xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2020 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
16 #include <fcntl.h>
17 
18 #include <rte_malloc.h>
19 #include <ethdev_driver.h>
20 #include <ethdev_pci.h>
21 #include <rte_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_bus_auxiliary.h>
24 #include <rte_common.h>
25 #include <rte_kvargs.h>
26 #include <rte_rwlock.h>
27 #include <rte_spinlock.h>
28 #include <rte_string_fns.h>
29 #include <rte_alarm.h>
30 #include <rte_eal_paging.h>
31 
32 #include <mlx5_glue.h>
33 #include <mlx5_devx_cmds.h>
34 #include <mlx5_common.h>
35 #include <mlx5_common_mp.h>
36 #include <mlx5_common_mr.h>
37 #include <mlx5_malloc.h>
38 
39 #include "mlx5_defs.h"
40 #include "mlx5.h"
41 #include "mlx5_common_os.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_rx.h"
45 #include "mlx5_tx.h"
46 #include "mlx5_autoconf.h"
47 #include "mlx5_flow.h"
48 #include "rte_pmd_mlx5.h"
49 #include "mlx5_verbs.h"
50 #include "mlx5_nl.h"
51 #include "mlx5_devx.h"
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /* rte flow indexed pool configuration. */
71 static struct mlx5_indexed_pool_config icfg[] = {
72 	{
73 		.size = sizeof(struct rte_flow),
74 		.trunk_size = 64,
75 		.need_lock = 1,
76 		.release_mem_en = 0,
77 		.malloc = mlx5_malloc,
78 		.free = mlx5_free,
79 		.per_core_cache = 0,
80 		.type = "ctl_flow_ipool",
81 	},
82 	{
83 		.size = sizeof(struct rte_flow),
84 		.trunk_size = 64,
85 		.grow_trunk = 3,
86 		.grow_shift = 2,
87 		.need_lock = 1,
88 		.release_mem_en = 0,
89 		.malloc = mlx5_malloc,
90 		.free = mlx5_free,
91 		.per_core_cache = 1 << 14,
92 		.type = "rte_flow_ipool",
93 	},
94 	{
95 		.size = sizeof(struct rte_flow),
96 		.trunk_size = 64,
97 		.grow_trunk = 3,
98 		.grow_shift = 2,
99 		.need_lock = 1,
100 		.release_mem_en = 0,
101 		.malloc = mlx5_malloc,
102 		.free = mlx5_free,
103 		.per_core_cache = 0,
104 		.type = "mcp_flow_ipool",
105 	},
106 };
107 
108 /**
109  * Set the completion channel file descriptor interrupt as non-blocking.
110  *
111  * @param[in] rxq_obj
112  *   Pointer to RQ channel object, which includes the channel fd
113  *
114  * @param[out] fd
115  *   The file descriptor (representing the intetrrupt) used in this channel.
116  *
117  * @return
118  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
119  */
120 int
121 mlx5_os_set_nonblock_channel_fd(int fd)
122 {
123 	int flags;
124 
125 	flags = fcntl(fd, F_GETFL);
126 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
127 }
128 
129 /**
130  * Get mlx5 device attributes. The glue function query_device_ex() is called
131  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
132  * device attributes from the glue out parameter.
133  *
134  * @param cdev
135  *   Pointer to mlx5 device.
136  *
137  * @param device_attr
138  *   Pointer to mlx5 device attributes.
139  *
140  * @return
141  *   0 on success, non zero error number otherwise
142  */
143 int
144 mlx5_os_get_dev_attr(struct mlx5_common_device *cdev,
145 		     struct mlx5_dev_attr *device_attr)
146 {
147 	int err;
148 	struct ibv_context *ctx = cdev->ctx;
149 	struct ibv_device_attr_ex attr_ex;
150 
151 	memset(device_attr, 0, sizeof(*device_attr));
152 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
153 	if (err)
154 		return err;
155 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
156 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
157 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
158 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
159 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
160 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
161 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
162 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
163 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
164 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
165 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
166 	device_attr->max_rwq_indirection_table_size =
167 		attr_ex.rss_caps.max_rwq_indirection_table_size;
168 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
169 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
170 
171 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
172 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
173 	if (err)
174 		return err;
175 
176 	device_attr->flags = dv_attr.flags;
177 	device_attr->comp_mask = dv_attr.comp_mask;
178 #ifdef HAVE_IBV_MLX5_MOD_SWP
179 	device_attr->sw_parsing_offloads =
180 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
181 #endif
182 	device_attr->min_single_stride_log_num_of_bytes =
183 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
184 	device_attr->max_single_stride_log_num_of_bytes =
185 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
186 	device_attr->min_single_wqe_log_num_of_strides =
187 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
188 	device_attr->max_single_wqe_log_num_of_strides =
189 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
190 	device_attr->stride_supported_qpts =
191 		dv_attr.striding_rq_caps.supported_qpts;
192 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
193 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
194 #endif
195 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
196 		sizeof(device_attr->fw_ver));
197 
198 	return err;
199 }
200 
201 /**
202  * Detect misc5 support or not
203  *
204  * @param[in] priv
205  *   Device private data pointer
206  */
207 #ifdef HAVE_MLX5DV_DR
208 static void
209 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
210 {
211 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
212 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
213 	 * Case: IPv4--->UDP--->VxLAN--->vni
214 	 */
215 	void *tbl;
216 	struct mlx5_flow_dv_match_params matcher_mask;
217 	void *match_m;
218 	void *matcher;
219 	void *headers_m;
220 	void *misc5_m;
221 	uint32_t *tunnel_header_m;
222 	struct mlx5dv_flow_matcher_attr dv_attr;
223 
224 	memset(&matcher_mask, 0, sizeof(matcher_mask));
225 	matcher_mask.size = sizeof(matcher_mask.buf);
226 	match_m = matcher_mask.buf;
227 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
228 	misc5_m = MLX5_ADDR_OF(fte_match_param,
229 			       match_m, misc_parameters_5);
230 	tunnel_header_m = (uint32_t *)
231 				MLX5_ADDR_OF(fte_match_set_misc5,
232 				misc5_m, tunnel_header_1);
233 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
234 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
235 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
236 	*tunnel_header_m = 0xffffff;
237 
238 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
239 	if (!tbl) {
240 		DRV_LOG(INFO, "No SW steering support");
241 		return;
242 	}
243 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
244 	dv_attr.match_mask = (void *)&matcher_mask,
245 	dv_attr.match_criteria_enable =
246 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
247 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
248 	dv_attr.priority = 3;
249 #ifdef HAVE_MLX5DV_DR_ESWITCH
250 	void *misc2_m;
251 	if (priv->config.dv_esw_en) {
252 		/* FDB enabled reg_c_0 */
253 		dv_attr.match_criteria_enable |=
254 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
255 		misc2_m = MLX5_ADDR_OF(fte_match_param,
256 				       match_m, misc_parameters_2);
257 		MLX5_SET(fte_match_set_misc2, misc2_m,
258 			 metadata_reg_c_0, 0xffff);
259 	}
260 #endif
261 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
262 						    &dv_attr, tbl);
263 	if (matcher) {
264 		priv->sh->misc5_cap = 1;
265 		mlx5_glue->dv_destroy_flow_matcher(matcher);
266 	}
267 	mlx5_glue->dr_destroy_flow_tbl(tbl);
268 #else
269 	RTE_SET_USED(priv);
270 #endif
271 }
272 #endif
273 
274 /**
275  * Initialize DR related data within private structure.
276  * Routine checks the reference counter and does actual
277  * resources creation/initialization only if counter is zero.
278  *
279  * @param[in] priv
280  *   Pointer to the private device data structure.
281  *
282  * @return
283  *   Zero on success, positive error code otherwise.
284  */
285 static int
286 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
287 {
288 	struct mlx5_dev_ctx_shared *sh = priv->sh;
289 	char s[MLX5_NAME_SIZE] __rte_unused;
290 	int err;
291 
292 	MLX5_ASSERT(sh && sh->refcnt);
293 	if (sh->refcnt > 1)
294 		return 0;
295 	err = mlx5_alloc_table_hash_list(priv);
296 	if (err)
297 		goto error;
298 	/* The resources below are only valid with DV support. */
299 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
300 	/* Init port id action list. */
301 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
302 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
303 						   flow_dv_port_id_create_cb,
304 						   flow_dv_port_id_match_cb,
305 						   flow_dv_port_id_remove_cb,
306 						   flow_dv_port_id_clone_cb,
307 						 flow_dv_port_id_clone_free_cb);
308 	if (!sh->port_id_action_list)
309 		goto error;
310 	/* Init push vlan action list. */
311 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
312 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
313 						    flow_dv_push_vlan_create_cb,
314 						    flow_dv_push_vlan_match_cb,
315 						    flow_dv_push_vlan_remove_cb,
316 						    flow_dv_push_vlan_clone_cb,
317 					       flow_dv_push_vlan_clone_free_cb);
318 	if (!sh->push_vlan_action_list)
319 		goto error;
320 	/* Init sample action list. */
321 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
322 	sh->sample_action_list = mlx5_list_create(s, sh, true,
323 						  flow_dv_sample_create_cb,
324 						  flow_dv_sample_match_cb,
325 						  flow_dv_sample_remove_cb,
326 						  flow_dv_sample_clone_cb,
327 						  flow_dv_sample_clone_free_cb);
328 	if (!sh->sample_action_list)
329 		goto error;
330 	/* Init dest array action list. */
331 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
332 	sh->dest_array_list = mlx5_list_create(s, sh, true,
333 					       flow_dv_dest_array_create_cb,
334 					       flow_dv_dest_array_match_cb,
335 					       flow_dv_dest_array_remove_cb,
336 					       flow_dv_dest_array_clone_cb,
337 					      flow_dv_dest_array_clone_free_cb);
338 	if (!sh->dest_array_list)
339 		goto error;
340 #endif
341 #ifdef HAVE_MLX5DV_DR
342 	void *domain;
343 
344 	/* Reference counter is zero, we should initialize structures. */
345 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
346 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
347 	if (!domain) {
348 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
349 		err = errno;
350 		goto error;
351 	}
352 	sh->rx_domain = domain;
353 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
354 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
355 	if (!domain) {
356 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
357 		err = errno;
358 		goto error;
359 	}
360 	sh->tx_domain = domain;
361 #ifdef HAVE_MLX5DV_DR_ESWITCH
362 	if (priv->config.dv_esw_en) {
363 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
364 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
365 		if (!domain) {
366 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
367 			err = errno;
368 			goto error;
369 		}
370 		sh->fdb_domain = domain;
371 	}
372 	/*
373 	 * The drop action is just some dummy placeholder in rdma-core. It
374 	 * does not belong to domains and has no any attributes, and, can be
375 	 * shared by the entire device.
