X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fmlx5%2Fmlx5_txq.c;h=97991f0c04f9d0b14a134f2c2561e2990f804c4f;hb=9bf26e1318e3cd7a04115c8201255505e90cab83;hp=23002cf86ba6aedcc1d05e5a6d6611e88f840e0c;hpb=8fd92a66c60a7310cf5ab91996b9b09447512a61;p=dpdk.git diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c index 23002cf86b..97991f0c04 100644 --- a/drivers/net/mlx5/mlx5_txq.c +++ b/drivers/net/mlx5/mlx5_txq.c @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: BSD-3-Clause * Copyright 2015 6WIND S.A. - * Copyright 2015 Mellanox. + * Copyright 2015 Mellanox Technologies, Ltd */ #include @@ -10,6 +10,7 @@ #include #include #include +#include /* Verbs header. */ /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */ @@ -46,8 +47,9 @@ txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl) unsigned int i; for (i = 0; (i != elts_n); ++i) - (*txq_ctrl->txq.elts)[i] = NULL; - DEBUG("%p: allocated and configured %u WRs", (void *)txq_ctrl, elts_n); + txq_ctrl->txq.elts[i] = NULL; + DRV_LOG(DEBUG, "port %u Tx queue %u allocated and configured %u WRs", + PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx, elts_n); txq_ctrl->txq.elts_head = 0; txq_ctrl->txq.elts_tail = 0; txq_ctrl->txq.elts_comp = 0; @@ -66,9 +68,10 @@ txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl) const uint16_t elts_m = elts_n - 1; uint16_t elts_head = txq_ctrl->txq.elts_head; uint16_t elts_tail = txq_ctrl->txq.elts_tail; - struct rte_mbuf *(*elts)[elts_n] = txq_ctrl->txq.elts; + struct rte_mbuf *(*elts)[elts_n] = &txq_ctrl->txq.elts; - DEBUG("%p: freeing WRs", (void *)txq_ctrl); + DRV_LOG(DEBUG, "port %u Tx queue %u freeing WRs", + PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx); txq_ctrl->txq.elts_head = 0; txq_ctrl->txq.elts_tail = 0; txq_ctrl->txq.elts_comp = 0; @@ -91,15 +94,16 @@ txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl) /** * Returns the per-port supported offloads. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * * @return * Supported Tx offloads. */ uint64_t -mlx5_priv_get_tx_port_offloads(struct priv *priv) +mlx5_get_tx_port_offloads(struct rte_eth_dev *dev) { + struct mlx5_priv *priv = dev->data->dev_private; uint64_t offloads = (DEV_TX_OFFLOAD_MULTI_SEGS | DEV_TX_OFFLOAD_VLAN_INSERT); struct mlx5_dev_config *config = &priv->config; @@ -110,6 +114,13 @@ mlx5_priv_get_tx_port_offloads(struct priv *priv) DEV_TX_OFFLOAD_TCP_CKSUM); if (config->tso) offloads |= DEV_TX_OFFLOAD_TCP_TSO; + if (config->swp) { + if (config->hw_csum) + offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM; + if (config->tso) + offloads |= (DEV_TX_OFFLOAD_IP_TNL_TSO | + DEV_TX_OFFLOAD_UDP_TNL_TSO); + } if (config->tunnel_en) { if (config->hw_csum) offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM; @@ -121,30 +132,55 @@ mlx5_priv_get_tx_port_offloads(struct priv *priv) } /** - * Checks if the per-queue offload configuration is valid. + * Tx queue presetup checks. * - * @param priv - * Pointer to private structure. - * @param offloads - * Per-queue offloads configuration. + * @param dev + * Pointer to Ethernet device structure. + * @param idx + * Tx queue index. + * @param desc + * Number of descriptors to configure in queue. * * @return - * 1 if the configuration is valid, 0 otherwise. + * 0 on success, a negative errno value otherwise and rte_errno is set. */ static int -priv_is_tx_queue_offloads_allowed(struct priv *priv, uint64_t offloads) +mlx5_tx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc) { - uint64_t port_offloads = priv->dev->data->dev_conf.txmode.offloads; - uint64_t port_supp_offloads = mlx5_priv_get_tx_port_offloads(priv); + struct mlx5_priv *priv = dev->data->dev_private; - /* There are no Tx offloads which are per queue. */ - if ((offloads & port_supp_offloads) != offloads) - return 0; - if ((port_offloads ^ offloads) & port_supp_offloads) - return 0; - return 1; + 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; + } + 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); + } + DRV_LOG(DEBUG, "port %u configuring queue %u for %u descriptors", + 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); + rte_errno = EOVERFLOW; + return -rte_errno; + } + if (!mlx5_txq_releasable(dev, idx)) { + rte_errno = EBUSY; + DRV_LOG(ERR, "port %u unable to release queue index %u", + dev->data->port_id, idx); + return -rte_errno; + } + mlx5_txq_release(dev, idx); + return 0; } - /** * DPDK callback to configure a TX queue. * @@ -160,74 +196,82 @@ priv_is_tx_queue_offloads_allowed(struct priv *priv, uint64_t offloads) * Thresholds parameters. * * @return - * 0 on success, negative errno value on failure. + * 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 priv *priv = dev->data->dev_private; + 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 ret = 0; + int res; - priv_lock(priv); - /* - * Don't verify port offloads for application which - * use the old API. - */ - if (!!