Copyright © 2010-2017 Linux Foundation
Permission is granted to copy, distribute and/or modify this document under the terms of the Creative Commons Attribution-Share Alike 2.0 UK: England & Wales as published by Creative Commons.
This version of the Yocto Project Linux Kernel Development Manual is for the 2.1.3 release of the Yocto Project. To be sure you have the latest version of the manual for this release, go to the Yocto Project documentation page and select the manual from that site. Manuals from the site are more up-to-date than manuals derived from the Yocto Project released TAR files.
If you located this manual through a web search, the version of the manual might not be the one you want (e.g. the search might have returned a manual much older than the Yocto Project version with which you are working). You can see all Yocto Project major releases by visiting the Releases page. If you need a version of this manual for a different Yocto Project release, visit the Yocto Project documentation page and select the manual set by using the "ACTIVE RELEASES DOCUMENTATION" or "DOCUMENTS ARCHIVE" pull-down menus.
                       To report any inaccuracies or problems with this
                       manual, send an email to the Yocto Project
                       discussion group at
                       yocto@yoctoproject.com or log into
                       the freenode #yocto channel.
                       
| Revision History | |
|---|---|
| Revision 1.4 | April 2013 | 
| Released with the Yocto Project 1.4 Release. | |
| Revision 1.5 | October 2013 | 
| Released with the Yocto Project 1.5 Release. | |
| Revision 1.5.1 | January 2014 | 
| Released with the Yocto Project 1.5.1 Release. | |
| Revision 1.6 | April 2014 | 
| Released with the Yocto Project 1.6 Release. | |
| Revision 1.7 | October 2014 | 
| Released with the Yocto Project 1.7 Release. | |
| Revision 1.8 | April 2015 | 
| Released with the Yocto Project 1.8 Release. | |
| Revision 2.0 | October 2015 | 
| Released with the Yocto Project 2.0 Release. | |
| Revision 2.1 | April 2016 | 
| Released with the Yocto Project 2.1 Release. | |
| Revision 2.1.1 | August 2016 | 
| Released with the Yocto Project 2.1.1 Release. | |
| Revision 2.1.2 | December 2016 | 
| Released with the Yocto Project 2.1.2 Release. | |
| Revision 2.1.3 | June 2017 | 
| Released with the Yocto Project 2.1.3 Release. | |
Table of Contents
Table of Contents
Regardless of how you intend to make use of the Yocto Project, chances are you will work with the Linux kernel. This manual provides background information on the Yocto Linux kernel Metadata, describes common tasks you can perform using the kernel tools, and shows you how to use the kernel Metadata needed to work with the kernel inside the Yocto Project.
            Each Yocto Project release has a set of linux-yocto recipes, whose
            Git repositories you can view in the Yocto
            Source Repositories under
            the "Yocto Linux Kernel" heading.
            New recipes for the release track the latest upstream developments
            and introduce newly-supported platforms.
            Previous recipes in the release are refreshed and supported for at
            least one additional release.
            As they align, these previous releases are updated to include the
            latest from the
            Long Term Support Initiative
            (LTSI) project.
            Also included is a linux-yocto development recipe
            (linux-yocto-dev.bb) should you want to work
            with the very latest in upstream Linux kernel development and
            kernel Metadata development.
        
The Yocto Project also provides a powerful set of kernel tools for managing Linux kernel sources and configuration data. You can use these tools to make a single configuration change, apply multiple patches, or work with your own kernel sources.
            In particular, the kernel tools allow you to generate configuration
            fragments that specify only what you must, and nothing more.
            Configuration fragments only need to contain the highest level
            visible CONFIG options as presented by the Linux
            kernel menuconfig system.
            Contrast this against a complete Linux kernel
            .config, which includes all the automatically
            selected CONFIG options.
            This efficiency reduces your maintenance effort and allows you
            to further separate your configuration in ways that make sense for
            your project.
            A common split separates policy and hardware.
            For example, all your kernels might support
            the proc and sys filesystems,
            but only specific boards require sound, USB, or specific drivers.
            Specifying these configurations individually allows you to aggregate
            them together as needed, but maintains them in only one place.
            Similar logic applies to separating source changes.
        
If you do not maintain your own kernel sources and need to make only minimal changes to the sources, the released recipes provide a vetted base upon which to layer your changes. Doing so allows you to benefit from the continual kernel integration and testing performed during development of the Yocto Project.
If, instead, you have a very specific Linux kernel source tree and are unable to align with one of the official linux-yocto recipes, an alternative exists by which you can use the Yocto Project Linux kernel tools with your own kernel sources.
The sections that follow provide instructions for completing specific Linux kernel development tasks. These instructions assume you are comfortable working with BitBake recipes and basic open-source development tools. Understanding these concepts will facilitate the process of working with the kernel recipes. If you find you need some additional background, please be sure to review and understand the following documentation:
The "Modifying Source Code" section in the Yocto Project Development Manual
The "Understanding and Creating Layers" section in the Yocto Project Development Manual
The "Modifying the Kernel" section in the Yocto Project Development Manual.
Finally, while this document focuses on the manual creation of recipes, patches, and configuration files, the Yocto Project Board Support Package (BSP) tools are available to automate this process with existing content and work well to create the initial framework and boilerplate code. For details on these tools, see the "Using the Yocto Project's BSP Tools" section in the Yocto Project Board Support Package (BSP) Developer's Guide.
Table of Contents
This chapter presents several common tasks you perform when you work with the Yocto Project Linux kernel. These tasks include preparing a layer, modifying an existing recipe, iterative development, working with your own sources, and incorporating out-of-tree modules.
            If you are going to be modifying kernel recipes, it is recommended
            that you create and prepare your own layer in which to do your
            work.
            Your layer contains its own
            BitBake
            append files
            (.bbappend) and provides a convenient
            mechanism to create your own recipe files
            (.bb).
            For details on how to create and work with layers, see the following
            sections in the Yocto Project Development Manual:
            
"Understanding and Creating Layers" for general information on layers and how to create layers.
"Set Up Your Layer for the Build" for specific instructions on setting up a layer for kernel development.
            In many cases, you can customize an existing linux-yocto recipe to
            meet the needs of your project.
            Each release of the Yocto Project provides a few Linux
            kernel recipes from which you can choose.
            These are located in the
            Source Directory
            in meta/recipes-kernel/linux.
        
Modifying an existing recipe can consist of the following:
Creating the append file
Applying patches
Changing the configuration
Before modifying an existing recipe, be sure that you have created a minimal, custom layer from which you can work. See the "Creating and Preparing a Layer" section for some general resources. You can also see the "Set Up Your Layer for the Build" section of the Yocto Project Development Manual for a detailed example.
                You create this file in your custom layer.
                You also name it accordingly based on the linux-yocto recipe
                you are using.
                For example, if you are modifying the
                meta/recipes-kernel/linux/linux-yocto_3.19.bb
                recipe, the append file will typically be located as follows
                within your custom layer:
                
     your-layer/recipes-kernel/linux/linux-yocto_3.19.bbappend
                
                The append file should initially extend the
                FILESPATH
                search path by prepending the directory that contains your
                files to the
                FILESEXTRAPATHS
                variable as follows:
                
     FILESEXTRAPATHS_prepend := "${THISDIR}/${PN}:"
                
                The path ${THISDIR}/${PN}
                expands to "linux-yocto" in the current directory for this
                example.
                If you add any new files that modify the kernel recipe and you
                have extended FILESPATH as
                described above, you must place the files in your layer in the
                following area:
                
     your-layer/recipes-kernel/linux/linux-yocto/
                
                If you have a single patch or a small series of patches
                that you want to apply to the Linux kernel source, you
                can do so just as you would with any other recipe.
                You first copy the patches to the path added to
                FILESEXTRAPATHS
                in your .bbappend file as described in
                the previous section, and then reference them in
                SRC_URI
                statements.
            
                For example, you can apply a three-patch series by adding the
                following lines to your linux-yocto
                .bbappend file in your layer:
                
     SRC_URI += "file://0001-first-change.patch"
     SRC_URI += "file://0002-second-change.patch"
     SRC_URI += "file://0003-third-change.patch"
                The next time you run BitBake to build the Linux kernel, BitBake detects the change in the recipe and fetches and applies the patches before building the kernel.
For a detailed example showing how to patch the kernel, see the "Patching the Kernel" section in the Yocto Project Development Manual.
                You can make wholesale or incremental changes to the final
                .config file used for the eventual
                Linux kernel configuration by including a
                defconfig file and by specifying
                configuration fragments in the
                SRC_URI
                to be applied to that file.
            
