RFR/RFC: 8220342: Remove scavenge_root_nmethods_do from VM_HeapWalkOperation::collect_simple_roots

[Stefan Karlsson]

On 2019-03-14 10:21, Erik Helin wrote:
>> On 12 Mar 2019, at 16:19, Stefan Karlsson <stefan.karlsson at oracle.com> wrote:
>>
>> Hi all,
> 
> Hey StefanK,
> 
>> Please review and/or comment on this change to remove CodeCache::scavenge_root_nmehods_do from VM_HeapWalkOperation::collect_simple_roots.
>>
>> http://cr.openjdk.java.net/~stefank/8220342/webrev.01/
> 
> I think this patch makes sense, removing the call CodeCache::scavenge_root_nmethods_do makes the semantics of the JVMTI operations you mentioned much more clear. Reviewed.

Thanks!

> 
>> On 12 Mar 2019, at 16:19, Stefan Karlsson <stefan.karlsson at oracle.com> wrote:
>> There has been some brief, internal discussions that maybe we want to visit all sets of roots in the vm, both strong and weak. A quick
>> implementation of that causes problem in testing when objects tagged by JVMTI, and JNI weak global handles, gets reported as roots.
>> Because of that, such change requires more investigation and work than simply extending the set of roots.
> 
> It would be nice if we could add a note somewhere about how HotSpot interprets “...and other objects used as roots for the purposes of garbage collection” in the JVMTI specification. As you have highlighted, and as we discussed, this wording in the JVMTI specification is not precise enough. Different GC algorithms (and different phases of a single GC algorithm) often have different root sets, rendering the wording “…objects used as roots for the purpose of garbage collection” unclear, since the “objects used as roots” can differ depending on GC algorithm and phase. An implementation of the JVMTI specification has to interpret this sentence and choose which objects to include in the root set for these JVMTI operations. I think it would be helpful for HotSpot JVMTI users if there is a place where we could document which objects HotSpot includes in this root set (we should preferably use the same set of root objects for all GC algorithms).

Should we create a hotspot/jvmti RFE for this?

StefanK

> 
> Thanks,
> Erik
>