1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
9 #include <ethdev_driver.h>
10 #include <rte_interrupts.h>
11 #include <rte_alarm.h>
12 #include <rte_cycles.h>
14 #include <mlx5_malloc.h>
17 #include "mlx5_flow.h"
20 #include "mlx5_utils.h"
21 #include "rte_pmd_mlx5.h"
24 * Stop traffic on Tx queues.
27 * Pointer to Ethernet device structure.
30 mlx5_txq_stop(struct rte_eth_dev *dev)
32 struct mlx5_priv *priv = dev->data->dev_private;
35 for (i = 0; i != priv->txqs_n; ++i)
36 mlx5_txq_release(dev, i);
40 * Start traffic on Tx queues.
43 * Pointer to Ethernet device structure.
46 * 0 on success, a negative errno value otherwise and rte_errno is set.
49 mlx5_txq_start(struct rte_eth_dev *dev)
51 struct mlx5_priv *priv = dev->data->dev_private;
55 for (i = 0; i != priv->txqs_n; ++i) {
56 struct mlx5_txq_ctrl *txq_ctrl = mlx5_txq_get(dev, i);
57 struct mlx5_txq_data *txq_data = &txq_ctrl->txq;
58 uint32_t flags = MLX5_MEM_RTE | MLX5_MEM_ZERO;
62 if (!txq_ctrl->is_hairpin)
63 txq_alloc_elts(txq_ctrl);
64 MLX5_ASSERT(!txq_ctrl->obj);
65 txq_ctrl->obj = mlx5_malloc(flags, sizeof(struct mlx5_txq_obj),
68 DRV_LOG(ERR, "Port %u Tx queue %u cannot allocate "
69 "memory resources.", dev->data->port_id,
74 ret = priv->obj_ops.txq_obj_new(dev, i);
76 mlx5_free(txq_ctrl->obj);
80 if (!txq_ctrl->is_hairpin) {
81 size_t size = txq_data->cqe_s * sizeof(*txq_data->fcqs);
83 txq_data->fcqs = mlx5_malloc(flags, size,
86 if (!txq_data->fcqs) {
87 DRV_LOG(ERR, "Port %u Tx queue %u cannot "
88 "allocate memory (FCQ).",
89 dev->data->port_id, i);
94 DRV_LOG(DEBUG, "Port %u txq %u updated with %p.",
95 dev->data->port_id, i, (void *)&txq_ctrl->obj);
96 LIST_INSERT_HEAD(&priv->txqsobj, txq_ctrl->obj, next);
100 ret = rte_errno; /* Save rte_errno before cleanup. */
102 mlx5_txq_release(dev, i);
104 rte_errno = ret; /* Restore rte_errno. */
109 * Register Rx queue mempools and fill the Rx queue cache.
110 * This function tolerates repeated mempool registration.
112 * @param[in] rxq_ctrl
113 * Rx queue control data.
116 * 0 on success, (-1) on failure and rte_errno is set.
119 mlx5_rxq_mempool_register(struct mlx5_rxq_ctrl *rxq_ctrl)
121 struct rte_mempool *mp;
125 mlx5_mr_flush_local_cache(&rxq_ctrl->rxq.mr_ctrl);
126 /* MPRQ mempool is registered on creation, just fill the cache. */
127 if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
128 return mlx5_mr_mempool_populate_cache(&rxq_ctrl->rxq.mr_ctrl,
129 rxq_ctrl->rxq.mprq_mp);
130 for (s = 0; s < rxq_ctrl->rxq.rxseg_n; s++) {
133 mp = rxq_ctrl->rxq.rxseg[s].mp;
134 is_extmem = (rte_pktmbuf_priv_flags(mp) &
135 RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF) != 0;
136 ret = mlx5_mr_mempool_register(rxq_ctrl->sh->cdev, mp,
138 if (ret < 0 && rte_errno != EEXIST)
140 ret = mlx5_mr_mempool_populate_cache(&rxq_ctrl->rxq.mr_ctrl,
149 * Stop traffic on Rx queues.
152 * Pointer to Ethernet device structure.
155 mlx5_rxq_stop(struct rte_eth_dev *dev)
157 struct mlx5_priv *priv = dev->data->dev_private;
160 for (i = 0; i != priv->rxqs_n; ++i)
161 mlx5_rxq_release(dev, i);
165 mlx5_rxq_ctrl_prepare(struct rte_eth_dev *dev, struct mlx5_rxq_ctrl *rxq_ctrl,
170 if (!rxq_ctrl->is_hairpin) {
172 * Pre-register the mempools. Regardless of whether
173 * the implicit registration is enabled or not,
174 * Rx mempool destruction is tracked to free MRs.
176 if (mlx5_rxq_mempool_register(rxq_ctrl) < 0)
178 ret = rxq_alloc_elts(rxq_ctrl);
182 MLX5_ASSERT(!rxq_ctrl->obj);
183 rxq_ctrl->obj = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
184 sizeof(*rxq_ctrl->obj), 0,
186 if (!rxq_ctrl->obj) {
187 DRV_LOG(ERR, "Port %u Rx queue %u can't allocate resources.",
188 dev->data->port_id, idx);
192 DRV_LOG(DEBUG, "Port %u rxq %u updated with %p.", dev->data->port_id,
193 idx, (void *)&rxq_ctrl->obj);
198 * Start traffic on Rx queues.
201 * Pointer to Ethernet device structure.
204 * 0 on success, a negative errno value otherwise and rte_errno is set.
