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)
275 DRV_LOG(DEBUG, "port %u starting device", dev->data->port_id);
276 fine_inline = rte_mbuf_dynflag_lookup
277 (RTE_PMD_MLX5_FINE_GRANULARITY_INLINE, NULL);
279 rte_net_mlx5_dynf_inline_mask = 1UL << fine_inline;
281 rte_net_mlx5_dynf_inline_mask = 0;
282 if (dev->data->nb_rx_queues > 0) {
283 ret = mlx5_dev_configure_rss_reta(dev);
285 DRV_LOG(ERR, "port %u reta config failed: %s",
286 dev->data->port_id, strerror(rte_errno));
290 ret = mlx5_txq_start(dev);
292 DRV_LOG(ERR, "port %u Tx queue allocation failed: %s",
293 dev->data->port_id, strerror(rte_errno));
296 ret = mlx5_rxq_start(dev);
298 DRV_LOG(ERR, "port %u Rx queue allocation failed: %s",
299 dev->data->port_id, strerror(rte_errno));
303 ret = mlx5_hairpin_bind(dev);
305 DRV_LOG(ERR, "port %u hairpin binding failed: %s",
306 dev->data->port_id, strerror(rte_errno));
310 /* Set started flag here for the following steps like control flow. */
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(ERR, "port %u failed to set defaults flows",
326 * In non-cached mode, it only needs to start the default mreg copy
327 * action and no flow created by application exists anymore.
328 * But it is worth wrapping the interface for further usage.
330 ret = mlx5_flow_start_default(dev);
332 DRV_LOG(DEBUG, "port %u failed to start default actions: %s",
333 dev->data->port_id, strerror(rte_errno));
337 dev->tx_pkt_burst = mlx5_select_tx_function(dev);
338 dev->rx_pkt_burst = mlx5_select_rx_function(dev);
339 /* Enable datapath on secondary process. */
340 mlx5_mp_req_start_rxtx(dev);
341 mlx5_dev_interrupt_handler_install(dev);
344 ret = rte_errno; /* Save rte_errno before cleanup. */
346 dev->data->dev_started = 0;
347 mlx5_flow_stop_default(dev);
348 mlx5_traffic_disable(dev);
351 rte_errno = ret; /* Restore rte_errno. */
356 * DPDK callback to stop the device.
358 * Simulate device stop by detaching all configured flows.
361 * Pointer to Ethernet device structure.
364 mlx5_dev_stop(struct rte_eth_dev *dev)
366 struct mlx5_priv *priv = dev->data->dev_private;
368 dev->data->dev_started = 0;
369 /* Prevent crashes when queues are still in use. */
370 dev->rx_pkt_burst = removed_rx_burst;
371 dev->tx_pkt_burst = removed_tx_burst;
373 /* Disable datapath on secondary process. */
374 mlx5_mp_req_stop_rxtx(dev);
375 usleep(1000 * priv->rxqs_n);
376 DRV_LOG(DEBUG, "port %u stopping device", dev->data->port_id);
377 mlx5_flow_stop_default(dev);
378 /* Control flows for default traffic can be removed firstly. */
379 mlx5_traffic_disable(dev);
380 /* All RX queue flags will be cleared in the flush interface. */
381 mlx5_flow_list_flush(dev, &priv->flows, true);
382 mlx5_rx_intr_vec_disable(dev);
383 mlx5_dev_interrupt_handler_uninstall(dev);
389 * Enable traffic flows configured by control plane
392 * Pointer to Ethernet device private data.
394 * Pointer to Ethernet device structure.
397 * 0 on success, a negative errno value otherwise and rte_errno is set.
400 mlx5_traffic_enable(struct rte_eth_dev *dev)
402 struct mlx5_priv *priv = dev->data->dev_private;
403 struct rte_flow_item_eth bcast = {
404 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
406 struct rte_flow_item_eth ipv6_multi_spec = {
407 .dst.addr_bytes = "\x33\x33\x00\x00\x00\x00",
409 struct rte_flow_item_eth ipv6_multi_mask = {
410 .dst.addr_bytes = "\xff\xff\x00\x00\x00\x00",
412 struct rte_flow_item_eth unicast = {
413 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
415 struct rte_flow_item_eth unicast_mask = {
416 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
418 const unsigned int vlan_filter_n = priv->vlan_filter_n;
419 const struct rte_ether_addr cmp = {
420 .addr_bytes = "\x00\x00\x00\x00\x00\x00",
427 * Hairpin txq default flow should be created no matter if it is
428 * isolation mode. Or else all the packets to be sent will be sent
429 * out directly without the TX flow actions, e.g. encapsulation.
