xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision bb85a78d)
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 tunnel_en = 0;
967 	unsigned int mpls_en = 0;
968 	unsigned int swp = 0;
969 	unsigned int mprq = 0;
970 	unsigned int mprq_min_stride_size_n = 0;
971 	unsigned int mprq_max_stride_size_n = 0;
972 	unsigned int mprq_min_stride_num_n = 0;
973 	unsigned int mprq_max_stride_num_n = 0;
974 	struct rte_ether_addr mac;
975 	char name[RTE_ETH_NAME_MAX_LEN];
976 	int own_domain_id = 0;
977 	uint16_t port_id;
978 	struct mlx5_port_info vport_info = { .query_flags = 0 };
979 	int i;
980 
981 	/* Determine if this port representor is supposed to be spawned. */
982 	if (switch_info->representor && dpdk_dev->devargs &&
983 	    !mlx5_representor_match(spawn, eth_da))
984 		return NULL;
985 	/* Build device name. */
986 	if (spawn->pf_bond < 0) {
987 		/* Single device. */
988 		if (!switch_info->representor)
989 			strlcpy(name, dpdk_dev->name, sizeof(name));
990 		else
991 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
992 				 dpdk_dev->name,
993 				 switch_info->name_type ==
994 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
995 				 switch_info->port_name);
996 	} else {
997 		/* Bonding device. */
998 		if (!switch_info->representor) {
999 			err = snprintf(name, sizeof(name), "%s_%s",
1000 				 dpdk_dev->name,
1001 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
1002 		} else {
1003 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
1004 				dpdk_dev->name,
1005 				mlx5_os_get_dev_device_name(spawn->phys_dev),
1006 				switch_info->ctrl_num,
1007 				switch_info->pf_num,
1008 				switch_info->name_type ==
1009 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
1010 				switch_info->port_name);
1011 		}
1012 	}
1013 	if (err >= (int)sizeof(name))
1014 		DRV_LOG(WARNING, "device name overflow %s", name);
1015 	/* check if the device is already spawned */
1016 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1017 		rte_errno = EEXIST;
1018 		return NULL;
1019 	}
1020 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1021 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1022 		struct mlx5_mp_id mp_id;
1023 
1024 		eth_dev = rte_eth_dev_attach_secondary(name);
1025 		if (eth_dev == NULL) {
1026 			DRV_LOG(ERR, "can not attach rte ethdev");
1027 			rte_errno = ENOMEM;
1028 			return NULL;
1029 		}
1030 		eth_dev->device = dpdk_dev;
1031 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1032 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1033 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1034 		err = mlx5_proc_priv_init(eth_dev);
1035 		if (err)
1036 			return NULL;
1037 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
1038 		/* Receive command fd from primary process */
1039 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1040 		if (err < 0)
1041 			goto err_secondary;
1042 		/* Remap UAR for Tx queues. */
1043 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1044 		if (err)
1045 			goto err_secondary;
1046 		/*
1047 		 * Ethdev pointer is still required as input since
1048 		 * the primary device is not accessible from the
1049 		 * secondary process.
1050 		 */
1051 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1052 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1053 		return eth_dev;
1054 err_secondary:
1055 		mlx5_dev_close(eth_dev);
1056 		return NULL;
1057 	}
1058 	/*
1059 	 * Some parameters ("tx_db_nc" in particularly) are needed in
1060 	 * advance to create dv/verbs device context. We proceed the
1061 	 * devargs here to get ones, and later proceed devargs again
1062 	 * to override some hardware settings.
1063 	 */
1064 	err = mlx5_args(config, dpdk_dev->devargs);
1065 	if (err) {
1066 		err = rte_errno;
1067 		DRV_LOG(ERR, "failed to process device arguments: %s",
1068 			strerror(rte_errno));
1069 		goto error;
1070 	}
1071 	if (config->dv_miss_info) {
1072 		if (switch_info->master || switch_info->representor)
1073 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
1074 	}
1075 	mlx5_malloc_mem_select(config->sys_mem_en);
1076 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
1077 	if (!sh)
1078 		return NULL;
1079 	config->devx = sh->devx;
1080 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1081 	config->dest_tir = 1;
1082 #endif
1083 #ifdef HAVE_IBV_MLX5_MOD_SWP
1084 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1085 #endif
1086 	/*
1087 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1088 	 * as all ConnectX-5 devices.
1089 	 */
1090 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1091 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1092 #endif
1093 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1094 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1095 #endif
1096 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1097 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1098 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1099 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1100 			mps = MLX5_MPW_ENHANCED;
1101 		} else {
1102 			DRV_LOG(DEBUG, "MPW is supported");
1103 			mps = MLX5_MPW;
1104 		}
1105 	} else {
1106 		DRV_LOG(DEBUG, "MPW isn't supported");
1107 		mps = MLX5_MPW_DISABLED;
1108 	}
1109 #ifdef HAVE_IBV_MLX5_MOD_SWP
1110 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1111 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1112 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1113 #endif
1114 	config->swp = !!swp;
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 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1147 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1148 			     (dv_attr.tunnel_offloads_caps &
1149 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
1150 			     (dv_attr.tunnel_offloads_caps &
1151 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1152 	}
1153 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1154 		tunnel_en ? "" : "not ");
1155 #else
1156 	DRV_LOG(WARNING,
1157 		"tunnel offloading disabled due to old OFED/rdma-core version");
1158 #endif
1159 	config->tunnel_en = tunnel_en;
1160 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1161 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1162 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1163 		   (dv_attr.tunnel_offloads_caps &
1164 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1165 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1166 		mpls_en ? "" : "not ");
1167 #else
1168 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1169 		" old OFED/rdma-core version or firmware configuration");
1170 #endif
1171 	config->mpls_en = mpls_en;
1172 	/* Check port status. */
1173 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1174 	if (err) {
1175 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1176 		goto error;
1177 	}
1178 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1179 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1180 		err = EINVAL;
1181 		goto error;
1182 	}
1183 	if (port_attr.state != IBV_PORT_ACTIVE)
1184 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1185 			mlx5_glue->port_state_str(port_attr.state),
1186 			port_attr.state);
1187 	/* Allocate private eth device data. */
1188 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1189 			   sizeof(*priv),
1190 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1191 	if (priv == NULL) {
1192 		DRV_LOG(ERR, "priv allocation failure");
1193 		err = ENOMEM;
1194 		goto error;
1195 	}
1196 	priv->sh = sh;
1197 	priv->dev_port = spawn->phys_port;
1198 	priv->pci_dev = spawn->pci_dev;
1199 	priv->mtu = RTE_ETHER_MTU;
1200 	/* Some internal functions rely on Netlink sockets, open them now. */
1201 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1202 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1203 	priv->representor = !!switch_info->representor;
1204 	priv->master = !!switch_info->master;
1205 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1206 	priv->vport_meta_tag = 0;
1207 	priv->vport_meta_mask = 0;
1208 	priv->pf_bond = spawn->pf_bond;
1209 
1210 	DRV_LOG(DEBUG,
1211 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1212 		priv->dev_port, dpdk_dev->bus->name,
1213 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1214 		priv->master, priv->representor, priv->pf_bond);
1215 
1216 	/*
1217 	 * If we have E-Switch we should determine the vport attributes.
1218 	 * E-Switch may use either source vport field or reg_c[0] metadata
1219 	 * register to match on vport index. The engaged part of metadata
1220 	 * register is defined by mask.
1221 	 */
1222 	if (switch_info->representor || switch_info->master) {
1223 		err = mlx5_glue->devx_port_query(sh->ctx,
1224 						 spawn->phys_port,
1225 						 &vport_info);
1226 		if (err) {
1227 			DRV_LOG(WARNING,
1228 				"can't query devx port %d on device %s",
1229 				spawn->phys_port,
1230 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1231 			vport_info.query_flags = 0;
1232 		}
1233 	}
1234 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1235 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1236 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1237 		if (!priv->vport_meta_mask) {
1238 			DRV_LOG(ERR, "vport zero mask for port %d"
1239 				     " on bonding device %s",
1240 				     spawn->phys_port,
1241 				     mlx5_os_get_dev_device_name
1242 							(spawn->phys_dev));
1243 			err = ENOTSUP;
1244 			goto error;
1245 		}
1246 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1247 			DRV_LOG(ERR, "invalid vport tag for port %d"
1248 				     " on bonding device %s",
1249 				     spawn->phys_port,
1250 				     mlx5_os_get_dev_device_name
1251 							(spawn->phys_dev));
1252 			err = ENOTSUP;
1253 			goto error;
1254 		}
1255 	}
1256 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1257 		priv->vport_id = vport_info.vport_id;
1258 	} else if (spawn->pf_bond >= 0 &&
1259 		   (switch_info->representor || switch_info->master)) {
1260 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1261 			     " on bonding device %s",
1262 			     spawn->phys_port,
1263 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1264 		err = ENOTSUP;
1265 		goto error;
1266 	} else {
1267 		/*
1268 		 * Suppose vport index in compatible way. Kernel/rdma_core
1269 		 * support single E-Switch per PF configurations only and
1270 		 * vport_id field contains the vport index for associated VF,
1271 		 * which is deduced from representor port name.
1272 		 * For example, let's have the IB device port 10, it has
1273 		 * attached network device eth0, which has port name attribute
1274 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1275 		 * as 3 (2+1). This assigning schema should be changed if the
1276 		 * multiple E-Switch instances per PF configurations or/and PCI
1277 		 * subfunctions are added.
