I think I've changed my mind regarding this patch.
At least with first caller wins, the function pointers don't bounce
around as they presently do.
-- Kurt
On Sep 24, 2008, at 7:38 AM, Kurt Zeilenga wrote:
Eric,
I think this patch doesn't offer a general solution to the problem.
The root problem here is that neither Cyrus SASL or OpenLDAP
libraries are truly reentrant. Each does offer some amount of
thread-safety, but are each limited by their design. Obviously,
both libraries could be improved by redesigning them to have less,
or at least, less irritating limitations.
Anyways, all this patch does is changes which caller wins the
initialization conflict. Where as before, last initializer won, now
first initializer wins. So this patch will lead to same sorts of
'weird crash' behavior as before when two callers make incompatible
initializations.
I think you need to rework your program to avoid incompatible
initialization. This basically means reworking the code so that
only one call to sasl_set_mutex is made. Unfortunately, if you are
using multiple libraries that each make their own calls to
sasl_set_mutex, you are kind of screwed. If you only are using one
such library (such as libldap_r), you need to make sure it's the
only caller and that it's called early enough for all uses of Cyrus
SASL.
Unfortunately, things like NSS and PAM can really cause a lot of
incompatible initialization that outside the control of the program.
-- Kurt
[resent from subscribed address]
On Sep 20, 2008, at 5:41 AM, Eric Leblond wrote:
Hi,
When working on NuFW (http://www.nufw.org), I've encounter some
weird crash
when calling sasl_dispose. The problem was in fact a logic problem in
sasl_set_mutex.
NuFW uses sasl and libldap_r. NuFW has to do a call to sasl_set_mutex
because it is multithreaded. One of the NuFW module uses libldap_r
which
also does a call to sasl_set_mutex. By doing this, we run into a
problem
because sasl_MUTEX_* function change during run time. Thus we can
allocate a mutex with NuFW function and destroy it with libldap_r
function. This lead to a crash in almost all cases.
IMHO, the only clean workaround is to modify sasl_set_mutex(): it
should
not be run twice in the same program. I attach a simple patch to this
mail which implement this behaviour.
BR,
--
Eric Leblond
INL: http://www.inl.fr/
NuFW: http://www.nufw.org/
<sasl_set_mutex_fix.diff>
