【4】Netty4源码分析-NioEventLoop实现的线程运行逻辑

转自 http://xw-z1985.iteye.com/blog/1928244

在netty服务端启动源码分析-线程创建一文中已分析SingleThreadEventExecutor所持有的线程的运行逻辑由NioEventLoop实现，那么本文就着手分析NioEventLoop实现的线程运行逻辑：

// NioEventLoop  
protected void run() {  
        for (;;) {  
            oldWakenUp = wakenUp.getAndSet(false);  
            try {  
                if (hasTasks()) {  
                    selectNow();  
                } else {  
                    select();  
                    if (wakenUp.get()) {  
                        selector.wakeup();  
                    }  
                }  
  
                cancelledKeys = 0;  
  
                final long ioStartTime = System.nanoTime();  
                needsToSelectAgain = false;  
                if (selectedKeys != null) {  
                    processSelectedKeysOptimized(selectedKeys.flip());  
                } else {  
                    processSelectedKeysPlain(selector.selectedKeys());  
                }  
                final long ioTime = System.nanoTime() - ioStartTime;  
  
                final int ioRatio = this.ioRatio;  
                runAllTasks(ioTime * (100 - ioRatio) / ioRatio);  
  
                if (isShuttingDown()) {  
                    closeAll();  
                    if (confirmShutdown()) {  
                        break;  
                    }  
                }  
            } catch (Throwable t) {  
                logger.warn("Unexpected exception in the selector loop.", t);  
  
                // Prevent possible consecutive immediate failures that lead to  
                // excessive CPU consumption.  
                try {  
                    Thread.sleep(1000);  
                } catch (InterruptedException e) {  
                    // Ignore.  
                }  
            }  
        }  
    }  

分析如下：

1. ioEventLoop执行的任务分为两大类：IO任务和非IO任务。IO任务即selectionKey中ready的事件，譬如accept、connect、read、write等；非IO任务则为添加到taskQueue中的任务，譬如之前文章中分析到的register0、bind、channelActive等任务
2. 两类任务的执行先后顺序为：IO任务->非IO任务。IO任务由processSelectedKeysOptimized(selectedKeys.flip())或processSelectedKeysPlain(selector.selectedKeys())触发；非IO任务由runAllTasks(ioTime * (100 - ioRatio) / ioRatio)触发
3. 两类任务的执行时间比由变量ioRatio控制，譬如：ioRatio=50（该值为默认值），则表示允许非IO任务执行的时间与IO任务的执行时间相等
4. 执行IO任务前，需要先进行select，以判断之前注册过的channel是否已经有感兴趣的事件ready
5. 如果任务队列中存在非IO任务，则执行非阻塞的selectNow()方法

// NioEventLoop  
  void selectNow() throws IOException {  
        try {  
            selector.selectNow();  
        } finally {  
            // restore wakup state if needed  
            if (wakenUp.get()) {  
                selector.wakeup();  
            }  
        }  
    }  

否则，执行阻塞的select()方法

// NioEventLoop  
   private void select() throws IOException {  
        Selector selector = this.selector;  
        try {  
            int selectCnt = 0;  
            long currentTimeNanos = System.nanoTime();  
            long selectDeadLineNanos = currentTimeNanos + delayNanos(currentTimeNanos);  
            for (;;) {  
                long timeoutMillis = (selectDeadLineNanos - currentTimeNanos + 500000L) / 1000000L;  
                if (timeoutMillis <= 0) {  
                    if (selectCnt == 0) {  
                        selector.selectNow();  
                        selectCnt = 1;  
                    }  
                    break;  
                }  
                int selectedKeys = selector.select(timeoutMillis);  
                selectCnt ++;  
                if (selectedKeys != 0 || oldWakenUp || wakenUp.get() || hasTasks()) {  
                    // Selected something,  
                    // waken up by user, or  
                    // the task queue has a pending task.  
                    break;  
                }  
  