207 mlx5_rxq_start(struct rte_eth_dev *dev)
209 struct mlx5_priv *priv = dev->data->dev_private;
213 /* Allocate/reuse/resize mempool for Multi-Packet RQ. */
214 if (mlx5_mprq_alloc_mp(dev)) {
215 /* Should not release Rx queues but return immediately. */
218 DRV_LOG(DEBUG, "Port %u dev_cap.max_qp_wr is %d.",
219 dev->data->port_id, priv->sh->dev_cap.max_qp_wr);
220 DRV_LOG(DEBUG, "Port %u dev_cap.max_sge is %d.",
221 dev->data->port_id, priv->sh->dev_cap.max_sge);
222 for (i = 0; i != priv->rxqs_n; ++i) {
223 struct mlx5_rxq_priv *rxq = mlx5_rxq_ref(dev, i);
224 struct mlx5_rxq_ctrl *rxq_ctrl;
228 rxq_ctrl = rxq->ctrl;
229 if (!rxq_ctrl->started) {
230 if (mlx5_rxq_ctrl_prepare(dev, rxq_ctrl, i) < 0)
232 LIST_INSERT_HEAD(&priv->rxqsobj, rxq_ctrl->obj, next);
234 ret = priv->obj_ops.rxq_obj_new(rxq);
236 mlx5_free(rxq_ctrl->obj);
237 rxq_ctrl->obj = NULL;
240 rxq_ctrl->started = true;
244 ret = rte_errno; /* Save rte_errno before cleanup. */
246 mlx5_rxq_release(dev, i);
248 rte_errno = ret; /* Restore rte_errno. */
253 * Binds Tx queues to Rx queues for hairpin.
255 * Binds Tx queues to the target Rx queues.
258 * Pointer to Ethernet device structure.
261 * 0 on success, a negative errno value otherwise and rte_errno is set.
264 mlx5_hairpin_auto_bind(struct rte_eth_dev *dev)
266 struct mlx5_priv *priv = dev->data->dev_private;
267 struct mlx5_devx_modify_sq_attr sq_attr = { 0 };
268 struct mlx5_devx_modify_rq_attr rq_attr = { 0 };
269 struct mlx5_txq_ctrl *txq_ctrl;
270 struct mlx5_rxq_priv *rxq;
271 struct mlx5_rxq_ctrl *rxq_ctrl;
272 struct mlx5_devx_obj *sq;
273 struct mlx5_devx_obj *rq;
276 bool need_auto = false;
277 uint16_t self_port = dev->data->port_id;
279 for (i = 0; i != priv->txqs_n; ++i) {
280 txq_ctrl = mlx5_txq_get(dev, i);
283 if (!txq_ctrl->is_hairpin ||
284 txq_ctrl->hairpin_conf.peers[0].port != self_port) {
285 mlx5_txq_release(dev, i);
288 if (txq_ctrl->hairpin_conf.manual_bind) {
289 mlx5_txq_release(dev, i);
293 mlx5_txq_release(dev, i);
297 for (i = 0; i != priv->txqs_n; ++i) {
298 txq_ctrl = mlx5_txq_get(dev, i);
301 /* Skip hairpin queues with other peer ports. */
302 if (!txq_ctrl->is_hairpin ||
303 txq_ctrl->hairpin_conf.peers[0].port != self_port) {
304 mlx5_txq_release(dev, i);
307 if (!txq_ctrl->obj) {
309 DRV_LOG(ERR, "port %u no txq object found: %d",
310 dev->data->port_id, i);
311 mlx5_txq_release(dev, i);
314 sq = txq_ctrl->obj->sq;
315 rxq = mlx5_rxq_get(dev, txq_ctrl->hairpin_conf.peers[0].queue);
317 mlx5_txq_release(dev, i);
319 DRV_LOG(ERR, "port %u no rxq object found: %d",
321 txq_ctrl->hairpin_conf.peers[0].queue);
324 rxq_ctrl = rxq->ctrl;
325 if (!rxq_ctrl->is_hairpin ||
326 rxq->hairpin_conf.peers[0].queue != i) {
328 DRV_LOG(ERR, "port %u Tx queue %d can't be binded to "
329 "Rx queue %d", dev->data->port_id,
330 i, txq_ctrl->hairpin_conf.peers[0].queue);
333 rq = rxq_ctrl->obj->rq;
336 DRV_LOG(ERR, "port %u hairpin no matching rxq: %d",
338 txq_ctrl->hairpin_conf.peers[0].queue);
341 sq_attr.state = MLX5_SQC_STATE_RDY;
342 sq_attr.sq_state = MLX5_SQC_STATE_RST;
343 sq_attr.hairpin_peer_rq = rq->id;
344 sq_attr.hairpin_peer_vhca =
345 priv->sh->cdev->config.hca_attr.vhca_id;
346 ret = mlx5_devx_cmd_modify_sq(sq, &sq_attr);
349 rq_attr.state = MLX5_SQC_STATE_RDY;
350 rq_attr.rq_state = MLX5_SQC_STATE_RST;
351 rq_attr.hairpin_peer_sq = sq->id;
352 rq_attr.hairpin_peer_vhca =
353 priv->sh->cdev->config.hca_attr.vhca_id;
354 ret = mlx5_devx_cmd_modify_rq(rq, &rq_attr);
357 /* Qs with auto-bind will be destroyed directly. */
358 rxq->hairpin_status = 1;
359 txq_ctrl->hairpin_status = 1;
360 mlx5_txq_release(dev, i);
364 mlx5_txq_release(dev, i);
369 * Fetch the peer queue's SW & HW information.
372 * Pointer to Ethernet device structure.
374 * Index of the queue to fetch the information.
375 * @param current_info
376 * Pointer to the input peer information, not used currently.
378 * Pointer to the structure to store the information, output.
380 * Positive to get the RxQ information, zero to get the TxQ information.
383 * 0 on success, a negative errno value otherwise and rte_errno is set.
386 mlx5_hairpin_queue_peer_update(struct rte_eth_dev *dev, uint16_t peer_queue,
387 struct rte_hairpin_peer_info *current_info,
388 struct rte_hairpin_peer_info *peer_info,
391 struct mlx5_priv *priv = dev->data->dev_private;
392 RTE_SET_USED(current_info);
394 if (dev->data->dev_started == 0) {
396 DRV_LOG(ERR, "peer port %u is not started",
401 * Peer port used as egress. In the current design, hairpin Tx queue
402 * will be bound to the peer Rx queue. Indeed, only the information of
403 * peer Rx queue needs to be fetched.
