[dpdk-dev] [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and Tx offload reconfig cmd

Ferruh Yigit ferruh.yigit at intel.com
Tue Apr 20 10:22:17 CEST 2021


On 4/20/2021 9:12 AM, Yu, DapengX wrote:
> 
> 
>> -----Original Message-----
>> From: Yigit, Ferruh <ferruh.yigit at intel.com>
>> Sent: Monday, April 19, 2021 11:46 PM
>> To: Yu, DapengX <dapengx.yu at intel.com>; Li, Xiaoyun
>> <xiaoyun.li at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>> Cc: dev at dpdk.org; stable at dpdk.org
>> Subject: Re: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and Tx offload
>> reconfig cmd
>>
>> On 4/15/2021 7:04 AM, Yu, DapengX wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Ferruh Yigit <ferruh.yigit at intel.com>
>>>> Sent: Monday, April 12, 2021 8:46 PM
>>>> To: Li, Xiaoyun <xiaoyun.li at intel.com>; Yu, DapengX
>>>> <dapengx.yu at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>>>> Cc: dev at dpdk.org; stable at dpdk.org
>>>> Subject: Re: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and Tx
>>>> offload reconfig cmd
>>>>
>>>> On 4/12/2021 3:21 AM, Li, Xiaoyun wrote:
>>>>>
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: Yu, DapengX <dapengx.yu at intel.com>
>>>>>> Sent: Friday, April 9, 2021 18:29
>>>>>> To: Li, Xiaoyun <xiaoyun.li at intel.com>; Yigit, Ferruh
>>>>>> <ferruh.yigit at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>>>>>> Cc: dev at dpdk.org; stable at dpdk.org
>>>>>> Subject: RE: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and Tx
>>>>>> offload reconfig cmd
>>>>>>
>>>>>>
>>>>>>
>>>>>>> -----Original Message-----
>>>>>>> From: Li, Xiaoyun <xiaoyun.li at intel.com>
>>>>>>> Sent: Friday, April 9, 2021 3:50 PM
>>>>>>> To: Yu, DapengX <dapengx.yu at intel.com>; Yigit, Ferruh
>>>>>>> <ferruh.yigit at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>>>>>>> Cc: dev at dpdk.org; stable at dpdk.org
>>>>>>> Subject: RE: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and
>>>>>>> Tx offload reconfig cmd
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> -----Original Message-----
>>>>>>>> From: Yu, DapengX <dapengx.yu at intel.com>
>>>>>>>> Sent: Friday, April 9, 2021 13:25
>>>>>>>> To: Yigit, Ferruh <ferruh.yigit at intel.com>; Li, Xiaoyun
>>>>>>>> <xiaoyun.li at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>>>>>>>> Cc: dev at dpdk.org; stable at dpdk.org
>>>>>>>> Subject: RE: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and
>>>>>>>> Tx offload reconfig cmd
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>> -----Original Message-----
>>>>>>>>> From: Yigit, Ferruh <ferruh.yigit at intel.com>
>>>>>>>>> Sent: Thursday, April 8, 2021 11:42 PM
>>>>>>>>> To: Yu, DapengX <dapengx.yu at intel.com>; Li, Xiaoyun
>>>>>>>>> <xiaoyun.li at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
>>>>>>>>> Cc: dev at dpdk.org; stable at dpdk.org
>>>>>>>>> Subject: Re: [dpdk-stable] [PATCH] app/testpmd: fix queue Rx and
>>>>>>>>> Tx offload reconfig cmd
>>>>>>>>>
>>>>>>>>> On 4/1/2021 9:28 AM, dapengx.yu at intel.com wrote:
>>>>>>>>>> From: Dapeng Yu <dapengx.yu at intel.com>
>>>>>>>>>>
>>>>>>>>>> Configure per queue rx offloading and per queue tx offloading
>>>>>>>>>> command shouldn't trigger the rte_eth_dev_configure() to
>>>>>>>>>> reconfigure
>>>>>>>> device.
>>>>>>>>>>
>>>>>>>>>> The patch sets the queue reconfiguration flag only, and does
>>>>>>>>>> not set the device reconfiguration flag. Therefore after port
>>>>>>>>>> is restarted,
>>>>>>>>>> rte_eth_dev_configure() will not be called again.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Just to clarify the impact, was calling 'rte_eth_dev_configure()'
>>>>>>>>> causing any problem, is this fixing any issue?
>>>>>>>>> Or is this patch an optimization to eliminate an unnecessary call?
>>>>>>>>>
>>>>>>>> This patch does fix an issue, and it also eliminates an unnecessary call.
>>>>>>>>
>>>>>>>> The issue is:
>>>>>>>> per-queue configuration, for example: port 0 rxq 0 rx_offload
>>>>>>>> jumbo_frame off triggers the per-device configuration change: the
>>>>>>>> RSS key is reconfigured and changes after rte_eth_dev_configure()
>>>>>>>> is called on ICE PMD driver, that cause a test case failure.
>>>>>>>>
>>>>>>>
>>>>>>> Hmmm. I agree on the following. It doesn't need dev_configure.
>>>>>>> That's why I give you ack.
>>>>>>>
