Linux, Open Source & Unrelated Topics by Paul Barker

Rebooting meta-linux-mainline

Published:
Tags: yocto, embedded-linux, meta-linux-mainline

Introduction

The meta-linux-mainline Yocto Project layer contains recipes for building the Linux kernel from unmodified sources as released on kernel.org. It can be used to develop a BSP which uses a mainline kernel by default, to replace the default vendor kernel in an existing BSP (which may be obsolete, insecure or otherwise broken), to support upstream kernel development or simply for testing. It provides linux-stable recipes for all stable release series (including LTS releases) currently supported upstream as well as a linux-mainline recipe for those who want to live on the bleeding edge. It is compatible with all currently supported Yocto Project releases ("dunfell" and "hardknott" at the time of this post) as well as the Yocto Project master branch. Further details can be found in the project's readme file.

Use Cases

The meta-linux-mainline layer has two primary use cases: supporting the development of new BSPs which use upstream kernel sources by default and overriding the kernel recipes in existing BSPs to use upstream kernel sources instead of vendor kernel sources.

When developing a new Yocto Project BSP for a hardware platform supported by the mainline kernel, it should not be necessary to maintain your own mainline kernel recipe. Users should be given the option of using the latest stable kernel, an LTS release series or even a bleeding-edge mainline kernel without the BSP maintainer needing to implement all these options. Users should also get regular updates to the kernel recipes without needing to distract the maintainer from focusing on the hardware-specific details of their BSP. These objectives can be achieved by using meta-linux-mainline as a dependency of the BSP layer and where necessary specifying the earliest mainline kernel version required to support the target hardware.

Many existing BSPs default to the use of a "vendor kernel" which incorporates many (sometimes several thousand) patches which have not been subject to upstream review, testing and integration by the kernel community. In many cases the target hardware is supported well enough by the mainline kernel for the intended use-case without the need for such patching. This can be particularly frustrating when the vendor kernel in question is obsolete, doesn't receive security updates or introduces compatibility issues. To switch away from a vendor kernel, the meta-linux-mainline layer can be added to the build alongside the relevant BSP layer(s) and the relevant upstream kernel can be selected in the local or distro config.

In both of these cases the linux-yocto recipes present in the core Yocto Project metadata could be used instead of the recipes in meta-linux-mainline. However, linux-yocto recipes are typically provided for a narrower set of kernel releases than those currently supported upstream and recipes are not added to stable Yocto Project branches for newer kernel release series (as the kernel community places an incredibly high value on backwards compatibility it is generally safe to update to new stable kernel releases). The linux-yocto kernels also include many patches which may not have been through a full upstream review by the kernel community. Using unpatched upstream kernel sources also has major benefits when supporting multiple Linux distributions on the target hardware as it is then possible to standardise on the upstream kernel. Lastly, if you have customer requirements or preferences for a mainline kernel then these can be met using the meta-linux-mainline layer.

This layer also supports two other secondary use cases. The mainline kernel recipe provided in this layer can be used to support upstream kernel development as it can be easily modified to point to an alternative source repository and branch or commit. The recipes may also be used as part of a regular testing process to ensure that an embedded device works as expected with new upstream kernel releases.

Defining Goals and non-Goals

For a project like meta-linux-mainline to succeed, it needs a clearly defined set of goals and non-goals. Goals identify the features and attributes which I want to see in this project, non-goals identify things which may in theory be possible to achieve but which I have chosen to exclude from the scope of this project and which will not be implemented or accepted as contributions without a major change to the project's scope. The goals are obviously important, they should be aligned with the intended use cases for the layer and drive the project forwards. The non-goals for a project are often overlooked but I think they are equally important, they help the project to avoid bloat and stay on track. Potential contributors can review the project's goals and non-goals and find out upfront if their changes are likely to be accepted into the project. The goals and non-goals for meta-linux-mainline are listed prominently in the project's readme file.

The current goals for meta-linux-mainline are as follows:

  • We provide recipes for all Linux kernel releases currently supported on kernel.org. These recipes are regularly updated to make it easy to follow mainline releases, the latest stable series or a chosen LTS release series.
  • We aim to be compatible with all currently supported Yocto Project releases as well as the upstream master branch.
  • We provide examples of how to use this layer in the form of BSP configurations for various QEMU and Raspberry Pi targets.

The current non-goals for meta-linux-mainline are as follows:

  • We do not carry patches against upstream kernel releases without a documented, exceptionally good reason.
  • We do not support obsolete kernel versions. Recipes are only provided for the latest patch release within a given release series. Once a release series becomes End-Of-Life (EOL) on kernel.org, the corresponding recipe will be removed from this layer.
  • We provide no guarantees that kernels built with this layer will boot successfully on your hardware or that particular features (e.g. perf) will work out of the box. The example BSP configurations are not intended to be directly used in production. To use this layer in production, create your own layer for configuration & integration and use this layer as a dependency.
  • We do not aim to replace the linux-yocto kernel from the Yocto Project.

Recent Changes

Project maintenance is now focused on a single main branch which aims to be compatible with the master branch and all currently maintained releases of the Yocto Project. The "master" and stable branches ("dunfell", etc) of meta-linux-mainline will simply follow the "main" branch. This simplification, along with improved automation of updates to the kernel recipes, should result in more regular updates to this layer while ensuring that the level of maintainer effort required remains small and sustainable.

Several other changes have been made to the project as highlighted in the project's ChangeLog. These include switching the example hardware BSP to Raspberry Pi 4, adding more QEMU example targets, switching the default LTS kernel series to 5.10, improving how stable kernels are downloaded and overhauling the scripts used to test and update this layer.

Future Plans

At this point the meta-linux-mainline layer is in a pretty good shape and meets the use cases which I have. Project maintenance is expected to be fairly straightforward as updates to the kernel recipes are fully automated. There are no major changes expected in the near future, the project will just tick over with minor improvements and regular recipe updates as needed. If you have any feature requests, please feel free to submit them via the issue tracker.

At some point I would like to see recipes for RT kernels added to the layer. This isn't something I immediately need myself, so I'd encourage anyone who has an immediate need for vanilla RT kernel recipes to contribute this feature to the project. I'm actually hoping that by the time I next need to play with realtime features I'll find that the RT patches have been merged fully into mainline Linux and no separate kernel recipes are actually needed in this layer.

The next Yocto Project release, 3.4 "honister", is expected in October this year. It's expected that upstream support for the 3.3 "hardknott" release series will end in November. This layer will be updated around those times to add support for the "honister" release and remove support for the "hardknott" release. The current Yocto Project LTS release, 3.1 "dunfell", is expected to be supported upstream until at least April 2022 and my intention is to continue supporting the "dunfell" release in this layer until upstream support ends.