[dpdk-stable] [PATCH 3/4] net/mlx5: optimize Tx doorbell write

[dpdk-stable] [PATCH 3/4] net/mlx5: optimize Tx doorbell write

[Yongseok Koh]

Unnecessary volatile attribute keeps compiler from further optimizing the
code and this results in a little performance drop (~2%). Because of memory
barriers, it is safe to remove.

Fixes: 6bf10ab69be0 ("net/mlx5: support 32-bit systems")
Cc: stable@dpdk.org

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
 drivers/net/mlx5/mlx5_rxtx.h | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index 59fb43fefe..e210453fe0 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -379,17 +379,16 @@ uint32_t mlx5_tx_update_ext_mp(struct mlx5_txq_data *txq, uintptr_t addr,
  *   Address of the lock to use for that UAR access.
  */
 static __rte_always_inline void
-__mlx5_uar_write64_relaxed(uint64_t val, volatile void *addr,
+__mlx5_uar_write64_relaxed(uint64_t val, void *addr,
 			   rte_spinlock_t *lock __rte_unused)
 {
 #ifdef RTE_ARCH_64
-	rte_write64_relaxed(val, addr);
+	*(uint64_t *)addr = val;
 #else /* !RTE_ARCH_64 */
 	rte_spinlock_lock(lock);
-	rte_write32_relaxed(val, addr);
+	*(uint32_t *)addr = val;
 	rte_io_wmb();
-	rte_write32_relaxed(val >> 32,
-			    (volatile void *)((volatile char *)addr + 4));
+	*((uint32_t *)addr + 1) = val >> 32;
 	rte_spinlock_unlock(lock);
 #endif
 }
@@ -407,7 +406,7 @@ __mlx5_uar_write64_relaxed(uint64_t val, volatile void *addr,
  *   Address of the lock to use for that UAR access.
  */
 static __rte_always_inline void
-__mlx5_uar_write64(uint64_t val, volatile void *addr, rte_spinlock_t *lock)
+__mlx5_uar_write64(uint64_t val, void *addr, rte_spinlock_t *lock)
 {
 	rte_io_wmb();
 	__mlx5_uar_write64_relaxed(val, addr, lock);
-- 
2.11.0

[dpdk-stable] [PATCH 4/4] net/mlx5: optimize Rx buffer replenishment threshold

[Yongseok Koh]

Due to redundant calculation per every burst, performance drops a little.

Fixes: e10245a13b2e ("net/mlx5: fix Rx buffer replenishment threshold")
Cc: stable@dpdk.org

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
 drivers/net/mlx5/mlx5_rxq.c           | 2 ++
 drivers/net/mlx5/mlx5_rxtx.h          | 1 +
 drivers/net/mlx5/mlx5_rxtx_vec_neon.h | 2 +-
 drivers/net/mlx5/mlx5_rxtx_vec_sse.h  | 2 +-
 4 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index b27fc4798d..183da0e282 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -1468,6 +1468,8 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	tmpl->rxq.mp = mp;
 	tmpl->rxq.stats.idx = idx;
 	tmpl->rxq.elts_n = log2above(desc);
+	tmpl->rxq.rq_repl_thresh =
+		MLX5_VPMD_RXQ_RPLNSH_THRESH(1 << tmpl->rxq.elts_n);
 	tmpl->rxq.elts =
 		(struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
 #ifndef RTE_ARCH_64
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index e210453fe0..f47d327cfb 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -101,6 +101,7 @@ struct mlx5_rxq_data {
 	uint16_t consumed_strd; /* Number of consumed strides in WQE. */
 	uint32_t rq_pi;
 	uint32_t cq_ci;
+	uint16_t rq_repl_thresh; /* Threshold for buffer replenishment. */
 	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
 	uint16_t mprq_max_memcpy_len; /* Maximum size of packet to memcpy. */
 	volatile void *wqes;
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
index 0b729f1859..883fe1bf91 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_neon.h
@@ -732,7 +732,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
 	 */
 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
-	if (repl_n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n))
+	if (repl_n >= rxq->rq_repl_thresh)
 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
 	/* See if there're unreturned mbufs from compressed CQE. */
 	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
index e0f95f923d..14117c4bb4 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec_sse.h
@@ -716,7 +716,7 @@ rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
 	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
 	 */
 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
-	if (repl_n >= MLX5_VPMD_RXQ_RPLNSH_THRESH(q_n))
+	if (repl_n >= rxq->rq_repl_thresh)
 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
 	/* See if there're unreturned mbufs from compressed CQE. */
 	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
-- 
2.11.0