xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision cf004fd33acca27a506b65904564eaf1cde4eea7)
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 interrupt) 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, a negative errno value otherwise and rte_errno is set.
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 		rte_errno = errno;
155 		return -rte_errno;
156 	}
157 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
158 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
159 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
160 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
161 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
162 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
163 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
164 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
165 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
166 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
167 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
168 	device_attr->max_rwq_indirection_table_size =
169 		attr_ex.rss_caps.max_rwq_indirection_table_size;
170 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
171 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
172 
173 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
174 #ifdef HAVE_IBV_MLX5_MOD_SWP
175 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
176 #endif
177 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
178 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
179 #endif
180 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
181 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
182 #endif
183 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
184 	if (err) {
185 		rte_errno = errno;
186 		return -rte_errno;
187 	}
188 
189 	device_attr->flags = dv_attr.flags;
190 	device_attr->comp_mask = dv_attr.comp_mask;
191 #ifdef HAVE_IBV_MLX5_MOD_SWP
192 	device_attr->sw_parsing_offloads =
193 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
194 #endif
195 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
196 	device_attr->min_single_stride_log_num_of_bytes =
197 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
198 	device_attr->max_single_stride_log_num_of_bytes =
199 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
200 	device_attr->min_single_wqe_log_num_of_strides =
201 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
202 	device_attr->max_single_wqe_log_num_of_strides =
203 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
204 	device_attr->stride_supported_qpts =
205 		dv_attr.striding_rq_caps.supported_qpts;
206 #endif
207 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
208 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
209 #endif
210 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
211 		sizeof(device_attr->fw_ver));
212 
213 	return 0;
214 }
215 
216 /**
217  * Detect misc5 support or not
218  *
219  * @param[in] priv
220  *   Device private data pointer
221  */
222 #ifdef HAVE_MLX5DV_DR
223 static void
224 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
225 {
226 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
227 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
228 	 * Case: IPv4--->UDP--->VxLAN--->vni
229 	 */
230 	void *tbl;
231 	struct mlx5_flow_dv_match_params matcher_mask;
232 	void *match_m;
233 	void *matcher;
234 	void *headers_m;
235 	void *misc5_m;
236 	uint32_t *tunnel_header_m;
237 	struct mlx5dv_flow_matcher_attr dv_attr;
238 
239 	memset(&matcher_mask, 0, sizeof(matcher_mask));
240 	matcher_mask.size = sizeof(matcher_mask.buf);
241 	match_m = matcher_mask.buf;
242 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
243 	misc5_m = MLX5_ADDR_OF(fte_match_param,
244 			       match_m, misc_parameters_5);
245 	tunnel_header_m = (uint32_t *)
246 				MLX5_ADDR_OF(fte_match_set_misc5,
247 				misc5_m, tunnel_header_1);
248 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
249 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
250 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
251 	*tunnel_header_m = 0xffffff;
252 
253 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
254 	if (!tbl) {
255 		DRV_LOG(INFO, "No SW steering support");
256 		return;
257 	}
258 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
259 	dv_attr.match_mask = (void *)&matcher_mask,
260 	dv_attr.match_criteria_enable =
261 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
262 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
263 	dv_attr.priority = 3;
264 #ifdef HAVE_MLX5DV_DR_ESWITCH
265 	void *misc2_m;
266 	if (priv->config.dv_esw_en) {
267 		/* FDB enabled reg_c_0 */
268 		dv_attr.match_criteria_enable |=
269 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
270 		misc2_m = MLX5_ADDR_OF(fte_match_param,
271 				       match_m, misc_parameters_2);
272 		MLX5_SET(fte_match_set_misc2, misc2_m,
273 			 metadata_reg_c_0, 0xffff);
274 	}
275 #endif
276 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
277 						    &dv_attr, tbl);
278 	if (matcher) {
279 		priv->sh->misc5_cap = 1;
280 		mlx5_glue->dv_destroy_flow_matcher(matcher);
281 	}
282 	mlx5_glue->dr_destroy_flow_tbl(tbl);
283 #else
284 	RTE_SET_USED(priv);
285 #endif
286 }
287 #endif
288 
289 /**
290  * Initialize DR related data within private structure.
291  * Routine checks the reference counter and does actual
292  * resources creation/initialization only if counter is zero.
293  *
294  * @param[in] priv
295  *   Pointer to the private device data structure.
296  *
297  * @return
298  *   Zero on success, positive error code otherwise.
299  */
300 static int
301 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
302 {
303 	struct mlx5_dev_ctx_shared *sh = priv->sh;
304 	char s[MLX5_NAME_SIZE] __rte_unused;
305 	int err;
306 
307 	MLX5_ASSERT(sh && sh->refcnt);
308 	if (sh->refcnt > 1)
309 		return 0;
310 	err = mlx5_alloc_table_hash_list(priv);
311 	if (err)
312 		goto error;
313 	/* The resources below are only valid with DV support. */
314 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
315 	/* Init port id action list. */
316 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
317 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
318 						   flow_dv_port_id_create_cb,
319 						   flow_dv_port_id_match_cb,
320 						   flow_dv_port_id_remove_cb,
321 						   flow_dv_port_id_clone_cb,
322 						 flow_dv_port_id_clone_free_cb);
323 	if (!sh->port_id_action_list)
324 		goto error;
325 	/* Init push vlan action list. */
326 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
327 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
328 						    flow_dv_push_vlan_create_cb,
329 						    flow_dv_push_vlan_match_cb,
330 						    flow_dv_push_vlan_remove_cb,
331 						    flow_dv_push_vlan_clone_cb,
332 					       flow_dv_push_vlan_clone_free_cb);
333 	if (!sh->push_vlan_action_list)
334 		goto error;
335 	/* Init sample action list. */
336 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
337 	sh->sample_action_list = mlx5_list_create(s, sh, true,
338 						  flow_dv_sample_create_cb,
339 						  flow_dv_sample_match_cb,
340 						  flow_dv_sample_remove_cb,
341 						  flow_dv_sample_clone_cb,
342 						  flow_dv_sample_clone_free_cb);
343 	if (!sh->sample_action_list)
344 		goto error;
345 	/* Init dest array action list. */
346 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
347 	sh->dest_array_list = mlx5_list_create(s, sh, true,
348 					       flow_dv_dest_array_create_cb,
349 					       flow_dv_dest_array_match_cb,
350 					       flow_dv_dest_array_remove_cb,
351 					       flow_dv_dest_array_clone_cb,
352 					      flow_dv_dest_array_clone_free_cb);
353 	if (!sh->dest_array_list)
354 		goto error;
355 	/* Init shared flex parsers list, no need lcore_share */
356 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
357 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
358 					       mlx5_flex_parser_create_cb,
359 					       mlx5_flex_parser_match_cb,
360 					       mlx5_flex_parser_remove_cb,
361 					       mlx5_flex_parser_clone_cb,
362 					       mlx5_flex_parser_clone_free_cb);
363 	if (!sh->flex_parsers_dv)
364 		goto error;
365 #endif
366 #ifdef HAVE_MLX5DV_DR
367 	void *domain;
368 
369 	/* Reference counter is zero, we should initialize structures. */
370 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
371 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
372 	if (!domain) {
373 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
374 		err = errno;
375 		goto error;
376 	}
377 	sh->rx_domain = domain;
378 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
379 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
380 	if (!domain) {
381 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
382 		err = errno;
383 		goto error;
384 	}
385 	sh->tx_domain = domain;
386 #ifdef HAVE_MLX5DV_DR_ESWITCH
387 	if (priv->config.dv_esw_en) {
388 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
389 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
390 		if (!domain) {
391 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
392 			err = errno;
393 			goto error;
394 		}
395 		sh->fdb_domain = domain;
396 	}
397 	/*
398 	 * The drop action is just some dummy placeholder in rdma-core. It
399 	 * does not belong to domains and has no any attributes, and, can be
400 	 * shared by the entire device.
401 	 */
402 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
403 	if (!sh->dr_drop_action) {
404 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
405 		err = errno;
406 		goto error;
407 	}
408 #endif
409 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
410 		err = mlx5_alloc_tunnel_hub(sh);
411 	if (err) {
412 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
413 		goto error;
414 	}
415 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
416 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
417 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
418 		if (sh->fdb_domain)
419 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
420 	}
421 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
422 	if (!priv->config.allow_duplicate_pattern) {
423 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
424 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
425 #endif
426 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
427 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
428 		if (sh->fdb_domain)
429 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
430 	}
431 
432 	__mlx5_discovery_misc5_cap(priv);
433 #endif /* HAVE_MLX5DV_DR */
434 	sh->default_miss_action =
435 			mlx5_glue->dr_create_flow_action_default_miss();
436 	if (!sh->default_miss_action)
437 		DRV_LOG(WARNING, "Default miss action is not supported.");
438 	LIST_INIT(&sh->shared_rxqs);
439 	return 0;
440 error:
441 	/* Rollback the created objects. */
442 	if (sh->rx_domain) {
443 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
444 		sh->rx_domain = NULL;
445 	}
446 	if (sh->tx_domain) {
447 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
448 		sh->tx_domain = NULL;
449 	}
450 	if (sh->fdb_domain) {
451 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
452 		sh->fdb_domain = NULL;
453 	}
454 	if (sh->dr_drop_action) {
455 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
456 		sh->dr_drop_action = NULL;
457 	}
458 	if (sh->pop_vlan_action) {
459 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
460 		sh->pop_vlan_action = NULL;
461 	}
462 	if (sh->encaps_decaps) {
463 		mlx5_hlist_destroy(sh->encaps_decaps);
464 		sh->encaps_decaps = NULL;
465 	}
466 	if (sh->modify_cmds) {
467 		mlx5_hlist_destroy(sh->modify_cmds);
468 		sh->modify_cmds = NULL;
469 	}
470 	if (sh->tag_table) {
471 		/* tags should be destroyed with flow before. */
472 		mlx5_hlist_destroy(sh->tag_table);
473 		sh->tag_table = NULL;
474 	}
475 	if (sh->tunnel_hub) {
476 		mlx5_release_tunnel_hub(sh, priv->dev_port);
477 		sh->tunnel_hub = NULL;
478 	}
479 	mlx5_free_table_hash_list(priv);
480 	if (sh->port_id_action_list) {
481 		mlx5_list_destroy(sh->port_id_action_list);
482 		sh->port_id_action_list = NULL;
483 	}
484 	if (sh->push_vlan_action_list) {
485 		mlx5_list_destroy(sh->push_vlan_action_list);
486 		sh->push_vlan_action_list = NULL;
487 	}
488 	if (sh->sample_action_list) {
489 		mlx5_list_destroy(sh->sample_action_list);
490 		sh->sample_action_list = NULL;
491 	}
492 	if (sh->dest_array_list) {
493 		mlx5_list_destroy(sh->dest_array_list);
494 		sh->dest_array_list = NULL;
495 	}
496 	return err;
497 }
498 
499 /**
500  * Destroy DR related data within private structure.
501  *
502  * @param[in] priv
503  *   Pointer to the private device data structure.
