xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 9f1d636f3ef08fd69b40cfe6eb954b0c9d1820d2)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2020 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 #include <net/if.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/sockios.h>
15 #include <linux/ethtool.h>
16 #include <fcntl.h>
17 
18 #include <rte_malloc.h>
19 #include <ethdev_driver.h>
20 #include <ethdev_pci.h>
21 #include <rte_pci.h>
22 #include <rte_bus_pci.h>
23 #include <rte_bus_auxiliary.h>
24 #include <rte_common.h>
25 #include <rte_kvargs.h>
26 #include <rte_rwlock.h>
27 #include <rte_spinlock.h>
28 #include <rte_string_fns.h>
29 #include <rte_alarm.h>
30 #include <rte_eal_paging.h>
31 
32 #include <mlx5_glue.h>
33 #include <mlx5_devx_cmds.h>
34 #include <mlx5_common.h>
35 #include <mlx5_common_mp.h>
36 #include <mlx5_common_mr.h>
37 #include <mlx5_malloc.h>
38 
39 #include "mlx5_defs.h"
40 #include "mlx5.h"
41 #include "mlx5_common_os.h"
42 #include "mlx5_utils.h"
43 #include "mlx5_rxtx.h"
44 #include "mlx5_rx.h"
45 #include "mlx5_tx.h"
46 #include "mlx5_autoconf.h"
47 #include "mlx5_flow.h"
48 #include "rte_pmd_mlx5.h"
49 #include "mlx5_verbs.h"
50 #include "mlx5_nl.h"
51 #include "mlx5_devx.h"
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /* rte flow indexed pool configuration. */
71 static struct mlx5_indexed_pool_config icfg[] = {
72 	{
73 		.size = sizeof(struct rte_flow),
74 		.trunk_size = 64,
75 		.need_lock = 1,
76 		.release_mem_en = 0,
77 		.malloc = mlx5_malloc,
78 		.free = mlx5_free,
79 		.per_core_cache = 0,
80 		.type = "ctl_flow_ipool",
81 	},
82 	{
83 		.size = sizeof(struct rte_flow),
84 		.trunk_size = 64,
85 		.grow_trunk = 3,
86 		.grow_shift = 2,
87 		.need_lock = 1,
88 		.release_mem_en = 0,
89 		.malloc = mlx5_malloc,
90 		.free = mlx5_free,
91 		.per_core_cache = 1 << 14,
92 		.type = "rte_flow_ipool",
93 	},
94 	{
95 		.size = sizeof(struct rte_flow),
96 		.trunk_size = 64,
97 		.grow_trunk = 3,
98 		.grow_shift = 2,
99 		.need_lock = 1,
100 		.release_mem_en = 0,
101 		.malloc = mlx5_malloc,
102 		.free = mlx5_free,
103 		.per_core_cache = 0,
104 		.type = "mcp_flow_ipool",
105 	},
106 };
107 
108 /**
109  * Set the completion channel file descriptor interrupt as non-blocking.
110  *
111  * @param[in] rxq_obj
112  *   Pointer to RQ channel object, which includes the channel fd
113  *
114  * @param[out] fd
115  *   The file descriptor (representing the intetrrupt) used in this channel.
116  *
117  * @return
118  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
119  */
120 int
121 mlx5_os_set_nonblock_channel_fd(int fd)
122 {
123 	int flags;
124 
125 	flags = fcntl(fd, F_GETFL);
126 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
127 }
128 
129 /**
130  * Get mlx5 device attributes. The glue function query_device_ex() is called
131  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
132  * device attributes from the glue out parameter.
133  *
134  * @param cdev
135  *   Pointer to mlx5 device.
136  *
137  * @param device_attr
138  *   Pointer to mlx5 device attributes.
139  *
140  * @return
141  *   0 on success, non zero error number otherwise
142  */
143 int
144 mlx5_os_get_dev_attr(struct mlx5_common_device *cdev,
145 		     struct mlx5_dev_attr *device_attr)
146 {
147 	int err;
148 	struct ibv_context *ctx = cdev->ctx;
149 	struct ibv_device_attr_ex attr_ex;
150 
151 	memset(device_attr, 0, sizeof(*device_attr));
152 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
153 	if (err)
154 		return err;
155 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
156 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
157 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
158 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
159 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
160 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
161 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
162 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
163 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
164 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
165 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
166 	device_attr->max_rwq_indirection_table_size =
167 		attr_ex.rss_caps.max_rwq_indirection_table_size;
168 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
169 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
170 
171 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
172 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
173 	if (err)
174 		return err;
175 
176 	device_attr->flags = dv_attr.flags;
177 	device_attr->comp_mask = dv_attr.comp_mask;
178 #ifdef HAVE_IBV_MLX5_MOD_SWP
179 	device_attr->sw_parsing_offloads =
180 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
181 #endif
182 	device_attr->min_single_stride_log_num_of_bytes =
183 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
184 	device_attr->max_single_stride_log_num_of_bytes =
185 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
186 	device_attr->min_single_wqe_log_num_of_strides =
187 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
188 	device_attr->max_single_wqe_log_num_of_strides =
189 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
190 	device_attr->stride_supported_qpts =
191 		dv_attr.striding_rq_caps.supported_qpts;
192 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
193 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
194 #endif
195 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
196 		sizeof(device_attr->fw_ver));
197 
198 	return err;
199 }
200 
201 /**
202  * Detect misc5 support or not
203  *
204  * @param[in] priv
205  *   Device private data pointer
206  */
207 #ifdef HAVE_MLX5DV_DR
208 static void
209 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
210 {
211 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
