RFR: 8255987: JDI tests fail with com.sun.jdi.ObjectCollectedException [v2]

[Chris Plummer]

On Mon, 7 Dec 2020 22:14:28 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> Ok, I understand now. `ReferenceType.instances()` only counts objects "reachable for the purposes of garbage collection". This change in behavior does concern me a little bit. I think the expectation is that the instances created by `ClassType.newInstance()` will not show up in this count unless `disableCollection()` is called, even when under a "suspend all". Clearly that's the expectation of this test, so the question is whether or not it is a reasonable expectation.
>> 
>> Note that `ClassType.newInstance()` says nothing about the state of the returned object w.r.t. GC. It makes no mention of the need to call `disableCollection()` before resuming the VM, so I guess this gives us some wiggle room. However, the argument against the object being strongly reachable is comes from user asking the question "who has the strong reference that makes it strongly reachable?". It's not obvious to the user why there is a strong reference, and why it seemingly goes a way once `VM.resumeAll()` is called.
>> 
>> I still think overall this is the right approach (baring a better approach being presented), but we may need to include some spec clarifications, and be prepared for some push back if this breaks anything.
>
> I don't follow your reasoning here Chris. All ObjectReferences can be GC'd at any time unless GC has been disallowed. So a reference create via newInstance is no different to any other reference. If it is currently reachable then instances() should return it. Are you treating "reachable for the purposes of garbage collection" as-if it said "strongly reachable"? It doesn't so I think you are reading too much into this. I think there is a lot of flexibility in this API in terms of what it may return regarding weak references.

I read "reachable for the purposes of garbage collection" as not including objects reachable only via weak reference. So if the only reference to an object is a weak reference, which is normally what you have after calling `ClassType.newInstance()`, then the object is not considered reachable. At the very least, his is how `ReferenceType.instances()` is implemented, and is based on JVMTI [FollowReferences](https://docs.oracle.com/en/java/javase/14/docs/specs/jvmti.html#FollowReferences)().

So given that, the expectation would be that an object returned `ClassType.newInstance()` would not be counted by `ReferenceType.instances()` unless something is done to add a strong reference to the object, such as calling `ObjectReference.disableCollection()`. Now with Per's changes a strong reference is also created with doing a VM.suspend(). The test doesn't expect this behavior, and it's understandable why.

-------------

PR: https://git.openjdk.java.net/jdk/pull/1595