xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision a2999c7bfe395708a6a68ebb6f3199af7d25cffe)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2020 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
16 #include <fcntl.h>
17 
18 #include <rte_malloc.h>
19 #include <rte_ethdev_driver.h>
20 #include <rte_ethdev_pci.h>
21 #include <rte_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_common.h>
24 #include <rte_kvargs.h>
25 #include <rte_rwlock.h>
26 #include <rte_spinlock.h>
27 #include <rte_string_fns.h>
28 #include <rte_alarm.h>
29 #include <rte_eal_paging.h>
30 
31 #include <mlx5_glue.h>
32 #include <mlx5_devx_cmds.h>
33 #include <mlx5_common.h>
34 #include <mlx5_common_mp.h>
35 #include <mlx5_common_mr.h>
36 #include <mlx5_malloc.h>
37 
38 #include "mlx5_defs.h"
39 #include "mlx5.h"
40 #include "mlx5_common_os.h"
41 #include "mlx5_utils.h"
42 #include "mlx5_rxtx.h"
43 #include "mlx5_autoconf.h"
44 #include "mlx5_mr.h"
45 #include "mlx5_flow.h"
46 #include "rte_pmd_mlx5.h"
47 #include "mlx5_verbs.h"
48 #include "mlx5_nl.h"
49 #include "mlx5_devx.h"
50 
51 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /**
71  * Set the completion channel file descriptor interrupt as non-blocking.
72  *
73  * @param[in] rxq_obj
74  *   Pointer to RQ channel object, which includes the channel fd
75  *
76  * @param[out] fd
77  *   The file descriptor (representing the intetrrupt) used in this channel.
78  *
79  * @return
80  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
81  */
82 int
83 mlx5_os_set_nonblock_channel_fd(int fd)
84 {
85 	int flags;
86 
87 	flags = fcntl(fd, F_GETFL);
88 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
89 }
90 
91 /**
92  * Get mlx5 device attributes. The glue function query_device_ex() is called
93  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
94  * device attributes from the glue out parameter.
95  *
96  * @param dev
97  *   Pointer to ibv context.
98  *
99  * @param device_attr
100  *   Pointer to mlx5 device attributes.
101  *
102  * @return
103  *   0 on success, non zero error number otherwise
104  */
105 int
106 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
107 {
108 	int err;
109 	struct ibv_device_attr_ex attr_ex;
110 	memset(device_attr, 0, sizeof(*device_attr));
111 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
112 	if (err)
113 		return err;
114 
115 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
116 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
117 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
118 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
119 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
120 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
121 	device_attr->max_rwq_indirection_table_size =
122 		attr_ex.rss_caps.max_rwq_indirection_table_size;
123 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
124 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
125 
126 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
127 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
128 	if (err)
129 		return err;
130 
131 	device_attr->flags = dv_attr.flags;
132 	device_attr->comp_mask = dv_attr.comp_mask;
133 #ifdef HAVE_IBV_MLX5_MOD_SWP
134 	device_attr->sw_parsing_offloads =
135 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
136 #endif
137 	device_attr->min_single_stride_log_num_of_bytes =
138 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
139 	device_attr->max_single_stride_log_num_of_bytes =
140 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
141 	device_attr->min_single_wqe_log_num_of_strides =
142 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
143 	device_attr->max_single_wqe_log_num_of_strides =
144 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
145 	device_attr->stride_supported_qpts =
146 		dv_attr.striding_rq_caps.supported_qpts;
147 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
148 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
149 #endif
150 
151 	return err;
152 }
153 
154 /**
155  * Verbs callback to allocate a memory. This function should allocate the space
156  * according to the size provided residing inside a huge page.
157  * Please note that all allocation must respect the alignment from libmlx5
158  * (i.e. currently rte_mem_page_size()).
159  *
160  * @param[in] size
161  *   The size in bytes of the memory to allocate.
162  * @param[in] data
163  *   A pointer to the callback data.
164  *
165  * @return
166  *   Allocated buffer, NULL otherwise and rte_errno is set.
167  */
168 static void *
169 mlx5_alloc_verbs_buf(size_t size, void *data)
170 {
171 	struct mlx5_priv *priv = data;
172 	void *ret;
173 	unsigned int socket = SOCKET_ID_ANY;
174 	size_t alignment = rte_mem_page_size();
175 	if (alignment == (size_t)-1) {
176 		DRV_LOG(ERR, "Failed to get mem page size");
177 		rte_errno = ENOMEM;
178 		return NULL;
179 	}
180 
181 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
182 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
183 
184 		socket = ctrl->socket;
185 	} else if (priv->verbs_alloc_ctx.type ==
186 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
187 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
188 
189 		socket = ctrl->socket;
190 	}
191 	MLX5_ASSERT(data != NULL);
192 	ret = mlx5_malloc(0, size, alignment, socket);
193 	if (!ret && size)
194 		rte_errno = ENOMEM;
195 	return ret;
196 }
197 
198 /**
199  * Verbs callback to free a memory.
200  *
201  * @param[in] ptr
202  *   A pointer to the memory to free.
203  * @param[in] data
204  *   A pointer to the callback data.
205  */
206 static void
207 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
208 {
209 	MLX5_ASSERT(data != NULL);
210 	mlx5_free(ptr);
211 }
212 
213 /**
214  * Initialize DR related data within private structure.
215  * Routine checks the reference counter and does actual
216  * resources creation/initialization only if counter is zero.
217  *
218  * @param[in] priv
219  *   Pointer to the private device data structure.
220  *
221  * @return
222  *   Zero on success, positive error code otherwise.
223  */
224 static int
225 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
226 {
227 	struct mlx5_dev_ctx_shared *sh = priv->sh;
228 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
229 	int err;
230 
231 	MLX5_ASSERT(sh && sh->refcnt);
232 	if (sh->refcnt > 1)
233 		return 0;
234 	err = mlx5_alloc_table_hash_list(priv);
235 	if (err)
236 		goto error;
237 	/* The resources below are only valid with DV support. */
238 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
239 	/* Init port id action cache list. */
240 	snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
241 	mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
242 			     flow_dv_port_id_create_cb,
243 			     flow_dv_port_id_match_cb,
244 			     flow_dv_port_id_remove_cb);
245 	/* Init push vlan action cache list. */
246 	snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
247 	mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
248 			     flow_dv_push_vlan_create_cb,
249 			     flow_dv_push_vlan_match_cb,
250 			     flow_dv_push_vlan_remove_cb);
251 	/* Init sample action cache list. */
252 	snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
253 	mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
254 			     flow_dv_sample_create_cb,
255 			     flow_dv_sample_match_cb,
256 			     flow_dv_sample_remove_cb);
257 	/* Init dest array action cache list. */
258 	snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
259 	mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
260 			     flow_dv_dest_array_create_cb,
261 			     flow_dv_dest_array_match_cb,
262 			     flow_dv_dest_array_remove_cb);
263 	/* Create tags hash list table. */
264 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
265 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
266 					  MLX5_HLIST_WRITE_MOST,
267 					  flow_dv_tag_create_cb, NULL,
268 					  flow_dv_tag_remove_cb);
269 	if (!sh->tag_table) {
270 		DRV_LOG(ERR, "tags with hash creation failed.");
271 		err = ENOMEM;
272 		goto error;
273 	}
274 	sh->tag_table->ctx = sh;
275 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
276 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
277 					    0, MLX5_HLIST_WRITE_MOST |
278 					    MLX5_HLIST_DIRECT_KEY,
279 					    flow_dv_modify_create_cb,
280 					    flow_dv_modify_match_cb,
281 					    flow_dv_modify_remove_cb);
282 	if (!sh->modify_cmds) {
283 		DRV_LOG(ERR, "hdr modify hash creation failed");
284 		err = ENOMEM;
285 		goto error;
286 	}
287 	sh->modify_cmds->ctx = sh;
288 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
289 	sh->encaps_decaps = mlx5_hlist_create(s,
290 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
291 					      0, MLX5_HLIST_DIRECT_KEY |
292 					      MLX5_HLIST_WRITE_MOST,
293 					      flow_dv_encap_decap_create_cb,
294 					      flow_dv_encap_decap_match_cb,
295 					      flow_dv_encap_decap_remove_cb);
296 	if (!sh->encaps_decaps) {
297 		DRV_LOG(ERR, "encap decap hash creation failed");
298 		err = ENOMEM;
299 		goto error;
300 	}
301 	sh->encaps_decaps->ctx = sh;
302 #endif
303 #ifdef HAVE_MLX5DV_DR
304 	void *domain;
305 
306 	/* Reference counter is zero, we should initialize structures. */
307 	domain = mlx5_glue->dr_create_domain(sh->ctx,
308 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
309 	if (!domain) {
310 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
311 		err = errno;
312 		goto error;
313 	}
314 	sh->rx_domain = domain;
315 	domain = mlx5_glue->dr_create_domain(sh->ctx,
316 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
317 	if (!domain) {
318 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
319 		err = errno;
320 		goto error;
321 	}
322 	sh->tx_domain = domain;
323 #ifdef HAVE_MLX5DV_DR_ESWITCH
324 	if (priv->config.dv_esw_en) {
325 		domain  = mlx5_glue->dr_create_domain
326 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
327 		if (!domain) {
328 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
329 			err = errno;
330 			goto error;
331 		}
332 		sh->fdb_domain = domain;
333 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
334 	}
335 #endif
336 	if (!sh->tunnel_hub)
337 		err = mlx5_alloc_tunnel_hub(sh);
338 	if (err) {
339 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
340 		goto error;
341 	}
342 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
343 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
344 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
345 		if (sh->fdb_domain)
346 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
347 	}
348 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
349 #endif /* HAVE_MLX5DV_DR */
350 	sh->default_miss_action =
351 			mlx5_glue->dr_create_flow_action_default_miss();
352 	if (!sh->default_miss_action)
353 		DRV_LOG(WARNING, "Default miss action is not supported.");
354 	return 0;
355 error:
356 	/* Rollback the created objects. */
357 	if (sh->rx_domain) {
358 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
359 		sh->rx_domain = NULL;
360 	}
361 	if (sh->tx_domain) {
362 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
363 		sh->tx_domain = NULL;
364 	}
365 	if (sh->fdb_domain) {
366 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
367 		sh->fdb_domain = NULL;
368 	}
369 	if (sh->esw_drop_action) {
370 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
371 		sh->esw_drop_action = NULL;
372 	}
373 	if (sh->pop_vlan_action) {
374 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
375 		sh->pop_vlan_action = NULL;
376 	}
377 	if (sh->encaps_decaps) {
378 		mlx5_hlist_destroy(sh->encaps_decaps);
379 		sh->encaps_decaps = NULL;
380 	}
381 	if (sh->modify_cmds) {
382 		mlx5_hlist_destroy(sh->modify_cmds);
383 		sh->modify_cmds = NULL;
384 	}
385 	if (sh->tag_table) {
386 		/* tags should be destroyed with flow before. */
387 		mlx5_hlist_destroy(sh->tag_table);
388 		sh->tag_table = NULL;
389 	}
390 	if (sh->tunnel_hub) {
391 		mlx5_release_tunnel_hub(sh, priv->dev_port);
392 		sh->tunnel_hub = NULL;
393 	}
394 	mlx5_free_table_hash_list(priv);
395 	return err;
396 }
397 
398 /**
399  * Destroy DR related data within private structure.
400  *
401  * @param[in] priv
402  *   Pointer to the private device data structure.
