[EXT] Re: [PATCH v5 2/2] app/testpmd: add command to process Rx metadata negotiation

[Ivan Malov]

Hi Thomas,

On Wed, 1 Feb 2023, Thomas Monjalon wrote:

> 01/02/2023 11:58, Andrew Rybchenko:
>> On 2/1/23 13:48, Jerin Jacob wrote:
>>> On Wed, Feb 1, 2023 at 2:59 PM Andrew Rybchenko
>>> <andrew.rybchenko at oktetlabs.ru> wrote:
>>>> Frankly speaking I don't understand why default value is so
>>>> important if we have a way to change it. Reasons should be
>>>> really strong to change existing defaults.
>>>
>>> The only reason is, typically testpmd will be used performance
>>> benchmarking as an industry standard. It is difficult to tell/educate
>>> the QA or customers
>>> that, "BTW if you need to get better performance add more flag to
>>> testpmd command line".
>
> I disagree.
> When you do performance benchmark, you tune settings accordingly.
>
>>> To make that worst, only some PMD needs to give the additional
>>> parameter to get better number.
>>> And also, testpmd usage will be treated as application modeling.
>>>
>>> Since this feature only used on sfc and cnxk driver, What is the
>>> situation with sfc driver?
>>> Keeping it as negotiated and not use the feature, will impact the per
>>> core performance of sfc or
>>> is it just PCI bandwidth thing which really dont show any difference in testpmd?
>>
>> Yes, sfc could run faster if no Rx metadata are negotiated. So,
>> it is better to negotiate nothing by default. But it is always
>> painful to change defaults. You need to explain that now you
>> need to negotiate Rx metadata to use mark, flag and tunnel offloads.
>> Yes, it will be required on sfc and cnxk only.
>> As an sfc maintainer I don't mind to change testpmd defaults.
>
> If we change testpmd defaults to "do nothing",
> then we should disable MBUF_FAST_FREE as well.

No. These are completely different. MBUF_FAST_FREE is an
optimisation technique, it's in the offload namespace.
Whilst "negotiate metadata" does not offload anything.
It just tells the PMD to enable *delivery* of some
data *when it is present*. And whether it will be
present or not, - this is decided by flow actions.
And, flow actions, in turn, belong in the domain
of decisions made by a specific person operating
the application. So no need to remove FAST_FREE.

>
>
>