xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision d0cf77e8c2b64319057f5f629a7a595ce6e8b556)
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 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
158 		sizeof(device_attr->fw_ver));
159 
160 	return err;
161 }
162 
163 /**
164  * Verbs callback to allocate a memory. This function should allocate the space
165  * according to the size provided residing inside a huge page.
166  * Please note that all allocation must respect the alignment from libmlx5
167  * (i.e. currently rte_mem_page_size()).
168  *
169  * @param[in] size
170  *   The size in bytes of the memory to allocate.
171  * @param[in] data
172  *   A pointer to the callback data.
173  *
174  * @return
175  *   Allocated buffer, NULL otherwise and rte_errno is set.
176  */
177 static void *
178 mlx5_alloc_verbs_buf(size_t size, void *data)
179 {
180 	struct mlx5_dev_ctx_shared *sh = data;
181 	void *ret;
182 	size_t alignment = rte_mem_page_size();
183 	if (alignment == (size_t)-1) {
184 		DRV_LOG(ERR, "Failed to get mem page size");
185 		rte_errno = ENOMEM;
186 		return NULL;
187 	}
188 
189 	MLX5_ASSERT(data != NULL);
190 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
191 	if (!ret && size)
192 		rte_errno = ENOMEM;
193 	return ret;
194 }
195 
196 /**
197  * Verbs callback to free a memory.
198  *
199  * @param[in] ptr
200  *   A pointer to the memory to free.
201  * @param[in] data
202  *   A pointer to the callback data.
203  */
204 static void
205 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
206 {
207 	MLX5_ASSERT(data != NULL);
208 	mlx5_free(ptr);
209 }
210 
211 /**
212  * Initialize DR related data within private structure.
213  * Routine checks the reference counter and does actual
214  * resources creation/initialization only if counter is zero.
215  *
216  * @param[in] priv
217  *   Pointer to the private device data structure.
218  *
219  * @return
220  *   Zero on success, positive error code otherwise.
221  */
222 static int
223 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
224 {
225 	struct mlx5_dev_ctx_shared *sh = priv->sh;
226 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
227 	int err;
228 
229 	MLX5_ASSERT(sh && sh->refcnt);
230 	if (sh->refcnt > 1)
231 		return 0;
232 	err = mlx5_alloc_table_hash_list(priv);
233 	if (err)
234 		goto error;
235 	/* The resources below are only valid with DV support. */
236 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
237 	/* Init port id action cache list. */
238 	snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
239 	mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
240 			     flow_dv_port_id_create_cb,
241 			     flow_dv_port_id_match_cb,
242 			     flow_dv_port_id_remove_cb);
243 	/* Init push vlan action cache list. */
244 	snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
245 	mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
246 			     flow_dv_push_vlan_create_cb,
247 			     flow_dv_push_vlan_match_cb,
248 			     flow_dv_push_vlan_remove_cb);
249 	/* Init sample action cache list. */
250 	snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
251 	mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
252 			     flow_dv_sample_create_cb,
253 			     flow_dv_sample_match_cb,
254 			     flow_dv_sample_remove_cb);
255 	/* Init dest array action cache list. */
256 	snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
257 	mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
258 			     flow_dv_dest_array_create_cb,
259 			     flow_dv_dest_array_match_cb,
260 			     flow_dv_dest_array_remove_cb);
261 	/* Create tags hash list table. */
262 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
263 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
264 					  MLX5_HLIST_WRITE_MOST,
265 					  flow_dv_tag_create_cb,
266 					  flow_dv_tag_match_cb,
267 					  flow_dv_tag_remove_cb);
268 	if (!sh->tag_table) {
269 		DRV_LOG(ERR, "tags with hash creation failed.");
270 		err = ENOMEM;
271 		goto error;
272 	}
273 	sh->tag_table->ctx = sh;
274 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
275 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
276 					    0, MLX5_HLIST_WRITE_MOST |
277 					    MLX5_HLIST_DIRECT_KEY,
278 					    flow_dv_modify_create_cb,
279 					    flow_dv_modify_match_cb,
280 					    flow_dv_modify_remove_cb);
281 	if (!sh->modify_cmds) {
282 		DRV_LOG(ERR, "hdr modify hash creation failed");
283 		err = ENOMEM;
284 		goto error;
285 	}
286 	sh->modify_cmds->ctx = sh;
287 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
288 	sh->encaps_decaps = mlx5_hlist_create(s,
289 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
290 					      0, MLX5_HLIST_DIRECT_KEY |
291 					      MLX5_HLIST_WRITE_MOST,
292 					      flow_dv_encap_decap_create_cb,
293 					      flow_dv_encap_decap_match_cb,
294 					      flow_dv_encap_decap_remove_cb);
295 	if (!sh->encaps_decaps) {
296 		DRV_LOG(ERR, "encap decap hash creation failed");
297 		err = ENOMEM;
298 		goto error;
299 	}
300 	sh->encaps_decaps->ctx = sh;
301 #endif
302 #ifdef HAVE_MLX5DV_DR
303 	void *domain;
304 
305 	/* Reference counter is zero, we should initialize structures. */
306 	domain = mlx5_glue->dr_create_domain(sh->ctx,
307 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
308 	if (!domain) {
309 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
310 		err = errno;
311 		goto error;
312 	}
313 	sh->rx_domain = domain;
314 	domain = mlx5_glue->dr_create_domain(sh->ctx,
315 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
316 	if (!domain) {
317 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
318 		err = errno;
319 		goto error;
320 	}
321 	sh->tx_domain = domain;
322 #ifdef HAVE_MLX5DV_DR_ESWITCH
323 	if (priv->config.dv_esw_en) {
324 		domain  = mlx5_glue->dr_create_domain
325 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
326 		if (!domain) {
327 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
328 			err = errno;
329 			goto error;
330 		}
331 		sh->fdb_domain = domain;
332 	}
333 	/*
334 	 * The drop action is just some dummy placeholder in rdma-core. It
335 	 * does not belong to domains and has no any attributes, and, can be
336 	 * shared by the entire device.
337 	 */
338 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
339 	if (!sh->dr_drop_action) {
340 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
341 		err = errno;
342 		goto error;
343 	}
344 #endif
345 	if (!sh->tunnel_hub)
346 		err = mlx5_alloc_tunnel_hub(sh);
347 	if (err) {
348 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
349 		goto error;
350 	}
351 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
352 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
353 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
354 		if (sh->fdb_domain)
355 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
356 	}
357 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
358 	if (!priv->config.allow_duplicate_pattern) {
359 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
360 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
361 #endif
362 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
363 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
364 		if (sh->fdb_domain)
365 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
366 	}
367 #endif /* HAVE_MLX5DV_DR */
368 	sh->default_miss_action =
369 			mlx5_glue->dr_create_flow_action_default_miss();
370 	if (!sh->default_miss_action)
371 		DRV_LOG(WARNING, "Default miss action is not supported.");
372 	return 0;
373 error:
374 	/* Rollback the created objects. */
375 	if (sh->rx_domain) {
376 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
377 		sh->rx_domain = NULL;
378 	}
379 	if (sh->tx_domain) {
380 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
381 		sh->tx_domain = NULL;
382 	}
383 	if (sh->fdb_domain) {
384 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
385 		sh->fdb_domain = NULL;
386 	}
387 	if (sh->dr_drop_action) {
388 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
389 		sh->dr_drop_action = NULL;
390 	}
391 	if (sh->pop_vlan_action) {
392 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
393 		sh->pop_vlan_action = NULL;
394 	}
395 	if (sh->encaps_decaps) {
396 		mlx5_hlist_destroy(sh->encaps_decaps);
397 		sh->encaps_decaps = NULL;
398 	}
399 	if (sh->modify_cmds) {
400 		mlx5_hlist_destroy(sh->modify_cmds);
401 		sh->modify_cmds = NULL;
402 	}
403 	if (sh->tag_table) {
404 		/* tags should be destroyed with flow before. */
405 		mlx5_hlist_destroy(sh->tag_table);
406 		sh->tag_table = NULL;
407 	}
408 	if (sh->tunnel_hub) {
409 		mlx5_release_tunnel_hub(sh, priv->dev_port);
410 		sh->tunnel_hub = NULL;
411 	}
412 	mlx5_free_table_hash_list(priv);
413 	return err;
414 }
415 
416 /**
417  * Destroy DR related data within private structure.
418  *
419  * @param[in] priv
420  *   Pointer to the private device data structure.
421  */
422 void
423 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
424 {
425 	struct mlx5_dev_ctx_shared *sh = priv->sh;
426 
427 	MLX5_ASSERT(sh && sh->refcnt);
428 	if (sh->refcnt > 1)
429 		return;
430 #ifdef HAVE_MLX5DV_DR
431 	if (sh->rx_domain) {
432 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
433 		sh->rx_domain = NULL;
434 	}
435 	if (sh->tx_domain) {
436 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
437 		sh->tx_domain = NULL;
438 	}
439 #ifdef HAVE_MLX5DV_DR_ESWITCH
440 	if (sh->fdb_domain) {
441 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
442 		sh->fdb_domain = NULL;
443 	}
444 	if (sh->dr_drop_action) {
445 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
446 		sh->dr_drop_action = NULL;
447 	}
448 #endif
449 	if (sh->pop_vlan_action) {
450 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
451 		sh->pop_vlan_action = NULL;
452 	}
453 #endif /* HAVE_MLX5DV_DR */
454 	if (sh->default_miss_action)
455 		mlx5_glue->destroy_flow_action
456 				(sh->default_miss_action);
457 	if (sh->encaps_decaps) {
458 		mlx5_hlist_destroy(sh->encaps_decaps);
459 		sh->encaps_decaps = NULL;
460 	}
461 	if (sh->modify_cmds) {
462 		mlx5_hlist_destroy(sh->modify_cmds);
463 		sh->modify_cmds = NULL;
464 	}
465 	if (sh->tag_table) {
466 		/* tags should be destroyed with flow before. */
467 		mlx5_hlist_destroy(sh->tag_table);
468 		sh->tag_table = NULL;
469 	}
470 	if (sh->tunnel_hub) {
471 		mlx5_release_tunnel_hub(sh, priv->dev_port);
472 		sh->tunnel_hub = NULL;
473 	}
474 	mlx5_cache_list_destroy(&sh->port_id_action_list);
475 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
476 	mlx5_free_table_hash_list(priv);
477 }
478 
479 /**
480  * Initialize shared data between primary and secondary process.
481  *
482  * A memzone is reserved by primary process and secondary processes attach to
483  * the memzone.
484  *
485  * @return
486  *   0 on success, a negative errno value otherwise and rte_errno is set.
487  */
488 static int
489 mlx5_init_shared_data(void)
490 {
491 	const struct rte_memzone *mz;
492 	int ret = 0;
493 
494 	rte_spinlock_lock(&mlx5_shared_data_lock);
495 	if (mlx5_shared_data == NULL) {
496 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
497 			/* Allocate shared memory. */
498 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
499 						 sizeof(*mlx5_shared_data),
500 						 SOCKET_ID_ANY, 0);
501 			if (mz == NULL) {
502 				DRV_LOG(ERR,
503 					"Cannot allocate mlx5 shared data");
504 				ret = -rte_errno;
505 				goto error;
506 			}
507 			mlx5_shared_data = mz->addr;
508 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
509 			rte_spinlock_init(&mlx5_shared_data->lock);
510 		} else {
511 			/* Lookup allocated shared memory. */
512 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
513 			if (mz == NULL) {
514 				DRV_LOG(ERR,
515 					"Cannot attach mlx5 shared data");
516 				ret = -rte_errno;
517 				goto error;
518 			}
519 			mlx5_shared_data = mz->addr;
520 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
521 		}
522 	}
523 error:
524 	rte_spinlock_unlock(&mlx5_shared_data_lock);
525 	return ret;
526 }
527 
528 /**
529  * PMD global initialization.
