C/arm-baremetal-ebook.pdf.txt

Bare-metal programming for ARM

A hands-on guide

Daniels Umanovskis

Contents

0 Introduction

.

.

.

.

.

.

.
Target audience .
.
.
Formatting and terminology .
.
.
.
.
Source code .
Licensing .
.
.
.
.
.
Credits and acknowledgments .

.
.

.
.

.
.

.
.

.
.

.
.

1 Environment setup
.
Linux
.
.
.
.
.
.
.
.
QEMU .
GCC cross-compiler toolchain .
.
Build system essentials .

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.

.

.

2 The first boot

.

.

.

.

.
.

.
.
.

.
.
.

The first hang .

.
.
.
.
Writing some code .
.
.
.
Assembling it .
.
.
.
And... Blastoff! .
.
.
.
What we did wrong .
.
Memory mappings .
.
Creating the vector table .
.
Creating the linker script .
What’s a linker anyway?
.
.
Hanging again - but better .

.

3 Adding a bootloader
.
Introduction .
.
.
Preparing U-Boot .
.
.
.
Creating a SD card .
Creating the uImage .
.
Booting everything .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

5
6
6
7
7
7

8
8
8
9
10

11
12
12
13
14
14
14
16
17
17
18

19
19
19
21
23
24

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

2

Bare-metal programming for ARM

ac090d8bfc

.

.

.

.

.
.

.
.

.
.

New startup code .

4 Preparing a C environment
.
.
Setting up the stack .
.
Handling sections and data .
.
Handing over to C .
.
.
.
.
.
.
.

.
.
Into the C .
.
Building and running .
.
Bonus: exploring the ELF file .

.
.
.
.

.
.
.
.

.
.
.

.
.

.

.

.

5 Build & debug system
.
Building and running .
Debugging in QEMU with GDB .

.

.

.

.

.
.

.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.

.
.

6 UART driver development
.
Doing the homework .

Writing the driver .

.

.

.

.

.
.
.

.
.
.

.
.
.

.
.
.

.
.
.

.
.
.

.
.
.

.
.
.
Basic UART operation .
Key PL011 registers .
.
.
PL011 - Versatile Express integration .
.
.
.
.
.

.
.
.
.
.
.
.
.
.
What’s in the box?
.
.
Exposing the SFRs
Register access width .
.
Initializing and configuring the UART .
.
Read and write functions .
.
.
Putting it to use .
.
.
Doing a test run .
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.

.
.
.

.
.
.

.
.
.

.

.

.

Summary .

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

7 Interrupts

.

.

.

.

.

.

.

Interrupt handling in ARMv7-A .
.
Generic Interrupt Controller of the Cortex-A9 .
.
.
First GIC implementation .
.
.
Handling an interrupt
.
.
.
.
Surviving the IRQ handler
.
Adapting the UART driver .
.
.
Handling different interrupt sources .
.
.
.
Summary .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.

.

.

.

.

.

.

.

.

.

.

.

Daniels Umanovskis

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

26
26
26
29
33
33
35
37

39
42
43

46
46
47
47
48
49
49
49
52
53
57
59
60
61

62
62
64
67
74
77
79
80
83

3

Bare-metal programming for ARM

ac090d8bfc

.

.
.

.
.

.
.

8 Simple scheduling
.
Private Timer Driver
System Time .
.
.
Overflows and spaceships .
.
.
Scheduler types .
.
Cooperative scheduler .
.
.
.
Summary .

.
.
.

.

.

.

.

.

.

.

.

.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

84
84
87
88
89
90
95

Daniels Umanovskis

4

0 Introduction

Modern programming takes many forms. There’s web development, desktop application development,
mobile development and more. Embedded programming is one of the areas of programming, and can
be radically different from the others. Embedded programming means programming for a computer
that’s mainly intended to be embedded within a larger system, and the embedded computer is usually
responsible for a particular task, instead of being a general-purpose computing device. The system
where it’s embedded might be a simple pocket calculator, or an industrial robot, or a spaceship. Some
embedded devices are microcontrollers with very little memory and low frequencies, others are more
powerful.

Embedded computers may be running a fully-fleged operating system, or a minimalistic system that
just provides some scheduling of real-time functions. In cases when there’s no operating system at all,
the computer is said to be bare metal, and consequently bare metal programming is programming
directly for a (micro-)computer that lacks an operating system. Bare metal programming can be both
highly challenging and very different from other types of programming. Code interfaces directly with
the underlying hardware, and common abstractions aren’t available. There are no files, processes or
command lines. You cannot even get the simplest C code to work without some preparatory steps. And,
in one of the biggest challenges, failures tend to be absolute and mysterious. It’s not uncommon to see
embedded developers break out tools such as voltmeters and oscilloscopes to debug their software.

Modern embedded hardware comes in very many types, but the field is dominated by CPUs
implementing an ARM architecture. Smartphones and other mobile devices often run Qualcomm
Snapdragon or Apple A-series CPUs, which are all based on the ARM architecture. Among
microcontrollers, ARM Cortex-M and Cortex-R series CPU cores are very popular. The ARM architectures
play a very significant role in modern computing.

The subject of this ebook is bare-metal programming in C for an ARM system. Specifically, the ARMv7-A
architecture is used, which is the last purely 32-bit ARM architecture, unlike the newer ARMv8/AArch64.
The -A suffix in ARMv7-A indicates the A profile, which is intended for more resource-intensive
applications. The corresponding microcontroller architecture is ARMv7-M.

Note that this is not a tutorial on how to write an OS. Some of the topics covered in this ebook are
relevant for OS development, but there are many OS-specific aspects that are not covered here.

5

Bare-metal programming for ARM

ac090d8bfc

Target audience

This ebook is aimed at people who have an interest in low-level programming, and in seeing how
to build a system from the ground up. Topics covered include system startup, driver development
and low-level memory management. For the most part, the chapters cover things from a practical
perspective, by building something, although background theory is provided.

The reader should be familiar with C programming. This is not a C tutorial, and even though there are
occasional notes on the language, the ebook is probably difficult to follow without having programmed
in C. Some minimal exposure to an assembly language and understanding of computer architecture
are very useful, though the assembly code presented is explained line by line as it’s introduced.

It’s also helpful to be familiar with Linux on a basic level. You should be able to navigate directories,
run shell scripts and do basic troubleshooting if something doesn’t work - fortunately, even for an
inexperienced Linux user, a simple online search is often enough to solve a problem. The ebook
assumes all development is done on Linux, although it should be possible to do it on OS X and even on
Windows with some creativity.

Experienced embedded developers are unlikely to find much value in this text.

Formatting and terminology

The ebook tries to for the most part follow usual conventions for a programming-related text.
Commands, bits of code or references to variables are formatted like code, with bigger code
snippets presented separately like this:

1 void do_amazing_things(void) {
2
3
4 }

int answer = 42;
/* A lot happens here! */

If you are reading the PDF version, note that longer lines of code have to get wrapped to fit within the
page, but the indentation and line numbers inside each code block should help keep things clear.

Due to some unfortunate historical legacy, there are two different definitions for data sizes in common
use. There’s the binary definition, where a kilobyte is 1024 bytes, and the metric definition, where a
kilobyte is 1000 bytes. Throughout this ebook, all references to data quantities are in the binary sense.
The meaning of “billion” in the book is 10^9.

Daniels Umanovskis

6

Bare-metal programming for ARM

ac090d8bfc

Source code

For each chapter, the corresponding source code is available. If you’re reading this on GitHub, you
can explore the repository and check its readme file for more information. If you’re reading the PDF
version or other standalone copy, you can head to the GitHub repository to access the source code,
and perhaps updates to this ebook. The repository URL is https://github.com/umanovskis/
baremetal-arm/.

Licensing

The ebook is licensed under Creative Commons Attribution-Share Alike (CC-BY-SA) license - see the Git
repository for more details on licensing.

Credits and acknowledgments

The PDF version of this ebook is typeset using the Eisvogel LaTeX template by Pascal Wagler.

Daniels Umanovskis

7

1 Environment setup

In this chapter, I’ll cover the basic environment setup to get started with ARM bare-metal programming
using an emulator. Some familiarity with using Linux is assumed. You don’t need to be a Linux expert,
but you should be able to use the command line and do some basic troubleshooting if the system
doesn’t behave as expected.

Linux

The first prerequisite is getting a Linux system up and running. Hopefully you are already familiar with
Linux and have some kind of Linux running. Otherwise you should install Linux, or set up a virtual
machine running Linux.

If you are running Windows and want to run a virtual Linux, VirtualBox is recommended. As for Linux
distributions, any modern distribution should be fine, although in some cases you might need to install
software manually. I use Linux Mint Debian Edition, and double-check most of the work in a virtual
machine running Ubuntu, which is the most popular Linux distribution for beginners.

QEMU

To emulate an ARM machine, we will be using QEMU, a powerful emulation and virtualization tool that
works with a variety of architectures. While the code we write should eventually be able to boot on a
real ARM device, it is much easier to start with an emulator. Why?

• No additional hardware is needed.

• You don’t have to worry about software flashing / download process.

• You have much better tools to inspect the state of the emulated hardware. When working with
real hardware, you would need a few drivers to get meaningful information from the software, or
use other more difficult methods.

8

Bare-metal programming for ARM

ac090d8bfc

Since QEMU supports a wide range of systems, we’ll need to install the ARM version. On Debian/Ubuntu
based systems, the qemu-system-arm package will provide what you need, so let’s just go ahead and
install it:

1 sudo apt-get install qemu-system-arm

GCC cross-compiler toolchain

The next step is the installation of a cross-compiler toolchain. You cannot use the regular gcc compiler
to build code that will run on an ARM system, instead you’ll need a cross-compiler. What is a cross-
compiler exactly? It’s simply a compiler that runs on one platform but creates executables for another
platform. In our case, we’re running Linux on the x86-64 platform, and we want executables for ARM,
so a cross compiler is the solution to that.

The GNU build tools, and by extension GCC, use the concept of a target triplet to describe a platform.
The triplet lists the platform’s architecture, vendor and operating system or binary interface type. The
vendor part of target triplets is generally irrelevant. You can look up your own machine’s target triplet
by running gcc -dumpmachine. I get x86_64-linux-gnu, yours will likely be the same or similar.

To compile for ARM, we need to select the correct cross-compiler toolchain, that is, the toolchain with a
target triplet matching our actual target. The fairly widespread gcc-arm-linux-gnueabi toolchain
will not work for our needs, and you can probably guess why – the name indicates that the toolchain is
intended to compile code for ARM devices running Linux. We’re going to do bare-metal programming,
so no Linux running on the target system.

The toolchain we need is gcc-arm-none-eabi. We will need a version with GCC 6 or newer, for when
we later use U-Boot. On Ubuntu, you should be able to simply install the toolchain:

1 sudo apt-get install gcc-arm-none-eabi

You can run arm-none-eabi-gcc --version to check the version number.
If you’re using a
distribution that offers an old version of the package, you can download the toolchain from ARM
directly. In that case, it’s recommended that you add the toolchain’s folder to your environment’s
PATH after extracting it somewhere.

Daniels Umanovskis

9

Bare-metal programming for ARM

ac090d8bfc

Build system essentials

Finally, we need the essential components of a build system. In the coming examples, we’ll be using
the standard Make build tool, as well as CMake. Debian-based systems provide a handy package called
build-essential, which installs Make and other relevant programs. CMake is available in a package
called cmake, so the installation is simple:

1 sudo apt-get install build-essential cmake

On some Linux variants, you might also need to install bison and flex if they are not already present.
Those tools are also required to build U-Boot.

1 sudo apt-get install bison flex

sort -R ~/facts-and-trivia | head -n1

The flex program is an implementation of lex, a standard lexical analyzer first developed in the
mid-1970s by Mike Lesk and Eric Schmidt, who served as the chairman of Google for some years.

With this, your system should now have everything that is necessary to compile programs for ARM and
run them in an emulated machine. In the next chapter, we’ll continue our introduction by booting the
emulated machine and giving some of the just-installed tools a spin.

Daniels Umanovskis

10

2 The first boot

We’ll continue our introduction to bare-metal ARM by starting an emulated ARM machine in QEMU,
and using the cross-compiler toolchain to load the simplest possible code into it.

Let us run QEMU for the very first time, with the following command:

1 qemu-system-arm -M vexpress-a9 -m 32M -no-reboot -nographic -monitor

telnet:127.0.0.1:1234,server,nowait

The QEMU machine will spin up, briefly seem to do nothing, and then crash with an error message.
The crash is to be expected - we did not provide any executable to run so of course our emulated
system cannot accomplish anything. For documenation of the QEMU command line, you can check
man qemu-doc and online, but let’s go through the command we used and break it down into parts.

• -M vexpress-a9. The -M switch selects the specific machine to be emulated. The ARM Versatile
Express is an ARM platform intended for prototyping and testing. Hardware boards with the
Versatile Express platform exist, and the platform is also a common choice for testing with
emulation. The vexpress-a9 variant has the Cortex A9 CPU, which powers a wide variety of
embedded devices that perform computationally-intensive tasks.

• -m 32M. This simply sets the RAM of the emulated machine to 32 megabytes.

• -no-reboot. Don’t reboot the system if it crashes.

• -nographic. Run QEMU as a command-line application, outputting everything to a terminal

console. The serial port is also redirected to the terminal.

• -monitor telnet:127.0.0.1:1234,server,nowait. One of the advantages of QEMU is
that it comes with a powerful QEMU monitor, an interface to examine the emulated machine and
control it. Here we say that the monitor should run on localhost, port 1234, with the server,
nowait options meaning that QEMU will provide a telnet server but will continue running even
if nobody connects to it.

That’s it for the first step - now you have a command to start an ARM system, although it will not
do anything except crashing with a message like qemu-system-arm: Trying to execute code

outside RAM or ROM at 0x04000000.

11

Bare-metal programming for ARM

ac090d8bfc

The first hang

Writing some code

Now we want to write some code, turn it into an executable that can run on an ARM system, and run it
in QEMU. This will also serve as our first cross-compilation attempt, so let’s go with the simplest code
possible - we will write a value to one of the CPU registers, and then enter an infinite loop, letting our
(emulated) hardware hang indefinitely. Since we’re doing bare-metal programming and have no form
of runtime or operating system, we have to write the code in assembly language. Create a file called
startup.s and write the following code:

1 ldr r2,str1
2 b .
3 str1: .word 0xDEADBEEF

Line by line:

1. We load the value at label str1 (which we will define shortly) into the register R2, which is one

of the general-purpose registers in the ARM architecture.

2. We enter an infinite loop. The period . is short-hand for current address, so b . means “branch

to the current instruction address”, which is an infinite loop.

3. We allocate a word (4 bytes) with the value 0xDEADBEEF, and give it the label str1. The value
0xDEADBEEF is a distinctive value that we should easily notice. Writing such values is a common
trick in low-level debugging, and 0xDEADBEEF is often used to indicate free memory or a general
software crash. Why will this work if we’re in an infinite loop? Because this is not executable
code, it’s just an instruction to the assembler to allocate the 4-byte word here.

sort -R ~/facts-and-trivia | head -n1

Memorable hexadecimal values like 0xDEADBEEF have a long tradition, with different vendors and
systems having their own constants. Wikipedia has a separate article on these magic values.

Daniels Umanovskis

12

Bare-metal programming for ARM

ac090d8bfc

Assembling it

Next we need to compile the code. Since we only wrote assembly code, compilation is not actually
relevant, so we will just assemble and link the code. The first time, let’s do this manually to see how to
use the cross-compiler toolchain correctly. What we’re doing here is very much not optimal even for an
example as simple as this, and we’ll improve the state of things soon.

First, we need to assemble startup.s, which we can do like this, telling the GNU Assembler (as) to
place the output in startup.o.

1 arm-none-eabi-as -o startup.o startup.s

We do not yet have any C files, so we can go ahead and link the object file, obtaining an executable.

1 arm-none-eabi-ld -o first-hang.elf startup.o

This will create the executable file first-hang.elf. You will also see a warning about missing _start.
The linker expects your code to include a _start symbol, which is normally where the execution
would start from. We can ignore this now because we only need the ELF file as an intermediate step
anyway. The first-hang.elf which you obtained is a sizable executable, reaching 33 kilobytes on my
system. ELF executables are the standard for Linux and other Unix-like systems, but they need to be
loaded and executed by an operating system, which we do not have. An ELF executable is therefore
not something we can use on bare metal, so the next step is to convert the ELF to a raw binary dump of
the code, like this:

1 arm-none-eabi-objcopy -O binary first-hang.elf first-hang.bin

The resulting first-hang.bin is just a 12-byte file, that’s all the space necessary for the code we wrote
in startup.s. If we look at the hexdump of this file, we’ll see 00000000 00 20 9f e5 fe ff ff
ea ef be ad de. You can recognize our 0xDEADBEEF constant at the end. The first eight bytes
are our assembly instructions in raw form. The code starts with 0020 9fe5. The ldr instruction has
the opcode e5, then 9f is a reference to the program counter (PC) register, and the 20 refers to R2,
meaning this is in fact ldr r2, [pc] encoded.

Looking at the hexdump of a binary and trying to match the bytes to assembly instructions is uncommon
even for low-level programming. It is somewhat useful here as an illustration, to see how we can go
from writing startup.s to code executable by an ARM CPU, but this is more detail than you would
typically need.

Daniels Umanovskis

13

Bare-metal programming for ARM

ac090d8bfc

And... Blastoff!

We can finally run our code on the ARM machine! Let’s do so.

1 qemu-system-arm -M vexpress-a9 -m 32M -no-reboot -nographic -monitor
telnet:127.0.0.1:1234,server,nowait -kernel first-hang.bin

This runs QEMU like previously, but we also pass -kernel first-hang.bin, indicating that we want
to load our binary file into the emulated machine. This time you should see QEMU hang indefinitely.
The QEMU monitor allows us to read the emulated machine’s registers, among other things, so we
can check whether our code has actually executed. Open a telnet connection in another terminal with
telnet localhost 1234, which should drop you into the QEMU monitor’s command line, looking
something like this:

1 QEMU 2.8.1 monitor - type 'help' for more information
2 (qemu)

At the (qemu) prompt, type info registers. That’s the monitor command to view registers. Near
the beginning of the output, you should spot our 0xDEADBEEF constant that has been loaded into
R2:

1 R00=00000000 R01=000008e0 R02=deadbeef R03=00000000

This means that yes indeed, QEMU has successfully executed the code we wrote. Not at all fancy, but it
worked. We have our first register write and hang.

What we did wrong

Our code worked, but even in this small example we didn’t really do things the right way.

Memory mappings

One issue is that we didn’t explicitly specify any start symbol that would show where our program
should begin executing. It works because when the CPU starts up, it begins executing from address
0x0, and we have placed a valid instruction at that address. But it could easily go wrong. Consider this
variation of startup.s, where we move the third line to the beginning.

Daniels Umanovskis

14

Bare-metal programming for ARM

ac090d8bfc

1 str1: .word 0xDEADBEEF
2 ldr r2,str1
3 b .

That is still valid assembly and feels like it should work, but it wouldn’t - the constant 0xDEADBEEF
would end up at address 0x0, and that’s not a valid instruction to begin the program with. Moreover,
even starting at address 0x0 isn’t really correct. On a real system, the interrupt vector table should be
located at address 0x0, and the general boot process should first have the bootloader starting, and
after that switch execution to your code, which is loaded somewhere else in memory.

QEMU is primarily used for running Linux or other Unix-like kernels, which is reflected in how it’s
normally started. When we start QEMU with -kernel first-hang.bin, QEMU acts as if booting such
a kernel. It copies our code to the memory location 0x10000, that is, a 64 kilobyte offset from the
beginning of RAM. Then it starts executing from the address 0x0, where QEMU already has some startup
code meant to prepare the machine and jump to the kernel.

Sounds like we should be able to find our first-hang.bin at 0x10000 in the QEMU memory then.
Let’s try do to that in the QEMU monitor, using the xp command which displays physical memory.
In the QEMU monitor prompt, type xp /4w 0x100000 to display the four words starting with that
memory address.

1 0000000000100000: 0x00000000 0x00000000 0x00000000 0x00000000

Everything is zero! If you check the address 0x0, you will find the same. How come?

The answer is memory mapping - the address space of the device encompasses more than just the
RAM. It’s time to consult the most important document when developing for a particular device, its
Technical Reference Manual, or TRM for short. The TRM for any embedded device is likely to have a
section called “memory map” or something along those lines. The TRM for our device is available from
ARM, and it indeed contains a memory map. (Note: when working with any device, downloading a
PDF version of the TRM is a very good idea.) In this memory map, we can see that the device’s RAM
(denoted as “local DDR2”) begins at 0x60000000.

