Re: [RFC PATCHv2 1/3] Add support IORING_SETUP_SQE_MIXED

[Caleb Sander Mateos <csander@purestorage.com>]

On Thu, Sep 4, 2025 at 12:27 PM Keith Busch <kbusch@meta.com> wrote:
>
> From: Keith Busch <kbusch@kernel.org>
>
> This adds core support for mixed sized SQEs in the same SQ ring. Before
> this, SQEs were either 64b in size (the normal size), or 128b if
> IORING_SETUP_SQE128 was set in the ring initialization. With the mixed
> support, an SQE may be either 64b or 128b on the same SQ ring. If the
> SQE is 128b in size, then IOSQE_SQE_128B will be set in the sqe flags.
> The client may post a NOP SQE with IOSQE_CQE_SKIP_SUCCESS set that the
> kernel should simply ignore as it's just a pad filler that is posted
> when required.
>
> Signed-off-by: Keith Busch <kbusch@kernel.org>
> ---
>  src/include/liburing.h          | 31 +++++++++++++++++++++++++++++++
>  src/include/liburing/io_uring.h |  9 +++++++++
>  2 files changed, 40 insertions(+)
>
> diff --git a/src/include/liburing.h b/src/include/liburing.h
> index 7ea876e1..97c70fa7 100644
> --- a/src/include/liburing.h
> +++ b/src/include/liburing.h
> @@ -1853,6 +1853,37 @@ IOURINGINLINE struct io_uring_sqe *_io_uring_get_sqe(struct io_uring *ring)
>         return sqe;
>  }
>
> +IOURINGINLINE struct io_uring_sqe *io_uring_get_sqe128_mixed(struct io_uring *ring)
> +       LIBURING_NOEXCEPT
> +{
> +       struct io_uring_sq *sq = &ring->sq;
> +       unsigned head = io_uring_load_sq_head(ring), tail = sq->sqe_tail;
> +       struct io_uring_sqe *sqe;
> +
> +       if (!(ring->flags & IORING_SETUP_SQE_MIXED))
> +               return NULL;
> +
> +       if ((tail & sq->ring_mask) + 1 == sq->ring_entries) {

The condition you used on the kernel side is probably a bit more efficient:
(tail + 1) & sq->ring_mask == 0

> +               sqe = _io_uring_get_sqe(ring);
> +               if (!sqe)
> +                       return NULL;
> +
> +               io_uring_prep_nop(sqe);
> +               sqe->flags |= IOSQE_CQE_SKIP_SUCCESS;
> +               tail = sq->sqe_tail;
> +       }
> +
> +       if ((tail + 1) - head >= sq->ring_entries)
> +               return NULL;

Would it make sense to check for a full SQ before creating a NOP SQE
to avoid wasted work if the actual SQE can't be posted?

Best,
Caleb

> +
> +       sqe = &sq->sqes[tail & sq->ring_mask];
> +       sq->sqe_tail = tail + 2;
> +       io_uring_initialize_sqe(sqe);
> +       sqe->flags |= IOSQE_SQE_128B;
> +
> +       return sqe;
> +}
> +
>  /*
>   * Return the appropriate mask for a buffer ring of size 'ring_entries'
>   */
> diff --git a/src/include/liburing/io_uring.h b/src/include/liburing/io_uring.h
> index 643514e5..fd02fa52 100644
> --- a/src/include/liburing/io_uring.h
> +++ b/src/include/liburing/io_uring.h
> @@ -126,6 +126,7 @@ enum io_uring_sqe_flags_bit {
>         IOSQE_ASYNC_BIT,
>         IOSQE_BUFFER_SELECT_BIT,
>         IOSQE_CQE_SKIP_SUCCESS_BIT,
> +       IOSQE_SQE_128B_BIT,
>  };
>
>  /*
> @@ -145,6 +146,8 @@ enum io_uring_sqe_flags_bit {
>  #define IOSQE_BUFFER_SELECT    (1U << IOSQE_BUFFER_SELECT_BIT)
>  /* don't post CQE if request succeeded */
>  #define IOSQE_CQE_SKIP_SUCCESS (1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
> +/* this is a 128b/big-sqe posting */
> +#define IOSQE_SQE_128B          (1U << IOSQE_SQE_128B_BIT)
>
>  /*
>   * io_uring_setup() flags
> @@ -211,6 +214,12 @@ enum io_uring_sqe_flags_bit {
>   */
>  #define IORING_SETUP_CQE_MIXED         (1U << 18)
>
> +/*
> + *  Allow both 64b and 128b SQEs. If a 128b SQE is posted, it will have
> + *  IOSQE_SQE_128B set in sqe->flags.
> + */
> +#define IORING_SETUP_SQE_MIXED         (1U << 19)
> +
>  enum io_uring_op {
>         IORING_OP_NOP,
>         IORING_OP_READV,
> --
> 2.47.3
>