ledisdb/rpl/file_table.go

package rpl

import (
	"bytes"
	"encoding/binary"
	"errors"
	"fmt"
	"github.com/edsrzf/mmap-go"
	"github.com/siddontang/go/log"
	"github.com/siddontang/go/num"
	"github.com/siddontang/go/sync2"
	"io"
	"os"
	"path"
	"sync"
	"time"
)

var (
	magic             = []byte("\x1c\x1d\xb8\x88\xff\x9e\x45\x55\x40\xf0\x4c\xda\xe0\xce\x47\xde\x65\x48\x71\x17")
	log0Data          = make([]byte, 17)
	errTableNeedFlush = errors.New("write table need flush")
	errTableFrozen    = errors.New("write table is frozen")
)

const tableReaderKeepaliveInterval int64 = 30

func fmtTableName(index int64) string {
	return fmt.Sprintf("%08d.ldb", index)
}

type tableReader struct {
	sync.Mutex

	name  string
	index int64

	f *os.File
	m mmap.MMap

	first uint64
	last  uint64

	lastTime uint32

	offsetStartPos int64
	offsetLen      uint32

	lastReadTime sync2.AtomicInt64
}

func newTableReader(base string, index int64) (*tableReader, error) {
	t := new(tableReader)
	t.name = path.Join(base, fmtTableName(index))
	t.index = index

	var err error

	if err = t.check(); err != nil {
		log.Error("check %s error: %s, try to repair", t.name, err.Error())

		if err = t.repair(); err != nil {
			log.Error("repair %s error: %s", t.name, err.Error())
			return nil, err
		}
	}

	t.close()

	return t, nil
}

func (t *tableReader) Close() {
	t.Lock()
	defer t.Unlock()

	t.close()
}

func (t *tableReader) close() {
	if t.m != nil {
		t.m.Unmap()
		t.m = nil
	}

	if t.f != nil {
		t.f.Close()
		t.f = nil
	}
}

func (t *tableReader) Keepalived() bool {
	l := t.lastReadTime.Get()
	if l > 0 && time.Now().Unix()-l > tableReaderKeepaliveInterval {
		return false
	}

	return true
}

func (t *tableReader) check() error {
	var err error

	if t.f, err = os.Open(t.name); err != nil {
		return err
	}

	st, _ := t.f.Stat()

	if st.Size() < 32 {
		return fmt.Errorf("file size %d too short", st.Size())
	}

	if _, err = t.f.Seek(-32, os.SEEK_END); err != nil {
		return err
	}

	if err = binary.Read(t.f, binary.BigEndian, &t.offsetStartPos); err != nil {
		return err
	} else if t.offsetStartPos >= st.Size() {
		return fmt.Errorf("invalid offset start pos %d, file size %d", t.offsetStartPos, st.Size())
	}

	if err = binary.Read(t.f, binary.BigEndian, &t.offsetLen); err != nil {
		return err
	} else if int64(t.offsetLen) >= st.Size() || t.offsetLen == 0 {
		return fmt.Errorf("invalid offset len %d, file size %d", t.offsetLen, st.Size())
	} else if t.offsetLen%4 != 0 {
		return fmt.Errorf("invalid offset len %d, must 4 multiple", t.offsetLen)
	}

	b := make([]byte, 20)
	if _, err = t.f.Read(b); err != nil {
		return err
	} else if !bytes.Equal(b, magic) {
		return fmt.Errorf("invalid magic data %q", b)
	}

	if t.m, err = mmap.MapRegion(t.f, int(t.offsetLen), mmap.RDONLY, 0, t.offsetStartPos); err != nil {
		return err
	}

	firstLogPos := binary.BigEndian.Uint32(t.m)
	lastLogPos := binary.BigEndian.Uint32(t.m[len(t.m)-4:])

	if firstLogPos != 0 {
		return fmt.Errorf("invalid first log pos %d, must 0", firstLogPos)
	} else if int64(lastLogPos) > t.offsetStartPos {
		return fmt.Errorf("invalid last log pos %d", lastLogPos)
	}

	var l Log
	if _, err = t.decodeLogHead(&l, int64(firstLogPos)); err != nil {
		return fmt.Errorf("decode first log err %s", err.Error())
	}