That means we have to add 0x60000000 to RAM addresses to obtain the physical address, so our
0x10000 where we expect the binary code to be loaded is at physical address 0x60010000. Let’s
check if we can find the code at that address: xp /4w 0x60010000 shows us:

1 0000000060010000: 0xe59f2000 0xeafffffe 0xdeadbeef 0x00000000

Daniels Umanovskis

15

Bare-metal programming for ARM

ac090d8bfc

There it is indeed, our first-hang.bin loaded into memory!

Creating the vector table

Having our code start at address 0x0 isn’t acceptable as explained before, as that is the address where
the interrupt vector table is expected. We should also not rely on things just working out, with help
from QEMU or without, and should explicitly specify the entry point for our program. Finally, we should
separate code and data, placing them in separate sections. Let’s start by improving our startup.s a
bit:

b Reset_Handler
b . /* 0x4 Undefined Instruction */
b . /* 0x8 Software Interrupt */
b . /* 0xC Prefetch Abort */
b . /* 0x10 Data Abort */
b . /* 0x14 Reserved */
b . /* 0x18 IRQ */
b . /* 0x1C FIQ */

1 .section .vector_table, "x"
2 .global _Reset
3 _Reset:
4
5
6
7
8
9
10
11
12
13 .section .text
14 Reset_Handler:
15
16
17

ldr r2, str1
b .
str1: .word 0xDEADBEEF

Here are the things we’re doing differently this time:

1. We’re creating the vector table at address 0x0, putting it in a separate section called .
vector_table, and declaring the global symbol _Reset to point to its beginning. We leave
most items in the vector table undefined, except for the reset vector, where we place the
instruction b Reset_Handler.

2. We moved our executable code to the .text section, which is the standard section for code. The
Reset_Handler label points to the code so that the reset interrupt vector will jump to it.

Daniels Umanovskis

16

Bare-metal programming for ARM

ac090d8bfc

Creating the linker script

Linker scripts are a key component in building embedded software. They provide the linker with
information on how to place the various sections in memory, among other things. Let’s create a linker
script for our simple program, call it linkscript.ld.

1 ENTRY(_Reset)
2
3 SECTIONS
4 {
5
6
7
8 }

. = 0x0;
.text : { startup.o (.vector_table) *(.text) }
. = ALIGN(8);

This script tells the linker that the program’s entry point is at the global symbol _Entry, which we
export from startup.s. Then the script goes on to list the section layout. Starting at address 0x0, we
create the .text section for code, consisting first of the .vector_table section from startup.o,
and then any and all other .text sections. We align the code to an 8-byte boundary as well.

What’s a linker anyway?

Indeed, what’s a linker and why are we using one? As developers, we often say “compilation” when
referring to the process by which source code turns into an executable. More accurately though,
compilation is just one step, normally followed by linking, and it’s this compile-and-link process that
often gets simply called compilation. The confusion isn’t helped by the fact that compiling and linking
are usually invoked with the same command in most tools. Whether you’re using an IDE or using GCC
from the command line, compilation and linking will usually occur together.

The compiler takes source code and produces object files as the output, these files usually have the .o
extension. A linker takes one or more object files, possibly adds external libraries, and links it all into an
executable. In any non-trivial program, each object file is likely to refer to functions that are contained
in other object files, and resolving those dependencies is part of the linker’s job. So linkers themselves
are nothing specific to embedded programming, but due to the low abstraction level available when
programming for bare metal, it’s common to require more control over the linker’s actions.

Linker scripts like the one above, in the broadest terms, tell the linker how to do its job. For now we’re
just giving it simple instructions, but later we’ll write a more sophisticated linker script.

Daniels Umanovskis

17

Bare-metal programming for ARM

ac090d8bfc

Hanging again - but better

We can now build the updated software. Let’s do that in a similar manner to before:

1 arm-none-eabi-as -o startup.o startup.s
2 arm-none-eabi-ld -T linkscript.ld -o better-hang.elf startup.o
3 arm-none-eabi-objcopy -O binary better-hang.elf better-hang.bin

Note the addition of -T linkscript.ld to the linker command, specifying to use our newly created
linker script. We still cannot use the ELF file directly, but we could use objdump to verify that our
linkscript changed things. Call arm-none-eabi-objdump -h better-hang.elf to see the list of
sections. You’ll notice the .text section. And if you use objdump to view startup.o, you’ll also see
.vector_table. You can even observe that the sizes of .vector_table and .text in startup.o
add up to the size of .text in the ELF file, further indicating that things are probably as we wanted.

We can now once again run the software in QEMU with qemu-system-arm -M vexpress-a9 -

m 32M -no-reboot -nographic -monitor telnet:127.0.0.1:1234,server,nowait -
kernel better-hang.bin and observe the same results as before, and happily knowing things are
now done in a more proper way.

In the next chapter, we will continue by introducing a bootloader into our experiments.

Daniels Umanovskis

18

3 Adding a bootloader

Introduction

The bootloader is a critical piece of software that is necessary to get the hardware into a usable state,
and load other more useful programs, such as an operating system. On PCs and other fully-featured
devices, common bootloaders include GNU GRUB (which is most likely used to boot your Linux system),
bootmgr (for modern versions of MS Windows) and others. Developing bootloaders is a separate and
complicated subject. Bootloaders are generally full of esoteric, highly architecture-specific code, and
in my opinion learning about bootloaders if fun, but the knowledge is also somewhat less transferable
to other areas of development.

Writing your own bootloader is certainly something that could be attempted, but in this series of posts
we will continue by doing something that is usually done in embedded development, namely using
Das U-Boot.

Das U-Boot, usually referred to just as U-Boot, is a very popular bootloader for embedded devices. It
supports a number of architectures, including ARM, and has pre-made configurations for a very large
number of devices, including the Versatile Express series that we’re using. Our goal in this article will
be to build U-Boot, and combine it with our previously built software. This will not, strictly speaking,
change anything significant while we are running on an emulated target in QEMU, but we would need
a bootloader in order to run on real hardware.

We will, in this article, change our boot sequence so that U-Boot starts, and then finds our program on
a simulated SD card, and subsequently boots it. I will only provide basic explanations for some of the
steps, because we’ll mostly be dealing with QEMU and Linux specifics here, not really related to ARM
programming.

Preparing U-Boot

First, you should download U-Boot. You could clone the project’s source tree, but the easiest way is to
download a release from the official FTP server. For writing this, I used u-boot-2018.09. This is also

19

Bare-metal programming for ARM

ac090d8bfc

the reason why your cross-compiler toolchain needs gcc of at least version 6 - earlier versions cannot
compile U-Boot.

After downloading U-Boot and extracting the sources (or cloning them), you need to run two commands
in the U-Boot folder.

1 make vexpress_ca9x4_config ARCH=arm CROSS_COMPILE=arm-none-eabi-

This command will prepare some U-Boot configuration, indicating that we want it for the ARM
architecture and, more specifically, we want to use the vexpress_ca9x4 configuration, which
corresponds to the CoreTile Express A9x4 implementation of the Versatile Express platform that
we’re using.‘ The configuration command should only take a few seconds to run, after which we can
build U-Boot:

1 make all ARCH=arm CROSS_COMPILE=arm-none-eabi-

If everything goes well, you should, after a short build process, see the file u-boot and u-boot.bin
created. You can quickly test by running QEMU as usual, except you start U-Boot, providing -kernel
u-boot on the command line (note that you’re booting u-boot and not u-boot.bin). You should
see U-Boot output some information, and you can drop into the U-Boot command mode if you hit a
key when prompted.

Having confirmed that you can run U-Boot, make a couple of small modifications to it. In configs/
vexpress_ca9x4_defconfig, change the CONFIG_BOOTCOMMAND line to the following:

1 CONFIG_BOOTCOMMAND="run mmc_elf_bootcmd"

The purpose of that will become clear a bit later on. Then open include/config_distro_bootcmd.h
and go to the end of the file. Find the last line that says done\0 and edit from there so that the file
looks like this:

1
2
3
4
5
6
7

"done\0"

\
"bootcmd_bare_arm="
"mmc dev 0;"
"ext2load mmc 0 0x60000000 bare-arm.uimg;"
"bootm 0x60000000;"
"\0"

\

\
\
\
\

Daniels Umanovskis

20

Bare-metal programming for ARM

ac090d8bfc

Note that in the above snippet, the first line with done\0 was already in the file, but we add a backslash
\ to the end, and then we add the subsequent lines. See the edited file in the repository. Regenerate
the U-Boot config and rebuild it:

1 make vexpress_ca9x4_config ARCH=arm CROSS_COMPILE=arm-none-eabi-
2 make all ARCH=arm CROSS_COMPILE=arm-none-eabi-

Now would be a good time to start U-Boot in QEMU and verify that everything works. Start QEMU by
passing the built U-Boot binary to it in the -kernel parameter, like this (where u-boot-2018.09 is a
subfolder name that you might need to change):

1 qemu-system-arm -M vexpress-a9 -m 32M -no-reboot -nographic -monitor
telnet:127.0.0.1:1234,server,nowait -kernel u-boot-2018.09/u-boot

QEMU should show U-Boot starting up, and if you hit a key when U-Boot prompts Hit any key to
stop autoboot, you’ll be dropped into the U-Boot command line. With that, we can be satisfied
that U-Boot was built correctly and works, so next we can tell it to boot something specific, like our
program.

Creating a SD card

On a real hardware board, you would probably have U-Boot and your program stored in the program
flash. This doesn’t comfortably work with QEMU and the Versatile Express series, so we’ll take another
approach that is very similar to what you could on hardware. We will create a SD card image, place
our program there, and tell U-Boot to boot it. What follows is again not particularly related to ARM
programming, but rather a convenient way of preparing an image.

First we’ll need an additional package that can be installed with sudo apt-get install qemu-
utils.

Next we need the SD card image itself, which we can create with qemu-img. Then we will create an
ext2 partition on the SD card, and finally copy the uImage containing our code to the card (we’ll create
the uImage in the next section). It is not easily possible to manipulate partitions directly inside an
image file, so we will need to mount it using qemu-nbd, a tool that makes it possible to mount QEMU
images as network block devices. The following script, which I called create-sd.sh, can be used to
automate the process:

Daniels Umanovskis

21

Bare-metal programming for ARM

ac090d8bfc

1 #!/bin/bash
2
3 SDNAME="$1"
4 UIMGNAME="$2"
5
6 if [ "$#" -ne 2 ]; then
7
8
9 fi
10
11 command -v qemu-img >/dev/null || { echo "qemu-img not installed"; exit

echo "Usage: "$0" sdimage uimage"
exit 1

1; }

12 command -v qemu-nbd >/dev/null || { echo "qemu-nbd not installed"; exit

1; }

13
14 qemu-img create "$SDNAME" 64M
15 sudo qemu-nbd -c /dev/nbd0 "$SDNAME"
16 (echo o;
17 echo n; echo p
18 echo 1
19 echo ; echo
20 echo w; echo p) | sudo fdisk /dev/nbd0
21 sudo mkfs.ext2 /dev/nbd0p1
22
23 mkdir tmp || true
24 sudo mount -o user /dev/nbd0p1 tmp/
25 sudo cp "$UIMGNAME" tmp/
26 sudo umount /dev/nbd0p1
27 rmdir tmp || true
28 sudo qemu-nbd -d /dev/nbd0

The script creates a 64 megabyte SD card image, mounts it as a network block device, creates a single
ext2 partition spanning the entire drive, and copies the supplied uImage to it. From the command line,
the script could then be used like

1 ./create-sd.sh sdcard.img bare-arm.uimg

to create an image called sdcard.img and copy the bare-arm.uimg uImage onto the emulated SD
card (we’ll create the image below, running the command at this point will fail).

NOTE

Depending on your system, you might get an error about /dev/nbd0 being unavailable when you run
the SD card creation script. The most likely cause of such an error is that you don’t have the nbd kernel

Daniels Umanovskis

22

Bare-metal programming for ARM

ac090d8bfc

module loaded. Loading it with sudo modprobe nbd should create /dev/nbd0. To permanently
add the module to the load list, you can do echo "nbd"| sudo tee -a /etc/modules

Creating the uImage

Now that we have created a SD card and can copy an uImage to it, we have to create the uImage itself.

First of all, what is an uImage? The U-Boot bootloader can load applications from different types of
images. These images can consist of multiple parts, and be fairly complex, like how Linux gets booted.
We are not trying to boot a Linux kernel or anything else complicated, so we’ll be using an older image
format for U-Boot, which is then the uImage format. The uImage format consists of simply the raw data
and a header that describes the image. Such images can be created with the mkimage utility, which is
part of U-Boot itself. When we built U-Boot, mkimage should have been built as well.

Let’s call mkimage and ask it to create an U-Boot uImage out of the application we had previously, the
“better hang” one. From now on, we’ll also be able to use ELF files instead of the raw binary dumps
because U-Boot knows how to load ELF files. mkimage should be located in the tools subfolder of
the U-Boot folder. Assuming our better-hang.elf is still present, we can do the following:

1 u-boot-2018.09/tools/mkimage -A arm -C none -T kernel -a 0x60000000 -e

0x60000000 -d better-hang.elf bare-arm.uimg

With that, we say that we want an uncompressed (-C none) image for ARM (-A arm), the image will
contain an OS kernel (-T kernel). With -d better-hang.bin we tell mkimage to put that .bin
file into the image. We told U-Boot that our image will be a kernel, which is not really true because we
don’t have an operating system. But the kernel image type indicates to U-Boot that the application is
not going to return control to U-Boot, and that it will manage interrupts and other low-level things by
itself. This is what we want since we’re looking at how to do low-level programming in bare metal.

We also indicate that the image should be loaded at 0x60000000 (with -a) and that the entry point for
the code will be at the same address (with -e). This choice of address is because we want to load the
image into the RAM of our device, and in the previous chapter we found that RAM starts at 0x60000000
on the board. Is it safe to place our code into the beginning of RAM? Will it not overwrite U-Boot itself
and prevent a proper boot? Fortunately, we don’t have that complication. U-Boot is initially executed
from ROM, and then, on ARM system, it copies itself to the end of the RAM before continuing from
there.

Daniels Umanovskis

23

Bare-metal programming for ARM

ac090d8bfc

When the uImage is created, we need to copy it to the SD card. As noted previously, it can be done just
by executing ./create-sd.sh sdcard.img bare-arm.uimg thanks to the script created before.
If everything went well, we now have a SD card image that can be supplied to QEMU.

Booting everything

We’re ready to boot! Let’s start QEMU as usual, except that this time we’ll also add an extra parameter
telling QEMU that we want to use a SD card.

1 qemu-system-arm -M vexpress-a9 -m 32M -no-reboot -nographic -monitor

telnet:127.0.0.1:1234,server,nowait -kernel u-boot-2018.09/u-boot -
sd sdcard.img

Hit a key when U-Boot prompts you to, in order to use the U-Boot command line interface. We can
now use a few commands to examine the state of things and confirm that everything is as we wanted.
First type mmc list and you should get a response like MMC: 0. This confirms the presence of an
emulated SD card. Then type ext2ls mmc 0. That is the equivalent of running ls on the SD card’s
filesystem, and you should see a response that includes the bare-arm.uimg file - our uImage.

Let’s load the uImage into memory. We can tell U-Boot to do that with ext2load mmc 0 0x60000000
bare-arm.uimg, which as can probably guess means to load bare-arm.uimg from the ext2 system
on the first MMC device into address 0x60000000. U-Boot should report success, and then we can
use iminfo 0x60000000 to verify whether the image is located at that address now. If everything
went well, U-Boot should report that a legacy ARM image has been found at the address, along with a
bit more information about the image. Now we can go and boot the image from memory: bootm 0
x60000000.

U-Boot will print Starting kernel... and seemingly hang. You can now check the QEMU monitor
(recall that you can connect to it with telnet localhost 1234), and issue the info registers
command to see that R2 is once again equal to 0xDEADBEEF. Success! Our program has now been
loaded from a SD card image and started through U-Boot!

Better yet, the modifications we made earlier to U-Boot allow it to perform this boot sequence
automatically. With those modifications, we added a new environment variable to U-Boot,
bootcmd_bare_arm, which contains the boot commands. If you type printenv bootcmd_bare_arm
in the U-Boot command-line, you’ll see the boot sequence.

If you start QEMU again and don’t press any keys to pause U-Boot, you should see If you type printenv
bootcmd_bare_arm in the U-Boot command-line, you’ll see the boot sequence. If you start QEMU

again and don’t press any keys to pause U-Boot, you should see the boot continue automatically.

Daniels Umanovskis

24

Bare-metal programming for ARM

ac090d8bfc

NOTE

Modifying the U-Boot source code in order to save a command sequence may seem strange, and indeed
we’re doing that because of QEMU emulation. Normally, running U-Boot from a writable device such
as a SD card would let us use U-Boot’s setenv and saveenv commands to permanently save changes
without recompiling the whole bootloader.

Having now completed a boot sequence fairly similar to real hardware, we can continue with our own
programming. In the next chapter, we’ll continue by getting some C code running.

Daniels Umanovskis

25

4 Preparing a C environment

In this part, we will do some significant work to get a proper application up and running. What we have
done so far is hardly an application, we only execute one instruction before hanging, and everything is
done in the reset handler. Also, we haven’t written or run any C code so far. Programming in assembly is
harder than in C, so generally bare-metal programs will do a small amount of initialization in assembly
and then hand control over to C code as soon as possible.

We are now going to write startup code that prepares a C environment and runs the main function in
C. This will require setting up the stack and also handling data relocation, that is, copying some data
from ROM to RAM. The first C code that we run will print some strings to the terminal by accessing the
UART peripheral device. This isn’t really using a proper driver, but performing some UART prints is a
very good way of seeing that things are working properly.

New startup code

Setting up the stack

Our startup code is getting a major rework. It will do several new and exciting things, the most basic
of which is to prepare the stack. C code will not execute properly without a stack, which is necessary
among other things to have working function calls.

Conceptually, preparing the stack is quite simple. We just pick two addresses in our memory space that
will be the start and end of the stack, and then set the initial stack pointer to the start of the stack. By
convention, the stack grows towards lower addresses, that is, your stack could be between addresses
like 0x60020000 and 0x60021000, which would give a stack of 0x1000 bytes, and the stack pointer
would initially point to the “end” address 0x600210000. This is called a descending stack. The ARM
Procedure Call Standard also specifies that a descending stack should be used.

The ARMv7A architecture has, on a system level, several stack pointers and the CPU has several
processor modes. Consulting the ARMv7A reference manual, we can see that processor modes are:
user, FIQ, IRQ, supervisor, monitor, abort, undefined and system. To simplify things, we will only care
about three modes now - the FIQ and IRQ modes, which execute fast interrupt and normal interrupt
code respectively - and the supervisor mode, which is the default mode the processor starts in.

26

Bare-metal programming for ARM

ac090d8bfc

Further consulting the manual (section B1.3.2 ARM core registers), we see that the ARM core registers,
R0 to R15, differ somewhat depending on the mode. We are now interested in the stack pointer register,
SP or R13, and the manual indicates that the “current” SP register an application sees depends on the
processor mode, with the supervisor, IRQ and FIQ modes each having their own SP register.

NOTE

Referring to ARM registers, SP and R13 always means the same thing. Similarly, LR is R14 and PC is R15.
The name used depends on the documentation or the tool you’re using, but don’t get confused by the
three special registers R13-R15 having multiple names.

In our startup code, we are going to switch to all the relevant modes in order and set the initial stack
pointer in the SP register to a memory address that we allocate for the purpose. We’ll also fill the stack
with a garbage pattern to indicate it’s unused.

First we add some defines at the beginning of our startup.s, with values for the different modes
taken from the ARMv7A manual.

1 /* Some defines */
2 .equ MODE_FIQ, 0x11
3 .equ MODE_IRQ, 0x12
4 .equ MODE_SVC, 0x13

Now we change our entry code in Reset_Handler so that it starts like this:

/* FIQ stack */
msr cpsr_c, MODE_FIQ
ldr r1, =_fiq_stack_start
ldr sp, =_fiq_stack_end
movw r0, #0xFEFE
movt r0, #0xFEFE

1 Reset_Handler:
2
3
4
5
6
7
8
9 fiq_loop:
10
11
12

cmp r1, sp
strlt r0, [r1], #4
blt fiq_loop

Daniels Umanovskis

27

Bare-metal programming for ARM

ac090d8bfc

Let’s walk through the code in more detail.

1

msr cpsr_c, MODE_FIQ

One of the most important registers in an ARMv7 CPU is the CPSR register, which stands for Current
Program Status Register. This register can be written with the special msr instruction - a regular mov
will not work. cpsr_c is used in the instruction in order to change CPSR without affecting the condition
flags in bits 28-31 of the CPSR value. The least significant five bits of CPSR form the mode field, so
writing to it is the way to switch processor modes. By writing 0x11 to the CPSR mode field, we switch
the processor to FIQ mode.

