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

[Andrew Rybchenko]

On 2/1/23 12:14, Jerin Jacob wrote:
> On Wed, Feb 1, 2023 at 2:37 PM Andrew Rybchenko
> <andrew.rybchenko at oktetlabs.ru> wrote:
>>
>> On 2/1/23 12:05, Thomas Monjalon wrote:
>>> 01/02/2023 10:00, Ori Kam:
>>>> Hi all,
>>>>
>>>> Sorry for jumping in late,
>>>>
>>>>> -----Original Message-----
>>>>> From: Jerin Jacob <jerinjacobk at gmail.com>
>>>>> Sent: Wednesday, 1 February 2023 10:53
>>>>>
>>>>> On Wed, Feb 1, 2023 at 12:46 PM Andrew Rybchenko
>>>>> <andrew.rybchenko at oktetlabs.ru> wrote:
>>>>>>
>>>>>> On 2/1/23 09:10, Ivan Malov wrote:
>>>>>>> Hello everyone,
>>>>>>>
>>>>>>> Since making automatic, or implicit, offload decisions does
>>>>>>> not belong in testpmd responsibility domain, it should be
>>>>>>> safer to avoid calling the "negotiate metadata delivery"
>>>>>>> API with some default selection unless the user asks to
>>>>>>> do so explicitly, via internal CLI or app options.
>>>>>>>
>>>>>>> With that in mind, port config <port_id> ... sounds OK.
>>>>>>>
>>>>>>> PMDs that support flow primitives which can generate metadata
>>>>>>> but, if in started state, can't enable its delivery may pass
>>>>>>> appropriate rte_error messages to the user suggesting
>>>>>>> they enable delivery of such metadata from NIC to PMD
>>>>>>> first. This way, if the person operating testpmd
>>>>>>> enters a flow create command and that fails,
>>>>>>> they can figure out the inconsistency, stop
>>>>>>> the port, negotiate, start and try again.
>>>>>>>
>>>>>>> As for non-debug applications, their developers shall
>>>>>>> be properly informed about the problem of enabling
>>>>>>> delivery of metadata from NIC to PMD. This way,
>>>>>>> they will invoke the negotiate API by default
>>>>>>> in their apps, with the feature selection (eg.
>>>>>>> MARK) as per nature of the app's business.
>>>>>>>
>>>>>>> This API should indeed be helpful to some PMDs with
>>>>>>> regard to collecting upfront knowledge like this.
>>>>>>> At the same time, should be benign to those PMDs
>>>>>>> who do not need this knowledge and can enable
>>>>>>> delivery of metadata right when inserting the
>>>>>>> flow rules. So I hope the API does not create
>>>>>>> too much discomfort to vendors not needing it.
>>>>>>>
>>>>>>> Thank you.
>>>>>>>
>>>>>>> On Wed, 1 Feb 2023, Thomas Monjalon wrote:
>>>>>>>
>>>>>>>> 31/01/2023 17:17, Jerin Jacob:
>>>>>>>>> On Fri, Jan 27, 2023 at 8:31 PM Thomas Monjalon
>>>>> <thomas at monjalon.net>
>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>> 27/01/2023 11:42, Nithin Kumar Dabilpuram:
>>>>>>>>>>> From: Thomas Monjalon <thomas at monjalon.net>
>>>>>>>>>>>> 27/01/2023 06:02, Nithin Kumar Dabilpuram:
>>>>>>>>>>>>> From: Thomas Monjalon <thomas at monjalon.net>
>>>>>>>>>>>>>> Ferruh is proposing to have a command "port config <port_id>
>>>>> ..."
>>>>>>>>>>>>>> to configure the flags to negotiate.
>>>>>>>>>>>>>> Are you OK with this approach?
>>>>>>>>>>>>>
>>>>>>>>>>>>> Yes, we are fine to have such command to enable and disable the
>>>>>>>>>>>>> feature
>>>>>>>>>>>>> with default being it disabled if supported by PMD.
>>>>>>>>>>>>> Is default being disabled fine if the feature is supported by a
>>>>>>>>>>>>> PMD ?
>>>>>>>>>>>>
>>>>>>>>>>>> I think the default should be enabled for ease of use.
>>>>>>>>>>>
>>>>>>>>>>> Since testpmd is used extensively for benchmarking purposes, we
>>>>>>>>>>> thought it should have minimum features
>>>>>>>>>>> enabled by default. The default testpmd doesn't have any Rx/Tx
>>>>>>>>>>> offloads enabled(except for FAST FREE),  default
>>>>>>>>>>> fwd mode being "iofwd" and the Rx metadata is only referenced
>>>>> when
>>>>>>>>>>> dumping packets.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>> Do we have similar features disables by default?
>>>>>>>>>>>> I mean do we know features in testpmd which require a "double
>>>>>>>>>>>> enablement"
>>>>>>>>>>>> like one configuration command + one rte_flow rule?
>>>>>>>>>>>
>>>>>>>>>>> Spec itself is that way i.e "RTE_FLOW_RULE +
>>>>>>>>>>> RX_METADATA_NEGOTIATE(once)"
>>>>>>>>>>>
>>>>>>>>>>> Isn't it enough if
>>>>>>>>>>>
>>>>>>>>>>> #1 We have enough print when rte_flow is being create without
>>>>>>>>>>> negotiation done and ask user to enable rx metadata using
>>>>>>>>>>> "port config <port_id>..."
>>>>>>>>>>> #2 Provide testpmd app arg to enable Rx metadata(for example "
>>>>>>>>>>> --rx-metadata") like other features to avoid calling another
>>>>>>>>>>> command before rte flow create.
>>>>>>>>>>
>>>>>>>>>> I'm not sure what is best.
>>>>>>>>>> I will let others discuss this part.
>>>>>>>>>
>>>>>>>>> IMO, enabling something always defeat the purpose to negotiate. IMO,
>>>>>>>>> someone needs to negotiate
>>>>>>>>> for a feature if the feature is needed. I think, the double enablement
>>>>>>>>> is part of the spec itself. In case, The PMD
>>>>>>>>> drivers won't like double enablement, no need to implement the PMD
>>>>>>>>> callback. That way, there is no change in the existing flow.
>>>>>>>>>
>>>>>>>>> The reason why cnxk driver thought of leveraging negotiate() feature
>>>>>>>>> so that it helps for optimization. e.s.p
>>>>>>>>> function template for multiprocess case as we know what features
>>>>>>>>> needed in fastpath upfront.
>>>>>>>>>
>>>>>>>>> If there still concerns with patch we can take up this to TB decide
>>>>>>>>> one way or another to make forward progress. Let us know.
>>>>>>>>
>>>>>>>> Ferruh, Andrew, Ori, Ivan, what is your opinion?
>>>>>>>> Let's progress with this patch to make it in -rc1.
>>>>>>
>>>>>> As I understand all agreed that we need testpmd command to
>>>>>> control negotiated Rx metadata. May be even command-line
>>>>>> option would be useful.
>>>>>>
>>>>>> So, remaining question is what should be the default value in
>>>>>> testpmd. Note that it is just testpmd question since default
>>>>>> value in an abstract application is nothing negotiated
>>>>>> (if I'm not mistaken).
>>>>>>
>>>>>> The key advantaan ge of the current behaviour is to avoid
>>>>>> "double-enabling" in testpmd. It preserves behaviour which
>>>>>> we had before before the API addition. It is a strong
>>>>>> argument. Basically it puts the feature into the same
>>>>>> basket as FAST_FREE - need an action to run faster.
>>>>>
>>>>> I think, there is a disconnect here. FAST_FREE is enabled by default.
>>
>> Sorry, I'm lost here. Don't we need to enable FAST_FREE
>> offload? As far as I know all offloads are disabled by default.
> 
> Not the case for FAST_FREE as disabling needs "more cycles on processor" side.
> 
> See app/test-pmd/testpmd.c
> /*
>   * Ethernet device configuration.
>   */
> struct rte_eth_rxmode rx_mode;
> 
> struct rte_eth_txmode tx_mode = {
>          .offloads = RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE,
> };
> 

