[HADOOP-12532] Data race in IPC client Client.stop()

• Type: Bug
• Status: In Progress
• Priority: Major
• Resolution: Unresolved
• Component/s: ipc
• Assignee: Wei-Chiu Chuang
• Reporter: Wei-Chiu Chuang
• Created: 30/Oct/15 17:02
• Updated: 30/Oct/15 22:47

Description

I found a data race in ipc.Client.stop()

ipc.Client maintains a hash map of connection threads. When stop() is called, it interrupts all connection threads; the threads are supposed to remove itself from the hash map as part of the clean up work; and stop() periodically checks to see if the hash map is empty and then returns.

The bug is, this checking operation is not synchronized, and the connection thread actually removes itself from the hash map before terminating connections.

This bug causes regression for HDFS-4925. In fact, the fix in HDFS-4925 may not be correct, because it assumes when it returns from QuorumJournalManager.close(), IPC client connection threads are terminated. But the reality is the IPC code assumes connections are closed, not the IPC connection threads (which in any case is buggy as well).

This is also likely related to the bug reported in HDFS-4925 (TestQuorumJournalManager.testPurgeLogs intermittently Fails assertNoThreadsMatching)

Issue Links

• blocks: HDFS-9347 Invariant assumption in TestQuorumJournalManager.shutdown() is wrong
• breaks: HDFS-4925 TestQuorumJournalManager.testPurgeLogs intermittently Fails assertNoThreadsMatching