net/bnxt: support NEON
authorLance Richardson <lance.richardson@broadcom.com>
Tue, 7 Jul 2020 22:22:25 +0000 (15:22 -0700)
committerFerruh Yigit <ferruh.yigit@intel.com>
Sat, 11 Jul 2020 04:18:52 +0000 (06:18 +0200)
Add bnxt vector PMD support using NEON SIMD instructions.
Also update the 20.08 release notes with this information.

Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
13 files changed:
doc/guides/nics/bnxt.rst
doc/guides/rel_notes/release_20_08.rst
drivers/net/bnxt/Makefile
drivers/net/bnxt/bnxt_ethdev.c
drivers/net/bnxt/bnxt_ring.c
drivers/net/bnxt/bnxt_rxq.h
drivers/net/bnxt/bnxt_rxr.c
drivers/net/bnxt/bnxt_rxr.h
drivers/net/bnxt/bnxt_rxtx_vec_common.h [new file with mode: 0644]
drivers/net/bnxt/bnxt_rxtx_vec_neon.c [new file with mode: 0644]
drivers/net/bnxt/bnxt_rxtx_vec_sse.c
drivers/net/bnxt/bnxt_txr.h
drivers/net/bnxt/meson.build

index a53cdad..6ff75d0 100644 (file)
@@ -774,9 +774,9 @@ Vector processing provides significantly improved performance over scalar
 processing (see Vector Processor, here).
 
 The BNXT PMD supports the vector processing using SSE (Streaming SIMD
-Extensions) instructions on x86 platforms. The BNXT vPMD (vector mode PMD) is
-currently limited to Intel/AMD CPU architecture. Support for ARM is *not*
-currently implemented.
+Extensions) instructions on x86 platforms. It also supports NEON intrinsics for
+vector processing on ARM CPUs. The BNXT vPMD (vector mode PMD) is available for
+Intel/AMD and ARM CPU architectures.
 
 This improved performance comes from several optimizations:
 
index dc32edb..7d8f462 100644 (file)
@@ -97,6 +97,7 @@ New Features
   * Added support for VXLAN encap/decap.
   * Added support for rte_flow_query for COUNT action.
   * Added support for rx_burst_mode_get and tx_burst_mode_get.
+  * Added vector mode support for ARM CPUs.
 
 * **Updated Mellanox mlx5 driver.**
 
index 6b9544b..c54fd10 100644 (file)
@@ -44,6 +44,9 @@ SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += rte_pmd_bnxt.c
 ifeq ($(CONFIG_RTE_ARCH_X86), y)
 SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += bnxt_rxtx_vec_sse.c
 endif
+ifeq ($(CONFIG_RTE_ARCH_ARM64), y)
+SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += bnxt_rxtx_vec_neon.c
+endif
 
 ifeq ($(CONFIG_RTE_LIBRTE_BNXT_PMD), y)
 CFLAGS += -I$(SRCDIR) -I$(SRCDIR)/tf_ulp -I$(SRCDIR)/tf_core -I$(SRCDIR)/hcapi
index 1dc85e6..3b9d3fb 100644 (file)
@@ -1076,7 +1076,7 @@ bnxt_receive_function(struct rte_eth_dev *eth_dev)
 {
        struct bnxt *bp = eth_dev->data->dev_private;
 
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 #ifndef RTE_LIBRTE_IEEE1588
        /*
         * Vector mode receive can be enabled only if scatter rx is not
@@ -1116,7 +1116,7 @@ bnxt_receive_function(struct rte_eth_dev *eth_dev)
 static eth_tx_burst_t
 bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)
 {
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 #ifndef RTE_LIBRTE_IEEE1588
        struct bnxt *bp = eth_dev->data->dev_private;
 
