按照《Unix网络编程》的划分,IO模型可以分为:阻塞IO、非阻塞IO、IO复用、信号驱动IO和异步IO,按照POSIX标准来划分只分为两类:同步IO和异步IO。如何区分呢?首先一个IO操作其实分成了两个步骤:发起IO请求和实际的IO操作,同步IO和异步IO的区别就在于第二个步骤是否阻塞,如果实际的IO读写阻塞请求进程,那么就是同步IO,因此阻塞IO、非阻塞IO、IO服用、信号驱动IO都是同步IO,如果不阻塞,而是操作系统帮你做完IO操作再将结果返回给你,那么就是异步IO。阻塞IO和非阻塞IO的区别在于第一步,发起IO请求是否会被阻塞,如果阻塞直到完成那么就是传统的阻塞IO,如果不阻塞,那么就是非阻塞IO。
Java nio 2.0的主要改进就是引入了异步IO(包括文件和网络),这里主要介绍下异步网络IO API的使用以及框架的设计,以TCP服务端为例。首先看下为了支持AIO引入的新的类和接口:
java.nio.channels.AsynchronousChannel
标记一个channel支持异步IO操作。
java.nio.channels.AsynchronousServerSocketChannel
ServerSocket的aio版本,创建TCP服务端,绑定地址,监听端口等。
java.nio.channels.AsynchronousSocketChannel
面向流的异步socket channel,表示一个连接。
java.nio.channels.AsynchronousChannelGroup
异步channel的分组管理,目的是为了资源共享。一个AsynchronousChannelGroup绑定一个线程池,这个线程池执行两个任务:处理IO事件和派发CompletionHandler。AsynchronousServerSocketChannel创建的时候可以传入一个 AsynchronousChannelGroup,那么通过AsynchronousServerSocketChannel创建的 AsynchronousSocketChannel将同属于一个组,共享资源。
java.nio.channels.CompletionHandler
异步IO操作结果的回调接口,用于定义在IO操作完成后所作的回调工作。AIO的API允许两种方式来处理异步操作的结果:返回的Future模式或者注册CompletionHandler,我更推荐用CompletionHandler的方式,这些handler的调用是由 AsynchronousChannelGroup的线程池派发的。显然,线程池的大小是性能的关键因素。AsynchronousChannelGroup允许绑定不同的线程池,通过三个静态方法来创建:
- public static AsynchronousChannelGroup withFixedThreadPool(int nThreads,
- ThreadFactory threadFactory)
- throws IOException
-
- public static AsynchronousChannelGroup withCachedThreadPool(ExecutorService executor,
- int initialSize)
-
- public static AsynchronousChannelGroup withThreadPool(ExecutorService executor)
- throws IOException
public static AsynchronousChannelGroup withFixedThreadPool(int nThreads,
ThreadFactory threadFactory)
throws IOException
public static AsynchronousChannelGroup withCachedThreadPool(ExecutorService executor,
int initialSize)
public static AsynchronousChannelGroup withThreadPool(ExecutorService executor)
throws IOException
需要根据具体应用相应调整,从框架角度出发,需要暴露这样的配置选项给用户。
在介绍完了aio引入的TCP的主要接口和类之后,我们来设想下一个aio框架应该怎么设计。参考非阻塞nio框架的设计,一般都是采用Reactor模式,Reacot负责事件的注册、select、事件的派发;相应地,异步IO有个Proactor模式,Proactor负责 CompletionHandler的派发,查看一个典型的IO写操作的流程来看两者的区别:
Reactor: send(msg) -> 消息队列是否为空,如果为空 -> 向Reactor注册OP_WRITE,然后返回 -> Reactor select -> 触发Writable,通知用户线程去处理 ->先注销Writable(很多人遇到的cpu 100%的问题就在于没有注销),处理Writeable,如果没有完全写入,继续注册OP_WRITE。注意到,写入的工作还是用户线程在处理。
Proactor: send(msg) -> 消息队列是否为空,如果为空,发起read异步调用,并注册CompletionHandler,然后返回。 -> 操作系统负责将你的消息写入,并返回结果(写入的字节数)给Proactor -> Proactor派发CompletionHandler。可见,写入的工作是操作系统在处理,无需用户线程参与。事实上在aio的API 中,AsynchronousChannelGroup就扮演了Proactor的角色。
CompletionHandler有三个方法,分别对应于处理成功、失败、被取消(通过返回的Future)情况下的回调处理:
- public interface CompletionHandler<V,A> {
-
- void completed(V result, A attachment);
-
- void failed(Throwable exc, A attachment);
-
-
- void cancelled(A attachment);
- }
public interface CompletionHandler<V,A> {
void completed(V result, A attachment);
void failed(Throwable exc, A attachment);
void cancelled(A attachment);
}
其中的泛型参数V表示IO调用的结果,而A是发起调用时传入的attchment。
在初步介绍完aio引入的类和接口后,我们看看一个典型的tcp服务端是怎么启动的,怎么接受连接并处理读和写,这里引用的代码都是yanf4j 的aio分支中的代码,可以从svn checkout,svn地址:http://yanf4j.googlecode.com/svn/branches/yanf4j-aio
第一步,创建一个AsynchronousServerSocketChannel,创建之前先创建一个 AsynchronousChannelGroup,上文提到AsynchronousServerSocketChannel可以绑定一个 AsynchronousChannelGroup,那么通过这个AsynchronousServerSocketChannel建立的连接都将同属于一个AsynchronousChannelGroup并共享资源:
- this.asynchronousChannelGroup = AsynchronousChannelGroup
- .withCachedThreadPool(Executors.newCachedThreadPool(),
- this.threadPoolSize);
this.asynchronousChannelGroup = AsynchronousChannelGroup
.withCachedThreadPool(Executors.newCachedThreadPool(),
this.threadPoolSize);
然后初始化一个AsynchronousServerSocketChannel,通过open方法:
- this.serverSocketChannel = AsynchronousServerSocketChannel
- .open(this.asynchronousChannelGroup);
this.serverSocketChannel = AsynchronousServerSocketChannel
.open(this.asynchronousChannelGroup);
通过nio 2.0引入的SocketOption类设置一些TCP选项:
- this.serverSocketChannel
- .setOption(
- StandardSocketOption.SO_REUSEADDR,true);
- this.serverSocketChannel
- .setOption(
- StandardSocketOption.SO_RCVBUF,16*1024);
this.serverSocketChannel
.setOption(
StandardSocketOption.SO_REUSEADDR,true);
this.serverSocketChannel
.setOption(
StandardSocketOption.SO_RCVBUF,16*1024);
绑定本地地址:
- this.serverSocketChannel
- .bind(new InetSocketAddress("localhost",8080), 100);
this.serverSocketChannel
.bind(new InetSocketAddress("localhost",8080), 100);
其中的100用于指定等待连接的队列大小(backlog)。完了吗?还没有,最重要的监听工作还没开始,监听端口是为了等待连接上来以便accept产生一个AsynchronousSocketChannel来表示一个新建立的连接,因此需要发起一个accept调用,调用是异步的,操作系统将在连接建立后,将最后的结果——AsynchronousSocketChannel返回给你:
- public void pendingAccept() {
- if (this.started && this.serverSocketChannel.isOpen()) {
- this.acceptFuture = this.serverSocketChannel.accept(null,
- new AcceptCompletionHandler());
-
- } else {
- throw new IllegalStateException("Controller has been closed");
- }
- }
public void pendingAccept() {
if (this.started && this.serverSocketChannel.isOpen()) {
this.acceptFuture = this.serverSocketChannel.accept(null,
new AcceptCompletionHandler());
} else {
throw new IllegalStateException("Controller has been closed");
}
}
注意,重复的accept调用将会抛出PendingAcceptException,后文提到的read和write也是如此。accept方法的第一个参数是你想传给CompletionHandler的attchment,第二个参数就是注册的用于回调的CompletionHandler,最后返回结果Future<AsynchronousSocketChannel>。你可以对future做处理,这里采用更推荐的方式就是注册一个CompletionHandler。那么accept的CompletionHandler中做些什么工作呢?显然一个赤裸裸的 AsynchronousSocketChannel是不够的,我们需要将它封装成session,一个session表示一个连接(mina里就叫 IoSession了),里面带了一个缓冲的消息队列以及一些其他资源等。在连接建立后,除非你的服务器只准备接受一个连接,不然你需要在后面继续调用pendingAccept来发起另一个accept请求:
- private final class AcceptCompletionHandler implements
- CompletionHandler<AsynchronousSocketChannel, Object> {
-
- @Override
- public void cancelled(Object attachment) {
- logger.warn("Accept operation was canceled");
- }
-
- @Override
- public void completed(AsynchronousSocketChannel socketChannel,
- Object attachment) {
- try {
- logger.debug("Accept connection from "
- + socketChannel.getRemoteAddress());
- configureChannel(socketChannel);
- AioSessionConfig sessionConfig = buildSessionConfig(socketChannel);
- Session session = new AioTCPSession(sessionConfig,
- AioTCPController.this.configuration
- .getSessionReadBufferSize(),
- AioTCPController.this.sessionTimeout);
- session.start();
- registerSession(session);
- } catch (Exception e) {
- e.printStackTrace();
- logger.error("Accept error", e);
- notifyException(e);
- } finally {
- <strong>pendingAccept</strong>();
- }
- }
-
- @Override
