xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 91d1cfafc977ccb224d4632b12952360934499f2)
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 sh
135  *   Pointer to shared device context.
136  *
137  * @return
138  *   0 on success, a negative errno value otherwise and rte_errno is set.
139  */
140 int
141 mlx5_os_capabilities_prepare(struct mlx5_dev_ctx_shared *sh)
142 {
143 	int err;
144 	struct ibv_context *ctx = sh->cdev->ctx;
145 	struct ibv_device_attr_ex attr_ex = { .comp_mask = 0 };
146 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
147 
148 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
149 	if (err) {
150 		rte_errno = errno;
151 		return -rte_errno;
152 	}
153 #ifdef HAVE_IBV_MLX5_MOD_SWP
154 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
155 #endif
156 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
157 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
158 #endif
159 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
160 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
161 #endif
162 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
163 	if (err) {
164 		rte_errno = errno;
165 		return -rte_errno;
166 	}
167 	memset(&sh->dev_cap, 0, sizeof(struct mlx5_dev_cap));
168 	sh->dev_cap.device_cap_flags_ex = attr_ex.device_cap_flags_ex;
169 	sh->dev_cap.max_qp_wr = attr_ex.orig_attr.max_qp_wr;
170 	sh->dev_cap.max_sge = attr_ex.orig_attr.max_sge;
171 	sh->dev_cap.max_cq = attr_ex.orig_attr.max_cq;
172 	sh->dev_cap.max_qp = attr_ex.orig_attr.max_qp;
173 	sh->dev_cap.raw_packet_caps = attr_ex.raw_packet_caps;
174 	sh->dev_cap.max_rwq_indirection_table_size =
175 		attr_ex.rss_caps.max_rwq_indirection_table_size;
176 	sh->dev_cap.max_tso = attr_ex.tso_caps.max_tso;
177 	sh->dev_cap.tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
178 	strlcpy(sh->dev_cap.fw_ver, attr_ex.orig_attr.fw_ver,
179 		sizeof(sh->dev_cap.fw_ver));
180 	sh->dev_cap.flags = dv_attr.flags;
181 	sh->dev_cap.comp_mask = dv_attr.comp_mask;
182 #ifdef HAVE_IBV_MLX5_MOD_SWP
183 	sh->dev_cap.sw_parsing_offloads =
184 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
185 #endif
186 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
187 	sh->dev_cap.min_single_stride_log_num_of_bytes =
188 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
189 	sh->dev_cap.max_single_stride_log_num_of_bytes =
190 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
191 	sh->dev_cap.min_single_wqe_log_num_of_strides =
192 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
193 	sh->dev_cap.max_single_wqe_log_num_of_strides =
194 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
195 	sh->dev_cap.stride_supported_qpts =
196 		dv_attr.striding_rq_caps.supported_qpts;
197 #endif
198 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
199 	sh->dev_cap.tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
200 #endif
201 	return 0;
202 }
203 
204 /**
205  * Detect misc5 support or not
206  *
207  * @param[in] priv
208  *   Device private data pointer
209  */
210 #ifdef HAVE_MLX5DV_DR
211 static void
212 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
213 {
214 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
215 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
216 	 * Case: IPv4--->UDP--->VxLAN--->vni
217 	 */
218 	void *tbl;
219 	struct mlx5_flow_dv_match_params matcher_mask;
220 	void *match_m;
221 	void *matcher;
222 	void *headers_m;
223 	void *misc5_m;
224 	uint32_t *tunnel_header_m;
225 	struct mlx5dv_flow_matcher_attr dv_attr;
226 
227 	memset(&matcher_mask, 0, sizeof(matcher_mask));
228 	matcher_mask.size = sizeof(matcher_mask.buf);
229 	match_m = matcher_mask.buf;
230 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
231 	misc5_m = MLX5_ADDR_OF(fte_match_param,
232 			       match_m, misc_parameters_5);
233 	tunnel_header_m = (uint32_t *)
234 				MLX5_ADDR_OF(fte_match_set_misc5,
235 				misc5_m, tunnel_header_1);
236 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
237 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
238 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
239 	*tunnel_header_m = 0xffffff;
240 
241 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
242 	if (!tbl) {
243 		DRV_LOG(INFO, "No SW steering support");
244 		return;
245 	}
246 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
247 	dv_attr.match_mask = (void *)&matcher_mask,
248 	dv_attr.match_criteria_enable =
249 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
250 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
251 	dv_attr.priority = 3;
252 #ifdef HAVE_MLX5DV_DR_ESWITCH
253 	void *misc2_m;
254 	if (priv->config.dv_esw_en) {
255 		/* FDB enabled reg_c_0 */
256 		dv_attr.match_criteria_enable |=
257 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
258 		misc2_m = MLX5_ADDR_OF(fte_match_param,
259 				       match_m, misc_parameters_2);
260 		MLX5_SET(fte_match_set_misc2, misc2_m,
261 			 metadata_reg_c_0, 0xffff);
262 	}
263 #endif
264 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
265 						    &dv_attr, tbl);
266 	if (matcher) {
267 		priv->sh->misc5_cap = 1;
268 		mlx5_glue->dv_destroy_flow_matcher(matcher);
269 	}
270 	mlx5_glue->dr_destroy_flow_tbl(tbl);
271 #else
272 	RTE_SET_USED(priv);
273 #endif
274 }
275 #endif
276 
277 /**
278  * Initialize DR related data within private structure.
279  * Routine checks the reference counter and does actual
280  * resources creation/initialization only if counter is zero.
281  *
282  * @param[in] priv
283  *   Pointer to the private device data structure.
284  *
285  * @return
286  *   Zero on success, positive error code otherwise.
287  */
288 static int
289 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
290 {
291 	struct mlx5_dev_ctx_shared *sh = priv->sh;
292 	char s[MLX5_NAME_SIZE] __rte_unused;
293 	int err;
294 
295 	MLX5_ASSERT(sh && sh->refcnt);
296 	if (sh->refcnt > 1)
297 		return 0;
298 	err = mlx5_alloc_table_hash_list(priv);
299 	if (err)
300 		goto error;
301 	/* The resources below are only valid with DV support. */
302 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
303 	/* Init port id action list. */
304 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
305 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
306 						   flow_dv_port_id_create_cb,
307 						   flow_dv_port_id_match_cb,
308 						   flow_dv_port_id_remove_cb,
309 						   flow_dv_port_id_clone_cb,
310 						 flow_dv_port_id_clone_free_cb);
311 	if (!sh->port_id_action_list)
312 		goto error;
313 	/* Init push vlan action list. */
314 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
315 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
316 						    flow_dv_push_vlan_create_cb,
317 						    flow_dv_push_vlan_match_cb,
318 						    flow_dv_push_vlan_remove_cb,
319 						    flow_dv_push_vlan_clone_cb,
320 					       flow_dv_push_vlan_clone_free_cb);
321 	if (!sh->push_vlan_action_list)
322 		goto error;
323 	/* Init sample action list. */
324 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
325 	sh->sample_action_list = mlx5_list_create(s, sh, true,
326 						  flow_dv_sample_create_cb,
327 						  flow_dv_sample_match_cb,
328 						  flow_dv_sample_remove_cb,
329 						  flow_dv_sample_clone_cb,
330 						  flow_dv_sample_clone_free_cb);
331 	if (!sh->sample_action_list)
332 		goto error;
333 	/* Init dest array action list. */
334 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
335 	sh->dest_array_list = mlx5_list_create(s, sh, true,
336 					       flow_dv_dest_array_create_cb,
337 					       flow_dv_dest_array_match_cb,
338 					       flow_dv_dest_array_remove_cb,
339 					       flow_dv_dest_array_clone_cb,
340 					      flow_dv_dest_array_clone_free_cb);
341 	if (!sh->dest_array_list)
342 		goto error;
343 	/* Init shared flex parsers list, no need lcore_share */
344 	snprintf(s, sizeof(s), "%s_flex_parsers_list", sh->ibdev_name);
345 	sh->flex_parsers_dv = mlx5_list_create(s, sh, false,
346 					       mlx5_flex_parser_create_cb,
347 					       mlx5_flex_parser_match_cb,
348 					       mlx5_flex_parser_remove_cb,
349 					       mlx5_flex_parser_clone_cb,
350 					       mlx5_flex_parser_clone_free_cb);
351 	if (!sh->flex_parsers_dv)
352 		goto error;
353 #endif
354 #ifdef HAVE_MLX5DV_DR
355 	void *domain;
356 
357 	/* Reference counter is zero, we should initialize structures. */
358 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
359 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
360 	if (!domain) {
361 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
362 		err = errno;
363 		goto error;
364 	}
365 	sh->rx_domain = domain;
366 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
367 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
368 	if (!domain) {
369 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
370 		err = errno;
371 		goto error;
372 	}
373 	sh->tx_domain = domain;
374 #ifdef HAVE_MLX5DV_DR_ESWITCH
375 	if (priv->config.dv_esw_en) {
376 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
377 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
378 		if (!domain) {
379 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
380 			err = errno;
381 			goto error;
382 		}
383 		sh->fdb_domain = domain;
384 	}
385 	/*
386 	 * The drop action is just some dummy placeholder in rdma-core. It
387 	 * does not belong to domains and has no any attributes, and, can be
388 	 * shared by the entire device.
389 	 */
390 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
391 	if (!sh->dr_drop_action) {
392 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
393 		err = errno;
394 		goto error;
395 	}
396 #endif
397 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
398 		err = mlx5_alloc_tunnel_hub(sh);
399 	if (err) {
400 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
401 		goto error;
402 	}
403 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
404 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
405 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
406 		if (sh->fdb_domain)
407 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
408 	}
409 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
410 	if (!priv->config.allow_duplicate_pattern) {
411 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
412 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
413 #endif
414 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
415 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
416 		if (sh->fdb_domain)
417 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
418 	}
419 
420 	__mlx5_discovery_misc5_cap(priv);
421 #endif /* HAVE_MLX5DV_DR */
422 	sh->default_miss_action =
423 			mlx5_glue->dr_create_flow_action_default_miss();
424 	if (!sh->default_miss_action)
425 		DRV_LOG(WARNING, "Default miss action is not supported.");
426 	LIST_INIT(&sh->shared_rxqs);
427 	return 0;
428 error:
429 	/* Rollback the created objects. */
430 	if (sh->rx_domain) {
431 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
432 		sh->rx_domain = NULL;
433 	}
434 	if (sh->tx_domain) {
435 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
436 		sh->tx_domain = NULL;
437 	}
438 	if (sh->fdb_domain) {
439 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
440 		sh->fdb_domain = NULL;
441 	}
442 	if (sh->dr_drop_action) {
443 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
444 		sh->dr_drop_action = NULL;
445 	}
446 	if (sh->pop_vlan_action) {
447 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
448 		sh->pop_vlan_action = NULL;
449 	}
450 	if (sh->encaps_decaps) {
451 		mlx5_hlist_destroy(sh->encaps_decaps);
452 		sh->encaps_decaps = NULL;
453 	}
454 	if (sh->modify_cmds) {
455 		mlx5_hlist_destroy(sh->modify_cmds);
456 		sh->modify_cmds = NULL;
457 	}
458 	if (sh->tag_table) {
459 		/* tags should be destroyed with flow before. */
460 		mlx5_hlist_destroy(sh->tag_table);
461 		sh->tag_table = NULL;
462 	}
463 	if (sh->tunnel_hub) {
464 		mlx5_release_tunnel_hub(sh, priv->dev_port);
465 		sh->tunnel_hub = NULL;
466 	}
467 	mlx5_free_table_hash_list(priv);
468 	if (sh->port_id_action_list) {
469 		mlx5_list_destroy(sh->port_id_action_list);
470 		sh->port_id_action_list = NULL;
471 	}
472 	if (sh->push_vlan_action_list) {
473 		mlx5_list_destroy(sh->push_vlan_action_list);
474 		sh->push_vlan_action_list = NULL;
475 	}
476 	if (sh->sample_action_list) {
477 		mlx5_list_destroy(sh->sample_action_list);
478 		sh->sample_action_list = NULL;
479 	}
480 	if (sh->dest_array_list) {
481 		mlx5_list_destroy(sh->dest_array_list);
482 		sh->dest_array_list = NULL;
483 	}
484 	return err;
485 }
486 
487 /**
488  * Destroy DR related data within private structure.
