RocketMQ中Broker的HA策略源码分析
- 2019 年 10 月 3 日
- 笔记
Broker的HA策略分为两部分
①同步元数据
②同步消息数据
同步元数据
在Slave启动时，会启动一个定时任务用来从master同步元数据
1 if (role == BrokerRole.SLAVE) { 2 if (null != slaveSyncFuture) { 3 slaveSyncFuture.cancel(false); 4 } 5 this.slaveSynchronize.setMasterAddr(null); 6 slaveSyncFuture = this.scheduledExecutorService.scheduleAtFixedRate(new Runnable() { 7 @Override 8 public void run() { 9 try { 10 BrokerController.this.slaveSynchronize.syncAll(); 11 } 12 catch (Throwable e) { 13 log.error("ScheduledTask SlaveSynchronize syncAll error.", e); 14 } 15 } 16 }, 1000 * 3, 1000 * 10, TimeUnit.MILLISECONDS); 17 }
这里设置了定时任务，执行slaveSynchronize的syncAll方法
可以注意在之前会通过setMasterAddr将Master的地址设为null，这是由于在后面会通过另一个定时任务registerBrokerAll来向NameServer获取Master的地址，详见：
SlaveSynchronize的syncAll方法：
1 public void syncAll() { 2 this.syncTopicConfig(); 3 this.syncConsumerOffset(); 4 this.syncDelayOffset(); 5 this.syncSubscriptionGroupConfig(); 6 }
这个方法会依次调用四个方法，来同步相应信息：
syncTopicConfig：同步topic的配置信息
syncConsumerOffset：同步Consumer的Offset信息
syncDelayOffset：同步延迟队列信息
syncSubscriptionGroupConfig：同步订阅信息
由于这几个方法的实现是类似的，这里就只看下syncTopicConfig的实现：
syncTopicConfig方法：
1 private void syncTopicConfig() { 2 String masterAddrBak = this.masterAddr; 3 if (masterAddrBak != null && !masterAddrBak.equals(brokerController.getBrokerAddr())) { 4 try { 5 TopicConfigSerializeWrapper topicWrapper = 6 this.brokerController.getBrokerOuterAPI().getAllTopicConfig(masterAddrBak); 7 if (!this.brokerController.getTopicConfigManager().getDataVersion() 8 .equals(topicWrapper.getDataVersion())) { 9 10 this.brokerController.getTopicConfigManager().getDataVersion() 11 .assignNewOne(topicWrapper.getDataVersion()); 12 this.brokerController.getTopicConfigManager().getTopicConfigTable().clear(); 13 this.brokerController.getTopicConfigManager().getTopicConfigTable() 14 .putAll(topicWrapper.getTopicConfigTable()); 15 this.brokerController.getTopicConfigManager().persist(); 16 17 log.info("Update slave topic config from master, {}", masterAddrBak); 18 } 19 } catch (Exception e) { 20 log.error("SyncTopicConfig Exception, {}", masterAddrBak, e); 21 } 22 } 23 }
这里首先获取master的地址masterAddr，由于registerBrokerAll定时任务的存在，即便这一次没有获取到masterAddr，只要节点中有master，总会在后面定时执行时从NameServer中获取到
当获取到master地址后，通过BrokerOuterAPI的getAllTopicConfig方法，向master请求
BrokerOuterAPI的getAllTopicConfig方法：
1 public TopicConfigSerializeWrapper getAllTopicConfig( 2 final String addr) throws RemotingConnectException, RemotingSendRequestException, 3 RemotingTimeoutException, InterruptedException, MQBrokerException { 4 RemotingCommand request = RemotingCommand.createRequestCommand(RequestCode.GET_ALL_TOPIC_CONFIG, null); 5 6 RemotingCommand response = this.remotingClient.invokeSync(MixAll.brokerVIPChannel(true, addr), request, 3000); 7 assert response != null; 8 switch (response.getCode()) { 9 case ResponseCode.SUCCESS: { 10 return TopicConfigSerializeWrapper.decode(response.getBody(), TopicConfigSerializeWrapper.class); 11 } 12 default: 13 break; 14 } 15 16 throw new MQBrokerException(response.getCode(), response.getRemark()); 17 }
首先构建GET_ALL_TOPIC_CONFIG求情指令，然后通过remotingClient的invokeSync进行同步发送，注意这里会通过MixAll的brokerVIPChannel方法，得到对应的master地址的VIP通道地址，就是端口号减2，这在我之前的博客中介绍过
有关同步发送在 【RocketMQ中Producer消息的发送源码分析】 中详细介绍过
请求发送给master后，来看看master是怎么处理的
master端在收到请求后会通过AdminBrokerProcessor的processRequest方法判别请求指令：
1 case RequestCode.GET_ALL_TOPIC_CONFIG: 2 return this.getAllTopicConfig(ctx, request);
执行getAllTopicConfig方法：
