Over winter break, I successfully modified Kristina Brook’s free Raspberry Pi firmware, a low-level program for hardware initialisation, which with my contributions is able to bring-up and boot a Linux kernel. Most of the bits and pieces were already there from when the firmware was actively developed in June of 2016; that being said, it was purely a proof-of-concept showing ARM initialisation – no payloads were actually loaded, let alone Linux. In this post, we’ll walk through the steps of writing the necessary bootloader (and all the pitfalls to which we were victims.)

First things first, what exactly is it that we’re trying to accomplish? Well, we need to know a bit about how computers boot. Essentially it’s a chain of programs, each larger than the last whose job is to load the next one. Over years of legacy code piling up, it is admittedly rather convoluted, especially on x86. But the theory is the same: the bootrom (burned into the chip) loads the firmware, the firmware loads the bootloader, the bootloader loads the kernel, the kernel loads the userspace. The exact contents of each step vary wildly – and before you ask, yes, this is the reason your computer takes so long to load in the morning :-)

The Raspberry Pi is somewhat… special in this regard. Usually, for systems with a graphics processor (like my Intel-based laptop), the CPU boots up and later triggers bring-up of the GPU. For various (strange) historical reasons, the Raspberry Pi instead boots from the VideoCore 4 GPU, which loads the CPU at-will. Another caveat is that the firmware is not actually on a ROM like you would expect – it’s on an SD card, or potentially even stranger mediums. Finally, while third-party bootloaders like U-Boot exist for the Pi, the stock firmware is setup to boot Linux directly. So, the boot chain for the official firmware is something like: bootrom (VC4) loads the GPU operating system, the GPU loads the kernel into memory, the GPU loads ARM with a stub program, the ARM stub jumps to Linux, and Linux loads user-space. It’s pretty crazy.

Some of these, er, peculiarities are due to fundamental differences in the system-on-chip used in the Raspberry Pi. Much of it was simply design of the firmware. Kristina’s firmware used a slightly different path, from which we’ll work: the bootrom loads her minimal GPU program, the GPU brings up ARM with an embedded ARM program, the ARM program initialises a handful of additional peripherals and then hangs. What peripherals, you may ask? Well, most notably she was kind enough to include an eMMC driver (for SD cards) and she bundled a library for reading FAT filesystems. That is, the infrastructure is setup to read files from the SD card.

The (simplest) path from here is clear. Rather than hang at the end of ARM program, load the kernel from the SD card and chuck it somewhere into memory. Then, in theory, you should be able to jump to it on ARM, and let Torvalds’ crowd do the rest. It’s easy… right?

Unfortunately, Linux is rather… demanding, hence why bootloaders are used at all. If you’re curious at the exact details on ARM – which are comparatively trivial next to certain other architectures, ahem – see the official documentation. See, before we can just jump to the kernel image in memory, it’s necessary to pass a variety of arguments to it. There are the easy ones, like setting a few registers to magic values, but something stands out as particularly irritating: the legacy “tagged list” or the more modern device tree. To understand the purpose – and associated nightmares – of these, it’s necessary to take another detour into the typical boot process.

Essentially, the kernel has two jobs: managing resources for users via scheduling, permission systems, and the like (interesting to other people), and managing the hardware (interesting to us). The former category is easy, and it’s fairly similar across architectures – when people say a kernel is “a UNIX system”, they are generally referring to this component. The latter, however, is necessarily different for each and every hardware configuration. Actually implementing these drivers is a nightmare we’ll approach later; for now, it’s just necessary to know it’s there, and to ask the inevitable question: precisely what hardware does the kernel manage?

It turns out there are a few obvious ways of handling this. The kernel could guess (bad idea). It could hard code what hardware is available (simple but unmaintainable). It could probe the system at run-time (sounds nice but is inevitably difficult in practice). Certain architectures, of course, favour certain methods. x86, for one, favours probing, encouraged by (highly criticized) systems like ACPI, EFI, and the BIOS. On the other hand, ARM, used in the Raspberry Pi and focused on embedded systems, assumes the kernel automagically knows what hardware to use, by hard-coding I suppose. Linux on ARM takes a middle ground: it uses a configurable device tree that contains a variety of information about the hardware, loaded at runtime. Normally I’d write this off as something for the geeks on the mailing list to sort out, but apparently information “loaded at runtime” is our responsibility now.

In a (rare) stroke of luck, the device tree files are the same with the stock firmware… mostly. We can get much of what we need from the official trees, anyways. It’s tempting to throw that on the SD card along with the kernel, load that into memory along with the kernel, and now jump. Oh, and pass the address of the blob in a register before jumping, as per the Linux specification. Surely we’re done, right?