489  *
490  * @param[in] priv
491  *   Pointer to the private device data structure.
492  */
493 void
494 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
495 {
496 	struct mlx5_dev_ctx_shared *sh = priv->sh;
497 
498 	MLX5_ASSERT(sh && sh->refcnt);
499 	if (sh->refcnt > 1)
500 		return;
501 	MLX5_ASSERT(LIST_EMPTY(&sh->shared_rxqs));
502 #ifdef HAVE_MLX5DV_DR
503 	if (sh->rx_domain) {
504 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
505 		sh->rx_domain = NULL;
506 	}
507 	if (sh->tx_domain) {
508 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
509 		sh->tx_domain = NULL;
510 	}
511 #ifdef HAVE_MLX5DV_DR_ESWITCH
512 	if (sh->fdb_domain) {
513 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
514 		sh->fdb_domain = NULL;
515 	}
516 	if (sh->dr_drop_action) {
517 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
518 		sh->dr_drop_action = NULL;
519 	}
520 #endif
521 	if (sh->pop_vlan_action) {
522 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
523 		sh->pop_vlan_action = NULL;
524 	}
525 #endif /* HAVE_MLX5DV_DR */
526 	if (sh->default_miss_action)
527 		mlx5_glue->destroy_flow_action
528 				(sh->default_miss_action);
529 	if (sh->encaps_decaps) {
530 		mlx5_hlist_destroy(sh->encaps_decaps);
531 		sh->encaps_decaps = NULL;
532 	}
533 	if (sh->modify_cmds) {
534 		mlx5_hlist_destroy(sh->modify_cmds);
535 		sh->modify_cmds = NULL;
536 	}
537 	if (sh->tag_table) {
538 		/* tags should be destroyed with flow before. */
539 		mlx5_hlist_destroy(sh->tag_table);
540 		sh->tag_table = NULL;
541 	}
542 	if (sh->tunnel_hub) {
543 		mlx5_release_tunnel_hub(sh, priv->dev_port);
544 		sh->tunnel_hub = NULL;
545 	}
546 	mlx5_free_table_hash_list(priv);
547 	if (sh->port_id_action_list) {
548 		mlx5_list_destroy(sh->port_id_action_list);
549 		sh->port_id_action_list = NULL;
550 	}
551 	if (sh->push_vlan_action_list) {
552 		mlx5_list_destroy(sh->push_vlan_action_list);
553 		sh->push_vlan_action_list = NULL;
554 	}
555 	if (sh->sample_action_list) {
556 		mlx5_list_destroy(sh->sample_action_list);
557 		sh->sample_action_list = NULL;
558 	}
559 	if (sh->dest_array_list) {
560 		mlx5_list_destroy(sh->dest_array_list);
561 		sh->dest_array_list = NULL;
562 	}
563 }
564 
565 /**
566  * Initialize shared data between primary and secondary process.
567  *
568  * A memzone is reserved by primary process and secondary processes attach to
569  * the memzone.
570  *
571  * @return
572  *   0 on success, a negative errno value otherwise and rte_errno is set.
573  */
574 static int
575 mlx5_init_shared_data(void)
576 {
577 	const struct rte_memzone *mz;
578 	int ret = 0;
579 
580 	rte_spinlock_lock(&mlx5_shared_data_lock);
581 	if (mlx5_shared_data == NULL) {
582 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
583 			/* Allocate shared memory. */
584 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
585 						 sizeof(*mlx5_shared_data),
586 						 SOCKET_ID_ANY, 0);
587 			if (mz == NULL) {
588 				DRV_LOG(ERR,
589 					"Cannot allocate mlx5 shared data");
590 				ret = -rte_errno;
591 				goto error;
592 			}
593 			mlx5_shared_data = mz->addr;
594 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
595 			rte_spinlock_init(&mlx5_shared_data->lock);
596 		} else {
597 			/* Lookup allocated shared memory. */
598 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
599 			if (mz == NULL) {
600 				DRV_LOG(ERR,
601 					"Cannot attach mlx5 shared data");
602 				ret = -rte_errno;
603 				goto error;
604 			}
605 			mlx5_shared_data = mz->addr;
606 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
607 		}
608 	}
609 error:
610 	rte_spinlock_unlock(&mlx5_shared_data_lock);
611 	return ret;
612 }
613 
614 /**
615  * PMD global initialization.
616  *
617  * Independent from individual device, this function initializes global
618  * per-PMD data structures distinguishing primary and secondary processes.
619  * Hence, each initialization is called once per a process.
620  *
621  * @return
622  *   0 on success, a negative errno value otherwise and rte_errno is set.
623  */
624 static int
625 mlx5_init_once(void)
626 {
627 	struct mlx5_shared_data *sd;
628 	struct mlx5_local_data *ld = &mlx5_local_data;
629 	int ret = 0;
630 
631 	if (mlx5_init_shared_data())
632 		return -rte_errno;
633 	sd = mlx5_shared_data;
634 	MLX5_ASSERT(sd);
635 	rte_spinlock_lock(&sd->lock);
636 	switch (rte_eal_process_type()) {
637 	case RTE_PROC_PRIMARY:
638 		if (sd->init_done)
639 			break;
640 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
641 					   mlx5_mp_os_primary_handle);
642 		if (ret)
643 			goto out;
644 		sd->init_done = true;
645 		break;
646 	case RTE_PROC_SECONDARY:
647 		if (ld->init_done)
648 			break;
649 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
650 					     mlx5_mp_os_secondary_handle);
651 		if (ret)
652 			goto out;
653 		++sd->secondary_cnt;
654 		ld->init_done = true;
655 		break;
656 	default:
657 		break;
658 	}
659 out:
660 	rte_spinlock_unlock(&sd->lock);
661 	return ret;
662 }
663 
664 /**
665  * DR flow drop action support detect.
666  *
667  * @param dev
668  *   Pointer to rte_eth_dev structure.
669  *
670  */
671 static void
672 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
673 {
674 #ifdef HAVE_MLX5DV_DR
675 	struct mlx5_priv *priv = dev->data->dev_private;
676 
677 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
678 		return;
679 	/**
680 	 * DR supports drop action placeholder when it is supported;
681 	 * otherwise, use the queue drop action.
682 	 */
683 	if (!priv->sh->drop_action_check_flag) {
684 		if (!mlx5_flow_discover_dr_action_support(dev))
685 			priv->sh->dr_drop_action_en = 1;
686 		priv->sh->drop_action_check_flag = 1;
687 	}
688 	if (priv->sh->dr_drop_action_en)
689 		priv->root_drop_action = priv->sh->dr_drop_action;
690 	else
691 		priv->root_drop_action = priv->drop_queue.hrxq->action;
692 #endif
693 }
694 
695 static void
696 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
697 {
698 	struct mlx5_priv *priv = dev->data->dev_private;
699 	void *ctx = priv->sh->cdev->ctx;
700 
701 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
702 	if (!priv->q_counters) {
703 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
704 		struct ibv_wq *wq;
705 
706 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
707 			"by DevX - fall-back to use the kernel driver global "
708 			"queue counter.", dev->data->port_id);
709 		/* Create WQ by kernel and query its queue counter ID. */
710 		if (cq) {
711 			wq = mlx5_glue->create_wq(ctx,
712 						  &(struct ibv_wq_init_attr){
713 						    .wq_type = IBV_WQT_RQ,
714 						    .max_wr = 1,
715 						    .max_sge = 1,
716 						    .pd = priv->sh->cdev->pd,
717 						    .cq = cq,
718 						});
719 			if (wq) {
720 				/* Counter is assigned only on RDY state. */
721 				int ret = mlx5_glue->modify_wq(wq,
722 						 &(struct ibv_wq_attr){
723 						 .attr_mask = IBV_WQ_ATTR_STATE,
724 						 .wq_state = IBV_WQS_RDY,
725 						});
726 
727 				if (ret == 0)
728 					mlx5_devx_cmd_wq_query(wq,
729 							 &priv->counter_set_id);
730 				claim_zero(mlx5_glue->destroy_wq(wq));
731 			}
732 			claim_zero(mlx5_glue->destroy_cq(cq));
733 		}
734 	} else {
735 		priv->counter_set_id = priv->q_counters->id;
736 	}
737 	if (priv->counter_set_id == 0)
738 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
739 			"available.", dev->data->port_id);
740 }
741 
742 /**
743  * Check if representor spawn info match devargs.
744  *
745  * @param spawn
746  *   Verbs device parameters (name, port, switch_info) to spawn.
747  * @param eth_da
748  *   Device devargs to probe.
749  *
750  * @return
751  *   Match result.
752  */
753 static bool
754 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
755 		       struct rte_eth_devargs *eth_da)
756 {
757 	struct mlx5_switch_info *switch_info = &spawn->info;
758 	unsigned int p, f;
759 	uint16_t id;
760 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
761 						      eth_da->type);
762 
763 	switch (eth_da->type) {
764 	case RTE_ETH_REPRESENTOR_SF:
765 		if (!(spawn->info.port_name == -1 &&
766 		      switch_info->name_type ==
767 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
768 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
769 			rte_errno = EBUSY;
770 			return false;
771 		}
772 		break;
773 	case RTE_ETH_REPRESENTOR_VF:
774 		/* Allows HPF representor index -1 as exception. */
775 		if (!(spawn->info.port_name == -1 &&
776 		      switch_info->name_type ==
777 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
778 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
779 			rte_errno = EBUSY;
780 			return false;
781 		}
782 		break;
783 	case RTE_ETH_REPRESENTOR_NONE:
784 		rte_errno = EBUSY;
785 		return false;
786 	default:
787 		rte_errno = ENOTSUP;
788 		DRV_LOG(ERR, "unsupported representor type");
789 		return false;
790 	}
791 	/* Check representor ID: */
792 	for (p = 0; p < eth_da->nb_ports; ++p) {
793 		if (spawn->pf_bond < 0) {
794 			/* For non-LAG mode, allow and ignore pf. */
795 			switch_info->pf_num = eth_da->ports[p];
796 			repr_id = mlx5_representor_id_encode(switch_info,
797 							     eth_da->type);
798 		}
799 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
800 			id = MLX5_REPRESENTOR_ID
801 				(eth_da->ports[p], eth_da->type,
802 				 eth_da->representor_ports[f]);
803 			if (repr_id == id)
804 				return true;
805 		}
806 	}
807 	rte_errno = EBUSY;
808 	return false;
809 }
810 
811 /**
812  * Spawn an Ethernet device from Verbs information.
813  *
814  * @param dpdk_dev
815  *   Backing DPDK device.
816  * @param spawn
817  *   Verbs device parameters (name, port, switch_info) to spawn.
818  * @param config
819  *   Device configuration parameters.
820  * @param eth_da
821  *   Device arguments.
822  *
823  * @return
824  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
825  *   is set. The following errors are defined:
826  *
827  *   EBUSY: device is not supposed to be spawned.
