engines/io_uring: remove unnecessary SQ full check

fio_ioring_queue() bails out if the SQ's tail + 1 == head. This will
always be false, since 0 <= tail - head <= entries. Probably it was
meant to check whether the SQ is full, i.e. tail == head + entries.
(The head index should be loaded with acquire ordering in that case.)
Checking for a full SQ isn't necessary anyways, as the prior check for
ld->queued == td->o.iodepth already ensures the SQ isn't full.

So remove the unnecessary and misleading tail + 1 == head check.

Signed-off-by: Caleb Sander Mateos <csander@purestorage.com>

diff --git a/engines/io_uring.c b/engines/io_uring.c
index 08f220c89053bb99497043b3c21fee090b10650a..512280377434b368077a80dbd8aec878cb9ebc82 100644 (file)
--- a/engines/io_uring.c
+++ b/engines/io_uring.c
@@ -775,7 +775,7 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
        struct ioring_data *ld = td->io_ops_data;
        struct ioring_options *o = td->eo;
        struct io_sq_ring *ring = &ld->sq_ring;
-       unsigned tail, next_tail;
+       unsigned tail;
 
        fio_ro_check(td, io_u);
 
@@ -795,11 +795,6 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
                return FIO_Q_COMPLETED;
        }
 
-       tail = *ring->tail;
-       next_tail = tail + 1;
-       if (next_tail == atomic_load_relaxed(ring->head))
-               return FIO_Q_BUSY;
-
        if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
                fio_ioring_cmdprio_prep(td, io_u);
 
@@ -807,8 +802,9 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
                o->cmd_type == FIO_URING_CMD_NVME)
                fio_ioring_cmd_nvme_pi(td, io_u);
 
+       tail = *ring->tail;
        ring->array[tail & ld->sq_ring_mask] = io_u->index;
-       atomic_store_release(ring->tail, next_tail);
+       atomic_store_release(ring->tail, tail + 1);
 
        ld->queued++;
        return FIO_Q_QUEUED;