In this document we will look at a simple definition of inetd(8), how several files that relate to inetd(8) work (not that these files are not related to other software), how to add a service to inetd(8) and some considerations both to use inetd(8) for a particular service and times when a service might be better off running outside of inetd(8).

In traditional Unix scenarios, one server (daemon) process watches for connections on a particular port, and handles incoming requests. Now if a machine offers many services, many daemon processes would be needed, mostly running idle but still wasting resources like memory. The internet super server, inetd, is an approach to this problem. It listens on a number of ports, and when it receives a request it then determines which program to run to handle the request and starts an instance of that program.

Following is a very simple diagram to illustrate inetd(8):

In the above diagram you can see the general idea. The inetd(8) process receives a request and then starts the appropriate server process. What inetd(8) is doing is software multiplexing. An important note here, regarding security: On many other UNIX-like systems, a package called tcpwrappers is used as a security enhancement for inetd(8). On NetBSD the tcpwrapper functionality is built into inetd(8) using libwrap.

The operation of inetd(8) is controlled by its own config file, surprisingly named /etc/inetd.conf, see inetd.conf(5). The inetd.conf file basically provides enabling and mapping of services the systems administrator would like to have multiplexed through inetd(8), indicating which program should be started for incoming requests on which port.

inetd.conf(5) is an ascii file containing one service per line, and several fields per line. The basic field layout is:

service-name socket-type protocol wait/nowait user:group server-program arguments

That is all a lot to digest and there are other things the systems administrator can do with some of the fields. Here is a sample line from an inetd.conf file:

ftp       stream  tcp    nowait  root   /usr/libexec/ftpd    ftpd -ll

From the left, the service-name is "ftp", socket-type is "stream", protocol is "tcp", inetd(8) won't wait for the server process to terminate ("nowait"), the process runs as user "root", path is /usr/libexec/ftpd and program name and arguments are "ftpd -ll". Notice in the last field, the program name is different from the service-name.

The next file to consider is the service name data base that can be found in /etc/services. This file basically contains information mapping a service name to a port number. The format of the /etc/services file is:

service-name port-number/protocol-name [aliases]

"service-name" is the name of the service, "port-number" is the port number assigned to the service, "protocol-name" is either "tcp" or "udp", and if alias names for a port are needed, they can be added as "aliases", separated by white spaces. Comments may be added after a hash mark (#).

Let's take a look at the "ssh" entries as an example:

ssh             22/tcp           # Secure Shell
ssh             22/udp

As we can see, from the left, the service name is "ssh", the port number is "22", the protocols are both "tcp" and "udp". Notice that there is a separate entry for every protocol a service can use (even on the same port).

Another file read by inetd(8) is /etc/protocols. This file has the information pertaining to DARPA Internet protocols. The format of the protocols name data base is:

protocol-name number [aliases]

where "protocol-name" describes the payload of an IP packet, e.g. "tcp" or "udp". "number" is the official protocol number assigned by IANA, and optional alias names can be added after that.

Let's look at the seventh entry in the /etc/protocols db as an example:

tcp     6       TCP             # transmission control protocol

Starting from the left, we see that the protocol name is "tcp", the number is "6" and the only aliases listed is "TCP", belonging to the Transmission Control Protocol as indicated by the comment in that line.

The rpc program number data base used by services with the "rpc" protocol type in inetd.conf(5) is kept in /etc/rpc and contains name mappings to rpc program numbers. The format of the file is:

server-name program-number aliases

For example, here is the nfs entry:

nfs             100003  nfsprog

As mentioned above, NetBSD's inetd(8) has the tcpwrapper package built in via the libwrap library. As such, inetd(8) can allow or deny access to each service on a more fine-grained base than just allowing a service to everyone, or not enabling it at all. The access control is defined in the files /etc/hosts.allow and /etc/hosts.deny, see the hosts_access(5) manpage.

Each of the two files contains several lines that describe access restrictions for a certain server. Access is allowed if permission is given in /etc/hosts.allow. If the service is not listened in /etc/hosts.allow but in /etc/hosts.deny, it is denied. If a service is listed in neither file, it is allowed, giving standard inetd(8) behaviour.