828  *   EEXIST: device is already spawned
829  */
830 static struct rte_eth_dev *
831 mlx5_dev_spawn(struct rte_device *dpdk_dev,
832 	       struct mlx5_dev_spawn_data *spawn,
833 	       struct mlx5_dev_config *config,
834 	       struct rte_eth_devargs *eth_da)
835 {
836 	const struct mlx5_switch_info *switch_info = &spawn->info;
837 	struct mlx5_dev_ctx_shared *sh = NULL;
838 	struct mlx5_hca_attr *hca_attr = &spawn->cdev->config.hca_attr;
839 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
840 	struct rte_eth_dev *eth_dev = NULL;
841 	struct mlx5_priv *priv = NULL;
842 	int err = 0;
843 	unsigned int hw_padding = 0;
844 	unsigned int mps;
845 	unsigned int mpls_en = 0;
846 	unsigned int swp = 0;
847 	unsigned int mprq = 0;
848 	struct rte_ether_addr mac;
849 	char name[RTE_ETH_NAME_MAX_LEN];
850 	int own_domain_id = 0;
851 	uint16_t port_id;
852 	struct mlx5_port_info vport_info = { .query_flags = 0 };
853 	int nl_rdma = -1;
854 	int i;
855 
856 	/* Determine if this port representor is supposed to be spawned. */
857 	if (switch_info->representor && dpdk_dev->devargs &&
858 	    !mlx5_representor_match(spawn, eth_da))
859 		return NULL;
860 	/* Build device name. */
861 	if (spawn->pf_bond < 0) {
862 		/* Single device. */
863 		if (!switch_info->representor)
864 			strlcpy(name, dpdk_dev->name, sizeof(name));
865 		else
866 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
867 				 dpdk_dev->name,
868 				 switch_info->name_type ==
869 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
870 				 switch_info->port_name);
871 	} else {
872 		/* Bonding device. */
873 		if (!switch_info->representor) {
874 			err = snprintf(name, sizeof(name), "%s_%s",
875 				       dpdk_dev->name, spawn->phys_dev_name);
876 		} else {
877 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
878 				dpdk_dev->name, spawn->phys_dev_name,
879 				switch_info->ctrl_num,
880 				switch_info->pf_num,
881 				switch_info->name_type ==
882 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
883 				switch_info->port_name);
884 		}
885 	}
886 	if (err >= (int)sizeof(name))
887 		DRV_LOG(WARNING, "device name overflow %s", name);
888 	/* check if the device is already spawned */
889 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
890 		rte_errno = EEXIST;
891 		return NULL;
892 	}
893 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
894 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
895 		struct mlx5_mp_id mp_id;
896 
897 		eth_dev = rte_eth_dev_attach_secondary(name);
898 		if (eth_dev == NULL) {
899 			DRV_LOG(ERR, "can not attach rte ethdev");
900 			rte_errno = ENOMEM;
901 			return NULL;
902 		}
903 		eth_dev->device = dpdk_dev;
904 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
905 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
906 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
907 		err = mlx5_proc_priv_init(eth_dev);
908 		if (err)
909 			return NULL;
910 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
911 		/* Receive command fd from primary process */
912 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
913 		if (err < 0)
914 			goto err_secondary;
915 		/* Remap UAR for Tx queues. */
916 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
917 		if (err)
918 			goto err_secondary;
919 		/*
920 		 * Ethdev pointer is still required as input since
921 		 * the primary device is not accessible from the
922 		 * secondary process.
923 		 */
924 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
925 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
926 		return eth_dev;
927 err_secondary:
928 		mlx5_dev_close(eth_dev);
929 		return NULL;
930 	}
931 	/* Process parameters. */
932 	err = mlx5_args(config, dpdk_dev->devargs);
933 	if (err) {
934 		DRV_LOG(ERR, "failed to process device arguments: %s",
935 			strerror(rte_errno));
936 		return NULL;
937 	}
938 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
939 	if (!sh)
940 		return NULL;
941 	/* Update final values for devargs before check sibling config. */
942 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
943 	if (config->dv_flow_en) {
944 		DRV_LOG(WARNING, "DV flow is not supported.");
945 		config->dv_flow_en = 0;
946 	}
947 #endif
948 #ifdef HAVE_MLX5DV_DR_ESWITCH
949 	if (!(hca_attr->eswitch_manager && config->dv_flow_en && sh->esw_mode))
950 		config->dv_esw_en = 0;
951 #else
952 	config->dv_esw_en = 0;
953 #endif
954 	if (config->dv_miss_info && config->dv_esw_en)
955 		config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
956 	if (!config->dv_esw_en &&
957 	    config->dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
958 		DRV_LOG(WARNING,
959 			"Metadata mode %u is not supported (no E-Switch).",
960 			config->dv_xmeta_en);
961 		config->dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
962 	}
963 	/* Check sibling device configurations. */
964 	err = mlx5_dev_check_sibling_config(sh, config, dpdk_dev);
965 	if (err)
966 		goto error;
967 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
968 	config->dest_tir = 1;
969 #endif
970 	/*
971 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
972 	 * as all ConnectX-5 devices.
973 	 */
974 	if (sh->dev_cap.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
975 		if (sh->dev_cap.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
976 			DRV_LOG(DEBUG, "enhanced MPW is supported");
977 			mps = MLX5_MPW_ENHANCED;
978 		} else {
979 			DRV_LOG(DEBUG, "MPW is supported");
980 			mps = MLX5_MPW;
981 		}
982 	} else {
983 		DRV_LOG(DEBUG, "MPW isn't supported");
984 		mps = MLX5_MPW_DISABLED;
985 	}
986 #ifdef HAVE_IBV_MLX5_MOD_SWP
987 	if (sh->dev_cap.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
988 		swp = sh->dev_cap.sw_parsing_offloads;
989 	DRV_LOG(DEBUG, "SWP support: %u", swp);
990 #endif
991 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
992 		MLX5_SW_PARSING_TSO_CAP);
993 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
994 	if (sh->dev_cap.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
995 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
996 			sh->dev_cap.min_single_stride_log_num_of_bytes);
997 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
998 			sh->dev_cap.max_single_stride_log_num_of_bytes);
999 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1000 			sh->dev_cap.min_single_wqe_log_num_of_strides);
1001 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1002 			sh->dev_cap.max_single_wqe_log_num_of_strides);
1003 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1004 			sh->dev_cap.stride_supported_qpts);
1005 		DRV_LOG(DEBUG, "\tmin_stride_wqe_log_size: %d",
1006 			config->mprq.log_min_stride_wqe_size);
1007 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1008 		mprq = 1;
1009 		config->mprq.log_min_stride_size =
1010 			sh->dev_cap.min_single_stride_log_num_of_bytes;
1011 		config->mprq.log_max_stride_size =
1012 			sh->dev_cap.max_single_stride_log_num_of_bytes;
1013 		config->mprq.log_min_stride_num =
1014 			sh->dev_cap.min_single_wqe_log_num_of_strides;
1015 		config->mprq.log_max_stride_num =
1016 			sh->dev_cap.max_single_wqe_log_num_of_strides;
1017 	}
1018 #endif
1019 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1020 	if (sh->dev_cap.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1021 		config->tunnel_en = sh->dev_cap.tunnel_offloads_caps &
1022 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1023 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1024 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1025 	}
1026 	if (config->tunnel_en) {
1027 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1028 		config->tunnel_en &
1029 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1030 		config->tunnel_en &
1031 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1032 		config->tunnel_en &
1033 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1034 		);
1035 	} else {
1036 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1037 	}
1038 #else
1039 	DRV_LOG(WARNING,
1040 		"tunnel offloading disabled due to old OFED/rdma-core version");
1041 #endif
1042 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1043 	mpls_en = ((sh->dev_cap.tunnel_offloads_caps &
1044 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1045 		   (sh->dev_cap.tunnel_offloads_caps &
1046 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1047 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1048 		mpls_en ? "" : "not ");
1049 #else
1050 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1051 		" old OFED/rdma-core version or firmware configuration");
1052 #endif
1053 	config->mpls_en = mpls_en;
1054 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1055 	/* Check port status. */
1056 	if (spawn->phys_port <= UINT8_MAX) {
1057 		/* Legacy Verbs api only support u8 port number. */
1058 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1059 					    &port_attr);
1060 		if (err) {
1061 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1062 			goto error;
1063 		}
1064 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1065 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1066 			err = EINVAL;
1067 			goto error;
1068 		}
1069 	} else if (nl_rdma >= 0) {
1070 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1071 		err = mlx5_nl_port_state(nl_rdma,
1072 			spawn->phys_dev_name,
1073 			spawn->phys_port);
1074 		if (err < 0) {
1075 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1076 				strerror(rte_errno));
1077 			err = -rte_errno;
1078 			goto error;
1079 		}
1080 		port_attr.state = (enum ibv_port_state)err;
1081 	}
1082 	if (port_attr.state != IBV_PORT_ACTIVE)
1083 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1084 			mlx5_glue->port_state_str(port_attr.state),
1085 			port_attr.state);
1086 	/* Allocate private eth device data. */
1087 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1088 			   sizeof(*priv),
1089 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1090 	if (priv == NULL) {
1091 		DRV_LOG(ERR, "priv allocation failure");
1092 		err = ENOMEM;
1093 		goto error;
1094 	}
1095 	priv->sh = sh;
1096 	priv->dev_port = spawn->phys_port;
1097 	priv->pci_dev = spawn->pci_dev;
1098 	priv->mtu = RTE_ETHER_MTU;
1099 	/* Some internal functions rely on Netlink sockets, open them now. */
1100 	priv->nl_socket_rdma = nl_rdma;
1101 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1102 	priv->representor = !!switch_info->representor;
1103 	priv->master = !!switch_info->master;
1104 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1105 	priv->vport_meta_tag = 0;
1106 	priv->vport_meta_mask = 0;
1107 	priv->pf_bond = spawn->pf_bond;
1108 
1109 	DRV_LOG(DEBUG,
1110 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1111 		priv->dev_port, dpdk_dev->bus->name,
1112 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1113 		priv->master, priv->representor, priv->pf_bond);
1114 
1115 	/*
1116 	 * If we have E-Switch we should determine the vport attributes.
1117 	 * E-Switch may use either source vport field or reg_c[0] metadata
1118 	 * register to match on vport index. The engaged part of metadata
1119 	 * register is defined by mask.
1120 	 */
1121 	if (sh->esw_mode) {
1122 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1123 						 spawn->phys_port,
1124 						 &vport_info);
1125 		if (err) {
1126 			DRV_LOG(WARNING,
1127 				"Cannot query devx port %d on device %s",
1128 				spawn->phys_port, spawn->phys_dev_name);
1129 			vport_info.query_flags = 0;
1130 		}
1131 	}
1132 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1133 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1134 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1135 		if (!priv->vport_meta_mask) {
1136 			DRV_LOG(ERR,
1137 				"vport zero mask for port %d on bonding device %s",
1138 				spawn->phys_port, spawn->phys_dev_name);
1139 			err = ENOTSUP;
1140 			goto error;
1141 		}
1142 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1143 			DRV_LOG(ERR,
1144 				"Invalid vport tag for port %d on bonding device %s",
1145 				spawn->phys_port, spawn->phys_dev_name);
1146 			err = ENOTSUP;
1147 			goto error;
1148 		}
1149 	}
1150 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1151 		priv->vport_id = vport_info.vport_id;
1152 	} else if (spawn->pf_bond >= 0 && sh->esw_mode) {
1153 		DRV_LOG(ERR,
1154 			"Cannot deduce vport index for port %d on bonding device %s",
1155 			spawn->phys_port, spawn->phys_dev_name);
1156 		err = ENOTSUP;
1157 		goto error;
1158 	} else {
1159 		/*
1160 		 * Suppose vport index in compatible way. Kernel/rdma_core
1161 		 * support single E-Switch per PF configurations only and
1162 		 * vport_id field contains the vport index for associated VF,
1163 		 * which is deduced from representor port name.
1164 		 * For example, let's have the IB device port 10, it has
1165 		 * attached network device eth0, which has port name attribute
1166 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1167 		 * as 3 (2+1). This assigning schema should be changed if the
1168 		 * multiple E-Switch instances per PF configurations or/and PCI
1169 		 * subfunctions are added.
1170 		 */
1171 		priv->vport_id = switch_info->representor ?
1172 				 switch_info->port_name + 1 : -1;
1173 	}
1174 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1175 							  eth_da->type);
1176 	/*
1177 	 * Look for sibling devices in order to reuse their switch domain
1178 	 * if any, otherwise allocate one.
