xref: /dpdk/drivers/net/mlx5/linux/mlx5_verbs.c (revision db4e81351fb85ff623bd0438d1b5a8fb55fe9fee)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2020 Mellanox Technologies, Ltd
3  */
4 
5 #include <stddef.h>
6 #include <errno.h>
7 #include <stdbool.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <unistd.h>
11 #include <inttypes.h>
12 #include <sys/queue.h>
13 
14 #include "mlx5_autoconf.h"
15 
16 #include <rte_mbuf.h>
17 #include <rte_malloc.h>
18 #include <rte_ethdev_driver.h>
19 #include <rte_common.h>
20 
21 #include <mlx5_glue.h>
22 #include <mlx5_common.h>
23 #include <mlx5_common_mr.h>
24 #include <mlx5_rxtx.h>
25 #include <mlx5_verbs.h>
26 #include <mlx5_utils.h>
27 #include <mlx5_malloc.h>
28 
29 /**
30  * Register mr. Given protection domain pointer, pointer to addr and length
31  * register the memory region.
32  *
33  * @param[in] pd
34  *   Pointer to protection domain context.
35  * @param[in] addr
36  *   Pointer to memory start address.
37  * @param[in] length
38  *   Length of the memory to register.
39  * @param[out] pmd_mr
40  *   pmd_mr struct set with lkey, address, length and pointer to mr object
41  *
42  * @return
43  *   0 on successful registration, -1 otherwise
44  */
45 static int
46 mlx5_reg_mr(void *pd, void *addr, size_t length,
47 		 struct mlx5_pmd_mr *pmd_mr)
48 {
49 	return mlx5_common_verbs_reg_mr(pd, addr, length, pmd_mr);
50 }
51 
52 /**
53  * Deregister mr. Given the mlx5 pmd MR - deregister the MR
54  *
55  * @param[in] pmd_mr
56  *   pmd_mr struct set with lkey, address, length and pointer to mr object
57  *
58  */
59 static void
60 mlx5_dereg_mr(struct mlx5_pmd_mr *pmd_mr)
61 {
62 	mlx5_common_verbs_dereg_mr(pmd_mr);
63 }
64 
65 /* verbs operations. */
66 const struct mlx5_verbs_ops mlx5_verbs_ops = {
67 	.reg_mr = mlx5_reg_mr,
68 	.dereg_mr = mlx5_dereg_mr,
69 };
70 
71 /**
72  * Modify Rx WQ vlan stripping offload
73  *
74  * @param rxq_obj
75  *   Rx queue object.
76  *
77  * @return 0 on success, non-0 otherwise
78  */
79 static int
80 mlx5_rxq_obj_modify_wq_vlan_strip(struct mlx5_rxq_obj *rxq_obj, int on)
81 {
82 	uint16_t vlan_offloads =
83 		(on ? IBV_WQ_FLAGS_CVLAN_STRIPPING : 0) |
84 		0;
85 	struct ibv_wq_attr mod;
86 	mod = (struct ibv_wq_attr){
87 		.attr_mask = IBV_WQ_ATTR_FLAGS,
88 		.flags_mask = IBV_WQ_FLAGS_CVLAN_STRIPPING,
89 		.flags = vlan_offloads,
90 	};
91 
92 	return mlx5_glue->modify_wq(rxq_obj->wq, &mod);
93 }
94 
95 /**
96  * Modifies the attributes for the specified WQ.
97  *
98  * @param rxq_obj
99  *   Verbs Rx queue object.
100  *
101  * @return
102  *   0 on success, a negative errno value otherwise and rte_errno is set.
103  */
104 static int
105 mlx5_ibv_modify_wq(struct mlx5_rxq_obj *rxq_obj, bool is_start)
106 {
107 	struct ibv_wq_attr mod = {
108 		.attr_mask = IBV_WQ_ATTR_STATE,
109 		.wq_state = is_start ? IBV_WQS_RDY : IBV_WQS_RESET,
110 	};
111 
112 	return mlx5_glue->modify_wq(rxq_obj->wq, &mod);
113 }
114 
115 /**
116  * Create a CQ Verbs object.
