Re: [Bug libgcj/13212] JNI/CNI AttachCurrentThread does not register thread with garbage collector

[Richard Guenther <rguenther at suse dot de>]

On Sat, 25 Mar 2006, Richard Guenther wrote:

> On Fri, 24 Mar 2006, aph at gcc dot gnu dot org wrote:
> > ------- Comment #28 from aph at gcc dot gnu dot org  2006-03-24 17:41 -------
> > Richard Guenther: should this be posted to java@gcc.gnu.org for discussion?
> > 
> > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13212
> 
> Hmm, I guess so.  Me and Michael Matz have tried to somehow work around
> the gjc GC issues with OpenOffice which does open libgcj via dlopen and
> such suffers from the thread registration problem.  On the 4.1 branch
> there is the workaround to have libgcj provide pthread_create which should
> in theory override libpthread.so pthread_create.  This works only if
> you link against libgcj, not if you later dlopen it, because you then
> bound to say pthread_create@GLIBC_2.1.  To work around this issue we made
> libgcj provide pthread_create with the same symver as glibc does.  This
> exposes another issue, namely boehm GC being configured with
> GC_all_interior_pointers set to 1 by default - which causes GC_bl_init
> to not initialize GC_old_normal_bl and others.  libgcj unfortunately
> sets GC_all_interior_pointers to 0 at init time, which then causes the
> uninitialized GC_old_normal_bl to be used and we segfault.  To work around
> this we changed boehm to default to GC_all_interior_pointers 0 instead.
> 
> Now, while this seems to solve the OpenOffice problems if you LD_PRELOAD
> libgcj, the change seems to cause libjava testsuite failures (which we
> didn't yet investigate):

Michael just noticed I messed up during cleaning up the patch:

> Index: boehm-gc/pthread_support.c
> ===================================================================
> *** boehm-gc/pthread_support.c	(revision 112314)
> --- boehm-gc/pthread_support.c	(working copy)
> *************** constr (void)
> *** 1231,1237 ****
>   
>   
>   int
> ! GC_PTHREAD_CREATE_NAME(pthread_t *new_thread,
>   		  const pthread_attr_t *attr,
>                     void *(*start_routine)(void *), void *arg)
>   {
> --- 1231,1237 ----
>   
>   
>   int
> ! GC_pthread_create (pthread_t *new_thread,
>   		  const pthread_attr_t *attr,
>                     void *(*start_routine)(void *), void *arg)
>   {
> *************** GC_PTHREAD_CREATE_NAME(pthread_t *new_th
> *** 1308,1313 ****
> --- 1308,1315 ----
>       return(result);
>   }
>   
> + __asm__ (".symver pthread_create, pthread_create@@" GC_PTHREAD_SYM_VERSION);
> + 
>   #ifdef GENERIC_COMPARE_AND_SWAP
>     pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
>   

Should read

__asm__ (".symver GC_pthread_create, pthread_create@@" GC_PTHREAD_SYM_VERSION);

This way libjava tests fine.

Richard.