Rajini Sivaram wrote:
Thank you, Richard. The fix will be of great help to me.
I imagine that the first check will need to be removed to support delayed
services with Declarative Services for Felix since presumably DS also
installs proxy services in order to implement delayed services.
Proxies are not needed in DS because it can just use the ServiceFactory
interface...
Even then, it seems that your scenario is only relevant if the registrar
also has access to the same class of the provider's service object,
right? In general, DS, for example, would not have access to any service
object interfaces so the first check would always fail and the second
test would be used. At least, that is my understanding.
I haven't committed the changes yet, but I should be the end of the day,
I imagine...I just want to think about it a little bit more before doing
so...
Thanks again.
-> richard
Thank you...
Regards,
Rajini
On 6/15/07, Richard S. Hall <[EMAIL PROTECTED]> wrote:
Rajini Sivaram wrote:
> Hello,
>
> I am working on SCA-OSGi integration to enable OSGi bundles to form
the
> implementation type for Apache Tuscany SCA components. I have run into
> some
> problems because of the way service references are filtered in Felix.
>
> This is the scenario. We startup Felix runtime in embedded mode inside
> Tuscany. To intercept an OSGi service, a proxy is registered into the
> OSGi
> service registry. The interface used by the proxy is of the same
> classloader
> as the one seen by the requestor bundle (it was loaded using
> requestorBundle.loadClass). The classloader of the system bundle
> however has
> access to this interface loaded by the application classloader. The
> requestor has a wire for the interface, but the provider (system
bundle)
> doesn't. The test performed for ServiceReference.isAssignableTo seems
> to be
> different for Equinox, Knopflerfish and Felix. In both Equinox and
> Knopflerfish, the proxy service is returned to the requestor bundle
when
> getServiceReference is called. Felix, however first tests if the class
> loaded by the provider is the same as the class loaded by the
requestor
(
> ServiceRegistrationImpl.isClassAccessible()). If the class cannot be
> loaded
> by the provider, then the class of the registered object is tested. Is
> the
> first test which prevents bundles from registering services of objects
> using
> different classloaders strictly necessary? Wouldn't it be sufficient
> to test
> the class of the object registered?
Yes, you could be right, because if the service object class is loaded
from the provider, then it should be discovered by the second check
anyway, right? Thus, it seems like it should only be necessary to do the
second check which will cover both cases...I think... :-)
Anyone want to venture a guess as to whether my logic is correct? If so,
I can commit a fix...
-> richard
