Re: [PATCH net-next 5/7] net: marvell: prestera: add LAG support

[Tobias Waldekranz <tobias@waldekranz.com>]

On Mon, Feb 08, 2021 at 13:05, Jakub Kicinski <kuba@kernel.org> wrote:
> On Mon, 08 Feb 2021 20:54:29 +0100 Tobias Waldekranz wrote:
>> On Thu, Feb 04, 2021 at 21:16, Jakub Kicinski <kuba@kernel.org> wrote:
>> > On Wed,  3 Feb 2021 18:54:56 +0200 Vadym Kochan wrote:  
>> >> From: Serhiy Boiko <serhiy.boiko@plvision.eu>
>> >> 
>> >> The following features are supported:
>> >> 
>> >>     - LAG basic operations
>> >>         - create/delete LAG
>> >>         - add/remove a member to LAG
>> >>         - enable/disable member in LAG
>> >>     - LAG Bridge support
>> >>     - LAG VLAN support
>> >>     - LAG FDB support
>> >> 
>> >> Limitations:
>> >> 
>> >>     - Only HASH lag tx type is supported
>> >>     - The Hash parameters are not configurable. They are applied
>> >>       during the LAG creation stage.
>> >>     - Enslaving a port to the LAG device that already has an
>> >>       upper device is not supported.  
>> >
>> > Tobias, Vladimir, you worked on LAG support recently, would you mind
>> > taking a look at this one?  
>> 
>> I took a quick look at it, and what I found left me very puzzled. I hope
>> you do not mind me asking a generic question about the policy around
>> switchdev drivers. If someone published a driver using something similar
>> to the following configuration flow:
>> 
>> iproute2  daemon(SDK)
>>    |        ^    |
>>    :        :    : user/kernel boundary
>>    v        |    |
>> netlink     |    |
>>    |        |    |
>>    v        |    |
>>  driver     |    |
>>    |        |    |
>>    '--------'    |
>>                  : kernel/hardware boundary
>>                  v
>>                 ASIC
>> 
>> My guess is that they would be (rightly IMO) told something along the
>> lines of "we do not accept drivers that are just shims for proprietary
>> SDKs".
>> 
>> But it seems like if that same someone has enough area to spare in their
>> ASIC to embed a CPU, it is perfectly fine to run that same SDK on it,
>> call it "firmware", and then push a shim driver into the kernel tree.
>> 
>> iproute2
>>    |
>>    :               user/kernel boundary
>>    v
>> netlink
>>    |
>>    v
>>  driver
>>    |
>>    |
>>    :               kernel/hardware boundary
>>    '-------------.
>>                  v
>>              daemon(SDK)
>>                  |
>>                  v
>>                 ASIC
>> 
>> What have we, the community, gained by this? In the old world, the
>> vendor usually at least had to ship me the SDK in source form. Having
>> seen the inside of some of those sausage factories, they are not the
>> kinds of code bases that I want at the bottom of my stack; even less so
>> in binary form where I am entirely at the vendor's mercy for bugfixes.
>> 
>> We are talking about a pure Ethernet fabric here, so there is no fig
>> leaf of "regulatory requirements" to hide behind, in contrast to WiFi
>> for example.
>> 
>> Is it the opinion of the netdev community that it is OK for vendors to
>> use this model?
>
> I ask myself that question pretty much every day. Sadly I have no clear
> answer.

Thank you for your candid answer, really appreciate it. I do not envy
you one bit, making those decisions must be extremely hard.

> Silicon is cheap, you can embed a reasonable ARM or Risc-V core in the
> chip for the area and power draw comparable to one high speed serdes
> lane.
>
> The drivers landing in the kernel are increasingly meaningless. My day
> job is working for a hyperscaler. Even though we have one of the most
> capable kernel teams on the planet most of issues with HW we face
> result in "something is wrong with the FW, let's call the vendor".

Right, and being a hyperscaler probably at least gets you some attention
when you call your vendor. My day job is working for a nanoscaler, so my
experience is that we must be prepared to solve all issues in-house; if
we get any help from the vendor that is just a bonus.

> And even when I say "drivers landing" it is an overstatement.
> If you look at high speed anything these days the drivers cover
> multiple generations of hardware, seems like ~5 years ago most
> NIC vendors reached sufficient FW saturation to cover up differences
> between HW generations.
>
> At the same time some FW is necessary. Certain chip functions, are 
> best driven by a micro-controller running a tight control loop. 

I agree. But I still do not understand why vendors cling to the source
of these like it was their wallet. That is the beauty of selling
silicon; you can fully leverage OSS and still have a very straight
forward business model.

> The complexity of FW is a spectrum, from basic to Qualcomm. 
> The problem is there is no way for us to know what FW is hiding
> by just looking at the driver.
>
> Where do we draw the line? 

Yeah it is a very hard problem. In this particular case though, the
vendor explicitly said that what they have done is compiled their
existing SDK to run on the ASIC:

https://lore.kernel.org/netdev/BN6PR18MB1587EB225C6B80BF35A44EBFBA5A0@BN6PR18MB1587.namprd18.prod.outlook.com

So there is no reason that it could not be done as a proper driver.

> Personally I'd really like to see us pushing back stronger.

Hear, hear!