@@ -2679,7 +2679,7 @@ bnxt_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
                         "Scalar");
                return 0;
        }
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        if (pkt_burst == bnxt_recv_pkts_vec) {
                snprintf(mode->info, sizeof(mode->info), "%s",
                         "Vector SSE");
@@ -2701,7 +2701,7 @@ bnxt_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
                         "Scalar");
                return 0;
        }
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        if (pkt_burst == bnxt_xmit_pkts_vec) {
                snprintf(mode->info, sizeof(mode->info), "%s",
                         "Vector SSE");
@@ -2730,7 +2730,7 @@ int bnxt_mtu_set_op(struct rte_eth_dev *eth_dev, uint16_t new_mtu)
        new_pkt_size = new_mtu + RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN +
                       VLAN_TAG_SIZE * BNXT_NUM_VLANS;
 
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        /*
         * If vector-mode tx/rx is active, disallow any MTU change that would
         * require scattered receive support.
index 24a947f..54f6547 100644 (file)
@@ -608,7 +608,7 @@ int bnxt_alloc_hwrm_rx_ring(struct bnxt *bp, int queue_index)
                bnxt_db_write(&rxr->ag_db, rxr->ag_prod);
        }
        rxq->index = queue_index;
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        bnxt_rxq_vec_setup(rxq);
 #endif
 
@@ -713,7 +713,7 @@ int bnxt_alloc_hwrm_rings(struct bnxt *bp)
                bnxt_db_write(&rxr->rx_db, rxr->rx_prod);
                bnxt_db_write(&rxr->ag_db, rxr->ag_prod);
                rxq->index = i;
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
                bnxt_rxq_vec_setup(rxq);
 #endif
        }
index 4f5182d..d5ce3b6 100644 (file)
@@ -22,7 +22,7 @@ struct bnxt_rx_queue {
        uint16_t                nb_rx_hold; /* num held free RX desc */
        uint16_t                rx_free_thresh; /* max free RX desc to hold */
        uint16_t                queue_id; /* RX queue index */
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        uint16_t                rxrearm_nb; /* number of descs to reinit. */
        uint16_t                rxrearm_start; /* next desc index to reinit. */
 #endif
index eb6f33c..baf73cb 100644 (file)
@@ -782,7 +782,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                     !rte_spinlock_trylock(&rxq->lock)))
                return 0;
 
-#if defined(RTE_ARCH_X86)
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
        /*
         * Replenish buffers if needed when a transition has been made from
         * vector- to non-vector- receive processing.
index e60c97f..2bf46cd 100644 (file)
@@ -221,7 +221,7 @@ int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq);
 int bnxt_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id);
 int bnxt_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id);
 
