xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision a295c69a8b24f34f9d98959853265b31bcbcc133)
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 #endif /* HAVE_MLX5DV_DR */
359 	sh->default_miss_action =
360 			mlx5_glue->dr_create_flow_action_default_miss();
361 	if (!sh->default_miss_action)
362 		DRV_LOG(WARNING, "Default miss action is not supported.");
363 	return 0;
364 error:
365 	/* Rollback the created objects. */
366 	if (sh->rx_domain) {
367 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
368 		sh->rx_domain = NULL;
369 	}
370 	if (sh->tx_domain) {
371 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
372 		sh->tx_domain = NULL;
373 	}
374 	if (sh->fdb_domain) {
375 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
376 		sh->fdb_domain = NULL;
377 	}
378 	if (sh->dr_drop_action) {
379 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
380 		sh->dr_drop_action = NULL;
381 	}
382 	if (sh->pop_vlan_action) {
383 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
384 		sh->pop_vlan_action = NULL;
385 	}
386 	if (sh->encaps_decaps) {
387 		mlx5_hlist_destroy(sh->encaps_decaps);
388 		sh->encaps_decaps = NULL;
389 	}
390 	if (sh->modify_cmds) {
391 		mlx5_hlist_destroy(sh->modify_cmds);
392 		sh->modify_cmds = NULL;
393 	}
394 	if (sh->tag_table) {
395 		/* tags should be destroyed with flow before. */
396 		mlx5_hlist_destroy(sh->tag_table);
397 		sh->tag_table = NULL;
398 	}
399 	if (sh->tunnel_hub) {
400 		mlx5_release_tunnel_hub(sh, priv->dev_port);
401 		sh->tunnel_hub = NULL;
402 	}
403 	mlx5_free_table_hash_list(priv);
404 	return err;
405 }
406 
407 /**
408  * Destroy DR related data within private structure.
409  *
410  * @param[in] priv
411  *   Pointer to the private device data structure.
412  */
413 void
414 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
415 {
416 	struct mlx5_dev_ctx_shared *sh = priv->sh;
417 
418 	MLX5_ASSERT(sh && sh->refcnt);
419 	if (sh->refcnt > 1)
420 		return;
421 #ifdef HAVE_MLX5DV_DR
422 	if (sh->rx_domain) {
423 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
424 		sh->rx_domain = NULL;
425 	}
426 	if (sh->tx_domain) {
427 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
428 		sh->tx_domain = NULL;
429 	}
430 #ifdef HAVE_MLX5DV_DR_ESWITCH
431 	if (sh->fdb_domain) {
432 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
433 		sh->fdb_domain = NULL;
434 	}
435 	if (sh->dr_drop_action) {
436 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
437 		sh->dr_drop_action = NULL;
438 	}
439 #endif
440 	if (sh->pop_vlan_action) {
441 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
442 		sh->pop_vlan_action = NULL;
443 	}
444 #endif /* HAVE_MLX5DV_DR */
445 	if (sh->default_miss_action)
446 		mlx5_glue->destroy_flow_action
447 				(sh->default_miss_action);
448 	if (sh->encaps_decaps) {
449 		mlx5_hlist_destroy(sh->encaps_decaps);
450 		sh->encaps_decaps = NULL;
451 	}
452 	if (sh->modify_cmds) {
453 		mlx5_hlist_destroy(sh->modify_cmds);
454 		sh->modify_cmds = NULL;
455 	}
456 	if (sh->tag_table) {
457 		/* tags should be destroyed with flow before. */
458 		mlx5_hlist_destroy(sh->tag_table);
459 		sh->tag_table = NULL;
460 	}
461 	if (sh->tunnel_hub) {
462 		mlx5_release_tunnel_hub(sh, priv->dev_port);
463 		sh->tunnel_hub = NULL;
464 	}
465 	mlx5_cache_list_destroy(&sh->port_id_action_list);
466 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
467 	mlx5_free_table_hash_list(priv);
468 }
469 
470 /**
471  * Initialize shared data between primary and secondary process.
472  *
473  * A memzone is reserved by primary process and secondary processes attach to
474  * the memzone.
475  *
476  * @return
477  *   0 on success, a negative errno value otherwise and rte_errno is set.
478  */
479 static int
480 mlx5_init_shared_data(void)
481 {
482 	const struct rte_memzone *mz;
483 	int ret = 0;
484 
485 	rte_spinlock_lock(&mlx5_shared_data_lock);
486 	if (mlx5_shared_data == NULL) {
487 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
488 			/* Allocate shared memory. */
489 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
490 						 sizeof(*mlx5_shared_data),
491 						 SOCKET_ID_ANY, 0);
492 			if (mz == NULL) {
493 				DRV_LOG(ERR,
494 					"Cannot allocate mlx5 shared data");
495 				ret = -rte_errno;
496 				goto error;
497 			}
498 			mlx5_shared_data = mz->addr;
499 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
500 			rte_spinlock_init(&mlx5_shared_data->lock);
501 		} else {
502 			/* Lookup allocated shared memory. */
503 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
504 			if (mz == NULL) {
505 				DRV_LOG(ERR,
506 					"Cannot attach mlx5 shared data");
507 				ret = -rte_errno;
508 				goto error;
509 			}
510 			mlx5_shared_data = mz->addr;
511 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
512 		}
513 	}
514 error:
515 	rte_spinlock_unlock(&mlx5_shared_data_lock);
516 	return ret;
517 }
518 
519 /**
520  * PMD global initialization.
521  *
522  * Independent from individual device, this function initializes global
523  * per-PMD data structures distinguishing primary and secondary processes.
524  * Hence, each initialization is called once per a process.
525  *
526  * @return
527  *   0 on success, a negative errno value otherwise and rte_errno is set.
528  */
529 static int
530 mlx5_init_once(void)
531 {
532 	struct mlx5_shared_data *sd;
533 	struct mlx5_local_data *ld = &mlx5_local_data;
534 	int ret = 0;
535 
536 	if (mlx5_init_shared_data())
537 		return -rte_errno;
538 	sd = mlx5_shared_data;
539 	MLX5_ASSERT(sd);
540 	rte_spinlock_lock(&sd->lock);
541 	switch (rte_eal_process_type()) {
542 	case RTE_PROC_PRIMARY:
543 		if (sd->init_done)
544 			break;
545 		LIST_INIT(&sd->mem_event_cb_list);
546 		rte_rwlock_init(&sd->mem_event_rwlock);
547 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
548 						mlx5_mr_mem_event_cb, NULL);
549 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
550 					   mlx5_mp_os_primary_handle);
551 		if (ret)
552 			goto out;
553 		sd->init_done = true;
554 		break;
555 	case RTE_PROC_SECONDARY:
556 		if (ld->init_done)
557 			break;
558 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
559 					     mlx5_mp_os_secondary_handle);
560 		if (ret)
561 			goto out;
562 		++sd->secondary_cnt;
563 		ld->init_done = true;
564 		break;
565 	default:
566 		break;
567 	}
568 out:
569 	rte_spinlock_unlock(&sd->lock);
570 	return ret;
571 }
572 
573 /**
574  * Create the Tx queue DevX/Verbs object.
575  *
576  * @param dev
577  *   Pointer to Ethernet device.
578  * @param idx
579  *   Queue index in DPDK Tx queue array.
580  *
581  * @return
582  *   0 on success, a negative errno value otherwise and rte_errno is set.
583  */
584 static int
585 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
586 {
587 	struct mlx5_priv *priv = dev->data->dev_private;
588 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
589 	struct mlx5_txq_ctrl *txq_ctrl =
590 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
591 
592 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
593 		return mlx5_txq_devx_obj_new(dev, idx);
594 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
595 	if (!priv->config.dv_esw_en)
596 		return mlx5_txq_devx_obj_new(dev, idx);
597 #endif
598 	return mlx5_txq_ibv_obj_new(dev, idx);
599 }
600 
601 /**
602  * Release an Tx DevX/verbs queue object.
603  *
604  * @param txq_obj
605  *   DevX/Verbs Tx queue object.
606  */
607 static void
608 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
609 {
610 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
611 		mlx5_txq_devx_obj_release(txq_obj);
612 		return;
613 	}
614 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
615 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
616 		mlx5_txq_devx_obj_release(txq_obj);
617 		return;
618 	}
619 #endif
620 	mlx5_txq_ibv_obj_release(txq_obj);
621 }
622 
623 /**
624  * DV flow counter mode detect and config.
625  *
626  * @param dev
627  *   Pointer to rte_eth_dev structure.
628  *
629  */
630 static void
631 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
632 {
633 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
634 	struct mlx5_priv *priv = dev->data->dev_private;
635 	struct mlx5_dev_ctx_shared *sh = priv->sh;
636 	bool fallback;
637 
638 #ifndef HAVE_IBV_DEVX_ASYNC
639 	fallback = true;
640 #else
641 	fallback = false;
642 	if (!priv->config.devx || !priv->config.dv_flow_en ||
643 	    !priv->config.hca_attr.flow_counters_dump ||
644 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
645 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
646 		fallback = true;
647 #endif
648 	if (fallback)
649 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
650 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
651 			priv->config.hca_attr.flow_counters_dump,
652 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
653 	/* Initialize fallback mode only on the port initializes sh. */
654 	if (sh->refcnt == 1)
655 		sh->cmng.counter_fallback = fallback;
656 	else if (fallback != sh->cmng.counter_fallback)
657 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
658 			"with others:%d.", PORT_ID(priv), fallback);
659 #endif
660 }
661 
662 static void
663 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
664 {
665 	struct mlx5_priv *priv = dev->data->dev_private;
666 	void *ctx = priv->sh->ctx;
667 
668 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
669 	if (!priv->q_counters) {
670 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
671 		struct ibv_wq *wq;
672 
673 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
674 			"by DevX - fall-back to use the kernel driver global "
675 			"queue counter.", dev->data->port_id);
676 		/* Create WQ by kernel and query its queue counter ID. */
677 		if (cq) {
678 			wq = mlx5_glue->create_wq(ctx,
679 						  &(struct ibv_wq_init_attr){
680 						    .wq_type = IBV_WQT_RQ,
681 						    .max_wr = 1,
682 						    .max_sge = 1,
683 						    .pd = priv->sh->pd,
684 						    .cq = cq,
685 						});
686 			if (wq) {
687 				/* Counter is assigned only on RDY state. */
688 				int ret = mlx5_glue->modify_wq(wq,
689 						 &(struct ibv_wq_attr){
690 						 .attr_mask = IBV_WQ_ATTR_STATE,
691 						 .wq_state = IBV_WQS_RDY,
692 						});
693 
694 				if (ret == 0)
695 					mlx5_devx_cmd_wq_query(wq,
696 							 &priv->counter_set_id);
697 				claim_zero(mlx5_glue->destroy_wq(wq));
698 			}
699 			claim_zero(mlx5_glue->destroy_cq(cq));
700 		}
701 	} else {
702 		priv->counter_set_id = priv->q_counters->id;
703 	}
704 	if (priv->counter_set_id == 0)
705 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
706 			"available.", dev->data->port_id);
707 }
708 
709 /**
710  * Check if representor spawn info match devargs.
