KernelMaintenance

Differences between revisions 2 and 3
⇤ ← Revision 2 as of 2007-04-21 15:05:58 →
Size: 15923
Editor: collinsap1
Comment: Updates for gutsy build system.
← Revision 3 as of 2007-05-10 12:58:05 → ⇥
Size: 16149
Editor: 195
Comment: Add comments to security section.
Deletions are marked like this. Additions are marked like this.
Line 226: Line 226:
If possible (especially for critical code paths), attempt to reproduce the bug via supplied reproducers (usually from the security team) to verify it actually fixes the bug. Make sure that there are no regressions as well.

Include(KernelTeamHeader)

This document is geared towards gutsy and later releases. For feisty, please see KernelMaintenanceFeisty.

This is meant as a reference for the kernel team.

Getting the source

Ubuntu kernel source is available via git. More information can be obtained from KernelGitGuide.

Build system breakdown

This requires some knowledge of the Debian build system (including the debian/rules, etc). The gutsy build system was rewritten from the ground up, to mainly exclude kernel-package, since it was getting too heavy and restrictive for the kernels builds. We still use kernel-wedge for udeb creation.

Flavours

Flavours are how we define different types of kernels on the same architecture. Each flavour can be targeted to a specific CPU, or provide some functionality for a specific purpose.

For example, the amd64 (x86_64) architecture has these flavours:

  • generic - The default install kernel.
  • server - Optimized for server installations.

When building on the amd64 architecture, these are the kernels that are built. In reality, they only differ in slight variations of kernel config options.

Configs

Each flavour is built using a pre-defined configuration file (used by the kernel build system). Each architecture contains it's config files in debian/config/<arch>/. The files are split for each flavour in debian/config/<arch>/config.<flavour>, and common options are shared in debian/config/<arch>/config.

When the kernels are built, the common and flavour specific config files are concatenated together.

debian/control

The debian/control file is generated from debian/control.stub by the kernel-wedge command.

The debian/control.stub file is generated from debian/control.stub.in by a debian/rules target. This is populated by files defining the flavours in debian/control.d/vars.*.

The vars files define package descriptions for each flavour.

debian/changelog

The changelog file is where we track all changes made to the repository. This is a standard debian/changelog. However, changes are not added to it manually. Instead, it is done automatically at release time using git-log. See below for process.

ABI

The ABI is probably one of the most important features of the build system. In the kernel package naming, the ABI is the numerical component immediately following the kernel version. E.g. 

2.6.20-6-generic

The ABI in the above example is 6. The ABI is represented in the debian/changelog as: 

linux-source-2.6.20 (2.6.22-1.3) feisty; urgency=low

The minor version after the ABI is an ever increasing counter. It basically denotes how many uploads have been made for this kernel version.

The kernel ABI is derived from the Modules.symver file produced during the kernel build. Each exported function is given a hash defining the function's arguments and return value, making it easy to detect changes which can break modules built against the kernel (such as our own linux-restricted-modules package).

The ABI files for the previous build are located in debian/abi/<prev-rev>/. The layout is similar to the debian/config/ directory in that each architecture is a subdirectory, and each flavour is a file with the ABI for that flavour in it.

When the kernel build process finishes, it checks each flavour against the prev ABI. If the previous and current ABI's are the same version (e.g. 6), and the ABI has changed, the build will fail. This signifies that an ABI bump is needed (see below).

As of now, a single ABI change in any flavour results in an ABI bump for the entire build. There is no per-flavour or per-architecture ABI number. This is for sanity reasons.

Module lists

Module lists are tracked similarly to the ABI, and are kept in debian/abi/<prev-ver>/<arch>/<flavour>.modules. The file is a basename sorted list, sans .ko extension, of all the modules in a particular flavour. This is checked regardless of ABI, since it's meant to track release-to-release to avoid modules going missing by mistake.

The check will honor debian/abi/<prev-ver>/<arch>/ignore to ignore an entire architecture, or debian/abi/<pre-ver>/<arch>/<flavour>.ignore for a specific kernel flavour.

Development cycle

This section describes all the steps involved in a common development cycles. That is, the cycles from one upload (release) of the Ubuntu kernel to the next.

Updating ABI and modules lists

To start a new release, you need to download the ABI and module files from the previous release. To do this, use: 

debian/bin/misc/getabis 2.6.22 6.11

This will download all the previous deb's, and extract the ABI files, and module lists. You may find it easier to do this from a DC machine where the archive is closer too (downloading all the kernel deb's is a bandwidth eater).

