Hadoop之HDFS原理及文件上传下载源码分析（下）

Posted 2020-09-12 Devil丶俊锅

　　上篇Hadoop之HDFS原理及文件上传下载源码分析（上）楼主主要介绍了hdfs原理及FileSystem的初始化源码解析， Client如何与NameNode建立RPC通信。本篇将继续介绍hdfs文件上传、下载源解析。

文件上传

　　先上文件上传的方法调用过程时序图：

　　

　　

 　　其主要执行过程：

1. 　　 FileSystem初始化，Client拿到NameNodeRpcServer代理对象，建立与NameNode的RPC通信（楼主上篇已经介绍过了）
2. 　　 调用FileSystem的create()方法，由于实现类为DistributedFileSystem,所有是调用该类中的create()方法
3. 　　 DistributedFileSystem持有DFSClient的引用，继续调用DFSClient中的create()方法
4. 　　 DFSOutputStream提供的静态newStreamForCreate()方法中调用NameNodeRpcServer服务端的create()方法并创建DFSOutputStream输出流对象返回
5. 　　 通过hadoop提供的IOUtil工具类将输出流输出到本地

　　下面我们来看下源码：

　　首先初始化文件系统，建立与服务端的RPC通信

　　

1 HDFSDemo.java
2 OutputStream os = fs.create(new Path("/test.log"));

 

　　调用FileSystem的create()方法，由于FileSystem是一个抽象类，这里实际上是调用的该类的子类create()方法

　　

1  //FileSystem.java
2 public abstract FSDataOutputStream create(Path f,
3       FsPermission permission,
4       boolean overwrite,
5       int bufferSize,
6       short replication,
7       long blockSize,
8       Progressable progress) throws IOException;

　　　前面我们已经说过FileSystem.get()返回的是DistributedFileSystem对象，所以这里我们直接进入DistributedFileSystem：

　

　　

 1   //DistributedFileSystem.java
 2 @Override
 3   public FSDataOutputStream create(final Path f, final FsPermission permission,
 4     final EnumSet<CreateFlag> cflags, final int bufferSize,
 5     final short replication, final long blockSize, final Progressable progress,
 6     final ChecksumOpt checksumOpt) throws IOException {
 7     statistics.incrementWriteOps(1);
 8     Path absF = fixRelativePart(f);
 9     return new FileSystemLinkResolver<FSDataOutputStream>() {
10       @Override
11       public FSDataOutputStream doCall(final Path p)
12           throws IOException, UnresolvedLinkException {
13         final DFSOutputStream dfsos = dfs.create(getPathName(p), permission,
14                 cflags, replication, blockSize, progress, bufferSize,
15                 checksumOpt);
16         //dfs为DistributedFileSystem所持有的DFSClient对象，这里调用DFSClient中的create()方法
17         return dfs.createWrappedOutputStream(dfsos, statistics);
18       }
19       @Override
20       public FSDataOutputStream next(final FileSystem fs, final Path p)
21           throws IOException {
22         return fs.create(p, permission, cflags, bufferSize,
23             replication, blockSize, progress, checksumOpt);
24       }
25     }.resolve(this, absF);
26   }

　　DFSClient的create()返回一个DFSOutputStream对象：

　　

 1  //DFSClient.java
 2 public DFSOutputStream create(String src, 
 3                              FsPermission permission,
 4                              EnumSet<CreateFlag> flag, 
 5                              boolean createParent,
 6                              short replication,
 7                              long blockSize,
 8                              Progressable progress,
 9                              int buffersize,
10                              ChecksumOpt checksumOpt,
11                              InetSocketAddress[] favoredNodes) throws IOException {
12     checkOpen();
13     if (permission == null) {
14       permission = FsPermission.getFileDefault();
15     }
16     FsPermission masked = permission.applyUMask(dfsClientConf.uMask);
17     if(LOG.isDebugEnabled()) {
18       LOG.debug(src + ": masked=" + masked);
19     }
20     //调用DFSOutputStream的静态方法newStreamForCreate，返回输出流
21     final DFSOutputStream result = DFSOutputStream.newStreamForCreate(this,
22         src, masked, flag, createParent, replication, blockSize, progress,
23         buffersize, dfsClientConf.createChecksum(checksumOpt),
24         getFavoredNodesStr(favoredNodes));
25     beginFileLease(result.getFileId(), result);
26     return result;
27   }

　　我们继续看下newStreamForCreate()中的业务逻辑：

　　

