xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision b53d106d34b5c638f5a2cbdfee0da5bd42d4383f)
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_bus_auxiliary.h>
24 #include <rte_common.h>
25 #include <rte_kvargs.h>
26 #include <rte_rwlock.h>
27 #include <rte_spinlock.h>
28 #include <rte_string_fns.h>
29 #include <rte_alarm.h>
30 #include <rte_eal_paging.h>
31 
32 #include <mlx5_glue.h>
33 #include <mlx5_devx_cmds.h>
34 #include <mlx5_common.h>
35 #include <mlx5_common_mp.h>
36 #include <mlx5_common_mr.h>
37 #include <mlx5_malloc.h>
38 
39 #include "mlx5_defs.h"
40 #include "mlx5.h"
41 #include "mlx5_common_os.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_rx.h"
45 #include "mlx5_tx.h"
46 #include "mlx5_autoconf.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 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /* rte flow indexed pool configuration. */
71 static struct mlx5_indexed_pool_config icfg[] = {
72 	{
73 		.size = sizeof(struct rte_flow),
74 		.trunk_size = 64,
75 		.need_lock = 1,
76 		.release_mem_en = 0,
77 		.malloc = mlx5_malloc,
78 		.free = mlx5_free,
79 		.per_core_cache = 0,
80 		.type = "ctl_flow_ipool",
81 	},
82 	{
83 		.size = sizeof(struct rte_flow),
84 		.trunk_size = 64,
85 		.grow_trunk = 3,
86 		.grow_shift = 2,
87 		.need_lock = 1,
88 		.release_mem_en = 0,
89 		.malloc = mlx5_malloc,
90 		.free = mlx5_free,
91 		.per_core_cache = 1 << 14,
92 		.type = "rte_flow_ipool",
93 	},
94 	{
95 		.size = sizeof(struct rte_flow),
96 		.trunk_size = 64,
97 		.grow_trunk = 3,
98 		.grow_shift = 2,
99 		.need_lock = 1,
100 		.release_mem_en = 0,
101 		.malloc = mlx5_malloc,
102 		.free = mlx5_free,
103 		.per_core_cache = 0,
104 		.type = "mcp_flow_ipool",
105 	},
106 };
107 
108 /**
109  * Set the completion channel file descriptor interrupt as non-blocking.
110  *
111  * @param[in] rxq_obj
112  *   Pointer to RQ channel object, which includes the channel fd
113  *
114  * @param[out] fd
115  *   The file descriptor (representing the intetrrupt) used in this channel.
116  *
117  * @return
118  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
119  */
120 int
121 mlx5_os_set_nonblock_channel_fd(int fd)
122 {
123 	int flags;
124 
125 	flags = fcntl(fd, F_GETFL);
126 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
127 }
128 
129 /**
130  * Get mlx5 device attributes. The glue function query_device_ex() is called
131  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
132  * device attributes from the glue out parameter.
133  *
134  * @param cdev
135  *   Pointer to mlx5 device.
136  *
137  * @param device_attr
138  *   Pointer to mlx5 device attributes.
139  *
140  * @return
141  *   0 on success, non zero error number otherwise
142  */
143 int
144 mlx5_os_get_dev_attr(struct mlx5_common_device *cdev,
145 		     struct mlx5_dev_attr *device_attr)
146 {
147 	int err;
148 	struct ibv_context *ctx = cdev->ctx;
149 	struct ibv_device_attr_ex attr_ex;
150 
151 	memset(device_attr, 0, sizeof(*device_attr));
152 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
153 	if (err)
154 		return err;
155 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
156 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
157 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
158 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
159 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
160 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
161 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
162 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
163 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
164 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
165 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
166 	device_attr->max_rwq_indirection_table_size =
167 		attr_ex.rss_caps.max_rwq_indirection_table_size;
168 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
169 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
170 
171 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
172 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
173 	if (err)
174 		return err;
175 
176 	device_attr->flags = dv_attr.flags;
177 	device_attr->comp_mask = dv_attr.comp_mask;
178 #ifdef HAVE_IBV_MLX5_MOD_SWP
179 	device_attr->sw_parsing_offloads =
180 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
181 #endif
182 	device_attr->min_single_stride_log_num_of_bytes =
183 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
184 	device_attr->max_single_stride_log_num_of_bytes =
185 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
186 	device_attr->min_single_wqe_log_num_of_strides =
187 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
188 	device_attr->max_single_wqe_log_num_of_strides =
189 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
190 	device_attr->stride_supported_qpts =
191 		dv_attr.striding_rq_caps.supported_qpts;
192 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
193 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
194 #endif
195 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
196 		sizeof(device_attr->fw_ver));
197 
198 	return err;
199 }
200 
201 /**
202  * Detect misc5 support or not
203  *
204  * @param[in] priv
205  *   Device private data pointer
206  */
207 #ifdef HAVE_MLX5DV_DR
208 static void
209 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
210 {
211 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
212 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
213 	 * Case: IPv4--->UDP--->VxLAN--->vni
214 	 */
215 	void *tbl;
216 	struct mlx5_flow_dv_match_params matcher_mask;
217 	void *match_m;
218 	void *matcher;
219 	void *headers_m;
220 	void *misc5_m;
221 	uint32_t *tunnel_header_m;
222 	struct mlx5dv_flow_matcher_attr dv_attr;
223 
224 	memset(&matcher_mask, 0, sizeof(matcher_mask));
225 	matcher_mask.size = sizeof(matcher_mask.buf);
226 	match_m = matcher_mask.buf;
227 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
228 	misc5_m = MLX5_ADDR_OF(fte_match_param,
229 			       match_m, misc_parameters_5);
230 	tunnel_header_m = (uint32_t *)
231 				MLX5_ADDR_OF(fte_match_set_misc5,
232 				misc5_m, tunnel_header_1);
233 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
234 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
235 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
236 	*tunnel_header_m = 0xffffff;
237 
238 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
239 	if (!tbl) {
240 		DRV_LOG(INFO, "No SW steering support");
241 		return;
242 	}
243 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
244 	dv_attr.match_mask = (void *)&matcher_mask,
245 	dv_attr.match_criteria_enable =
246 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
247 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
248 	dv_attr.priority = 3;
249 #ifdef HAVE_MLX5DV_DR_ESWITCH
250 	void *misc2_m;
251 	if (priv->config.dv_esw_en) {
252 		/* FDB enabled reg_c_0 */
253 		dv_attr.match_criteria_enable |=
254 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
255 		misc2_m = MLX5_ADDR_OF(fte_match_param,
256 				       match_m, misc_parameters_2);
257 		MLX5_SET(fte_match_set_misc2, misc2_m,
258 			 metadata_reg_c_0, 0xffff);
259 	}
260 #endif
261 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
262 						    &dv_attr, tbl);
263 	if (matcher) {
264 		priv->sh->misc5_cap = 1;
265 		mlx5_glue->dv_destroy_flow_matcher(matcher);
266 	}
267 	mlx5_glue->dr_destroy_flow_tbl(tbl);
268 #else
269 	RTE_SET_USED(priv);
270 #endif
271 }
272 #endif
273 
274 /**
275  * Initialize DR related data within private structure.
276  * Routine checks the reference counter and does actual
277  * resources creation/initialization only if counter is zero.
278  *
279  * @param[in] priv
280  *   Pointer to the private device data structure.
281  *
282  * @return
283  *   Zero on success, positive error code otherwise.
284  */
285 static int
286 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
287 {
288 	struct mlx5_dev_ctx_shared *sh = priv->sh;
289 	char s[MLX5_NAME_SIZE] __rte_unused;
290 	int err;
291 
292 	MLX5_ASSERT(sh && sh->refcnt);
293 	if (sh->refcnt > 1)
294 		return 0;
295 	err = mlx5_alloc_table_hash_list(priv);
296 	if (err)
297 		goto error;
298 	/* The resources below are only valid with DV support. */
299 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
300 	/* Init port id action list. */
301 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
302 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
303 						   flow_dv_port_id_create_cb,
304 						   flow_dv_port_id_match_cb,
305 						   flow_dv_port_id_remove_cb,
306 						   flow_dv_port_id_clone_cb,
307 						 flow_dv_port_id_clone_free_cb);
308 	if (!sh->port_id_action_list)
309 		goto error;
310 	/* Init push vlan action list. */
311 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
312 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
313 						    flow_dv_push_vlan_create_cb,
314 						    flow_dv_push_vlan_match_cb,
315 						    flow_dv_push_vlan_remove_cb,
316 						    flow_dv_push_vlan_clone_cb,
317 					       flow_dv_push_vlan_clone_free_cb);
318 	if (!sh->push_vlan_action_list)
319 		goto error;
320 	/* Init sample action list. */
321 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
322 	sh->sample_action_list = mlx5_list_create(s, sh, true,
323 						  flow_dv_sample_create_cb,
324 						  flow_dv_sample_match_cb,
325 						  flow_dv_sample_remove_cb,
326 						  flow_dv_sample_clone_cb,
327 						  flow_dv_sample_clone_free_cb);
328 	if (!sh->sample_action_list)
329 		goto error;
330 	/* Init dest array action list. */
331 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
332 	sh->dest_array_list = mlx5_list_create(s, sh, true,
333 					       flow_dv_dest_array_create_cb,
334 					       flow_dv_dest_array_match_cb,
335 					       flow_dv_dest_array_remove_cb,
336 					       flow_dv_dest_array_clone_cb,
337 					      flow_dv_dest_array_clone_free_cb);
338 	if (!sh->dest_array_list)
339 		goto error;
340 	/* Init shared flex parsers list, no need lcore_share */
341 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
342 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
343 					       mlx5_flex_parser_create_cb,
344 					       mlx5_flex_parser_match_cb,
345 					       mlx5_flex_parser_remove_cb,
346 					       mlx5_flex_parser_clone_cb,
347 					       mlx5_flex_parser_clone_free_cb);
348 	if (!sh->flex_parsers_dv)
349 		goto error;
350 #endif
351 #ifdef HAVE_MLX5DV_DR
352 	void *domain;
353 
354 	/* Reference counter is zero, we should initialize structures. */
355 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
356 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
357 	if (!domain) {
358 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
359 		err = errno;
360 		goto error;
361 	}
362 	sh->rx_domain = domain;
363 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
364 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
365 	if (!domain) {
366 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
367 		err = errno;
368 		goto error;
369 	}
370 	sh->tx_domain = domain;
371 #ifdef HAVE_MLX5DV_DR_ESWITCH
372 	if (priv->config.dv_esw_en) {
373 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
374 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
375 		if (!domain) {
376 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
377 			err = errno;
378 			goto error;
379 		}
380 		sh->fdb_domain = domain;
381 	}
382 	/*
383 	 * The drop action is just some dummy placeholder in rdma-core. It
384 	 * does not belong to domains and has no any attributes, and, can be
385 	 * shared by the entire device.
386 	 */
387 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
388 	if (!sh->dr_drop_action) {
389 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
390 		err = errno;
391 		goto error;
392 	}
393 #endif
394 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
395 		err = mlx5_alloc_tunnel_hub(sh);
396 	if (err) {
397 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
398 		goto error;
399 	}
400 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
401 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
402 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
403 		if (sh->fdb_domain)
404 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
405 	}
406 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
407 	if (!priv->config.allow_duplicate_pattern) {
408 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
409 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
410 #endif
411 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
412 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
413 		if (sh->fdb_domain)
414 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
415 	}
416 
417 	__mlx5_discovery_misc5_cap(priv);
418 #endif /* HAVE_MLX5DV_DR */
419 	sh->default_miss_action =
420 			mlx5_glue->dr_create_flow_action_default_miss();
421 	if (!sh->default_miss_action)
422 		DRV_LOG(WARNING, "Default miss action is not supported.");
423 	LIST_INIT(&sh->shared_rxqs);
424 	return 0;
425 error:
426 	/* Rollback the created objects. */
427 	if (sh->rx_domain) {
428 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
429 		sh->rx_domain = NULL;
430 	}
431 	if (sh->tx_domain) {
432 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
433 		sh->tx_domain = NULL;
434 	}
435 	if (sh->fdb_domain) {
436 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
437 		sh->fdb_domain = NULL;
438 	}
439 	if (sh->dr_drop_action) {
440 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
441 		sh->dr_drop_action = NULL;
442 	}
443 	if (sh->pop_vlan_action) {
444 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
445 		sh->pop_vlan_action = NULL;
446 	}
447 	if (sh->encaps_decaps) {
448 		mlx5_hlist_destroy(sh->encaps_decaps);
449 		sh->encaps_decaps = NULL;
450 	}
451 	if (sh->modify_cmds) {
452 		mlx5_hlist_destroy(sh->modify_cmds);
453 		sh->modify_cmds = NULL;
454 	}
455 	if (sh->tag_table) {
456 		/* tags should be destroyed with flow before. */
457 		mlx5_hlist_destroy(sh->tag_table);
458 		sh->tag_table = NULL;
459 	}
460 	if (sh->tunnel_hub) {
461 		mlx5_release_tunnel_hub(sh, priv->dev_port);
462 		sh->tunnel_hub = NULL;
463 	}
464 	mlx5_free_table_hash_list(priv);
465 	if (sh->port_id_action_list) {
466 		mlx5_list_destroy(sh->port_id_action_list);
467 		sh->port_id_action_list = NULL;
468 	}
469 	if (sh->push_vlan_action_list) {
470 		mlx5_list_destroy(sh->push_vlan_action_list);
471 		sh->push_vlan_action_list = NULL;
472 	}
473 	if (sh->sample_action_list) {
474 		mlx5_list_destroy(sh->sample_action_list);
475 		sh->sample_action_list = NULL;
476 	}
477 	if (sh->dest_array_list) {
478 		mlx5_list_destroy(sh->dest_array_list);
479 		sh->dest_array_list = NULL;
480 	}
481 	return err;
482 }
483 
484 /**
485  * Destroy DR related data within private structure.
