Hadoop学习总结之二：HDFS读写过程解析(转)

一、文件的打开
1.1、客户端
HDFS打开一个文件，需要在客户端调用DistributedFileSystem.open(Path f, int bufferSize)，其实现为：
public FSDataInputStream open(Path f, int bufferSize) throws IOException {
  return new DFSClient.DFSDataInputStream(
        dfs.open(getPathName(f), bufferSize, verifyChecksum, statistics));
}
其中dfs为DistributedFileSystem的成员变量DFSClient，其open函数被调用，其中创建一个DFSInputStream(src, buffersize, verifyChecksum)并返回。
在DFSInputStream的构造函数中，openInfo函数被调用，其主要从namenode中得到要打开的文件所对应的blocks的信息，实现如下：
 
synchronized void openInfo() throws IOException {
  LocatedBlocks newInfo = callGetBlockLocations(namenode, src, 0, prefetchSize);
  this.locatedBlocks = newInfo;
  this.currentNode = null;
}
private static LocatedBlocks callGetBlockLocations(ClientProtocol namenode,
    String src, long start, long length) throws IOException {
    return namenode.getBlockLocations(src, start, length);
}
LocatedBlocks主要包含一个链表的List<LocatedBlock> blocks，其中每个LocatedBlock包含如下信息：
Block b：此block的信息
long offset：此block在文件中的偏移量
DatanodeInfo[] locs：此block位于哪些DataNode上
上面namenode.getBlockLocations是一个RPC调用，最终调用NameNode类的getBlockLocations函数。
1.2、NameNode
NameNode.getBlockLocations实现如下：
public LocatedBlocks   getBlockLocations(String src,
                                        long offset,
                                        long length) throws IOException {
  return namesystem.getBlockLocations(getClientMachine(),
                                      src, offset, length);
}
namesystem是NameNode一个成员变量，其类型为FSNamesystem，保存的是NameNode的name space树，其中一个重要的成员变量为FSDirectory dir。
FSDirectory和Lucene中的FSDirectory没有任何关系，其主要包括FSImage fsImage，用于读写硬盘上的fsimage文件，FSImage类有成员变量FSEditLog editLog，用于读写硬盘上的edit文件，这两个文件的关系在上一篇文章中已经解释过。
FSDirectory还有一个重要的成员变量INodeDirectoryWithQuota rootDir，INodeDirectoryWithQuota的父类为INodeDirectory，实现如下：
public class INodeDirectory extends INode {
  ……
  private List<INode> children;
  ……
｝
由此可见INodeDirectory本身是一个INode，其中包含一个链表的INode，此链表中，如果仍为文件夹，则是类型INodeDirectory，如果是文件，则是类型INodeFile，INodeFile中有成员变量BlockInfo blocks[]，是此文件包含的block的信息。显然这是一棵树形的结构。
FSNamesystem.getBlockLocations函数如下：
public LocatedBlocks getBlockLocations(String src, long offset, long length,
    boolean doAccessTime) throws IOException {
  final LocatedBlocks ret = getBlockLocationsInternal(src, dir.getFileINode(src),
      offset, length, Integer.MAX_VALUE, doAccessTime); 
  return ret;
}
dir.getFileINode(src)通过路径名从文件系统树中找到INodeFile，其中保存的是要打开的文件的INode的信息。
getBlockLocationsInternal的实现如下：
 