212 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
213 	 * Case: IPv4--->UDP--->VxLAN--->vni
214 	 */
215 	void *tbl;
216 	struct mlx5_flow_dv_match_params matcher_mask;
217 	void *match_m;
218 	void *matcher;
219 	void *headers_m;
220 	void *misc5_m;
221 	uint32_t *tunnel_header_m;
222 	struct mlx5dv_flow_matcher_attr dv_attr;
223 
224 	memset(&matcher_mask, 0, sizeof(matcher_mask));
225 	matcher_mask.size = sizeof(matcher_mask.buf);
226 	match_m = matcher_mask.buf;
227 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
228 	misc5_m = MLX5_ADDR_OF(fte_match_param,
229 			       match_m, misc_parameters_5);
230 	tunnel_header_m = (uint32_t *)
231 				MLX5_ADDR_OF(fte_match_set_misc5,
232 				misc5_m, tunnel_header_1);
233 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
234 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
235 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
236 	*tunnel_header_m = 0xffffff;
237 
238 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
239 	if (!tbl) {
240 		DRV_LOG(INFO, "No SW steering support");
241 		return;
242 	}
243 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
244 	dv_attr.match_mask = (void *)&matcher_mask,
245 	dv_attr.match_criteria_enable =
246 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
247 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
248 	dv_attr.priority = 3;
249 #ifdef HAVE_MLX5DV_DR_ESWITCH
250 	void *misc2_m;
251 	if (priv->config.dv_esw_en) {
252 		/* FDB enabled reg_c_0 */
253 		dv_attr.match_criteria_enable |=
254 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
255 		misc2_m = MLX5_ADDR_OF(fte_match_param,
256 				       match_m, misc_parameters_2);
257 		MLX5_SET(fte_match_set_misc2, misc2_m,
258 			 metadata_reg_c_0, 0xffff);
259 	}
260 #endif
261 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->cdev->ctx,
262 						    &dv_attr, tbl);
263 	if (matcher) {
264 		priv->sh->misc5_cap = 1;
265 		mlx5_glue->dv_destroy_flow_matcher(matcher);
266 	}
267 	mlx5_glue->dr_destroy_flow_tbl(tbl);
268 #else
269 	RTE_SET_USED(priv);
270 #endif
271 }
272 #endif
273 
274 /**
275  * Initialize DR related data within private structure.
276  * Routine checks the reference counter and does actual
277  * resources creation/initialization only if counter is zero.
278  *
279  * @param[in] priv
280  *   Pointer to the private device data structure.
281  *
282  * @return
283  *   Zero on success, positive error code otherwise.
284  */
285 static int
286 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
287 {
288 	struct mlx5_dev_ctx_shared *sh = priv->sh;
289 	char s[MLX5_NAME_SIZE] __rte_unused;
290 	int err;
291 
292 	MLX5_ASSERT(sh && sh->refcnt);
293 	if (sh->refcnt > 1)
294 		return 0;
295 	err = mlx5_alloc_table_hash_list(priv);
296 	if (err)
297 		goto error;
298 	/* The resources below are only valid with DV support. */
299 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
300 	/* Init port id action list. */
301 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
302 	sh->port_id_action_list = mlx5_list_create(s, sh, true,
303 						   flow_dv_port_id_create_cb,
304 						   flow_dv_port_id_match_cb,
305 						   flow_dv_port_id_remove_cb,
306 						   flow_dv_port_id_clone_cb,
307 						 flow_dv_port_id_clone_free_cb);
308 	if (!sh->port_id_action_list)
309 		goto error;
310 	/* Init push vlan action list. */
311 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
312 	sh->push_vlan_action_list = mlx5_list_create(s, sh, true,
313 						    flow_dv_push_vlan_create_cb,
314 						    flow_dv_push_vlan_match_cb,
315 						    flow_dv_push_vlan_remove_cb,
316 						    flow_dv_push_vlan_clone_cb,
317 					       flow_dv_push_vlan_clone_free_cb);
318 	if (!sh->push_vlan_action_list)
319 		goto error;
320 	/* Init sample action list. */
321 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
322 	sh->sample_action_list = mlx5_list_create(s, sh, true,
323 						  flow_dv_sample_create_cb,
324 						  flow_dv_sample_match_cb,
325 						  flow_dv_sample_remove_cb,
326 						  flow_dv_sample_clone_cb,
327 						  flow_dv_sample_clone_free_cb);
328 	if (!sh->sample_action_list)
329 		goto error;
330 	/* Init dest array action list. */
331 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
332 	sh->dest_array_list = mlx5_list_create(s, sh, true,
333 					       flow_dv_dest_array_create_cb,
334 					       flow_dv_dest_array_match_cb,
335 					       flow_dv_dest_array_remove_cb,
336 					       flow_dv_dest_array_clone_cb,
337 					      flow_dv_dest_array_clone_free_cb);
338 	if (!sh->dest_array_list)
339 		goto error;
340 #endif
341 #ifdef HAVE_MLX5DV_DR
342 	void *domain;
343 
344 	/* Reference counter is zero, we should initialize structures. */
345 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
346 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
347 	if (!domain) {
348 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
349 		err = errno;
350 		goto error;
351 	}
352 	sh->rx_domain = domain;
353 	domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
354 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
355 	if (!domain) {
356 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
357 		err = errno;
358 		goto error;
359 	}
360 	sh->tx_domain = domain;
361 #ifdef HAVE_MLX5DV_DR_ESWITCH
362 	if (priv->config.dv_esw_en) {
363 		domain = mlx5_glue->dr_create_domain(sh->cdev->ctx,
364 						     MLX5DV_DR_DOMAIN_TYPE_FDB);
365 		if (!domain) {
366 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
367 			err = errno;
368 			goto error;
369 		}
370 		sh->fdb_domain = domain;
371 	}
372 	/*
373 	 * The drop action is just some dummy placeholder in rdma-core. It
374 	 * does not belong to domains and has no any attributes, and, can be
375 	 * shared by the entire device.
