diskless -- booting a system over the network
The ability to boot a machine over the network is useful for diskless or
dataless machines, or as a temporary measure while repairing or reinstalling
file systems on a local disk. This file provides a general
description of the interactions between a client and its server when a
client is booting over the network.
When booting a system over the network, there are three phases of interaction
between client and server:
1. The stage-1 bootstrap, typically PXE built into your Ethernet card,
loads a second-stage boot program.
2. The second-stage boot program, typically pxeboot(8), loads modules
and the kernel, and boots the kernel.
3. The kernel NFS mounts the root directory and continues from there.
Each of these phases are described in further detail below.
First, the stage-1 bootstrap loads the stage-2 boot program over the network.
The stage-1 bootstrap typically uses BOOTP or DHCP to obtain the
filename to load, then uses TFTP to load the file. This file is typically
called pxeboot, and should be copied from /boot/pxeboot into the
TFTP directory on the server, which is typically /tftpdir.
The stage-2 boot program then loads additional modules and the kernel.
These files may not exist on the DHCP or BOOTP server. You can use the
next-server option available in DHCP configurations to specify the server
holding the second stage boot files and kernel. The stage-2 program uses
NFS or TFTP to obtain these files. By default, NFS is used. If you are
using pxeboot(8), you can install a version that uses TFTP by setting
LOADER_TFTP_SUPPORT=YES in your /etc/make.conf, then recompiling and
reinstalling pxeboot(8) via the command listed below. It is often necessary
to use TFTP here so you can place a custom kernel in /tftpdir/. If
you use NFS and do not have a custom root file system for the diskless
client, the stage-2 boot will load your server's kernel as the kernel for
the diskless machine, which may not be what you want to have happen.
cd /usr/src/sys/i386/boot
make clean; make; make install
cp /boot/pxeboot /tftpdir/
In phase 3, the kernel again uses DHCP or BOOTP to acquire configuration
information, and proceeds to mount the root file system and start operation.
The boot scripts recognize a diskless startup and perform the
actions found in /etc/rc.d/initdiskless and /etc/rc.d/diskless. On older
systems, the scripts are located in /etc/rc.diskless1 and
/etc/rc.diskless2.
In order to run a diskless client, you need the following:
+o An NFS server which exports a root and /usr partitions with appropriate
permissions. The diskless scripts work with read-only partitions,
as long as root is exported with -maproot=0 so that some system
files can be accessed. As an example, /etc/exports can contain
the following lines:
<ROOT> -ro -maproot=0 -alldirs <list of diskless clients>
/usr -ro -alldirs <list of diskless clients>
where <ROOT> is the mount point on the server of the root partition.
The script /usr/share/examples/diskless/clone_root can be used to
create a shared read-only root partition, but in many cases you may
decide to export (again as read-only) the root directory used by the
server itself.
+o A BOOTP or DHCP server. bootpd(8) can be enabled by uncommenting the
``bootps'' line in /etc/inetd.conf. A sample /etc/bootptab can be
the following:
.default:\
hn:ht=1:vm=rfc1048:\
:sm=255.255.255.0:\
:sa=<SERVER>:\
:gw=<GATEWAY>:\
:rp="<SERVER>:<ROOT>":
<CLIENT>:ha=0123456789ab:tc=.default
where <SERVER>, <GATEWAY> and <ROOT> have the obvious meanings.
+o A properly initialized root partition. The script
/usr/share/examples/diskless/clone_root can help in creating it,
using the server's root partition as a reference. If you are just
starting out, you should simply use the server's own root directory,
/, and not try to clone it.
