xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision ecaee305b811036d297eda04ad7c91d8349de77e)
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 <ethdev_driver.h>
20 #include <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_rx.h"
44 #include "mlx5_tx.h"
45 #include "mlx5_autoconf.h"
46 #include "mlx5_mr.h"
47 #include "mlx5_flow.h"
48 #include "rte_pmd_mlx5.h"
49 #include "mlx5_verbs.h"
50 #include "mlx5_nl.h"
51 #include "mlx5_devx.h"
52 
53 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
54 
55 #ifndef HAVE_IBV_MLX5_MOD_MPW
56 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
57 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
58 #endif
59 
60 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
61 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
62 #endif
63 
64 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
65 
66 /* Spinlock for mlx5_shared_data allocation. */
67 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
68 
69 /* Process local data for secondary processes. */
70 static struct mlx5_local_data mlx5_local_data;
71 
72 /**
73  * Set the completion channel file descriptor interrupt as non-blocking.
74  *
75  * @param[in] rxq_obj
76  *   Pointer to RQ channel object, which includes the channel fd
77  *
78  * @param[out] fd
79  *   The file descriptor (representing the intetrrupt) used in this channel.
80  *
81  * @return
82  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
83  */
84 int
85 mlx5_os_set_nonblock_channel_fd(int fd)
86 {
87 	int flags;
88 
89 	flags = fcntl(fd, F_GETFL);
90 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
91 }
92 
93 /**
94  * Get mlx5 device attributes. The glue function query_device_ex() is called
95  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
96  * device attributes from the glue out parameter.
97  *
98  * @param dev
99  *   Pointer to ibv context.
100  *
101  * @param device_attr
102  *   Pointer to mlx5 device attributes.
103  *
104  * @return
105  *   0 on success, non zero error number otherwise
106  */
107 int
108 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
109 {
110 	int err;
111 	struct ibv_device_attr_ex attr_ex;
112 	memset(device_attr, 0, sizeof(*device_attr));
113 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
114 	if (err)
115 		return err;
116 
117 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
118 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
119 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
120 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
121 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
122 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
123 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
124 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
125 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
126 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
127 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
128 	device_attr->max_rwq_indirection_table_size =
129 		attr_ex.rss_caps.max_rwq_indirection_table_size;
130 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
131 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
132 
133 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
134 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
135 	if (err)
136 		return err;
137 
138 	device_attr->flags = dv_attr.flags;
139 	device_attr->comp_mask = dv_attr.comp_mask;
140 #ifdef HAVE_IBV_MLX5_MOD_SWP
141 	device_attr->sw_parsing_offloads =
142 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
143 #endif
144 	device_attr->min_single_stride_log_num_of_bytes =
145 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
146 	device_attr->max_single_stride_log_num_of_bytes =
147 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
148 	device_attr->min_single_wqe_log_num_of_strides =
149 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
150 	device_attr->max_single_wqe_log_num_of_strides =
151 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
152 	device_attr->stride_supported_qpts =
153 		dv_attr.striding_rq_caps.supported_qpts;
154 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
155 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
156 #endif
157 
158 	return err;
159 }
160 
161 /**
162  * Verbs callback to allocate a memory. This function should allocate the space
163  * according to the size provided residing inside a huge page.
164  * Please note that all allocation must respect the alignment from libmlx5
165  * (i.e. currently rte_mem_page_size()).
166  *
167  * @param[in] size
168  *   The size in bytes of the memory to allocate.
169  * @param[in] data
170  *   A pointer to the callback data.
171  *
172  * @return
173  *   Allocated buffer, NULL otherwise and rte_errno is set.
174  */
175 static void *
176 mlx5_alloc_verbs_buf(size_t size, void *data)
177 {
178 	struct mlx5_dev_ctx_shared *sh = data;
179 	void *ret;
180 	size_t alignment = rte_mem_page_size();
181 	if (alignment == (size_t)-1) {
182 		DRV_LOG(ERR, "Failed to get mem page size");
183 		rte_errno = ENOMEM;
184 		return NULL;
185 	}
186 
187 	MLX5_ASSERT(data != NULL);
188 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
189 	if (!ret && size)
190 		rte_errno = ENOMEM;
191 	return ret;
192 }
193 
194 /**
195  * Verbs callback to free a memory.
196  *
197  * @param[in] ptr
198  *   A pointer to the memory to free.
199  * @param[in] data
200  *   A pointer to the callback data.
201  */
202 static void
203 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
204 {
205 	MLX5_ASSERT(data != NULL);
206 	mlx5_free(ptr);
207 }
208 
209 /**
210  * Initialize DR related data within private structure.
211  * Routine checks the reference counter and does actual
212  * resources creation/initialization only if counter is zero.
213  *
214  * @param[in] priv
215  *   Pointer to the private device data structure.
216  *
217  * @return
218  *   Zero on success, positive error code otherwise.
219  */
220 static int
221 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
222 {
223 	struct mlx5_dev_ctx_shared *sh = priv->sh;
224 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
225 	int err;
226 
227 	MLX5_ASSERT(sh && sh->refcnt);
228 	if (sh->refcnt > 1)
229 		return 0;
230 	err = mlx5_alloc_table_hash_list(priv);
231 	if (err)
232 		goto error;
233 	/* The resources below are only valid with DV support. */
234 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
235 	/* Init port id action cache list. */
236 	snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
237 	mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
238 			     flow_dv_port_id_create_cb,
239 			     flow_dv_port_id_match_cb,
240 			     flow_dv_port_id_remove_cb);
241 	/* Init push vlan action cache list. */
242 	snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
243 	mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
244 			     flow_dv_push_vlan_create_cb,
245 			     flow_dv_push_vlan_match_cb,
246 			     flow_dv_push_vlan_remove_cb);
247 	/* Init sample action cache list. */
248 	snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
249 	mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
250 			     flow_dv_sample_create_cb,
251 			     flow_dv_sample_match_cb,
252 			     flow_dv_sample_remove_cb);
253 	/* Init dest array action cache list. */
254 	snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
255 	mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
256 			     flow_dv_dest_array_create_cb,
257 			     flow_dv_dest_array_match_cb,
258 			     flow_dv_dest_array_remove_cb);
259 	/* Create tags hash list table. */
260 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
261 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
262 					  MLX5_HLIST_WRITE_MOST,
263 					  flow_dv_tag_create_cb,
264 					  flow_dv_tag_match_cb,
265 					  flow_dv_tag_remove_cb);
266 	if (!sh->tag_table) {
267 		DRV_LOG(ERR, "tags with hash creation failed.");
268 		err = ENOMEM;
269 		goto error;
270 	}
271 	sh->tag_table->ctx = sh;
272 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
273 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
274 					    0, MLX5_HLIST_WRITE_MOST |
275 					    MLX5_HLIST_DIRECT_KEY,
276 					    flow_dv_modify_create_cb,
277 					    flow_dv_modify_match_cb,
278 					    flow_dv_modify_remove_cb);
279 	if (!sh->modify_cmds) {
280 		DRV_LOG(ERR, "hdr modify hash creation failed");
281 		err = ENOMEM;
282 		goto error;
283 	}
284 	sh->modify_cmds->ctx = sh;
285 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
286 	sh->encaps_decaps = mlx5_hlist_create(s,
287 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
288 					      0, MLX5_HLIST_DIRECT_KEY |
289 					      MLX5_HLIST_WRITE_MOST,
290 					      flow_dv_encap_decap_create_cb,
291 					      flow_dv_encap_decap_match_cb,
292 					      flow_dv_encap_decap_remove_cb);
293 	if (!sh->encaps_decaps) {
294 		DRV_LOG(ERR, "encap decap hash creation failed");
295 		err = ENOMEM;
296 		goto error;
297 	}
298 	sh->encaps_decaps->ctx = sh;
299 #endif
300 #ifdef HAVE_MLX5DV_DR
301 	void *domain;
302 
303 	/* Reference counter is zero, we should initialize structures. */
304 	domain = mlx5_glue->dr_create_domain(sh->ctx,
305 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
306 	if (!domain) {
307 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
308 		err = errno;
309 		goto error;
310 	}
311 	sh->rx_domain = domain;
312 	domain = mlx5_glue->dr_create_domain(sh->ctx,
313 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
314 	if (!domain) {
315 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
316 		err = errno;
317 		goto error;
318 	}
319 	sh->tx_domain = domain;
320 #ifdef HAVE_MLX5DV_DR_ESWITCH
321 	if (priv->config.dv_esw_en) {
322 		domain  = mlx5_glue->dr_create_domain
323 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
324 		if (!domain) {
325 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
326 			err = errno;
327 			goto error;
328 		}
329 		sh->fdb_domain = domain;
330 	}
331 	/*
332 	 * The drop action is just some dummy placeholder in rdma-core. It
333 	 * does not belong to domains and has no any attributes, and, can be
334 	 * shared by the entire device.
335 	 */
336 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
337 	if (!sh->dr_drop_action) {
338 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
339 		err = errno;
340 		goto error;
341 	}
342 #endif
343 	if (!sh->tunnel_hub)
344 		err = mlx5_alloc_tunnel_hub(sh);
345 	if (err) {
346 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
347 		goto error;
348 	}
349 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
350 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
351 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
352 		if (sh->fdb_domain)
353 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
354 	}
355 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
356 #endif /* HAVE_MLX5DV_DR */
357 	sh->default_miss_action =
358 			mlx5_glue->dr_create_flow_action_default_miss();
359 	if (!sh->default_miss_action)
360 		DRV_LOG(WARNING, "Default miss action is not supported.");
361 	return 0;
362 error:
363 	/* Rollback the created objects. */
364 	if (sh->rx_domain) {
365 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
366 		sh->rx_domain = NULL;
367 	}
368 	if (sh->tx_domain) {
369 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
370 		sh->tx_domain = NULL;
371 	}
372 	if (sh->fdb_domain) {
373 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
374 		sh->fdb_domain = NULL;
375 	}
376 	if (sh->dr_drop_action) {
377 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
378 		sh->dr_drop_action = NULL;
379 	}
380 	if (sh->pop_vlan_action) {
381 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
382 		sh->pop_vlan_action = NULL;
383 	}
384 	if (sh->encaps_decaps) {
385 		mlx5_hlist_destroy(sh->encaps_decaps);
386 		sh->encaps_decaps = NULL;
387 	}
388 	if (sh->modify_cmds) {
389 		mlx5_hlist_destroy(sh->modify_cmds);
390 		sh->modify_cmds = NULL;
391 	}
392 	if (sh->tag_table) {
393 		/* tags should be destroyed with flow before. */
394 		mlx5_hlist_destroy(sh->tag_table);
395 		sh->tag_table = NULL;
396 	}
397 	if (sh->tunnel_hub) {
398 		mlx5_release_tunnel_hub(sh, priv->dev_port);
399 		sh->tunnel_hub = NULL;
400 	}
401 	mlx5_free_table_hash_list(priv);
402 	return err;
403 }
404 
405 /**
406  * Destroy DR related data within private structure.
407  *
408  * @param[in] priv
409  *   Pointer to the private device data structure.
