Re: [PATCH v2 net-next 3/3] net: phy: aquantia: add support for aqr115c

[Daniel Golle <daniel@makrotopia.org>]

On Fri, Jun 28, 2024 at 02:18:45AM +0200, Andrew Lunn wrote:
> On Thu, Jun 27, 2024 at 11:42:45PM +0100, Daniel Golle wrote:
> > Hi Bartosz,
> > 
> > On Thu, Jun 27, 2024 at 01:30:17PM +0200, Bartosz Golaszewski wrote:
> > > From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> > > 
> > > Add support for a new model to the Aquantia driver. This PHY supports
> > > Overlocked SGMII mode with 2.5G speeds.
> > 
> > I don't think that there is such a thing as "Overclocked SGMII mode with
> > 2.5G speed".
> 
> Unfortunately, there is. A number of vendors say they do this, without
> saying quite what they actually do.  As you point out, symbol
> replication does not work, and in-band signalling also makes no
> sense. So they throw all that away. Leaving just the higher clock
> rate, single speed, and no in-band signalling.
> 
> In the end, that looks very similar to 2500BaseX with broken inband
> signalling.

Let's call it that then: "2500Base-X with broken in-band signalling".

MaxLinear describes that quite clearly in their (open!) datasheets[1],
and gives some insight into the (mis-)use of the term "SGMII" in the
industry as synonymous to just any type of serialized Ethernet MII:

"
3.4 SGMII Interface

The GPY211 implements a serial data interface, called SGMII or SerDes,
to connect to another chip implementing the MAC layer (MAC SoC).
"
(page 32)

Later on they mention that
"
3.4.7 Auto-negotiation Modes Supported by SGMII

Two modes are supported for the SGMII auto-negotiation protocol:
 * Cisco* Serial-GMII Specification 1.8 [4]
 * 1000BX IEEE 802.3 following IEEE Clause 37 [2]
"
(page 37)

Aquantia's datasheets are only available under NDA, so I cannot quote
them directly, but I can tell you that their definition of "SGMII" is
pretty similar to that of MaxLinear.

> 
> > Hence I assume that what you meant to say here is that the PHY uses
> > 2500Base-X as interface mode and performs rate-adaptation for speeds
> > less than 2500M (or half-duplex) using pause frames.
> 
> Not all systems assume rate adaptation. Some are known to use SGMII
> for 10/100/1G with inband signalling, and then swap to 2500BaseX
> without inband-signalling for 2.5G operation!

Yes, most 2.5G PHYs out there (MaxLinear, RealTek) actually support both,
with interface-mode switching being the better option compared to often
rather problematic rate-adaptation...

When it comes to Aquantia we are using 2500Base-X with rate adaptation
for the older 2.5G PHYs, so I assume the newer ones would not differ in
that regard. Or rather: If we were to introduce interface-mode-switching
also for the Aquantia 2.5G PHYs then we should try doing it for all of
them at least.

> 
> 2.5G is a mess.

+1

[1]: https://assets.maxlinear.com/web/documents/617810_gpy211b1vc_gpy211c0vc_ds_rev1.4.pdf