Revision 19 as of 2012-01-23 23:30:02

Clear message
Edit Info Other
Login
Packaging/KernelModules/

Kmods2

Packaging of kernel modules with the kmod-Standard 2

Introduction

Kernel modules are a special case in rpm packaging and need careful handling. There are a lot of ways to package kernel modules -- to avoid confusion for the users and reviewers as well as to make it easier for RPM depsolvers to support kernel-modules the Fedora (Extras|Engineering) Steering Commitee (FESCo) in 2005/2006 worked out the kmod concept for packaging kernel-modules in Fedora. By mid 2007 packages with kernel modules (like kmods or dkms) were completely banned from Fedora. Livna developers (which started shipping kmod packages soon after FESCo developed the concept) at that point picked up the kmod standard and improved it for Fedora 8. For Fedora 9 it was later a improved a bit more to provide let standard kmod srcs build a akmod rpm, that together with the akmods scripts provides a dkms-like functionality.

This document describes the enhanced version which is also known as kmod 2.x. The description of the first version can be found in this wiki as well. The initial version of this document further is based on that document, thus you can use the wiki-internal diff tools to see the differences by comparing the first version with the latest one. You can also read the section that explains the main differences roughly.

Split

There are always at least two SRPMS when a kernel-module gets packaged -- one builds a userland package from the source while the other builds packages with *only* the kernel-module(s) in it. That way new kernel-modules (build for newer kernels or modified because new patches were needed) can be shipped without shipping new userland packages, which avoids unnecessary downloads and updates for our users.

userland package

The binary packages built from the userland SRPM contains tools, documentation, license files, udev configuration etc. There always has to exist such a package -- even if the packaged software only builds kernel-modules it has at least some docs and a license file that need to be packaged in the userland package.

The packager is free to split the userland-package further into those with the general userland parts, that works fine without the kernel-modules (docs, general tools, devel-files), and one with the kernel-module related parts (udev rules for example).

The userland packages need to follow the usual Fedora Packaging Guidelines. Two additional rules MUST be followed for packages with parts related to the kernel-module(s):

  1. MUST: The package must tie itself to the kernel-module using 'Requires: %{name}-kmod  >= %{version}'

  2. MUST: The package must provide %{name}-kmod-common using 'Provides: %{name}-kmod-common = %{version}'; it is possible to name the package directly %{name}-kmod-common, but it should be avoided, as that's meaningless for users and confusing.

kernel-module package

Besides the normal packaging rules there are several additional rules for the package that contains the modules, which is called kmod package from now on. Instead of writing all those down a specfile template was created.

That template (and thus all kmod packages) make use of a bash script called kmodtool, which normally should be in a package called kmodtool that gets provided by the repo. There are further buildsys-build-<repo>-{newest,current} packages needed, which is used when the kmods get build in the buildsys of the repo; those packages holds the information for which kernel versions the kmods get built.

All kmod packages should use the template(see below) as a base. Reviewers of kmod packages should diff the proposed packages against the template. Normally only the names and the way the modules itself are built should differ. There shouldn't be other differences without a good reason.

kmod template

# (un)define the next line to either build for the newest or all current kernels
%define buildforkernels newest
#define buildforkernels current
#define buildforkernels akmod

# name should have a -kmod suffix
Name:           

Version:        
Release:        1%{?dist}.1
Summary:        Kernel module(s)

Group:          System Environment/Kernel

License:        
URL:            
Source0:        
BuildRoot:      %{_tmppath}/%{name}-%{version}-%{release}-root-%(%{__id_u} -n)

BuildRequires:  %{_bindir}/kmodtool


# needed for plague to make sure it builds for i586 and i686
ExclusiveArch:  i586 i686 x86_64 ppc ppc64

# get the proper build-sysbuild package from the repo, which
# tracks in all the kernel-devel packages
BuildRequires:  %{_bindir}/kmodtool

