[HADOOP-11957] if an IOException error is thrown in DomainSocket.close we go into infinite loop.

• Type: Bug
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.7.1
• Component/s: net
• Target Version/s: 2.7.8
• Assignee: Anu Engineer
• Reporter: Anu Engineer
• Created: 11/May/15 18:51
• Updated: 02/Jun/18 17:40

Description

if an IOException error is thrown in DomainSocket.close we go into infinite loop.

Issue : If the shutdown0(fd) call throws an IOException we break out of the if shutdown call but will continue to loop in the while loop infinitely since we have no way of decrementing the counter. Please scroll down and see the comment marked with Bug Bug to see where the issue is.

DomainSocket.java

  @Override
  public void close() throws IOException {
    // Set the closed bit on this DomainSocket
    int count = 0;
    try {
      count = refCount.setClosed();
    } catch (ClosedChannelException e) {
      // Someone else already closed the DomainSocket.
      return;
    }
    // Wait for all references to go away
    boolean didShutdown = false;
    boolean interrupted = false;
    while (count > 0) {
      if (!didShutdown) {
        try {
          // Calling shutdown on the socket will interrupt blocking system
          // calls like accept, write, and read that are going on in a
          // different thread.
          shutdown0(fd);
        } catch (IOException e) {
          LOG.error("shutdown error: ", e);
        }
        didShutdown = true; 

        // *BUG BUG* <-- Here the code will never exit the loop
        // if the count is greater then 0. we need to break out
        // of the while loop in case of IOException Error

      }
      try {
        Thread.sleep(10);
      } catch (InterruptedException e) {
        interrupted = true;
      }
      count = refCount.getReferenceCount();
    }

    // At this point, nobody has a reference to the file descriptor, 
    // and nobody will be able to get one in the future either.
    // We now call close(2) on the file descriptor.
    // After this point, the file descriptor number will be reused by 
    // something else.  Although this DomainSocket object continues to hold 
    // the old file descriptor number (it's a final field), we never use it 
    // again because this DomainSocket is closed.
    close0(fd);
    if (interrupted) {
      Thread.currentThread().interrupt();
    }
  }

Issue Links

• blocks: HDFS-3296 Running libhdfs tests in mac fails