ActiveMQ producer 流量控制

http://activemq.apache.org/producer-flow-control.html

翻译：

流量控制是指：如果broker检测到destination的内存限制、temp文件限制、file store限制被超过了，就会减慢消息的流动。producer会被阻塞直到有可用资源，或者收到一个JMSException：这些行为是可以配置的。

值得一提的是：默认的<systemUsage>设置会导致producer阻塞，当达到memoryLimit或<systemUsage>的限制值时，这种阻塞行为有时被误认为是一个"挂起的producer"，而事实上producer只是在等待可用空间。

同步发送的消息会自动使用按producer的流量控制；通常会应用到同步发送持久化消息，除非你开启useAsyncSend标志。

使用异步发送的producer（通常讲，发送非持久化消息的producer）不需要等待来自broker的确认；所以，当达到memory limit时，你不会被通知到。如果你想知道broker的限制值达到了，你需要配置 ProducerWindowSize 连接参数，然后异步的消息也能按producer控制流量了。

ActiveMQConnectionFactory connctionFactory = ... 
connctionFactory.setProducerWindowSize(1024000);

 ProducerWindowSize 是producer发送的最大字节数，在等待来自broker的消息确认前。 如果你在发送非持久化消息（默认异步发送），并且希望知道queue或topic的memory limit是否达到了。那么你需要设置connection factory 为 'alwaysSyncSend'。但是，这会降低速度，它能保证你的producer马上知道内存问题。 如果你喜欢，你可以关掉指定jms queue和topic的流量控制，例如：

<destinationPolicy>  
    <policyMap>    
        <policyEntries>      
            <policyEntry topic="FOO.>" producerFlowControl="false"/> 
            </policyEntries>  
    </policyMap>
</destinationPolicy>

注意，在ActiveMQ 5.x中引入了新的file cursor，非持久化消息会被刷到临时文件存储中来减少内存使用量。所以，你会发现queue的memoryLimit永远达不到，因为file cursor花不了多少内存，如果你真的要把所有非持久化消息保存在内存中，并且当memoryLimit达到时停止producer，你应该配置<vmQueueCursor>。

<policyEntry queue=">" producerFlowControl="true" memoryLimit="1mb">      
    <pendingQueuePolicy>    
        <vmQueueCursor/>  
    </pendingQueuePolicy>
</policyEntry>

上面的片段能保证，所有的消息保存在内存中，并且每一个队列只有1Mb的限制。

How Producer Flow Control works

如果你在发送持久化消息，broker会发送一个ProducerAck消息给producer，它告知producer前一个发送窗口已经被处理了，所以producer现在可以发送下一个窗口。这和consumer的消息确认很像。当没有空间可用时，调用send()操作会无限阻塞，另一种方法是在客户端抛出异常。通过设置sendFailIfNoSpace为true，broker会导致send()抛出javax.jms.ResourceAllocationException，异常会传播到客户端。下面是配置示例：

<systemUsage> 
    <systemUsage sendFailIfNoSpace="true">   
        <memoryUsage>     
            <memoryUsage limit="20 mb"/>   
        </memoryUsage> 
    </systemUsage>
</systemUsage>

 这种做法的好处是，客户端会捕获一个javax.jms.ResourceAllocationException异常，等一会然后重试send()，而不再是无限地等待。

从5.3.1开始，加入了sendFailIfNoSpaceAfterTimeout属性，如果broker在配置的时间内仍然没有空余空间，此时send()才会失败，并且把异常传递到客户端，下面是配置：

<systemUsage> 
    <systemUsage sendFailIfNoSpaceAfterTimeout="3000">   
        <memoryUsage>     
            <memoryUsage limit="20 mb"/>   
        </memoryUsage> 
    </systemUsage>
</systemUsage>

关闭流量控制

通常的需求是关闭流量控制，这样消息分发可以一直进行直到磁盘空间被 pending messages 耗尽。

通过配置<systemUsage>元素的某些属性，你可以降低producer的速率。

<systemUsage>  
    <systemUsage>    
        <memoryUsage>      
            <memoryUsage limit="64 mb" />    
        </memoryUsage>    
        <storeUsage>      
            <storeUsage limit="100 gb" />    
        </storeUsage>    
        <tempUsage>      
            <tempUsage limit="10 gb" />    
        </tempUsage>  
    </systemUsage>
</systemUsage>