                If you have a complete, working Linux kernel
                .config
                file you want to use for the configuration, as before, copy
                that file to the appropriate ${PN}
                directory in your layer's
                recipes-kernel/linux directory,
                and rename the copied file to "defconfig".
                Then, add the following lines to the linux-yocto
                .bbappend file in your layer:
                
     FILESEXTRAPATHS_prepend := "${THISDIR}/${PN}:"
     SRC_URI += "file://defconfig"
                
                The SRC_URI tells the build system how to
                search for the file, while the
                FILESEXTRAPATHS
                extends the
                FILESPATH
                variable (search directories) to include the
                ${PN} directory you created to hold the
                configuration changes.
            
defconfig file before applying any
                subsequent configuration fragments.
                The final kernel configuration is a combination of the
                configurations in the defconfig file and
                any configuration fragments you provide.
                You need to realize that if you have any configuration
                fragments, the build system applies these on top of and
                after applying the existing defconfig
                file configurations.
            
                Generally speaking, the preferred approach is to determine the
                incremental change you want to make and add that as a
                configuration fragment.
                For example, if you want to add support for a basic serial
                console, create a file named 8250.cfg in
                the ${PN} directory with the following
                content (without indentation):
                
     CONFIG_SERIAL_8250=y
     CONFIG_SERIAL_8250_CONSOLE=y
     CONFIG_SERIAL_8250_PCI=y
     CONFIG_SERIAL_8250_NR_UARTS=4
     CONFIG_SERIAL_8250_RUNTIME_UARTS=4
     CONFIG_SERIAL_CORE=y
     CONFIG_SERIAL_CORE_CONSOLE=y
                
                Next, include this configuration fragment and extend the
                FILESPATH variable in your
                .bbappend file:
                
     FILESEXTRAPATHS_prepend := "${THISDIR}/${PN}:"
     SRC_URI += "file://8250.cfg"
                The next time you run BitBake to build the Linux kernel, BitBake detects the change in the recipe and fetches and applies the new configuration before building the kernel.
For a detailed example showing how to configure the kernel, see the "Configuring the Kernel" section in the Yocto Project Development Manual.
defconfig File¶
                It might be desirable to have kernel configuration fragment
                support through a defconfig file that
                is pulled from the kernel source tree for the configured
                machine.
                By default, the OpenEmbedded build system looks for
                defconfig files in the layer used for
                Metadata, which is "out-of-tree", and then configures them
                using the following:
                
     SRC_URI += "file://defconfig"
                
                If you do not want to maintain copies of
                defconfig files in your layer but would
                rather allow users to use the default configuration from the
                kernel tree and still be able to add configuration fragments
                to the
                SRC_URI
                through, for example, append files, you can direct the
                OpenEmbedded build system to use a
                defconfig file that is "in-tree".
            
                To specify an "in-tree" defconfig file,
                edit the recipe that builds your kernel so that it has the
                following command form:
                
     KBUILD_DEFCONFIG_KMACHINE ?= defconfig_file
                
                You need to append the variable with
                KMACHINE
                and then supply the path to your "in-tree"
                defconfig file.
            
                Aside from modifying your kernel recipe and providing your own
                defconfig file, you need to be sure no
                files or statements set SRC_URI to use a
                defconfig other than your "in-tree"
                file (e.g. a kernel's linux-machine.inc
                file).
                In other words, if the build system detects a statement
                that identifies an "out-of-tree"
                defconfig file, that statement
                will override your
                KBUILD_DEFCONFIG variable.
            
                See the
                KBUILD_DEFCONFIG
                variable description for more information.
            
If you do not have existing patches or configuration files, you can iteratively generate them from within the BitBake build environment as described within this section. During an iterative workflow, running a previously completed BitBake task causes BitBake to invalidate the tasks that follow the completed task in the build sequence. Invalidated tasks rebuild the next time you run the build using BitBake.
As you read this section, be sure to substitute the name of your Linux kernel recipe for the term "linux-yocto".
If kernel images are being built with "-dirty" on the end of the version string, this simply means that modifications in the source directory have not been committed.
     $ git status
                
You can use the above Git command to report modified, removed, or added files. You should commit those changes to the tree regardless of whether they will be saved, exported, or used. Once you commit the changes, you need to rebuild the kernel.
To force a pickup and commit of all such pending changes, enter the following:
     $ git add .
     $ git commit -s -a -m "getting rid of -dirty"
                
Next, rebuild the kernel.
                You can manipulate the .config file
                used to build a linux-yocto recipe with the
                menuconfig command as follows:
                
     $ bitbake linux-yocto -c menuconfig
                
                This command starts the Linux kernel configuration tool,
                which allows you to prepare a new
                .config file for the build.
                When you exit the tool, be sure to save your changes
                at the prompt.
            
                The resulting .config file is
                located in the build directory,
                ${B},
                which expands to
                ${WORKDIR}/linux-${PACKAGE_ARCH}-${LINUX_KERNEL_TYPE}-build.
                You can use the entire .config file as the
                defconfig file as described in the
                "Changing the Configuration" section.
                For more information on the .config file,
                see the
                "Using menuconfig"
                section in the Yocto Project Development Manual.
                
bitbake -e
                    command:
                    
     $ bitbake -e virtual/kernel
                    
                    Search the output for the variable in which you are
                    interested to see exactly how it is expanded and used.
                
                A better method is to create a configuration fragment using the
                differences between two configuration files: one previously
                created and saved, and one freshly created using the
                menuconfig tool.
            
To create a configuration fragment using this method, follow these steps:
Complete a build at least through the kernel configuration task as follows:
     $ bitbake linux-yocto -c kernel_configme -f
                        
                        This step ensures that you will be creating a
                        .config file from a known state.
                        Because situations exist where your build state might
                        become unknown, it is best to run the previous
                        command prior to starting up
                        menuconfig.
                        
Run the menuconfig
                        command:
                        
     $ bitbake linux-yocto -c menuconfig
                        Run the diffconfig
                        command to prepare a configuration fragment.
                        The resulting file fragment.cfg
                        will be placed in the
                        ${WORKDIR} directory:
                        
     $ bitbake linux-yocto -c diffconfig
                        
                The diffconfig command creates a file that is a
                list of Linux kernel CONFIG_ assignments.
                See the "Changing the Configuration"
                section for information on how to use the output as a
                configuration fragment.
                
                The kernel tools also provide configuration validation.
                You can use these tools to produce warnings for when a
                requested configuration does not appear in the final
                .config file or when you override a
                policy configuration in a hardware configuration fragment.
                Here is an example with some sample output of the command
                that runs these tools:
                
     $ bitbake linux-yocto -c kernel_configcheck -f
     ...
     NOTE: validating kernel configuration
     This BSP sets 3 invalid/obsolete kernel options.
     These config options are not offered anywhere within this kernel.
     The full list can be found in your kernel src dir at:
     meta/cfg/standard/mybsp/invalid.cfg
     This BSP sets 21 kernel options that are possibly non-hardware related.
     The full list can be found in your kernel src dir at:
     meta/cfg/standard/mybsp/specified_non_hdw.cfg
     WARNING: There were 2 hardware options requested that do not
              have a corresponding value present in the final ".config" file.
              This probably means you are not getting the config you wanted.
              The full list can be found in your kernel src dir at:
              meta/cfg/standard/mybsp/mismatch.cfg
                
                The output describes the various problems that you can
                encounter along with where to find the offending configuration
                items.
                You can use the information in the logs to adjust your
                configuration files and then repeat the
                kernel_configme and
                kernel_configcheck commands until
                they produce no warnings.
            
                For more information on how to use the
                menuconfig tool, see the
                "Using menuconfig"
                section in the Yocto Project Development Manual.
            
You can experiment with source code changes and create a simple patch without leaving the BitBake environment. To get started, be sure to complete a build at least through the kernel configuration task:
     $ bitbake linux-yocto -c kernel_configme -f
                
                Taking this step ensures you have the sources prepared
                and the configuration completed.
                You can find the sources in the build directory within the
                source/ directory, which is a symlink
                (i.e. ${B}/source).
                The source/ directory expands to
                ${WORKDIR}/linux-${PACKAGE_ARCH}-${LINUX_KERNEL_TYPE}-build/source.
                The directory pointed to by the
                source/ symlink is also known as
                ${STAGING_KERNEL_DIR}.
            
                You can edit the sources as you would any other Linux source
                tree.
                However, keep in mind that you will lose changes if you
                trigger the
                do_fetch
                task for the recipe.
                You can avoid triggering this task by not using BitBake to
                run the
                cleanall,
                cleansstate,
                or forced
                fetch
                commands.
                Also, do not modify the recipe itself while working
                with temporary changes or BitBake might run the
                fetch command depending on the
                changes to the recipe.
            
