On Wed, Feb 15, 2023 at 02:34:23AM +0200, Cristian Ciocaltea wrote: > On 2/11/23 18:01, Andrew Lunn wrote: > > > + starfive,gtxclk-dlychain: > > > + $ref: /schemas/types.yaml#/definitions/uint32 > > > + description: GTX clock delay chain setting > > > > Please could you add more details to this. Is this controlling the > > RGMII delays? 0ns or 2ns? > > This is what gets written to JH7100_SYSMAIN_REGISTER49 and it's currently > set to 4 in patch 12/12. As already mentioned, I don't have the register > information in the datasheet, but I'll update this as soon as we get some > details. I have seen what happens to this value, but i have no idea what it actually means. And without knowing what it means, i cannot say if it is being used correctly or not. And it could be related to the next part of my comment... > > > > + gmac: ethernet@10020000 { > > > + compatible = "starfive,jh7100-dwmac", "snps,dwmac"; > > > + reg = <0x0 0x10020000 0x0 0x10000>; > > > + clocks = <&clkgen JH7100_CLK_GMAC_ROOT_DIV>, > > > + <&clkgen JH7100_CLK_GMAC_AHB>, > > > + <&clkgen JH7100_CLK_GMAC_PTP_REF>, > > > + <&clkgen JH7100_CLK_GMAC_GTX>, > > > + <&clkgen JH7100_CLK_GMAC_TX_INV>; > > > + clock-names = "stmmaceth", "pclk", "ptp_ref", "gtxc", "tx"; > > > + resets = <&rstgen JH7100_RSTN_GMAC_AHB>; > > > + reset-names = "ahb"; > > > + interrupts = <6>, <7>; > > > + interrupt-names = "macirq", "eth_wake_irq"; > > > + max-frame-size = <9000>; > > > + phy-mode = "rgmii-txid"; > > > > This is unusual. Does your board have a really long RX clock line to > > insert the 2ns delay needed on the RX side? > > Just tested with "rgmii" and didn't notice any issues. If I'm not missing > anything, I'll do the change in the next revision. rgmii-id is generally the value to be used. That indicates the board needs 2ns delays adding by something, either the MAC or the PHY. And then i always recommend the MAC driver does nothing, pass the value to the PHY and let the PHY add the delays. So try both rgmii and rgmii-id and do a lot of bi directional transfers. Then look at the reported ethernet frame check sum error counts, both local and the link peer. I would expect one setting gives you lots of errors, and the other works much better. Andrew