403  */
404 void
405 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
406 {
407 	struct mlx5_dev_ctx_shared *sh = priv->sh;
408 
409 	MLX5_ASSERT(sh && sh->refcnt);
410 	if (sh->refcnt > 1)
411 		return;
412 #ifdef HAVE_MLX5DV_DR
413 	if (sh->rx_domain) {
414 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
415 		sh->rx_domain = NULL;
416 	}
417 	if (sh->tx_domain) {
418 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
419 		sh->tx_domain = NULL;
420 	}
421 #ifdef HAVE_MLX5DV_DR_ESWITCH
422 	if (sh->fdb_domain) {
423 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
424 		sh->fdb_domain = NULL;
425 	}
426 	if (sh->esw_drop_action) {
427 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
428 		sh->esw_drop_action = NULL;
429 	}
430 #endif
431 	if (sh->pop_vlan_action) {
432 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
433 		sh->pop_vlan_action = NULL;
434 	}
435 #endif /* HAVE_MLX5DV_DR */
436 	if (sh->default_miss_action)
437 		mlx5_glue->destroy_flow_action
438 				(sh->default_miss_action);
439 	if (sh->encaps_decaps) {
440 		mlx5_hlist_destroy(sh->encaps_decaps);
441 		sh->encaps_decaps = NULL;
442 	}
443 	if (sh->modify_cmds) {
444 		mlx5_hlist_destroy(sh->modify_cmds);
445 		sh->modify_cmds = NULL;
446 	}
447 	if (sh->tag_table) {
448 		/* tags should be destroyed with flow before. */
449 		mlx5_hlist_destroy(sh->tag_table);
450 		sh->tag_table = NULL;
451 	}
452 	if (sh->tunnel_hub) {
453 		mlx5_release_tunnel_hub(sh, priv->dev_port);
454 		sh->tunnel_hub = NULL;
455 	}
456 	mlx5_cache_list_destroy(&sh->port_id_action_list);
457 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
458 	mlx5_free_table_hash_list(priv);
459 }
460 
461 /**
462  * Initialize shared data between primary and secondary process.
463  *
464  * A memzone is reserved by primary process and secondary processes attach to
465  * the memzone.
466  *
467  * @return
468  *   0 on success, a negative errno value otherwise and rte_errno is set.
469  */
470 static int
471 mlx5_init_shared_data(void)
472 {
473 	const struct rte_memzone *mz;
474 	int ret = 0;
475 
476 	rte_spinlock_lock(&mlx5_shared_data_lock);
477 	if (mlx5_shared_data == NULL) {
478 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
479 			/* Allocate shared memory. */
480 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
481 						 sizeof(*mlx5_shared_data),
482 						 SOCKET_ID_ANY, 0);
483 			if (mz == NULL) {
484 				DRV_LOG(ERR,
485 					"Cannot allocate mlx5 shared data");
486 				ret = -rte_errno;
487 				goto error;
488 			}
489 			mlx5_shared_data = mz->addr;
490 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
491 			rte_spinlock_init(&mlx5_shared_data->lock);
492 		} else {
493 			/* Lookup allocated shared memory. */
494 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
495 			if (mz == NULL) {
496 				DRV_LOG(ERR,
497 					"Cannot attach mlx5 shared data");
498 				ret = -rte_errno;
499 				goto error;
500 			}
501 			mlx5_shared_data = mz->addr;
502 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
503 		}
504 	}
505 error:
506 	rte_spinlock_unlock(&mlx5_shared_data_lock);
507 	return ret;
508 }
509 
510 /**
511  * PMD global initialization.
512  *
513  * Independent from individual device, this function initializes global
514  * per-PMD data structures distinguishing primary and secondary processes.
515  * Hence, each initialization is called once per a process.
516  *
517  * @return
518  *   0 on success, a negative errno value otherwise and rte_errno is set.
519  */
520 static int
521 mlx5_init_once(void)
522 {
523 	struct mlx5_shared_data *sd;
524 	struct mlx5_local_data *ld = &mlx5_local_data;
525 	int ret = 0;
526 
527 	if (mlx5_init_shared_data())
528 		return -rte_errno;
529 	sd = mlx5_shared_data;
530 	MLX5_ASSERT(sd);
531 	rte_spinlock_lock(&sd->lock);
532 	switch (rte_eal_process_type()) {
533 	case RTE_PROC_PRIMARY:
534 		if (sd->init_done)
535 			break;
536 		LIST_INIT(&sd->mem_event_cb_list);
537 		rte_rwlock_init(&sd->mem_event_rwlock);
538 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
539 						mlx5_mr_mem_event_cb, NULL);
540 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
541 					   mlx5_mp_os_primary_handle);
542 		if (ret)
543 			goto out;
544 		sd->init_done = true;
545 		break;
546 	case RTE_PROC_SECONDARY:
547 		if (ld->init_done)
548 			break;
549 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
550 					     mlx5_mp_os_secondary_handle);
551 		if (ret)
552 			goto out;
553 		++sd->secondary_cnt;
554 		ld->init_done = true;
555 		break;
556 	default:
557 		break;
558 	}
559 out:
560 	rte_spinlock_unlock(&sd->lock);
561 	return ret;
562 }
563 
564 /**
565  * Create the Tx queue DevX/Verbs object.
566  *
567  * @param dev
568  *   Pointer to Ethernet device.
569  * @param idx
570  *   Queue index in DPDK Tx queue array.
571  *
572  * @return
573  *   0 on success, a negative errno value otherwise and rte_errno is set.
574  */
575 static int
576 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
577 {
578 	struct mlx5_priv *priv = dev->data->dev_private;
579 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
580 	struct mlx5_txq_ctrl *txq_ctrl =
581 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
582 
583 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
584 		return mlx5_txq_devx_obj_new(dev, idx);
585 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
586 	if (!priv->config.dv_esw_en)
587 		return mlx5_txq_devx_obj_new(dev, idx);
588 #endif
589 	return mlx5_txq_ibv_obj_new(dev, idx);
590 }
591 
592 /**
593  * Release an Tx DevX/verbs queue object.
594  *
595  * @param txq_obj
596  *   DevX/Verbs Tx queue object.
597  */
598 static void
599 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
600 {
601 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
602 		mlx5_txq_devx_obj_release(txq_obj);
603 		return;
604 	}
605 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
606 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
607 		mlx5_txq_devx_obj_release(txq_obj);
608 		return;
609 	}
610 #endif
611 	mlx5_txq_ibv_obj_release(txq_obj);
612 }
613 
614 /**
615  * DV flow counter mode detect and config.
616  *
617  * @param dev
618  *   Pointer to rte_eth_dev structure.
619  *
620  */
621 static void
622 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
623 {
624 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
625 	struct mlx5_priv *priv = dev->data->dev_private;
626 	struct mlx5_dev_ctx_shared *sh = priv->sh;
627 	bool fallback;
628 
629 #ifndef HAVE_IBV_DEVX_ASYNC
630 	fallback = true;
631 #else
632 	fallback = false;
633 	if (!priv->config.devx || !priv->config.dv_flow_en ||
634 	    !priv->config.hca_attr.flow_counters_dump ||
635 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
636 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
637 		fallback = true;
638 #endif
639 	if (fallback)
640 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
641 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
642 			priv->config.hca_attr.flow_counters_dump,
643 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
644 	/* Initialize fallback mode only on the port initializes sh. */
645 	if (sh->refcnt == 1)
646 		sh->cmng.counter_fallback = fallback;
647 	else if (fallback != sh->cmng.counter_fallback)
648 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
649 			"with others:%d.", PORT_ID(priv), fallback);
650 #endif
651 }
652 
653 /**
654  * Spawn an Ethernet device from Verbs information.
655  *
656  * @param dpdk_dev
657  *   Backing DPDK device.
658  * @param spawn
659  *   Verbs device parameters (name, port, switch_info) to spawn.
660  * @param config
661  *   Device configuration parameters.
662  *
663  * @return
664  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
665  *   is set. The following errors are defined:
666  *
667  *   EBUSY: device is not supposed to be spawned.
668  *   EEXIST: device is already spawned
669  */
670 static struct rte_eth_dev *
671 mlx5_dev_spawn(struct rte_device *dpdk_dev,
672 	       struct mlx5_dev_spawn_data *spawn,
673 	       struct mlx5_dev_config *config)
674 {
675 	const struct mlx5_switch_info *switch_info = &spawn->info;
676 	struct mlx5_dev_ctx_shared *sh = NULL;
677 	struct ibv_port_attr port_attr;
678 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
679 	struct rte_eth_dev *eth_dev = NULL;
680 	struct mlx5_priv *priv = NULL;
681 	int err = 0;
682 	unsigned int hw_padding = 0;
683 	unsigned int mps;
684 	unsigned int cqe_comp;
685 	unsigned int cqe_pad = 0;
686 	unsigned int tunnel_en = 0;
687 	unsigned int mpls_en = 0;
688 	unsigned int swp = 0;
689 	unsigned int mprq = 0;
690 	unsigned int mprq_min_stride_size_n = 0;
691 	unsigned int mprq_max_stride_size_n = 0;
692 	unsigned int mprq_min_stride_num_n = 0;
693 	unsigned int mprq_max_stride_num_n = 0;
694 	struct rte_ether_addr mac;
695 	char name[RTE_ETH_NAME_MAX_LEN];
696 	int own_domain_id = 0;
697 	uint16_t port_id;
698 	unsigned int i;
699 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
700 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
701 #endif
702 
703 	/* Determine if this port representor is supposed to be spawned. */
704 	if (switch_info->representor && dpdk_dev->devargs) {
705 		struct rte_eth_devargs eth_da;
706 
707 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
708 		if (err) {
709 			rte_errno = -err;
710 			DRV_LOG(ERR, "failed to process device arguments: %s",
711 				strerror(rte_errno));
712 			return NULL;
713 		}
714 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
715 			if (eth_da.representor_ports[i] ==
716 			    (uint16_t)switch_info->port_name)
717 				break;
718 		if (i == eth_da.nb_representor_ports) {
719 			rte_errno = EBUSY;
720 			return NULL;
721 		}
722 	}
723 	/* Build device name. */
724 	if (spawn->pf_bond <  0) {
725 		/* Single device. */
726 		if (!switch_info->representor)
727 			strlcpy(name, dpdk_dev->name, sizeof(name));
728 		else
729 			snprintf(name, sizeof(name), "%s_representor_%u",
730 				 dpdk_dev->name, switch_info->port_name);
731 	} else {
732 		/* Bonding device. */
733 		if (!switch_info->representor)
734 			snprintf(name, sizeof(name), "%s_%s",
735 				 dpdk_dev->name,
736 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
737 		else
738 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
739 				 dpdk_dev->name,
740 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
741 				 switch_info->port_name);
742 	}
743 	/* check if the device is already spawned */
744 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
745 		rte_errno = EEXIST;
746 		return NULL;
747 	}
748 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
749 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
750 		struct mlx5_mp_id mp_id;
751 
752 		eth_dev = rte_eth_dev_attach_secondary(name);
753 		if (eth_dev == NULL) {
754 			DRV_LOG(ERR, "can not attach rte ethdev");
755 			rte_errno = ENOMEM;
756 			return NULL;
757 		}
758 		priv = eth_dev->data->dev_private;
759 		if (priv->sh->bond_dev != UINT16_MAX)
760 			/* For bonding port, use primary PCI device. */
761 			eth_dev->device =
762 				rte_eth_devices[priv->sh->bond_dev].device;
763 		else
764 			eth_dev->device = dpdk_dev;
765 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
766 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
767 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
768 		err = mlx5_proc_priv_init(eth_dev);
769 		if (err)
770 			return NULL;
771 		mp_id.port_id = eth_dev->data->port_id;
772 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
773 		/* Receive command fd from primary process */
774 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
775 		if (err < 0)
776 			goto err_secondary;
777 		/* Remap UAR for Tx queues. */
778 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
779 		if (err)
780 			goto err_secondary;
781 		/*
782 		 * Ethdev pointer is still required as input since
783 		 * the primary device is not accessible from the
784 		 * secondary process.
