xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 081e42dab11d1add2d038fdf2bd4c86b20043d08)
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_mr.h"
48 #include "mlx5_flow.h"
49 #include "rte_pmd_mlx5.h"
50 #include "mlx5_verbs.h"
51 #include "mlx5_nl.h"
52 #include "mlx5_devx.h"
53 
54 #ifndef HAVE_IBV_MLX5_MOD_MPW
55 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
56 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
57 #endif
58 
59 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
60 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
61 #endif
62 
63 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
64 
65 /* Spinlock for mlx5_shared_data allocation. */
66 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
67 
68 /* Process local data for secondary processes. */
69 static struct mlx5_local_data mlx5_local_data;
70 
71 /* rte flow indexed pool configuration. */
72 static struct mlx5_indexed_pool_config icfg[] = {
73 	{
74 		.size = sizeof(struct rte_flow),
75 		.trunk_size = 64,
76 		.need_lock = 1,
77 		.release_mem_en = 0,
78 		.malloc = mlx5_malloc,
79 		.free = mlx5_free,
80 		.per_core_cache = 0,
81 		.type = "ctl_flow_ipool",
82 	},
83 	{
84 		.size = sizeof(struct rte_flow),
85 		.trunk_size = 64,
86 		.grow_trunk = 3,
87 		.grow_shift = 2,
88 		.need_lock = 1,
89 		.release_mem_en = 0,
90 		.malloc = mlx5_malloc,
91 		.free = mlx5_free,
92 		.per_core_cache = 1 << 14,
93 		.type = "rte_flow_ipool",
94 	},
95 	{
96 		.size = sizeof(struct rte_flow),
97 		.trunk_size = 64,
98 		.grow_trunk = 3,
99 		.grow_shift = 2,
100 		.need_lock = 1,
101 		.release_mem_en = 0,
102 		.malloc = mlx5_malloc,
103 		.free = mlx5_free,
104 		.per_core_cache = 0,
105 		.type = "mcp_flow_ipool",
106 	},
107 };
108 
109 /**
110  * Set the completion channel file descriptor interrupt as non-blocking.
111  *
112  * @param[in] rxq_obj
113  *   Pointer to RQ channel object, which includes the channel fd
114  *
115  * @param[out] fd
116  *   The file descriptor (representing the intetrrupt) used in this channel.
117  *
118  * @return
119  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
120  */
121 int
122 mlx5_os_set_nonblock_channel_fd(int fd)
123 {
124 	int flags;
125 
126 	flags = fcntl(fd, F_GETFL);
127 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
128 }
129 
130 /**
131  * Get mlx5 device attributes. The glue function query_device_ex() is called
132  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
133  * device attributes from the glue out parameter.
134  *
135  * @param dev
136  *   Pointer to ibv context.
137  *
138  * @param device_attr
139  *   Pointer to mlx5 device attributes.
140  *
141  * @return
142  *   0 on success, non zero error number otherwise
143  */
144 int
145 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
146 {
147 	int err;
148 	struct ibv_device_attr_ex attr_ex;
149 	memset(device_attr, 0, sizeof(*device_attr));
150 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
151 	if (err)
152 		return err;
153 
154 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
155 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
156 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
157 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
158 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
159 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
160 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
161 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
162 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
163 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
164 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
165 	device_attr->max_rwq_indirection_table_size =
166 		attr_ex.rss_caps.max_rwq_indirection_table_size;
167 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
168 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
169 
170 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
171 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
172 	if (err)
173 		return err;
174 
175 	device_attr->flags = dv_attr.flags;
176 	device_attr->comp_mask = dv_attr.comp_mask;
177 #ifdef HAVE_IBV_MLX5_MOD_SWP
178 	device_attr->sw_parsing_offloads =
179 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
180 #endif
181 	device_attr->min_single_stride_log_num_of_bytes =
182 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
183 	device_attr->max_single_stride_log_num_of_bytes =
184 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
185 	device_attr->min_single_wqe_log_num_of_strides =
186 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
187 	device_attr->max_single_wqe_log_num_of_strides =
188 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
189 	device_attr->stride_supported_qpts =
190 		dv_attr.striding_rq_caps.supported_qpts;
191 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
192 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
193 #endif
194 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
195 		sizeof(device_attr->fw_ver));
196 
197 	return err;
198 }
199 
200 /**
201  * Verbs callback to allocate a memory. This function should allocate the space
202  * according to the size provided residing inside a huge page.
203  * Please note that all allocation must respect the alignment from libmlx5
204  * (i.e. currently rte_mem_page_size()).
205  *
206  * @param[in] size
207  *   The size in bytes of the memory to allocate.
208  * @param[in] data
209  *   A pointer to the callback data.
210  *
211  * @return
212  *   Allocated buffer, NULL otherwise and rte_errno is set.
213  */
214 static void *
215 mlx5_alloc_verbs_buf(size_t size, void *data)
216 {
217 	struct mlx5_dev_ctx_shared *sh = data;
218 	void *ret;
219 	size_t alignment = rte_mem_page_size();
220 	if (alignment == (size_t)-1) {
221 		DRV_LOG(ERR, "Failed to get mem page size");
222 		rte_errno = ENOMEM;
223 		return NULL;
224 	}
225 
226 	MLX5_ASSERT(data != NULL);
227 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
228 	if (!ret && size)
229 		rte_errno = ENOMEM;
230 	return ret;
231 }
232 
233 /**
234  * Detect misc5 support or not
235  *
236  * @param[in] priv
237  *   Device private data pointer
238  */
239 #ifdef HAVE_MLX5DV_DR
240 static void
241 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
242 {
243 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
244 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
245 	 * Case: IPv4--->UDP--->VxLAN--->vni
246 	 */
247 	void *tbl;
248 	struct mlx5_flow_dv_match_params matcher_mask;
249 	void *match_m;
250 	void *matcher;
251 	void *headers_m;
252 	void *misc5_m;
253 	uint32_t *tunnel_header_m;
254 	struct mlx5dv_flow_matcher_attr dv_attr;
255 
256 	memset(&matcher_mask, 0, sizeof(matcher_mask));
257 	matcher_mask.size = sizeof(matcher_mask.buf);
258 	match_m = matcher_mask.buf;
259 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
260 	misc5_m = MLX5_ADDR_OF(fte_match_param,
261 			       match_m, misc_parameters_5);
262 	tunnel_header_m = (uint32_t *)
263 				MLX5_ADDR_OF(fte_match_set_misc5,
264 				misc5_m, tunnel_header_1);
265 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
266 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
267 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
268 	*tunnel_header_m = 0xffffff;
269 
270 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
271 	if (!tbl) {
272 		DRV_LOG(INFO, "No SW steering support");
273 		return;
274 	}
275 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
276 	dv_attr.match_mask = (void *)&matcher_mask,
277 	dv_attr.match_criteria_enable =
278 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
279 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
280 	dv_attr.priority = 3;
281 #ifdef HAVE_MLX5DV_DR_ESWITCH
282 	void *misc2_m;
283 	if (priv->config.dv_esw_en) {
284 		/* FDB enabled reg_c_0 */
285 		dv_attr.match_criteria_enable |=
286 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
287 		misc2_m = MLX5_ADDR_OF(fte_match_param,
288 				       match_m, misc_parameters_2);
289 		MLX5_SET(fte_match_set_misc2, misc2_m,
290 			 metadata_reg_c_0, 0xffff);
291 	}
292 #endif
293 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->ctx,
294 						    &dv_attr, tbl);
295 	if (matcher) {
296 		priv->sh->misc5_cap = 1;
297 		mlx5_glue->dv_destroy_flow_matcher(matcher);
298 	}
299 	mlx5_glue->dr_destroy_flow_tbl(tbl);
300 #else
301 	RTE_SET_USED(priv);
302 #endif
303 }
304 #endif
305 
306 /**
307  * Verbs callback to free a memory.
308  *
309  * @param[in] ptr
310  *   A pointer to the memory to free.
311  * @param[in] data
312  *   A pointer to the callback data.
313  */
314 static void
315 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
316 {
317 	MLX5_ASSERT(data != NULL);
318 	mlx5_free(ptr);
319 }
320 
321 /**
322  * Initialize DR related data within private structure.
323  * Routine checks the reference counter and does actual
324  * resources creation/initialization only if counter is zero.
325  *
326  * @param[in] priv
327  *   Pointer to the private device data structure.
328  *
329  * @return
330  *   Zero on success, positive error code otherwise.
331  */
332 static int
333 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
334 {
335 	struct mlx5_dev_ctx_shared *sh = priv->sh;
336 	char s[MLX5_NAME_SIZE] __rte_unused;
337 	int err;
338 
339 	MLX5_ASSERT(sh && sh->refcnt);
340 	if (sh->refcnt > 1)
341 		return 0;
342 	err = mlx5_alloc_table_hash_list(priv);
343 	if (err)
344 		goto error;
345 	/* The resources below are only valid with DV support. */
346 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
347 	/* Init port id action list. */
348 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
349 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
350 						   flow_dv_port_id_create_cb,
351 						   flow_dv_port_id_match_cb,
352 						   flow_dv_port_id_remove_cb,
353 						   flow_dv_port_id_clone_cb,
354 						 flow_dv_port_id_clone_free_cb);
355 	if (!sh->port_id_action_list)
356 		goto error;
357 	/* Init push vlan action list. */
358 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
359 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
360 						    flow_dv_push_vlan_create_cb,
361 						    flow_dv_push_vlan_match_cb,
362 						    flow_dv_push_vlan_remove_cb,
363 						    flow_dv_push_vlan_clone_cb,
364 					       flow_dv_push_vlan_clone_free_cb);
365 	if (!sh->push_vlan_action_list)
366 		goto error;
367 	/* Init sample action list. */
368 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
369 	sh->sample_action_list = mlx5_list_create(s, sh, true,
370 						  flow_dv_sample_create_cb,
371 						  flow_dv_sample_match_cb,
372 						  flow_dv_sample_remove_cb,
373 						  flow_dv_sample_clone_cb,
374 						  flow_dv_sample_clone_free_cb);
375 	if (!sh->sample_action_list)
376 		goto error;
377 	/* Init dest array action list. */
378 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
379 	sh->dest_array_list = mlx5_list_create(s, sh, true,
380 					       flow_dv_dest_array_create_cb,
381 					       flow_dv_dest_array_match_cb,
382 					       flow_dv_dest_array_remove_cb,
383 					       flow_dv_dest_array_clone_cb,
384 					      flow_dv_dest_array_clone_free_cb);
385 	if (!sh->dest_array_list)
386 		goto error;
387 #endif
388 #ifdef HAVE_MLX5DV_DR
389 	void *domain;
390 
391 	/* Reference counter is zero, we should initialize structures. */
392 	domain = mlx5_glue->dr_create_domain(sh->ctx,
393 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
394 	if (!domain) {
395 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
396 		err = errno;
397 		goto error;
398 	}
399 	sh->rx_domain = domain;
400 	domain = mlx5_glue->dr_create_domain(sh->ctx,
401 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
402 	if (!domain) {
403 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
404 		err = errno;
405 		goto error;
406 	}
407 	sh->tx_domain = domain;
408 #ifdef HAVE_MLX5DV_DR_ESWITCH
409 	if (priv->config.dv_esw_en) {
410 		domain  = mlx5_glue->dr_create_domain
411 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
412 		if (!domain) {
413 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
414 			err = errno;
415 			goto error;
416 		}
417 		sh->fdb_domain = domain;
418 	}
419 	/*
420 	 * The drop action is just some dummy placeholder in rdma-core. It
421 	 * does not belong to domains and has no any attributes, and, can be
422 	 * shared by the entire device.
