X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;ds=sidebyside;f=drivers%2Fnet%2Fmlx5%2Fmlx5_txq.c;h=4ab6ac1611027d5a6c4eb6814cae73ace7961135;hb=0febfcce36934d7265d37be818690d0fe8a2e9ec;hp=fe3b4ec76d74cd28d2bcbeec06f715bbba843251;hpb=8791ff42ef0222c8cccd1092485102734184b658;p=dpdk.git diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c index fe3b4ec76d..4ab6ac1611 100644 --- a/drivers/net/mlx5/mlx5_txq.c +++ b/drivers/net/mlx5/mlx5_txq.c @@ -4,7 +4,6 @@ */ #include -#include #include #include #include @@ -18,6 +17,7 @@ #pragma GCC diagnostic ignored "-Wpedantic" #endif #include +#include #ifdef PEDANTIC #pragma GCC diagnostic error "-Wpedantic" #endif @@ -27,12 +27,17 @@ #include #include -#include "mlx5_utils.h" +#include +#include +#include +#include +#include + #include "mlx5_defs.h" +#include "mlx5_utils.h" #include "mlx5.h" #include "mlx5_rxtx.h" #include "mlx5_autoconf.h" -#include "mlx5_glue.h" /** * Allocate TX queue elements. @@ -61,7 +66,7 @@ txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl) * @param txq_ctrl * Pointer to TX queue structure. */ -static void +void txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl) { const uint16_t elts_n = 1 << txq_ctrl->txq.elts_n; @@ -79,9 +84,9 @@ txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl) while (elts_tail != elts_head) { struct rte_mbuf *elt = (*elts)[elts_tail & elts_m]; - assert(elt != NULL); + MLX5_ASSERT(elt != NULL); rte_pktmbuf_free_seg(elt); -#ifndef NDEBUG +#ifdef RTE_LIBRTE_MLX5_DEBUG /* Poisoning. */ memset(&(*elts)[elts_tail & elts_m], 0x77, @@ -114,6 +119,8 @@ mlx5_get_tx_port_offloads(struct rte_eth_dev *dev) DEV_TX_OFFLOAD_TCP_CKSUM); if (config->tso) offloads |= DEV_TX_OFFLOAD_TCP_TSO; + if (config->tx_pp) + offloads |= DEV_TX_OFFLOAD_SEND_ON_TIMESTAMP; if (config->swp) { if (config->hw_csum) offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM; @@ -126,58 +133,47 @@ mlx5_get_tx_port_offloads(struct rte_eth_dev *dev) offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM; if (config->tso) offloads |= (DEV_TX_OFFLOAD_VXLAN_TNL_TSO | - DEV_TX_OFFLOAD_GRE_TNL_TSO); + DEV_TX_OFFLOAD_GRE_TNL_TSO | + DEV_TX_OFFLOAD_GENEVE_TNL_TSO); } -#ifdef HAVE_IBV_FLOW_DV_SUPPORT - if (config->dv_flow_en) - offloads |= DEV_TX_OFFLOAD_MATCH_METADATA; -#endif return offloads; } /** - * DPDK callback to configure a TX queue. + * Tx queue presetup checks. * * @param dev * Pointer to Ethernet device structure. * @param idx - * TX queue index. + * Tx queue index. * @param desc * Number of descriptors to configure in queue. - * @param socket - * NUMA socket on which memory must be allocated. - * @param[in] conf - * Thresholds parameters. * * @return * 0 on success, a negative errno value otherwise and rte_errno is set. */ -int -mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, - unsigned int socket, const struct rte_eth_txconf *conf) +static int +mlx5_tx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc) { struct mlx5_priv *priv = dev->data->dev_private; - struct mlx5_txq_data *txq = (*priv->txqs)[idx]; - struct mlx5_txq_ctrl *txq_ctrl = - container_of(txq, struct mlx5_txq_ctrl, txq); - if (desc <= MLX5_TX_COMP_THRESH) { + if (*desc <= MLX5_TX_COMP_THRESH) { DRV_LOG(WARNING, "port %u number of descriptors requested for Tx queue" " %u must be higher than MLX5_TX_COMP_THRESH, using %u" - " instead of %u", - dev->data->port_id, idx, MLX5_TX_COMP_THRESH + 1, desc); - desc = MLX5_TX_COMP_THRESH + 1; + " instead of %u", dev->data->port_id, idx, + MLX5_TX_COMP_THRESH + 1, *desc); + *desc = MLX5_TX_COMP_THRESH + 1; } - if (!rte_is_power_of_2(desc)) { - desc = 1 << log2above(desc); + if (!rte_is_power_of_2(*desc)) { + *desc = 1 << log2above(*desc); DRV_LOG(WARNING, "port %u increased number of descriptors in Tx queue" " %u to the next power of two (%d)", - dev->data->port_id, idx, desc); + dev->data->port_id, idx, *desc); } DRV_LOG(DEBUG, "port %u configuring queue %u for %u descriptors", - dev->data->port_id, idx, desc); + dev->data->port_id, idx, *desc); if (idx >= priv->txqs_n) { DRV_LOG(ERR, "port %u Tx queue index out of range (%u >= %u)", dev->data->port_id, idx, priv->txqs_n); @@ -191,6 +187,38 @@ mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, return -rte_errno; } mlx5_txq_release(dev, idx); + return 0; +} +/** + * DPDK callback to configure a TX queue. + * + * @param dev + * Pointer to Ethernet device structure. + * @param idx + * TX queue index. + * @param desc + * Number of descriptors to configure in queue. + * @param socket + * NUMA socket on which memory must be allocated. + * @param[in] conf + * Thresholds parameters. + * + * @return + * 0 on success, a negative errno value otherwise and rte_errno is set. + */ +int +mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, + unsigned int socket, const struct rte_eth_txconf *conf) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq = (*priv->txqs)[idx]; + struct mlx5_txq_ctrl *txq_ctrl = + container_of(txq, struct mlx5_txq_ctrl, txq); + int res; + + res = mlx5_tx_queue_pre_setup(dev, idx, &desc); + if (res) + return res; txq_ctrl = mlx5_txq_new(dev, idx, desc, socket, conf); if (!txq_ctrl) { DRV_LOG(ERR, "port %u unable to allocate queue index %u", @@ -203,6 +231,56 @@ mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, return 0; } +/** + * DPDK callback to configure a TX hairpin queue. + * + * @param dev + * Pointer to Ethernet device structure. + * @param idx + * TX queue index. + * @param desc + * Number of descriptors to configure in queue. + * @param[in] hairpin_conf + * The hairpin binding configuration. + * + * @return + * 0 on success, a negative errno value otherwise and rte_errno is set. + */ +int +mlx5_tx_hairpin_queue_setup(struct rte_eth_dev *dev, uint16_t idx, + uint16_t desc, + const struct rte_eth_hairpin_conf *hairpin_conf) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq = (*priv->txqs)[idx]; + struct mlx5_txq_ctrl *txq_ctrl = + container_of(txq, struct mlx5_txq_ctrl, txq); + int res; + + res = mlx5_tx_queue_pre_setup(dev, idx, &desc); + if (res) + return res; + if (hairpin_conf->peer_count != 1 || + hairpin_conf->peers[0].port != dev->data->port_id || + hairpin_conf->peers[0].queue >= priv->rxqs_n) { + DRV_LOG(ERR, "port %u unable to setup hairpin queue index %u " + " invalid hairpind configuration", dev->data->port_id, + idx); + rte_errno = EINVAL; + return -rte_errno; + } + txq_ctrl = mlx5_txq_hairpin_new(dev, idx, desc, hairpin_conf); + if (!txq_ctrl) { + DRV_LOG(ERR, "port %u unable to allocate queue index %u", + dev->data->port_id, idx); + return -rte_errno; + } + DRV_LOG(DEBUG, "port %u adding Tx queue %u to list", + dev->data->port_id, idx); + (*priv->txqs)[idx] = &txq_ctrl->txq; + return 0; +} + /** * DPDK callback to release a TX queue. * @@ -223,13 +301,37 @@ mlx5_tx_queue_release(void *dpdk_txq) priv = txq_ctrl->priv; for (i = 0; (i != priv->txqs_n); ++i) if ((*priv->txqs)[i] == txq) { - mlx5_txq_release(ETH_DEV(priv), i); DRV_LOG(DEBUG, "port %u removing Tx queue %u from list", PORT_ID(priv), txq->idx); + mlx5_txq_release(ETH_DEV(priv), i); break; } } +/** + * Configure the doorbell register non-cached attribute. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + * @param page_size + * Systme page size + */ +static void +txq_uar_ncattr_init(struct mlx5_txq_ctrl *txq_ctrl, size_t page_size) +{ + struct mlx5_priv *priv = txq_ctrl->priv; + off_t cmd; + + txq_ctrl->txq.db_heu = priv->config.dbnc == MLX5_TXDB_HEURISTIC; + txq_ctrl->txq.db_nc = 0; + /* Check the doorbell register mapping type. */ + cmd = txq_ctrl->uar_mmap_offset / page_size; + cmd >>= MLX5_UAR_MMAP_CMD_SHIFT; + cmd &= MLX5_UAR_MMAP_CMD_MASK; + if (cmd == MLX5_MMAP_GET_NC_PAGES_CMD) + txq_ctrl->txq.db_nc = 1; +} + /** * Initialize Tx UAR registers for primary process. * @@ -241,19 +343,22 @@ txq_uar_init(struct mlx5_txq_ctrl *txq_ctrl) { struct mlx5_priv *priv = txq_ctrl->priv; struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv)); + const size_t page_size = sysconf(_SC_PAGESIZE); #ifndef RTE_ARCH_64 unsigned int lock_idx; - const size_t page_size = sysconf(_SC_PAGESIZE); #endif - assert(rte_eal_process_type() == RTE_PROC_PRIMARY); - assert(ppriv); + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + return; + MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY); + MLX5_ASSERT(ppriv); ppriv->uar_table[txq_ctrl->txq.idx] = txq_ctrl->bf_reg; + txq_uar_ncattr_init(txq_ctrl, page_size); #ifndef RTE_ARCH_64 /* Assign an UAR lock according to UAR page number */ lock_idx = (txq_ctrl->uar_mmap_offset / page_size) & MLX5_UAR_PAGE_NUM_MASK; - txq_ctrl->txq.uar_lock = &priv->uar_lock[lock_idx]; + txq_ctrl->txq.uar_lock = &priv->sh->uar_lock[lock_idx]; #endif } @@ -282,7 +387,9 @@ txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd) uintptr_t offset; const size_t page_size = sysconf(_SC_PAGESIZE); - assert(ppriv); + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + return 0; + MLX5_ASSERT(ppriv); /* * As rdma-core, UARs are mapped in size of OS page * size. Ref to libmlx5 function: mlx5_init_context() @@ -300,6 +407,7 @@ txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd) } addr = RTE_PTR_ADD(addr, offset); ppriv->uar_table[txq->idx] = addr; + txq_uar_ncattr_init(txq_ctrl, page_size); return 0; } @@ -316,10 +424,36 @@ txq_uar_uninit_secondary(struct mlx5_txq_ctrl *txq_ctrl) const size_t page_size = sysconf(_SC_PAGESIZE); void *addr; + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + return; addr = ppriv->uar_table[txq_ctrl->txq.idx]; munmap(RTE_PTR_ALIGN_FLOOR(addr, page_size), page_size); } +/** + * Deinitialize Tx UAR registers for secondary process. + * + * @param dev + * Pointer to Ethernet device. + */ +void +mlx5_tx_uar_uninit_secondary(struct rte_eth_dev *dev) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq; + struct mlx5_txq_ctrl *txq_ctrl; + unsigned int i; + + MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_SECONDARY); + for (i = 0; i != priv->txqs_n; ++i) { + if (!(*priv->txqs)[i]) + continue; + txq = (*priv->txqs)[i]; + txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq); + txq_uar_uninit_secondary(txq_ctrl); + } +} + /** * Initialize Tx UAR registers for secondary process. * @@ -340,13 +474,15 @@ mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd) unsigned int i; int ret; - assert(rte_eal_process_type() == RTE_PROC_SECONDARY); + MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_SECONDARY); for (i = 0; i != priv->txqs_n; ++i) { if (!