530  *
531  * Independent from individual device, this function initializes global
532  * per-PMD data structures distinguishing primary and secondary processes.
533  * Hence, each initialization is called once per a process.
534  *
535  * @return
536  *   0 on success, a negative errno value otherwise and rte_errno is set.
537  */
538 static int
539 mlx5_init_once(void)
540 {
541 	struct mlx5_shared_data *sd;
542 	struct mlx5_local_data *ld = &mlx5_local_data;
543 	int ret = 0;
544 
545 	if (mlx5_init_shared_data())
546 		return -rte_errno;
547 	sd = mlx5_shared_data;
548 	MLX5_ASSERT(sd);
549 	rte_spinlock_lock(&sd->lock);
550 	switch (rte_eal_process_type()) {
551 	case RTE_PROC_PRIMARY:
552 		if (sd->init_done)
553 			break;
554 		LIST_INIT(&sd->mem_event_cb_list);
555 		rte_rwlock_init(&sd->mem_event_rwlock);
556 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
557 						mlx5_mr_mem_event_cb, NULL);
558 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
559 					   mlx5_mp_os_primary_handle);
560 		if (ret)
561 			goto out;
562 		sd->init_done = true;
563 		break;
564 	case RTE_PROC_SECONDARY:
565 		if (ld->init_done)
566 			break;
567 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
568 					     mlx5_mp_os_secondary_handle);
569 		if (ret)
570 			goto out;
571 		++sd->secondary_cnt;
572 		ld->init_done = true;
573 		break;
574 	default:
575 		break;
576 	}
577 out:
578 	rte_spinlock_unlock(&sd->lock);
579 	return ret;
580 }
581 
582 /**
583  * Create the Tx queue DevX/Verbs object.
584  *
585  * @param dev
586  *   Pointer to Ethernet device.
587  * @param idx
588  *   Queue index in DPDK Tx queue array.
589  *
590  * @return
591  *   0 on success, a negative errno value otherwise and rte_errno is set.
592  */
593 static int
594 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
595 {
596 	struct mlx5_priv *priv = dev->data->dev_private;
597 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
598 	struct mlx5_txq_ctrl *txq_ctrl =
599 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
600 
601 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
602 		return mlx5_txq_devx_obj_new(dev, idx);
603 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
604 	if (!priv->config.dv_esw_en)
605 		return mlx5_txq_devx_obj_new(dev, idx);
606 #endif
607 	return mlx5_txq_ibv_obj_new(dev, idx);
608 }
609 
610 /**
611  * Release an Tx DevX/verbs queue object.
612  *
613  * @param txq_obj
614  *   DevX/Verbs Tx queue object.
615  */
616 static void
617 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
618 {
619 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
620 		mlx5_txq_devx_obj_release(txq_obj);
621 		return;
622 	}
623 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
624 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
625 		mlx5_txq_devx_obj_release(txq_obj);
626 		return;
627 	}
628 #endif
629 	mlx5_txq_ibv_obj_release(txq_obj);
630 }
631 
632 /**
633  * DV flow counter mode detect and config.
634  *
635  * @param dev
636  *   Pointer to rte_eth_dev structure.
637  *
638  */
639 static void
640 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
641 {
642 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
643 	struct mlx5_priv *priv = dev->data->dev_private;
644 	struct mlx5_dev_ctx_shared *sh = priv->sh;
645 	bool fallback;
646 
647 #ifndef HAVE_IBV_DEVX_ASYNC
648 	fallback = true;
649 #else
650 	fallback = false;
651 	if (!priv->config.devx || !priv->config.dv_flow_en ||
652 	    !priv->config.hca_attr.flow_counters_dump ||
653 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
654 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
655 		fallback = true;
656 #endif
657 	if (fallback)
658 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
659 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
660 			priv->config.hca_attr.flow_counters_dump,
661 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
662 	/* Initialize fallback mode only on the port initializes sh. */
663 	if (sh->refcnt == 1)
664 		sh->cmng.counter_fallback = fallback;
665 	else if (fallback != sh->cmng.counter_fallback)
666 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
667 			"with others:%d.", PORT_ID(priv), fallback);
668 #endif
669 }
670 
671 static void
672 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
673 {
674 	struct mlx5_priv *priv = dev->data->dev_private;
675 	void *ctx = priv->sh->ctx;
676 
677 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
678 	if (!priv->q_counters) {
679 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
680 		struct ibv_wq *wq;
681 
682 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
683 			"by DevX - fall-back to use the kernel driver global "
684 			"queue counter.", dev->data->port_id);
685 		/* Create WQ by kernel and query its queue counter ID. */
686 		if (cq) {
687 			wq = mlx5_glue->create_wq(ctx,
688 						  &(struct ibv_wq_init_attr){
689 						    .wq_type = IBV_WQT_RQ,
690 						    .max_wr = 1,
691 						    .max_sge = 1,
692 						    .pd = priv->sh->pd,
693 						    .cq = cq,
694 						});
695 			if (wq) {
696 				/* Counter is assigned only on RDY state. */
697 				int ret = mlx5_glue->modify_wq(wq,
698 						 &(struct ibv_wq_attr){
699 						 .attr_mask = IBV_WQ_ATTR_STATE,
700 						 .wq_state = IBV_WQS_RDY,
701 						});
702 
703 				if (ret == 0)
704 					mlx5_devx_cmd_wq_query(wq,
705 							 &priv->counter_set_id);
706 				claim_zero(mlx5_glue->destroy_wq(wq));
707 			}
708 			claim_zero(mlx5_glue->destroy_cq(cq));
709 		}
710 	} else {
711 		priv->counter_set_id = priv->q_counters->id;
712 	}
713 	if (priv->counter_set_id == 0)
714 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
715 			"available.", dev->data->port_id);
716 }
717 
718 /**
719  * Check if representor spawn info match devargs.
720  *
721  * @param spawn
722  *   Verbs device parameters (name, port, switch_info) to spawn.
723  * @param eth_da
724  *   Device devargs to probe.
725  *
726  * @return
727  *   Match result.
728  */
729 static bool
730 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
731 		       struct rte_eth_devargs *eth_da)
732 {
733 	struct mlx5_switch_info *switch_info = &spawn->info;
734 	unsigned int p, f;
735 	uint16_t id;
736 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
737 						      eth_da->type);
738 
739 	switch (eth_da->type) {
740 	case RTE_ETH_REPRESENTOR_SF:
741 		if (!(spawn->info.port_name == -1 &&
742 		      switch_info->name_type ==
743 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
744 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
745 			rte_errno = EBUSY;
746 			return false;
747 		}
748 		break;
749 	case RTE_ETH_REPRESENTOR_VF:
750 		/* Allows HPF representor index -1 as exception. */
751 		if (!(spawn->info.port_name == -1 &&
752 		      switch_info->name_type ==
753 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
754 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
755 			rte_errno = EBUSY;
756 			return false;
757 		}
758 		break;
759 	case RTE_ETH_REPRESENTOR_NONE:
760 		rte_errno = EBUSY;
761 		return false;
762 	default:
763 		rte_errno = ENOTSUP;
764 		DRV_LOG(ERR, "unsupported representor type");
765 		return false;
766 	}
767 	/* Check representor ID: */
768 	for (p = 0; p < eth_da->nb_ports; ++p) {
769 		if (spawn->pf_bond < 0) {
770 			/* For non-LAG mode, allow and ignore pf. */
771 			switch_info->pf_num = eth_da->ports[p];
772 			repr_id = mlx5_representor_id_encode(switch_info,
773 							     eth_da->type);
774 		}
775 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
776 			id = MLX5_REPRESENTOR_ID
777 				(eth_da->ports[p], eth_da->type,
778 				 eth_da->representor_ports[f]);
779 			if (repr_id == id)
780 				return true;
781 		}
782 	}
783 	rte_errno = EBUSY;
784 	return false;
785 }
786 
787 
788 /**
789  * Spawn an Ethernet device from Verbs information.
790  *
791  * @param dpdk_dev
792  *   Backing DPDK device.
793  * @param spawn
794  *   Verbs device parameters (name, port, switch_info) to spawn.
795  * @param config
796  *   Device configuration parameters.
797  * @param config
798  *   Device arguments.
799  *
800  * @return
801  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
802  *   is set. The following errors are defined:
803  *
804  *   EBUSY: device is not supposed to be spawned.
805  *   EEXIST: device is already spawned
806  */
807 static struct rte_eth_dev *
808 mlx5_dev_spawn(struct rte_device *dpdk_dev,
809 	       struct mlx5_dev_spawn_data *spawn,
810 	       struct mlx5_dev_config *config,
811 	       struct rte_eth_devargs *eth_da)
812 {
813 	const struct mlx5_switch_info *switch_info = &spawn->info;
814 	struct mlx5_dev_ctx_shared *sh = NULL;
815 	struct ibv_port_attr port_attr;
816 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
817 	struct rte_eth_dev *eth_dev = NULL;
818 	struct mlx5_priv *priv = NULL;
819 	int err = 0;
820 	unsigned int hw_padding = 0;
821 	unsigned int mps;
822 	unsigned int tunnel_en = 0;
823 	unsigned int mpls_en = 0;
824 	unsigned int swp = 0;
825 	unsigned int mprq = 0;
826 	unsigned int mprq_min_stride_size_n = 0;
827 	unsigned int mprq_max_stride_size_n = 0;
828 	unsigned int mprq_min_stride_num_n = 0;
829 	unsigned int mprq_max_stride_num_n = 0;
830 	struct rte_ether_addr mac;
831 	char name[RTE_ETH_NAME_MAX_LEN];
832 	int own_domain_id = 0;
833 	uint16_t port_id;
834 	struct mlx5_port_info vport_info = { .query_flags = 0 };
835 
836 	/* Determine if this port representor is supposed to be spawned. */
837 	if (switch_info->representor && dpdk_dev->devargs &&
838 	    !mlx5_representor_match(spawn, eth_da))
839 		return NULL;
840 	/* Build device name. */
841 	if (spawn->pf_bond < 0) {
842 		/* Single device. */
843 		if (!switch_info->representor)
844 			strlcpy(name, dpdk_dev->name, sizeof(name));
845 		else
846 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
847 				 dpdk_dev->name,
848 				 switch_info->name_type ==
849 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
850 				 switch_info->port_name);
851 	} else {
852 		/* Bonding device. */
853 		if (!switch_info->representor) {
854 			err = snprintf(name, sizeof(name), "%s_%s",
855 				 dpdk_dev->name,
856 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
857 		} else {
858 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
859 				dpdk_dev->name,
860 				mlx5_os_get_dev_device_name(spawn->phys_dev),
861 				switch_info->ctrl_num,
862 				switch_info->pf_num,
863 				switch_info->name_type ==
864 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
865 				switch_info->port_name);
866 		}
867 	}
868 	if (err >= (int)sizeof(name))
869 		DRV_LOG(WARNING, "device name overflow %s", name);
870 	/* check if the device is already spawned */
871 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
872 		rte_errno = EEXIST;
873 		return NULL;
874 	}
875 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
876 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
877 		struct mlx5_mp_id mp_id;
878 
879 		eth_dev = rte_eth_dev_attach_secondary(name);
880 		if (eth_dev == NULL) {
881 			DRV_LOG(ERR, "can not attach rte ethdev");
882 			rte_errno = ENOMEM;
883 			return NULL;
884 		}
885 		eth_dev->device = dpdk_dev;
886 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
887 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
888 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
889 		err = mlx5_proc_priv_init(eth_dev);
890 		if (err)
891 			return NULL;
892 		mp_id.port_id = eth_dev->data->port_id;
893 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
894 		/* Receive command fd from primary process */
895 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
896 		if (err < 0)
897 			goto err_secondary;
898 		/* Remap UAR for Tx queues. */
899 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
900 		if (err)
901 			goto err_secondary;
902 		/*
903 		 * Ethdev pointer is still required as input since
904 		 * the primary device is not accessible from the
905 		 * secondary process.
906 		 */
907 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
908 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
909 		return eth_dev;
910 err_secondary:
911 		mlx5_dev_close(eth_dev);
912 		return NULL;
913 	}
914 	/*
915 	 * Some parameters ("tx_db_nc" in particularly) are needed in
916 	 * advance to create dv/verbs device context. We proceed the
917 	 * devargs here to get ones, and later proceed devargs again
918 	 * to override some hardware settings.
919 	 */
920 	err = mlx5_args(config, dpdk_dev->devargs);
921 	if (err) {
922 		err = rte_errno;
923 		DRV_LOG(ERR, "failed to process device arguments: %s",
924 			strerror(rte_errno));
925 		goto error;
926 	}
927 	if (config->dv_miss_info) {
928 		if (switch_info->master || switch_info->representor)
929 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
930 	}
931 	mlx5_malloc_mem_select(config->sys_mem_en);
932 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
933 	if (!sh)
934 		return NULL;
935 	config->devx = sh->devx;
936 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
937 	config->dest_tir = 1;
938 #endif
939 #ifdef HAVE_IBV_MLX5_MOD_SWP
940 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
941 #endif
942 	/*
943 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
944 	 * as all ConnectX-5 devices.
945 	 */
946 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
947 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
948 #endif
949 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
950 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
951 #endif
952 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
953 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
954 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
955 			DRV_LOG(DEBUG, "enhanced MPW is supported");
956 			mps = MLX5_MPW_ENHANCED;
957 		} else {
958 			DRV_LOG(DEBUG, "MPW is supported");
959 			mps = MLX5_MPW;
960 		}
961 	} else {
962 		DRV_LOG(DEBUG, "MPW isn't supported");
963 		mps = MLX5_MPW_DISABLED;
964 	}
965 #ifdef HAVE_IBV_MLX5_MOD_SWP
966 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
967 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
968 	DRV_LOG(DEBUG, "SWP support: %u", swp);
969 #endif
970 	config->swp = !!swp;
971 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
972 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
973 		struct mlx5dv_striding_rq_caps mprq_caps =
974 			dv_attr.striding_rq_caps;
975 
976 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
977 			mprq_caps.min_single_stride_log_num_of_bytes);
978 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
979 			mprq_caps.max_single_stride_log_num_of_bytes);
980 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
981 			mprq_caps.min_single_wqe_log_num_of_strides);
982 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
983 			mprq_caps.max_single_wqe_log_num_of_strides);
984 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
985 			mprq_caps.supported_qpts);
986 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
987 		mprq = 1;
988 		mprq_min_stride_size_n =
989 			mprq_caps.min_single_stride_log_num_of_bytes;
990 		mprq_max_stride_size_n =
991 			mprq_caps.max_single_stride_log_num_of_bytes;
992 		mprq_min_stride_num_n =
993 			mprq_caps.min_single_wqe_log_num_of_strides;
994 		mprq_max_stride_num_n =
995 			mprq_caps.max_single_wqe_log_num_of_strides;
996 	}
997 #endif
998 	/* Rx CQE compression is enabled by default. */
999 	config->cqe_comp = 1;
1000 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1001 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1002 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1003 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1004 			     (dv_attr.tunnel_offloads_caps &
1005 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
1006 			     (dv_attr.tunnel_offloads_caps &
1007 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1008 	}
1009 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1010 		tunnel_en ? "" : "not ");
1011 #else
1012 	DRV_LOG(WARNING,
1013 		"tunnel offloading disabled due to old OFED/rdma-core version");
1014 #endif
1015 	config->tunnel_en = tunnel_en;
1016 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1017 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1018 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1019 		   (dv_attr.tunnel_offloads_caps &
1020 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1021 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1022 		mpls_en ? "" : "not ");
1023 #else
1024 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1025 		" old OFED/rdma-core version or firmware configuration");
1026 #endif
1027 	config->mpls_en = mpls_en;
1028 	/* Check port status. */
1029 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1030 	if (err) {
1031 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1032 		goto error;
1033 	}
1034 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1035 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1036 		err = EINVAL;
1037 		goto error;
1038 	}
1039 	if (port_attr.state != IBV_PORT_ACTIVE)
1040 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1041 			mlx5_glue->port_state_str(port_attr.state),
1042 			port_attr.state);
1043 	/* Allocate private eth device data. */
1044 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1045 			   sizeof(*priv),
1046 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1047 	if (priv == NULL) {
1048 		DRV_LOG(ERR, "priv allocation failure");
1049 		err = ENOMEM;
1050 		goto error;
1051 	}
1052 	priv->sh = sh;
1053 	priv->dev_port = spawn->phys_port;
1054 	priv->pci_dev = spawn->pci_dev;
1055 	priv->mtu = RTE_ETHER_MTU;
1056 	/* Some internal functions rely on Netlink sockets, open them now. */
1057 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1058 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1059 	priv->representor = !!switch_info->representor;
1060 	priv->master = !!switch_info->master;
1061 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1062 	priv->vport_meta_tag = 0;
1063 	priv->vport_meta_mask = 0;
1064 	priv->pf_bond = spawn->pf_bond;
1065 	/*
1066 	 * If we have E-Switch we should determine the vport attributes.
1067 	 * E-Switch may use either source vport field or reg_c[0] metadata
1068 	 * register to match on vport index. The engaged part of metadata
1069 	 * register is defined by mask.
1070 	 */
1071 	if (switch_info->representor || switch_info->master) {
1072 		err = mlx5_glue->devx_port_query(sh->ctx,
1073 						 spawn->phys_port,
1074 						 &vport_info);
1075 		if (err) {
1076 			DRV_LOG(WARNING,
1077 				"can't query devx port %d on device %s",
1078 				spawn->phys_port,
1079 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1080 			vport_info.query_flags = 0;
1081 		}
1082 	}
1083 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1084 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1085 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1086 		if (!priv->vport_meta_mask) {
1087 			DRV_LOG(ERR, "vport zero mask for port %d"
1088 				     " on bonding device %s",
1089 				     spawn->phys_port,
1090 				     mlx5_os_get_dev_device_name
1091 							(spawn->phys_dev));
1092 			err = ENOTSUP;
1093 			goto error;
1094 		}
1095 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1096 			DRV_LOG(ERR, "invalid vport tag for port %d"
1097 				     " on bonding device %s",
1098 				     spawn->phys_port,
1099 				     mlx5_os_get_dev_device_name
1100 							(spawn->phys_dev));
1101 			err = ENOTSUP;
1102 			goto error;
1103 		}
1104 	}
1105 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1106 		priv->vport_id = vport_info.vport_id;
1107 	} else if (spawn->pf_bond >= 0 &&
1108 		   (switch_info->representor || switch_info->master)) {
1109 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1110 			     " on bonding device %s",
1111 			     spawn->phys_port,
1112 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1113 		err = ENOTSUP;
1114 		goto error;
1115 	} else {
1116 		/*
1117 		 * Suppose vport index in compatible way. Kernel/rdma_core
1118 		 * support single E-Switch per PF configurations only and
1119 		 * vport_id field contains the vport index for associated VF,
1120 		 * which is deduced from representor port name.
1121 		 * For example, let's have the IB device port 10, it has
1122 		 * attached network device eth0, which has port name attribute
1123 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1124 		 * as 3 (2+1). This assigning schema should be changed if the
1125 		 * multiple E-Switch instances per PF configurations or/and PCI
1126 		 * subfunctions are added.
1127 		 */
1128 		priv->vport_id = switch_info->representor ?
1129 				 switch_info->port_name + 1 : -1;
1130 	}
1131 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1132 							  eth_da->type);
1133 	/*
1134 	 * Look for sibling devices in order to reuse their switch domain
1135 	 * if any, otherwise allocate one.
1136 	 */
1137 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1138 		const struct mlx5_priv *opriv =
1139 			rte_eth_devices[port_id].data->dev_private;
1140 
1141 		if (!opriv ||
1142 		    opriv->sh != priv->sh ||
1143 			opriv->domain_id ==
1144 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1145 			continue;
1146 		priv->domain_id = opriv->domain_id;
1147 		break;
1148 	}
1149 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1150 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1151 		if (err) {
1152 			err = rte_errno;
1153 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1154 				strerror(rte_errno));
1155 			goto error;
1156 		}
1157 		own_domain_id = 1;
1158 	}
1159 	/* Override some values set by hardware configuration. */
1160 	mlx5_args(config, dpdk_dev->devargs);
1161 	err = mlx5_dev_check_sibling_config(priv, config);
1162 	if (err)
1163 		goto error;
1164 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1165 			    IBV_DEVICE_RAW_IP_CSUM);
1166 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1167 		(config->hw_csum ? "" : "not "));
1168 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1169 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1170 	DRV_LOG(DEBUG, "counters are not supported");
1171 #endif
1172 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1173 	if (config->dv_flow_en) {
1174 		DRV_LOG(WARNING, "DV flow is not supported");
1175 		config->dv_flow_en = 0;
1176 	}
1177 #endif
1178 	config->ind_table_max_size =
1179 		sh->device_attr.max_rwq_indirection_table_size;
1180 	/*
1181 	 * Remove this check once DPDK supports larger/variable
1182 	 * indirection tables.
1183 	 */
1184 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1185 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1186 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1187 		config->ind_table_max_size);
1188 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1189 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1190 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1191 		(config->hw_vlan_strip ? "" : "not "));
1192 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1193 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1194 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1195 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1196 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1197 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1198 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1199 #endif
1200 	if (config->hw_padding && !hw_padding) {
1201 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1202 		config->hw_padding = 0;
1203 	} else if (config->hw_padding) {
1204 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1205 	}
1206 	config->tso = (sh->device_attr.max_tso > 0 &&
1207 		      (sh->device_attr.tso_supported_qpts &
1208 		       (1 << IBV_QPT_RAW_PACKET)));
1209 	if (config->tso)
1210 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1211 	/*
1212 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1213 	 * by default.