486  *
487  * @param[in] priv
488  *   Pointer to the private device data structure.
489  */
490 void
491 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
492 {
493 	struct mlx5_dev_ctx_shared *sh = priv->sh;
494 
495 	MLX5_ASSERT(sh && sh->refcnt);
496 	if (sh->refcnt > 1)
497 		return;
498 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
499 #ifdef HAVE_MLX5DV_DR
500 	if (sh->rx_domain) {
501 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
502 		sh->rx_domain = NULL;
503 	}
504 	if (sh->tx_domain) {
505 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
506 		sh->tx_domain = NULL;
507 	}
508 #ifdef HAVE_MLX5DV_DR_ESWITCH
509 	if (sh->fdb_domain) {
510 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
511 		sh->fdb_domain = NULL;
512 	}
513 	if (sh->dr_drop_action) {
514 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
515 		sh->dr_drop_action = NULL;
516 	}
517 #endif
518 	if (sh->pop_vlan_action) {
519 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
520 		sh->pop_vlan_action = NULL;
521 	}
522 #endif /* HAVE_MLX5DV_DR */
523 	if (sh->default_miss_action)
524 		mlx5_glue->destroy_flow_action
525 				(sh->default_miss_action);
526 	if (sh->encaps_decaps) {
527 		mlx5_hlist_destroy(sh->encaps_decaps);
528 		sh->encaps_decaps = NULL;
529 	}
530 	if (sh->modify_cmds) {
531 		mlx5_hlist_destroy(sh->modify_cmds);
532 		sh->modify_cmds = NULL;
533 	}
534 	if (sh->tag_table) {
535 		/* tags should be destroyed with flow before. */
536 		mlx5_hlist_destroy(sh->tag_table);
537 		sh->tag_table = NULL;
538 	}
539 	if (sh->tunnel_hub) {
540 		mlx5_release_tunnel_hub(sh, priv->dev_port);
541 		sh->tunnel_hub = NULL;
542 	}
543 	mlx5_free_table_hash_list(priv);
544 	if (sh->port_id_action_list) {
545 		mlx5_list_destroy(sh->port_id_action_list);
546 		sh->port_id_action_list = NULL;
547 	}
548 	if (sh->push_vlan_action_list) {
549 		mlx5_list_destroy(sh->push_vlan_action_list);
550 		sh->push_vlan_action_list = NULL;
551 	}
552 	if (sh->sample_action_list) {
553 		mlx5_list_destroy(sh->sample_action_list);
554 		sh->sample_action_list = NULL;
555 	}
556 	if (sh->dest_array_list) {
557 		mlx5_list_destroy(sh->dest_array_list);
558 		sh->dest_array_list = NULL;
559 	}
560 }
561 
562 /**
563  * Initialize shared data between primary and secondary process.
564  *
565  * A memzone is reserved by primary process and secondary processes attach to
566  * the memzone.
567  *
568  * @return
569  *   0 on success, a negative errno value otherwise and rte_errno is set.
570  */
571 static int
572 mlx5_init_shared_data(void)
573 {
574 	const struct rte_memzone *mz;
575 	int ret = 0;
576 
577 	rte_spinlock_lock(&mlx5_shared_data_lock);
578 	if (mlx5_shared_data == NULL) {
579 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
580 			/* Allocate shared memory. */
581 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
582 						 sizeof(*mlx5_shared_data),
583 						 SOCKET_ID_ANY, 0);
584 			if (mz == NULL) {
585 				DRV_LOG(ERR,
586 					"Cannot allocate mlx5 shared data");
587 				ret = -rte_errno;
588 				goto error;
589 			}
590 			mlx5_shared_data = mz->addr;
591 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
592 			rte_spinlock_init(&mlx5_shared_data->lock);
593 		} else {
594 			/* Lookup allocated shared memory. */
595 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
596 			if (mz == NULL) {
597 				DRV_LOG(ERR,
598 					"Cannot attach mlx5 shared data");
599 				ret = -rte_errno;
600 				goto error;
601 			}
602 			mlx5_shared_data = mz->addr;
603 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
604 		}
605 	}
606 error:
607 	rte_spinlock_unlock(&mlx5_shared_data_lock);
608 	return ret;
609 }
610 
611 /**
612  * PMD global initialization.
613  *
614  * Independent from individual device, this function initializes global
615  * per-PMD data structures distinguishing primary and secondary processes.
616  * Hence, each initialization is called once per a process.
617  *
618  * @return
619  *   0 on success, a negative errno value otherwise and rte_errno is set.
620  */
621 static int
622 mlx5_init_once(void)
623 {
624 	struct mlx5_shared_data *sd;
625 	struct mlx5_local_data *ld = &mlx5_local_data;
626 	int ret = 0;
627 
628 	if (mlx5_init_shared_data())
629 		return -rte_errno;
630 	sd = mlx5_shared_data;
631 	MLX5_ASSERT(sd);
632 	rte_spinlock_lock(&sd->lock);
633 	switch (rte_eal_process_type()) {
634 	case RTE_PROC_PRIMARY:
635 		if (sd->init_done)
636 			break;
637 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
638 					   mlx5_mp_os_primary_handle);
639 		if (ret)
640 			goto out;
641 		sd->init_done = true;
642 		break;
643 	case RTE_PROC_SECONDARY:
644 		if (ld->init_done)
645 			break;
646 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
647 					     mlx5_mp_os_secondary_handle);
648 		if (ret)
649 			goto out;
650 		++sd->secondary_cnt;
651 		ld->init_done = true;
652 		break;
653 	default:
654 		break;
655 	}
656 out:
657 	rte_spinlock_unlock(&sd->lock);
658 	return ret;
659 }
660 
661 /**
662  * DV flow counter mode detect and config.
663  *
664  * @param dev
665  *   Pointer to rte_eth_dev structure.
666  *
667  */
668 static void
669 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
670 {
671 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
672 	struct mlx5_priv *priv = dev->data->dev_private;
673 	struct mlx5_dev_ctx_shared *sh = priv->sh;
674 	bool fallback;
675 
676 #ifndef HAVE_IBV_DEVX_ASYNC
677 	fallback = true;
678 #else
679 	fallback = false;
680 	if (!sh->devx || !priv->config.dv_flow_en ||
681 	    !priv->config.hca_attr.flow_counters_dump ||
682 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
683 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
684 		fallback = true;
685 #endif
686 	if (fallback)
687 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
688 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
689 			priv->config.hca_attr.flow_counters_dump,
690 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
691 	/* Initialize fallback mode only on the port initializes sh. */
692 	if (sh->refcnt == 1)
693 		sh->cmng.counter_fallback = fallback;
694 	else if (fallback != sh->cmng.counter_fallback)
695 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
696 			"with others:%d.", PORT_ID(priv), fallback);
697 #endif
698 }
699 
700 /**
701  * DR flow drop action support detect.
702  *
703  * @param dev
704  *   Pointer to rte_eth_dev structure.
705  *
706  */
707 static void
708 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
709 {
710 #ifdef HAVE_MLX5DV_DR
711 	struct mlx5_priv *priv = dev->data->dev_private;
712 
713 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
714 		return;
715 	/**
716 	 * DR supports drop action placeholder when it is supported;
717 	 * otherwise, use the queue drop action.
718 	 */
719 	if (!priv->sh->drop_action_check_flag) {
720 		if (!mlx5_flow_discover_dr_action_support(dev))
721 			priv->sh->dr_drop_action_en = 1;
722 		priv->sh->drop_action_check_flag = 1;
723 	}
724 	if (priv->sh->dr_drop_action_en)
725 		priv->root_drop_action = priv->sh->dr_drop_action;
726 	else
727 		priv->root_drop_action = priv->drop_queue.hrxq->action;
728 #endif
729 }
730 
731 static void
732 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
733 {
734 	struct mlx5_priv *priv = dev->data->dev_private;
735 	void *ctx = priv->sh->cdev->ctx;
736 
737 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
738 	if (!priv->q_counters) {
739 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
740 		struct ibv_wq *wq;
741 
742 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
743 			"by DevX - fall-back to use the kernel driver global "
744 			"queue counter.", dev->data->port_id);
745 		/* Create WQ by kernel and query its queue counter ID. */
746 		if (cq) {
747 			wq = mlx5_glue->create_wq(ctx,
748 						  &(struct ibv_wq_init_attr){
749 						    .wq_type = IBV_WQT_RQ,
750 						    .max_wr = 1,
751 						    .max_sge = 1,
752 						    .pd = priv->sh->cdev->pd,
753 						    .cq = cq,
754 						});
755 			if (wq) {
756 				/* Counter is assigned only on RDY state. */
757 				int ret = mlx5_glue->modify_wq(wq,
758 						 &(struct ibv_wq_attr){
759 						 .attr_mask = IBV_WQ_ATTR_STATE,
760 						 .wq_state = IBV_WQS_RDY,
761 						});
762 
763 				if (ret == 0)
764 					mlx5_devx_cmd_wq_query(wq,
765 							 &priv->counter_set_id);
766 				claim_zero(mlx5_glue->destroy_wq(wq));
767 			}
768 			claim_zero(mlx5_glue->destroy_cq(cq));
769 		}
770 	} else {
771 		priv->counter_set_id = priv->q_counters->id;
772 	}
773 	if (priv->counter_set_id == 0)
774 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
775 			"available.", dev->data->port_id);
776 }
777 
778 /**
779  * Check if representor spawn info match devargs.
780  *
781  * @param spawn
782  *   Verbs device parameters (name, port, switch_info) to spawn.
783  * @param eth_da
784  *   Device devargs to probe.
785  *
786  * @return
787  *   Match result.
788  */
789 static bool
790 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
791 		       struct rte_eth_devargs *eth_da)
792 {
793 	struct mlx5_switch_info *switch_info = &spawn->info;
794 	unsigned int p, f;
795 	uint16_t id;
796 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
797 						      eth_da->type);
798 
799 	switch (eth_da->type) {
800 	case RTE_ETH_REPRESENTOR_SF:
801 		if (!(spawn->info.port_name == -1 &&
802 		      switch_info->name_type ==
803 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
804 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
805 			rte_errno = EBUSY;
806 			return false;
807 		}
808 		break;
809 	case RTE_ETH_REPRESENTOR_VF:
810 		/* Allows HPF representor index -1 as exception. */
811 		if (!(spawn->info.port_name == -1 &&
812 		      switch_info->name_type ==
813 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
814 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
815 			rte_errno = EBUSY;
816 			return false;
817 		}
818 		break;
819 	case RTE_ETH_REPRESENTOR_NONE:
820 		rte_errno = EBUSY;
821 		return false;
822 	default:
823 		rte_errno = ENOTSUP;
824 		DRV_LOG(ERR, "unsupported representor type");
825 		return false;
826 	}
827 	/* Check representor ID: */
828 	for (p = 0; p < eth_da->nb_ports; ++p) {
829 		if (spawn->pf_bond < 0) {
830 			/* For non-LAG mode, allow and ignore pf. */
831 			switch_info->pf_num = eth_da->ports[p];
832 			repr_id = mlx5_representor_id_encode(switch_info,
833 							     eth_da->type);
834 		}
835 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
836 			id = MLX5_REPRESENTOR_ID
837 				(eth_da->ports[p], eth_da->type,
838 				 eth_da->representor_ports[f]);
839 			if (repr_id == id)
840 				return true;
841 		}
842 	}
843 	rte_errno = EBUSY;
844 	return false;
845 }
846 
847 /**
848  * Spawn an Ethernet device from Verbs information.
849  *
850  * @param dpdk_dev
851  *   Backing DPDK device.
852  * @param spawn
853  *   Verbs device parameters (name, port, switch_info) to spawn.
854  * @param config
855  *   Device configuration parameters.
856  * @param eth_da
857  *   Device arguments.
858  *
859  * @return
860  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
861  *   is set. The following errors are defined:
862  *
863  *   EBUSY: device is not supposed to be spawned.
864  *   EEXIST: device is already spawned
865  */
866 static struct rte_eth_dev *
867 mlx5_dev_spawn(struct rte_device *dpdk_dev,
868 	       struct mlx5_dev_spawn_data *spawn,
869 	       struct mlx5_dev_config *config,
870 	       struct rte_eth_devargs *eth_da)
871 {
872 	const struct mlx5_switch_info *switch_info = &spawn->info;
873 	struct mlx5_dev_ctx_shared *sh = NULL;
874 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
875 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
876 	struct rte_eth_dev *eth_dev = NULL;
877 	struct mlx5_priv *priv = NULL;
878 	int err = 0;
879 	unsigned int hw_padding = 0;
880 	unsigned int mps;
881 	unsigned int mpls_en = 0;
882 	unsigned int swp = 0;
883 	unsigned int mprq = 0;
884 	unsigned int mprq_min_stride_size_n = 0;
885 	unsigned int mprq_max_stride_size_n = 0;
886 	unsigned int mprq_min_stride_num_n = 0;
887 	unsigned int mprq_max_stride_num_n = 0;
888 	struct rte_ether_addr mac;
889 	char name[RTE_ETH_NAME_MAX_LEN];
890 	int own_domain_id = 0;
891 	uint16_t port_id;
892 	struct mlx5_port_info vport_info = { .query_flags = 0 };
893 	int nl_rdma = -1;
894 	int i;
895 
896 	/* Determine if this port representor is supposed to be spawned. */
897 	if (switch_info->representor && dpdk_dev->devargs &&
898 	    !mlx5_representor_match(spawn, eth_da))
899 		return NULL;
900 	/* Build device name. */
901 	if (spawn->pf_bond < 0) {
902 		/* Single device. */
903 		if (!switch_info->representor)
904 			strlcpy(name, dpdk_dev->name, sizeof(name));
905 		else
906 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
907 				 dpdk_dev->name,
908 				 switch_info->name_type ==
909 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
910 				 switch_info->port_name);
911 	} else {
912 		/* Bonding device. */
913 		if (!switch_info->representor) {
914 			err = snprintf(name, sizeof(name), "%s_%s",
915 				       dpdk_dev->name, spawn->phys_dev_name);
916 		} else {
917 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
918 				dpdk_dev->name, spawn->phys_dev_name,
919 				switch_info->ctrl_num,
920 				switch_info->pf_num,
921 				switch_info->name_type ==
922 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
923 				switch_info->port_name);
924 		}
925 	}
926 	if (err >= (int)sizeof(name))
927 		DRV_LOG(WARNING, "device name overflow %s", name);
928 	/* check if the device is already spawned */
929 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
930 		rte_errno = EEXIST;
931 		return NULL;
932 	}
933 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
934 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
935 		struct mlx5_mp_id mp_id;
936 
937 		eth_dev = rte_eth_dev_attach_secondary(name);
938 		if (eth_dev == NULL) {
939 			DRV_LOG(ERR, "can not attach rte ethdev");
940 			rte_errno = ENOMEM;
941 			return NULL;
942 		}
943 		eth_dev->device = dpdk_dev;
944 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
945 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
946 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
947 		err = mlx5_proc_priv_init(eth_dev);
948 		if (err)
949 			return NULL;
950 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
951 		/* Receive command fd from primary process */
952 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
953 		if (err < 0)
954 			goto err_secondary;
955 		/* Remap UAR for Tx queues. */
956 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
957 		if (err)
958 			goto err_secondary;
959 		/*
960 		 * Ethdev pointer is still required as input since
961 		 * the primary device is not accessible from the
962 		 * secondary process.
963 		 */
964 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
965 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
966 		return eth_dev;
967 err_secondary:
968 		mlx5_dev_close(eth_dev);
969 		return NULL;
970 	}
971 	/*
972 	 * Some parameters ("tx_db_nc" in particularly) are needed in
973 	 * advance to create dv/verbs device context. We proceed the
974 	 * devargs here to get ones, and later proceed devargs again
975 	 * to override some hardware settings.
976 	 */
977 	err = mlx5_args(config, dpdk_dev->devargs);
978 	if (err) {
979 		err = rte_errno;
980 		DRV_LOG(ERR, "failed to process device arguments: %s",
981 			strerror(rte_errno));
982 		goto error;
983 	}
984 	if (config->dv_miss_info) {
985 		if (switch_info->master || switch_info->representor)
986 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
987 	}
988 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
989 	if (!sh)
990 		return NULL;
991 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
992 	config->dest_tir = 1;
993 #endif
994 #ifdef HAVE_IBV_MLX5_MOD_SWP
995 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
996 #endif
997 	/*
998 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
999 	 * as all ConnectX-5 devices.
1000 	 */
1001 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1002 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1003 #endif
1004 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1005 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1006 #endif
1007 	mlx5_glue->dv_query_device(sh->cdev->ctx, &dv_attr);
1008 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1009 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1010 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1011 			mps = MLX5_MPW_ENHANCED;
1012 		} else {
1013 			DRV_LOG(DEBUG, "MPW is supported");
1014 			mps = MLX5_MPW;
1015 		}
1016 	} else {
1017 		DRV_LOG(DEBUG, "MPW isn't supported");
1018 		mps = MLX5_MPW_DISABLED;
1019 	}
1020 #ifdef HAVE_IBV_MLX5_MOD_SWP
1021 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1022 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1023 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1024 #endif
1025 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
1026 		MLX5_SW_PARSING_TSO_CAP);
1027 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1028 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1029 		struct mlx5dv_striding_rq_caps mprq_caps =
1030 			dv_attr.striding_rq_caps;
1031 
1032 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1033 			mprq_caps.min_single_stride_log_num_of_bytes);
1034 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1035 			mprq_caps.max_single_stride_log_num_of_bytes);
1036 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1037 			mprq_caps.min_single_wqe_log_num_of_strides);
1038 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1039 			mprq_caps.max_single_wqe_log_num_of_strides);
1040 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1041 			mprq_caps.supported_qpts);
1042 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1043 		mprq = 1;
1044 		mprq_min_stride_size_n =
1045 			mprq_caps.min_single_stride_log_num_of_bytes;
1046 		mprq_max_stride_size_n =
1047 			mprq_caps.max_single_stride_log_num_of_bytes;
1048 		mprq_min_stride_num_n =
1049 			mprq_caps.min_single_wqe_log_num_of_strides;
1050 		mprq_max_stride_num_n =
1051 			mprq_caps.max_single_wqe_log_num_of_strides;
1052 	}
1053 #endif
1054 	/* Rx CQE compression is enabled by default. */
1055 	config->cqe_comp = 1;
1056 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1057 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1058 		config->tunnel_en = dv_attr.tunnel_offloads_caps &
1059 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1060 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1061 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1062 	}
1063 	if (config->tunnel_en) {
1064 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1065 		config->tunnel_en &
1066 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1067 		config->tunnel_en &
1068 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1069 		config->tunnel_en &
1070 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1071 		);
1072 	} else {
1073 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1074 	}
1075 #else
1076 	DRV_LOG(WARNING,
1077 		"tunnel offloading disabled due to old OFED/rdma-core version");
1078 #endif
1079 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1080 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1081 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1082 		   (dv_attr.tunnel_offloads_caps &
1083 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1084 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1085 		mpls_en ? "" : "not ");
1086 #else
1087 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1088 		" old OFED/rdma-core version or firmware configuration");
1089 #endif
1090 	config->mpls_en = mpls_en;
1091 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1092 	/* Check port status. */
1093 	if (spawn->phys_port <= UINT8_MAX) {
1094 		/* Legacy Verbs api only support u8 port number. */
1095 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1096 					    &port_attr);
1097 		if (err) {
1098 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1099 			goto error;
1100 		}
1101 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1102 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1103 			err = EINVAL;
1104 			goto error;
1105 		}
1106 	} else if (nl_rdma >= 0) {
1107 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1108 		err = mlx5_nl_port_state(nl_rdma,
1109 			spawn->phys_dev_name,
1110 			spawn->phys_port);
1111 		if (err < 0) {
1112 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1113 				strerror(rte_errno));
1114 			err = -rte_errno;
1115 			goto error;
1116 		}
1117 		port_attr.state = (enum ibv_port_state)err;
1118 	}
1119 	if (port_attr.state != IBV_PORT_ACTIVE)
1120 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1121 			mlx5_glue->port_state_str(port_attr.state),
1122 			port_attr.state);
1123 	/* Allocate private eth device data. */
1124 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1125 			   sizeof(*priv),
1126 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1127 	if (priv == NULL) {
1128 		DRV_LOG(ERR, "priv allocation failure");
1129 		err = ENOMEM;
1130 		goto error;
1131 	}
1132 	priv->sh = sh;
1133 	priv->dev_port = spawn->phys_port;
1134 	priv->pci_dev = spawn->pci_dev;
1135 	priv->mtu = RTE_ETHER_MTU;
1136 	/* Some internal functions rely on Netlink sockets, open them now. */
1137 	priv->nl_socket_rdma = nl_rdma;
1138 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1139 	priv->representor = !!switch_info->representor;
1140 	priv->master = !!switch_info->master;
1141 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1142 	priv->vport_meta_tag = 0;
1143 	priv->vport_meta_mask = 0;
1144 	priv->pf_bond = spawn->pf_bond;
1145 
1146 	DRV_LOG(DEBUG,
1147 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1148 		priv->dev_port, dpdk_dev->bus->name,
1149 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1150 		priv->master, priv->representor, priv->pf_bond);
1151 
1152 	/*
1153 	 * If we have E-Switch we should determine the vport attributes.
1154 	 * E-Switch may use either source vport field or reg_c[0] metadata
1155 	 * register to match on vport index. The engaged part of metadata
1156 	 * register is defined by mask.
1157 	 */
1158 	if (switch_info->representor || switch_info->master) {
1159 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1160 						 spawn->phys_port,
1161 						 &vport_info);
1162 		if (err) {
1163 			DRV_LOG(WARNING,
1164 				"Cannot query devx port %d on device %s",
1165 				spawn->phys_port, spawn->phys_dev_name);
1166 			vport_info.query_flags = 0;
1167 		}
1168 	}
1169 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1170 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1171 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1172 		if (!priv->vport_meta_mask) {
1173 			DRV_LOG(ERR,
1174 				"vport zero mask for port %d on bonding device %s",
1175 				spawn->phys_port, spawn->phys_dev_name);
1176 			err = ENOTSUP;
1177 			goto error;
1178 		}
1179 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1180 			DRV_LOG(ERR,
1181 				"Invalid vport tag for port %d on bonding device %s",
1182 				spawn->phys_port, spawn->phys_dev_name);
1183 			err = ENOTSUP;
1184 			goto error;
1185 		}
1186 	}
1187 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1188 		priv->vport_id = vport_info.vport_id;
1189 	} else if (spawn->pf_bond >= 0 &&
1190 		   (switch_info->representor || switch_info->master)) {
1191 		DRV_LOG(ERR,
1192 			"Cannot deduce vport index for port %d on bonding device %s",
1193 			spawn->phys_port, spawn->phys_dev_name);
1194 		err = ENOTSUP;
1195 		goto error;
1196 	} else {
1197 		/*
1198 		 * Suppose vport index in compatible way. Kernel/rdma_core
1199 		 * support single E-Switch per PF configurations only and
1200 		 * vport_id field contains the vport index for associated VF,
1201 		 * which is deduced from representor port name.
