【源起Netty 正传】Netty Channel

Channel定位

注意:如无特别说明,文中的Channel都指的是Netty Channel(io.netty.channel)

一周时间的Channel家族学习,一度让我怀疑人生——研究这个方法有没有用?学习Netty是不是有点儿下了高速走乡间小路的意思?我为啥要读源码?
之所以产生这些疑问,除了我本身心理活动丰富以外,主要病因在于没搞清楚Channel在Netty体系中的定位。而没能清晰理解Netty的定位,也默默的送出了一记助攻。

作些本质思考:Netty是一个NIO框架,是一个嫁接在java NIO基础上的框架

宏观上可以这么理解,见下图:

【源起Netty 正传】Netty Channel

先不急着聊Channel,回顾下IO演进过程,重点关注IO框架的结构变化。搞懂了这部分后,我们将明白Channel在IO世界中扮演的角色!

进击的IO

BIO

【源起Netty 正传】Netty Channel

此图展示的已经算是优化后的BIO了——用到了线程池。显然,每一个client都需要server端付出一个Thread的代价,即使你通过线程池做了优化,由于受到线程个数的制约,激增的客户端依旧表现的“欲求不满”。

NIO

【源起Netty 正传】Netty Channel

  • Acceptor注册Selector,监听accept事件
  • 当客户端连接后,触发accept事件
  • 服务器构建对应的Channel,并在其上注册Selector,监听读写事件
  • 当发生读写事件后,进行相应的读写处理

Reactor单线程

【源起Netty 正传】Netty Channel

与NIO模型相似,当然也就有和NIO同样的问题:selector/reactor单个线程处理多个channel的各种操作,如果其中一个channel的事件处理延缓了,将影响其它channel。

Reactor多线程

将read/write这种io处理操作分隔出来,非io型操作(业务操作)配备以线程池,进化成reactor多线程模型:

【源起Netty 正传】Netty Channel

这样的架构,系统瓶颈转移至Reactor部分。而目前劳苦功高的Reactor作了两件事:
1.接收客户端链接请求
2.处理IO型读写操作

主从Reactor

将接收client链接的功能再次拆分出来:

【源起Netty 正传】Netty Channel

Netty恰恰就是主从Reactor模型的实践者,想想服务端创建时的代码:

EventLoopGroup bossGroup = new NioEventLoopGroup(1);
EventLoopGroup workerGroup = new NioEventLoopGroup();

ServerBootstrap b = new ServerBootstrap();
b.group(bossGroup, workerGroup)

...

从nio时代的模型图上开始出现channel(java channel),它的定位就是进行诸如connect、write、read、close等底层交互。概括一下,java channel是上承selector下连socket的存在。而netty channel,则把java channel当作了底层

源码分析

类结构

清楚了Channel的定位,接下来对其常用api进行分析。

首先拍出类图:
【源起Netty 正传】Netty Channel

其实Channel内部还有一套体系,Unsafe家族:
【源起Netty 正传】Netty Channel

Unsafe是Channel的内置类(接口),与java channel交互的重任最终会落到Unsafe身上。

write方法

write只是将数据写入到了ChannelOutboundBuffer中,并没有真正的发送出去,到flush方法调用时,才写入到java channel中发送给对方。

下面列出AbstractChannel的write方法,值得关注的地方已打上中文注释:

@Override
public final void write(Object msg, ChannelPromise promise) {
    assertEventLoop();

    ChannelOutboundBuffer outboundBuffer = this.outboundBuffer;
    if (outboundBuffer == null) {
        // If the outboundBuffer is null we know the channel was closed and so
        // need to fail the future right away. If it is not null the handling of the rest
        // will be done in flush0()
        // See https://github.com/netty/netty/issues/2362
        safeSetFailure(promise, WRITE_CLOSED_CHANNEL_EXCEPTION);
        // release message now to prevent resource-leak
        ReferenceCountUtil.release(msg);
        return;
    }

    int size;
    try {
        msg = filterOutboundMessage(msg);   //作消息的包装,转换成ByteBuf等
        size = pipeline.estimatorHandle().size(msg);
        if (size < 0) {
            size = 0;
        }
    } catch (Throwable t) {
        safeSetFailure(promise, t);
        ReferenceCountUtil.release(msg);
        return;
    }

    outboundBuffer.addMessage(msg, size, promise);    //msg消息写入ChannelOutboundBuffer
}

上述代码最后一行,msg写入了ChannelOutboundBuffer的尾节点tailEntry,同时将unflushedEntry赋值暂存。代码展开如下:

public void addMessage(Object msg, int size, ChannelPromise promise) {
    Entry entry = Entry.newInstance(msg, size, total(msg), promise);
    if (tailEntry == null) {
        flushedEntry = null;
        tailEntry = entry;
    } else {
        Entry tail = tailEntry;
        tail.next = entry;
        tailEntry = entry;
    }
    if (unflushedEntry == null) {    //注释一、标记成“未刷新”的数据
        unflushedEntry = entry;
    }

    incrementPendingOutboundBytes(entry.pendingSize, false);
}

ChannelOutboundBuffer类

这里对ChannelOutboundBuffer类进行简单说明,按惯例先看类注释。

/**
 * (Transport implementors only) an internal data structure used by {@link AbstractChannel} to store its pending
 * outbound write requests.
 *
 *  省略...
 */

前文提到过,write方法将消息写到ChannelOutboundBuffer,算是数据暂存;之后的flush再将消息刷到java channel乃至客户端。

来张示意图,方便理解:
【源起Netty 正传】Netty Channel

图中列出的三个属性,在write->ChannelOutboundBuffer->flush的数据流转过程中比较关键。Entry是啥?ChannelOutboundBuffer的静态内部类,典型的链表结构数据:

static final class Entry {

        Entry next;
        
