[ 
https://issues.apache.org/jira/browse/ARIES-1399?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

John Ross reopened ARIES-1399:
------------------------------

I'm reopening this because the chosen solution is not sufficient to address all 
possible scenarios. Two of the three CT tests referenced in the description are 
failing again.

org.osgi.test.cases.subsystem.junit.SharingPolicySubsystemTests.test3A1b_BundleIsolationFeatures
org.osgi.test.cases.subsystem.junit.SharingPolicySubsystemTests.test3A1a_BundleIsolationComposite

They are failing for a similar but not the same reason as before. This time, 
the INSTALLED and RESOLVED events are both received but in the wrong order. The 
reason is that during the synchronous processing of the INSTALLED event, the 
system repository gets called which retrieves all capabilities for the bundle, 
including those for services, in order to cache them in the capability set. 
This bypasses the protection of lazy computation previously added in order to 
avoid having region digraph (which uses a bundle event hook) filter out the 
RESOLVED event. All synchronous bundle listeners called after the system 
repository will see the RESOLVED event before the INSTALLED event.

The second solution mentioned in the description is still an option. Another 
option would be to use an asynchronous bundle listener for the system 
repository.

> Trunk fails OSGi R6 CT
> ----------------------
>
>                 Key: ARIES-1399
>                 URL: https://issues.apache.org/jira/browse/ARIES-1399
>             Project: Aries
>          Issue Type: Bug
>            Reporter: John Ross
>            Assignee: John Ross
>         Attachments: TEST-org.osgi.test.cases.subsystem-6.0.0.html, 
> TEST-org.osgi.test.cases.subsystem-6.0.0.html
>
>
> The Subsystems implementation is currently failing three tests in the R6 CT 
> as a result of some of the recent performance tunings:
> org.osgi.test.cases.subsystem.junit.SharingPolicySubsystemTests.test3A1a_BundleIsolationApplication
> org.osgi.test.cases.subsystem.junit.SharingPolicySubsystemTests.test3A1b_BundleIsolationFeatures
> org.osgi.test.cases.subsystem.junit.SharingPolicySubsystemTests.test3A1a_BundleIsolationComposite
> It fails all three for the same reason. The tests explicitly install a bundle 
> using the region bundle context of the root subsystem. The bundle is 
> subsequently started. Using a synchronous bundle listener, the tests 
> reasonably expect to see INSTALLED, RESOLVED, STARTING, and ACTIVE bundle 
> events. However, the RESOLVED event is missing. It is missing because:
> (1) The Bundle Event Hook, registered with a service ranking of 
> Integer.MAX_VALUE, correctly identifies the event as pertaining to an 
> explicitly installed bundle and acts accordingly.
> (2) As normal, the bundle is wrapped by a Bundle Revision Resource in order 
> to compute the service capabilities and requirements if the Application 
> Modeller service is present. Previously, the capabilities and requirements 
> were not cached but (lazily) computed with each request. Today, they are 
> computed once and cached at instantiation.
> (3) Computing the service capabilities and requirements requires searching 
> for Blueprint XML files through Bundle.findEntries which results in the 
> framework attempting to resolve the bundle.
> (4) Region Digraph, through its own bundle event hook, receives the RESOLVED 
> event before the INSTALLED event. Because this is an unknown bundle, the 
> event is filtered and not delivered to any listeners, synchronous or 
> otherwise.
> There are at least a few possible solutions to the problem.
> (1) Have the Bundle Event Hook explicitly add the bundle to the region using 
> Region.addBundle(Bundle) as part of handling the INSTALLED event. This would 
> (+) solve the immediate issue but (-) promote out of order event delivery to 
> other hooks and listeners.
> (2) Prevent bundles under these circumstances from resolving altogether. This 
> would (+) solve the immediate issue and (+) prevent out of order event 
> delivery, but (-) eliminate the possibility of searching fragment bundles for 
> Blueprint XML files. 
> (3) Lazily compute service dependencies and requirements in Bundle Revision 
> Resource as before but still cache them. This would (+) solve the immediate 
> issue, (+) resemble how things already were before performance tuning, (+) 
> not contribute to out of order event delivery, and (+) leave open the 
> possibility of searching fragments for Blueprint XML files.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to