Compiling Sendmail

Introduction

Read sendmail/SECURITY for important installation information!

Sendmail is a trademark of Sendmail, Inc.

Read below for more details on building sendmail.

Important: do not use optimization (``-O'') if you are running gcc 2.4.x or 2.5.x. There is a bug in the gcc optimizer that causes sendmail compiles to fail miserably.

Jim Wilson of Cygnus believes he has found the problem -- it will probably be fixed in GCC 2.5.6 -- but until this is verified, be very suspicious of gcc -O.

This problem is reported to have been fixed in gcc 2.6.

Important: Read the appropriate paragraphs in the section on Operating System and Compile Quirks.

For detailed instructions, please read the document ../doc/op/op.me:

cd ../doc/op ; make op.ps op.txt

Building Sendmail

By far, the easiest way to compile sendmail is to use the "Build" script:

sh Build

This uses the "uname" command to figure out what architecture you are on and creates a proper Makefile accordingly. It also creates a subdirectory per object format, so that multiarchitecture support is easy. In general this should be all you need. IRIX 6.x users should read the note below in the Operating System and Compile Quirks section.

If you need to look at other include or library directories, use the -I or -L flags on the command line, e.g.,

sh Build -I/usr/sww/include -L/usr/sww/lib

It's also possible to create local site configuration in the file site.config.m4 (or another file settable with the -f flag). This file contains M4 definitions for various compilation values; the most useful are:

confMAPDEF
-D flags to specify database types to be included (see below)
confENVDEF
-D flags to specify other environment information
confINCDIRS
-I flags for finding include files during compilation
confLIBDIRS
-L flags for finding libraries during linking
confLIBS
-l flags for selecting libraries during linking
confLDOPTS
other ld(1) linker options

Others can be found by examining Makefile.m4. Please read ../devtools/README for more information about the site.config.m4 file.

You can recompile from scratch using the -c flag with the Build command. This removes the existing compilation directory for the current platform and builds a new one. The -c flag must also be used if any site.*.m4 file in devtools/Site/ is changed.

Porting to a new Unix-based system should be a matter of creating an appropriate configuration file in the devtools/OS/ directory.

Database Definitions

There are several database formats that can be used for the alias files and for general maps. When used for alias files they interact in an attempt to be backward compatible.

The options are:

NEWDB
The new Berkeley DB package. Some systems (e.g., BSD/OS and Digital UNIX 4.0) have some version of this package pre-installed. If your system does not have Berkeley DB pre-installed, or the version installed is not version 2.0 or greater (e.g., is Berkeley DB 1.85 or 1.86), get the current version from http://www.sleepycat.com/. DO NOT use a version from any of the University of California, Berkeley "Net" or other distributions. If you are still running BSD/386 1.x, you will need to upgrade the included Berkeley DB library to a current version. NEWDB is included automatically if the Build script can find a library named libdb.a or libdb.so. See also OPERATING SYSTEM AND COMPILE QUIRKS about Berkeley DB versions, e.g., DB 4.1.x.
NDBM
The older NDBM implementation -- the very old V7 DBM implementation is no longer supported.
NIS
Network Information Services. To use this you must have NIS support on your system.
NISPLUS
NIS+ (the revised NIS released with Solaris 2). You must have NIS+ support on your system to use this flag.
HESIOD
Support for Hesiod (from the DEC/Athena distribution). You must already have Hesiod support on your system for this to work. You may be able to get this to work with the MIT/Athena version of Hesiod, but that's likely to be a lot of work. BIND 8.X also includes Hesiod support.
LDAPMAP
Lightweight Directory Access Protocol support. You will have to install the UMich or OpenLDAP (http://www.openldap.org/) ldap and lber libraries to use this flag.
MAP_REGEX
Regular Expression support. You will need to use an operating system which comes with the POSIX regex() routines or install a regexp library such as libregex from the Free Software Foundation.
DNSMAP
DNS map support. Requires NAMED_BIND.
PH_MAP
PH map support. You will need the libphclient library from the nph package. MAP_NSD nsd map support (IRIX 6.5 and later).

NOTE WELL for NEWDB support: If you want to get ndbm support, for Berkeley DB versions under 2.0, it is CRITICAL that you remove ndbm.o from libdb.a before you install it and DO NOT install ndbm.h; for Berkeley DB versions 2.0 through 2.3.14, remove dbm.o from libdb.a before you install it. If you don't delete these, there is absolutely no point to including -DNDBM, since it will just get you another (inferior) API to the same format database. These files OVERRIDE calls to ndbm routines -- in particular, if you leave ndbm.h in, you can find yourself using the new db package even if you don't define NEWDB. Berkeley DB versions later than 2.3.14 do not need to be modified. Please also consult the README in the top level directory of the sendmail distribution for other important information.

Further note: DO NOT remove your existing /usr/include/ndbm.h -- you need that one. But do not install an updated ndbm.h in /usr/include, /usr/local/include, or anywhere else.

If NEWDB and NDBM are defined (but not NIS), then sendmail will read NDBM format alias files, but the next time a newaliases is run the format will be converted to NEWDB; that format will be used forever more. This is intended as a transition feature.

If NEWDB, NDBM, and NIS are all defined and the name of the file includes the string "/yp/", sendmail will rebuild BOTH the NEWDB and NDBM format alias files. However, it will only read the NEWDB file; the NDBM format file is used only by the NIS subsystem. This is needed because the NIS maps on an NIS server are built directly from the NDBM files.

If NDBM and NIS are defined (regardless of the definition of NEWDB), and the filename includes the string "/yp/", sendmail adds the special tokens "YP_LAST_MODIFIED" and "YP_MASTER_NAME", both of which are required if the NDBM file is to be used as an NIS map.

All of these flags are normally defined in a confMAPDEF setting in your site.config.m4.

If you define NEWDB or HESIOD you get the User Database (USERDB) automatically. Generally you do want to have NEWDB for it to do anything interesting. See above for getting the Berkeley DB package (i.e., NEWDB). There is no separate "user database" package -- don't bother searching for it on the net.

Hesiod and LDAP require libraries that may not be installed with your system. These are outside of my ability to provide support. See the Operating System and Compile Quirks section for more information.

The regex map can be used to see if an address matches a certain regular expression. For example, all-numerics local parts are common spam addresses, so "^[0-9]+$" would match this. By using such a map in a check_* rule-set, you can block a certain range of addresses that would otherwise be considered valid.

Compile Flags

Wherever possible, I try to make sendmail pull in the correct compilation options needed to compile on various environments based on automatically defined symbols. Some machines don't seem to have useful symbols available, requiring that a compilation flag be defined in the Makefile; see the devtools/OS subdirectory for the supported architectures.

If you are a system to which sendmail has already been ported you should not have to touch the following symbols. But if you are porting, you may have to tweak the following compilation flags in conf.h in order to get it to compile and link properly:

