
// if (opt.IndexPageSize % 2) != 0 {
|
||
// return nil, errors.New("IndexPageSize must be multiple of 2")
|
||
// }
|
||
// if (opt.DataPageSize % 2) != 0 {
|
||
// return nil, errors.New("DataPageSize must be multiple of 2")
|
||
// }
|
||
if opt.Inbox == nil {
|
||
return nil, errors.New("Inbox option is required")
|
||
}
|
||
if opt.WorkerInbox == nil {
|
||
return nil, errors.New("WorkerInbox 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.AppendToWorkerQueue == nil {
|
||
// return nil, errors.New("AppendToWorkerQueue 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")
|
||
}
|
||
if opt.DataFile == nil {
|
||
return nil, errors.New("DataFile option is required")
|
||
}
|
||
if opt.IndexFile == nil {
|
||
return nil, errors.New("IndexFile 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")
|
||
}
|
||
|
||
indexPageNoProvider, err := NewPageNoProvider(PageNoProviderOptions{
|
||
FilePath: qb.GetIndexFilePath(opt.Dir, opt.DatabaseName),
|
||
PageSize: IndexPageSize,
|
||
FreeList: opt.IndexFreeList,
|
||
})
|
||
|
||
dataPageNoProvider, err := NewPageNoProvider(PageNoProviderOptions{
|
||
FilePath: qb.GetDataFilePath(opt.Dir, opt.DatabaseName),
|
||
PageSize: DataPageSize,
|
||
FreeList: opt.DataFreeList,
|
||
})
|
||
|
||
pageIngester, err := NewPageIngester(PageIngesterOptions{
|
||
GetIndexPageNumber: indexPageNoProvider.GetPageNumber,
|
||
GetDataPageNumber: dataPageNoProvider.GetPageNumber,
|
||
})
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
|
||
writePreparer := NewWritePreparer(WritePreparerOptions{
|
||
//MinWALBufferSize: 128,
|
||
PageIngester: pageIngester,
|
||
})
|
||
|
||
s := &Writer{
|
||
inbox: opt.Inbox,
|
||
workerInbox: opt.WorkerInbox,
|
||
dir: opt.Dir,
|
||
databaseName: opt.DatabaseName,
|
||
dataFile: opt.DataFile,
|
||
indexFile: opt.IndexFile,
|
||
dataFreeList: opt.DataFreeList,
|
||
indexFreeList: opt.IndexFreeList,
|
||
writePreparer: writePreparer,
|
||
exitCh: opt.ExitCh,
|
||
waitGroup: opt.WaitGroup,
|
||
}
|
||
|
||
if opt.WAL != "" {
|
||
s.wal, err = os.OpenFile(filepath.Join(opt.Dir, opt.WAL), os.O_APPEND|os.O_WRONLY, filePerm)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
} else {
|
||
s.wal, err = os.OpenFile(
|
||
qb.GetWALFilePath(opt.Dir, opt.DatabaseName, 0), // .wal_0
|
||
os.O_CREATE|os.O_APPEND|os.O_WRONLY,
|
||
filePerm,
|
||
)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
}
|
||
|
||
return s, nil
|
||
}
|
||
|
||
func (s *Writer) Run() {
|
||
for {
|
||
select {
|
||
case <-s.inbox.Ready():
|
||
if err := s.packAndWrite(); err != nil {
|
||
qb.Abort(qb.FailedWriteToTxLog, err)
|
||
}
|
||
|
||
case <-s.exitCh:
|
||
s.exit()
|
||
return
|
||
}
|
||
}
|
||
}
|
||
|
||
func (s *Writer) packAndWrite() (err error) {
|
||
//fmt.Println("packAndWrite")
|
||
//s.mutex.Lock()
|
||
//isExited := s.isExited
|
||
|
||
// var exitWaitGroup *sync.WaitGroup
|
||
// if s.isExited {
|
||
// exitWaitGroup = s.waitGroup
|
||
// }
|
||
//s.mutex.Unlock()
|
||
|
||
input := s.inbox.Drain()
|
||
|
||
//fmt.Printf("Drain: %#v\n", input)
|
||
|
||
prepared := s.writePreparer.Prepare(input)
|
||
|
||
//fmt.Printf("prepared: %#v\n", prepared)
|
||
fmt.Println("prepared")
|
||
|
||
// 3. Пишу на диск WAL (append)
|
||
n, err := s.wal.Write(prepared.Packet)
|
||
if err != nil {
|
||
return
|
||
}
|
||
//fmt.Printf("written to wal: % x\n", prepared.Packet)
|
||
if n != len(prepared.Packet) {
|
||
return fmt.Errorf("written %d != packet size %d", n, len(prepared.Packet))
|
||
}
|
||
if err = s.wal.Sync(); err != nil {
|
||
return
|
||
}
|
||
fmt.Println("synced to wal")
|
||
|
||
// 4. Пишу в atree сторінки
|
||
if len(prepared.WriteToData) > 0 {
|
||
err = WriteDataPages(s.dataFile, prepared.WriteToData)
|
||
if err != nil {
|
||
return
|
||
}
|
||
if err = s.dataFile.Sync(); err != nil {
|
||
return
|
||
}
|
||
fmt.Println("written to data")
|
||
}
|
||
if len(prepared.WriteToIndex) > 0 {
