Linux Tiny-FAQ

From eLinux.org
Revision as of 03:30, 27 October 2011 by Peter Huewe (talk | contribs) (Credits: consistent formatting)
Jump to: navigation, search

FAQ: The Linux-Tiny Tree

This FAQ was originally written as a contribution to the Linux-Tiny effort which aims to make it possible to create a lean-and-mean Linux kernel. This document is a copy of the original linux-tiny FAQ. It is intended to evolve into a more generic Embedded Linux FAQ over time.

A special note should be made here about Matt Mackall <mpm at selenic.com> who originally put linux-tiny together and made it into something that others could download.


Credits

Contributors:

   * Mahendra <Mahendra_M at infosys.com>
   * Matt Mackall <mpm at selenic.com>
   * Rob Landley <rob at landley.net>
   * Tim Bird <tim.bird at am.sony.com>

Author:

* Onno Benschop <onno at itmaze.com.au>

What is this all about?

This project is a collection of patches to the Linux kernel to remove different aspects of the kernel in order to save space. For example, removing all the error messages from the kernel removes about 300k from it.

A kernel uses space in all manner of different ways, the space it takes up on disk, the space it takes to load the code into memory and the space it takes to store information while it is running. Linux-Tiny is a set of patches to the mainline kernel (at kernel.org) that addresses the way the kernel uses space. The patches you'll find here are intended to be included into the main kernel-tree and some have already been submitted for such use.

This is useful in places where space is limited, or in the case of error messages, there is no-one to see them - like in embedded systems. Linux-Tiny is also for users of small or legacy machines such as 386's and handhelds.

A description on the how and why is available here: Linux_Tiny

and here is a really detailed description:

http://www.selenic.com/tiny-about/Reprint-Mackall-OLS2004.pdf

How do I use it?

This project will not supply you with an actual kernel, it gives you a set of patches that you need to apply to a kernel source tree, after which you can configure and compile your new trimmed kernel.

The Linux-Tiny patches default to the standard kernel configuration, so you'll need to turn on or off those settings you want to change.

The process to apply Linux-Tiny is pretty straightforward if you've patched a kernel before, but if you haven't there are a few steps you'll need to take: (and you should read the question "Will it work on any kernel?")

  1. Download kernel sources
  2. Download the appropriate patch set for that kernel
  3. cd /path/to/my/linux/sources/
  4. patch -p1 </path/to/patch/Linux-Tiny.patch
  5. Check for reject files and fix where appropriate
  6. Configure the kernel
  7. Compile the kernel


What Difference does it make?

Well, that depends entirely on what you're trying to achieve. Different patches and different ideas in size reduction are suited to different purposes. The various size reductions have varying impacts on different aspects of the kernel. Efforts are being made to document and compare sizes.


Config Size Test

Tim Bird has written an program to automatically test the kernel size under different configurations. As an example, this table was produced for a test run conducted on Oct 1st, 2007, on an i386 target (called 'nut'): http://testlab.celinuxforum.org/otlwiki/ConfigSizeTestResultsNutOct1

One of the main purposes of this program is to find problems with the Linux-tiny patches. Also, this data could be used to help identify the options having the largest effect on the kernel size.


Bloat Watch

Matt Maccall wrote bloatwatch at CELF 2006, and the data that was gathered then is still available, at: http://testlab.celinuxforum.org/bloatwatch/index.cgi

It provides a great amount of detail, and the ability to compare versions over time. As such, it really was intended to be a kernel size regression analysis tool.


What is the difference between the broken-out and the patch files?

The broken out patches contain all the patches as individual patches, so you can choose to apply one or all of them, where the patch file contains all the patches as one big patch, applied mostly in the intended merge order from simple to complex.

For example, one set of patches turns off all error messages. Another set of patches changes the way memory is allocated. The broken out set contains the two different patches as separate entities, the complete patch contains both of the changes together.

The reason it is like that, is because not everyone will want to apply all patches, or will be able to apply all patches - see below.


Will it work on any kernel?

Yes.

Seriously, the answer is in typical computing fashion: "That depends"

If you're new to patching, maybe a brief explanation is needed. Pardon me if this is too remedial, but I wanted to write low level since I don't know your experience. Sorry in advance for the length.

Patches are created by diff'ing two source bases against each other. A patch is like a set of instructions for converting one source base into the other. Inside the patch are both the changes themselves, and context information which helps patch to identify the correct place to make each change.

