When the kernel has been loaded by a boot loader (either by loadlin,
LILO or a network boot program) it has to be told what root fs device
to use, and where to find the server and the name of the directory
on the server to mount as root. This can be established by a couple
of kernel command line parameters:
root=/dev/nfs
This is necessary to enable the pseudo-NFS-device. Note that it's not a
real device but just a synonym to tell the kernel to use NFS instead of
a real device.
nfsroot=[:][,]
If the `nfsroot' parameter is NOT given on the command line, the default
"/tftpboot/%s" will be used.
Specifies the IP address of the NFS server. If this field
is not given, the default address as determined by the
`ip' variable (see below) is used. One use of this
parameter is for example to allow using different servers
for RARP and NFS. Usually you can leave this blank.
Name of the directory on the server to mount as root. If
there is a "%s" token in the string, the token will be
replaced by the ASCII-representation of the client's IP
address.
Standard NFS options. All options are separated by commas.
If the options field is not given, the following defaults
will be used:
port = as given by server portmap daemon
rsize = 1024
wsize = 1024
timeo = 7
retrans = 3
acregmin = 3
acregmax = 60
acdirmin = 30
acdirmax = 60
flags = hard, nointr, noposix, cto, ac
This parameter tells the kernel how to configure IP addresses of devices
and also how to set up the IP routing table. It was originally called `nfsaddrs',
but now the boot-time IP configuration works independently of NFS, so it
was renamed to `ip' and the old name remained as an alias for compatibility
reasons.
If this parameter is missing from the kernel command line, all fields are
assumed to be empty, and the defaults mentioned below apply. In general
this means that the kernel tries to configure everything using both
RARP and BOOTP (depending on what has been enabled during kernel confi-
guration, and if both what protocol answer got in first).
IP address of the client. If empty, the address will either
be determined by RARP or BOOTP. What protocol is used de-
pends on what has been enabled during kernel configuration
and on the parameter. If this parameter is not
empty, neither RARP nor BOOTP will be used.
IP address of the NFS server. If RARP is used to determine
the client address and this parameter is NOT empty only
replies from the specified server are accepted. To use
different RARP and NFS server, specify your RARP server
here (or leave it blank), and specify your NFS server in
the `nfsroot' parameter (see above). If this entry is blank
the address of the server is used which answered the RARP
or BOOTP request.
IP address of a gateway if the server is on a different
subnet. If this entry is empty no gateway is used and the
server is assumed to be on the local network, unless a
value has been received by BOOTP.
Netmask for local network interface. If this is empty,
the netmask is derived from the client IP address assuming
classful addressing, unless overridden in BOOTP reply.
Name of the client. If empty, the client IP address is
used in ASCII notation, or the value received by BOOTP.
Name of network device to use. If this is empty, all
devices are used for RARP and BOOTP requests, and the
first one we receive a reply on is configured. If you have
only one device, you can safely leave this blank.
Method to use for autoconfiguration. If this is either
'rarp' or 'bootp', the specified protocol is used.
If the value is 'both' or empty, both protocols are used
so far as they have been enabled during kernel configura-
tion. 'off' means no autoconfiguration.
The parameter can appear alone as the value to the `ip'
parameter (without all the ':' characters before) in which case auto-
configuration is used.
3.4) Using a boot ROM
This is probably the most elegant way of booting a diskless
client. With a boot ROM the kernel gets loaded using the TFTP
protocol. As far as I know, no commercial boot ROMs yet
support booting Linux over the network, but there are two
free implementations of a boot ROM available on sunsite.unc.edu
and its mirrors. They are called 'netboot-nfs' and 'etherboot'.
Both contain everything you need to boot a diskless Linux client.
When using initrd, the system typically boots as follows:
1) the boot loader loads the kernel and the initial RAM disk
2) the kernel converts initrd into a "normal" RAM disk and
frees the memory used by initrd
3) initrd is mounted read-write as root
4) /linuxrc is executed (this can be any valid executable, including
shell scripts; it is run with uid 0 and can do basically everything
init can do)
5) linuxrc mounts the "real" root file system
6) linuxrc places the root file system at the root directory using the
pivot_root system call
7) the usual boot sequence (e.g. invocation of /sbin/init) is performed
on the root file system
8) the initrd file system is removed
Note that changing the root directory does not involve unmounting it.
It is therefore possible to leave processes running on initrd during that
procedure. Also note that file systems mounted under initrd continue to
be accessible.
Boot command-line options
-------------------------
initrd adds the following new options:
Loads the specified file as the initial RAM disk. When using LILO, you
have to specify the RAM disk image file in /etc/lilo.conf, using the
INITRD configuration variable.
noinitrd
initrd data is preserved but it is not converted to a RAM disk and
the "normal" root file system is mounted. initrd data can be read
from /dev/initrd. Note that the data in initrd can have any structure
in this case and doesn't necessarily have to be a file system image.