711  *
712  * @param spawn
713  *   Verbs device parameters (name, port, switch_info) to spawn.
714  * @param eth_da
715  *   Device devargs to probe.
716  *
717  * @return
718  *   Match result.
719  */
720 static bool
721 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
722 		       struct rte_eth_devargs *eth_da)
723 {
724 	struct mlx5_switch_info *switch_info = &spawn->info;
725 	unsigned int p, f;
726 	uint16_t id;
727 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
728 						      eth_da->type);
729 
730 	switch (eth_da->type) {
731 	case RTE_ETH_REPRESENTOR_SF:
732 		if (!(spawn->info.port_name == -1 &&
733 		      switch_info->name_type ==
734 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
735 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
736 			rte_errno = EBUSY;
737 			return false;
738 		}
739 		break;
740 	case RTE_ETH_REPRESENTOR_VF:
741 		/* Allows HPF representor index -1 as exception. */
742 		if (!(spawn->info.port_name == -1 &&
743 		      switch_info->name_type ==
744 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
745 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
746 			rte_errno = EBUSY;
747 			return false;
748 		}
749 		break;
750 	case RTE_ETH_REPRESENTOR_NONE:
751 		rte_errno = EBUSY;
752 		return false;
753 	default:
754 		rte_errno = ENOTSUP;
755 		DRV_LOG(ERR, "unsupported representor type");
756 		return false;
757 	}
758 	/* Check representor ID: */
759 	for (p = 0; p < eth_da->nb_ports; ++p) {
760 		if (spawn->pf_bond < 0) {
761 			/* For non-LAG mode, allow and ignore pf. */
762 			switch_info->pf_num = eth_da->ports[p];
763 			repr_id = mlx5_representor_id_encode(switch_info,
764 							     eth_da->type);
765 		}
766 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
767 			id = MLX5_REPRESENTOR_ID
768 				(eth_da->ports[p], eth_da->type,
769 				 eth_da->representor_ports[f]);
770 			if (repr_id == id)
771 				return true;
772 		}
773 	}
774 	rte_errno = EBUSY;
775 	return false;
776 }
777 
778 
779 /**
780  * Spawn an Ethernet device from Verbs information.
781  *
782  * @param dpdk_dev
783  *   Backing DPDK device.
784  * @param spawn
785  *   Verbs device parameters (name, port, switch_info) to spawn.
786  * @param config
787  *   Device configuration parameters.
788  * @param config
789  *   Device arguments.
790  *
791  * @return
792  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
793  *   is set. The following errors are defined:
794  *
795  *   EBUSY: device is not supposed to be spawned.
796  *   EEXIST: device is already spawned
797  */
798 static struct rte_eth_dev *
799 mlx5_dev_spawn(struct rte_device *dpdk_dev,
800 	       struct mlx5_dev_spawn_data *spawn,
801 	       struct mlx5_dev_config *config,
802 	       struct rte_eth_devargs *eth_da)
803 {
804 	const struct mlx5_switch_info *switch_info = &spawn->info;
805 	struct mlx5_dev_ctx_shared *sh = NULL;
806 	struct ibv_port_attr port_attr;
807 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
808 	struct rte_eth_dev *eth_dev = NULL;
809 	struct mlx5_priv *priv = NULL;
810 	int err = 0;
811 	unsigned int hw_padding = 0;
812 	unsigned int mps;
813 	unsigned int tunnel_en = 0;
814 	unsigned int mpls_en = 0;
815 	unsigned int swp = 0;
816 	unsigned int mprq = 0;
817 	unsigned int mprq_min_stride_size_n = 0;
818 	unsigned int mprq_max_stride_size_n = 0;
819 	unsigned int mprq_min_stride_num_n = 0;
820 	unsigned int mprq_max_stride_num_n = 0;
821 	struct rte_ether_addr mac;
822 	char name[RTE_ETH_NAME_MAX_LEN];
823 	int own_domain_id = 0;
824 	uint16_t port_id;
825 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
826 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
827 #endif
828 
829 	/* Determine if this port representor is supposed to be spawned. */
830 	if (switch_info->representor && dpdk_dev->devargs &&
831 	    !mlx5_representor_match(spawn, eth_da))
832 		return NULL;
833 	/* Build device name. */
834 	if (spawn->pf_bond < 0) {
835 		/* Single device. */
836 		if (!switch_info->representor)
837 			strlcpy(name, dpdk_dev->name, sizeof(name));
838 		else
839 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
840 				 dpdk_dev->name,
841 				 switch_info->name_type ==
842 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
843 				 switch_info->port_name);
844 	} else {
845 		/* Bonding device. */
846 		if (!switch_info->representor) {
847 			err = snprintf(name, sizeof(name), "%s_%s",
848 				 dpdk_dev->name,
849 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
850 		} else {
851 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
852 				dpdk_dev->name,
853 				mlx5_os_get_dev_device_name(spawn->phys_dev),
854 				switch_info->ctrl_num,
855 				switch_info->pf_num,
856 				switch_info->name_type ==
857 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
858 				switch_info->port_name);
859 		}
860 	}
861 	if (err >= (int)sizeof(name))
862 		DRV_LOG(WARNING, "device name overflow %s", name);
863 	/* check if the device is already spawned */
864 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
865 		rte_errno = EEXIST;
866 		return NULL;
867 	}
868 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
869 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
870 		struct mlx5_mp_id mp_id;
871 
872 		eth_dev = rte_eth_dev_attach_secondary(name);
873 		if (eth_dev == NULL) {
874 			DRV_LOG(ERR, "can not attach rte ethdev");
875 			rte_errno = ENOMEM;
876 			return NULL;
877 		}
878 		eth_dev->device = dpdk_dev;
879 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
880 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
881 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
882 		err = mlx5_proc_priv_init(eth_dev);
883 		if (err)
884 			return NULL;
885 		mp_id.port_id = eth_dev->data->port_id;
886 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
887 		/* Receive command fd from primary process */
888 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
889 		if (err < 0)
890 			goto err_secondary;
891 		/* Remap UAR for Tx queues. */
892 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
893 		if (err)
894 			goto err_secondary;
895 		/*
896 		 * Ethdev pointer is still required as input since
897 		 * the primary device is not accessible from the
898 		 * secondary process.
899 		 */
900 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
901 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
902 		return eth_dev;
903 err_secondary:
904 		mlx5_dev_close(eth_dev);
905 		return NULL;
906 	}
907 	/*
908 	 * Some parameters ("tx_db_nc" in particularly) are needed in
909 	 * advance to create dv/verbs device context. We proceed the
910 	 * devargs here to get ones, and later proceed devargs again
911 	 * to override some hardware settings.
912 	 */
913 	err = mlx5_args(config, dpdk_dev->devargs);
914 	if (err) {
915 		err = rte_errno;
916 		DRV_LOG(ERR, "failed to process device arguments: %s",
917 			strerror(rte_errno));
918 		goto error;
919 	}
920 	if (config->dv_miss_info) {
921 		if (switch_info->master || switch_info->representor)
922 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
923 	}
924 	mlx5_malloc_mem_select(config->sys_mem_en);
925 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
926 	if (!sh)
927 		return NULL;
928 	config->devx = sh->devx;
929 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
930 	config->dest_tir = 1;
931 #endif
932 #ifdef HAVE_IBV_MLX5_MOD_SWP
933 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
934 #endif
935 	/*
936 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
937 	 * as all ConnectX-5 devices.
938 	 */
939 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
940 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
941 #endif
942 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
943 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
944 #endif
945 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
946 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
947 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
948 			DRV_LOG(DEBUG, "enhanced MPW is supported");
949 			mps = MLX5_MPW_ENHANCED;
950 		} else {
951 			DRV_LOG(DEBUG, "MPW is supported");
952 			mps = MLX5_MPW;
953 		}
954 	} else {
955 		DRV_LOG(DEBUG, "MPW isn't supported");
956 		mps = MLX5_MPW_DISABLED;
957 	}
958 #ifdef HAVE_IBV_MLX5_MOD_SWP
959 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
960 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
961 	DRV_LOG(DEBUG, "SWP support: %u", swp);
962 #endif
963 	config->swp = !!swp;
964 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
965 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
966 		struct mlx5dv_striding_rq_caps mprq_caps =
967 			dv_attr.striding_rq_caps;
968 
969 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
970 			mprq_caps.min_single_stride_log_num_of_bytes);
971 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
972 			mprq_caps.max_single_stride_log_num_of_bytes);
973 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
974 			mprq_caps.min_single_wqe_log_num_of_strides);
975 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
976 			mprq_caps.max_single_wqe_log_num_of_strides);
977 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
978 			mprq_caps.supported_qpts);
979 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
980 		mprq = 1;
981 		mprq_min_stride_size_n =
982 			mprq_caps.min_single_stride_log_num_of_bytes;
983 		mprq_max_stride_size_n =
984 			mprq_caps.max_single_stride_log_num_of_bytes;
985 		mprq_min_stride_num_n =
986 			mprq_caps.min_single_wqe_log_num_of_strides;
987 		mprq_max_stride_num_n =
988 			mprq_caps.max_single_wqe_log_num_of_strides;
989 	}
990 #endif
991 	/* Rx CQE compression is enabled by default. */
992 	config->cqe_comp = 1;
993 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
994 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
995 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
996 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
997 			     (dv_attr.tunnel_offloads_caps &
998 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
999 			     (dv_attr.tunnel_offloads_caps &
1000 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1001 	}
1002 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1003 		tunnel_en ? "" : "not ");
1004 #else
1005 	DRV_LOG(WARNING,
1006 		"tunnel offloading disabled due to old OFED/rdma-core version");
1007 #endif
1008 	config->tunnel_en = tunnel_en;
1009 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1010 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1011 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1012 		   (dv_attr.tunnel_offloads_caps &
1013 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1014 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1015 		mpls_en ? "" : "not ");
1016 #else
1017 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1018 		" old OFED/rdma-core version or firmware configuration");
1019 #endif
1020 	config->mpls_en = mpls_en;
1021 	/* Check port status. */
1022 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1023 	if (err) {
1024 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1025 		goto error;
1026 	}
1027 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1028 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1029 		err = EINVAL;
1030 		goto error;
1031 	}
1032 	if (port_attr.state != IBV_PORT_ACTIVE)
1033 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1034 			mlx5_glue->port_state_str(port_attr.state),
1035 			port_attr.state);
1036 	/* Allocate private eth device data. */
1037 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1038 			   sizeof(*priv),
1039 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1040 	if (priv == NULL) {
1041 		DRV_LOG(ERR, "priv allocation failure");
1042 		err = ENOMEM;
1043 		goto error;
1044 	}
1045 	priv->sh = sh;
1046 	priv->dev_port = spawn->phys_port;
1047 	priv->pci_dev = spawn->pci_dev;
1048 	priv->mtu = RTE_ETHER_MTU;
1049 	/* Some internal functions rely on Netlink sockets, open them now. */
1050 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1051 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1052 	priv->representor = !!switch_info->representor;
1053 	priv->master = !!switch_info->master;
1054 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1055 	priv->vport_meta_tag = 0;
1056 	priv->vport_meta_mask = 0;
1057 	priv->pf_bond = spawn->pf_bond;
1058 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1059 	/*
1060 	 * The DevX port query API is implemented. E-Switch may use
1061 	 * either vport or reg_c[0] metadata register to match on
1062 	 * vport index. The engaged part of metadata register is
1063 	 * defined by mask.