1278 		 */
1279 		priv->vport_id = switch_info->representor ?
1280 				 switch_info->port_name + 1 : -1;
1281 	}
1282 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1283 							  eth_da->type);
1284 	/*
1285 	 * Look for sibling devices in order to reuse their switch domain
1286 	 * if any, otherwise allocate one.
1287 	 */
1288 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1289 		const struct mlx5_priv *opriv =
1290 			rte_eth_devices[port_id].data->dev_private;
1291 
1292 		if (!opriv ||
1293 		    opriv->sh != priv->sh ||
1294 			opriv->domain_id ==
1295 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1296 			continue;
1297 		priv->domain_id = opriv->domain_id;
1298 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1299 			priv->dev_port, priv->domain_id);
1300 		break;
1301 	}
1302 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1303 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1304 		if (err) {
1305 			err = rte_errno;
1306 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1307 				strerror(rte_errno));
1308 			goto error;
1309 		}
1310 		own_domain_id = 1;
1311 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1312 			priv->dev_port, priv->domain_id);
1313 	}
1314 	/* Override some values set by hardware configuration. */
1315 	mlx5_args(config, dpdk_dev->devargs);
1316 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
1317 	if (err)
1318 		goto error;
1319 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1320 			    IBV_DEVICE_RAW_IP_CSUM);
1321 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1322 		(config->hw_csum ? "" : "not "));
1323 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1324 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1325 	DRV_LOG(DEBUG, "counters are not supported");
1326 #endif
1327 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1328 	if (config->dv_flow_en) {
1329 		DRV_LOG(WARNING, "DV flow is not supported");
1330 		config->dv_flow_en = 0;
1331 	}
1332 #endif
1333 	if (spawn->max_port > UINT8_MAX) {
1334 		/* Verbs can't support ports larger than 255 by design. */
1335 		DRV_LOG(ERR, "can't support IB ports > UINT8_MAX");
1336 		err = EINVAL;
1337 		goto error;
1338 	}
1339 	config->ind_table_max_size =
1340 		sh->device_attr.max_rwq_indirection_table_size;
1341 	/*
1342 	 * Remove this check once DPDK supports larger/variable
1343 	 * indirection tables.
1344 	 */
1345 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1346 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1347 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1348 		config->ind_table_max_size);
1349 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1350 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1351 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1352 		(config->hw_vlan_strip ? "" : "not "));
1353 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1354 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1355 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1356 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1357 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1358 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1359 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1360 #endif
1361 	if (config->hw_padding && !hw_padding) {
1362 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1363 		config->hw_padding = 0;
1364 	} else if (config->hw_padding) {
1365 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1366 	}
1367 	config->tso = (sh->device_attr.max_tso > 0 &&
1368 		      (sh->device_attr.tso_supported_qpts &
1369 		       (1 << IBV_QPT_RAW_PACKET)));
1370 	if (config->tso)
1371 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1372 	/*
1373 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1374 	 * by default.
1375 	 */
1376 	if (config->mps == MLX5_ARG_UNSET)
1377 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1378 							  MLX5_MPW_DISABLED;
1379 	else
1380 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1381 	DRV_LOG(INFO, "%sMPS is %s",
1382 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1383 		config->mps == MLX5_MPW ? "legacy " : "",
1384 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1385 	if (config->devx) {
1386 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1387 		if (err) {
1388 			err = -err;
1389 			goto error;
1390 		}
1391 		/* Check relax ordering support. */
1392 		if (!haswell_broadwell_cpu) {
1393 			sh->cmng.relaxed_ordering_write =
1394 				config->hca_attr.relaxed_ordering_write;
1395 			sh->cmng.relaxed_ordering_read =
1396 				config->hca_attr.relaxed_ordering_read;
1397 		} else {
1398 			sh->cmng.relaxed_ordering_read = 0;
1399 			sh->cmng.relaxed_ordering_write = 0;
1400 		}
1401 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1402 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1403 		sh->steering_format_version =
1404 			config->hca_attr.steering_format_version;
1405 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1406 		/* Check for LRO support. */
1407 		if (config->dest_tir && config->hca_attr.lro_cap &&
1408 		    config->dv_flow_en) {
1409 			/* TBD check tunnel lro caps. */
1410 			config->lro.supported = config->hca_attr.lro_cap;
1411 			DRV_LOG(DEBUG, "Device supports LRO");
1412 			/*
1413 			 * If LRO timeout is not configured by application,
1414 			 * use the minimal supported value.
1415 			 */
1416 			if (!config->lro.timeout)
1417 				config->lro.timeout =
1418 				config->hca_attr.lro_timer_supported_periods[0];
1419 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1420 				config->lro.timeout);
1421 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1422 				"required for coalescing is %d bytes",
1423 				config->hca_attr.lro_min_mss_size);
1424 		}
1425 #if defined(HAVE_MLX5DV_DR) && \
1426 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1427 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1428 		if (config->hca_attr.qos.sup &&
1429 		    config->hca_attr.qos.flow_meter_old &&
1430 		    config->dv_flow_en) {
1431 			uint8_t reg_c_mask =
1432 				config->hca_attr.qos.flow_meter_reg_c_ids;
1433 			/*
1434 			 * Meter needs two REG_C's for color match and pre-sfx
1435 			 * flow match. Here get the REG_C for color match.
1436 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1437 			 */
1438 			reg_c_mask &= 0xfc;
1439 			if (__builtin_popcount(reg_c_mask) < 1) {
1440 				priv->mtr_en = 0;
1441 				DRV_LOG(WARNING, "No available register for"
1442 					" meter.");
1443 			} else {
1444 				/*
1445 				 * The meter color register is used by the
1446 				 * flow-hit feature as well.
1447 				 * The flow-hit feature must use REG_C_3
1448 				 * Prefer REG_C_3 if it is available.
1449 				 */
1450 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1451 					priv->mtr_color_reg = REG_C_3;
1452 				else
1453 					priv->mtr_color_reg = ffs(reg_c_mask)
1454 							      - 1 + REG_C_0;
1455 				priv->mtr_en = 1;
1456 				priv->mtr_reg_share =
1457 				      config->hca_attr.qos.flow_meter;
1458 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1459 					priv->mtr_color_reg);
1460 			}
1461 		}
1462 		if (config->hca_attr.qos.sup &&
1463 			config->hca_attr.qos.flow_meter_aso_sup) {
1464 			uint32_t log_obj_size =
1465 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1466 			if (log_obj_size >=
1467 			config->hca_attr.qos.log_meter_aso_granularity &&
1468 			log_obj_size <=
1469 			config->hca_attr.qos.log_meter_aso_max_alloc)
1470 				sh->meter_aso_en = 1;
1471 		}
1472 		if (priv->mtr_en) {
1473 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1474 			if (err) {
1475 				err = -err;
1476 				goto error;
1477 			}
1478 		}
1479 		if (config->hca_attr.flow.tunnel_header_0_1)
1480 			sh->tunnel_header_0_1 = 1;
1481 #endif
1482 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1483 		if (config->hca_attr.flow_hit_aso &&
1484 		    priv->mtr_color_reg == REG_C_3) {
1485 			sh->flow_hit_aso_en = 1;
1486 			err = mlx5_flow_aso_age_mng_init(sh);
1487 			if (err) {
1488 				err = -err;
1489 				goto error;
1490 			}
1491 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1492 		}
1493 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1494 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1495 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1496 		if (config->hca_attr.ct_offload &&
1497 		    priv->mtr_color_reg == REG_C_3) {
1498 			err = mlx5_flow_aso_ct_mng_init(sh);
1499 			if (err) {
1500 				err = -err;
1501 				goto error;
1502 			}
1503 			DRV_LOG(DEBUG, "CT ASO is supported.");
1504 			sh->ct_aso_en = 1;
1505 		}
1506 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1507 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1508 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1509 		    config->dv_flow_en) {
1510 			priv->sampler_en = 1;
1511 			DRV_LOG(DEBUG, "Sampler enabled!");
1512 		} else {
1513 			priv->sampler_en = 0;
1514 			if (!config->hca_attr.log_max_ft_sampler_num)
1515 				DRV_LOG(WARNING,
1516 					"No available register for sampler.");
1517 			else
1518 				DRV_LOG(DEBUG, "DV flow is not supported!");
1519 		}
1520 #endif
1521 	}
1522 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1523 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1524 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1525 		config->cqe_comp = 0;
1526 	}
1527 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1528 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1529 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1530 				 " format isn't supported.");
1531 		config->cqe_comp = 0;
1532 	}
1533 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1534 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1535 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1536 				 " format isn't supported.");
1537 		config->cqe_comp = 0;
1538 	}
1539 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1540 			config->cqe_comp ? "" : "not ");
1541 	if (config->tx_pp) {
1542 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1543 			config->hca_attr.dev_freq_khz);
1544 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1545 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1546 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1547 			config->hca_attr.cross_channel ? "" : "not ");
1548 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1549 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1550 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1551 			config->hca_attr.non_wire_sq ? "" : "not ");
1552 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1553 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1554 			config->hca_attr.log_max_static_sq_wq);
1555 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1556 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1557 		if (!config->devx) {
1558 			DRV_LOG(ERR, "DevX is required for packet pacing");
1559 			err = ENODEV;
1560 			goto error;
1561 		}
1562 		if (!config->hca_attr.qos.packet_pacing) {
1563 			DRV_LOG(ERR, "Packet pacing is not supported");
1564 			err = ENODEV;
1565 			goto error;
1566 		}
1567 		if (!config->hca_attr.cross_channel) {
1568 			DRV_LOG(ERR, "Cross channel operations are"
1569 				     " required for packet pacing");
1570 			err = ENODEV;
1571 			goto error;
1572 		}
1573 		if (!config->hca_attr.wqe_index_ignore) {
1574 			DRV_LOG(ERR, "WQE index ignore feature is"
1575 				     " required for packet pacing");
1576 			err = ENODEV;
1577 			goto error;
1578 		}
1579 		if (!config->hca_attr.non_wire_sq) {
1580 			DRV_LOG(ERR, "Non-wire SQ feature is"
1581 				     " required for packet pacing");
1582 			err = ENODEV;
1583 			goto error;
1584 		}
1585 		if (!config->hca_attr.log_max_static_sq_wq) {
1586 			DRV_LOG(ERR, "Static WQE SQ feature is"
1587 				     " required for packet pacing");
1588 			err = ENODEV;
1589 			goto error;
1590 		}
1591 		if (!config->hca_attr.qos.wqe_rate_pp) {
1592 			DRV_LOG(ERR, "WQE rate mode is required"
1593 				     " for packet pacing");
1594 			err = ENODEV;
1595 			goto error;
1596 		}
1597 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1598 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1599 			     " can't create queues for packet pacing");
1600 		err = ENODEV;
1601 		goto error;
1602 #endif
1603 	}
1604 	if (config->devx) {
1605 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1606 
1607 		err = config->hca_attr.access_register_user ?