423 	 */
424 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
425 	if (!sh->dr_drop_action) {
426 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
427 		err = errno;
428 		goto error;
429 	}
430 #endif
431 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
432 		err = mlx5_alloc_tunnel_hub(sh);
433 	if (err) {
434 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
435 		goto error;
436 	}
437 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
438 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
439 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
440 		if (sh->fdb_domain)
441 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
442 	}
443 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
444 	if (!priv->config.allow_duplicate_pattern) {
445 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
446 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
447 #endif
448 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
449 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
450 		if (sh->fdb_domain)
451 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
452 	}
453 
454 	__mlx5_discovery_misc5_cap(priv);
455 #endif /* HAVE_MLX5DV_DR */
456 	sh->default_miss_action =
457 			mlx5_glue->dr_create_flow_action_default_miss();
458 	if (!sh->default_miss_action)
459 		DRV_LOG(WARNING, "Default miss action is not supported.");
460 	return 0;
461 error:
462 	/* Rollback the created objects. */
463 	if (sh->rx_domain) {
464 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
465 		sh->rx_domain = NULL;
466 	}
467 	if (sh->tx_domain) {
468 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
469 		sh->tx_domain = NULL;
470 	}
471 	if (sh->fdb_domain) {
472 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
473 		sh->fdb_domain = NULL;
474 	}
475 	if (sh->dr_drop_action) {
476 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
477 		sh->dr_drop_action = NULL;
478 	}
479 	if (sh->pop_vlan_action) {
480 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
481 		sh->pop_vlan_action = NULL;
482 	}
483 	if (sh->encaps_decaps) {
484 		mlx5_hlist_destroy(sh->encaps_decaps);
485 		sh->encaps_decaps = NULL;
486 	}
487 	if (sh->modify_cmds) {
488 		mlx5_hlist_destroy(sh->modify_cmds);
489 		sh->modify_cmds = NULL;
490 	}
491 	if (sh->tag_table) {
492 		/* tags should be destroyed with flow before. */
493 		mlx5_hlist_destroy(sh->tag_table);
494 		sh->tag_table = NULL;
495 	}
496 	if (sh->tunnel_hub) {
497 		mlx5_release_tunnel_hub(sh, priv->dev_port);
498 		sh->tunnel_hub = NULL;
499 	}
500 	mlx5_free_table_hash_list(priv);
501 	if (sh->port_id_action_list) {
502 		mlx5_list_destroy(sh->port_id_action_list);
503 		sh->port_id_action_list = NULL;
504 	}
505 	if (sh->push_vlan_action_list) {
506 		mlx5_list_destroy(sh->push_vlan_action_list);
507 		sh->push_vlan_action_list = NULL;
508 	}
509 	if (sh->sample_action_list) {
510 		mlx5_list_destroy(sh->sample_action_list);
511 		sh->sample_action_list = NULL;
512 	}
513 	if (sh->dest_array_list) {
514 		mlx5_list_destroy(sh->dest_array_list);
515 		sh->dest_array_list = NULL;
516 	}
517 	return err;
518 }
519 
520 /**
521  * Destroy DR related data within private structure.
522  *
523  * @param[in] priv
524  *   Pointer to the private device data structure.
525  */
526 void
527 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
528 {
529 	struct mlx5_dev_ctx_shared *sh = priv->sh;
530 
531 	MLX5_ASSERT(sh && sh->refcnt);
532 	if (sh->refcnt > 1)
533 		return;
534 #ifdef HAVE_MLX5DV_DR
535 	if (sh->rx_domain) {
536 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
537 		sh->rx_domain = NULL;
538 	}
539 	if (sh->tx_domain) {
540 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
541 		sh->tx_domain = NULL;
542 	}
543 #ifdef HAVE_MLX5DV_DR_ESWITCH
544 	if (sh->fdb_domain) {
545 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
546 		sh->fdb_domain = NULL;
547 	}
548 	if (sh->dr_drop_action) {
549 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
550 		sh->dr_drop_action = NULL;
551 	}
552 #endif
553 	if (sh->pop_vlan_action) {
554 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
555 		sh->pop_vlan_action = NULL;
556 	}
557 #endif /* HAVE_MLX5DV_DR */
558 	if (sh->default_miss_action)
559 		mlx5_glue->destroy_flow_action
560 				(sh->default_miss_action);
561 	if (sh->encaps_decaps) {
562 		mlx5_hlist_destroy(sh->encaps_decaps);
563 		sh->encaps_decaps = NULL;
564 	}
565 	if (sh->modify_cmds) {
566 		mlx5_hlist_destroy(sh->modify_cmds);
567 		sh->modify_cmds = NULL;
568 	}
569 	if (sh->tag_table) {
570 		/* tags should be destroyed with flow before. */
571 		mlx5_hlist_destroy(sh->tag_table);
572 		sh->tag_table = NULL;
573 	}
574 	if (sh->tunnel_hub) {
575 		mlx5_release_tunnel_hub(sh, priv->dev_port);
576 		sh->tunnel_hub = NULL;
577 	}
578 	mlx5_free_table_hash_list(priv);
579 	if (sh->port_id_action_list) {
580 		mlx5_list_destroy(sh->port_id_action_list);
581 		sh->port_id_action_list = NULL;
582 	}
583 	if (sh->push_vlan_action_list) {
584 		mlx5_list_destroy(sh->push_vlan_action_list);
585 		sh->push_vlan_action_list = NULL;
586 	}
587 	if (sh->sample_action_list) {
588 		mlx5_list_destroy(sh->sample_action_list);
589 		sh->sample_action_list = NULL;
590 	}
591 	if (sh->dest_array_list) {
592 		mlx5_list_destroy(sh->dest_array_list);
593 		sh->dest_array_list = NULL;
594 	}
595 }
596 
597 /**
598  * Initialize shared data between primary and secondary process.
599  *
600  * A memzone is reserved by primary process and secondary processes attach to
601  * the memzone.
602  *
603  * @return
604  *   0 on success, a negative errno value otherwise and rte_errno is set.
605  */
606 static int
607 mlx5_init_shared_data(void)
608 {
609 	const struct rte_memzone *mz;
610 	int ret = 0;
611 
612 	rte_spinlock_lock(&mlx5_shared_data_lock);
613 	if (mlx5_shared_data == NULL) {
614 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
615 			/* Allocate shared memory. */
616 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
617 						 sizeof(*mlx5_shared_data),
618 						 SOCKET_ID_ANY, 0);
619 			if (mz == NULL) {
620 				DRV_LOG(ERR,
621 					"Cannot allocate mlx5 shared data");
622 				ret = -rte_errno;
623 				goto error;
624 			}
625 			mlx5_shared_data = mz->addr;
626 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
627 			rte_spinlock_init(&mlx5_shared_data->lock);
628 		} else {
629 			/* Lookup allocated shared memory. */
630 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
631 			if (mz == NULL) {
632 				DRV_LOG(ERR,
633 					"Cannot attach mlx5 shared data");
634 				ret = -rte_errno;
635 				goto error;
636 			}
637 			mlx5_shared_data = mz->addr;
638 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
639 		}
640 	}
641 error:
642 	rte_spinlock_unlock(&mlx5_shared_data_lock);
643 	return ret;
644 }
645 
646 /**
647  * PMD global initialization.
648  *
649  * Independent from individual device, this function initializes global
650  * per-PMD data structures distinguishing primary and secondary processes.
651  * Hence, each initialization is called once per a process.
652  *
653  * @return
654  *   0 on success, a negative errno value otherwise and rte_errno is set.
655  */
656 static int
657 mlx5_init_once(void)
658 {
659 	struct mlx5_shared_data *sd;
660 	struct mlx5_local_data *ld = &mlx5_local_data;
661 	int ret = 0;
662 
663 	if (mlx5_init_shared_data())
664 		return -rte_errno;
665 	sd = mlx5_shared_data;
666 	MLX5_ASSERT(sd);
667 	rte_spinlock_lock(&sd->lock);
668 	switch (rte_eal_process_type()) {
669 	case RTE_PROC_PRIMARY:
670 		if (sd->init_done)
671 			break;
672 		LIST_INIT(&sd->mem_event_cb_list);
673 		rte_rwlock_init(&sd->mem_event_rwlock);
674 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
675 						mlx5_mr_mem_event_cb, NULL);
676 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
677 					   mlx5_mp_os_primary_handle);
678 		if (ret)
679 			goto out;
680 		sd->init_done = true;
681 		break;
682 	case RTE_PROC_SECONDARY:
683 		if (ld->init_done)
684 			break;
685 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
686 					     mlx5_mp_os_secondary_handle);
687 		if (ret)
688 			goto out;
689 		++sd->secondary_cnt;
690 		ld->init_done = true;
691 		break;
692 	default:
693 		break;
694 	}
695 out:
696 	rte_spinlock_unlock(&sd->lock);
697 	return ret;
698 }
699 
700 /**
701  * Create the Tx queue DevX/Verbs object.
702  *
703  * @param dev
704  *   Pointer to Ethernet device.
705  * @param idx
706  *   Queue index in DPDK Tx queue array.
707  *
708  * @return
709  *   0 on success, a negative errno value otherwise and rte_errno is set.
710  */
711 static int
712 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
713 {
714 	struct mlx5_priv *priv = dev->data->dev_private;
715 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
716 	struct mlx5_txq_ctrl *txq_ctrl =
717 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
718 
719 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
720 		return mlx5_txq_devx_obj_new(dev, idx);
721 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
722 	if (!priv->config.dv_esw_en)
723 		return mlx5_txq_devx_obj_new(dev, idx);
724 #endif
725 	return mlx5_txq_ibv_obj_new(dev, idx);
726 }
727 
728 /**
729  * Release an Tx DevX/verbs queue object.
730  *
731  * @param txq_obj
732  *   DevX/Verbs Tx queue object.
733  */
734 static void
735 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
736 {
737 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
738 		mlx5_txq_devx_obj_release(txq_obj);
739 		return;
740 	}
741 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
742 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
743 		mlx5_txq_devx_obj_release(txq_obj);
744 		return;
745 	}
746 #endif
747 	mlx5_txq_ibv_obj_release(txq_obj);
748 }
749 
750 /**
751  * DV flow counter mode detect and config.
752  *
753  * @param dev
754  *   Pointer to rte_eth_dev structure.
755  *
756  */
757 static void
758 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
759 {
760 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
761 	struct mlx5_priv *priv = dev->data->dev_private;
762 	struct mlx5_dev_ctx_shared *sh = priv->sh;
763 	bool fallback;
764 
765 #ifndef HAVE_IBV_DEVX_ASYNC
766 	fallback = true;
767 #else
768 	fallback = false;
769 	if (!priv->config.devx || !priv->config.dv_flow_en ||
770 	    !priv->config.hca_attr.flow_counters_dump ||
771 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
772 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
773 		fallback = true;
774 #endif
775 	if (fallback)
776 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
777 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
778 			priv->config.hca_attr.flow_counters_dump,
779 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
780 	/* Initialize fallback mode only on the port initializes sh. */
781 	if (sh->refcnt == 1)
782 		sh->cmng.counter_fallback = fallback;
783 	else if (fallback != sh->cmng.counter_fallback)
784 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
785 			"with others:%d.", PORT_ID(priv), fallback);
786 #endif
787 }
788 
789 /**
790  * DR flow drop action support detect.
791  *
792  * @param dev
793  *   Pointer to rte_eth_dev structure.
794  *
795  */
796 static void
797 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
798 {
799 #ifdef HAVE_MLX5DV_DR
800 	struct mlx5_priv *priv = dev->data->dev_private;
801 
802 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
803 		return;
804 	/**
805 	 * DR supports drop action placeholder when it is supported;
806 	 * otherwise, use the queue drop action.
807 	 */
808 	if (mlx5_flow_discover_dr_action_support(dev))
809 		priv->root_drop_action = priv->drop_queue.hrxq->action;
810 	else
811 		priv->root_drop_action = priv->sh->dr_drop_action;
812 #endif
813 }
814 
815 static void
816 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
817 {
818 	struct mlx5_priv *priv = dev->data->dev_private;
819 	void *ctx = priv->sh->ctx;
820 
821 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
822 	if (!priv->q_counters) {
823 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
824 		struct ibv_wq *wq;
825 
826 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
827 			"by DevX - fall-back to use the kernel driver global "
828 			"queue counter.", dev->data->port_id);
829 		/* Create WQ by kernel and query its queue counter ID. */
830 		if (cq) {
831 			wq = mlx5_glue->create_wq(ctx,
832 						  &(struct ibv_wq_init_attr){
833 						    .wq_type = IBV_WQT_RQ,
834 						    .max_wr = 1,
835 						    .max_sge = 1,
836 						    .pd = priv->sh->pd,
837 						    .cq = cq,
838 						});
839 			if (wq) {
840 				/* Counter is assigned only on RDY state. */
841 				int ret = mlx5_glue->modify_wq(wq,
842 						 &(struct ibv_wq_attr){
843 						 .attr_mask = IBV_WQ_ATTR_STATE,
844 						 .wq_state = IBV_WQS_RDY,
845 						});
846 
847 				if (ret == 0)
848 					mlx5_devx_cmd_wq_query(wq,
849 							 &priv->counter_set_id);
850 				claim_zero(mlx5_glue->destroy_wq(wq));
851 			}
852 			claim_zero(mlx5_glue->destroy_cq(cq));
853 		}
854 	} else {
855 		priv->counter_set_id = priv->q_counters->id;
856 	}
857 	if (priv->counter_set_id == 0)
858 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
859 			"available.", dev->data->port_id);
860 }
861 
862 /**
863  * Check if representor spawn info match devargs.
864  *
865  * @param spawn
866  *   Verbs device parameters (name, port, switch_info) to spawn.
867  * @param eth_da
868  *   Device devargs to probe.
869  *
870  * @return
871  *   Match result.
