The reason for the requirement of a barrier between the txq writes
and the doorbell record writes is to avoid a case where the device
reads the doorbell record's new value before the txq writes are flushed
to memory.
The current use of rte_wmb is not necessary, and can be replaced by
rte_io_wmb which is more relaxed.
Replacing the rte_wmb is also expected to improve the throughput.
on v2:
* replace compiler barrier with rte_io_wmb.
Signed-off-by: Shahaf Shuler <shah...@mellanox.com>
Signed-off-by: Yongseok Koh <ys...@mellanox.com>
Signed-off-by: Alexander Solganik <solga...@gmail.com>
Signed-off-by: Sagi Grimberg <s...@grimberg.me>
Acked-by: Nelio Laranjeiro <nelio.laranje...@6wind.com>
---
drivers/net/mlx5/mlx5_rxtx.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index b3b161da5..e9895a9c0 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -591,7 +591,7 @@ mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe
*wqe)
uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
volatile uint64_t *src = ((volatile uint64_t *)wqe);
- rte_wmb();
+ rte_io_wmb();
*txq->qp_db = htonl(txq->wqe_ci);
/* Ensure ordering between DB record and BF copy. */
rte_wmb();
--
2.12.0
