javaagent中redefine和retransform的区别

Question

When packing a jar file with custom java agent, you can add the following properties:

• Can-Redefine-Classes
• Can-Retransform-Classes

What is the difference between those two?

If redefinition happens before class is loaded and retransformation after, then when does exactly retransformation happen?

Answer 1

They seem almost redundant in the functionality they offer us. The main difference seems to be that when we redefine a class, we supply a byte[] with the new definition out of the blue, whereas when we retransform , we get a byte[] containing the current definition via the same API, and we return a modified byte[] .

Therefore, to redefine , we need to know more about the class. Consider the use-case of injecting profiling trace statements. With retransform you can do that more directly: just look at the bytecode given, modify it, and return it. But if we went the redefine route, we would need to fetch the original byte[] from somewhere (like getResourceAsStream() ).

Another apparent difference is in how we interact with other class transformers; who goes first. Transforms are applied to the original or a redefined class, so several transforms can be additive, for example.

Historically, if we look at the Since comments in the API documentation , or on page 238 of this book (Friesen 2007 Beginning Java SE 6 Platform ), we notice that redefinition capabilities were introduced in Java 5, and retransformation in Java 6. My guess is that retransformation was introduced as a more general capability, but redefinition had to be retained for backwards compatibility.

Quoting the key sentence about retransformation methods from the book linked above:

Agents use these methods to retransform previously loaded classes without needing to access their class files.

---

The second part of the question:

If redefinition happens before class is loaded and retransformation after, then when does exactly retransformation happen?

No, redefinition happens after the class is loaded, as well as retransformation. They happen when you call your Instrumentation instance's redefineClasses(..) and retransformClasses(..) methods, respectively.

Here is a question to any experts passing by: is there anything you can do by redefining classes, that you can't do by retransforming them? My guess is that the answer is "nothing".

Answer 2

Redefinition means that at an arbitrary point of time an agent will invoke Instrumentation. redefineClasses to change the actual definition of existing (and already loaded) classes. The agent will provide the bytecode for the new definition.

Retransformation refers to the process of class file transformation which is normally applied at class loading time. Agents can register ClassFileTransformers which are called one after another to apply transformations to the byte code before the class will be initialized. So Retransformation refers to the capability of the JVM to repeat this process for already loaded classes. In this case an agent may invoke Instrumentation.retransformClasses specifying which classes to retransform but no bytecode. Instead the JVM will call all registered retransforming capable ClassFileTransformers providing the actual bytecode (or the result of the previous transformer for a chained transformer).

Answer 3

Retransformation:

Transformation refers to the process of class file transformation which is normally applied at class loading time. Agents can register ClassFileTransformers which are called one after another to apply transformations to the byte code before the class will be initialized. So Retransformation refers to the capability of the JVM to repeat this process for already loaded classes.

Retransformation is the same capabilities as instrumenting the classes/methods but for the loaded classes.

Issues with Retransformation: JVM stores bytecode in PermGen (Java 7 or lower) or Metaspace (Java 8 or higher). Every time an agent retransforms the class/method, JVM saves a copy of the modified bytecode in PermGen or Metaspace. Excessive retransformation can lead memory leaks.

IBM Java has additional 10% CPU overhead for retransformation. That is why retransformation is disabled in IBM Java agent. ~80% of all JVM crashes are due to retransformation - This does not mean that JVM crashes most of the time but whenever we face JVM crash issue, ~80% of the time is because of retransformation.

Disable Retransformation: There is a java option system property (-Dappdynamics.agent.disable.retransformation=true) to disable retransformation