SYSTEM5 Adjust for System V (not necessarily Release 4).
SYS5SIGNALS Use System V signal semantics -- the signal handler is automatically dropped when the signal is caught. If this is not set, use POSIX/BSD semantics, where the signal handler stays in force until an exec or an explicit delete. Implied by SYSTEM5.
SYS5SETPGRP Use System V setpgrp() semantics. Implied by SYSTEM5.
HASNICE Define this to zero if you lack the nice(2) system call.
HASRRESVPORT Define this to zero if you lack the rresvport(3) system call.
HASFCHMOD Define this to one if you have the fchmod(2) system call. This improves security.
HASFCHOWN Define this to one if you have the fchown(2) system call. This is required for the TrustedUser option if sendmail must rebuild an (alias) map.
HASFLOCK Set this if you prefer to use the flock(2) system call rather than using fcntl-based locking. Fcntl locking has some semantic gotchas, but many vendor systems also interface it to lockd(8) to do NFS-style locking. Unfortunately, may vendors implementations of fcntl locking is just plain broken (e.g., locks are never released, causing your sendmail to deadlock; when the kernel runs out of locks your system crashes). For this reason, I recommend always defining this unless you are absolutely certain that your fcntl locking implementation really works.
HASUNAME Set if you have the uname system call. Implied by SYSTEM5.
HASUNSETENV Define this if your system library has the unsetenv subroutine.
HASSETSID Define this if you have the setsid(2) system call. This is implied if your system appears to be POSIX compliant.
HASINITGROUPS Define this if you have the initgroups(3) routine.
HASSETVBUF Define this if you have the setvbuf(3) library call. If you don't, setlinebuf will be used instead. This defaults on if your compiler defines __STDC__.
HASSETREUID Define this if you have setreuid(2) ***AND*** root can use setreuid to change to an arbitrary user. This second condition is not satisfied on AIX 3.x. You may find that your system has setresuid(2), (for example, on HP-UX) in which case you will also have to #define setreuid(r, e) to be the appropriate call. Some systems (such as Solaris) have a compatibility routine that doesn't work properly, but may have "saved user ids" properly implemented so you can ``#define setreuid(r, e) seteuid(e)'' and have it work. The important thing is that you have a call that will set the effective uid independently of the real or saved uid and be able to set the effective uid back again when done. There's a test program in ../test/t_setreuid.c that will try things on your system. Setting this improves the security, since sendmail doesn't have to read .forward and :include: files as root. There are certain attacks that may be unpreventable without this call.
USESETEUID Define this to 1 if you have a seteuid(2) system call that will allow root to set only the effective user id to an arbitrary value ***AND*** you have saved user ids. This is preferable to HASSETREUID if these conditions are fulfilled. These are the semantics of the to-be-released revision of Posix.1. The test program ../test/t_seteuid.c will try this out on your system. If you define both HASSETREUID and USESETEUID the former is ignored.
HASSETEGID Define this if you have setegid(2) and it can be used to set the saved gid. Please run t_dropgid in test/ if you are not sure whether the call works.
HASSETREGID Define this if you have setregid(2) and it can be used to set the saved gid. Please run t_dropgid in test/ if you are not sure whether the call works.
HASSETRESGID Define this if you have setresgid(2) and it can be used to set the saved gid. Please run t_dropgid in test/ if you are not sure whether the call works.
HASLSTAT Define this if you have symbolic links (and thus the lstat(2) system call). This improves security. Unlike most other options, this one is on by default, so you need to #undef it in conf.h if you don't have symbolic links (these days everyone does).
HASSETRLIMIT Define this to 1 if you have the setrlimit(2) syscall. You can define it to 0 to force it off. It is assumed if you are running a BSD-like system.
HASULIMIT Define this if you have the ulimit(2) syscall (System V style systems). HASSETRLIMIT overrides, as it is more general.
HASWAITPID Define this if you have the waitpid(2) syscall.
HASGETDTABLESIZE Define this if you have the getdtablesize(2) syscall.
HASSRANDOMDEV Define this if your system has the srandomdev(3) function call.
HASURANDOMDEV Define this if your system has /dev/urandom(4).
HASSTRERROR Define this if you have the libc strerror function (which should be declared in <errno.h>), and it should be used instead of sys_errlist.
SM_CONF_GETOPT Define this as 0 if you need a reimplementation of getopt(3). On some systems, getopt does very odd things if called to scan the arguments twice. This flag will ask sendmail to compile in a local version of getopt that works properly.
NEEDSTRTOL Define this if your standard C library does not define strtol(3). This will compile in a local version.
NEEDFSYNC Define this if your standard C library does not define fsync(2). This will try to simulate the operation using fcntl(2); if that is not available it does nothing, which isn't great, but at least it compiles and runs.
HASGETUSERSHELL Define this to 1 if you have getusershell(3) in your standard C library. If this is not defined, or is defined to be 0, sendmail will scan the /etc/shells file (no NIS-style support, defaults to /bin/sh and /bin/csh if that file does not exist) to get a list of unrestricted user shells. This is used to determine whether users are allowed to forward their mail to a program or a file.
NEEDPUTENV Define this if your system needs am emulation of the putenv(3) call. Define to 1 to implement it in terms of setenv(3) or to 2 to do it in terms of primitives.
NOFTRUNCATE Define this if you don't have the ftruncate(2) syscall. If you don't have this system call, there is an unavoidable race condition that occurs when creating alias databases.
GIDSET_T The type of entries in a gidset passed as the second argument to getgroups(2). Historically this has been an int, so this is the default, but some systems (such as IRIX) pass it as a gid_t, which is an unsigned short. This will make a difference, so it is important to get this right! However, it is only an issue if you have group sets.
SLEEP_T The type returned by the system sleep() function. Defaults to "unsigned int". Don't worry about this if you don't have compilation problems.
ARBPTR_T The type of an arbitrary pointer -- defaults to "void *". If you have a very old compiler you may need to define this to be "char *".
LA_TYPE The type of load average your kernel supports. These can be one of:
  • LA_ZERO (1) -- it always returns the load average as "zero" (and does so on all architectures).
  • LA_INT (2) to read /dev/kmem for the symbol avenrun and interpret as a long integer.
  • LA_FLOAT (3) same, but interpret the result as a floating point number.
  • LA_SHORT (6) to interpret as a short integer.
  • LA_SUBR (4) if you have the getloadavg(3) routine in your system library.
  • LA_MACH (5) to use MACH-style load averages (calls processor_set_info()),
  • LA_PROCSTR (7) to read /proc/loadavg and interpret it as a string representing a floating-point number (Linux-style).
  • LA_READKSYM (8) is an implementation suitable for some versions of SVr4 that uses the MIOC_READKSYM ioctl call to read /dev/kmem.
  • LA_DGUX (9) is a special implementation for DG/UX that uses the dg_sys_info system call.
  • LA_HPUX (10) is an HP-UX specific version that uses the pstat_getdynamic system call.
  • LA_IRIX6 (11) is an IRIX 6.x specific version that adapts to 32 or 64 bit kernels; it is otherwise very similar to LA_INT.
  • LA_KSTAT (12) uses the (Solaris-specific) kstat(3k) implementation.
  • LA_DEVSHORT (13) reads a short from a system file (default: /dev/table/avenrun) and scales it in the same manner as LA_SHORT.
LA_INT, LA_SHORT, LA_FLOAT, and LA_READKSYM have several other parameters that they try to divine: the name of your kernel, the name of the variable in the kernel to examine, the number of bits of precision in a fixed point load average, and so forth. LA_DEVSHORT uses _PATH_AVENRUN to find the device to be read to find the load average. In desperation, use LA_ZERO. The actual code is in conf.c -- it can be tweaked if you are brave.
FSHIFT For LA_INT, LA_SHORT, and LA_READKSYM, this is the number of bits of load average after the binary point -- i.e., the number of bits to shift right in order to scale the integer to get the true integer load average. Defaults to 8.
_PATH_UNIX The path to your kernel. Needed only for LA_INT, LA_SHORT, and LA_FLOAT. Defaults to "/unix" on System V, "/vmunix" everywhere else.
LA_AVENRUN For LA_INT, LA_SHORT, and LA_FLOAT, the name of the kernel variable that holds the load average. Defaults to "avenrun" on System V, "_avenrun" everywhere else.
SFS_TYPE Encodes how your kernel can locate the amount of free space on a disk partition. This can be set to
  • SFS_NONE (0) if you have no way of getting this information,
  • SFS_USTAT (1) if you have the ustat(2) system call,
  • SFS_4ARGS (2) if you have a four-argument statfs(2) system call (and the include file is <sys/statfs.h>),
  • SFS_VFS (3),
  • SFS_MOUNT (4),
  • SFS_STATFS (5) if you have the two-argument statfs(2) system call with includes in <sys/vfs.h>, <sys/mount.h>, or <sys/statfs.h>; respectively, or
  • SFS_STATVFS (6) if you have the two-argument statvfs(2) call.
The default if nothing is defined is SFS_NONE.
SFS_BAVAIL with SFS_4ARGS you can also set SFS_BAVAIL to the field name in the statfs structure that holds the useful information; this defaults to f_bavail.
SPT_TYPE Encodes how your system can display what a process is doing on a ps(1) command (SPT stands for Set Process Title). Can be set to:
  • SPT_NONE (0) -- Don't try to set the process title at all.
  • SPT_REUSEARGV (1) -- Pad out your argv with the information; this is the default if none specified.
  • SPT_BUILTIN (2) -- The system library has setproctitle.
  • SPT_PSTAT (3) -- Use the PSTAT_SETCMD option to pstat(2) to set the process title; this is used by HP-UX.
  • SPT_PSSTRINGS (4) -- Use the magic PS_STRINGS pointer (4.4BSD).
  • SPT_SYSMIPS (5) -- Use sysmips() supported by NEWS-OS 6.
  • SPT_SCO (6) -- Write kernel u. area.
  • SPT_CHANGEARGV (7) -- Write pointers to our own strings into the existing argv vector.
SPT_PADCHAR Character used to pad the process title; if undefined, the space character (0x20) is used. This is ignored if SPT_TYPE != SPT_REUSEARGV.
ERRLIST_PREDEFINED If set, assumes that some header file defines sys_errlist. This may be needed if you get type conflicts on this variable -- otherwise don't worry about it.
WAITUNION The wait(2) routine takes a "union wait" argument instead of an integer argument. This is for compatibility with old versions of BSD.
SCANF You can set this to extend the F command to accept a scanf string -- this gives you a primitive parser for class definitions -- BUT it can make you vulnerable to core dumps if the target file is poorly formed.
SYSLOG_BUFSIZE You can define this to be the size of the buffer that syslog accepts. If it is not defined, it assumes a 1024-byte buffer. If the buffer is very small (under 256 bytes) the log message format changes -- each e-mail message will log many more messages, since it will log each piece of information as a separate line in syslog.
BROKEN_RES_SEARCH On Ultrix (and maybe other systems?) if you use the res_search routine with an unknown host name, it returns -1 but sets h_errno to 0 instead of HOST_NOT_FOUND. If you set this, sendmail considers 0 to be the same as HOST_NOT_FOUND.
NAMELISTMASK If defined, values returned by nlist(3) are masked against this value before use -- a common value is 0x7fffffff to strip off the top bit.
BSD4_4_SOCKADDR If defined, socket addresses have an sa_len field that defines the length of this address.
SAFENFSPATHCONF Set this to 1 if and only if you have verified that a pathconf(2) call with _PC_CHOWN_RESTRICTED argument on an NFS filesystem where the underlying system allows users to give away files to other users returns <= 0. Be sure you try both on NFS V2 and V3. Some systems assume that their local policy apply to NFS servers -- this is a bad assumption! The test/t_pathconf.c program will try this for you -- you have to run it in a directory that is mounted from a server that allows file giveaway.
SIOCGIFCONF_IS_BROKEN Set this if your system has an SIOCGIFCONF ioctl defined, but it doesn't behave the same way as "most" systems (BSD, Solaris, SunOS, HP-UX, etc.)
SIOCGIFNUM_IS_BROKEN Set this if your system has an SIOCGIFNUM ioctl defined, but it doesn't behave the same way as "most" systems (Solaris, HP-UX).
FAST_PID_RECYCLE Set this if your system can reuse the same PID in the same second.
SO_REUSEADDR_IS_BROKEN Set this if your system has a setsockopt() SO_REUSEADDR flag but doesn't pay attention to it when trying to bind a socket to a recently closed port.
NEEDSGETIPNODE Set this if your system supports IPv6 but doesn't include the getipnodeby{name,addr}() functions. Set automatically for Linux's glibc.
PIPELINING Support SMTP PIPELINING (set by default).
USING_NETSCAPE_LDAP Deprecated in favor of SM_CONF_LDAP_MEMFREE. See libsm/README.
NEEDLINK Set this if your system doesn't have a link() call. It will create a copy of the file instead of a hardlink.
USE_ENVIRON Set this to 1 to access process environment variables from the external variable environ instead of the third parameter of main().
USE_DOUBLE_FORK By default this is on (1). Set it to 0 to suppress the extra fork() used to avoid intermediate zombies.