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 uint16_t bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
                            uint16_t nb_pkts);
 int bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq);
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_common.h b/drivers/net/bnxt/bnxt_rxtx_vec_common.h
new file mode 100644 (file)
index 0000000..3da3c48
--- /dev/null
@@ -0,0 +1,57 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Broadcom
+ * All rights reserved.
+ */
+
+#ifndef _BNXT_RXTX_VEC_COMMON_H_
+#define _BNXT_RXTX_VEC_COMMON_H_
+
+#define RTE_BNXT_MAX_RX_BURST          32
+#define RTE_BNXT_MAX_TX_BURST          32
+#define RTE_BNXT_RXQ_REARM_THRESH      32
+#define RTE_BNXT_DESCS_PER_LOOP                4
+
+#define TX_BD_FLAGS_CMPL ((1 << TX_BD_LONG_FLAGS_BD_CNT_SFT) | \
+                         TX_BD_SHORT_FLAGS_COAL_NOW | \
+                         TX_BD_SHORT_TYPE_TX_BD_SHORT | \
+                         TX_BD_LONG_FLAGS_PACKET_END)
+
+#define TX_BD_FLAGS_NOCMPL (TX_BD_FLAGS_CMPL | TX_BD_LONG_FLAGS_NO_CMPL)
+
+static inline uint32_t
+bnxt_xmit_flags_len(uint16_t len, uint16_t flags)
+{
+       switch (len >> 9) {
+       case 0:
+               return flags | TX_BD_LONG_FLAGS_LHINT_LT512;
+       case 1:
+               return flags | TX_BD_LONG_FLAGS_LHINT_LT1K;
+       case 2:
+               return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
+       case 3:
+               return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
+       default:
+               return flags | TX_BD_LONG_FLAGS_LHINT_GTE2K;
+       }
+}
+
+static inline int
+bnxt_rxq_vec_setup_common(struct bnxt_rx_queue *rxq)
+{
+       uintptr_t p;
+       struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
+
+       mb_def.nb_segs = 1;
+       mb_def.data_off = RTE_PKTMBUF_HEADROOM;
+       mb_def.port = rxq->port_id;
+       rte_mbuf_refcnt_set(&mb_def, 1);
+
+       /* prevent compiler reordering: rearm_data covers previous fields */
+       rte_compiler_barrier();
+       p = (uintptr_t)&mb_def.rearm_data;
+       rxq->mbuf_initializer = *(uint64_t *)p;
+       rxq->rxrearm_nb = 0;
+       rxq->rxrearm_start = 0;
+       return 0;
+}
+#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
new file mode 100644 (file)
index 0000000..488a0b4
--- /dev/null
@@ -0,0 +1,469 @@
+/* SPDX-License-Identifier: BSD-3-Clause */
+/* Copyright(c) 2019-2020 Broadcom All rights reserved. */
+
+#include <inttypes.h>
+#include <stdbool.h>
+
+#include <rte_bitmap.h>
+#include <rte_byteorder.h>
+#include <rte_malloc.h>
+#include <rte_memory.h>
+#include <rte_vect.h>
+
+#include "bnxt.h"
+#include "bnxt_cpr.h"
+#include "bnxt_ring.h"
+#include "bnxt_rxr.h"
+#include "bnxt_rxq.h"
+#include "hsi_struct_def_dpdk.h"
+#include "bnxt_rxtx_vec_common.h"
+
+#include "bnxt_txq.h"
+#include "bnxt_txr.h"
+
+/*
+ * 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 bnxt_sw_rx_bd *rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
+       struct rte_mbuf *mb0, *mb1;
+       int i;
+
+       const uint64x2_t hdr_room = {0, RTE_PKTMBUF_HEADROOM};
+       const uint64x2_t addrmask = {0, UINT64_MAX};
+
+       /* Pull RTE_BNXT_RXQ_REARM_THRESH more mbufs into the software ring */
+       if (rte_mempool_get_bulk(rxq->mb_pool,
+                                (void *)rx_bufs,
+                                RTE_BNXT_RXQ_REARM_THRESH) < 0) {
+               rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+                       RTE_BNXT_RXQ_REARM_THRESH;
+
+               return;
+       }
+
+       /* Initialize the mbufs in vector, process 2 mbufs in one loop */
+       for (i = 0; i < RTE_BNXT_RXQ_REARM_THRESH; i += 2, rx_bufs += 2) {
+               uint64x2_t buf_addr0, buf_addr1;
+               uint64x2_t rxbd0, rxbd1;
+
+               mb0 = rx_bufs[0].mbuf;
+               mb1 = rx_bufs[1].mbuf;
+
+               /* 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 += RTE_BNXT_RXQ_REARM_THRESH;
+       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 -= RTE_BNXT_RXQ_REARM_THRESH;
+}
+
+static uint32_t
+bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
+{
+       uint32_t l3, pkt_type = 0;
+       uint32_t t_ipcs = 0, ip6 = 0, vlan = 0;
+       uint32_t flags_type;
+
+       vlan = !!(rxcmp1->flags2 &
+               rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN));
+       pkt_type |= vlan ? RTE_PTYPE_L2_ETHER_VLAN : RTE_PTYPE_L2_ETHER;
+
+       t_ipcs = !!(rxcmp1->flags2 &
+               rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC));
+       ip6 = !!(rxcmp1->flags2 &
