From: Caleb Sander Mateos <csander@purestorage.com>
Date: Sat, 31 May 2025 17:25:16 +0000 (-0600)
Subject: engines/io_uring: remove unnecessary SQ full check
X-Git-Url: https://git.kernel.dk/?a=commitdiff_plain;h=d88e8c91a26150de2be19fcb776ec308e77f78a3;p=fio.git

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 08f220c8..51228037 100644
--- 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;