1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
14 #include <rte_malloc.h>
15 #include <rte_ethdev_driver.h>
16 #include <rte_common.h>
17 #include <rte_eal_paging.h>
19 #include <mlx5_glue.h>
20 #include <mlx5_devx_cmds.h>
21 #include <mlx5_common.h>
22 #include <mlx5_common_mr.h>
23 #include <mlx5_common_os.h>
24 #include <mlx5_malloc.h>
26 #include "mlx5_defs.h"
27 #include "mlx5_utils.h"
29 #include "mlx5_rxtx.h"
30 #include "mlx5_autoconf.h"
33 * Allocate TX queue elements.
36 * Pointer to TX queue structure.
39 txq_alloc_elts(struct mlx5_txq_ctrl *txq_ctrl)
41 const unsigned int elts_n = 1 << txq_ctrl->txq.elts_n;
44 for (i = 0; (i != elts_n); ++i)
45 txq_ctrl->txq.elts[i] = NULL;
46 DRV_LOG(DEBUG, "port %u Tx queue %u allocated and configured %u WRs",
47 PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx, elts_n);
48 txq_ctrl->txq.elts_head = 0;
49 txq_ctrl->txq.elts_tail = 0;
50 txq_ctrl->txq.elts_comp = 0;
54 * Free TX queue elements.
57 * Pointer to TX queue structure.
60 txq_free_elts(struct mlx5_txq_ctrl *txq_ctrl)
62 const uint16_t elts_n = 1 << txq_ctrl->txq.elts_n;
63 const uint16_t elts_m = elts_n - 1;
64 uint16_t elts_head = txq_ctrl->txq.elts_head;
65 uint16_t elts_tail = txq_ctrl->txq.elts_tail;
66 struct rte_mbuf *(*elts)[elts_n] = &txq_ctrl->txq.elts;
68 DRV_LOG(DEBUG, "port %u Tx queue %u freeing WRs",
69 PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx);
70 txq_ctrl->txq.elts_head = 0;
71 txq_ctrl->txq.elts_tail = 0;
72 txq_ctrl->txq.elts_comp = 0;
74 while (elts_tail != elts_head) {
75 struct rte_mbuf *elt = (*elts)[elts_tail & elts_m];
77 MLX5_ASSERT(elt != NULL);
78 rte_pktmbuf_free_seg(elt);
79 #ifdef RTE_LIBRTE_MLX5_DEBUG
81 memset(&(*elts)[elts_tail & elts_m],
83 sizeof((*elts)[elts_tail & elts_m]));
90 * Returns the per-port supported offloads.
93 * Pointer to Ethernet device.
96 * Supported Tx offloads.
99 mlx5_get_tx_port_offloads(struct rte_eth_dev *dev)
101 struct mlx5_priv *priv = dev->data->dev_private;
102 uint64_t offloads = (DEV_TX_OFFLOAD_MULTI_SEGS |
103 DEV_TX_OFFLOAD_VLAN_INSERT);
104 struct mlx5_dev_config *config = &priv->config;
107 offloads |= (DEV_TX_OFFLOAD_IPV4_CKSUM |
108 DEV_TX_OFFLOAD_UDP_CKSUM |
109 DEV_TX_OFFLOAD_TCP_CKSUM);
111 offloads |= DEV_TX_OFFLOAD_TCP_TSO;
113 offloads |= DEV_TX_OFFLOAD_SEND_ON_TIMESTAMP;
116 offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
118 offloads |= (DEV_TX_OFFLOAD_IP_TNL_TSO |
119 DEV_TX_OFFLOAD_UDP_TNL_TSO);
121 if (config->tunnel_en) {
123 offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
125 offloads |= (DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
126 DEV_TX_OFFLOAD_GRE_TNL_TSO |
127 DEV_TX_OFFLOAD_GENEVE_TNL_TSO);
132 /* Fetches and drops all SW-owned and error CQEs to synchronize CQ. */
134 txq_sync_cq(struct mlx5_txq_data *txq)
136 volatile struct mlx5_cqe *cqe;
141 cqe = &txq->cqes[txq->cq_ci & txq->cqe_m];
142 ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci);
143 if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
144 if (likely(ret != MLX5_CQE_STATUS_ERR)) {
145 /* No new CQEs in completion queue. */
146 MLX5_ASSERT(ret == MLX5_CQE_STATUS_HW_OWN);
152 /* Move all CQEs to HW ownership. */
153 for (i = 0; i < txq->cqe_s; i++) {
155 cqe->op_own = MLX5_CQE_INVALIDATE;
157 /* Resync CQE and WQE (WQ in reset state). */
159 *txq->cq_db = rte_cpu_to_be_32(txq->cq_ci);
164 * Tx queue stop. Device queue goes to the idle state,
165 * all involved mbufs are freed from elts/WQ.
168 * Pointer to Ethernet device structure.
173 * 0 on success, a negative errno value otherwise and rte_errno is set.
176 mlx5_tx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t idx)
178 struct mlx5_priv *priv = dev->data->dev_private;
179 struct mlx5_txq_data *txq = (*priv->txqs)[idx];
180 struct mlx5_txq_ctrl *txq_ctrl =
181 container_of(txq, struct mlx5_txq_ctrl, txq);
184 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
185 /* Move QP to RESET state. */
186 if (txq_ctrl->obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ) {
187 struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
189 /* Change queue state to reset with DevX. */
190 msq_attr.sq_state = MLX5_SQC_STATE_RDY;
191 msq_attr.state = MLX5_SQC_STATE_RST;
192 ret = mlx5_devx_cmd_modify_sq(txq_ctrl->obj->sq_devx,
195 DRV_LOG(ERR, "Cannot change the "
196 "Tx QP state to RESET %s",
202 struct ibv_qp_attr mod = {
203 .qp_state = IBV_QPS_RESET,
204 .port_num = (uint8_t)priv->dev_port,
206 struct ibv_qp *qp = txq_ctrl->obj->qp;
208 /* Change queue state to reset with Verbs. */
209 ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
211 DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
212 "%s", strerror(errno));
217 /* Handle all send completions. */
219 /* Free elts stored in the SQ. */
220 txq_free_elts(txq_ctrl);
221 /* Prevent writing new pkts to SQ by setting no free WQE.*/
222 txq->wqe_ci = txq->wqe_s;
225 /* Set the actual queue state. */
226 dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STOPPED;
231 * Tx queue stop. Device queue goes to the idle state,
232 * all involved mbufs are freed from elts/WQ.
235 * Pointer to Ethernet device structure.
240 * 0 on success, a negative errno value otherwise and rte_errno is set.
243 mlx5_tx_queue_stop(struct rte_eth_dev *dev, uint16_t idx)
247 if (dev->data->tx_queue_state[idx] == RTE_ETH_QUEUE_STATE_HAIRPIN) {
248 DRV_LOG(ERR, "Hairpin queue can't be stopped");
252 if (dev->data->tx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STOPPED)
254 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
255 ret = mlx5_mp_os_req_queue_control(dev, idx,
256 MLX5_MP_REQ_QUEUE_TX_STOP);
258 ret = mlx5_tx_queue_stop_primary(dev, idx);
264 * Rx queue start. Device queue goes to the ready state,
265 * all required mbufs are allocated and WQ is replenished.
268 * Pointer to Ethernet device structure.
273 * 0 on success, a negative errno value otherwise and rte_errno is set.
276 mlx5_tx_queue_start_primary(struct rte_eth_dev *dev, uint16_t idx)
278 struct mlx5_priv *priv = dev->data->dev_private;
279 struct mlx5_txq_data *txq = (*priv->txqs)[idx];
280 struct mlx5_txq_ctrl *txq_ctrl =
281 container_of(txq, struct mlx5_txq_ctrl, txq);
284 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
285 if (txq_ctrl->obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ) {
286 struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
287 struct mlx5_txq_obj *obj = txq_ctrl->obj;
289 msq_attr.sq_state = MLX5_SQC_STATE_RDY;
290 msq_attr.state = MLX5_SQC_STATE_RST;
291 ret = mlx5_devx_cmd_modify_sq(obj->sq_devx, &msq_attr);
295 "Cannot change the Tx QP state to RESET "
296 "%s", strerror(errno));
299 msq_attr.sq_state = MLX5_SQC_STATE_RST;
300 msq_attr.state = MLX5_SQC_STATE_RDY;
301 ret = mlx5_devx_cmd_modify_sq(obj->sq_devx, &msq_attr);
305 "Cannot change the Tx QP state to READY "
306 "%s", strerror(errno));
310 struct ibv_qp_attr mod = {
311 .qp_state = IBV_QPS_RESET,
312 .port_num = (uint8_t)priv->dev_port,
314 struct ibv_qp *qp = txq_ctrl->obj->qp;
316 ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
318 DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
319 "%s", strerror(errno));
323 mod.qp_state = IBV_QPS_INIT;
324 ret = mlx5_glue->modify_qp(qp, &mod,
325 (IBV_QP_STATE | IBV_QP_PORT));
327 DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s",
332 mod.qp_state = IBV_QPS_RTR;
333 ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
335 DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s",
340 mod.qp_state = IBV_QPS_RTS;
341 ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
343 DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s",
349 txq_ctrl->txq.wqe_ci = 0;
350 txq_ctrl->txq.wqe_pi = 0;
351 txq_ctrl->txq.elts_comp = 0;
352 /* Set the actual queue state. */
353 dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
358 * Rx queue start. Device queue goes to the ready state,
359 * all required mbufs are allocated and WQ is replenished.