1064 	 */
1065 	if (switch_info->representor || switch_info->master) {
1066 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1067 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1068 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1069 						 &devx_port);
1070 		if (err) {
1071 			DRV_LOG(WARNING,
1072 				"can't query devx port %d on device %s",
1073 				spawn->phys_port,
1074 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1075 			devx_port.comp_mask = 0;
1076 		}
1077 	}
1078 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1079 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1080 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1081 		if (!priv->vport_meta_mask) {
1082 			DRV_LOG(ERR, "vport zero mask for port %d"
1083 				     " on bonding device %s",
1084 				     spawn->phys_port,
1085 				     mlx5_os_get_dev_device_name
1086 							(spawn->phys_dev));
1087 			err = ENOTSUP;
1088 			goto error;
1089 		}
1090 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1091 			DRV_LOG(ERR, "invalid vport tag for port %d"
1092 				     " on bonding device %s",
1093 				     spawn->phys_port,
1094 				     mlx5_os_get_dev_device_name
1095 							(spawn->phys_dev));
1096 			err = ENOTSUP;
1097 			goto error;
1098 		}
1099 	}
1100 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1101 		priv->vport_id = devx_port.vport_num;
1102 	} else if (spawn->pf_bond >= 0 &&
1103 		   (switch_info->representor || switch_info->master)) {
1104 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1105 			     " on bonding device %s",
1106 			     spawn->phys_port,
1107 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1108 		err = ENOTSUP;
1109 		goto error;
1110 	} else {
1111 		/* Suppose vport index in compatible way. */
1112 		priv->vport_id = switch_info->representor ?
1113 				 switch_info->port_name + 1 : -1;
1114 	}
1115 #else
1116 	/*
1117 	 * Kernel/rdma_core support single E-Switch per PF configurations
1118 	 * only and vport_id field contains the vport index for
1119 	 * associated VF, which is deduced from representor port name.
1120 	 * For example, let's have the IB device port 10, it has
1121 	 * attached network device eth0, which has port name attribute
1122 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1123 	 * as 3 (2+1). This assigning schema should be changed if the
1124 	 * multiple E-Switch instances per PF configurations or/and PCI
1125 	 * subfunctions are added.
1126 	 */
1127 	priv->vport_id = switch_info->representor ?
1128 			 switch_info->port_name + 1 : -1;
1129 #endif
1130 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1131 							  eth_da->type);
1132 	/*
1133 	 * Look for sibling devices in order to reuse their switch domain
1134 	 * if any, otherwise allocate one.
1135 	 */
1136 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1137 		const struct mlx5_priv *opriv =
1138 			rte_eth_devices[port_id].data->dev_private;
1139 
1140 		if (!opriv ||
1141 		    opriv->sh != priv->sh ||
1142 			opriv->domain_id ==
1143 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1144 			continue;
1145 		priv->domain_id = opriv->domain_id;
1146 		break;
1147 	}
1148 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1149 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1150 		if (err) {
1151 			err = rte_errno;
1152 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1153 				strerror(rte_errno));
1154 			goto error;
1155 		}
1156 		own_domain_id = 1;
1157 	}
1158 	/* Override some values set by hardware configuration. */
1159 	mlx5_args(config, dpdk_dev->devargs);
1160 	err = mlx5_dev_check_sibling_config(priv, config);
1161 	if (err)
1162 		goto error;
1163 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1164 			    IBV_DEVICE_RAW_IP_CSUM);
1165 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1166 		(config->hw_csum ? "" : "not "));
1167 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1168 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1169 	DRV_LOG(DEBUG, "counters are not supported");
1170 #endif
1171 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1172 	if (config->dv_flow_en) {
1173 		DRV_LOG(WARNING, "DV flow is not supported");
1174 		config->dv_flow_en = 0;
1175 	}
1176 #endif
1177 	config->ind_table_max_size =
1178 		sh->device_attr.max_rwq_indirection_table_size;
1179 	/*
1180 	 * Remove this check once DPDK supports larger/variable
1181 	 * indirection tables.
1182 	 */
1183 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1184 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1185 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1186 		config->ind_table_max_size);
1187 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1188 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1189 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1190 		(config->hw_vlan_strip ? "" : "not "));
1191 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1192 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1193 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1194 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1195 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1196 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1197 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1198 #endif
1199 	if (config->hw_padding && !hw_padding) {
1200 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1201 		config->hw_padding = 0;
1202 	} else if (config->hw_padding) {
1203 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1204 	}
1205 	config->tso = (sh->device_attr.max_tso > 0 &&
1206 		      (sh->device_attr.tso_supported_qpts &
1207 		       (1 << IBV_QPT_RAW_PACKET)));
1208 	if (config->tso)
1209 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1210 	/*
1211 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1212 	 * by default.
1213 	 */
1214 	if (config->mps == MLX5_ARG_UNSET)
1215 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1216 							  MLX5_MPW_DISABLED;
1217 	else
1218 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1219 	DRV_LOG(INFO, "%sMPS is %s",
1220 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1221 		config->mps == MLX5_MPW ? "legacy " : "",
1222 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1223 	if (config->devx) {
1224 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1225 		if (err) {
1226 			err = -err;
1227 			goto error;
1228 		}
1229 		/* Check relax ordering support. */
1230 		if (!haswell_broadwell_cpu) {
1231 			sh->cmng.relaxed_ordering_write =
1232 				config->hca_attr.relaxed_ordering_write;
1233 			sh->cmng.relaxed_ordering_read =
1234 				config->hca_attr.relaxed_ordering_read;
1235 		} else {
1236 			sh->cmng.relaxed_ordering_read = 0;
1237 			sh->cmng.relaxed_ordering_write = 0;
1238 		}
1239 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1240 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1241 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1242 		/* Check for LRO support. */
1243 		if (config->dest_tir && config->hca_attr.lro_cap &&
1244 		    config->dv_flow_en) {
1245 			/* TBD check tunnel lro caps. */
1246 			config->lro.supported = config->hca_attr.lro_cap;
1247 			DRV_LOG(DEBUG, "Device supports LRO");
1248 			/*
1249 			 * If LRO timeout is not configured by application,
1250 			 * use the minimal supported value.
1251 			 */
1252 			if (!config->lro.timeout)
1253 				config->lro.timeout =
1254 				config->hca_attr.lro_timer_supported_periods[0];
1255 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1256 				config->lro.timeout);
1257 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1258 				"required for coalescing is %d bytes",
1259 				config->hca_attr.lro_min_mss_size);
1260 		}
1261 #if defined(HAVE_MLX5DV_DR) && \
1262 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1263 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1264 		if (config->hca_attr.qos.sup &&
1265 		    config->hca_attr.qos.flow_meter_old &&
1266 		    config->dv_flow_en) {
1267 			uint8_t reg_c_mask =
1268 				config->hca_attr.qos.flow_meter_reg_c_ids;
1269 			/*
1270 			 * Meter needs two REG_C's for color match and pre-sfx
1271 			 * flow match. Here get the REG_C for color match.
1272 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1273 			 */
1274 			reg_c_mask &= 0xfc;
1275 			if (__builtin_popcount(reg_c_mask) < 1) {
1276 				priv->mtr_en = 0;
1277 				DRV_LOG(WARNING, "No available register for"
1278 					" meter.");
1279 			} else {
1280 				/*
1281 				 * The meter color register is used by the
1282 				 * flow-hit feature as well.
1283 				 * The flow-hit feature must use REG_C_3
1284 				 * Prefer REG_C_3 if it is available.
