-- [I]AttributeMap, ChannelOutboundInvoker, Comparable -- [I]AttributeMap ---- [I]Channel ---- [C]DefaultAttributeMap -- [I]Channel, [C]DefaultAttributeMap ---- [AC]AbstractChannel ------ [AC]AbstractNioChannel -------- [AC]AbstractNioMessageChannel, [AC]AbstractNioByteChannel -- [I]Channel -- [I]Channel ---- [I]ServerChannel ---- [I]DuplexChannel ------ [I](N)ServerSocketChannel ------ [I](N)SocketChannel -- [I](N)ServerSocketChannel, [AC]AbstractNioMessageChannel -- (N)SocketChannel, AbstractNioByteChannel ---- [C]NioServerSocketChannel ---- [C]NioSocketChannel
这个实现仍是蛮严格的:html
protected void doRegister() throws Exception { boolean selected = false; for (;;) { try { selectionKey = javaChannel().register(eventLoop().unwrappedSelector(), 0, this); return; } catch (CancelledKeyException e) { if (!selected) { // Force the Selector to select now as the "canceled" SelectionKey may still be // cached and not removed because no Select.select(..) operation was called yet. eventLoop().selectNow(); selected = true; } else { // We forced a select operation on the selector before but the SelectionKey is still cached // for whatever reason. JDK bug ? throw e; } } } }
注册这个事情看起来不复杂,可是要严格写好写对就要注意。
注册原本只须要用:java
/** * Registers this channel with the given selector, returning a selection * key. * * <p> If this channel is currently registered with the given selector then * the selection key representing that registration is returned. The key's * interest set will have been changed to <tt>ops</tt>, as if by invoking * the {@link SelectionKey#interestOps(int) interestOps(int)} method. If * the <tt>att</tt> argument is not <tt>null</tt> then the key's attachment * will have been set to that value. A {@link CancelledKeyException} will * be thrown if the key has already been cancelled. * * <p> Otherwise this channel has not yet been registered with the given * selector, so it is registered and the resulting new key is returned. * The key's initial interest set will be <tt>ops</tt> and its attachment * will be <tt>att</tt>. * * <p> This method may be invoked at any time. If this method is invoked * while another invocation of this method or of the {@link * #configureBlocking(boolean) configureBlocking} method is in progress * then it will first block until the other operation is complete. This * method will then synchronize on the selector's key set and therefore may * block if invoked concurrently with another registration or selection * operation involving the same selector. </p> * * <p> If this channel is closed while this operation is in progress then * the key returned by this method will have been cancelled and will * therefore be invalid. </p> * * @param sel * The selector with which this channel is to be registered * * @param ops * The interest set for the resulting key * * @param att * The attachment for the resulting key; may be <tt>null</tt> * * @throws ClosedChannelException * If this channel is closed * * @throws ClosedSelectorException * If the selector is closed * * @throws IllegalBlockingModeException * If this channel is in blocking mode * * @throws IllegalSelectorException * If this channel was not created by the same provider * as the given selector * * @throws CancelledKeyException * If this channel is currently registered with the given selector * but the corresponding key has already been cancelled * * @throws IllegalArgumentException * If a bit in the <tt>ops</tt> set does not correspond to an * operation that is supported by this channel, that is, if * {@code set & ~validOps() != 0} * * @return A key representing the registration of this channel with * the given selector */ java.nio.channels.SelectableChannel.register(Selector sel, int ops, Object att) throws ClosedChannelException;
Selector实例,用的是EventLoop的unwrappedSelector实例
int ops 此处目前注册时送的是0(能够送定义的4个以外的?),SelectionKey定义了4个:ios
- OP_READ 1
- OP_WRITE 4
- OP_CONNECT 8
- OP_ACCEPT 16
Object att 表示要塞给Selector实例的附件,此处送的是AbstractNioChannelgit
doRegister异常的处理:github
+ | +----------v----------+ +-----------------------> SelectableChannel | | | .register | | +----------+----------+ | | | | | +----------v----------+ +-----------------+ | | is success +---Y---> return | | +---------------------+ +-----------------+ | N N | | | | +------------------v-----+ +---v----------------+ | | CancelledKeyException | | other exception | | +-----------+------------+ +------+-------------+ | | | | | | +-----------------+ | +--------v--------+ +-------> throw exp | | | selected? | +-------^---------+ | +-------------+---+ | | N | | | | +--------Y--------------------------+ | +---------v---+ | | selectNow | +----+ selected. | +-------------+
有几个注意点sql
JDK SelectionKey定义的4个操做没有0,可是注册时ops时为啥送0 ?能够参见io.netty.channel.socket.nio.AbstractNioChannel doRegister() ?? #1836,有netty做者,和写这个代码的commiter的解释。大体意思就是为了解决潜在的JDK的bug。promise
It's an intentional stuff(有意设计) to deal with a potential(潜在的) JDK bug where it returns a selection key with readyOps set to 0 for no good reason, leading to a spin loop. So, I'd leave it as it is.