1179 	 */
1180 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1181 		const struct mlx5_priv *opriv =
1182 			rte_eth_devices[port_id].data->dev_private;
1183 
1184 		if (!opriv ||
1185 		    opriv->sh != priv->sh ||
1186 			opriv->domain_id ==
1187 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1188 			continue;
1189 		priv->domain_id = opriv->domain_id;
1190 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1191 			priv->dev_port, priv->domain_id);
1192 		break;
1193 	}
1194 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1195 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1196 		if (err) {
1197 			err = rte_errno;
1198 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1199 				strerror(rte_errno));
1200 			goto error;
1201 		}
1202 		own_domain_id = 1;
1203 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1204 			priv->dev_port, priv->domain_id);
1205 	}
1206 	config->hw_csum = !!(sh->dev_cap.device_cap_flags_ex &
1207 			    IBV_DEVICE_RAW_IP_CSUM);
1208 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1209 		(config->hw_csum ? "" : "not "));
1210 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1211 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1212 	DRV_LOG(DEBUG, "counters are not supported");
1213 #endif
1214 	config->ind_table_max_size =
1215 		sh->dev_cap.max_rwq_indirection_table_size;
1216 	/*
1217 	 * Remove this check once DPDK supports larger/variable
1218 	 * indirection tables.
1219 	 */
1220 	if (config->ind_table_max_size > (unsigned int)RTE_ETH_RSS_RETA_SIZE_512)
1221 		config->ind_table_max_size = RTE_ETH_RSS_RETA_SIZE_512;
1222 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1223 		config->ind_table_max_size);
1224 	config->hw_vlan_strip = !!(sh->dev_cap.raw_packet_caps &
1225 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1226 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1227 		(config->hw_vlan_strip ? "" : "not "));
1228 	config->hw_fcs_strip = !!(sh->dev_cap.raw_packet_caps &
1229 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1230 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1231 	hw_padding = !!sh->dev_cap.rx_pad_end_addr_align;
1232 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1233 	hw_padding = !!(sh->dev_cap.device_cap_flags_ex &
1234 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1235 #endif
1236 	if (config->hw_padding && !hw_padding) {
1237 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1238 		config->hw_padding = 0;
1239 	} else if (config->hw_padding) {
1240 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1241 	}
1242 	config->tso = (sh->dev_cap.max_tso > 0 &&
1243 		      (sh->dev_cap.tso_supported_qpts &
1244 		       (1 << IBV_QPT_RAW_PACKET)));
1245 	if (config->tso)
1246 		config->tso_max_payload_sz = sh->dev_cap.max_tso;
1247 	/*
1248 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1249 	 * by default.
1250 	 */
1251 	if (config->mps == MLX5_ARG_UNSET)
1252 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1253 							  MLX5_MPW_DISABLED;
1254 	else
1255 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1256 	DRV_LOG(INFO, "%sMPS is %s",
1257 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1258 		config->mps == MLX5_MPW ? "legacy " : "",
1259 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1260 	if (sh->cdev->config.devx) {
1261 		sh->steering_format_version = hca_attr->steering_format_version;
1262 		/* Check for LRO support. */
1263 		if (config->dest_tir && hca_attr->lro_cap &&
1264 		    config->dv_flow_en) {
1265 			/* TBD check tunnel lro caps. */
1266 			config->lro.supported = hca_attr->lro_cap;
1267 			DRV_LOG(DEBUG, "Device supports LRO");
1268 			/*
1269 			 * If LRO timeout is not configured by application,
1270 			 * use the minimal supported value.
1271 			 */
1272 			if (!config->lro.timeout)
1273 				config->lro.timeout =
1274 				       hca_attr->lro_timer_supported_periods[0];
1275 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1276 				config->lro.timeout);
1277 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1278 				"required for coalescing is %d bytes",
1279 				hca_attr->lro_min_mss_size);
1280 		}
1281 #if defined(HAVE_MLX5DV_DR) && \
1282 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1283 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1284 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_old &&
1285 		    config->dv_flow_en) {
1286 			uint8_t reg_c_mask = hca_attr->qos.flow_meter_reg_c_ids;
1287 			/*
1288 			 * Meter needs two REG_C's for color match and pre-sfx
1289 			 * flow match. Here get the REG_C for color match.
1290 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1291 			 */
1292 			reg_c_mask &= 0xfc;
1293 			if (__builtin_popcount(reg_c_mask) < 1) {
1294 				priv->mtr_en = 0;
1295 				DRV_LOG(WARNING, "No available register for"
1296 					" meter.");
1297 			} else {
1298 				/*
1299 				 * The meter color register is used by the
1300 				 * flow-hit feature as well.
1301 				 * The flow-hit feature must use REG_C_3
1302 				 * Prefer REG_C_3 if it is available.
1303 				 */
1304 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1305 					priv->mtr_color_reg = REG_C_3;
1306 				else
1307 					priv->mtr_color_reg = ffs(reg_c_mask)
1308 							      - 1 + REG_C_0;
1309 				priv->mtr_en = 1;
1310 				priv->mtr_reg_share = hca_attr->qos.flow_meter;
1311 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1312 					priv->mtr_color_reg);
1313 			}
1314 		}
1315 		if (hca_attr->qos.sup && hca_attr->qos.flow_meter_aso_sup) {
1316 			uint32_t log_obj_size =
1317 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1318 			if (log_obj_size >=
1319 			    hca_attr->qos.log_meter_aso_granularity &&
1320 			    log_obj_size <=
1321 			    hca_attr->qos.log_meter_aso_max_alloc)
1322 				sh->meter_aso_en = 1;
1323 		}
1324 		if (priv->mtr_en) {
1325 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1326 			if (err) {
1327 				err = -err;
1328 				goto error;
1329 			}
1330 		}
1331 		if (hca_attr->flow.tunnel_header_0_1)
1332 			sh->tunnel_header_0_1 = 1;
1333 #endif
1334 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1335 		if (hca_attr->flow_hit_aso && priv->mtr_color_reg == REG_C_3) {
1336 			sh->flow_hit_aso_en = 1;
1337 			err = mlx5_flow_aso_age_mng_init(sh);
1338 			if (err) {
1339 				err = -err;
1340 				goto error;
1341 			}
1342 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1343 		}
1344 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1345 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1346 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1347 		if (hca_attr->ct_offload && priv->mtr_color_reg == REG_C_3) {
1348 			err = mlx5_flow_aso_ct_mng_init(sh);
1349 			if (err) {
1350 				err = -err;
1351 				goto error;
1352 			}
1353 			DRV_LOG(DEBUG, "CT ASO is supported.");
1354 			sh->ct_aso_en = 1;
1355 		}
1356 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1357 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1358 		if (hca_attr->log_max_ft_sampler_num > 0  &&
1359 		    config->dv_flow_en) {
1360 			priv->sampler_en = 1;
1361 			DRV_LOG(DEBUG, "Sampler enabled!");
1362 		} else {
1363 			priv->sampler_en = 0;
1364 			if (!hca_attr->log_max_ft_sampler_num)
1365 				DRV_LOG(WARNING,
1366 					"No available register for sampler.");
1367 			else
1368 				DRV_LOG(DEBUG, "DV flow is not supported!");
1369 		}
1370 #endif
1371 	}
1372 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1373 	    !(sh->dev_cap.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1374 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1375 		config->cqe_comp = 0;
1376 	}
1377 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1378 	    (!sh->cdev->config.devx || !hca_attr->mini_cqe_resp_flow_tag)) {
1379 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1380 				 " format isn't supported.");
1381 		config->cqe_comp = 0;
1382 	}
1383 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1384 	    (!sh->cdev->config.devx || !hca_attr->mini_cqe_resp_l3_l4_tag)) {
1385 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1386 				 " format isn't supported.");
1387 		config->cqe_comp = 0;
1388 	}
1389 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1390 			config->cqe_comp ? "" : "not ");
1391 	if (config->tx_pp) {
1392 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1393 			hca_attr->dev_freq_khz);
1394 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1395 			hca_attr->qos.packet_pacing ? "" : "not ");
1396 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1397 			hca_attr->cross_channel ? "" : "not ");
1398 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1399 			hca_attr->wqe_index_ignore ? "" : "not ");
1400 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1401 			hca_attr->non_wire_sq ? "" : "not ");
1402 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1403 			hca_attr->log_max_static_sq_wq ? "" : "not ",
1404 			hca_attr->log_max_static_sq_wq);
1405 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1406 			hca_attr->qos.wqe_rate_pp ? "" : "not ");
1407 		if (!sh->cdev->config.devx) {
1408 			DRV_LOG(ERR, "DevX is required for packet pacing");
1409 			err = ENODEV;
1410 			goto error;
1411 		}
1412 		if (!hca_attr->qos.packet_pacing) {
1413 			DRV_LOG(ERR, "Packet pacing is not supported");
1414 			err = ENODEV;
1415 			goto error;
1416 		}
1417 		if (!hca_attr->cross_channel) {
1418 			DRV_LOG(ERR, "Cross channel operations are"
1419 				     " required for packet pacing");
1420 			err = ENODEV;
1421 			goto error;
1422 		}
1423 		if (!hca_attr->wqe_index_ignore) {
1424 			DRV_LOG(ERR, "WQE index ignore feature is"
1425 				     " required for packet pacing");
1426 			err = ENODEV;
1427 			goto error;
1428 		}
1429 		if (!hca_attr->non_wire_sq) {
1430 			DRV_LOG(ERR, "Non-wire SQ feature is"
1431 				     " required for packet pacing");
1432 			err = ENODEV;
1433 			goto error;
1434 		}
1435 		if (!hca_attr->log_max_static_sq_wq) {
1436 			DRV_LOG(ERR, "Static WQE SQ feature is"
1437 				     " required for packet pacing");
1438 			err = ENODEV;
1439 			goto error;
1440 		}
1441 		if (!hca_attr->qos.wqe_rate_pp) {
1442 			DRV_LOG(ERR, "WQE rate mode is required"
1443 				     " for packet pacing");
1444 			err = ENODEV;
1445 			goto error;
1446 		}
1447 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1448 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1449 			     " can't create queues for packet pacing");
1450 		err = ENODEV;
1451 		goto error;
1452 #endif
1453 	}
1454 	if (config->std_delay_drop || config->hp_delay_drop) {
1455 		if (!hca_attr->rq_delay_drop) {
1456 			config->std_delay_drop = 0;
1457 			config->hp_delay_drop = 0;
1458 			DRV_LOG(WARNING,
1459 				"dev_port-%u: Rxq delay drop is not supported",
1460 				priv->dev_port);
1461 		}
1462 	}
1463 	if (sh->cdev->config.devx)
1464 		mlx5_rt_timestamp_config(sh, config, hca_attr);
1465 	/*
1466 	 * If HW has bug working with tunnel packet decapsulation and
1467 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1468 	 * bit. Then RTE_ETH_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1469 	 */
1470 	if (hca_attr->scatter_fcs_w_decap_disable && config->decap_en)
1471 		config->hw_fcs_strip = 0;
1472 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1473 		(config->hw_fcs_strip ? "" : "not "));
1474 	if (config->mprq.enabled && !mprq) {
1475 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1476 		config->mprq.enabled = 0;
1477 	}
1478 	if (config->max_dump_files_num == 0)
1479 		config->max_dump_files_num = 128;
1480 	eth_dev = rte_eth_dev_allocate(name);
1481 	if (eth_dev == NULL) {
1482 		DRV_LOG(ERR, "can not allocate rte ethdev");
1483 		err = ENOMEM;
1484 		goto error;
1485 	}
1486 	if (priv->representor) {
1487 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1488 		eth_dev->data->representor_id = priv->representor_id;
1489 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1490 			struct mlx5_priv *opriv =
1491 				rte_eth_devices[port_id].data->dev_private;
1492 			if (opriv &&
1493 			    opriv->master &&
1494 			    opriv->domain_id == priv->domain_id &&
1495 			    opriv->sh == priv->sh) {
1496 				eth_dev->data->backer_port_id = port_id;
1497 				break;
1498 			}
1499 		}
1500 		if (port_id >= RTE_MAX_ETHPORTS)
1501 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1502 	}
1503 	priv->mp_id.port_id = eth_dev->data->port_id;
1504 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1505 	/*
1506 	 * Store associated network device interface index. This index
1507 	 * is permanent throughout the lifetime of device. So, we may store
1508 	 * the ifindex here and use the cached value further.
