---
title: "my mainframe runs arch btw"
subtitle: "porting Arch Linux to IBM s390x"
author: "Josephine Pfeiffer"
date: "FOSDEM 2026"
---

## the premise

\only<1>{I use Arch}\only<2->{I use Arch (btw)}

. . .

on an IBM mainframe

. . .

an architecture lineage from 1964

::: notes
Pause for recognition after "I use Arch btw." Then pause again after "on an IBM mainframe." IBM System/360, now z/Architecture. The machine your bank still runs its core software on.
:::

---

## s390x in 30 seconds

- IBM Z / IBM LinuxONE
- 64-bit big-endian
- boots via **IPL**
- your bank runs on this

::: notes
s390x is IBM Z, also branded LinuxONE. Basically the last big-endian architecture anyone still manufactures. It doesn't boot, it does IPL -- Initial Program Load. Storage isn't a disk, it's DASD. Networking is QETH -- QDIO Ethernet, because apparently regular Ethernet wasn't complicated enough.
:::

---

## "why?"

---

## "why?"


I thought it would be funny :p

---

## the problem

Arch Linux assumes you build entirely on the target architecture

. . .

I do not own a mainframe

. . .

(but I have access to one)

![](diagrams/mainframe.jpeg){ height=45% }

::: notes
I am building on a laptop
:::

---

## the solution: a hybrid build

![](diagrams/hybrid-build.png){ height=80% }

::: notes
Why not build everything natively on the mainframe? The z/VM runs RHEL 9.6, not Arch. To build Arch packages natively on s390x, you'd first need pacman, makepkg, and all the build dependencies -- but those don't exist for s390x yet. That's the chicken-and-egg problem this hybrid approach solves.
:::

---

## step 1: the kernel

cross-compile with `s390x-linux-gnu-gcc` in a Fedora container

```bash
make ARCH=s390 CROSS_COMPILE=s390x-linux-gnu- \
     -j$(nproc) bzImage modules
```

this part actually works fine

. . .

this is the last time I will say that

::: notes
The kernel cross-compiles cleanly with the s390x GCC toolchain. This is the last step that just works. I want you to appreciate this moment.
:::

---

## step 2: busybox

busybox must be statically linked for the initramfs

. . .

reliably static-linking s390x binaries requires native compilation

. . .

you need a real mainframe

```bash
$ ssh mainframe
$ make CONFIG_STATIC=y
$ scp busybox-s390x-static laptop:boot/
```

2.3 MB. 60+ applets

::: notes
busybox is a single binary that provides stripped-down implementations of common Unix utilities (sh, cp, mount, etc). It replaces coreutils, util-linux, and a shell in a single statically-linked binary. The initramfs needs these tools before the real root filesystem is mounted, and busybox is the standard way to provide them in a minimal footprint. The "ash" applet in the later slide is busybox's Almquist shell implementation -- the shell that runs as /bin/sh in early boot.
:::

---

## step 3: mkinitcpio

Arch's initramfs generator assumes it can *run* the binaries it packs

. . .

it cannot run s390x binaries on x86_64.

. . .

```bash
# Can't do this on x86_64:
busybox --list
# So instead: hardcode all 60+ applet names
local busybox_applets=(
    "[" "[[" "ash" "awk" "basename" "cat"
    "chmod" "cp" "dd" "df" "echo" "env"
    # ... 50 more lines of this
)
```

::: notes
mkinitcpio has a reasonable assumption: it can execute the binaries it's packing. Except mine are s390x and I'm on x86. So it can't run `busybox --list` to discover applets. My fix: hardcode all 60-plus applet names in a bash array. Elegant? No. Works? Yes.
:::

---

## step 4: the CONFIG_UNIX incident

systemd boots. systemd immediately dies.

. . .

```bash
CONFIG_UNIX=y
# AF_UNIX sockets. systemd needs these.
```

::: notes
No error. No log. Just... nothing. After a few hours of staring at QEMU output, I find it: CONFIG_UNIX=y. AF_UNIX socket support. Systemd will not function without it. It will also not tell you this. You just have to know.
:::

---

## the boot sequence

![](diagrams/boot-sequence.png){ width=100% }

![](diagrams/boot-sequence-2.png){ width=100% }

total system image: **11 MB** smaller than most Electron apps.

---

## it works :D

```bash
$ uname -a
Linux archlinux 6.18.6-arch1-1-s390x #1 SMP s390x GNU/Linux

$ cat /etc/os-release
NAME="Arch Linux"
PRETTY_NAME="Arch Linux"

$ busybox | head -1
BusyBox v1.37.0 (s390x) multi-call binary.

$ systemctl status
  State: running
  Units: 12 loaded
```

::: notes
uname says s390x. os-release says Arch Linux. Busybox and systemd are running. This is a real, booting, functional Arch Linux system on mainframe architecture.
:::

---

## what I learned

- cross-compilation has hard limits so sometimes you need the real hardware
- distribution tooling has architecture assumptions everywhere
- `CONFIG_UNIX=y`

::: notes
And IBM actually gives away free mainframe access through the LinuxONE Community Cloud, so you can do this too.
:::

---

## Arch Linux on s390x

**working today:**

- kernel, initramfs, busybox, systemd
- full boot to systemd shell
- QEMU emulation + real z/VM hardware

**next:**

- pacman and makepkg for s390x
- pacstrap and base packages

---

## thank you

github.com/pfeifferj/archlinux-s390x

josie.lol

::: notes
And yes, I use Arch btw.
:::