Name: kb87695 Date: 07/11/2001


Please evaluate the bug below to determine if it works as designed or if this is
actually a bug.

This bug could be reproduced with Solaris Java 1.3.1 Build 1.3.1-b24. The bug
cannot be reporduced in Merlin latest.

Basically, the problem is that the JVMPI_OBJ_ALLOC of java.lang.String have a
posted object size which does not match the actual instance size. The same
problem could happen to other classes, not necessarily only String.

The testcase bwlow will print out all CLASS_LOAD/OBJ_ALLOC/OBJ_FREE/OBJ_MOVE
events concering java.lang.String class. Note for this small example, No GC kick in
at all, so OBJ_FREE/OBJ_MOVE just never happens.

The problem is related to the TLE usage. If you turn off TLE by -XX:-UseTLE, it
will print out same size for all instances.

The problem seems only occuring with the server VM, but not with Cleint VM.

hppdl694$ java -server -version
java version "1.3.1"
Java(TM) 2 Runtime Environment, Standard Edition (build 1.3.1-b24)
Java HotSpot(TM) Server VM (build 1.3.1-b24, mixed mode)

1, Save the objects.java file

public class objects {
        public static void main(String[] args) {
                int N = 1000;
                for (int i = 0; i < N; i++) {
                        String str = new String();
                }
        }
}

2, Save the jvmpi_tool.cpp file

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <jvmpi.h>
#ifdef _WINDOWS
# include <windows.h>
# define SLEEP(s) Sleep(s * 1000);
#else
# define SLEEP(s) sleep(s);
#endif

#define JVMPI(m) (jvmpi->m)

static JVMPI_Interface *jvmpi;
static void notify(JVMPI_Event *);


extern "C" JNIEXPORT jint JNICALL
JVM_OnLoad(JavaVM *jvm, char *options, void *reserved)
{
        int res = jvm->GetEnv((void **)&jvmpi, JVMPI_VERSION_1);

        if (res < 0) {
                return JNI_ERR;
        }

        jvmpi->NotifyEvent = notify;

        JVMPI(EnableEvent)(JVMPI_EVENT_JVM_INIT_DONE, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_JVM_SHUT_DOWN, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_CLASS_LOAD, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_OBJECT_ALLOC, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_OBJECT_MOVE, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_OBJECT_FREE, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_GC_START, NULL);
        JVMPI(EnableEvent)(JVMPI_EVENT_GC_FINISH, NULL);

        return JNI_OK;
}

static unsigned long moved, freed;
static jobjectID stringClassID=(jobjectID)-1;

void
notify(JVMPI_Event * event)
{
        jint result;

        switch (event->event_type) {
        case JVMPI_EVENT_JVM_INIT_DONE:
                break;
        case JVMPI_EVENT_JVM_SHUT_DOWN:
                printf("\nSHUT_DOWN");
                break;
        case JVMPI_EVENT_CLASS_LOAD:
                if (!strcmp(event->u.class_load.class_name, "java.lang.String") ||
                        !strcmp(event->u.class_load.class_name, "java/lang/String")) {
                        stringClassID=event->u.class_load.class_id;
                        printf("\nCLASS_LOAD: class_name: %s source_name: %s class_id %ld",
                                event->u.class_load.class_name,
                                event->u.class_load.source_name,
                                event->u.class_load.class_id);
                }
                break;
        case JVMPI_EVENT_OBJECT_ALLOC:
                if (event->u.obj_alloc.class_id==stringClassID) {
                   //printf("\nOBJECT_ALLOC: arena: %d class_id: %ld is_array: %d size %d obj_id %ld",
                        printf("\nOBJECT_ALLOC: arena: %d class_id: %p is_array: %d size %d obj_id %p",
                                event->u.obj_alloc.arena_id,
                                event->u.obj_alloc.class_id,
                                event->u.obj_alloc.is_array,
                                event->u.obj_alloc.size,
                                event->u.obj_alloc.obj_id);
                }
                break;
        case JVMPI_EVENT_GC_START:
                printf("\nGC_START");
                moved = freed = 0;
                break;
        case JVMPI_EVENT_OBJECT_MOVE:
                if (event->u.obj_alloc.class_id==stringClassID) {
                        printf("\nOBJECT_MOVE: arena: %d obj_id: %ld new_arena_id: %d new_obj_id %ld",
                                event->u.obj_move.arena_id,
                                event->u.obj_move.obj_id,
                                event->u.obj_move.new_arena_id,
                                event->u.obj_move.new_obj_id);
                }
                ++moved;
                break;
        case JVMPI_EVENT_OBJECT_FREE:
                if (event->u.obj_alloc.class_id==stringClassID) {
                        printf("\nOBJECT_FREE: obj_id: %ld",
                                event->u.obj_free.obj_id);
                }
                ++freed;
                break;
        case JVMPI_EVENT_GC_FINISH:
                printf("\nGC_FINISH: moved %lu, freed %lu\n", moved, freed);
                fflush(stdout);
                break;
        case JVMPI_REQUESTED_EVENT|JVMPI_EVENT_HEAP_DUMP:
                printf("\nGot HEAP_DUMP");
                break;
        default:
                break;
        }
}

