1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
9 #include <rte_ethdev_driver.h>
10 #include <rte_interrupts.h>
11 #include <rte_alarm.h>
14 #include "mlx5_rxtx.h"
15 #include "mlx5_utils.h"
16 #include "rte_pmd_mlx5.h"
19 * Stop traffic on Tx queues.
22 * Pointer to Ethernet device structure.
25 mlx5_txq_stop(struct rte_eth_dev *dev)
27 struct mlx5_priv *priv = dev->data->dev_private;
30 for (i = 0; i != priv->txqs_n; ++i)
31 mlx5_txq_release(dev, i);
35 * Start traffic on Tx queues.
38 * Pointer to Ethernet device structure.
41 * 0 on success, a negative errno value otherwise and rte_errno is set.
44 mlx5_txq_start(struct rte_eth_dev *dev)
46 struct mlx5_priv *priv = dev->data->dev_private;
50 for (i = 0; i != priv->txqs_n; ++i) {
51 struct mlx5_txq_ctrl *txq_ctrl = mlx5_txq_get(dev, i);
55 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
56 txq_ctrl->obj = mlx5_txq_obj_new
57 (dev, i, MLX5_TXQ_OBJ_TYPE_DEVX_HAIRPIN);
59 txq_alloc_elts(txq_ctrl);
60 txq_ctrl->obj = mlx5_txq_obj_new
61 (dev, i, MLX5_TXQ_OBJ_TYPE_IBV);
70 ret = rte_errno; /* Save rte_errno before cleanup. */
72 mlx5_txq_release(dev, i);
74 rte_errno = ret; /* Restore rte_errno. */
79 * Stop traffic on Rx queues.
82 * Pointer to Ethernet device structure.
85 mlx5_rxq_stop(struct rte_eth_dev *dev)
87 struct mlx5_priv *priv = dev->data->dev_private;
90 for (i = 0; i != priv->rxqs_n; ++i)
91 mlx5_rxq_release(dev, i);
95 * Start traffic on Rx queues.
98 * Pointer to Ethernet device structure.
101 * 0 on success, a negative errno value otherwise and rte_errno is set.
104 mlx5_rxq_start(struct rte_eth_dev *dev)
106 struct mlx5_priv *priv = dev->data->dev_private;
109 enum mlx5_rxq_obj_type obj_type = MLX5_RXQ_OBJ_TYPE_IBV;
110 struct mlx5_rxq_data *rxq = NULL;
112 for (i = 0; i < priv->rxqs_n; ++i) {
113 rxq = (*priv->rxqs)[i];
115 if (rxq && rxq->lro) {
116 obj_type = MLX5_RXQ_OBJ_TYPE_DEVX_RQ;
120 /* Allocate/reuse/resize mempool for Multi-Packet RQ. */
121 if (mlx5_mprq_alloc_mp(dev)) {
122 /* Should not release Rx queues but return immediately. */
125 for (i = 0; i != priv->rxqs_n; ++i) {
126 struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_get(dev, i);
127 struct rte_mempool *mp;
131 if (rxq_ctrl->type == MLX5_RXQ_TYPE_HAIRPIN) {
132 rxq_ctrl->obj = mlx5_rxq_obj_new
133 (dev, i, MLX5_RXQ_OBJ_TYPE_DEVX_HAIRPIN);
138 /* Pre-register Rx mempool. */
139 mp = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
140 rxq_ctrl->rxq.mprq_mp : rxq_ctrl->rxq.mp;
142 "port %u Rx queue %u registering"
143 " mp %s having %u chunks",
144 dev->data->port_id, rxq_ctrl->rxq.idx,
145 mp->name, mp->nb_mem_chunks);
146 mlx5_mr_update_mp(dev, &rxq_ctrl->rxq.mr_ctrl, mp);
147 ret = rxq_alloc_elts(rxq_ctrl);
150 rxq_ctrl->obj = mlx5_rxq_obj_new(dev, i, obj_type);
153 if (obj_type == MLX5_RXQ_OBJ_TYPE_IBV)
154 rxq_ctrl->wqn = rxq_ctrl->obj->wq->wq_num;
155 else if (obj_type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ)
156 rxq_ctrl->wqn = rxq_ctrl->obj->rq->id;
160 ret = rte_errno; /* Save rte_errno before cleanup. */
162 mlx5_rxq_release(dev, i);
164 rte_errno = ret; /* Restore rte_errno. */
169 * Binds Tx queues to Rx queues for hairpin.