private synchronized LocatedBlocks getBlockLocationsInternal(String src,
                                                     INodeFile inode,
                                                     long offset,
                                                     long length,
                                                     int nrBlocksToReturn,
                                                     boolean doAccessTime)
                                                     throws IOException {
  //得到此文件的block信息
  Block[] blocks = inode.getBlocks();
  List<LocatedBlock> results = new ArrayList<LocatedBlock>(blocks.length);
  //计算从offset开始，长度为length所涉及的blocks
  int curBlk = 0;
  long curPos = 0, blkSize = 0;
  int nrBlocks = (blocks[0].getNumBytes() == 0) ? 0 : blocks.length;
  for (curBlk = 0; curBlk < nrBlocks; curBlk++) {
    blkSize = blocks[curBlk].getNumBytes();
    if (curPos + blkSize > offset) {
      //当offset在curPos和curPos + blkSize之间的时候，curBlk指向offset所在的block
      break;
    }
    curPos += blkSize;
  }
  long endOff = offset + length;
  //循环，依次遍历从curBlk开始的每个block，直到当前位置curPos越过endOff
  do {
    int numNodes = blocksMap.numNodes(blocks[curBlk]);
    int numCorruptNodes = countNodes(blocks[curBlk]).corruptReplicas();
    int numCorruptReplicas = corruptReplicas.numCorruptReplicas(blocks[curBlk]);
    boolean blockCorrupt = (numCorruptNodes == numNodes);
    int numMachineSet = blockCorrupt ? numNodes :
                          (numNodes - numCorruptNodes);
    //依次找到此block所对应的datanode，将其中没有损坏的放入machineSet中
    DatanodeDescriptor[] machineSet = new DatanodeDescriptor[numMachineSet];
    if (numMachineSet > 0) {
      numNodes = 0;
      for(Iterator<DatanodeDescriptor> it =
          blocksMap.nodeIterator(blocks[curBlk]); it.hasNext();) {
        DatanodeDescriptor dn = it.next();
        boolean replicaCorrupt = corruptReplicas.isReplicaCorrupt(blocks[curBlk], dn);
        if (blockCorrupt || (!blockCorrupt && !replicaCorrupt))
          machineSet[numNodes++] = dn;
      }
    }
    //使用此machineSet和当前的block构造一个LocatedBlock
    results.add(new LocatedBlock(blocks[curBlk], machineSet, curPos,
                blockCorrupt));
    curPos += blocks[curBlk].getNumBytes();
    curBlk++;
  } while (curPos < endOff
        && curBlk < blocks.length
        && results.size() < nrBlocksToReturn);
  //使用此LocatedBlock链表构造一个LocatedBlocks对象返回
  return inode.createLocatedBlocks(results);
}
1.3、客户端
通过RPC调用，在NameNode得到的LocatedBlocks对象，作为成员变量构造DFSInputStream对象，最后包装为FSDataInputStream返回给用户。
 
二、文件的读取
2.1、客户端
文件读取的时候，客户端利用文件打开的时候得到的FSDataInputStream.read(long position, byte[] buffer, int offset, int length)函数进行文件读操作。
FSDataInputStream会调用其封装的DFSInputStream的read(long position, byte[] buffer, int offset, int length)函数，实现如下：
 
public int read(long position, byte[] buffer, int offset, int length)
  throws IOException {
  long filelen = getFileLength();
  int realLen = length;
  if ((position + length) > filelen) {
    realLen = (int)(filelen - position);
  }
  //首先得到包含从offset到offset + length内容的block列表
  //比如对于64M一个block的文件系统来说，欲读取从100M开始，长度为128M的数据，则block列表包括第2，3，4块block
  List<LocatedBlock> blockRange = getBlockRange(position, realLen);
  int remaining = realLen;
  //对每一个block，从中读取内容
  //对于上面的例子，对于第2块block，读取从36M开始，读取长度28M，对于第3块，读取整一块64M，对于第4块，读取从0开始，长度为36M，共128M数据
  for (LocatedBlock blk : blockRange) {
    long targetStart = position - blk.getStartOffset();
    long bytesToRead = Math.min(remaining, blk.getBlockSize() - targetStart);
    fetchBlockByteRange(blk, targetStart,
                        targetStart + bytesToRead - 1, buffer, offset);
    remaining -= bytesToRead;
    position += bytesToRead;
    offset += bytesToRead;
  }
  assert remaining == 0 : "Wrong number of bytes read.";
  if (stats != null) {
    stats.incrementBytesRead(realLen);
  }
  return realLen;
}
其中getBlockRange函数如下：
 
