Buck's Blog

  • January 21, 2014

JRockit R28 issue with exact profiling and generics

David Buck
Principal Member of Technical Staff

Some users of JRockit R28 may have noticed problems using Mission Control's exact profiling on methods that use Java's generics facility. Invocation counters for these methods would simply not respond; calling each method would fail to increment the corresponding call count.

For exact method profiling in R28, we replaced our homegrown bytecode instrumentation solution with Apache's Byte Code Engineering Library (BCEL). A version of BCEL was already included in the JDK as an internal component of JAXP, and using BCEL helped provide a cleaner code generation pipeline.

The problem was that while the internal version of BCEL contained in the JDK is very stable and works fine for the very narrow use cases JAXP requires, there were problems using it to instrument arbitrary code as needed by Mission Control Console's profiling tool.

One of those issues was that support for Java generics was never fully implemented in BCEL. In particular, instrumenting a method with generic code could produce bytecode with inconsistencies between the LocalVariableTable (LVT) attribute and the LocalVariableTypeTable (LVTT) attribute (Please see the class file format for details). Thankfully, this issue was found and fixed (in development branches) by the BCEL project:

Bug 39695 - java.lang.ClassFormatError: LVTT entry for 'local' in class file org/shiftone/jrat/test/dummy/CrashTestDummy does not match any LVT entry

Unfortunately, the JDK version of BCEL predated this fix. So when JRockit tried to instrument such methods using BCEL, the new method's bytecode would be invalid and fail subsequent bytecode validation, leaving the original, uninstrumented, version of the method in place.

While I briefly considered adding code on the JRockit side to work around the BCEL issue, it seemed that fixing the version of BCEL in the JDK itself was The Right Thing to do here. Unfortunately for me, the fix for bug 39695 was based on of a version of BCEL that was much more recent than the one contained in the JDK, so I needed to port over a lot of other code to get things working.

JDK-8003147: port fix for BCEL bug 39695 to our copy bundled as part of jaxp

My port of the BCEL fix and other needed code went into 5u45, 6u60 and 7u40. Note that for Java 5, our only option was to put the fix into a CPU release, as we no longer provide non-CPU releases for Java 5. This means that the exact version of JRockit this fix made it into depends on the major Java version: For Java 5: R28.2.7. For Java 6: R28.2.9. As the LVT/LVTT would normally only be included with debug builds, recompiling affected classes without the -g flag should also be a viable workaround for users of earlier releases.

Hopefully, not too many users were impacted by this issue. As explained very well in JR:TDG, sampling-based profiling (like that provided by Flight Recorder), is almost always a better choice than exact profiling. But this story is interesting for another reason: it is a great example of how depending on internal (not officially documented) classes within the JDK is almost always a really bad idea (*1). Even we have been known to get bit.

(*1) Upcoming modularization of the Java Class Library is expected to do a better job preventing outside use of internal JDK classes. Not that it would have helped here.

Be the first to comment

Comments ( 0 )
Please enter your name.Please provide a valid email address.Please enter a comment.CAPTCHA challenge response provided was incorrect. Please try again.