Re: [PATCH net] net: phylink: set phy_state interface when attaching SFP

[Daniel Golle]

On Sat, Nov 25, 2023 at 08:28:20AM +0000, Russell King (Oracle) wrote:
> On Sat, Nov 25, 2023 at 04:56:20AM +0000, Daniel Golle wrote:
> > Assume 'usxgmii' being set as initial interface mode in DTS. Now plug
> > a 2.5GBase-T SFP module with exposed PHY. Currently this results in
> > a rather bizare situation:
> > 
> > RTL8221B-VB-CG 2.5Gbps PHY (C45) i2c:sfp1-wan:11: rtl822x_c45_get_features: supported=00,00000000,00008000,000080ef
> > mtk_soc_eth 15100000.ethernet eth2: requesting link mode phy/2500base-x with support 00,00000000,00008000,0000e0ef
> > mtk_soc_eth 15100000.ethernet eth2: switched to phy/2500base-x link mode   <<< !!!!!!
> > mtk_soc_eth 15100000.ethernet eth2: major config usxgmii    <<< !!!!!!
> > mtk_soc_eth 15100000.ethernet eth2: phylink_mac_config: mode=phy/usxgmii/none adv=00,00000000,00000000,00000000 pause=00
> > mtk_soc_eth 15100000.ethernet eth2: PHY [i2c:sfp1-wan:11] driver [RTL8221B-VB-CG 2.5Gbps PHY (C45)] (irq=POLL)
> > mtk_soc_eth 15100000.ethernet eth2: phy: 2500base-x setting supported 00,00000000,00008000,0000e0ef advertising 00,00000000,00008000,0000e0ef
> > 
> > Then the link seemingly comes up (but is dead) because no subsequent
> > call to phylink_major_config actually configured MAC and PCS for
> > 2500base-x mode.
> > 
> > This is because phylink_mac_initial_config() considers
> > pl->phy_state.interface if in MLO_AN_PHY mode while
> > phylink_sfp_set_config() only sets pl->link_config.interface.
> > 
> > Also set pl->phy_state.interface in phylink_sfp_set_config().
> 
> Does it _actually_ matter?

It does matter, I'm suggesting this change because the SFP module
won't work without it.

> When the PHY's link comes up, doesn't it get sorted out for the real
> mode that will be used?

That does happen once the interface mode *changes* to anything else,
than 2500M/Full and the PHY driver sets state->interface to SGMII.
However, initially it does *not* happen, probably because phylink
ends up believing MAC and PCS are already operating in 2500base-x
mode (but that's not the case).

Previously (eg. with the BananaPi R3) this has not been a problem
because the default interface mode defined in device tree anyway has
also been 2500base-x. Now that I'm testing the same module with the
R4 which uses usxgmii by default the problem surfaced.