410  */
411 void
412 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
413 {
414 	struct mlx5_dev_ctx_shared *sh = priv->sh;
415 
416 	MLX5_ASSERT(sh && sh->refcnt);
417 	if (sh->refcnt > 1)
418 		return;
419 #ifdef HAVE_MLX5DV_DR
420 	if (sh->rx_domain) {
421 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
422 		sh->rx_domain = NULL;
423 	}
424 	if (sh->tx_domain) {
425 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
426 		sh->tx_domain = NULL;
427 	}
428 #ifdef HAVE_MLX5DV_DR_ESWITCH
429 	if (sh->fdb_domain) {
430 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
431 		sh->fdb_domain = NULL;
432 	}
433 	if (sh->dr_drop_action) {
434 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
435 		sh->dr_drop_action = NULL;
436 	}
437 #endif
438 	if (sh->pop_vlan_action) {
439 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
440 		sh->pop_vlan_action = NULL;
441 	}
442 #endif /* HAVE_MLX5DV_DR */
443 	if (sh->default_miss_action)
444 		mlx5_glue->destroy_flow_action
445 				(sh->default_miss_action);
446 	if (sh->encaps_decaps) {
447 		mlx5_hlist_destroy(sh->encaps_decaps);
448 		sh->encaps_decaps = NULL;
449 	}
450 	if (sh->modify_cmds) {
451 		mlx5_hlist_destroy(sh->modify_cmds);
452 		sh->modify_cmds = NULL;
453 	}
454 	if (sh->tag_table) {
455 		/* tags should be destroyed with flow before. */
456 		mlx5_hlist_destroy(sh->tag_table);
457 		sh->tag_table = NULL;
458 	}
459 	if (sh->tunnel_hub) {
460 		mlx5_release_tunnel_hub(sh, priv->dev_port);
461 		sh->tunnel_hub = NULL;
462 	}
463 	mlx5_cache_list_destroy(&sh->port_id_action_list);
464 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
465 	mlx5_free_table_hash_list(priv);
466 }
467 
468 /**
469  * Initialize shared data between primary and secondary process.
470  *
471  * A memzone is reserved by primary process and secondary processes attach to
472  * the memzone.
473  *
474  * @return
475  *   0 on success, a negative errno value otherwise and rte_errno is set.
476  */
477 static int
478 mlx5_init_shared_data(void)
479 {
480 	const struct rte_memzone *mz;
481 	int ret = 0;
482 
483 	rte_spinlock_lock(&mlx5_shared_data_lock);
484 	if (mlx5_shared_data == NULL) {
485 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
486 			/* Allocate shared memory. */
487 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
488 						 sizeof(*mlx5_shared_data),
489 						 SOCKET_ID_ANY, 0);
490 			if (mz == NULL) {
491 				DRV_LOG(ERR,
492 					"Cannot allocate mlx5 shared data");
493 				ret = -rte_errno;
494 				goto error;
495 			}
496 			mlx5_shared_data = mz->addr;
497 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
498 			rte_spinlock_init(&mlx5_shared_data->lock);
499 		} else {
500 			/* Lookup allocated shared memory. */
501 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
502 			if (mz == NULL) {
503 				DRV_LOG(ERR,
504 					"Cannot attach mlx5 shared data");
505 				ret = -rte_errno;
506 				goto error;
507 			}
508 			mlx5_shared_data = mz->addr;
509 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
510 		}
511 	}
512 error:
513 	rte_spinlock_unlock(&mlx5_shared_data_lock);
514 	return ret;
515 }
516 
517 /**
518  * PMD global initialization.
519  *
520  * Independent from individual device, this function initializes global
521  * per-PMD data structures distinguishing primary and secondary processes.
522  * Hence, each initialization is called once per a process.
523  *
524  * @return
525  *   0 on success, a negative errno value otherwise and rte_errno is set.
526  */
527 static int
528 mlx5_init_once(void)
529 {
530 	struct mlx5_shared_data *sd;
531 	struct mlx5_local_data *ld = &mlx5_local_data;
532 	int ret = 0;
533 
534 	if (mlx5_init_shared_data())
535 		return -rte_errno;
536 	sd = mlx5_shared_data;
537 	MLX5_ASSERT(sd);
538 	rte_spinlock_lock(&sd->lock);
539 	switch (rte_eal_process_type()) {
540 	case RTE_PROC_PRIMARY:
541 		if (sd->init_done)
542 			break;
543 		LIST_INIT(&sd->mem_event_cb_list);
544 		rte_rwlock_init(&sd->mem_event_rwlock);
545 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
546 						mlx5_mr_mem_event_cb, NULL);
547 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
548 					   mlx5_mp_os_primary_handle);
549 		if (ret)
550 			goto out;
551 		sd->init_done = true;
552 		break;
553 	case RTE_PROC_SECONDARY:
554 		if (ld->init_done)
555 			break;
556 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
557 					     mlx5_mp_os_secondary_handle);
558 		if (ret)
559 			goto out;
560 		++sd->secondary_cnt;
561 		ld->init_done = true;
562 		break;
563 	default:
564 		break;
565 	}
566 out:
567 	rte_spinlock_unlock(&sd->lock);
568 	return ret;
569 }
570 
571 /**
572  * Create the Tx queue DevX/Verbs object.
573  *
574  * @param dev
575  *   Pointer to Ethernet device.
576  * @param idx
577  *   Queue index in DPDK Tx queue array.
578  *
579  * @return
580  *   0 on success, a negative errno value otherwise and rte_errno is set.
581  */
582 static int
583 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
584 {
585 	struct mlx5_priv *priv = dev->data->dev_private;
586 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
587 	struct mlx5_txq_ctrl *txq_ctrl =
588 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
589 
590 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
591 		return mlx5_txq_devx_obj_new(dev, idx);
592 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
593 	if (!priv->config.dv_esw_en)
594 		return mlx5_txq_devx_obj_new(dev, idx);
595 #endif
596 	return mlx5_txq_ibv_obj_new(dev, idx);
597 }
598 
599 /**
600  * Release an Tx DevX/verbs queue object.
601  *
602  * @param txq_obj
603  *   DevX/Verbs Tx queue object.
604  */
605 static void
606 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
607 {
608 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
609 		mlx5_txq_devx_obj_release(txq_obj);
610 		return;
611 	}
612 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
613 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
614 		mlx5_txq_devx_obj_release(txq_obj);
615 		return;
616 	}
617 #endif
618 	mlx5_txq_ibv_obj_release(txq_obj);
619 }
620 
621 /**
622  * DV flow counter mode detect and config.
623  *
624  * @param dev
625  *   Pointer to rte_eth_dev structure.
626  *
627  */
628 static void
629 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
630 {
631 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
632 	struct mlx5_priv *priv = dev->data->dev_private;
633 	struct mlx5_dev_ctx_shared *sh = priv->sh;
634 	bool fallback;
635 
636 #ifndef HAVE_IBV_DEVX_ASYNC
637 	fallback = true;
638 #else
639 	fallback = false;
640 	if (!priv->config.devx || !priv->config.dv_flow_en ||
641 	    !priv->config.hca_attr.flow_counters_dump ||
642 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
643 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
644 		fallback = true;
645 #endif
646 	if (fallback)
647 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
648 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
649 			priv->config.hca_attr.flow_counters_dump,
650 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
651 	/* Initialize fallback mode only on the port initializes sh. */
652 	if (sh->refcnt == 1)
653 		sh->cmng.counter_fallback = fallback;
654 	else if (fallback != sh->cmng.counter_fallback)
655 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
656 			"with others:%d.", PORT_ID(priv), fallback);
657 #endif
658 }
659 
660 static void
661 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
662 {
663 	struct mlx5_priv *priv = dev->data->dev_private;
664 	void *ctx = priv->sh->ctx;
665 
666 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
667 	if (!priv->q_counters) {
668 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
669 		struct ibv_wq *wq;
670 
671 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
672 			"by DevX - fall-back to use the kernel driver global "
673 			"queue counter.", dev->data->port_id);
674 		/* Create WQ by kernel and query its queue counter ID. */
675 		if (cq) {
676 			wq = mlx5_glue->create_wq(ctx,
677 						  &(struct ibv_wq_init_attr){
678 						    .wq_type = IBV_WQT_RQ,
679 						    .max_wr = 1,
680 						    .max_sge = 1,
681 						    .pd = priv->sh->pd,
682 						    .cq = cq,
683 						});
684 			if (wq) {
685 				/* Counter is assigned only on RDY state. */
686 				int ret = mlx5_glue->modify_wq(wq,
687 						 &(struct ibv_wq_attr){
688 						 .attr_mask = IBV_WQ_ATTR_STATE,
689 						 .wq_state = IBV_WQS_RDY,
690 						});
691 
692 				if (ret == 0)
693 					mlx5_devx_cmd_wq_query(wq,
694 							 &priv->counter_set_id);
695 				claim_zero(mlx5_glue->destroy_wq(wq));
696 			}
697 			claim_zero(mlx5_glue->destroy_cq(cq));
698 		}
699 	} else {
700 		priv->counter_set_id = priv->q_counters->id;
701 	}
702 	if (priv->counter_set_id == 0)
703 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
704 			"available.", dev->data->port_id);
705 }
706 
707 /**
708  * Check if representor spawn info match devargs.
709  *
710  * @param spawn
711  *   Verbs device parameters (name, port, switch_info) to spawn.
712  * @param eth_da
713  *   Device devargs to probe.
714  *
715  * @return
716  *   Match result.
717  */
718 static bool
719 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
720 		       struct rte_eth_devargs *eth_da)
721 {
722 	struct mlx5_switch_info *switch_info = &spawn->info;
723 	unsigned int p, f;
724 	uint16_t id;
725 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
726 						      eth_da->type);
727 
728 	switch (eth_da->type) {
729 	case RTE_ETH_REPRESENTOR_SF:
730 		if (!(spawn->info.port_name == -1 &&
731 		      switch_info->name_type ==
732 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
733 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
734 			rte_errno = EBUSY;
735 			return false;
736 		}
737 		break;
738 	case RTE_ETH_REPRESENTOR_VF:
739 		/* Allows HPF representor index -1 as exception. */
740 		if (!(spawn->info.port_name == -1 &&
741 		      switch_info->name_type ==
742 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
743 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
744 			rte_errno = EBUSY;
745 			return false;
746 		}
747 		break;
748 	case RTE_ETH_REPRESENTOR_NONE:
749 		rte_errno = EBUSY;
750 		return false;
751 	default:
752 		rte_errno = ENOTSUP;
753 		DRV_LOG(ERR, "unsupported representor type");
754 		return false;
755 	}
756 	/* Check representor ID: */
757 	for (p = 0; p < eth_da->nb_ports; ++p) {
758 		if (spawn->pf_bond < 0) {
759 			/* For non-LAG mode, allow and ignore pf. */
760 			switch_info->pf_num = eth_da->ports[p];
761 			repr_id = mlx5_representor_id_encode(switch_info,
762 							     eth_da->type);
763 		}
764 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
765 			id = MLX5_REPRESENTOR_ID
766 				(eth_da->ports[p], eth_da->type,
767 				 eth_da->representor_ports[f]);
768 			if (repr_id == id)
769 				return true;
770 		}
771 	}
772 	rte_errno = EBUSY;
773 	return false;
774 }
775 
776 
777 /**
778  * Spawn an Ethernet device from Verbs information.
779  *
780  * @param dpdk_dev
781  *   Backing DPDK device.
782  * @param spawn
783  *   Verbs device parameters (name, port, switch_info) to spawn.
784  * @param config
785  *   Device configuration parameters.
786  * @param config
787  *   Device arguments.
788  *
789  * @return
790  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
791  *   is set. The following errors are defined:
792  *
793  *   EBUSY: device is not supposed to be spawned.
