Updated by Jean-Fran�ois Dock�s.Reorganized and enhanced by Alex
Dupre.
A FreeBSD machine can boot over the network and operate without a local disk, using
file systems mounted from an NFS server. No system
modification is necessary, beyond standard configuration files. Such a system is
relatively easy to set up because all the necessary elements are readily available:
There are at least two possible methods to load the kernel over the network:
PXE: The Intel®
Preboot eXecution Environment system is a form of smart boot ROM built into some
networking cards or motherboards. See pxeboot(8) for more
details.
The Etherboot port (net/etherboot) produces ROM-able code to boot kernels over the
network. The code can be either burnt into a boot PROM on a network card, or loaded from
a local floppy (or hard) disk drive, or from a running MS-DOS® system. Many network cards are supported.
A sample script (/usr/share/examples/diskless/clone_root)
eases the creation and maintenance of the workstation's root file system on the server.
The script will probably require a little customization but it will get you started very
quickly.
Standard system startup files exist in /etc to detect and
support a diskless system startup.
Swapping, if needed, can be done either to an NFS
file or to a local disk.
There are many ways to set up diskless workstations. Many elements are involved, and
most can be customized to suit local taste. The following will describe variations on the
setup of a complete system, emphasizing simplicity and compatibility with the standard
FreeBSD startup scripts. The system described has the following characteristics:
The diskless workstations use a shared read-only / file
system, and a shared read-only /usr.
The root file system is a copy of a standard FreeBSD root (typically the server's),
with some configuration files overridden by ones specific to diskless operation or,
possibly, to the workstation they belong to.
The parts of the root which have to be writable are overlaid with md(4) file systems.
Any changes will be lost when the system reboots.
The kernel is transferred and loaded either with Etherboot
or PXE as some situations may mandate the use of
either method.
Caution: As described, this system is insecure. It should live in a protected
area of a network, and be untrusted by other hosts.
All the information in this section has been tested using FreeBSD 5.2.1-RELEASE.
Setting up diskless workstations is both relatively straightforward and prone to
errors. These are sometimes difficult to diagnose for a number of reasons. For
example:
Compile time options may determine different behaviors at runtime.
Error messages are often cryptic or totally absent.
In this context, having some knowledge of the background mechanisms involved is very
useful to solve the problems that may arise.
Several operations need to be performed for a successful bootstrap:
The machine needs to obtain initial parameters such as its IP address, executable
filename, server name, root path. This is done using the DHCP or BOOTP protocols. DHCP is a compatible extension of BOOTP, and uses the same port
numbers and basic packet format.
It is possible to configure a system to use only BOOTP. The bootpd(8) server
program is included in the base FreeBSD system.
However, DHCP has a number of advantages over BOOTP
(nicer configuration files, possibility of using PXE,
plus many others not directly related to diskless operation), and we will describe mainly
a DHCP configuration, with equivalent examples using
bootpd(8) when
possible. The sample configuration will use the ISC DHCP
software package (release 3.0.1.r12 was installed on the test server).
The machine needs to transfer one or several programs to local memory. Either TFTP or NFS are used. The
choice between TFTP and NFS is a compile time option in several places. A common source
of error is to specify filenames for the wrong protocol: TFTP typically transfers all files from a single directory on
the server, and would expect filenames relative to this directory. NFS needs absolute file paths.
The possible intermediate bootstrap programs and the kernel need to be initialized and
executed. There are several important variations in this area:
PXE will load pxeboot(8), which is a
modified version of the FreeBSD third stage loader. The loader(8) will obtain
most parameters necessary to system startup, and leave them in the kernel environment
before transferring control. It is possible to use a GENERIC
kernel in this case.
Etherboot, will directly load the kernel, with less
preparation. You will need to build a kernel with specific options.
PXE and Etherboot work
equally well; however, because kernels normally let the loader(8) do more work
for them, PXE is the preferred method.
If your BIOS and network cards support PXE, you should probably use it.
Finally, the machine needs to access its file systems. NFS is used in all cases.
The ISC DHCP server can answer both BOOTP and DHCP requests.
ISC DHCP 3.0 is not part of the base system. You will first
need to install the net/isc-dhcp3-server port or the corresponding package.
Once ISC DHCP is installed, it needs a configuration file
to run (normally named /usr/local/etc/dhcpd.conf). Here follows
a commented example, where host margaux uses Etherboot and host corbieres uses
PXE:
This option tells dhcpd to send the value in the host declarations as the hostname for the diskless host. An
alternate way would be to add an option host-name margaux inside the host
declarations.
The next-server directive designates the TFTP or NFS server to use
for loading loader or kernel file (the default is to use the same host as the DHCP server).
The filename directive defines the file that Etherboot or PXE will load for
the next execution step. It must be specified according to the transfer method used. Etherboot can be compiled to use NFS or TFTP. The FreeBSD
port configures NFS by default. PXE uses TFTP, which is why
a relative filename is used here (this may depend on the TFTP server configuration, but would be fairly typical). Also,
PXE loads pxeboot, not the
kernel. There are other interesting possibilities, like loading pxeboot from a FreeBSD CD-ROM /boot
directory (as pxeboot(8) can load a
GENERIC kernel, this makes it possible to use PXE to boot from a remote CD-ROM).