376 	 */
377 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
378 	if (!sh->dr_drop_action) {
379 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
380 		err = errno;
381 		goto error;
382 	}
383 #endif
384 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
385 		err = mlx5_alloc_tunnel_hub(sh);
386 	if (err) {
387 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
388 		goto error;
389 	}
390 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
391 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
392 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
393 		if (sh->fdb_domain)
394 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
395 	}
396 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
397 	if (!priv->config.allow_duplicate_pattern) {
398 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
399 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
400 #endif
401 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
402 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
403 		if (sh->fdb_domain)
404 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
405 	}
406 
407 	__mlx5_discovery_misc5_cap(priv);
408 #endif /* HAVE_MLX5DV_DR */
409 	sh->default_miss_action =
410 			mlx5_glue->dr_create_flow_action_default_miss();
411 	if (!sh->default_miss_action)
412 		DRV_LOG(WARNING, "Default miss action is not supported.");
413 	return 0;
414 error:
415 	/* Rollback the created objects. */
416 	if (sh->rx_domain) {
417 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
418 		sh->rx_domain = NULL;
419 	}
420 	if (sh->tx_domain) {
421 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
422 		sh->tx_domain = NULL;
423 	}
424 	if (sh->fdb_domain) {
425 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
426 		sh->fdb_domain = NULL;
427 	}
428 	if (sh->dr_drop_action) {
429 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
430 		sh->dr_drop_action = NULL;
431 	}
432 	if (sh->pop_vlan_action) {
433 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
434 		sh->pop_vlan_action = NULL;
435 	}
436 	if (sh->encaps_decaps) {
437 		mlx5_hlist_destroy(sh->encaps_decaps);
438 		sh->encaps_decaps = NULL;
439 	}
440 	if (sh->modify_cmds) {
441 		mlx5_hlist_destroy(sh->modify_cmds);
442 		sh->modify_cmds = NULL;
443 	}
444 	if (sh->tag_table) {
445 		/* tags should be destroyed with flow before. */
446 		mlx5_hlist_destroy(sh->tag_table);
447 		sh->tag_table = NULL;
448 	}
449 	if (sh->tunnel_hub) {
450 		mlx5_release_tunnel_hub(sh, priv->dev_port);
451 		sh->tunnel_hub = NULL;
452 	}
453 	mlx5_free_table_hash_list(priv);
454 	if (sh->port_id_action_list) {
455 		mlx5_list_destroy(sh->port_id_action_list);
456 		sh->port_id_action_list = NULL;
457 	}
458 	if (sh->push_vlan_action_list) {
459 		mlx5_list_destroy(sh->push_vlan_action_list);
460 		sh->push_vlan_action_list = NULL;
461 	}
462 	if (sh->sample_action_list) {
463 		mlx5_list_destroy(sh->sample_action_list);
464 		sh->sample_action_list = NULL;
465 	}
466 	if (sh->dest_array_list) {
467 		mlx5_list_destroy(sh->dest_array_list);
468 		sh->dest_array_list = NULL;
469 	}
470 	return err;
471 }
472 
473 /**
474  * Destroy DR related data within private structure.
475  *
476  * @param[in] priv
477  *   Pointer to the private device data structure.
478  */
479 void
480 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
481 {
482 	struct mlx5_dev_ctx_shared *sh = priv->sh;
483 
484 	MLX5_ASSERT(sh && sh->refcnt);
485 	if (sh->refcnt > 1)
486 		return;
487 #ifdef HAVE_MLX5DV_DR
488 	if (sh->rx_domain) {
489 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
490 		sh->rx_domain = NULL;
491 	}
492 	if (sh->tx_domain) {
493 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
494 		sh->tx_domain = NULL;
495 	}
496 #ifdef HAVE_MLX5DV_DR_ESWITCH
497 	if (sh->fdb_domain) {
498 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
499 		sh->fdb_domain = NULL;
500 	}
501 	if (sh->dr_drop_action) {
502 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
503 		sh->dr_drop_action = NULL;
504 	}
505 #endif
506 	if (sh->pop_vlan_action) {
507 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
508 		sh->pop_vlan_action = NULL;
509 	}
510 #endif /* HAVE_MLX5DV_DR */
511 	if (sh->default_miss_action)
512 		mlx5_glue->destroy_flow_action
513 				(sh->default_miss_action);
514 	if (sh->encaps_decaps) {
515 		mlx5_hlist_destroy(sh->encaps_decaps);
516 		sh->encaps_decaps = NULL;
517 	}
518 	if (sh->modify_cmds) {
519 		mlx5_hlist_destroy(sh->modify_cmds);
520 		sh->modify_cmds = NULL;
521 	}
522 	if (sh->tag_table) {
523 		/* tags should be destroyed with flow before. */
524 		mlx5_hlist_destroy(sh->tag_table);
525 		sh->tag_table = NULL;
526 	}
527 	if (sh->tunnel_hub) {
528 		mlx5_release_tunnel_hub(sh, priv->dev_port);
529 		sh->tunnel_hub = NULL;
530 	}
531 	mlx5_free_table_hash_list(priv);
532 	if (sh->port_id_action_list) {
533 		mlx5_list_destroy(sh->port_id_action_list);
534 		sh->port_id_action_list = NULL;
535 	}
536 	if (sh->push_vlan_action_list) {
537 		mlx5_list_destroy(sh->push_vlan_action_list);
538 		sh->push_vlan_action_list = NULL;
539 	}
540 	if (sh->sample_action_list) {
541 		mlx5_list_destroy(sh->sample_action_list);
542 		sh->sample_action_list = NULL;
543 	}
544 	if (sh->dest_array_list) {
545 		mlx5_list_destroy(sh->dest_array_list);
546 		sh->dest_array_list = NULL;
547 	}
548 }
549 
550 /**
551  * Initialize shared data between primary and secondary process.
552  *
553  * A memzone is reserved by primary process and secondary processes attach to
554  * the memzone.
555  *
556  * @return
557  *   0 on success, a negative errno value otherwise and rte_errno is set.
558  */
559 static int
560 mlx5_init_shared_data(void)
561 {
562 	const struct rte_memzone *mz;
563 	int ret = 0;
564 
565 	rte_spinlock_lock(&mlx5_shared_data_lock);
566 	if (mlx5_shared_data == NULL) {
567 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
568 			/* Allocate shared memory. */
569 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
570 						 sizeof(*mlx5_shared_data),
571 						 SOCKET_ID_ANY, 0);
572 			if (mz == NULL) {
573 				DRV_LOG(ERR,
574 					"Cannot allocate mlx5 shared data");
575 				ret = -rte_errno;
576 				goto error;
577 			}
578 			mlx5_shared_data = mz->addr;
579 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
580 			rte_spinlock_init(&mlx5_shared_data->lock);
581 		} else {
582 			/* Lookup allocated shared memory. */
583 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
584 			if (mz == NULL) {
585 				DRV_LOG(ERR,
586 					"Cannot attach mlx5 shared data");
587 				ret = -rte_errno;
588 				goto error;
589 			}
590 			mlx5_shared_data = mz->addr;
591 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
592 		}
593 	}
594 error:
595 	rte_spinlock_unlock(&mlx5_shared_data_lock);
596 	return ret;
597 }
598 
599 /**
600  * PMD global initialization.
601  *
602  * Independent from individual device, this function initializes global
603  * per-PMD data structures distinguishing primary and secondary processes.
604  * Hence, each initialization is called once per a process.
605  *
606  * @return
607  *   0 on success, a negative errno value otherwise and rte_errno is set.
608  */
609 static int
610 mlx5_init_once(void)
611 {
612 	struct mlx5_shared_data *sd;
613 	struct mlx5_local_data *ld = &mlx5_local_data;
614 	int ret = 0;
615 
616 	if (mlx5_init_shared_data())
617 		return -rte_errno;
618 	sd = mlx5_shared_data;
619 	MLX5_ASSERT(sd);
620 	rte_spinlock_lock(&sd->lock);
621 	switch (rte_eal_process_type()) {
622 	case RTE_PROC_PRIMARY:
623 		if (sd->init_done)
624 			break;
625 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
626 					   mlx5_mp_os_primary_handle);
627 		if (ret)
628 			goto out;
629 		sd->init_done = true;
630 		break;
631 	case RTE_PROC_SECONDARY:
632 		if (ld->init_done)
633 			break;
634 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
635 					     mlx5_mp_os_secondary_handle);
636 		if (ret)
637 			goto out;
638 		++sd->secondary_cnt;
639 		ld->init_done = true;
640 		break;
641 	default:
642 		break;
643 	}
644 out:
645 	rte_spinlock_unlock(&sd->lock);
646 	return ret;
647 }
648 
649 /**
650  * DV flow counter mode detect and config.
651  *
652  * @param dev
653  *   Pointer to rte_eth_dev structure.
654  *
655  */
656 static void
657 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
658 {
659 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
660 	struct mlx5_priv *priv = dev->data->dev_private;
661 	struct mlx5_dev_ctx_shared *sh = priv->sh;
662 	bool fallback;
663 
664 #ifndef HAVE_IBV_DEVX_ASYNC
665 	fallback = true;
666 #else
667 	fallback = false;
668 	if (!sh->devx || !priv->config.dv_flow_en ||
669 	    !priv->config.hca_attr.flow_counters_dump ||
670 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
671 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
672 		fallback = true;
673 #endif
674 	if (fallback)
675 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
676 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
677 			priv->config.hca_attr.flow_counters_dump,
678 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
679 	/* Initialize fallback mode only on the port initializes sh. */
680 	if (sh->refcnt == 1)
681 		sh->cmng.counter_fallback = fallback;
682 	else if (fallback != sh->cmng.counter_fallback)
683 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
684 			"with others:%d.", PORT_ID(priv), fallback);
685 #endif
686 }
687 
688 /**
689  * DR flow drop action support detect.
690  *
691  * @param dev
692  *   Pointer to rte_eth_dev structure.
693  *
694  */
695 static void
696 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
697 {
698 #ifdef HAVE_MLX5DV_DR
699 	struct mlx5_priv *priv = dev->data->dev_private;
700 
701 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
702 		return;
703 	/**
704 	 * DR supports drop action placeholder when it is supported;
705 	 * otherwise, use the queue drop action.
706 	 */
707 	if (mlx5_flow_discover_dr_action_support(dev))
708 		priv->root_drop_action = priv->drop_queue.hrxq->action;
709 	else
710 		priv->root_drop_action = priv->sh->dr_drop_action;
711 #endif
712 }
713 
714 static void
715 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
716 {
717 	struct mlx5_priv *priv = dev->data->dev_private;
718 	void *ctx = priv->sh->cdev->ctx;
719 
720 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
721 	if (!priv->q_counters) {
722 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
723 		struct ibv_wq *wq;
724 
725 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
726 			"by DevX - fall-back to use the kernel driver global "
727 			"queue counter.", dev->data->port_id);
728 		/* Create WQ by kernel and query its queue counter ID. */
729 		if (cq) {
730 			wq = mlx5_glue->create_wq(ctx,
731 						  &(struct ibv_wq_init_attr){
732 						    .wq_type = IBV_WQT_RQ,
733 						    .max_wr = 1,
734 						    .max_sge = 1,
735 						    .pd = priv->sh->cdev->pd,
736 						    .cq = cq,
737 						});
738 			if (wq) {
739 				/* Counter is assigned only on RDY state. */
740 				int ret = mlx5_glue->modify_wq(wq,
741 						 &(struct ibv_wq_attr){
742 						 .attr_mask = IBV_WQ_ATTR_STATE,
743 						 .wq_state = IBV_WQS_RDY,
744 						});
745 
746 				if (ret == 0)
747 					mlx5_devx_cmd_wq_query(wq,
748 							 &priv->counter_set_id);
749 				claim_zero(mlx5_glue->destroy_wq(wq));
750 			}
751 			claim_zero(mlx5_glue->destroy_cq(cq));
752 		}
753 	} else {
754 		priv->counter_set_id = priv->q_counters->id;
755 	}
756 	if (priv->counter_set_id == 0)
757 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
758 			"available.", dev->data->port_id);
759 }
760 
761 /**
762  * Check if representor spawn info match devargs.
763  *
764  * @param spawn
765  *   Verbs device parameters (name, port, switch_info) to spawn.
766  * @param eth_da
767  *   Device devargs to probe.
768  *
769  * @return
770  *   Match result.
771  */
772 static bool
773 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
774 		       struct rte_eth_devargs *eth_da)
775 {
776 	struct mlx5_switch_info *switch_info = &spawn->info;
777 	unsigned int p, f;
778 	uint16_t id;
779 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
780 						      eth_da->type);
781 
782 	switch (eth_da->type) {
783 	case RTE_ETH_REPRESENTOR_SF:
784 		if (!(spawn->info.port_name == -1 &&
785 		      switch_info->name_type ==
786 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
787 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
788 			rte_errno = EBUSY;
789 			return false;
790 		}
791 		break;
792 	case RTE_ETH_REPRESENTOR_VF:
793 		/* Allows HPF representor index -1 as exception. */
794 		if (!(spawn->info.port_name == -1 &&
795 		      switch_info->name_type ==
796 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
797 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
798 			rte_errno = EBUSY;
799 			return false;
800 		}
801 		break;
802 	case RTE_ETH_REPRESENTOR_NONE:
803 		rte_errno = EBUSY;
804 		return false;
805 	default:
806 		rte_errno = ENOTSUP;
807 		DRV_LOG(ERR, "unsupported representor type");
808 		return false;
809 	}
810 	/* Check representor ID: */
811 	for (p = 0; p < eth_da->nb_ports; ++p) {
812 		if (spawn->pf_bond < 0) {
813 			/* For non-LAG mode, allow and ignore pf. */
814 			switch_info->pf_num = eth_da->ports[p];
815 			repr_id = mlx5_representor_id_encode(switch_info,
816 							     eth_da->type);
817 		}
818 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
819 			id = MLX5_REPRESENTOR_ID
820 				(eth_da->ports[p], eth_da->type,
821 				 eth_da->representor_ports[f]);
822 			if (repr_id == id)
823 				return true;
824 		}
825 	}
826 	rte_errno = EBUSY;
827 	return false;
828 }
829 
830 /**
831  * Spawn an Ethernet device from Verbs information.
832  *
833  * @param dpdk_dev
834  *   Backing DPDK device.
835  * @param spawn
836  *   Verbs device parameters (name, port, switch_info) to spawn.
837  * @param config
838  *   Device configuration parameters.
839  * @param eth_da
840  *   Device arguments.
841  *
842  * @return
843  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
844  *   is set. The following errors are defined:
845  *
846  *   EBUSY: device is not supposed to be spawned.
847  *   EEXIST: device is already spawned
848  */
849 static struct rte_eth_dev *
850 mlx5_dev_spawn(struct rte_device *dpdk_dev,
851 	       struct mlx5_dev_spawn_data *spawn,
852 	       struct mlx5_dev_config *config,
853 	       struct rte_eth_devargs *eth_da)
854 {
855 	const struct mlx5_switch_info *switch_info = &spawn->info;
856 	struct mlx5_dev_ctx_shared *sh = NULL;
857 	struct ibv_port_attr port_attr = { .state = IBV_PORT_NOP };
858 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
859 	struct rte_eth_dev *eth_dev = NULL;
860 	struct mlx5_priv *priv = NULL;
861 	int err = 0;
862 	unsigned int hw_padding = 0;
863 	unsigned int mps;
864 	unsigned int mpls_en = 0;
865 	unsigned int swp = 0;
866 	unsigned int mprq = 0;
867 	unsigned int mprq_min_stride_size_n = 0;
868 	unsigned int mprq_max_stride_size_n = 0;
869 	unsigned int mprq_min_stride_num_n = 0;
870 	unsigned int mprq_max_stride_num_n = 0;
871 	struct rte_ether_addr mac;
872 	char name[RTE_ETH_NAME_MAX_LEN];
873 	int own_domain_id = 0;
874 	uint16_t port_id;
875 	struct mlx5_port_info vport_info = { .query_flags = 0 };
876 	int nl_rdma = -1;
877 	int i;
878 
879 	/* Determine if this port representor is supposed to be spawned. */
880 	if (switch_info->representor && dpdk_dev->devargs &&
881 	    !mlx5_representor_match(spawn, eth_da))
882 		return NULL;
883 	/* Build device name. */
884 	if (spawn->pf_bond < 0) {
885 		/* Single device. */
886 		if (!switch_info->representor)
887 			strlcpy(name, dpdk_dev->name, sizeof(name));
888 		else
889 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
890 				 dpdk_dev->name,
891 				 switch_info->name_type ==
892 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
893 				 switch_info->port_name);
894 	} else {
895 		/* Bonding device. */
896 		if (!switch_info->representor) {
897 			err = snprintf(name, sizeof(name), "%s_%s",
898 				       dpdk_dev->name, spawn->phys_dev_name);
899 		} else {
900 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
901 				dpdk_dev->name, spawn->phys_dev_name,
902 				switch_info->ctrl_num,
903 				switch_info->pf_num,
904 				switch_info->name_type ==
905 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
906 				switch_info->port_name);
907 		}
908 	}
909 	if (err >= (int)sizeof(name))
910 		DRV_LOG(WARNING, "device name overflow %s", name);
911 	/* check if the device is already spawned */
912 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
913 		rte_errno = EEXIST;
914 		return NULL;
915 	}
916 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
917 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
918 		struct mlx5_mp_id mp_id;
919 
920 		eth_dev = rte_eth_dev_attach_secondary(name);
921 		if (eth_dev == NULL) {
922 			DRV_LOG(ERR, "can not attach rte ethdev");
923 			rte_errno = ENOMEM;
924 			return NULL;
925 		}
926 		eth_dev->device = dpdk_dev;
927 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
928 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
929 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
930 		err = mlx5_proc_priv_init(eth_dev);
931 		if (err)
932 			return NULL;
933 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
934 		/* Receive command fd from primary process */
935 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
936 		if (err < 0)
937 			goto err_secondary;
938 		/* Remap UAR for Tx queues. */
939 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
940 		if (err)
941 			goto err_secondary;
942 		/*
943 		 * Ethdev pointer is still required as input since
944 		 * the primary device is not accessible from the
945 		 * secondary process.
946 		 */
947 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
948 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
949 		return eth_dev;
950 err_secondary:
951 		mlx5_dev_close(eth_dev);
952 		return NULL;
953 	}
954 	/*
955 	 * Some parameters ("tx_db_nc" in particularly) are needed in
956 	 * advance to create dv/verbs device context. We proceed the
957 	 * devargs here to get ones, and later proceed devargs again
958 	 * to override some hardware settings.
959 	 */
960 	err = mlx5_args(config, dpdk_dev->devargs);
961 	if (err) {
962 		err = rte_errno;
963 		DRV_LOG(ERR, "failed to process device arguments: %s",
964 			strerror(rte_errno));
965 		goto error;
966 	}
967 	if (config->dv_miss_info) {
968 		if (switch_info->master || switch_info->representor)
969 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
970 	}
971 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
972 	if (!sh)
973 		return NULL;
974 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
975 	config->dest_tir = 1;
976 #endif
977 #ifdef HAVE_IBV_MLX5_MOD_SWP
978 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
979 #endif
980 	/*
981 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
982 	 * as all ConnectX-5 devices.
983 	 */
984 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
985 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
986 #endif
987 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
988 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
989 #endif
990 	mlx5_glue->dv_query_device(sh->cdev->ctx, &dv_attr);
991 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
992 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
993 			DRV_LOG(DEBUG, "enhanced MPW is supported");
994 			mps = MLX5_MPW_ENHANCED;
995 		} else {
996 			DRV_LOG(DEBUG, "MPW is supported");
997 			mps = MLX5_MPW;
998 		}
999 	} else {
1000 		DRV_LOG(DEBUG, "MPW isn't supported");
1001 		mps = MLX5_MPW_DISABLED;
1002 	}
1003 #ifdef HAVE_IBV_MLX5_MOD_SWP
1004 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1005 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1006 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1007 #endif
1008 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
1009 		MLX5_SW_PARSING_TSO_CAP);
1010 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1011 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1012 		struct mlx5dv_striding_rq_caps mprq_caps =
1013 			dv_attr.striding_rq_caps;
1014 
1015 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1016 			mprq_caps.min_single_stride_log_num_of_bytes);
1017 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1018 			mprq_caps.max_single_stride_log_num_of_bytes);
1019 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1020 			mprq_caps.min_single_wqe_log_num_of_strides);
1021 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1022 			mprq_caps.max_single_wqe_log_num_of_strides);
1023 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1024 			mprq_caps.supported_qpts);
1025 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1026 		mprq = 1;
1027 		mprq_min_stride_size_n =
1028 			mprq_caps.min_single_stride_log_num_of_bytes;
1029 		mprq_max_stride_size_n =
1030 			mprq_caps.max_single_stride_log_num_of_bytes;
1031 		mprq_min_stride_num_n =
1032 			mprq_caps.min_single_wqe_log_num_of_strides;
1033 		mprq_max_stride_num_n =
1034 			mprq_caps.max_single_wqe_log_num_of_strides;
1035 	}
1036 #endif
1037 	/* Rx CQE compression is enabled by default. */
1038 	config->cqe_comp = 1;
1039 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1040 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1041 		config->tunnel_en = dv_attr.tunnel_offloads_caps &
1042 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1043 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1044 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1045 	}
1046 	if (config->tunnel_en) {
1047 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1048 		config->tunnel_en &
1049 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1050 		config->tunnel_en &
1051 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1052 		config->tunnel_en &
1053 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1054 		);
1055 	} else {
1056 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1057 	}
1058 #else
1059 	DRV_LOG(WARNING,
1060 		"tunnel offloading disabled due to old OFED/rdma-core version");
1061 #endif
1062 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1063 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1064 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1065 		   (dv_attr.tunnel_offloads_caps &
1066 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1067 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1068 		mpls_en ? "" : "not ");
1069 #else
1070 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1071 		" old OFED/rdma-core version or firmware configuration");
1072 #endif
1073 	config->mpls_en = mpls_en;
1074 	nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1075 	/* Check port status. */
1076 	if (spawn->phys_port <= UINT8_MAX) {
1077 		/* Legacy Verbs api only support u8 port number. */
1078 		err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1079 					    &port_attr);
1080 		if (err) {
1081 			DRV_LOG(ERR, "port query failed: %s", strerror(err));
1082 			goto error;
1083 		}
1084 		if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1085 			DRV_LOG(ERR, "port is not configured in Ethernet mode");
1086 			err = EINVAL;
1087 			goto error;
1088 		}
1089 	} else if (nl_rdma >= 0) {
1090 		/* IB doesn't allow more than 255 ports, must be Ethernet. */
1091 		err = mlx5_nl_port_state(nl_rdma,
1092 			spawn->phys_dev_name,
1093 			spawn->phys_port);
1094 		if (err < 0) {
1095 			DRV_LOG(INFO, "Failed to get netlink port state: %s",
1096 				strerror(rte_errno));
1097 			err = -rte_errno;
1098 			goto error;
1099 		}
1100 		port_attr.state = (enum ibv_port_state)err;
1101 	}
1102 	if (port_attr.state != IBV_PORT_ACTIVE)
1103 		DRV_LOG(INFO, "port is not active: \"%s\" (%d)",
1104 			mlx5_glue->port_state_str(port_attr.state),
1105 			port_attr.state);
1106 	/* Allocate private eth device data. */
1107 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1108 			   sizeof(*priv),
1109 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1110 	if (priv == NULL) {
1111 		DRV_LOG(ERR, "priv allocation failure");
1112 		err = ENOMEM;
1113 		goto error;
1114 	}
1115 	priv->sh = sh;
1116 	priv->dev_port = spawn->phys_port;
1117 	priv->pci_dev = spawn->pci_dev;
1118 	priv->mtu = RTE_ETHER_MTU;
1119 	/* Some internal functions rely on Netlink sockets, open them now. */
1120 	priv->nl_socket_rdma = nl_rdma;
1121 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1122 	priv->representor = !!switch_info->representor;
1123 	priv->master = !!switch_info->master;
1124 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1125 	priv->vport_meta_tag = 0;
1126 	priv->vport_meta_mask = 0;
1127 	priv->pf_bond = spawn->pf_bond;
1128 
1129 	DRV_LOG(DEBUG,
1130 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1131 		priv->dev_port, dpdk_dev->bus->name,
1132 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1133 		priv->master, priv->representor, priv->pf_bond);
1134 
1135 	/*
1136 	 * If we have E-Switch we should determine the vport attributes.
1137 	 * E-Switch may use either source vport field or reg_c[0] metadata
1138 	 * register to match on vport index. The engaged part of metadata
1139 	 * register is defined by mask.
1140 	 */
1141 	if (switch_info->representor || switch_info->master) {
1142 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1143 						 spawn->phys_port,
1144 						 &vport_info);
1145 		if (err) {
1146 			DRV_LOG(WARNING,
1147 				"Cannot query devx port %d on device %s",
1148 				spawn->phys_port, spawn->phys_dev_name);
1149 			vport_info.query_flags = 0;
1150 		}
1151 	}
1152 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1153 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1154 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1155 		if (!priv->vport_meta_mask) {
1156 			DRV_LOG(ERR,
1157 				"vport zero mask for port %d on bonding device %s",
1158 				spawn->phys_port, spawn->phys_dev_name);
1159 			err = ENOTSUP;
1160 			goto error;
1161 		}
1162 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1163 			DRV_LOG(ERR,
1164 				"Invalid vport tag for port %d on bonding device %s",
1165 				spawn->phys_port, spawn->phys_dev_name);
1166 			err = ENOTSUP;
1167 			goto error;
1168 		}
1169 	}
1170 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1171 		priv->vport_id = vport_info.vport_id;
1172 	} else if (spawn->pf_bond >= 0 &&
1173 		   (switch_info->representor || switch_info->master)) {
1174 		DRV_LOG(ERR,
1175 			"Cannot deduce vport index for port %d on bonding device %s",
1176 			spawn->phys_port, spawn->phys_dev_name);
1177 		err = ENOTSUP;
1178 		goto error;
1179 	} else {
1180 		/*
1181 		 * Suppose vport index in compatible way. Kernel/rdma_core
1182 		 * support single E-Switch per PF configurations only and
1183 		 * vport_id field contains the vport index for associated VF,
1184 		 * which is deduced from representor port name.
