hadoop2.7.3源码解析之hdfs删除文件全流程分析

客户端删除文件

先来一段简单的代码,用java的api删除hdfs的 文件

    Configuration conf = new Configuration();    FileSystem fs = FileSystem.get(conf);        Path p = new Path("hdfs://127.0.0.1:9000/demo021.txt");        fs.delete(p,true);        fs.close();// 释放资源        System.out.println("删除成功.....");

namenode删除文件

客户端通过ClientProtocol.delete(String, boolean)方法来删除文件,最终实现是NameNodeRpcServer.delete(String, boolean)方法.

之后调用了FSNamesystem的delete来删除namesystem中的相应的文件.,这里总共分为两步,第一步,从namespace删除相应的文件信息并收集删除的文件的数据块.第二步,将收集到的待删除的数据块加到blockmanage的invalidateBlocks中,等待datanode下次心跳的时候生成删除命令发给datanode,然后删除具体的数据块.

  boolean delete(String src, boolean recursive, boolean logRetryCache)      throws IOException {    waitForLoadingFSImage();    BlocksMapUpdateInfo toRemovedBlocks = null;    writeLock();    boolean ret = false;    try {      //检查是否有写的权限      checkOperation(OperationCategory.WRITE);      //检查是否处于安全模式      checkNameNodeSafeMode("Cannot delete " + src);      //从命名空间删除相应的文件      toRemovedBlocks = FSDirDeleteOp.delete(          this, src, recursive, logRetryCache);      ret = toRemovedBlocks != null;    } catch (AccessControlException e) {      logAuditEvent(false, "delete", src);      throw e;    } finally {      writeUnlock();    }    //将删除操作记录到editlog中    getEditLog().logSync();    if (toRemovedBlocks != null) {     //删除数据块操作      removeBlocks(toRemovedBlocks); // Incremental deletion of blocks    }    logAuditEvent(true, "delete", src);    return ret;  }

从命名空间删除文件

通过工具类FSDirDeleteOp的静态方法delete来删除文件,并且收集该文件的要删除的block.

最终通过FSDirDeleteOp类的unprotectedDelete(FSDirectory, INodesInPath, BlocksMapUpdateInfo, List, long)方法来执行删除操作.之所以叫做unprotectedDelet,是因为这个时候删除只是将该文件从命名空间中删除,并没有真正的写入editlog.

删除过程分为以下几个步骤:
1.检查文件是否存在
2,修改快照记录
3.从namespace中移除文件,也就是FSDirectory记录的INodeDirectory 类型的rootDir中删除;
4.设置父文件夹的最后修改时间
5更新删除的记录数
6 收集要删除的block

  /**   * Delete a path from the name space   * Update the count at each ancestor directory with quota   * @param iip the inodes resolved from the path   * @param collectedBlocks blocks collected from the deleted path   * @param removedINodes inodes that should be removed from inodeMap   * @param mtime the time the inode is removed   * @return the number of inodes deleted; 0 if no inodes are deleted.   */  private static long unprotectedDelete(      FSDirectory fsd, INodesInPath iip, BlocksMapUpdateInfo collectedBlocks,      List<INode> removedINodes, long mtime) {    assert fsd.hasWriteLock();    // check if target node exists    //检查是否存在    INode targetNode = iip.getLastINode();    if (targetNode == null) {      return -1;    }//修改快照    // record modification    final int latestSnapshot = iip.getLatestSnapshotId();    targetNode.recordModification(latestSnapshot);//最核心的代码,从命名空间删除    // Remove the node from the namespace    long removed = fsd.removeLastINode(iip);    if (removed == -1) {      return -1;    }//设置父文件夹的最后修改时间    // set the parent's modification time    final INodeDirectory parent = targetNode.getParent();    parent.updateModificationTime(mtime, latestSnapshot);//更新记录数    fsd.updateCountForDelete(targetNode, iip);    if (removed == 0) {      return 0;    }//收集要删除的block    // collect block and update quota    if (!targetNode.isInLatestSnapshot(latestSnapshot)) {    //收集INodeFile中的blocks变量存放的block信息      targetNode.destroyAndCollectBlocks(fsd.getBlockStoragePolicySuite(),        collectedBlocks, removedINodes);    } else {      QuotaCounts counts = targetNode.cleanSubtree(        fsd.getBlockStoragePolicySuite(), CURRENT_STATE_ID,          latestSnapshot, collectedBlocks, removedINodes);      removed = counts.getNameSpace();      fsd.updateCountNoQuotaCheck(iip, iip.length() -1, counts.negation());    }    if (NameNode.stateChangeLog.isDebugEnabled()) {      NameNode.stateChangeLog.debug("DIR* FSDirectory.unprotectedDelete: "          + iip.getPath() + " is removed");    }    return removed;  }}