794  *   EEXIST: device is already spawned
795  */
796 static struct rte_eth_dev *
797 mlx5_dev_spawn(struct rte_device *dpdk_dev,
798 	       struct mlx5_dev_spawn_data *spawn,
799 	       struct mlx5_dev_config *config,
800 	       struct rte_eth_devargs *eth_da)
801 {
802 	const struct mlx5_switch_info *switch_info = &spawn->info;
803 	struct mlx5_dev_ctx_shared *sh = NULL;
804 	struct ibv_port_attr port_attr;
805 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
806 	struct rte_eth_dev *eth_dev = NULL;
807 	struct mlx5_priv *priv = NULL;
808 	int err = 0;
809 	unsigned int hw_padding = 0;
810 	unsigned int mps;
811 	unsigned int tunnel_en = 0;
812 	unsigned int mpls_en = 0;
813 	unsigned int swp = 0;
814 	unsigned int mprq = 0;
815 	unsigned int mprq_min_stride_size_n = 0;
816 	unsigned int mprq_max_stride_size_n = 0;
817 	unsigned int mprq_min_stride_num_n = 0;
818 	unsigned int mprq_max_stride_num_n = 0;
819 	struct rte_ether_addr mac;
820 	char name[RTE_ETH_NAME_MAX_LEN];
821 	int own_domain_id = 0;
822 	uint16_t port_id;
823 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
824 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
825 #endif
826 
827 	/* Determine if this port representor is supposed to be spawned. */
828 	if (switch_info->representor && dpdk_dev->devargs &&
829 	    !mlx5_representor_match(spawn, eth_da))
830 		return NULL;
831 	/* Build device name. */
832 	if (spawn->pf_bond < 0) {
833 		/* Single device. */
834 		if (!switch_info->representor)
835 			strlcpy(name, dpdk_dev->name, sizeof(name));
836 		else
837 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
838 				 dpdk_dev->name,
839 				 switch_info->name_type ==
840 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
841 				 switch_info->port_name);
842 	} else {
843 		/* Bonding device. */
844 		if (!switch_info->representor) {
845 			err = snprintf(name, sizeof(name), "%s_%s",
846 				 dpdk_dev->name,
847 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
848 		} else {
849 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
850 				dpdk_dev->name,
851 				mlx5_os_get_dev_device_name(spawn->phys_dev),
852 				switch_info->ctrl_num,
853 				switch_info->pf_num,
854 				switch_info->name_type ==
855 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
856 				switch_info->port_name);
857 		}
858 	}
859 	if (err >= (int)sizeof(name))
860 		DRV_LOG(WARNING, "device name overflow %s", name);
861 	/* check if the device is already spawned */
862 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
863 		rte_errno = EEXIST;
864 		return NULL;
865 	}
866 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
867 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
868 		struct mlx5_mp_id mp_id;
869 
870 		eth_dev = rte_eth_dev_attach_secondary(name);
871 		if (eth_dev == NULL) {
872 			DRV_LOG(ERR, "can not attach rte ethdev");
873 			rte_errno = ENOMEM;
874 			return NULL;
875 		}
876 		eth_dev->device = dpdk_dev;
877 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
878 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
879 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
880 		err = mlx5_proc_priv_init(eth_dev);
881 		if (err)
882 			return NULL;
883 		mp_id.port_id = eth_dev->data->port_id;
884 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
885 		/* Receive command fd from primary process */
886 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
887 		if (err < 0)
888 			goto err_secondary;
889 		/* Remap UAR for Tx queues. */
890 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
891 		if (err)
892 			goto err_secondary;
893 		/*
894 		 * Ethdev pointer is still required as input since
895 		 * the primary device is not accessible from the
896 		 * secondary process.
897 		 */
898 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
899 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
900 		return eth_dev;
901 err_secondary:
902 		mlx5_dev_close(eth_dev);
903 		return NULL;
904 	}
905 	/*
906 	 * Some parameters ("tx_db_nc" in particularly) are needed in
907 	 * advance to create dv/verbs device context. We proceed the
908 	 * devargs here to get ones, and later proceed devargs again
909 	 * to override some hardware settings.
910 	 */
911 	err = mlx5_args(config, dpdk_dev->devargs);
912 	if (err) {
913 		err = rte_errno;
914 		DRV_LOG(ERR, "failed to process device arguments: %s",
915 			strerror(rte_errno));
916 		goto error;
917 	}
918 	if (config->dv_miss_info) {
919 		if (switch_info->master || switch_info->representor)
920 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
921 	}
922 	mlx5_malloc_mem_select(config->sys_mem_en);
923 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
924 	if (!sh)
925 		return NULL;
926 	config->devx = sh->devx;
927 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
928 	config->dest_tir = 1;
929 #endif
930 #ifdef HAVE_IBV_MLX5_MOD_SWP
931 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
932 #endif
933 	/*
934 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
935 	 * as all ConnectX-5 devices.
936 	 */
937 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
938 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
939 #endif
940 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
941 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
942 #endif
943 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
944 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
945 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
946 			DRV_LOG(DEBUG, "enhanced MPW is supported");
947 			mps = MLX5_MPW_ENHANCED;
948 		} else {
949 			DRV_LOG(DEBUG, "MPW is supported");
950 			mps = MLX5_MPW;
951 		}
952 	} else {
953 		DRV_LOG(DEBUG, "MPW isn't supported");
954 		mps = MLX5_MPW_DISABLED;
955 	}
956 #ifdef HAVE_IBV_MLX5_MOD_SWP
957 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
958 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
959 	DRV_LOG(DEBUG, "SWP support: %u", swp);
960 #endif
961 	config->swp = !!swp;
962 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
963 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
964 		struct mlx5dv_striding_rq_caps mprq_caps =
965 			dv_attr.striding_rq_caps;
966 
967 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
968 			mprq_caps.min_single_stride_log_num_of_bytes);
969 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
970 			mprq_caps.max_single_stride_log_num_of_bytes);
971 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
972 			mprq_caps.min_single_wqe_log_num_of_strides);
973 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
974 			mprq_caps.max_single_wqe_log_num_of_strides);
975 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
976 			mprq_caps.supported_qpts);
977 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
978 		mprq = 1;
979 		mprq_min_stride_size_n =
980 			mprq_caps.min_single_stride_log_num_of_bytes;
981 		mprq_max_stride_size_n =
982 			mprq_caps.max_single_stride_log_num_of_bytes;
983 		mprq_min_stride_num_n =
984 			mprq_caps.min_single_wqe_log_num_of_strides;
985 		mprq_max_stride_num_n =
986 			mprq_caps.max_single_wqe_log_num_of_strides;
987 	}
988 #endif
989 	/* Rx CQE compression is enabled by default. */
990 	config->cqe_comp = 1;
991 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
992 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
993 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
994 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
995 			     (dv_attr.tunnel_offloads_caps &
996 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
997 			     (dv_attr.tunnel_offloads_caps &
998 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
999 	}
1000 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1001 		tunnel_en ? "" : "not ");
1002 #else
1003 	DRV_LOG(WARNING,
1004 		"tunnel offloading disabled due to old OFED/rdma-core version");
1005 #endif
1006 	config->tunnel_en = tunnel_en;
1007 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1008 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1009 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1010 		   (dv_attr.tunnel_offloads_caps &
1011 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1012 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1013 		mpls_en ? "" : "not ");
1014 #else
1015 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1016 		" old OFED/rdma-core version or firmware configuration");
1017 #endif
1018 	config->mpls_en = mpls_en;
1019 	/* Check port status. */
1020 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1021 	if (err) {
1022 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1023 		goto error;
1024 	}
1025 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1026 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1027 		err = EINVAL;
1028 		goto error;
1029 	}
1030 	if (port_attr.state != IBV_PORT_ACTIVE)
1031 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1032 			mlx5_glue->port_state_str(port_attr.state),
1033 			port_attr.state);
1034 	/* Allocate private eth device data. */
1035 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1036 			   sizeof(*priv),
1037 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1038 	if (priv == NULL) {
1039 		DRV_LOG(ERR, "priv allocation failure");
1040 		err = ENOMEM;
1041 		goto error;
1042 	}
1043 	priv->sh = sh;
1044 	priv->dev_port = spawn->phys_port;
1045 	priv->pci_dev = spawn->pci_dev;
1046 	priv->mtu = RTE_ETHER_MTU;
1047 	/* Some internal functions rely on Netlink sockets, open them now. */
1048 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1049 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1050 	priv->representor = !!switch_info->representor;
1051 	priv->master = !!switch_info->master;
1052 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1053 	priv->vport_meta_tag = 0;
1054 	priv->vport_meta_mask = 0;
1055 	priv->pf_bond = spawn->pf_bond;
1056 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1057 	/*
1058 	 * The DevX port query API is implemented. E-Switch may use
1059 	 * either vport or reg_c[0] metadata register to match on
1060 	 * vport index. The engaged part of metadata register is
1061 	 * defined by mask.
1062 	 */
1063 	if (switch_info->representor || switch_info->master) {
1064 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1065 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1066 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1067 						 &devx_port);
1068 		if (err) {
1069 			DRV_LOG(WARNING,
1070 				"can't query devx port %d on device %s",
1071 				spawn->phys_port,
1072 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1073 			devx_port.comp_mask = 0;
1074 		}
1075 	}
1076 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1077 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1078 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1079 		if (!priv->vport_meta_mask) {
1080 			DRV_LOG(ERR, "vport zero mask for port %d"
1081 				     " on bonding device %s",
1082 				     spawn->phys_port,
1083 				     mlx5_os_get_dev_device_name
1084 							(spawn->phys_dev));
1085 			err = ENOTSUP;
1086 			goto error;
1087 		}
1088 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1089 			DRV_LOG(ERR, "invalid vport tag for port %d"
1090 				     " on bonding device %s",
1091 				     spawn->phys_port,
1092 				     mlx5_os_get_dev_device_name
1093 							(spawn->phys_dev));
1094 			err = ENOTSUP;
1095 			goto error;
1096 		}
1097 	}
1098 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1099 		priv->vport_id = devx_port.vport_num;
1100 	} else if (spawn->pf_bond >= 0 &&
1101 		   (switch_info->representor || switch_info->master)) {
1102 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1103 			     " on bonding device %s",
1104 			     spawn->phys_port,
1105 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1106 		err = ENOTSUP;
1107 		goto error;
1108 	} else {
1109 		/* Suppose vport index in compatible way. */
1110 		priv->vport_id = switch_info->representor ?
1111 				 switch_info->port_name + 1 : -1;
1112 	}
1113 #else
1114 	/*
1115 	 * Kernel/rdma_core support single E-Switch per PF configurations
1116 	 * only and vport_id field contains the vport index for
1117 	 * associated VF, which is deduced from representor port name.
1118 	 * For example, let's have the IB device port 10, it has
1119 	 * attached network device eth0, which has port name attribute
1120 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1121 	 * as 3 (2+1). This assigning schema should be changed if the
1122 	 * multiple E-Switch instances per PF configurations or/and PCI
1123 	 * subfunctions are added.
1124 	 */
1125 	priv->vport_id = switch_info->representor ?
1126 			 switch_info->port_name + 1 : -1;
1127 #endif
1128 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1129 							  eth_da->type);
1130 	/*
1131 	 * Look for sibling devices in order to reuse their switch domain
1132 	 * if any, otherwise allocate one.
