↓Skip to main content
  1. Notes/

Building a Kernel Lab, Then a Plan to Use It

·4 mins
Table of Contents

I’ve wanted to contribute to the Linux kernel for a while. The gap was never desire — it was that I had no safe place to break things. This is the note of closing that gap in an afternoon, and the plan I wrote so the momentum doesn’t evaporate.

First, finish what was half-done #

The lab box was supposed to be on Debian 13 (trixie). It wasn’t — not really. base-files reported trixie, but sources.list still pointed at bookworm and 401 packages were stranded at Debian 12 versions. A dist-upgrade had stalled partway and had its sources quietly reverted. A split-brain machine.

The lesson I keep relearning with this box: verify state from the box, never from memory. The saved notes said “cleanly on trixie.” The live apt update output said otherwise.

The fix was a staged, dry-run-first upgrade:

sudo apt-get -s full-upgrade | grep '^Remv'   # read the removals BEFORE committing
sudo apt-get upgrade --without-new-pkgs
sudo apt-get full-upgrade                       # 1332 upgraded, 405 new, 118 removed
sudo reboot

The 118 removals were t64 library renames and expected transitions, not data loss — which is exactly why you read the list before you run it live. After the reboot:

$ cat /etc/debian_version
13.6
$ uname -r
6.12.95+deb13-amd64

0 not upgraded is the number that tells you the job is clean.

Turning on the hypervisor #

The virt stack came across in the upgrade — libvirt 11.3, QEMU 10.0 — it just wasn’t switched on. Three small things stood between me and running VMs:

sudo usermod -aG kvm,libvirt tester        # permission to touch /dev/kvm
sudo systemctl start libvirtd
sudo virsh net-start default && sudo virsh net-autostart default

One gotcha cost a few minutes: virsh net-list came back empty even though the network clearly existed. The unprivileged virsh talks to a per-user daemon (qemu:///session); the network lives in the system one. Point it at the right URI and it’s there:

$ virsh -c qemu:///system net-list --all
 Name      State    Autostart   Persistent
--------------------------------------------
 default   active   yes         yes

A kernel that boots in a second #

Here’s the part I actually enjoy. To test a kernel you don’t need a full OS install — you need a root filesystem with one program in it. So I built a minimal BusyBox initramfs: a ~1.5 MB gzipped cpio archive containing BusyBox, its shared libraries, and an /init script that mounts the basics and hands you a shell.

#!/bin/busybox sh
/bin/busybox --install -s /bin
mount -t proc none /proc
mount -t sysfs none /sys
mount -t devtmpfs none /dev
exec setsid cttyhack sh

Pack it, then boot any compiled kernel against it:

qemu-system-x86_64 -enable-kvm -m 512 -nographic -no-reboot \
  -kernel arch/x86/boot/bzImage -initrd initramfs.cpio.gz \
  -append "console=ttyS0 rdinit=/init"

The payoff:

[    0.000000] Linux version 6.12.95+deb13-amd64 ...
[    0.000000] Hypervisor detected: KVM
...
=== busybox initramfs up — kernel: 6.12.95+deb13-amd64 ===
~ #

From make to a shell in about a second, in RAM, with nothing on disk to corrupt. When I break a kernel now, I break it here.

The harder part: not stopping #

Setup is the easy dopamine. The failure mode for a project like this is week three, when novelty runs out. So before touching a single line of kernel code I wrote the plan — twelve weeks, three phases:

  1. Foundations — shell fluency, C refresh, git and the patch workflow, and building/booting that first kernel.
  2. Internals — LFD103, writing loadable modules, and reading real subsystem code until it stops looking like hieroglyphics.
  3. Contribution — find a genuine small fix in drivers/staging/, run checkpatch.pl until it’s silent, and git send-email it to the right maintainer.

It leans almost entirely on free material — the Linux Foundation’s LFD103, KernelNewbies, Bootlin’s slides, LWN, and the kernel’s own Documentation/process/.

The one design decision I’m proud of is a recovery rule baked into the plan itself: don’t restart the clock, resume the checkbox. Missing days is assumed, not punished. On a low-energy day the bar is one git pull && make-sized thing — read a doc section, tick one box. Momentum beats intensity, and a plan that shames you for slipping is a plan you’ll abandon.

I built it as an interactive checklist with timestamped check-ins so there’s a dated record of the journey — it exports back to a version-controlled Markdown file whenever I want a permanent copy. If a merged patch is at the end of it, I’ll write that note too.