	t.first = l.ID
	var n int64
	if n, err = t.decodeLogHead(&l, int64(lastLogPos)); err != nil {
		return fmt.Errorf("decode last log err %s", err.Error())
	} else if n+int64(len(log0Data)) != t.offsetStartPos {
		return fmt.Errorf("invalid last log, no proper log0")
	}

	t.last = l.ID
	t.lastTime = l.CreateTime

	if t.first > t.last {
		return fmt.Errorf("invalid log table first %d > last %d", t.first, t.last)
	} else if (t.last - t.first + 1) != uint64(t.offsetLen/4) {
		return fmt.Errorf("invalid log table, first %d, last %d, and log num %d", t.first, t.last, t.offsetLen/4)
	}

	return nil
}

func (t *tableReader) repair() error {
	t.close()

	var err error
	if t.f, err = os.Open(t.name); err != nil {
		return err
	}

	defer t.close()

	tw := newTableWriter(path.Dir(t.name), t.index, maxLogFileSize)

	tmpName := tw.name + ".tmp"
	tw.name = tmpName
	os.Remove(tmpName)

	defer func() {
		tw.Close()
		os.Remove(tmpName)
	}()

	var l Log

	for {
		if err := l.Decode(t.f); err != nil {
			return err
		}

		if l.ID == 0 {
			break
		}

		t.lastTime = l.CreateTime

		if err := tw.StoreLog(&l); err != nil {
			return err
		}
	}

	t.close()

	var tr *tableReader
	if tr, err = tw.Flush(); err != nil {
		return err
	}

	t.first = tr.first
	t.last = tr.last
	t.offsetStartPos = tr.offsetStartPos
	t.offsetLen = tr.offsetLen

	defer tr.Close()

	os.Remove(t.name)

	if err := os.Rename(tmpName, t.name); err != nil {
		return err
	}

	return nil
}

func (t *tableReader) decodeLogHead(l *Log, pos int64) (int64, error) {
	_, err := t.f.Seek(int64(pos), os.SEEK_SET)
	if err != nil {
		return 0, err
	}

	dataLen, err := l.DecodeHead(t.f)
	if err != nil {
		return 0, err
	}

	return pos + int64(l.HeadSize()) + int64(dataLen), nil
}

func (t *tableReader) GetLog(id uint64, l *Log) error {
	if id < t.first || id > t.last {
		return fmt.Errorf("log %d not in [%d=%d]", id, t.first, t.last)
	}

	t.lastReadTime.Set(time.Now().Unix())

	t.Lock()
	defer t.Unlock()

	if err := t.openTable(); err != nil {
		t.close()
		return err
	}

	pos := binary.BigEndian.Uint32(t.m[(id-t.first)*4:])

	if _, err := t.f.Seek(int64(pos), os.SEEK_SET); err != nil {
		return err
	}

	if err := l.Decode(t.f); err != nil {
		return err
	} else if l.ID != id {
		return fmt.Errorf("invalid log id %d != %d", l.ID, id)
	}

	return nil
}

func (t *tableReader) openTable() error {
	var err error
	if t.f == nil {
		if t.f, err = os.Open(t.name); err != nil {
			return err
		}
	}

	if t.m == nil {
		if t.m, err = mmap.MapRegion(t.f, int(t.offsetLen), mmap.RDONLY, 0, t.offsetStartPos); err != nil {
			return err
		}
	}

	return nil
}

type tableWriter struct {
	sync.RWMutex

	wf *os.File
	rf *os.File

	rm sync.Mutex

	base  string
	name  string
	index int64

	first uint64
	last  uint64

	offsetBuf []byte

	frozen bool

	maxLogSize int64
}

func newTableWriter(base string, index int64, maxLogSize int64) *tableWriter {
	t := new(tableWriter)

	t.base = base
	t.name = path.Join(base, fmtTableName(index))
	t.index = index

	t.maxLogSize = maxLogSize

	return t
}

func (t *tableWriter) close() {
	if t.rf != nil {
		t.rf.Close()
		t.rf = nil
	}

	if t.wf != nil {
		t.wf.Close()
		t.wf = nil
	}
}

func (t *tableWriter) Close() {
	t.Lock()
	defer t.Unlock()

	t.close()
}

func (t *tableWriter) reset() {
	t.close()

