net/mlx5: add vectorized Rx/Tx burst for x86
authorYongseok Koh <yskoh@mellanox.com>
Thu, 6 Jul 2017 18:41:10 +0000 (11:41 -0700)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 7 Jul 2017 09:56:03 +0000 (11:56 +0200)
To make vectorized burst routines enabled, it is required to run on x86_64
architecture. If all the conditions are met, the vectorized burst functions
are enabled automatically. The decision is made individually on RX and TX.
There's no PMD option to make a selection.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
drivers/net/mlx5/Makefile
drivers/net/mlx5/mlx5_defs.h
drivers/net/mlx5/mlx5_ethdev.c
drivers/net/mlx5/mlx5_rxq.c
drivers/net/mlx5/mlx5_rxtx.c
drivers/net/mlx5/mlx5_rxtx.h
drivers/net/mlx5/mlx5_rxtx_vec_sse.c [new file with mode: 0644]
drivers/net/mlx5/mlx5_txq.c

index 8b8f6ea..8736de5 100644 (file)
@@ -39,6 +39,9 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_txq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx.c
+ifeq ($(CONFIG_RTE_ARCH_X86_64),y)
+SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx_vec_sse.c
+endif
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_trigger.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_ethdev.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_mac.c
index 201bb33..a76bc6f 100644 (file)
 /* Maximum Packet headers size (L2+L3+L4) for TSO. */
 #define MLX5_MAX_TSO_HEADER 128
 
+/* Default minimum number of Tx queues for vectorized Tx. */
+#define MLX5_VPMD_MIN_TXQS 4
+
+/* Threshold of buffer replenishment for vectorized Rx. */
+#define MLX5_VPMD_RXQ_RPLNSH_THRESH   64U
+
+/* Maximum size of burst for vectorized Rx. */
+#define MLX5_VPMD_RX_MAX_BURST        MLX5_VPMD_RXQ_RPLNSH_THRESH
+
+/*
+ * Maximum size of burst for vectorized Tx. This is related to the maximum size
+ * of Enhaned MPW (eMPW) WQE as vectorized Tx is supported with eMPW.
+ */
+#define MLX5_VPMD_TX_MAX_BURST        32U
+
+/* Number of packets vectorized Rx can simultaneously process in a loop. */
+#define MLX5_VPMD_DESCS_PER_LOOP      4
+
 #endif /* RTE_PMD_MLX5_DEFS_H_ */
index 54c4c6c..b70b7b9 100644 (file)
@@ -723,7 +723,8 @@ mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
 
        };
 
-       if (dev->rx_pkt_burst == mlx5_rx_burst)
+       if (dev->rx_pkt_burst == mlx5_rx_burst ||
+           dev->rx_pkt_burst == mlx5_rx_burst_vec)
                return ptypes;
        return NULL;
 }
@@ -1585,9 +1586,16 @@ priv_select_tx_function(struct priv *priv)
        priv->dev->tx_pkt_burst = mlx5_tx_burst;
        /* Select appropriate TX function. */
        if (priv->mps == MLX5_MPW_ENHANCED) {
-               priv->dev->tx_pkt_burst =
-                       mlx5_tx_burst_empw;
-               DEBUG("selected Enhanced MPW TX function");
+               if (priv_check_vec_tx_support(priv) > 0) {
+                       if (priv_check_raw_vec_tx_support(priv) > 0)
+                               priv->dev->tx_pkt_burst = mlx5_tx_burst_raw_vec;
+                       else
+                               priv->dev->tx_pkt_burst = mlx5_tx_burst_vec;
+                       DEBUG("selected Enhanced MPW TX vectorized function");
+               } else {
+                       priv->dev->tx_pkt_burst = mlx5_tx_burst_empw;
+                       DEBUG("selected Enhanced MPW TX function");
+               }
        } else if (priv->mps && priv->txq_inline) {
                priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
                DEBUG("selected MPW inline TX function");
@@ -1606,5 +1614,11 @@ priv_select_tx_function(struct priv *priv)
 void
 priv_select_rx_function(struct priv *priv)
 {
-       priv->dev->rx_pkt_burst = mlx5_rx_burst;
+       if (priv_check_vec_rx_support(priv) > 0) {
+               priv_prep_vec_rx_function(priv);
+               priv->dev->rx_pkt_burst = mlx5_rx_burst_vec;
+               DEBUG("selected RX vectorized function");
+       } else {
+               priv->dev->rx_pkt_burst = mlx5_rx_burst;
+       }
 }
index 98b20eb..34ec95b 100644 (file)
@@ -633,6 +633,41 @@ priv_rehash_flows(struct priv *priv)
        return 0;
 }
 
+/**
+ * Unlike regular Rx function, vPMD Rx doesn't replace mbufs immediately when
+ * receiving packets. Instead it replaces later in bulk. In rxq->elts[], entries
+ * from rq_pi to rq_ci are owned by device but the rest is already delivered to
+ * application. In order not to reuse those mbufs by rxq_alloc_elts(), this
+ * function must be called to replace used mbufs.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+rxq_trim_elts(struct rxq *rxq)
+{
+       const uint16_t q_n = (1 << rxq->elts_n);
+       const uint16_t q_mask = q_n - 1;
+       uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
+       uint16_t i;
+
+       if (!rxq->trim_elts)
+               return 0;
+       for (i = 0; i < used; ++i) {
+               struct rte_mbuf *buf;
+               buf = rte_pktmbuf_alloc(rxq->mp);
+               if (!buf)
+                       return ENOMEM;
+               (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = buf;
+       }
+       rxq->rq_pi = rxq->rq_ci;
+       rxq->trim_elts = 0;
+       return 0;
+}
+
 /**
  * Allocate RX queue elements.
  *
@@ -800,6 +835,7 @@ rxq_rehash(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl)
                return err;
        }
        /* Snatch mbufs from original queue. */
+       claim_zero(rxq_trim_elts(&rxq_ctrl->rxq));
        claim_zero(rxq_alloc_elts(rxq_ctrl, elts_n, rxq_ctrl->rxq.elts));
        for (i = 0; i != elts_n; ++i) {
                struct rte_mbuf *buf = (*rxq_ctrl->rxq.elts)[i];
@@ -860,6 +896,7 @@ rxq_setup(struct rxq_ctrl *tmpl)
        tmpl->rxq.cqe_n = log2above(cq_info.cqe_cnt);
        tmpl->rxq.cq_ci = 0;
        tmpl->rxq.rq_ci = 0;
+       tmpl->rxq.rq_pi = 0;
        tmpl->rxq.cq_db = cq_info.dbrec;
        tmpl->rxq.wqes =
                (volatile struct mlx5_wqe_data_seg (*)[])
@@ -993,7 +1030,12 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        if (priv->cqe_comp) {
                attr.cq.comp_mask |= IBV_EXP_CQ_INIT_ATTR_FLAGS;
                attr.cq.flags |= IBV_EXP_CQ_COMPRESSED_CQE;
-               cqe_n = (desc * 2) - 1; /* Double the number of CQEs. */
+               /*
+                * For vectorized Rx, it must not be doubled in order to
+                * make cq_ci and rq_ci aligned.
+                */
+               if (rxq_check_vec_support(&tmpl.rxq) < 0)
+                       cqe_n = (desc * 2) - 1; /* Double the number of CQEs. */
        }
        tmpl.cq = ibv_exp_create_cq(priv->ctx, cqe_n, NULL, tmpl.channel, 0,
                                    &attr.cq);
@@ -1103,7 +1145,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        if (rxq_ctrl->rxq.elts_n) {
                assert(1 << rxq_ctrl->rxq.elts_n == desc);
                assert(rxq_ctrl->rxq.elts != tmpl.rxq.elts);
-               ret = rxq_alloc_elts(&tmpl, desc, rxq_ctrl->rxq.elts);
+               ret = rxq_trim_elts(&rxq_ctrl->rxq);
+               if (!ret)
+                       ret = rxq_alloc_elts(&tmpl, desc, rxq_ctrl->rxq.elts);
        } else
                ret = rxq_alloc_elts(&tmpl, desc, NULL);
        if (ret) {
@@ -1165,6 +1209,7 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
        struct priv *priv = dev->data->dev_private;
        struct rxq *rxq = (*priv->rxqs)[idx];
        struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+       const uint16_t desc_pad = MLX5_VPMD_DESCS_PER_LOOP; /* For vPMD. */
        int ret;
 
        if (mlx5_is_secondary())
@@ -1198,7 +1243,8 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                if (rxq_ctrl->rxq.elts_n != log2above(desc)) {
                        rxq_ctrl = rte_realloc(rxq_ctrl,
                                               sizeof(*rxq_ctrl) +
-                                              desc * sizeof(struct rte_mbuf *),
+                                              (desc + desc_pad) *
+                                               sizeof(struct rte_mbuf *),
                                               RTE_CACHE_LINE_SIZE);
                        if (!rxq_ctrl) {
                                ERROR("%p: unable to reallocate queue index %u",
@@ -1209,7 +1255,8 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                }
        } else {
                rxq_ctrl = rte_calloc_socket("RXQ", 1, sizeof(*rxq_ctrl) +
-                                            desc * sizeof(struct rte_mbuf *),
+                                            (desc + desc_pad) *
+                                             sizeof(struct rte_mbuf *),
                                             0, socket);
                if (rxq_ctrl == NULL) {
                        ERROR("%p: unable to allocate queue index %u",
index 688ee90..ab6df19 100644 (file)
 #include "mlx5_defs.h"
 #include "mlx5_prm.h"
 
-static __rte_always_inline int
-check_cqe(volatile struct mlx5_cqe *cqe,
-         unsigned int cqes_n, const uint16_t ci);
-
-static __rte_always_inline void
-txq_complete(struct txq *txq);
-
-static __rte_always_inline uint32_t
-txq_mb2mr(struct txq *txq, struct rte_mbuf *mb);
-
-static __rte_always_inline void
-mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe);
-
 static __rte_always_inline uint32_t
 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe);
 
@@ -92,101 +79,29 @@ mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
 static __rte_always_inline uint32_t
 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe);
 
