/* SPDX-License-Identifier: BSD-3-Clause
* Copyright 2015 6WIND S.A.
- * Copyright 2015 Mellanox.
+ * Copyright 2015 Mellanox Technologies, Ltd
*/
#include <stddef.h>
#include <stdint.h>
#include <unistd.h>
#include <sys/mman.h>
+#include <inttypes.h>
/* Verbs header. */
/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
#pragma GCC diagnostic ignored "-Wpedantic"
#endif
#include <infiniband/verbs.h>
+#include <infiniband/mlx5dv.h>
#ifdef PEDANTIC
#pragma GCC diagnostic error "-Wpedantic"
#endif
#include <rte_ethdev_driver.h>
#include <rte_common.h>
-#include "mlx5_utils.h"
+#include <mlx5_glue.h>
+#include <mlx5_devx_cmds.h>
+
#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.
unsigned int i;
for (i = 0; (i != elts_n); ++i)
- (*txq_ctrl->txq.elts)[i] = NULL;
+ txq_ctrl->txq.elts[i] = NULL;
DRV_LOG(DEBUG, "port %u Tx queue %u allocated and configured %u WRs",
- txq_ctrl->priv->dev->data->port_id, txq_ctrl->idx, elts_n);
+ 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;
* @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;
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;
DRV_LOG(DEBUG, "port %u Tx queue %u freeing WRs",
- txq_ctrl->priv->dev->data->port_id, txq_ctrl->idx);
+ 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;
uint64_t
mlx5_get_tx_port_offloads(struct rte_eth_dev *dev)
{
- struct priv *priv = dev->data->dev_private;
+ 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;
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;
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);
}
return offloads;
}
/**
- * Checks if the per-queue offload configuration is valid.
- *
- * @param dev
- * Pointer to Ethernet device.
- * @param offloads
- * Per-queue offloads configuration.
- *
- * @return
- * 1 if the configuration is valid, 0 otherwise.
- */
-static int
-mlx5_is_tx_queue_offloads_allowed(struct rte_eth_dev *dev, uint64_t offloads)
-{
- uint64_t port_offloads = dev->data->dev_conf.txmode.offloads;
- uint64_t port_supp_offloads = mlx5_get_tx_port_offloads(dev);
-
- /* 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;
-}
-
-/**
- * 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 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);
+ struct mlx5_priv *priv = dev->data->dev_private;
- /*
- * Don't verify port offloads for application which
- * use the old API.
- */
- if (!!(conf->txq_flags & ETH_TXQ_FLAGS_IGNORE) &&
- !mlx5_is_tx_queue_offloads_allowed(dev, conf->offloads)) {
- rte_errno = ENOTSUP;
- DRV_LOG(ERR,
- "port %u Tx queue offloads 0x%" PRIx64 " don't match"
- " port offloads 0x%" PRIx64 " or supported offloads 0x%"
- PRIx64,
- dev->data->port_id, conf->offloads,
- dev->data->dev_conf.txmode.offloads,
- mlx5_get_tx_port_offloads(dev));
- return -rte_errno;
- }
if (desc <= MLX5_TX_COMP_THRESH) {
DRV_LOG(WARNING,
"port %u number of descriptors requested for Tx queue"
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",
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.
*
{
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)
priv = txq_ctrl->priv;
for (i = 0; (i != priv->txqs_n); ++i)
if ((*priv->txqs)[i] == txq) {
- mlx5_txq_release(priv->dev, i);
+ mlx5_txq_release(ETH_DEV(priv), i);
DRV_LOG(DEBUG, "port %u removing Tx queue %u from list",
- priv->dev->data->port_id, txq_ctrl->idx);
+ PORT_ID(priv), txq->idx);
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;
+}
/**
- * Mmap TX UAR(HW doorbell) pages into reserved UAR address space.
- * Both primary and secondary process do mmap to make UAR address
- * aligned.
+ * Initialize Tx UAR registers for primary process.