This option is used mainly for debugging.
Note: /dev/initrd is read-only and it can only be used once. As soon
as the last process has closed it, all data is freed and /dev/initrd
can't be opened anymore.
root=/dev/ram0 (without devfs)
root=/dev/rd/0 (with devfs)
initrd is mounted as root, and the normal boot procedure is followed,
with the RAM disk still mounted as root.
Installation
------------
First, a directory for the initrd file system has to be created on the
"normal" root file system, e.g.
# mkdir /initrd
The name is not relevant. More details can be found on the pivot_root(2)
man page.
If the root file system is created during the boot procedure (i.e. if
you're building an install floppy), the root file system creation
procedure should create the /initrd directory.
If initrd will not be mounted in some cases, its content is still
accessible if the following device has been created (note that this
does not work if using devfs):
# mknod /dev/initrd b 1 250
# chmod 400 /dev/initrd
Second, the kernel has to be compiled with RAM disk support and with
support for the initial RAM disk enabled. Also, at least all components
needed to execute programs from initrd (e.g. executable format and file
system) must be compiled into the kernel.
Third, you have to create the RAM disk image. This is done by creating a
file system on a block device, copying files to it as needed, and then
copying the content of the block device to the initrd file. With recent
kernels, at least three types of devices are suitable for that:
- a floppy disk (works everywhere but it's painfully slow)
- a RAM disk (fast, but allocates physical memory)
- a loopback device (the most elegant solution)
We'll describe the loopback device method:
1) make sure loopback block devices are configured into the kernel
2) create an empty file system of the appropriate size, e.g.
# dd if=/dev/zero of=initrd bs=300k count=1
# mke2fs -F -m0 -b 1024 initrd
(if space is critical, you may want to use the Minix FS instead of Ext2)
(Note that due to a problem elsewhere in the kernel, you _must_ use a
1024-byte blocksize when creating your file system. If any other
value is used, the kernel will be unable to mount the initrd at boot
time, causing a kernel panic.)
3) mount the file system, e.g.
# mount -t ext2 -o loop initrd /mnt
4) create the console device (not necessary if using devfs, but it can't
hurt to do it anyway):
# mkdir /mnt/dev
# mknod /mnt/dev/console c 5 1
5) copy all the files that are needed to properly use the initrd
environment. Don't forget the most important file, /linuxrc
Note that /linuxrc's permissions must include "x" (execute).
6) correct operation the initrd environment can frequently be tested
even without rebooting with the command
# chroot /mnt /linuxrc
This is of course limited to initrds that do not interfere with the
general system state (e.g. by reconfiguring network interfaces,
overwriting mounted devices, trying to start already running demons,
etc. Note however that it is usually possible to use pivot_root in
such a chroot'ed initrd environment.)
7) unmount the file system
# umount /mnt
8) the initrd is now in the file "initrd". Optionally, it can now be
compressed
# gzip -9 initrd
For experimenting with initrd, you may want to take a rescue floppy and
only add a symbolic link from /linuxrc to /bin/sh. Alternatively, you
can try the experimental newlib environment [2] to create a small
initrd.
Finally, you have to boot the kernel and load initrd. Almost all Linux
boot loaders support initrd. Since the boot process is still compatible
with an older mechanism, the following boot command line parameters
have to be given:
root=/dev/ram0 init=/linuxrc rw
if not using devfs, or
root=/dev/rd/0 init=/linuxrc rw
if using devfs. (rw is only necessary if writing to the initrd file
system.)
With LOADLIN, you simply execute
LOADLIN initrd=
e.g. LOADLIN C:\LINUX\BZIMAGE initrd=C:\LINUX\INITRD.GZ root=/dev/ram0 init=/linuxrc rw
With LILO, you add the option INITRD= to either the global section
or to the section of the respective kernel in /etc/lilo.conf, and pass
the options using APPEND, e.g.
image = /bzImage
initrd = /boot/initrd.gz
append = "root=/dev/ram0 init=/linuxrc rw"
and run /sbin/lilo
For other boot loaders, please refer to the respective documentation.
Now you can boot and enjoy using initrd.
Changing the root device
------------------------
When finished with its duties, linuxrc typically changes the root device
and proceeds with starting the Linux system on the "real" root device.