(conf->txq_flags & ETH_TXQ_FLAGS_IGNORE) && - !priv_is_tx_queue_offloads_allowed(priv, conf->offloads)) { - ret = ENOTSUP; - ERROR("%p: Tx queue offloads 0x%" PRIx64 " don't match port " - "offloads 0x%" PRIx64 " or supported offloads 0x%" PRIx64, - (void *)dev, conf->offloads, - dev->data->dev_conf.txmode.offloads, - mlx5_priv_get_tx_port_offloads(priv)); - goto out; - } - if (desc <= MLX5_TX_COMP_THRESH) { - WARN("%p: number of descriptors requested for TX queue %u" - " must be higher than MLX5_TX_COMP_THRESH, using" - " %u instead of %u", - (void *)dev, idx, MLX5_TX_COMP_THRESH + 1, desc); - desc = MLX5_TX_COMP_THRESH + 1; + 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", + dev->data->port_id, idx); + return -rte_errno; } - if (!rte_is_power_of_2(desc)) { - desc = 1 << log2above(desc); - WARN("%p: increased number of descriptors in TX queue %u" - " to the next power of two (%d)", - (void *)dev, idx, desc); + 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 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; } - DEBUG("%p: configuring queue %u for %u descriptors", - (void *)dev, idx, desc); - if (idx >= priv->txqs_n) { - ERROR("%p: queue index out of range (%u >= %u)", - (void *)dev, idx, priv->txqs_n); - priv_unlock(priv); - return -EOVERFLOW; - } - if (!mlx5_priv_txq_releasable(priv, idx)) { - ret = EBUSY; - ERROR("%p: unable to release queue index %u", - (void *)dev, idx); - goto out; - } - mlx5_priv_txq_release(priv, idx); - txq_ctrl = mlx5_priv_txq_new(priv, idx, desc, socket, conf); + txq_ctrl = mlx5_txq_hairpin_new(dev, idx, desc, hairpin_conf); if (!txq_ctrl) { - ERROR("%p: unable to allocate queue index %u", - (void *)dev, idx); - ret = ENOMEM; - goto out; + DRV_LOG(ERR, "port %u unable to allocate queue index %u", + dev->data->port_id, idx); + return -rte_errno; } - DEBUG("%p: adding TX queue %p to list", - (void *)dev, (void *)txq_ctrl); + DRV_LOG(DEBUG, "port %u adding Tx queue %u to list", + dev->data->port_id, idx); (*priv->txqs)[idx] = &txq_ctrl->txq; -out: - priv_unlock(priv); - return -ret; + txq_ctrl->type = MLX5_TXQ_TYPE_HAIRPIN; + return 0; } /** @@ -241,173 +285,293 @@ mlx5_tx_queue_release(void *dpdk_txq) { struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq; struct mlx5_txq_ctrl *txq_ctrl; - struct priv *priv; + struct mlx5_priv *priv; unsigned int i; if (txq == NULL) return; txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq); priv = txq_ctrl->priv; - priv_lock(priv); for (i = 0; (i != priv->txqs_n); ++i) if ((*priv->txqs)[i] == txq) { - DEBUG("%p: removing TX queue %p from list", - (void *)priv->dev, (void *)txq_ctrl); - mlx5_priv_txq_release(priv, i); + mlx5_txq_release(ETH_DEV(priv), i); + DRV_LOG(DEBUG, "port %u removing Tx queue %u from list", + PORT_ID(priv), txq->idx); break; } - priv_unlock(priv); } +/** + * Initialize Tx UAR registers for primary process. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + */ +static void +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)); +#ifndef RTE_ARCH_64 + unsigned int lock_idx; + const size_t page_size = sysconf(_SC_PAGESIZE); +#endif + + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + return; + assert(rte_eal_process_type() == RTE_PROC_PRIMARY); + assert(ppriv); + ppriv->uar_table[txq_ctrl->txq.idx] = txq_ctrl->bf_reg; +#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]; +#endif +} /** - * Mmap TX UAR(HW doorbell) pages into reserved UAR address space. - * Both primary and secondary process do mmap to make UAR address - * aligned. + * Remap UAR register of a Tx queue for secondary process. + * + * Remapped address is stored at the table in the process private structure of + * the device, indexed by queue index. * - * @param[in] priv - * Pointer to private structure. + * @param txq_ctrl + * Pointer to Tx queue control structure. * @param fd * Verbs file descriptor to map UAR pages. * * @return - * 0 on success, errno value on failure. + * 0 on success, a negative errno value otherwise and rte_errno is set. */ -int -priv_tx_uar_remap(struct priv *priv, int fd) +static int +txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd) { - unsigned int i, j; - uintptr_t pages[priv->txqs_n]; - unsigned int pages_n = 0; + struct mlx5_priv *priv = txq_ctrl->priv; + struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv)); + struct mlx5_txq_data *txq = &txq_ctrl->txq; + void *addr; uintptr_t uar_va; - uintptr_t off; + uintptr_t offset; + const size_t page_size = sysconf(_SC_PAGESIZE); + + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + return 0; + assert(ppriv); + /* + * As rdma-core, UARs are mapped in size of OS page + * size. Ref to libmlx5 function: mlx5_init_context() + */ + uar_va = (uintptr_t)txq_ctrl->bf_reg; + offset = uar_va & (page_size - 1); /* Offset in page. */ + addr = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, fd, + txq_ctrl->uar_mmap_offset); + if (addr == MAP_FAILED) { + DRV_LOG(ERR, + "port %u mmap failed for BF reg of txq %u", + txq->port_id, txq->idx); + rte_errno = ENXIO; + return -rte_errno; + } + addr = RTE_PTR_ADD(addr, offset); + ppriv->uar_table[txq->idx] = addr; + return 0; +} + +/** + * Unmap UAR register of a Tx queue for secondary process. + * + * @param txq_ctrl + * Pointer to Tx queue control structure. + */ +static void +txq_uar_uninit_secondary(struct mlx5_txq_ctrl *txq_ctrl) +{ + struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(txq_ctrl->priv)); + const size_t page_size = sysconf(_SC_PAGESIZE); void *addr; - void *ret; + + 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); +} + +/** + * Initialize Tx UAR registers for secondary process. + * + * @param dev + * Pointer to Ethernet device. + * @param fd + * Verbs file descriptor to map UAR pages. + * + * @return + * 0 on success, a negative errno value otherwise and rte_errno is set. + */ +int +mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd) +{ + struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_data *txq; struct mlx5_txq_ctrl *txq_ctrl; - int already_mapped; - size_t page_size = sysconf(_SC_PAGESIZE); - int r; + unsigned int i; + int ret; - memset(pages, 0, priv->txqs_n * sizeof(uintptr_t)); - /* - * As rdma-core, UARs are mapped in size of OS page size. - * Use aligned address to avoid duplicate mmap. - * Ref to libmlx5 function: mlx5_init_context() - */ + 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); - /* UAR addr form verbs used to find dup and offset in page. */ - uar_va = (uintptr_t)txq_ctrl->bf_reg_orig; - off = uar_va & (page_size - 1); /* offset in page. */ - uar_va = RTE_ALIGN_FLOOR(uar_va, page_size); /* page addr. */ - already_mapped = 0; - for (j = 0; j != pages_n; ++j) { - if (pages[j] == uar_va) { - already_mapped = 1; - break; - } - } - /* new address in reserved UAR address space. */ - addr = RTE_PTR_ADD(priv->uar_base, - uar_va & (MLX5_UAR_SIZE - 1)); - if (!already_mapped) { - pages[pages_n++] = uar_va; - /* fixed mmap to specified address in reserved - * address space. - */ - ret = mmap(addr, page_size, - PROT_WRITE, MAP_FIXED | MAP_SHARED, fd, - txq_ctrl->uar_mmap_offset); - if (ret != addr) { - /* fixed mmap have to return same address */ - ERROR("call to mmap failed on UAR for txq %d\n", - i); - r = ENXIO; - return r; - } - } - if (rte_eal_process_type() == RTE_PROC_PRIMARY) /* save once */ - txq_ctrl->txq.bf_reg = RTE_PTR_ADD((void *)addr, off); - else - assert(txq_ctrl->txq.bf_reg == - RTE_PTR_ADD((void *)addr, off)); + if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD) + continue; + assert(txq->idx == (uint16_t)i); + ret = txq_uar_init_secondary(txq_ctrl, fd); + if (ret) + goto error; } return 0; +error: + /* Rollback. */ + do { + 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); + } while (i--); + return -rte_errno; } /** - * Check if the burst function is using eMPW. + * Create the Tx hairpin queue object. * - * @param tx_pkt_burst - * Tx burst function pointer. + * @param dev + * Pointer to Ethernet device. + * @param idx + * Queue index in DPDK Tx queue array * * @return - * 1 if the burst function is using eMPW, 0 otherwise. + * The hairpin DevX object initialised, NULL otherwise and rte_errno is set. */ -static int -is_empw_burst_func(eth_tx_burst_t tx_pkt_burst) +static struct mlx5_txq_obj * +mlx5_txq_obj_hairpin_new(struct rte_eth_dev *dev, uint16_t idx) { - if (tx_pkt_burst == mlx5_tx_burst_raw_vec || - tx_pkt_burst == mlx5_tx_burst_vec || - tx_pkt_burst == mlx5_tx_burst_empw) - return 1; - return 0; + 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; + int ret = 0; + + assert(txq_data); + 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; + goto error; + } + tmpl->type = MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN; + tmpl->txq_ctrl = txq_ctrl; + attr.hairpin = 1; + attr.tis_lst_sz = 1; + /* Workaround for hairpin startup */ + attr.wq_attr.log_hairpin_num_packets = log2above(32); + /* Workaround for packets larger than 1KB */ + attr.wq_attr.log_hairpin_data_sz = + priv->config.hca_attr.log_max_hairpin_wq_data_sz; + 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_errno = errno; + goto error; + } + 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; +error: + ret = rte_errno; /* Save rte_errno before cleanup. */ + if (tmpl->tis) + mlx5_devx_cmd_destroy(tmpl->tis); + if (tmpl->sq) + mlx5_devx_cmd_destroy(tmpl->sq); + rte_errno = ret; /* Restore rte_errno. */ + return NULL; } /** * Create the Tx queue Verbs object. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx - * Queue index in DPDK Rx queue array + * Queue index in DPDK Tx queue array. + * @param type + * Type of the Tx queue object to create. * * @return - * The Verbs object initialised if it can be created. + * The Verbs object initialised, NULL otherwise and rte_errno is set. */ -struct mlx5_txq_ibv* -mlx5_priv_txq_ibv_new(struct priv *priv, 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; + 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; struct ibv_qp_attr mod; - struct ibv_cq_ex cq_attr; } attr; unsigned int cqe_n; struct mlx5dv_qp qp = { .comp_mask = MLX5DV_QP_MASK_UAR_MMAP_OFFSET }; struct mlx5dv_cq cq_info; struct mlx5dv_obj obj; const int desc = 1 << txq_data->elts_n; - eth_tx_burst_t tx_pkt_burst = priv_select_tx_function(priv, priv->dev); int ret = 0; + if (type == MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN) + return mlx5_txq_obj_hairpin_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); 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")) { - ERROR("MLX5_ENABLE_CQE_COMPRESSION must never be set"); - goto error; + DRV_LOG(ERR, + "port %u MLX5_ENABLE_CQE_COMPRESSION must never be set", + dev->data->port_id); + rte_errno = EINVAL; + return NULL; } - memset(&tmpl, 0, sizeof(struct mlx5_txq_ibv)); - /* MRs will be registered in mp2mr[] later. */ + memset(&tmpl, 0, sizeof(struct mlx5_txq_obj)); attr.cq = (struct ibv_cq_init_attr_ex){ .comp_mask = 0, }; - cqe_n = ((desc / MLX5_TX_COMP_THRESH) - 1) ? - ((desc / MLX5_TX_COMP_THRESH) - 1) : 1; - if (is_empw_burst_func(tx_pkt_burst)) - cqe_n += MLX5_TX_COMP_THRESH_INLINE_DIV; - tmpl.cq = mlx5_glue->create_cq(priv->ctx, cqe_n, NULL, NULL, 0); + cqe_n = desc / MLX5_TX_COMP_THRESH + + 1 + MLX5_TX_COMP_THRESH_INLINE_DIV; + tmpl.cq = mlx5_glue->create_cq(priv->sh->ctx, cqe_n, NULL, NULL, 0); if (tmpl.cq == NULL) { - ERROR("%p: CQ creation failure", (void *)txq_ctrl); + DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure", + dev->data->port_id, idx); + rte_errno = errno; goto error; } attr.init = (struct ibv_qp_init_attr_ex){ @@ -418,15 +582,15 @@ mlx5_priv_txq_ibv_new(struct priv *priv, uint16_t idx) .cap = { /* Max number of outstanding WRs. */ .max_send_wr = - ((priv->device_attr.orig_attr.max_qp_wr < + ((priv->sh->device_attr.orig_attr.max_qp_wr < desc) ? - priv->device_attr.orig_attr.max_qp_wr : + priv->sh->device_attr.orig_attr.max_qp_wr : desc), /* * Max number of scatter/gather elements in a WR, * must be 1 to prevent libmlx5 from trying to affect * too much memory. TX gather is not impacted by the - * priv->device_attr.max_sge limit and will still work + * device_attr.max_sge limit and will still work * properly. */ .max_send_sge = 1, @@ -437,30 +601,35 @@ mlx5_priv_txq_ibv_new(struct priv *priv, uint16_t idx) * Tx burst. */ .sq_sig_all = 0, - .pd = priv->pd, + .pd = priv->sh->pd, .comp_mask = IBV_QP_INIT_ATTR_PD, }; - if (txq_data->max_inline) + if (txq_data->inlen_send) attr.init.cap.max_inline_data = txq_ctrl->max_inline_data; if (txq_data->tso_en) { attr.init.max_tso_header = txq_ctrl->max_tso_header; attr.init.comp_mask |= IBV_QP_INIT_ATTR_MAX_TSO_HEADER; } - tmpl.qp = mlx5_glue->create_qp_ex(priv->ctx, &attr.init); + tmpl.qp = mlx5_glue->create_qp_ex(priv->sh->ctx, &attr.init); if (tmpl.qp == NULL) { - ERROR("%p: QP creation failure", (void *)txq_ctrl); + DRV_LOG(ERR, "port %u Tx queue %u QP creation failure", + dev->data->port_id, idx); + rte_errno = errno; goto error; } attr.mod = (struct ibv_qp_attr){ /* Move the QP to this state. */ .qp_state = IBV_QPS_INIT, - /* Primary port number. */ - .port_num = priv->port + /* IB device port number. */ + .port_num = (uint8_t)priv->ibv_port, }; ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, (IBV_QP_STATE | IBV_QP_PORT)); if (ret) { - ERROR("%p: QP state to IBV_QPS_INIT failed", (void *)txq_ctrl); + DRV_LOG(ERR, + "port %u Tx queue %u QP state to IBV_QPS_INIT failed", + dev->data->port_id, idx); + rte_errno = errno; goto error; } attr.mod = (struct ibv_qp_attr){ @@ -468,19 +637,27 @@ mlx5_priv_txq_ibv_new(struct priv *priv, uint16_t idx) }; ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE); if (ret) { - ERROR("%p: QP state to IBV_QPS_RTR failed", (void *)txq_ctrl); + DRV_LOG(ERR, + "port %u Tx queue %u QP state to IBV_QPS_RTR failed", + dev->data->port_id, idx); + rte_errno = errno; goto error; } attr.mod.qp_state = IBV_QPS_RTS; ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE); if (ret) { - ERROR("%p: QP state to IBV_QPS_RTS failed", (void *)txq_ctrl); + DRV_LOG(ERR, + "port %u Tx queue %u QP state to IBV_QPS_RTS failed", + dev->data->port_id, 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) { - ERROR("%p: cannot allocate memory", (void *)txq_ctrl); + if (!txq_obj) { + DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory", + dev->data->port_id, idx); + rte_errno = ENOMEM; goto error; } obj.cq.in = tmpl.cq; @@ -488,66 +665,108 @@ mlx5_priv_txq_ibv_new(struct priv *priv, uint16_t idx) obj.qp.in = tmpl.qp; obj.qp.out = &qp; ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_QP); - if (ret != 0) + if (ret != 0) { + rte_errno = errno; goto error; + } if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) { - ERROR("Wrong MLX5_CQE_SIZE environment variable value: " - "it should be set to %u", RTE_CACHE_LINE_SIZE); + DRV_LOG(ERR, + "port %u wrong MLX5_CQE_SIZE environment variable" + " value: it should be set to %u", + dev->data->port_id, RTE_CACHE_LINE_SIZE); + rte_errno = EINVAL; goto error; } txq_data->cqe_n = log2above(cq_info.cqe_cnt); + txq_data->cqe_s = 1 << txq_data->cqe_n; + txq_data->cqe_m = txq_data->cqe_s - 1; txq_data->qp_num_8s = tmpl.qp->qp_num << 8; txq_data->wqes = qp.sq.buf; txq_data->wqe_n = log2above(qp.sq.wqe_cnt); + txq_data->wqe_s = 1 << txq_data->wqe_n; + txq_data->wqe_m = txq_data->wqe_s - 1; + txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s; txq_data->qp_db = &qp.dbrec[MLX5_SND_DBR]; - txq_ctrl->bf_reg_orig = qp.bf.reg; txq_data->cq_db = cq_info.dbrec; - txq_data->cqes = - (volatile struct mlx5_cqe (*)[]) - (uintptr_t)cq_info.buf; + 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_ibv->qp = tmpl.qp; - txq_ibv->cq = tmpl.cq; - rte_atomic32_inc(&txq_ibv->refcnt); + txq_data->wqe_comp = 0; + txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV; +#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) { + ret = mlx5_devx_cmd_qp_query_tis_td(tmpl.qp, qp.tisn, + &priv->sh->tdn); + if (ret) { + DRV_LOG(ERR, "Fail to query port %u Tx queue %u QP TIS " + "transport domain", dev->data->port_id, idx); + rte_errno = EINVAL; + goto error; + } else { + DRV_LOG(DEBUG, "port %u Tx queue %u TIS number %d " + "transport domain %d", dev->data->port_id, + idx, qp.tisn, priv->sh->tdn); + } + } +#endif + txq_obj->qp = tmpl.qp; + txq_obj->cq = tmpl.cq; + rte_atomic32_inc(&txq_obj->refcnt); + txq_ctrl->bf_reg = qp.bf.reg; 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, + dev->data->port_id, txq_ctrl->uar_mmap_offset); } else { - ERROR("Failed to retrieve UAR info, invalid libmlx5.so version"); + DRV_LOG(ERR, + "port %u failed to retrieve UAR info, invalid" + " libmlx5.so", + dev->data->port_id); + rte_errno = EINVAL; goto error; } - DEBUG("%p: Verbs Tx queue %p: refcnt %d", (void *)priv, - (void *)txq_ibv, rte_atomic32_read(&txq_ibv->refcnt)); - LIST_INSERT_HEAD(&priv->txqsibv, txq_ibv, next); + 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_obj) + rte_free(txq_obj); priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE; + rte_errno = ret; /* Restore rte_errno. */ return NULL; } /** * Get an Tx queue Verbs object. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx - * Queue index in DPDK Rx queue array + * Queue index in DPDK Tx queue array. * * @return * The Verbs object if it exists. */ -struct mlx5_txq_ibv* -mlx5_priv_txq_ibv_get(struct priv *priv, 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; if (idx >= priv->txqs_n) @@ -555,79 +774,118 @@ mlx5_priv_txq_ibv_get(struct priv *priv, 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); - DEBUG("%p: Verbs Tx queue %p: refcnt %d", (void *)priv, - (void *)txq_ctrl->ibv, - rte_atomic32_read(&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 priv - * Pointer to private structure. - * @param txq_ibv + * @param txq_obj * Verbs Tx queue object. * * @return - * 0 on success, errno on failure. + * 1 while a reference on it exists, 0 when freed. */ int -mlx5_priv_txq_ibv_release(struct priv *priv, struct mlx5_txq_ibv *txq_ibv) +mlx5_txq_obj_release(struct mlx5_txq_obj *txq_obj) { - (void)priv; - assert(txq_ibv); - DEBUG("%p: Verbs Tx queue %p: refcnt %d", (void *)priv, - (void *)txq_ibv, rte_atomic32_read(&txq_ibv->refcnt)); - 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); + 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 { + claim_zero(mlx5_glue->destroy_qp(txq_obj->qp)); + claim_zero(mlx5_glue->destroy_cq(txq_obj->cq)); + } + LIST_REMOVE(txq_obj, next); + rte_free(txq_obj); return 0; } - return EBUSY; + return 1; } /** - * Return true if a single reference exists on the object. + * Verify the Verbs Tx queue list is empty * - * @param priv - * Pointer to private structure. - * @param txq_ibv - * Verbs Tx queue object. + * @param dev + * Pointer to Ethernet device. + * + * @return + * The number of object not released. */ int -mlx5_priv_txq_ibv_releasable(struct priv *priv, struct mlx5_txq_ibv *txq_ibv) +mlx5_txq_obj_verify(struct rte_eth_dev *dev) { - (void)priv; - assert(txq_ibv); - return (rte_atomic32_read(&txq_ibv->refcnt) == 1); + struct mlx5_priv *priv = dev->data->dev_private; + int ret = 0; + struct mlx5_txq_obj *txq_obj; + + LIST_FOREACH(txq_obj, &priv->txqsobj, next) { + DRV_LOG(DEBUG, "port %u Verbs Tx queue %u still referenced", + dev->data->port_id, txq_obj->txq_ctrl->txq.idx); + ++ret; + } + return ret; } /** - * Verify the Verbs Tx queue list is empty + * Calculate the total number of WQEBB for Tx queue. * - * @param priv - * Pointer to private structure. + * Simplified version of calc_sq_size() in rdma-core. * - * @return the number of object not released. + * @param txq_ctrl + * Pointer to Tx queue control structure. + * + * @return + * The number of WQEBB. */ -int -mlx5_priv_txq_ibv_verify(struct priv *priv) +static int +txq_calc_wqebb_cnt(struct mlx5_txq_ctrl *txq_ctrl) { - int ret = 0; - struct mlx5_txq_ibv *txq_ibv; + unsigned int wqe_size; + const unsigned int desc = 1 << txq_ctrl->txq.elts_n; - LIST_FOREACH(txq_ibv, &priv->txqsibv, next) { - DEBUG("%p: Verbs Tx queue %p still referenced", (void *)priv, - (void *)txq_ibv); - ++ret; - } - return ret; + wqe_size = MLX5_WQE_CSEG_SIZE + + MLX5_WQE_ESEG_SIZE + + MLX5_WSEG_SIZE - + MLX5_ESEG_MIN_INLINE_SIZE + + txq_ctrl->max_inline_data; + 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.orig_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; } /** @@ -639,102 +897,306 @@ mlx5_priv_txq_ibv_verify(struct priv *priv) static void txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) { - struct priv *priv = txq_ctrl->priv; + struct mlx5_priv *priv = txq_ctrl->priv; struct mlx5_dev_config *config = &priv->config; - const unsigned int max_tso_inline = - ((MLX5_MAX_TSO_HEADER + (RTE_CACHE_LINE_SIZE - 1)) / - RTE_CACHE_LINE_SIZE); - unsigned int txq_inline; - unsigned int txqs_inline; - unsigned int inline_max_packet_sz; - eth_tx_burst_t tx_pkt_burst = priv_select_tx_function(priv, priv->dev); - int is_empw_func = is_empw_burst_func(tx_pkt_burst); - int tso = !!(txq_ctrl->txq.offloads & DEV_TX_OFFLOAD_TCP_TSO); - - txq_inline = (config->txq_inline == MLX5_ARG_UNSET) ? - 0 : config->txq_inline; - txqs_inline = (config->txqs_inline == MLX5_ARG_UNSET) ? - 0 : config->txqs_inline; - inline_max_packet_sz = - (config->inline_max_packet_sz == MLX5_ARG_UNSET) ? - 0 : config->inline_max_packet_sz; - if (is_empw_func) { - if (config->txq_inline == MLX5_ARG_UNSET) - txq_inline = MLX5_WQE_SIZE_MAX - MLX5_WQE_SIZE; - if (config->txqs_inline == MLX5_ARG_UNSET) - txqs_inline = MLX5_EMPW_MIN_TXQS; - if (config->inline_max_packet_sz == MLX5_ARG_UNSET) - inline_max_packet_sz = MLX5_EMPW_MAX_INLINE_LEN; - txq_ctrl->txq.mpw_hdr_dseg = config->mpw_hdr_dseg; - txq_ctrl->txq.inline_max_packet_sz = inline_max_packet_sz; - } - if (txq_inline && priv->txqs_n >= txqs_inline) { - unsigned int ds_cnt; - - txq_ctrl->txq.max_inline = - ((txq_inline + (RTE_CACHE_LINE_SIZE - 1)) / - RTE_CACHE_LINE_SIZE); - if (is_empw_func) { - /* To minimize the size of data set, avoid requesting - * too large WQ. - */ - txq_ctrl->max_inline_data = - ((RTE_MIN(txq_inline, - inline_max_packet_sz) + - (RTE_CACHE_LINE_SIZE - 1)) / - RTE_CACHE_LINE_SIZE) * RTE_CACHE_LINE_SIZE; - } else if (tso) { - int inline_diff = txq_ctrl->txq.max_inline - - max_tso_inline; + unsigned int inlen_send; /* Inline data for ordinary SEND.*/ + unsigned int inlen_empw; /* Inline data for enhanced MPW. */ + unsigned int inlen_mode; /* Minimal required Inline data. */ + unsigned int txqs_inline; /* Min Tx queues to enable inline. */ + uint64_t dev_txoff = priv->dev_data->dev_conf.txmode.offloads; + bool tso = txq_ctrl->txq.offloads & (DEV_TX_OFFLOAD_TCP_TSO | + DEV_TX_OFFLOAD_VXLAN_TNL_TSO | + DEV_TX_OFFLOAD_GRE_TNL_TSO | + DEV_TX_OFFLOAD_IP_TNL_TSO | + DEV_TX_OFFLOAD_UDP_TNL_TSO); + bool vlan_inline; + unsigned int temp; + if (config->txqs_inline == MLX5_ARG_UNSET) + txqs_inline = +#if defined(RTE_ARCH_ARM64) + (priv->pci_dev->id.device_id == + PCI_DEVICE_ID_MELLANOX_CONNECTX5BF) ? + MLX5_INLINE_MAX_TXQS_BLUEFIELD : +#endif + MLX5_INLINE_MAX_TXQS; + else + txqs_inline = (unsigned int)config->txqs_inline; + inlen_send = (config->txq_inline_max == MLX5_ARG_UNSET) ? + MLX5_SEND_DEF_INLINE_LEN : + (unsigned int)config->txq_inline_max; + inlen_empw = (config->txq_inline_mpw == MLX5_ARG_UNSET) ? + MLX5_EMPW_DEF_INLINE_LEN : + (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) + 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 + * support E-Switch Flows. + */ + if (inlen_mode) { + if (inlen_mode <= MLX5_ESEG_MIN_INLINE_SIZE) { /* - * Adjust inline value as Verbs aggregates - * tso_inline and txq_inline fields. + * Optimize minimal inlining for single + * segment packets to fill one WQEBB + * without gaps. */ - txq_ctrl->max_inline_data = inline_diff > 0 ? - inline_diff * - RTE_CACHE_LINE_SIZE : - 0; + temp = MLX5_ESEG_MIN_INLINE_SIZE; } else { - txq_ctrl->max_inline_data = - txq_ctrl->txq.max_inline * RTE_CACHE_LINE_SIZE; + temp = inlen_mode - MLX5_ESEG_MIN_INLINE_SIZE; + temp = RTE_ALIGN(temp, MLX5_WSEG_SIZE) + + MLX5_ESEG_MIN_INLINE_SIZE; + temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN); } + if (temp != inlen_mode) { + DRV_LOG(INFO, + "port %u minimal required inline setting" + " aligned from %u to %u", + PORT_ID(priv), inlen_mode, temp); + inlen_mode = temp; + } + } + /* + * If port is configured to support VLAN insertion and device + * does not support this feature by HW (for NICs before ConnectX-5 + * or in case of wqe_vlan_insert flag is not set) we must enable + * data inline on all queues because it is supported by single + * tx_burst routine. + */ + txq_ctrl->txq.vlan_en = config->hw_vlan_insert; + vlan_inline = (dev_txoff & DEV_TX_OFFLOAD_VLAN_INSERT) && + !config->hw_vlan_insert; + /* + * 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) { /* - * Check if the inline size is too large in a way which - * can make the WQE DS to overflow. - * Considering in calculation: - * WQE CTRL (1 DS) - * WQE ETH (1 DS) - * Inline part (N DS) + * The data sent with ordinal MLX5_OPCODE_SEND + * may be inlined in Ethernet Segment, align the + * length accordingly to fit entire WQEBBs. */ - ds_cnt = 2 + (txq_ctrl->txq.max_inline / MLX5_WQE_DWORD_SIZE); - if (ds_cnt > MLX5_DSEG_MAX) { - unsigned int max_inline = (MLX5_DSEG_MAX - 2) * - MLX5_WQE_DWORD_SIZE; - - max_inline = max_inline - (max_inline % - RTE_CACHE_LINE_SIZE); - WARN("txq inline is too large (%d) setting it to " - "the maximum possible: %d\n", - txq_inline, max_inline); - txq_ctrl->txq.max_inline = max_inline / - RTE_CACHE_LINE_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 - + MLX5_WQE_ESEG_SIZE - + MLX5_WQE_DSEG_SIZE * 2); + temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN); + temp = RTE_MAX(temp, inlen_mode); + if (temp != inlen_send) { + DRV_LOG(INFO, + "port %u ordinary send inline setting" + " aligned from %u to %u", + PORT_ID(priv), inlen_send, temp); + inlen_send = temp; } + /* + * Not aligned to cache lines, but to WQEs. + * First bytes of data (initial alignment) + * is going to be copied explicitly at the + * 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); + } else if (inlen_mode) { + /* + * If minimal inlining is requested we must + * enable inlining in general, despite the + * number of configured queues. Ignore the + * txq_inline_max devarg, this is not + * full-featured inline. + */ + inlen_send = inlen_mode; + 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 = 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 - + MLX5_WQE_ESEG_SIZE - + MLX5_WQE_DSEG_SIZE); + temp = RTE_MIN(temp, MLX5_EMPW_MAX_INLINE_LEN); + if (temp != inlen_empw) { + DRV_LOG(INFO, + "port %u enhanced empw inline setting" + " aligned from %u to %u", + 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); + txq_ctrl->txq.inlen_empw = inlen_empw; + } + txq_ctrl->max_inline_data = RTE_MAX(inlen_send, inlen_empw); if (tso) { - txq_ctrl->max_tso_header = max_tso_inline * RTE_CACHE_LINE_SIZE; - txq_ctrl->txq.max_inline = RTE_MAX(txq_ctrl->txq.max_inline, - max_tso_inline); + txq_ctrl->max_tso_header = MLX5_MAX_TSO_HEADER; + txq_ctrl->max_inline_data = RTE_MAX(txq_ctrl->max_inline_data, + MLX5_MAX_TSO_HEADER); txq_ctrl->txq.tso_en = 1; } - txq_ctrl->txq.tunnel_en = config->tunnel_en; + txq_ctrl->txq.tunnel_en = config->tunnel_en | config->swp; + txq_ctrl->txq.swp_en = ((DEV_TX_OFFLOAD_IP_TNL_TSO | + DEV_TX_OFFLOAD_UDP_TNL_TSO | + DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM) & + 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.orig_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.orig_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.orig_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.orig_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); + assert(txq_ctrl->max_inline_data <= max_inline); + assert(txq_ctrl->txq.inlen_mode <= max_inline); + assert(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_send); + assert(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_empw); + return 0; +error: + rte_errno = ENOMEM; + return -ENOMEM; } /** * Create a DPDK Tx queue. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx * TX queue index. * @param desc @@ -745,38 +1207,98 @@ txq_set_params(struct mlx5_txq_ctrl *txq_ctrl) * Thresholds parameters. * * @return - * A DPDK queue object on success. + * A DPDK queue object on success, NULL otherwise and rte_errno is set. */ -struct mlx5_txq_ctrl* -mlx5_priv_txq_new(struct priv *priv, uint16_t idx, uint16_t desc, - unsigned int socket, - const struct rte_eth_txconf *conf) +struct mlx5_txq_ctrl * +mlx5_txq_new(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_ctrl *tmpl; tmpl = rte_calloc_socket("TXQ", 1, sizeof(*tmpl) + desc * sizeof(struct rte_mbuf *), 0, socket); - if (!tmpl) + if (!tmpl) { + rte_errno = ENOMEM; return NULL; + } + if (mlx5_mr_btree_init(&tmpl->txq.mr_ctrl.cache_bh, + MLX5_MR_BTREE_CACHE_N, socket)) { + /* rte_errno