1214 	 */
1215 	if (config->mps == MLX5_ARG_UNSET)
1216 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1217 							  MLX5_MPW_DISABLED;
1218 	else
1219 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1220 	DRV_LOG(INFO, "%sMPS is %s",
1221 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1222 		config->mps == MLX5_MPW ? "legacy " : "",
1223 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1224 	if (config->devx) {
1225 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1226 		if (err) {
1227 			err = -err;
1228 			goto error;
1229 		}
1230 		/* Check relax ordering support. */
1231 		if (!haswell_broadwell_cpu) {
1232 			sh->cmng.relaxed_ordering_write =
1233 				config->hca_attr.relaxed_ordering_write;
1234 			sh->cmng.relaxed_ordering_read =
1235 				config->hca_attr.relaxed_ordering_read;
1236 		} else {
1237 			sh->cmng.relaxed_ordering_read = 0;
1238 			sh->cmng.relaxed_ordering_write = 0;
1239 		}
1240 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1241 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1242 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1243 		/* Check for LRO support. */
1244 		if (config->dest_tir && config->hca_attr.lro_cap &&
1245 		    config->dv_flow_en) {
1246 			/* TBD check tunnel lro caps. */
1247 			config->lro.supported = config->hca_attr.lro_cap;
1248 			DRV_LOG(DEBUG, "Device supports LRO");
1249 			/*
1250 			 * If LRO timeout is not configured by application,
1251 			 * use the minimal supported value.
1252 			 */
1253 			if (!config->lro.timeout)
1254 				config->lro.timeout =
1255 				config->hca_attr.lro_timer_supported_periods[0];
1256 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1257 				config->lro.timeout);
1258 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1259 				"required for coalescing is %d bytes",
1260 				config->hca_attr.lro_min_mss_size);
1261 		}
1262 #if defined(HAVE_MLX5DV_DR) && \
1263 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1264 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1265 		if (config->hca_attr.qos.sup &&
1266 		    config->hca_attr.qos.flow_meter_old &&
1267 		    config->dv_flow_en) {
1268 			uint8_t reg_c_mask =
1269 				config->hca_attr.qos.flow_meter_reg_c_ids;
1270 			/*
1271 			 * Meter needs two REG_C's for color match and pre-sfx
1272 			 * flow match. Here get the REG_C for color match.
1273 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1274 			 */
1275 			reg_c_mask &= 0xfc;
1276 			if (__builtin_popcount(reg_c_mask) < 1) {
1277 				priv->mtr_en = 0;
1278 				DRV_LOG(WARNING, "No available register for"
1279 					" meter.");
1280 			} else {
1281 				/*
1282 				 * The meter color register is used by the
1283 				 * flow-hit feature as well.
1284 				 * The flow-hit feature must use REG_C_3
1285 				 * Prefer REG_C_3 if it is available.
1286 				 */
1287 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1288 					priv->mtr_color_reg = REG_C_3;
1289 				else
1290 					priv->mtr_color_reg = ffs(reg_c_mask)
1291 							      - 1 + REG_C_0;
1292 				priv->mtr_en = 1;
1293 				priv->mtr_reg_share =
1294 				      config->hca_attr.qos.flow_meter;
1295 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1296 					priv->mtr_color_reg);
1297 			}
1298 		}
1299 		if (config->hca_attr.qos.sup &&
1300 			config->hca_attr.qos.flow_meter_aso_sup) {
1301 			uint32_t log_obj_size =
1302 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1303 			if (log_obj_size >=
1304 			config->hca_attr.qos.log_meter_aso_granularity &&
1305 			log_obj_size <=
1306 			config->hca_attr.qos.log_meter_aso_max_alloc)
1307 				sh->meter_aso_en = 1;
1308 		}
1309 		if (priv->mtr_en) {
1310 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1311 			if (err) {
1312 				err = -err;
1313 				goto error;
1314 			}
1315 		}
1316 #endif
1317 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1318 		if (config->hca_attr.flow_hit_aso &&
1319 		    priv->mtr_color_reg == REG_C_3) {
1320 			sh->flow_hit_aso_en = 1;
1321 			err = mlx5_flow_aso_age_mng_init(sh);
1322 			if (err) {
1323 				err = -err;
1324 				goto error;
1325 			}
1326 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1327 		}
1328 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1329 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1330 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1331 		if (config->hca_attr.ct_offload &&
1332 		    priv->mtr_color_reg == REG_C_3) {
1333 			err = mlx5_flow_aso_ct_mng_init(sh);
1334 			if (err) {
1335 				err = -err;
1336 				goto error;
1337 			}
1338 			DRV_LOG(DEBUG, "CT ASO is supported.");
1339 			sh->ct_aso_en = 1;
1340 		}
1341 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1342 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1343 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1344 		    config->dv_flow_en) {
1345 			priv->sampler_en = 1;
1346 			DRV_LOG(DEBUG, "Sampler enabled!");
1347 		} else {
1348 			priv->sampler_en = 0;
1349 			if (!config->hca_attr.log_max_ft_sampler_num)
1350 				DRV_LOG(WARNING,
1351 					"No available register for sampler.");
1352 			else
1353 				DRV_LOG(DEBUG, "DV flow is not supported!");
1354 		}
1355 #endif
1356 	}
1357 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1358 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1359 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1360 		config->cqe_comp = 0;
1361 	}
1362 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1363 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1364 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1365 				 " format isn't supported.");
1366 		config->cqe_comp = 0;
1367 	}
1368 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1369 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1370 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1371 				 " format isn't supported.");
1372 		config->cqe_comp = 0;
1373 	}
1374 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1375 			config->cqe_comp ? "" : "not ");
1376 	if (config->tx_pp) {
1377 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1378 			config->hca_attr.dev_freq_khz);
1379 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1380 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1381 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1382 			config->hca_attr.cross_channel ? "" : "not ");
1383 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1384 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1385 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1386 			config->hca_attr.non_wire_sq ? "" : "not ");
1387 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1388 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1389 			config->hca_attr.log_max_static_sq_wq);
1390 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1391 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1392 		if (!config->devx) {
1393 			DRV_LOG(ERR, "DevX is required for packet pacing");
1394 			err = ENODEV;
1395 			goto error;
1396 		}
1397 		if (!config->hca_attr.qos.packet_pacing) {
1398 			DRV_LOG(ERR, "Packet pacing is not supported");
1399 			err = ENODEV;
1400 			goto error;
1401 		}
1402 		if (!config->hca_attr.cross_channel) {
1403 			DRV_LOG(ERR, "Cross channel operations are"
1404 				     " required for packet pacing");
1405 			err = ENODEV;
1406 			goto error;
1407 		}
1408 		if (!config->hca_attr.wqe_index_ignore) {
1409 			DRV_LOG(ERR, "WQE index ignore feature is"
1410 				     " required for packet pacing");
1411 			err = ENODEV;
1412 			goto error;
1413 		}
1414 		if (!config->hca_attr.non_wire_sq) {
1415 			DRV_LOG(ERR, "Non-wire SQ feature is"
1416 				     " required for packet pacing");
1417 			err = ENODEV;
1418 			goto error;
1419 		}
1420 		if (!config->hca_attr.log_max_static_sq_wq) {
1421 			DRV_LOG(ERR, "Static WQE SQ feature is"
1422 				     " required for packet pacing");
1423 			err = ENODEV;
1424 			goto error;
1425 		}
1426 		if (!config->hca_attr.qos.wqe_rate_pp) {
1427 			DRV_LOG(ERR, "WQE rate mode is required"
1428 				     " for packet pacing");
1429 			err = ENODEV;
1430 			goto error;
1431 		}
1432 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1433 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1434 			     " can't create queues for packet pacing");
1435 		err = ENODEV;
1436 		goto error;
1437 #endif
1438 	}
1439 	if (config->devx) {
1440 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1441 
1442 		err = config->hca_attr.access_register_user ?
1443 			mlx5_devx_cmd_register_read
1444 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1445 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1446 		if (!err) {
1447 			uint32_t ts_mode;
1448 
1449 			/* MTUTC register is read successfully. */
1450 			ts_mode = MLX5_GET(register_mtutc, reg,
1451 					   time_stamp_mode);
1452 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1453 				config->rt_timestamp = 1;
1454 		} else {
1455 			/* Kernel does not support register reading. */
1456 			if (config->hca_attr.dev_freq_khz ==
1457 						 (NS_PER_S / MS_PER_S))
1458 				config->rt_timestamp = 1;
1459 		}
1460 	}
1461 	/*
1462 	 * If HW has bug working with tunnel packet decapsulation and
1463 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1464 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1465 	 */
1466 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1467 		config->hw_fcs_strip = 0;
1468 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1469 		(config->hw_fcs_strip ? "" : "not "));
1470 	if (config->mprq.enabled && mprq) {
1471 		if (config->mprq.stride_num_n &&
1472 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1473 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1474 			config->mprq.stride_num_n =
1475 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1476 						mprq_min_stride_num_n),
1477 					mprq_max_stride_num_n);
1478 			DRV_LOG(WARNING,
1479 				"the number of strides"
1480 				" for Multi-Packet RQ is out of range,"
1481 				" setting default value (%u)",
1482 				1 << config->mprq.stride_num_n);
1483 		}
1484 		if (config->mprq.stride_size_n &&
1485 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1486 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1487 			config->mprq.stride_size_n =
1488 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1489 						mprq_min_stride_size_n),
1490 					mprq_max_stride_size_n);
1491 			DRV_LOG(WARNING,
1492 				"the size of a stride"
1493 				" for Multi-Packet RQ is out of range,"
1494 				" setting default value (%u)",
1495 				1 << config->mprq.stride_size_n);
1496 		}
1497 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1498 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1499 	} else if (config->mprq.enabled && !mprq) {
1500 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1501 		config->mprq.enabled = 0;
1502 	}
1503 	if (config->max_dump_files_num == 0)
1504 		config->max_dump_files_num = 128;
1505 	eth_dev = rte_eth_dev_allocate(name);
1506 	if (eth_dev == NULL) {
1507 		DRV_LOG(ERR, "can not allocate rte ethdev");
1508 		err = ENOMEM;
1509 		goto error;
1510 	}
1511 	if (priv->representor) {
1512 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1513 		eth_dev->data->representor_id = priv->representor_id;
1514 	}
1515 	priv->mp_id.port_id = eth_dev->data->port_id;
1516 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1517 	/*
1518 	 * Store associated network device interface index. This index
1519 	 * is permanent throughout the lifetime of device. So, we may store
1520 	 * the ifindex here and use the cached value further.