1509 	 */
1510 	MLX5_ASSERT(spawn->ifindex);
1511 	priv->if_index = spawn->ifindex;
1512 	priv->lag_affinity_idx = sh->refcnt - 1;
1513 	eth_dev->data->dev_private = priv;
1514 	priv->dev_data = eth_dev->data;
1515 	eth_dev->data->mac_addrs = priv->mac;
1516 	eth_dev->device = dpdk_dev;
1517 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1518 	/* Configure the first MAC address by default. */
1519 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1520 		DRV_LOG(ERR,
1521 			"port %u cannot get MAC address, is mlx5_en"
1522 			" loaded? (errno: %s)",
1523 			eth_dev->data->port_id, strerror(rte_errno));
1524 		err = ENODEV;
1525 		goto error;
1526 	}
1527 	DRV_LOG(INFO,
1528 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1529 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1530 #ifdef RTE_LIBRTE_MLX5_DEBUG
1531 	{
1532 		char ifname[MLX5_NAMESIZE];
1533 
1534 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1535 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1536 				eth_dev->data->port_id, ifname);
1537 		else
1538 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1539 				eth_dev->data->port_id);
1540 	}
1541 #endif
1542 	/* Get actual MTU if possible. */
1543 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1544 	if (err) {
1545 		err = rte_errno;
1546 		goto error;
1547 	}
1548 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1549 		priv->mtu);
1550 	/* Initialize burst functions to prevent crashes before link-up. */
1551 	eth_dev->rx_pkt_burst = rte_eth_pkt_burst_dummy;
1552 	eth_dev->tx_pkt_burst = rte_eth_pkt_burst_dummy;
1553 	eth_dev->dev_ops = &mlx5_dev_ops;
1554 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1555 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1556 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1557 	/* Register MAC address. */
1558 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1559 	if (config->vf && config->vf_nl_en)
1560 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1561 				      mlx5_ifindex(eth_dev),
1562 				      eth_dev->data->mac_addrs,
1563 				      MLX5_MAX_MAC_ADDRESSES);
1564 	priv->ctrl_flows = 0;
1565 	rte_spinlock_init(&priv->flow_list_lock);
1566 	TAILQ_INIT(&priv->flow_meters);
1567 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1568 	if (!priv->mtr_profile_tbl)
1569 		goto error;
1570 	/* Bring Ethernet device up. */
1571 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1572 		eth_dev->data->port_id);
1573 	mlx5_set_link_up(eth_dev);
1574 	/*
1575 	 * Even though the interrupt handler is not installed yet,
1576 	 * interrupts will still trigger on the async_fd from
1577 	 * Verbs context returned by ibv_open_device().
1578 	 */
1579 	mlx5_link_update(eth_dev, 0);
1580 	/* Detect minimal data bytes to inline. */
1581 	mlx5_set_min_inline(spawn, config);
1582 	/* Store device configuration on private structure. */
1583 	priv->config = *config;
1584 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1585 		icfg[i].release_mem_en = !!config->reclaim_mode;
1586 		if (config->reclaim_mode)
1587 			icfg[i].per_core_cache = 0;
1588 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1589 		if (!priv->flows[i])
1590 			goto error;
1591 	}
1592 	/* Create context for virtual machine VLAN workaround. */
1593 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1594 	if (config->dv_flow_en) {
1595 		err = mlx5_alloc_shared_dr(priv);
1596 		if (err)
1597 			goto error;
1598 		if (mlx5_flex_item_port_init(eth_dev) < 0)
1599 			goto error;
1600 	}
1601 	if (sh->cdev->config.devx && config->dv_flow_en && config->dest_tir) {
1602 		priv->obj_ops = devx_obj_ops;
1603 		mlx5_queue_counter_id_prepare(eth_dev);
1604 		priv->obj_ops.lb_dummy_queue_create =
1605 					mlx5_rxq_ibv_obj_dummy_lb_create;
1606 		priv->obj_ops.lb_dummy_queue_release =
1607 					mlx5_rxq_ibv_obj_dummy_lb_release;
1608 	} else if (spawn->max_port > UINT8_MAX) {
1609 		/* Verbs can't support ports larger than 255 by design. */
1610 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1611 		err = ENOTSUP;
1612 		goto error;
1613 	} else {
1614 		priv->obj_ops = ibv_obj_ops;
1615 	}
1616 	if (config->tx_pp &&
1617 	    priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new) {
1618 		/*
1619 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1620 		 * packet pacing and already checked above.
1621 		 * Hence, we should only make sure the SQs will be created
1622 		 * with DevX, not with Verbs.
1623 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1624 		 * with Clock Queue UAR as required for Tx scheduling.
1625 		 */
1626 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1627 		err = ENODEV;
1628 		goto error;
1629 	}
1630 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1631 	if (!priv->drop_queue.hrxq)
1632 		goto error;
1633 	/* Port representor shares the same max priority with pf port. */
1634 	if (!priv->sh->flow_priority_check_flag) {
1635 		/* Supported Verbs flow priority number detection. */
1636 		err = mlx5_flow_discover_priorities(eth_dev);
1637 		priv->sh->flow_max_priority = err;
1638 		priv->sh->flow_priority_check_flag = 1;
1639 	} else {
1640 		err = priv->sh->flow_max_priority;
1641 	}
1642 	if (err < 0) {
1643 		err = -err;
1644 		goto error;
1645 	}
1646 	mlx5_set_metadata_mask(eth_dev);
1647 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1648 	    !priv->sh->dv_regc0_mask) {
1649 		DRV_LOG(ERR, "metadata mode %u is not supported "
1650 			     "(no metadata reg_c[0] is available)",
1651 			     priv->config.dv_xmeta_en);
1652 			err = ENOTSUP;
1653 			goto error;
1654 	}
1655 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1656 				       mlx5_hrxq_create_cb,
1657 				       mlx5_hrxq_match_cb,
1658 				       mlx5_hrxq_remove_cb,
1659 				       mlx5_hrxq_clone_cb,
1660 				       mlx5_hrxq_clone_free_cb);
1661 	if (!priv->hrxqs)
1662 		goto error;
1663 	rte_rwlock_init(&priv->ind_tbls_lock);
1664 	/* Query availability of metadata reg_c's. */
1665 	if (!priv->sh->metadata_regc_check_flag) {
1666 		err = mlx5_flow_discover_mreg_c(eth_dev);
1667 		if (err < 0) {
1668 			err = -err;
1669 			goto error;
1670 		}
1671 	}
1672 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1673 		DRV_LOG(DEBUG,
1674 			"port %u extensive metadata register is not supported",
1675 			eth_dev->data->port_id);
1676 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1677 			DRV_LOG(ERR, "metadata mode %u is not supported "
1678 				     "(no metadata registers available)",
1679 				     priv->config.dv_xmeta_en);
1680 			err = ENOTSUP;
1681 			goto error;
1682 		}
1683 	}
1684 	if (priv->config.dv_flow_en &&
1685 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1686 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1687 	    priv->sh->dv_regc0_mask) {
1688 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1689 						      MLX5_FLOW_MREG_HTABLE_SZ,
1690 						      false, true, eth_dev,
1691 						      flow_dv_mreg_create_cb,
1692 						      flow_dv_mreg_match_cb,
1693 						      flow_dv_mreg_remove_cb,
1694 						      flow_dv_mreg_clone_cb,
1695 						    flow_dv_mreg_clone_free_cb);
1696 		if (!priv->mreg_cp_tbl) {
1697 			err = ENOMEM;
1698 			goto error;
1699 		}
1700 	}
1701 	rte_spinlock_init(&priv->shared_act_sl);
1702 	mlx5_flow_counter_mode_config(eth_dev);
1703 	mlx5_flow_drop_action_config(eth_dev);
1704 	if (priv->config.dv_flow_en)
1705 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1706 	return eth_dev;
1707 error:
1708 	if (priv) {
1709 		if (priv->mreg_cp_tbl)
1710 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1711 		if (priv->sh)
1712 			mlx5_os_free_shared_dr(priv);
1713 		if (priv->nl_socket_route >= 0)
1714 			close(priv->nl_socket_route);
1715 		if (priv->vmwa_context)
1716 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1717 		if (eth_dev && priv->drop_queue.hrxq)
1718 			mlx5_drop_action_destroy(eth_dev);
1719 		if (priv->mtr_profile_tbl)
1720 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1721 		if (own_domain_id)
1722 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1723 		if (priv->hrxqs)
1724 			mlx5_list_destroy(priv->hrxqs);
1725 		if (eth_dev && priv->flex_item_map)
1726 			mlx5_flex_item_port_cleanup(eth_dev);
1727 		mlx5_free(priv);
1728 		if (eth_dev != NULL)
1729 			eth_dev->data->dev_private = NULL;
1730 	}
1731 	if (eth_dev != NULL) {
1732 		/* mac_addrs must not be freed alone because part of
1733 		 * dev_private
1734 		 **/
1735 		eth_dev->data->mac_addrs = NULL;
1736 		rte_eth_dev_release_port(eth_dev);
1737 	}
1738 	if (sh)
1739 		mlx5_free_shared_dev_ctx(sh);
1740 	if (nl_rdma >= 0)
1741 		close(nl_rdma);
1742 	MLX5_ASSERT(err > 0);
1743 	rte_errno = err;
1744 	return NULL;
1745 }
1746 
1747 /**
1748  * Comparison callback to sort device data.
1749  *
1750  * This is meant to be used with qsort().
1751  *
1752  * @param a[in]
1753  *   Pointer to pointer to first data object.
1754  * @param b[in]
1755  *   Pointer to pointer to second data object.
1756  *
1757  * @return
1758  *   0 if both objects are equal, less than 0 if the first argument is less
1759  *   than the second, greater than 0 otherwise.
1760  */
1761 static int
1762 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1763 {
1764 	const struct mlx5_switch_info *si_a =
1765 		&((const struct mlx5_dev_spawn_data *)a)->info;
1766 	const struct mlx5_switch_info *si_b =
1767 		&((const struct mlx5_dev_spawn_data *)b)->info;
1768 	int ret;
1769 
1770 	/* Master device first. */
1771 	ret = si_b->master - si_a->master;
1772 	if (ret)
1773 		return ret;
1774 	/* Then representor devices. */
1775 	ret = si_b->representor - si_a->representor;
1776 	if (ret)
1777 		return ret;
1778 	/* Unidentified devices come last in no specific order. */
1779 	if (!si_a->representor)
1780 		return 0;
1781 	/* Order representors by name. */
1782 	return si_a->port_name - si_b->port_name;
1783 }
1784 
1785 /**
1786  * Match PCI information for possible slaves of bonding device.
1787  *
1788  * @param[in] ibdev_name
1789  *   Name of Infiniband device.
1790  * @param[in] pci_dev
1791  *   Pointer to primary PCI address structure to match.
1792  * @param[in] nl_rdma
1793  *   Netlink RDMA group socket handle.
1794  * @param[in] owner
1795  *   Representor owner PF index.
1796  * @param[out] bond_info
1797  *   Pointer to bonding information.
1798  *
1799  * @return
1800  *   negative value if no bonding device found, otherwise
1801  *   positive index of slave PF in bonding.
1802  */
1803 static int
1804 mlx5_device_bond_pci_match(const char *ibdev_name,
1805 			   const struct rte_pci_addr *pci_dev,
1806 			   int nl_rdma, uint16_t owner,
1807 			   struct mlx5_bond_info *bond_info)
1808 {
1809 	char ifname[IF_NAMESIZE + 1];
1810 	unsigned int ifindex;
1811 	unsigned int np, i;
1812 	FILE *bond_file = NULL, *file;
1813 	int pf = -1;
1814 	int ret;
1815 	uint8_t cur_guid[32] = {0};
1816 	uint8_t guid[32] = {0};
1817 
1818 	/*
1819 	 * Try to get master device name. If something goes wrong suppose
1820 	 * the lack of kernel support and no bonding devices.
1821 	 */
1822 	memset(bond_info, 0, sizeof(*bond_info));
1823 	if (nl_rdma < 0)
1824 		return -1;
1825 	if (!strstr(ibdev_name, "bond"))
1826 		return -1;
1827 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1828 	if (!np)
1829 		return -1;
1830 	if (mlx5_get_device_guid(pci_dev, cur_guid, sizeof(cur_guid)) < 0)
1831 		return -1;
1832 	/*
1833 	 * The master device might not be on the predefined port(not on port
1834 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1835 	 * device ports and find master.
