On Mon, Jul 06, 2020 at 08:04:16PM -0300, Jason Gunthorpe wrote:
> On Tue, Jun 30, 2020 at 01:18:53PM +0300, Leon Romanovsky wrote:
> > @@ -916,21 +916,24 @@ static int ib_uverbs_alloc_mw(struct uverbs_attr_bundle *attrs)
> > goto err_put;
> > }
> >
> > - mw = pd->device->ops.alloc_mw(pd, cmd.mw_type, &attrs->driver_udata);
> > - if (IS_ERR(mw)) {
> > - ret = PTR_ERR(mw);
> > + mw = rdma_zalloc_drv_obj(ib_dev, ib_mw);
> > + if (!mw) {
> > + ret = -ENOMEM;
> > goto err_put;
> > }
> >
> > - mw->device = pd->device;
> > - mw->pd = pd;
> > + mw->device = ib_dev;
> > + mw->pd = pd;
> > mw->uobject = uobj;
> > - atomic_inc(&pd->usecnt);
> > -
> > uobj->object = mw;
> > + mw->type = cmd.mw_type;
> >
> > - memset(&resp, 0, sizeof(resp));
> > - resp.rkey = mw->rkey;
> > + ret = pd->device->ops.alloc_mw(mw, &mw->rkey, &attrs->driver_udata);
>
> Why the strange &mw->rkey ? Can't the drivers just do mw->rkey = foo ?
We can, if we want to allow drivers set fields in ib_* structures that
there passed as part of alloc_* flows. It doesn't feel right to me to
mix different layers.
Thanks
>
> Jason