-#ifndef NDEBUG
-
-/**
- * Verify or set magic value in CQE.
- *
- * @param cqe
- *   Pointer to CQE.
- *
- * @return
- *   0 the first time.
- */
-static inline int
-check_cqe_seen(volatile struct mlx5_cqe *cqe)
-{
-       static const uint8_t magic[] = "seen";
-       volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
-       int ret = 1;
-       unsigned int i;
-
-       for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
-               if (!ret || (*buf)[i] != magic[i]) {
-                       ret = 0;
-                       (*buf)[i] = magic[i];
-               }
-       return ret;
-}
-
-#endif /* NDEBUG */
-
-/**
- * Check whether CQE is valid.
- *
- * @param cqe
- *   Pointer to CQE.
- * @param cqes_n
- *   Size of completion queue.
- * @param ci
- *   Consumer index.
- *
- * @return
- *   0 on success, 1 on failure.
- */
-static inline int
-check_cqe(volatile struct mlx5_cqe *cqe,
-         unsigned int cqes_n, const uint16_t ci)
-{
-       uint16_t idx = ci & cqes_n;
-       uint8_t op_own = cqe->op_own;
-       uint8_t op_owner = MLX5_CQE_OWNER(op_own);
-       uint8_t op_code = MLX5_CQE_OPCODE(op_own);
-
-       if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
-               return 1; /* No CQE. */
-#ifndef NDEBUG
-       if ((op_code == MLX5_CQE_RESP_ERR) ||
-           (op_code == MLX5_CQE_REQ_ERR)) {
-               volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
-               uint8_t syndrome = err_cqe->syndrome;
-
-               if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
-                   (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
-                       return 0;
-               if (!check_cqe_seen(cqe))
-                       ERROR("unexpected CQE error %u (0x%02x)"
-                             " syndrome 0x%02x",
-                             op_code, op_code, syndrome);
-               return 1;
-       } else if ((op_code != MLX5_CQE_RESP_SEND) &&
-                  (op_code != MLX5_CQE_REQ)) {
-               if (!check_cqe_seen(cqe))
-                       ERROR("unexpected CQE opcode %u (0x%02x)",
-                             op_code, op_code);
-               return 1;
-       }
-#endif /* NDEBUG */
-       return 0;
-}
-
-/**
- * Return the address of the WQE.
- *
- * @param txq
- *   Pointer to TX queue structure.
- * @param  wqe_ci
- *   WQE consumer index.
- *
- * @return
- *   WQE address.
+/*
+ * The index to the array should have:
+ * bit[1:0] = l3_hdr_type, bit[2] = tunneled, bit[3] = outer_l3_type
  */
-static inline uintptr_t *
-tx_mlx5_wqe(struct txq *txq, uint16_t ci)
-{
-       ci &= ((1 << txq->wqe_n) - 1);
-       return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
-}
+const uint32_t mlx5_ptype_table[] = {
+       RTE_PTYPE_UNKNOWN,
+       RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,               /* b0001 */
+       RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,               /* b0010 */
+       RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+       RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
+               RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b0101 */
+       RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
+               RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b0110 */
+       RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+       RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,               /* b1001 */
+       RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,               /* b1010 */
+       RTE_PTYPE_UNKNOWN, RTE_PTYPE_UNKNOWN,
+       RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
+               RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN, /* b1101 */
+       RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
+               RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN, /* b1110 */
+       RTE_PTYPE_ALL_MASK                           /* b1111 */
+};
 
 /**
  * Return the size of tailroom of WQ.
@@ -244,174 +159,6 @@ mlx5_copy_to_wq(void *dst, const void *src, size_t n,
        return ret;
 }
 
-/**
- * Manage TX completions.
- *
- * When sending a burst, mlx5_tx_burst() posts several WRs.
- *
- * @param txq
- *   Pointer to TX queue structure.
- */
-static inline void
-txq_complete(struct txq *txq)
-{
-       const uint16_t elts_n = 1 << txq->elts_n;
-       const uint16_t elts_m = elts_n - 1;
-       const unsigned int cqe_n = 1 << txq->cqe_n;
-       const unsigned int cqe_cnt = cqe_n - 1;
-       uint16_t elts_free = txq->elts_tail;
-       uint16_t elts_tail;
-       uint16_t cq_ci = txq->cq_ci;
-       volatile struct mlx5_cqe *cqe = NULL;
-       volatile struct mlx5_wqe_ctrl *ctrl;
-       struct rte_mbuf *m, *free[elts_n];
-       struct rte_mempool *pool = NULL;
-       unsigned int blk_n = 0;
-
-       do {
-               volatile struct mlx5_cqe *tmp;
-
-               tmp = &(*txq->cqes)[cq_ci & cqe_cnt];
-               if (check_cqe(tmp, cqe_n, cq_ci))
-                       break;
-               cqe = tmp;
-#ifndef NDEBUG
-               if (MLX5_CQE_FORMAT(cqe->op_own) == MLX5_COMPRESSED) {
-                       if (!check_cqe_seen(cqe))
-                               ERROR("unexpected compressed CQE, TX stopped");
-                       return;
-               }
-               if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
-                   (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
-                       if (!check_cqe_seen(cqe))
-                               ERROR("unexpected error CQE, TX stopped");
-                       return;
-               }
-#endif /* NDEBUG */
-               ++cq_ci;
-       } while (1);
-       if (unlikely(cqe == NULL))
-               return;
-       txq->wqe_pi = ntohs(cqe->wqe_counter);
-       ctrl = (volatile struct mlx5_wqe_ctrl *)
-               tx_mlx5_wqe(txq, txq->wqe_pi);
-       elts_tail = ctrl->ctrl3;
-       assert((elts_tail & elts_m) < (1 << txq->wqe_n));
-       /* Free buffers. */
-       while (elts_free != elts_tail) {
-               m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
-               if (likely(m != NULL)) {
-                       if (likely(m->pool == pool)) {
-                               free[blk_n++] = m;
-                       } else {
-                               if (likely(pool != NULL))
-                                       rte_mempool_put_bulk(pool,
-                                                            (void *)free,
-                                                            blk_n);
-                               free[0] = m;
-                               pool = m->pool;
-                               blk_n = 1;
-                       }
-               }
-       }
-       if (blk_n)
-               rte_mempool_put_bulk(pool, (void *)free, blk_n);
-#ifndef NDEBUG
-       elts_free = txq->elts_tail;
-       /* Poisoning. */
-       while (elts_free != elts_tail) {
-               memset(&(*txq->elts)[elts_free & elts_m],
-                      0x66,
-                      sizeof((*txq->elts)[elts_free & elts_m]));
-               ++elts_free;
-       }
-#endif
-       txq->cq_ci = cq_ci;
-       txq->elts_tail = elts_tail;
-       /* Update the consumer index. */
-       rte_wmb();
-       *txq->cq_db = htonl(cq_ci);
-}
-
-/**
- * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
- * the cloned mbuf is allocated is returned instead.
- *
- * @param buf
- *   Pointer to mbuf.
- *
- * @return
- *   Memory pool where data is located for given mbuf.
- */
-static struct rte_mempool *
-txq_mb2mp(struct rte_mbuf *buf)
-{
-       if (unlikely(RTE_MBUF_INDIRECT(buf)))
-               return rte_mbuf_from_indirect(buf)->pool;
-       return buf->pool;
-}
-
-/**
- * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
- * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
- * remove an entry first.
- *
- * @param txq
- *   Pointer to TX queue structure.
- * @param[in] mp
- *   Memory Pool for which a Memory Region lkey must be returned.
- *
- * @return
- *   mr->lkey on success, (uint32_t)-1 on failure.
- */
-static inline uint32_t
-txq_mb2mr(struct txq *txq, struct rte_mbuf *mb)
-{
-       uint16_t i = txq->mr_cache_idx;
-       uintptr_t addr = rte_pktmbuf_mtod(mb, uintptr_t);
-
-       assert(i < RTE_DIM(txq->mp2mr));
-       if (likely(txq->mp2mr[i].start <= addr && txq->mp2mr[i].end >= addr))
-               return txq->mp2mr[i].lkey;
-       for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
-               if (unlikely(txq->mp2mr[i].mr == NULL)) {
-                       /* Unknown MP, add a new MR for it. */
-                       break;
-               }
-               if (txq->mp2mr[i].start <= addr &&
-                   txq->mp2mr[i].end >= addr) {
-                       assert(txq->mp2mr[i].lkey != (uint32_t)-1);
-                       assert(htonl(txq->mp2mr[i].mr->lkey) ==
-                              txq->mp2mr[i].lkey);
-                       txq->mr_cache_idx = i;
-                       return txq->mp2mr[i].lkey;
-               }
-       }
-       txq->mr_cache_idx = 0;
-       return txq_mp2mr_reg(txq, txq_mb2mp(mb), i);
-}
-
-/**
- * Ring TX queue doorbell.
- *
- * @param txq
- *   Pointer to TX queue structure.
- * @param wqe
- *   Pointer to the last WQE posted in the NIC.
- */
-static inline void
-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();
-       *txq->qp_db = htonl(txq->wqe_ci);
-       /* Ensure ordering between DB record and BF copy. */
-       rte_wmb();
-       *dst = *src;
-}
-
 /**
  * DPDK callback to check the status of a tx descriptor.
  *
@@ -429,7 +176,7 @@ mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
        struct txq *txq = tx_queue;
        uint16_t used;
 