1836 	 */
1837 	for (i = 1; i <= np; ++i) {
1838 		/* Check whether Infiniband port is populated. */
1839 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1840 		if (!ifindex)
1841 			continue;
1842 		if (!if_indextoname(ifindex, ifname))
1843 			continue;
1844 		/* Try to read bonding slave names from sysfs. */
1845 		MKSTR(slaves,
1846 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1847 		bond_file = fopen(slaves, "r");
1848 		if (bond_file)
1849 			break;
1850 	}
1851 	if (!bond_file)
1852 		return -1;
1853 	/* Use safe format to check maximal buffer length. */
1854 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1855 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1856 		char tmp_str[IF_NAMESIZE + 32];
1857 		struct rte_pci_addr pci_addr;
1858 		struct mlx5_switch_info	info;
1859 		int ret;
1860 
1861 		/* Process slave interface names in the loop. */
1862 		snprintf(tmp_str, sizeof(tmp_str),
1863 			 "/sys/class/net/%s", ifname);
1864 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1865 			DRV_LOG(WARNING,
1866 				"Cannot get PCI address for netdev \"%s\".",
1867 				ifname);
1868 			continue;
1869 		}
1870 		/* Slave interface PCI address match found. */
1871 		snprintf(tmp_str, sizeof(tmp_str),
1872 			 "/sys/class/net/%s/phys_port_name", ifname);
1873 		file = fopen(tmp_str, "rb");
1874 		if (!file)
1875 			break;
1876 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1877 		if (fscanf(file, "%32s", tmp_str) == 1)
1878 			mlx5_translate_port_name(tmp_str, &info);
1879 		fclose(file);
1880 		/* Only process PF ports. */
1881 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1882 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1883 			continue;
1884 		/* Check max bonding member. */
1885 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1886 			DRV_LOG(WARNING, "bonding index out of range, "
1887 				"please increase MLX5_BOND_MAX_PORTS: %s",
1888 				tmp_str);
1889 			break;
1890 		}
1891 		/* Get ifindex. */
1892 		snprintf(tmp_str, sizeof(tmp_str),
1893 			 "/sys/class/net/%s/ifindex", ifname);
1894 		file = fopen(tmp_str, "rb");
1895 		if (!file)
1896 			break;
1897 		ret = fscanf(file, "%u", &ifindex);
1898 		fclose(file);
1899 		if (ret != 1)
1900 			break;
1901 		/* Save bonding info. */
1902 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1903 			sizeof(bond_info->ports[0].ifname));
1904 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1905 		bond_info->ports[info.port_name].ifindex = ifindex;
1906 		bond_info->n_port++;
1907 		/*
1908 		 * Under socket direct mode, bonding will use
1909 		 * system_image_guid as identification.
1910 		 * After OFED 5.4, guid is readable (ret >= 0) under sysfs.
1911 		 * All bonding members should have the same guid even if driver
1912 		 * is using PCIe BDF.
1913 		 */
1914 		ret = mlx5_get_device_guid(&pci_addr, guid, sizeof(guid));
1915 		if (ret < 0)
1916 			break;
1917 		else if (ret > 0) {
1918 			if (!memcmp(guid, cur_guid, sizeof(guid)) &&
1919 			    owner == info.port_name &&
1920 			    (owner != 0 || (owner == 0 &&
1921 			    !rte_pci_addr_cmp(pci_dev, &pci_addr))))
1922 				pf = info.port_name;
1923 		} else if (pci_dev->domain == pci_addr.domain &&
1924 		    pci_dev->bus == pci_addr.bus &&
1925 		    pci_dev->devid == pci_addr.devid &&
1926 		    ((pci_dev->function == 0 &&
1927 		      pci_dev->function + owner == pci_addr.function) ||
1928 		     (pci_dev->function == owner &&
1929 		      pci_addr.function == owner)))
1930 			pf = info.port_name;
1931 	}
1932 	if (pf >= 0) {
1933 		/* Get bond interface info */
1934 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1935 					   bond_info->ifname);
1936 		if (ret)
1937 			DRV_LOG(ERR, "unable to get bond info: %s",
1938 				strerror(rte_errno));
1939 		else
1940 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1941 				ifindex, bond_info->ifindex, bond_info->ifname);
1942 	}
1943 	if (owner == 0 && pf != 0) {
1944 		DRV_LOG(INFO, "PCIe instance %04x:%02x:%02x.%x isn't bonding owner",
1945 				pci_dev->domain, pci_dev->bus, pci_dev->devid,
1946 				pci_dev->function);
1947 	}
1948 	return pf;
1949 }
1950 
1951 static void
1952 mlx5_os_config_default(struct mlx5_dev_config *config,
1953 		       struct mlx5_common_dev_config *cconf)
1954 {
1955 	memset(config, 0, sizeof(*config));
1956 	config->mps = MLX5_ARG_UNSET;
1957 	config->cqe_comp = 1;
1958 	config->rx_vec_en = 1;
1959 	config->txq_inline_max = MLX5_ARG_UNSET;
1960 	config->txq_inline_min = MLX5_ARG_UNSET;
1961 	config->txq_inline_mpw = MLX5_ARG_UNSET;
1962 	config->txqs_inline = MLX5_ARG_UNSET;
1963 	config->vf_nl_en = 1;
1964 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1965 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1966 	config->mprq.log_min_stride_wqe_size = cconf->devx ?
1967 					cconf->hca_attr.log_min_stride_wqe_sz :
1968 					MLX5_MPRQ_LOG_MIN_STRIDE_WQE_SIZE;
1969 	config->mprq.log_stride_num = MLX5_MPRQ_DEFAULT_LOG_STRIDE_NUM;
1970 	config->dv_esw_en = 1;
1971 	config->dv_flow_en = 1;
1972 	config->decap_en = 1;
1973 	config->log_hp_size = MLX5_ARG_UNSET;
1974 	config->allow_duplicate_pattern = 1;
1975 	config->std_delay_drop = 0;
1976 	config->hp_delay_drop = 0;
1977 }
1978 
1979 /**
1980  * Register a PCI device within bonding.
1981  *
1982  * This function spawns Ethernet devices out of a given PCI device and
1983  * bonding owner PF index.
1984  *
1985  * @param[in] cdev
1986  *   Pointer to common mlx5 device structure.
1987  * @param[in] req_eth_da
1988  *   Requested ethdev device argument.
1989  * @param[in] owner_id
1990  *   Requested owner PF port ID within bonding device, default to 0.
1991  *
1992  * @return
1993  *   0 on success, a negative errno value otherwise and rte_errno is set.
1994  */
1995 static int
1996 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
1997 		     struct rte_eth_devargs *req_eth_da,
1998 		     uint16_t owner_id)
1999 {
2000 	struct ibv_device **ibv_list;
2001 	/*
2002 	 * Number of found IB Devices matching with requested PCI BDF.
2003 	 * nd != 1 means there are multiple IB devices over the same
2004 	 * PCI device and we have representors and master.
2005 	 */
2006 	unsigned int nd = 0;
2007 	/*
2008 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2009 	 * we have the single multiport IB device, and there may be
2010 	 * representors attached to some of found ports.
2011 	 */
2012 	unsigned int np = 0;
2013 	/*
2014 	 * Number of DPDK ethernet devices to Spawn - either over
2015 	 * multiple IB devices or multiple ports of single IB device.
2016 	 * Actually this is the number of iterations to spawn.
2017 	 */
2018 	unsigned int ns = 0;
2019 	/*
2020 	 * Bonding device
2021 	 *   < 0 - no bonding device (single one)
2022 	 *  >= 0 - bonding device (value is slave PF index)
2023 	 */
2024 	int bd = -1;
2025 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2026 	struct mlx5_dev_spawn_data *list = NULL;
2027 	struct mlx5_dev_config dev_config;
2028 	struct rte_eth_devargs eth_da = *req_eth_da;
2029 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2030 	struct mlx5_bond_info bond_info;
2031 	int ret = -1;
2032 
2033 	errno = 0;
2034 	ibv_list = mlx5_glue->get_device_list(&ret);
2035 	if (!ibv_list) {
2036 		rte_errno = errno ? errno : ENOSYS;
2037 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2038 		return -rte_errno;
2039 	}
2040 	/*
2041 	 * First scan the list of all Infiniband devices to find
2042 	 * matching ones, gathering into the list.
2043 	 */
2044 	struct ibv_device *ibv_match[ret + 1];
2045 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2046 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2047 	unsigned int i;
2048 
2049 	while (ret-- > 0) {
2050 		struct rte_pci_addr pci_addr;
2051 
2052 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2053 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2054 						nl_rdma, owner_id, &bond_info);
2055 		if (bd >= 0) {
2056 			/*
2057 			 * Bonding device detected. Only one match is allowed,
2058 			 * the bonding is supported over multi-port IB device,
2059 			 * there should be no matches on representor PCI
2060 			 * functions or non VF LAG bonding devices with
2061 			 * specified address.
2062 			 */
2063 			if (nd) {
2064 				DRV_LOG(ERR,
2065 					"multiple PCI match on bonding device"
2066 					"\"%s\" found", ibv_list[ret]->name);
2067 				rte_errno = ENOENT;
2068 				ret = -rte_errno;
2069 				goto exit;
2070 			}
2071 			/* Amend owner pci address if owner PF ID specified. */
2072 			if (eth_da.nb_representor_ports)
2073 				owner_pci.function += owner_id;
2074 			DRV_LOG(INFO,
2075 				"PCI information matches for slave %d bonding device \"%s\"",
2076 				bd, ibv_list[ret]->name);
2077 			ibv_match[nd++] = ibv_list[ret];
2078 			break;
2079 		} else {
2080 			/* Bonding device not found. */
2081 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2082 					      &pci_addr))
2083 				continue;
2084 			if (owner_pci.domain != pci_addr.domain ||
2085 			    owner_pci.bus != pci_addr.bus ||
2086 			    owner_pci.devid != pci_addr.devid ||
2087 			    owner_pci.function != pci_addr.function)
2088 				continue;
2089 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2090 				ibv_list[ret]->name);
2091 			ibv_match[nd++] = ibv_list[ret];
2092 		}
2093 	}
2094 	ibv_match[nd] = NULL;
2095 	if (!nd) {
2096 		/* No device matches, just complain and bail out. */
2097 		DRV_LOG(WARNING,
2098 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
2099 			" are kernel drivers loaded?",
2100 			owner_pci.domain, owner_pci.bus,
2101 			owner_pci.devid, owner_pci.function);
2102 		rte_errno = ENOENT;
2103 		ret = -rte_errno;
2104 		goto exit;
2105 	}
2106 	if (nd == 1) {
2107 		/*
2108 		 * Found single matching device may have multiple ports.
2109 		 * Each port may be representor, we have to check the port
2110 		 * number and check the representors existence.
2111 		 */
2112 		if (nl_rdma >= 0)
2113 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2114 		if (!np)
2115 			DRV_LOG(WARNING,
2116 				"Cannot get IB device \"%s\" ports number.",
2117 				ibv_match[0]->name);
2118 		if (bd >= 0 && !np) {
2119 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2120 			rte_errno = ENOENT;
2121 			ret = -rte_errno;
2122 			goto exit;
2123 		}
2124 	}
2125 	/* Now we can determine the maximal amount of devices to be spawned. */
2126 	list = mlx5_malloc(MLX5_MEM_ZERO,
2127 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2128 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2129 	if (!list) {
2130 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2131 		rte_errno = ENOMEM;
2132 		ret = -rte_errno;
2133 		goto exit;
2134 	}
2135 	if (bd >= 0 || np > 1) {
2136 		/*
2137 		 * Single IB device with multiple ports found,
2138 		 * it may be E-Switch master device and representors.
2139 		 * We have to perform identification through the ports.