872  */
873 static bool
874 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
875 		       struct rte_eth_devargs *eth_da)
876 {
877 	struct mlx5_switch_info *switch_info = &spawn->info;
878 	unsigned int p, f;
879 	uint16_t id;
880 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
881 						      eth_da->type);
882 
883 	switch (eth_da->type) {
884 	case RTE_ETH_REPRESENTOR_SF:
885 		if (!(spawn->info.port_name == -1 &&
886 		      switch_info->name_type ==
887 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
888 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
889 			rte_errno = EBUSY;
890 			return false;
891 		}
892 		break;
893 	case RTE_ETH_REPRESENTOR_VF:
894 		/* Allows HPF representor index -1 as exception. */
895 		if (!(spawn->info.port_name == -1 &&
896 		      switch_info->name_type ==
897 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
898 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
899 			rte_errno = EBUSY;
900 			return false;
901 		}
902 		break;
903 	case RTE_ETH_REPRESENTOR_NONE:
904 		rte_errno = EBUSY;
905 		return false;
906 	default:
907 		rte_errno = ENOTSUP;
908 		DRV_LOG(ERR, "unsupported representor type");
909 		return false;
910 	}
911 	/* Check representor ID: */
912 	for (p = 0; p < eth_da->nb_ports; ++p) {
913 		if (spawn->pf_bond < 0) {
914 			/* For non-LAG mode, allow and ignore pf. */
915 			switch_info->pf_num = eth_da->ports[p];
916 			repr_id = mlx5_representor_id_encode(switch_info,
917 							     eth_da->type);
918 		}
919 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
920 			id = MLX5_REPRESENTOR_ID
921 				(eth_da->ports[p], eth_da->type,
922 				 eth_da->representor_ports[f]);
923 			if (repr_id == id)
924 				return true;
925 		}
926 	}
927 	rte_errno = EBUSY;
928 	return false;
929 }
930 
931 
932 /**
933  * Spawn an Ethernet device from Verbs information.
934  *
935  * @param dpdk_dev
936  *   Backing DPDK device.
937  * @param spawn
938  *   Verbs device parameters (name, port, switch_info) to spawn.
939  * @param config
940  *   Device configuration parameters.
941  * @param config
942  *   Device arguments.
943  *
944  * @return
945  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
946  *   is set. The following errors are defined:
947  *
948  *   EBUSY: device is not supposed to be spawned.
949  *   EEXIST: device is already spawned
950  */
951 static struct rte_eth_dev *
952 mlx5_dev_spawn(struct rte_device *dpdk_dev,
953 	       struct mlx5_dev_spawn_data *spawn,
954 	       struct mlx5_dev_config *config,
955 	       struct rte_eth_devargs *eth_da)
956 {
957 	const struct mlx5_switch_info *switch_info = &spawn->info;
958 	struct mlx5_dev_ctx_shared *sh = NULL;
959 	struct ibv_port_attr port_attr;
960 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
961 	struct rte_eth_dev *eth_dev = NULL;
962 	struct mlx5_priv *priv = NULL;
963 	int err = 0;
964 	unsigned int hw_padding = 0;
965 	unsigned int mps;
966 	unsigned int mpls_en = 0;
967 	unsigned int swp = 0;
968 	unsigned int mprq = 0;
969 	unsigned int mprq_min_stride_size_n = 0;
970 	unsigned int mprq_max_stride_size_n = 0;
971 	unsigned int mprq_min_stride_num_n = 0;
972 	unsigned int mprq_max_stride_num_n = 0;
973 	struct rte_ether_addr mac;
974 	char name[RTE_ETH_NAME_MAX_LEN];
975 	int own_domain_id = 0;
976 	uint16_t port_id;
977 	struct mlx5_port_info vport_info = { .query_flags = 0 };
978 	int i;
979 
980 	/* Determine if this port representor is supposed to be spawned. */
981 	if (switch_info->representor && dpdk_dev->devargs &&
982 	    !mlx5_representor_match(spawn, eth_da))
983 		return NULL;
984 	/* Build device name. */
985 	if (spawn->pf_bond < 0) {
986 		/* Single device. */
987 		if (!switch_info->representor)
988 			strlcpy(name, dpdk_dev->name, sizeof(name));
989 		else
990 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
991 				 dpdk_dev->name,
992 				 switch_info->name_type ==
993 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
994 				 switch_info->port_name);
995 	} else {
996 		/* Bonding device. */
997 		if (!switch_info->representor) {
998 			err = snprintf(name, sizeof(name), "%s_%s",
999 				 dpdk_dev->name,
1000 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
1001 		} else {
1002 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1003 				dpdk_dev->name,
1004 				mlx5_os_get_dev_device_name(spawn->phys_dev),
1005 				switch_info->ctrl_num,
1006 				switch_info->pf_num,
1007 				switch_info->name_type ==
1008 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1009 				switch_info->port_name);
1010 		}
1011 	}
1012 	if (err >= (int)sizeof(name))
1013 		DRV_LOG(WARNING, "device name overflow %s", name);
1014 	/* check if the device is already spawned */
1015 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1016 		rte_errno = EEXIST;
1017 		return NULL;
1018 	}
1019 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1020 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1021 		struct mlx5_mp_id mp_id;
1022 
1023 		eth_dev = rte_eth_dev_attach_secondary(name);
1024 		if (eth_dev == NULL) {
1025 			DRV_LOG(ERR, "can not attach rte ethdev");
1026 			rte_errno = ENOMEM;
1027 			return NULL;
1028 		}
1029 		eth_dev->device = dpdk_dev;
1030 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1031 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1032 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1033 		err = mlx5_proc_priv_init(eth_dev);
1034 		if (err)
1035 			return NULL;
1036 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1037 		/* Receive command fd from primary process */
1038 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1039 		if (err < 0)
1040 			goto err_secondary;
1041 		/* Remap UAR for Tx queues. */
1042 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1043 		if (err)
1044 			goto err_secondary;
1045 		/*
1046 		 * Ethdev pointer is still required as input since
1047 		 * the primary device is not accessible from the
1048 		 * secondary process.
1049 		 */
1050 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1051 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1052 		return eth_dev;
1053 err_secondary:
1054 		mlx5_dev_close(eth_dev);
1055 		return NULL;
1056 	}
1057 	/*
1058 	 * Some parameters ("tx_db_nc" in particularly) are needed in
1059 	 * advance to create dv/verbs device context. We proceed the
1060 	 * devargs here to get ones, and later proceed devargs again
1061 	 * to override some hardware settings.
1062 	 */
1063 	err = mlx5_args(config, dpdk_dev->devargs);
1064 	if (err) {
1065 		err = rte_errno;
1066 		DRV_LOG(ERR, "failed to process device arguments: %s",
1067 			strerror(rte_errno));
1068 		goto error;
1069 	}
1070 	if (config->dv_miss_info) {
1071 		if (switch_info->master || switch_info->representor)
1072 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
1073 	}
1074 	mlx5_malloc_mem_select(config->sys_mem_en);
1075 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
1076 	if (!sh)
1077 		return NULL;
1078 	config->devx = sh->devx;
1079 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1080 	config->dest_tir = 1;
1081 #endif
1082 #ifdef HAVE_IBV_MLX5_MOD_SWP
1083 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1084 #endif
1085 	/*
1086 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1087 	 * as all ConnectX-5 devices.
1088 	 */
1089 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1090 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1091 #endif
1092 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1093 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1094 #endif
1095 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1096 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1097 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1098 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1099 			mps = MLX5_MPW_ENHANCED;
1100 		} else {
1101 			DRV_LOG(DEBUG, "MPW is supported");
1102 			mps = MLX5_MPW;
1103 		}
1104 	} else {
1105 		DRV_LOG(DEBUG, "MPW isn't supported");
1106 		mps = MLX5_MPW_DISABLED;
1107 	}
1108 #ifdef HAVE_IBV_MLX5_MOD_SWP
1109 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1110 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1111 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1112 #endif
1113 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
1114 		MLX5_SW_PARSING_TSO_CAP);
1115 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1116 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1117 		struct mlx5dv_striding_rq_caps mprq_caps =
1118 			dv_attr.striding_rq_caps;
1119 
1120 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1121 			mprq_caps.min_single_stride_log_num_of_bytes);
1122 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1123 			mprq_caps.max_single_stride_log_num_of_bytes);
1124 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1125 			mprq_caps.min_single_wqe_log_num_of_strides);
1126 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1127 			mprq_caps.max_single_wqe_log_num_of_strides);
1128 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1129 			mprq_caps.supported_qpts);
1130 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1131 		mprq = 1;
1132 		mprq_min_stride_size_n =
1133 			mprq_caps.min_single_stride_log_num_of_bytes;
1134 		mprq_max_stride_size_n =
1135 			mprq_caps.max_single_stride_log_num_of_bytes;
1136 		mprq_min_stride_num_n =
1137 			mprq_caps.min_single_wqe_log_num_of_strides;
1138 		mprq_max_stride_num_n =
1139 			mprq_caps.max_single_wqe_log_num_of_strides;
1140 	}
1141 #endif
1142 	/* Rx CQE compression is enabled by default. */
1143 	config->cqe_comp = 1;
1144 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1145 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1146 		config->tunnel_en = dv_attr.tunnel_offloads_caps &
1147 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1148 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1149 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1150 	}
1151 	if (config->tunnel_en) {
1152 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1153 		config->tunnel_en &
1154 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1155 		config->tunnel_en &
1156 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1157 		config->tunnel_en &
1158 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1159 		);
1160 	} else {
1161 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1162 	}
1163 #else
1164 	DRV_LOG(WARNING,
1165 		"tunnel offloading disabled due to old OFED/rdma-core version");
1166 #endif
1167 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1168 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1169 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1170 		   (dv_attr.tunnel_offloads_caps &
1171 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1172 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1173 		mpls_en ? "" : "not ");
1174 #else
1175 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1176 		" old OFED/rdma-core version or firmware configuration");
1177 #endif
1178 	config->mpls_en = mpls_en;
1179 	/* Check port status. */
1180 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1181 	if (err) {
1182 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1183 		goto error;
1184 	}
1185 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1186 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1187 		err = EINVAL;
1188 		goto error;
1189 	}
1190 	if (port_attr.state != IBV_PORT_ACTIVE)
1191 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1192 			mlx5_glue->port_state_str(port_attr.state),
1193 			port_attr.state);
1194 	/* Allocate private eth device data. */
1195 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1196 			   sizeof(*priv),
1197 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1198 	if (priv == NULL) {
1199 		DRV_LOG(ERR, "priv allocation failure");
1200 		err = ENOMEM;
1201 		goto error;
1202 	}
1203 	priv->sh = sh;
1204 	priv->dev_port = spawn->phys_port;
1205 	priv->pci_dev = spawn->pci_dev;
1206 	priv->mtu = RTE_ETHER_MTU;
1207 	/* Some internal functions rely on Netlink sockets, open them now. */
1208 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1209 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1210 	priv->representor = !!switch_info->representor;
1211 	priv->master = !!switch_info->master;
1212 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1213 	priv->vport_meta_tag = 0;
1214 	priv->vport_meta_mask = 0;
1215 	priv->pf_bond = spawn->pf_bond;
1216 
1217 	DRV_LOG(DEBUG,
1218 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1219 		priv->dev_port, dpdk_dev->bus->name,
1220 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1221 		priv->master, priv->representor, priv->pf_bond);
1222 
1223 	/*
1224 	 * If we have E-Switch we should determine the vport attributes.
1225 	 * E-Switch may use either source vport field or reg_c[0] metadata
1226 	 * register to match on vport index. The engaged part of metadata
1227 	 * register is defined by mask.
1228 	 */
1229 	if (switch_info->representor || switch_info->master) {
1230 		err = mlx5_glue->devx_port_query(sh->ctx,
1231 						 spawn->phys_port,
1232 						 &vport_info);
1233 		if (err) {
1234 			DRV_LOG(WARNING,
1235 				"can't query devx port %d on device %s",
1236 				spawn->phys_port,
1237 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1238 			vport_info.query_flags = 0;
1239 		}
1240 	}
1241 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1242 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1243 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1244 		if (!priv->vport_meta_mask) {
1245 			DRV_LOG(ERR, "vport zero mask for port %d"
1246 				     " on bonding device %s",
1247 				     spawn->phys_port,
1248 				     mlx5_os_get_dev_device_name
1249 							(spawn->phys_dev));
1250 			err = ENOTSUP;
1251 			goto error;
1252 		}
1253 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1254 			DRV_LOG(ERR, "invalid vport tag for port %d"
1255 				     " on bonding device %s",
1256 				     spawn->phys_port,
1257 				     mlx5_os_get_dev_device_name
1258 							(spawn->phys_dev));
1259 			err = ENOTSUP;
1260 			goto error;
1261 		}
1262 	}
1263 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1264 		priv->vport_id = vport_info.vport_id;
1265 	} else if (spawn->pf_bond >= 0 &&
1266 		   (switch_info->representor || switch_info->master)) {
1267 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1268 			     " on bonding device %s",
1269 			     spawn->phys_port,
1270 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1271 		err = ENOTSUP;
1272 		goto error;
1273 	} else {
1274 		/*
1275 		 * Suppose vport index in compatible way. Kernel/rdma_core
1276 		 * support single E-Switch per PF configurations only and
1277 		 * vport_id field contains the vport index for associated VF,
1278 		 * which is deduced from representor port name.
1279 		 * For example, let's have the IB device port 10, it has
1280 		 * attached network device eth0, which has port name attribute
1281 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1282 		 * as 3 (2+1). This assigning schema should be changed if the
1283 		 * multiple E-Switch instances per PF configurations or/and PCI
1284 		 * subfunctions are added.
