[jira] [Updated] (DERBY-5258) btree post commit releases latch before committing/aborting purges, possibly allowing other operation on page

["Mike Matrigali (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/DERBY-5258?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Mike Matrigali updated DERBY-5258:
----------------------------------

    Attachment: DERBY-5258_diff.txt

Attaching proposed fix.  This has not passed tests yet.
I have not been able to reproduce the problem in my environment so the fix is just from code inspection.  One change from previous comments is that this fix is to 
purgeRowLevelCommittedDeletes().  The purgeCommittedDeletes() routine is ok, as it holds a table
level exclusive lock until end of transaction.

I will run full set of tests before committing to trunk.

I will try for another day to reproduce, but if I can't will probably go ahead and check in to trunk unless anyone
feels I should not.

> btree post commit releases latch before committing/aborting purges, possibly allowing other operation on page
> -------------------------------------------------------------------------------------------------------------
>
>                 Key: DERBY-5258
>                 URL: https://issues.apache.org/jira/browse/DERBY-5258
>             Project: Derby
>          Issue Type: Bug
>          Components: Store
>    Affects Versions: 10.1.3.1, 10.2.2.0, 10.3.3.0, 10.4.2.0, 10.5.3.0, 10.6.1.0, 10.7.1.1, 10.8.1.2
>            Reporter: Mike Matrigali
>            Assignee: Mike Matrigali
>         Attachments: DERBY-5258_diff.txt
>
>
> From code inspection found the following problem.  BTreePostCommit.purgeCommittedDeletes gives up the latch in it's finally block, before the internal transaction is
> committed.  The transaction is committed no sync upon return from this routine leaving a very small window when some other thread could get latch on the page and
> perform operations on the page.
> This can be a problem if for some reason the internal transaction is never committed.  Purges actually return space to the page, unlike deletes.  In order to backout the 
> purges one must add the rows back, taking up space on the page.  If another tranaction comes in before the internal transaction is committed and does inserts there may
> be no space for the backout of the purges.  This is why normal delete processing only sets flags on the rows and purge processing is handled differently.  
> I found this problem while debugging a database submitted as part of DERBY-5248.  I believe this issue can cause the problem there, but since we have no repro have
> decided to create a new issue to target this specific problem/solution.  Later can close the other issue if it can never be reproduce after the fix.   In DEBY-5248 there 
> are purges without a commit followed immediated by an insert in another transaction that is commited and the purge transaction is never committed.  On recovery the
> system tries to abort the internal transaction and eventually trashes the page when it does not actually have enough space to abort the purge.  See that issue for more
> detail.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira