xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 9c30a6f3c9a456e8111a2b1e5f6c2c02a62025b6)
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 static void
790 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
791 {
792 	struct mlx5_priv *priv = dev->data->dev_private;
793 	void *ctx = priv->sh->ctx;
794 
795 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
796 	if (!priv->q_counters) {
797 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
798 		struct ibv_wq *wq;
799 
800 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
801 			"by DevX - fall-back to use the kernel driver global "
802 			"queue counter.", dev->data->port_id);
803 		/* Create WQ by kernel and query its queue counter ID. */
804 		if (cq) {
805 			wq = mlx5_glue->create_wq(ctx,
806 						  &(struct ibv_wq_init_attr){
807 						    .wq_type = IBV_WQT_RQ,
808 						    .max_wr = 1,
809 						    .max_sge = 1,
810 						    .pd = priv->sh->pd,
811 						    .cq = cq,
812 						});
813 			if (wq) {
814 				/* Counter is assigned only on RDY state. */
815 				int ret = mlx5_glue->modify_wq(wq,
816 						 &(struct ibv_wq_attr){
817 						 .attr_mask = IBV_WQ_ATTR_STATE,
818 						 .wq_state = IBV_WQS_RDY,
819 						});
820 
821 				if (ret == 0)
822 					mlx5_devx_cmd_wq_query(wq,
823 							 &priv->counter_set_id);
824 				claim_zero(mlx5_glue->destroy_wq(wq));
825 			}
826 			claim_zero(mlx5_glue->destroy_cq(cq));
827 		}
828 	} else {
829 		priv->counter_set_id = priv->q_counters->id;
830 	}
831 	if (priv->counter_set_id == 0)
832 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
833 			"available.", dev->data->port_id);
834 }
835 
836 /**
837  * Check if representor spawn info match devargs.
838  *
839  * @param spawn
840  *   Verbs device parameters (name, port, switch_info) to spawn.
841  * @param eth_da
842  *   Device devargs to probe.
843  *
844  * @return
845  *   Match result.
846  */
847 static bool
848 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
849 		       struct rte_eth_devargs *eth_da)
850 {
851 	struct mlx5_switch_info *switch_info = &spawn->info;
852 	unsigned int p, f;
853 	uint16_t id;
854 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
855 						      eth_da->type);
856 
857 	switch (eth_da->type) {
858 	case RTE_ETH_REPRESENTOR_SF:
859 		if (!(spawn->info.port_name == -1 &&
860 		      switch_info->name_type ==
861 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
862 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
863 			rte_errno = EBUSY;
864 			return false;
865 		}
866 		break;
867 	case RTE_ETH_REPRESENTOR_VF:
868 		/* Allows HPF representor index -1 as exception. */
869 		if (!(spawn->info.port_name == -1 &&
870 		      switch_info->name_type ==
871 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
872 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
873 			rte_errno = EBUSY;
874 			return false;
875 		}
876 		break;
877 	case RTE_ETH_REPRESENTOR_NONE:
878 		rte_errno = EBUSY;
879 		return false;
880 	default:
881 		rte_errno = ENOTSUP;
882 		DRV_LOG(ERR, "unsupported representor type");
883 		return false;
884 	}
885 	/* Check representor ID: */
886 	for (p = 0; p < eth_da->nb_ports; ++p) {
887 		if (spawn->pf_bond < 0) {
888 			/* For non-LAG mode, allow and ignore pf. */
889 			switch_info->pf_num = eth_da->ports[p];
890 			repr_id = mlx5_representor_id_encode(switch_info,
891 							     eth_da->type);
892 		}
893 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
894 			id = MLX5_REPRESENTOR_ID
895 				(eth_da->ports[p], eth_da->type,
896 				 eth_da->representor_ports[f]);
897 			if (repr_id == id)
898 				return true;
899 		}
900 	}
901 	rte_errno = EBUSY;
902 	return false;
903 }
904 
905 
906 /**
907  * Spawn an Ethernet device from Verbs information.
908  *
909  * @param dpdk_dev
910  *   Backing DPDK device.
911  * @param spawn
912  *   Verbs device parameters (name, port, switch_info) to spawn.
913  * @param config
914  *   Device configuration parameters.
915  * @param config
916  *   Device arguments.
917  *
918  * @return
919  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
920  *   is set. The following errors are defined:
921  *
922  *   EBUSY: device is not supposed to be spawned.
923  *   EEXIST: device is already spawned
924  */
925 static struct rte_eth_dev *
926 mlx5_dev_spawn(struct rte_device *dpdk_dev,
927 	       struct mlx5_dev_spawn_data *spawn,
928 	       struct mlx5_dev_config *config,
929 	       struct rte_eth_devargs *eth_da)
930 {
931 	const struct mlx5_switch_info *switch_info = &spawn->info;
932 	struct mlx5_dev_ctx_shared *sh = NULL;
933 	struct ibv_port_attr port_attr;
934 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
935 	struct rte_eth_dev *eth_dev = NULL;
936 	struct mlx5_priv *priv = NULL;
937 	int err = 0;
938 	unsigned int hw_padding = 0;
939 	unsigned int mps;
940 	unsigned int tunnel_en = 0;
941 	unsigned int mpls_en = 0;
942 	unsigned int swp = 0;
943 	unsigned int mprq = 0;
944 	unsigned int mprq_min_stride_size_n = 0;
945 	unsigned int mprq_max_stride_size_n = 0;
946 	unsigned int mprq_min_stride_num_n = 0;
947 	unsigned int mprq_max_stride_num_n = 0;
948 	struct rte_ether_addr mac;
949 	char name[RTE_ETH_NAME_MAX_LEN];
950 	int own_domain_id = 0;
951 	uint16_t port_id;
952 	struct mlx5_port_info vport_info = { .query_flags = 0 };
953 	int i;
954 
955 	/* Determine if this port representor is supposed to be spawned. */
956 	if (switch_info->representor && dpdk_dev->devargs &&
957 	    !mlx5_representor_match(spawn, eth_da))
958 		return NULL;
959 	/* Build device name. */
960 	if (spawn->pf_bond < 0) {
961 		/* Single device. */
962 		if (!switch_info->representor)
963 			strlcpy(name, dpdk_dev->name, sizeof(name));
964 		else
965 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
966 				 dpdk_dev->name,
967 				 switch_info->name_type ==
968 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
969 				 switch_info->port_name);
970 	} else {
971 		/* Bonding device. */
972 		if (!switch_info->representor) {
973 			err = snprintf(name, sizeof(name), "%s_%s",
974 				 dpdk_dev->name,
975 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
976 		} else {
977 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
978 				dpdk_dev->name,
979 				mlx5_os_get_dev_device_name(spawn->phys_dev),
980 				switch_info->ctrl_num,
981 				switch_info->pf_num,
982 				switch_info->name_type ==
983 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
984 				switch_info->port_name);
985 		}
986 	}
987 	if (err >= (int)sizeof(name))
988 		DRV_LOG(WARNING, "device name overflow %s", name);
989 	/* check if the device is already spawned */
990 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
991 		rte_errno = EEXIST;
992 		return NULL;
993 	}
994 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
995 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
996 		struct mlx5_mp_id mp_id;
997 
998 		eth_dev = rte_eth_dev_attach_secondary(name);
999 		if (eth_dev == NULL) {
1000 			DRV_LOG(ERR, "can not attach rte ethdev");
1001 			rte_errno = ENOMEM;
1002 			return NULL;
1003 		}
1004 		eth_dev->device = dpdk_dev;
1005 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1006 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1007 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1008 		err = mlx5_proc_priv_init(eth_dev);
1009 		if (err)
1010 			return NULL;
1011 		mp_id.port_id = eth_dev->data->port_id;
1012 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1013 		/* Receive command fd from primary process */
1014 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1015 		if (err < 0)
1016 			goto err_secondary;
1017 		/* Remap UAR for Tx queues. */
1018 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1019 		if (err)
1020 			goto err_secondary;
1021 		/*
1022 		 * Ethdev pointer is still required as input since
1023 		 * the primary device is not accessible from the
1024 		 * secondary process.
1025 		 */
1026 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1027 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1028 		return eth_dev;
1029 err_secondary:
1030 		mlx5_dev_close(eth_dev);
1031 		return NULL;
1032 	}
1033 	/*
1034 	 * Some parameters ("tx_db_nc" in particularly) are needed in
1035 	 * advance to create dv/verbs device context. We proceed the
1036 	 * devargs here to get ones, and later proceed devargs again
1037 	 * to override some hardware settings.
1038 	 */
1039 	err = mlx5_args(config, dpdk_dev->devargs);
1040 	if (err) {
1041 		err = rte_errno;
1042 		DRV_LOG(ERR, "failed to process device arguments: %s",
1043 			strerror(rte_errno));
1044 		goto error;
1045 	}
1046 	if (config->dv_miss_info) {
1047 		if (switch_info->master || switch_info->representor)
1048 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
1049 	}
1050 	mlx5_malloc_mem_select(config->sys_mem_en);
1051 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
1052 	if (!sh)
1053 		return NULL;
1054 	config->devx = sh->devx;
1055 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1056 	config->dest_tir = 1;
1057 #endif
1058 #ifdef HAVE_IBV_MLX5_MOD_SWP
1059 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1060 #endif
1061 	/*
1062 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1063 	 * as all ConnectX-5 devices.