117  *
118  * @param dev
119  *   Pointer to Ethernet device.
120  * @param idx
121  *   Queue index in DPDK Rx queue array.
122  *
123  * @return
124  *   The Verbs CQ object initialized, NULL otherwise and rte_errno is set.
125  */
126 static struct ibv_cq *
127 mlx5_rxq_ibv_cq_create(struct rte_eth_dev *dev, uint16_t idx)
128 {
129 	struct mlx5_priv *priv = dev->data->dev_private;
130 	struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
131 	struct mlx5_rxq_ctrl *rxq_ctrl =
132 		container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
133 	struct mlx5_rxq_obj *rxq_obj = rxq_ctrl->obj;
134 	unsigned int cqe_n = mlx5_rxq_cqe_num(rxq_data);
135 	struct {
136 		struct ibv_cq_init_attr_ex ibv;
137 		struct mlx5dv_cq_init_attr mlx5;
138 	} cq_attr;
139 
140 	cq_attr.ibv = (struct ibv_cq_init_attr_ex){
141 		.cqe = cqe_n,
142 		.channel = rxq_obj->ibv_channel,
143 		.comp_mask = 0,
144 	};
145 	cq_attr.mlx5 = (struct mlx5dv_cq_init_attr){
146 		.comp_mask = 0,
147 	};
148 	if (priv->config.cqe_comp && !rxq_data->hw_timestamp) {
149 		cq_attr.mlx5.comp_mask |=
150 				MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
151 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
152 		cq_attr.mlx5.cqe_comp_res_format =
153 				mlx5_rxq_mprq_enabled(rxq_data) ?
154 				MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX :
155 				MLX5DV_CQE_RES_FORMAT_HASH;
156 #else
157 		cq_attr.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
158 #endif
159 		/*
160 		 * For vectorized Rx, it must not be doubled in order to
161 		 * make cq_ci and rq_ci aligned.
162 		 */
163 		if (mlx5_rxq_check_vec_support(rxq_data) < 0)
164 			cq_attr.ibv.cqe *= 2;
165 	} else if (priv->config.cqe_comp && rxq_data->hw_timestamp) {
166 		DRV_LOG(DEBUG,
167 			"Port %u Rx CQE compression is disabled for HW"
168 			" timestamp.",
169 			dev->data->port_id);
170 	}
171 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
172 	if (priv->config.cqe_pad) {
173 		cq_attr.mlx5.comp_mask |= MLX5DV_CQ_INIT_ATTR_MASK_FLAGS;
174 		cq_attr.mlx5.flags |= MLX5DV_CQ_INIT_ATTR_FLAGS_CQE_PAD;
175 	}
176 #endif
177 	return mlx5_glue->cq_ex_to_cq(mlx5_glue->dv_create_cq(priv->sh->ctx,
178 							      &cq_attr.ibv,
179 							      &cq_attr.mlx5));
180 }
181 
182 /**
183  * Create a WQ Verbs object.
184  *
185  * @param dev
186  *   Pointer to Ethernet device.
187  * @param idx
188  *   Queue index in DPDK Rx queue array.
189  *
190  * @return
191  *   The Verbs WQ object initialized, NULL otherwise and rte_errno is set.
192  */
193 static struct ibv_wq *
194 mlx5_rxq_ibv_wq_create(struct rte_eth_dev *dev, uint16_t idx)
195 {
196 	struct mlx5_priv *priv = dev->data->dev_private;
197 	struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
198 	struct mlx5_rxq_ctrl *rxq_ctrl =
199 		container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
200 	struct mlx5_rxq_obj *rxq_obj = rxq_ctrl->obj;
201 	unsigned int wqe_n = 1 << rxq_data->elts_n;
202 	struct {
203 		struct ibv_wq_init_attr ibv;
204 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
205 		struct mlx5dv_wq_init_attr mlx5;
206 #endif
207 	} wq_attr;
208 
209 	wq_attr.ibv = (struct ibv_wq_init_attr){
210 		.wq_context = NULL, /* Could be useful in the future. */
211 		.wq_type = IBV_WQT_RQ,
212 		/* Max number of outstanding WRs. */
213 		.max_wr = wqe_n >> rxq_data->sges_n,
214 		/* Max number of scatter/gather elements in a WR. */
215 		.max_sge = 1 << rxq_data->sges_n,
216 		.pd = priv->sh->pd,
217 		.cq = rxq_obj->ibv_cq,
218 		.comp_mask = IBV_WQ_FLAGS_CVLAN_STRIPPING | 0,
219 		.create_flags = (rxq_data->vlan_strip ?
