void Connect()
{
try
{
socket = GetSocket();
if (socket == null)
{
HandleError();
return;
}
if (pageTimestamps != null && socket != null)
{
int size = pageTimestamps.Length;
byte[] psBuf = new byte[4 + size * 4];
Bytes.pack4(psBuf, 0, size);
for (int i = 0; i < size; i++)
{
Bytes.pack4(psBuf, (i + 1) * 4, pageTimestamps[i]);
}
socket.Send(psBuf, 0, psBuf.Length, SocketFlags.None);
}
return;
}
catch (SocketException x)
{
Console.WriteLine("Connection failure: " + x);
}
HandleError();
socket = null;
}
public int packString(int offs, string s)
{
if (s == null)
{
extend(offs + 4);
Bytes.pack4(arr, offs, -1);
offs += 4;
}
else
{
int len = s.Length;
if (encoding == null)
{
extend(offs + 4 + len * 2);
Bytes.pack4(arr, offs, len);
offs += 4;
for (int i = 0; i < len; i++)
{
Bytes.pack2(arr, offs, (short)s[i]);
offs += 2;
}
}
else
{
byte[] bytes = encoding.GetBytes(s);
extend(offs + 4 + bytes.Length);
Bytes.pack4(arr, offs, -2 - bytes.Length);
Array.Copy(bytes, 0, arr, offs + 4, bytes.Length);
offs += 4 + bytes.Length;
}
}
return(offs);
}
internal void pack(Page pg, int i)
{
byte[] dst = pg.data;
switch (key.type)
{
case ClassDescriptor.FieldType.tpBoolean:
case ClassDescriptor.FieldType.tpSByte:
case ClassDescriptor.FieldType.tpByte:
dst[BtreePage.firstKeyOffs + i] = (byte)key.ival;
break;
case ClassDescriptor.FieldType.tpShort:
case ClassDescriptor.FieldType.tpUShort:
case ClassDescriptor.FieldType.tpChar:
Bytes.pack2(dst, BtreePage.firstKeyOffs + i * 2, (short)key.ival);
break;
case ClassDescriptor.FieldType.tpInt:
case ClassDescriptor.FieldType.tpUInt:
case ClassDescriptor.FieldType.tpEnum:
case ClassDescriptor.FieldType.tpObject:
case ClassDescriptor.FieldType.tpOid:
Bytes.pack4(dst, BtreePage.firstKeyOffs + i * 4, key.ival);
break;
case ClassDescriptor.FieldType.tpLong:
case ClassDescriptor.FieldType.tpULong:
case ClassDescriptor.FieldType.tpDate:
Bytes.pack8(dst, BtreePage.firstKeyOffs + i * 8, key.lval);
break;
case ClassDescriptor.FieldType.tpFloat:
Bytes.packF4(dst, BtreePage.firstKeyOffs + i * 4, (float)key.dval);
break;
case ClassDescriptor.FieldType.tpDouble:
Bytes.packF8(dst, BtreePage.firstKeyOffs + i * 8, key.dval);
break;
case ClassDescriptor.FieldType.tpDecimal:
Bytes.packDecimal(dst, BtreePage.firstKeyOffs + i * 16, key.dec);
break;
case ClassDescriptor.FieldType.tpGuid:
Bytes.packGuid(dst, BtreePage.firstKeyOffs + i * 16, key.guid);
break;
default:
Debug.Assert(false, "Invalid type");
break;
}
Bytes.pack4(dst, BtreePage.firstKeyOffs + (BtreePage.maxItems - i - 1) * 4, oid);
}
public virtual void Write(long pos, byte[] buf)
{
lock (mutex)
{
if (pageTimestamps != null)
{
int pageNo = (int)(pos >> Page.pageSizeLog);
if (pageNo >= pageTimestamps.Length)
{
int newLength = pageNo >= pageTimestamps.Length * 2 ? pageNo + 1 : pageTimestamps.Length * 2;
int[] newPageTimestamps = new int[newLength];
Array.Copy(pageTimestamps, 0, newPageTimestamps, 0, pageTimestamps.Length);
pageTimestamps = newPageTimestamps;
int[] newDirtyPageTimestampMap = new int[(((newLength * 4 + Page.pageSize - 1) >> Page.pageSizeLog) + 31) >> 5];
Array.Copy(dirtyPageTimestampMap, 0, newDirtyPageTimestampMap, 0, dirtyPageTimestampMap.Length);
dirtyPageTimestampMap = newDirtyPageTimestampMap;
}
pageTimestamps[pageNo] = ++timestamp;
dirtyPageTimestampMap[pageNo >> (Page.pageSizeLog - 2 + 5)] |= 1 << ((pageNo >> (Page.pageSizeLog - 2)) & 31);