1064 	 */
1065 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1066 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1067 #endif
1068 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1069 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1070 #endif
1071 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1072 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1073 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1074 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1075 			mps = MLX5_MPW_ENHANCED;
1076 		} else {
1077 			DRV_LOG(DEBUG, "MPW is supported");
1078 			mps = MLX5_MPW;
1079 		}
1080 	} else {
1081 		DRV_LOG(DEBUG, "MPW isn't supported");
1082 		mps = MLX5_MPW_DISABLED;
1083 	}
1084 #ifdef HAVE_IBV_MLX5_MOD_SWP
1085 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1086 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1087 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1088 #endif
1089 	config->swp = !!swp;
1090 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1091 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1092 		struct mlx5dv_striding_rq_caps mprq_caps =
1093 			dv_attr.striding_rq_caps;
1094 
1095 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1096 			mprq_caps.min_single_stride_log_num_of_bytes);
1097 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1098 			mprq_caps.max_single_stride_log_num_of_bytes);
1099 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1100 			mprq_caps.min_single_wqe_log_num_of_strides);
1101 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1102 			mprq_caps.max_single_wqe_log_num_of_strides);
1103 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1104 			mprq_caps.supported_qpts);
1105 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1106 		mprq = 1;
1107 		mprq_min_stride_size_n =
1108 			mprq_caps.min_single_stride_log_num_of_bytes;
1109 		mprq_max_stride_size_n =
1110 			mprq_caps.max_single_stride_log_num_of_bytes;
1111 		mprq_min_stride_num_n =
1112 			mprq_caps.min_single_wqe_log_num_of_strides;
1113 		mprq_max_stride_num_n =
1114 			mprq_caps.max_single_wqe_log_num_of_strides;
1115 	}
1116 #endif
1117 	/* Rx CQE compression is enabled by default. */
1118 	config->cqe_comp = 1;
1119 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1120 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1121 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1122 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1123 			     (dv_attr.tunnel_offloads_caps &
1124 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
1125 			     (dv_attr.tunnel_offloads_caps &
1126 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1127 	}
1128 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1129 		tunnel_en ? "" : "not ");
1130 #else
1131 	DRV_LOG(WARNING,
1132 		"tunnel offloading disabled due to old OFED/rdma-core version");
1133 #endif
1134 	config->tunnel_en = tunnel_en;
1135 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1136 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1137 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1138 		   (dv_attr.tunnel_offloads_caps &
1139 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1140 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1141 		mpls_en ? "" : "not ");
1142 #else
1143 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1144 		" old OFED/rdma-core version or firmware configuration");
1145 #endif
1146 	config->mpls_en = mpls_en;
1147 	/* Check port status. */
1148 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1149 	if (err) {
1150 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1151 		goto error;
1152 	}
1153 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1154 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1155 		err = EINVAL;
1156 		goto error;
1157 	}
1158 	if (port_attr.state != IBV_PORT_ACTIVE)
1159 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1160 			mlx5_glue->port_state_str(port_attr.state),
1161 			port_attr.state);
1162 	/* Allocate private eth device data. */
1163 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1164 			   sizeof(*priv),
1165 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1166 	if (priv == NULL) {
1167 		DRV_LOG(ERR, "priv allocation failure");
1168 		err = ENOMEM;
1169 		goto error;
1170 	}
1171 	priv->sh = sh;
1172 	priv->dev_port = spawn->phys_port;
1173 	priv->pci_dev = spawn->pci_dev;
1174 	priv->mtu = RTE_ETHER_MTU;
1175 	/* Some internal functions rely on Netlink sockets, open them now. */
1176 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1177 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1178 	priv->representor = !!switch_info->representor;
1179 	priv->master = !!switch_info->master;
1180 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1181 	priv->vport_meta_tag = 0;
1182 	priv->vport_meta_mask = 0;
1183 	priv->pf_bond = spawn->pf_bond;
1184 	/*
1185 	 * If we have E-Switch we should determine the vport attributes.
1186 	 * E-Switch may use either source vport field or reg_c[0] metadata
1187 	 * register to match on vport index. The engaged part of metadata
1188 	 * register is defined by mask.
1189 	 */
1190 	if (switch_info->representor || switch_info->master) {
1191 		err = mlx5_glue->devx_port_query(sh->ctx,
1192 						 spawn->phys_port,
1193 						 &vport_info);
1194 		if (err) {
1195 			DRV_LOG(WARNING,
1196 				"can't query devx port %d on device %s",
1197 				spawn->phys_port,
1198 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1199 			vport_info.query_flags = 0;
1200 		}
1201 	}
1202 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1203 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1204 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1205 		if (!priv->vport_meta_mask) {
1206 			DRV_LOG(ERR, "vport zero mask for port %d"
1207 				     " on bonding device %s",
1208 				     spawn->phys_port,
1209 				     mlx5_os_get_dev_device_name
1210 							(spawn->phys_dev));
1211 			err = ENOTSUP;
1212 			goto error;
1213 		}
1214 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1215 			DRV_LOG(ERR, "invalid vport tag for port %d"
1216 				     " on bonding device %s",
1217 				     spawn->phys_port,
1218 				     mlx5_os_get_dev_device_name
1219 							(spawn->phys_dev));
1220 			err = ENOTSUP;
1221 			goto error;
1222 		}
1223 	}
1224 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1225 		priv->vport_id = vport_info.vport_id;
1226 	} else if (spawn->pf_bond >= 0 &&
1227 		   (switch_info->representor || switch_info->master)) {
1228 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1229 			     " on bonding device %s",
1230 			     spawn->phys_port,
1231 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1232 		err = ENOTSUP;
1233 		goto error;
1234 	} else {
1235 		/*
1236 		 * Suppose vport index in compatible way. Kernel/rdma_core
1237 		 * support single E-Switch per PF configurations only and
1238 		 * vport_id field contains the vport index for associated VF,
1239 		 * which is deduced from representor port name.
1240 		 * For example, let's have the IB device port 10, it has
1241 		 * attached network device eth0, which has port name attribute
1242 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1243 		 * as 3 (2+1). This assigning schema should be changed if the
1244 		 * multiple E-Switch instances per PF configurations or/and PCI
1245 		 * subfunctions are added.
1246 		 */
1247 		priv->vport_id = switch_info->representor ?
1248 				 switch_info->port_name + 1 : -1;
1249 	}
1250 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1251 							  eth_da->type);
1252 	/*
1253 	 * Look for sibling devices in order to reuse their switch domain
1254 	 * if any, otherwise allocate one.
1255 	 */
1256 	MLX5_ETH_FOREACH_DEV(port_id, NULL) {
1257 		const struct mlx5_priv *opriv =
1258 			rte_eth_devices[port_id].data->dev_private;
1259 
1260 		if (!opriv ||
1261 		    opriv->sh != priv->sh ||
1262 			opriv->domain_id ==
1263 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1264 			continue;
1265 		priv->domain_id = opriv->domain_id;
1266 		break;
1267 	}
1268 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1269 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1270 		if (err) {
1271 			err = rte_errno;
1272 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1273 				strerror(rte_errno));
1274 			goto error;
1275 		}
1276 		own_domain_id = 1;
1277 	}
1278 	/* Override some values set by hardware configuration. */
1279 	mlx5_args(config, dpdk_dev->devargs);
1280 	err = mlx5_dev_check_sibling_config(priv, config);
1281 	if (err)
1282 		goto error;
1283 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1284 			    IBV_DEVICE_RAW_IP_CSUM);
1285 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1286 		(config->hw_csum ? "" : "not "));
1287 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1288 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1289 	DRV_LOG(DEBUG, "counters are not supported");
1290 #endif
1291 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1292 	if (config->dv_flow_en) {
1293 		DRV_LOG(WARNING, "DV flow is not supported");
1294 		config->dv_flow_en = 0;
1295 	}
1296 #endif
1297 	if (spawn->max_port > UINT8_MAX) {
1298 		/* Verbs can't support ports larger than 255 by design. */
1299 		DRV_LOG(ERR, "can't support IB ports > UINT8_MAX");
1300 		err = EINVAL;
1301 		goto error;
1302 	}
1303 	config->ind_table_max_size =
1304 		sh->device_attr.max_rwq_indirection_table_size;
1305 	/*
1306 	 * Remove this check once DPDK supports larger/variable
1307 	 * indirection tables.
1308 	 */
1309 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1310 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1311 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1312 		config->ind_table_max_size);
1313 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1314 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1315 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1316 		(config->hw_vlan_strip ? "" : "not "));
1317 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1318 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1319 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1320 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1321 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1322 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1323 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1324 #endif
1325 	if (config->hw_padding && !hw_padding) {
1326 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1327 		config->hw_padding = 0;
1328 	} else if (config->hw_padding) {
1329 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1330 	}
1331 	config->tso = (sh->device_attr.max_tso > 0 &&
1332 		      (sh->device_attr.tso_supported_qpts &
1333 		       (1 << IBV_QPT_RAW_PACKET)));
1334 	if (config->tso)
1335 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1336 	/*
1337 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1338 	 * by default.
1339 	 */
1340 	if (config->mps == MLX5_ARG_UNSET)
1341 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1342 							  MLX5_MPW_DISABLED;
1343 	else
1344 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1345 	DRV_LOG(INFO, "%sMPS is %s",
1346 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1347 		config->mps == MLX5_MPW ? "legacy " : "",
1348 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1349 	if (config->devx) {
1350 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1351 		if (err) {
1352 			err = -err;
1353 			goto error;
1354 		}
1355 		/* Check relax ordering support. */
1356 		if (!haswell_broadwell_cpu) {
1357 			sh->cmng.relaxed_ordering_write =
1358 				config->hca_attr.relaxed_ordering_write;
1359 			sh->cmng.relaxed_ordering_read =
1360 				config->hca_attr.relaxed_ordering_read;
1361 		} else {
1362 			sh->cmng.relaxed_ordering_read = 0;
1363 			sh->cmng.relaxed_ordering_write = 0;
1364 		}
1365 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1366 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1367 		sh->steering_format_version =
1368 			config->hca_attr.steering_format_version;
1369 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1370 		/* Check for LRO support. */
1371 		if (config->dest_tir && config->hca_attr.lro_cap &&
1372 		    config->dv_flow_en) {
1373 			/* TBD check tunnel lro caps. */
1374 			config->lro.supported = config->hca_attr.lro_cap;
1375 			DRV_LOG(DEBUG, "Device supports LRO");
1376 			/*
1377 			 * If LRO timeout is not configured by application,
1378 			 * use the minimal supported value.