- public void failed(Throwable exc, Object attachment) {
- logger.error("Accept error", exc);
- try {
- notifyException(exc);
- } finally {
- <strong>pendingAccept</strong>();
- }
- }
- }
private final class AcceptCompletionHandler implements
CompletionHandler<AsynchronousSocketChannel, Object> {
@Override
public void cancelled(Object attachment) {
logger.warn("Accept operation was canceled");
}
@Override
public void completed(AsynchronousSocketChannel socketChannel,
Object attachment) {
try {
logger.debug("Accept connection from "
+ socketChannel.getRemoteAddress());
configureChannel(socketChannel);
AioSessionConfig sessionConfig = buildSessionConfig(socketChannel);
Session session = new AioTCPSession(sessionConfig,
AioTCPController.this.configuration
.getSessionReadBufferSize(),
AioTCPController.this.sessionTimeout);
session.start();
registerSession(session);
} catch (Exception e) {
e.printStackTrace();
logger.error("Accept error", e);
notifyException(e);
} finally {
pendingAccept();
}
}
@Override
public void failed(Throwable exc, Object attachment) {
logger.error("Accept error", exc);
try {
notifyException(exc);
} finally {
pendingAccept();
}
}
}
注意到了吧,我们在failed和completed方法中在最后都调用了pendingAccept来继续发起accept调用,等待新的连接上来。有的同学可能要说了,这样搞是不是递归调用,会不会堆栈溢出?实际上不会,因为发起accept调用的线程与CompletionHandler回调的线程并非同一个,不是一个上下文中,两者之间没有耦合关系。要注意到,CompletionHandler的回调共用的是 AsynchronousChannelGroup绑定的线程池,因此千万别在CompletionHandler回调方法中调用阻塞或者长时间的操作,例如sleep,回调方法最好能支持超时,防止线程池耗尽。
连接建立后,怎么读和写呢?回忆下在nonblocking nio框架中,连接建立后的第一件事是干什么?注册OP_READ事件等待socket可读。异步IO也同样如此,连接建立后马上发起一个异步read调用,等待socket可读,这个是Session.start方法中所做的事情:
- public class AioTCPSession {
- protected void start0() {
- pendingRead();
- }
-
- protected final void pendingRead() {
- if (!isClosed() && this.asynchronousSocketChannel.isOpen()) {
- if (!this.readBuffer.hasRemaining()) {
- this.readBuffer = ByteBufferUtils
- .increaseBufferCapatity(this.readBuffer);
- }
- this.readFuture = this.asynchronousSocketChannel.read(
- this.readBuffer, this, this.readCompletionHandler);
- } else {
- throw new IllegalStateException(
- "Session Or Channel has been closed");
- }
- }
-
- }
public class AioTCPSession {
protected void start0() {
pendingRead();
}
protected final void pendingRead() {
if (!isClosed() && this.asynchronousSocketChannel.isOpen()) {
if (!this.readBuffer.hasRemaining()) {
this.readBuffer = ByteBufferUtils
.increaseBufferCapatity(this.readBuffer);
}
this.readFuture = this.asynchronousSocketChannel.read(
this.readBuffer, this, this.readCompletionHandler);
} else {
throw new IllegalStateException(
"Session Or Channel has been closed");
}
}
}
AsynchronousSocketChannel的read调用与AsynchronousServerSocketChannel的accept调用类似,同样是非阻塞的,返回结果也是一个Future,但是写的结果是整数,表示写入了多少字节,因此read调用返回的是Future<Integer>,方法的第一个参数是读的缓冲区,操作系统将IO读到数据拷贝到这个缓冲区,第二个参数是传递给 CompletionHandler的attchment,第三个参数就是注册的用于回调的CompletionHandler。这里保存了read的结果Future,这是为了在关闭连接的时候能够主动取消调用,accept也是如此。现在可以看看read的CompletionHandler的实现:
- public final class ReadCompletionHandler implements
- CompletionHandler<Integer, AbstractAioSession> {
-
- private static final Logger log = LoggerFactory
- .getLogger(ReadCompletionHandler.class);
- protected final AioTCPController controller;
-
