Re: [net-next PATCH 3/4] net: sched: cls_u32 add bit to specify software only rules

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

On 16-02-25 03:05 PM, Jamal Hadi Salim wrote:
> On 16-02-25 04:56 PM, John Fastabend wrote:
>> On 16-02-25 04:56 AM, Jamal Hadi Salim wrote:
> 
>>
>> decoding that is not a problem. The ixgbe driver code already applied
>> can decode that without much trouble. The thing I want to avoid is
>> requiring my driver to do the inverse translation because although it
>> is possible its entirely unnecessary.
>>
>> To do the above example without a software cache for example
>> means I read out row 2048 of a hardware table then you get a bunch of
>> bits. From those bits I consult what fields are being loaded into
>> the table in the packet processing pipeline. I learn its the src_ip
>> fields then I have the value/mask and can unwind it. Finally if I
>> collapsed some hash tables onto this hardware table I have to do the
>> inverse of my collapsing scheme. The ixgbe one is sort of simple just
>> because I only have one table in hardware but with multiple tables
>> its a bit more difficult. Finally I've unwound the thing and can
>> print something back out of 'tc' but it seems easier to just cache
>> the hardware rules somewhere. Maybe other driver/hardware will have
>> a different opinion though depending on how much your firmware can
>> store and how ambitious you are. Personally I don't see any need
>> for the above code.
>>
> 
> I think if you can cache the rules and have a way to easily map to
> the hardware then this would work fine.

Yep that is the goal. I think the debate is if its acceptable to do
it with an entirely new filter list ingress_hw Jiri's opinion is that
it would be best to do it inline inside the classifier. At the moment
I'm looking at the code to see if there is a clean way to do it. IMO
using a ingress_hw classifier is a nice solution.

In the meantime I just respun patches 1-3 with the feedback and will
submit those while I work out patch 4.

> 
> cheers,
> jamal