On Mon, 27 Sep 2021 12:02:53 GMT, Lin Zang <[email protected]> wrote:
>> The root cause for crash in ZGC is that the JNIHandles are processed before
>> object iteration. And ZGC would update the JNIHandles at object iteration
>> with read barrier. So the crash is cause by accessing the invalid address
>> which can be dummy info after zgc, and hence crash.
>>
>> The lock rank issue can be fixed because the related mutexes are acquired in
>> safepoint. so the safepoint_check_required could be safepoint_check_always.
>>
>> The Epsilon issue is caused by wrong _num_dumper_thread calculated when the
>> gang==NULL.
>
> Lin Zang has updated the pull request incrementally with one additional
> commit since the last revision:
>
> remove redundant empty line
src/hotspot/share/services/heapDumper.cpp line 1601:
> 1599: void JNILocalsDumper::do_oop(oop* obj_p) {
> 1600: // ignore null handles
> 1601: oop o = NativeAccess<AS_NO_KEEPALIVE>::oop_load(obj_p);
The JNI Local roots do not need a load barrier, only JNI Global roots need
that. The JNI Local roots are processed on safepoint entry as part of the
"thread head" (via `ZStackWatermark::ZStackWatermark::start_processing_impl()`
-> `Thread::oops_do_no_frames()`), so once you are in
`VM_HeapDumper::do_thread()` the JNI Local roots have already passed a load
barrier.
-------------
PR: https://git.openjdk.java.net/jdk/pull/5681