1 private RemotingCommand getAllTopicConfig(ChannelHandlerContext ctx, RemotingCommand request) { 2 final RemotingCommand response = RemotingCommand.createResponseCommand(GetAllTopicConfigResponseHeader.class); 3 // final GetAllTopicConfigResponseHeader responseHeader = 4 // (GetAllTopicConfigResponseHeader) response.readCustomHeader(); 5 6 String content = this.brokerController.getTopicConfigManager().encode(); 7 if (content != null && content.length() > 0) { 8 try { 9 response.setBody(content.getBytes(MixAll.DEFAULT_CHARSET)); 10 } catch (UnsupportedEncodingException e) { 11 log.error("", e); 12 13 response.setCode(ResponseCode.SYSTEM_ERROR); 14 response.setRemark("UnsupportedEncodingException " + e); 15 return response; 16 } 17 } else { 18 log.error("No topic in this broker, client: {}", ctx.channel().remoteAddress()); 19 response.setCode(ResponseCode.SYSTEM_ERROR); 20 response.setRemark("No topic in this broker"); 21 return response; 22 } 23 24 response.setCode(ResponseCode.SUCCESS); 25 response.setRemark(null); 26 27 return response; 28 }
这里会将TopicConfigManager中保存的topicConfigTable：
1 private final ConcurrentMap<String, TopicConfig> topicConfigTable = 2 new ConcurrentHashMap<String, TopicConfig>(1024);
将这个map通过encode方法转换成json字符串，再通过Netty发送给slave
回到slave中，在同步发送的情况下，会等待会送响应，收到响应后：
1 switch (response.getCode()) { 2 case ResponseCode.SUCCESS: { 3 return TopicConfigSerializeWrapper.decode(response.getBody(), TopicConfigSerializeWrapper.class); 4 } 5 default: 6 break; 7 }
通过decode解码，将json字符串转换为map封装在 TopicConfigSerializeWrapper中
回到syncTopicConfig方法中：
得到TopicConfigSerializeWrapper实例后
1 if (!this.brokerController.getTopicConfigManager().getDataVersion() 2 .equals(topicWrapper.getDataVersion())) { 3 4 this.brokerController.getTopicConfigManager().getDataVersion() 5 .assignNewOne(topicWrapper.getDataVersion()); 6 this.brokerController.getTopicConfigManager().getTopicConfigTable().clear(); 7 this.brokerController.getTopicConfigManager().getTopicConfigTable() 8 .putAll(topicWrapper.getTopicConfigTable()); 9 this.brokerController.getTopicConfigManager().persist(); 10 11 log.info("Update slave topic config from master, {}", masterAddrBak); 12 }
判断版本是否一致，若不一致，会进行替换，这样slave的Topic配置信息就和master保持同步了
其他三种信息的同步同理
同步消息数据
在master启动时，会通过JDK的NIO方式启动一个HA服务线程，用以处理slave的连接：
1 public void run() { 2 log.info(this.getServiceName() + " service started"); 3 4 while (!this.isStopped()) { 5 try { 6 this.selector.select(1000); 7 Set<SelectionKey> selected = this.selector.selectedKeys(); 8 9 if (selected != null) { 10 for (SelectionKey k : selected) { 11 if ((k.readyOps() & SelectionKey.OP_ACCEPT) != 0) { 12 SocketChannel sc = ((ServerSocketChannel) k.channel()).accept(); 13 14 if (sc != null) { 15 HAService.log.info("HAService receive new connection, " 16 + sc.socket().getRemoteSocketAddress()); 17 18 try { 19 HAConnection conn = new HAConnection(HAService.this, sc); 20 conn.start(); 21 HAService.this.addConnection(conn); 22 } catch (Exception e) { 23 log.error("new HAConnection exception", e); 24 sc.close(); 25 } 26 } 27 } else { 28 log.warn("Unexpected ops in select " + k.readyOps()); 29 } 30 } 31 32 selected.clear(); 33 } 34 } catch (Exception e) { 35 log.error(this.getServiceName() + " service has exception.", e); 36 } 37 } 38 39 log.info(this.getServiceName() + " service end"); 40 }
这里就是非常典型的JDK NIO的使用，在侦听到连接取得SocketChannel后，将其封装为HAConnection
1 public HAConnection(final HAService haService, final SocketChannel socketChannel) throws IOException { 2 this.haService = haService; 3 this.socketChannel = socketChannel; 4 this.clientAddr = this.socketChannel.socket().getRemoteSocketAddress().toString(); 5 this.socketChannel.configureBlocking(false); 6 this.socketChannel.socket().setSoLinger(false, -1); 7 this.socketChannel.socket().setTcpNoDelay(true); 8 this.socketChannel.socket().setReceiveBufferSize(1024 * 64); 9 this.socketChannel.socket().setSendBufferSize(1024 * 64); 10 this.writeSocketService = new WriteSocketService(this.socketChannel); 11 this.readSocketService = new ReadSocketService(this.socketChannel); 12 this.haService.getConnectionCount().incrementAndGet(); 13 }
在构造方法内进行了对socketChannel的一些配置，还创建了一个WriteSocketService和一个ReadSocketService，这两个是后续处理消息同步的基础