                if (SELECTOR_AUTO_REBUILD_THRESHOLD > 0 &&  
                        selectCnt >= SELECTOR_AUTO_REBUILD_THRESHOLD) {  
                    // The selector returned prematurely many times in a row.  
                    // Rebuild the selector to work around the problem.  
                    logger.warn(  
                            "Selector.select() returned prematurely {} times in a row; rebuilding selector.",  
                            selectCnt);  
                    rebuildSelector();  
                    selector = this.selector;  
                    // Select again to populate selectedKeys.  
                    selector.selectNow();  
                    selectCnt = 1;  
                    break;  
                }  
                currentTimeNanos = System.nanoTime();  
            }  
            if (selectCnt > MIN_PREMATURE_SELECTOR_RETURNS) {  
                if (logger.isDebugEnabled()) {  
                    logger.debug("Selector.select() returned prematurely {} times in a row.", selectCnt - 1);  
                }  
            }  
        } catch (CancelledKeyException e) {  
            if (logger.isDebugEnabled()) {  
                logger.debug(CancelledKeyException.class.getSimpleName() + " raised by a Selector - JDK bug?", e);  
            }  
            // Harmless exception - log anyway  
        }  
    }  

下面分析阻塞的select方法：

• 首先执行delayNanos(currentTimeNanos)：计算延迟任务队列中第一个任务的到期执行时间（即最晚还能延迟执行的时间）.注意：（每个SingleThreadEventExecutor都持有一个延迟执行任务的优先队列：final Queue<ScheduledFutureTask<?>> delayedTaskQueue = new PriorityQueue

//SingleThreadEventExecutor  
protected long delayNanos(long currentTimeNanos) {  
        ScheduledFutureTask<?> delayedTask = delayedTaskQueue.peek();  
        if (delayedTask == null) {  
            return SCHEDULE_PURGE_INTERVAL;  
        }  
  
        return delayedTask.delayNanos(currentTimeNanos);  
}  
  
//ScheduledFutureTask  
public long delayNanos(long currentTimeNanos) {  
        return Math.max(0, deadlineNanos() - (currentTimeNanos - START_TIME));  
    }  
public long deadlineNanos() {  
        return deadlineNanos;  
    }  

• 如果当前时间已经超过到期执行时间后的500000纳秒（这个数字是如何定的？），则说明被延迟执行的任务不能再延迟了：如果在进入这个方法后还没有执行过selectNow方法（由标记selectCnt是否为0来判断），则先执行非阻塞的selectNow方法，然后立即返回；否则，立即返回
• 如果当前时间没有超过到期执行时间后的500000L纳秒，则说明被延迟执行的任务还可以再延迟，所以可以让select的阻塞时间长一点（说不定多出的这点时间就能select到一个ready的IO任务），故执行阻塞的selector.select(timeoutMillis)方法
• 如果已经存在ready的selectionKey，或者该selector被唤醒，或者此时非IO任务队列加入了新的任务，则立即返回
• 否则，本次执行selector.select(timeoutMillis)方法后的结果selectedKeys肯定为0，如果连续返回0的select次数还没有超过SELECTOR_AUTO_REBUILD_THRESHOLD（默认值为512），则继续下一次for循环。注意，根据以下算法：long timeoutMillis = (selectDeadLineNanos - currentTimeNanos + 500000L) / 1000000L。随着currentTimeNanos的增大，在进入第二次for循环时，正常情况下（即：在没有selectionKey已ready的情况下，selector.select(timeoutMillis)确实阻塞了timeoutMillis毫秒才返回0）计算出的timeoutMillis肯定小于0，计算如下：

假设第一次和第二次进入for循环时的当前时间分currentTimeNanos1，currentTimeNanos2，由于在第一次循环中select阻塞了timeoutMillis1毫秒，所以currentTimeNanons2纳秒 > currentTimeNanos1纳秒+timeoutMillis1毫秒. 那么，第二次的timeoutMillis2 = (selectDeadLineNanos – currentTimeNanos2 + 500000) / 1000000 < (selectDeadLineNanos – (currentTimeNanos1+timeoutMillis1*1000000)+ 500000) / 1000000 =
timeoutMillis1- timeoutMillis1=0

