[RFC,0/4] io_uring/splice: extend splice for supporting ublk zero copy

Message

Ming Lei Nov. 3, 2022, 8:50 a.m. UTC

Hello Guys,

This patch extends io_uring/splice by adding two flags(SPLICE_F_DIRECT &
SPLICE_F_READ_TO_READ) for supporting ublk zero copy, and fuse could benefit
from the change too.

- SPLICE_F_DIRECT is for using do_splice_direct() to support zero copy

- SPLICE_F_READ_TO_READ is for supporting ublk READ zero copy, the plain
splice can support WRITE zero copy by:

	splice(ublkc_fd, ublkc_pos, pipe_wr_fd, NULL, len, flags)
	splice(pipe_rd_fd, NULL, backing_fd, backing_off, len, flags)

but can't support READ zc. Extend splice to allow to splice from the 1st
->splice_read()(producer) to the 2nd ->splice_read()(consumer), then READ
request pages reference can be transferred to backing IO code path.

The initial idea is suggested by Miklos Szeredi & Stefan Hajnoczi.

The patchset has been verified basically by ublk builtin tests(loop/008,
loop/009, generic/003), and basic mount, git clone, kernel building, umount
tests on ublk-loop[1] which is created by 'ublk add -t loop -f $backing -z'.

The next step is to allow io_uring to run do_splice_direct*()
in async style like normal async RW instead of offloading to
iowq context, so that top performance can be reached, and that
depends on current work.

Any comments are welcome.

[1] https://github.com/ming1/ubdsrv/commits/splice-zc


Ming Lei (4):
  io_uring/splice: support do_splice_direct
  fs/splice: add helper of splice_dismiss_pipe()
  io_uring/splice: support splice from ->splice_read to ->splice_read
  ublk_drv: support splice based read/write zero copy

 drivers/block/ublk_drv.c      | 151 +++++++++++++++++++++++++-
 fs/read_write.c               |   5 +-
 fs/splice.c                   | 193 +++++++++++++++++++++++++++++-----
 include/linux/fs.h            |   2 +
 include/linux/pipe_fs_i.h     |   9 ++
 include/linux/splice.h        |  14 +++
 include/uapi/linux/ublk_cmd.h |  34 +++++-
 io_uring/splice.c             |  16 ++-
 8 files changed, 392 insertions(+), 32 deletions(-)

Comments

Ziyang Zhang Nov. 4, 2022, 9:15 a.m. UTC | #1

On 2022/11/3 16:50, Ming Lei wrote:
> Hello Guys,
> 
> This patch extends io_uring/splice by adding two flags(SPLICE_F_DIRECT &
> SPLICE_F_READ_TO_READ) for supporting ublk zero copy, and fuse could benefit
> from the change too.
> 
> - SPLICE_F_DIRECT is for using do_splice_direct() to support zero copy
> 
> - SPLICE_F_READ_TO_READ is for supporting ublk READ zero copy, the plain
> splice can support WRITE zero copy by:
> 
> 	splice(ublkc_fd, ublkc_pos, pipe_wr_fd, NULL, len, flags)
> 	splice(pipe_rd_fd, NULL, backing_fd, backing_off, len, flags)
> 
> but can't support READ zc. Extend splice to allow to splice from the 1st
> ->splice_read()(producer) to the 2nd ->splice_read()(consumer), then READ
> request pages reference can be transferred to backing IO code path.
> 
> The initial idea is suggested by Miklos Szeredi & Stefan Hajnoczi.
> 
> The patchset has been verified basically by ublk builtin tests(loop/008,
> loop/009, generic/003), and basic mount, git clone, kernel building, umount
> tests on ublk-loop[1] which is created by 'ublk add -t loop -f $backing -z'.
> 
> The next step is to allow io_uring to run do_splice_direct*()
> in async style like normal async RW instead of offloading to
> iowq context, so that top performance can be reached, and that
> depends on current work.
> 
> Any comments are welcome.
> 
> [1] https://github.com/ming1/ubdsrv/commits/splice-zc
> 
> 
> Ming Lei (4):
>   io_uring/splice: support do_splice_direct
>   fs/splice: add helper of splice_dismiss_pipe()
>   io_uring/splice: support splice from ->splice_read to ->splice_read
>   ublk_drv: support splice based read/write zero copy
> 
>  drivers/block/ublk_drv.c      | 151 +++++++++++++++++++++++++-
>  fs/read_write.c               |   5 +-
>  fs/splice.c                   | 193 +++++++++++++++++++++++++++++-----
>  include/linux/fs.h            |   2 +
>  include/linux/pipe_fs_i.h     |   9 ++
>  include/linux/splice.h        |  14 +++
>  include/uapi/linux/ublk_cmd.h |  34 +++++-
>  io_uring/splice.c             |  16 ++-
>  8 files changed, 392 insertions(+), 32 deletions(-)
>

Hi, Ming

I have quickly scanned the code. It seems like biovec pages are successfully
passed to the backend for READ/WRITE. I will learn your code and run some
tests. I will give more feedback next week.

Regards,
Zhang