Compile-Time Features

There are a bunch of features that you can decide to compile in, such as selecting various database packages and special protocol support. Several are assumed based on other compilation flags -- if you want to "un-assume" something, you probably need to edit conf.h. Compilation flags that add support for special features include:

NDBM Include support for "new" DBM library for aliases and maps. Normally defined in the Makefile.
NEWDB Include support for Berkeley DB package (hash & btree) for aliases and maps. Normally defined in the Makefile. If the version of NEWDB you have is the old one that does not include the "fd" call (this call was added in version 1.5 of the Berkeley DB code), you must upgrade to the current version of Berkeley DB.
NIS Define this to get NIS (YP) support for aliases and maps. Normally defined in the Makefile.
NISPLUS Define this to get NIS+ support for aliases and maps. Normally defined in the Makefile.
HESIOD Define this to get Hesiod support for aliases and maps. Normally defined in the Makefile.
NETINFO Define this to get NeXT NetInfo support for aliases and maps. Normally defined in the Makefile.
LDAPMAP Define this to get LDAP support for maps.
PH_MAP Define this to get PH support for maps.
MAP_NSD Define this to get nsd support for maps.
USERDB Define this to 1 to include support for the User Information Database. Implied by NEWDB or HESIOD. You can use -DUSERDB=0 to explicitly turn it off.
IDENTPROTO Define this as 1 to get IDENT (RFC 1413) protocol support. This is assumed unless you are running on Ultrix or HP-UX, both of which have a problem in the UDP implementation. You can define it to be 0 to explicitly turn off IDENT protocol support. If defined off, the code is actually still compiled in, but it defaults off; you can turn it on by setting the IDENT timeout in the configuration file.
IP_SRCROUTE Define this to 1 to get IP source routing information displayed in the Received: header. This is assumed on most systems, but some (e.g., Ultrix) apparently have a broken version of getsockopt that doesn't properly support the IP_OPTIONS call. You probably want this if your OS can cope with it. Symptoms of failure will be that it won't compile properly (that is, no support for fetching IP_OPTIONs), or it compiles but source-routed TCP connections either refuse to open or open and hang for no apparent reason. Ultrix and AIX3 are known to fail this way.
LOG Set this to get syslog(3) support. Defined by default in conf.h. You want this if at all possible.
NETINET Set this to get TCP/IP support. Defined by default in conf.h. You probably want this.
NETINET6 Set this to get IPv6 support. Other configuration may be needed in conf.h for your particular operating system. Also, DaemonPortOptions must be set appropriately for sendmail to accept IPv6 connections.
NETISO Define this to get ISO networking support.
NETUNIX Define this to get Unix domain networking support. Defined by default. A few bizarre systems (SCO, ISC, Altos) don't support this networking domain.
NETNS Define this to get NS networking support.
NETX25 Define this to get X.25 networking support.
NAMED_BIND If non-zero, include DNS (name daemon) support, including MX support. The specs say you must use this if you run SMTP. You don't have to be running a name server daemon on your machine to need this -- any use of the DNS resolver, including remote access to another machine, requires this option. Defined by default in conf.h. Define it to zero ONLY on machines that do not use DNS in any way.
MATCHGECOS Permit fuzzy matching of user names against the full name (GECOS) field in the /etc/passwd file. This should probably be on, since you can disable it from the config file if you want to. Defined by default in conf.h.
MIME8TO7 If non-zero, include 8 to 7 bit MIME conversions. This also controls advertisement of 8BITMIME in the ESMTP startup dialogue.
MIME7TO8 If non-zero, include 7 to 8 bit MIME conversions.
HES_GETMAILHOST Define this to 1 if you are using Hesiod with the hes_getmailhost() routine. This is included with the MIT Hesiod distribution, but not with the DEC Hesiod distribution.
XDEBUG Do additional internal checking. These don't cost too much; you might as well leave this on.
TCPWRAPPERS Turns on support for the TCP wrappers library (-lwrap). See below for further information.
SECUREWARE Enable calls to the SecureWare luid enabling/changing routines. SecureWare is a C2 security package added to several UNIX's (notably ConvexOS) to get a C2 Secure system. This option causes mail delivery to be done with the luid of the recipient.
SHARE_V1 Support for the fair share scheduler, version 1. Setting to 1 causes final delivery to be done using the recipients resource limitations. So far as I know, this is only supported on ConvexOS.
SASL Enables SMTP AUTH (RFC 2554). This requires the Cyrus SASL library (ftp://ftp.andrew.cmu.edu/pub/cyrus-mail/). Please install at least version 1.5.13. See below for further information: SASL Compilation and configuration. If your SASL library is older than 1.5.10, you have to set this to its version number using a simple conversion: a.b.c -> c + b*100 + a*10000, e.g. for 1.5.9 define SASL=10509. Note: Using an older version than 1.5.5 of Cyrus SASL is not supported. Starting with version 1.5.10, setting SASL=1 is sufficient. Any value other than 1 (or 0) will be compared with the actual version found and if there is a mismatch, compilation will fail.
EGD Define this if your system has EGD installed, see http://egd.sourceforge.net/. It should be used to seed the PRNG for STARTTLS if HASURANDOMDEV is not defined.
STARTTLS Enables SMTP STARTTLS (RFC 2487). This requires OpenSSL; use OpenSSL 0.9.5a or later (if compatible with this version), do not use 0.9.3. See STARTTLS Compilation and Configuration for further information.
TLS_NO_RSA Turn off support for RSA algorithms in STARTTLS.
MILTER Turn on support for external filters using the Milter API. See libmilter/README for more information.
REQUIRES_DIR_FSYNC Turn on support for file systems that require to call fsync() for a directory if the meta-data in it has been changed. This should be turned on at least for older versions of ReiserFS; it is enabled by default for Linux. According to some information this flag is not needed anymore for kernel 2.4.16 and newer. We would appreciate feedback about the semantics of the various file systems available for Linux. An alternative to this compile time flag is to mount the queue directory without the -async option, or using chattr +S on Linux.
DBMMODE The default file permissions to use when creating new database files for maps and aliases. Defaults to 0640.

Generic notice: If you enable a compile time option that needs libraries or include files that don't come with sendmail or are installed in a location that your C compiler doesn't use by default you should set confINCDIRS and confLIBDIRS as explained in the first section: Building Sendmail.

DNS/Resolver Issues

Many systems have old versions of the resolver library. At a minimum, you should be running BIND 4.8.3; older versions may compile, but they have known bugs that should give you pause.

Common problems in old versions include "undefined" errors for dn_skipname.

Some people have had a problem with BIND 4.9; it uses some routines that it expects to be externally defined such as strerror(). It may help to link with "-l44bsd" to solve this problem. This has apparently been fixed in later versions of BIND, starting around 4.9.3. In other words, if you use 4.9.0 through 4.9.2, you need -l44bsd; for earlier or later versions, you do not.

Please be sure to link with the same version of the resolver as the header files you used -- some people have used the 4.9 headers and linked with BIND 4.8 or vice versa, and it doesn't work. Unfortunately, it doesn't fail in an obvious way -- things just subtly don't work.

Wildcard MX records are a bad idea! The only situation in which they work reliably is if you have two versions of DNS, one in the real world which has a wildcard pointing to your firewall, and a completely different version of the database internally that does not include wildcard MX records that match your domain. Anything else will give you headaches!

When attempting to canonify a hostname, some broken name servers will return SERVFAIL (a temporary failure) on T_AAAA (IPv6) lookups. If you want to excuse this behavior, include WorkAroundBrokenAAAA in ResolverOptions. However, instead, we recommend catching the problem and reporting it to the name server administrator so we can rid the world of broken name servers.