362 * Pointer to Ethernet device structure.
367 * 0 on success, a negative errno value otherwise and rte_errno is set.
370 mlx5_tx_queue_start(struct rte_eth_dev *dev, uint16_t idx)
374 if (dev->data->tx_queue_state[idx] == RTE_ETH_QUEUE_STATE_HAIRPIN) {
375 DRV_LOG(ERR, "Hairpin queue can't be started");
379 if (dev->data->tx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STARTED)
381 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
382 ret = mlx5_mp_os_req_queue_control(dev, idx,
383 MLX5_MP_REQ_QUEUE_TX_START);
385 ret = mlx5_tx_queue_start_primary(dev, idx);
391 * Tx queue presetup checks.
394 * Pointer to Ethernet device structure.
398 * Number of descriptors to configure in queue.
401 * 0 on success, a negative errno value otherwise and rte_errno is set.
404 mlx5_tx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc)
406 struct mlx5_priv *priv = dev->data->dev_private;
408 if (*desc <= MLX5_TX_COMP_THRESH) {
410 "port %u number of descriptors requested for Tx queue"
411 " %u must be higher than MLX5_TX_COMP_THRESH, using %u"
412 " instead of %u", dev->data->port_id, idx,
413 MLX5_TX_COMP_THRESH + 1, *desc);
414 *desc = MLX5_TX_COMP_THRESH + 1;
416 if (!rte_is_power_of_2(*desc)) {
417 *desc = 1 << log2above(*desc);
419 "port %u increased number of descriptors in Tx queue"
420 " %u to the next power of two (%d)",
421 dev->data->port_id, idx, *desc);
423 DRV_LOG(DEBUG, "port %u configuring queue %u for %u descriptors",
424 dev->data->port_id, idx, *desc);
425 if (idx >= priv->txqs_n) {
426 DRV_LOG(ERR, "port %u Tx queue index out of range (%u >= %u)",
427 dev->data->port_id, idx, priv->txqs_n);
428 rte_errno = EOVERFLOW;
431 if (!mlx5_txq_releasable(dev, idx)) {
433 DRV_LOG(ERR, "port %u unable to release queue index %u",
434 dev->data->port_id, idx);
437 mlx5_txq_release(dev, idx);
441 * DPDK callback to configure a TX queue.
444 * Pointer to Ethernet device structure.
448 * Number of descriptors to configure in queue.
450 * NUMA socket on which memory must be allocated.
452 * Thresholds parameters.
455 * 0 on success, a negative errno value otherwise and rte_errno is set.
458 mlx5_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
459 unsigned int socket, const struct rte_eth_txconf *conf)
461 struct mlx5_priv *priv = dev->data->dev_private;
462 struct mlx5_txq_data *txq = (*priv->txqs)[idx];
463 struct mlx5_txq_ctrl *txq_ctrl =
464 container_of(txq, struct mlx5_txq_ctrl, txq);
467 res = mlx5_tx_queue_pre_setup(dev, idx, &desc);
470 txq_ctrl = mlx5_txq_new(dev, idx, desc, socket, conf);
472 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
473 dev->data->port_id, idx);
476 DRV_LOG(DEBUG, "port %u adding Tx queue %u to list",
477 dev->data->port_id, idx);
478 (*priv->txqs)[idx] = &txq_ctrl->txq;
479 dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
484 * DPDK callback to configure a TX hairpin queue.
487 * Pointer to Ethernet device structure.
491 * Number of descriptors to configure in queue.
492 * @param[in] hairpin_conf
493 * The hairpin binding configuration.
496 * 0 on success, a negative errno value otherwise and rte_errno is set.
499 mlx5_tx_hairpin_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
501 const struct rte_eth_hairpin_conf *hairpin_conf)
503 struct mlx5_priv *priv = dev->data->dev_private;
504 struct mlx5_txq_data *txq = (*priv->txqs)[idx];
505 struct mlx5_txq_ctrl *txq_ctrl =
506 container_of(txq, struct mlx5_txq_ctrl, txq);
509 res = mlx5_tx_queue_pre_setup(dev, idx, &desc);
512 if (hairpin_conf->peer_count != 1 ||
513 hairpin_conf->peers[0].port != dev->data->port_id ||
514 hairpin_conf->peers[0].queue >= priv->rxqs_n) {
515 DRV_LOG(ERR, "port %u unable to setup hairpin queue index %u "
516 " invalid hairpind configuration", dev->data->port_id,
521 txq_ctrl = mlx5_txq_hairpin_new(dev, idx, desc, hairpin_conf);
523 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
524 dev->data->port_id, idx);
527 DRV_LOG(DEBUG, "port %u adding Tx queue %u to list",
528 dev->data->port_id, idx);
529 (*priv->txqs)[idx] = &txq_ctrl->txq;
530 dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_HAIRPIN;
535 * DPDK callback to release a TX queue.
538 * Generic TX queue pointer.
541 mlx5_tx_queue_release(void *dpdk_txq)
543 struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
544 struct mlx5_txq_ctrl *txq_ctrl;
545 struct mlx5_priv *priv;
550 txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq);
551 priv = txq_ctrl->priv;
552 for (i = 0; (i != priv->txqs_n); ++i)
553 if ((*priv->txqs)[i] == txq) {
554 DRV_LOG(DEBUG, "port %u removing Tx queue %u from list",
555 PORT_ID(priv), txq->idx);
556 mlx5_txq_release(ETH_DEV(priv), i);
562 * Configure the doorbell register non-cached attribute.
565 * Pointer to Tx queue control structure.
570 txq_uar_ncattr_init(struct mlx5_txq_ctrl *txq_ctrl, size_t page_size)
572 struct mlx5_priv *priv = txq_ctrl->priv;
575 txq_ctrl->txq.db_heu = priv->config.dbnc == MLX5_TXDB_HEURISTIC;
576 txq_ctrl->txq.db_nc = 0;
577 /* Check the doorbell register mapping type. */
578 cmd = txq_ctrl->uar_mmap_offset / page_size;
579 cmd >>= MLX5_UAR_MMAP_CMD_SHIFT;
580 cmd &= MLX5_UAR_MMAP_CMD_MASK;
581 if (cmd == MLX5_MMAP_GET_NC_PAGES_CMD)
582 txq_ctrl->txq.db_nc = 1;
586 * Initialize Tx UAR registers for primary process.
589 * Pointer to Tx queue control structure.
592 txq_uar_init(struct mlx5_txq_ctrl *txq_ctrl)
594 struct mlx5_priv *priv = txq_ctrl->priv;
595 struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv));
597 unsigned int lock_idx;
599 const size_t page_size = rte_mem_page_size();
600 if (page_size == (size_t)-1) {
601 DRV_LOG(ERR, "Failed to get mem page size");
605 if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD)
607 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
609 ppriv->uar_table[txq_ctrl->txq.idx] = txq_ctrl->bf_reg;
610 txq_uar_ncattr_init(txq_ctrl, page_size);
612 /* Assign an UAR lock according to UAR page number */
613 lock_idx = (txq_ctrl->uar_mmap_offset / page_size) &
614 MLX5_UAR_PAGE_NUM_MASK;
615 txq_ctrl->txq.uar_lock = &priv->sh->uar_lock[lock_idx];
620 * Remap UAR register of a Tx queue for secondary process.
622 * Remapped address is stored at the table in the process private structure of
623 * the device, indexed by queue index.
626 * Pointer to Tx queue control structure.
628 * Verbs file descriptor to map UAR pages.
