The traditional Gentoo way of getting a kernel is to install the sources, and then configure and build one yourself. For those who didn’t want to go through the tedious process of configuring it manually, an alternative route of using genkernel was provided. However, neither of those variants was able to really provide the equivalent of kernels provided by binary distributions.
I have manually configured the kernels for my private systems long time ago. Today, I wouldn’t really have bothered. In fact, I realized that for some time I’m really hesitant to even upgrade them because of the effort needed to update configuration. The worst part is, whenever a new kernel does not boot, I have to ask myself: is it a real bug, or is it my fault for configuring it wrong?
I’m not alone in this. Recently Михаил Коляда (zlogene) has talked to me about providing binary kernels for Gentoo. While I have not strictly implemented what he had in mind, he inspired me to start working on a distribution kernel. The goal was to create a kernel package that users can install to get a working kernel with minimal effort, and that would be upgraded automatically as part of regular @world upgrades.
Pros and cons of your own kernel
If I am to justify switching from the old tradition of custom kernels to a universal kernel package, I should start by discussing the reasons why you may want to configure a custom kernel in the first place.
In my opinion, the most important feature of a custom kernel is that you can fine-tune it to your hardware. You just have to build the drivers you need (or may need), and the features you care about. The modules for my last custom kernel have occupied 44 MiB. The modules for the distribution kernel occupy 294 MiB. Such a difference in size also comes with a proportional increase of build time. This can be an important argument for people with low-end hardware. On the other hand, the distribution kernel permits building reusable binary packages that can save more computing power.
The traditional Gentoo argument is performance. However, these days I would be very careful arguing about that. I suppose you are able to reap benefits if you know how to configure your kernel towards a specific workload. But then — a misconfiguration can have the exact opposite effect. We must not forget that binary distributions are important players in the field — and the kernel must also be able to achieve good performance when not using a dedicated configuration.
At some point I have worked on achieving a very fast startup. For this reason I’ve switched to using LILO as the bootloader, and a kernel suitable for booting my system without an initramfs. A universal kernel naturally needs an initramfs, and is slower to boot.
The main counterargument is the effort. As mentioned earlier, I’ve personally grown tired of having to manually deal with my kernel. Do the potential gains mentioned outweigh the loss of human time on configuring and maintaining a custom kernel?
Creating a truly universal kernel
A distribution kernel makes sense only if it works on a wide range of systems. Furthermore, I didn’t forget the original idea of binary kernel packages. I didn’t want to write an ebuild that can install a working kernel anywhere. I wanted to create an ebuild that can be used to build a binary package that’s going to work on a wide range of setups — including not only different hardware but also bootloaders and /boot layout. A package that would work fine both for my ‘traditional’ LILO setup and UEFI systemd-boot setup.
The first part of a distribution kernel is the right configuration. I wanted to use a well-tested configuration known to build kernels used on many systems, while at the same time minimizing the maintenance effort on our end. Reusing the configuration from a binary distro was the obvious solution. I went for using the config from Arch Linux’s kernel package with minimal changes (e.g. changing the default hostname to Gentoo).
The second part is an initramfs. Since we need to support a wide variety of setups, we can’t get away without it. To follow the configuration used, Dracut was the natural choice.
The third and hardest part is installing it all. Since I’ve already set a goal of reusing the same binary package on different filesystem layouts, the actual installation needed to be moved to postinst phase. Our distribution kernel package installs the kernel into an interim location which is entirely setup-independent, rendering the binary packages setup-agnostic as well. The initramfs is created and installed into the final location along with the kernel in pkg_postinst.
Support for different install layouts is provided by reusing the installkernel tool, originally installed by debianutils. As part of the effort, it was extended with initramfs support and moved into a separate sys-kernel/installkernel-gentoo package. Furthermore, an alternative sys-kernel/installkernel-systemd-boot package was created to provide an out-of-the-box support for systemd-boot layout. If neither of those two work for you, you can easily create your own /usr/local/bin/installkernel that follows your own layout.
The experimental versions of the distribution kernel are packaged as sys-kernel/vanilla-kernel (in distinction from sys-kernel/vanilla-sources that install the sources). Besides providing the default zero-effort setup, the package supports using your own configuration via savedconfig (but no easy way to update it at the moment). It also provides a forced flag that can be used by expert users to disable the initramfs.
The primary goal at the moment is to test the package and find bugs that could prevent our users from using it. In the future, we’re planning to extend it to other architectures, kernel variants (Gentoo patch set in particular) and LTS versions. We’re also considering providing prebuilt binary packages — however, this will probably be a part of a bigger effort into providing an official Gentoo binhost.