[dpdk-dev] [PATCH 1/2] app/testpmd: update Rx RSS HASH offload when setting MQ RSS

[Wei Hu (Xavier)]

Hi, Ferruh Yigit

On 2020/9/23 17:35, Ferruh Yigit wrote:
> On 9/23/2020 8:04 AM, Wei Hu (Xavier) wrote:
>> Hi, Ferruh Yigit
>>
>> On 2020/9/23 0:21, Ferruh Yigit wrote:
>>> On 9/8/2020 3:16 AM, Wei Hu (Xavier) wrote:
>>>> From: Huisong Li <lihuisong at huawei.com>
>>>>
>>>> Currently, when starting testpmd application without 
>>>> '--disable-rss' and
>>>> the number of Rx queue configured is greater than 1, ETH_MQ_RX_RSS 
>>>> flag
>>>> is set in port->dev_conf.rxmode.mq_mode in testpmd application, and
>>>> DEV_RX_OFFLOAD_RSS_HASH flag is set in rx_offloads
>>>> (dev->data->dev_conf.rxmode.offloads) according to the ETH_MQ_RX_RSS
>>>> flag of rxmode.mq_mode in PMD drivers.
>>>>
>>>> However, DEV_RX_OFFLOAD_RSS_HASH is not set to rx_offloads maintained
>>>> in testpmd application, this will cause the inconsistent problem that
>>>> rx_offloads is different for testpmd and PMD drivers.
>>>
>>> Yes for DEV_RX_OFFLOAD_RSS_HASH, application rx_offload config and 
>>> PMD one diverges, for *some* PMDs.
>>>
>>> This is done to have the backward compatibility of the PMD behavior.
>>>
>>> The PMDs that would like to write the provide the calculated hash 
>>> value back, overwrites the offload config to enable 
>>> 'DEV_RX_OFFLOAD_RSS_HASH'. And does more work than user requested, 
>>> this shouldn't have any side affect.
>>> Some doesn't provide the hash unless user explicitly requests it.
>>>
>>> Applications shouldn't set it blindly, at least first check if PMD 
>>> supports it but even that case, unless it is needed I think HASH 
>>> offload shouldn't be requested by default.
>>
>> OK, we are going to do modification to add check if PMD support 
>> DEV_RX_OFFLOAD_RSS_HASH before setting it as below:
>>
>>   @@ -3356,11 +3356,13 @@ init_port_config(void)
>>             }
>>             if (port->dcb_flag == 0) {
>>                if (port->dev_conf.rx_adv_conf.rss_conf.rss_hf != 0) {
>>                     port->dev_conf.rxmode.mq_mode =
>>                         (enum rte_eth_rx_mq_mode)
>>                             (rx_mq_mode & ETH_MQ_RX_RSS);
>>             if (port->dev_info.rx_offload_capa &
>>                DEV_RX_OFFLOAD_RSS_HASH)
>>               port->dev_conf.rxmode.offloads |=
>>                          DEV_RX_OFFLOAD_RSS_HASH;
>>               } else
>>                     port->dev_conf.rxmode.mq_mode = ETH_MQ_RX_NONE;
>>             }
>>
>
> Hi Xavier,
>
> The capability check is correct thing to do, but even in that case I 
> am not sure if we should set the config by default.
>
> The reason to extract the 'DEV_RX_OFFLOAD_RSS_HASH' offload is to gain 
> performance for some NICs, enabling it by default defeats the purpose.
>
> The offload should be enabled when application needs the provided hash 
> value.
>
> I can see you are trying to remove the divergence between PMD and 
> application config, this can be fixed when all PMDs take this user 
> offload request into account instead of overwriting, but for some PMDs 
> it is cheaper to provide the hash value instead of checks, so this 
> divergence is not easy to address.
>
OK, please ignore this patch and the V2 patch.

http://patches.dpdk.org/patch/78502/

Thanks.


Regards

Xavier