private synchronized List<LocatedBlock> getBlockRange(long offset,
                                                      long length)
                                                    throws IOException {
  List<LocatedBlock> blockRange = new ArrayList<LocatedBlock>();
  //首先从缓存的locatedBlocks中查找offset所在的block在缓存链表中的位置
  int blockIdx = locatedBlocks.findBlock(offset);
  if (blockIdx < 0) { // block is not cached
    blockIdx = LocatedBlocks.getInsertIndex(blockIdx);
  }
  long remaining = length;
  long curOff = offset;
  while(remaining > 0) {
    LocatedBlock blk = null;
    //按照blockIdx的位置找到block
    if(blockIdx < locatedBlocks.locatedBlockCount())
      blk = locatedBlocks.get(blockIdx);
    //如果block为空，则缓存中没有此block，则直接从NameNode中查找这些block，并加入缓存
    if (blk == null || curOff < blk.getStartOffset()) {
      LocatedBlocks newBlocks;
      newBlocks = callGetBlockLocations(namenode, src, curOff, remaining);
      locatedBlocks.insertRange(blockIdx, newBlocks.getLocatedBlocks());
      continue;
    }
    //如果block找到，则放入结果集
    blockRange.add(blk);
    long bytesRead = blk.getStartOffset() + blk.getBlockSize() - curOff;
    remaining -= bytesRead;
    curOff += bytesRead;
    //取下一个block
    blockIdx++;
  }
  return blockRange;
}
其中fetchBlockByteRange实现如下：
 
private void fetchBlockByteRange(LocatedBlock block, long start,
                                 long end, byte[] buf, int offset) throws IOException {
  Socket dn = null;
  int numAttempts = block.getLocations().length;
  //此while循环为读取失败后的重试次数
  while (dn == null && numAttempts-- > 0 ) {
    //选择一个DataNode来读取数据
    DNAddrPair retval = chooseDataNode(block);
    DatanodeInfo chosenNode = retval.info;
    InetSocketAddress targetAddr = retval.addr;
    BlockReader reader = null;
    try {
      //创建Socket连接到DataNode
      dn = socketFactory.createSocket();
      dn.connect(targetAddr, socketTimeout);
      dn.setSoTimeout(socketTimeout);
      int len = (int) (end - start + 1);
      //利用建立的Socket链接，生成一个reader负责从DataNode读取数据
      reader = BlockReader.newBlockReader(dn, src,
                                          block.getBlock().getBlockId(),
                                          block.getBlock().getGenerationStamp(),
                                          start, len, buffersize,
                                          verifyChecksum, clientName);
      //读取数据
      int nread = reader.readAll(buf, offset, len);
      return;
    } finally {
      IOUtils.closeStream(reader);
      IOUtils.closeSocket(dn);
      dn = null;
    }
    //如果读取失败，则将此DataNode标记为失败节点
    addToDeadNodes(chosenNode);
  }
}
BlockReader.newBlockReader函数实现如下：
 
public static BlockReader newBlockReader( Socket sock, String file,
                                   long blockId,
                                   long genStamp,
                                   long startOffset, long len,
                                   int bufferSize, boolean verifyChecksum,
                                   String clientName)
                                   throws IOException {
  //使用Socket建立写入流，向DataNode发送读指令
  DataOutputStream out = new DataOutputStream(
    new BufferedOutputStream(NetUtils.getOutputStream(sock,HdfsConstants.WRITE_TIMEOUT)));
  out.writeShort( DataTransferProtocol.DATA_TRANSFER_VERSION );
  out.write( DataTransferProtocol.OP_READ_BLOCK );
  out.writeLong( blockId );
  out.writeLong( genStamp );
  out.writeLong( startOffset );
  out.writeLong( len );
  Text.writeString(out, clientName);
  out.flush();
  //使用Socket建立读入流，用于从DataNode读取数据
  DataInputStream in = new DataInputStream(
      new BufferedInputStream(NetUtils.getInputStream(sock),
                              bufferSize));
  DataChecksum checksum = DataChecksum.newDataChecksum( in );
  long firstChunkOffset = in.readLong();
  //生成一个reader，主要包含读入流，用于读取数据
  return new BlockReader( file, blockId, in, checksum, verifyChecksum,
                          startOffset, firstChunkOffset, sock );
}
BlockReader的readAll函数就是用上面生成的DataInputStream读取数据。
2.2、DataNode
在DataNode启动的时候，会调用函数startDataNode，其中与数据读取有关的逻辑如下：
 
