X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fmlx5%2Fmlx5_rxtx.c;h=a21201038ca0273362e14e4ed29633ee87cfe203;hb=611faa5f46cc67449f272e14450fc6a0a275767d;hp=c7be532b94a277e4ab0b778267d34601580be469;hpb=318ea4cfa1b1c1393181a0a70ba8fd9776de341a;p=dpdk.git diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c index c7be532b94..a21201038c 100644 --- a/drivers/net/mlx5/mlx5_rxtx.c +++ b/drivers/net/mlx5/mlx5_rxtx.c @@ -3,7 +3,6 @@ * Copyright 2015-2019 Mellanox Technologies, Ltd */ -#include #include #include #include @@ -26,13 +25,18 @@ #include #include #include +#include +#include +#include +#include + +#include "mlx5_defs.h" #include "mlx5.h" +#include "mlx5_mr.h" #include "mlx5_utils.h" #include "mlx5_rxtx.h" #include "mlx5_autoconf.h" -#include "mlx5_defs.h" -#include "mlx5_prm.h" /* TX burst subroutines return codes. */ enum mlx5_txcmp_code { @@ -61,6 +65,7 @@ enum mlx5_txcmp_code { #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.*/ +#define MLX5_TXOFF_CONFIG_MPW (1u << 9) /* Legacy MPW supported.*/ /* The most common offloads groups. */ #define MLX5_TXOFF_CONFIG_NONE 0 @@ -124,6 +129,9 @@ uint32_t mlx5_ptype_table[] __rte_cache_aligned = { uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned; uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned; +uint64_t rte_net_mlx5_dynf_inline_mask; +#define PKT_TX_DYNF_NOINLINE rte_net_mlx5_dynf_inline_mask + /** * Build a table to translate Rx completion flags to packet type. * @@ -529,6 +537,89 @@ mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset) return RTE_ETH_RX_DESC_AVAIL; } +/** + * DPDK callback to get the RX queue information + * + * @param dev + * Pointer to the device structure. + * + * @param rx_queue_id + * Rx queue identificator. + * + * @param qinfo + * Pointer to the RX queue information structure. + * + * @return + * None. + */ + +void +mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id, + struct rte_eth_rxq_info *qinfo) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_rxq_data *rxq = (*priv->rxqs)[rx_queue_id]; + struct mlx5_rxq_ctrl *rxq_ctrl = + container_of(rxq, struct mlx5_rxq_ctrl, rxq); + + if (!rxq) + return; + qinfo->mp = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ? + rxq->mprq_mp : rxq->mp; + qinfo->conf.rx_thresh.pthresh = 0; + qinfo->conf.rx_thresh.hthresh = 0; + qinfo->conf.rx_thresh.wthresh = 0; + qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh; + qinfo->conf.rx_drop_en = 1; + qinfo->conf.rx_deferred_start = rxq_ctrl ? 0 : 1; + qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads; + qinfo->scattered_rx = dev->data->scattered_rx; + qinfo->nb_desc = 1 << rxq->elts_n; +} + +/** + * DPDK callback to get the RX packet burst mode information + * + * @param dev + * Pointer to the device structure. + * + * @param rx_queue_id + * Rx queue identificatior. + * + * @param mode + * Pointer to the burts mode information. + * + * @return + * 0 as success, -EINVAL as failure. + */ + +int +mlx5_rx_burst_mode_get(struct rte_eth_dev *dev, + uint16_t rx_queue_id __rte_unused, + struct rte_eth_burst_mode *mode) +{ + eth_rx_burst_t pkt_burst = dev->rx_pkt_burst; + + if (pkt_burst == mlx5_rx_burst) { + snprintf(mode->info, sizeof(mode->info), "%s", "Scalar"); + } else if (pkt_burst == mlx5_rx_burst_mprq) { + snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ"); + } else if (pkt_burst == mlx5_rx_burst_vec) { +#if defined RTE_ARCH_X86_64 + snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE"); +#elif defined RTE_ARCH_ARM64 + snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon"); +#elif defined RTE_ARCH_PPC_64 + snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec"); +#else + return -EINVAL; +#endif + } else { + return -EINVAL; + } + return 0; +} + /** * DPDK callback to get the number of used descriptors in a RX queue * @@ -583,18 +674,16 @@ 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); @@ -654,9 +743,10 @@ 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, otherwise + * the error completion entry is handled successfully. */ -uint16_t +static int mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq, volatile struct mlx5_err_cqe *err_cqe) { @@ -700,17 +790,14 @@ mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq, */ txq->stats.oerrors += ((txq->wqe_ci & wqe_m) - new_wqe_pi) & wqe_m; - if (tx_recover_qp(txq_ctrl) == 0) { - txq->cq_ci++; - /* Release all the remaining buffers. */ - return txq->elts_head; + if (tx_recover_qp(txq_ctrl)) { + /* Recovering failed - retry later on the same WQE. */ + return -1; } - /* Recovering failed - try again later on the same WQE. */ - } else { - txq->cq_ci++; + /* Release all the remaining buffers. */ + txq_free_elts(txq_ctrl); } - /* Do not release buffers. */ - return txq->elts_tail; + return 0; } /** @@ -780,7 +867,7 @@ mlx5_rxq_initialize(struct mlx5_rxq_data *rxq) byte_count = DATA_LEN(buf); } /* scat->addr must be able to store a pointer. */ - assert(sizeof(scat->addr) >= sizeof(uintptr_t)); + MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t)); *scat = (struct mlx5_wqe_data_seg){ .addr = rte_cpu_to_be_64(addr), .byte_count = rte_cpu_to_be_32(byte_count), @@ -848,7 +935,7 @@ mlx5_queue_state_modify_primary(struct rte_eth_dev *dev, &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; @@ -861,12 +948,12 @@ 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; } @@ -874,7 +961,7 @@ 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; @@ -882,7 +969,7 @@ 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; @@ -890,7 +977,7 @@ 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; @@ -914,6 +1001,7 @@ static int mlx5_queue_state_modify(struct rte_eth_dev *dev, struct mlx5_mp_arg_queue_state_modify *sm) { + struct mlx5_priv *priv = dev->data->dev_private; int ret = 0; switch (rte_eal_process_type()) { @@ -921,7 +1009,7 @@ mlx5_queue_state_modify(struct rte_eth_dev *dev, ret = mlx5_queue_state_modify_primary(dev, sm); break; case RTE_PROC_SECONDARY: - ret = mlx5_mp_req_queue_state_modify(dev, sm); + ret = mlx5_mp_req_queue_state_modify(&priv->mp_id, sm); break; default: break; @@ -938,14 +1026,15 @@ 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; @@ -1012,7 +1101,7 @@ 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; @@ -1036,6 +1125,16 @@ 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; @@ -1240,6 +1339,11 @@ rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt, pkt->hash.fdir.hi = mlx5_flow_mark_get(mark); } } + if (rxq->dynf_meta && cqe->flow_table_metadata) { + pkt->ol_flags |= rxq->flow_meta_mask; + *RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset, uint32_t *) = + cqe->flow_table_metadata; + } if (rxq->csum) pkt->ol_flags |= rxq_cq_to_ol_flags(cqe); if (rxq->vlan_strip && @@ -1306,7 +1410,7 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) break; } while (pkt != seg) { - assert(pkt != (*rxq->elts)[idx]); + MLX5_ASSERT(pkt != (*rxq->elts)[idx]); rep = NEXT(pkt); NEXT(pkt) = NULL; NB_SEGS(pkt) = 1; @@ -1323,8 +1427,8 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) break; } pkt = seg; - assert(len >= (rxq->crc_present << 2)); - pkt->ol_flags = 0; + MLX5_ASSERT(len >= (rxq->crc_present << 2)); + pkt->ol_flags &= EXT_ATTACHED_MBUF; /* If compressed, take hash result from mini-CQE. */ rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ? cqe->rx_hash_res : @@ -1514,7 +1618,7 @@ mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx, &((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg; void *addr; - assert(rep != NULL); + MLX5_ASSERT(rep != NULL); /* Replace MPRQ buf. */ (*rxq->mprq_bufs)[rq_idx] = rep; /* Replace WQE. */ @@ -1557,21 +1661,20 @@ 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) { struct rte_mbuf *pkt; void *addr; int ret; - unsigned int len; + uint32_t len; uint16_t strd_cnt; uint16_t strd_idx; uint32_t offset; uint32_t byte_cnt; + int32_t