504  */
505 void
506 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
507 {
508 	struct mlx5_dev_ctx_shared *sh = priv->sh;
509 
510 	MLX5_ASSERT(sh && sh->refcnt);
511 	if (sh->refcnt > 1)
512 		return;
513 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
514 #ifdef HAVE_MLX5DV_DR
515 	if (sh->rx_domain) {
516 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
517 		sh->rx_domain = NULL;
518 	}
519 	if (sh->tx_domain) {
520 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
521 		sh->tx_domain = NULL;
522 	}
523 #ifdef HAVE_MLX5DV_DR_ESWITCH
524 	if (sh->fdb_domain) {
525 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
526 		sh->fdb_domain = NULL;
527 	}
528 	if (sh->dr_drop_action) {
529 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
530 		sh->dr_drop_action = NULL;
531 	}
532 #endif
533 	if (sh->pop_vlan_action) {
534 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
535 		sh->pop_vlan_action = NULL;
536 	}
537 #endif /* HAVE_MLX5DV_DR */
538 	if (sh->default_miss_action)
539 		mlx5_glue->destroy_flow_action
540 				(sh->default_miss_action);
541 	if (sh->encaps_decaps) {
542 		mlx5_hlist_destroy(sh->encaps_decaps);
543 		sh->encaps_decaps = NULL;
544 	}
545 	if (sh->modify_cmds) {
546 		mlx5_hlist_destroy(sh->modify_cmds);
547 		sh->modify_cmds = NULL;
548 	}
549 	if (sh->tag_table) {
550 		/* tags should be destroyed with flow before. */
551 		mlx5_hlist_destroy(sh->tag_table);
552 		sh->tag_table = NULL;
553 	}
554 	if (sh->tunnel_hub) {
555 		mlx5_release_tunnel_hub(sh, priv->dev_port);
556 		sh->tunnel_hub = NULL;
557 	}
558 	mlx5_free_table_hash_list(priv);
559 	if (sh->port_id_action_list) {
560 		mlx5_list_destroy(sh->port_id_action_list);
561 		sh->port_id_action_list = NULL;
562 	}
563 	if (sh->push_vlan_action_list) {
564 		mlx5_list_destroy(sh->push_vlan_action_list);
565 		sh->push_vlan_action_list = NULL;
566 	}
567 	if (sh->sample_action_list) {
568 		mlx5_list_destroy(sh->sample_action_list);
569 		sh->sample_action_list = NULL;
570 	}
571 	if (sh->dest_array_list) {
572 		mlx5_list_destroy(sh->dest_array_list);
573 		sh->dest_array_list = NULL;
574 	}
575 }
576 
577 /**
578  * Initialize shared data between primary and secondary process.
579  *
580  * A memzone is reserved by primary process and secondary processes attach to
581  * the memzone.
582  *
583  * @return
584  *   0 on success, a negative errno value otherwise and rte_errno is set.
585  */
586 static int
587 mlx5_init_shared_data(void)
588 {
589 	const struct rte_memzone *mz;
590 	int ret = 0;
591 
592 	rte_spinlock_lock(&mlx5_shared_data_lock);
593 	if (mlx5_shared_data == NULL) {
594 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
595 			/* Allocate shared memory. */
596 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
597 						 sizeof(*mlx5_shared_data),
598 						 SOCKET_ID_ANY, 0);
599 			if (mz == NULL) {
600 				DRV_LOG(ERR,
601 					"Cannot allocate mlx5 shared data");
602 				ret = -rte_errno;
603 				goto error;
604 			}
605 			mlx5_shared_data = mz->addr;
606 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
607 			rte_spinlock_init(&mlx5_shared_data->lock);
608 		} else {
609 			/* Lookup allocated shared memory. */
610 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
611 			if (mz == NULL) {
612 				DRV_LOG(ERR,
613 					"Cannot attach mlx5 shared data");
614 				ret = -rte_errno;
615 				goto error;
616 			}
617 			mlx5_shared_data = mz->addr;
618 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
619 		}
620 	}
621 error:
622 	rte_spinlock_unlock(&mlx5_shared_data_lock);
623 	return ret;
624 }
625 
626 /**
627  * PMD global initialization.
628  *
629  * Independent from individual device, this function initializes global
630  * per-PMD data structures distinguishing primary and secondary processes.
631  * Hence, each initialization is called once per a process.
632  *
633  * @return
634  *   0 on success, a negative errno value otherwise and rte_errno is set.
635  */
636 static int
637 mlx5_init_once(void)
638 {
639 	struct mlx5_shared_data *sd;
640 	struct mlx5_local_data *ld = &mlx5_local_data;
641 	int ret = 0;
642 
643 	if (mlx5_init_shared_data())
644 		return -rte_errno;
645 	sd = mlx5_shared_data;
646 	MLX5_ASSERT(sd);
647 	rte_spinlock_lock(&sd->lock);
648 	switch (rte_eal_process_type()) {
649 	case RTE_PROC_PRIMARY:
650 		if (sd->init_done)
651 			break;
652 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
653 					   mlx5_mp_os_primary_handle);
654 		if (ret)
655 			goto out;
656 		sd->init_done = true;
657 		break;
658 	case RTE_PROC_SECONDARY:
659 		if (ld->init_done)
660 			break;
661 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
662 					     mlx5_mp_os_secondary_handle);
663 		if (ret)
664 			goto out;
665 		++sd->secondary_cnt;
666 		ld->init_done = true;
667 		break;
668 	default:
669 		break;
670 	}
671 out:
672 	rte_spinlock_unlock(&sd->lock);
673 	return ret;
674 }
675 
676 /**
677  * DR flow drop action support detect.
678  *
679  * @param dev
680  *   Pointer to rte_eth_dev structure.
681  *
682  */
683 static void
684 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
685 {
686 #ifdef HAVE_MLX5DV_DR
687 	struct mlx5_priv *priv = dev->data->dev_private;
688 
689 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
690 		return;
691 	/**
692 	 * DR supports drop action placeholder when it is supported;
693 	 * otherwise, use the queue drop action.
694 	 */
695 	if (!priv->sh->drop_action_check_flag) {
696 		if (!mlx5_flow_discover_dr_action_support(dev))
697 			priv->sh->dr_drop_action_en = 1;
698 		priv->sh->drop_action_check_flag = 1;
699 	}
700 	if (priv->sh->dr_drop_action_en)
701 		priv->root_drop_action = priv->sh->dr_drop_action;
702 	else
703 		priv->root_drop_action = priv->drop_queue.hrxq->action;
704 #endif
705 }
706 
707 static void
708 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
709 {
710 	struct mlx5_priv *priv = dev->data->dev_private;
711 	void *ctx = priv->sh->cdev->ctx;
712 
713 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
714 	if (!priv->q_counters) {
715 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
716 		struct ibv_wq *wq;
717 
718 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
719 			"by DevX - fall-back to use the kernel driver global "
720 			"queue counter.", dev->data->port_id);
721 		/* Create WQ by kernel and query its queue counter ID. */
722 		if (cq) {
723 			wq = mlx5_glue->create_wq(ctx,
724 						  &(struct ibv_wq_init_attr){
725 						    .wq_type = IBV_WQT_RQ,
726 						    .max_wr = 1,
727 						    .max_sge = 1,
728 						    .pd = priv->sh->cdev->pd,
729 						    .cq = cq,
730 						});
731 			if (wq) {
732 				/* Counter is assigned only on RDY state. */
733 				int ret = mlx5_glue->modify_wq(wq,
734 						 &(struct ibv_wq_attr){
735 						 .attr_mask = IBV_WQ_ATTR_STATE,
736 						 .wq_state = IBV_WQS_RDY,
737 						});
738 
739 				if (ret == 0)
740 					mlx5_devx_cmd_wq_query(wq,
741 							 &priv->counter_set_id);
742 				claim_zero(mlx5_glue->destroy_wq(wq));
743 			}
744 			claim_zero(mlx5_glue->destroy_cq(cq));
745 		}
746 	} else {
747 		priv->counter_set_id = priv->q_counters->id;
748 	}
749 	if (priv->counter_set_id == 0)
750 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
751 			"available.", dev->data->port_id);
752 }
753 
754 /**
755  * Check if representor spawn info match devargs.
756  *
757  * @param spawn
758  *   Verbs device parameters (name, port, switch_info) to spawn.
759  * @param eth_da
760  *   Device devargs to probe.
761  *
762  * @return
763  *   Match result.
764  */
765 static bool
766 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
767 		       struct rte_eth_devargs *eth_da)
768 {
769 	struct mlx5_switch_info *switch_info = &spawn->info;
770 	unsigned int p, f;
771 	uint16_t id;
772 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
773 						      eth_da->type);
774 
775 	switch (eth_da->type) {
776 	case RTE_ETH_REPRESENTOR_SF:
777 		if (!(spawn->info.port_name == -1 &&
778 		      switch_info->name_type ==
779 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
780 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
781 			rte_errno = EBUSY;
782 			return false;
783 		}
784 		break;
785 	case RTE_ETH_REPRESENTOR_VF:
786 		/* Allows HPF representor index -1 as exception. */
787 		if (!(spawn->info.port_name == -1 &&
788 		      switch_info->name_type ==
789 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
790 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
791 			rte_errno = EBUSY;
792 			return false;
793 		}
794 		break;
795 	case RTE_ETH_REPRESENTOR_NONE:
796 		rte_errno = EBUSY;
797 		return false;
798 	default:
799 		rte_errno = ENOTSUP;
800 		DRV_LOG(ERR, "unsupported representor type");
801 		return false;
802 	}
803 	/* Check representor ID: */
804 	for (p = 0; p < eth_da->nb_ports; ++p) {
805 		if (spawn->pf_bond < 0) {
806 			/* For non-LAG mode, allow and ignore pf. */
807 			switch_info->pf_num = eth_da->ports[p];
808 			repr_id = mlx5_representor_id_encode(switch_info,
809 							     eth_da->type);
810 		}
811 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
812 			id = MLX5_REPRESENTOR_ID
813 				(eth_da->ports[p], eth_da->type,
814 				 eth_da->representor_ports[f]);
815 			if (repr_id == id)
816 				return true;
817 		}
818 	}
819 	rte_errno = EBUSY;
820 	return false;
821 }
822 
823 /**
824  * Spawn an Ethernet device from Verbs information.
825  *
826  * @param dpdk_dev
827  *   Backing DPDK device.
828  * @param spawn
829  *   Verbs device parameters (name, port, switch_info) to spawn.
830  * @param config
831  *   Device configuration parameters.
832  * @param eth_da
833  *   Device arguments.
834  *
835  * @return
836  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
837  *   is set. The following errors are defined:
838  *
839  *   EBUSY: device is not supposed to be spawned.
