RFR(S) 8212200 assert when shared java.lang.Object is redefined by JVMTI agent
ioi.lam at oracle.com
Mon Oct 22 01:11:57 UTC 2018
CDS has various built-in assumptions that classes loaded by
SystemDictionary::resolve_well_known_classes must not be replaced
by JVMTI ClassFileLoadHook during run time, including
- field offsets computed in JavaClasses::compute_offsets
- the layout of the strings objects in the shared strings table
The "well-known" classes can be replaced by ClassFileLoadHook only
when JvmtiExport::early_class_hook_env() is true. Therefore, the
fix is to disable CDS under this condition.
I have added a few test cases to try to replace shared classes,
including well-known classes and other classes. See
comments in ReplaceCriticalClasses.java for details.
As a clean up, I also renamed all use of "preloaded" in
the source code to "well-known". They refer to the same thing
in HotSpot, so there's no need to use 2 terms. Also, The word
"preloaded" is ambiguous -- it's unclear when "preloading" happens,
and could be confused with what CDS does during archive dump time.
> In early e-mails Jiangli wrote:
> We should consider including more classes from the default classlist
> in the test. Archived classes loaded during both 'early' phase and after
> should be tested.
> For future optimizations, we might want to prevent loading additional
> shared classes if any of the archived system classes is changed.
What's the benefit of doing this? Today we already stop loading a shared
class if its super class was not loaded from the archive.
More information about the hotspot-runtime-dev