2140 		 */
2141 		MLX5_ASSERT(nl_rdma >= 0);
2142 		MLX5_ASSERT(ns == 0);
2143 		MLX5_ASSERT(nd == 1);
2144 		MLX5_ASSERT(np);
2145 		for (i = 1; i <= np; ++i) {
2146 			list[ns].bond_info = &bond_info;
2147 			list[ns].max_port = np;
2148 			list[ns].phys_port = i;
2149 			list[ns].phys_dev_name = ibv_match[0]->name;
2150 			list[ns].eth_dev = NULL;
2151 			list[ns].pci_dev = pci_dev;
2152 			list[ns].cdev = cdev;
2153 			list[ns].pf_bond = bd;
2154 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2155 							   ibv_match[0]->name,
2156 							   i);
2157 			if (!list[ns].ifindex) {
2158 				/*
2159 				 * No network interface index found for the
2160 				 * specified port, it means there is no
2161 				 * representor on this port. It's OK,
2162 				 * there can be disabled ports, for example
2163 				 * if sriov_numvfs < sriov_totalvfs.
2164 				 */
2165 				continue;
2166 			}
2167 			ret = -1;
2168 			if (nl_route >= 0)
2169 				ret = mlx5_nl_switch_info(nl_route,
2170 							  list[ns].ifindex,
2171 							  &list[ns].info);
2172 			if (ret || (!list[ns].info.representor &&
2173 				    !list[ns].info.master)) {
2174 				/*
2175 				 * We failed to recognize representors with
2176 				 * Netlink, let's try to perform the task
2177 				 * with sysfs.
2178 				 */
2179 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2180 							     &list[ns].info);
2181 			}
2182 			if (!ret && bd >= 0) {
2183 				switch (list[ns].info.name_type) {
2184 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2185 					if (np == 1) {
2186 						/*
2187 						 * Force standalone bonding
2188 						 * device for ROCE LAG
2189 						 * configurations.
2190 						 */
2191 						list[ns].info.master = 0;
2192 						list[ns].info.representor = 0;
2193 					}
2194 					if (list[ns].info.port_name == bd)
2195 						ns++;
2196 					break;
2197 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2198 					/* Fallthrough */
2199 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2200 					/* Fallthrough */
2201 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2202 					if (list[ns].info.pf_num == bd)
2203 						ns++;
2204 					break;
2205 				default:
2206 					break;
2207 				}
2208 				continue;
2209 			}
2210 			if (!ret && (list[ns].info.representor ^
2211 				     list[ns].info.master))
2212 				ns++;
2213 		}
2214 		if (!ns) {
2215 			DRV_LOG(ERR,
2216 				"Unable to recognize master/representors on the IB device with multiple ports.");
2217 			rte_errno = ENOENT;
2218 			ret = -rte_errno;
2219 			goto exit;
2220 		}
2221 	} else {
2222 		/*
2223 		 * The existence of several matching entries (nd > 1) means
2224 		 * port representors have been instantiated. No existing Verbs
2225 		 * call nor sysfs entries can tell them apart, this can only
2226 		 * be done through Netlink calls assuming kernel drivers are
2227 		 * recent enough to support them.
2228 		 *
2229 		 * In the event of identification failure through Netlink,
2230 		 * try again through sysfs, then:
2231 		 *
2232 		 * 1. A single IB device matches (nd == 1) with single
2233 		 *    port (np=0/1) and is not a representor, assume
2234 		 *    no switch support.
2235 		 *
2236 		 * 2. Otherwise no safe assumptions can be made;
2237 		 *    complain louder and bail out.
2238 		 */
2239 		for (i = 0; i != nd; ++i) {
2240 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2241 			list[ns].bond_info = NULL;
2242 			list[ns].max_port = 1;
2243 			list[ns].phys_port = 1;
2244 			list[ns].phys_dev_name = ibv_match[i]->name;
2245 			list[ns].eth_dev = NULL;
2246 			list[ns].pci_dev = pci_dev;
2247 			list[ns].cdev = cdev;
2248 			list[ns].pf_bond = -1;
2249 			list[ns].ifindex = 0;
2250 			if (nl_rdma >= 0)
2251 				list[ns].ifindex = mlx5_nl_ifindex
2252 							    (nl_rdma,
2253 							     ibv_match[i]->name,
2254 							     1);
2255 			if (!list[ns].ifindex) {
2256 				char ifname[IF_NAMESIZE];
2257 
2258 				/*
2259 				 * Netlink failed, it may happen with old
2260 				 * ib_core kernel driver (before 4.16).
2261 				 * We can assume there is old driver because
2262 				 * here we are processing single ports IB
2263 				 * devices. Let's try sysfs to retrieve
2264 				 * the ifindex. The method works for
2265 				 * master device only.
2266 				 */
2267 				if (nd > 1) {
2268 					/*
2269 					 * Multiple devices found, assume
2270 					 * representors, can not distinguish
2271 					 * master/representor and retrieve
2272 					 * ifindex via sysfs.
2273 					 */
2274 					continue;
2275 				}
2276 				ret = mlx5_get_ifname_sysfs
2277 					(ibv_match[i]->ibdev_path, ifname);
2278 				if (!ret)
2279 					list[ns].ifindex =
2280 						if_nametoindex(ifname);
2281 				if (!list[ns].ifindex) {
2282 					/*
2283 					 * No network interface index found
2284 					 * for the specified device, it means
2285 					 * there it is neither representor
2286 					 * nor master.
2287 					 */
2288 					continue;
2289 				}
2290 			}
2291 			ret = -1;
2292 			if (nl_route >= 0)
2293 				ret = mlx5_nl_switch_info(nl_route,
2294 							  list[ns].ifindex,
2295 							  &list[ns].info);
2296 			if (ret || (!list[ns].info.representor &&
2297 				    !list[ns].info.master)) {
2298 				/*
2299 				 * We failed to recognize representors with
2300 				 * Netlink, let's try to perform the task
2301 				 * with sysfs.
2302 				 */
2303 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2304 							     &list[ns].info);
2305 			}
2306 			if (!ret && (list[ns].info.representor ^
2307 				     list[ns].info.master)) {
2308 				ns++;
2309 			} else if ((nd == 1) &&
2310 				   !list[ns].info.representor &&
2311 				   !list[ns].info.master) {
2312 				/*
2313 				 * Single IB device with one physical port and
2314 				 * attached network device.
2315 				 * May be SRIOV is not enabled or there is no
2316 				 * representors.
2317 				 */
2318 				DRV_LOG(INFO, "No E-Switch support detected.");
2319 				ns++;
2320 				break;
2321 			}
2322 		}
2323 		if (!ns) {
2324 			DRV_LOG(ERR,
2325 				"Unable to recognize master/representors on the multiple IB devices.");
2326 			rte_errno = ENOENT;
2327 			ret = -rte_errno;
2328 			goto exit;
2329 		}
2330 		/*
2331 		 * New kernels may add the switch_id attribute for the case
2332 		 * there is no E-Switch and we wrongly recognized the only
2333 		 * device as master. Override this if there is the single
2334 		 * device with single port and new device name format present.
2335 		 */
2336 		if (nd == 1 &&
2337 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2338 			list[0].info.master = 0;
2339 			list[0].info.representor = 0;
2340 		}
2341 	}
2342 	MLX5_ASSERT(ns);
2343 	/*
2344 	 * Sort list to probe devices in natural order for users convenience
2345 	 * (i.e. master first, then representors from lowest to highest ID).
2346 	 */
2347 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2348 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2349 		/* Set devargs default values. */
2350 		if (eth_da.nb_mh_controllers == 0) {
2351 			eth_da.nb_mh_controllers = 1;
2352 			eth_da.mh_controllers[0] = 0;
2353 		}
2354 		if (eth_da.nb_ports == 0 && ns > 0) {
2355 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2356 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2357 					pci_dev->device.devargs->args);
2358 			eth_da.nb_ports = 1;
2359 			eth_da.ports[0] = list[0].info.pf_num;
2360 		}
2361 		if (eth_da.nb_representor_ports == 0) {
2362 			eth_da.nb_representor_ports = 1;
2363 			eth_da.representor_ports[0] = 0;
2364 		}
2365 	}
2366 	for (i = 0; i != ns; ++i) {
2367 		uint32_t restore;
2368 
2369 		/* Default configuration. */
2370 		mlx5_os_config_default(&dev_config, &cdev->config);
2371 		dev_config.vf = mlx5_dev_is_vf_pci(pci_dev);
2372 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i],
2373 						 &dev_config, &eth_da);
2374 		if (!list[i].eth_dev) {
2375 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2376 				break;
2377 			/* Device is disabled or already spawned. Ignore it. */
2378 			continue;
2379 		}
2380 		restore = list[i].eth_dev->data->dev_flags;
2381 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2382 		/**
2383 		 * Each representor has a dedicated interrupts vector.
2384 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2385 		 * representor eth_dev object because representor and PF
2386 		 * share the same PCI address.
2387 		 * Override representor device with a dedicated
2388 		 * interrupts handle here.
2389 		 * Representor interrupts handle is released in mlx5_dev_stop().
2390 		 */
2391 		if (list[i].info.representor) {
2392 			struct rte_intr_handle *intr_handle =
2393 				rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2394 			if (intr_handle == NULL) {
2395 				DRV_LOG(ERR,
2396 					"port %u failed to allocate memory for interrupt handler "
2397 					"Rx interrupts will not be supported",
2398 					i);
2399 				rte_errno = ENOMEM;
2400 				ret = -rte_errno;
2401 				goto exit;
2402 			}
2403 			list[i].eth_dev->intr_handle = intr_handle;
2404 		}
2405 		/* Restore non-PCI flags cleared by the above call. */
2406 		list[i].eth_dev->data->dev_flags |= restore;
2407 		rte_eth_dev_probing_finish(list[i].eth_dev);
2408 	}
2409 	if (i != ns) {
2410 		DRV_LOG(ERR,
2411 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2412 			" encountering an error: %s",
2413 			owner_pci.domain, owner_pci.bus,
2414 			owner_pci.devid, owner_pci.function,
2415 			strerror(rte_errno));
2416 		ret = -rte_errno;
2417 		/* Roll back. */
2418 		while (i--) {
2419 			if (!list[i].eth_dev)
2420 				continue;
2421 			mlx5_dev_close(list[i].eth_dev);
2422 			/* mac_addrs must not be freed because in dev_private */
2423 			list[i].eth_dev->data->mac_addrs = NULL;
2424 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2425 		}
2426 		/* Restore original error. */
2427 		rte_errno = -ret;
2428 	} else {
2429 		ret = 0;
2430 	}
2431 exit:
2432 	/*
2433 	 * Do the routine cleanup:
2434 	 * - close opened Netlink sockets
2435 	 * - free allocated spawn data array
2436 	 * - free the Infiniband device list
2437 	 */
2438 	if (nl_rdma >= 0)
2439 		close(nl_rdma);
2440 	if (nl_route >= 0)
2441 		close(nl_route);
2442 	if (list)
2443 		mlx5_free(list);
2444 	MLX5_ASSERT(ibv_list);
2445 	mlx5_glue->free_device_list(ibv_list);
2446 	return ret;
2447 }
2448 
2449 static int
2450 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2451 			  struct rte_eth_devargs *eth_da)
2452 {
2453 	int ret = 0;
2454 
2455 	if (dev->devargs == NULL)
2456 		return 0;
2457 	memset(eth_da, 0, sizeof(*eth_da));
2458 	/* Parse representor information first from class argument. */
2459 	if (dev->devargs->cls_str)
2460 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2461 	if (ret != 0) {
2462 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2463 			dev->devargs->cls_str);
2464 		return -rte_errno;
2465 	}
2466 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2467 		/* Parse legacy device argument */
2468 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2469 		if (ret) {
2470 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2471 				dev->devargs->args);
2472 			return -rte_errno;
2473 		}
2474 	}
2475 	return 0;
2476 }
2477 
2478 /**
2479  * Callback to register a PCI device.
2480  *
2481  * This function spawns Ethernet devices out of a given PCI device.
2482  *
2483  * @param[in] cdev
2484  *   Pointer to common mlx5 device structure.
2485  *
2486  * @return
2487  *   0 on success, a negative errno value otherwise and rte_errno is set.