STARTTLS Compilation and Configuration

Please read the documentation accompanying the OpenSSL library. You have to compile and install the OpenSSL libraries before you can compile sendmail. See devtools/README how to set the correct compile time parameters; you should at least set the following variables:

APPENDDEF(`conf_sendmail_ENVDEF', `-DSTARTTLS') APPENDDEF(`conf_sendmail_LIBS', `-lssl -lcrypto')

If you have installed the OpenSSL libraries and include files in a location that your C compiler doesn't use by default you should set

confINCDIRS and confLIBDIRS as explained in the first section: Building Sendmail.

Configuration information can be found in doc/op/op.me (required certificates) and cf/README (how to tell sendmail about certificates).

To perform an initial test, connect to your sendmail daemon (telnet localhost 25) and issue a EHLO localhost and see whether

250-STARTTLS

is in the response. If it isn't, run the daemon with

-O LogLevel=14

and try again. Then take a look at the logfile and see whether there are any problems listed about permissions (unsafe files) or the validity of X.509 certificates.

From: Garrett Wollman <wollman@lcs.mit.edu>

If your certificate authority is hierarchical, and you only include the top-level CA certificate in the CACertFile file, some mail clients may be unable to infer the proper certificate chain when selecting a client certificate. Including the bottom-level CA certificate(s) in the CACertFile file will allow these clients to work properly. This is not necessary if you are not using client certificates for authentication, or if all your clients are running Sendmail or other programs using the OpenSSL library (which get it right automatically). In addition, some mail clients are totally incapable of using certificate authentication -- even some of those which already support SSL/TLS for confidentiality.

Further information can be found via on our tips pages.

SASL Compilation and Configuration

Please read the documentation accompanying the Cyrus SASL library (INSTALL and README). If you use Berkeley DB for Cyrus SASL then you must compile sendmail with the same version of Berkeley DB. See devtools/README for how to set the correct compile time parameters; you should at least set the following variables:

APPENDDEF(`conf_sendmail_ENVDEF', `-DSASL') APPENDDEF(`conf_sendmail_LIBS', `-lsasl')

If you have installed the Cyrus SASL library and include files in a location that your C compiler doesn't use by default you should set confINCDIRS and confLIBDIRS as explained in the first section: Building Sendmail.

You have to select and install authentication mechanisms and tell sendmail where to find the sasl library and the include files (see devtools/README for the parameters to set). Set up the required users and passwords as explained in the SASL documentation. See also cf/README for authentication related options (especially DefaultAuthInfo if you want authentication between MTAs).

To perform an initial test, connect to your sendmail daemon (telnet localhost 25) and issue a EHLO localhost and see whether

250-AUTH ....

is in the response. If it isn't, run the daemon with

-O LogLevel=14

and try again. Then take a look at the logfile and see whether there are any security related problems listed (unsafe files).

Further information can be found via on our tips pages.

Operating System and Compile Quirks

GCC problems

When compiling with "gcc -O -Wall" specify "-DSM_OMIT_BOGUS_WARNINGS" too (see include/sm/cdefs.h for more info).

***************************************************************** ** IMPORTANT: DO NOT USE OPTIMIZATION (``-O'') IF YOU ARE ** ** RUNNING GCC 2.4.x or 2.5.x. THERE IS A BUG IN THE GCC ** ** OPTIMIZER THAT CAUSES SENDMAIL COMPILES TO FAIL MISERABLY. ** *****************************************************************

Jim Wilson of Cygnus believes he has found the problem -- it will probably be fixed in GCC 2.5.6 -- but until this is verified, be very suspicious of gcc -O. This problem is reported to have been fixed in gcc 2.6.

A bug in gcc 2.5.5 caused problems compiling sendmail 8.6.5 with optimization on a Sparc. If you are using gcc 2.5.5, you should upgrade to the latest version of gcc.

Apparently GCC 2.7.0 on the Pentium processor has optimization problems. I recommend against using -O on that architecture. This has been seen on FreeBSD 2.0.5 RELEASE.

Solaris 2.X users should use version 2.7.2.3 over 2.7.2.

We have been told there are problems with gcc 2.8.0. If you are using this version, you should upgrade to 2.8.1 or later.

BerkeleyDB

Berkeley DB 4.1.x with x <= 24 does not work with sendmail. You need at least 4.1.25.

GDBM

GDBM does not work with sendmail because the additional security checks and file locking cause problems. Unfortunately, gdbm does not provide a compile flag in its version of ndbm.h so the code can adapt. Until the GDBM authors can fix these problems, GDBM will not be supported. Please use Berkeley DB instead.

Configuration file location

Up to 8.6, sendmail tried to find the sendmail.cf file in the same place as the vendors had put it, even when this was obviously stupid. As of 8.7, sendmail ALWAYS looks for /etc/sendmail.cf. Beginning with 8.10, sendmail uses /etc/mail/sendmail.cf. You can get sendmail to use the stupid vendor .cf location by adding -DUSE_VENDOR_CF_PATH during compilation, but this may break support programs and scripts that need to find sendmail.cf. You are STRONGLY urged to use symbolic links if you want to use the vendor location rather than changing the location in the sendmail binary.

NETINFO systems use NETINFO to determine the location of sendmail.cf. The full path to sendmail.cf is stored as the value of the "sendmail.cf" property in the "/locations/sendmail" subdirectory of NETINFO. Set the value of this property to "/etc/mail/sendmail.cf" (without the quotes) to use this new default location for Sendmail 8.10.0 and higher.

BIND 8.1 Compilation Warnings and Link Errors

Since the release of BIND version 8.1, many people building sendmail have experienced problems compiling and linking with the new BIND include files and libraries under /usr/local/. The following message explains:

From: Bob Halley <halley@vix.com>
Cc: bind-workers@vix.com
Subject: Re: bind 8.1 linking problems.
Date: Thu, 19 Jun 1997 16:21:41 -0700

I agree that it's not an Apache problem.

Until BIND 8 can handle shared libraries and do edits on system libraries, the plan has been to put BIND 8 header files and libraries some place where they would be out of the way. Applications that hadn't had their Makefiles changed would continue to use the system's native .h files and resolver.

Unfortunately, on some systems /usr/local/include is not as out of the way as we thought it was. We have the bad situation of applications using BIND 8 header files without linking against libbind.a, which prevents them from linking.

The next BIND 8.1.1 test kit will install .h files into /usr/local/bind/include, and libbind.a into /usr/local/bind/lib.

Also, the missing __ansi_realloc() on SunOS was a bug. __ansi_realloc() is something we wrote because a number our our library routines need ANSI C semantics for realloc(), which the native SunOS realloc() doesn't provide. The bug was that this routine was in a separate library, libport.a, instead of in libbind.a. The next 8.1.1 test kit will also fix this problem, by putting the contents of libport.a into libbind.a, and eliminating libport.a altogether.

/Bob

ControlSocket Permissions

Paraphrased from BIND 8.2.1's README:

Solaris and other pre-4.4BSD kernels do not respect ownership or protections on UNIX-domain sockets. The short term fix for this is to override the default path and put such control sockets into root- owned directories which do not permit non-root to r/w/x through them. The long term fix is for all kernels to upgrade to 4.4BSD semantics.

HP MPE/iX

The MPE-specific code within sendmail emulates a set-user-id root environment for the sendmail binary. But there is no root uid 0 on MPE, nor is there any support for set-user-id programs. Even when sendmail thinks it is running as uid 0, it will still have the file access rights of the underlying non-zero uid, but because sendmail is an MPE priv-mode program it will still be able to call setuid() to successfully switch to a new uid.

MPE setgid() semantics don't quite work the way sendmail expects, so special emulation is done here also.

This uid/gid emulation is enabled via the setuid/setgid file mode bits which are not currently used by MPE. Code in libsm/mpeix.c examines these bits and enables emulation if they have been set, i.e., chmod u+s,g+s /SENDMAIL/CURRENT/SENDMAIL.

SunOS 4.x (Solaris 1.x)

You may have to use -lresolv on SunOS. However, beware that this links in a new version of gethostbyname that does not understand NIS, so you must have all of your hosts in DNS.

Some people have reported problems with the SunOS version of -lresolv and/or in.named, and suggest that you get a newer version. The symptoms are delays when you connect to the SMTP server on a SunOS machine or having your domain added to addresses inappropriately. There is a version of BIND version 4.9 on gatekeeper.DEC.COM in pub/BSD/bind/4.9.

There is substantial disagreement about whether you can make this work with resolv+, which allows you to specify a search-path of services. Some people report that it works fine, others claim it doesn't work at all (including causing sendmail to drop core when it tries to do multiple resolv+ lookups for a single job). I haven't tried resolv+, as we use DNS exclusively.

Should you want to try resolv+, it is on ftp.uu.net in /networking/ip/dns.

Apparently getservbyname() can fail under moderate to high load under some circumstances. This will exhibit itself as the message ``554 makeconnection: service "smtp" unknown''. The problem has been traced to one or more blank lines in /etc/services on the NIS server machine. Delete these and it should work. This info is thanks to Brian Bartholomew <bb@math.ufl.edu> of I-Kinetics, Inc.

NOTE: The SunOS 4.X linker uses library paths specified during compilation using -L for run-time shared library searches. Therefore, it is vital that relative and unsafe directory paths not be used when compiling sendmail.

