linux-next/Documentation/early-userspace

Early userspace support
=======================

Last update: 2004-12-20 tlh


"Early userspace" is a set of libraries and programs that provide
various pieces of functionality that are important enough to be
available while a Linux kernel is coming up, but that don't need to be
run inside the kernel itself.

It consists of several major infrastructure components:

- gen_init_cpio, a program that builds a cpio-format archive
  containing a root filesystem image.  This archive is compressed, and
  the compressed image is linked into the kernel image.
- initramfs, a chunk of code that unpacks the compressed cpio image
  midway through the kernel boot process.
- klibc, a userspace C library, currently packaged separately, that is
  optimized for correctness and small size.

The cpio file format used by initramfs is the "newc" (aka "cpio -H newc")
format, and is documented in the file "buffer-format.txt".  There are
two ways to add an early userspace image: specify an existing cpio
archive to be used as the image or have the kernel build process build
the image from specifications.

CPIO ARCHIVE method

You can create a cpio archive that contains the early userspace image.
Your cpio archive should be specified in CONFIG_INITRAMFS_SOURCE and it
will be used directly.  Only a single cpio file may be specified in
CONFIG_INITRAMFS_SOURCE and directory and file names are not allowed in
combination with a cpio archive.

IMAGE BUILDING method

The kernel build process can also build an early userspace image from
source parts rather than supplying a cpio archive.  This method provides
a way to create images with root-owned files even though the image was
built by an unprivileged user.

The image is specified as one or more sources in
CONFIG_INITRAMFS_SOURCE.  Sources can be either directories or files -
cpio archives are *not* allowed when building from sources.

A source directory will have it and all of its contents packaged.  The
specified directory name will be mapped to '/'.  When packaging a
directory, limited user and group ID translation can be performed.
INITRAMFS_ROOT_UID can be set to a user ID that needs to be mapped to
user root (0).  INITRAMFS_ROOT_GID can be set to a group ID that needs
to be mapped to group root (0).

A source file must be directives in the format required by the
usr/gen_init_cpio utility (run 'usr/gen_init_cpio --help' to get the
file format).  The directives in the file will be passed directly to
usr/gen_init_cpio.

When a combination of directories and files are specified then the
initramfs image will be an aggregate of all of them.  In this way a user
can create a 'root-image' directory and install all files into it.
Because device-special files cannot be created by a unprivileged user,
special files can be listed in a 'root-files' file.  Both 'root-image'
and 'root-files' can be listed in CONFIG_INITRAMFS_SOURCE and a complete
early userspace image can be built by an unprivileged user.

As a technical note, when directories and files are specified, the
entire CONFIG_INITRAMFS_SOURCE is passed to
scripts/gen_initramfs_list.sh.  This means that CONFIG_INITRAMFS_SOURCE
can really be interpreted as any legal argument to
gen_initramfs_list.sh.  If a directory is specified as an argument then
the contents are scanned, uid/gid translation is performed, and
usr/gen_init_cpio file directives are output.  If a directory is
specified as an arugemnt to scripts/gen_initramfs_list.sh then the
contents of the file are simply copied to the output.  All of the output
directives from directory scanning and file contents copying are
processed by usr/gen_init_cpio.

See also 'scripts/gen_initramfs_list.sh -h'.

Where's this all leading?
=========================

The klibc distribution contains some of the necessary software to make
early userspace useful.  The klibc distribution is currently
maintained separately from the kernel, but this may change early in
the 2.7 era (it missed the boat for 2.5).

You can obtain somewhat infrequent snapshots of klibc from
ftp://ftp.kernel.org/pub/linux/libs/klibc/

For active users, you are better off using the klibc git
repository, at http://git.kernel.org/?p=libs/klibc/klibc.git

The standalone klibc distribution currently provides three components,
in addition to the klibc library:

- ipconfig, a program that configures network interfaces.  It can
  configure them statically, or use DHCP to obtain information
  dynamically (aka "IP autoconfiguration").
- nfsmount, a program that can mount an NFS filesystem.
- kinit, the "glue" that uses ipconfig and nfsmount to replace the old
  support for IP autoconfig, mount a filesystem over NFS, and continue
  system boot using that filesystem as root.

kinit is built as a single statically linked binary to save space.

Eventually, several more chunks of kernel functionality will hopefully
move to early userspace:

- Almost all of init/do_mounts* (the beginning of this is already in
  place)
- ACPI table parsing
- Insert unwieldy subsystem that doesn't really need to be in kernel
  space here

If kinit doesn't meet your current needs and you've got bytes to burn,
the klibc distribution includes a small Bourne-compatible shell (ash)
and a number of other utilities, so you can replace kinit and build
custom initramfs images that meet your needs exactly.

For questions and help, you can sign up for the early userspace
mailing list at http://www.zytor.com/mailman/listinfo/klibc

How does it work?
=================

The kernel has currently 3 ways to mount the root filesystem:

a) all required device and filesystem drivers compiled into the kernel, no
   initrd.  init/main.c:init() will call prepare_namespace() to mount the
   final root filesystem, based on the root= option and optional init= to run
   some other init binary than listed at the end of init/main.c:init().

b) some device and filesystem drivers built as modules and stored in an
   initrd.  The initrd must contain a binary '/linuxrc' which is supposed to
   load these driver modules.  It is also possible to mount the final root
   filesystem via linuxrc and use the pivot_root syscall.  The initrd is
   mounted and executed via prepare_namespace().

c) using initramfs.  The call to prepare_namespace() must be skipped.
   This means that a binary must do all the work.  Said binary can be stored
   into initramfs either via modifying usr/gen_init_cpio.c or via the new
   initrd format, an cpio archive.  It must be called "/init".  This binary
   is responsible to do all the things prepare_namespace() would do.

   To maintain backwards compatibility, the /init binary will only run if it
   comes via an initramfs cpio archive.  If this is not the case,
   init/main.c:init() will run prepare_namespace() to mount the final root
   and exec one of the predefined init binaries.

Bryan O'Sullivan <bos@serpentine.com>