1379 			 */
1380 			if (!config->lro.timeout)
1381 				config->lro.timeout =
1382 				config->hca_attr.lro_timer_supported_periods[0];
1383 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1384 				config->lro.timeout);
1385 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1386 				"required for coalescing is %d bytes",
1387 				config->hca_attr.lro_min_mss_size);
1388 		}
1389 #if defined(HAVE_MLX5DV_DR) && \
1390 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1391 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1392 		if (config->hca_attr.qos.sup &&
1393 		    config->hca_attr.qos.flow_meter_old &&
1394 		    config->dv_flow_en) {
1395 			uint8_t reg_c_mask =
1396 				config->hca_attr.qos.flow_meter_reg_c_ids;
1397 			/*
1398 			 * Meter needs two REG_C's for color match and pre-sfx
1399 			 * flow match. Here get the REG_C for color match.
1400 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1401 			 */
1402 			reg_c_mask &= 0xfc;
1403 			if (__builtin_popcount(reg_c_mask) < 1) {
1404 				priv->mtr_en = 0;
1405 				DRV_LOG(WARNING, "No available register for"
1406 					" meter.");
1407 			} else {
1408 				/*
1409 				 * The meter color register is used by the
1410 				 * flow-hit feature as well.
1411 				 * The flow-hit feature must use REG_C_3
1412 				 * Prefer REG_C_3 if it is available.
1413 				 */
1414 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1415 					priv->mtr_color_reg = REG_C_3;
1416 				else
1417 					priv->mtr_color_reg = ffs(reg_c_mask)
1418 							      - 1 + REG_C_0;
1419 				priv->mtr_en = 1;
1420 				priv->mtr_reg_share =
1421 				      config->hca_attr.qos.flow_meter;
1422 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1423 					priv->mtr_color_reg);
1424 			}
1425 		}
1426 		if (config->hca_attr.qos.sup &&
1427 			config->hca_attr.qos.flow_meter_aso_sup) {
1428 			uint32_t log_obj_size =
1429 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1430 			if (log_obj_size >=
1431 			config->hca_attr.qos.log_meter_aso_granularity &&
1432 			log_obj_size <=
1433 			config->hca_attr.qos.log_meter_aso_max_alloc)
1434 				sh->meter_aso_en = 1;
1435 		}
1436 		if (priv->mtr_en) {
1437 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1438 			if (err) {
1439 				err = -err;
1440 				goto error;
1441 			}
1442 		}
1443 		if (config->hca_attr.flow.tunnel_header_0_1)
1444 			sh->tunnel_header_0_1 = 1;
1445 #endif
1446 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1447 		if (config->hca_attr.flow_hit_aso &&
1448 		    priv->mtr_color_reg == REG_C_3) {
1449 			sh->flow_hit_aso_en = 1;
1450 			err = mlx5_flow_aso_age_mng_init(sh);
1451 			if (err) {
1452 				err = -err;
1453 				goto error;
1454 			}
1455 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1456 		}
1457 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1458 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1459 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1460 		if (config->hca_attr.ct_offload &&
1461 		    priv->mtr_color_reg == REG_C_3) {
1462 			err = mlx5_flow_aso_ct_mng_init(sh);
1463 			if (err) {
1464 				err = -err;
1465 				goto error;
1466 			}
1467 			DRV_LOG(DEBUG, "CT ASO is supported.");
1468 			sh->ct_aso_en = 1;
1469 		}
1470 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1471 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1472 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1473 		    config->dv_flow_en) {
1474 			priv->sampler_en = 1;
1475 			DRV_LOG(DEBUG, "Sampler enabled!");
1476 		} else {
1477 			priv->sampler_en = 0;
1478 			if (!config->hca_attr.log_max_ft_sampler_num)
1479 				DRV_LOG(WARNING,
1480 					"No available register for sampler.");
1481 			else
1482 				DRV_LOG(DEBUG, "DV flow is not supported!");
1483 		}
1484 #endif
1485 	}
1486 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1487 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1488 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1489 		config->cqe_comp = 0;
1490 	}
1491 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1492 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1493 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1494 				 " format isn't supported.");
1495 		config->cqe_comp = 0;
1496 	}
1497 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1498 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1499 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1500 				 " format isn't supported.");
1501 		config->cqe_comp = 0;
1502 	}
1503 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1504 			config->cqe_comp ? "" : "not ");
1505 	if (config->tx_pp) {
1506 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1507 			config->hca_attr.dev_freq_khz);
1508 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1509 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1510 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1511 			config->hca_attr.cross_channel ? "" : "not ");
1512 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1513 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1514 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1515 			config->hca_attr.non_wire_sq ? "" : "not ");
1516 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1517 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1518 			config->hca_attr.log_max_static_sq_wq);
1519 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1520 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1521 		if (!config->devx) {
1522 			DRV_LOG(ERR, "DevX is required for packet pacing");
1523 			err = ENODEV;
1524 			goto error;
1525 		}
1526 		if (!config->hca_attr.qos.packet_pacing) {
1527 			DRV_LOG(ERR, "Packet pacing is not supported");
1528 			err = ENODEV;
1529 			goto error;
1530 		}
1531 		if (!config->hca_attr.cross_channel) {
1532 			DRV_LOG(ERR, "Cross channel operations are"
1533 				     " required for packet pacing");
1534 			err = ENODEV;
1535 			goto error;
1536 		}
1537 		if (!config->hca_attr.wqe_index_ignore) {
1538 			DRV_LOG(ERR, "WQE index ignore feature is"
1539 				     " required for packet pacing");
1540 			err = ENODEV;
1541 			goto error;
1542 		}
1543 		if (!config->hca_attr.non_wire_sq) {
1544 			DRV_LOG(ERR, "Non-wire SQ feature is"
1545 				     " required for packet pacing");
1546 			err = ENODEV;
1547 			goto error;
1548 		}
1549 		if (!config->hca_attr.log_max_static_sq_wq) {
1550 			DRV_LOG(ERR, "Static WQE SQ feature is"
1551 				     " required for packet pacing");
1552 			err = ENODEV;
1553 			goto error;
1554 		}
1555 		if (!config->hca_attr.qos.wqe_rate_pp) {
1556 			DRV_LOG(ERR, "WQE rate mode is required"
1557 				     " for packet pacing");
1558 			err = ENODEV;
1559 			goto error;
1560 		}
1561 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1562 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1563 			     " can't create queues for packet pacing");
1564 		err = ENODEV;
1565 		goto error;
1566 #endif
1567 	}
1568 	if (config->devx) {
1569 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1570 
1571 		err = config->hca_attr.access_register_user ?
1572 			mlx5_devx_cmd_register_read
1573 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1574 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1575 		if (!err) {
1576 			uint32_t ts_mode;
1577 
1578 			/* MTUTC register is read successfully. */
1579 			ts_mode = MLX5_GET(register_mtutc, reg,
1580 					   time_stamp_mode);
1581 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1582 				config->rt_timestamp = 1;
1583 		} else {
1584 			/* Kernel does not support register reading. */
1585 			if (config->hca_attr.dev_freq_khz ==
1586 						 (NS_PER_S / MS_PER_S))
1587 				config->rt_timestamp = 1;
1588 		}
1589 	}
1590 	/*
1591 	 * If HW has bug working with tunnel packet decapsulation and
1592 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1593 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1594 	 */
1595 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1596 		config->hw_fcs_strip = 0;
1597 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1598 		(config->hw_fcs_strip ? "" : "not "));
1599 	if (config->mprq.enabled && mprq) {
1600 		if (config->mprq.stride_num_n &&
1601 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1602 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1603 			config->mprq.stride_num_n =
1604 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1605 						mprq_min_stride_num_n),
1606 					mprq_max_stride_num_n);
1607 			DRV_LOG(WARNING,
1608 				"the number of strides"
1609 				" for Multi-Packet RQ is out of range,"
1610 				" setting default value (%u)",
1611 				1 << config->mprq.stride_num_n);
1612 		}
1613 		if (config->mprq.stride_size_n &&
1614 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1615 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1616 			config->mprq.stride_size_n =
1617 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1618 						mprq_min_stride_size_n),
1619 					mprq_max_stride_size_n);
1620 			DRV_LOG(WARNING,
1621 				"the size of a stride"
1622 				" for Multi-Packet RQ is out of range,"
1623 				" setting default value (%u)",
1624 				1 << config->mprq.stride_size_n);
1625 		}
1626 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1627 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1628 	} else if (config->mprq.enabled && !mprq) {
1629 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1630 		config->mprq.enabled = 0;
1631 	}
1632 	if (config->max_dump_files_num == 0)
1633 		config->max_dump_files_num = 128;
1634 	eth_dev = rte_eth_dev_allocate(name);
1635 	if (eth_dev == NULL) {
1636 		DRV_LOG(ERR, "can not allocate rte ethdev");
1637 		err = ENOMEM;
1638 		goto error;
1639 	}
1640 	if (priv->representor) {
1641 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1642 		eth_dev->data->representor_id = priv->representor_id;
1643 	}
1644 	priv->mp_id.port_id = eth_dev->data->port_id;
1645 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1646 	/*
1647 	 * Store associated network device interface index. This index
1648 	 * is permanent throughout the lifetime of device. So, we may store
1649 	 * the ifindex here and use the cached value further.
