Re: [PATCH net-next RFC v3 01/14] devlink: Add reload action option to devlink reload command

[Jiri Pirko]

Thu, Sep 03, 2020 at 09:47:19PM CEST, kuba@xxxxxxxxxx wrote:
>On Thu, 3 Sep 2020 07:57:29 +0200 Jiri Pirko wrote:
>> Wed, Sep 02, 2020 at 05:30:25PM CEST, kuba@xxxxxxxxxx wrote:
>> >On Wed, 2 Sep 2020 11:46:27 +0200 Jiri Pirko wrote:  
>> >> >? Do we need such change there too or keep it as is, each action by itself
>> >> >and return what was performed ?    
>> >> 
>> >> Well, I don't know. User asks for X, X should be performed, not Y or Z.
>> >> So perhaps the return value is not needed.
>> >> Just driver advertizes it supports X, Y, Z and the users says:
>> >> 1) do X, driver does X
>> >> 2) do Y, driver does Y
>> >> 3) do Z, driver does Z
>> >> [
>> >> I think this kindof circles back to the original proposal...  
>> >
>> >Why? User does not care if you activate new devlink params when
>> >activating new firmware. Trust me. So why make the user figure out
>> >which of all possible reset option they should select? If there is 
>> >a legitimate use case to limit what is reset - it should be handled
>> >by a separate negative attribute, like --live which says don't reset
>> >anything.  
>> 
>> I see. Okay. Could you please sum-up the interface as you propose it?
>
>What I proposed on v1, pass requested actions as a bitfield, driver may
>perform more actions, we can return performed actions in the response.

Okay. So for example for mlxsw, user might say:
1) I want driver reinit
    kernel reports: fw reset and driver reinit was done
2) I want fw reset
    kernel reports: fw reset and driver reinit was done
3) I want fw reset and driver reinit
    kernel reports: fw reset and driver reinit was done

>
>Then separate attribute to carry constraints for the request, like
>--live.

Hmm, this is a bit unclear how it is supposed to work. The constraints
apply for all? I mean, the actions are requested by a bitfield.
So the user can say:
I want fw reset and driver reinit --live. "--live" applies to both fw
reset and driver reinit? That is odd.

>
>I'd think the supported actions in devlink_ops would be fine as a
>bitfield, too. Combinations are often hard to capture in static data.