Re: [PATCH 2/3] RDMA/hfi1, rdmavt: open-code rvt_set_ibdev_name()

[Dennis Dalessandro]

On 3/23/26 7:01 AM, Leon Romanovsky wrote:
> On Mon, Mar 23, 2026 at 09:48:59AM +0100, Arnd Bergmann wrote:
> > On Mon, Mar 23, 2026, at 09:08, Leon Romanovsky wrote:
> > > On Fri, Mar 20, 2026 at 04:53:04PM +0100, Arnd Bergmann wrote:
> > > > On Fri, Mar 20, 2026, at 16:12, Arnd Bergmann wrote:
> > > >
> > > > > +	 */
> > > > > +	ibdev = &dd->verbs_dev.rdi.ibdev;
> > > > > +	dev_set_name(&ibdev->dev, "%s_%d", class_name(), dd->unit);
> > > > > +	strscpy(&ibdev->name, dev_name(&ibdev->dev), IB_DEVICE_NAME_MAX);
> > > > > +
> > > >
> > > > I messed this up during a rebase, that should have been
> > > >
> > > >         strscpy(ibdev->name, dev_name(&ibdev->dev), IB_DEVICE_NAME_MAX);
> > > >
> > > > (without the extra &). I'll wait for comments before resending.
> > >
> > > The hfi1 driver is scheduled for removal. Dennis has already posted the
> > > hfi2 driver, which serves as its replacement.
> >
> > Ok, that does sound like a sensible decision, and I'll just drop
> > patches 1 and 3 then, which are just cleanups.
> >
> > The cover letter at [1] suggests that the two drivers will still
> > coexist for a bit though, so I think we'd still want patch 2/3
> > in order to get a clean 'allmodconfig' build when the
> > -Wmissing-format-attribute is enabled by defaultt. I have a couple
> > of patches in flight.
>
> Sure, builds need to be fixed.
>
> > I took a quick look at the hfi2 driver, and noticed a few things
> > that that may be worth addressing before it gets merged, mostly
> > stuff copied from hfi1:
> >
> > - A few global functions with questionable namespacing:
> >    user_event_ack, ctxt_reset, iowait_init, register_pinning_interface,
> >    sc_{alloc,free,enable,disable}, pio_copy, acquire_hw_mutex,
> >    load_firmware, cap_mask.
> >    It would make sense to prefix all global identifiers with 'hfi2_',
> >    both out of principle, and to allow building hfi1 and hfi2 into
> >    an allyesconfig kernel without link failures.
> >
> > - The use of INFINIBAND_RDMAVT seems unnecessary: right now
> >    this is only used by hfi1, now shared with hfi2 but later to
> >    be exclusive to the latter. Since it is unlikely to ever
> >    be used by another driver again, this may be a good time
> >    to drop the abstraction again and integrate it all into
> >    hfi2, with the old version getting dropped along with hfi1.
>
> The best approach is to drop rdmavt as well, since hfi2 is expected to
> align with the other drivers in drivers/infiniband/hw.
>
> Dennis, is this feasible?

Feasible yes. I'd like to get hfi2 crossed off the list and in the tree 
first though. Then come back and do that. I'd like to do more than just 
plop rdmavt inside hfi2 and call it a day. There is a lot of code 
cleanup/simplification that we can do.

-Denny