On Fri, Apr 25, 2025 at 03:46:06PM +0530, Abhijit Gangurde wrote:
> On 4/24/25 18:38, Jason Gunthorpe wrote:
> > On Wed, Apr 23, 2025 at 03:59:07PM +0530, Abhijit Gangurde wrote:
> > > +static int ionic_aux_probe(struct auxiliary_device *adev,
> > > + const struct auxiliary_device_id *id)
> > > +{
> > > + struct ionic_aux_dev *ionic_adev;
> > > + struct net_device *ndev;
> > > + struct ionic_ibdev *dev;
> > > +
> > > + ionic_adev = container_of(adev, struct ionic_aux_dev, adev);
> > > + ndev = ionic_api_get_netdev_from_handle(ionic_adev->handle);
> > It must not do this, the net_device should not go into the IB driver,
> > like this that will create a huge complex tangled mess.
> >
> > The netdev(s) come in indirectly through the gid table and through the
> > net notifiers and ib_device_set_netdev() and they should only be
> > touched in paths dealing with specific areas.
> >
> > So don't use things like netdev_err, we have ib_err/dev_err and
> > related instead for IB drivers to use.
>
> Sure. Will remove storing of net_device in the IB driver and its
> references in the next spin. Will wait for some more feedback
> before rolling out v2.
The problem is that coupling with net_device is so distracting that
both of us are not really invested time into deep review of this series.
Another problematic pattern is usage of "void *handle" to convey
information between aux devices. Please use struct pointer instead of
void for that.
Thanks
>
> Thanks,
> Abhijit
>
> >
> > > +struct ionic_ibdev {
> > > + struct ib_device ibdev;
> > > +
> > > + struct device *hwdev;
> > > + struct net_device *ndev;
> > Same here, this member should not exist, and it didn't hold a
> > refcount for this pointer.
> >
> > Jason