 1 //DFSOutputStream.java
 2  static DFSOutputStream newStreamForCreate(DFSClient dfsClient, String src,
 3       FsPermission masked, EnumSet<CreateFlag> flag, boolean createParent,
 4       short replication, long blockSize, Progressable progress, int buffersize,
 5       DataChecksum checksum, String[] favoredNodes) throws IOException {
 6     TraceScope scope =
 7         dfsClient.getPathTraceScope("newStreamForCreate", src);
 8     try {
 9       HdfsFileStatus stat = null;
10       boolean shouldRetry = true;
11       int retryCount = CREATE_RETRY_COUNT;
12       while (shouldRetry) {
13         shouldRetry = false;
14         try {
15           //这里通过dfsClient的NameNode代理对象调用NameNodeRpcServer中实现的create()方法
16           stat = dfsClient.namenode.create(src, masked, dfsClient.clientName,
17               new EnumSetWritable<CreateFlag>(flag), createParent, replication,
18               blockSize, SUPPORTED_CRYPTO_VERSIONS);
19           break;
20         } catch (RemoteException re) {
21           IOException e = re.unwrapRemoteException(
22               AccessControlException.class,
23               DSQuotaExceededException.class,
24               FileAlreadyExistsException.class,
25               FileNotFoundException.class,
26               ParentNotDirectoryException.class,
27               NSQuotaExceededException.class,
28               RetryStartFileException.class,
29               SafeModeException.class,
30               UnresolvedPathException.class,
31               SnapshotAccessControlException.class,
32               UnknownCryptoProtocolVersionException.class);
33           if (e instanceof RetryStartFileException) {
34             if (retryCount > 0) {
35               shouldRetry = true;
36               retryCount--;
37             } else {
38               throw new IOException("Too many retries because of encryption" +
39                   " zone operations", e);
40             }
41           } else {
42             throw e;
43           }
44         }
45       }
46       Preconditions.checkNotNull(stat, "HdfsFileStatus should not be null!");
47      //new输出流对象
48       final DFSOutputStream out = new DFSOutputStream(dfsClient, src, stat,
49           flag, progress, checksum, favoredNodes);
50       out.start();//调用内部类DataStreamer的start()方法，DataStreamer继承Thread，所以说这是一个线程，从NameNode中申请新的block信息；
　　　　　　　　　　　　　　　　同时前面我们介绍hdfs原理的时候提到的流水线作业（Pipeline）也是在这里实现，有兴趣的同学可以去研究下，这里就不带大家看了
51       return out;
52     } finally {
53       scope.close();
54     }
55   }

　　　　

　　到此，Client拿到了服务端的输出流对象，那么后面就容易了，都是一些简答的文件输出，输入流的操作（hadoop提供的IOUitl）。

文件下载

　　文件上传的大致流程与文件下载类似，与上传一样，我们先上程序方法调用时序图：

　　

　　主要执行过程：　　

1. 　　 FileSystem初始化，Client拿到NameNodeRpcServer代理对象，建立与NameNode的RPC通信（与前面一样）
2. 　　 调用FileSystem的open()方法，由于实现类为DistributedFileSystem,所有是调用该类中的open()方法
3. 　　 DistributedFileSystem持有DFSClient的引用，继续调用DFSClient中的open()方法
4. 　　 实例化DFSInputStream输入流
5. 　　 调用openinfo()方法
6. 　　 调用fetchLocatedBlocksAndGetLastBlockLength()方法，抓取block信息并获取最后block长度
7.        调用DFSClient中的getLocatedBlocks()方法，获取block信息
8. 　　 在callGetBlockLocations()方法中通过NameNode代理对象调用NameNodeRpcServer的getBlockLocations()方法
9.        将block信息写入输出流
10.        交给IOUtil,下载文件到本地

　　接下来，我们开始看源码：

　　首先任然是FileSystem的初始化，前面有，这里就不贴出来了，我们直接从DistributedFileSystem的open()开始看。

　　

 1 //DistributedFifeSystem.java
 2 @Override
 3   public FSDataInputStream open(Path f, final int bufferSize)
 4       throws IOException {
 5     statistics.incrementReadOps(1);
 6     Path absF = fixRelativePart(f);
 7     return new FileSystemLinkResolver<FSDataInputStream>() {
 8       @Override
 9       public FSDataInputStream doCall(final Path p)
10           throws IOException, UnresolvedLinkException {
11         final DFSInputStream dfsis =
12           dfs.open(getPathName(p), bufferSize, verifyChecksum);
13         //dfs为DFSClient对象，调用open()返回输入流
14         return dfs.createWrappedInputStream(dfsis);
15       }
16       @Override
17       public FSDataInputStream next(final FileSystem fs, final Path p)
18           throws IOException {
19         return fs.open(p, bufferSize);
20       }
21     }.resolve(this, absF);
22   }

　　DFSClient中并没有直接使用NameNode的代理对象，而是传给了DFSInputStream：

　　

 1 //DFSClient.java
 2 public DFSInputStream open(String src, int buffersize, boolean verifyChecksum)
 3       throws IOException, UnresolvedLinkException {
 4     checkOpen();   
 5     TraceScope scope = getPathTraceScope("newDFSInputStream", src);
 6     try {
 7       //这里并没有直接通过NameNode的代理对象调用服务端的方法，直接new输入流并把当前对象作为参数传入
 8       return new DFSInputStream(this, src, verifyChecksum);
 9     } finally {
10       scope.close();
11     }
12   }

　　那么在DFSInputStream必须持有DFSClient的引用：

　　