-       txq_complete(txq);
+       mlx5_tx_complete(txq);
        used = txq->elts_head - txq->elts_tail;
        if (offset < used)
                return RTE_ETH_TX_DESC_FULL;
@@ -525,7 +272,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        /* Prefetch first packet cacheline. */
        rte_prefetch0(*pkts);
        /* Start processing. */
-       txq_complete(txq);
+       mlx5_tx_complete(txq);
        max_elts = (elts_n - (elts_head - txq->elts_tail));
        max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
        if (unlikely(!max_wqe))
@@ -773,7 +520,7 @@ use_dseg:
                        naddr = htonll(addr);
                        *dseg = (rte_v128u32_t){
                                htonl(length),
-                               txq_mb2mr(txq, buf),
+                               mlx5_tx_mb2mr(txq, buf),
                                naddr,
                                naddr >> 32,
                        };
@@ -812,7 +559,7 @@ next_seg:
                naddr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
                *dseg = (rte_v128u32_t){
                        htonl(length),
-                       txq_mb2mr(txq, buf),
+                       mlx5_tx_mb2mr(txq, buf),
                        naddr,
                        naddr >> 32,
                };
@@ -991,7 +738,7 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
        rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
        /* Start processing. */
-       txq_complete(txq);
+       mlx5_tx_complete(txq);
        max_elts = (elts_n - (elts_head - txq->elts_tail));
        max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
        if (unlikely(!max_wqe))
@@ -1054,7 +801,7 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        addr = rte_pktmbuf_mtod(buf, uintptr_t);
                        *dseg = (struct mlx5_wqe_data_seg){
                                .byte_count = htonl(DATA_LEN(buf)),
-                               .lkey = txq_mb2mr(txq, buf),
+                               .lkey = mlx5_tx_mb2mr(txq, buf),
                                .addr = htonll(addr),
                        };
 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
@@ -1214,7 +961,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
        rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
        rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
        /* Start processing. */
-       txq_complete(txq);
+       mlx5_tx_complete(txq);
        max_elts = (elts_n - (elts_head - txq->elts_tail));
        do {
                struct rte_mbuf *buf = *(pkts++);
@@ -1300,7 +1047,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
                                addr = rte_pktmbuf_mtod(buf, uintptr_t);
                                *dseg = (struct mlx5_wqe_data_seg){
                                        .byte_count = htonl(DATA_LEN(buf)),
-                                       .lkey = txq_mb2mr(txq, buf),
+                                       .lkey = mlx5_tx_mb2mr(txq, buf),
                                        .addr = htonll(addr),
                                };
 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
@@ -1495,7 +1242,7 @@ mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
        if (unlikely(!pkts_n))
                return 0;
        /* Start processing. */
-       txq_complete(txq);
+       mlx5_tx_complete(txq);
        max_elts = (elts_n - (elts_head - txq->elts_tail));
        /* A CQE slot must always be available. */
        assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
@@ -1607,7 +1354,7 @@ mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                                addr = rte_pktmbuf_mtod(buf, uintptr_t);
                                *dseg = (struct mlx5_wqe_data_seg){
                                        .byte_count = htonl(DATA_LEN(buf)),
-                                       .lkey = txq_mb2mr(txq, buf),
+                                       .lkey = mlx5_tx_mb2mr(txq, buf),
                                        .addr = htonll(addr),
                                };
 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
@@ -1690,7 +1437,7 @@ mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        naddr = htonll(addr);
                        *dseg = (rte_v128u32_t) {
                                htonl(length),
-                               txq_mb2mr(txq, buf),
+                               mlx5_tx_mb2mr(txq, buf),
                                naddr,
                                naddr >> 32,
                        };
@@ -2138,3 +1885,71 @@ removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
        (void)pkts_n;
        return 0;
 }
+
+/*
+ * Vectorized Rx/Tx routines are not compiled in when required vector
+ * instructions are not supported on a target architecture. The following null
+ * stubs are needed for linkage when those are not included outside of this file
+ * (e.g.  mlx5_rxtx_vec_sse.c for x86).
+ */
+
+uint16_t __attribute__((weak))
+mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       (void)dpdk_txq;
+       (void)pkts;
+       (void)pkts_n;
+       return 0;
+}
+
+uint16_t __attribute__((weak))
+mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       (void)dpdk_txq;
+       (void)pkts;
+       (void)pkts_n;
+       return 0;
+}
+
+uint16_t __attribute__((weak))
+mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       (void)dpdk_rxq;
+       (void)pkts;
+       (void)pkts_n;
+       return 0;
+}
+
+int __attribute__((weak))
+priv_check_raw_vec_tx_support(struct priv *priv)
+{
+       (void)priv;
+       return -ENOTSUP;
+}
+
+int __attribute__((weak))
+priv_check_vec_tx_support(struct priv *priv)
+{
+       (void)priv;
+       return -ENOTSUP;
+}
+
+int __attribute__((weak))
+rxq_check_vec_support(struct rxq *rxq)
+{
+       (void)rxq;
+       return -ENOTSUP;
+}
+
+int __attribute__((weak))
+priv_check_vec_rx_support(struct priv *priv)
+{
+       (void)priv;
+       return -ENOTSUP;
+}
+
+void __attribute__((weak))
+priv_prep_vec_rx_function(struct priv *priv)
+{
+       (void)priv;
+}
index d0f508e..534aaeb 100644 (file)
@@ -115,10 +115,13 @@ struct rxq {
        unsigned int port_id:8;
        unsigned int rss_hash:1; /* RSS hash result is enabled. */
        unsigned int mark:1; /* Marked flow available on the queue. */
-       unsigned int :8; /* Remaining bits. */
+       unsigned int pending_err:1; /* CQE error needs to be handled. */
+       unsigned int trim_elts:1; /* Whether elts needs clean-up. */
+       unsigned int :6; /* Remaining bits. */
        volatile uint32_t *rq_db;
        volatile uint32_t *cq_db;
        uint16_t rq_ci;
+       uint16_t rq_pi;
        uint16_t cq_ci;
        volatile struct mlx5_wqe_data_seg(*wqes)[];
        volatile struct mlx5_cqe(*cqes)[];
@@ -126,6 +129,8 @@ struct rxq {
        struct rte_mbuf *(*elts)[];
        struct rte_mempool *mp;
        struct mlx5_rxq_stats stats;
+       uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
+       struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
 } __rte_cache_aligned;
 
 /* RX queue control descriptor. */
@@ -261,6 +266,7 @@ struct txq {
        uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
        uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
        uint32_t qp_num_8s; /* QP number shifted by 8. */
+       uint32_t flags; /* Flags for Tx Queue. */
        volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
        volatile void *wqes; /* Work queue (use volatile to write into). */
        volatile uint32_t *qp_db; /* Work queue doorbell. */
@@ -328,6 +334,8 @@ uint16_t mlx5_tx_burst_secondary_setup(void *, struct rte_mbuf **, uint16_t);
 
 /* mlx5_rxtx.c */
 
+extern const uint32_t mlx5_ptype_table[];
+
 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
@@ -338,10 +346,282 @@ uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
 int mlx5_rx_descriptor_status(void *, uint16_t);
 int mlx5_tx_descriptor_status(void *, uint16_t);
 
+/* Vectorized version of mlx5_rxtx.c */
+int priv_check_raw_vec_tx_support(struct priv *);
+int priv_check_vec_tx_support(struct priv *);
+int rxq_check_vec_support(struct rxq *);
+int priv_check_vec_rx_support(struct priv *);
+void priv_prep_vec_rx_function(struct priv *);
+uint16_t mlx5_tx_burst_raw_vec(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_tx_burst_vec(void *, struct rte_mbuf **, uint16_t);
+uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
+
 /* mlx5_mr.c */
 
 struct ibv_mr *mlx5_mp2mr(struct ibv_pd *, struct rte_mempool *);
 void txq_mp2mr_iter(struct rte_mempool *, void *);
 uint32_t txq_mp2mr_reg(struct txq *, struct rte_mempool *, unsigned int);
 