785 		 */
786 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
787 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
788 		return eth_dev;
789 err_secondary:
790 		mlx5_dev_close(eth_dev);
791 		return NULL;
792 	}
793 	/*
794 	 * Some parameters ("tx_db_nc" in particularly) are needed in
795 	 * advance to create dv/verbs device context. We proceed the
796 	 * devargs here to get ones, and later proceed devargs again
797 	 * to override some hardware settings.
798 	 */
799 	err = mlx5_args(config, dpdk_dev->devargs);
800 	if (err) {
801 		err = rte_errno;
802 		DRV_LOG(ERR, "failed to process device arguments: %s",
803 			strerror(rte_errno));
804 		goto error;
805 	}
806 	if (config->dv_miss_info) {
807 		if (switch_info->master || switch_info->representor)
808 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
809 	}
810 	mlx5_malloc_mem_select(config->sys_mem_en);
811 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
812 	if (!sh)
813 		return NULL;
814 	config->devx = sh->devx;
815 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
816 	config->dest_tir = 1;
817 #endif
818 #ifdef HAVE_IBV_MLX5_MOD_SWP
819 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
820 #endif
821 	/*
822 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
823 	 * as all ConnectX-5 devices.
824 	 */
825 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
826 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
827 #endif
828 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
829 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
830 #endif
831 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
832 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
833 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
834 			DRV_LOG(DEBUG, "enhanced MPW is supported");
835 			mps = MLX5_MPW_ENHANCED;
836 		} else {
837 			DRV_LOG(DEBUG, "MPW is supported");
838 			mps = MLX5_MPW;
839 		}
840 	} else {
841 		DRV_LOG(DEBUG, "MPW isn't supported");
842 		mps = MLX5_MPW_DISABLED;
843 	}
844 #ifdef HAVE_IBV_MLX5_MOD_SWP
845 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
846 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
847 	DRV_LOG(DEBUG, "SWP support: %u", swp);
848 #endif
849 	config->swp = !!swp;
850 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
851 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
852 		struct mlx5dv_striding_rq_caps mprq_caps =
853 			dv_attr.striding_rq_caps;
854 
855 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
856 			mprq_caps.min_single_stride_log_num_of_bytes);
857 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
858 			mprq_caps.max_single_stride_log_num_of_bytes);
859 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
860 			mprq_caps.min_single_wqe_log_num_of_strides);
861 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
862 			mprq_caps.max_single_wqe_log_num_of_strides);
863 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
864 			mprq_caps.supported_qpts);
865 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
866 		mprq = 1;
867 		mprq_min_stride_size_n =
868 			mprq_caps.min_single_stride_log_num_of_bytes;
869 		mprq_max_stride_size_n =
870 			mprq_caps.max_single_stride_log_num_of_bytes;
871 		mprq_min_stride_num_n =
872 			mprq_caps.min_single_wqe_log_num_of_strides;
873 		mprq_max_stride_num_n =
874 			mprq_caps.max_single_wqe_log_num_of_strides;
875 	}
876 #endif
877 	if (RTE_CACHE_LINE_SIZE == 128 &&
878 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
879 		cqe_comp = 0;
880 	else
881 		cqe_comp = 1;
882 	config->cqe_comp = cqe_comp;
883 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
884 	/* Whether device supports 128B Rx CQE padding. */
885 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
886 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
887 #endif
888 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
889 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
890 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
891 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
892 			     (dv_attr.tunnel_offloads_caps &
893 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
894 			     (dv_attr.tunnel_offloads_caps &
895 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
896 	}
897 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
898 		tunnel_en ? "" : "not ");
899 #else
900 	DRV_LOG(WARNING,
901 		"tunnel offloading disabled due to old OFED/rdma-core version");
902 #endif
903 	config->tunnel_en = tunnel_en;
904 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
905 	mpls_en = ((dv_attr.tunnel_offloads_caps &
906 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
907 		   (dv_attr.tunnel_offloads_caps &
908 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
909 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
910 		mpls_en ? "" : "not ");
911 #else
912 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
913 		" old OFED/rdma-core version or firmware configuration");
914 #endif
915 	config->mpls_en = mpls_en;
916 	/* Check port status. */
917 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
918 	if (err) {
919 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
920 		goto error;
921 	}
922 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
923 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
924 		err = EINVAL;
925 		goto error;
926 	}
927 	if (port_attr.state != IBV_PORT_ACTIVE)
928 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
929 			mlx5_glue->port_state_str(port_attr.state),
930 			port_attr.state);
931 	/* Allocate private eth device data. */
932 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
933 			   sizeof(*priv),
934 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
935 	if (priv == NULL) {
936 		DRV_LOG(ERR, "priv allocation failure");
937 		err = ENOMEM;
938 		goto error;
939 	}
940 	priv->sh = sh;
941 	priv->dev_port = spawn->phys_port;
942 	priv->pci_dev = spawn->pci_dev;
943 	priv->mtu = RTE_ETHER_MTU;
944 	priv->mp_id.port_id = port_id;
945 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
946 	/* Some internal functions rely on Netlink sockets, open them now. */
947 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
948 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
949 	priv->representor = !!switch_info->representor;
950 	priv->master = !!switch_info->master;
951 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
952 	priv->vport_meta_tag = 0;
953 	priv->vport_meta_mask = 0;
954 	priv->pf_bond = spawn->pf_bond;
955 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
956 	/*
957 	 * The DevX port query API is implemented. E-Switch may use
958 	 * either vport or reg_c[0] metadata register to match on
959 	 * vport index. The engaged part of metadata register is
960 	 * defined by mask.
961 	 */
962 	if (switch_info->representor || switch_info->master) {
963 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
964 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
965 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
966 						 &devx_port);
967 		if (err) {
968 			DRV_LOG(WARNING,
969 				"can't query devx port %d on device %s",
970 				spawn->phys_port,
971 				mlx5_os_get_dev_device_name(spawn->phys_dev));
972 			devx_port.comp_mask = 0;
973 		}
974 	}
975 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
976 		priv->vport_meta_tag = devx_port.reg_c_0.value;
977 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
978 		if (!priv->vport_meta_mask) {
979 			DRV_LOG(ERR, "vport zero mask for port %d"
980 				     " on bonding device %s",
981 				     spawn->phys_port,
982 				     mlx5_os_get_dev_device_name
983 							(spawn->phys_dev));
984 			err = ENOTSUP;
985 			goto error;
986 		}
987 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
988 			DRV_LOG(ERR, "invalid vport tag for port %d"
989 				     " on bonding device %s",
990 				     spawn->phys_port,
991 				     mlx5_os_get_dev_device_name
992 							(spawn->phys_dev));
993 			err = ENOTSUP;
994 			goto error;
995 		}
996 	}
997 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
998 		priv->vport_id = devx_port.vport_num;
999 	} else if (spawn->pf_bond >= 0) {
1000 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1001 			     " on bonding device %s",
1002 			     spawn->phys_port,
1003 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1004 		err = ENOTSUP;
1005 		goto error;
1006 	} else {
1007 		/* Suppose vport index in compatible way. */
1008 		priv->vport_id = switch_info->representor ?
1009 				 switch_info->port_name + 1 : -1;
1010 	}
1011 #else
1012 	/*
1013 	 * Kernel/rdma_core support single E-Switch per PF configurations
1014 	 * only and vport_id field contains the vport index for
1015 	 * associated VF, which is deduced from representor port name.
1016 	 * For example, let's have the IB device port 10, it has
1017 	 * attached network device eth0, which has port name attribute
1018 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1019 	 * as 3 (2+1). This assigning schema should be changed if the
1020 	 * multiple E-Switch instances per PF configurations or/and PCI
1021 	 * subfunctions are added.
1022 	 */
1023 	priv->vport_id = switch_info->representor ?
1024 			 switch_info->port_name + 1 : -1;
1025 #endif
1026 	/* representor_id field keeps the unmodified VF index. */
1027 	priv->representor_id = switch_info->representor ?
1028 			       switch_info->port_name : -1;
1029 	/*
1030 	 * Look for sibling devices in order to reuse their switch domain
1031 	 * if any, otherwise allocate one.
1032 	 */
1033 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1034 		const struct mlx5_priv *opriv =
1035 			rte_eth_devices[port_id].data->dev_private;
1036 
1037 		if (!opriv ||
1038 		    opriv->sh != priv->sh ||
1039 			opriv->domain_id ==
1040 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1041 			continue;
1042 		priv->domain_id = opriv->domain_id;
1043 		break;
1044 	}
1045 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1046 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1047 		if (err) {
1048 			err = rte_errno;
1049 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1050 				strerror(rte_errno));
1051 			goto error;
1052 		}
1053 		own_domain_id = 1;
1054 	}
1055 	/* Override some values set by hardware configuration. */
1056 	mlx5_args(config, dpdk_dev->devargs);
1057 	err = mlx5_dev_check_sibling_config(priv, config);
1058 	if (err)
1059 		goto error;
1060 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1061 			    IBV_DEVICE_RAW_IP_CSUM);
1062 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1063 		(config->hw_csum ? "" : "not "));
1064 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1065 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1066 	DRV_LOG(DEBUG, "counters are not supported");
1067 #endif
1068 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1069 	if (config->dv_flow_en) {
1070 		DRV_LOG(WARNING, "DV flow is not supported");
1071 		config->dv_flow_en = 0;
1072 	}
1073 #endif
1074 	config->ind_table_max_size =
1075 		sh->device_attr.max_rwq_indirection_table_size;
1076 	/*
1077 	 * Remove this check once DPDK supports larger/variable
1078 	 * indirection tables.
1079 	 */
1080 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1081 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1082 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1083 		config->ind_table_max_size);
1084 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1085 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1086 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1087 		(config->hw_vlan_strip ? "" : "not "));
1088 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1089 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1090 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1091 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1092 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1093 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1094 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1095 #endif
1096 	if (config->hw_padding && !hw_padding) {
1097 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1098 		config->hw_padding = 0;
1099 	} else if (config->hw_padding) {
1100 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1101 	}
1102 	config->tso = (sh->device_attr.max_tso > 0 &&
1103 		      (sh->device_attr.tso_supported_qpts &
1104 		       (1 << IBV_QPT_RAW_PACKET)));
1105 	if (config->tso)
1106 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1107 	/*
1108 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1109 	 * by default.
