[dpdk-dev] [PATCH] mlx4: use dummy rxqs when a non-pow2 number is requested

[Olivier Matz]

Hi,

On 03/22/2016 03:27 PM, Wiles, Keith wrote:
> Hi Olivier,
>
>> Hi Keith,
>>
>> On 03/21/2016 06:38 PM, Wiles, Keith wrote:
>>>> On Mar 21, 2016, at 11:10 AM, Olivier Matz <olivier.matz at 6wind.com> wrote:
>>>>
>>>> When using RSS, the number of rxqs has to be a power of two.
>>>> This is a problem because there is no API is dpdk that makes
>>>> the application aware of that.
>>>>
>>>> A good compromise is to allow the application to request a
>>>> number of rxqs that is not a power of 2, but having inactive
>>>> queues that will never receive packets. In this configuration,
>>>> a warning will be issued to users to let them know that
>>>> this is not an optimal configuration.
>>>
>>> Not sure I like this solution. I think an error should be returned with a log message instead. What if the next driver needs power of three or must be odd or even number.
>>>
>>> The bigger problem is the application is no longer portable for any given nic configuration.
>>>
>>> We need a method for the application to query the system for these types of information. But as we do not have that API we need to just error the request off.
>>
>>
>> The initial problem is that the driver says "I support a maximum
>> of X queues" and if the application configures a lower number, it
>> gets an error.
>>
>> There is no API in DPDK to tell that only specific number of queues
>> are supported. Adding an API is a solution, but in this case it's
>> probably overkill. With this patch, the driver can present the proper
>> number of queues to the application, knowing that the spreading of
>> the packets won't be ideal (some queues won't receive packets), but
>> it will work.
>>
>> A step further in this direction would be to configure more queues
>> than asked in hardware to do a better spreading, almost similar to
>> what is done with RETA tables in mlx5. But this is more complicated
>> to do, especially if we want it for 16.04.
>
> Well I guess I must agree with the solution, but I am not real happy. Can we mark this a temp fix until we figure out a cleaner solution as I would not want this type of solution forever or be the standard way to handle these problems.

Back on this issue, I agree that a cleaner solution may be needed,
probably in the ethdev API. I did a quick look in the drivers directory
and, from what I remember, vmxnet3 also need to have a number of rxq
and txq to be a power of 2.

>From what I see in the code, if an application tries to configure a
number of queue which is not a power of 2 on vmxnet3, the driver will
fail to start without any log.

Yong, do you feel a patch similar to what was done on mlx4 is
feasable/suitable in vmxnet3 driver? Shouldn't at least have some
error logs saying that the number of rxq/txq is invalid?

It cleanup or rework is planned in the ethdev API for 16.11, maybe
this is a problem that should be addressed.


Regards,
Olivier