376 	 */
377 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
378 	if (!sh->dr_drop_action) {
379 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
380 		err = errno;
381 		goto error;
382 	}
383 #endif
384 	if (!sh->tunnel_hub && priv->config.dv_miss_info)
385 		err = mlx5_alloc_tunnel_hub(sh);
386 	if (err) {
387 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
388 		goto error;
389 	}
390 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
391 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
392 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
393 		if (sh->fdb_domain)
394 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
395 	}
396 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
397 	if (!priv->config.allow_duplicate_pattern) {
398 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
399 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
400 #endif
401 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
402 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
403 		if (sh->fdb_domain)
404 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
405 	}
406 
407 	__mlx5_discovery_misc5_cap(priv);
408 #endif /* HAVE_MLX5DV_DR */
409 	sh->default_miss_action =
410 			mlx5_glue->dr_create_flow_action_default_miss();
411 	if (!sh->default_miss_action)
412 		DRV_LOG(WARNING, "Default miss action is not supported.");
413 	return 0;
414 error:
415 	/* Rollback the created objects. */
416 	if (sh->rx_domain) {
417 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
418 		sh->rx_domain = NULL;
419 	}
420 	if (sh->tx_domain) {
421 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
422 		sh->tx_domain = NULL;
423 	}
424 	if (sh->fdb_domain) {
425 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
426 		sh->fdb_domain = NULL;
427 	}
428 	if (sh->dr_drop_action) {
429 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
430 		sh->dr_drop_action = NULL;
431 	}
432 	if (sh->pop_vlan_action) {
433 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
434 		sh->pop_vlan_action = NULL;
435 	}
436 	if (sh->encaps_decaps) {
437 		mlx5_hlist_destroy(sh->encaps_decaps);
438 		sh->encaps_decaps = NULL;
439 	}
440 	if (sh->modify_cmds) {
441 		mlx5_hlist_destroy(sh->modify_cmds);
442 		sh->modify_cmds = NULL;
443 	}
444 	if (sh->tag_table) {
445 		/* tags should be destroyed with flow before. */
446 		mlx5_hlist_destroy(sh->tag_table);
447 		sh->tag_table = NULL;
448 	}
449 	if (sh->tunnel_hub) {
450 		mlx5_release_tunnel_hub(sh, priv->dev_port);
451 		sh->tunnel_hub = NULL;
452 	}
453 	mlx5_free_table_hash_list(priv);
454 	if (sh->port_id_action_list) {
455 		mlx5_list_destroy(sh->port_id_action_list);
456 		sh->port_id_action_list = NULL;
457 	}
458 	if (sh->push_vlan_action_list) {
459 		mlx5_list_destroy(sh->push_vlan_action_list);
460 		sh->push_vlan_action_list = NULL;
461 	}
462 	if (sh->sample_action_list) {
463 		mlx5_list_destroy(sh->sample_action_list);
464 		sh->sample_action_list = NULL;
465 	}
466 	if (sh->dest_array_list) {
467 		mlx5_list_destroy(sh->dest_array_list);
468 		sh->dest_array_list = NULL;
469 	}
470 	return err;
471 }
472 
473 /**
474  * Destroy DR related data within private structure.
475  *
476  * @param[in] priv
477  *   Pointer to the private device data structure.
478  */
479 void
480 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
481 {
482 	struct mlx5_dev_ctx_shared *sh = priv->sh;
483 
484 	MLX5_ASSERT(sh && sh->refcnt);
485 	if (sh->refcnt > 1)
486 		return;
487 #ifdef HAVE_MLX5DV_DR
488 	if (sh->rx_domain) {
489 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
490 		sh->rx_domain = NULL;
491 	}
492 	if (sh->tx_domain) {
493 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
494 		sh->tx_domain = NULL;
495 	}
496 #ifdef HAVE_MLX5DV_DR_ESWITCH
497 	if (sh->fdb_domain) {
498 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
499 		sh->fdb_domain = NULL;
500 	}
501 	if (sh->dr_drop_action) {
502 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
503 		sh->dr_drop_action = NULL;
504 	}
505 #endif
506 	if (sh->pop_vlan_action) {
507 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
508 		sh->pop_vlan_action = NULL;
509 	}
510 #endif /* HAVE_MLX5DV_DR */
511 	if (sh->default_miss_action)
512 		mlx5_glue->destroy_flow_action
513 				(sh->default_miss_action);
514 	if (sh->encaps_decaps) {
515 		mlx5_hlist_destroy(sh->encaps_decaps);
516 		sh->encaps_decaps = NULL;
517 	}
518 	if (sh->modify_cmds) {
519 		mlx5_hlist_destroy(sh->modify_cmds);
520 		sh->modify_cmds = NULL;
521 	}
522 	if (sh->tag_table) {
523 		/* tags should be destroyed with flow before. */
524 		mlx5_hlist_destroy(sh->tag_table);
525 		sh->tag_table = NULL;
526 	}
527 	if (sh->tunnel_hub) {
528 		mlx5_release_tunnel_hub(sh, priv->dev_port);
529 		sh->tunnel_hub = NULL;
530 	}
531 	mlx5_free_table_hash_list(priv);
532 	if (sh->port_id_action_list) {
533 		mlx5_list_destroy(sh->port_id_action_list);
534 		sh->port_id_action_list = NULL;
535 	}
536 	if (sh->push_vlan_action_list) {
537 		mlx5_list_destroy(sh->push_vlan_action_list);
538 		sh->push_vlan_action_list = NULL;
539 	}
540 	if (sh->sample_action_list) {
541 		mlx5_list_destroy(sh->sample_action_list);
542 		sh->sample_action_list = NULL;
543 	}
544 	if (sh->dest_array_list) {
545 		mlx5_list_destroy(sh->dest_array_list);
546 		sh->dest_array_list = NULL;
547 	}
548 }
549 
550 /**
551  * Initialize shared data between primary and secondary process.