631 * 0 on success, a negative errno value otherwise and rte_errno is set.
634 txq_uar_init_secondary(struct mlx5_txq_ctrl *txq_ctrl, int fd)
636 struct mlx5_priv *priv = txq_ctrl->priv;
637 struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv));
638 struct mlx5_txq_data *txq = &txq_ctrl->txq;
642 const size_t page_size = rte_mem_page_size();
643 if (page_size == (size_t)-1) {
644 DRV_LOG(ERR, "Failed to get mem page size");
649 if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD)
653 * As rdma-core, UARs are mapped in size of OS page
654 * size. Ref to libmlx5 function: mlx5_init_context()
656 uar_va = (uintptr_t)txq_ctrl->bf_reg;
657 offset = uar_va & (page_size - 1); /* Offset in page. */
658 addr = rte_mem_map(NULL, page_size, RTE_PROT_WRITE, RTE_MAP_SHARED,
659 fd, txq_ctrl->uar_mmap_offset);
662 "port %u mmap failed for BF reg of txq %u",
663 txq->port_id, txq->idx);
667 addr = RTE_PTR_ADD(addr, offset);
668 ppriv->uar_table[txq->idx] = addr;
669 txq_uar_ncattr_init(txq_ctrl, page_size);
674 * Unmap UAR register of a Tx queue for secondary process.
677 * Pointer to Tx queue control structure.
680 txq_uar_uninit_secondary(struct mlx5_txq_ctrl *txq_ctrl)
682 struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(txq_ctrl->priv));
684 const size_t page_size = rte_mem_page_size();
685 if (page_size == (size_t)-1) {
686 DRV_LOG(ERR, "Failed to get mem page size");
690 if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD)
692 addr = ppriv->uar_table[txq_ctrl->txq.idx];
693 rte_mem_unmap(RTE_PTR_ALIGN_FLOOR(addr, page_size), page_size);
697 * Deinitialize Tx UAR registers for secondary process.
700 * Pointer to Ethernet device.
703 mlx5_tx_uar_uninit_secondary(struct rte_eth_dev *dev)
705 struct mlx5_priv *priv = dev->data->dev_private;
706 struct mlx5_txq_data *txq;
707 struct mlx5_txq_ctrl *txq_ctrl;
710 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_SECONDARY);
711 for (i = 0; i != priv->txqs_n; ++i) {
712 if (!(*priv->txqs)[i])
714 txq = (*priv->txqs)[i];
715 txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq);
716 txq_uar_uninit_secondary(txq_ctrl);
721 * Initialize Tx UAR registers for secondary process.
724 * Pointer to Ethernet device.
726 * Verbs file descriptor to map UAR pages.
729 * 0 on success, a negative errno value otherwise and rte_errno is set.
732 mlx5_tx_uar_init_secondary(struct rte_eth_dev *dev, int fd)
734 struct mlx5_priv *priv = dev->data->dev_private;
735 struct mlx5_txq_data *txq;
736 struct mlx5_txq_ctrl *txq_ctrl;
740 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_SECONDARY);
741 for (i = 0; i != priv->txqs_n; ++i) {
742 if (!(*priv->txqs)[i])
744 txq = (*priv->txqs)[i];
745 txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq);
746 if (txq_ctrl->type != MLX5_TXQ_TYPE_STANDARD)
748 MLX5_ASSERT(txq->idx == (uint16_t)i);
749 ret = txq_uar_init_secondary(txq_ctrl, fd);
757 if (!(*priv->txqs)[i])
759 txq = (*priv->txqs)[i];
760 txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq);
761 txq_uar_uninit_secondary(txq_ctrl);
767 * Create the Tx hairpin queue object.
770 * Pointer to Ethernet device.
772 * Queue index in DPDK Tx queue array
775 * The hairpin DevX object initialised, NULL otherwise and rte_errno is set.
777 static struct mlx5_txq_obj *
778 mlx5_txq_obj_hairpin_new(struct rte_eth_dev *dev, uint16_t idx)
780 struct mlx5_priv *priv = dev->data->dev_private;
781 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
782 struct mlx5_txq_ctrl *txq_ctrl =
783 container_of(txq_data, struct mlx5_txq_ctrl, txq);
784 struct mlx5_devx_create_sq_attr attr = { 0 };
785 struct mlx5_txq_obj *tmpl = NULL;
786 uint32_t max_wq_data;
788 MLX5_ASSERT(txq_data);
789 MLX5_ASSERT(!txq_ctrl->obj);
790 tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl), 0,
794 "port %u Tx queue %u cannot allocate memory resources",
795 dev->data->port_id, txq_data->idx);
799 tmpl->type = MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN;
800 tmpl->txq_ctrl = txq_ctrl;
803 max_wq_data = priv->config.hca_attr.log_max_hairpin_wq_data_sz;
804 /* Jumbo frames > 9KB should be supported, and more packets. */
805 if (priv->config.log_hp_size != (uint32_t)MLX5_ARG_UNSET) {
806 if (priv->config.log_hp_size > max_wq_data) {
807 DRV_LOG(ERR, "total data size %u power of 2 is "
808 "too large for hairpin",
809 priv->config.log_hp_size);
814 attr.wq_attr.log_hairpin_data_sz = priv->config.log_hp_size;
816 attr.wq_attr.log_hairpin_data_sz =
817 (max_wq_data < MLX5_HAIRPIN_JUMBO_LOG_SIZE) ?
818 max_wq_data : MLX5_HAIRPIN_JUMBO_LOG_SIZE;
820 /* Set the packets number to the maximum value for performance. */
821 attr.wq_attr.log_hairpin_num_packets =
822 attr.wq_attr.log_hairpin_data_sz -
823 MLX5_HAIRPIN_QUEUE_STRIDE;
824 attr.tis_num = priv->sh->tis->id;
825 tmpl->sq = mlx5_devx_cmd_create_sq(priv->sh->ctx, &attr);
828 "port %u tx hairpin queue %u can't create sq object",
829 dev->data->port_id, idx);
834 DRV_LOG(DEBUG, "port %u sxq %u updated with %p", dev->data->port_id,
836 rte_atomic32_inc(&tmpl->refcnt);
837 LIST_INSERT_HEAD(&priv->txqsobj, tmpl, next);
842 * Destroy the Tx queue DevX object.
845 * Txq object to destroy
848 txq_release_sq_resources(struct mlx5_txq_obj *txq_obj)
850 MLX5_ASSERT(txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ);
852 if (txq_obj->sq_devx)
853 claim_zero(mlx5_devx_cmd_destroy(txq_obj->sq_devx));
854 if (txq_obj->sq_dbrec_page)
855 claim_zero(mlx5_release_dbr
856 (&txq_obj->txq_ctrl->priv->dbrpgs,
858 (txq_obj->sq_dbrec_page->umem),
859 txq_obj->sq_dbrec_offset));
860 if (txq_obj->sq_umem)
861 claim_zero(mlx5_glue->devx_umem_dereg(txq_obj->sq_umem));
863 mlx5_free(txq_obj->sq_buf);
864 if (txq_obj->cq_devx)
865 claim_zero(mlx5_devx_cmd_destroy(txq_obj->cq_devx));
866 if (txq_obj->cq_dbrec_page)
867 claim_zero(mlx5_release_dbr
868 (&txq_obj->txq_ctrl->priv->dbrpgs,
870 (txq_obj->cq_dbrec_page->umem),
871 txq_obj->cq_dbrec_offset));
872 if (txq_obj->cq_umem)
873 claim_zero(mlx5_glue->devx_umem_dereg(txq_obj->cq_umem));
875 mlx5_free(txq_obj->cq_buf);
879 * Create the Tx queue DevX object.
882 * Pointer to Ethernet device.
884 * Queue index in DPDK Tx queue array
887 * The DevX object initialised, NULL otherwise and rte_errno is set.
