build crash after extending oopDesc

Tue Jun 12 06:31:27 PDT 2012

FYI, increasing the size of the oopDesc indeed became more complex that 
in the early version of HotSpot.

We had to do that work in a JDK5 code base (to implement the RTGC in 
JavaRTS).

There are indeed a few issues with perm gen objects. If I remember 
correctly, in addition to the problems mentioned by Coleen, these 
objects contains a few variable size arrays used by the JVM and some of 
the offsets were not taking into account the extra fields in the header.
You should double check all the size computations in these classes.

In addition, you may have some alignment issues with 64-bit fields if
your header is no longer aligned on a 8-bytes boundary.

Bertrand.

On 6/12/12 3:08 PM, Coleen Phillimore wrote:
>
> It looks like the hotspot code expects the Klass vtable to be the third
> word. Can you send the stack where it crashed?
> Do you want to add this field for Java objects or just for the metadata
> objects like klassOop and methodOop?
> thanks,
> Coleen
>
> On 6/12/2012 7:57 AM, Simmen David wrote:
>> No matter where I put the field it always crashes at the same point
>>
>> MfG David
>>
>> -----Original Message-----
>> From: hotspot-dev-bounces at openjdk.java.net
>> [mailto:hotspot-dev-bounces at openjdk.java.net] On Behalf Of Coleen
>> Phillimore
>> Sent: Monday, June 11, 2012 3:19 PM
>> To: hotspot-dev at openjdk.java.net
>> Subject: Re: build crash after extending oopDesc
>>
>>
>> The entire VM knows that the _klass field is after the mark field. Try
>> moving your new field after that.
>> Coleen
>>
>> On 6/11/2012 4:31 AM, Simmen David wrote:
>>> Hello
>>> Thanks so far your help.
>>> Moving the new field after the _mark field did not help.
>>>
>>> The error still occurs when the build process tries to link hotspot.
>>> Alltough the error message now difers a bit it occurs still at the
>>> same place
>>>
>>> Here is the build output
>>> Linking vm...
>>> /usr/bin/objcopy --only-keep-debug libjvm.so libjvm.debuginfo
>>> /usr/bin/objcopy --add-gnu-debuglink=libjvm.debuginfo libjvm.so strip
>>> -g libjvm.so [ -f libjvm.debuginfo ] || ln -s libjvm.debuginfo
>>> libjvm.debuginfo zip -q -y libjvm.diz libjvm.debuginfo
>>> libjvm.debuginfo rm -f libjvm.debuginfo libjvm.debuginfo [ -f
>>> libjvm.diz ] || { ln -s libjvm.diz libjvm.diz; } echo Linking
>>> launcher...
>>> Linking launcher...
>>> gcc -m32 -march=i586 -Xlinker -O1 -Wl,--hash-style=both -Xlinker -z
>>> -Xlinker noexecstack -m32 -march=i586 -Xlinker -export-dynamic -L
>>> `pwd` -o gamma launcher/java_md.o launcher/jli_util.o
>>> launcher/wildcard.o launcher/java.o -ljvm -lm -ldl -lpthread
>>> make[6]: Leaving directory
>>> `/home/jdk/openjdku4/build/linux-i586/hotspot/outputdir/linux_i486_compiler2/product'
>>>
>>> All done.
>>> make[5]: Leaving directory
>>> `/home/jdk/openjdku4/build/linux-i586/hotspot/outputdir/linux_i486_compiler2/product'
>>>
>>> cd linux_i486_compiler2/product&& ./test_gamma Using java runtime at:
>>> /usr/lib/jvm/java-6-openjdk/jre # # A fatal error has been detected by
>>> the Java Runtime Environment:
>>> #
>>> # SIGSEGV (0xb) at pc=0x00000000, pid=6318, tid=1090440000 # # JRE
>>> version: 6.0_24-b24 # Java VM: OpenJDK Server VM (23.2-b03 mixed mode
>>> linux-x86 ) # Problematic frame:
>>> # C 0x00000000
>>> #
>>> # Failed to write core dump. Core dumps have been disabled. To enable
>>> core dumping, try "ulimit -c unlimited" before starting Java again # #
>>> An error report file with more information is saved as:
>>> #
>>> /home/jdk/openjdku4/build/linux-i586/hotspot/outputdir/linux_i486_comp
>>> iler2/product/hs_err_pid6318.log