即：timeoutMillis2 < 0。因此第二次不会再进行select，直接跳出循环并返回

• 否则，如果连续多次返回0，说明每次调用selector.select(timeoutMillis)后根本就没有阻塞timeoutMillis时间，而是立即就返回了，且结果为0. 这说明触发了epool cpu100%的bug（https://github.com/netty/netty/issues/327）。解决方案就是对selector重新rebuild

public void rebuildSelector() {  
        if (!inEventLoop()) {  
            execute(new Runnable() {  
                @Override  
                public void run() {  
                    rebuildSelector();  
                }  
            });  
            return;  
        }  
        final Selector oldSelector = selector;  
        final Selector newSelector;  
        if (oldSelector == null) {  
            return;  
        }  
        try {  
            newSelector = openSelector();  
        } catch (Exception e) {  
            logger.warn("Failed to create a new Selector.", e);  
            return;  
        }  
        // Register all channels to the new Selector.  
        int nChannels = 0;  
        for (;;) {  
            try {  
                for (SelectionKey key: oldSelector.keys()) {  
                    Object a = key.attachment();  
                    try {  
                        if (key.channel().keyFor(newSelector) != null) {  
                            continue;  
                        }  
                        int interestOps = key.interestOps();  
                        key.cancel();  
                        key.channel().register(newSelector, interestOps, a);  
                        nChannels ++;  
                    } catch (Exception e) {  
                        logger.warn("Failed to re-register a Channel to the new Selector.", e);  
                        if (a instanceof AbstractNioChannel) {  
                            AbstractNioChannel ch = (AbstractNioChannel) a;  
                            ch.unsafe().close(ch.unsafe().voidPromise());  
                        } else {  
                            @SuppressWarnings("unchecked")  
                            NioTask<SelectableChannel> task = (NioTask<SelectableChannel>) a;  
                            invokeChannelUnregistered(task, key, e);  
                        }  
                    }  
                }  
            } catch (ConcurrentModificationException e) {  
                // Probably due to concurrent modification of the key set.  
                continue;  
            }  
  
            break;  
        }  
  
        selector = newSelector;  
  
        try {  
            // time to close the old selector as everything else is registered to the new one  
            oldSelector.close();  
        } catch (Throwable t) {  
            if (logger.isWarnEnabled()) {  
                logger.warn("Failed to close the old Selector.", t);  
            }  
        }  
  
        logger.info("Migrated " + nChannels + " channel(s) to the new Selector.");  
    }  

Rebuild的本质：其实就是重新创建一个selector，然后将原来的那个selector中已注册的所有channel重新注册到新的selector中，并将老的selectionKey全部cancel掉，最后将的selector关闭。对selector进行rebuild之后，还需要重新调用selectNow方法，检查是否有已ready的selectionKey.

6. 执行select（）或者selectNow()后，如果已经有已ready的selectionKey，则开始执行IO操。processSelectedKeysOptimized和processSelectedKeysPlain的执行逻辑是很相似的

// NioEventLoop  
  private void processSelectedKeysOptimized(SelectionKey[] selectedKeys) {  
        for (int i = 0;; i ++) {  
            final SelectionKey k = selectedKeys[i];  
            if (k == null) {  
                break;  
            }  
  
            final Object a = k.attachment();  
  
            if (a instanceof AbstractNioChannel) {  
                processSelectedKey(k, (AbstractNioChannel) a);  
            } else {  
                @SuppressWarnings("unchecked")  
                NioTask<SelectableChannel> task = (NioTask<SelectableChannel>) a;  
                processSelectedKey(k, task);  
            }  
  
            if (needsToSelectAgain) {  
                selectAgain();  
                // Need to flip the optimized selectedKeys to get the right reference to the array  
                // and reset the index to -1 which will then set to 0 on the for loop  
                // to start over again.  
                //  
                // See https://github.com/netty/netty/issues/1523  
                selectedKeys = this.selectedKeys.flip();  
                i = -1;  
            }  
        }  
    }  

private void processSelectedKeysPlain(Set<SelectionKey> selectedKeys) {  
        // check if the set is empty and if so just return to not create garbage by  
        // creating a new Iterator every time even if there is nothing to process.  
        // See https://github.com/netty/netty/issues/597  
        if (selectedKeys.isEmpty()) {  
            return;  
        }  
  