1521 	 */
1522 	MLX5_ASSERT(spawn->ifindex);
1523 	priv->if_index = spawn->ifindex;
1524 	eth_dev->data->dev_private = priv;
1525 	priv->dev_data = eth_dev->data;
1526 	eth_dev->data->mac_addrs = priv->mac;
1527 	eth_dev->device = dpdk_dev;
1528 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1529 	/* Configure the first MAC address by default. */
1530 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1531 		DRV_LOG(ERR,
1532 			"port %u cannot get MAC address, is mlx5_en"
1533 			" loaded? (errno: %s)",
1534 			eth_dev->data->port_id, strerror(rte_errno));
1535 		err = ENODEV;
1536 		goto error;
1537 	}
1538 	DRV_LOG(INFO,
1539 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1540 		eth_dev->data->port_id,
1541 		mac.addr_bytes[0], mac.addr_bytes[1],
1542 		mac.addr_bytes[2], mac.addr_bytes[3],
1543 		mac.addr_bytes[4], mac.addr_bytes[5]);
1544 #ifdef RTE_LIBRTE_MLX5_DEBUG
1545 	{
1546 		char ifname[MLX5_NAMESIZE];
1547 
1548 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1549 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1550 				eth_dev->data->port_id, ifname);
1551 		else
1552 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1553 				eth_dev->data->port_id);
1554 	}
1555 #endif
1556 	/* Get actual MTU if possible. */
1557 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1558 	if (err) {
1559 		err = rte_errno;
1560 		goto error;
1561 	}
1562 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1563 		priv->mtu);
1564 	/* Initialize burst functions to prevent crashes before link-up. */
1565 	eth_dev->rx_pkt_burst = removed_rx_burst;
1566 	eth_dev->tx_pkt_burst = removed_tx_burst;
1567 	eth_dev->dev_ops = &mlx5_dev_ops;
1568 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1569 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1570 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1571 	/* Register MAC address. */
1572 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1573 	if (config->vf && config->vf_nl_en)
1574 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1575 				      mlx5_ifindex(eth_dev),
1576 				      eth_dev->data->mac_addrs,
1577 				      MLX5_MAX_MAC_ADDRESSES);
1578 	priv->flows = 0;
1579 	priv->ctrl_flows = 0;
1580 	rte_spinlock_init(&priv->flow_list_lock);
1581 	TAILQ_INIT(&priv->flow_meters);
1582 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1583 	if (!priv->mtr_profile_tbl)
1584 		goto error;
1585 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1586 	mlx5_glue->dv_set_context_attr(sh->ctx,
1587 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1588 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1589 				.alloc = &mlx5_alloc_verbs_buf,
1590 				.free = &mlx5_free_verbs_buf,
1591 				.data = sh,
1592 			}));
1593 	/* Bring Ethernet device up. */
1594 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1595 		eth_dev->data->port_id);
1596 	mlx5_set_link_up(eth_dev);
1597 	/*
1598 	 * Even though the interrupt handler is not installed yet,
1599 	 * interrupts will still trigger on the async_fd from
1600 	 * Verbs context returned by ibv_open_device().
1601 	 */
1602 	mlx5_link_update(eth_dev, 0);
1603 #ifdef HAVE_MLX5DV_DR_ESWITCH
1604 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1605 	      (switch_info->representor || switch_info->master)))
1606 		config->dv_esw_en = 0;
1607 #else
1608 	config->dv_esw_en = 0;
1609 #endif
1610 	/* Detect minimal data bytes to inline. */
1611 	mlx5_set_min_inline(spawn, config);
1612 	/* Store device configuration on private structure. */
1613 	priv->config = *config;
1614 	/* Create context for virtual machine VLAN workaround. */
1615 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1616 	if (config->dv_flow_en) {
1617 		err = mlx5_alloc_shared_dr(priv);
1618 		if (err)
1619 			goto error;
1620 	}
1621 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1622 		priv->obj_ops = devx_obj_ops;
1623 		priv->obj_ops.drop_action_create =
1624 						ibv_obj_ops.drop_action_create;
1625 		priv->obj_ops.drop_action_destroy =
1626 						ibv_obj_ops.drop_action_destroy;
1627 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1628 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1629 #else
1630 		if (config->dv_esw_en)
1631 			priv->obj_ops.txq_obj_modify =
1632 						ibv_obj_ops.txq_obj_modify;
1633 #endif
1634 		/* Use specific wrappers for Tx object. */
1635 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1636 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1637 		mlx5_queue_counter_id_prepare(eth_dev);
1638 		priv->obj_ops.lb_dummy_queue_create =
1639 					mlx5_rxq_ibv_obj_dummy_lb_create;
1640 		priv->obj_ops.lb_dummy_queue_release =
1641 					mlx5_rxq_ibv_obj_dummy_lb_release;
1642 	} else {
1643 		priv->obj_ops = ibv_obj_ops;
1644 	}
1645 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1646 	if (!priv->drop_queue.hrxq)
1647 		goto error;
1648 	/* Supported Verbs flow priority number detection. */
1649 	err = mlx5_flow_discover_priorities(eth_dev);
1650 	if (err < 0) {
1651 		err = -err;
1652 		goto error;
1653 	}
1654 	priv->config.flow_prio = err;
1655 	if (!priv->config.dv_esw_en &&
1656 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1657 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1658 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1659 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1660 	}
1661 	mlx5_set_metadata_mask(eth_dev);
1662 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1663 	    !priv->sh->dv_regc0_mask) {
1664 		DRV_LOG(ERR, "metadata mode %u is not supported "
1665 			     "(no metadata reg_c[0] is available)",
1666 			     priv->config.dv_xmeta_en);
1667 			err = ENOTSUP;
1668 			goto error;
1669 	}
1670 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1671 			     mlx5_hrxq_create_cb,
1672 			     mlx5_hrxq_match_cb,
1673 			     mlx5_hrxq_remove_cb);
1674 	/* Query availability of metadata reg_c's. */
1675 	err = mlx5_flow_discover_mreg_c(eth_dev);
1676 	if (err < 0) {
1677 		err = -err;
1678 		goto error;
1679 	}
1680 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1681 		DRV_LOG(DEBUG,
1682 			"port %u extensive metadata register is not supported",
1683 			eth_dev->data->port_id);
1684 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1685 			DRV_LOG(ERR, "metadata mode %u is not supported "
1686 				     "(no metadata registers available)",
1687 				     priv->config.dv_xmeta_en);
1688 			err = ENOTSUP;
1689 			goto error;
1690 		}
1691 	}
1692 	if (priv->config.dv_flow_en &&
1693 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1694 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1695 	    priv->sh->dv_regc0_mask) {
1696 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1697 						      MLX5_FLOW_MREG_HTABLE_SZ,
1698 						      0, 0,
1699 						      flow_dv_mreg_create_cb,
1700 						      flow_dv_mreg_match_cb,
1701 						      flow_dv_mreg_remove_cb);
1702 		if (!priv->mreg_cp_tbl) {
1703 			err = ENOMEM;
1704 			goto error;
1705 		}
1706 		priv->mreg_cp_tbl->ctx = eth_dev;
1707 	}
1708 	rte_spinlock_init(&priv->shared_act_sl);
1709 	mlx5_flow_counter_mode_config(eth_dev);
1710 	if (priv->config.dv_flow_en)
1711 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1712 	return eth_dev;
1713 error:
1714 	if (priv) {
1715 		if (priv->mreg_cp_tbl)
1716 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1717 		if (priv->sh)
1718 			mlx5_os_free_shared_dr(priv);
1719 		if (priv->nl_socket_route >= 0)
1720 			close(priv->nl_socket_route);
1721 		if (priv->nl_socket_rdma >= 0)
1722 			close(priv->nl_socket_rdma);
1723 		if (priv->vmwa_context)
1724 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1725 		if (eth_dev && priv->drop_queue.hrxq)
1726 			mlx5_drop_action_destroy(eth_dev);
1727 		if (priv->mtr_profile_tbl)
1728 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1729 		if (own_domain_id)
1730 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1731 		mlx5_cache_list_destroy(&priv->hrxqs);
1732 		mlx5_free(priv);
1733 		if (eth_dev != NULL)
1734 			eth_dev->data->dev_private = NULL;
1735 	}
1736 	if (eth_dev != NULL) {
1737 		/* mac_addrs must not be freed alone because part of
1738 		 * dev_private
1739 		 **/
1740 		eth_dev->data->mac_addrs = NULL;
1741 		rte_eth_dev_release_port(eth_dev);
1742 	}
1743 	if (sh)
1744 		mlx5_free_shared_dev_ctx(sh);
1745 	MLX5_ASSERT(err > 0);
1746 	rte_errno = err;
1747 	return NULL;
1748 }
1749 
1750 /**
1751  * Comparison callback to sort device data.
1752  *
1753  * This is meant to be used with qsort().
1754  *
1755  * @param a[in]
1756  *   Pointer to pointer to first data object.
1757  * @param b[in]
1758  *   Pointer to pointer to second data object.
1759  *
1760  * @return
1761  *   0 if both objects are equal, less than 0 if the first argument is less
1762  *   than the second, greater than 0 otherwise.
1763  */
1764 static int
1765 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1766 {
1767 	const struct mlx5_switch_info *si_a =
1768 		&((const struct mlx5_dev_spawn_data *)a)->info;
1769 	const struct mlx5_switch_info *si_b =
1770 		&((const struct mlx5_dev_spawn_data *)b)->info;
1771 	int ret;
1772 
1773 	/* Master device first. */
1774 	ret = si_b->master - si_a->master;
1775 	if (ret)
1776 		return ret;
1777 	/* Then representor devices. */
1778 	ret = si_b->representor - si_a->representor;
1779 	if (ret)
1780 		return ret;
1781 	/* Unidentified devices come last in no specific order. */
1782 	if (!si_a->representor)
1783 		return 0;
1784 	/* Order representors by name. */
1785 	return si_a->port_name - si_b->port_name;
1786 }
1787 
1788 /**
1789  * Match PCI information for possible slaves of bonding device.
1790  *
1791  * @param[in] ibv_dev
1792  *   Pointer to Infiniband device structure.
1793  * @param[in] pci_dev
1794  *   Pointer to primary PCI address structure to match.
1795  * @param[in] nl_rdma
1796  *   Netlink RDMA group socket handle.
1797  * @param[in] owner
1798  *   Rerepsentor owner PF index.
1799  * @param[out] bond_info
1800  *   Pointer to bonding information.
1801  *
1802  * @return
1803  *   negative value if no bonding device found, otherwise
1804  *   positive index of slave PF in bonding.
1805  */
1806 static int
1807 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1808 			   const struct rte_pci_addr *pci_dev,
1809 			   int nl_rdma, uint16_t owner,
1810 			   struct mlx5_bond_info *bond_info)
1811 {
1812 	char ifname[IF_NAMESIZE + 1];
1813 	unsigned int ifindex;
1814 	unsigned int np, i;
1815 	FILE *bond_file = NULL, *file;
1816 	int pf = -1;
1817 	int ret;
1818 
1819 	/*
1820 	 * Try to get master device name. If something goes
1821 	 * wrong suppose the lack of kernel support and no
1822 	 * bonding devices.
1823 	 */
1824 	memset(bond_info, 0, sizeof(*bond_info));
1825 	if (nl_rdma < 0)
1826 		return -1;
1827 	if (!strstr(ibv_dev->name, "bond"))
1828 		return -1;
1829 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1830 	if (!np)
1831 		return -1;
1832 	/*
1833 	 * The Master device might not be on the predefined
1834 	 * port (not on port index 1, it is not garanted),
1835 	 * we have to scan all Infiniband device port and
1836 	 * find master.