1110 	 */
1111 	if (config->mps == MLX5_ARG_UNSET)
1112 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1113 							  MLX5_MPW_DISABLED;
1114 	else
1115 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1116 	DRV_LOG(INFO, "%sMPS is %s",
1117 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1118 		config->mps == MLX5_MPW ? "legacy " : "",
1119 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1120 	if (config->cqe_comp && !cqe_comp) {
1121 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1122 		config->cqe_comp = 0;
1123 	}
1124 	if (config->cqe_pad && !cqe_pad) {
1125 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1126 		config->cqe_pad = 0;
1127 	} else if (config->cqe_pad) {
1128 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1129 	}
1130 	if (config->devx) {
1131 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1132 		if (err) {
1133 			err = -err;
1134 			goto error;
1135 		}
1136 		/* Check relax ordering support. */
1137 		if (!haswell_broadwell_cpu) {
1138 			sh->cmng.relaxed_ordering_write =
1139 				config->hca_attr.relaxed_ordering_write;
1140 			sh->cmng.relaxed_ordering_read =
1141 				config->hca_attr.relaxed_ordering_read;
1142 		} else {
1143 			sh->cmng.relaxed_ordering_read = 0;
1144 			sh->cmng.relaxed_ordering_write = 0;
1145 		}
1146 		/* Check for LRO support. */
1147 		if (config->dest_tir && config->hca_attr.lro_cap &&
1148 		    config->dv_flow_en) {
1149 			/* TBD check tunnel lro caps. */
1150 			config->lro.supported = config->hca_attr.lro_cap;
1151 			DRV_LOG(DEBUG, "Device supports LRO");
1152 			/*
1153 			 * If LRO timeout is not configured by application,
1154 			 * use the minimal supported value.
1155 			 */
1156 			if (!config->lro.timeout)
1157 				config->lro.timeout =
1158 				config->hca_attr.lro_timer_supported_periods[0];
1159 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1160 				config->lro.timeout);
1161 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1162 				"required for coalescing is %d bytes",
1163 				config->hca_attr.lro_min_mss_size);
1164 		}
1165 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1166 		if (config->hca_attr.qos.sup &&
1167 		    config->hca_attr.qos.srtcm_sup &&
1168 		    config->dv_flow_en) {
1169 			uint8_t reg_c_mask =
1170 				config->hca_attr.qos.flow_meter_reg_c_ids;
1171 			/*
1172 			 * Meter needs two REG_C's for color match and pre-sfx
1173 			 * flow match. Here get the REG_C for color match.
1174 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1175 			 */
1176 			reg_c_mask &= 0xfc;
1177 			if (__builtin_popcount(reg_c_mask) < 1) {
1178 				priv->mtr_en = 0;
1179 				DRV_LOG(WARNING, "No available register for"
1180 					" meter.");
1181 			} else {
1182 				/*
1183 				 * The meter color register is used by the
1184 				 * flow-hit feature as well.
1185 				 * The flow-hit feature must use REG_C_3
1186 				 * Prefer REG_C_3 if it is available.
1187 				 */
1188 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1189 					priv->mtr_color_reg = REG_C_3;
1190 				else
1191 					priv->mtr_color_reg = ffs(reg_c_mask)
1192 							      - 1 + REG_C_0;
1193 				priv->mtr_en = 1;
1194 				priv->mtr_reg_share =
1195 				      config->hca_attr.qos.flow_meter_reg_share;
1196 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1197 					priv->mtr_color_reg);
1198 			}
1199 		}
1200 #endif
1201 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1202 		if (config->hca_attr.flow_hit_aso &&
1203 		    priv->mtr_color_reg == REG_C_3) {
1204 			sh->flow_hit_aso_en = 1;
1205 			err = mlx5_flow_aso_age_mng_init(sh);
1206 			if (err) {
1207 				err = -err;
1208 				goto error;
1209 			}
1210 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1211 		}
1212 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1213 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1214 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1215 		    config->dv_flow_en) {
1216 			priv->sampler_en = 1;
1217 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1218 		} else {
1219 			priv->sampler_en = 0;
1220 			if (!config->hca_attr.log_max_ft_sampler_num)
1221 				DRV_LOG(WARNING, "No available register for"
1222 						" Sampler.");
1223 			else
1224 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1225 		}
1226 #endif
1227 	}
1228 	if (config->tx_pp) {
1229 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1230 			config->hca_attr.dev_freq_khz);
1231 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1232 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1233 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1234 			config->hca_attr.cross_channel ? "" : "not ");
1235 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1236 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1237 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1238 			config->hca_attr.non_wire_sq ? "" : "not ");
1239 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1240 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1241 			config->hca_attr.log_max_static_sq_wq);
1242 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1243 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1244 		if (!config->devx) {
1245 			DRV_LOG(ERR, "DevX is required for packet pacing");
1246 			err = ENODEV;
1247 			goto error;
1248 		}
1249 		if (!config->hca_attr.qos.packet_pacing) {
1250 			DRV_LOG(ERR, "Packet pacing is not supported");
1251 			err = ENODEV;
1252 			goto error;
1253 		}
1254 		if (!config->hca_attr.cross_channel) {
1255 			DRV_LOG(ERR, "Cross channel operations are"
1256 				     " required for packet pacing");
1257 			err = ENODEV;
1258 			goto error;
1259 		}
1260 		if (!config->hca_attr.wqe_index_ignore) {
1261 			DRV_LOG(ERR, "WQE index ignore feature is"
1262 				     " required for packet pacing");
1263 			err = ENODEV;
1264 			goto error;
1265 		}
1266 		if (!config->hca_attr.non_wire_sq) {
1267 			DRV_LOG(ERR, "Non-wire SQ feature is"
1268 				     " required for packet pacing");
1269 			err = ENODEV;
1270 			goto error;
1271 		}
1272 		if (!config->hca_attr.log_max_static_sq_wq) {
1273 			DRV_LOG(ERR, "Static WQE SQ feature is"
1274 				     " required for packet pacing");
1275 			err = ENODEV;
1276 			goto error;
1277 		}
1278 		if (!config->hca_attr.qos.wqe_rate_pp) {
1279 			DRV_LOG(ERR, "WQE rate mode is required"
1280 				     " for packet pacing");
1281 			err = ENODEV;
1282 			goto error;
1283 		}
1284 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1285 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1286 			     " can't create queues for packet pacing");
1287 		err = ENODEV;
1288 		goto error;
1289 #endif
1290 	}
1291 	if (config->devx) {
1292 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1293 
1294 		err = config->hca_attr.access_register_user ?
1295 			mlx5_devx_cmd_register_read
1296 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1297 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1298 		if (!err) {
1299 			uint32_t ts_mode;
1300 
1301 			/* MTUTC register is read successfully. */
1302 			ts_mode = MLX5_GET(register_mtutc, reg,
1303 					   time_stamp_mode);
1304 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1305 				config->rt_timestamp = 1;
1306 		} else {
1307 			/* Kernel does not support register reading. */
1308 			if (config->hca_attr.dev_freq_khz ==
1309 						 (NS_PER_S / MS_PER_S))
1310 				config->rt_timestamp = 1;
1311 		}
1312 	}
1313 	/*
1314 	 * If HW has bug working with tunnel packet decapsulation and
1315 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1316 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1317 	 */
1318 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1319 		config->hw_fcs_strip = 0;
1320 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1321 		(config->hw_fcs_strip ? "" : "not "));
1322 	if (config->mprq.enabled && mprq) {
1323 		if (config->mprq.stride_num_n &&
1324 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1325 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1326 			config->mprq.stride_num_n =
1327 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1328 						mprq_min_stride_num_n),
1329 					mprq_max_stride_num_n);
1330 			DRV_LOG(WARNING,
1331 				"the number of strides"
1332 				" for Multi-Packet RQ is out of range,"
1333 				" setting default value (%u)",
1334 				1 << config->mprq.stride_num_n);
1335 		}
1336 		if (config->mprq.stride_size_n &&
1337 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1338 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1339 			config->mprq.stride_size_n =
1340 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1341 						mprq_min_stride_size_n),
1342 					mprq_max_stride_size_n);
1343 			DRV_LOG(WARNING,
1344 				"the size of a stride"
1345 				" for Multi-Packet RQ is out of range,"
1346 				" setting default value (%u)",
1347 				1 << config->mprq.stride_size_n);
1348 		}
1349 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1350 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1351 	} else if (config->mprq.enabled && !mprq) {
1352 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1353 		config->mprq.enabled = 0;
1354 	}
1355 	if (config->max_dump_files_num == 0)
1356 		config->max_dump_files_num = 128;
1357 	eth_dev = rte_eth_dev_allocate(name);
1358 	if (eth_dev == NULL) {
1359 		DRV_LOG(ERR, "can not allocate rte ethdev");
1360 		err = ENOMEM;
1361 		goto error;
1362 	}
1363 	if (priv->representor) {
1364 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1365 		eth_dev->data->representor_id = priv->representor_id;
1366 	}
1367 	/*
1368 	 * Store associated network device interface index. This index
1369 	 * is permanent throughout the lifetime of device. So, we may store
1370 	 * the ifindex here and use the cached value further.
1371 	 */
1372 	MLX5_ASSERT(spawn->ifindex);
1373 	priv->if_index = spawn->ifindex;
1374 	if (priv->pf_bond >= 0 && priv->master) {
1375 		/* Get bond interface info */
1376 		err = mlx5_sysfs_bond_info(priv->if_index,
1377 				     &priv->bond_ifindex,
1378 				     priv->bond_name);
1379 		if (err)
1380 			DRV_LOG(ERR, "unable to get bond info: %s",
1381 				strerror(rte_errno));
1382 		else
1383 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1384 				priv->if_index, priv->bond_ifindex,
1385 				priv->bond_name);
1386 	}
1387 	eth_dev->data->dev_private = priv;
1388 	priv->dev_data = eth_dev->data;
1389 	eth_dev->data->mac_addrs = priv->mac;
1390 	if (spawn->pf_bond < 0) {
1391 		eth_dev->device = dpdk_dev;
1392 	} else {
1393 		/* Use primary bond PCI as device. */
1394 		if (sh->bond_dev == UINT16_MAX) {
1395 			sh->bond_dev = eth_dev->data->port_id;
1396 			eth_dev->device = dpdk_dev;
1397 		} else {
1398 			eth_dev->device = rte_eth_devices[sh->bond_dev].device;
1399 		}
1400 	}
1401 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1402 	/* Configure the first MAC address by default. */
1403 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1404 		DRV_LOG(ERR,
1405 			"port %u cannot get MAC address, is mlx5_en"
1406 			" loaded? (errno: %s)",
1407 			eth_dev->data->port_id, strerror(rte_errno));
1408 		err = ENODEV;
1409 		goto error;
1410 	}
1411 	DRV_LOG(INFO,
1412 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1413 		eth_dev->data->port_id,
1414 		mac.addr_bytes[0], mac.addr_bytes[1],
1415 		mac.addr_bytes[2], mac.addr_bytes[3],
1416 		mac.addr_bytes[4], mac.addr_bytes[5]);
1417 #ifdef RTE_LIBRTE_MLX5_DEBUG
1418 	{
1419 		char ifname[IF_NAMESIZE];
1420 
1421 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1422 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1423 				eth_dev->data->port_id, ifname);
1424 		else
1425 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1426 				eth_dev->data->port_id);
1427 	}
1428 #endif
1429 	/* Get actual MTU if possible. */
1430 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1431 	if (err) {
1432 		err = rte_errno;
1433 		goto error;
1434 	}
1435 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1436 		priv->mtu);
1437 	/* Initialize burst functions to prevent crashes before link-up. */
1438 	eth_dev->rx_pkt_burst = removed_rx_burst;
1439 	eth_dev->tx_pkt_burst = removed_tx_burst;
1440 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1441 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1442 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1443 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1444 	/* Register MAC address. */
1445 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1446 	if (config->vf && config->vf_nl_en)
1447 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1448 				      mlx5_ifindex(eth_dev),
1449 				      eth_dev->data->mac_addrs,
1450 				      MLX5_MAX_MAC_ADDRESSES);
1451 	priv->flows = 0;
1452 	priv->ctrl_flows = 0;
1453 	rte_spinlock_init(&priv->flow_list_lock);
1454 	TAILQ_INIT(&priv->flow_meters);
1455 	TAILQ_INIT(&priv->flow_meter_profiles);
1456 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1457 	mlx5_glue->dv_set_context_attr(sh->ctx,
1458 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1459 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1460 				.alloc = &mlx5_alloc_verbs_buf,
1461 				.free = &mlx5_free_verbs_buf,
1462 				.data = priv,
1463 			}));
1464 	/* Bring Ethernet device up. */
1465 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1466 		eth_dev->data->port_id);
1467 	mlx5_set_link_up(eth_dev);
1468 	/*
1469 	 * Even though the interrupt handler is not installed yet,
1470 	 * interrupts will still trigger on the async_fd from
1471 	 * Verbs context returned by ibv_open_device().
