Re: [PATCH V2 net 00/11] Extending the ena driver to support new features and enhance performance

[Jakub Kicinski <jakub.kicinski@netronome.com>]

On Tue, 04 Jun 2019 07:57:48 +0100, David Woodhouse wrote:
> On Tue, 2019-06-04 at 02:15 +0000, Bshara, Nafea wrote:
> > On Jun 3, 2019, at 6:52 PM, Andrew Lunn <andrew@lunn.ch> wrote:
> > > > Any "SmartNIC" vendor has temptation of uAPI-level hand off to the
> > > > firmware (including my employer), we all run pretty beefy processors
> > > > inside "the NIC" after all.  The device centric ethtool configuration
> > > > can be implemented by just forwarding the uAPI structures as they are
> > > > to the FW.  I'm sure Andrew and others who would like to see Linux
> > > > takes more control over PHYs etc. would not like this scenario, either.  
> > > 
> > > No, i would not. There are a few good examples of both firmware and
> > > open drivers being used to control the same PHY, on different
> > > boards. The PHY driver was developed by the community, and has more
> > > features than the firmware driver. And it keeps gaining features. The
> > > firmware i stuck, no updates. The community driver can be debugged,
> > > the firmware is a black box, no chance of the community fixing any
> > > bugs in it.
> > > 
> > > And PHYs are commodity devices. I doubt there is any value add in the
> > > firmware for a PHY, any real IPR which makes the product better, magic
> > > sauce related to the PHY. So just save the cost of writing and
> > > maintaining firmware, export the MDIO bus, and let Linux control it.
> > > Concentrate the engineers on the interesting parts of the NIC, the
> > > Smart parts, where there can be real IPR.
> > > 
> > > And i would say this is true for any NIC. Let Linux control the PHY.
> > 
> > It may be true for old GbE PHYs where it’s a discrete chip from the
> > likes of Marvell or broadcom
> > 
> > But at 25/50/100G, the PHy is actually part of the nic. It’s a very
> > complex SERDES.  Cloud providers like us spend enormous amount of
> > time testing the PHY across process and voltage variations, all cable
> > types, length and manufacturing variations, and against all switches
> > we use.  Community drivers won’t be able to validate and tune all
> > this.
> > 
> > Plus we would need exact same setting for Linux, including all
> > distributions even 10year old like RHEL6, for all Windows, ESX, DPDK,
> > FreeBSD,  and support millions of different customers with different
> > sets of Machine images. 
> > 
> > In this case, there is no practical choice by have the firmware to
> > manage the PHY  
> 
> I don't quite know why we're talking about PHYs in this context.
> ENA is basically a virtio NIC. It has no PHY.

I brought it up as an example, to illustrate that we'd rather see less
trust and control being blindly handed over to the firmware.

Would you mind answering what are the examples of very important flags
you need to expose to users with 10yo kernels?  Or any examples for that
matter..