1608 			mlx5_devx_cmd_register_read
1609 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1610 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1611 		if (!err) {
1612 			uint32_t ts_mode;
1613 
1614 			/* MTUTC register is read successfully. */
1615 			ts_mode = MLX5_GET(register_mtutc, reg,
1616 					   time_stamp_mode);
1617 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1618 				config->rt_timestamp = 1;
1619 		} else {
1620 			/* Kernel does not support register reading. */
1621 			if (config->hca_attr.dev_freq_khz ==
1622 						 (NS_PER_S / MS_PER_S))
1623 				config->rt_timestamp = 1;
1624 		}
1625 	}
1626 	/*
1627 	 * If HW has bug working with tunnel packet decapsulation and
1628 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1629 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1630 	 */
1631 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1632 		config->hw_fcs_strip = 0;
1633 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1634 		(config->hw_fcs_strip ? "" : "not "));
1635 	if (config->mprq.enabled && mprq) {
1636 		if (config->mprq.stride_num_n &&
1637 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1638 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1639 			config->mprq.stride_num_n =
1640 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1641 						mprq_min_stride_num_n),
1642 					mprq_max_stride_num_n);
1643 			DRV_LOG(WARNING,
1644 				"the number of strides"
1645 				" for Multi-Packet RQ is out of range,"
1646 				" setting default value (%u)",
1647 				1 << config->mprq.stride_num_n);
1648 		}
1649 		if (config->mprq.stride_size_n &&
1650 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1651 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1652 			config->mprq.stride_size_n =
1653 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1654 						mprq_min_stride_size_n),
1655 					mprq_max_stride_size_n);
1656 			DRV_LOG(WARNING,
1657 				"the size of a stride"
1658 				" for Multi-Packet RQ is out of range,"
1659 				" setting default value (%u)",
1660 				1 << config->mprq.stride_size_n);
1661 		}
1662 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1663 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1664 	} else if (config->mprq.enabled && !mprq) {
1665 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1666 		config->mprq.enabled = 0;
1667 	}
1668 	if (config->max_dump_files_num == 0)
1669 		config->max_dump_files_num = 128;
1670 	eth_dev = rte_eth_dev_allocate(name);
1671 	if (eth_dev == NULL) {
1672 		DRV_LOG(ERR, "can not allocate rte ethdev");
1673 		err = ENOMEM;
1674 		goto error;
1675 	}
1676 	if (priv->representor) {
1677 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1678 		eth_dev->data->representor_id = priv->representor_id;
1679 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1680 			struct mlx5_priv *opriv =
1681 				rte_eth_devices[port_id].data->dev_private;
1682 			if (opriv &&
1683 			    opriv->master &&
1684 			    opriv->domain_id == priv->domain_id &&
1685 			    opriv->sh == priv->sh) {
1686 				eth_dev->data->backer_port_id = port_id;
1687 				break;
1688 			}
1689 		}
1690 		if (port_id >= RTE_MAX_ETHPORTS)
1691 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1692 	}
1693 	priv->mp_id.port_id = eth_dev->data->port_id;
1694 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1695 	/*
1696 	 * Store associated network device interface index. This index
1697 	 * is permanent throughout the lifetime of device. So, we may store
1698 	 * the ifindex here and use the cached value further.
1699 	 */
1700 	MLX5_ASSERT(spawn->ifindex);
1701 	priv->if_index = spawn->ifindex;
1702 	eth_dev->data->dev_private = priv;
1703 	priv->dev_data = eth_dev->data;
1704 	eth_dev->data->mac_addrs = priv->mac;
1705 	eth_dev->device = dpdk_dev;
1706 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1707 	/* Configure the first MAC address by default. */
1708 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1709 		DRV_LOG(ERR,
1710 			"port %u cannot get MAC address, is mlx5_en"
1711 			" loaded? (errno: %s)",
1712 			eth_dev->data->port_id, strerror(rte_errno));
1713 		err = ENODEV;
1714 		goto error;
1715 	}
1716 	DRV_LOG(INFO,
1717 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1718 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1719 #ifdef RTE_LIBRTE_MLX5_DEBUG
1720 	{
1721 		char ifname[MLX5_NAMESIZE];
1722 
1723 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1724 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1725 				eth_dev->data->port_id, ifname);
1726 		else
1727 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1728 				eth_dev->data->port_id);
1729 	}
1730 #endif
1731 	/* Get actual MTU if possible. */
1732 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1733 	if (err) {
1734 		err = rte_errno;
1735 		goto error;
1736 	}
1737 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1738 		priv->mtu);
1739 	/* Initialize burst functions to prevent crashes before link-up. */
1740 	eth_dev->rx_pkt_burst = removed_rx_burst;
1741 	eth_dev->tx_pkt_burst = removed_tx_burst;
1742 	eth_dev->dev_ops = &mlx5_dev_ops;
1743 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1744 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1745 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1746 	/* Register MAC address. */
1747 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1748 	if (config->vf && config->vf_nl_en)
1749 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1750 				      mlx5_ifindex(eth_dev),
1751 				      eth_dev->data->mac_addrs,
1752 				      MLX5_MAX_MAC_ADDRESSES);
1753 	priv->ctrl_flows = 0;
1754 	rte_spinlock_init(&priv->flow_list_lock);
1755 	TAILQ_INIT(&priv->flow_meters);
1756 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1757 	if (!priv->mtr_profile_tbl)
1758 		goto error;
1759 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1760 	mlx5_glue->dv_set_context_attr(sh->ctx,
1761 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1762 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1763 				.alloc = &mlx5_alloc_verbs_buf,
1764 				.free = &mlx5_free_verbs_buf,
1765 				.data = sh,
1766 			}));
1767 	/* Bring Ethernet device up. */
1768 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1769 		eth_dev->data->port_id);
1770 	mlx5_set_link_up(eth_dev);
1771 	/*
1772 	 * Even though the interrupt handler is not installed yet,
1773 	 * interrupts will still trigger on the async_fd from
1774 	 * Verbs context returned by ibv_open_device().
1775 	 */
1776 	mlx5_link_update(eth_dev, 0);
1777 #ifdef HAVE_MLX5DV_DR_ESWITCH
1778 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1779 	      (switch_info->representor || switch_info->master)))
1780 		config->dv_esw_en = 0;
1781 #else
1782 	config->dv_esw_en = 0;
1783 #endif
1784 	/* Detect minimal data bytes to inline. */
1785 	mlx5_set_min_inline(spawn, config);
1786 	/* Store device configuration on private structure. */
1787 	priv->config = *config;
1788 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1789 		icfg[i].release_mem_en = !!config->reclaim_mode;
1790 		if (config->reclaim_mode)
1791 			icfg[i].per_core_cache = 0;
1792 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1793 		if (!priv->flows[i])
1794 			goto error;
1795 	}
1796 	/* Create context for virtual machine VLAN workaround. */
1797 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1798 	if (config->dv_flow_en) {
1799 		err = mlx5_alloc_shared_dr(priv);
1800 		if (err)
1801 			goto error;
1802 	}
1803 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1804 		priv->obj_ops = devx_obj_ops;
1805 		priv->obj_ops.drop_action_create =
1806 						ibv_obj_ops.drop_action_create;
1807 		priv->obj_ops.drop_action_destroy =
1808 						ibv_obj_ops.drop_action_destroy;
1809 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1810 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1811 #else
1812 		if (config->dv_esw_en)
1813 			priv->obj_ops.txq_obj_modify =
1814 						ibv_obj_ops.txq_obj_modify;
1815 #endif
1816 		/* Use specific wrappers for Tx object. */
1817 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1818 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1819 		mlx5_queue_counter_id_prepare(eth_dev);
1820 		priv->obj_ops.lb_dummy_queue_create =
1821 					mlx5_rxq_ibv_obj_dummy_lb_create;
1822 		priv->obj_ops.lb_dummy_queue_release =
1823 					mlx5_rxq_ibv_obj_dummy_lb_release;
1824 	} else {
1825 		priv->obj_ops = ibv_obj_ops;
1826 	}
1827 	if (config->tx_pp &&
1828 	    (priv->config.dv_esw_en ||
1829 	     priv->obj_ops.txq_obj_new != mlx5_os_txq_obj_new)) {
1830 		/*
1831 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1832 		 * packet pacing and already checked above.