220 				 IBV_WQ_FLAGS_CVLAN_STRIPPING : 0),
221 	};
222 	/* By default, FCS (CRC) is stripped by hardware. */
223 	if (rxq_data->crc_present) {
224 		wq_attr.ibv.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
225 		wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
226 	}
227 	if (priv->config.hw_padding) {
228 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
229 		wq_attr.ibv.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
230 		wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
231 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
232 		wq_attr.ibv.create_flags |= IBV_WQ_FLAGS_PCI_WRITE_END_PADDING;
233 		wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
234 #endif
235 	}
236 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
237 	wq_attr.mlx5 = (struct mlx5dv_wq_init_attr){
238 		.comp_mask = 0,
239 	};
240 	if (mlx5_rxq_mprq_enabled(rxq_data)) {
241 		struct mlx5dv_striding_rq_init_attr *mprq_attr =
242 						&wq_attr.mlx5.striding_rq_attrs;
243 
244 		wq_attr.mlx5.comp_mask |= MLX5DV_WQ_INIT_ATTR_MASK_STRIDING_RQ;
245 		*mprq_attr = (struct mlx5dv_striding_rq_init_attr){
246 			.single_stride_log_num_of_bytes = rxq_data->strd_sz_n,
247 			.single_wqe_log_num_of_strides = rxq_data->strd_num_n,
248 			.two_byte_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT,
249 		};
250 	}
251 	rxq_obj->wq = mlx5_glue->dv_create_wq(priv->sh->ctx, &wq_attr.ibv,
252 					      &wq_attr.mlx5);
253 #else
254 	rxq_obj->wq = mlx5_glue->create_wq(priv->sh->ctx, &wq_attr.ibv);
255 #endif
256 	if (rxq_obj->wq) {
257 		/*
258 		 * Make sure number of WRs*SGEs match expectations since a queue
259 		 * cannot allocate more than "desc" buffers.
260 		 */
261 		if (wq_attr.ibv.max_wr != (wqe_n >> rxq_data->sges_n) ||
262 		    wq_attr.ibv.max_sge != (1u << rxq_data->sges_n)) {
263 			DRV_LOG(ERR,
264 				"Port %u Rx queue %u requested %u*%u but got"
265 				" %u*%u WRs*SGEs.",
266 				dev->data->port_id, idx,
267 				wqe_n >> rxq_data->sges_n,
268 				(1 << rxq_data->sges_n),
269 				wq_attr.ibv.max_wr, wq_attr.ibv.max_sge);
270 			claim_zero(mlx5_glue->destroy_wq(rxq_obj->wq));
271 			rxq_obj->wq = NULL;
272 			rte_errno = EINVAL;
273 		}
274 	}
275 	return rxq_obj->wq;
276 }
277 
278 /**
279  * Create the Rx queue Verbs object.
280  *
281  * @param dev
282  *   Pointer to Ethernet device.
283  * @param idx
284  *   Queue index in DPDK Rx queue array.
285  *
286  * @return
287  *   0 on success, a negative errno value otherwise and rte_errno is set.