1133 	 */
1134 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1135 		const struct mlx5_priv *opriv =
1136 			rte_eth_devices[port_id].data->dev_private;
1137 
1138 		if (!opriv ||
1139 		    opriv->sh != priv->sh ||
1140 			opriv->domain_id ==
1141 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1142 			continue;
1143 		priv->domain_id = opriv->domain_id;
1144 		break;
1145 	}
1146 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1147 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1148 		if (err) {
1149 			err = rte_errno;
1150 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1151 				strerror(rte_errno));
1152 			goto error;
1153 		}
1154 		own_domain_id = 1;
1155 	}
1156 	/* Override some values set by hardware configuration. */
1157 	mlx5_args(config, dpdk_dev->devargs);
1158 	err = mlx5_dev_check_sibling_config(priv, config);
1159 	if (err)
1160 		goto error;
1161 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1162 			    IBV_DEVICE_RAW_IP_CSUM);
1163 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1164 		(config->hw_csum ? "" : "not "));
1165 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1166 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1167 	DRV_LOG(DEBUG, "counters are not supported");
1168 #endif
1169 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1170 	if (config->dv_flow_en) {
1171 		DRV_LOG(WARNING, "DV flow is not supported");
1172 		config->dv_flow_en = 0;
1173 	}
1174 #endif
1175 	config->ind_table_max_size =
1176 		sh->device_attr.max_rwq_indirection_table_size;
1177 	/*
1178 	 * Remove this check once DPDK supports larger/variable
1179 	 * indirection tables.
1180 	 */
1181 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1182 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1183 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1184 		config->ind_table_max_size);
1185 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1186 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1187 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1188 		(config->hw_vlan_strip ? "" : "not "));
1189 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1190 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1191 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1192 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1193 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1194 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1195 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1196 #endif
1197 	if (config->hw_padding && !hw_padding) {
1198 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1199 		config->hw_padding = 0;
1200 	} else if (config->hw_padding) {
1201 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1202 	}
1203 	config->tso = (sh->device_attr.max_tso > 0 &&
1204 		      (sh->device_attr.tso_supported_qpts &
1205 		       (1 << IBV_QPT_RAW_PACKET)));
1206 	if (config->tso)
1207 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1208 	/*
1209 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1210 	 * by default.
1211 	 */
1212 	if (config->mps == MLX5_ARG_UNSET)
1213 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1214 							  MLX5_MPW_DISABLED;
1215 	else
1216 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1217 	DRV_LOG(INFO, "%sMPS is %s",
1218 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1219 		config->mps == MLX5_MPW ? "legacy " : "",
1220 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1221 	if (config->devx) {
1222 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1223 		if (err) {
1224 			err = -err;
1225 			goto error;
1226 		}
1227 		/* Check relax ordering support. */
1228 		if (!haswell_broadwell_cpu) {
1229 			sh->cmng.relaxed_ordering_write =
1230 				config->hca_attr.relaxed_ordering_write;
1231 			sh->cmng.relaxed_ordering_read =
1232 				config->hca_attr.relaxed_ordering_read;
1233 		} else {
1234 			sh->cmng.relaxed_ordering_read = 0;
1235 			sh->cmng.relaxed_ordering_write = 0;
1236 		}
1237 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1238 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1239 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1240 		/* Check for LRO support. */
1241 		if (config->dest_tir && config->hca_attr.lro_cap &&
1242 		    config->dv_flow_en) {
1243 			/* TBD check tunnel lro caps. */
1244 			config->lro.supported = config->hca_attr.lro_cap;
1245 			DRV_LOG(DEBUG, "Device supports LRO");
1246 			/*
1247 			 * If LRO timeout is not configured by application,
1248 			 * use the minimal supported value.
1249 			 */
1250 			if (!config->lro.timeout)
1251 				config->lro.timeout =
1252 				config->hca_attr.lro_timer_supported_periods[0];
1253 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1254 				config->lro.timeout);
1255 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1256 				"required for coalescing is %d bytes",
1257 				config->hca_attr.lro_min_mss_size);
1258 		}
1259 #if defined(HAVE_MLX5DV_DR) && \
1260 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1261 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1262 		if (config->hca_attr.qos.sup &&
1263 		    config->hca_attr.qos.flow_meter_old &&
1264 		    config->dv_flow_en) {
1265 			uint8_t reg_c_mask =
1266 				config->hca_attr.qos.flow_meter_reg_c_ids;
1267 			/*
1268 			 * Meter needs two REG_C's for color match and pre-sfx
1269 			 * flow match. Here get the REG_C for color match.
1270 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1271 			 */
1272 			reg_c_mask &= 0xfc;
1273 			if (__builtin_popcount(reg_c_mask) < 1) {
1274 				priv->mtr_en = 0;
1275 				DRV_LOG(WARNING, "No available register for"
1276 					" meter.");
1277 			} else {
1278 				/*
1279 				 * The meter color register is used by the
1280 				 * flow-hit feature as well.
1281 				 * The flow-hit feature must use REG_C_3
1282 				 * Prefer REG_C_3 if it is available.
1283 				 */
1284 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1285 					priv->mtr_color_reg = REG_C_3;
1286 				else
1287 					priv->mtr_color_reg = ffs(reg_c_mask)
1288 							      - 1 + REG_C_0;
1289 				priv->mtr_en = 1;
1290 				priv->mtr_reg_share =
1291 				      config->hca_attr.qos.flow_meter;
1292 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1293 					priv->mtr_color_reg);
1294 			}
1295 		}
1296 		if (config->hca_attr.qos.sup &&
1297 			config->hca_attr.qos.flow_meter_aso_sup) {
1298 			uint32_t log_obj_size =
1299 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1300 			if (log_obj_size >=
1301 			config->hca_attr.qos.log_meter_aso_granularity &&
1302 			log_obj_size <=
1303 			config->hca_attr.qos.log_meter_aso_max_alloc)
1304 				sh->meter_aso_en = 1;
1305 		}
1306 		if (priv->mtr_en) {
1307 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1308 			if (err) {
1309 				err = -err;
1310 				goto error;
1311 			}
1312 		}
1313 #endif
1314 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1315 		if (config->hca_attr.flow_hit_aso &&
1316 		    priv->mtr_color_reg == REG_C_3) {
1317 			sh->flow_hit_aso_en = 1;
1318 			err = mlx5_flow_aso_age_mng_init(sh);
1319 			if (err) {
1320 				err = -err;
1321 				goto error;
1322 			}
1323 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1324 		}
1325 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1326 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1327 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1328 		    config->dv_flow_en) {
1329 			priv->sampler_en = 1;
1330 			DRV_LOG(DEBUG, "Sampler enabled!");
1331 		} else {
1332 			priv->sampler_en = 0;
1333 			if (!config->hca_attr.log_max_ft_sampler_num)
1334 				DRV_LOG(WARNING,
1335 					"No available register for sampler.");
1336 			else
1337 				DRV_LOG(DEBUG, "DV flow is not supported!");
1338 		}
1339 #endif
1340 	}
1341 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1342 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1343 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1344 		config->cqe_comp = 0;
1345 	}
1346 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1347 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1348 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1349 				 " format isn't supported.");
1350 		config->cqe_comp = 0;
1351 	}
1352 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1353 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1354 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1355 				 " format isn't supported.");
1356 		config->cqe_comp = 0;
1357 	}
1358 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1359 			config->cqe_comp ? "" : "not ");
1360 	if (config->tx_pp) {
1361 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1362 			config->hca_attr.dev_freq_khz);
1363 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1364 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1365 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1366 			config->hca_attr.cross_channel ? "" : "not ");
1367 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1368 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1369 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1370 			config->hca_attr.non_wire_sq ? "" : "not ");
1371 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1372 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1373 			config->hca_attr.log_max_static_sq_wq);
1374 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1375 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1376 		if (!config->devx) {
1377 			DRV_LOG(ERR, "DevX is required for packet pacing");
1378 			err = ENODEV;
1379 			goto error;
1380 		}
1381 		if (!config->hca_attr.qos.packet_pacing) {
1382 			DRV_LOG(ERR, "Packet pacing is not supported");
1383 			err = ENODEV;
1384 			goto error;
1385 		}
1386 		if (!config->hca_attr.cross_channel) {
1387 			DRV_LOG(ERR, "Cross channel operations are"
1388 				     " required for packet pacing");
1389 			err = ENODEV;
1390 			goto error;
1391 		}
1392 		if (!config->hca_attr.wqe_index_ignore) {
1393 			DRV_LOG(ERR, "WQE index ignore feature is"
1394 				     " required for packet pacing");
1395 			err = ENODEV;
1396 			goto error;
1397 		}
1398 		if (!config->hca_attr.non_wire_sq) {
1399 			DRV_LOG(ERR, "Non-wire SQ feature is"
1400 				     " required for packet pacing");
1401 			err = ENODEV;
1402 			goto error;
1403 		}
1404 		if (!config->hca_attr.log_max_static_sq_wq) {
1405 			DRV_LOG(ERR, "Static WQE SQ feature is"
1406 				     " required for packet pacing");
1407 			err = ENODEV;
1408 			goto error;
1409 		}
1410 		if (!config->hca_attr.qos.wqe_rate_pp) {
1411 			DRV_LOG(ERR, "WQE rate mode is required"
1412 				     " for packet pacing");
1413 			err = ENODEV;
1414 			goto error;
1415 		}
1416 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1417 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1418 			     " can't create queues for packet pacing");
1419 		err = ENODEV;
1420 		goto error;
1421 #endif
1422 	}
1423 	if (config->devx) {
1424 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1425 
1426 		err = config->hca_attr.access_register_user ?
1427 			mlx5_devx_cmd_register_read
1428 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1429 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1430 		if (!err) {
1431 			uint32_t ts_mode;
1432 
1433 			/* MTUTC register is read successfully. */
1434 			ts_mode = MLX5_GET(register_mtutc, reg,
1435 					   time_stamp_mode);
1436 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1437 				config->rt_timestamp = 1;
1438 		} else {
1439 			/* Kernel does not support register reading. */
1440 			if (config->hca_attr.dev_freq_khz ==
1441 						 (NS_PER_S / MS_PER_S))
1442 				config->rt_timestamp = 1;
1443 		}
1444 	}
1445 	/*
1446 	 * If HW has bug working with tunnel packet decapsulation and
1447 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1448 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1449 	 */
1450 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1451 		config->hw_fcs_strip = 0;
1452 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1453 		(config->hw_fcs_strip ? "" : "not "));
1454 	if (config->mprq.enabled && mprq) {
1455 		if (config->mprq.stride_num_n &&
1456 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1457 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1458 			config->mprq.stride_num_n =
1459 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1460 						mprq_min_stride_num_n),
1461 					mprq_max_stride_num_n);
1462 			DRV_LOG(WARNING,
1463 				"the number of strides"
1464 				" for Multi-Packet RQ is out of range,"
1465 				" setting default value (%u)",
1466 				1 << config->mprq.stride_num_n);
1467 		}
1468 		if (config->mprq.stride_size_n &&
1469 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1470 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1471 			config->mprq.stride_size_n =
1472 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1473 						mprq_min_stride_size_n),
1474 					mprq_max_stride_size_n);
1475 			DRV_LOG(WARNING,
1476 				"the size of a stride"
1477 				" for Multi-Packet RQ is out of range,"
1478 				" setting default value (%u)",
1479 				1 << config->mprq.stride_size_n);
1480 		}
1481 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1482 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1483 	} else if (config->mprq.enabled && !mprq) {
1484 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1485 		config->mprq.enabled = 0;
1486 	}
1487 	if (config->max_dump_files_num == 0)
1488 		config->max_dump_files_num = 128;
1489 	eth_dev = rte_eth_dev_allocate(name);
1490 	if (eth_dev == NULL) {
1491 		DRV_LOG(ERR, "can not allocate rte ethdev");
1492 		err = ENOMEM;
1493 		goto error;
1494 	}
1495 	if (priv->representor) {
1496 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1497 		eth_dev->data->representor_id = priv->representor_id;
1498 	}
1499 	priv->mp_id.port_id = eth_dev->data->port_id;
1500 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1501 	/*
1502 	 * Store associated network device interface index. This index
1503 	 * is permanent throughout the lifetime of device. So, we may store
1504 	 * the ifindex here and use the cached value further.