3, Save the makefile

CXX = CC
COMMON = -c
CXXFLAGS = -I. -I/opt/jdk1.3.1/include -I/opt/jdk1.3.1/include/solaris $(COMMON)

LDFLAGS = -G
LD = ld

all: libtest.so objects.class

jvmpi_tool.o : jvmpi_tool.cpp
        $(CXX) $(CXXFLAGS) jvmpi_tool.cpp

libtest.so: jvmpi_tool.o
        $(LD) $(LDFLAGS) -o libtest.so jvmpi_tool.o -lrt -lpthread

objects.class : objects.java
        javac objects.java

clean :
        rm *.class *.so *.o

I assume the jdk1.3.1 is installed under /opt/jdk1.3.1 directory, and CC and LD installed
properly in your PATH.

4, Build everything by typing make. You should get the libtest.so and the objects.class
files.

5, Run the testcase now by typing
        java -server -cp . -Xruntest objects

It is not happening with the default client VM, only with the server VM.

You will see the output from the libtest.so giving the OBJ_ALLOC events for all instances
of java.lang.String. The fields of each line is
        Arena ID, class ID, is_array indicator, bytesize of objects, object ID

6, What is the problem?

Part of the out:

OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d72960
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d729a8
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d729f0
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 2 size 16 obj_id f3d72a08
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d72a18
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 16384 obj_id f3d74358 <--- 16384 but not 24
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d74358
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 16384 obj_id f3d78358 <--- 16384 but not 24
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d78358
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 16384 obj_id f3d7c358 <--- 16384 but not 24
OBJECT_ALLOC: arena: 4 class_id: f6809f98 is_array: 0 size 24 obj_id f3d7c358
OBJECT_ALLOC: arena: 5 class_id: f6809f98 is_array: 0 size 24 obj_id f68a6798

The spurious ones are pointed by the arrow. While the instance size is actually
24 bytes on solaris, three String object allocation is posted with 16384 bytes.

7, How did this happen?

In the debugger I saw that the to-be-allocated object could not fit into the
TLE, hence a TLE expansion is supposed to happen. In universe::clear_allocate_local,
the new_tle_size is computed, this new_tle_size is of 16384 bytes. Then the VM does
Universe::clear_allocate_shared(klass, new_tle_size, CHECK_0) to allocate the
String object. In clear_allocate_shared it will post the JVMPI_OBJ_ALLOC_EVENT
with the new_tle_size. This is how this non-matched String object size comes into
the picture. If turning off TLE by flag -XX:-UseTLE, it will go away.

I don't know if this is a design issue or an impl issue. If it is a design issue
and it is working as designed, then it's ok, I will withdraw the bug report and
tell our customer. Otherwise we'd like to have a fix ASAP.
(Review ID: 127896)
======================================================================