1833 		 * Hence, we should only make sure the SQs will be created
1834 		 * with DevX, not with Verbs.
1835 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1836 		 * with Clock Queue UAR as required for Tx scheduling.
1837 		 */
1838 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1839 		err = ENODEV;
1840 		goto error;
1841 	}
1842 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1843 	if (!priv->drop_queue.hrxq)
1844 		goto error;
1845 	/* Supported Verbs flow priority number detection. */
1846 	err = mlx5_flow_discover_priorities(eth_dev);
1847 	if (err < 0) {
1848 		err = -err;
1849 		goto error;
1850 	}
1851 	priv->config.flow_prio = err;
1852 	if (!priv->config.dv_esw_en &&
1853 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1854 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1855 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1856 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1857 	}
1858 	mlx5_set_metadata_mask(eth_dev);
1859 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1860 	    !priv->sh->dv_regc0_mask) {
1861 		DRV_LOG(ERR, "metadata mode %u is not supported "
1862 			     "(no metadata reg_c[0] is available)",
1863 			     priv->config.dv_xmeta_en);
1864 			err = ENOTSUP;
1865 			goto error;
1866 	}
1867 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1868 				       mlx5_hrxq_create_cb,
1869 				       mlx5_hrxq_match_cb,
1870 				       mlx5_hrxq_remove_cb,
1871 				       mlx5_hrxq_clone_cb,
1872 				       mlx5_hrxq_clone_free_cb);
1873 	if (!priv->hrxqs)
1874 		goto error;
1875 	rte_rwlock_init(&priv->ind_tbls_lock);
1876 	/* Query availability of metadata reg_c's. */
1877 	err = mlx5_flow_discover_mreg_c(eth_dev);
1878 	if (err < 0) {
1879 		err = -err;
1880 		goto error;
1881 	}
1882 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1883 		DRV_LOG(DEBUG,
1884 			"port %u extensive metadata register is not supported",
1885 			eth_dev->data->port_id);
1886 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1887 			DRV_LOG(ERR, "metadata mode %u is not supported "
1888 				     "(no metadata registers available)",
1889 				     priv->config.dv_xmeta_en);
1890 			err = ENOTSUP;
1891 			goto error;
1892 		}
1893 	}
1894 	if (priv->config.dv_flow_en &&
1895 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1896 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1897 	    priv->sh->dv_regc0_mask) {
1898 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1899 						      MLX5_FLOW_MREG_HTABLE_SZ,
1900 						      false, true, eth_dev,
1901 						      flow_dv_mreg_create_cb,
1902 						      flow_dv_mreg_match_cb,
1903 						      flow_dv_mreg_remove_cb,
1904 						      flow_dv_mreg_clone_cb,
1905 						    flow_dv_mreg_clone_free_cb);
1906 		if (!priv->mreg_cp_tbl) {
1907 			err = ENOMEM;
1908 			goto error;
1909 		}
1910 	}
1911 	rte_spinlock_init(&priv->shared_act_sl);
1912 	mlx5_flow_counter_mode_config(eth_dev);
1913 	mlx5_flow_drop_action_config(eth_dev);
1914 	if (priv->config.dv_flow_en)
1915 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1916 	return eth_dev;
1917 error:
1918 	if (priv) {
1919 		if (priv->mreg_cp_tbl)
1920 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1921 		if (priv->sh)
1922 			mlx5_os_free_shared_dr(priv);
1923 		if (priv->nl_socket_route >= 0)
1924 			close(priv->nl_socket_route);
1925 		if (priv->nl_socket_rdma >= 0)
1926 			close(priv->nl_socket_rdma);
1927 		if (priv->vmwa_context)
1928 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1929 		if (eth_dev && priv->drop_queue.hrxq)
1930 			mlx5_drop_action_destroy(eth_dev);
1931 		if (priv->mtr_profile_tbl)
1932 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1933 		if (own_domain_id)
1934 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1935 		if (priv->hrxqs)
1936 			mlx5_list_destroy(priv->hrxqs);
1937 		mlx5_free(priv);
1938 		if (eth_dev != NULL)
1939 			eth_dev->data->dev_private = NULL;
1940 	}
1941 	if (eth_dev != NULL) {
1942 		/* mac_addrs must not be freed alone because part of
1943 		 * dev_private
1944 		 **/
1945 		eth_dev->data->mac_addrs = NULL;
1946 		rte_eth_dev_release_port(eth_dev);
1947 	}
1948 	if (sh)
1949 		mlx5_free_shared_dev_ctx(sh);
1950 	MLX5_ASSERT(err > 0);
1951 	rte_errno = err;
1952 	return NULL;
1953 }
1954 
1955 /**
1956  * Comparison callback to sort device data.
1957  *
1958  * This is meant to be used with qsort().
1959  *
1960  * @param a[in]
1961  *   Pointer to pointer to first data object.
1962  * @param b[in]
1963  *   Pointer to pointer to second data object.
1964  *
1965  * @return
1966  *   0 if both objects are equal, less than 0 if the first argument is less
1967  *   than the second, greater than 0 otherwise.
1968  */
1969 static int
1970 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1971 {
1972 	const struct mlx5_switch_info *si_a =
1973 		&((const struct mlx5_dev_spawn_data *)a)->info;
1974 	const struct mlx5_switch_info *si_b =
1975 		&((const struct mlx5_dev_spawn_data *)b)->info;
1976 	int ret;
1977 
1978 	/* Master device first. */
1979 	ret = si_b->master - si_a->master;
1980 	if (ret)
1981 		return ret;
1982 	/* Then representor devices. */
1983 	ret = si_b->representor - si_a->representor;
1984 	if (ret)
1985 		return ret;
1986 	/* Unidentified devices come last in no specific order. */
1987 	if (!si_a->representor)
1988 		return 0;
1989 	/* Order representors by name. */
1990 	return si_a->port_name - si_b->port_name;
1991 }
1992 
1993 /**
1994  * Match PCI information for possible slaves of bonding device.
1995  *
1996  * @param[in] ibv_dev
1997  *   Pointer to Infiniband device structure.
1998  * @param[in] pci_dev
1999  *   Pointer to primary PCI address structure to match.
2000  * @param[in] nl_rdma
2001  *   Netlink RDMA group socket handle.
2002  * @param[in] owner
2003  *   Rerepsentor owner PF index.
2004  * @param[out] bond_info
2005  *   Pointer to bonding information.
2006  *
2007  * @return
2008  *   negative value if no bonding device found, otherwise
2009  *   positive index of slave PF in bonding.
2010  */
2011 static int
2012 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2013 			   const struct rte_pci_addr *pci_dev,
2014 			   int nl_rdma, uint16_t owner,
2015 			   struct mlx5_bond_info *bond_info)
2016 {
2017 	char ifname[IF_NAMESIZE + 1];
2018 	unsigned int ifindex;
2019 	unsigned int np, i;
2020 	FILE *bond_file = NULL, *file;
2021 	int pf = -1;
2022 	int ret;
2023 
2024 	/*
2025 	 * Try to get master device name. If something goes
2026 	 * wrong suppose the lack of kernel support and no
2027 	 * bonding devices.
2028 	 */
2029 	memset(bond_info, 0, sizeof(*bond_info));
2030 	if (nl_rdma < 0)
2031 		return -1;
2032 	if (!strstr(ibv_dev->name, "bond"))
2033 		return -1;
2034 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2035 	if (!np)
2036 		return -1;
2037 	/*
2038 	 * The Master device might not be on the predefined
2039 	 * port (not on port index 1, it is not garanted),
2040 	 * we have to scan all Infiniband device port and
2041 	 * find master.