1202 		 * For example, let's have the IB device port 10, it has
1203 		 * attached network device eth0, which has port name attribute
1204 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1205 		 * as 3 (2+1). This assigning schema should be changed if the
1206 		 * multiple E-Switch instances per PF configurations or/and PCI
1207 		 * subfunctions are added.
1208 		 */
1209 		priv->vport_id = switch_info->representor ?
1210 				 switch_info->port_name + 1 : -1;
1211 	}
1212 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1213 							  eth_da->type);
1214 	/*
1215 	 * Look for sibling devices in order to reuse their switch domain
1216 	 * if any, otherwise allocate one.
1217 	 */
1218 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1219 		const struct mlx5_priv *opriv =
1220 			rte_eth_devices[port_id].data->dev_private;
1221 
1222 		if (!opriv ||
1223 		    opriv->sh != priv->sh ||
1224 			opriv->domain_id ==
1225 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1226 			continue;
1227 		priv->domain_id = opriv->domain_id;
1228 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1229 			priv->dev_port, priv->domain_id);
1230 		break;
1231 	}
1232 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1233 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1234 		if (err) {
1235 			err = rte_errno;
1236 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1237 				strerror(rte_errno));
1238 			goto error;
1239 		}
1240 		own_domain_id = 1;
1241 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1242 			priv->dev_port, priv->domain_id);
1243 	}
1244 	/* Override some values set by hardware configuration. */
1245 	mlx5_args(config, dpdk_dev->devargs);
1246 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
1247 	if (err)
1248 		goto error;
1249 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1250 			    IBV_DEVICE_RAW_IP_CSUM);
1251 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1252 		(config->hw_csum ? "" : "not "));
1253 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1254 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1255 	DRV_LOG(DEBUG, "counters are not supported");
1256 #endif
1257 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1258 	if (config->dv_flow_en) {
1259 		DRV_LOG(WARNING, "DV flow is not supported");
1260 		config->dv_flow_en = 0;
1261 	}
1262 #endif
1263 	config->ind_table_max_size =
1264 		sh->device_attr.max_rwq_indirection_table_size;
1265 	/*
1266 	 * Remove this check once DPDK supports larger/variable
1267 	 * indirection tables.
1268 	 */
1269 	if (config->ind_table_max_size > (unsigned int)RTE_ETH_RSS_RETA_SIZE_512)
1270 		config->ind_table_max_size = RTE_ETH_RSS_RETA_SIZE_512;
1271 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1272 		config->ind_table_max_size);
1273 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1274 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1275 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1276 		(config->hw_vlan_strip ? "" : "not "));
1277 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1278 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1279 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1280 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1281 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1282 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1283 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1284 #endif
1285 	if (config->hw_padding && !hw_padding) {
1286 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1287 		config->hw_padding = 0;
1288 	} else if (config->hw_padding) {
1289 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1290 	}
1291 	config->tso = (sh->device_attr.max_tso > 0 &&
1292 		      (sh->device_attr.tso_supported_qpts &
1293 		       (1 << IBV_QPT_RAW_PACKET)));
1294 	if (config->tso)
1295 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1296 	/*
1297 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1298 	 * by default.
1299 	 */
1300 	if (config->mps == MLX5_ARG_UNSET)
1301 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1302 							  MLX5_MPW_DISABLED;
1303 	else
1304 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1305 	DRV_LOG(INFO, "%sMPS is %s",
1306 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1307 		config->mps == MLX5_MPW ? "legacy " : "",
1308 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1309 	if (sh->devx) {
1310 		config->hca_attr = sh->cdev->config.hca_attr;
1311 		sh->steering_format_version =
1312 			config->hca_attr.steering_format_version;
1313 		/* Check for LRO support. */
1314 		if (config->dest_tir && config->hca_attr.lro_cap &&
1315 		    config->dv_flow_en) {
1316 			/* TBD check tunnel lro caps. */
1317 			config->lro.supported = config->hca_attr.lro_cap;
1318 			DRV_LOG(DEBUG, "Device supports LRO");
1319 			/*
1320 			 * If LRO timeout is not configured by application,
1321 			 * use the minimal supported value.
1322 			 */
1323 			if (!config->lro.timeout)
1324 				config->lro.timeout =
1325 				config->hca_attr.lro_timer_supported_periods[0];
1326 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1327 				config->lro.timeout);
1328 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1329 				"required for coalescing is %d bytes",
1330 				config->hca_attr.lro_min_mss_size);
1331 		}
1332 #if defined(HAVE_MLX5DV_DR) && \
1333 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1334 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1335 		if (config->hca_attr.qos.sup &&
1336 		    config->hca_attr.qos.flow_meter_old &&
1337 		    config->dv_flow_en) {
1338 			uint8_t reg_c_mask =
1339 				config->hca_attr.qos.flow_meter_reg_c_ids;
1340 			/*
1341 			 * Meter needs two REG_C's for color match and pre-sfx
1342 			 * flow match. Here get the REG_C for color match.
1343 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1344 			 */
1345 			reg_c_mask &= 0xfc;
1346 			if (__builtin_popcount(reg_c_mask) < 1) {
1347 				priv->mtr_en = 0;
1348 				DRV_LOG(WARNING, "No available register for"
1349 					" meter.");
1350 			} else {
1351 				/*
1352 				 * The meter color register is used by the
1353 				 * flow-hit feature as well.
1354 				 * The flow-hit feature must use REG_C_3
1355 				 * Prefer REG_C_3 if it is available.
1356 				 */
1357 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1358 					priv->mtr_color_reg = REG_C_3;
1359 				else
1360 					priv->mtr_color_reg = ffs(reg_c_mask)
1361 							      - 1 + REG_C_0;
1362 				priv->mtr_en = 1;
1363 				priv->mtr_reg_share =
1364 				      config->hca_attr.qos.flow_meter;
1365 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1366 					priv->mtr_color_reg);
1367 			}
1368 		}
1369 		if (config->hca_attr.qos.sup &&
1370 			config->hca_attr.qos.flow_meter_aso_sup) {
1371 			uint32_t log_obj_size =
1372 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1373 			if (log_obj_size >=
1374 			config->hca_attr.qos.log_meter_aso_granularity &&
1375 			log_obj_size <=
1376 			config->hca_attr.qos.log_meter_aso_max_alloc)
1377 				sh->meter_aso_en = 1;
1378 		}
1379 		if (priv->mtr_en) {
1380 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1381 			if (err) {
1382 				err = -err;
1383 				goto error;
1384 			}
1385 		}
1386 		if (config->hca_attr.flow.tunnel_header_0_1)
1387 			sh->tunnel_header_0_1 = 1;
1388 #endif
1389 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1390 		if (config->hca_attr.flow_hit_aso &&
1391 		    priv->mtr_color_reg == REG_C_3) {
1392 			sh->flow_hit_aso_en = 1;
1393 			err = mlx5_flow_aso_age_mng_init(sh);
1394 			if (err) {
1395 				err = -err;
1396 				goto error;
1397 			}
1398 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1399 		}
1400 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1401 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1402 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1403 		if (config->hca_attr.ct_offload &&
1404 		    priv->mtr_color_reg == REG_C_3) {
1405 			err = mlx5_flow_aso_ct_mng_init(sh);
1406 			if (err) {
1407 				err = -err;
1408 				goto error;
1409 			}
1410 			DRV_LOG(DEBUG, "CT ASO is supported.");
1411 			sh->ct_aso_en = 1;
1412 		}
1413 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1414 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1415 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1416 		    config->dv_flow_en) {
1417 			priv->sampler_en = 1;
1418 			DRV_LOG(DEBUG, "Sampler enabled!");
1419 		} else {
1420 			priv->sampler_en = 0;
1421 			if (!config->hca_attr.log_max_ft_sampler_num)
1422 				DRV_LOG(WARNING,
1423 					"No available register for sampler.");
1424 			else
1425 				DRV_LOG(DEBUG, "DV flow is not supported!");
1426 		}
1427 #endif
1428 	}
1429 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1430 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1431 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1432 		config->cqe_comp = 0;
1433 	}
1434 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1435 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1436 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1437 				 " format isn't supported.");
1438 		config->cqe_comp = 0;
1439 	}
1440 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1441 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1442 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1443 				 " format isn't supported.");
1444 		config->cqe_comp = 0;
1445 	}
1446 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1447 			config->cqe_comp ? "" : "not ");
1448 	if (config->tx_pp) {
1449 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1450 			config->hca_attr.dev_freq_khz);
1451 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1452 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1453 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1454 			config->hca_attr.cross_channel ? "" : "not ");
1455 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1456 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1457 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1458 			config->hca_attr.non_wire_sq ? "" : "not ");
1459 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1460 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1461 			config->hca_attr.log_max_static_sq_wq);
1462 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1463 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1464 		if (!sh->devx) {
1465 			DRV_LOG(ERR, "DevX is required for packet pacing");
1466 			err = ENODEV;
1467 			goto error;
1468 		}
1469 		if (!config->hca_attr.qos.packet_pacing) {
1470 			DRV_LOG(ERR, "Packet pacing is not supported");
1471 			err = ENODEV;
1472 			goto error;
1473 		}
1474 		if (!config->hca_attr.cross_channel) {
1475 			DRV_LOG(ERR, "Cross channel operations are"
1476 				     " required for packet pacing");
1477 			err = ENODEV;
1478 			goto error;
1479 		}
1480 		if (!config->hca_attr.wqe_index_ignore) {
1481 			DRV_LOG(ERR, "WQE index ignore feature is"
1482 				     " required for packet pacing");
1483 			err = ENODEV;
1484 			goto error;
1485 		}
1486 		if (!config->hca_attr.non_wire_sq) {
1487 			DRV_LOG(ERR, "Non-wire SQ feature is"
1488 				     " required for packet pacing");
1489 			err = ENODEV;
1490 			goto error;
1491 		}
1492 		if (!config->hca_attr.log_max_static_sq_wq) {
1493 			DRV_LOG(ERR, "Static WQE SQ feature is"
1494 				     " required for packet pacing");
1495 			err = ENODEV;
1496 			goto error;
1497 		}
1498 		if (!config->hca_attr.qos.wqe_rate_pp) {
1499 			DRV_LOG(ERR, "WQE rate mode is required"
1500 				     " for packet pacing");
1501 			err = ENODEV;
1502 			goto error;
1503 		}
1504 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1505 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1506 			     " can't create queues for packet pacing");
1507 		err = ENODEV;
1508 		goto error;
1509 #endif
1510 	}
1511 	if (config->std_delay_drop || config->hp_delay_drop) {
1512 		if (!config->hca_attr.rq_delay_drop) {
1513 			config->std_delay_drop = 0;
1514 			config->hp_delay_drop = 0;
1515 			DRV_LOG(WARNING,
1516 				"dev_port-%u: Rxq delay drop is not supported",
1517 				priv->dev_port);
1518 		}
1519 	}
1520 	if (sh->devx) {
1521 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1522 
1523 		err = config->hca_attr.access_register_user ?