405 if (direction == 0) {
406 struct mlx5_txq_ctrl *txq_ctrl;
408 txq_ctrl = mlx5_txq_get(dev, peer_queue);
409 if (txq_ctrl == NULL) {
411 DRV_LOG(ERR, "Failed to get port %u Tx queue %d",
412 dev->data->port_id, peer_queue);
415 if (!txq_ctrl->is_hairpin) {
417 DRV_LOG(ERR, "port %u queue %d is not a hairpin Txq",
418 dev->data->port_id, peer_queue);
419 mlx5_txq_release(dev, peer_queue);
422 if (txq_ctrl->obj == NULL || txq_ctrl->obj->sq == NULL) {
424 DRV_LOG(ERR, "port %u no Txq object found: %d",
425 dev->data->port_id, peer_queue);
426 mlx5_txq_release(dev, peer_queue);
429 peer_info->qp_id = txq_ctrl->obj->sq->id;
430 peer_info->vhca_id = priv->sh->cdev->config.hca_attr.vhca_id;
431 /* 1-to-1 mapping, only the first one is used. */
432 peer_info->peer_q = txq_ctrl->hairpin_conf.peers[0].queue;
433 peer_info->tx_explicit = txq_ctrl->hairpin_conf.tx_explicit;
434 peer_info->manual_bind = txq_ctrl->hairpin_conf.manual_bind;
435 mlx5_txq_release(dev, peer_queue);
436 } else { /* Peer port used as ingress. */
437 struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, peer_queue);
438 struct mlx5_rxq_ctrl *rxq_ctrl;
442 DRV_LOG(ERR, "Failed to get port %u Rx queue %d",
443 dev->data->port_id, peer_queue);
446 rxq_ctrl = rxq->ctrl;
447 if (!rxq_ctrl->is_hairpin) {
449 DRV_LOG(ERR, "port %u queue %d is not a hairpin Rxq",
450 dev->data->port_id, peer_queue);
453 if (rxq_ctrl->obj == NULL || rxq_ctrl->obj->rq == NULL) {
455 DRV_LOG(ERR, "port %u no Rxq object found: %d",
456 dev->data->port_id, peer_queue);
459 peer_info->qp_id = rxq_ctrl->obj->rq->id;
460 peer_info->vhca_id = priv->sh->cdev->config.hca_attr.vhca_id;
461 peer_info->peer_q = rxq->hairpin_conf.peers[0].queue;
462 peer_info->tx_explicit = rxq->hairpin_conf.tx_explicit;
463 peer_info->manual_bind = rxq->hairpin_conf.manual_bind;
469 * Bind the hairpin queue with the peer HW information.
470 * This needs to be called twice both for Tx and Rx queues of a pair.
471 * If the queue is already bound, it is considered successful.
474 * Pointer to Ethernet device structure.
476 * Index of the queue to change the HW configuration to bind.
478 * Pointer to information of the peer queue.
480 * Positive to configure the TxQ, zero to configure the RxQ.
483 * 0 on success, a negative errno value otherwise and rte_errno is set.
486 mlx5_hairpin_queue_peer_bind(struct rte_eth_dev *dev, uint16_t cur_queue,
487 struct rte_hairpin_peer_info *peer_info,
493 * Consistency checking of the peer queue: opposite direction is used
494 * to get the peer queue info with ethdev port ID, no need to check.
496 if (peer_info->peer_q != cur_queue) {
498 DRV_LOG(ERR, "port %u queue %d and peer queue %d mismatch",
499 dev->data->port_id, cur_queue, peer_info->peer_q);
502 if (direction != 0) {
503 struct mlx5_txq_ctrl *txq_ctrl;
504 struct mlx5_devx_modify_sq_attr sq_attr = { 0 };
506 txq_ctrl = mlx5_txq_get(dev, cur_queue);
507 if (txq_ctrl == NULL) {
509 DRV_LOG(ERR, "Failed to get port %u Tx queue %d",
510 dev->data->port_id, cur_queue);
513 if (!txq_ctrl->is_hairpin) {
515 DRV_LOG(ERR, "port %u queue %d not a hairpin Txq",
516 dev->data->port_id, cur_queue);
517 mlx5_txq_release(dev, cur_queue);
520 if (txq_ctrl->obj == NULL || txq_ctrl->obj->sq == NULL) {
522 DRV_LOG(ERR, "port %u no Txq object found: %d",
523 dev->data->port_id, cur_queue);
524 mlx5_txq_release(dev, cur_queue);
527 if (txq_ctrl->hairpin_status != 0) {
528 DRV_LOG(DEBUG, "port %u Tx queue %d is already bound",
529 dev->data->port_id, cur_queue);
530 mlx5_txq_release(dev, cur_queue);
534 * All queues' of one port consistency checking is done in the
535 * bind() function, and that is optional.