                To test your temporary changes, instruct BitBake to run the
                compile again.
                The -f option forces the command to run
                even though BitBake might think it has already done so:
                
     $ bitbake linux-yocto -c compile -f
                
                If the compile fails, you can update the sources and repeat
                the compile.
                Once compilation is successful, you can inspect and test
                the resulting build (i.e. kernel, modules, and so forth) from
                the following build directory:
                
     ${WORKDIR}/linux-${PACKAGE_ARCH}-${LINUX_KERNEL_TYPE}-build
                
                Alternatively, you can run the deploy
                command to place the kernel image in the
                tmp/deploy/images directory:
                
	$ bitbake linux-yocto -c deploy
                And, of course, you can perform the remaining installation and packaging steps by issuing:
	$ bitbake linux-yocto
                
For rapid iterative development, the edit-compile-repeat loop described in this section is preferable to rebuilding the entire recipe because the installation and packaging tasks are very time consuming.
                Once you are satisfied with your source code modifications,
                you can make them permanent by generating patches and
                applying them to the
                SRC_URI
                statement as described in the
                "Applying Patches"
                section.
                If you are not familiar with generating patches, refer to the
                "Creating the Patch"
                section in the Yocto Project Development Manual.
            
If you cannot work with one of the Linux kernel versions supported by existing linux-yocto recipes, you can still make use of the Yocto Project Linux kernel tooling by working with your own sources. When you use your own sources, you will not be able to leverage the existing kernel Metadata and stabilization work of the linux-yocto sources. However, you will be able to manage your own Metadata in the same format as the linux-yocto sources. Maintaining format compatibility facilitates converging with linux-yocto on a future, mutually-supported kernel version.
            To help you use your own sources, the Yocto Project provides a
            linux-yocto custom recipe
            (linux-yocto-custom.bb) that uses
            kernel.org sources
            and the Yocto Project Linux kernel tools for managing
            kernel Metadata.
            You can find this recipe in the
            poky Git repository of the
            Yocto Project Source Repository
            at:
            
     poky/meta-skeleton/recipes-kernel/linux/linux-yocto-custom.bb
            
Here are some basic steps you can use to work with your own sources:
Copy the linux-yocto-custom.bb
                    recipe to your layer and give it a meaningful name.
                    The name should include the version of the Linux kernel you
                    are using (e.g.
                    linux-yocto-myproject_3.19.bb,
                    where "3.19" is the base version of the Linux kernel
                    with which you would be working).
In the same directory inside your layer,
                    create a matching directory
                    to store your patches and configuration files (e.g.
                    linux-yocto-myproject).
                    
Make sure you have either a
                    defconfig file or configuration
                    fragment files.
                    When you use the linux-yocto-custom.bb
                    recipe, you must specify a configuration.
                    If you do not have a defconfig file,
                    you can run the following:
                    
     $ make defconfig
                    
                    After running the command, copy the resulting
                    .config to the
                    files directory as "defconfig" and
                    then add it to the
                    SRC_URI
                    variable in the recipe.
Running the make defconfig
                    command results in the default configuration for your
                    architecture as defined by your kernel.
                    However, no guarantee exists that this configuration is
                    valid for your use case, or that your board will even boot.
                    This is particularly true for non-x86 architectures.
                    To use non-x86 defconfig files, you
                    need to be more specific and find one that matches your
                    board (i.e. for arm, you look in
                    arch/arm/configs and use the one that
                    is the best starting point for your board).
                    
Edit the following variables in your recipe as appropriate for your project:
SRC_URI:
                            The SRC_URI should specify
                            a Git repository that uses one of the supported Git
                            fetcher protocols (i.e. file,
                            git, http,
                            and so forth).
                            The SRC_URI variable should
                            also specify either a defconfig
                            file or some configuration fragment files.
                            The skeleton recipe provides an example
                            SRC_URI as a syntax reference.
                            
LINUX_VERSION:
                            The Linux kernel version you are using (e.g.
                            "3.19").
LINUX_VERSION_EXTENSION:
                            The Linux kernel CONFIG_LOCALVERSION
                            that is compiled into the resulting kernel and visible
                            through the uname command.
                            
SRCREV:
                            The commit ID from which you want to build.
                            
PR:
                            Treat this variable the same as you would in any other
                            recipe.
                            Increment the variable to indicate to the OpenEmbedded
                            build system that the recipe has changed.
                            
PV:
                            The default PV assignment is
                            typically adequate.
                            It combines the LINUX_VERSION
                            with the Source Control Manager (SCM) revision
                            as derived from the
                            SRCPV
                            variable.
                            The combined results are a string with
                            the following form:
                            
     3.19.11+git1+68a635bf8dfb64b02263c1ac80c948647cc76d5f_1+218bd8d2022b9852c60d32f0d770931e3cf343e2
                            
                            While lengthy, the extra verbosity in PV
                            helps ensure you are using the exact
                            sources from which you intend to build.
                            
COMPATIBLE_MACHINE:
                            A list of the machines supported by your new recipe.
                            This variable in the example recipe is set
                            by default to a regular expression that matches
                            only the empty string, "(^$)".
                            This default setting triggers an explicit build
                            failure.
                            You must change it to match a list of the machines
                            that your new recipe supports.
                            For example, to support the qemux86
                            and qemux86-64 machines, use
                            the following form:
                            
     COMPATIBLE_MACHINE = "qemux86|qemux86-64"
                            Provide further customizations to your recipe as needed just as you would customize an existing linux-yocto recipe. See the "Modifying an Existing Recipe" section for information.
This section describes steps to build out-of-tree modules on your target and describes how to incorporate out-of-tree modules in the build.
While the traditional Yocto Project development model would be to include kernel modules as part of the normal build process, you might find it useful to build modules on the target. This could be the case if your target system is capable and powerful enough to handle the necessary compilation. Before deciding to build on your target, however, you should consider the benefits of using a proper cross-development environment from your build host.
                If you want to be able to build out-of-tree modules on
                the target, there are some steps you need to take
                on the target that is running your SDK image.
                Briefly, the kernel-dev package
                is installed by default on all
                *.sdk images and the
                kernel-devsrc package is installed
                on many of the *.sdk images.
                However, you need to create some scripts prior to
                attempting to build the out-of-tree modules on the target
                that is running that image.
            
                Prior to attempting to build the out-of-tree modules,
                you need to be on the target as root and you need to
                change to the /usr/src/kernel directory.
                Next, make the scripts:
                
     # cd /usr/src/kernel
     # make scripts
                
                Because all SDK image recipes include
                dev-pkgs, the
                kernel-dev packages will be installed
                as part of the SDK image and the
                kernel-devsrc packages will be installed
                as part of applicable SDK images.
                The SDK uses the scripts when building out-of-tree
                modules.
                Once you have switched to that directory and created the
                scripts, you should be able to build your out-of-tree modules
                on the target.
            
                While it is always preferable to work with sources integrated
                into the Linux kernel sources, if you need an external kernel
                module, the hello-mod.bb recipe is
                available as a template from which you can create your
                own out-of-tree Linux kernel module recipe.
            
                This template recipe is located in the
                poky Git repository of the
                Yocto Project Source Repository
                at:
                
     poky/meta-skeleton/recipes-kernel/hello-mod/hello-mod_0.1.bb
                
                To get started, copy this recipe to your layer and give it a
                meaningful name (e.g. mymodule_1.0.bb).
                In the same directory, create a new directory named
                files where you can store any source files,
                patches, or other files necessary for building
                the module that do not come with the sources.
                Finally, update the recipe as needed for the module.
                Typically, you will need to set the following variables:
                
                Depending on the build system used by the module sources,
                you might need to make some adjustments.
                For example, a typical module Makefile
                looks much like the one provided with the
                hello-mod template:
                
     obj-m := hello.o
     SRC := $(shell pwd)
     all:
         $(MAKE) -C $(KERNEL_SRC) M=$(SRC)
     modules_install:
         $(MAKE) -C $(KERNEL_SRC) M=$(SRC) modules_install
     ...
                
                The important point to note here is the
                KERNEL_SRC
                variable.
                The
                module
                class sets this variable and the
                KERNEL_PATH
                variable to
                ${
                with the necessary Linux kernel build information to build
                modules.
                If your module STAGING_KERNEL_DIR}Makefile uses a different
                variable, you might want to override the
                do_compile()
                step, or create a patch to
                the Makefile to work with the more typical
                KERNEL_SRC or
                KERNEL_PATH variables.
            
After you have prepared your recipe, you will likely want to include the module in your images. To do this, see the documentation for the following variables in the Yocto Project Reference Manual and set one of them appropriately for your machine configuration file:
Modules are often not required for boot and can be excluded from certain build configurations. The following allows for the most flexibility:
     MACHINE_EXTRA_RRECOMMENDS += "kernel-module-mymodule"
                
                The value is derived by appending the module filename without
                the .ko extension to the string
                "kernel-module-".
            