1
2

ldr r1, =_fiq_stack_start
ldr sp, =_fiq_stack_end

We load the end address of the FIQ stack into SP, and the start address into R1. Just setting the SP
register would be sufficient, but we’ll use the start address in R1 in order to fill the stack. The actual
addresses for _fiq_stack_end and other symbols we’re using in stack initialization will be output by
the linker with the help of our linkscript, covered later.

1
2

movw r0, #0xFEFE
movt r0, #0xFEFE

These two lines just write the value 0xFEFEFEFE to R0. ARM assembly has limitations on what values
can be directly loaded into a register with the mov instruction, so one common way to load a 4-byte
constant into a register is to use movw and movt together. In general movt r0, x; movw r0, y
corresponds to loading x << 16 | y into R0.

1 fiq_loop:
2
3
4

cmp r1, sp
strlt r0, [r1], #4
blt fiq_loop

NOTE

The strlt and blt assembly instructions you see above are examples of ARM’s conditional execution
instructions. Many instructions have conditional forms, in which the instruction has a condition code

Daniels Umanovskis

28

Bare-metal programming for ARM

ac090d8bfc

suffix. These instructions are only executed if certain flags are set. The store instruction is normally
str, and then strlt is the conditional form that will only execute if the lt condition code is met,
standing for less-than. So a simplified way to state things is that strlt will only execute if the previous
compare instruction resulted in a less-than status.

This is the loop that actually fills the stack with 0xFEFEFEFE. First it compares the value in R1 to the
value in SP. If R1 is less than SP, the value in R0 will be written to the address stored in R1, and R1 gets
increased by 4. Then the loop continues as long as R1 is less than SP.

Once the loop is over and the FIQ stack is ready, we repeat the process with the IRQ stack, and finally
the supervisor mode stack (see the code in the full listing of startup.s at the end).

Handling sections and data

A rundown on sections

To get the next steps right, we have to understand the main segments that a program normally contains,
and how they normally appear in ELF file sections.

• The code segment, normally called .text. It contains the program’s executable code, which

normally means it’s read-only and has a known size.

• The data segment, normally called .data. It contains data that can be modified by the program
at runtime. In C terms, global and static variables that have a non-zero initial value will normally
go here.

• The read-only data segment, usually with a corresponding section called .rodata, sometimes
merged with .text. Contains data that cannot be modified at runtime, in C terms constants will
be stored here.

• The unitialized data segment, also known as the BSS segment and with the corresponding
section being .bss. Don’t worry about the strange name, it has a history dating back to the
1950s. C variables that are static and have no initial value will end up here. The BSS segment
is expected to be filled with zeroes, so variables here will end up initialized to zero. Modern
compilers are also smart enough to assign variables explicitly initialized with zero to BSS, that is,
static int x = 0; would end up in .bss.

Thinking now about our program being stored in some kind of read-only memory, like the onboard
flash memory of a device, we can reason about which sections have to be copied into the RAM of the

Daniels Umanovskis

29

Bare-metal programming for ARM

ac090d8bfc

device. There’s no need to copy .text - code can be executed directly from ROM. Same for .rodata,
constants can be read from the ROM. However, .data is for modifiable variables and therefore needs
to be in RAM. .bss needs to be created in RAM and zeroed out.

Most embedded programs need to take care of ROM-to-RAM data copying. The specifics depend on
the hardware and the use case. On larger, more capable single-board computers, like the popular
Raspberry Pi series, it’s reasonable to copy even the code (.text section) to RAM and run from RAM
because of the performance benefits as opposed to reading the ROM. On microcontrollers, copying the
code to RAM isn’t even an option much of the time, as a typical ARM-based microcontroller using a
Cortex-M3 CPU or similar might have around 1 megabyte flash memory but only 96 kilobytes of RAM.

New section layout

Previously we had written linkscript.ld, a small linker script. Now we will need a more complete
linker script that will define the data sections as well. It will also need to export the start and end
addresses of sections so that copying code can use them, and finally, the linker script will also export
some symbols for the stack, like the _fiq_stack_start that we used in our stack setup code.

Normally, we would expect the program to be stored in flash or other form of ROM, as mentioned
before. With QEMU it is possible to emulate flash memory on a number of platforms, but unfortunately
not on the Versatile Express series. We’ll do something different then and pretend that a section of the
emulated RAM is actually ROM. Let’s say that the area at 0x60000000, where RAM actually starts, is
going to be treated as ROM. And let’s then say that we’ll use 0x70000000 as the pretend starting point
of RAM. There’s no need to skip so much memory - the two points are 256 MB apart - but it’s then very
easy to look at addresses and immediately know if it’s RAM from the first digit.

The first thing to do in the linker script, then, is to define the two different memory areas, our (pretend)
ROM and RAM. We do this after the ENTRY directive, using a MEMORY block.

1 MEMORY
2 {
3
4
5 }

ROM (rx) : ORIGIN = 0x60000000, LENGTH = 1M
RAM (rwx): ORIGIN = 0x70000000, LENGTH = 32M

NOTE

Daniels Umanovskis

30

Bare-metal programming for ARM

ac090d8bfc

The GNU linker, ld, can occasionally appear to be picky with the syntax. In the snippet above, the spaces
are also significant. If you write LENGTH=1M without spaces, it won’t work, and you’ll be rewarded with
a terse “syntax error” message from the linker.

Our section definitions will now be like this:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

.text : {

startup.o (.vector_table)
*(.text)
*(.rodata)

} > ROM

_text_end = .;
.data : AT(ADDR(.text) + SIZEOF(.text))
{

_data_start = .;
*(.data)
. = ALIGN(8);
_data_end = .;

} > RAM
.bss : {

_bss_start = .;
*(.bss)
. = ALIGN(8);
_bss_end = .;

} > RAM

The .text section is similar to what we had before, but we’re also going to append .rodata to it
to make life easier with one section less. We write > ROM to indicate that .text should be linked to
ROM. The _text_end symbol exports the address where .text ends in ROM and hence where the
next section starts.

.data follows next, and we use AT to specify the load address as being right after .text. We collect all
input .data sections in our output .data section and link it all to RAM, defining _data_start and
_data_end as the RAM addresses where the .data section will reside at runtime. These two symbols
are written inside the block that ends with > RAM, hence they will be RAM addresses.

.bss is handled in a similar manner, linking it to RAM and exporting the start and end addresses.

Copying ROM to RAM

As discussed, we need to copy the .data section from ROM to RAM in our startup code. Thanks to the
linker script, we know where .data starts in ROM, and where it should start and end in RAM, which is

Daniels Umanovskis

31

Bare-metal programming for ARM

ac090d8bfc

all the information we need to perform the copying. In startup.s, after dealing with the stacks we
now have the following code:

/* Start copying data */
ldr r0, =_text_end
ldr r1, =_data_start
ldr r2, =_data_end

1
2
3
4
5
6 data_loop:
7
8
9
10

cmp r1, r2
ldrlt r3, [r0], #4
strlt r3, [r1], #4
blt data_loop

We begin by preparing some data in registers. We load _text_end into R0, with _text_end being
the address in ROM where .text has ended and .data starts. _data_start is the address in RAM
at which .data should start, and _data_end is correspondingly the end address in RAM. Then the
loop itself compares R1 and R2 registers and, as long as R1 is smaller (meaning we haven’t reached
_data_end), we first load 4 bytes of data from ROM into R3, and then store these bytes at the memory
address in R1. The #4 operand in the load and store instructions ensures we’re increasing the values in
R0 and R1 correspondingly so that loop continues over the entirety of .data in ROM.

With that done, we also initialize the .bss section with this small snippet:

1
2
3
4
5 bss_loop:
6
7
8

mov r0, #0
ldr r1, =_bss_start
ldr r2, =_bss_end

cmp r1, r2
strlt r0, [r1], #4
blt bss_loop

First we store the value 0 in R0 and then loop over memory between the addresses _bss_start and
_bss_end, writing 0 to each memory address. Note how this loop is simpler than the one for .data -
there is no ROM address stored in any registers. This is because there’s no need to copy anything from
ROM for .bss, it’s all going to be zeroes anyway. Indeed, the zeroes aren’t even stored in the binary
because they would just take up space.

Daniels Umanovskis

32

Bare-metal programming for ARM

ac090d8bfc

Handing over to C

To summarize, to hand control over to C code, we need to make sure the stack is initialized, and that
the memory for modifiable variables is initialized as needed. Now that our startup code handles that,
there is no additional magic in how C code is started. It’s just a matter of calling the main function in C.
So we just do that in assembly and that’s it:

1

bl main

By using the branch-with-link instruction bl, we can continue running in case the main function returns.
However, we don’t actually want it to return, as it wouldn’t make any sense. We want to continue
running our application as long as the hardware is powered on (or QEMU is running), so a bare-metal
application will have an infinite loop of some sorts. In case main returns, we’ll just indicate an error,
same as with internal CPU exceptions.

b Abort_Exception

1
2
3 Abort_Exception:
4

swi 0xFF

And the above branch should never execute.

Into the C

We’re finally ready to leave the complexities of linker scripts and assembly, and write some code in
good old C. Create a new file, such as cstart.c with the following code:

1 #include <stdint.h>
2
3 volatile uint8_t* uart0 = (uint8_t*)0x10009000;
4
5 void write(const char* str)
6 {
7
8
9
10 }
11
12 int main() {

while (*str) {

*uart0 = *str++;

}

Daniels Umanovskis

33

Bare-metal programming for ARM

ac090d8bfc

13
14
15
16
17
18
19
20
21
22
23
24
25
26 }

const char* s = "Hello world from bare-metal!\n";
write(s);
*uart0 = 'A';
*uart0 = 'B';
*uart0 = 'C';
*uart0 = '\n';
while (*s != '\0') {
*uart0 = *s;
s++;

}
while (1) {};

return 0;

The code is simple and just outputs some text through the device’s Universal Asynchronous Receiver-
Transmitter (UART). The next part will discuss writing a UART driver in more detail, so let’s not worry
about any UART specifics for now, just note that the hardware’s UART0 (there are several UARTs)
control register is located at 0x10009000, and that a single character can be printed by writing to that
address.

It’s pretty clear that the expected output is:

1 Hello world from bare-metal!
2 ABC
3 Hello world from bare-metal!

with the first line coming from the call to write and the other two being printed from main.

The more interesting thing about this code is that it tests the stack, the read-only data section and the
regular data section. Let’s consider how the different sections would be used when linking the above
code.

1 volatile uint8_t* uart0 = (uint8_t*)0x10009000;

The uart0 variable is global, not constant, and is initialized with a non-zero value. It will therefore be
stored in the .data section, which means it will be copied to RAM by our startup code.

1 const char* s = "Hello world from bare-metal!\n";

Here we have a string, which will be stored in .rodata because that’s how GCC handles most strings.

And the lines printing individual letters, like *uart0 = 'A'; shouldn’t cause anything to be stored in

Daniels Umanovskis

34

Bare-metal programming for ARM

ac090d8bfc

any of the data sections, they will just be compiled to instructions storing the corresponding character
code in a register directly. The line printing A will do mov r2, #0x41 when compiled, with 0x41 or
65 in decimal being the ASCII code for A.

Finally, the C code also uses the stack because of the write function. If the stack has not been set up
correctly, write will fail to receive any arguments when called like write(s), so the first line of the
expected output wouldn’t appear. The stack is also used to allocate the s pointer itself, meaning the
third line wouldn’t appear either.

Building and running

There are a few changes to how we need to build the application. Assembling the startup code in
startup.s is not going to change:

1 arm-none-eabi-as -o startup.o startup.s

The new C source file needs to be compiled, and a few special link options need to be passed to GCC:

1 arm-none-eabi-gcc -c -nostdlib -nostartfiles -lgcc -o cstart.o cstart.c

With -nostdlib we indicate that we’re not using the standard C library, or any other standard libraries
that GCC would like to link against. The C standard library provides the very useful standard functions
like printf, but it also assumes that the system implements certain requirements for those functions.
We have nothing of the sort, so we don’t link with the C library at all. However, since -nostdlib
disables all default libraries, we explicitly re-add libgcc with the -lgcc flag. libgcc doesn’t provide
standard C functions, but instead provides code to deal with CPU or architecture-specific issues. One
such issue on ARM is that there is no ARM instruction for division, so the compiler normally has to
provide a division routine, which is something GCC does in libgcc. We don’t really need it now but
include it anyway, which is good practice when compiling bare-metal ARM software with GCC.

The -nostartfiles option tells GCC to omit standard startup code, since we are providing our own in
startup.s.

We’re also providing the -c switch to stop after the compilation phase. We don’t want GCC to use its
default linker script, and we’ll perform the linking in the next step with ld. Later, when defining proper
build targets, this will become streamlined.

Linking everything to obtain an ELF file has not undergone any changes, except for the addition of
cstart.o:

Daniels Umanovskis

35

Bare-metal programming for ARM

ac090d8bfc

1 arm-none-eabi-ld -T linkscript.ld -o cenv.elf startup.o cstart.o

Even though previously we booted a hang example through U-Boot by using an ELF file directly, now
we’ll need to go back to using a plain binary, which means calling objcopy to convert the ELF file into
a binary. Why is this necessary? There’s an educational aspect and the practical aspect. From the
educational side, having a raw binary is much closer to the situation we would have on real hardware,
where the on-board flash memory would be written with the raw binary contents. The practical aspect
is that U-Boot’s support of ELF files is limited. It can load an ELF into memory and boot it, but U-Boot
doesn’t handle ELF sections correctly, as it doesn’t perform any relocation. When loading an ELF file,
U-Boot just copies it to RAM and ignores how the sections should be placed. This creates problems
starting with the .text section, which will not be in the expected memory location because U-Boot
retains the ELF file’s header and any padding that might exist between it and .text. Workarounds for
these problems are possible, but using a binary file is simpler and much more reasonable.

We convert cenv.elf into a raw binary as follows:

1 arm-none-eabi-objcopy -O binary cenv.elf cenv.bin

Finally, when invoking mkimage to create the U-Boot image, we specify the binary as the input. After
that we can create the SD card image using the same create-sd.sh script we made in the previous
part.

1 mkimage -A arm -C none -T kernel -a 0x60000000 -e 0x60000000 -d cenv.

bin bare-arm.uimg

2 ./create-sd.sh sdcard.img bare-arm.uimg

That’s it for building the application! All that remains is to run QEMU (remember to specify the right
path to the U-Boot binary). One change in the QEMU command-line is that we will now use -m 512M
to provide the machine with 512 megabytes of RAM. Since we’re using RAM to also emulate ROM, we
need the memory at 0x60000000 to be accessible, but also the memory at 0x70000000. With those
addresses being 256 megabytes apart, we need to tell QEMU to emulate at least that much memory.

1 qemu-system-arm -M vexpress-a9 -m 512M -no-reboot -nographic -monitor

telnet:127.0.0.1:1234,server,nowait -kernel ../common_uboot/u-boot -
sd sdcard.img

Run QEMU as above, and you should see the three lines written to UART by our C code. There’s now a
real program running on our ARM Versatile Express!

Daniels Umanovskis

36

Bare-metal programming for ARM

ac090d8bfc

Bonus: exploring the ELF file

Dealing with linker scripts, ELF sections and relocations can be difficult. One indispensable tool is
objdump, capable of displaying all kinds of information about object files, as well as disassembling
them. Let’s look at some of the useful commands to run on our cenv.elf. First is the -h option, which
summarizes sections headers.

0 .text

1 arm-none-eabi-objdump -h cenv.elf
2
3 Sections:
4 Idx Name
5
6
7
8
9
10

Size
000001ea 60000000
CONTENTS, ALLOC, LOAD, READONLY, CODE
00000008 70000000
CONTENTS, ALLOC, LOAD, DATA
00000000 70000008
ALLOC

LMA
60000000

600001f2

600001ea

1 .data

2 .bss

File off
00010000

00020008

00020000

VMA

Algn
2**2

2**2

2**0

Of particular interest are the load address (LMA), which indicates where the section would be in ROM,
and the virtual address (VMA), which indicates where the section would be during runtime, which in
our case means RAM. We can also look at the contents of an individual section. Suppose we want to
know what’s in .data:

1 arm-none-eabi-objdump -s -j .data cenv.elf
2
3 Contents of section .data:
4

70000000 00900010 00000000

........

The 70000000 in the beginning is the address, and then we see the actual data - 00900010 can be
recognized as the little-endian 4-byte encoding of 0x10009000, the address of UART0.

Running objdump with the -t switch shows the symbol table, so for arm-none-eabi-objdump -t
cenv.elf we would see quite a bit of output, some of it containing:

1 00000011 l
2 00000012 l
3 00000013 l
4 60000034 l
5 6000004c l
6 600001ea g
7 70000008 g
8 00001000 g

*ABS* 00000000 MODE_FIQ
*ABS* 00000000 MODE_IRQ
*ABS* 00000000 MODE_SVC
.text 00000000 fiq_loop
.text 00000000 irq_loop
.text 00000000 _text_end
.bss
*ABS* 00000000 _fiq_stack_size

00000000 _bss_start

Daniels Umanovskis

37

Bare-metal programming for ARM

ac090d8bfc

9 70000008 g
10 600000dc g

.bss

00000000 _bss_end

F .text 00000054 write

You can see how smybols defined in the assembly, C function names and linker-exported symbols are
all available in the output.

Finally, running arm-none-eabi-objdump -d cenv.elf will disassemble the code, something that
usually ends up being necessary at some point in low-level embedded development.

Daniels Umanovskis

38

5 Build & debug system

This part is going to be a detour from bare-metal programming in order to quickly set up a build system
using CMake, and briefly show how our program can be debugged while running in QEMU. If you’re not
interested, you can skip this part, though at least skimming it would be recommended.

We will use CMake as our build manager. CMake provides a powerful language to define projects, and
doesn’t actually build them itself - CMake generates input for another build system, which will be GNU
Make since we’re developing on Linux. It’s also possible to use Make by itself, but for new projects I
prefer to use CMake even when its cross-platform capabilities and other powerful features are not
needed.

It should also be noted here that I am far from a CMake expert, and build systems aren’t the focus of
these articles, so this could certainly be done better.

To begin with, we’ll organize our project folder with some subfolders, and create a top-level
CMakeLists.txt, which is the input file CMake normally expects. The better-organized project
structure looks like this:

- - create-sd.sh

1 |-- CMakeLists.txt
2 |-- scripts
3 |
4
5
6
7

|-- cstart.c
|-- linkscript.ld
- - startup.s

- - src

The src folder contains all the source code, the scripts folder is for utility scripts like the one creating
our SD card image, and at the top there’s CMakeLists.txt.

We want CMake to handle the following for us:

• Rebuild U-Boot if necessary

• Build our program, including the binary converstion with objcopy

• Create the SD card image for QEMU

• Provide a way of running QEMU

39

Bare-metal programming for ARM

ac090d8bfc

To accomplish that, CMakeLists.txt can contain the following:

1 cmake_minimum_required (VERSION 2.8)
2
3 set(CMAKE_SYSTEM_NAME Generic)
4 set(CMAKE_SYSTEM_PROCESSOR arm)
5 set(CMAKE_CROSSCOMPILING TRUE)
6
7 set(UBOOT_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../common_uboot")
8 set(MKIMAGE "${UBOOT_PATH}/tools/mkimage")
9
10 project (bare-metal-arm C ASM)
11
12 set(CMAKE_C_COMPILER "arm-none-eabi-gcc")
13 set(CMAKE_ASM_COMPILER "arm-none-eabi-as")
14 set(CMAKE_OBJCOPY "arm-none-eabi-objcopy")
15
16 file(GLOB LINKSCRIPT "src/linkscript.ld")
17 set(ASMFILES src/startup.s)
18 set(SRCLIST src/cstart.c)
19
20 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -nostartfiles -nostdlib -g -Wall")
21 set(CMAKE_EXE_LINKER_FLAGS "-T ${LINKSCRIPT}")
22
23 add_custom_target(u-boot
24

COMMAND make vexpress_ca9x4_config ARCH=arm CROSS_COMPILE=

arm-none-eabi-

COMMAND make all ARCH=arm CROSS_COMPILE=arm-none-eabi-
WORKING_DIRECTORY ${UBOOT_PATH})

25
26
27
28 add_executable(bare-metal ${SRCLIST} ${ASMFILES})
29 set_target_properties(bare-metal PROPERTIES OUTPUT_NAME "bare-metal.elf

")

30 add_dependencies(bare-metal u-boot)
31
32 add_custom_command(TARGET bare-metal POST_BUILD COMMAND ${CMAKE_OBJCOPY

33

}
-O binary bare-metal.elf bare-metal.bin COMMENT "Converting ELF to

binary")

34
35 add_custom_command(TARGET bare-metal POST_BUILD COMMAND ${MKIMAGE}
36

-A arm -C none -T kernel -a 0x60000000 -e 0x60000000 -d bare-metal.

bin bare-arm.uimg

COMMENT "Building U-Boot image")

37
38
39 add_custom_command(TARGET bare-metal POST_BUILD COMMAND bash ${

40
41

CMAKE_CURRENT_SOURCE_DIR}/scripts/create-sd.sh
sdcard.img bare-arm.uimg
COMMENT "Creating SD card image")

Daniels Umanovskis

40

Bare-metal programming for ARM

ac090d8bfc

42
43 add_custom_target(run)
44 add_custom_command(TARGET run POST_BUILD COMMAND
45

qemu-system-arm -M vexpress-a9 -m 512M -no-reboot -

46

47

nographic

-monitor telnet:127.0.0.1:1234,server,nowait -kernel $
{UBOOT_PATH}/u-boot -sd sdcard.img -serial mon:
stdio

COMMENT "Running QEMU...")