	t.first = 0
	t.last = 0
	t.index = t.index + 1
	t.name = path.Join(t.base, fmtTableName(t.index))
	t.offsetBuf = t.offsetBuf[0:0]
}

func (t *tableWriter) Flush() (*tableReader, error) {
	t.Lock()
	defer t.Unlock()

	t.frozen = true

	if t.wf == nil {
		return nil, fmt.Errorf("nil write handler")
	}

	defer t.reset()

	tr := new(tableReader)
	tr.name = t.name
	tr.index = t.index

	st, _ := t.wf.Stat()

	tr.offsetStartPos = st.Size() + int64(len(log0Data))
	tr.offsetLen = uint32(len(t.offsetBuf))

	tr.first = t.first
	tr.last = t.last

	if n, err := t.wf.Write(log0Data); err != nil {
		log.Error("flush log0data error %s", err.Error())
		return nil, err
	} else if n != len(log0Data) {
		log.Error("flush log0data only %d != %d", n, len(log0Data))
		return nil, io.ErrShortWrite
	}

	if n, err := t.wf.Write(t.offsetBuf); err != nil {
		log.Error("flush offset buffer error %s", err.Error())
		return nil, err
	} else if n != len(t.offsetBuf) {
		log.Error("flush offset buffer only %d != %d", n, len(t.offsetBuf))
		return nil, io.ErrShortWrite
	}

	if err := binary.Write(t.wf, binary.BigEndian, tr.offsetStartPos); err != nil {
		log.Error("flush offset start pos error %s", err.Error())
		return nil, err
	}

	if err := binary.Write(t.wf, binary.BigEndian, tr.offsetLen); err != nil {
		log.Error("flush offset len error %s", err.Error())
		return nil, err
	}

	if n, err := t.wf.Write(magic); err != nil {
		log.Error("flush magic data error %s", err.Error())
		return nil, err
	} else if n != len(magic) {
		log.Error("flush magic data only %d != %d", n, len(magic))
		return nil, io.ErrShortWrite
	}

	return tr, nil
}

func (t *tableWriter) StoreLog(l *Log) error {
	if l.ID == 0 {
		return ErrStoreLogID
	}

	t.Lock()
	defer t.Unlock()

	if t.last > 0 && l.ID != t.last+1 {
		return ErrStoreLogID
	}

	if t.last-t.first+1 > maxLogNumInFile {
		return errTableNeedFlush
	}

	var err error
	if t.wf == nil {
		if t.wf, err = os.OpenFile(t.name, os.O_CREATE|os.O_WRONLY, 0644); err != nil {
			return err
		}
	}

	if t.offsetBuf == nil {
		t.offsetBuf = make([]byte, 0, maxLogNumInFile*4)
	}

	st, _ := t.wf.Stat()
	if st.Size() >= t.maxLogSize {
		return errTableNeedFlush
	}

	offsetPos := uint32(st.Size())

	if err := l.Encode(t.wf); err != nil {
		return err
	}

	t.offsetBuf = append(t.offsetBuf, num.Uint32ToBytes(offsetPos)...)
	if t.first == 0 {
		t.first = l.ID
	}

	t.last = l.ID

	//todo add LRU cache

	return nil
}

func (t *tableWriter) GetLog(id uint64, l *Log) error {
	t.RLock()
	defer t.RUnlock()

	if t.frozen {
		return errTableFrozen
	}

	if id < t.first || id > t.last {
		return fmt.Errorf("log %d not in [%d=%d]", id, t.first, t.last)
	}

	//todo memory cache

	offset := binary.BigEndian.Uint32(t.offsetBuf[(id-t.first)*4:])

	if err := t.getLog(l, int64(offset)); err != nil {
		return err
	} else if l.ID != id {
		return fmt.Errorf("invalid log id %d != %d", id, l.ID)
	}

	return nil
}

func (t *tableWriter) getLog(l *Log, pos int64) error {
	t.rm.Lock()
	defer t.rm.Unlock()

	var err error
	if t.rf == nil {
		if t.rf, err = os.Open(t.name); err != nil {
			return err
		}
	}

	if _, err = t.rf.Seek(pos, os.SEEK_SET); err != nil {
		return err
	}

	if err = l.Decode(t.rf); err != nil {
		return err
	}

	return nil
}