1505 	 */
1506 	MLX5_ASSERT(spawn->ifindex);
1507 	priv->if_index = spawn->ifindex;
1508 	eth_dev->data->dev_private = priv;
1509 	priv->dev_data = eth_dev->data;
1510 	eth_dev->data->mac_addrs = priv->mac;
1511 	eth_dev->device = dpdk_dev;
1512 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1513 	/* Configure the first MAC address by default. */
1514 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1515 		DRV_LOG(ERR,
1516 			"port %u cannot get MAC address, is mlx5_en"
1517 			" loaded? (errno: %s)",
1518 			eth_dev->data->port_id, strerror(rte_errno));
1519 		err = ENODEV;
1520 		goto error;
1521 	}
1522 	DRV_LOG(INFO,
1523 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1524 		eth_dev->data->port_id,
1525 		mac.addr_bytes[0], mac.addr_bytes[1],
1526 		mac.addr_bytes[2], mac.addr_bytes[3],
1527 		mac.addr_bytes[4], mac.addr_bytes[5]);
1528 #ifdef RTE_LIBRTE_MLX5_DEBUG
1529 	{
1530 		char ifname[MLX5_NAMESIZE];
1531 
1532 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1533 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1534 				eth_dev->data->port_id, ifname);
1535 		else
1536 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1537 				eth_dev->data->port_id);
1538 	}
1539 #endif
1540 	/* Get actual MTU if possible. */
1541 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1542 	if (err) {
1543 		err = rte_errno;
1544 		goto error;
1545 	}
1546 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1547 		priv->mtu);
1548 	/* Initialize burst functions to prevent crashes before link-up. */
1549 	eth_dev->rx_pkt_burst = removed_rx_burst;
1550 	eth_dev->tx_pkt_burst = removed_tx_burst;
1551 	eth_dev->dev_ops = &mlx5_dev_ops;
1552 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1553 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1554 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1555 	/* Register MAC address. */
1556 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1557 	if (config->vf && config->vf_nl_en)
1558 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1559 				      mlx5_ifindex(eth_dev),
1560 				      eth_dev->data->mac_addrs,
1561 				      MLX5_MAX_MAC_ADDRESSES);
1562 	priv->flows = 0;
1563 	priv->ctrl_flows = 0;
1564 	rte_spinlock_init(&priv->flow_list_lock);
1565 	TAILQ_INIT(&priv->flow_meters);
1566 	TAILQ_INIT(&priv->flow_meter_profiles);
1567 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1568 	mlx5_glue->dv_set_context_attr(sh->ctx,
1569 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1570 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1571 				.alloc = &mlx5_alloc_verbs_buf,
1572 				.free = &mlx5_free_verbs_buf,
1573 				.data = sh,
1574 			}));
1575 	/* Bring Ethernet device up. */
1576 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1577 		eth_dev->data->port_id);
1578 	mlx5_set_link_up(eth_dev);
1579 	/*
1580 	 * Even though the interrupt handler is not installed yet,
1581 	 * interrupts will still trigger on the async_fd from
1582 	 * Verbs context returned by ibv_open_device().
1583 	 */
1584 	mlx5_link_update(eth_dev, 0);
1585 #ifdef HAVE_MLX5DV_DR_ESWITCH
1586 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1587 	      (switch_info->representor || switch_info->master)))
1588 		config->dv_esw_en = 0;
1589 #else
1590 	config->dv_esw_en = 0;
1591 #endif
1592 	/* Detect minimal data bytes to inline. */
1593 	mlx5_set_min_inline(spawn, config);
1594 	/* Store device configuration on private structure. */
1595 	priv->config = *config;
1596 	/* Create context for virtual machine VLAN workaround. */
1597 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1598 	if (config->dv_flow_en) {
1599 		err = mlx5_alloc_shared_dr(priv);
1600 		if (err)
1601 			goto error;
1602 	}
1603 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1604 		priv->obj_ops = devx_obj_ops;
1605 		priv->obj_ops.drop_action_create =
1606 						ibv_obj_ops.drop_action_create;
1607 		priv->obj_ops.drop_action_destroy =
1608 						ibv_obj_ops.drop_action_destroy;
1609 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1610 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1611 #else
1612 		if (config->dv_esw_en)
1613 			priv->obj_ops.txq_obj_modify =
1614 						ibv_obj_ops.txq_obj_modify;
1615 #endif
1616 		/* Use specific wrappers for Tx object. */
1617 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1618 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1619 		mlx5_queue_counter_id_prepare(eth_dev);
1620 
1621 	} else {
1622 		priv->obj_ops = ibv_obj_ops;
1623 	}
1624 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1625 	if (!priv->drop_queue.hrxq)
1626 		goto error;
1627 	/* Supported Verbs flow priority number detection. */
1628 	err = mlx5_flow_discover_priorities(eth_dev);
1629 	if (err < 0) {
1630 		err = -err;
1631 		goto error;
1632 	}
1633 	priv->config.flow_prio = err;
1634 	if (!priv->config.dv_esw_en &&
1635 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1636 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1637 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1638 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1639 	}
1640 	mlx5_set_metadata_mask(eth_dev);
1641 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1642 	    !priv->sh->dv_regc0_mask) {
1643 		DRV_LOG(ERR, "metadata mode %u is not supported "
1644 			     "(no metadata reg_c[0] is available)",
1645 			     priv->config.dv_xmeta_en);
1646 			err = ENOTSUP;
1647 			goto error;
1648 	}
1649 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1650 			     mlx5_hrxq_create_cb,
1651 			     mlx5_hrxq_match_cb,
1652 			     mlx5_hrxq_remove_cb);
1653 	/* Query availability of metadata reg_c's. */
1654 	err = mlx5_flow_discover_mreg_c(eth_dev);
1655 	if (err < 0) {
1656 		err = -err;
1657 		goto error;
1658 	}
1659 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1660 		DRV_LOG(DEBUG,
1661 			"port %u extensive metadata register is not supported",
1662 			eth_dev->data->port_id);
1663 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1664 			DRV_LOG(ERR, "metadata mode %u is not supported "
1665 				     "(no metadata registers available)",
1666 				     priv->config.dv_xmeta_en);
1667 			err = ENOTSUP;
1668 			goto error;
1669 		}
1670 	}
1671 	if (priv->config.dv_flow_en &&
1672 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1673 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1674 	    priv->sh->dv_regc0_mask) {
1675 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1676 						      MLX5_FLOW_MREG_HTABLE_SZ,
1677 						      0, 0,
1678 						      flow_dv_mreg_create_cb,
1679 						      flow_dv_mreg_match_cb,
1680 						      flow_dv_mreg_remove_cb);
1681 		if (!priv->mreg_cp_tbl) {
1682 			err = ENOMEM;
1683 			goto error;
1684 		}
1685 		priv->mreg_cp_tbl->ctx = eth_dev;
1686 	}
1687 	rte_spinlock_init(&priv->shared_act_sl);
1688 	mlx5_flow_counter_mode_config(eth_dev);
1689 	if (priv->config.dv_flow_en)
1690 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1691 	return eth_dev;
1692 error:
1693 	if (priv) {
1694 		if (priv->mreg_cp_tbl)
1695 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1696 		if (priv->sh)
1697 			mlx5_os_free_shared_dr(priv);
1698 		if (priv->nl_socket_route >= 0)
1699 			close(priv->nl_socket_route);
1700 		if (priv->nl_socket_rdma >= 0)
1701 			close(priv->nl_socket_rdma);
1702 		if (priv->vmwa_context)
1703 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1704 		if (eth_dev && priv->drop_queue.hrxq)
1705 			mlx5_drop_action_destroy(eth_dev);
1706 		if (own_domain_id)
1707 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1708 		mlx5_cache_list_destroy(&priv->hrxqs);
1709 		mlx5_free(priv);
1710 		if (eth_dev != NULL)
1711 			eth_dev->data->dev_private = NULL;
1712 	}
1713 	if (eth_dev != NULL) {
1714 		/* mac_addrs must not be freed alone because part of
1715 		 * dev_private
1716 		 **/
1717 		eth_dev->data->mac_addrs = NULL;
1718 		rte_eth_dev_release_port(eth_dev);
1719 	}
1720 	if (sh)
1721 		mlx5_free_shared_dev_ctx(sh);
1722 	MLX5_ASSERT(err > 0);
1723 	rte_errno = err;
1724 	return NULL;
1725 }
1726 
1727 /**
1728  * Comparison callback to sort device data.
1729  *
1730  * This is meant to be used with qsort().
1731  *
1732  * @param a[in]
1733  *   Pointer to pointer to first data object.
1734  * @param b[in]
1735  *   Pointer to pointer to second data object.
1736  *
1737  * @return
1738  *   0 if both objects are equal, less than 0 if the first argument is less
1739  *   than the second, greater than 0 otherwise.
1740  */
1741 static int
1742 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1743 {
1744 	const struct mlx5_switch_info *si_a =
1745 		&((const struct mlx5_dev_spawn_data *)a)->info;
1746 	const struct mlx5_switch_info *si_b =
1747 		&((const struct mlx5_dev_spawn_data *)b)->info;
1748 	int ret;
1749 
1750 	/* Master device first. */
1751 	ret = si_b->master - si_a->master;
1752 	if (ret)
1753 		return ret;
1754 	/* Then representor devices. */
1755 	ret = si_b->representor - si_a->representor;
1756 	if (ret)
1757 		return ret;
1758 	/* Unidentified devices come last in no specific order. */
1759 	if (!si_a->representor)
1760 		return 0;
1761 	/* Order representors by name. */
1762 	return si_a->port_name - si_b->port_name;
1763 }
1764 
1765 /**
1766  * Match PCI information for possible slaves of bonding device.
1767  *
1768  * @param[in] ibv_dev
1769  *   Pointer to Infiniband device structure.
1770  * @param[in] pci_dev
1771  *   Pointer to primary PCI address structure to match.
1772  * @param[in] nl_rdma
1773  *   Netlink RDMA group socket handle.
1774  * @param[in] owner
1775  *   Rerepsentor owner PF index.
1776  * @param[out] bond_info
1777  *   Pointer to bonding information.
1778  *
1779  * @return
1780  *   negative value if no bonding device found, otherwise
1781  *   positive index of slave PF in bonding.
1782  */
1783 static int
1784 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1785 			   const struct rte_pci_addr *pci_dev,
1786 			   int nl_rdma, uint16_t owner,
1787 			   struct mlx5_bond_info *bond_info)
1788 {
1789 	char ifname[IF_NAMESIZE + 1];
1790 	unsigned int ifindex;
1791 	unsigned int np, i;
1792 	FILE *bond_file = NULL, *file;
1793 	int pf = -1;
1794 	int ret;
1795 
1796 	/*
1797 	 * Try to get master device name. If something goes
1798 	 * wrong suppose the lack of kernel support and no
1799 	 * bonding devices.
1800 	 */
1801 	memset(bond_info, 0, sizeof(*bond_info));
1802 	if (nl_rdma < 0)
1803 		return -1;
1804 	if (!strstr(ibv_dev->name, "bond"))
1805 		return -1;
1806 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1807 	if (!np)
1808 		return -1;
1809 	/*
1810 	 * The Master device might not be on the predefined
1811 	 * port (not on port index 1, it is not garanted),
1812 	 * we have to scan all Infiniband device port and
1813 	 * find master.