552  *
553  * A memzone is reserved by primary process and secondary processes attach to
554  * the memzone.
555  *
556  * @return
557  *   0 on success, a negative errno value otherwise and rte_errno is set.
558  */
559 static int
560 mlx5_init_shared_data(void)
561 {
562 	const struct rte_memzone *mz;
563 	int ret = 0;
564 
565 	rte_spinlock_lock(&mlx5_shared_data_lock);
566 	if (mlx5_shared_data == NULL) {
567 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
568 			/* Allocate shared memory. */
569 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
570 						 sizeof(*mlx5_shared_data),
571 						 SOCKET_ID_ANY, 0);
572 			if (mz == NULL) {
573 				DRV_LOG(ERR,
574 					"Cannot allocate mlx5 shared data");
575 				ret = -rte_errno;
576 				goto error;
577 			}
578 			mlx5_shared_data = mz->addr;
579 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
580 			rte_spinlock_init(&mlx5_shared_data->lock);
581 		} else {
582 			/* Lookup allocated shared memory. */
583 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
584 			if (mz == NULL) {
585 				DRV_LOG(ERR,
586 					"Cannot attach mlx5 shared data");
587 				ret = -rte_errno;
588 				goto error;
589 			}
590 			mlx5_shared_data = mz->addr;
591 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
592 		}
593 	}
594 error:
595 	rte_spinlock_unlock(&mlx5_shared_data_lock);
596 	return ret;
597 }
598 
599 /**
600  * PMD global initialization.
601  *
602  * Independent from individual device, this function initializes global
603  * per-PMD data structures distinguishing primary and secondary processes.
604  * Hence, each initialization is called once per a process.
605  *
606  * @return
607  *   0 on success, a negative errno value otherwise and rte_errno is set.
608  */
609 static int
610 mlx5_init_once(void)
611 {
612 	struct mlx5_shared_data *sd;
613 	struct mlx5_local_data *ld = &mlx5_local_data;
614 	int ret = 0;
615 
616 	if (mlx5_init_shared_data())
617 		return -rte_errno;
618 	sd = mlx5_shared_data;
619 	MLX5_ASSERT(sd);
620 	rte_spinlock_lock(&sd->lock);
621 	switch (rte_eal_process_type()) {
622 	case RTE_PROC_PRIMARY:
623 		if (sd->init_done)
624 			break;
625 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
626 					   mlx5_mp_os_primary_handle);
627 		if (ret)
628 			goto out;
629 		sd->init_done = true;
630 		break;
631 	case RTE_PROC_SECONDARY:
632 		if (ld->init_done)
633 			break;
634 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
635 					     mlx5_mp_os_secondary_handle);
636 		if (ret)
637 			goto out;
638 		++sd->secondary_cnt;
639 		ld->init_done = true;
640 		break;
641 	default:
642 		break;
643 	}
644 out:
645 	rte_spinlock_unlock(&sd->lock);
646 	return ret;
647 }
648 
649 /**
650  * Create the Tx queue DevX/Verbs object.
651  *
652  * @param dev
653  *   Pointer to Ethernet device.
654  * @param idx
655  *   Queue index in DPDK Tx queue array.
656  *
657  * @return
658  *   0 on success, a negative errno value otherwise and rte_errno is set.
659  */
660 static int
661 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
662 {
663 	struct mlx5_priv *priv = dev->data->dev_private;
664 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
665 	struct mlx5_txq_ctrl *txq_ctrl =
666 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
667 
668 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
669 		return mlx5_txq_devx_obj_new(dev, idx);
670 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
671 	if (!priv->config.dv_esw_en)
672 		return mlx5_txq_devx_obj_new(dev, idx);
673 #endif
674 	return mlx5_txq_ibv_obj_new(dev, idx);
675 }
676 
677 /**
678  * Release an Tx DevX/verbs queue object.
679  *
680  * @param txq_obj
681  *   DevX/Verbs Tx queue object.
682  */
683 static void
684 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
685 {
686 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
687 		mlx5_txq_devx_obj_release(txq_obj);
688 		return;
689 	}
690 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
691 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
692 		mlx5_txq_devx_obj_release(txq_obj);
693 		return;
694 	}
695 #endif
696 	mlx5_txq_ibv_obj_release(txq_obj);
697 }
698 
699 /**
700  * DV flow counter mode detect and config.
701  *
702  * @param dev
703  *   Pointer to rte_eth_dev structure.
704  *
705  */
706 static void
707 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
708 {
709 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
710 	struct mlx5_priv *priv = dev->data->dev_private;
711 	struct mlx5_dev_ctx_shared *sh = priv->sh;
712 	bool fallback;
713 
714 #ifndef HAVE_IBV_DEVX_ASYNC
715 	fallback = true;
716 #else
717 	fallback = false;
718 	if (!sh->devx || !priv->config.dv_flow_en ||
719 	    !priv->config.hca_attr.flow_counters_dump ||
720 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
721 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
722 		fallback = true;
723 #endif
724 	if (fallback)
725 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
726 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
727 			priv->config.hca_attr.flow_counters_dump,
728 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
729 	/* Initialize fallback mode only on the port initializes sh. */
730 	if (sh->refcnt == 1)
731 		sh->cmng.counter_fallback = fallback;
732 	else if (fallback != sh->cmng.counter_fallback)
733 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
734 			"with others:%d.", PORT_ID(priv), fallback);
735 #endif
736 }
737 
738 /**
739  * DR flow drop action support detect.
740  *
741  * @param dev
742  *   Pointer to rte_eth_dev structure.
