Mailing-List: contact dev-help@jackrabbit.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@jackrabbit.apache.org
Message-ID: <1110166108.1236163556942.JavaMail.jira@brutus>
Date: Wed, 4 Mar 2009 02:45:56 -0800 (PST)
From: "Jukka Zitting (JIRA)" <jira@apache.org>
To: dev@jackrabbit.apache.org
Subject: [jira] Commented: (JCR-2000) Deadlock on concurrent commits
In-Reply-To: <1999483538.1235576941753.JavaMail.jira@brutus>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/JCR-2000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12678683#action_12678683 ] 

Jukka Zitting commented on JCR-2000:
------------------------------------

The thread deaths I saw were caused by the RandomOperationTest class killing them. At least in a few cases the threads were still holding on locks, which made the test block indefinitely in the cleanup phase after all test threads had been killed.

Why do we need the thread killing in the first place? A normal join() call with no timeout should be fine enough. If there's a deadlock, then a manually generated thread dump is much more accurate (it contains all the locks held) than the stack traces that the test case now dumps.

> Hmm, I understand, but the above mentioned test works just fine in the 1.4 branch.

The proposed change here extends the scope of the versioning lock acquired in a transaction commit, which makes this deadlock (A: commit() -> versioning lock -> workspace lock; B: save() -> workspace lock -> versioning lock) more likely to happen, but the deadlock scenario already existed before, it probably just was never triggered due to lucky timing.

> Deadlock on concurrent commits
> ------------------------------
>
>                 Key: JCR-2000
>                 URL: https://issues.apache.org/jira/browse/JCR-2000
>             Project: Jackrabbit Content Repository
>          Issue Type: Bug
>          Components: jackrabbit-core, transactions
>    Affects Versions: 1.5.3
>            Reporter: Jukka Zitting
>            Assignee: Jukka Zitting
>             Fix For: 1.5.4
>
>         Attachments: JCR-2000.patch, JCR-2000.patch
>
>
> As reported in the followup to JCR-1979, there's a case where two transactions may be concurrently inside a commit. This is bad as it breaks the main assumption in http://jackrabbit.apache.org/concurrency-control.html about all transactions first acquiring the versioning write lock.
> Looking deeper into this I find that the versioning write lock is only acquired if the transaction being committed contains versioning operations. This is incorrect as all transactions in any case need to access the version store when checking for references.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.