+                rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_IP_TYPE));
+
+       flags_type = rxcmp->flags_type &
+               rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
+
+       if (!t_ipcs && !ip6)
+               l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
+       else if (!t_ipcs && ip6)
+               l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
+       else if (t_ipcs && !ip6)
+               l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
+       else
+               l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
+
+       switch (flags_type) {
+       case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_ICMP):
+               if (!t_ipcs)
+                       pkt_type |= l3 | RTE_PTYPE_L4_ICMP;
+               else
+                       pkt_type |= l3 | RTE_PTYPE_INNER_L4_ICMP;
+               break;
+
+       case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_TCP):
+               if (!t_ipcs)
+                       pkt_type |= l3 | RTE_PTYPE_L4_TCP;
+               else
+                       pkt_type |= l3 | RTE_PTYPE_INNER_L4_TCP;
+               break;
+
+       case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_UDP):
+               if (!t_ipcs)
+                       pkt_type |= l3 | RTE_PTYPE_L4_UDP;
+               else
+                       pkt_type |= l3 | RTE_PTYPE_INNER_L4_UDP;
+               break;
+
+       case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_IP):
+               pkt_type |= l3;
+               break;
+       }
+
+       return pkt_type;
+}
+
+static void
+bnxt_parse_csum(struct rte_mbuf *mbuf, struct rx_pkt_cmpl_hi *rxcmp1)
+{
+       uint32_t flags;
+
+       flags = flags2_0xf(rxcmp1);
+       /* IP Checksum */
+       if (likely(IS_IP_NONTUNNEL_PKT(flags))) {
+               if (unlikely(RX_CMP_IP_CS_ERROR(rxcmp1)))
+                       mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+               else
+                       mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
+       } else if (IS_IP_TUNNEL_PKT(flags)) {
+               if (unlikely(RX_CMP_IP_OUTER_CS_ERROR(rxcmp1) ||
+                            RX_CMP_IP_CS_ERROR(rxcmp1)))
+                       mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+               else
+                       mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
+       } else if (unlikely(RX_CMP_IP_CS_UNKNOWN(rxcmp1))) {
+               mbuf->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
+       }
+
+       /* L4 Checksum */
+       if (likely(IS_L4_NONTUNNEL_PKT(flags))) {
+               if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
+                       mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+               else
+                       mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+       } else if (IS_L4_TUNNEL_PKT(flags)) {
+               if (unlikely(RX_CMP_L4_INNER_CS_ERR2(rxcmp1)))
+                       mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+               else
+                       mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+               if (unlikely(RX_CMP_L4_OUTER_CS_ERR2(rxcmp1))) {
+                       mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_BAD;
+               } else if (unlikely(IS_L4_TUNNEL_PKT_ONLY_INNER_L4_CS
+                                   (flags))) {
+                       mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_UNKNOWN;
+               } else {
+                       mbuf->ol_flags |= PKT_RX_OUTER_L4_CKSUM_GOOD;
+               }
+       } else if (unlikely(RX_CMP_L4_CS_UNKNOWN(rxcmp1))) {
+               mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
+       }
+}
+
+uint16_t
+bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+                  uint16_t nb_pkts)
+{
+       struct bnxt_rx_queue *rxq = rx_queue;
+       struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+       struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+       uint32_t raw_cons = cpr->cp_raw_cons;
+       uint32_t cons;
+       int nb_rx_pkts = 0;
+       struct rx_pkt_cmpl *rxcmp;
+       bool evt = false;
+       const uint64x2_t mbuf_init = {rxq->mbuf_initializer, 0};
+       const uint8x16_t shuf_msk = {
+               0xFF, 0xFF, 0xFF, 0xFF,    /* pkt_type (zeroes) */
+               2, 3, 0xFF, 0xFF,          /* pkt_len */
+               2, 3,                      /* data_len */
+               0xFF, 0xFF,                /* vlan_tci (zeroes) */
+               12, 13, 14, 15             /* rss hash */
+       };
+
+       /* If Rx Q was stopped return */
