Re: [PATCH v4 1/1] can: add pruss CAN driver.

[Wolfgang Grandegger <wg@grandegger.com>]

On 05/04/2011 04:48 PM, Arnd Bergmann wrote:
> On Wednesday 04 May 2011, Wolfgang Grandegger wrote:
>> On 05/04/2011 03:11 PM, Arnd Bergmann wrote:
> 
>>> Wolfgang, I'm a bit worried by the API being split between sockets and sysfs.
>>> The problem is that once the sysfs API is established, users will start
>>> relying on it, and you can no longer migrate away from it, even when
>>> a later version of the Socket CAN also supports setting through a different
>>> interface. What is the current interface to set mail box IDs in software?
>>
>> Note that this CAN controller is *very* special. It cannot handle all
>> CAN id's due to a lack or resources. The PRUSS firmware is able to
>> manage just up to 8 different CAN identifiers out of the usual 4096
>> (12-bit) or even more for the extended CAN ids using 29 bits. 
> 
> So for other controllers, they can simply access every ID within
> the range (12 or 29 bits), but there is no need to configure?

Yes, 11 or 29 bits, to be correct.

> What are these IDs for? Do they identify a local port, a remote address,
> a connection, or something else?

It's a message identifier, which is used for bus arbitration and which
other CAN nodes can listen to. See also:

http://lxr.linux.no/#linux+v2.6.38/Documentation/networking/can.txt#L146
http://en.wikipedia.org/wiki/Controller_area_network

>> There is
>> no other CAN controller with such rather serious limitations and
>> therefore there exists also no appropriate interface. I think using
>> sysfs is OK for such device-specific parameters, at least for the time
>> being.
> 
> It sounds like it's not very scalable, especially since the implementation
> is done completely in firmware. Imagine a new firmware version suddenly
> supporting 256 ids instead of 8 -- you'd then have to create 256 sysfs
> files to be compatible if I understand you correctly.

Well, than an array of CAN identifiers per file would indeed be more
appropriate.

>>> How hard would it be to implement that feature in Socket CAN?
>>
>> CAN controllers usually provide some kind of hardware CAN id filtering,
>> but in a very hardware dependent way. A generic interface may be able to
>> handle the PRUSS restrictions as well. CAN devices are usually
>> configured through the netlink interface. e.g.
>>
>>   $ ip link set can0 up type can bitrate 125000
>>
>> and such a common interface would be netlink based as well.
> 
> Agreed.
> 
>>> Is that something that Subhasish or someone else could to as a prerequisite
>>> to merging the driver?
>>
>> Any ideas on how to handle hardware filtering in a generic way are
>> welcome. I will try to come up with a proposal sooner than later.
> 
> It sounds to me like the best solution would be change the firmware
> to lift that restriction and simply allow all IDs, in case it's not
> actually a hardware limitation (which sounds unlikely).

Yes, that would be best but they told us, that it's not possible with
the available hardware resources. Subhasish?

> If that's not possible, maybe it's possible to define a generic
> filtering interface using netlink, and then either do it completely
> in the kernel, or using the hardware support.

Well, I hesitate to implement an interface especially for such an exotic
device. Fine if it could be handled by a generic CAN hardware filter
interface, which is especially useful for normal CAN controllers.

Wolfgang.