[
https://issues.apache.org/jira/browse/HADOOP-10404?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13932142#comment-13932142
]
Colin Patrick McCabe commented on HADOOP-10404:
-----------------------------------------------
Hi guys. Thanks for looking at this. Although we could fix it like you did,
by taking the lock in {{NotificationHandler#handle}} and
{{DomainSocketWatcher#kick}}, I would prefer to ensure that the lock is already
taken in the watcher thread before these functions are invoked. The lock is
supposed to be already held at that point-- the only exception is that I forgot
to hold the lock in the cleanup logic of watcherThread's run method. It seems
more consistent to just fix the cleanup logic.
I posted a v3 where I just take the lock there, and add
assert(lock.isHeldByCurrentThread()) in {{NotificationHandler#handle}} and
{{DomainSocketWatcher#kick}}.
> Some accesses to DomainSocketWatcher#closed are not protected by lock
> ---------------------------------------------------------------------
>
> Key: HADOOP-10404
> URL: https://issues.apache.org/jira/browse/HADOOP-10404
> Project: Hadoop Common
> Issue Type: Bug
> Reporter: Ted Yu
> Assignee: Tsuyoshi OZAWA
> Priority: Minor
> Attachments: HADOOP-10404.003.patch, HADOOP-10404.1.patch,
> HADOOP-10404.2.patch
>
>
> {code}
> * Lock which protects toAdd, toRemove, and closed.
> */
> private final ReentrantLock lock = new ReentrantLock();
> {code}
> There're two places, NotificationHandler.handle() and kick(), where access to
> closed is without holding lock.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
