[RFC PATCHv2 0/1]

[Keith Busch <kbusch@meta.com>]

From: Keith Busch <kbusch@kernel.org>

The CQ supports mixed size entries, so why not do it for SQ's too? There
are use cases that currently allocate different queues just to keep
these things separated, but we can efficiently handle both cases in a
single ring.

This RFC is taking the last SQE flags bit for the purpose. That might
not be okay, but serves as a proof of concept for an initial look.

Changes from v1:

  - Fixed up the kernel's handling for big SQEs on a mixed SQ for
    uring_cmd and fdinfo

  - Added liburing tests for nvme uring_cmd usage and inserting a 128b
    SQE in the last entry to test the kernel's handling for a bad wrap.

kernel:

Keith Busch (1):
  io_uring: add support for IORING_SETUP_SQE_MIXED

 include/linux/io_uring_types.h |  3 +++
 include/uapi/linux/io_uring.h  |  9 +++++++++
 io_uring/fdinfo.c              | 32 +++++++++++++++++++++++++-------
 io_uring/io_uring.c            | 20 +++++++++++++++++++-
 io_uring/register.c            |  2 +-
 io_uring/uring_cmd.c           |  4 +++-
 6 files changed, 60 insertions(+), 10 deletions(-)

liburing:

Keith Busch (3):
  Add support IORING_SETUP_SQE_MIXED
  Add nop testing for IORING_SETUP_SQE_MIXED
  Add mixed sqe test for uring commands

 src/include/liburing.h          |  31 ++++++++
 src/include/liburing/io_uring.h |   9 +++
 test/Makefile                   |   3 +
 test/sqe-mixed-bad-wrap.c       | 121 +++++++++++++++++++++++++++++
 test/sqe-mixed-nop.c            | 104 +++++++++++++++++++++++++
 test/sqe-mixed-uring_cmd.c      | 168 +++++++++++++++++++++++++++++++++++++++++
 6 files changed, 436 insertions(+)

-- 
2.47.3