An MX record is simply the method used by DNS to route mail bound for one machine to another instead. An MX record is created by a single line in one of your named(8) files:
hostA IN MX 10 hostB
This line says that all mail destined for hostA in your domain should instead be delivered to hostB in your domain. The IN says that this is an Internet-type record, and the 10 is the cost for using this MX record.
An MX record can point to another host or to the original host:
hostA IN MX 0 hostA
This line says that mail for hostA will be delivered to hostA. Such records might seem redundant, but they are not because a host can have many MX records (one of which can point to itself):
hostA IN MX 0 hostA IN MX 10 hostB
Here, hostA has the lowest cost (0 versus 10 for hostB), so the first delivery attempt will be to hostA. If hostA is unavailable, delivery will be attempted to hostB instead.
Usually, MX records point to hosts inside the same domain. Therefore, managing them does not require the cooperation of others. But it is legal for MX records to point to hosts in different domains:
hostA IN MX 0 hostA IN MX 10 host.other.domain.
Here, you must contact the administrator at other.domain and obtain permission before creating this MX record. We cover this concept in more detail when we discuss disaster preparation later in this chapter.
Although MX records are usually straightforward, there can be a few problems associated with them.
The A and AAAA records for a host are lines that give the host's IP address or addresses:
hostC IN A 184.108.40.206 IPv4 hostC IN AAAA 3ffe:8050:201:1860:42::1IPv6
Here, hostC is the host's name. The IN says this is an Internet-type record. The A marks this as an IPv4 A record, with the IP address 220.127.116.11. The AAAA marks this as an IPv6 AAAA record, with the IP address 3ffe:8050:201:1860:42::1.
An MX record must point to a hostname that has an A or AAAA record. To illustrate, consider the following:
hostA IN MX 10 hostB illegal IN MX 20 hostC hostB IN MX 10 hostC hostC IN A 18.104.22.168
Note that hostB lacks an A record but hostC has one. It is illegal to point an MX record at a host that lacks an A or AAAA record. Therefore, the first line in the preceding example is illegal, whereas the second line is legal.
Although such a mistake is difficult to make when maintaining your own domain tables, it can easily happen if you rely on a name server in someone else's domain, as shown:
hostA IN MX 10 mail.other.domain.
The other administrator might, for example, retire the machine mail and replace its A record with an MX record that points to a different machine. Unless you are notified of the change, your MX record will suddenly become illegal.
Note that although an MX record must point to a hostname that has an A or AAAA record, it is illegal for an MX record to point directly to an A or AAAA record:
hostA IN MX 10 22.214.171.124 illegal
The sendmail program is frequently more forgiving than other MTAs because it accepts an MX record that points to a CNAME record. The presumption is that, eventually, the CNAME will correctly point to an A or AAAA record. But beware, this kind of indirection can cost additional DNS lookups. Consider this example of an exceptionally bad setup:
hostA IN MX 10 mailhub mailhub IN CNAME nfsmast nfsmast IN CNAME hostB hostB IN A 126.96.36.199
First, sendmail looks up hostA and gets an MX record pointing to mailhub. Because there is only a single MX record, sendmail considers mailhub to be official. Next, mailhub is looked up to find an A or AAAA record (IP address), but instead a CNAME (nfsmast) is returned. Now, sendmail must look up the CNAME nfsmast to find its A or AAAA record. But again a CNAME is returned instead. So sendmail must again look for an A or AAAA record (this time with hostB). Finally, sendmail succeeds by finding the A record for hostB, but only after far too many lookups.
 Most of this happens inside the gethostbyname(3) or getipnodebyname(3) C-library routine.
The correct way to form the previous DNS file entries is as follows:
hostA IN MX 10 hostB mailhub IN CNAME hostB nfsmast IN CNAME hostB hostB IN A 188.8.131.52
In general, try to construct DNS records in such a way that the fewest lookups are required to resolve any records.
