8212205 (XS) VM asserts after CDS archive has been unmapped

[Ioi Lam]

Hi Jiangli,

Thanks for your comments. How about this:

http://cr.openjdk.java.net/~iklam/jdk12/8212205-crash-after-cds-unmap.v02/

- Ioi

On 10/25/18 3:11 PM, Jiangli Zhou wrote:
> On 10/25/18 1:49 PM, Ioi Lam wrote:
>
>> MetaspaceObj::_shared_metaspace_top is a private field. 
>> MetaspaceShared can modify the field because it's a friend class of 
>> MetaspaceObj. I don't want to add more friend classes. Also, I think 
>> FileMapInfo class is more like an internal class for MetaspaceShared 
>> and other classes should avoid using it.
> How about adding something like 
> MetaspaceShared::clear_shared_metaspace (or 
> MetaspaceShared::reset_shared_metaspace)? It could be called from 
> MetaspaceShared::stop_sharing_and_unmap (or from metaspace.cpp 
> directly). The reason I bring it up is because having two public 
> stop_sharing_and_unmap() APIs is a bit confusing and 
> FileMapInfo::stop_sharing_and_unmap() is no longer intended to be used 
> by others for general cases. Or how about making 
> FileMapInfo::stop_sharing_and_unmap() private (add MetaspaceShared as 
> a friend to FileMapInfo)?
>
> If you want to keep both public APIs, probably consider at least 
> adding a comment warning about the use of 
> FileMapInfo::stop_sharing_and_unmap()?
>
> Thanks,
> Jiangli
>>
>> Thanks
>>
>> - Ioi
>>
>> On 10/25/18 12:20 PM, Jiangli Zhou wrote:
>>> Hi Ioi,
>>>
>>> The fix looks good, but it's probably better to fix it in 
>>> FileMapInfo::stop_sharing_and_unmap() directly?
>>>
>>> Thanks,
>>>
>>> Jiangli
>>>
>>>
>>> On 10/25/18 11:22 AM, Ioi Lam wrote:
>>>> https://bugs.openjdk.java.net/browse/JDK-8212205
>>>> http://cr.openjdk.java.net/~iklam/jdk12/8212205-crash-after-cds-unmap.v01/ 
>>>>
>>>>
>>>> This is a silly bug but it happens only rarely, when address space 
>>>> randomization
>>>> causes insufficient memory to be reservable at addresses required 
>>>> by CDS. In this
>>>> case, the CDS archive is unmapped, but we forget to clean up
>>>> _shared_metaspace_{base,top}. As a result, the JVM will misidentify 
>>>> metaspace
>>>> objects allocated in certain address ranges to be "shared", in the 
>>>> following function:
>>>>
>>>> class MetaspaceObj {
>>>>   bool is_shared() const {
>>>>     return (((void*)this) < _shared_metaspace_top && ((void*)this) 
>>>> >= _shared_metaspace_base);
>>>>   }
>>>> }
>>>>
>>>> The fix is to reset those pointers after CDS has been unmapped.
>>>>
>>>> Thanks
>>>> - Ioi
>>>>
>>>
>>
>