>>> #
>>> # If you would like to submit a bug report, please visit:
>>> # http://bugreport.sun.com/bugreport/crash.jsp
>>> #
>>> Aborted (core dumped)
>>> make[4]: *** [product] Error 134
>>>
>>>
>>> and here the code
>>> class oopDesc {
>>> friend class VMStructs;
>>> private:
>>> volatile markOop _mark;
>>> jint NewField;
>>> union _metadata {
>>> wideKlassOop _klass;
>>> narrowOop _compressed_klass;
>>> } _metadata;
>>>
>>>
>>> MfG David
>>>
>>> -----Original Message-----
>>> From: hotspot-dev-bounces at openjdk.java.net
>>> [mailto:hotspot-dev-bounces at openjdk.java.net] On Behalf Of Thomas
>>> Wuerthinger
>>> Sent: Sunday, June 10, 2012 8:23 PM
>>> To: Remi Forax
>>> Cc: hotspot-dev at openjdk.java.net; Christian Thalinger
>>> Subject: Re: build crash after extending oopDesc
>>>
>>> Extensibility and flexibility are at least main goals of the Graal
>>> project. We are currently doing experiments around removing the
>>> restriction on the JIT by writing the interpreter in Java too. We
>>> will present our results at this year's JVM Language Summit.
>>>
>>> - thomas
>>>
>>>
>>> On 10.06.2012 12:55, Remi Forax wrote:
>>>> I think Graal (the OpenJDK project) is nowadays close to the original
>>>> hotspot in term of hackability even if it's restricted to the JIT.
>>>>
>>>> Rémi
>>>>
>>>> Sent from my Phone
>>>>
>>>> ----- Reply message -----
>>>> From: "John Rose"<john.r.rose at oracle.com>
>>>> To: "Krystal Mok"<rednaxelafx at gmail.com>
>>>> Cc: "hotspot-dev at openjdk.java.net"<hotspot-dev at openjdk.java.net>,
>>>> "Christian Thalinger"<christian.thalinger at oracle.com>
>>>> Subject: build crash after extending oopDesc
>>>> Date: Sat, Jun 9, 2012 09:37
>>>>
>>>>
>>>> On Jun 8, 2012, at 11:51 AM, Krystal Mok wrote:
>>>>
>>>>> David,
>>>>>
>>>>> If this patch [1] resembles your change, then what I saw was that
>>>> the VM would build all right, but would fail running test_gamma right
>>>> at the end of the build process.
>>>>> The problem was caused by arrayOopDesc::length_offset_in_bytes(), as
>>>> noted [2]. test_gamma runs fine if I drop a .hotspotrc file
>>>> specifying -UseCompressedOops in it, which confirms the problem.
>>>>
>>>> Good analysis, Kris.
>>>>
>>>> What happens if you add the extraneous field after the _mark field?
>>>> I think that would be more robust, although there is still likely to
>>>> be a bug tail.
>>>>
>>>> Here's an old war-story: When the original Hotspot team brought
>>>> their prototype to Sun long ago, I was on the technical evaluation
>>>> team.
>>>> One of my tests was to inject an extra field into class oopDesc.
>>>> (Specifically, the field held a string representation of the thread
>>>> backtrace at the point of the object's allocation. A crude
>>>> allocation
>>>> profiler!) The system worked with a few edits and no further
>>>> adjustment. This demonstrated to me that the software architecture
>>>> was robust and ready to evolve.
>>>>
>>>> (Things are different now. The system is more than 10x larger. It
>>>> contains multiple platform ports, two good JITs, instrumentation,
>>>> several GCs, and compressed pointers, none of which the prototype had.
>>>> I wish our system were still as malleable as it was when it was
>>>> young, but it's a different engineering problem today.)
>>>>
>>>> — John
>>>>
>>>> diff --git a/src/share/vm/oops/oop.hpp b/src/share/vm/oops/oop.hpp
>>>> --- a/src/share/vm/oops/oop.hpp
>>>> +++ b/src/share/vm/oops/oop.hpp
>>>> @@ -62,6 +62,7 @@
>>>> friend class VMStructs;
>>>> private:
>>>> volatile markOop _mark;
>>>> + intptr_t _extraneous;
>>>> union _metadata {
>>>> wideKlassOop _klass;
>>>> narrowOop _compressed_klass;
>>>>