Each line in /etc/hosts.allow and /etc/hosts.deny contains a service either by name (as given in the field for argv[0] in /etc/inetd.conf, e.g. "ftpd" instead of "ftp"), or the special service "ALL" which obviously applies to all services. Following the service name is - separated by a colon - a number of access restrictions, which can be hostnames, domains, single IP addresses, whole IP subnets or some other restrictions, please check hosts_access(5) for all the details.

An example configuration that is mostly open but denies access to services to a certain host and all machines from a certain domain would look like this:

# /etc/hostname.deny:
ALL: some.host.name, .some.domain

Another example that would be mostly closed, denying access to all but very few machines would need entries in both /etc/hosts.allow and /etc/hosts.deny. The entry for /etc/hosts.deny would be:

# /etc/hosts.deny
ALL: ALL

The entry to allow a few hosts would be put into /etc/hosts.allow:

# /etc/hosts.allow
ALL: friend.host.domain otherfriend.otherhost.otherdomain

Many times a systems administrator will find that they need to add a service to their system that is not already in inetd(8) or they may wish to move a service to it because it does not get very much traffic. This is usually pretty simple, so as an example we will look at adding a version of POP3 on a NetBSD system.

In this case we have retrieved and installed the "cucipop" package, which can be found in pkgsrc/mail/cucipop. This server is pretty simple to use, the only oddities are different path locations. Since it is POP3 we know it is a stream oriented connection with "nowait". Running as "root" will be fine, the only item that is different is the location of the program and the name of the program itself.

So the first half of the new entry in /etc/inetd.conflooks like this:

pop3   stream  tcp     nowait  root

After installation, pkgsrc deposited cucipop in /usr/pkg/sbin/cucipop. So with the next field we have:

pop3   stream  tcp     nowait  root /usr/pkg/sbin/cucipop

Last, we want to use the Berkeley mailbox format, so our server program must be called with the -Y option. This leaves the entire entry looking like so:

pop3   stream  tcp     nowait  root /usr/pkg/sbin/cucipop cucipop -Y

We have added the service named "pop3" to /etc/inetd.conf. Next item to check is that the system can map the service name to a port number in /etc/services:

# grep ^pop3 /etc/services
pop3            110/tcp         # POP version 3
pop3            110/udp
pop3s           995/tcp                 # pop3 protocol over TLS/SSL (was spop3)
pop3s           995/udp                 # pop3 protocol over TLS/SSL (was spop3)

The "pop3" entries here are of interest, i.e. they are already contained in the /etc/services file shipped with NetBSD.

Now, to have inetd(8) use the new entry, we simply restart it using the rc script:

# sh /etc/rc.d/inetd restart

All done, in most cases, the software you are using has documentation that will specify the entry, in the off case it does not, sometimes it helps to try and find something similar to the server program you will be adding. A classic example of this is a MUD server which has built-in telnet. You can pretty much borrow the telnet entry and change parts where needed.

The decision to add or move a service into or out of inetd(8) is usually based on server load. As an example, on most systems the telnet daemon does not require as many new connections as say a mail server. Most of the time the administrator has to feel out if a service should be moved.

A good example I have seen is mail services such as smtp and pop. I had setup a mail server in which pop3 was in inetd(8) and exim was running in standalone, I mistakenly assumed it would run fine since there was a low amount of users, namely myself and a diagnostic account. The server was also setup to act as a backup MX and relay in case another heavily used one went down. When I ran some tests I discovered a huge time lag for pop connections remotely. This was because of my steady fetching of mail and the diagnostic user constantly mailing diagnostics back and forth. In the end I had to move the pop3 service out of inetd(8).

The reason for moving the service is actually quite interesting. When a particular service becomes heavily used, of course, it causes a load on the system. In the case of a service that runs within the inetd(8) meta daemon the effects of a heavily loaded service can also harm other services that use inetd(8). If the multiplexor is getting too many requests for one particular service, it will begin to affect the performance of other services that use inetd(8). The fix, in a situation like that, is to make the offending service run outside of inetd(8) so the response time of both the service and inetd(8) will increase.

Following is some additional reading and information about topics covered in this document.

NetBSD manual pages:

Miscellaneous links: