Prakash Saraf created AMQ-7016:
----------------------------------
Summary: Deadlock in KahaDB journal caused during rotation and
getting current data file
Key: AMQ-7016
URL: https://issues.apache.org/jira/browse/AMQ-7016
Project: ActiveMQ
Issue Type: Bug
Components: KahaDB
Affects Versions: 5.14.0
Environment: Apache AMQ 5.14.0 with Java Spring.
Reporter: Prakash Saraf
Attachments: amq_deadlock.log
Hi,
We encountered a deadlock in AMQ. From the thread analysis it looks the
deadlock happened during one thread doing reservation of journal file (journal
compaction thread), while other was getting the current data file (general data
appender).
*Thread 1 (any of thread that add/remove mesg):*
The getCurrentDataFile(int capacity) method, first locks the currentDataFile
and then does rotation, if current capacity is not enough. The rotateWriteFile
first locks the fileId and then the current file.
Locks in sequence -> currentDataFile -> dataFileIdlock -> currentDataFile
*Thread 2:*
The reserveDataFile method first locks dataFileIdlock and then currentDataFile.
Locks in sequence -> dataFileIdlock -> currentDataFile
Please also see the thread dumps attached.
Shouldn't we always first lock the currentDataFile and then dataFildIdLock? Is
there a potential reason, why the reserveDataFile and rotateWriteFile first
locks fileId and then currentDataFile?
Can we also consider using read/write locks rather complete object lock?
Please let me know if any other details are required.
Thanks,
Prakash Saraf
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