1472 	 */
1473 	mlx5_link_update(eth_dev, 0);
1474 #ifdef HAVE_MLX5DV_DR_ESWITCH
1475 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1476 	      (switch_info->representor || switch_info->master)))
1477 		config->dv_esw_en = 0;
1478 #else
1479 	config->dv_esw_en = 0;
1480 #endif
1481 	/* Detect minimal data bytes to inline. */
1482 	mlx5_set_min_inline(spawn, config);
1483 	/* Store device configuration on private structure. */
1484 	priv->config = *config;
1485 	/* Create context for virtual machine VLAN workaround. */
1486 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1487 	if (config->dv_flow_en) {
1488 		err = mlx5_alloc_shared_dr(priv);
1489 		if (err)
1490 			goto error;
1491 	}
1492 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1493 		priv->obj_ops = devx_obj_ops;
1494 		priv->obj_ops.drop_action_create =
1495 						ibv_obj_ops.drop_action_create;
1496 		priv->obj_ops.drop_action_destroy =
1497 						ibv_obj_ops.drop_action_destroy;
1498 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1499 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1500 #else
1501 		if (config->dv_esw_en)
1502 			priv->obj_ops.txq_obj_modify =
1503 						ibv_obj_ops.txq_obj_modify;
1504 #endif
1505 		/* Use specific wrappers for Tx object. */
1506 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1507 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1508 
1509 	} else {
1510 		priv->obj_ops = ibv_obj_ops;
1511 	}
1512 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1513 	if (!priv->drop_queue.hrxq)
1514 		goto error;
1515 	/* Supported Verbs flow priority number detection. */
1516 	err = mlx5_flow_discover_priorities(eth_dev);
1517 	if (err < 0) {
1518 		err = -err;
1519 		goto error;
1520 	}
1521 	priv->config.flow_prio = err;
1522 	if (!priv->config.dv_esw_en &&
1523 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1524 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1525 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1526 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1527 	}
1528 	mlx5_set_metadata_mask(eth_dev);
1529 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1530 	    !priv->sh->dv_regc0_mask) {
1531 		DRV_LOG(ERR, "metadata mode %u is not supported "
1532 			     "(no metadata reg_c[0] is available)",
1533 			     priv->config.dv_xmeta_en);
1534 			err = ENOTSUP;
1535 			goto error;
1536 	}
1537 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1538 			     mlx5_hrxq_create_cb,
1539 			     mlx5_hrxq_match_cb,
1540 			     mlx5_hrxq_remove_cb);
1541 	/* Query availability of metadata reg_c's. */
1542 	err = mlx5_flow_discover_mreg_c(eth_dev);
1543 	if (err < 0) {
1544 		err = -err;
1545 		goto error;
1546 	}
1547 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1548 		DRV_LOG(DEBUG,
1549 			"port %u extensive metadata register is not supported",
1550 			eth_dev->data->port_id);
1551 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1552 			DRV_LOG(ERR, "metadata mode %u is not supported "
1553 				     "(no metadata registers available)",
1554 				     priv->config.dv_xmeta_en);
1555 			err = ENOTSUP;
1556 			goto error;
1557 		}
1558 	}
1559 	if (priv->config.dv_flow_en &&
1560 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1561 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1562 	    priv->sh->dv_regc0_mask) {
1563 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1564 						      MLX5_FLOW_MREG_HTABLE_SZ,
1565 						      0, 0,
1566 						      flow_dv_mreg_create_cb,
1567 						      NULL,
1568 						      flow_dv_mreg_remove_cb);
1569 		if (!priv->mreg_cp_tbl) {
1570 			err = ENOMEM;
1571 			goto error;
1572 		}
1573 		priv->mreg_cp_tbl->ctx = eth_dev;
1574 	}
1575 	rte_spinlock_init(&priv->shared_act_sl);
1576 	mlx5_flow_counter_mode_config(eth_dev);
1577 	if (priv->config.dv_flow_en)
1578 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1579 	return eth_dev;
1580 error:
1581 	if (priv) {
1582 		if (priv->mreg_cp_tbl)
1583 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1584 		if (priv->sh)
1585 			mlx5_os_free_shared_dr(priv);
1586 		if (priv->nl_socket_route >= 0)
1587 			close(priv->nl_socket_route);
1588 		if (priv->nl_socket_rdma >= 0)
1589 			close(priv->nl_socket_rdma);
1590 		if (priv->vmwa_context)
1591 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1592 		if (eth_dev && priv->drop_queue.hrxq)
1593 			mlx5_drop_action_destroy(eth_dev);
1594 		if (own_domain_id)
1595 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1596 		mlx5_cache_list_destroy(&priv->hrxqs);
1597 		mlx5_free(priv);
1598 		if (eth_dev != NULL)
1599 			eth_dev->data->dev_private = NULL;
1600 	}
1601 	if (eth_dev != NULL) {
1602 		/* mac_addrs must not be freed alone because part of
1603 		 * dev_private
1604 		 **/
1605 		eth_dev->data->mac_addrs = NULL;
1606 		rte_eth_dev_release_port(eth_dev);
1607 	}
1608 	if (sh)
1609 		mlx5_free_shared_dev_ctx(sh);
1610 	MLX5_ASSERT(err > 0);
1611 	rte_errno = err;
1612 	return NULL;
1613 }
1614 
1615 /**
1616  * Comparison callback to sort device data.
1617  *
1618  * This is meant to be used with qsort().
1619  *
1620  * @param a[in]
1621  *   Pointer to pointer to first data object.
1622  * @param b[in]
1623  *   Pointer to pointer to second data object.
1624  *
1625  * @return
1626  *   0 if both objects are equal, less than 0 if the first argument is less
1627  *   than the second, greater than 0 otherwise.
1628  */
1629 static int
1630 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1631 {
1632 	const struct mlx5_switch_info *si_a =
1633 		&((const struct mlx5_dev_spawn_data *)a)->info;
1634 	const struct mlx5_switch_info *si_b =
1635 		&((const struct mlx5_dev_spawn_data *)b)->info;
1636 	int ret;
1637 
1638 	/* Master device first. */
1639 	ret = si_b->master - si_a->master;
1640 	if (ret)
1641 		return ret;
1642 	/* Then representor devices. */
1643 	ret = si_b->representor - si_a->representor;
1644 	if (ret)
1645 		return ret;
1646 	/* Unidentified devices come last in no specific order. */
1647 	if (!si_a->representor)
1648 		return 0;
1649 	/* Order representors by name. */
1650 	return si_a->port_name - si_b->port_name;
1651 }
1652 
1653 /**
1654  * Match PCI information for possible slaves of bonding device.
1655  *
1656  * @param[in] ibv_dev
1657  *   Pointer to Infiniband device structure.
1658  * @param[in] pci_dev
1659  *   Pointer to PCI device structure to match PCI address.
1660  * @param[in] nl_rdma
1661  *   Netlink RDMA group socket handle.
1662  *
1663  * @return
1664  *   negative value if no bonding device found, otherwise
1665  *   positive index of slave PF in bonding.
1666  */
1667 static int
1668 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1669 			   const struct rte_pci_device *pci_dev,
1670 			   int nl_rdma)
1671 {
1672 	char ifname[IF_NAMESIZE + 1];
1673 	unsigned int ifindex;
1674 	unsigned int np, i;
1675 	FILE *file = NULL;
1676 	int pf = -1;
1677 
1678 	/*
1679 	 * Try to get master device name. If something goes
1680 	 * wrong suppose the lack of kernel support and no
1681 	 * bonding devices.
1682 	 */
1683 	if (nl_rdma < 0)
1684 		return -1;
1685 	if (!strstr(ibv_dev->name, "bond"))
1686 		return -1;
1687 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1688 	if (!np)
1689 		return -1;
1690 	/*
1691 	 * The Master device might not be on the predefined
1692 	 * port (not on port index 1, it is not garanted),
1693 	 * we have to scan all Infiniband device port and
1694 	 * find master.
1695 	 */
1696 	for (i = 1; i <= np; ++i) {
1697 		/* Check whether Infiniband port is populated. */
1698 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1699 		if (!ifindex)
1700 			continue;
1701 		if (!if_indextoname(ifindex, ifname))
1702 			continue;
1703 		/* Try to read bonding slave names from sysfs. */
1704 		MKSTR(slaves,
1705 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1706 		file = fopen(slaves, "r");
1707 		if (file)
1708 			break;
1709 	}
1710 	if (!file)
1711 		return -1;
1712 	/* Use safe format to check maximal buffer length. */
1713 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1714 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1715 		char tmp_str[IF_NAMESIZE + 32];
1716 		struct rte_pci_addr pci_addr;
1717 		struct mlx5_switch_info	info;
1718 
1719 		/* Process slave interface names in the loop. */
1720 		snprintf(tmp_str, sizeof(tmp_str),
1721 			 "/sys/class/net/%s", ifname);
1722 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1723 			DRV_LOG(WARNING, "can not get PCI address"
1724 					 " for netdev \"%s\"", ifname);
1725 			continue;
1726 		}
1727 		if (pci_dev->addr.domain != pci_addr.domain ||
1728 		    pci_dev->addr.bus != pci_addr.bus ||
1729 		    pci_dev->addr.devid != pci_addr.devid ||
1730 		    pci_dev->addr.function != pci_addr.function)
1731 			continue;
1732 		/* Slave interface PCI address match found. */
1733 		fclose(file);
1734 		snprintf(tmp_str, sizeof(tmp_str),
1735 			 "/sys/class/net/%s/phys_port_name", ifname);
1736 		file = fopen(tmp_str, "rb");
1737 		if (!file)
1738 			break;
1739 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1740 		if (fscanf(file, "%32s", tmp_str) == 1)
1741 			mlx5_translate_port_name(tmp_str, &info);
1742 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1743 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1744 			pf = info.port_name;
1745 		break;
1746 	}
1747 	if (file)
1748 		fclose(file);
1749 	return pf;
1750 }
1751 
1752 /**
1753  * DPDK callback to register a PCI device.
1754  *
1755  * This function spawns Ethernet devices out of a given PCI device.
1756  *
1757  * @param[in] pci_drv
1758  *   PCI driver structure (mlx5_driver).
1759  * @param[in] pci_dev
1760  *   PCI device information.
1761  *
1762  * @return
1763  *   0 on success, a negative errno value otherwise and rte_errno is set.
