java VM SEGV on 2.0.98

[pw@dancer.ca.sandia.gov (Pete Wyckoff)]

Something (presumably) in the jdk1.1.6v5 VM is corrupting glibc
structures during dynamic symbol lookup.  I know, what do you care?
But a hack to glibc "fix"es it.  linux-2.1.125, ix86,
glibc-2.0.98 with "-g -O", all shared libraries.


The java executable (from blackdown, built against glibc-2.0.7, they claim)
fails with stack trace:

#0  0x400a01e9 in dlsym (handle=0xffffffff, name=0x4007b3e4 "getmsg")
    at dlsym.c:87
#1  0x40067839 in socket (domain=-1073744800, type=1024, protocol=1024)
    at ../../../../src/genunix/java/green_threads/src/iomgr.c:255
#2  0x40067952 in InitializeAsyncIO ()
    at ../../../../src/genunix/java/green_threads/src/iomgr.c:272
#3  0x4006307b in sysInitializeJavaVM ()
    at ../../../../src/genunix/java/javai/javai.c:801
#4  0x40058561 in JNI_CreateJavaVM ()
    at ../../../../src/share/java/runtime/jni.c:3924
#5  0x40062834 in java_main ()
    at ../../../../src/genunix/java/javai/javai.c:149
#6  0x8061db6 in main (argc=2, argv=0xbffff5a4, envp=0xbffff5b0)
    at ../../../../src/genunix/java/javai/java.c:31


I blindly modified elf/dlsym.c to avoid the problem as follows:

dlsym (void *handle, const char *name)
{     
  struct dlsym_args args;
  args.caller = (ElfW(Addr)) __builtin_return_address (0);
  args.handle = handle;
  args.name = name;
  
#if 0  /* original */
  return (_dlerror_run (dlsym_doit, &args)
          ? NULL : (void *) (args.loadbase + args.ref->st_value));
#else
  if (_dlerror_run (dlsym_doit, &args)) {
    return NULL;
  } else {
    if (args.ref)  /* this dereference was causing SEGV */
      return (void *) (args.loadbase + args.ref->st_value);
    else {
      printf("dlsym: not expecting NULL args, name = %s\n", name);
      return NULL;
    }
  }
#endif
}


The symbols which cause problems are then printed out:

dlsym: not expecting NULL args, name = getmsg
dlsym: not expecting NULL args, name = putmsg
dlsym: not expecting NULL args, name = Java_java_lang_System_currentTimeMillis_stub
dlsym: not expecting NULL args, name = Java_java_lang_System_arraycopy_stub
dlsym: not expecting NULL args, name = Java_java_lang_Class_getPrimitiveClass_stub
...


Perhaps there's something about the "stub"-ness of these things.  get/putmsg
are the only glibc symbols which appear, the rest are from jdk libs.  An
older jdk (1.1.3) used to work on this machine under linux-2.1.x and some
single-digit glibc-2.0.x, but not with the new glibc-2.0.98.

Sorry for the lack of info.  Just a warning in case more reports start
rolling in.

			-- Pete
---------------------------------------------
Pete Wyckoff          | wyckoff@ca.sandia.gov
Sandia National Labs  | 925 294 3503 (voice)
MS 9011, P.O. Box 969 | 925 294 1225 (fax)
Livermore, CA  94551  |