                Because the variable is
                RRECOMMENDS
                and not a
                RDEPENDS
                variable, the build will not fail if this module is not
                available to include in the image.
            
A common question when working with a kernel is: "What changes have been applied to this tree?" Rather than using "grep" across directories to see what has changed, you can use Git to inspect or search the kernel tree. Using Git is an efficient way to see what has changed in the tree.
Following are a few examples that show how to use Git commands to examine changes. These examples are by no means the only way to see changes.
kernel.org history is blended
                    with Yocto Project kernel changes.
                    You can form ranges by using branch names from the
                    kernel tree as the upper and lower commit markers with
                    the Git commands.
                    You can see the branch names through the web interface
                    to the Yocto Project source repositories at
                    http://git.yoctoproject.org/cgit.cgi.
                
                To see a full range of the changes, use the
                git whatchanged command and specify a
                commit range for the branch
                (commit..commit).
            
                Here is an example that looks at what has changed in the
                emenlow branch of the
                linux-yocto-3.19 kernel.
                The lower commit range is the commit associated with the
                standard/base branch, while
                the upper commit range is the commit associated with the
                standard/emenlow branch.
                
     $ git whatchanged origin/standard/base..origin/standard/emenlow
                
                To see short, one line summaries of changes use the
                git log command:
                
     $ git log --oneline origin/standard/base..origin/standard/emenlow
                
Use this command to see code differences for the changes:
     $ git diff origin/standard/base..origin/standard/emenlow
                
Use this command to see the commit log messages and the text differences:
     $ git show origin/standard/base..origin/standard/emenlow
                
                Use this command to create individual patches for
                each change.
                Here is an example that that creates patch files for each
                commit and places them in your Documents
                directory:
                
     $ git format-patch -o $HOME/Documents origin/standard/base..origin/standard/emenlow
                
                Tags in the Yocto Project kernel tree divide changes for
                significant features or branches.
                The git show tag
                command shows changes based on a tag.
                Here is an example that shows systemtap
                changes:
                
     $ git show systemtap
                
                You can use the
                git branch --contains tag
                command to show the branches that contain a particular feature.
                This command shows the branches that contain the
                systemtap feature:
                
     $ git branch --contains systemtap
                
            You can add kernel features in the
            recipe-space by
            using the
            KERNEL_FEATURES
            variable and by specifying the feature's .scc
            file path in the
            SRC_URI
            statement.
            When you add features using this method, the OpenEmbedded build
            system checks to be sure the features are present.
            If the features are not present, the build stops.
            Kernel features are the last elements processed for configuring
            and patching the kernel.
            Therefore, adding features in this manner is a way
            to enforce specific features are present and enabled
            without needing to do a full audit of any other layer's additions
            to the SRC_URI statement.
        
            You add a kernel feature by providing the feature as part of the
            KERNEL_FEATURES variable and by providing the
            path to the feature's .scc file, which is
            relative to the root of the kernel Metadata.
            The OpenEmbedded build system searches all forms of kernel
            Metadata on the SRC_URI statement regardless
            of whether the Metadata is in the "kernel-cache", system kernel
            Metadata, or a recipe-space Metadata.
            See the
            "Kernel Metadata Location"
            section for additional information.
        
            When you specify the feature's .scc file
            on the SRC_URI statement, the OpenEmbedded
            build system adds the directory of that
            .scc file along with all its subdirectories
            to the kernel feature search path.
            Because subdirectories are searched, you can reference a single
            .scc file in the
            SRC_URI statement to reference multiple kernel
            features.
        
Consider the following example that adds the "test.scc" feature to the build.
                    Create a .scc file and locate it
                    just as you would any other patch file,
                    .cfg file, or fetcher item
                    you specify in the SRC_URI
                    statement.
                    
                                You must add the directory of the
                                .scc file to the fetcher's
                                search path in the same manner as you would
                                add a .patch file.
                                
                                You can create additional
                                .scc files beneath the
                                directory that contains the file you are
                                adding.
                                All subdirectories are searched during the
                                build as potential feature directories.
                                
                    Continuing with the example, suppose the "test.scc"
                    feature you are adding has a
                    test.scc file in the following
                    directory:
                    
     my_recipe
        |
        +-linux-yocto
           |
           +-test.cfg
           +-test.scc
                    
                    In this example, the linux-yocto
                    directory has both the feature
                    test.scc file and a similarly
                    named configuration fragment file
                    test.cfg.
                    
                    Add the .scc file to the
                    recipe's SRC_URI statement:
                    
     SRC_URI_append = " file://test.scc"
                    The leading space before the path is important as the path is appended to the existing path.
Specify the feature as a kernel feature:
     KERNEL_FEATURES_append = " test.scc"
                    The OpenEmbedded build system processes the kernel feature when it builds the kernel.
test.scc file.
                    
Table of Contents
        In addition to supporting configuration fragments and patches, the
        Yocto Project kernel tools also support rich
        Metadata that you can
        use to define complex policies and Board Support Package (BSP) support.
        The purpose of the Metadata and the tools that manage it, known as
        the kern-tools (kern-tools-native_git.bb), is
        to help you manage the complexity of the configuration and sources
        used to support multiple BSPs and Linux kernel types.
    
        The kernel sources in the Yocto Project contain kernel Metadata, which
        is located in the meta branches of the kernel
        source Git repositories.
        This Metadata defines Board Support Packages (BSPs) that
        correspond to definitions in linux-yocto recipes for the same BSPs.
        A BSP consists of an aggregation of kernel policy and enabled
        hardware-specific features.
        The BSP can be influenced from within the linux-yocto recipe.
        
linux-yocto.inc include file is said to be a
            "linux-yocto style" recipe.
        
        Every linux-yocto style recipe must define the
        KMACHINE
        variable.
        This variable is typically set to the same value as the
        MACHINE
        variable, which is used by
        BitBake.
        However, in some cases, the variable might instead refer to the
        underlying platform of the MACHINE.
    
        Multiple BSPs can reuse the same KMACHINE
        name if they are built using the same BSP description.
        The "ep108-zynqmp" and "qemuzynqmp" BSP combination
        in the meta-xilinx
        layer is a good example of two BSPs using the same
        KMACHINE value (i.e. "zynqmp").
        See the BSP Descriptions section
        for more information.
    
        Every linux-yocto style recipe must also indicate the Linux kernel
        source repository branch used to build the Linux kernel.
        The KBRANCH
        variable must be set to indicate the branch.
        
KBRANCH value to define an
            alternate branch typically with a machine override as shown here
            from the meta-emenlow layer:
            
     KBRANCH_emenlow-noemgd = "standard/base"
            
The linux-yocto style recipes can optionally define the following variables:
     KERNEL_FEATURES
     LINUX_KERNEL_TYPE
        
        LINUX_KERNEL_TYPE defines the kernel type to be
        used in assembling the configuration.
        If you do not specify a LINUX_KERNEL_TYPE,
        it defaults to "standard".
        Together with
        KMACHINE,
        LINUX_KERNEL_TYPE defines the search
        arguments used by the kernel tools to find the
        appropriate description within the kernel Metadata with which to
        build out the sources and configuration.
        The linux-yocto recipes define "standard", "tiny", and "preempt-rt"
        kernel types.
        See the "Kernel Types" section
        for more information on kernel types.
    
        During the build, the kern-tools search for the BSP description
        file that most closely matches the KMACHINE
        and LINUX_KERNEL_TYPE variables passed in from the
        recipe.
        The tools use the first BSP description it finds that match
        both variables.
        If the tools cannot find a match, they issue a warning such as
        the following:
        
     WARNING: Can't find any BSP hardware or required configuration fragments.
     WARNING: Looked at meta/cfg/broken/emenlow-broken/hdw_frags.txt and
              meta/cfg/broken/emenlow-broken/required_frags.txt in directory:
              meta/cfg/broken/emenlow-broken
        
        In this example, KMACHINE was set to "emenlow-broken"
        and LINUX_KERNEL_TYPE was set to "broken".
    
        The tools first search for the KMACHINE and
        then for the LINUX_KERNEL_TYPE.
        If the tools cannot find a partial match, they will use the
        sources from the KBRANCH and any configuration
        specified in the
        SRC_URI.
    
        You can use the KERNEL_FEATURES variable
        to include features (configuration fragments, patches, or both) that
        are not already included by the KMACHINE and
        LINUX_KERNEL_TYPE variable combination.
        For example, to include a feature specified as
        "features/netfilter/netfilter.scc",
        specify:
        
     KERNEL_FEATURES += "features/netfilter/netfilter.scc"
        
        To include a feature called "cfg/sound.scc" just for the
        qemux86 machine, specify:
        
     KERNEL_FEATURES_append_qemux86 = " cfg/sound.scc"
        
        The value of the entries in KERNEL_FEATURES
        are dependent on their location within the kernel Metadata itself.
        The examples here are taken from the meta
        branch of the linux-yocto-3.19 repository.
        Within that branch, "features" and "cfg" are subdirectories of the
        meta/cfg/kernel-cache directory.
        For more information, see the
        "Kernel Metadata Syntax" section.
        