*
- * @param[in] dev
- * Pointer to Ethernet device.
+ * @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));
+ const size_t page_size = sysconf(_SC_PAGESIZE);
+#ifndef RTE_ARCH_64
+ unsigned int lock_idx;
+#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;
+ 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];
+#endif
+}
+
+/**
+ * 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 txq_ctrl
+ * Pointer to Tx queue control structure.
* @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_remap(struct rte_eth_dev *dev, int fd)
+static int
+txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd)
{
- struct priv *priv = dev->data->dev_private;
- 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;
+ txq_uar_ncattr_init(txq_ctrl, page_size);
+ 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);
+ 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);
- assert(txq_ctrl->idx == (uint16_t)i);
- /* 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 */
- DRV_LOG(ERR,
- "port %u call to mmap failed on UAR"
- " for txq %u",
- dev->data->port_id, txq_ctrl->idx);
- rte_errno = ENXIO;
- return -rte_errno;
- }
- }
- 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;
}
/**
* @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, 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 priv *priv = dev->data->dev_private;
+ 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 = mlx5_select_tx_function(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;
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) {
DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure",
dev->data->port_id, 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,
* 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) {
DRV_LOG(ERR, "port %u Tx queue %u QP creation failure",
dev->data->port_id, idx);
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));
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;
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;
+ 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) {
+ 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 {
DRV_LOG(ERR,
"port %u failed to retrieve UAR info, invalid"
rte_errno = EINVAL;
goto error;
}
- DRV_LOG(DEBUG, "port %u Verbs Tx queue %u: refcnt %d",
- dev->data->port_id, idx, rte_atomic32_read(&txq_ibv->refcnt));
- 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_data && txq_data->fcqs)
+ rte_free(txq_data->fcqs);
+ 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;
* @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_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 priv *priv = dev->data->dev_private;
+ struct mlx5_priv *priv = dev->data->dev_private;
struct mlx5_txq_ctrl *txq_ctrl;
if (idx >= priv->txqs_n)
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);
- DRV_LOG(DEBUG, "port %u Verbs Tx queue %u: refcnt %d",
- dev->data->port_id, txq_ctrl->idx,
- 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 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);
- DRV_LOG(DEBUG, "port %u Verbs Tx queue %u: refcnt %d",
- txq_ibv->txq_ctrl->priv->dev->data->port_id,
- txq_ibv->txq_ctrl->idx, 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));
+ if (txq_obj->txq_ctrl->txq.fcqs)
+ rte_free(txq_obj->txq_ctrl->txq.fcqs);
+ }
+ LIST_REMOVE(txq_obj, next);
+ rte_free(txq_obj);
return 0;
}
return 1;
}
-/**
- * Return true if a single reference exists on the object.
- *
- * @param txq_ibv
- * Verbs Tx queue object.
- */
-int
-mlx5_txq_ibv_releasable(struct mlx5_txq_ibv *txq_ibv)
-{
- assert(txq_ibv);
- return (rte_atomic32_read(&txq_ibv->refcnt) == 1);
-}
-
/**
* Verify the Verbs Tx queue list is empty
*
* 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 priv *priv = dev->data->dev_private;
+ 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->idx);
+ dev->data->port_id, txq_obj->txq_ctrl->txq.idx);
++ret;
}
return ret;
}
+/**
+ * Calculate the total number of WQEBB for Tx queue.
+ *
+ * Simplified version of calc_sq_size() in rdma-core.
+ *
+ * @param txq_ctrl
+ * Pointer to Tx queue control structure.
+ *
+ * @return
+ * The number of WQEBB.
+ */
+static int
+txq_calc_wqebb_cnt(struct mlx5_txq_ctrl *txq_ctrl)
+{
+ unsigned int wqe_size;
+ const unsigned int desc = 1 << txq_ctrl->txq.elts_n;
+
+ 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;
+}
+
/**
* Set Tx queue parameters from device configuration.
