Fwd: Re: Request for feedback for CAN message structure

[Pavel Pisa]

I am forwarding Oliver Hartkopp's founded reply
to the list for archival and rest of the RTEMS
community

----------  Forwarded Message  ----------

Subject: Re: Request for feedback for CAN message structure
Date: Tuesday 21 of June 2022, 08:55:03
From: Oliver Hartkopp <socketcan at hartkopp.net>
To: Pavel Pisa <ppisa4lists at pikron.com>, devel at rtems.org
CC: Prashanth S <fishesprashanth at gmail.com>

Hello all,

I'm definitely with Pavel's suggestions here!

I added some small remarks regarding the current CAN XL discussion in-line:

On 20.06.22 22:57, Pavel Pisa wrote:
> Hello Prashanth S and others,
> 
> excuse me for lower activity last weeks. We have exams period
> and I have lot of other duties and was even ill. I am at Embedded
> World in Nuremberg now, so may it be there is some chance to meet
> somebody other from RTEMS community.
> 
> As for the e-mail it is better to add one of my personal e-mails
> (pisa at cmp.felk.cvut.cz or ppisa at pikron.com)
> to notice me for something important, I do not check every
> message comming through my list subscription. I go through
> RTEMS, NuttX, CAN, etc. forums only when I have spare time.
> 
> On Monday 20 of June 2022 18:37:38 Prashanth S wrote:
>> This is a request for suggestions and feedback for the CAN message
>> structure.
>>
>> *CAN message structure that represents the can messages from application to
>> driver:*
>>
>> struct can_message {
>> uint32_t timestamp;
>> uint32_t id;                 // 32 bits to support extended id (29 bits)
>> uint16_t flags;            // RTR | BRS | EXT ...
>> uint8_t dlc;                 // (0 - 8) | 12 | 16 | 20 | 24 | 32 | 48 | 64
>> uint8_t res;                 // reserved for alignment.
>> uint8_t data[64];         // For supporting data transfer up to 64 bytes.
>> };
>>
>> This is the CAN messages structure created based on the suggestions in the
>> mail chain and looking through other CAN solutions (Nuttx, GRCAN, LinCAN).
> 
> Yes I like solution with explicit and aligned fields.
> I confirm that I think that length representing real data length is better
> and it has been chosen for CAN FD by SocketCAN.

Yes. Handling the DLC/Length conversion in every user program is a mess. 
That's why we actually changed the dlc to len even for Classical CAN 
frames in the struct can_frame:

struct can_frame {
         canid_t can_id;  /* 32 bit CAN_ID + EFF/RTR/ERR flags */
         union {
                 /* CAN frame payload length in byte (0 .. CAN_MAX_DLEN)
                  * was previously named can_dlc so we need to carry that
                  * name for legacy support
                  */
                 __u8 len;
                 __u8 can_dlc; /* deprecated */
         } __attribute__((packed)); /* disable padding added in some ABIs */
         __u8 __pad; /* padding */
         __u8 __res0; /* reserved / padding */
         __u8 len8_dlc; /* optional DLC for 8 byte payload length (9 .. 
15) */
         __u8 data[CAN_MAX_DLEN] __attribute__((aligned(8)));
};

You might also think about a union to carry the Classical CAN length and 
the CAN FD length in one 8 bit value and add the extra Classical CAN raw 
DLC in another 8 bit value - and make a union with a 16 bit length value 
which is needed for CAN XL.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/include/uapi/linux/can.h?id=ea7800565a128c1adafa1791ce80afd6016fe21c

But in fact the raw DLC value has been introduced only because of some 
people doing security testing. The Classical CAN raw DLC value has no 
value for normal CAN applications and therefore using a simple 16 bit 
length value would be the best choice IMHO (see below).

> I confirm that flags implemented with explicit sized field seem to be more
> safe to me. Bitfields are tricky due to endianing and have problem with
> initialization and copying when new field is added.
> 
> I am not sure if the field name "dlc" (data length code) is right
> when the field represents real length, may be "len" is a better
> match.

Yes, definitely!

> I think that fields order makes sense as well (alignment, purpose etc..).
> Only think to consider is possibility to swap
> 
>    uint32_t timestamp;
>    uint32_t id;
> 
> because id is the first and the most important and accessed field,
> so when the pointer to structure is give it can be small win
> on some targets that it can be accessed without offset addition.
> But that is probably neglectable and there can be arguments against.

The 'id' really defines the CAN frame so it can be seen as a 'key' - 
like in a data base. It makes sense to put it first.

Btw. adding RTR and EXT to the flags field is something I would do today 
too. Good choice!

> Another think to consider is a size of "dlc" and or "len" code.
> There is ongoing standardization process for CAN XL and it
> would allow CAN frames up to 2kB and the length will be with
> byte granularity. But CAN XL adds datatype filed and much more
> options so it is probably out of the actual RTEMS scope
> and would require variable length read and write systemcalls
> because 2kB overhead for some 1 or 2 byte message is too high.
> 
> So even 8 bit length field is OK for now and may be res field
> can solve CAN XL compatibility one day.

Why not defining the len as an uint16_t right now?

In fact we (at SocketCAN) need to move the length structure element for 
CAN XL now which is no fun.

As you would have a 16 bit length field and a 16 bit flags field you 
will be CAN XL ready from the beginning.

The only element that needs to be introduced 'before' the data[] is the 
32 bit acceptance field ('AF').

The fact whether AF is present or not could probably be determined by a 
CAN XL flag in the flags element.

The 8 bit virtual CAN ID (vcid) of CAN XL will likely not be visible in 
struct canxl_frame as the vcid will be transported and accessed similar 
to the vlan id in ethernet in Linux. You will need to create a tagged 
network interface analogue to the ethernet vlan handling.

> So generally I agree with the message structure, I declare
> minor weight vote to switch from "dlc" to "len".
> I think that switch to common message structure based API
> on RTEMS would be big win and that (at least for
> now) keeping character device like API with IOCTLs
> is simpler and matches better actual RTEMS CAN implementations.
> 
> Best wishes,
> 
> Pavel

Best regards,
Oliver