void startDataNode(Configuration conf,
                   AbstractList<File> dataDirs
                   ) throws IOException {
  ……
  // 建立一个ServerSocket，并生成一个DataXceiverServer来监控客户端的链接
  ServerSocket ss = (socketWriteTimeout > 0) ?
        ServerSocketChannel.open().socket() : new ServerSocket();
  Server.bind(ss, socAddr, 0);
  ss.setReceiveBufferSize(DEFAULT_DATA_SOCKET_SIZE);
  // adjust machine name with the actual port
  tmpPort = ss.getLocalPort();
  selfAddr = new InetSocketAddress(ss.getInetAddress().getHostAddress(),
                                   tmpPort);
  this.dnRegistration.setName(machineName + ":" + tmpPort);
  this.threadGroup = new ThreadGroup("dataXceiverServer");
  this.dataXceiverServer = new Daemon(threadGroup,
      new DataXceiverServer(ss, conf, this));
  this.threadGroup.setDaemon(true); // auto destroy when empty
  ……
}
DataXceiverServer.run()函数如下：
 
public void run() {
  while (datanode.shouldRun) {
      //接受客户端的链接
      Socket s = ss.accept();
      s.setTcpNoDelay(true);
      //生成一个线程DataXceiver来对建立的链接提供服务
      new Daemon(datanode.threadGroup,
          new DataXceiver(s, datanode, this)).start();
  }
  try {
    ss.close();
  } catch (IOException ie) {
    LOG.warn(datanode.dnRegistration + ":DataXceiveServer: "
                            + StringUtils.stringifyException(ie));
  }
}
DataXceiver.run()函数如下：
 
public void run() {
  DataInputStream in=null;
  try {
    //建立一个输入流，读取客户端发送的指令
    in = new DataInputStream(
        new BufferedInputStream(NetUtils.getInputStream(s),
                                SMALL_BUFFER_SIZE));
    short version = in.readShort();
    boolean local = s.getInetAddress().equals(s.getLocalAddress());
    byte op = in.readByte();
    // Make sure the xciver count is not exceeded
    int curXceiverCount = datanode.getXceiverCount();
    long startTime = DataNode.now();
    switch ( op ) {
    //读取
    case DataTransferProtocol.OP_READ_BLOCK:
      //真正的读取数据
      readBlock( in );
      datanode.myMetrics.readBlockOp.inc(DataNode.now() - startTime);
      if (local)
        datanode.myMetrics.readsFromLocalClient.inc();
      else
        datanode.myMetrics.readsFromRemoteClient.inc();
      break;
    //写入
    case DataTransferProtocol.OP_WRITE_BLOCK:
      //真正的写入数据
      writeBlock( in );
      datanode.myMetrics.writeBlockOp.inc(DataNode.now() - startTime);
      if (local)
        datanode.myMetrics.writesFromLocalClient.inc();
      else
        datanode.myMetrics.writesFromRemoteClient.inc();
      break;
    //其他的指令
    ……
    }
  } catch (Throwable t) {
    LOG.error(datanode.dnRegistration + ":DataXceiver",t);
  } finally {
    IOUtils.closeStream(in);
    IOUtils.closeSocket(s);
    dataXceiverServer.childSockets.remove(s);
  }
}
 
private void readBlock(DataInputStream in) throws IOException {
  //读取指令
  long blockId = in.readLong();         
  Block block = new Block( blockId, 0 , in.readLong());
  long startOffset = in.readLong();
  long length = in.readLong();
  String clientName = Text.readString(in);
  //创建一个写入流，用于向客户端写数据
  OutputStream baseStream = NetUtils.getOutputStream(s,
      datanode.socketWriteTimeout);
  DataOutputStream out = new DataOutputStream(
               new BufferedOutputStream(baseStream, SMALL_BUFFER_SIZE));
  //生成BlockSender用于读取本地的block的数据，并发送给客户端
  //BlockSender有一个成员变量InputStream blockIn用于读取本地block的数据
  BlockSender blockSender = new BlockSender(block, startOffset, length,
          true, true, false, datanode, clientTraceFmt);
   out.writeShort(DataTransferProtocol.OP_STATUS_SUCCESS); // send op status
   //向客户端写入数据
   long read = blockSender.sendBlock(out, baseStream, null);
   ……
  } finally {
    IOUtils.closeStream(out);
    IOUtils.closeStream(blockSender);
  }
}
三、文件的写入
下面解析向hdfs上传一个文件的过程。
3.1、客户端
上传一个文件到hdfs，一般会调用DistributedFileSystem.create，其实现如下：
 
