On 16-03-02 05:22 AM, Jiri Pirko wrote:
> Wed, Mar 02, 2016 at 12:14:39PM CET, gerlitz...@gmail.com wrote:
>> On Tue, Mar 1, 2016 at 7:01 PM, Jiri Pirko <j...@resnulli.us> wrote:
>>> Tue, Mar 01, 2016 at 05:49:27PM CET, a...@vadai.me wrote:
>>>> On Tue, Mar 01, 2016 at 03:47:19PM +0100, Jiri Pirko wrote:
>>>>> Tue, Mar 01, 2016 at 03:24:43PM CET, a...@vadai.me wrote:
>>>>>> This patch is based on a patch made by John Fastabend.
>>
>>>>>> It adds support for offloading cls_flower.
>>>>>> A filter that is offloaded successfully by hardware, will not be added to
>>>>>> the hashtable and won't be processed by software.
>>
>>>>> That is wrong. User should explicitly specify to not include rule into sw
>>>>> by SKIP_KERNEL flag (does not exist now, with John's recent patch we'll
>>>>> have only SKIP_HW). Please add that in this patchset.
>>
>>>> Why? If a rule is offloaded, why would the user want to reprocess it by
>>>> software?
>>
>>>> If the user use SKIP_HW, it will be processed by SW. Else, the user
>>>> would want it to be processed by HW or fallback to SW. I don't
>>>> understand in which case the user would like to have it done twice.
>>
>>> For example if you turn on the offloading by unsetting NETIF_F_HW_TC.
>>> Or if someone inserts skbs into rx path directly, for example pktgen.
>>> We need SKIP_KERNEL to be set by user, not implicit.
>>
>> As discussed in netdev, we want to have three modes for TC offloads
>>
>> 1. SW only
>> 2. HW only (and err if can't)
>> 3. HW and if not supported fallback to SW
>>
>> Now, from your reply, I understand we want a fourth mode
>>
>> 4. Both (HW and SW)
>
> I would perhaps do it a litte bit differently:
> NO FLAG (default)- insert into kernel and HW now:
> if NETIF_F_HW_TC is off (default)
> -> push to kernel only (current behaviour)
> if NETIF_F_HW_TC is on AND push to HW fails
> -> return error
> SKIP_HW - flag to tell kernel not to insert into HW
> SKIP_SW - flag to tell kernel not to insert into kernel
>
> to achieve hw only, user has to turn on the NETIF_F_HW_TC and
> pass SKIP_SW flag.
>
The modes Jiri describes here is exactly how I planned to build
this. And at the moment the only one we are missing is SKIP_HW
which I'm reworking now and should have in a few days.
To resolve the error handling if the rule is SKIP_HW or NO_FLAG
an error will be thrown if it can not be applied to software.
Notice if an error happens on the software insert with NO_FLAG then
the hardware insert is not attempted either. With SKIP_SW I will
throw an error if the hardware insert fails because there is
no software fallback in this mode.
The only mode I haven't looked at doing is
3. HW and if not supported fallback to SW
I'm not sure I have a use case for it at the moment. It is sufficient
for me to just do a SKIP_SW command followed by a SKIP_HW command
if needed. I guess someone else could implement it if they really need
it.
>
>>
>> Do we agree that these four policies/modes make sense?
>>
>> So you want #4 to be the default? can you elaborate why? note that for
>> the HW marking
>> case a "both" mode will be very inefficient, also for actions like
>> vlan push/pop, encap/decap, etc
>
> Well when you push/pop vlan of encap/decap tunnel header in hw, you won't
> match the packet in kernel again. Most likely. But anyway, if you turn
> on NETIF_F_HW_TC you know what you are doing and you adjust the
> included rules (flags) accordingly.
>
>
>> the result will be just wrong... I don't think this should be the default.
>>
>> Or.