537 if (peer_info->tx_explicit !=
538 txq_ctrl->hairpin_conf.tx_explicit) {
540 DRV_LOG(ERR, "port %u Tx queue %d and peer Tx rule mode"
541 " mismatch", dev->data->port_id, cur_queue);
542 mlx5_txq_release(dev, cur_queue);
545 if (peer_info->manual_bind !=
546 txq_ctrl->hairpin_conf.manual_bind) {
548 DRV_LOG(ERR, "port %u Tx queue %d and peer binding mode"
549 " mismatch", dev->data->port_id, cur_queue);
550 mlx5_txq_release(dev, cur_queue);
553 sq_attr.state = MLX5_SQC_STATE_RDY;
554 sq_attr.sq_state = MLX5_SQC_STATE_RST;
555 sq_attr.hairpin_peer_rq = peer_info->qp_id;
556 sq_attr.hairpin_peer_vhca = peer_info->vhca_id;
557 ret = mlx5_devx_cmd_modify_sq(txq_ctrl->obj->sq, &sq_attr);
559 txq_ctrl->hairpin_status = 1;
560 mlx5_txq_release(dev, cur_queue);
562 struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, cur_queue);
563 struct mlx5_rxq_ctrl *rxq_ctrl;
564 struct mlx5_devx_modify_rq_attr rq_attr = { 0 };
568 DRV_LOG(ERR, "Failed to get port %u Rx queue %d",
569 dev->data->port_id, cur_queue);
572 rxq_ctrl = rxq->ctrl;
573 if (!rxq_ctrl->is_hairpin) {
575 DRV_LOG(ERR, "port %u queue %d not a hairpin Rxq",
576 dev->data->port_id, cur_queue);
579 if (rxq_ctrl->obj == NULL || rxq_ctrl->obj->rq == NULL) {
581 DRV_LOG(ERR, "port %u no Rxq object found: %d",
582 dev->data->port_id, cur_queue);
585 if (rxq->hairpin_status != 0) {
586 DRV_LOG(DEBUG, "port %u Rx queue %d is already bound",
587 dev->data->port_id, cur_queue);
590 if (peer_info->tx_explicit !=
591 rxq->hairpin_conf.tx_explicit) {
593 DRV_LOG(ERR, "port %u Rx queue %d and peer Tx rule mode"
594 " mismatch", dev->data->port_id, cur_queue);
597 if (peer_info->manual_bind !=
598 rxq->hairpin_conf.manual_bind) {
600 DRV_LOG(ERR, "port %u Rx queue %d and peer binding mode"
601 " mismatch", dev->data->port_id, cur_queue);
604 rq_attr.state = MLX5_SQC_STATE_RDY;
605 rq_attr.rq_state = MLX5_SQC_STATE_RST;
606 rq_attr.hairpin_peer_sq = peer_info->qp_id;
607 rq_attr.hairpin_peer_vhca = peer_info->vhca_id;
608 ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq, &rq_attr);
610 rxq->hairpin_status = 1;
616 * Unbind the hairpin queue and reset its HW configuration.
617 * This needs to be called twice both for Tx and Rx queues of a pair.
618 * If the queue is already unbound, it is considered successful.
621 * Pointer to Ethernet device structure.
623 * Index of the queue to change the HW configuration to unbind.
625 * Positive to reset the TxQ, zero to reset the RxQ.
628 * 0 on success, a negative errno value otherwise and rte_errno is set.
631 mlx5_hairpin_queue_peer_unbind(struct rte_eth_dev *dev, uint16_t cur_queue,
636 if (direction != 0) {
637 struct mlx5_txq_ctrl *txq_ctrl;
638 struct mlx5_devx_modify_sq_attr sq_attr = { 0 };
640 txq_ctrl = mlx5_txq_get(dev, cur_queue);
641 if (txq_ctrl == NULL) {
643 DRV_LOG(ERR, "Failed to get port %u Tx queue %d",
644 dev->data->port_id, cur_queue);
647 if (!txq_ctrl->is_hairpin) {
649 DRV_LOG(ERR, "port %u queue %d not a hairpin Txq",
650 dev->data->port_id, cur_queue);
651 mlx5_txq_release(dev, cur_queue);
654 /* Already unbound, return success before obj checking. */
655 if (txq_ctrl->hairpin_status == 0) {
656 DRV_LOG(DEBUG, "port %u Tx queue %d is already unbound",
657 dev->data->port_id, cur_queue);
658 mlx5_txq_release(dev, cur_queue);
661 if (!txq_ctrl->obj || !txq_ctrl->obj->sq) {
663 DRV_LOG(ERR, "port %u no Txq object found: %d",
664 dev->data->port_id, cur_queue);
665 mlx5_txq_release(dev, cur_queue);
668 sq_attr.state = MLX5_SQC_STATE_RST;
669 sq_attr.sq_state = MLX5_SQC_STATE_RST;
670 ret = mlx5_devx_cmd_modify_sq(txq_ctrl->obj->sq, &sq_attr);
672 txq_ctrl->hairpin_status = 0;
673 mlx5_txq_release(dev, cur_queue);
675 struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, cur_queue);
676 struct mlx5_rxq_ctrl *rxq_ctrl;
677 struct mlx5_devx_modify_rq_attr rq_attr = { 0 };
681 DRV_LOG(ERR, "Failed to get port %u Rx queue %d",
682 dev->data->port_id, cur_queue);
685 rxq_ctrl = rxq->ctrl;
686 if (!rxq_ctrl->is_hairpin) {
688 DRV_LOG(ERR, "port %u queue %d not a hairpin Rxq",
689 dev->data->port_id, cur_queue);
692 if (rxq->hairpin_status == 0) {
693 DRV_LOG(DEBUG, "port %u Rx queue %d is already unbound",
694 dev->data->port_id, cur_queue);
697 if (rxq_ctrl->obj == NULL || rxq_ctrl->obj->rq == NULL) {
699 DRV_LOG(ERR, "port %u no Rxq object found: %d",
700 dev->data->port_id, cur_queue);
703 rq_attr.state = MLX5_SQC_STATE_RST;
704 rq_attr.rq_state = MLX5_SQC_STATE_RST;
705 ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq, &rq_attr);
707 rxq->hairpin_status = 0;
713 * Bind the hairpin port pairs, from the Tx to the peer Rx.
714 * This function only supports to bind the Tx to one Rx.
717 * Pointer to Ethernet device structure.
719 * Port identifier of the Rx port.
722 * 0 on success, a negative errno value otherwise and rte_errno is set.
725 mlx5_hairpin_bind_single_port(struct rte_eth_dev *dev, uint16_t rx_port)
727 struct mlx5_priv *priv = dev->data->dev_private;
729 struct mlx5_txq_ctrl *txq_ctrl;
731 struct rte_hairpin_peer_info peer = {0xffffff};
732 struct rte_hairpin_peer_info cur;
733 const struct rte_eth_hairpin_conf *conf;
735 uint16_t local_port = priv->dev_data->port_id;
740 if (mlx5_eth_find_next(rx_port, dev->device) != rx_port) {
742 DRV_LOG(ERR, "Rx port %u does not belong to mlx5", rx_port);
746 * Before binding TxQ to peer RxQ, first round loop will be used for
747 * checking the queues' configuration consistency. This would be a
748 * little time consuming but better than doing the rollback.