+#ifndef NDEBUG
+/**
+ * Verify or set magic value in CQE.
+ *
+ * @param cqe
+ *   Pointer to CQE.
+ *
+ * @return
+ *   0 the first time.
+ */
+static inline int
+check_cqe_seen(volatile struct mlx5_cqe *cqe)
+{
+       static const uint8_t magic[] = "seen";
+       volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
+       int ret = 1;
+       unsigned int i;
+
+       for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
+               if (!ret || (*buf)[i] != magic[i]) {
+                       ret = 0;
+                       (*buf)[i] = magic[i];
+               }
+       return ret;
+}
+#endif /* NDEBUG */
+
+/**
+ * Check whether CQE is valid.
+ *
+ * @param cqe
+ *   Pointer to CQE.
+ * @param cqes_n
+ *   Size of completion queue.
+ * @param ci
+ *   Consumer index.
+ *
+ * @return
+ *   0 on success, 1 on failure.
+ */
+static __rte_always_inline int
+check_cqe(volatile struct mlx5_cqe *cqe,
+         unsigned int cqes_n, const uint16_t ci)
+{
+       uint16_t idx = ci & cqes_n;
+       uint8_t op_own = cqe->op_own;
+       uint8_t op_owner = MLX5_CQE_OWNER(op_own);
+       uint8_t op_code = MLX5_CQE_OPCODE(op_own);
+
+       if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
+               return 1; /* No CQE. */
+#ifndef NDEBUG
+       if ((op_code == MLX5_CQE_RESP_ERR) ||
+           (op_code == MLX5_CQE_REQ_ERR)) {
+               volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
+               uint8_t syndrome = err_cqe->syndrome;
+
+               if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
+                   (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
+                       return 0;
+               if (!check_cqe_seen(cqe))
+                       ERROR("unexpected CQE error %u (0x%02x)"
+                             " syndrome 0x%02x",
+                             op_code, op_code, syndrome);
+               return 1;
+       } else if ((op_code != MLX5_CQE_RESP_SEND) &&
+                  (op_code != MLX5_CQE_REQ)) {
+               if (!check_cqe_seen(cqe))
+                       ERROR("unexpected CQE opcode %u (0x%02x)",
+                             op_code, op_code);
+               return 1;
+       }
+#endif /* NDEBUG */
+       return 0;
+}
+
+/**
+ * Return the address of the WQE.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param  wqe_ci
+ *   WQE consumer index.
+ *
+ * @return
+ *   WQE address.
+ */
+static inline uintptr_t *
+tx_mlx5_wqe(struct txq *txq, uint16_t ci)
+{
+       ci &= ((1 << txq->wqe_n) - 1);
+       return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
+}
+
+/**
+ * Manage TX completions.
+ *
+ * When sending a burst, mlx5_tx_burst() posts several WRs.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ */
+static __rte_always_inline void
+mlx5_tx_complete(struct txq *txq)
+{
+       const uint16_t elts_n = 1 << txq->elts_n;
+       const uint16_t elts_m = elts_n - 1;
+       const unsigned int cqe_n = 1 << txq->cqe_n;
+       const unsigned int cqe_cnt = cqe_n - 1;
+       uint16_t elts_free = txq->elts_tail;
+       uint16_t elts_tail;
+       uint16_t cq_ci = txq->cq_ci;
+       volatile struct mlx5_cqe *cqe = NULL;
+       volatile struct mlx5_wqe_ctrl *ctrl;
+       struct rte_mbuf *m, *free[elts_n];
+       struct rte_mempool *pool = NULL;
+       unsigned int blk_n = 0;
+
+       do {
+               volatile struct mlx5_cqe *tmp;
+
+               tmp = &(*txq->cqes)[cq_ci & cqe_cnt];
+               if (check_cqe(tmp, cqe_n, cq_ci))
+                       break;
+               cqe = tmp;
+#ifndef NDEBUG
+               if (MLX5_CQE_FORMAT(cqe->op_own) == MLX5_COMPRESSED) {
+                       if (!check_cqe_seen(cqe))
+                               ERROR("unexpected compressed CQE, TX stopped");
+                       return;
+               }
+               if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
+                   (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
+                       if (!check_cqe_seen(cqe))
+                               ERROR("unexpected error CQE, TX stopped");
+                       return;
+               }
+#endif /* NDEBUG */
+               ++cq_ci;
+       } while (1);
+       if (unlikely(cqe == NULL))
+               return;
+       txq->wqe_pi = ntohs(cqe->wqe_counter);
+       ctrl = (volatile struct mlx5_wqe_ctrl *)
+               tx_mlx5_wqe(txq, txq->wqe_pi);
+       elts_tail = ctrl->ctrl3;
+       assert((elts_tail & elts_m) < (1 << txq->wqe_n));
+       /* Free buffers. */
+       while (elts_free != elts_tail) {
+               m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
+               if (likely(m != NULL)) {
+                       if (likely(m->pool == pool)) {
+                               free[blk_n++] = m;
+                       } else {
+                               if (likely(pool != NULL))
+                                       rte_mempool_put_bulk(pool,
+                                                            (void *)free,
+                                                            blk_n);
+                               free[0] = m;
+                               pool = m->pool;
+                               blk_n = 1;
+                       }
+               }
+       }
+       if (blk_n)
+               rte_mempool_put_bulk(pool, (void *)free, blk_n);
+#ifndef NDEBUG
+       elts_free = txq->elts_tail;
+       /* Poisoning. */
+       while (elts_free != elts_tail) {
+               memset(&(*txq->elts)[elts_free & elts_m],
+                      0x66,
+                      sizeof((*txq->elts)[elts_free & elts_m]));
+               ++elts_free;
+       }
+#endif
+       txq->cq_ci = cq_ci;
+       txq->elts_tail = elts_tail;
+       /* Update the consumer index. */
+       rte_wmb();
+       *txq->cq_db = htonl(cq_ci);
+}
+
+/**
+ * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
+ * the cloned mbuf is allocated is returned instead.
+ *
+ * @param buf
+ *   Pointer to mbuf.
+ *
+ * @return
+ *   Memory pool where data is located for given mbuf.
+ */
+static struct rte_mempool *
+mlx5_tx_mb2mp(struct rte_mbuf *buf)
+{
+       if (unlikely(RTE_MBUF_INDIRECT(buf)))
+               return rte_mbuf_from_indirect(buf)->pool;
+       return buf->pool;
+}
+
+/**
+ * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
+ * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
+ * remove an entry first.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param[in] mp
+ *   Memory Pool for which a Memory Region lkey must be returned.
+ *
+ * @return
+ *   mr->lkey on success, (uint32_t)-1 on failure.
+ */
+static __rte_always_inline uint32_t
+mlx5_tx_mb2mr(struct txq *txq, struct rte_mbuf *mb)
+{
+       uint16_t i = txq->mr_cache_idx;
+       uintptr_t addr = rte_pktmbuf_mtod(mb, uintptr_t);
+
+       assert(i < RTE_DIM(txq->mp2mr));
+       if (likely(txq->mp2mr[i].start <= addr && txq->mp2mr[i].end >= addr))
+               return txq->mp2mr[i].lkey;
+       for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
+               if (unlikely(txq->mp2mr[i].mr == NULL)) {
+                       /* Unknown MP, add a new MR for it. */
+                       break;
+               }
+               if (txq->mp2mr[i].start <= addr &&
+                   txq->mp2mr[i].end >= addr) {
+                       assert(txq->mp2mr[i].lkey != (uint32_t)-1);
+                       assert(htonl(txq->mp2mr[i].mr->lkey) ==
+                              txq->mp2mr[i].lkey);
+                       txq->mr_cache_idx = i;
+                       return txq->mp2mr[i].lkey;
+               }
+       }
+       txq->mr_cache_idx = 0;
+       return txq_mp2mr_reg(txq, mlx5_tx_mb2mp(mb), i);
+}
+
+/**
+ * Ring TX queue doorbell.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param wqe
+ *   Pointer to the last WQE posted in the NIC.
+ */
+static __rte_always_inline void
+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();
+       *txq->qp_db = htonl(txq->wqe_ci);
+       /* Ensure ordering between DB record and BF copy. */
+       rte_wmb();
+       *dst = *src;
+}
+
 #endif /* RTE_PMD_MLX5_RXTX_H_ */
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec_sse.c b/drivers/net/mlx5/mlx5_rxtx_vec_sse.c
new file mode 100644 (file)
index 0000000..95e96ba
--- /dev/null
@@ -0,0 +1,1394 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright 2017 6WIND S.A.
+ *   Copyright 2017 Mellanox.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of 6WIND S.A. nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <assert.h>
+#include <stdint.h>
+#include <string.h>
+#include <stdlib.h>
+#include <smmintrin.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx5_hw.h>
+#include <infiniband/arch.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
+
+/* DPDK headers don't like -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <rte_mbuf.h>
+#include <rte_mempool.h>
+#include <rte_prefetch.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
+
+#include "mlx5.h"
+#include "mlx5_utils.h"
+#include "mlx5_rxtx.h"
+#include "mlx5_autoconf.h"
+#include "mlx5_defs.h"
+#include "mlx5_prm.h"
+
+#ifndef __INTEL_COMPILER
+#pragma GCC diagnostic ignored "-Wcast-qual"
+#endif
+
+/**
+ * Fill in buffer descriptors in a multi-packet send descriptor.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param dseg
+ *   Pointer to buffer descriptor to be writen.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param n
+ *   Number of packets to be filled.
+ */
+static inline void
+txq_wr_dseg_v(struct txq *txq, __m128i *dseg,
+             struct rte_mbuf **pkts, unsigned int n)
+{
+       unsigned int pos;
+       uintptr_t addr;
+       const __m128i shuf_mask_dseg =
+               _mm_set_epi8(8,  9, 10, 11, /* addr, bswap64 */
+                           12, 13, 14, 15,
+                            7,  6,  5,  4, /* lkey */
+                            0,  1,  2,  3  /* length, bswap32 */);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       uint32_t tx_byte = 0;
+#endif
+
+       for (pos = 0; pos < n; ++pos, ++dseg) {
+               __m128i desc;
+               struct rte_mbuf *pkt = pkts[pos];
+
+               addr = rte_pktmbuf_mtod(pkt, uintptr_t);
+               desc = _mm_set_epi32(addr >> 32,
+                                    addr,
+                                    mlx5_tx_mb2mr(txq, pkt),
+                                    DATA_LEN(pkt));
+               desc = _mm_shuffle_epi8(desc, shuf_mask_dseg);
+               _mm_store_si128(dseg, desc);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               tx_byte += DATA_LEN(pkt);
+#endif
+       }
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       txq->stats.obytes += tx_byte;
+#endif
+}
+
+/**
+ * Count the number of continuous single segment packets. The first packet must
+ * be a single segment packet.
+ *
+ * @param pkts
+ *   Pointer to array of packets.
+ * @param pkts_n
+ *   Number of packets.
+ *
+ * @return
+ *   Number of continuous single segment packets.
+ */
+static inline unsigned int
+txq_check_multiseg(struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       unsigned int pos;
+
+       if (!pkts_n)
+               return 0;
+       assert(NB_SEGS(pkts[0]) == 1);
+       /* Count the number of continuous single segment packets. */
+       for (pos = 1; pos < pkts_n; ++pos)
+               if (NB_SEGS(pkts[pos]) > 1)
+                       break;
+       return pos;
+}
+
+/**
+ * Count the number of packets having same ol_flags and calculate cs_flags.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets.
+ * @param pkts_n
+ *   Number of packets.
+ * @param cs_flags
+ *   Pointer of flags to be returned.
+ *
+ * @return
+ *   Number of packets having same ol_flags.
+ */
+static inline unsigned int
+txq_calc_offload(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
+                uint8_t *cs_flags)
+{
+       unsigned int pos;
+       const uint64_t ol_mask =
+               PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM |
+               PKT_TX_UDP_CKSUM | PKT_TX_TUNNEL_GRE |
+               PKT_TX_TUNNEL_VXLAN | PKT_TX_OUTER_IP_CKSUM;
+
+       if (!pkts_n)
+               return 0;
+       /* Count the number of packets having same ol_flags. */
+       for (pos = 1; pos < pkts_n; ++pos)
+               if ((pkts[pos]->ol_flags ^ pkts[0]->ol_flags) & ol_mask)
+                       break;
+       /* Should open another MPW session for the rest. */
+       if (pkts[0]->ol_flags &
+           (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+               const uint64_t is_tunneled =
+                       pkts[0]->ol_flags &
+                       (PKT_TX_TUNNEL_GRE |
+                        PKT_TX_TUNNEL_VXLAN);
+
+               if (is_tunneled && txq->tunnel_en) {