840  *   EEXIST: device is already spawned
841  */
842 static struct rte_eth_dev *
843 mlx5_dev_spawn(struct rte_device *dpdk_dev,
844 	       struct mlx5_dev_spawn_data *spawn,
845 	       struct mlx5_dev_config *config,
846 	       struct rte_eth_devargs *eth_da)
847 {
848 	const struct mlx5_switch_info *switch_info = &spawn->info;
849 	struct mlx5_dev_ctx_shared *sh = NULL;
850 	struct mlx5_hca_attr *hca_attr = &spawn->cdev->config.hca_attr;
851 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
852 	struct rte_eth_dev *eth_dev = NULL;
853 	struct mlx5_priv *priv = NULL;
854 	int err = 0;
855 	unsigned int hw_padding = 0;
856 	unsigned int mps;
857 	unsigned int mpls_en = 0;
858 	unsigned int swp = 0;
859 	unsigned int mprq = 0;
860 	struct rte_ether_addr mac;
861 	char name[RTE_ETH_NAME_MAX_LEN];
862 	int own_domain_id = 0;
863 	uint16_t port_id;
864 	struct mlx5_port_info vport_info = { .query_flags = 0 };
865 	int nl_rdma = -1;
866 	int i;
867 
868 	/* Determine if this port representor is supposed to be spawned. */
869 	if (switch_info->representor && dpdk_dev->devargs &&
870 	    !mlx5_representor_match(spawn, eth_da))
871 		return NULL;
872 	/* Build device name. */
873 	if (spawn->pf_bond < 0) {
874 		/* Single device. */
875 		if (!switch_info->representor)
876 			strlcpy(name, dpdk_dev->name, sizeof(name));
877 		else
878 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
879 				 dpdk_dev->name,
880 				 switch_info->name_type ==
881 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
882 				 switch_info->port_name);
883 	} else {
884 		/* Bonding device. */
885 		if (!switch_info->representor) {
886 			err = snprintf(name, sizeof(name), "%s_%s",
887 				       dpdk_dev->name, spawn->phys_dev_name);
888 		} else {
889 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
890 				dpdk_dev->name, spawn->phys_dev_name,
891 				switch_info->ctrl_num,
892 				switch_info->pf_num,
893 				switch_info->name_type ==
894 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
895 				switch_info->port_name);
896 		}
897 	}
898 	if (err >= (int)sizeof(name))
899 		DRV_LOG(WARNING, "device name overflow %s", name);
900 	/* check if the device is already spawned */
901 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
902 		rte_errno = EEXIST;
903 		return NULL;
904 	}
905 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
906 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
907 		struct mlx5_mp_id mp_id;
908 
909 		eth_dev = rte_eth_dev_attach_secondary(name);
910 		if (eth_dev == NULL) {
911 			DRV_LOG(ERR, "can not attach rte ethdev");
912 			rte_errno = ENOMEM;
913 			return NULL;
914 		}
915 		eth_dev->device = dpdk_dev;
916 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
917 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
918 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
919 		err = mlx5_proc_priv_init(eth_dev);
920 		if (err)
921 			return NULL;
922 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
923 		/* Receive command fd from primary process */
924 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
925 		if (err < 0)
926 			goto err_secondary;
927 		/* Remap UAR for Tx queues. */
928 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
929 		if (err)
930 			goto err_secondary;
931 		/*
932 		 * Ethdev pointer is still required as input since
933 		 * the primary device is not accessible from the
934 		 * secondary process.
935 		 */
936 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
937 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
938 		return eth_dev;
939 err_secondary:
940 		mlx5_dev_close(eth_dev);
941 		return NULL;
942 	}
943 	/* Process parameters. */
944 	err = mlx5_args(config, dpdk_dev->devargs);
945 	if (err) {
946 		DRV_LOG(ERR, "failed to process device arguments: %s",
947 			strerror(rte_errno));
948 		return NULL;
949 	}
950 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
951 	if (!sh)
952 		return NULL;
953 	/* Update final values for devargs before check sibling config. */
954 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
955 	if (config->dv_flow_en) {
956 		DRV_LOG(WARNING, "DV flow is not supported.");
957 		config->dv_flow_en = 0;
958 	}
959 #endif
960 #ifdef HAVE_MLX5DV_DR_ESWITCH
961 	if (!(hca_attr->eswitch_manager && config->dv_flow_en && sh->esw_mode))
962 		config->dv_esw_en = 0;
963 #else
964 	config->dv_esw_en = 0;
965 #endif
966 	if (config->dv_miss_info && config->dv_esw_en)
967 		config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
968 	if (!config->dv_esw_en &&
969 	    config->dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
970 		DRV_LOG(WARNING,
971 			"Metadata mode %u is not supported (no E-Switch).",
972 			config->dv_xmeta_en);
973 		config->dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
974 	}
975 	/* Check sibling device configurations. */
976 	err = mlx5_dev_check_sibling_config(sh, config, dpdk_dev);
977 	if (err)
978 		goto error;
979 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
980 	config->dest_tir = 1;
981 #endif
982 	/*
983 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
984 	 * as all ConnectX-5 devices.
985 	 */
986 	if (sh->device_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
987 		if (sh->device_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
988 			DRV_LOG(DEBUG, "enhanced MPW is supported");
989 			mps = MLX5_MPW_ENHANCED;
990 		} else {
991 			DRV_LOG(DEBUG, "MPW is supported");
992 			mps = MLX5_MPW;
993 		}
994 	} else {
995 		DRV_LOG(DEBUG, "MPW isn't supported");
996 		mps = MLX5_MPW_DISABLED;
997 	}
998 #ifdef HAVE_IBV_MLX5_MOD_SWP
999 	if (sh->device_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1000 		swp = sh->device_attr.sw_parsing_offloads;
1001 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1002 #endif
1003 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
1004 		MLX5_SW_PARSING_TSO_CAP);
1005 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1006 	if (sh->device_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1007 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1008 			sh->device_attr.min_single_stride_log_num_of_bytes);
1009 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1010 			sh->device_attr.max_single_stride_log_num_of_bytes);
1011 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1012 			sh->device_attr.min_single_wqe_log_num_of_strides);
1013 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1014 			sh->device_attr.max_single_wqe_log_num_of_strides);
1015 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1016 			sh->device_attr.stride_supported_qpts);
1017 		DRV_LOG(DEBUG, "\tmin_stride_wqe_log_size: %d",
1018 			config->mprq.log_min_stride_wqe_size);
1019 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1020 		mprq = 1;
1021 		config->mprq.log_min_stride_size =
1022 			sh->device_attr.min_single_stride_log_num_of_bytes;
1023 		config->mprq.log_max_stride_size =
1024 			sh->device_attr.max_single_stride_log_num_of_bytes;
1025 		config->mprq.log_min_stride_num =
1026 			sh->device_attr.min_single_wqe_log_num_of_strides;
1027 		config->mprq.log_max_stride_num =
1028 			sh->device_attr.max_single_wqe_log_num_of_strides;
1029 	}
1030 #endif
1031 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1032 	if (sh->device_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1033 		config->tunnel_en = sh->device_attr.tunnel_offloads_caps &
1034 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1035 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1036 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1037 	}
1038 	if (config->tunnel_en) {
1039 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1040 		config->tunnel_en &
1041 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1042 		config->tunnel_en &
1043 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1044 		config->tunnel_en &
1045 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1046 		);
1047 	} else {
1048 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1049 	}
1050 #else
1051 	DRV_LOG(WARNING,
1052 		"tunnel offloading disabled due to old OFED/rdma-core version");
1053 #endif
1054 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1055 	mpls_en = ((sh->device_attr.tunnel_offloads_caps &
1056 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1057 		   (sh->device_attr.tunnel_offloads_caps &
1058 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1059 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1060 		mpls_en ? "" : "not ");
1061 #else
1062 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1063 		" old OFED/rdma-core version or firmware configuration");
1064 #endif
1065 	config->mpls_en = mpls_en;
1066 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1067 	/* Check port status. */
1068 	if (spawn->phys_port <= UINT8_MAX) {
1069 		/* Legacy Verbs api only support u8 port number. */
1070 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1071 					    &port_attr);
1072 		if (err) {
1073 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1074 			goto error;
1075 		}
1076 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1077 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1078 			err = EINVAL;
1079 			goto error;
1080 		}
1081 	} else if (nl_rdma >= 0) {
1082 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1083 		err = mlx5_nl_port_state(nl_rdma,
1084 			spawn->phys_dev_name,
1085 			spawn->phys_port);
1086 		if (err < 0) {
1087 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1088 				strerror(rte_errno));
1089 			err = -rte_errno;
1090 			goto error;
1091 		}
1092 		port_attr.state = (enum ibv_port_state)err;
1093 	}
1094 	if (port_attr.state != IBV_PORT_ACTIVE)
1095 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1096 			mlx5_glue->port_state_str(port_attr.state),
1097 			port_attr.state);
1098 	/* Allocate private eth device data. */
1099 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1100 			   sizeof(*priv),
1101 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1102 	if (priv == NULL) {
1103 		DRV_LOG(ERR, "priv allocation failure");
1104 		err = ENOMEM;
1105 		goto error;
1106 	}
1107 	priv->sh = sh;
1108 	priv->dev_port = spawn->phys_port;
1109 	priv->pci_dev = spawn->pci_dev;
1110 	priv->mtu = RTE_ETHER_MTU;
1111 	/* Some internal functions rely on Netlink sockets, open them now. */
1112 	priv->nl_socket_rdma = nl_rdma;
1113 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1114 	priv->representor = !!switch_info->representor;
1115 	priv->master = !!switch_info->master;
1116 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1117 	priv->vport_meta_tag = 0;
1118 	priv->vport_meta_mask = 0;
1119 	priv->pf_bond = spawn->pf_bond;
1120 
1121 	DRV_LOG(DEBUG,
1122 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1123 		priv->dev_port, dpdk_dev->bus->name,
1124 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1125 		priv->master, priv->representor, priv->pf_bond);
1126 
1127 	/*
1128 	 * If we have E-Switch we should determine the vport attributes.
1129 	 * E-Switch may use either source vport field or reg_c[0] metadata
1130 	 * register to match on vport index. The engaged part of metadata
1131 	 * register is defined by mask.
1132 	 */
1133 	if (sh->esw_mode) {
1134 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1135 						 spawn->phys_port,
1136 						 &vport_info);
1137 		if (err) {
1138 			DRV_LOG(WARNING,
1139 				"Cannot query devx port %d on device %s",
1140 				spawn->phys_port, spawn->phys_dev_name);
1141 			vport_info.query_flags = 0;
1142 		}
1143 	}
1144 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1145 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1146 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1147 		if (!priv->vport_meta_mask) {
1148 			DRV_LOG(ERR,
1149 				"vport zero mask for port %d on bonding device %s",
1150 				spawn->phys_port, spawn->phys_dev_name);
1151 			err = ENOTSUP;
1152 			goto error;
1153 		}
1154 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1155 			DRV_LOG(ERR,
1156 				"Invalid vport tag for port %d on bonding device %s",
1157 				spawn->phys_port, spawn->phys_dev_name);
1158 			err = ENOTSUP;
1159 			goto error;
1160 		}
1161 	}
1162 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1163 		priv->vport_id = vport_info.vport_id;
1164 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1165 		DRV_LOG(ERR,
1166 			"Cannot deduce vport index for port %d on bonding device %s",
1167 			spawn->phys_port, spawn->phys_dev_name);
1168 		err = ENOTSUP;
1169 		goto error;
1170 	} else {
1171 		/*
1172 		 * Suppose vport index in compatible way. Kernel/rdma_core
1173 		 * support single E-Switch per PF configurations only and
1174 		 * vport_id field contains the vport index for associated VF,
1175 		 * which is deduced from representor port name.
1176 		 * For example, let's have the IB device port 10, it has
1177 		 * attached network device eth0, which has port name attribute
1178 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1179 		 * as 3 (2+1). This assigning schema should be changed if the
1180 		 * multiple E-Switch instances per PF configurations or/and PCI
1181 		 * subfunctions are added.
1182 		 */
1183 		priv->vport_id = switch_info->representor ?
1184 				 switch_info->port_name + 1 : -1;
1185 	}
1186 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1187 							  eth_da->type);
1188 	/*
1189 	 * Look for sibling devices in order to reuse their switch domain
1190 	 * if any, otherwise allocate one.
1191 	 */
1192 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1193 		const struct mlx5_priv *opriv =
1194 			rte_eth_devices[port_id].data->dev_private;
1195 
1196 		if (!opriv ||
1197 		    opriv->sh != priv->sh ||
1198 			opriv->domain_id ==
1199 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1200 			continue;
1201 		priv->domain_id = opriv->domain_id;
1202 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1203 			priv->dev_port, priv->domain_id);
1204 		break;
1205 	}
1206 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1207 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1208 		if (err) {
1209 			err = rte_errno;
1210 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1211 				strerror(rte_errno));
1212 			goto error;
1213 		}
1214 		own_domain_id = 1;
1215 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1216 			priv->dev_port, priv->domain_id);
1217 	}
1218 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1219 			    IBV_DEVICE_RAW_IP_CSUM);
1220 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1221 		(config->hw_csum ? "" : "not "));
1222 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1223 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1224 	DRV_LOG(DEBUG, "counters are not supported");
1225 #endif
1226 	config->ind_table_max_size =
1227 		sh->device_attr.max_rwq_indirection_table_size;
1228 	/*
1229 	 * Remove this check once DPDK supports larger/variable
1230 	 * indirection tables.