750 for (i = 0; i != priv->txqs_n; i++) {
751 txq_ctrl = mlx5_txq_get(dev, i);
752 if (txq_ctrl == NULL)
754 if (!txq_ctrl->is_hairpin) {
755 mlx5_txq_release(dev, i);
759 * All hairpin Tx queues of a single port that connected to the
760 * same peer Rx port should have the same "auto binding" and
761 * "implicit Tx flow" modes.
762 * Peer consistency checking will be done in per queue binding.
764 conf = &txq_ctrl->hairpin_conf;
765 if (conf->peers[0].port == rx_port) {
767 manual = conf->manual_bind;
768 explicit = conf->tx_explicit;
770 if (manual != conf->manual_bind ||
771 explicit != conf->tx_explicit) {
773 DRV_LOG(ERR, "port %u queue %d mode"
774 " mismatch: %u %u, %u %u",
775 local_port, i, manual,
776 conf->manual_bind, explicit,
778 mlx5_txq_release(dev, i);
784 mlx5_txq_release(dev, i);
786 /* Once no queue is configured, success is returned directly. */
789 /* All the hairpin TX queues need to be traversed again. */
790 for (i = 0; i != priv->txqs_n; i++) {
791 txq_ctrl = mlx5_txq_get(dev, i);
792 if (txq_ctrl == NULL)
794 if (!txq_ctrl->is_hairpin) {
795 mlx5_txq_release(dev, i);
798 if (txq_ctrl->hairpin_conf.peers[0].port != rx_port) {
799 mlx5_txq_release(dev, i);
802 rx_queue = txq_ctrl->hairpin_conf.peers[0].queue;
804 * Fetch peer RxQ's information.
805 * No need to pass the information of the current queue.
807 ret = rte_eth_hairpin_queue_peer_update(rx_port, rx_queue,
810 mlx5_txq_release(dev, i);
813 /* Accessing its own device, inside mlx5 PMD. */
814 ret = mlx5_hairpin_queue_peer_bind(dev, i, &peer, 1);
816 mlx5_txq_release(dev, i);
819 /* Pass TxQ's information to peer RxQ and try binding. */
820 cur.peer_q = rx_queue;
821 cur.qp_id = txq_ctrl->obj->sq->id;
822 cur.vhca_id = priv->sh->cdev->config.hca_attr.vhca_id;
823 cur.tx_explicit = txq_ctrl->hairpin_conf.tx_explicit;
824 cur.manual_bind = txq_ctrl->hairpin_conf.manual_bind;
826 * In order to access another device in a proper way, RTE level
827 * private function is needed.
829 ret = rte_eth_hairpin_queue_peer_bind(rx_port, rx_queue,
832 mlx5_txq_release(dev, i);
835 mlx5_txq_release(dev, i);
840 * Do roll-back process for the queues already bound.
841 * No need to check the return value of the queue unbind function.
844 /* No validation is needed here. */
845 txq_ctrl = mlx5_txq_get(dev, i);
846 if (txq_ctrl == NULL)
848 rx_queue = txq_ctrl->hairpin_conf.peers[0].queue;
849 rte_eth_hairpin_queue_peer_unbind(rx_port, rx_queue, 0);
850 mlx5_hairpin_queue_peer_unbind(dev, i, 1);
851 mlx5_txq_release(dev, i);
857 * Unbind the hairpin port pair, HW configuration of both devices will be clear
858 * and status will be reset for all the queues used between them.
859 * This function only supports to unbind the Tx from one Rx.
862 * Pointer to Ethernet device structure.
864 * Port identifier of the Rx port.
867 * 0 on success, a negative errno value otherwise and rte_errno is set.
870 mlx5_hairpin_unbind_single_port(struct rte_eth_dev *dev, uint16_t rx_port)
872 struct mlx5_priv *priv = dev->data->dev_private;
873 struct mlx5_txq_ctrl *txq_ctrl;
876 uint16_t cur_port = priv->dev_data->port_id;
878 if (mlx5_eth_find_next(rx_port, dev->device) != rx_port) {
880 DRV_LOG(ERR, "Rx port %u does not belong to mlx5", rx_port);
883 for (i = 0; i != priv->txqs_n; i++) {
886 txq_ctrl = mlx5_txq_get(dev, i);
887 if (txq_ctrl == NULL)
889 if (!txq_ctrl->is_hairpin) {
890 mlx5_txq_release(dev, i);
893 if (txq_ctrl->hairpin_conf.peers[0].port != rx_port) {
894 mlx5_txq_release(dev, i);
897 /* Indeed, only the first used queue needs to be checked. */
898 if (txq_ctrl->hairpin_conf.manual_bind == 0) {
899 if (cur_port != rx_port) {
901 DRV_LOG(ERR, "port %u and port %u are in"
902 " auto-bind mode", cur_port, rx_port);
903 mlx5_txq_release(dev, i);
909 rx_queue = txq_ctrl->hairpin_conf.peers[0].queue;
910 mlx5_txq_release(dev, i);
911 ret = rte_eth_hairpin_queue_peer_unbind(rx_port, rx_queue, 0);
913 DRV_LOG(ERR, "port %u Rx queue %d unbind - failure",
917 ret = mlx5_hairpin_queue_peer_unbind(dev, i, 1);
919 DRV_LOG(ERR, "port %u Tx queue %d unbind - failure",
928 * Bind hairpin ports, Rx could be all ports when using RTE_MAX_ETHPORTS.
