[
https://issues.apache.org/jira/browse/SOLR-7836?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14697858#comment-14697858
]
Yonik Seeley commented on SOLR-7836:
------------------------------------
Here's the scenario w/ the old code.
thread1 does an add:
1) in DUH2, calls coreState.getIndexWriter()
- this increments the ref count on the writer and sets
coreState.writerFree=false
2) calls UpdateLog.add
thread2 calls core reload():
3) calls coreState.newIndexWriter()
- coreState.pauseWriter is set to true, so no new references will be
handed out
- goes into a loop waiting for writerFree=true (for all other references
that were handed out to be returned)
thread1:
4) UpdateLog.add continues and indirectly causes coreState.getIndexWriter()
to be called
- sees coreState.pauseWriter set to true and thus does into wait loop
So: one can't call getIndexWriter() and then do anything else that will
eventually call getIndexWriter() or newIndexWriter()
If we keep that restriction, then moving the ulog.add outside of the
getIndexWriter/release block was correct.
Don't know about the other changes... the extra sync added in DUH2 does still
seem unnecessary. And I haven't looked at what changes were made to
SolrCoreState yet.
> Possible deadlock when closing refcounted index writers.
> --------------------------------------------------------
>
> Key: SOLR-7836
> URL: https://issues.apache.org/jira/browse/SOLR-7836
> Project: Solr
> Issue Type: Bug
> Reporter: Erick Erickson
> Assignee: Erick Erickson
> Fix For: Trunk, 5.4
>
> Attachments: SOLR-7836-synch.patch, SOLR-7836.patch, SOLR-7836.patch,
> SOLR-7836.patch, deadlock_3.res.zip, deadlock_5_pass_iw.res.zip, deadlock_test
>
>
> Preliminary patch for what looks like a possible race condition between
> writerFree and pauseWriter in DefaultSorlCoreState.
> Looking for comments and/or why I'm completely missing the boat.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