1231 	 */
1232 	if (config->ind_table_max_size > (unsigned int)RTE_ETH_RSS_RETA_SIZE_512)
1233 		config->ind_table_max_size = RTE_ETH_RSS_RETA_SIZE_512;
1234 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1235 		config->ind_table_max_size);
1236 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1237 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1238 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1239 		(config->hw_vlan_strip ? "" : "not "));
1240 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1241 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1242 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1243 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1244 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1245 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1246 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1247 #endif
1248 	if (config->hw_padding && !hw_padding) {
1249 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1250 		config->hw_padding = 0;
1251 	} else if (config->hw_padding) {
1252 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1253 	}
1254 	config->tso = (sh->device_attr.max_tso > 0 &&
1255 		      (sh->device_attr.tso_supported_qpts &
1256 		       (1 << IBV_QPT_RAW_PACKET)));
1257 	if (config->tso)
1258 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1259 	/*
1260 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1261 	 * by default.
1262 	 */
1263 	if (config->mps == MLX5_ARG_UNSET)
1264 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1265 							  MLX5_MPW_DISABLED;
1266 	else
1267 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1268 	DRV_LOG(INFO, "%sMPS is %s",
1269 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1270 		config->mps == MLX5_MPW ? "legacy " : "",
1271 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1272 	if (sh->cdev->config.devx) {
1273 		sh->steering_format_version = hca_attr->steering_format_version;
1274 		/* Check for LRO support. */
1275 		if (config->dest_tir && hca_attr->lro_cap &&
1276 		    config->dv_flow_en) {
1277 			/* TBD check tunnel lro caps. */
1278 			config->lro.supported = hca_attr->lro_cap;
1279 			DRV_LOG(DEBUG, "Device supports LRO");
1280 			/*
1281 			 * If LRO timeout is not configured by application,
1282 			 * use the minimal supported value.
1283 			 */
1284 			if (!config->lro.timeout)
1285 				config->lro.timeout =
1286 				       hca_attr->lro_timer_supported_periods[0];
1287 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1288 				config->lro.timeout);
1289 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1290 				"required for coalescing is %d bytes",
1291 				hca_attr->lro_min_mss_size);
1292 		}
1293 #if defined(HAVE_MLX5DV_DR) && \
1294 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1295 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1296 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1297 		    config->dv_flow_en) {
1298 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1299 			/*
1300 			 * Meter needs two REG_C's for color match and pre-sfx
1301 			 * flow match. Here get the REG_C for color match.
1302 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1303 			 */
1304 			reg_c_mask &= 0xfc;
1305 			if (__builtin_popcount(reg_c_mask) < 1) {
1306 				priv->mtr_en = 0;
1307 				DRV_LOG(WARNING, "No available register for"
1308 					" meter.");
1309 			} else {
1310 				/*
1311 				 * The meter color register is used by the
1312 				 * flow-hit feature as well.
1313 				 * The flow-hit feature must use REG_C_3
1314 				 * Prefer REG_C_3 if it is available.
1315 				 */
1316 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1317 					priv->mtr_color_reg = REG_C_3;
1318 				else
1319 					priv->mtr_color_reg = ffs(reg_c_mask)
1320 							      - 1 + REG_C_0;
1321 				priv->mtr_en = 1;
1322 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1323 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1324 					priv->mtr_color_reg);
1325 			}
1326 		}
1327 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1328 			uint32_t log_obj_size =
1329 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1330 			if (log_obj_size >=
1331 			    hca_attr->qos.log_meter_aso_granularity &&
1332 			    log_obj_size <=
1333 			    hca_attr->qos.log_meter_aso_max_alloc)
1334 				sh->meter_aso_en = 1;
1335 		}
1336 		if (priv->mtr_en) {
1337 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1338 			if (err) {
1339 				err = -err;
1340 				goto error;
1341 			}
1342 		}
1343 		if (hca_attr->flow.tunnel_header_0_1)
1344 			sh->tunnel_header_0_1 = 1;
1345 #endif
1346 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1347 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1348 			sh->flow_hit_aso_en = 1;
1349 			err = mlx5_flow_aso_age_mng_init(sh);
1350 			if (err) {
1351 				err = -err;
1352 				goto error;
1353 			}
1354 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1355 		}
1356 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1357 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1358 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1359 		if (hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1360 			err = mlx5_flow_aso_ct_mng_init(sh);
1361 			if (err) {
1362 				err = -err;
1363 				goto error;
1364 			}
1365 			DRV_LOG(DEBUG, "CT ASO is supported.");
1366 			sh->ct_aso_en = 1;
1367 		}
1368 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1369 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1370 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1371 		    config->dv_flow_en) {
1372 			priv->sampler_en = 1;
1373 			DRV_LOG(DEBUG, "Sampler enabled!");
1374 		} else {
1375 			priv->sampler_en = 0;
1376 			if (!hca_attr->log_max_ft_sampler_num)
1377 				DRV_LOG(WARNING,
1378 					"No available register for sampler.");
1379 			else
1380 				DRV_LOG(DEBUG, "DV flow is not supported!");
1381 		}
1382 #endif
1383 	}
1384 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1385 	    !(sh->device_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1386 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1387 		config->cqe_comp = 0;
1388 	}
1389 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1390 	    (!sh->cdev->config.devx || !hca_attr->mini_cqe_resp_flow_tag)) {
1391 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1392 				 " format isn't supported.");
1393 		config->cqe_comp = 0;
1394 	}
1395 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1396 	    (!sh->cdev->config.devx || !hca_attr->mini_cqe_resp_l3_l4_tag)) {
1397 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1398 				 " format isn't supported.");
1399 		config->cqe_comp = 0;
1400 	}
1401 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1402 			config->cqe_comp ? "" : "not ");
1403 	if (config->tx_pp) {
1404 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1405 			hca_attr->dev_freq_khz);
1406 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1407 			hca_attr->qos.packet_pacing ? "" : "not ");
1408 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1409 			hca_attr->cross_channel ? "" : "not ");
1410 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1411 			hca_attr->wqe_index_ignore ? "" : "not ");
1412 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1413 			hca_attr->non_wire_sq ? "" : "not ");
1414 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1415 			hca_attr->log_max_static_sq_wq ? "" : "not ",
1416 			hca_attr->log_max_static_sq_wq);
1417 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1418 			hca_attr->qos.wqe_rate_pp ? "" : "not ");
1419 		if (!sh->cdev->config.devx) {
1420 			DRV_LOG(ERR, "DevX is required for packet pacing");
1421 			err = ENODEV;
1422 			goto error;
1423 		}
1424 		if (!hca_attr->qos.packet_pacing) {
1425 			DRV_LOG(ERR, "Packet pacing is not supported");
1426 			err = ENODEV;
1427 			goto error;
1428 		}
1429 		if (!hca_attr->cross_channel) {
1430 			DRV_LOG(ERR, "Cross channel operations are"
1431 				     " required for packet pacing");
1432 			err = ENODEV;
1433 			goto error;
1434 		}
1435 		if (!hca_attr->wqe_index_ignore) {
1436 			DRV_LOG(ERR, "WQE index ignore feature is"
1437 				     " required for packet pacing");
1438 			err = ENODEV;
1439 			goto error;
1440 		}
1441 		if (!hca_attr->non_wire_sq) {
1442 			DRV_LOG(ERR, "Non-wire SQ feature is"
1443 				     " required for packet pacing");
1444 			err = ENODEV;
1445 			goto error;
1446 		}
1447 		if (!hca_attr->log_max_static_sq_wq) {
1448 			DRV_LOG(ERR, "Static WQE SQ feature is"
1449 				     " required for packet pacing");
1450 			err = ENODEV;
1451 			goto error;
1452 		}
1453 		if (!hca_attr->qos.wqe_rate_pp) {
1454 			DRV_LOG(ERR, "WQE rate mode is required"
1455 				     " for packet pacing");
1456 			err = ENODEV;
1457 			goto error;
1458 		}
1459 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1460 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1461 			     " can't create queues for packet pacing");
1462 		err = ENODEV;
1463 		goto error;
1464 #endif
1465 	}
1466 	if (config->std_delay_drop || config->hp_delay_drop) {
1467 		if (!hca_attr->rq_delay_drop) {
1468 			config->std_delay_drop = 0;
1469 			config->hp_delay_drop = 0;
1470 			DRV_LOG(WARNING,
1471 				"dev_port-%u: Rxq delay drop is not supported",
1472 				priv->dev_port);
1473 		}
1474 	}
1475 	if (sh->cdev->config.devx)
1476 		mlx5_rt_timestamp_config(sh, config, hca_attr);
1477 	/*
1478 	 * If HW has bug working with tunnel packet decapsulation and
1479 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1480 	 * bit. Then RTE_ETH_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1481 	 */
1482 	if (hca_attr->scatter_fcs_w_decap_disable && config->decap_en)
1483 		config->hw_fcs_strip = 0;
1484 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1485 		(config->hw_fcs_strip ? "" : "not "));
1486 	if (config->mprq.enabled && !mprq) {
1487 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1488 		config->mprq.enabled = 0;
1489 	}
1490 	if (config->max_dump_files_num == 0)
1491 		config->max_dump_files_num = 128;
1492 	eth_dev = rte_eth_dev_allocate(name);
1493 	if (eth_dev == NULL) {
1494 		DRV_LOG(ERR, "can not allocate rte ethdev");
1495 		err = ENOMEM;
1496 		goto error;
1497 	}
1498 	if (priv->representor) {
1499 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1500 		eth_dev->data->representor_id = priv->representor_id;
1501 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1502 			struct mlx5_priv *opriv =
1503 				rte_eth_devices[port_id].data->dev_private;
1504 			if (opriv &&
1505 			    opriv->master &&
1506 			    opriv->domain_id == priv->domain_id &&
1507 			    opriv->sh == priv->sh) {
1508 				eth_dev->data->backer_port_id = port_id;
1509 				break;
1510 			}
1511 		}
1512 		if (port_id >= RTE_MAX_ETHPORTS)
1513 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1514 	}
1515 	priv->mp_id.port_id = eth_dev->data->port_id;
1516 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1517 	/*
1518 	 * Store associated network device interface index. This index
1519 	 * is permanent throughout the lifetime of device. So, we may store
1520 	 * the ifindex here and use the cached value further.