在创建完HAConnection后，调用其start方法：
1 public void start() { 2 this.readSocketService.start(); 3 this.writeSocketService.start(); 4 }
这里会启动两个线程，分别处理读取slave发送的数据，以及向slave发送数据
到这里，先不急着分析master了，来看看slave端
slave在启动时，会启动HAClient的线程：
1 public void run() { 2 log.info(this.getServiceName() + " service started"); 3 4 while (!this.isStopped()) { 5 try { 6 if (this.connectMaster()) { 7 8 if (this.isTimeToReportOffset()) { 9 boolean result = this.reportSlaveMaxOffset(this.currentReportedOffset); 10 if (!result) { 11 this.closeMaster(); 12 } 13 } 14 15 this.selector.select(1000); 16 17 boolean ok = this.processReadEvent(); 18 if (!ok) { 19 this.closeMaster(); 20 } 21 22 if (!reportSlaveMaxOffsetPlus()) { 23 continue; 24 } 25 26 long interval = 27 HAService.this.getDefaultMessageStore().getSystemClock().now() 28 - this.lastWriteTimestamp; 29 if (interval > HAService.this.getDefaultMessageStore().getMessageStoreConfig() 30 .getHaHousekeepingInterval()) { 31 log.warn("HAClient, housekeeping, found this connection[" + this.masterAddress 32 + "] expired, " + interval); 33 this.closeMaster(); 34 log.warn("HAClient, master not response some time, so close connection"); 35 } 36 } else { 37 this.waitForRunning(1000 * 5); 38 } 39 } catch (Exception e) { 40 log.warn(this.getServiceName() + " service has exception. ", e); 41 this.waitForRunning(1000 * 5); 42 } 43 } 44 45 log.info(this.getServiceName() + " service end"); 46 }
在这个while循环中，首先通过connectMaster检查是否和master连接了
connectMaster方法：
1 private boolean connectMaster() throws ClosedChannelException { 2 if (null == socketChannel) { 3 String addr = this.masterAddress.get(); 4 if (addr != null) { 5 6 SocketAddress socketAddress = RemotingUtil.string2SocketAddress(addr); 7 if (socketAddress != null) { 8 this.socketChannel = RemotingUtil.connect(socketAddress); 9 if (this.socketChannel != null) { 10 this.socketChannel.register(this.selector, SelectionKey.OP_READ); 11 } 12 } 13 } 14 15 this.currentReportedOffset = HAService.this.defaultMessageStore.getMaxPhyOffset(); 16 17 this.lastWriteTimestamp = System.currentTimeMillis(); 18 } 19 20 return this.socketChannel != null; 21 }
若是socketChannel为null，意味着并没有产生连接，或者连接断开
需要重新根据masterAddress建立网络连接
只要是需要建立连接，都需要通过defaultMessageStore的getMaxPhyOffset方法，获取本地最大的Offset，由currentReportedOffset保存，后续用于向master报告；以及保存了一个时间戳lastWriteTimestamp，用于之后的校对
当确保与master的连接建立成功后，通过isTimeToReportOffset方法，检查是否需要向master报告当前的最大Offset
isTimeToReportOffset方法：
1 private boolean isTimeToReportOffset() { 2 long interval = 3 HAService.this.defaultMessageStore.getSystemClock().now() - this.lastWriteTimestamp; 4 boolean needHeart = interval > HAService.this.defaultMessageStore.getMessageStoreConfig() 5 .getHaSendHeartbeatInterval(); 6 7 return needHeart; 8 }
这里就通过lastWriteTimestamp和当前时间检查，判断是否达到了报告时间间隔HaSendHeartbeatInterval，默认5s
若是达到了，就需要通过reportSlaveMaxOffset方法，将记录的currentReportedOffset这个最大的offset发送给master
reportSlaveMaxOffset方法：
1 private boolean reportSlaveMaxOffset(final long maxOffset) { 2 this.reportOffset.position(0); 3 this.reportOffset.limit(8); 4 this.reportOffset.putLong(maxOffset); 5 this.reportOffset.position(0); 6 this.reportOffset.limit(8); 7 8 for (int i = 0; i < 3 && this.reportOffset.hasRemaining(); i++) { 9 try { 10 this.socketChannel.write(this.reportOffset); 11 } catch (IOException e) { 12 log.error(this.getServiceName() 13 + "reportSlaveMaxOffset this.socketChannel.write exception", e); 14 return false; 15 } 16 } 17 18 return !this.reportOffset.hasRemaining(); 19 }
其中reportOffset是专门用来缓存offset的ByteBuffer
1 private final ByteBuffer reportOffset = ByteBuffer.allocate(8);
将maxOffset存放在reportOffset中，然后通过socketChannel的write方法，完成向master的发送
其中hasRemaining方法用来检查当前位置是否已经达到缓冲区极限limit，确保reportOffset 中的内容能被完全发送出去
发送成功后，会调用selector的select方法，在超时时间内进行NIO的轮询，等待master的回送
通过这我们可以看出slave在和master建立连接后，会定时向master报告自己当前的offset
来看看master收到offset后是如何处理的：
在master端会通过前面提到的ReadSocketService线程进行处理：
1 public void run() { 2 HAConnection.log.info(this.getServiceName() + " service started"); 3 4 while (!this.isStopped()) { 5 try { 6 this.selector.select(1000); 7 boolean ok = this.processReadEvent(); 8 if (!ok) { 9 HAConnection.log.error("processReadEvent error"); 10 break; 11 } 12 13 long interval = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now() - this.lastReadTimestamp; 14 if (interval > HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaHousekeepingInterval()) { 15 log.warn("ha housekeeping, found this connection[" + HAConnection.this.clientAddr + "] expired, " + interval); 16 break; 17 } 18 } catch (Exception e) { 19 HAConnection.log.error(this.getServiceName() + " service has exception.", e); 20 break; 21 } 22 } 23 24 this.makeStop(); 25 26 writeSocketService.makeStop(); 27 28 haService.removeConnection(HAConnection.this); 29 30 HAConnection.this.haService.getConnectionCount().decrementAndGet(); 31 32 SelectionKey sk = this.socketChannel.keyFor(this.selector); 33 if (sk != null) { 34 sk.cancel(); 35 } 36 37 try { 38 this.selector.close(); 39 this.socketChannel.close(); 40 } catch (IOException e) { 41 HAConnection.log.error("", e); 42 } 43 44 HAConnection.log.info(this.getServiceName() + " service end"); 45 }
这里的while循环中首先也是通过selector的select方法，在超时时间内进行NIO的轮询
轮询结束后的进一步的处理由processReadEvent来完成：
1 private boolean processReadEvent() { 2 int readSizeZeroTimes = 0; 3 4 if (!this.byteBufferRead.hasRemaining()) { 5 this.byteBufferRead.flip(); 6 this.processPostion = 0; 7 } 8 9 while (this.byteBufferRead.hasRemaining()) { 10 try { 11 int readSize = this.socketChannel.read(this.byteBufferRead); 12 if (readSize > 0) { 13 readSizeZeroTimes = 0; 14 this.lastReadTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now(); 15 if ((this.byteBufferRead.position() - this.processPostion) >= 8) { 16 int pos = this.byteBufferRead.position() - (this.byteBufferRead.position() % 8); 17 long readOffset = this.byteBufferRead.getLong(pos - 8); 18 this.processPostion = pos; 19 20 HAConnection.this.slaveAckOffset = readOffset; 21 if (HAConnection.this.slaveRequestOffset < 0) { 22 HAConnection.this.slaveRequestOffset = readOffset; 23 log.info("slave[" + HAConnection.this.clientAddr + "] request offset " + readOffset); 24 } 25 26 HAConnection.this.haService.notifyTransferSome(HAConnection.this.slaveAckOffset); 27 } 28 } else if (readSize == 0) { 29 if (++readSizeZeroTimes >= 3) { 30 break; 31 } 32 } else { 33 log.error("read socket[" + HAConnection.this.clientAddr + "] < 0"); 34 return false; 35 } 36 } catch (IOException e) { 37 log.error("processReadEvent exception", e); 38 return false; 39 } 40 } 41 42 return true; 43 } 44 }
这个方法其实就是通过socketChannel的read方法，将slave发送过来的数据存入byteBufferRead中
在确保发送过来的数据能达到8字节时，取出long类型的offset值，然后交给HAConnection的slaveAckOffset成员进行保存
其中slaveRequestOffset是用来处理第一次连接时的同步
notifyTransferSome方法是作为同步master时，进行相应的唤醒操作，异步master则没有要求，在后面具体分析
也就是说ReadSocketService这个线程，只是不断地读取并更新slave发送来的offset数据
再来看看WriteSocketService线程是如何进行向slave的发送：
1 public void run() { 2 HAConnection.log.info(this.getServiceName() + " service started"); 3 4 while (!this.isStopped()) { 5 try { 6 this.selector.select(1000); 7 8 if (-1 == HAConnection.this.slaveRequestOffset) { 9 Thread.sleep(10); 10 continue; 11 } 12 13 if (-1 == this.nextTransferFromWhere) { 14 if (0 == HAConnection.this.slaveRequestOffset) { 15 long masterOffset = HAConnection.this.haService.getDefaultMessageStore().getCommitLog().getMaxOffset(); 16 masterOffset = 17 masterOffset 18 - (masterOffset % HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig() 19 .getMapedFileSizeCommitLog()); 20 21 if (masterOffset < 0) { 22 masterOffset = 0; 23 } 24 25 this.nextTransferFromWhere = masterOffset; 26 } else { 27 this.nextTransferFromWhere = HAConnection.this.slaveRequestOffset; 28 } 29 30 log.info("master transfer data from " + this.nextTransferFromWhere + " to slave[" + HAConnection.this.clientAddr 31 + "], and slave request " + HAConnection.this.slaveRequestOffset); 32 } 33 34 if (this.lastWriteOver) { 35 36 long interval = 37 HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now() - this.lastWriteTimestamp; 38 39 if (interval > HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig() 40 .getHaSendHeartbeatInterval()) { 41 42 // Build Header 43 this.byteBufferHeader.position(0); 44 this.byteBufferHeader.limit(headerSize); 45 this.byteBufferHeader.putLong(this.nextTransferFromWhere); 46 this.byteBufferHeader.putInt(0); 47 this.byteBufferHeader.flip(); 48 49 this.lastWriteOver = this.transferData(); 50 if (!this.lastWriteOver) 51 continue; 52 } 53 } else { 54 this.lastWriteOver = this.transferData(); 55 if (!this.lastWriteOver) 56 continue; 57 } 58 59 SelectMappedBufferResult selectResult = 60 HAConnection.this.haService.getDefaultMessageStore().getCommitLogData(this.nextTransferFromWhere); 61 if (selectResult != null) { 62 int size = selectResult.getSize(); 63 if (size > HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaTransferBatchSize()) { 64 size = HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaTransferBatchSize(); 65 } 66 67 long thisOffset = this.nextTransferFromWhere; 68 this.nextTransferFromWhere += size; 69 70 selectResult.getByteBuffer().limit(size); 71 this.selectMappedBufferResult = selectResult; 72 73 // Build Header 74 this.byteBufferHeader.position(0); 75 this.byteBufferHeader.limit(headerSize); 76 this.byteBufferHeader.putLong(thisOffset); 77 this.byteBufferHeader.putInt(size); 78 this.byteBufferHeader.flip(); 79 80 this.lastWriteOver = this.transferData(); 81 } else { 82 83 HAConnection.this.haService.getWaitNotifyObject().allWaitForRunning(100); 84 } 85 } catch (Exception e) { 86 87 HAConnection.log.error(this.getServiceName() + " service has exception.", e); 88 break; 89 } 90 } 91 92 HAConnection.this.haService.getWaitNotifyObject().removeFromWaitingThreadTable(); 93 94 if (this.selectMappedBufferResult != null) { 95 this.selectMappedBufferResult.release(); 96 } 97 98 this.makeStop(); 99 100 readSocketService.makeStop(); 101 102 haService.removeConnection(HAConnection.this); 103 104 SelectionKey sk = this.socketChannel.keyFor(this.selector); 105 if (sk != null) { 106 sk.cancel(); 107 } 108 109 try { 110 this.selector.close(); 111 this.socketChannel.close(); 112 } catch (IOException e) { 113 HAConnection.log.error("", e); 114 } 115 116 HAConnection.log.info(this.getServiceName() + " service end"); 117 }
这里一开始会对slaveRequestOffset进行一次判断，当且仅当slaveRequestOffset初始化的时候是才是-1
也就是说当slave还没有发送过来offset时，WriteSocketService线程只会干等
当slave发送来offset后
首先对nextTransferFromWhere进行了判断，nextTransferFromWhere和slaveRequestOffset一样，在初始化的时候为-1
也就代表着master和slave刚刚建立连接，并没有进行过一次消息的同步！
此时会对修改了的slaveRequestOffset进行判断
若是等于0，说明slave没有任何消息的历史记录，那么此时master会取得自身的MaxOffset，根据这个MaxOffset，通过：
1 masterOffset = masterOffset 2 - (masterOffset % HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig() 3 .getMapedFileSizeCommitLog() /* 1G */);
计算出最后一个文件开始的offset
也就是说，当slave没有消息的历史记录，master只会从本地最后一个CommitLog文件开始的地方，将消息数据发送给slave
若是slave有数据，就从slave发送来的offset的位置起，进行发送，通过nextTransferFromWhere记录这个offset值
接着对lastWriteOver进行了判断，lastWriteOver是一个状态量，用来表示上次发送是否传输完毕，初始化是true
若是true，这里会进行一次时间检查，lastWriteTimestamp记录最后一次发送的时间
一次来判断是否超过了时间间隔haSendHeartbeatInterval（默认5s）
也就是说至少有5s，master没有向slave发送任何消息
那么此时就会发送一个心跳包
其中byteBufferHeader是一个12字节的ByteBuffer：
1 private final int headerSize = 8 + 4; 2 private final ByteBuffer byteBufferHeader = ByteBuffer.allocate(headerSize);
这里就简单地构造了一个心跳包，后续通过transferData方法来完成数据的发送
若是 lastWriteOver为false，则表示上次数据没有发送完，就需要通过transferData方法，将剩余数据继续发送，只要没发送完，只会重复循环，直到发完
先继续往下看，下面就是发送具体的消息数据了：