1185 		 * For example, let's have the IB device port 10, it has
1186 		 * attached network device eth0, which has port name attribute
1187 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1188 		 * as 3 (2+1). This assigning schema should be changed if the
1189 		 * multiple E-Switch instances per PF configurations or/and PCI
1190 		 * subfunctions are added.
1191 		 */
1192 		priv->vport_id = switch_info->representor ?
1193 				 switch_info->port_name + 1 : -1;
1194 	}
1195 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1196 							  eth_da->type);
1197 	/*
1198 	 * Look for sibling devices in order to reuse their switch domain
1199 	 * if any, otherwise allocate one.
1200 	 */
1201 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1202 		const struct mlx5_priv *opriv =
1203 			rte_eth_devices[port_id].data->dev_private;
1204 
1205 		if (!opriv ||
1206 		    opriv->sh != priv->sh ||
1207 			opriv->domain_id ==
1208 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1209 			continue;
1210 		priv->domain_id = opriv->domain_id;
1211 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1212 			priv->dev_port, priv->domain_id);
1213 		break;
1214 	}
1215 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1216 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1217 		if (err) {
1218 			err = rte_errno;
1219 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1220 				strerror(rte_errno));
1221 			goto error;
1222 		}
1223 		own_domain_id = 1;
1224 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1225 			priv->dev_port, priv->domain_id);
1226 	}
1227 	/* Override some values set by hardware configuration. */
1228 	mlx5_args(config, dpdk_dev->devargs);
1229 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
1230 	if (err)
1231 		goto error;
1232 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1233 			    IBV_DEVICE_RAW_IP_CSUM);
1234 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1235 		(config->hw_csum ? "" : "not "));
1236 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1237 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1238 	DRV_LOG(DEBUG, "counters are not supported");
1239 #endif
1240 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1241 	if (config->dv_flow_en) {
1242 		DRV_LOG(WARNING, "DV flow is not supported");
1243 		config->dv_flow_en = 0;
1244 	}
1245 #endif
1246 	config->ind_table_max_size =
1247 		sh->device_attr.max_rwq_indirection_table_size;
1248 	/*
1249 	 * Remove this check once DPDK supports larger/variable
1250 	 * indirection tables.
1251 	 */
1252 	if (config->ind_table_max_size > (unsigned int)RTE_ETH_RSS_RETA_SIZE_512)
1253 		config->ind_table_max_size = RTE_ETH_RSS_RETA_SIZE_512;
1254 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1255 		config->ind_table_max_size);
1256 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1257 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1258 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1259 		(config->hw_vlan_strip ? "" : "not "));
1260 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1261 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1262 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1263 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1264 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1265 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1266 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1267 #endif
1268 	if (config->hw_padding && !hw_padding) {
1269 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1270 		config->hw_padding = 0;
1271 	} else if (config->hw_padding) {
1272 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1273 	}
1274 	config->tso = (sh->device_attr.max_tso > 0 &&
1275 		      (sh->device_attr.tso_supported_qpts &
1276 		       (1 << IBV_QPT_RAW_PACKET)));
1277 	if (config->tso)
1278 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1279 	/*
1280 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1281 	 * by default.
1282 	 */
1283 	if (config->mps == MLX5_ARG_UNSET)
1284 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1285 							  MLX5_MPW_DISABLED;
1286 	else
1287 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1288 	DRV_LOG(INFO, "%sMPS is %s",
1289 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1290 		config->mps == MLX5_MPW ? "legacy " : "",
1291 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1292 	if (sh->devx) {
1293 		config->hca_attr = sh->cdev->config.hca_attr;
1294 		sh->steering_format_version =
1295 			config->hca_attr.steering_format_version;
1296 		/* Check for LRO support. */
1297 		if (config->dest_tir && config->hca_attr.lro_cap &&
1298 		    config->dv_flow_en) {
1299 			/* TBD check tunnel lro caps. */
1300 			config->lro.supported = config->hca_attr.lro_cap;
1301 			DRV_LOG(DEBUG, "Device supports LRO");
1302 			/*
1303 			 * If LRO timeout is not configured by application,
1304 			 * use the minimal supported value.
1305 			 */
1306 			if (!config->lro.timeout)
1307 				config->lro.timeout =
1308 				config->hca_attr.lro_timer_supported_periods[0];
1309 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1310 				config->lro.timeout);
1311 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1312 				"required for coalescing is %d bytes",
1313 				config->hca_attr.lro_min_mss_size);
1314 		}
1315 #if defined(HAVE_MLX5DV_DR) && \
1316 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1317 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1318 		if (config->hca_attr.qos.sup &&
1319 		    config->hca_attr.qos.flow_meter_old &&
1320 		    config->dv_flow_en) {
1321 			uint8_t reg_c_mask =
1322 				config->hca_attr.qos.flow_meter_reg_c_ids;
1323 			/*
1324 			 * Meter needs two REG_C's for color match and pre-sfx
1325 			 * flow match. Here get the REG_C for color match.
1326 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1327 			 */
1328 			reg_c_mask &= 0xfc;
1329 			if (__builtin_popcount(reg_c_mask) < 1) {
1330 				priv->mtr_en = 0;
1331 				DRV_LOG(WARNING, "No available register for"
1332 					" meter.");
1333 			} else {
1334 				/*
1335 				 * The meter color register is used by the
1336 				 * flow-hit feature as well.
1337 				 * The flow-hit feature must use REG_C_3
1338 				 * Prefer REG_C_3 if it is available.
1339 				 */
1340 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1341 					priv->mtr_color_reg = REG_C_3;
1342 				else
1343 					priv->mtr_color_reg = ffs(reg_c_mask)
1344 							      - 1 + REG_C_0;
1345 				priv->mtr_en = 1;
1346 				priv->mtr_reg_share =
1347 				      config->hca_attr.qos.flow_meter;
1348 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1349 					priv->mtr_color_reg);
1350 			}
1351 		}
1352 		if (config->hca_attr.qos.sup &&
1353 			config->hca_attr.qos.flow_meter_aso_sup) {
1354 			uint32_t log_obj_size =
1355 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1356 			if (log_obj_size >=
1357 			config->hca_attr.qos.log_meter_aso_granularity &&
1358 			log_obj_size <=
1359 			config->hca_attr.qos.log_meter_aso_max_alloc)
1360 				sh->meter_aso_en = 1;
1361 		}
1362 		if (priv->mtr_en) {
1363 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1364 			if (err) {
1365 				err = -err;
1366 				goto error;
1367 			}
1368 		}
1369 		if (config->hca_attr.flow.tunnel_header_0_1)
1370 			sh->tunnel_header_0_1 = 1;
1371 #endif
1372 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1373 		if (config->hca_attr.flow_hit_aso &&
1374 		    priv->mtr_color_reg == REG_C_3) {
1375 			sh->flow_hit_aso_en = 1;
1376 			err = mlx5_flow_aso_age_mng_init(sh);
1377 			if (err) {
1378 				err = -err;
1379 				goto error;
1380 			}
1381 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1382 		}
1383 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1384 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1385 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1386 		if (config->hca_attr.ct_offload &&
1387 		    priv->mtr_color_reg == REG_C_3) {
1388 			err = mlx5_flow_aso_ct_mng_init(sh);
1389 			if (err) {
1390 				err = -err;
1391 				goto error;
1392 			}
1393 			DRV_LOG(DEBUG, "CT ASO is supported.");
1394 			sh->ct_aso_en = 1;
1395 		}
1396 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1397 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1398 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1399 		    config->dv_flow_en) {
1400 			priv->sampler_en = 1;
1401 			DRV_LOG(DEBUG, "Sampler enabled!");
1402 		} else {
1403 			priv->sampler_en = 0;
1404 			if (!config->hca_attr.log_max_ft_sampler_num)
1405 				DRV_LOG(WARNING,
1406 					"No available register for sampler.");
1407 			else
1408 				DRV_LOG(DEBUG, "DV flow is not supported!");
1409 		}
1410 #endif
1411 	}
1412 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1413 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1414 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1415 		config->cqe_comp = 0;
1416 	}
1417 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1418 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1419 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1420 				 " format isn't supported.");
1421 		config->cqe_comp = 0;
1422 	}
1423 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1424 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1425 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1426 				 " format isn't supported.");
1427 		config->cqe_comp = 0;
1428 	}
1429 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1430 			config->cqe_comp ? "" : "not ");
1431 	if (config->tx_pp) {
1432 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1433 			config->hca_attr.dev_freq_khz);
1434 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1435 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1436 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1437 			config->hca_attr.cross_channel ? "" : "not ");
1438 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1439 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1440 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1441 			config->hca_attr.non_wire_sq ? "" : "not ");
1442 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1443 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1444 			config->hca_attr.log_max_static_sq_wq);
1445 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1446 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1447 		if (!sh->devx) {
1448 			DRV_LOG(ERR, "DevX is required for packet pacing");
1449 			err = ENODEV;
1450 			goto error;
1451 		}
1452 		if (!config->hca_attr.qos.packet_pacing) {
1453 			DRV_LOG(ERR, "Packet pacing is not supported");
1454 			err = ENODEV;
1455 			goto error;
1456 		}
1457 		if (!config->hca_attr.cross_channel) {
1458 			DRV_LOG(ERR, "Cross channel operations are"
1459 				     " required for packet pacing");
1460 			err = ENODEV;
1461 			goto error;
1462 		}
1463 		if (!config->hca_attr.wqe_index_ignore) {
1464 			DRV_LOG(ERR, "WQE index ignore feature is"
1465 				     " required for packet pacing");
1466 			err = ENODEV;
1467 			goto error;
1468 		}
1469 		if (!config->hca_attr.non_wire_sq) {
1470 			DRV_LOG(ERR, "Non-wire SQ feature is"
1471 				     " required for packet pacing");
1472 			err = ENODEV;
1473 			goto error;
1474 		}
1475 		if (!config->hca_attr.log_max_static_sq_wq) {
1476 			DRV_LOG(ERR, "Static WQE SQ feature is"
1477 				     " required for packet pacing");
1478 			err = ENODEV;
1479 			goto error;
1480 		}
1481 		if (!config->hca_attr.qos.wqe_rate_pp) {
1482 			DRV_LOG(ERR, "WQE rate mode is required"
1483 				     " for packet pacing");
1484 			err = ENODEV;
1485 			goto error;
1486 		}
1487 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1488 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1489 			     " can't create queues for packet pacing");
1490 		err = ENODEV;
1491 		goto error;
1492 #endif
1493 	}
1494 	if (sh->devx) {
1495 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1496 
1497 		err = config->hca_attr.access_register_user ?