1650 	 */
1651 	MLX5_ASSERT(spawn->ifindex);
1652 	priv->if_index = spawn->ifindex;
1653 	eth_dev->data->dev_private = priv;
1654 	priv->dev_data = eth_dev->data;
1655 	eth_dev->data->mac_addrs = priv->mac;
1656 	eth_dev->device = dpdk_dev;
1657 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1658 	/* Configure the first MAC address by default. */
1659 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1660 		DRV_LOG(ERR,
1661 			"port %u cannot get MAC address, is mlx5_en"
1662 			" loaded? (errno: %s)",
1663 			eth_dev->data->port_id, strerror(rte_errno));
1664 		err = ENODEV;
1665 		goto error;
1666 	}
1667 	DRV_LOG(INFO,
1668 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1669 		eth_dev->data->port_id,
1670 		mac.addr_bytes[0], mac.addr_bytes[1],
1671 		mac.addr_bytes[2], mac.addr_bytes[3],
1672 		mac.addr_bytes[4], mac.addr_bytes[5]);
1673 #ifdef RTE_LIBRTE_MLX5_DEBUG
1674 	{
1675 		char ifname[MLX5_NAMESIZE];
1676 
1677 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1678 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1679 				eth_dev->data->port_id, ifname);
1680 		else
1681 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1682 				eth_dev->data->port_id);
1683 	}
1684 #endif
1685 	/* Get actual MTU if possible. */
1686 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1687 	if (err) {
1688 		err = rte_errno;
1689 		goto error;
1690 	}
1691 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1692 		priv->mtu);
1693 	/* Initialize burst functions to prevent crashes before link-up. */
1694 	eth_dev->rx_pkt_burst = removed_rx_burst;
1695 	eth_dev->tx_pkt_burst = removed_tx_burst;
1696 	eth_dev->dev_ops = &mlx5_dev_ops;
1697 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1698 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1699 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1700 	/* Register MAC address. */
1701 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1702 	if (config->vf && config->vf_nl_en)
1703 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1704 				      mlx5_ifindex(eth_dev),
1705 				      eth_dev->data->mac_addrs,
1706 				      MLX5_MAX_MAC_ADDRESSES);
1707 	priv->ctrl_flows = 0;
1708 	rte_spinlock_init(&priv->flow_list_lock);
1709 	TAILQ_INIT(&priv->flow_meters);
1710 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1711 	if (!priv->mtr_profile_tbl)
1712 		goto error;
1713 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1714 	mlx5_glue->dv_set_context_attr(sh->ctx,
1715 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1716 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1717 				.alloc = &mlx5_alloc_verbs_buf,
1718 				.free = &mlx5_free_verbs_buf,
1719 				.data = sh,
1720 			}));
1721 	/* Bring Ethernet device up. */
1722 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1723 		eth_dev->data->port_id);
1724 	mlx5_set_link_up(eth_dev);
1725 	/*
1726 	 * Even though the interrupt handler is not installed yet,
1727 	 * interrupts will still trigger on the async_fd from
1728 	 * Verbs context returned by ibv_open_device().
1729 	 */
1730 	mlx5_link_update(eth_dev, 0);
1731 #ifdef HAVE_MLX5DV_DR_ESWITCH
1732 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1733 	      (switch_info->representor || switch_info->master)))
1734 		config->dv_esw_en = 0;
1735 #else
1736 	config->dv_esw_en = 0;
1737 #endif
1738 	/* Detect minimal data bytes to inline. */
1739 	mlx5_set_min_inline(spawn, config);
1740 	/* Store device configuration on private structure. */
1741 	priv->config = *config;
1742 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1743 		icfg[i].release_mem_en = !!config->reclaim_mode;
1744 		if (config->reclaim_mode)
1745 			icfg[i].per_core_cache = 0;
1746 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1747 		if (!priv->flows[i])
1748 			goto error;
1749 	}
1750 	/* Create context for virtual machine VLAN workaround. */
1751 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1752 	if (config->dv_flow_en) {
1753 		err = mlx5_alloc_shared_dr(priv);
1754 		if (err)
1755 			goto error;
1756 	}
1757 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1758 		priv->obj_ops = devx_obj_ops;
1759 		priv->obj_ops.drop_action_create =
1760 						ibv_obj_ops.drop_action_create;
1761 		priv->obj_ops.drop_action_destroy =
1762 						ibv_obj_ops.drop_action_destroy;
1763 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1764 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1765 #else
1766 		if (config->dv_esw_en)
1767 			priv->obj_ops.txq_obj_modify =
1768 						ibv_obj_ops.txq_obj_modify;
1769 #endif
1770 		/* Use specific wrappers for Tx object. */
1771 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1772 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1773 		mlx5_queue_counter_id_prepare(eth_dev);
1774 		priv->obj_ops.lb_dummy_queue_create =
1775 					mlx5_rxq_ibv_obj_dummy_lb_create;
1776 		priv->obj_ops.lb_dummy_queue_release =
1777 					mlx5_rxq_ibv_obj_dummy_lb_release;
1778 	} else {
1779 		priv->obj_ops = ibv_obj_ops;
1780 	}
1781 	if (config->tx_pp &&
1782 	    (priv->config.dv_esw_en ||
1783 	     priv->obj_ops.txq_obj_new != mlx5_os_txq_obj_new)) {
1784 		/*
1785 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1786 		 * packet pacing and already checked above.
1787 		 * Hence, we should only make sure the SQs will be created
1788 		 * with DevX, not with Verbs.
1789 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1790 		 * with Clock Queue UAR as required for Tx scheduling.
1791 		 */
1792 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1793 		err = ENODEV;
1794 		goto error;
1795 	}
1796 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1797 	if (!priv->drop_queue.hrxq)
1798 		goto error;
1799 	/* Supported Verbs flow priority number detection. */
1800 	err = mlx5_flow_discover_priorities(eth_dev);
1801 	if (err < 0) {
1802 		err = -err;
1803 		goto error;
1804 	}
1805 	priv->config.flow_prio = err;
1806 	if (!priv->config.dv_esw_en &&
1807 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1808 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1809 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1810 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1811 	}
1812 	mlx5_set_metadata_mask(eth_dev);
1813 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1814 	    !priv->sh->dv_regc0_mask) {
1815 		DRV_LOG(ERR, "metadata mode %u is not supported "
1816 			     "(no metadata reg_c[0] is available)",
1817 			     priv->config.dv_xmeta_en);
1818 			err = ENOTSUP;
1819 			goto error;
1820 	}
1821 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1822 				       mlx5_hrxq_create_cb,
1823 				       mlx5_hrxq_match_cb,
1824 				       mlx5_hrxq_remove_cb,
1825 				       mlx5_hrxq_clone_cb,
1826 				       mlx5_hrxq_clone_free_cb);
1827 	if (!priv->hrxqs)
1828 		goto error;
1829 	rte_rwlock_init(&priv->ind_tbls_lock);
1830 	/* Query availability of metadata reg_c's. */
1831 	err = mlx5_flow_discover_mreg_c(eth_dev);
1832 	if (err < 0) {
1833 		err = -err;
1834 		goto error;
1835 	}
1836 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1837 		DRV_LOG(DEBUG,
1838 			"port %u extensive metadata register is not supported",
1839 			eth_dev->data->port_id);
1840 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1841 			DRV_LOG(ERR, "metadata mode %u is not supported "
1842 				     "(no metadata registers available)",
1843 				     priv->config.dv_xmeta_en);
1844 			err = ENOTSUP;
1845 			goto error;
1846 		}
1847 	}
1848 	if (priv->config.dv_flow_en &&
1849 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1850 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1851 	    priv->sh->dv_regc0_mask) {
1852 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1853 						      MLX5_FLOW_MREG_HTABLE_SZ,
1854 						      false, true, eth_dev,
1855 						      flow_dv_mreg_create_cb,
1856 						      flow_dv_mreg_match_cb,
1857 						      flow_dv_mreg_remove_cb,
1858 						      flow_dv_mreg_clone_cb,
1859 						    flow_dv_mreg_clone_free_cb);
1860 		if (!priv->mreg_cp_tbl) {
1861 			err = ENOMEM;
1862 			goto error;
1863 		}
1864 	}
1865 	rte_spinlock_init(&priv->shared_act_sl);
1866 	mlx5_flow_counter_mode_config(eth_dev);
1867 	if (priv->config.dv_flow_en)
1868 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1869 	return eth_dev;
1870 error:
1871 	if (priv) {
1872 		if (priv->mreg_cp_tbl)
1873 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1874 		if (priv->sh)
1875 			mlx5_os_free_shared_dr(priv);
1876 		if (priv->nl_socket_route >= 0)
1877 			close(priv->nl_socket_route);
1878 		if (priv->nl_socket_rdma >= 0)
1879 			close(priv->nl_socket_rdma);
1880 		if (priv->vmwa_context)
1881 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1882 		if (eth_dev && priv->drop_queue.hrxq)
1883 			mlx5_drop_action_destroy(eth_dev);
1884 		if (priv->mtr_profile_tbl)
1885 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1886 		if (own_domain_id)
1887 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1888 		if (priv->hrxqs)
1889 			mlx5_list_destroy(priv->hrxqs);
1890 		mlx5_free(priv);
1891 		if (eth_dev != NULL)
1892 			eth_dev->data->dev_private = NULL;
1893 	}
1894 	if (eth_dev != NULL) {
1895 		/* mac_addrs must not be freed alone because part of
1896 		 * dev_private
1897 		 **/
1898 		eth_dev->data->mac_addrs = NULL;
1899 		rte_eth_dev_release_port(eth_dev);
1900 	}
1901 	if (sh)
1902 		mlx5_free_shared_dev_ctx(sh);
1903 	MLX5_ASSERT(err > 0);
1904 	rte_errno = err;
1905 	return NULL;
1906 }
1907 
1908 /**
1909  * Comparison callback to sort device data.
1910  *
1911  * This is meant to be used with qsort().
1912  *
1913  * @param a[in]
1914  *   Pointer to pointer to first data object.
1915  * @param b[in]
1916  *   Pointer to pointer to second data object.
1917  *
1918  * @return
1919  *   0 if both objects are equal, less than 0 if the first argument is less
1920  *   than the second, greater than 0 otherwise.
1921  */
1922 static int
1923 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1924 {
1925 	const struct mlx5_switch_info *si_a =
1926 		&((const struct mlx5_dev_spawn_data *)a)->info;
1927 	const struct mlx5_switch_info *si_b =
1928 		&((const struct mlx5_dev_spawn_data *)b)->info;
1929 	int ret;
1930 
1931 	/* Master device first. */
1932 	ret = si_b->master - si_a->master;
1933 	if (ret)
1934 		return ret;
1935 	/* Then representor devices. */
1936 	ret = si_b->representor - si_a->representor;
1937 	if (ret)
1938 		return ret;
1939 	/* Unidentified devices come last in no specific order. */
1940 	if (!si_a->representor)
1941 		return 0;
1942 	/* Order representors by name. */
1943 	return si_a->port_name - si_b->port_name;
1944 }
1945 
1946 /**
1947  * Match PCI information for possible slaves of bonding device.
