Re: [PATCH v5 3/5] io_uring: count CQEs in io_iopoll_check()

[Jens Axboe <axboe@kernel.dk>]

On 3/6/26 7:35 PM, Ming Lei wrote:
> On Fri, Mar 06, 2026 at 05:38:15PM -0800, Caleb Sander Mateos wrote:
>> On Wed, Mar 4, 2026 at 8:29?AM Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>> On 3/4/26 8:46 AM, Caleb Sander Mateos wrote:
>>>> On Wed, Mar 4, 2026 at 2:33?AM Ming Lei <ming.lei@redhat.com> wrote:
>>>>>
>>>>> On Mon, Mar 02, 2026 at 10:29:12AM -0700, Caleb Sander Mateos wrote:
>>>>>> A subsequent commit will allow uring_cmds that don't use iopoll on
>>>>>> IORING_SETUP_IOPOLL io_urings. As a result, CQEs can be posted without
>>>>>> setting the iopoll_completed flag for a request in iopoll_list or going
>>>>>> through task work. For example, a UBLK_U_IO_FETCH_IO_CMDS command could
>>>>>> call io_uring_mshot_cmd_post_cqe() to directly post a CQE. The
>>>>>> io_iopoll_check() loop currently only counts completions posted in
>>>>>> io_do_iopoll() when determining whether the min_events threshold has
>>>>>> been met. It also exits early if there are any existing CQEs before
>>>>>> polling, or if any CQEs are posted while running task work. CQEs posted
>>>>>> via io_uring_mshot_cmd_post_cqe() or other mechanisms won't be counted
>>>>>> against min_events.
>>>>>>
>>>>>> Explicitly check the available CQEs in each io_iopoll_check() loop
>>>>>> iteration to account for CQEs posted in any fashion.
>>>>>>
>>>>>> Signed-off-by: Caleb Sander Mateos <csander@purestorage.com>
>>>>>> ---
>>>>>>  io_uring/io_uring.c | 9 ++-------
>>>>>>  1 file changed, 2 insertions(+), 7 deletions(-)
>>>>>>
>>>>>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
>>>>>> index 46f39831d27c..b4625695bb3a 100644
>>>>>> --- a/io_uring/io_uring.c
>>>>>> +++ b/io_uring/io_uring.c
>>>>>> @@ -1184,11 +1184,10 @@ __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
>>>>>>               io_move_task_work_from_local(ctx);
>>>>>>  }
>>>>>>
>>>>>>  static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned int min_events)
>>>>>>  {
>>>>>> -     unsigned int nr_events = 0;
>>>>>>       unsigned long check_cq;
>>>>>>
>>>>>>       min_events = min(min_events, ctx->cq_entries);
>>>>>>
>>>>>>       lockdep_assert_held(&ctx->uring_lock);
>>>>>> @@ -1227,34 +1226,30 @@ static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned int min_events)
>>>>>>                * the poll to the issued list. Otherwise we can spin here
>>>>>>                * forever, while the workqueue is stuck trying to acquire the
>>>>>>                * very same mutex.
>>>>>>                */
>>>>>>               if (list_empty(&ctx->iopoll_list) || io_task_work_pending(ctx)) {
>>>>>> -                     u32 tail = ctx->cached_cq_tail;
>>>>>> -
>>>>>>                       (void) io_run_local_work_locked(ctx, min_events);
>>>>>>
>>>>>>                       if (task_work_pending(current) || list_empty(&ctx->iopoll_list)) {
>>>>>>                               mutex_unlock(&ctx->uring_lock);
>>>>>>                               io_run_task_work();
>>>>>>                               mutex_lock(&ctx->uring_lock);
>>>>>>                       }
>>>>>>                       /* some requests don't go through iopoll_list */
>>>>>> -                     if (tail != ctx->cached_cq_tail || list_empty(&ctx->iopoll_list))
>>>>>> +                     if (list_empty(&ctx->iopoll_list))
>>>>>>                               break;
>>>>>>               }
>>>>>>               ret = io_do_iopoll(ctx, !min_events);
>>>>>>               if (unlikely(ret < 0))
>>>>>>                       return ret;
>>>>>>
>>>>>>               if (task_sigpending(current))
>>>>>>                       return -EINTR;
>>>>>>               if (need_resched())
>>>>>>                       break;
>>>>>> -
>>>>>> -             nr_events += ret;
>>>>>> -     } while (nr_events < min_events);
>>>>>> +     } while (io_cqring_events(ctx) < min_events);
>>>>>
>>>>> Before entering the loop, if io_cqring_events() finds any queued CQE,
>>>>> io_iopoll_check() returns immediately without polling.
>>>>>
>>>>> If the queued CQE is originated from non-iopoll uring_cmd, iopoll request
>>>>> will not be polled, may this be one issue?
>>>>
>>>> I also noticed that logic and thought it seemed odd. I would think
>>>> we'd always want to wait for min_events CQEs (and iopoll once even if
>>>> min_events is 0). Looks like Jens added the early return in commit
>>>> a3a0e43fd770 ("io_uring: don't enter poll loop if we have CQEs
>>>> pending"), perhaps he can shed some light on it?
>>>
>>> I don't  recall the bug in question, it's been a while... But it always
>>> makes sense to return events that are ready, and skip polling. It should
>>> only be done if there are no ready events to reap.
>>
>> Ming, are you okay with preserving that behavior in this patch then? I
>> guess there's a potential fairness concern where REQ_F_IOPOLL requests
>> may not be polled for some time if non-REQ_F_IOPOLL requests continue
>> to frequently post CQEs.
> 
> IMO, the fairness may not a big deal given userspace should keep polling
> if the iopoll IO isn't done.

I think so.

> But forget to mention, if non-iopoll CQE is posted and ->cq_flush becomes
> true, io_submit_flush_completions() may not get chance to run in case of
> the early return. 
> 
> Maybe something like below is needed:
> 
> @@ -1556,8 +1556,10 @@ static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned int min_events)
>          * If we do, we can potentially be spinning for commands that
>          * already triggered a CQE (eg in error).
>          */
> -       if (io_cqring_events(ctx))
> +       if (io_cqring_events(ctx)) {
> +               io_submit_flush_completions(ctx);
>                 return 0;
> +       }

We should probably experiment with this a bit, I don't think it's a
showstopper for merging this.

Caleb, I think we can stage this for 7.1 and see how it goes.

-- 
Jens Axboe