1285 				 */
1286 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1287 					priv->mtr_color_reg = REG_C_3;
1288 				else
1289 					priv->mtr_color_reg = ffs(reg_c_mask)
1290 							      - 1 + REG_C_0;
1291 				priv->mtr_en = 1;
1292 				priv->mtr_reg_share =
1293 				      config->hca_attr.qos.flow_meter;
1294 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1295 					priv->mtr_color_reg);
1296 			}
1297 		}
1298 		if (config->hca_attr.qos.sup &&
1299 			config->hca_attr.qos.flow_meter_aso_sup) {
1300 			uint32_t log_obj_size =
1301 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1302 			if (log_obj_size >=
1303 			config->hca_attr.qos.log_meter_aso_granularity &&
1304 			log_obj_size <=
1305 			config->hca_attr.qos.log_meter_aso_max_alloc)
1306 				sh->meter_aso_en = 1;
1307 		}
1308 		if (priv->mtr_en) {
1309 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1310 			if (err) {
1311 				err = -err;
1312 				goto error;
1313 			}
1314 		}
1315 #endif
1316 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1317 		if (config->hca_attr.flow_hit_aso &&
1318 		    priv->mtr_color_reg == REG_C_3) {
1319 			sh->flow_hit_aso_en = 1;
1320 			err = mlx5_flow_aso_age_mng_init(sh);
1321 			if (err) {
1322 				err = -err;
1323 				goto error;
1324 			}
1325 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1326 		}
1327 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1328 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1329 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1330 		if (config->hca_attr.ct_offload &&
1331 		    priv->mtr_color_reg == REG_C_3) {
1332 			err = mlx5_flow_aso_ct_mng_init(sh);
1333 			if (err) {
1334 				err = -err;
1335 				goto error;
1336 			}
1337 			DRV_LOG(DEBUG, "CT ASO is supported.");
1338 			sh->ct_aso_en = 1;
1339 		}
1340 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1341 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1342 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1343 		    config->dv_flow_en) {
1344 			priv->sampler_en = 1;
1345 			DRV_LOG(DEBUG, "Sampler enabled!");
1346 		} else {
1347 			priv->sampler_en = 0;
1348 			if (!config->hca_attr.log_max_ft_sampler_num)
1349 				DRV_LOG(WARNING,
1350 					"No available register for sampler.");
1351 			else
1352 				DRV_LOG(DEBUG, "DV flow is not supported!");
1353 		}
1354 #endif
1355 	}
1356 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1357 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1358 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1359 		config->cqe_comp = 0;
1360 	}
1361 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1362 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1363 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1364 				 " format isn't supported.");
1365 		config->cqe_comp = 0;
1366 	}
1367 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1368 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1369 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1370 				 " format isn't supported.");
1371 		config->cqe_comp = 0;
1372 	}
1373 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1374 			config->cqe_comp ? "" : "not ");
1375 	if (config->tx_pp) {
1376 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1377 			config->hca_attr.dev_freq_khz);
1378 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1379 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1380 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1381 			config->hca_attr.cross_channel ? "" : "not ");
1382 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1383 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1384 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1385 			config->hca_attr.non_wire_sq ? "" : "not ");
1386 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1387 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1388 			config->hca_attr.log_max_static_sq_wq);
1389 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1390 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1391 		if (!config->devx) {
1392 			DRV_LOG(ERR, "DevX is required for packet pacing");
1393 			err = ENODEV;
1394 			goto error;
1395 		}
1396 		if (!config->hca_attr.qos.packet_pacing) {
1397 			DRV_LOG(ERR, "Packet pacing is not supported");
1398 			err = ENODEV;
1399 			goto error;
1400 		}
1401 		if (!config->hca_attr.cross_channel) {
1402 			DRV_LOG(ERR, "Cross channel operations are"
1403 				     " required for packet pacing");
1404 			err = ENODEV;
1405 			goto error;
1406 		}
1407 		if (!config->hca_attr.wqe_index_ignore) {
1408 			DRV_LOG(ERR, "WQE index ignore feature is"
1409 				     " required for packet pacing");
1410 			err = ENODEV;
1411 			goto error;
1412 		}
1413 		if (!config->hca_attr.non_wire_sq) {
1414 			DRV_LOG(ERR, "Non-wire SQ feature is"
1415 				     " required for packet pacing");
1416 			err = ENODEV;
1417 			goto error;
1418 		}
1419 		if (!config->hca_attr.log_max_static_sq_wq) {
1420 			DRV_LOG(ERR, "Static WQE SQ feature is"
1421 				     " required for packet pacing");
1422 			err = ENODEV;
1423 			goto error;
1424 		}
1425 		if (!config->hca_attr.qos.wqe_rate_pp) {
1426 			DRV_LOG(ERR, "WQE rate mode is required"
1427 				     " for packet pacing");
1428 			err = ENODEV;
1429 			goto error;
1430 		}
1431 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1432 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1433 			     " can't create queues for packet pacing");
1434 		err = ENODEV;
1435 		goto error;
1436 #endif
1437 	}
1438 	if (config->devx) {
1439 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1440 
1441 		err = config->hca_attr.access_register_user ?
1442 			mlx5_devx_cmd_register_read
1443 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1444 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1445 		if (!err) {
1446 			uint32_t ts_mode;
1447 
1448 			/* MTUTC register is read successfully. */
1449 			ts_mode = MLX5_GET(register_mtutc, reg,
1450 					   time_stamp_mode);
1451 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1452 				config->rt_timestamp = 1;
1453 		} else {
1454 			/* Kernel does not support register reading. */
1455 			if (config->hca_attr.dev_freq_khz ==
1456 						 (NS_PER_S / MS_PER_S))
1457 				config->rt_timestamp = 1;
1458 		}
1459 	}
1460 	/*
1461 	 * If HW has bug working with tunnel packet decapsulation and
1462 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1463 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1464 	 */
1465 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1466 		config->hw_fcs_strip = 0;
1467 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1468 		(config->hw_fcs_strip ? "" : "not "));
1469 	if (config->mprq.enabled && mprq) {
1470 		if (config->mprq.stride_num_n &&
1471 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1472 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1473 			config->mprq.stride_num_n =
1474 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1475 						mprq_min_stride_num_n),
1476 					mprq_max_stride_num_n);
1477 			DRV_LOG(WARNING,
1478 				"the number of strides"
1479 				" for Multi-Packet RQ is out of range,"
1480 				" setting default value (%u)",
1481 				1 << config->mprq.stride_num_n);
1482 		}
1483 		if (config->mprq.stride_size_n &&
1484 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1485 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1486 			config->mprq.stride_size_n =
1487 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1488 						mprq_min_stride_size_n),
1489 					mprq_max_stride_size_n);
1490 			DRV_LOG(WARNING,
1491 				"the size of a stride"
1492 				" for Multi-Packet RQ is out of range,"
1493 				" setting default value (%u)",
1494 				1 << config->mprq.stride_size_n);
1495 		}
1496 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1497 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1498 	} else if (config->mprq.enabled && !mprq) {
1499 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1500 		config->mprq.enabled = 0;
1501 	}
1502 	if (config->max_dump_files_num == 0)
1503 		config->max_dump_files_num = 128;
1504 	eth_dev = rte_eth_dev_allocate(name);
1505 	if (eth_dev == NULL) {
1506 		DRV_LOG(ERR, "can not allocate rte ethdev");
1507 		err = ENOMEM;
1508 		goto error;
1509 	}
1510 	if (priv->representor) {
1511 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1512 		eth_dev->data->representor_id = priv->representor_id;
1513 	}
1514 	priv->mp_id.port_id = eth_dev->data->port_id;
1515 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1516 	/*
1517 	 * Store associated network device interface index. This index
1518 	 * is permanent throughout the lifetime of device. So, we may store
1519 	 * the ifindex here and use the cached value further.
1520 	 */
1521 	MLX5_ASSERT(spawn->ifindex);
1522 	priv->if_index = spawn->ifindex;
1523 	eth_dev->data->dev_private = priv;
1524 	priv->dev_data = eth_dev->data;
1525 	eth_dev->data->mac_addrs = priv->mac;
1526 	eth_dev->device = dpdk_dev;
1527 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1528 	/* Configure the first MAC address by default. */
1529 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1530 		DRV_LOG(ERR,
1531 			"port %u cannot get MAC address, is mlx5_en"
1532 			" loaded? (errno: %s)",
1533 			eth_dev->data->port_id, strerror(rte_errno));
1534 		err = ENODEV;
1535 		goto error;
1536 	}
1537 	DRV_LOG(INFO,
1538 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1539 		eth_dev->data->port_id,
1540 		mac.addr_bytes[0], mac.addr_bytes[1],
1541 		mac.addr_bytes[2], mac.addr_bytes[3],
1542 		mac.addr_bytes[4], mac.addr_bytes[5]);
1543 #ifdef RTE_LIBRTE_MLX5_DEBUG
1544 	{
1545 		char ifname[MLX5_NAMESIZE];
1546 
1547 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1548 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1549 				eth_dev->data->port_id, ifname);
1550 		else
1551 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1552 				eth_dev->data->port_id);
1553 	}
1554 #endif
1555 	/* Get actual MTU if possible. */
1556 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1557 	if (err) {
1558 		err = rte_errno;
1559 		goto error;
1560 	}
1561 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1562 		priv->mtu);
1563 	/* Initialize burst functions to prevent crashes before link-up. */
1564 	eth_dev->rx_pkt_burst = removed_rx_burst;
1565 	eth_dev->tx_pkt_burst = removed_tx_burst;
1566 	eth_dev->dev_ops = &mlx5_dev_ops;
1567 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1568 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1569 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1570 	/* Register MAC address. */
1571 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1572 	if (config->vf && config->vf_nl_en)
1573 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1574 				      mlx5_ifindex(eth_dev),
1575 				      eth_dev->data->mac_addrs,
1576 				      MLX5_MAX_MAC_ADDRESSES);
1577 	priv->flows = 0;
1578 	priv->ctrl_flows = 0;
1579 	rte_spinlock_init(&priv->flow_list_lock);
1580 	TAILQ_INIT(&priv->flow_meters);
1581 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1582 	if (!priv->mtr_profile_tbl)
1583 		goto error;
1584 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1585 	mlx5_glue->dv_set_context_attr(sh->ctx,
1586 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1587 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1588 				.alloc = &mlx5_alloc_verbs_buf,
1589 				.free = &mlx5_free_verbs_buf,
1590 				.data = sh,
1591 			}));
1592 	/* Bring Ethernet device up. */
1593 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1594 		eth_dev->data->port_id);
1595 	mlx5_set_link_up(eth_dev);
1596 	/*
1597 	 * Even though the interrupt handler is not installed yet,
1598 	 * interrupts will still trigger on the async_fd from
1599 	 * Verbs context returned by ibv_open_device().