1498 			mlx5_devx_cmd_register_read
1499 				(sh->cdev->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1500 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1501 		if (!err) {
1502 			uint32_t ts_mode;
1503 
1504 			/* MTUTC register is read successfully. */
1505 			ts_mode = MLX5_GET(register_mtutc, reg,
1506 					   time_stamp_mode);
1507 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1508 				config->rt_timestamp = 1;
1509 		} else {
1510 			/* Kernel does not support register reading. */
1511 			if (config->hca_attr.dev_freq_khz ==
1512 						 (NS_PER_S / MS_PER_S))
1513 				config->rt_timestamp = 1;
1514 		}
1515 	}
1516 	/*
1517 	 * If HW has bug working with tunnel packet decapsulation and
1518 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1519 	 * bit. Then RTE_ETH_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1520 	 */
1521 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1522 		config->hw_fcs_strip = 0;
1523 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1524 		(config->hw_fcs_strip ? "" : "not "));
1525 	if (config->mprq.enabled && mprq) {
1526 		if (config->mprq.stride_num_n &&
1527 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1528 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1529 			config->mprq.stride_num_n =
1530 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1531 						mprq_min_stride_num_n),
1532 					mprq_max_stride_num_n);
1533 			DRV_LOG(WARNING,
1534 				"the number of strides"
1535 				" for Multi-Packet RQ is out of range,"
1536 				" setting default value (%u)",
1537 				1 << config->mprq.stride_num_n);
1538 		}
1539 		if (config->mprq.stride_size_n &&
1540 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1541 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1542 			config->mprq.stride_size_n =
1543 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1544 						mprq_min_stride_size_n),
1545 					mprq_max_stride_size_n);
1546 			DRV_LOG(WARNING,
1547 				"the size of a stride"
1548 				" for Multi-Packet RQ is out of range,"
1549 				" setting default value (%u)",
1550 				1 << config->mprq.stride_size_n);
1551 		}
1552 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1553 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1554 	} else if (config->mprq.enabled && !mprq) {
1555 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1556 		config->mprq.enabled = 0;
1557 	}
1558 	if (config->max_dump_files_num == 0)
1559 		config->max_dump_files_num = 128;
1560 	eth_dev = rte_eth_dev_allocate(name);
1561 	if (eth_dev == NULL) {
1562 		DRV_LOG(ERR, "can not allocate rte ethdev");
1563 		err = ENOMEM;
1564 		goto error;
1565 	}
1566 	if (priv->representor) {
1567 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1568 		eth_dev->data->representor_id = priv->representor_id;
1569 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1570 			struct mlx5_priv *opriv =
1571 				rte_eth_devices[port_id].data->dev_private;
1572 			if (opriv &&
1573 			    opriv->master &&
1574 			    opriv->domain_id == priv->domain_id &&
1575 			    opriv->sh == priv->sh) {
1576 				eth_dev->data->backer_port_id = port_id;
1577 				break;
1578 			}
1579 		}
1580 		if (port_id >= RTE_MAX_ETHPORTS)
1581 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1582 	}
1583 	priv->mp_id.port_id = eth_dev->data->port_id;
1584 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1585 	/*
1586 	 * Store associated network device interface index. This index
1587 	 * is permanent throughout the lifetime of device. So, we may store
1588 	 * the ifindex here and use the cached value further.
1589 	 */
1590 	MLX5_ASSERT(spawn->ifindex);
1591 	priv->if_index = spawn->ifindex;
1592 	priv->lag_affinity_idx = sh->refcnt - 1;
1593 	eth_dev->data->dev_private = priv;
1594 	priv->dev_data = eth_dev->data;
1595 	eth_dev->data->mac_addrs = priv->mac;
1596 	eth_dev->device = dpdk_dev;
1597 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1598 	/* Configure the first MAC address by default. */
1599 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1600 		DRV_LOG(ERR,
1601 			"port %u cannot get MAC address, is mlx5_en"
1602 			" loaded? (errno: %s)",
1603 			eth_dev->data->port_id, strerror(rte_errno));
1604 		err = ENODEV;
1605 		goto error;
1606 	}
1607 	DRV_LOG(INFO,
1608 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1609 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1610 #ifdef RTE_LIBRTE_MLX5_DEBUG
1611 	{
1612 		char ifname[MLX5_NAMESIZE];
1613 
1614 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1615 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1616 				eth_dev->data->port_id, ifname);
1617 		else
1618 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1619 				eth_dev->data->port_id);
1620 	}
1621 #endif
1622 	/* Get actual MTU if possible. */
1623 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1624 	if (err) {
1625 		err = rte_errno;
1626 		goto error;
1627 	}
1628 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1629 		priv->mtu);
1630 	/* Initialize burst functions to prevent crashes before link-up. */
1631 	eth_dev->rx_pkt_burst = removed_rx_burst;
1632 	eth_dev->tx_pkt_burst = removed_tx_burst;
1633 	eth_dev->dev_ops = &mlx5_dev_ops;
1634 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1635 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1636 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1637 	/* Register MAC address. */
1638 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1639 	if (config->vf && config->vf_nl_en)
1640 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1641 				      mlx5_ifindex(eth_dev),
1642 				      eth_dev->data->mac_addrs,
1643 				      MLX5_MAX_MAC_ADDRESSES);
1644 	priv->ctrl_flows = 0;
1645 	rte_spinlock_init(&priv->flow_list_lock);
1646 	TAILQ_INIT(&priv->flow_meters);
1647 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1648 	if (!priv->mtr_profile_tbl)
1649 		goto error;
1650 	/* Bring Ethernet device up. */
1651 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1652 		eth_dev->data->port_id);
1653 	mlx5_set_link_up(eth_dev);
1654 	/*
1655 	 * Even though the interrupt handler is not installed yet,
1656 	 * interrupts will still trigger on the async_fd from
1657 	 * Verbs context returned by ibv_open_device().
1658 	 */
1659 	mlx5_link_update(eth_dev, 0);
1660 #ifdef HAVE_MLX5DV_DR_ESWITCH
1661 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1662 	      (switch_info->representor || switch_info->master)))
1663 		config->dv_esw_en = 0;
1664 #else
1665 	config->dv_esw_en = 0;
1666 #endif
1667 	/* Detect minimal data bytes to inline. */
1668 	mlx5_set_min_inline(spawn, config);
1669 	/* Store device configuration on private structure. */
1670 	priv->config = *config;
1671 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1672 		icfg[i].release_mem_en = !!config->reclaim_mode;
1673 		if (config->reclaim_mode)
1674 			icfg[i].per_core_cache = 0;
1675 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1676 		if (!priv->flows[i])
1677 			goto error;
1678 	}
1679 	/* Create context for virtual machine VLAN workaround. */
1680 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1681 	if (config->dv_flow_en) {
1682 		err = mlx5_alloc_shared_dr(priv);
1683 		if (err)
1684 			goto error;
1685 	}
1686 	if (sh->devx && config->dv_flow_en && config->dest_tir) {
1687 		priv->obj_ops = devx_obj_ops;
1688 		priv->obj_ops.drop_action_create =
1689 						ibv_obj_ops.drop_action_create;
1690 		priv->obj_ops.drop_action_destroy =
1691 						ibv_obj_ops.drop_action_destroy;
1692 		mlx5_queue_counter_id_prepare(eth_dev);
1693 		priv->obj_ops.lb_dummy_queue_create =
1694 					mlx5_rxq_ibv_obj_dummy_lb_create;
1695 		priv->obj_ops.lb_dummy_queue_release =
1696 					mlx5_rxq_ibv_obj_dummy_lb_release;
1697 	} else if (spawn->max_port > UINT8_MAX) {
1698 		/* Verbs can't support ports larger than 255 by design. */
1699 		DRV_LOG(ERR, "must enable DV and ESW when RDMA link ports > 255");
1700 		err = ENOTSUP;
1701 		goto error;
1702 	} else {
1703 		priv->obj_ops = ibv_obj_ops;
1704 	}
1705 	if (config->tx_pp &&
1706 	    (priv->config.dv_esw_en ||
1707 	     priv->obj_ops.txq_obj_new != mlx5_txq_devx_obj_new)) {
1708 		/*
1709 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1710 		 * packet pacing and already checked above.
1711 		 * Hence, we should only make sure the SQs will be created
1712 		 * with DevX, not with Verbs.
1713 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1714 		 * with Clock Queue UAR as required for Tx scheduling.
1715 		 */
1716 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1717 		err = ENODEV;
1718 		goto error;
1719 	}
1720 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1721 	if (!priv->drop_queue.hrxq)
1722 		goto error;
1723 	/* Supported Verbs flow priority number detection. */
1724 	err = mlx5_flow_discover_priorities(eth_dev);
1725 	if (err < 0) {
1726 		err = -err;
1727 		goto error;
1728 	}
1729 	priv->config.flow_prio = err;
1730 	if (!priv->config.dv_esw_en &&
1731 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1732 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1733 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1734 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1735 	}
1736 	mlx5_set_metadata_mask(eth_dev);
1737 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1738 	    !priv->sh->dv_regc0_mask) {
1739 		DRV_LOG(ERR, "metadata mode %u is not supported "
1740 			     "(no metadata reg_c[0] is available)",
1741 			     priv->config.dv_xmeta_en);
1742 			err = ENOTSUP;
1743 			goto error;
1744 	}
1745 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1746 				       mlx5_hrxq_create_cb,
1747 				       mlx5_hrxq_match_cb,
1748 				       mlx5_hrxq_remove_cb,
1749 				       mlx5_hrxq_clone_cb,
1750 				       mlx5_hrxq_clone_free_cb);
1751 	if (!priv->hrxqs)
1752 		goto error;
1753 	rte_rwlock_init(&priv->ind_tbls_lock);
1754 	/* Query availability of metadata reg_c's. */
1755 	err = mlx5_flow_discover_mreg_c(eth_dev);
1756 	if (err < 0) {
1757 		err = -err;
1758 		goto error;
1759 	}
1760 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1761 		DRV_LOG(DEBUG,
1762 			"port %u extensive metadata register is not supported",
1763 			eth_dev->data->port_id);
1764 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1765 			DRV_LOG(ERR, "metadata mode %u is not supported "
1766 				     "(no metadata registers available)",
1767 				     priv->config.dv_xmeta_en);
1768 			err = ENOTSUP;
1769 			goto error;
1770 		}
1771 	}
1772 	if (priv->config.dv_flow_en &&
1773 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1774 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1775 	    priv->sh->dv_regc0_mask) {
1776 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1777 						      MLX5_FLOW_MREG_HTABLE_SZ,
1778 						      false, true, eth_dev,
1779 						      flow_dv_mreg_create_cb,
1780 						      flow_dv_mreg_match_cb,
1781 						      flow_dv_mreg_remove_cb,
1782 						      flow_dv_mreg_clone_cb,
1783 						    flow_dv_mreg_clone_free_cb);
1784 		if (!priv->mreg_cp_tbl) {
1785 			err = ENOMEM;
1786 			goto error;
1787 		}
1788 	}
1789 	rte_spinlock_init(&priv->shared_act_sl);
1790 	mlx5_flow_counter_mode_config(eth_dev);
1791 	mlx5_flow_drop_action_config(eth_dev);
1792 	if (priv->config.dv_flow_en)
1793 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1794 	return eth_dev;
1795 error:
1796 	if (priv) {
1797 		if (priv->mreg_cp_tbl)
1798 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1799 		if (priv->sh)
1800 			mlx5_os_free_shared_dr(priv);
1801 		if (priv->nl_socket_route >= 0)
1802 			close(priv->nl_socket_route);
1803 		if (priv->vmwa_context)
1804 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1805 		if (eth_dev && priv->drop_queue.hrxq)
1806 			mlx5_drop_action_destroy(eth_dev);
1807 		if (priv->mtr_profile_tbl)
1808 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1809 		if (own_domain_id)
1810 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1811 		if (priv->hrxqs)
1812 			mlx5_list_destroy(priv->hrxqs);
1813 		mlx5_free(priv);
1814 		if (eth_dev != NULL)
1815 			eth_dev->data->dev_private = NULL;
1816 	}
1817 	if (eth_dev != NULL) {
1818 		/* mac_addrs must not be freed alone because part of
1819 		 * dev_private
1820 		 **/
1821 		eth_dev->data->mac_addrs = NULL;
1822 		rte_eth_dev_release_port(eth_dev);
1823 	}
1824 	if (sh)
1825 		mlx5_free_shared_dev_ctx(sh);
1826 	if (nl_rdma >= 0)
1827 		close(nl_rdma);
1828 	MLX5_ASSERT(err > 0);
1829 	rte_errno = err;
1830 	return NULL;
1831 }
1832 
1833 /**
1834  * Comparison callback to sort device data.
1835  *
1836  * This is meant to be used with qsort().
1837  *
1838  * @param a[in]
1839  *   Pointer to pointer to first data object.
1840  * @param b[in]
1841  *   Pointer to pointer to second data object.