hdrm_overlap; 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. */ @@ -1604,7 +1707,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) byte_cnt = ret; strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >> MLX5_MPRQ_STRIDE_NUM_SHIFT; - assert(strd_cnt); + MLX5_ASSERT(strd_cnt); consumed_strd += strd_cnt; if (byte_cnt & MLX5_MPRQ_FILLER_MASK) continue; @@ -1615,48 +1718,71 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) /* mini-CQE for MPRQ doesn't have hash result. */ strd_idx = rte_be_to_cpu_16(mcqe->stride_idx); } - 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 - * consume multiple strides without raising an error. In this - * case, the packet should be dropped because it is bigger than - * the max_rx_pkt_len. - */ - if (unlikely(!lro_num_seg && strd_cnt > 1)) { - ++rxq->stats.idropped; - continue; - } + MLX5_ASSERT(strd_idx < strd_n); + MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) & + wq_mask)); pkt = rte_pktmbuf_alloc(rxq->mp); if (unlikely(pkt == NULL)) { ++rxq->stats.rx_nombuf; break; } len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT; - assert((int)len >= (rxq->crc_present << 2)); + MLX5_ASSERT((int)len >= (rxq->crc_present << 2)); 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, strd_n), offset); + hdrm_overlap = len + RTE_PKTMBUF_HEADROOM - strd_cnt * strd_sz; /* * Memcpy packets to the target mbuf if: * - The size of packet is smaller than mprq_max_memcpy_len. * - Out of buffer in the Mempool for Multi-Packet RQ. + * - The packet's stride overlaps a headroom and scatter is off. */ - if (len <= rxq->mprq_max_memcpy_len || rxq->mprq_repl == NULL) { - /* - * When memcpy'ing packet due to out-of-buffer, the - * packet must be smaller than the target mbuf. - */ - if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) { + if (len <= rxq->mprq_max_memcpy_len || + rxq->mprq_repl == NULL || + (hdrm_overlap > 0 && !rxq->strd_scatter_en)) { + if (likely(rte_pktmbuf_tailroom(pkt) >= len)) { + rte_memcpy(rte_pktmbuf_mtod(pkt, void *), + addr, len); + DATA_LEN(pkt) = len; + } else if (rxq->strd_scatter_en) { + struct rte_mbuf *prev = pkt; + uint32_t seg_len = + RTE_MIN(rte_pktmbuf_tailroom(pkt), len); + uint32_t rem_len = len - seg_len; + + rte_memcpy(rte_pktmbuf_mtod(pkt, void *), + addr, seg_len); + DATA_LEN(pkt) = seg_len; + while (rem_len) { + struct rte_mbuf *next = + rte_pktmbuf_alloc(rxq->mp); + + if (unlikely(next == NULL)) { + rte_pktmbuf_free(pkt); + ++rxq->stats.rx_nombuf; + goto out; + } + NEXT(prev) = next; + SET_DATA_OFF(next, 0); + addr = RTE_PTR_ADD(addr, seg_len); + seg_len = RTE_MIN + (rte_pktmbuf_tailroom(next), + rem_len); + rte_memcpy + (rte_pktmbuf_mtod(next, void *), + addr, seg_len); + DATA_LEN(next) = seg_len; + rem_len -= seg_len; + prev = next; + ++NB_SEGS(pkt); + } + } else { rte_pktmbuf_free_seg(pkt); ++rxq->stats.idropped; 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; @@ -1665,9 +1791,9 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) /* 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); - buf_addr = RTE_PTR_SUB(addr, headroom_sz); + MLX5_ASSERT((uint16_t)rte_atomic16_read(&buf->refcnt) <= + strd_n + 1); + buf_addr = RTE_PTR_SUB(addr, RTE_PKTMBUF_HEADROOM); /* * MLX5 device doesn't use iova but it is necessary in a * case where the Rx packet is transmitted via a @@ -1686,43 +1812,42 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) 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 - * kernel interface. - */ - if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) { - rte_pktmbuf_free_seg(pkt); - ++rxq->stats.idropped; - continue; - } + SET_DATA_OFF(pkt, RTE_PKTMBUF_HEADROOM); + MLX5_ASSERT(pkt->ol_flags == EXT_ATTACHED_MBUF); + MLX5_ASSERT(rte_pktmbuf_tailroom(pkt) < + len - (hdrm_overlap > 0 ? hdrm_overlap : 0)); 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. + * Copy the last fragment of a packet (up to headroom + * size bytes) in case there is a stride overlap with + * a next packet's headroom. Allocate a separate mbuf + * to store this fragment and link it. Scatter is on. */ - if (!rxq->strd_headroom_en) { - struct rte_mbuf *headroom_mbuf = - rte_pktmbuf_alloc(rxq->mp); + if (hdrm_overlap > 0) { + MLX5_ASSERT(rxq->strd_scatter_en); + struct rte_mbuf *seg = + rte_pktmbuf_alloc(rxq->mp); - if (unlikely(headroom_mbuf == NULL)) { + if (unlikely(seg == NULL)) { rte_pktmbuf_free_seg(pkt); ++rxq->stats.rx_nombuf; break; } - PORT(pkt) = rxq->port_id; - NEXT(headroom_mbuf) = pkt; - pkt = headroom_mbuf; + SET_DATA_OFF(seg, 0); + rte_memcpy(rte_pktmbuf_mtod(seg, void *), + RTE_PTR_ADD(addr, len - hdrm_overlap), + hdrm_overlap); + DATA_LEN(seg) = hdrm_overlap; + DATA_LEN(pkt) = len - hdrm_overlap; + NEXT(pkt) = seg; NB_SEGS(pkt) = 2; } } rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res); - if (lro_num_seg > 1) { + if (cqe->lro_num_seg > 1) { mlx5_lro_update_hdr(addr, cqe, len); pkt->ol_flags |= PKT_RX_LRO; - pkt->tso_segsz = strd_sz; + pkt->tso_segsz = len / cqe->lro_num_seg; } PKT_LEN(pkt) = len; PORT(pkt) = rxq->port_id; @@ -1734,6 +1859,7 @@ mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n) *(pkts++) = pkt; ++i; } +out: /* Update the consumer indexes. */ rxq->consumed_strd = consumed_strd; rte_cio_wmb(); @@ -1853,8 +1979,8 @@ mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts, * copying pointers to temporary array * for rte_mempool_put_bulk() calls. */ - assert(pkts); - assert(pkts_n); + MLX5_ASSERT(pkts); + MLX5_ASSERT(pkts_n); for (;;) { for (;;) { /* @@ -1863,7 +1989,7 @@ mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts, */ mbuf = rte_pktmbuf_prefree_seg(*pkts); if (likely(mbuf != NULL)) { - assert(mbuf == *pkts); + MLX5_ASSERT(mbuf == *pkts); if (likely(n_free != 0)) { if (unlikely(pool != mbuf->pool)) /* From different pool. */ @@ -1900,9 +2026,9 @@ mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts, * This loop is implemented to avoid multiple * inlining of rte_mempool_put_bulk(). */ - assert(pool); - assert(p_free); - assert(n_free); + MLX5_ASSERT(pool); + MLX5_ASSERT(p_free); + MLX5_ASSERT(n_free); /* * Free the array of pre-freed mbufs * belonging to the same memory pool. @@ -1950,8 +2076,8 @@ mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq, { uint16_t n_elts = tail - txq->elts_tail; - assert(n_elts); - assert(n_elts <= txq->elts_s); + MLX5_ASSERT(n_elts); + MLX5_ASSERT(n_elts <= txq->elts_s); /* * Implement a loop to support ring buffer wraparound * with single inlining of mlx5_tx_free_mbuf(). @@ -1961,8 +2087,8 @@ mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq, part = txq->elts_s - (txq->elts_tail & txq->elts_m); part = RTE_MIN(part, n_elts); - assert(part); - assert(part <= txq->elts_s); + MLX5_ASSERT(part); + MLX5_ASSERT(part <= txq->elts_s); mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m], part, olx); txq->elts_tail += part; @@ -1993,11 +2119,11 @@ mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq, unsigned int part; struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts; - assert(pkts); - assert(pkts_n); + MLX5_ASSERT(pkts); + MLX5_ASSERT(pkts_n); part = txq->elts_s - (txq->elts_head & txq->elts_m); - assert(part); - assert(part <= txq->elts_s); + MLX5_ASSERT(part); + MLX5_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, @@ -2009,6 +2135,35 @@ mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq, (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 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, + unsigned int olx __rte_unused) +{ + if (likely(last_cqe != NULL)) { + uint16_t tail; + + txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter); + tail = txq->fcqs[(txq->cq_ci - 1) & txq->cqe_m]; + if (likely(tail != txq->elts_tail)) { + mlx5_tx_free_elts(txq, tail, olx); + MLX5_ASSERT(tail == txq->elts_tail); + } + } +} + /** * Manage TX completions. This routine checks the CQ for * arrived CQEs, deduces the last accomplished WQE in SQ, @@ -2028,10 +2183,12 @@ mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq, unsigned int olx __rte_unused) { unsigned int count = MLX5_TX_COMP_MAX_CQE; - bool update = false; - uint16_t tail = txq->elts_tail; + volatile struct mlx5_cqe *last_cqe = NULL; + bool ring_doorbell = false; 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; @@ -2040,54 +2197,60 @@ mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq, 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); + MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN); break; } - /* Some error occurred, try to restart. */ + /* + * 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(); - tail = mlx5_tx_error_cqe_handle + ret = mlx5_tx_error_cqe_handle (txq, (volatile struct mlx5_err_cqe *)cqe); - if (likely(tail != txq->elts_tail)) { - mlx5_tx_free_elts(txq, tail, olx); - assert(tail == txq->elts_tail); + if (unlikely(ret < 0)) { + /* + * Some error occurred on queue error + * handling, we do not advance the index + * here, allowing to retry on next call. + */ + return; } - /* Allow flushing all CQEs from the queue. */ - count = txq->cqe_s; - } else { - volatile struct mlx5_wqe_cseg *cseg; - - /* Normal transmit completion. */ + /* + * We are going to fetch all entries with + * MLX5_CQE_SYNDROME_WR_FLUSH_ERR status. + * The send queue is supposed to be empty. + */ + ring_doorbell = true; ++txq->cq_ci; - rte_cio_rmb(); - txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter); - cseg = (volatile struct mlx5_wqe_cseg *) - (txq->wqes + (txq->wqe_pi & txq->wqe_m)); - tail = cseg->misc; + txq->cq_pi = txq->cq_ci; + last_cqe = NULL; + continue; } -#ifndef NDEBUG - if (txq->cq_pi) - --txq->cq_pi; -#endif - update = true; - /* - * 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); - if (likely(tail != txq->elts_tail)) { - /* Free data buffers from elts. */ - mlx5_tx_free_elts(txq, tail, olx); - assert(tail == txq->elts_tail); - } - if (likely(update)) { - /* Update the consumer index. */ + /* Normal transmit completion. */ + MLX5_ASSERT(txq->cq_ci != txq->cq_pi); + MLX5_ASSERT((txq->fcqs[txq->cq_ci & txq->cqe_m] >> 16) == + cqe->wqe_counter); + ring_doorbell = true; + ++txq->cq_ci; + last_cqe = cqe; + /* + * 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. + */ + if (likely(--count == 0)) + break; + } while (true); + if (likely(ring_doorbell)) { + /* Ring doorbell to notify hardware. */ rte_compiler_barrier(); - *txq->cq_db = - rte_cpu_to_be_32(txq->cq_ci); + *txq->cq_db = rte_cpu_to_be_32(txq->cq_ci); + mlx5_tx_comp_flush(txq, last_cqe, olx); } } @@ -2098,8 +2261,6 @@ mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq, * * @param txq * Pointer to TX queue structure. - * @param n_mbuf - * Number of mbuf not stored yet in elts array. * @param loc * Pointer to burst routine local context. * @param olx @@ -2108,30 +2269,36 @@ mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq, */ static __rte_always_inline void mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq, - unsigned int n_mbuf, struct mlx5_txq_local *restrict loc, - unsigned int olx __rte_unused) + unsigned int olx) { - uint16_t head = txq->elts_head + n_mbuf; + uint16_t head = txq->elts_head; + unsigned int part; + part = MLX5_TXOFF_CONFIG(INLINE) ? + 0 : loc->pkts_sent - loc->pkts_copy; + head += part; if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH || - (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres) { + (MLX5_TXOFF_CONFIG(INLINE) && + (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) { volatile struct mlx5_wqe *last = loc->wqe_last; + MLX5_ASSERT(last); txq->elts_comp = head; - txq->wqe_comp = txq->wqe_ci; + 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); + /* Save elts_head in dedicated free on completion queue. */ +#ifdef RTE_LIBRTE_MLX5_DEBUG + txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head | + (last->cseg.opcode >> 8) << 16; +#else + txq->fcqs[txq->cq_pi++ & txq->cqe_m] = head; +#endif + /* A CQE slot must always be available. */ + MLX5_ASSERT((txq->cq_pi - txq->cq_ci) <= txq->cqe_s); } } @@ -2189,6 +2356,9 @@ mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq, { struct mlx5_wqe_cseg *restrict cs = &wqe->cseg; + /* For legacy MPW replace the EMPW by TSO with modifier. */ + if (MLX5_TXOFF_CONFIG(MPW) && opcode == MLX5_OPCODE_ENHANCED_MPSW) + opcode = MLX5_OPCODE_TSO | MLX5_OPC_MOD_MPW << 24; 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 << @@ -2233,8 +2403,8 @@ mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused, 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_METADATA ? - loc->mbuf->tx_metadata : 0 : 0; + 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) { @@ -2242,7 +2412,7 @@ mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused, * We should get here only if device support * this feature correctly. */ - assert(txq->vlan_en); + MLX5_ASSERT(txq->vlan_en); es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT | loc->mbuf->vlan_tci); } else { @@ -2293,8 +2463,8 @@ mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused, 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_METADATA ? - loc->mbuf->tx_metadata : 0 : 0; + 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)), @@ -2320,7 +2490,7 @@ mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused, 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))); + MLX5_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. */ @@ -2386,8 +2556,8 @@ mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq, 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_METADATA ? - loc->mbuf->tx_metadata : 0 : 0; + 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)), @@ -2413,7 +2583,7 @@ mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq, 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))); + MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t))); *(uint16_t *)pdst = *(unaligned_uint16_t *)psrc; psrc += sizeof(uint16_t); } else { @@ -2422,11 +2592,11 @@ mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq, psrc += sizeof(rte_v128u32_t); } pdst = (uint8_t *)(es + 2); - assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE); - assert(pdst < (uint8_t *)txq->wqes_end); + MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_ASSERT(pdst < (uint8_t *)txq->wqes_end); inlen -= MLX5_ESEG_MIN_INLINE_SIZE; if (!inlen) { - assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE)); + MLX5_ASSERT(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE)); return (struct mlx5_wqe_dseg *)pdst; } /* @@ -2467,21 +2637,30 @@ mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq, * Pointer to burst routine local context. * @param len * Length of data to be copied. + * @param must + * Length of data to be copied ignoring no inline hint. * @param olx * Configured Tx offloads mask. It is fully defined at * compile time and may be used for optimization. + * + * @return + * Number of actual copied data bytes. This is always greater than or + * equal to must parameter and might be lesser than len in no inline + * hint flag is encountered. */ -static __rte_always_inline void +static __rte_always_inline unsigned int mlx5_tx_mseg_memcpy(uint8_t *pdst, struct mlx5_txq_local *restrict loc, unsigned int len, + unsigned int must, unsigned int olx __rte_unused) { struct rte_mbuf *mbuf; - unsigned int part, dlen; + unsigned int part, dlen, copy = 0; uint8_t *psrc; - assert(len); + MLX5_ASSERT(len); + MLX5_ASSERT(must <= len); do { /* Allow zero length packets, must check first. */ dlen = rte_pktmbuf_data_len(loc->mbuf); @@ -2491,9 +2670,28 @@ mlx5_tx_mseg_memcpy(uint8_t *pdst, loc->mbuf = mbuf->next; rte_pktmbuf_free_seg(mbuf); loc->mbuf_off = 0; - assert(loc->mbuf_nseg > 1); - assert(loc->mbuf); + MLX5_ASSERT(loc->mbuf_nseg > 