- public ReadCompletionHandler(AioTCPController controller) {
- this.controller = controller;
- }
-
- @Override
- public void cancelled(AbstractAioSession session) {
- log.warn("Session(" + session.getRemoteSocketAddress()
- + ") read operation was canceled");
- }
-
- @Override
- public void completed(Integer result, AbstractAioSession session) {
- if (log.isDebugEnabled())
- log.debug("Session(" + session.getRemoteSocketAddress()
- + ") read +" + result + " bytes");
- if (result < 0) {
- session.close();
- return;
- }
- try {
- if (result > 0) {
- session.updateTimeStamp();
- session.getReadBuffer().flip();
- session.decode();
- session.getReadBuffer().compact();
- }
- } finally {
- try {
- session.pendingRead();
- } catch (IOException e) {
- session.onException(e);
- session.close();
- }
- }
- controller.checkSessionTimeout();
- }
-
- @Override
- public void failed(Throwable exc, AbstractAioSession session) {
- log.error("Session read error", exc);
- session.onException(exc);
- session.close();
- }
-
- }
public final class ReadCompletionHandler implements
CompletionHandler<Integer, AbstractAioSession> {
private static final Logger log = LoggerFactory
.getLogger(ReadCompletionHandler.class);
protected final AioTCPController controller;
public ReadCompletionHandler(AioTCPController controller) {
this.controller = controller;
}
@Override
public void cancelled(AbstractAioSession session) {
log.warn("Session(" + session.getRemoteSocketAddress()
+ ") read operation was canceled");
}
@Override
public void completed(Integer result, AbstractAioSession session) {
if (log.isDebugEnabled())
log.debug("Session(" + session.getRemoteSocketAddress()
+ ") read +" + result + " bytes");
if (result < 0) {
session.close();
return;
}
try {
if (result > 0) {
session.updateTimeStamp();
session.getReadBuffer().flip();
session.decode();
session.getReadBuffer().compact();
}
} finally {
try {
session.pendingRead();
} catch (IOException e) {
session.onException(e);
session.close();
}
}
controller.checkSessionTimeout();
}
@Override
public void failed(Throwable exc, AbstractAioSession session) {
log.error("Session read error", exc);
session.onException(exc);
session.close();
}
}
如果IO读失败,会返回失败产生的异常,这种情况下我们就主动关闭连接,通过session.close()方法,这个方法干了两件事情:关闭channel和取消read调用:
- if (null != this.readFuture) {
- this.readFuture.cancel(true);
- }
- this.asynchronousSocketChannel.close();
if (null != this.readFuture) {
this.readFuture.cancel(true);
}
this.asynchronousSocketChannel.close();
在读成功的情况下,我们还需要判断结果result是否小于0,如果小于0就表示对端关闭了,这种情况下我们也主动关闭连接并返回。如果读到一定字节,也就是result大于0的情况下,我们就尝试从读缓冲区中decode出消息,并派发给业务处理器的回调方法,最终通过pendingRead继续发起read调用等待socket的下一次可读。可见,我们并不需要自己去调用channel来进行IO读,而是操作系统帮你直接读到了缓冲区,然后给你一个结果表示读入了多少字节,你处理这个结果即可。而nonblocking IO框架中,是reactor通知用户线程socket可读了,然后用户线程自己去调用read进行实际读操作。这里还有个需要注意的地方,就是decode出来的消息的派发给业务处理器工作最好交给一个线程池来处理,避免阻塞group绑定的线程池。
IO写的操作与此类似,不过通常写的话我们会在session中关联一个缓冲队列来处理,没有完全写入或者等待写入的消息都存放在队列中,队列为空的情况下发起write调用:
- protected void write0(WriteMessage message) {
- boolean needWrite = false;
- synchronized (this.writeQueue) {
- needWrite = this.writeQueue.isEmpty();
- this.writeQueue.offer(message);
- }
- if (needWrite) {
- pendingWrite(message);
- }
- }
-
- protected final void pendingWrite(WriteMessage message) {
