That does not work in the general case but only if the initiating class is
also on the class path. A class can only be identified uniquely as a pair
of name and class loader.
I experimented with this and ended up iterating over
Instrumentation.getAllLoadedClasses which resulted in rather poor
performance and ambiguity if there exist classes with the same name on
different class loaders.
Again, I argue that this problem is therefore worth addressing.
Best regards, Rafael
2016-09-30 15:51 GMT+02:00 Andrew Dinn <ad...@redhat.com>:
> On 30/09/16 14:27, Rafael Winterhalter wrote:
> > A Java agent ends up on the class path. The Byte Buddy agent (or any
> > similar library) basically adds a single class:
> >
> > package net.bytebuddy.agent;
> > public class Installer {
> > public static volatile Instrumentation instrumentation;
> > public static void agentMain(String argument, Instrumentation
> > instrumentation) {
> > Agent.instrumentation = instrumentation
> > }
> > }
> >
> > Since the class is added using self-attachment, the agentmain method is
> > called by the VM and the field value is set. In order to keep the field
> > accessible, it needs to be public such that any class loader can call:
> >
> > Instrumentation instrumentation = (Instrumentation)
> > Class.forName("net.bytebuddy.agent.Installer", false,
> > ClassLoader.getSystemClassLoader()).getDeclaredField("
> instrumentation").get(null);
> >
> > Any library on the class path can now also call the above code without
> > requiring any priviledges as the Instrumentation instance is exposed
> > without constraints. Adding a proper method for reading an instance of
> > Instrumentation would prevent this.
>
> Well, that's easily fixed. Make the agent push the Instrumentation
> instance to the class which loaded the agent.
>
> For example, provide the name of the class and name of a public setter
> method as arguments to agentMain (and, if you want, a shared key to
> validate the set). Then get the agent to locate the class, lookup the
> setter method and hand over the instance.
>
> regards,
>
>
> Andrew Dinn
> -----------
>
>
