On Thu, Aug 27, 2020 at 04:10:27PM +0300, Gal Pressman wrote: > On 27/08/2020 16:01, Jason Gunthorpe wrote: > > On Thu, Aug 27, 2020 at 03:58:23PM +0300, Gal Pressman wrote: > >> On 27/08/2020 15:20, Håkon Bugge wrote: > >>>> On 27 Aug 2020, at 14:01, Jason Gunthorpe <jgg@xxxxxxxx> wrote: > >>>> On Thu, Aug 27, 2020 at 11:20:16AM +0300, Gal Pressman wrote: > >>>>> On 27/08/2020 10:53, Kamal Heib wrote: > >>>>>> On Thu, Aug 20, 2020 at 04:53:38PM +0300, Kamal Heib wrote: > >>>>>>> On Thu, Aug 20, 2020 at 04:11:23PM +0300, Gal Pressman wrote: > >>>>>>>> On 20/08/2020 15:53, Kamal Heib wrote: > >>>>>>>>> Now that the query_pkey() isn't mandatory by the RDMA core, this > >>>>>>>>> callback can be removed from the usnic provider. > >>>>>>>> > >>>>>>>> Not directly related to this patch, but pyverbs has a test which verifies that > >>>>>>>> max_pkeys > 0, maybe this check should be removed. > >>>>>>> > >>>>>>> Or changed to work only for node_type == e.IBV_NODE_CA? > >>>>>>> > >>>>>>> Thanks, > >>>>>>> Kamal > >>>>>> > >>>>>> BTW, do the efa care about pkey? > >>>>> > >>>>> Depends.. We only support the default pkey so it doesn't do much in terms of > >>>>> functionality, but we still need to support it as part of the QP state machine > >>>>> for modify QP. > >>>> > >>>> Does the pkey appear on the wire, or is it just some cruft for API sake? > >>> > >>> On the wire. Included in the BTH (Base Transfer Header). > >> > >> He was probably asking specifically about EFA. > >> I can't share any details about the wire protocol, does it matter? > > > > If it isn't actually used for anything then the driver shouldn't > > expose PKEY at all, if you do use it then leave it. > > How would that work? How can you not expose pkeys and still have ibv_ud_pingpong > work? > You mean remove query_pkey, but still support IBV_QP_PKEY_INDEX in modify_qp? iWarp doesn't use PKEY at all, so you'd have to follow their model. Some of the userspace that assumes IB would probably need fixing (eg ud pingpong does not support iwarp) Jason