1521 	 */
1522 	MLX5_ASSERT(spawn->ifindex);
1523 	priv->if_index = spawn->ifindex;
1524 	priv->lag_affinity_idx = sh->refcnt - 1;
1525 	eth_dev->data->dev_private = priv;
1526 	priv->dev_data = eth_dev->data;
1527 	eth_dev->data->mac_addrs = priv->mac;
1528 	eth_dev->device = dpdk_dev;
1529 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1530 	/* Configure the first MAC address by default. */
1531 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1532 		DRV_LOG(ERR,
1533 			"port %u cannot get MAC address, is mlx5_en"
1534 			" loaded? (errno: %s)",
1535 			eth_dev->data->port_id, strerror(rte_errno));
1536 		err = ENODEV;
1537 		goto error;
1538 	}
1539 	DRV_LOG(INFO,
1540 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1541 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1542 #ifdef RTE_LIBRTE_MLX5_DEBUG
1543 	{
1544 		char ifname[MLX5_NAMESIZE];
1545 
1546 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1547 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1548 				eth_dev->data->port_id, ifname);
1549 		else
1550 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1551 				eth_dev->data->port_id);
1552 	}
1553 #endif
1554 	/* Get actual MTU if possible. */
1555 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1556 	if (err) {
1557 		err = rte_errno;
1558 		goto error;
1559 	}
1560 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1561 		priv->mtu);
1562 	/* Initialize burst functions to prevent crashes before link-up. */
1563 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1564 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1565 	eth_dev->dev_ops = &mlx5_dev_ops;
1566 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1567 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1568 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1569 	/* Register MAC address. */
1570 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1571 	if (config->vf && config->vf_nl_en)
1572 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1573 				      mlx5_ifindex(eth_dev),
1574 				      eth_dev->data->mac_addrs,
1575 				      MLX5_MAX_MAC_ADDRESSES);
1576 	priv->ctrl_flows = 0;
1577 	rte_spinlock_init(&priv->flow_list_lock);
1578 	TAILQ_INIT(&priv->flow_meters);
1579 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1580 	if (!priv->mtr_profile_tbl)
1581 		goto error;
1582 	/* Bring Ethernet device up. */
1583 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1584 		eth_dev->data->port_id);
1585 	mlx5_set_link_up(eth_dev);
1586 	/*
1587 	 * Even though the interrupt handler is not installed yet,
1588 	 * interrupts will still trigger on the async_fd from
1589 	 * Verbs context returned by ibv_open_device().
1590 	 */
1591 	mlx5_link_update(eth_dev, 0);
1592 	/* Detect minimal data bytes to inline. */
1593 	mlx5_set_min_inline(spawn, config);
1594 	/* Store device configuration on private structure. */
1595 	priv->config = *config;
1596 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1597 		icfg[i].release_mem_en = !!config->reclaim_mode;
1598 		if (config->reclaim_mode)
1599 			icfg[i].per_core_cache = 0;
1600 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1601 		if (!priv->flows[i])
1602 			goto error;
1603 	}
1604 	/* Create context for virtual machine VLAN workaround. */
1605 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1606 	if (config->dv_flow_en) {
1607 		err = mlx5_alloc_shared_dr(priv);
1608 		if (err)
1609 			goto error;
1610 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1611 			goto error;
1612 	}
1613 	if (sh->cdev->config.devx && config->dv_flow_en && config->dest_tir) {
1614 		priv->obj_ops = devx_obj_ops;
1615 		mlx5_queue_counter_id_prepare(eth_dev);
1616 		priv->obj_ops.lb_dummy_queue_create =
1617 					mlx5_rxq_ibv_obj_dummy_lb_create;
1618 		priv->obj_ops.lb_dummy_queue_release =
1619 					mlx5_rxq_ibv_obj_dummy_lb_release;
1620 	} else if (spawn->max_port > UINT8_MAX) {
1621 		/* Verbs can't support ports larger than 255 by design. */
1622 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1623 		err = ENOTSUP;
1624 		goto error;
1625 	} else {
1626 		priv->obj_ops = ibv_obj_ops;
1627 	}
1628 	if (config->tx_pp &&
1629 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1630 		/*
1631 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1632 		 * packet pacing and already checked above.
1633 		 * Hence, we should only make sure the SQs will be created
1634 		 * with DevX, not with Verbs.
1635 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1636 		 * with Clock Queue UAR as required for Tx scheduling.
1637 		 */
1638 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1639 		err = ENODEV;
1640 		goto error;
1641 	}
1642 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1643 	if (!priv->drop_queue.hrxq)
1644 		goto error;
1645 	/* Port representor shares the same max priority with pf port. */
1646 	if (!priv->sh->flow_priority_check_flag) {
1647 		/* Supported Verbs flow priority number detection. */
1648 		err = mlx5_flow_discover_priorities(eth_dev);
1649 		priv->sh->flow_max_priority = err;
1650 		priv->sh->flow_priority_check_flag = 1;
1651 	} else {
1652 		err = priv->sh->flow_max_priority;
1653 	}
1654 	if (err < 0) {
1655 		err = -err;
1656 		goto error;
1657 	}
1658 	mlx5_set_metadata_mask(eth_dev);
1659 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1660 	    !priv->sh->dv_regc0_mask) {
1661 		DRV_LOG(ERR, "metadata mode %u is not supported "
1662 			     "(no metadata reg_c[0] is available)",
1663 			     priv->config.dv_xmeta_en);
1664 			err = ENOTSUP;
1665 			goto error;
1666 	}
1667 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1668 				       mlx5_hrxq_create_cb,
1669 				       mlx5_hrxq_match_cb,
1670 				       mlx5_hrxq_remove_cb,
1671 				       mlx5_hrxq_clone_cb,
1672 				       mlx5_hrxq_clone_free_cb);
1673 	if (!priv->hrxqs)
1674 		goto error;
1675 	rte_rwlock_init(&priv->ind_tbls_lock);
1676 	/* Query availability of metadata reg_c's. */
1677 	if (!priv->sh->metadata_regc_check_flag) {
1678 		err = mlx5_flow_discover_mreg_c(eth_dev);
1679 		if (err < 0) {
1680 			err = -err;
1681 			goto error;
1682 		}
1683 	}
1684 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1685 		DRV_LOG(DEBUG,
1686 			"port %u extensive metadata register is not supported",
1687 			eth_dev->data->port_id);
1688 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1689 			DRV_LOG(ERR, "metadata mode %u is not supported "
1690 				     "(no metadata registers available)",
1691 				     priv->config.dv_xmeta_en);
1692 			err = ENOTSUP;
1693 			goto error;
1694 		}
1695 	}
1696 	if (priv->config.dv_flow_en &&
1697 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1698 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1699 	    priv->sh->dv_regc0_mask) {
1700 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1701 						      MLX5_FLOW_MREG_HTABLE_SZ,
1702 						      false, true, eth_dev,
1703 						      flow_dv_mreg_create_cb,
1704 						      flow_dv_mreg_match_cb,
1705 						      flow_dv_mreg_remove_cb,
1706 						      flow_dv_mreg_clone_cb,
1707 						    flow_dv_mreg_clone_free_cb);
1708 		if (!priv->mreg_cp_tbl) {
1709 			err = ENOMEM;
1710 			goto error;
1711 		}
1712 	}
1713 	rte_spinlock_init(&priv->shared_act_sl);
1714 	mlx5_flow_counter_mode_config(eth_dev);
1715 	mlx5_flow_drop_action_config(eth_dev);
1716 	if (priv->config.dv_flow_en)
1717 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1718 	return eth_dev;
1719 error:
1720 	if (priv) {
1721 		if (priv->mreg_cp_tbl)
1722 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1723 		if (priv->sh)
1724 			mlx5_os_free_shared_dr(priv);
1725 		if (priv->nl_socket_route >= 0)
1726 			close(priv->nl_socket_route);
1727 		if (priv->vmwa_context)
1728 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1729 		if (eth_dev && priv->drop_queue.hrxq)
1730 			mlx5_drop_action_destroy(eth_dev);
1731 		if (priv->mtr_profile_tbl)
1732 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1733 		if (own_domain_id)
1734 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1735 		if (priv->hrxqs)
1736 			mlx5_list_destroy(priv->hrxqs);
1737 		if (eth_dev && priv->flex_item_map)
1738 			mlx5_flex_item_port_cleanup(eth_dev);
1739 		mlx5_free(priv);
1740 		if (eth_dev != NULL)
1741 			eth_dev->data->dev_private = NULL;
1742 	}
1743 	if (eth_dev != NULL) {
1744 		/* mac_addrs must not be freed alone because part of
1745 		 * dev_private
1746 		 **/
1747 		eth_dev->data->mac_addrs = NULL;
1748 		rte_eth_dev_release_port(eth_dev);
1749 	}
1750 	if (sh)
1751 		mlx5_free_shared_dev_ctx(sh);
1752 	if (nl_rdma >= 0)
1753 		close(nl_rdma);
1754 	MLX5_ASSERT(err > 0);
1755 	rte_errno = err;
1756 	return NULL;
1757 }
1758 
1759 /**
1760  * Comparison callback to sort device data.
1761  *
1762  * This is meant to be used with qsort().
1763  *
1764  * @param a[in]
1765  *   Pointer to pointer to first data object.
1766  * @param b[in]
1767  *   Pointer to pointer to second data object.
1768  *
1769  * @return
1770  *   0 if both objects are equal, less than 0 if the first argument is less
1771  *   than the second, greater than 0 otherwise.
1772  */
1773 static int
1774 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1775 {
1776 	const struct mlx5_switch_info *si_a =
1777 		&((const struct mlx5_dev_spawn_data *)a)->info;
1778 	const struct mlx5_switch_info *si_b =
1779 		&((const struct mlx5_dev_spawn_data *)b)->info;
1780 	int ret;
1781 
1782 	/* Master device first. */
1783 	ret = si_b->master - si_a->master;
1784 	if (ret)
1785 		return ret;
1786 	/* Then representor devices. */
1787 	ret = si_b->representor - si_a->representor;
1788 	if (ret)
1789 		return ret;
1790 	/* Unidentified devices come last in no specific order. */
1791 	if (!si_a->representor)
1792 		return 0;
1793 	/* Order representors by name. */
1794 	return si_a->port_name - si_b->port_name;
1795 }
1796 
1797 /**
1798  * Match PCI information for possible slaves of bonding device.
1799  *
1800  * @param[in] ibdev_name
1801  *   Name of Infiniband device.
1802  * @param[in] pci_dev
1803  *   Pointer to primary PCI address structure to match.
1804  * @param[in] nl_rdma
1805  *   Netlink RDMA group socket handle.
1806  * @param[in] owner
1807  *   Representor owner PF index.
1808  * @param[out] bond_info
1809  *   Pointer to bonding information.
1810  *
1811  * @return
1812  *   negative value if no bonding device found, otherwise
1813  *   positive index of slave PF in bonding.
1814  */
1815 static int
1816 mlx5_device_bond_pci_match(const char *ibdev_name,
1817 			   const struct rte_pci_addr *pci_dev,
1818 			   int nl_rdma, uint16_t owner,
1819 			   struct mlx5_bond_info *bond_info)
1820 {
1821 	char ifname[IF_NAMESIZE + 1];
1822 	unsigned int ifindex;
1823 	unsigned int np, i;
1824 	FILE *bond_file = NULL, *file;
1825 	int pf = -1;
1826 	int ret;
1827 	uint8_t cur_guid[32] = {0};
1828 	uint8_t guid[32] = {0};
1829 
1830 	/*
1831 	 * Try to get master device name. If something goes wrong suppose
1832 	 * the lack of kernel support and no bonding devices.
1833 	 */
1834 	memset(bond_info, 0, sizeof(*bond_info));
1835 	if (nl_rdma < 0)
1836 		return -1;
1837 	if (!strstr(ibdev_name, "bond"))
1838 		return -1;
1839 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1840 	if (!np)
1841 		return -1;
1842 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1843 		return -1;
1844 	/*
1845 	 * The master device might not be on the predefined port(not on port
1846 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1847 	 * device ports and find master.