The root-path option defines the path to the root file
system, in usual NFS notation. When using PXE, it is possible to leave off the host's IP as long as you
do not enable the kernel option BOOTP. The NFS server
will then be the same as the TFTP one.
Here follows an equivalent bootpd configuration (reduced to
one client). This would be found in /etc/bootptab.
Please note that Etherboot must be compiled with the
non-default option NO_DHCP_SUPPORT in order to use BOOTP, and
that PXE needs DHCP. The only
obvious advantage of bootpd is that it exists in the base
system.
Etherboot's Web site
contains extensive documentation mainly intended for Linux systems, but
nonetheless containing useful information. The following will just outline how you would
use Etherboot on a FreeBSD system.
You must first install the net/etherboot package or port.
You can change the Etherboot configuration (i.e. to use
TFTP instead of NFS) by editing the Config file in
the Etherboot source directory.
For our setup, we shall use a boot floppy. For other methods (PROM, or MS-DOS program), please refer to the Etherboot documentation.
To make a boot floppy, insert a floppy in the drive on the machine where you installed
Etherboot, then change your current directory to the src directory in the Etherboot tree and
type:
#gmake bin32/devicetype.fd0
devicetype depends on the type of the Ethernet
card in the diskless workstation. Refer to the NIC file in the
same directory to determine the right devicetype.
By default, the pxeboot(8) loader
loads the kernel via NFS. It can be compiled to use
TFTP instead by specifying the LOADER_TFTP_SUPPORT option in /etc/make.conf. See the comments in /usr/share/examples/etc/make.conf for instructions.
There are two other make.conf options which may be useful
for setting up a serial console diskless machine: BOOT_PXELDR_PROBE_KEYBOARD, and BOOT_PXELDR_ALWAYS_SERIAL.
To use PXE when the machine starts, you will
usually need to select the Boot from network option in your
BIOS setup, or type a function key during the PC
initialization.
Note: It appears that at least some PXE
versions want the TCP version of TFTP. In this case, add a second line, replacing dgram udp with stream tcp.
Tell inetd to reread its configuration file. The inetd_enable="YES" must be in the /etc/rc.conf file for this command to execute correctly:
#/etc/rc.d/inetd restart
You can place the tftpboot directory anywhere on the server.
Make sure that the location is set in both inetd.conf and dhcpd.conf.
In all cases, you also need to enable NFS and
export the appropriate file system on the NFS
server.
Add this to /etc/rc.conf:
nfs_server_enable="YES"
Export the file system where the diskless root directory is located by adding the
following to /etc/exports (adjust the volume mount point and
replace margaux corbieres with the names of the
diskless workstations):
/data/misc -alldirs -ro margaux corbieres
Tell mountd to reread its configuration file. If you
actually needed to enable NFS in /etc/rc.conf at the first step, you probably want to reboot
instead.
If using Etherboot, you need to create a kernel
configuration file for the diskless client with the following options (in addition to the
usual ones):
options BOOTP # Use BOOTP to obtain IP address/hostname
options BOOTP_NFSROOT # NFS mount root file system using BOOTP info
You may also want to use BOOTP_NFSV3, BOOT_COMPAT and BOOTP_WIRED_TO (refer to
NOTES).
These option names are historical and slightly misleading as they actually enable
indifferent use of DHCP and BOOTP inside the kernel
(it is also possible to force strict BOOTP or DHCP
use).
Build the kernel (see Chapter 8), and copy it to the
place specified in dhcpd.conf.
Note: When using PXE, building a kernel with
the above options is not strictly necessary (though suggested). Enabling them will cause
more DHCP requests to be issued during kernel startup,
with a small risk of inconsistency between the new values and those retrieved by pxeboot(8) in some
special cases. The advantage of using them is that the host name will be set as a side
effect. Otherwise you will need to set the host name by another method, for example in a
client-specific rc.conf file.
Note: In order to be loadable with Etherboot, a
kernel needs to have the device hints compiled in. You would typically set the following
option in the configuration file (see the NOTES configuration
comments file):
This method is quick and will install a complete virgin system (not only the root file
system) into DESTDIR. All you have to do is simply execute the
following script:
#!/bin/sh
export DESTDIR=/data/misc/diskless
mkdir -p ${DESTDIR}
cd /usr/src; make buildworld && make buildkernel
cd /usr/src/etc; make distribution
Once done, you may need to customize your /etc/rc.conf and
/etc/fstab placed into DESTDIR according
to your needs.
The kernel does not support enabling NFS swap at
boot time. Swap must be enabled by the startup scripts, by mounting a writable file
system and creating and enabling a swap file. To create a swap file of appropriate size,
you can do like this:
When the server for the root file system is not running FreeBSD, you will have to
create the root file system on a FreeBSD machine, then copy it to its destination, using
tar or cpio.
In this situation, there are sometimes problems with the special files in /dev, due to differing major/minor integer sizes. A solution to
this problem is to export a directory from the non-FreeBSD server, mount this directory
onto a FreeBSD machine, and use devfs(5) to
allocate device nodes transparently for the user.