On October 12, 2017 8:04:19 AM PDT, Jiri Pirko <j...@resnulli.us> wrote: >Thu, Oct 12, 2017 at 04:46:24PM CEST, ro...@cumulusnetworks.com wrote: >>On Thu, Oct 12, 2017 at 7:40 AM, Jiri Pirko <j...@resnulli.us> wrote: >>> Thu, Oct 12, 2017 at 04:35:10PM CEST, ro...@cumulusnetworks.com >wrote: >>>>On Thu, Oct 12, 2017 at 6:34 AM, Steve Lin ><steven.l...@broadcom.com> wrote: >>>>> Adds a devlink command for getting & setting device configuration >>>>> parameters, and enumerates a bunch of those parameters as devlink >>>>> attributes. Also introduces an attribute that can be set by a >>>>> driver to indicate that the config change doesn't take effect >>>>> until the next restart (as in the case of the bnxt driver changes >>>>> in this patchset, for which all the configuration changes affect >NVM >>>>> only, and aren't loaded until the next restart.) >>>>> >>>>> bnxt driver patches make use of these new devlink cmds/attributes. >>>>> >>>>> Steve Lin (3): >>>>> devlink: Add config parameter get/set operations >>>>> bnxt: Move generic devlink code to new file >>>>> bnxt: Add devlink support for config get/set >>>>> >>>> >>>>Is the goal here to move all ethtool operations to devlink (I saw >some >>>>attrs related to speed etc). ?. >>>>We do need to move ethtool attrs to netlink and devlink is a good >>>>place (and of-course leave the current ethtool api around for >backward >>>>compatibility). >>> >>> We need to make sure we are not moving things to devlink which don't >>> belong there. All options that use "netdev" as a handle should go >into >>> rtnetlink instead. >>> >> >>Any reason you want to keep that restriction ?. >>FWIS, devlink is a driver api just like ethtool is. >>and ethtool needs to move to netlink soon...and It would be better to >>not put the rtnl_lock burden on ethtool driver operations. Instead of >>adding yet another driver api, extending devlink seems like a great >>fit to me. > >Hmm, the original purpose of devlink was to obtain iface for things >that >could not use "netdev" as a handle. I try to stick with it as we >already >have iface for things that could use "netdev" as a handle - rtnetlink. > >Not sure we want to go this way and add "netdev"-handle things into >devlink. Thoughts?
In the current situation where we have ethtool and devlink operating separately on different objects as entry points to the kernel, I agree with that design. Once we move ethtool (or however we name its successor) over to netlink there is an opportunity for accessing objects that do and do not have a netdevice representor today (e.g: management ports on switches) with the same interface, and devlink could be used for that. In terms of compatibility though we should have a pseudo generic layer that can take ethtool ioctl() and transform that into a netlink message and then call back down to drivers with the existing ethtool_ops that are implemented. It is reasonably simple to use coccinelle to update these ethtool_ops with possibly updated signatures to support netlink attributes and whatnot, but forcing drivers to quit doing ethtool_ops entitely and implement new sets of "ethtool over netlink" ops is a non starter IMHO. -- Florian