  public FSDataOutputStream create(Path f, FsPermission permission,
    boolean overwrite,
    int bufferSize, short replication, long blockSize,
    Progressable progress) throws IOException {
    return new FSDataOutputStream
       (dfs.create(getPathName(f), permission,
                   overwrite, replication, blockSize, progress, bufferSize),
        statistics);
  }
其最终生成一个FSDataOutputStream用于向新生成的文件中写入数据。其成员变量dfs的类型为DFSClient，DFSClient的create函数如下：
  public OutputStream create(String src,
                             FsPermission permission,
                             boolean overwrite,
                             short replication,
                             long blockSize,
                             Progressable progress,
                             int buffersize
                             ) throws IOException {
    checkOpen();
    if (permission == null) {
      permission = FsPermission.getDefault();
    }
    FsPermission masked = permission.applyUMask(FsPermission.getUMask(conf));
    OutputStream result = new DFSOutputStream(src, masked,
        overwrite, replication, blockSize, progress, buffersize,
        conf.getInt("io.bytes.per.checksum", 512));
    leasechecker.put(src, result);
    return result;
  }
其中构造了一个DFSOutputStream，在其构造函数中，同过RPC调用NameNode的create来创建一个文件。 
当然，构造函数中还做了一件重要的事情，就是streamer.start()，也即启动了一个pipeline，用于写数据，在写入数据的过程中，我们会仔细分析。
  DFSOutputStream(String src, FsPermission masked, boolean overwrite,
      short replication, long blockSize, Progressable progress,
      int buffersize, int bytesPerChecksum) throws IOException {
    this(src, blockSize, progress, bytesPerChecksum);
    computePacketChunkSize(writePacketSize, bytesPerChecksum);
    try {
      namenode.create(
          src, masked, clientName, overwrite, replication, blockSize);
    } catch(RemoteException re) {
      throw re.unwrapRemoteException(AccessControlException.class,
                                     QuotaExceededException.class);
    }
    streamer.start();
  }
 
3.2、NameNode
NameNode的create函数调用namesystem.startFile函数，其又调用startFileInternal函数，实现如下：
  private synchronized void startFileInternal(String src,
                                              PermissionStatus permissions,
                                              String holder,
                                              String clientMachine,
                                              boolean overwrite,
                                              boolean append,
                                              short replication,
                                              long blockSize
                                              ) throws IOException {
    ......
   //创建一个新的文件，状态为under construction，没有任何data block与之对应
   long genstamp = nextGenerationStamp();
   INodeFileUnderConstruction newNode = dir.addFile(src, permissions,
      replication, blockSize, holder, clientMachine, clientNode, genstamp);
   ......
  }
 
3.3、客户端
下面轮到客户端向新创建的文件中写入数据了，一般会使用FSDataOutputStream的write函数，最终会调用DFSOutputStream的writeChunk函数：
按照hdfs的设计，对block的数据写入使用的是pipeline的方式，也即将数据分成一个个的package，如果需要复制三分，分别写入DataNode 1, 2, 3，则会进行如下的过程：
首先将package 1写入DataNode 1
然后由DataNode 1负责将package 1写入DataNode 2，同时客户端可以将pacage 2写入DataNode 1
然后DataNode 2负责将package 1写入DataNode 3, 同时客户端可以讲package 3写入DataNode 1，DataNode 1将package 2写入DataNode 2
就这样将一个个package排着队的传递下去，直到所有的数据全部写入并复制完毕
  protected synchronized void writeChunk(byte[] b, int offset, int len, byte[] checksum)
                                                        throws IOException {
      //创建一个package，并写入数据
      currentPacket = new Packet(packetSize, chunksPerPacket,
                                   bytesCurBlock);
      currentPacket.writeChecksum(checksum, 0, cklen);
      currentPacket.writeData(b, offset, len);
      currentPacket.numChunks++;
      bytesCurBlock += len;
      //如果此package已满，则放入队列中准备发送
      if (currentPacket.numChunks == currentPacket.maxChunks ||
          bytesCurBlock == blockSize) {
          ......
          dataQueue.addLast(currentPacket);
          //唤醒等待dataqueue的传输线程，也即DataStreamer
          dataQueue.notifyAll();
          currentPacket = null;
          ......
      }
  }