1524 			mlx5_devx_cmd_register_read
1525 				(sh->cdev->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1526 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1527 		if (!err) {
1528 			uint32_t ts_mode;
1529 
1530 			/* MTUTC register is read successfully. */
1531 			ts_mode = MLX5_GET(register_mtutc, reg,
1532 					   time_stamp_mode);
1533 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1534 				config->rt_timestamp = 1;
1535 		} else {
1536 			/* Kernel does not support register reading. */
1537 			if (config->hca_attr.dev_freq_khz ==
1538 						 (NS_PER_S / MS_PER_S))
1539 				config->rt_timestamp = 1;
1540 		}
1541 	}
1542 	/*
1543 	 * If HW has bug working with tunnel packet decapsulation and
1544 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1545 	 * bit. Then RTE_ETH_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1546 	 */
1547 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1548 		config->hw_fcs_strip = 0;
1549 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1550 		(config->hw_fcs_strip ? "" : "not "));
1551 	if (config->mprq.enabled && mprq) {
1552 		if (config->mprq.stride_num_n &&
1553 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1554 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1555 			config->mprq.stride_num_n =
1556 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1557 						mprq_min_stride_num_n),
1558 					mprq_max_stride_num_n);
1559 			DRV_LOG(WARNING,
1560 				"the number of strides"
1561 				" for Multi-Packet RQ is out of range,"
1562 				" setting default value (%u)",
1563 				1 << config->mprq.stride_num_n);
1564 		}
1565 		if (config->mprq.stride_size_n &&
1566 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1567 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1568 			config->mprq.stride_size_n =
1569 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1570 						mprq_min_stride_size_n),
1571 					mprq_max_stride_size_n);
1572 			DRV_LOG(WARNING,
1573 				"the size of a stride"
1574 				" for Multi-Packet RQ is out of range,"
1575 				" setting default value (%u)",
1576 				1 << config->mprq.stride_size_n);
1577 		}
1578 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1579 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1580 	} else if (config->mprq.enabled && !mprq) {
1581 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1582 		config->mprq.enabled = 0;
1583 	}
1584 	if (config->max_dump_files_num == 0)
1585 		config->max_dump_files_num = 128;
1586 	eth_dev = rte_eth_dev_allocate(name);
1587 	if (eth_dev == NULL) {
1588 		DRV_LOG(ERR, "can not allocate rte ethdev");
1589 		err = ENOMEM;
1590 		goto error;
1591 	}
1592 	if (priv->representor) {
1593 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1594 		eth_dev->data->representor_id = priv->representor_id;
1595 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1596 			struct mlx5_priv *opriv =
1597 				rte_eth_devices[port_id].data->dev_private;
1598 			if (opriv &&
1599 			    opriv->master &&
1600 			    opriv->domain_id == priv->domain_id &&
1601 			    opriv->sh == priv->sh) {
1602 				eth_dev->data->backer_port_id = port_id;
1603 				break;
1604 			}
1605 		}
1606 		if (port_id >= RTE_MAX_ETHPORTS)
1607 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1608 	}
1609 	priv->mp_id.port_id = eth_dev->data->port_id;
1610 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1611 	/*
1612 	 * Store associated network device interface index. This index
1613 	 * is permanent throughout the lifetime of device. So, we may store
1614 	 * the ifindex here and use the cached value further.
1615 	 */
1616 	MLX5_ASSERT(spawn->ifindex);
1617 	priv->if_index = spawn->ifindex;
1618 	priv->lag_affinity_idx = sh->refcnt - 1;
1619 	eth_dev->data->dev_private = priv;
1620 	priv->dev_data = eth_dev->data;
1621 	eth_dev->data->mac_addrs = priv->mac;
1622 	eth_dev->device = dpdk_dev;
1623 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1624 	/* Configure the first MAC address by default. */
1625 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1626 		DRV_LOG(ERR,
1627 			"port %u cannot get MAC address, is mlx5_en"
1628 			" loaded? (errno: %s)",
1629 			eth_dev->data->port_id, strerror(rte_errno));
1630 		err = ENODEV;
1631 		goto error;
1632 	}
1633 	DRV_LOG(INFO,
1634 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1635 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1636 #ifdef RTE_LIBRTE_MLX5_DEBUG
1637 	{
1638 		char ifname[MLX5_NAMESIZE];
1639 
1640 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1641 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1642 				eth_dev->data->port_id, ifname);
1643 		else
1644 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1645 				eth_dev->data->port_id);
1646 	}
1647 #endif
1648 	/* Get actual MTU if possible. */
1649 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1650 	if (err) {
1651 		err = rte_errno;
1652 		goto error;
1653 	}
1654 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1655 		priv->mtu);
1656 	/* Initialize burst functions to prevent crashes before link-up. */
1657 	eth_dev->rx_pkt_burst = removed_rx_burst;
1658 	eth_dev->tx_pkt_burst = removed_tx_burst;
1659 	eth_dev->dev_ops = &mlx5_dev_ops;
1660 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1661 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1662 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1663 	/* Register MAC address. */
1664 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1665 	if (config->vf && config->vf_nl_en)
1666 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1667 				      mlx5_ifindex(eth_dev),
1668 				      eth_dev->data->mac_addrs,
1669 				      MLX5_MAX_MAC_ADDRESSES);
1670 	priv->ctrl_flows = 0;
1671 	rte_spinlock_init(&priv->flow_list_lock);
1672 	TAILQ_INIT(&priv->flow_meters);
1673 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1674 	if (!priv->mtr_profile_tbl)
1675 		goto error;
1676 	/* Bring Ethernet device up. */
1677 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1678 		eth_dev->data->port_id);
1679 	mlx5_set_link_up(eth_dev);
1680 	/*
1681 	 * Even though the interrupt handler is not installed yet,
1682 	 * interrupts will still trigger on the async_fd from
1683 	 * Verbs context returned by ibv_open_device().
1684 	 */
1685 	mlx5_link_update(eth_dev, 0);
1686 #ifdef HAVE_MLX5DV_DR_ESWITCH
1687 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1688 	      (switch_info->representor || switch_info->master)))
1689 		config->dv_esw_en = 0;
1690 #else
1691 	config->dv_esw_en = 0;
1692 #endif
1693 	/* Detect minimal data bytes to inline. */
1694 	mlx5_set_min_inline(spawn, config);
1695 	/* Store device configuration on private structure. */
1696 	priv->config = *config;
1697 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1698 		icfg[i].release_mem_en = !!config->reclaim_mode;
1699 		if (config->reclaim_mode)
1700 			icfg[i].per_core_cache = 0;
1701 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1702 		if (!priv->flows[i])
1703 			goto error;
1704 	}
1705 	/* Create context for virtual machine VLAN workaround. */
1706 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1707 	if (config->dv_flow_en) {
1708 		err = mlx5_alloc_shared_dr(priv);
1709 		if (err)
1710 			goto error;
1711 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1712 			goto error;
1713 	}
1714 	if (sh->devx && config->dv_flow_en && config->dest_tir) {
1715 		priv->obj_ops = devx_obj_ops;
1716 		mlx5_queue_counter_id_prepare(eth_dev);
1717 		priv->obj_ops.lb_dummy_queue_create =
1718 					mlx5_rxq_ibv_obj_dummy_lb_create;
1719 		priv->obj_ops.lb_dummy_queue_release =
1720 					mlx5_rxq_ibv_obj_dummy_lb_release;
1721 	} else if (spawn->max_port > UINT8_MAX) {
1722 		/* Verbs can't support ports larger than 255 by design. */
1723 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1724 		err = ENOTSUP;
1725 		goto error;
1726 	} else {
1727 		priv->obj_ops = ibv_obj_ops;
1728 	}
1729 	if (config->tx_pp &&
1730 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1731 		/*
1732 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1733 		 * packet pacing and already checked above.
1734 		 * Hence, we should only make sure the SQs will be created
1735 		 * with DevX, not with Verbs.
1736 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1737 		 * with Clock Queue UAR as required for Tx scheduling.
1738 		 */
1739 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1740 		err = ENODEV;
1741 		goto error;
1742 	}
1743 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1744 	if (!priv->drop_queue.hrxq)
1745 		goto error;
1746 	/* Port representor shares the same max prioirity with pf port. */
1747 	if (!priv->sh->flow_priority_check_flag) {
1748 		/* Supported Verbs flow priority number detection. */
1749 		err = mlx5_flow_discover_priorities(eth_dev);
1750 		priv->sh->flow_max_priority = err;
1751 		priv->sh->flow_priority_check_flag = 1;
1752 	} else {
1753 		err = priv->sh->flow_max_priority;
1754 	}
1755 	if (err < 0) {
1756 		err = -err;
1757 		goto error;
1758 	}
1759 	if (!priv->config.dv_esw_en &&
1760 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1761 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1762 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1763 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1764 	}
1765 	mlx5_set_metadata_mask(eth_dev);
1766 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1767 	    !priv->sh->dv_regc0_mask) {
1768 		DRV_LOG(ERR, "metadata mode %u is not supported "
1769 			     "(no metadata reg_c[0] is available)",
1770 			     priv->config.dv_xmeta_en);
1771 			err = ENOTSUP;
1772 			goto error;
1773 	}
1774 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1775 				       mlx5_hrxq_create_cb,
1776 				       mlx5_hrxq_match_cb,
1777 				       mlx5_hrxq_remove_cb,
1778 				       mlx5_hrxq_clone_cb,
1779 				       mlx5_hrxq_clone_free_cb);
1780 	if (!priv->hrxqs)
1781 		goto error;
1782 	rte_rwlock_init(&priv->ind_tbls_lock);
1783 	/* Query availability of metadata reg_c's. */
1784 	if (!priv->sh->metadata_regc_check_flag) {
1785 		err = mlx5_flow_discover_mreg_c(eth_dev);
1786 		if (err < 0) {
1787 			err = -err;
1788 			goto error;
1789 		}
1790 	}
1791 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1792 		DRV_LOG(DEBUG,
1793 			"port %u extensive metadata register is not supported",
1794 			eth_dev->data->port_id);
1795 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1796 			DRV_LOG(ERR, "metadata mode %u is not supported "
1797 				     "(no metadata registers available)",
1798 				     priv->config.dv_xmeta_en);
1799 			err = ENOTSUP;
1800 			goto error;
1801 		}
1802 	}
1803 	if (priv->config.dv_flow_en &&
1804 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1805 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1806 	    priv->sh->dv_regc0_mask) {
1807 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1808 						      MLX5_FLOW_MREG_HTABLE_SZ,
1809 						      false, true, eth_dev,
1810 						      flow_dv_mreg_create_cb,
1811 						      flow_dv_mreg_match_cb,
1812 						      flow_dv_mreg_remove_cb,
1813 						      flow_dv_mreg_clone_cb,
1814 						    flow_dv_mreg_clone_free_cb);
1815 		if (!priv->mreg_cp_tbl) {
1816 			err = ENOMEM;
1817 			goto error;
1818 		}
1819 	}
1820 	rte_spinlock_init(&priv->shared_act_sl);
1821 	mlx5_flow_counter_mode_config(eth_dev);
1822 	mlx5_flow_drop_action_config(eth_dev);
1823 	if (priv->config.dv_flow_en)
1824 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1825 	return eth_dev;
1826 error:
1827 	if (priv) {
1828 		if (priv->mreg_cp_tbl)
1829 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1830 		if (priv->sh)
1831 			mlx5_os_free_shared_dr(priv);
1832 		if (priv->nl_socket_route >= 0)
1833 			close(priv->nl_socket_route);
1834 		if (priv->vmwa_context)
1835 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1836 		if (eth_dev && priv->drop_queue.hrxq)
1837 			mlx5_drop_action_destroy(eth_dev);
1838 		if (priv->mtr_profile_tbl)
1839 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1840 		if (own_domain_id)
1841 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1842 		if (priv->hrxqs)
1843 			mlx5_list_destroy(priv->hrxqs);
1844 		if (eth_dev && priv->flex_item_map)
1845 			mlx5_flex_item_port_cleanup(eth_dev);
1846 		mlx5_free(priv);
1847 		if (eth_dev != NULL)
1848 			eth_dev->data->dev_private = NULL;
1849 	}
1850 	if (eth_dev != NULL) {
1851 		/* mac_addrs must not be freed alone because part of
1852 		 * dev_private
1853 		 **/
1854 		eth_dev->data->mac_addrs = NULL;
1855 		rte_eth_dev_release_port(eth_dev);
1856 	}
1857 	if (sh)
1858 		mlx5_free_shared_dev_ctx(sh);
1859 	if (nl_rdma >= 0)
1860 		close(nl_rdma);
1861 	MLX5_ASSERT(err > 0);
1862 	rte_errno = err;
1863 	return NULL;
1864 }
1865 
1866 /**
1867  * Comparison callback to sort device data.
1868  *
1869  * This is meant to be used with qsort().
1870  *
1871  * @param a[in]
1872  *   Pointer to pointer to first data object.
1873  * @param b[in]
1874  *   Pointer to pointer to second data object.
1875  *
1876  * @return
1877  *   0 if both objects are equal, less than 0 if the first argument is less
1878  *   than the second, greater than 0 otherwise.
1879  */
1880 static int
1881 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1882 {
1883 	const struct mlx5_switch_info *si_a =
1884 		&((const struct mlx5_dev_spawn_data *)a)->info;
1885 	const struct mlx5_switch_info *si_b =
1886 		&((const struct mlx5_dev_spawn_data *)b)->info;
1887 	int ret;
1888 
1889 	/* Master device first. */
1890 	ret = si_b->master - si_a->master;
1891 	if (ret)
1892 		return ret;
1893 	/* Then representor devices. */
1894 	ret = si_b->representor - si_a->representor;
1895 	if (ret)
1896 		return ret;
1897 	/* Unidentified devices come last in no specific order. */
1898 	if (!si_a->representor)
1899 		return 0;
1900 	/* Order representors by name. */
1901 	return si_a->port_name - si_b->port_name;
1902 }
1903 
1904 /**
1905  * Match PCI information for possible slaves of bonding device.