1837 	 */
1838 	for (i = 1; i <= np; ++i) {
1839 		/* Check whether Infiniband port is populated. */
1840 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1841 		if (!ifindex)
1842 			continue;
1843 		if (!if_indextoname(ifindex, ifname))
1844 			continue;
1845 		/* Try to read bonding slave names from sysfs. */
1846 		MKSTR(slaves,
1847 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1848 		bond_file = fopen(slaves, "r");
1849 		if (bond_file)
1850 			break;
1851 	}
1852 	if (!bond_file)
1853 		return -1;
1854 	/* Use safe format to check maximal buffer length. */
1855 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1856 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1857 		char tmp_str[IF_NAMESIZE + 32];
1858 		struct rte_pci_addr pci_addr;
1859 		struct mlx5_switch_info	info;
1860 
1861 		/* Process slave interface names in the loop. */
1862 		snprintf(tmp_str, sizeof(tmp_str),
1863 			 "/sys/class/net/%s", ifname);
1864 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1865 			DRV_LOG(WARNING, "can not get PCI address"
1866 					 " for netdev \"%s\"", ifname);
1867 			continue;
1868 		}
1869 		/* Slave interface PCI address match found. */
1870 		snprintf(tmp_str, sizeof(tmp_str),
1871 			 "/sys/class/net/%s/phys_port_name", ifname);
1872 		file = fopen(tmp_str, "rb");
1873 		if (!file)
1874 			break;
1875 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1876 		if (fscanf(file, "%32s", tmp_str) == 1)
1877 			mlx5_translate_port_name(tmp_str, &info);
1878 		fclose(file);
1879 		/* Only process PF ports. */
1880 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1881 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1882 			continue;
1883 		/* Check max bonding member. */
1884 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1885 			DRV_LOG(WARNING, "bonding index out of range, "
1886 				"please increase MLX5_BOND_MAX_PORTS: %s",
1887 				tmp_str);
1888 			break;
1889 		}
1890 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
1891 		if (pci_dev->domain == pci_addr.domain &&
1892 		    pci_dev->bus == pci_addr.bus &&
1893 		    pci_dev->devid == pci_addr.devid &&
1894 		    ((pci_dev->function == 0 &&
1895 		      pci_dev->function + owner == pci_addr.function) ||
1896 		     (pci_dev->function == owner &&
1897 		      pci_addr.function == owner)))
1898 			pf = info.port_name;
1899 		/* Get ifindex. */
1900 		snprintf(tmp_str, sizeof(tmp_str),
1901 			 "/sys/class/net/%s/ifindex", ifname);
1902 		file = fopen(tmp_str, "rb");
1903 		if (!file)
1904 			break;
1905 		ret = fscanf(file, "%u", &ifindex);
1906 		fclose(file);
1907 		if (ret != 1)
1908 			break;
1909 		/* Save bonding info. */
1910 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1911 			sizeof(bond_info->ports[0].ifname));
1912 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1913 		bond_info->ports[info.port_name].ifindex = ifindex;
1914 		bond_info->n_port++;
1915 	}
1916 	if (pf >= 0) {
1917 		/* Get bond interface info */
1918 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1919 					   bond_info->ifname);
1920 		if (ret)
1921 			DRV_LOG(ERR, "unable to get bond info: %s",
1922 				strerror(rte_errno));
1923 		else
1924 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1925 				ifindex, bond_info->ifindex, bond_info->ifname);
1926 	}
1927 	return pf;
1928 }
1929 
1930 /**
1931  * Register a PCI device within bonding.
1932  *
1933  * This function spawns Ethernet devices out of a given PCI device and
1934  * bonding owner PF index.
1935  *
1936  * @param[in] pci_dev
1937  *   PCI device information.
1938  * @param[in] req_eth_da
1939  *   Requested ethdev device argument.
1940  * @param[in] owner_id
1941  *   Requested owner PF port ID within bonding device, default to 0.
1942  *
1943  * @return
1944  *   0 on success, a negative errno value otherwise and rte_errno is set.
1945  */
1946 static int
1947 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1948 		     struct rte_eth_devargs *req_eth_da,
1949 		     uint16_t owner_id)
1950 {
1951 	struct ibv_device **ibv_list;
1952 	/*
1953 	 * Number of found IB Devices matching with requested PCI BDF.
1954 	 * nd != 1 means there are multiple IB devices over the same
1955 	 * PCI device and we have representors and master.
1956 	 */
1957 	unsigned int nd = 0;
1958 	/*
1959 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1960 	 * we have the single multiport IB device, and there may be
1961 	 * representors attached to some of found ports.
1962 	 */
1963 	unsigned int np = 0;
1964 	/*
1965 	 * Number of DPDK ethernet devices to Spawn - either over
1966 	 * multiple IB devices or multiple ports of single IB device.
1967 	 * Actually this is the number of iterations to spawn.
1968 	 */
1969 	unsigned int ns = 0;
1970 	/*
1971 	 * Bonding device
1972 	 *   < 0 - no bonding device (single one)
1973 	 *  >= 0 - bonding device (value is slave PF index)
1974 	 */
1975 	int bd = -1;
1976 	struct mlx5_dev_spawn_data *list = NULL;
1977 	struct mlx5_dev_config dev_config;
1978 	unsigned int dev_config_vf;
1979 	struct rte_eth_devargs eth_da = *req_eth_da;
1980 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1981 	struct mlx5_bond_info bond_info;
1982 	int ret = -1;
1983 
1984 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1985 		mlx5_pmd_socket_init();
1986 	ret = mlx5_init_once();
1987 	if (ret) {
1988 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1989 			strerror(rte_errno));
1990 		return -rte_errno;
1991 	}
1992 	errno = 0;
1993 	ibv_list = mlx5_glue->get_device_list(&ret);
1994 	if (!ibv_list) {
1995 		rte_errno = errno ? errno : ENOSYS;
1996 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1997 		return -rte_errno;
1998 	}
1999 	/*
2000 	 * First scan the list of all Infiniband devices to find
2001 	 * matching ones, gathering into the list.
2002 	 */
2003 	struct ibv_device *ibv_match[ret + 1];
2004 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2005 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2006 	unsigned int i;
2007 
2008 	while (ret-- > 0) {
2009 		struct rte_pci_addr pci_addr;
2010 
2011 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2012 		bd = mlx5_device_bond_pci_match
2013 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2014 				 &bond_info);
2015 		if (bd >= 0) {
2016 			/*
2017 			 * Bonding device detected. Only one match is allowed,
2018 			 * the bonding is supported over multi-port IB device,
2019 			 * there should be no matches on representor PCI
2020 			 * functions or non VF LAG bonding devices with
2021 			 * specified address.
2022 			 */
2023 			if (nd) {
2024 				DRV_LOG(ERR,
2025 					"multiple PCI match on bonding device"
2026 					"\"%s\" found", ibv_list[ret]->name);
2027 				rte_errno = ENOENT;
2028 				ret = -rte_errno;
2029 				goto exit;
2030 			}
2031 			/* Amend owner pci address if owner PF ID specified. */
2032 			if (eth_da.nb_representor_ports)
2033 				owner_pci.function += owner_id;
2034 			DRV_LOG(INFO, "PCI information matches for"
2035 				      " slave %d bonding device \"%s\"",
2036 				      bd, ibv_list[ret]->name);
2037 			ibv_match[nd++] = ibv_list[ret];
2038 			break;
2039 		} else {
2040 			/* Bonding device not found. */
2041 			if (mlx5_dev_to_pci_addr
2042 				(ibv_list[ret]->ibdev_path, &pci_addr))
2043 				continue;
2044 			if (owner_pci.domain != pci_addr.domain ||
2045 			    owner_pci.bus != pci_addr.bus ||
2046 			    owner_pci.devid != pci_addr.devid ||
2047 			    owner_pci.function != pci_addr.function)
2048 				continue;
2049 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2050 				ibv_list[ret]->name);
2051 			ibv_match[nd++] = ibv_list[ret];
2052 		}
2053 	}
2054 	ibv_match[nd] = NULL;
2055 	if (!nd) {
2056 		/* No device matches, just complain and bail out. */
2057 		DRV_LOG(WARNING,
2058 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2059 			" are kernel drivers loaded?",
2060 			owner_pci.domain, owner_pci.bus,
2061 			owner_pci.devid, owner_pci.function);
2062 		rte_errno = ENOENT;
2063 		ret = -rte_errno;
2064 		goto exit;
2065 	}
2066 	if (nd == 1) {
2067 		/*
2068 		 * Found single matching device may have multiple ports.
2069 		 * Each port may be representor, we have to check the port
2070 		 * number and check the representors existence.
2071 		 */
2072 		if (nl_rdma >= 0)
2073 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2074 		if (!np)
2075 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2076 					 " ports number", ibv_match[0]->name);
2077 		if (bd >= 0 && !np) {
2078 			DRV_LOG(ERR, "can not get ports"
2079 				     " for bonding device");
2080 			rte_errno = ENOENT;
2081 			ret = -rte_errno;
2082 			goto exit;
2083 		}
2084 	}
2085 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2086 	if (bd >= 0) {
2087 		/*
2088 		 * This may happen if there is VF LAG kernel support and
2089 		 * application is compiled with older rdma_core library.
2090 		 */
2091 		DRV_LOG(ERR,
2092 			"No kernel/verbs support for VF LAG bonding found.");
2093 		rte_errno = ENOTSUP;
2094 		ret = -rte_errno;
2095 		goto exit;
2096 	}
2097 #endif
2098 	/*
2099 	 * Now we can determine the maximal
2100 	 * amount of devices to be spawned.
2101 	 */
2102 	list = mlx5_malloc(MLX5_MEM_ZERO,
2103 			   sizeof(struct mlx5_dev_spawn_data) *
2104 			   (np ? np : nd),
2105 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2106 	if (!list) {
2107 		DRV_LOG(ERR, "spawn data array allocation failure");
2108 		rte_errno = ENOMEM;
2109 		ret = -rte_errno;
2110 		goto exit;
2111 	}
2112 	if (bd >= 0 || np > 1) {
2113 		/*
2114 		 * Single IB device with multiple ports found,
2115 		 * it may be E-Switch master device and representors.
2116 		 * We have to perform identification through the ports.
2117 		 */
2118 		MLX5_ASSERT(nl_rdma >= 0);
2119 		MLX5_ASSERT(ns == 0);
2120 		MLX5_ASSERT(nd == 1);
2121 		MLX5_ASSERT(np);
2122 		for (i = 1; i <= np; ++i) {
2123 			list[ns].bond_info = &bond_info;
2124 			list[ns].max_port = np;
2125 			list[ns].phys_port = i;
2126 			list[ns].phys_dev = ibv_match[0];
2127 			list[ns].eth_dev = NULL;
2128 			list[ns].pci_dev = pci_dev;
2129 			list[ns].pf_bond = bd;
2130 			list[ns].ifindex = mlx5_nl_ifindex
2131 				(nl_rdma,
2132 				mlx5_os_get_dev_device_name
2133 						(list[ns].phys_dev), i);
2134 			if (!list[ns].ifindex) {
2135 				/*
2136 				 * No network interface index found for the
2137 				 * specified port, it means there is no
2138 				 * representor on this port. It's OK,
2139 				 * there can be disabled ports, for example
2140 				 * if sriov_numvfs < sriov_totalvfs.
2141 				 */
2142 				continue;
2143 			}
2144 			ret = -1;
2145 			if (nl_route >= 0)
2146 				ret = mlx5_nl_switch_info
2147 					       (nl_route,
2148 						list[ns].ifindex,
2149 						&list[ns].info);
2150 			if (ret || (!list[ns].info.representor &&
2151 				    !list[ns].info.master)) {
2152 				/*
2153 				 * We failed to recognize representors with
2154 				 * Netlink, let's try to perform the task
2155 				 * with sysfs.