首先根据nextTransferFromWhere，也就是刚才保存的offset，通过DefaultMessageStore的getCommitLogData方法，其实际上调用的是CommitLog的getData方法，这个方法在
【RocketMQ中Broker的启动源码分析（二）】中关于消息调度（ReputMessageService）时详细介绍过
根据offset找到对应的CommitLog文件，将其从offset对应起始处所有数据读入ByteBuffer中，由SelectMappedBufferResult封装
这里若是master已将将所有本地数据同步给了slave，那么得到的SelectMappedBufferResult就会为null，会调用：
1 HAConnection.this.haService.getWaitNotifyObject().allWaitForRunning(100);
将自身阻塞，超时等待100ms，要么一直等到超时时间到了，要么就会在后面所讲的同步双传中被同步master唤醒
在得到SelectMappedBufferResult后，这里会对读取到的数据大小进行一次判断，若是大于haTransferBatchSize（默认32K），将size改为32K，实际上就是对发送数据大小的限制，大于32K会切割，每次最多只允许发送32k
通过thisOffset记录nextTransferFromWhere即offset
更新nextTransferFromWhere值，以便下一次定位
还会将读取到的数据结果selectResult交给selectMappedBufferResult保存
然后构建消息头，这里就和心跳包格式一样，前八字节存放offset，后四字节存放数据大小
最后调用transferData方法，进行发送：
1 private boolean transferData() throws Exception { 2 int writeSizeZeroTimes = 0; 3 // Write Header 4 while (this.byteBufferHeader.hasRemaining()) { 5 int writeSize = this.socketChannel.write(this.byteBufferHeader); 6 if (writeSize > 0) { 7 writeSizeZeroTimes = 0; 8 this.lastWriteTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now(); 9 } else if (writeSize == 0) { 10 if (++writeSizeZeroTimes >= 3) { 11 break; 12 } 13 } else { 14 throw new Exception("ha master write header error < 0"); 15 } 16 } 17 18 if (null == this.selectMappedBufferResult) { 19 return !this.byteBufferHeader.hasRemaining(); 20 } 21 22 writeSizeZeroTimes = 0; 23 24 // Write Body 25 if (!this.byteBufferHeader.hasRemaining()) { 26 while (this.selectMappedBufferResult.getByteBuffer().hasRemaining()) { 27 int writeSize = this.socketChannel.write(this.selectMappedBufferResult.getByteBuffer()); 28 if (writeSize > 0) { 29 writeSizeZeroTimes = 0; 30 this.lastWriteTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now(); 31 } else if (writeSize == 0) { 32 if (++writeSizeZeroTimes >= 3) { 33 break; 34 } 35 } else { 36 throw new Exception("ha master write body error < 0"); 37 } 38 } 39 } 40 41 boolean result = !this.byteBufferHeader.hasRemaining() && !this.selectMappedBufferResult.getByteBuffer().hasRemaining(); 42 43 if (!this.selectMappedBufferResult.getByteBuffer().hasRemaining()) { 44 this.selectMappedBufferResult.release(); 45 this.selectMappedBufferResult = null; 46 } 47 48 return result; 49 }
首先将byteBufferHeader中的12字节消息头通过socketChannel的write方法发送出去
然后将selectMappedBufferResult中的ByteBuffer的消息数据发送出去
若是selectMappedBufferResult等于null，说明是心跳包，只发送消息头
无论发送什么都会将时间记录在lastWriteTimestamp中，以便后续发送心跳包的判断
看到这里其实就会发现WriteSocketService线程开启后，只要slave向master发出了第一个offset后，WriteSocketService线程都会不断地将对应位置自己本地的CommitLog文件中的内容发送给slave，直到完全同步后，WriteSocketService线程才会稍微缓缓，进入阻塞100ms以及每隔五秒发一次心跳包的状态
但是只要当Producer向master发送来消息后，由刷盘线程完成持久化后，WriteSocketService线程又会忙碌起来，此时也才是体现同步双写和异步复制的时候
先不急着说这个，来看看slave接收到消息是如何处理的：
是在HAClient的线程中的processReadEvent方法处理的：
1 private boolean processReadEvent() { 2 int readSizeZeroTimes = 0; 3 while (this.byteBufferRead.hasRemaining()) { 4 try { 5 int readSize = this.socketChannel.read(this.byteBufferRead); 6 if (readSize > 0) { 7 lastWriteTimestamp = HAService.this.defaultMessageStore.getSystemClock().now(); 8 readSizeZeroTimes = 0; 9 boolean result = this.dispatchReadRequest(); 10 if (!result) { 11 log.error("HAClient, dispatchReadRequest error"); 12 return false; 13 } 14 } else if (readSize == 0) { 15 if (++readSizeZeroTimes >= 3) { 16 break; 17 } 18 } else { 19 log.info("HAClient, processReadEvent read socket < 0"); 20 return false; 21 } 22 } catch (IOException e) { 23 log.info("HAClient, processReadEvent read socket exception", e); 24 return false; 25 } 26 } 27 28 return true; 29 }
在socketChannel通过read方法将master发送的数据读取到byteBufferRead缓冲区后，由dispatchReadRequest方法做进一步处理
dispatchReadRequest方法：