If you try to apply a patch to a different version of software than it was originally created from, you may encounter problems. Patch may not find the correct place to make a change. When this happens, it leaves a reject file, showing the change that it could not make.

The Linux-Tiny patches were created based on a kernel.org version of Linux. Almost all kernels shipped by the major Linux distribution vendors are significantly different from the kernel.org version of Linux. Distribution vendors add many (often hundreds) of patches to their kernels, in order to provide extra features to their products. This includes drivers, protocols, file systems, and other enhancements.

Since Linux-Tiny is a fairly big patch-set, it is very likely that there will be overlap between some of the changes it makes and some of the changes made by your Linux distribution.

If you get rejects while patching, it means that not all of the patch could apply. It is possible that the rejects don't matter, but you can't know that without looking at them. Basically, any time you get rejects you need to examine them and either 1) fix them or 2) decide they can be ignored, before proceeding.

Many times, rejects can be fixed pretty easily. A common cause of rejects is multiple additions to the beginning or end of something. Often, in these cases, the changes themselves don't really interfere with each other. But the change in text from one patch causes the patch program to be unable to match the context for a change from another patch.

For example, take the following text and patches. Starting with a simple text file describing a fish, there are two patches, one of which adds stuff about a dog and one adds stuff about a cat. Semantically, these patches don't interfere with each other, and there is no harm in applying both changes to the file. However, patch has problems with them.

file 'A':

I have a fish, named Charlie.
He swims in a fishbowl.
He doesn't eat much.

file '1.patch':

--- A   2004-10-21 10:49:53.547578239 -0700
+++ B   2004-10-21 10:50:40.395525796 -0700
@@ -1,3 +1,5 @@
 I have a fish, named Charlie.
 He swims in a fishbowl.
 He doesn't eat much.
+My dog is named Spot.
+Spot is friendly and wags his tail.

file '2.patch':

--- A   2004-10-21 10:49:53.547578239 -0700
+++ C   2004-10-21 10:51:05.435687319 -0700
@@ -1,3 +1,5 @@
 I have a fish, named Charlie.
 He swims in a fishbowl.
 He doesn't eat much.
+My cat is named cleo.
+Cleo scratches the couch.

Each of these patches could be applied successfully to file A, individually. However, if you try to apply these patches in sequence, like so:

cp A D   
patch D <1.patch   
patch D <2.patch

You'll get a reject on the second patch. The place where the description of the cat is supposed to be is now different, and patch gives up.

To fix this, you need to add the rejected changes manually, taking into account the differences caused by the other changes. (Think of 1.patch as your Linux distribution changes and 2.patch as the Linux-Tiny patch set.)

Doing this is usually pretty easy. You look at the rejected hunks, and compare the lines they intended to patch from the original file with the lines in your source base. Compare a kernel.org version of Linux, with your distribution version of Linux, for the file and lines mentioned in the reject file. In the example above, I would have to decide if I wanted the description of the dog to come before or after the description of the cat. I would make the change manually. (If this were source, I would then recompile and test extensively... :-)

For extra credit, once you have made your changes, you can create a new Linux-Tiny patch set by diff'ing your distribution-tiny with your original distribution (which you should have saved off earlier). Your distribution users may appreciate having such a "personal-tiny" patch set to work with.


How do I get access to this information?

Access has been created to a repository of the Linux-tiny patches using the Mercurial SCM:

   * http://selenic.com/repo/tiny

This has all of the old -tiny releases nicely tagged and new changes will show up here first.

Assuming you have Mercurial and quilt installed, grabbing the latest -tiny work is as simple as:

cd linux
hg clone http://selenic.com/repo/tiny patches
quilt push -a


This will hopefully also make it easier for people to keep in sync with Matt Mackall's changes and send him fixes, and make it easier for him to track changes in the upstream kernel.

More information about Mercurial here:

   * http://selenic.com/mercurial/wiki/


Is there more that I can read?

Tim Bird has done some work with Linux-Tiny on an ARM OSK board. He has written some raw notes on the subject.

Thomas Lundquist has done some work using lzma rather than bzip to do further compression and a diff for a 2.4 Kernel is available.

Ming-Ching Tiew has completed some 2.4 and 2.6 patches for lzma. Both vmlinuz and init are available. Patches for busybox are included as well as a utility called lzmacat. (Note, this requires the LZMA SDK.)

Some work is being done with LinuxBIOS and Linux-Tiny.