743  *
744  */
745 static void
746 mlx5_flow_drop_action_config(struct rte_eth_dev *dev __rte_unused)
747 {
748 #ifdef HAVE_MLX5DV_DR
749 	struct mlx5_priv *priv = dev->data->dev_private;
750 
751 	if (!priv->config.dv_flow_en || !priv->sh->dr_drop_action)
752 		return;
753 	/**
754 	 * DR supports drop action placeholder when it is supported;
755 	 * otherwise, use the queue drop action.
756 	 */
757 	if (mlx5_flow_discover_dr_action_support(dev))
758 		priv->root_drop_action = priv->drop_queue.hrxq->action;
759 	else
760 		priv->root_drop_action = priv->sh->dr_drop_action;
761 #endif
762 }
763 
764 static void
765 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
766 {
767 	struct mlx5_priv *priv = dev->data->dev_private;
768 	void *ctx = priv->sh->cdev->ctx;
769 
770 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
771 	if (!priv->q_counters) {
772 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
773 		struct ibv_wq *wq;
774 
775 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
776 			"by DevX - fall-back to use the kernel driver global "
777 			"queue counter.", dev->data->port_id);
778 		/* Create WQ by kernel and query its queue counter ID. */
779 		if (cq) {
780 			wq = mlx5_glue->create_wq(ctx,
781 						  &(struct ibv_wq_init_attr){
782 						    .wq_type = IBV_WQT_RQ,
783 						    .max_wr = 1,
784 						    .max_sge = 1,
785 						    .pd = priv->sh->cdev->pd,
786 						    .cq = cq,
787 						});
788 			if (wq) {
789 				/* Counter is assigned only on RDY state. */
790 				int ret = mlx5_glue->modify_wq(wq,
791 						 &(struct ibv_wq_attr){
792 						 .attr_mask = IBV_WQ_ATTR_STATE,
793 						 .wq_state = IBV_WQS_RDY,
794 						});
795 
796 				if (ret == 0)
797 					mlx5_devx_cmd_wq_query(wq,
798 							 &priv->counter_set_id);
799 				claim_zero(mlx5_glue->destroy_wq(wq));
800 			}
801 			claim_zero(mlx5_glue->destroy_cq(cq));
802 		}
803 	} else {
804 		priv->counter_set_id = priv->q_counters->id;
805 	}
806 	if (priv->counter_set_id == 0)
807 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
808 			"available.", dev->data->port_id);
809 }
810 
811 /**
812  * Check if representor spawn info match devargs.
813  *
814  * @param spawn
815  *   Verbs device parameters (name, port, switch_info) to spawn.
816  * @param eth_da
817  *   Device devargs to probe.
818  *
819  * @return
820  *   Match result.
821  */
822 static bool
823 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
824 		       struct rte_eth_devargs *eth_da)
825 {
826 	struct mlx5_switch_info *switch_info = &spawn->info;
827 	unsigned int p, f;
828 	uint16_t id;
829 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
830 						      eth_da->type);
831 
832 	switch (eth_da->type) {
833 	case RTE_ETH_REPRESENTOR_SF:
834 		if (!(spawn->info.port_name == -1 &&
835 		      switch_info->name_type ==
836 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
837 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
838 			rte_errno = EBUSY;
839 			return false;
840 		}
841 		break;
842 	case RTE_ETH_REPRESENTOR_VF:
843 		/* Allows HPF representor index -1 as exception. */
844 		if (!(spawn->info.port_name == -1 &&
845 		      switch_info->name_type ==
846 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
847 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
848 			rte_errno = EBUSY;
849 			return false;
850 		}
851 		break;
852 	case RTE_ETH_REPRESENTOR_NONE:
853 		rte_errno = EBUSY;
854 		return false;
855 	default:
856 		rte_errno = ENOTSUP;
857 		DRV_LOG(ERR, "unsupported representor type");
858 		return false;
859 	}
860 	/* Check representor ID: */
861 	for (p = 0; p < eth_da->nb_ports; ++p) {
862 		if (spawn->pf_bond < 0) {
863 			/* For non-LAG mode, allow and ignore pf. */
864 			switch_info->pf_num = eth_da->ports[p];
865 			repr_id = mlx5_representor_id_encode(switch_info,
866 							     eth_da->type);
867 		}
868 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
869 			id = MLX5_REPRESENTOR_ID
870 				(eth_da->ports[p], eth_da->type,
871 				 eth_da->representor_ports[f]);
872 			if (repr_id == id)
873 				return true;
874 		}
875 	}
876 	rte_errno = EBUSY;
877 	return false;
878 }
879 
880 
881 /**
882  * Spawn an Ethernet device from Verbs information.
883  *
884  * @param dpdk_dev
885  *   Backing DPDK device.
886  * @param spawn
887  *   Verbs device parameters (name, port, switch_info) to spawn.
888  * @param config
889  *   Device configuration parameters.
890  * @param eth_da
891  *   Device arguments.
892  *
893  * @return
894  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
895  *   is set. The following errors are defined:
896  *
897  *   EBUSY: device is not supposed to be spawned.