1848 	 */
1849 	for (i = 1; i <= np; ++i) {
1850 		/* Check whether Infiniband port is populated. */
1851 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1852 		if (!ifindex)
1853 			continue;
1854 		if (!if_indextoname(ifindex, ifname))
1855 			continue;
1856 		/* Try to read bonding slave names from sysfs. */
1857 		MKSTR(slaves,
1858 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1859 		bond_file = fopen(slaves, "r");
1860 		if (bond_file)
1861 			break;
1862 	}
1863 	if (!bond_file)
1864 		return -1;
1865 	/* Use safe format to check maximal buffer length. */
1866 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1867 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1868 		char tmp_str[IF_NAMESIZE + 32];
1869 		struct rte_pci_addr pci_addr;
1870 		struct mlx5_switch_info	info;
1871 		int ret;
1872 
1873 		/* Process slave interface names in the loop. */
1874 		snprintf(tmp_str, sizeof(tmp_str),
1875 			 "/sys/class/net/%s", ifname);
1876 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1877 			DRV_LOG(WARNING,
1878 				"Cannot get PCI address for netdev \"%s\".",
1879 				ifname);
1880 			continue;
1881 		}
1882 		/* Slave interface PCI address match found. */
1883 		snprintf(tmp_str, sizeof(tmp_str),
1884 			 "/sys/class/net/%s/phys_port_name", ifname);
1885 		file = fopen(tmp_str, "rb");
1886 		if (!file)
1887 			break;
1888 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1889 		if (fscanf(file, "%32s", tmp_str) == 1)
1890 			mlx5_translate_port_name(tmp_str, &info);
1891 		fclose(file);
1892 		/* Only process PF ports. */
1893 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1894 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1895 			continue;
1896 		/* Check max bonding member. */
1897 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1898 			DRV_LOG(WARNING, "bonding index out of range, "
1899 				"please increase MLX5_BOND_MAX_PORTS: %s",
1900 				tmp_str);
1901 			break;
1902 		}
1903 		/* Get ifindex. */
1904 		snprintf(tmp_str, sizeof(tmp_str),
1905 			 "/sys/class/net/%s/ifindex", ifname);
1906 		file = fopen(tmp_str, "rb");
1907 		if (!file)
1908 			break;
1909 		ret = fscanf(file, "%u", &ifindex);
1910 		fclose(file);
1911 		if (ret != 1)
1912 			break;
1913 		/* Save bonding info. */
1914 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1915 			sizeof(bond_info->ports[0].ifname));
1916 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1917 		bond_info->ports[info.port_name].ifindex = ifindex;
1918 		bond_info->n_port++;
1919 		/*
1920 		 * Under socket direct mode, bonding will use
1921 		 * system_image_guid as identification.
1922 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1923 		 * All bonding members should have the same guid even if driver
1924 		 * is using PCIe BDF.
1925 		 */
1926 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1927 		if (ret < 0)
1928 			break;
1929 		else if (ret > 0) {
1930 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1931 			    owner == info.port_name &&
1932 			    (owner != 0 || (owner == 0 &&
1933 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1934 				pf = info.port_name;
1935 		} else if (pci_dev->domain == pci_addr.domain &&
1936 		    pci_dev->bus == pci_addr.bus &&
1937 		    pci_dev->devid == pci_addr.devid &&
1938 		    ((pci_dev->function == 0 &&
1939 		      pci_dev->function + owner == pci_addr.function) ||
1940 		     (pci_dev->function == owner &&
1941 		      pci_addr.function == owner)))
1942 			pf = info.port_name;
1943 	}
1944 	if (pf >= 0) {
1945 		/* Get bond interface info */
1946 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1947 					   bond_info->ifname);
1948 		if (ret)
1949 			DRV_LOG(ERR, "unable to get bond info: %s",
1950 				strerror(rte_errno));
1951 		else
1952 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1953 				ifindex, bond_info->ifindex, bond_info->ifname);
1954 	}
1955 	if (owner == 0 && pf != 0) {
1956 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1957 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1958 				pci_dev->function);
1959 	}
1960 	return pf;
1961 }
1962 
1963 static void
1964 mlx5_os_config_default(struct mlx5_dev_config *config,
1965 		       struct mlx5_common_dev_config *cconf)
1966 {
1967 	memset(config, 0, sizeof(*config));
1968 	config->mps = MLX5_ARG_UNSET;
1969 	config->cqe_comp = 1;
1970 	config->rx_vec_en = 1;
1971 	config->txq_inline_max = MLX5_ARG_UNSET;
1972 	config->txq_inline_min = MLX5_ARG_UNSET;
1973 	config->txq_inline_mpw = MLX5_ARG_UNSET;
1974 	config->txqs_inline = MLX5_ARG_UNSET;
1975 	config->vf_nl_en = 1;
1976 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1977 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1978 	config->mprq.log_min_stride_wqe_size = cconf->devx ?
1979 					cconf->hca_attr.log_min_stride_wqe_sz :
1980 					MLX5_MPRQ_LOG_MIN_STRIDE_WQE_SIZE;
1981 	config->mprq.log_stride_num = MLX5_MPRQ_DEFAULT_LOG_STRIDE_NUM;
1982 	config->dv_esw_en = 1;
1983 	config->dv_flow_en = 1;
1984 	config->decap_en = 1;
1985 	config->log_hp_size = MLX5_ARG_UNSET;
1986 	config->allow_duplicate_pattern = 1;
1987 	config->std_delay_drop = 0;
1988 	config->hp_delay_drop = 0;
1989 }
1990 
1991 /**
1992  * Register a PCI device within bonding.
1993  *
1994  * This function spawns Ethernet devices out of a given PCI device and
1995  * bonding owner PF index.
1996  *
1997  * @param[in] cdev
1998  *   Pointer to common mlx5 device structure.
1999  * @param[in] req_eth_da
2000  *   Requested ethdev device argument.
2001  * @param[in] owner_id
2002  *   Requested owner PF port ID within bonding device, default to 0.
2003  *
2004  * @return
2005  *   0 on success, a negative errno value otherwise and rte_errno is set.
2006  */
2007 static int
2008 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
2009 		     struct rte_eth_devargs *req_eth_da,
2010 		     uint16_t owner_id)
2011 {
2012 	struct ibv_device **ibv_list;
2013 	/*
2014 	 * Number of found IB Devices matching with requested PCI BDF.
2015 	 * nd != 1 means there are multiple IB devices over the same
2016 	 * PCI device and we have representors and master.
2017 	 */
2018 	unsigned int nd = 0;
2019 	/*
2020 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2021 	 * we have the single multiport IB device, and there may be
2022 	 * representors attached to some of found ports.
2023 	 */
2024 	unsigned int np = 0;
2025 	/*
2026 	 * Number of DPDK ethernet devices to Spawn - either over
2027 	 * multiple IB devices or multiple ports of single IB device.
2028 	 * Actually this is the number of iterations to spawn.
2029 	 */
2030 	unsigned int ns = 0;
2031 	/*
2032 	 * Bonding device
2033 	 *   < 0 - no bonding device (single one)
2034 	 *  >= 0 - bonding device (value is slave PF index)
2035 	 */
2036 	int bd = -1;
2037 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2038 	struct mlx5_dev_spawn_data *list = NULL;
2039 	struct mlx5_dev_config dev_config;
2040 	struct rte_eth_devargs eth_da = *req_eth_da;
2041 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2042 	struct mlx5_bond_info bond_info;
2043 	int ret = -1;
2044 
2045 	errno = 0;
2046 	ibv_list = mlx5_glue->get_device_list(&ret);
2047 	if (!ibv_list) {
2048 		rte_errno = errno ? errno : ENOSYS;
2049 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2050 		return -rte_errno;
2051 	}
2052 	/*
2053 	 * First scan the list of all Infiniband devices to find
2054 	 * matching ones, gathering into the list.
2055 	 */
2056 	struct ibv_device *ibv_match[ret + 1];
2057 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2058 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2059 	unsigned int i;
2060 
2061 	while (ret-- > 0) {
2062 		struct rte_pci_addr pci_addr;
2063 
2064 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2065 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2066 						nl_rdma, owner_id, &bond_info);
2067 		if (bd >= 0) {
2068 			/*
2069 			 * Bonding device detected. Only one match is allowed,
2070 			 * the bonding is supported over multi-port IB device,
2071 			 * there should be no matches on representor PCI
2072 			 * functions or non VF LAG bonding devices with
2073 			 * specified address.
2074 			 */
2075 			if (nd) {
2076 				DRV_LOG(ERR,
2077 					"multiple PCI match on bonding device"
2078 					"\"%s\" found", ibv_list[ret]->name);
2079 				rte_errno = ENOENT;
2080 				ret = -rte_errno;
2081 				goto exit;
2082 			}
2083 			/* Amend owner pci address if owner PF ID specified. */
2084 			if (eth_da.nb_representor_ports)
2085 				owner_pci.function += owner_id;
2086 			DRV_LOG(INFO,
2087 				"PCI information matches for slave %d bonding device \"%s\"",
2088 				bd, ibv_list[ret]->name);
2089 			ibv_match[nd++] = ibv_list[ret];
2090 			break;
2091 		} else {
2092 			/* Bonding device not found. */
2093 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2094 					      &pci_addr))
2095 				continue;
2096 			if (owner_pci.domain != pci_addr.domain ||
2097 			    owner_pci.bus != pci_addr.bus ||
2098 			    owner_pci.devid != pci_addr.devid ||
2099 			    owner_pci.function != pci_addr.function)
2100 				continue;
2101 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2102 				ibv_list[ret]->name);
2103 			ibv_match[nd++] = ibv_list[ret];
2104 		}
2105 	}
2106 	ibv_match[nd] = NULL;
2107 	if (!nd) {
2108 		/* No device matches, just complain and bail out. */
2109 		DRV_LOG(WARNING,
2110 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
2111 			" are kernel drivers loaded?",
2112 			owner_pci.domain, owner_pci.bus,
2113 			owner_pci.devid, owner_pci.function);
2114 		rte_errno = ENOENT;
2115 		ret = -rte_errno;
2116 		goto exit;
2117 	}
2118 	if (nd == 1) {
2119 		/*
2120 		 * Found single matching device may have multiple ports.
2121 		 * Each port may be representor, we have to check the port
2122 		 * number and check the representors existence.
2123 		 */
2124 		if (nl_rdma >= 0)
2125 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2126 		if (!np)
2127 			DRV_LOG(WARNING,
2128 				"Cannot get IB device \"%s\" ports number.",
2129 				ibv_match[0]->name);
2130 		if (bd >= 0 && !np) {
2131 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2132 			rte_errno = ENOENT;
2133 			ret = -rte_errno;
2134 			goto exit;
2135 		}
2136 	}
2137 	/* Now we can determine the maximal amount of devices to be spawned. */
2138 	list = mlx5_malloc(MLX5_MEM_ZERO,
2139 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2140 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2141 	if (!list) {
2142 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2143 		rte_errno = ENOMEM;
2144 		ret = -rte_errno;
2145 		goto exit;
2146 	}
2147 	if (bd >= 0 || np > 1) {
2148 		/*
2149 		 * Single IB device with multiple ports found,
2150 		 * it may be E-Switch master device and representors.
2151 		 * We have to perform identification through the ports.
2152 		 */
2153 		MLX5_ASSERT(nl_rdma >= 0);
2154 		MLX5_ASSERT(ns == 0);
2155 		MLX5_ASSERT(nd == 1);
2156 		MLX5_ASSERT(np);
2157 		for (i = 1; i <= np; ++i) {
2158 			list[ns].bond_info = &bond_info;
2159 			list[ns].max_port = np;
2160 			list[ns].phys_port = i;
2161 			list[ns].phys_dev_name = ibv_match[0]->name;
2162 			list[ns].eth_dev = NULL;
2163 			list[ns].pci_dev = pci_dev;
2164 			list[ns].cdev = cdev;
2165 			list[ns].pf_bond = bd;
2166 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2167 							   ibv_match[0]->name,
2168 							   i);
2169 			if (!list[ns].ifindex) {
2170 				/*
2171 				 * No network interface index found for the
2172 				 * specified port, it means there is no
2173 				 * representor on this port. It's OK,
2174 				 * there can be disabled ports, for example
2175 				 * if sriov_numvfs < sriov_totalvfs.
