Re: [dpdk-dev] [PATCH 02/12] net/bnxt: fix rxq/txq get information

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 9/18/2020 7:41 PM, Lance Richardson wrote:
> On Fri, Sep 11, 2020 at 10:41 AM Ferruh Yigit <ferruh.yigit@intel.com> wrote:
>>
>>> -     qinfo->conf.rx_drop_en = 0;
>>> +     qinfo->conf.rx_drop_en = 1;
>>
>> Why 0 is wrong but 1 is correct?
>>
>> Technically 'rx_drop_en' is a user configuration, which is set via
>> 'rte_eth_rx_queue_setup()' API.
>>
>> bnxt seems not honoring this config option at all.
>>
>> Based on HW capability, I think two things can be done,
>> 1) Configure the HW based on config request, and return configured value in
>> 'bnxt_rxq_info_get_op()'. see 'ixgbe'.
>>
>> 2) If HW is not configurable, check the value in 'rte_eth_rx_queue_setup()'
>> a) return error if unsupported value requested. see 'sfc'.
>> b) log a warning and overwrite the requested config with whatever supported.
>> And for both a & b, return current config in the 'bnxt_rxq_info_get_op()'
>>
>>>        qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
>>> +     qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
>>
> Hi Ferruh,
> 
> Apologies, this somehow didn't make it to my inbox.
> 
> I believe case (2) applies here, rx_drop_en is not currently configurable in hw,
> so this change was intended to accurately report the effective value. I'm not
> sure whether (2a) or (2b) would be better, but (2b) seems less likely to cause
> issues for existing applications.
 >

I guess (2a) may cause more issues with existing applications, 
applications previously running without problem may now start failing if 
PMD starts returning error.

In (2b) execution will be same, and if PMD already doesn't support NOT 
dropping, functionality will be same, only configuration will reflect 
what is actually happening instead of what user thinks happening.

> 
>>
>> This is for queue specific offloads, you are returning port offloads.
>> As far as I can see bnxt doesn't have any queue specific offload, so this can be
>> dropped.
> 
> It wasn't clear to me whether this was intended to report the difference between
> the offload configuration for the queue and the offload configuration
> for the port
> or the effective offload configuration for the queue. I noticed that
> several other
> PMDs (e.g. mlx5, netvsc, sfc) report the offload configuration for the port in
> rx/tx_queue_info_get(). The sfc PMD reports the offload configuration for the
> port combined with queue-specific offloads, based on those examples this
> seemed to be correct. I guess you're saying those are also incorrect?
> 

May bad, they look OK, "the effective offload configuration for the 
queue" makes more sense here.
I guess I confused same fields should hold the queue specific offloads 
when used for 'rte_eth_rx_queue_setup()'