Nope! See, the device tree blob does quite a bit more than simply enumerating peripherals. One field that the bootloader (that is, our code) needs to fill is the kernel arguments. For those of you who aren’t familiar with Linux, the kernel accepts a handful of parameters, much like a userspace program, controlling all sorts of behaviours; importantly, they determine the root device, which will matter later. So, the strategy is, like the stock firmware, throw the command-line options into a file on the SD card, load them into memory, and pass it – err, how do we pass this exactly? I did say it was in the device tree. Yes, we need to modify the device tree at run-time1. Pull in libfdt, and patch in the correct field (chosen/bootargs), and that’s that. Additionally, a memory map needs to be specified under the /memory node, which can be patched in the same way. Onwards!

If we had a dynamic initramfs we wanted to load, that would be yet a third blob to load from the SD card and chuck at the kernel, though to keep everything simple, we can skip this (for now). Indeed, at this point everything is place for the boot. According to the specification, we need to set a few registers to magic values and the device tree address, and jump to the kernel. Wait a minute, registers? This is C code! Ugh, I guess we’ll need to drop down into assembly for the final jump…

Alternatively, we can appreciate the Linux’s choice of registers, r0 through r2. Depending on your familiarity with low-level ARM code, you might recognize these registers as storage for the first three arguments in the standard ARM calling convention. That is, with the right type specification, you can cast the kernel blob in memory to a function, and just “call” it! The first two arguments are magic numbers, zero and all ones, and the third argument is the DTB pointer, so we define it like so:

typedef void (*linux_t)(uint32_t, uint32_t, void*);

Later, we cast the kernel blob:

linux_t kernel = reinterpret_cast<linux_t>(zImage);

And finally, everything simply falls into place, in an elegance atypical of embedded systems development:

kernel(0, ~0, dtb_address);

Woo-hoo! We’re done, right? I wish. Compiling the kernel itself is not such a big deal, if enough peripherals are disabled to the point of uselessness. Minimally, we should expect to see something. I’d settle for an error message right now… but nope! Just silence. I swear, I made sure earlyprintk was enabled and setup for the Raspberry Pi’s UART, the PL011 – oh, shoot. Houston, we have a problem.

For those who are not familiar, a UART is a chip that lets a computer easily communicate with a serial port. This is nice, because it is really easy to setup a UART on most architectures, and the protocol is ubiquitous, so it can be used for debugging all sorts of appliances. Kristina’s firmware had alreay setup the UART for debug, and with the appropriate USB-to-serial cable, it’s a piece of cake to play with in screen or minicom. So, why won’t Linux use it? Well, it turns out that the Raspberry Pi actually has two UARTs on board, not just one. The main UART (a PL011) is used by Linux, both for earlyprintk and generally for the system. The auxillary UART, on the other hand, is not normally used, although for simplicity Kristina chose to use it instead, ignoring bring-up for the “real” UART. At this point, we’ll need to write a new driver for this chip and migrate the low-level, two-thousand line project to it. OK, I’m exaggerating the complexity a bit; the registers are accessed slightly differently, and initialisation requires additional clocking, but honestly it’s not a big deal, and the code is fairly standard for bare-metal Raspberry Pi development. OK, now we should be done… right?

Well, depends how you define ‘done’, of course! Indeed, we can now verify that Linux is beginning its early boot process, and if you’re lucky, your console will be flooded by debug information from the kernel. But it won’t boot, yet, since there isn’t a userspace to boot. Remember, on our massive purge to compile a minimal kernel as fast as possible, we ignored even the SD card driver. That’s a bit of a problem, no?

So, first things first, let’s build a minimal environment, since initialising a full GNU/Linux system will likely convolute this further. Instead, it is pretty easy to build a BusyBox binary, which is enough to act as an initial ramdisk (statically linked into the kernel via a special configuration option, to avoid adding another phase to the bootloader). And yes, in theory if the command-line option rdinit=/bin/sh is added, you will boot to a shell! Hooray Linux!

But let’s not stop here – I did mention the SD card driver was a big deal. We can enable this driver from menuconfig, and compile the kernel, and — oh no, missing symbols! rpi_firmware_xyz where xyz is a handful of functions. Did we forget a driver? Not quite; these are from the mailbox interface. In particular, the stock firmware exposes a handful of helper functions, akin to the BIOS on the PC architecture, to provide for functions that the kernel can’t (or in this case, won’t) support itself. Now, this is a problem, since we’re not using the stock firmware. We could mimic the API, of course, and implement all of those methods ourselves, and while this would let us use, say, U-Boot without ports, it would be a huge amount of work, for little gain. Instead, it turns out much of this is also implemented in Linux, and it just optionally uses the firmware. Reading between the lines a bit, we can just comment out a bit of code, and the SD card driver will compile happily, and even work, mostly. There are some outstanding bugs regarding the SD card driver, although I suspect this issue is in Kristina’s territory now. Nonetheless, you can mount the SD card from the BusyBox ramdisk and chroot into Debian, if you’d like. vim works surprisingly well at 115200 baud! Similar patches (and additional firmware bring-up) would be necessary in the future for more peripherals, like USB or Ethernet.

But hey, two thousand words later, Linux boots. Next time, remind me to try a microkernel.


  1. I am aware there are other ways to approach this in a super minimalist setup, although this is the approach used in the stock firmware and is therefore how users will expect it to work.