1600 	 */
1601 	mlx5_link_update(eth_dev, 0);
1602 #ifdef HAVE_MLX5DV_DR_ESWITCH
1603 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1604 	      (switch_info->representor || switch_info->master)))
1605 		config->dv_esw_en = 0;
1606 #else
1607 	config->dv_esw_en = 0;
1608 #endif
1609 	/* Detect minimal data bytes to inline. */
1610 	mlx5_set_min_inline(spawn, config);
1611 	/* Store device configuration on private structure. */
1612 	priv->config = *config;
1613 	/* Create context for virtual machine VLAN workaround. */
1614 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1615 	if (config->dv_flow_en) {
1616 		err = mlx5_alloc_shared_dr(priv);
1617 		if (err)
1618 			goto error;
1619 	}
1620 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1621 		priv->obj_ops = devx_obj_ops;
1622 		priv->obj_ops.drop_action_create =
1623 						ibv_obj_ops.drop_action_create;
1624 		priv->obj_ops.drop_action_destroy =
1625 						ibv_obj_ops.drop_action_destroy;
1626 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1627 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1628 #else
1629 		if (config->dv_esw_en)
1630 			priv->obj_ops.txq_obj_modify =
1631 						ibv_obj_ops.txq_obj_modify;
1632 #endif
1633 		/* Use specific wrappers for Tx object. */
1634 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1635 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1636 		mlx5_queue_counter_id_prepare(eth_dev);
1637 		priv->obj_ops.lb_dummy_queue_create =
1638 					mlx5_rxq_ibv_obj_dummy_lb_create;
1639 		priv->obj_ops.lb_dummy_queue_release =
1640 					mlx5_rxq_ibv_obj_dummy_lb_release;
1641 	} else {
1642 		priv->obj_ops = ibv_obj_ops;
1643 	}
1644 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1645 	if (!priv->drop_queue.hrxq)
1646 		goto error;
1647 	/* Supported Verbs flow priority number detection. */
1648 	err = mlx5_flow_discover_priorities(eth_dev);
1649 	if (err < 0) {
1650 		err = -err;
1651 		goto error;
1652 	}
1653 	priv->config.flow_prio = err;
1654 	if (!priv->config.dv_esw_en &&
1655 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1656 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1657 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1658 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1659 	}
1660 	mlx5_set_metadata_mask(eth_dev);
1661 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1662 	    !priv->sh->dv_regc0_mask) {
1663 		DRV_LOG(ERR, "metadata mode %u is not supported "
1664 			     "(no metadata reg_c[0] is available)",
1665 			     priv->config.dv_xmeta_en);
1666 			err = ENOTSUP;
1667 			goto error;
1668 	}
1669 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1670 			     mlx5_hrxq_create_cb,
1671 			     mlx5_hrxq_match_cb,
1672 			     mlx5_hrxq_remove_cb);
1673 	/* Query availability of metadata reg_c's. */
1674 	err = mlx5_flow_discover_mreg_c(eth_dev);
1675 	if (err < 0) {
1676 		err = -err;
1677 		goto error;
1678 	}
1679 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1680 		DRV_LOG(DEBUG,
1681 			"port %u extensive metadata register is not supported",
1682 			eth_dev->data->port_id);
1683 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1684 			DRV_LOG(ERR, "metadata mode %u is not supported "
1685 				     "(no metadata registers available)",
1686 				     priv->config.dv_xmeta_en);
1687 			err = ENOTSUP;
1688 			goto error;
1689 		}
1690 	}
1691 	if (priv->config.dv_flow_en &&
1692 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1693 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1694 	    priv->sh->dv_regc0_mask) {
1695 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1696 						      MLX5_FLOW_MREG_HTABLE_SZ,
1697 						      0, 0,
1698 						      flow_dv_mreg_create_cb,
1699 						      flow_dv_mreg_match_cb,
1700 						      flow_dv_mreg_remove_cb);
1701 		if (!priv->mreg_cp_tbl) {
1702 			err = ENOMEM;
1703 			goto error;
1704 		}
1705 		priv->mreg_cp_tbl->ctx = eth_dev;
1706 	}
1707 	rte_spinlock_init(&priv->shared_act_sl);
1708 	mlx5_flow_counter_mode_config(eth_dev);
1709 	if (priv->config.dv_flow_en)
1710 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1711 	return eth_dev;
1712 error:
1713 	if (priv) {
1714 		if (priv->mreg_cp_tbl)
1715 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1716 		if (priv->sh)
1717 			mlx5_os_free_shared_dr(priv);
1718 		if (priv->nl_socket_route >= 0)
1719 			close(priv->nl_socket_route);
1720 		if (priv->nl_socket_rdma >= 0)
1721 			close(priv->nl_socket_rdma);
1722 		if (priv->vmwa_context)
1723 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1724 		if (eth_dev && priv->drop_queue.hrxq)
1725 			mlx5_drop_action_destroy(eth_dev);
1726 		if (priv->mtr_profile_tbl)
1727 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1728 		if (own_domain_id)
1729 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1730 		mlx5_cache_list_destroy(&priv->hrxqs);
1731 		mlx5_free(priv);
1732 		if (eth_dev != NULL)
1733 			eth_dev->data->dev_private = NULL;
1734 	}
1735 	if (eth_dev != NULL) {
1736 		/* mac_addrs must not be freed alone because part of
1737 		 * dev_private
1738 		 **/
1739 		eth_dev->data->mac_addrs = NULL;
1740 		rte_eth_dev_release_port(eth_dev);
1741 	}
1742 	if (sh)
1743 		mlx5_free_shared_dev_ctx(sh);
1744 	MLX5_ASSERT(err > 0);
1745 	rte_errno = err;
1746 	return NULL;
1747 }
1748 
1749 /**
1750  * Comparison callback to sort device data.
1751  *
1752  * This is meant to be used with qsort().
1753  *
1754  * @param a[in]
1755  *   Pointer to pointer to first data object.
1756  * @param b[in]
1757  *   Pointer to pointer to second data object.
1758  *
1759  * @return
1760  *   0 if both objects are equal, less than 0 if the first argument is less
1761  *   than the second, greater than 0 otherwise.
1762  */
1763 static int
1764 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1765 {
1766 	const struct mlx5_switch_info *si_a =
1767 		&((const struct mlx5_dev_spawn_data *)a)->info;
1768 	const struct mlx5_switch_info *si_b =
1769 		&((const struct mlx5_dev_spawn_data *)b)->info;
1770 	int ret;
1771 
1772 	/* Master device first. */
1773 	ret = si_b->master - si_a->master;
1774 	if (ret)
1775 		return ret;
1776 	/* Then representor devices. */
1777 	ret = si_b->representor - si_a->representor;
1778 	if (ret)
1779 		return ret;
1780 	/* Unidentified devices come last in no specific order. */
1781 	if (!si_a->representor)
1782 		return 0;
1783 	/* Order representors by name. */
1784 	return si_a->port_name - si_b->port_name;
1785 }
1786 
1787 /**
1788  * Match PCI information for possible slaves of bonding device.
1789  *
1790  * @param[in] ibv_dev
1791  *   Pointer to Infiniband device structure.
1792  * @param[in] pci_dev
1793  *   Pointer to primary PCI address structure to match.
1794  * @param[in] nl_rdma
1795  *   Netlink RDMA group socket handle.
1796  * @param[in] owner
1797  *   Rerepsentor owner PF index.
1798  * @param[out] bond_info
1799  *   Pointer to bonding information.
1800  *
1801  * @return
1802  *   negative value if no bonding device found, otherwise
1803  *   positive index of slave PF in bonding.
1804  */
1805 static int
1806 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1807 			   const struct rte_pci_addr *pci_dev,
1808 			   int nl_rdma, uint16_t owner,
1809 			   struct mlx5_bond_info *bond_info)
1810 {
1811 	char ifname[IF_NAMESIZE + 1];
1812 	unsigned int ifindex;
1813 	unsigned int np, i;
1814 	FILE *bond_file = NULL, *file;
1815 	int pf = -1;
1816 	int ret;
1817 
1818 	/*
1819 	 * Try to get master device name. If something goes
1820 	 * wrong suppose the lack of kernel support and no
1821 	 * bonding devices.
1822 	 */
1823 	memset(bond_info, 0, sizeof(*bond_info));
1824 	if (nl_rdma < 0)
1825 		return -1;
1826 	if (!strstr(ibv_dev->name, "bond"))
1827 		return -1;
1828 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1829 	if (!np)
1830 		return -1;
1831 	/*
1832 	 * The Master device might not be on the predefined
1833 	 * port (not on port index 1, it is not garanted),
1834 	 * we have to scan all Infiniband device port and
1835 	 * find master.
1836 	 */
1837 	for (i = 1; i <= np; ++i) {
1838 		/* Check whether Infiniband port is populated. */
1839 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1840 		if (!ifindex)
1841 			continue;
1842 		if (!if_indextoname(ifindex, ifname))
1843 			continue;
1844 		/* Try to read bonding slave names from sysfs. */
1845 		MKSTR(slaves,
1846 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1847 		bond_file = fopen(slaves, "r");
1848 		if (bond_file)
1849 			break;
1850 	}
1851 	if (!bond_file)
1852 		return -1;
1853 	/* Use safe format to check maximal buffer length. */
1854 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1855 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1856 		char tmp_str[IF_NAMESIZE + 32];
1857 		struct rte_pci_addr pci_addr;
1858 		struct mlx5_switch_info	info;
1859 
1860 		/* Process slave interface names in the loop. */
1861 		snprintf(tmp_str, sizeof(tmp_str),
1862 			 "/sys/class/net/%s", ifname);
1863 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1864 			DRV_LOG(WARNING, "can not get PCI address"
1865 					 " for netdev \"%s\"", ifname);
1866 			continue;
1867 		}
1868 		/* Slave interface PCI address match found. */
1869 		snprintf(tmp_str, sizeof(tmp_str),
1870 			 "/sys/class/net/%s/phys_port_name", ifname);
1871 		file = fopen(tmp_str, "rb");
1872 		if (!file)
1873 			break;
1874 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1875 		if (fscanf(file, "%32s", tmp_str) == 1)
1876 			mlx5_translate_port_name(tmp_str, &info);
1877 		fclose(file);
1878 		/* Only process PF ports. */
1879 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1880 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1881 			continue;
1882 		/* Check max bonding member. */
1883 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1884 			DRV_LOG(WARNING, "bonding index out of range, "
1885 				"please increase MLX5_BOND_MAX_PORTS: %s",
1886 				tmp_str);
1887 			break;
1888 		}
1889 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
1890 		if (pci_dev->domain == pci_addr.domain &&
1891 		    pci_dev->bus == pci_addr.bus &&
1892 		    pci_dev->devid == pci_addr.devid &&
1893 		    ((pci_dev->function == 0 &&
1894 		      pci_dev->function + owner == pci_addr.function) ||
1895 		     (pci_dev->function == owner &&
1896 		      pci_addr.function == owner)))
1897 			pf = info.port_name;
1898 		/* Get ifindex. */
1899 		snprintf(tmp_str, sizeof(tmp_str),
1900 			 "/sys/class/net/%s/ifindex", ifname);
1901 		file = fopen(tmp_str, "rb");
1902 		if (!file)
1903 			break;
1904 		ret = fscanf(file, "%u", &ifindex);
1905 		fclose(file);
1906 		if (ret != 1)
1907 			break;
1908 		/* Save bonding info. */
1909 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1910 			sizeof(bond_info->ports[0].ifname));
1911 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1912 		bond_info->ports[info.port_name].ifindex = ifindex;
1913 		bond_info->n_port++;
1914 	}
1915 	if (pf >= 0) {
1916 		/* Get bond interface info */
1917 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1918 					   bond_info->ifname);
1919 		if (ret)
1920 			DRV_LOG(ERR, "unable to get bond info: %s",
1921 				strerror(rte_errno));
1922 		else
1923 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1924 				ifindex, bond_info->ifindex, bond_info->ifname);
1925 	}
1926 	return pf;
1927 }
1928 
1929 /**
1930  * Register a PCI device within bonding.
