RFR(M): 8203824: Chain exception from initialization in later NoClassDefFoundErrors.

[David Holmes]

Hi Goetz,

On 5/06/2018 11:21 PM, Lindenmaier, Goetz wrote:
> Hi David,
> 
> I also would prefer to keep this field within hotspot.
> 
> The reason we use a Java function is that that
> also calls out to a routine that transforms the
> backtrace to stackTrace recursively for the chained
> exceptions, which was implemented in Java.
> We did this to make sure we don't keep any classes
> alive that are referenced by the backtrace.
> I didn't include that (yet), because that calls back into
> the VM ...
> Do you think I should address this issue right away,
> or do you think it's overly cautious to free the backtrace?

Hadn't thought too much about this aspect - seems to open a bit of a can 
of worms. The thread that originally hit the exception could be 
operating entirely within a child loader of the defining loader of the 
class in question. That child loader and all its classes could be 
unloaded by the time the next thread tries to load the class and 
retrieve the original exception. So having the backtrace keep those 
classes alive seems bad. Converting the backtrace to strackTraceElements 
addresses that - but at the expense of a complicated Java-level 
transformation.

I'm very reluctant to add such complexities here. My basic position is 
that whichever thread encountered the original 
ExceptionInInitializerError should be reporting it some way, not 
silently swallowing it - so diagnosis of the problem should be apparent 
from logs. With that in mind I would suggest clearing the backtrace so 
we only report the type of exception and the message - _but_ that 
doesn't work because the original exception may still be inflight in the 
original thread that encountered it! So maybe the original exception 
type and message should simply be added to the message for the NCDFE?

Thanks,
David

> Best regards,
>    Goetz.
> 
> 
>> -----Original Message-----
>> From: David Holmes [mailto:david.holmes at oracle.com]
>> Sent: Dienstag, 5. Juni 2018 04:36
>> To: Lindenmaier, Goetz <goetz.lindenmaier at sap.com>; hotspot-runtime-
>> dev at openjdk.java.net
>> Subject: Re: RFR(M): 8203824: Chain exception from initialization in later
>> NoClassDefFoundErrors.
>>
>> Hi Goetz,
>>
>> On 2/06/2018 2:02 AM, Lindenmaier, Goetz wrote:
>>> Hi,
>>>
>>> If static initialization of a class fails, an ExceptionInInitializerError
>>> is thrown. Later accesses to this class will fail, too. They report
>>> a NoClassDefFoundError. This is misleading, as the class has been
>>> found in first place but the initialization failed.
>>>
>>> To give more context, retain the previous ExceptionInInitializerError
>>> and chain it into the NoClassDefFoundError:
>>>
>>> I added a field to java/lang/Class to retain the exception.
>>> This is the design we followed in SAP JVM. I am not sure whether
>>> the additional memory consumption of Class is acceptable, or whether
>>> a hash map holding the exceptions internally would be a better
>>> design.
>>
>> I suggest moving the field to instanceKlass and keeping this entirely
>> within the VM. I would not want to see the change on the Java side or
>> the need for a Java upcall to set this - you might hit more exceptions
>> in the process! Any change on the Java side has to go through the
>> core-libs folk anyway.
>>
>> Thanks,
>> David
>> -----
>>
>>> I also left out code that transforms the Throwable::backtrace in the
>> retained
>>> ExceptionInInitializerError and it's recursively chained exceptions into
>>> Throwable::stackTrace and limits the recursively chained exceptions.
>>> If this change is appreciated, I'll add that in a later change.
>>>
>>> I found a test for NoClassDefFoundError messages and extended that.
>>> I moved the existing test into the canonical location under
>> /exceptionMsgs/
>>>
>>> Best regards,
>>>     Goetz.
>>>
>>>
>>>
>>>
>>>
>>>
>>> Failing class initialization throws a ExceptionInInitializerError.
>>> Later accesses to this class throw a NoClassDefFoundErrors
>>> hiding the original cause.
>>>
>>> This change adds code to chain the initial EIIE in the NCDFE.
>>> http://cr.openjdk.java.net/~goetz/wr18/8203826-exMsg-
>> NoClassDefFoundError/webrev
>>>
>>> The initial EIIE is stored in the Class that failed to initialize.
>>> I thought
>>>