2042 	 */
2043 	for (i = 1; i <= np; ++i) {
2044 		/* Check whether Infiniband port is populated. */
2045 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2046 		if (!ifindex)
2047 			continue;
2048 		if (!if_indextoname(ifindex, ifname))
2049 			continue;
2050 		/* Try to read bonding slave names from sysfs. */
2051 		MKSTR(slaves,
2052 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2053 		bond_file = fopen(slaves, "r");
2054 		if (bond_file)
2055 			break;
2056 	}
2057 	if (!bond_file)
2058 		return -1;
2059 	/* Use safe format to check maximal buffer length. */
2060 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2061 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2062 		char tmp_str[IF_NAMESIZE + 32];
2063 		struct rte_pci_addr pci_addr;
2064 		struct mlx5_switch_info	info;
2065 
2066 		/* Process slave interface names in the loop. */
2067 		snprintf(tmp_str, sizeof(tmp_str),
2068 			 "/sys/class/net/%s", ifname);
2069 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
2070 			DRV_LOG(WARNING, "can not get PCI address"
2071 					 " for netdev \"%s\"", ifname);
2072 			continue;
2073 		}
2074 		/* Slave interface PCI address match found. */
2075 		snprintf(tmp_str, sizeof(tmp_str),
2076 			 "/sys/class/net/%s/phys_port_name", ifname);
2077 		file = fopen(tmp_str, "rb");
2078 		if (!file)
2079 			break;
2080 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2081 		if (fscanf(file, "%32s", tmp_str) == 1)
2082 			mlx5_translate_port_name(tmp_str, &info);
2083 		fclose(file);
2084 		/* Only process PF ports. */
2085 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2086 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2087 			continue;
2088 		/* Check max bonding member. */
2089 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2090 			DRV_LOG(WARNING, "bonding index out of range, "
2091 				"please increase MLX5_BOND_MAX_PORTS: %s",
2092 				tmp_str);
2093 			break;
2094 		}
2095 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
2096 		if (pci_dev->domain == pci_addr.domain &&
2097 		    pci_dev->bus == pci_addr.bus &&
2098 		    pci_dev->devid == pci_addr.devid &&
2099 		    ((pci_dev->function == 0 &&
2100 		      pci_dev->function + owner == pci_addr.function) ||
2101 		     (pci_dev->function == owner &&
2102 		      pci_addr.function == owner)))
2103 			pf = info.port_name;
2104 		/* Get ifindex. */
2105 		snprintf(tmp_str, sizeof(tmp_str),
2106 			 "/sys/class/net/%s/ifindex", ifname);
2107 		file = fopen(tmp_str, "rb");
2108 		if (!file)
2109 			break;
2110 		ret = fscanf(file, "%u", &ifindex);
2111 		fclose(file);
2112 		if (ret != 1)
2113 			break;
2114 		/* Save bonding info. */
2115 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2116 			sizeof(bond_info->ports[0].ifname));
2117 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2118 		bond_info->ports[info.port_name].ifindex = ifindex;
2119 		bond_info->n_port++;
2120 	}
2121 	if (pf >= 0) {
2122 		/* Get bond interface info */
2123 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2124 					   bond_info->ifname);
2125 		if (ret)
2126 			DRV_LOG(ERR, "unable to get bond info: %s",
2127 				strerror(rte_errno));
2128 		else
2129 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2130 				ifindex, bond_info->ifindex, bond_info->ifname);
2131 	}
2132 	return pf;
2133 }
2134 
2135 static void
2136 mlx5_os_config_default(struct mlx5_dev_config *config)
2137 {
2138 	memset(config, 0, sizeof(*config));
2139 	config->mps = MLX5_ARG_UNSET;
2140 	config->dbnc = MLX5_ARG_UNSET;
2141 	config->rx_vec_en = 1;
2142 	config->txq_inline_max = MLX5_ARG_UNSET;
2143 	config->txq_inline_min = MLX5_ARG_UNSET;
2144 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2145 	config->txqs_inline = MLX5_ARG_UNSET;
2146 	config->vf_nl_en = 1;
2147 	config->mr_ext_memseg_en = 1;
2148 	config->mr_mempool_reg_en = 1;
2149 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2150 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2151 	config->dv_esw_en = 1;
2152 	config->dv_flow_en = 1;
2153 	config->decap_en = 1;
2154 	config->log_hp_size = MLX5_ARG_UNSET;
2155 	config->allow_duplicate_pattern = 1;
2156 }
2157 
2158 /**
2159  * Register a PCI device within bonding.
2160  *
2161  * This function spawns Ethernet devices out of a given PCI device and
2162  * bonding owner PF index.
2163  *
2164  * @param[in] pci_dev
2165  *   PCI device information.
2166  * @param[in] req_eth_da
2167  *   Requested ethdev device argument.
2168  * @param[in] owner_id
2169  *   Requested owner PF port ID within bonding device, default to 0.
2170  *
2171  * @return
2172  *   0 on success, a negative errno value otherwise and rte_errno is set.
2173  */
2174 static int
2175 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
2176 		     struct rte_eth_devargs *req_eth_da,
2177 		     uint16_t owner_id)
2178 {
2179 	struct ibv_device **ibv_list;
2180 	/*
2181 	 * Number of found IB Devices matching with requested PCI BDF.
2182 	 * nd != 1 means there are multiple IB devices over the same
2183 	 * PCI device and we have representors and master.
2184 	 */
2185 	unsigned int nd = 0;
2186 	/*
2187 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2188 	 * we have the single multiport IB device, and there may be
2189 	 * representors attached to some of found ports.
2190 	 */
2191 	unsigned int np = 0;
2192 	/*
2193 	 * Number of DPDK ethernet devices to Spawn - either over
2194 	 * multiple IB devices or multiple ports of single IB device.
2195 	 * Actually this is the number of iterations to spawn.
2196 	 */
2197 	unsigned int ns = 0;
2198 	/*
2199 	 * Bonding device
2200 	 *   < 0 - no bonding device (single one)
2201 	 *  >= 0 - bonding device (value is slave PF index)
2202 	 */
2203 	int bd = -1;
2204 	struct mlx5_dev_spawn_data *list = NULL;
2205 	struct mlx5_dev_config dev_config;
2206 	unsigned int dev_config_vf;
2207 	struct rte_eth_devargs eth_da = *req_eth_da;
2208 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2209 	struct mlx5_bond_info bond_info;
2210 	int ret = -1;
2211 
2212 	errno = 0;
2213 	ibv_list = mlx5_glue->get_device_list(&ret);
2214 	if (!ibv_list) {
2215 		rte_errno = errno ? errno : ENOSYS;
2216 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2217 		return -rte_errno;
2218 	}
2219 	/*
2220 	 * First scan the list of all Infiniband devices to find
2221 	 * matching ones, gathering into the list.
2222 	 */
2223 	struct ibv_device *ibv_match[ret + 1];
2224 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2225 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2226 	unsigned int i;
2227 
2228 	while (ret-- > 0) {
2229 		struct rte_pci_addr pci_addr;
2230 
2231 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2232 		bd = mlx5_device_bond_pci_match
2233 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2234 				 &bond_info);
2235 		if (bd >= 0) {
2236 			/*
2237 			 * Bonding device detected. Only one match is allowed,
2238 			 * the bonding is supported over multi-port IB device,
2239 			 * there should be no matches on representor PCI
2240 			 * functions or non VF LAG bonding devices with
2241 			 * specified address.
2242 			 */
2243 			if (nd) {
2244 				DRV_LOG(ERR,
2245 					"multiple PCI match on bonding device"
2246 					"\"%s\" found", ibv_list[ret]->name);
2247 				rte_errno = ENOENT;
2248 				ret = -rte_errno;
2249 				goto exit;
2250 			}
2251 			/* Amend owner pci address if owner PF ID specified. */
2252 			if (eth_da.nb_representor_ports)
2253 				owner_pci.function += owner_id;
2254 			DRV_LOG(INFO, "PCI information matches for"
2255 				      " slave %d bonding device \"%s\"",
2256 				      bd, ibv_list[ret]->name);
2257 			ibv_match[nd++] = ibv_list[ret];
2258 			break;
2259 		} else {
2260 			/* Bonding device not found. */
2261 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2262 					      &pci_addr))
2263 				continue;
2264 			if (owner_pci.domain != pci_addr.domain ||
2265 			    owner_pci.bus != pci_addr.bus ||
2266 			    owner_pci.devid != pci_addr.devid ||
2267 			    owner_pci.function != pci_addr.function)
2268 				continue;
2269 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2270 				ibv_list[ret]->name);
2271 			ibv_match[nd++] = ibv_list[ret];
2272 		}
2273 	}
2274 	ibv_match[nd] = NULL;
2275 	if (!nd) {
2276 		/* No device matches, just complain and bail out. */
2277 		DRV_LOG(WARNING,
2278 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2279 			" are kernel drivers loaded?",
2280 			owner_pci.domain, owner_pci.bus,
2281 			owner_pci.devid, owner_pci.function);
2282 		rte_errno = ENOENT;
2283 		ret = -rte_errno;
2284 		goto exit;
2285 	}
2286 	if (nd == 1) {
2287 		/*
2288 		 * Found single matching device may have multiple ports.
2289 		 * Each port may be representor, we have to check the port
2290 		 * number and check the representors existence.
2291 		 */
2292 		if (nl_rdma >= 0)
2293 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2294 		if (!np)
2295 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2296 					 " ports number", ibv_match[0]->name);
2297 		if (bd >= 0 && !np) {
2298 			DRV_LOG(ERR, "can not get ports"
2299 				     " for bonding device");
2300 			rte_errno = ENOENT;
2301 			ret = -rte_errno;
2302 			goto exit;
2303 		}
2304 	}
2305 	/*
2306 	 * Now we can determine the maximal
2307 	 * amount of devices to be spawned.
2308 	 */
2309 	list = mlx5_malloc(MLX5_MEM_ZERO,
2310 			   sizeof(struct mlx5_dev_spawn_data) *
2311 			   (np ? np : nd),
2312 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2313 	if (!list) {
2314 		DRV_LOG(ERR, "spawn data array allocation failure");
2315 		rte_errno = ENOMEM;
2316 		ret = -rte_errno;
2317 		goto exit;
2318 	}
2319 	if (bd >= 0 || np > 1) {
2320 		/*
2321 		 * Single IB device with multiple ports found,
2322 		 * it may be E-Switch master device and representors.
2323 		 * We have to perform identification through the ports.
