Re: [PATCH net-next 03/10] devlink: Serialize access to rate domains

[Jiri Pirko]

Wed, Mar 05, 2025 at 01:04:12AM +0100, kuba@xxxxxxxxxx wrote:
>On Tue, 4 Mar 2025 14:11:40 +0100 Jiri Pirko wrote:
>> Mon, Mar 03, 2025 at 11:06:23PM +0100, kuba@xxxxxxxxxx wrote:
>> >On Thu, 27 Feb 2025 13:22:25 +0100 Jiri Pirko wrote:  
>> >> Depends. On normal host sr-iov, no. On smartnic where you have PF in
>> >> host, yes.  
>> >
>> >Yet another "great choice" in mlx5 other drivers have foreseen
>> >problems with and avoided.  
>> 
>> What do you mean? How else to model it? Do you suggest having PF devlink
>> port for the PF that instantiates? That would sound like Uroboros to me.
>
>I reckon it was always more obvious to those of us working on
>NPU-derived devices, to which a PCIe port is just a PCIe port,
>with no PCIe<>MAC "pipeline" to speak of.
>
>The reason why having the "PF port" is a good idea is exactly
>why we're having this conversation. If you don't you'll assign
>to the global scope attributes which are really just port attributes.

Well, we have devlink port for uplink for this purpose. Why isn't that
enough?


>
>> >> Looks like pretty much all current NICs are multi-PFs, aren't they?  
>> >
>> >Not in a way which requires cross-port state sharing, no.
>> >You should know this.  
>> 
>> This is not about cross-port state sharing. This is about per-PF
>> configuration. What am I missing?
>
>Maybe we lost the thread of the conversation.. :)
>I'm looking at the next patch in this series and it says:
>
>  devlink: Introduce shared rate domains
>
>  The underlying idea is modeling a piece of hardware which:
>  1. Exposes multiple functions as separate devlink objects.
>  2. Is capable of instantiating a transmit scheduling tree spanning
>     multiple functions.
>
>  Modeling this requires devlink rate nodes with parents across other
>  devlink objects.
>
>Are these domains are not cross port?

Sure. Cross PF even. What I suggest is, if we have devlink instance of
which these 2 PFs are nested, we have this "domain" explicitly defined
and we don't need any other construct.