Netty(四)NIO-优化与源码

Netty优化与源码

1. 优化

1.1 扩展序列化算法

序列化,反序列化主要用于消息正文的转换。
序列化:将java对象转为要传输对象(byte[]或json,最终都是byte[])
反序列化:将正文还原成java对象。

//java自带的序列化
// 反序列化
byte[] body = new byte[bodyLength];
byteByf.readBytes(body);
ObjectInputStream in = new ObjectInputStream(new ByteArrayInputStream(body));
Message message = (Message) in.readObject();
message.setSequenceId(sequenceId);
// 序列化
ByteArrayOutputStream out = new ByteArrayOutputStream();
new ObjectOutputStream(out).writeObject(message);
byte[] bytes = out.toByteArray();

为了支持更多序列化算法,抽象一个 Serializer 接口,提供两个实现,将实现加入了枚举类 Serializer.Algorithm 中:

enum SerializerAlgorithm implements Serializer {
	// Java 实现
    Java {
        @Override
        public <T> T deserialize(Class<T> clazz, byte[] bytes) {
            try {
                ObjectInputStream in = 
                    new ObjectInputStream(new ByteArrayInputStream(bytes));
                Object object = in.readObject();
                return (T) object;
            } catch (IOException | ClassNotFoundException e) {
                throw new RuntimeException("SerializerAlgorithm.Java 反序列化错误", e);
            }
        }

        @Override
        public <T> byte[] serialize(T object) {
            try {
                ByteArrayOutputStream out = new ByteArrayOutputStream();
                new ObjectOutputStream(out).writeObject(object);
                return out.toByteArray();
            } catch (IOException e) {
                throw new RuntimeException("SerializerAlgorithm.Java 序列化错误", e);
            }
        }
    }, 
    // Json 实现(引入了 Gson 依赖)
    Json {
        @Override
        public <T> T deserialize(Class<T> clazz, byte[] bytes) {
            return new Gson().fromJson(new String(bytes, StandardCharsets.UTF_8), clazz);
        }

        @Override
        public <T> byte[] serialize(T object) {
            return new Gson().toJson(object).getBytes(StandardCharsets.UTF_8);
        }
    };

    // 需要从协议的字节中得到是哪种序列化算法
    public static SerializerAlgorithm getByInt(int type) {
        SerializerAlgorithm[] array = SerializerAlgorithm.values();
        if (type < 0 || type > array.length - 1) {
            throw new IllegalArgumentException("超过 SerializerAlgorithm 范围");
        }
        return array[type];
    }
}

增加配置类和配置文件:

public abstract class Config {
    static Properties properties;
    static {
        try (InputStream in = Config.class.getResourceAsStream("/application.properties")) {
            properties = new Properties();
            properties.load(in);
        } catch (IOException e) {
            throw new ExceptionInInitializerError(e);
        }
    }
    public static int getServerPort() {
        String value = properties.getProperty("server.port");
        if(value == null) {
            return 8080;
        } else {
            return Integer.parseInt(value);
        }
    }
    public static Serializer.Algorithm getSerializerAlgorithm() {
        String value = properties.getProperty("serializer.algorithm");
        if(value == null) {
            return Serializer.Algorithm.Java;
        } else {
            return Serializer.Algorithm.valueOf(value);
        }
    }
}

配置文件

serializer.algorithm=Json

修改编解码器

/**
 * 必须和 LengthFieldBasedFrameDecoder 一起使用,确保接到的 ByteBuf 消息是完整的
 */
public class MessageCodecSharable extends MessageToMessageCodec<ByteBuf, Message> {
    @Override
    public void encode(ChannelHandlerContext ctx, Message msg, List<Object> outList) throws Exception {
        // 3. 1 字节的序列化方式 jdk 0 , json 1
        out.writeByte(Config.getSerializerAlgorithm().ordinal());
        byte[] bytes = Config.getSerializerAlgorithm().serialize(msg);
    }