1931  *
1932  * This function spawns Ethernet devices out of a given PCI device and
1933  * bonding owner PF index.
1934  *
1935  * @param[in] pci_dev
1936  *   PCI device information.
1937  * @param[in] req_eth_da
1938  *   Requested ethdev device argument.
1939  * @param[in] owner_id
1940  *   Requested owner PF port ID within bonding device, default to 0.
1941  *
1942  * @return
1943  *   0 on success, a negative errno value otherwise and rte_errno is set.
1944  */
1945 static int
1946 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1947 		     struct rte_eth_devargs *req_eth_da,
1948 		     uint16_t owner_id)
1949 {
1950 	struct ibv_device **ibv_list;
1951 	/*
1952 	 * Number of found IB Devices matching with requested PCI BDF.
1953 	 * nd != 1 means there are multiple IB devices over the same
1954 	 * PCI device and we have representors and master.
1955 	 */
1956 	unsigned int nd = 0;
1957 	/*
1958 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1959 	 * we have the single multiport IB device, and there may be
1960 	 * representors attached to some of found ports.
1961 	 */
1962 	unsigned int np = 0;
1963 	/*
1964 	 * Number of DPDK ethernet devices to Spawn - either over
1965 	 * multiple IB devices or multiple ports of single IB device.
1966 	 * Actually this is the number of iterations to spawn.
1967 	 */
1968 	unsigned int ns = 0;
1969 	/*
1970 	 * Bonding device
1971 	 *   < 0 - no bonding device (single one)
1972 	 *  >= 0 - bonding device (value is slave PF index)
1973 	 */
1974 	int bd = -1;
1975 	struct mlx5_dev_spawn_data *list = NULL;
1976 	struct mlx5_dev_config dev_config;
1977 	unsigned int dev_config_vf;
1978 	struct rte_eth_devargs eth_da = *req_eth_da;
1979 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1980 	struct mlx5_bond_info bond_info;
1981 	int ret = -1;
1982 
1983 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1984 		mlx5_pmd_socket_init();
1985 	ret = mlx5_init_once();
1986 	if (ret) {
1987 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1988 			strerror(rte_errno));
1989 		return -rte_errno;
1990 	}
1991 	errno = 0;
1992 	ibv_list = mlx5_glue->get_device_list(&ret);
1993 	if (!ibv_list) {
1994 		rte_errno = errno ? errno : ENOSYS;
1995 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1996 		return -rte_errno;
1997 	}
1998 	/*
1999 	 * First scan the list of all Infiniband devices to find
2000 	 * matching ones, gathering into the list.
2001 	 */
2002 	struct ibv_device *ibv_match[ret + 1];
2003 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2004 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2005 	unsigned int i;
2006 
2007 	while (ret-- > 0) {
2008 		struct rte_pci_addr pci_addr;
2009 
2010 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2011 		bd = mlx5_device_bond_pci_match
2012 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2013 				 &bond_info);
2014 		if (bd >= 0) {
2015 			/*
2016 			 * Bonding device detected. Only one match is allowed,
2017 			 * the bonding is supported over multi-port IB device,
2018 			 * there should be no matches on representor PCI
2019 			 * functions or non VF LAG bonding devices with
2020 			 * specified address.
2021 			 */
2022 			if (nd) {
2023 				DRV_LOG(ERR,
2024 					"multiple PCI match on bonding device"
2025 					"\"%s\" found", ibv_list[ret]->name);
2026 				rte_errno = ENOENT;
2027 				ret = -rte_errno;
2028 				goto exit;
2029 			}
2030 			/* Amend owner pci address if owner PF ID specified. */
2031 			if (eth_da.nb_representor_ports)
2032 				owner_pci.function += owner_id;
2033 			DRV_LOG(INFO, "PCI information matches for"
2034 				      " slave %d bonding device \"%s\"",
2035 				      bd, ibv_list[ret]->name);
2036 			ibv_match[nd++] = ibv_list[ret];
2037 			break;
2038 		} else {
2039 			/* Bonding device not found. */
2040 			if (mlx5_dev_to_pci_addr
2041 				(ibv_list[ret]->ibdev_path, &pci_addr))
2042 				continue;
2043 			if (owner_pci.domain != pci_addr.domain ||
2044 			    owner_pci.bus != pci_addr.bus ||
2045 			    owner_pci.devid != pci_addr.devid ||
2046 			    owner_pci.function != pci_addr.function)
2047 				continue;
2048 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2049 				ibv_list[ret]->name);
2050 			ibv_match[nd++] = ibv_list[ret];
2051 		}
2052 	}
2053 	ibv_match[nd] = NULL;
2054 	if (!nd) {
2055 		/* No device matches, just complain and bail out. */
2056 		DRV_LOG(WARNING,
2057 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2058 			" are kernel drivers loaded?",
2059 			owner_pci.domain, owner_pci.bus,
2060 			owner_pci.devid, owner_pci.function);
2061 		rte_errno = ENOENT;
2062 		ret = -rte_errno;
2063 		goto exit;
2064 	}
2065 	if (nd == 1) {
2066 		/*
2067 		 * Found single matching device may have multiple ports.
2068 		 * Each port may be representor, we have to check the port
2069 		 * number and check the representors existence.
2070 		 */
2071 		if (nl_rdma >= 0)
2072 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2073 		if (!np)
2074 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2075 					 " ports number", ibv_match[0]->name);
2076 		if (bd >= 0 && !np) {
2077 			DRV_LOG(ERR, "can not get ports"
2078 				     " for bonding device");
2079 			rte_errno = ENOENT;
2080 			ret = -rte_errno;
2081 			goto exit;
2082 		}
2083 	}
2084 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2085 	if (bd >= 0) {
2086 		/*
2087 		 * This may happen if there is VF LAG kernel support and
2088 		 * application is compiled with older rdma_core library.
2089 		 */
2090 		DRV_LOG(ERR,
2091 			"No kernel/verbs support for VF LAG bonding found.");
2092 		rte_errno = ENOTSUP;
2093 		ret = -rte_errno;
2094 		goto exit;
2095 	}
2096 #endif
2097 	/*
2098 	 * Now we can determine the maximal
2099 	 * amount of devices to be spawned.
2100 	 */
2101 	list = mlx5_malloc(MLX5_MEM_ZERO,
2102 			   sizeof(struct mlx5_dev_spawn_data) *
2103 			   (np ? np : nd),
2104 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2105 	if (!list) {
2106 		DRV_LOG(ERR, "spawn data array allocation failure");
2107 		rte_errno = ENOMEM;
2108 		ret = -rte_errno;
2109 		goto exit;
2110 	}
2111 	if (bd >= 0 || np > 1) {
2112 		/*
2113 		 * Single IB device with multiple ports found,
2114 		 * it may be E-Switch master device and representors.
2115 		 * We have to perform identification through the ports.
2116 		 */
2117 		MLX5_ASSERT(nl_rdma >= 0);
2118 		MLX5_ASSERT(ns == 0);
2119 		MLX5_ASSERT(nd == 1);
2120 		MLX5_ASSERT(np);
2121 		for (i = 1; i <= np; ++i) {
2122 			list[ns].bond_info = &bond_info;
2123 			list[ns].max_port = np;
2124 			list[ns].phys_port = i;
2125 			list[ns].phys_dev = ibv_match[0];
2126 			list[ns].eth_dev = NULL;
2127 			list[ns].pci_dev = pci_dev;
2128 			list[ns].pf_bond = bd;
2129 			list[ns].ifindex = mlx5_nl_ifindex
2130 				(nl_rdma,
2131 				mlx5_os_get_dev_device_name
2132 						(list[ns].phys_dev), i);
2133 			if (!list[ns].ifindex) {
2134 				/*
2135 				 * No network interface index found for the
2136 				 * specified port, it means there is no
2137 				 * representor on this port. It's OK,
2138 				 * there can be disabled ports, for example
2139 				 * if sriov_numvfs < sriov_totalvfs.
2140 				 */
2141 				continue;
2142 			}
2143 			ret = -1;
2144 			if (nl_route >= 0)
2145 				ret = mlx5_nl_switch_info
2146 					       (nl_route,
2147 						list[ns].ifindex,
2148 						&list[ns].info);
2149 			if (ret || (!list[ns].info.representor &&
2150 				    !list[ns].info.master)) {
2151 				/*
2152 				 * We failed to recognize representors with
2153 				 * Netlink, let's try to perform the task
2154 				 * with sysfs.
2155 				 */
2156 				ret =  mlx5_sysfs_switch_info
2157 						(list[ns].ifindex,
2158 						 &list[ns].info);
2159 			}
2160 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2161 			if (!ret && bd >= 0) {
2162 				switch (list[ns].info.name_type) {
2163 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2164 					if (list[ns].info.port_name == bd)
2165 						ns++;
2166 					break;
2167 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2168 					/* Fallthrough */
2169 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2170 					/* Fallthrough */
2171 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2172 					if (list[ns].info.pf_num == bd)
2173 						ns++;
2174 					break;
2175 				default:
2176 					break;
2177 				}
2178 				continue;
2179 			}
2180 #endif
2181 			if (!ret && (list[ns].info.representor ^
2182 				     list[ns].info.master))
2183 				ns++;
2184 		}
2185 		if (!ns) {
2186 			DRV_LOG(ERR,
2187 				"unable to recognize master/representors"
2188 				" on the IB device with multiple ports");
2189 			rte_errno = ENOENT;
2190 			ret = -rte_errno;
2191 			goto exit;
2192 		}
2193 	} else {
2194 		/*
2195 		 * The existence of several matching entries (nd > 1) means
2196 		 * port representors have been instantiated. No existing Verbs
2197 		 * call nor sysfs entries can tell them apart, this can only
2198 		 * be done through Netlink calls assuming kernel drivers are
2199 		 * recent enough to support them.
2200 		 *
2201 		 * In the event of identification failure through Netlink,
2202 		 * try again through sysfs, then:
2203 		 *
2204 		 * 1. A single IB device matches (nd == 1) with single
2205 		 *    port (np=0/1) and is not a representor, assume
2206 		 *    no switch support.
2207 		 *
2208 		 * 2. Otherwise no safe assumptions can be made;
2209 		 *    complain louder and bail out.