+                       *cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
+                                   MLX5_ETH_WQE_L4_INNER_CSUM;
+                       if (pkts[0]->ol_flags & PKT_TX_OUTER_IP_CKSUM)
+                               *cs_flags |= MLX5_ETH_WQE_L3_CSUM;
+               } else {
+                       *cs_flags = MLX5_ETH_WQE_L3_CSUM |
+                                   MLX5_ETH_WQE_L4_CSUM;
+               }
+       }
+       return pos;
+}
+
+/**
+ * Send multi-segmented packets until it encounters a single segment packet in
+ * the pkts list.
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param pkts_n
+ *   Number of packets to be sent.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+static uint16_t
+txq_scatter_v(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       uint16_t elts_head = txq->elts_head;
+       const uint16_t elts_n = 1 << txq->elts_n;
+       const uint16_t elts_m = elts_n - 1;
+       const uint16_t wq_n = 1 << txq->wqe_n;
+       const uint16_t wq_mask = wq_n - 1;
+       const unsigned int nb_dword_per_wqebb =
+               MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
+       const unsigned int nb_dword_in_hdr =
+               sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
+       unsigned int n;
+       volatile struct mlx5_wqe *wqe = NULL;
+
+       assert(elts_n > pkts_n);
+       mlx5_tx_complete(txq);
+       if (unlikely(!pkts_n))
+               return 0;
+       for (n = 0; n < pkts_n; ++n) {
+               struct rte_mbuf *buf = pkts[n];
+               unsigned int segs_n = buf->nb_segs;
+               unsigned int ds = nb_dword_in_hdr;
+               unsigned int len = PKT_LEN(buf);
+               uint16_t wqe_ci = txq->wqe_ci;
+               const __m128i shuf_mask_ctrl =
+                       _mm_set_epi8(15, 14, 13, 12,
+                                     8,  9, 10, 11, /* bswap32 */
+                                     4,  5,  6,  7, /* bswap32 */
+                                     0,  1,  2,  3  /* bswap32 */);
+               uint8_t cs_flags = 0;
+               uint16_t max_elts;
+               uint16_t max_wqe;
+               __m128i *t_wqe, *dseg;
+               __m128i ctrl;
+
+               assert(segs_n);
+               max_elts = elts_n - (elts_head - txq->elts_tail);
+               max_wqe = wq_n - (txq->wqe_ci - txq->wqe_pi);
+               /*
+                * A MPW session consumes 2 WQEs at most to
+                * include MLX5_MPW_DSEG_MAX pointers.
+                */
+               if (segs_n == 1 ||
+                   max_elts < segs_n || max_wqe < 2)
+                       break;
+               wqe = &((volatile struct mlx5_wqe64 *)
+                        txq->wqes)[wqe_ci & wq_mask].hdr;
+               if (buf->ol_flags &
+                    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
+                       const uint64_t is_tunneled = buf->ol_flags &
+                                                     (PKT_TX_TUNNEL_GRE |
+                                                      PKT_TX_TUNNEL_VXLAN);
+
+                       if (is_tunneled && txq->tunnel_en) {
+                               cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
+                                          MLX5_ETH_WQE_L4_INNER_CSUM;
+                               if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
+                                       cs_flags |= MLX5_ETH_WQE_L3_CSUM;
+                       } else {
+                               cs_flags = MLX5_ETH_WQE_L3_CSUM |
+                                          MLX5_ETH_WQE_L4_CSUM;
+                       }
+               }
+               /* Title WQEBB pointer. */
+               t_wqe = (__m128i *)wqe;
+               dseg = (__m128i *)(wqe + 1);
+               do {
+                       if (!(ds++ % nb_dword_per_wqebb)) {
+                               dseg = (__m128i *)
+                                       &((volatile struct mlx5_wqe64 *)
+                                          txq->wqes)[++wqe_ci & wq_mask];
+                       }
+                       txq_wr_dseg_v(txq, dseg++, &buf, 1);
+                       (*txq->elts)[elts_head++ & elts_m] = buf;
+                       buf = buf->next;
+               } while (--segs_n);
+               if (ds % nb_dword_per_wqebb)
+                       ++wqe_ci;
+               /* Fill CTRL in the header. */
+               ctrl = _mm_set_epi32(0, 0, txq->qp_num_8s | ds,
+                                    MLX5_OPC_MOD_MPW << 24 |
+                                    txq->wqe_ci << 8 | MLX5_OPCODE_TSO);
+               ctrl = _mm_shuffle_epi8(ctrl, shuf_mask_ctrl);
+               _mm_store_si128(t_wqe, ctrl);
+               /* Fill ESEG in the header. */
+               _mm_store_si128(t_wqe + 1,
+                               _mm_set_epi16(0, 0, 0, 0,
+                                             htons(len), cs_flags,
+                                             0, 0));
+               txq->wqe_ci = wqe_ci;
+       }
+       if (!n)
+               return 0;
+       txq->elts_comp += (uint16_t)(elts_head - txq->elts_head);
+       txq->elts_head = elts_head;
+       if (txq->elts_comp >= MLX5_TX_COMP_THRESH) {
+               wqe->ctrl[2] = htonl(8);
+               wqe->ctrl[3] = txq->elts_head;
+               txq->elts_comp = 0;
+               ++txq->cq_pi;
+       }
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       txq->stats.opackets += n;
+#endif
+       mlx5_tx_dbrec(txq, wqe);
+       return n;
+}
+
+/**
+ * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
+ * it returns to make it processed by txq_scatter_v(). All the packets in
+ * the pkts list should be single segment packets having same offload flags.
+ * This must be checked by txq_check_multiseg() and txq_calc_offload().
+ *
+ * @param txq
+ *   Pointer to TX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be sent.
+ * @param pkts_n
+ *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
+ * @param cs_flags
+ *   Checksum offload flags to be written in the descriptor.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+static inline uint16_t
+txq_burst_v(struct txq *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
+           uint8_t cs_flags)
+{
+       struct rte_mbuf **elts;
+       uint16_t elts_head = txq->elts_head;
+       const uint16_t elts_n = 1 << txq->elts_n;
+       const uint16_t elts_m = elts_n - 1;
+       const unsigned int nb_dword_per_wqebb =
+               MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
+       const unsigned int nb_dword_in_hdr =
+               sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
+       unsigned int n = 0;
+       unsigned int pos;
+       uint16_t max_elts;
+       uint16_t max_wqe;
+       uint32_t comp_req = 0;
+       const uint16_t wq_n = 1 << txq->wqe_n;
+       const uint16_t wq_mask = wq_n - 1;
+       uint16_t wq_idx = txq->wqe_ci & wq_mask;
+       volatile struct mlx5_wqe64 *wq =
+               &((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
+       volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
+       const __m128i shuf_mask_ctrl =
+               _mm_set_epi8(15, 14, 13, 12,
+                             8,  9, 10, 11, /* bswap32 */
+                             4,  5,  6,  7, /* bswap32 */
+                             0,  1,  2,  3  /* bswap32 */);
+       __m128i *t_wqe, *dseg;
+       __m128i ctrl;
+
+       /* Make sure all packets can fit into a single WQE. */
+       assert(elts_n > pkts_n);
+       mlx5_tx_complete(txq);
+       max_elts = (elts_n - (elts_head - txq->elts_tail));
+       max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
+       pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
+       if (unlikely(!pkts_n))
+               return 0;
+       elts = &(*txq->elts)[elts_head & elts_m];
+       /* Loop for available tailroom first. */
+       n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
+       for (pos = 0; pos < (n & -2); pos += 2)
+               _mm_storeu_si128((__m128i *)&elts[pos],
+                                _mm_loadu_si128((__m128i *)&pkts[pos]));
+       if (n & 1)
+               elts[pos] = pkts[pos];
+       /* Check if it crosses the end of the queue. */
+       if (unlikely(n < pkts_n)) {
+               elts = &(*txq->elts)[0];
+               for (pos = 0; pos < pkts_n - n; ++pos)
+                       elts[pos] = pkts[n + pos];
+       }
+       txq->elts_head += pkts_n;
+       /* Save title WQEBB pointer. */
+       t_wqe = (__m128i *)wqe;
+       dseg = (__m128i *)(wqe + 1);
+       /* Calculate the number of entries to the end. */
+       n = RTE_MIN(
+               (wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
+               pkts_n);
+       /* Fill DSEGs. */
+       txq_wr_dseg_v(txq, dseg, pkts, n);
+       /* Check if it crosses the end of the queue. */
+       if (n < pkts_n) {
+               dseg = (__m128i *)txq->wqes;
+               txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
+       }
+       if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
+               txq->elts_comp += pkts_n;
+       } else {
+               /* Request a completion. */
+               txq->elts_comp = 0;
+               ++txq->cq_pi;
+               comp_req = 8;
+       }
+       /* Fill CTRL in the header. */
+       ctrl = _mm_set_epi32(txq->elts_head, comp_req,
+                            txq->qp_num_8s | (pkts_n + 2),
+                            MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
+                               txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW);
+       ctrl = _mm_shuffle_epi8(ctrl, shuf_mask_ctrl);
+       _mm_store_si128(t_wqe, ctrl);
+       /* Fill ESEG in the header. */
+       _mm_store_si128(t_wqe + 1,
+                       _mm_set_epi8(0, 0, 0, 0,
+                                    0, 0, 0, 0,
+                                    0, 0, 0, cs_flags,
+                                    0, 0, 0, 0));
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       txq->stats.opackets += pkts_n;
+#endif
+       txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
+                      nb_dword_per_wqebb;
+       /* Ring QP doorbell. */
+       mlx5_tx_dbrec(txq, wqe);
+       return pkts_n;
+}
+
+/**
+ * DPDK callback for vectorized TX.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
+                     uint16_t pkts_n)
+{
+       struct txq *txq = (struct txq *)dpdk_txq;
+       uint16_t nb_tx = 0;
+
+       while (pkts_n > nb_tx) {
+               uint16_t n;
+               uint16_t ret;
+
+               n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
+               ret = txq_burst_v(txq, &pkts[nb_tx], n, 0);
+               nb_tx += ret;
+               if (!ret)
+                       break;
+       }
+       return nb_tx;
+}
+
+/**
+ * DPDK callback for vectorized TX with multi-seg packets and offload.
+ *
+ * @param dpdk_txq
+ *   Generic pointer to TX queue structure.
+ * @param[in] pkts
+ *   Packets to transmit.
+ * @param pkts_n
+ *   Number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully transmitted (<= pkts_n).
+ */
+uint16_t
+mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       struct txq *txq = (struct txq *)dpdk_txq;
+       uint16_t nb_tx = 0;
+
+       while (pkts_n > nb_tx) {
+               uint8_t cs_flags = 0;
+               uint16_t n;
+               uint16_t ret;
+
+               /* Transmit multi-seg packets in the head of pkts list. */
+               if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) &&
+                   NB_SEGS(pkts[nb_tx]) > 1)
+                       nb_tx += txq_scatter_v(txq,
+                                              &pkts[nb_tx],
+                                              pkts_n - nb_tx);
+               n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
+               if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS))
+                       n = txq_check_multiseg(&pkts[nb_tx], n);
+               if (!(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
+                       n = txq_calc_offload(txq, &pkts[nb_tx], n, &cs_flags);
+               ret = txq_burst_v(txq, &pkts[nb_tx], n, cs_flags);
+               nb_tx += ret;
+               if (!ret)
+                       break;
+       }
+       return nb_tx;
+}
+
+/**
+ * Store free buffers to RX SW ring.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param pkts
+ *   Pointer to array of packets to be stored.
+ * @param pkts_n
+ *   Number of packets to be stored.
+ */
+static inline void
+rxq_copy_mbuf_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t n)
+{
+       const uint16_t q_mask = (1 << rxq->elts_n) - 1;
