kernel_optimize_test/Documentation/filesystems/ramfs-rootfs-initramfs.txt

ramfs, rootfs and initramfs
October 17, 2005
Rob Landley <rob@landley.net>
=============================

What is ramfs?
--------------

Ramfs is a very simple filesystem that exports Linux's disk caching
mechanisms (the page cache and dentry cache) as a dynamically resizable
ram-based filesystem.

Normally all files are cached in memory by Linux.  Pages of data read from
backing store (usually the block device the filesystem is mounted on) are kept
around in case it's needed again, but marked as clean (freeable) in case the
Virtual Memory system needs the memory for something else.  Similarly, data
written to files is marked clean as soon as it has been written to backing
store, but kept around for caching purposes until the VM reallocates the
memory.  A similar mechanism (the dentry cache) greatly speeds up access to
directories.

With ramfs, there is no backing store.  Files written into ramfs allocate
dentries and page cache as usual, but there's nowhere to write them to.
This means the pages are never marked clean, so they can't be freed by the
VM when it's looking to recycle memory.

The amount of code required to implement ramfs is tiny, because all the
work is done by the existing Linux caching infrastructure.  Basically,
you're mounting the disk cache as a filesystem.  Because of this, ramfs is not
an optional component removable via menuconfig, since there would be negligible
space savings.

ramfs and ramdisk:
------------------

The older "ram disk" mechanism created a synthetic block device out of
an area of ram and used it as backing store for a filesystem.  This block
device was of fixed size, so the filesystem mounted on it was of fixed
size.  Using a ram disk also required unnecessarily copying memory from the
fake block device into the page cache (and copying changes back out), as well
as creating and destroying dentries.  Plus it needed a filesystem driver
(such as ext2) to format and interpret this data.

Compared to ramfs, this wastes memory (and memory bus bandwidth), creates
unnecessary work for the CPU, and pollutes the CPU caches.  (There are tricks
to avoid this copying by playing with the page tables, but they're unpleasantly
complicated and turn out to be about as expensive as the copying anyway.)
More to the point, all the work ramfs is doing has to happen _anyway_,
since all file access goes through the page and dentry caches.  The ram
disk is simply unnecessary, ramfs is internally much simpler.

Another reason ramdisks are semi-obsolete is that the introduction of
loopback devices offered a more flexible and convenient way to create
synthetic block devices, now from files instead of from chunks of memory.
See losetup (8) for details.

ramfs and tmpfs:
----------------

One downside of ramfs is you can keep writing data into it until you fill
up all memory, and the VM can't free it because the VM thinks that files
should get written to backing store (rather than swap space), but ramfs hasn't
got any backing store.  Because of this, only root (or a trusted user) should
be allowed write access to a ramfs mount.

A ramfs derivative called tmpfs was created to add size limits, and the ability
to write the data to swap space.  Normal users can be allowed write access to
tmpfs mounts.  See Documentation/filesystems/tmpfs.txt for more information.

What is rootfs?
---------------

Rootfs is a special instance of ramfs, which is always present in 2.6 systems.
(It's used internally as the starting and stopping point for searches of the
kernel's doubly-linked list of mount points.)

Most systems just mount another filesystem over it and ignore it.  The
amount of space an empty instance of ramfs takes up is tiny.

What is initramfs?
------------------

All 2.6 Linux kernels contain a gzipped "cpio" format archive, which is
extracted into rootfs when the kernel boots up.  After extracting, the kernel
checks to see if rootfs contains a file "init", and if so it executes it as PID
1.  If found, this init process is responsible for bringing the system the
rest of the way up, including locating and mounting the real root device (if
any).  If rootfs does not contain an init program after the embedded cpio
archive is extracted into it, the kernel will fall through to the older code
to locate and mount a root partition, then exec some variant of /sbin/init
out of that.

All this differs from the old initrd in several ways:

  - The old initrd was a separate file, while the initramfs archive is linked
    into the linux kernel image.  (The directory linux-*/usr is devoted to
    generating this archive during the build.)

  - The old initrd file was a gzipped filesystem image (in some file format,
    such as ext2, that had to be built into the kernel), while the new
    initramfs archive is a gzipped cpio archive (like tar only simpler,
    see cpio(1) and Documentation/early-userspace/buffer-format.txt).

  - The program run by the old initrd (which was called /initrd, not /init) did
    some setup and then returned to the kernel, while the init program from
    initramfs is not expected to return to the kernel.  (If /init needs to hand
    off control it can overmount / with a new root device and exec another init
    program.  See the switch_root utility, below.)

  - When switching another root device, initrd would pivot_root and then
    umount the ramdisk.  But initramfs is rootfs: you can neither pivot_root
    rootfs, nor unmount it.  Instead delete everything out of rootfs to
    free up the space (find -xdev / -exec rm '{}' ';'), overmount rootfs
    with the new root (cd /newmount; mount --move . /; chroot .), attach
    stdin/stdout/stderr to the new /dev/console, and exec the new init.

    Since this is a remarkably persnickity process (and involves deleting
    commands before you can run them), the klibc package introduced a helper
    program (utils/run_init.c) to do all this for you.  Most other packages
    (such as busybox) have named this command "switch_root".