1764  */
1765 int
1766 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1767 		  struct rte_pci_device *pci_dev)
1768 {
1769 	struct ibv_device **ibv_list;
1770 	/*
1771 	 * Number of found IB Devices matching with requested PCI BDF.
1772 	 * nd != 1 means there are multiple IB devices over the same
1773 	 * PCI device and we have representors and master.
1774 	 */
1775 	unsigned int nd = 0;
1776 	/*
1777 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1778 	 * we have the single multiport IB device, and there may be
1779 	 * representors attached to some of found ports.
1780 	 */
1781 	unsigned int np = 0;
1782 	/*
1783 	 * Number of DPDK ethernet devices to Spawn - either over
1784 	 * multiple IB devices or multiple ports of single IB device.
1785 	 * Actually this is the number of iterations to spawn.
1786 	 */
1787 	unsigned int ns = 0;
1788 	/*
1789 	 * Bonding device
1790 	 *   < 0 - no bonding device (single one)
1791 	 *  >= 0 - bonding device (value is slave PF index)
1792 	 */
1793 	int bd = -1;
1794 	struct mlx5_dev_spawn_data *list = NULL;
1795 	struct mlx5_dev_config dev_config;
1796 	unsigned int dev_config_vf;
1797 	int ret;
1798 
1799 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1800 		mlx5_pmd_socket_init();
1801 	ret = mlx5_init_once();
1802 	if (ret) {
1803 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1804 			strerror(rte_errno));
1805 		return -rte_errno;
1806 	}
1807 	errno = 0;
1808 	ibv_list = mlx5_glue->get_device_list(&ret);
1809 	if (!ibv_list) {
1810 		rte_errno = errno ? errno : ENOSYS;
1811 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1812 		return -rte_errno;
1813 	}
1814 	/*
1815 	 * First scan the list of all Infiniband devices to find
1816 	 * matching ones, gathering into the list.
1817 	 */
1818 	struct ibv_device *ibv_match[ret + 1];
1819 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1820 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1821 	unsigned int i;
1822 
1823 	while (ret-- > 0) {
1824 		struct rte_pci_addr pci_addr;
1825 
1826 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1827 		bd = mlx5_device_bond_pci_match
1828 				(ibv_list[ret], pci_dev, nl_rdma);
1829 		if (bd >= 0) {
1830 			/*
1831 			 * Bonding device detected. Only one match is allowed,
1832 			 * the bonding is supported over multi-port IB device,
1833 			 * there should be no matches on representor PCI
1834 			 * functions or non VF LAG bonding devices with
1835 			 * specified address.
1836 			 */
1837 			if (nd) {
1838 				DRV_LOG(ERR,
1839 					"multiple PCI match on bonding device"
1840 					"\"%s\" found", ibv_list[ret]->name);
1841 				rte_errno = ENOENT;
1842 				ret = -rte_errno;
1843 				goto exit;
1844 			}
1845 			DRV_LOG(INFO, "PCI information matches for"
1846 				      " slave %d bonding device \"%s\"",
1847 				      bd, ibv_list[ret]->name);
1848 			ibv_match[nd++] = ibv_list[ret];
1849 			break;
1850 		}
1851 		if (mlx5_dev_to_pci_addr
1852 			(ibv_list[ret]->ibdev_path, &pci_addr))
1853 			continue;
1854 		if (pci_dev->addr.domain != pci_addr.domain ||
1855 		    pci_dev->addr.bus != pci_addr.bus ||
1856 		    pci_dev->addr.devid != pci_addr.devid ||
1857 		    pci_dev->addr.function != pci_addr.function)
1858 			continue;
1859 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1860 			ibv_list[ret]->name);
1861 		ibv_match[nd++] = ibv_list[ret];
1862 	}
1863 	ibv_match[nd] = NULL;
1864 	if (!nd) {
1865 		/* No device matches, just complain and bail out. */
1866 		DRV_LOG(WARNING,
1867 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1868 			" are kernel drivers loaded?",
1869 			pci_dev->addr.domain, pci_dev->addr.bus,
1870 			pci_dev->addr.devid, pci_dev->addr.function);
1871 		rte_errno = ENOENT;
1872 		ret = -rte_errno;
1873 		goto exit;
1874 	}
1875 	if (nd == 1) {
1876 		/*
1877 		 * Found single matching device may have multiple ports.
1878 		 * Each port may be representor, we have to check the port
1879 		 * number and check the representors existence.
1880 		 */
1881 		if (nl_rdma >= 0)
1882 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1883 		if (!np)
1884 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1885 					 " ports number", ibv_match[0]->name);
1886 		if (bd >= 0 && !np) {
1887 			DRV_LOG(ERR, "can not get ports"
1888 				     " for bonding device");
1889 			rte_errno = ENOENT;
1890 			ret = -rte_errno;
1891 			goto exit;
1892 		}
1893 	}
1894 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1895 	if (bd >= 0) {
1896 		/*
1897 		 * This may happen if there is VF LAG kernel support and
1898 		 * application is compiled with older rdma_core library.
1899 		 */
1900 		DRV_LOG(ERR,
1901 			"No kernel/verbs support for VF LAG bonding found.");
1902 		rte_errno = ENOTSUP;
1903 		ret = -rte_errno;
1904 		goto exit;
1905 	}
1906 #endif
1907 	/*
1908 	 * Now we can determine the maximal
1909 	 * amount of devices to be spawned.
1910 	 */
1911 	list = mlx5_malloc(MLX5_MEM_ZERO,
1912 			   sizeof(struct mlx5_dev_spawn_data) *
1913 			   (np ? np : nd),
1914 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1915 	if (!list) {
1916 		DRV_LOG(ERR, "spawn data array allocation failure");
1917 		rte_errno = ENOMEM;
1918 		ret = -rte_errno;
1919 		goto exit;
1920 	}
1921 	if (bd >= 0 || np > 1) {
1922 		/*
1923 		 * Single IB device with multiple ports found,
1924 		 * it may be E-Switch master device and representors.
1925 		 * We have to perform identification through the ports.
1926 		 */
1927 		MLX5_ASSERT(nl_rdma >= 0);
1928 		MLX5_ASSERT(ns == 0);
1929 		MLX5_ASSERT(nd == 1);
1930 		MLX5_ASSERT(np);
1931 		for (i = 1; i <= np; ++i) {
1932 			list[ns].max_port = np;
1933 			list[ns].phys_port = i;
1934 			list[ns].phys_dev = ibv_match[0];
1935 			list[ns].eth_dev = NULL;
1936 			list[ns].pci_dev = pci_dev;
1937 			list[ns].pf_bond = bd;
1938 			list[ns].ifindex = mlx5_nl_ifindex
1939 				(nl_rdma,
1940 				mlx5_os_get_dev_device_name
1941 						(list[ns].phys_dev), i);
1942 			if (!list[ns].ifindex) {
1943 				/*
1944 				 * No network interface index found for the
1945 				 * specified port, it means there is no
1946 				 * representor on this port. It's OK,
1947 				 * there can be disabled ports, for example
1948 				 * if sriov_numvfs < sriov_totalvfs.
1949 				 */
1950 				continue;
1951 			}
1952 			ret = -1;
1953 			if (nl_route >= 0)
1954 				ret = mlx5_nl_switch_info
1955 					       (nl_route,
1956 						list[ns].ifindex,
1957 						&list[ns].info);
1958 			if (ret || (!list[ns].info.representor &&
1959 				    !list[ns].info.master)) {
1960 				/*
1961 				 * We failed to recognize representors with
1962 				 * Netlink, let's try to perform the task
1963 				 * with sysfs.
1964 				 */
1965 				ret =  mlx5_sysfs_switch_info
1966 						(list[ns].ifindex,
1967 						 &list[ns].info);
1968 			}
1969 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1970 			if (!ret && bd >= 0) {
1971 				switch (list[ns].info.name_type) {
1972 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1973 					if (list[ns].info.port_name == bd)
1974 						ns++;
1975 					break;
1976 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1977 					/* Fallthrough */
1978 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1979 					if (list[ns].info.pf_num == bd)
1980 						ns++;
1981 					break;
1982 				default:
1983 					break;
1984 				}
1985 				continue;
1986 			}
1987 #endif
1988 			if (!ret && (list[ns].info.representor ^
1989 				     list[ns].info.master))
1990 				ns++;
1991 		}
1992 		if (!ns) {
1993 			DRV_LOG(ERR,
1994 				"unable to recognize master/representors"
1995 				" on the IB device with multiple ports");
1996 			rte_errno = ENOENT;
1997 			ret = -rte_errno;
1998 			goto exit;
1999 		}
2000 	} else {
2001 		/*
2002 		 * The existence of several matching entries (nd > 1) means
2003 		 * port representors have been instantiated. No existing Verbs
2004 		 * call nor sysfs entries can tell them apart, this can only
2005 		 * be done through Netlink calls assuming kernel drivers are
2006 		 * recent enough to support them.
2007 		 *
2008 		 * In the event of identification failure through Netlink,
2009 		 * try again through sysfs, then:
2010 		 *
2011 		 * 1. A single IB device matches (nd == 1) with single
2012 		 *    port (np=0/1) and is not a representor, assume
2013 		 *    no switch support.
2014 		 *
2015 		 * 2. Otherwise no safe assumptions can be made;
2016 		 *    complain louder and bail out.
2017 		 */
2018 		for (i = 0; i != nd; ++i) {
2019 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2020 			list[ns].max_port = 1;
2021 			list[ns].phys_port = 1;
2022 			list[ns].phys_dev = ibv_match[i];
2023 			list[ns].eth_dev = NULL;
2024 			list[ns].pci_dev = pci_dev;
2025 			list[ns].pf_bond = -1;
2026 			list[ns].ifindex = 0;
2027 			if (nl_rdma >= 0)
2028 				list[ns].ifindex = mlx5_nl_ifindex
2029 				(nl_rdma,
2030 				mlx5_os_get_dev_device_name
2031 						(list[ns].phys_dev), 1);
2032 			if (!list[ns].ifindex) {
2033 				char ifname[IF_NAMESIZE];
2034 
2035 				/*
2036 				 * Netlink failed, it may happen with old
2037 				 * ib_core kernel driver (before 4.16).
2038 				 * We can assume there is old driver because
2039 				 * here we are processing single ports IB
2040 				 * devices. Let's try sysfs to retrieve
2041 				 * the ifindex. The method works for
2042 				 * master device only.
2043 				 */
2044 				if (nd > 1) {
2045 					/*
2046 					 * Multiple devices found, assume
2047 					 * representors, can not distinguish
2048 					 * master/representor and retrieve
2049 					 * ifindex via sysfs.
2050 					 */
2051 					continue;
2052 				}
2053 				ret = mlx5_get_ifname_sysfs
2054 					(ibv_match[i]->ibdev_path, ifname);
2055 				if (!ret)
2056 					list[ns].ifindex =
2057 						if_nametoindex(ifname);
2058 				if (!list[ns].ifindex) {
2059 					/*
2060 					 * No network interface index found
2061 					 * for the specified device, it means
2062 					 * there it is neither representor
2063 					 * nor master.
2064 					 */
2065 					continue;
2066 				}
2067 			}
2068 			ret = -1;
2069 			if (nl_route >= 0)
2070 				ret = mlx5_nl_switch_info
2071 					       (nl_route,
2072 						list[ns].ifindex,
2073 						&list[ns].info);
2074 			if (ret || (!list[ns].info.representor &&
2075 				    !list[ns].info.master)) {
2076 				/*
2077 				 * We failed to recognize representors with
2078 				 * Netlink, let's try to perform the task
2079 				 * with sysfs.