+       struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
+       unsigned int pos;
+       uint16_t p = n & -2;
+
+       for (pos = 0; pos < p; pos += 2) {
+               __m128i mbp;
+
+               mbp = _mm_loadu_si128((__m128i *)&elts[pos]);
+               _mm_storeu_si128((__m128i *)&pkts[pos], mbp);
+       }
+       if (n & 1)
+               pkts[pos] = elts[pos];
+}
+
+/**
+ * Replenish buffers for RX in bulk.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param n
+ *   Number of buffers to be replenished.
+ */
+static inline void
+rxq_replenish_bulk_mbuf(struct rxq *rxq, uint16_t n)
+{
+       const uint16_t q_n = 1 << rxq->elts_n;
+       const uint16_t q_mask = q_n - 1;
+       const uint16_t elts_idx = rxq->rq_ci & q_mask;
+       struct rte_mbuf **elts = &(*rxq->elts)[elts_idx];
+       volatile struct mlx5_wqe_data_seg *wq = &(*rxq->wqes)[elts_idx];
+       unsigned int i;
+
+       assert(n >= MLX5_VPMD_RXQ_RPLNSH_THRESH);
+       assert(n <= (uint16_t)(q_n - (rxq->rq_ci - rxq->rq_pi)));
+       assert(MLX5_VPMD_RXQ_RPLNSH_THRESH > MLX5_VPMD_DESCS_PER_LOOP);
+       /* Not to cross queue end. */
+       n = RTE_MIN(n - MLX5_VPMD_DESCS_PER_LOOP, q_n - elts_idx);
+       if (rte_mempool_get_bulk(rxq->mp, (void *)elts, n) < 0) {
+               rxq->stats.rx_nombuf += n;
+               return;
+       }
+       for (i = 0; i < n; ++i)
+               wq[i].addr = htonll(rte_pktmbuf_mtod(elts[i], uintptr_t));
+       rxq->rq_ci += n;
+       rte_wmb();
+       *rxq->rq_db = htonl(rxq->rq_ci);
+}
+
+/**
+ * Decompress a compressed completion and fill in mbufs in RX SW ring with data
+ * extracted from the title completion descriptor.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param cq
+ *   Pointer to completion array having a compressed completion at first.
+ * @param elts
+ *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
+ *   the title completion descriptor to be copied to the rest of mbufs.
+ */
+static inline void
+rxq_cq_decompress_v(struct rxq *rxq,
+                   volatile struct mlx5_cqe *cq,
+                   struct rte_mbuf **elts)
+{
+       volatile struct mlx5_mini_cqe8 *mcq = (void *)(cq + 1);
+       struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
+       unsigned int pos;
+       unsigned int i;
+       unsigned int inv = 0;
+       /* Mask to shuffle from extracted mini CQE to mbuf. */
+       const __m128i shuf_mask1 =
+               _mm_set_epi8(0,  1,  2,  3, /* rss, bswap32 */
+                           -1, -1,         /* skip vlan_tci */
+                            6,  7,         /* data_len, bswap16 */
+                           -1, -1,  6,  7, /* pkt_len, bswap16 */
+                           -1, -1, -1, -1  /* skip packet_type */);
+       const __m128i shuf_mask2 =
+               _mm_set_epi8(8,  9, 10, 11, /* rss, bswap32 */
+                           -1, -1,         /* skip vlan_tci */
+                           14, 15,         /* data_len, bswap16 */
+                           -1, -1, 14, 15, /* pkt_len, bswap16 */
+                           -1, -1, -1, -1  /* skip packet_type */);
+       /* Restore the compressed count. Must be 16 bits. */
+       const uint16_t mcqe_n = t_pkt->data_len +
+                               (rxq->crc_present * ETHER_CRC_LEN);
+       const __m128i rearm =
+               _mm_loadu_si128((__m128i *)&t_pkt->rearm_data);
+       const __m128i rxdf =
+               _mm_loadu_si128((__m128i *)&t_pkt->rx_descriptor_fields1);
+       const __m128i crc_adj =
+               _mm_set_epi16(0, 0, 0,
+                             rxq->crc_present * ETHER_CRC_LEN,
+                             0,
+                             rxq->crc_present * ETHER_CRC_LEN,
+                             0, 0);
+       const uint32_t flow_tag = t_pkt->hash.fdir.hi;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       const __m128i zero = _mm_setzero_si128();
+       const __m128i ones = _mm_cmpeq_epi32(zero, zero);
+       uint32_t rcvd_byte = 0;
+       /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
+       const __m128i len_shuf_mask =
+               _mm_set_epi8(-1, -1, -1, -1,
+                            -1, -1, -1, -1,
+                            14, 15,  6,  7,
+                            10, 11,  2,  3);
+#endif
+
+       /* Compile time sanity check for this function. */
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+                        offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+                        offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
+                        offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
+       /*
+        * A. load mCQEs into a 128bit register.
+        * B. store rearm data to mbuf.
+        * C. combine data from mCQEs with rx_descriptor_fields1.
+        * D. store rx_descriptor_fields1.
+        * E. store flow tag (rte_flow mark).
+        */
+       for (pos = 0; pos < mcqe_n; ) {
+               __m128i mcqe1, mcqe2;
+               __m128i rxdf1, rxdf2;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               __m128i byte_cnt, invalid_mask;
+#endif
+
+               if (!(pos & 0x7) && pos + 8 < mcqe_n)
+                       rte_prefetch0((void *)(cq + pos + 8));
+               /* A.1 load mCQEs into a 128bit register. */
+               mcqe1 = _mm_loadu_si128((__m128i *)&mcq[pos % 8]);
+               mcqe2 = _mm_loadu_si128((__m128i *)&mcq[pos % 8 + 2]);
+               /* B.1 store rearm data to mbuf. */
+               _mm_storeu_si128((__m128i *)&elts[pos]->rearm_data, rearm);
+               _mm_storeu_si128((__m128i *)&elts[pos + 1]->rearm_data, rearm);
+               /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+               rxdf1 = _mm_shuffle_epi8(mcqe1, shuf_mask1);
+               rxdf2 = _mm_shuffle_epi8(mcqe1, shuf_mask2);
+               rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
+               rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
+               rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
+               rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
+               /* D.1 store rx_descriptor_fields1. */
+               _mm_storeu_si128((__m128i *)
+                                 &elts[pos]->rx_descriptor_fields1,
+                                rxdf1);
+               _mm_storeu_si128((__m128i *)
+                                 &elts[pos + 1]->rx_descriptor_fields1,
+                                rxdf2);
+               /* B.1 store rearm data to mbuf. */
+               _mm_storeu_si128((__m128i *)&elts[pos + 2]->rearm_data, rearm);
+               _mm_storeu_si128((__m128i *)&elts[pos + 3]->rearm_data, rearm);
+               /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
+               rxdf1 = _mm_shuffle_epi8(mcqe2, shuf_mask1);
+               rxdf2 = _mm_shuffle_epi8(mcqe2, shuf_mask2);
+               rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
+               rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
+               rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
+               rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
+               /* D.1 store rx_descriptor_fields1. */
+               _mm_storeu_si128((__m128i *)
+                                 &elts[pos + 2]->rx_descriptor_fields1,
+                                rxdf1);
+               _mm_storeu_si128((__m128i *)
+                                 &elts[pos + 3]->rx_descriptor_fields1,
+                                rxdf2);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               invalid_mask = _mm_set_epi64x(0,
+                                             (mcqe_n - pos) *
+                                             sizeof(uint16_t) * 8);
+               invalid_mask = _mm_sll_epi64(ones, invalid_mask);
+               mcqe1 = _mm_srli_si128(mcqe1, 4);
+               byte_cnt = _mm_blend_epi16(mcqe1, mcqe2, 0xcc);
+               byte_cnt = _mm_shuffle_epi8(byte_cnt, len_shuf_mask);
+               byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
+               byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
+               rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
+#endif
+               if (rxq->mark) {
+                       /* E.1 store flow tag (rte_flow mark). */
+                       elts[pos]->hash.fdir.hi = flow_tag;
+                       elts[pos + 1]->hash.fdir.hi = flow_tag;
+                       elts[pos + 2]->hash.fdir.hi = flow_tag;
+                       elts[pos + 3]->hash.fdir.hi = flow_tag;
+               }
+               pos += MLX5_VPMD_DESCS_PER_LOOP;
+               /* Move to next CQE and invalidate consumed CQEs. */
+               if (!(pos & 0x7) && pos < mcqe_n) {
+                       mcq = (void *)(cq + pos);
+                       for (i = 0; i < 8; ++i)
+                               cq[inv++].op_own = MLX5_CQE_INVALIDATE;
+               }
+       }
+       /* Invalidate the rest of CQEs. */
+       for (; inv < mcqe_n; ++inv)
+               cq[inv].op_own = MLX5_CQE_INVALIDATE;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       rxq->stats.ipackets += mcqe_n;
+       rxq->stats.ibytes += rcvd_byte;
+#endif
+       rxq->cq_ci += mcqe_n;
+}
+
+/**
+ * Calculate packet type and offload flag for mbuf and store it.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param cqes[4]
+ *   Array of four 16bytes completions extracted from the original completion
+ *   descriptor.
+ * @param op_err
+ *   Opcode vector having responder error status. Each field is 4B.
+ * @param pkts
+ *   Pointer to array of packets to be filled.
+ */
+static inline void
+rxq_cq_to_ptype_oflags_v(struct rxq *rxq, __m128i cqes[4], __m128i op_err,
+                        struct rte_mbuf **pkts)
+{
+       __m128i pinfo0, pinfo1;
+       __m128i pinfo, ptype;
+       __m128i ol_flags = _mm_set1_epi32(rxq->rss_hash * PKT_RX_RSS_HASH);
+       __m128i cv_flags;
+       const __m128i zero = _mm_setzero_si128();
+       const __m128i ptype_mask =
+               _mm_set_epi32(0xd06, 0xd06, 0xd06, 0xd06);
+       const __m128i ptype_ol_mask =
+               _mm_set_epi32(0x106, 0x106, 0x106, 0x106);
+       const __m128i pinfo_mask =
+               _mm_set_epi32(0x3, 0x3, 0x3, 0x3);
+       const __m128i cv_flag_sel =
+               _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0,
+                            (uint8_t)((PKT_RX_IP_CKSUM_GOOD |
+                                       PKT_RX_L4_CKSUM_GOOD) >> 1),
+                            0,
+                            (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
+                            0,
+                            (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
+                            (uint8_t)(PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED),
+                            0);
+       const __m128i cv_mask =
+               _mm_set_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+                             PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+                             PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+                             PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+                             PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+                             PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED,
+                             PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
+                             PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED);
+       const __m128i mbuf_init =
+               _mm_loadl_epi64((__m128i *)&rxq->mbuf_initializer);
+       __m128i rearm0, rearm1, rearm2, rearm3;
+
+       /* Extract pkt_info field. */
+       pinfo0 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
+       pinfo1 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
+       pinfo = _mm_unpacklo_epi64(pinfo0, pinfo1);
+       /* Extract hdr_type_etc field. */
+       pinfo0 = _mm_unpackhi_epi32(cqes[0], cqes[1]);
+       pinfo1 = _mm_unpackhi_epi32(cqes[2], cqes[3]);
+       ptype = _mm_unpacklo_epi64(pinfo0, pinfo1);
+       if (rxq->mark) {
+               const __m128i pinfo_ft_mask =
+                       _mm_set_epi32(0xffffff00, 0xffffff00,
+                                     0xffffff00, 0xffffff00);
+               const __m128i fdir_flags = _mm_set1_epi32(PKT_RX_FDIR);
+               const __m128i fdir_id_flags = _mm_set1_epi32(PKT_RX_FDIR_ID);
+               __m128i flow_tag, invalid_mask;
+
+               flow_tag = _mm_and_si128(pinfo, pinfo_ft_mask);
+               /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
+               invalid_mask = _mm_cmpeq_epi32(flow_tag, zero);