2156 				 */
2157 				ret =  mlx5_sysfs_switch_info
2158 						(list[ns].ifindex,
2159 						 &list[ns].info);
2160 			}
2161 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2162 			if (!ret && bd >= 0) {
2163 				switch (list[ns].info.name_type) {
2164 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2165 					if (list[ns].info.port_name == bd)
2166 						ns++;
2167 					break;
2168 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2169 					/* Fallthrough */
2170 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2171 					/* Fallthrough */
2172 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2173 					if (list[ns].info.pf_num == bd)
2174 						ns++;
2175 					break;
2176 				default:
2177 					break;
2178 				}
2179 				continue;
2180 			}
2181 #endif
2182 			if (!ret && (list[ns].info.representor ^
2183 				     list[ns].info.master))
2184 				ns++;
2185 		}
2186 		if (!ns) {
2187 			DRV_LOG(ERR,
2188 				"unable to recognize master/representors"
2189 				" on the IB device with multiple ports");
2190 			rte_errno = ENOENT;
2191 			ret = -rte_errno;
2192 			goto exit;
2193 		}
2194 	} else {
2195 		/*
2196 		 * The existence of several matching entries (nd > 1) means
2197 		 * port representors have been instantiated. No existing Verbs
2198 		 * call nor sysfs entries can tell them apart, this can only
2199 		 * be done through Netlink calls assuming kernel drivers are
2200 		 * recent enough to support them.
2201 		 *
2202 		 * In the event of identification failure through Netlink,
2203 		 * try again through sysfs, then:
2204 		 *
2205 		 * 1. A single IB device matches (nd == 1) with single
2206 		 *    port (np=0/1) and is not a representor, assume
2207 		 *    no switch support.
2208 		 *
2209 		 * 2. Otherwise no safe assumptions can be made;
2210 		 *    complain louder and bail out.
2211 		 */
2212 		for (i = 0; i != nd; ++i) {
2213 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2214 			list[ns].bond_info = NULL;
2215 			list[ns].max_port = 1;
2216 			list[ns].phys_port = 1;
2217 			list[ns].phys_dev = ibv_match[i];
2218 			list[ns].eth_dev = NULL;
2219 			list[ns].pci_dev = pci_dev;
2220 			list[ns].pf_bond = -1;
2221 			list[ns].ifindex = 0;
2222 			if (nl_rdma >= 0)
2223 				list[ns].ifindex = mlx5_nl_ifindex
2224 				(nl_rdma,
2225 				mlx5_os_get_dev_device_name
2226 						(list[ns].phys_dev), 1);
2227 			if (!list[ns].ifindex) {
2228 				char ifname[IF_NAMESIZE];
2229 
2230 				/*
2231 				 * Netlink failed, it may happen with old
2232 				 * ib_core kernel driver (before 4.16).
2233 				 * We can assume there is old driver because
2234 				 * here we are processing single ports IB
2235 				 * devices. Let's try sysfs to retrieve
2236 				 * the ifindex. The method works for
2237 				 * master device only.
2238 				 */
2239 				if (nd > 1) {
2240 					/*
2241 					 * Multiple devices found, assume
2242 					 * representors, can not distinguish
2243 					 * master/representor and retrieve
2244 					 * ifindex via sysfs.
2245 					 */
2246 					continue;
2247 				}
2248 				ret = mlx5_get_ifname_sysfs
2249 					(ibv_match[i]->ibdev_path, ifname);
2250 				if (!ret)
2251 					list[ns].ifindex =
2252 						if_nametoindex(ifname);
2253 				if (!list[ns].ifindex) {
2254 					/*
2255 					 * No network interface index found
2256 					 * for the specified device, it means
2257 					 * there it is neither representor
2258 					 * nor master.
2259 					 */
2260 					continue;
2261 				}
2262 			}
2263 			ret = -1;
2264 			if (nl_route >= 0)
2265 				ret = mlx5_nl_switch_info
2266 					       (nl_route,
2267 						list[ns].ifindex,
2268 						&list[ns].info);
2269 			if (ret || (!list[ns].info.representor &&
2270 				    !list[ns].info.master)) {
2271 				/*
2272 				 * We failed to recognize representors with
2273 				 * Netlink, let's try to perform the task
2274 				 * with sysfs.
2275 				 */
2276 				ret =  mlx5_sysfs_switch_info
2277 						(list[ns].ifindex,
2278 						 &list[ns].info);
2279 			}
2280 			if (!ret && (list[ns].info.representor ^
2281 				     list[ns].info.master)) {
2282 				ns++;
2283 			} else if ((nd == 1) &&
2284 				   !list[ns].info.representor &&
2285 				   !list[ns].info.master) {
2286 				/*
2287 				 * Single IB device with
2288 				 * one physical port and
2289 				 * attached network device.
2290 				 * May be SRIOV is not enabled
2291 				 * or there is no representors.
2292 				 */
2293 				DRV_LOG(INFO, "no E-Switch support detected");
2294 				ns++;
2295 				break;
2296 			}
2297 		}
2298 		if (!ns) {
2299 			DRV_LOG(ERR,
2300 				"unable to recognize master/representors"
2301 				" on the multiple IB devices");
2302 			rte_errno = ENOENT;
2303 			ret = -rte_errno;
2304 			goto exit;
2305 		}
2306 		/*
2307 		 * New kernels may add the switch_id attribute for the case
2308 		 * there is no E-Switch and we wrongly recognized the
2309 		 * only device as master. Override this if there is the
2310 		 * single device with single port and new device name
2311 		 * format present.
2312 		 */
2313 		if (nd == 1 &&
2314 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2315 			list[0].info.master = 0;
2316 			list[0].info.representor = 0;
2317 		}
2318 	}
2319 	MLX5_ASSERT(ns);
2320 	/*
2321 	 * Sort list to probe devices in natural order for users convenience
2322 	 * (i.e. master first, then representors from lowest to highest ID).
2323 	 */
2324 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2325 	/* Device specific configuration. */
2326 	switch (pci_dev->id.device_id) {
2327 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2328 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2329 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2330 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2331 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2332 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2333 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2334 		dev_config_vf = 1;
2335 		break;
2336 	default:
2337 		dev_config_vf = 0;
2338 		break;
2339 	}
2340 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2341 		/* Set devargs default values. */
2342 		if (eth_da.nb_mh_controllers == 0) {
2343 			eth_da.nb_mh_controllers = 1;
2344 			eth_da.mh_controllers[0] = 0;
2345 		}
2346 		if (eth_da.nb_ports == 0 && ns > 0) {
2347 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2348 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2349 					pci_dev->device.devargs->args);
2350 			eth_da.nb_ports = 1;
2351 			eth_da.ports[0] = list[0].info.pf_num;
2352 		}
2353 		if (eth_da.nb_representor_ports == 0) {
2354 			eth_da.nb_representor_ports = 1;
2355 			eth_da.representor_ports[0] = 0;
2356 		}
2357 	}
2358 	for (i = 0; i != ns; ++i) {
2359 		uint32_t restore;
2360 
2361 		/* Default configuration. */
2362 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2363 		dev_config.vf = dev_config_vf;
2364 		dev_config.mps = MLX5_ARG_UNSET;
2365 		dev_config.dbnc = MLX5_ARG_UNSET;
2366 		dev_config.rx_vec_en = 1;
2367 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2368 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2369 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2370 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2371 		dev_config.vf_nl_en = 1;
2372 		dev_config.mr_ext_memseg_en = 1;
2373 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2374 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2375 		dev_config.dv_esw_en = 1;
2376 		dev_config.dv_flow_en = 1;
2377 		dev_config.decap_en = 1;
2378 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2379 		dev_config.allow_duplicate_pattern = 1;
2380 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2381 						 &list[i],
2382 						 &dev_config,
2383 						 &eth_da);
2384 		if (!list[i].eth_dev) {
2385 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2386 				break;
2387 			/* Device is disabled or already spawned. Ignore it. */
2388 			continue;
2389 		}
2390 		restore = list[i].eth_dev->data->dev_flags;
2391 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2392 		/* Restore non-PCI flags cleared by the above call. */
2393 		list[i].eth_dev->data->dev_flags |= restore;
2394 		rte_eth_dev_probing_finish(list[i].eth_dev);
2395 	}
2396 	if (i != ns) {
2397 		DRV_LOG(ERR,
2398 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2399 			" encountering an error: %s",
2400 			owner_pci.domain, owner_pci.bus,
2401 			owner_pci.devid, owner_pci.function,
2402 			strerror(rte_errno));
2403 		ret = -rte_errno;
2404 		/* Roll back. */
2405 		while (i--) {
2406 			if (!list[i].eth_dev)
2407 				continue;
2408 			mlx5_dev_close(list[i].eth_dev);
2409 			/* mac_addrs must not be freed because in dev_private */
2410 			list[i].eth_dev->data->mac_addrs = NULL;
2411 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2412 		}
2413 		/* Restore original error. */
2414 		rte_errno = -ret;
2415 	} else {
2416 		ret = 0;
2417 	}
2418 exit:
2419 	/*
2420 	 * Do the routine cleanup:
2421 	 * - close opened Netlink sockets
2422 	 * - free allocated spawn data array
2423 	 * - free the Infiniband device list
2424 	 */
2425 	if (nl_rdma >= 0)
2426 		close(nl_rdma);
2427 	if (nl_route >= 0)
2428 		close(nl_route);
2429 	if (list)
2430 		mlx5_free(list);
2431 	MLX5_ASSERT(ibv_list);
2432 	mlx5_glue->free_device_list(ibv_list);
2433 	return ret;
2434 }
2435 
2436 /**
2437  * DPDK callback to register a PCI device.
2438  *
2439  * This function spawns Ethernet devices out of a given PCI device.
2440  *
2441  * @param[in] pci_drv
2442  *   PCI driver structure (mlx5_driver).
2443  * @param[in] pci_dev
2444  *   PCI device information.
2445  *
2446  * @return
2447  *   0 on success, a negative errno value otherwise and rte_errno is set.
2448  */
2449 int
2450 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2451 		  struct rte_pci_device *pci_dev)
2452 {
2453 	struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2454 	int ret = 0;
2455 	uint16_t p;
2456 
2457 	if (pci_dev->device.devargs) {
2458 		/* Parse representor information from device argument. */
2459 		if (pci_dev->device.devargs->cls_str)
2460 			ret = rte_eth_devargs_parse
2461 				(pci_dev->device.devargs->cls_str, &eth_da);
2462 		if (ret) {
2463 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2464 				pci_dev->device.devargs->cls_str);
2465 			return -rte_errno;
2466 		}
2467 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2468 			/* Support legacy device argument */
2469 			ret = rte_eth_devargs_parse
2470 				(pci_dev->device.devargs->args, &eth_da);
2471 			if (ret) {
2472 				DRV_LOG(ERR, "failed to parse device arguments: %s",
2473 					pci_dev->device.devargs->args);
2474 				return -rte_errno;
2475 			}
2476 		}
2477 	}
2478 
2479 	if (eth_da.nb_ports > 0) {
2480 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2481 		for (p = 0; p < eth_da.nb_ports; p++)
2482 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2483 						   eth_da.ports[p]);
2484 	} else {
2485 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2486 	}
2487 	return ret;
2488 }
2489 
2490 static int
2491 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2492 {
2493 	char *env;
2494 	int value;
2495 
2496 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2497 	/* Get environment variable to store. */
2498 	env = getenv(MLX5_SHUT_UP_BF);
2499 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2500 	if (config->dbnc == MLX5_ARG_UNSET)
2501 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2502 	else
2503 		setenv(MLX5_SHUT_UP_BF,
2504 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2505 	return value;
2506 }
2507 
2508 static void
2509 mlx5_restore_doorbell_mapping_env(int value)
2510 {
2511 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2512 	/* Restore the original environment variable state. */
2513 	if (value == MLX5_ARG_UNSET)
2514 		unsetenv(MLX5_SHUT_UP_BF);
2515 	else
2516 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2517 }
2518 
2519 /**
2520  * Extract pdn of PD object using DV API.