1285 		 */
1286 		priv->vport_id = switch_info->representor ?
1287 				 switch_info->port_name + 1 : -1;
1288 	}
1289 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1290 							  eth_da->type);
1291 	/*
1292 	 * Look for sibling devices in order to reuse their switch domain
1293 	 * if any, otherwise allocate one.
1294 	 */
1295 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1296 		const struct mlx5_priv *opriv =
1297 			rte_eth_devices[port_id].data->dev_private;
1298 
1299 		if (!opriv ||
1300 		    opriv->sh != priv->sh ||
1301 			opriv->domain_id ==
1302 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1303 			continue;
1304 		priv->domain_id = opriv->domain_id;
1305 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1306 			priv->dev_port, priv->domain_id);
1307 		break;
1308 	}
1309 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1310 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1311 		if (err) {
1312 			err = rte_errno;
1313 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1314 				strerror(rte_errno));
1315 			goto error;
1316 		}
1317 		own_domain_id = 1;
1318 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1319 			priv->dev_port, priv->domain_id);
1320 	}
1321 	/* Override some values set by hardware configuration. */
1322 	mlx5_args(config, dpdk_dev->devargs);
1323 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
1324 	if (err)
1325 		goto error;
1326 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1327 			    IBV_DEVICE_RAW_IP_CSUM);
1328 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1329 		(config->hw_csum ? "" : "not "));
1330 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1331 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1332 	DRV_LOG(DEBUG, "counters are not supported");
1333 #endif
1334 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1335 	if (config->dv_flow_en) {
1336 		DRV_LOG(WARNING, "DV flow is not supported");
1337 		config->dv_flow_en = 0;
1338 	}
1339 #endif
1340 	if (spawn->max_port > UINT8_MAX) {
1341 		/* Verbs can't support ports larger than 255 by design. */
1342 		DRV_LOG(ERR, "can't support IB ports > UINT8_MAX");
1343 		err = EINVAL;
1344 		goto error;
1345 	}
1346 	config->ind_table_max_size =
1347 		sh->device_attr.max_rwq_indirection_table_size;
1348 	/*
1349 	 * Remove this check once DPDK supports larger/variable
1350 	 * indirection tables.
1351 	 */
1352 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1353 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1354 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1355 		config->ind_table_max_size);
1356 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1357 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1358 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1359 		(config->hw_vlan_strip ? "" : "not "));
1360 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1361 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1362 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1363 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1364 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1365 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1366 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1367 #endif
1368 	if (config->hw_padding && !hw_padding) {
1369 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1370 		config->hw_padding = 0;
1371 	} else if (config->hw_padding) {
1372 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1373 	}
1374 	config->tso = (sh->device_attr.max_tso > 0 &&
1375 		      (sh->device_attr.tso_supported_qpts &
1376 		       (1 << IBV_QPT_RAW_PACKET)));
1377 	if (config->tso)
1378 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1379 	/*
1380 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1381 	 * by default.
1382 	 */
1383 	if (config->mps == MLX5_ARG_UNSET)
1384 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1385 							  MLX5_MPW_DISABLED;
1386 	else
1387 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1388 	DRV_LOG(INFO, "%sMPS is %s",
1389 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1390 		config->mps == MLX5_MPW ? "legacy " : "",
1391 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1392 	if (config->devx) {
1393 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1394 		if (err) {
1395 			err = -err;
1396 			goto error;
1397 		}
1398 		/* Check relax ordering support. */
1399 		if (!haswell_broadwell_cpu) {
1400 			sh->cmng.relaxed_ordering_write =
1401 				config->hca_attr.relaxed_ordering_write;
1402 			sh->cmng.relaxed_ordering_read =
1403 				config->hca_attr.relaxed_ordering_read;
1404 		} else {
1405 			sh->cmng.relaxed_ordering_read = 0;
1406 			sh->cmng.relaxed_ordering_write = 0;
1407 		}
1408 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1409 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1410 		sh->steering_format_version =
1411 			config->hca_attr.steering_format_version;
1412 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1413 		/* Check for LRO support. */
1414 		if (config->dest_tir && config->hca_attr.lro_cap &&
1415 		    config->dv_flow_en) {
1416 			/* TBD check tunnel lro caps. */
1417 			config->lro.supported = config->hca_attr.lro_cap;
1418 			DRV_LOG(DEBUG, "Device supports LRO");
1419 			/*
1420 			 * If LRO timeout is not configured by application,
1421 			 * use the minimal supported value.
1422 			 */
1423 			if (!config->lro.timeout)
1424 				config->lro.timeout =
1425 				config->hca_attr.lro_timer_supported_periods[0];
1426 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1427 				config->lro.timeout);
1428 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1429 				"required for coalescing is %d bytes",
1430 				config->hca_attr.lro_min_mss_size);
1431 		}
1432 #if defined(HAVE_MLX5DV_DR) && \
1433 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1434 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1435 		if (config->hca_attr.qos.sup &&
1436 		    config->hca_attr.qos.flow_meter_old &&
1437 		    config->dv_flow_en) {
1438 			uint8_t reg_c_mask =
1439 				config->hca_attr.qos.flow_meter_reg_c_ids;
1440 			/*
1441 			 * Meter needs two REG_C's for color match and pre-sfx
1442 			 * flow match. Here get the REG_C for color match.
1443 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1444 			 */
1445 			reg_c_mask &= 0xfc;
1446 			if (__builtin_popcount(reg_c_mask) < 1) {
1447 				priv->mtr_en = 0;
1448 				DRV_LOG(WARNING, "No available register for"
1449 					" meter.");
1450 			} else {
1451 				/*
1452 				 * The meter color register is used by the
1453 				 * flow-hit feature as well.
1454 				 * The flow-hit feature must use REG_C_3
1455 				 * Prefer REG_C_3 if it is available.
1456 				 */
1457 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1458 					priv->mtr_color_reg = REG_C_3;
1459 				else
1460 					priv->mtr_color_reg = ffs(reg_c_mask)
1461 							      - 1 + REG_C_0;
1462 				priv->mtr_en = 1;
1463 				priv->mtr_reg_share =
1464 				      config->hca_attr.qos.flow_meter;
1465 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1466 					priv->mtr_color_reg);
1467 			}
1468 		}
1469 		if (config->hca_attr.qos.sup &&
1470 			config->hca_attr.qos.flow_meter_aso_sup) {
1471 			uint32_t log_obj_size =
1472 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1473 			if (log_obj_size >=
1474 			config->hca_attr.qos.log_meter_aso_granularity &&
1475 			log_obj_size <=
1476 			config->hca_attr.qos.log_meter_aso_max_alloc)
1477 				sh->meter_aso_en = 1;
1478 		}
1479 		if (priv->mtr_en) {
1480 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1481 			if (err) {
1482 				err = -err;
1483 				goto error;
1484 			}
1485 		}
1486 		if (config->hca_attr.flow.tunnel_header_0_1)
1487 			sh->tunnel_header_0_1 = 1;
1488 #endif
1489 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1490 		if (config->hca_attr.flow_hit_aso &&
1491 		    priv->mtr_color_reg == REG_C_3) {
1492 			sh->flow_hit_aso_en = 1;
1493 			err = mlx5_flow_aso_age_mng_init(sh);
1494 			if (err) {
1495 				err = -err;
1496 				goto error;
1497 			}
1498 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1499 		}
1500 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1501 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1502 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1503 		if (config->hca_attr.ct_offload &&
1504 		    priv->mtr_color_reg == REG_C_3) {
1505 			err = mlx5_flow_aso_ct_mng_init(sh);
1506 			if (err) {
1507 				err = -err;
1508 				goto error;
1509 			}
1510 			DRV_LOG(DEBUG, "CT ASO is supported.");
1511 			sh->ct_aso_en = 1;
1512 		}
1513 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1514 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1515 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1516 		    config->dv_flow_en) {
1517 			priv->sampler_en = 1;
1518 			DRV_LOG(DEBUG, "Sampler enabled!");
1519 		} else {
1520 			priv->sampler_en = 0;
1521 			if (!config->hca_attr.log_max_ft_sampler_num)
1522 				DRV_LOG(WARNING,
1523 					"No available register for sampler.");
1524 			else
1525 				DRV_LOG(DEBUG, "DV flow is not supported!");
1526 		}
1527 #endif
1528 	}
1529 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1530 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1531 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1532 		config->cqe_comp = 0;
1533 	}
1534 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1535 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1536 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1537 				 " format isn't supported.");
1538 		config->cqe_comp = 0;
1539 	}
1540 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1541 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1542 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1543 				 " format isn't supported.");
1544 		config->cqe_comp = 0;
1545 	}
1546 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1547 			config->cqe_comp ? "" : "not ");
1548 	if (config->tx_pp) {
1549 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1550 			config->hca_attr.dev_freq_khz);
1551 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1552 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1553 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1554 			config->hca_attr.cross_channel ? "" : "not ");
1555 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1556 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1557 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1558 			config->hca_attr.non_wire_sq ? "" : "not ");
1559 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1560 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1561 			config->hca_attr.log_max_static_sq_wq);
1562 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1563 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1564 		if (!config->devx) {
1565 			DRV_LOG(ERR, "DevX is required for packet pacing");
1566 			err = ENODEV;
1567 			goto error;
1568 		}
1569 		if (!config->hca_attr.qos.packet_pacing) {
1570 			DRV_LOG(ERR, "Packet pacing is not supported");
1571 			err = ENODEV;
1572 			goto error;
1573 		}
1574 		if (!config->hca_attr.cross_channel) {
1575 			DRV_LOG(ERR, "Cross channel operations are"
1576 				     " required for packet pacing");
1577 			err = ENODEV;
1578 			goto error;
1579 		}
1580 		if (!config->hca_attr.wqe_index_ignore) {
1581 			DRV_LOG(ERR, "WQE index ignore feature is"
1582 				     " required for packet pacing");
1583 			err = ENODEV;
1584 			goto error;
1585 		}
1586 		if (!config->hca_attr.non_wire_sq) {
1587 			DRV_LOG(ERR, "Non-wire SQ feature is"
1588 				     " required for packet pacing");
1589 			err = ENODEV;
1590 			goto error;
1591 		}
1592 		if (!config->hca_attr.log_max_static_sq_wq) {
1593 			DRV_LOG(ERR, "Static WQE SQ feature is"
1594 				     " required for packet pacing");
1595 			err = ENODEV;
1596 			goto error;
1597 		}
1598 		if (!config->hca_attr.qos.wqe_rate_pp) {
1599 			DRV_LOG(ERR, "WQE rate mode is required"
1600 				     " for packet pacing");
1601 			err = ENODEV;
1602 			goto error;
1603 		}
1604 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1605 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1606 			     " can't create queues for packet pacing");
1607 		err = ENODEV;
1608 		goto error;
1609 #endif
1610 	}
1611 	if (config->devx) {
1612 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1613 
1614 		err = config->hca_attr.access_register_user ?
1615 			mlx5_devx_cmd_register_read
1616 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1617 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1618 		if (!err) {
1619 			uint32_t ts_mode;
1620 
1621 			/* MTUTC register is read successfully. */
1622 			ts_mode = MLX5_GET(register_mtutc, reg,
1623 					   time_stamp_mode);
1624 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1625 				config->rt_timestamp = 1;
1626 		} else {
1627 			/* Kernel does not support register reading. */
1628 			if (config->hca_attr.dev_freq_khz ==
1629 						 (NS_PER_S / MS_PER_S))
1630 				config->rt_timestamp = 1;
1631 		}
1632 	}
1633 	/*
1634 	 * If HW has bug working with tunnel packet decapsulation and
1635 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1636 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1637 	 */
1638 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1639 		config->hw_fcs_strip = 0;
1640 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1641 		(config->hw_fcs_strip ? "" : "not "));
1642 	if (config->mprq.enabled && mprq) {
1643 		if (config->mprq.stride_num_n &&
1644 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1645 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1646 			config->mprq.stride_num_n =
1647 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1648 						mprq_min_stride_num_n),
1649 					mprq_max_stride_num_n);
1650 			DRV_LOG(WARNING,
1651 				"the number of strides"
1652 				" for Multi-Packet RQ is out of range,"
1653 				" setting default value (%u)",
1654 				1 << config->mprq.stride_num_n);
1655 		}
1656 		if (config->mprq.stride_size_n &&
1657 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1658 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1659 			config->mprq.stride_size_n =
1660 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1661 						mprq_min_stride_size_n),
1662 					mprq_max_stride_size_n);
1663 			DRV_LOG(WARNING,
1664 				"the size of a stride"
1665 				" for Multi-Packet RQ is out of range,"
1666 				" setting default value (%u)",
1667 				1 << config->mprq.stride_size_n);
1668 		}
1669 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1670 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1671 	} else if (config->mprq.enabled && !mprq) {
1672 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1673 		config->mprq.enabled = 0;
1674 	}
1675 	if (config->max_dump_files_num == 0)
1676 		config->max_dump_files_num = 128;
1677 	eth_dev = rte_eth_dev_allocate(name);
1678 	if (eth_dev == NULL) {
1679 		DRV_LOG(ERR, "can not allocate rte ethdev");
1680 		err = ENOMEM;
1681 		goto error;
1682 	}
1683 	if (priv->representor) {
1684 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1685 		eth_dev->data->representor_id = priv->representor_id;
1686 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1687 			struct mlx5_priv *opriv =
1688 				rte_eth_devices[port_id].data->dev_private;
1689 			if (opriv &&
1690 			    opriv->master &&
1691 			    opriv->domain_id == priv->domain_id &&
1692 			    opriv->sh == priv->sh) {
1693 				eth_dev->data->backer_port_id = port_id;
1694 				break;
1695 			}
1696 		}
1697 		if (port_id >= RTE_MAX_ETHPORTS)
1698 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1699 	}
1700 	priv->mp_id.port_id = eth_dev->data->port_id;
1701 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1702 	/*
1703 	 * Store associated network device interface index. This index
1704 	 * is permanent throughout the lifetime of device. So, we may store
1705 	 * the ifindex here and use the cached value further.