288  */
289 static int
290 mlx5_rxq_ibv_obj_new(struct rte_eth_dev *dev, uint16_t idx)
291 {
292 	struct mlx5_priv *priv = dev->data->dev_private;
293 	struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
294 	struct mlx5_rxq_ctrl *rxq_ctrl =
295 		container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
296 	struct mlx5_rxq_obj *tmpl = rxq_ctrl->obj;
297 	struct mlx5dv_cq cq_info;
298 	struct mlx5dv_rwq rwq;
299 	int ret = 0;
300 	struct mlx5dv_obj obj;
301 
302 	MLX5_ASSERT(rxq_data);
303 	MLX5_ASSERT(tmpl);
304 	priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
305 	priv->verbs_alloc_ctx.obj = rxq_ctrl;
306 	tmpl->type = MLX5_RXQ_OBJ_TYPE_IBV;
307 	tmpl->rxq_ctrl = rxq_ctrl;
308 	if (rxq_ctrl->irq) {
309 		tmpl->ibv_channel =
310 				mlx5_glue->create_comp_channel(priv->sh->ctx);
311 		if (!tmpl->ibv_channel) {
312 			DRV_LOG(ERR, "Port %u: comp channel creation failure.",
313 				dev->data->port_id);
314 			rte_errno = ENOMEM;
315 			goto error;
316 		}
317 		tmpl->fd = ((struct ibv_comp_channel *)(tmpl->ibv_channel))->fd;
318 	}
319 	/* Create CQ using Verbs API. */
320 	tmpl->ibv_cq = mlx5_rxq_ibv_cq_create(dev, idx);
321 	if (!tmpl->ibv_cq) {
322 		DRV_LOG(ERR, "Port %u Rx queue %u CQ creation failure.",
323 			dev->data->port_id, idx);
324 		rte_errno = ENOMEM;
325 		goto error;
326 	}
327 	obj.cq.in = tmpl->ibv_cq;
328 	obj.cq.out = &cq_info;
329 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ);
330 	if (ret) {
331 		rte_errno = ret;
332 		goto error;
333 	}
334 	if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
335 		DRV_LOG(ERR,
336 			"Port %u wrong MLX5_CQE_SIZE environment "
337 			"variable value: it should be set to %u.",
338 			dev->data->port_id, RTE_CACHE_LINE_SIZE);
339 		rte_errno = EINVAL;
340 		goto error;
341 	}
342 	/* Fill the rings. */
343 	rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
344 	rxq_data->cq_db = cq_info.dbrec;
345 	rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
346 	rxq_data->cq_uar = cq_info.cq_uar;
347 	rxq_data->cqn = cq_info.cqn;
348 	/* Create WQ (RQ) using Verbs API. */
349 	tmpl->wq = mlx5_rxq_ibv_wq_create(dev, idx);
350 	if (!tmpl->wq) {
351 		DRV_LOG(ERR, "Port %u Rx queue %u WQ creation failure.",
352 			dev->data->port_id, idx);
353 		rte_errno = ENOMEM;
354 		goto error;
355 	}
356 	/* Change queue state to ready. */
357 	ret = mlx5_ibv_modify_wq(tmpl, true);
358 	if (ret) {
359 		DRV_LOG(ERR,
360 			"Port %u Rx queue %u WQ state to IBV_WQS_RDY failed.",
361 			dev->data->port_id, idx);
362 		rte_errno = ret;
363 		goto error;
364 	}
365 	obj.rwq.in = tmpl->wq;
366 	obj.rwq.out = &rwq;
367 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_RWQ);
368 	if (ret) {
369 		rte_errno = ret;
370 		goto error;
371 	}
372 	rxq_data->wqes = rwq.buf;
373 	rxq_data->rq_db = rwq.dbrec;
374 	rxq_data->cq_arm_sn = 0;
375 	mlx5_rxq_initialize(rxq_data);
376 	rxq_data->cq_ci = 0;
377 	priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
378 	dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
379 	rxq_ctrl->wqn = ((struct ibv_wq *)(tmpl->wq))->wq_num;
380 	return 0;
381 error:
382 	ret = rte_errno; /* Save rte_errno before cleanup. */
383 	if (tmpl->wq)
384 		claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
385 	if (tmpl->ibv_cq)
386 		claim_zero(mlx5_glue->destroy_cq(tmpl->ibv_cq));
387 	if (tmpl->ibv_channel)
388 		claim_zero(mlx5_glue->destroy_comp_channel(tmpl->ibv_channel));
389 	rte_errno = ret; /* Restore rte_errno. */
390 	priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
391 	return -rte_errno;
392 }
393 
394 /**
395  * Release an Rx verbs queue object.