// Also check for readOps of 0 to workaround possible JDK bug which may otherwise lead to a spin loop缓存
有个dengyuankai272同窗说到以下,感受解释的更靠谱,做者解释的那个0不是注册的这个地方app
There are two way to register interested events: Channel.register() SelectionKey.interestOps(int) netty use 2. After binding channel to selector with interestOps 0, netty will invoke fireChannelActive() ->fire a read event ->HeadContext.unsafe.doBeginRead() ->SelectionKey.interestOps(int) to set interestOps.
用了eventLoop().cancel(selectionKey());
跟进去逻辑以下:socket
+-------------------+ | SelectionKey | | .cancel | +--------+----------+ | | +--------v----------+ | cancelledKey add 1| +--------+----------+ | | | +----------v-----------+ | cancelledKeys | +----+ >= CLEANUP_INTERVAL+----+ | | (256,hard-code) | | N +----------------------+ Y | | | | +-----v--------+ +-------------v-------+ | return | | needsToSelectAgain| | | | = true | +--------------+ +-----------+---------+ | | // poll task,processSelectedKeysPlain,closeAll等地方触发 +-----------v---------+ | selector | | .selectNow | +---------------------+
+------------------+ | SelectionKey +----N------+ | .isValid | | +------------------+ +------v------+ Y | return | | +-------------+ +--------v---------+ | readPending | | = true | +--------+---------+ | | | | +-----------------v----------------+ | selectionKey.interestOps | | (interestOps | readInterestOp) | +--------------+-------------------+
connectPromise.tryFailure(DO_CLOSE_CLOSED_CHANNEL_EXCEPTION);
connectTimeoutFuture.cancel(false);
将链接的promise和链接超时的future设置成失败和取消。
是将非direct的bug包装转换成direct的buf,好比PooledHeapByteBuf
就不是direct的。
分配direct buf,一是靠ByteBufAllocator.directBuffer,另外一是靠ByteBufUtil.threadLocalDirectBuffer()。
这个仍是有点意思的,还会涉及到 ReferenceCountUtil
,待分析。
newDirectBuffer
会用在AbstractNioByteChannel.filterOutboundMessage(Object)
逻辑中。
在RestExpress项目中,这个方法的实际调用栈示例与简单分析以下:
ServerBootstrapFactory$2.newThread(Runnable) line: 90 // 开始建立worker线程 ThreadPerTaskExecutor.execute(Runnable) line: 33 NioEventLoop(SingleThreadEventExecutor).doStartThread() line: 894 NioEventLoop(SingleThreadEventExecutor).startThread() line: 865 NioEventLoop(SingleThreadEventExecutor).execute(Runnable) line: 758 NioSocketChannel$NioSocketChannelUnsafe(AbstractChannel$AbstractUnsafe).register(EventLoop, ChannelPromise) line: 483 NioEventLoop(SingleThreadEventLoop).register(ChannelPromise) line: 80 NioEventLoop(SingleThreadEventLoop).register(Channel) line: 74 // return register(new DefaultChannelPromise(channel, this)); 拿Channel new了一个DefaultChannelPromise,这个Channel对象就是上面boss线程read出来message Object,也就是NioSocketChannel实例。 NioEventLoopGroup(MultithreadEventLoopGroup).register(Channel) line: 86 ServerBootstrap$ServerBootstrapAcceptor.channelRead(ChannelHandlerContext, Object) line: 255 // 这个是boss线程accept以后准备投递到worker线程的地方 childGroup.register(child).addListener(new ChannelFutureListener() {... DefaultChannelHandlerContext(AbstractChannelHandlerContext).invokeChannelRead(Object) line: 362 AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext, Object) line: 348 DefaultChannelPipeline$HeadContext(AbstractChannelHandlerContext).fireChannelRead(Object) line: 340 DefaultChannelPipeline$HeadContext.channelRead(ChannelHandlerContext, Object) line: 1408 DefaultChannelPipeline$HeadContext(AbstractChannelHandlerContext).invokeChannelRead(Object) line: 362 AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext, Object) line: 348 DefaultChannelPipeline.fireChannelRead(Object) line: 930 AbstractNioMessageChannel$NioMessageUnsafe.read() line: 93 // ** 正在分析的这个方法 NioEventLoop.processSelectedKey(SelectionKey, AbstractNioChannel) line: 677 // 是accept或者read事件 或者是0 触发boss线程上的read动做。