DataStreamer的run函数如下：
  public void run() {
    while (!closed && clientRunning) {
      Packet one = null;
      synchronized (dataQueue) {
        //如果队列中没有package，则等待
        while ((!closed && !hasError && clientRunning
               && dataQueue.size() == 0) || doSleep) {
          try {
            dataQueue.wait(1000);
          } catch (InterruptedException  e) {
          }
          doSleep = false;
        }
        try {
          //得到队列中的第一个package
          one = dataQueue.getFirst();
          long offsetInBlock = one.offsetInBlock;
          //由NameNode分配block，并生成一个写入流指向此block
          if (blockStream == null) {
            nodes = nextBlockOutputStream(src);
            response = new ResponseProcessor(nodes);
            response.start();
          }
          ByteBuffer buf = one.getBuffer();
          //将package从dataQueue移至ackQueue,等待确认
          dataQueue.removeFirst();
          dataQueue.notifyAll();
          synchronized (ackQueue) {
            ackQueue.addLast(one);
            ackQueue.notifyAll();
          }
          //利用生成的写入流将数据写入DataNode中的block
          blockStream.write(buf.array(), buf.position(), buf.remaining());
          if (one.lastPacketInBlock) {
            blockStream.writeInt(0); //表示此block写入完毕
          }
          blockStream.flush();
        } catch (Throwable e) {
        }
      }
      ......
  }
 
其中重要的一个函数是nextBlockOutputStream，实现如下：
  private DatanodeInfo[] nextBlockOutputStream(String client) throws IOException {
    LocatedBlock lb = null;
    boolean retry = false;
    DatanodeInfo[] nodes;
    int count = conf.getInt("dfs.client.block.write.retries", 3);
    boolean success;
    do {
      ......
      //由NameNode为文件分配DataNode和block
      lb = locateFollowingBlock(startTime);
      block = lb.getBlock();
      nodes = lb.getLocations();
      //创建向DataNode的写入流
      success = createBlockOutputStream(nodes, clientName, false);
      ......
    } while (retry && --count >= 0);
    return nodes;
  }
 
locateFollowingBlock中通过RPC调用namenode.addBlock(src, clientName)函数
 
3.4、NameNode
NameNode的addBlock函数实现如下：
  public LocatedBlock addBlock(String src,
                               String clientName) throws IOException {
    LocatedBlock locatedBlock = namesystem.getAdditionalBlock(src, clientName);
    return locatedBlock;
  }
FSNamesystem的getAdditionalBlock实现如下：
  public LocatedBlock getAdditionalBlock(String src,
                                         String clientName
                                         ) throws IOException {
    long fileLength, blockSize;
    int replication;
    DatanodeDescriptor clientNode = null;
    Block newBlock = null;
    ......
    //为新的block选择DataNode
    DatanodeDescriptor targets[] = replicator.chooseTarget(replication,
                                                           clientNode,
                                                           null,
                                                           blockSize);
    ......
    //得到文件路径中所有path的INode，其中最后一个是新添加的文件对的INode，状态为under construction
    INode[] pathINodes = dir.getExistingPathINodes(src);
    int inodesLen = pathINodes.length;
    INodeFileUnderConstruction pendingFile  = (INodeFileUnderConstruction)
                                                pathINodes[inodesLen - 1];
    //为文件分配block, 并设置在那写DataNode上
    newBlock = allocateBlock(src, pathINodes);
    pendingFile.setTargets(targets);
    ......
    return new LocatedBlock(newBlock, targets, fileLength);
  }
 