Kernel Metadata always exists outside of the kernel tree either defined in a kernel recipe (recipe-space) or outside of the recipe. Where you choose to define the Metadata depends on what you want to do and how you intend to work. Regardless of where you define the kernel Metadata, the syntax used applies equally.
If you are unfamiliar with the Linux kernel and only wish to apply a configuration and possibly a couple of patches provided to you by others, the recipe-space method is recommended. This method is also a good approach if you are working with Linux kernel sources you do not control or if you just do not want to maintain a Linux kernel Git repository on your own. For partial information on how you can define kernel Metadata in the recipe-space, see the "Modifying an Existing Recipe" section.
Conversely, if you are actively developing a kernel and are already maintaining a Linux kernel Git repository of your own, you might find it more convenient to work with kernel Metadata kept outside the recipe-space. Working with Metadata in this area can make iterative development of the Linux kernel more efficient outside of the BitBake environment.
            When stored in recipe-space, the kernel Metadata files reside in a
            directory hierarchy below
            FILESEXTRAPATHS.
            For a linux-yocto recipe or for a Linux kernel recipe derived
            by copying and modifying
            oe-core/meta-skeleton/recipes-kernel/linux/linux-yocto-custom.bb
            to a recipe in your layer, FILESEXTRAPATHS
            is typically set to
            ${THISDIR}/${PN}.
            See the "Modifying an Existing Recipe"
            section for more information.
        
Here is an example that shows a trivial tree of kernel Metadata stored in recipe-space within a BSP layer:
     meta-my_bsp_layer/
     `-- recipes-kernel
         `-- linux
             `-- linux-yocto
                 |-- bsp-standard.scc
                 |-- bsp.cfg
                 `-- standard.cfg
            
            When the Metadata is stored in recipe-space, you must take
            steps to ensure BitBake has the necessary information to decide
            what files to fetch and when they need to be fetched again.
            It is only necessary to specify the .scc
            files on the
            SRC_URI.
            BitBake parses them and fetches any files referenced in the
            .scc files by the include,
            patch, or kconf commands.
            Because of this, it is necessary to bump the recipe
            PR
            value when changing the content of files not explicitly listed
            in the SRC_URI.
        
            When stored outside of the recipe-space, the kernel Metadata
            files reside in a separate repository.
            The OpenEmbedded build system adds the Metadata to the build as
            a "ktype=meta" repository through the
            SRC_URI
            variable.
            As an example, consider the following SRC_URI
            statement from the linux-yocto_4.4.bb
            kernel recipe:
            
     SRC_URI = "git://git.yoctoproject.org/linux-yocto-4.4.git;name=machine;branch=${KBRANCH}; \
                git://git.yoctoproject.org/yocto-kernel-cache;type=kmeta;name=meta;branch=yocto-4.4;destsuffix=${KMETA}"
            
            ${KMETA}, in this context, is simply used to
            name the directory into which the Git fetcher places the Metadata.
            This behavior is no different than any multi-repository
            SRC_URI statement used in a recipe.
        
            You can keep kernel Metadata in a "kernel-cache", which is a
            directory containing configuration fragments.
            As with any Metadata kept outside the recipe-space, you simply
            need to use the SRC_URI statement with the
            "type=kmeta" attribute.
            Doing so makes the kernel Metadata available during the
            configuration phase.
        
            If you modify the Metadata, you must not forget to update the
            SRCREV
            statements in the kernel's recipe.
            In particular, you need to update the
            SRCREV_meta variable to match the commit in
            the KMETA branch you wish to use.
            Changing the data in these branches and not updating the
            SRCREV statements to match will cause the
            build to fetch an older commit.
        
        The kernel Metadata consists of three primary types of files:
        scc
        [1]
        description files, configuration fragments, and patches.
        The scc files define variables and include or
        otherwise reference any of the three file types.
        The description files are used to aggregate all types of kernel
        Metadata into
        what ultimately describes the sources and the configuration required
        to build a Linux kernel tailored to a specific machine.
    
        The scc description files are used to define two
        fundamental types of kernel Metadata:
        
Features
Board Support Packages (BSPs)
Features aggregate sources in the form of patches and configuration fragments into a modular reusable unit. You can use features to implement conceptually separate kernel Metadata descriptions such as pure configuration fragments, simple patches, complex features, and kernel types. Kernel types define general kernel features and policy to be reused in the BSPs.
BSPs define hardware-specific features and aggregate them with kernel types to form the final description of what will be assembled and built.
While the kernel Metadata syntax does not enforce any logical separation of configuration fragments, patches, features or kernel types, best practices dictate a logical separation of these types of Metadata. The following Metadata file hierarchy is recommended:
     base/
        bsp/
        cfg/
        features/
        ktypes/
        patches/
        
        The bsp directory contains the
        BSP descriptions.
        The remaining directories all contain "features".
        Separating bsp from the rest of the structure
        aids conceptualizing intended usage.
    
        Use these guidelines to help place your scc
        description files within the structure:
        
If your file contains
                only configuration fragments, place the file in the
                cfg directory.
If your file contains
                only source-code fixes, place the file in the
                patches directory.
If your file encapsulates
                a major feature, often combining sources and configurations,
                place the file in features directory.
                
If your file aggregates
                non-hardware configuration and patches in order to define a
                base kernel policy or major kernel type to be reused across
                multiple BSPs, place the file in ktypes
                directory.
                
        These distinctions can easily become blurred - especially as
        out-of-tree features slowly merge upstream over time.
        Also, remember that how the description files are placed is
        a purely logical organization and has no impact on the functionality
        of the kernel Metadata.
        There is no impact because all of cfg,
        features, patches, and
        ktypes, contain "features" as far as the kernel
        tools are concerned.
    
        Paths used in kernel Metadata files are relative to
        <base>, which is either
        FILESEXTRAPATHS
        if you are creating Metadata in
        recipe-space,
        or meta/cfg/kernel-cache/ if you are creating
        Metadata outside of the recipe-space.
    
            The simplest unit of kernel Metadata is the configuration-only
            feature.
            This feature consists of one or more Linux kernel configuration
            parameters in a configuration fragment file
            (.cfg) and a .scc file
            that describes the fragment.
        
            The Symmetric Multi-Processing (SMP) fragment included in the
            linux-yocto-3.19 Git repository
            consists of the following two files:
            
     cfg/smp.scc:
        define KFEATURE_DESCRIPTION "Enable SMP"
        define KFEATURE_COMPATIBILITY all
        kconf hardware smp.cfg
     cfg/smp.cfg:
        CONFIG_SMP=y
        CONFIG_SCHED_SMT=y
        # Increase default NR_CPUS from 8 to 64 so that platform with
        # more than 8 processors can be all activated at boot time
        CONFIG_NR_CPUS=64
            You can find information on configuration fragment files in the "Creating Configuration Fragments" section of the Yocto Project Development Manual and in the "Generating Configuration Files" section earlier in this manual.
            KFEATURE_DESCRIPTION
            provides a short description of the fragment.
            Higher level kernel tools use this description.
        
            The kconf command is used to include the
            actual configuration fragment in an .scc
            file, and the "hardware" keyword identifies the fragment as
            being hardware enabling, as opposed to general policy,
            which would use the "non-hardware" keyword.
            The distinction is made for the benefit of the configuration
            validation tools, which warn you if a hardware fragment
            overrides a policy set by a non-hardware fragment.
            
kconf statements, one per fragment.
            
As described in the "Generating Configuration Files" section, you can use the following BitBake command to audit your configuration:
     $ bitbake linux-yocto -c kernel_configcheck -f
            
            Patch descriptions are very similar to configuration fragment
            descriptions, which are described in the previous section.
            However, instead of a .cfg file, these
            descriptions work with source patches.
        
A typical patch includes a description file and the patch itself:
     patches/mypatch.scc:
        patch mypatch.patch
     patches/mypatch.patch:
        typical-patch
            
            You can create the typical .patch
            file using diff -Nurp or
            git format-patch.
        
The description file can include multiple patch statements, one per patch.
            Features are complex kernel Metadata types that consist
            of configuration fragments (kconf), patches
            (patch), and possibly other feature
            description files (include).
        