}
Bytes.pack8(txBuf, 0, pos);
Array.Copy(buf, 0, txBuf, 8, buf.Length);
if (pageTimestamps != null)
{
Bytes.pack4(txBuf, Page.pageSize + 8, timestamp);
}
for (int i = 0; i < replicas.Length; i++)
{
if (Send(i, txBuf))
{
if (ack && pos == 0 && !Receive(i, rcBuf))
{
Console.WriteLine("Failed to receive ACK from node " + replicas[i].host);
}
}
}
}
file.Write(pos, buf);
}
public void Sync()
{
if (pageTimestamps != null)
{
lock (mutex) {
byte[] page = new byte[Page.pageSize];
for (int i = 0; i < dirtyPageTimestampMap.Length; i++)
{
if (dirtyPageTimestampMap[i] != 0)
{
for (int j = 0; j < 32; j++)
{
if ((dirtyPageTimestampMap[i] & (1 << j)) != 0)
{
int pageNo = (i << 5) + j;
int beg = pageNo << (Page.pageSizeLog - 2);
int end = beg + Page.pageSize / 4;
if (end > pageTimestamps.Length)
{
end = pageTimestamps.Length;
}
int offs = 0;
while (beg < end)
{
Bytes.pack4(page, offs, pageTimestamps[beg]);
beg += 1;
offs += 4;
}
long pos = pageNo << Page.pageSizeLog;
pageTimestampFile.Write(pos, page);
}
}
dirtyPageTimestampMap[i] = 0;
}
}
}
pageTimestampFile.Sync();
}
file.Sync();
}
public int packI4(int offs, int val)
{
extend(offs + 4);
Bytes.pack4(arr, offs, val);
return(offs + 4);
}
private int packKeyPart(ByteBuffer buf, int dst, ClassDescriptor.FieldType type, object val)
{
switch (type)
{
case ClassDescriptor.FieldType.tpBoolean:
dst = buf.packBool(dst, (bool)val);
break;
case ClassDescriptor.FieldType.tpByte:
dst = buf.packI1(dst, (byte)val);
break;
case ClassDescriptor.FieldType.tpSByte:
dst = buf.packI1(dst, (sbyte)val);
break;
case ClassDescriptor.FieldType.tpShort:
dst = buf.packI2(dst, (short)val);
break;
case ClassDescriptor.FieldType.tpUShort:
dst = buf.packI2(dst, (ushort)val);
break;
case ClassDescriptor.FieldType.tpChar:
dst = buf.packI2(dst, (char)val);
break;
case ClassDescriptor.FieldType.tpInt:
case ClassDescriptor.FieldType.tpOid:
dst = buf.packI4(dst, (int)val);
break;
case ClassDescriptor.FieldType.tpEnum:
dst = buf.packI4(dst, Convert.ToInt32(val));
break;
case ClassDescriptor.FieldType.tpUInt:
dst = buf.packI4(dst, (int)(uint)val);
break;
case ClassDescriptor.FieldType.tpObject:
if (val != null)
{
Storage.MakePersistent(val);
dst = buf.packI4(dst, Storage.GetOid(val));
}
else
{
dst = buf.packI4(dst, 0);
}
break;
case ClassDescriptor.FieldType.tpLong:
dst = buf.packI8(dst, (long)val);
break;
case ClassDescriptor.FieldType.tpULong:
dst = buf.packI8(dst, (long)(ulong)val);
break;
case ClassDescriptor.FieldType.tpDate:
dst = buf.packDate(dst, (DateTime)val);
break;
case ClassDescriptor.FieldType.tpFloat:
dst = buf.packF4(dst, (float)val);
break;
case ClassDescriptor.FieldType.tpDouble:
dst = buf.packF8(dst, (double)val);
break;
case ClassDescriptor.FieldType.tpDecimal:
dst = buf.packDecimal(dst, (decimal)val);
break;
case ClassDescriptor.FieldType.tpGuid:
dst = buf.packGuid(dst, (Guid)val);
break;
case ClassDescriptor.FieldType.tpString:
dst = buf.packString(dst, (string)val);
break;
case ClassDescriptor.FieldType.tpArrayOfByte:
buf.extend(dst + 4);
if (val != null)
{
byte[] arr = (byte[])val;
int len = arr.Length;
Bytes.pack4(buf.arr, dst, len);
dst += 4;
buf.extend(dst + len);
Array.Copy(arr, 0, buf.arr, dst, len);
dst += len;
}
else
{
Bytes.pack4(buf.arr, dst, 0);
dst += 4;
}
break;
default:
Debug.Assert(false, "Invalid type");
break;
}
return(dst);
}
