University of Waikato
ian.mcdonald@jandi.co.nz
Abstract. This paper focuses on the process of development for the
Linux kernel. This topic is approached both from a technical viewpoint
and also the interactions with the open source community. The aim of
the paper is for readers not to have the same diculties the author
experienced!
Keywords: Operating Systems
1 Introduction
This paper discusses the social aspects of kernel development, followed by tips
for building a kernel. The paper then focusses on releasing code and lastly, but
not least important, testing and debugging.
The paper highlights some of the lessons learned by the author and the
methodology used in developing within the kernel.
2 Building a kernel
2.1 Maintaining a source code tree
Keeping in synchronisation with the Linux kernel source code tree takes a nontrivial
amount of time. It is necessary to keep synchronised if it is intended to
have the code merged into the tree or released as an ongoing codebase.
The Linux source code is maintained in a git [1] tree. Git is a new source
code management tool created by Linus Torvalds. Git stores every change as a
patch addressed by a SHA1 hash. Developers can make copies of git trees and
because each patch is unique it is relatively simple for developers to synchronise
their code base with other developers.
For developing in the kernel it is recommended to make a copy of Linus
Torvalds’ source code tree by issuing a statement similar to:
git-clone \
git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git \
~/linuxsrc/linus
and then using git-pull to keep the git tree up to date. If the maintainer for
the area that is being developed keeps a separate git tree, as most do, then you
will need to create a copy of it by issuing a statement similar to:
git-clone --reference ~/linuxsrc/linus \
git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6.git \
~/linuxsrc/davem
When the above syntax is used it uses the local copy where possible making
synchronisation quicker and uses a fraction of the bandwidth.
It is often useful to use a patch management tool on top of git such as stgit
[2] or quilt [3]. These tools allow a series of patches to be maintained as a “stack”
of patches which can be applied against dierent or updated trees.
Often a maintainer will “rebase” their tree which means they delete their
original tree, make a fresh clone of Linus’ tree and then they will reapply any
outstanding patches to that tree. If a maintainer rebases their tree then it is
usually not possible to update it using the git-pull or stg pull command.
Presuming the use of stgit, issue the command stg export to export the patches
(this should also be done before any pull in case of a problem), save the files that
are under patches-branch in a temporary directory, clone the git tree again to
the latest tree and then import the patches into the new tree by stg import.
2.2 Use of distcc
To speed up compilation if multiple machines are available then distcc [4] can be
used which is a distributed C compiler, and then specifying to make how many
parallel threads to run using the -j parameter. For kernel development there
are a couple of caveats to be aware of. The same version of the C compiler gcc
must be installed on the machines that are being used as a compile pool. It is
also important to specify only one target on the command line or else the kernel
will continue to rebuild from scratch each time the make command is issued.
For example to build a kernel and prepare for it to be installed the following
commands could be issued:
make -j6 CC=distcc all
make modules_install INSTALL_MOD_PATH=~/tmp
2.3 Other resources
For an introduction to Linux development there are excellent resources available
such as [5] and [6]. The use of simple tools should also not be underestimated
such as the use of grep -n -r phrase to find a symbol in the kernel. The author
also maintains a Wiki page at http://wlug.org.nz/KernelDevelopment which
contains other tips.
3 Releasing code
An often cited mantra in the open source community is “release early and release
often” [7]. This concept is useful even prior to code release.
It can productive to talk about what you are planning to implement. This
can be useful as other people can inform you if they are working on similar
code as is often the case. It also gives people an opportunity to give feedback on
whether the ideas are good or need refining.
Once the ideas have been crystallised it is often useful to release a code as a
request for comment (RFC) if the code is not ready to be merged. This, again,
allows discussion of the ideas being implemented.
The code will probably go through many, many iterations so if the code
is rejected then it is important to work on the areas highlighted rather than
becoming disheartened.
When developing code for the kernel it needs to be submitted to the maintainer
of the subsystem. The MAINTAINERS file in the top level directory of the
Linux source code tree contains the list of the maintainers. A posting to the
linux-kernel list or other kernel mailing lists will usually produce a response on
how the code is maintained if it is not clear.