Surprised, thanks. So, it one more difference of the testpmd
defaults from an abstract application.

>>
>>>>> i.e We don't need any specific action to run faster. To align with performance
>>>>> test application, by default the configuration should be run faster. User
>>>>> needs to give explicit configuration to allow more offload or the one causes
>>>>> the mpps drops. IMO, That is the theme followed in testpmd.
>>>>>
>>>>>
>>>> I agree with Andrew, the default should stay the same, as now, PMD may already implement
>>>> logic to only enable the feature if there is a flow rule.
>>>> Changing the default will result in breaking applications.
>>>
>>> That's not what is discussed here.
>>> We are talking only about testpmd default.
>>>
>>>> I want to suggest new approach for this feature,
>>>> maybe we can use the rte_flow_configure, and add a new bit that says if those
>>>> actions are going to be used.
>>>> What do you think?
>>>
>>> Let's not change the API please.
>>>
>>>
>>>>>> I see no problem in such approach.
>>>>>>
>>>>>> The key disadvantage is the difference in testpmd and
>>>>>> other applications default behaviour.
>>>>>>
>>>>>> I'd look at the feature in the following way:
>>>>>> if an application theoretically wants to use
>>>>>> USER_FLAG, USER_MARK or TUNNEL_ID it must negotiate
>>>>>> corresponding Rx metadata to ensure that the feature is
>>>>>> available and HW is informed that application may need it.
>>>>>> Since testpmd supports corresponding flow API actions and
>>>>>> flow tunnels, it tries to negotiate it by default, but do
>>>>>> not fail if the negotiation fails.
>>>>>>
>>>>>> So, I'd would vote to keeping the default value as is.

Two above paragraphs still stand.

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.