2176 				 */
2177 				continue;
2178 			}
2179 			ret = -1;
2180 			if (nl_route >= 0)
2181 				ret = mlx5_nl_switch_info(nl_route,
2182 							  list[ns].ifindex,
2183 							  &list[ns].info);
2184 			if (ret || (!list[ns].info.representor &&
2185 				    !list[ns].info.master)) {
2186 				/*
2187 				 * We failed to recognize representors with
2188 				 * Netlink, let's try to perform the task
2189 				 * with sysfs.
2190 				 */
2191 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2192 							     &list[ns].info);
2193 			}
2194 			if (!ret && bd >= 0) {
2195 				switch (list[ns].info.name_type) {
2196 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2197 					if (np == 1) {
2198 						/*
2199 						 * Force standalone bonding
2200 						 * device for ROCE LAG
2201 						 * configurations.
2202 						 */
2203 						list[ns].info.master = 0;
2204 						list[ns].info.representor = 0;
2205 					}
2206 					if (list[ns].info.port_name == bd)
2207 						ns++;
2208 					break;
2209 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2210 					/* Fallthrough */
2211 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2212 					/* Fallthrough */
2213 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2214 					if (list[ns].info.pf_num == bd)
2215 						ns++;
2216 					break;
2217 				default:
2218 					break;
2219 				}
2220 				continue;
2221 			}
2222 			if (!ret && (list[ns].info.representor ^
2223 				     list[ns].info.master))
2224 				ns++;
2225 		}
2226 		if (!ns) {
2227 			DRV_LOG(ERR,
2228 				"Unable to recognize master/representors on the IB device with multiple ports.");
2229 			rte_errno = ENOENT;
2230 			ret = -rte_errno;
2231 			goto exit;
2232 		}
2233 	} else {
2234 		/*
2235 		 * The existence of several matching entries (nd > 1) means
2236 		 * port representors have been instantiated. No existing Verbs
2237 		 * call nor sysfs entries can tell them apart, this can only
2238 		 * be done through Netlink calls assuming kernel drivers are
2239 		 * recent enough to support them.
2240 		 *
2241 		 * In the event of identification failure through Netlink,
2242 		 * try again through sysfs, then:
2243 		 *
2244 		 * 1. A single IB device matches (nd == 1) with single
2245 		 *    port (np=0/1) and is not a representor, assume
2246 		 *    no switch support.
2247 		 *
2248 		 * 2. Otherwise no safe assumptions can be made;
2249 		 *    complain louder and bail out.
2250 		 */
2251 		for (i = 0; i != nd; ++i) {
2252 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2253 			list[ns].bond_info = NULL;
2254 			list[ns].max_port = 1;
2255 			list[ns].phys_port = 1;
2256 			list[ns].phys_dev_name = ibv_match[i]->name;
2257 			list[ns].eth_dev = NULL;
2258 			list[ns].pci_dev = pci_dev;
2259 			list[ns].cdev = cdev;
2260 			list[ns].pf_bond = -1;
2261 			list[ns].ifindex = 0;
2262 			if (nl_rdma >= 0)
2263 				list[ns].ifindex = mlx5_nl_ifindex
2264 							    (nl_rdma,
2265 							     ibv_match[i]->name,
2266 							     1);
2267 			if (!list[ns].ifindex) {
2268 				char ifname[IF_NAMESIZE];
2269 
2270 				/*
2271 				 * Netlink failed, it may happen with old
2272 				 * ib_core kernel driver (before 4.16).
2273 				 * We can assume there is old driver because
2274 				 * here we are processing single ports IB
2275 				 * devices. Let's try sysfs to retrieve
2276 				 * the ifindex. The method works for
2277 				 * master device only.
2278 				 */
2279 				if (nd > 1) {
2280 					/*
2281 					 * Multiple devices found, assume
2282 					 * representors, can not distinguish
2283 					 * master/representor and retrieve
2284 					 * ifindex via sysfs.
2285 					 */
2286 					continue;
2287 				}
2288 				ret = mlx5_get_ifname_sysfs
2289 					(ibv_match[i]->ibdev_path, ifname);
2290 				if (!ret)
2291 					list[ns].ifindex =
2292 						if_nametoindex(ifname);
2293 				if (!list[ns].ifindex) {
2294 					/*
2295 					 * No network interface index found
2296 					 * for the specified device, it means
2297 					 * there it is neither representor
2298 					 * nor master.
2299 					 */
2300 					continue;
2301 				}
2302 			}
2303 			ret = -1;
2304 			if (nl_route >= 0)
2305 				ret = mlx5_nl_switch_info(nl_route,
2306 							  list[ns].ifindex,
2307 							  &list[ns].info);
2308 			if (ret || (!list[ns].info.representor &&
2309 				    !list[ns].info.master)) {
2310 				/*
2311 				 * We failed to recognize representors with
2312 				 * Netlink, let's try to perform the task
2313 				 * with sysfs.
2314 				 */
2315 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2316 							     &list[ns].info);
2317 			}
2318 			if (!ret && (list[ns].info.representor ^
2319 				     list[ns].info.master)) {
2320 				ns++;
2321 			} else if ((nd == 1) &&
2322 				   !list[ns].info.representor &&
2323 				   !list[ns].info.master) {
2324 				/*
2325 				 * Single IB device with one physical port and
2326 				 * attached network device.
2327 				 * May be SRIOV is not enabled or there is no
2328 				 * representors.
2329 				 */
2330 				DRV_LOG(INFO, "No E-Switch support detected.");
2331 				ns++;
2332 				break;
2333 			}
2334 		}
2335 		if (!ns) {
2336 			DRV_LOG(ERR,
2337 				"Unable to recognize master/representors on the multiple IB devices.");
2338 			rte_errno = ENOENT;
2339 			ret = -rte_errno;
2340 			goto exit;
2341 		}
2342 		/*
2343 		 * New kernels may add the switch_id attribute for the case
2344 		 * there is no E-Switch and we wrongly recognized the only
2345 		 * device as master. Override this if there is the single
2346 		 * device with single port and new device name format present.
2347 		 */
2348 		if (nd == 1 &&
2349 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2350 			list[0].info.master = 0;
2351 			list[0].info.representor = 0;
2352 		}
2353 	}
2354 	MLX5_ASSERT(ns);
2355 	/*
2356 	 * Sort list to probe devices in natural order for users convenience
2357 	 * (i.e. master first, then representors from lowest to highest ID).
2358 	 */
2359 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2360 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2361 		/* Set devargs default values. */
2362 		if (eth_da.nb_mh_controllers == 0) {
2363 			eth_da.nb_mh_controllers = 1;
2364 			eth_da.mh_controllers[0] = 0;
2365 		}
2366 		if (eth_da.nb_ports == 0 && ns > 0) {
2367 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2368 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2369 					pci_dev->device.devargs->args);
2370 			eth_da.nb_ports = 1;
2371 			eth_da.ports[0] = list[0].info.pf_num;
2372 		}
2373 		if (eth_da.nb_representor_ports == 0) {
2374 			eth_da.nb_representor_ports = 1;
2375 			eth_da.representor_ports[0] = 0;
2376 		}
2377 	}
2378 	for (i = 0; i != ns; ++i) {
2379 		uint32_t restore;
2380 
2381 		/* Default configuration. */
2382 		mlx5_os_config_default(&dev_config, &cdev->config);
2383 		dev_config.vf = mlx5_dev_is_vf_pci(pci_dev);
2384 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i],
2385 						 &dev_config, &eth_da);
2386 		if (!list[i].eth_dev) {
2387 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2388 				break;
2389 			/* Device is disabled or already spawned. Ignore it. */
2390 			continue;
2391 		}
2392 		restore = list[i].eth_dev->data->dev_flags;
2393 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2394 		/**
2395 		 * Each representor has a dedicated interrupts vector.
2396 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2397 		 * representor eth_dev object because representor and PF
2398 		 * share the same PCI address.
2399 		 * Override representor device with a dedicated
2400 		 * interrupts handle here.
2401 		 * Representor interrupts handle is released in mlx5_dev_stop().
2402 		 */
2403 		if (list[i].info.representor) {
2404 			struct rte_intr_handle *intr_handle =
2405 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2406 			if (intr_handle == NULL) {
2407 				DRV_LOG(ERR,
2408 					"port %u failed to allocate memory for interrupt handler "
2409 					"Rx interrupts will not be supported",
2410 					i);
2411 				rte_errno = ENOMEM;
2412 				ret = -rte_errno;
2413 				goto exit;
2414 			}
2415 			list[i].eth_dev->intr_handle = intr_handle;
2416 		}
2417 		/* Restore non-PCI flags cleared by the above call. */
2418 		list[i].eth_dev->data->dev_flags |= restore;
2419 		rte_eth_dev_probing_finish(list[i].eth_dev);
2420 	}
2421 	if (i != ns) {
2422 		DRV_LOG(ERR,
2423 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2424 			" encountering an error: %s",
2425 			owner_pci.domain, owner_pci.bus,
2426 			owner_pci.devid, owner_pci.function,
2427 			strerror(rte_errno));
2428 		ret = -rte_errno;
2429 		/* Roll back. */
2430 		while (i--) {
2431 			if (!list[i].eth_dev)
2432 				continue;
2433 			mlx5_dev_close(list[i].eth_dev);
2434 			/* mac_addrs must not be freed because in dev_private */
2435 			list[i].eth_dev->data->mac_addrs = NULL;
2436 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2437 		}
2438 		/* Restore original error. */
2439 		rte_errno = -ret;
2440 	} else {
2441 		ret = 0;
2442 	}
2443 exit:
2444 	/*
2445 	 * Do the routine cleanup:
2446 	 * - close opened Netlink sockets
2447 	 * - free allocated spawn data array
2448 	 * - free the Infiniband device list
2449 	 */
2450 	if (nl_rdma >= 0)
2451 		close(nl_rdma);
2452 	if (nl_route >= 0)
2453 		close(nl_route);
2454 	if (list)
2455 		mlx5_free(list);
2456 	MLX5_ASSERT(ibv_list);
2457 	mlx5_glue->free_device_list(ibv_list);
2458 	return ret;
2459 }
2460 
2461 static int
2462 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2463 			  struct rte_eth_devargs *eth_da)
2464 {
2465 	int ret = 0;
2466 
2467 	if (dev->devargs == NULL)
2468 		return 0;
2469 	memset(eth_da, 0, sizeof(*eth_da));
2470 	/* Parse representor information first from class argument. */
2471 	if (dev->devargs->cls_str)
2472 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2473 	if (ret != 0) {
2474 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2475 			dev->devargs->cls_str);
2476 		return -rte_errno;
2477 	}
2478 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2479 		/* Parse legacy device argument */
2480 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2481 		if (ret) {
2482 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2483 				dev->devargs->args);
2484 			return -rte_errno;
2485 		}
2486 	}
2487 	return 0;
2488 }
2489 
2490 /**
2491  * Callback to register a PCI device.
2492  *
2493  * This function spawns Ethernet devices out of a given PCI device.
2494  *
2495  * @param[in] cdev
2496  *   Pointer to common mlx5 device structure.
2497  *
2498  * @return
2499  *   0 on success, a negative errno value otherwise and rte_errno is set.