1906  *
1907  * @param[in] ibdev_name
1908  *   Name of Infiniband device.
1909  * @param[in] pci_dev
1910  *   Pointer to primary PCI address structure to match.
1911  * @param[in] nl_rdma
1912  *   Netlink RDMA group socket handle.
1913  * @param[in] owner
1914  *   Representor owner PF index.
1915  * @param[out] bond_info
1916  *   Pointer to bonding information.
1917  *
1918  * @return
1919  *   negative value if no bonding device found, otherwise
1920  *   positive index of slave PF in bonding.
1921  */
1922 static int
1923 mlx5_device_bond_pci_match(const char *ibdev_name,
1924 			   const struct rte_pci_addr *pci_dev,
1925 			   int nl_rdma, uint16_t owner,
1926 			   struct mlx5_bond_info *bond_info)
1927 {
1928 	char ifname[IF_NAMESIZE + 1];
1929 	unsigned int ifindex;
1930 	unsigned int np, i;
1931 	FILE *bond_file = NULL, *file;
1932 	int pf = -1;
1933 	int ret;
1934 	uint8_t cur_guid[32] = {0};
1935 	uint8_t guid[32] = {0};
1936 
1937 	/*
1938 	 * Try to get master device name. If something goes wrong suppose
1939 	 * the lack of kernel support and no bonding devices.
1940 	 */
1941 	memset(bond_info, 0, sizeof(*bond_info));
1942 	if (nl_rdma < 0)
1943 		return -1;
1944 	if (!strstr(ibdev_name, "bond"))
1945 		return -1;
1946 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1947 	if (!np)
1948 		return -1;
1949 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1950 		return -1;
1951 	/*
1952 	 * The master device might not be on the predefined port(not on port
1953 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1954 	 * device ports and find master.
1955 	 */
1956 	for (i = 1; i <= np; ++i) {
1957 		/* Check whether Infiniband port is populated. */
1958 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1959 		if (!ifindex)
1960 			continue;
1961 		if (!if_indextoname(ifindex, ifname))
1962 			continue;
1963 		/* Try to read bonding slave names from sysfs. */
1964 		MKSTR(slaves,
1965 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1966 		bond_file = fopen(slaves, "r");
1967 		if (bond_file)
1968 			break;
1969 	}
1970 	if (!bond_file)
1971 		return -1;
1972 	/* Use safe format to check maximal buffer length. */
1973 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1974 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1975 		char tmp_str[IF_NAMESIZE + 32];
1976 		struct rte_pci_addr pci_addr;
1977 		struct mlx5_switch_info	info;
1978 		int ret;
1979 
1980 		/* Process slave interface names in the loop. */
1981 		snprintf(tmp_str, sizeof(tmp_str),
1982 			 "/sys/class/net/%s", ifname);
1983 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1984 			DRV_LOG(WARNING,
1985 				"Cannot get PCI address for netdev \"%s\".",
1986 				ifname);
1987 			continue;
1988 		}
1989 		/* Slave interface PCI address match found. */
1990 		snprintf(tmp_str, sizeof(tmp_str),
1991 			 "/sys/class/net/%s/phys_port_name", ifname);
1992 		file = fopen(tmp_str, "rb");
1993 		if (!file)
1994 			break;
1995 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1996 		if (fscanf(file, "%32s", tmp_str) == 1)
1997 			mlx5_translate_port_name(tmp_str, &info);
1998 		fclose(file);
1999 		/* Only process PF ports. */
2000 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2001 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2002 			continue;
2003 		/* Check max bonding member. */
2004 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2005 			DRV_LOG(WARNING, "bonding index out of range, "
2006 				"please increase MLX5_BOND_MAX_PORTS: %s",
2007 				tmp_str);
2008 			break;
2009 		}
2010 		/* Get ifindex. */
2011 		snprintf(tmp_str, sizeof(tmp_str),
2012 			 "/sys/class/net/%s/ifindex", ifname);
2013 		file = fopen(tmp_str, "rb");
2014 		if (!file)
2015 			break;
2016 		ret = fscanf(file, "%u", &ifindex);
2017 		fclose(file);
2018 		if (ret != 1)
2019 			break;
2020 		/* Save bonding info. */
2021 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2022 			sizeof(bond_info->ports[0].ifname));
2023 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2024 		bond_info->ports[info.port_name].ifindex = ifindex;
2025 		bond_info->n_port++;
2026 		/*
2027 		 * Under socket direct mode, bonding will use
2028 		 * system_image_guid as identification.
2029 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
2030 		 * All bonding members should have the same guid even if driver
2031 		 * is using PCIe BDF.
2032 		 */
2033 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
2034 		if (ret < 0)
2035 			break;
2036 		else if (ret > 0) {
2037 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
2038 			    owner == info.port_name &&
2039 			    (owner != 0 || (owner == 0 &&
2040 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
2041 				pf = info.port_name;
2042 		} else if (pci_dev->domain == pci_addr.domain &&
2043 		    pci_dev->bus == pci_addr.bus &&
2044 		    pci_dev->devid == pci_addr.devid &&
2045 		    ((pci_dev->function == 0 &&
2046 		      pci_dev->function + owner == pci_addr.function) ||
2047 		     (pci_dev->function == owner &&
2048 		      pci_addr.function == owner)))
2049 			pf = info.port_name;
2050 	}
2051 	if (pf >= 0) {
2052 		/* Get bond interface info */
2053 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2054 					   bond_info->ifname);
2055 		if (ret)
2056 			DRV_LOG(ERR, "unable to get bond info: %s",
2057 				strerror(rte_errno));
2058 		else
2059 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2060 				ifindex, bond_info->ifindex, bond_info->ifname);
2061 	}
2062 	if (owner == 0 && pf != 0) {
2063 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
2064 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
2065 				pci_dev->function);
2066 	}
2067 	return pf;
2068 }
2069 
2070 static void
2071 mlx5_os_config_default(struct mlx5_dev_config *config)
2072 {
2073 	memset(config, 0, sizeof(*config));
2074 	config->mps = MLX5_ARG_UNSET;
2075 	config->rx_vec_en = 1;
2076 	config->txq_inline_max = MLX5_ARG_UNSET;
2077 	config->txq_inline_min = MLX5_ARG_UNSET;
2078 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2079 	config->txqs_inline = MLX5_ARG_UNSET;
2080 	config->vf_nl_en = 1;
2081 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2082 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2083 	config->dv_esw_en = 1;
2084 	config->dv_flow_en = 1;
2085 	config->decap_en = 1;
2086 	config->log_hp_size = MLX5_ARG_UNSET;
2087 	config->allow_duplicate_pattern = 1;
2088 	config->std_delay_drop = 0;
2089 	config->hp_delay_drop = 0;
2090 }
2091 
2092 /**
2093  * Register a PCI device within bonding.
2094  *
2095  * This function spawns Ethernet devices out of a given PCI device and
2096  * bonding owner PF index.
2097  *
2098  * @param[in] cdev
2099  *   Pointer to common mlx5 device structure.
2100  * @param[in] req_eth_da
2101  *   Requested ethdev device argument.
2102  * @param[in] owner_id
2103  *   Requested owner PF port ID within bonding device, default to 0.
2104  *
2105  * @return
2106  *   0 on success, a negative errno value otherwise and rte_errno is set.
2107  */
2108 static int
2109 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
2110 		     struct rte_eth_devargs *req_eth_da,
2111 		     uint16_t owner_id)
2112 {
2113 	struct ibv_device **ibv_list;
2114 	/*
2115 	 * Number of found IB Devices matching with requested PCI BDF.
2116 	 * nd != 1 means there are multiple IB devices over the same
2117 	 * PCI device and we have representors and master.
2118 	 */
2119 	unsigned int nd = 0;
2120 	/*
2121 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2122 	 * we have the single multiport IB device, and there may be
2123 	 * representors attached to some of found ports.
2124 	 */
2125 	unsigned int np = 0;
2126 	/*
2127 	 * Number of DPDK ethernet devices to Spawn - either over
2128 	 * multiple IB devices or multiple ports of single IB device.
2129 	 * Actually this is the number of iterations to spawn.
2130 	 */
2131 	unsigned int ns = 0;
2132 	/*
2133 	 * Bonding device
2134 	 *   < 0 - no bonding device (single one)
2135 	 *  >= 0 - bonding device (value is slave PF index)
2136 	 */
2137 	int bd = -1;
2138 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2139 	struct mlx5_dev_spawn_data *list = NULL;
2140 	struct mlx5_dev_config dev_config;
2141 	unsigned int dev_config_vf;
2142 	struct rte_eth_devargs eth_da = *req_eth_da;
2143 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2144 	struct mlx5_bond_info bond_info;
2145 	int ret = -1;
2146 
2147 	errno = 0;
2148 	ibv_list = mlx5_glue->get_device_list(&ret);
2149 	if (!ibv_list) {
2150 		rte_errno = errno ? errno : ENOSYS;
2151 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2152 		return -rte_errno;
2153 	}
2154 	/*
2155 	 * First scan the list of all Infiniband devices to find
2156 	 * matching ones, gathering into the list.
2157 	 */
2158 	struct ibv_device *ibv_match[ret + 1];
2159 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2160 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2161 	unsigned int i;
2162 
2163 	while (ret-- > 0) {
2164 		struct rte_pci_addr pci_addr;
2165 
2166 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2167 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2168 						nl_rdma, owner_id, &bond_info);
2169 		if (bd >= 0) {
2170 			/*
2171 			 * Bonding device detected. Only one match is allowed,
2172 			 * the bonding is supported over multi-port IB device,
2173 			 * there should be no matches on representor PCI
2174 			 * functions or non VF LAG bonding devices with
2175 			 * specified address.
2176 			 */
2177 			if (nd) {
2178 				DRV_LOG(ERR,
2179 					"multiple PCI match on bonding device"
2180 					"\"%s\" found", ibv_list[ret]->name);
2181 				rte_errno = ENOENT;
2182 				ret = -rte_errno;
2183 				goto exit;
2184 			}
2185 			/* Amend owner pci address if owner PF ID specified. */
2186 			if (eth_da.nb_representor_ports)
2187 				owner_pci.function += owner_id;
2188 			DRV_LOG(INFO,
2189 				"PCI information matches for slave %d bonding device \"%s\"",
2190 				bd, ibv_list[ret]->name);
2191 			ibv_match[nd++] = ibv_list[ret];
2192 			break;
2193 		} else {
2194 			/* Bonding device not found. */
2195 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2196 					      &pci_addr))
2197 				continue;
2198 			if (owner_pci.domain != pci_addr.domain ||
2199 			    owner_pci.bus != pci_addr.bus ||
2200 			    owner_pci.devid != pci_addr.devid ||
2201 			    owner_pci.function != pci_addr.function)
2202 				continue;
2203 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2204 				ibv_list[ret]->name);
2205 			ibv_match[nd++] = ibv_list[ret];
2206 		}
2207 	}
2208 	ibv_match[nd] = NULL;
2209 	if (!nd) {
2210 		/* No device matches, just complain and bail out. */
2211 		DRV_LOG(WARNING,
2212 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
2213 			" are kernel drivers loaded?",
2214 			owner_pci.domain, owner_pci.bus,
2215 			owner_pci.devid, owner_pci.function);
2216 		rte_errno = ENOENT;
2217 		ret = -rte_errno;
2218 		goto exit;
2219 	}
2220 	if (nd == 1) {
2221 		/*
2222 		 * Found single matching device may have multiple ports.
2223 		 * Each port may be representor, we have to check the port
2224 		 * number and check the representors existence.
2225 		 */
2226 		if (nl_rdma >= 0)
2227 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2228 		if (!np)
2229 			DRV_LOG(WARNING,
2230 				"Cannot get IB device \"%s\" ports number.",
2231 				ibv_match[0]->name);
2232 		if (bd >= 0 && !np) {
2233 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2234 			rte_errno = ENOENT;
2235 			ret = -rte_errno;
2236 			goto exit;
2237 		}
2238 	}
2239 	/* Now we can determine the maximal amount of devices to be spawned. */
2240 	list = mlx5_malloc(MLX5_MEM_ZERO,
2241 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2242 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2243 	if (!list) {
2244 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2245 		rte_errno = ENOMEM;
2246 		ret = -rte_errno;
2247 		goto exit;
2248 	}
2249 	if (bd >= 0 || np > 1) {
2250 		/*
2251 		 * Single IB device with multiple ports found,
2252 		 * it may be E-Switch master device and representors.
2253 		 * We have to perform identification through the ports.
2254 		 */
2255 		MLX5_ASSERT(nl_rdma >= 0);
2256 		MLX5_ASSERT(ns == 0);
2257 		MLX5_ASSERT(nd == 1);
2258 		MLX5_ASSERT(np);
2259 		for (i = 1; i <= np; ++i) {
2260 			list[ns].bond_info = &bond_info;
2261 			list[ns].max_port = np;
2262 			list[ns].phys_port = i;
2263 			list[ns].phys_dev_name = ibv_match[0]->name;
2264 			list[ns].eth_dev = NULL;
2265 			list[ns].pci_dev = pci_dev;
2266 			list[ns].cdev = cdev;
2267 			list[ns].pf_bond = bd;
2268 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2269 							   ibv_match[0]->name,
2270 							   i);
2271 			if (!list[ns].ifindex) {
2272 				/*
2273 				 * No network interface index found for the
2274 				 * specified port, it means there is no
2275 				 * representor on this port. It's OK,
2276 				 * there can be disabled ports, for example
2277 				 * if sriov_numvfs < sriov_totalvfs.
2278 				 */
2279 				continue;
2280 			}
2281 			ret = -1;
2282 			if (nl_route >= 0)
2283 				ret = mlx5_nl_switch_info(nl_route,
2284 							  list[ns].ifindex,
2285 							  &list[ns].info);
2286 			if (ret || (!list[ns].info.representor &&
2287 				    !list[ns].info.master)) {
2288 				/*
2289 				 * We failed to recognize representors with
2290 				 * Netlink, let's try to perform the task
2291 				 * with sysfs.
2292 				 */
2293 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2294 							     &list[ns].info);
2295 			}
2296 			if (!ret && bd >= 0) {
2297 				switch (list[ns].info.name_type) {
2298 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2299 					if (np == 1) {
2300 						/*
2301 						 * Force standalone bonding
2302 						 * device for ROCE LAG
2303 						 * confgiurations.
2304 						 */
2305 						list[ns].info.master = 0;
2306 						list[ns].info.representor = 0;
2307 					}
2308 					if (list[ns].info.port_name == bd)
2309 						ns++;
2310 					break;
2311 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2312 					/* Fallthrough */
2313 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2314 					/* Fallthrough */
2315 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2316 					if (list[ns].info.pf_num == bd)
2317 						ns++;
2318 					break;
2319 				default:
2320 					break;
2321 				}
2322 				continue;
2323 			}
2324 			if (!ret && (list[ns].info.representor ^
2325 				     list[ns].info.master))
2326 				ns++;
2327 		}
2328 		if (!ns) {
2329 			DRV_LOG(ERR,
2330 				"Unable to recognize master/representors on the IB device with multiple ports.");
2331 			rte_errno = ENOENT;
2332 			ret = -rte_errno;
2333 			goto exit;
2334 		}
2335 	} else {
2336 		/*
2337 		 * The existence of several matching entries (nd > 1) means
2338 		 * port representors have been instantiated. No existing Verbs
2339 		 * call nor sysfs entries can tell them apart, this can only
2340 		 * be done through Netlink calls assuming kernel drivers are
2341 		 * recent enough to support them.
2342 		 *
2343 		 * In the event of identification failure through Netlink,
2344 		 * try again through sysfs, then:
2345 		 *
2346 		 * 1. A single IB device matches (nd == 1) with single
2347 		 *    port (np=0/1) and is not a representor, assume
2348 		 *    no switch support.
2349 		 *
2350 		 * 2. Otherwise no safe assumptions can be made;
2351 		 *    complain louder and bail out.
2352 		 */
2353 		for (i = 0; i != nd; ++i) {
2354 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2355 			list[ns].bond_info = NULL;
2356 			list[ns].max_port = 1;
2357 			list[ns].phys_port = 1;
2358 			list[ns].phys_dev_name = ibv_match[i]->name;
2359 			list[ns].eth_dev = NULL;
2360 			list[ns].pci_dev = pci_dev;
2361 			list[ns].cdev = cdev;
2362 			list[ns].pf_bond = -1;
2363 			list[ns].ifindex = 0;
2364 			if (nl_rdma >= 0)
2365 				list[ns].ifindex = mlx5_nl_ifindex
2366 							    (nl_rdma,
2367 							     ibv_match[i]->name,
2368 							     1);
2369 			if (!list[ns].ifindex) {
2370 				char ifname[IF_NAMESIZE];
2371 
2372 				/*
2373 				 * Netlink failed, it may happen with old
2374 				 * ib_core kernel driver (before 4.16).
2375 				 * We can assume there is old driver because
2376 				 * here we are processing single ports IB
2377 				 * devices. Let's try sysfs to retrieve
2378 				 * the ifindex. The method works for
2379 				 * master device only.
2380 				 */
2381 				if (nd > 1) {
2382 					/*
2383 					 * Multiple devices found, assume
2384 					 * representors, can not distinguish
2385 					 * master/representor and retrieve
2386 					 * ifindex via sysfs.
2387 					 */
2388 					continue;
2389 				}
2390 				ret = mlx5_get_ifname_sysfs
2391 					(ibv_match[i]->ibdev_path, ifname);
2392 				if (!ret)
2393 					list[ns].ifindex =
2394 						if_nametoindex(ifname);
2395 				if (!list[ns].ifindex) {
2396 					/*
2397 					 * No network interface index found
2398 					 * for the specified device, it means
2399 					 * there it is neither representor
2400 					 * nor master.
2401 					 */
2402 					continue;
2403 				}
2404 			}
2405 			ret = -1;
2406 			if (nl_route >= 0)
2407 				ret = mlx5_nl_switch_info(nl_route,
2408 							  list[ns].ifindex,
2409 							  &list[ns].info);
2410 			if (ret || (!list[ns].info.representor &&
2411 				    !list[ns].info.master)) {
2412 				/*
2413 				 * We failed to recognize representors with
2414 				 * Netlink, let's try to perform the task
2415 				 * with sysfs.
2416 				 */
2417 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2418 							     &list[ns].info);
2419 			}
2420 			if (!ret && (list[ns].info.representor ^
2421 				     list[ns].info.master)) {
2422 				ns++;
2423 			} else if ((nd == 1) &&
2424 				   !list[ns].info.representor &&
2425 				   !list[ns].info.master) {
2426 				/*
2427 				 * Single IB device with one physical port and
2428 				 * attached network device.
2429 				 * May be SRIOV is not enabled or there is no
2430 				 * representors.
2431 				 */
2432 				DRV_LOG(INFO, "No E-Switch support detected.");
2433 				ns++;
2434 				break;
2435 			}
2436 		}
2437 		if (!ns) {
2438 			DRV_LOG(ERR,
2439 				"Unable to recognize master/representors on the multiple IB devices.");
2440 			rte_errno = ENOENT;
2441 			ret = -rte_errno;
2442 			goto exit;
2443 		}
2444 		/*
2445 		 * New kernels may add the switch_id attribute for the case
2446 		 * there is no E-Switch and we wrongly recognized the only
2447 		 * device as master. Override this if there is the single
2448 		 * device with single port and new device name format present.
2449 		 */
2450 		if (nd == 1 &&
2451 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2452 			list[0].info.master = 0;
2453 			list[0].info.representor = 0;
2454 		}
2455 	}
2456 	MLX5_ASSERT(ns);
2457 	/*
2458 	 * Sort list to probe devices in natural order for users convenience
2459 	 * (i.e. master first, then representors from lowest to highest ID).
2460 	 */
2461 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2462 	/* Device specific configuration. */
2463 	switch (pci_dev->id.device_id) {
2464 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2465 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2466 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2467 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2468 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2469 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2470 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2471 		dev_config_vf = 1;
2472 		break;
2473 	default:
2474 		dev_config_vf = 0;
2475 		break;
2476 	}
2477 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2478 		/* Set devargs default values. */
2479 		if (eth_da.nb_mh_controllers == 0) {
2480 			eth_da.nb_mh_controllers = 1;
2481 			eth_da.mh_controllers[0] = 0;
2482 		}
2483 		if (eth_da.nb_ports == 0 && ns > 0) {
2484 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2485 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2486 					pci_dev->device.devargs->args);
2487 			eth_da.nb_ports = 1;
2488 			eth_da.ports[0] = list[0].info.pf_num;
2489 		}
2490 		if (eth_da.nb_representor_ports == 0) {
2491 			eth_da.nb_representor_ports = 1;
2492 			eth_da.representor_ports[0] = 0;
2493 		}
2494 	}
2495 	for (i = 0; i != ns; ++i) {
2496 		uint32_t restore;
2497 
2498 		/* Default configuration. */
2499 		mlx5_os_config_default(&dev_config);
2500 		dev_config.vf = dev_config_vf;
2501 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i],
2502 						 &dev_config, &eth_da);
2503 		if (!list[i].eth_dev) {
2504 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2505 				break;
2506 			/* Device is disabled or already spawned. Ignore it. */
2507 			continue;
2508 		}
2509 		restore = list[i].eth_dev->data->dev_flags;
2510 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2511 		/**
2512 		 * Each representor has a dedicated interrupts vector.
2513 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2514 		 * representor eth_dev object because representor and PF
2515 		 * share the same PCI address.
2516 		 * Override representor device with a dedicated
2517 		 * interrupts handle here.
2518 		 * Representor interrupts handle is released in mlx5_dev_stop().
2519 		 */
2520 		if (list[i].info.representor) {
2521 			struct rte_intr_handle *intr_handle =
2522 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2523 			if (intr_handle == NULL) {
2524 				DRV_LOG(ERR,
2525 					"port %u failed to allocate memory for interrupt handler "
2526 					"Rx interrupts will not be supported",
2527 					i);
2528 				rte_errno = ENOMEM;
2529 				ret = -rte_errno;
2530 				goto exit;
2531 			}
2532 			list[i].eth_dev->intr_handle = intr_handle;
2533 		}
2534 		/* Restore non-PCI flags cleared by the above call. */
2535 		list[i].eth_dev->data->dev_flags |= restore;
2536 		rte_eth_dev_probing_finish(list[i].eth_dev);
2537 	}
2538 	if (i != ns) {
2539 		DRV_LOG(ERR,
2540 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2541 			" encountering an error: %s",
2542 			owner_pci.domain, owner_pci.bus,
2543 			owner_pci.devid, owner_pci.function,
2544 			strerror(rte_errno));
2545 		ret = -rte_errno;
2546 		/* Roll back. */
2547 		while (i--) {
2548 			if (!list[i].eth_dev)
2549 				continue;
2550 			mlx5_dev_close(list[i].eth_dev);
2551 			/* mac_addrs must not be freed because in dev_private */
2552 			list[i].eth_dev->data->mac_addrs = NULL;
2553 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2554 		}
2555 		/* Restore original error. */
2556 		rte_errno = -ret;
2557 	} else {
2558 		ret = 0;
2559 	}
2560 exit:
2561 	/*
2562 	 * Do the routine cleanup:
2563 	 * - close opened Netlink sockets
2564 	 * - free allocated spawn data array
2565 	 * - free the Infiniband device list
2566 	 */
2567 	if (nl_rdma >= 0)
2568 		close(nl_rdma);
2569 	if (nl_route >= 0)
2570 		close(nl_route);
2571 	if (list)
2572 		mlx5_free(list);
2573 	MLX5_ASSERT(ibv_list);
2574 	mlx5_glue->free_device_list(ibv_list);
2575 	return ret;
2576 }
2577 
2578 static int
2579 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2580 			  struct rte_eth_devargs *eth_da)
2581 {
2582 	int ret = 0;
2583 
2584 	if (dev->devargs == NULL)
2585 		return 0;
2586 	memset(eth_da, 0, sizeof(*eth_da));
2587 	/* Parse representor information first from class argument. */
2588 	if (dev->devargs->cls_str)
2589 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2590 	if (ret != 0) {
2591 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2592 			dev->devargs->cls_str);
2593 		return -rte_errno;
2594 	}
2595 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2596 		/* Parse legacy device argument */
2597 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2598 		if (ret) {
2599 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2600 				dev->devargs->args);
2601 			return -rte_errno;
2602 		}
2603 	}
2604 	return 0;
2605 }
2606 
2607 /**
2608  * Callback to register a PCI device.
2609  *
2610  * This function spawns Ethernet devices out of a given PCI device.