1814 	 */
1815 	for (i = 1; i <= np; ++i) {
1816 		/* Check whether Infiniband port is populated. */
1817 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1818 		if (!ifindex)
1819 			continue;
1820 		if (!if_indextoname(ifindex, ifname))
1821 			continue;
1822 		/* Try to read bonding slave names from sysfs. */
1823 		MKSTR(slaves,
1824 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1825 		bond_file = fopen(slaves, "r");
1826 		if (bond_file)
1827 			break;
1828 	}
1829 	if (!bond_file)
1830 		return -1;
1831 	/* Use safe format to check maximal buffer length. */
1832 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1833 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1834 		char tmp_str[IF_NAMESIZE + 32];
1835 		struct rte_pci_addr pci_addr;
1836 		struct mlx5_switch_info	info;
1837 
1838 		/* Process slave interface names in the loop. */
1839 		snprintf(tmp_str, sizeof(tmp_str),
1840 			 "/sys/class/net/%s", ifname);
1841 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1842 			DRV_LOG(WARNING, "can not get PCI address"
1843 					 " for netdev \"%s\"", ifname);
1844 			continue;
1845 		}
1846 		/* Slave interface PCI address match found. */
1847 		snprintf(tmp_str, sizeof(tmp_str),
1848 			 "/sys/class/net/%s/phys_port_name", ifname);
1849 		file = fopen(tmp_str, "rb");
1850 		if (!file)
1851 			break;
1852 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1853 		if (fscanf(file, "%32s", tmp_str) == 1)
1854 			mlx5_translate_port_name(tmp_str, &info);
1855 		fclose(file);
1856 		/* Only process PF ports. */
1857 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1858 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1859 			continue;
1860 		/* Check max bonding member. */
1861 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1862 			DRV_LOG(WARNING, "bonding index out of range, "
1863 				"please increase MLX5_BOND_MAX_PORTS: %s",
1864 				tmp_str);
1865 			break;
1866 		}
1867 		/* Match PCI address. */
1868 		if (pci_dev->domain == pci_addr.domain &&
1869 		    pci_dev->bus == pci_addr.bus &&
1870 		    pci_dev->devid == pci_addr.devid &&
1871 		    pci_dev->function + owner == pci_addr.function)
1872 			pf = info.port_name;
1873 		/* Get ifindex. */
1874 		snprintf(tmp_str, sizeof(tmp_str),
1875 			 "/sys/class/net/%s/ifindex", ifname);
1876 		file = fopen(tmp_str, "rb");
1877 		if (!file)
1878 			break;
1879 		ret = fscanf(file, "%u", &ifindex);
1880 		fclose(file);
1881 		if (ret != 1)
1882 			break;
1883 		/* Save bonding info. */
1884 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1885 			sizeof(bond_info->ports[0].ifname));
1886 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1887 		bond_info->ports[info.port_name].ifindex = ifindex;
1888 		bond_info->n_port++;
1889 	}
1890 	if (pf >= 0) {
1891 		/* Get bond interface info */
1892 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1893 					   bond_info->ifname);
1894 		if (ret)
1895 			DRV_LOG(ERR, "unable to get bond info: %s",
1896 				strerror(rte_errno));
1897 		else
1898 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1899 				ifindex, bond_info->ifindex, bond_info->ifname);
1900 	}
1901 	return pf;
1902 }
1903 
1904 /**
1905  * Register a PCI device within bonding.
1906  *
1907  * This function spawns Ethernet devices out of a given PCI device and
1908  * bonding owner PF index.
1909  *
1910  * @param[in] pci_dev
1911  *   PCI device information.
1912  * @param[in] req_eth_da
1913  *   Requested ethdev device argument.
1914  * @param[in] owner_id
1915  *   Requested owner PF port ID within bonding device, default to 0.
1916  *
1917  * @return
1918  *   0 on success, a negative errno value otherwise and rte_errno is set.
1919  */
1920 static int
1921 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1922 		     struct rte_eth_devargs *req_eth_da,
1923 		     uint16_t owner_id)
1924 {
1925 	struct ibv_device **ibv_list;
1926 	/*
1927 	 * Number of found IB Devices matching with requested PCI BDF.
1928 	 * nd != 1 means there are multiple IB devices over the same
1929 	 * PCI device and we have representors and master.
1930 	 */
1931 	unsigned int nd = 0;
1932 	/*
1933 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1934 	 * we have the single multiport IB device, and there may be
1935 	 * representors attached to some of found ports.
1936 	 */
1937 	unsigned int np = 0;
1938 	/*
1939 	 * Number of DPDK ethernet devices to Spawn - either over
1940 	 * multiple IB devices or multiple ports of single IB device.
1941 	 * Actually this is the number of iterations to spawn.
1942 	 */
1943 	unsigned int ns = 0;
1944 	/*
1945 	 * Bonding device
1946 	 *   < 0 - no bonding device (single one)
1947 	 *  >= 0 - bonding device (value is slave PF index)
1948 	 */
1949 	int bd = -1;
1950 	struct mlx5_dev_spawn_data *list = NULL;
1951 	struct mlx5_dev_config dev_config;
1952 	unsigned int dev_config_vf;
1953 	struct rte_eth_devargs eth_da = *req_eth_da;
1954 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1955 	struct mlx5_bond_info bond_info;
1956 	int ret = -1;
1957 
1958 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1959 		mlx5_pmd_socket_init();
1960 	ret = mlx5_init_once();
1961 	if (ret) {
1962 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1963 			strerror(rte_errno));
1964 		return -rte_errno;
1965 	}
1966 	errno = 0;
1967 	ibv_list = mlx5_glue->get_device_list(&ret);
1968 	if (!ibv_list) {
1969 		rte_errno = errno ? errno : ENOSYS;
1970 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1971 		return -rte_errno;
1972 	}
1973 	/*
1974 	 * First scan the list of all Infiniband devices to find
1975 	 * matching ones, gathering into the list.
1976 	 */
1977 	struct ibv_device *ibv_match[ret + 1];
1978 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1979 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1980 	unsigned int i;
1981 
1982 	while (ret-- > 0) {
1983 		struct rte_pci_addr pci_addr;
1984 
1985 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1986 		bd = mlx5_device_bond_pci_match
1987 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
1988 				 &bond_info);
1989 		if (bd >= 0) {
1990 			/*
1991 			 * Bonding device detected. Only one match is allowed,
1992 			 * the bonding is supported over multi-port IB device,
1993 			 * there should be no matches on representor PCI
1994 			 * functions or non VF LAG bonding devices with
1995 			 * specified address.
1996 			 */
1997 			if (nd) {
1998 				DRV_LOG(ERR,
1999 					"multiple PCI match on bonding device"
2000 					"\"%s\" found", ibv_list[ret]->name);
2001 				rte_errno = ENOENT;
2002 				ret = -rte_errno;
2003 				goto exit;
2004 			}
2005 			/* Amend owner pci address if owner PF ID specified. */
2006 			if (eth_da.nb_representor_ports)
2007 				owner_pci.function += owner_id;
2008 			DRV_LOG(INFO, "PCI information matches for"
2009 				      " slave %d bonding device \"%s\"",
2010 				      bd, ibv_list[ret]->name);
2011 			ibv_match[nd++] = ibv_list[ret];
2012 			break;
2013 		} else {
2014 			/* Bonding device not found. */
2015 			if (mlx5_dev_to_pci_addr
2016 				(ibv_list[ret]->ibdev_path, &pci_addr))
2017 				continue;
2018 			if (owner_pci.domain != pci_addr.domain ||
2019 			    owner_pci.bus != pci_addr.bus ||
2020 			    owner_pci.devid != pci_addr.devid ||
2021 			    owner_pci.function != pci_addr.function)
2022 				continue;
2023 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2024 				ibv_list[ret]->name);
2025 			ibv_match[nd++] = ibv_list[ret];
2026 		}
2027 	}
2028 	ibv_match[nd] = NULL;
2029 	if (!nd) {
2030 		/* No device matches, just complain and bail out. */
2031 		DRV_LOG(WARNING,
2032 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2033 			" are kernel drivers loaded?",
2034 			owner_pci.domain, owner_pci.bus,
2035 			owner_pci.devid, owner_pci.function);
2036 		rte_errno = ENOENT;
2037 		ret = -rte_errno;
2038 		goto exit;
2039 	}
2040 	if (nd == 1) {
2041 		/*
2042 		 * Found single matching device may have multiple ports.
2043 		 * Each port may be representor, we have to check the port
2044 		 * number and check the representors existence.
2045 		 */
2046 		if (nl_rdma >= 0)
2047 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2048 		if (!np)
2049 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2050 					 " ports number", ibv_match[0]->name);
2051 		if (bd >= 0 && !np) {
2052 			DRV_LOG(ERR, "can not get ports"
2053 				     " for bonding device");
2054 			rte_errno = ENOENT;
2055 			ret = -rte_errno;
2056 			goto exit;
2057 		}
2058 	}
2059 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2060 	if (bd >= 0) {
2061 		/*
2062 		 * This may happen if there is VF LAG kernel support and
2063 		 * application is compiled with older rdma_core library.
2064 		 */
2065 		DRV_LOG(ERR,
2066 			"No kernel/verbs support for VF LAG bonding found.");
2067 		rte_errno = ENOTSUP;
2068 		ret = -rte_errno;
2069 		goto exit;
2070 	}
2071 #endif
2072 	/*
2073 	 * Now we can determine the maximal
2074 	 * amount of devices to be spawned.
2075 	 */
2076 	list = mlx5_malloc(MLX5_MEM_ZERO,
2077 			   sizeof(struct mlx5_dev_spawn_data) *
2078 			   (np ? np : nd),
2079 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2080 	if (!list) {
2081 		DRV_LOG(ERR, "spawn data array allocation failure");
2082 		rte_errno = ENOMEM;
2083 		ret = -rte_errno;
2084 		goto exit;
2085 	}
2086 	if (bd >= 0 || np > 1) {
2087 		/*
2088 		 * Single IB device with multiple ports found,
2089 		 * it may be E-Switch master device and representors.
2090 		 * We have to perform identification through the ports.
2091 		 */
2092 		MLX5_ASSERT(nl_rdma >= 0);
2093 		MLX5_ASSERT(ns == 0);
2094 		MLX5_ASSERT(nd == 1);
2095 		MLX5_ASSERT(np);
2096 		for (i = 1; i <= np; ++i) {
2097 			list[ns].bond_info = &bond_info;
2098 			list[ns].max_port = np;
2099 			list[ns].phys_port = i;
2100 			list[ns].phys_dev = ibv_match[0];
2101 			list[ns].eth_dev = NULL;
2102 			list[ns].pci_dev = pci_dev;
2103 			list[ns].pf_bond = bd;
2104 			list[ns].ifindex = mlx5_nl_ifindex
2105 				(nl_rdma,
2106 				mlx5_os_get_dev_device_name
2107 						(list[ns].phys_dev), i);
2108 			if (!list[ns].ifindex) {
2109 				/*
2110 				 * No network interface index found for the
2111 				 * specified port, it means there is no
2112 				 * representor on this port. It's OK,
2113 				 * there can be disabled ports, for example
2114 				 * if sriov_numvfs < sriov_totalvfs.
