On Tue, Jan 15, 2019 at 08:17:37PM +0200, Leon Romanovsky wrote:
> On Tue, Jan 15, 2019 at 09:53:13AM -0700, Jason Gunthorpe wrote:
> > On Tue, Jan 15, 2019 at 12:56:10PM +0200, Leon Romanovsky wrote:
> > > > > - ret = xa_insert(xa, res_to_id(res), res, GFP_KERNEL);
> > > > > - WARN_ONCE(ret == -EEXIST, "Tried to add non-unique type %d entry\n",
> > > > > - res->type);
> > > > > + if (rt[res->type].max) {
> > > > > + /* Not HW-capable device */
> > > > > + res->id = rt[res->type].reserved;
> > > > > + ret = xa_alloc(xa, &res->id, rt[res->type].max, res,
> > > > > + GFP_KERNEL);
> > > >
> > > > I feel like the SW ids should by allocated cyclicly?
> > >
> > > It was my initial thought too, but after I saw that most of the drivers
> > > are using find_first_bit()/find_first_zero_bit() in various bitmap,
> > > I realized that adding complexity for cycling allocation doesn't worth it.
> >
> > Those drivers didn't contemplate exposing these numbers of net link to
> > an agent that isn't part of the create/destroy cycle.
> >
> > Using cyclic is the best way to avoid various races and confusion for
> > the netlink side, and it is straightforward to do.
>
> No problem, I'll change it and will drop that patch with "user" field.
Actually after thinking more, in order to implement this cyclic
allocator, I will need to store next index and update it after every
xa_alloc call and it means that we will need to add lock to maintain
that index.
Are you sure that we really need it?
Thanks
>
> >
> > Jason
Attachment:
signature.asc
Description: PGP signature