1842  *
1843  * @return
1844  *   0 if both objects are equal, less than 0 if the first argument is less
1845  *   than the second, greater than 0 otherwise.
1846  */
1847 static int
1848 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1849 {
1850 	const struct mlx5_switch_info *si_a =
1851 		&((const struct mlx5_dev_spawn_data *)a)->info;
1852 	const struct mlx5_switch_info *si_b =
1853 		&((const struct mlx5_dev_spawn_data *)b)->info;
1854 	int ret;
1855 
1856 	/* Master device first. */
1857 	ret = si_b->master - si_a->master;
1858 	if (ret)
1859 		return ret;
1860 	/* Then representor devices. */
1861 	ret = si_b->representor - si_a->representor;
1862 	if (ret)
1863 		return ret;
1864 	/* Unidentified devices come last in no specific order. */
1865 	if (!si_a->representor)
1866 		return 0;
1867 	/* Order representors by name. */
1868 	return si_a->port_name - si_b->port_name;
1869 }
1870 
1871 /**
1872  * Match PCI information for possible slaves of bonding device.
1873  *
1874  * @param[in] ibdev_name
1875  *   Name of Infiniband device.
1876  * @param[in] pci_dev
1877  *   Pointer to primary PCI address structure to match.
1878  * @param[in] nl_rdma
1879  *   Netlink RDMA group socket handle.
1880  * @param[in] owner
1881  *   Representor owner PF index.
1882  * @param[out] bond_info
1883  *   Pointer to bonding information.
1884  *
1885  * @return
1886  *   negative value if no bonding device found, otherwise
1887  *   positive index of slave PF in bonding.
1888  */
1889 static int
1890 mlx5_device_bond_pci_match(const char *ibdev_name,
1891 			   const struct rte_pci_addr *pci_dev,
1892 			   int nl_rdma, uint16_t owner,
1893 			   struct mlx5_bond_info *bond_info)
1894 {
1895 	char ifname[IF_NAMESIZE + 1];
1896 	unsigned int ifindex;
1897 	unsigned int np, i;
1898 	FILE *bond_file = NULL, *file;
1899 	int pf = -1;
1900 	int ret;
1901 
1902 	/*
1903 	 * Try to get master device name. If something goes wrong suppose
1904 	 * the lack of kernel support and no bonding devices.
1905 	 */
1906 	memset(bond_info, 0, sizeof(*bond_info));
1907 	if (nl_rdma < 0)
1908 		return -1;
1909 	if (!strstr(ibdev_name, "bond"))
1910 		return -1;
1911 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1912 	if (!np)
1913 		return -1;
1914 	/*
1915 	 * The master device might not be on the predefined port(not on port
1916 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1917 	 * device ports and find master.
1918 	 */
1919 	for (i = 1; i <= np; ++i) {
1920 		/* Check whether Infiniband port is populated. */
1921 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1922 		if (!ifindex)
1923 			continue;
1924 		if (!if_indextoname(ifindex, ifname))
1925 			continue;
1926 		/* Try to read bonding slave names from sysfs. */
1927 		MKSTR(slaves,
1928 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1929 		bond_file = fopen(slaves, "r");
1930 		if (bond_file)
1931 			break;
1932 	}
1933 	if (!bond_file)
1934 		return -1;
1935 	/* Use safe format to check maximal buffer length. */
1936 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1937 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1938 		char tmp_str[IF_NAMESIZE + 32];
1939 		struct rte_pci_addr pci_addr;
1940 		struct mlx5_switch_info	info;
1941 
1942 		/* Process slave interface names in the loop. */
1943 		snprintf(tmp_str, sizeof(tmp_str),
1944 			 "/sys/class/net/%s", ifname);
1945 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1946 			DRV_LOG(WARNING,
1947 				"Cannot get PCI address for netdev \"%s\".",
1948 				ifname);
1949 			continue;
1950 		}
1951 		/* Slave interface PCI address match found. */
1952 		snprintf(tmp_str, sizeof(tmp_str),
1953 			 "/sys/class/net/%s/phys_port_name", ifname);
1954 		file = fopen(tmp_str, "rb");
1955 		if (!file)
1956 			break;
1957 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1958 		if (fscanf(file, "%32s", tmp_str) == 1)
1959 			mlx5_translate_port_name(tmp_str, &info);
1960 		fclose(file);
1961 		/* Only process PF ports. */
1962 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1963 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1964 			continue;
1965 		/* Check max bonding member. */
1966 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1967 			DRV_LOG(WARNING, "bonding index out of range, "
1968 				"please increase MLX5_BOND_MAX_PORTS: %s",
1969 				tmp_str);
1970 			break;
1971 		}
1972 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
1973 		if (pci_dev->domain == pci_addr.domain &&
1974 		    pci_dev->bus == pci_addr.bus &&
1975 		    pci_dev->devid == pci_addr.devid &&
1976 		    ((pci_dev->function == 0 &&
1977 		      pci_dev->function + owner == pci_addr.function) ||
1978 		     (pci_dev->function == owner &&
1979 		      pci_addr.function == owner)))
1980 			pf = info.port_name;
1981 		/* Get ifindex. */
1982 		snprintf(tmp_str, sizeof(tmp_str),
1983 			 "/sys/class/net/%s/ifindex", ifname);
1984 		file = fopen(tmp_str, "rb");
1985 		if (!file)
1986 			break;
1987 		ret = fscanf(file, "%u", &ifindex);
1988 		fclose(file);
1989 		if (ret != 1)
1990 			break;
1991 		/* Save bonding info. */
1992 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1993 			sizeof(bond_info->ports[0].ifname));
1994 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1995 		bond_info->ports[info.port_name].ifindex = ifindex;
1996 		bond_info->n_port++;
1997 	}
1998 	if (pf >= 0) {
1999 		/* Get bond interface info */
2000 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2001 					   bond_info->ifname);
2002 		if (ret)
2003 			DRV_LOG(ERR, "unable to get bond info: %s",
2004 				strerror(rte_errno));
2005 		else
2006 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2007 				ifindex, bond_info->ifindex, bond_info->ifname);
2008 	}
2009 	return pf;
2010 }
2011 
2012 static void
2013 mlx5_os_config_default(struct mlx5_dev_config *config)
2014 {
2015 	memset(config, 0, sizeof(*config));
2016 	config->mps = MLX5_ARG_UNSET;
2017 	config->rx_vec_en = 1;
2018 	config->txq_inline_max = MLX5_ARG_UNSET;
2019 	config->txq_inline_min = MLX5_ARG_UNSET;
2020 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2021 	config->txqs_inline = MLX5_ARG_UNSET;
2022 	config->vf_nl_en = 1;
2023 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2024 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2025 	config->dv_esw_en = 1;
2026 	config->dv_flow_en = 1;
2027 	config->decap_en = 1;
2028 	config->log_hp_size = MLX5_ARG_UNSET;
2029 	config->allow_duplicate_pattern = 1;
2030 }
2031 
2032 /**
2033  * Register a PCI device within bonding.
2034  *
2035  * This function spawns Ethernet devices out of a given PCI device and
2036  * bonding owner PF index.
2037  *
2038  * @param[in] cdev
2039  *   Pointer to common mlx5 device structure.
2040  * @param[in] req_eth_da
2041  *   Requested ethdev device argument.
2042  * @param[in] owner_id
2043  *   Requested owner PF port ID within bonding device, default to 0.
2044  *
2045  * @return
2046  *   0 on success, a negative errno value otherwise and rte_errno is set.
2047  */
2048 static int
2049 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
2050 		     struct rte_eth_devargs *req_eth_da,
2051 		     uint16_t owner_id)
2052 {
2053 	struct ibv_device **ibv_list;
2054 	/*
2055 	 * Number of found IB Devices matching with requested PCI BDF.
2056 	 * nd != 1 means there are multiple IB devices over the same
2057 	 * PCI device and we have representors and master.
2058 	 */
2059 	unsigned int nd = 0;
2060 	/*
2061 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2062 	 * we have the single multiport IB device, and there may be
2063 	 * representors attached to some of found ports.
2064 	 */
2065 	unsigned int np = 0;
2066 	/*
2067 	 * Number of DPDK ethernet devices to Spawn - either over
2068 	 * multiple IB devices or multiple ports of single IB device.
2069 	 * Actually this is the number of iterations to spawn.
2070 	 */
2071 	unsigned int ns = 0;
2072 	/*
2073 	 * Bonding device
2074 	 *   < 0 - no bonding device (single one)
2075 	 *  >= 0 - bonding device (value is slave PF index)
2076 	 */
2077 	int bd = -1;
2078 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2079 	struct mlx5_dev_spawn_data *list = NULL;
2080 	struct mlx5_dev_config dev_config;
2081 	unsigned int dev_config_vf;
2082 	struct rte_eth_devargs eth_da = *req_eth_da;
2083 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2084 	struct mlx5_bond_info bond_info;
2085 	int ret = -1;
2086 
2087 	errno = 0;
2088 	ibv_list = mlx5_glue->get_device_list(&ret);
2089 	if (!ibv_list) {
2090 		rte_errno = errno ? errno : ENOSYS;
2091 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2092 		return -rte_errno;
2093 	}
2094 	/*
2095 	 * First scan the list of all Infiniband devices to find
2096 	 * matching ones, gathering into the list.
2097 	 */
2098 	struct ibv_device *ibv_match[ret + 1];
2099 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2100 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2101 	unsigned int i;
2102 
2103 	while (ret-- > 0) {
2104 		struct rte_pci_addr pci_addr;
2105 
2106 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2107 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2108 						nl_rdma, owner_id, &bond_info);
2109 		if (bd >= 0) {
2110 			/*
2111 			 * Bonding device detected. Only one match is allowed,
2112 			 * the bonding is supported over multi-port IB device,
2113 			 * there should be no matches on representor PCI
2114 			 * functions or non VF LAG bonding devices with
2115 			 * specified address.
2116 			 */
2117 			if (nd) {
2118 				DRV_LOG(ERR,
2119 					"multiple PCI match on bonding device"
2120 					"\"%s\" found", ibv_list[ret]->name);
2121 				rte_errno = ENOENT;
2122 				ret = -rte_errno;
2123 				goto exit;
2124 			}
2125 			/* Amend owner pci address if owner PF ID specified. */
2126 			if (eth_da.nb_representor_ports)
2127 				owner_pci.function += owner_id;
2128 			DRV_LOG(INFO,
2129 				"PCI information matches for slave %d bonding device \"%s\"",
2130 				bd, ibv_list[ret]->name);
2131 			ibv_match[nd++] = ibv_list[ret];
2132 			break;
2133 		} else {
2134 			/* Bonding device not found. */
2135 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2136 					      &pci_addr))
2137 				continue;
2138 			if (owner_pci.domain != pci_addr.domain ||
2139 			    owner_pci.bus != pci_addr.bus ||
2140 			    owner_pci.devid != pci_addr.devid ||
2141 			    owner_pci.function != pci_addr.function)
2142 				continue;
2143 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2144 				ibv_list[ret]->name);
2145 			ibv_match[nd++] = ibv_list[ret];
2146 		}
2147 	}
2148 	ibv_match[nd] = NULL;
2149 	if (!nd) {
2150 		/* No device matches, just complain and bail out. */
2151 		DRV_LOG(WARNING,
2152 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
2153 			" are kernel drivers loaded?",
2154 			owner_pci.domain, owner_pci.bus,
2155 			owner_pci.devid, owner_pci.function);
2156 		rte_errno = ENOENT;
2157 		ret = -rte_errno;
2158 		goto exit;
2159 	}
2160 	if (nd == 1) {
2161 		/*
2162 		 * Found single matching device may have multiple ports.
2163 		 * Each port may be representor, we have to check the port
2164 		 * number and check the representors existence.