2324 		 */
2325 		MLX5_ASSERT(nl_rdma >= 0);
2326 		MLX5_ASSERT(ns == 0);
2327 		MLX5_ASSERT(nd == 1);
2328 		MLX5_ASSERT(np);
2329 		for (i = 1; i <= np; ++i) {
2330 			list[ns].bond_info = &bond_info;
2331 			list[ns].max_port = np;
2332 			list[ns].phys_port = i;
2333 			list[ns].phys_dev = ibv_match[0];
2334 			list[ns].eth_dev = NULL;
2335 			list[ns].pci_dev = pci_dev;
2336 			list[ns].pf_bond = bd;
2337 			list[ns].ifindex = mlx5_nl_ifindex
2338 				(nl_rdma,
2339 				mlx5_os_get_dev_device_name
2340 						(list[ns].phys_dev), i);
2341 			if (!list[ns].ifindex) {
2342 				/*
2343 				 * No network interface index found for the
2344 				 * specified port, it means there is no
2345 				 * representor on this port. It's OK,
2346 				 * there can be disabled ports, for example
2347 				 * if sriov_numvfs < sriov_totalvfs.
2348 				 */
2349 				continue;
2350 			}
2351 			ret = -1;
2352 			if (nl_route >= 0)
2353 				ret = mlx5_nl_switch_info
2354 					       (nl_route,
2355 						list[ns].ifindex,
2356 						&list[ns].info);
2357 			if (ret || (!list[ns].info.representor &&
2358 				    !list[ns].info.master)) {
2359 				/*
2360 				 * We failed to recognize representors with
2361 				 * Netlink, let's try to perform the task
2362 				 * with sysfs.
2363 				 */
2364 				ret =  mlx5_sysfs_switch_info
2365 						(list[ns].ifindex,
2366 						 &list[ns].info);
2367 			}
2368 			if (!ret && bd >= 0) {
2369 				switch (list[ns].info.name_type) {
2370 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2371 					if (np == 1) {
2372 						/*
2373 						 * Force standalone bonding
2374 						 * device for ROCE LAG
2375 						 * confgiurations.
2376 						 */
2377 						list[ns].info.master = 0;
2378 						list[ns].info.representor = 0;
2379 					}
2380 					if (list[ns].info.port_name == bd)
2381 						ns++;
2382 					break;
2383 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2384 					/* Fallthrough */
2385 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2386 					/* Fallthrough */
2387 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2388 					if (list[ns].info.pf_num == bd)
2389 						ns++;
2390 					break;
2391 				default:
2392 					break;
2393 				}
2394 				continue;
2395 			}
2396 			if (!ret && (list[ns].info.representor ^
2397 				     list[ns].info.master))
2398 				ns++;
2399 		}
2400 		if (!ns) {
2401 			DRV_LOG(ERR,
2402 				"unable to recognize master/representors"
2403 				" on the IB device with multiple ports");
2404 			rte_errno = ENOENT;
2405 			ret = -rte_errno;
2406 			goto exit;
2407 		}
2408 	} else {
2409 		/*
2410 		 * The existence of several matching entries (nd > 1) means
2411 		 * port representors have been instantiated. No existing Verbs
2412 		 * call nor sysfs entries can tell them apart, this can only
2413 		 * be done through Netlink calls assuming kernel drivers are
2414 		 * recent enough to support them.
2415 		 *
2416 		 * In the event of identification failure through Netlink,
2417 		 * try again through sysfs, then:
2418 		 *
2419 		 * 1. A single IB device matches (nd == 1) with single
2420 		 *    port (np=0/1) and is not a representor, assume
2421 		 *    no switch support.
2422 		 *
2423 		 * 2. Otherwise no safe assumptions can be made;
2424 		 *    complain louder and bail out.
2425 		 */
2426 		for (i = 0; i != nd; ++i) {
2427 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2428 			list[ns].bond_info = NULL;
2429 			list[ns].max_port = 1;
2430 			list[ns].phys_port = 1;
2431 			list[ns].phys_dev = ibv_match[i];
2432 			list[ns].eth_dev = NULL;
2433 			list[ns].pci_dev = pci_dev;
2434 			list[ns].pf_bond = -1;
2435 			list[ns].ifindex = 0;
2436 			if (nl_rdma >= 0)
2437 				list[ns].ifindex = mlx5_nl_ifindex
2438 				(nl_rdma,
2439 				mlx5_os_get_dev_device_name
2440 						(list[ns].phys_dev), 1);
2441 			if (!list[ns].ifindex) {
2442 				char ifname[IF_NAMESIZE];
2443 
2444 				/*
2445 				 * Netlink failed, it may happen with old
2446 				 * ib_core kernel driver (before 4.16).
2447 				 * We can assume there is old driver because
2448 				 * here we are processing single ports IB
2449 				 * devices. Let's try sysfs to retrieve
2450 				 * the ifindex. The method works for
2451 				 * master device only.
2452 				 */
2453 				if (nd > 1) {
2454 					/*
2455 					 * Multiple devices found, assume
2456 					 * representors, can not distinguish
2457 					 * master/representor and retrieve
2458 					 * ifindex via sysfs.
2459 					 */
2460 					continue;
2461 				}
2462 				ret = mlx5_get_ifname_sysfs
2463 					(ibv_match[i]->ibdev_path, ifname);
2464 				if (!ret)
2465 					list[ns].ifindex =
2466 						if_nametoindex(ifname);
2467 				if (!list[ns].ifindex) {
2468 					/*
2469 					 * No network interface index found
2470 					 * for the specified device, it means
2471 					 * there it is neither representor
2472 					 * nor master.
2473 					 */
2474 					continue;
2475 				}
2476 			}
2477 			ret = -1;
2478 			if (nl_route >= 0)
2479 				ret = mlx5_nl_switch_info
2480 					       (nl_route,
2481 						list[ns].ifindex,
2482 						&list[ns].info);
2483 			if (ret || (!list[ns].info.representor &&
2484 				    !list[ns].info.master)) {
2485 				/*
2486 				 * We failed to recognize representors with
2487 				 * Netlink, let's try to perform the task
2488 				 * with sysfs.
2489 				 */
2490 				ret =  mlx5_sysfs_switch_info
2491 						(list[ns].ifindex,
2492 						 &list[ns].info);
2493 			}
2494 			if (!ret && (list[ns].info.representor ^
2495 				     list[ns].info.master)) {
2496 				ns++;
2497 			} else if ((nd == 1) &&
2498 				   !list[ns].info.representor &&
2499 				   !list[ns].info.master) {
2500 				/*
2501 				 * Single IB device with
2502 				 * one physical port and
2503 				 * attached network device.
2504 				 * May be SRIOV is not enabled
2505 				 * or there is no representors.
2506 				 */
2507 				DRV_LOG(INFO, "no E-Switch support detected");
2508 				ns++;
2509 				break;
2510 			}
2511 		}
2512 		if (!ns) {
2513 			DRV_LOG(ERR,
2514 				"unable to recognize master/representors"
2515 				" on the multiple IB devices");
2516 			rte_errno = ENOENT;
2517 			ret = -rte_errno;
2518 			goto exit;
2519 		}
2520 		/*
2521 		 * New kernels may add the switch_id attribute for the case
2522 		 * there is no E-Switch and we wrongly recognized the
2523 		 * only device as master. Override this if there is the
2524 		 * single device with single port and new device name
2525 		 * format present.
2526 		 */
2527 		if (nd == 1 &&
2528 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2529 			list[0].info.master = 0;
2530 			list[0].info.representor = 0;
2531 		}
2532 	}
2533 	MLX5_ASSERT(ns);
2534 	/*
2535 	 * Sort list to probe devices in natural order for users convenience
2536 	 * (i.e. master first, then representors from lowest to highest ID).
2537 	 */
2538 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2539 	/* Device specific configuration. */
2540 	switch (pci_dev->id.device_id) {
2541 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2542 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2543 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2544 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2545 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2546 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2547 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2548 		dev_config_vf = 1;
2549 		break;
2550 	default:
2551 		dev_config_vf = 0;
2552 		break;
2553 	}
2554 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2555 		/* Set devargs default values. */
2556 		if (eth_da.nb_mh_controllers == 0) {
2557 			eth_da.nb_mh_controllers = 1;
2558 			eth_da.mh_controllers[0] = 0;
2559 		}
2560 		if (eth_da.nb_ports == 0 && ns > 0) {
2561 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2562 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2563 					pci_dev->device.devargs->args);
2564 			eth_da.nb_ports = 1;
2565 			eth_da.ports[0] = list[0].info.pf_num;
2566 		}
2567 		if (eth_da.nb_representor_ports == 0) {
2568 			eth_da.nb_representor_ports = 1;
2569 			eth_da.representor_ports[0] = 0;
2570 		}
2571 	}
2572 	for (i = 0; i != ns; ++i) {
2573 		uint32_t restore;
2574 
2575 		/* Default configuration. */
2576 		mlx5_os_config_default(&dev_config);
2577 		dev_config.vf = dev_config_vf;
2578 		list[i].numa_node = pci_dev->device.numa_node;
2579 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2580 						 &list[i],
2581 						 &dev_config,
2582 						 &eth_da);
2583 		if (!list[i].eth_dev) {
2584 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2585 				break;
2586 			/* Device is disabled or already spawned. Ignore it. */
2587 			continue;
2588 		}
2589 		restore = list[i].eth_dev->data->dev_flags;
2590 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2591 		/**
2592 		 * Each representor has a dedicated interrupts vector.
2593 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2594 		 * representor eth_dev object because representor and PF
2595 		 * share the same PCI address.
2596 		 * Override representor device with a dedicated
2597 		 * interrupts handle here.