1948  *
1949  * @param[in] ibv_dev
1950  *   Pointer to Infiniband device structure.
1951  * @param[in] pci_dev
1952  *   Pointer to primary PCI address structure to match.
1953  * @param[in] nl_rdma
1954  *   Netlink RDMA group socket handle.
1955  * @param[in] owner
1956  *   Rerepsentor owner PF index.
1957  * @param[out] bond_info
1958  *   Pointer to bonding information.
1959  *
1960  * @return
1961  *   negative value if no bonding device found, otherwise
1962  *   positive index of slave PF in bonding.
1963  */
1964 static int
1965 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1966 			   const struct rte_pci_addr *pci_dev,
1967 			   int nl_rdma, uint16_t owner,
1968 			   struct mlx5_bond_info *bond_info)
1969 {
1970 	char ifname[IF_NAMESIZE + 1];
1971 	unsigned int ifindex;
1972 	unsigned int np, i;
1973 	FILE *bond_file = NULL, *file;
1974 	int pf = -1;
1975 	int ret;
1976 
1977 	/*
1978 	 * Try to get master device name. If something goes
1979 	 * wrong suppose the lack of kernel support and no
1980 	 * bonding devices.
1981 	 */
1982 	memset(bond_info, 0, sizeof(*bond_info));
1983 	if (nl_rdma < 0)
1984 		return -1;
1985 	if (!strstr(ibv_dev->name, "bond"))
1986 		return -1;
1987 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1988 	if (!np)
1989 		return -1;
1990 	/*
1991 	 * The Master device might not be on the predefined
1992 	 * port (not on port index 1, it is not garanted),
1993 	 * we have to scan all Infiniband device port and
1994 	 * find master.
1995 	 */
1996 	for (i = 1; i <= np; ++i) {
1997 		/* Check whether Infiniband port is populated. */
1998 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1999 		if (!ifindex)
2000 			continue;
2001 		if (!if_indextoname(ifindex, ifname))
2002 			continue;
2003 		/* Try to read bonding slave names from sysfs. */
2004 		MKSTR(slaves,
2005 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
2006 		bond_file = fopen(slaves, "r");
2007 		if (bond_file)
2008 			break;
2009 	}
2010 	if (!bond_file)
2011 		return -1;
2012 	/* Use safe format to check maximal buffer length. */
2013 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2014 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2015 		char tmp_str[IF_NAMESIZE + 32];
2016 		struct rte_pci_addr pci_addr;
2017 		struct mlx5_switch_info	info;
2018 
2019 		/* Process slave interface names in the loop. */
2020 		snprintf(tmp_str, sizeof(tmp_str),
2021 			 "/sys/class/net/%s", ifname);
2022 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
2023 			DRV_LOG(WARNING, "can not get PCI address"
2024 					 " for netdev \"%s\"", ifname);
2025 			continue;
2026 		}
2027 		/* Slave interface PCI address match found. */
2028 		snprintf(tmp_str, sizeof(tmp_str),
2029 			 "/sys/class/net/%s/phys_port_name", ifname);
2030 		file = fopen(tmp_str, "rb");
2031 		if (!file)
2032 			break;
2033 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2034 		if (fscanf(file, "%32s", tmp_str) == 1)
2035 			mlx5_translate_port_name(tmp_str, &info);
2036 		fclose(file);
2037 		/* Only process PF ports. */
2038 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2039 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2040 			continue;
2041 		/* Check max bonding member. */
2042 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2043 			DRV_LOG(WARNING, "bonding index out of range, "
2044 				"please increase MLX5_BOND_MAX_PORTS: %s",
2045 				tmp_str);
2046 			break;
2047 		}
2048 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
2049 		if (pci_dev->domain == pci_addr.domain &&
2050 		    pci_dev->bus == pci_addr.bus &&
2051 		    pci_dev->devid == pci_addr.devid &&
2052 		    ((pci_dev->function == 0 &&
2053 		      pci_dev->function + owner == pci_addr.function) ||
2054 		     (pci_dev->function == owner &&
2055 		      pci_addr.function == owner)))
2056 			pf = info.port_name;
2057 		/* Get ifindex. */
2058 		snprintf(tmp_str, sizeof(tmp_str),
2059 			 "/sys/class/net/%s/ifindex", ifname);
2060 		file = fopen(tmp_str, "rb");
2061 		if (!file)
2062 			break;
2063 		ret = fscanf(file, "%u", &ifindex);
2064 		fclose(file);
2065 		if (ret != 1)
2066 			break;
2067 		/* Save bonding info. */
2068 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2069 			sizeof(bond_info->ports[0].ifname));
2070 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2071 		bond_info->ports[info.port_name].ifindex = ifindex;
2072 		bond_info->n_port++;
2073 	}
2074 	if (pf >= 0) {
2075 		/* Get bond interface info */
2076 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2077 					   bond_info->ifname);
2078 		if (ret)
2079 			DRV_LOG(ERR, "unable to get bond info: %s",
2080 				strerror(rte_errno));
2081 		else
2082 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2083 				ifindex, bond_info->ifindex, bond_info->ifname);
2084 	}
2085 	return pf;
2086 }
2087 
2088 static void
2089 mlx5_os_config_default(struct mlx5_dev_config *config)
2090 {
2091 	memset(config, 0, sizeof(*config));
2092 	config->mps = MLX5_ARG_UNSET;
2093 	config->dbnc = MLX5_ARG_UNSET;
2094 	config->rx_vec_en = 1;
2095 	config->txq_inline_max = MLX5_ARG_UNSET;
2096 	config->txq_inline_min = MLX5_ARG_UNSET;
2097 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2098 	config->txqs_inline = MLX5_ARG_UNSET;
2099 	config->vf_nl_en = 1;
2100 	config->mr_ext_memseg_en = 1;
2101 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2102 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2103 	config->dv_esw_en = 1;
2104 	config->dv_flow_en = 1;
2105 	config->decap_en = 1;
2106 	config->log_hp_size = MLX5_ARG_UNSET;
2107 }
2108 
2109 /**
2110  * Register a PCI device within bonding.
2111  *
2112  * This function spawns Ethernet devices out of a given PCI device and
2113  * bonding owner PF index.
2114  *
2115  * @param[in] pci_dev
2116  *   PCI device information.
2117  * @param[in] req_eth_da
2118  *   Requested ethdev device argument.
2119  * @param[in] owner_id
2120  *   Requested owner PF port ID within bonding device, default to 0.
2121  *
2122  * @return
2123  *   0 on success, a negative errno value otherwise and rte_errno is set.
2124  */
2125 static int
2126 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
2127 		     struct rte_eth_devargs *req_eth_da,
2128 		     uint16_t owner_id)
2129 {
2130 	struct ibv_device **ibv_list;
2131 	/*
2132 	 * Number of found IB Devices matching with requested PCI BDF.
2133 	 * nd != 1 means there are multiple IB devices over the same
2134 	 * PCI device and we have representors and master.
2135 	 */
2136 	unsigned int nd = 0;
2137 	/*
2138 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2139 	 * we have the single multiport IB device, and there may be
2140 	 * representors attached to some of found ports.
2141 	 */
2142 	unsigned int np = 0;
2143 	/*
2144 	 * Number of DPDK ethernet devices to Spawn - either over
2145 	 * multiple IB devices or multiple ports of single IB device.
2146 	 * Actually this is the number of iterations to spawn.
2147 	 */
2148 	unsigned int ns = 0;
2149 	/*
2150 	 * Bonding device
2151 	 *   < 0 - no bonding device (single one)
2152 	 *  >= 0 - bonding device (value is slave PF index)
2153 	 */
2154 	int bd = -1;
2155 	struct mlx5_dev_spawn_data *list = NULL;
2156 	struct mlx5_dev_config dev_config;
2157 	unsigned int dev_config_vf;
2158 	struct rte_eth_devargs eth_da = *req_eth_da;
2159 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2160 	struct mlx5_bond_info bond_info;
2161 	int ret = -1;
2162 
2163 	errno = 0;
2164 	ibv_list = mlx5_glue->get_device_list(&ret);
2165 	if (!ibv_list) {
2166 		rte_errno = errno ? errno : ENOSYS;
2167 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2168 		return -rte_errno;
2169 	}
2170 	/*
2171 	 * First scan the list of all Infiniband devices to find
2172 	 * matching ones, gathering into the list.
2173 	 */
2174 	struct ibv_device *ibv_match[ret + 1];
2175 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2176 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2177 	unsigned int i;
2178 
2179 	while (ret-- > 0) {
2180 		struct rte_pci_addr pci_addr;
2181 
2182 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2183 		bd = mlx5_device_bond_pci_match
2184 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2185 				 &bond_info);
2186 		if (bd >= 0) {
2187 			/*
2188 			 * Bonding device detected. Only one match is allowed,
2189 			 * the bonding is supported over multi-port IB device,
2190 			 * there should be no matches on representor PCI
2191 			 * functions or non VF LAG bonding devices with
2192 			 * specified address.
