Re: [dpdk-dev] [PATCH] librte_ethdev: extend dpdk api led control to query capability

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 1/8/2020 1:58 PM, Laurent Hardy wrote:
> 
> On 1/8/20 2:06 PM, Thomas Monjalon wrote:
>> 08/01/2020 13:59, Ferruh Yigit:
>>> On 1/8/2020 10:31 AM, Laurent Hardy wrote:
>>>> Hi all,
>>>>
>>>> On 1/8/20 10:55 AM, David Marchand wrote:
>>>>> On Wed, Jan 8, 2020 at 10:09 AM Ferruh Yigit <ferruh.yigit@intel.com> wrote:
>>>>>> On 1/8/2020 8:56 AM, David Marchand wrote:
>>>>>>> Hello Laurent,
>>>>>>>
>>>>>>> Bonne année.
>>>>>>>
>>>>>>> Cc: maintainers.
>>>>>>>
>>>>>>> On Tue, Jan 7, 2020 at 3:57 PM Laurent Hardy <laurent.hardy@6wind.com> wrote:
>>>>>>>> In current led control API we have no way to know if a device is able
>>>>>>>> to handle on/off requests coming from the application.
>>>>>>>> Knowing if the device is led control capable could be useful to avoid
>>>>>>>> exchanges between application and kernel.
>>>>>>>> Using the on/off requests to flag if the device is led control capable
>>>>>>>> (based on the ENOSUP returned error) is not convenient as such request
>>>>>>>> can change the led state on device.
>>>>>>>>
>>>>>>>> This patch adds a new function rte_eth_led_ctrl_capable() that will look
>>>>>>>> for led_off/on dev ops availability on the related pmd, to know if the
>>>>>>>> device is able to handle such led control requests (on/off).
>>>>>>> This patch breaks the ABI, which is BAD :-).
>>>>>> Why it is an ABI break, dev_ops should be between library and drivers, so it
>>>>>> should be out of the ABI concern, isn't it.
>>>>> You are right.
>>>>> So in our context, this is not an ABI breakage.
>>>>> But abidiff still reports it, so maybe some filtering is required to
>>>>> avoid this false positive.
>>>>>
>>>>> Note that if we insert an ops before rx_queue_count, we would have a
>>>>> real ABI breakage, as this ops is accessed via an inline wrapper by
>>>>> applications.
>>>>>
>>>>>
>>>>>>> This new api only needs to look at the existing ops, so you can remove
>>>>>>> the (unused in your patch) dev_led_ctrl_capable ops.
>>>>>>>
>>>>>>> OTOH, would it make sense to expose this capability in dev_flags?
>>>>>>>
>>>>>> 'rte_eth_led_on()' & 'rte_eth_led_off()' APIs returns '-ENOTSUP' when the not
>>>>>> supported, can that help application to understand?
>>>>> You might want to know it is supported without changing the state.
>>>>> Laurent?
>>>> First, happy new year :)
>>>>
>>>> Yes exactly, the purpose of this patch is to query if the device is led
>>>> control capable or not without changing the led state.
>>>>
>>>> About exposing the capability through a dev_flags, means to make some
>>>> modification in each pmds. It looks more easy in term of pmds
>>>> maintenance to relying on the rte_eth_led_off()/on() dev ops
>>>> availability at rte_ethdev level, right ?
>>>>
>>> 'dev_flag' definition is not clear, right now it holds the combination of status
>>> and capability. And we have 'rte_eth_dev_info' struct, which is again
>>> combination of device capability and status.
>> I agree capabilities in ethdev are a bit of a mess.
>> I would appreciate someone makes a complete audit of it
>> so we can discuss how to improve the situation.
>>
>>
>>> Perhaps we should have explicit capabilities and status fields, even in the
>>> rte_device level which inherited by net/crypto devices etc..
>> No, ethdev capabilities should stay in ethdev.
>>
>>
>>> But for dev_ops, instead of having another capabilities indicator, which
>>> requires PMDs to keep this synchronized, I think it is better if we can self
>>> contain this information within dev_ops, like not implementing dev_ops would
>>> mean it is not supported, this way it is easier to maintain and less error prone.
>> It means the dev_ops is resetted at init if a device does not support the feature.
>> It is against having const dev_ops.
>>
>>
>>> Only we should have it without side effect,
>>>
>>> 1- adding an additional 'dry-run' parameter can work, but this means breaking
>>> ABI and updating majority of the ethdev APIs :)
>>> 2- Adding 'is_supported' versions of the APIs as we need can be an option, like
>>> 'rte_eth_led_on_is_supported()'
>>> 3- Olivier's suggestion to add a new API to get the led status, so that this
>>> information can be used select led API which won't cause side affect and let us
>>> learn if it is supported.
>>>
>>> Any other alternatives?
>>>
>>> I would prefer the 2) in above ones, which is very similar to the original patch.
> 
> I can provide a V2 which will remove the useless dev_led_ctrl_capable ops.

+1, dev_led_ctrl_capable is not used.

> 
> About the 'is_supported()' versions of APIs, in the current patch I 
> factorize
> the check on dev ops on and off availability in a same function named
> "led_ctrl_capable" but I can rename it if required.
> 
> Just in this specific case I don't dissociate on and off capability, as 
> being
> able to set the led off without a way to set it on again sounds a bit 
> unusual :)

What about following,
Right now there is not way to get led status, only have on/off
We can store status in ethdev layer add a 'rte_eth_led_status' which can return
status.
If the on/off dev_ops are not set, it can return 'unavailable' which covers your
usecase.