Re: [PATCHv3 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

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

On Thu, Sep 25, 2025 at 8:03 AM Jens Axboe <axboe@kernel.dk> wrote:
>
> On 9/24/25 9:12 AM, Keith Busch wrote:
> > contiguous in the SQ ring, a 128b SQE cannot wrap the ring. For this
> > case, a single NOP SQE should be posted with the SKIP_SUCCESS flag set.
> > The kernel should simply ignore those.
>
> I think this mirrors the CQE side too much - the kernel doesn't ignore
> then, they get processed just like any other NOP that has SKIP_SUCCESS
> set. They don't post a CQE, but that's not because they are ignored,
> that's just the nature of a successful NOP w/SKIP_SUCCESS set.
>
> > @@ -2179,6 +2179,14 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
> >       opcode = array_index_nospec(opcode, IORING_OP_LAST);
> >
> >       def = &io_issue_defs[opcode];
> > +     if (def->is_128) {
> > +             if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
> > +                 (ctx->cached_sq_head & (ctx->sq_entries - 1)) == 0)
> > +                     return io_init_fail_req(req, -EINVAL);
> > +             ctx->cached_sq_head++;
> > +             (*left)--;
> > +     }
>
> This could do with a comment!
>
> > @@ -582,9 +583,10 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
> >   * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
> >   * slot.
> >   */
> > -static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
> > +static inline size_t uring_sqe_size(struct io_kiocb *req)
> >  {
> > -     if (ctx->flags & IORING_SETUP_SQE128)
> > +     if (req->ctx->flags & IORING_SETUP_SQE128 ||
> > +         req->opcode == IORING_OP_URING_CMD128)
> >               return 2 * sizeof(struct io_uring_sqe);
> >       return sizeof(struct io_uring_sqe);
>
> This one really confused me, but then I grep'ed, and it's uring_cmd
> specific. Should probably move this one to uring_cmd.c rather than have
> it elsewhere.
>
> > +int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
> > +{
> > +     if (!(req->ctx->flags & IORING_SETUP_SQE_MIXED))
> > +             return -EINVAL;
> > +     return io_uring_cmd_prep(req, sqe);
> > +}
>
> Why isn't this just allowed for SQE128 as well? There should be no
> reason to disallow explicitly 128b sqe commands in SQE128 mode, they
> should work for any mode that supports 128b SQEs which is either
> SQE_MIXED or SQE128?

Not to mention, the check in io_init_req() should already have
rejected a 128-byte operation on a non-IORING_SETUP_SQE_MIXED
io_ring_ctx.

Best,
Caleb