将相应的数据块加到InvalidateBlocks中

FSNamesystem的removeBlocks循环刚才收集到的blocks,然后调用blockManager的removeBlock来处理要删除的数据块.

在blockManager的removeBlock中,首先获取到相应的block对应的DatanodeDescriptor,然后将其加到invalidateBlocks里面,然后从blocksMap,corruptReplicas,pendingReplications ,neededReplications 中删除相应的block.

  public void removeBlock(Block block) {    assert namesystem.hasWriteLock();    // No need to ACK blocks that are being removed entirely    // from the namespace, since the removal of the associated    // file already removes them from the block map below.    block.setNumBytes(BlockCommand.NO_ACK);    addToInvalidates(block); //加到invalidateBlocks中    removeBlockFromMap(block);//从blocksMap删除    // Remove the block from pendingReplications and neededReplications    pendingReplications.remove(block);    neededReplications.remove(block, UnderReplicatedBlocks.LEVEL);    if (postponedMisreplicatedBlocks.remove(block)) {      postponedMisreplicatedBlocksCount.decrementAndGet();    }  }

ReplicationMonitor监控线程

BlockManage里面有一个ReplicationMonitor线程,不断的计算块的副本信息和无效的块信息,以便生成相应的命令,等下次心跳的时候传给datanode.在这里我们只是看下相应的删除的方法.

通过run方法我们找到计算无效的块信息的方法computeInvalidateWork,在这里会循环invalidateBlocks中的所有datanode,然后循环调用invalidateWorkForOneNode方法一个一个的datanode来处理.

在invalidateWorkForOneNode中,首先将相应的datanode从invalidateBlocks中删除,然后调用invalidateBlocks.invalidateWork将该DatanodeDescriptor相应的无效的块加到DatanodeDescriptor类中LightWeightHashSet类型的变量invalidateBlocks中,等待下次心跳生成删除命令.

心跳生成删除命令

具体生成删除相关命令的代码在以下方法中,DatanodeManager.handleHeartbeat(DatanodeRegistration, StorageReport[], String, long, long, int, int, int, VolumeFailureSummary).

        //check block invalidation        Block[] blks = nodeinfo.getInvalidateBlocks(blockInvalidateLimit);        if (blks != null) {          cmds.add(new BlockCommand(DatanodeProtocol.DNA_INVALIDATE,              blockPoolId, blks));        }

有关hdfs心跳相关的信息请参考
http://blog.csdn.net/zhangjun5965/article/details/75579238

datanode删除相应的block并汇报

心跳处理删除命令

datanode方面是通过BPServiceActor的offerService方法进行心跳相关的操作,报告心跳之后,会依次处理从namenode接收的命令,最终处理的方法落在BPOfferService.processCommandFromActive(DatanodeCommand, BPServiceActor)方法上.

在这个方法中,通过switch来判断传过来的是哪种命令,来分别进行处理,删除数据块对应的是DatanodeProtocol.DNA_INVALIDATE,最终进入了FsDatasetImpl.invalidate(String, Block[])方法来从磁盘删除具体的数据块.

异步单独开启线程删除磁盘数据

具体的操作方法是调用了asyncDiskService.deleteAsync异步的开启线程删除数据块,以便提高效率.