+       if (unlikely(!rxq->rx_started))
+               return 0;
+
+       if (rxq->rxrearm_nb >= RTE_BNXT_RXQ_REARM_THRESH)
+               bnxt_rxq_rearm(rxq, rxr);
+
+       /* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
+       nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
+
+       /* Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP */
+       nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
+       if (!nb_pkts)
+               return 0;
+
+       /* Handle RX burst request */
+       while (1) {
+               cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
+
+               rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
+
+               if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
+                       break;
+
+               if (likely(CMP_TYPE(rxcmp) == RX_PKT_CMPL_TYPE_RX_L2)) {
+                       struct rx_pkt_cmpl_hi *rxcmp1;
+                       uint32_t tmp_raw_cons;
+                       uint16_t cp_cons;
+                       struct rte_mbuf *mbuf;
+                       uint64x2_t mm_rxcmp;
+                       uint8x16_t pkt_mb;
+
+                       tmp_raw_cons = NEXT_RAW_CMP(raw_cons);
+                       cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
+                       rxcmp1 = (struct rx_pkt_cmpl_hi *)
+                                               &cpr->cp_desc_ring[cp_cons];
+
+                       if (!CMP_VALID(rxcmp1, tmp_raw_cons,
+                                      cpr->cp_ring_struct))
+                               break;
+
+                       raw_cons = tmp_raw_cons;
+                       cons = rxcmp->opaque;
+
+                       mbuf = rxr->rx_buf_ring[cons].mbuf;
+                       rte_prefetch0(mbuf);
+                       rxr->rx_buf_ring[cons].mbuf = NULL;
+
+                       /* Set constant fields from mbuf initializer. */
+                       vst1q_u64((uint64_t *)&mbuf->rearm_data, mbuf_init);
+
+                       /* Set mbuf pkt_len, data_len, and rss_hash fields. */
+                       mm_rxcmp = vld1q_u64((uint64_t *)rxcmp);
+                       pkt_mb = vqtbl1q_u8(vreinterpretq_u8_u64(mm_rxcmp),
+                                           shuf_msk);
+                       vst1q_u64((uint64_t *)&mbuf->rx_descriptor_fields1,
+                                 vreinterpretq_u64_u8(pkt_mb));
+
+                       rte_compiler_barrier();
+
+                       if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
+                               mbuf->ol_flags |= PKT_RX_RSS_HASH;
+
+                       if (rxcmp1->flags2 &
+                           RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
+                               mbuf->vlan_tci = rxcmp1->metadata &
+                                       (RX_PKT_CMPL_METADATA_VID_MASK |
+                                       RX_PKT_CMPL_METADATA_DE |
+                                       RX_PKT_CMPL_METADATA_PRI_MASK);
+                               mbuf->ol_flags |=
+                                       PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
+                       }
+
+                       bnxt_parse_csum(mbuf, rxcmp1);
+                       mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
+
+                       rx_pkts[nb_rx_pkts++] = mbuf;
+               } else if (!BNXT_NUM_ASYNC_CPR(rxq->bp)) {
+                       evt =
+                       bnxt_event_hwrm_resp_handler(rxq->bp,
+                                                    (struct cmpl_base *)rxcmp);
+               }
+
+               raw_cons = NEXT_RAW_CMP(raw_cons);
+               if (nb_rx_pkts == nb_pkts || evt)
+                       break;
+       }
+       rxr->rx_prod = RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
+
+       rxq->rxrearm_nb += nb_rx_pkts;
+       cpr->cp_raw_cons = raw_cons;
+       cpr->valid = !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
+       if (nb_rx_pkts || evt)
+               bnxt_db_cq(cpr);
+
+       return nb_rx_pkts;
+}
+
+static void
+bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
+{
+       struct bnxt_tx_ring_info *txr = txq->tx_ring;
+       struct rte_mbuf **free = txq->free;
+       uint16_t cons = txr->tx_cons;
+       unsigned int blk = 0;
+
+       while (nr_pkts--) {
+               struct bnxt_sw_tx_bd *tx_buf;
+               struct rte_mbuf *mbuf;
+
+               tx_buf = &txr->tx_buf_ring[cons];
+               cons = RING_NEXT(txr->tx_ring_struct, cons);
+               mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