2165 		 */
2166 		if (nl_rdma >= 0)
2167 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2168 		if (!np)
2169 			DRV_LOG(WARNING,
2170 				"Cannot get IB device \"%s\" ports number.",
2171 				ibv_match[0]->name);
2172 		if (bd >= 0 && !np) {
2173 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2174 			rte_errno = ENOENT;
2175 			ret = -rte_errno;
2176 			goto exit;
2177 		}
2178 	}
2179 	/* Now we can determine the maximal amount of devices to be spawned. */
2180 	list = mlx5_malloc(MLX5_MEM_ZERO,
2181 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2182 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2183 	if (!list) {
2184 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2185 		rte_errno = ENOMEM;
2186 		ret = -rte_errno;
2187 		goto exit;
2188 	}
2189 	if (bd >= 0 || np > 1) {
2190 		/*
2191 		 * Single IB device with multiple ports found,
2192 		 * it may be E-Switch master device and representors.
2193 		 * We have to perform identification through the ports.
2194 		 */
2195 		MLX5_ASSERT(nl_rdma >= 0);
2196 		MLX5_ASSERT(ns == 0);
2197 		MLX5_ASSERT(nd == 1);
2198 		MLX5_ASSERT(np);
2199 		for (i = 1; i <= np; ++i) {
2200 			list[ns].bond_info = &bond_info;
2201 			list[ns].max_port = np;
2202 			list[ns].phys_port = i;
2203 			list[ns].phys_dev_name = ibv_match[0]->name;
2204 			list[ns].eth_dev = NULL;
2205 			list[ns].pci_dev = pci_dev;
2206 			list[ns].cdev = cdev;
2207 			list[ns].pf_bond = bd;
2208 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2209 							   ibv_match[0]->name,
2210 							   i);
2211 			if (!list[ns].ifindex) {
2212 				/*
2213 				 * No network interface index found for the
2214 				 * specified port, it means there is no
2215 				 * representor on this port. It's OK,
2216 				 * there can be disabled ports, for example
2217 				 * if sriov_numvfs < sriov_totalvfs.
2218 				 */
2219 				continue;
2220 			}
2221 			ret = -1;
2222 			if (nl_route >= 0)
2223 				ret = mlx5_nl_switch_info(nl_route,
2224 							  list[ns].ifindex,
2225 							  &list[ns].info);
2226 			if (ret || (!list[ns].info.representor &&
2227 				    !list[ns].info.master)) {
2228 				/*
2229 				 * We failed to recognize representors with
2230 				 * Netlink, let's try to perform the task
2231 				 * with sysfs.
2232 				 */
2233 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2234 							     &list[ns].info);
2235 			}
2236 			if (!ret && bd >= 0) {
2237 				switch (list[ns].info.name_type) {
2238 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2239 					if (np == 1) {
2240 						/*
2241 						 * Force standalone bonding
2242 						 * device for ROCE LAG
2243 						 * confgiurations.
2244 						 */
2245 						list[ns].info.master = 0;
2246 						list[ns].info.representor = 0;
2247 					}
2248 					if (list[ns].info.port_name == bd)
2249 						ns++;
2250 					break;
2251 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2252 					/* Fallthrough */
2253 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2254 					/* Fallthrough */
2255 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2256 					if (list[ns].info.pf_num == bd)
2257 						ns++;
2258 					break;
2259 				default:
2260 					break;
2261 				}
2262 				continue;
2263 			}
2264 			if (!ret && (list[ns].info.representor ^
2265 				     list[ns].info.master))
2266 				ns++;
2267 		}
2268 		if (!ns) {
2269 			DRV_LOG(ERR,
2270 				"Unable to recognize master/representors on the IB device with multiple ports.");
2271 			rte_errno = ENOENT;
2272 			ret = -rte_errno;
2273 			goto exit;
2274 		}
2275 	} else {
2276 		/*
2277 		 * The existence of several matching entries (nd > 1) means
2278 		 * port representors have been instantiated. No existing Verbs
2279 		 * call nor sysfs entries can tell them apart, this can only
2280 		 * be done through Netlink calls assuming kernel drivers are
2281 		 * recent enough to support them.
2282 		 *
2283 		 * In the event of identification failure through Netlink,
2284 		 * try again through sysfs, then:
2285 		 *
2286 		 * 1. A single IB device matches (nd == 1) with single
2287 		 *    port (np=0/1) and is not a representor, assume
2288 		 *    no switch support.
2289 		 *
2290 		 * 2. Otherwise no safe assumptions can be made;
2291 		 *    complain louder and bail out.
2292 		 */
2293 		for (i = 0; i != nd; ++i) {
2294 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2295 			list[ns].bond_info = NULL;
2296 			list[ns].max_port = 1;
2297 			list[ns].phys_port = 1;
2298 			list[ns].phys_dev_name = ibv_match[i]->name;
2299 			list[ns].eth_dev = NULL;
2300 			list[ns].pci_dev = pci_dev;
2301 			list[ns].cdev = cdev;
2302 			list[ns].pf_bond = -1;
2303 			list[ns].ifindex = 0;
2304 			if (nl_rdma >= 0)
2305 				list[ns].ifindex = mlx5_nl_ifindex
2306 							    (nl_rdma,
2307 							     ibv_match[i]->name,
2308 							     1);
2309 			if (!list[ns].ifindex) {
2310 				char ifname[IF_NAMESIZE];
2311 
2312 				/*
2313 				 * Netlink failed, it may happen with old
2314 				 * ib_core kernel driver (before 4.16).
2315 				 * We can assume there is old driver because
2316 				 * here we are processing single ports IB
2317 				 * devices. Let's try sysfs to retrieve
2318 				 * the ifindex. The method works for
2319 				 * master device only.
2320 				 */
2321 				if (nd > 1) {
2322 					/*
2323 					 * Multiple devices found, assume
2324 					 * representors, can not distinguish
2325 					 * master/representor and retrieve
2326 					 * ifindex via sysfs.
2327 					 */
2328 					continue;
2329 				}
2330 				ret = mlx5_get_ifname_sysfs
2331 					(ibv_match[i]->ibdev_path, ifname);
2332 				if (!ret)
2333 					list[ns].ifindex =
2334 						if_nametoindex(ifname);
2335 				if (!list[ns].ifindex) {
2336 					/*
2337 					 * No network interface index found
2338 					 * for the specified device, it means
2339 					 * there it is neither representor
2340 					 * nor master.
2341 					 */
2342 					continue;
2343 				}
2344 			}
2345 			ret = -1;
2346 			if (nl_route >= 0)
2347 				ret = mlx5_nl_switch_info(nl_route,
2348 							  list[ns].ifindex,
2349 							  &list[ns].info);
2350 			if (ret || (!list[ns].info.representor &&
2351 				    !list[ns].info.master)) {
2352 				/*
2353 				 * We failed to recognize representors with
2354 				 * Netlink, let's try to perform the task
2355 				 * with sysfs.
2356 				 */
2357 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2358 							     &list[ns].info);
2359 			}
2360 			if (!ret && (list[ns].info.representor ^
2361 				     list[ns].info.master)) {
2362 				ns++;
2363 			} else if ((nd == 1) &&
2364 				   !list[ns].info.representor &&
2365 				   !list[ns].info.master) {
2366 				/*
2367 				 * Single IB device with one physical port and
2368 				 * attached network device.
2369 				 * May be SRIOV is not enabled or there is no
2370 				 * representors.
2371 				 */
2372 				DRV_LOG(INFO, "No E-Switch support detected.");
2373 				ns++;
2374 				break;
2375 			}
2376 		}
2377 		if (!ns) {
2378 			DRV_LOG(ERR,
2379 				"Unable to recognize master/representors on the multiple IB devices.");
2380 			rte_errno = ENOENT;
2381 			ret = -rte_errno;
2382 			goto exit;
2383 		}
2384 		/*
2385 		 * New kernels may add the switch_id attribute for the case
2386 		 * there is no E-Switch and we wrongly recognized the only
2387 		 * device as master. Override this if there is the single
2388 		 * device with single port and new device name format present.
2389 		 */
2390 		if (nd == 1 &&
2391 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2392 			list[0].info.master = 0;
2393 			list[0].info.representor = 0;
2394 		}
2395 	}
2396 	MLX5_ASSERT(ns);
2397 	/*
2398 	 * Sort list to probe devices in natural order for users convenience
2399 	 * (i.e. master first, then representors from lowest to highest ID).
2400 	 */
2401 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2402 	/* Device specific configuration. */
2403 	switch (pci_dev->id.device_id) {
2404 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2405 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2406 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2407 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2408 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2409 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2410 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2411 		dev_config_vf = 1;
2412 		break;
2413 	default:
2414 		dev_config_vf = 0;
2415 		break;
2416 	}
2417 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2418 		/* Set devargs default values. */
2419 		if (eth_da.nb_mh_controllers == 0) {
2420 			eth_da.nb_mh_controllers = 1;
2421 			eth_da.mh_controllers[0] = 0;
2422 		}
2423 		if (eth_da.nb_ports == 0 && ns > 0) {
2424 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2425 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2426 					pci_dev->device.devargs->args);
2427 			eth_da.nb_ports = 1;
2428 			eth_da.ports[0] = list[0].info.pf_num;
2429 		}
2430 		if (eth_da.nb_representor_ports == 0) {
2431 			eth_da.nb_representor_ports = 1;
2432 			eth_da.representor_ports[0] = 0;
2433 		}
2434 	}
2435 	for (i = 0; i != ns; ++i) {
2436 		uint32_t restore;
2437 
2438 		/* Default configuration. */
2439 		mlx5_os_config_default(&dev_config);
2440 		dev_config.vf = dev_config_vf;
2441 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i],
2442 						 &dev_config, &eth_da);
2443 		if (!list[i].eth_dev) {
2444 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2445 				break;
2446 			/* Device is disabled or already spawned. Ignore it. */
2447 			continue;
2448 		}
2449 		restore = list[i].eth_dev->data->dev_flags;
2450 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2451 		/**
2452 		 * Each representor has a dedicated interrupts vector.
2453 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2454 		 * representor eth_dev object because representor and PF
2455 		 * share the same PCI address.
2456 		 * Override representor device with a dedicated
2457 		 * interrupts handle here.
2458 		 * Representor interrupts handle is released in mlx5_dev_stop().
2459 		 */
2460 		if (list[i].info.representor) {
2461 			struct rte_intr_handle *intr_handle;
2462 			intr_handle = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
2463 						  sizeof(*intr_handle), 0,
2464 						  SOCKET_ID_ANY);
2465 			if (!intr_handle) {
2466 				DRV_LOG(ERR,
2467 					"port %u failed to allocate memory for interrupt handler "
2468 					"Rx interrupts will not be supported",
2469 					i);
2470 				rte_errno = ENOMEM;
2471 				ret = -rte_errno;
2472 				goto exit;
2473 			}
2474 			list[i].eth_dev->intr_handle = intr_handle;
2475 		}
2476 		/* Restore non-PCI flags cleared by the above call. */
2477 		list[i].eth_dev->data->dev_flags |= restore;
2478 		rte_eth_dev_probing_finish(list[i].eth_dev);
2479 	}
2480 	if (i != ns) {
2481 		DRV_LOG(ERR,
2482 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2483 			" encountering an error: %s",
2484 			owner_pci.domain, owner_pci.bus,
2485 			owner_pci.devid, owner_pci.function,
2486 			strerror(rte_errno));
2487 		ret = -rte_errno;
2488 		/* Roll back. */
2489 		while (i--) {
2490 			if (!list[i].eth_dev)
2491 				continue;
2492 			mlx5_dev_close(list[i].eth_dev);
2493 			/* mac_addrs must not be freed because in dev_private */
2494 			list[i].eth_dev->data->mac_addrs = NULL;
2495 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2496 		}
2497 		/* Restore original error. */
2498 		rte_errno = -ret;
2499 	} else {
2500 		ret = 0;
2501 	}
2502 exit:
2503 	/*
2504 	 * Do the routine cleanup:
2505 	 * - close opened Netlink sockets
2506 	 * - free allocated spawn data array
2507 	 * - free the Infiniband device list
2508 	 */
2509 	if (nl_rdma >= 0)
2510 		close(nl_rdma);
2511 	if (nl_route >= 0)
2512 		close(nl_route);
2513 	if (list)
2514 		mlx5_free(list);
2515 	MLX5_ASSERT(ibv_list);
2516 	mlx5_glue->free_device_list(ibv_list);
2517 	return ret;
2518 }
2519 
2520 static int
2521 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2522 			  struct rte_eth_devargs *eth_da)
2523 {
2524 	int ret = 0;
2525 
2526 	if (dev->devargs == NULL)
2527 		return 0;
2528 	memset(eth_da, 0, sizeof(*eth_da));
2529 	/* Parse representor information first from class argument. */
2530 	if (dev->devargs->cls_str)
2531 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2532 	if (ret != 0) {
2533 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2534 			dev->devargs->cls_str);
2535 		return -rte_errno;
2536 	}
2537 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2538 		/* Parse legacy device argument */
2539 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2540 		if (ret) {
2541 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2542 				dev->devargs->args);
2543 			return -rte_errno;
2544 		}
2545 	}
2546 	return 0;
2547 }
2548 
2549 /**
2550  * Callback to register a PCI device.