2598 		 * Representor interrupts handle is released in mlx5_dev_stop().
2599 		 */
2600 		if (list[i].info.representor) {
2601 			struct rte_intr_handle *intr_handle;
2602 			intr_handle = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
2603 						  sizeof(*intr_handle), 0,
2604 						  SOCKET_ID_ANY);
2605 			if (!intr_handle) {
2606 				DRV_LOG(ERR,
2607 					"port %u failed to allocate memory for interrupt handler "
2608 					"Rx interrupts will not be supported",
2609 					i);
2610 				rte_errno = ENOMEM;
2611 				ret = -rte_errno;
2612 				goto exit;
2613 			}
2614 			list[i].eth_dev->intr_handle = intr_handle;
2615 		}
2616 		/* Restore non-PCI flags cleared by the above call. */
2617 		list[i].eth_dev->data->dev_flags |= restore;
2618 		rte_eth_dev_probing_finish(list[i].eth_dev);
2619 	}
2620 	if (i != ns) {
2621 		DRV_LOG(ERR,
2622 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2623 			" encountering an error: %s",
2624 			owner_pci.domain, owner_pci.bus,
2625 			owner_pci.devid, owner_pci.function,
2626 			strerror(rte_errno));
2627 		ret = -rte_errno;
2628 		/* Roll back. */
2629 		while (i--) {
2630 			if (!list[i].eth_dev)
2631 				continue;
2632 			mlx5_dev_close(list[i].eth_dev);
2633 			/* mac_addrs must not be freed because in dev_private */
2634 			list[i].eth_dev->data->mac_addrs = NULL;
2635 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2636 		}
2637 		/* Restore original error. */
2638 		rte_errno = -ret;
2639 	} else {
2640 		ret = 0;
2641 	}
2642 exit:
2643 	/*
2644 	 * Do the routine cleanup:
2645 	 * - close opened Netlink sockets
2646 	 * - free allocated spawn data array
2647 	 * - free the Infiniband device list
2648 	 */
2649 	if (nl_rdma >= 0)
2650 		close(nl_rdma);
2651 	if (nl_route >= 0)
2652 		close(nl_route);
2653 	if (list)
2654 		mlx5_free(list);
2655 	MLX5_ASSERT(ibv_list);
2656 	mlx5_glue->free_device_list(ibv_list);
2657 	return ret;
2658 }
2659 
2660 static int
2661 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2662 			  struct rte_eth_devargs *eth_da)
2663 {
2664 	int ret = 0;
2665 
2666 	if (dev->devargs == NULL)
2667 		return 0;
2668 	memset(eth_da, 0, sizeof(*eth_da));
2669 	/* Parse representor information first from class argument. */
2670 	if (dev->devargs->cls_str)
2671 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2672 	if (ret != 0) {
2673 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2674 			dev->devargs->cls_str);
2675 		return -rte_errno;
2676 	}
2677 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2678 		/* Parse legacy device argument */
2679 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2680 		if (ret) {
2681 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2682 				dev->devargs->args);
2683 			return -rte_errno;
2684 		}
2685 	}
2686 	return 0;
2687 }
2688 
2689 /**
2690  * Callback to register a PCI device.
2691  *
2692  * This function spawns Ethernet devices out of a given PCI device.
2693  *
2694  * @param[in] pci_dev
2695  *   PCI device information.
2696  *
2697  * @return
2698  *   0 on success, a negative errno value otherwise and rte_errno is set.
2699  */
2700 static int
2701 mlx5_os_pci_probe(struct rte_pci_device *pci_dev)
2702 {
2703 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2704 	int ret = 0;
2705 	uint16_t p;
2706 
2707 	ret = mlx5_os_parse_eth_devargs(&pci_dev->device, &eth_da);
2708 	if (ret != 0)
2709 		return ret;
2710 
2711 	if (eth_da.nb_ports > 0) {
2712 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2713 		for (p = 0; p < eth_da.nb_ports; p++) {
2714 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2715 						   eth_da.ports[p]);
2716 			if (ret)
2717 				break;
2718 		}
2719 		if (ret) {
2720 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2721 				"aborted due to proding failure of PF %u",
2722 				pci_dev->addr.domain, pci_dev->addr.bus,
2723 				pci_dev->addr.devid, pci_dev->addr.function,
2724 				eth_da.ports[p]);
2725 			mlx5_net_remove(&pci_dev->device);
2726 		}
2727 	} else {
2728 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2729 	}
2730 	return ret;
2731 }
2732 
2733 /* Probe a single SF device on auxiliary bus, no representor support. */
2734 static int
2735 mlx5_os_auxiliary_probe(struct rte_device *dev)
2736 {
2737 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2738 	struct mlx5_dev_config config;
2739 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2740 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2741 	struct rte_eth_dev *eth_dev;
2742 	int ret = 0;
2743 
2744 	/* Parse ethdev devargs. */
2745 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2746 	if (ret != 0)
2747 		return ret;
2748 	/* Set default config data. */
2749 	mlx5_os_config_default(&config);
2750 	config.sf = 1;
2751 	/* Init spawn data. */
2752 	spawn.max_port = 1;
2753 	spawn.phys_port = 1;
2754 	spawn.phys_dev = mlx5_os_get_ibv_dev(dev);
2755 	if (spawn.phys_dev == NULL)
2756 		return -rte_errno;
2757 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2758 	if (ret < 0) {
2759 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2760 		return ret;
2761 	}
2762 	spawn.ifindex = ret;
2763 	spawn.numa_node = dev->numa_node;
2764 	/* Spawn device. */
2765 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2766 	if (eth_dev == NULL)
2767 		return -rte_errno;
2768 	/* Post create. */
2769 	eth_dev->intr_handle = &adev->intr_handle;
2770 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2771 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2772 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2773 		eth_dev->data->numa_node = dev->numa_node;
2774 	}
2775 	rte_eth_dev_probing_finish(eth_dev);
2776 	return 0;
2777 }
2778 
2779 /**
2780  * Net class driver callback to probe a device.
2781  *
2782  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2783  *
2784  * @param[in] dev
2785  *   Pointer to the generic device.
2786  *
2787  * @return
2788  *   0 on success, the function cannot fail.
2789  */
2790 int
2791 mlx5_os_net_probe(struct rte_device *dev)
2792 {
2793 	int ret;
2794 
2795 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2796 		mlx5_pmd_socket_init();
2797 	ret = mlx5_init_once();
2798 	if (ret) {
2799 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2800 			strerror(rte_errno));
2801 		return -rte_errno;
2802 	}
2803 	if (mlx5_dev_is_pci(dev))
2804 		return mlx5_os_pci_probe(RTE_DEV_TO_PCI(dev));
2805 	else
2806 		return mlx5_os_auxiliary_probe(dev);
2807 }
2808 
2809 static int
2810 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2811 {
2812 	char *env;
2813 	int value;
2814 
2815 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2816 	/* Get environment variable to store. */
2817 	env = getenv(MLX5_SHUT_UP_BF);
2818 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2819 	if (config->dbnc == MLX5_ARG_UNSET)
2820 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2821 	else
2822 		setenv(MLX5_SHUT_UP_BF,
2823 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2824 	return value;
2825 }
2826 
2827 static void
2828 mlx5_restore_doorbell_mapping_env(int value)
2829 {
2830 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2831 	/* Restore the original environment variable state. */
2832 	if (value == MLX5_ARG_UNSET)
2833 		unsetenv(MLX5_SHUT_UP_BF);
2834 	else
2835 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2836 }
2837 
2838 /**
2839  * Extract pdn of PD object using DV API.
2840  *
2841  * @param[in] pd
2842  *   Pointer to the verbs PD object.
2843  * @param[out] pdn
2844  *   Pointer to the PD object number variable.
2845  *
2846  * @return
2847  *   0 on success, error value otherwise.
2848  */
2849 int
2850 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2851 {
2852 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2853 	struct mlx5dv_obj obj;
2854 	struct mlx5dv_pd pd_info;
2855 	int ret = 0;
2856 
2857 	obj.pd.in = pd;
2858 	obj.pd.out = &pd_info;
2859 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2860 	if (ret) {
2861 		DRV_LOG(DEBUG, "Fail to get PD object info");
2862 		return ret;
2863 	}
2864 	*pdn = pd_info.pdn;
2865 	return 0;
2866 #else
2867 	(void)pd;
2868 	(void)pdn;
2869 	return -ENOTSUP;
2870 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2871 }
2872 
2873 /**
2874  * Function API to open IB device.
2875  *
2876  * This function calls the Linux glue APIs to open a device.
2877  *
2878  * @param[in] spawn
2879  *   Pointer to the IB device attributes (name, port, etc).
2880  * @param[out] config
2881  *   Pointer to device configuration structure.
2882  * @param[out] sh
2883  *   Pointer to shared context structure.
2884  *
2885  * @return
2886  *   0 on success, a positive error value otherwise.
2887  */
2888 int
2889 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2890 		     const struct mlx5_dev_config *config,
2891 		     struct mlx5_dev_ctx_shared *sh)
2892 {
2893 	int dbmap_env;
2894 	int err = 0;
2895 
2896 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2897 	/*
2898 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2899 	 * before the device creation. The rdma_core library
2900 	 * checks the variable at device creation and
2901 	 * stores the result internally.