    @Override
    protected void decode(ChannelHandlerContext ctx, ByteBuf in, List<Object> out) throws Exception {
        byte serializerAlgorithm = in.readByte(); // 0 或 1
        // 找到反序列化算法
        Serializer.Algorithm algorithm = Serializer.Algorithm.values()[serializerAlgorithm];
        // 确定具体消息类型
        Class<? extends Message> messageClass = Message.getMessageClass(messageType);
        Message message = algorithm.deserialize(messageClass, bytes);

        out.add(message);
    }
}

1.2 参数调优

CONNECT_TIMEOUT_MILLIS
  • 属于SocketChannel参数,用在客户端建立连接时,如超时则抛出timeout异常
  • SO_TIMEOUT主要用在阻塞IO,阻塞IO中accept,read等都是无限等待的
Bootstrap bootstrap = new Bootstrap().group(group)
                    .option(ChannelOption.CONNECT_TIMEOUT_MILLIS, 5000)
                    .channel(NioServerSocketChannel.class).handler(new LoggingHandler());

附源码部分 io.netty.channel.nio.AbstractNioChannel.AbstractNioUnsafe#connect

@Override
public final void connect(
        final SocketAddress remoteAddress, final SocketAddress localAddress, final ChannelPromise promise) {
    // ...
    // Schedule connect timeout.
    int connectTimeoutMillis = config().getConnectTimeoutMillis();
    if (connectTimeoutMillis > 0) {
        connectTimeoutFuture = eventLoop().schedule(new Runnable() {
            @Override
            public void run() {                
                ChannelPromise connectPromise = AbstractNioChannel.this.connectPromise;
                ConnectTimeoutException cause =
                    new ConnectTimeoutException("connection timed out: " + remoteAddress); // 断点2
                if (connectPromise != null && connectPromise.tryFailure(cause)) {
                    close(voidPromise());
                }
            }
        }, connectTimeoutMillis, TimeUnit.MILLISECONDS);
    }
	// ...
}
SO_BACKLOG
  • 属于ServerSocketChannel参数
    Netty(四)NIO-优化与源码_第1张图片
    sync queue - 半连接队列
    • 大小通过 /proc/sys/net/ipv4/tcp_max_syn_backlog 指定,在 syncookies 启用的情况下,逻辑上没有最大值限制,这个设置便被忽略
      accept queue - 全连接队列
    • 其大小通过 /proc/sys/net/core/somaxconn 指定,在使用 listen 函数时,内核会根据传入的 backlog 参数与系统参数,取二者的较小值
    • 如果 accpet queue 队列满了,server 将发送一个拒绝连接的错误信息到 client
      netty中可通过option(ChannelOption.SO_BACKLOG,值)来设置大小
public class DefaultServerSocketChannelConfig extends DefaultChannelConfig
                                              implements ServerSocketChannelConfig {
    private volatile int backlog = NetUtil.SOMAXCONN;
    // ...默认大小
}
ulimit -n
  • 限制一个进程打开最大文件描述符的数目,属于操作系统参数
TCP_NODELAY
  • nagle算法的延迟,一般设为true不延迟,数据赞属于 SocketChannal 参数
SO_SNDBUF & SO_RECVBUF

滑动接口的参数,现在的操作系统会根据实际情况自动调整。

  • SO_SNDBUF 属于 SocketChannal 参数
  • SO_RCVBUF 既可用于 SocketChannal 参数,也可以用于 ServerSocketChannal 参数(建议设置到 ServerSocketChannal 上)
ALLOCATOR

ByteBuf分配器,属于 SocketChannal 参数,用来分配 ByteBuf, ctx.alloc()。源码详解P128

RCVBUF_ALLOCATOR
  • 属于 SocketChannal 参数,控制 netty 接收缓冲区大小。源码详解:P129
  • 负责入站数据的分配,决定入站缓冲区的大小(并可动态调整),统一采用 direct 直接内存,具体池化还是非池化由 allocator 决定

1.3 RPC 框架

通过反射获取配置

public class ServicesFactory {
    static Properties properties;
    static Map<Class<?>, Object> map = new ConcurrentHashMap<>();

    static {
        try {
            InputStream in = Config.class.getResourceAsStream("/application.properties");
            properties = new Properties();
            properties.load(in);
            Set<String> names = properties.stringPropertyNames();
            for (String name : names) {
                if (name.endsWith("Services")) {
                    Class<?> interfaceClass = Class.forName(name);
                    Class<?> instanceClass = Class.forName(properties.getProperty(name));
                    map.put(interfaceClass, instanceClass.newInstance());
                }
            }
        } catch (IOException | ClassNotFoundException | InstantiationException | IllegalAccessException e) {
            throw new ExceptionInInitializerError(e);
        }
    }
    
    public static <T> T getService(Class<T> interfaceClass) {
        return (T) map.get(interfaceClass);
    }
}