2210 		 */
2211 		for (i = 0; i != nd; ++i) {
2212 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2213 			list[ns].bond_info = NULL;
2214 			list[ns].max_port = 1;
2215 			list[ns].phys_port = 1;
2216 			list[ns].phys_dev = ibv_match[i];
2217 			list[ns].eth_dev = NULL;
2218 			list[ns].pci_dev = pci_dev;
2219 			list[ns].pf_bond = -1;
2220 			list[ns].ifindex = 0;
2221 			if (nl_rdma >= 0)
2222 				list[ns].ifindex = mlx5_nl_ifindex
2223 				(nl_rdma,
2224 				mlx5_os_get_dev_device_name
2225 						(list[ns].phys_dev), 1);
2226 			if (!list[ns].ifindex) {
2227 				char ifname[IF_NAMESIZE];
2228 
2229 				/*
2230 				 * Netlink failed, it may happen with old
2231 				 * ib_core kernel driver (before 4.16).
2232 				 * We can assume there is old driver because
2233 				 * here we are processing single ports IB
2234 				 * devices. Let's try sysfs to retrieve
2235 				 * the ifindex. The method works for
2236 				 * master device only.
2237 				 */
2238 				if (nd > 1) {
2239 					/*
2240 					 * Multiple devices found, assume
2241 					 * representors, can not distinguish
2242 					 * master/representor and retrieve
2243 					 * ifindex via sysfs.
2244 					 */
2245 					continue;
2246 				}
2247 				ret = mlx5_get_ifname_sysfs
2248 					(ibv_match[i]->ibdev_path, ifname);
2249 				if (!ret)
2250 					list[ns].ifindex =
2251 						if_nametoindex(ifname);
2252 				if (!list[ns].ifindex) {
2253 					/*
2254 					 * No network interface index found
2255 					 * for the specified device, it means
2256 					 * there it is neither representor
2257 					 * nor master.
2258 					 */
2259 					continue;
2260 				}
2261 			}
2262 			ret = -1;
2263 			if (nl_route >= 0)
2264 				ret = mlx5_nl_switch_info
2265 					       (nl_route,
2266 						list[ns].ifindex,
2267 						&list[ns].info);
2268 			if (ret || (!list[ns].info.representor &&
2269 				    !list[ns].info.master)) {
2270 				/*
2271 				 * We failed to recognize representors with
2272 				 * Netlink, let's try to perform the task
2273 				 * with sysfs.
2274 				 */
2275 				ret =  mlx5_sysfs_switch_info
2276 						(list[ns].ifindex,
2277 						 &list[ns].info);
2278 			}
2279 			if (!ret && (list[ns].info.representor ^
2280 				     list[ns].info.master)) {
2281 				ns++;
2282 			} else if ((nd == 1) &&
2283 				   !list[ns].info.representor &&
2284 				   !list[ns].info.master) {
2285 				/*
2286 				 * Single IB device with
2287 				 * one physical port and
2288 				 * attached network device.
2289 				 * May be SRIOV is not enabled
2290 				 * or there is no representors.
2291 				 */
2292 				DRV_LOG(INFO, "no E-Switch support detected");
2293 				ns++;
2294 				break;
2295 			}
2296 		}
2297 		if (!ns) {
2298 			DRV_LOG(ERR,
2299 				"unable to recognize master/representors"
2300 				" on the multiple IB devices");
2301 			rte_errno = ENOENT;
2302 			ret = -rte_errno;
2303 			goto exit;
2304 		}
2305 		/*
2306 		 * New kernels may add the switch_id attribute for the case
2307 		 * there is no E-Switch and we wrongly recognized the
2308 		 * only device as master. Override this if there is the
2309 		 * single device with single port and new device name
2310 		 * format present.
2311 		 */
2312 		if (nd == 1 &&
2313 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2314 			list[0].info.master = 0;
2315 			list[0].info.representor = 0;
2316 		}
2317 	}
2318 	MLX5_ASSERT(ns);
2319 	/*
2320 	 * Sort list to probe devices in natural order for users convenience
2321 	 * (i.e. master first, then representors from lowest to highest ID).
2322 	 */
2323 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2324 	/* Device specific configuration. */
2325 	switch (pci_dev->id.device_id) {
2326 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2327 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2328 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2329 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2330 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2331 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2332 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2333 		dev_config_vf = 1;
2334 		break;
2335 	default:
2336 		dev_config_vf = 0;
2337 		break;
2338 	}
2339 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2340 		/* Set devargs default values. */
2341 		if (eth_da.nb_mh_controllers == 0) {
2342 			eth_da.nb_mh_controllers = 1;
2343 			eth_da.mh_controllers[0] = 0;
2344 		}
2345 		if (eth_da.nb_ports == 0 && ns > 0) {
2346 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2347 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2348 					pci_dev->device.devargs->args);
2349 			eth_da.nb_ports = 1;
2350 			eth_da.ports[0] = list[0].info.pf_num;
2351 		}
2352 		if (eth_da.nb_representor_ports == 0) {
2353 			eth_da.nb_representor_ports = 1;
2354 			eth_da.representor_ports[0] = 0;
2355 		}
2356 	}
2357 	for (i = 0; i != ns; ++i) {
2358 		uint32_t restore;
2359 
2360 		/* Default configuration. */
2361 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2362 		dev_config.vf = dev_config_vf;
2363 		dev_config.mps = MLX5_ARG_UNSET;
2364 		dev_config.dbnc = MLX5_ARG_UNSET;
2365 		dev_config.rx_vec_en = 1;
2366 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2367 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2368 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2369 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2370 		dev_config.vf_nl_en = 1;
2371 		dev_config.mr_ext_memseg_en = 1;
2372 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2373 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2374 		dev_config.dv_esw_en = 1;
2375 		dev_config.dv_flow_en = 1;
2376 		dev_config.decap_en = 1;
2377 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2378 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2379 						 &list[i],
2380 						 &dev_config,
2381 						 &eth_da);
2382 		if (!list[i].eth_dev) {
2383 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2384 				break;
2385 			/* Device is disabled or already spawned. Ignore it. */
2386 			continue;
2387 		}
2388 		restore = list[i].eth_dev->data->dev_flags;
2389 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2390 		/* Restore non-PCI flags cleared by the above call. */
2391 		list[i].eth_dev->data->dev_flags |= restore;
2392 		rte_eth_dev_probing_finish(list[i].eth_dev);
2393 	}
2394 	if (i != ns) {
2395 		DRV_LOG(ERR,
2396 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2397 			" encountering an error: %s",
2398 			owner_pci.domain, owner_pci.bus,
2399 			owner_pci.devid, owner_pci.function,
2400 			strerror(rte_errno));
2401 		ret = -rte_errno;
2402 		/* Roll back. */
2403 		while (i--) {
2404 			if (!list[i].eth_dev)
2405 				continue;
2406 			mlx5_dev_close(list[i].eth_dev);
2407 			/* mac_addrs must not be freed because in dev_private */
2408 			list[i].eth_dev->data->mac_addrs = NULL;
2409 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2410 		}
2411 		/* Restore original error. */
2412 		rte_errno = -ret;
2413 	} else {
2414 		ret = 0;
2415 	}
2416 exit:
2417 	/*
2418 	 * Do the routine cleanup:
2419 	 * - close opened Netlink sockets
2420 	 * - free allocated spawn data array
2421 	 * - free the Infiniband device list
2422 	 */
2423 	if (nl_rdma >= 0)
2424 		close(nl_rdma);
2425 	if (nl_route >= 0)
2426 		close(nl_route);
2427 	if (list)
2428 		mlx5_free(list);
2429 	MLX5_ASSERT(ibv_list);
2430 	mlx5_glue->free_device_list(ibv_list);
2431 	return ret;
2432 }
2433 
2434 /**
2435  * DPDK callback to register a PCI device.
2436  *
2437  * This function spawns Ethernet devices out of a given PCI device.
2438  *
2439  * @param[in] pci_drv
2440  *   PCI driver structure (mlx5_driver).
2441  * @param[in] pci_dev
2442  *   PCI device information.
2443  *
2444  * @return
2445  *   0 on success, a negative errno value otherwise and rte_errno is set.
2446  */
2447 int
2448 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2449 		  struct rte_pci_device *pci_dev)
2450 {
2451 	struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2452 	int ret = 0;
2453 	uint16_t p;
2454 
2455 	if (pci_dev->device.devargs) {
2456 		/* Parse representor information from device argument. */
2457 		if (pci_dev->device.devargs->cls_str)
2458 			ret = rte_eth_devargs_parse
2459 				(pci_dev->device.devargs->cls_str, &eth_da);
2460 		if (ret) {
2461 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2462 				pci_dev->device.devargs->cls_str);
2463 			return -rte_errno;
2464 		}
2465 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2466 			/* Support legacy device argument */
2467 			ret = rte_eth_devargs_parse
2468 				(pci_dev->device.devargs->args, &eth_da);
2469 			if (ret) {
2470 				DRV_LOG(ERR, "failed to parse device arguments: %s",
2471 					pci_dev->device.devargs->args);
2472 				return -rte_errno;
2473 			}
2474 		}
2475 	}
2476 
2477 	if (eth_da.nb_ports > 0) {
2478 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2479 		for (p = 0; p < eth_da.nb_ports; p++)
2480 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2481 						   eth_da.ports[p]);
2482 	} else {
2483 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2484 	}
2485 	return ret;
2486 }
2487 
2488 static int
2489 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2490 {
2491 	char *env;
2492 	int value;
2493 
2494 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2495 	/* Get environment variable to store. */
2496 	env = getenv(MLX5_SHUT_UP_BF);
2497 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2498 	if (config->dbnc == MLX5_ARG_UNSET)
2499 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2500 	else
2501 		setenv(MLX5_SHUT_UP_BF,
2502 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2503 	return value;
2504 }
2505 
2506 static void
2507 mlx5_restore_doorbell_mapping_env(int value)
2508 {
2509 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2510 	/* Restore the original environment variable state. */
2511 	if (value == MLX5_ARG_UNSET)
2512 		unsetenv(MLX5_SHUT_UP_BF);
2513 	else
2514 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2515 }
2516 
2517 /**
2518  * Extract pdn of PD object using DV API.
2519  *
2520  * @param[in] pd
2521  *   Pointer to the verbs PD object.
2522  * @param[out] pdn
2523  *   Pointer to the PD object number variable.
2524  *
2525  * @return
2526  *   0 on success, error value otherwise.
2527  */
2528 int
2529 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2530 {
2531 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2532 	struct mlx5dv_obj obj;
2533 	struct mlx5dv_pd pd_info;
2534 	int ret = 0;
2535 
2536 	obj.pd.in = pd;
2537 	obj.pd.out = &pd_info;
2538 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2539 	if (ret) {
2540 		DRV_LOG(DEBUG, "Fail to get PD object info");
2541 		return ret;
2542 	}
2543 	*pdn = pd_info.pdn;
2544 	return 0;
2545 #else
2546 	(void)pd;
2547 	(void)pdn;
2548 	return -ENOTSUP;
2549 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2550 }
2551 
2552 /**
2553  * Function API to open IB device.