*
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 =
- mlx5_select_tx_function(txq_ctrl->priv->dev);
- int is_empw_func = is_empw_burst_func(tx_pkt_burst);
- int tso = !!(txq_ctrl->txq.offloads & (DEV_TX_OFFLOAD_TCP_TSO |
- DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
- DEV_TX_OFFLOAD_GRE_TNL_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 && 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
+ * 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) {
+ /*
+ * The data sent with ordinal MLX5_OPCODE_SEND
+ * may be inlined in Ethernet Segment, align the
+ * length accordingly to fit entire WQEBBs.
+ */
+ 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) {
/*
- * 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 MLX5_OPCODE_ENHANCED_MPSW
+ * may be inlined in Data 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);
- DRV_LOG(WARNING,
- "port %u txq inline is too large (%d) setting"
- " it to the maximum possible: %d\n",
- priv->dev->data->port_id, txq_inline,
- max_inline);
- txq_ctrl->txq.max_inline = max_inline /
- RTE_CACHE_LINE_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 -
+ 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 ||
+ !txq_ctrl->txq.inlen_empw);
+ return 0;
+error:
+ rte_errno = ENOMEM;
+ return -ENOMEM;
}
/**
mlx5_txq_new(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_ctrl *tmpl;
tmpl = rte_calloc_socket("TXQ", 1,
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->idx = idx;
+ 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. */
- DRV_LOG(DEBUG, "port %u priv->device_attr.max_qp_wr is %d",
- dev->data->port_id, priv->device_attr.orig_attr.max_qp_wr);
- DRV_LOG(DEBUG, "port %u priv->device_attr.max_sge is %d",
- dev->data->port_id, 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);
- DRV_LOG(DEBUG, "port %u Tx queue %u: refcnt %d", dev->data->port_id,
- idx, rte_atomic32_read(&tmpl->refcnt));
LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next);
return tmpl;
}
struct mlx5_txq_ctrl *
mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx)
{
- struct priv *priv = dev->data->dev_private;
+ 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_txq_ibv_get(dev, idx);
- for (i = 0; i != MLX5_PMD_TX_MP_CACHE; ++i) {
- if (ctrl->txq.mp2mr[i])
- claim_nonzero
- (mlx5_mr_get(dev,
- ctrl->txq.mp2mr[i]->mp));
- }
+ mlx5_txq_obj_get(dev, idx);
rte_atomic32_inc(&ctrl->refcnt);
- DRV_LOG(DEBUG, "port %u Tx queue %u refcnt %d",
- dev->data->port_id,
- ctrl->idx, rte_atomic32_read(&ctrl->refcnt));
}
return ctrl;
}
int
mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx)
{
- struct priv *priv = dev->data->dev_private;
- 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);
- DRV_LOG(DEBUG, "port %u Tx queue %u: refcnt %d", dev->data->port_id,
- txq->idx, rte_atomic32_read(&txq->refcnt));
- if (txq->ibv && !mlx5_txq_ibv_release(txq->ibv))
- txq->ibv = NULL;
- for (i = 0; i != MLX5_PMD_TX_MP_CACHE; ++i) {
- if (txq->txq.mp2mr[i]) {
- mlx5_mr_release(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;
int
mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx)
{
- struct priv *priv = dev->data->dev_private;
+ struct mlx5_priv *priv = dev->data->dev_private;
struct mlx5_txq_ctrl *txq;
if (!(*priv->txqs)[idx])
int
mlx5_txq_verify(struct rte_eth_dev *dev)
{
- struct priv *priv = dev->data->dev_private;
- 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) {
+ LIST_FOREACH(txq_ctrl, &priv->txqsctrl, next) {
DRV_LOG(DEBUG, "port %u Tx queue %u still referenced",
- dev->data->port_id, txq->idx);
+ dev->data->port_id, txq_ctrl->txq.idx);
++ret;
}
return ret;