2193 			 */
2194 			if (nd) {
2195 				DRV_LOG(ERR,
2196 					"multiple PCI match on bonding device"
2197 					"\"%s\" found", ibv_list[ret]->name);
2198 				rte_errno = ENOENT;
2199 				ret = -rte_errno;
2200 				goto exit;
2201 			}
2202 			/* Amend owner pci address if owner PF ID specified. */
2203 			if (eth_da.nb_representor_ports)
2204 				owner_pci.function += owner_id;
2205 			DRV_LOG(INFO, "PCI information matches for"
2206 				      " slave %d bonding device \"%s\"",
2207 				      bd, ibv_list[ret]->name);
2208 			ibv_match[nd++] = ibv_list[ret];
2209 			break;
2210 		} else {
2211 			/* Bonding device not found. */
2212 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2213 					      &pci_addr))
2214 				continue;
2215 			if (owner_pci.domain != pci_addr.domain ||
2216 			    owner_pci.bus != pci_addr.bus ||
2217 			    owner_pci.devid != pci_addr.devid ||
2218 			    owner_pci.function != pci_addr.function)
2219 				continue;
2220 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2221 				ibv_list[ret]->name);
2222 			ibv_match[nd++] = ibv_list[ret];
2223 		}
2224 	}
2225 	ibv_match[nd] = NULL;
2226 	if (!nd) {
2227 		/* No device matches, just complain and bail out. */
2228 		DRV_LOG(WARNING,
2229 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2230 			" are kernel drivers loaded?",
2231 			owner_pci.domain, owner_pci.bus,
2232 			owner_pci.devid, owner_pci.function);
2233 		rte_errno = ENOENT;
2234 		ret = -rte_errno;
2235 		goto exit;
2236 	}
2237 	if (nd == 1) {
2238 		/*
2239 		 * Found single matching device may have multiple ports.
2240 		 * Each port may be representor, we have to check the port
2241 		 * number and check the representors existence.
2242 		 */
2243 		if (nl_rdma >= 0)
2244 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2245 		if (!np)
2246 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2247 					 " ports number", ibv_match[0]->name);
2248 		if (bd >= 0 && !np) {
2249 			DRV_LOG(ERR, "can not get ports"
2250 				     " for bonding device");
2251 			rte_errno = ENOENT;
2252 			ret = -rte_errno;
2253 			goto exit;
2254 		}
2255 	}
2256 	/*
2257 	 * Now we can determine the maximal
2258 	 * amount of devices to be spawned.
2259 	 */
2260 	list = mlx5_malloc(MLX5_MEM_ZERO,
2261 			   sizeof(struct mlx5_dev_spawn_data) *
2262 			   (np ? np : nd),
2263 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2264 	if (!list) {
2265 		DRV_LOG(ERR, "spawn data array allocation failure");
2266 		rte_errno = ENOMEM;
2267 		ret = -rte_errno;
2268 		goto exit;
2269 	}
2270 	if (bd >= 0 || np > 1) {
2271 		/*
2272 		 * Single IB device with multiple ports found,
2273 		 * it may be E-Switch master device and representors.
2274 		 * We have to perform identification through the ports.
2275 		 */
2276 		MLX5_ASSERT(nl_rdma >= 0);
2277 		MLX5_ASSERT(ns == 0);
2278 		MLX5_ASSERT(nd == 1);
2279 		MLX5_ASSERT(np);
2280 		for (i = 1; i <= np; ++i) {
2281 			list[ns].bond_info = &bond_info;
2282 			list[ns].max_port = np;
2283 			list[ns].phys_port = i;
2284 			list[ns].phys_dev = ibv_match[0];
2285 			list[ns].eth_dev = NULL;
2286 			list[ns].pci_dev = pci_dev;
2287 			list[ns].pf_bond = bd;
2288 			list[ns].ifindex = mlx5_nl_ifindex
2289 				(nl_rdma,
2290 				mlx5_os_get_dev_device_name
2291 						(list[ns].phys_dev), i);
2292 			if (!list[ns].ifindex) {
2293 				/*
2294 				 * No network interface index found for the
2295 				 * specified port, it means there is no
2296 				 * representor on this port. It's OK,
2297 				 * there can be disabled ports, for example
2298 				 * if sriov_numvfs < sriov_totalvfs.
2299 				 */
2300 				continue;
2301 			}
2302 			ret = -1;
2303 			if (nl_route >= 0)
2304 				ret = mlx5_nl_switch_info
2305 					       (nl_route,
2306 						list[ns].ifindex,
2307 						&list[ns].info);
2308 			if (ret || (!list[ns].info.representor &&
2309 				    !list[ns].info.master)) {
2310 				/*
2311 				 * We failed to recognize representors with
2312 				 * Netlink, let's try to perform the task
2313 				 * with sysfs.
2314 				 */
2315 				ret =  mlx5_sysfs_switch_info
2316 						(list[ns].ifindex,
2317 						 &list[ns].info);
2318 			}
2319 			if (!ret && bd >= 0) {
2320 				switch (list[ns].info.name_type) {
2321 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2322 					if (np == 1) {
2323 						/*
2324 						 * Force standalone bonding
2325 						 * device for ROCE LAG
2326 						 * confgiurations.
2327 						 */
2328 						list[ns].info.master = 0;
2329 						list[ns].info.representor = 0;
2330 					}
2331 					if (list[ns].info.port_name == bd)
2332 						ns++;
2333 					break;
2334 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2335 					/* Fallthrough */
2336 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2337 					/* Fallthrough */
2338 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2339 					if (list[ns].info.pf_num == bd)
2340 						ns++;
2341 					break;
2342 				default:
2343 					break;
2344 				}
2345 				continue;
2346 			}
2347 			if (!ret && (list[ns].info.representor ^
2348 				     list[ns].info.master))
2349 				ns++;
2350 		}
2351 		if (!ns) {
2352 			DRV_LOG(ERR,
2353 				"unable to recognize master/representors"
2354 				" on the IB device with multiple ports");
2355 			rte_errno = ENOENT;
2356 			ret = -rte_errno;
2357 			goto exit;
2358 		}
2359 	} else {
2360 		/*
2361 		 * The existence of several matching entries (nd > 1) means
2362 		 * port representors have been instantiated. No existing Verbs
2363 		 * call nor sysfs entries can tell them apart, this can only
2364 		 * be done through Netlink calls assuming kernel drivers are
2365 		 * recent enough to support them.
2366 		 *
2367 		 * In the event of identification failure through Netlink,
2368 		 * try again through sysfs, then:
2369 		 *
2370 		 * 1. A single IB device matches (nd == 1) with single
2371 		 *    port (np=0/1) and is not a representor, assume
2372 		 *    no switch support.
2373 		 *
2374 		 * 2. Otherwise no safe assumptions can be made;
2375 		 *    complain louder and bail out.
2376 		 */
2377 		for (i = 0; i != nd; ++i) {
2378 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2379 			list[ns].bond_info = NULL;
2380 			list[ns].max_port = 1;
2381 			list[ns].phys_port = 1;
2382 			list[ns].phys_dev = ibv_match[i];
2383 			list[ns].eth_dev = NULL;
2384 			list[ns].pci_dev = pci_dev;
2385 			list[ns].pf_bond = -1;
2386 			list[ns].ifindex = 0;
2387 			if (nl_rdma >= 0)
2388 				list[ns].ifindex = mlx5_nl_ifindex
2389 				(nl_rdma,
2390 				mlx5_os_get_dev_device_name
2391 						(list[ns].phys_dev), 1);
2392 			if (!list[ns].ifindex) {
2393 				char ifname[IF_NAMESIZE];
2394 
2395 				/*
2396 				 * Netlink failed, it may happen with old
2397 				 * ib_core kernel driver (before 4.16).
2398 				 * We can assume there is old driver because
2399 				 * here we are processing single ports IB
2400 				 * devices. Let's try sysfs to retrieve
2401 				 * the ifindex. The method works for
2402 				 * master device only.
2403 				 */
2404 				if (nd > 1) {
2405 					/*
2406 					 * Multiple devices found, assume
2407 					 * representors, can not distinguish
2408 					 * master/representor and retrieve
2409 					 * ifindex via sysfs.
2410 					 */
2411 					continue;
2412 				}
2413 				ret = mlx5_get_ifname_sysfs
2414 					(ibv_match[i]->ibdev_path, ifname);
2415 				if (!ret)
2416 					list[ns].ifindex =
2417 						if_nametoindex(ifname);
2418 				if (!list[ns].ifindex) {
2419 					/*
2420 					 * No network interface index found
2421 					 * for the specified device, it means
2422 					 * there it is neither representor
2423 					 * nor master.
2424 					 */
2425 					continue;
2426 				}
2427 			}
2428 			ret = -1;
2429 			if (nl_route >= 0)
2430 				ret = mlx5_nl_switch_info
2431 					       (nl_route,
2432 						list[ns].ifindex,
2433 						&list[ns].info);
2434 			if (ret || (!list[ns].info.representor &&
2435 				    !list[ns].info.master)) {
2436 				/*
2437 				 * We failed to recognize representors with
2438 				 * Netlink, let's try to perform the task
2439 				 * with sysfs.
2440 				 */
2441 				ret =  mlx5_sysfs_switch_info
2442 						(list[ns].ifindex,
2443 						 &list[ns].info);
2444 			}
2445 			if (!ret && (list[ns].info.representor ^
2446 				     list[ns].info.master)) {
2447 				ns++;
2448 			} else if ((nd == 1) &&
2449 				   !list[ns].info.representor &&
2450 				   !list[ns].info.master) {
2451 				/*
2452 				 * Single IB device with
2453 				 * one physical port and
2454 				 * attached network device.
2455 				 * May be SRIOV is not enabled
2456 				 * or there is no representors.
2457 				 */
2458 				DRV_LOG(INFO, "no E-Switch support detected");
2459 				ns++;
2460 				break;
2461 			}
2462 		}
2463 		if (!ns) {
2464 			DRV_LOG(ERR,
2465 				"unable to recognize master/representors"
2466 				" on the multiple IB devices");
2467 			rte_errno = ENOENT;
2468 			ret = -rte_errno;
2469 			goto exit;
2470 		}
2471 		/*
2472 		 * New kernels may add the switch_id attribute for the case
2473 		 * there is no E-Switch and we wrongly recognized the
2474 		 * only device as master. Override this if there is the
2475 		 * single device with single port and new device name
2476 		 * format present.
2477 		 */
2478 		if (nd == 1 &&
2479 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2480 			list[0].info.master = 0;
2481 			list[0].info.representor = 0;
2482 		}
2483 	}
2484 	MLX5_ASSERT(ns);
2485 	/*
2486 	 * Sort list to probe devices in natural order for users convenience
2487 	 * (i.e. master first, then representors from lowest to highest ID).
