X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fmlx5%2Fmlx5_rxtx.c;h=887e283c180301a3ed99cfd0b195ed1fd05c4df9;hb=9bf26e1318e3cd7a04115c8201255505e90cab83;hp=13f943177cfc2a07e32f4956ad11d61b3976f9b2;hpb=38b4b397a57dfb0f892c8289ce1e724f6a37ba0b;p=dpdk.git diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index 13f943177c..887e283c18 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: BSD-3-Clause * Copyright 2015 6WIND S.A. - * Copyright 2015 Mellanox Technologies, Ltd + * Copyright 2015-2019 Mellanox Technologies, Ltd */ #include @@ -34,6 +34,57 @@ #include "mlx5_defs.h" #include "mlx5_prm.h" +/* TX burst subroutines return codes. */ +enum mlx5_txcmp_code { + MLX5_TXCMP_CODE_EXIT = 0, + MLX5_TXCMP_CODE_ERROR, + MLX5_TXCMP_CODE_SINGLE, + MLX5_TXCMP_CODE_MULTI, + MLX5_TXCMP_CODE_TSO, + MLX5_TXCMP_CODE_EMPW, +}; + +/* + * These defines are used to configure Tx burst routine option set + * supported at compile time. The not specified options are optimized out + * out due to if conditions can be explicitly calculated at compile time. + * The offloads with bigger runtime check (require more CPU cycles to + * skip) overhead should have the bigger index - this is needed to + * select the better matching routine function if no exact match and + * some offloads are not actually requested. + */ +#define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/ +#define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/ +#define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/ +#define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */ +#define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */ +#define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/ +#define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */ +#define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/ + +/* The most common offloads groups. */ +#define MLX5_TXOFF_CONFIG_NONE 0 +#define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \ + MLX5_TXOFF_CONFIG_TSO | \ + MLX5_TXOFF_CONFIG_SWP | \ + MLX5_TXOFF_CONFIG_CSUM | \ + MLX5_TXOFF_CONFIG_INLINE | \ + MLX5_TXOFF_CONFIG_VLAN | \ + MLX5_TXOFF_CONFIG_METADATA) + +#define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask) + +#define MLX5_TXOFF_DECL(func, olx) \ +static uint16_t mlx5_tx_burst_##func(void *txq, \ + struct rte_mbuf **pkts, \ + uint16_t pkts_n) \ +{ \ + return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \ + pkts, pkts_n, (olx)); \ +} + +#define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx}, + static __rte_always_inline uint32_t rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe); @@ -49,12 +100,23 @@ rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt, volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res); static __rte_always_inline void -mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx); +mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx, + const unsigned int strd_n); static int mlx5_queue_state_modify(struct rte_eth_dev *dev, struct mlx5_mp_arg_queue_state_modify *sm); +static inline void +mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp, + volatile struct mlx5_cqe *restrict cqe, + uint32_t phcsum); + +static inline void +mlx5_lro_update_hdr(uint8_t *restrict padd, + volatile struct mlx5_cqe *restrict cqe, + uint32_t len); + uint32_t mlx5_ptype_table[] __rte_cache_aligned = { [0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */ }; @@ -287,6 +349,109 @@ mlx5_set_swp_types_table(void) } } +/** + * Set Software Parser flags and offsets in Ethernet Segment of WQE. + * Flags must be preliminary initialized to zero. + * + * @param loc + * Pointer to burst routine local context. + * @param swp_flags + * Pointer to store Software Parser flags + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Software Parser offsets packed in dword. + * Software Parser flags are set by pointer. + */ +static __rte_always_inline uint32_t +txq_mbuf_to_swp(struct mlx5_txq_local *restrict loc, + uint8_t *swp_flags, + unsigned int olx) +{ + uint64_t ol, tunnel; + unsigned int idx, off; + uint32_t set; + + if (!MLX5_TXOFF_CONFIG(SWP)) + return 0; + ol = loc->mbuf->ol_flags; + tunnel = ol & PKT_TX_TUNNEL_MASK; + /* + * Check whether Software Parser is required. + * Only customized tunnels may ask for. + */ + if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP)) + return 0; + /* + * The index should have: + * bit[0:1] = PKT_TX_L4_MASK + * bit[4] = PKT_TX_IPV6 + * bit[8] = PKT_TX_OUTER_IPV6 + * bit[9] = PKT_TX_OUTER_UDP + */ + idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52; + idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0; + *swp_flags = mlx5_swp_types_table[idx]; + /* + * Set offsets for SW parser. Since ConnectX-5, SW parser just + * complements HW parser. SW parser starts to engage only if HW parser + * can't reach a header. For the older devices, HW parser will not kick + * in if any of SWP offsets is set. Therefore, all of the L3 offsets + * should be set regardless of HW offload. + */ + off = loc->mbuf->outer_l2_len; + if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT) + off += sizeof(struct rte_vlan_hdr); + set = (off >> 1) << 8; /* Outer L3 offset. */ + off += loc->mbuf->outer_l3_len; + if (tunnel == PKT_TX_TUNNEL_UDP) + set |= off >> 1; /* Outer L4 offset. */ + if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */ + const uint64_t csum = ol & PKT_TX_L4_MASK; + off += loc->mbuf->l2_len; + set |= (off >> 1) << 24; /* Inner L3 offset. */ + if (csum == PKT_TX_TCP_CKSUM || + csum == PKT_TX_UDP_CKSUM || + (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) { + off += loc->mbuf->l3_len; + set |= (off >> 1) << 16; /* Inner L4 offset. */ + } + } + set = rte_cpu_to_le_32(set); + return set; +} + +/** + * Convert the Checksum offloads to Verbs. + * + * @param buf + * Pointer to the mbuf. + * + * @return + * Converted checksum flags. + */ +static __rte_always_inline uint8_t +txq_ol_cksum_to_cs(struct rte_mbuf *buf) +{ + uint32_t idx; + uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK); + const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK | + PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM; + + /* + * The index should have: + * bit[0] = PKT_TX_TCP_SEG + * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM + * bit[4] = PKT_TX_IP_CKSUM + * bit[8] = PKT_TX_OUTER_IP_CKSUM + * bit[9] = tunnel + */ + idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9); + return mlx5_cksum_table[idx]; +} + /** * Internal function to compute the number of used descriptors in an RX queue * @@ -418,18 +583,16 @@ mlx5_dump_debug_information(const char *fname, const char *hex_title, MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname); fd = fopen(path, "a+"); if (!fd) { - DRV_LOG(WARNING, "cannot open %s for debug dump\n", - path); + DRV_LOG(WARNING, "cannot open %s for debug dump", path); MKSTR(path2, "./%s", fname); fd = fopen(path2, "a+"); if (!fd) { - DRV_LOG(ERR, "cannot open %s for debug dump\n", - path2); + DRV_LOG(ERR, "cannot open %s for debug dump", path2); return; } - DRV_LOG(INFO, "New debug dump in file %s\n", path2); + DRV_LOG(INFO, "New debug dump in file %s", path2); } else { - DRV_LOG(INFO, "New debug dump in file %s\n", path); + DRV_LOG(INFO, "New debug dump in file %s", path); } if (hex_title) rte_hexdump(fd, hex_title, buf, hex_len); @@ -489,10 +652,11 @@ check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe) * Pointer to the error CQE. * * @return - * The last Tx buffer element to free. + * Negative value if queue recovery failed, + * the last Tx buffer element to free otherwise. */ -uint16_t -mlx5_tx_error_cqe_handle(struct mlx5_txq_data *txq, +int +mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq, volatile struct mlx5_err_cqe *err_cqe) { if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) { @@ -541,6 +705,7 @@ mlx5_tx_error_cqe_handle(struct mlx5_txq_data *txq, return txq->elts_head; } /* Recovering failed - try again later on the same WQE. */ + return -1; } else { txq->cq_ci++; } @@ -602,7 +767,8 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq) scat = &((volatile struct mlx5_wqe_mprq *) rxq->wqes)[i].dseg; - addr = (uintptr_t)mlx5_mprq_buf_addr(buf); + addr = (uintptr_t)mlx5_mprq_buf_addr(buf, + 1 << rxq->strd_num_n); byte_count = (1 << rxq->strd_sz_n) * (1 << rxq->strd_num_n); } else { @@ -634,7 +800,7 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq) } /** - * Modify a Verbs queue state. + * Modify a Verbs/DevX queue state. * This must be called from the primary process. * * @param dev @@ -653,17 +819,36 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, struct mlx5_priv *priv = dev->data->dev_private; if (sm->is_wq) { - struct ibv_wq_attr mod = { - .attr_mask = IBV_WQ_ATTR_STATE, - .wq_state = sm->state, - }; struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id]; struct mlx5_rxq_ctrl *rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq); - ret = mlx5_glue->modify_wq(rxq_ctrl->ibv->wq, &mod); + if (rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV) { + struct ibv_wq_attr mod = { + .attr_mask = IBV_WQ_ATTR_STATE, + .wq_state = sm->state, + }; + + ret = mlx5_glue->modify_wq(rxq_ctrl->obj->wq, &mod); + } else { /* rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ. */ + struct mlx5_devx_modify_rq_attr rq_attr; + + memset(&rq_attr, 0, sizeof(rq_attr)); + if (sm->state == IBV_WQS_RESET) { + rq_attr.rq_state = MLX5_RQC_STATE_ERR; + rq_attr.state = MLX5_RQC_STATE_RST; + } else if (sm->state == IBV_WQS_RDY) { + rq_attr.rq_state = MLX5_RQC_STATE_RST; + rq_attr.state = MLX5_RQC_STATE_RDY; + } else if (sm->state == IBV_WQS_ERR) { + rq_attr.rq_state = MLX5_RQC_STATE_RDY; + rq_attr.state = MLX5_RQC_STATE_ERR; + } + ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq, + &rq_attr); + } if (ret) { - DRV_LOG(ERR, "Cannot change Rx WQ state to %u - %s\n", + DRV_LOG(ERR, "Cannot change Rx WQ state to %u - %s", sm->state, strerror(errno)); rte_errno = errno; return ret; @@ -676,12 +861,12 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, .qp_state = IBV_QPS_RESET, .port_num = (uint8_t)priv->ibv_port, }; - struct ibv_qp *qp = txq_ctrl->ibv->qp; + struct ibv_qp *qp = txq_ctrl->obj->qp; ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE); if (ret) { DRV_LOG(ERR, "Cannot change the Tx QP state to RESET " - "%s\n", strerror(errno)); + "%s", strerror(errno)); rte_errno = errno; return ret; } @@ -689,7 +874,7 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, ret = mlx5_glue->modify_qp(qp, &mod, (IBV_QP_STATE | IBV_QP_PORT)); if (ret) { - DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s\n", + DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s", strerror(errno)); rte_errno = errno; return ret; @@ -697,7 +882,7 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, mod.qp_state = IBV_QPS_RTR; ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE); if (ret) { - DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s\n", + DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s", strerror(errno)); rte_errno = errno; return ret; @@ -705,7 +890,7 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, mod.qp_state = IBV_QPS_RTS; ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE); if (ret) { - DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s\n", + DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s", strerror(errno)); rte_errno = errno; return ret; @@ -753,14 +938,15 @@ mlx5_queue_state_modify(struct rte_eth_dev *dev, * * @param[in] rxq * Pointer to RX queue structure. - * @param[in] mbuf_prepare - * Whether to prepare mbufs for the RQ. + * @param[in] vec + * 1 when called from vectorized Rx burst, need to prepare mbufs for the RQ. + * 0 when called from non-vectorized Rx burst. * * @return * -1 in case of recovery error, otherwise the CQE status. */ int -mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t mbuf_prepare) +mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec) { const uint16_t cqe_n = 1 << rxq->cqe_n; const uint16_t cqe_mask = cqe_n - 1; @@ -827,7 +1013,7 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t mbuf_prepare) if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm)) return -1; - if (mbuf_prepare) { + if (vec) { const uint16_t q_mask = wqe_n - 1; uint16_t elt_idx; struct rte_mbuf **elt; @@ -851,6 +1037,16 @@ mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t mbuf_prepare) return -1; } } + for (i = 0; i < (int)wqe_n; ++i) { + elt = &(*rxq->elts)[i]; + DATA_LEN(*elt) = + (uint16_t)((*elt)->buf_len - + rte_pktmbuf_headroom(*elt)); + } + /* Padding with a fake mbuf for vec Rx. */ + for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i) + (*rxq->elts)[wqe_n + i] = + &rxq->fake_mbuf; } mlx5_rxq_initialize(rxq); rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR; @@ -1148,6 +1344,13 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) if (rxq->crc_present) len -= RTE_ETHER_CRC_LEN; PKT_LEN(pkt) = len; + if (cqe->lro_num_seg > 1) { + mlx5_lro_update_hdr + (rte_pktmbuf_mtod(pkt, uint8_t *), cqe, + len); + pkt->ol_flags |= PKT_RX_LRO; + pkt->tso_segsz = len / cqe->lro_num_seg; + } } DATA_LEN(rep) = DATA_LEN(seg); PKT_LEN(rep) = PKT_LEN(seg); @@ -1199,6 +1402,101 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) return i; } +/** + * Update LRO packet TCP header. + * The HW LRO feature doesn't update the TCP header after coalescing the + * TCP segments but supplies information in CQE to fill it by SW. + * + * @param tcp + * Pointer to the TCP header. + * @param cqe + * Pointer to the completion entry.. + * @param phcsum + * The L3 pseudo-header checksum. + */ +static inline void +mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp, + volatile struct mlx5_cqe *restrict cqe, + uint32_t phcsum) +{ + uint8_t