(*priv->txqs)[i]) continue; txq = (*priv->txqs)[i]; txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq); - assert(txq->idx == (uint16_t)i); + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + continue; + MLX5_ASSERT(txq->idx == (uint16_t)i); ret = txq_uar_init_secondary(txq_ctrl, fd); if (ret) goto error; @@ -364,6 +500,372 @@ error: return -rte_errno; } +/** + * Create the Tx hairpin queue object. + * + * @param dev + * Pointer to Ethernet device. + * @param idx + * Queue index in DPDK Tx queue array + * + * @return + * The hairpin DevX object initialised, NULL otherwise and rte_errno is set. + */ +static struct mlx5_txq_obj * +mlx5_txq_obj_hairpin_new(struct rte_eth_dev *dev, uint16_t idx) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_data *txq_data = (*priv->txqs)[idx]; + struct mlx5_txq_ctrl *txq_ctrl = + container_of(txq_data, struct mlx5_txq_ctrl, txq); + struct mlx5_devx_create_sq_attr attr = { 0 }; + struct mlx5_txq_obj *tmpl = NULL; + uint32_t max_wq_data; + + MLX5_ASSERT(txq_data); + MLX5_ASSERT(!txq_ctrl->obj); + tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0, + txq_ctrl->socket); + if (!tmpl) { + DRV_LOG(ERR, + "port %u Tx queue %u cannot allocate memory resources", + dev->data->port_id, txq_data->idx); + rte_errno = ENOMEM; + return NULL; + } + tmpl->type = MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN; + tmpl->txq_ctrl = txq_ctrl; + attr.hairpin = 1; + attr.tis_lst_sz = 1; + max_wq_data = priv->config.hca_attr.log_max_hairpin_wq_data_sz; + /* Jumbo frames > 9KB should be supported, and more packets. */ + if (priv->config.log_hp_size != (uint32_t)MLX5_ARG_UNSET) { + if (priv->config.log_hp_size > max_wq_data) { + DRV_LOG(ERR, "total data size %u power of 2 is " + "too large for hairpin", + priv->config.log_hp_size); + rte_free(tmpl); + rte_errno = ERANGE; + return NULL; + } + attr.wq_attr.log_hairpin_data_sz = priv->config.log_hp_size; + } else { + attr.wq_attr.log_hairpin_data_sz = + (max_wq_data < MLX5_HAIRPIN_JUMBO_LOG_SIZE) ? + max_wq_data : MLX5_HAIRPIN_JUMBO_LOG_SIZE; + } + /* Set the packets number to the maximum value for performance. */ + attr.wq_attr.log_hairpin_num_packets = + attr.wq_attr.log_hairpin_data_sz - + MLX5_HAIRPIN_QUEUE_STRIDE; + attr.tis_num = priv->sh->tis->id; + tmpl->sq = mlx5_devx_cmd_create_sq(priv->sh->ctx, &attr); + if (!tmpl->sq) { + DRV_LOG(ERR, + "port %u tx hairpin queue %u can't create sq object", + dev->data->port_id, idx); + rte_free(tmpl); + rte_errno = errno; + return NULL; + } + DRV_LOG(DEBUG, "port %u sxq %u updated with %p", dev->data->port_id, + idx, (void *)&tmpl); + rte_atomic32_inc(&tmpl->refcnt); + LIST_INSERT_HEAD(&priv->txqsobj, tmpl, next); + return tmpl; +} + +/** + * Destroy the Tx queue DevX object. + * + * @param txq_obj + * Txq object to destroy + */ +static void +txq_release_sq_resources(struct mlx5_txq_obj *txq_obj) +{ + MLX5_ASSERT(txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ); + + if (txq_obj->sq_devx) + claim_zero(mlx5_devx_cmd_destroy(txq_obj->sq_devx)); + if (txq_obj->sq_dbrec_page) + claim_zero(mlx5_release_dbr + (&txq_obj->txq_ctrl->priv->dbrpgs, + mlx5_os_get_umem_id + (txq_obj->sq_dbrec_page->umem), + txq_obj->sq_dbrec_offset)); + if (txq_obj->sq_umem) + claim_zero(mlx5_glue->devx_umem_dereg(txq_obj->sq_umem)); + if (txq_obj->sq_buf) + rte_free(txq_obj->sq_buf); + if (txq_obj->cq_devx) + claim_zero(mlx5_devx_cmd_destroy(txq_obj->cq_devx)); + if (txq_obj->cq_dbrec_page) + claim_zero(mlx5_release_dbr + (&txq_obj->txq_ctrl->priv->dbrpgs, + mlx5_os_get_umem_id + (txq_obj->cq_dbrec_page->umem), + txq_obj->cq_dbrec_offset)); + if (txq_obj->cq_umem) + claim_zero(mlx5_glue->devx_umem_dereg(txq_obj->cq_umem)); + if (txq_obj->cq_buf) + rte_free(txq_obj->cq_buf); +} + +/** + * Create the Tx queue DevX object. + * + * @param dev + * Pointer to Ethernet device. + * @param idx + * Queue index in DPDK Tx queue array + * + * @return + * The DevX object initialised, NULL otherwise and rte_errno is set. + */ +static struct mlx5_txq_obj * +mlx5_txq_obj_devx_new(struct rte_eth_dev *dev, uint16_t idx) +{ +#ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET + DRV_LOG(ERR, "port %u Tx queue %u cannot create with DevX, no UAR", + dev->data->port_id, idx); + rte_errno = ENOMEM; + return NULL; +#else + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_dev_ctx_shared *sh = priv->sh; + struct mlx5_txq_data *txq_data = (*priv->txqs)[idx]; + struct mlx5_txq_ctrl *txq_ctrl = + container_of(txq_data, struct mlx5_txq_ctrl, txq); + struct mlx5_devx_create_sq_attr sq_attr = { 0 }; + struct mlx5_devx_modify_sq_attr msq_attr = { 0 }; + struct mlx5_devx_cq_attr cq_attr = { 0 }; + struct mlx5_txq_obj *txq_obj = NULL; + size_t page_size = sysconf(_SC_PAGESIZE); + struct mlx5_cqe *cqe; + uint32_t i, nqe; + int ret = 0; + + MLX5_ASSERT(txq_data); + MLX5_ASSERT(!txq_ctrl->obj); + txq_obj = rte_calloc_socket(__func__, 1, + sizeof(struct mlx5_txq_obj), 0, + txq_ctrl->socket); + if (!txq_obj) { + DRV_LOG(ERR, + "port %u Tx queue %u cannot allocate memory resources", + dev->data->port_id, txq_data->idx); + rte_errno = ENOMEM; + return NULL; + } + txq_obj->type = MLX5_TXQ_OBJ_TYPE_DEVX_SQ; + txq_obj->txq_ctrl = txq_ctrl; + txq_obj->dev = dev; + /* Create the Completion Queue. */ + nqe = (1UL << txq_data->elts_n) / MLX5_TX_COMP_THRESH + + 1 + MLX5_TX_COMP_THRESH_INLINE_DIV; + nqe = 1UL << log2above(nqe); + if (nqe > UINT16_MAX) { + DRV_LOG(ERR, + "port %u Tx queue %u requests to many CQEs %u", + dev->data->port_id, txq_data->idx, nqe); + rte_errno = EINVAL; + goto error; + } + /* Allocate memory buffer for CQEs. */ + txq_obj->cq_buf = rte_zmalloc_socket(__func__, + nqe * sizeof(struct mlx5_cqe), + MLX5_CQE_BUF_ALIGNMENT, + sh->numa_node); + if (!txq_obj->cq_buf) { + DRV_LOG(ERR, + "port %u Tx queue %u cannot allocate memory (CQ)", + dev->data->port_id, txq_data->idx); + rte_errno = ENOMEM; + goto error; + } + txq_data->cqe_n = log2above(nqe); + txq_data->cqe_s = 1 << txq_data->cqe_n; + txq_data->cqe_m = txq_data->cqe_s - 1; + txq_data->cqes = (volatile struct mlx5_cqe *)txq_obj->cq_buf; + txq_data->cq_ci = 0; + txq_data->cq_pi = 0; + /* Register allocated buffer in user space with DevX. */ + txq_obj->cq_umem = mlx5_glue->devx_umem_reg + (sh->ctx, + (void *)txq_obj->cq_buf, + nqe * sizeof(struct mlx5_cqe), + IBV_ACCESS_LOCAL_WRITE); + if (!txq_obj->cq_umem) { + rte_errno = errno; + DRV_LOG(ERR, + "port %u Tx queue %u cannot register memory (CQ)", + dev->data->port_id, txq_data->idx); + goto error; + } + /* Allocate doorbell record for completion queue. */ + txq_obj->cq_dbrec_offset = mlx5_get_dbr(sh->ctx, + &priv->dbrpgs, + &txq_obj->cq_dbrec_page); + if (txq_obj->cq_dbrec_offset < 0) + goto error; + txq_data->cq_db = (volatile uint32_t *)(txq_obj->cq_dbrec_page->dbrs + + txq_obj->cq_dbrec_offset); + *txq_data->cq_db = 0; + /* Create completion queue object with DevX. */ + cq_attr.cqe_size = (sizeof(struct mlx5_cqe) == 128) ? + MLX5_CQE_SIZE_128B : MLX5_CQE_SIZE_64B; + cq_attr.uar_page_id = sh->tx_uar->page_id; + cq_attr.eqn = sh->txpp.eqn; + cq_attr.q_umem_valid = 1; + cq_attr.q_umem_offset = (uintptr_t)txq_obj->cq_buf % page_size; + cq_attr.q_umem_id = txq_obj->cq_umem->umem_id; + cq_attr.db_umem_valid = 1; + cq_attr.db_umem_offset = txq_obj->cq_dbrec_offset; + cq_attr.db_umem_id = mlx5_os_get_umem_id(txq_obj->cq_dbrec_page->umem); + cq_attr.log_cq_size = rte_log2_u32(nqe); + cq_attr.log_page_size = rte_log2_u32(page_size); + txq_obj->cq_devx = mlx5_devx_cmd_create_cq(sh->ctx, &cq_attr); + if (!txq_obj->cq_devx) { + rte_errno = errno; + DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure", + dev->data->port_id, idx); + goto error; + } + /* Initial fill CQ buffer with invalid CQE opcode. */ + cqe = (struct mlx5_cqe *)txq_obj->cq_buf; + for (i = 0; i < txq_data->cqe_s; i++) { + cqe->op_own = (MLX5_CQE_INVALID << 4) | MLX5_CQE_OWNER_MASK; + ++cqe; + } + /* Create the Work Queue. */ + nqe = RTE_MIN(1UL << txq_data->elts_n, + (uint32_t)sh->device_attr.max_qp_wr); + txq_obj->sq_buf = rte_zmalloc_socket(__func__, + nqe * sizeof(struct mlx5_wqe), + page_size, + sh->numa_node); + if (!txq_obj->sq_buf) { + DRV_LOG(ERR, + "port %u Tx queue %u cannot allocate memory (SQ)", + dev->data->port_id, txq_data->idx); + rte_errno = ENOMEM; + goto error; + } + txq_data->wqe_n = log2above(nqe); + txq_data->wqe_s = 1 << txq_data->wqe_n; + txq_data->wqe_m = txq_data->wqe_s - 1; + txq_data->wqes = (struct mlx5_wqe *)txq_obj->sq_buf; + txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s; + txq_data->wqe_ci = 0; + txq_data->wqe_pi = 0; + txq_data->wqe_comp = 0; + txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV; + /* Register allocated buffer in user space with DevX. */ + txq_obj->sq_umem = mlx5_glue->devx_umem_reg + (sh->ctx, + (void *)txq_obj->sq_buf, + nqe * sizeof(struct mlx5_wqe), + IBV_ACCESS_LOCAL_WRITE); + if (!txq_obj->sq_umem) { + rte_errno = errno; + DRV_LOG(ERR, + "port %u Tx queue %u cannot register memory (SQ)", + dev->data->port_id, txq_data->idx); + goto error; + } + /* Allocate doorbell record for completion queue. */ + txq_obj->cq_dbrec_offset = mlx5_get_dbr(sh->ctx, + &priv->dbrpgs, + &txq_obj->sq_dbrec_page); + if (txq_obj->sq_dbrec_offset < 0) + goto error; + txq_data->qp_db = (volatile uint32_t *) + (txq_obj->sq_dbrec_page->dbrs + + txq_obj->sq_dbrec_offset + + MLX5_SND_DBR * sizeof(uint32_t)); + *txq_data->qp_db = 0; + /* Create Send Queue object with DevX. */ + sq_attr.tis_lst_sz = 1; + sq_attr.tis_num = sh->tis->id; + sq_attr.state = MLX5_SQC_STATE_RST; + sq_attr.cqn = txq_obj->cq_devx->id; + sq_attr.flush_in_error_en = 1; + sq_attr.allow_multi_pkt_send_wqe = !!priv->config.mps; + sq_attr.allow_swp = !!priv->config.swp; + sq_attr.min_wqe_inline_mode = priv->config.hca_attr.vport_inline_mode; + sq_attr.wq_attr.uar_page = sh->tx_uar->page_id; + sq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC; + sq_attr.wq_attr.pd = sh->pdn; + sq_attr.wq_attr.log_wq_stride = rte_log2_u32(MLX5_WQE_SIZE); + sq_attr.wq_attr.log_wq_sz = txq_data->wqe_n; + sq_attr.wq_attr.dbr_umem_valid = 1; + sq_attr.wq_attr.dbr_addr = txq_obj->cq_dbrec_offset; + sq_attr.wq_attr.dbr_umem_id = + mlx5_os_get_umem_id(txq_obj->cq_dbrec_page->umem); + sq_attr.wq_attr.wq_umem_valid = 1; + sq_attr.wq_attr.wq_umem_id = txq_obj->sq_umem->umem_id; + sq_attr.wq_attr.wq_umem_offset = (uintptr_t)txq_obj->sq_buf % page_size; + txq_obj->sq_devx = mlx5_devx_cmd_create_sq(sh->ctx, &sq_attr); + if (!txq_obj->sq_devx) { + rte_errno = errno; + DRV_LOG(ERR, "port %u Tx queue %u SQ creation failure", + dev->data->port_id, idx); + goto error; + } + txq_data->qp_num_8s = txq_obj->sq_devx->id << 8; + /* Change Send Queue state to Ready-to-Send. */ + msq_attr.sq_state = MLX5_SQC_STATE_RST; + msq_attr.state = MLX5_SQC_STATE_RDY; + ret = mlx5_devx_cmd_modify_sq(txq_obj->sq_devx, &msq_attr); + if (ret) { + rte_errno = errno; + DRV_LOG(ERR, + "port %u Tx queue %u SP state to SQC_STATE_RDY failed", + dev->data->port_id, idx); + goto error; + } + txq_data->fcqs = rte_calloc_socket(__func__, + txq_data->cqe_s, + sizeof(*txq_data->fcqs), + RTE_CACHE_LINE_SIZE, + txq_ctrl->socket); + if (!txq_data->fcqs) { + DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory (FCQ)", + dev->data->port_id, idx); + rte_errno = ENOMEM; + goto error; + } +#ifdef HAVE_IBV_FLOW_DV_SUPPORT + /* + * If using DevX need to query and store