Re: [RFC PATCHv2 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

[Keith Busch <kbusch@kernel.org>]

On Wed, Sep 17, 2025 at 08:44:13AM -0600, Jens Axboe wrote:
> On 9/11/25 7:07 AM, Ming Lei wrote:
> > On Thu, Sep 11, 2025 at 9:02?PM Keith Busch <kbusch@kernel.org> wrote:
> >>
> >> On Thu, Sep 11, 2025 at 10:19:06AM +0800, Ming Lei wrote:
> >>> On Thu, Sep 11, 2025 at 10:11:47AM +0800, Ming Lei wrote:
> >>>> SQE128 is used for uring_cmd only, so it could be one uring_cmd
> >>>> private flag. However, the implementation may be ugly and fragile.
> >>>
> >>> Or in case of IORING_SETUP_SQE_MIXED, IORING_OP_URING_CMD is always interpreted
> >>> as plain 64bit SQE, also add IORING_OP_URING_CMD128 for SQE128 only.
> >>
> >> Maybe that's good enough, but I was looking for more flexibility to have
> >> big SQEs for read/write too. Not that I have a strong use case for it
> >> now, but in hindsight, that's where "io_uring_attr_pi" should have been
> >> placed instead of outide the submission queue.
> > 
> > Then you can add READ128/WRITE128...
> 
> Yeah, I do think this is the best approach - make it implied by the
> opcode. Doesn't mean we have to bifurcate the whole opcode space,
> as generally not a lot of opcodes will want/need an 128b SQE.
> 
> And it also nicely solves the issue of needing to solve the flags space
> issue.
> 
> So maybe spin a v3 with that approach?

Yep, almost got it ready. I had to introduce NOP128 because that's a
very convenient op for testing. I hope that's okay.