Here is an example that shows a feature description file:
     features/myfeature.scc
        define KFEATURE_DESCRIPTION "Enable myfeature"
        patch 0001-myfeature-core.patch
        patch 0002-myfeature-interface.patch
        include cfg/myfeature_dependency.scc
        kconf non-hardware myfeature.cfg
            
            This example shows how the patch and
            kconf commands are used as well as
            how an additional feature description file is included.
        
            Typically, features are less granular than configuration
            fragments and are more likely than configuration fragments
            and patches to be the types of things you want to specify
            in the KERNEL_FEATURES variable of the
            Linux kernel recipe.
            See the "Using Kernel Metadata in a Recipe"
            section earlier in the manual.
        
            A kernel type defines a high-level kernel policy by
            aggregating non-hardware configuration fragments with
            patches you want to use when building a Linux kernels of a
            specific type.
            Syntactically, kernel types are no different than features
            as described in the "Features"
            section.
            The LINUX_KERNEL_TYPE variable in the kernel
            recipe selects the kernel type.
            See the "Using Kernel Metadata in a Recipe"
            section for more information.
        
            As an example, the linux-yocto-3.19
            tree defines three kernel types: "standard",
            "tiny", and "preempt-rt":
            
"standard": Includes the generic Linux kernel policy of the Yocto Project linux-yocto kernel recipes. This policy includes, among other things, which file systems, networking options, core kernel features, and debugging and tracing options are supported.
"preempt-rt":
                    Applies the PREEMPT_RT
                    patches and the configuration options required to
                    build a real-time Linux kernel.
                    This kernel type inherits from the "standard" kernel type.
                    
"tiny": Defines a bare minimum configuration meant to serve as a base for very small Linux kernels. The "tiny" kernel type is independent from the "standard" configuration. Although the "tiny" kernel type does not currently include any source changes, it might in the future.
            The "standard" kernel type is defined by
            standard.scc:
            
     # Include this kernel type fragment to get the standard features and
     # configuration values.
     # Include all standard features
     include standard-nocfg.scc
     kconf non-hardware standard.cfg
     # individual cfg block section
     include cfg/fs/devtmpfs.scc
     include cfg/fs/debugfs.scc
     include cfg/fs/btrfs.scc
     include cfg/fs/ext2.scc
     include cfg/fs/ext3.scc
     include cfg/fs/ext4.scc
     include cfg/net/ipv6.scc
     include cfg/net/ip_nf.scc
     include cfg/net/ip6_nf.scc
     include cfg/net/bridge.scc
            
            As with any .scc file, a
            kernel type definition can aggregate other
            .scc files with
            include commands.
            These definitions can also directly pull in
            configuration fragments and patches with the
            kconf and patch
            commands, respectively.
        
.scc file.
            The Board Support Package (BSP) file can implicitly define
            the kernel type using a define
            KTYPE myktype
            line.
            See the "BSP Descriptions"
            section for more information.
        
            BSP descriptions combine kernel types with hardware-specific
            features.
            The hardware-specific portion is typically defined
            independently, and then aggregated with each supported kernel
            type.
            Consider this simple BSP description that supports the
            mybsp machine:
            
     mybsp.scc:
        define KMACHINE mybsp
        define KTYPE standard
        define KARCH i386
        kconf mybsp.cfg
            
            Every BSP description should define the
            KMACHINE,
            KTYPE,
            and KARCH
            variables.
            These variables allow the OpenEmbedded build system to identify
            the description as meeting the criteria set by the recipe being
            built.
            This simple example supports the "mybsp" machine for the "standard"
            kernel and the "i386" architecture.
        
            Be aware that a hard link between the
            KTYPE variable and a kernel type
            description file does not exist.
            Thus, if you do not have kernel types defined in your kernel
            Metadata, you only need to ensure that the kernel recipe's
            LINUX_KERNEL_TYPE
            variable and the KTYPE variable in the
            BSP description file match.
            
KTYPE in the BSP optional.
            
If you did want to separate your kernel policy from your hardware configuration, you could do so by specifying a kernel type, such as "standard" and including that description file in the BSP description file. See the "Kernel Types" section for more information.
            You might also have multiple hardware configurations that you
            aggregate into a single hardware description file that you
            could include in the BSP description file, rather than referencing
            a single .cfg file.
            Consider the following:
            
     mybsp.scc:
        define KMACHINE mybsp
        define KTYPE standard
        define KARCH i386
        include standard.scc
        include mybsp-hw.scc
            
            In the above example, standard.scc
            aggregates all the configuration fragments, patches, and
            features that make up your standard kernel policy whereas
            mybsp-hw.scc
            aggregates all those necessary
            to support the hardware available on the
            mybsp machine.
            For information on how to break a complete
            .config file into the various
            configuration fragments, see the
            "Generating Configuration Files"
            section.
        
            Many real-world examples are more complex.
            Like any other .scc file, BSP
            descriptions can aggregate features.
            Consider the Minnow BSP definition from the
            linux-yocto-3.19
            Git repository:
            
     minnow.scc:
        include cfg/x86.scc
        include features/eg20t/eg20t.scc
        include cfg/dmaengine.scc
        include features/power/intel.scc
        include cfg/efi.scc
        include features/usb/ehci-hcd.scc
        include features/usb/ohci-hcd.scc
        include features/usb/usb-gadgets.scc
        include features/usb/touchscreen-composite.scc
        include cfg/timer/hpet.scc
        include cfg/timer/rtc.scc
        include features/leds/leds.scc
        include features/spi/spidev.scc
        include features/i2c/i2cdev.scc
        # Earlyprintk and port debug requires 8250
        kconf hardware cfg/8250.cfg
        kconf hardware minnow.cfg
        kconf hardware minnow-dev.cfg
            
            The minnow.scc description file includes
            a hardware configuration fragment
            (minnow.cfg) specific to the Minnow
            BSP as well as several more general configuration
            fragments and features enabling hardware found on the
            machine.
            This description file is then included in each of the three
            "minnow" description files for the supported kernel types
            (i.e. "standard", "preempt-rt", and "tiny").
            Consider the "minnow" description for the "standard" kernel
            type:
            
     minnow-standard.scc:
        define KMACHINE minnow
        define KTYPE standard
        define KARCH i386
        include ktypes/standard
        include minnow.scc
        # Extra minnow configs above the minimal defined in minnow.scc
        include cfg/efi-ext.scc
        include features/media/media-all.scc
        include features/sound/snd_hda_intel.scc
        # The following should really be in standard.scc
        # USB live-image support
        include cfg/usb-mass-storage.scc
        include cfg/boot-live.scc
        # Basic profiling
        include features/latencytop/latencytop.scc
        include features/profiling/profiling.scc
        # Requested drivers that don't have an existing scc
        kconf hardware minnow-drivers-extra.cfg
            
            The include command midway through the file
            includes the minnow.scc description that
            defines all hardware enablements for the BSP that is common to all
            kernel types.
            Using this command significantly reduces duplication.
        
Now consider the "minnow" description for the "tiny" kernel type:
     minnow-tiny.scc:
        define KMACHINE minnow
        define KTYPE tiny
        define KARCH i386
        include ktypes/tiny
        include minnow.scc
            As you might expect, the "tiny" description includes quite a bit less. In fact, it includes only the minimal policy defined by the "tiny" kernel type and the hardware-specific configuration required for booting the machine along with the most basic functionality of the system as defined in the base "minnow" description file.
            Notice again the three critical variables:
            KMACHINE, KTYPE,
            and KARCH.
            Of these variables, only the KTYPE has changed.
            It is now set to "tiny".
        
        Many recipes based on the linux-yocto-custom.bb
        recipe use Linux kernel sources that have only a single
        branch - "master".
        This type of repository structure is fine for linear development
        supporting a single machine and architecture.
        However, if you work with multiple boards and architectures,
        a kernel source repository with multiple branches is more
        efficient.
        For example, suppose you need a series of patches for one board to boot.
        Sometimes, these patches are works-in-progress or fundamentally wrong,
        yet they are still necessary for specific boards.
        In these situations, you most likely do not want to include these
        patches in every kernel you build (i.e. have the patches as part of
        the lone "master" branch).
        It is situations like these that give rise to multiple branches used
        within a Linux kernel sources Git repository.
    