171 * Binds Tx queues to the target Rx queues.
174 * Pointer to Ethernet device structure.
177 * 0 on success, a negative errno value otherwise and rte_errno is set.
180 mlx5_hairpin_bind(struct rte_eth_dev *dev)
182 struct mlx5_priv *priv = dev->data->dev_private;
183 struct mlx5_devx_modify_sq_attr sq_attr = { 0 };
184 struct mlx5_devx_modify_rq_attr rq_attr = { 0 };
185 struct mlx5_txq_ctrl *txq_ctrl;
186 struct mlx5_rxq_ctrl *rxq_ctrl;
187 struct mlx5_devx_obj *sq;
188 struct mlx5_devx_obj *rq;
192 for (i = 0; i != priv->txqs_n; ++i) {
193 txq_ctrl = mlx5_txq_get(dev, i);
196 if (txq_ctrl->type != MLX5_TXQ_TYPE_HAIRPIN) {
197 mlx5_txq_release(dev, i);
200 if (!txq_ctrl->obj) {
202 DRV_LOG(ERR, "port %u no txq object found: %d",
203 dev->data->port_id, i);
204 mlx5_txq_release(dev, i);
207 sq = txq_ctrl->obj->sq;
208 rxq_ctrl = mlx5_rxq_get(dev,
209 txq_ctrl->hairpin_conf.peers[0].queue);
211 mlx5_txq_release(dev, i);
213 DRV_LOG(ERR, "port %u no rxq object found: %d",
215 txq_ctrl->hairpin_conf.peers[0].queue);
218 if (rxq_ctrl->type != MLX5_RXQ_TYPE_HAIRPIN ||
219 rxq_ctrl->hairpin_conf.peers[0].queue != i) {
221 DRV_LOG(ERR, "port %u Tx queue %d can't be binded to "
222 "Rx queue %d", dev->data->port_id,
223 i, txq_ctrl->hairpin_conf.peers[0].queue);
226 rq = rxq_ctrl->obj->rq;
229 DRV_LOG(ERR, "port %u hairpin no matching rxq: %d",
231 txq_ctrl->hairpin_conf.peers[0].queue);
234 sq_attr.state = MLX5_SQC_STATE_RDY;
235 sq_attr.sq_state = MLX5_SQC_STATE_RST;
236 sq_attr.hairpin_peer_rq = rq->id;
237 sq_attr.hairpin_peer_vhca = priv->config.hca_attr.vhca_id;
238 ret = mlx5_devx_cmd_modify_sq(sq, &sq_attr);
241 rq_attr.state = MLX5_SQC_STATE_RDY;
242 rq_attr.rq_state = MLX5_SQC_STATE_RST;
243 rq_attr.hairpin_peer_sq = sq->id;
244 rq_attr.hairpin_peer_vhca = priv->config.hca_attr.vhca_id;
245 ret = mlx5_devx_cmd_modify_rq(rq, &rq_attr);
248 mlx5_txq_release(dev, i);
249 mlx5_rxq_release(dev, txq_ctrl->hairpin_conf.peers[0].queue);
253 mlx5_txq_release(dev, i);
254 mlx5_rxq_release(dev, txq_ctrl->hairpin_conf.peers[0].queue);
259 * DPDK callback to start the device.
261 * Simulate device start by attaching all configured flows.
264 * Pointer to Ethernet device structure.
267 * 0 on success, a negative errno value otherwise and rte_errno is set.