TIS transport domain value. + * This is done once per port. + * Will use this value on Rx, when creating matching TIR. + */ + if (priv->config.devx && !priv->sh->tdn) + priv->sh->tdn = priv->sh->td->id; +#endif + MLX5_ASSERT(sh->tx_uar); + MLX5_ASSERT(sh->tx_uar->reg_addr); + txq_ctrl->bf_reg = sh->tx_uar->reg_addr; + txq_ctrl->uar_mmap_offset = sh->tx_uar->mmap_off; + rte_atomic32_set(&txq_obj->refcnt, 1); + txq_uar_init(txq_ctrl); + LIST_INSERT_HEAD(&priv->txqsobj, txq_obj, next); + return txq_obj; +error: + ret = rte_errno; /* Save rte_errno before cleanup. */ + txq_release_sq_resources(txq_obj); + if (txq_data->fcqs) { + rte_free(txq_data->fcqs); + txq_data->fcqs = NULL; + } + rte_free(txq_obj); + rte_errno = ret; /* Restore rte_errno. */ + return NULL; +#endif +} + /** * Create the Tx queue Verbs object. * @@ -371,19 +873,22 @@ error: * Pointer to Ethernet device. * @param idx * Queue index in DPDK Tx queue array. + * @param type + * Type of the Tx queue object to create. * * @return * The Verbs object initialised, NULL otherwise and rte_errno is set. */ -struct mlx5_txq_ibv * -mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) +struct mlx5_txq_obj * +mlx5_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx, + enum mlx5_txq_obj_type type) { struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_data *txq_data = (*priv->txqs)[idx]; struct mlx5_txq_ctrl *txq_ctrl = container_of(txq_data, struct mlx5_txq_ctrl, txq); - struct mlx5_txq_ibv tmpl; - struct mlx5_txq_ibv *txq_ibv = NULL; + struct mlx5_txq_obj tmpl; + struct mlx5_txq_obj *txq_obj = NULL; union { struct ibv_qp_init_attr_ex init; struct ibv_cq_init_attr_ex cq; @@ -396,12 +901,16 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) const int desc = 1 << txq_data->elts_n; int ret = 0; + if (type == MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN) + return mlx5_txq_obj_hairpin_new(dev, idx); + if (type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ) + return mlx5_txq_obj_devx_new(dev, idx); #ifdef HAVE_IBV_FLOW_DV_SUPPORT /* If using DevX, need additional mask to read tisn value. */ if (priv->config.devx && !priv->sh->tdn) qp.comp_mask |= MLX5DV_QP_MASK_RAW_QP_HANDLES; #endif - assert(txq_data); + MLX5_ASSERT(txq_data); priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_TX_QUEUE; priv->verbs_alloc_ctx.obj = txq_ctrl; if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) { @@ -411,7 +920,7 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) rte_errno = EINVAL; return NULL; } - memset(&tmpl, 0, sizeof(struct mlx5_txq_ibv)); + memset(&tmpl, 0, sizeof(struct mlx5_txq_obj)); attr.cq = (struct ibv_cq_init_attr_ex){ .comp_mask = 0, }; @@ -432,9 +941,9 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) .cap = { /* Max number of outstanding WRs. */ .max_send_wr = - ((priv->sh->device_attr.orig_attr.max_qp_wr < + ((priv->sh->device_attr.max_qp_wr < desc) ? - priv->sh->device_attr.orig_attr.max_qp_wr : + priv->sh->device_attr.max_qp_wr : desc), /* * Max number of scatter/gather elements in a WR, @@ -471,7 +980,7 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) /* Move the QP to this state. */ .qp_state = IBV_QPS_INIT, /* IB device port number. */ - .port_num = (uint8_t)priv->ibv_port, + .port_num = (uint8_t)priv->dev_port, }; ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, (IBV_QP_STATE | IBV_QP_PORT)); @@ -502,9 +1011,9 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) rte_errno = errno; goto error; } - txq_ibv = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_txq_ibv), 0, + txq_obj = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_txq_obj), 0, txq_ctrl->socket); - if (!txq_ibv) { + if (!txq_obj) { DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory", dev->data->port_id, idx); rte_errno = ENOMEM; @@ -540,13 +1049,22 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) txq_data->cq_db = cq_info.dbrec; txq_data->cqes = (volatile struct mlx5_cqe *)cq_info.buf; txq_data->cq_ci = 0; -#ifndef NDEBUG txq_data->cq_pi = 0; -#endif txq_data->wqe_ci = 0; txq_data->wqe_pi = 0; txq_data->wqe_comp = 0; txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV; + txq_data->fcqs = rte_calloc_socket(__func__, + txq_data->cqe_s, + sizeof(*txq_data->fcqs), + RTE_CACHE_LINE_SIZE, + txq_ctrl->socket); + if (!txq_data->fcqs) { + DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory (FCQ)", + dev->data->port_id, idx); + rte_errno = ENOMEM; + goto error; + } #ifdef HAVE_IBV_FLOW_DV_SUPPORT /* * If using DevX need to query and store TIS transport domain value. @@ -568,11 +1086,10 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) } } #endif - txq_ibv->qp = tmpl.qp; - txq_ibv->cq = tmpl.cq; - rte_atomic32_inc(&txq_ibv->refcnt); + txq_obj->qp = tmpl.qp; + txq_obj->cq = tmpl.cq; + rte_atomic32_inc(&txq_obj->refcnt); txq_ctrl->bf_reg = qp.bf.reg; - txq_uar_init(txq_ctrl); if (qp.comp_mask & MLX5DV_QP_MASK_UAR_MMAP_OFFSET) { txq_ctrl->uar_mmap_offset = qp.uar_mmap_offset; DRV_LOG(DEBUG, "port %u: uar_mmap_offset 0x%"PRIx64, @@ -585,18 +1102,23 @@ mlx5_txq_ibv_new(struct rte_eth_dev *dev, uint16_t idx) rte_errno = EINVAL; goto error; } - LIST_INSERT_HEAD(&priv->txqsibv, txq_ibv, next); - txq_ibv->txq_ctrl = txq_ctrl; + txq_uar_init(txq_ctrl); + LIST_INSERT_HEAD(&priv->txqsobj, txq_obj, next); + txq_obj->txq_ctrl = txq_ctrl; priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE; - return txq_ibv; + return txq_obj; error: ret = rte_errno; /* Save rte_errno before cleanup. */ if (tmpl.cq) claim_zero(mlx5_glue->destroy_cq(tmpl.cq)); if (tmpl.qp) claim_zero(mlx5_glue->destroy_qp(tmpl.qp)); - if (txq_ibv) - rte_free(txq_ibv); + if (txq_data && txq_data->fcqs) { + rte_free(txq_data->fcqs); + txq_data->fcqs = NULL; + } + if (txq_obj) + rte_free(txq_obj); priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE; rte_errno = ret; /* Restore rte_errno. */ return NULL; @@ -613,8 +1135,8 @@ error: * @return * The Verbs object if it exists. */ -struct mlx5_txq_ibv * -mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx) +struct mlx5_txq_obj * +mlx5_txq_obj_get(struct rte_eth_dev *dev, uint16_t idx) { struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_ctrl *txq_ctrl; @@ -624,29 +1146,40 @@ mlx5_txq_ibv_get(struct rte_eth_dev *dev, uint16_t idx) if (!