<memoryUsage>对应NON_PERSISTENT消息的内存容量，<storeUsage> 对应PERSITENT消息的磁盘容量，<tempUsage>对应临时文件的磁盘容量。

结合代码分析：

client-side： org.apache.activemq.ActiveMQMessageProducer.send

broker-side：从 org.apache.activemq.broker.region.Queue.send 开始

//org.apache.activemq.broker.region.Queue
public void send(final ProducerBrokerExchange producerExchange, final Message message) throws Exception {
    final ConnectionContext context = producerExchange.getConnectionContext();
    // There is delay between the client sending it and it arriving at the
    // destination.. it may have expired.
    message.setRegionDestination(this);
    ProducerState state = producerExchange.getProducerState();
    if (state == null) {
        LOG.warn("Send failed for: {}, missing producer state for: {}", message, producerExchange);
        throw new JMSException("Cannot send message to " + getActiveMQDestination() + " with invalid (null) producer state");
    }
    final ProducerInfo producerInfo = producerExchange.getProducerState().getInfo();
    //是否发送ProducerAck
    final boolean sendProducerAck = !message.isResponseRequired() && producerInfo.getWindowSize() > 0
            && !context.isInRecoveryMode();
    if (message.isExpired()) {
        // message not stored - or added to stats yet - so check here
        broker.getRoot().messageExpired(context, message, null);
        if (sendProducerAck) {
            ProducerAck ack = new ProducerAck(producerInfo.getProducerId(), message.getSize());
            context.getConnection().dispatchAsync(ack);
        }
        return;
    }
    if (memoryUsage.isFull()) { //如果内存耗尽
        // 尽管这里有大段代码，但是调试没进这儿
    }
    doMessageSend(producerExchange, message);
    if (sendProducerAck) {
        ProducerAck ack = new ProducerAck(producerInfo.getProducerId(), message.getSize());
        context.getConnection().dispatchAsync(ack);
    }
}

把消息放进 broker 的 pendingList 之前，会检查可用空间：

//org.apache.activemq.broker.region.Queue
void doMessageSend(final ProducerBrokerExchange producerExchange, final Message message) 
    throws IOException,    Exception {
    final ConnectionContext context = producerExchange.getConnectionContext();
    ListenableFuture<Object> result = null;
    boolean needsOrderingWithTransactions = context.isInTransaction();

    producerExchange.incrementSend();
    //检查使用空间
    checkUsage(context, producerExchange, message);
    sendLock.lockInterruptibly();
    try {
        if (store != null && message.isPersistent()) {
            try {
                message.getMessageId().setBrokerSequenceId(getDestinationSequenceId());
                if (messages.isCacheEnabled()) {
                    result = store.asyncAddQueueMessage(context, message, isOptimizeStorage());
                    result.addListener(new PendingMarshalUsageTracker(message));
                } else {
                    store.addMessage(context, message);
                }
                if (isReduceMemoryFootprint()) {
                    message.clearMarshalledState();
                }
            } catch (Exception e) {
                // we may have a store in inconsistent state, so reset the cursor
                // before restarting normal broker operations
                resetNeeded = true;
                throw e;
            }
        }
        // did a transaction commit beat us to the index?
        synchronized (orderIndexUpdates) {
            needsOrderingWithTransactions |= !orderIndexUpdates.isEmpty();
        }
        if (needsOrderingWithTransactions ) {
            // If this is a transacted message.. increase the usage now so that
            // a big TX does not blow up
            // our memory. This increment is decremented once the tx finishes..
            message.incrementReferenceCount();

            registerSendSync(message, context);
        } else {
            // Add to the pending list, this takes care of incrementing the
            // usage manager.
            sendMessage(message);
        }
    } finally {
        sendLock.unlock();
    }
    if (!needsOrderingWithTransactions) {
        messageSent(context, message);
    }
    if (result != null && message.isResponseRequired() && !result.isCancelled()) {
        try {
            result.get();
        } catch (CancellationException e) {
            // ignore - the task has been cancelled if the message
            // has already been deleted
        }
    }
}

