Re: [PATCH/RFC net-next 1/2] devlink: expose port function commands to assign VFs to multiple netdevs

From: Jiri Pirko <jiri@resnulli.us>
To: Yinjun Zhang <yinjun.zhang@corigine.com>
Cc: Saeed Mahameed <saeed@kernel.org>,
	Jakub Kicinski <kuba@kernel.org>,
	Simon Horman <simon.horman@corigine.com>,
	Leon Romanovsky <leon@kernel.org>,
	David Miller <davem@davemloft.net>,
	Paolo Abeni <pabeni@redhat.com>,
	Michael Chan <michael.chan@broadcom.com>,
	Andy Gospodarek <andy@greyhouse.net>,
	Gal Pressman <gal@nvidia.com>,
	Jesse Brandeburg <jesse.brandeburg@intel.com>,
	Tony Nguyen <anthony.l.nguyen@intel.com>,
	Edward Cree <ecree.xilinx@gmail.com>,
	Vladimir Oltean <vladimir.oltean@nxp.com>,
	Andrew Lunn <andrew@lunn.ch>, Fei Qin <fei.qin@corigine.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	oss-drivers <oss-drivers@corigine.com>
Subject: Re: [PATCH/RFC net-next 1/2] devlink: expose port function commands to assign VFs to multiple netdevs
Date: Thu, 9 Feb 2023 16:15:58 +0100	[thread overview]
Message-ID: <Y+UOLkAWD0yCJHCb@nanopsycho> (raw)
In-Reply-To: <DM6PR13MB37055FC589B66F4F06EF264FFCD99@DM6PR13MB3705.namprd13.prod.outlook.com>

Thu, Feb 09, 2023 at 03:20:48AM CET, yinjun.zhang@corigine.com wrote:
>On Wed, 8 Feb 2023 16:55:12 -0800, Saeed Mahameed wrote:
>> On 08 Feb 15:35, Jakub Kicinski wrote:
>> >On Wed, 8 Feb 2023 13:37:08 -0800 Saeed Mahameed wrote:
>> >> I don't understand the difference between the two modes,
>> >> 1) "where VFs are associated with physical ports"
>> >> 2) "another mode where all VFs are associated with one physical port"
>> >>
>> >> anyway here how it works for ConnectX devices, and i think the model
>> should
>> >> be generalized to others as it simplifies the user life in my opinion.
>> >
>> >I'm guessing the version of the NFP Simon posted this for behaves
>> >much like CX3 / mlx4. One PF, multiple Ethernet ports.
>> 
>> Then the question is, can they do PF per port and avoid such complex APIs ?
>> 
>
>To answer your last question, it needs silicon support, so we can't for some old products.
>
>Then let me clarify something more for this patch-set's purpose. 
>Indeed, one port per PF is current mainstream. In this case, all the VFs created from PF0
>use physical port 0 as the uplink port(outlet to external world), and all the VFs from PF1
>use p1 as the uplink port. Let me call them two switch-sets. And they're isolated, you can't 
>make the traffic input from VFs of PF0 output to p1 or VFs of PF1, right? Even with TC in
>switchdev mode, the two switch-sets are still isolated, right? Correct me if I'm wrong here.
>And the posted configuration in this patch-set is useless in this case, it's for one PF with
>multi ports.
>
>Let me take NFP implementation for example here, all the VFs created from the single PF
>use p0 as the uplink port by default. In legacy mode, by no means we can choose other

Legacy is legacy. I believe it is like 5 years already no knobs for
legacy mode are accepted. You should not use it for new features.
Why this is any different?

Implement TC offloading and then you can ballance the hell out of the
thing :)

>ports as outlet. So what we're doing here is try to simulate one-port-per-PF case, to split
>one switch-set to several switch-sets with every physical port as the uplink port respectively,
>by grouping the VFs and assigning them to physical ports.