        // 省略...
}

write方法的最后部分(注释一位置)调用outboundBuffer.addMessage(msg, size, promise),已将封装msg的Entry赋值给tailEntry和unflushedEntry;而flush方法,通过调用outboundBuffer.addFlush()(下文,注释二位置),将unflushedEntry间接赋值给了flushedEntry

public void addFlush() {
    Entry entry = unflushedEntry;
    if (entry != null) {
        if (flushedEntry == null) {
            // there is no flushedEntry yet, so start with the entry
            flushedEntry = entry;
        }
        do {
            flushed ++;
            if (!entry.promise.setUncancellable()) {
                // Was cancelled so make sure we free up memory and notify about the freed bytes
                int pending = entry.cancel();
                decrementPendingOutboundBytes(pending, false, true);
            }
            entry = entry.next;
        } while (entry != null);

        // All flushed so reset unflushedEntry
        unflushedEntry = null;
    }
}

flush方法

直接从AbstractChannel的flush方法开始(若以Channel的flush为开端会经pipeline,将有很长调用链,省略):

public final void flush() {
    assertEventLoop();

    ChannelOutboundBuffer outboundBuffer = this.outboundBuffer;
    if (outboundBuffer == null) {
        return;
    }

    outboundBuffer.addFlush();    //注释二、标记成“已刷新”数据
    flush0();    //数据处理
}

outboundBuffer.addFlush()方法已经分析过了,跟踪调用链flush0->doWrite,我们看下AbstractNioByteChanneldoWrite方法:

@Override
protected void doWrite(ChannelOutboundBuffer in) throws Exception {
    int writeSpinCount = config().getWriteSpinCount();    //自旋计数,限制循环次数,默认16
    do {
        Object msg = in.current();    //flushedEntry的msg
        if (msg == null) {
            // Wrote all messages.
            clearOpWrite();
            // Directly return here so incompleteWrite(...) is not called.
            return;
        }
        writeSpinCount -= doWriteInternal(in, msg);
    } while (writeSpinCount > 0);

    incompleteWrite(writeSpinCount < 0);
}

writeSpinCount是个自旋计数,类似于自旋锁的设定,防止当前IO线程由于网络等原因无尽执行写操作,而使得线程假死,造成资源浪费

观察doWriteInternal方法,关键处依旧中文注释伺候:

private int doWriteInternal(ChannelOutboundBuffer in, Object msg) throws Exception {
    if (msg instanceof ByteBuf) {
        ByteBuf buf = (ByteBuf) msg;
        if (!buf.isReadable()) {    //writerIndex - readerIndex >0 ? true: flase
            in.remove();
            return 0;
        }

        final int localFlushedAmount = doWriteBytes(buf);   //返回实际写入到java channel的字节数
        if (localFlushedAmount > 0) {   //写入成功
            in.progress(localFlushedAmount);
            /**
             * 1.已经全部写完,执行in.remove()
             * 2.“写半包”场景,直接返回1。
             *   外层方法的自旋变量writeSpinCount递减成15,轮询再次执行本方法
             */
            if (!buf.isReadable()) {
                in.remove();
            }
            return 1;
        }
    } else if (msg instanceof FileRegion) {
    
        //“文件型”消息处理逻辑省略..
        
    } else {
        // Should not reach here.
        throw new Error();
    }
    return WRITE_STATUS_SNDBUF_FULL;    //发送缓冲区满,值=Integer.MAX_VALUE
}

回到doWrite方法,最后执行了incompleteWrite(writeSpinCount < 0)

protected final void incompleteWrite(boolean setOpWrite) {
    // Did not write completely.
    if (setOpWrite) {
        setOpWrite();
    } else {
        // Schedule flush again later so other tasks can be picked up in the meantime
        Runnable flushTask = this.flushTask;
        if (flushTask == null) {
            flushTask = this.flushTask = new Runnable() {
                @Override
                public void run() {
                    flush();
                }
            };
        }
        eventLoop().execute(flushTask);
    }
}

这里的设定挺有意思:

  • 如果 setOpWrite = writeSpinCount < 0 = true,即 doWriteInternal方法返回值 = WRITE_STATUS_SNDBUF_FULL(发送缓冲区满)时,设置写操作位:
protected final void setOpWrite() {
    final SelectionKey key = selectionKey();
    // Check first if the key is still valid as it may be canceled as part of the deregistration
    // from the EventLoop
    // See https://github.com/netty/netty/issues/2104
    if (!key.isValid()) {
        return;
    }
    final int interestOps = key.interestOps();
    if ((interestOps & SelectionKey.OP_WRITE) == 0) {
        key.interestOps(interestOps | SelectionKey.OP_WRITE);
    }
}

其实就是设置SelectionKey的OP_WRITE操作位,在selector/reactor下次轮询的时候,将再次执行写操作

  • 如果 setOpWrite = writeSpinCount < 0 = false,即 doWriteInternal方法返回值 = 1,16次写半包仍旧没将消息发送出去,则通过定时器再次执行flush:
public Channel flush() {
    pipeline.flush();
    return this;
}

结论:前者由于发送缓冲区满,已无法写入数据,于是继希望于selector的下次轮询;后者则可能只是因为自旋次数少,引起的数据发送不完全,直接将任务再次放入pipeline,而无需等待selector。
这无疑是种优化,细节之处,功力尽显!

感谢

相关推荐