SQPOLL: io_sq_thread taking all the CPU

[jmillan]

uname -r: 6.1.0-13-amd64 liburing: 2.4

I'm successfully using liburing to send out UDP and TCP packets using io_uring_prep_sendto and io_uring_prep_write respectively.

I want to test the SQPOLL, so I initialize the ring with the given flag:

 unsigned int flags = IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_SQPOLL;

  // Initialize io_uring.
  auto err = io_uring_queue_init(DepLibUring::QueueDepth, std::addressof(this->ring), flags);

Once after calling io_uring_prep_sendto() or io_uring_prep_write(), if IORING_SQ_NEED_WAKEUP flag is set, I call io_uring_submit():

if (IO_URING_READ_ONCE(*this->ring.sq.kflags) & IORING_SQ_NEED_WAKEUP)
{
    auto err = io_uring_submit(std::addressof(this->ring));
}

This is working properly in the sense that packets are sent as expected, but an io_sq_thread is taking all the CPU.

When the thread is idle there is no CPU usage, but once it's awaken with io_uring_submit, it takes all the CPU again. Am I using it incorrectly?

My use case is a media server relaying network packets from one endpoint to one or many other endpoints.

[jmillan]

Maybe this happens for this scenario (and services alike) where the cadence of packets being sent does not allow the io_sq_thread sleep and keeps it awaken all the time. Maybe the scenario does not fit SQPOLL design. I just wonder if it is me misusing the API.

[axboe]

This is how polling works, it remains active until the desired idle time. It's impossible to poll for new SQ events without actually busy looping!

[jmillan]

It's impossible to poll for new SQ events without actually busy looping!

Yes, I understand it. I think SQ polling does definitely not suit my use case.

Thanks @axboe!