SunOS 4.0.2 (Sun 386i)

Date: Fri, 25 Aug 1995 11:13:58 +0200 (MET DST)
From: teus@oce.nl

Sendmail 8.7.Beta.12 compiles and runs nearly out of the box with the following changes:

  • Don't use /usr/5bin in your PATH, but make /usr/5bin/uname available as "uname" command.
  • Use the defines "-DBSD4_3 -DNAMED_BIND=0" in the devtools/OS/SunOS.4.0, which is selected via the "uname" command.

I recommend to make available the db-library on the system first (and change the Makefile to use this library).

Note that the sendmail.cf and aliases files are found in /etc.

SunOS 4.1.3, 4.1.3_U1

Sendmail causes crashes on SunOS 4.1.3 and 4.1.3_U1. According to Sun bug number 1077939:

If an application does a getsockopt() on a SOCK_STREAM (TCP) socket after the other side of the connection has sent a TCP RESET for the stream, the kernel gets a Bus Trap in the tcp_ctloutput() or ip_ctloutput() routine.

For 4.1.3, this is fixed in patch 100584-08, available on the Sunsolve 2.7.1 or later CDs. For 4.1.3_U1, this was fixed in patch 101790-01 (SunOS 4.1.3_U1: TCP socket and reset problems), later obsoleted by patch 102010-05.

Sun patch 100584-08 is not currently publicly available on their ftp site but a user has reported it can be found at other sites using a web search engine.

Solaris 2.x (SunOS 5.x)

To compile for Solaris, the Makefile built by Build must include a SOLARIS definition which reflects the Solaris version (i.e. -DSOLARIS=20400 for 2.4 or -DSOLARIS=20501 for 2.5.1). If you are using gcc, make sure -I/usr/include is not used (or it might complain about TopFrame). If you are using Sun's cc, make sure /opt/SUNWspro/bin/cc is used instead of /usr/ucb/cc (or it might complain about tm_zone).

The Solaris 2.x (x <= 3) "syslog" function is apparently limited to something about 90 characters because of a kernel limitation. If you have source code, you can probably up this number. You can get patches that fix this problem: the patch ids are:

Solaris 2.1 100834 Solaris 2.2 100999 Solaris 2.3 101318

Be sure you have the appropriate patch installed or you won't see system logging.

Solaris 2.4 (SunOS 5.4)

If you include /usr/lib at the end of your LD_LIBRARY_PATH you run the risk of getting the wrong libraries under some circumstances. This is because of a new feature in Solaris 2.4, described by Rod.Evans@Eng.Sun.COM:

Prior to SunOS 5.4, any LD_LIBRARY_PATH setting was ignored by the runtime linker if the application was setxid (secure), thus your applications search path would be:

/usr/local/lib LD_LIBRARY_PATH component - IGNORED /usr/lib LD_LIBRARY_PATH component - IGNORED /usr/local/lib RPATH - honored /usr/lib RPATH - honored

the effect is that path 3 would be the first used, and this would satisfy your resolv.so lookup.

In SunOS 5.4 we made the LD_LIBRARY_PATH a little more flexible. People who developed setxid applications wanted to be able to alter the library search path to some degree to allow for their own testing and debugging mechanisms. It was decided that the only secure way to do this was to allow a `trusted' path to be used in LD_LIBRARY_PATH. The only trusted directory we presently define is /usr/lib. Thus a set-user-ID root developer could play with some alternative shared object implementations and place them in /usr/lib (being root we assume they'ed have access to write in this directory). This change was made as part of 1155380 - after a *huge* amount of discussion regarding the security aspect of things.

So, in SunOS 5.4 your applications search path would be:

/usr/local/lib from LD_LIBRARY_PATH - IGNORED (untrustworthy) /usr/lib from LD_LIBRARY_PATH - honored (trustworthy) /usr/local/lib from RPATH - honored /usr/lib from RPATH - honored

here, path 2 would be the first used.

Solaris 2.5.1 (SunOS 5.5.1) and 2.6 (SunOS 5.6)

Apparently Solaris 2.5.1 patch 103663-01 installs a new /usr/include/resolv.h file that defines the __P macro without checking to see if it is already defined. This new resolv.h is also included in the Solaris 2.6 distribution. This causes compile warnings such as:

In file included from daemon.c:51: /usr/include/resolv.h:208: warning: `__P' redefined cdefs.h:58: warning: this is the location of the previous definition

These warnings can be safely ignored or you can create a resolv.h file in the obj.SunOS.5.5.1.* or obj.SunOS.5.6.* directory that reads

#undef __P #include "/usr/include/resolv.h"

This problem was fixed in Solaris 7 (Sun bug ID 4081053).

Solaris 7 (SunOS 5.7)

Solaris 7 includes LDAP libraries but the implementation was lacking a few things. The following settings can be placed in devtools/Site/site.SunOS.5.7.m4 if you plan on using those libraries.

APPENDDEF(`confMAPDEF', `-DLDAPMAP') APPENDDEF(`confENVDEF', `-DLDAP_VERSION_MAX=3') APPENDDEF(`confLIBS', `-lldap')

Also, Sun's patch 107555 is needed to prevent a crash in the call to ldap_set_option for LDAP_OPT_REFERRALS in ldapmap_setopts if LDAP support is compiled in sendmail.

Solaris 8 and later (SunOS 5.8 and later)

Solaris 8 and later can optionally install LDAP support. If you have installed the Entire Distribution meta-cluster, you can use the following in devtools/Site/site.SunOS.5.8.m4 (or other appropriately versioned file) to enable LDAP:

APPENDDEF(`confMAPDEF', `-DLDAPMAP') APPENDDEF(`confLIBS', `-lldap')

Solaris 9 and later (SunOS 5.9 and later)

Solaris 9 and later have a revised LDAP library, libldap.so.5, which is derived from a Netscape implementation, thus requiring that SM_CONF_LDAP_MEMFREE be defined in conjunction with LDAPMAP:

APPENDDEF(`confMAPDEF', `-DLDAPMAP') APPENDDEF(`confENVDEF', `-DSM_CONF_LDAP_MEMFREE') APPENDDEF(`confLIBS', `-lldap')

Solaris

If you are using dns for hostname resolution on Solaris, make sure that the 'dns' entry is last on the hosts line in '/etc/nsswitch.conf'. For example, use:

hosts: nisplus files dns

Do not use:

hosts: nisplus dns [NOTFOUND=return] files

Note that 'nisplus' above is an illustration. The same comment applies no matter what naming services you are using. If you have anything other than dns last, even after "[NOTFOUND=return]", sendmail may not be able to determine whether an error was temporary or permanent. The error returned by the solaris gethostbyname() is the error for the last lookup used, and other naming services do not have the same concept of temporary failure.

Ultrix

By default, the IDENT protocol is turned off on Ultrix. If you are running Ultrix 4.4 or later, or if you have included patch CXO-8919 for Ultrix 4.2 or 4.3 to fix the TCP problem, you can turn IDENT on in the configuration file by setting the "ident" timeout. The Ultrix 4.5 Y2K patch (ULTV45-022-1) has changed the resolver included in libc.a. Unfortunately, the __RES symbol hasn't changed and therefore, sendmail can no longer automatically detect the newer version. If you get a compiler error:

/lib/libc.a(gethostent.o): local_hostname_length: multiply defined

Then rebuild with this in devtools/Site/site.ULTRIX.m4:

APPENDDEF(`conf_sendmail_ENVDEF', `-DNEEDLOCAL_HOSTNAME_LENGTH=0')

Digital Unix (formerly DEC OSF/1)

If you are compiling on OSF/1 (DEC Alpha), you must use -L/usr/shlib (otherwise it core dumps on startup). You may also need -mld to get the nlist() function, although some versions apparently don't need this.

Also, the enclosed makefile removed /usr/sbin/smtpd; if you need it, just create the link to the sendmail binary.

On DEC OSF/1 3.2 or earlier, the MatchGECOS option doesn't work properly due to a bug in the getpw* routines. If you want to use this, use -DDEC_OSF_BROKEN_GETPWENT=1. The problem is fixed in 3.2C.

Digital's mail delivery agent, /bin/mail (aka /bin/binmail), will only preserve the envelope sender in the "From " header if DefaultUserID is set to daemon. Setting this to mailnull will cause all mail to have the header "From mailnull ...". To use a different DefaultUserID, you will need to use a different mail delivery agent (such as mail.local found in the sendmail distribution).

On Digital UNIX 4.0 and later, Berkeley DB 1.85 is included with the operating system and already has the ndbm.o module removed. However, Digital has modified the original Berkeley DB db.h include file. This results in the following warning while compiling map.c and udb.c:

cc: Warning: /usr/include/db.h, line 74: The redefinition of the macro "__signed" conflicts with a current definition because the replacement lists differ. The redefinition is now in effect. #define __signed signed ------------------------^

This warning can be ignored.

Digital UNIX's linker checks /usr/ccs/lib/ before /usr/lib/. If you have installed a new version of BIND in /usr/include and /usr/lib, you will experience difficulties as Digital ships libresolv.a in /usr/ccs/lib/ as well. Be sure to replace both copies of libresolv.a.

IRIX

The header files on SGI IRIX are completely prototyped, and as a result you can sometimes get some warning messages during compilation. These can be ignored. There are two errors in deliver only if you are using gcc, both of the form ``warning: passing arg N of `execve' from incompatible pointer type''. Also, if you compile with -DNIS, you will get a complaint about a declaration of struct dom_binding in a prototype when compiling map.c; this is not important because the function being prototyped is not used in that file.