2080 				 */
2081 				ret =  mlx5_sysfs_switch_info
2082 						(list[ns].ifindex,
2083 						 &list[ns].info);
2084 			}
2085 			if (!ret && (list[ns].info.representor ^
2086 				     list[ns].info.master)) {
2087 				ns++;
2088 			} else if ((nd == 1) &&
2089 				   !list[ns].info.representor &&
2090 				   !list[ns].info.master) {
2091 				/*
2092 				 * Single IB device with
2093 				 * one physical port and
2094 				 * attached network device.
2095 				 * May be SRIOV is not enabled
2096 				 * or there is no representors.
2097 				 */
2098 				DRV_LOG(INFO, "no E-Switch support detected");
2099 				ns++;
2100 				break;
2101 			}
2102 		}
2103 		if (!ns) {
2104 			DRV_LOG(ERR,
2105 				"unable to recognize master/representors"
2106 				" on the multiple IB devices");
2107 			rte_errno = ENOENT;
2108 			ret = -rte_errno;
2109 			goto exit;
2110 		}
2111 	}
2112 	MLX5_ASSERT(ns);
2113 	/*
2114 	 * Sort list to probe devices in natural order for users convenience
2115 	 * (i.e. master first, then representors from lowest to highest ID).
2116 	 */
2117 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2118 	/* Device specific configuration. */
2119 	switch (pci_dev->id.device_id) {
2120 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2121 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2122 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2123 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2124 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2125 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2126 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2127 		dev_config_vf = 1;
2128 		break;
2129 	default:
2130 		dev_config_vf = 0;
2131 		break;
2132 	}
2133 	for (i = 0; i != ns; ++i) {
2134 		uint32_t restore;
2135 
2136 		/* Default configuration. */
2137 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2138 		dev_config.vf = dev_config_vf;
2139 		dev_config.mps = MLX5_ARG_UNSET;
2140 		dev_config.dbnc = MLX5_ARG_UNSET;
2141 		dev_config.rx_vec_en = 1;
2142 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2143 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2144 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2145 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2146 		dev_config.vf_nl_en = 1;
2147 		dev_config.mr_ext_memseg_en = 1;
2148 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2149 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2150 		dev_config.dv_esw_en = 1;
2151 		dev_config.dv_flow_en = 1;
2152 		dev_config.decap_en = 1;
2153 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2154 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2155 						 &list[i],
2156 						 &dev_config);
2157 		if (!list[i].eth_dev) {
2158 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2159 				break;
2160 			/* Device is disabled or already spawned. Ignore it. */
2161 			continue;
2162 		}
2163 		restore = list[i].eth_dev->data->dev_flags;
2164 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2165 		/* Restore non-PCI flags cleared by the above call. */
2166 		list[i].eth_dev->data->dev_flags |= restore;
2167 		rte_eth_dev_probing_finish(list[i].eth_dev);
2168 	}
2169 	if (i != ns) {
2170 		DRV_LOG(ERR,
2171 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2172 			" encountering an error: %s",
2173 			pci_dev->addr.domain, pci_dev->addr.bus,
2174 			pci_dev->addr.devid, pci_dev->addr.function,
2175 			strerror(rte_errno));
2176 		ret = -rte_errno;
2177 		/* Roll back. */
2178 		while (i--) {
2179 			if (!list[i].eth_dev)
2180 				continue;
2181 			mlx5_dev_close(list[i].eth_dev);
2182 			/* mac_addrs must not be freed because in dev_private */
2183 			list[i].eth_dev->data->mac_addrs = NULL;
2184 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2185 		}
2186 		/* Restore original error. */
2187 		rte_errno = -ret;
2188 	} else {
2189 		ret = 0;
2190 	}
2191 exit:
2192 	/*
2193 	 * Do the routine cleanup:
2194 	 * - close opened Netlink sockets
2195 	 * - free allocated spawn data array
2196 	 * - free the Infiniband device list
2197 	 */
2198 	if (nl_rdma >= 0)
2199 		close(nl_rdma);
2200 	if (nl_route >= 0)
2201 		close(nl_route);
2202 	if (list)
2203 		mlx5_free(list);
2204 	MLX5_ASSERT(ibv_list);
2205 	mlx5_glue->free_device_list(ibv_list);
2206 	return ret;
2207 }
2208 
2209 static int
2210 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2211 {
2212 	char *env;
2213 	int value;
2214 
2215 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2216 	/* Get environment variable to store. */
2217 	env = getenv(MLX5_SHUT_UP_BF);
2218 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2219 	if (config->dbnc == MLX5_ARG_UNSET)
2220 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2221 	else
2222 		setenv(MLX5_SHUT_UP_BF,
2223 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2224 	return value;
2225 }
2226 
2227 static void
2228 mlx5_restore_doorbell_mapping_env(int value)
2229 {
2230 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2231 	/* Restore the original environment variable state. */
2232 	if (value == MLX5_ARG_UNSET)
2233 		unsetenv(MLX5_SHUT_UP_BF);
2234 	else
2235 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2236 }
2237 
2238 /**
2239  * Extract pdn of PD object using DV API.
2240  *
2241  * @param[in] pd
2242  *   Pointer to the verbs PD object.
2243  * @param[out] pdn
2244  *   Pointer to the PD object number variable.
2245  *
2246  * @return
2247  *   0 on success, error value otherwise.
2248  */
2249 int
2250 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2251 {
2252 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2253 	struct mlx5dv_obj obj;
2254 	struct mlx5dv_pd pd_info;
2255 	int ret = 0;
2256 
2257 	obj.pd.in = pd;
2258 	obj.pd.out = &pd_info;
2259 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2260 	if (ret) {
2261 		DRV_LOG(DEBUG, "Fail to get PD object info");
2262 		return ret;
2263 	}
2264 	*pdn = pd_info.pdn;
2265 	return 0;
2266 #else
2267 	(void)pd;
2268 	(void)pdn;
2269 	return -ENOTSUP;
2270 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2271 }
2272 
2273 /**
2274  * Function API to open IB device.
2275  *
2276  * This function calls the Linux glue APIs to open a device.
2277  *
2278  * @param[in] spawn
2279  *   Pointer to the IB device attributes (name, port, etc).
2280  * @param[out] config
2281  *   Pointer to device configuration structure.
2282  * @param[out] sh
2283  *   Pointer to shared context structure.
2284  *
2285  * @return
2286  *   0 on success, a positive error value otherwise.
2287  */
2288 int
2289 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2290 		     const struct mlx5_dev_config *config,
2291 		     struct mlx5_dev_ctx_shared *sh)
2292 {
2293 	int dbmap_env;
2294 	int err = 0;
2295 
2296 	sh->numa_node = spawn->pci_dev->device.numa_node;
2297 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2298 	/*
2299 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2300 	 * before the device creation. The rdma_core library
2301 	 * checks the variable at device creation and
2302 	 * stores the result internally.
2303 	 */
2304 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2305 	/* Try to open IB device with DV first, then usual Verbs. */
2306 	errno = 0;
2307 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2308 	if (sh->ctx) {
2309 		sh->devx = 1;
2310 		DRV_LOG(DEBUG, "DevX is supported");
2311 		/* The device is created, no need for environment. */
2312 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2313 	} else {
2314 		/* The environment variable is still configured. */
2315 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2316 		err = errno ? errno : ENODEV;
2317 		/*
2318 		 * The environment variable is not needed anymore,
2319 		 * all device creation attempts are completed.
2320 		 */
2321 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2322 		if (!sh->ctx)
2323 			return err;
2324 		DRV_LOG(DEBUG, "DevX is NOT supported");
2325 		err = 0;
2326 	}
2327 	return err;
2328 }
2329 
2330 /**
2331  * Install shared asynchronous device events handler.
2332  * This function is implemented to support event sharing
2333  * between multiple ports of single IB device.
2334  *
2335  * @param sh
2336  *   Pointer to mlx5_dev_ctx_shared object.
2337  */
2338 void
2339 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2340 {
2341 	int ret;
2342 	int flags;
2343 
2344 	sh->intr_handle.fd = -1;
2345 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2346 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2347 		    F_SETFL, flags | O_NONBLOCK);
2348 	if (ret) {
2349 		DRV_LOG(INFO, "failed to change file descriptor async event"
2350 			" queue");
2351 	} else {
2352 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2353 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2354 		if (rte_intr_callback_register(&sh->intr_handle,
2355 					mlx5_dev_interrupt_handler, sh)) {
2356 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2357 			sh->intr_handle.fd = -1;
2358 		}
2359 	}
2360 	if (sh->devx) {
2361 #ifdef HAVE_IBV_DEVX_ASYNC
2362 		sh->intr_handle_devx.fd = -1;
2363 		sh->devx_comp =
2364 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2365 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2366 		if (!devx_comp) {
2367 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2368 			return;
2369 		}
2370 		flags = fcntl(devx_comp->fd, F_GETFL);
2371 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2372 		if (ret) {
2373 			DRV_LOG(INFO, "failed to change file descriptor"
2374 				" devx comp");
2375 			return;
2376 		}
2377 		sh->intr_handle_devx.fd = devx_comp->fd;
2378 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2379 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2380 					mlx5_dev_interrupt_handler_devx, sh)) {
2381 			DRV_LOG(INFO, "Fail to install the devx shared"
2382 				" interrupt.");
2383 			sh->intr_handle_devx.fd = -1;
2384 		}
2385 #endif /* HAVE_IBV_DEVX_ASYNC */
2386 	}
2387 }
2388 
2389 /**
2390  * Uninstall shared asynchronous device events handler.
2391  * This function is implemented to support event sharing
2392  * between multiple ports of single IB device.
2393  *
2394  * @param dev
2395  *   Pointer to mlx5_dev_ctx_shared object.
2396  */
2397 void
2398 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2399 {
2400 	if (sh->intr_handle.fd >= 0)
2401 		mlx5_intr_callback_unregister(&sh->intr_handle,
2402 					      mlx5_dev_interrupt_handler, sh);
2403 #ifdef HAVE_IBV_DEVX_ASYNC
2404 	if (sh->intr_handle_devx.fd >= 0)
2405 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2406 				  mlx5_dev_interrupt_handler_devx, sh);
2407 	if (sh->devx_comp)
2408 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2409 #endif
2410 }
2411 
2412 /**
2413  * Read statistics by a named counter.
2414  *
2415  * @param[in] priv
2416  *   Pointer to the private device data structure.
2417  * @param[in] ctr_name
2418  *   Pointer to the name of the statistic counter to read
2419  * @param[out] stat
2420  *   Pointer to read statistic value.
2421  * @return
2422  *   0 on success and stat is valud, 1 if failed to read the value
2423  *   rte_errno is set.