1706 	 */
1707 	MLX5_ASSERT(spawn->ifindex);
1708 	priv->if_index = spawn->ifindex;
1709 	eth_dev->data->dev_private = priv;
1710 	priv->dev_data = eth_dev->data;
1711 	eth_dev->data->mac_addrs = priv->mac;
1712 	eth_dev->device = dpdk_dev;
1713 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1714 	/* Configure the first MAC address by default. */
1715 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1716 		DRV_LOG(ERR,
1717 			"port %u cannot get MAC address, is mlx5_en"
1718 			" loaded? (errno: %s)",
1719 			eth_dev->data->port_id, strerror(rte_errno));
1720 		err = ENODEV;
1721 		goto error;
1722 	}
1723 	DRV_LOG(INFO,
1724 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1725 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1726 #ifdef RTE_LIBRTE_MLX5_DEBUG
1727 	{
1728 		char ifname[MLX5_NAMESIZE];
1729 
1730 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1731 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1732 				eth_dev->data->port_id, ifname);
1733 		else
1734 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1735 				eth_dev->data->port_id);
1736 	}
1737 #endif
1738 	/* Get actual MTU if possible. */
1739 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1740 	if (err) {
1741 		err = rte_errno;
1742 		goto error;
1743 	}
1744 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1745 		priv->mtu);
1746 	/* Initialize burst functions to prevent crashes before link-up. */
1747 	eth_dev->rx_pkt_burst = removed_rx_burst;
1748 	eth_dev->tx_pkt_burst = removed_tx_burst;
1749 	eth_dev->dev_ops = &mlx5_dev_ops;
1750 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1751 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1752 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1753 	/* Register MAC address. */
1754 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1755 	if (config->vf && config->vf_nl_en)
1756 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1757 				      mlx5_ifindex(eth_dev),
1758 				      eth_dev->data->mac_addrs,
1759 				      MLX5_MAX_MAC_ADDRESSES);
1760 	priv->ctrl_flows = 0;
1761 	rte_spinlock_init(&priv->flow_list_lock);
1762 	TAILQ_INIT(&priv->flow_meters);
1763 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1764 	if (!priv->mtr_profile_tbl)
1765 		goto error;
1766 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1767 	mlx5_glue->dv_set_context_attr(sh->ctx,
1768 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1769 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1770 				.alloc = &mlx5_alloc_verbs_buf,
1771 				.free = &mlx5_free_verbs_buf,
1772 				.data = sh,
1773 			}));
1774 	/* Bring Ethernet device up. */
1775 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1776 		eth_dev->data->port_id);
1777 	mlx5_set_link_up(eth_dev);
1778 	/*
1779 	 * Even though the interrupt handler is not installed yet,
1780 	 * interrupts will still trigger on the async_fd from
1781 	 * Verbs context returned by ibv_open_device().
1782 	 */
1783 	mlx5_link_update(eth_dev, 0);
1784 #ifdef HAVE_MLX5DV_DR_ESWITCH
1785 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1786 	      (switch_info->representor || switch_info->master)))
1787 		config->dv_esw_en = 0;
1788 #else
1789 	config->dv_esw_en = 0;
1790 #endif
1791 	/* Detect minimal data bytes to inline. */
1792 	mlx5_set_min_inline(spawn, config);
1793 	/* Store device configuration on private structure. */
1794 	priv->config = *config;
1795 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1796 		icfg[i].release_mem_en = !!config->reclaim_mode;
1797 		if (config->reclaim_mode)
1798 			icfg[i].per_core_cache = 0;
1799 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1800 		if (!priv->flows[i])
1801 			goto error;
1802 	}
1803 	/* Create context for virtual machine VLAN workaround. */
1804 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1805 	if (config->dv_flow_en) {
1806 		err = mlx5_alloc_shared_dr(priv);
1807 		if (err)
1808 			goto error;
1809 	}
1810 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1811 		priv->obj_ops = devx_obj_ops;
1812 		priv->obj_ops.drop_action_create =
1813 						ibv_obj_ops.drop_action_create;
1814 		priv->obj_ops.drop_action_destroy =
1815 						ibv_obj_ops.drop_action_destroy;
1816 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1817 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1818 #else
1819 		if (config->dv_esw_en)
1820 			priv->obj_ops.txq_obj_modify =
1821 						ibv_obj_ops.txq_obj_modify;
1822 #endif
1823 		/* Use specific wrappers for Tx object. */
1824 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1825 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1826 		mlx5_queue_counter_id_prepare(eth_dev);
1827 		priv->obj_ops.lb_dummy_queue_create =
1828 					mlx5_rxq_ibv_obj_dummy_lb_create;
1829 		priv->obj_ops.lb_dummy_queue_release =
1830 					mlx5_rxq_ibv_obj_dummy_lb_release;
1831 	} else {
1832 		priv->obj_ops = ibv_obj_ops;
1833 	}
1834 	if (config->tx_pp &&
1835 	    (priv->config.dv_esw_en ||
1836 	     priv->obj_ops.txq_obj_new != mlx5_os_txq_obj_new)) {
1837 		/*
1838 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1839 		 * packet pacing and already checked above.
1840 		 * Hence, we should only make sure the SQs will be created
1841 		 * with DevX, not with Verbs.
1842 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1843 		 * with Clock Queue UAR as required for Tx scheduling.
1844 		 */
1845 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1846 		err = ENODEV;
1847 		goto error;
1848 	}
1849 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1850 	if (!priv->drop_queue.hrxq)
1851 		goto error;
1852 	/* Supported Verbs flow priority number detection. */
1853 	err = mlx5_flow_discover_priorities(eth_dev);
1854 	if (err < 0) {
1855 		err = -err;
1856 		goto error;
1857 	}
1858 	priv->config.flow_prio = err;
1859 	if (!priv->config.dv_esw_en &&
1860 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1861 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1862 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1863 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1864 	}
1865 	mlx5_set_metadata_mask(eth_dev);
1866 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1867 	    !priv->sh->dv_regc0_mask) {
1868 		DRV_LOG(ERR, "metadata mode %u is not supported "
1869 			     "(no metadata reg_c[0] is available)",
1870 			     priv->config.dv_xmeta_en);
1871 			err = ENOTSUP;
1872 			goto error;
1873 	}
1874 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1875 				       mlx5_hrxq_create_cb,
1876 				       mlx5_hrxq_match_cb,
1877 				       mlx5_hrxq_remove_cb,
1878 				       mlx5_hrxq_clone_cb,
1879 				       mlx5_hrxq_clone_free_cb);
1880 	if (!priv->hrxqs)
1881 		goto error;
1882 	rte_rwlock_init(&priv->ind_tbls_lock);
1883 	/* Query availability of metadata reg_c's. */
1884 	err = mlx5_flow_discover_mreg_c(eth_dev);
1885 	if (err < 0) {
1886 		err = -err;
1887 		goto error;
1888 	}
1889 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1890 		DRV_LOG(DEBUG,
1891 			"port %u extensive metadata register is not supported",
1892 			eth_dev->data->port_id);
1893 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1894 			DRV_LOG(ERR, "metadata mode %u is not supported "
1895 				     "(no metadata registers available)",
1896 				     priv->config.dv_xmeta_en);
1897 			err = ENOTSUP;
1898 			goto error;
1899 		}
1900 	}
1901 	if (priv->config.dv_flow_en &&
1902 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1903 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1904 	    priv->sh->dv_regc0_mask) {
1905 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1906 						      MLX5_FLOW_MREG_HTABLE_SZ,
1907 						      false, true, eth_dev,
1908 						      flow_dv_mreg_create_cb,
1909 						      flow_dv_mreg_match_cb,
1910 						      flow_dv_mreg_remove_cb,
1911 						      flow_dv_mreg_clone_cb,
1912 						    flow_dv_mreg_clone_free_cb);
1913 		if (!priv->mreg_cp_tbl) {
1914 			err = ENOMEM;
1915 			goto error;
1916 		}
1917 	}
1918 	rte_spinlock_init(&priv->shared_act_sl);
1919 	mlx5_flow_counter_mode_config(eth_dev);
1920 	mlx5_flow_drop_action_config(eth_dev);
1921 	if (priv->config.dv_flow_en)
1922 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1923 	return eth_dev;
1924 error:
1925 	if (priv) {
1926 		if (priv->mreg_cp_tbl)
1927 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1928 		if (priv->sh)
1929 			mlx5_os_free_shared_dr(priv);
1930 		if (priv->nl_socket_route >= 0)
1931 			close(priv->nl_socket_route);
1932 		if (priv->nl_socket_rdma >= 0)
1933 			close(priv->nl_socket_rdma);
1934 		if (priv->vmwa_context)
1935 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1936 		if (eth_dev && priv->drop_queue.hrxq)
1937 			mlx5_drop_action_destroy(eth_dev);
1938 		if (priv->mtr_profile_tbl)
1939 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1940 		if (own_domain_id)
1941 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1942 		if (priv->hrxqs)
1943 			mlx5_list_destroy(priv->hrxqs);
1944 		mlx5_free(priv);
1945 		if (eth_dev != NULL)
1946 			eth_dev->data->dev_private = NULL;
1947 	}
1948 	if (eth_dev != NULL) {
1949 		/* mac_addrs must not be freed alone because part of
1950 		 * dev_private
1951 		 **/
1952 		eth_dev->data->mac_addrs = NULL;
1953 		rte_eth_dev_release_port(eth_dev);
1954 	}
1955 	if (sh)
1956 		mlx5_free_shared_dev_ctx(sh);
1957 	MLX5_ASSERT(err > 0);
1958 	rte_errno = err;
1959 	return NULL;
1960 }
1961 
1962 /**
1963  * Comparison callback to sort device data.
1964  *
1965  * This is meant to be used with qsort().
1966  *
1967  * @param a[in]
1968  *   Pointer to pointer to first data object.
1969  * @param b[in]
1970  *   Pointer to pointer to second data object.
1971  *
1972  * @return
1973  *   0 if both objects are equal, less than 0 if the first argument is less
1974  *   than the second, greater than 0 otherwise.
1975  */
1976 static int
1977 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1978 {
1979 	const struct mlx5_switch_info *si_a =
1980 		&((const struct mlx5_dev_spawn_data *)a)->info;
1981 	const struct mlx5_switch_info *si_b =
1982 		&((const struct mlx5_dev_spawn_data *)b)->info;
1983 	int ret;
1984 
1985 	/* Master device first. */
1986 	ret = si_b->master - si_a->master;
1987 	if (ret)
1988 		return ret;
1989 	/* Then representor devices. */
1990 	ret = si_b->representor - si_a->representor;
1991 	if (ret)
1992 		return ret;
1993 	/* Unidentified devices come last in no specific order. */
1994 	if (!si_a->representor)
1995 		return 0;
1996 	/* Order representors by name. */
1997 	return si_a->port_name - si_b->port_name;
1998 }
1999 
2000 /**
2001  * Match PCI information for possible slaves of bonding device.
2002  *
2003  * @param[in] ibv_dev
2004  *   Pointer to Infiniband device structure.
2005  * @param[in] pci_dev
2006  *   Pointer to primary PCI address structure to match.
2007  * @param[in] nl_rdma
2008  *   Netlink RDMA group socket handle.
2009  * @param[in] owner
2010  *   Rerepsentor owner PF index.
2011  * @param[out] bond_info
2012  *   Pointer to bonding information.
2013  *
2014  * @return
2015  *   negative value if no bonding device found, otherwise
2016  *   positive index of slave PF in bonding.
2017  */
2018 static int
2019 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2020 			   const struct rte_pci_addr *pci_dev,
2021 			   int nl_rdma, uint16_t owner,
2022 			   struct mlx5_bond_info *bond_info)
2023 {
2024 	char ifname[IF_NAMESIZE + 1];
2025 	unsigned int ifindex;
2026 	unsigned int np, i;
2027 	FILE *bond_file = NULL, *file;
2028 	int pf = -1;
2029 	int ret;
2030 
2031 	/*
2032 	 * Try to get master device name. If something goes
2033 	 * wrong suppose the lack of kernel support and no
2034 	 * bonding devices.
