Re: [PATCH net-next 3/6] net: lan966x: Add QUSGMII support for lan966x

[Maxime Chevallier]

Hello Russell,

On Thu, 19 May 2022 15:26:23 +0100
"Russell King (Oracle)" <linux@xxxxxxxxxxxxxxx> wrote:

> Hi,
> 
> On Thu, May 19, 2022 at 03:56:44PM +0200, Maxime Chevallier wrote:
> > diff --git a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h
> > b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h index
> > e6642083ab9e..304c784f48f6 100644 ---
> > a/drivers/net/ethernet/microchip/lan966x/lan966x_main.h +++
> > b/drivers/net/ethernet/microchip/lan966x/lan966x_main.h @@ -452,4
> > +452,10 @@ static inline void lan_rmw(u32 val, u32 mask, struct
> > lan966x *lan966x, gcnt, gwidth, raddr, rinst, rcnt, rwidth)); }
> >  
> > +static inline bool lan966x_is_qsgmii(phy_interface_t mode)
> > +{
> > +	return (mode == PHY_INTERFACE_MODE_QSGMII) ||
> > +	       (mode == PHY_INTERFACE_MODE_QUSGMII);
> > +}  
>
> Maybe linux/phy.h should provide a helper, something like:
> 
> 	phy_interface_serdes_lanes()
> 
> that returns how many serdes lanes the interface mode uses?

Sorry about the delayed answer, I was resuming the work on this, and
realised that although a helper would be indeed great, especially for
generic PHY drivers, it won't help much in this case since
QSGMII/QUSGMII both use 1 serdes lane, as SGMII and such. If I'm not
mistaken, QSGMII is SGMII clocked at 5Gbps with a specific preamble
allowing to identify the src/dst port.

We could however imagine a helper identifying the number of links, or
lanes (or another terminology) that is carried by a given mode. I know
that besides QSGMII for 4 ports, there exists PSGMII for 5 ports, and
OSGMII for 8 ports, so this would definitely prove useful in the
future.

Sorry if this ends-up being a misunderstanding on the terminology,
we're probably already talking about the same thing, but I think that
"serdes lane" would better describe the number of physical differential
pairs that creates the link (like, 1 for SGMII, 2 for RXAUI, 4 for XAUI
and so on).

maybe something like

	phy_interface_lines() or
	phy_interface_num_ports() or simply
	phy_interface_lanes()

> > diff --git
> > a/drivers/net/ethernet/microchip/lan966x/lan966x_phylink.c
> > b/drivers/net/ethernet/microchip/lan966x/lan966x_phylink.c index
> > 38a7e95d69b4..96708352f53e 100644 ---
> > a/drivers/net/ethernet/microchip/lan966x/lan966x_phylink.c +++
> > b/drivers/net/ethernet/microchip/lan966x/lan966x_phylink.c @@
> > -28,11 +28,18 @@ static int lan966x_phylink_mac_prepare(struct
> > phylink_config *config, phy_interface_t iface) { struct
> > lan966x_port *port = netdev_priv(to_net_dev(config->dev));
> > +	phy_interface_t serdes_mode = iface;
> >  	int err;
> >  
> >  	if (port->serdes) {
> > +		/* As far as the SerDes is concerned, QUSGMII is
> > the same as
> > +		 * QSGMII.
> > +		 */
> > +		if (lan966x_is_qsgmii(iface))
> > +			serdes_mode = PHY_INTERFACE_MODE_QSGMII;
> > +
> >  		err = phy_set_mode_ext(port->serdes,
> > PHY_MODE_ETHERNET,
> > -				       iface);
> > +				       serdes_mode);  
> 
> I don't think that the ethernet MAC driver should be changing the
> interface mode before passing it down to the generic PHY layer -
> phy_set_mode_ext() is defined to take the phy interface mode, and any
> aliasing of modes should really be up to the generic PHY driver not
> the ethernet MAC driver.

Indeed, I'll split the series so that we first add support for the new
mode, and then send separate series for the generic PHY driver on one
side, and inband extensions on the other one.

Thanks,

Maxime

> Thanks.
>