+               tx_buf->mbuf = NULL;
+
+               if (blk && mbuf->pool != free[0]->pool) {
+                       rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
+                       blk = 0;
+               }
+               free[blk++] = mbuf;
+       }
+       if (blk)
+               rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
+
+       txr->tx_cons = cons;
+}
+
+static void
+bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
+{
+       struct bnxt_cp_ring_info *cpr = txq->cp_ring;
+       uint32_t raw_cons = cpr->cp_raw_cons;
+       uint32_t cons;
+       uint32_t nb_tx_pkts = 0;
+       struct tx_cmpl *txcmp;
+       struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
+       struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
+       uint32_t ring_mask = cp_ring_struct->ring_mask;
+
+       do {
+               cons = RING_CMPL(ring_mask, raw_cons);
+               txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
+
+               if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
+                       break;
+
+               if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
+                       nb_tx_pkts += txcmp->opaque;
+               else
+                       RTE_LOG_DP(ERR, PMD,
+                                  "Unhandled CMP type %02x\n",
+                                  CMP_TYPE(txcmp));
+               raw_cons = NEXT_RAW_CMP(raw_cons);
+       } while (nb_tx_pkts < ring_mask);
+
+       cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
+       if (nb_tx_pkts) {
+               bnxt_tx_cmp_vec(txq, nb_tx_pkts);
+               cpr->cp_raw_cons = raw_cons;
+               bnxt_db_cq(cpr);
+       }
+}
+
+static uint16_t
+bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+                         uint16_t nb_pkts)
+{
+       struct bnxt_tx_queue *txq = tx_queue;
+       struct bnxt_tx_ring_info *txr = txq->tx_ring;
+       uint16_t prod = txr->tx_prod;
+       struct rte_mbuf *tx_mbuf;
+       struct tx_bd_long *txbd = NULL;
+       struct bnxt_sw_tx_bd *tx_buf;
+       uint16_t to_send;
+
+       nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
+
+       if (unlikely(nb_pkts == 0))
+               return 0;
+
+       /* Handle TX burst request */
+       to_send = nb_pkts;
+       while (to_send) {
+               tx_mbuf = *tx_pkts++;
+               rte_prefetch0(tx_mbuf);
+
+               tx_buf = &txr->tx_buf_ring[prod];
+               tx_buf->mbuf = tx_mbuf;
+               tx_buf->nr_bds = 1;
+
+               txbd = &txr->tx_desc_ring[prod];
+               txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
+               txbd->len = tx_mbuf->data_len;
+               txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
+                                                      TX_BD_FLAGS_NOCMPL);
+               prod = RING_NEXT(txr->tx_ring_struct, prod);
+               to_send--;
+       }
+
+       /* Request a completion for last packet in burst */
+       if (txbd) {
+               txbd->opaque = nb_pkts;
+               txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
+       }
+
+       rte_compiler_barrier();
+       bnxt_db_write(&txr->tx_db, prod);
+
+       txr->tx_prod = prod;
+
+       return nb_pkts;
+}
+
+uint16_t
+bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+                  uint16_t nb_pkts)
+{
+       int nb_sent = 0;
+       struct bnxt_tx_queue *txq = tx_queue;
+
+       /* Tx queue was stopped; wait for it to be restarted */
+       if (unlikely(!txq->tx_started)) {
+               PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
+               return 0;
+       }
+
+       /* Handle TX completions */
+       if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
+               bnxt_handle_tx_cp_vec(txq);
+
+       while (nb_pkts) {
+               uint16_t ret, num;
+
+               num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
+               ret = bnxt_xmit_fixed_burst_vec(tx_queue,
+                                               &tx_pkts[nb_sent],
+                                               num);
+               nb_sent += ret;
+               nb_pkts -= ret;
+               if (ret < num)
+                       break;
+       }
+
+       return nb_sent;
+}
+
+int __rte_cold
+bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
+{
+       return bnxt_rxq_vec_setup_common(rxq);
+}
index 8f73add..c4ca5cf 100644 (file)
@@ -20,6 +20,7 @@
 #include "bnxt_rxr.h"
 #include "bnxt_rxq.h"
 #include "hsi_struct_def_dpdk.h"
+#include "bnxt_rxtx_vec_common.h"
 