        Iterator<SelectionKey> i = selectedKeys.iterator();  
        for (;;) {  
            final SelectionKey k = i.next();  
            final Object a = k.attachment();  
            i.remove();  
  
            if (a instanceof AbstractNioChannel) {  
                processSelectedKey(k, (AbstractNioChannel) a);  
            } else {  
                @SuppressWarnings("unchecked")  
                NioTask<SelectableChannel> task = (NioTask<SelectableChannel>) a;  
                processSelectedKey(k, task);  
            }  
  
            if (!i.hasNext()) {  
                break;  
            }  
  
            if (needsToSelectAgain) {  
                selectAgain();  
                selectedKeys = selector.selectedKeys();  
  
                // Create the iterator again to avoid ConcurrentModificationException  
                if (selectedKeys.isEmpty()) {  
                    break;  
                } else {  
                    i = selectedKeys.iterator();  
                }  
            }  
        }  
    }  

此处仅分析processSelectedKeysOptimized方法，对于这两个方法的区别暂时放下，后续再分析吧。processSelectedKeysOptimized的执行逻辑基本上就是循环处理每个select出来的selectionKey，每个selectionKey的处理首先根据attachment的类型来进行分发处理发：如果类型为AbstractNioChannel，则执行一种逻辑；其他，则执行另外一种逻辑。此处，本文仅分析类型为AbstractNioChannel的处理逻辑，另一种逻辑的分析暂时放下，后续再分析。
在判断attachment的类型前，首先需要弄清楚这个attatchment是何时关联到selectionKey上的？还记得socket一文中分析的register0任务吗？ AbstractNioChannel类中有如下代码：

selectionKey = javaChannel().register(eventLoop().selector, 0, this);

此处将this（即AbstractNioChannel）作为attachment关联到selectionKey

现在开始分析类型为AbstractNioChannel的处理逻辑，首先看processSelectedKey(k, (AbstractNioChannel) a)的实现：

//NioEventLoop  
private static void processSelectedKey(SelectionKey k, AbstractNioChannel ch) {  
        final NioUnsafe unsafe = ch.unsafe();  
        if (!k.isValid()) {  
            // close the channel if the key is not valid anymore  
            unsafe.close(unsafe.voidPromise());  
            return;  
        }  
  
        int readyOps = -1;  
        try {  
            readyOps = k.readyOps();  
            if ((readyOps & (SelectionKey.OP_READ | SelectionKey.OP_ACCEPT)) != 0 || readyOps == 0) {  
                unsafe.read();  
                if (!ch.isOpen()) {  
                    // Connection already closed - no need to handle write.  
                    return;  
                }  
            }  
            if ((readyOps & SelectionKey.OP_WRITE) != 0) {  
                processWritable(ch);  
            }  
            if ((readyOps & SelectionKey.OP_CONNECT) != 0) {  
                // remove OP_CONNECT as otherwise Selector.select(..) will always return without blocking  
                // See https://github.com/netty/netty/issues/924  
                int ops = k.interestOps();  
                ops &= ~SelectionKey.OP_CONNECT;  
                k.interestOps(ops);  
  
                unsafe.finishConnect();  
            }  
        } catch (CancelledKeyException e) {  
            if (readyOps != -1 && (readyOps & SelectionKey.OP_WRITE) != 0) {  
                unregisterWritableTasks(ch);  
            }  
            unsafe.close(unsafe.voidPromise());  
        }  
    }  