2035 	 */
2036 	memset(bond_info, 0, sizeof(*bond_info));
2037 	if (nl_rdma < 0)
2038 		return -1;
2039 	if (!strstr(ibv_dev->name, "bond"))
2040 		return -1;
2041 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2042 	if (!np)
2043 		return -1;
2044 	/*
2045 	 * The Master device might not be on the predefined
2046 	 * port (not on port index 1, it is not garanted),
2047 	 * we have to scan all Infiniband device port and
2048 	 * find master.
2049 	 */
2050 	for (i = 1; i <= np; ++i) {
2051 		/* Check whether Infiniband port is populated. */
2052 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2053 		if (!ifindex)
2054 			continue;
2055 		if (!if_indextoname(ifindex, ifname))
2056 			continue;
2057 		/* Try to read bonding slave names from sysfs. */
2058 		MKSTR(slaves,
2059 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2060 		bond_file = fopen(slaves, "r");
2061 		if (bond_file)
2062 			break;
2063 	}
2064 	if (!bond_file)
2065 		return -1;
2066 	/* Use safe format to check maximal buffer length. */
2067 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2068 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2069 		char tmp_str[IF_NAMESIZE + 32];
2070 		struct rte_pci_addr pci_addr;
2071 		struct mlx5_switch_info	info;
2072 
2073 		/* Process slave interface names in the loop. */
2074 		snprintf(tmp_str, sizeof(tmp_str),
2075 			 "/sys/class/net/%s", ifname);
2076 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
2077 			DRV_LOG(WARNING, "can not get PCI address"
2078 					 " for netdev \"%s\"", ifname);
2079 			continue;
2080 		}
2081 		/* Slave interface PCI address match found. */
2082 		snprintf(tmp_str, sizeof(tmp_str),
2083 			 "/sys/class/net/%s/phys_port_name", ifname);
2084 		file = fopen(tmp_str, "rb");
2085 		if (!file)
2086 			break;
2087 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2088 		if (fscanf(file, "%32s", tmp_str) == 1)
2089 			mlx5_translate_port_name(tmp_str, &info);
2090 		fclose(file);
2091 		/* Only process PF ports. */
2092 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2093 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2094 			continue;
2095 		/* Check max bonding member. */
2096 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2097 			DRV_LOG(WARNING, "bonding index out of range, "
2098 				"please increase MLX5_BOND_MAX_PORTS: %s",
2099 				tmp_str);
2100 			break;
2101 		}
2102 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
2103 		if (pci_dev->domain == pci_addr.domain &&
2104 		    pci_dev->bus == pci_addr.bus &&
2105 		    pci_dev->devid == pci_addr.devid &&
2106 		    ((pci_dev->function == 0 &&
2107 		      pci_dev->function + owner == pci_addr.function) ||
2108 		     (pci_dev->function == owner &&
2109 		      pci_addr.function == owner)))
2110 			pf = info.port_name;
2111 		/* Get ifindex. */
2112 		snprintf(tmp_str, sizeof(tmp_str),
2113 			 "/sys/class/net/%s/ifindex", ifname);
2114 		file = fopen(tmp_str, "rb");
2115 		if (!file)
2116 			break;
2117 		ret = fscanf(file, "%u", &ifindex);
2118 		fclose(file);
2119 		if (ret != 1)
2120 			break;
2121 		/* Save bonding info. */
2122 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2123 			sizeof(bond_info->ports[0].ifname));
2124 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2125 		bond_info->ports[info.port_name].ifindex = ifindex;
2126 		bond_info->n_port++;
2127 	}
2128 	if (pf >= 0) {
2129 		/* Get bond interface info */
2130 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2131 					   bond_info->ifname);
2132 		if (ret)
2133 			DRV_LOG(ERR, "unable to get bond info: %s",
2134 				strerror(rte_errno));
2135 		else
2136 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2137 				ifindex, bond_info->ifindex, bond_info->ifname);
2138 	}
2139 	return pf;
2140 }
2141 
2142 static void
2143 mlx5_os_config_default(struct mlx5_dev_config *config)
2144 {
2145 	memset(config, 0, sizeof(*config));
2146 	config->mps = MLX5_ARG_UNSET;
2147 	config->dbnc = MLX5_ARG_UNSET;
2148 	config->rx_vec_en = 1;
2149 	config->txq_inline_max = MLX5_ARG_UNSET;
2150 	config->txq_inline_min = MLX5_ARG_UNSET;
2151 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2152 	config->txqs_inline = MLX5_ARG_UNSET;
2153 	config->vf_nl_en = 1;
2154 	config->mr_ext_memseg_en = 1;
2155 	config->mr_mempool_reg_en = 1;
2156 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2157 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2158 	config->dv_esw_en = 1;
2159 	config->dv_flow_en = 1;
2160 	config->decap_en = 1;
2161 	config->log_hp_size = MLX5_ARG_UNSET;
2162 	config->allow_duplicate_pattern = 1;
2163 }
2164 
2165 /**
2166  * Register a PCI device within bonding.
2167  *
2168  * This function spawns Ethernet devices out of a given PCI device and
2169  * bonding owner PF index.
2170  *
2171  * @param[in] pci_dev
2172  *   PCI device information.
2173  * @param[in] req_eth_da
2174  *   Requested ethdev device argument.
2175  * @param[in] owner_id
2176  *   Requested owner PF port ID within bonding device, default to 0.
2177  *
2178  * @return
2179  *   0 on success, a negative errno value otherwise and rte_errno is set.
2180  */
2181 static int
2182 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
2183 		     struct rte_eth_devargs *req_eth_da,
2184 		     uint16_t owner_id)
2185 {
2186 	struct ibv_device **ibv_list;
2187 	/*
2188 	 * Number of found IB Devices matching with requested PCI BDF.
2189 	 * nd != 1 means there are multiple IB devices over the same
2190 	 * PCI device and we have representors and master.
2191 	 */
2192 	unsigned int nd = 0;
2193 	/*
2194 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2195 	 * we have the single multiport IB device, and there may be
2196 	 * representors attached to some of found ports.
2197 	 */
2198 	unsigned int np = 0;
2199 	/*
2200 	 * Number of DPDK ethernet devices to Spawn - either over
2201 	 * multiple IB devices or multiple ports of single IB device.
2202 	 * Actually this is the number of iterations to spawn.
2203 	 */
2204 	unsigned int ns = 0;
2205 	/*
2206 	 * Bonding device
2207 	 *   < 0 - no bonding device (single one)
2208 	 *  >= 0 - bonding device (value is slave PF index)
2209 	 */
2210 	int bd = -1;
2211 	struct mlx5_dev_spawn_data *list = NULL;
2212 	struct mlx5_dev_config dev_config;
2213 	unsigned int dev_config_vf;
2214 	struct rte_eth_devargs eth_da = *req_eth_da;
2215 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2216 	struct mlx5_bond_info bond_info;
2217 	int ret = -1;
2218 
2219 	errno = 0;
2220 	ibv_list = mlx5_glue->get_device_list(&ret);
2221 	if (!ibv_list) {
2222 		rte_errno = errno ? errno : ENOSYS;
2223 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2224 		return -rte_errno;
2225 	}
2226 	/*
2227 	 * First scan the list of all Infiniband devices to find
2228 	 * matching ones, gathering into the list.
2229 	 */
2230 	struct ibv_device *ibv_match[ret + 1];
2231 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2232 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2233 	unsigned int i;
2234 
2235 	while (ret-- > 0) {
2236 		struct rte_pci_addr pci_addr;
2237 
2238 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2239 		bd = mlx5_device_bond_pci_match
2240 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2241 				 &bond_info);
2242 		if (bd >= 0) {
2243 			/*
2244 			 * Bonding device detected. Only one match is allowed,
2245 			 * the bonding is supported over multi-port IB device,
2246 			 * there should be no matches on representor PCI
2247 			 * functions or non VF LAG bonding devices with
2248 			 * specified address.
2249 			 */
2250 			if (nd) {
2251 				DRV_LOG(ERR,
2252 					"multiple PCI match on bonding device"
2253 					"\"%s\" found", ibv_list[ret]->name);
2254 				rte_errno = ENOENT;
2255 				ret = -rte_errno;
2256 				goto exit;
2257 			}
2258 			/* Amend owner pci address if owner PF ID specified. */
2259 			if (eth_da.nb_representor_ports)
2260 				owner_pci.function += owner_id;
2261 			DRV_LOG(INFO, "PCI information matches for"
2262 				      " slave %d bonding device \"%s\"",
2263 				      bd, ibv_list[ret]->name);
2264 			ibv_match[nd++] = ibv_list[ret];
2265 			break;
2266 		} else {
2267 			/* Bonding device not found. */
2268 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2269 					      &pci_addr))
2270 				continue;
2271 			if (owner_pci.domain != pci_addr.domain ||
2272 			    owner_pci.bus != pci_addr.bus ||
2273 			    owner_pci.devid != pci_addr.devid ||
2274 			    owner_pci.function != pci_addr.function)
2275 				continue;
2276 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2277 				ibv_list[ret]->name);
2278 			ibv_match[nd++] = ibv_list[ret];
2279 		}
2280 	}
2281 	ibv_match[nd] = NULL;
2282 	if (!nd) {
2283 		/* No device matches, just complain and bail out. */
2284 		DRV_LOG(WARNING,
2285 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2286 			" are kernel drivers loaded?",
2287 			owner_pci.domain, owner_pci.bus,
2288 			owner_pci.devid, owner_pci.function);
2289 		rte_errno = ENOENT;
2290 		ret = -rte_errno;
2291 		goto exit;
2292 	}
2293 	if (nd == 1) {
2294 		/*
2295 		 * Found single matching device may have multiple ports.
2296 		 * Each port may be representor, we have to check the port
2297 		 * number and check the representors existence.
2298 		 */
2299 		if (nl_rdma >= 0)
2300 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2301 		if (!np)
2302 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2303 					 " ports number", ibv_match[0]->name);
2304 		if (bd >= 0 && !np) {
2305 			DRV_LOG(ERR, "can not get ports"
2306 				     " for bonding device");
2307 			rte_errno = ENOENT;
2308 			ret = -rte_errno;
2309 			goto exit;
2310 		}
2311 	}
2312 	/*
2313 	 * Now we can determine the maximal
2314 	 * amount of devices to be spawned.
2315 	 */
2316 	list = mlx5_malloc(MLX5_MEM_ZERO,
2317 			   sizeof(struct mlx5_dev_spawn_data) *
2318 			   (np ? np : nd),
2319 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2320 	if (!list) {
2321 		DRV_LOG(ERR, "spawn data array allocation failure");
2322 		rte_errno = ENOMEM;
2323 		ret = -rte_errno;
2324 		goto exit;
2325 	}
2326 	if (bd >= 0 || np > 1) {
2327 		/*
2328 		 * Single IB device with multiple ports found,
2329 		 * it may be E-Switch master device and representors.
2330 		 * We have to perform identification through the ports.
2331 		 */
2332 		MLX5_ASSERT(nl_rdma >= 0);
2333 		MLX5_ASSERT(ns == 0);
2334 		MLX5_ASSERT(nd == 1);
2335 		MLX5_ASSERT(np);
2336 		for (i = 1; i <= np; ++i) {
2337 			list[ns].bond_info = &bond_info;
2338 			list[ns].max_port = np;
2339 			list[ns].phys_port = i;
2340 			list[ns].phys_dev = ibv_match[0];
2341 			list[ns].eth_dev = NULL;
2342 			list[ns].pci_dev = pci_dev;
2343 			list[ns].pf_bond = bd;
2344 			list[ns].ifindex = mlx5_nl_ifindex
2345 				(nl_rdma,
2346 				mlx5_os_get_dev_device_name
2347 						(list[ns].phys_dev), i);
2348 			if (!list[ns].ifindex) {
2349 				/*
2350 				 * No network interface index found for the
2351 				 * specified port, it means there is no
2352 				 * representor on this port. It's OK,
2353 				 * there can be disabled ports, for example
2354 				 * if sriov_numvfs < sriov_totalvfs.
2355 				 */
2356 				continue;
2357 			}
2358 			ret = -1;
2359 			if (nl_route >= 0)
2360 				ret = mlx5_nl_switch_info
2361 					       (nl_route,
2362 						list[ns].ifindex,
2363 						&list[ns].info);
2364 			if (ret || (!list[ns].info.representor &&
2365 				    !list[ns].info.master)) {
2366 				/*
2367 				 * We failed to recognize representors with
2368 				 * Netlink, let's try to perform the task
2369 				 * with sysfs.
2370 				 */
2371 				ret =  mlx5_sysfs_switch_info
2372 						(list[ns].ifindex,
2373 						 &list[ns].info);
2374 			}
2375 			if (!ret && bd >= 0) {
2376 				switch (list[ns].info.name_type) {
2377 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2378 					if (np == 1) {
2379 						/*
2380 						 * Force standalone bonding
2381 						 * device for ROCE LAG
2382 						 * confgiurations.
