[
https://issues.apache.org/jira/browse/ACE-105?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dennis Geurts updated ACE-105:
------------------------------
Description:
During a commit of changes to the repository, I have reached a deadlock.
Basically, what happens is the following:
The commit() action triggers a call to commit on all repository sets. This
triggers a serialization to XML of the repositories. During this action, each
repository is set to busy, which eventually sets all contained RepositoryObject
to busy ( i.e. RepositoryObjectImpl.setBusy() is called). This is where a lock
is created, first on its attributes, then on its associations.
Meanwhile, based on an asynchronous EventAdmin event, it is decided that a new
deployment version is required for a particular target. During this deployment
creation process, it is possible that we loop over all objects in the
repository. During this process we end up in the getAssociations() call of the
same RepositoryObjectImpl, that was just set to busy. Unfortunately, this call
to getAssociations() first locks the contained associations. Then, during the
equals() call, a lock is created on the contained attributes.
This can lead to a deadlock. I will attach a threaddump that reproduces the
issue.
I propose the following patch (see attachemnt) where in the
RepositoryObjectImpl.setBusy() call the lock order is reversed.
After applying this patch locally, I have not been able to reproduce this
deadlock anymore.
was:
During a commit of changes to the repository, I have reached a deadlock.
Basically, what happens is the following:
The commit() action triggers a call to commit on all repository sets. This
triggers a serialization to XML of the repositories. During this action, each
repository is set to busy, which eventually sets all contained RepositoryObject
to busy ( i.e. RepositoryObjectImpl.setBusy() is called). This is where a lock
is created, first on its attributes, then on its associations.
Meanwhile, based on a synchronous EventAdmin event, it is decided that a new
deployment version is required for a particular target. During this deployment
creation process, it is possible that we loop over all objects in the
repository. During this process we end up in the getAssociations() call of the
same RepositoryObjectImpl, that was just set to busy. Unfortunately, this call
to getAssociations() first locks the contained associations. Then, during the
equals() call, a lock is created on the contained attributes.
This can lead to a deadlock. I will attach a threaddump that reproduces the
issue.
I propose the following patch (see attachemnt) where in the
RepositoryObjectImpl.setBusy() call the lock order is reversed.
After applying this patch locally, I have not been able to reproduce this
deadlock anymore.
> Deadlock on (client) RepositoryAdmin.commit()
> ---------------------------------------------
>
> Key: ACE-105
> URL: https://issues.apache.org/jira/browse/ACE-105
> Project: Ace
> Issue Type: Bug
> Affects Versions: 0.8.0
> Reporter: Dennis Geurts
> Attachments: ace-roi.patch, dump.txt
>
>
> During a commit of changes to the repository, I have reached a deadlock.
> Basically, what happens is the following:
> The commit() action triggers a call to commit on all repository sets. This
> triggers a serialization to XML of the repositories. During this action, each
> repository is set to busy, which eventually sets all contained
> RepositoryObject to busy ( i.e. RepositoryObjectImpl.setBusy() is called).
> This is where a lock is created, first on its attributes, then on its
> associations.
> Meanwhile, based on an asynchronous EventAdmin event, it is decided that a
> new deployment version is required for a particular target. During this
> deployment creation process, it is possible that we loop over all objects in
> the repository. During this process we end up in the getAssociations() call
> of the same RepositoryObjectImpl, that was just set to busy. Unfortunately,
> this call to getAssociations() first locks the contained associations. Then,
> during the equals() call, a lock is created on the contained attributes.
> This can lead to a deadlock. I will attach a threaddump that reproduces the
> issue.
> I propose the following patch (see attachemnt) where in the
> RepositoryObjectImpl.setBusy() call the lock order is reversed.
> After applying this patch locally, I have not been able to reproduce this
> deadlock anymore.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
