Reactor模型是典型的事件驱动模型。在网络编程中,所谓的事件当然就是read、write、bind、connect、close等这些动作了。Reactor模型的实现有很多种,下面介绍最基本的三种:
- 单线程版
- 多线程版
- 主从多线程版
Key Word:Java NIO,Reactor模型,Java并发编程,Event-Driven
单线程版本
结构图(引用自Doug Lea的Scalable IO in Java)如下:
Reactor模型图
上图中Reactor是一个典型的事件驱动中心,客户端发起请求并建立连接时,会触发注册在多路复用器Selector上的SelectionKey.OP_ACCEPT事件,绑定在该事件上的Acceptor对象的职责就是接受请求,为接下来的读写操作做准备。
Reactor设计如下:
代码语言:javascript复制/**
* Reactor
*
* @author wqx
*
*/
public class Reactor implements Runnable {
private static final Logger LOG = LoggerFactory.getLogger(Reactor.class);
private Selector selector;
private ServerSocketChannel ssc;
private Handler DEFAULT_HANDLER = new Handler(){
@Override
public void processRequest(Processor processor, ByteBuffer msg) {
//NOOP
}
};
private Handler handler = DEFAULT_HANDLER;
/**
* 启动阶段
* @param port
* @throws IOException
*/
public Reactor(int port, int maxClients, Handler serverHandler) throws IOException{
selector = Selector.open();
ssc = ServerSocketChannel.open();
ssc.configureBlocking(false);
ssc.socket().bind(new InetSocketAddress(port));
this.handler = serverHandler;
SelectionKey sk = ssc.register(selector, SelectionKey.OP_ACCEPT);
sk.attach(new Acceptor());
}
/**
* 轮询阶段
*/
@Override
public void run() {
while(!ssc.socket().isClosed()){
try {
selector.select(1000);
Set<SelectionKey> keys;
synchronized(this){
keys = selector.selectedKeys();
}
Iterator<SelectionKey> it = keys.iterator();
while(it.hasNext()){
SelectionKey key = it.next();
dispatch(key);
it.remove();
}
} catch (IOException e) {
e.printStackTrace();
}
}
close();
}
public void dispatch(SelectionKey key){
Runnable r = (Runnable)key.attachment();
if(r != null)
r.run();
}
/**
* 用于接受TCP连接的Acceptor
*
*/
class Acceptor implements Runnable{
@Override
public void run() {
SocketChannel sc;
try {
sc = ssc.accept();
if(sc != null){
new Processor(Reactor.this,selector,sc);
}
} catch (IOException e) {
e.printStackTrace();
}
}
}
public void close(){
try {
selector.close();
if(LOG.isDebugEnabled()){
LOG.debug("Close selector");
}
} catch (IOException e) {
LOG.warn("Ignoring exception during close selector, e=" e);
}
}
public void processRequest(Processor processor, ByteBuffer msg){
if(handler != DEFAULT_HANDLER){
handler.processRequest(processor, msg);
}
}
}
上面是典型的单线程版本的Reactor实现,实例化Reactor对象的过程中,在当前多路复用器Selector上注册了OP_ACCEPT事件,当OP_ACCEPT事件发生后,Reactor通过dispatch方法执行Acceptor的run方法,Acceptor类的主要功能就是接受请求,建立连接,并将代表连接建立的SocketChannel以参数的形式构造Processor对象。
Processor的任务就是进行I/O操作。
下面是Processor的源码:
代码语言:javascript复制/**
* Server Processor
*
* @author wqx
*/
public class Processor implements Runnable {
private static final Logger LOG = LoggerFactory.getLogger(Processor.class);
Reactor reactor;
private SocketChannel sc;
private final SelectionKey sk;
private final ByteBuffer lenBuffer = ByteBuffer.allocate(4);
private ByteBuffer inputBuffer = lenBuffer;
private ByteBuffer outputDirectBuffer = ByteBuffer.allocateDirect(1024 * 64);
private LinkedBlockingQueue<ByteBuffer> outputQueue = new LinkedBlockingQueue<ByteBuffer>();
public Processor(Reactor reactor, Selector sel,SocketChannel channel) throws IOException{
this.reactor = reactor;
sc = channel;
sc.configureBlocking(false);
sk = sc.register(sel, SelectionKey.OP_READ);
sk.attach(this);
sel.wakeup();
}
@Override
public void run() {
if(sc.isOpen() && sk.isValid()){
if(sk.isReadable()){
doRead();
}else if(sk.isWritable()){
doSend();
}
}else{
LOG.error("try to do read/write operation on null socket");
try {
if(sc != null)
sc.close();
} catch (IOException e) {}
}
}
private void doRead(){
try {
int byteSize = sc.read(inputBuffer);
if(byteSize < 0){
LOG.error("Unable to read additional data");
