[
https://issues.apache.org/jira/browse/CAMEL-27322?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Work on CAMEL-27322 started by shashank.
----------------------------------------
> camel-file - a FileLockClusterView follower in the same JVM releases the
> leader's file lock, so a member in another JVM can become a second leader
> --------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: CAMEL-27322
> URL: https://issues.apache.org/jira/browse/CAMEL-27322
> Project: Camel
> Issue Type: Bug
> Components: camel-file
> Reporter: shashank
> Assignee: shashank
> Priority: Major
>
> h3. What happens
> {{FileLockClusterView.acquireLock}} opens a new {{RandomAccessFile}} on the
> lock file and calls {{tryLock}}. When another view of the same JVM holds the
> lock, the JDK throws {{OverlappingFileLockException}} and the new channel is
> closed. The file locks are locks of the process (POSIX fcntl on Linux and
> macOS): closing any channel of the file releases all the locks of the process
> on it ({{FileLock}} javadoc). The leader view stays LEADER and its
> {{FileLock}} still says it is valid, but the file is no longer locked at OS
> level, so a member in another JVM can take the lock and become LEADER too.
> The follower only tries the lock when the leader's heartbeat is stale (a GC
> pause or a blocking NFS call of the leader) or the data file is empty (cold
> start), which is when the lock is needed. Several CamelContexts with a
> {{FileLockClusterService}} on the same root in one JVM is a supported setup
> (the camel-master failover tests use it).
> h3. How to reproduce
> Two CamelContexts in one JVM with a {{FileLockClusterService}} on one root,
> plus a member in another JVM. Stall the leader's checks so its heartbeat gets
> stale: the follower in the same JVM tries the lock and closes its channel, an
> external fcntl probe then acquires the lock while view 1 is LEADER, and the
> member in the other JVM becomes LEADER as well.
> Found with a TLA+ model of the file lock cluster view.
> h3. Fix
> A JVM-wide registry of the lock files that a view holds or is trying: a view
> does not open a lock file that another view of its JVM has registered; the
> entry is removed after the lock is released and the files are closed.
> _Filed with Claude Code on behalf of allthingssecurity._
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)