dlopen(), atexit() on FreeBSD

Hi,

I bumped into a platform-specific problem when using the
Firebird
database client library in a dlopen()ed module on FreeBSD.
libdbi
(http://libdbi.sourcefor
ge.net) is a database abstraction layer which
dlopen()s available database drivers at runtime to provide
connectivity to various database engines. This design works
without
problems on a variety of platforms and with a variety of
database
client libraries, but causes a segfault with Firebird on
FreeBSD:

#0  0x28514fe4 in ?? ()
#1  0x281507c3 in __cxa_finalize () from /lib/libc.so.6
#2  0x281503fe in exit () from /lib/libc.so.6
#3  0x0804a40f in main (argc=1, argv=0xbfbfe754) at
test_dbi.c:419

The application crashes when exit() is called. Googling told
me that
__cxa_finalize () is invoked by atexit(). Our drivers and
apps do not
use this function, but the firebird client libraries do:

markusyeti:~/prog/libdbi-drivers/tests# grep atexit
/usr/local/lib/libfb*
Binary file /usr/local/lib/libfbclient.so matches
Binary file /usr/local/lib/libfbembed.so matches

Googling also told me that the conflict between atexit() and
dlopen()
on FreeBSD is a known problem, see e.g.:

http://www.imagemagick.org/pipermai
l/magick-developers/2006-March/002523.html

Is there anything I can do about this from my end?

regards,
Markus
-- 
Markus Hoenicka
markus.hoenickacats.de
(Spam-protected email: replace the quadrupeds with
"mhoenicka")
http://www.mhoenicka.de

_______________________________________________
freebsd-questionsfreebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-que
stions
To unsubscribe, send any mail to
"freebsd-questions-unsubscribefreebsd.org"

On 2007-12-30 18:49, Markus Hoenicka <markus.hoenickamhoenicka.de> wrote:
> Hi,
> I bumped into a platform-specific problem when using
the Firebird
> database client library in a dlopen()ed module on
FreeBSD. libdbi
> (http://libdbi.sourcefor
ge.net) is a database abstraction layer which
> dlopen()s available database drivers at runtime to
provide
> connectivity to various database engines. This design
works without
> problems on a variety of platforms and with a variety
of database
> client libraries, but causes a segfault with Firebird
on FreeBSD:
> 
> #0  0x28514fe4 in ?? ()
> #1  0x281507c3 in __cxa_finalize () from
/lib/libc.so.6
> #2  0x281503fe in exit () from /lib/libc.so.6
> #3  0x0804a40f in main (argc=1, argv=0xbfbfe754) at
test_dbi.c:419
> 
> The application crashes when exit() is called. Googling
told me that
> __cxa_finalize () is invoked by atexit(). Our drivers
and apps do not
> use this function, but the firebird client libraries
do:
> 
> markusyeti:~/prog/libdbi-drivers/tests# grep atexit
/usr/local/lib/libfb*
> Binary file /usr/local/lib/libfbclient.so matches
> Binary file /usr/local/lib/libfbembed.so matches
> 
> Googling also told me that the conflict between
atexit() and dlopen()
> on FreeBSD is a known problem, see e.g.:
> 
> http://www.imagemagick.org/pipermai
l/magick-developers/2006-March/002523.html
> 
> Is there anything I can do about this from my end?

The __cxa_finalize() function is not called by atexit(), but
by exit()
itself.  I don't know of any way to `unregister' exit
handlers, so one
way of `fixing' this is to avoid calling dlclose() before
exit() in the
example code shown at the URL above:

% #include <stdio.h>
% #include <dlfcn.h>
%
% int main( int argc, char ** argv)
% {
%         void * handle;
%         const char * mod =
"/usr/local/lib/libMagick.so";
%         void (*InitializeMagick)(const char*);
%
%         handle = dlopen( mod, RTLD_LAZY);
%         if ( !handle) {
%                 fprintf( stderr, "cannot load
%sn", mod);
%                 exit(1);
%         }
%
%         (void*) InitializeMagick = dlsym( handle,
"InitializeMagick");
%         if ( !InitializeMagick) {
%                 fprintf( stderr, "cannot resolve
InitializeMagickn");
%                 exit(1);
%         }
%
%         InitializeMagick( "moo");
%         dlclose( handle);
% }

Since the program is going to exit and have all its dlopened
shared
objects be unmapped, it's probably ok to skip the dlclose()
step in
this example.

In the case of the firebird libraries, since they are
libraries things
are a bit trickier, because the consumers of these libraries
don't
really know that a dlopened object has called atexit() :(

I think this is probably something that the freebsd-hackers
list will
be interested in.  Can you post a description of the problem
there too?

- Giorgos

_______________________________________________
freebsd-questionsfreebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-que
stions
To unsubscribe, send any mail to
"freebsd-questions-unsubscribefreebsd.org"

Giorgos Keramidas writes:
 > The __cxa_finalize() function is not called by
atexit(), but by exit()

I see.

 > Since the program is going to exit and have all its
dlopened shared
 > objects be unmapped, it's probably ok to skip the
dlclose() step in
 > this example.
 > 

Just for the record: I've tried disabling the dlclose() call
in our
code. It both prevents the segfaults and does not seem to
inflict any
harm. I'll use the code that way for the time being, any
real solution
for this problem notwithstanding.

 > I think this is probably something that the
freebsd-hackers list will
 > be interested in.  Can you post a description of the
problem there too?
 > 

I'll do. I haven't read up the full history of this problem,
but
apparently other programs and libraries are affected too, so
it is
probably fair to seek assistance.

regards,
Markus

-- 
Markus Hoenicka
markus.hoenickacats.de
(Spam-protected email: replace the quadrupeds with
"mhoenicka")
http://www.mhoenicka.de

_______________________________________________
freebsd-questionsfreebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-que
stions
To unsubscribe, send any mail to
"freebsd-questions-unsubscribefreebsd.org"