Re: [PATCH v2 0/5] io_uring: avoid uring_lock for IORING_SETUP_SINGLE_ISSUER

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

On Wed, Sep 10, 2025 at 8:36 AM Jens Axboe <axboe@kernel.dk> wrote:
>
> On 9/10/25 5:57 AM, Pavel Begunkov wrote:
> > On 9/9/25 14:35, Jens Axboe wrote:
> >>
> >> On Thu, 04 Sep 2025 11:08:57 -0600, Caleb Sander Mateos wrote:
> >>> As far as I can tell, setting IORING_SETUP_SINGLE_ISSUER when creating
> >>> an io_uring doesn't actually enable any additional optimizations (aside
> >>> from being a requirement for IORING_SETUP_DEFER_TASKRUN). This series
> >>> leverages IORING_SETUP_SINGLE_ISSUER's guarantee that only one task
> >>> submits SQEs to skip taking the uring_lock mutex in the submission and
> >>> task work paths.
> >>>
> >>> [...]
> >>
> >> Applied, thanks!
> >>
> >> [1/5] io_uring: don't include filetable.h in io_uring.h
> >>        commit: 5d4c52bfa8cdc1dc1ff701246e662be3f43a3fe1
> >> [2/5] io_uring/rsrc: respect submitter_task in io_register_clone_buffers()
> >>        commit: 2f076a453f75de691a081c89bce31b530153d53b
> >> [3/5] io_uring: clear IORING_SETUP_SINGLE_ISSUER for IORING_SETUP_SQPOLL
> >>        commit: 6f5a203998fcf43df1d43f60657d264d1918cdcd
> >> [4/5] io_uring: factor out uring_lock helpers
> >>        commit: 7940a4f3394a6af801af3f2bcd1d491a71a7631d
> >> [5/5] io_uring: avoid uring_lock for IORING_SETUP_SINGLE_ISSUER
> >>        commit: 4cc292a0faf1f0755935aebc9b288ce578d0ced2
> >
> > FWIW, from a glance that should be quite broken, there is a bunch of
> > bits protected from parallel use by the lock. I described this
> > optimisation few years back around when first introduced SINGLE_ISSUER
> > and the DEFER_TASKRUN locking model, but to this day think it's not
> > worth it as it'll be a major pain for any future changes. It would've
> > been more feasible if links wasn't a thing. Though, none of it is
> > my problem anymore, and I'm not insisting.
>
> Hmm yes, was actually pondering this last night as well and was going
> to take a closer look today as I have a flight coming up. I'll leave
> them in there for now just to see if syzbot finds anything, and take
> that closer look and see if it's salvageable for now or if we just need
> a new revised take on this.

Is the concern the various IO_URING_F_UNLOCKED contexts (e.g. io_uring
worker threads) relying on uring_lock to synchronize access to the
io_ring_ctx with submitter_task? I think it would be possible to
provide mutual exclusion in those contexts using a task work item to
suspend submitter_task. When submitter_task picks up the task work, it
can unblock the thread running in IO_URING_F_UNLOCKED context, which
can then take the uring_lock as usual. Once it releases the
uring_lock, it can unblock submitter_task.
This approach could certainly add latency to taking uring_lock in
IO_URING_F_UNLOCKED contexts, though I don't expect that is very
common in applications using io_uring. We could certainly add a new
setup flag to avoid changing the behavior for existing
IORING_SETUP_SINGLE_ISSUER users. What are your thoughts on this
approach?

Thanks,
Caleb