898  *   EEXIST: device is already spawned
899  */
900 static struct rte_eth_dev *
901 mlx5_dev_spawn(struct rte_device *dpdk_dev,
902 	       struct mlx5_dev_spawn_data *spawn,
903 	       struct mlx5_dev_config *config,
904 	       struct rte_eth_devargs *eth_da)
905 {
906 	const struct mlx5_switch_info *switch_info = &spawn->info;
907 	struct mlx5_dev_ctx_shared *sh = NULL;
908 	struct ibv_port_attr port_attr;
909 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
910 	struct rte_eth_dev *eth_dev = NULL;
911 	struct mlx5_priv *priv = NULL;
912 	int err = 0;
913 	unsigned int hw_padding = 0;
914 	unsigned int mps;
915 	unsigned int mpls_en = 0;
916 	unsigned int swp = 0;
917 	unsigned int mprq = 0;
918 	unsigned int mprq_min_stride_size_n = 0;
919 	unsigned int mprq_max_stride_size_n = 0;
920 	unsigned int mprq_min_stride_num_n = 0;
921 	unsigned int mprq_max_stride_num_n = 0;
922 	struct rte_ether_addr mac;
923 	char name[RTE_ETH_NAME_MAX_LEN];
924 	int own_domain_id = 0;
925 	uint16_t port_id;
926 	struct mlx5_port_info vport_info = { .query_flags = 0 };
927 	int i;
928 
929 	/* Determine if this port representor is supposed to be spawned. */
930 	if (switch_info->representor && dpdk_dev->devargs &&
931 	    !mlx5_representor_match(spawn, eth_da))
932 		return NULL;
933 	/* Build device name. */
934 	if (spawn->pf_bond < 0) {
935 		/* Single device. */
936 		if (!switch_info->representor)
937 			strlcpy(name, dpdk_dev->name, sizeof(name));
938 		else
939 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
940 				 dpdk_dev->name,
941 				 switch_info->name_type ==
942 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
943 				 switch_info->port_name);
944 	} else {
945 		/* Bonding device. */
946 		if (!switch_info->representor) {
947 			err = snprintf(name, sizeof(name), "%s_%s",
948 				       dpdk_dev->name, spawn->phys_dev_name);
949 		} else {
950 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
951 				dpdk_dev->name, spawn->phys_dev_name,
952 				switch_info->ctrl_num,
953 				switch_info->pf_num,
954 				switch_info->name_type ==
955 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
956 				switch_info->port_name);
957 		}
958 	}
959 	if (err >= (int)sizeof(name))
960 		DRV_LOG(WARNING, "device name overflow %s", name);
961 	/* check if the device is already spawned */
962 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
963 		rte_errno = EEXIST;
964 		return NULL;
965 	}
966 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
967 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
968 		struct mlx5_mp_id mp_id;
969 
970 		eth_dev = rte_eth_dev_attach_secondary(name);
971 		if (eth_dev == NULL) {
972 			DRV_LOG(ERR, "can not attach rte ethdev");
973 			rte_errno = ENOMEM;
974 			return NULL;
975 		}
976 		eth_dev->device = dpdk_dev;
977 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
978 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
979 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
980 		err = mlx5_proc_priv_init(eth_dev);
981 		if (err)
982 			return NULL;
983 		mlx5_mp_id_init(&mp_id, eth_dev->data->port_id);
984 		/* Receive command fd from primary process */
985 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
986 		if (err < 0)
987 			goto err_secondary;
988 		/* Remap UAR for Tx queues. */
989 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
990 		if (err)
991 			goto err_secondary;
992 		/*
993 		 * Ethdev pointer is still required as input since
994 		 * the primary device is not accessible from the
995 		 * secondary process.
996 		 */
997 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
998 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
999 		return eth_dev;
1000 err_secondary:
1001 		mlx5_dev_close(eth_dev);
1002 		return NULL;
1003 	}
1004 	/*
1005 	 * Some parameters ("tx_db_nc" in particularly) are needed in
1006 	 * advance to create dv/verbs device context. We proceed the
1007 	 * devargs here to get ones, and later proceed devargs again
1008 	 * to override some hardware settings.
1009 	 */
1010 	err = mlx5_args(config, dpdk_dev->devargs);
1011 	if (err) {
1012 		err = rte_errno;
1013 		DRV_LOG(ERR, "failed to process device arguments: %s",
1014 			strerror(rte_errno));
1015 		goto error;
1016 	}
1017 	if (config->dv_miss_info) {
1018 		if (switch_info->master || switch_info->representor)
1019 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
1020 	}
1021 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
1022 	if (!sh)
1023 		return NULL;
1024 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1025 	config->dest_tir = 1;
1026 #endif
1027 #ifdef HAVE_IBV_MLX5_MOD_SWP
1028 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1029 #endif
1030 	/*
1031 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1032 	 * as all ConnectX-5 devices.
1033 	 */
1034 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1035 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1036 #endif
1037 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1038 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1039 #endif
1040 	mlx5_glue->dv_query_device(sh->cdev->ctx, &dv_attr);
1041 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1042 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1043 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1044 			mps = MLX5_MPW_ENHANCED;
1045 		} else {
1046 			DRV_LOG(DEBUG, "MPW is supported");
1047 			mps = MLX5_MPW;
1048 		}
1049 	} else {
1050 		DRV_LOG(DEBUG, "MPW isn't supported");
1051 		mps = MLX5_MPW_DISABLED;
1052 	}
1053 #ifdef HAVE_IBV_MLX5_MOD_SWP
1054 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1055 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1056 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1057 #endif
1058 	config->swp = swp & (MLX5_SW_PARSING_CAP | MLX5_SW_PARSING_CSUM_CAP |
1059 		MLX5_SW_PARSING_TSO_CAP);
1060 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1061 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1062 		struct mlx5dv_striding_rq_caps mprq_caps =
1063 			dv_attr.striding_rq_caps;
1064 
1065 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1066 			mprq_caps.min_single_stride_log_num_of_bytes);
1067 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1068 			mprq_caps.max_single_stride_log_num_of_bytes);
1069 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1070 			mprq_caps.min_single_wqe_log_num_of_strides);
1071 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1072 			mprq_caps.max_single_wqe_log_num_of_strides);