In order to compile sendmail you will have had to install the developers' option in order to get the necessary include files.

If you compile with -lmalloc (the fast memory allocator), you may get warning messages such as the following:

ld32: WARNING 85: definition of _calloc in /usr/lib32/libmalloc.so preempts that definition in /usr/lib32/mips3/libc.so. ld32: WARNING 85: definition of _malloc in /usr/lib32/libmalloc.so preempts that definition in /usr/lib32/mips3/libc.so. ld32: WARNING 85: definition of _realloc in /usr/lib32/libmalloc.so preempts that definition in /usr/lib32/mips3/libc.so. ld32: WARNING 85: definition of _free in /usr/lib32/libmalloc.so preempts that definition in /usr/lib32/mips3/libc.so. ld32: WARNING 85: definition of _cfree in /usr/lib32/libmalloc.so preempts that definition in /usr/lib32/mips3/libc.so.

These are unavoidable and innocuous -- just ignore them.

According to Dave Sill <de5@ornl.gov>, there is a version of the Berkeley DB library patched to run on Irix 6.2 available from http://reality.sgi.com/ariel/freeware/#db .

IRIX 6.x

If you are using XFS filesystem, avoid using the -32 ABI switch to the cc compiler if possible.

Broken inet_aton and inet_ntoa on IRIX using gcc: There's a problem with gcc on IRIX, i.e., gcc can't pass structs less than 16 bits long unless they are 8 bits; IRIX 6.2 has some other sized structs. See http://www.bitmechanic.com/mail-archives/mysql/current/0418.html This problem seems to be fixed by gcc v2.95.2, gcc v2.8.1 is reported as broken. Check your gcc version for this bug before installing sendmail.

IRIX 6.4

The IRIX 6.5.4 version of /bin/m4 does not work properly with sendmail. Either install fw_m4.sw.m4 off the Freeware_May99 CD and /usr/freeware/bin/m4 or install and use GNU m4.

NeXT or NEXTSTEP

NEXTSTEP 3.3 and earlier ship with the old DBM library. Also, Berkeley DB does not currently run on NEXTSTEP.

If you are compiling on NEXTSTEP, you will have to create an empty file "unistd.h" and create a file "dirent.h" containing:

#include <sys/dir.h> #define dirent direct

(The devtools/OS/NeXT should try to do both of these for you.)

Apparently, there is a bug in getservbyname on Nextstep 3.0 that causes it to fail under some circumstances with the message "SYSERR: service "smtp" unknown" logged. You should be able to work around this by including the line:

OOPort=25

in your .cf file.

BSDI (BSD/386) 1.0, NetBSD 0.9, FreeBSD 1.0

The "m4" from BSDI won't handle the config files properly. I haven't had a chance to test this myself.

The M4 shipped in FreeBSD and NetBSD 0.9 don't handle the config files properly. One must use either GNU m4 1.1 or the PD-M4 recently posted in comp.os.386bsd.bugs (and maybe others). NetBSD-current includes the PD-M4 (as stated in the NetBSD file CHANGES).

FreeBSD 1.0 RELEASE has uname(2) now. Use -DUSEUNAME in order to use it (look into devtools/OS/FreeBSD). NetBSD-current may have it too but it has not been verified.

The latest version of Berkeley DB uses a different naming scheme than the version that is supplied with your release. This means you will be able to use the current version of Berkeley DB with sendmail as long you use the new db.h when compiling sendmail and link it against the new libdb.a or libdb.so. You should probably keep the original db.h in /usr/include and the new db.h in /usr/local/include.

4.3BSD

If you are running a "virgin" version of 4.3BSD, you'll have a very old resolver and be missing some header files. The header files are simple -- create empty versions and everything will work fine. For the resolver you should really port a new version (4.8.3 or later) of the resolver; 4.9 is available on gatekeeper.DEC.COM in pub/BSD/bind/4.9. If you are really determined to continue to use your old, buggy version (or as a shortcut to get sendmail working -- I'm sure you have the best intentions to port a modern version of BIND), you can copy ../contrib/oldbind.compat.c into sendmail and add the following to devtools/Site/site.config.m4:

APPENDDEF(`confOBJADD', `oldbind.compat.o')

OpenBSD (up to 2.9 Release), NetBSD, FreeBSD (up to 4.3-RELEASE)

m4 from *BSD won't handle libsm/Makefile.m4 properly, since the maximum length for strings is too short. You need to use GNU m4 or patch m4, for example.

A/UX

Date: Tue, 12 Oct 1993 18:28:28 -0400 (EDT)
From: "Eric C. Hagberg" <hagberg@med.cornell.edu>
Subject: Fix for A/UX ndbm

I guess this isn't really a sendmail bug, however, it is something that A/UX users should be aware of when compiling sendmail 8.6.

Apparently, the calls that sendmail is using to the ndbm routines in A/UX 3.0.x contain calls to "broken" routines, in that the aliases database will break when it gets "just a little big" (sorry I don't have exact numbers here, but it broke somewhere around 20-25 aliases for me.), making all aliases non-functional after exceeding this point.

What I did was to get the gnu-dbm-1.6 package, compile it, and then re-compile sendmail with "-lgdbm", "-DNDBM", and using the ndbm.h header file that comes with the gnu-package. This makes things behave properly.

[NOTE: see comment above about GDBM]

I suppose porting the New Berkeley DB package is another route, however, I made a quick attempt at it, and found it difficult (not easy at least); the gnu-dbm package "configured" and compiled easily.

[NOTE: Berkeley DB version 2.X runs on A/UX and can be used for database maps].

SCO Unix

From: Thomas Essebier <tom@stallion.oz.au>
Organization: Stallion Technologies Pty Ltd.

It will probably help those who are trying to configure sendmail 8.6.9 to know that if they are on SCO, they had better set OI-dnsrch or they will core dump as soon as they try to use the resolver. i.e., although SCO has _res.dnsrch defined, and is kinda BIND 4.8.3, it does not initialize it, nor does it understand 'search' in /etc/named.boot.

- sigh -

According to SCO, the m4 which ships with UnixWare 2.1.2 is broken. We recommend installing GNU m4 before attempting to build sendmail.

On some versions a bogus error value is listed if connections time out (large negative number). To avoid this explicitly set Timeout.connect to a reasonable value (several minutes).

DG/UX

Doug Anderson <dlander@afterlife.ncsc.mil> has successfully run V8 on the DG/UX 5.4.2 and 5.4R3.x platforms under heavy usage. Originally, the DG /bin/mail program wasn't compatible with the V8 sendmail, since the DG /bin/mail requires the environment variable "_FORCE_MAIL_LOCAL_=yes" be set. Version 8.7 now includes this in the environment before invoking the local mailer. Some have used procmail to avoid this problem in the past. It works but some have experienced file locking problems with their DG/UX ports of procmail.

Apollo DomainOS

If you are compiling on Apollo, you will have to create an empty file "unistd.h" (for DomainOS 10.3 and earlier) and create a file "dirent.h" containing:

#include <sys/dir.h> #define dirent direct

(devtools/OS/DomainOS will attempt to do both of these for you.)

HP-UX 8.00

Date: Mon, 24 Jan 1994 13:25:45 +0200
From: Kimmo Suominen <Kimmo.Suominen@lut.fi>
Subject: 8.6.5 w/ HP-UX 8.00 on s300

Just compiled and fought with sendmail 8.6.5 on a HP9000/360 (i.e., a series 300 machine) running HP-UX 8.00.

I was getting segmentation fault when delivering to a local user. With debugging I saw it was faulting when doing _free@libc... *sigh* It seems the new implementation of malloc on s300 is buggy as of 8.0, so I tried out the one in -lmalloc (malloc(3X)). With that it seems to work just dandy.

When linking, you will get the following error:

ld: multiply defined symbol _freespace in file /usr/lib/libmalloc.a

but you can just ignore it. You might want to add this info to the README file for the future...

Linux

Something broke between versions 0.99.13 and 0.99.14 of Linux: the flock() system call gives errors. If you are running .14, you must not use flock. You can do this with -DHASFLOCK=0. We have also been getting complaints since version 2.4.X was released. Unless the bug is fixed before sendmail 8.13 is shipped, 8.13 will change the default locking method to fcntl() for Linux kernel version 2.4 and later. Be sure to update other sendmail related programs to match locking techniques (some examples, besides makemap and mail.local, include procmail, mailx, mutt, elm, etc).

Around the inclusion of bind-4.9.3 & Linux libc-4.6.20, the initialization of the _res structure changed. If /etc/hosts.conf was configured as "hosts, bind" the resolver code could return "Name server failure" errors. This is supposedly fixed in later versions of libc (>= 4.6.29?), and later versions of sendmail (> 8.6.10) try to work around the problem.

Some older versions (< 4.6.20?) of the libc/include files conflict with sendmail's version of cdefs.h. Deleting sendmail's version on those systems should be non-harmful, and new versions don't care.

