Re: [PATCH v4] net: macb: Fix several edge cases in validate

On 10/26/21 2:28 PM, Russell King (Oracle) wrote:

On Tue, Oct 26, 2021 at 01:49:03PM -0400, Sean Anderson wrote:

quoted

On 10/26/21 1:46 PM, Russell King (Oracle) wrote:

quoted

On Tue, Oct 26, 2021 at 01:28:15PM -0400, Sean Anderson wrote:

quoted

Actually, according to the Zynq UltraScale+ Devices Register Reference
[1], the PCS does not support 10/100. So should SGMII even fall through
here?

[1] https://www.xilinx.com/html_docs/registers/ug1087/gem___pcs_control.html

Hmm. That brings with it fundamental question: if the PCS supports 1G
only, does it _actually_ support Cisco SGMII, or does it only support
1000base-X?

Of course, in the technical reference manual [1], they say

quoted

The line rate is 1 Gb/s as SGMII hardwired to function at 1 Gb/s only.
However, the data transfer rate can be forced down to 100 Mb/s or 10
Mb/s if the link partner is not capable.

which sounds like the normal byte-repetition of SGMII...

And they also talk about how the autonegotiation timeout and control words are different.

This document looks a little comical. "GEM supports the serial gigabit
media-independent interface (SGMII, 1000BASE-SX and 1000BASE-LX) at
1000 Gb/s using the PS-GTR interface."

So:
a) it supports terabyte speeds?
b) it provides an optical connection direct from the SoC?
    (the L and S in 1000BASE-X refer to the laser wavelength!)

They really should just be saying "1000BASE-X" rather than specifying
an optical standard, but lets ignore that fundamental mistake for now.

In the section "SGMII, 1000BASE-SX, or 1000BASE-LX" it says:

"When bit [27] (SGMII mode) in the network configuration register
(GEM(0:3).network_config[sgmii_mode_enable]) is set, it changes the
behavior of the auto-negotiation advertisement and link partner ability
registers to meet the requirements of SGMII. Additionally, the time
duration of the link timer is reduced from 10ms to 1.6ms."

That bodes well for Cisco SGMII support, but it says nothing about how
that affects the PCS registers themselves.

As you say above, it goes on to say:

"The line rate is 1 Gb/s as SGMII hardwired to function at 1 GB/s
only."

That statement is confused. Cisco SGMII and 1000Base-X actually operate
at 1.25Gbaud line rate due to the 4B5B encoding on the Serdes. However,
the underlying data rate is 1Gbps, with 100 and 10Mbps achieved by
symbol replication of only the data portions of the packet transfer.
This replication does not, however, apply to non-data symbols though.

I suppose they _could_ have implemented Cisco SGMII by having the PCS
fixed in 1G mode, and then replicate the data prior to the PCS. That
would be rather peculiar though, and I'm not sure whether that could
work for the non-data symbols. I suppose it depends whether they just
slow down the transmission rate into the PCS, or do only data portion
replication before the PCS.

I've also just found the register listing in HTML form (so less
searchable than a PDF),

Unfortunately, AFAIK this is the only listing they have :l

(OTOH it is easy to link to)

and the PCS register listing only shows
1000base-X layout for the advertisement and link partner registers.
It seems to make no mention of "SGMII" mode.

The autonegotiation registers [2, 3] both mention SGMII "non SGMII".

The mux to turn on SGMII is in the network_config register [3].

[1] https://www.xilinx.com/html_docs/registers/ug1087/gem___pcs_an_adv.html
[2] https://www.xilinx.com/html_docs/registers/ug1087/gem___pcs_an_lp_base.html
[3] https://www.xilinx.com/html_docs/registers/ug1087/gem___network_config.html

So we have an open question: do 10 and 100M speeds actually work on
GEM, and if they do, how does one program it to operate at these
speeds. Currently, the driver seems to change bits in NCFGR to change
speed, and also reconfigure the transmit clock rate.

Going back to the first point I mentioned above, how much should we
take from these documents as actually being correct? Should we not
assume anything, but instead just experiment with the hardware and
see what works.

For example, are the two speed bits in the PCS control register
really read-only when in Cisco SGMII mode, or can they be changed -
and if they can be changed, does that have an effect on the ethernet
link?

Keep in mind that it is not only Zynq(MP) parts with have GEMs, but
several other SoCs as well. I have not reviewed their datasheets (except
for SiFive's which just say "go read the Linux driver"). It is possible
that other SoCs may not have these limitations. So any experimental
program will need to also experiment with e.g. sama.

Perhaps someone from cadence could comment on what is actually supported
by gem/macb?

--Sean