[
https://issues.apache.org/jira/browse/FELIX-911?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12669653#action_12669653
]
Richard S. Hall commented on FELIX-911:
---------------------------------------
The main issue with the patch is that it only narrows the possibility of a
deadlock. I think I have an idea on how we could eliminate it in this
situation, so I will try to look into it this week.
The main gist is to have threads waiting for bundle locks get notified whenever
the bundle state changes so they can check to see if the bundle is in the state
they need and fail if it isn't. In your example, the registering thread would
need the bundle to be active, when it was notified that the stopping thread
caused the bundle's state to become stopping, it would return a failure for the
register operation.
> Potential deadlock between Bundle.stop() and BundleContext.registerService()
> ----------------------------------------------------------------------------
>
> Key: FELIX-911
> URL: https://issues.apache.org/jira/browse/FELIX-911
> Project: Felix
> Issue Type: Bug
> Components: Framework
> Affects Versions: felix-1.0.0, felix-1.0.1, felix-1.0.3, felix-1.0.4,
> felix-1.2.0, felix-1.2.1
, felix-1.2.2, felix-1.4.0, felix-1.4.1, felix-1.6.0
> Reporter: Felix Meschberger
> Assignee: Richard S. Hall
> Fix For: felix-1.6.0
>
> Attachments: FELIX-911.patch
>
>
> When a bundle tries to register a service (in thread T1) and the framework is
> at the same time stopping the bundle (in thread T2), it may be that thread T2
> holds the bundle's lock. T1 trying to register the service is thus blocked
> since it cannot get the bundle lock.
> This may cause a deadlock in Felix SCR, which also works with synchronized
> instances: T1 may be SCR Actor thread starting a component and registering a
> service and thus holding a lock on the component instance. T2 may be the
> StartLevel service trying to stop the bundle, which causes the SCR to
> immediately stop the component. Since the component is locked by T1, T2
> cannot acquire the lock. But since T2 already has the bundle lock, T1 cannot
> continue.
> The problem is, that in this concrete case it is the Component's activate()
> method which registers the service and not the SCR ComponentManager (the
> latter deadlock situation has been taken care of in FELIX-384).
> Looking at the code of Felix.registerService it looks like we could check for
> the bundle state before holding the bundle lock (and we could recheck after
> the lock just to be sure, if needed) and thus prevent the deadlock situation.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
