[ovs-dev] [PATCHv6 2/3] ofproto-dpif-sflow: Add snaplen for sample action and sFlow.

[William Tu]

Hi Ben,
No significant cost. Since the implementation is simple, let me add it
in next version and we can decide to have it or not.
William

On Fri, Jun 10, 2016 at 8:01 AM, Ben Pfaff <blp at ovn.org> wrote:
> Is there a significant cost to truncating in the userspace datapath, or
> does it just "look weird"?
>
> On Thu, Jun 09, 2016 at 10:20:31PM -0700, William Tu wrote:
>> Hi Ben,
>>
>> Because for sFlow, it doesn't have any benefit to do
>> "sample(truncate(n), userspace(...))" in userspace datapath. I tried
>> to implement it by truncating the packet at OVS_ACTION_ATTR_USERSPACE
>> in dp_execute_cb() but it looks weird. And currently I couldn't think
>> of any other use case so I let sFlow translates differently for kernel
>> and userspace dp.
>>
>> Regards,
>> William
>>
>>
>> On Thu, Jun 9, 2016 at 8:16 PM, Ben Pfaff <blp at ovn.org> wrote:
>> > On Thu, Jun 09, 2016 at 06:06:47PM -0700, William Tu wrote:
>> >> > I'm not sure why CHECK_TRUNC_USERSPACE exists, because I think that your
>> >> > patch implements the new action in userspace.
>> >>
>> >> When translating sflow header config, only if it runs kernel datapath,
>> >> I will program truncate to sample's actions list, ex:
>> >> "sample(trunc(64), userspace(...))". If it runs userspace datapath,
>> >> then it falls back to original way "sample(userspace(...))" and let
>> >> sflow code cut the packet to 'header' size.
>> >
>> > Why are the two datapaths treated differently?  Software is easier to
>> > test if all the cases are similar, when possible.