396  *
397  * @param rxq_obj
398  *   Verbs Rx queue object.
399  */
400 static void
401 mlx5_rxq_ibv_obj_release(struct mlx5_rxq_obj *rxq_obj)
402 {
403 	MLX5_ASSERT(rxq_obj);
404 	MLX5_ASSERT(rxq_obj->wq);
405 	MLX5_ASSERT(rxq_obj->ibv_cq);
406 	claim_zero(mlx5_glue->destroy_wq(rxq_obj->wq));
407 	claim_zero(mlx5_glue->destroy_cq(rxq_obj->ibv_cq));
408 	if (rxq_obj->ibv_channel)
409 		claim_zero(mlx5_glue->destroy_comp_channel
410 							(rxq_obj->ibv_channel));
411 }
412 
413 /**
414  * Get event for an Rx verbs queue object.
415  *
416  * @param rxq_obj
417  *   Verbs Rx queue object.
418  *
419  * @return
420  *   0 on success, a negative errno value otherwise and rte_errno is set.
421  */
422 static int
423 mlx5_rx_ibv_get_event(struct mlx5_rxq_obj *rxq_obj)
424 {
425 	struct ibv_cq *ev_cq;
426 	void *ev_ctx;
427 	int ret = mlx5_glue->get_cq_event(rxq_obj->ibv_channel,
428 					  &ev_cq, &ev_ctx);
429 
430 	if (ret < 0 || ev_cq != rxq_obj->ibv_cq)
431 		goto exit;
432 	mlx5_glue->ack_cq_events(rxq_obj->ibv_cq, 1);
433 	return 0;
434 exit:
435 	if (ret < 0)
436 		rte_errno = errno;
437 	else
438 		rte_errno = EINVAL;
439 	return -rte_errno;
440 }
441 
442 /**
443  * Creates a receive work queue as a filed of indirection table.
444  *
445  * @param dev
446  *   Pointer to Ethernet device.
447  * @param log_n
448  *   Log of number of queues in the array.
449  * @param ind_tbl
450  *   Verbs indirection table object.
451  *
452  * @return
453  *   0 on success, a negative errno value otherwise and rte_errno is set.
454  */
455 static int
456 mlx5_ibv_ind_table_new(struct rte_eth_dev *dev, const unsigned int log_n,
457 		       struct mlx5_ind_table_obj *ind_tbl)
458 {
459 	struct mlx5_priv *priv = dev->data->dev_private;
460 	struct ibv_wq *wq[1 << log_n];
461 	unsigned int i, j;
462 
463 	MLX5_ASSERT(ind_tbl);
464 	for (i = 0; i != ind_tbl->queues_n; ++i) {
465 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[ind_tbl->queues[i]];
466 		struct mlx5_rxq_ctrl *rxq_ctrl =
467 				container_of(rxq, struct mlx5_rxq_ctrl, rxq);
468 
469 		wq[i] = rxq_ctrl->obj->wq;
470 	}
471 	MLX5_ASSERT(i > 0);
472 	/* Finalise indirection table. */
473 	for (j = 0; i != (unsigned int)(1 << log_n); ++j, ++i)
474 		wq[i] = wq[j];
475 	ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table(priv->sh->ctx,
476 					&(struct ibv_rwq_ind_table_init_attr){
477 						.log_ind_tbl_size = log_n,
478 						.ind_tbl = wq,
479 						.comp_mask = 0,
480 					});
481 	if (!ind_tbl->ind_table) {
482 		rte_errno = errno;
483 		return -rte_errno;
484 	}
485 	return 0;
486 }
487 
488 /**
489  * Destroys the specified Indirection Table.
490  *
491  * @param ind_table
492  *   Indirection table to release.