Repository organization strategies exist that maximize source reuse, remove redundancy, and logically order your changes. This section presents strategies for the following cases:
Encapsulating patches in a feature description and only including the patches in the BSP descriptions of the applicable boards.
Creating a machine branch in your kernel source repository and applying the patches on that branch only.
Creating a feature branch in your kernel source repository and merging that branch into your BSP when needed.
The approach you take is entirely up to you and depends on what works best for your development model.
if you are reusing patches from an external tree and are not working on the patches, you might find the encapsulated feature to be appropriate. Given this scenario, you do not need to create any branches in the source repository. Rather, you just take the static patches you need and encapsulate them within a feature description. Once you have the feature description, you simply include that into the BSP description as described in the "BSP Descriptions" section.
You can find information on how to create patches and BSP descriptions in the "Patches" and "BSP Descriptions" sections.
When you have multiple machines and architectures to support, or you are actively working on board support, it is more efficient to create branches in the repository based on individual machines. Having machine branches allows common source to remain in the "master" branch with any features specific to a machine stored in the appropriate machine branch. This organization method frees you from continually reintegrating your patches into a feature.
            Once you have a new branch, you can set up your kernel Metadata
            to use the branch a couple different ways.
            In the recipe, you can specify the new branch as the
            KBRANCH to use for the board as
            follows:
            
     KBRANCH = "mynewbranch"
            
            Another method is to use the branch command
            in the BSP description:
            
     mybsp.scc:
        define KMACHINE mybsp
        define KTYPE standard
        define KARCH i386
        include standard.scc
        branch mynewbranch
        include mybsp-hw.scc
            
If you find yourself with numerous branches, you might consider using a hierarchical branching system similar to what the linux-yocto Linux kernel repositories use:
     common/kernel_type/machine
            
            If you had two kernel types, "standard" and "small" for
            instance, three machines, and common
            as mydir, the branches in your
            Git repository might look like this:
            
     mydir/base
     mydir/standard/base
     mydir/standard/machine_a
     mydir/standard/machine_b
     mydir/standard/machine_c
     mydir/small/base
     mydir/small/machine_a
            
            This organization can help clarify the branch relationships.
            In this case, mydir/standard/machine_a
            includes everything in mydir/base and
            mydir/standard/base.
            The "standard" and "small" branches add sources specific to those
            kernel types that for whatever reason are not appropriate for the
            other branches.
            
mydir/standard and
                mydir/standard/machine_a because it
                would have to create a file and a directory named "standard".
            
            When you are actively developing new features, it can be more
            efficient to work with that feature as a branch, rather than
            as a set of patches that have to be regularly updated.
            The Yocto Project Linux kernel tools provide for this with
            the git merge command.
        
            To merge a feature branch into a BSP, insert the
            git merge command after any
            branch commands:
            
     mybsp.scc:
        define KMACHINE mybsp
        define KTYPE standard
        define KARCH i386
        include standard.scc
        branch mynewbranch
        git merge myfeature
        include mybsp-hw.scc
            
        This section provides a brief reference for the commands you can use
        within an SCC description file (.scc):
        
branch [ref]:
                Creates a new branch relative to the current branch
                (typically ${KTYPE}) using
                the currently checked-out branch, or "ref" if specified.
                
define:
                Defines variables, such as KMACHINE,
                KTYPE, KARCH,
                and KFEATURE_DESCRIPTION.
include SCC_FILE:
                Includes an SCC file in the current file.
                The file is parsed as if you had inserted it inline.
                
kconf [hardware|non-hardware] CFG_FILE:
                Queues a configuration fragment for merging into the final
                Linux .config file.
git merge GIT_BRANCH:
                Merges the feature branch into the current branch.
                
patch PATCH_FILE:
                Applies the patch to the current Git branch.
[1] 
                scc stands for Series Configuration
                Control, but the naming has less significance in the
                current implementation of the tooling than it had in the
                past.
                Consider scc files to be description files.
            
Table of Contents
            Kernels available through the Yocto Project, like other kernels, are based off the Linux
            kernel releases from http://www.kernel.org.
            At the beginning of a major development cycle, the Yocto Project team
            chooses its kernel based on factors such as release timing, the anticipated release
            timing of final upstream kernel.org versions, and Yocto Project
            feature requirements.
            Typically, the kernel chosen is in the
            final stages of development by the community.
            In other words, the kernel is in the release
            candidate or "rc" phase and not yet a final release.
            But, by being in the final stages of external development, the team knows that the
            kernel.org final release will clearly be within the early stages of
            the Yocto Project development window.
        
This balance allows the team to deliver the most up-to-date kernel possible, while still ensuring that the team has a stable official release for the baseline Linux kernel version.
            The ultimate source for kernels available through the Yocto Project are released kernels
            from kernel.org.
            In addition to a foundational kernel from kernel.org, the
            kernels available contain a mix of important new mainline
            developments, non-mainline developments (when there is no alternative),
            Board Support Package (BSP) developments,
            and custom features.
            These additions result in a commercially released Yocto Project Linux kernel that caters
            to specific embedded designer needs for targeted hardware.
        
Once a kernel is officially released, the Yocto Project team goes into their next development cycle, or upward revision (uprev) cycle, while still continuing maintenance on the released kernel. It is important to note that the most sustainable and stable way to include feature development upstream is through a kernel uprev process. Back-porting hundreds of individual fixes and minor features from various kernel versions is not sustainable and can easily compromise quality.
            During the uprev cycle, the Yocto Project team uses an ongoing analysis of
            kernel development, BSP support, and release timing to select the best
            possible kernel.org version.
            The team continually monitors community kernel
            development to look for significant features of interest.
            The team does consider back-porting large features if they have a significant advantage.
            User or community demand can also trigger a back-port or creation of new
            functionality in the Yocto Project baseline kernel during the uprev cycle.
        
Generally speaking, every new kernel both adds features and introduces new bugs. These consequences are the basic properties of upstream kernel development and are managed by the Yocto Project team's kernel strategy. It is the Yocto Project team's policy to not back-port minor features to the released kernel. They only consider back-porting significant technological jumps - and, that is done after a complete gap analysis. The reason for this policy is that back-porting any small to medium sized change from an evolving kernel can easily create mismatches, incompatibilities and very subtle errors.
These policies result in both a stable and a cutting edge kernel that mixes forward ports of existing features and significant and critical new functionality. Forward porting functionality in the kernels available through the Yocto Project kernel can be thought of as a "micro uprev." The many “micro uprevs” produce a kernel version with a mix of important new mainline, non-mainline, BSP developments and feature integrations. This kernel gives insight into new features and allows focused amounts of testing to be done on the kernel, which prevents surprises when selecting the next major uprev. The quality of these cutting edge kernels is evolving and the kernels are used in leading edge feature and BSP development.
This section describes the architecture of the kernels available through the Yocto Project and provides information on the mechanisms used to achieve that architecture.
                As mentioned earlier, a key goal of the Yocto Project is to present the
                developer with
                a kernel that has a clear and continuous history that is visible to the user.
                The architecture and mechanisms used achieve that goal in a manner similar to the
                upstream kernel.org.
            
You can think of a Yocto Project kernel as consisting of a baseline Linux kernel with added features logically structured on top of the baseline. The features are tagged and organized by way of a branching strategy implemented by the source code manager (SCM) Git. For information on Git as applied to the Yocto Project, see the "Git" section in the Yocto Project Development Manual.
The result is that the user has the ability to see the added features and the commits that make up those features. In addition to being able to see added features, the user can also view the history of what made up the baseline kernel.
The following illustration shows the conceptual Yocto Project kernel.
|  | 
In the illustration, the "Kernel.org Branch Point" marks the specific spot (or release) from which the Yocto Project kernel is created. From this point "up" in the tree, features and differences are organized and tagged.
The "Yocto Project Baseline Kernel" contains functionality that is common to every kernel type and BSP that is organized further up the tree. Placing these common features in the tree this way means features do not have to be duplicated along individual branches of the structure.
From the Yocto Project Baseline Kernel, branch points represent specific functionality for individual BSPs as well as real-time kernels. The illustration represents this through three BSP-specific branches and a real-time kernel branch. Each branch represents some unique functionality for the BSP or a real-time kernel.
In this example structure, the real-time kernel branch has common features for all real-time kernels and contains more branches for individual BSP-specific real-time kernels. The illustration shows three branches as an example. Each branch points the way to specific, unique features for a respective real-time kernel as they apply to a given BSP.
The resulting tree structure presents a clear path of markers (or branches) to the developer that, for all practical purposes, is the kernel needed for any given set of requirements.
The Yocto Project team creates kernel branches at points where functionality is no longer shared and thus, needs to be isolated. For example, board-specific incompatibilities would require different functionality and would require a branch to separate the features. Likewise, for specific kernel features, the same branching strategy is used.
This branching strategy results in a tree that has features organized to be specific for particular functionality, single kernel types, or a subset of kernel types. This strategy also results in not having to store the same feature twice internally in the tree. Rather, the kernel team stores the unique differences required to apply the feature onto the kernel type in question.
BSP-specific code additions are handled in a similar manner to kernel-specific additions. Some BSPs only make sense given certain kernel types. So, for these types, the team creates branches off the end of that kernel type for all of the BSPs that are supported on that kernel type. From the perspective of the tools that create the BSP branch, the BSP is really no different than a feature. Consequently, the same branching strategy applies to BSPs as it does to features. So again, rather than store the BSP twice, the team only stores the unique differences for the BSP across the supported multiple kernels.
                While this strategy can result in a tree with a significant number of branches, it is
                important to realize that from the developer's point of view, there is a linear
                path that travels from the baseline kernel.org, through a select
                group of features and ends with their BSP-specific commits.
                In other words, the divisions of the kernel are transparent and are not relevant
                to the developer on a day-to-day basis.
                From the developer's perspective, this path is the "master" branch.
                The developer does not need to be aware of the existence of any other branches at all.
                Of course, there is value in the existence of these branches
                in the tree, should a person decide to explore them.
                For example, a comparison between two BSPs at either the commit level or at the line-by-line
                code diff level is now a trivial operation.
            