2488 	 */
2489 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2490 	/* Device specific configuration. */
2491 	switch (pci_dev->id.device_id) {
2492 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2493 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2494 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2495 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2496 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2497 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2498 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2499 		dev_config_vf = 1;
2500 		break;
2501 	default:
2502 		dev_config_vf = 0;
2503 		break;
2504 	}
2505 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2506 		/* Set devargs default values. */
2507 		if (eth_da.nb_mh_controllers == 0) {
2508 			eth_da.nb_mh_controllers = 1;
2509 			eth_da.mh_controllers[0] = 0;
2510 		}
2511 		if (eth_da.nb_ports == 0 && ns > 0) {
2512 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2513 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2514 					pci_dev->device.devargs->args);
2515 			eth_da.nb_ports = 1;
2516 			eth_da.ports[0] = list[0].info.pf_num;
2517 		}
2518 		if (eth_da.nb_representor_ports == 0) {
2519 			eth_da.nb_representor_ports = 1;
2520 			eth_da.representor_ports[0] = 0;
2521 		}
2522 	}
2523 	for (i = 0; i != ns; ++i) {
2524 		uint32_t restore;
2525 
2526 		/* Default configuration. */
2527 		mlx5_os_config_default(&dev_config);
2528 		dev_config.vf = dev_config_vf;
2529 		dev_config.allow_duplicate_pattern = 1;
2530 		list[i].numa_node = pci_dev->device.numa_node;
2531 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2532 						 &list[i],
2533 						 &dev_config,
2534 						 &eth_da);
2535 		if (!list[i].eth_dev) {
2536 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2537 				break;
2538 			/* Device is disabled or already spawned. Ignore it. */
2539 			continue;
2540 		}
2541 		restore = list[i].eth_dev->data->dev_flags;
2542 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2543 		/**
2544 		 * Each representor has a dedicated interrupts vector.
2545 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2546 		 * representor eth_dev object because representor and PF
2547 		 * share the same PCI address.
2548 		 * Override representor device with a dedicated
2549 		 * interrupts handle here.
2550 		 * Representor interrupts handle is released in mlx5_dev_stop().
2551 		 */
2552 		if (list[i].info.representor) {
2553 			struct rte_intr_handle *intr_handle;
2554 			intr_handle = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
2555 						  sizeof(*intr_handle), 0,
2556 						  SOCKET_ID_ANY);
2557 			if (!intr_handle) {
2558 				DRV_LOG(ERR,
2559 					"port %u failed to allocate memory for interrupt handler "
2560 					"Rx interrupts will not be supported",
2561 					i);
2562 				rte_errno = ENOMEM;
2563 				ret = -rte_errno;
2564 				goto exit;
2565 			}
2566 			list[i].eth_dev->intr_handle = intr_handle;
2567 		}
2568 		/* Restore non-PCI flags cleared by the above call. */
2569 		list[i].eth_dev->data->dev_flags |= restore;
2570 		rte_eth_dev_probing_finish(list[i].eth_dev);
2571 	}
2572 	if (i != ns) {
2573 		DRV_LOG(ERR,
2574 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2575 			" encountering an error: %s",
2576 			owner_pci.domain, owner_pci.bus,
2577 			owner_pci.devid, owner_pci.function,
2578 			strerror(rte_errno));
2579 		ret = -rte_errno;
2580 		/* Roll back. */
2581 		while (i--) {
2582 			if (!list[i].eth_dev)
2583 				continue;
2584 			mlx5_dev_close(list[i].eth_dev);
2585 			/* mac_addrs must not be freed because in dev_private */
2586 			list[i].eth_dev->data->mac_addrs = NULL;
2587 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2588 		}
2589 		/* Restore original error. */
2590 		rte_errno = -ret;
2591 	} else {
2592 		ret = 0;
2593 	}
2594 exit:
2595 	/*
2596 	 * Do the routine cleanup:
2597 	 * - close opened Netlink sockets
2598 	 * - free allocated spawn data array
2599 	 * - free the Infiniband device list
2600 	 */
2601 	if (nl_rdma >= 0)
2602 		close(nl_rdma);
2603 	if (nl_route >= 0)
2604 		close(nl_route);
2605 	if (list)
2606 		mlx5_free(list);
2607 	MLX5_ASSERT(ibv_list);
2608 	mlx5_glue->free_device_list(ibv_list);
2609 	return ret;
2610 }
2611 
2612 static int
2613 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2614 			  struct rte_eth_devargs *eth_da)
2615 {
2616 	int ret = 0;
2617 
2618 	if (dev->devargs == NULL)
2619 		return 0;
2620 	memset(eth_da, 0, sizeof(*eth_da));
2621 	/* Parse representor information first from class argument. */
2622 	if (dev->devargs->cls_str)
2623 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2624 	if (ret != 0) {
2625 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2626 			dev->devargs->cls_str);
2627 		return -rte_errno;
2628 	}
2629 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2630 		/* Parse legacy device argument */
2631 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2632 		if (ret) {
2633 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2634 				dev->devargs->args);
2635 			return -rte_errno;
2636 		}
2637 	}
2638 	return 0;
2639 }
2640 
2641 /**
2642  * Callback to register a PCI device.
2643  *
2644  * This function spawns Ethernet devices out of a given PCI device.
2645  *
2646  * @param[in] pci_dev
2647  *   PCI device information.
2648  *
2649  * @return
2650  *   0 on success, a negative errno value otherwise and rte_errno is set.
2651  */
2652 static int
2653 mlx5_os_pci_probe(struct rte_pci_device *pci_dev)
2654 {
2655 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2656 	int ret = 0;
2657 	uint16_t p;
2658 
2659 	ret = mlx5_os_parse_eth_devargs(&pci_dev->device, &eth_da);
2660 	if (ret != 0)
2661 		return ret;
2662 
2663 	if (eth_da.nb_ports > 0) {
2664 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2665 		for (p = 0; p < eth_da.nb_ports; p++)
2666 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2667 						   eth_da.ports[p]);
2668 	} else {
2669 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2670 	}
2671 	return ret;
2672 }
2673 
2674 /* Probe a single SF device on auxiliary bus, no representor support. */
2675 static int
2676 mlx5_os_auxiliary_probe(struct rte_device *dev)
2677 {
2678 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2679 	struct mlx5_dev_config config;
2680 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2681 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2682 	struct rte_eth_dev *eth_dev;
2683 	int ret = 0;
2684 
2685 	/* Parse ethdev devargs. */
2686 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2687 	if (ret != 0)
2688 		return ret;
2689 	/* Set default config data. */
2690 	mlx5_os_config_default(&config);
2691 	config.sf = 1;
2692 	/* Init spawn data. */
2693 	spawn.max_port = 1;
2694 	spawn.phys_port = 1;
2695 	spawn.phys_dev = mlx5_os_get_ibv_dev(dev);
2696 	if (spawn.phys_dev == NULL)
2697 		return -rte_errno;
2698 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2699 	if (ret < 0) {
2700 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2701 		return ret;
2702 	}
2703 	spawn.ifindex = ret;
2704 	spawn.numa_node = dev->numa_node;
2705 	/* Spawn device. */
2706 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2707 	if (eth_dev == NULL)
2708 		return -rte_errno;
2709 	/* Post create. */
2710 	eth_dev->intr_handle = &adev->intr_handle;
2711 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2712 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2713 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2714 		eth_dev->data->numa_node = dev->numa_node;
2715 	}
2716 	rte_eth_dev_probing_finish(eth_dev);
2717 	return 0;
2718 }
2719 
2720 /**
2721  * Net class driver callback to probe a device.
2722  *
2723  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2724  *
2725  * @param[in] dev
2726  *   Pointer to the generic device.
2727  *
2728  * @return
2729  *   0 on success, the function cannot fail.
2730  */
2731 int
2732 mlx5_os_net_probe(struct rte_device *dev)
2733 {
2734 	int ret;
2735 
2736 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2737 		mlx5_pmd_socket_init();
2738 	ret = mlx5_init_once();
2739 	if (ret) {
2740 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2741 			strerror(rte_errno));
2742 		return -rte_errno;
2743 	}
2744 	if (mlx5_dev_is_pci(dev))
2745 		return mlx5_os_pci_probe(RTE_DEV_TO_PCI(dev));
2746 	else
2747 		return mlx5_os_auxiliary_probe(dev);
2748 }
2749 
2750 static int
2751 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2752 {
2753 	char *env;
2754 	int value;
2755 
2756 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2757 	/* Get environment variable to store. */
2758 	env = getenv(MLX5_SHUT_UP_BF);
2759 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2760 	if (config->dbnc == MLX5_ARG_UNSET)
2761 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2762 	else
2763 		setenv(MLX5_SHUT_UP_BF,
2764 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2765 	return value;
2766 }
2767 
2768 static void
2769 mlx5_restore_doorbell_mapping_env(int value)
2770 {
2771 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2772 	/* Restore the original environment variable state. */
2773 	if (value == MLX5_ARG_UNSET)
2774 		unsetenv(MLX5_SHUT_UP_BF);
2775 	else
2776 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2777 }
2778 
2779 /**
2780  * Extract pdn of PD object using DV API.
2781  *
2782  * @param[in] pd
2783  *   Pointer to the verbs PD object.
2784  * @param[out] pdn
2785  *   Pointer to the PD object number variable.
2786  *
2787  * @return
2788  *   0 on success, error value otherwise.
2789  */
2790 int
2791 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2792 {
2793 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2794 	struct mlx5dv_obj obj;
2795 	struct mlx5dv_pd pd_info;
2796 	int ret = 0;
2797 
2798 	obj.pd.in = pd;
2799 	obj.pd.out = &pd_info;
2800 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2801 	if (ret) {
2802 		DRV_LOG(DEBUG, "Fail to get PD object info");
2803 		return ret;
2804 	}
2805 	*pdn = pd_info.pdn;
2806 	return 0;
2807 #else
2808 	(void)pd;
2809 	(void)pdn;
2810 	return -ENOTSUP;
2811 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2812 }
2813 
2814 /**
2815  * Function API to open IB device.
2816  *
2817  * This function calls the Linux glue APIs to open a device.
