Configuring Exim for different Operating Systems
------------------------------------------------

These notes describe the way in which Exim is configured at the C program level
for different operating systems. The normal configuration options that apply
independently of the operating system are specified by creating files in the
Local directory, as described in chapter 4 of the manual.

These notes cover the os.* files in the OS directory, and contain information
for people who want to port the program to some new OS, or to modify the
configuration for an existing port. If you are just wanting to compile Exim on
a system that it already knows about, you do not need to read further unless
there are problems.

The os.c-<ostype> files
-----------------------

There may be an os.c-<ostype> file for each operating system, but for many of
them it is not necessary. No error occurs is there isn't one. There is a
generic file called os.c which contains code that is common to two or more OS
for setting a restarting or a non-restarting signal, for computing the load
average, and for finding all the network interface addresses. A few OS have
their own individual code for one or more of these. When they do, the code is
put into an os.c-<ostype> file, which also defines a macro such as
OS_RESTARTING_SIGNAL (for example) to cut out the common code in the generic
os.c.

The os.h-<ostype> files
-----------------------

For each OS that Exim knows about, there is an os.h-<ostype> file, where
<ostype> is the OS name. The relevant file is included as a C header file for
all Exim compilation by pointing a symbolic link called os.h at it from the
build directory. The settings are as follows:

The select() function
---------------------

There is a difference in the data type for the second argument to the select()
function in some OS. The macro SELECT_ARG2_TYPE can be used to define the type.
If it is not defined in os.h, then it is defaulted to fs_set in exim.h.

The dn_expand() function
------------------------

There is a difference in the data type for the fourth argument to the
dn_expand() function in some OS. The macro DN_EXPAND_ARG4_TYPE can be used to
define the type. If it is not defined in os.h, then it is defaulted to char *
in exim.h.

The h_errno variable
--------------------

If NEED_H_ERRNO is defined, then a definition of the form

extern int h_errno

is included in the compiled code of Exim.

The strerror() function
-----------------------

Most systems provide the ANSI standard strerror() function; older systems may
instead have an errlist[] variable in which to look up error texts. Defining
STRERROR_FROM_ERRLIST causes Exim to build its own strerror() function that
mimics the ANSI function by lookup up the error code in errlist.

Truncating files
----------------

The fcntl() option for truncating the length of a file is called F_FREESP in
most systems; in some, however, it is called O_TRUNC. Some os.h files define
F_FREESP to be O_TRUNC for this reason.

Finding local interfaces
------------------------

The SIOCGIFCONF ioctl for finding local interfaces behaves differently on BSD
systems. It returns a vector of ifreq blocks containing sockaddr structures
that can be longer than their sizeof definition, making the returned ifreq
blocks longer than their sizeof definitions. BSD sockaddrs structures contain
an sa_len field giving the actual size. To cope with difference, there is a
macro called HAVE_SA_LEN. If it is defined, code that works on BSD systems is
used. Otherwise, the objects returned by SIOCGIFCONF are assumed to be of
length sizeof(struct ifreq).

On some operating systems, the SIOCGIFCONF ioctl returns the IP addresses
with the list of interfaces, and there is no need to call SIOCGIFADDR for each
individual address. Mostly, making the second call does no harm, but on Linux
when there are IP aliases, it causes things to go wrong. This also happens on
BSDI and GNU Hurd. Therefore, there is now a macro to cut it out, called
SIOCGIFCONF_GIVES_ADDR.

Note that, if IPv6 support is configured, Exim cannot find the IPv6 addresses
on local interfaces by itself. You need to set the local_interfaces option in
this situation.

Computing load averages
-----------------------

There are several different ways that load averages are computed. One-off code
is put in the os.c-<ostype>, but several OS use similar methods, and these
are coded in the generic os.c, using a number of parameters to make variations
between OS.

Sometimes the load average is not available to unprivileged callers. If
LOAD_AVG_NEEDS_ROOT is set, Exim ensures that it is root before trying to
obtain a load average value.