We begin with some preparation, telling CMake that we’ll be cross-compiling so it doesn’t assume it’s
building a Linux application, and we store the U-Boot path in a variable. Then there’s the specification
of the project itself:

1 project (bare-metal-arm C ASM)

bare-metal-arm is an arbitrary project name, and we indicate that C and assembly files are used in
it. The next few lines explicitly specify the cross-compiler toolchain elements to be used. After that, we
list the source files included in the project:

1 file(GLOB LINKSCRIPT "src/linkscript.ld")
2 set(ASMFILES src/startup.s)
3 set(SRCLIST src/cstart.c)

Some people prefer to specify wildcards so that all .c files are included automatically, but CMake
guidelines recommend against this approach. This is a matter of debate, but here I chose to go with
explicitly listed files, meaning that new files will need to manually be added here.

CMake lets us specify flags to be passed to the compiler and linker, which we do:

1 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -nostartfiles -nostdlib -g -Wall")
2 set(CMAKE_EXE_LINKER_FLAGS "-T ${LINKSCRIPT}")

Those are like what we used previously, except for the addition of -g -Wall. The -g switch enables
generation of debug symbols, which we’ll need soon, and -Wall enables all compiler warnings and is
very good to use as a matter of general practice.

We then define a custom target to build U-Boot, which just runs U-Boot’s regular make commands.
After that we’re ready to define our main target, which we say is an executable that should be built out
of source files in the SRCLIST and ASMFILES variables, and should be called bare-metal.elf:

Daniels Umanovskis

41

Bare-metal programming for ARM

ac090d8bfc

1 add_executable(bare-metal ${SRCLIST} ${ASMFILES})
2 set_target_properties(bare-metal PROPERTIES OUTPUT_NAME "bare-metal.elf

")

The two subsequent uses of add_custom_command are to invoke objcopy and call the create-sd.
sh script to build the U-Boot uimage.

That’s everything we need to build our entire program, all the way to the U-Boot image containing it. It
is, however, also convenient to have a way of running QEMU from the same environment, so we also
define a target called run and provide the QEMU command-line to it. The addition of -serial mon:
stdio in the QEMU command line means that we’ll be able to issue certain QEMU monitor commands
directly in the terminal, giving us a cleaner shutdown option.

Building and running

When building the project, I strongly recommend doing what’s known as an out of source build. It
simply means that all the files resulting from the build should go into a separate folder, and not your
source folder. This is cleaner (no mixing of source and object files), easier to use with Git (just ignore
the whole build folder), and allows you to have several builds at the same time.

To build with CMake, first you need to tell CMake to generate the build configuration from CMakeLists
.txt. The easiest way to perform a build is:

1 cmake -S . -Bbuild

The -S . option says to look for the source, starting with CMakeLists.txt, in the current folder,
and -Bbuild specifies that a folder called build should contain the generated configuration. After
running that command, you’ll have the build folder containing configurations generated by CMake.
You can then use make inside that folder. There’s no need to call CMake itself again unless the build
configuration is supposed to change. If you, for instance, add a new source file to the project, you need
to include it in CMakeLists.txt and call CMake again.

From the newly created build folder, simply invoking the default make target will build everything.

1 make

You’ll see compilation of U-Boot and our own project, and the other steps culminating in the creation
of sdcard.img. Since we also defined a CMake target to run QEMU, it can also be invoked directly
after building. Just do:

Daniels Umanovskis

42

Bare-metal programming for ARM

ac090d8bfc

1 make run

and QEMU will run our program. Far more convenent than what we had been doing.

HINT

If you run QEMU with make run and then terminate it with Ctrl-C, you’ll get messages about the target
having failed. This is harmless but doesn’t look nice. Instead, you can cleanly quit QEMU with Ctrl-A,
X (that is Ctrl-A first and then X without the Ctrl key). It’s a feature of the QEMU monitor, and works
because of adding -serial mon:stdio to the QEMU command-line.

Debugging in QEMU with GDB

While the QEMU monitor provides many useful features, it’s not a proper debugger. When running
software on a PC through QEMU, as opposed to running on real hardware, it would be a waste not to
take advantage of the superior debug capabilities available. We can debug our bare-metal program
using the GDB, the GNU debugger. GDB provides remote debugging capabilities, with a server called
gdbserver running on the machine to be debugged, and then the main gdb client communicatng
with the server. QEMU is able to start an instance of gdbserver along with the program it’s emulating,
so remote debugging is a possibility with QEMU and GDB.

Starting gdbserver when running QEMU is as easy as adding -gdb tcp::2159 to the QEMU
command line (2159 is the standard port for GDB remote debugging). Given that we’re using
CMake, we can use it to define a new target for a debug run of QEMU. These are the additions in
CMakeLists.txt:

1 string(CONCAT GDBSCRIPT "target remote localhost:2159\n"
2
3 file(WRITE ${CMAKE_BINARY_DIR}/gdbscript ${GDBSCRIPT})
4
5 add_custom_target(drun)
6 add_custom_command(TARGET drun PRE_BUILD COMMAND ${CMAKE_COMMAND} -E

"file bare-metal.elf")

cmake_echo_color --cyan

7

"To connect the debugger, run arm-none-eabi-gdb -x

gdbscript")

Daniels Umanovskis

43

Bare-metal programming for ARM

ac090d8bfc

8 add_custom_command(TARGET drun PRE_BUILD COMMAND ${CMAKE_COMMAND} -E

cmake_echo_color --cyan

9
10
11 add_custom_command(TARGET drun POST_BUILD COMMAND
12

"To start execution, type continue in gdb")

qemu-system-arm -S -M vexpress-a9 -m 512M -no-reboot -

13

14

nographic -gdb tcp::2159

-monitor telnet:127.0.0.1:1234,server,nowait -kernel $
{UBOOT_PATH}/u-boot -sd sdcard.img -serial mon:
stdio

COMMENT "Running QEMU with debug server...")

The drun target (for debug run) adds -gdb tcp::2159 to start gdbserver, and -S, which tells QEMU
not to start execution after loading. That option is useful for debugging because it gives you the time
to set breakpoints, letting you debug the code very early if you need to.

When debugging remotely, GDB needs to know what server to connect to, and where to get the debug
symbols. We can connect using the GDB command target remote localost:2159 and then load
the ELF file using file bare-metal.elf. To avoid typing those commands manually all the time, we
ask CMake to put them into a file called gdbscript that GDB can read when started.

Let’s rebuild and try a quick debug session.

1 cmake -S . -Bbuild
2 cd build
3 make
4 make drun

You should see CMake print some hints that we provided, and then QEMU will wait doing nothing. In
another terminal now, you can start GDB from the build folder, telling it to read commands from the
gdbscript file:

1 arm-none-eabi-gdb -x gdbscript

If you’re using a Linux distribution from 2018 or later (Ubuntu 18.04 or Debian 10 for example), there
might be no arm-none-eabi-gdb. In that case, run gdb-multiarch instead (after installing with
sudo apt-get install gdb-multiarch if needed).

Now you have GDB running and displaying (gdb), its command prompt. You can set a breakpoint
using the break command, let’s try to set one in the main function, and then continue execution with
c (short for continue):

Daniels Umanovskis

44

Bare-metal programming for ARM

ac090d8bfc

1 (gdb) break main
2 (gdb) c

As soon as you issue the c command, you’ll see QEMU running. After U-Boot output, it will stop again,
and GDB will show that it hit a breakpoint, something like the following:

1 Breakpoint 1, main () at /some/path/to/repo/src/05_cmake/src/cstart.c

:13

2 13

const char* s = "Hello world from bare-metal!\n";

From there, you can use the n command to step through the source code, info stack to see the stack
trace, and any other GDB commands.

I won’t be covering GDB in additional detail here, that’s outside the scope of these tutorials. GDB has a
comprehensive, if overwhelming, manual, and there’s a lot more material available online. Beej’s guide
to GDB, authored by Brian Hall, is perhaps the best getting-started guide for GDB. If you’d rather use a
graphical front-end, there is also a large selection. When looking for GDB-related information online,
don’t be alarmed if you find old articles - GDB dates back to the 1980s, and while it keeps getting new
features, the basics haven’t changed in a very long time.

Our project now has a half-decent build system, we are no longer relying on manual steps, and can
debug our program. This is a good place to continue from!

Daniels Umanovskis

45

6 UART driver development

This chapter will concern driver development, a crucial part of bare-metal programming. We will walk
through writing a UART driver for the Versatile Express series, but the ambition here is not so much to
cover that particular UART in detail as it is to show the general approach and patterns when writing a
similar driver. As always with programming, there is a lot that can be debated, and there are parts that
can be done differently. Starting with a UART driver specifically has its advantages. UARTs are very
common peripherals, they’re much simpler than other serial buses such as SPI or I2C, and the UART
pretty much corresponds to standard input/output when run in QEMU.

Doing the homework

Writing a peripheral driver is not something you should just jump into. You need to understand the
device itself, and how it integrates with the rest of the hardware. If you start coding off the hip, you’re
likely to end up with major design issues, or just a driver that mysteriously fails to work because you
missed a small but crucial detail. Thinking before doing should apply to most programming, but driver
programming is particularly unforgiving if you fail to follow that rule.

Before writing a peripheral device driver, we need to understand, in broad strokes, the following about
the device:

• How it performs its function(s). Whether it’s a communication device, a signal converter, or
anything else, there are going to be many details of how the device operates. In the case of a
UART device, some of the things that fall here are, what baud rates does it support? Are there
input and output buffers? When does it sample incoming data?

• How it is controlled. Most of the time, the peripheral device will have several registers, writing

and reading them is what controls the device. You need to know what the registers do.

• How it integrates with the hardware. When the device is part of a larger system, which could be a
system-on-a-chip or a motherboard-based design, it somehow connects to the rest of the system.
Does the device take an external input clock and, if so, where from? Does enabling the device
require some other system conditions to be met? The registers for controlling the device are
somehow accessible from the CPU, typically by being mapped to a particular memory address.

46

Bare-metal programming for ARM

ac090d8bfc

From a CPU perspective, registers that control peripherals are Special Function Registers (SFR),
though not all SFRs correspond to peripherals.

Let’s then look at the UART of the Versatile Express and learn enough about it to be ready to create the
driver.

Basic UART operation

UART is a fairly simple communications bus. Data is sent out on one wire, and received on another. Two
UARTs can thus communicate directly, and there is no clock signal or synchronization of any kind, it’s
instead expected that both UARTs are configured to use the same baud rate. UART data is transmitted
in packets, which always begin with a start bit, followed by 5 to 9 data bits, then optionally a parity bit,
then 1 or 2 stop bits. A packet could look like this:

1 +-------+---+---+---+---+---+---+---+---+---+---+
2 | Start | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | P | S |
3 | bit
|
|
4 |
5 +-------+---+---+---+---+---+---+---+---+---+---+

|
|

|
|

|
|

|
|

|
|

|
|

|
|

|
|

|
|

|
|

A pair of UARTs needs to use the same frame formatting for successful communication. In practice, the
most common format is 8 bits of data, no parity bit, and 1 stop bit. This is sometimes written as 8-N-1
in shorthand, and can be written together with the baud rate like 115200/8-N-1 to describe a UART’s
configuration.

On to the specific UART device that we have. The Versatile Express hardware series comes with the
PrimeCell UART PL011, the reference manual is also available from the ARM website. Reading through
the manual, we can see that the PL011 is a typical UART with programmable baud rate and packet
format settings, that it supports infrared transmission, and FIFO buffers both for transmission and
reception. Additionally there’s Direct Memory Access (DMA) support, and support for hardware flow
control. In the context of UART, hardware flow control means making use of two additional physical
signals so that one UART can inform another when it’s ready to send, or ready to receive, data.

Key PL011 registers

The PL011 UART manual also describes the registers that control the peripheral, and we can identify
the most important registers that our driver will need to access in order to make the UART work. We
will need to work with:

• Data register (DR). The data received, or to be transmitted, will be accessed through DR.

Daniels Umanovskis

47

Bare-metal programming for ARM

ac090d8bfc

• Receive status / error clear register (RSRECR). Errors are indicated here, and the error flag can

also be cleared from here.

• Flag register (FR). Various flags indicating the UART’s status are collected in FR.

• Integer baud rate register (IBRD) and Fractional baud rate register (FBRD). Used together to set

the baud rate.

• Line control register (LCR_H). Used primarily to program the frame format.

• Control register (CR). Global control of the peripheral, including turning it on and off.

In addition, there are registers related to interrupts, but we will begin by using polling, which is
inefficient but simpler. We will also not care about the DMA feature.

As is often the case, reading register descriptions in the manual also reveals some special considerations
that apply to the particular hardware. For example, it turns out that writing the IBRD or FBRD registers
will not actually have any effect until writing the LCR_H - so even if you only want to update IBRD, you
need to perform a sequence of two writes, one to IBRD and another to LCR_H. It is very common in
embedded programming to run into such special rules for reading or writing registers, which is one of
the reasons reading the manual for the device you’re about to program is so important.

PL011 - Versatile Express integration

Now that we are somewhat familiar with the PL011 UART peripheral itself, it’s time to look at how
integrates with the Versatile Express hardware. The VE hardware itself consists of a motherboard and
daughter board, and the UARTs are on the motherboard, which is called the Motherboard Express µATX
and of course has its own reference manual.

One important thing from the PL011 manual is the reference clock, UARTCLK. Some peripherals have
their own independent clock, but most of them, especially simpler peripherals, use an external
reference clock that is then often divided further as needed. For external clocks, the peripheral’s
manual cannot provide specifics, so the information on what the clock is has to be found elsewhere. In
our case, the motherboard’s documentation has a separate section on clocks (2.3 Clock architecture in
the linked PDF), where we can see that UARTs are clocked by the OSC2 clock from the motherboard,
which has a frequency of 24 MHz. This is very convenient, we will not need to worry about the reference
clock possibly having different values, we can just say it’s 24 MHz.

Next we need to find where the UART SFRs are located from the CPU’s perspective. The motherboard’s
manual has memory maps, which differ depending on the daughter board. We’re using the CoreTile
Express A9x4, so it has what the manual calls the ARM Legacy memory map in section 4.2. It says that
the address for UART0 is 0x9000, using SMB (System Memory Bus) chip select CS7, with the chip select
introducing an additional offset that the daughter board defines. Then it’s on to the CoreTile Express

Daniels Umanovskis

48

Bare-metal programming for ARM

ac090d8bfc

A9x4 manual, which explains that the board’s memory controller places most motherboard peripherals
under CS7, and in 3.2 Daughterboard memory map we see that CS7 memory mappings for accessing
the motherboard’s peripherals start at 0x10000000. Thus the address of UART0 from the perspective
of the CPU running our code is CS7 base 0x10000000 plus an offset of 0x9000, so 0x10009000 is
ultimately the address we need.

Yes, this means that we have to check two different manuals just to find the peripheral’s address. This
is, once again, nothing unusual in an embedded context.

Writing the driver

What’s in the box?

In higher-level programming, you can usually treat drivers as a black box, if you even give them any
consideration. They’re there, they do things with hardware, and they only have a few functions you’re
exposed to. Now that we’re writing a driver, we have to consider what it consists of, the things we need
to implement. Broadly, we can say that a driver has:

• An initialization function. It starts the device, performing whatever steps are needed. This is

usually relatively simple.

• Configuration functions. Most devices can be configured to perform their functions differently.
For a UART, programming the baud rate and frame format would fall here. Configuration can be
simple or very complex.

• Runtime functions. These are the reason for having the driver in the first place, the interesting

stuff happens here. In the case of UART, this means functions to transmit and read data.

• A deinitialization function. It turns the device off, and is quite often omitted.

• Interrupt handlers. Most peripherals have some interrupts, which need to be handled in special
functions called interrupt handlers, or interrupt service routines. We won’t be covering that for
now.

Now we have a rough outline of what we need to implement. We will need code to start and configure
the UART, and to send and receive data. Let’s get on with the implementation.

Exposing the SFRs

We know by now that programming the UART will be done by accessing the SFRs. It is possible, of
course, to access the memory locations directly, but a better way is to define a C struct that reflects

Daniels Umanovskis

49

Bare-metal programming for ARM

ac090d8bfc

what the SFRs look like. We again refer to the PL011 manual for the register summary. It begins like
this:

Figure 6.1: PL011 register summary

Looking at the table, we can define it in code as follows:

1 typedef volatile struct __attribute__((packed)) {
2
3

/* 0x0 Data Register */
/* 0x4 Receive status / error clear

uint32_t DR;
uint32_t RSRECR;
register */

4
5
6
7
8
9
10
11
12 } uart_registers;

uint32_t _reserved0[4];
const uint32_t FR;
uint32_t _reserved1;
uint32_t ILPR;
uint32_t IBRD;
uint32_t FBRD;
uint32_t LCRH;
uint32_t CR;

/* 0x8 - 0x14 reserved */
/* 0x18 Flag register */
/* 0x1C reserved */
/* 0x20 Low-power counter register */
/* 0x24 Integer baudrate register */
/* 0x28 Fractional baudrate register */
/* 0x2C Line control register */
/* 0x30 Control register */

There are several things to note about the code. One is that it uses fixed-size types like uint32_t.
Since the C99 standard was adopted, C has included the stdint.h header that defines exact-width

Daniels Umanovskis

50

Bare-metal programming for ARM

ac090d8bfc

integer types. So uint32_t is guaranteed to be a 32-bit type, as opposed to unsigned int, for which
there is no guaranteed fixed size. The layout and size of the SFRs is fixed, as described in the manual,
so the struct has to match in terms of field sizes.

For the same reason, __attribute__((packed)) is provided. Normally, the compiler is allowed
to insert padding between struct fields in order to align the whole struct to some size suited for the
architecture. Consider the following example:

1 typedef struct {
2
3
4
5 } example;

char a; /* 1 byte */
int b; /* 4 bytes */
char c; /* 1 byte */

If you compile that struct for a typical x86 system where int is 4 bytes, the compiler will probably try
to align the struct to a 4-byte boundary, and align the individual members to such a boundary as well,
inserting 3 bytes after a and another 3 bytes after c, giving the struct a total size of 12 bytes.

When working with SFRs, we definitely don’t want the compiler to insert any padding bytes or take
any other liberties with the code. __attribute__((packed)) is a GCC attribute (also recognized by
some other compilers like clang) that tells the compiler to use the struct as it is written, using the least
amount of memory possible to represent it. Forcing structs to be packed is generally not a great idea
when working with “normal” data, but it’s very good practice for structs that represent SFRs.

Sometimes there might be reserved memory locations between various registers. In our case of
the PL011 UART, there are reserved bytes between the RSRECR and FR registers, and four more after
FR. There’s no general way in C to mark such struct fields as unusable, so giving them names like
_reserved0 indicates the purpose. In our struct definition, we define uint32_t _reserved0[4];
to skip 16 bytes, and uint32_t _reserved1; to skip another 4 bytes later.

Some SFRs are read-only, like the FR, in which case it’s helpful to declare the corresponding field as
const. Attempts to write a read-only register would fail anyway (the register would remain unchanged),
but marking it as const lets the compiler check for attempts to write the register.

Having defined a struct that mimics the SFR layout, we can create a static variable in our driver that
will point to the UART0 device:

1 static uart_registers* uart0 = (uart_registers*)0x10009000u;

A possible alternative to the above would be to declare a macro that would point to the SFRs,
such as #define UART0 ((uart_registers*)0x10009000u). That choice is largely a matter of
preference.

Daniels Umanovskis

51

Bare-metal programming for ARM

ac090d8bfc

Register access width

An important, but easy to overlook, aspect of writing to SFRs is access width. A device might require
its SFRs to be written all at once, with one write instruction, or the device might introduce other
constraints. The corresponding device manual should indicate how it expects the registers to be
accessed. Commonly encountered types of access are:

• Word access. The access operation should have the same width as the machine word, e.g. 32

bits on a 32-bit system. A SFR is usually the size of one word.