NOTE ON LINUX & BIND: By default, the Makefile generated for Linux includes header files in /usr/local/include and libraries in /usr/local/lib. If you've installed BIND on your system, the header files typically end up in the search path and you need to add "-lresolv" to the LIBS line in your Makefile. Really old versions may need to include "-l44bsd" as well (particularly if the link phase complains about missing strcasecmp, strncasecmp or strpbrk). Complaints about an undefined reference to `__dn_skipname' in domain.o are a sure sign that you need to add -lresolv to LIBS. Newer versions of Linux are basically threaded BIND, so you may or may not see complaints if you accidentally mix BIND headers/libraries with virginal libc. If you have BIND headers in /usr/local/include (resolv.h, etc) you *should* be adding -lresolv to LIBS. Data structures may change and you'd be asking for a core dump.

A number of problems have been reported regarding the Linux 2.2.0 kernel. So far, these problems have been tracked down to syslog() and DNS resolution. We believe the problem is with the poll() implementation in the Linux 2.2.0 kernel and poll()-aware versions of glib (at least up to 2.0.111).

glibc

glibc 2.2.1 (and possibly other versions) changed the value of __RES in resolv.h but failed to actually provide the IPv6 API changes that the change implied. Therefore, compiling with -DNETINET6 fails.

Workarounds:

  1. Compile without -DNETINET6
  2. Build against a real BIND 8.2.2 include/lib tree
  3. Wait for glibc to fix it

AIX 4.X

The AIX 4.X linker uses library paths specified during compilation using -L for run-time shared library searches. Therefore, it is vital that relative and unsafe directory paths not be using when compiling sendmail. Because of this danger, by default, compiles on AIX use the -blibpath option to limit shared libraries to /usr/lib and /lib. If you need to allow more directories, such as /usr/local/lib, modify your devtools/Site/site.AIX.4.2.m4, site.AIX.4.3.m4, and/or site.AIX.4.x.m4 file(s) and set confLDOPTS appropriately. For example:

define(`confLDOPTS', `-blibpath:/usr/lib:/lib:/usr/local/lib')

Be sure to only add (safe) system directories.

The AIX version of GNU ld also exhibits this problem. If you are using that version, instead of -blibpath, use its -rpath option. For example:

gcc -Wl,-rpath /usr/lib -Wl,-rpath /lib -Wl,-rpath /usr/local/lib

AIX 4.X

If the test program t-event (and most others) in libsm fails, check your compiler settings. It seems that the flags -qnoro or -qnoroconst on some AIX versions trigger a compiler bug. Check your compiler settings or use cc instead of xlc.

AIX 4.0-4.2, maybe some AIX 4.3 versions

The AIX m4 implements a different mechanism for ifdef which is inconsistent with other versions of m4. Therefore, it will not work properly with the sendmail Build architecture or m4 configuration method. To work around this problem, please use GNU m4 from ftp.gnu.org. The problem seems to be solved in AIX 4.3.3 at least.

AIX 4.3.3

From: Valdis.Kletnieks@vt.edu Date: Sun, 02 Jul 2000 03:58:02 -0400

Under AIX 4.3.3, after applying bos.adt.include 4.3.3.12 to close the BIND 8.2.2 security holes, you can no longer build with -DNETINET6 because they changed the value of __RES in resolv.h but failed to actually provide the API changes that the change implied.

Workarounds:

  1. Compile without -DNETINET6
  2. Build against a real BIND 8.2.2 include/lib tree
  3. Wait for IBM to fix it

AIX 3.x

This version of sendmail does not support MB, MG, and MR resource records, which are supported by AIX sendmail.

Several people have reported that the IBM-supplied named returns fairly random results -- the named should be replaced. It is not necessary to replace the resolver, which will simplify installation. A new BIND resolver can be found at http://www.isc.org/isc/.

AIX 3.1.x

The supplied load average code only works correctly for AIX 3.2.x. For 3.1, use -DLA_TYPE=LA_SUBR and get the latest ``monitor'' package by Jussi Maki <jmaki@hut.fi> from ftp.funet.fi in the directory pub/unix/AIX/rs6000/monitor-1.12.tar.Z; use the loadavgd daemon, and the getloadavg subroutine supplied with that package. If you don't care about load average throttling, just turn off load average checking using -DLA_TYPE=LA_ZERO.

RISC/os

RISC/os from MIPS is a merged AT&T/Berkeley system. When you compile on that platform you will get duplicate definitions on many files. You can ignore these.

System V Release 4 Based Systems

There is a single devtools OS that is intended for all SVR4-based systems (built from devtools/OS/ SVR4). It defines __svr4__, which is predefined by some compilers. If your compiler already defines this compile variable, you can delete the definition from the generated Makefile or create a devtools/Site/site.config.m4 file.

It's been tested on Dell Issue 2.2.

DELL SVR4

Date: Mon, 06 Dec 1993 10:42:29 EST
From: "Kimmo Suominen" <kim@grendel.lut.fi>
Message-ID: <2d0352f9.lento29@lento29.UUCP>
To: eric@cs.berkeley.edu
Cc: sendmail@cs.berkeley.edu
Subject: Notes for DELL SVR4

Eric,

Here are some notes for compiling Sendmail 8.6.4 on DELL SVR4. I ran across these things when helping out some people who contacted me by e-mail.

  1. Use gcc 2.4.5 (or later?). Dell distributes gcc 2.1 with their Issue 2.2 Unix. It is too old, and gives you problems with clock.c, because sigset_t won't get defined in <sys/signal.h>. This is due to a problematic protection rule in there, and is fixed with gcc 2.4.5.
  2. If you don't use the new Berkeley DB (-DNEWDB), then you need to add "-lc -lucb" to the libraries to link with. This is because the -ldbm distributed by Dell needs the bcopy, bcmp and bzero functions. It is important that you specify both libraries in the given order to be sure you only get the BSTRING functions from the UCB library (and not the signal routines etc.).
  3. Don't leave out "-lelf" even if compiling with "-lc -lucb". The UCB library also has another copy of the nlist routines, but we do want the ones from "-lelf".

If anyone needs a compiled gcc 2.4.5 and/or a ported DB library, they can use anonymous ftp to fetch them from lut.fi in the /kim directory. They are copies of what I use on grendel.lut.fi, and offering them does not imply that I would also support them. I have sent the DB port for SVR4 back to Keith Bostic for inclusion in the official distribution, but I haven't heard anything from him as of today.

  • gcc-2.4.5-svr4.tar.gz: (gcc 2.4.5 and the corresponding libg++)
  • db-1.72.tar.gz: (with source, objects and a installed copy)

Cheers
+ Kim

--
* Kimmo.Suominen@lut.fi * SysVr4 enthusiast at GRENDEL.LUT.FI *
* KIM@FINFILES.BITNET * Postmaster and Hostmaster at LUT.FI *
* + 358 200 865 718 * Unix area moderator at NIC.FUNET.FI *

ConvexOS 10.1 and below

In order to use the name server, you must create the file /etc/use_nameserver. If this file does not exist, the call to res_init() will fail and you will have absolutely no access to DNS, including MX records.

Amdahl UTS 2.1.5

In order to get UTS to work, you will have to port BIND 4.9. The vendor's BIND is reported to be ``totally inadequate.'' See sendmail/contrib/AmdahlUTS.patch for the patches necessary to get BIND 4.9 compiled for UTS.

UnixWare

According to Alexander Kolbasov <sasha@unitech.gamma.ru>, the m4 on UnixWare 2.0 (still in Beta) will core dump on the config files. GNU m4 and the m4 from UnixWare 1.x both work. According to Larry Rosenman <ler@lerami.lerctr.org>:

UnixWare 2.1.[23]'s m4 chokes (not obviously) when processing the 8.9.0 cf files.

I had a LOCAL_RULE_0 that wound up AFTER the SBasic_check_rcpt rules using the SCO supplied M4. GNU M4 works fine.

UNICOS 8.0.3.4

Some people have reported that the -O flag on UNICOS can cause problems. You may want to turn this off if you have problems running sendmail. Reported by Jerry G. DeLapp <jgd@acl.lanl.gov>.

Darwin/Mac OS X (10.X.X)

The linker errors produced regarding getopt() and its associated variables can safely be ignored.

From Mike Zimmerman <zimmy@torrentnet.com>:

From scratch here is what Darwin users need to do to the standard 10.0.0, 10.0.1 install to get sendmail working. From MacOSX.com Forums:

  1. chmod g-w / /private /private/etc
  2. Properly set HOSTNAME in /etc/hostconfig to your FQDN: HOSTNAME=-my.domain.com-
  3. Edit /etc/rc.boot: hostname my.domain.com domainname domain.com
  4. Edit /System/Library/StartupItems/Sendmail/Sendmail: Remove the "&" after the sendmail command: /usr/sbin/sendmail -bd -q1h

From Carsten Klapp <carsten.klapp@home.com>:

The easiest workaround is to remove the group-writable permission for the root directory and the symbolic /etc inherits this change. While this does fix sendmail, the unfortunate side-effect is the OS X admin will no longer be able to manipulate icons in the top level of the Startup disk unless logged into the GUI as the superuser.

In applying the alternate workaround, care must be taken while swapping the symlink /etc with the directory /private/etc. In all likelihood any admin who is concerned with this sendmail error has enough experience to not accidentally harm anything in the process.

