On Wed, 29 Sep 2021 18:31:51 +0300 Leon Romanovsky wrote: > On Wed, Sep 29, 2021 at 07:39:40AM -0700, Jakub Kicinski wrote: > > On Wed, 29 Sep 2021 17:13:28 +0300 Leon Romanovsky wrote: > > > We don't need to advertise counters for feature that is not supported. > > > In multiport mlx5 devices, the reload functionality is not supported, so > > > this change at least make that device to behave like all other netdev > > > devices that don't support devlink reload. > > > > > > The ops structure is set very early to make sure that internal devlink > > > routines will be able access driver back during initialization (btw very > > > questionable design choice) > > > > Indeed, is this fixable? Or now that devlink_register() was moved to > > the end of probe netdev can call ops before instance is registered? > > > > > and at that stage the driver doesn't know > > > yet which device type it is going to drive. > > > > > > So the answer is: > > > 1. Can't have two structures. > > > > I still don't understand why. To be clear - swapping full op structures > > is probably acceptable if it's a pure upgrade (existing pointers match). > > Poking new ops into a structure (in alphabetical order if I understand > > your reply to Greg, not destructor-before-contructor) is what I deem > > questionable. > > It is sorted simply for readability and not for any other technical > reason. > > Regarding new ops, this is how we are setting callbacks in RDMA based on > actual device support. It works like a charm. > > > > 2. Same behaviour across all netdev devices. > > > > Unclear what this is referring to. > > If your device doesn't support devlink reload, it won't print any > reload counters at all. It is not the case for the multiport mlx5 > device. It doesn't support, but still present these counters. There's myriad ways you can hide features. Swapping ops is heavy handed and prone to data races, I don't like it.