2115 				 */
2116 				continue;
2117 			}
2118 			ret = -1;
2119 			if (nl_route >= 0)
2120 				ret = mlx5_nl_switch_info
2121 					       (nl_route,
2122 						list[ns].ifindex,
2123 						&list[ns].info);
2124 			if (ret || (!list[ns].info.representor &&
2125 				    !list[ns].info.master)) {
2126 				/*
2127 				 * We failed to recognize representors with
2128 				 * Netlink, let's try to perform the task
2129 				 * with sysfs.
2130 				 */
2131 				ret =  mlx5_sysfs_switch_info
2132 						(list[ns].ifindex,
2133 						 &list[ns].info);
2134 			}
2135 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2136 			if (!ret && bd >= 0) {
2137 				switch (list[ns].info.name_type) {
2138 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2139 					if (list[ns].info.port_name == bd)
2140 						ns++;
2141 					break;
2142 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2143 					/* Fallthrough */
2144 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2145 					/* Fallthrough */
2146 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2147 					if (list[ns].info.pf_num == bd)
2148 						ns++;
2149 					break;
2150 				default:
2151 					break;
2152 				}
2153 				continue;
2154 			}
2155 #endif
2156 			if (!ret && (list[ns].info.representor ^
2157 				     list[ns].info.master))
2158 				ns++;
2159 		}
2160 		if (!ns) {
2161 			DRV_LOG(ERR,
2162 				"unable to recognize master/representors"
2163 				" on the IB device with multiple ports");
2164 			rte_errno = ENOENT;
2165 			ret = -rte_errno;
2166 			goto exit;
2167 		}
2168 	} else {
2169 		/*
2170 		 * The existence of several matching entries (nd > 1) means
2171 		 * port representors have been instantiated. No existing Verbs
2172 		 * call nor sysfs entries can tell them apart, this can only
2173 		 * be done through Netlink calls assuming kernel drivers are
2174 		 * recent enough to support them.
2175 		 *
2176 		 * In the event of identification failure through Netlink,
2177 		 * try again through sysfs, then:
2178 		 *
2179 		 * 1. A single IB device matches (nd == 1) with single
2180 		 *    port (np=0/1) and is not a representor, assume
2181 		 *    no switch support.
2182 		 *
2183 		 * 2. Otherwise no safe assumptions can be made;
2184 		 *    complain louder and bail out.
2185 		 */
2186 		for (i = 0; i != nd; ++i) {
2187 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2188 			list[ns].bond_info = NULL;
2189 			list[ns].max_port = 1;
2190 			list[ns].phys_port = 1;
2191 			list[ns].phys_dev = ibv_match[i];
2192 			list[ns].eth_dev = NULL;
2193 			list[ns].pci_dev = pci_dev;
2194 			list[ns].pf_bond = -1;
2195 			list[ns].ifindex = 0;
2196 			if (nl_rdma >= 0)
2197 				list[ns].ifindex = mlx5_nl_ifindex
2198 				(nl_rdma,
2199 				mlx5_os_get_dev_device_name
2200 						(list[ns].phys_dev), 1);
2201 			if (!list[ns].ifindex) {
2202 				char ifname[IF_NAMESIZE];
2203 
2204 				/*
2205 				 * Netlink failed, it may happen with old
2206 				 * ib_core kernel driver (before 4.16).
2207 				 * We can assume there is old driver because
2208 				 * here we are processing single ports IB
2209 				 * devices. Let's try sysfs to retrieve
2210 				 * the ifindex. The method works for
2211 				 * master device only.
2212 				 */
2213 				if (nd > 1) {
2214 					/*
2215 					 * Multiple devices found, assume
2216 					 * representors, can not distinguish
2217 					 * master/representor and retrieve
2218 					 * ifindex via sysfs.
2219 					 */
2220 					continue;
2221 				}
2222 				ret = mlx5_get_ifname_sysfs
2223 					(ibv_match[i]->ibdev_path, ifname);
2224 				if (!ret)
2225 					list[ns].ifindex =
2226 						if_nametoindex(ifname);
2227 				if (!list[ns].ifindex) {
2228 					/*
2229 					 * No network interface index found
2230 					 * for the specified device, it means
2231 					 * there it is neither representor
2232 					 * nor master.
2233 					 */
2234 					continue;
2235 				}
2236 			}
2237 			ret = -1;
2238 			if (nl_route >= 0)
2239 				ret = mlx5_nl_switch_info
2240 					       (nl_route,
2241 						list[ns].ifindex,
2242 						&list[ns].info);
2243 			if (ret || (!list[ns].info.representor &&
2244 				    !list[ns].info.master)) {
2245 				/*
2246 				 * We failed to recognize representors with
2247 				 * Netlink, let's try to perform the task
2248 				 * with sysfs.
2249 				 */
2250 				ret =  mlx5_sysfs_switch_info
2251 						(list[ns].ifindex,
2252 						 &list[ns].info);
2253 			}
2254 			if (!ret && (list[ns].info.representor ^
2255 				     list[ns].info.master)) {
2256 				ns++;
2257 			} else if ((nd == 1) &&
2258 				   !list[ns].info.representor &&
2259 				   !list[ns].info.master) {
2260 				/*
2261 				 * Single IB device with
2262 				 * one physical port and
2263 				 * attached network device.
2264 				 * May be SRIOV is not enabled
2265 				 * or there is no representors.
2266 				 */
2267 				DRV_LOG(INFO, "no E-Switch support detected");
2268 				ns++;
2269 				break;
2270 			}
2271 		}
2272 		if (!ns) {
2273 			DRV_LOG(ERR,
2274 				"unable to recognize master/representors"
2275 				" on the multiple IB devices");
2276 			rte_errno = ENOENT;
2277 			ret = -rte_errno;
2278 			goto exit;
2279 		}
2280 	}
2281 	MLX5_ASSERT(ns);
2282 	/*
2283 	 * Sort list to probe devices in natural order for users convenience
2284 	 * (i.e. master first, then representors from lowest to highest ID).
2285 	 */
2286 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2287 	/* Device specific configuration. */
2288 	switch (pci_dev->id.device_id) {
2289 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2290 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2291 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2292 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2293 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2294 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2295 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2296 		dev_config_vf = 1;
2297 		break;
2298 	default:
2299 		dev_config_vf = 0;
2300 		break;
2301 	}
2302 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2303 		/* Set devargs default values. */
2304 		if (eth_da.nb_mh_controllers == 0) {
2305 			eth_da.nb_mh_controllers = 1;
2306 			eth_da.mh_controllers[0] = 0;
2307 		}
2308 		if (eth_da.nb_ports == 0 && ns > 0) {
2309 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2310 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2311 					pci_dev->device.devargs->args);
2312 			eth_da.nb_ports = 1;
2313 			eth_da.ports[0] = list[0].info.pf_num;
2314 		}
2315 		if (eth_da.nb_representor_ports == 0) {
2316 			eth_da.nb_representor_ports = 1;
2317 			eth_da.representor_ports[0] = 0;
2318 		}
2319 	}
2320 	for (i = 0; i != ns; ++i) {
2321 		uint32_t restore;
2322 
2323 		/* Default configuration. */
2324 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2325 		dev_config.vf = dev_config_vf;
2326 		dev_config.mps = MLX5_ARG_UNSET;
2327 		dev_config.dbnc = MLX5_ARG_UNSET;
2328 		dev_config.rx_vec_en = 1;
2329 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2330 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2331 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2332 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2333 		dev_config.vf_nl_en = 1;
2334 		dev_config.mr_ext_memseg_en = 1;
2335 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2336 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2337 		dev_config.dv_esw_en = 1;
2338 		dev_config.dv_flow_en = 1;
2339 		dev_config.decap_en = 1;
2340 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2341 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2342 						 &list[i],
2343 						 &dev_config,
2344 						 &eth_da);
2345 		if (!list[i].eth_dev) {
2346 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2347 				break;
2348 			/* Device is disabled or already spawned. Ignore it. */
2349 			continue;
2350 		}
2351 		restore = list[i].eth_dev->data->dev_flags;
2352 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2353 		/* Restore non-PCI flags cleared by the above call. */
2354 		list[i].eth_dev->data->dev_flags |= restore;
2355 		rte_eth_dev_probing_finish(list[i].eth_dev);
2356 	}
2357 	if (i != ns) {
2358 		DRV_LOG(ERR,
2359 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2360 			" encountering an error: %s",
2361 			owner_pci.domain, owner_pci.bus,
2362 			owner_pci.devid, owner_pci.function,
2363 			strerror(rte_errno));
2364 		ret = -rte_errno;
2365 		/* Roll back. */
2366 		while (i--) {
2367 			if (!list[i].eth_dev)
2368 				continue;
2369 			mlx5_dev_close(list[i].eth_dev);
2370 			/* mac_addrs must not be freed because in dev_private */
2371 			list[i].eth_dev->data->mac_addrs = NULL;
2372 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2373 		}
2374 		/* Restore original error. */
2375 		rte_errno = -ret;
2376 	} else {
2377 		ret = 0;
2378 	}
2379 exit:
2380 	/*
2381 	 * Do the routine cleanup:
2382 	 * - close opened Netlink sockets
2383 	 * - free allocated spawn data array
2384 	 * - free the Infiniband device list
2385 	 */
2386 	if (nl_rdma >= 0)
2387 		close(nl_rdma);
2388 	if (nl_route >= 0)
2389 		close(nl_route);
2390 	if (list)
2391 		mlx5_free(list);
2392 	MLX5_ASSERT(ibv_list);
2393 	mlx5_glue->free_device_list(ibv_list);
2394 	return ret;
2395 }
2396 
2397 /**
2398  * DPDK callback to register a PCI device.
2399  *
2400  * This function spawns Ethernet devices out of a given PCI device.
2401  *
2402  * @param[in] pci_drv
2403  *   PCI driver structure (mlx5_driver).
2404  * @param[in] pci_dev
2405  *   PCI device information.
2406  *
2407  * @return
2408  *   0 on success, a negative errno value otherwise and rte_errno is set.
2409  */
2410 int
2411 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2412 		  struct rte_pci_device *pci_dev)
2413 {
2414 	struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2415 	int ret = 0;
2416 	uint16_t p;
2417 
2418 	if (pci_dev->device.devargs) {
2419 		/* Parse representor information from device argument. */
2420 		if (pci_dev->device.devargs->cls_str)
2421 			ret = rte_eth_devargs_parse
2422 				(pci_dev->device.devargs->cls_str, &eth_da);
2423 		if (ret) {
2424 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2425 				pci_dev->device.devargs->cls_str);
2426 			return -rte_errno;
2427 		}
2428 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2429 			/* Support legacy device argument */
2430 			ret = rte_eth_devargs_parse
2431 				(pci_dev->device.devargs->args, &eth_da);
2432 			if (ret) {
2433 				DRV_LOG(ERR, "failed to parse device arguments: %s",
2434 					pci_dev->device.devargs->args);
2435 				return -rte_errno;
2436 			}
2437 		}
2438 	}
2439 
2440 	if (eth_da.nb_ports > 0) {
2441 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2442 		for (p = 0; p < eth_da.nb_ports; p++)
2443 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2444 						   eth_da.ports[p]);
2445 	} else {
2446 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2447 	}
2448 	return ret;
2449 }
2450 
2451 static int
2452 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2453 {
2454 	char *env;
2455 	int value;
2456 
2457 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2458 	/* Get environment variable to store. */
2459 	env = getenv(MLX5_SHUT_UP_BF);
2460 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2461 	if (config->dbnc == MLX5_ARG_UNSET)
2462 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2463 	else
2464 		setenv(MLX5_SHUT_UP_BF,
2465 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2466 	return value;
2467 }
2468 
2469 static void
2470 mlx5_restore_doorbell_mapping_env(int value)
2471 {
2472 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2473 	/* Restore the original environment variable state. */
2474 	if (value == MLX5_ARG_UNSET)
2475 		unsetenv(MLX5_SHUT_UP_BF);
2476 	else
2477 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2478 }
2479 
2480 /**
2481  * Extract pdn of PD object using DV API.