929 * @see mlx5_hairpin_bind_single_port()
932 mlx5_hairpin_bind(struct rte_eth_dev *dev, uint16_t rx_port)
938 * If the Rx port has no hairpin configuration with the current port,
939 * the binding will be skipped in the called function of single port.
940 * Device started status will be checked only before the queue
941 * information updating.
943 if (rx_port == RTE_MAX_ETHPORTS) {
944 MLX5_ETH_FOREACH_DEV(p, dev->device) {
945 ret = mlx5_hairpin_bind_single_port(dev, p);
951 return mlx5_hairpin_bind_single_port(dev, rx_port);
954 MLX5_ETH_FOREACH_DEV(pp, dev->device)
956 mlx5_hairpin_unbind_single_port(dev, pp);
961 * Unbind hairpin ports, Rx could be all ports when using RTE_MAX_ETHPORTS.
962 * @see mlx5_hairpin_unbind_single_port()
965 mlx5_hairpin_unbind(struct rte_eth_dev *dev, uint16_t rx_port)
970 if (rx_port == RTE_MAX_ETHPORTS)
971 MLX5_ETH_FOREACH_DEV(p, dev->device) {
972 ret = mlx5_hairpin_unbind_single_port(dev, p);
977 ret = mlx5_hairpin_unbind_single_port(dev, rx_port);
982 * DPDK callback to get the hairpin peer ports list.
983 * This will return the actual number of peer ports and save the identifiers
984 * into the array (sorted, may be different from that when setting up the
985 * hairpin peer queues).
986 * The peer port ID could be the same as the port ID of the current device.
989 * Pointer to Ethernet device structure.
991 * Pointer to array to save the port identifiers.
993 * The length of the array.
995 * Current port to peer port direction.
996 * positive - current used as Tx to get all peer Rx ports.
997 * zero - current used as Rx to get all peer Tx ports.
1000 * 0 or positive value on success, actual number of peer ports.
1001 * a negative errno value otherwise and rte_errno is set.
1004 mlx5_hairpin_get_peer_ports(struct rte_eth_dev *dev, uint16_t *peer_ports,
1005 size_t len, uint32_t direction)
1007 struct mlx5_priv *priv = dev->data->dev_private;
1008 struct mlx5_txq_ctrl *txq_ctrl;
1011 uint32_t bits[(RTE_MAX_ETHPORTS + 31) / 32] = {0};
1015 for (i = 0; i < priv->txqs_n; i++) {
1016 txq_ctrl = mlx5_txq_get(dev, i);
1019 if (!txq_ctrl->is_hairpin) {
1020 mlx5_txq_release(dev, i);
1023 pp = txq_ctrl->hairpin_conf.peers[0].port;
1024 if (pp >= RTE_MAX_ETHPORTS) {
1026 mlx5_txq_release(dev, i);
1027 DRV_LOG(ERR, "port %hu queue %u peer port "
1029 priv->dev_data->port_id, i, pp);
1032 bits[pp / 32] |= 1 << (pp % 32);
1033 mlx5_txq_release(dev, i);
1036 for (i = 0; i < priv->rxqs_n; i++) {
1037 struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, i);
1038 struct mlx5_rxq_ctrl *rxq_ctrl;
1042 rxq_ctrl = rxq->ctrl;
1043 if (!rxq_ctrl->is_hairpin)
1045 pp = rxq->hairpin_conf.peers[0].port;
1046 if (pp >= RTE_MAX_ETHPORTS) {
1048 DRV_LOG(ERR, "port %hu queue %u peer port "
1050 priv->dev_data->port_id, i, pp);
1053 bits[pp / 32] |= 1 << (pp % 32);
1056 for (i = 0; i < RTE_MAX_ETHPORTS; i++) {
1057 if (bits[i / 32] & (1 << (i % 32))) {
1058 if ((size_t)ret >= len) {
1062 peer_ports[ret++] = i;
1069 * DPDK callback to start the device.
1071 * Simulate device start by attaching all configured flows.
1074 * Pointer to Ethernet device structure.
1077 * 0 on success, a negative errno value otherwise and rte_errno is set.
1080 mlx5_dev_start(struct rte_eth_dev *dev)
1082 struct mlx5_priv *priv = dev->data->dev_private;
1086 DRV_LOG(DEBUG, "port %u starting device", dev->data->port_id);
1087 fine_inline = rte_mbuf_dynflag_lookup
1088 (RTE_PMD_MLX5_FINE_GRANULARITY_INLINE, NULL);
1089 if (fine_inline >= 0)
1090 rte_net_mlx5_dynf_inline_mask = 1UL << fine_inline;
1092 rte_net_mlx5_dynf_inline_mask = 0;
1093 if (dev->data->nb_rx_queues > 0) {
1094 ret = mlx5_dev_configure_rss_reta(dev);
1096 DRV_LOG(ERR, "port %u reta config failed: %s",
1097 dev->data->port_id, strerror(rte_errno));
1101 ret = mlx5_txpp_start(dev);
1103 DRV_LOG(ERR, "port %u Tx packet pacing init failed: %s",
1104 dev->data->port_id, strerror(rte_errno));
1107 if (mlx5_devx_obj_ops_en(priv->sh) &&
1108 priv->obj_ops.lb_dummy_queue_create) {
1109 ret = priv->obj_ops.lb_dummy_queue_create(dev);
1113 ret = mlx5_txq_start(dev);
1115 DRV_LOG(ERR, "port %u Tx queue allocation failed: %s",
1116 dev->data->port_id, strerror(rte_errno));
1119 if (priv->config.std_delay_drop || priv->config.hp_delay_drop) {
1120 if (!priv->sh->dev_cap.vf && !priv->sh->dev_cap.sf &&
1121 !priv->representor) {
1122 ret = mlx5_get_flag_dropless_rq(dev);
1125 "port %u cannot query dropless flag",
1126 dev->data->port_id);
1129 "port %u dropless_rq OFF, no rearming",
1130 dev->data->port_id);
1133 "port %u doesn't support dropless_rq flag",
1134 dev->data->port_id);
1137 ret = mlx5_rxq_start(dev);
1139 DRV_LOG(ERR, "port %u Rx queue allocation failed: %s",
1140 dev->data->port_id, strerror(rte_errno));
1144 * Such step will be skipped if there is no hairpin TX queue configured
1145 * with RX peer queue from the same device.