889 static struct mlx5_txq_obj *
890 mlx5_txq_obj_devx_new(struct rte_eth_dev *dev, uint16_t idx)
892 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
893 DRV_LOG(ERR, "port %u Tx queue %u cannot create with DevX, no UAR",
894 dev->data->port_id, idx);
898 struct mlx5_priv *priv = dev->data->dev_private;
899 struct mlx5_dev_ctx_shared *sh = priv->sh;
900 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
901 struct mlx5_txq_ctrl *txq_ctrl =
902 container_of(txq_data, struct mlx5_txq_ctrl, txq);
903 struct mlx5_devx_create_sq_attr sq_attr = { 0 };
904 struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
905 struct mlx5_devx_cq_attr cq_attr = { 0 };
906 struct mlx5_txq_obj *txq_obj = NULL;
908 struct mlx5_cqe *cqe;
911 size_t alignment = (size_t)-1;
914 MLX5_ASSERT(txq_data);
915 MLX5_ASSERT(!txq_ctrl->obj);
916 page_size = rte_mem_page_size();
917 if (page_size == (size_t)-1) {
918 DRV_LOG(ERR, "Failed to get mem page size");
922 txq_obj = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
923 sizeof(struct mlx5_txq_obj), 0,
927 "port %u Tx queue %u cannot allocate memory resources",
928 dev->data->port_id, txq_data->idx);
932 txq_obj->type = MLX5_TXQ_OBJ_TYPE_DEVX_SQ;
933 txq_obj->txq_ctrl = txq_ctrl;
935 /* Create the Completion Queue. */
936 nqe = (1UL << txq_data->elts_n) / MLX5_TX_COMP_THRESH +
937 1 + MLX5_TX_COMP_THRESH_INLINE_DIV;
938 nqe = 1UL << log2above(nqe);
939 if (nqe > UINT16_MAX) {
941 "port %u Tx queue %u requests to many CQEs %u",
942 dev->data->port_id, txq_data->idx, nqe);
946 /* Allocate memory buffer for CQEs. */
947 alignment = MLX5_CQE_BUF_ALIGNMENT;
948 if (alignment == (size_t)-1) {
949 DRV_LOG(ERR, "Failed to get mem page size");
953 txq_obj->cq_buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
954 nqe * sizeof(struct mlx5_cqe),
957 if (!txq_obj->cq_buf) {
959 "port %u Tx queue %u cannot allocate memory (CQ)",
960 dev->data->port_id, txq_data->idx);
964 txq_data->cqe_n = log2above(nqe);
965 txq_data->cqe_s = 1 << txq_data->cqe_n;
966 txq_data->cqe_m = txq_data->cqe_s - 1;
967 txq_data->cqes = (volatile struct mlx5_cqe *)txq_obj->cq_buf;
970 /* Register allocated buffer in user space with DevX. */
971 txq_obj->cq_umem = mlx5_glue->devx_umem_reg
973 (void *)txq_obj->cq_buf,
974 nqe * sizeof(struct mlx5_cqe),
975 IBV_ACCESS_LOCAL_WRITE);
976 if (!txq_obj->cq_umem) {
979 "port %u Tx queue %u cannot register memory (CQ)",
980 dev->data->port_id, txq_data->idx);
983 /* Allocate doorbell record for completion queue. */
984 txq_obj->cq_dbrec_offset = mlx5_get_dbr(sh->ctx,
986 &txq_obj->cq_dbrec_page);
987 if (txq_obj->cq_dbrec_offset < 0)
989 txq_data->cq_db = (volatile uint32_t *)(txq_obj->cq_dbrec_page->dbrs +
990 txq_obj->cq_dbrec_offset);
991 *txq_data->cq_db = 0;
992 /* Create completion queue object with DevX. */
993 cq_attr.cqe_size = (sizeof(struct mlx5_cqe) == 128) ?
994 MLX5_CQE_SIZE_128B : MLX5_CQE_SIZE_64B;
995 cq_attr.uar_page_id = mlx5_os_get_devx_uar_page_id(sh->tx_uar);
996 cq_attr.eqn = sh->txpp.eqn;
997 cq_attr.q_umem_valid = 1;
998 cq_attr.q_umem_offset = (uintptr_t)txq_obj->cq_buf % page_size;
999 cq_attr.q_umem_id = mlx5_os_get_umem_id(txq_obj->cq_umem);
1000 cq_attr.db_umem_valid = 1;
1001 cq_attr.db_umem_offset = txq_obj->cq_dbrec_offset;
1002 cq_attr.db_umem_id = mlx5_os_get_umem_id(txq_obj->cq_dbrec_page->umem);
1003 cq_attr.log_cq_size = rte_log2_u32(nqe);
1004 cq_attr.log_page_size = rte_log2_u32(page_size);
1005 txq_obj->cq_devx = mlx5_devx_cmd_create_cq(sh->ctx, &cq_attr);
1006 if (!txq_obj->cq_devx) {
1008 DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure",
1009 dev->data->port_id, idx);
1012 /* Initial fill CQ buffer with invalid CQE opcode. */
1013 cqe = (struct mlx5_cqe *)txq_obj->cq_buf;
1014 for (i = 0; i < txq_data->cqe_s; i++) {
1015 cqe->op_own = (MLX5_CQE_INVALID << 4) | MLX5_CQE_OWNER_MASK;
1018 /* Create the Work Queue. */
1019 nqe = RTE_MIN(1UL << txq_data->elts_n,
1020 (uint32_t)sh->device_attr.max_qp_wr);
1021 txq_obj->sq_buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1022 nqe * sizeof(struct mlx5_wqe),
1023 page_size, sh->numa_node);
1024 if (!txq_obj->sq_buf) {
1026 "port %u Tx queue %u cannot allocate memory (SQ)",
1027 dev->data->port_id, txq_data->idx);
1031 txq_data->wqe_n = log2above(nqe);
1032 txq_data->wqe_s = 1 << txq_data->wqe_n;
1033 txq_data->wqe_m = txq_data->wqe_s - 1;
1034 txq_data->wqes = (struct mlx5_wqe *)txq_obj->sq_buf;
1035 txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s;
1036 txq_data->wqe_ci = 0;
1037 txq_data->wqe_pi = 0;
1038 txq_data->wqe_comp = 0;
1039 txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV;
1040 /* Register allocated buffer in user space with DevX. */
1041 txq_obj->sq_umem = mlx5_glue->devx_umem_reg
1043 (void *)txq_obj->sq_buf,
1044 nqe * sizeof(struct mlx5_wqe),
1045 IBV_ACCESS_LOCAL_WRITE);
1046 if (!txq_obj->sq_umem) {
1049 "port %u Tx queue %u cannot register memory (SQ)",
1050 dev->data->port_id, txq_data->idx);
1053 /* Allocate doorbell record for completion queue. */
1054 txq_obj->cq_dbrec_offset = mlx5_get_dbr(sh->ctx,
1056 &txq_obj->sq_dbrec_page);
1057 if (txq_obj->sq_dbrec_offset < 0)
1059 txq_data->qp_db = (volatile uint32_t *)
1060 (txq_obj->sq_dbrec_page->dbrs +
1061 txq_obj->sq_dbrec_offset +
1062 MLX5_SND_DBR * sizeof(uint32_t));
1063 *txq_data->qp_db = 0;
1064 /* Create Send Queue object with DevX. */
1065 sq_attr.tis_lst_sz = 1;
1066 sq_attr.tis_num = sh->tis->id;
1067 sq_attr.state = MLX5_SQC_STATE_RST;
1068 sq_attr.cqn = txq_obj->cq_devx->id;
1069 sq_attr.flush_in_error_en = 1;
1070 sq_attr.allow_multi_pkt_send_wqe = !!priv->config.mps;
1071 sq_attr.allow_swp = !!priv->config.swp;
1072 sq_attr.min_wqe_inline_mode = priv->config.hca_attr.vport_inline_mode;
1073 sq_attr.wq_attr.uar_page = mlx5_os_get_devx_uar_page_id(sh->tx_uar);
1074 sq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
1075 sq_attr.wq_attr.pd = sh->pdn;
1076 sq_attr.wq_attr.log_wq_stride = rte_log2_u32(MLX5_WQE_SIZE);
1077 sq_attr.wq_attr.log_wq_sz = txq_data->wqe_n;
1078 sq_attr.wq_attr.dbr_umem_valid = 1;
1079 sq_attr.wq_attr.dbr_addr = txq_obj->cq_dbrec_offset;
1080 sq_attr.wq_attr.dbr_umem_id =
1081 mlx5_os_get_umem_id(txq_obj->cq_dbrec_page->umem);
1082 sq_attr.wq_attr.wq_umem_valid = 1;
1083 sq_attr.wq_attr.wq_umem_id = mlx5_os_get_umem_id(txq_obj->sq_umem);
1084 sq_attr.wq_attr.wq_umem_offset = (uintptr_t)txq_obj->sq_buf % page_size;
1085 txq_obj->sq_devx = mlx5_devx_cmd_create_sq(sh->ctx, &sq_attr);
1086 if (!txq_obj->sq_devx) {
1088 DRV_LOG(ERR, "port %u Tx queue %u SQ creation failure",
1089 dev->data->port_id, idx);
1092 txq_data->qp_num_8s = txq_obj->sq_devx->id << 8;
1093 /* Change Send Queue state to Ready-to-Send. */
1094 msq_attr.sq_state = MLX5_SQC_STATE_RST;
1095 msq_attr.state = MLX5_SQC_STATE_RDY;
1096 ret = mlx5_devx_cmd_modify_sq(txq_obj->sq_devx, &msq_attr);
1100 "port %u Tx queue %u SP state to SQC_STATE_RDY failed",
1101 dev->data->port_id, idx);
1104 txq_data->fcqs = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1105 txq_data->cqe_s * sizeof(*txq_data->fcqs),
1106 RTE_CACHE_LINE_SIZE,
1108 if (!txq_data->fcqs) {
1109 DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory (FCQ)",
1110 dev->data->port_id, idx);
1114 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1116 * If using DevX need to query and store TIS transport domain value.