Starting a new release

Immediately after an upload you should run: 

debian/rules startnewrelease

This will generate a new debian/changelog entry for the next revision number, and modify appropriate other files under the debian/ directory.

Bumping the ABI

If it becomes necessary to increase the ABI, you simply need to edit debian/changelog, increase the ABI in the version, then run this command: 

debian/rules debian/control.stub

This will regenerate all of the control files with the appropriate ABI.

Commit templates

In debian/commit-templates/ are files used for making git-commit's easier.

You would use these files in this way: 

git-commit -s -F debian/commit-templates/<template> -e

Always commit using -s to add your Signed-off-by to the commit log. The -e flag says to edit the log before committing. Some templates do not require editing (e.g. the bumpabi and updateconfigs templates).

The primary one you will use is the patch template. It is commented heavily, so should be self explanatory.

Upstream syncing

During an Ubuntu development cycles, the kernel is constantly synced against upstream Linux GIT head. This involves merging conflicts, possibly deprecating local patches, and performing test builds.

For stable releases, we cherry pick and sometimes outright sync to the 2.6.x.y GIT tree for that particular release.

Third party drivers

In prior releases, third party drivers were kept as part of the main kernel tree in the ubuntu/ directory. This has changed in gutsy. Third party drivers are now separate, and in the linux-ubuntu-modules package.

Updating configs

As new drivers are added and merges are done with upstream, it becomes necessary to update our configs in debian/config/. This is easily done for all architectures with one simple command: 

debian/rules updateconfigs

This runs silentoldconfig for all flavours, so if any possible config options unanswered, it will ask you. Alternatively, if you know of a new config option, you can alleviate answering the same question for it on all flavours by prepending it to the debian/config/<arch>/config files, and running the above command.

Once the configs are updated, you simply commit the new changes: 

git-update-index debian/config/*/*
git-commit -s -F debian/commit-templates/update-configs

Preparing an upload

Once the tree is ready for upload, follow these steps to complete the package for uploading:

  • Run git-status to be sure you have no outstanding commits, or extraneous files in your local tree.

  • Create the changelog entries, with the following commands: 
     debian/rules insertchanges
 <open edit debian/changelog in editor to verify release and date>
 git-update-index debian/changelog
 git-commit -s -F debian/commit-templates/updateconfigs

  • Tag the release. The -m option is the message for the tag commit, and matches the actual tag name. 

     git-tag -s -m Ubuntu-2.6.20-6.11 Ubuntu-2.6.20-6.11
  • Push the changes and tag to our repo.

  • Run dpkg-buildpackage. NOTE: Requires latest dpkg-dev so that it ignores GIT meta-data. 

     dpkg-buildpackage -S -rfakeroot

  • Upload using whatever tools you wish (dupload for example).

It should be noted that for development releases, a new tarball is generated at each upload. However at release time, a .orig.tar.gz is generated, sans debian/ directory (but with debian/firmware/, to more easily track changes after release.

Pushing to our repo

The kernel team should push changes to our internal repo (see KernelGitGuide for public repo). To make this easy, you can setup a script called push-gutsy with this: 

branches="master"
remotes="rookery.ubuntu.com:/srv/kernel-team/private/ubuntu-gutsy.git"

cd /path/to/your/repo

for repo in $remotes; do
        echo "Pushing to: $repo"
        git-send-pack $repo $branches $@
        echo
done

For normal pushes, just call the script, but if you need to push a tag, add the tag to the command line, e.g.: 

push-gutsy Ubuntu-2.6.22-1.3

Security and Proposed Updates

On occasion, security patches are applied to stable kernels. For these kernels, you will set the archive in debian/changelog to <rel>-security, where <rel> is some release like edgy.

The security team generally provides patches in the form of CVE's, with GIT SHA URL's to an upstream patch that fixes the problem. We usually use will fetch from the upstream GIT repo, and git-cherry-pick to get the fix into the local tree.

The easiest way to fetch the remote repo's objects is to create a remote file such as .git/remotes/upstream-linux-2.6.20, containing: 

URL: ssh://master.kernel.org/pub/scm/linux/kernel/git/stable/linux-2.6.20.y.git
Pull: master:upstream-linux-2.6.20

Then to pull the objects, and cherry pick an SHA, do this: 

git-fetch upstream-linux-2.6.20
git-cherry-pick -e -x aa1812b3e344dbb2c7f30e1b95661cbd36d0cd2b

The -e option will bring up an editor to add any Ubuntu related info to the commit (perhaps a bug number, or maybe even add the CVE number). The -x will keep the original GIT SHA in the commit message to track where it came from.

