[11] RFR(M) 8201850: [AOT] vm crash when run test compiler/aot/fingerprint/SelfChangedCDS.java

[Vladimir Kozlov]

That what should current code do. I modified AOTLoader::universe_init() to only check shift values 
(setting shift was moved to new function) in each loaded AOT libraries. If a library has different 
shift recorded it is marked as invalided and will be unloaded:

http://cr.openjdk.java.net/~kvn/8201850/webrev.00/src/hotspot/share/aot/aotLoader.cpp.udiff.html

And such check is done unconditionally now. In original code the problem was that we reset shift if 
it was 0.

Thanks,
Vladimir

On 4/19/18 12:54 PM, Ioi Lam wrote:
> Also, would be it possible to generate code inside the .so library to record the shift size, and 
> trigger an assert if the runtime shift size is different than expected?
> 
> That way, even if the VM loads the .so file by mistake (e.g., in the future someone introduced a bug 
> in AOTLoader::universe_init), the problem will be caught early on and won't cause mysterious crashes?
> 
> Thanks
> 
> - Ioi
> 
> 
> On 4/19/18 12:46 PM, Ioi Lam wrote:
>> Hi Vladimir,
>>
>> The changes look good.
>>
>> I think (?) that the original failure happened only on machines with small amount of memory so 
>> default heap size is small. To ensure that this scenario is always tested, in SelfChangedCDS.java, 
>> maybe add a new test case that specifically uses a small heap during both dump time and run time, 
>> so CDS will use 0 shift?
>>
>> Thanks
>> - Ioi
>>
>> On 4/19/18 9:56 AM, Vladimir Kozlov wrote:
>>> http://cr.openjdk.java.net/~kvn/8201850/webrev.00/
>>> https://bugs.openjdk.java.net/browse/JDK-8201850
>>>
>>> The problem is AOT enforces not 0 compressed oop shift value (AOT uses logMinObjAlignment [1]) 
>>> regardless heap size to cover more cases. CDS does not do that - if heap is small (as in this 
>>> case) it will record 0 shift during dump.
>>> Later when CDS archive is loaded into VM with the same small heap it verifies that oop shift is 
>>> still 0. And it is 0 at that time because Metaspace::global_initialize() which loads archive is 
>>> called before AOTLoader::universe_init() which reset shift value.
>>> As result shared java objects (strings) in CDS archive are mapped into java heap with wrong 
>>> (after AOT change shift value) compressed oops encoding. GC dies because of that.
>>>
>>> The propose fix is add new AOTLoader::set_narrow_oop_shift() method which is called in 
>>> Universe::initialize_heap() before Metaspace::global_initialize(). AOTLoader::universe_init() is 
>>> changed to only verify that shift value is consistent. I also added code to unload AOT library 
>>> with non matching configs.
>>>
>>> Tested with small heap which reproduced the problem and full Mach5 testing.
>>>
>>> I want especially to thank Jiangli Zhou from runtime group who did initial investigation and 
>>> found root cause of this problem. She also suggested and discussed with me several solutions.
>>>
>>> Regards,
>>> Vladimir
>>>
>>> [1] 
>>> http://hg.openjdk.java.net/jdk/jdk/file/d5c43e9f08fb/src/jdk.internal.vm.compiler/share/classes/org.graalvm.compiler.hotspot/src/org/graalvm/compiler/hotspot/AOTGraalHotSpotVMConfig.java#l33 
>>>
>>
>