Re: dlopen(), atexit() crash on FreeBSD (testcase included)

On Mon, 31 Dec 2007 17:35:10 +0100
"Markus Hoenicka" <markus.hoenicka@xxxxxxxxxxxx> wrote:

Hi,

I've been redirected by Giorgos Keramidas to this list after reporting
a problem on the freebsd-questions list. I'd greatly appreciate if you
could have a look at the following problem. Apparently programs are
doomed to segfault on FreeBSD if dlopen()ed modules install exit
handlers via atexit(). Similar problem reports have cropped up before,
see e.g.

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

My system runs:

FreeBSD yeti.mininet 6.1-RELEASE FreeBSD 6.1-RELEASE #1: Mon Aug 28
22:24:48 CEST 2006
markus@xxxxxxxxxxxx:/usr/src/sys/i386/compile/YETI i386

I'm one of the developers of libdbi, a database abstraction layer for
C, see http://libdbi.sourceforge.net.

libdbi is a library for programs which are supposed to be able to
access different database engines with a unified API. libdbi
essentially maps generic API calls to the specific database client
library calls of a particular database engine. To do this, libdbi
loads available database drivers at runtime via dlopen() calls. Each
of these drivers is linked against one database client
library. E.g. the Firebird driver is linked against
libfbclient.so. When libdbi is properly shut down, it unloads all
loaded drivers by calling dlclose() on each of them.

This design works well on all supported platforms and with all
supported database engines, with one exception: the Firebird driver on
FreeBSD invariably causes a segfault when the application linked
against libdbi exits:

#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 reason appears to be that the Firebird client libraries install
exit handlers via atexit(). Remember that due to libdbi's design to
load all available drivers whether or not they are used later, libdbi
will cause a crash even if no Firebird database is accessed - it is
sufficient that the driver has been loaded. As per Giorgos' suggestion
it is simple to circumvent this segfault by avoiding the call to
dlclose() before exiting, but I wonder whether there is a more robust
solution for this problem.

The attached minimal testcase is sufficient to illustrate the
problem. atexitmod.c defines a module which is loaded by datest.c Make
sure to fix the hardcoded path in datest.c before building the app. To
build the test program and watch it crash, do the following:

gcc -shared -o atexitmod.so atexitmod.c
gcc -o datest datest.c
./datest

Commenting out either the atexit() call in atexitmod.c or the
dlclose() call in datest.c prevent the segfault.

If you find some solution, please cc me as I'm not subscribed to
freebsd-hackers.

regards,
Markus

As designed. atexit should not be used by shared objects that do not
expect themselves to live until actual exit() happens. ELF provides
proper _init/_fini sections to support shared object
initialization/destruction.

--
Alexander Kabaev

Attachment: signature.asc
Description: PGP signature

Relevant Pages