I have been able to solve the issue with gigabit speed that some TV box have, see here: T95Z max ethernet problem
Sam from OMSC says that according to him, the cali_val 0x20000 has always a bad impact on gigabit speed and stability, but I am far from having tested the removal of this dts binding on many boards
So at least, it is time to add another dts, or regress test on as many as S912 or gigabit board we can the ethernet performance using dts with no cali_val binding
Don’t see any regression on my X92 device after this change
I have checked the mainline dwmac_meson8b.c driver: they don’t set the ETH_REG2 registers, so basically they run without any cali_val. According to the people I have talked with, there is zero documentation of this registers, looking at the symptoms of what writing 0x20000 does, most probably it set RX clock screw, that is board specific and cannot be good for all.
@anon88919003 what do you think? Shall we just remove it, or make a new dts?
I think we can remove it and see what the feedback is in the nightlies as the next beta won’t be due for anywhere upto a month.
I’m all for giving it a shot.
The changes have been made and will appear in tomorrows nightly.
The reason for doing it this way was so it could be tested in nightlies before we made it a global change, we have had no negative reports about it yet. I also checked and cali_val is never used on S905 which is the majority of the userbase so if there is a negative impact it would be on a relatively low number of users, time will tell though.
In amlogic 4.9 bsp they set it also for internal phy (used by S905X and in general 100mbit box). Someone with contacts in amlogic should ask for register description