}
if(!inputBuffer.hasRemaining()){
if(inputBuffer == lenBuffer){
//read length
inputBuffer.flip();
int len = inputBuffer.getInt();
if(len < 0){
throw new IllegalArgumentException("Illegal data length");
}
//prepare for receiving data
inputBuffer = ByteBuffer.allocate(len);
}else{
//read data
if(inputBuffer.hasRemaining()){
sc.read(inputBuffer);
}
if(!inputBuffer.hasRemaining()){
inputBuffer.flip();
processRequest();
//clear lenBuffer and waiting for next reading operation
lenBuffer.clear();
inputBuffer = lenBuffer;
}
}
}
} catch (IOException e) {
LOG.error("Unexcepted Exception during read. e=" e);
try {
if(sc != null)
sc.close();
} catch (IOException e1) {
LOG.warn("Ignoring exception when close socketChannel");
}
}
}
/**
* process request and get response
*
* @param request
* @return
*/
private void processRequest(){
reactor.processRequest(this,inputBuffer);
}
private void doSend(){
try{
/**
* write data to channel:
* step 1: write the length of data(occupy 4 byte)
* step 2: data content
*/
if(outputQueue.size() > 0){
ByteBuffer directBuffer = outputDirectBuffer;
directBuffer.clear();
for(ByteBuffer buf : outputQueue){
buf.flip();
if(buf.remaining() > directBuffer.remaining()){
//prevent BufferOverflowException
buf = (ByteBuffer) buf.slice().limit(directBuffer.remaining());
}
//transfers the bytes remaining in buf into directBuffer
int p = buf.position();
directBuffer.put(buf);
//reset position
buf.position(p);
if(!directBuffer.hasRemaining()){
break;
}
}
directBuffer.flip();
int sendSize = sc.write(directBuffer);
while(!outputQueue.isEmpty()){
ByteBuffer buf = outputQueue.peek();
int left = buf.remaining() - sendSize;
if(left > 0){
buf.position(buf.position() sendSize);
break;
}
sendSize -= buf.remaining();
outputQueue.remove();
}
}
synchronized(reactor){
if(outputQueue.size() == 0){
//disable write
disableWrite();
}else{
//enable write
enableWrite();
}
}
} catch (CancelledKeyException e) {
LOG.warn("CancelledKeyException occur e=" e);
} catch (IOException e) {
LOG.warn("Exception causing close, due to " e);
}
}
public void sendBuffer(ByteBuffer bb){
try{
synchronized(this.reactor){
if(LOG.isDebugEnabled()){
LOG.debug("add sendable bytebuffer into outputQueue");
}
//wrap ByteBuffer with length header
ByteBuffer wrapped = wrap(bb);
outputQueue.add(wrapped);
enableWrite();
}
}catch(Exception e){
LOG.error("Unexcepted Exception: ", e);
}
}
private ByteBuffer wrap(ByteBuffer bb){
bb.flip();
lenBuffer.clear();
int len = bb.remaining();
lenBuffer.putInt(len);
ByteBuffer resp = ByteBuffer.allocate(len 4);
lenBuffer.flip();
resp.put(lenBuffer);
resp.put(bb);
return resp;
}
private void enableWrite(){
int i = sk.interestOps();
if((i & SelectionKey.OP_WRITE) == 0){
sk.interestOps(i | SelectionKey.OP_WRITE);
}
}
private void disableWrite(){
int i = sk.interestOps();
if((i & SelectionKey.OP_WRITE) == 1){
sk.interestOps(i & (~SelectionKey.OP_WRITE));
}
}
}
其实Processor要做的事情很简单,就是向selector注册感兴趣的读写时间,OP_READ或OP_WRITE,然后等待事件触发,做相应的操作。
代码语言:javascript复制 @Override
public void run() {
if(sc.isOpen() && sk.isValid()){
if(sk.isReadable()){
doRead();
}else if(sk.isWritable()){
doSend();
}
}else{
LOG.error("try to do read/write operation on null socket");
try {
if(sc != null)
sc.close();
} catch (IOException e) {}
}
}
而doRead()和doSend()方法稍微复杂了一点,这里其实处理了用TCP协议进行通信时必须要解决的问题:TCP粘包拆包问题
TCP粘包拆包问题
我们都知道TCP协议是面向字节流的,而字节流是连续的,无法有效识别应用层数据的边界。如下图:
粘包拆包示意图
上图显示的应用层有三个数据包,D1,D2,D3.当应用层数据传到传输层后,可能会出现粘包拆包现象。
TCP协议的基本传输单位是报文段,而每个报文段最大有效载荷是有限制的,一般以太网MTU为1500,去除IP头20B,TCP头20B,那么剩下的1460B就是传输层最大报文段的有效载荷。如果应用层数据大于该值(如上图中的数据块D2),那么传输层就会进行拆分重组。
解决方案
- 消息定长(通信双方发送的消息固定长度,缺点很明显:浪费可耻!!!)