You often do not want to use the same rc.conf or rc.local files for
the diskless boot as you do on the server. The diskless boot scripts
provide a mechanism through which you can override various files in
/etc (as well as other subdirectories of root). The scripts provide
four overriding directories situated in /conf/base, /conf/default,
/conf/<broadcast-ip>, and /conf/<machine-ip>. You should always create
/conf/base/etc, which will entirely replace the server's /etc on
the diskless machine. You can clone the server's /etc here or you
can create a special file which tells the diskless boot scripts to
remount the server's /etc onto /conf/base/etc. You do this by creating
the file /conf/base/etc/diskless_remount containing the mount
point to use as a basis of the diskless machine's /etc. For example,
the file might contain:
10.0.0.1:/etc
Alternativly, if the server contains several independent roots, the
file might contain:
10.0.0.1:/usr/diskless/4.7-RELEASE/etc
This would work, but if you copied /usr/diskless/4.7-RELEASE to
/usr/diskless/4.8-RELEASE and upgraded the installation, you would
need to modify the diskless_remount files to reflect that move. To
avoid that, paths in diskless_remount files begining with / have the
actual path of the client's root prepended to them so the file could
instead contain:
/etc
The diskless scripts create memory file systems to hold the overriden
directories. Only a 2MB partition is created by default, which may
not be sufficient for your purposes. To override this, you can create
the file /conf/base/etc/md_size containing the size, in 512 byte
sectors, of the memory disk to create for that directory.
You then typically provide file-by-file overrides in the
/conf/default/etc directory. At a minimum, you must provide overrides
for /etc/fstab, /etc/rc.conf, and /etc/rc.local via
/conf/default/etc/fstab, /conf/default/etc/rc.conf, and
/conf/default/etc/rc.local.
Overrides are hierarchical. You can supply network-specific defaults
in the /conf/<BROADCASTIP>/etc directory, where <BROADCASTIP> represents
the broadcast IP address of the diskless system as given to it
via BOOTP. The diskless_remount and md_size features work in any of
these directories. The configuration feature works on directories
other then /etc, you simply create the directory you wish to replace
or override in /conf/{base,default,<broadcast>,<ip>}/* and work it in
the same way that you work /etc.
Since you normally clone the server's /etc using the
/conf/base/etc/diskless_remount, you might wish to remove unneeded
files from the memory file system. For example, if the server has a
firewall but you do not, you might wish to remove /etc/ipfw.conf.
You can do this by creating a /conf/base/<DIRECTORY>.remove file.
For example, /conf/base/etc.remove, which contains a list of relative
paths that the boot scripts should remove from the memory file systems.
As a minimum, you normally need to have the following in
/conf/default/etc/fstab
<SERVER>:<ROOT> / nfs ro 0 0
<SERVER>:/usr /usr nfs ro 0 0
proc /proc procfs rw 0 0
You also need to create a customized version of
/conf/default/etc/rc.conf which should contain the startup options
for the diskless client, and /conf/default/etc/rc.local which could
be empty but prevents the server's own /etc/rc.local from leaking
onto the diskless system.
In rc.conf, most likely you will not need to set hostname and
ifconfig_* because these will be already set by the startup code.
Finally, it might be convenient to use a case statement using
`hostname` as the switch variable to do machine-specific configuration
in case a number of diskless clients share the same configuration
files.
+o The kernel for the diskless clients, which will be loaded using NFS
or TFTP, should be built with at least the following options:
options BOOTP
options BOOTP_NFSROOT
options BOOTP_COMPAT
In the devices section add:
device md
If you use the firewall, remember to default to ``open'', or your
kernel will not be able to send/receive the BOOTP packets.
Be warned that using unencrypted NFS to mount root and user partitions
may expose information such as encryption keys.
This manpage is probably incomplete.
FreeBSD sometimes requires to write onto the root partition, so the
startup scripts mount MFS file systems on some locations (e.g. /etc and
/var), while trying to preserve the original content. The process might
not handle all cases.
ethers(5), exports(5), bootpd(8), mountd(8), nfsd(8), pxeboot(8),
reboot(8), tftpd(8)
ports/net/etherboot
FreeBSD 5.2.1 December 23, 2002 FreeBSD 5.2.1 [ Back ] |
