diff --git a/ports/cortex-m/README.md b/ports/cortex-m/README.md
index 2114a47..ae3ad3f 100644
--- a/ports/cortex-m/README.md
+++ b/ports/cortex-m/README.md
@@ -1 +1,118 @@
-# Comming soon...
\ No newline at end of file
+# ARM Cortex-M Port
+
+## Summary and Prerequisites
+This port should run on any Cortex-M3/4/4F (but not on M0, sorry). It uses
+RedHat's [newlib](https://sourceware.org/newlib/) and [libopencm3]
+(https://github.com/libopencm3/libopencm3). You will also need an ARM compiler
+toolchain. If you want to flash or debug your target, [openocd](http://openocd.org)
+ is also a must. For STM32 targets [stlink](https://github.com/texane/stlink)
+might also be helpful.
+
+On Debian systems, compiler, newlib and openocd can easily be installed by the
+package manager (untested, not going to set up a new system just to test this):
+
+``` 
+apt-get install gcc-arm-none-eabi binutils-arm-none-eabi
+apt-get install libnewlib-arm-none-eabi libnewlib-dev
+apt-get install openocd
+```
+
+## Code Layout
+The "classic" port components (code needed for task setup and context
+switching and the atomport{-private}.h headers) are residing in the
+top level port directory.
+
+There are additional subdirectories:
+
+* **boards** contains subdirectories for specific hardware. Each
+board needs at least a Makefile, which defines certain variables describing
+the hardware used, as well as a list of extra object files needed. There will
+usually be at least a board_setup.c, which contains code to initialise the
+hardware properly (clocks, uart, systick timer)
+
+* **common** contains code needed by multiple boards, such as
+stub functions for newlib or routines shared by multiple boards using the
+same family of target MCUs.
+
+* **linker** contains linker script fragments for specific target MCUs. These
+just define the memory areas (RAM, Flash) of the chip and include libopencm3's
+main linker script for the appropriate chip family
+
+* **libopencm3** unless you compile and link against an external installation
+of libopencm3, this will be a Git submodule containing, surprise!, libopencm3
+
+* **build** this is a temporary directory where all object files end up in and
+which will be removed by ``make clean``  
+
+## Build Preparation
+Unless you decide to use an external installation of libopencm3, you will have
+to set up the libopencm3 sub-module:
+```
+git submodule add https://github.com/libopencm3/libopencm3.git
+git submodule init
+git submodule update
+```
+Optional: As of 2015-07-08 the libopencm3 API has not been declared stable. If
+future changes break the build, you can check out the older revision used while
+developing this port:
+```
+cd libopencm3
+git checkout a4bb8f7e240c9f238384cf86d009002ba42a25ed
+```
+
+## Building and Flashing
+To build the test suite, run
+```
+make BOARD=*yourboard* all
+```
+where *yourboard* is the name of a directory in **boards**. If no BOARD is
+given, it defaults to nucleo-f103rb.
+
+To build with an external libopencm3, run
+```
+make BOARD=*yourboard* OPENCM3_DIR=*/path/to/libopencm3* all
+```
+If your board Makefile also set up the necessary openocd variables, you can
+use it to flash the application image:
+```
+make BOARD=*yourboard* BINARY=*yourimage* flash
+```
+N.B.: with the ek-lm4f120xl board openocd will report an error after flashing.
+I think it is because it can not handle the changed core clock after the 
+application starts, but I simply can't be bothered to further investigate this.
+The application gets flashed and you can ignore the error.
+
+## Adding New Boards
+*TODO*
+
+## Port Internals
+The Cortex-M port is different from the other architectures insofar as it makes
+use of two particular features. First, it uses separate stacks for thread and
+exception context. When the core enters exception mode, it first pushes xPSR,
+PC, LR, r0-r3 and r12 on the currently active stack (probably the thread stack
+in PSP) and then switches to the main stack stored in MSP. It also stores a
+special EXC_RETURN code in LR which, when loaded into the PC, will determine
+if on return program execution continues to use the MSP or switches over to 
+the PSP and, on cores with an FPU, whether FPU registers need to be restored.
+The Cortex-M also implements a nested vectored interrupt controller (NVIC),
+which means that a running ISR may be preempted by an exception of higher
+priority.
+
+The use of separate stacks for thread and exception context has the nice
+implication that you do not have to reserve space on every task's stack for 
+possible use by ISRs. But it also means that it breaks atomthreads concept 
+of simply swapping task stacks, regardless of if atomSched() was called from
+thread or interrupt context. We would have to implement different 
+archContextSwitch() functions called from thread or exception context and
+also do messy stack manipulations depending on whether the task to be
+scheduled in was scheduled out in in the same context or not. Yuck!
+And don't get me started on nested exceptions calling atomIntExit()...   
+
+This is where the second feature comes handy, the PendSV mechanism.
+PendSV is an asynchronous exception with the lowest possible priority, which
+means that, when triggered, it will be called by the NVIC if there are no
+other exceptions pending or running. We use it by having archContextSwitch()
+set up a pointer to the TCB that should be scheduled in and then trigger the
+PendSv exception. As soon as program flow leaves the critical section or 
+performs the outermost exception return, the pend_sv_handler() will be called
+and the thread context switch takes place.