Populating initramfs:
---------------------

The 2.6 kernel build process always creates a gzipped cpio format initramfs
archive and links it into the resulting kernel binary.  By default, this
archive is empty (consuming 134 bytes on x86).  The config option
CONFIG_INITRAMFS_SOURCE (for some reason buried under devices->block devices
in menuconfig, and living in usr/Kconfig) can be used to specify a source for
the initramfs archive, which will automatically be incorporated into the
resulting binary.  This option can point to an existing gzipped cpio archive, a
directory containing files to be archived, or a text file specification such
as the following example:

  dir /dev 755 0 0
  nod /dev/console 644 0 0 c 5 1
  nod /dev/loop0 644 0 0 b 7 0
  dir /bin 755 1000 1000
  slink /bin/sh busybox 777 0 0
  file /bin/busybox initramfs/busybox 755 0 0
  dir /proc 755 0 0
  dir /sys 755 0 0
  dir /mnt 755 0 0
  file /init initramfs/init.sh 755 0 0

Run "usr/gen_init_cpio" (after the kernel build) to get a usage message
documenting the above file format.

One advantage of the text file is that root access is not required to
set permissions or create device nodes in the new archive.  (Note that those
two example "file" entries expect to find files named "init.sh" and "busybox" in
a directory called "initramfs", under the linux-2.6.* directory.  See
Documentation/early-userspace/README for more details.)

The kernel does not depend on external cpio tools, gen_init_cpio is created
from usr/gen_init_cpio.c which is entirely self-contained, and the kernel's
boot-time extractor is also (obviously) self-contained.  However, if you _do_
happen to have cpio installed, the following command line can extract the
generated cpio image back into its component files:

  cpio -i -d -H newc -F initramfs_data.cpio --no-absolute-filenames

Contents of initramfs:
----------------------

If you don't already understand what shared libraries, devices, and paths
you need to get a minimal root filesystem up and running, here are some
references:
http://www.tldp.org/HOWTO/Bootdisk-HOWTO/
http://www.tldp.org/HOWTO/From-PowerUp-To-Bash-Prompt-HOWTO.html
http://www.linuxfromscratch.org/lfs/view/stable/

The "klibc" package (http://www.kernel.org/pub/linux/libs/klibc) is
designed to be a tiny C library to statically link early userspace
code against, along with some related utilities.  It is BSD licensed.

I use uClibc (http://www.uclibc.org) and busybox (http://www.busybox.net)
myself.  These are LGPL and GPL, respectively.  (A self-contained initramfs
package is planned for the busybox 1.2 release.)

In theory you could use glibc, but that's not well suited for small embedded
uses like this.  (A "hello world" program statically linked against glibc is
over 400k.  With uClibc it's 7k.  Also note that glibc dlopens libnss to do
name lookups, even when otherwise statically linked.)

Why cpio rather than tar?
-------------------------

This decision was made back in December, 2001.  The discussion started here:

  http://www.uwsg.iu.edu/hypermail/linux/kernel/0112.2/1538.html

And spawned a second thread (specifically on tar vs cpio), starting here:

  http://www.uwsg.iu.edu/hypermail/linux/kernel/0112.2/1587.html

The quick and dirty summary version (which is no substitute for reading
the above threads) is:

1) cpio is a standard.  It's decades old (from the AT&T days), and already
   widely used on Linux (inside RPM, Red Hat's device driver disks).  Here's
   a Linux Journal article about it from 1996:

      http://www.linuxjournal.com/article/1213

   It's not as popular as tar because the traditional cpio command line tools
   require _truly_hideous_ command line arguments.  But that says nothing
   either way about the archive format, and there are alternative tools,
   such as:

     http://freshmeat.net/projects/afio/

2) The cpio archive format chosen by the kernel is simpler and cleaner (and
   thus easier to create and parse) than any of the (literally dozens of)
   various tar archive formats.  The complete initramfs archive format is
   explained in buffer-format.txt, created in usr/gen_init_cpio.c, and
   extracted in init/initramfs.c.  All three together come to less than 26k
   total of human-readable text.

3) The GNU project standardizing on tar is approximately as relevant as
   Windows standardizing on zip.  Linux is not part of either, and is free
   to make its own technical decisions.

4) Since this is a kernel internal format, it could easily have been
   something brand new.  The kernel provides its own tools to create and
   extract this format anyway.  Using an existing standard was preferable,
   but not essential.

5) Al Viro made the decision (quote: "tar is ugly as hell and not going to be
   supported on the kernel side"):

      http://www.uwsg.iu.edu/hypermail/linux/kernel/0112.2/1540.html

   explained his reasoning:

      http://www.uwsg.iu.edu/hypermail/linux/kernel/0112.2/1550.html
      http://www.uwsg.iu.edu/hypermail/linux/kernel/0112.2/1638.html

   and, most importantly, designed and implemented the initramfs code.

Future directions:
------------------

Today (2.6.14), initramfs is always compiled in, but not always used.  The
kernel falls back to legacy boot code that is reached only if initramfs does
not contain an /init program.  The fallback is legacy code, there to ensure a
smooth transition and allowing early boot functionality to gradually move to
"early userspace" (I.E. initramfs).

The move to early userspace is necessary because finding and mounting the real
root device is complex.  Root partitions can span multiple devices (raid or
separate journal).  They can be out on the network (requiring dhcp, setting a
specific mac address, logging into a server, etc).  They can live on removable
media, with dynamically allocated major/minor numbers and persistent naming
issues requiring a full udev implementation to sort out.  They can be
compressed, encrypted, copy-on-write, loopback mounted, strangely partitioned,
and so on.

This kind of complexity (which inevitably includes policy) is rightly handled
in userspace.  Both klibc and busybox/uClibc are working on simple initramfs
packages to drop into a kernel build, and when standard solutions are ready
and widely deployed, the kernel's legacy early boot code will become obsolete
and a candidate for the feature removal schedule.

But that's a while off yet.