• Half-word access. As the name indicates, this means accessing half of a word at a time, so 16 bits

on a 32-bit system.

• Byte access. Any byte within the register can be written individually.

Requiring word access, possibly with allowing half-word access, is common. Allowing byte access to
SFRs is somewhat less common.

What does this mean in practice, and how to make sure access is of the right width? You have to be
aware of how your code is going to access SFRs, consider the following:

1 sfr->SOMEFIELD |= 0xFu;

Most likely, the line would result in assembly code that performs a whole-word write, such as, if the
address of SOMEFIELD is in register R0

1 ldr r1, [r0] ; load value in SOMEFIELD into R1
2 orr r2, r1, #15 ; save SOMEFIELD | 0xF into R2
3 str r2, [r0] ; write back to SOMEFIELD

or other optimized code that would be even better. However, it’s possible that the line would be
compiled into code that performs multiple accesses, i.e. one for each byte of SOMEFIELD.

A single word-wide write can be ensured by explicitly asking for a write to a uint32_t* such as:

1 *(uint32_t*)sfr->SOMEFIELD = some_val;

This is not particularly important for the PL011 UART specifically, which does not specify any restrictions
on register access, but using word access explicitly would be good practice nonetheless. In the next
chapter, dealing with the interrupt controller, register access width becomes more important.

Daniels Umanovskis

52

Bare-metal programming for ARM

ac090d8bfc

Initializing and configuring the UART

Let’s now write uart_configure(), which will initialize and configure the UART. For some drivers
you might want a separate init function, but a uart_init() here wouldn’t make much sense, the
device is quite simple. The functionitself is not particularly complex either, but can showcase some
patterns.

First we need to define a couple of extra types. For the return type, we want something that can indicate
failure or success. It’s very useful for functions to be able to indicate success or failure, and driver
functions can often fail in many ways. Protecting against possible programmer errors is of particular
interest - it’s definitely possible to use the driver incorrectly! So one of the approaches is to define error
codes for each driver (or each driver type perhaps), like the following:

1 typedef enum {
2
3
4
5
6
7
8 } uart_error;

UART_OK = 0,
UART_INVALID_ARGUMENT_BAUDRATE,
UART_INVALID_ARGUMENT_WORDSIZE,
UART_INVALID_ARGUMENT_STOP_BITS,
UART_RECEIVE_ERROR,
UART_NO_DATA

A common convention is to give the success code a value of 0, and then we add some more error
codes to the enumeration. Let’s use this uart_error type as the return type for our configuration
function.

Then we need to pass some configuration to the driver, in order to set the baud rate, word size, etc.
One possibility is to define the following struct describing a config:

1 typedef struct {
uint8_t
2
uint8_t
3
bool
4
5
uint32_t
6 } uart_config;

data_bits;
stop_bits;
parity;
baudrate;

This approach, of course, dictates that uart_configure would take a parameter of the uart_config
type, giving us:

1 uart_error uart_configure(uart_config* config)

Daniels Umanovskis

53

Bare-metal programming for ARM

ac090d8bfc

There are other possible design choices. You could omit the struct, and just pass in multiple
parameters, like uart_configure(uint8_t data_bits, uint8_t stop_bits, bool parity
, unit32_t baudrate).
I prefer a struct because those values logically belong together. Yet
another option would be to have separate functions per parameter, such as uart_set_baudrate
and uart_set_data_bits, but I think that is a weaker choice, as it can create issues with the order
in which those functions are called.

On to the function body. You can see the entire source in the corresponding file for this chapter, and
here I’ll go through it block by block.

First, we perform some validation of the configuration, returning the appropriate error code if some
parameter is outside the acceptable range.

1
2
3
4
5
6
7
8
9

if (config->data_bits < 5u || config->data_bits > 8u) {

return UART_INVALID_ARGUMENT_WORDSIZE;

}
if (config->stop_bits == 0u || config->stop_bits > 2u) {

return UART_INVALID_ARGUMENT_STOP_BITS;

}
if (config->baudrate < 110u || config->baudrate > 460800u) {

return UART_INVALID_ARGUMENT_BAUDRATE;

}

UART only allows 5 to 8 bits as the data size, and the only choices for the stop bit is to have one or two.
For the baudrate check, we just constrain the baudrate to be between two standard values.

With validation done, the rest of the function essentially follows the PL011 UART’s manual for how to
configure it. First the UART needs to be disabled, allowed to finish an ongoing transmission, if any, and
its transmit FIFO should be flushed. Here’s the code:

1
2
3
4
5

/* Disable the UART */
uart0->CR &= ~CR_UARTEN;
/* Finish any current transmission, and flush the FIFO */
while (uart0->FR & FR_BUSY);
uart0->LCRH &= ~LCRH_FEN;

Having a similar while loop is common in driver code when waiting on some hardware process. In
this case, the PL011’s FR has a BUSY bit that indicates if a transmission is ongoing. Setting the FEN bit
in LCRH to 0 is the way to flush the transmit queue.

What about all those defines like CR_UARTEN in the lines above though? Here they are from the
corresponding header file:

Daniels Umanovskis

54

Bare-metal programming for ARM

ac090d8bfc

1 #define FR_BUSY
2 #define LCRH_FEN
3 #define CR_UARTEN

(1 << 3u)
(1 << 4u)
(1 << 0u)

Typically, one SFR has many individual settings, with one setting often using just one or two bits.
The locations of the bits are always in the corresponding manual, but using them directly doesn’t
make for the most readable code. Consider uart0->CR &= ~(1u) or while (uart0->FR & (1
<< 3u)). Any time you read such a line, you’d have to refer to the manual to check what the bit or
mask means. Symbolic names make such code much more readable, and here I use the pattern of
SFRNAME_BITNAME, so CR_UARTEN is the bit called UARTEN in the CR SFR. I won’t include more of
those defines in this chapter, but they’re all in the full header file.

NOTE

Bit manipulation is usually a very important part of driver code, such as the above snippet. Bitwise
operators and shift operators are a part of C, and I won’t be covering them here. Hopefully you’re
familiar enough with bit manipulation to read the code presented here. Just in case though, this cheat
sheet might be handy:

Assuming that b is one bit,

x |= b sets bit b in x

x &= ~b clears bit b in x

x & b checks if b is set

One bit in position n can be conveniently written as 1 left-shifted n places. E.g. bit 4 is 1 << 4 and bit
0 is 1 << 0

Next we configure the UART’s baudrate. This is another operation that translates to fairly simple code,
but requires a careful reading of the manual. To obtain a certain baudrate, we need to divide the
(input) reference clock with a certain divisor value. The divisor value is stored in two SFRs, IBRD for
the integer part and FBRD for the fractional part. Accordig to the manual, baudrate divisor =
reference clock / (16 * baudrate). The integer part of that result is used directly, and the
fractional part needs to be converted to a 6-bit number m, where m = integer((fractional part

* 64)+ 0.5). We can translate that into C code as follows:

Daniels Umanovskis

55

Bare-metal programming for ARM

ac090d8bfc

1
2

3
4
5
6

double intpart, fractpart;
double baudrate_divisor = (double)refclock / (16u * config->

baudrate);

fractpart = modf(baudrate_divisor, &intpart);

uart0->IBRD = (uint16_t)intpart;
uart0->FBRD = (uint8_t)((fractpart * 64u) + 0.5);

It’s possible to obtain the fractional part with some arithmetics, but we can just use the standard C
modf function that exists for that purpose and is available after including <math.h>. While we cannot
use the entire C standard library on bare-metal without performing some extra work, mathematical
functions do not require anything extra, so we can use them.

Since our reference clock is 24 MHz, as we established before, the refclock variable is 24000000u.
Assuming that we want to set a baudrate of 9600, first the baudrate_divisor will be calculated
as 24000000 / (16 * 9600), giving 156.25. The modf function will helpfully set intpart to 156
and fractpart to 0.25. Following the manual’s instructions, we directly write the 156 to IBRD, and
convert 0.25 to a 6-bit number. 0.25 * 64 + 0.5 is 16.5, we only take the integer part of that,
so 16 goes into FBRD. Note that 16 makes sense as a representation of 0.25 if you consider that the
largest 6-bit number is 63.

We continue now by setting up the rest of the configuration - data bits, parity and the stop bit.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

uint32_t lcrh = 0u;

/* Set data word size */
switch (config->data_bits)
{
case 5:

lcrh |= LCRH_WLEN_5BITS;
break;

case 6:

lcrh |= LCRH_WLEN_6BITS;
break;

case 7:

lcrh |= LCRH_WLEN_7BITS;
break;

case 8:

lcrh |= LCRH_WLEN_8BITS;
break;

}

/* Set parity. If enabled, use even parity */
if (config->parity) {
lcrh |= LCRH_PEN;
lcrh |= LCRH_EPS;

Daniels Umanovskis

56

Bare-metal programming for ARM

ac090d8bfc

24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

lcrh |= LCRH_SPS;

} else {

lcrh &= ~LCRH_PEN;
lcrh &= ~LCRH_EPS;
lcrh &= ~LCRH_SPS;

}

/* Set stop bits */
if (config->stop_bits == 1u) {

lcrh &= ~LCRH_STP2;

} else if (config->stop_bits == 2u) {

lcrh |= LCRH_STP2;

}

/* Enable FIFOs */
lcrh |= LCRH_FEN;

uart0->LCRH = lcrh;

That is a longer piece of code, but there’s not much remarkable about it. For the most part it’s just
picking the correct bits to set or clear depending on the provided configuration. One thing to note is
the use of the temporary lcrh variable where the value is built, before actually writing it to the LCRH
register at the end. It is sometimes necessary to make sure an entire register is written at once, in which
case this is the technique to use. In the case of this particular device, LCRH can be written bit-by-bit,
but writing to it also triggers updates of IBRD and FBRD, so we might as well avoid doing that many
times.

At the end of the above snippet, we enable FIFOs for potentially better performance, and write the
LCRH as discussed. After that, everything is configured, and all that remains is to actually turn the UART
on:

1

uart0->CR |= CR_UARTEN;

Read and write functions

We can start the UART with our preferred configuration now, so it’s a good time to implement functions
that actually perform useful work - that is, send and receive data.

Code for sending is very straightforward:

1 void uart_putchar(char c) {
2

while (uart0->FR & FR_TXFF);

Daniels Umanovskis

57

Bare-metal programming for ARM

ac090d8bfc

uart0->DR = c;

3
4 }
5
6 void uart_write(const char* data) {
7
8
9
10 }

uart_putchar(*data++);

while (*data) {

}

Given any string, we just output it one character at a time. The TXFF bit in FR that uart_putchar()
waits for indicates a full transmit queue - we just wait until that’s no longer the case.

These two functions have void return type, they don’t return uart_error. Why? It’s again a design
decision, meaning you could argue against it, but the write functions don’t have any meaningful way of
detecting errors anyway. Data is sent out on the bus and that’s it. The UART doesn’t know if anybody’s
receiving it, and it doesn’t have any error flags that are useful when transmitting. So the void return
type here is intended to suggest that the function isn’t capable of providing any useful information
regarding its own status.

The data reception code is a bit more interesting because it actually has error checks:

}

return UART_NO_DATA;

if (uart0->FR & FR_RXFE) {

1 uart_error uart_getchar(char* c) {
2
3
4
5
6
7
8
9
10
11
12
13 }

}
return UART_OK;

/* The character had an error */
uart0->RSRECR &= RSRECR_ERR_MASK;
return UART_RECEIVE_ERROR;

*c = uart0->DR & DR_DATA_MASK;
if (uart0->RSRECR & RSRECR_ERR_MASK) {

First it checks if the receive FIFO is empty, using the RXFE bit in FR. Returning UART_NO_DATA in that
case tells the user of this code not to expect any character. Otherwise, if data is present, the function
first reads it from the data register DR, and then checks the corresponding error status - it has to be
done in this order, once again according to the all-knowing manual. The PL011 UART can distinguish
between several kinds of errors (framing, parity, break, overrun) but here we treat them all the same,
using RSRECR_ERR_MASK as a bitmask to check if any error is present. In that case, a write to the
RSRECR register is performed to reset the error flags.

Daniels Umanovskis

58

Bare-metal programming for ARM

ac090d8bfc

Putting it to use

We need some code to make use of our new driver! One possibility is to rewrite cstart.c like the
following:

if (!strncmp("help\r", buf, strlen("help\r"))) {

uart_write("Just type and see what happens!\n");
} else if (!strncmp("uname\r", buf, strlen("uname\r"))) {

uart_write("bare-metal arm 06_uart\n");

}

uart_config config = {
.data_bits = 8,
.stop_bits = 1,
.parity = false,
.baudrate = 9600

1 #include <stdint.h>
2 #include <stdbool.h>
3 #include <string.h>
4 #include "uart_pl011.h"
5
6 char buf[64];
7 uint8_t buf_idx = 0u;
8
9 static void parse_cmd(void) {
10
11
12
13
14
15 }
16
17 int main() {
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

uart_putchar('A');
uart_putchar('B');
uart_putchar('C');
uart_putchar('\n');

}

};
uart_configure(&config);

uart_write("I love drivers!\n");
uart_write("Type below...\n");

while (1) {
char c;
if (uart_getchar(&c) == UART_OK) {

uart_putchar(c);
buf[buf_idx % 64] = c;
buf_idx++;
if (c == '\r') {

uart_write("\n");
buf_idx = 0u;
parse_cmd();

Daniels Umanovskis

59

Bare-metal programming for ARM

ac090d8bfc

45
46
47
48
49 }

}

}

return 0;

The main function asks the UART driver to configure it for 9600/8-N-1, a commonly used mode,
and then outputs some text to the screen much as the previous chapter’s example did. Some more
interesting things happen within the while loop now though - it constantly polls the UART for incoming
data and appends any characters read to a buffer, and prints that same character back to the screen.
Then, when a carriage return (\r) is read, it calls parse_cmd().That’s a very basic method of waiting
for something to be input and reacting on the Enter key.

parse_cmd() is a simple function that has responses in case the input line was help or uname. This
way, without writing anything fancy, we grant our bare-metal program the ability to respond to user
input!

Doing a test run

To build our program with the new driver, it needs to be added to the source file list in CMake, and we
need to change a couple of flags as well. To add the driver file, just add the file to the appropriate list in
CMakeLists.txt:

1 set(SRCLIST src/cstart.c src/uart_pl011.c)

We are using the C standard library now, and we need to link against libm, so the compiler and linker
flags should now look like this:

1 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -nostartfiles -g -Wall")
2 set(CMAKE_EXE_LINKER_FLAGS "-T ${LINKSCRIPT} -lgcc -lm")

When building C programs with GCC, the -lm flag is necessary if the program uses mathematical
functions declared in the standard header math.h. For the rest of the standard library, no separate
flag is needed. This special treatment of libm stems from technical decisions made a long time ago,
and is by now best treated simply as a bit of legacy to remember.

Rebuild everything (since we changed CMakeLists.txt that means you also need to invoke cmake
), run the program in QEMU and, after some output, you should be able to type into the terminal
and see what you type. That’s the UART driver at work! Each character you type gets returned by

Daniels Umanovskis

60

Bare-metal programming for ARM

ac090d8bfc

uart_getchar and then output to the screen by uart_putchar. Try writing help and hitting the
Enter key - you should see the output as defined in parse_cmd.

Unfortunately, this is also the first point where using QEMU is a disadvantage. QEMU emulation of
devices has limitations, and the PL011 UART is no exception. Our driver will work in QEMU no matter
the baudrate. No matter if we connect the UART to standard input/output (as now) or a Linux character
device, the baudrate will not matter. There’s also no enable/disable mechanism for the emulated UART
- the driver would work even if we never enabled the device in the CR register.

There’s no real way around those issues short of running on real hardware, the best we can do is try
and write the driver as if for the actual hardware device.

Summary

We’ve written our first driver that interfaces with the hardware directly, and we wrote some proof-of-
concept code making use of the driver. A big takeaway is that carefully reading the manual is at least
half the work involved in writing a driver. In writing the actual driver code, we also saw how it can be
quite different from most non-driver code. There’s a lot of bit manipulation, and operations that are
order-sensitive in ways that may not be intuitive. The fact that the same location in memory, like the
DR SFR, acts differently when being read versus written is also something rarely encountered outside
of driver code.

Unsurprisingly, the driver written in this chapter is not perfect. Some possibilities for improvement:

• Interrupt handling. Currently the driver is being used in polling mode, constantly asking it if new
characters have been received. In most practical cases, polling is too inefficient and interrupts
are desired. This is something that the next chapter handles.

• More robustness. Error handling, sanity checks and other measures preventing the driver from
being incorrectly are good! The driver could return different error codes for different types of
receive errors instead of lumping them all together. The driver could keep track of its own status
and prevent functions like uart_write from executing before the configuration has been done
with uart_configure.

• The reference clock is hardcoded as 24 MHz now, the driver should instead query the hardware

to find the reference clock’s frequency.

Daniels Umanovskis

61

7 Interrupts

There’s no reasonable way of handling systems programming, such as embedded development or
operating system development, without interrupts being a major consideration.

What’s an interrupt, anyway? It’s a kind of notification signal that a CPU receives as an indication that
something important needs to be handled. An interrupt is often sent by another hardware device, in
which case that’s a hardware interrupt. The CPU responds to an interrupt by interrupting its current
activity (hence the name), and switching to a special function called an interrupt handler or an interrupt
service routine - ISR for short. After dealing with the interrupt, the CPU will resume whatever it was
doing previously.

There are also software interrupts, which can be triggered by the CPU itself upon detecting an error, or
may be possible for the programmer to trigger with code.

Interrupts are used primarily for performance reasons. A polling-based approach, where external
devices are continuously asked for some kind of status, are inefficient. The UART driver we wrote in
the previous chapter is a prime example of that. We use it to let the user send data to our program by
typing, and typing is far slower than the frequency at which the CPU can check for new data. Interrupts
solve this problem by instead letting the device notify the CPU of an event, such as the UART receiving
a new data byte.

If you read the manual for the PL011 UART in the previous chapter, you probably remember seeing
some registers that control interrupt settings, which we ignored at the time. So, changing the driver to
work with interrupts should be just a matter of setting some registers to enable interrupts and writing
an ISR to handle them, right?

No, not even close. Inerrupt handling is often quite complicated, and there’s work to be done before
any interrupts can be used at all, and then there are additional considerations for any ISRs. Let’s get to
it.

Interrupt handling in ARMv7-A

Interrupt handling is very hardware-dependent. We need to look into the general interrupt handling
procedure of the particular architecture, and then into specifics of a particular implementation like a

62

Bare-metal programming for ARM

ac090d8bfc

specific CPU. The ARMv7-A manual provides quite a lot of useful information about interrupt handling
on that architecture.

ARMv7-A uses the generic term exception to refer, in general terms, to interrupts and some other
exception types like CPU errors. An interrupt is called an IRQ exception in ARMv7-A, so that’s the term
the manual names a lot.When an ARMv7-A CPU takes an exception, it transfers control to an instruction
located at the appropriate location in the vector table, depending on the exception type. The very first
code we wrote for startup began with the vector table.

As a reminder:

1 _Reset:
2
3
4
5
6
7
8

b Reset_Handler
b . /* 0x4 Undefined Instruction */
b . /* 0x8 Software Interrupt */
b . /* 0xC Prefetch Abort */
b . /* 0x10 Data Abort */
b . /* 0x14 Reserved */
b . /* 0x18 IRQ */

In the code above, we had an instruction at offset 0x0, for the reset exception, and dead loops for
the other exception types, including IRQs at 0x18. So normally, an ARMv7-A CPU will execute the
instruction at 0x18 starting from the vector table’s beginning when it takes an IRQ exception.

There’s more that happens, too. When an IRQ exception is taken, the CPU will switch its mode to the
IRQ mode, affecting how some registers are seen by the CPU. When we were initially preparing the
stack for the C environment, we set several stacks up for different CPU modes, IRQ mode being one of
them.

At this point it’s worth noting that IRQ (and FIQ) exceptions can be disabled or enabled globally. The
CPSR register, which you might recall we used to explicitly switch to different modes in Chapter 4, also
holds the I and F bits that control whether IRQs and FIQs are enabled respectively.

Ignoring some advanced ARMv7 features like monitor and hypervisor modes, the sequence upon taking
an IRQ exception is the following:

1. Figure out the address of the next instruction to be executed after handling the interrupt, and

write it into the LR register.

2. Save the CPSR register, which contains the current processor status, into the SPSR register.

3. Switch to IRQ mode, by changing the mode bits in CPSR to 0x12.

4. Make some additional changes in CPSR, such as clearing conditional execution flags.

Daniels Umanovskis

63

Bare-metal programming for ARM

ac090d8bfc

5. Check the VE (Interrupt Vectors Enabled) bit in the SCTLR (system control register). If VE is 0, go
to the start of the vector table plus 0x18. If VE is 1, go to the appropriate implementation-defined
location for the interrupt vector.