- 每个消息之间加分割符(缺点:消息编解码耗时,并且如果消息体中本省就包含分隔字符,需要进行转义,效率低)
- 每个数据包加个Header!!!(header中指定后面数据的长度,这不就是Tcp、Ip协议通用的做法么。。。哈哈)
采用方案三
示意图如下:
数据包结构
header区占用4B,内容为数据的长度。too simple。。。-_-
理论有了,下面具体分析下Read、Write的实现过程:
doRead
inputBuffer负责接受数据,lenBuffer负责接受数据长度,初始化的时候,将lenBuffer赋给inputBuffer,定义如下:
代码语言:javascript复制private final ByteBuffer lenBuffer = ByteBuffer.allocate(4);
private ByteBuffer inputBuffer = lenBuffer;
- 如果inputBuffer == lenBuffer,那么从inputBuffer中读取出一个整型值len,这个值就是接下来要接受的数据的长度。同时分配一个大小为len的内存空间,并复制给inputBuffer,准备接受数据!!!
private void doRead(){
try {
int byteSize = sc.read(inputBuffer);
if(byteSize < 0){
LOG.error("Unable to read additional data");
}
if(!inputBuffer.hasRemaining()){
if(inputBuffer == lenBuffer){
//read length
inputBuffer.flip();
int len = inputBuffer.getInt();
if(len < 0){
throw new IllegalArgumentException("Illegal data length");
}
//prepare for receiving data
inputBuffer = ByteBuffer.allocate(len);
else{...}
- 如果inputBuffer != lenBuffer,那么开始接受数据吧!
if(inputBuffer == lenBuffer){
//。。。
}else{
//read data
if(inputBuffer.hasRemaining()){
sc.read(inputBuffer);
}
if(!inputBuffer.hasRemaining()){
inputBuffer.flip();
processRequest();
//clear lenBuffer and waiting for next reading operation
lenBuffer.clear();
inputBuffer = lenBuffer;
}
}
注意:
- 必须保证缓冲区是满的,即inputBuffer.hasRemaining()=false
- processRequest后,将inputBuffer重新赋值为lenBuffer,为下一次读操作做准备。
doWrite
用户调用sendBuffer方法发送数据,其实就是将数据加入outputQueue,这个outputQueue就是一个发送缓冲队列。
代码语言:javascript复制public void sendBuffer(ByteBuffer bb){
try{
synchronized(this.reactor){
if(LOG.isDebugEnabled()){
LOG.debug("add sendable bytebuffer into outputQueue");
}
//wrap ByteBuffer with length header
ByteBuffer wrapped = wrap(bb);
outputQueue.add(wrapped);
enableWrite();
}
}catch(Exception e){
LOG.error("Unexcepted Exception: ", e);
}
}
doSend方法就很好理解了,无非就是不断从outputQueue中取数据,然后写入channel中即可。过程如下:
将发送队列outputQueue中的数据写入缓冲区outputDirectBuffer:
- 清空outputDirectBuffer,为发送数据做准备
- 将outputQueue数据写入outputDirectBuffer
- 调用socketChannel.write(outputDirectBuffer);将outputDirectBuffer写入socket缓冲区
执行步骤2的时候,我们可能会遇到这么几种情况: 1.某个数据包大小超过了outputDirectBuffer剩余空间大小 2.outputDirectBuffer已被填满,但是outputQueue仍有待发送的数据 执行步骤3的时候,也可能出现下面两种情况: 1.outputDirectBuffer被全部写入socket缓冲区 2.outputDirectBuffer只有部分数据或者压根就没有数据被写入socket缓冲区
实现过程可以结合源码,这里重点分析下面几个点:
代码语言:javascript复制为什么需要重置buf的position
int p = buf.position();
directBuffer.put(buf);
//reset position
buf.position(p);
写入directBuffer的数据是即将被写入SocketChannel的数据,问题就在于:当我们调用
代码语言:javascript复制int sendSize = sc.write(directBuffer);