|
||
err = WriteIndexPages(s.indexFile, prepared.WriteToIndex)
|
||
if err != nil {
|
||
return
|
||
}
|
||
if err = s.indexFile.Sync(); err != nil {
|
||
return
|
||
}
|
||
fmt.Println("written to index")
|
||
}
|
||
// 6. відправляю input - worker-у
|
||
|
||
var (
|
||
forceSnapshot bool
|
||
)
|
||
|
||
if s.written > dumpSnapshotAfterNBytes {
|
||
forceSnapshot = true
|
||
}
|
||
|
||
// if isExited && s.written > 0 {
|
||
// forceSnapshot = true
|
||
// }
|
||
|
||
if forceSnapshot {
|
||
if err := s.wal.Close(); err != nil {
|
||
return fmt.Errorf("close wal file: %s", err)
|
||
}
|
||
s.wal = nil
|
||
s.written = 0
|
||
|
||
snapshotNumberCh := make(chan int, 1)
|
||
|
||
s.workerInbox.Push(Changes{
|
||
Commits: prepared.Commits,
|
||
SnapshotNumberCh: snapshotNumberCh,
|
||
FrozenIndexPagesCount: s.indexFreeList.Pages(),
|
||
IndexPageNumbers: s.indexFreeList.Cached(),
|
||
FrozenDataPagesCount: s.dataFreeList.Pages(),
|
||
DataPageNumbers: s.dataFreeList.Cached(),
|
||
})
|
||
|
||
snapshotNumber := <-snapshotNumberCh
|
||
|
||
// копіюю сторінки із delta файла в base файл і потім роблю Truncate
|
||
err = s.indexFreeList.Merge()
|
||
if err != nil {
|
||
return
|
||
}
|
||
err = s.dataFreeList.Merge()
|
||
if err != nil {
|
||
return
|
||
}
|
||
|
||
var err error
|
||
s.wal, err = os.OpenFile(
|
||
qb.GetWALFilePath(s.dir, s.databaseName, snapshotNumber),
|
||
os.O_CREATE|os.O_WRONLY,
|
||
filePerm,
|
||
)
|
||
if err != nil {
|
||
return fmt.Errorf("create new changes file: %s", err)
|
||
}
|
||
} else {
|
||
s.workerInbox.Push(Changes{
|
||
Commits: prepared.Commits,
|
||
})
|
||
fmt.Println("pushed commits")
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// func (s *Writer) reset() {
|
||
// s.buf.Reset()
|
||
// s.buf.Write([]byte{
|
||
// 0, 0, 0, 0, // packet length
|
||
// 0, 0, 0, 0, // crc32
|
||
// 0, 0, 0, 0, // lsn
|
||
|
||
// })
|
||
// s.pagesToWrite = nil
|
||
// s.workerReqs = nil
|
||
// s.waitCh = make(chan struct{})
|
||
// }
|
||
|
||
// func (s *Writer) flush(w *bytes.Buffer) error {
|
||
|
||
// return nil
|
||
// }
|
||
|
||
func (s *Writer) exit() {
|
||
s.mutex.Lock()
|
||
s.isExited = true
|
||
s.mutex.Unlock()
|
||
|
||
if err := s.packAndWrite(); err != nil {
|
||
qb.Abort(qb.FailedWriteToTxLog, err)
|
||
}
|
||
}
|
||
|
||
// type IndexPayloadToWrite struct {
|
||
// PageNo uint32
|
||
// Payload []byte
|
||
// ZeroLevel bool
|
||
// }
|
||
|
||
type PageToWrite struct {
|
||
PageNo uint32
|
||
Content []byte
|
||
}
|
||
|
||
func WriteDataPages(file *os.File, pages []PageToWrite) (err error) {
|
||
for _, p := range pages {
|
||
if len(p.Content) != DataPageSize {
|
||
return fmt.Errorf("wrong data page size: %d", len(p.Content))
|
||
}
|
||
var (
|
||
off = int(p.PageNo-1) * DataPageSize
|
||
n int
|
||
)
|
||
n, err = file.WriteAt(p.Content, int64(off))
|
||
if err != nil {
|
||
return fmt.Errorf("file.WriteAt: %s; off=%d", err, off)
|
||
}
|
||
if n != DataPageSize {
|
||
return fmt.Errorf("write %d instead of %d", n, DataPageSize)
|
||
}
|
||
}
|
||
return
|
||
}
|
||
|
||
func WriteIndexPages(file *os.File, pages []PageToWrite) (err error) {
|
||
for _, p := range pages {
|
||
if len(p.Content) != IndexPageSize {
|
||
return fmt.Errorf("wrong index page size: %d", len(p.Content))
|
||
}
|
||
var (
|
||
off = int(p.PageNo-1) * IndexPageSize
|
||
n int
|
||
)
|
||
n, err = file.WriteAt(p.Content, int64(off))
|
||
if err != nil {
|
||
return
|
||
}
|
||
if n != IndexPageSize {
|
||
return fmt.Errorf("write %d instead of %d", n, IndexPageSize)
|
||
}
|
||
}
|
||
return
|
||
}
|
||
|
||
// API
|
||
|
||
// FIX - add
|
||
// type DeletedMeasuresSince struct {
|
||
// MetricID uint32
|
||
// LastPageNo uint32
|
||
// IsRootChanged bool
|
||
// RootPageNo uint32
|
||
// FreeDataPages []uint32
|
||
// FreeIndexPages []uint32
|
||
// TimestampsBuf []byte
|
||
// ValuesBuf []byte
|
||
// }
|