      // Delete the block asynchronously to make sure we can do it fast enough.      // It's ok to unlink the block file before the uncache operation      // finishes.      try {        asyncDiskService.deleteAsync(v.obtainReference(), f,            FsDatasetUtil.getMetaFile(f, invalidBlks[i].getGenerationStamp()),            new ExtendedBlock(bpid, invalidBlks[i]),            dataStorage.getTrashDirectoryForBlockFile(bpid, f));      } catch (ClosedChannelException e) {        LOG.warn("Volume " + v + " is closed, ignore the deletion task for " +            "block " + invalidBlks[i]);      }

多线程删除具体是开启了一个ReplicaFileDeleteTask线程来做删除的操作,这个方法会先删除数据块信息和meta信息,删除之后调用 datanode.notifyNamenodeDeletedBlock(block, volume.getStorageID());向namenode报告最近删除的数据块.
但是这个时候并不是将这些信息直接发给namenode,而是要删除的blocks和其对应的DatanodeStorage生成ReceivedDeletedBlockInfo对象存在了BPServiceActor的Map

向namenode汇报删除的块

在BPServiceActor的心跳处理方法offerService中,会通过 reportReceivedDeletedBlocks();读取pendingIncrementalBRperStorage变量中的blocks信息,向namenode汇报刚刚删除的数据块信息.

        if (sendImmediateIBR ||            (startTime - lastDeletedReport > dnConf.deleteReportInterval)) {          reportReceivedDeletedBlocks();          lastDeletedReport = startTime;        }

最终通过datanode和namenode之间的协议DatanodeProtocol.blockReceivedAndDeleted(DatanodeRegistration, String, StorageReceivedDeletedBlocks[])来向namenode报告刚才删除的数据块.

namenode处理删除block的汇报

namenode处理最近删除的块的方法是在NameNodeRpcServer的同民个的方法blockReceivedAndDeleted中,通过跟踪代码,最终到了BlockManager.removeStoredBlock(Block, DatanodeDescriptor)中.

首先从blocksMap中移除相应的块信息,然后判断是否是因为datanode挂掉而导致的block被移除,并做相应的处理,然后从excessReplicateMap,corruptReplicas队列中将其移除.

  /**   * Modify (block-->datanode) map. Possibly generate replication tasks, if the   * removed block is still valid.   */  public void removeStoredBlock(Block block, DatanodeDescriptor node) {    blockLog.debug("BLOCK* removeStoredBlock: {} from {}", block, node);    assert (namesystem.hasWriteLock());    {      //从blocksMap移除      if (!blocksMap.removeNode(block, node)) {        blockLog.debug("BLOCK* removeStoredBlock: {} has already been" +            " removed from node {}", block, node);        return;      }      //判断是否是因为datanode失败而移除的数据块,如果block仍然有效,检查副本是不是必要的,在这种情况下,需要将block加到待复制的block列表中.      // It's possible that the block was removed because of a datanode      // failure. If the block is still valid, check if replication is      // necessary. In that case, put block on a possibly-will-      // be-replicated list.      //      BlockCollection bc = blocksMap.getBlockCollection(block);      if (bc != null) {        namesystem.decrementSafeBlockCount(block);        updateNeededReplications(block, -1, 0);      }      //      // We've removed a block from a node, so it's definitely no longer      // in "excess" there.      //从excessReplicateMap移除      LightWeightLinkedSet<Block> excessBlocks = excessReplicateMap.get(node          .getDatanodeUuid());      if (excessBlocks != null) {        if (excessBlocks.remove(block)) {          excessBlocksCount.decrementAndGet();          blockLog.debug("BLOCK* removeStoredBlock: {} is removed from " +              "excessBlocks", block);          if (excessBlocks.size() == 0) {            excessReplicateMap.remove(node.getDatanodeUuid());          }        }      }       //从corruptReplicas移除      // Remove the replica from corruptReplicas      corruptReplicas.removeFromCorruptReplicasMap(block, node);    }  }

总结

上述只是基于hadoop2.7.3源码自己的一些学习笔记,如有不对的地方,还请见谅