Re: [PATCH net-next RFC 01/13] devlink: Add reload level option to devlink reload command

[Jiri Pirko]

Tue, Aug 04, 2020 at 10:39:46PM CEST, kuba@xxxxxxxxxx wrote:
>On Tue, 4 Aug 2020 12:04:18 +0200 Jiri Pirko wrote:
>> Mon, Aug 03, 2020 at 10:57:03PM CEST, kuba@xxxxxxxxxx wrote:
>> >I was trying to avoid having to provide a Cartesian product of
>> >operation and system disruption level, if any other action can
>> >be done "live" at some point.
>> >
>> >But no strong feelings about that one.
>> >
>> >Really, as long as there is no driver-specific defaults (or as 
>> >little driver-specific anything as possible) and user actions 
>> >are clearly expressed (fw-reset does not necessarily imply
>> >fw-activation) - the API will be fine IMO.  
>> 
>> Clear actions, that is what I'm fine with.
>> 
>> But not sure how you think we can achieve no driver-specific defaults.
>> We have them already :/ I don't think we can easily remove them and not
>> break user expectations.
>
>AFAIU the per-driver default is needed because we went too low 
>level with what the action constitutes. We need maintain the higher
>level actions.
>
>The user clearly did not care if FW was reset during devlink reload
>before this set, so what has changed? The objective user has is to

Well for mlxsw, the user is used to this flow:
devlink dev flash - flash new fw
devlink dev reload - new fw is activated and reset and driver instances
are re-created.


>activate their config / FW / move to different net ns. 
>
>Reloading the driver or resetting FW is a low level detail which
>achieves different things for different implementations. So it's 
>not a suitable abstraction -> IOW we need the driver default.

I'm confused. So you think we need the driver default?


>
>
>The work flow for the user is:
>
>0. download fw to /lib/firmware
>1. devlink flash $dev $fw
>2. if live activation is enabled
>   yes - devlink reload $dev $live-activate
>   no - report machine has to be drained for reboot
>
>fw-reset can't be $live-activate, because as Jake said fw-reset does
>not activate the new image for Intel. So will we end up per-driver
>defaults in the kernel space, and user space maintaining a mapping from

Well, that is what what is Moshe's proposal. Per-driver kernel default..
I'm not sure what we are arguing about then :/


>a driver to what a "level" of reset implies.
>
>I hope this makes things crystal clear. Please explain what problems
>you're seeing and extensions you're expecting. A list of user scenarios
>you foresee would be v. useful.