3.5、客户端
在分配了DataNode和block以后，createBlockOutputStream开始写入数据。
  private boolean createBlockOutputStream(DatanodeInfo[] nodes, String client,
                  boolean recoveryFlag) {
      //创建一个socket，链接DataNode
      InetSocketAddress target = NetUtils.createSocketAddr(nodes[0].getName());
      s = socketFactory.createSocket();
      int timeoutValue = 3000 * nodes.length + socketTimeout;
      s.connect(target, timeoutValue);
      s.setSoTimeout(timeoutValue);
      s.setSendBufferSize(DEFAULT_DATA_SOCKET_SIZE);
      long writeTimeout = HdfsConstants.WRITE_TIMEOUT_EXTENSION * nodes.length +
                          datanodeWriteTimeout;
      DataOutputStream out = new DataOutputStream(
          new BufferedOutputStream(NetUtils.getOutputStream(s, writeTimeout),
                                   DataNode.SMALL_BUFFER_SIZE));
      blockReplyStream = new DataInputStream(NetUtils.getInputStream(s));
      //写入指令
      out.writeShort( DataTransferProtocol.DATA_TRANSFER_VERSION );
      out.write( DataTransferProtocol.OP_WRITE_BLOCK );
      out.writeLong( block.getBlockId() );
      out.writeLong( block.getGenerationStamp() );
      out.writeInt( nodes.length );
      out.writeBoolean( recoveryFlag );
      Text.writeString( out, client );
      out.writeBoolean(false);
      out.writeInt( nodes.length - 1 );
      //注意，次循环从1开始，而非从0开始。将除了第一个DataNode以外的另外两个DataNode的信息发送给第一个DataNode, 第一个DataNode可以根据此信息将数据写给另两个DataNode
      for (int i = 1; i < nodes.length; i++) {
        nodes[i].write(out);
      }
      checksum.writeHeader( out );
      out.flush();
      firstBadLink = Text.readString(blockReplyStream);
      if (firstBadLink.length() != 0) {
        throw new IOException("Bad connect ack with firstBadLink " + firstBadLink);
      }
      blockStream = out;
  }
 
客户端在DataStreamer的run函数中创建了写入流后，调用blockStream.write将数据写入DataNode
 
3.6、DataNode
DataNode的DataXceiver中，收到指令DataTransferProtocol.OP_WRITE_BLOCK则调用writeBlock函数：
  private void writeBlock(DataInputStream in) throws IOException {
    DatanodeInfo srcDataNode = null;
    //读入头信息
    Block block = new Block(in.readLong(),
        dataXceiverServer.estimateBlockSize, in.readLong());
    int pipelineSize = in.readInt(); // num of datanodes in entire pipeline
    boolean isRecovery = in.readBoolean(); // is this part of recovery?
    String client = Text.readString(in); // working on behalf of this client
    boolean hasSrcDataNode = in.readBoolean(); // is src node info present
    if (hasSrcDataNode) {
      srcDataNode = new DatanodeInfo();
      srcDataNode.readFields(in);
    }
    int numTargets = in.readInt();
    if (numTargets < 0) {
      throw new IOException("Mislabelled incoming datastream.");
    }
    //读入剩下的DataNode列表，如果当前是第一个DataNode，则此列表中收到的是第二个，第三个DataNode的信息，如果当前是第二个DataNode，则受到的是第三个DataNode的信息
    DatanodeInfo targets[] = new DatanodeInfo[numTargets];
    for (int i = 0; i < targets.length; i++) {
      DatanodeInfo tmp = new DatanodeInfo();
      tmp.readFields(in);
      targets[i] = tmp;
    }
    DataOutputStream mirrorOut = null;  // stream to next target
    DataInputStream mirrorIn = null;    // reply from next target
    DataOutputStream replyOut = null;   // stream to prev target
    Socket mirrorSock = null;           // socket to next target
    BlockReceiver blockReceiver = null; // responsible for data handling
    String mirrorNode = null;           // the name:port of next target
    String firstBadLink = "";           // first datanode that failed in connection setup
    try {
      //生成一个BlockReceiver, 其有成员变量DataInputStream in为从客户端或者上一个DataNode读取数据，还有成员变量DataOutputStream mirrorOut，用于向下一个DataNode写入数据，还有成员变量OutputStream out用于将数据写入本地。
      blockReceiver = new BlockReceiver(block, in,
          s.getRemoteSocketAddress().toString(),
          s.getLocalSocketAddress().toString(),
          isRecovery, client, srcDataNode, datanode);
      // get a connection back to the previous target
      replyOut = new DataOutputStream(
                     NetUtils.getOutputStream(s, datanode.socketWriteTimeout));
      //如果当前不是最后一个DataNode，则同下一个DataNode建立socket连接
      if (targets.length > 0) {
        InetSocketAddress mirrorTarget = null;
        // Connect to backup machine
        mirrorNode = targets[0].getName();
        mirrorTarget = NetUtils.createSocketAddr(mirrorNode);
        mirrorSock = datanode.newSocket();
        int timeoutValue = numTargets * datanode.socketTimeout;
        int writeTimeout = datanode.socketWriteTimeout +
                             (HdfsConstants.WRITE_TIMEOUT_EXTENSION * numTargets);
        mirrorSock.connect(mirrorTarget, timeoutValue);
        mirrorSock.setSoTimeout(timeoutValue);
        mirrorSock.setSendBufferSize(DEFAULT_DATA_SOCKET_SIZE);
        //创建向下一个DataNode写入数据的流
        mirrorOut = new DataOutputStream(
             new BufferedOutputStream(
                         NetUtils.getOutputStream(mirrorSock, writeTimeout),
                         SMALL_BUFFER_SIZE));
        mirrorIn = new DataInputStream(NetUtils.getInputStream(mirrorSock));
        mirrorOut.writeShort( DataTransferProtocol.DATA_TRANSFER_VERSION );
        mirrorOut.write( DataTransferProtocol.OP_WRITE_BLOCK );
        mirrorOut.writeLong( block.getBlockId() );
        mirrorOut.writeLong( block.getGenerationStamp() );
        mirrorOut.writeInt( pipelineSize );
        mirrorOut.writeBoolean( isRecovery );
        Text.writeString( mirrorOut, client );
        mirrorOut.writeBoolean(hasSrcDataNode);
        if (hasSrcDataNode) { // pass src node information
          srcDataNode.write(mirrorOut);
        }
        mirrorOut.writeInt( targets.length - 1 );
        //此出也是从1开始，将除了下一个DataNode的其他DataNode信息发送给下一个DataNode
        for ( int i = 1; i < targets.length; i++ ) {
          targets[i].write( mirrorOut );
        }
        blockReceiver.writeChecksumHeader(mirrorOut);
        mirrorOut.flush();
      }
      //使用BlockReceiver接受block
      String mirrorAddr = (mirrorSock == null) ? null : mirrorNode;
      blockReceiver.receiveBlock(mirrorOut, mirrorIn, replyOut,
                                 mirrorAddr, null, targets.length);
      ......
    } finally {
      // close all opened streams
      IOUtils.closeStream(mirrorOut);
      IOUtils.closeStream(mirrorIn);
      IOUtils.closeStream(replyOut);
      IOUtils.closeSocket(mirrorSock);
      IOUtils.closeStream(blockReceiver);
    }
  }
 