+               ol_flags = _mm_or_si128(ol_flags,
+                                       _mm_andnot_si128(invalid_mask,
+                                                        fdir_flags));
+               /* Mask out invalid entries. */
+               flow_tag = _mm_andnot_si128(invalid_mask, flow_tag);
+               /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
+               ol_flags = _mm_or_si128(ol_flags,
+                                       _mm_andnot_si128(
+                                               _mm_cmpeq_epi32(flow_tag,
+                                                               pinfo_ft_mask),
+                                               fdir_id_flags));
+       }
+       /*
+        * Merge the two fields to generate the following:
+        * bit[1]  = l3_ok,    bit[2]     = l4_ok
+        * bit[8]  = cv,       bit[11:10] = l3_hdr_type
+        * bit[12] = tunneled, bit[13]    = outer_l3_type
+        */
+       ptype = _mm_and_si128(ptype, ptype_mask);
+       pinfo = _mm_and_si128(pinfo, pinfo_mask);
+       pinfo = _mm_slli_epi32(pinfo, 12);
+       ptype = _mm_or_si128(ptype, pinfo);
+       ptype = _mm_srli_epi32(ptype, 10);
+       ptype = _mm_packs_epi32(ptype, zero);
+       /* Errored packets will have RTE_PTYPE_ALL_MASK. */
+       op_err = _mm_srli_epi16(op_err, 12);
+       ptype = _mm_or_si128(ptype, op_err);
+       pkts[0]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 0)];
+       pkts[1]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 2)];
+       pkts[2]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 4)];
+       pkts[3]->packet_type = mlx5_ptype_table[_mm_extract_epi8(ptype, 6)];
+       /* Fill flags for checksum and VLAN. */
+       pinfo = _mm_and_si128(pinfo, ptype_ol_mask);
+       pinfo = _mm_shuffle_epi8(cv_flag_sel, pinfo);
+       /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
+       cv_flags = _mm_slli_epi32(pinfo, 9);
+       cv_flags = _mm_or_si128(pinfo, cv_flags);
+       /* Move back flags to start from byte[0]. */
+       cv_flags = _mm_srli_epi32(cv_flags, 8);
+       /* Mask out garbage bits. */
+       cv_flags = _mm_and_si128(cv_flags, cv_mask);
+       /* Merge to ol_flags. */
+       ol_flags = _mm_or_si128(ol_flags, cv_flags);
+       /* Merge mbuf_init and ol_flags. */
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
+                        offsetof(struct rte_mbuf, rearm_data) + 8);
+       rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 8), 0x30);
+       rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 4), 0x30);
+       rearm2 = _mm_blend_epi16(mbuf_init, ol_flags, 0x30);
+       rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(ol_flags, 4), 0x30);
+       /* Write 8B rearm_data and 8B ol_flags. */
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
+                        RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
+       _mm_store_si128((__m128i *)&pkts[0]->rearm_data, rearm0);
+       _mm_store_si128((__m128i *)&pkts[1]->rearm_data, rearm1);
+       _mm_store_si128((__m128i *)&pkts[2]->rearm_data, rearm2);
+       _mm_store_si128((__m128i *)&pkts[3]->rearm_data, rearm3);
+}
+
+/**
+ * Skip error packets.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+static uint16_t
+rxq_handle_pending_error(struct rxq *rxq, struct rte_mbuf **pkts,
+                        uint16_t pkts_n)
+{
+       uint16_t n = 0;
+       unsigned int i;
+
+       for (i = 0; i < pkts_n; ++i) {
+               struct rte_mbuf *pkt = pkts[i];
+
+               if (pkt->packet_type == RTE_PTYPE_ALL_MASK)
+                       rte_pktmbuf_free_seg(pkt);
+               else
+                       pkts[n++] = pkt;
+       }
+       rxq->stats.idropped += (pkts_n - n);
+       rxq->pending_err = 0;
+       return n;
+}
+
+/**
+ * Receive burst of packets. An errored completion also consumes a mbuf, but the
+ * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
+ * before returning to application.
+ *
+ * @param rxq
+ *   Pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets received including errors (<= pkts_n).
+ */
+static inline uint16_t
+rxq_burst_v(struct rxq *rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       const uint16_t q_n = 1 << rxq->cqe_n;
+       const uint16_t q_mask = q_n - 1;
+       volatile struct mlx5_cqe *cq;
+       struct rte_mbuf **elts;
+       unsigned int pos;
+       uint64_t n;
+       uint16_t repl_n;
+       uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
+       uint16_t nocmp_n = 0;
+       uint16_t rcvd_pkt = 0;
+       unsigned int cq_idx = rxq->cq_ci & q_mask;
+       unsigned int elts_idx;
+       unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
+       const __m128i owner_check =
+               _mm_set_epi64x(0x0100000001000000LL, 0x0100000001000000LL);
+       const __m128i opcode_check =
+               _mm_set_epi64x(0xf0000000f0000000LL, 0xf0000000f0000000LL);
+       const __m128i format_check =
+               _mm_set_epi64x(0x0c0000000c000000LL, 0x0c0000000c000000LL);
+       const __m128i resp_err_check =
+               _mm_set_epi64x(0xe0000000e0000000LL, 0xe0000000e0000000LL);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       uint32_t rcvd_byte = 0;
+       /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
+       const __m128i len_shuf_mask =
+               _mm_set_epi8(-1, -1, -1, -1,
+                            -1, -1, -1, -1,
+                            12, 13,  8,  9,
+                             4,  5,  0,  1);
+#endif
+       /* Mask to shuffle from extracted CQE to mbuf. */
+       const __m128i shuf_mask =
+               _mm_set_epi8(-1,  3,  2,  1, /* fdir.hi */
+                            12, 13, 14, 15, /* rss, bswap32 */
+                            10, 11,         /* vlan_tci, bswap16 */
+                             4,  5,         /* data_len, bswap16 */
+                            -1, -1,         /* zero out 2nd half of pkt_len */
+                             4,  5          /* pkt_len, bswap16 */);
+       /* Mask to blend from the last Qword to the first DQword. */
+       const __m128i blend_mask =
+               _mm_set_epi8(-1, -1, -1, -1,
+                            -1, -1, -1, -1,
+                             0,  0,  0,  0,
+                             0,  0,  0, -1);
+       const __m128i zero = _mm_setzero_si128();
+       const __m128i ones = _mm_cmpeq_epi32(zero, zero);
+       const __m128i crc_adj =
+               _mm_set_epi16(0, 0, 0, 0, 0,
+                             rxq->crc_present * ETHER_CRC_LEN,
+                             0,
+                             rxq->crc_present * ETHER_CRC_LEN);
+       const __m128i flow_mark_adj = _mm_set_epi32(rxq->mark * (-1), 0, 0, 0);
+
+       /* Compile time sanity check for this function. */
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
+                        offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
+       RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
+                        offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, pkt_info) != 0);
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rx_hash_res) !=
+                        offsetof(struct mlx5_cqe, pkt_info) + 12);
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd1) +
+                         sizeof(((struct mlx5_cqe *)0)->rsvd1) !=
+                        offsetof(struct mlx5_cqe, hdr_type_etc));
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, vlan_info) !=
+                        offsetof(struct mlx5_cqe, hdr_type_etc) + 2);
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, rsvd2) +
+                         sizeof(((struct mlx5_cqe *)0)->rsvd2) !=
+                        offsetof(struct mlx5_cqe, byte_cnt));
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, sop_drop_qpn) !=
+                        RTE_ALIGN(offsetof(struct mlx5_cqe, sop_drop_qpn), 8));
+       RTE_BUILD_BUG_ON(offsetof(struct mlx5_cqe, op_own) !=
+                        offsetof(struct mlx5_cqe, sop_drop_qpn) + 7);
+       assert(rxq->sges_n == 0);
+       assert(rxq->cqe_n == rxq->elts_n);
+       cq = &(*rxq->cqes)[cq_idx];
+       rte_prefetch0(cq);
+       rte_prefetch0(cq + 1);
+       rte_prefetch0(cq + 2);
+       rte_prefetch0(cq + 3);
+       pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
+       /*
+        * Order of indexes:
+        *   rq_ci >= cq_ci >= rq_pi
+        * Definition of indexes:
+        *   rq_ci - cq_ci := # of buffers owned by HW (posted).
+        *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
+        *   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)
+               rxq_replenish_bulk_mbuf(rxq, repl_n);
+       /* See if there're unreturned mbufs from compressed CQE. */
+       rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
+       if (rcvd_pkt > 0) {
+               rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
+               rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
+               rxq->rq_pi += rcvd_pkt;
+               pkts += rcvd_pkt;
+       }
+       elts_idx = rxq->rq_pi & q_mask;
+       elts = &(*rxq->elts)[elts_idx];
+       /* Not to overflow pkts array. */
+       pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
+       /* Not to cross queue end. */
+       pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
+       if (!pkts_n)
+               return rcvd_pkt;
+       /* At this point, there shouldn't be any remained packets. */
+       assert(rxq->rq_pi == rxq->cq_ci);
+       /*
+        * A. load first Qword (8bytes) in one loop.
+        * B. copy 4 mbuf pointers from elts ring to returing pkts.
+        * C. load remained CQE data and extract necessary fields.
+        *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
+        *    following structure:
+        *        struct {
+        *          uint8_t  pkt_info;
+        *          uint8_t  flow_tag[3];
+        *          uint16_t byte_cnt;
+        *          uint8_t  rsvd4;
+        *          uint8_t  op_own;
+        *          uint16_t hdr_type_etc;
+        *          uint16_t vlan_info;
+        *          uint32_t rx_has_res;
+        *        } c;
+        * D. fill in mbuf.
+        * E. get valid CQEs.
+        * F. find compressed CQE.
+        */
+       for (pos = 0;
+            pos < pkts_n;
+            pos += MLX5_VPMD_DESCS_PER_LOOP) {
+               __m128i cqes[MLX5_VPMD_DESCS_PER_LOOP];
+               __m128i cqe_tmp1, cqe_tmp2;
+               __m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
+               __m128i op_own, op_own_tmp1, op_own_tmp2;
+               __m128i opcode, owner_mask, invalid_mask;
+               __m128i comp_mask;
+               __m128i mask;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               __m128i byte_cnt;
+#endif
+               __m128i mbp1, mbp2;
+               __m128i p = _mm_set_epi16(0, 0, 0, 0, 3, 2, 1, 0);
+               unsigned int p1, p2, p3;
+
+               /* Prefetch next 4 CQEs. */
+               if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
+                       rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
+                       rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
+                       rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
+                       rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
+               }
+               /* A.0 do not cross the end of CQ. */
+               mask = _mm_set_epi64x(0, (pkts_n - pos) * sizeof(uint16_t) * 8);
+               mask = _mm_sll_epi64(ones, mask);
+               p = _mm_andnot_si128(mask, p);
+               /* A.1 load cqes. */
+               p3 = _mm_extract_epi16(p, 3);
+               cqes[3] = _mm_loadl_epi64((__m128i *)
+                                          &cq[pos + p3].sop_drop_qpn);
+               rte_compiler_barrier();
+               p2 = _mm_extract_epi16(p, 2);
+               cqes[2] = _mm_loadl_epi64((__m128i *)
+                                          &cq[pos + p2].sop_drop_qpn);
+               rte_compiler_barrier();
+               /* B.1 load mbuf pointers. */
+               mbp1 = _mm_loadu_si128((__m128i *)&elts[pos]);
+               mbp2 = _mm_loadu_si128((__m128i *)&elts[pos + 2]);
+               /* A.1 load a block having op_own. */
+               p1 = _mm_extract_epi16(p, 1);
+               cqes[1] = _mm_loadl_epi64((__m128i *)
+                                          &cq[pos + p1].sop_drop_qpn);
+               rte_compiler_barrier();
+               cqes[0] = _mm_loadl_epi64((__m128i *)
+                                          &cq[pos].sop_drop_qpn);
+               /* B.2 copy mbuf pointers. */
+               _mm_storeu_si128((__m128i *)&pkts[pos], mbp1);
+               _mm_storeu_si128((__m128i *)&pkts[pos + 2], mbp2);