%{!?kernels:BuildRequires: buildsys-build-rpmfusion-kerneldevpkgs-%{?buildforkernels:%{buildforkernels}}%{!?buildforkernels:current}-%{_target_cpu} }

# kmodtool does its magic here
%{expand:%(kmodtool --target %{_target_cpu} --repo rpmfusion --kmodname %{name} %{?buildforkernels:--%{buildforkernels}} %{?kernels:--for-kernels "%{?kernels}"} 2>/dev/null) }


%description


%prep
# error out if there was something wrong with kmodtool
%{?kmodtool_check}

# print kmodtool output for debugging purposes:
kmodtool  --target %{_target_cpu}  --repo %{repo} --kmodname %{name} %{?buildforkernels:--%{buildforkernels}} %{?kernels:--for-kernels "%{?kernels}"} 2>/dev/null

%setup -q -c -T -a 0

# apply patches and do other stuff here
# pushd foo-%{version}
# #patch0 -p1 -b .suffix
# popd

for kernel_version in %{?kernel_versions} ; do
    cp -a foo-%{version} _kmod_build_${kernel_version%%___*}
done


%build
for kernel_version in %{?kernel_versions}; do
    make %{?_smp_mflags} -C "${kernel_version##*___}" SUBDIRS=${PWD}/_kmod_build_${kernel_version%%___*} modules
done


%install
rm -rf ${RPM_BUILD_ROOT}

for kernel_version in %{?kernel_versions}; do
    make install DESTDIR=${RPM_BUILD_ROOT} KMODPATH=%{kmodinstdir_prefix}/${kernel_version%%___*}/%{kmodinstdir_postfix}
    # install -D -m 755 _kmod_build_${kernel_version%%___*}/foo/foo.ko  ${RPM_BUILD_ROOT}%{kmodinstdir_prefix}/${kernel_version%%___*}/%{kmodinstdir_postfix}/foo.kmod
done
%{?akmod_install}


%clean
rm -rf $RPM_BUILD_ROOT


%changelog

So how does it work? let's go through important parts and describe the different bits:

%install
rm -rf ${RPM_BUILD_ROOT
}
for kernel_version in %{kernel_versions}; do
    install -D -m 755 _kmod_build_${kernel_version%%___*}/foo/foo.ko ${RPM_BUILD_ROOT}%{kmodinstdir_prefix}/foo.ko
done
chmod u+x ${RPM_BUILD_ROOT}/lib/modules/*/extra/*/*

kmodtool

kmodtool is a simple bash script that creates parts of the spec file dynamically depending on how you call it -- that's why the kmod template above has no files section. kmodtool gets called like this

# kmodtool does its magic here
%{expand:%(kmodtool --target %{_target_cpu} --repo %{repo} --kmodname foo %{?buildforkernels:--%{buildforkernels}} %{?kernels:--for-kernels "%{?kernels}"} 2>/dev/null) }

kmodtool itself is no real magic -- if you understand a bit of bash scripting you should be able to see how it works. It created three major parts:

  1. one that defines some macros
  2. a meta-package kmod-foo per kernel variant, and
  3. a package kmod-foo-<kernel-version> per kernel variant which contains the module itself and gets tracked in for new kernels by new kmod-foo meta package.

kmod-foo-<kernel-version> package

The most important part of kmodtool is the function print_rpmtemplate_per_kmodpkg, which contains a template for the package which will hold the kmod later:

%package       -n kmod-${kmodname}-${verrel}${variant}

Summary:          ${kmodname} kernel module(s) for ${verrel}${variant}

Group:            System Environment/Kernel

Provides:         kernel-modules-for-kernel = ${verrel}${variant}

Provides:         ${kmodname}-kmod = %{?epoch:%{epoch}:}%{version}-%{release}

Requires:         ${kmodname}-kmod-common >= %{?epoch:%{epoch}:}%{version}

Requires(post):   /sbin/depmod

Requires(postun): /sbin/depmod

Requires:         kernel-%{_target_cpu} = ${verrel}${variant}
BuildRequires:    kernel${dashvariant}-devel-%{_target_cpu} = ${verrel}