2521  *
2522  * @param[in] pd
2523  *   Pointer to the verbs PD object.
2524  * @param[out] pdn
2525  *   Pointer to the PD object number variable.
2526  *
2527  * @return
2528  *   0 on success, error value otherwise.
2529  */
2530 int
2531 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2532 {
2533 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2534 	struct mlx5dv_obj obj;
2535 	struct mlx5dv_pd pd_info;
2536 	int ret = 0;
2537 
2538 	obj.pd.in = pd;
2539 	obj.pd.out = &pd_info;
2540 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2541 	if (ret) {
2542 		DRV_LOG(DEBUG, "Fail to get PD object info");
2543 		return ret;
2544 	}
2545 	*pdn = pd_info.pdn;
2546 	return 0;
2547 #else
2548 	(void)pd;
2549 	(void)pdn;
2550 	return -ENOTSUP;
2551 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2552 }
2553 
2554 /**
2555  * Function API to open IB device.
2556  *
2557  * This function calls the Linux glue APIs to open a device.
2558  *
2559  * @param[in] spawn
2560  *   Pointer to the IB device attributes (name, port, etc).
2561  * @param[out] config
2562  *   Pointer to device configuration structure.
2563  * @param[out] sh
2564  *   Pointer to shared context structure.
2565  *
2566  * @return
2567  *   0 on success, a positive error value otherwise.
2568  */
2569 int
2570 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2571 		     const struct mlx5_dev_config *config,
2572 		     struct mlx5_dev_ctx_shared *sh)
2573 {
2574 	int dbmap_env;
2575 	int err = 0;
2576 
2577 	sh->numa_node = spawn->pci_dev->device.numa_node;
2578 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2579 	/*
2580 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2581 	 * before the device creation. The rdma_core library
2582 	 * checks the variable at device creation and
2583 	 * stores the result internally.
2584 	 */
2585 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2586 	/* Try to open IB device with DV first, then usual Verbs. */
2587 	errno = 0;
2588 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2589 	if (sh->ctx) {
2590 		sh->devx = 1;
2591 		DRV_LOG(DEBUG, "DevX is supported");
2592 		/* The device is created, no need for environment. */
2593 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2594 	} else {
2595 		/* The environment variable is still configured. */
2596 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2597 		err = errno ? errno : ENODEV;
2598 		/*
2599 		 * The environment variable is not needed anymore,
2600 		 * all device creation attempts are completed.
2601 		 */
2602 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2603 		if (!sh->ctx)
2604 			return err;
2605 		DRV_LOG(DEBUG, "DevX is NOT supported");
2606 		err = 0;
2607 	}
2608 	if (!err && sh->ctx) {
2609 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2610 		mlx5_glue->dv_set_context_attr(sh->ctx,
2611 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2612 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2613 				.alloc = &mlx5_alloc_verbs_buf,
2614 				.free = &mlx5_free_verbs_buf,
2615 				.data = sh,
2616 			}));
2617 	}
2618 	return err;
2619 }
2620 
2621 /**
2622  * Install shared asynchronous device events handler.
2623  * This function is implemented to support event sharing
2624  * between multiple ports of single IB device.
2625  *
2626  * @param sh
2627  *   Pointer to mlx5_dev_ctx_shared object.
2628  */
2629 void
2630 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2631 {
2632 	int ret;
2633 	int flags;
2634 
2635 	sh->intr_handle.fd = -1;
2636 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2637 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2638 		    F_SETFL, flags | O_NONBLOCK);
2639 	if (ret) {
2640 		DRV_LOG(INFO, "failed to change file descriptor async event"
2641 			" queue");
2642 	} else {
2643 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2644 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2645 		if (rte_intr_callback_register(&sh->intr_handle,
2646 					mlx5_dev_interrupt_handler, sh)) {
2647 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2648 			sh->intr_handle.fd = -1;
2649 		}
2650 	}
2651 	if (sh->devx) {
2652 #ifdef HAVE_IBV_DEVX_ASYNC
2653 		sh->intr_handle_devx.fd = -1;
2654 		sh->devx_comp =
2655 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2656 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2657 		if (!devx_comp) {
2658 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2659 			return;
2660 		}
2661 		flags = fcntl(devx_comp->fd, F_GETFL);
2662 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2663 		if (ret) {
2664 			DRV_LOG(INFO, "failed to change file descriptor"
2665 				" devx comp");
2666 			return;
2667 		}
2668 		sh->intr_handle_devx.fd = devx_comp->fd;
2669 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2670 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2671 					mlx5_dev_interrupt_handler_devx, sh)) {
2672 			DRV_LOG(INFO, "Fail to install the devx shared"
2673 				" interrupt.");
2674 			sh->intr_handle_devx.fd = -1;
2675 		}
2676 #endif /* HAVE_IBV_DEVX_ASYNC */
2677 	}
2678 }
2679 
2680 /**
2681  * Uninstall shared asynchronous device events handler.
2682  * This function is implemented to support event sharing
2683  * between multiple ports of single IB device.
2684  *
2685  * @param dev
2686  *   Pointer to mlx5_dev_ctx_shared object.
2687  */
2688 void
2689 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2690 {
2691 	if (sh->intr_handle.fd >= 0)
2692 		mlx5_intr_callback_unregister(&sh->intr_handle,
2693 					      mlx5_dev_interrupt_handler, sh);
2694 #ifdef HAVE_IBV_DEVX_ASYNC
2695 	if (sh->intr_handle_devx.fd >= 0)
2696 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2697 				  mlx5_dev_interrupt_handler_devx, sh);
2698 	if (sh->devx_comp)
2699 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2700 #endif
2701 }
2702 
2703 /**
2704  * Read statistics by a named counter.
2705  *
2706  * @param[in] priv
2707  *   Pointer to the private device data structure.
2708  * @param[in] ctr_name
2709  *   Pointer to the name of the statistic counter to read
2710  * @param[out] stat
2711  *   Pointer to read statistic value.
2712  * @return
2713  *   0 on success and stat is valud, 1 if failed to read the value
2714  *   rte_errno is set.
2715  *
2716  */
2717 int
2718 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2719 		      uint64_t *stat)
2720 {
2721 	int fd;
2722 
2723 	if (priv->sh) {
2724 		if (priv->q_counters != NULL &&
2725 		    strcmp(ctr_name, "out_of_buffer") == 0)
2726 			return mlx5_devx_cmd_queue_counter_query
2727 					(priv->q_counters, 0, (uint32_t *)stat);
2728 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2729 		      priv->sh->ibdev_path,
2730 		      priv->dev_port,
2731 		      ctr_name);
2732 		fd = open(path, O_RDONLY);
2733 		/*
2734 		 * in switchdev the file location is not per port
2735 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2736 		 */
2737 		if (fd == -1) {
2738 			MKSTR(path1, "%s/hw_counters/%s",
2739 			      priv->sh->ibdev_path,
2740 			      ctr_name);
2741 			fd = open(path1, O_RDONLY);
2742 		}
2743 		if (fd != -1) {
2744 			char buf[21] = {'\0'};
2745 			ssize_t n = read(fd, buf, sizeof(buf));
2746 
2747 			close(fd);
2748 			if (n != -1) {
2749 				*stat = strtoull(buf, NULL, 10);
2750 				return 0;
2751 			}
2752 		}
2753 	}
2754 	*stat = 0;
2755 	return 1;
2756 }
2757 
2758 /**
2759  * Set the reg_mr and dereg_mr call backs
2760  *
2761  * @param reg_mr_cb[out]
2762  *   Pointer to reg_mr func
2763  * @param dereg_mr_cb[out]
2764  *   Pointer to dereg_mr func
2765  *
2766  */
2767 void
2768 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2769 		      mlx5_dereg_mr_t *dereg_mr_cb)
2770 {
2771 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2772 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2773 }
2774 
2775 /**
2776  * Remove a MAC address from device
2777  *
2778  * @param dev
2779  *   Pointer to Ethernet device structure.
2780  * @param index
2781  *   MAC address index.
2782  */
2783 void
2784 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2785 {
2786 	struct mlx5_priv *priv = dev->data->dev_private;
2787 	const int vf = priv->config.vf;
2788 
2789 	if (vf)
2790 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2791 					mlx5_ifindex(dev), priv->mac_own,
2792 					&dev->data->mac_addrs[index], index);
2793 }
2794 
2795 /**
2796  * Adds a MAC address to the device
2797  *
2798  * @param dev
2799  *   Pointer to Ethernet device structure.
2800  * @param mac_addr
2801  *   MAC address to register.
2802  * @param index
2803  *   MAC address index.
2804  *
2805  * @return
2806  *   0 on success, a negative errno value otherwise
2807  */
2808 int
2809 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2810 		     uint32_t index)
2811 {
2812 	struct mlx5_priv *priv = dev->data->dev_private;
2813 	const int vf = priv->config.vf;
2814 	int ret = 0;
2815 
2816 	if (vf)
2817 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2818 					   mlx5_ifindex(dev), priv->mac_own,
2819 					   mac, index);
2820 	return ret;
2821 }
2822 
2823 /**
2824  * Modify a VF MAC address
2825  *
2826  * @param priv
2827  *   Pointer to device private data.
2828  * @param mac_addr
2829  *   MAC address to modify into.
2830  * @param iface_idx
2831  *   Net device interface index
2832  * @param vf_index
2833  *   VF index
2834  *
2835  * @return
2836  *   0 on success, a negative errno value otherwise
2837  */
2838 int
2839 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2840 			   unsigned int iface_idx,
2841 			   struct rte_ether_addr *mac_addr,
2842 			   int vf_index)
2843 {
2844 	return mlx5_nl_vf_mac_addr_modify
2845 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2846 }
2847 
2848 /**
2849  * Set device promiscuous mode
2850  *
2851  * @param dev
2852  *   Pointer to Ethernet device structure.
2853  * @param enable
2854  *   0 - promiscuous is disabled, otherwise - enabled
2855  *
2856  * @return
2857  *   0 on success, a negative error value otherwise
2858  */
2859 int
2860 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2861 {
2862 	struct mlx5_priv *priv = dev->data->dev_private;
2863 
2864 	return mlx5_nl_promisc(priv->nl_socket_route,
2865 			       mlx5_ifindex(dev), !!enable);
2866 }
2867 
2868 /**
2869  * Set device promiscuous mode
2870  *
2871  * @param dev
2872  *   Pointer to Ethernet device structure.
2873  * @param enable
2874  *   0 - all multicase is disabled, otherwise - enabled
2875  *
2876  * @return
2877  *   0 on success, a negative error value otherwise
2878  */
2879 int
2880 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2881 {
2882 	struct mlx5_priv *priv = dev->data->dev_private;
2883 
2884 	return mlx5_nl_allmulti(priv->nl_socket_route,
2885 				mlx5_ifindex(dev), !!enable);
2886 }
2887 
2888 /**
2889  * Flush device MAC addresses
2890  *
2891  * @param dev
2892  *   Pointer to Ethernet device structure.
2893  *
2894  */
2895 void
2896 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2897 {
2898 	struct mlx5_priv *priv = dev->data->dev_private;
2899 
2900 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2901 			       dev->data->mac_addrs,
2902 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2903 }
2904