The procedure involves the following steps:
- mounting the new root file system
- turning it into the root file system
- removing all accesses to the old (initrd) root file system
- unmounting the initrd file system and de-allocating the RAM disk
Mounting the new root file system is easy: it just needs to be mounted on
a directory under the current root. Example:
# mkdir /new-root
# mount -o ro /dev/hda1 /new-root
The root change is accomplished with the pivot_root system call, which
is also available via the pivot_root utility (see pivot_root(8) man
page; pivot_root is distributed with util-linux version 2.10h or higher
[3]). pivot_root moves the current root to a directory under the new
root, and puts the new root at its place. The directory for the old root
must exist before calling pivot_root. Example:
# cd /new-root
# mkdir initrd
# pivot_root . initrd
Now, the linuxrc process may still access the old root via its
executable, shared libraries, standard input/output/error, and its
current root directory. All these references are dropped by the
following command:
# exec chroot . what-follows dev/console 2>&1
Where what-follows is a program under the new root, e.g. /sbin/init
If the new root file system will be used with devfs and has no valid
/dev directory, devfs must be mounted before invoking chroot in order to
provide /dev/console.
Note: implementation details of pivot_root may change with time. In order
to ensure compatibility, the following points should be observed:
- before calling pivot_root, the current directory of the invoking
process should point to the new root directory
- use . as the first argument, and the _relative_ path of the directory
for the old root as the second argument
- a chroot program must be available under the old and the new root
- chroot to the new root afterwards
- use relative paths for dev/console in the exec command
Now, the initrd can be unmounted and the memory allocated by the RAM
disk can be freed:
# umount /initrd
# blockdev --flushbufs /dev/ram0
# /dev/rd/0 if using devfs
It is also possible to use initrd with an NFS-mounted root, see the
pivot_root(8) man page for details.
Note: if linuxrc or any program exec'ed from it terminates for some
reason, the old change_root mechanism is invoked (see section "Obsolete
root change mechanism").
Usage scenarios
---------------
The main motivation for implementing initrd was to allow for modular
kernel configuration at system installation. The procedure would work
as follows:
1) system boots from floppy or other media with a minimal kernel
(e.g. support for RAM disks, initrd, a.out, and the Ext2 FS) and
loads initrd
2) /linuxrc determines what is needed to (1) mount the "real" root FS
(i.e. device type, device drivers, file system) and (2) the
distribution media (e.g. CD-ROM, network, tape, ...). This can be
done by asking the user, by auto-probing, or by using a hybrid
approach.
3) /linuxrc loads the necessary kernel modules
4) /linuxrc creates and populates the root file system (this doesn't
have to be a very usable system yet)
5) /linuxrc invokes pivot_root to change the root file system and
execs - via chroot - a program that continues the installation
6) the boot loader is installed
7) the boot loader is configured to load an initrd with the set of
modules that was used to bring up the system (e.g. /initrd can be
modified, then unmounted, and finally, the image is written from
/dev/ram0 or /dev/rd/0 to a file)
8) now the system is bootable and additional installation tasks can be
performed
The key role of initrd here is to re-use the configuration data during
normal system operation without requiring the use of a bloated "generic"
kernel or re-compiling or re-linking the kernel.
A second scenario is for installations where Linux runs on systems with
different hardware configurations in a single administrative domain. In
such cases, it is desirable to generate only a small set of kernels
(ideally only one) and to keep the system-specific part of configuration
information as small as possible. In this case, a common initrd could be
generated with all the necessary modules. Then, only /linuxrc or a file
read by it would have to be different.
A third scenario are more convenient recovery disks, because information
like the location of the root FS partition doesn't have to be provided at
boot time, but the system loaded from initrd can invoke a user-friendly
dialog and it can also perform some sanity checks (or even some form of
auto-detection).
Last not least, CD-ROM distributors may use it for better installation
from CD, e.g. by using a boot floppy and bootstrapping a bigger RAM disk
via initrd from CD; or by booting via a loader like LOADLIN or directly
from the CD-ROM, and loading the RAM disk from CD without need of
floppies.
Obsolete root change mechanism
------------------------------
The following mechanism was used before the introduction of pivot_root.
Current kernels still support it, but you should _not_ rely on its
continued availability.
It works by mounting the "real" root device (i.e. the one set with rdev
in the kernel image or with root=... at the boot command line) as the
root file system when linuxrc exits. The initrd file system is then
unmounted, or, if it is still busy, moved to a directory /initrd, if
such a directory exists on the new root file system.
In order to use this mechanism, you do not have to specify the boot
command options root, init, or rw. (If specified, they will affect
the real root file system, not the initrd environment.)
If /proc is mounted, the "real" root device can be changed from within
linuxrc by writing the number of the new root FS device to the special
file /proc/sys/kernel/real-root-dev, e.g.
# echo 0x301 >/proc/sys/kernel/real-root-dev
Note that the mechanism is incompatible with NFS and similar file
systems.
This old, deprecated mechanism is commonly called "change_root", while
the new, supported mechanism is called "pivot_root".
Resources
---------