%post          -n kmod-${kmodname}-${verrel}${variant}

/sbin/depmod -aeF /boot/System.map-${verrel}${variant} ${verrel}${variant} > /dev/null || :

%postun        -n kmod-${kmodname}-${verrel}${variant}

/sbin/depmod  -aF /boot/System.map-${verrel}${variant} ${verrel}${variant} &> /dev/null || :

%description  -n kmod-${kmodname}-${verrel}${variant}

This package provides the ${kmodname} kernel modules built for the Linux

kernel ${verrel}${variant} for the %{_target_cpu} family of processors.


%files        -n kmod-${kmodname}-${verrel}${variant}

%defattr(644,root,root,755)

/lib/modules/${verrel}${variant}/extra/${kmodname}/

This macro later expands to something like the following when building for a standard kernel 2.6.23.1-42.fc8 on i686 and %{version} = 1.5; the output get inserted into the spec file before building :

%package       -n kmod-foo-2.6.23.1-42.fc8

Summary:          foo kernel module(s) for 2.6.23.1-42.fc8

Group:            System Environment/Kernel

Provides:         kernel-modules-for-kernel = 2.6.23.1-42.fc8

Provides:         foo-kmod = 1.5-1
Requires:         foo-kmod-common >= 1.5-1
Requires(post):   /sbin/depmod

Requires(postun): /sbin/depmod



Requires:         kernel-i686 = 2.6.23.1-42.fc8

BuildRequires:    kernel-devel- i686 = 2.6.23.1-42.fc8

%post          -n kmod-foo-2.6.23.1-42.fc8

/sbin/depmod -aeF /boot/System.map-2.6.23.1-42.fc8 2.6.23.1-42.fc8 > /dev/null || :

%postun        -n kmod-foo-2.6.23.1-42.fc8

/sbin/depmod  -aF /boot/System.map-2.6.23.1-42.fc8 2.6.23.1-42.fc8 &> /dev/null || :



%description  -n kmod-foo-2.6.23.1-42.fc8

This package provides the foo kernel modules built for the Linux

kernel 2.6.23.1-42.fc8 for the %{_target_cpu} family of processors.


%files        -n kmod-foo-2.6.23.1-42.fc8

%defattr(644,root,root,755)

/lib/modules/2.6.23.1-42.fc8/extra/foo/

Why all that? Let's go though the interesting bits in detail:

All kernel modules need to have the prefix kmod (that's a bit shorter than kernel-module) and need to be in a seperate package

Via this provides depsolvers or scripts can check for which kernel a module was build.

The userland-package that depends on a package that provides that to make sure that yum and other depsolvers install a proper kernel-module if you install a userland package that requires a kernel-module.

A kernel module without the kernel it was built for is useless. Don't use /boot/vmlinuz-*, it's not portable.

Kernel modules without the userland part is useless in most cases. There are rare packages when kernel modules don't need a part in userland, but we require it anyway -- at least the license and the docs needs to be placed somewhere in any case and a userland package is the right place for them.

Needed for building kernel-modules. This is just here if we are building for a Fedora kernel and will not show up when building for a custom kernel

Kernel modules shall not be executable -- but they need to be after %install to allow /usr/lib/rpm/find-debuginfo.sh to strip them.

Separate location -- don't mess up with the rest of the kernel. "extra" was picked because of upstream kernel documentation. Only kernel modules in that dir are allowed -- nothing else, because otherwise they might conflict between different versions!

kmod-foo

This is a meta-package which only purpose is to depend on the latest kmod-foo-<kernel-version> package. Without this meta-pacakge the later would be tracked in automatically for newly released kernels, because yum can't know that kmod-ntfs-2.6.23.1-49.fc8.x86_64 is a update for kmod-ntfs-2.6.23.1-42.fc8.x86_64 and that the latter should remain installed.

