This is a very rough description of a very rough build process. At some point we may fix things up in configure to handle this all automatically. Sorry for the crude nature of this build, it's certainly not for the faint of heart. [The build process has been updated, please read this entire topic for the full details.]

 Overview
 --------

For this port, I chose to keep all of the LDAP internals in ASCII, and only deal with translation of EBCDIC from the host interfaces - reading config files, parsing command line arguments, and writing log files. The V2R9 C Compiler tools have two features to assist in this approach; a CONVLIT option which lets you specify the character set to use when compiling character and string literals, and a __LIBASCII mode for the C library, which replaces a number of library functions with code that accepts arguments in ASCII instead of EBCDIC.

With the combination of these two features, most of the porting effort is taken care of, but there are still some issues.

The configure script will trip up on detecting the Pthreads support. This is because the OS/390 implementation of Pthreads uses function signatures that are different from other published standards. The correct thing to do here is fix the configure script to recognize this IBM variant. I haven't fixed the configure script yet, instead I bypassed the test by setting some environment variables.

Some of the __LIBASCII functions don't work correctly, the getopt() function always segfaulted, and the regcomp() never generated working regular expression matches. So, I used the supplied liblutil/getopt.c and a specially compiled copy of Henry Spencer's regex library instead of the __LIBASCII functions.

In order to handle the ASCII/EBCDIC translation for config file reading and log message writing, I defined several replacement functions (in liblutil/stdio.c) and use macros to make the regular source use these functions. I haven't found a clean location to define these macros, so at present I manually insert them into include/portable.h after configure has run.

There's a lot of other manual tweaks in this process. In time this will all get absorbed into the automated configure script.

 Build Prerequisites
 -------------------

The regex library must be compiled in ASCII mode. This is done by invoking c89 as:

        c89 -Wc,CONVLIT(ISO8859-1)

        #! /bin/sh
        exec /bin/c89 -Wc,CONVLIT\(ISO8859-1\) -Wc,LANGLVL\(EXTENDED\) -D__LIBASCII $*

Defining __LIBASCII activates the definitions for the ASCII-compatibility library.

Note that you must use this script approach, because libtool will strip most of these options and that will give you an unusable build.

The gdbm library should be built in the default EBCDIC mode. The libldbm library has been patched to translate incoming ASCII database names into EBCDIC so the files can be opened properly.

 Running Configure
 -----------------

        CC=acc (the shell script mentioned above, it must be on your $PATH)
        CPP="c89 -E" (all preprocessing is done thru the C compiler driver)
        LD=c89 (all linking is done thru the C compiler driver)
        _C89_CCMODE=1 (allows arbitrary order of commandline arguments)
        CPPFLAGS="-I/usr/local/include -D_ALL_SOURCE"
        LDFLAGS=-L/usr/local/lib
        LIBS=-lregex
        ol_cv_pthread_create=yes
        ol_cv_pthread_create_works=yes
        ac_cv_func_getopt=no

Since the configure script doesn't handle the OS/390 version of Pthreads yet, I used the "ol_cv_pthread_XXX" macros to bypass those parts of the tests. The OS/390 library does provide getopt but the __LIBASCII version crashes so we force the use of the liblutil version instead.

For this build, I only had gdbm and not Berkeley DB. I also did not have SASL, or OpenSSL or Kerberos. I ran configure with --enable-ldbm --disable-bdb.

Note that configure will not detect support for shared libraries. OS/390 does support shared libraries, but libtool doesn't know about it yet. Perhaps in the future this support will be added.

I found no "makedepend" or equivalent tool, so I skipped that step after configure completed.

When configure finishes, you still need to perform some manual edits before you can run make. In include/portable.h, add these definitions:

    #ifdef HAVE_EBCDIC
    #define fputs lutil_fputs
    #define fgets lutil_fgets
    #define printf lutil_printf
    #define fprintf lutil_fprintf
    #define vfprintf lutil_vfprintf
    #define vsprintf lutil_vsprintf
    #define BOTH_STRINGS_H 1
    #define HAVE_PTHREADS_OS390 1
    #endif

In libraries/liblutil/Makefile, add stdio.c to the UNIX_SRCS macro and add stdio.o to the UNIX_OBJS macro. Eventually the configure script will be fixed to do this automatically, but for now you need to do this by hand.

 Conclusion
 ----------

The OS/390 runtime environment allocates stacks in the 24-bit address space by default. This is referred to as memory "below the 16M line." This is the reason why thread creation gets ENOMEM using the default 16M stack size that OpenLDAP uses. Instead of setting LDAP_PVT_THREAD_STACK_SIZE=0 as mentioned above, change the runtime STACK option to allow stack space to be allocated anywhere in the 31-bit address space. This is done by setting the _CEE_RUNOPT environment variable:

        _CEE_RUNOPT="STACK(,,ANY)"

See the OS/390 Language Environment Programming Reference for more details on the STACK option.

It appears that the OpenLDAP default LDAP_PVT_THREAD_STACK_SIZE of 16MB is rather excessive. 2MB is a good starting point, and the stack can still grow incrementally as needed. Also, instead of relying on an environment variable, you can compile in your desired settings using #pragma runopts. This pragma must be specified only once for any given application, and generally goes in the source file where main() is defined. One way or another, it's critical to set the STACK runtime option for slapd and slurpd since they use threads and thus make heavy use of the stack.
hyc@openldap.org

• You still must use the ASCII cc (acc) script for actual compilation.
• You no longer need to bypass the pthreads tests in the configure script, and the script automatically handles getopt, so you don't need to set the ol_cv_pthread_create, ol_cv_pthread_create_works, or ac_cv_func_getopt environment variables before running configure.
• The mkdep script has been extended to be able to produce dependencies from regular C preprocessor output. You need to set these environment variables before running configure:

          MKDEP=c89
          MKDEP_FLAGS=-EC
  

  You'll also need to set this environment variable before running the make depend step:

          _C89_ELINES=1
  

• The necessary #defines for fputs and the other stdio functions are now integrated into portable.h so you don't need to manually edit it.
• You no longer need to edit libraries/liblutil/Makefile manually.
• You still need to override the default thread stack size definition of 16M. Setting it to 2M works best.

        CC=acc (the shell script mentioned above, it must be on your $PATH)
        CPP="c89 -E" (all preprocessing is done thru the C compiler driver)
        LD=c89 (all linking is done thru the C compiler driver)
        _C89_CCMODE=1 (allows arbitrary order of commandline arguments)
        CPPFLAGS="-I/usr/local/include -D_ALL_SOURCE"
        LDFLAGS=-L/usr/local/lib
        LIBS="-lregex -lcrypto"
        MKDEP=c89
        MKDEP_FLAGS=-EC
        _C89_ELINES=1 (needed by the mkdep process)