RFR(M): 8203826: Chain class initialization exceptions into later NoClassDefFoundErrors

[Volker Simonis]

Hi,

can I please have a review for the following change which saves
ExceptionInInitializerError thrown during class initialization and
chains them as cause into potential NoClassDefFoundErrors for the same
class. We are using this features since years in our commercial SAP
JVM and it proved extremely useful for detecting and fixing errors
especially in big deployments.

This is a follow-up on a discussion previously started by Goetz [1].
His first proposal (which is close to our current, internal
implementation) inserted an additional field into java.lang.Class
objects to save potential ExceptionInInitializerErrors. This was
considered to much overhead in the initial discussion [1].

http://cr.openjdk.java.net/~simonis/webrevs/2018/8203826.v2/
https://bugs.openjdk.java.net/browse/JDK-8203826

So in this change, I've completely re-implemented the feature by using
a java.lang.Hashtable which is attached to the ClassLoaderData object.
The Hashtable is lazily created when the first
ExceptionInInitializerError is thrown and maps the Class which
triggered the ExceptionInInitializerError during the execution of its
static initializer to the corresponding ExceptionInInitializerError.

If the same class will be accessed once again, this will directly lead
to a plain NoClassDefFoundError (as per the JVMS, 5.5 Initialization)
because the static initializer won't be executed a second time. Until
now, this NoClassDefFoundError wasn't linked in any way to the root
cause of the problem (i.e. the first ExceptionInInitializerError
together with the chained exception that happened during the execution
of the static initializer). With this change, the NoClassDefFoundError
will now chain the initial ExceptionInInitializerError as cause,
making it much easier to detect the problem which lead to the
NoClassDefFoundError.

Following is an example from the new JTreg tests which comes which
this change to demonstrate the feature. Until know, a typical stack
trace from a NoClassDefFoundError looked as follows:

java.lang.NoClassDefFoundError: Could not initialize class
NoClassDefFound$ClassWithFailedInitializer
    at java.base/java.lang.Class.forName0(Native Method)
    at java.base/java.lang.Class.forName(Class.java:291)
    at NoClassDefFound.main(NoClassDefFound.java:38)

With this change, the same stack trace now looks as follows:

java.lang.NoClassDefFoundError: Could not initialize class
NoClassDefFound$ClassWithFailedInitializer
    at java.base/java.lang.Class.forName0(Native Method)
    at java.base/java.lang.Class.forName(Class.java:315)
    at NoClassDefFound.main(NoClassDefFound.java:38)
Caused by: java.lang.ExceptionInInitializerError
    at java.base/jdk.internal.reflect.NativeConstructorAccessorImpl.newInstance0(Native
Method)
    at java.base/jdk.internal.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)
    at java.base/jdk.internal.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)
    at java.base/java.lang.reflect.Constructor.newInstance(Constructor.java:490)
    at java.base/java.lang.Class.newInstance(Class.java:584)
    at NoClassDefFound$ClassWithFailedInitializer.<clinit>(NoClassDefFound.java:20)
    at java.base/java.lang.Class.forName0(Native Method)
    at java.base/java.lang.Class.forName(Class.java:315)
    at NoClassDefFound.main(NoClassDefFound.java:30)
Caused by: java.lang.ArrayIndexOutOfBoundsException: Index 2 out of
bounds for length 1
    at NoClassDefFound$A.<clinit>(NoClassDefFound.java:9)
    ... 9 more

As you can see, the reason for the NoClassDefFoundError when accessing
the class 'NoClassDefFound$ClassWithFailedInitializer' is actually not
even in the class or its static initializer itself, but in the class
'NoClassDefFound$A' which is a base class of
'NoClassDefFound$ClassWithFailedInitializer'. This is not easily
detectible from the old, plain NoClassDefFoundError.

As I wrote, the only overhead we have with the new implementation is
an additional OopHandle field per ClassLoaderData which I think is
acceptable. The Hashtable object itself is only created lazily, after
the first occurrence of an ExceptionInInitializerError in the
corresponding class loader. The whole Hashtable creation and
storing/quering of ExceptionInInitializerErrors in
ClassLoaderData::record_init_exception()/ClassLoaderData::query_init_exception()
is optional in the sense that any errors/exceptions occurring during
the execution of these functions are ignored and cleared.

Finally, we also take care to recursively convert all native
backtraces in the stored ExceptionInInitializerErrors (and their
suppressed and chained exceptions) into symbolic stack traces in order
to avoid holding references to classes and prevent their unloading.
This is implemented in the new private, static method
java.lang.Throwable::removeNativeBacktrace() which is called for each
ExceptionInInitializerError before it is stored in the Hashtable.

Thank you and best regards,
Volker

[1] http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2018-June/028310.html