2818  *
2819  * @param[in] spawn
2820  *   Pointer to the IB device attributes (name, port, etc).
2821  * @param[out] config
2822  *   Pointer to device configuration structure.
2823  * @param[out] sh
2824  *   Pointer to shared context structure.
2825  *
2826  * @return
2827  *   0 on success, a positive error value otherwise.
2828  */
2829 int
2830 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2831 		     const struct mlx5_dev_config *config,
2832 		     struct mlx5_dev_ctx_shared *sh)
2833 {
2834 	int dbmap_env;
2835 	int err = 0;
2836 
2837 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2838 	/*
2839 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2840 	 * before the device creation. The rdma_core library
2841 	 * checks the variable at device creation and
2842 	 * stores the result internally.
2843 	 */
2844 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2845 	/* Try to open IB device with DV first, then usual Verbs. */
2846 	errno = 0;
2847 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2848 	if (sh->ctx) {
2849 		sh->devx = 1;
2850 		DRV_LOG(DEBUG, "DevX is supported");
2851 		/* The device is created, no need for environment. */
2852 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2853 	} else {
2854 		/* The environment variable is still configured. */
2855 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2856 		err = errno ? errno : ENODEV;
2857 		/*
2858 		 * The environment variable is not needed anymore,
2859 		 * all device creation attempts are completed.
2860 		 */
2861 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2862 		if (!sh->ctx)
2863 			return err;
2864 		DRV_LOG(DEBUG, "DevX is NOT supported");
2865 		err = 0;
2866 	}
2867 	if (!err && sh->ctx) {
2868 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2869 		mlx5_glue->dv_set_context_attr(sh->ctx,
2870 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2871 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2872 				.alloc = &mlx5_alloc_verbs_buf,
2873 				.free = &mlx5_free_verbs_buf,
2874 				.data = sh,
2875 			}));
2876 	}
2877 	return err;
2878 }
2879 
2880 /**
2881  * Install shared asynchronous device events handler.
2882  * This function is implemented to support event sharing
2883  * between multiple ports of single IB device.
2884  *
2885  * @param sh
2886  *   Pointer to mlx5_dev_ctx_shared object.
2887  */
2888 void
2889 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2890 {
2891 	int ret;
2892 	int flags;
2893 
2894 	sh->intr_handle.fd = -1;
2895 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2896 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2897 		    F_SETFL, flags | O_NONBLOCK);
2898 	if (ret) {
2899 		DRV_LOG(INFO, "failed to change file descriptor async event"
2900 			" queue");
2901 	} else {
2902 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2903 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2904 		if (rte_intr_callback_register(&sh->intr_handle,
2905 					mlx5_dev_interrupt_handler, sh)) {
2906 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2907 			sh->intr_handle.fd = -1;
2908 		}
2909 	}
2910 	if (sh->devx) {
2911 #ifdef HAVE_IBV_DEVX_ASYNC
2912 		sh->intr_handle_devx.fd = -1;
2913 		sh->devx_comp =
2914 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2915 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2916 		if (!devx_comp) {
2917 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2918 			return;
2919 		}
2920 		flags = fcntl(devx_comp->fd, F_GETFL);
2921 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2922 		if (ret) {
2923 			DRV_LOG(INFO, "failed to change file descriptor"
2924 				" devx comp");
2925 			return;
2926 		}
2927 		sh->intr_handle_devx.fd = devx_comp->fd;
2928 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2929 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2930 					mlx5_dev_interrupt_handler_devx, sh)) {
2931 			DRV_LOG(INFO, "Fail to install the devx shared"
2932 				" interrupt.");
2933 			sh->intr_handle_devx.fd = -1;
2934 		}
2935 #endif /* HAVE_IBV_DEVX_ASYNC */
2936 	}
2937 }
2938 
2939 /**
2940  * Uninstall 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 dev
2945  *   Pointer to mlx5_dev_ctx_shared object.
2946  */
2947 void
2948 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2949 {
2950 	if (sh->intr_handle.fd >= 0)
2951 		mlx5_intr_callback_unregister(&sh->intr_handle,
2952 					      mlx5_dev_interrupt_handler, sh);
2953 #ifdef HAVE_IBV_DEVX_ASYNC
2954 	if (sh->intr_handle_devx.fd >= 0)
2955 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2956 				  mlx5_dev_interrupt_handler_devx, sh);
2957 	if (sh->devx_comp)
2958 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2959 #endif
2960 }
2961 
2962 /**
2963  * Read statistics by a named counter.
2964  *
2965  * @param[in] priv
2966  *   Pointer to the private device data structure.
2967  * @param[in] ctr_name
2968  *   Pointer to the name of the statistic counter to read
2969  * @param[out] stat
2970  *   Pointer to read statistic value.
2971  * @return
2972  *   0 on success and stat is valud, 1 if failed to read the value
2973  *   rte_errno is set.
2974  *
2975  */
2976 int
2977 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2978 		      uint64_t *stat)
2979 {
2980 	int fd;
2981 
2982 	if (priv->sh) {
2983 		if (priv->q_counters != NULL &&
2984 		    strcmp(ctr_name, "out_of_buffer") == 0)
2985 			return mlx5_devx_cmd_queue_counter_query
2986 					(priv->q_counters, 0, (uint32_t *)stat);
2987 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2988 		      priv->sh->ibdev_path,
2989 		      priv->dev_port,
2990 		      ctr_name);
2991 		fd = open(path, O_RDONLY);
2992 		/*
2993 		 * in switchdev the file location is not per port
2994 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2995 		 */
2996 		if (fd == -1) {
2997 			MKSTR(path1, "%s/hw_counters/%s",
2998 			      priv->sh->ibdev_path,
2999 			      ctr_name);
3000 			fd = open(path1, O_RDONLY);
3001 		}
3002 		if (fd != -1) {
3003 			char buf[21] = {'\0'};
3004 			ssize_t n = read(fd, buf, sizeof(buf));
3005 
3006 			close(fd);
3007 			if (n != -1) {
3008 				*stat = strtoull(buf, NULL, 10);
3009 				return 0;
3010 			}
3011 		}
3012 	}
3013 	*stat = 0;
3014 	return 1;
3015 }
3016 
3017 /**
3018  * Set the reg_mr and dereg_mr call backs
3019  *
3020  * @param reg_mr_cb[out]
3021  *   Pointer to reg_mr func
3022  * @param dereg_mr_cb[out]
3023  *   Pointer to dereg_mr func
3024  *
3025  */
3026 void
3027 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
3028 		      mlx5_dereg_mr_t *dereg_mr_cb)
3029 {
3030 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
3031 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
3032 }
3033 
3034 /**
3035  * Remove a MAC address from device
3036  *
3037  * @param dev
3038  *   Pointer to Ethernet device structure.
3039  * @param index
3040  *   MAC address index.
3041  */
3042 void
3043 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3044 {
3045 	struct mlx5_priv *priv = dev->data->dev_private;
3046 	const int vf = priv->config.vf;
3047 
3048 	if (vf)
3049 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
3050 					mlx5_ifindex(dev), priv->mac_own,
3051 					&dev->data->mac_addrs[index], index);
3052 }
3053 
3054 /**
3055  * Adds a MAC address to the device
3056  *
3057  * @param dev
3058  *   Pointer to Ethernet device structure.
3059  * @param mac_addr
3060  *   MAC address to register.
3061  * @param index
3062  *   MAC address index.
3063  *
3064  * @return
3065  *   0 on success, a negative errno value otherwise
3066  */
3067 int
3068 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
3069 		     uint32_t index)
3070 {
3071 	struct mlx5_priv *priv = dev->data->dev_private;
3072 	const int vf = priv->config.vf;
3073 	int ret = 0;
3074 
3075 	if (vf)
3076 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
3077 					   mlx5_ifindex(dev), priv->mac_own,
3078 					   mac, index);
3079 	return ret;
3080 }
3081 
3082 /**
3083  * Modify a VF MAC address
3084  *
3085  * @param priv
3086  *   Pointer to device private data.
3087  * @param mac_addr
3088  *   MAC address to modify into.
3089  * @param iface_idx
3090  *   Net device interface index
3091  * @param vf_index
3092  *   VF index
3093  *
3094  * @return
3095  *   0 on success, a negative errno value otherwise
3096  */
3097 int
3098 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
3099 			   unsigned int iface_idx,
3100 			   struct rte_ether_addr *mac_addr,
3101 			   int vf_index)
3102 {
3103 	return mlx5_nl_vf_mac_addr_modify
3104 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
3105 }
3106 
3107 /**
3108  * Set device promiscuous mode
3109  *
3110  * @param dev
3111  *   Pointer to Ethernet device structure.
3112  * @param enable
3113  *   0 - promiscuous is disabled, otherwise - enabled
3114  *
3115  * @return
3116  *   0 on success, a negative error value otherwise
3117  */
3118 int
3119 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
3120 {
3121 	struct mlx5_priv *priv = dev->data->dev_private;
3122 
3123 	return mlx5_nl_promisc(priv->nl_socket_route,
3124 			       mlx5_ifindex(dev), !!enable);
3125 }
3126 
3127 /**
3128  * Set device promiscuous mode
3129  *
3130  * @param dev
3131  *   Pointer to Ethernet device structure.
3132  * @param enable
3133  *   0 - all multicase is disabled, otherwise - enabled
3134  *
3135  * @return
3136  *   0 on success, a negative error value otherwise
3137  */
3138 int
3139 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
3140 {
3141 	struct mlx5_priv *priv = dev->data->dev_private;
3142 
3143 	return mlx5_nl_allmulti(priv->nl_socket_route,
3144 				mlx5_ifindex(dev), !!enable);
3145 }
3146 
3147 /**
3148  * Flush device MAC addresses
3149  *
3150  * @param dev
3151  *   Pointer to Ethernet device structure.
3152  *
3153  */
3154 void
3155 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3156 {
3157 	struct mlx5_priv *priv = dev->data->dev_private;
3158 
3159 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3160 			       dev->data->mac_addrs,
3161 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3162 }
3163