注意是 boss线程。在boss线程上accept等同于read。 if ((readyOps & (SelectionKey.OP_READ | SelectionKey.OP_ACCEPT)) != 0 || readyOps == 0) { unsafe.read(); } NioEventLoop.processSelectedKeysOptimized() line: 612 NioEventLoop.processSelectedKeys() line: 529 NioEventLoop.run() line: 491 SingleThreadEventExecutor$5.run() line: 905 Thread.run() line: 748
这个堆栈实际上也解释了,boss线程从接到IO上的accept事件到交给work线程去作后续的读取的整个过程。
更详细的用sql查trace数据:
select * from trace_data where id>=15309 and THREAD_ID=13 and STACK_NUM>5 order by id
逻辑流程:
靠子类的doReadMessages将须要读取的message读取出来放到readBuf (一个List)
迭代读取到的message挨个通知pipeline channelRead事件,pipeline.fireChannelRead
清理readBuf
通知pipeline channelReadComplete事件,pipeline.fireChannelReadComplete
若是异常则关闭 closeOnReadError 并通知pipeline fireExceptionCaught事件
子类实现
此处咱们用到的AbstractNioMessageChannel
有个子类实现是NioServerSocketChannel
doReadMessages实先的内容是: 作accept操做 并建立NioSocketChannel实例做为read到的message,这个Channel实例看上面的的调用栈,会在NioEventLoop(SingleThreadEventLoop).register(Channel) line: 74用到。
看下trace数据:
select * from trace_data where class_name like '%AbstractNioByteChannel%'
数据:
ID THREAD_ID STACK_NUM THREAD_NAME METHOD_ID CLASS_NAME METHOD_NAME LINE_NUM 15336 13 7 boss-0 14143 io/netty/channel/nio/AbstractNioByteChannel <clinit> 45 15350 13 8 boss-0 14123 io/netty/channel/nio/AbstractNioByteChannel <init> 67 15374 13 15 boss-0 14144 io/netty/channel/nio/AbstractNioByteChannel$NioByteUnsafe <init> 97 15407 13 9 boss-0 14186 io/netty/channel/nio/AbstractNioByteChannel$1 <init> 49 15420 13 13 boss-0 14127 io/netty/channel/nio/AbstractNioByteChannel metadata 85 20790 15 5 worker-0 14147 io/netty/channel/nio/AbstractNioByteChannel$NioByteUnsafe read 186 20793 15 6 worker-0 14128 io/netty/channel/nio/AbstractNioByteChannel shouldBreakReadReady 89 26019 15 30 worker-0 14133 io/netty/channel/nio/AbstractNioByteChannel filterOutboundMessage 283 26288 15 30 worker-0 14139 io/netty/channel/nio/AbstractNioByteChannel clearOpWrite 348 26357 15 5 worker-0 14147 io/netty/channel/nio/AbstractNioByteChannel$NioByteUnsafe read 186 26360 15 6 worker-0 14128 io/netty/channel/nio/AbstractNioByteChannel shouldBreakReadReady 89 26757 15 6 worker-0 14145 io/netty/channel/nio/AbstractNioByteChannel$NioByteUnsafe closeOnRead 111 26767 15 7 worker-0 14141 io/netty/channel/nio/AbstractNioByteChannel access$000 43 26768 15 8 worker-0 14129 io/netty/channel/nio/AbstractNioByteChannel isAllowHalfClosure 93
建立是在boss线程,工做是在work线程,主要用到的方法是read。
tarce 数据中的记录:
ID THREAD_ID STACK_NUM THREAD_NAME METHOD_ID CLASS_NAME METHOD_NAME LINE_NUM 20790 15 5 worker-0 14147 io/netty/channel/nio/AbstractNioByteChannel$NioByteUnsafe read 186
从NioByteUnsafe.read到http object解码的调用栈:
HttpRequestDecoder(HttpObjectDecoder).decode(ChannelHandlerContext, ByteBuf, List<Object>) line: 196 HttpRequestDecoder(ByteToMessageDecoder).decodeRemovalReentryProtection(ChannelHandlerContext, ByteBuf, List<Object>) line: 502 HttpRequestDecoder(ByteToMessageDecoder).callDecode(ChannelHandlerContext, ByteBuf, List<Object>) line: 441 HttpRequestDecoder(ByteToMessageDecoder).channelRead(ChannelHandlerContext, Object) line: 278 DefaultChannelHandlerContext(AbstractChannelHandlerContext).invokeChannelRead(Object) line: 362 AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext, Object) line: 348 DefaultChannelHandlerContext(AbstractChannelHandlerContext).fireChannelRead(Object) line: 340 ReadTimeoutHandler(IdleStateHandler).channelRead(ChannelHandlerContext, Object) line: 286 DefaultChannelHandlerContext(AbstractChannelHandlerContext).invokeChannelRead(Object) line: 362 AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext, Object) line: 348 DefaultChannelPipeline$HeadContext(AbstractChannelHandlerContext).fireChannelRead(Object) line: 340 DefaultChannelPipeline$HeadContext.channelRead(ChannelHandlerContext, Object) line: 1408 DefaultChannelPipeline$HeadContext(AbstractChannelHandlerContext).invokeChannelRead(Object) line: 362 AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext, Object) line: 348 DefaultChannelPipeline.fireChannelRead(Object) line: 930 NioSocketChannel$NioSocketChannelUnsafe(AbstractNioByteChannel$NioByteUnsafe).read() line: 163 NioEventLoop.processSelectedKey(SelectionKey, AbstractNioChannel) line: 677 NioEventLoop.processSelectedKeysOptimized() line: 612 NioEventLoop.processSelectedKeys() line: 529 NioEventLoop.run() line: 491 SingleThreadEventExecutor$5.run() line: 905 Thread.run() line: 748
逻辑实现:
shouldBreakReadReady检查, 判断是否应该中断 读ready的处理,针对的状况能够参见其代码判断逻辑
调用子类实现的doReadBytes方法,读取byte的动做
流水线触发读事件 pipeline.fireChannelRead,fireChannelRead接受的参数是Object类型...,因此此处给的ByteBuf和上面NioMessageUnsafe read时给送Channel均可以的。
流水线触发读完成事件 pipeline.fireChannelReadComplete()
DefaultChannelHandlerContext 这个等到后面Handler相关分析的地方再继续分析
nio的select()的时候,只要数据通道容许写,每次select()返回的OP_WRITE都是true。因此在nio的写数据里面,咱们在每次须要写数据以前把数据放到缓冲区,而且注册OP_WRITE,对selector进行wakeup(),这样这一轮select()发现有OP_WRITE以后,将缓冲区数据写入channel,清空缓冲区,而且反注册OP_WRITE,写数据完成。
protected final void clearOpWrite() { 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); // 反注册?? } }
构造函数指定这个类感兴趣的I/O事件是OP_ACCEPT
这个类实现了doBind具体细节
对于不一样的JDK版本用了不一样的bind方法
JDK7以上(含) java.nio.channels.ServerSocketChannel.bind(SocketAddress, int)
JDK7如下 java.net.ServerSocket.bind(SocketAddress, int)
@Override protected void doBind(SocketAddress localAddress) throws Exception { if (PlatformDependent.javaVersion() >= 7) { javaChannel().bind(localAddress, config.getBacklog()); } else { javaChannel().socket().bind(localAddress, config.getBacklog()); } }
直接调用ServerSocketChannel的close方法,没有额外逻辑。
用NIO的ServerSocketChannel实例accept出NIO的SocketChannel实例,再用NIO的SocketChannel实例建立了netty对应的NioSocketChannel实例,并将此实例做为message返回到上下文处理逻辑中。(为了整个封装,messgae不必定指消息报文,此处就是指从ServerSocketChannel-->SocketChannel-->NioSocketChannel)
NioSocketChannel建立实例初始化时又作了很大一堆逻辑处理。
NioServerSocketChannel端不支持doConnect操做,不支持doFinishConnect操做,不支持doDisconnect操做,不支持doWriteMessage操做,不支持filterOutboundMessage操做。
AbstractNioUnsafe的实现逻辑,诸如connect等
io.netty.channel.nio.AbstractNioMessageChannel.doWrite(ChannelOutboundBuffer)实现逻辑待分析,在rest server这个项目中并无用到这个逻辑。
ServerSocketChannel 这个类1.4就有了,可是它的bind方法在1.7才提供,详细参见其注释。可是JDK1.7为啥要新增这个方法呢?为了解决啥问题?// TODO