Consider the following MX setup, which causes all mail for hostA to be sent to hostB and all mail for hostB to be sent to hostB, or to hostC if hostB is down:
 We are fudging for the sake of simplicity. Here, we assume that all the hosts also have A records.
hostA IN MX 10 hostB hostB IN MX 10 hostB IN MX 20 hostC
One might expect sendmail to be smart and deliver mail for hostA to hostC if hostB is down. But sendmail won't do that. The RFC standards do not allow it to recursively look up additional MX records. If sendmail did, it could get hopelessly entangled in MX loops. Consider the following:
hostA IN MX 10 hostB hostB IN MX 10 hostB IN MX 20 hostC hostC IN MX 10 hostA potential loop
If your intention is to have hostA MX to two other hosts, you must state that explicitly:
hostA IN MX 10 hostB IN MX 20 hostC hostB IN MX 10 hostB IN MX 20 hostC
Another reason sendmail refuses to follow MX records beyond the target host is that costs in such a situation are undefined. Consider the previous example with the potential loop. What is the cost of hostA when MX'd by hostB to hostC? Should it be the minimum of 10, the maximum of 20, the mean of 15, or the sum of 30?
Wildcard MX records should not be used unless you understand all the possible risks. They can provide a shorthand way of MX'ing many hosts with a single MX record, but it is a shorthand that can be easily abused. For example:
*.dc.gov. IN MX 10 hostB
This says that any host in the domain .dc.gov (where that host doesn't have any record of its own) should have its mail forwarded to hostB.
; domain is .dc.gov *.dc.gov. IN MX 10 hostB hostA IN MX 10 hostC hostB IN A 184.108.40.206
Here, mail to hostD (no record at all) will be forwarded to hostB. But the wildcard MX record will be ignored for hostA and hostB because each has its own record.
Extreme care must be exercised in setting up wildcard MX records. It is easy to create ambiguous situations that DNS might not be able to handle correctly. Consider the following, for example:
; domain is sub.dc.gov *.dc.gov. IN MX 10 hostB.dc.gov. *.sub.dc.gov. IN MX 10 hostC.dc.gov.
Here, an unqualified name such as the plain hostD matches both wildcard records. This is ambiguous, so DNS automatically picks the most complete one (*.sub.dc.gov.) and supplies that MX record to sendmail.
One compelling weakness of wildcard MX records is that they match any hostname at all, even for machines that don't exist:
; domain is sub.dc.gov *.dc.gov. IN MX 10 hostB.dc.gov.
Here, mail to foo.dc.gov will be forwarded to hostB.dc.gov, even if there is no host foo in that domain.
Wildcard MX records almost never have any appropriate use on the Internet. They are often misunderstood and are often used just to save the effort of typing hundreds of MX records. They do, however, have legitimate uses behind firewall machines and on non-Internet networks.
Many older MTAs on the network ignore MX records. Some pre-Solaris Sun sites, for example, wrongly run the non-MX version of sendmail when they should use /usr/lib/sendmail.mx. Some Solaris sites wrongly do all host lookups with NIS when they should list dns on the hosts line of their /etc/nsswitch.conf file. Because of these and other mistakes, you will occasionally find some sites that insist on sending mail to a host even though that host has been explicitly MX'd to another.
To illustrate why this is bad, consider a UUCP host that has only an MX record. It has no A record because it is not on the network:
uuhost IN MX 10 uucpserver
Here, mail to uuhost will be sent to uucpserver, which will forward the message to uuhost with UUCP software. An attempt to ignore this MX record will fail because uuhost has no other records. Similar problems can arise for printers with direct network connections, terminal servers, and even workstations that don't run an SMTP daemon such as sendmail.
If you believe in DNS and disdain sites that don't, you can simply ignore the offending sites. In this case the mail will fail if your MX'd host doesn't run a sendmail daemon (or another MTA). This is not as nasty as it sounds. There is actually considerable support for this approach; failure to obey MX records is a clear violation of published network protocols. RFC1123, Host Requirements, section 5.3.5, notes that obeying MX records is mandatory. RFC1123 has existed for more than 12 years.
On the other hand, if you want to ensure that all mail is received, even on a workstation whose mail is MX'd elsewhere, you can run the sendmail daemon on every machine.