1073 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1074 			mprq_caps.supported_qpts);
1075 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1076 		mprq = 1;
1077 		mprq_min_stride_size_n =
1078 			mprq_caps.min_single_stride_log_num_of_bytes;
1079 		mprq_max_stride_size_n =
1080 			mprq_caps.max_single_stride_log_num_of_bytes;
1081 		mprq_min_stride_num_n =
1082 			mprq_caps.min_single_wqe_log_num_of_strides;
1083 		mprq_max_stride_num_n =
1084 			mprq_caps.max_single_wqe_log_num_of_strides;
1085 	}
1086 #endif
1087 	/* Rx CQE compression is enabled by default. */
1088 	config->cqe_comp = 1;
1089 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1090 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1091 		config->tunnel_en = dv_attr.tunnel_offloads_caps &
1092 			     (MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN |
1093 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE |
1094 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE);
1095 	}
1096 	if (config->tunnel_en) {
1097 		DRV_LOG(DEBUG, "tunnel offloading is supported for %s%s%s",
1098 		config->tunnel_en &
1099 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN ? "[VXLAN]" : "",
1100 		config->tunnel_en &
1101 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE ? "[GRE]" : "",
1102 		config->tunnel_en &
1103 		MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE ? "[GENEVE]" : ""
1104 		);
1105 	} else {
1106 		DRV_LOG(DEBUG, "tunnel offloading is not supported");
1107 	}
1108 #else
1109 	DRV_LOG(WARNING,
1110 		"tunnel offloading disabled due to old OFED/rdma-core version");
1111 #endif
1112 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1113 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1114 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1115 		   (dv_attr.tunnel_offloads_caps &
1116 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1117 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1118 		mpls_en ? "" : "not ");
1119 #else
1120 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1121 		" old OFED/rdma-core version or firmware configuration");
1122 #endif
1123 	config->mpls_en = mpls_en;
1124 	/* Check port status. */
1125 	err = mlx5_glue->query_port(sh->cdev->ctx, spawn->phys_port,
1126 				    &port_attr);
1127 	if (err) {
1128 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1129 		goto error;
1130 	}
1131 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1132 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1133 		err = EINVAL;
1134 		goto error;
1135 	}
1136 	if (port_attr.state != IBV_PORT_ACTIVE)
1137 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1138 			mlx5_glue->port_state_str(port_attr.state),
1139 			port_attr.state);
1140 	/* Allocate private eth device data. */
1141 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1142 			   sizeof(*priv),
1143 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1144 	if (priv == NULL) {
1145 		DRV_LOG(ERR, "priv allocation failure");
1146 		err = ENOMEM;
1147 		goto error;
1148 	}
1149 	priv->sh = sh;
1150 	priv->dev_port = spawn->phys_port;
1151 	priv->pci_dev = spawn->pci_dev;
1152 	priv->mtu = RTE_ETHER_MTU;
1153 	/* Some internal functions rely on Netlink sockets, open them now. */
1154 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1155 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1156 	priv->representor = !!switch_info->representor;
1157 	priv->master = !!switch_info->master;
1158 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1159 	priv->vport_meta_tag = 0;
1160 	priv->vport_meta_mask = 0;
1161 	priv->pf_bond = spawn->pf_bond;
1162 
1163 	DRV_LOG(DEBUG,
1164 		"dev_port=%u bus=%s pci=%s master=%d representor=%d pf_bond=%d\n",
1165 		priv->dev_port, dpdk_dev->bus->name,
1166 		priv->pci_dev ? priv->pci_dev->name : "NONE",
1167 		priv->master, priv->representor, priv->pf_bond);
1168 
1169 	/*
1170 	 * If we have E-Switch we should determine the vport attributes.
1171 	 * E-Switch may use either source vport field or reg_c[0] metadata
1172 	 * register to match on vport index. The engaged part of metadata
1173 	 * register is defined by mask.
1174 	 */
1175 	if (switch_info->representor || switch_info->master) {
1176 		err = mlx5_glue->devx_port_query(sh->cdev->ctx,
1177 						 spawn->phys_port,
1178 						 &vport_info);
1179 		if (err) {
1180 			DRV_LOG(WARNING,
1181 				"Cannot query devx port %d on device %s",
1182 				spawn->phys_port, spawn->phys_dev_name);
1183 			vport_info.query_flags = 0;
1184 		}
1185 	}
1186 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1187 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1188 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1189 		if (!priv->vport_meta_mask) {
1190 			DRV_LOG(ERR,
1191 				"vport zero mask for port %d on bonding device %s",
1192 				spawn->phys_port, spawn->phys_dev_name);
1193 			err = ENOTSUP;
1194 			goto error;
1195 		}
1196 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1197 			DRV_LOG(ERR,
1198 				"Invalid vport tag for port %d on bonding device %s",
1199 				spawn->phys_port, spawn->phys_dev_name);
1200 			err = ENOTSUP;
1201 			goto error;
1202 		}
1203 	}
1204 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1205 		priv->vport_id = vport_info.vport_id;
1206 	} else if (spawn->pf_bond >= 0 &&
1207 		   (switch_info->representor || switch_info->master)) {
1208 		DRV_LOG(ERR,
1209 			"Cannot deduce vport index for port %d on bonding device %s",
1210 			spawn->phys_port, spawn->phys_dev_name);
1211 		err = ENOTSUP;
1212 		goto error;
1213 	} else {
1214 		/*
1215 		 * Suppose vport index in compatible way. Kernel/rdma_core
1216 		 * support single E-Switch per PF configurations only and
1217 		 * vport_id field contains the vport index for associated VF,
1218 		 * which is deduced from representor port name.
1219 		 * For example, let's have the IB device port 10, it has
1220 		 * attached network device eth0, which has port name attribute
1221 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1222 		 * as 3 (2+1). This assigning schema should be changed if the
1223 		 * multiple E-Switch instances per PF configurations or/and PCI
1224 		 * subfunctions are added.
1225 		 */
1226 		priv->vport_id = switch_info->representor ?
1227 				 switch_info->port_name + 1 : -1;
1228 	}
1229 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1230 							  eth_da->type);
1231 	/*
1232 	 * Look for sibling devices in order to reuse their switch domain
1233 	 * if any, otherwise allocate one.