1 private boolean dispatchReadRequest() { 2 final int msgHeaderSize = 8 + 4; // phyoffset + size 3 int readSocketPos = this.byteBufferRead.position(); 4 5 while (true) { 6 int diff = this.byteBufferRead.position() - this.dispatchPostion; 7 if (diff >= msgHeaderSize) { 8 long masterPhyOffset = this.byteBufferRead.getLong(this.dispatchPostion); 9 int bodySize = this.byteBufferRead.getInt(this.dispatchPostion + 8); 10 11 long slavePhyOffset = HAService.this.defaultMessageStore.getMaxPhyOffset(); 12 13 if (slavePhyOffset != 0) { 14 if (slavePhyOffset != masterPhyOffset) { 15 log.error("master pushed offset not equal the max phy offset in slave, SLAVE: " 16 + slavePhyOffset + " MASTER: " + masterPhyOffset); 17 return false; 18 } 19 } 20 21 if (diff >= (msgHeaderSize + bodySize)) { 22 byte[] bodyData = new byte[bodySize]; 23 this.byteBufferRead.position(this.dispatchPostion + msgHeaderSize); 24 this.byteBufferRead.get(bodyData); 25 26 HAService.this.defaultMessageStore.appendToCommitLog(masterPhyOffset, bodyData); 27 28 this.byteBufferRead.position(readSocketPos); 29 this.dispatchPostion += msgHeaderSize + bodySize; 30 31 if (!reportSlaveMaxOffsetPlus()) { 32 return false; 33 } 34 35 continue; 36 } 37 } 38 39 if (!this.byteBufferRead.hasRemaining()) { 40 this.reallocateByteBuffer(); 41 } 42 43 break; 44 } 45 46 return true; 47 }
这里就首先将12字节的消息头取出来
masterPhyOffset：8字节offset ，bodySize ：4字节消息大小
根据master发来的masterPhyOffset会和自己本地的slavePhyOffset进行校验，以便安全备份
之后就会将byteBufferRead中存放在消息头后面的消息数据取出来，调用appendToCommitLog方法持久化到的CommitLog中
1 public boolean appendToCommitLog(long startOffset, byte[] data) { 2 if (this.shutdown) { 3 log.warn("message store has shutdown, so appendToPhyQueue is forbidden"); 4 return false; 5 } 6 7 boolean result = this.commitLog.appendData(startOffset, data); 8 if (result) { 9 this.reputMessageService.wakeup(); 10 } else { 11 log.error("appendToPhyQueue failed " + startOffset + " " + data.length); 12 } 13 14 return result; 15 }
实际上调用了commitLog的appendData方法将其写入磁盘，这个方法我在前面博客中介绍过
在完成写入后，需要唤醒reputMessageService消息调度，以便Consumer的消费
关于消息调度详见 【RocketMQ中Broker的启动源码分析（二）】
当然前面说过master还会发送心跳消息，但这里明显没对心跳消息进行处理，只是appendToCommitLog调用时，传入了一个大小为0的byte数组，显然有些不合理，想不通
在完成后，还会调用reportSlaveMaxOffsetPlus方法：
1 private boolean reportSlaveMaxOffsetPlus() { 2 boolean result = true; 3 long currentPhyOffset = HAService.this.defaultMessageStore.getMaxPhyOffset(); 4 if (currentPhyOffset > this.currentReportedOffset) { 5 this.currentReportedOffset = currentPhyOffset; 6 result = this.reportSlaveMaxOffset(this.currentReportedOffset); 7 if (!result) { 8 this.closeMaster(); 9 log.error("HAClient, reportSlaveMaxOffset error, " + this.currentReportedOffset); 10 } 11 } 12 13 return result; 14 }
由于完成了写入，那么此时获取到的offset肯定比currentReportedOffset中保存的大，然后再次通过reportSlaveMaxOffset方法，将当前的offset报告给master
这其实上已经完成了异步master的异步复制过程
再来看看同步双写是如何实现的：
和刷盘一样，都是在Producer发送完消息，Broker进行完消息的存储后进行的
在CommitLog的handleHA方法：
1 public void handleHA(AppendMessageResult result, PutMessageResult putMessageResult, MessageExt messageExt) { 2 if (BrokerRole.SYNC_MASTER == this.defaultMessageStore.getMessageStoreConfig().getBrokerRole()) { 3 HAService service = this.defaultMessageStore.getHaService(); 4 if (messageExt.isWaitStoreMsgOK()) { 5 // Determine whether to wait 6 if (service.isSlaveOK(result.getWroteOffset() + result.getWroteBytes())) { 7 GroupCommitRequest request = new GroupCommitRequest(result.getWroteOffset() + result.getWroteBytes()); 8 service.putRequest(request); 9 service.getWaitNotifyObject().wakeupAll(); 10 boolean flushOK = 11 request.waitForFlush(this.defaultMessageStore.getMessageStoreConfig().getSyncFlushTimeout()); 12 if (!flushOK) { 13 log.error("do sync transfer other node, wait return, but failed, topic: " + messageExt.getTopic() + " tags: " 14 + messageExt.getTags() + " client address: " + messageExt.getBornHostNameString()); 15 putMessageResult.setPutMessageStatus(PutMessageStatus.FLUSH_SLAVE_TIMEOUT); 16 } 17 } 18 // Slave problem 19 else { 20 // Tell the producer, slave not available 21 putMessageResult.setPutMessageStatus(PutMessageStatus.SLAVE_NOT_AVAILABLE); 22 } 23 } 24 } 25 26 }