2383 						 */
2384 						list[ns].info.master = 0;
2385 						list[ns].info.representor = 0;
2386 					}
2387 					if (list[ns].info.port_name == bd)
2388 						ns++;
2389 					break;
2390 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2391 					/* Fallthrough */
2392 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2393 					/* Fallthrough */
2394 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2395 					if (list[ns].info.pf_num == bd)
2396 						ns++;
2397 					break;
2398 				default:
2399 					break;
2400 				}
2401 				continue;
2402 			}
2403 			if (!ret && (list[ns].info.representor ^
2404 				     list[ns].info.master))
2405 				ns++;
2406 		}
2407 		if (!ns) {
2408 			DRV_LOG(ERR,
2409 				"unable to recognize master/representors"
2410 				" on the IB device with multiple ports");
2411 			rte_errno = ENOENT;
2412 			ret = -rte_errno;
2413 			goto exit;
2414 		}
2415 	} else {
2416 		/*
2417 		 * The existence of several matching entries (nd > 1) means
2418 		 * port representors have been instantiated. No existing Verbs
2419 		 * call nor sysfs entries can tell them apart, this can only
2420 		 * be done through Netlink calls assuming kernel drivers are
2421 		 * recent enough to support them.
2422 		 *
2423 		 * In the event of identification failure through Netlink,
2424 		 * try again through sysfs, then:
2425 		 *
2426 		 * 1. A single IB device matches (nd == 1) with single
2427 		 *    port (np=0/1) and is not a representor, assume
2428 		 *    no switch support.
2429 		 *
2430 		 * 2. Otherwise no safe assumptions can be made;
2431 		 *    complain louder and bail out.
2432 		 */
2433 		for (i = 0; i != nd; ++i) {
2434 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2435 			list[ns].bond_info = NULL;
2436 			list[ns].max_port = 1;
2437 			list[ns].phys_port = 1;
2438 			list[ns].phys_dev = ibv_match[i];
2439 			list[ns].eth_dev = NULL;
2440 			list[ns].pci_dev = pci_dev;
2441 			list[ns].pf_bond = -1;
2442 			list[ns].ifindex = 0;
2443 			if (nl_rdma >= 0)
2444 				list[ns].ifindex = mlx5_nl_ifindex
2445 				(nl_rdma,
2446 				mlx5_os_get_dev_device_name
2447 						(list[ns].phys_dev), 1);
2448 			if (!list[ns].ifindex) {
2449 				char ifname[IF_NAMESIZE];
2450 
2451 				/*
2452 				 * Netlink failed, it may happen with old
2453 				 * ib_core kernel driver (before 4.16).
2454 				 * We can assume there is old driver because
2455 				 * here we are processing single ports IB
2456 				 * devices. Let's try sysfs to retrieve
2457 				 * the ifindex. The method works for
2458 				 * master device only.
2459 				 */
2460 				if (nd > 1) {
2461 					/*
2462 					 * Multiple devices found, assume
2463 					 * representors, can not distinguish
2464 					 * master/representor and retrieve
2465 					 * ifindex via sysfs.
2466 					 */
2467 					continue;
2468 				}
2469 				ret = mlx5_get_ifname_sysfs
2470 					(ibv_match[i]->ibdev_path, ifname);
2471 				if (!ret)
2472 					list[ns].ifindex =
2473 						if_nametoindex(ifname);
2474 				if (!list[ns].ifindex) {
2475 					/*
2476 					 * No network interface index found
2477 					 * for the specified device, it means
2478 					 * there it is neither representor
2479 					 * nor master.
2480 					 */
2481 					continue;
2482 				}
2483 			}
2484 			ret = -1;
2485 			if (nl_route >= 0)
2486 				ret = mlx5_nl_switch_info
2487 					       (nl_route,
2488 						list[ns].ifindex,
2489 						&list[ns].info);
2490 			if (ret || (!list[ns].info.representor &&
2491 				    !list[ns].info.master)) {
2492 				/*
2493 				 * We failed to recognize representors with
2494 				 * Netlink, let's try to perform the task
2495 				 * with sysfs.
2496 				 */
2497 				ret =  mlx5_sysfs_switch_info
2498 						(list[ns].ifindex,
2499 						 &list[ns].info);
2500 			}
2501 			if (!ret && (list[ns].info.representor ^
2502 				     list[ns].info.master)) {
2503 				ns++;
2504 			} else if ((nd == 1) &&
2505 				   !list[ns].info.representor &&
2506 				   !list[ns].info.master) {
2507 				/*
2508 				 * Single IB device with
2509 				 * one physical port and
2510 				 * attached network device.
2511 				 * May be SRIOV is not enabled
2512 				 * or there is no representors.
2513 				 */
2514 				DRV_LOG(INFO, "no E-Switch support detected");
2515 				ns++;
2516 				break;
2517 			}
2518 		}
2519 		if (!ns) {
2520 			DRV_LOG(ERR,
2521 				"unable to recognize master/representors"
2522 				" on the multiple IB devices");
2523 			rte_errno = ENOENT;
2524 			ret = -rte_errno;
2525 			goto exit;
2526 		}
2527 		/*
2528 		 * New kernels may add the switch_id attribute for the case
2529 		 * there is no E-Switch and we wrongly recognized the
2530 		 * only device as master. Override this if there is the
2531 		 * single device with single port and new device name
2532 		 * format present.
2533 		 */
2534 		if (nd == 1 &&
2535 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2536 			list[0].info.master = 0;
2537 			list[0].info.representor = 0;
2538 		}
2539 	}
2540 	MLX5_ASSERT(ns);
2541 	/*
2542 	 * Sort list to probe devices in natural order for users convenience
2543 	 * (i.e. master first, then representors from lowest to highest ID).
2544 	 */
2545 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2546 	/* Device specific configuration. */
2547 	switch (pci_dev->id.device_id) {
2548 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2549 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2550 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2551 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2552 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2553 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2554 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2555 		dev_config_vf = 1;
2556 		break;
2557 	default:
2558 		dev_config_vf = 0;
2559 		break;
2560 	}
2561 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2562 		/* Set devargs default values. */
2563 		if (eth_da.nb_mh_controllers == 0) {
2564 			eth_da.nb_mh_controllers = 1;
2565 			eth_da.mh_controllers[0] = 0;
2566 		}
2567 		if (eth_da.nb_ports == 0 && ns > 0) {
2568 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2569 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2570 					pci_dev->device.devargs->args);
2571 			eth_da.nb_ports = 1;
2572 			eth_da.ports[0] = list[0].info.pf_num;
2573 		}
2574 		if (eth_da.nb_representor_ports == 0) {
2575 			eth_da.nb_representor_ports = 1;
2576 			eth_da.representor_ports[0] = 0;
2577 		}
2578 	}
2579 	for (i = 0; i != ns; ++i) {
2580 		uint32_t restore;
2581 
2582 		/* Default configuration. */
2583 		mlx5_os_config_default(&dev_config);
2584 		dev_config.vf = dev_config_vf;
2585 		list[i].numa_node = pci_dev->device.numa_node;
2586 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2587 						 &list[i],
2588 						 &dev_config,
2589 						 &eth_da);
2590 		if (!list[i].eth_dev) {
2591 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2592 				break;
2593 			/* Device is disabled or already spawned. Ignore it. */
2594 			continue;
2595 		}
2596 		restore = list[i].eth_dev->data->dev_flags;
2597 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2598 		/**
2599 		 * Each representor has a dedicated interrupts vector.
2600 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2601 		 * representor eth_dev object because representor and PF
2602 		 * share the same PCI address.
2603 		 * Override representor device with a dedicated
2604 		 * interrupts handle here.
2605 		 * Representor interrupts handle is released in mlx5_dev_stop().
2606 		 */
2607 		if (list[i].info.representor) {
2608 			struct rte_intr_handle *intr_handle;
2609 			intr_handle = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
2610 						  sizeof(*intr_handle), 0,
2611 						  SOCKET_ID_ANY);
2612 			if (!intr_handle) {
2613 				DRV_LOG(ERR,
2614 					"port %u failed to allocate memory for interrupt handler "
2615 					"Rx interrupts will not be supported",
2616 					i);
2617 				rte_errno = ENOMEM;
2618 				ret = -rte_errno;
2619 				goto exit;
2620 			}
2621 			list[i].eth_dev->intr_handle = intr_handle;
2622 		}
2623 		/* Restore non-PCI flags cleared by the above call. */
2624 		list[i].eth_dev->data->dev_flags |= restore;
2625 		rte_eth_dev_probing_finish(list[i].eth_dev);
2626 	}
2627 	if (i != ns) {
2628 		DRV_LOG(ERR,
2629 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2630 			" encountering an error: %s",
2631 			owner_pci.domain, owner_pci.bus,
2632 			owner_pci.devid, owner_pci.function,
2633 			strerror(rte_errno));
2634 		ret = -rte_errno;
2635 		/* Roll back. */
2636 		while (i--) {
2637 			if (!list[i].eth_dev)
2638 				continue;
2639 			mlx5_dev_close(list[i].eth_dev);
2640 			/* mac_addrs must not be freed because in dev_private */
2641 			list[i].eth_dev->data->mac_addrs = NULL;
2642 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2643 		}
2644 		/* Restore original error. */
2645 		rte_errno = -ret;
2646 	} else {
2647 		ret = 0;
2648 	}
2649 exit:
2650 	/*
2651 	 * Do the routine cleanup:
2652 	 * - close opened Netlink sockets
2653 	 * - free allocated spawn data array
2654 	 * - free the Infiniband device list
2655 	 */
2656 	if (nl_rdma >= 0)
2657 		close(nl_rdma);
2658 	if (nl_route >= 0)
2659 		close(nl_route);
2660 	if (list)
2661 		mlx5_free(list);
2662 	MLX5_ASSERT(ibv_list);
2663 	mlx5_glue->free_device_list(ibv_list);
2664 	return ret;
2665 }
2666 
2667 static int
2668 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2669 			  struct rte_eth_devargs *eth_da)
2670 {
2671 	int ret = 0;
2672 
2673 	if (dev->devargs == NULL)
2674 		return 0;
2675 	memset(eth_da, 0, sizeof(*eth_da));
2676 	/* Parse representor information first from class argument. */
2677 	if (dev->devargs->cls_str)
2678 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2679 	if (ret != 0) {
2680 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2681 			dev->devargs->cls_str);
2682 		return -rte_errno;
2683 	}
2684 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2685 		/* Parse legacy device argument */
2686 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2687 		if (ret) {
2688 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2689 				dev->devargs->args);
2690 			return -rte_errno;
2691 		}
2692 	}
2693 	return 0;
2694 }
2695 
2696 /**
2697  * Callback to register a PCI device.
2698  *
2699  * This function spawns Ethernet devices out of a given PCI device.
2700  *
2701  * @param[in] pci_dev
2702  *   PCI device information.
2703  *
2704  * @return
2705  *   0 on success, a negative errno value otherwise and rte_errno is set.
2706  */
2707 static int
2708 mlx5_os_pci_probe(struct rte_pci_device *pci_dev)
2709 {
2710 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2711 	int ret = 0;
2712 	uint16_t p;
2713 
2714 	ret = mlx5_os_parse_eth_devargs(&pci_dev->device, &eth_da);
2715 	if (ret != 0)
2716 		return ret;
2717 
2718 	if (eth_da.nb_ports > 0) {
2719 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2720 		for (p = 0; p < eth_da.nb_ports; p++) {
2721 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2722 						   eth_da.ports[p]);
2723 			if (ret)
2724 				break;
2725 		}
2726 		if (ret) {
2727 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2728 				"aborted due to proding failure of PF %u",
2729 				pci_dev->addr.domain, pci_dev->addr.bus,
2730 				pci_dev->addr.devid, pci_dev->addr.function,
2731 				eth_da.ports[p]);
2732 			mlx5_net_remove(&pci_dev->device);
2733 		}
2734 	} else {
2735 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2736 	}
2737 	return ret;
2738 }
2739 
2740 /* Probe a single SF device on auxiliary bus, no representor support. */
2741 static int
2742 mlx5_os_auxiliary_probe(struct rte_device *dev)
2743 {
2744 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2745 	struct mlx5_dev_config config;
2746 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2747 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2748 	struct rte_eth_dev *eth_dev;
2749 	int ret = 0;
2750 
2751 	/* Parse ethdev devargs. */
2752 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2753 	if (ret != 0)
2754 		return ret;
2755 	/* Set default config data. */
2756 	mlx5_os_config_default(&config);
2757 	config.sf = 1;
2758 	/* Init spawn data. */
2759 	spawn.max_port = 1;
2760 	spawn.phys_port = 1;
2761 	spawn.phys_dev = mlx5_os_get_ibv_dev(dev);
2762 	if (spawn.phys_dev == NULL)
2763 		return -rte_errno;
2764 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2765 	if (ret < 0) {
2766 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2767 		return ret;
2768 	}
2769 	spawn.ifindex = ret;
2770 	spawn.numa_node = dev->numa_node;
2771 	/* Spawn device. */
2772 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2773 	if (eth_dev == NULL)
2774 		return -rte_errno;
2775 	/* Post create. */
2776 	eth_dev->intr_handle = &adev->intr_handle;
2777 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2778 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2779 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2780 		eth_dev->data->numa_node = dev->numa_node;
2781 	}
2782 	rte_eth_dev_probing_finish(eth_dev);
2783 	return 0;
2784 }
2785 
2786 /**
2787  * Net class driver callback to probe a device.
