[
https://issues.apache.org/jira/browse/HADOOP-6269?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Todd Lipcon updated HADOOP-6269:
--------------------------------
Attachment: hadoop-6269.txt
OK. Here's the summary of a correct (I believe) fix:
- To avoid deadlock we need a lock ordering
- We'll order instance locks before class locks (ie you must never lock an
instance when holding a class lock)
The issue is that addDefaultResource is currently static synchronized and locks
a class instance by calling conf.reloadConfiguration (synchronized non-static)
Attaching a patch which breaks this synchronization by holding the lock only
long enough to make a defensive copy of the shared list. I don't think this is
a performance issue since addDefaultResource is called very rarely.
Tests pass. Will test on the application that keeps turning up this bug here.
Review would be appreciated.
> Missing synchronization for defaultResources in Configuration.addResource
> -------------------------------------------------------------------------
>
> Key: HADOOP-6269
> URL: https://issues.apache.org/jira/browse/HADOOP-6269
> Project: Hadoop Common
> Issue Type: Bug
> Components: conf
> Affects Versions: 0.20.1
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Fix For: 0.20.2, 0.21.0
>
> Attachments: hadoop-6269.txt, hadoop-6269.txt
>
>
> Configuration.defaultResources is a simple ArrayList. In two places in
> Configuration it is accessed without appropriate synchronization, which we've
> seen to occasionally result in ConcurrentModificationExceptions.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