That last part sounds confusing. What’s with that implementation-defined location?

Remember that ARMv7-A is not a CPU. It’s a CPU architecture. In this architecture, interrupts are
supported as just discussed, and there’s always the possibility to use an interrupt handler at 0x18
bytes into the vector table. That is, however, not always convenient. Consider that there can be many
different interrupt sources, while the vector table can only contain one branch instruction at 0x18.
This means that the the function taking care of interrupts would first have to figure out which interrupt
was triggered, and then act appropriately. Such an approach puts extra burden on the CPU as it has to
check all possible interrupt sources.

The solution to that is known as vectored interrupts. In a vectored interrupt system, each interrupt has
its own vector (a unique ID). Then some kind of vectored interrupt controller is in place that knows
which ISR to route each interrupt vector to.

The ARMv7-A architecture has numerous implementations, as in specific CPUs. The architecture
description says that vectored interrupts may be supported, but the details are left up to the
implementation. The choice of which interrupt system to use, though, is controlled by the architecture-
defined SCTLR register. In our case, implementation-defined will mean that vectored interrupts are
not supported - the CPU we’re using doesn’t allow vectored interrupts.

Generic Interrupt Controller of the Cortex-A9

We’re programming for a CoreTile Express A9x4 daughterboard, which contains the Cortex-A9 MPCore
CPU. The MPCore means it’s a CPU that can consist of one to four individual Cortex-A9 cores. So it’s
the Cortex-A9 MPCore manual that becomes our next stop. There’s a chapter in the manual for the
interrupt controller - so far so good - but it immediately refers to another manual. Turns out that the
Cortex-A9 has an interrupt controller of the ARM Generic Interrupt Controller type, for which there’s a
separate manual (note that GIC version 4.0 makes a lot of references to the ARMv8 architecture). The
Cortex-A9 manual refers to version 1.0 of the GIC specification, but reading version 2.0 is also fine, there
aren’t too many differences and none in the basic features.

The GIC is one of the major interrupt controller implementations. This is one of the area where the
difference between A-profile and R-profile of ARMv7 matters. ARMv7-R CPUs such as the Cortex-R4
normally use a vectored controller like the appropriately named VIC.

The GIC has its own set of SFRs that control its operation, and the GIC as a whole is responsible for
forwarding interrupt requests to the correct A9 core in the A9-MPCore. There are two main components

Daniels Umanovskis

64

Bare-metal programming for ARM

ac090d8bfc

in the GIC - the Distributor and the CPU interfaces. The Distributor receives interrupt requests, prioritizes
them, and forwards them to the CPU interfaces, each of which corresponds to an A9 core.

Let’s clarify with a schematic drawing. The Distributor and the CPU interfaces are all part of the GIC, with
each CPU then using its own assigned CPU interface to communicate with the GIC. The communication
is two-way because CPUs need to not only receive interrupts but also, at least, to inform the GIC when
interrupt handling completes.

| |

ARM GIC

IRQ source | | Distrib- | | CPU

IRQ source +------------------------+
|
| +-------+ |

1
2
3 +-------------> +----------+
4
5
6 +-------------> | utor
| |
7
IRQ source | |
8
9 +-------------> |
10
| |
11
IRQ source | |
12 +-------------> |
13
| 1
14
IRQ source |
|
15 +------------->
+-------+ |
|
16
+------------------------+
17

| | I-face| |
| |
| | 0
<-----+
| |
| +-------+ |
|
|
| +-------+ |
| |
| | CPU

+-----------+
| +----------+ | I-face+-----> Cortex A-9|
|
| |
|
|
<-----+ CPU 1
|
|
+-----------+

+-----------+
+-----> Cortex A-9|
|
|
|
+-----------+

|
| CPU 0

To enable interrupts, we’ll need to program the GIC Distributor, telling it to enable certain interrupts,
and forward them to our CPU. Once we have some form of working interrupt handling, we’ll need
to tell our program to report back to the GIC, using the CPU Interface part, when the handling of an
interrupt has been finished.

The general sequence for an interrupt is as follows:

1. The GIC receives an interrupt request. That particular interrupt is now considered pending.

2. If the specific interrupt is enabled in the GIC, the Distributor determines the core or cores to

forward it to.

3. Among all pending interrupts, the Distributor chooses the one with the highest priority for each

CPU interface.

4. The GIC’s CPU interface forwards the interrupt to the processor, if priority rules tell it to do so.

5. The processor acknowledges the interrupt, informing the GIC. The interrupt is now active or,

possibly, active and pending if the interrupt has been requested again.

Daniels Umanovskis

65

Bare-metal programming for ARM

ac090d8bfc

6. The software running on the processor handles the interrupt and then informs the GIC that the

handling is complete. The interrupt is now inactive.

Note that interrupts can also be preempted, that is, a higher-priority interrupt can be forwarded to a
CPU while it’s already processing an active lower-priority interrupt.

Just as with the UART driver previously, it’s wise to identify some key registers of the GIC that we will
need to program to process interrupts. I’ll once again omit the GIC prefix in register names for brevity.
Registers whose names start with D (or GICD in full) belong to the Distributor system, those with C
names belong to the CPU interface system.

For the Distributor, key registers include:

• DCTLR - the global Distributor Control Register, containing the enable bit - no interrupts will be

forwarded to CPUs without turning that bit on.

• DISENABLERn - interrupt set-enable registers. There are multiple such registers, hence the n at

the end. Writing to these registers enables specific interrupts.

• DICENABLERn - interrupt clear-enable registers. Like the above, but writing to these registers

disables interrupts.

• DIPRIORITYRn - interrupt priorty registers. Lets each interrupt have a different priority level, with
these priorities determining which interrupt actually gets forwarded to a CPU when there are
multiple pending interrupts.

• DITARGETSRn - interrupt processor target registers. These determine which CPU will get notified

for each interrupt.

• DICFGRn - interrupt configuration registers. They identify whether each interrupt is edge-
triggered or level-sensitive. Edge-triggered interrupts can be deasserted (marked as no longer
pending) by the peripheral that triggered them in the first place, level-sensitive interrupts can
only be cleared by the CPU.

There are more Distributor registers but the ones above would let us get some interrupt handling in
place. That’s just the Distributor part of the GIC though, there’s also the CPU interface part, with key
registers including:

• CCTLR - CPU interface control register, enabling or disabling interrupt forwarding to the particular

CPU connected to that interface.

• CCPMR - interrupt priority mask register. Acts as a filter of sorts between the Distributor and the
CPUs - this register defines the minimum priority level for an intrrupt to be forwarded to the CPU.

• CIAR - interrupt acknowledge register. The CPU receiving the interrupt is expected to read from

this register in order to obtain the interrupt ID, and thereby acknowledge the interrupt.

Daniels Umanovskis

66

Bare-metal programming for ARM

ac090d8bfc

• CEOIR - end of interrupt register. The CPU is expected to write to this register after completing

the handling of an interrupt.

First GIC implementation

Let us say that the first goal is to successfully react to an interrupt. For that, we will need a basic GIC
driver and an interrupt handler, as well as some specific interrupt to enable and react to. The UART can
act as an interrupt source, as a UART data reception (keypress in the terminal) triggers an interrupt.
From there, we’ll be able to iterate and improve the implementation with better interrupt hanlders
and the use of vectorized interrupts.

This section has quite a lot of information and again refers to multiple manuals, so do not worry if it
initially seems complicated!

We begin by defining the appropriate structures in a header file that could be called gic.h, taking the
register map from the GIC manual as the source of information. The result looks something like this:

1 typedef volatile struct __attribute__((packed)) {
2

uint32_t DCTLR;

/* 0x0 Distributor Control register

3
4

5

6

7

8

9

10

11

12

13

14
15

*/

const uint32_t DTYPER;
const uint32_t DIIDR;

register */

/* 0x4 Controller type register */
/* 0x8 Implementer identification

uint32_t _reserved0[29];

/* 0xC - 0x80; reserved and

implementation-defined */

uint32_t DIGROUPR[32];

/* 0x80 - 0xFC Interrupt group

registers */

uint32_t DISENABLER[32];
enable registers */
uint32_t DICENABLER[32];
enable registers */
uint32_t DISPENDR[32];

pending registers */

uint32_t DICPENDR[32];

pending registers */

uint32_t DICDABR[32];

Registers (GIC v1) */
uint32_t _reserved1[32];

*/

/* 0x100 - 0x17C Interrupt set-

/* 0x180 - 0x1FC Interrupt clear-

/* 0x200 - 0x27C Interrupt set-

/* 0x280 - 0x2FC Interrupt clear-

/* 0x300 - 0x3FC Active Bit

/* 0x380 - 0x3FC reserved on GIC v1

uint32_t DIPRIORITY[255];

/* 0x400 - 0x7F8 Interrupt priority

registers */
uint32_t _reserved2;
const uint32_t DITARGETSRO[8];

targets, RO */

/* 0x7FC reserved */
/* 0x800 - 0x81C Interrupt CPU

Daniels Umanovskis

67

25

26
27

28

29

30

31

32

33

34

Bare-metal programming for ARM

ac090d8bfc

16

17
18

19

uint32_t DITARGETSR[246];

/* 0x820 - 0xBF8 Interrupt CPU

targets */

uint32_t _reserved3;
uint32_t DICFGR[64];
registers */

/* 0xBFC reserved */
/* 0xC00 - 0xCFC Interrupt config

/* Some PPI, SPI status registers and identification registers

beyond this.
Don't care about them */

20
21 } gic_distributor_registers;
22
23 typedef volatile struct __attribute__((packed)) {
24

uint32_t CCTLR;
register */
uint32_t CCPMR;
register */
uint32_t CBPR;
const uint32_t CIAR;

register */
uint32_t CEOIR;

*/

const uint32_t CRPR;

*/

const uint32_t CHPPIR;

interrupt register */

uint32_t CABPR;
register */

const uint32_t CAIAR;

acknowledge register */

uint32_t CAEOIR;
register */

/* 0x0 CPU Interface control

/* 0x4 Interrupt priority mask

/* 0x8 Binary point register */
/* 0xC Interrupt acknowledge

/* 0x10 End of interrupt register

/* 0x14 Running priority register

/* 0x18 Higher priority pending

/* 0x1C Aliased binary point

/* 0x20 Aliased interrupt

/* 0x24 Aliased end of interrupt

const uint32_t CAHPPIR;

/* 0x28 Aliased highest priority

pending interrupt register */

35 } gic_cpu_interface_registers;

There is nothing particularly noteworthy about the structs, they follow the same patterns as explained
in the previous chapter. Note that Distributor and CPU Interface stuctures cannot be joined together
because they may not be contiguous in memory (and indeed aren’t on the Cortex-A CPUs).

When that’s done, we need to write gic.c, our implementation file. The first version can be really
simple, but it will nonetheless reveal several things that we had not had to consider before. JHere’s
how gic.c begins:

1 #include "gic.h"
2
3 static gic_distributor_registers* gic_dregs;
4 static gic_cpu_interface_registers* gic_ifregs;
5
6 void gic_init(void) {

Daniels Umanovskis

68

Bare-metal programming for ARM

ac090d8bfc

7
8
9
10

11

12
13

14

gic_ifregs = (gic_cpu_interface_registers*)GIC_IFACE_BASE;
gic_dregs = (gic_distributor_registers*)GIC_DIST_BASE;

WRITE32(gic_ifregs->CCPMR, 0xFFFFu); /* Enable all interrupt

priorities */

WRITE32(gic_ifregs->CCTLR, CCTRL_ENABLE); /* Enable interrupt

forwarding to this CPU */

gic_distributor_registers* gic_dregs = (gic_distributor_registers*)

GIC_DIST_BASE;

WRITE32(gic_dregs->DCTLR, DCTRL_ENABLE); /* Enable the interrupt

distributor */

15 }

We define static variables to hold pointers to the Distributor and the CPU Interface, and write an
initialization function. Here you might already notice one difference from the UART driver earlier. The
UART driver had its pointer initialized to the hardware address the hardware uses, like this:

1 static uart_registers* uart0 = (uart_registers*)0x10009000u;

With GIC registers, we cannot do the same because their address is implementation-dependent.
Hardcoding the address for a particular board is possible (and it is what QEMU itself does) but we
can implement the more correct way, setting those register addresses in gic_init. The Cortex-A9
MPCore manual states that the GIC is within the CPU’s private memory region, specifically the CPU
interface is at 0x0100 from PERIPHBASE and the Distributor is at 0x1000 from PERIPHBASE. What’s
this PERIPHBASE then? The A9 MPCore manual also states that:

Figure 7.1: Description of PERIPHBASE

It should be clear that the GIC Distributor is located at PERIPHBASE + 0x1000 but obtaining
PERIPHBASE seems confusing. Let’s take a look at the GIC_DIST_BASE and GIC_IFACE_BASE
macros that gic_init uses.

1 #define GIC_DIST_BASE

((cpu_get_periphbase() + GIC_DISTRIBUTOR_OFFSET

))

2 #define GIC_IFACE_BASE ((cpu_get_periphbase() + GIC_IFACE_OFFSET))

Daniels Umanovskis

69

Bare-metal programming for ARM

ac090d8bfc

I put the offsets themselves into a different CPU-specific header file cpu_a9.h, but it can of course be
organized however you want. The cpu_get_periphbase function is implemented like this:

1 inline uint32_t cpu_get_periphbase(void) {
2
3
4
5 }

uint32_t result;
asm ("mrc p15, #4, %0, c15, c0, #0" : "=r" (result));
return result;

Just what is going on there? It’s a function with a weirdly-formatted assembly line, and the assembly
itself refers to strange things like p15. Let’s break this down.

C functions can use what is known as inline assembly in order to include assembly code directly. Inline
assembly is generally used either for manual optimization of critical code, or to perform operations
that are not exposed to ordinary code. We have the latter case. When writing inline assembly for GCC,
you can use the extended assembly syntax, letting you read or write C variables. When you see a colon
: in an inline assembly block, that’s extended assembly syntax, which is documented by GCC and
in the simplest case looks like asm("asm-code-here": "output"), where the output refers to C
variables that will be modified.

The %0 part in our extended assembly block is just a placeholder, and will be replaced by the first (and,
in this case, the only) output operand, which is "=r"(result). That output syntax in turn means that
we want to use some register (=r) and that it should write to the result variable. The choice of the
specific register is left to GCC. If we were writing in pure assembly, the instruction would be, assuming
the R0 register gets used for output

1 mrc p15, #4, r0, c15, c0, #0

That’s still one strange-looking instruction. ARM processors (not just ARMv7 but also older architectures)
support coprocessors, which may include additional functionality outside the core processor chip
itself. Coprocessor 15, or CP15 for short, is dedicated to important control and configuration functions.
Coprocessors are accessed through the special instructions mrc (read) and mcr (write). Those
instructions contain additional opcodes, the meaning of which depends on the coprocessor.

The A9 MPCore manual makes a reference to the “CP15 c15 Configuration Base Address Register” when
describing PERIPHBASE. CP15 is, as we now know, coprocessor 15, but the c15 part refers, confusingly,
to something else, namely to a specific register in CP15. The mrc instruction has a generic format,
which is:

Daniels Umanovskis

70

Bare-metal programming for ARM

ac090d8bfc

1 mrc coproc, op1, Rd, CRn, CRm [,op2]

So the coprocessor number comes first, Rd refers to the ARM register to read data to, while op1 and
optionally op2 are operation codes defined by the coprocessor, and CRn and CRm are coprocessor
registers. This means that, in order to do something with a coprocessor, we need to look up its own
documentation. The coprocessor’s features fall under the corresponding processor features, and we
can find what interests us in the Cortex A9 manual. Chapter 4, System Control concerns CP15, and a
register summary lists the various operations and registers that are available. Under c15, we find the
following:

Figure 7.2: CP15 c15 register summary

Looking through the table, we can finally find out that reading the Configuration Base Register, which
contains the PERIPHBASE value, requires accessing CP15 with Rn=c15, op1 = 4, CRm = c0, and
op2 = 0. Putting it all together gives the mrc instruction that we use in cpu_get_periphbase.

The remainder of gic_init is quite unremarkable. We enable forwarding of interrupts with all
priorities to the current CPU, and enable the GIC Distributor so that interrupts from external sources
could reach the CPU. Note the use of the WRITE32 macro. Register access width was mentioned in the
previous chapter, and unlike the PL011 UART, the GIC explicitly states that all registers permit 32-bit
word access, with only a few Distributor registers allowing byte access. So we should take care to write
the registers with one 32-bit write with this macro.

Daniels Umanovskis

71

Bare-metal programming for ARM

ac090d8bfc

1 #define WRITE32(_reg, _val) (*(volatile uint32_t*)&_reg = _val)

The next order of business is to let specific interrupts be enabled. Initializing the GIC means we can
now receive interrupts in general. As said before, upon receiving an interrupt, the GIC Distributor
checks if the particular interrupt is enabled before forwarding it to the CPU interface. The Set-Enable
registers, GICD_ISENABLER[n], control whether a particular interrupt is enabled. Each ISENABLER
register can enable up to 32 interrupts, and having many such registers allows the hardware to have
more than 32 different interrupt sources. Given an interrupt with id N, enabling it means setting the bit
N % 32 in register N / 32, where integer division is used. For example, interrupt 45 would be bit 13
(45 % 32 = 13) in ISENABLER[1] (45 / 32 = 1).

For each interrupt, you also need to select which CPU interface(s) to forward the interrupt to, done in
the GICD_ITARGETSR[n] registers. The calculation for these registers is slightly different, for interrupt
with id N the register is N / 4, and the target list has to be written to byte N % 4 in that register. The
target list is just a byte where bit 0 represents CPU interface 0, bit 1 represents CPU interface 1 and so on.
We don’t need anything fancy here, we just want to forward any enabled interrupts to CPU Interface
0.

With that knowledge, writing the following function becomes quite simple:

uint32_t reg_val = gic_dregs->DISENABLER[reg];
reg_val |= (1u << bit);
WRITE32(gic_dregs->DISENABLER[reg], reg_val);

/* Enable the interrupt */
uint8_t reg = number / 32;
uint8_t bit = number % 32;

1 void gic_enable_interrupt(uint8_t number) {
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16 }

/* Forward interrupt to CPU Interface 0 */
reg = number / 4;
bit = (number % 4) * 8; /* Can result in bit 0, 8, 16 or 24 */
reg_val = gic_dregs->DITARGETSR[reg];
reg_val |= (1u << bit);
WRITE32(gic_dregs->DITARGETSR[reg], reg_val);

Now we have gic_init to initialize the GIC and gic_enable_interrupt to enable a specific
interrupt. The preparation is almost done, we just need functions to globally disable and enable
interrupts. When using an interrupt controller, it’s a good idea to disable interrupts on startup, and
then enable them after the interrupt controller is ready.

Disabling interrupts is easy, we can do it somewhere in the assembly startup code in startup.s. At
some point when the CPU is in supervisor mode, add the cpsid if instruction to disable all interrupts

Daniels Umanovskis

72

Bare-metal programming for ARM

ac090d8bfc

- the if part means both IRQs and FIQs. One possible place to do that would be right before the
bl main instruction that jumps to C code.

Enabling interrupts is done similarly, with the cpsie if instruction. We’ll want to call this from C code
eventually so it’s convenient to create a C function with inline assembly in some header file, like this:

1 inline void cpu_enable_interrupts(void) {
2
3 }

asm ("cpsie if");

Looks like we’re done! Just to make sure the new functions are getting used, call gic_init() and then
cpu_enable_interrupts() from somewhere in the main function (after the initial UART outputs
perhaps). At this point you can try building the program (remember to add gic.c to the source file
list in CMakeLists.txt), but surprisingly enough, the program won’t compile, and you’ll get an error
like

1 /tmp/ccluurNJ.s:146: Error: selected processor does not support

cpsie

if' in ARM mode

This is our first practical encounter with the fact that ARMv7 (same goes for some other ARM
architectures) has two instruction sets, ARM and Thumb (Thumb version 2 to be exact). Thumb
instructions are smaller at 16 bits compared to the 32 bits of an ARM instruction, and so can be
more efficient, at the cost of losing some flexibility. ARM CPUs can freely switch between the two
instruction sets, but Thumb should be the primary choice. In the above error message, GCC is telling
us that cpsie if is not available as an ARM instruction. It is indeed not, it’s a Thumb instruction. We
need to change the compiler options and add -mthumb to request generation of Thumb code. In
CMakeLists.txt, that means editing the line that sets CMAKE_C_FLAGS. After adding -mthumb to it
we can try to recompile. The interrupt-enabling instruction no longer causes any problems but another
issue crops up:

1 /tmp/ccC72j7I.s:37: Error: selected processor does not support

mrc p15

,#4,r2,c15,c0,#0' in Thumb mode

Indeed, accessing the coprocessors is only possible with ARM instructions. The mrc instruction does
not exist in the Thumb instruction set. It’s possible to control the CPU’s instruction set and freely switch
between the two, but fortunately, GCC can figure things out by itself if we tell it what specific CPU we’re
using. So far we’ve just been compiling for a generic ARM CPU, but we can easily specify the CPU by
also adding -mcpu=cortex-a9 to the compilation flags. So now with -mthumb -mcpu=cortex-a9
added to the compile flags, we can finally compile and run the application just as before.