493  */
494 static void
495 mlx5_ibv_ind_table_destroy(struct mlx5_ind_table_obj *ind_tbl)
496 {
497 	claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl->ind_table));
498 }
499 
500 /**
501  * Create an Rx Hash queue.
502  *
503  * @param dev
504  *   Pointer to Ethernet device.
505  * @param hrxq
506  *   Pointer to Rx Hash queue.
507  * @param tunnel
508  *   Tunnel type.
509  *
510  * @return
511  *   0 on success, a negative errno value otherwise and rte_errno is set.
512  */
513 static int
514 mlx5_ibv_hrxq_new(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq,
515 		  int tunnel __rte_unused)
516 {
517 	struct mlx5_priv *priv = dev->data->dev_private;
518 	struct ibv_qp *qp = NULL;
519 	struct mlx5_ind_table_obj *ind_tbl = hrxq->ind_table;
520 	const uint8_t *rss_key = hrxq->rss_key;
521 	uint64_t hash_fields = hrxq->hash_fields;
522 	int err;
523 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
524 	struct mlx5dv_qp_init_attr qp_init_attr;
525 
526 	memset(&qp_init_attr, 0, sizeof(qp_init_attr));
527 	if (tunnel) {
528 		qp_init_attr.comp_mask =
529 				       MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
530 		qp_init_attr.create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
531 	}
532 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
533 	if (dev->data->dev_conf.lpbk_mode) {
534 		/* Allow packet sent from NIC loop back w/o source MAC check. */
535 		qp_init_attr.comp_mask |=
536 				MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
537 		qp_init_attr.create_flags |=
538 				MLX5DV_QP_CREATE_TIR_ALLOW_SELF_LOOPBACK_UC;
539 	}
540 #endif
541 	qp = mlx5_glue->dv_create_qp
542 			(priv->sh->ctx,
543 			 &(struct ibv_qp_init_attr_ex){
544 				.qp_type = IBV_QPT_RAW_PACKET,
545 				.comp_mask =
546 					IBV_QP_INIT_ATTR_PD |
547 					IBV_QP_INIT_ATTR_IND_TABLE |
548 					IBV_QP_INIT_ATTR_RX_HASH,
549 				.rx_hash_conf = (struct ibv_rx_hash_conf){
550 					.rx_hash_function =
551 						IBV_RX_HASH_FUNC_TOEPLITZ,
552 					.rx_hash_key_len = hrxq->rss_key_len,
553 					.rx_hash_key =
554 						(void *)(uintptr_t)rss_key,
555 					.rx_hash_fields_mask = hash_fields,
556 				},
557 				.rwq_ind_tbl = ind_tbl->ind_table,
558 				.pd = priv->sh->pd,
559 			  },
560 			  &qp_init_attr);
561 #else
562 	qp = mlx5_glue->create_qp_ex
563 			(priv->sh->ctx,
564 			 &(struct ibv_qp_init_attr_ex){
565 				.qp_type = IBV_QPT_RAW_PACKET,
566 				.comp_mask =
567 					IBV_QP_INIT_ATTR_PD |
568 					IBV_QP_INIT_ATTR_IND_TABLE |
569 					IBV_QP_INIT_ATTR_RX_HASH,
570 				.rx_hash_conf = (struct ibv_rx_hash_conf){
571 					.rx_hash_function =
572 						IBV_RX_HASH_FUNC_TOEPLITZ,
573 					.rx_hash_key_len = hrxq->rss_key_len,
574 					.rx_hash_key =
575 						(void *)(uintptr_t)rss_key,
576 					.rx_hash_fields_mask = hash_fields,
577 				},
578 				.rwq_ind_tbl = ind_tbl->ind_table,
579 				.pd = priv->sh->pd,
580 			 });
581 #endif
582 	if (!qp) {
583 		rte_errno = errno;
584 		goto error;
585 	}
586 	hrxq->qp = qp;
587 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
588 	hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
589 	if (!hrxq->action) {
590 		rte_errno = errno;
591 		goto error;
592 	}
593 #endif
594 	return 0;
595 error:
596 	err = rte_errno; /* Save rte_errno before cleanup. */
597 	if (qp)
598 		claim_zero(mlx5_glue->destroy_qp(qp));
599 	rte_errno = err; /* Restore rte_errno. */
600 	return -rte_errno;
601 }
602 
603 /**
604  * Destroy a Verbs queue pair.