RPC消息处理器

@ChannelHandler.Sharable
public class RpcRequestMessageHandler extends SimpleChannelInboundHandler<RpcRequestMessage> {
    @Override
    protected void channelRead0(ChannelHandlerContext ctx, RpcRequestMessage message) {
        RpcResponseMessage response = new RpcResponseMessage();
        try {
            HelloService service = (HelloService) ServicesFactory.getService(Class.forName(message.getInterfaceName()));
            Method method = service.getClass().getMethod(message.getMethodName(), message.getParameterTypes());
            Object invoke = method.invoke(service, message.getParameterValue());

            response.setReturnValue(invoke);
        } catch (Exception e) {
            e.printStackTrace();
            response.setExceptionValue(e);
        }
        ctx.writeAndFlush(response);
    }

    //本地调试
    public static void main(String[] args) throws ClassNotFoundException, NoSuchMethodException, InvocationTargetException, IllegalAccessException {
        RpcRequestMessage message = new RpcRequestMessage(1,
                "com.aric.server.service.HelloService",
                "sayHello",
                String.class,
                new Class[]{String.class},
                new Object[]{"aric"});
        HelloService service = (HelloService) ServicesFactory.getService(Class.forName(message.getInterfaceName()));
        Method method = service.getClass().getMethod(message.getMethodName(), message.getParameterTypes());
        Object invoke = method.invoke(service, message.getParameterValue());
        System.out.println(invoke);
    }
}

客户端优化,抽取使用代理对象发送消息

/**
 * 使用代理对象替换,主线程发送
 * NioEventLoop线程接收结果,需要线程间通信,使用promise对象接收结果
 * @author
 * @created by xuyu on 2023/9/23-23:10
 */
@Slf4j
public class RpcClientManager {

    public static void main(String[] args) {
        //后期创建代理类优化发送结构
        getChannel().writeAndFlush(new RpcRequestMessage(
                1,
                "com.aric.server.service.HelloService",
                "sayHello",
                String.class,
                new Class[]{String.class},
                new Object[]{"test"}
        ));

        //使用代理发送
        HelloService service = getProxyService(HelloService.class);
        service.sayHello("test");
    }

    //创建代理类
    public static <T> T getProxyService(Class<T> serviceClass) {
        ClassLoader loader = serviceClass.getClassLoader();  //当前类加载器
        Class[] interfaces = new Class[]{serviceClass};//代理类要实现的接口
        //jdk自带的代理
        Object o = Proxy.newProxyInstance(loader, interfaces, (proxy, method, arg) -> {
            //proxy代理对象,method:代理方法,arg:代理参数
            //1.将方法调用转换为消息对象
            RpcRequestMessage message = new RpcRequestMessage(
                    SequenceIdGenerator.nextId(),
                    serviceClass.getName(),
                    method.getName(),
                    method.getReturnType(),
                    method.getParameterTypes(),
                    arg
            );
            //2.将消息对象发送出去
            getChannel().writeAndFlush(message);
            //3.TODO:待优化异步等待返回结果
            return null;
        });
        return (T)o;
    }

    private static Channel channel = null;
    private static final Object LOCK = new Object();

    //单例构造获取唯一channel对象
    public static Channel getChannel() {
        if (channel != null) {
            return channel;
        }
        synchronized (LOCK) {
            if (channel != null) {
                return channel;
            }
            initChannel();
            return channel;
        }
    }

    //初始化channel方法
    private static void initChannel() {
        NioEventLoopGroup group = new NioEventLoopGroup();
        LoggingHandler LOGGING_HANDLER = new LoggingHandler(LogLevel.DEBUG);
        MessageCodecSharable MESSAGE_CODEC = new MessageCodecSharable();
        RpcResponseMessageHandler RPC_HANDLER = new RpcResponseMessageHandler();