1117 * This is done once per port.
1118 * Will use this value on Rx, when creating matching TIR.
1120 if (priv->config.devx && !priv->sh->tdn)
1121 priv->sh->tdn = priv->sh->td->id;
1123 MLX5_ASSERT(sh->tx_uar);
1124 reg_addr = mlx5_os_get_devx_uar_reg_addr(sh->tx_uar);
1125 MLX5_ASSERT(reg_addr);
1126 txq_ctrl->bf_reg = reg_addr;
1127 txq_ctrl->uar_mmap_offset =
1128 mlx5_os_get_devx_uar_mmap_offset(sh->tx_uar);
1129 rte_atomic32_set(&txq_obj->refcnt, 1);
1130 txq_uar_init(txq_ctrl);
1131 LIST_INSERT_HEAD(&priv->txqsobj, txq_obj, next);
1134 ret = rte_errno; /* Save rte_errno before cleanup. */
1135 txq_release_sq_resources(txq_obj);
1136 if (txq_data->fcqs) {
1137 mlx5_free(txq_data->fcqs);
1138 txq_data->fcqs = NULL;
1141 rte_errno = ret; /* Restore rte_errno. */
1147 * Create the Tx queue Verbs object.
1150 * Pointer to Ethernet device.
1152 * Queue index in DPDK Tx queue array.
1154 * Type of the Tx queue object to create.
1157 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1159 struct mlx5_txq_obj *
1160 mlx5_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx,
1161 enum mlx5_txq_obj_type type)
1163 struct mlx5_priv *priv = dev->data->dev_private;
1164 struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
1165 struct mlx5_txq_ctrl *txq_ctrl =
1166 container_of(txq_data, struct mlx5_txq_ctrl, txq);
1167 struct mlx5_txq_obj tmpl;
1168 struct mlx5_txq_obj *txq_obj = NULL;
1170 struct ibv_qp_init_attr_ex init;
1171 struct ibv_cq_init_attr_ex cq;
1172 struct ibv_qp_attr mod;
1175 struct mlx5dv_qp qp = { .comp_mask = MLX5DV_QP_MASK_UAR_MMAP_OFFSET };
1176 struct mlx5dv_cq cq_info;
1177 struct mlx5dv_obj obj;
1178 const int desc = 1 << txq_data->elts_n;
1181 if (type == MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN)
1182 return mlx5_txq_obj_hairpin_new(dev, idx);
1183 if (type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ)
1184 return mlx5_txq_obj_devx_new(dev, idx);
1185 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1186 /* If using DevX, need additional mask to read tisn value. */
1187 if (priv->config.devx && !priv->sh->tdn)
1188 qp.comp_mask |= MLX5DV_QP_MASK_RAW_QP_HANDLES;
1190 MLX5_ASSERT(txq_data);
1191 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_TX_QUEUE;
1192 priv->verbs_alloc_ctx.obj = txq_ctrl;
1193 if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) {
1195 "port %u MLX5_ENABLE_CQE_COMPRESSION must never be set",
1196 dev->data->port_id);
1200 memset(&tmpl, 0, sizeof(struct mlx5_txq_obj));
1201 attr.cq = (struct ibv_cq_init_attr_ex){
1204 cqe_n = desc / MLX5_TX_COMP_THRESH +
1205 1 + MLX5_TX_COMP_THRESH_INLINE_DIV;
1206 tmpl.cq = mlx5_glue->create_cq(priv->sh->ctx, cqe_n, NULL, NULL, 0);
1207 if (tmpl.cq == NULL) {
1208 DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure",
1209 dev->data->port_id, idx);
1213 attr.init = (struct ibv_qp_init_attr_ex){
1214 /* CQ to be associated with the send queue. */
1216 /* CQ to be associated with the receive queue. */
1219 /* Max number of outstanding WRs. */
1221 ((priv->sh->device_attr.max_qp_wr <
1223 priv->sh->device_attr.max_qp_wr :
1226 * Max number of scatter/gather elements in a WR,
1227 * must be 1 to prevent libmlx5 from trying to affect
1228 * too much memory. TX gather is not impacted by the
1229 * device_attr.max_sge limit and will still work
1234 .qp_type = IBV_QPT_RAW_PACKET,
1236 * Do *NOT* enable this, completions events are managed per
1241 .comp_mask = IBV_QP_INIT_ATTR_PD,
1243 if (txq_data->inlen_send)
1244 attr.init.cap.max_inline_data = txq_ctrl->max_inline_data;
1245 if (txq_data->tso_en) {
1246 attr.init.max_tso_header = txq_ctrl->max_tso_header;
1247 attr.init.comp_mask |= IBV_QP_INIT_ATTR_MAX_TSO_HEADER;
1249 tmpl.qp = mlx5_glue->create_qp_ex(priv->sh->ctx, &attr.init);
1250 if (tmpl.qp == NULL) {
1251 DRV_LOG(ERR, "port %u Tx queue %u QP creation failure",
1252 dev->data->port_id, idx);
1256 attr.mod = (struct ibv_qp_attr){
1257 /* Move the QP to this state. */
1258 .qp_state = IBV_QPS_INIT,
1259 /* IB device port number. */
1260 .port_num = (uint8_t)priv->dev_port,
1262 ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod,
1263 (IBV_QP_STATE | IBV_QP_PORT));
1266 "port %u Tx queue %u QP state to IBV_QPS_INIT failed",
1267 dev->data->port_id, idx);
1271 attr.mod = (struct ibv_qp_attr){
1272 .qp_state = IBV_QPS_RTR
1274 ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1277 "port %u Tx queue %u QP state to IBV_QPS_RTR failed",
1278 dev->data->port_id, idx);
1282 attr.mod.qp_state = IBV_QPS_RTS;
1283 ret = mlx5_glue->modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1286 "port %u Tx queue %u QP state to IBV_QPS_RTS failed",
1287 dev->data->port_id, idx);
1291 txq_obj = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1292 sizeof(struct mlx5_txq_obj), 0,
1295 DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory",
1296 dev->data->port_id, idx);
1300 obj.cq.in = tmpl.cq;
1301 obj.cq.out = &cq_info;
1302 obj.qp.in = tmpl.qp;
1304 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_QP);
1309 if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
1311 "port %u wrong MLX5_CQE_SIZE environment variable"
1312 " value: it should be set to %u",
1313 dev->data->port_id, RTE_CACHE_LINE_SIZE);
1317 txq_data->cqe_n = log2above(cq_info.cqe_cnt);
1318 txq_data->cqe_s = 1 << txq_data->cqe_n;
1319 txq_data->cqe_m = txq_data->cqe_s - 1;
1320 txq_data->qp_num_8s = ((struct ibv_qp *)tmpl.qp)->qp_num << 8;
1321 txq_data->wqes = qp.sq.buf;
1322 txq_data->wqe_n = log2above(qp.sq.wqe_cnt);
1323 txq_data->wqe_s = 1 << txq_data->wqe_n;
1324 txq_data->wqe_m = txq_data->wqe_s - 1;
1325 txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s;
1326 txq_data->qp_db = &qp.dbrec[MLX5_SND_DBR];
1327 txq_data->cq_db = cq_info.dbrec;
1328 txq_data->cqes = (volatile struct mlx5_cqe *)cq_info.buf;
1329 txq_data->cq_ci = 0;
1330 txq_data->cq_pi = 0;
1331 txq_data->wqe_ci = 0;
1332 txq_data->wqe_pi = 0;
1333 txq_data->wqe_comp = 0;
1334 txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV;
1335 txq_data->fcqs = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1336 txq_data->cqe_s * sizeof(*txq_data->fcqs),
1337 RTE_CACHE_LINE_SIZE, txq_ctrl->socket);
1338 if (!txq_data->fcqs) {
1339 DRV_LOG(ERR, "port %u Tx queue %u cannot allocate memory (FCQ)",
1340 dev->data->port_id, idx);
1344 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1346 * If using DevX need to query and store TIS transport domain value.
