icedtea-1.14 JamVM + OpenJDK 7 - oj7-remove-sun.misc.PerfCounter-debug-code.patch

[Xerxes Rånby]

fre 2011-05-27 klockan 14:03 +0100 skrev Robert Lougher:
> On 27 May 2011 12:55, Xerxes Rånby <xerxes at zafena.se> wrote:
> > fre 2011-05-27 klockan 12:48 +0200 skrev Xerxes Rånby:
> >> > What I currently observe are a
> >> > unsatisfied link error.
> >> >
> >> > xranby at xranby-ESPRIMO-P7935:~/test
> >> > $ ../icedtea-1.14/openjdk.build/j2sdk-image/bin/java Linpack
> >> > java.lang.UnsatisfiedLinkError: registerNatives
> >> >     at sun.misc.Perf.registerNatives(Native Method)
> >> >     at sun.misc.Perf.<clinit>(Perf.java:536)
> >> ...
> >> >     at sun.misc.PerfCounter.<clinit>(PerfCounter.java:51)
> >> >     at java.net.URLClassLoader.defineClass(URLClassLoader.java:448)
> >> >
> >> > The URLClassLoader uses sun.misc.Perf to register startup time.
> >> >
> >> > I start to think this are a bug in OpenJDK 7 that the sun.misc.Perf
> >> > constructor expects to be able to use the native registerNatives method
> >> > in libhprof.so without explicilty making sure that this library have
> >> > already been loaded.
> >> >
> >> > The libhprof.so are part of the Hotspot JVMTI implementation so it are
> >> > no surprise that JamVM currently dont not load it.
> >
> > The attached patch fix this UnsatisfiedLinkError by simply removing the
> > use of sun.misc from the OpenJDK 7 ClassLoader and URLClassLoader. This
> > makes it possible to use the OpenJDK 7 classes in combination with other
> > JVM's than Hotspot.
...
> >
> > If someone like this debug code and have a suggestion how to properly
> > re-engineer sun.misc to support other JVM's than Hotspot im all ear.
> >
> 
> I was in the middle of writing the previous email when I was dragged
> off to lunch.  I think this is a valid fix.  However, I would like to
> understand the mechanism which is enabling HotSpot to work.
hard to tell the use of registerNatives in OpenJDK 7 sun.misc.Perf
constructor are part of the initial "duke" commit
http://hg.openjdk.java.net/jdk7/jdk7/jdk/annotate/c344779fab58/src/share/classes/sun/misc/Perf.java

I have been able to track down the use of sun.misc.PerfCounter in the
OpenJDK7 classes  to this commit:
http://hg.openjdk.java.net/jdk7/jdk7/jdk/rev/c7e469ae3edb
http://bugs.sun.com/view_bug.do?bug_id=6878481
by following the bugreport chain its clear that this commit was added to
find the root of the "On Windows System.nanoTime() may be 25x slower
than System.currentTimeMillis()" bug
http://bugs.sun.com/view_bug.do?bug_id=6440250


> Of
> course, once the native functions can be found, there's no guarantee
> that they will work, as they may use other unimplemented features.
> 
> Rob.

sun.misc.PerfCounter debug code have been fount to be used in four
places of openjdk 7:
ClassLoader, URLClassLoader, ZipFile and D3DGraphicsDevice

With the attached oj7-remove-sun.misc.PerfCounter-debug-code.patch in
place jamvm + oj7 now passes most of the hotspot jtreg test:


hotspot results: passed: 101; failed: 6; error: 1


* BUGS:
FAILED: compiler/7009231/Test7009231.java  this looks like a genuine
bug, it fails using openjdk 6 + jamvm as well while it pass using
hotspot and cacao on i386.
FAILED: runtime/6929067/Test6929067.sh    fails because it try to open
libjvm.so.0  instead of libjvm.so this are fault in the test itself or
an icedtea bug. I can make this test pass by creating a symlink from
libjvm.so.0 to libjvm.so
FAILED: runtime/6819213/TestBootNativeLibraryPath.java    looks to be a
bootclasspath bug   Error initialising natives: couldn't open
libjava.so: use -verbose:jni for more information

* Needs more investigation, possible bugs.
Error:  compiler/6901572/Test.java a speed timeout..   contains a lot of
loops that do not finish in 120seconds.
FAILED: runtime/6626217/Test6626217.sh I am not sure what this are
supposed to do..  got a warning: [options] bootstrap class path not set
in conjunction with -source 1.4

* Safe to ignore:
FAILED: gc/6581734/Test6581734.java can be ignored, test looks for some
specific debug info reported by -verbose:gc -XX:+UseConcMarkSweepGC
FAILED: runtime/6981737/Test6981737.java execStatus=Failed. Execution
failed: `main' threw exception: java.lang.RuntimeException: FAIL: Wrong
value for java.vm.vendor property, "Robert Lougher", expected to be of
form: "Oracle Corporation"



When I try to run the langtools or jdk tests using JamVM + oj7 then I
hit another UnsatisfiedLinkError for sun.misc.Unsafe.copyMemory. This
method looks possible to implement by using the existing JamVM copyArray
method that happens to take similar arguments.

Exception in thread "main" java.lang.UnsatisfiedLinkError: copyMemory
	at sun.misc.Unsafe.copyMemory(Native Method)
	at
sun.nio.fs.NativeBuffers.copyCStringToNativeBuffer(NativeBuffers.java:127)
	at
sun.nio.fs.UnixNativeDispatcher.copyToNativeBuffer(UnixNativeDispatcher.java:50)
...

Xerxes

-------------- next part --------------
A non-text attachment was scrubbed...
Name: oj7-remove-sun.misc.PerfCounter-debug-code.patch
Type: text/x-patch
Size: 3401 bytes
Desc: not available
Url : http://mail.openjdk.java.net/pipermail/distro-pkg-dev/attachments/20110527/5094365d/oj7-remove-sun.misc.PerfCounter-debug-code.patch