(1) If HAVE_BSD_GETLOADAVG is defined, Exim uses a simple call to the
getloadavg() function.

(2) If HAVE_KSTAT is defined, Exim uses the kstat package as found in Solaris 2
(but nowhere else as yet). It uses some supplementary definitions:

  LOAD_AVG_KSTAT          the kstat to use
  LOAD_AVG_KSTAT_MODULE   the module to access
  LOAD_AVG_KSTAT_SYMBOL   the symbol containing the value we want
  LOAD_AVG_KSTAT_FIELD    the field identity

(3) If HAVE_DEV_KMEM is defined, Exim reads load average values from the
/dev/kmem device. It uses some supplementary definitions:

  LOAD_AVG_TYPE           the data type
  LOAD_AVG_SYMBOL         the symbol to look up
  KERNEL_PATH             the name of the kernel
  FSCALE                  a scaling factor

Sometimes FSCALE is defined in system headers so need not be defined in the
os.h-<ostype> file.

Glibc systems and IP options
----------------------------

The code for inspecting IP options is the same in all OS except for systems
using glibc (e.g. Linux), which uses a different structure to return data from
getsockopt(). To handle this, there is a macro called

  GLIBC_IP_OPTIONS

which should be set for Linux (in os.h-Linux) and any other operating system
that uses glibc.

Options for statvfs()
---------------------

The following settings apply to the compilation of the Exim monitor as well as
to the main Exim binary.

#undefine HAVE_STATFS

Exim has options for checking the amount of space in the spool partition
before accepting a message, and the monitor has the ability to display a
stripchart of the percentage fullness of a particular disc partition, usually
/var/spool/mail. The standard way of finding out the data is to call the
statvfs() function, but some operating systems use statfs() and some may not
have the ability at all. The Exim code uses STATVFS() for this function and
this gets defined appropriately. HAVE_STATFS is defined before including the
os.h file; undefining it suppresses the code for checking a partition in the
main binary, and for monitoring disc partition in the monitor.

When HAVE_STATFS is defined, the distinction between statvfs() and statfs() is
made by checking HAVE_SYS_STATVFS_H. If it is defined, then sys/statvfs.h is
included. Otherwise, STATVFS() is defined as a macro for statfs(), and some
further includes are done, according to the following definitions:

#define HAVE_SYS_MOUNT_H
#define HAVE_VFS_H

Each of those definitions causes the inclusion of the corresponding system
header file in the Exim monitor compilation. For example, the first one causes

#include <sys/mount.h>

to be obeyed. Different systems may require different combinations of these
headers.

The sys/resource.h header
-------------------------

One OS does not have the sys/resource.h header. If NO_SYS_RESOURCE_H is defined
in an os.h-<ostype> file, then the #include for this header is skipped in
exim.h.

Support for login_cap functions
-------------------------------

Some of the BSD systems support functions for controlling the resources that
user processes can use (e.g. login_getpwclass). If HAVE_SETCLASSRESOURCES is
defined, Exim supports this feature for running pipe deliveries, using the
setclassresources() function.

The crypt_h header
------------------

Some OS require crypt.h to be included to get a prototype for the crypt()
function. This is needed only when compiling with AUTH support. If CRYPT_H is
defined, then this header is included.

mmap() support
--------------

The CDB support includes the option of handling file operations by using
mmap()/munmap(). This gives a reasonable performance increase which will
probably scale over multiple processes (since the files are mapped read-only
shared). The vast majority of modern operating systems will support mmap
(certainly in the simplified way that it is being used here). For example any
BSD 4.x derived or POSIX compliant system will support it, as will pretty much
any system using dynamically shared link libraries.

If the OS is believed to support mmap() then the symbol HAVE_MMAP is defined.
Not all systems that support mmap will have had their config files updated to
reflect this. Currently Linux, Sun, BSD and SGI/mips systems have been updated.

*** End ***