2902 	 */
2903 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2904 	/* Try to open IB device with DV first, then usual Verbs. */
2905 	errno = 0;
2906 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2907 	if (sh->ctx) {
2908 		sh->devx = 1;
2909 		DRV_LOG(DEBUG, "DevX is supported");
2910 		/* The device is created, no need for environment. */
2911 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2912 	} else {
2913 		/* The environment variable is still configured. */
2914 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2915 		err = errno ? errno : ENODEV;
2916 		/*
2917 		 * The environment variable is not needed anymore,
2918 		 * all device creation attempts are completed.
2919 		 */
2920 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2921 		if (!sh->ctx)
2922 			return err;
2923 		DRV_LOG(DEBUG, "DevX is NOT supported");
2924 		err = 0;
2925 	}
2926 	if (!err && sh->ctx) {
2927 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2928 		mlx5_glue->dv_set_context_attr(sh->ctx,
2929 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2930 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2931 				.alloc = &mlx5_alloc_verbs_buf,
2932 				.free = &mlx5_free_verbs_buf,
2933 				.data = sh,
2934 			}));
2935 	}
2936 	return err;
2937 }
2938 
2939 /**
2940  * Install shared asynchronous device events handler.
2941  * This function is implemented to support event sharing
2942  * between multiple ports of single IB device.
2943  *
2944  * @param sh
2945  *   Pointer to mlx5_dev_ctx_shared object.
2946  */
2947 void
2948 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2949 {
2950 	int ret;
2951 	int flags;
2952 
2953 	sh->intr_handle.fd = -1;
2954 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2955 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2956 		    F_SETFL, flags | O_NONBLOCK);
2957 	if (ret) {
2958 		DRV_LOG(INFO, "failed to change file descriptor async event"
2959 			" queue");
2960 	} else {
2961 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2962 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2963 		if (rte_intr_callback_register(&sh->intr_handle,
2964 					mlx5_dev_interrupt_handler, sh)) {
2965 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2966 			sh->intr_handle.fd = -1;
2967 		}
2968 	}
2969 	if (sh->devx) {
2970 #ifdef HAVE_IBV_DEVX_ASYNC
2971 		sh->intr_handle_devx.fd = -1;
2972 		sh->devx_comp =
2973 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2974 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2975 		if (!devx_comp) {
2976 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2977 			return;
2978 		}
2979 		flags = fcntl(devx_comp->fd, F_GETFL);
2980 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2981 		if (ret) {
2982 			DRV_LOG(INFO, "failed to change file descriptor"
2983 				" devx comp");
2984 			return;
2985 		}
2986 		sh->intr_handle_devx.fd = devx_comp->fd;
2987 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2988 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2989 					mlx5_dev_interrupt_handler_devx, sh)) {
2990 			DRV_LOG(INFO, "Fail to install the devx shared"
2991 				" interrupt.");
2992 			sh->intr_handle_devx.fd = -1;
2993 		}
2994 #endif /* HAVE_IBV_DEVX_ASYNC */
2995 	}
2996 }
2997 
2998 /**
2999  * Uninstall shared asynchronous device events handler.
3000  * This function is implemented to support event sharing
3001  * between multiple ports of single IB device.
3002  *
3003  * @param dev
3004  *   Pointer to mlx5_dev_ctx_shared object.
3005  */
3006 void
3007 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
3008 {
3009 	if (sh->intr_handle.fd >= 0)
3010 		mlx5_intr_callback_unregister(&sh->intr_handle,
3011 					      mlx5_dev_interrupt_handler, sh);
3012 #ifdef HAVE_IBV_DEVX_ASYNC
3013 	if (sh->intr_handle_devx.fd >= 0)
3014 		rte_intr_callback_unregister(&sh->intr_handle_devx,
3015 				  mlx5_dev_interrupt_handler_devx, sh);
3016 	if (sh->devx_comp)
3017 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
3018 #endif
3019 }
3020 
3021 /**
3022  * Read statistics by a named counter.
3023  *
3024  * @param[in] priv
3025  *   Pointer to the private device data structure.
3026  * @param[in] ctr_name
3027  *   Pointer to the name of the statistic counter to read
3028  * @param[out] stat
3029  *   Pointer to read statistic value.
3030  * @return
3031  *   0 on success and stat is valud, 1 if failed to read the value
3032  *   rte_errno is set.
3033  *
3034  */
3035 int
3036 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
3037 		      uint64_t *stat)
3038 {
3039 	int fd;
3040 
3041 	if (priv->sh) {
3042 		if (priv->q_counters != NULL &&
3043 		    strcmp(ctr_name, "out_of_buffer") == 0)
3044 			return mlx5_devx_cmd_queue_counter_query
3045 					(priv->q_counters, 0, (uint32_t *)stat);
3046 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
3047 		      priv->sh->ibdev_path,
3048 		      priv->dev_port,
3049 		      ctr_name);
3050 		fd = open(path, O_RDONLY);
3051 		/*
3052 		 * in switchdev the file location is not per port
3053 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
3054 		 */
3055 		if (fd == -1) {
3056 			MKSTR(path1, "%s/hw_counters/%s",
3057 			      priv->sh->ibdev_path,
3058 			      ctr_name);
3059 			fd = open(path1, O_RDONLY);
3060 		}
3061 		if (fd != -1) {
3062 			char buf[21] = {'\0'};
3063 			ssize_t n = read(fd, buf, sizeof(buf));
3064 
3065 			close(fd);
3066 			if (n != -1) {
3067 				*stat = strtoull(buf, NULL, 10);
3068 				return 0;
3069 			}
3070 		}
3071 	}
3072 	*stat = 0;
3073 	return 1;
3074 }
3075 
3076 /**
3077  * Set the reg_mr and dereg_mr call backs
3078  *
3079  * @param reg_mr_cb[out]
3080  *   Pointer to reg_mr func
3081  * @param dereg_mr_cb[out]
3082  *   Pointer to dereg_mr func
3083  *
3084  */
3085 void
3086 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
3087 		      mlx5_dereg_mr_t *dereg_mr_cb)
3088 {
3089 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
3090 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
3091 }
3092 
3093 /**
3094  * Remove a MAC address from device
3095  *
3096  * @param dev
3097  *   Pointer to Ethernet device structure.
3098  * @param index
3099  *   MAC address index.
3100  */
3101 void
3102 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3103 {
3104 	struct mlx5_priv *priv = dev->data->dev_private;
3105 	const int vf = priv->config.vf;
3106 
3107 	if (vf)
3108 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
3109 					mlx5_ifindex(dev), priv->mac_own,
3110 					&dev->data->mac_addrs[index], index);
3111 }
3112 
3113 /**
3114  * Adds a MAC address to the device
3115  *
3116  * @param dev
3117  *   Pointer to Ethernet device structure.
3118  * @param mac_addr
3119  *   MAC address to register.
3120  * @param index
3121  *   MAC address index.
3122  *
3123  * @return
3124  *   0 on success, a negative errno value otherwise
3125  */
3126 int
3127 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
3128 		     uint32_t index)
3129 {
3130 	struct mlx5_priv *priv = dev->data->dev_private;
3131 	const int vf = priv->config.vf;
3132 	int ret = 0;
3133 
3134 	if (vf)
3135 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
3136 					   mlx5_ifindex(dev), priv->mac_own,
3137 					   mac, index);
3138 	return ret;
3139 }
3140 
3141 /**
3142  * Modify a VF MAC address
3143  *
3144  * @param priv
3145  *   Pointer to device private data.
3146  * @param mac_addr
3147  *   MAC address to modify into.
3148  * @param iface_idx
3149  *   Net device interface index
3150  * @param vf_index
3151  *   VF index
3152  *
3153  * @return
3154  *   0 on success, a negative errno value otherwise
3155  */
3156 int
3157 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
3158 			   unsigned int iface_idx,
3159 			   struct rte_ether_addr *mac_addr,
3160 			   int vf_index)
3161 {
3162 	return mlx5_nl_vf_mac_addr_modify
3163 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
3164 }
3165 
3166 /**
3167  * Set device promiscuous mode
3168  *
3169  * @param dev
3170  *   Pointer to Ethernet device structure.
3171  * @param enable
3172  *   0 - promiscuous is disabled, otherwise - enabled
3173  *
3174  * @return
3175  *   0 on success, a negative error value otherwise
3176  */
3177 int
3178 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
3179 {
3180 	struct mlx5_priv *priv = dev->data->dev_private;
3181 
3182 	return mlx5_nl_promisc(priv->nl_socket_route,
3183 			       mlx5_ifindex(dev), !!enable);
3184 }
3185 
3186 /**
3187  * Set device promiscuous mode
3188  *
3189  * @param dev
3190  *   Pointer to Ethernet device structure.
3191  * @param enable
3192  *   0 - all multicase is disabled, otherwise - enabled
3193  *
3194  * @return
3195  *   0 on success, a negative error value otherwise
3196  */
3197 int
3198 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
3199 {
3200 	struct mlx5_priv *priv = dev->data->dev_private;
3201 
3202 	return mlx5_nl_allmulti(priv->nl_socket_route,
3203 				mlx5_ifindex(dev), !!enable);
3204 }
3205 
3206 /**
3207  * Flush device MAC addresses
3208  *
3209  * @param dev
3210  *   Pointer to Ethernet device structure.
3211  *
3212  */
3213 void
3214 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3215 {
3216 	struct mlx5_priv *priv = dev->data->dev_private;
3217 
3218 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3219 			       dev->data->mac_addrs,
3220 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3221 }
3222