1347 * This is done once per port.
1348 * Will use this value on Rx, when creating matching TIR.
1350 if (priv->config.devx && !priv->sh->tdn) {
1351 ret = mlx5_devx_cmd_qp_query_tis_td(tmpl.qp, qp.tisn,
1354 DRV_LOG(ERR, "Fail to query port %u Tx queue %u QP TIS "
1355 "transport domain", dev->data->port_id, idx);
1359 DRV_LOG(DEBUG, "port %u Tx queue %u TIS number %d "
1360 "transport domain %d", dev->data->port_id,
1361 idx, qp.tisn, priv->sh->tdn);
1365 txq_obj->qp = tmpl.qp;
1366 txq_obj->cq = tmpl.cq;
1367 rte_atomic32_inc(&txq_obj->refcnt);
1368 txq_ctrl->bf_reg = qp.bf.reg;
1369 if (qp.comp_mask & MLX5DV_QP_MASK_UAR_MMAP_OFFSET) {
1370 txq_ctrl->uar_mmap_offset = qp.uar_mmap_offset;
1371 DRV_LOG(DEBUG, "port %u: uar_mmap_offset 0x%"PRIx64,
1372 dev->data->port_id, txq_ctrl->uar_mmap_offset);
1375 "port %u failed to retrieve UAR info, invalid"
1377 dev->data->port_id);
1381 txq_uar_init(txq_ctrl);
1382 LIST_INSERT_HEAD(&priv->txqsobj, txq_obj, next);
1383 txq_obj->txq_ctrl = txq_ctrl;
1384 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1387 ret = rte_errno; /* Save rte_errno before cleanup. */
1389 claim_zero(mlx5_glue->destroy_cq(tmpl.cq));
1391 claim_zero(mlx5_glue->destroy_qp(tmpl.qp));
1392 if (txq_data && txq_data->fcqs) {
1393 mlx5_free(txq_data->fcqs);
1394 txq_data->fcqs = NULL;
1398 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1399 rte_errno = ret; /* Restore rte_errno. */
1404 * Get an Tx queue Verbs object.
1407 * Pointer to Ethernet device.
1409 * Queue index in DPDK Tx queue array.
1412 * The Verbs object if it exists.
1414 struct mlx5_txq_obj *
1415 mlx5_txq_obj_get(struct rte_eth_dev *dev, uint16_t idx)
1417 struct mlx5_priv *priv = dev->data->dev_private;
1418 struct mlx5_txq_ctrl *txq_ctrl;
1420 if (idx >= priv->txqs_n)
1422 if (!(*priv->txqs)[idx])
1424 txq_ctrl = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq);
1426 rte_atomic32_inc(&txq_ctrl->obj->refcnt);
1427 return txq_ctrl->obj;
1431 * Release an Tx verbs queue object.
1434 * Verbs Tx queue object.
1437 * 1 while a reference on it exists, 0 when freed.
1440 mlx5_txq_obj_release(struct mlx5_txq_obj *txq_obj)
1442 MLX5_ASSERT(txq_obj);
1443 if (rte_atomic32_dec_and_test(&txq_obj->refcnt)) {
1444 if (txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN) {
1446 claim_zero(mlx5_devx_cmd_destroy(txq_obj->tis));
1447 } else if (txq_obj->type == MLX5_TXQ_OBJ_TYPE_DEVX_SQ) {
1448 txq_release_sq_resources(txq_obj);
1450 claim_zero(mlx5_glue->destroy_qp(txq_obj->qp));
1451 claim_zero(mlx5_glue->destroy_cq(txq_obj->cq));
1453 if (txq_obj->txq_ctrl->txq.fcqs) {
1454 mlx5_free(txq_obj->txq_ctrl->txq.fcqs);
1455 txq_obj->txq_ctrl->txq.fcqs = NULL;
1457 LIST_REMOVE(txq_obj, next);
1465 * Verify the Verbs Tx queue list is empty
1468 * Pointer to Ethernet device.
1471 * The number of object not released.
1474 mlx5_txq_obj_verify(struct rte_eth_dev *dev)
1476 struct mlx5_priv *priv = dev->data->dev_private;
1478 struct mlx5_txq_obj *txq_obj;
1480 LIST_FOREACH(txq_obj, &priv->txqsobj, next) {
1481 DRV_LOG(DEBUG, "port %u Verbs Tx queue %u still referenced",
1482 dev->data->port_id, txq_obj->txq_ctrl->txq.idx);
1489 * Calculate the total number of WQEBB for Tx queue.
1491 * Simplified version of calc_sq_size() in rdma-core.
1494 * Pointer to Tx queue control structure.
1497 * The number of WQEBB.
1500 txq_calc_wqebb_cnt(struct mlx5_txq_ctrl *txq_ctrl)
1502 unsigned int wqe_size;
1503 const unsigned int desc = 1 << txq_ctrl->txq.elts_n;
1505 wqe_size = MLX5_WQE_CSEG_SIZE +
1506 MLX5_WQE_ESEG_SIZE +
1508 MLX5_ESEG_MIN_INLINE_SIZE +
1509 txq_ctrl->max_inline_data;
1510 return rte_align32pow2(wqe_size * desc) / MLX5_WQE_SIZE;
1514 * Calculate the maximal inline data size for Tx queue.
1517 * Pointer to Tx queue control structure.
1520 * The maximal inline data size.
1523 txq_calc_inline_max(struct mlx5_txq_ctrl *txq_ctrl)
1525 const unsigned int desc = 1 << txq_ctrl->txq.elts_n;
1526 struct mlx5_priv *priv = txq_ctrl->priv;
1527 unsigned int wqe_size;
1529 wqe_size = priv->sh->device_attr.max_qp_wr / desc;
1533 * This calculation is derived from tthe source of
1534 * mlx5_calc_send_wqe() in rdma_core library.
1536 wqe_size = wqe_size * MLX5_WQE_SIZE -
1537 MLX5_WQE_CSEG_SIZE -
1538 MLX5_WQE_ESEG_SIZE -
1541 MLX5_DSEG_MIN_INLINE_SIZE;
1546 * Set Tx queue parameters from device configuration.
1549 * Pointer to Tx queue control structure.
1552 txq_set_params(struct mlx5_txq_ctrl *txq_ctrl)
1554 struct mlx5_priv *priv = txq_ctrl->priv;
1555 struct mlx5_dev_config *config = &priv->config;
1556 unsigned int inlen_send; /* Inline data for ordinary SEND.*/
1557 unsigned int inlen_empw; /* Inline data for enhanced MPW. */
1558 unsigned int inlen_mode; /* Minimal required Inline data. */
1559 unsigned int txqs_inline; /* Min Tx queues to enable inline. */
1560 uint64_t dev_txoff = priv->dev_data->dev_conf.txmode.offloads;
1561 bool tso = txq_ctrl->txq.offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1562 DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1563 DEV_TX_OFFLOAD_GRE_TNL_TSO |
1564 DEV_TX_OFFLOAD_IP_TNL_TSO |
1565 DEV_TX_OFFLOAD_UDP_TNL_TSO);
1569 if (config->txqs_inline == MLX5_ARG_UNSET)
1571 #if defined(RTE_ARCH_ARM64)
1572 (priv->pci_dev->id.device_id ==
1573 PCI_DEVICE_ID_MELLANOX_CONNECTX5BF) ?
1574 MLX5_INLINE_MAX_TXQS_BLUEFIELD :
1576 MLX5_INLINE_MAX_TXQS;
1578 txqs_inline = (unsigned int)config->txqs_inline;
1579 inlen_send = (config->txq_inline_max == MLX5_ARG_UNSET) ?
1580 MLX5_SEND_DEF_INLINE_LEN :
1581 (unsigned int)config->txq_inline_max;
1582 inlen_empw = (config->txq_inline_mpw == MLX5_ARG_UNSET) ?
1583 MLX5_EMPW_DEF_INLINE_LEN :
1584 (unsigned int)config->txq_inline_mpw;
1585 inlen_mode = (config->txq_inline_min == MLX5_ARG_UNSET) ?