270 mlx5_dev_start(struct rte_eth_dev *dev)
272 struct mlx5_priv *priv = dev->data->dev_private;
276 DRV_LOG(DEBUG, "port %u starting device", dev->data->port_id);
277 fine_inline = rte_mbuf_dynflag_lookup
278 (RTE_PMD_MLX5_FINE_GRANULARITY_INLINE, NULL);
280 rte_net_mlx5_dynf_inline_mask = 1UL << fine_inline;
282 rte_net_mlx5_dynf_inline_mask = 0;
283 if (dev->data->nb_rx_queues > 0) {
284 ret = mlx5_dev_configure_rss_reta(dev);
286 DRV_LOG(ERR, "port %u reta config failed: %s",
287 dev->data->port_id, strerror(rte_errno));
291 ret = mlx5_txq_start(dev);
293 DRV_LOG(ERR, "port %u Tx queue allocation failed: %s",
294 dev->data->port_id, strerror(rte_errno));
297 ret = mlx5_rxq_start(dev);
299 DRV_LOG(ERR, "port %u Rx queue allocation failed: %s",
300 dev->data->port_id, strerror(rte_errno));
304 ret = mlx5_hairpin_bind(dev);
306 DRV_LOG(ERR, "port %u hairpin binding failed: %s",
307 dev->data->port_id, strerror(rte_errno));
311 dev->data->dev_started = 1;
312 ret = mlx5_rx_intr_vec_enable(dev);
314 DRV_LOG(ERR, "port %u Rx interrupt vector creation failed",
318 mlx5_stats_init(dev);
319 ret = mlx5_traffic_enable(dev);
321 DRV_LOG(DEBUG, "port %u failed to set defaults flows",
325 ret = mlx5_flow_start(dev, &priv->flows);
327 DRV_LOG(DEBUG, "port %u failed to set flows",
332 dev->tx_pkt_burst = mlx5_select_tx_function(dev);
333 dev->rx_pkt_burst = mlx5_select_rx_function(dev);
334 /* Enable datapath on secondary process. */
335 mlx5_mp_req_start_rxtx(dev);
336 mlx5_dev_interrupt_handler_install(dev);
339 ret = rte_errno; /* Save rte_errno before cleanup. */
341 dev->data->dev_started = 0;
342 mlx5_flow_stop(dev, &priv->flows);
343 mlx5_traffic_disable(dev);
346 rte_errno = ret; /* Restore rte_errno. */
351 * DPDK callback to stop the device.
353 * Simulate device stop by detaching all configured flows.
356 * Pointer to Ethernet device structure.
359 mlx5_dev_stop(struct rte_eth_dev *dev)
361 struct mlx5_priv *priv = dev->data->dev_private;
363 dev->data->dev_started = 0;
364 /* Prevent crashes when queues are still in use. */
365 dev->rx_pkt_burst = removed_rx_burst;
366 dev->tx_pkt_burst = removed_tx_burst;
368 /* Disable datapath on secondary process. */
369 mlx5_mp_req_stop_rxtx(dev);
370 usleep(1000 * priv->rxqs_n);
371 DRV_LOG(DEBUG, "port %u stopping device", dev->data->port_id);
372 mlx5_flow_stop(dev, &priv->flows);
373 mlx5_traffic_disable(dev);
374 mlx5_rx_intr_vec_disable(dev);
375 mlx5_dev_interrupt_handler_uninstall(dev);
381 * Enable traffic flows configured by control plane
384 * Pointer to Ethernet device private data.
386 * Pointer to Ethernet device structure.
389 * 0 on success, a negative errno value otherwise and rte_errno is set.
392 mlx5_traffic_enable(struct rte_eth_dev *dev)
394 struct mlx5_priv *priv = dev->data->dev_private;
395 struct rte_flow_item_eth bcast = {
396 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
398 struct rte_flow_item_eth ipv6_multi_spec = {
399 .dst.addr_bytes = "\x33\x33\x00\x00\x00\x00",
401 struct rte_flow_item_eth ipv6_multi_mask = {
402 .dst.addr_bytes = "\xff\xff\x00\x00\x00\x00",
404 struct rte_flow_item_eth unicast = {
405 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
407 struct rte_flow_item_eth unicast_mask = {
408 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
410 const unsigned int vlan_filter_n = priv->vlan_filter_n;
411 const struct rte_ether_addr cmp = {
412 .addr_bytes = "\x00\x00\x00\x00\x00\x00",
419 * Hairpin txq default flow should be created no matter if it is
420 * isolation mode. Or else all the packets to be sent will be sent
421 * out directly without the TX flow actions, e.g. encapsulation.