Any code released should always be against the latest tree of the maintainer.
One of the biggest reasons that code isn’t accepted is that the maintainer cannot
use it as it applies against an older version of Linux.
When code is submitted it must be able to be applied without problems by
the maintainer. Prior to submitting test each patch with:
git-apply --check --whitespace=error < mypatch.diff
It is also required to choose a mail client to send the patch which does not
alter whitespace such as spaces being substituted for tabs, breaking lines up or
sending as HTML. The author uses KMail as he has found this is one of the few
that will meet these criteria.
The code and patches should also fit into the existing kernel code style. The
Documentation directory of the Linux source code contains many useful guides
to assist with this.
It is recommended for a developer to submit kernel changes in multiple small
patches rather than one large patch. This is that so the change can be reviewed
in a series of steps which are simpler to read and find bugs in. If multiple patches
are provided then part of the change can be accepted, rather than all rejected if
one large patch is used.
4 Testing and debugging
It is important that code is well tested before it is released.
There are a number of virtualisation tools available to assist with testing
software such as qemu [8], UML [9], VMWare [10] and Xen [11]. These tools can
make testing easier if computers with sucient processing power are used. Any
software developed should also be tested on computers natively as well as this
can uncover separate bugs.
There are debugging tools available that can be used on the kernel such as
gdb. The simplest debugging tool to use for timing dependent code is the printk
statement in the code which acts like a printf statement, except the output goes
to the kernel logging daemon. If it is desired to find which patch caused a bug
then git-bisect can be used to “divide and conquer” the code base.
In the kernel there is also a mechanism for tracing code execution and capturing
data called kprobes which allows hooks to be added into any function entry
or exit dynamically without any changes being made to the original function. In
our research we have taken an implementation of this for monitoring TCP and
implemented it for DCCP.
With testing it is often necessary to transfer newly built kernels to multiple
machines. This can be automated through a script such as:
#! /bin/bash
# syntax m machine_name directory version
H=$HOME
SRC=$H/linuxsrc/$2
VER=$3
rm $H/tmp/lib/modules/$VER/build
rm $H/tmp/lib/modules/$VER/source
rsync $SRC/System.map root@$1:/boot/System.map-$VER
rsync $SRC/arch/i386/boot/bzImage root@$1:/boot/vmlinuz-$VER
rsync -av $H/tmp/lib/modules/$VER root@$1:/lib/modules
5 Conclusion
Development in the Linux kernel is more than simply editing code and typing
make all. It is the hope of the authour that this paper helps more people develop
in the Linux kernel by taking into account other considerations and applying
these.
References
1. Web: git. http://git.or.cz/ (Accessed 2006)
2. Web: Stacked git. http://www.procode.org/stgit/ (Accessed 2006)
3. Web: Quilt patch management tools. http://savannah.nongnu.org/projects/quilt/
(Accessed 2006)
4. Pool, M., et al.: Distcc: a fast, free distributed c/c++ compiler, 2002–. URL
http://distcc. samba. org
5. Love, R.: Linux Kernel Development. Second edn. Novell Press (2005)
6. Web: Lxr. http://lxr.linux.no/ (Accessed 2006)
7. Raymond, E.: Cathedral & the Bazaar: Musings on Linux and Open Source by an
Accidental Revolutionary. (2001)
8. QEMU, C.: Emulator. URL: http://fabrice. bellard. free. fr/qemu
9. Dike, J.: A user-mode port of the Linux kernel. Proceedings of the Annual Linux
Showcase. Atlanta, GA, Oct (2000)
10. VMware, I.: The VMWare software package. See http://www. vmware. com
11. Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T., Ho, A., Neugebauer,
R., Pratt, I., Warfield, A.: Xen and the art of virtualization. Proceedings of the
nineteenth ACM symposium on Operating systems principles (2003) 164–177
No comments:
Post a Comment