Security uploads should be build tested prior to upload (this is the case for any upload, but especially so for security and proposed updates). You and one other person should boot test the resulting kernel.

If possible (especially for critical code paths), attempt to reproduce the bug via supplied reproducers (usually from the security team) to verify it actually fixes the bug. Make sure that there are no regressions as well.

After upload, notify the security team that you have made the upload and include a copy of the changelog.

Performing builds

Aside from the standard dpkg-buildpackage build method, there are several easier ways to do build for developers. When I actually add new code, I generally do an out-of-tree build for test, like so: 

$ cd ubuntu-gutsy
$ mkdir ../build
$ cat debian/config/i386/config{,.generic} > ../build/.config
$ make O=`pwd`/../build oldconfig
$ make O=`pwd`/../build

You can add -jX as needed.

One other helpful tool for a developer is to install the ccache package. This speeds rebuild by a large factor.

Normal full build

To test normal package building, one can use this command from the kernel git repo: 

fakeroot debian/rules binary-arch

This will build every kernel flavour.

Partial build

You can also do single or multiple specific flavours during the build, using this syntax: 

fakeroot debian/rules binary-generic

So you can change generic with any flavour you want to build.

Build options

There are several build options, in the form of environment variables and make variables. Make options are passed on the debian/rules command line.

  • Environment

    • CONCURRENCY_LEVEL=X where X is a number to pass using make's -j option.

    • AUTOBUILD - (can be passed to make as well). Tells the build to create a non-official package with generated ABI. Used for daily builds of the kernel.

  • Make

    • skipabi - Used for test only, causes the build to skip the ABI consistency check.

    • skipmodule - Same as above, but for the module check.

    • NOKERNLOG - Used for the insertchanges or printchanges target. Causes git-ubuntu-log to not include upstream kernel changes in changelog output.

Post upload

There are several things to do after an upload. These depend greatly on what the upload involved. Usually, for a non-ABI changing upload, you do not need to do anything further.

However, for an ABI changing upload, packages need to be rebuilt against the linux-headers. Note that these packages have a extra version who's major matches the ABI of the kernel it is being built against.

linux-restricted-modules

This package contains modules that are proprietary. They are distributed separately so that it is easy for users to not have to use them.

For an ABI bump, edit debian/rules and find abi_version. Change the value to the new ABI number. Use dpkg-buildpackage as normal after creating a new changelog entry.

linux-ubuntu-modules

Same as with linux-restricted-modules.

linux-meta

The linux-meta package provides meta packages for easier user upgrades. Because our kernel packages are named with full ABI and flavour, users normally would have to manually upgrade to a new kernel (if it had an ABI bump). The meta packages always point to the latest kernel, so if users install these packages (and they are installed by default on new systems), then upgrades will always pull in everything they need to be current.

For an ABI bump, edit debian/rules and find KERNEL_ABI. Do the same as for linux-restricted-modules.

linux-backports-modules

This package is usually empty during development. It is meant for post-release updates of new drivers. It is packages separately from the kernel so that it is easier to provide sometimes untested drivers for new hardware on a stable release. Users do not have this package installed by default.

The ABI bump method is the same as for linux-meta.

main, proposed and security, OH MY!

We have some differing processes for different types of uploads. Described here for the sanity challenged.

main

No differences for this. Uploading to main (the current development release) is the primary target for work.

proposed

Technically known as <rel>-proposed (e.g. edgy-proposed) is where we can push changes that need testing, and are unsuited for a wide audience of users who expect a rock solid distribution.

Typical changes in this tree are new drivers, and large updates to support new hardware.

The GIT tree is named ubuntu-<rel>-updates.git. It's ABI remains a constant -50, and the changelog for this release never rolls, meaning the it only has one changelog set for every upload. This changelog set is updated every upload to reflect new patches, and to remove ones that are either broken, or have been merged to the stable kernel tree.

See the KernelUpdates page for more information.

security

This is self explanatory, or so it would seem. While the ubuntu-<rel>.git tree is meant primarily for security updates, occasionally stable changes are merged into it from proposed.

CategoryKernel

KernelTeam/KernelMaintenance (last edited 2012-02-22 17:20:37 by eth0)