Re: [dpdk-dev] [PATCH 1/2] ethdev: add flow action type update as an offload

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 10/31/19 5:49 PM, Thomas Monjalon wrote:
> 31/10/2019 10:49, Andrew Rybchenko:
>> On 10/28/19 5:00 PM, Ori Kam wrote:
>>>> -----Original Message-----
>>> From: Andrew Rybchenko <arybchenko@solarflare.com>
>>>> On 10/28/19 1:50 PM, Ori Kam wrote:
>>>>> Hi Pavan,
>>>>>
>>>>> Sorry for jumping in late.
>>>>>
>>>>> I don't understand why we need this feature. If the user didn't set any flow
>>>> with MARK
>>>>> then the user doesn't need to check it.
>>>> There is pretty long discussion on the topic already, please, read [1].
>>>>
>>>> [1]
>>>> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Finbox.dpdk
>>>> .org%2Fdev%2F3251fc00-7598-1c4f-fc2a-
>>>> 380065f0a435%40solarflare.com%2F&amp;data=02%7C01%7Corika%40mellan
>>>> ox.com%7Ce3f779d4b7c44b682d6508d75b9d8688%7Ca652971c7d2e4d9ba6a4
>>>> d149256f461b%7C0%7C0%7C637078604439019114&amp;sdata=sYooc%2FQ3C
>>>> kUZG3gRFPlZrm8xMfMB9gOWWex5YIkWhMc%3D&amp;reserved=0
>>>>
>>> Thanks for the link, it was an interesting reading.
>>>
>>>>> Also it breaks compatibility.
>>>> Yes, there is a deprecation notice for it.
>>>>
>>>>> If my understanding is correct the MARK field is going to be moved to
>>>> dynamic field, and this
>>>>> will be way to control the use of MARK.
>>>> Yes and I think the offload should used to request dynamic
>>>> field register. Similar to timestamp in dynamic mbuf examples.
>>>> Application requests Rx timestamp offload, PMD registers dynamic
>>>> filed.
>>>>
>>> In general it was decided that there will be no capability for rte_flow API, due to the fact that
>>> it is impossible to support all possible combinations. For example a PMD can allow mark on Rx
>>> while not supporting it on e-switch (transfer) or on Tx.
>>> The only way to validate it is validating a flow. If the flow is validated then the action is supported.
>>> This is the exact approach we are implementing with the Meta feature.
>>> So as I see it, the logic should be something like this:
>>> 1. run devconfigure.
>>> 2. allocate mempool
>>> 3. setup queues.
>>> 4. run rte_flow_validate with mark action.
>>> If flow validated register mark in mbuf else don't register.
>>> If the PMD needs some special setting for mark he can update the queue when he gets the flow to validate.
>>> At this stage the device is not started so any change is allowed.
>> I understand why there is capability reporting in rte_flow API when
>> it is about rte_flow API itself. The problem appears when rte_flow
>> API starts to interact with other functionality.
>> Which pattern/actions should application try in order to decide
>> if MARK is supported or not.
> Why application should decide whether MARK is supported or not?
> In my understanding it can be enabled dynamically per flow.

Yes, it is per flow right now, but it is resource consuming to
make a flow rule just to discard it and work without offload.
The application already suffers and attempt to use hardware
offload makes it suffer even more. Of course, hardware offload
in application may be simply globall disabled, but presence of
MARK offload allows to do it dynamically based on offload
reported by PMD.

Also I think that Qi has a good example for vPMD why
MARK offload would be useful.

>> The right answer is a pattern/action
>> which will be really used, but what to do if there are many
>> combinations or if these combinations are not know in advance.
>> Minimal? But I easily imagine cases when minimal is not supported,
>> but more complex real life patterns are supported.
>>
>> The main idea behind the offload is as much as you know in advance
>> as much you can optimize without overcomplicating drivers and HW.
>>
>> In the case of OVS, absence MARK offload would mean that OVS
>> should not even try to use partial offload even if it is enabled.
>> So, no efforts are required to try to convert flow into pattern and
>> validate the flow rule.
> That's an interesting feedback.
> I would like to understand why OVS cannot adapt its datapath on demand
> per port, per queue and per flow?

I guess there is a misunderstanding here. What I'm trying to say
is that introduction of MARK offload would make code a bit more
simple and efficient. Basically it would be possible to enable
so-called hardware offload in OVS by default, but finally make
a decision per port based on MARK offload availability
(should it try to make rte_flow rule by flow and insert it?)