I agree designing HALs that play well across vendors is difficult since the peripherals have all kinds of limitations and edge cases, sometimes even in device families from the same silicon vendor. :(

STOP is the main thing you need full control over though, yes (based on my own experience anyway). You can have repeated starts, but some sensors will expect: START - READ - STOP - START - WRITE - STOP ... and some will expect: START - READ - START - WRITE - STOP, and they often won't work as expected using the wrong pattern.

The two gotchas I consistently come across are needing control over when the STOP bit is inserted (i.e. having that in the right place in the drivers), and more rarely on sensors that use clock stretching making sure that the master handles that properly. There are several 'I2C' peripheral blocks out there that don't handle clock stretching properly, though this isn't an API level issue, rather it's a HW problem (unless you are bit-banging I2C of course).


On 18/10/16 23:50, will sanfilippo wrote:
Well, I do agree that this new API I am proposing is not as clean as the old 
API; just dont know what else to do. I guess this is part of the problem trying 
to make generic HALs.

At least you will be able to control when the STOP occurs which is the most 
important thing I think.

On Oct 18, 2016, at 1:24 PM, Kevin Townsend <ke...@adafruit.com> wrote:

Hi Will,

Having control over when the STOP happens is important since different devices 
will behave differently, especially if you are performing a series of events 
like read/write/read or something similar. Some devices expect the STOP to 
happen in different places in a command sequence.

My vote would be a less attractive AP where the stop is a flag. Since we 
apparently aren't using _begin to set the START we may as well drop the STOP 
being handled in _end, but I'm obviously curious what other people think here.

Sending multiple STARTS is perfectly valid BTW: 


On 18/10/16 21:58, will sanfilippo wrote:

I am new to i2c so please bear with me. Hopefully this makes sense to the i2c 
experts out there.

In adding i2c to the nordic platforms I ran across an issue that I am not sure 
how to solve. Well, I know one way to solve it but it would require changes to 
the API. Here is the issue:

In order to get the nordic HW to generate a NACK bit (well, the opposite of an 
ACK) on the 9th bit after a read, you need to do something special to the HW 
prior to reading the last received character. When you do this, the HW will 
generate the correct 9th bit and will also put a STOP condition on the line. 
While it is possible to get the HW to generate a STOP independently, I dont see 
how I can get it to generate the correct 9th bit unless you know it is the last 
character you want to read.

Another issue that is bothersome is the nordic SDK itself: when you are done 
with the read it automatically generates the STOP. That is fine in and of 
itself; the SDK could be modified, but the issue is that dang 9th bit.

Our API has the following functions:
hal_i2c_master_begin(): Not exactly sure what this does on all platforms but 
currently is a NOOP on the nordic platforms.
hal_i2c_master_write(): generates a start but no stop.
hal_i2c_master_read(): generates a start but no stop.
hal_i2c_master_end(): generates a stop

At first glance, this seems to be a fine API and is easy to see a 
“transaction”: a begin, writes and/or reads chained together, and an end. 
Unfortunately, I could not see how to do this on the nordic platforms.

One way to do this would be to get rid of begin/end and have a flag in the 
read/write API. The flag would be “generate stop”. This way users can chain 
reads/writes together and end them with a STOP whenever they want. The only con 
to this is that it is not so easy to look at the code to see that transactions 
have a stop.

Given that I dont have alot of experience wth many i2c devices, I dont know if 
having the ability to skip START and/or address when the read/write APIs are 
called is useful. So far in my limited experience the start/address condition 
between the read/writes is not an issue.

So to summarize, here is what I suggest:
* Remove the begin/end API.
* Add a flag to write/read that will cause a STOP to be generated at the end of 
the read write.

Thanks! Comments greatly appreciated.

Reply via email to