Re: [PATCH 07/12] dt-bindings: net: Add StarFive JH7100 SoC

[Cristian Ciocaltea <cristian.ciocaltea@collabora.com>]

On 2/17/23 17:25, Cristian Ciocaltea wrote:
> On 2/17/23 15:30, Andrew Lunn wrote:
>>>> I would actually say it shows we don't understand what is going on
>>>> with delays. "rgmii" is not every often the correct value. The fact it
>>>> works suggests the MAC is adding delays.
>>>>
>>>> What value are you using for starfive,gtxclk-dlychain ?
>>>
>>> This is set to '4' in patch 12/12.
>>>
>>>> Try 0 and then "rgmii-id"
>>>
>>> I made some more tests and it seems the only stable configuration is
>>> "rgmii"
>>> with "starfive,gtxclk-dlychain" set to 4:
>>>
>>> phy-mode | dlychain | status
>>> ---------+----------+--------------------------------------------
>>> rgmii    |        4 | OK (no issues observed)
>>> rgmii-id |        4 | BROKEN (errors reported [1])
>>> rgmii    |        0 | UNRELIABLE (no errors, but frequent stalls)
>>> rgmii-id |        0 | BROKEN (errors reported)
>>>
>>> [1] Reported errors in case of BROKEN status:
>>> $ grep '' /sys/class/net/eth0/statistics/* | grep -v ':0$'
>>
>> Thanks for the testing.
>>
>> So it seems like something is adding delays when it probably should
>> not. Ideally we want to know what.
>>
>> There is a danger here, something which has happened in the past. A
>> PHY which ignored "rgmii" and actually did power on defaults which was
>> "rgmii-id". As a result, lots of boards put "rmgii" in there DT blob,
>> which 'worked'. Until a board came along which really did need
>> "rgmii". The developer bringing that board up debugged the PHY, found
>> the problem and made it respect "rgmii" so their board worked. And the
>> fix broke a number of 'working' boards which had the wrong "rgmii"
>> instead of "rgmii-id".
> 
> Thanks for the heads-up.
> 
>> So you have a choice. Go with 4 and "rgmii", but put in a big fat
>> warning, "Works somehow but is technically wrong and will probably
>> break sometime in the future". Or try to understand what is really
>> going on here, were are the delays coming from, and fix the issue.
>>
>>        Andrew
> 
> I will try to analyze this further.

As the non-coherent DMA work this series depended on has been completed,
I started to investigate further the "rgmii-id" issue.
I couldn't spot anything wrong in the Motorcomm PHY driver, but
eventually got this working by adjusting rx-internal-delay-ps.

Will do some more testing before submitting v2.

Thanks,
Cristian