[BUG] MTK SoC Ethernet throughput TX only ~620Mbit/s since 6.2-rc1

[Daniel Golle]

Hi Felix,

On Tue, Mar 14, 2023 at 11:30:53AM +0100, Felix Fietkau wrote:
> On 07.03.23 19:32, Frank Wunderlich wrote:
> > Hi,
> > 
> > i've noticed that beginning on 6.2-rc1 the throughput on my Bananapi-R2 and Bananapi-R3 goes from 940Mbit/s down do 620Mbit/s since 6.2-rc1.
> > Only TX (from SBC PoV) is affected, RX is still 940Mbit/s.
> > 
> > i bisected this to this commit:
> > 
> > f63959c7eec3151c30a2ee0d351827b62e742dcb ("net: ethernet: mtk_eth_soc: implement multi-queue support for per-port queues")
> > 
> > Daniel reported me that this is known so far and they need assistance from MTK and i should report it officially.
> > 
> > As far as i understand it the commit should fix problems with clients using non-GBE speeds (10/100 Mbit/s) on the Gbit-capable dsa
> > interfaces (mt753x connected) behind the mac, but now the Gigabit speed is no more reached.
> > I see no CRC/dropped packets, retransmitts or similar.
> > 
> > after reverting the commit above i get 940Mbit like in rx direction, but this will introduce the problems mentioned above so this not a complete fix.
> I don't have a BPI-R2, but I tested on BPI-R3 and can't reproduce this
> issue. Do you see it on all ports, or only on specific ones?

I also can't reproduce this if unsing any of the gigE ports wired via
MT7531 on the R3. However, I can reproduce the issue if using a 1 GBit/s
SFP module in slot SFP1 of the R3 (connected directly to GMAC2/eth1).

Users have reported this also to be a problem also on MT7622 on devices
directly connecting a PHY (and not using MT7531).

In all cases, reverting the commit above fixes the issue.


Cheers


Daniel