423 for (i = 0; i != priv->txqs_n; ++i) {
424 struct mlx5_txq_ctrl *txq_ctrl = mlx5_txq_get(dev, i);
427 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
428 ret = mlx5_ctrl_flow_source_queue(dev, i);
430 mlx5_txq_release(dev, i);
434 mlx5_txq_release(dev, i);
436 if (priv->config.dv_esw_en && !priv->config.vf) {
437 if (mlx5_flow_create_esw_table_zero_flow(dev))
438 priv->fdb_def_rule = 1;
440 DRV_LOG(INFO, "port %u FDB default rule cannot be"
441 " configured - only Eswitch group 0 flows are"
442 " supported.", dev->data->port_id);
446 if (dev->data->promiscuous) {
447 struct rte_flow_item_eth promisc = {
448 .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
449 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
453 ret = mlx5_ctrl_flow(dev, &promisc, &promisc);
457 if (dev->data->all_multicast) {
458 struct rte_flow_item_eth multicast = {
459 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
460 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
464 ret = mlx5_ctrl_flow(dev, &multicast, &multicast);
468 /* Add broadcast/multicast flows. */
469 for (i = 0; i != vlan_filter_n; ++i) {
470 uint16_t vlan = priv->vlan_filter[i];
472 struct rte_flow_item_vlan vlan_spec = {
473 .tci = rte_cpu_to_be_16(vlan),
475 struct rte_flow_item_vlan vlan_mask =
476 rte_flow_item_vlan_mask;
478 ret = mlx5_ctrl_flow_vlan(dev, &bcast, &bcast,
479 &vlan_spec, &vlan_mask);
482 ret = mlx5_ctrl_flow_vlan(dev, &ipv6_multi_spec,
484 &vlan_spec, &vlan_mask);
488 if (!vlan_filter_n) {
489 ret = mlx5_ctrl_flow(dev, &bcast, &bcast);
492 ret = mlx5_ctrl_flow(dev, &ipv6_multi_spec,
498 /* Add MAC address flows. */
499 for (i = 0; i != MLX5_MAX_MAC_ADDRESSES; ++i) {
500 struct rte_ether_addr *mac = &dev->data->mac_addrs[i];
502 if (!memcmp(mac, &cmp, sizeof(*mac)))
504 memcpy(&unicast.dst.addr_bytes,
507 for (j = 0; j != vlan_filter_n; ++j) {
508 uint16_t vlan = priv->vlan_filter[j];
510 struct rte_flow_item_vlan vlan_spec = {
511 .tci = rte_cpu_to_be_16(vlan),
513 struct rte_flow_item_vlan vlan_mask =
514 rte_flow_item_vlan_mask;
516 ret = mlx5_ctrl_flow_vlan(dev, &unicast,
523 if (!vlan_filter_n) {
524 ret = mlx5_ctrl_flow(dev, &unicast, &unicast_mask);
531 ret = rte_errno; /* Save rte_errno before cleanup. */
532 mlx5_flow_list_flush(dev, &priv->ctrl_flows);
533 rte_errno = ret; /* Restore rte_errno. */
539 * Disable traffic flows configured by control plane
542 * Pointer to Ethernet device private data.
545 mlx5_traffic_disable(struct rte_eth_dev *dev)
547 struct mlx5_priv *priv = dev->data->dev_private;
549 mlx5_flow_list_flush(dev, &priv->ctrl_flows);
553 * Restart traffic flows configured by control plane
556 * Pointer to Ethernet device private data.
559 * 0 on success, a negative errno value otherwise and rte_errno is set.
562 mlx5_traffic_restart(struct rte_eth_dev *dev)
564 if (dev->data->dev_started) {
565 mlx5_traffic_disable(dev);
566 return mlx5_traffic_enable(dev);