- message = preprocessWriteMessage(message);
- if (!isClosed() && this.asynchronousSocketChannel.isOpen()) {
- this.asynchronousSocketChannel.write(message.getWriteBuffer(),
- this, this.writeCompletionHandler);
- } else {
- throw new IllegalStateException(
- "Session Or Channel has been closed");
- }
- }
protected void write0(WriteMessage message) {
boolean needWrite = false;
synchronized (this.writeQueue) {
needWrite = this.writeQueue.isEmpty();
this.writeQueue.offer(message);
}
if (needWrite) {
pendingWrite(message);
}
}
protected final void pendingWrite(WriteMessage message) {
message = preprocessWriteMessage(message);
if (!isClosed() && this.asynchronousSocketChannel.isOpen()) {
this.asynchronousSocketChannel.write(message.getWriteBuffer(),
this, this.writeCompletionHandler);
} else {
throw new IllegalStateException(
"Session Or Channel has been closed");
}
}
write调用返回的结果与read一样是一个Future<Integer>,而write的CompletionHandler处理的核心逻辑大概是这样:
- @Override
- public void completed(Integer result, AbstractAioSession session) {
- if (log.isDebugEnabled())
- log.debug("Session(" + session.getRemoteSocketAddress()
- + ") writen " + result + " bytes");
-
- WriteMessage writeMessage;
- Queue<WriteMessage> writeQueue = session.getWriteQueue();
- synchronized (writeQueue) {
- writeMessage = writeQueue.peek();
- if (writeMessage.getWriteBuffer() == null
- || !writeMessage.getWriteBuffer().hasRemaining()) {
- writeQueue.remove();
- if (writeMessage.getWriteFuture() != null) {
- writeMessage.getWriteFuture().setResult(Boolean.TRUE);
- }
- try {
- session.getHandler().onMessageSent(session,
- writeMessage.getMessage());
- } catch (Exception e) {
- session.onException(e);
- }
- writeMessage = writeQueue.peek();
- }
- }
- if (writeMessage != null) {
- try {
- session.pendingWrite(writeMessage);
- } catch (IOException e) {
- session.onException(e);
- session.close();
- }
- }
- }
@Override
public void completed(Integer result, AbstractAioSession session) {
if (log.isDebugEnabled())
log.debug("Session(" + session.getRemoteSocketAddress()
+ ") writen " + result + " bytes");
WriteMessage writeMessage;
Queue<WriteMessage> writeQueue = session.getWriteQueue();
synchronized (writeQueue) {
writeMessage = writeQueue.peek();
if (writeMessage.getWriteBuffer() == null
|| !writeMessage.getWriteBuffer().hasRemaining()) {
writeQueue.remove();
if (writeMessage.getWriteFuture() != null) {
writeMessage.getWriteFuture().setResult(Boolean.TRUE);
}
try {
session.getHandler().onMessageSent(session,
writeMessage.getMessage());
} catch (Exception e) {
session.onException(e);
}
writeMessage = writeQueue.peek();
}
}
if (writeMessage != null) {
try {
session.pendingWrite(writeMessage);
} catch (IOException e) {
session.onException(e);
session.close();
}
}
}
compete方法中的result就是实际写入的字节数,然后我们判断消息的缓冲区是否还有剩余,如果没有就将消息从队列中移除,如果队列中还有消息,那么继续发起write调用。
重复一下,这里引用的代码都是yanf4j aio分支中的源码,感兴趣的朋友可以直接check out出来看看: http://yanf4j.googlecode.com/svn/branches/yanf4j-aio。
在引入了aio之后,java对于网络层的支持已经非常完善,该有的都有了,java也已经成为服务器开发的首选语言之一。java的弱项在于对内存的管理上,由于这一切都交给了GC,因此在高性能的网络服务器上还是Cpp的天下。java这种单一堆模型比之erlang的进程内堆模型还是有差距,很难做到高效的垃圾回收和细粒度的内存管理。
这里仅仅是介绍了aio开发的核心流程,对于一个网络框架来说,还需要考虑超时的处理、缓冲buffer的处理、业务层和网络层的切分、可扩展性、性能的可调性以及一定的通用性要求。