l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) & + MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT; + /* + * The HW calculates only the TCP payload checksum, need to complete + * the TCP header checksum and the L3 pseudo-header checksum. + */ + uint32_t csum = phcsum + cqe->csum; + + if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK || + l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) { + tcp->tcp_flags |= RTE_TCP_ACK_FLAG; + tcp->recv_ack = cqe->lro_ack_seq_num; + tcp->rx_win = cqe->lro_tcp_win; + } + if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK) + tcp->tcp_flags |= RTE_TCP_PSH_FLAG; + tcp->cksum = 0; + csum += rte_raw_cksum(tcp, (tcp->data_off & 0xF) * 4); + csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff); + csum = (~csum) & 0xffff; + if (csum == 0) + csum = 0xffff; + tcp->cksum = csum; +} + +/** + * Update LRO packet headers. + * The HW LRO feature doesn't update the L3/TCP headers after coalescing the + * TCP segments but supply information in CQE to fill it by SW. + * + * @param padd + * The packet address. + * @param cqe + * Pointer to the completion entry.. + * @param len + * The packet length. + */ +static inline void +mlx5_lro_update_hdr(uint8_t *restrict padd, + volatile struct mlx5_cqe *restrict cqe, + uint32_t len) +{ + union { + struct rte_ether_hdr *eth; + struct rte_vlan_hdr *vlan; + struct rte_ipv4_hdr *ipv4; + struct rte_ipv6_hdr *ipv6; + struct rte_tcp_hdr *tcp; + uint8_t *hdr; + } h = { + .hdr = padd, + }; + uint16_t proto = h.eth->ether_type; + uint32_t phcsum; + + h.eth++; + while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) || + proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) { + proto = h.vlan->eth_proto; + h.vlan++; + } + if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) { + h.ipv4->time_to_live = cqe->lro_min_ttl; + h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd)); + h.ipv4->hdr_checksum = 0; + h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4); + phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0); + h.ipv4++; + } else { + h.ipv6->hop_limits = cqe->lro_min_ttl; + h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) - + sizeof(*h.ipv6)); + phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0); + h.ipv6++; + } + mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum); +} + void mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque) { @@ -1219,7 +1517,8 @@ mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf) } static inline void -mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx) +mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx, + const unsigned int strd_n) { struct mlx5_mprq_buf *rep = rxq->mprq_repl; volatile struct mlx5_wqe_data_seg *wqe = @@ -1230,7 +1529,7 @@ mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx) /* Replace MPRQ buf. */ (*rxq->mprq_bufs)[rq_idx] = rep; /* Replace WQE. */ - addr = mlx5_mprq_buf_addr(rep); + addr = mlx5_mprq_buf_addr(rep, strd_n); wqe->addr = rte_cpu_to_be_64((uintptr_t)addr); /* If there's only one MR, no need to replace LKey in WQE. */ if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1)) @@ -1269,6 +1568,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) unsigned int i = 0; uint32_t rq_ci = rxq->rq_ci; uint16_t consumed_strd = rxq->consumed_strd; + uint16_t headroom_sz = rxq->strd_headroom_en * RTE_PKTMBUF_HEADROOM; struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask]; while (i < pkts_n) { @@ -1282,11 +1582,12 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) uint32_t byte_cnt; volatile struct mlx5_mini_cqe8 *mcqe = NULL; uint32_t rss_hash_res = 0; + uint8_t lro_num_seg; if (consumed_strd == strd_n) { /* Replace WQE only if the buffer is still in use. */ if (rte_atomic16_read(&buf->refcnt) > 1) { - mprq_buf_replace(rxq, rq_ci & wq_mask); + mprq_buf_replace(rxq, rq_ci & wq_mask, strd_n); /* Release the old buffer. */ mlx5_mprq_buf_free(buf); } else if (unlikely(rxq->mprq_repl == NULL)) { @@ -1327,6 +1628,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) } assert(strd_idx < strd_n); assert(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) & wq_mask)); + lro_num_seg = cqe->lro_num_seg; /* * Currently configured to receive a packet per a stride. But if * MTU is adjusted through kernel interface, device could @@ -1334,7 +1636,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) * case, the packet should be dropped because it is bigger than * the max_rx_pkt_len. */ - if (unlikely(strd_cnt > 1)) { + if (unlikely(!lro_num_seg && strd_cnt > 1)) { ++rxq->stats.idropped; continue; } @@ -1348,9 +1650,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) if (rxq->crc_present) len -= RTE_ETHER_CRC_LEN; offset = strd_idx * strd_sz + strd_shift; - addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf), offset); - /* Initialize the offload flag. */ - pkt->ol_flags = 0; + addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset); /* * Memcpy packets to the target mbuf if: * - The size of packet is smaller than mprq_max_memcpy_len. @@ -1367,34 +1667,37 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) continue; } rte_memcpy(rte_pktmbuf_mtod(pkt, void *), addr, len); + DATA_LEN(pkt) = len; } else { rte_iova_t buf_iova; struct rte_mbuf_ext_shared_info *shinfo; uint16_t buf_len = strd_cnt * strd_sz; + void *buf_addr; /* Increment the refcnt of the whole chunk. */ rte_atomic16_add_return(&buf->refcnt, 1); assert((uint16_t)rte_atomic16_read(&buf->refcnt) <= strd_n + 1); - addr = RTE_PTR_SUB(addr, RTE_PKTMBUF_HEADROOM); + buf_addr = RTE_PTR_SUB(addr, headroom_sz); /* * MLX5 device doesn't use iova but it is necessary in a * case where the Rx packet is transmitted via a * different PMD. */ buf_iova = rte_mempool_virt2iova(buf) + - RTE_PTR_DIFF(addr, buf); - shinfo = rte_pktmbuf_ext_shinfo_init_helper(addr, - &buf_len, mlx5_mprq_buf_free_cb, buf); + RTE_PTR_DIFF(buf_addr, buf); + shinfo = &buf->shinfos[strd_idx]; + rte_mbuf_ext_refcnt_set(shinfo, 1); /* * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when * attaching the stride to mbuf and more offload flags * will be added below by calling rxq_cq_to_mbuf(). * Other fields will be overwritten. */ - rte_pktmbuf_attach_extbuf(pkt, addr, buf_iova, buf_len, - shinfo); - rte_pktmbuf_reset_headroom(pkt); + rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova, + buf_len, shinfo); + /* Set mbuf head-room. */ + pkt->data_off = headroom_sz; assert(pkt->ol_flags == EXT_ATTACHED_MBUF); /* * Prevent potential overflow due to MTU change through @@ -1405,10 +1708,34 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) ++rxq->stats.idropped; continue; } + DATA_LEN(pkt) = len; + /* + * LRO packet may consume all the stride memory, in this + * case packet head-room space is not guaranteed so must + * to add an empty mbuf for the head-room. + */ + if (!rxq->strd_headroom_en) { + struct rte_mbuf *headroom_mbuf = + rte_pktmbuf_alloc(rxq->mp); + + if (unlikely(headroom_mbuf == NULL)) { + rte_pktmbuf_free_seg(pkt); + ++rxq->stats.rx_nombuf; + break; + } + PORT(pkt) = rxq->port_id; + NEXT(headroom_mbuf) = pkt; + pkt = headroom_mbuf; + NB_SEGS(pkt) = 2; + } } rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res); + if (lro_num_seg > 1) { + mlx5_lro_update_hdr(addr, cqe, len); + pkt->ol_flags |= PKT_RX_LRO; + pkt->tso_segsz = strd_sz; + } PKT_LEN(pkt) = len; - DATA_LEN(pkt) = len; PORT(pkt) = rxq->port_id; #ifdef MLX5_PMD_SOFT_COUNTERS /* Increment bytes counter. */ @@ -1511,6 +1838,349 @@ mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused) return -ENOTSUP; } +/** + * Free the mbufs from the linear array of pointers. + * + * @param pkts + * Pointer to array of packets to be free. + * @param pkts_n + * Number of packets to be freed. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + unsigned int olx __rte_unused) +{ + struct rte_mempool *pool = NULL; + struct rte_mbuf **p_free = NULL; + struct rte_mbuf *mbuf; + unsigned int n_free = 0; + + /* + * The implemented algorithm eliminates + * copying pointers to temporary array + * for rte_mempool_put_bulk() calls. + */ + assert(pkts); + assert(pkts_n); + for (;;) { + for (;;) { + /* + * Decrement mbuf reference counter, detach + * indirect and external buffers if needed. + */ + mbuf = rte_pktmbuf_prefree_seg(*pkts); + if (likely(mbuf != NULL)) { + assert(mbuf == *pkts); + if (likely(n_free != 0)) { + if (unlikely(pool != mbuf->pool)) + /* From different pool. */ + break; + } else { + /* Start new scan array. */ + pool = mbuf->pool; + p_free = pkts; + } + ++n_free; + ++pkts; + --pkts_n; + if (unlikely(pkts_n == 0)) { + mbuf = NULL; + break; + } + } else { + /* + * This happens if mbuf is still referenced. + * We can't put it back to the pool, skip. + */ + ++pkts; + --pkts_n; + if (unlikely(n_free != 0)) + /* There is some array to free.*/ + break; + if (unlikely(pkts_n == 0)) + /* Last mbuf, nothing to free. */ + return; + } + } + for (;;) { + /* + * This loop is implemented to avoid multiple + * inlining of rte_mempool_put_bulk(). + */ + assert(pool); + assert(p_free); + assert(n_free); + /* + * Free the array of pre-freed mbufs + * belonging to the same memory pool. + */ + rte_mempool_put_bulk(pool, (void *)p_free, n_free); + if (unlikely(mbuf != NULL)) { + /* There is the request to start new scan. */ + pool = mbuf->pool; + p_free = pkts++; + n_free = 1; + --pkts_n; + if (likely(pkts_n != 0)) + break; + /* + * This is the last mbuf to be freed. + * Do one more loop iteration to complete. + * This is rare case of the last unique mbuf. + */ + mbuf = NULL; + continue; + } + if (likely(pkts_n == 0)) + return; + n_free = 0; + break; + } + } +} + +/** + * Free the mbuf from the elts ring buffer till new tail. + * + * @param txq + * Pointer to Tx queue structure. + * @param tail + * Index in elts to free up to, becomes new elts tail. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq, + uint16_t tail, + unsigned int olx __rte_unused) +{ + uint16_t n_elts = tail - txq->elts_tail; + + assert(n_elts); + assert(n_elts <= txq->elts_s); + /* + * Implement a loop to support ring buffer wraparound + * with single inlining of mlx5_tx_free_mbuf(). + */ + do { + unsigned int part; + + part = txq->elts_s - (txq->elts_tail & txq->elts_m); + part = RTE_MIN(part, n_elts); + assert(part); + assert(part <= txq->elts_s); + mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m], + part, olx); + txq->elts_tail += part; + n_elts -= part; + } while (n_elts); +} + +/** + * Store the mbuf being sent into elts ring buffer. + * On Tx completion these mbufs will be freed. + * + * @param txq + * Pointer to Tx queue structure. + * @param pkts + * Pointer to array of packets to be stored. + * @param pkts_n + * Number of packets to be stored. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + unsigned int olx __rte_unused) +{ + unsigned int part; + struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts; + + assert(pkts); + assert(pkts_n); + part = txq->elts_s - (txq->elts_head & txq->elts_m); + assert(part); + assert(part <= txq->elts_s); + /* This code is a good candidate for vectorizing with SIMD. */ + rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)), + (void *)pkts, + RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *)); + txq->elts_head += pkts_n; + if (unlikely(part < pkts_n)) + /* The copy is wrapping around the elts array. */ + rte_memcpy((void *)elts, (void *)(pkts + part), + (pkts_n - part) * sizeof(struct rte_mbuf *)); +} + +/** + * Update completion queue consuming index via doorbell + * and flush the completed data buffers. + * + * @param txq + * Pointer to TX queue structure. + * @param valid CQE pointer + * if not NULL update txq->wqe_pi and flush the buffers + * @param itail + * if not negative - flush the buffers till this index. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_comp_flush(struct mlx5_txq_data *restrict txq, + volatile struct mlx5_cqe *last_cqe, + int itail, + unsigned int olx __rte_unused) +{ + uint16_t tail; + + if (likely(last_cqe != NULL)) { + txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter); + tail = ((volatile struct mlx5_wqe_cseg *) + (txq->wqes + (txq->wqe_pi & txq->wqe_m)))->misc; + } else if (itail >= 0) { + tail = (uint16_t)itail; + } else { + return; + } + rte_compiler_barrier(); + *txq->cq_db = rte_cpu_to_be_32(txq->cq_ci); + if (likely(tail != txq->elts_tail)) { + mlx5_tx_free_elts(txq, tail, olx); + assert(tail == txq->elts_tail); + } +} + +/** + * Manage TX completions. This routine checks the CQ for + * arrived CQEs, deduces the last accomplished WQE in SQ, + * updates SQ producing index and frees all completed mbufs. + * + * @param txq + * Pointer to TX queue structure. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * NOTE: not inlined intentionally, it makes tx_burst + * routine smaller, simple and faster - from experiments. + */ +static void +mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq, + unsigned int olx __rte_unused) +{ + unsigned int count = MLX5_TX_COMP_MAX_CQE; + volatile struct mlx5_cqe *last_cqe = NULL; + int ret; + + static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value"); + static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value"); + do { + volatile struct mlx5_cqe *cqe; + + cqe = &txq->cqes[txq->cq_ci & txq->cqe_m]; + ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci); + if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) { + if (likely(ret != MLX5_CQE_STATUS_ERR)) { + /* No new CQEs in completion queue. */ + assert(ret == MLX5_CQE_STATUS_HW_OWN); + break; + } + /* + * Some error occurred, try to restart. + * We have no barrier after WQE related Doorbell + * written, make sure all writes are completed + * here, before we might perform SQ reset. + */ + rte_wmb(); + ret = mlx5_tx_error_cqe_handle + (txq, (volatile struct mlx5_err_cqe *)cqe); + /* + * Flush buffers, update consuming index + * if recovery succeeded. Otherwise + * just try to recover later. + */ + last_cqe = NULL; + break; + } + /* Normal transmit completion. */ + ++txq->cq_ci; + last_cqe = cqe; +#ifndef NDEBUG + if (txq->cq_pi) + --txq->cq_pi; +#endif + /* + * We have to restrict the amount of processed CQEs + * in one tx_burst routine call. The CQ may be large + * and many CQEs may be updated by the NIC in one + * transaction. Buffers freeing is time consuming, + * multiple iterations may introduce significant + * latency. + */ + } while (--count); + mlx5_tx_comp_flush(txq, last_cqe, ret, olx); +} + +/** + * Check if the completion request flag should be set in the last WQE. + * Both pushed mbufs and WQEs are monitored and the completion request + * flag is set if any of thresholds is reached. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param multi, + * Routine is called from multi-segment sending loop, + * do not correct the elts_head according to the pkts_copy. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + bool multi, + unsigned int olx) +{ + uint16_t head = txq->elts_head; + unsigned int part; + + part = (MLX5_TXOFF_CONFIG(INLINE) || multi) ? + 0 : loc->pkts_sent - loc->pkts_copy; + head += part; + if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH || + (MLX5_TXOFF_CONFIG(INLINE) && + (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) { + volatile struct mlx5_wqe *last = loc->wqe_last; + + txq->elts_comp = head; + if (MLX5_TXOFF_CONFIG(INLINE)) + txq->wqe_comp = txq->wqe_ci; + /* Request unconditional completion on last WQE. */ + last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS << + MLX5_COMP_MODE_OFFSET); + /* Save elts_head in unused "immediate" field of WQE. */ + last->cseg.misc = head; + /* + * A CQE slot must always be available. Count the + * issued CEQ "always" request instead of production + * index due to here can be CQE with errors and + * difference with ci may become inconsistent. + */ + assert(txq->cqe_s > ++txq->cq_pi); + } +} + /** * DPDK callback to check the status of a tx descriptor. * @@ -1525,25 +2195,3031 @@ mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused) int mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset) { - (void)tx_queue; - (void)offset; - return RTE_ETH_TX_DESC_FULL; + struct mlx5_txq_data *restrict txq = tx_queue; + uint16_t used; + + mlx5_tx_handle_completion(txq, 0); + used = txq->elts_head - txq->elts_tail; + if (offset < used) + return RTE_ETH_TX_DESC_FULL; + return RTE_ETH_TX_DESC_DONE; } /** - * Configure the TX function to use. + * Build the Control Segment with specified opcode: + * - MLX5_OPCODE_SEND + * - MLX5_OPCODE_ENHANCED_MPSW + * - MLX5_OPCODE_TSO * - * @param dev - * Pointer to private data structure. + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Control Segment. + * @param ds + * Supposed length of WQE in segments. + * @param opcode + * SQ WQE opcode to put into Control Segment. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc __rte_unused, + struct mlx5_wqe *restrict wqe, + unsigned int ds, + unsigned int opcode, + unsigned int olx __rte_unused) +{ + struct mlx5_wqe_cseg *restrict cs = &wqe->cseg; + + cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode); + cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds); + cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR << + MLX5_COMP_MODE_OFFSET); + cs->misc = RTE_BE32(0); +} + +/** + * Build the Ethernet Segment without inlined data. + * Supports Software Parser, Checksums and VLAN + * insertion Tx offload features. * - * @return - * Pointer to selected Tx burst function. + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Ethernet Segment. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. */ -eth_tx_burst_t -mlx5_select_tx_function(struct rte_eth_dev *dev) +static __rte_always_inline void +mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe *restrict wqe, + unsigned int olx) { - (void)dev; - return removed_tx_burst; + struct mlx5_wqe_eseg *restrict es = &wqe->eseg; + uint32_t csum; + + /* + * Calculate and set check sum flags first, dword field + * in segment may be shared with Software Parser flags. + */ + csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0; + es->flags = rte_cpu_to_le_32(csum); + /* + * Calculate and set Software Parser offsets and flags. + * These flags a set for custom UDP and IP tunnel packets. + */ + es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx); + /* Fill metadata field if needed. */ + es->metadata = MLX5_TXOFF_CONFIG(METADATA) ? + loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0; + /* Engage VLAN tag insertion feature if requested. */ + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) { + /* + * We should get here only if device support + * this feature correctly. + */ + assert(txq->vlan_en); + es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT | + loc->mbuf->vlan_tci); + } else { + es->inline_hdr = RTE_BE32(0); + } } +/** + * Build the Ethernet Segment with minimal inlined data + * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is + * used to fill the gap in single WQEBB WQEs. + * Supports Software Parser, Checksums and VLAN + * insertion Tx offload features. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Ethernet Segment. + * @param vlan + * Length of VLAN tag insertion if any. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe *restrict wqe, + unsigned int vlan, + unsigned int olx) +{ + struct mlx5_wqe_eseg *restrict es = &wqe->eseg; + uint32_t csum; + uint8_t *psrc, *pdst; + + /* + * Calculate and set check sum flags first, dword field + * in segment may be shared with Software Parser flags. + */ + csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0; + es->flags = rte_cpu_to_le_32(csum); + /* + * Calculate and set Software Parser offsets and flags. + * These flags a set for custom UDP and IP tunnel packets. + */ + es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx); + /* Fill metadata field if needed. */ + es->metadata = MLX5_TXOFF_CONFIG(METADATA) ? + loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0; + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(rte_v128u32_t)), + "invalid Ethernet Segment data size"); + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(struct rte_vlan_hdr) + + 2 * RTE_ETHER_ADDR_LEN), + "invalid Ethernet Segment data size"); + psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *); + es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE); + es->inline_data = *(unaligned_uint16_t *)psrc; + psrc += sizeof(uint16_t); + pdst = (uint8_t *)(es + 1); + if (MLX5_TXOFF_CONFIG(VLAN) && vlan) { + /* Implement VLAN tag insertion as part inline data. */ + memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t)); + pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t); + psrc += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t); + /* Insert VLAN ethertype + VLAN tag. */ + *(unaligned_uint32_t *)pdst = rte_cpu_to_be_32 + ((RTE_ETHER_TYPE_VLAN << 16) | + loc->mbuf->vlan_tci); + pdst += sizeof(struct rte_vlan_hdr); + /* Copy the rest two bytes from packet data. */ + assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t))); + *(uint16_t *)pdst = *(unaligned_uint16_t *)psrc; + } else { + /* Fill the gap in the title WQEBB with inline data. */ + rte_mov16(pdst, psrc); + } +} +/** + * Build the Ethernet Segment with entire packet + * data inlining. Checks the boundary of WQEBB and + * ring buffer wrapping, supports Software Parser, + * Checksums and VLAN insertion Tx offload features. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Ethernet Segment. + * @param vlan + * Length of VLAN tag insertion if any. + * @param inlen + * Length of data to inline (VLAN included, if any). + * @param tso + * TSO flag, set mss field from the packet. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Pointer to the next Data Segment (aligned and wrapped around). + */ +static __rte_always_inline struct mlx5_wqe_dseg * +mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe *restrict wqe, + unsigned int vlan, + unsigned int inlen, + unsigned int tso, + unsigned int olx) +{ + struct mlx5_wqe_eseg *restrict es = &wqe->eseg; + uint32_t csum; + uint8_t *psrc, *pdst; + unsigned int part; + + /* + * Calculate and set check sum flags first, dword field + * in segment may be shared with Software Parser flags. + */ + csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0; + if (tso) { + csum <<= 24; + csum |= loc->mbuf->tso_segsz; + es->flags = rte_cpu_to_be_32(csum); + } else { + es->flags = rte_cpu_to_le_32(csum); + } + /* + * Calculate and set Software Parser offsets and flags. + * These flags a set for custom UDP and IP tunnel packets. + */ + es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx); + /* Fill metadata field if needed. */ + es->metadata = MLX5_TXOFF_CONFIG(METADATA) ? + loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0; + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(rte_v128u32_t)), + "invalid Ethernet Segment data size"); + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(struct rte_vlan_hdr) + + 2 * RTE_ETHER_ADDR_LEN), + "invalid Ethernet Segment data size"); + psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *); + es->inline_hdr_sz = rte_cpu_to_be_16(inlen); + es->inline_data = *(unaligned_uint16_t *)psrc; + psrc += sizeof(uint16_t); + pdst = (uint8_t *)(es + 1); + if (MLX5_TXOFF_CONFIG(VLAN) && vlan) { + /* Implement VLAN tag insertion as part inline data. */ + memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t)); + pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t); + psrc += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t); + /* Insert VLAN ethertype + VLAN tag. */ + *(unaligned_uint32_t *)pdst = rte_cpu_to_be_32 + ((RTE_ETHER_TYPE_VLAN << 16) | + loc->mbuf->vlan_tci); + pdst += sizeof(struct rte_vlan_hdr); + /* Copy the rest two bytes from packet data. */ + assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t))); + *(uint16_t *)pdst = *(unaligned_uint16_t *)psrc; + psrc += sizeof(uint16_t); + } else { + /* Fill the gap in the title WQEBB with inline data. */ + rte_mov16(pdst, psrc); + psrc += sizeof(rte_v128u32_t); + } + pdst = (uint8_t *)(es + 2); + assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE); + assert(pdst < (uint8_t *)txq->wqes_end); + inlen -= MLX5_ESEG_MIN_INLINE_SIZE; + if (!inlen) { + assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE)); + return (struct mlx5_wqe_dseg *)pdst; + } + /* + * The WQEBB space availability is checked by caller. + * Here we should be aware of WQE ring buffer wraparound only. + */ + part = (uint8_t *)txq->wqes_end - pdst; + part = RTE_MIN(part, inlen); + do { + rte_memcpy(pdst, psrc, part); + inlen -= part; + if (likely(!inlen)) { + /* + * If return value is not used by the caller + * the code below will be optimized out. + */ + pdst += part; + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + if (unlikely(pdst >= (uint8_t *)txq->wqes_end)) + pdst = (uint8_t *)txq->wqes; + return (struct mlx5_wqe_dseg *)pdst; + } + pdst = (uint8_t *)txq->wqes; + psrc += part; + part = inlen; + } while (true); +} + +/** + * Copy data from chain of mbuf to the specified linear buffer. + * Checksums and VLAN insertion Tx offload features. If data + * from some mbuf copied completely this mbuf is freed. Local + * structure is used to keep the byte stream state. + * + * @param pdst + * Pointer to the destination linear buffer. + * @param loc + * Pointer to burst routine local context. + * @param len + * Length of data to be copied. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_mseg_memcpy(uint8_t *pdst, + struct mlx5_txq_local *restrict loc, + unsigned int len, + unsigned int olx __rte_unused) +{ + struct rte_mbuf *mbuf; + unsigned int part, dlen; + uint8_t *psrc; + + assert(len); + do { + /* Allow zero length packets, must check first. */ + dlen = rte_pktmbuf_data_len(loc->mbuf); + if (dlen <= loc->mbuf_off) { + /* Exhausted packet, just free. */ + mbuf = loc->mbuf; + loc->mbuf = mbuf->next; + rte_pktmbuf_free_seg(mbuf); + loc->mbuf_off = 0; + assert(loc->mbuf_nseg > 1); + assert(loc->mbuf); + --loc->mbuf_nseg; + continue; + } + dlen -= loc->mbuf_off; + psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *, + loc->mbuf_off); + part = RTE_MIN(len, dlen); + rte_memcpy(pdst, psrc, part); + loc->mbuf_off += part; + len -= part; + if (!len) { + if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) { + loc->mbuf_off = 0; + /* Exhausted packet, just free. */ + mbuf = loc->mbuf; + loc->mbuf = mbuf->next; + rte_pktmbuf_free_seg(mbuf); + loc->mbuf_off = 0; + assert(loc->mbuf_nseg >= 1); + --loc->mbuf_nseg; + } + return; + } + pdst += part; + } while (true); +} + +/** + * Build the Ethernet Segment with inlined data from + * multi-segment packet. Checks the boundary of WQEBB + * and ring buffer wrapping, supports Software Parser, + * Checksums and VLAN insertion Tx offload features. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Ethernet Segment. + * @param vlan + * Length of VLAN tag insertion if any. + * @param inlen + * Length of data to inline (VLAN included, if any). + * @param tso + * TSO flag, set mss field from the packet. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Pointer to the next Data Segment (aligned and + * possible NOT wrapped around - caller should do + * wrapping check on its own). + */ +static __rte_always_inline struct mlx5_wqe_dseg * +mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe *restrict wqe, + unsigned int vlan, + unsigned int inlen, + unsigned int tso, + unsigned int olx) +{ + struct mlx5_wqe_eseg *restrict es = &wqe->eseg; + uint32_t csum; + uint8_t *pdst; + unsigned int part; + + /* + * Calculate and set check sum flags first, uint32_t field + * in segment may be shared with Software Parser flags. + */ + csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0; + if (tso) { + csum <<= 24; + csum |= loc->mbuf->tso_segsz; + es->flags = rte_cpu_to_be_32(csum); + } else { + es->flags = rte_cpu_to_le_32(csum); + } + /* + * Calculate and set Software Parser offsets and flags. + * These flags a set for custom UDP and IP tunnel packets. + */ + es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx); + /* Fill metadata field if needed. */ + es->metadata = MLX5_TXOFF_CONFIG(METADATA) ? + loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0 : 0; + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(rte_v128u32_t)), + "invalid Ethernet Segment data size"); + static_assert(MLX5_ESEG_MIN_INLINE_SIZE == + (sizeof(uint16_t) + + sizeof(struct rte_vlan_hdr) + + 2 * RTE_ETHER_ADDR_LEN), + "invalid Ethernet Segment data size"); + assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE); + es->inline_hdr_sz = rte_cpu_to_be_16(inlen); + pdst = (uint8_t *)&es->inline_data; + if (MLX5_TXOFF_CONFIG(VLAN) && vlan) { + /* Implement VLAN tag insertion as part inline data. */ + mlx5_tx_mseg_memcpy(pdst, loc, 2 * RTE_ETHER_ADDR_LEN, olx); + pdst += 2 * RTE_ETHER_ADDR_LEN; + *(unaligned_uint32_t *)pdst = rte_cpu_to_be_32 + ((RTE_ETHER_TYPE_VLAN << 16) | + loc->mbuf->vlan_tci); + pdst += sizeof(struct rte_vlan_hdr); + inlen -= 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr); + } + assert(pdst < (uint8_t *)txq->wqes_end); + /* + * The WQEBB space availability is checked by caller. + * Here we should be aware of WQE ring buffer wraparound only. + */ + part = (uint8_t *)txq->wqes_end - pdst; + part = RTE_MIN(part, inlen); + assert(part); + do { + mlx5_tx_mseg_memcpy(pdst, loc, part, olx); + inlen -= part; + if (likely(!inlen)) { + pdst += part; + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + return (struct mlx5_wqe_dseg *)pdst; + } + pdst = (uint8_t *)txq->wqes; + part = inlen; + } while (true); +} + +/** + * Build the Data Segment of pointer type. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param dseg + * Pointer to WQE to fill with built Data Segment. + * @param buf + * Data buffer to point. + * @param len + * Data buffer length. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe_dseg *restrict dseg, + uint8_t *buf, + unsigned int len, + unsigned int olx __rte_unused) + +{ + assert(len); + dseg->bcount = rte_cpu_to_be_32(len); + dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf); + dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf); +} + +/** + * Build the Data Segment of pointer type or inline + * if data length is less than buffer in minimal + * Data Segment size. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param dseg + * Pointer to WQE to fill with built Data Segment. + * @param buf + * Data buffer to point. + * @param len + * Data buffer length. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + */ +static __rte_always_inline void +mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe_dseg *restrict dseg, + uint8_t *buf, + unsigned int len, + unsigned int olx __rte_unused) + +{ + uintptr_t dst, src; + + assert(len); + if (len > MLX5_DSEG_MIN_INLINE_SIZE) { + dseg->bcount = rte_cpu_to_be_32(len); + dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf); + dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf); + + return; + } + dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE); + /* Unrolled implementation of generic rte_memcpy. */ + dst = (uintptr_t)&dseg->inline_data[0]; + src = (uintptr_t)buf; + if (len & 0x08) { +#ifdef RTE_ARCH_STRICT_ALIGN + assert(dst == RTE_PTR_ALIGN(dst, sizeof(uint32_t))); + *(uint32_t *)dst = *(unaligned_uint32_t *)src; + dst += sizeof(uint32_t); + src += sizeof(uint32_t); + *(uint32_t *)dst = *(unaligned_uint32_t *)src; + dst += sizeof(uint32_t); + src += sizeof(uint32_t); +#else + *(uint64_t *)dst = *(unaligned_uint64_t *)src; + dst += sizeof(uint64_t); + src += sizeof(uint64_t); +#endif + } + if (len & 0x04) { + *(uint32_t *)dst = *(unaligned_uint32_t *)src; + dst += sizeof(uint32_t); + src += sizeof(uint32_t); + } + if (len & 0x02) { + *(uint16_t *)dst = *(unaligned_uint16_t *)src; + dst += sizeof(uint16_t); + src += sizeof(uint16_t); + } + if (len & 0x01) + *(uint8_t *)dst = *(uint8_t *)src; +} + +/** + * Build the Data Segment of inlined data from single + * segment packet, no VLAN insertion. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param dseg + * Pointer to WQE to fill with built Data Segment. + * @param buf + * Data buffer to point. + * @param len + * Data buffer length. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Pointer to the next Data Segment after inlined data. + * Ring buffer wraparound check is needed. We do not + * do it here because it may not be needed for the + * last packet in the eMPW session. + */ +static __rte_always_inline struct mlx5_wqe_dseg * +mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc __rte_unused, + struct mlx5_wqe_dseg *restrict dseg, + uint8_t *buf, + unsigned int len, + unsigned int olx __rte_unused) +{ + unsigned int part; + uint8_t *pdst; + + dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE); + pdst = &dseg->inline_data[0]; + /* + * The WQEBB space availability is checked by caller. + * Here we should be aware of WQE ring buffer wraparound only. + */ + part = (uint8_t *)txq->wqes_end - pdst; + part = RTE_MIN(part, len); + do { + rte_memcpy(pdst, buf, part); + len -= part; + if (likely(!len)) { + pdst += part; + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + /* Note: no final wraparound check here. */ + return (struct mlx5_wqe_dseg *)pdst; + } + pdst = (uint8_t *)txq->wqes; + buf += part; + part = len; + } while (true); +} + +/** + * Build the Data Segment of inlined data from single + * segment packet with VLAN insertion. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param dseg + * Pointer to the dseg fill with built Data Segment. + * @param buf + * Data buffer to point. + * @param len + * Data buffer length. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Pointer to the next Data Segment after inlined data. + * Ring buffer wraparound check is needed. + */ +static __rte_always_inline struct mlx5_wqe_dseg * +mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc __rte_unused, + struct mlx5_wqe_dseg *restrict dseg, + uint8_t *buf, + unsigned int len, + unsigned int olx __rte_unused) + +{ + unsigned int part; + uint8_t *pdst; + + assert(len > MLX5_ESEG_MIN_INLINE_SIZE); + static_assert(MLX5_DSEG_MIN_INLINE_SIZE == + (2 * RTE_ETHER_ADDR_LEN), + "invalid Data Segment data size"); + dseg->bcount = rte_cpu_to_be_32((len + sizeof(struct rte_vlan_hdr)) | + MLX5_ETH_WQE_DATA_INLINE); + pdst = &dseg->inline_data[0]; + memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE); + buf += MLX5_DSEG_MIN_INLINE_SIZE; + pdst += MLX5_DSEG_MIN_INLINE_SIZE; + len -= MLX5_DSEG_MIN_INLINE_SIZE; + /* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */ + assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE)); + if (unlikely(pdst >= (uint8_t *)txq->wqes_end)) + pdst = (uint8_t *)txq->wqes; + *(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) | + loc->mbuf->vlan_tci); + pdst += sizeof(struct rte_vlan_hdr); + /* + * The WQEBB space availability is checked by caller. + * Here we should be aware of WQE ring buffer wraparound only. + */ + part = (uint8_t *)txq->wqes_end - pdst; + part = RTE_MIN(part, len); + do { + rte_memcpy(pdst, buf, part); + len -= part; + if (likely(!len)) { + pdst += part; + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + /* Note: no final wraparound check here. */ + return (struct mlx5_wqe_dseg *)pdst; + } + pdst = (uint8_t *)txq->wqes; + buf += part; + part = len; + } while (true); +} + +/** + * Build the Ethernet Segment with optionally inlined data with + * VLAN insertion and following Data Segments (if any) from + * multi-segment packet. Used by ordinary send and TSO. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param wqe + * Pointer to WQE to fill with built Ethernet/Data Segments. + * @param vlan + * Length of VLAN header to insert, 0 means no VLAN insertion. + * @param inlen + * Data length to inline. For TSO this parameter specifies + * exact value, for ordinary send routine can be aligned by + * caller to provide better WQE space saving and data buffer + * start address alignment. This length includes VLAN header + * being inserted. + * @param tso + * Zero means ordinary send, inlined data can be extended, + * otherwise this is TSO, inlined data length is fixed. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * Actual size of built WQE in segments. + */ +static __rte_always_inline unsigned int +mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + struct mlx5_wqe *restrict wqe, + unsigned int vlan, + unsigned int inlen, + unsigned int tso, + unsigned int olx __rte_unused) +{ + struct mlx5_wqe_dseg *restrict dseg; + unsigned int ds; + + assert((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen); + loc->mbuf_nseg = NB_SEGS(loc->mbuf); + loc->mbuf_off = 0; + + dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx); + if (!loc->mbuf_nseg) + goto dseg_done; + /* + * There are still some mbuf remaining, not inlined. + * The first mbuf may be partially inlined and we + * must process the possible non-zero data offset. + */ + if (loc->mbuf_off) { + unsigned int dlen; + uint8_t *dptr; + + /* + * Exhausted packets must be dropped before. + * Non-zero offset means there are some data + * remained in the packet. + */ + assert(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf)); + assert(rte_pktmbuf_data_len(loc->mbuf)); + dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *, + loc->mbuf_off); + dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off; + /* + * Build the pointer/minimal data Data Segment. + * Do ring buffer wrapping check in advance. + */ + if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) + dseg = (struct mlx5_wqe_dseg *)txq->wqes; + mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx); + /* Store the mbuf to be freed on completion. */ + assert(loc->elts_free); + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + --loc->elts_free; + ++dseg; + if (--loc->mbuf_nseg == 0) + goto dseg_done; + loc->mbuf = loc->mbuf->next; + loc->mbuf_off = 0; + } + do { + if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) { + struct rte_mbuf *mbuf; + + /* Zero length segment found, just skip. */ + mbuf = loc->mbuf; + loc->mbuf = loc->mbuf->next; + rte_pktmbuf_free_seg(mbuf); + if (--loc->mbuf_nseg == 0) + break; + } else { + if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) + dseg = (struct mlx5_wqe_dseg *)txq->wqes; + mlx5_tx_dseg_iptr + (txq, loc, dseg, + rte_pktmbuf_mtod(loc->mbuf, uint8_t *), + rte_pktmbuf_data_len(loc->mbuf), olx); + assert(loc->elts_free); + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + --loc->elts_free; + ++dseg; + if (--loc->mbuf_nseg == 0) + break; + loc->mbuf = loc->mbuf->next; + } + } while (true); + +dseg_done: + /* Calculate actual segments used from the dseg pointer. */ + if ((uintptr_t)wqe < (uintptr_t)dseg) + ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE; + else + ds = (((uintptr_t)dseg - (uintptr_t)wqe) + + txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE; + return ds; +} + +/** + * Tx one packet function for multi-segment TSO. Supports all + * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs, + * sends one packet per WQE. + * + * This routine is responsible for storing processed mbuf + * into elts ring buffer and update elts_head. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * Local context variables partially updated. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + struct mlx5_wqe *restrict wqe; + unsigned int ds, dlen, inlen, ntcp, vlan = 0; + + /* + * Calculate data length to be inlined to estimate + * the required space in WQE ring buffer. + */ + dlen = rte_pktmbuf_pkt_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) + vlan = sizeof(struct rte_vlan_hdr); + inlen = loc->mbuf->l2_len + vlan + + loc->mbuf->l3_len + loc->mbuf->l4_len; + if (unlikely((!inlen || !loc->mbuf->tso_segsz))) + return MLX5_TXCMP_CODE_ERROR; + if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK) + inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len; + /* Packet must contain all TSO headers. */ + if (unlikely(inlen > MLX5_MAX_TSO_HEADER || + inlen <= MLX5_ESEG_MIN_INLINE_SIZE || + inlen > (dlen + vlan))) + return MLX5_TXCMP_CODE_ERROR; + assert(inlen >= txq->inlen_mode); + /* + * Check whether there are enough free WQEBBs: + * - Control Segment + * - Ethernet Segment + * - First Segment of inlined Ethernet data + * - ... data continued ... + * - Data Segments of pointer/min inline type + */ + ds = NB_SEGS(loc->mbuf) + 2 + (inlen - + MLX5_ESEG_MIN_INLINE_SIZE + + MLX5_WSEG_SIZE + + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE; + if (unlikely(loc->wqe_free < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_EXIT; + /* Check for maximal WQE size. */ + if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_ERROR; +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes/packets counters. */ + ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) / + loc->mbuf->tso_segsz; + /* + * One will be added for mbuf itself + * at the end of the mlx5_tx_burst from + * loc->pkts_sent field. + */ + --ntcp; + txq->stats.opackets += ntcp; + txq->stats.obytes += dlen + vlan + ntcp * inlen; +#endif + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx); + ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx); + wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds); + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, true, olx); + return MLX5_TXCMP_CODE_MULTI; +} + +/** + * Tx one packet function for multi-segment SEND. Supports all + * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs, + * sends one packet per WQE, without any data inlining in + * Ethernet Segment. + * + * This routine is responsible for storing processed mbuf + * into elts ring buffer and update elts_head. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * Local context variables partially updated. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + struct mlx5_wqe_dseg *restrict dseg; + struct mlx5_wqe *restrict wqe; + unsigned int ds, nseg; + + assert(NB_SEGS(loc->mbuf) > 1); + /* + * No inline at all, it means the CPU cycles saving + * is prioritized at configuration, we should not + * copy any packet data to WQE. + */ + nseg = NB_SEGS(loc->mbuf); + ds = 2 + nseg; + if (unlikely(loc->wqe_free < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_EXIT; + /* Check for maximal WQE size. */ + if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_ERROR; + /* + * Some Tx offloads may cause an error if + * packet is not long enough, check against + * assumed minimal length. + */ + if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE) + return MLX5_TXCMP_CODE_ERROR; +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) + txq->stats.obytes += sizeof(struct rte_vlan_hdr); +#endif + /* + * SEND WQE, one WQEBB: + * - Control Segment, SEND opcode + * - Ethernet Segment, optional VLAN, no inline + * - Data Segments, pointer only type + */ + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx); + mlx5_tx_eseg_none(txq, loc, wqe, olx); + dseg = &wqe->dseg[0]; + do { + if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) { + struct rte_mbuf *mbuf; + + /* + * Zero length segment found, have to + * correct total size of WQE in segments. + * It is supposed to be rare occasion, so + * in normal case (no zero length segments) + * we avoid extra writing to the Control + * Segment. + */ + --ds; + wqe->cseg.sq_ds -= RTE_BE32(1); + mbuf = loc->mbuf; + loc->mbuf = mbuf->next; + rte_pktmbuf_free_seg(mbuf); + if (--nseg == 0) + break; + } else { + mlx5_tx_dseg_ptr + (txq, loc, dseg, + rte_pktmbuf_mtod(loc->mbuf, uint8_t *), + rte_pktmbuf_data_len(loc->mbuf), olx); + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + --loc->elts_free; + if (--nseg == 0) + break; + ++dseg; + if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) + dseg = (struct mlx5_wqe_dseg *)txq->wqes; + loc->mbuf = loc->mbuf->next; + } + } while (true); + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, true, olx); + return MLX5_TXCMP_CODE_MULTI; +} + +/** + * Tx one packet function for multi-segment SEND. Supports all + * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs, + * sends one packet per WQE, with data inlining in + * Ethernet Segment and minimal Data Segments. + * + * This routine is responsible for storing processed mbuf + * into elts ring buffer and update elts_head. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * Local context variables partially updated. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + struct mlx5_wqe *restrict wqe; + unsigned int ds, inlen, dlen, vlan = 0; + + assert(MLX5_TXOFF_CONFIG(INLINE)); + assert(NB_SEGS(loc->mbuf) > 1); + /* + * First calculate data length to be inlined + * to estimate the required space for WQE. + */ + dlen = rte_pktmbuf_pkt_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) + vlan = sizeof(struct rte_vlan_hdr); + inlen = dlen + vlan; + /* Check against minimal length. */ + if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE) + return MLX5_TXCMP_CODE_ERROR; + assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE); + if (inlen > txq->inlen_send) { + struct rte_mbuf *mbuf; + unsigned int nxlen; + uintptr_t start; + + /* + * Packet length exceeds the allowed inline + * data length, check whether the minimal + * inlining is required. + */ + if (txq->inlen_mode) { + assert(txq->inlen_mode >= MLX5_ESEG_MIN_INLINE_SIZE); + assert(txq->inlen_mode <= txq->inlen_send); + inlen = txq->inlen_mode; + } else { + if (!vlan || txq->vlan_en) { + /* + * VLAN insertion will be done inside by HW. + * It is not utmost effective - VLAN flag is + * checked twice, but we should proceed the + * inlining length correctly and take into + * account the VLAN header being inserted. + */ + return mlx5_tx_packet_multi_send + (txq, loc, olx); + } + inlen = MLX5_ESEG_MIN_INLINE_SIZE; + } + /* + * Now we know the minimal amount of data is requested + * to inline. Check whether we should inline the buffers + * from the chain beginning to eliminate some mbufs. + */ + mbuf = loc->mbuf; + nxlen = rte_pktmbuf_data_len(mbuf); + if (unlikely(nxlen <= txq->inlen_send)) { + /* We can inline first mbuf at least. */ + if (nxlen < inlen) { + unsigned int smlen; + + /* Scan mbufs till inlen filled. */ + do { + smlen = nxlen; + mbuf = NEXT(mbuf); + assert(mbuf); + nxlen = rte_pktmbuf_data_len(mbuf); + nxlen += smlen; + } while (unlikely(nxlen < inlen)); + if (unlikely(nxlen > txq->inlen_send)) { + /* We cannot inline entire mbuf. */ + smlen = inlen - smlen; + start = rte_pktmbuf_mtod_offset + (mbuf, uintptr_t, smlen); + goto do_align; + } + } + do { + inlen = nxlen; + mbuf = NEXT(mbuf); + /* There should be not end of packet. */ + assert(mbuf); + nxlen = inlen + rte_pktmbuf_data_len(mbuf); + } while (unlikely(nxlen < txq->inlen_send)); + } + start = rte_pktmbuf_mtod(mbuf, uintptr_t); + /* + * Check whether we can do inline to align start + * address of data buffer to cacheline. + */ +do_align: + start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1); + if (unlikely(start)) { + start += inlen; + if (start <= txq->inlen_send) + inlen = start; + } + } + /* + * Check whether there are enough free WQEBBs: + * - Control Segment + * - Ethernet Segment + * - First Segment of inlined Ethernet data + * - ... data continued ... + * - Data Segments of pointer/min inline type + * + * Estimate the number of Data Segments conservatively, + * supposing no any mbufs is being freed during inlining. + */ + assert(inlen <= txq->inlen_send); + ds = NB_SEGS(loc->mbuf) + 2 + (inlen - + MLX5_ESEG_MIN_INLINE_SIZE + + MLX5_WSEG_SIZE + + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE; + if (unlikely(loc->wqe_free < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_EXIT; + /* Check for maximal WQE size. */ + if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4))) + return MLX5_TXCMP_CODE_ERROR; +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes/packets counters. */ + txq->stats.obytes += dlen + vlan; +#endif + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx); + ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx); + wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds); + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, true, olx); + return MLX5_TXCMP_CODE_MULTI; +} + +/** + * Tx burst function for multi-segment packets. Supports all + * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs, + * sends one packet per WQE. Function stops sending if it + * encounters the single-segment packet. + * + * This routine is responsible for storing processed mbuf + * into elts ring buffer and update elts_head. + * + * @param txq + * Pointer to TX queue structure. + * @param[in] pkts + * Packets to transmit. + * @param pkts_n + * Number of packets in array. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered. + * MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered. + * Local context variables updated. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + assert(loc->elts_free && loc->wqe_free); + assert(pkts_n > loc->pkts_sent); + pkts += loc->pkts_sent + 1; + pkts_n -= loc->pkts_sent; + for (;;) { + enum mlx5_txcmp_code ret; + + assert(NB_SEGS(loc->mbuf) > 1); + /* + * Estimate the number of free elts quickly but + * conservatively. Some segment may be fully inlined + * and freed, ignore this here - precise estimation + * is costly. + */ + if (loc->elts_free < NB_SEGS(loc->mbuf)) + return MLX5_TXCMP_CODE_EXIT; + if (MLX5_TXOFF_CONFIG(TSO) && + unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) { + /* Proceed with multi-segment TSO. */ + ret = mlx5_tx_packet_multi_tso(txq, loc, olx); + } else if (MLX5_TXOFF_CONFIG(INLINE)) { + /* Proceed with multi-segment SEND with inlining. */ + ret = mlx5_tx_packet_multi_inline(txq, loc, olx); + } else { + /* Proceed with multi-segment SEND w/o inlining. */ + ret = mlx5_tx_packet_multi_send(txq, loc, olx); + } + if (ret == MLX5_TXCMP_CODE_EXIT) + return MLX5_TXCMP_CODE_EXIT; + if (ret == MLX5_TXCMP_CODE_ERROR) + return MLX5_TXCMP_CODE_ERROR; + /* WQE is built, go to the next packet. */ + ++loc->pkts_sent; + --pkts_n; + if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + loc->mbuf = *pkts++; + if (pkts_n > 1) + rte_prefetch0(*pkts); + if (likely(NB_SEGS(loc->mbuf) > 1)) + continue; + /* Here ends the series of multi-segment packets. */ + if (MLX5_TXOFF_CONFIG(TSO) && + unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) + return MLX5_TXCMP_CODE_TSO; + return MLX5_TXCMP_CODE_SINGLE; + } + assert(false); +} + +/** + * Tx burst function for single-segment packets with TSO. + * Supports all types of Tx offloads, except multi-packets. + * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE. + * Function stops sending if it encounters the multi-segment + * packet or packet without TSO requested. + * + * The routine is responsible for storing processed mbuf + * into elts ring buffer and update elts_head if inline + * offloads is requested due to possible early freeing + * of the inlined mbufs (can not store pkts array in elts + * as a batch). + * + * @param txq + * Pointer to TX queue structure. + * @param[in] pkts + * Packets to transmit. + * @param pkts_n + * Number of packets in array. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered. + * MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered. + * Local context variables updated. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + assert(loc->elts_free && loc->wqe_free); + assert(pkts_n > loc->pkts_sent); + pkts += loc->pkts_sent + 1; + pkts_n -= loc->pkts_sent; + for (;;) { + struct mlx5_wqe_dseg *restrict dseg; + struct mlx5_wqe *restrict wqe; + unsigned int ds, dlen, hlen, ntcp, vlan = 0; + uint8_t *dptr; + + assert(NB_SEGS(loc->mbuf) == 1); + dlen = rte_pktmbuf_data_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) { + vlan = sizeof(struct rte_vlan_hdr); + } + /* + * First calculate the WQE size to check + * whether we have enough space in ring buffer. + */ + hlen = loc->mbuf->l2_len + vlan + + loc->mbuf->l3_len + loc->mbuf->l4_len; + if (unlikely((!hlen || !loc->mbuf->tso_segsz))) + return MLX5_TXCMP_CODE_ERROR; + if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK) + hlen += loc->mbuf->outer_l2_len + + loc->mbuf->outer_l3_len; + /* Segment must contain all TSO headers. */ + if (unlikely(hlen > MLX5_MAX_TSO_HEADER || + hlen <= MLX5_ESEG_MIN_INLINE_SIZE || + hlen > (dlen + vlan))) + return MLX5_TXCMP_CODE_ERROR; + /* + * Check whether there are enough free WQEBBs: + * - Control Segment + * - Ethernet Segment + * - First Segment of inlined Ethernet data + * - ... data continued ... + * - Finishing Data Segment of pointer type + */ + ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE + + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE; + if (loc->wqe_free < ((ds + 3) / 4)) + return MLX5_TXCMP_CODE_EXIT; +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes/packets counters. */ + ntcp = (dlen + vlan - hlen + + loc->mbuf->tso_segsz - 1) / + loc->mbuf->tso_segsz; + /* + * One will be added for mbuf itself at the end + * of the mlx5_tx_burst from loc->pkts_sent field. + */ + --ntcp; + txq->stats.opackets += ntcp; + txq->stats.obytes += dlen + vlan + ntcp * hlen; +#endif + /* + * Build the TSO WQE: + * - Control Segment + * - Ethernet Segment with hlen bytes inlined + * - Data Segment of pointer type + */ + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, ds, + MLX5_OPCODE_TSO, olx); + dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx); + dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan; + dlen -= hlen - vlan; + mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx); + /* + * WQE is built, update the loop parameters + * and go to the next packet. + */ + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + if (MLX5_TXOFF_CONFIG(INLINE)) + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + --loc->elts_free; + ++loc->pkts_sent; + --pkts_n; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, false, olx); + if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + loc->mbuf = *pkts++; + if (pkts_n > 1) + rte_prefetch0(*pkts); + if (MLX5_TXOFF_CONFIG(MULTI) && + unlikely(NB_SEGS(loc->mbuf) > 1)) + return MLX5_TXCMP_CODE_MULTI; + if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))) + return MLX5_TXCMP_CODE_SINGLE; + /* Continue with the next TSO packet. */ + } + assert(false); +} + +/** + * Analyze the packet and select the best method to send. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * @param newp + * The predefined flag whether do complete check for + * multi-segment packets and TSO. + * + * @return + * MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered. + * MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO. + * MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND. + * MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int olx, + bool newp) +{ + /* Check for multi-segment packet. */ + if (newp && + MLX5_TXOFF_CONFIG(MULTI) && + unlikely(NB_SEGS(loc->mbuf) > 1)) + return MLX5_TXCMP_CODE_MULTI; + /* Check for TSO packet. */ + if (newp && + MLX5_TXOFF_CONFIG(TSO) && + unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) + return MLX5_TXCMP_CODE_TSO; + /* Check if eMPW is enabled at all. */ + if (!MLX5_TXOFF_CONFIG(EMPW)) + return MLX5_TXCMP_CODE_SINGLE; + /* Check if eMPW can be engaged. */ + if (MLX5_TXOFF_CONFIG(VLAN) && + unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) && + (!MLX5_TXOFF_CONFIG(INLINE) || + unlikely((rte_pktmbuf_data_len(loc->mbuf) + + sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) { + /* + * eMPW does not support VLAN insertion offload, + * we have to inline the entire packet but + * packet is too long for inlining. + */ + return MLX5_TXCMP_CODE_SINGLE; + } + return MLX5_TXCMP_CODE_EMPW; +} + +/** + * Check the next packet attributes to match with the eMPW batch ones. + * + * @param txq + * Pointer to TX queue structure. + * @param es + * Pointer to Ethernet Segment of eMPW batch. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * true - packet match with eMPW batch attributes. + * false - no match, eMPW should be restarted. + */ +static __rte_always_inline bool +mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused, + struct mlx5_wqe_eseg *restrict es, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + uint8_t swp_flags = 0; + + /* Compare the checksum flags, if any. */ + if (MLX5_TXOFF_CONFIG(CSUM) && + txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags) + return false; + /* Compare the Software Parser offsets and flags. */ + if (MLX5_TXOFF_CONFIG(SWP) && + (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) || + es->swp_flags != swp_flags)) + return false; + /* Fill metadata field if needed. */ + if (MLX5_TXOFF_CONFIG(METADATA) && + es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0)) + return false; + /* There must be no VLAN packets in eMPW loop. */ + if (MLX5_TXOFF_CONFIG(VLAN)) + assert(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)); + return true; +} + +/* + * Update send loop variables and WQE for eMPW loop + * without data inlining. Number of Data Segments is + * equal to the number of sent packets. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param ds + * Number of packets/Data Segments/Packets. + * @param slen + * Accumulated statistics, bytes sent + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * true - packet match with eMPW batch attributes. + * false - no match, eMPW should be restarted. + */ +static __rte_always_inline void +mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int ds, + unsigned int slen, + unsigned int olx) +{ + assert(!MLX5_TXOFF_CONFIG(INLINE)); +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += slen; +#else + (void)slen; +#endif + loc->elts_free -= ds; + loc->pkts_sent += ds; + ds += 2; + loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds); + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, false, olx); +} + +/* + * Update send loop variables and WQE for eMPW loop + * with data inlining. Gets the size of pushed descriptors + * and data to the WQE. + * + * @param txq + * Pointer to TX queue structure. + * @param loc + * Pointer to burst routine local context. + * @param len + * Total size of descriptor/data in bytes. + * @param slen + * Accumulated statistics, data bytes sent. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * true - packet match with eMPW batch attributes. + * false - no match, eMPW should be restarted. + */ +static __rte_always_inline void +mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq, + struct mlx5_txq_local *restrict loc, + unsigned int len, + unsigned int slen, + unsigned int olx __rte_unused) +{ + assert(MLX5_TXOFF_CONFIG(INLINE)); + assert((len % MLX5_WSEG_SIZE) == 0); +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += slen; +#else + (void)slen; +#endif + len = len / MLX5_WSEG_SIZE + 2; + loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len); + txq->wqe_ci += (len + 3) / 4; + loc->wqe_free -= (len + 3) / 4; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, false, olx); +} + +/** + * The set of Tx burst functions for single-segment packets + * without TSO and with Multi-Packet Writing feature support. + * Supports all types of Tx offloads, except multi-packets + * and TSO. + * + * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends + * as many packet per WQE as it can. If eMPW is not configured + * or packet can not be sent with eMPW (VLAN insertion) the + * ordinary SEND opcode is used and only one packet placed + * in WQE. + * + * Functions stop sending if it encounters the multi-segment + * packet or packet with TSO requested. + * + * The routines are responsible for storing processed mbuf + * into elts ring buffer and update elts_head if inlining + * offload is requested. Otherwise the copying mbufs to elts + * can be postponed and completed at the end of burst routine. + * + * @param txq + * Pointer to TX queue structure. + * @param[in] pkts + * Packets to transmit. + * @param pkts_n + * Number of packets in array. + * @param loc + * Pointer to burst routine local context. + * @param olx + * Configured Tx offloads mask. It is fully defined at + * compile time and may be used for optimization. + * + * @return + * MLX5_TXCMP_CODE_EXIT - sending is done or impossible. + * MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred. + * MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered. + * MLX5_TXCMP_CODE_TSO - TSO packet encountered. + * MLX5_TXCMP_CODE_SINGLE - used inside functions set. + * MLX5_TXCMP_CODE_EMPW - used inside functions set. + * + * Local context variables updated. + * + * + * The routine sends packets with MLX5_OPCODE_EMPW + * without inlining, this is dedicated optimized branch. + * No VLAN insertion is supported. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + /* + * Subroutine is the part of mlx5_tx_burst_single() + * and sends single-segment packet with eMPW opcode + * without data inlining. + */ + assert(!MLX5_TXOFF_CONFIG(INLINE)); + assert(MLX5_TXOFF_CONFIG(EMPW)); + assert(loc->elts_free && loc->wqe_free); + assert(pkts_n > loc->pkts_sent); + static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size"); + pkts += loc->pkts_sent + 1; + pkts_n -= loc->pkts_sent; + for (;;) { + struct mlx5_wqe_dseg *restrict dseg; + struct mlx5_wqe_eseg *restrict eseg; + enum mlx5_txcmp_code ret; + unsigned int part, loop; + unsigned int slen = 0; + +next_empw: + part = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS); + if (unlikely(loc->elts_free < part)) { + /* We have no enough elts to save all mbufs. */ + if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS)) + return MLX5_TXCMP_CODE_EXIT; + /* But we still able to send at least minimal eMPW. */ + part = loc->elts_free; + } + /* Check whether we have enough WQEs */ + if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) { + if (unlikely(loc->wqe_free < + ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4))) + return MLX5_TXCMP_CODE_EXIT; + part = (loc->wqe_free * 4) - 2; + } + if (likely(part > 1)) + rte_prefetch0(*pkts); + loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m); + /* + * Build eMPW title WQEBB: + * - Control Segment, eMPW opcode + * - Ethernet Segment, no inline + */ + mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2, + MLX5_OPCODE_ENHANCED_MPSW, olx); + mlx5_tx_eseg_none(txq, loc, loc->wqe_last, + olx & ~MLX5_TXOFF_CONFIG_VLAN); + eseg = &loc->wqe_last->eseg; + dseg = &loc->wqe_last->dseg[0]; + loop = part; + for (;;) { + uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf); +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + slen += dlen; +#endif + mlx5_tx_dseg_ptr + (txq, loc, dseg, + rte_pktmbuf_mtod(loc->mbuf, uint8_t *), + dlen, olx); + if (unlikely(--loop == 0)) + break; + loc->mbuf = *pkts++; + if (likely(loop > 1)) + rte_prefetch0(*pkts); + ret = mlx5_tx_able_to_empw(txq, loc, olx, true); + /* + * Unroll the completion code to avoid + * returning variable value - it results in + * unoptimized sequent checking in caller. + */ + if (ret == MLX5_TXCMP_CODE_MULTI) { + part -= loop; + mlx5_tx_sdone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_MULTI; + } + if (ret == MLX5_TXCMP_CODE_TSO) { + part -= loop; + mlx5_tx_sdone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_TSO; + } + if (ret == MLX5_TXCMP_CODE_SINGLE) { + part -= loop; + mlx5_tx_sdone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_SINGLE; + } + if (ret != MLX5_TXCMP_CODE_EMPW) { + assert(false); + part -= loop; + mlx5_tx_sdone_empw(txq, loc, part, slen, olx); + return MLX5_TXCMP_CODE_ERROR; + } + /* + * Check whether packet parameters coincide + * within assumed eMPW batch: + * - check sum settings + * - metadata value + * - software parser settings + */ + if (!mlx5_tx_match_empw(txq, eseg, loc, olx)) { + assert(loop); + part -= loop; + mlx5_tx_sdone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + pkts_n -= part; + goto next_empw; + } + /* Packet attributes match, continue the same eMPW. */ + ++dseg; + if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) + dseg = (struct mlx5_wqe_dseg *)txq->wqes; + } + /* eMPW is built successfully, update loop parameters. */ + assert(!loop); + assert(pkts_n >= part); +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += slen; +#endif + loc->elts_free -= part; + loc->pkts_sent += part; + txq->wqe_ci += (2 + part + 3) / 4; + loc->wqe_free -= (2 + part + 3) / 4; + pkts_n -= part; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, false, olx); + if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + loc->mbuf = *pkts++; + ret = mlx5_tx_able_to_empw(txq, loc, olx, true); + if (unlikely(ret != MLX5_TXCMP_CODE_EMPW)) + return ret; + /* Continue sending eMPW batches. */ + } + assert(false); +} + +/** + * The routine sends packets with MLX5_OPCODE_EMPW + * with inlining, optionally supports VLAN insertion. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + /* + * Subroutine is the part of mlx5_tx_burst_single() + * and sends single-segment packet with eMPW opcode + * with data inlining. + */ + assert(MLX5_TXOFF_CONFIG(INLINE)); + assert(MLX5_TXOFF_CONFIG(EMPW)); + assert(loc->elts_free && loc->wqe_free); + assert(pkts_n > loc->pkts_sent); + static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size"); + pkts += loc->pkts_sent + 1; + pkts_n -= loc->pkts_sent; + for (;;) { + struct mlx5_wqe_dseg *restrict dseg; + struct mlx5_wqe_eseg *restrict eseg; + enum mlx5_txcmp_code ret; + unsigned int room, part, nlim; + unsigned int slen = 0; + + /* + * Limits the amount of packets in one WQE + * to improve CQE latency generation. + */ + nlim = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS); + /* Check whether we have minimal amount WQEs */ + if (unlikely(loc->wqe_free < + ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4))) + return MLX5_TXCMP_CODE_EXIT; + if (likely(pkts_n > 1)) + rte_prefetch0(*pkts); + loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m); + /* + * Build eMPW title WQEBB: + * - Control Segment, eMPW opcode, zero DS + * - Ethernet Segment, no inline + */ + mlx5_tx_cseg_init(txq, loc, loc->wqe_last, 0, + MLX5_OPCODE_ENHANCED_MPSW, olx); + mlx5_tx_eseg_none(txq, loc, loc->wqe_last, + olx & ~MLX5_TXOFF_CONFIG_VLAN); + eseg = &loc->wqe_last->eseg; + dseg = &loc->wqe_last->dseg[0]; + room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE, + loc->wqe_free) * MLX5_WQE_SIZE - + MLX5_WQE_CSEG_SIZE - + MLX5_WQE_ESEG_SIZE; + /* Build WQE till we have space, packets and resources. */ + part = room; + for (;;) { + uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf); + uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *); + unsigned int tlen; + + assert(room >= MLX5_WQE_DSEG_SIZE); + assert((room % MLX5_WQE_DSEG_SIZE) == 0); + assert((uintptr_t)dseg < (uintptr_t)txq->wqes_end); + /* + * Some Tx offloads may cause an error if + * packet is not long enough, check against + * assumed minimal length. + */ + if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) { + part -= room; + if (unlikely(!part)) + return MLX5_TXCMP_CODE_ERROR; + /* + * We have some successfully built + * packet Data Segments to send. + */ + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + return MLX5_TXCMP_CODE_ERROR; + } + /* Inline or not inline - that's the Question. */ + if (dlen > txq->inlen_empw) + goto pointer_empw; + /* Inline entire packet, optional VLAN insertion. */ + tlen = sizeof(dseg->bcount) + dlen; + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) { + /* + * The packet length must be checked in + * mlx5_tx_able_to_empw() and packet + * fits into inline length guaranteed. + */ + assert((dlen + sizeof(struct rte_vlan_hdr)) <= + txq->inlen_empw); + tlen += sizeof(struct rte_vlan_hdr); + if (room < tlen) + break; + dseg = mlx5_tx_dseg_vlan(txq, loc, dseg, + dptr, dlen, olx); +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + slen += sizeof(struct rte_vlan_hdr); +#endif + } else { + if (room < tlen) + break; + dseg = mlx5_tx_dseg_empw(txq, loc, dseg, + dptr, dlen, olx); + } + tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE); + assert(room >= tlen); + room -= tlen; + /* + * Packet data are completely inlined, + * free the packet immediately. + */ + rte_pktmbuf_free_seg(loc->mbuf); + goto next_mbuf; +pointer_empw: + /* + * Not inlinable VLAN packets are + * proceeded outside of this routine. + */ + assert(room >= MLX5_WQE_DSEG_SIZE); + if (MLX5_TXOFF_CONFIG(VLAN)) + assert(!(loc->mbuf->ol_flags & + PKT_TX_VLAN_PKT)); + mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx); + /* We have to store mbuf in elts.*/ + txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; + room -= MLX5_WQE_DSEG_SIZE; + /* Ring buffer wraparound is checked at the loop end.