internal static void setItem(Page pg, int index, int mask)
{
Bytes.pack4(pg.data, firstKeyOffs + index * 4, mask);
}
internal static void setType(byte[] arr, int offs, int type)
{
Bytes.pack4(arr, offs + 4, type);
}
internal static void setSize(byte[] arr, int offs, int size)
{
Bytes.pack4(arr, offs, size);
}
public void run()
{
try
{
byte[] buf = new byte[Page.pageSize + PAGE_DATA_OFFSET + (pageTimestamps != null ? 4 : 0)];
while (listening)
{
int offs = 0;
do
{
int rc = -1;
try
{
if (socket != null)
{
ArrayList recvSocks = new ArrayList(1);
recvSocks.Add(socket);
Socket.Select(recvSocks, null, null, replicationReceiveTimeout * 2000);
if (recvSocks.Count == 1)
{
rc = socket.Receive(buf, offs, buf.Length - offs, SocketFlags.None);
}
else
{
Console.WriteLine("Receive timeout expired");
}
}
}
catch (SocketException x)
{
Console.WriteLine("Failed to receive data from master: " + x);
rc = -1;
}
lock (done)
{
if (!listening)
{
return;
}
}
if (rc < 0)
{
HandleError();
hangup();
return;
}
else
{
offs += rc;
}
} while (offs < buf.Length);
long pos = Bytes.unpack8(buf, 0);
bool transactionCommit = false;
if (pos == 0)
{
if (replicationAck)
{
try
{
socket.Send(buf, 0, 1, SocketFlags.None);
}
catch (SocketException x)
{
Console.WriteLine("Failed to send request to master: " + x);
HandleError();
hangup();
return;
}
}
if (buf[PAGE_DATA_OFFSET + DB_HDR_CURR_INDEX_OFFSET] != prevIndex)
{
prevIndex = buf[PAGE_DATA_OFFSET + DB_HDR_CURR_INDEX_OFFSET];
lck.ExclusiveLock();
transactionCommit = true;
}
}
else if (pos == REPL_SYNC)
{
lock (sync)
{
outOfSync = false;
Monitor.Pulse(sync);
}
continue;
}
else if (pos == REPL_PING)
{
continue;
}
else if (pos == REPL_CLOSE)
{
hangup();
return;
}
if (pageTimestamps != null)
{
int pageNo = (int)(pos >> Page.pageSizeLog);
if (pageNo >= pageTimestamps.Length)
{
int newLength = pageNo >= pageTimestamps.Length * 2 ? pageNo + 1 : pageTimestamps.Length * 2;
int[] newPageTimestamps = new int[newLength];
Array.Copy(pageTimestamps, 0, newPageTimestamps, 0, pageTimestamps.Length);
pageTimestamps = newPageTimestamps;
int[] newDirtyPageTimestampMap = new int[(((newLength * 4 + Page.pageSize - 1) >> Page.pageSizeLog) + 31) >> 5];
Array.Copy(dirtyPageTimestampMap, 0, newDirtyPageTimestampMap, 0, dirtyPageTimestampMap.Length);
dirtyPageTimestampMap = newDirtyPageTimestampMap;
}
int timestamp = Bytes.unpack4(buf, Page.pageSize + PAGE_DATA_OFFSET);
pageTimestamps[pageNo] = timestamp;
dirtyPageTimestampMap[pageNo >> (Page.pageSizeLog - 2 + 5)] |= 1 << ((pageNo >> (Page.pageSizeLog - 2)) & 31);
}
Page pg = pool.putPage(pos);
Array.Copy(buf, PAGE_DATA_OFFSET, pg.data, 0, Page.pageSize);
pool.unfix(pg);
if (pos == 0)
{
if (!initialized && buf[PAGE_DATA_OFFSET + DB_HDR_INITIALIZED_OFFSET] != 0)
{
lock (init)
{
initialized = true;
Monitor.Pulse(init);
}
}
if (transactionCommit)
{
lck.Unlock();
lock (commit)
{
Monitor.PulseAll(commit);
}
if (listener != null)
{
listener.OnMasterDatabaseUpdate();
}
pool.flush();
if (pageTimestamps != null)
{
byte[] page = new byte[Page.pageSize];
for (int i = 0; i < dirtyPageTimestampMap.Length; i++)
{
if (dirtyPageTimestampMap[i] != 0)
{
for (int j = 0; j < 32; j++)
{
if ((dirtyPageTimestampMap[i] & (1 << j)) != 0)
{
int pageNo = (i << 5) + j;
int beg = pageNo << (Page.pageSizeLog - 2);
int end = beg + Page.pageSize / 4;
if (end > pageTimestamps.Length)
{
end = pageTimestamps.Length;
