On Thu, May 30, 2019 at 09:58:53PM +0000, Adit Ranadive wrote: > That's interesting. So the guest would still do a ibv_reg_mr and the device > would re-send the mappings using ibv_reg_mr_iova? > How would you prevent applications from using/misusing ibv_reg_mr_iova by > passing arbitrary values to hca_va or is that even a problem? It is not a problem, the iova & mkey is the unique identifier for the memory. There is nothing that says iova has to be anything special. > What if the underlying device doesn't support zero-based MRs? All devices have to support iova to handle the existing user cases. Thus all devices also support a 0 iova. The foolishness was making zero-based into something special at the driver level. The core code should have handled it. Jason
