Re: Segmentation fault in s23_clnt.c when used in a library

On 2014-10-15 13:57, Aaron Richton wrote:

On Wed, 15 Oct 2014, Tobias Ljunggren wrote:

...

I built with moznss instead of openSSL. That solves the problem and I don't try to fix someone else mistakes.

It might hide the problem, but that doesn't fix anything. It's worth reiterating up front that MozNSS isn't ideal, especially if you're planning to use slapd(8) as a server.

But more importantly, you're probably just kicking the can. You're noticing a conflict on SSL_new today, but it could just as easily be malloc(3) or strcmp(3) tomorrow. Are you seriously going to change OpenLDAP to use uClibc or who-knows-what as a way of tricking the linker? The long-term answer is ensuring that everything is linked to "the normal libraries" in the way that you'd expect.

You are right, and I would if I could. Problem is that the application that starts the process (which loads my module) is closed source and I can't do anything about it. This is a module that will update a ldap directory based on events from another system (or update the system based on changes in a LDAP directory), this only concerns the libraries (liblber and libldap_r) not slapd. I've checked with both ldd(1) and nm(1) and system libraries like libc*, libgcc*, libstdc++ and libpthread* are linked correctly.

How to accomplish this is somewhat platform-dependent, although the principles are always the same. On Solaris, for example, -Bdirect and mapfiles would help in this situation. ld from GNU binutils has other -B options that may be relevant, and "dynamic-list" functionality. etc etc. Bottom line, this is a fundamental build issue, and changing the color of the Legos doesn't really address the root cause.

As far as I understand, today, I can't really do anything about this since the application that cause the openSSL conflicts is closed source?

...

Best regards, Tobias

On 2014-10-08 14:59, Tobias Ljunggren wrote: Hello,

  I just found what's causing the crash. The application that loads my library already have some of the openSSL 

functions. Either they have copied the source code or statically linked to openSSL. Because of this openLDAP sometimes uses the wrong functions (SSL_new is one of them).

  I guess my only option is to build openLDAP with a static linkage to openSSL?

  Best regards,
  Tobias

  On 2014-10-07 17:06, Tobias Ljunggren wrote:
        Hello,

        I've been trying to track a segmentation fault for a couple of days but I'm running out of ideas and need some 

help or suggestion on how to proceed.

        Perhaps this is the wrong forum and if so I apologize. It's also worth mentioning that I don't have any knowledge 

about the openSSL library source code but I've been working with ldap libraries for quite some time (about 10 years).

        What I have is shared object (.so) that is loaded from another program (nothing I have control of) using a public 

API. My library loads openLDAP (libldap_r and liblber) using dlopen/dlsym. After some initialization code I try to bind to a ldap source (ldaps://ldaphost:636) but I get a segmentation fault in s23_clnt.c at line 159: else if (s->ctx->info_callback != NULL)

        For some reason s->ctx is a null pointer. My first thoughts was that I do something I shouldn't prior to 

ldap_sasl_bind_s which corrupts the LDAP handle but after days with gdb and valgrind I still can't find anything.

        openLDAP version used is 2.4.40 built with openSSL 1.0.1i.

        If I copy the ldap code to a test executable it works fine.

        Is there anything special to consider when loading the libraries from a shared object?

        I understand that it is almost impossible to give me any good answers but since I'm a bit frustrated I hope 

someone can give me some hints.

        Best regards,
        Tobias