(*priv->txqs)[idx]) return NULL; txq_ctrl = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq); - if (txq_ctrl->ibv) - rte_atomic32_inc(&txq_ctrl->ibv->refcnt); - return txq_ctrl->ibv; + if (txq_ctrl->obj) + rte_atomic32_inc(&txq_ctrl->obj->refcnt); + return txq_ctrl->obj; } /** * Release an Tx verbs queue object. * - * @param txq_ibv + * @param txq_obj * Verbs Tx queue object. * * @return * 1 while a reference on it exists, 0 when freed. */ int -mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv) +mlx5_txq_obj_release(struct mlx5_txq_obj *txq_obj) { - assert(txq_ibv); - if (rte_atomic32_dec_and_test(&txq_ibv->refcnt)) { - claim_zero(mlx5_glue->destroy_qp(txq_ibv->qp)); - claim_zero(mlx5_glue->destroy_cq(txq_ibv->cq)); - LIST_REMOVE(txq_ibv, next); - rte_free(txq_ibv); + MLX5_ASSERT(txq_obj); + if (rte_atomic32_dec_and_test(&txq_obj->refcnt)) { + if (txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN) { + if (txq_obj->tis) + claim_zero(mlx5_devx_cmd_destroy(txq_obj->tis)); + } else if (txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ) { + txq_release_sq_resources(txq_obj); + } else { + claim_zero(mlx5_glue->destroy_qp(txq_obj->qp)); + claim_zero(mlx5_glue->destroy_cq(txq_obj->cq)); + } + if (txq_obj->txq_ctrl->txq.fcqs) { + rte_free(txq_obj->txq_ctrl->txq.fcqs); + txq_obj->txq_ctrl->txq.fcqs = NULL; + } + LIST_REMOVE(txq_obj, next); + rte_free(txq_obj); return 0; } return 1; @@ -662,15 +1195,15 @@ mlx5_txq_ibv_release(struct mlx5_txq_ibv *txq_ibv) * The number of object not released. */ int -mlx5_txq_ibv_verify(struct rte_eth_dev *dev) +mlx5_txq_obj_verify(struct rte_eth_dev *dev) { struct mlx5_priv *priv = dev->data->dev_private; int ret = 0; - struct mlx5_txq_ibv *txq_ibv; + struct mlx5_txq_obj *txq_obj; - LIST_FOREACH(txq_ibv, &priv->txqsibv, next) { + LIST_FOREACH(txq_obj, &priv->txqsobj, next) { DRV_LOG(DEBUG, "port %u Verbs Tx queue %u still referenced", - dev->data->port_id, txq_ibv->txq_ctrl->txq.idx); + dev->data->port_id, txq_obj->txq_ctrl->txq.idx); ++ret; } return ret; @@ -701,6 +1234,38 @@ txq_calc_wqebb_cnt(struct mlx5_txq_ctrl *txq_ctrl) return rte_align32pow2(wqe_size * desc) / MLX5_WQE_SIZE; } +/** + * Calculate the maximal inline data size for Tx queue. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + * + * @return + * The maximal inline data size. + */ +static unsigned int +txq_calc_inline_max(struct mlx5_txq_ctrl *txq_ctrl) +{ + const unsigned int desc = 1 << txq_ctrl->txq.elts_n; + struct mlx5_priv *priv = txq_ctrl->priv; + unsigned int wqe_size; + + wqe_size = priv->sh->device_attr.max_qp_wr / desc; + if (!wqe_size) + return 0; + /* + * This calculation is derived from tthe source of + * mlx5_calc_send_wqe() in rdma_core library. + */ + wqe_size = wqe_size * MLX5_WQE_SIZE - + MLX5_WQE_CSEG_SIZE - + MLX5_WQE_ESEG_SIZE - + MLX5_WSEG_SIZE - + MLX5_WSEG_SIZE + + MLX5_DSEG_MIN_INLINE_SIZE; + return wqe_size; +} + /** * Set Tx queue parameters from device configuration. * @@ -728,7 +1293,7 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) if (config->txqs_inline == MLX5_ARG_UNSET) txqs_inline = #if defined(RTE_ARCH_ARM64) - (priv->sh->pci_dev->id.device_id == + (priv->pci_dev->id.device_id == PCI_DEVICE_ID_MELLANOX_CONNECTX5BF) ? MLX5_INLINE_MAX_TXQS_BLUEFIELD : #endif @@ -743,13 +1308,13 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) (unsigned int)config->txq_inline_mpw; inlen_mode = (config->txq_inline_min == MLX5_ARG_UNSET) ? 0 : (unsigned int)config->txq_inline_min; - if (config->mps != MLX5_MPW_ENHANCED) + if (config->mps != MLX5_MPW_ENHANCED && config->mps != MLX5_MPW) inlen_empw = 0; /* * If there is requested minimal amount of data to inline * we MUST enable inlining. This is a case for ConnectX-4 * which usually requires L2 inlined for correct operating - * and ConnectX-4LX which requires L2-L4 inlined to + * and ConnectX-4 Lx which requires L2-L4 inlined to * support E-Switch Flows. */ if (inlen_mode) { @@ -784,20 +1349,21 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) txq_ctrl->txq.vlan_en = config->hw_vlan_insert; vlan_inline = (dev_txoff & DEV_TX_OFFLOAD_VLAN_INSERT) && !config->hw_vlan_insert; - if (vlan_inline) - inlen_send = RTE_MAX(inlen_send, MLX5_ESEG_MIN_INLINE_SIZE); /* * If there are few Tx queues it is prioritized * to save CPU cycles and disable data inlining at all. */ - if ((inlen_send && priv->txqs_n >= txqs_inline) || vlan_inline) { + if (inlen_send && priv->txqs_n >= txqs_inline) { /* * The data sent with ordinal MLX5_OPCODE_SEND * may be inlined in Ethernet Segment, align the * length accordingly to fit entire WQEBBs. */ - temp = (inlen_send / MLX5_WQE_SIZE) * MLX5_WQE_SIZE + - MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE; + temp = RTE_MAX(inlen_send, + MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE); + temp -= MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE; + temp = RTE_ALIGN(temp, MLX5_WQE_SIZE); + temp += MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE; temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX + MLX5_ESEG_MIN_INLINE_SIZE - MLX5_WQE_CSEG_SIZE - @@ -819,47 +1385,48 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) * beginning of inlining buffer in Ethernet * Segment. */ - assert(inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE); - assert(inlen_send <= MLX5_WQE_SIZE_MAX + - MLX5_ESEG_MIN_INLINE_SIZE - - MLX5_WQE_CSEG_SIZE - - MLX5_WQE_ESEG_SIZE - - MLX5_WQE_DSEG_SIZE * 2); - txq_ctrl->txq.inlen_send = inlen_send; - txq_ctrl->txq.inlen_mode = inlen_mode; - txq_ctrl->txq.inlen_empw = 0; - } else { + MLX5_ASSERT(inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_ASSERT(inlen_send <= MLX5_WQE_SIZE_MAX + + MLX5_ESEG_MIN_INLINE_SIZE - + MLX5_WQE_CSEG_SIZE - + MLX5_WQE_ESEG_SIZE - + MLX5_WQE_DSEG_SIZE * 2); + } else if (inlen_mode) { /* * If minimal inlining is requested we must * enable inlining in general, despite the - * number of configured queues. + * number of configured queues. Ignore the + * txq_inline_max devarg, this is not + * full-featured inline. */ inlen_send = inlen_mode; - if (inlen_mode) { - /* - * Extend space for inline data to allow - * optional alignment of data buffer - * start address, it may improve PCIe - * performance. - */ - inlen_send = RTE_MIN(inlen_send + MLX5_WQE_SIZE, - MLX5_SEND_MAX_INLINE_LEN); - } - txq_ctrl->txq.inlen_send = inlen_send; - txq_ctrl->txq.inlen_mode = inlen_mode; - txq_ctrl->txq.inlen_empw = 0; + inlen_empw = 0; + } else if (vlan_inline) { + /* + * Hardware does not report offload for + * VLAN insertion, we must enable data inline + * to implement feature by software. + */ + inlen_send = MLX5_ESEG_MIN_INLINE_SIZE; + inlen_empw = 0; + } else { inlen_send = 0; inlen_empw = 0; } + txq_ctrl->txq.inlen_send = inlen_send; + txq_ctrl->txq.inlen_mode = inlen_mode; + txq_ctrl->txq.inlen_empw = 0; if (inlen_send && inlen_empw && priv->txqs_n >= txqs_inline) { /* * The data sent with MLX5_OPCODE_ENHANCED_MPSW * may be inlined in Data Segment, align the * length accordingly to fit entire WQEBBs. */ - temp = (inlen_empw + MLX5_WQE_SIZE - 1) / MLX5_WQE_SIZE; - temp = temp * MLX5_WQE_SIZE + - MLX5_DSEG_MIN_INLINE_SIZE - MLX5_WQE_DSEG_SIZE; + temp = RTE_MAX(inlen_empw, + MLX5_WQE_SIZE + MLX5_DSEG_MIN_INLINE_SIZE); + temp -= MLX5_DSEG_MIN_INLINE_SIZE; + temp = RTE_ALIGN(temp, MLX5_WQE_SIZE); + temp += MLX5_DSEG_MIN_INLINE_SIZE; temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX + MLX5_DSEG_MIN_INLINE_SIZE - MLX5_WQE_CSEG_SIZE - @@ -873,12 +1440,12 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) PORT_ID(priv), inlen_empw, temp); inlen_empw = temp; } - assert(inlen_empw >= MLX5_ESEG_MIN_INLINE_SIZE); - assert(inlen_empw <= MLX5_WQE_SIZE_MAX + - MLX5_DSEG_MIN_INLINE_SIZE - - MLX5_WQE_CSEG_SIZE - - MLX5_WQE_ESEG_SIZE - - MLX5_WQE_DSEG_SIZE); + MLX5_ASSERT(inlen_empw >= MLX5_ESEG_MIN_INLINE_SIZE); + MLX5_ASSERT(inlen_empw <= MLX5_WQE_SIZE_MAX + + MLX5_DSEG_MIN_INLINE_SIZE - + MLX5_WQE_CSEG_SIZE - + MLX5_WQE_ESEG_SIZE - + MLX5_WQE_DSEG_SIZE); txq_ctrl->txq.inlen_empw = inlen_empw; } txq_ctrl->max_inline_data = RTE_MAX(inlen_send, inlen_empw); @@ -895,6 +1462,115 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) txq_ctrl->txq.offloads) && config->swp; } +/** + * Adjust Tx queue data inline parameters for large queue sizes. + * The data inline feature requires multiple WQEs to fit the packets, + * and if the large amount of Tx descriptors is requested by application + * the total WQE amount may exceed the hardware capabilities. If the + * default inline setting are used we can try to adjust these ones and + * meet the hardware requirements and not exceed the queue size. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + * + * @return + * Zero on success, otherwise the parameters can not be adjusted. + */ +static int +txq_adjust_params(struct mlx5_txq_ctrl *txq_ctrl) +{ + struct mlx5_priv *priv = txq_ctrl->priv; + struct mlx5_dev_config *config = &priv->config; + unsigned int max_inline; + + max_inline = txq_calc_inline_max(txq_ctrl); + if (!txq_ctrl->txq.inlen_send) { + /* + * Inline data feature is not engaged at all. + * There is nothing to adjust. + */ + return 0; + } + if (txq_ctrl->max_inline_data <= max_inline) { + /* + * The requested inline data length does not + * exceed queue capabilities. + */ + return 0; + } + if (txq_ctrl->txq.inlen_mode > max_inline) { + DRV_LOG(ERR, + "minimal data inline requirements (%u) are not" + " satisfied (%u) on port %u, try the smaller" + " Tx queue size (%d)", + txq_ctrl->txq.inlen_mode, max_inline, + priv->dev_data->port_id, + priv->sh->device_attr.max_qp_wr); + goto error; + } + if (txq_ctrl->txq.inlen_send > max_inline && + config->txq_inline_max != MLX5_ARG_UNSET && + config->txq_inline_max > (int)max_inline) { + DRV_LOG(ERR, + "txq_inline_max requirements (%u) are not" + " satisfied (%u) on port %u, try the smaller" + " Tx queue size (%d)", + txq_ctrl->txq.inlen_send, max_inline, + priv->dev_data->port_id, + priv->sh->device_attr.max_qp_wr); + goto error; + } + if (txq_ctrl->txq.inlen_empw > max_inline && + config->txq_inline_mpw != MLX5_ARG_UNSET && + config->txq_inline_mpw > (int)max_inline) { + DRV_LOG(ERR, + "txq_inline_mpw requirements (%u) are not" + " satisfied (%u) on port %u, try the smaller" + " Tx queue size (%d)", + txq_ctrl->txq.inlen_empw, max_inline, + priv->dev_data->port_id, + priv->sh->device_attr.max_qp_wr); + goto error; + } + if (txq_ctrl->txq.tso_en && max_inline < MLX5_MAX_TSO_HEADER) { + DRV_LOG(ERR, + "tso header inline requirements (%u) are not" + " satisfied (%u) on port %u, try the smaller" + " Tx queue size (%d)", + MLX5_MAX_TSO_HEADER, max_inline, + priv->dev_data->port_id, + priv->sh->device_attr.max_qp_wr); + goto error; + } + if (txq_ctrl->txq.inlen_send > max_inline) { + DRV_LOG(WARNING, + "adjust txq_inline_max (%u->%u)" + " due to large Tx queue on port %u", + txq_ctrl->txq.inlen_send, max_inline, + priv->dev_data->port_id); + txq_ctrl->txq.inlen_send = max_inline; + } + if (txq_ctrl->txq.inlen_empw > max_inline) { + DRV_LOG(WARNING, + "adjust txq_inline_mpw (%u->%u)" + "due to large Tx queue on port %u", + txq_ctrl->txq.inlen_empw, max_inline, + priv->dev_data->port_id); + txq_ctrl->txq.inlen_empw = max_inline; + } + txq_ctrl->max_inline_data = RTE_MAX(txq_ctrl->txq.inlen_send, + txq_ctrl->txq.inlen_empw); + MLX5_ASSERT(txq_ctrl->max_inline_data <= max_inline); + MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= max_inline); + MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_send); + MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_empw || + !txq_ctrl->txq.inlen_empw); + return 0; +error: + rte_errno = ENOMEM; + return -ENOMEM; +} + /** * Create a DPDK Tx queue. * @@ -933,8 +1609,8 @@ mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, goto error; } /* Save pointer of global generation number to check memory event. */ - tmpl->txq.mr_ctrl.dev_gen_ptr = &priv->sh->mr.dev_gen; - assert(desc > MLX5_TX_COMP_THRESH); + tmpl->txq.mr_ctrl.dev_gen_ptr = &priv->sh->share_cache.dev_gen; + MLX5_ASSERT(desc > MLX5_TX_COMP_THRESH); tmpl->txq.offloads = conf->offloads | dev->data->dev_conf.txmode.offloads; tmpl->priv = priv; @@ -945,17 +1621,20 @@ mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, tmpl->txq.port_id = dev->data->port_id; tmpl->txq.idx = idx; txq_set_params(tmpl); + if (txq_adjust_params(tmpl)) + goto error; if (txq_calc_wqebb_cnt(tmpl) > - priv->sh->device_attr.orig_attr.max_qp_wr) { + priv->sh->device_attr.max_qp_wr) { DRV_LOG(ERR, "port %u Tx WQEBB count (%d) exceeds the limit (%d)," " try smaller queue size", dev->data->port_id, txq_calc_wqebb_cnt(tmpl), - priv->sh->device_attr.orig_attr.max_qp_wr); + priv->sh->device_attr.max_qp_wr); rte_errno = ENOMEM; goto error; } rte_atomic32_inc(&tmpl->refcnt); + tmpl->type = MLX5_TXQ_TYPE_STANDARD; LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next); return tmpl; error: @@ -963,6 +1642,46 @@ error: return NULL; } +/** + * Create a DPDK Tx hairpin queue. + * + * @param dev + * Pointer to Ethernet device. + * @param idx + * TX queue index. + * @param desc + * Number of descriptors to configure in queue. + * @param hairpin_conf + * The hairpin configuration. + * + * @return + * A DPDK queue object on success, NULL otherwise and rte_errno is set. + */ +struct mlx5_txq_ctrl * +mlx5_txq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc, + const struct rte_eth_hairpin_conf *hairpin_conf) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_ctrl *tmpl; + + tmpl = rte_calloc_socket("TXQ", 1, + sizeof(*tmpl), 0, SOCKET_ID_ANY); + if (!tmpl) { + rte_errno = ENOMEM; + return NULL; + } + tmpl->priv = priv; + tmpl->socket = SOCKET_ID_ANY; + tmpl->txq.elts_n = log2above(desc); + tmpl->txq.port_id = dev->data->port_id; + tmpl->txq.idx = idx; + tmpl->hairpin_conf = *hairpin_conf; + tmpl->type = MLX5_TXQ_TYPE_HAIRPIN; + rte_atomic32_inc(&tmpl->refcnt); + LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next); + return tmpl; +} + /** * Get a Tx queue. * @@ -983,7 +1702,7 @@ mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx) if ((*priv->txqs)[idx]) { ctrl = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq); - mlx5_txq_ibv_get(dev, idx); + mlx5_txq_obj_get(dev, idx); rte_atomic32_inc(&ctrl->refcnt); } return ctrl; @@ -1009,8 +1728,8 @@ mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx) if (!(*priv->txqs)[idx]) return 0; txq = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq); - if (txq->ibv && !mlx5_txq_ibv_release(txq->ibv)) - txq->ibv = NULL; + if (txq->obj && !mlx5_txq_obj_release(txq->obj)) + txq->obj = NULL; if (rte_atomic32_dec_and_test(&txq->refcnt)) { txq_free_elts(txq); mlx5_mr_btree_free(&txq->txq.mr_ctrl.cache_bh); @@ -1068,3 +1787,35 @@ mlx5_txq_verify(struct rte_eth_dev *dev) } return ret; } + +/** + * Set the Tx queue dynamic timestamp (mask and offset) + * + * @param[in] dev + * Pointer to the Ethernet device structure. + */ +void +mlx5_txq_dynf_timestamp_set(struct rte_eth_dev *dev) +{ + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_dev_ctx_shared *sh = priv->sh; + struct mlx5_txq_data *data; + int off, nbit; + unsigned int i; + uint64_t mask = 0; + + nbit = rte_mbuf_dynflag_lookup + (RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL); + off = rte_mbuf_dynfield_lookup + (RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL); + if (nbit > 0 && off >= 0 && sh->txpp.refcnt) + mask = 1ULL << nbit; + for (i = 0; i != priv->txqs_n; ++i) { + data = (*priv->txqs)[i]; + if (!data) + continue; + data->sh = sh; + data->ts_mask = mask; + data->ts_offset = off; + } +}