}
offs = 0;
while (beg < end)
{
Bytes.pack4(page, offs, pageTimestamps[beg]);
beg += 1;
offs += 4;
}
pageTimestampFile.Write(pageNo << Page.pageSizeLog, page);
}
}
}
dirtyPageTimestampMap[i] = 0;
}
pageTimestampFile.Sync();
}
}
}
}
}
finally
{
listening = false;
lock (init)
{
Monitor.Pulse(init);
}
lock (sync)
{
Monitor.Pulse(sync);
}
}
}
public override void Write(long pos, byte[] buf)
{
file.Write(pos, buf);
lock (mutex)
{
if (pageTimestamps != null)
{
int pageNo = (int)(pos >> Page.pageSizeLog);
if (pageNo >= pageTimestamps.Length)
{
int newLength = pageNo >= pageTimestamps.Length * 2 ? pageNo + 1 : pageTimestamps.Length * 2;
int[] newPageTimestamps = new int[newLength];
Array.Copy(pageTimestamps, 0, newPageTimestamps, 0, pageTimestamps.Length);
pageTimestamps = newPageTimestamps;
int[] newDirtyPageTimestampMap = new int[(((newLength * 4 + Page.pageSize - 1) >> Page.pageSizeLog) + 31) >> 5];
Array.Copy(dirtyPageTimestampMap, 0, newDirtyPageTimestampMap, 0, dirtyPageTimestampMap.Length);
dirtyPageTimestampMap = newDirtyPageTimestampMap;
}
pageTimestamps[pageNo] = ++timestamp;
dirtyPageTimestampMap[pageNo >> (Page.pageSizeLog - 2 + 5)] |= 1 << ((pageNo >> (Page.pageSizeLog - 2)) & 31);
}
}
for (int i = 0; i < replicas.Length; i++)
{
if (replicas[i].socket != null)
{
byte[] data = new byte[txBuf.Length];
Bytes.pack8(data, 0, pos);
Array.Copy(buf, 0, data, 8, buf.Length);
if (pageTimestamps != null)
{
Bytes.pack4(data, Page.pageSize + 8, timestamp);
}
Parcel p = new Parcel();
p.data = data;
p.pos = pos;
p.host = i;
lock (async)
{
buffered += data.Length;
while (buffered > asyncBufSize && buffered != data.Length)
{
Monitor.Wait(async);
}
}
lock (go)
{
if (head == null)
{
head = tail = p;
}
else
{
tail = tail.next = p;
}
Monitor.Pulse(go);
}
}
}
}
void synchronizeNewSlaveNode(int i)
{
long size = storage.DatabaseSize;
int[] syncNodeTimestamps = null;
byte[] txBuf;
if (pageTimestamps != null)
{
txBuf = new byte[12 + Page.pageSize];
byte[] psBuf = new byte[4];
if (!Receive(i, psBuf))
{
Console.WriteLine("Failed to receive page timestamps length from slave node " + replicas[i].host);
return;
}
int psSize = Bytes.unpack4(psBuf, 0);
psBuf = new byte[psSize * 4];
if (!Receive(i, psBuf))
{
Console.WriteLine("Failed to receive page timestamps from slave node " + replicas[i].host);
return;
}
syncNodeTimestamps = new int[psSize];
for (int j = 0; j < psSize; j++)
{
syncNodeTimestamps[j] = Bytes.unpack4(psBuf, j * 4);
}
}
else
{
txBuf = new byte[8 + Page.pageSize];
}
for (long pos = 0; pos < size; pos += Page.pageSize)
{
int pageNo = (int)(pos >> Page.pageSizeLog);
if (syncNodeTimestamps != null)
{
if (pageNo < syncNodeTimestamps.Length &&
pageNo < pageTimestamps.Length &&
syncNodeTimestamps[pageNo] == pageTimestamps[pageNo])
{
continue;
}
}
lock (storage)
{
lock (storage.objectCache)
{
Page pg = storage.pool.getPage(pos);
Bytes.pack8(txBuf, 0, pos);
Array.Copy(pg.data, 0, txBuf, 8, Page.pageSize);
storage.pool.unfix(pg);
if (syncNodeTimestamps != null)
{
Bytes.pack4(txBuf, Page.pageSize + 8, pageNo < pageTimestamps.Length ? pageTimestamps[pageNo] : 0);
}
}
}
if (!Send(i, txBuf))
{
return;
}
if (ack && pos == 0 && !Receive(i, rcBuf))
{
Console.WriteLine("Failed to receive ACK from node " + replicas[i].host);
return;
}
}
Bytes.pack8(txBuf, 0, ReplicationSlaveStorageImpl.REPL_SYNC);
Send(i, txBuf);
}