is already set. */ + 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.offloads = conf->offloads; + tmpl->txq.offloads = conf->offloads | + dev->data->dev_conf.txmode.offloads; tmpl->priv = priv; tmpl->socket = socket; tmpl->txq.elts_n = log2above(desc); + tmpl->txq.elts_s = desc; + tmpl->txq.elts_m = desc - 1; + tmpl->txq.port_id = dev->data->port_id; + tmpl->txq.idx = idx; txq_set_params(tmpl); - /* MRs will be registered in mp2mr[] later. */ - DEBUG("priv->device_attr.max_qp_wr is %d", - priv->device_attr.orig_attr.max_qp_wr); - DEBUG("priv->device_attr.max_sge is %d", - priv->device_attr.orig_attr.max_sge); - tmpl->txq.elts = - (struct rte_mbuf *(*)[1 << tmpl->txq.elts_n])(tmpl + 1); - tmpl->txq.stats.idx = idx; + if (txq_adjust_params(tmpl)) + goto error; + if (txq_calc_wqebb_cnt(tmpl) > + priv->sh->device_attr.orig_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); + 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: + rte_free(tmpl); + 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); - DEBUG("%p: Tx queue %p: refcnt %d", (void *)priv, - (void *)tmpl, rte_atomic32_read(&tmpl->refcnt)); LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next); return tmpl; } @@ -784,37 +1306,25 @@ mlx5_priv_txq_new(struct priv *priv, uint16_t idx, uint16_t desc, /** * Get a Tx queue. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx * TX queue index. * * @return * A pointer to the queue if it exists. */ -struct mlx5_txq_ctrl* -mlx5_priv_txq_get(struct priv *priv, uint16_t idx) +struct mlx5_txq_ctrl * +mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx) { + struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_ctrl *ctrl = NULL; if ((*priv->txqs)[idx]) { ctrl = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq); - unsigned int i; - - mlx5_priv_txq_ibv_get(priv, idx); - for (i = 0; i != MLX5_PMD_TX_MP_CACHE; ++i) { - struct mlx5_mr *mr = NULL; - - (void)mr; - if (ctrl->txq.mp2mr[i]) { - mr = priv_mr_get(priv, ctrl->txq.mp2mr[i]->mp); - assert(mr); - } - } + mlx5_txq_obj_get(dev, idx); rte_atomic32_inc(&ctrl->refcnt); - DEBUG("%p: Tx queue %p: refcnt %d", (void *)priv, - (void *)ctrl, rte_atomic32_read(&ctrl->refcnt)); } return ctrl; } @@ -822,57 +1332,41 @@ mlx5_priv_txq_get(struct priv *priv, uint16_t idx) /** * Release a Tx queue. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx * TX queue index. * * @return - * 0 on success, errno on failure. + * 1 while a reference on it exists, 0 when freed. */ int -mlx5_priv_txq_release(struct priv *priv, uint16_t idx) +mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx) { - unsigned int i; + struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_ctrl *txq; - size_t page_size = sysconf(_SC_PAGESIZE); if (!(*priv->txqs)[idx]) return 0; txq = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq); - DEBUG("%p: Tx queue %p: refcnt %d", (void *)priv, - (void *)txq, rte_atomic32_read(&txq->refcnt)); - if (txq->ibv) { - int ret; - - ret = mlx5_priv_txq_ibv_release(priv, txq->ibv); - if (!ret) - txq->ibv = NULL; - } - for (i = 0; i != MLX5_PMD_TX_MP_CACHE; ++i) { - if (txq->txq.mp2mr[i]) { - priv_mr_release(priv, txq->txq.mp2mr[i]); - txq->txq.mp2mr[i] = NULL; - } - } - if (priv->uar_base) - munmap((void *)RTE_ALIGN_FLOOR((uintptr_t)txq->txq.bf_reg, - page_size), page_size); + 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); LIST_REMOVE(txq, next); rte_free(txq); (*priv->txqs)[idx] = NULL; return 0; } - return EBUSY; + return 1; } /** * Verify if the queue can be released. * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * @param idx * TX queue index. * @@ -880,8 +1374,9 @@ mlx5_priv_txq_release(struct priv *priv, uint16_t idx) * 1 if the queue can be released. */ int -mlx5_priv_txq_releasable(struct priv *priv, uint16_t idx) +mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx) { + struct mlx5_priv *priv = dev->data->dev_private; struct mlx5_txq_ctrl *txq; if (!(*priv->txqs)[idx]) @@ -893,20 +1388,22 @@ mlx5_priv_txq_releasable(struct priv *priv, uint16_t idx) /** * Verify the Tx Queue list is empty * - * @param priv - * Pointer to private structure. + * @param dev + * Pointer to Ethernet device. * - * @return the number of object not released. + * @return + * The number of object not released. */ int -mlx5_priv_txq_verify(struct priv *priv) +mlx5_txq_verify(struct rte_eth_dev *dev) { - struct mlx5_txq_ctrl *txq; + struct mlx5_priv *priv = dev->data->dev_private; + struct mlx5_txq_ctrl *txq_ctrl; int ret = 0; - LIST_FOREACH(txq, &priv->txqsctrl, next) { - DEBUG("%p: Tx Queue %p still referenced", (void *)priv, - (void *)txq); + LIST_FOREACH(txq_ctrl, &priv->txqsctrl, next) { + DRV_LOG(DEBUG, "port %u Tx queue %u still referenced", + dev->data->port_id, txq_ctrl->txq.idx); ++ret; } return ret;