[10/12] net/bnxt: optimize vector mode mbuf allocation
diff mbox series

Message ID 20200909155700.29016-1-lance.richardson@broadcom.com
State Accepted
Delegated to: Ajit Khaparde
Headers show
Series
  • net/bnxt: vector PMD improvements
Related show

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

Lance Richardson Sept. 9, 2020, 3:57 p.m. UTC
Simplify and optimize receive mbuf allocation function used
by the vector mode PMDs.

Reviewed-by: Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
---
 drivers/net/bnxt/bnxt_rxtx_vec_common.h | 40 ++++++++++++++
 drivers/net/bnxt/bnxt_rxtx_vec_neon.c   | 70 -------------------------
 drivers/net/bnxt/bnxt_rxtx_vec_sse.c    | 70 -------------------------
 3 files changed, 40 insertions(+), 140 deletions(-)

Patch
diff mbox series

diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_common.h b/drivers/net/bnxt/bnxt_rxtx_vec_common.h
index fc2a12272b..819b8290e4 100644
--- a/drivers/net/bnxt/bnxt_rxtx_vec_common.h
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_common.h
@@ -56,4 +56,44 @@  bnxt_rxq_vec_setup_common(struct bnxt_rx_queue *rxq)
 	rxq->rxrearm_start = 0;
 	return 0;
 }
+
+static inline void
+bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
+{
+	struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
+	struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
+	int nb, i;
+
+	/*
+	 * Number of mbufs to allocate must be a multiple of four. The
+	 * allocation must not go past the end of the ring.
+	 */
+	nb = RTE_MIN(rxq->rxrearm_nb & ~0x3,
+		     rxq->nb_rx_desc - rxq->rxrearm_start);
+
+	/* Allocate new mbufs into the software ring. */
+	if (rte_mempool_get_bulk(rxq->mb_pool, (void *)rx_bufs, nb) < 0) {
+		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed += nb;
+
+		return;
+	}
+
+	/* Initialize the mbufs in vector, process 4 mbufs per loop. */
+	for (i = 0; i < nb; i += 4) {
+		rxbds[0].address = rte_mbuf_data_iova_default(rx_bufs[0]);
+		rxbds[1].address = rte_mbuf_data_iova_default(rx_bufs[1]);
+		rxbds[2].address = rte_mbuf_data_iova_default(rx_bufs[2]);
+		rxbds[3].address = rte_mbuf_data_iova_default(rx_bufs[3]);
+
+		rxbds += 4;
+		rx_bufs += 4;
+	}
+
+	rxq->rxrearm_start += nb;
+	bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= nb;
+}
 #endif /* _BNXT_RXTX_VEC_COMMON_H_ */
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_neon.c b/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
index 37b8c83656..24f9fc3c39 100644
--- a/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_neon.c
@@ -22,76 +22,6 @@ 
  * RX Ring handling
  */
 
-static inline void
-bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
-{
-	struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
-	struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
-	struct rte_mbuf *mb0, *mb1;
-	int nb, i;
-
-	const uint64x2_t hdr_room = {0, RTE_PKTMBUF_HEADROOM};
-	const uint64x2_t addrmask = {0, UINT64_MAX};
-
-	/*
-	 * Number of mbufs to allocate must be a multiple of two. The
-	 * allocation must not go past the end of the ring.
-	 */
-	nb = RTE_MIN(rxq->rxrearm_nb & ~0x1,
-		     rxq->nb_rx_desc - rxq->rxrearm_start);
-
-	/* Allocate new mbufs into the software ring */
-	if (rte_mempool_get_bulk(rxq->mb_pool, (void *)rx_bufs, nb) < 0) {
-		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed += nb;
-
-		return;
-	}
-
-	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
-	for (i = 0; i < nb; i += 2, rx_bufs += 2) {
-		uint64x2_t buf_addr0, buf_addr1;
-		uint64x2_t rxbd0, rxbd1;
-
-		mb0 = rx_bufs[0];
-		mb1 = rx_bufs[1];
-
-		/* Load address fields from both mbufs */
-		buf_addr0 = vld1q_u64((uint64_t *)&mb0->buf_addr);
-		buf_addr1 = vld1q_u64((uint64_t *)&mb1->buf_addr);
-
-		/* Load both rx descriptors (preserving some existing fields) */
-		rxbd0 = vld1q_u64((uint64_t *)(rxbds + 0));
-		rxbd1 = vld1q_u64((uint64_t *)(rxbds + 1));
-
-		/* Add default offset to buffer address. */
-		buf_addr0 = vaddq_u64(buf_addr0, hdr_room);
-		buf_addr1 = vaddq_u64(buf_addr1, hdr_room);
-
-		/* Clear all fields except address. */
-		buf_addr0 =  vandq_u64(buf_addr0, addrmask);
-		buf_addr1 =  vandq_u64(buf_addr1, addrmask);
-
-		/* Clear address field in descriptor. */
-		rxbd0 = vbicq_u64(rxbd0, addrmask);
-		rxbd1 = vbicq_u64(rxbd1, addrmask);
-
-		/* Set address field in descriptor. */
-		rxbd0 = vaddq_u64(rxbd0, buf_addr0);
-		rxbd1 = vaddq_u64(rxbd1, buf_addr1);
-
-		/* Store descriptors to memory. */
-		vst1q_u64((uint64_t *)(rxbds++), rxbd0);
-		vst1q_u64((uint64_t *)(rxbds++), rxbd1);
-	}
-
-	rxq->rxrearm_start += nb;
-	bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
-	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
-		rxq->rxrearm_start = 0;
-
-	rxq->rxrearm_nb -= nb;
-}
-
 static uint32_t
 bnxt_parse_pkt_type(uint32x4_t mm_rxcmp, uint32x4_t mm_rxcmp1)
 {
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
index 761d835963..7e87555408 100644
--- a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
@@ -26,76 +26,6 @@ 
  * RX Ring handling
  */
 