对持久化消息和非持久化消息分类检查：

// org.apache.activemq.broker.region.Queue
private void checkUsage(ConnectionContext context,ProducerBrokerExchange producerBrokerExchange, Message message) 
    throws ResourceAllocationException, IOException, InterruptedException {
    if (message.isPersistent()) { // 持久化消息
        if (store != null && systemUsage.getStoreUsage().isFull(getStoreUsageHighWaterMark())) {
            final String logMessage = "Persistent store is Full, " + getStoreUsageHighWaterMark() + "% of "
                + systemUsage.getStoreUsage().getLimit() + ". Stopping producer ("
                + message.getProducerId() + ") to prevent flooding "
                + getActiveMQDestination().getQualifiedName() + "."
                + " See http://activemq.apache.org/producer-flow-control.html for more info";

            waitForSpace(context, producerBrokerExchange, systemUsage.getStoreUsage(), getStoreUsageHighWaterMark(), logMessage);
        }
    } else if (messages.getSystemUsage() != null && systemUsage.getTempUsage().isFull()) {
        // 非持久化消息
        final String logMessage = "Temp Store is Full ("
                + systemUsage.getTempUsage().getPercentUsage() + "% of " + systemUsage.getTempUsage().getLimit()
                +"). Stopping producer (" + message.getProducerId()
            + ") to prevent flooding " + getActiveMQDestination().getQualifiedName() + "."
            + " See http://activemq.apache.org/producer-flow-control.html for more info";

        waitForSpace(context, producerBrokerExchange, messages.getSystemUsage().getTempUsage(), logMessage);
    }
}

最后进入具体的处理逻辑：

// org.apache.activemq.broker.region.BaseDestination
protected final void waitForSpace(ConnectionContext context, ProducerBrokerExchange producerBrokerExchange, 
        Usage<?> usage, int highWaterMark, String warning) 
        throws IOException, InterruptedException, ResourceAllocationException {
    // 如果配置了sendFailIfNoSpace="true"
    if (!context.isNetworkConnection() && systemUsage.isSendFailIfNoSpace()) {
        getLog().debug("sendFailIfNoSpace, forcing exception on send, usage: {}: {}", usage, warning);
        throw new ResourceAllocationException(warning);
    }
    if (!context.isNetworkConnection() && systemUsage.getSendFailIfNoSpaceAfterTimeout() != 0) {
        if (!usage.waitForSpace(systemUsage.getSendFailIfNoSpaceAfterTimeout(), highWaterMark)) {
            getLog().debug("sendFailIfNoSpaceAfterTimeout expired, forcing exception on send, usage: {}: {}", usage, warning);
            throw new ResourceAllocationException(warning);
        }
    } else {
        long start = System.currentTimeMillis();
        long nextWarn = start;
        producerBrokerExchange.blockingOnFlowControl(true);
        destinationStatistics.getBlockedSends().increment();
        while (!usage.waitForSpace(1000, highWaterMark)) {
            if (context.getStopping().get()) {
                throw new IOException("Connection closed, send aborted.");
            }

            long now = System.currentTimeMillis();
            if (now >= nextWarn) {
                getLog().info("{}: {} (blocking for: {}s)", new Object[]{ usage, warning, new Long(((now - start) / 1000))});
                nextWarn = now + blockedProducerWarningInterval;
            }
        }
        long finish = System.currentTimeMillis();
        long totalTimeBlocked = finish - start;
        destinationStatistics.getBlockedTime().addTime(totalTimeBlocked);
        producerBrokerExchange.incrementTimeBlocked(this,totalTimeBlocked);
        producerBrokerExchange.blockingOnFlowControl(false);
    }
}

如果配置了sendFailIfNoSpace="true"，并且抛出异常了，处理异常的调用栈如下：

//org.apache.activemq.broker.TransportConnection
public void serviceException(Throwable e) {
    if (...) {
        ...
    }
    else if (!stopping.get() && !inServiceException) {
        inServiceException = true;
        try {
            SERVICELOG.warn("Async error occurred: ", e);
            ConnectionError ce = new ConnectionError();
            ce.setException(e);
            if (pendingStop) {
                dispatchSync(ce);
            } else {
                dispatchAsync(ce);
            }
        } finally {
            inServiceException = false;
        }
    }
}

那么 producer 是如何处理 ConnectionError 消息呢？

在org.apache.activemq.ActiveMQConnection.onCommand(Object o) 方法中：

public Response processConnectionError(final ConnectionError error) throws Exception {
    executor.execute(new Runnable() {
        @Override
        public void run() {
            onAsyncException(error.getException());
        }
    });
    return null;
}