的时候,directBuffer中的数据都被写入Channel了吗?明显是不确定的(具体可以看java.nio.channels.SocketChannel.write(ByteBuffer src)的doc文档)
上面的问题如何解决
思路很简单,根据write方法返回值sendSize,遍历outputQueue中的ByteBuffer,根据buf.remaining()和sendSize的大小,才可以确定buf是否真的被发送了。如下所示:
代码语言:javascript复制while(!outputQueue.isEmpty()){
ByteBuffer buf = outputQueue.peek();
int left = buf.remaining() - sendSize;
if(left > 0){
buf.position(buf.position() sendSize);
break;
}
sendSize -= buf.remaining();
outputQueue.remove();
}
网络通信基本解决,上面的处理思路是参照Zookeeper网络模块的实现,有兴趣可以看Zookeeper相应源码。
测试
Server端:
代码语言:javascript复制public class ServerTest {
private static int PORT = 8888;
public static void main(String[] args) throws IOException, InterruptedException {
Thread t = new Thread(new Reactor(PORT,1024,new MyHandler()));
t.start();
System.out.println("server start");
t.join();
}
}
用户自定义Handler:
代码语言:javascript复制public class MyHandler implements Handler {
@Override
public void processRequest(Processor processor, ByteBuffer msg) {
byte[] con = new byte[msg.remaining()];
msg.get(con);
String str = new String(con,0,con.length);
String resp = "";
switch(str){
case "request1":resp = "response1";break;
case "request2":resp = "response2";break;
case "request3":resp = "response3";break;
default :resp = "";
}
ByteBuffer buf = ByteBuffer.allocate(resp.getBytes().length);
buf.put(resp.getBytes());
processor.sendBuffer(buf);
}
}
client端
代码语言:javascript复制public class ClientTest {
private static String HOST = "localhost";
private static int PORT = 8888;
public static void main(String[] args) throws IOException {
Client client = new Client();
client.socket().setTcpNoDelay(true);
client.connect(
new InetSocketAddress(HOST,PORT));
ByteBuffer msg;
for(int i = 1; i <= 3; i ){
msg = ByteBuffer.wrap(("request" i).getBytes());
System.out.println("send-" "request" i);
ByteBuffer resp = client.send(msg);
byte[] retVal = new byte[resp.remaining()];
resp.get(retVal);
System.out.println("receive-" new String(retVal,0,retVal.length));
}
}
}
输出:
代码语言:javascript复制send-request1
receive-response1
send-request2
receive-response2
send-request3
receive-response3
Client是一个客户端工具类,简单封装了发送ByteBuffer前,添加header的逻辑。详见源码。Client.java
总结
在这种实现方式中,dispatch方法是同步阻塞的!!!所有的IO操作和业务逻辑处理都在NIO线程(即Reactor线程)中完成。如果业务处理很快,那么这种实现方式没什么问题,不用切换到用户线程。但是,想象一下如果业务处理很耗时(涉及很多数据库操作、磁盘操作等),那么这种情况下Reactor将被阻塞,这肯定是我们不希望看到的。解决方法很简单,业务逻辑进行异步处理,即交给用户线程处理。
下面分析下单线程版的Reactor模型的缺点:
- 自始自终都只有一个Reactor线程,缺点很明显:Reactor意外挂了,整个系统也就无法正常工作,可靠性太差。
- 单线程的另外一个问题是在大负载的情况下,Reactor的处理速度必然会成为系统性能的瓶颈。
如何解决上述问题呢?下文详解Reactor多线程版本
GitHub完整源码