[dpdk-dev] [PATCH 1/2] net/i40e: desc loading is unnecessarily ordered for aarch64

Gavin Hu gavin.hu at arm.com
Tue Aug 13 12:43:30 CEST 2019

For x86, the descriptors needs to be loaded in order, so in between two
descriptors loading, there is a compiler barrier in place.[1]
For aarch64, a patch [2] is in place to survive with discontinuous DD bits,
the barriers can be removed to take full advantage of out-of-order

50% performance gain in the RFC2544 NDR test was measured on ThunderX2.
12.50% performan gain in the RFC2544 NDR test was measured on Ampere
eMAG80 platform.

[1] http://inbox.dpdk.org/users/039ED4275CED7440929022BC67E7061153D71548@
[2] https://mails.dpdk.org/archives/stable/2017-October/003324.html

Fixes: ae0eb310f253 ("net/i40e: implement vector PMD for ARM")
Cc: stable at dpdk.org

Signed-off-by: Gavin Hu <gavin.hu at arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang at arm.com>
Reviewed-by: Steve Capper <steve.capper at arm.com>
 drivers/net/i40e/i40e_rxtx_vec_neon.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/net/i40e/i40e_rxtx_vec_neon.c b/drivers/net/i40e/i40e_rxtx_vec_neon.c
index 83572ef..5555e9b 100644
--- a/drivers/net/i40e/i40e_rxtx_vec_neon.c
+++ b/drivers/net/i40e/i40e_rxtx_vec_neon.c
@@ -285,7 +285,6 @@ _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
 		/* Read desc statuses backwards to avoid race condition */
 		/* A.1 load 4 pkts desc */
 		descs[3] =  vld1q_u64((uint64_t *)(rxdp + 3));
-		rte_rmb();
 		/* B.2 copy 2 mbuf point into rx_pkts  */
 		vst1q_u64((uint64_t *)&rx_pkts[pos], mbp1);

More information about the dev mailing list