Daniels Umanovskis

73

Bare-metal programming for ARM

ac090d8bfc

You should see that the program works just like it did at the end of the previous chapter. Indeed, we’ve
enabled the GIC, and have interrupts enabled globally for the CPU, but we haven’t enabled any specific
interrupts yet, so the GIC will never forward any interrupts that may get triggered.

NOTE

With the GIC enabled, you can view its registers with a debugger or in the QEMU monitor, with some
caveats. If you’re using QEMU older than version 3.0, then the Distributor’s control register will show
the value 0 when read that way, even if the Distributor is actually enabled. And if you try to access the
CPU Interface registers (starting at 0x1e000100) with an external debugger like GDB, QEMU will crash,
at least up to and including version 3.1.0

Handling an interrupt

Let’s now put the GIC to use and enable the UART interrupt, which should be triggered any time the
UART receives data, which corresponds to us pressing a key in the terminal when running with QEMU.
After receiving an interrupt, we’ll need to properly handle it to continue program execution.

Enabling the UART interrupt should be easy since we already wrote the gic_enable_interrupt
function, all we need to do now is to call it with the correct interrupt number. That means once again
going back to the manuals to find the interrupt ID numbe we need to use. Interrupt numbers usually
differ depending on the board, and in our case the CoreTile Express A9x4 manual can be the first stop.
The section 2.6 Interrupts explains that the integrated test chip for the Cortex-A9 MPCore on this board
is directly connected to the motherboard (where the UART is located as we remember from the previous
chapter), and that motherboard interrupts 0-42 map to interrupts 32-74 on the daughterboard. This
means we need to check the motherboard manual and add 32 to the interrupt number we find there.

The Motherboard Express µATX manual explains in 2.6 Interrupt signals that the motherboard has no
interrupt controller, but connects interrupt signals to the daughterboard. The signal list says that
UART0INTR, the interrupt signal for UART0, is number 5. Since the daughterboard remaps interrupts,
we’ll need to enable interrupt 37 in order to receive UART interrupts in our program. The following
snippet in main should do just fine:

Daniels Umanovskis

74

Bare-metal programming for ARM

ac090d8bfc

1 gic_init();
2 gic_enable_interrupt(37);
3 cpu_enable_interrupts();

And we need an interrupt handler, which we need to point out in the vector table in startup.s. It
should now look something like

1 _Reset:
2
3
4
5
6
7
8
9

b Reset_Handler
b Abort_Exception /* 0x4
b . /* 0x8 Software Interrupt */
b Abort_Exception /* 0xC
b Abort_Exception /* 0x10 Data Abort */
b . /* 0x14 Reserved */
b IrqHandler /* 0x18 IRQ */
b . /* 0x1C FIQ */

Prefetch Abort */

Undefined Instruction */

The seventh entry in the vector table, at offset 0x18, will now jump to IrqHandler. We can add it to
the end of startup.s, and the simplest implementation that would tell us things are working fine
can just store the data that the UART received in some register and hang.

1 IrqHandler:
2
3
4

ldr r0, =0x10009000
ldr r1, [r0]
b .

Reading from the UART register at 0x10009000 gives the data byte that was received, and we proceed
to store it in R1 before hanging. Why hang? Continuing execution isn’t as simple as just returning from
the IRQ handler, you have to take care to save the program state before the IRQ, then restore it, which
we’re not doing. Our handler, the way it’s written above, breaks the program state completely.

Let’s compile and test now! Once the program has started in QEMU and written its greetings to the
UART, press a key in the terminal to trigger the now-enabled UART interrupt. Then check the registers
with info registers in QEMU monitors, and unfortunately you’ll notice a problem. The IRQ handler
doesn’t seem to be running and the program is just hanging. Output could be something similar to:

1 (qemu) info registers
2 R00=00000005 R01=00000000 R02=00000008 R03=00000090
3 R04=00000000 R05=7ffd864c R06=60000000 R07=00000000
4 R08=00000400 R09=7fef5ef8 R10=00000001 R11=00000001
5 R12=00000000 R13=00000013 R14=7ff96264 R15=7ff96240
6 PSR=00000192 ---- A S irq32

Daniels Umanovskis

75

Bare-metal programming for ARM

ac090d8bfc

Good news first, the program status register PSR indicates that the CPU is running in IRQ mode (0x192
& 0x1F is 0x12, which is IRQ mode, but QEMU helpfully points it out by writing irq32 on the same
line). The bad news is that R0 and R1 don’t contain the values we would expect from IrqHandler,
and the CPU seems to be currently running code at some strange address in R15 (remember that R15 is
just another name for PC, the program counter register). The address doesn’t correspond to anything
we’ve loaded into memory so the conclusion is that the CPU did receive an interrupt, but failed to run
IrqHandler.

This is one more detail that happens due to QEMU emulation not being perfect. If you remember
the discussion about memory and section layout from Chapter 4, we’re pretending that our ROM
starts at 0x60000000. The Cortex-A9 CPU, however, expects the vector table to be located at address
0x0, according to the architecture, and IRQ handling starts by executing the instruction at 0x18 from
the vector table base. Unfortunately, our vector table is actually at 0x60000000 and address 0x0 is
reserved by QEMU for the program flash memory, which we cannot emulate.

We then need to make a QEMU-specific modification to our code and indicate that the vector table
base is at 0x60000000. This is a very low-level modification of the CPU configuration, so you might be
able to guess that the system control coprocessor, CP15, is involved again. We previously used its c15
register to read PERIPHBASE, and the ARMv7-A manual will reveal that the c12 register contains the
vector table base address, which may also be modified. To write to the coprocessor, we use the mcr
instruction (as opposed to mrc for reading), and the instructions we need will be:

1 ldr r0, =0x60000000
2 mcr p15, #0, r0, c12, c0, #0

Those two instructions should be somewhere early in the startup code, such as right after the
Reset_Handler label. Having done that modification, we can perform another rebuild and test run.
Press a key in the terminal, and check the registers in the QEMU monitor. Now you should see that R0
contains the UART address, and R1 contains the code code of the key you pressed, such as 0x66 for f
or 0x61 for a.

1 R00=10009000 R01=00000066 R02=00000008 R03=00000090

With that, we have correctly jumped into an interrupt handler after an external interrupt triggers, which
is a major step towards improving our bare-metal system.

Daniels Umanovskis

76

Bare-metal programming for ARM

ac090d8bfc

Surviving the IRQ handler

Our basic implementation of the IRQ handler isn’t good for much, the biggest issue being that the
program hangs completely and never leaves the IRQ mode.

Interrupt handlers, as the name suggests, interrupt whatever the program was doing previously. This
means that state needs to be saved before the handler, and restored after. The general-purpose ARM
registers, for example, are shared between modes, so if your register R0 contains something, and then
an interrupt handler writes to it, the old value is lost. This is part of the reason why a separate IRQ stack
is needed (which we prepare in the startup code), as the IRQ stack is normally where the context would
be saved.

When writing interrupt handlers in assembly, we have to take care of context saving and restoring, and
correctly returning from the handler. Hand-written assembly interrupt handlers should be reserved for
cases where fine-tuned assembly is critical, but generally it’s much easier to write interrupt handlers
in C, where they become regular functions for the most part. The compiler can handle context save
and restore, and everything else that’s needed for interrupt handling, if told that a particular function
is an interrupt handler. In GCC, __attribute__((interrupt)) is a decoration that can be used to
indicate that a function is an interrupt handler.

We can write a new function in the UART driver that would respond to the interrupt.

1 void __attribute__((interrupt)) uart_isr(void) {
2
3 }

uart_write("Interrupt!\n");

Then just changing b IrqHandler to b uart_isr in the vector table will ensure that the uart_isr
function is the one called when interrupts occur. If you test this, you’ll see that the program just
keeps spamming Interrupt! endlessly after a keypress. Our ISR needs to communicate with the GIC,
acknowledge the interrupt and signal the GIC when the ISR is done. In the GIC, we need a function that
acknowledges an interrupt.

1 uint32_t gic_acknowledge_interrupt(void) {
2
3 }

return gic_ifregs->CIAR & CIAR_ID_MASK;

CIAR_ID_MASK is 0x3FF because the lowest 9 bits of CIAR contain the interrupt ID of the interrupt
that the GIC is signaling. After a read from CIAR, the interrupt is said to change from pending state to
active. Another function is necessary to signal the end of the interrupt, which is done by writing the
interrupt ID to the EOIR register.

Daniels Umanovskis

77

Bare-metal programming for ARM

ac090d8bfc

1 void gic_end_interrupt(uint16_t number) {
2
3 }

WRITE32(gic_ifregs->CEOIR, (number & CEOIR_ID_MASK));

The ISR could then use those two functions and do something along the lines of:

uint16_t irq = gic_acknowledge_interrupt();
if (irq == 37) {

1 void __attribute__((interrupt)) uart_isr(void) {
2
3
4
5
6
7 }

}
gic_end_interrupt(37);

uart_write("Interrupt!\n");

This implementation is better but would still result in endless outputs. The end-of-interrupt would
be correctly signaled to the GIC, but the GIC would forward a new UART interrupt to the CPU. This is
because the interrupt is generated by the UART peripheral, the GIC just forwards it. The code above lets
the GIC know we’re done handling the interrupt, but doesn’t inform the UART peripheral of that. The
PL011 UART has an interrupt clear register, ICR, which is already in the header file from the last chapter.
Clearing all interrupts can be done by writing 1 to bits 0-10, meaning the mask is 0x7FF. If we clear all
interrupt sources in the UART before signaling end-of-interrupt to the GIC, everything will work.

uint16_t irq = gic_acknowledge_interrupt();
if (irq == 37) {

1 void __attribute__((interrupt)) uart_isr(void) {
2
3
4
5
6
7
8 }

}
uart0->ICR = ICR_ALL_MASK;
gic_end_interrupt(37);

uart_write("Interrupt!\n");

With that interrupt handler, our program will write Interrupt! every time you press a key in the
terminal, after which it will resume normal execution. You can verify for yourself that the CPU returns
to the supervisor (SVC) mode after handling the interrupt. It can also be interesting to disassemble the
ELF file and note how the code for uart_isr differs from any other functions - GCC will have generated
stmdb and ldmia instructions to save several registers to the stack and restore them later.

Daniels Umanovskis

78

Bare-metal programming for ARM

ac090d8bfc

Adapting the UART driver

We now finally have working interrupt handling with a properly functional ISR that handles an interrupt,
clears the interrupt source and passes control back to whatever code was running before the interrupt.
Next let us apply interrupts in a useful manner, by adapting the UART driver and making the interrupts
do something useful.

The first thing to note is that what we’ve been calling “the UART interrupt” is a specific interrupt signal,
UART0INT that the motherboard forwards to the GIC. From the point of view of the PL011 UART itself
though, several different interrupts exist. The PL011 manual has a section devoted to interrupts, which
lists eleven different interrupts that the peripheral can generate, and it also generates an interrupt
UARTINTR that is an OR of the individual interrupts (that is, UARTINTR is active if any of the others is).
It’s this UARTINTR that corresponds to the interrupt number 37 which we enabled, but our driver code
should check which interrupt occurred specifically and react accordingly.

The UARTMIS register can be used to read the masked interrupt status, with the UARTRIS providing
the raw interrupt status. The difference between those is that, if an interrupt is masked (disabled) in
the UART’s configuration, it can still show as active in the raw register but not the masked one. By
default all interrupts all unmasked (enabled) on the PL011 so this distinction doesn’t matter for us. Of
the individual UART interrupts, only the receive interrupt is really interesting in the basic use case, so
let’s implement that one properly, as well as one of the error interrupts.

All interrupt-related PL011 registers use the same pattern, where bits 0-10 correspond to the same
interrupts. The receive (RX) interrupt is bit 4, the break error (BE) interrupt is bit 9. We can express that
nicely with a couple of defines:

1 #define RX_INTERRUPT
2 #define BE_INTERRUPT

(1u << 4u)
(1u << 9u)

We’re using the UART as a terminal, so when the receive interrupt occurs, we’d like to print the character
that was received. If the break error occurs, we can’t do much except clear the error flag (in the RSRECR
register) and write an error message. Let’s write a new ISR that checks for the actual underlying UART
interrupt and reacts accordingly.

(void)gic_acknowledge_interrupt();

1 void __attribute__((interrupt)) uart_isr(void) {
2
3
4
5
6

uint32_t status = uart0->MIS;
if (status & RX_INTERRUPT) {

/* Read the received character and print it back*/

Daniels Umanovskis

79

Bare-metal programming for ARM

ac090d8bfc

7
8
9
10
11
12
13
14
15
16
17
18
19
20
21 }

char c = uart0->DR & DR_DATA_MASK;
uart_putchar(c);
if (c == '\r') {

uart_write("\n");

}

} else if (status & BE_INTERRUPT) {

uart_write("Break error detected!\n");
/* Clear the error flag */
uart0->RSRECR = ECR_BE;
/* Clear the interrupt */
uart0->ICR = BE_INTERRUPT;

}

gic_end_interrupt(UART0_INTERRUPT);

In the previous chapter, we had a loop in main that polled the UART. That is no longer necessary, but
remember that main should not terminate so the while (1) loop should still be there. The terminal
functionality is now available and interrupt-driven!

Handling different interrupt sources

The interrupt handling solution at this point has a major flaw. No matter what interrupt the CPU
receives, the b uart_isr from the vector table will take us to that interrupt handler, which is of course
only suitable for the UART interrupt. Early on in this chapter, there was mention of vectored interrupts,
which we cannot use since our hardware uses the GIC, a non-vectored interrupt controller. Therefore
we’ll need to use a software solution, writing a top-level interrupt handler that will be responsible for
finding out which interrupt got triggered and then calling the appropriate function.

In the simplest case, we’d then write a function like the following:

1 void __attribute__((interrupt)) irq_handler(void) {
uint16_t irq = gic_acknowledge_interrupt();
2
switch (irq) {
3
4
case UART0_INTERRUPT:
5
6
7
8
9
10
11
12 }

uart_write("Unknown interrupt!\n");
break;

}
gic_end_interrupt(irq);

uart_isr();
break;

default:

Daniels Umanovskis

80

Bare-metal programming for ARM

ac090d8bfc

This top-level irq_handler should then be pointed to by the vector table, and adding support for
new interrupts would just mean adding them to the switch statement. The top-level handler takes
care of the GIC acknowledge/end-of-interrupt calls, so individual handlers like uart_isr no longer
have to do it, nor do they need the __attribute__((interrupt)) anymore because the top-level
handler is where the switch to IRQ mode should happen.

Purely from an embedded code perspective, there’s no problem with such a handler and having a
long list of interrupts in the switch statement. It’s not a great solution in terms of general software
design though. It creates quite tight coupling between the top-level IRQ handler, which should be
considered to be a separate module from the GIC, and the handler would have to know about all
other relevant modules. If we place the above handler into a separate file like irq.c, it would have
to include uart_pl011.h for the header’s declaration of uart_isr. If we then add a timer module,
irq.c would also need to include timer.h and irq_handler would have to be modified to call some
timer ISR, which is not a good, maintainable way to structure the code.

A better solution is to make the IRQ handler use callbacks, and allow individual modules to register
those callbacks. We can then offload some important work to a separate IRQ component, with irq.h:

(1024)
(ISR_COUNT - 1)

1 #ifndef IRQ_H
2 #define IRQ_H
3
4 #include <stdint.h>
5
6 typedef void (*isr_ptr)(void);
7
8 #define ISR_COUNT
9 #define MAX_ISR
10
11
12 typedef enum {
13
14
15
16 } irq_error;
17
18 irq_error irq_register_isr(uint16_t irq_number, isr_ptr callback);
19
20 #endif

IRQ_OK = 0,
IRQ_INVALID_IRQ_ID,
IRQ_ALREADY_REGISTERED

The header defines a function irq_register_isr that other modules would then call to register their
own ISRs. The isr_ptr type is a function pointer to an ISR - typedef void (*isr_ptr)(void);
means that isr_ptr is a pointer to a function that returns void and takes no parameters. If the syntax
is confusing, take a moment to read up on C function pointers online - conceptually function pointers
are not difficult but the syntax tends to feel obscure until you get used to it.

Daniels Umanovskis

81

Bare-metal programming for ARM

ac090d8bfc

The implementation in irq.c is:

isr();

}
gic_end_interrupt(irq);

uint16_t irq = gic_acknowledge_interrupt();
isr_ptr isr = callback(irq);
if (isr != NULL) {

1 #include <stddef.h>
2 #include "irq.h"
3 #include "gic.h"
4
5 static isr_ptr callbacks[1024] = { NULL };
6
7 static isr_ptr callback(uint16_t number);
8
9 void __attribute__((interrupt)) irq_handler(void) {
10
11
12
13
14
15
16 }
17
18 irq_error irq_register_isr(uint16_t irq_number, isr_ptr callback) {
19
20
21
22
23
24
25
26
27 }
28
29 static isr_ptr callback(uint16_t number) {
30
31
32
33
34 }

}
return callbacks[number];

} else if (callbacks[irq_number] != NULL) {

return IRQ_ALREADY_REGISTERED;

}
return IRQ_OK;

callbacks[irq_number] = callback;

if (irq_number > MAX_ISR) {

return IRQ_INVALID_IRQ_ID;

if (number > MAX_ISR) {

return NULL;

} else {

We use an array that can store up 1024 ISRs, which is enough to use all the interrupts the GIC supports
if desired. The top-level irq_handler talks to the GIC and calls whatever ISR has been registered for
the particular interrupt. The UART driver then registers its own ISR in uart_init just before enabling
the UART peripheral, like this:

1 /* Register the interrupt */
2 (void)irq_register_isr(UART0_INTERRUPT, uart_isr);

Such a solution no longer requires the IRQ handler to know which specific ISRs exist beforehand, and

Daniels Umanovskis

82

Bare-metal programming for ARM

ac090d8bfc

is easier to maintain.

Summary

In this chapter, we went over interrupt handling in general, the ARM Generic Interrupt Controller, and
we wrote some interrupt handlers.

Interrupt controllers
Interrupts are often among the trickiest topics in embedded development.
themselves are quite complicated - we used the GIC in pretty much the simplest way possible, but it
can quickly get complicated once you start grouping interrupts, working with their priorities and so
on. Another complication arises from the hard-to-predict nature of interrupts. You don’t know what
regular code will be executed when an interrupt happens. Many interrupts in the real world depend on
timing or external data sources, so debugging with breakpoints affects the behavior of the program.

As a broad generalization, interrupt handling becomes trickier and more important as you develop
on more limited hardware. When dealing with microcontrollers, you often have to understand the
amount of time spent in interrupts, and may also find that the switching between normal and IRQ
modes creates real performance issues. Fast interrupts, FIQs, which we didn’t cover in this chapter
exist in ARMv7 to help alleviate the overhead of regular IRQs.

In a real embedded system that does something useful, interrupts are likely to drive some critical parts
of functionality. For example, most systems need some way of measuring time or triggering some code
in a time-based manner, and that usually happens by having a timer that generates interrupts.

Daniels Umanovskis

83

8 Simple scheduling

Very few embedded applications can be useful without some kind of time-keeping and scheduling.
Being able to program things like “do this every X seconds” or “measure the time between events A
and B” is a key aspect of nearly all practically useful applications.

In this chapter, we’ll look at two related concepts. Timers, which are a hardware feature that allows
software to keep track of time, and scheduling, which is how you program a system to run some code,
or a task, on some kind of time-based schedule - hence the name. Scheduling in particular is a complex
subject, some discussion of which will follow later, but first we’ll need to set up some kind of time
measurement system.

In order to keep track of time in our system, we’re going to use two tiers of “ticks”. First, closer to the
hardware, we’ll have a timer driver for a hardware timer of the Cortex-A9 CPU. This driver will generate
interrupts at regular intervals. We will use those interrupts to keep track of system time, a separate
counter that we’ll use in the rest of the system as “the time”.

Such a split is not necessary in a simple tutorial system, but is good practice due to the system time
thus not being directly connected to a particular hardware clock or driver implementation. This allows
better portability as it becomes possible to switch the underlying timer driver without affecting uses of
system time.

The first task then is to create a timer driver. Since its purpose will be to generate regular interrupts,
note that this work builds directly on the previous chapter, where interrupt handling capability was
added.

