Hello Jens,
This patch tries to address limitation from in-tree ublk zero copy:
- one IO needs two extra uring_cmd for register/unregister bvec buffer, not
efficient
- introduced dependency on the two uring_cmd, so the buffer consumer has to
linked with the two uring_cmd, hard to use & less efficient
This patchset adds feature UBLK_F_AUTO_BUF_REG:
- register request buffer automatically before delivering io command to ublk server
- unregister request buffer automatically when completing the request
- both io_uring context and buffer index can be specified from ublk
uring_cmd header
With this way, 'fio/t/io_uring -p0 /dev/ublkb0' shows that IOPS is improved
by 50% compared with F_SUPPORT_ZERO_COPY in my test VM.
kernel selftests are added for covering both function & stress test.
Please review & comment!
Thanks,
Ming
Ming Lei (7):
io_uring: add 'struct io_buf_data' for register/unregister bvec buffer
io_uring: add helper __io_buffer_[un]register_bvec
io_uring: support to register bvec buffer to specified io_uring
ublk: convert to refcount_t
ublk: prepare for supporting to register request buffer automatically
ublk: register buffer to specified io_uring & buf index via
UBLK_F_AUTO_BUF_REG
selftests: ublk: support UBLK_F_AUTO_BUF_REG
drivers/block/ublk_drv.c | 165 ++++++++++++++----
include/linux/io_uring/cmd.h | 15 +-
include/uapi/linux/ublk_cmd.h | 38 ++++
io_uring/rsrc.c | 141 ++++++++++-----
tools/testing/selftests/ublk/Makefile | 2 +
tools/testing/selftests/ublk/file_backed.c | 12 +-
tools/testing/selftests/ublk/kublk.c | 24 ++-
tools/testing/selftests/ublk/kublk.h | 6 +
tools/testing/selftests/ublk/null.c | 43 +++--
tools/testing/selftests/ublk/stripe.c | 21 +--
.../testing/selftests/ublk/test_generic_08.sh | 32 ++++
.../testing/selftests/ublk/test_stress_03.sh | 6 +
.../testing/selftests/ublk/test_stress_04.sh | 6 +
.../testing/selftests/ublk/test_stress_05.sh | 8 +
14 files changed, 409 insertions(+), 110 deletions(-)
create mode 100755 tools/testing/selftests/ublk/test_generic_08.sh
--
2.47.0