2611  *
2612  * @param[in] cdev
2613  *   Pointer to common mlx5 device structure.
2614  *
2615  * @return
2616  *   0 on success, a negative errno value otherwise and rte_errno is set.
2617  */
2618 static int
2619 mlx5_os_pci_probe(struct mlx5_common_device *cdev)
2620 {
2621 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2622 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2623 	int ret = 0;
2624 	uint16_t p;
2625 
2626 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2627 	if (ret != 0)
2628 		return ret;
2629 
2630 	if (eth_da.nb_ports > 0) {
2631 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2632 		for (p = 0; p < eth_da.nb_ports; p++) {
2633 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2634 						   eth_da.ports[p]);
2635 			if (ret)
2636 				break;
2637 		}
2638 		if (ret) {
2639 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2640 				"aborted due to proding failure of PF %u",
2641 				pci_dev->addr.domain, pci_dev->addr.bus,
2642 				pci_dev->addr.devid, pci_dev->addr.function,
2643 				eth_da.ports[p]);
2644 			mlx5_net_remove(cdev);
2645 		}
2646 	} else {
2647 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0);
2648 	}
2649 	return ret;
2650 }
2651 
2652 /* Probe a single SF device on auxiliary bus, no representor support. */
2653 static int
2654 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev)
2655 {
2656 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2657 	struct mlx5_dev_config config;
2658 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2659 	struct rte_device *dev = cdev->dev;
2660 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2661 	struct rte_eth_dev *eth_dev;
2662 	int ret = 0;
2663 
2664 	/* Parse ethdev devargs. */
2665 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2666 	if (ret != 0)
2667 		return ret;
2668 	/* Set default config data. */
2669 	mlx5_os_config_default(&config);
2670 	config.sf = 1;
2671 	/* Init spawn data. */
2672 	spawn.max_port = 1;
2673 	spawn.phys_port = 1;
2674 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2675 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2676 	if (ret < 0) {
2677 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2678 		return ret;
2679 	}
2680 	spawn.ifindex = ret;
2681 	spawn.cdev = cdev;
2682 	/* Spawn device. */
2683 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2684 	if (eth_dev == NULL)
2685 		return -rte_errno;
2686 	/* Post create. */
2687 	eth_dev->intr_handle = adev->intr_handle;
2688 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2689 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2690 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2691 		eth_dev->data->numa_node = dev->numa_node;
2692 	}
2693 	rte_eth_dev_probing_finish(eth_dev);
2694 	return 0;
2695 }
2696 
2697 /**
2698  * Net class driver callback to probe a device.
2699  *
2700  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2701  *
2702  * @param[in] cdev
2703  *   Pointer to the common mlx5 device.
2704  *
2705  * @return
2706  *   0 on success, a negative errno value otherwise and rte_errno is set.
2707  */
2708 int
2709 mlx5_os_net_probe(struct mlx5_common_device *cdev)
2710 {
2711 	int ret;
2712 
2713 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2714 		mlx5_pmd_socket_init();
2715 	ret = mlx5_init_once();
2716 	if (ret) {
2717 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2718 			strerror(rte_errno));
2719 		return -rte_errno;
2720 	}
2721 	if (mlx5_dev_is_pci(cdev->dev))
2722 		return mlx5_os_pci_probe(cdev);
2723 	else
2724 		return mlx5_os_auxiliary_probe(cdev);
2725 }
2726 
2727 /**
2728  * Cleanup resources when the last device is closed.
2729  */
2730 void
2731 mlx5_os_net_cleanup(void)
2732 {
2733 	mlx5_pmd_socket_uninit();
2734 }
2735 
2736 /**
2737  * Install shared asynchronous device events handler.
2738  * This function is implemented to support event sharing
2739  * between multiple ports of single IB device.
2740  *
2741  * @param sh
2742  *   Pointer to mlx5_dev_ctx_shared object.
2743  */
2744 void
2745 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2746 {
2747 	int ret;
2748 	int flags;
2749 	struct ibv_context *ctx = sh->cdev->ctx;
2750 
2751 	sh->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2752 	if (sh->intr_handle == NULL) {
2753 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2754 		rte_errno = ENOMEM;
2755 		return;
2756 	}
2757 	rte_intr_fd_set(sh->intr_handle, -1);
2758 
2759 	flags = fcntl(ctx->async_fd, F_GETFL);
2760 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2761 	if (ret) {
2762 		DRV_LOG(INFO, "failed to change file descriptor async event"
2763 			" queue");
2764 	} else {
2765 		rte_intr_fd_set(sh->intr_handle, ctx->async_fd);
2766 		rte_intr_type_set(sh->intr_handle, RTE_INTR_HANDLE_EXT);
2767 		if (rte_intr_callback_register(sh->intr_handle,
2768 					mlx5_dev_interrupt_handler, sh)) {
2769 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2770 			rte_intr_fd_set(sh->intr_handle, -1);
2771 		}
2772 	}
2773 	if (sh->devx) {
2774 #ifdef HAVE_IBV_DEVX_ASYNC
2775 		sh->intr_handle_devx =
2776 			rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2777 		if (!sh->intr_handle_devx) {
2778 			DRV_LOG(ERR, "Fail to allocate intr_handle");
2779 			rte_errno = ENOMEM;
2780 			return;
2781 		}
2782 		rte_intr_fd_set(sh->intr_handle_devx, -1);
2783 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2784 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2785 		if (!devx_comp) {
2786 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2787 			return;
2788 		}
2789 		flags = fcntl(devx_comp->fd, F_GETFL);
2790 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2791 		if (ret) {
2792 			DRV_LOG(INFO, "failed to change file descriptor"
2793 				" devx comp");
2794 			return;
2795 		}
2796 		rte_intr_fd_set(sh->intr_handle_devx, devx_comp->fd);
2797 		rte_intr_type_set(sh->intr_handle_devx,
2798 					 RTE_INTR_HANDLE_EXT);
2799 		if (rte_intr_callback_register(sh->intr_handle_devx,
2800 					mlx5_dev_interrupt_handler_devx, sh)) {
2801 			DRV_LOG(INFO, "Fail to install the devx shared"
2802 				" interrupt.");
2803 			rte_intr_fd_set(sh->intr_handle_devx, -1);
2804 		}
2805 #endif /* HAVE_IBV_DEVX_ASYNC */
2806 	}
2807 }
2808 
2809 /**
2810  * Uninstall shared asynchronous device events handler.
2811  * This function is implemented to support event sharing
2812  * between multiple ports of single IB device.
2813  *
2814  * @param dev
2815  *   Pointer to mlx5_dev_ctx_shared object.
2816  */
2817 void
2818 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2819 {
2820 	if (rte_intr_fd_get(sh->intr_handle) >= 0)
2821 		mlx5_intr_callback_unregister(sh->intr_handle,
2822 					      mlx5_dev_interrupt_handler, sh);
2823 	rte_intr_instance_free(sh->intr_handle);
2824 #ifdef HAVE_IBV_DEVX_ASYNC
2825 	if (rte_intr_fd_get(sh->intr_handle_devx) >= 0)
2826 		rte_intr_callback_unregister(sh->intr_handle_devx,
2827 				  mlx5_dev_interrupt_handler_devx, sh);
2828 	rte_intr_instance_free(sh->intr_handle_devx);
2829 	if (sh->devx_comp)
2830 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2831 #endif
2832 }
2833 
2834 /**
2835  * Read statistics by a named counter.
2836  *
2837  * @param[in] priv
2838  *   Pointer to the private device data structure.
2839  * @param[in] ctr_name
2840  *   Pointer to the name of the statistic counter to read
2841  * @param[out] stat
2842  *   Pointer to read statistic value.
2843  * @return
2844  *   0 on success and stat is valud, 1 if failed to read the value
2845  *   rte_errno is set.
2846  *
2847  */
2848 int
2849 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2850 		      uint64_t *stat)
2851 {
2852 	int fd;
2853 
2854 	if (priv->sh) {
2855 		if (priv->q_counters != NULL &&
2856 		    strcmp(ctr_name, "out_of_buffer") == 0)
2857 			return mlx5_devx_cmd_queue_counter_query
2858 					(priv->q_counters, 0, (uint32_t *)stat);
2859 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2860 		      priv->sh->ibdev_path,
2861 		      priv->dev_port,
2862 		      ctr_name);
2863 		fd = open(path, O_RDONLY);
2864 		/*
2865 		 * in switchdev the file location is not per port
2866 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2867 		 */
2868 		if (fd == -1) {
2869 			MKSTR(path1, "%s/hw_counters/%s",
2870 			      priv->sh->ibdev_path,
2871 			      ctr_name);
2872 			fd = open(path1, O_RDONLY);
2873 		}
2874 		if (fd != -1) {
2875 			char buf[21] = {'\0'};
2876 			ssize_t n = read(fd, buf, sizeof(buf));
2877 
2878 			close(fd);
2879 			if (n != -1) {
2880 				*stat = strtoull(buf, NULL, 10);
2881 				return 0;
2882 			}
2883 		}
2884 	}
2885 	*stat = 0;
2886 	return 1;
2887 }
2888 
2889 /**
2890  * Remove a MAC address from device
2891  *
2892  * @param dev
2893  *   Pointer to Ethernet device structure.
2894  * @param index
2895  *   MAC address index.
2896  */
2897 void
2898 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2899 {
2900 	struct mlx5_priv *priv = dev->data->dev_private;
2901 	const int vf = priv->config.vf;
2902 
2903 	if (vf)
2904 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2905 					mlx5_ifindex(dev), priv->mac_own,
2906 					&dev->data->mac_addrs[index], index);
2907 }
2908 
2909 /**
2910  * Adds a MAC address to the device
2911  *
2912  * @param dev
2913  *   Pointer to Ethernet device structure.
2914  * @param mac_addr
2915  *   MAC address to register.
2916  * @param index
2917  *   MAC address index.
2918  *
2919  * @return
2920  *   0 on success, a negative errno value otherwise
2921  */
2922 int
2923 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2924 		     uint32_t index)
2925 {
2926 	struct mlx5_priv *priv = dev->data->dev_private;
2927 	const int vf = priv->config.vf;
2928 	int ret = 0;
2929 
2930 	if (vf)
2931 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2932 					   mlx5_ifindex(dev), priv->mac_own,
2933 					   mac, index);
2934 	return ret;
2935 }
2936 
2937 /**
2938  * Modify a VF MAC address
2939  *
2940  * @param priv
2941  *   Pointer to device private data.
2942  * @param mac_addr
2943  *   MAC address to modify into.
2944  * @param iface_idx
2945  *   Net device interface index
2946  * @param vf_index
2947  *   VF index
2948  *
2949  * @return
2950  *   0 on success, a negative errno value otherwise
2951  */
2952 int
2953 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2954 			   unsigned int iface_idx,
2955 			   struct rte_ether_addr *mac_addr,
2956 			   int vf_index)
2957 {
2958 	return mlx5_nl_vf_mac_addr_modify
2959 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2960 }
2961 
2962 /**
2963  * Set device promiscuous mode
2964  *
2965  * @param dev
2966  *   Pointer to Ethernet device structure.
2967  * @param enable
2968  *   0 - promiscuous is disabled, otherwise - enabled
2969  *
2970  * @return
2971  *   0 on success, a negative error value otherwise
2972  */
2973 int
2974 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2975 {
2976 	struct mlx5_priv *priv = dev->data->dev_private;
2977 
2978 	return mlx5_nl_promisc(priv->nl_socket_route,
2979 			       mlx5_ifindex(dev), !!enable);
2980 }
2981 
2982 /**
2983  * Set device promiscuous mode
2984  *
2985  * @param dev
2986  *   Pointer to Ethernet device structure.
2987  * @param enable
2988  *   0 - all multicase is disabled, otherwise - enabled
2989  *
2990  * @return
2991  *   0 on success, a negative error value otherwise
2992  */
2993 int
2994 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2995 {
2996 	struct mlx5_priv *priv = dev->data->dev_private;
2997 
2998 	return mlx5_nl_allmulti(priv->nl_socket_route,
2999 				mlx5_ifindex(dev), !!enable);
3000 }
3001 
3002 /**
3003  * Flush device MAC addresses
3004  *
3005  * @param dev
3006  *   Pointer to Ethernet device structure.
3007  *
3008  */
3009 void
3010 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3011 {
3012 	struct mlx5_priv *priv = dev->data->dev_private;
3013 
3014 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3015 			       dev->data->mac_addrs,
3016 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3017 }
3018