2551  *
2552  * This function spawns Ethernet devices out of a given PCI device.
2553  *
2554  * @param[in] cdev
2555  *   Pointer to common mlx5 device structure.
2556  *
2557  * @return
2558  *   0 on success, a negative errno value otherwise and rte_errno is set.
2559  */
2560 static int
2561 mlx5_os_pci_probe(struct mlx5_common_device *cdev)
2562 {
2563 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2564 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2565 	int ret = 0;
2566 	uint16_t p;
2567 
2568 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2569 	if (ret != 0)
2570 		return ret;
2571 
2572 	if (eth_da.nb_ports > 0) {
2573 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2574 		for (p = 0; p < eth_da.nb_ports; p++) {
2575 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2576 						   eth_da.ports[p]);
2577 			if (ret)
2578 				break;
2579 		}
2580 		if (ret) {
2581 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2582 				"aborted due to proding failure of PF %u",
2583 				pci_dev->addr.domain, pci_dev->addr.bus,
2584 				pci_dev->addr.devid, pci_dev->addr.function,
2585 				eth_da.ports[p]);
2586 			mlx5_net_remove(cdev);
2587 		}
2588 	} else {
2589 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0);
2590 	}
2591 	return ret;
2592 }
2593 
2594 /* Probe a single SF device on auxiliary bus, no representor support. */
2595 static int
2596 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev)
2597 {
2598 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2599 	struct mlx5_dev_config config;
2600 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2601 	struct rte_device *dev = cdev->dev;
2602 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2603 	struct rte_eth_dev *eth_dev;
2604 	int ret = 0;
2605 
2606 	/* Parse ethdev devargs. */
2607 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2608 	if (ret != 0)
2609 		return ret;
2610 	/* Set default config data. */
2611 	mlx5_os_config_default(&config);
2612 	config.sf = 1;
2613 	/* Init spawn data. */
2614 	spawn.max_port = 1;
2615 	spawn.phys_port = 1;
2616 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2617 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2618 	if (ret < 0) {
2619 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2620 		return ret;
2621 	}
2622 	spawn.ifindex = ret;
2623 	spawn.cdev = cdev;
2624 	/* Spawn device. */
2625 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2626 	if (eth_dev == NULL)
2627 		return -rte_errno;
2628 	/* Post create. */
2629 	eth_dev->intr_handle = &adev->intr_handle;
2630 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2631 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2632 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2633 		eth_dev->data->numa_node = dev->numa_node;
2634 	}
2635 	rte_eth_dev_probing_finish(eth_dev);
2636 	return 0;
2637 }
2638 
2639 /**
2640  * Net class driver callback to probe a device.
2641  *
2642  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2643  *
2644  * @param[in] cdev
2645  *   Pointer to the common mlx5 device.
2646  *
2647  * @return
2648  *   0 on success, a negative errno value otherwise and rte_errno is set.
2649  */
2650 int
2651 mlx5_os_net_probe(struct mlx5_common_device *cdev)
2652 {
2653 	int ret;
2654 
2655 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2656 		mlx5_pmd_socket_init();
2657 	ret = mlx5_init_once();
2658 	if (ret) {
2659 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2660 			strerror(rte_errno));
2661 		return -rte_errno;
2662 	}
2663 	if (mlx5_dev_is_pci(cdev->dev))
2664 		return mlx5_os_pci_probe(cdev);
2665 	else
2666 		return mlx5_os_auxiliary_probe(cdev);
2667 }
2668 
2669 /**
2670  * Cleanup resources when the last device is closed.
2671  */
2672 void
2673 mlx5_os_net_cleanup(void)
2674 {
2675 	mlx5_pmd_socket_uninit();
2676 }
2677 
2678 /**
2679  * Install shared asynchronous device events handler.
2680  * This function is implemented to support event sharing
2681  * between multiple ports of single IB device.
2682  *
2683  * @param sh
2684  *   Pointer to mlx5_dev_ctx_shared object.
2685  */
2686 void
2687 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2688 {
2689 	int ret;
2690 	int flags;
2691 	struct ibv_context *ctx = sh->cdev->ctx;
2692 
2693 	sh->intr_handle.fd = -1;
2694 	flags = fcntl(ctx->async_fd, F_GETFL);
2695 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2696 	if (ret) {
2697 		DRV_LOG(INFO, "failed to change file descriptor async event"
2698 			" queue");
2699 	} else {
2700 		sh->intr_handle.fd = ctx->async_fd;
2701 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2702 		if (rte_intr_callback_register(&sh->intr_handle,
2703 					mlx5_dev_interrupt_handler, sh)) {
2704 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2705 			sh->intr_handle.fd = -1;
2706 		}
2707 	}
2708 	if (sh->devx) {
2709 #ifdef HAVE_IBV_DEVX_ASYNC
2710 		sh->intr_handle_devx.fd = -1;
2711 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2712 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2713 		if (!devx_comp) {
2714 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2715 			return;
2716 		}
2717 		flags = fcntl(devx_comp->fd, F_GETFL);
2718 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2719 		if (ret) {
2720 			DRV_LOG(INFO, "failed to change file descriptor"
2721 				" devx comp");
2722 			return;
2723 		}
2724 		sh->intr_handle_devx.fd = devx_comp->fd;
2725 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2726 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2727 					mlx5_dev_interrupt_handler_devx, sh)) {
2728 			DRV_LOG(INFO, "Fail to install the devx shared"
2729 				" interrupt.");
2730 			sh->intr_handle_devx.fd = -1;
2731 		}
2732 #endif /* HAVE_IBV_DEVX_ASYNC */
2733 	}
2734 }
2735 
2736 /**
2737  * Uninstall shared asynchronous device events handler.
2738  * This function is implemented to support event sharing
2739  * between multiple ports of single IB device.
2740  *
2741  * @param dev
2742  *   Pointer to mlx5_dev_ctx_shared object.
2743  */
2744 void
2745 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2746 {
2747 	if (sh->intr_handle.fd >= 0)
2748 		mlx5_intr_callback_unregister(&sh->intr_handle,
2749 					      mlx5_dev_interrupt_handler, sh);
2750 #ifdef HAVE_IBV_DEVX_ASYNC
2751 	if (sh->intr_handle_devx.fd >= 0)
2752 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2753 				  mlx5_dev_interrupt_handler_devx, sh);
2754 	if (sh->devx_comp)
2755 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2756 #endif
2757 }
2758 
2759 /**
2760  * Read statistics by a named counter.
2761  *
2762  * @param[in] priv
2763  *   Pointer to the private device data structure.
2764  * @param[in] ctr_name
2765  *   Pointer to the name of the statistic counter to read
2766  * @param[out] stat
2767  *   Pointer to read statistic value.
2768  * @return
2769  *   0 on success and stat is valud, 1 if failed to read the value
2770  *   rte_errno is set.
2771  *
2772  */
2773 int
2774 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2775 		      uint64_t *stat)
2776 {
2777 	int fd;
2778 
2779 	if (priv->sh) {
2780 		if (priv->q_counters != NULL &&
2781 		    strcmp(ctr_name, "out_of_buffer") == 0)
2782 			return mlx5_devx_cmd_queue_counter_query
2783 					(priv->q_counters, 0, (uint32_t *)stat);
2784 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2785 		      priv->sh->ibdev_path,
2786 		      priv->dev_port,
2787 		      ctr_name);
2788 		fd = open(path, O_RDONLY);
2789 		/*
2790 		 * in switchdev the file location is not per port
2791 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2792 		 */
2793 		if (fd == -1) {
2794 			MKSTR(path1, "%s/hw_counters/%s",
2795 			      priv->sh->ibdev_path,
2796 			      ctr_name);
2797 			fd = open(path1, O_RDONLY);
2798 		}
2799 		if (fd != -1) {
2800 			char buf[21] = {'\0'};
2801 			ssize_t n = read(fd, buf, sizeof(buf));
2802 
2803 			close(fd);
2804 			if (n != -1) {
2805 				*stat = strtoull(buf, NULL, 10);
2806 				return 0;
2807 			}
2808 		}
2809 	}
2810 	*stat = 0;
2811 	return 1;
2812 }
2813 
2814 /**
2815  * Remove a MAC address from device
2816  *
2817  * @param dev
2818  *   Pointer to Ethernet device structure.
2819  * @param index
2820  *   MAC address index.
2821  */
2822 void
2823 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2824 {
2825 	struct mlx5_priv *priv = dev->data->dev_private;
2826 	const int vf = priv->config.vf;
2827 
2828 	if (vf)
2829 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2830 					mlx5_ifindex(dev), priv->mac_own,
2831 					&dev->data->mac_addrs[index], index);
2832 }
2833 
2834 /**
2835  * Adds a MAC address to the device
2836  *
2837  * @param dev
2838  *   Pointer to Ethernet device structure.
2839  * @param mac_addr
2840  *   MAC address to register.
2841  * @param index
2842  *   MAC address index.
2843  *
2844  * @return
2845  *   0 on success, a negative errno value otherwise
2846  */
2847 int
2848 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2849 		     uint32_t index)
2850 {
2851 	struct mlx5_priv *priv = dev->data->dev_private;
2852 	const int vf = priv->config.vf;
2853 	int ret = 0;
2854 
2855 	if (vf)
2856 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2857 					   mlx5_ifindex(dev), priv->mac_own,
2858 					   mac, index);
2859 	return ret;
2860 }
2861 
2862 /**
2863  * Modify a VF MAC address
2864  *
2865  * @param priv
2866  *   Pointer to device private data.
2867  * @param mac_addr
2868  *   MAC address to modify into.
2869  * @param iface_idx
2870  *   Net device interface index
2871  * @param vf_index
2872  *   VF index
2873  *
2874  * @return
2875  *   0 on success, a negative errno value otherwise
2876  */
2877 int
2878 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2879 			   unsigned int iface_idx,
2880 			   struct rte_ether_addr *mac_addr,
2881 			   int vf_index)
2882 {
2883 	return mlx5_nl_vf_mac_addr_modify
2884 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2885 }
2886 
2887 /**
2888  * Set device promiscuous mode
2889  *
2890  * @param dev
2891  *   Pointer to Ethernet device structure.
2892  * @param enable
2893  *   0 - promiscuous is disabled, otherwise - enabled
2894  *
2895  * @return
2896  *   0 on success, a negative error value otherwise
2897  */
2898 int
2899 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2900 {
2901 	struct mlx5_priv *priv = dev->data->dev_private;
2902 
2903 	return mlx5_nl_promisc(priv->nl_socket_route,
2904 			       mlx5_ifindex(dev), !!enable);
2905 }
2906 
2907 /**
2908  * Set device promiscuous mode
2909  *
2910  * @param dev
2911  *   Pointer to Ethernet device structure.
2912  * @param enable
2913  *   0 - all multicase is disabled, otherwise - enabled
2914  *
2915  * @return
2916  *   0 on success, a negative error value otherwise
2917  */
2918 int
2919 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2920 {
2921 	struct mlx5_priv *priv = dev->data->dev_private;
2922 
2923 	return mlx5_nl_allmulti(priv->nl_socket_route,
2924 				mlx5_ifindex(dev), !!enable);
2925 }
2926 
2927 /**
2928  * Flush device MAC addresses
2929  *
2930  * @param dev
2931  *   Pointer to Ethernet device structure.
2932  *
2933  */
2934 void
2935 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2936 {
2937 	struct mlx5_priv *priv = dev->data->dev_private;
2938 
2939 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2940 			       dev->data->mac_addrs,
2941 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2942 }
2943