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

[Ming Lei <ming.lei@redhat.com>]

On Wed, Sep 10, 2025 at 06:28:43PM -0600, Jens Axboe wrote:
> On 9/10/25 11:44 AM, Caleb Sander Mateos wrote:
> >> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> >> index 04ebff33d0e62..9cef9085f52ee 100644
> >> --- a/include/uapi/linux/io_uring.h
> >> +++ b/include/uapi/linux/io_uring.h
> >> @@ -146,6 +146,7 @@ enum io_uring_sqe_flags_bit {
> >>         IOSQE_ASYNC_BIT,
> >>         IOSQE_BUFFER_SELECT_BIT,
> >>         IOSQE_CQE_SKIP_SUCCESS_BIT,
> >> +       IOSQE_SQE_128B_BIT,
> > 
> > Have you given any thought to how we would handle the likely scenario
> > that we want to define more SQE flags in the future? Are there
> > existing unused bytes of the SQE where the new flags could go? If not,
> > we may need to repurpose some existing but rarely used field. And then
> > we'd likely want to reserve this last flag bit to specify whether the
> > SQE is using this "extended flags" field.
> 
> Yep this is my main problem with this change. If you search the io_uring
> list on lore you can find discussions about this in relation to when
> Ming had his SQE grouping patches that also used this last bit. My
> suggestion then was indeed to have this last flag be "look at XX for
> IOSQE2_* flags". But it never quite got finalized. IIRC, my suggestion
> back then was to use the personality field, since it's a pretty
> specialized use case. Only issue with that is that you could then not
> use IOSQE2_* flags with personality.
> 
> IOW, I think the IOSQE_SQE_128B flag is fine for prototyping and testing
> these patches, but we unfortunately do need to iron out how on earth
> we'll expose some more flags before this can go in.

SQE128 is used for uring_cmd only, so it could be one uring_cmd
private flag. However, the implementation may be ugly and fragile.


Thanks,
Ming