605  *
606  * @param hrxq
607  *   Hash Rx queue to release its qp.
608  */
609 static void
610 mlx5_ibv_qp_destroy(struct mlx5_hrxq *hrxq)
611 {
612 	claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
613 }
614 
615 /**
616  * Release a drop Rx queue Verbs object.
617  *
618  * @param dev
619  *   Pointer to Ethernet device.
620  */
621 static void
622 mlx5_rxq_ibv_obj_drop_release(struct rte_eth_dev *dev)
623 {
624 	struct mlx5_priv *priv = dev->data->dev_private;
625 	struct mlx5_rxq_obj *rxq = priv->drop_queue.rxq;
626 
627 	if (rxq->wq)
628 		claim_zero(mlx5_glue->destroy_wq(rxq->wq));
629 	if (rxq->ibv_cq)
630 		claim_zero(mlx5_glue->destroy_cq(rxq->ibv_cq));
631 	mlx5_free(rxq);
632 	priv->drop_queue.rxq = NULL;
633 }
634 
635 /**
636  * Create a drop Rx queue Verbs object.
637  *
638  * @param dev
639  *   Pointer to Ethernet device.
640  *
641  * @return
642  *   0 on success, a negative errno value otherwise and rte_errno is set.
643  */
644 static int
645 mlx5_rxq_ibv_obj_drop_create(struct rte_eth_dev *dev)
646 {
647 	struct mlx5_priv *priv = dev->data->dev_private;
648 	struct ibv_context *ctx = priv->sh->ctx;
649 	struct mlx5_rxq_obj *rxq = priv->drop_queue.rxq;
650 
651 	if (rxq)
652 		return 0;
653 	rxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rxq), 0, SOCKET_ID_ANY);
654 	if (!rxq) {
655 		DEBUG("Port %u cannot allocate drop Rx queue memory.",
656 		      dev->data->port_id);
657 		rte_errno = ENOMEM;
658 		return -rte_errno;
659 	}
660 	priv->drop_queue.rxq = rxq;
661 	rxq->ibv_cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
662 	if (!rxq->ibv_cq) {
663 		DEBUG("Port %u cannot allocate CQ for drop queue.",
664 		      dev->data->port_id);
665 		rte_errno = errno;
666 		goto error;
667 	}
668 	rxq->wq = mlx5_glue->create_wq(ctx, &(struct ibv_wq_init_attr){
669 						    .wq_type = IBV_WQT_RQ,
670 						    .max_wr = 1,
671 						    .max_sge = 1,
672 						    .pd = priv->sh->pd,
673 						    .cq = rxq->ibv_cq,
674 					      });
675 	if (!rxq->wq) {
676 		DEBUG("Port %u cannot allocate WQ for drop queue.",
677 		      dev->data->port_id);
678 		rte_errno = errno;
679 		goto error;
680 	}
681 	priv->drop_queue.rxq = rxq;
682 	return 0;
683 error:
684 	mlx5_rxq_ibv_obj_drop_release(dev);
685 	return -rte_errno;
686 }
687 
688 /**
689  * Create a Verbs drop action for Rx Hash queue.
690  *
691  * @param dev
692  *   Pointer to Ethernet device.
693  *
694  * @return
695  *   0 on success, a negative errno value otherwise and rte_errno is set.