2424  *
2425  */
2426 int
2427 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2428 		      uint64_t *stat)
2429 {
2430 	int fd;
2431 
2432 	if (priv->sh) {
2433 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2434 		      priv->sh->ibdev_path,
2435 		      priv->dev_port,
2436 		      ctr_name);
2437 		fd = open(path, O_RDONLY);
2438 		/*
2439 		 * in switchdev the file location is not per port
2440 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2441 		 */
2442 		if (fd == -1) {
2443 			MKSTR(path1, "%s/hw_counters/%s",
2444 			      priv->sh->ibdev_path,
2445 			      ctr_name);
2446 			fd = open(path1, O_RDONLY);
2447 		}
2448 		if (fd != -1) {
2449 			char buf[21] = {'\0'};
2450 			ssize_t n = read(fd, buf, sizeof(buf));
2451 
2452 			close(fd);
2453 			if (n != -1) {
2454 				*stat = strtoull(buf, NULL, 10);
2455 				return 0;
2456 			}
2457 		}
2458 	}
2459 	*stat = 0;
2460 	return 1;
2461 }
2462 
2463 /**
2464  * Set the reg_mr and dereg_mr call backs
2465  *
2466  * @param reg_mr_cb[out]
2467  *   Pointer to reg_mr func
2468  * @param dereg_mr_cb[out]
2469  *   Pointer to dereg_mr func
2470  *
2471  */
2472 void
2473 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2474 		      mlx5_dereg_mr_t *dereg_mr_cb)
2475 {
2476 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2477 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2478 }
2479 
2480 /**
2481  * Remove a MAC address from device
2482  *
2483  * @param dev
2484  *   Pointer to Ethernet device structure.
2485  * @param index
2486  *   MAC address index.
2487  */
2488 void
2489 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2490 {
2491 	struct mlx5_priv *priv = dev->data->dev_private;
2492 	const int vf = priv->config.vf;
2493 
2494 	if (vf)
2495 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2496 					mlx5_ifindex(dev), priv->mac_own,
2497 					&dev->data->mac_addrs[index], index);
2498 }
2499 
2500 /**
2501  * Adds a MAC address to the device
2502  *
2503  * @param dev
2504  *   Pointer to Ethernet device structure.
2505  * @param mac_addr
2506  *   MAC address to register.
2507  * @param index
2508  *   MAC address index.
2509  *
2510  * @return
2511  *   0 on success, a negative errno value otherwise
2512  */
2513 int
2514 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2515 		     uint32_t index)
2516 {
2517 	struct mlx5_priv *priv = dev->data->dev_private;
2518 	const int vf = priv->config.vf;
2519 	int ret = 0;
2520 
2521 	if (vf)
2522 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2523 					   mlx5_ifindex(dev), priv->mac_own,
2524 					   mac, index);
2525 	return ret;
2526 }
2527 
2528 /**
2529  * Modify a VF MAC address
2530  *
2531  * @param priv
2532  *   Pointer to device private data.
2533  * @param mac_addr
2534  *   MAC address to modify into.
2535  * @param iface_idx
2536  *   Net device interface index
2537  * @param vf_index
2538  *   VF index
2539  *
2540  * @return
2541  *   0 on success, a negative errno value otherwise
2542  */
2543 int
2544 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2545 			   unsigned int iface_idx,
2546 			   struct rte_ether_addr *mac_addr,
2547 			   int vf_index)
2548 {
2549 	return mlx5_nl_vf_mac_addr_modify
2550 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2551 }
2552 
2553 /**
2554  * Set device promiscuous mode
2555  *
2556  * @param dev
2557  *   Pointer to Ethernet device structure.
2558  * @param enable
2559  *   0 - promiscuous is disabled, otherwise - enabled
2560  *
2561  * @return
2562  *   0 on success, a negative error value otherwise
2563  */
2564 int
2565 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2566 {
2567 	struct mlx5_priv *priv = dev->data->dev_private;
2568 
2569 	return mlx5_nl_promisc(priv->nl_socket_route,
2570 			       mlx5_ifindex(dev), !!enable);
2571 }
2572 
2573 /**
2574  * Set device promiscuous mode
2575  *
2576  * @param dev
2577  *   Pointer to Ethernet device structure.
2578  * @param enable
2579  *   0 - all multicase is disabled, otherwise - enabled
2580  *
2581  * @return
2582  *   0 on success, a negative error value otherwise
2583  */
2584 int
2585 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2586 {
2587 	struct mlx5_priv *priv = dev->data->dev_private;
2588 
2589 	return mlx5_nl_allmulti(priv->nl_socket_route,
2590 				mlx5_ifindex(dev), !!enable);
2591 }
2592 
2593 /**
2594  * Flush device MAC addresses
2595  *
2596  * @param dev
2597  *   Pointer to Ethernet device structure.
2598  *
2599  */
2600 void
2601 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2602 {
2603 	struct mlx5_priv *priv = dev->data->dev_private;
2604 
2605 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2606 			       dev->data->mac_addrs,
2607 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2608 }
2609 
2610 const struct eth_dev_ops mlx5_os_dev_ops = {
2611 	.dev_configure = mlx5_dev_configure,
2612 	.dev_start = mlx5_dev_start,
2613 	.dev_stop = mlx5_dev_stop,
2614 	.dev_set_link_down = mlx5_set_link_down,
2615 	.dev_set_link_up = mlx5_set_link_up,
2616 	.dev_close = mlx5_dev_close,
2617 	.promiscuous_enable = mlx5_promiscuous_enable,
2618 	.promiscuous_disable = mlx5_promiscuous_disable,
2619 	.allmulticast_enable = mlx5_allmulticast_enable,
2620 	.allmulticast_disable = mlx5_allmulticast_disable,
2621 	.link_update = mlx5_link_update,
2622 	.stats_get = mlx5_stats_get,
2623 	.stats_reset = mlx5_stats_reset,
2624 	.xstats_get = mlx5_xstats_get,
2625 	.xstats_reset = mlx5_xstats_reset,
2626 	.xstats_get_names = mlx5_xstats_get_names,
2627 	.fw_version_get = mlx5_fw_version_get,
2628 	.dev_infos_get = mlx5_dev_infos_get,
2629 	.read_clock = mlx5_txpp_read_clock,
2630 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2631 	.vlan_filter_set = mlx5_vlan_filter_set,
2632 	.rx_queue_setup = mlx5_rx_queue_setup,
2633 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2634 	.tx_queue_setup = mlx5_tx_queue_setup,
2635 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2636 	.rx_queue_release = mlx5_rx_queue_release,
2637 	.tx_queue_release = mlx5_tx_queue_release,
2638 	.rx_queue_start = mlx5_rx_queue_start,
2639 	.rx_queue_stop = mlx5_rx_queue_stop,
2640 	.tx_queue_start = mlx5_tx_queue_start,
2641 	.tx_queue_stop = mlx5_tx_queue_stop,
2642 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2643 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2644 	.mac_addr_remove = mlx5_mac_addr_remove,
2645 	.mac_addr_add = mlx5_mac_addr_add,
2646 	.mac_addr_set = mlx5_mac_addr_set,
2647 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2648 	.mtu_set = mlx5_dev_set_mtu,
2649 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2650 	.vlan_offload_set = mlx5_vlan_offload_set,
2651 	.reta_update = mlx5_dev_rss_reta_update,
2652 	.reta_query = mlx5_dev_rss_reta_query,
2653 	.rss_hash_update = mlx5_rss_hash_update,
2654 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2655 	.filter_ctrl = mlx5_dev_filter_ctrl,
2656 	.rxq_info_get = mlx5_rxq_info_get,
2657 	.txq_info_get = mlx5_txq_info_get,
2658 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2659 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2660 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2661 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2662 	.is_removed = mlx5_is_removed,
2663 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2664 	.get_module_info = mlx5_get_module_info,
2665 	.get_module_eeprom = mlx5_get_module_eeprom,
2666 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2667 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2668 	.hairpin_bind = mlx5_hairpin_bind,
2669 	.hairpin_unbind = mlx5_hairpin_unbind,
2670 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2671 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2672 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2673 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2674 };
2675 
2676 /* Available operations from secondary process. */
2677 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2678 	.stats_get = mlx5_stats_get,
2679 	.stats_reset = mlx5_stats_reset,
2680 	.xstats_get = mlx5_xstats_get,
2681 	.xstats_reset = mlx5_xstats_reset,
2682 	.xstats_get_names = mlx5_xstats_get_names,
2683 	.fw_version_get = mlx5_fw_version_get,
2684 	.dev_infos_get = mlx5_dev_infos_get,
2685 	.read_clock = mlx5_txpp_read_clock,
2686 	.rx_queue_start = mlx5_rx_queue_start,
2687 	.rx_queue_stop = mlx5_rx_queue_stop,
2688 	.tx_queue_start = mlx5_tx_queue_start,
2689 	.tx_queue_stop = mlx5_tx_queue_stop,
2690 	.rxq_info_get = mlx5_rxq_info_get,
2691 	.txq_info_get = mlx5_txq_info_get,
2692 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2693 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2694 	.get_module_info = mlx5_get_module_info,
2695 	.get_module_eeprom = mlx5_get_module_eeprom,
2696 };
2697 
2698 /* Available operations in flow isolated mode. */
2699 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2700 	.dev_configure = mlx5_dev_configure,
2701 	.dev_start = mlx5_dev_start,
2702 	.dev_stop = mlx5_dev_stop,
2703 	.dev_set_link_down = mlx5_set_link_down,
2704 	.dev_set_link_up = mlx5_set_link_up,
2705 	.dev_close = mlx5_dev_close,
2706 	.promiscuous_enable = mlx5_promiscuous_enable,
2707 	.promiscuous_disable = mlx5_promiscuous_disable,
2708 	.allmulticast_enable = mlx5_allmulticast_enable,
2709 	.allmulticast_disable = mlx5_allmulticast_disable,
2710 	.link_update = mlx5_link_update,
2711 	.stats_get = mlx5_stats_get,
2712 	.stats_reset = mlx5_stats_reset,
2713 	.xstats_get = mlx5_xstats_get,
2714 	.xstats_reset = mlx5_xstats_reset,
2715 	.xstats_get_names = mlx5_xstats_get_names,
2716 	.fw_version_get = mlx5_fw_version_get,
2717 	.dev_infos_get = mlx5_dev_infos_get,
2718 	.read_clock = mlx5_txpp_read_clock,
2719 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2720 	.vlan_filter_set = mlx5_vlan_filter_set,
2721 	.rx_queue_setup = mlx5_rx_queue_setup,
2722 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2723 	.tx_queue_setup = mlx5_tx_queue_setup,
2724 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2725 	.rx_queue_release = mlx5_rx_queue_release,
2726 	.tx_queue_release = mlx5_tx_queue_release,
2727 	.rx_queue_start = mlx5_rx_queue_start,
2728 	.rx_queue_stop = mlx5_rx_queue_stop,
2729 	.tx_queue_start = mlx5_tx_queue_start,
2730 	.tx_queue_stop = mlx5_tx_queue_stop,
2731 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2732 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2733 	.mac_addr_remove = mlx5_mac_addr_remove,
2734 	.mac_addr_add = mlx5_mac_addr_add,
2735 	.mac_addr_set = mlx5_mac_addr_set,
2736 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2737 	.mtu_set = mlx5_dev_set_mtu,
2738 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2739 	.vlan_offload_set = mlx5_vlan_offload_set,
2740 	.filter_ctrl = mlx5_dev_filter_ctrl,
2741 	.rxq_info_get = mlx5_rxq_info_get,
2742 	.txq_info_get = mlx5_txq_info_get,
2743 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2744 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2745 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2746 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2747 	.is_removed = mlx5_is_removed,
2748 	.get_module_info = mlx5_get_module_info,
2749 	.get_module_eeprom = mlx5_get_module_eeprom,
2750 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2751 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2752 	.hairpin_bind = mlx5_hairpin_bind,
2753 	.hairpin_unbind = mlx5_hairpin_unbind,
2754 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2755 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2756 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2757 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2758 };
2759