 #include "bnxt_txq.h"
 #include "bnxt_txr.h"
  * RX Ring handling
  */
 
-#define RTE_BNXT_MAX_RX_BURST          32
-#define RTE_BNXT_MAX_TX_BURST          32
-#define RTE_BNXT_RXQ_REARM_THRESH      32
-#define RTE_BNXT_DESCS_PER_LOOP                4
-
 static inline void
 bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
 {
@@ -388,30 +384,6 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
        }
 }
 
-#define TX_BD_FLAGS_CMPL ((1 << TX_BD_LONG_FLAGS_BD_CNT_SFT) | \
-                         TX_BD_SHORT_FLAGS_COAL_NOW | \
-                         TX_BD_SHORT_TYPE_TX_BD_SHORT | \
-                         TX_BD_LONG_FLAGS_PACKET_END)
-
-#define TX_BD_FLAGS_NOCMPL (TX_BD_FLAGS_CMPL | TX_BD_LONG_FLAGS_NO_CMPL)
-
-static inline uint32_t
-bnxt_xmit_flags_len(uint16_t len, uint16_t flags)
-{
-       switch (len >> 9) {
-       case 0:
-               return flags | TX_BD_LONG_FLAGS_LHINT_LT512;
-       case 1:
-               return flags | TX_BD_LONG_FLAGS_LHINT_LT1K;
-       case 2:
-               return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
-       case 3:
-               return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
-       default:
-               return flags | TX_BD_LONG_FLAGS_LHINT_GTE2K;
-       }
-}
-
 static uint16_t
 bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
                          uint16_t nb_pkts)
@@ -498,19 +470,5 @@ bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 int __rte_cold
 bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
 {
-       uintptr_t p;
-       struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
-
-       mb_def.nb_segs = 1;
-       mb_def.data_off = RTE_PKTMBUF_HEADROOM;
-       mb_def.port = rxq->port_id;
-       rte_mbuf_refcnt_set(&mb_def, 1);
-
-       /* prevent compiler reordering: rearm_data covers previous fields */
-       rte_compiler_barrier();
-       p = (uintptr_t)&mb_def.rearm_data;
-       rxq->mbuf_initializer = *(uint64_t *)p;
-       rxq->rxrearm_nb = 0;
-       rxq->rxrearm_start = 0;
-       return 0;
+       return bnxt_rxq_vec_setup_common(rxq);
 }
index e7f43f9..7715c11 100644 (file)
@@ -59,7 +59,7 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
                               uint16_t nb_pkts);
 uint16_t bnxt_dummy_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
                              uint16_t nb_pkts);
-#ifdef RTE_ARCH_X86
+#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 uint16_t bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
                            uint16_t nb_pkts);
 #endif
index 5fb0ed3..0b93c31 100644 (file)
@@ -68,4 +68,6 @@ sources = files('bnxt_cpr.c',
 
 if arch_subdir == 'x86'
        sources += files('bnxt_rxtx_vec_sse.c')
+elif arch_subdir == 'arm' and host_machine.cpu_family().startswith('aarch64')
+       sources += files('bnxt_rxtx_vec_neon.c')
 endif