On Tue, 28 May 2024 12:36:41 GMT, Thomas Stuefe <stu...@openjdk.org> wrote:
> In `JvmtiUtil::single_threaded_resource_area()`, we create a resource area > that is supposed to work even if the current thread is not attached yet and > there is no associated Thread or the Thread has no valid ResourceArea. > > It contains a workaround: > > > // lazily create the single threaded resource area > // pick a size which is not a standard since the pools don't exist yet > _single_threaded_resource_area = new (mtInternal) > ResourceArea(Chunk::non_pool_size); > > > It specifies a non-standard chunk size to circumvent the chunk-pool-based > allocation in the RA constructor, ensuring that only malloc is used. This is > because in the old days the ChunkPools had been allocated from C-Heap and > there was a time window when no chunk pools were live yet. > > This is quirky and a bit ugly. It is also unnecessary since > [JDK-8272112](https://bugs.openjdk.org/browse/JDK-8272112) (since JDK 18). We > now create chunk pools as global objects, so they are live as soon as the > libjvm C++ initialization ran. We can remove this workaround and the comment. > > --- > > Tests: GHAs. > I also manually called this function, and allocated from the resulting > ResourceArea, at the very beginning of CreateJavaVM. I made sure that both > allocations and follow-up-chunk-allocation worked even this early in VM life. @jdksjolen could you take a look? You know the Arena coding behind it, and this PR is, in a very circumvent way, one of the prerequisites for NMT simplifications I plan. ------------- PR Comment: https://git.openjdk.org/jdk/pull/19425#issuecomment-2135835425
