Re: make_services dumps core

[David Ayers]

Chris B. Vetter schrieb:

> anyone also experiencing core dumps of make_services when installing
> the latest CVS versions of -gui? I first noticed the problem with
> yesterday's checkout. Same problem today.
> 
> The interesting thing is that I added a simple fprintf() as the very
> first line (even before variable definitions) in main(), and core is
> dumped before fprintf() even gets called...
> 
> The system hasn't been changed, no required or optional dependencies
> have been updated within the last few days and -gui compiled fine with
> Friday's checkout. Still don't have internet at home so I dunno about
> the weekend versions.
> 
> Using gdb, all I can figure out is that it apparently receives a
> SIGSEGV when calling pthread_key_create(). The only reference to that
> call is in base/Source/thr-pthread.m
> 
> Any ideas what's happening?

Hello Chris,

I assume this is the libpthread in conjunction with libobjc issue:

http://savannah.gnu.org/bugs/?func=detailitem&item_id=13800

The problem is that some (i.g. FreeBSD's) implementations of libpthread
rely on a constructor function:

void _thread_init_hack(void) __attribute__ ((constructor));
void
_thread_init_hack(void)
{
_libpthread_init(NULL);
}

to have the internal structures of libpthread initialzed instead of
guardinging the public api (like pthread_key_create())with:

if (_thr_initial == NULL)
  _libpthread_init(NULL);

libobjc also uses a constructor function to setup the threading
environment and happens to call pthread_key_create() first.  Now whether
the libpthread constructor is called before the libobjc constructur is
(also?) dependent on the order -l link directives.  I'm not aware of any
recent changes in -make though, so it might also be dependent on the
phase of the moon ;-).

The bug is in the libpthread implementation that should guard it's
public API.  It seems there was a bug filed for this for FreeBSD but I'm
not  sure what the state of that is.

Cheers,
David