Need help with my IOCP client code #2

#include "assert.h"
#include <winsock2.h>
#include <mswsock.h>
#include <process.h>
#include <stdio.h>
#pragma comment(lib,"ws2_32.lib")
#pragma comment(lib,"mswsock.lib")

#include "iostream"
using namespace std;

struct IODATA
{
	OVERLAPPED ol;
	DWORD op;
	SOCKET s;
	HANDLE hIocp;
	char *buff;				// I/O操作使用的缓冲区
	int nLen;				// buff缓冲区（使用的）大小
};

char szSendMsg[4096]={0};
char szRevMsg[4096]={0};
#define CONNCHECKDATA "hello0123456"
IODATA io={0};

DWORD WINAPI SendThread(LPVOID lpparam);

int main(int argc, char* argv[])
{
	int ret=0;
	DWORD bytes=0;
	WSADATA data={0};
	static LPFN_CONNECTEX lpfnConnectEx=NULL;
	int iRet=WSAStartup(0x0202,&data);

	HANDLE hIocp=CreateIoCompletionPort(INVALID_HANDLE_VALUE,NULL,0,0);
	
	io.s=socket(AF_INET,SOCK_STREAM,0);
	io.op=FD_CONNECT;
	io.hIocp=hIocp;
	

	SOCKADDR_IN lAddr,rAddr;
	lAddr.sin_family=AF_INET;
	lAddr.sin_port=INADDR_ANY;
	lAddr.sin_addr.s_addr = INADDR_ANY;

	rAddr.sin_family=AF_INET;
	rAddr.sin_port=htons(4567);
	rAddr.sin_addr.s_addr=inet_addr("127.0.0.1");

	ret=bind(io.s,(SOCKADDR*)&lAddr,sizeof(lAddr));
	CreateIoCompletionPort((HANDLE)io.s,hIocp,(ULONG_PTR)&io,0);

	if(NULL==lpfnConnectEx)
	{
		GUID guid=WSAID_CONNECTEX;
		ret=WSAIoctl(io.s,SIO_GET_EXTENSION_FUNCTION_POINTER,&guid,sizeof(guid),&lpfnConnectEx,sizeof(lpfnConnectEx),&bytes,NULL,NULL);
	}

	//    _beginthreadex(NULL,0,_IoHandle,&io,0,NULL);  

	printf("connecting...\n");

	char SendBuffer[4096]={0};
	char* pMask = (char*)SendBuffer;
	*pMask=128;
	strcpy(pMask+3,CONNCHECKDATA);
	*(WORD*)(pMask+1) = strlen(CONNCHECKDATA)+1;
	ret=lpfnConnectEx(io.s,(SOCKADDR*)&rAddr,sizeof(rAddr),SendBuffer,*(WORD*)(pMask+1) + 3,&bytes,&io.ol);

	/*
	BOOL bResult = m_lpfnConnectEx (m_sockClient , 
		(sockaddr *)&addrPeer ,  // [in] 对方地址 
		nLen ,               // [in] 对方地址长度 
		lpSendBuffer ,       // [in] 连接后要发送的内容，这里不用 
		dwSendDataLength ,   // [in] 发送内容的字节数 ，这里不用 
		&dwBytesSent ,       // [out] 发送了多少个字节，这里不用 
		(OVERLAPPED *)pmyoverlapped ); // [in] 这东西复杂，下一篇有详解 
		*/


	DWORD dwTick=GetTickCount();
	while(1)
	{
		IODATA* pio=NULL;
		LPOVERLAPPED pol=NULL;
		//ret=GetQueuedCompletionStatus(hIocp,&bytes,(PULONG_PTR)&pio,&pol,1000);
		ret = ::GetQueuedCompletionStatus(hIocp,&bytes,(PULONG_PTR)&pio,&pol,WSA_INFINITE);

		printf("GQCS returned  ---  ");
		if(!pio)
		{
			printf("GQCS timeout.closesocket[socket=%d] -%d\n",io.s,GetTickCount()-dwTick);
			if(ret==0)
			{
				shutdown(io.s,SD_BOTH);
				closesocket(io.s);                
			}
		}
		else
		{
			if(pio->op==FD_CONNECT)
			{
				if(ret)
				{
					printf("connect OK, recving[socket=%d]...\n",pio->s);

					IODATA *pData = new IODATA;
					*pData = *pio;
					WSABUF wb;
					wb.buf=szRevMsg;
					wb.len=4096;
					pData->op = FD_READ;
					pData->buff = wb.buf;
					pData->nLen = wb.len;
					// Post another receive operation
					DWORD dwFlags = 0;
					DWORD dwRecvBytes;

					// Set up the per-I/O operation data for the next
					// overlapped call
					ZeroMemory(&(pData->ol), sizeof(OVERLAPPED));

					ret=WSARecv(pData->s,&wb,1,&dwRecvBytes,&dwFlags,&pData->ol,NULL);
					if (ret!=0)
					{
						if(::WSAGetLastError() != WSA_IO_PENDING)
						{
							printf("rev err:%d\n",WSAGetLastError());
						}
					}		

					HANDLE hthread;
					DWORD dwthreadid;
					hthread=CreateThread(NULL,0,SendThread,(LPVOID)0,0,&dwthreadid);
					if(hthread==NULL)
					{
						printf("fail55\n");
					}
					CloseHandle(hthread);
					continue;
				}
				else
				{
					printf("%s[%d] failed %d\n",pio->op==FD_CONNECT ? "Connect" : "Recv",pio->s,WSAGetLastError());
					break;
				}
			}
			else if(pio->op==FD_READ)
			{
				long *pLen = (long *)pio->buff;
				//assert(*pLen == bytes);

				pio->buff[bytes]=0;
				printf("printf returned[%d] MSG:%s\n",bytes,pio->buff+sizeof(long));

				continue;
			}
			else if (pio->op==FD_WRITE)
			{
// 				WSABUF wb;
// 				wb.buf=new char[4096];
// 				char* pMask = (char*)wb.buf;
// 				*pMask=128;
// 				strcpy(wb.buf+3,CONNCHECKDATA);
// 				*(WORD*)(pMask+1) = strlen(CONNCHECKDATA)+1;
// 				wb.len=*(WORD*)(pMask+1)+3;
// 				pio->op = FD_WRITE;
// 				ret=WSASend(pio->s,&wb,1,&bytes,0,&pio->ol,NULL);

				
				continue;
			}
		}
	}

	system("pause");
	return 0;
}

DWORD WINAPI SendThread(LPVOID lpparam)
{
	//int num=(int)lpparam;

	while (true)
	{
		cout<<"发送..."<<endl;
		WSABUF wb;
		wb.buf=szSendMsg;
		char* pMask = (char*)wb.buf;
		*pMask=128;
		strcpy(wb.buf+3,"测试数据abcd...");
		*(WORD*)(pMask+1) = strlen("测试数据abcd...")+1;
		wb.len=*(WORD*)(pMask+1)+3;
		io.op = FD_WRITE;
		io.buff = wb.buf;
		io.nLen = wb.len;
		DWORD bytes = 0 ;
		int ret=WSASend(io.s,&wb,1,&bytes,0,&io.ol,NULL);
		if (ret!=0)
		{
			printf("rev err:%d\n",WSAGetLastError());
		}		
		cin.get();
	}
	


	return 0;
}