*/ + ++dseg; +next_mbuf: +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + slen += dlen; +#endif + loc->pkts_sent++; + loc->elts_free--; + pkts_n--; + if (unlikely(!pkts_n || !loc->elts_free)) { + /* + * We have no resources/packets to + * continue build descriptors. + */ + part -= room; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + return MLX5_TXCMP_CODE_EXIT; + } + loc->mbuf = *pkts++; + if (likely(pkts_n > 1)) + rte_prefetch0(*pkts); + ret = mlx5_tx_able_to_empw(txq, loc, olx, true); + /* + * Unroll the completion code to avoid + * returning variable value - it results in + * unoptimized sequent checking in caller. + */ + if (ret == MLX5_TXCMP_CODE_MULTI) { + part -= room; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_MULTI; + } + if (ret == MLX5_TXCMP_CODE_TSO) { + part -= room; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_TSO; + } + if (ret == MLX5_TXCMP_CODE_SINGLE) { + part -= room; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + return MLX5_TXCMP_CODE_SINGLE; + } + if (ret != MLX5_TXCMP_CODE_EMPW) { + assert(false); + part -= room; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + return MLX5_TXCMP_CODE_ERROR; + } + /* Check if we have minimal room left. */ + nlim--; + if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE)) + break; + /* + * Check whether packet parameters coincide + * within assumed eMPW batch: + * - check sum settings + * - metadata value + * - software parser settings + */ + if (!mlx5_tx_match_empw(txq, eseg, loc, olx)) + break; + /* Packet attributes match, continue the same eMPW. */ + if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) + dseg = (struct mlx5_wqe_dseg *)txq->wqes; + } + /* + * We get here to close an existing eMPW + * session and start the new one. + */ + assert(pkts_n); + part -= room; + if (unlikely(!part)) + return MLX5_TXCMP_CODE_EXIT; + mlx5_tx_idone_empw(txq, loc, part, slen, olx); + if (unlikely(!loc->elts_free || + !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + /* Continue the loop with new eMPW session. */ + } + assert(false); +} + +/** + * The routine sends packets with ordinary MLX5_OPCODE_SEND. + * Data inlining and VLAN insertion are supported. + */ +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + /* + * Subroutine is the part of mlx5_tx_burst_single() + * and sends single-segment packet with SEND opcode. + */ + assert(loc->elts_free && loc->wqe_free); + assert(pkts_n > loc->pkts_sent); + pkts += loc->pkts_sent + 1; + pkts_n -= loc->pkts_sent; + for (;;) { + struct mlx5_wqe *restrict wqe; + enum mlx5_txcmp_code ret; + + assert(NB_SEGS(loc->mbuf) == 1); + if (MLX5_TXOFF_CONFIG(INLINE)) { + unsigned int inlen, vlan = 0; + + inlen = rte_pktmbuf_data_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) { + vlan = sizeof(struct rte_vlan_hdr); + inlen += vlan; + static_assert((sizeof(struct rte_vlan_hdr) + + sizeof(struct rte_ether_hdr)) == + MLX5_ESEG_MIN_INLINE_SIZE, + "invalid min inline data size"); + } + /* + * If inlining is enabled at configuration time + * the limit must be not less than minimal size. + * Otherwise we would do extra check for data + * size to avoid crashes due to length overflow. + */ + assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE); + if (inlen <= txq->inlen_send) { + unsigned int seg_n, wqe_n; + + rte_prefetch0(rte_pktmbuf_mtod + (loc->mbuf, uint8_t *)); + /* Check against minimal length. */ + if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE) + return MLX5_TXCMP_CODE_ERROR; + /* + * Completely inlined packet data WQE: + * - Control Segment, SEND opcode + * - Ethernet Segment, no VLAN insertion + * - Data inlined, VLAN optionally inserted + * - Alignment to MLX5_WSEG_SIZE + * Have to estimate amount of WQEBBs + */ + seg_n = (inlen + 3 * MLX5_WSEG_SIZE - + MLX5_ESEG_MIN_INLINE_SIZE + + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE; + /* Check if there are enough WQEBBs. */ + wqe_n = (seg_n + 3) / 4; + if (wqe_n > loc->wqe_free) + return MLX5_TXCMP_CODE_EXIT; + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, seg_n, + MLX5_OPCODE_SEND, olx); + mlx5_tx_eseg_data(txq, loc, wqe, + vlan, inlen, 0, olx); + txq->wqe_ci += wqe_n; + loc->wqe_free -= wqe_n; + /* + * Packet data are completely inlined, + * free the packet immediately. + */ + rte_pktmbuf_free_seg(loc->mbuf); + } else if (!MLX5_TXOFF_CONFIG(EMPW) && + txq->inlen_mode) { + /* + * If minimal inlining is requested the eMPW + * feature should be disabled due to data is + * inlined into Ethernet Segment, which can + * not contain inlined data for eMPW due to + * segment shared for all packets. + */ + struct mlx5_wqe_dseg *restrict dseg; + unsigned int ds; + uint8_t *dptr; + + /* + * The inline-mode settings require + * to inline the specified amount of + * data bytes to the Ethernet Segment. + * We should check the free space in + * WQE ring buffer to inline partially. + */ + assert(txq->inlen_send >= txq->inlen_mode); + assert(inlen > txq->inlen_mode); + assert(txq->inlen_mode >= + MLX5_ESEG_MIN_INLINE_SIZE); + /* + * Check whether there are enough free WQEBBs: + * - Control Segment + * - Ethernet Segment + * - First Segment of inlined Ethernet data + * - ... data continued ... + * - Finishing Data Segment of pointer type + */ + ds = (MLX5_WQE_CSEG_SIZE + + MLX5_WQE_ESEG_SIZE + + MLX5_WQE_DSEG_SIZE + + txq->inlen_mode - + MLX5_ESEG_MIN_INLINE_SIZE + + MLX5_WQE_DSEG_SIZE + + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE; + if (loc->wqe_free < ((ds + 3) / 4)) + return MLX5_TXCMP_CODE_EXIT; + /* + * Build the ordinary SEND WQE: + * - Control Segment + * - Ethernet Segment, inline inlen_mode bytes + * - Data Segment of pointer type + */ + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, ds, + MLX5_OPCODE_SEND, olx); + dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, + txq->inlen_mode, + 0, olx); + dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + + txq->inlen_mode - vlan; + inlen -= txq->inlen_mode; + mlx5_tx_dseg_ptr(txq, loc, dseg, + dptr, inlen, olx); + /* + * WQE is built, update the loop parameters + * and got to the next packet. + */ + txq->wqe_ci += (ds + 3) / 4; + loc->wqe_free -= (ds + 3) / 4; + /* We have to store mbuf in elts.*/ + assert(MLX5_TXOFF_CONFIG(INLINE)); + txq->elts[txq->elts_head++ & txq->elts_m] = + loc->mbuf; + --loc->elts_free; + } else { + uint8_t *dptr; + unsigned int dlen; + + /* + * Partially inlined packet data WQE, we have + * some space in title WQEBB, we can fill it + * with some packet data. It takes one WQEBB, + * it is available, no extra space check: + * - Control Segment, SEND opcode + * - Ethernet Segment, no VLAN insertion + * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data + * - Data Segment, pointer type + * + * We also get here if VLAN insertion is not + * supported by HW, the inline is enabled. + */ + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, 4, + MLX5_OPCODE_SEND, olx); + mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx); + dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + + MLX5_ESEG_MIN_INLINE_SIZE - vlan; + /* + * The length check is performed above, by + * comparing with txq->inlen_send. We should + * not get overflow here. + */ + assert(inlen > MLX5_ESEG_MIN_INLINE_SIZE); + dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE; + mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1], + dptr, dlen, olx); + ++txq->wqe_ci; + --loc->wqe_free; + /* We have to store mbuf in elts.*/ + assert(MLX5_TXOFF_CONFIG(INLINE)); + txq->elts[txq->elts_head++ & txq->elts_m] = + loc->mbuf; + --loc->elts_free; + } +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += vlan + + rte_pktmbuf_data_len(loc->mbuf); +#endif + } else { + /* + * No inline at all, it means the CPU cycles saving + * is prioritized at configuration, we should not + * copy any packet data to WQE. + * + * SEND WQE, one WQEBB: + * - Control Segment, SEND opcode + * - Ethernet Segment, optional VLAN, no inline + * - Data Segment, pointer type + */ + wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); + loc->wqe_last = wqe; + mlx5_tx_cseg_init(txq, loc, wqe, 3, + MLX5_OPCODE_SEND, olx); + mlx5_tx_eseg_none(txq, loc, wqe, olx); + mlx5_tx_dseg_ptr + (txq, loc, &wqe->dseg[0], + rte_pktmbuf_mtod(loc->mbuf, uint8_t *), + rte_pktmbuf_data_len(loc->mbuf), olx); + ++txq->wqe_ci; + --loc->wqe_free; + /* + * We should not store mbuf pointer in elts + * if no inlining is configured, this is done + * by calling routine in a batch copy. + */ + assert(!MLX5_TXOFF_CONFIG(INLINE)); + --loc->elts_free; +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Update sent data bytes counter. */ + txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf); + if (MLX5_TXOFF_CONFIG(VLAN) && + loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) + txq->stats.obytes += + sizeof(struct rte_vlan_hdr); +#endif + } + ++loc->pkts_sent; + --pkts_n; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, loc, false, olx); + if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free)) + return MLX5_TXCMP_CODE_EXIT; + loc->mbuf = *pkts++; + if (pkts_n > 1) + rte_prefetch0(*pkts); + ret = mlx5_tx_able_to_empw(txq, loc, olx, true); + if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE)) + return ret; + } + assert(false); +} + +static __rte_always_inline enum mlx5_txcmp_code +mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + unsigned int pkts_n, + struct mlx5_txq_local *restrict loc, + unsigned int olx) +{ + enum mlx5_txcmp_code ret; + + ret = mlx5_tx_able_to_empw(txq, loc, olx, false); + if (ret == MLX5_TXCMP_CODE_SINGLE) + goto ordinary_send; + assert(ret == MLX5_TXCMP_CODE_EMPW); + for (;;) { + /* Optimize for inline/no inline eMPW send. */ + ret = (MLX5_TXOFF_CONFIG(INLINE)) ? + mlx5_tx_burst_empw_inline + (txq, pkts, pkts_n, loc, olx) : + mlx5_tx_burst_empw_simple + (txq, pkts, pkts_n, loc, olx); + if (ret != MLX5_TXCMP_CODE_SINGLE) + return ret; + /* The resources to send one packet should remain. */ + assert(loc->elts_free && loc->wqe_free); +ordinary_send: + ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx); + assert(ret != MLX5_TXCMP_CODE_SINGLE); + if (ret != MLX5_TXCMP_CODE_EMPW) + return ret; + /* The resources to send one packet should remain. */ + assert(loc->elts_free && loc->wqe_free); + } +} + +/** + * DPDK Tx callback template. This is configured template + * used to generate routines optimized for specified offload setup. + * One of this generated functions is chosen at SQ configuration + * time. + * + * @param txq + * Generic pointer to TX queue structure. + * @param[in] pkts + * Packets to transmit. + * @param pkts_n + * Number of packets in array. + * @param olx + * Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx + * values. Should be static to take compile time static configuration + * advantages. + * + * @return + * Number of packets successfully transmitted (<= pkts_n). + */ +static __rte_always_inline uint16_t +mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq, + struct rte_mbuf **restrict pkts, + uint16_t pkts_n, + unsigned int olx) +{ + struct mlx5_txq_local loc; + enum mlx5_txcmp_code ret; + unsigned int part; + + assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); + assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi)); + if (unlikely(!pkts_n)) + return 0; + loc.pkts_sent = 0; + loc.pkts_copy = 0; + loc.wqe_last = NULL; + +send_loop: + loc.pkts_loop = loc.pkts_sent; + /* + * Check if there are some CQEs, if any: + * - process an encountered errors + * - process the completed WQEs + * - free related mbufs + * - doorbell the NIC about processed CQEs + */ + rte_prefetch0(*(pkts + loc.pkts_sent)); + mlx5_tx_handle_completion(txq, olx); + /* + * Calculate the number of available resources - elts and WQEs. + * There are two possible different scenarios: + * - no data inlining into WQEs, one WQEBB may contains upto + * four packets, in this case elts become scarce resource + * - data inlining into WQEs, one packet may require multiple + * WQEBBs, the WQEs become the limiting factor. + */ + assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); + loc.elts_free = txq->elts_s - + (uint16_t)(txq->elts_head - txq->elts_tail); + assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi)); + loc.wqe_free = txq->wqe_s - + (uint16_t)(txq->wqe_ci - txq->wqe_pi); + if (unlikely(!loc.elts_free || !loc.wqe_free)) + goto burst_exit; + for (;;) { + /* + * Fetch the packet from array. Usually this is + * the first packet in series of multi/single + * segment packets. + */ + loc.mbuf = *(pkts + loc.pkts_sent); + /* Dedicated branch for multi-segment packets. */ + if (MLX5_TXOFF_CONFIG(MULTI) && + unlikely(NB_SEGS(loc.mbuf) > 1)) { + /* + * Multi-segment packet encountered. + * Hardware is able to process it only + * with SEND/TSO opcodes, one packet + * per WQE, do it in dedicated routine. + */ +enter_send_multi: + assert(loc.pkts_sent >= loc.pkts_copy); + part = loc.pkts_sent - loc.pkts_copy; + if (!MLX5_TXOFF_CONFIG(INLINE) && part) { + /* + * There are some single-segment mbufs not + * stored in elts. The mbufs must be in the + * same order as WQEs, so we must copy the + * mbufs to elts here, before the coming + * multi-segment packet mbufs is appended. + */ + mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, + part, olx); + loc.pkts_copy = loc.pkts_sent; + } + assert(pkts_n > loc.pkts_sent); + ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx); + if (!MLX5_TXOFF_CONFIG(INLINE)) + loc.pkts_copy = loc.pkts_sent; + /* + * These returned code checks are supposed + * to be optimized out due to routine inlining. + */ + if (ret == MLX5_TXCMP_CODE_EXIT) { + /* + * The routine returns this code when + * all packets are sent or there is no + * enough resources to complete request. + */ + break; + } + if (ret == MLX5_TXCMP_CODE_ERROR) { + /* + * The routine returns this code when + * some error in the incoming packets + * format occurred. + */ + txq->stats.oerrors++; + break; + } + if (ret == MLX5_TXCMP_CODE_SINGLE) { + /* + * The single-segment packet was encountered + * in the array, try to send it with the + * best optimized way, possible engaging eMPW. + */ + goto enter_send_single; + } + if (MLX5_TXOFF_CONFIG(TSO) && + ret == MLX5_TXCMP_CODE_TSO) { + /* + * The single-segment TSO packet was + * encountered in the array. + */ + goto enter_send_tso; + } + /* We must not get here. Something is going wrong. */ + assert(false); + txq->stats.oerrors++; + break; + } + /* Dedicated branch for single-segment TSO packets. */ + if (MLX5_TXOFF_CONFIG(TSO) && + unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) { + /* + * TSO might require special way for inlining + * (dedicated parameters) and is sent with + * MLX5_OPCODE_TSO opcode only, provide this + * in dedicated branch. + */ +enter_send_tso: + assert(NB_SEGS(loc.mbuf) == 1); + assert(pkts_n > loc.pkts_sent); + ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx); + /* + * These returned code checks are supposed + * to be optimized out due to routine inlining. + */ + if (ret == MLX5_TXCMP_CODE_EXIT) + break; + if (ret == MLX5_TXCMP_CODE_ERROR) { + txq->stats.oerrors++; + break; + } + if (ret == MLX5_TXCMP_CODE_SINGLE) + goto enter_send_single; + if (MLX5_TXOFF_CONFIG(MULTI) && + ret == MLX5_TXCMP_CODE_MULTI) { + /* + * The multi-segment packet was + * encountered in the array. + */ + goto enter_send_multi; + } + /* We must not get here. Something is going wrong. */ + assert(false); + txq->stats.oerrors++; + break; + } + /* + * The dedicated branch for the single-segment packets + * without TSO. Often these ones can be sent using + * MLX5_OPCODE_EMPW with multiple packets in one WQE. + * The routine builds the WQEs till it encounters + * the TSO or multi-segment packet (in case if these + * offloads are requested at SQ configuration time). + */ +enter_send_single: + assert(pkts_n > loc.pkts_sent); + ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx); + /* + * These returned code checks are supposed + * to be optimized out due to routine inlining. + */ + if (ret == MLX5_TXCMP_CODE_EXIT) + break; + if (ret == MLX5_TXCMP_CODE_ERROR) { + txq->stats.oerrors++; + break; + } + if (MLX5_TXOFF_CONFIG(MULTI) && + ret == MLX5_TXCMP_CODE_MULTI) { + /* + * The multi-segment packet was + * encountered in the array. + */ + goto enter_send_multi; + } + if (MLX5_TXOFF_CONFIG(TSO) && + ret == MLX5_TXCMP_CODE_TSO) { + /* + * The single-segment TSO packet was + * encountered in the array. + */ + goto enter_send_tso; + } + /* We must not get here. Something is going wrong. */ + assert(false); + txq->stats.oerrors++; + break; + } + /* + * Main Tx loop is completed, do the rest: + * - set completion request if thresholds are reached + * - doorbell the hardware + * - copy the rest of mbufs to elts (if any) + */ + assert(MLX5_TXOFF_CONFIG(INLINE) || loc.pkts_sent >= loc.pkts_copy); + /* Take a shortcut if nothing is sent. */ + if (unlikely(loc.pkts_sent == loc.pkts_loop)) + goto burst_exit; + /* + * Ring QP doorbell immediately after WQE building completion + * to improve latencies. The pure software related data treatment + * can be completed after doorbell. Tx CQEs for this SQ are + * processed in this thread only by the polling. + */ + mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, 0); + /* Not all of the mbufs may be stored into elts yet. */ + part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy; + if (!MLX5_TXOFF_CONFIG(INLINE) && part) { + /* + * There are some single-segment mbufs not stored in elts. + * It can be only if the last packet was single-segment. + * The copying is gathered into one place due to it is + * a good opportunity to optimize that with SIMD. + * Unfortunately if inlining is enabled the gaps in + * pointer array may happen due to early freeing of the + * inlined mbufs. + */ + mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx); + loc.pkts_copy = loc.pkts_sent; + } + assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); + assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi)); + if (pkts_n > loc.pkts_sent) { + /* + * If burst size is large there might be no enough CQE + * fetched from completion queue and no enough resources + * freed to send all the packets. + */ + goto send_loop; + } +burst_exit: +#ifdef MLX5_PMD_SOFT_COUNTERS + /* Increment sent packets counter. */ + txq->stats.opackets += loc.pkts_sent; +#endif + return loc.pkts_sent; +} + +/* Generate routines with Enhanced Multi-Packet Write support. */ +MLX5_TXOFF_DECL(full_empw, + MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(none_empw, + MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(md_empw, + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(mt_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(mtsc_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(mti_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(mtv_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(mtiv_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(sc_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(sci_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(scv_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(sciv_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(i_empw, + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(v_empw, + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_DECL(iv_empw, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +/* Generate routines without Enhanced Multi-Packet Write support. */ +MLX5_TXOFF_DECL(full, + MLX5_TXOFF_CONFIG_FULL) + +MLX5_TXOFF_DECL(none, + MLX5_TXOFF_CONFIG_NONE) + +MLX5_TXOFF_DECL(md, + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(mt, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(mtsc, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(mti, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + + +MLX5_TXOFF_DECL(mtv, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + + +MLX5_TXOFF_DECL(mtiv, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(sc, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(sci, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + + +MLX5_TXOFF_DECL(scv, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + + +MLX5_TXOFF_DECL(sciv, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(i, + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(v, + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_DECL(iv, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +/* + * Array of declared and compiled Tx burst function and corresponding + * supported offloads set. The array is used to select the Tx burst + * function for specified offloads set at Tx queue configuration time. + */ +const struct { + eth_tx_burst_t func; + unsigned int olx; +} txoff_func[] = { +MLX5_TXOFF_INFO(full_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(none_empw, + MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(md_empw, + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(mt_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(mtsc_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(mti_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(mtv_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(mtiv_empw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(sc_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(sci_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(scv_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(sciv_empw, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(i_empw, + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(v_empw, + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(iv_empw, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW) + +MLX5_TXOFF_INFO(full, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(none, + MLX5_TXOFF_CONFIG_NONE) + +MLX5_TXOFF_INFO(md, + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(mt, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(mtsc, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(mti, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + + +MLX5_TXOFF_INFO(mtv, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(mtiv, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(sc, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(sci, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(scv, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(sciv, + MLX5_TXOFF_CONFIG_SWP | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(i, + MLX5_TXOFF_CONFIG_INLINE | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(v, + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(iv, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA) +}; + +/** + * Configure the Tx function to use. The routine checks configured + * Tx offloads for the device and selects appropriate Tx burst + * routine. There are multiple Tx burst routines compiled from + * the same template in the most optimal way for the dedicated + * Tx offloads set. + * + * @param dev + * Pointer to private data structure. + * + * @return + * Pointer to selected Tx burst function. + */ +eth_tx_burst_t +mlx5_select_tx_function(struct rte_eth_dev *dev) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_dev_config *config = &priv->config; + uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads; + unsigned int diff = 0, olx = 0, i, m; + + static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <= + MLX5_DSEG_MAX, "invalid WQE max size"); + static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE, + "invalid WQE Control Segment size"); + static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE, + "invalid WQE Ethernet Segment size"); + static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE, + "invalid WQE Data Segment size"); + static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE, + "invalid WQE size"); + assert(priv); + if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) { + /* We should support Multi-Segment Packets. */ + olx |= MLX5_TXOFF_CONFIG_MULTI; + } + if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO | + DEV_TX_OFFLOAD_VXLAN_TNL_TSO | + DEV_TX_OFFLOAD_GRE_TNL_TSO | + DEV_TX_OFFLOAD_IP_TNL_TSO | + DEV_TX_OFFLOAD_UDP_TNL_TSO)) { + /* We should support TCP Send Offload. */ + olx |= MLX5_TXOFF_CONFIG_TSO; + } + if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO | + DEV_TX_OFFLOAD_UDP_TNL_TSO | + DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) { + /* We should support Software Parser for Tunnels. */ + olx |= MLX5_TXOFF_CONFIG_SWP; + } + if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM | + DEV_TX_OFFLOAD_UDP_CKSUM | + DEV_TX_OFFLOAD_TCP_CKSUM | + DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) { + /* We should support IP/TCP/UDP Checksums. */ + olx |= MLX5_TXOFF_CONFIG_CSUM; + } + if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) { + /* We should support VLAN insertion. */ + olx |= MLX5_TXOFF_CONFIG_VLAN; + } + if (priv->txqs_n && (*priv->txqs)[0]) { + struct mlx5_txq_data *txd = (*priv->txqs)[0]; + + if (txd->inlen_send) { + /* + * Check the data inline requirements. Data inline + * is enabled on per device basis, we can check + * the first Tx queue only. + * + * If device does not support VLAN insertion in WQE + * and some queues are requested to perform VLAN + * insertion offload than inline must be enabled. + */ + olx |= MLX5_TXOFF_CONFIG_INLINE; + } + } + if (config->mps == MLX5_MPW_ENHANCED && + config->txq_inline_min <= 0) { + /* + * The NIC supports Enhanced Multi-Packet Write. + * We do not support legacy MPW due to its + * hardware related problems, so we just ignore + * legacy MLX5_MPW settings. There should be no + * minimal required inline data. + */ + olx |= MLX5_TXOFF_CONFIG_EMPW; + } + if (rte_flow_dynf_metadata_avail()) { + /* We should support Flow metadata. */ + olx |= MLX5_TXOFF_CONFIG_METADATA; + } + /* + * Scan the routines table to find the minimal + * satisfying routine with requested offloads. + */ + m = RTE_DIM(txoff_func); + for (i = 0; i < RTE_DIM(txoff_func); i++) { + unsigned int tmp; + + tmp = txoff_func[i].olx; + if (tmp == olx) { + /* Meets requested offloads exactly.*/ + m = i; + break; + } + if ((tmp & olx) != olx) { + /* Does not meet requested offloads at all. */ + continue; + } + if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW) + /* Do not enable eMPW if not configured. */ + continue; + if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE) + /* Do not enable inlining if not configured. */ + continue; + /* + * Some routine meets the requirements. + * Check whether it has minimal amount + * of not requested offloads. + */ + tmp = __builtin_popcountl(tmp & ~olx); + if (m >= RTE_DIM(txoff_func) || tmp < diff) { + /* First or better match, save and continue. */ + m = i; + diff = tmp; + continue; + } + if (tmp == diff) { + tmp = txoff_func[i].olx ^ txoff_func[m].olx; + if (__builtin_ffsl(txoff_func[i].olx & ~tmp) < + __builtin_ffsl(txoff_func[m].olx & ~tmp)) { + /* Lighter not requested offload. */ + m = i; + } + } + } + if (m >= RTE_DIM(txoff_func)) { + DRV_LOG(DEBUG, "port %u has no selected Tx function" + " for requested offloads %04X", + dev->data->port_id, olx); + return NULL; + } + DRV_LOG(DEBUG, "port %u has selected Tx function" + " supporting offloads %04X/%04X", + dev->data->port_id, olx, txoff_func[m].olx); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI) + DRV_LOG(DEBUG, "\tMULTI (multi segment)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO) + DRV_LOG(DEBUG, "\tTSO (TCP send offload)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP) + DRV_LOG(DEBUG, "\tSWP (software parser)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM) + DRV_LOG(DEBUG, "\tCSUM (checksum offload)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE) + DRV_LOG(DEBUG, "\tINLIN (inline data)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN) + DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA) + DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) + DRV_LOG(DEBUG, "\tEMPW (Enhanced MPW)"); + return txoff_func[m].func; +}