This repo serves as a collection of low level examples.  No operating
system, embedded or low level embedded or deeply embedded or bare metal,
whatever your term is for this.

I am in no way shape or form associated with the raspberry pi organization
nor broadcom.  I just happen to own one (some) and am sharing my
experiences.  The raspberry pi is about education, and I feel low
level education is just as important as Python programming.

From what we know so far there is a gpu on chip which:

1) boots off of an on chip rom of some sort
2) reads the sd card and looks for additional gpu specific boot files
bootcode.bin, loader.bin, start.elf in the root dir of the first partition
(fat formatted)
3) in the same dir it looks for config.txt which you can do things like
change the arm speed from the default 700MHz, change the address where
to load kernel.img, and many others
4) it reads kernel.img the arm boot binary file and copies it to memory
5) releases reset on the arm such that it runs from the address where
the kernel.img data was written

I have tried a few things for example incorrectly assuming kernel.img
was loaded at address 0x00000000 as a default.  If you tell it zero you
still get 0x8000.  The arm and gpu share memory I am guessing the gpu
is using that first part of memory, dont really know.  0x8000 being
familiar to linux folks and this is intended to be first a linux
computer so this makes sense.

You will want to go here
http://elinux.org/RPi_Hardware
And get the datasheet for the part
http://www.raspberrypi.org/wp-content/uploads/2012/02/BCM2835-ARM-Peripherals.pdf
(might be an old link, find the one on the wiki page)
And the schematic for the board
http://www.raspberrypi.org/wp-content/uploads/2012/04/Raspberry-Pi-Schematics-R1.0.pdf
(might be an old link, find the one on the wiki page)

Early in the BCM2835 document you see a memory map.  I am going to
operate based on the middle map, this is how the ARM comes up.  The
left side is the system which we dont have direct access to in that
form, the gpu probably, not the ARM.  The ARM comes up with a memory
space that is basically 0x40000000 bytes in size as it mentions in
the middle chart.  256MBytes is 0x10000000, I am guessing they gave
room to add memory as peripherals are mapped into arm address space at
0x20000000.  When you see 0x7Exxxxxx in the manual replace that with
0x20xxxxxx as your ARM physical address.  Experimentally I have seen
the memory repeats every 0x40000000, read 0x40008000 and you see the
data from 0x8000.  I wouldnt rely on this, just an observation (likely
ignoring the upper address bits in the memory controller).

Now this memory map shows that somewhere within the SDRAM address space
there is a boundary between the ARM memory in the lower addresses and
the GPU memory in the upper addresses.  That boundary is not explained
in that document.  There are at the moment three gpu start.elf files
to choose from.  A or the difference between them is where this boundary
lies.  The default is a 50/50 split the arm gets 128MBytes, and the gpu
128MBytes.  arm224_start.elf for example gives 224MBytes to the ARM
and 32MBytes to the GPU.  They say that you dont get full gpu performance
if you limit the gpu memory this much.  These examples are all going
to assume that the ARM only has 128MBytes and the default boot setting
of 0x8000 for the kernel_address.

I do not normally zero out .bss or use .data (see the bssdata example)
nor gcc libraries nor C libraries so you can build most if not all of
my examples using a gcc cross compiler.  Basically it doesnt matter if
you use arm-none-linux-gnueabi or arm-none-eabi.  What was formerly
codesourcery.com still has a LITE version of their toolchain which is
easy to come by, easy to install and well maybe not easy to use but you
can use it.  Building your own toolchain from gnu sources (binutils and
gcc) is fairly straight forward see the build_gcc directory for a build
script.

As far as we know so far the Raspberry Pi is not "brickable".  Normally
what brickable means is the processor relies on a boot flash and with
that flash it is possible to change/erase it such that the processor will
not boot normally.  Brickable and bricked sometimes excludes things
like jtag or speciall programming headers.  From the customers perspective
a bricked board is...bricked.  But on return to the vendor they may
have other equipment that is able to recover the board without doing
any soldering, perhaps plugging in jtag or some other cable on pins/pads
they have declared as reserved.  Take apart a tv remote control or
calculator, etc and you may see some holes or pads on the circuit board,
for some of these devices just opening the battery case you have a view
of some of the pcboard.  This is no doubt a programming header.  Long
story short, so far as I know the Raspberry Pi is not brickable because
the rom/flash that performs the initial boot is for the gpu and we dont
have access to the gpu nor its boot rom/flash.  The gpu relies on the
sd card to complete the boot, so the sd card flash is really the boot
flash for the system.  And it is very easy for the customer to remove
and replace/modify that boot flash.  So from a software perspective
unless you/we accidentally figure out how to change/erase the gpu boot
code (my guess is it is a one time programmable) you cant brick it.