1); + MLX5_ASSERT(loc->mbuf); --loc->mbuf_nseg; + if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) { + unsigned int diff; + + if (copy >= must) { + /* + * We already copied the minimal + * requested amount of data. + */ + return copy; + } + diff = must - copy; + if (diff <= rte_pktmbuf_data_len(loc->mbuf)) { + /* + * Copy only the minimal required + * part of the data buffer. + */ + len = diff; + } + } continue; } dlen -= loc->mbuf_off; @@ -2501,6 +2699,7 @@ mlx5_tx_mseg_memcpy(uint8_t *pdst, loc->mbuf_off); part = RTE_MIN(len, dlen); rte_memcpy(pdst, psrc, part); + copy += part; loc->mbuf_off += part; len -= part; if (!len) { @@ -2511,10 +2710,10 @@ mlx5_tx_mseg_memcpy(uint8_t *pdst, loc->mbuf = mbuf->next; rte_pktmbuf_free_seg(mbuf); loc->mbuf_off = 0; - assert(loc->mbuf_nseg >= 1); + MLX5_ASSERT(loc->mbuf_nseg >= 1); --loc->mbuf_nseg; } - return; + return copy; } pdst += part; } while (true); @@ -2559,7 +2758,7 @@ mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq, struct mlx5_wqe_eseg *restrict es = &wqe->eseg; uint32_t csum; uint8_t *pdst; - unsigned int part; + unsigned int part, tlen = 0; /* * Calculate and set check sum flags first, uint32_t field @@ -2580,8 +2779,8 @@ mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq, 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_METADATA ? - loc->mbuf->tx_metadata : 0 : 0; + 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)), @@ -2591,37 +2790,46 @@ mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq, 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); + MLX5_ASSERT(inlen >= MLX5_ESEG_MIN_INLINE_SIZE); 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); + mlx5_tx_mseg_memcpy(pdst, loc, + 2 * RTE_ETHER_ADDR_LEN, + 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); + tlen += 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr); } - assert(pdst < (uint8_t *)txq->wqes_end); + MLX5_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); + part = RTE_MIN(part, inlen - tlen); + MLX5_ASSERT(part); do { - mlx5_tx_mseg_memcpy(pdst, loc, part, olx); - inlen -= part; - if (likely(!inlen)) { - pdst += part; + unsigned int copy; + + /* + * Copying may be interrupted inside the routine + * if run into no inline hint flag. + */ + copy = tlen >= txq->inlen_mode ? 0 : (txq->inlen_mode - tlen); + copy = mlx5_tx_mseg_memcpy(pdst, loc, part, copy, olx); + tlen += copy; + if (likely(inlen <= tlen) || copy < part) { + es->inline_hdr_sz = rte_cpu_to_be_16(tlen); + pdst += copy; pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); return (struct mlx5_wqe_dseg *)pdst; } pdst = (uint8_t *)txq->wqes; - part = inlen; + part = inlen - tlen; } while (true); } @@ -2651,7 +2859,7 @@ mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq, unsigned int olx __rte_unused) { - assert(len); + MLX5_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); @@ -2687,7 +2895,7 @@ mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq, { uintptr_t dst, src; - assert(len); + MLX5_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); @@ -2699,27 +2907,33 @@ mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq, /* 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 - memcpy(dst, src, len); + MLX5_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 - if (len & 0x08) { - *(uint64_t *)dst = *(uint64_t *)src; + *(uint64_t *)dst = *(unaligned_uint64_t *)src; dst += sizeof(uint64_t); src += sizeof(uint64_t); +#endif } if (len & 0x04) { - *(uint32_t *)dst = *(uint32_t *)src; + *(uint32_t *)dst = *(unaligned_uint32_t *)src; dst += sizeof(uint32_t); src += sizeof(uint32_t); } if (len & 0x02) { - *(uint16_t *)dst = *(uint16_t *)src; + *(uint16_t *)dst = *(unaligned_uint16_t *)src; dst += sizeof(uint16_t); src += sizeof(uint16_t); } if (len & 0x01) *(uint8_t *)dst = *(uint8_t *)src; -#endif } /** @@ -2757,8 +2971,14 @@ mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq, unsigned int part; uint8_t *pdst; - dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE); - pdst = &dseg->inline_data[0]; + if (!MLX5_TXOFF_CONFIG(MPW)) { + /* Store the descriptor byte counter for eMPW sessions. */ + dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE); + pdst = &dseg->inline_data[0]; + } else { + /* The entire legacy MPW session counter is stored on close. */ + pdst = (uint8_t *)dseg; + } /* * The WQEBB space availability is checked by caller. * Here we should be aware of WQE ring buffer wraparound only. @@ -2770,7 +2990,8 @@ mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq, len -= part; if (likely(!len)) { pdst += part; - pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + if (!MLX5_TXOFF_CONFIG(MPW)) + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); /* Note: no final wraparound check here. */ return (struct mlx5_wqe_dseg *)pdst; } @@ -2814,23 +3035,31 @@ mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq, unsigned int part; uint8_t *pdst; - assert(len > MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_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]; + if (!MLX5_TXOFF_CONFIG(MPW)) { + /* Store the descriptor byte counter for eMPW sessions. */ + dseg->bcount = rte_cpu_to_be_32 + ((len + sizeof(struct rte_vlan_hdr)) | + MLX5_ETH_WQE_DATA_INLINE); + pdst = &dseg->inline_data[0]; + } else { + /* The entire legacy MPW session counter is stored on close. */ + pdst = (uint8_t *)dseg; + } 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)); + MLX5_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); - if (unlikely(pdst >= (uint8_t *)txq->wqes_end)) - pdst = (uint8_t *)txq->wqes; /* * The WQEBB space availability is checked by caller. * Here we should be aware of WQE ring buffer wraparound only. @@ -2842,7 +3071,8 @@ mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq, len -= part; if (likely(!len)) { pdst += part; - pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); + if (!MLX5_TXOFF_CONFIG(MPW)) + pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE); /* Note: no final wraparound check here. */ return (struct mlx5_wqe_dseg *)pdst; } @@ -2893,7 +3123,7 @@ mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq, struct mlx5_wqe_dseg *restrict dseg; unsigned int ds; - assert((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen); + MLX5_ASSERT((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen); loc->mbuf_nseg = NB_SEGS(loc->mbuf); loc->mbuf_off = 0; @@ -2914,8 +3144,8 @@ mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq, * 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)); + MLX5_ASSERT(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf)); + MLX5_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; @@ -2927,7 +3157,7 @@ mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq, 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); + MLX5_ASSERT(loc->elts_free); txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; --loc->elts_free; ++dseg; @@ -2953,7 +3183,7 @@ mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq, (txq, loc, dseg, rte_pktmbuf_mtod(loc->mbuf, uint8_t *), rte_pktmbuf_data_len(loc->mbuf), olx); - assert(loc->elts_free); + MLX5_ASSERT(loc->elts_free); txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; --loc->elts_free; ++dseg; @@ -3020,7 +3250,7 @@ mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq, inlen <= MLX5_ESEG_MIN_INLINE_SIZE || inlen > (dlen + vlan))) return MLX5_TXCMP_CODE_ERROR; - assert(inlen >= txq->inlen_mode); + MLX5_ASSERT(inlen >= txq->inlen_mode); /* * Check whether there are enough free WQEBBs: * - Control Segment @@ -3092,7 +3322,7 @@ mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq, struct mlx5_wqe *restrict wqe; unsigned int ds, nseg; - assert(NB_SEGS(loc->mbuf) > 1); + MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1); /* * No inline at all, it means the CPU cycles saving * is prioritized at configuration, we should not @@ -3199,8 +3429,8 @@ mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, struct mlx5_wqe *restrict wqe; unsigned int ds, inlen, dlen, vlan = 0; - assert(MLX5_TXOFF_CONFIG(INLINE)); - assert(NB_SEGS(loc->mbuf) > 1); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1); /* * First calculate data length to be inlined * to estimate the required space for WQE. @@ -3212,8 +3442,9 @@ mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, /* 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) { + MLX5_ASSERT(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE); + if (inlen > txq->inlen_send || + loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) { struct rte_mbuf *mbuf; unsigned int nxlen; uintptr_t start; @@ -3224,11 +3455,13 @@ mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, * inlining is required. */ if (txq->inlen_mode) { - assert(txq->inlen_mode >= MLX5_ESEG_MIN_INLINE_SIZE); - assert(txq->inlen_mode <= txq->inlen_send); + MLX5_ASSERT(txq->inlen_mode >= + MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_ASSERT(txq->inlen_mode <= txq->inlen_send); inlen = txq->inlen_mode; } else { - if (!vlan || txq->vlan_en) { + if (loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE || + !vlan || txq->vlan_en) { /* * VLAN insertion will be done inside by HW. * It is not utmost effective - VLAN flag is @@ -3257,7 +3490,7 @@ mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, do { smlen = nxlen; mbuf = NEXT(mbuf); - assert(mbuf); + MLX5_ASSERT(mbuf); nxlen = rte_pktmbuf_data_len(mbuf); nxlen += smlen; } while (unlikely(nxlen < inlen)); @@ -3273,7 +3506,7 @@ mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq, inlen = nxlen; mbuf = NEXT(mbuf); /* There should be not end of packet. */ - assert(mbuf); + MLX5_ASSERT(mbuf); nxlen = inlen + rte_pktmbuf_data_len(mbuf); } while (unlikely(nxlen < txq->inlen_send)); } @@ -3301,7 +3534,7 @@ do_align: * Estimate the number of Data Segments conservatively, * supposing no any mbufs is being freed during inlining. */ - assert(inlen <= txq->inlen_send); + MLX5_ASSERT(inlen <= txq->inlen_send); ds = NB_SEGS(loc->mbuf) + 2 + (inlen - MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WSEG_SIZE + @@ -3360,14 +3593,14 @@ mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq, struct mlx5_txq_local *restrict loc, unsigned int olx) { - assert(loc->elts_free && loc->wqe_free); - assert(pkts_n > loc->pkts_sent); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); + MLX5_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); + MLX5_ASSERT(NB_SEGS(loc->mbuf) > 1); /* * Estimate the number of free elts quickly but * conservatively. Some segment may be fully inlined @@ -3403,11 +3636,11 @@ mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq, continue; /* Here ends the series of multi-segment packets. */ if (MLX5_TXOFF_CONFIG(TSO) && - unlikely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))) + unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) return MLX5_TXCMP_CODE_TSO; return MLX5_TXCMP_CODE_SINGLE; } - assert(false); + MLX5_ASSERT(false); } /** @@ -3449,8 +3682,8 @@ mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq, struct mlx5_txq_local *restrict loc, unsigned int olx) { - assert(loc->elts_free && loc->wqe_free); - assert(pkts_n > loc->pkts_sent); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); + MLX5_ASSERT(pkts_n > loc->pkts_sent); pkts += loc->pkts_sent + 1; pkts_n -= loc->pkts_sent; for (;;) { @@ -3459,7 +3692,7 @@ mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq, unsigned int ds, dlen, hlen, ntcp, vlan = 0; uint8_t *dptr; - assert(NB_SEGS(loc->mbuf) == 1); + MLX5_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) { @@ -3539,11 +3772,11 @@ mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq, if (MLX5_TXOFF_CONFIG(MULTI) && unlikely(NB_SEGS(loc->mbuf) > 1)) return MLX5_TXCMP_CODE_MULTI; - if (unlikely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))) + if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))) return MLX5_TXCMP_CODE_SINGLE; /* Continue with the next TSO packet. */ } - assert(false); + MLX5_ASSERT(false); } /** @@ -3603,6 +3836,7 @@ mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq, /** * Check the next packet attributes to match with the eMPW batch ones. + * In addition, for legacy MPW the packet length is checked either. * * @param txq * Pointer to TX queue structure. @@ -3610,6 +3844,8 @@ mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq, * Pointer to Ethernet Segment of eMPW batch. * @param loc * Pointer to burst routine local context. + * @param dlen + * Length of previous packet in MPW descriptor. * @param olx * Configured Tx offloads mask. It is fully defined at * compile time and may be used for optimization. @@ -3622,6 +3858,7 @@ 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, + uint32_t dlen, unsigned int olx) { uint8_t swp_flags = 0; @@ -3637,12 +3874,16 @@ mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused, return false; /* Fill metadata field if needed. */ if (MLX5_TXOFF_CONFIG(METADATA) && - es->metadata != (loc->mbuf->ol_flags & PKT_TX_METADATA ? - loc->mbuf->tx_metadata : 0)) + es->metadata != (loc->mbuf->ol_flags & PKT_TX_DYNF_METADATA ? + *RTE_FLOW_DYNF_METADATA(loc->mbuf) : 0)) + return false; + /* Legacy MPW can send packets with the same lengt only. */ + if (MLX5_TXOFF_CONFIG(MPW) && + dlen != rte_pktmbuf_data_len(loc->mbuf)) 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)); + MLX5_ASSERT(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)); return true; } @@ -3674,7 +3915,7 @@ mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq, unsigned int slen, unsigned int olx __rte_unused) { - assert(!MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE)); #ifdef MLX5_PMD_SOFT_COUNTERS /* Update sent data bytes counter. */ txq->stats.obytes += slen; @@ -3702,6 +3943,8 @@ mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq, * Total size of descriptor/data in bytes. * @param slen * Accumulated statistics, data bytes sent. + * @param wqem + * The base WQE for the eMPW/MPW descriptor. * @param olx * Configured Tx offloads mask. It is fully defined at * compile time and may be used for optimization. @@ -3715,20 +3958,40 @@ mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq, struct mlx5_txq_local *restrict loc, unsigned int len, unsigned int slen, + struct mlx5_wqe *restrict wqem, unsigned int olx __rte_unused) { - assert(MLX5_TXOFF_CONFIG(INLINE)); - assert((len % MLX5_WSEG_SIZE) == 0); + struct mlx5_wqe_dseg *dseg = &wqem->dseg[0]; + + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE)); #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); + if (MLX5_TXOFF_CONFIG(MPW) && dseg->bcount == RTE_BE32(0)) { + /* + * If the legacy MPW session contains the inline packets + * we should set the only inline data segment length + * and align the total length to the segment size. + */ + MLX5_ASSERT(len > sizeof(dseg->bcount)); + dseg->bcount = rte_cpu_to_be_32((len - sizeof(dseg->bcount)) | + MLX5_ETH_WQE_DATA_INLINE); + len = (len + MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE + 2; + } else { + /* + * The session is not legacy MPW or contains the + * data buffer pointer segments. + */ + MLX5_ASSERT((len % MLX5_WSEG_SIZE) == 0); + len = len / MLX5_WSEG_SIZE + 2; + } + wqem->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; + loc->wqe_last = wqem; } /** @@ -3790,10 +4053,10 @@ mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq, * 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); + MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW)); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); + MLX5_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; @@ -3805,7 +4068,10 @@ mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq, unsigned int slen = 0; next_empw: - part = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS); + MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1); + part = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ? + MLX5_MPW_MAX_PACKETS : + 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)) @@ -3835,6 +4101,10 @@ next_empw: eseg = &loc->wqe_last->eseg; dseg = &loc->wqe_last->dseg[0]; loop = part; + /* Store the packet length for legacy MPW. */ + if (MLX5_TXOFF_CONFIG(MPW)) + eseg->mss = rte_cpu_to_be_16 + (rte_pktmbuf_data_len(loc->mbuf)); for (;;) { uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf); #ifdef MLX5_PMD_SOFT_COUNTERS @@ -3864,6 +4134,7 @@ next_empw: return MLX5_TXCMP_CODE_EXIT; return MLX5_TXCMP_CODE_MULTI; } + MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1); if (ret == MLX5_TXCMP_CODE_TSO) { part -= loop; mlx5_tx_sdone_empw(txq, loc, part, slen, olx); @@ -3881,7 +4152,7 @@ next_empw: return MLX5_TXCMP_CODE_SINGLE; } if (ret != MLX5_TXCMP_CODE_EMPW) { - assert(false); + MLX5_ASSERT(false); part -= loop; mlx5_tx_sdone_empw(txq, loc, part, slen, olx); return MLX5_TXCMP_CODE_ERROR; @@ -3892,14 +4163,16 @@ next_empw: * - check sum settings * - metadata value * - software parser settings + * - packets length (legacy MPW only) */ - if (!mlx5_tx_match_empw(txq, eseg, loc, olx)) { - assert(loop); + if (!mlx5_tx_match_empw(txq, eseg, loc, dlen, olx)) { + MLX5_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. */ @@ -3908,8 +4181,8 @@ next_empw: dseg = (struct mlx5_wqe_dseg *)txq->wqes; } /* eMPW is built successfully, update loop parameters. */ - assert(!loop); - assert(pkts_n >= part); + MLX5_ASSERT(!loop); + MLX5_ASSERT(pkts_n >= part); #ifdef MLX5_PMD_SOFT_COUNTERS /* Update sent data bytes counter. */ txq->stats.obytes += slen; @@ -3927,7 +4200,7 @@ next_empw: return ret; /* Continue sending eMPW batches. */ } - assert(false); + MLX5_ASSERT(false); } /** @@ -3946,43 +4219,62 @@ mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq, * 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); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(EMPW)); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); + MLX5_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; + struct mlx5_wqe *restrict wqem; enum mlx5_txcmp_code ret; - unsigned int room, part; + unsigned int room, part, nlim; unsigned int slen = 0; -next_empw: + MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1); + /* + * Limits the amount of packets in one WQE + * to improve CQE latency generation. + */ + nlim = RTE_MIN(pkts_n, MLX5_TXOFF_CONFIG(MPW) ? + MLX5_MPW_INLINE_MAX_PACKETS : + 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); + wqem = 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_tx_cseg_init(txq, loc, wqem, 0, MLX5_OPCODE_ENHANCED_MPSW, olx); - mlx5_tx_eseg_none(txq, loc, loc->wqe_last, + mlx5_tx_eseg_none(txq, loc, wqem, olx & ~MLX5_TXOFF_CONFIG_VLAN); - eseg = &loc->wqe_last->eseg; - dseg = &loc->wqe_last->dseg[0]; + dseg = &wqem->dseg[0]; + /* Store the packet length for legacy MPW. */ + if (MLX5_TXOFF_CONFIG(MPW)) + wqem->eseg.mss = rte_cpu_to_be_16 + (rte_pktmbuf_data_len(loc->mbuf)); room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE, loc->wqe_free) * MLX5_WQE_SIZE - MLX5_WQE_CSEG_SIZE - MLX5_WQE_ESEG_SIZE; + /* Limit the room for legacy MPW sessions for performance. */ + if (MLX5_TXOFF_CONFIG(MPW)) + room = RTE_MIN(room, + RTE_MAX(txq->inlen_empw + + sizeof(dseg->bcount) + + (MLX5_TXOFF_CONFIG(VLAN) ? + sizeof(struct rte_vlan_hdr) : 0), + MLX5_MPW_INLINE_MAX_PACKETS * + MLX5_WQE_DSEG_SIZE)); /* Build WQE till we have space, packets and resources. */ part = room; for (;;) { @@ -3990,9 +4282,9 @@ next_empw: 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); + MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE); + MLX5_ASSERT((room % MLX5_WQE_DSEG_SIZE) == 0); + MLX5_ASSERT((uintptr_t)dseg < (uintptr_t)txq->wqes_end); /* * Some Tx offloads may cause an error if * packet is not long enough, check against @@ -4006,14 +4298,36 @@ next_empw: * We have some successfully built * packet Data Segments to send. */ - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, olx); return MLX5_TXCMP_CODE_ERROR; } /* Inline or not inline - that's the Question. */ - if (dlen > txq->inlen_empw) + if (dlen > txq->inlen_empw || + loc->mbuf->ol_flags & PKT_TX_DYNF_NOINLINE) goto pointer_empw; + if (MLX5_TXOFF_CONFIG(MPW)) { + if (dlen > txq->inlen_send) + goto pointer_empw; + tlen = dlen; + if (part == room) { + /* Open new inline MPW session. */ + tlen += sizeof(dseg->bcount); + dseg->bcount = RTE_BE32(0); + dseg = RTE_PTR_ADD + (dseg, sizeof(dseg->bcount)); + } else { + /* + * No pointer and inline descriptor + * intermix for legacy MPW sessions. + */ + if (wqem->dseg[0].bcount) + break; + } + } else { + tlen = sizeof(dseg->bcount) + dlen; + } /* Inline entire packet, optional VLAN insertion. */ - tlen = sizeof(dseg->bcount) + dlen; if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) { /* @@ -4021,8 +4335,9 @@ next_empw: * mlx5_tx_able_to_empw() and packet * fits into inline length guaranteed. */ - assert((dlen + sizeof(struct rte_vlan_hdr)) <= - txq->inlen_empw); + MLX5_ASSERT((dlen + + sizeof(struct rte_vlan_hdr)) <= + txq->inlen_empw); tlen += sizeof(struct rte_vlan_hdr); if (room < tlen) break; @@ -4038,8 +4353,9 @@ next_empw: dseg = mlx5_tx_dseg_empw(txq, loc, dseg, dptr, dlen, olx); } - tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE); - assert(room >= tlen); + if (!MLX5_TXOFF_CONFIG(MPW)) + tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE); + MLX5_ASSERT(room >= tlen); room -= tlen; /* * Packet data are completely inlined, @@ -4048,14 +4364,22 @@ next_empw: rte_pktmbuf_free_seg(loc->mbuf); goto next_mbuf; pointer_empw: + /* + * No pointer and inline descriptor + * intermix for legacy MPW sessions. + */ + if (MLX5_TXOFF_CONFIG(MPW) && + part != room && + wqem->dseg[0].bcount == RTE_BE32(0)) + break; /* * Not inlinable VLAN packets are * proceeded outside of this routine. */ - assert(room >= MLX5_WQE_DSEG_SIZE); + MLX5_ASSERT(room >= MLX5_WQE_DSEG_SIZE); if (MLX5_TXOFF_CONFIG(VLAN)) - assert(!(loc->mbuf->ol_flags & - PKT_TX_VLAN_PKT)); + MLX5_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; @@ -4076,15 +4400,10 @@ next_mbuf: * continue build descriptors. */ part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, olx); return MLX5_TXCMP_CODE_EXIT; } - /* Check if we have minimal room left. */ - if (room < MLX5_WQE_DSEG_SIZE) { - part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); - goto next_empw; - } loc->mbuf = *pkts++; if (likely(pkts_n > 1)) rte_prefetch0(*pkts); @@ -4096,15 +4415,18 @@ next_mbuf: */ if (ret == MLX5_TXCMP_CODE_MULTI) { part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, olx); if (unlikely(!loc->elts_free || !loc->wqe_free)) return MLX5_TXCMP_CODE_EXIT; return MLX5_TXCMP_CODE_MULTI; } + MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1); if (ret == MLX5_TXCMP_CODE_TSO) { part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, olx); if (unlikely(!loc->elts_free || !loc->wqe_free)) return MLX5_TXCMP_CODE_EXIT; @@ -4112,26 +4434,34 @@ next_mbuf: } if (ret == MLX5_TXCMP_CODE_SINGLE) { part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, 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); + MLX5_ASSERT(false); part -= room; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, + slen, wqem, 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 + * - packets length (legacy MPW only) */ - if (!mlx5_tx_match_empw(txq, eseg, loc, olx)) + if (!mlx5_tx_match_empw(txq, &wqem->eseg, + loc, dlen, olx)) break; /* Packet attributes match, continue the same eMPW. */ if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end) @@ -4141,17 +4471,17 @@ next_mbuf: * We get here to close an existing eMPW * session and start the new one. */ - assert(pkts_n); + MLX5_ASSERT(pkts_n); part -= room; if (unlikely(!part)) return