> -----Original Message-----
> From: Michal Kubecek <mkube...@suse.cz>
> Sent: Wednesday, November 25, 2020 12:12 AM
> To: Danielle Ratson <daniel...@nvidia.com>
> Cc: Jiri Pirko <j...@resnulli.us>; Andrew Lunn <and...@lunn.ch>; Jakub
> Kicinski <k...@kernel.org>; Ido Schimmel
> <ido...@idosch.org>; netdev@vger.kernel.org; da...@davemloft.net; Jiri Pirko
> <j...@nvidia.com>; f.faine...@gmail.com; mlxsw
> <ml...@nvidia.com>; Ido Schimmel <ido...@nvidia.com>;
> johan...@sipsolutions.net
> Subject: Re: [PATCH net-next 1/6] ethtool: Extend link modes settings uAPI
> with lanes
>
> On Mon, Nov 23, 2020 at 09:47:53AM +0000, Danielle Ratson wrote:
> >
> >
> > > -----Original Message-----
> > > From: Michal Kubecek <mkube...@suse.cz>
> > > Sent: Thursday, October 22, 2020 7:28 PM
> > > To: Danielle Ratson <daniel...@nvidia.com>
> > > Cc: Jiri Pirko <j...@resnulli.us>; Andrew Lunn <and...@lunn.ch>;
> > > Jakub Kicinski <k...@kernel.org>; Ido Schimmel <ido...@idosch.org>;
> > > netdev@vger.kernel.org; da...@davemloft.net; Jiri Pirko
> > > <j...@nvidia.com>; f.faine...@gmail.com; mlxsw <ml...@nvidia.com>;
> > > Ido Schimmel <ido...@nvidia.com>; johan...@sipsolutions.net
> > > Subject: Re: [PATCH net-next 1/6] ethtool: Extend link modes
> > > settings uAPI with lanes
> > >
> > > On Thu, Oct 22, 2020 at 06:15:48AM +0000, Danielle Ratson wrote:
> > > > > -----Original Message-----
> > > > > From: Michal Kubecek <mkube...@suse.cz>
> > > > > Sent: Wednesday, October 21, 2020 11:48 AM
> > > > >
> > > > > Ah, right, it does. But as you extend struct
> > > > > ethtool_link_ksettings and drivers will need to be updated to
> > > > > provide this information, wouldn't it be more useful to let the
> > > > > driver provide link mode in use instead (and derive number of lanes
> > > > > from it)?
> > > >
> > > > This is the way it is done with the speed parameter, so I have
> > > > aligned it to it. Why the lanes should be done differently
> > > > comparing to the speed?
> > >
> > > Speed and duplex have worked this way since ages and the interface
> > > was probably introduced back in times when combination of speed and
> > > duplex was sufficient to identify the link mode. This is no longer
> > > the case and even adding number of lanes wouldn't make the combination
> > > unique. So if we are going to extend the interface now
> and update drivers to provide extra information, I believe it would be more
> useful to provide full information.
> > >
> > > Michal
> >
> > Hi Michal,
> >
> > What do you think of passing the link modes you have suggested as a
> > bitmask, similar to "supported", that contains only one positive bit?
> > Something like that:
Hi Michal,
Thanks for your response.
Actually what I said is not very accurate.
In ethtool, for speed 100G and 4 lanes for example, there are few link modes
that fits:
ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT
ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT
ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT
ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT
The difference is the media. And in the driver we shrink into one bit.
But maybe that makes passing a bitmask more sense, or am I missing something?
> >
> > diff --git a/include/linux/ethtool.h b/include/linux/ethtool.h index
> > afae2beacbc3..dd946c88daa3 100644
> > --- a/include/linux/ethtool.h
> > +++ b/include/linux/ethtool.h
> > @@ -127,6 +127,7 @@ struct ethtool_link_ksettings {
> > __ETHTOOL_DECLARE_LINK_MODE_MASK(supported);
> > __ETHTOOL_DECLARE_LINK_MODE_MASK(advertising);
> > __ETHTOOL_DECLARE_LINK_MODE_MASK(lp_advertising);
> > + __ETHTOOL_DECLARE_LINK_MODE_MASK(chosen);
> > } link_modes;
> > u32 lanes;
> > };
> >
> > Do you have perhaps a better suggestion?
>
> Not sure if it's better but as we know there is only one mode, we could
> simply pass the index. We would still need to reserve a special
> value for none/unknown but getting an index would make lookup easier.
>
> > And the speed and duplex parameters should be removed from being
> > passed like as well, right?
>
> We cannot remove them from struct ethtool_link_settings and the ioctl and
> netlink messages as those are part of UAPI and we have
> to preserve backward compatibility. But drivers which provide link mode would
> not need to fill speed and duplex in their -
> >get_link_ksettings() as the common code could do that for them.
Yes of course I didn't mean to remove the parameters from the struct, just to
not prepare them for passing to ethtool when getting the link settings.
Thanks.
>
> Michal
