Re: [PATCH 3/4] ublk: add feature UBLK_F_AUTO_ZERO_COPY

[Ming Lei <ming.lei@redhat.com>]

On Sat, Apr 26, 2025 at 08:34:40PM -0600, Keith Busch wrote:
> On Sat, Apr 26, 2025 at 03:42:59PM -0700, Caleb Sander Mateos wrote:
> > On Sat, Apr 26, 2025 at 2:41 AM Ming Lei <ming.lei@redhat.com> wrote:
> > >
> > > UBLK_F_SUPPORT_ZERO_COPY requires ublk server to issue explicit buffer
> > > register/unregister uring_cmd for each IO, this way is not only inefficient,
> > > but also introduce dependency between buffer consumer and buffer register/
> > > unregister uring_cmd, please see tools/testing/selftests/ublk/stripe.c
> > > in which backing file IO has to be issued one by one by IOSQE_IO_LINK.
> > 
> > This is a great idea!
> 
> This is very similiar to something I proposed off-list, and the feedback

Looks we both think of it, :-)

> back then was this won't work because the back-end ring that wants to
> use the zero-copy buffer isn't the same as the ublk server ring
> recieving notification of a new command; the ublk driver has no idea
> which uring to register the bvec with. Also, this is using the request
> "tag" as the io_uring buf index, which wouldn't work when the ublk
> server ring handles multiple ublk devices due to the tag collisions.
> 
> If you're can make those trade-offs, then this is a great simplification
> to the whole thing.

The io_uring fd & buffer index can be provided from 'ublksrv_io_cmd'.

https://lore.kernel.org/linux-block/aA2RNG3-WzuQqEN6@fedora/

If we only support IORING_ENTER_REGISTERED_RING, 32bit is enough for
io_uring fd & buffer index, and there is still 64bits available if not
taking UBLK_F_ZONED into account.


Thanks, 
Ming