1147 ret = mlx5_hairpin_auto_bind(dev);
1149 DRV_LOG(ERR, "port %u hairpin auto binding failed: %s",
1150 dev->data->port_id, strerror(rte_errno));
1153 /* Set started flag here for the following steps like control flow. */
1154 dev->data->dev_started = 1;
1155 ret = mlx5_rx_intr_vec_enable(dev);
1157 DRV_LOG(ERR, "port %u Rx interrupt vector creation failed",
1158 dev->data->port_id);
1161 mlx5_os_stats_init(dev);
1163 * Attach indirection table objects detached on port stop.
1164 * They may be needed to create RSS in non-isolated mode.
1166 ret = mlx5_action_handle_attach(dev);
1169 "port %u failed to attach indirect actions: %s",
1170 dev->data->port_id, rte_strerror(rte_errno));
1173 ret = mlx5_traffic_enable(dev);
1175 DRV_LOG(ERR, "port %u failed to set defaults flows",
1176 dev->data->port_id);
1179 /* Set a mask and offset of dynamic metadata flows into Rx queues. */
1180 mlx5_flow_rxq_dynf_metadata_set(dev);
1181 /* Set flags and context to convert Rx timestamps. */
1182 mlx5_rxq_timestamp_set(dev);
1183 /* Set a mask and offset of scheduling on timestamp into Tx queues. */
1184 mlx5_txq_dynf_timestamp_set(dev);
1186 * In non-cached mode, it only needs to start the default mreg copy
1187 * action and no flow created by application exists anymore.
1188 * But it is worth wrapping the interface for further usage.
1190 ret = mlx5_flow_start_default(dev);
1192 DRV_LOG(DEBUG, "port %u failed to start default actions: %s",
1193 dev->data->port_id, strerror(rte_errno));
1196 if (mlx5_dev_ctx_shared_mempool_subscribe(dev) != 0) {
1197 DRV_LOG(ERR, "port %u failed to subscribe for mempool life cycle: %s",
1198 dev->data->port_id, rte_strerror(rte_errno));
1202 dev->tx_pkt_burst = mlx5_select_tx_function(dev);
1203 dev->rx_pkt_burst = mlx5_select_rx_function(dev);
1204 /* Enable datapath on secondary process. */
1205 mlx5_mp_os_req_start_rxtx(dev);
1206 if (rte_intr_fd_get(priv->sh->intr_handle) >= 0) {
1207 priv->sh->port[priv->dev_port - 1].ih_port_id =
1208 (uint32_t)dev->data->port_id;
1210 DRV_LOG(INFO, "port %u starts without RMV interrupts.",
1211 dev->data->port_id);
1212 dev->data->dev_conf.intr_conf.rmv = 0;
1214 if (rte_intr_fd_get(priv->sh->intr_handle_nl) >= 0) {
1215 priv->sh->port[priv->dev_port - 1].nl_ih_port_id =
1216 (uint32_t)dev->data->port_id;
1218 DRV_LOG(INFO, "port %u starts without LSC interrupts.",
1219 dev->data->port_id);
1220 dev->data->dev_conf.intr_conf.lsc = 0;
1222 if (rte_intr_fd_get(priv->sh->intr_handle_devx) >= 0)
1223 priv->sh->port[priv->dev_port - 1].devx_ih_port_id =
1224 (uint32_t)dev->data->port_id;
1227 ret = rte_errno; /* Save rte_errno before cleanup. */
1229 dev->data->dev_started = 0;
1230 mlx5_flow_stop_default(dev);
1231 mlx5_traffic_disable(dev);
1234 if (priv->obj_ops.lb_dummy_queue_release)
1235 priv->obj_ops.lb_dummy_queue_release(dev);
1236 mlx5_txpp_stop(dev); /* Stop last. */
1237 rte_errno = ret; /* Restore rte_errno. */
1242 * DPDK callback to stop the device.
1244 * Simulate device stop by detaching all configured flows.
1247 * Pointer to Ethernet device structure.
1250 mlx5_dev_stop(struct rte_eth_dev *dev)
1252 struct mlx5_priv *priv = dev->data->dev_private;
1254 dev->data->dev_started = 0;
1255 /* Prevent crashes when queues are still in use. */
1256 dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1257 dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1259 /* Disable datapath on secondary process. */
1260 mlx5_mp_os_req_stop_rxtx(dev);
1261 rte_delay_us_sleep(1000 * priv->rxqs_n);
1262 DRV_LOG(DEBUG, "port %u stopping device", dev->data->port_id);
1263 mlx5_flow_stop_default(dev);
1264 /* Control flows for default traffic can be removed firstly. */
1265 mlx5_traffic_disable(dev);
1266 /* All RX queue flags will be cleared in the flush interface. */
1267 mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_GEN, true);
1268 mlx5_flow_meter_rxq_flush(dev);
1269 mlx5_action_handle_detach(dev);
1270 mlx5_rx_intr_vec_disable(dev);
1271 priv->sh->port[priv->dev_port - 1].ih_port_id = RTE_MAX_ETHPORTS;
1272 priv->sh->port[priv->dev_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
1273 priv->sh->port[priv->dev_port - 1].nl_ih_port_id = RTE_MAX_ETHPORTS;
1276 if (priv->obj_ops.lb_dummy_queue_release)
1277 priv->obj_ops.lb_dummy_queue_release(dev);
1278 mlx5_txpp_stop(dev);
1284 * Enable traffic flows configured by control plane
1287 * Pointer to Ethernet device structure.
1290 * 0 on success, a negative errno value otherwise and rte_errno is set.
