[dpdk-dev] i40e and RSS woes

Vlad Zolotarov vladz at cloudius-systems.com
Mon Aug 24 13:40:16 CEST 2015



On 08/24/15 14:13, Vlad Zolotarov wrote:
>
>
> On 03/05/15 07:56, Zhang, Helin wrote:
>> Hi Gleb
>>
>> Sorry for late! I am struggling on my tasks for the following DPDK 
>> release these days.
>>
>>> -----Original Message-----
>>> From: Gleb Natapov [mailto:gleb at cloudius-systems.com]
>>> Sent: Monday, March 2, 2015 8:56 PM
>>> To: dev at dpdk.org
>>> Cc: Zhang, Helin
>>> Subject: Re: i40e and RSS woes
>>>
>>> Ping.
>>>
>>> On Thu, Feb 19, 2015 at 04:50:10PM +0200, Gleb Natapov wrote:
>>>> CCing i40e driver author in a hope to get an answer.
>>>>
>>>> On Mon, Feb 16, 2015 at 03:36:54PM +0200, Gleb Natapov wrote:
>>>>> I have an application that works reasonably well with ixgbe driver,
>>>>> but when I try to use it with i40e I encounter various RSS related 
>>>>> issues.
>>>>>
>>>>> First one is that for some reason i40e, when it builds default reta
>>>>> table, round down number of queues to power of two. Why is this? If
>> It seems because of i40e queue configuration. We will check it more 
>> and see
>> if it can be changed or improved later.
>
> Helin, hi!
> Sorry for bringing it back but it seems that the RSS queues number 
> issue (rounding it down to the nearest power of 2)
> still hasn't been addressed in the master branch.
>
> Could u, pls., clarify what is that "i40e queue configuration" that 
> requires this alignment u are referring above?
>
> From what i could see "num" parameter is not propagated outside the 
> i40e_pf_config_rss() in any form except for RSS table contents.
> This means that any code that would need to know the number of Rx 
> queues would use the dev_data->nb_rx_queues (e.g. i40e_dev_rx_init())
> and wouldn't be able to know that i40e_pf_config_rss() something 
> different except for scanning the RSS table in HW which is of course 
> not an option.
>
> Therefore, from the first look it seems that this rounding may be 
> safely removed unless I've missed something.
>
> Pls., comment.

Have just noticed this:

	/* Each of below queue pairs should be power of 2 since it's the
	   precondition after TC configuration applied */
	uint16_t lan_nb_qps; /* The number of queue pairs of LAN */

I still couldn't find any justification for either requiring the above 
or requiring any correlation between the number of Tx queues (whatever 
it is) and the
number of Rx queues in the HW spec. It seems that spec implies that Rx 
and Tx configuration is completely orthogonal (as it should be).

Could u, pls., clarify how TC configuration imposes a requirement on a 
number of Rx queues to be a power of 2?

thanks in advance,
vlad

>
> thanks,
> vlad
>
>>
>>>>> I configure reta by my own using all of the queues everything seams
>>>>> to be working. To add insult to injury I do not get any errors
>>>>> during configuration some queues just do not receive any traffic.
>>>>>
>>>>> The second problem is that for some reason i40e does not use 40 byte
>>>>> toeplitz hash key like any other driver, but it expects the key to
>>>>> be 52 bytes. And it would have being fine (if we ignore the fact
>>>>> that it contradicts MS spec), but how my high level code suppose 
>>>>> to know
>>> that?
>> Actually a rss_key_len was introduced in struct rte_eth_rss_conf 
>> recently. So the
>> length should be indicated clearly. But I found the annotations of 
>> that structure
>> should have been reworked. I will try to rework it with clear 
>> descriptions.
>>
>>>>> And again, device configuration does not fail when wrong key length
>>>>> is provided, it just uses some other key. Guys this kind of error
>>>>> handling is completely unacceptable.
>> If less length of key is provided, it will not be used at all, the 
>> default key will be used.
>> So there is no issue as you said. But we need to add more clear 
>> description for the
>> structure of rte_eth_rss_conf.
>>
>> Thank you very much for the good comments!
>>
>> Regards,
>> Helin
>>
>>>>> The last one is more of a question. Why interface to change RSS hash
>>>>> function (XOR or toeplitz) is part of a filter configuration and not
>>>>> rss config?
>>>>>
>>>>> -- 
>>>>>             Gleb.
>>>> -- 
>>>>             Gleb.
>>> -- 
>>>             Gleb.
>



More information about the dev mailing list