2500  */
2501 static int
2502 mlx5_os_pci_probe(struct mlx5_common_device *cdev)
2503 {
2504 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2505 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2506 	int ret = 0;
2507 	uint16_t p;
2508 
2509 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2510 	if (ret != 0)
2511 		return ret;
2512 
2513 	if (eth_da.nb_ports > 0) {
2514 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2515 		for (p = 0; p < eth_da.nb_ports; p++) {
2516 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2517 						   eth_da.ports[p]);
2518 			if (ret)
2519 				break;
2520 		}
2521 		if (ret) {
2522 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2523 				"aborted due to prodding failure of PF %u",
2524 				pci_dev->addr.domain, pci_dev->addr.bus,
2525 				pci_dev->addr.devid, pci_dev->addr.function,
2526 				eth_da.ports[p]);
2527 			mlx5_net_remove(cdev);
2528 		}
2529 	} else {
2530 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0);
2531 	}
2532 	return ret;
2533 }
2534 
2535 /* Probe a single SF device on auxiliary bus, no representor support. */
2536 static int
2537 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev)
2538 {
2539 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2540 	struct mlx5_dev_config config;
2541 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2542 	struct rte_device *dev = cdev->dev;
2543 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2544 	struct rte_eth_dev *eth_dev;
2545 	int ret = 0;
2546 
2547 	/* Parse ethdev devargs. */
2548 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2549 	if (ret != 0)
2550 		return ret;
2551 	/* Set default config data. */
2552 	mlx5_os_config_default(&config, &cdev->config);
2553 	config.sf = 1;
2554 	/* Init spawn data. */
2555 	spawn.max_port = 1;
2556 	spawn.phys_port = 1;
2557 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2558 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2559 	if (ret < 0) {
2560 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2561 		return ret;
2562 	}
2563 	spawn.ifindex = ret;
2564 	spawn.cdev = cdev;
2565 	/* Spawn device. */
2566 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2567 	if (eth_dev == NULL)
2568 		return -rte_errno;
2569 	/* Post create. */
2570 	eth_dev->intr_handle = adev->intr_handle;
2571 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2572 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2573 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2574 		eth_dev->data->numa_node = dev->numa_node;
2575 	}
2576 	rte_eth_dev_probing_finish(eth_dev);
2577 	return 0;
2578 }
2579 
2580 /**
2581  * Net class driver callback to probe a device.
2582  *
2583  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2584  *
2585  * @param[in] cdev
2586  *   Pointer to the common mlx5 device.
2587  *
2588  * @return
2589  *   0 on success, a negative errno value otherwise and rte_errno is set.
2590  */
2591 int
2592 mlx5_os_net_probe(struct mlx5_common_device *cdev)
2593 {
2594 	int ret;
2595 
2596 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2597 		mlx5_pmd_socket_init();
2598 	ret = mlx5_init_once();
2599 	if (ret) {
2600 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2601 			strerror(rte_errno));
2602 		return -rte_errno;
2603 	}
2604 	if (mlx5_dev_is_pci(cdev->dev))
2605 		return mlx5_os_pci_probe(cdev);
2606 	else
2607 		return mlx5_os_auxiliary_probe(cdev);
2608 }
2609 
2610 /**
2611  * Cleanup resources when the last device is closed.
2612  */
2613 void
2614 mlx5_os_net_cleanup(void)
2615 {
2616 	mlx5_pmd_socket_uninit();
2617 }
2618 
2619 /**
2620  * Install shared asynchronous device events handler.
2621  * This function is implemented to support event sharing
2622  * between multiple ports of single IB device.
2623  *
2624  * @param sh
2625  *   Pointer to mlx5_dev_ctx_shared object.
2626  */
2627 void
2628 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2629 {
2630 	int ret;
2631 	int flags;
2632 	struct ibv_context *ctx = sh->cdev->ctx;
2633 
2634 	sh->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2635 	if (sh->intr_handle == NULL) {
2636 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2637 		rte_errno = ENOMEM;
2638 		return;
2639 	}
2640 	rte_intr_fd_set(sh->intr_handle, -1);
2641 
2642 	flags = fcntl(ctx->async_fd, F_GETFL);
2643 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2644 	if (ret) {
2645 		DRV_LOG(INFO, "failed to change file descriptor async event"
2646 			" queue");
2647 	} else {
2648 		rte_intr_fd_set(sh->intr_handle, ctx->async_fd);
2649 		rte_intr_type_set(sh->intr_handle, RTE_INTR_HANDLE_EXT);
2650 		if (rte_intr_callback_register(sh->intr_handle,
2651 					mlx5_dev_interrupt_handler, sh)) {
2652 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2653 			rte_intr_fd_set(sh->intr_handle, -1);
2654 		}
2655 	}
2656 	if (sh->cdev->config.devx) {
2657 #ifdef HAVE_IBV_DEVX_ASYNC
2658 		sh->intr_handle_devx =
2659 			rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2660 		if (!sh->intr_handle_devx) {
2661 			DRV_LOG(ERR, "Fail to allocate intr_handle");
2662 			rte_errno = ENOMEM;
2663 			return;
2664 		}
2665 		rte_intr_fd_set(sh->intr_handle_devx, -1);
2666 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2667 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2668 		if (!devx_comp) {
2669 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2670 			return;
2671 		}
2672 		flags = fcntl(devx_comp->fd, F_GETFL);
2673 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2674 		if (ret) {
2675 			DRV_LOG(INFO, "failed to change file descriptor"
2676 				" devx comp");
2677 			return;
2678 		}
2679 		rte_intr_fd_set(sh->intr_handle_devx, devx_comp->fd);
2680 		rte_intr_type_set(sh->intr_handle_devx,
2681 					 RTE_INTR_HANDLE_EXT);
2682 		if (rte_intr_callback_register(sh->intr_handle_devx,
2683 					mlx5_dev_interrupt_handler_devx, sh)) {
2684 			DRV_LOG(INFO, "Fail to install the devx shared"
2685 				" interrupt.");
2686 			rte_intr_fd_set(sh->intr_handle_devx, -1);
2687 		}
2688 #endif /* HAVE_IBV_DEVX_ASYNC */
2689 	}
2690 }
2691 
2692 /**
2693  * Uninstall shared asynchronous device events handler.
2694  * This function is implemented to support event sharing
2695  * between multiple ports of single IB device.
2696  *
2697  * @param dev
2698  *   Pointer to mlx5_dev_ctx_shared object.
2699  */
2700 void
2701 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2702 {
2703 	if (rte_intr_fd_get(sh->intr_handle) >= 0)
2704 		mlx5_intr_callback_unregister(sh->intr_handle,
2705 					      mlx5_dev_interrupt_handler, sh);
2706 	rte_intr_instance_free(sh->intr_handle);
2707 #ifdef HAVE_IBV_DEVX_ASYNC
2708 	if (rte_intr_fd_get(sh->intr_handle_devx) >= 0)
2709 		rte_intr_callback_unregister(sh->intr_handle_devx,
2710 				  mlx5_dev_interrupt_handler_devx, sh);
2711 	rte_intr_instance_free(sh->intr_handle_devx);
2712 	if (sh->devx_comp)
2713 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2714 #endif
2715 }
2716 
2717 /**
2718  * Read statistics by a named counter.
2719  *
2720  * @param[in] priv
2721  *   Pointer to the private device data structure.
2722  * @param[in] ctr_name
2723  *   Pointer to the name of the statistic counter to read
2724  * @param[out] stat
2725  *   Pointer to read statistic value.
2726  * @return
2727  *   0 on success and stat is valud, 1 if failed to read the value
2728  *   rte_errno is set.
2729  *
2730  */
2731 int
2732 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2733 		      uint64_t *stat)
2734 {
2735 	int fd;
2736 
2737 	if (priv->sh) {
2738 		if (priv->q_counters != NULL &&
2739 		    strcmp(ctr_name, "out_of_buffer") == 0)
2740 			return mlx5_devx_cmd_queue_counter_query
2741 					(priv->q_counters, 0, (uint32_t *)stat);
2742 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2743 		      priv->sh->ibdev_path,
2744 		      priv->dev_port,
2745 		      ctr_name);
2746 		fd = open(path, O_RDONLY);
2747 		/*
2748 		 * in switchdev the file location is not per port
2749 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2750 		 */
2751 		if (fd == -1) {
2752 			MKSTR(path1, "%s/hw_counters/%s",
2753 			      priv->sh->ibdev_path,
2754 			      ctr_name);
2755 			fd = open(path1, O_RDONLY);
2756 		}
2757 		if (fd != -1) {
2758 			char buf[21] = {'\0'};
2759 			ssize_t n = read(fd, buf, sizeof(buf));
2760 
2761 			close(fd);
2762 			if (n != -1) {
2763 				*stat = strtoull(buf, NULL, 10);
2764 				return 0;
2765 			}
2766 		}
2767 	}
2768 	*stat = 0;
2769 	return 1;
2770 }
2771 
2772 /**
2773  * Remove a MAC address from device
2774  *
2775  * @param dev
2776  *   Pointer to Ethernet device structure.
2777  * @param index
2778  *   MAC address index.
2779  */
2780 void
2781 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2782 {
2783 	struct mlx5_priv *priv = dev->data->dev_private;
2784 	const int vf = priv->config.vf;
2785 
2786 	if (vf)
2787 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2788 					mlx5_ifindex(dev), priv->mac_own,
2789 					&dev->data->mac_addrs[index], index);
2790 }
2791 
2792 /**
2793  * Adds a MAC address to the device
2794  *
2795  * @param dev
2796  *   Pointer to Ethernet device structure.
2797  * @param mac_addr
2798  *   MAC address to register.
2799  * @param index
2800  *   MAC address index.
2801  *
2802  * @return
2803  *   0 on success, a negative errno value otherwise
2804  */
2805 int
2806 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2807 		     uint32_t index)
2808 {
2809 	struct mlx5_priv *priv = dev->data->dev_private;
2810 	const int vf = priv->config.vf;
2811 	int ret = 0;
2812 
2813 	if (vf)
2814 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2815 					   mlx5_ifindex(dev), priv->mac_own,
2816 					   mac, index);
2817 	return ret;
2818 }
2819 
2820 /**
2821  * Modify a VF MAC address
2822  *
2823  * @param priv
2824  *   Pointer to device private data.
2825  * @param mac_addr
2826  *   MAC address to modify into.
2827  * @param iface_idx
2828  *   Net device interface index
2829  * @param vf_index
2830  *   VF index
2831  *
2832  * @return
2833  *   0 on success, a negative errno value otherwise
2834  */
2835 int
2836 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2837 			   unsigned int iface_idx,
2838 			   struct rte_ether_addr *mac_addr,
2839 			   int vf_index)
2840 {
2841 	return mlx5_nl_vf_mac_addr_modify
2842 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2843 }
2844 
2845 /**
2846  * Set device promiscuous mode
2847  *
2848  * @param dev
2849  *   Pointer to Ethernet device structure.
2850  * @param enable
2851  *   0 - promiscuous is disabled, otherwise - enabled
2852  *
2853  * @return
2854  *   0 on success, a negative error value otherwise
2855  */
2856 int
2857 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2858 {
2859 	struct mlx5_priv *priv = dev->data->dev_private;
2860 
2861 	return mlx5_nl_promisc(priv->nl_socket_route,
2862 			       mlx5_ifindex(dev), !!enable);
2863 }
2864 
2865 /**
2866  * Set device promiscuous mode
2867  *
2868  * @param dev
2869  *   Pointer to Ethernet device structure.
2870  * @param enable
2871  *   0 - all multicase is disabled, otherwise - enabled
2872  *
2873  * @return
2874  *   0 on success, a negative error value otherwise
2875  */
2876 int
2877 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2878 {
2879 	struct mlx5_priv *priv = dev->data->dev_private;
2880 
2881 	return mlx5_nl_allmulti(priv->nl_socket_route,
2882 				mlx5_ifindex(dev), !!enable);
2883 }
2884 
2885 /**
2886  * Flush device MAC addresses
2887  *
2888  * @param dev
2889  *   Pointer to Ethernet device structure.
2890  *
2891  */
2892 void
2893 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2894 {
2895 	struct mlx5_priv *priv = dev->data->dev_private;
2896 
2897 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2898 			       dev->data->mac_addrs,
2899 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2900 }
2901