Although you are not required to have MX records for all hosts, there is good reason to consider doing so. To illustrate, consider the following host that has only an A record:
hostB IN A 220.127.116.11
When V8.12 and above sendmail first looks up this host, it asks the name server for that host's MX records. Because there are none, that request comes back empty. The sendmail program must then make a second lookup for the IP address.
When pre-V8.12 sendmail first looks up this host, it asks the local name server for all records. Because there is only an A record, that is all it gets. But note that asking for any record causes the local name server to cache the information.
The next time sendmail looks up this same host, the local name server will return the A record from its cache. This is faster and reduces Internet traffic. The cached information is "nonauthoritative" (because it is a copy) and includes no MX records (because there are none).
When pre-V8.12 sendmail gets a nonauthoritative reply that lacks MX records, it is forced to do another DNS lookup. This time, it specifically asks for MX records. In this case there are none, so it gets none.
Because hostB lacks an MX record, sendmail performs a DNS lookup each and every time mail is sent to that host. If hostB were a major mail-receiving site, its lack of an MX record would cause many sendmail programs, all over the world, to waste network bandwidth with useless DNS lookups.
We strongly recommend that every host on the Internet have at least one MX record. As a minimum, it can simply point to itself with a 0 cost:
hostB IN A 18.104.22.168 IN MX 0 hostB
This will not change how mail is routed to hostB but will reduce the number of DNS lookups required.
RFC974 leaves the treatment of ambiguous MX records to the implementor's discretion. This has generated much debate in sendmail circles. Consider the following:
foo IN MX 10 hostA foo IN MX 20 hostB mail from hostB to foo foo IN MX 30 hostC
When mail is sent from a host (hostB) that is an MX record for the receiving host (foo), all MX records that have a cost equal to or greater than that of hostB must be discarded. The mail is then delivered to the remaining MX host with the lowest cost (hostA). This is a sensible rule because it prevents hostB from wrongly trying to deliver to itself.
It is possible to configure hostB so that it views the name foo as a synonym for its own name. Such a configuration results in hostB never looking up any MX records because it recognizes mail to foo as local.
But what should happen if hostB does not recognize foo as local and if there is no hostA?
no hostA foo IN MX 20 hostB mail from hostB to foo foo IN MX 30 hostC
Again, RFC974 says that when mail is being sent from a host (hostB) that is an MX record for the receiving host (foo), all MX records that have a cost equal to or greater than that of hostB must be discarded. In this example, that leaves zero MX records. Three courses of action are now open to sendmail, but RFC974 doesn't say which it should use:
Assume that this is an error condition. Clearly, hostB should have been configured to recognize foo as local. It didn't (hence the MX lookup and discarding in the first place), so it must not have known what it was doing. V8 sendmail with the TryNullMXList option (TryNullMXList) not set (undeclared or declared as false) will bounce the mail message with this message:
553 5.3.5 hostconfig error: mail loops back to me (MX problem?)
Look to see whether foo has an A record. If it does, go ahead and try to deliver the mail message directly to foo. If it lacks an A record, bounce the message. This approach runs the risk that foo might not be configured to properly accept mail (thus causing mail to disappear down a black hole). Still, this approach can be desirable in some circumstances. V8 sendmail with the TryNullMXList option (TryNullMXList) set to true always tries to connect to foo.
 As does the UIUC version of IDA sendmail. Other versions of IDA (such as KJS) do not. Note that defining the TryNullMXList option to true has the undesirable side effect of allowing anyone on the Internet to use your host as a backup MX server, without your permission.
Assume (even though it has not been configured to do so) that foo should be treated as local to hostB. No version of sendmail makes this assumption.
This situation is not an idle exercise. Consider the MX record for uuhost presented in the previous section:
uuhost IN MX 10 uucpserver
Here, uuhost has no A or AAAA record because it is connected to uucpserver via a dial-up line. If uucpserver is not configured to recognize uuhost as one of its UUCP clients, and if mail is sent from uucpserver to uuhost, it will query DNS and get itself as the MX record for uuhost. As we have shown, that MX record is discarded, and an ambiguous situation has developed.