431 for (i = 0; i != priv->txqs_n; ++i) {
432 struct mlx5_txq_ctrl *txq_ctrl = mlx5_txq_get(dev, i);
435 if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
436 ret = mlx5_ctrl_flow_source_queue(dev, i);
438 mlx5_txq_release(dev, i);
442 mlx5_txq_release(dev, i);
444 if (priv->config.dv_esw_en && !priv->config.vf) {
445 if (mlx5_flow_create_esw_table_zero_flow(dev))
446 priv->fdb_def_rule = 1;
448 DRV_LOG(INFO, "port %u FDB default rule cannot be"
449 " configured - only Eswitch group 0 flows are"
450 " supported.", dev->data->port_id);
454 if (dev->data->promiscuous) {
455 struct rte_flow_item_eth promisc = {
456 .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
457 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
461 ret = mlx5_ctrl_flow(dev, &promisc, &promisc);
465 if (dev->data->all_multicast) {
466 struct rte_flow_item_eth multicast = {
467 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
468 .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
472 ret = mlx5_ctrl_flow(dev, &multicast, &multicast);
476 /* Add broadcast/multicast flows. */
477 for (i = 0; i != vlan_filter_n; ++i) {
478 uint16_t vlan = priv->vlan_filter[i];
480 struct rte_flow_item_vlan vlan_spec = {
481 .tci = rte_cpu_to_be_16(vlan),
483 struct rte_flow_item_vlan vlan_mask =
484 rte_flow_item_vlan_mask;
486 ret = mlx5_ctrl_flow_vlan(dev, &bcast, &bcast,
487 &vlan_spec, &vlan_mask);
490 ret = mlx5_ctrl_flow_vlan(dev, &ipv6_multi_spec,
492 &vlan_spec, &vlan_mask);
496 if (!vlan_filter_n) {
497 ret = mlx5_ctrl_flow(dev, &bcast, &bcast);
500 ret = mlx5_ctrl_flow(dev, &ipv6_multi_spec,
506 /* Add MAC address flows. */
507 for (i = 0; i != MLX5_MAX_MAC_ADDRESSES; ++i) {
508 struct rte_ether_addr *mac = &dev->data->mac_addrs[i];
510 if (!memcmp(mac, &cmp, sizeof(*mac)))
512 memcpy(&unicast.dst.addr_bytes,
515 for (j = 0; j != vlan_filter_n; ++j) {
516 uint16_t vlan = priv->vlan_filter[j];
518 struct rte_flow_item_vlan vlan_spec = {
519 .tci = rte_cpu_to_be_16(vlan),
521 struct rte_flow_item_vlan vlan_mask =
522 rte_flow_item_vlan_mask;
524 ret = mlx5_ctrl_flow_vlan(dev, &unicast,
531 if (!vlan_filter_n) {
532 ret = mlx5_ctrl_flow(dev, &unicast, &unicast_mask);
539 ret = rte_errno; /* Save rte_errno before cleanup. */
540 mlx5_flow_list_flush(dev, &priv->ctrl_flows, false);
541 rte_errno = ret; /* Restore rte_errno. */
547 * Disable traffic flows configured by control plane
550 * Pointer to Ethernet device private data.
553 mlx5_traffic_disable(struct rte_eth_dev *dev)
555 struct mlx5_priv *priv = dev->data->dev_private;
557 mlx5_flow_list_flush(dev, &priv->ctrl_flows, false);
561 * Restart traffic flows configured by control plane
564 * Pointer to Ethernet device private data.
567 * 0 on success, a negative errno value otherwise and rte_errno is set.
570 mlx5_traffic_restart(struct rte_eth_dev *dev)
572 if (dev->data->dev_started) {
573 mlx5_traffic_disable(dev);
574 return mlx5_traffic_enable(dev);