>>>>>>> But your issue. Shouldn't you fix the driver? The vsi->rss_key[]
>>>>>>> just updated itself as a random value when dev_conf doesn't
>>>>>>> contain
>>>> one.
>>>>>>> But your case is dev_conf doesn't contain new rss key, but
>>>>>>> vsi->rss_key is not null. In this case, it should keep the same
>>>>>>> vsi->not get a new
>>>>>> random key.
>>>>>>>
>>>>>> In current implementation, dev->data-
>>> dev_conf.rx_adv_conf.rss_conf
>>>>>> in PMD does not hold the rss_key value if user does not set it in
>>>>>> the input parameter of rte_eth_dev_configure().
>>>>>>
>>>>>> Even if PMD generate one automatically, it will not be saved in
>>>>>> dev->data-
>>>>>>> dev_conf.rx_adv_conf.rss_conf.
>>>>>>
>>>>>> When user want to get rss_key, it will be retrieved on the fly from
>>>>>> hardware, but not from any variable in PMD.
>>>>>>
>>>>>> So PMD (ice and i40e) think only rss_key  which is set by user via
>>>>>> rte_eth_dev_configure() can be reused when port is reconfigured.
>>>>>>
>>>>>> If it is not present, PMD will generate one by itself anyway even
>>>>>> if it is present in
>>>>>> vsi->rss_key.
>>>>>>
>>>>>> I don’t think this behavior is wrong, so did not fix ice PMD.
>>>>>
>>>>> If you want to recover the rss key. The driver should store the key
>>>>> in vsi- rss_key also in ice_rss_hash_update(). Then when user not
>>>>> config rss_key
>>>> in rss_conf and vs->rss_key is not 0, you should set the original
>>>> value not a random value to hw.
>>>>>
>>>>
>>>> +1
>>>
>>> The background of the patch is:
>>> A test case expects that RSS feature makes the same packet flow into
>>> same queue, after queue is reconfigured or device is reconfigured,
>>> even if user does not use a RSS key to configure a device in advance.
>>> The test case is actually over testing, and tester has agreed to
>>> modify test plan to make it pass. If we only care about the test case,
>>> testpmd and ice PMD does not need any patch.
>>>
>>> But tester found a hidden problem when run the above case:
>>> Configure per queue rx offloading and per queue tx offloading command
>>> trigger the rte_eth_dev_configure() to reconfigure device.
>>>
>>> The problem for the above test case can be resolved by setting RSS key in
>> advance.
>>> And looks like no imminent risk to deprecate the patch.
>>> But it does not mean that the per queue command triggering device
>> reconfiguration is correct.
>>> So leave behind the test case, the patch is still reasonable.
>>>
>>> @Ferruh, if there is still concern for this patch, please comments on it,
>> thanks!
>>> If there is no objection, I will change the patch back to new state, and move
>> on.
>>>
>>
>> Hi Dapeng,
>>
>> The patch is reasonable, that is why Xiaoyun ack'ed it at first place.
>>
>> But later we have find out that the change has potential to hide some other
>> driver error, so I think it make sense to wait for the driver fix fist, to be sure
>> this change won't hide that issue, later we can proceed with this patch.
>>
>> Calling an API redundantly on the control path is not so bad, we can live with
>> it until other issue solved.
>>
> 
> @Ferruh, thank you for sparing time to answer.
> The previous comments mentioned a controversy that there is error in ICE PMD driver(Random RSS key generation on device reconfiguration),
> Actually this is not bug, and it has been confirmed with maintainer(Zhang Qi), there is no need to fix it.

Wasn't the problem that device get a new random RSS key on each call of the 
configure API, is this expected behavior?

And will the RSS key change even if the user set the RSS key between two 
configure calls?

> Are there other driver errors which this patch may hide? If not, I think we can proceed with this patch.
> 
>>>>
>>>>> Otherwise, you are assuming the behavior rss to different queues is
>>>>> right
>>>> since user want random key. It's not an issue.
>>>>>
>>>>> Any scenario, it's not testpmd issue. Please withdraw your patch.
>>>>>
>>>>>>
>>>>>>>> There is an unnecessary call in original implementation because
>>>>>>>> both
>>>>>>>> cmd_config_per_queue_rx_offload_parsed() and
>>>>>>>> cmd_config_per_queue_tx_offload_parsed()
>>>>>>>> does not update the "port->dev_conf" which hold the port
>>>>>>>> configuration, therefore there is no need to call
>>>> rte_eth_dev_configure().
>>>>>
>>>
> 



More information about the dev mailing list