2788  *
2789  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2790  *
2791  * @param[in] dev
2792  *   Pointer to the generic device.
2793  *
2794  * @return
2795  *   0 on success, the function cannot fail.
2796  */
2797 int
2798 mlx5_os_net_probe(struct rte_device *dev)
2799 {
2800 	int ret;
2801 
2802 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2803 		mlx5_pmd_socket_init();
2804 	ret = mlx5_init_once();
2805 	if (ret) {
2806 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2807 			strerror(rte_errno));
2808 		return -rte_errno;
2809 	}
2810 	if (mlx5_dev_is_pci(dev))
2811 		return mlx5_os_pci_probe(RTE_DEV_TO_PCI(dev));
2812 	else
2813 		return mlx5_os_auxiliary_probe(dev);
2814 }
2815 
2816 static int
2817 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2818 {
2819 	char *env;
2820 	int value;
2821 
2822 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2823 	/* Get environment variable to store. */
2824 	env = getenv(MLX5_SHUT_UP_BF);
2825 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2826 	if (config->dbnc == MLX5_ARG_UNSET)
2827 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2828 	else
2829 		setenv(MLX5_SHUT_UP_BF,
2830 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2831 	return value;
2832 }
2833 
2834 static void
2835 mlx5_restore_doorbell_mapping_env(int value)
2836 {
2837 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2838 	/* Restore the original environment variable state. */
2839 	if (value == MLX5_ARG_UNSET)
2840 		unsetenv(MLX5_SHUT_UP_BF);
2841 	else
2842 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2843 }
2844 
2845 /**
2846  * Extract pdn of PD object using DV API.
2847  *
2848  * @param[in] pd
2849  *   Pointer to the verbs PD object.
2850  * @param[out] pdn
2851  *   Pointer to the PD object number variable.
2852  *
2853  * @return
2854  *   0 on success, error value otherwise.
2855  */
2856 int
2857 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2858 {
2859 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2860 	struct mlx5dv_obj obj;
2861 	struct mlx5dv_pd pd_info;
2862 	int ret = 0;
2863 
2864 	obj.pd.in = pd;
2865 	obj.pd.out = &pd_info;
2866 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2867 	if (ret) {
2868 		DRV_LOG(DEBUG, "Fail to get PD object info");
2869 		return ret;
2870 	}
2871 	*pdn = pd_info.pdn;
2872 	return 0;
2873 #else
2874 	(void)pd;
2875 	(void)pdn;
2876 	return -ENOTSUP;
2877 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2878 }
2879 
2880 /**
2881  * Function API to open IB device.
2882  *
2883  * This function calls the Linux glue APIs to open a device.
2884  *
2885  * @param[in] spawn
2886  *   Pointer to the IB device attributes (name, port, etc).
2887  * @param[out] config
2888  *   Pointer to device configuration structure.
2889  * @param[out] sh
2890  *   Pointer to shared context structure.
2891  *
2892  * @return
2893  *   0 on success, a positive error value otherwise.
2894  */
2895 int
2896 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2897 		     const struct mlx5_dev_config *config,
2898 		     struct mlx5_dev_ctx_shared *sh)
2899 {
2900 	int dbmap_env;
2901 	int err = 0;
2902 
2903 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2904 	/*
2905 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2906 	 * before the device creation. The rdma_core library
2907 	 * checks the variable at device creation and
2908 	 * stores the result internally.
2909 	 */
2910 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2911 	/* Try to open IB device with DV first, then usual Verbs. */
2912 	errno = 0;
2913 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2914 	if (sh->ctx) {
2915 		sh->devx = 1;
2916 		DRV_LOG(DEBUG, "DevX is supported");
2917 		/* The device is created, no need for environment. */
2918 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2919 	} else {
2920 		/* The environment variable is still configured. */
2921 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2922 		err = errno ? errno : ENODEV;
2923 		/*
2924 		 * The environment variable is not needed anymore,
2925 		 * all device creation attempts are completed.
2926 		 */
2927 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2928 		if (!sh->ctx)
2929 			return err;
2930 		DRV_LOG(DEBUG, "DevX is NOT supported");
2931 		err = 0;
2932 	}
2933 	if (!err && sh->ctx) {
2934 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2935 		mlx5_glue->dv_set_context_attr(sh->ctx,
2936 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2937 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2938 				.alloc = &mlx5_alloc_verbs_buf,
2939 				.free = &mlx5_free_verbs_buf,
2940 				.data = sh,
2941 			}));
2942 	}
2943 	return err;
2944 }
2945 
2946 /**
2947  * Install shared asynchronous device events handler.
2948  * This function is implemented to support event sharing
2949  * between multiple ports of single IB device.
2950  *
2951  * @param sh
2952  *   Pointer to mlx5_dev_ctx_shared object.
2953  */
2954 void
2955 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2956 {
2957 	int ret;
2958 	int flags;
2959 
2960 	sh->intr_handle.fd = -1;
2961 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2962 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2963 		    F_SETFL, flags | O_NONBLOCK);
2964 	if (ret) {
2965 		DRV_LOG(INFO, "failed to change file descriptor async event"
2966 			" queue");
2967 	} else {
2968 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2969 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2970 		if (rte_intr_callback_register(&sh->intr_handle,
2971 					mlx5_dev_interrupt_handler, sh)) {
2972 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2973 			sh->intr_handle.fd = -1;
2974 		}
2975 	}
2976 	if (sh->devx) {
2977 #ifdef HAVE_IBV_DEVX_ASYNC
2978 		sh->intr_handle_devx.fd = -1;
2979 		sh->devx_comp =
2980 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2981 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2982 		if (!devx_comp) {
2983 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2984 			return;
2985 		}
2986 		flags = fcntl(devx_comp->fd, F_GETFL);
2987 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2988 		if (ret) {
2989 			DRV_LOG(INFO, "failed to change file descriptor"
2990 				" devx comp");
2991 			return;
2992 		}
2993 		sh->intr_handle_devx.fd = devx_comp->fd;
2994 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2995 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2996 					mlx5_dev_interrupt_handler_devx, sh)) {
2997 			DRV_LOG(INFO, "Fail to install the devx shared"
2998 				" interrupt.");
2999 			sh->intr_handle_devx.fd = -1;
3000 		}
3001 #endif /* HAVE_IBV_DEVX_ASYNC */
3002 	}
3003 }
3004 
3005 /**
3006  * Uninstall shared asynchronous device events handler.
3007  * This function is implemented to support event sharing
3008  * between multiple ports of single IB device.
3009  *
3010  * @param dev
3011  *   Pointer to mlx5_dev_ctx_shared object.
3012  */
3013 void
3014 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
3015 {
3016 	if (sh->intr_handle.fd >= 0)
3017 		mlx5_intr_callback_unregister(&sh->intr_handle,
3018 					      mlx5_dev_interrupt_handler, sh);
3019 #ifdef HAVE_IBV_DEVX_ASYNC
3020 	if (sh->intr_handle_devx.fd >= 0)
3021 		rte_intr_callback_unregister(&sh->intr_handle_devx,
3022 				  mlx5_dev_interrupt_handler_devx, sh);
3023 	if (sh->devx_comp)
3024 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
3025 #endif
3026 }
3027 
3028 /**
3029  * Read statistics by a named counter.
3030  *
3031  * @param[in] priv
3032  *   Pointer to the private device data structure.
3033  * @param[in] ctr_name
3034  *   Pointer to the name of the statistic counter to read
3035  * @param[out] stat
3036  *   Pointer to read statistic value.
3037  * @return
3038  *   0 on success and stat is valud, 1 if failed to read the value
3039  *   rte_errno is set.
3040  *
3041  */
3042 int
3043 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
3044 		      uint64_t *stat)
3045 {
3046 	int fd;
3047 
3048 	if (priv->sh) {
3049 		if (priv->q_counters != NULL &&
3050 		    strcmp(ctr_name, "out_of_buffer") == 0)
3051 			return mlx5_devx_cmd_queue_counter_query
3052 					(priv->q_counters, 0, (uint32_t *)stat);
3053 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
3054 		      priv->sh->ibdev_path,
3055 		      priv->dev_port,
3056 		      ctr_name);
3057 		fd = open(path, O_RDONLY);
3058 		/*
3059 		 * in switchdev the file location is not per port
3060 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
3061 		 */
3062 		if (fd == -1) {
3063 			MKSTR(path1, "%s/hw_counters/%s",
3064 			      priv->sh->ibdev_path,
3065 			      ctr_name);
3066 			fd = open(path1, O_RDONLY);
3067 		}
3068 		if (fd != -1) {
3069 			char buf[21] = {'\0'};
3070 			ssize_t n = read(fd, buf, sizeof(buf));
3071 
3072 			close(fd);
3073 			if (n != -1) {
3074 				*stat = strtoull(buf, NULL, 10);
3075 				return 0;
3076 			}
3077 		}
3078 	}
3079 	*stat = 0;
3080 	return 1;
3081 }
3082 
3083 /**
3084  * Set the reg_mr and dereg_mr call backs
3085  *
3086  * @param reg_mr_cb[out]
3087  *   Pointer to reg_mr func
3088  * @param dereg_mr_cb[out]
3089  *   Pointer to dereg_mr func
3090  *
3091  */
3092 void
3093 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
3094 		      mlx5_dereg_mr_t *dereg_mr_cb)
3095 {
3096 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
3097 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
3098 }
3099 
3100 /**
3101  * Remove a MAC address from device
3102  *
3103  * @param dev
3104  *   Pointer to Ethernet device structure.
3105  * @param index
3106  *   MAC address index.
3107  */
3108 void
3109 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3110 {
3111 	struct mlx5_priv *priv = dev->data->dev_private;
3112 	const int vf = priv->config.vf;
3113 
3114 	if (vf)
3115 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
3116 					mlx5_ifindex(dev), priv->mac_own,
3117 					&dev->data->mac_addrs[index], index);
3118 }
3119 
3120 /**
3121  * Adds a MAC address to the device
3122  *
3123  * @param dev
3124  *   Pointer to Ethernet device structure.
3125  * @param mac_addr
3126  *   MAC address to register.
3127  * @param index
3128  *   MAC address index.
3129  *
3130  * @return
3131  *   0 on success, a negative errno value otherwise
3132  */
3133 int
3134 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
3135 		     uint32_t index)
3136 {
3137 	struct mlx5_priv *priv = dev->data->dev_private;
3138 	const int vf = priv->config.vf;
3139 	int ret = 0;
3140 
3141 	if (vf)
3142 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
3143 					   mlx5_ifindex(dev), priv->mac_own,
3144 					   mac, index);
3145 	return ret;
3146 }
3147 
3148 /**
3149  * Modify a VF MAC address
3150  *
3151  * @param priv
3152  *   Pointer to device private data.
3153  * @param mac_addr
3154  *   MAC address to modify into.
3155  * @param iface_idx
3156  *   Net device interface index
3157  * @param vf_index
3158  *   VF index
3159  *
3160  * @return
3161  *   0 on success, a negative errno value otherwise
3162  */
3163 int
3164 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
3165 			   unsigned int iface_idx,
3166 			   struct rte_ether_addr *mac_addr,
3167 			   int vf_index)
3168 {
3169 	return mlx5_nl_vf_mac_addr_modify
3170 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
3171 }
3172 
3173 /**
3174  * Set device promiscuous mode
3175  *
3176  * @param dev
3177  *   Pointer to Ethernet device structure.
3178  * @param enable
3179  *   0 - promiscuous is disabled, otherwise - enabled
3180  *
3181  * @return
3182  *   0 on success, a negative error value otherwise
3183  */
3184 int
3185 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
3186 {
3187 	struct mlx5_priv *priv = dev->data->dev_private;
3188 
3189 	return mlx5_nl_promisc(priv->nl_socket_route,
3190 			       mlx5_ifindex(dev), !!enable);
3191 }
3192 
3193 /**
3194  * Set device promiscuous mode
3195  *
3196  * @param dev
3197  *   Pointer to Ethernet device structure.
3198  * @param enable
3199  *   0 - all multicase is disabled, otherwise - enabled
3200  *
3201  * @return
3202  *   0 on success, a negative error value otherwise
3203  */
3204 int
3205 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
3206 {
3207 	struct mlx5_priv *priv = dev->data->dev_private;
3208 
3209 	return mlx5_nl_allmulti(priv->nl_socket_route,
3210 				mlx5_ifindex(dev), !!enable);
3211 }
3212 
3213 /**
3214  * Flush device MAC addresses
3215  *
3216  * @param dev
3217  *   Pointer to Ethernet device structure.
3218  *
3219  */
3220 void
3221 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3222 {
3223 	struct mlx5_priv *priv = dev->data->dev_private;
3224 
3225 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3226 			       dev->data->mac_addrs,
3227 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3228 }
3229