Working with the kernel as a structured tree follows recognized community best practices. In particular, the kernel as shipped with the product, should be considered an "upstream source" and viewed as a series of historical and documented modifications (commits). These modifications represent the development and stabilization done by the Yocto Project kernel development team.
Because commits only change at significant release points in the product life cycle, developers can work on a branch created from the last relevant commit in the shipped Yocto Project kernel. As mentioned previously, the structure is transparent to the developer because the kernel tree is left in this state after cloning and building the kernel.
                The Source Code Manager (SCM) is Git.
                This SCM is the obvious mechanism for meeting the previously mentioned goals.
                Not only is it the SCM for kernel.org but,
                Git continues to grow in popularity and supports many different work flows,
                front-ends and management techniques.
            
You can find documentation on Git at http://git-scm.com/documentation. You can also get an introduction to Git as it applies to the Yocto Project in the "Git" section in the Yocto Project Development Manual. These referenced sections overview Git and describe a minimal set of commands that allows you to be functional using Git.
Table of Contents
This section describes construction of the Yocto Project kernel source repositories as accomplished by the Yocto Project team to create kernel repositories. These kernel repositories are found under the heading "Yocto Linux Kernel" at http://git.yoctoproject.org/cgit.cgi and can be shipped as part of a Yocto Project release. The team creates these repositories by compiling and executing the set of feature descriptions for every BSP and feature in the product. Those feature descriptions list all necessary patches, configuration, branching, tagging and feature divisions found in a kernel. Thus, the Yocto Project kernel repository (or tree) is built.
The existence of this tree allows you to access and clone a particular Yocto Project kernel repository and use it to build images based on their configurations and features.
            You can find the files used to describe all the valid features and BSPs
            in the Yocto Project kernel in any clone of the Yocto Project kernel source repository
            Git tree.
            For example, the following command clones the Yocto Project baseline kernel that
            branched off of linux.org version 3.19:
            
     $ git clone git://git.yoctoproject.org/linux-yocto-3.19
            For another example of how to set up a local Git repository of the Yocto Project kernel files, see the "Yocto Project Kernel" bulleted item in the Yocto Project Development Manual.
            Once you have cloned the kernel Git repository on your local machine, you can
            switch to the meta branch within the repository.
            Here is an example that assumes the local Git repository for the kernel is in
            a top-level directory named linux-yocto-3.19:
            
     $ cd linux-yocto-3.19
     $ git checkout -b meta origin/meta
            
            Once you have checked out and switched to the meta branch,
            you can see a snapshot of all the kernel configuration and feature descriptions that are
            used to build that particular kernel repository.
            These descriptions are in the form of .scc files.
        
You should realize, however, that browsing your local kernel repository for feature descriptions and patches is not an effective way to determine what is in a particular kernel branch. Instead, you should use Git directly to discover the changes in a branch. Using Git is an efficient and flexible way to inspect changes to the kernel.
The following steps describe what happens when the Yocto Project Team constructs the Yocto Project kernel source Git repository (or tree) found at http://git.yoctoproject.org/cgit.cgi given the introduction of a new top-level kernel feature or BSP. These are the actions that effectively create the tree that includes the new feature, patch or BSP:
A top-level kernel feature is passed to the kernel build subsystem. Normally, this feature is a BSP for a particular kernel type.
The file that describes the top-level feature is located by searching these system directories:
The in-tree kernel-cache directories, which are located
                            in meta/cfg/kernel-cache
Areas pointed to by SRC_URI statements
                            found in recipes
                    For a typical build, the target of the search is a
                    feature description in an .scc file
                    whose name follows this format:
                    
     bsp_name-kernel_type.scc
                    
Once located, the feature description is either compiled into a simple script of actions, or into an existing equivalent script that is already part of the shipped kernel.
Extra features are appended to the top-level feature description.
                    These features can come from the
                    KERNEL_FEATURES
                    variable in recipes.
Each extra feature is located, compiled and appended to the script as described in step three.
The script is executed to produce a series of meta-*
                    directories.
                    These directories are descriptions of all the branches, tags, patches and configurations that
                    need to be applied to the base Git repository to completely create the
                    source (build) branch for the new BSP or feature.
The base repository is cloned, and the actions
                    listed in the meta-* directories are applied to the
                    tree.
The Git repository is left with the desired branch checked out and any required branching, patching and tagging has been performed.
The kernel tree is now ready for developer consumption to be locally cloned, configured, and built into a Yocto Project kernel specific to some target hardware.
The generated meta-* directories add to the kernel
                as shipped with the Yocto Project release.
                Any add-ons and configuration data are applied to the end of an existing branch.
                The full repository generation that is found in the
                official Yocto Project kernel repositories at
                http://git.yoctoproject.org/cgit.cgi
                is the combination of all supported boards and configurations.
The technique the Yocto Project team uses is flexible and allows for seamless blending of an immutable history with additional patches specific to a deployment. Any additions to the kernel become an integrated part of the branches.
Once a local Git repository of the Yocto Project kernel exists on a development system, you can consider the compilation phase of kernel development - building a kernel image. Some prerequisites exist that are validated by the build process before compilation starts:
The
                SRC_URI points
                to the kernel Git repository.
A BSP build branch exists. This branch has the following form:
     kernel_type/bsp_name
                The OpenEmbedded build system makes sure these conditions exist before attempting compilation. Other means, however, do exist, such as as bootstrapping a BSP.
            Before building a kernel, the build process verifies the tree
            and configures the kernel by processing all of the
            configuration "fragments" specified by feature descriptions in the .scc
            files.
            As the features are compiled, associated kernel configuration fragments are noted
            and recorded in the meta-* series of directories in their compilation order.
            The fragments are migrated, pre-processed and passed to the Linux Kernel
            Configuration subsystem (lkc) as raw input in the form
            of a .config file.
            The lkc uses its own internal dependency constraints to do the final
            processing of that information and generates the final .config file
            that is used during compilation.
        
Using the board's architecture and other relevant values from the board's template, kernel compilation is started and a kernel image is produced.
            The other thing that you notice once you configure a kernel is that
            the build process generates a build tree that is separate from your kernel's local Git
            source repository tree.
            This build tree has a name that uses the following form, where
            ${MACHINE} is the metadata name of the machine (BSP) and "kernel_type" is one
            of the Yocto Project supported kernel types (e.g. "standard"):
        
     linux-${MACHINE}-kernel_type-build
        
            The existing support in the kernel.org tree achieves this
            default functionality.
        
            This behavior means that all the generated files for a particular machine or BSP are now in
            the build tree directory.
            The files include the final .config file, all the .o
            files, the .a files, and so forth.
            Since each machine or BSP has its own separate
            Build Directory
            in its own separate branch
            of the Git repository, you can easily switch between different builds.
        
Table of Contents
The following lists some solutions for common questions.
| C.1.1. | 
                        How do I use my own Linux kernel  | 
| Refer to the "Changing the Configuration" section for information. | |
| C.1.2. | How do I create configuration fragments? | 
| Refer to the "Generating Configuration Files" section for information. | |
| C.1.3. | How do I use my own Linux kernel sources? | 
| Refer to the "Working With Your Own Sources" section for information. | |
| C.1.4. | How do I install/not-install the kernel image on the rootfs? | 
| 
                        The kernel image (e.g.  See the "Using .bbappend Files" section in the Yocto Project Development Manual for information on how to use an append file to override metadata. | |
| C.1.5. | How do I install a specific kernel module? | 
| 
                        Linux kernel modules are packaged individually.
                        To ensure a specific kernel module is included in an image,
                        include it in the appropriate machine
                         These other variables are useful for installing specific modules: 
     
                        For example, set the following in the  
     MACHINE_EXTRA_RRECOMMENDS += "kernel-module-ab123"
                        For more information, see the "Incorporating Out-of-Tree Modules" section. | |
| C.1.6. | How do I change the Linux kernel command line? | 
| 
                        The Linux kernel command line is typically specified in
                        the machine config using the  
     APPEND += "printk.time=y initcall_debug debug"
                        
 |