On Sun, Dec 09, 2018 at 11:26:02AM +0200, Gal Pressman wrote: > On 06-Dec-18 18:09, Jason Gunthorpe wrote: > > On Thu, Dec 06, 2018 at 12:40:26PM +0200, Gal Pressman wrote: > > > >> The opaque handle is reported in the RX work requests completions; > >> in case the address handle was not yet registered, the device > >> reports the GID in an extended completion entry. Therefore, we do > >> not need the 40 byte offset memory layout. > > > > If the protocol doesn't have the 40 byte offset then it isn't > > IBV_QPT_UD, and you'll need to also use a driver QP for this as well. > > > > Jason > > > > I need to figure out the exact details, but we can probably configure the device > to do the 40 bytes offset, although we don't really need it. > > Is this the only thing we're missing for the IBV_QPT_UD? You're asking me? It is your responsibility to make sure the device implements all mandatory IBTA verbs semantics for using IBV_QPT_UD Jason
