Question about add ethdev loopback set API

[fengchengwen]

On 2022/12/14 18:38, Ferruh Yigit wrote:
> On 12/14/2022 7:25 AM, fengchengwen wrote:
>> On 2022/12/13 19:25, Ferruh Yigit wrote:
>>> On 12/13/2022 10:04 AM, fengchengwen wrote:
>>>> Hi Ferruh,
>>>>
>>>>     During the test, we need to delineate where go wrong when encountered
>>>> e.g. CRC error. In this scenario, loopback is useful.
>>>>
>>>>     I think we can add a loopback set API which could set inner or outer loop,
>>>> and user can use telemetry to set the loopback in the above scenario.
>>>>
>>>>     I'd like to hear your opinion about add a loopback set API.
>>>>
>>>
>>> Hi Chengwen,
>>>
>>> Is the intention to test ethdev layer or driver?
>>>
>>> It is possible to use ring vdev to create a loopback and to test ethdev
>>> layer.
>>>
>>> For driver, it can be possible to create physical loopback connection,
>>> or even can implement loopback Rx/Tx burst functions in driver.
>>> Using another host to send/receive packets to DUT (device under test) is
>>> another approach.
>>>
>>>
>>> What kind of loopback implementation do you have in your mind?
>>
>> Mainly MAC layer and lower physical layer:
>>
>>    --------   ---------------        ------------        ----------                --------------------
>>    |      |   |        - rx |        | -  rx  - |        | - rx - |                |                  |
>>    | Host | - |   MAC       |   -    |  SerDes  |   -    |  PHY   |        ====    | Packet Generator |
>>    |      |   |        - tx |        | -  tx  - |        | - tx - |                |                  |
>>    --------   ---------------        ------------        ----------                --------------------
>>
>> The support loopback in hns3 platform:
>>    Inner loopback subtypes: which host send pkts and recv and then verify:
>>         Serdes tx to rx
>>         PHY tx to rx
>>
>>    Outer loopback subtypes: which Packet-Generator send pkts and recv and then verify:
>>         MAC tx to rx
>>
>> I think we could support the above loopback types, and maybe other PMD platform support
>> more loopback types.
>>
> 
> There is a 'lpbk_mode' field of "struct rte_eth_conf", to configure
> loopback in 'rte_eth_dev_configure()',
> but it isn't fine grained to define the possible modes you mentioned
> above. Currently '0' means loopback disabled and non zero means it is
> enabled, details left to drivers.
> 
> 'lpbk_mode' is 32bit, we have space to define detailed loopback modes.

Emm, I found the lpbk_mode is not defined in a unified manner, which is vendor specified.

> 
> 
> Having loopback configuration in 'rte_eth_dev_configure()' requires port
> to stop and reconfigure to enable/disable loopback.
> If it is possible to change loopback behavior without full
> reconfiguration cycle, we can add two new APIs to enable/disable
> loopback. But this has the downside of two different APIs change same
> config, and we had problems related this in the past.

Yes, the 'use rte_eth_dev_configure() to config loopback' is inflexible, and I also notice
the testpmd command "set tx loopback port-id on/off", and it invoke PMD's public API which
is rte_pmd_ixgbe/i40e/bnxt/dpaa_set_tx_loopback. According to this information, loopback
needs to be configured during running.

So I suggest add one API:

int rte_eth_dev_set_loopback(uint16_t port_id, uint32_t lpbk_mode), and costraint
this API can only invoke when port is started, if turn off (lpbk_mode==0) then should
recovering rte_eth_conf's lpbk_mode.

Also the lpbk_mode is vendor specified.

> 
> 
> 
> Also there is a hairpin feature in ethdev, that connects Rx queue back
> to Tx queue, but not sure if that justifies your "MAC tx ro rx" testcase.
> 
> 
> .
>