BlockReceiver的receiveBlock函数中，一段重要的逻辑如下：
  void receiveBlock(
      DataOutputStream mirrOut, // output to next datanode
      DataInputStream mirrIn,   // input from next datanode
      DataOutputStream replyOut,  // output to previous datanode
      String mirrAddr, BlockTransferThrottler throttlerArg,
      int numTargets) throws IOException {
      ......
      //不断的接受package，直到结束
      while (receivePacket() > 0) {}
      if (mirrorOut != null) {
        try {
          mirrorOut.writeInt(0); // mark the end of the block
          mirrorOut.flush();
        } catch (IOException e) {
          handleMirrorOutError(e);
        }
      }
      ......
  }
 
BlockReceiver的receivePacket函数如下：
  private int receivePacket() throws IOException {
    //从客户端或者上一个节点接收一个package
    int payloadLen = readNextPacket();
    buf.mark();
    //read the header
    buf.getInt(); // packet length
    offsetInBlock = buf.getLong(); // get offset of packet in block
    long seqno = buf.getLong();    // get seqno
    boolean lastPacketInBlock = (buf.get() != 0);
    int endOfHeader = buf.position();
    buf.reset();
    setBlockPosition(offsetInBlock);
    //将package写入下一个DataNode
    if (mirrorOut != null) {
      try {
        mirrorOut.write(buf.array(), buf.position(), buf.remaining());
        mirrorOut.flush();
      } catch (IOException e) {
        handleMirrorOutError(e);
      }
    }
    buf.position(endOfHeader);       
    int len = buf.getInt();
    offsetInBlock += len;
    int checksumLen = ((len + bytesPerChecksum - 1)/bytesPerChecksum)*
                                                            checksumSize;
    int checksumOff = buf.position();
    int dataOff = checksumOff + checksumLen;
    byte pktBuf[] = buf.array();
    buf.position(buf.limit()); // move to the end of the data.
    ......
    //将数据写入本地的block
    out.write(pktBuf, dataOff, len);
    /// flush entire packet before sending ack
    flush();
    // put in queue for pending acks
    if (responder != null) {
      ((PacketResponder)responder.getRunnable()).enqueue(seqno,
                                      lastPacketInBlock);
    }
    return payloadLen;
  }