-static inline void
-bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
-{
-	struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
-	struct rte_mbuf **rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
-	struct rte_mbuf *mb0, *mb1;
-	int nb, i;
-
-	const __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM, 0);
-	const __m128i addrmask = _mm_set_epi64x(UINT64_MAX, 0);
-
-	/*
-	 * Number of mbufs to allocate must be a multiple of two. The
-	 * allocation must not go past the end of the ring.
-	 */
-	nb = RTE_MIN(rxq->rxrearm_nb & ~0x1,
-		     rxq->nb_rx_desc - rxq->rxrearm_start);
-
-	/* Allocate new mbufs into the software ring */
-	if (rte_mempool_get_bulk(rxq->mb_pool, (void *)rx_bufs, nb) < 0) {
-		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed += nb;
-
-		return;
-	}
-
-	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
-	for (i = 0; i < nb; i += 2, rx_bufs += 2) {
-		__m128i buf_addr0, buf_addr1;
-		__m128i rxbd0, rxbd1;
-
-		mb0 = rx_bufs[0];
-		mb1 = rx_bufs[1];
-
-		/* Load address fields from both mbufs */
-		buf_addr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
-		buf_addr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
-
-		/* Load both rx descriptors (preserving some existing fields) */
-		rxbd0 = _mm_loadu_si128((__m128i *)(rxbds + 0));
-		rxbd1 = _mm_loadu_si128((__m128i *)(rxbds + 1));
-
-		/* Add default offset to buffer address. */
-		buf_addr0 = _mm_add_epi64(buf_addr0, hdr_room);
-		buf_addr1 = _mm_add_epi64(buf_addr1, hdr_room);
-
-		/* Clear all fields except address. */
-		buf_addr0 =  _mm_and_si128(buf_addr0, addrmask);
-		buf_addr1 =  _mm_and_si128(buf_addr1, addrmask);
-
-		/* Clear address field in descriptor. */
-		rxbd0 = _mm_andnot_si128(addrmask, rxbd0);
-		rxbd1 = _mm_andnot_si128(addrmask, rxbd1);
-
-		/* Set address field in descriptor. */
-		rxbd0 = _mm_add_epi64(rxbd0, buf_addr0);
-		rxbd1 = _mm_add_epi64(rxbd1, buf_addr1);
-
-		/* Store descriptors to memory. */
-		_mm_store_si128((__m128i *)(rxbds++), rxbd0);
-		_mm_store_si128((__m128i *)(rxbds++), rxbd1);
-	}
-
-	rxq->rxrearm_start += nb;
-	bnxt_db_write(&rxr->rx_db, rxq->rxrearm_start - 1);
-	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
-		rxq->rxrearm_start = 0;
-
-	rxq->rxrearm_nb -= nb;
-}
-
 static __m128i
 bnxt_parse_pkt_type(__m128i mm_rxcmp, __m128i mm_rxcmp1)
 {