终于见到熟悉nio处理代码了，它根据selecionKey的readyOps的值进行分发，下一篇文章将分析readyOps为accept时的处理逻辑。关于final NioUnsafe unsafe = ch.unsafe()，还记得socket一文中分析的：NioUnsafe由AbstractChannel的子类AbstractNioMessageChannel实例化，其类型为NioMessageUnsafe，它里面定义了read方法，即readyOps为accept的处理逻辑。
7. 执行完io任务后，接着执行非IO任务：runAllTasks(ioTime * (100 - ioRatio) / ioRatio)

//NioEventLoop  
protected boolean runAllTasks(long timeoutNanos) {  
        fetchFromDelayedQueue();  
        Runnable task = pollTask();  
        if (task == null) {  
            return false;  
        }  
        final long deadline = ScheduledFutureTask.nanoTime() + timeoutNanos;  
        long runTasks = 0;  
        long lastExecutionTime;  
        for (;;) {  
            try {  
                task.run();  
            } catch (Throwable t) {  
                logger.warn("A task raised an exception.", t);  
            }  
            runTasks ++;  
            // Check timeout every 64 tasks because nanoTime() is relatively expensive.  
            // XXX: Hard-coded value - will make it configurable if it is really a problem.  
            if ((runTasks & 0x3F) == 0) {  
                lastExecutionTime = ScheduledFutureTask.nanoTime();  
                if (lastExecutionTime >= deadline) {  
                    break;  
                }  
            }  
  
            task = pollTask();  
            if (task == null) {  
                lastExecutionTime = ScheduledFutureTask.nanoTime();  
                break;  
            }  
        }  
  
        this.lastExecutionTime = lastExecutionTime;  
        return true;  
    }  

首先分析fetchFromDelayedQueue()方法，由父类SingleThreadEventExecutor实现

// SingleThreadEventExecutor  
private void fetchFromDelayedQueue() {  
        long nanoTime = 0L;  
        for (;;) {  
            ScheduledFutureTask<?> delayedTask = delayedTaskQueue.peek();  
            if (delayedTask == null) {  
                break;  
            }  
  
            if (nanoTime == 0L) {  
                nanoTime = ScheduledFutureTask.nanoTime();  
            }  
  
            if (delayedTask.deadlineNanos() <= nanoTime) {  
                delayedTaskQueue.remove();  
                taskQueue.add(delayedTask);  
            } else {  
                break;  
            }  
        }  
    }  

其功能是将延迟任务队列（delayedTaskQueue）中已经超过延迟执行时间的任务迁移到非IO任务队列（taskQueue）中.然后依次从taskQueue取出任务执行，每执行64个任务，就进行耗时检查，如果已执行时间超过预先设定的执行时间，则停止执行非IO任务，避免非IO任务太多，影响IO任务的执行

总结：NioEventLoop实现的线程执行逻辑做了以下事情

1. 先后执行IO任务和非IO任务，两类任务的执行时间比由变量ioRatio控制，默认是非IO任务允许执行和IO任务相同的时间
2. 如果taskQueue存在非IO任务，或者delayedTaskQueue存在已经超时的任务，则执行非阻塞的selectNow()方法，否则执行阻塞的select(time)方法
3. 如果阻塞的select(time)方法立即返回0的次数超过某个值（默认为512次），说明触发了epoll的cpu 100% bug，通过对selector进行rebuild解决：即重新创建一个selector，然后将原来的selector中已注册的所有channel重新注册到新的selector中，并将老的selectionKey全部cancel掉，最后将老的selector关闭
4. 如果select的结果不为0，则依次处理每个ready的selectionKey，根据readyOps的值，进行不同的分发处理，譬如accept、read、write、connect等
5. 执行完IO任务后，再执行非IO任务，其中会将delayedTaskQueue已超时的任务加入到taskQueue中。每执行64个任务，就进行耗时检查，如果已执行时间超过通过ioRatio和之前执行IO任务的耗时计算出来的非IO任务预计执行时间，则停止执行剩下的非IO任务