On Mon, Aug 19, 2019 at 02:17:44PM -0700, Sagi Grimberg wrote: > > >>>> ----- On 16 Aug, 2019, at 15:16, Christoph Hellwig h...@lst.de wrote: > >>>>> Sorry for not replying to the earlier version, and thanks for doing > >>>>> this work. > >>>>> > >>>>> I wonder if instead of using our own structure we'd just use > >>>>> a full nvme SQE for the input and CQE for that output. Even if we > >>>>> reserve a few fields that means we are ready for any newly used > >>>>> field (at least until the SQE/CQE sizes are expanded..). > >>>> > >>>> We could do that, nvme_command and nvme_completion are already UAPI. > >>>> On the other hand that would mean not filling out certain fields like > >>>> command_id. Can do an approach like this. > >>> > >>> Well, we need to pass user space addresses and lengths, which isn't > >>> captured in struct nvme_command. > >> > >> Isn't simply having a 64 variant simpler? > > > > Could you provide more details on what you mean by this? > > Why would we need to pass addresses and lengths if userspace is > sending the 64 variant when it is expecting a 64 result? > > Or maybe I'm missing something...
The recommendation was to have user space provide an SQE, i.e. 'struct nvme_command', as input to the driver and receive 'struct nvme_completion' in response. I am only pointing out that 'struct nvme_command' is inappropriate for user space.