RFR (S): JDK-8152949: Jigsaw crash when Klass in _fixup_module_field_list is unloaded

[Stefan Karlsson]

On 2016-04-15 21:45, Alan Bateman wrote:
>
> On 15/04/2016 18:02, Lois Foltan wrote:
>>
>> Hi Stefan,
>>
>> In start up before module system initialization in complete I believe 
>> the VM is single threaded, so the increment/decrement reference 
>> counts do not need to be atomic.  Adding it is a defensive move in 
>> case the reference count is ever used passed start up in the future.  
>> It kind of does seem a bit excessive, sounds like you agree?
> There will be a number of threads running before the base module is 
> defined to the VM. As things stand the the java threads at this point 
> will be the Common-Cleaner, Finalizer, Reference Handler and Signal 
> Handler.

So, are you saying that we need the atomics?

The java_lang_Class::create_mirror function isn't multi-thread safe, and 
must already be guarded by a lock (SystemDictionary_lock AFAICT). The 
increment in Unsafe_DefineAnonymousClass0, will only be done once, for 
the single InstanceKlass instance in the CLD. And all reads of 
_keep_alive from the GC are done during safepoints.

How does ModuleEntryTable::patch_javabase_entries guard against 
concurrent inserts into the _fixup_module_field_list list?

thanks,
StefanK


>
> -Alan