Re: [PATCH 3/3] arm64: dts: rockchip: Add RTL8367RB-VB switch to Radxa E24C

[Jonas Karlman <jonas@kwiboo.se>]

Hi Andrew,

On 7/27/2025 9:16 PM, Andrew Lunn wrote:
> On Sun, Jul 27, 2025 at 06:03:00PM +0000, Jonas Karlman wrote:
>> The Radxa E24C has a Realtek RTL8367RB-VB switch with four usable ports
>> and is connected using a fixed-link to GMAC1 on the RK3528 SoC.
>>
>> Add an ethernet-switch node to describe the RTL8367RB-VB switch.
>>
>> Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
>> ---
>> Initial testing with iperf3 showed ~930-940 Mbits/sec in one direction
>> and only around ~1-2 Mbits/sec in the other direction.
>>
>> The RK3528 hardware design guide recommends that timing between TXCLK
>> and data is controlled by MAC, and timing between RXCLK and data is
>> controlled by PHY.
>>
>> Any mix of MAC (rx/tx delay) and switch (rx/tx internal delay) did not
>> seem to resolve this speed issue, however dropping snps,tso seems to fix
>> that issue.
> 
> It could well be that the Synopsis TSO code does not understand the
> DSA headers. When it takes a big block to TCP data and segments it,
> you need to have the DSA header on each segment. If it does not do
> that, only the first segment has the DSA header, the switch is going
> to be dropping all the other segments, causes TCP to do a lot of
> retries.

Thanks for your insights!

I can confirm that disable of TSO and RX checksum offload on the conduit
interface help fix any TCP speed issue and reduced UDP packet loss to a
minimum.

Regards,
Jonas

> 
>> Unsure what is best here, should MAC or switch add the delays?
> 
> It should not matter. 2ns is 2ns...
> 
> 	Andrew