Re: GETADDRINFO -- core dumped slapd (ITS#1304)

[peter.marschall@mayn.de]

Hi,

I think I have found one possible reason for some problems
with getaddrinfo() under Linux with GBU libc

Today (to be honest, it was 3 days ago), I had the same problem.
A freshly compiled slapd from openldap 2.0.17 crashed
during make test. 
I played around a bit and extracted a test case from 
tests/scripts/test000-rootdse:
  LD_LIBRARY_PATH=..... ../servers/slapd/slapd ....
This was the offending command. When I only called
  ../servers/slapd/slapd ....
and used the already installed (older) libraries it magically
worked.

Three days of work with a debugger and searching through
the whole system, I found the reason for this strange behaviour:
I have libbind.a and the corresponding header files from
bind 8.2.x on my system !!!

Unfortunately the definition struct addrinfo from GNU libc
collides with the same struct of bind 8.2.x: the fields ai_addr and
ai_canonname are swapped.

When being compiled, libldap*.so and slapd use the
/isr/include/netdb.h hader file form the system library, but
since since libldap_r.so is being linked together with libbind.a,
it provides the wrong version of getaddrinfo() to slapd.
The correct version of the funktion is in libc.so and is not being
called, since libldap_r.so is first in the list of dynamic libraries.

I filed a bug report to the bind people.
It  got the ticket number [BIND-BUGS #5027]

My workaround was wuite simple.
I simply changed the order in which libbind and libresolv
are checked for res_query and added a check for __res_query 
in libresolv
So, libbind does not get included in the link list for libldap_r.so,
and everything works O.K.

This is the patch:

--- configure.in@@ -709,6 +709,16 @@
+++ configure.in        Fri Oct 19 14:55:11 2001
 ol_link_dnssrv=no
 AC_CHECK_FUNC(res_query,:)
 if test $ac_cv_func_res_query = no ; then
+       AC_CHECK_LIB(resolv, res_query)
+       ac_cv_func_res_query=$ac_cv_lib_resolv_res_query
+fi
+
+if test $ac_cv_func_res_query = no ; then
+       AC_CHECK_LIB(resolv, __res_query)
+       ac_cv_func_res_query=$ac_cv_lib_resolv___res_query
+fi
@@ -718,10 +728,6 @@
        ac_cv_func_res_query=$ac_cv_lib_bind___res_query
 fi

-if test $ac_cv_func_res_query = no ; then
-       AC_CHECK_LIB(resolv, res_query)
-       ac_cv_func_res_query=$ac_cv_lib_resolv_res_query
-fi

 if test "$ac_cv_func_res_query" = yes ; then
        AC_DEFINE(HAVE_RES_QUERY,1,
+
+if test $ac_cv_func_res_query = no ; then
        AC_CHECK_LIB(bind, res_query)
        ac_cv_func_res_query=$ac_cv_lib_bind_res_query
 fi


Yours
Peter

PS: I read on the openldap ITS that one guy suggested writing
    your own implementation of etaddrinfo().
    Please do NOT do it !!!
    The collisions between GNU libc and bind 8 show the dangers
    of this approach !


-- 
Peter Marschall     |   eMail: peter.marschall@mayn.de
Scheffelstraße 15   |          peter.marschall@is-energy.de
97072 Würzburg      |   Tel:   0931/14721
PGP:  D7 FF 20 FE E6 6B 31 74  D1 10 88 E0 3C FE 28 35