696  */
697 static int
698 mlx5_ibv_drop_action_create(struct rte_eth_dev *dev)
699 {
700 	struct mlx5_priv *priv = dev->data->dev_private;
701 	struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
702 	struct ibv_rwq_ind_table *ind_tbl = NULL;
703 	struct mlx5_rxq_obj *rxq;
704 	int ret;
705 
706 	MLX5_ASSERT(hrxq && hrxq->ind_table);
707 	ret = mlx5_rxq_ibv_obj_drop_create(dev);
708 	if (ret < 0)
709 		goto error;
710 	rxq = priv->drop_queue.rxq;
711 	ind_tbl = mlx5_glue->create_rwq_ind_table
712 				(priv->sh->ctx,
713 				 &(struct ibv_rwq_ind_table_init_attr){
714 					.log_ind_tbl_size = 0,
715 					.ind_tbl = (struct ibv_wq **)&rxq->wq,
716 					.comp_mask = 0,
717 				 });
718 	if (!ind_tbl) {
719 		DEBUG("Port %u cannot allocate indirection table for drop"
720 		      " queue.", dev->data->port_id);
721 		rte_errno = errno;
722 		goto error;
723 	}
724 	hrxq->qp = mlx5_glue->create_qp_ex(priv->sh->ctx,
725 		 &(struct ibv_qp_init_attr_ex){
726 			.qp_type = IBV_QPT_RAW_PACKET,
727 			.comp_mask = IBV_QP_INIT_ATTR_PD |
728 				     IBV_QP_INIT_ATTR_IND_TABLE |
729 				     IBV_QP_INIT_ATTR_RX_HASH,
730 			.rx_hash_conf = (struct ibv_rx_hash_conf){
731 				.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
732 				.rx_hash_key_len = MLX5_RSS_HASH_KEY_LEN,
733 				.rx_hash_key = rss_hash_default_key,
734 				.rx_hash_fields_mask = 0,
735 				},
736 			.rwq_ind_tbl = ind_tbl,
737 			.pd = priv->sh->pd
738 		 });
739 	if (!hrxq->qp) {
740 		DEBUG("Port %u cannot allocate QP for drop queue.",
741 		      dev->data->port_id);
742 		rte_errno = errno;
743 		goto error;
744 	}
745 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
746 	hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
747 	if (!hrxq->action) {
748 		rte_errno = errno;
749 		goto error;
750 	}
751 #endif
752 	hrxq->ind_table->ind_table = ind_tbl;
753 	return 0;
754 error:
755 	if (hrxq->qp)
756 		claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
757 	if (ind_tbl)
758 		claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl));
759 	if (priv->drop_queue.rxq)
760 		mlx5_rxq_ibv_obj_drop_release(dev);
761 	return -rte_errno;
762 }
763 
764 /**
765  * Release a drop hash Rx queue.
766  *
767  * @param dev
768  *   Pointer to Ethernet device.
769  */
770 static void
771 mlx5_ibv_drop_action_destroy(struct rte_eth_dev *dev)
772 {
773 	struct mlx5_priv *priv = dev->data->dev_private;
774 	struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
775 	struct ibv_rwq_ind_table *ind_tbl = hrxq->ind_table->ind_table;
776 
777 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
778 	claim_zero(mlx5_glue->destroy_flow_action(hrxq->action));
779 #endif
780 	claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
781 	claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl));
782 	mlx5_rxq_ibv_obj_drop_release(dev);
783 }
784 
785 struct mlx5_obj_ops ibv_obj_ops = {
786 	.rxq_obj_modify_vlan_strip = mlx5_rxq_obj_modify_wq_vlan_strip,
787 	.rxq_obj_new = mlx5_rxq_ibv_obj_new,
788 	.rxq_event_get = mlx5_rx_ibv_get_event,
789 	.rxq_obj_modify = mlx5_ibv_modify_wq,
790 	.rxq_obj_release = mlx5_rxq_ibv_obj_release,
791 	.ind_table_new = mlx5_ibv_ind_table_new,
792 	.ind_table_destroy = mlx5_ibv_ind_table_destroy,
793 	.hrxq_new = mlx5_ibv_hrxq_new,
794 	.hrxq_destroy = mlx5_ibv_qp_destroy,
795 	.drop_action_create = mlx5_ibv_drop_action_create,
796 	.drop_action_destroy = mlx5_ibv_drop_action_destroy,
797 };
798