[concurrency-interest] JEP-159 Enhanced Class Redefinition vs Unsafe.___FieldOffset()

Nathan Reynolds nathan.reynolds at oracle.com
Tue Jun 4 16:19:22 EDT 2013


JEP 159 Enhanced Class Redefinition 
(seehttp://openjdk.java.net/jeps/159) will allow for adding and removing 
fields in a class.  Also, a field could be changed in size as well (e.g. 
long → int).  When the change happens, the heap is scanned for all of 
the instances of the class(es) and all of the instances are modified.  
The cost is about that of a full GC.

I think the original driver for JEP 159 was to enhance what debuggers 
can change on the fly as the program is running.  I envision using JEP 
159 to remove dead fields, deal with fields used in a small percentage 
of instances (rarely used) and reduce the size of oversized fields.  
Dead and rarely used fields account for 6.9% of an average heap.  
Oversized fields account for 1.5% of an average heap.  By _occasionally_ 
using JEP 159 at runtime, we could change the class definitions to save 
most of this space and hence improve GC times and perhaps response times.

If I understand correctly, Unsafe.objectFieldOffset() and 
Unsafe.staticFieldOffset() return the number bytes from the beginning of 
the instance or class where the field is located.  This offset is 
typically obtained when the class is loaded or the instance is 
constructed.  For example, 
java.util.concurrent.atomic.AtomicLongFieldUpdater$CASUpdater will get 
the offset in the constructor for the specified field and class.  It 
keeps this offset for the lifetime of the CASUpdater instance.  The 
field and class can be any loaded class.

So, what will happen when JEP 159...

 1. removes the field used by CASUpdater?  It seems like CASUpdater will
    corrupt the instance or the object header of the next object.
 2. adds or removes a different field and this shifts the field used by
    CASUpdater?
 3. changes the long field to an int?
 4. adds or removes @Contended?  This will rearrange the fields in all
    of the instances.

CASUpdater is just one class.  What about all of the other classes which 
use or _will_ use Unsafe.___FieldOffset?

Maybe JEP 159 has already solved the problem.  If so, I would like to 
hear how.  If not, how can this problem be solved?

If JEP 159 changes the returned value of ___FieldOffset() from an offset 
to a cookie, what code will break which assumes it is an offset?

The value returned from Unsafe.___FieldOffset seems to be used as a 
cookie.  This cookie can be an index into an array of field offsets.  
JEP 159 will then have to fix up the array of field offsets for the 
affected classes.  If a field is removed or altered in size, then the 
existing entry in the array should be changed to an error flag (i.e. 
-1).  In the case of adding or altering a field, then a new entry should 
be added to the array.  This will prevent corruptions since the Unsafe 
operations will check the entry for the error flag.

The engineer will have to be aware that if they change the size of the 
field from a long to an int, then they have to change the 
AtomicLongFieldUpdater to an AtomicIntegerFieldUpdater.  Currently, the 
compiler doesn't flag a problem if an Atomic__FieldUpdater specifies a 
non-existent field or a field of the wrong type or non-volatile.  These 
checks are left for runtime.  Maybe the compiler should make this check 
as well.

The code that deals with dead, rarely used and oversized fields can't 
change the fields that were used by Unsafe.___FieldOffsets. If a field 
is changed and then Unsafe.___FieldOffset is called on that field, then 
those fields will have to be reverted.

Is there a better solution?
-- 
Nathan Reynolds 
<http://psr.us.oracle.com/wiki/index.php/User:Nathan_Reynolds> | 
Architect | 602.333.9091
Oracle PSR Engineering <http://psr.us.oracle.com/> | Server Technology
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.oswego.edu/pipermail/concurrency-interest/attachments/20130604/0333b8c1/attachment.html>


More information about the Concurrency-interest mailing list