  1. Swap the /etc symlink with /private/etc (as superuser):
    1. rm /etc
    2. mv /private/etc /etc
    3. ln -s /etc /private/etc
  2. Set / to group unwritable (as superuser): chmod g-w /

Darwin/Mac OS X (10.1.5)

Apple's upgrade to sendmail 8.12 is incorrectly configured. You will need to manually fix it up by doing the following:

  1. chown smmsp:smmsp /var/spool/clientmqueue
  2. chmod 2770 /var/spool/clientmqueue
  3. chgrp smmsp /usr/sbin/sendmail
  4. chmod g+s /usr/sbin/sendmail

From Daniel J. Luke <dluke@geeklair.net>:

It appears that setting the sendmail.cf property in /locations/sendmail in NetInfo on Mac OS X 10.1.5 with sendmail 8.12.4 causes 'bad things' to happen.

Specifically sendmail instances that should be getting their config from /etc/mail/submit.cf don't (so mail/mutt/perl scripts which open pipes to sendmail stop working as sendmail tries to write to /var/spool/mqueue and cannot as sendmail is no longer suid root).

Removing the entry from NetInfo fixes this problem.

GNU getopt

I'm told that GNU getopt has a problem in that it gets confused by the double call. Use the version in conf.c instead.

BIND 4.9.2 and Ultrix

If you are running on Ultrix, be sure you read conf/Info.Ultrix in the BIND distribution very carefully -- there is information in there that you need to know in order to avoid errors of the form:

/lib/libc.a(gethostent.o): sethostent: multiply defined /lib/libc.a(gethostent.o): endhostent: multiply defined /lib/libc.a(gethostent.o): gethostbyname: multiply defined /lib/libc.a(gethostent.o): gethostbyaddr: multiply defined

during the link stage.

BIND 8.X

BIND 8.X returns HOST_NOT_FOUND instead of TRY_AGAIN on temporary DNS failures when trying to find the hostname associated with an IP address (gethostbyaddr()). This can cause problems as $&{client_name} based lookups in class R ($=R) and the access database won't succeed.

This will be fixed in BIND 8.2.1. For earlier versions, this can be fixed by making dns the last name service queried for host resolution in /etc/irs.conf:

hosts local continue hosts dns

strtoul

Some compilers (notably gcc) claim to be ANSI C but do not include the ANSI-required routine "strtoul". If your compiler has this problem, you will get an error in srvrsmtp.c on the code:

# ifdef defined(__STDC__) && !defined(BROKEN_ANSI_LIBRARY) e->e_msgsize = strtoul(vp, (char **) NULL, 10); # else e->e_msgsize = strtol(vp, (char **) NULL, 10); # endif

You can use -DBROKEN_ANSI_LIBRARY to get around this problem.

Listproc 6.0c

Date: 23 Sep 1995 23:56:07 GMT
Message-ID: <95925101334.~INN-AUMa00187.comp-news@dl.ac.uk>
From: alansz@mellers1.psych.berkeley.edu (Alan Schwartz)
Subject: Listproc 6.0c + Sendmail 8.7 [Helpful hint]

Just upgraded to sendmail 8.7, and discovered that listproc 6.0c breaks, because it, by default, sends a blank "HELO" rather than a "HELO hostname" when using the 'system' or 'telnet' mail method.

The fix is to include -DZMAILER in the compilation, which will cause it to use "HELO hostname" (which Z-mail apparently requires as well. :)

OpenSSL

OpenSSL versions prior to 0.9.6 use a macro named Free which conflicts with existing macro names on some platforms, such as AIX. Do not use 0.9.3, but OpenSSL 0.9.5a or later if compatible with 0.9.5a.

PH

PH support is provided by Mark Roth <roth@uiuc.edu>. The map is described at http://www-dev.cso.uiuc.edu/sendmail/

NOTE: The "spacedname" pseudo-field which was used by earlier versions of the PH map code is no longer supported! See the URL listed above for more information.

Please contact Mark Roth for support and questions regarding the map.

TCP Wrappers

If you are using -DTCPWRAPPERS to get TCP Wrappers support you will also need to install libwrap.a and modify your site.config.m4 file or the generated Makefile to include -lwrap in the LIBS line (make sure that INCDIRS and LIBDIRS point to where the tcpd.h and libwrap.a can be found).

TCP Wrappers is available at ftp://ftp.porcupine.org/pub/security/.

If you have alternate MX sites for your site, be sure that all of your MX sites reject the same set of hosts. If not, a bad guy whom you reject will connect to your site, fail, and move on to the next MX site, which will accept the mail for you and forward it on to you.

Regular Expressions (MAP_REGEX)

If sendmail linking fails with:

undefined reference to 'regcomp'

or sendmail gives an error about a regular expression with:

pattern-compile-error: : Operation not applicable

Your libc does not include a running version of POSIX-regex. Use librx or regex.o from the GNU Free Software Foundation, ftp://ftp.gnu.org/pub/gnu/rx-?.?.tar.gz or ftp://ftp.gnu.org/pub/gnu/regex-?.?.tar.gz. You can also use the regex-lib by Henry Spencer, ftp://ftp.funet.fi/pub/languages/C/spencer/regex.shar.gz Make sure, your compiler reads regex.h from the distribution, not from /usr/include, otherwise sendmail will dump a core.

Manual Pages

The manual pages have been written against the -man macros, and should format correctly with any reasonable *roff.

Debugging Hooks

As of 8.6.5, sendmail daemons will catch a SIGUSR1 signal and log some debugging output (logged at LOG_DEBUG severity). The information dumped is:

* The value of the $j macro. * A warning if $j is not in the set $=w. * A list of the open file descriptors. * The contents of the connection cache. * If ruleset 89 is defined, it is evaluated and the results printed.

This allows you to get information regarding the runtime state of the daemon on the fly. This should not be done too frequently, since the process of rewriting may lose memory which will not be recovered. Also, ruleset 89 may call non-reentrant routines, so there is a small non-zero probability that this will cause other problems. It is really only for debugging serious problems.

A typical formulation of ruleset 89 would be:

R$* $@ $>0 some test address

Description of Source Files

The following list describes the files in the sendmail directory:

Build
Shell script for building sendmail.
Makefile
A convenience for calling ./Build.
Makefile.m4
A template for constructing a makefile based on the information in the devtools directory.
README
This file.
TRACEFLAGS
My own personal list of the trace flags -- not guaranteed to be particularly up to date.
alias.c
Does name aliasing in all forms.
aliases.5
Man page describing the format of the aliases file.
arpadate.c
A subroutine which creates ARPANET standard dates.
bf.c
Routines to implement memory-buffered file system using hooks provided by libsm now (formerly Torek stdio library).
bf.h
Buffered file I/O function declarations and data structure and function declarations for bf.c.
collect.c
The routine that actually reads the mail into a temp file. It also does a certain amount of parsing of the header, etc.
conf.c
The configuration file. This contains information that is presumed to be quite static and non- controversial, or code compiled in for efficiency reasons. Most of the configuration is in sendmail.cf.
conf.h
Configuration that must be known everywhere.
control.c
Routines to implement control socket.
convtime.c
A routine to sanely process times.
daemon.c
Routines to implement daemon mode.
deliver.c
Routines to deliver mail.
domain.c
Routines that interface with DNS (the Domain Name System).
envelope.c
Routines to manipulate the envelope structure.
err.c
Routines to print error messages.
headers.c
Routines to process message headers.
helpfile
An example helpfile for the SMTP HELP command and -bt mode.
macro.c
The macro expander. This is used internally to insert information from the configuration file.
mailq.1
Man page for the mailq command.
main.c
The main routine to sendmail. This file also contains some miscellaneous routines.
makesendmail
A convenience for calling ./Build.
map.c
Support for database maps.
mci.c
Routines that handle mail connection information caching.
milter.c
MTA portions of the mail filter API.
mime.c
MIME conversion routines.
newaliases.1
Man page for the newaliases command.
parseaddr.c
The routines which do address parsing.
queue.c
Routines to implement message queueing.
readcf.c
The routine that reads the configuration file and translates it to internal form.
recipient.c
Routines that manipulate the recipient list.
sasl.c
Routines to interact with Cyrys-SASL.
savemail.c
Routines which save the letter on processing errors.
sendmail.8
Man page for the sendmail command.
sendmail.h
Main header file for sendmail.
sfsasl.c
I/O interface between SASL/TLS and the MTA.
sfsasl.h
Header file for sfsasl.c.
shmticklib.c
Routines for shared memory counters.
sm_resolve.c
Routines for DNS lookups (for DNS map type).
sm_resolve.h
Header file for sm_resolve.c.
srvrsmtp.c
Routines to implement server SMTP.
stab.c
Routines to manage the symbol table.
stats.c
Routines to collect and post the statistics.
statusd_shm.h
Data structure and function declarations for shmticklib.c.
sysexits.c
List of error messages associated with error codes in sysexits.h.
sysexits.h
List of error codes for systems that lack their own.
timers.c
Routines to provide microtimers.
timers.h
Data structure and function declarations for timers.h.
tls.c
Routines for TLS.
trace.c
The trace package. These routines allow setting and testing of trace flags with a high granularity.
udb.c
The user database interface module.
usersmtp.c
Routines to implement user SMTP.
util.c
Some general purpose routines used by sendmail.
version.c
The version number and information about this version of sendmail.

(Derived from sendmail/README Version 8.355.2.11, last updated 2002-Dec-28)