2488  */
2489 static int
2490 mlx5_os_pci_probe(struct mlx5_common_device *cdev)
2491 {
2492 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2493 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2494 	int ret = 0;
2495 	uint16_t p;
2496 
2497 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2498 	if (ret != 0)
2499 		return ret;
2500 
2501 	if (eth_da.nb_ports > 0) {
2502 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2503 		for (p = 0; p < eth_da.nb_ports; p++) {
2504 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2505 						   eth_da.ports[p]);
2506 			if (ret)
2507 				break;
2508 		}
2509 		if (ret) {
2510 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2511 				"aborted due to prodding failure of PF %u",
2512 				pci_dev->addr.domain, pci_dev->addr.bus,
2513 				pci_dev->addr.devid, pci_dev->addr.function,
2514 				eth_da.ports[p]);
2515 			mlx5_net_remove(cdev);
2516 		}
2517 	} else {
2518 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0);
2519 	}
2520 	return ret;
2521 }
2522 
2523 /* Probe a single SF device on auxiliary bus, no representor support. */
2524 static int
2525 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev)
2526 {
2527 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2528 	struct mlx5_dev_config config;
2529 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2530 	struct rte_device *dev = cdev->dev;
2531 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2532 	struct rte_eth_dev *eth_dev;
2533 	int ret = 0;
2534 
2535 	/* Parse ethdev devargs. */
2536 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2537 	if (ret != 0)
2538 		return ret;
2539 	/* Set default config data. */
2540 	mlx5_os_config_default(&config, &cdev->config);
2541 	config.sf = 1;
2542 	/* Init spawn data. */
2543 	spawn.max_port = 1;
2544 	spawn.phys_port = 1;
2545 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2546 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2547 	if (ret < 0) {
2548 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2549 		return ret;
2550 	}
2551 	spawn.ifindex = ret;
2552 	spawn.cdev = cdev;
2553 	/* Spawn device. */
2554 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2555 	if (eth_dev == NULL)
2556 		return -rte_errno;
2557 	/* Post create. */
2558 	eth_dev->intr_handle = adev->intr_handle;
2559 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2560 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2561 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2562 		eth_dev->data->numa_node = dev->numa_node;
2563 	}
2564 	rte_eth_dev_probing_finish(eth_dev);
2565 	return 0;
2566 }
2567 
2568 /**
2569  * Net class driver callback to probe a device.
2570  *
2571  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2572  *
2573  * @param[in] cdev
2574  *   Pointer to the common mlx5 device.
2575  *
2576  * @return
2577  *   0 on success, a negative errno value otherwise and rte_errno is set.
2578  */
2579 int
2580 mlx5_os_net_probe(struct mlx5_common_device *cdev)
2581 {
2582 	int ret;
2583 
2584 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2585 		mlx5_pmd_socket_init();
2586 	ret = mlx5_init_once();
2587 	if (ret) {
2588 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2589 			strerror(rte_errno));
2590 		return -rte_errno;
2591 	}
2592 	if (mlx5_dev_is_pci(cdev->dev))
2593 		return mlx5_os_pci_probe(cdev);
2594 	else
2595 		return mlx5_os_auxiliary_probe(cdev);
2596 }
2597 
2598 /**
2599  * Cleanup resources when the last device is closed.
2600  */
2601 void
2602 mlx5_os_net_cleanup(void)
2603 {
2604 	mlx5_pmd_socket_uninit();
2605 }
2606 
2607 /**
2608  * Install shared asynchronous device events handler.
2609  * This function is implemented to support event sharing
2610  * between multiple ports of single IB device.
2611  *
2612  * @param sh
2613  *   Pointer to mlx5_dev_ctx_shared object.
2614  */
2615 void
2616 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2617 {
2618 	int ret;
2619 	int flags;
2620 	struct ibv_context *ctx = sh->cdev->ctx;
2621 
2622 	sh->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2623 	if (sh->intr_handle == NULL) {
2624 		DRV_LOG(ERR, "Fail to allocate intr_handle");
2625 		rte_errno = ENOMEM;
2626 		return;
2627 	}
2628 	rte_intr_fd_set(sh->intr_handle, -1);
2629 
2630 	flags = fcntl(ctx->async_fd, F_GETFL);
2631 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2632 	if (ret) {
2633 		DRV_LOG(INFO, "failed to change file descriptor async event"
2634 			" queue");
2635 	} else {
2636 		rte_intr_fd_set(sh->intr_handle, ctx->async_fd);
2637 		rte_intr_type_set(sh->intr_handle, RTE_INTR_HANDLE_EXT);
2638 		if (rte_intr_callback_register(sh->intr_handle,
2639 					mlx5_dev_interrupt_handler, sh)) {
2640 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2641 			rte_intr_fd_set(sh->intr_handle, -1);
2642 		}
2643 	}
2644 	if (sh->cdev->config.devx) {
2645 #ifdef HAVE_IBV_DEVX_ASYNC
2646 		sh->intr_handle_devx =
2647 			rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
2648 		if (!sh->intr_handle_devx) {
2649 			DRV_LOG(ERR, "Fail to allocate intr_handle");
2650 			rte_errno = ENOMEM;
2651 			return;
2652 		}
2653 		rte_intr_fd_set(sh->intr_handle_devx, -1);
2654 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2655 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2656 		if (!devx_comp) {
2657 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2658 			return;
2659 		}
2660 		flags = fcntl(devx_comp->fd, F_GETFL);
2661 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2662 		if (ret) {
2663 			DRV_LOG(INFO, "failed to change file descriptor"
2664 				" devx comp");
2665 			return;
2666 		}
2667 		rte_intr_fd_set(sh->intr_handle_devx, devx_comp->fd);
2668 		rte_intr_type_set(sh->intr_handle_devx,
2669 					 RTE_INTR_HANDLE_EXT);
2670 		if (rte_intr_callback_register(sh->intr_handle_devx,
2671 					mlx5_dev_interrupt_handler_devx, sh)) {
2672 			DRV_LOG(INFO, "Fail to install the devx shared"
2673 				" interrupt.");
2674 			rte_intr_fd_set(sh->intr_handle_devx, -1);
2675 		}
2676 #endif /* HAVE_IBV_DEVX_ASYNC */
2677 	}
2678 }
2679 
2680 /**
2681  * Uninstall shared asynchronous device events handler.
2682  * This function is implemented to support event sharing
2683  * between multiple ports of single IB device.
2684  *
2685  * @param dev
2686  *   Pointer to mlx5_dev_ctx_shared object.
2687  */
2688 void
2689 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2690 {
2691 	if (rte_intr_fd_get(sh->intr_handle) >= 0)
2692 		mlx5_intr_callback_unregister(sh->intr_handle,
2693 					      mlx5_dev_interrupt_handler, sh);
2694 	rte_intr_instance_free(sh->intr_handle);
2695 #ifdef HAVE_IBV_DEVX_ASYNC
2696 	if (rte_intr_fd_get(sh->intr_handle_devx) >= 0)
2697 		rte_intr_callback_unregister(sh->intr_handle_devx,
2698 				  mlx5_dev_interrupt_handler_devx, sh);
2699 	rte_intr_instance_free(sh->intr_handle_devx);
2700 	if (sh->devx_comp)
2701 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2702 #endif
2703 }
2704 
2705 /**
2706  * Read statistics by a named counter.
2707  *
2708  * @param[in] priv
2709  *   Pointer to the private device data structure.
2710  * @param[in] ctr_name
2711  *   Pointer to the name of the statistic counter to read
2712  * @param[out] stat
2713  *   Pointer to read statistic value.
2714  * @return
2715  *   0 on success and stat is valud, 1 if failed to read the value
2716  *   rte_errno is set.
2717  *
2718  */
2719 int
2720 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2721 		      uint64_t *stat)
2722 {
2723 	int fd;
2724 
2725 	if (priv->sh) {
2726 		if (priv->q_counters != NULL &&
2727 		    strcmp(ctr_name, "out_of_buffer") == 0)
2728 			return mlx5_devx_cmd_queue_counter_query
2729 					(priv->q_counters, 0, (uint32_t *)stat);
2730 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2731 		      priv->sh->ibdev_path,
2732 		      priv->dev_port,
2733 		      ctr_name);
2734 		fd = open(path, O_RDONLY);
2735 		/*
2736 		 * in switchdev the file location is not per port
2737 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2738 		 */
2739 		if (fd == -1) {
2740 			MKSTR(path1, "%s/hw_counters/%s",
2741 			      priv->sh->ibdev_path,
2742 			      ctr_name);
2743 			fd = open(path1, O_RDONLY);
2744 		}
2745 		if (fd != -1) {
2746 			char buf[21] = {'\0'};
2747 			ssize_t n = read(fd, buf, sizeof(buf));
2748 
2749 			close(fd);
2750 			if (n != -1) {
2751 				*stat = strtoull(buf, NULL, 10);
2752 				return 0;
2753 			}
2754 		}
2755 	}
2756 	*stat = 0;
2757 	return 1;
2758 }
2759 
2760 /**
2761  * Remove a MAC address from device
2762  *
2763  * @param dev
2764  *   Pointer to Ethernet device structure.
2765  * @param index
2766  *   MAC address index.
2767  */
2768 void
2769 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2770 {
2771 	struct mlx5_priv *priv = dev->data->dev_private;
2772 	const int vf = priv->config.vf;
2773 
2774 	if (vf)
2775 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2776 					mlx5_ifindex(dev), priv->mac_own,
2777 					&dev->data->mac_addrs[index], index);
2778 }
2779 
2780 /**
2781  * Adds a MAC address to the device
2782  *
2783  * @param dev
2784  *   Pointer to Ethernet device structure.
2785  * @param mac_addr
2786  *   MAC address to register.
2787  * @param index
2788  *   MAC address index.
2789  *
2790  * @return
2791  *   0 on success, a negative errno value otherwise
2792  */
2793 int
2794 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2795 		     uint32_t index)
2796 {
2797 	struct mlx5_priv *priv = dev->data->dev_private;
2798 	const int vf = priv->config.vf;
2799 	int ret = 0;
2800 
2801 	if (vf)
2802 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2803 					   mlx5_ifindex(dev), priv->mac_own,
2804 					   mac, index);
2805 	return ret;
2806 }
2807 
2808 /**
2809  * Modify a VF MAC address
2810  *
2811  * @param priv
2812  *   Pointer to device private data.
2813  * @param mac_addr
2814  *   MAC address to modify into.
2815  * @param iface_idx
2816  *   Net device interface index
2817  * @param vf_index
2818  *   VF index
2819  *
2820  * @return
2821  *   0 on success, a negative errno value otherwise
2822  */
2823 int
2824 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2825 			   unsigned int iface_idx,
2826 			   struct rte_ether_addr *mac_addr,
2827 			   int vf_index)
2828 {
2829 	return mlx5_nl_vf_mac_addr_modify
2830 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2831 }
2832 
2833 /**
2834  * Set device promiscuous mode
2835  *
2836  * @param dev
2837  *   Pointer to Ethernet device structure.
2838  * @param enable
2839  *   0 - promiscuous is disabled, otherwise - enabled
2840  *
2841  * @return
2842  *   0 on success, a negative error value otherwise
2843  */
2844 int
2845 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2846 {
2847 	struct mlx5_priv *priv = dev->data->dev_private;
2848 
2849 	return mlx5_nl_promisc(priv->nl_socket_route,
2850 			       mlx5_ifindex(dev), !!enable);
2851 }
2852 
2853 /**
2854  * Set device promiscuous mode
2855  *
2856  * @param dev
2857  *   Pointer to Ethernet device structure.
2858  * @param enable
2859  *   0 - all multicase is disabled, otherwise - enabled
2860  *
2861  * @return
2862  *   0 on success, a negative error value otherwise
2863  */
2864 int
2865 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2866 {
2867 	struct mlx5_priv *priv = dev->data->dev_private;
2868 
2869 	return mlx5_nl_allmulti(priv->nl_socket_route,
2870 				mlx5_ifindex(dev), !!enable);
2871 }
2872 
2873 /**
2874  * Flush device MAC addresses
2875  *
2876  * @param dev
2877  *   Pointer to Ethernet device structure.
2878  *
2879  */
2880 void
2881 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2882 {
2883 	struct mlx5_priv *priv = dev->data->dev_private;
2884 
2885 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2886 			       dev->data->mac_addrs,
2887 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2888 }
2889