Re: [net-next PATCH v1 04/11] rocker: add pipeline model for rocker switch

[John Fastabend <john.fastabend@gmail.com>]

On 01/06/2015 09:16 AM, Scott Feldman wrote:
> On Tue, Jan 6, 2015 at 9:00 AM, John Fastabend <john.fastabend@gmail.com> wrote:
>> On 01/05/2015 11:01 PM, Scott Feldman wrote:
>>>> +
>>>> +struct net_flow_jump_table parse_ethernet[3] = {
>>>> +       {
>>>> +               .field = {
>>>> +                  .header = HEADER_ETHERNET,
>>>> +                  .field = HEADER_ETHERNET_ETHERTYPE,
>>>> +                  .type = NET_FLOW_FIELD_REF_ATTR_TYPE_U16,
>>>> +                  .value_u16 = 0x0800,
>
> ETH_P_IP, etc
>
>>>
>>>
>>> How is htons/ntohs conversions happening here?
>>
>>
>> my current stance is to leave everything in host order in the model
>> and let the drivers do conversions as needed. For example some drivers
>> want the vlan vid in host order others network order. I think its
>> more readable above then with hton*() throughout.
>
> Hmmm...I would argue adding htons/htonl makes it more readable in the
> sense that it's a reminder that this is a field in a network header,
> to be used for matching against packet headers, which use
> network-ordering.  Store the field in the order best for comparison
> with the raw pkt data.  Drivers may still need to do some conversion
> if the field is programmed in hardware in a diff order.
>

Easy enough here, but then when we set_flows what do we use
network-ordering or host? If it can be network-order in some
cases and host-order in others its hard to resolve pragmatically.
Humans at a CLI can most likely get it right for well known fields
such as VLAN IDs but for less common fields (maybe proprietary)
or management software it gets tricky.

I guess we could add a flag to indicate byte ordering.

-- 
John Fastabend         Intel Corporation