这里就会检查Broker的类型，看以看到只对SYNC_MASTER即同步master进行了操作
这个操作过程其实就和同步刷盘类似
根据Offset+WroteBytes创建一条记录GroupCommitRequest，然后会将添加在List中
然后调用getWaitNotifyObject的wakeupAll方法，把阻塞中的所有WriteSocketService线程唤醒
因为master和slave是一对多的关系，那么这里就会有多个slave连接，也就有多个WriteSocketService线程，保证消息能同步到所有slave中
在唤醒WriteSocketService线程工作后，调用request的waitForFlush方法，将自身阻塞，预示着同步复制的真正开启
在HAService开启时，还开启了一个GroupTransferService线程：
1 public void run() { 2 log.info(this.getServiceName() + " service started"); 3 4 while (!this.isStopped()) { 5 try { 6 this.waitForRunning(10); 7 this.doWaitTransfer(); 8 } catch (Exception e) { 9 log.warn(this.getServiceName() + " service has exception. ", e); 10 } 11 } 12 13 log.info(this.getServiceName() + " service end"); 14 }
这里的工作原理和同步刷盘GroupCommitService基本一致，相似的地方我就不仔细分析了
GroupTransferService同样保存两张List：
1 private volatile List<CommitLog.GroupCommitRequest> requestsWrite = new ArrayList<>(); 2 private volatile List<CommitLog.GroupCommitRequest> requestsRead = new ArrayList<>();
由这两张List做一个类似JVM新生代的复制算法
在handleHA方法中，就会将创建的GroupCommitRequest记录添加在requestsWrite这个List中
其中doWaitTransfer方法：
1 private void doWaitTransfer() { 2 synchronized (this.requestsRead) { 3 if (!this.requestsRead.isEmpty()) { 4 for (CommitLog.GroupCommitRequest req : this.requestsRead) { 5 boolean transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset(); 6 for (int i = 0; !transferOK && i < 5; i++) { 7 this.notifyTransferObject.waitForRunning(1000); 8 transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset(); 9 } 10 11 if (!transferOK) { 12 log.warn("transfer messsage to slave timeout, " + req.getNextOffset()); 13 } 14 15 req.wakeupCustomer(transferOK); 16 } 17 18 this.requestsRead.clear(); 19 } 20 } 21 }
和刷盘一样，这里会通过复制算法，将requestsWrite和requestsRead进行替换，那么这里的requestsRead实际上就存放着刚才添加的记录
首先取出记录中的NextOffset和push2SlaveMaxOffset比较
push2SlaveMaxOffset值是通过slave发送过来的，在之前说过的ReadSocketService线程中的：
1 HAConnection.this.haService.notifyTransferSome(HAConnection.this.slaveAckOffset);
notifyTransferSome方法：
1 public void notifyTransferSome(final long offset) { 2 for (long value = this.push2SlaveMaxOffset.get(); offset > value; ) { 3 boolean ok = this.push2SlaveMaxOffset.compareAndSet(value, offset); 4 if (ok) { 5 this.groupTransferService.notifyTransferSome(); 6 break; 7 } else { 8 value = this.push2SlaveMaxOffset.get(); 9 } 10 } 11 }
即便也多个slave连接，这里的push2SlaveMaxOffset永远会记录最大的那个offset
所以在doWaitTransfer中，根据当前NextOffset（完成写入后master本地的offset），进行判断
其实这里主要要考虑到WriteSocketService线程的工作原理，只要本地文件有更新，那么就会向slave发送数据，所以这里由于HA同步是发生在刷盘后的，那么就有可能在这个doWaitTransfer执行前，有slave已经将数据进行了同步，并且向master报告了自己offset，更新了push2SlaveMaxOffset的值
那么
1 boolean transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset(); 2 ```
这个判断就会为真，意味着节点中已经有了备份，所以就会直接调用
1 req.wakeupCustomer(transferOK);
以此来唤醒刚才在handleHA方法中的阻塞
若是判断为假，就说明没有一个slave完成同步，就需要
1 for (int i = 0; !transferOK && i < 5; i++) { 2 this.notifyTransferObject.waitForRunning(1000); 3 transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset(); 4 }
通过waitForRunning进行阻塞，超时等待，最多五次等待，超过时间会向Producer发送FLUSH_SLAVE_TIMEOUT
若是在超时时间内，有slave完成了同步，并向master发送了offset后，在notifyTransferSome方法中：
1 public void notifyTransferSome(final long offset) { 2 for (long value = this.push2SlaveMaxOffset.get(); offset > value; ) { 3 boolean ok = this.push2SlaveMaxOffset.compareAndSet(value, offset); 4 if (ok) { 5 this.groupTransferService.notifyTransferSome(); 6 break; 7 } else { 8 value = this.push2SlaveMaxOffset.get(); 9 } 10 } 11 }
就会更新push2SlaveMaxOffset，并通过notifyTransferSome唤醒上面所说的阻塞
然后再次判断push2SlaveMaxOffset和getNextOffset
成功后唤醒刚才在handleHA方法中的阻塞，同步master的主从复制也就结束
由于同步master的刷盘是在主从复制前发生的，所以同步双写意味着master和slave都会完成消息的持久化
至此，RocketMQ中Broker的HA策略分析到此结束
