Files
linuxkit/docs/kernels.md
Avi Deitcher c6b7b3c00a Update kernel build docs
Signed-off-by: Avi Deitcher <avi@deitcher.net>
2017-07-17 14:03:53 +03:00

13 KiB

Linux kernels

LinuxKit kernel images are distributed as hub images which contain the kernel, kernel modules, kernel config file, and optionally, kernel headers to compile kernel modules against. The repository containing the official LinuxKit kernels is at linuxkit/kernels.

The LinuxKit kernels are based on the latest stable releases and are updated frequently to include bug and security fixes. For some kernels we do carry additional patches, which are mostly back-ported fixes from newer kernels. The full kernel source with patches can be found on github. Each kernel image is tagged with the full kernel version (e.g., linuxkit/kernel:4.9.33) and with the full kernel version plus the hash of the files it was created from (git tree hash of the ./kernel directory). For selected kernels (mostly the LTS kernels and latest stable kernels) we also compile/push kernels with additional debugging enabled. The hub images for these kernels have the _dbg suffix in the tag. For some kernels, we also provide matching packages containing the perf utility for debugging and performance tracing. The perf package is called kernel-perf and is tagged the same way as the kernel packages.

In addition to the official kernel images, LinuxKit offers the ability to build bootable Linux images with kernels from various distributions. We mostly offer this mostly for testing purposes. "Foreign" kernel images are created by re-packing the native kernel packages into hub images. The hub images are typically tagged with the kernel version.

In summary, LinuxKit offers a choice of the following kernels:

Compiling external kernel modules

This section describes how to build external (out-of-tree) kernel modules. It is assumed you have the source available to those modules, and require the correct kernel version headers and compile tools.

The LinuxKit kernel packages include kernel-dev.tar which contains the headers and other files required to compile kernel modules against the specific version of the kernel. Currently, the headers are not included in the initial RAM disk, but it is possible to compile custom modules offline and include then include the modules in the initial RAM disk.

There is a example, but basically one can use a multi-stage build to compile the kernel modules:

FROM linuxkit/kernel:4.9.33 AS ksrc
# Extract headers and compile module
FROM linuxkit/kernel-compile:1b396c221af673757703258159ddc8539843b02b@sha256:6b32d205bfc6407568324337b707d195d027328dbfec554428ea93e7b0a8299b AS build
COPY --from=ksrc /kernel-dev.tar /
RUN tar xf kernel-dev.tar

# copy module source code and compile

To use the kernel module, we recommend adding a final stage to the Dockerfile above, which copies the kernel module from the build stage and performs a insmod as the entry point. You can add this package to the onboot section in your YAML file. kmod.yml contains an example for the configuration.

Compiling internal kernel modules

If you want to compile in-tree kernel modules, i.e. those whose source is already in the kernel tree but have not been included in linuxkit/kernel, you have two options:

  1. Follow the external kernel modules process from above
  2. Modify the kernel config in ../kernel/ and rebuild the kernel.

In general, if it is an in-tree module, we prefer to include it in the standard linuxkit kernel distribution, i.e. option 2 above. Once you have it working, please open a Pull Request to include it.

External Process

The kernel-dev.tar included with each kernel does not include the kernel sources, only the headers. To build those modules, you will need to download the kernel source separately and recompile. The in-container process that downloads the source is available in the Dockerfile.

Modify Config

Building an in-tree module is very similar to building a new modified kernel (see below):

  1. Modify the appropriate kernel.config-* file(s)
  2. Compile

Building and using custom kernels

To build and test locally modified kernels, e.g., to try a different kernel config or new patches, the existing kernel build system in the ../kernel can be re-used. For example, assuming the current 4.9 kernel is 4.9.33, you can build a local kernel with:

make build_4.9.x

This will create a local kernel image called linuxkit/kernel:4.9.33-<hash>-dirty assuming you haven't committed you local changes. You can then use this in your YAML file as:

kernel:
  image: linuxkit/kernel:4.9.33-<hash>-dirty

If you have committed your local changes, the -dirty will not be appended. Then you can also override the Hub organisation to use the image elsewhere with:

make ORG=<your hub org>

The image will be uploaded to Hub and can be use in a YAML file as <your hub org>/kernel:4.9.33 or as <your hub org>/kernel:4.9.33-<hash>.

Modifying the Config

Each series of kernels has a config file dedicated to it in ../kernel/, e.g. kernel.config-4.9.x. To build a particular series of kernel:

  1. Create a separate git branch (not required but strongly recommended)
  2. Modify the appropriate kernel.config, e.g. kernel.config-4.9.x
  3. Run make build_<series> with appropriate arguments per this section, e.g. make build_4.9.x ORG=foo HASH=bar
  4. Create a .yml, build and test

You can modify the config in one of two ways:

  • Manually, editing the config file
  • Using a standard config generator, like menuconfig

Generally, you will manually edit a file if you are a Linux kernel expert and fully understand all of the dependencies, or if the change is minor and you are highly confident there are no dependencies.

If you wish to use menuconfig, which figures out dependencies for you, you will need an environment in which to run it. Fortunately, the linuxkit project's kernel compile process already sets one up for you. To get an appropriate environment:

  1. cd kernel/
  2. Run a build for your desired kernel series, e.g. make build_4.9.x ORG=foo HASH=bar
  3. When you see the output from make defconfig && make oldconfig complete, hit Ctrl-C to stop the build
  4. Note the hash from the intermediate container. That intermediate container has all of the tools and source in it, and can be used to build.
  5. Get a shell in that intermediate container, mounting the current directory in: docker run -it --rm -v ${PWD}:/src <hash> sh

This will give you a read-to-run kernel build environment, with all of the config files in /src/.

