RFR: 8135322: ConstantPool::release_C_heap_structures not run in some circumstances

[Andreas Eriksson]

Hi,

Here is a new webrev where I do the filtering of the _deallocate_list.
http://cr.openjdk.java.net/~aeriksso/8135322/webrev.01/

Another possible version is to call free_deallocate_list in 
ClassLoaderDataGraph::do_unloading.
http://cr.openjdk.java.net/~aeriksso/8135322/webrev.do_unloading/

Both of these versions pass the test which crashed.
What do you think? Any preference for one of the versions?

Thanks,
Andreas

On 2016-04-01 18:22, Andreas Eriksson wrote:
> Hi,
>
> When running some additional testing I had a test crash.
> Looks like sometimes data we need to run a full deallocate_contents 
> call when we are in ~ClassLoaderData is already freed (seems to depend 
> on what GC we are running, and what we have in the _deallocate_list).
>
> Doing the free_deallocate_list call in 
> ClassLoaderDataGraph::do_unloading instead will most likely work, 
> since it's run at an earlier point in the GC.
> Or I could do what I mentioned in the initial email, only run 
> release_C_heap_structures() on constant pools in the list that do not 
> have a corresponding InstanceKlass.
>
> I'm leaning towards, doing the free_deallocate_list call in 
> ClassLoaderDataGraph::do_unloading, when we detect that the 
> ClassLoaderData is dead.
> I'll send out a new review once I've had some time to test that fix.
>
> Thanks,
> Andreas
>
> On 2016-04-01 02:15, serguei.spitsyn at oracle.com wrote:
>> Andreas,
>>
>> I agree with Coleen, it is a good find and a good and safe fix.
>>
>> Thanks!
>> Serguei
>>
>>
>> On 3/30/16 16:06, Coleen Phillimore wrote:
>>>
>>> Andreas,
>>>
>>> Good find!  This seems like a safe fix.  It's unfortunate to have to 
>>> walk the deallocate list, but I think it's short enough and less 
>>> work in comparison to having to call free_C_heap_structures on all 
>>> the other classes in the dead class loader data.   And 
>>> deallocate_contents removes the scratch_class from the CLD _klasses 
>>> list so won't walk it twice.
>>>
>>> This is really good.
>>>
>>> Coleen
>>>
>>> On 3/30/16 8:12 AM, Andreas Eriksson wrote:
>>>> Hi,
>>>>
>>>> Please review this fix for 8135322: 
>>>> ConstantPool::release_C_heap_structures not run in some circumstances.
>>>>
>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8135322
>>>> Webrev: http://cr.openjdk.java.net/~aeriksso/8135322/webrev.00
>>>>
>>>> The way this occurs is:
>>>>
>>>> 1. A ConstantPool, /merge_cp/, without a corresponding 
>>>> InstanceKlass is
>>>>    created during class redefinition.
>>>> 2. /Merge_cp/ is added to ClassLoaderData::_deallocate_list, but is 
>>>> not
>>>>    cleaned up by ClassLoaderDataGraph::do_unloading because the 
>>>> holding
>>>>    ClassLoaderData is already dead.
>>>> 3. ClassLoaderData::~ClassLoaderData is called, which usually runs
>>>>    InstanceKlass::release_C_heap_structures, which in turn will call
>>>>    ConstantPool::release_C_heap_structures. But since there is no
>>>>    corresponding InstanceKlass for /merge_cp/ it will never release 
>>>> its
>>>>    C heap structures.
>>>> 4. For JDK 8 this means we leak a Monitor (ConstantPool::_lock). For
>>>>    JDK 9 the Monitor has been removed, but we will miss doing a 
>>>> call to
>>>>    ConstantPool::unreference_symbols, which is probably not good 
>>>> either.
>>>>
>>>>
>>>> This change adds a call to free everything in the _deallocate_list 
>>>> first thing in ClassLoaderData::~ClassLoaderData.
>>>>
>>>> Note:
>>>> Running deallocate_contents() on everything in the list might be 
>>>> doing unnecessary work:
>>>> What we really would like to do is run release_C_heap_structures() 
>>>> on constant pools in the list that do not have a corresponding 
>>>> InstanceKlass. Not sure it's worth the effort to do the filtering, 
>>>> since I believe the _deallocate_list is usually not that long.
>>>>
>>>> Note 2:
>>>> After fixing in JDK 9 I'll backport this to JDK 8, where it has a 
>>>> more immediately visible effect.
>>>>
>>>> Regards,
>>>> Andreas
>>>>
>>>
>>
>