Edit on 01.01.2013: Seems to be that it is right what people say after all: The Raspberry Pi does not support repeated start. The stuff below still works for my device, but apparently only by chance; the behaviour of the controller is not correct. Here's a screenshot of the communication:So, since this is the open source part of the world, we can just go look at what they do, and steal it! ;)
|Communication generated by the code below. After the first data byte (the register address, 4) has been transferred, you can easily spot the spike in SDA (yellow) while SCL (teal) is high, meaning STOP START. This is incorrect.
The sources are easy to find (here, for example), and you can just modify i2cget.c to do what you need. Here is my code for accessing the magnetometer mentioned above, as an example.
Oh, and here's a fancy image of the board with the magnetometer mentioned above:
|Board with a three-axis magnetometer, in a 2mm x 2mm DFN package
I should also mention that the magnetometer itself is pretty awesome! It's a three-axis magnetometer with 16 bits resolution per axis (!), and it costs less than two euros at farnell (actually the whole board shown above is only about two euros in total if you add up the components, PCB, and the solder). The device is well sensitive enough to measure the earth's magnetic field; thus, you can use it to accurately determine its absolute rotation. I did a test which rotates meshes in blender in real-time when you rotate the device -- it worked quite well! It needs a lot of adjustment to the axes tough, they don't really match yet. I'll write about that again if I get it to work nicely.
Here's an example plot when rotating the device around a bit:
|Raw data received from magnetometer when rotating it by hand
|Moving a large magnet towards the device. t=0 corresponds to ~1m distance. The peaks in the y and z axes in the end originate from the magnet being rotated a bit while moving it away.