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.
See also diskless(8) manual page.
The ISC DHCP server can answer both BOOTP and DHCP requests.
ISC DHCP 3.1 is not part of the base system. You will first need to install the net/isc-dhcp31-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:
default-lease-time 600; max-lease-time 7200; authoritative; option domain-name "example.com"; option domain-name-servers 192.168.4.1; option routers 192.168.4.1; subnet 192.168.4.0 netmask 255.255.255.0 { use-host-decl-names on; option subnet-mask 255.255.255.0; option broadcast-address 192.168.4.255; host margaux { hardware ethernet 01:23:45:67:89:ab; fixed-address margaux.example.com; next-server 192.168.4.4; filename "/data/misc/kernel.diskless"; option root-path "192.168.4.4:/data/misc/diskless"; } host corbieres { hardware ethernet 00:02:b3:27:62:df; fixed-address corbieres.example.com; next-server 192.168.4.4; filename "pxeboot"; option root-path "192.168.4.4:/data/misc/diskless"; } }
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.
.def100:\ :hn:ht=1:sa=192.168.4.4:vm=rfc1048:\ :sm=255.255.255.0:\ :ds=192.168.4.1:\ :gw=192.168.4.1:\ :hd="/tftpboot":\ :bf="/kernel.diskless":\ :rp="192.168.4.4:/data/misc/diskless": margaux:ha=0123456789ab:tc=.def100
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.
If you are using PXE or Etherboot configured to use TFTP, you need to enable tftpd on the file server:
Create a directory from which tftpd will serve the files, e.g. /tftpboot.
Add this line to your /etc/inetd.conf:
tftp dgram udp wait root /usr/libexec/tftpd tftpd -l -s /tftpboot
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.
# /etc/rc.d/mountd restart
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):
hints "GENERIC.hints"
You need to create a root file system for the diskless workstations, in the location listed as root-path in dhcpd.conf.
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 make installworld && make installkernel 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.
If needed, a swap file located on the server can be accessed via NFS.
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:
# dd if=/dev/zero of=/path/to/swapfile bs=1k count=1 oseek=100000
To enable it you have to add the following line to your rc.conf:
swapfile=/path/to/swapfile
If the diskless workstation is configured to run X, you will have to adjust the XDM configuration file, which puts the error log on /usr by default.
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.