Private Timer Driver

A Cortex-A9 MPCore CPU provides a global timer and private timers. There’s one private timer per core.
The global timer is constantly counting up, even with the CPU paused in debug mode. The per-core
private timers count down from some starting value to zero, sending an interrupt when zero is reached.
It’s possible to use either timer for scheduling, but the typical solution is to use the private timer. It’s
somewhat easier to handle due to being 32 bits wide (the global timer is 64 bits) and due to stopping
when the CPU is stopped.

84

Bare-metal programming for ARM

ac090d8bfc

The Cortex-A9 MPCore manual explains the private timer in Chapter 4.1. The timer’s operation is quite
simple. A certain starting load value is loaded into the load register LR. When the timer is started, it
keeps counting down from that load value, and generates an interrupt when the counter reaches 0.
Then, if the auto-reload function is enabled, the counter automatically restarts from the load value.
As is common with other devices, the private timer has its own control register CTRL, which controls
whether auto-reload is enabled, whether interrupt generation is enabled, and whether the whole timer
is enabled.

From the same manual, we can see that the private timer’s registers are at offset 0x600 from
PERIPHBASE, and we already used PERIPHBASE in the previous chapter for GIC registers. Finally, the
manual gives the formula to calculate the timer’s period.

Figure 8.1: Private timer period formula

A prescaler can essentially reduce the incoming clock frequency, but using that is optional. If we simplify
with the assumption that prescaler is 0, we can get Load value = (period * PERIPHCLK)- 1.
The peripheral clock, PERIPHCLK, is the same 24 MHz clock from the motherboard that clocks the
UART.

I will not go through the timer driver in every detail here, as it just applies concepts from the previous
two chapters. As always, you can examine the full source in this chapter’s corresponding source code
folder. We call the driver ptimer, implement it in ptimer.h and ptimer.c, and the initialization
function is as follows:

}
uint32_t load_val = millisecs_to_timer_value(millisecs);
WRITE32(regs->LR, load_val); /* Load the initial timer value */

return PTIMER_INVALID_PERIOD;

regs = (private_timer_registers*)PTIMER_BASE;
if (!validate_config(millisecs)) {

1 ptimer_error ptimer_init(uint16_t millisecs) {
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16 }

return PTIMER_OK;

/* Register the interrupt */
(void)irq_register_isr(PTIMER_INTERRUPT, ptimer_isr);

uint32_t ctrl = CTRL_AUTORELOAD | CTRL_IRQ_ENABLE | CTRL_ENABLE;
WRITE32(regs->CTRL, ctrl); /* Configure and start the timer */

Daniels Umanovskis

85

Bare-metal programming for ARM

ac090d8bfc

The function accepts the desired timer tick period, in milliseconds, calculates the corresponding load
value for LR and then enables interrupt generation, auto-reload and starts the timer.

One bit of interest here is the ISR registration, for which the interrupt number is defined as #define
PTIMER_INTERRUPT (29u). The CPU manual says that the private timer generates interrupt 29
when the counter reaches zero. And in the code we’re actually using 29, unlike with the UART driver,
where we mapped UART0INTR, number 5, to 37 in the code. But this interrupt remapping only applies
to certain interrupts that are generated on the motherboard. The private timer is part of the A9 CPU
itself, so no remapping is needed.

The millisecs_to_timer_value function calculates the value to be written into LR from the desired
timer frequency in milliseconds. Normally it should look like this:

1 static uint32_t millisecs_to_timer_value(uint16_t millisecs) {
2
3
4 }

double period = millisecs * 0.001;
return (period * refclock) - 1;

However, things are quite different for us due to using QEMU. It doesn’t emulate the 24 MHz peripheral
clock, and QEMU in general does not attempt to provide timings that are similar to the hardware being
emulated. For our UART driver, this means that the baud rate settings don’t have any real effect, but
that wasn’t a problem. For the timer though, it means that the period won’t be the same as on real
hardware, so the actual implementation used for this tutorial is:

double period = millisecs * 0.001;
uint32_t value = (period * refclock) - 1;
value *= 3; /* Additional QEMU slowdown factor */

1 static uint32_t millisecs_to_timer_value(uint16_t millisecs) {
2
3
4
5
6
7 }

return value;

With the timer driver in place, the simplest way to test is to make the timer ISR print something out,
and initialize the timer with a one-second period. Here’s the straightforward ISR:

1 void ptimer_isr(void) {
2
3
4 }

uart_write("Ptimer!\n");
WRITE32(regs->ISR, ISR_CLEAR); /* Clear the interrupt */

And somewhere in our main function:

Daniels Umanovskis

86

Bare-metal programming for ARM

ac090d8bfc

1 gic_enable_interrupt(PTIMER_INTERRUPT);
2
3 if (ptimer_init(1000u) != PTIMER_OK) {
4
5 }

uart_write("Failed to initialize CPU timer!\n");

Note the call to gic_enable_interrupt, and recall that each interrupt needs to be enabled in the
GIC, it’s not enough to just register a handler with irq_register_isr. This code should result in the
private timer printing out a message every second or, due to the very approximate calculation in the
emulated version, approximately every second.

Build everything and run (if you’re implementing yourself as you read, remember to add the new C file
to CMakeLists.txt), and you should see regular outputs from the timer.

HINT

You can use gawk, the GNU version of awk to print timestamps to the terminal. Instead of just make
run, type make run | gawk '{ print strftime("[%H:%M:%S]"), $0 }' and you’ll see the
local time before every line of output. This is useful to ascertain that the private timer, when set to
1000 milliseconds, procudes output roughly every second.

System Time

As discussed previously, we want to use some kind of system time system-wide. This is going to have
a very straightforward implementation. The private timer will tick every millisecond, and its ISR
will increment the system time. So system time itself will also be measured in milliseconds. Then
systime.c is exceedingly simple:

1 #include "systime.h"
2
3 static volatile systime_t systime;
4
5 void systime_tick(void) {
6
7 }

systime++;

Daniels Umanovskis

87

Bare-metal programming for ARM

ac090d8bfc

8
9 systime_t systime_get(void) {
10
11 }

return systime;

The systime_t type is defined in the corresponding header file, as typedef uint32_t systime_t
.

To make use of this, the private timer’s ISR is modified so it simply calls systime_tick after clearing
the hardware interrupt flag.

1 void ptimer_isr(void) {
2
3
4 }

WRITE32(regs->ISR, ISR_CLEAR); /* Clear the interrupt */
systime_tick();

That’s it, just change the ptimer_init call in main to use a 1-millisecond period, and you have a
system-wide time that can be accessed whenever needed.

Overflows and spaceships

A discussion of timers is an opportune time to not only make a bad pun but also to mention overflows.
By no means limited to embedded programming, overflows are nonetheless more prominent in low-
level systems programming. As a refresher, an overflow occurs when the result of a calculation exceeds
the maximum range of its datatype. A uint8_t has the maximum value 255 and so 255 + 1 would
cause an overflow.

Timers in particular tend to overflow. For example, our use of uint32_t for system time means that
the maximum system timer value is 0xFF FF FF FF, or just shy of 4.3 billion in decimal. A timer that
ticks every millisecond will reach that number after 49 days. So code that assumes a timer will always
keep inceasing can break in mysterious ways after 49 days. This kind of bug is notoriously difficult to
track to down.

One solution is of course to use a bigger data type. Using a 64-bit integer to represent a millisecond
timer would be sufficient for 292 billion years. This does little to address problems in older systems,
however. Many UNIX-based systems begin counting time from the 1st of January, 1970, and use a 32-bit
integer, giving rise to what’s known as the Year 2038 problem, as such systems cannot represent any
time after January 19, 2038.

When overflows are possible, code should account for them. Sometimes overflows can be disregarded,
but saying that something “cannot happen” is dangerous. It’s reasonable to assume, for example, that

Daniels Umanovskis

88

Bare-metal programming for ARM

ac090d8bfc

a microwave oven won’t be running for 49 days in a row, but in some circumstances such assumptions
should not be made.

One example of an expensive, irrecoverable overflow bug is the NASA spacecraft Deep Impact. After
more than eight years in space, and multiple significant accomplishments including excavating a
comet, Deep Impact suddenly lost contact with Earth. That was due to a 32-bit timer overflowing and
causing the onboard computers to continuously reboot.

Overflow bugs can go unnoticed for many years. The binary search algorithm, which is very widely
used, is often implemented incorrectly due to an overflow bug, and that bug was not noticed for two
decades, in which it even made its way into the Java language’s standard library.

Scheduler types

A scheduler is responsible for allocating necessary resources to do some work. To make various bits
of code run on a schedule, CPU time is the resource to be allocated, and the various tasks comprise
work. Different types of schedulers and different scheduling algorithms exist, with a specific choice
depending on the use case and the system’s constraints.

One useful concept to understand is that of real-time systems. Such a system has constraints, or
deadlines, on some timings, and these deadlines are expressed in specific time measurements. A “fast
response” isn’t specific, “response within 2 milliseconds” is. Further, a real-time system is said to be
hard real-time if it cannot be allowed to miss any deadlines at all, that is, a single missed deadline
constitutes a system failure. Real-time systems are commonly found in embedded systems controlling
aircraft, cars, industrial or medical equipment, and so on.By contrast, most consumer-oriented software
isn’t real-time.

A real-time system requires a scheduler that can guarantee adherence to the required deadlines. Such
systems will typically run a real-time operating system (RTOS) which provides the necessary scheduling
support. We’re not dealing with real-time constraints, and we’re not writing an operating system, so
putting RTOS aside, there are two classes of schedulers to consider.

Cooperative schedulers provide cooperative (or non-preemptive) multitasking. In this case, every task is
allowed to run until it returns control to the scheduler, at which point the scheduler can start another
task. Cooperative scedulers are easy to implement, but their major downside is relying on each task
to make reasonable use of CPU resources. A poorly-written task can cause the entire system to slow
down or even hang entirely. Implementing individual tasks is also simpler in the cooperative case - the
task can assume that it will not be interrupted, and will instead run from start to finish.

Cooperative scheduling is fairly common in low-resource embedded systems, and the implementation
only requires that some kind of system-wide time exists. One example of a suitable system could be a

Daniels Umanovskis

89

Bare-metal programming for ARM

ac090d8bfc

thermostat. It can have several tasks (measure room temperature, calculate necessary output, read
user settings, log some data) running with the same periods indefinitely.

Preemptive schedulers use interrupts to preempt, or suspend, a task and hand control over to another
task. This ensures that one task is not able to hang the entire system, or just take too long before letting
other tasks run. Such schedulers implement some kind of algorithm for choosing when to interrupt a
task and what task to execute next, and implementing actual preemption is another challenge.

Cooperative scheduler

To implement a basic cooperative scheduler, we don’t need much code. We need to keep track of what
tasks exist in the system, how often they want to run, and then the scheduler should execute those
tasks. The scheduler’s header file can be written so:

task_entry_ptr entry;
systime_t period;
systime_t last_run;

1 #include "systime.h"
2
3 typedef void (*task_entry_ptr)(void);
4
5 typedef struct {
6
7
8
9 } task_desc;
10
11 typedef enum {
12
13
14 } sched_error;
15
16 #define MAX_NUM_TASKS (10u)
17
18 sched_error sched_add_task(task_entry_ptr entry, systime_t period);
19 void sched_run(void);

SCHED_OK = 0,
SCHED_TOO_MANY_TASKS

Each task should have an entry point, a function that returns void and has no parameters. A pointer
to the entry point, together with the desired task period and the time of the last run, form the task
descriptor in the task_desc type. The scheduler provides a function sched_add_task, which can
add tasks to the scheduler at run-time. Let’s look at the implementation. Here’s how sched.c starts:

1 #include <stddef.h>
2 #include <stdint.h>
3 #include "sched.h"

Daniels Umanovskis

90

Bare-metal programming for ARM

ac090d8bfc

}

return SCHED_TOO_MANY_TASKS;

if (table_idx >= MAX_NUM_TASKS) {

4
5 static task_desc task_table[MAX_NUM_TASKS] = {0};
6 static uint_8 table_idx = 0;
7
8 sched_error sched_add_task(task_entry_ptr entry, systime_t period) {
9
10
11
12
13
14
15
16
17
18
19
20
21 }

};
task_table[table_idx++] = task;

.entry = entry,
.period = period,
.last_run = 0

return SCHED_OK;

task_desc task = {

The task_table array is where all the task descriptors are kept. sched_add_task is pretty simple -
if there’s free space in the task table, the function creates a task descriptor and adds it to the table. The
task’s last run time is set to 0. Then the interesting work happens in the scheduler’s sched_run:

for (uint8_t i = 0; i < MAX_NUM_TASKS; i++) {

task_desc* task = &task_table[i];
if (task->entry == NULL) {

1 void sched_run(void) {
while (1) {
2
3
4
5
6
7
8
9

}

continue;

if (task->last_run + task->period <= systime_get()) { /*

Overflow bug! */
task->last_run = systime_get();
task->entry();

}

}

}

10
11
12
13
14
15 }

NOTE

The source code from the repository for this chapter includes sched.c with this example, but doesn’t
built it by default, instead it builds sched_preemptive.c for the preemptive scheduler that we will
cover later. Change CMakeLists.txt if you wish to build the cooperative scheduler.

Daniels Umanovskis

91

Bare-metal programming for ARM

ac090d8bfc

You may have noticed that the function is contained within an infinite while (1) loop. The scheduler
isn’t supposed to terminate, and sched_run will be able to replace the infinite loop that we’ve had in
main all along.

The actual work is done in the scheduler’s for-loop, which loops through the entire task table, and
looks for tasks whose time to run has come, that is, the system time has increased by at least period
since the task’s last run time stored in last_run. When the scheduler finds such a task, it updates the
last run time and executes the task by calling its entry point.

All that remains in order to test the scheduler is to add a couple of tasks, and schedule them in main.

1 void task0(void) {
2
3
4
5
6
7
8 }
9
10 void task1(void) {
11
12
13
14
15
16
17 }

systime_t start = systime_get();
uart_write("Entering task 0... systime ");
uart_write_uint(start);
uart_write("\n");
while (start + 1000u > systime_get());
uart_write("Exiting task 0...\n");

systime_t start = systime_get();
uart_write("Entering task 1... systime ");
uart_write_uint(start);
uart_write("\n");
while (start + 1000u > systime_get());
uart_write("Exiting task 1...\n");

The above defines two tasks. Note how there’s nothing special about those functions, it’s sufficient for
them to be void functions with no parameters for them to be scheduled with our implementation.
Both tasks have the same behavior - print a message with the current system time, wait for 1000
system time ticks and exit with another message. To actually schedule them and hand control over to
the scheduler, modify main to get rid of the infinite loop and instead do:

1 (void)sched_add_task(&task0, 5000u);
2 (void)sched_add_task(&task1, 2000u);
3
4 sched_run();

That will schedule task 0 to run every 5000 ticks (roughly 5 seconds) and task 1 for every 2000 ticks.

Daniels Umanovskis

92

Bare-metal programming for ARM

ac090d8bfc

NOTE

The tasks use uart_write_uint to print the current systime, a new function not part of the previously-
written UART driver. We cannot use standard C library functions such as sprintf without performing
additional work, so this new function is a quick way to output unsigned numbers like systime. For
completeness, here’s the implementation:

char buf[8];
int8_t i = 0;
do {

1 void uart_write_uint(uint32_t num) {
2
3
4
5
6
7
8
9
10
11
12 }

uint8_t remainder = num % 10;
buf[i++] = '0' + remainder;
num /= 10;
} while (num != 0);
for (i--; i >= 0; i--) {

uart_putchar(buf[i]);

}

Running the system now should produce some output indicating the scheduler’s hard at work.

1 Welcome to Chapter 8, Scheduling!
2 Entering task 1... systime 2000
3 Exiting task 1...
4 Entering task 1... systime 4000
5 Exiting task 1...
6 Entering task 0... systime 5000
7 Exiting task 0...
8 Entering task 1... systime 6000
9 Exiting task 1...
10 Entering task 1... systime 8000
11 Exiting task 1...
12 Entering task 0... systime 10000
13 Exiting task 0...
14 Entering task 1... systime 11000
15 Exiting task 1...
16 Entering task 1... systime 13000
17 Exiting task 1...
18 Entering task 0... systime 15000

Daniels Umanovskis

93

Bare-metal programming for ARM

ac090d8bfc

19 Exiting task 0...
20 Entering task 1... systime 16000
21 Exiting task 1...

What does that tell us? Well, the scheduler seems to be working fine. Task 0 first gets executed at
systime 5000, which is enough time for task 1 to run twice, starting at times 2000 and 4000. We can
also see that this is very much not a real-time system, as the schedule we’ve provided serves as a
suggestion for the scheduler but isn’t strictly enforced. At systime 10000, it’s time for both tasks to be
executed (task 0 for the 2nd time and task 1 for the 5th time), but task 0 gets the execution slot (due
to having its entry earlier in the task table), and task 1 gets delayed until systime 11000, when task 0
finishes.

In the scheduler’s sched_run loop, you may have noticed a comment about an overflow bug on
one line. After saying how spacecraft can get lost due to overflows, it wouldn’t feel right to omit an
example.

1 if (task->last_run + task->period <= systime_get()) { /* Overflow bug!

*/

The normal case for that line is straightforward - if at least period ticks have passed since last_run
, the task needs to be run. How about when some of these variables approach UINT32_MAX, the
maximum value they can hold? Suppose last_run is almost at the maximum value, e.g. UINT32_MAX
- 10 (that’s 0xFF FF FF F5). Let’s say period is 100. So the task ran at UINT32_MAX - 10
ticks, took some time to complete, and the scheduler loop runs against at, for instance, systime
UINT32_MAX - 8. Two ticks have passed since the last execution, so the task shouldn’t run. But the
calculation last_run + period is UINT32_MAX - 10 + 100, which overflows! Unsigned integers
wrap around to zero on overflow, and so the result becomes 89. That is less than the current system
time, and the task will run again. And the problem will repeat in the next iteration as well, until
eventually fixing itself after system time also overflows and wraps around to zero.

That’s a potentially serious bug that is much easier to introduce than to notice. How to perform that
calculation safely, then? Instead of adding the timestamps, you should use subtraction so that the
intermediate result is a duration. That check should be:

1 if (systime_get() - task->last_run >= task->period)

It might seem like such a calculation can go wrong. Supppose last_run is close to overflow, as in the
example before, and systime has recently rolled over. So you’d be subtracting a very large last_run
from the small positive result of systime_get, which would result in a large negative number under

Daniels Umanovskis

94

Bare-metal programming for ARM

ac090d8bfc

ordinary arithmetic. But with unsigned integers, that will still result in a correct calculation. So if the
above calculation amounts to something like:

1 if (20 - (UINT32_MAX - 10) >= 100)

the left side will evaluate to 31, and the if will evaluate to false.

Mathematically speaking, unsigned integers in C implement arithmetics modulo the type’s maximum
value plus one. So a uint8_t, holding the maximum value 255, performs all operations modulo 256.
Importantly, only unsigned integers have overflow behavior that is guaranteed by the C standard.
Signed integer overflow is undefined behavior, and you should avoid writing code that relies on signed
integer overflows.

Life is good when you have a working scheduler, and indeed this kind of simple scheduler is pretty
similar to what some embedded systems run in the real world. There are small improvements that can
be made to the scheduler without fundamentally changing it, such as allowing each task to have a
priority, so that the highest-priority task would be chosen at times when multiple tasks wish to run,
such as at systime 10000 above. Note the low memory overhead of this scheduler. It uses 1 byte for the
table_idx variable and creates a 12-byte task descriptor for each task. This kind of memory use is
one of the reasons why such simple cooperative schedulers are a valid choice for resource-constrained
embedded systems.

It is also easy to introduce some problems that a cooperative scheduler won’t be able to manage well.
For example, create a third task that just runs a while (1); loop. The scheduler will run the task
and the system will hang with no hope of recovery. In a real system, a watchdog would probably be
configured and reset everything, but a reset loop is not much better than a plain hang.

Summary

In this chapter, we looked at the simplest way of making a bare-metal system perform some useful
work on a schedule. Doing this required a small new driver, and we built an abstract system time on
top of it. From there, it was just a few more steps to have a working scheduler that runs predefined
tasks on a predefined schedule.

A real-world cooperative scheduler would be slightly more complex, but not by much. For all its
simplicity, a cooperative scheduler is appropriate in simple embedded systems that don’t have to
quickly react to external input, and have well-known tasks that don’t interfere with one another and
aren’t expected to run for long. At the same time, the scheduler is anything but robust - a single error
in a task can cause it to hang permanently.

Daniels Umanovskis

95

Bare-metal programming for ARM

ac090d8bfc

We also saw how overflow errors are easy to introduce, and happen to often be associated with timing
code.

Daniels Umanovskis

96