Yes, when the proxy extends a class, everything is taken care of by the bridge method in the superclass. But, when dealing with local EJB interfaces, the proxy's superclass is Object, so there is no bridge method at all.

For EJBs, the stacktrace would look something like this:
    StringFooImpl.doSth(String)
    StringFooImpl.doSth(Object)
    methodHandler.invoke("doSth(Object)")
    StringFoo$Proxy.doSth(Object)            <-- proxy extends Object, implements StringFoo interface (no bridge method)

For regular managed beans, the stacktrace would look like this:
    StringFooImpl.doSth(String)
    methodHandler.invoke("doSth(String)")
    StringFooImpl$Proxy.doSth(String)
    StringFooImpl$Proxy.doSth(Object)        <-- proxy extends StringFooImpl (and has bridge method)

Note the different methods methodHandler is handling.

Marko


On 7.9.2012 17:27, Stuart Douglas wrote:

IMHO the correct way to deal with this is to simply make bridge methods delegate to super(), which will then result in the actual intercepted method being called.

To be honest I thought we already did this, as we have had multiple related bugs in the past. Do you have a test case that I can look at?

Stuart

Marko Lukša

7 September 2012 7:19 PM

Hey all.

I've been working on https://issues.jboss.org/browse/WELD-1162 and need your opinion.

Say we have:

public interface Foo<T> {
   void doSomething(T t);
}
public interface StringFoo extends Foo<String> {}
public class StringFooImpl implements StringFoo {}

and

@Inject StringFoo stringFoo;

The proxy created by Weld is a subclass of StringFooImpl and therefore has two declared methods:

void doSomething(Object o) { doSomething((String) o); }
void doSomething(String) {...}

However, when StringFooImpl is a session bean, with StringFoo as its local interface, the proxy is a subclass of Object and therefore the proxy only has the following declared method:

void doSomething(Object o);

In both cases, when a client invokes stringFoo.doSomething("foo"), the method doSomething(Object) is invoked. But there's a difference in what happens next:

• In the non-ejb version, the bridge method then immediately invokes doSomething(String) and only then is the proxy's method handler invoked. The handler is therefore dealing with the method doSomething(String)
• in the EJB version, doSomething(Object) is not a bridge method, and so the method handler is invoked directly and it (the handler) is operating on doSomething(Object).

In the second case, this ultimately means that Weld will check whether doSomething(Object) is intercepted. It isn't, since Beans.getInterceptableMethods() is ignoring bridge methods. The interceptor will not be invoked. On the other hand, in the first case, the interceptor _will_ be invoked, since Weld will be checking whether doSomething(String) is intercepted.

My initial solution was to make Beans.getInterceptableMethods() also return bridge methods, but now I'm thinking the actual problem is in the proxy itself. IMO, when creating a proxy based on an interface, we should also generate bridge methods on the proxy (this should be either done by Weld or by Javassist directly). These bridge methods should be perfectly normal bridge methods and should not invoke the method handler directly. They should simply invoke the non-bridge method and the non-bridge method should then invoke the method handler.

The java compiler can't add bridge methods directly to interfaces which require them, so it adds them to all the classes implementing the interface (StringFooImpl in our case). Since we are creating StringFoo$Proxy, which is also a class implementing an interface which requires bridge methods, we should add the bridge methods
to it - exactly as the java compiler would.

This would solve the interceptor problem and possibly other similar problems as well.

What do you think?

Marko

_______________________________________________
weld-dev mailing list
weld-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/weld-dev