MLX5_TXCMP_CODE_EXIT; - mlx5_tx_idone_empw(txq, loc, part, slen, olx); + mlx5_tx_idone_empw(txq, loc, part, slen, wqem, olx); if (unlikely(!loc->elts_free || !loc->wqe_free)) return MLX5_TXCMP_CODE_EXIT; - goto next_empw; + /* Continue the loop with new eMPW session. */ } - assert(false); + MLX5_ASSERT(false); } /** @@ -4169,15 +4499,15 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * 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); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); + MLX5_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); + MLX5_ASSERT(NB_SEGS(loc->mbuf) == 1); if (MLX5_TXOFF_CONFIG(INLINE)) { unsigned int inlen, vlan = 0; @@ -4197,7 +4527,8 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * 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); + MLX5_ASSERT(txq->inlen_send >= + MLX5_ESEG_MIN_INLINE_SIZE); if (inlen <= txq->inlen_send) { unsigned int seg_n, wqe_n; @@ -4206,6 +4537,33 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, /* Check against minimal length. */ if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE) return MLX5_TXCMP_CODE_ERROR; + if (loc->mbuf->ol_flags & + PKT_TX_DYNF_NOINLINE) { + /* + * The hint flag not to inline packet + * data is set. Check whether we can + * follow the hint. + */ + if ((!MLX5_TXOFF_CONFIG(EMPW) && + txq->inlen_mode) || + (MLX5_TXOFF_CONFIG(MPW) && + txq->inlen_mode)) { + /* + * The hardware requires the + * minimal inline data header. + */ + goto single_min_inline; + } + if (MLX5_TXOFF_CONFIG(VLAN) && + vlan && !txq->vlan_en) { + /* + * We must insert VLAN tag + * by software means. + */ + goto single_part_inline; + } + goto single_no_inline; + } /* * Completely inlined packet data WQE: * - Control Segment, SEND opcode @@ -4234,8 +4592,9 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * free the packet immediately. */ rte_pktmbuf_free_seg(loc->mbuf); - } else if (!MLX5_TXOFF_CONFIG(EMPW) && - txq->inlen_mode) { + } else if ((!MLX5_TXOFF_CONFIG(EMPW) || + MLX5_TXOFF_CONFIG(MPW)) && + txq->inlen_mode) { /* * If minimal inlining is requested the eMPW * feature should be disabled due to data is @@ -4254,10 +4613,11 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * 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); +single_min_inline: + MLX5_ASSERT(txq->inlen_send >= txq->inlen_mode); + MLX5_ASSERT(inlen > txq->inlen_mode); + MLX5_ASSERT(txq->inlen_mode >= + MLX5_ESEG_MIN_INLINE_SIZE); /* * Check whether there are enough free WQEBBs: * - Control Segment @@ -4300,7 +4660,7 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, txq->wqe_ci += (ds + 3) / 4; loc->wqe_free -= (ds + 3) / 4; /* We have to store mbuf in elts.*/ - assert(MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE)); txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; --loc->elts_free; @@ -4321,6 +4681,7 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * We also get here if VLAN insertion is not * supported by HW, the inline is enabled. */ +single_part_inline: wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); loc->wqe_last = wqe; mlx5_tx_cseg_init(txq, loc, wqe, 4, @@ -4333,14 +4694,14 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * comparing with txq->inlen_send. We should * not get overflow here. */ - assert(inlen > MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_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)); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE)); txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf; --loc->elts_free; @@ -4361,6 +4722,7 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * - Ethernet Segment, optional VLAN, no inline * - Data Segment, pointer type */ +single_no_inline: wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m); loc->wqe_last = wqe; mlx5_tx_cseg_init(txq, loc, wqe, 3, @@ -4377,7 +4739,7 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, * if no inlining is configured, this is done * by calling routine in a batch copy. */ - assert(!MLX5_TXOFF_CONFIG(INLINE)); + MLX5_ASSERT(!MLX5_TXOFF_CONFIG(INLINE)); --loc->elts_free; #ifdef MLX5_PMD_SOFT_COUNTERS /* Update sent data bytes counter. */ @@ -4399,7 +4761,7 @@ mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq, if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE)) return ret; } - assert(false); + MLX5_ASSERT(false); } static __rte_always_inline enum mlx5_txcmp_code @@ -4414,7 +4776,7 @@ mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq, 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); + MLX5_ASSERT(ret == MLX5_TXCMP_CODE_EMPW); for (;;) { /* Optimize for inline/no inline eMPW send. */ ret = (MLX5_TXOFF_CONFIG(INLINE)) ? @@ -4425,14 +4787,14 @@ mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq, if (ret != MLX5_TXCMP_CODE_SINGLE) return ret; /* The resources to send one packet should remain. */ - assert(loc->elts_free && loc->wqe_free); + MLX5_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); + MLX5_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); + MLX5_ASSERT(loc->elts_free && loc->wqe_free); } } @@ -4466,8 +4828,16 @@ mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq, 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)); + MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); + MLX5_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 @@ -4475,29 +4845,24 @@ mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq, * - free related mbufs * - doorbell the NIC about processed CQEs */ - if (unlikely(!pkts_n)) - return 0; - rte_prefetch0(*pkts); + 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 + * - no data inlining into WQEs, one WQEBB may contains up to * 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)); + MLX5_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)); + MLX5_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)) - return 0; - loc.pkts_sent = 0; - loc.pkts_copy = 0; - loc.wqe_last = NULL; + goto burst_exit; for (;;) { /* * Fetch the packet from array. Usually this is @@ -4515,7 +4880,7 @@ mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq, * per WQE, do it in dedicated routine. */ enter_send_multi: - assert(loc.pkts_sent >= loc.pkts_copy); + MLX5_ASSERT(loc.pkts_sent >= loc.pkts_copy); part = loc.pkts_sent - loc.pkts_copy; if (!MLX5_TXOFF_CONFIG(INLINE) && part) { /* @@ -4529,7 +4894,7 @@ enter_send_multi: part, olx); loc.pkts_copy = loc.pkts_sent; } - assert(pkts_n > loc.pkts_sent); + MLX5_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; @@ -4571,7 +4936,7 @@ enter_send_multi: goto enter_send_tso; } /* We must not get here. Something is going wrong. */ - assert(false); + MLX5_ASSERT(false); txq->stats.oerrors++; break; } @@ -4585,8 +4950,8 @@ enter_send_multi: * in dedicated branch. */ enter_send_tso: - assert(NB_SEGS(loc.mbuf) == 1); - assert(pkts_n > loc.pkts_sent); + MLX5_ASSERT(NB_SEGS(loc.mbuf) == 1); + MLX5_ASSERT(pkts_n > loc.pkts_sent); ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx); /* * These returned code checks are supposed @@ -4609,7 +4974,7 @@ enter_send_tso: goto enter_send_multi; } /* We must not get here. Something is going wrong. */ - assert(false); + MLX5_ASSERT(false); txq->stats.oerrors++; break; } @@ -4622,7 +4987,7 @@ enter_send_tso: * offloads are requested at SQ configuration time). */ enter_send_single: - assert(pkts_n > loc.pkts_sent); + MLX5_ASSERT(pkts_n > loc.pkts_sent); ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx); /* * These returned code checks are supposed @@ -4651,7 +5016,7 @@ enter_send_single: goto enter_send_tso; } /* We must not get here. Something is going wrong. */ - assert(false); + MLX5_ASSERT(false); txq->stats.oerrors++; break; } @@ -4661,24 +5026,51 @@ enter_send_single: * - doorbell the hardware * - copy the rest of mbufs to elts (if any) */ - assert(MLX5_TXOFF_CONFIG(INLINE) || loc.pkts_sent >= loc.pkts_copy); + MLX5_ASSERT(MLX5_TXOFF_CONFIG(INLINE) || + loc.pkts_sent >= loc.pkts_copy); /* Take a shortcut if nothing is sent. */ - if (unlikely(loc.pkts_sent == 0)) - return 0; - /* Not all of the mbufs may be stored into elts yet. */ - part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy; - mlx5_tx_request_completion(txq, part, &loc, olx); + if (unlikely(loc.pkts_sent == loc.pkts_loop)) + goto burst_exit; + /* Request CQE generation if limits are reached. */ + mlx5_tx_request_completion(txq, &loc, olx); /* * 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. + * + * The rdma core library can map doorbell register in two ways, + * depending on the environment variable "MLX5_SHUT_UP_BF": + * + * - as regular cached memory, the variable is either missing or + * set to zero. This type of mapping may cause the significant + * doorbell register writing latency and requires explicit + * memory write barrier to mitigate this issue and prevent + * write combining. + * + * - as non-cached memory, the variable is present and set to + * not "0" value. This type of mapping may cause performance + * impact under heavy loading conditions but the explicit write + * memory barrier is not required and it may improve core + * performance. + * + * - the legacy behaviour (prior 19.08 release) was to use some + * heuristics to decide whether write memory barrier should + * be performed. This behavior is supported with specifying + * tx_db_nc=2, write barrier is skipped if application + * provides the full recommended burst of packets, it + * supposes the next packets are coming and the write barrier + * will be issued on the next burst (after descriptor writing, + * at least). */ - mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, 0); + mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, !txq->db_nc && + (!txq->db_heu || pkts_n % MLX5_TX_DEFAULT_BURST)); + /* 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 last packet was single-segment. + * 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 @@ -4686,13 +5078,23 @@ enter_send_single: * inlined mbufs. */ mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx); + loc.pkts_copy = loc.pkts_sent; + } + MLX5_ASSERT(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); + MLX5_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 - assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail)); - assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi)); return loc.pkts_sent; } @@ -4830,6 +5232,34 @@ MLX5_TXOFF_DECL(iv, MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA) +/* + * Generate routines with Legacy Multi-Packet Write support. + * This mode is supported by ConnectX-4 Lx only and imposes + * offload limitations, not supported: + * - ACL/Flows (metadata are becoming meaningless) + * - WQE Inline headers + * - SRIOV (E-Switch offloads) + * - VLAN insertion + * - tunnel encapsulation/decapsulation + * - TSO + */ +MLX5_TXOFF_DECL(none_mpw, + MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_DECL(mci_mpw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_DECL(mc_mpw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_DECL(i_mpw, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) + /* * Array of declared and compiled Tx burst function and corresponding * supported offloads set. The array is used to select the Tx burst @@ -4932,7 +5362,6 @@ MLX5_TXOFF_INFO(mti, MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_METADATA) - MLX5_TXOFF_INFO(mtv, MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO | MLX5_TXOFF_CONFIG_VLAN | @@ -4973,6 +5402,23 @@ MLX5_TXOFF_INFO(v, MLX5_TXOFF_INFO(iv, MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN | MLX5_TXOFF_CONFIG_METADATA) + +MLX5_TXOFF_INFO(none_mpw, + MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_INFO(mci_mpw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_INFO(mc_mpw, + MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_CSUM | + MLX5_TXOFF_CONFIG_EMPW | MLX5_TXOFF_CONFIG_MPW) + +MLX5_TXOFF_INFO(i_mpw, + MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW) }; /** @@ -5006,7 +5452,7 @@ mlx5_select_tx_function(struct rte_eth_dev *dev) "invalid WQE Data Segment size"); static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE, "invalid WQE size"); - assert(priv); + MLX5_ASSERT(priv); if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) { /* We should support Multi-Segment Packets. */ olx |= MLX5_TXOFF_CONFIG_MULTI; @@ -5055,18 +5501,29 @@ mlx5_select_tx_function(struct rte_eth_dev *dev) 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. + * The NIC supports Enhanced Multi-Packet Write + * and does not require minimal inline data. */ olx |= MLX5_TXOFF_CONFIG_EMPW; } - if (tx_offloads & DEV_TX_OFFLOAD_MATCH_METADATA) { + if (rte_flow_dynf_metadata_avail()) { /* We should support Flow metadata. */ olx |= MLX5_TXOFF_CONFIG_METADATA; } + if (config->mps == MLX5_MPW) { + /* + * The NIC supports Legacy Multi-Packet Write. + * The MLX5_TXOFF_CONFIG_MPW controls the + * descriptor building method in combination + * with MLX5_TXOFF_CONFIG_EMPW. + */ + if (!(olx & (MLX5_TXOFF_CONFIG_TSO | + MLX5_TXOFF_CONFIG_SWP | + MLX5_TXOFF_CONFIG_VLAN | + MLX5_TXOFF_CONFIG_METADATA))) + olx |= MLX5_TXOFF_CONFIG_EMPW | + MLX5_TXOFF_CONFIG_MPW; + } /* * Scan the routines table to find the minimal * satisfying routine with requested offloads. @@ -5135,9 +5592,100 @@ mlx5_select_tx_function(struct rte_eth_dev *dev) 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)"); + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW) { + if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MPW) + DRV_LOG(DEBUG, "\tMPW (Legacy MPW)"); + else + DRV_LOG(DEBUG, "\tEMPW (Enhanced MPW)"); + } return txoff_func[m].func; } +/** + * DPDK callback to get the TX queue information + * + * @param dev + * Pointer to the device structure. + * + * @param tx_queue_id + * Tx queue identificator. + * + * @param qinfo + * Pointer to the TX queue information structure. + * + * @return + * None. + */ + +void +mlx5_txq_info_get(struct rte_eth_dev *dev, uint16_t tx_queue_id, + struct rte_eth_txq_info *qinfo) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq = (*priv->txqs)[tx_queue_id]; + struct mlx5_txq_ctrl *txq_ctrl = + container_of(txq, struct mlx5_txq_ctrl, txq); + + if (!txq) + return; + qinfo->nb_desc = txq->elts_s; + qinfo->conf.tx_thresh.pthresh = 0; + qinfo->conf.tx_thresh.hthresh = 0; + qinfo->conf.tx_thresh.wthresh = 0; + qinfo->conf.tx_rs_thresh = 0; + qinfo->conf.tx_free_thresh = 0; + qinfo->conf.tx_deferred_start = txq_ctrl ? 0 : 1; + qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads; +} +/** + * DPDK callback to get the TX packet burst mode information + * + * @param dev + * Pointer to the device structure. + * + * @param tx_queue_id + * Tx queue identificatior. + * + * @param mode + * Pointer to the burts mode information. + * + * @return + * 0 as success, -EINVAL as failure. + */ + +int +mlx5_tx_burst_mode_get(struct rte_eth_dev *dev, + uint16_t tx_queue_id __rte_unused, + struct rte_eth_burst_mode *mode) +{ + eth_tx_burst_t pkt_burst = dev->tx_pkt_burst; + unsigned int i, olx; + + for (i = 0; i < RTE_DIM(txoff_func); i++) { + if (pkt_burst == txoff_func[i].func) { + olx = txoff_func[i].olx; + snprintf(mode->info, sizeof(mode->info), + "%s%s%s%s%s%s%s%s", + (olx & MLX5_TXOFF_CONFIG_EMPW) ? + ((olx & MLX5_TXOFF_CONFIG_MPW) ? + "Legacy MPW" : "Enhanced MPW") : "No MPW", + (olx & MLX5_TXOFF_CONFIG_MULTI) ? + " + MULTI" : "", + (olx & MLX5_TXOFF_CONFIG_TSO) ? + " + TSO" : "", + (olx & MLX5_TXOFF_CONFIG_SWP) ? + " + SWP" : "", + (olx & MLX5_TXOFF_CONFIG_CSUM) ? + " + CSUM" : "", + (olx & MLX5_TXOFF_CONFIG_INLINE) ? + " + INLINE" : "", + (olx & MLX5_TXOFF_CONFIG_VLAN) ? + " + VLAN" : "", + (olx & MLX5_TXOFF_CONFIG_METADATA) ? + " + METADATA" : ""); + return 0; + } + } + return -EINVAL; +}