Small question

[MikuChan03]

I'm wondering, do we really need the io_uring_smp_load_acquire barrier? As I understand it, while the code in the following if statement might be executed speculatively, the side effects of that execution are only supposed to be visible to other threads when it is clear that the condition is met.

Removing the barrier, I have compiled liburing on a weakly ordered armv7 and run the tests repeatedly. No error occurred.

/*
 * Return an sqe to fill. Application must later call io_uring_submit()
 * when it's ready to tell the kernel about it. The caller may call this
 * function multiple times before calling io_uring_submit().
 *
 * Returns a vacant sqe, or NULL if we're full.
 */
IOURINGINLINE struct io_uring_sqe *_io_uring_get_sqe(struct io_uring *ring)
{
        struct io_uring_sq *sq = &ring->sq;
        unsigned int head, next = sq->sqe_tail + 1;
        int shift = 0;

        if (ring->flags & IORING_SETUP_SQE128)
                shift = 1;
        if (!(ring->flags & IORING_SETUP_SQPOLL))
                head = *sq->khead;
        else
                head = io_uring_smp_load_acquire(sq->khead);

        if (next - head <= sq->ring_entries) {
                struct io_uring_sqe *sqe;

                sqe = &sq->sqes[(sq->sqe_tail & sq->ring_mask) << shift];
                sq->sqe_tail = next;
                io_uring_initialize_sqe(sqe);
                return sqe;
        }

        return NULL;
}

[axboe]

Don't think it's needed in practice, io_uring_get_sqe() isn't thread safe to begin with. Care to send a patch?

[MikuChan03]

Right away, you'll have it shortly!

[MikuChan03]

@axboe It's here.