 1 //DFSInputStream.java 构造
 2 DFSInputStream(DFSClient dfsClient, String src, boolean verifyChecksum
 3                  ) throws IOException, UnresolvedLinkException {
 4     this.dfsClient = dfsClient;//只有DFSClient的引用
 5     this.verifyChecksum = verifyChecksum;
 6     this.src = src;
 7     synchronized (infoLock) {
 8       this.cachingStrategy = dfsClient.getDefaultReadCachingStrategy();
 9     }
10     openInfo();//调openInfo()
11   }

　　openInfo()用来抓取block信息：

 1 void openInfo() throws IOException, UnresolvedLinkException {
 2     synchronized(infoLock) {
 3       lastBlockBeingWrittenLength = fetchLocatedBlocksAndGetLastBlockLength();//抓取block信息
 4       int retriesForLastBlockLength = dfsClient.getConf().retryTimesForGetLastBlockLength;//获取配置信息，尝试抓取的次数，楼主记得在2.6以前这里写的3；当然，现在的默认值也为3
 5       while (retriesForLastBlockLength > 0) {
 6         if (lastBlockBeingWrittenLength == -1) {
 7           DFSClient.LOG.warn("Last block locations not available. "
 8               + "Datanodes might not have reported blocks completely."
 9               + " Will retry for " + retriesForLastBlockLength + " times");
10           waitFor(dfsClient.getConf().retryIntervalForGetLastBlockLength);
11           lastBlockBeingWrittenLength = fetchLocatedBlocksAndGetLastBlockLength();
12         } else {
13           break;
14         }
15         retriesForLastBlockLength--;
16       }
17       if (retriesForLastBlockLength == 0) {
18         throw new IOException("Could not obtain the last block locations.");
19       }
20     }
21   }

 

　　获取block信息：

 1 //DFSInputStream.java
 2 private long fetchLocatedBlocksAndGetLastBlockLength() throws IOException {
 3     final LocatedBlocks newInfo = dfsClient.getLocatedBlocks(src, 0);
 4     //回到DFSClient中来获取当前block信息
 5     if (DFSClient.LOG.isDebugEnabled()) {
 6       DFSClient.LOG.debug("newInfo = " + newInfo);
 7     }
 8     if (newInfo == null) {
 9       throw new IOException("Cannot open filename " + src);
10     }
11 
12     if (locatedBlocks != null) {
13       Iterator<LocatedBlock> oldIter = locatedBlocks.getLocatedBlocks().iterator();
14       Iterator<LocatedBlock> newIter = newInfo.getLocatedBlocks().iterator();
15       while (oldIter.hasNext() && newIter.hasNext()) {
16         if (! oldIter.next().getBlock().equals(newIter.next().getBlock())) {
17           throw new IOException("Blocklist for " + src + " has changed!");
18         }
19       }
20     }
21     locatedBlocks = newInfo;
22     long lastBlockBeingWrittenLength = 0;
23     if (!locatedBlocks.isLastBlockComplete()) {
24       final LocatedBlock last = locatedBlocks.getLastLocatedBlock();
25       if (last != null) {
26         if (last.getLocations().length == 0) {
27           if (last.getBlockSize() == 0) {         
28             return 0;
29           }
30           return -1;
31         }
32         final long len = readBlockLength(last);
33         last.getBlock().setNumBytes(len);
34         lastBlockBeingWrittenLength = len; 
35       }
36     }
37 
38     fileEncryptionInfo = locatedBlocks.getFileEncryptionInfo();
39     //返回block开始写的位置
40     return lastBlockBeingWrittenLength;
41   }

　　回到DFSClient中：

　　

 1 DFSClient.java
 2 @VisibleForTesting
 3   public LocatedBlocks getLocatedBlocks(String src, long start, long length)
 4       throws IOException {
 5     TraceScope scope = getPathTraceScope("getBlockLocations", src);
 6     try {
 7       //这里NameNode作为参数传递到callGetBlockLocations()中
 8       return callGetBlockLocations(namenode, src, start, length);
 9     } finally {
10       scope.close();
11     }
12   }

　　调用服务端方法，返回block信息：

 1 //DFSClient.java
 2 static LocatedBlocks callGetBlockLocations(ClientProtocol namenode,
 3       String src, long start, long length) 
 4       throws IOException {
 5     try {
 6      //看到这里，不用做过多的解释了吧？
 7       return namenode.getBlockLocations(src, start, length);
 8     } catch(RemoteException re) {
 9       throw re.unwrapRemoteException(AccessControlException.class,
10                                      FileNotFoundException.class,
11                                      UnresolvedPathException.class);
12     }
13   }

　　最终将文件block相关信息写入输入流，通过工具类IOUtil输出到本地文件。

　　那关于hadoop之hdfs原理及文件上传下载源码解析就写到这里，下系列的文章，楼主会写一些关于mapreduce或者hive相关的文章分享给大家。

　　示例代码地址：https://github.com/LJunChina/hadoop