        Bootstrap bootstrap = new Bootstrap();
        bootstrap.channel(NioSocketChannel.class);
        bootstrap.group(group);
        bootstrap.handler(new ChannelInitializer<SocketChannel>() {
            @Override
            protected void initChannel(SocketChannel ch) throws Exception {
                ch.pipeline().addLast(new ProtocolFrameDecoder());
                ch.pipeline().addLast(LOGGING_HANDLER);
                ch.pipeline().addLast(MESSAGE_CODEC);
                ch.pipeline().addLast(RPC_HANDLER);
            }
        });
        try {
            channel = bootstrap.connect("localhost", 8080).sync().channel();
            //改为异步
            channel.closeFuture().addListener(future -> {
                group.shutdownGracefully();
            });
        } catch (InterruptedException e) {
            log.debug("client error", e);
        }
    }
}

优化:线程间通信:异步获取返回结果
Netty(四)NIO-优化与源码_第2张图片

//创建代理类
    public static <T> T getProxyService(Class<T> serviceClass) {
        ClassLoader loader = serviceClass.getClassLoader();  //当前类加载器
        Class[] interfaces = new Class[]{serviceClass};//代理类要实现的接口
        //jdk自带的代理
        Object o = Proxy.newProxyInstance(loader, interfaces, (proxy, method, arg) -> {
            //proxy代理对象,method:代理方法,arg:代理参数
            //1.将方法调用转换为消息对象
            int sequenceId = SequenceIdGenerator.nextId();
            RpcRequestMessage message = new RpcRequestMessage(
                    sequenceId,
                    serviceClass.getName(),
                    method.getName(),
                    method.getReturnType(),
                    method.getParameterTypes(),
                    arg
            );
            //2.将消息对象发送出去
            getChannel().writeAndFlush(message);
            //3.返回
            //准备好空的promise对象来接收结果,参数为指定promise对象异步接收结果的线程
            DefaultPromise<Object> promise = new DefaultPromise<>(getChannel().eventLoop());
            RpcResponseMessageHandler.PROMISE.put(sequenceId, promise);
//            promise.addListener(future -> {
//                //创建线程处理任务
//            });
            //原线程等待promise的结果
            promise.await();
            if (promise.isSuccess()) {
                return promise.getNow();
            } else {
                throw new RuntimeException(promise.cause());
            }
        });
        return (T) o;
    }
/**
 * rpc响应消息处理器
 */
@Slf4j
@ChannelHandler.Sharable
public class RpcResponseMessageHandler extends SimpleChannelInboundHandler<RpcResponseMessage> {
    //序号-promise<结果类型>,多个线程访问,用于异步接收rpc调用的返回结果
    public static final Map<Integer, Promise<Object>> PROMISE = new ConcurrentHashMap<>();
    @Override
    protected void channelRead0(ChannelHandlerContext ctx, RpcResponseMessage msg) throws Exception {
        //拿到空的promise
        Promise<Object> promise = PROMISE.remove(msg.getSequenceId());  //返回并移除
        if (promise != null) {
            Object returnValue = msg.getReturnValue();
            Exception exceptionValue = msg.getExceptionValue();
            if (exceptionValue != null) {
                promise.setFailure(exceptionValue);
            } else {
                promise.setSuccess(returnValue);
            }
        }
        System.out.println(msg);
    }

代码:https://gitee.com/xuyu294636185/netty-demo.git

2. 源码

2.1 netty启动剖析

        //1. netty中使用EventLoopGroup(Nio boss线程),来封装线程和selector
        Selector selector = Selector.open();
        //创建NioServerSocketChannel,同时初始化它关联的handler,以及为原生ssc存储config
        NioServerSocketChannel attachment = new NioServerSocketChannel();
        ServerSocketChannel ssc = ServerSocketChannel.open();
        ssc.configureBlocking(false);
        //2.启动nio boss线程执行
        //建立selector和channel的注册,sscKey是事件的句柄,是将来事件发生后,通过它可以知道事件和哪个channel的事件
        SelectionKey sscKey = ssc.register(selector, 0, attachment);
        ssc.bind(new InetSocketAddress(8080));
        //表示sscKey只关注accept事件
        sscKey.interestOps(SelectionKey.OP_ACCEPT);
启动流程

Netty(四)NIO-优化与源码_第3张图片

EventLoop

EventLoop重要组成:selector,线程,任务队列
EventLoop既会处理io事件,也会处理普通任务和定时任务

  1. selector何时创建?
    在构造方法创建时通过SelectorProvider.openSelector();
  2. eventloop为什么会有两个selector成员?
    为了在遍历selectedKey时提高性能。
    一个是原始的unwrappedselector(底层是hashset实现),一个是包装后的selector(底层是数组实现)
  3. eventLoop的nio线程在何时启动?
    在首次调用exectue方法时executor中将当前线程赋给nio线程,并通过state状态控制位只会启动一次
  4. 提交普通任务会不会结束select阻塞?