2554  *
2555  * This function calls the Linux glue APIs to open a device.
2556  *
2557  * @param[in] spawn
2558  *   Pointer to the IB device attributes (name, port, etc).
2559  * @param[out] config
2560  *   Pointer to device configuration structure.
2561  * @param[out] sh
2562  *   Pointer to shared context structure.
2563  *
2564  * @return
2565  *   0 on success, a positive error value otherwise.
2566  */
2567 int
2568 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2569 		     const struct mlx5_dev_config *config,
2570 		     struct mlx5_dev_ctx_shared *sh)
2571 {
2572 	int dbmap_env;
2573 	int err = 0;
2574 
2575 	sh->numa_node = spawn->pci_dev->device.numa_node;
2576 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2577 	/*
2578 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2579 	 * before the device creation. The rdma_core library
2580 	 * checks the variable at device creation and
2581 	 * stores the result internally.
2582 	 */
2583 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2584 	/* Try to open IB device with DV first, then usual Verbs. */
2585 	errno = 0;
2586 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2587 	if (sh->ctx) {
2588 		sh->devx = 1;
2589 		DRV_LOG(DEBUG, "DevX is supported");
2590 		/* The device is created, no need for environment. */
2591 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2592 	} else {
2593 		/* The environment variable is still configured. */
2594 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2595 		err = errno ? errno : ENODEV;
2596 		/*
2597 		 * The environment variable is not needed anymore,
2598 		 * all device creation attempts are completed.
2599 		 */
2600 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2601 		if (!sh->ctx)
2602 			return err;
2603 		DRV_LOG(DEBUG, "DevX is NOT supported");
2604 		err = 0;
2605 	}
2606 	if (!err && sh->ctx) {
2607 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2608 		mlx5_glue->dv_set_context_attr(sh->ctx,
2609 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2610 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2611 				.alloc = &mlx5_alloc_verbs_buf,
2612 				.free = &mlx5_free_verbs_buf,
2613 				.data = sh,
2614 			}));
2615 	}
2616 	return err;
2617 }
2618 
2619 /**
2620  * Install shared asynchronous device events handler.
2621  * This function is implemented to support event sharing
2622  * between multiple ports of single IB device.
2623  *
2624  * @param sh
2625  *   Pointer to mlx5_dev_ctx_shared object.
2626  */
2627 void
2628 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2629 {
2630 	int ret;
2631 	int flags;
2632 
2633 	sh->intr_handle.fd = -1;
2634 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2635 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2636 		    F_SETFL, flags | O_NONBLOCK);
2637 	if (ret) {
2638 		DRV_LOG(INFO, "failed to change file descriptor async event"
2639 			" queue");
2640 	} else {
2641 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2642 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2643 		if (rte_intr_callback_register(&sh->intr_handle,
2644 					mlx5_dev_interrupt_handler, sh)) {
2645 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2646 			sh->intr_handle.fd = -1;
2647 		}
2648 	}
2649 	if (sh->devx) {
2650 #ifdef HAVE_IBV_DEVX_ASYNC
2651 		sh->intr_handle_devx.fd = -1;
2652 		sh->devx_comp =
2653 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2654 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2655 		if (!devx_comp) {
2656 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2657 			return;
2658 		}
2659 		flags = fcntl(devx_comp->fd, F_GETFL);
2660 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2661 		if (ret) {
2662 			DRV_LOG(INFO, "failed to change file descriptor"
2663 				" devx comp");
2664 			return;
2665 		}
2666 		sh->intr_handle_devx.fd = devx_comp->fd;
2667 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2668 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2669 					mlx5_dev_interrupt_handler_devx, sh)) {
2670 			DRV_LOG(INFO, "Fail to install the devx shared"
2671 				" interrupt.");
2672 			sh->intr_handle_devx.fd = -1;
2673 		}
2674 #endif /* HAVE_IBV_DEVX_ASYNC */
2675 	}
2676 }
2677 
2678 /**
2679  * Uninstall shared asynchronous device events handler.
2680  * This function is implemented to support event sharing
2681  * between multiple ports of single IB device.
2682  *
2683  * @param dev
2684  *   Pointer to mlx5_dev_ctx_shared object.
2685  */
2686 void
2687 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2688 {
2689 	if (sh->intr_handle.fd >= 0)
2690 		mlx5_intr_callback_unregister(&sh->intr_handle,
2691 					      mlx5_dev_interrupt_handler, sh);
2692 #ifdef HAVE_IBV_DEVX_ASYNC
2693 	if (sh->intr_handle_devx.fd >= 0)
2694 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2695 				  mlx5_dev_interrupt_handler_devx, sh);
2696 	if (sh->devx_comp)
2697 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2698 #endif
2699 }
2700 
2701 /**
2702  * Read statistics by a named counter.
2703  *
2704  * @param[in] priv
2705  *   Pointer to the private device data structure.
2706  * @param[in] ctr_name
2707  *   Pointer to the name of the statistic counter to read
2708  * @param[out] stat
2709  *   Pointer to read statistic value.
2710  * @return
2711  *   0 on success and stat is valud, 1 if failed to read the value
2712  *   rte_errno is set.
2713  *
2714  */
2715 int
2716 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2717 		      uint64_t *stat)
2718 {
2719 	int fd;
2720 
2721 	if (priv->sh) {
2722 		if (priv->q_counters != NULL &&
2723 		    strcmp(ctr_name, "out_of_buffer") == 0)
2724 			return mlx5_devx_cmd_queue_counter_query
2725 					(priv->q_counters, 0, (uint32_t *)stat);
2726 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2727 		      priv->sh->ibdev_path,
2728 		      priv->dev_port,
2729 		      ctr_name);
2730 		fd = open(path, O_RDONLY);
2731 		/*
2732 		 * in switchdev the file location is not per port
2733 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2734 		 */
2735 		if (fd == -1) {
2736 			MKSTR(path1, "%s/hw_counters/%s",
2737 			      priv->sh->ibdev_path,
2738 			      ctr_name);
2739 			fd = open(path1, O_RDONLY);
2740 		}
2741 		if (fd != -1) {
2742 			char buf[21] = {'\0'};
2743 			ssize_t n = read(fd, buf, sizeof(buf));
2744 
2745 			close(fd);
2746 			if (n != -1) {
2747 				*stat = strtoull(buf, NULL, 10);
2748 				return 0;
2749 			}
2750 		}
2751 	}
2752 	*stat = 0;
2753 	return 1;
2754 }
2755 
2756 /**
2757  * Set the reg_mr and dereg_mr call backs
2758  *
2759  * @param reg_mr_cb[out]
2760  *   Pointer to reg_mr func
2761  * @param dereg_mr_cb[out]
2762  *   Pointer to dereg_mr func
2763  *
2764  */
2765 void
2766 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2767 		      mlx5_dereg_mr_t *dereg_mr_cb)
2768 {
2769 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2770 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2771 }
2772 
2773 /**
2774  * Remove a MAC address from device
2775  *
2776  * @param dev
2777  *   Pointer to Ethernet device structure.
2778  * @param index
2779  *   MAC address index.
2780  */
2781 void
2782 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2783 {
2784 	struct mlx5_priv *priv = dev->data->dev_private;
2785 	const int vf = priv->config.vf;
2786 
2787 	if (vf)
2788 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2789 					mlx5_ifindex(dev), priv->mac_own,
2790 					&dev->data->mac_addrs[index], index);
2791 }
2792 
2793 /**
2794  * Adds a MAC address to the device
2795  *
2796  * @param dev
2797  *   Pointer to Ethernet device structure.
2798  * @param mac_addr
2799  *   MAC address to register.
2800  * @param index
2801  *   MAC address index.
2802  *
2803  * @return
2804  *   0 on success, a negative errno value otherwise
2805  */
2806 int
2807 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2808 		     uint32_t index)
2809 {
2810 	struct mlx5_priv *priv = dev->data->dev_private;
2811 	const int vf = priv->config.vf;
2812 	int ret = 0;
2813 
2814 	if (vf)
2815 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2816 					   mlx5_ifindex(dev), priv->mac_own,
2817 					   mac, index);
2818 	return ret;
2819 }
2820 
2821 /**
2822  * Modify a VF MAC address
2823  *
2824  * @param priv
2825  *   Pointer to device private data.
2826  * @param mac_addr
2827  *   MAC address to modify into.
2828  * @param iface_idx
2829  *   Net device interface index
2830  * @param vf_index
2831  *   VF index
2832  *
2833  * @return
2834  *   0 on success, a negative errno value otherwise
2835  */
2836 int
2837 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2838 			   unsigned int iface_idx,
2839 			   struct rte_ether_addr *mac_addr,
2840 			   int vf_index)
2841 {
2842 	return mlx5_nl_vf_mac_addr_modify
2843 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2844 }
2845 
2846 /**
2847  * Set device promiscuous mode
2848  *
2849  * @param dev
2850  *   Pointer to Ethernet device structure.
2851  * @param enable
2852  *   0 - promiscuous is disabled, otherwise - enabled
2853  *
2854  * @return
2855  *   0 on success, a negative error value otherwise
2856  */
2857 int
2858 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2859 {
2860 	struct mlx5_priv *priv = dev->data->dev_private;
2861 
2862 	return mlx5_nl_promisc(priv->nl_socket_route,
2863 			       mlx5_ifindex(dev), !!enable);
2864 }
2865 
2866 /**
2867  * Set device promiscuous mode
2868  *
2869  * @param dev
2870  *   Pointer to Ethernet device structure.
2871  * @param enable
2872  *   0 - all multicase is disabled, otherwise - enabled
2873  *
2874  * @return
2875  *   0 on success, a negative error value otherwise
2876  */
2877 int
2878 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2879 {
2880 	struct mlx5_priv *priv = dev->data->dev_private;
2881 
2882 	return mlx5_nl_allmulti(priv->nl_socket_route,
2883 				mlx5_ifindex(dev), !!enable);
2884 }
2885 
2886 /**
2887  * Flush device MAC addresses
2888  *
2889  * @param dev
2890  *   Pointer to Ethernet device structure.
2891  *
2892  */
2893 void
2894 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2895 {
2896 	struct mlx5_priv *priv = dev->data->dev_private;
2897 
2898 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2899 			       dev->data->mac_addrs,
2900 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2901 }
2902