NullPointerException in JDWP during HotSwap

Question

I'm developing an application that needs to redefine classes at runtime, for which I'm using the HotSwap feature of the Sun JVM. I've written a HotSwap class that conects the application to itself via the SharedMemoryTransportService, obtains the VirtualMachine instance for the JVM and contains this method for swapping a single class:
     public void swap(String className, byte[] classBytes) throws IOException {
               Map<ReferenceType, byte[]> classDefinitions = new HashMap<ReferenceType, byte[]>();
               classDefinitions.put(getReferenceType(className, vm), classBytes);
               vm.redefineClasses(classDefinitions);
I then wrote a load of tests that take various classes and insert lines of code at the top of their methods. Most of these seemed to pass fine, with the inserted code executing successfully, but some caused the following NullPointerException to occur:
java.lang.NullPointerException
     at com.sun.tools.jdi.JDWP$VirtualMachine$RedefineClasses$ClassDef.write(JDWP.java:1473)
     at com.sun.tools.jdi.JDWP$VirtualMachine$RedefineClasses$ClassDef.access$300(JDWP.java:1452)
     at com.sun.tools.jdi.JDWP$VirtualMachine$RedefineClasses.enqueueCommand(JDWP.java:1508)
     at com.sun.tools.jdi.JDWP$VirtualMachine$RedefineClasses.process(JDWP.java:1490)
     at com.sun.tools.jdi.VirtualMachineImpl.redefineClasses(VirtualMachineImpl.java:289)
     at com.sun.tools.jdi.HotSwap.swap(HotSwap.java:71)
     at uk.ac.ic.doc.cuXca.persistence.core.persistencehandlers.ObjectClassEnhancer.enhance(ObjectClassEnhancer.java:92)
... my classes
For ages I thought that I was generating incorrect bytecode, but then I noticed that the majority of these exceptions disappeared if I made sure an instance of the swapped class existed before the swap (by adding a call to Class.newInstance()). However, some of my tests still throw the above exception and I'm confused as to why. To make matters worse, it seems to be caused by some kind of race condition, because different tests fail each time. My test code is not multithreaded, and I'm certain that the code I'm generating is valid. The whole newInstance thing makes me think that sometimes the classes have not been resolved prior to the swap. Has anyone who has used hotswapping before had any experience of this exception and been able to fix it?

Answer 1

The following tests produce some interesting results (each block is a separate run and enhance(..) is my method for altering the given class and swapping it):
enhance(A.class); // fails with NullPointerException
A a = new A();
enhance(A.class); // succeeds
A a = new A();
enhance(A.class); // succeeds
enhance(B.class); // fails
A a = new A();
enhance(A.class); // succeeds
B b = new B();
enhance(B.class); // fails with NullPointerException
A a = new A();
B b = new B();
enhance(A.class); // succeeds
enhance(B.class); // succeeds
From these, it seems to be that all classes to be swapped must be loaded before the first swap occurs, however then I found the following:
A a = new A();
enhance(A.class); // succeeds
B b = new B();
enhance(B.class); // fails with NullPointerException
C c = new C();
enhance(C.class); // fails with NullPointerException
A a = new A();
enhance(A.class); // succeeds
B b = new B();
C c = new C();
enhance(B.class); // fails with NullPointerException
enhance(C.class); // succeeds
Here swapping C succeeds if it was loaded before the swap of B failed! I'm guessing the state of whether classes have been loaded or not is being cached and only being refreshed on the first swap (when the connection to the JVM is first made) and when the exception occurs. Unfortunately I can't rely on loading all my classes before the first swap since I'm swapping user classes that I have no control over.
Any help greatly appreciated!