This code is fairly robust. The clean up at the end of each function is probably no longer needed. I have not retested with multimaster modes after making some changes to NACK handling, but I don't think I messed up loss of arbitration, so it should be ok. If it works properly with multimaster, then the final cleanup can probably be removed. However, one can always replace it with something better if one discovers the driver is not quite perfect or there is some other problem that can't be pinned down.
This code depends on a logging task that logs to NVM. I am not giving that up, so replace the log call with a call of your own or an empty function.
You may use it without warranty. If you make it better, please post your changes to this entry. It was not easy to close all the loop holes. In particular, a hard to find problem was the write word code. If I did not check the status between the two data bytes, and there was a NACK on the first byte, the task would get lost in the weeds. The symptom was I would halt in the debugger and the task would be off in assembly somewhere with no stack information available.