1586 0 : (unsigned int)config->txq_inline_min;
1587 if (config->mps != MLX5_MPW_ENHANCED && config->mps != MLX5_MPW)
1590 * If there is requested minimal amount of data to inline
1591 * we MUST enable inlining. This is a case for ConnectX-4
1592 * which usually requires L2 inlined for correct operating
1593 * and ConnectX-4 Lx which requires L2-L4 inlined to
1594 * support E-Switch Flows.
1597 if (inlen_mode <= MLX5_ESEG_MIN_INLINE_SIZE) {
1599 * Optimize minimal inlining for single
1600 * segment packets to fill one WQEBB
1603 temp = MLX5_ESEG_MIN_INLINE_SIZE;
1605 temp = inlen_mode - MLX5_ESEG_MIN_INLINE_SIZE;
1606 temp = RTE_ALIGN(temp, MLX5_WSEG_SIZE) +
1607 MLX5_ESEG_MIN_INLINE_SIZE;
1608 temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN);
1610 if (temp != inlen_mode) {
1612 "port %u minimal required inline setting"
1613 " aligned from %u to %u",
1614 PORT_ID(priv), inlen_mode, temp);
1619 * If port is configured to support VLAN insertion and device
1620 * does not support this feature by HW (for NICs before ConnectX-5
1621 * or in case of wqe_vlan_insert flag is not set) we must enable
1622 * data inline on all queues because it is supported by single
1625 txq_ctrl->txq.vlan_en = config->hw_vlan_insert;
1626 vlan_inline = (dev_txoff & DEV_TX_OFFLOAD_VLAN_INSERT) &&
1627 !config->hw_vlan_insert;
1629 * If there are few Tx queues it is prioritized
1630 * to save CPU cycles and disable data inlining at all.
1632 if (inlen_send && priv->txqs_n >= txqs_inline) {
1634 * The data sent with ordinal MLX5_OPCODE_SEND
1635 * may be inlined in Ethernet Segment, align the
1636 * length accordingly to fit entire WQEBBs.
1638 temp = RTE_MAX(inlen_send,
1639 MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE);
1640 temp -= MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE;
1641 temp = RTE_ALIGN(temp, MLX5_WQE_SIZE);
1642 temp += MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE;
1643 temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX +
1644 MLX5_ESEG_MIN_INLINE_SIZE -
1645 MLX5_WQE_CSEG_SIZE -
1646 MLX5_WQE_ESEG_SIZE -
1647 MLX5_WQE_DSEG_SIZE * 2);
1648 temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN);
1649 temp = RTE_MAX(temp, inlen_mode);
1650 if (temp != inlen_send) {
1652 "port %u ordinary send inline setting"
1653 " aligned from %u to %u",
1654 PORT_ID(priv), inlen_send, temp);
1658 * Not aligned to cache lines, but to WQEs.
1659 * First bytes of data (initial alignment)
1660 * is going to be copied explicitly at the
1661 * beginning of inlining buffer in Ethernet
1664 MLX5_ASSERT(inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
1665 MLX5_ASSERT(inlen_send <= MLX5_WQE_SIZE_MAX +
1666 MLX5_ESEG_MIN_INLINE_SIZE -
1667 MLX5_WQE_CSEG_SIZE -
1668 MLX5_WQE_ESEG_SIZE -
1669 MLX5_WQE_DSEG_SIZE * 2);
1670 } else if (inlen_mode) {
1672 * If minimal inlining is requested we must
1673 * enable inlining in general, despite the
1674 * number of configured queues. Ignore the
1675 * txq_inline_max devarg, this is not
1676 * full-featured inline.
1678 inlen_send = inlen_mode;
1680 } else if (vlan_inline) {
1682 * Hardware does not report offload for
1683 * VLAN insertion, we must enable data inline
1684 * to implement feature by software.
1686 inlen_send = MLX5_ESEG_MIN_INLINE_SIZE;
1692 txq_ctrl->txq.inlen_send = inlen_send;
1693 txq_ctrl->txq.inlen_mode = inlen_mode;
1694 txq_ctrl->txq.inlen_empw = 0;
1695 if (inlen_send && inlen_empw && priv->txqs_n >= txqs_inline) {
1697 * The data sent with MLX5_OPCODE_ENHANCED_MPSW
1698 * may be inlined in Data Segment, align the
1699 * length accordingly to fit entire WQEBBs.
1701 temp = RTE_MAX(inlen_empw,
1702 MLX5_WQE_SIZE + MLX5_DSEG_MIN_INLINE_SIZE);
1703 temp -= MLX5_DSEG_MIN_INLINE_SIZE;
1704 temp = RTE_ALIGN(temp, MLX5_WQE_SIZE);
1705 temp += MLX5_DSEG_MIN_INLINE_SIZE;
1706 temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX +
1707 MLX5_DSEG_MIN_INLINE_SIZE -
1708 MLX5_WQE_CSEG_SIZE -
1709 MLX5_WQE_ESEG_SIZE -
1710 MLX5_WQE_DSEG_SIZE);
1711 temp = RTE_MIN(temp, MLX5_EMPW_MAX_INLINE_LEN);
1712 if (temp != inlen_empw) {
1714 "port %u enhanced empw inline setting"
1715 " aligned from %u to %u",
1716 PORT_ID(priv), inlen_empw, temp);
1719 MLX5_ASSERT(inlen_empw >= MLX5_ESEG_MIN_INLINE_SIZE);
1720 MLX5_ASSERT(inlen_empw <= MLX5_WQE_SIZE_MAX +
1721 MLX5_DSEG_MIN_INLINE_SIZE -
1722 MLX5_WQE_CSEG_SIZE -
1723 MLX5_WQE_ESEG_SIZE -
1724 MLX5_WQE_DSEG_SIZE);
1725 txq_ctrl->txq.inlen_empw = inlen_empw;
1727 txq_ctrl->max_inline_data = RTE_MAX(inlen_send, inlen_empw);
1729 txq_ctrl->max_tso_header = MLX5_MAX_TSO_HEADER;
1730 txq_ctrl->max_inline_data = RTE_MAX(txq_ctrl->max_inline_data,
1731 MLX5_MAX_TSO_HEADER);
1732 txq_ctrl->txq.tso_en = 1;
1734 txq_ctrl->txq.tunnel_en = config->tunnel_en | config->swp;
1735 txq_ctrl->txq.swp_en = ((DEV_TX_OFFLOAD_IP_TNL_TSO |
1736 DEV_TX_OFFLOAD_UDP_TNL_TSO |
1737 DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM) &
1738 txq_ctrl->txq.offloads) && config->swp;
1742 * Adjust Tx queue data inline parameters for large queue sizes.
1743 * The data inline feature requires multiple WQEs to fit the packets,
1744 * and if the large amount of Tx descriptors is requested by application
1745 * the total WQE amount may exceed the hardware capabilities. If the
1746 * default inline setting are used we can try to adjust these ones and
1747 * meet the hardware requirements and not exceed the queue size.
1750 * Pointer to Tx queue control structure.
1753 * Zero on success, otherwise the parameters can not be adjusted.
1756 txq_adjust_params(struct mlx5_txq_ctrl *txq_ctrl)
1758 struct mlx5_priv *priv = txq_ctrl->priv;
1759 struct mlx5_dev_config *config = &priv->config;
1760 unsigned int max_inline;
1762 max_inline = txq_calc_inline_max(txq_ctrl);
1763 if (!txq_ctrl->txq.inlen_send) {
1765 * Inline data feature is not engaged at all.
1766 * There is nothing to adjust.
1770 if (txq_ctrl->max_inline_data <= max_inline) {
1772 * The requested inline data length does not
1773 * exceed queue capabilities.