1293 mlx5_traffic_enable(struct rte_eth_dev *dev)
1295 struct mlx5_priv *priv = dev->data->dev_private;
1296 struct rte_flow_item_eth bcast = {
1297 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1299 struct rte_flow_item_eth ipv6_multi_spec = {
1300 .dst.addr_bytes = "\x33\x33\x00\x00\x00\x00",
1302 struct rte_flow_item_eth ipv6_multi_mask = {
1303 .dst.addr_bytes = "\xff\xff\x00\x00\x00\x00",
1305 struct rte_flow_item_eth unicast = {
1306 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
1308 struct rte_flow_item_eth unicast_mask = {
1309 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1311 const unsigned int vlan_filter_n = priv->vlan_filter_n;
1312 const struct rte_ether_addr cmp = {
1313 .addr_bytes = "\x00\x00\x00\x00\x00\x00",
1320 * Hairpin txq default flow should be created no matter if it is
1321 * isolation mode. Or else all the packets to be sent will be sent
1322 * out directly without the TX flow actions, e.g. encapsulation.
1324 for (i = 0; i != priv->txqs_n; ++i) {
1325 struct mlx5_txq_ctrl *txq_ctrl = mlx5_txq_get(dev, i);
1328 /* Only Tx implicit mode requires the default Tx flow. */
1329 if (txq_ctrl->is_hairpin &&
1330 txq_ctrl->hairpin_conf.tx_explicit == 0 &&
1331 txq_ctrl->hairpin_conf.peers[0].port ==
1332 priv->dev_data->port_id) {
1333 ret = mlx5_ctrl_flow_source_queue(dev, i);
1335 mlx5_txq_release(dev, i);
1339 if (priv->sh->config.dv_esw_en) {
1340 if (mlx5_flow_create_devx_sq_miss_flow(dev, i) == 0) {
1342 "Port %u Tx queue %u SQ create representor devx default miss rule failed.",
1343 dev->data->port_id, i);
1347 mlx5_txq_release(dev, i);
1349 if (priv->sh->config.dv_esw_en) {
1350 if (mlx5_flow_create_esw_table_zero_flow(dev))
1351 priv->fdb_def_rule = 1;
1353 DRV_LOG(INFO, "port %u FDB default rule cannot be"
1354 " configured - only Eswitch group 0 flows are"
1355 " supported.", dev->data->port_id);
1357 if (!priv->sh->config.lacp_by_user && priv->pf_bond >= 0) {
1358 ret = mlx5_flow_lacp_miss(dev);
1360 DRV_LOG(INFO, "port %u LACP rule cannot be created - "
1361 "forward LACP to kernel.", dev->data->port_id);
1363 DRV_LOG(INFO, "LACP traffic will be missed in port %u."
1364 , dev->data->port_id);
1368 if (dev->data->promiscuous) {
1369 struct rte_flow_item_eth promisc = {
1370 .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
1371 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
1375 ret = mlx5_ctrl_flow(dev, &promisc, &promisc);
1379 if (dev->data->all_multicast) {
1380 struct rte_flow_item_eth multicast = {
1381 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1382 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
1386 ret = mlx5_ctrl_flow(dev, &multicast, &multicast);
1390 /* Add broadcast/multicast flows. */
1391 for (i = 0; i != vlan_filter_n; ++i) {
1392 uint16_t vlan = priv->vlan_filter[i];
1394 struct rte_flow_item_vlan vlan_spec = {
1395 .tci = rte_cpu_to_be_16(vlan),
1397 struct rte_flow_item_vlan vlan_mask =
1398 rte_flow_item_vlan_mask;
1400 ret = mlx5_ctrl_flow_vlan(dev, &bcast, &bcast,
1401 &vlan_spec, &vlan_mask);
1404 ret = mlx5_ctrl_flow_vlan(dev, &ipv6_multi_spec,
1406 &vlan_spec, &vlan_mask);
1410 if (!vlan_filter_n) {
1411 ret = mlx5_ctrl_flow(dev, &bcast, &bcast);
1414 ret = mlx5_ctrl_flow(dev, &ipv6_multi_spec,
1417 /* Do not fail on IPv6 broadcast creation failure. */
1419 "IPv6 broadcast is not supported");
1424 /* Add MAC address flows. */
1425 for (i = 0; i != MLX5_MAX_MAC_ADDRESSES; ++i) {
1426 struct rte_ether_addr *mac = &dev->data->mac_addrs[i];
1428 if (!memcmp(mac, &cmp, sizeof(*mac)))
1430 memcpy(&unicast.dst.addr_bytes,
1432 RTE_ETHER_ADDR_LEN);
1433 for (j = 0; j != vlan_filter_n; ++j) {
1434 uint16_t vlan = priv->vlan_filter[j];
1436 struct rte_flow_item_vlan vlan_spec = {
1437 .tci = rte_cpu_to_be_16(vlan),
1439 struct rte_flow_item_vlan vlan_mask =
1440 rte_flow_item_vlan_mask;
1442 ret = mlx5_ctrl_flow_vlan(dev, &unicast,
1449 if (!vlan_filter_n) {
1450 ret = mlx5_ctrl_flow(dev, &unicast, &unicast_mask);
1457 ret = rte_errno; /* Save rte_errno before cleanup. */
1458 mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_CTL, false);
1459 rte_errno = ret; /* Restore rte_errno. */
1465 * Disable traffic flows configured by control plane
1468 * Pointer to Ethernet device private data.
1471 mlx5_traffic_disable(struct rte_eth_dev *dev)
1473 mlx5_flow_list_flush(dev, MLX5_FLOW_TYPE_CTL, false);
1477 * Restart traffic flows configured by control plane
1480 * Pointer to Ethernet device private data.
1483 * 0 on success, a negative errno value otherwise and rte_errno is set.
1486 mlx5_traffic_restart(struct rte_eth_dev *dev)
1488 if (dev->data->dev_started) {
1489 mlx5_traffic_disable(dev);
1490 return mlx5_traffic_enable(dev);