Re: [PATCH v2 2/2] arm64: dts: rockchip: Add rk3576 evb2 board

[Chaoyi Chen <chaoyi.chen@rock-chips.com>]

On 1/8/2026 4:49 PM, Alexey Charkov wrote:
> On Thu, Jan 8, 2026 at 12:38 PM Chaoyi Chen <chaoyi.chen@rock-chips.com> wrote:
>>
>> On 1/8/2026 4:11 PM, Alexey Charkov wrote:
>>> On Thu, Jan 8, 2026 at 12:02 PM Chaoyi Chen <kernel@airkyi.com> wrote:
>>>>
>>>> On 1/8/2026 3:42 PM, Chaoyi Chen wrote:
>>>>> Hello Alexey, Andrew,
>>>>>
>>>>> On 1/8/2026 2:53 PM, Alexey Charkov wrote:
>>>>>> On Wed, Jan 7, 2026 at 10:18 PM Andrew Lunn <andrew@lunn.ch> wrote:
>>>>>>>
>>>>>>>> +&gmac0 {
>>>>>>>> +     clock_in_out = "output";
>>>>>>>> +     phy-mode = "rgmii-rxid";
>>>>>>>
>>>>>>> rgmii-rxid is odd. Does the PCB really have an extra long TX clock
>>>>>>> line, but a short RX clock line?
>>>>>>>
>>>>>>> Try changing this to rgmii-id, and drop the tx_delay property.
>>>>>>
>>>>>> Actually it would be great if Rockchip could clarify the delay
>>>>>> duration introduced by a single delay element in GMAC-IOMUX delay
>>>>>> lines, which are controlled in the GMAC driver by the {tx,rx}_delay
>>>>>> properties. Maybe we could then switch to using
>>>>>> {tx,rx}_internal_delay_ps for fine-tuning the delays on the GMAC side
>>>>>> as envisaged in DT bindings [1], and use phy-mode = "rgmii-id"
>>>>>> throughout. Chaoyi, any chance you could ask around in your hardware
>>>>>> team?
>>>>>>
>>>>>> Currently though removing the delays at GMAC side altogether causes
>>>>>> unstable link operation - see [2] for example.
>>>>>>
>>>>>> [1] https://github.com/torvalds/linux/blob/master/Documentation/devicetree/bindings/net/ethernet-controller.yaml#L342-L347
>>>>>> [2] https://gitlab.collabora.com/hardware-enablement/rockchip-3588/linux/-/commit/372f3e9ae62cc62cdf2543391ea57be6bb548a0c
>>>>>
>>>>> Sorry, this problem has been discussed many times before. It's because
>>>>> the gmac on the Rockchip platform currently relies on setting the
>>>>> corresponding delay via phy-mode [3].
>>>>>
>>>>> [3] https://lore.kernel.org/all/mqoyjn7mnq6tmt6n6oev4wa3herjaxlupml2fmcampwiajvj4a@r5zs4d3jdm5p/
>>>>>
>>>>> The delay introduced by the delay line is not absolute. In reality,
>>>>> it depends on factors such as the chip's design and process technology.
>>>>>
>>>>> And for RK3576, you can assume that:
>>>>>
>>>>>       time(ns) = 0.0579 * delay_line_count + 0.105
>>>>>
>>>>> For example, tx_delay = <0x20> means:
>>>>>
>>>>>       time = 0.0579 * 0x20 + 0.105 ns = 1.9578 ns
>>>>>
>>>>> And I believe {tx,rx}_internal_delay_ps is indeed a good idea.
>>>>> I'll try to add them in v3. Thanks.
>>>>>
>>>>
>>>> I've also see some dt that use {tx,rx}_internal_delay_ps inside the PHY,
>>>> and compared to doing it in the MAC, which one is the better choice?
>>>
>>> Your PHY defaults to 1950ps in rgmii-id [1], so adding anything on top
>>> of that on GMAC side would land you with a longer total TX delay than
>>> you currently get according to the coefficients you've just posted
>>> (1784.1ps). I would say go for "tx-internal-delay-ps = <1800>" on the
>>> PHY side for the closest match.
>>>
>>> [1] https://github.com/torvalds/linux/blob/master/Documentation/devicetree/bindings/net/motorcomm%2Cyt8xxx.yaml#L36
>>
>> Ah, I thought it was something like this:
>>
>> &gmac0 {
>>         phy-mode = "rgmii";
> 
> phy-mode = "rgmii-id";
> 
>>         tx-internal-delay-ps = <1784>;
> 
> Drop this, as recommended in the lengthy note at the bottom of the
> binding doc [1] (the GMAC shouldn't be adding a delay that is close to
> 2ns).
> 
> [1] https://github.com/torvalds/linux/blob/master/Documentation/devicetree/bindings/net/ethernet-controller.yaml#L342-L354
> 
>> };
>>
>> But what you actually said was this:
>>
>> &mdio1 {
> 
> &mdio0 I guess (but same applies to gmac1+mdio1)
> 
>>         rgmii_phy: ethernet-phy@1 {
>>                 compatible = "ethernet-phy-ieee802.3-c22";
>>                 reg = <0x1>;
>>                 tx-internal-delay-ps = <1784>;
> 
> The PHY binding [2] says it can't accept arbitrary values here - only
> 150ps increments. So you'll need either 1650 or 1800 to get close to
> your calculated value above.
> 
> [2] https://github.com/torvalds/linux/blob/master/Documentation/devicetree/bindings/net/motorcomm%2Cyt8xxx.yaml#L36
> 
>>         };
>> };
>>
>> These two `tx-internal-delay-ps` things shouldn't be the same, right?
> 
> They add up, which you don't need in this case AFAICT.
> 

Ah, got it. Thank you for the clarification :)

-- 
Best, 
Chaoyi