2482  *
2483  * @param[in] pd
2484  *   Pointer to the verbs PD object.
2485  * @param[out] pdn
2486  *   Pointer to the PD object number variable.
2487  *
2488  * @return
2489  *   0 on success, error value otherwise.
2490  */
2491 int
2492 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2493 {
2494 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2495 	struct mlx5dv_obj obj;
2496 	struct mlx5dv_pd pd_info;
2497 	int ret = 0;
2498 
2499 	obj.pd.in = pd;
2500 	obj.pd.out = &pd_info;
2501 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2502 	if (ret) {
2503 		DRV_LOG(DEBUG, "Fail to get PD object info");
2504 		return ret;
2505 	}
2506 	*pdn = pd_info.pdn;
2507 	return 0;
2508 #else
2509 	(void)pd;
2510 	(void)pdn;
2511 	return -ENOTSUP;
2512 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2513 }
2514 
2515 /**
2516  * Function API to open IB device.
2517  *
2518  * This function calls the Linux glue APIs to open a device.
2519  *
2520  * @param[in] spawn
2521  *   Pointer to the IB device attributes (name, port, etc).
2522  * @param[out] config
2523  *   Pointer to device configuration structure.
2524  * @param[out] sh
2525  *   Pointer to shared context structure.
2526  *
2527  * @return
2528  *   0 on success, a positive error value otherwise.
2529  */
2530 int
2531 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2532 		     const struct mlx5_dev_config *config,
2533 		     struct mlx5_dev_ctx_shared *sh)
2534 {
2535 	int dbmap_env;
2536 	int err = 0;
2537 
2538 	sh->numa_node = spawn->pci_dev->device.numa_node;
2539 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2540 	/*
2541 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2542 	 * before the device creation. The rdma_core library
2543 	 * checks the variable at device creation and
2544 	 * stores the result internally.
2545 	 */
2546 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2547 	/* Try to open IB device with DV first, then usual Verbs. */
2548 	errno = 0;
2549 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2550 	if (sh->ctx) {
2551 		sh->devx = 1;
2552 		DRV_LOG(DEBUG, "DevX is supported");
2553 		/* The device is created, no need for environment. */
2554 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2555 	} else {
2556 		/* The environment variable is still configured. */
2557 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2558 		err = errno ? errno : ENODEV;
2559 		/*
2560 		 * The environment variable is not needed anymore,
2561 		 * all device creation attempts are completed.
2562 		 */
2563 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2564 		if (!sh->ctx)
2565 			return err;
2566 		DRV_LOG(DEBUG, "DevX is NOT supported");
2567 		err = 0;
2568 	}
2569 	if (!err && sh->ctx) {
2570 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2571 		mlx5_glue->dv_set_context_attr(sh->ctx,
2572 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2573 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2574 				.alloc = &mlx5_alloc_verbs_buf,
2575 				.free = &mlx5_free_verbs_buf,
2576 				.data = sh,
2577 			}));
2578 	}
2579 	return err;
2580 }
2581 
2582 /**
2583  * Install shared asynchronous device events handler.
2584  * This function is implemented to support event sharing
2585  * between multiple ports of single IB device.
2586  *
2587  * @param sh
2588  *   Pointer to mlx5_dev_ctx_shared object.
2589  */
2590 void
2591 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2592 {
2593 	int ret;
2594 	int flags;
2595 
2596 	sh->intr_handle.fd = -1;
2597 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2598 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2599 		    F_SETFL, flags | O_NONBLOCK);
2600 	if (ret) {
2601 		DRV_LOG(INFO, "failed to change file descriptor async event"
2602 			" queue");
2603 	} else {
2604 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2605 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2606 		if (rte_intr_callback_register(&sh->intr_handle,
2607 					mlx5_dev_interrupt_handler, sh)) {
2608 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2609 			sh->intr_handle.fd = -1;
2610 		}
2611 	}
2612 	if (sh->devx) {
2613 #ifdef HAVE_IBV_DEVX_ASYNC
2614 		sh->intr_handle_devx.fd = -1;
2615 		sh->devx_comp =
2616 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2617 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2618 		if (!devx_comp) {
2619 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2620 			return;
2621 		}
2622 		flags = fcntl(devx_comp->fd, F_GETFL);
2623 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2624 		if (ret) {
2625 			DRV_LOG(INFO, "failed to change file descriptor"
2626 				" devx comp");
2627 			return;
2628 		}
2629 		sh->intr_handle_devx.fd = devx_comp->fd;
2630 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2631 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2632 					mlx5_dev_interrupt_handler_devx, sh)) {
2633 			DRV_LOG(INFO, "Fail to install the devx shared"
2634 				" interrupt.");
2635 			sh->intr_handle_devx.fd = -1;
2636 		}
2637 #endif /* HAVE_IBV_DEVX_ASYNC */
2638 	}
2639 }
2640 
2641 /**
2642  * Uninstall shared asynchronous device events handler.
2643  * This function is implemented to support event sharing
2644  * between multiple ports of single IB device.
2645  *
2646  * @param dev
2647  *   Pointer to mlx5_dev_ctx_shared object.
2648  */
2649 void
2650 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2651 {
2652 	if (sh->intr_handle.fd >= 0)
2653 		mlx5_intr_callback_unregister(&sh->intr_handle,
2654 					      mlx5_dev_interrupt_handler, sh);
2655 #ifdef HAVE_IBV_DEVX_ASYNC
2656 	if (sh->intr_handle_devx.fd >= 0)
2657 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2658 				  mlx5_dev_interrupt_handler_devx, sh);
2659 	if (sh->devx_comp)
2660 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2661 #endif
2662 }
2663 
2664 /**
2665  * Read statistics by a named counter.
2666  *
2667  * @param[in] priv
2668  *   Pointer to the private device data structure.
2669  * @param[in] ctr_name
2670  *   Pointer to the name of the statistic counter to read
2671  * @param[out] stat
2672  *   Pointer to read statistic value.
2673  * @return
2674  *   0 on success and stat is valud, 1 if failed to read the value
2675  *   rte_errno is set.
2676  *
2677  */
2678 int
2679 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2680 		      uint64_t *stat)
2681 {
2682 	int fd;
2683 
2684 	if (priv->sh) {
2685 		if (priv->q_counters != NULL &&
2686 		    strcmp(ctr_name, "out_of_buffer") == 0)
2687 			return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2688 							   0, (uint32_t *)stat);
2689 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2690 		      priv->sh->ibdev_path,
2691 		      priv->dev_port,
2692 		      ctr_name);
2693 		fd = open(path, O_RDONLY);
2694 		/*
2695 		 * in switchdev the file location is not per port
2696 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2697 		 */
2698 		if (fd == -1) {
2699 			MKSTR(path1, "%s/hw_counters/%s",
2700 			      priv->sh->ibdev_path,
2701 			      ctr_name);
2702 			fd = open(path1, O_RDONLY);
2703 		}
2704 		if (fd != -1) {
2705 			char buf[21] = {'\0'};
2706 			ssize_t n = read(fd, buf, sizeof(buf));
2707 
2708 			close(fd);
2709 			if (n != -1) {
2710 				*stat = strtoull(buf, NULL, 10);
2711 				return 0;
2712 			}
2713 		}
2714 	}
2715 	*stat = 0;
2716 	return 1;
2717 }
2718 
2719 /**
2720  * Set the reg_mr and dereg_mr call backs
2721  *
2722  * @param reg_mr_cb[out]
2723  *   Pointer to reg_mr func
2724  * @param dereg_mr_cb[out]
2725  *   Pointer to dereg_mr func
2726  *
2727  */
2728 void
2729 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2730 		      mlx5_dereg_mr_t *dereg_mr_cb)
2731 {
2732 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2733 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2734 }
2735 
2736 /**
2737  * Remove a MAC address from device
2738  *
2739  * @param dev
2740  *   Pointer to Ethernet device structure.
2741  * @param index
2742  *   MAC address index.
2743  */
2744 void
2745 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2746 {
2747 	struct mlx5_priv *priv = dev->data->dev_private;
2748 	const int vf = priv->config.vf;
2749 
2750 	if (vf)
2751 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2752 					mlx5_ifindex(dev), priv->mac_own,
2753 					&dev->data->mac_addrs[index], index);
2754 }
2755 
2756 /**
2757  * Adds a MAC address to the device
2758  *
2759  * @param dev
2760  *   Pointer to Ethernet device structure.
2761  * @param mac_addr
2762  *   MAC address to register.
2763  * @param index
2764  *   MAC address index.
2765  *
2766  * @return
2767  *   0 on success, a negative errno value otherwise
2768  */
2769 int
2770 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2771 		     uint32_t index)
2772 {
2773 	struct mlx5_priv *priv = dev->data->dev_private;
2774 	const int vf = priv->config.vf;
2775 	int ret = 0;
2776 
2777 	if (vf)
2778 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2779 					   mlx5_ifindex(dev), priv->mac_own,
2780 					   mac, index);
2781 	return ret;
2782 }
2783 
2784 /**
2785  * Modify a VF MAC address
2786  *
2787  * @param priv
2788  *   Pointer to device private data.
2789  * @param mac_addr
2790  *   MAC address to modify into.
2791  * @param iface_idx
2792  *   Net device interface index
2793  * @param vf_index
2794  *   VF index
2795  *
2796  * @return
2797  *   0 on success, a negative errno value otherwise
2798  */
2799 int
2800 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2801 			   unsigned int iface_idx,
2802 			   struct rte_ether_addr *mac_addr,
2803 			   int vf_index)
2804 {
2805 	return mlx5_nl_vf_mac_addr_modify
2806 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2807 }
2808 
2809 /**
2810  * Set device promiscuous mode
2811  *
2812  * @param dev
2813  *   Pointer to Ethernet device structure.
2814  * @param enable
2815  *   0 - promiscuous is disabled, otherwise - enabled
2816  *
2817  * @return
2818  *   0 on success, a negative error value otherwise
2819  */
2820 int
2821 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2822 {
2823 	struct mlx5_priv *priv = dev->data->dev_private;
2824 
2825 	return mlx5_nl_promisc(priv->nl_socket_route,
2826 			       mlx5_ifindex(dev), !!enable);
2827 }
2828 
2829 /**
2830  * Set device promiscuous mode
2831  *
2832  * @param dev
2833  *   Pointer to Ethernet device structure.
2834  * @param enable
2835  *   0 - all multicase is disabled, otherwise - enabled
2836  *
2837  * @return
2838  *   0 on success, a negative error value otherwise
2839  */
2840 int
2841 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2842 {
2843 	struct mlx5_priv *priv = dev->data->dev_private;
2844 
2845 	return mlx5_nl_allmulti(priv->nl_socket_route,
2846 				mlx5_ifindex(dev), !!enable);
2847 }
2848 
2849 /**
2850  * Flush device MAC addresses
2851  *
2852  * @param dev
2853  *   Pointer to Ethernet device structure.
2854  *
2855  */
2856 void
2857 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2858 {
2859 	struct mlx5_priv *priv = dev->data->dev_private;
2860 
2861 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2862 			       dev->data->mac_addrs,
2863 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2864 }
2865