Re: [PATCH net-next v1 0/9] Ethernet Cable test support

[Michael Walle <michael@walle.cc>]

Am 2020-04-30 22:19, schrieb Florian Fainelli:
> On 4/30/20 1:13 PM, Michael Walle wrote:
>> Am 2020-04-30 22:04, schrieb Florian Fainelli:
>>> On 4/30/20 12:41 PM, Andrew Lunn wrote:
>>>>> ECD. The registers looks exactly like the one from the Marvell 
>>>>> PHYs,
>>>>> which makes me wonder if both have the same building block or if 
>>>>> one
>>>>> imitated the registers of the other. There are subtle differences
>>>>> like one bit in the broadcom PHY is "break link" and is 
>>>>> self-clearing,
>>>>> while the bit on the Marvell PHY is described as "perform 
>>>>> diagnostics
>>>>> on link break".
>>>> 
>>>> Should we be sharing code between the two drivers?
>>> 
>>> Yes, I am amazed how how identical they are, nearly on a bit level
>>> identical, the coincidence is uncanny. The expansion registers are 
>>> also
>>> 0x10 - 0x15 just in the reverse order,
>> 
>> At what PHY are you looking? I've just found some datasheets where 
>> they
>> are at 0xC0 to 0xC5.
> 
> BCM54810 because that's what I have on my desk right now, but 0x10 
> would
> be the offset relative to the expansion register space, which would
> translate into this:
> 
> https://github.com/ffainelli/linux/commits/broadcom-cable-tests
> 
> (sorry for the mess it is a patchwork of tests on various platforms,
> based on an earlier branch from Andrew).

Ah thanks, now I see what you mean with CD vs ECD. In your latest WIP
you're using ECD, so did both actually work with a link?

Also according to you header files bit 14 of the ECD_CTRL is the
"run after aneg bit"; in the BCM54140 it is just marked as reserved.
I guess I'm trying the CD on the BCM54140 tomorrow.

-michael

> 
>> 
>>> you know, so as to make it not
>>> too obvious this looks about the same ;) I wonder if we managed to 
>>> find
>>> something here.
>>> 
>>>> 
>>>>> What do you mean by calibrate it?
>>>> 
>>>> Some of the Marvell documentation talks about calibrating for losses
>>>> on the PCB. Run a diagnostics with no cable plugged in, and get the
>>>> cable length to the 'fault'. This gives you the distance to the RJ45
>>>> socket. You should then subtract that from all subsequent results.
>>>> But since this is board design specific, i decided to ignore it. I
>>>> suppose it could be stuffed into a DT property, but i got the 
>>>> feeling
>>>> it is not worth it, given the measurement granularity of 80cm.
>>> 
>>> OK, accuracy is different here, they are said to be +/- 5m accurate 
>>> for
>>> cable faults and +/- 10m accurate for good cables.
>> 
>> Accuracy != granularity. But yes, if one digit correspond to 80cm it
>> doesn't really make sense to remove the PCB trace error if you assume
>> that it might add just one digit at most.
> 
> One of the test racks that I use has very short cables, but I guess it
> does not matter if they get reported as 80cm or 160cm...