    int selectedKeys = selector.select(timeoutMillis);
    protected void wakeup(boolean inEventLoop) {
    if(!inEventLoop && wakeUp.compareAndSet(false,true)) {
    selector.wakeup();
    }
    }
  5. wakeup方法理解
    inEventLoop:用于判断当前wakeup线程是否和nio线程是否相同,不同才能进入。
    wakeUp:原子Boolean变量,如果有多个线程来提交任务,为了避免wakeup被频繁调用。只有一个成功。
  6. 每次循环时,什么时候会进入SelectStrategy.SELECT分支?
    public void run(){
    for(; {
    switch(selectStrategy.calculateStrategy(selectNowSupplier, hasTask())) {
    case SelectStrategy.CONTINUE:
    continue;
    case SelectStrategy.BUSY_WAIT:
    case SelectStrategy.SELECT:
    select(wakeUp.getAndSet(false));
    if(wakeUp.get()) {…}
    default:
    }
    }
    }
    public int calculateStrategy(IntSupplier supplier,boolean hasTasks) {
    return hasTasks ? suppplier.get() : SelectStrategy.SELECT;
    }
    没有任务时,才会进入SELECT。
    当有任务时,会调用SelectNow方法,顺便拿到io事件。
  7. 何时会select阻塞,阻塞多久?
    long currentTimeNanos = System.nanoTime();
    long selectDeadLineNanos = currentTimeNanos + delayNanos(cuurrentTimeNanos);
    for(;{
    long timeoutMillis = (selectDeadLineNanos - currentTimeNanos + 500000L) / 1000000L;
    int selectedKeys = selector.select(timeoutMillis);
    }
    没有定时任务的情况
    selectDeadLineNanos:截至时间 = 当前时间 + 1s
    timeoutMillis:超时时间 = 1s + 0.5ms
  8. nio空轮询bug在哪体现,如何解决?
    for(;{
    long timeoutMillis = (selectDeadLineNanos - currentTimeNanos + 500000L) / 1000000L;
    int selectedKeys = selector.select(timeoutMillis);
    }
    在select中没有阻塞,一直在死循环
    解决:引入selectCnt,每循环一次++,当超过设置的阈值(默认512),selectRebuildSelector(selectCnt)重新创建一个selector,替换旧的。
  9. ioRatio控制什么,设置为100有何作用?
    if(ioRatio == 100) {
    processSelectedKeys(); //处理所有ioio事件
    runAllTasks();
    } else {
    long ioStartTime = System.nanoTime();
    processSelectedKeys();
    long ioTime = System.nanoTime() - ioStartTime;
    runAllTasks(ioTime * (100 - ioRatio) / ioRatio); //避免普通事件执行时间太长
    }
    ioRatio控制处理io事件所占用的事件比例50%,ioTime代表执行io事件处理耗时。
  10. selectedKeys优化,在哪区分不同事件类型。
    selectedKeys由hashset集合替换为数组实现。
    private void processSelectedKeys() [
    if(selectedKeys != null) {
    processSelectedKeysOptimized(); //优化后的
    } else {
    processSelectedKeysPlain(selector.selectedKeys()); //原始的
    }
    }
    private void processSelectedKeysOptimized() {
    for(int i = 0;i < selectedKeys.size; ++i) {
    SelectionKey k = selectedKeys.keys[i];
    selectedKeys.keys[i] = null;
    Objected a = k.attachment();
    if(a instanceof AbstractNioChannel) {
    processSelectedKey(k, (AbstractNioChannel) a); //处理具体的事件类型
    }
    }
    }
accept流程
  1. selector.select()阻塞直到事件发生
  2. 遍历处理selectedKeys
  3. 拿到一个key,判断事件类型是否为accpet
  4. 创建socketChannel,设置非阻塞
  5. 将socketChannel注册到selector
  6. 关注selectionKey的read事件。
read流程
  1. selector.select()阻塞直到事件发生
  2. 遍历处理selectedKeys
  3. 拿到一个key,判断事件类型是否为read
  4. 读取操作

你可能感兴趣的:(Netty,nio)