1777 if (txq_ctrl->txq.inlen_mode > max_inline) {
1779 "minimal data inline requirements (%u) are not"
1780 " satisfied (%u) on port %u, try the smaller"
1781 " Tx queue size (%d)",
1782 txq_ctrl->txq.inlen_mode, max_inline,
1783 priv->dev_data->port_id,
1784 priv->sh->device_attr.max_qp_wr);
1787 if (txq_ctrl->txq.inlen_send > max_inline &&
1788 config->txq_inline_max != MLX5_ARG_UNSET &&
1789 config->txq_inline_max > (int)max_inline) {
1791 "txq_inline_max requirements (%u) are not"
1792 " satisfied (%u) on port %u, try the smaller"
1793 " Tx queue size (%d)",
1794 txq_ctrl->txq.inlen_send, max_inline,
1795 priv->dev_data->port_id,
1796 priv->sh->device_attr.max_qp_wr);
1799 if (txq_ctrl->txq.inlen_empw > max_inline &&
1800 config->txq_inline_mpw != MLX5_ARG_UNSET &&
1801 config->txq_inline_mpw > (int)max_inline) {
1803 "txq_inline_mpw requirements (%u) are not"
1804 " satisfied (%u) on port %u, try the smaller"
1805 " Tx queue size (%d)",
1806 txq_ctrl->txq.inlen_empw, max_inline,
1807 priv->dev_data->port_id,
1808 priv->sh->device_attr.max_qp_wr);
1811 if (txq_ctrl->txq.tso_en && max_inline < MLX5_MAX_TSO_HEADER) {
1813 "tso header inline requirements (%u) are not"
1814 " satisfied (%u) on port %u, try the smaller"
1815 " Tx queue size (%d)",
1816 MLX5_MAX_TSO_HEADER, max_inline,
1817 priv->dev_data->port_id,
1818 priv->sh->device_attr.max_qp_wr);
1821 if (txq_ctrl->txq.inlen_send > max_inline) {
1823 "adjust txq_inline_max (%u->%u)"
1824 " due to large Tx queue on port %u",
1825 txq_ctrl->txq.inlen_send, max_inline,
1826 priv->dev_data->port_id);
1827 txq_ctrl->txq.inlen_send = max_inline;
1829 if (txq_ctrl->txq.inlen_empw > max_inline) {
1831 "adjust txq_inline_mpw (%u->%u)"
1832 "due to large Tx queue on port %u",
1833 txq_ctrl->txq.inlen_empw, max_inline,
1834 priv->dev_data->port_id);
1835 txq_ctrl->txq.inlen_empw = max_inline;
1837 txq_ctrl->max_inline_data = RTE_MAX(txq_ctrl->txq.inlen_send,
1838 txq_ctrl->txq.inlen_empw);
1839 MLX5_ASSERT(txq_ctrl->max_inline_data <= max_inline);
1840 MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= max_inline);
1841 MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_send);
1842 MLX5_ASSERT(txq_ctrl->txq.inlen_mode <= txq_ctrl->txq.inlen_empw ||
1843 !txq_ctrl->txq.inlen_empw);
1851 * Create a DPDK Tx queue.
1854 * Pointer to Ethernet device.
1858 * Number of descriptors to configure in queue.
1860 * NUMA socket on which memory must be allocated.
1862 * Thresholds parameters.
1865 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1867 struct mlx5_txq_ctrl *
1868 mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1869 unsigned int socket, const struct rte_eth_txconf *conf)
1871 struct mlx5_priv *priv = dev->data->dev_private;
1872 struct mlx5_txq_ctrl *tmpl;
1874 tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl) +
1875 desc * sizeof(struct rte_mbuf *), 0, socket);
1880 if (mlx5_mr_btree_init(&tmpl->txq.mr_ctrl.cache_bh,
1881 MLX5_MR_BTREE_CACHE_N, socket)) {
1882 /* rte_errno is already set. */
1885 /* Save pointer of global generation number to check memory event. */
1886 tmpl->txq.mr_ctrl.dev_gen_ptr = &priv->sh->share_cache.dev_gen;
1887 MLX5_ASSERT(desc > MLX5_TX_COMP_THRESH);
1888 tmpl->txq.offloads = conf->offloads |
1889 dev->data->dev_conf.txmode.offloads;
1891 tmpl->socket = socket;
1892 tmpl->txq.elts_n = log2above(desc);
1893 tmpl->txq.elts_s = desc;
1894 tmpl->txq.elts_m = desc - 1;
1895 tmpl->txq.port_id = dev->data->port_id;
1896 tmpl->txq.idx = idx;
1897 txq_set_params(tmpl);
1898 if (txq_adjust_params(tmpl))
1900 if (txq_calc_wqebb_cnt(tmpl) >
1901 priv->sh->device_attr.max_qp_wr) {
1903 "port %u Tx WQEBB count (%d) exceeds the limit (%d),"
1904 " try smaller queue size",
1905 dev->data->port_id, txq_calc_wqebb_cnt(tmpl),
1906 priv->sh->device_attr.max_qp_wr);
1910 rte_atomic32_inc(&tmpl->refcnt);
1911 tmpl->type = MLX5_TXQ_TYPE_STANDARD;
1912 LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next);
1920 * Create a DPDK Tx hairpin queue.
1923 * Pointer to Ethernet device.
1927 * Number of descriptors to configure in queue.
1928 * @param hairpin_conf
1929 * The hairpin configuration.
1932 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1934 struct mlx5_txq_ctrl *
1935 mlx5_txq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1936 const struct rte_eth_hairpin_conf *hairpin_conf)
1938 struct mlx5_priv *priv = dev->data->dev_private;
1939 struct mlx5_txq_ctrl *tmpl;
1941 tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl), 0,
1948 tmpl->socket = SOCKET_ID_ANY;
1949 tmpl->txq.elts_n = log2above(desc);
1950 tmpl->txq.port_id = dev->data->port_id;
1951 tmpl->txq.idx = idx;
1952 tmpl->hairpin_conf = *hairpin_conf;
1953 tmpl->type = MLX5_TXQ_TYPE_HAIRPIN;
1954 rte_atomic32_inc(&tmpl->refcnt);
1955 LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next);
1963 * Pointer to Ethernet device.
1968 * A pointer to the queue if it exists.
1970 struct mlx5_txq_ctrl *
1971 mlx5_txq_get(struct rte_eth_dev *dev, uint16_t idx)
1973 struct mlx5_priv *priv = dev->data->dev_private;
1974 struct mlx5_txq_ctrl *ctrl = NULL;
1976 if ((*priv->txqs)[idx]) {
1977 ctrl = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl,
1979 mlx5_txq_obj_get(dev, idx);
1980 rte_atomic32_inc(&ctrl->refcnt);
1986 * Release a Tx queue.
1989 * Pointer to Ethernet device.
1994 * 1 while a reference on it exists, 0 when freed.
1997 mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx)
1999 struct mlx5_priv *priv = dev->data->dev_private;
2000 struct mlx5_txq_ctrl *txq;
2002 if (!(*priv->txqs)[idx])
2004 txq = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq);
2005 if (txq->obj && !mlx5_txq_obj_release(txq->obj))
2007 if (rte_atomic32_dec_and_test(&txq->refcnt)) {
2009 mlx5_mr_btree_free(&txq->txq.mr_ctrl.cache_bh);
2010 LIST_REMOVE(txq, next);
2012 (*priv->txqs)[idx] = NULL;
2013 dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STOPPED;
2020 * Verify if the queue can be released.
2023 * Pointer to Ethernet device.
2028 * 1 if the queue can be released.
2031 mlx5_txq_releasable(struct rte_eth_dev *dev, uint16_t idx)
2033 struct mlx5_priv *priv = dev->data->dev_private;
2034 struct mlx5_txq_ctrl *txq;
2036 if (!(*priv->txqs)[idx])
2038 txq = container_of((*priv->txqs)[idx], struct mlx5_txq_ctrl, txq);
2039 return (rte_atomic32_read(&txq->refcnt) == 1);
2043 * Verify the Tx Queue list is empty
2046 * Pointer to Ethernet device.
2049 * The number of object not released.
2052 mlx5_txq_verify(struct rte_eth_dev *dev)
2054 struct mlx5_priv *priv = dev->data->dev_private;
2055 struct mlx5_txq_ctrl *txq_ctrl;
2058 LIST_FOREACH(txq_ctrl, &priv->txqsctrl, next) {
2059 DRV_LOG(DEBUG, "port %u Tx queue %u still referenced",
2060 dev->data->port_id, txq_ctrl->txq.idx);
2067 * Set the Tx queue dynamic timestamp (mask and offset)
2070 * Pointer to the Ethernet device structure.
2073 mlx5_txq_dynf_timestamp_set(struct rte_eth_dev *dev)
2075 struct mlx5_priv *priv = dev->data->dev_private;
2076 struct mlx5_dev_ctx_shared *sh = priv->sh;
2077 struct mlx5_txq_data *data;
2082 nbit = rte_mbuf_dynflag_lookup
2083 (RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME, NULL);
2084 off = rte_mbuf_dynfield_lookup
2085 (RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL);
2086 if (nbit > 0 && off >= 0 && sh->txpp.refcnt)
2087 mask = 1ULL << nbit;
2088 for (i = 0; i != priv->txqs_n; ++i) {
2089 data = (*priv->txqs)[i];
2093 data->ts_mask = mask;
2094 data->ts_offset = off;