The meta-package gets created in the function print_rpmtemplate_kmodmetapkg from kmodtool. It will create a package which looks like this:

%package      -n kmod-foo
Summary:         Metapackage which tracks in foo kernel module for newest kernel
Group:           System Environment/Kernel
Requires:        kmod-foo-2.6.23.1-49.fc8

%description  -n kmod-foo
This is a meta-package without payload which sole purpose is to require the
foo kernel module(s) for the newest kernel,
to make sure you get it together with a new kernel.

%files        -n kmod-foo
%defattr(644,root,root,755)

macros

kmodtool further creates three macros that can be used in the spec file during the %prep, %build and %install stages.

%define kmodinstdir_prefix  /lib/modules/
%define kmodinstdir_postfix  /extra/ntfs/
%define kernel_versions 2.6.23.1-49.fc8___%{_usrsrc}/kernels/2.6.23.1-49.fc8-%{_target_cpu} 2.6.23.1-49.fc8PAE___%{_usrsrc}/kernels/2.6.23.1-49.fc8PAE-%{_target_cpu}

Look above in the kmod spec file template for details how to use those three.

kmods v1 versus v2

There were several design issues that lead to the the overhaul named kmods v2:

Mini-FAQ

Is it possible to compile a kmod against self compiled or other kernels?

Yes, easily, it's just a few steps. If you don't have a rpm build environment set one up like this

$ su -c "yum -y install rpmdevtools kmodtool kernel-devel"
$ rpmdev-setuptree

Now download the kmod src.rpm to the local directory and rebuild it for the running kernel:

$ yumdownloader --source kmod-foo
$ rpmbuild --rebuild foo-kmod*.src.rpm --define "kernels $(uname -r)" --target $(uname -m)

At the end of the build output you will see the names of the RPMs rpmbuild built. Use yum to install it with a command like this

$ su -c "yum --nogpgcheck install ~/rpmbuild/RPMS/i686/kmod-foo-1.0-4.i686.rpm"

Some notes, as above commands sometimes need some adjustments:

What's the best way to test if the spec file actually works?

Build the kmod (with and without mock) against a different kernel than the one that is running on your system.

Will there be further enhancements for the kernel module proposal?

Probably yes. We probably didn't consider every possible scenario out there. But the general scheme/direction probably will stay.

Do I need a special plugin for yum or other depsolvers?

It works fine without any plugins. Maybe a plugin will be written to fix two corner cases:

Does it require special support in the buildsys of a repo?

No. There is just a minor glitch in some versions of mock that makes building for older kernels a two-cycle build (if you BuildRequire both kernel-devel-2.6.23.1-42.fc8 and kernel-devel-2.6.23.1-49.fc8 in a package mock/yum will only install kernel-devel-2.6.23.1-49.fc8)

This standard is stupid.

Maybe. Post a better one. No offense, more a FYI: FESCo and some people from Livna and RPM Fusion invested a lot of time in this standard and had to make a lot of compromises to make everyone happy -- doing bigger changes just "because I like my scheme better" probably won't get accepted. But technical advantages that are properly documented (and maybe even tested in real life) might convince us. But remember, the improvements need to be worth converting all the existing packages over.

Why not simply use dkms

The short answer: Dkms especially for experienced users that often change kernels is a nice concept that also has a lot of usage cases for enterprise customers. Thus repos should consider to ship dkms packages in parallel to kmods. But for ordinary users shipping a pre-compiled module has many advantages (don't need devel environment) instead of doing a Gentoo-Like (custom build) concept.

Examples

A good example for the kmod v2 standards are the madiwfi/madwifi-kmod package from Livna; the ntfs packages are also a good starting point. For an example of the "skip the build for the xen kernel" take a look at the kmods for the graphic drivers from AMD and Nvidia.


CategoryPackaging CategoryKmodPackaging