In case of failure halfway through the operation we get very different results depending upon the iov_iter flavour: iovec, pipe - advances by the amount actually copied, kvec, bvec - does *NOT* advance at all Which semantics is desired? AFAICS, the calls can be repeated - e.g. the loop in dax_iomap_actor() will call dax_copy_to_iter() again on the short read and with iovec-backed iter it will try to copy from the place of failure (presumably returning 0 that time around and terminating the loop), while with bvec or kvec it will go and paste the copies of the same chunk again until it runs out of destination.