+               rte_compiler_barrier();
+               /* C.1 load remained CQE data and extract necessary fields. */
+               cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p3]);
+               cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos + p2]);
+               cqes[3] = _mm_blendv_epi8(cqes[3], cqe_tmp2, blend_mask);
+               cqes[2] = _mm_blendv_epi8(cqes[2], cqe_tmp1, blend_mask);
+               cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p3].rsvd1[3]);
+               cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos + p2].rsvd1[3]);
+               cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x30);
+               cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x30);
+               cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p3].rsvd2[10]);
+               cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos + p2].rsvd2[10]);
+               cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x04);
+               cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x04);
+               /* C.2 generate final structure for mbuf with swapping bytes. */
+               pkt_mb3 = _mm_shuffle_epi8(cqes[3], shuf_mask);
+               pkt_mb2 = _mm_shuffle_epi8(cqes[2], shuf_mask);
+               /* C.3 adjust CRC length. */
+               pkt_mb3 = _mm_sub_epi16(pkt_mb3, crc_adj);
+               pkt_mb2 = _mm_sub_epi16(pkt_mb2, crc_adj);
+               /* C.4 adjust flow mark. */
+               pkt_mb3 = _mm_add_epi32(pkt_mb3, flow_mark_adj);
+               pkt_mb2 = _mm_add_epi32(pkt_mb2, flow_mark_adj);
+               /* D.1 fill in mbuf - rx_descriptor_fields1. */
+               _mm_storeu_si128((void *)&pkts[pos + 3]->pkt_len, pkt_mb3);
+               _mm_storeu_si128((void *)&pkts[pos + 2]->pkt_len, pkt_mb2);
+               /* E.1 extract op_own field. */
+               op_own_tmp2 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
+               /* C.1 load remained CQE data and extract necessary fields. */
+               cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p1]);
+               cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos]);
+               cqes[1] = _mm_blendv_epi8(cqes[1], cqe_tmp2, blend_mask);
+               cqes[0] = _mm_blendv_epi8(cqes[0], cqe_tmp1, blend_mask);
+               cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p1].rsvd1[3]);
+               cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos].rsvd1[3]);
+               cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x30);
+               cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x30);
+               cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p1].rsvd2[10]);
+               cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos].rsvd2[10]);
+               cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x04);
+               cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x04);
+               /* C.2 generate final structure for mbuf with swapping bytes. */
+               pkt_mb1 = _mm_shuffle_epi8(cqes[1], shuf_mask);
+               pkt_mb0 = _mm_shuffle_epi8(cqes[0], shuf_mask);
+               /* C.3 adjust CRC length. */
+               pkt_mb1 = _mm_sub_epi16(pkt_mb1, crc_adj);
+               pkt_mb0 = _mm_sub_epi16(pkt_mb0, crc_adj);
+               /* C.4 adjust flow mark. */
+               pkt_mb1 = _mm_add_epi32(pkt_mb1, flow_mark_adj);
+               pkt_mb0 = _mm_add_epi32(pkt_mb0, flow_mark_adj);
+               /* E.1 extract op_own byte. */
+               op_own_tmp1 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
+               op_own = _mm_unpackhi_epi64(op_own_tmp1, op_own_tmp2);
+               /* D.1 fill in mbuf - rx_descriptor_fields1. */
+               _mm_storeu_si128((void *)&pkts[pos + 1]->pkt_len, pkt_mb1);
+               _mm_storeu_si128((void *)&pkts[pos]->pkt_len, pkt_mb0);
+               /* E.2 flip owner bit to mark CQEs from last round. */
+               owner_mask = _mm_and_si128(op_own, owner_check);
+               if (ownership)
+                       owner_mask = _mm_xor_si128(owner_mask, owner_check);
+               owner_mask = _mm_cmpeq_epi32(owner_mask, owner_check);
+               owner_mask = _mm_packs_epi32(owner_mask, zero);
+               /* E.3 get mask for invalidated CQEs. */
+               opcode = _mm_and_si128(op_own, opcode_check);
+               invalid_mask = _mm_cmpeq_epi32(opcode_check, opcode);
+               invalid_mask = _mm_packs_epi32(invalid_mask, zero);
+               /* E.4 mask out beyond boundary. */
+               invalid_mask = _mm_or_si128(invalid_mask, mask);
+               /* E.5 merge invalid_mask with invalid owner. */
+               invalid_mask = _mm_or_si128(invalid_mask, owner_mask);
+               /* F.1 find compressed CQE format. */
+               comp_mask = _mm_and_si128(op_own, format_check);
+               comp_mask = _mm_cmpeq_epi32(comp_mask, format_check);
+               comp_mask = _mm_packs_epi32(comp_mask, zero);
+               /* F.2 mask out invalid entries. */
+               comp_mask = _mm_andnot_si128(invalid_mask, comp_mask);
+               comp_idx = _mm_cvtsi128_si64(comp_mask);
+               /* F.3 get the first compressed CQE. */
+               comp_idx = comp_idx ?
+                               __builtin_ctzll(comp_idx) /
+                                       (sizeof(uint16_t) * 8) :
+                               MLX5_VPMD_DESCS_PER_LOOP;
+               /* E.6 mask out entries after the compressed CQE. */
+               mask = _mm_set_epi64x(0, comp_idx * sizeof(uint16_t) * 8);
+               mask = _mm_sll_epi64(ones, mask);
+               invalid_mask = _mm_or_si128(invalid_mask, mask);
+               /* E.7 count non-compressed valid CQEs. */
+               n = _mm_cvtsi128_si64(invalid_mask);
+               n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
+                       MLX5_VPMD_DESCS_PER_LOOP;
+               nocmp_n += n;
+               /* D.2 get the final invalid mask. */
+               mask = _mm_set_epi64x(0, n * sizeof(uint16_t) * 8);
+               mask = _mm_sll_epi64(ones, mask);
+               invalid_mask = _mm_or_si128(invalid_mask, mask);
+               /* D.3 check error in opcode. */
+               opcode = _mm_cmpeq_epi32(resp_err_check, opcode);
+               opcode = _mm_packs_epi32(opcode, zero);
+               opcode = _mm_andnot_si128(invalid_mask, opcode);
+               /* D.4 mark if any error is set */
+               rxq->pending_err |= !!_mm_cvtsi128_si64(opcode);
+               /* D.5 fill in mbuf - rearm_data and packet_type. */
+               rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+               /* Add up received bytes count. */
+               byte_cnt = _mm_shuffle_epi8(op_own, len_shuf_mask);
+               byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
+               byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
+               rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
+#endif
+               /*
+                * Break the loop unless more valid CQE is expected, or if
+                * there's a compressed CQE.
+                */
+               if (n != MLX5_VPMD_DESCS_PER_LOOP)
+                       break;
+       }
+       /* If no new CQE seen, return without updating cq_db. */
+       if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
+               return rcvd_pkt;
+       /* Update the consumer indexes for non-compressed CQEs. */
+       assert(nocmp_n <= pkts_n);
+       rxq->cq_ci += nocmp_n;
+       rxq->rq_pi += nocmp_n;
+       rcvd_pkt += nocmp_n;
+#ifdef MLX5_PMD_SOFT_COUNTERS
+       rxq->stats.ipackets += nocmp_n;
+       rxq->stats.ibytes += rcvd_byte;
+#endif
+       /* Decompress the last CQE if compressed. */
+       if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
+               assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
+               rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
+               /* Return more packets if needed. */
+               if (nocmp_n < pkts_n) {
+                       uint16_t n = rxq->cq_ci - rxq->rq_pi;
+
+                       n = RTE_MIN(n, pkts_n - nocmp_n);
+                       rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
+                       rxq->rq_pi += n;
+                       rcvd_pkt += n;
+               }
+       }
+       rte_wmb();
+       *rxq->cq_db = htonl(rxq->cq_ci);
+       return rcvd_pkt;
+}
+
+/**
+ * DPDK callback for vectorized RX.
+ *
+ * @param dpdk_rxq
+ *   Generic pointer to RX queue structure.
+ * @param[out] pkts
+ *   Array to store received packets.
+ * @param pkts_n
+ *   Maximum number of packets in array.
+ *
+ * @return
+ *   Number of packets successfully received (<= pkts_n).
+ */
+uint16_t
+mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
+{
+       struct rxq *rxq = dpdk_rxq;
+       uint16_t nb_rx;
+
+       nb_rx = rxq_burst_v(rxq, pkts, pkts_n);
+       if (unlikely(rxq->pending_err))
+               nb_rx = rxq_handle_pending_error(rxq, pkts, nb_rx);
+       return nb_rx;
+}
+
+/**
+ * Check Tx queue flags are set for raw vectorized Tx.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_raw_vec_tx_support(struct priv *priv)
+{
+       uint16_t i;
+
+       /* All the configured queues should support. */
+       for (i = 0; i < priv->txqs_n; ++i) {
+               struct txq *txq = (*priv->txqs)[i];
+
+               if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) ||
+                   !(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
+                       break;
+       }
+       if (i != priv->txqs_n)
+               return -ENOTSUP;
+       return 1;
+}
+
+/**
+ * Check a device can support vectorized TX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_vec_tx_support(struct priv *priv)
+{
+       if (priv->txqs_n > MLX5_VPMD_MIN_TXQS ||
+           priv->mps != MLX5_MPW_ENHANCED ||
+           priv->tso)
+               return -ENOTSUP;
+       return 1;
+}
+
+/**
+ * Check a RX queue can support vectorized RX.
+ *
+ * @param rxq
+ *   Pointer to RX queue.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+rxq_check_vec_support(struct rxq *rxq)
+{
+       if (rxq->sges_n != 0)
+               return -ENOTSUP;
+       return 1;
+}
+
+/**
+ * Check a device can support vectorized RX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   1 if supported, negative errno value if not.
+ */
+int __attribute__((cold))
+priv_check_vec_rx_support(struct priv *priv)
+{
+       uint16_t i;
+
+       /* All the configured queues should support. */
+       for (i = 0; i < priv->rxqs_n; ++i) {
+               struct rxq *rxq = (*priv->rxqs)[i];
+
+               if (rxq_check_vec_support(rxq) < 0)
+                       break;
+       }
+       if (i != priv->rxqs_n)
+               return -ENOTSUP;
+       return 1;
+}
+
+/**
+ * Prepare for vectorized RX.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ */
+void
+priv_prep_vec_rx_function(struct priv *priv)
+{
+       uint16_t i;
+
+       for (i = 0; i < priv->rxqs_n; ++i) {
+               struct rxq *rxq = (*priv->rxqs)[i];
+               struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
+               const uint16_t desc = 1 << rxq->elts_n;
+               int j;
+
+               assert(rxq->elts_n == rxq->cqe_n);
+               /* Initialize default rearm_data for vPMD. */
+               mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
+               rte_mbuf_refcnt_set(mbuf_init, 1);
+               mbuf_init->nb_segs = 1;
+               mbuf_init->port = rxq->port_id;
+               /*
+                * prevent compiler reordering:
+                * rearm_data covers previous fields.
+                */
+               rte_compiler_barrier();
+               rxq->mbuf_initializer =
+                       *(uint64_t *)&mbuf_init->rearm_data;
+               /* Padding with a fake mbuf for vectorized Rx. */
+               for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
+                       (*rxq->elts)[desc + j] = &rxq->fake_mbuf;
+               /* Mark that it need to be cleaned up for rxq_alloc_elts(). */
+               rxq->trim_elts = 1;
+       }
+}
index 03e23f5..3ad6ac0 100644 (file)
@@ -243,7 +243,7 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
                ERROR("MLX5_ENABLE_CQE_COMPRESSION must never be set");
                goto error;
        }
-       (void)conf; /* Thresholds configuration (ignored). */
+       tmpl.txq.flags = conf->txq_flags;
        assert(desc > MLX5_TX_COMP_THRESH);
        tmpl.txq.elts_n = log2above(desc);
        if (priv->mps == MLX5_MPW_ENHANCED)