For the output of step 4, e.g.:

Step X/29 : COMMAND
 ---> b2a4a976d661

Once you have your shell, and you want to run the config, you can do the following. We assume you have launched your config container using the steps above, i.e. docker run -it --rm -v ${PWD}:/src <hash> sh. The kernel source is in /linux/, while the kernel/ directory from linuxkit is in /src/:

Unless you are building the config from scratch, you probably want to make small modifications to the existing config.

The appropriate config at /src/kernel.config-<series> was already copied over to /linux/.config by the build.

  1. cd /linux
  2. make menuconfig
  3. Load in the existing config: On the bottom menu, use the left-right arrow keys to Load
  4. Load it from .config
  5. Exit from the Load pop-up and make the desired changes
  6. Save the modified config: On the bottom menu, use the left-right arrow keys to Save
  7. Save it to .config
  8. Exit the menu by selecting Exit from the bottom meny as many times as necessary
  9. Copy the saved config to the mount location: cp /linux/.config /src/some-saved-name.config (replace with an appropriate name)
  10. Exit out of the container
  11. Check the differences generated by menuconfig with diff kernel.config-<series> some-saved-name.config.
    • If the changes are as you expected, proceed to the next step
    • If the changes are different, either return to the container and menuconfig, or edit manually
  12. Copy the new config file to the build location: cp some-saved-name.config kernel.config-<series>
  13. Run your build: make build_<series>

Working with Linux kernel patches for LinuxKit

We may apply patches to the Linux kernel used in LinuxKit, primarily to cherry-pick some upstream patches or to add some additional functionality, not yet accepted upstream.

Patches are located in kernel/patches-<kernel version> and should follow these rules:

  • Patches must be in git am format, i.e. they should contain a complete and sensible commit message.
  • Patches must contain a Developer's Certificate of Origin.
  • Patch files must have a numeric prefix to ensure the ordering in which they are applied.
  • If patches are cherry-picked, they must be cherry-picked with -x to contain the original commit ID.
  • If patches are from a different git tree (other than the stable tree), or from a mailing list posting they should contain an Origin: line with a link to the source.

This document outlines the recommended procedure to handle patches. The general process is to apply them to a branch of the Linux stable tree and then export them with git format-patch.

If you want to add or remove patches currently used, please also ping @rneugeba on the PR so that we can update our internal Linux tree to ensure that patches are carried forward if we update the kernel in the future.

Preparation

Patches are applied to point releases of the linux stable tree. You need an up-to-date copy of that tree:

git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git

Add it as a remote to a clone of the LinuxKit clone.

We use the following variables:

  • KITSRC: Base directory of LinuxKit repository
  • LINUXSRC: Base directory of Linux stable kernel repository e.g.:
KITSRC=~/src/linuxkit/linuxkit
LINUXSRC=~/src/linuxkit/linux

to refer to the location of the LinuxKit and Linux kernel trees.

Updating the patches to a new kernel version

There are different ways to do this, but we recommend applying the patches to the current version and then rebase to the new version. We define the following variables to refer to the current base tag and the new tag you want to rebase the patches to:

CURTAG=v4.9.14
NEWTAG=v4.9.15

If you don't already have a branch, it's best to import the current patch set and then rebase:

cd $LINUXSRC
git checkout -b ${NEWTAG}-linuxkit ${CURTAG}
git am ${KITSRC}/kernel/patches/*.patch
git rebase ${NEWTAG}-linuxkit ${NEWTAG}

The git am should not have any conflicts and if the rebase has conflicts resolve them, then git add <files> and git rebase --continue.

If you already have linux tree with a ${CURTAG}-linuxkit branch, you can rebase by creating a new branch from the current branch and then rebase:

cd $LINUXSRC
git checkout ${CURTAG}-linuxkit
git branch ${NEWTAG}-linuxkit ${CURTAG}-linuxkit
git rebase --onto ${NEWTAG} ${NEWTAG} ${NEWTAG}-linuxkit

Again, resolve any conflicts as described above.

Adding/Removing patches

If you want to add or remove patches make sure you have an up-to-date branch with the currently applied patches (see above). Then either any normal means (git cherry-pick -x, git am, or git commit, etc) to add new patches. For cherry-picked patches also please add a Origin: line after the DCO lines with a reference the git tree the patch was cherry-picked from.

If the patch is not cherry-picked try to include as much information in the commit message as possible as to where the patch originated from. The canonical form would be to add a Origin: line after the DCO lines, e.g.:

Origin: https://patchwork.ozlabs.org/patch/622404/

Export patches to LinuxKit

To export patches to LinuxKit, you should use git format-patch from the Linux tree, e.g., something along these lines:

cd $LINUXSRC
rm $KITSRC/kernel/patches-4.9.x/*
git format-patch -o $KITSRC/kernel/patches-4.9.x v4.9.15..HEAD

Then, create a PR for LinuxKit.

Using perf

The kernel-perf package contains a statically linked perf binary under /usr/bin which is matched with the kernel of the same tag. The simplest way to use the perf utility is to add the package to the init section in the YAML file. This adds the binary to the root filesystem.

To use the binary, you can either bind mount it into the getty or ssh service container or you can access the root filesystem from the getty container via nsenter:

nsenter -m/proc/1/ns/mnt ash

Alternatively, you can add the kernel-perf package as stage in a multi-stage build to add it to a custom package.