[HADOOP-14214] DomainSocketWatcher::add()/delete() should not self interrupt while looping await()

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 2.9.0, 2.7.4, 3.0.0-alpha4, 2.8.2
• Component/s: hdfs-client
• Target Version/s: 2.8.1, 3.0.0-alpha4
• Hadoop Flags: Reviewed
• Assignee: Mingliang Liu
• Reporter: Mingliang Liu
• Created: 22/Mar/17 20:08
• Updated: 26/Jul/17 14:46
• Resolved: 23/Mar/17 20:47

Description

Our hive team found a TPCDS job whose queries running on LLAP seem to be getting stuck. Dozens of threads were waiting for the DfsClientShmManager::lock, as following jstack:

Thread 251 (IO-Elevator-Thread-5):
  State: WAITING
  Blocked count: 3871
  Wtaited count: 4565
  Waiting on java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject@16ead198
  Stack:
    sun.misc.Unsafe.park(Native Method)
    java.util.concurrent.locks.LockSupport.park(LockSupport.java:175)
    java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitUninterruptibly(AbstractQueuedSynchronizer.java:1976)
    org.apache.hadoop.hdfs.shortcircuit.DfsClientShmManager$EndpointShmManager.allocSlot(DfsClientShmManager.java:255)
    org.apache.hadoop.hdfs.shortcircuit.DfsClientShmManager.allocSlot(DfsClientShmManager.java:434)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.allocShmSlot(ShortCircuitCache.java:1017)
    org.apache.hadoop.hdfs.BlockReaderFactory.createShortCircuitReplicaInfo(BlockReaderFactory.java:476)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.create(ShortCircuitCache.java:784)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.fetchOrCreate(ShortCircuitCache.java:718)
    org.apache.hadoop.hdfs.BlockReaderFactory.getBlockReaderLocal(BlockReaderFactory.java:422)
    org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:333)
    org.apache.hadoop.hdfs.DFSInputStream.actualGetFromOneDataNode(DFSInputStream.java:1181)
    org.apache.hadoop.hdfs.DFSInputStream.fetchBlockByteRange(DFSInputStream.java:1118)
    org.apache.hadoop.hdfs.DFSInputStream.pread(DFSInputStream.java:1478)
    org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:1441)
    org.apache.hadoop.fs.FSInputStream.readFully(FSInputStream.java:121)
    org.apache.hadoop.fs.FSDataInputStream.readFully(FSDataInputStream.java:111)
    org.apache.orc.impl.RecordReaderUtils$DefaultDataReader.readStripeFooter(RecordReaderUtils.java:166)
    org.apache.hadoop.hive.llap.io.metadata.OrcStripeMetadata.<init>(OrcStripeMetadata.java:64)
    org.apache.hadoop.hive.llap.io.encoded.OrcEncodedDataReader.readStripesMetadata(OrcEncodedDataReader.java:622)

The thread that is expected to signal those threads is calling DomainSocketWatcher::add() method, but it gets stuck there dealing with InterruptedException infinitely. The jstack is like:

Thread 44417 (TezTR-257387_2840_12_10_52_0):
  State: RUNNABLE
  Blocked count: 3
  Wtaited count: 5
  Stack:
    java.lang.Throwable.fillInStackTrace(Native Method)
    java.lang.Throwable.fillInStackTrace(Throwable.java:783)
    java.lang.Throwable.<init>(Throwable.java:250)
    java.lang.Exception.<init>(Exception.java:54)
    java.lang.InterruptedException.<init>(InterruptedException.java:57)
    java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:2034)
    org.apache.hadoop.net.unix.DomainSocketWatcher.add(DomainSocketWatcher.java:325)
    org.apache.hadoop.hdfs.shortcircuit.DfsClientShmManager$EndpointShmManager.allocSlot(DfsClientShmManager.java:266)
    org.apache.hadoop.hdfs.shortcircuit.DfsClientShmManager.allocSlot(DfsClientShmManager.java:434)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.allocShmSlot(ShortCircuitCache.java:1017)
    org.apache.hadoop.hdfs.BlockReaderFactory.createShortCircuitReplicaInfo(BlockReaderFactory.java:476)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.create(ShortCircuitCache.java:784)
    org.apache.hadoop.hdfs.shortcircuit.ShortCircuitCache.fetchOrCreate(ShortCircuitCache.java:718)
    org.apache.hadoop.hdfs.BlockReaderFactory.getBlockReaderLocal(BlockReaderFactory.java:422)
    org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:333)
    org.apache.hadoop.hdfs.DFSInputStream.actualGetFromOneDataNode(DFSInputStream.java:1181)
    org.apache.hadoop.hdfs.DFSInputStream.fetchBlockByteRange(DFSInputStream.java:1118)
    org.apache.hadoop.hdfs.DFSInputStream.pread(DFSInputStream.java:1478)
    org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:1441)
    org.apache.hadoop.fs.FSInputStream.readFully(FSInputStream.java:121)

The whole job makes no progress because of this.

The thread in DomainSocketWatcher::add() is expected to eventually break the while loop where it waits for the newly added entry being deleted by another thread. However, if this thread is ever interrupted, chances are that it will hold the lock forever so if(!toAdd.contains(entry)) always be false.

DomainSocketWatcher::add()

  public void add(DomainSocket sock, Handler handler) {
    lock.lock();
    try {
      ......
      toAdd.add(entry);
      kick();
      while (true) {
        try {
          processedCond.await();
        } catch (InterruptedException e) {
          Thread.currentThread().interrupt();
        }
        if (!toAdd.contains(entry)) {
          break;
        }
      }
    } finally {
      lock.unlock();
    }
  }

The reason here is that, this method catches the InterruptedException and self interrupts during await(). The await() method internally calls AbstractQueuedSynchronizer::await(), which will throw a new InterruptedException if it's interrupted.

AbstractQueuedSynchronizer::await()

        public final void await() throws InterruptedException {
            if (Thread.interrupted())
                throw new InterruptedException();
            Node node = addConditionWaiter();
            ...

Our code in DomainSocketWatcher::add() catches this exception (again) and self interrupt (again). Please note in this process, the associated lock is never released so that the other thread which is supposed to make if(!toAdd.contains(entry)) be true is still pending on the lock.

DomainSocketWatcher::delete() has similar code logic and should suffer from similar problems.

Thanks jdere for testing and reporting this.

Issue Links

• is related to: FLINK-3466 Job might get stuck in restoreState() from HDFS due to interrupt