To use my samples you do not need a huge sd card.  Nor do you need nor
want to download one of the linux images, takes a while to download,
takes a bigger sd card to use, and takes forever to write to the sd card.
AND I am not able to run with the firmware on those cards.  I use the
firmware from http://github.com/raspberrypi.  The minimum configuration
you need to get started at this level is:

go to http://github.com/raspberrypi, you DO NOT need to download
the repo, they have some large files in there you will not need (for
my examples).  go to the firmware directory and then the boot directory.
For each of these files, bootcode.bin, loader.bin, start.elf (NOT
kernel.img, dont need it, too big).  Click on the file name, it will
go to another page then click on View Raw and it will let you download
the file.  For reference, I do not use nor have a config.txt file on my
sd card.  I only have the four files.

bootcode.bin is about 2MBytes, the other files are smaller, so you will
want an sd card that is at least a few meg, probably a full 128MBytes or
256MBytes or larger (gigabytes are just fine) for the gpu files plus
the sample files here.  ARM memory for these samples is assumed 128MB
so they wont be even that large.

What I do is setup the sd card with a single partition, fat32.  And
copy the above files.  bootcode.bin, loader.bin and start.elf.  From
there you take .bin files from my examples and place them on the sd card
with the name kernel.img.  It wont take you very long to figure out this
is going to get painful.

1) power off raspi
2) remove sd card
3) insert sd card in reader
4) plug reader into computer
5) mount/wait
6) copy binary file to kernel.img
7) sync/wait
8) unmount
9) insert sd card in raspi
10) power raspi
11) repeat

There are ways to avoid this, one is jtag, which is not as expensive
as it used to be.  It used to be in the thousands of dollars, now it
is under $50 and the software tools are free.  Now the raspi does have
jtag on the arm, getting the jtag connected to that is going to require
some soldering.  (this is described later and in the armjtag sample).

Another method is a bootloader, typically you use a serial port connected
to the target processor.  That processor boots a bootloader program that
in some way, shape, or form allows you to interact with the bootloader
and load programs into memory (that the bootloader is not using itself)
and then the bootloader branches/jumps to your program.  If your program
is still being debugged and you want to try again, you reboot the processor
the bootloader comes up and you can try again without having to move any
sd cards, etc.  The sd card dance above is now replaced with the
bootloader dance:

1) power off raspi
2) power on raspi
3) type command to load and start new program

I have working bootloader examples.  bootloader03 is the currently
recommended version.  But you need more hardware (no soldering is
required).  For those old enough to know what a serial port is, you
CANNOT connect your raspberry pi directly to this port, you will fry
the raspberry pi.  You need some sort of serial port at 3.3V either
a level shifter of some sort (transceiver like a MAX232) or a usb
serial port where the signals are 3.3V (dont need to use RS232 just
stay at the logic level).  The solution I recommend is a non-solder
solution:

http://www.sparkfun.com/products/9873
plus some male/female wire
http://www.sparkfun.com/products/9140

Solutions that may involve soldering
http://www.sparkfun.com/products/718
http://www.sparkfun.com/products/8430

Or this for level shifting to a real com port.
http://www.sparkfun.com/products/449

On the raspberry pi, the connector with two rows of a bunch of pins is
P1.  Starting at that corner of the board, the outside corner pin is
pin 2.  From pin 2 heading toward the yellow rca connector the pins
are 2, 4, 6, 8, 10.  Pin 6 connect to gnd on the usb to serial board
pin 8 is TX out of the raspi connect that to RX on the usb to serial
board.  pin 10 is RX into the raspi, connect that to TX on the usb to
serial board.  Careful not to have metal items on the usb to serial
board touch metal items on the raspberry pi (other than the three
connections described).  On your host computer you will want to use
some sort of dumb terminal program, minicom, putty, etc.  Set the
serial port (usb to serial board) to 115200 baud, 8 data bits, no
parity, 1 stop bit.  NO flow control.  With minicom to get no flow
control you normally have to save the config, exit minicom, then
restart minicom and load the config in order for flow control
changes to take effect.  Once you have a saved config you dont have
to mess with it any more.

Read more about the bootloaders in their local README files.  Likewise
if you are able to do some soldering on electronics see the armjtag
README file.  Other than chewing up a few more GPIO pins, and another
thing you have to buy, the jtag solution is the most powerful and useful.
My typical setup is the armjtag binary as kernel.img, a usb to jtag
board like the amontec jtag-tiny and a usb to serial using minicom.

As far as these samples go I recommend starting with blinker01 then
follow the discovery of the chip into uart01, etc.  You will need some
sort of cross compiler (well maybe a native compiler on a raspberry pi
or other arm system).  See the build_gcc directory if you cant get
the LITE version from codesourcery.com (now mentor graphics).

UPDATE: I added the baremetal directory, it has a tutorial on one
aspect of the things you need to master to do bare metal programming.

UPDATE: I added the pitopi directory (pi to pi).  If you have two
Raspberry Pi boards, instead of buying one of the usb to serial or
other solutions above you can use one Raspberry Pi as the development
platform which already has a uart at the right line levels, and the
other Raspberry Pi for running the bare metal programs.