Blob Blame History Raw
This document tries to describe the software layout and design of
nss-pam-ldapd. It should provide some help for contributing code to this


Contributions to nss-pam-ldapd are most welcome. Integrating contributions
will be done on a best-effort basis and can be made easier if the following
are considered:

* for large changes it is a good idea to send an email first
* send your patches in unified diff (diff -u) format, Git patches or Git pull
* try to use the Git version of the software to develop the patch
* clearly state which problem you're trying to solve and how this is
* please follow the existing coding conventions
* please test the patch and include information on tested platforms, etc.
* add a copyright statement with the patch if you feel the contribution is
  significant enough (e.g. more than a few lines)
* ensure that the code you contribute can be integrated in the project
  under the LGPL (when including third-party code, retain copyright
  information license, you have permission to distribute the code, etc.)

Please email if you want to
contribute. All contributions will be acknowledged in the AUTHORS file.


For building Git snapshots the following tools are needed:

* autoconf (2.65 is currently used but 2.61 is minimal)
* automake (1.14 is currently used but older versions may also work)
* OpenLDAP libraries (2.4 is generally used)
* PAM libraries
* optionally a Kerberos library (MIT Kerberos is tested)
* optionally a SASL library (only Cyrus SASL is tested)
* docbook2x for generating the manual pages

To build the Git snapshot run the shell script to build the
configure script. When developing patches please use --enable-warnings with
configure and avoid introducing new warnings.


The versioning scheme of nss-pam-ldapd is a simple major.minor.micro
numbering. The idea is to keep a stable (x.y) branch that only gets bug fixes
and small enhancements while development goes in another branch. Backwards
incompatible changes should be announced clearly.


The basic design splits the functionality in three parts. The NSS part
interfaces with libc and translates the NSS requests into simple generic
requests (e.g. "get user with name test", "get group with gid 101" or "get all
shadow entries").

Another part is the PAM module which handles authentication requests from the
system. The PAM operations are also translated into atomic, stateless

Both these parts translate the queries in a simple protocol used to
communicate with the nslcd daemon. This daemon translates the requests into
LDAP searches. As a result, the NSS and PAM modules don't need to known
anything about LDAP (in fact replacing it with another lookup method should be
very simple) and don't have to link with the LDAP libraries.

  libc NSS ->
                  |->  nslcd  -> OpenLDAP -> LDAP server
  PAM stack ->

An alternative implementation of nslcd in Python is provided as pynslcd. This
implementation is less mature but it easier to add advanced features there. A
collection of utilities is also provided that communicate with nslcd that can
provide additional functions.

design goals
* make it as simple as possible
* simpler configuration and semantics
* simpler, clearer and completer documentation
* split source code into manageable parts
* get rid of unneeded code and complexity
* have a stable, easily maintainable piece of high quality software


The NSS module is implemented in the nss directory. The functions are split
into files according to the database they support. The files support multiple
NSS implementations.

The NSS interface is specific to the C library that is used. The original
implementation was for the GNU C Library but now also includes an
implementation for Solaris' C Library and has some support for FreeBSD.

GNU C Library notes

Function definitions for glibc look like:

  This function opens the connection to the nslcd (with a time-out), builds
  the correct data structures and does a request (write()) to the nslcd
  waiting for an answer (again with a time-out)

The complete list of exported functions can be found in exports.linux and

Currently a number of macros are used to build most of the function bodies for
these functions. Part of this is defined in the common/nslcd-prot.h file and
the NSS-specific stuff is in nss/common.h.

For memory management, the general mechanism that is expected to be used is to
return NSS_STATUS_TRYAGAIN and set errno to ERANGE. This causes glibc to retry
the request with a larger buffer.

Some useful links:

Solaris C Library notes

The Solaris C library uses a different mechanism. For each map a back-end
object is allocated per thread which is used to do queries. The object is
created with a constructor (e.g. _nss_ldap_passwd_constr()) that returns a
back-end that contains a list of function pointer to lookup methods and a

A buffer is passed with every request but a local buffer that is stored in the
back-end can presumably also be created.

Earlier versions of Solaris expected the NSS functions to return the binary
representation of the lookups (e.g. struct passwd) but later versions expect a
string representation of the data to be returned (just like a single line out
of /etc/passwd was read) but only if running from nscd. If args->buf.result is
NULL a string representation is requested (except for ether by address lookup
which is special).

Source and documentation pointers for Solaris NSS:

FreeBSD C Libarary notes

The FreeBSD C library seems to have support for exposing GNU C Library NSS
module functions through a wrapper function. This makes it very easy to
implement NSS support on FreeBSD.

Pointers for more documentation on this is welcome. Some information is
available here:


The PAM module is implemented in the pam directory. Implementation is fairly
straight-forward. The PAM module stores some state between calls to nslcd in a
struct. The calls to nslcd are however stateless. The PAM module may supply
some information that help lookups (most notably DNs of user entries).

Care must be taken with the communication because the nslcd requests are not
authenticated (e.g. changing passwords requests should include all
credentials). The PAM requests may result in state changes on the LDAP server
and this is where they are most notably different from the NSS requests.

Some useful links:


The protocol used for communicating between the NSS library and PAM module on
one end and the nslcd daemon on the other is very simple and almost fully
described in the nslcd.h header file. The common/nslcd-prot.h header file
defines some macros that are used for reading and writing protocol entities
(strings, 32-bit integers, etc).

Every NSS database has a corresponding source file in the nss and the nslcd
directory. The PAM module is built up of a single file in both the pam and
nslcd directories.

If the protocol is changed in an incompatible way the protocol version should
be incremented in nslcd.h. There is currently no versioning scheme available
for this.

A communications module (common/tio.c) was made so we can define simpler
semantics for time-out values and buffer sizes. All components use this module
which means that it includes functionality that is needed for both (e.g. large
write buffers for the server part and large resettable read buffers for the
NSS part). Maybe building two modules from the same source with different
features in them is an option (e.g. the NSS part needs the read buffers and
handling of SIGPIPE and the nslcd part needs the write buffers and possibly
flushing in the background).

The common directory also contains some other generally useful modules that
are used in some components.


At the server end a dispatcher picks up the request and delegates it to one of
the database specific functions.

  This functions fills in the correct parameters from the request. This
  function should write responses to the stream.

Big parts of the request handling functions are generated by macros because
the structure is very similar across the different NSS requests.


This design does open up the system to more potential security issues
(relative to nss_ldap) because there is now a local interface to a daemon with
privileges. With nss_ldad, processes could only potentially exploit bugs in
the library and gain the privileges of the process that was doing the name
lookups. In this case the privileges of the daemon are potentially exposed.

Extra care should be taken with processes that normally require extra
privileges (getting shadow entries, authentication, updating session
information, etc).

Any user on the system can perform nslcd queries so either the nslcd daemon
needs to check the userid of the caller or the request needs to contain the
needed credentials itself.

On the other hand the design also offers some security improvements. It is
much easier to handle security updates of the LDAP, SSL or related libraries
and access to privileged LDAP configuration information can be much better


In the test directory there are a number of tests available. See the file
README in the test directory for more details.