1234 	 */
1235 	MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1236 		const struct mlx5_priv *opriv =
1237 			rte_eth_devices[port_id].data->dev_private;
1238 
1239 		if (!opriv ||
1240 		    opriv->sh != priv->sh ||
1241 			opriv->domain_id ==
1242 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1243 			continue;
1244 		priv->domain_id = opriv->domain_id;
1245 		DRV_LOG(DEBUG, "dev_port-%u inherit domain_id=%u\n",
1246 			priv->dev_port, priv->domain_id);
1247 		break;
1248 	}
1249 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1250 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1251 		if (err) {
1252 			err = rte_errno;
1253 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1254 				strerror(rte_errno));
1255 			goto error;
1256 		}
1257 		own_domain_id = 1;
1258 		DRV_LOG(DEBUG, "dev_port-%u new domain_id=%u\n",
1259 			priv->dev_port, priv->domain_id);
1260 	}
1261 	/* Override some values set by hardware configuration. */
1262 	mlx5_args(config, dpdk_dev->devargs);
1263 	err = mlx5_dev_check_sibling_config(priv, config, dpdk_dev);
1264 	if (err)
1265 		goto error;
1266 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1267 			    IBV_DEVICE_RAW_IP_CSUM);
1268 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1269 		(config->hw_csum ? "" : "not "));
1270 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1271 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1272 	DRV_LOG(DEBUG, "counters are not supported");
1273 #endif
1274 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1275 	if (config->dv_flow_en) {
1276 		DRV_LOG(WARNING, "DV flow is not supported");
1277 		config->dv_flow_en = 0;
1278 	}
1279 #endif
1280 	if (spawn->max_port > UINT8_MAX) {
1281 		/* Verbs can't support ports larger than 255 by design. */
1282 		DRV_LOG(ERR, "can't support IB ports > UINT8_MAX");
1283 		err = EINVAL;
1284 		goto error;
1285 	}
1286 	config->ind_table_max_size =
1287 		sh->device_attr.max_rwq_indirection_table_size;
1288 	/*
1289 	 * Remove this check once DPDK supports larger/variable
1290 	 * indirection tables.
1291 	 */
1292 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1293 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1294 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1295 		config->ind_table_max_size);
1296 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1297 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1298 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1299 		(config->hw_vlan_strip ? "" : "not "));
1300 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1301 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1302 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1303 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1304 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1305 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1306 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1307 #endif
1308 	if (config->hw_padding && !hw_padding) {
1309 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1310 		config->hw_padding = 0;
1311 	} else if (config->hw_padding) {
1312 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1313 	}
1314 	config->tso = (sh->device_attr.max_tso > 0 &&
1315 		      (sh->device_attr.tso_supported_qpts &
1316 		       (1 << IBV_QPT_RAW_PACKET)));
1317 	if (config->tso)
1318 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1319 	/*
1320 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1321 	 * by default.
1322 	 */
1323 	if (config->mps == MLX5_ARG_UNSET)
1324 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1325 							  MLX5_MPW_DISABLED;
1326 	else
1327 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1328 	DRV_LOG(INFO, "%sMPS is %s",
1329 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1330 		config->mps == MLX5_MPW ? "legacy " : "",
1331 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1332 	if (sh->devx) {
1333 		config->hca_attr = sh->cdev->config.hca_attr;
1334 		sh->steering_format_version =
1335 			config->hca_attr.steering_format_version;
1336 		/* Check for LRO support. */
1337 		if (config->dest_tir && config->hca_attr.lro_cap &&
1338 		    config->dv_flow_en) {
1339 			/* TBD check tunnel lro caps. */
1340 			config->lro.supported = config->hca_attr.lro_cap;
1341 			DRV_LOG(DEBUG, "Device supports LRO");
1342 			/*
1343 			 * If LRO timeout is not configured by application,
1344 			 * use the minimal supported value.
1345 			 */
1346 			if (!config->lro.timeout)
1347 				config->lro.timeout =
1348 				config->hca_attr.lro_timer_supported_periods[0];
1349 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1350 				config->lro.timeout);
1351 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1352 				"required for coalescing is %d bytes",
1353 				config->hca_attr.lro_min_mss_size);
1354 		}
1355 #if defined(HAVE_MLX5DV_DR) && \
1356 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1357 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1358 		if (config->hca_attr.qos.sup &&
1359 		    config->hca_attr.qos.flow_meter_old &&
1360 		    config->dv_flow_en) {
1361 			uint8_t reg_c_mask =
1362 				config->hca_attr.qos.flow_meter_reg_c_ids;
1363 			/*
1364 			 * Meter needs two REG_C's for color match and pre-sfx
1365 			 * flow match. Here get the REG_C for color match.
1366 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1367 			 */
1368 			reg_c_mask &= 0xfc;
1369 			if (__builtin_popcount(reg_c_mask) < 1) {
1370 				priv->mtr_en = 0;
1371 				DRV_LOG(WARNING, "No available register for"
1372 					" meter.");
1373 			} else {
1374 				/*
1375 				 * The meter color register is used by the
1376 				 * flow-hit feature as well.
1377 				 * The flow-hit feature must use REG_C_3
1378 				 * Prefer REG_C_3 if it is available.
1379 				 */
1380 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1381 					priv->mtr_color_reg = REG_C_3;
1382 				else
1383 					priv->mtr_color_reg = ffs(reg_c_mask)
1384 							      - 1 + REG_C_0;
1385 				priv->mtr_en = 1;
1386 				priv->mtr_reg_share =
1387 				      config->hca_attr.qos.flow_meter;
1388 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1389 					priv->mtr_color_reg);
1390 			}
1391 		}
1392 		if (config->hca_attr.qos.sup &&
1393 			config->hca_attr.qos.flow_meter_aso_sup) {
1394 			uint32_t log_obj_size =
1395 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1396 			if (log_obj_size >=
1397 			config->hca_attr.qos.log_meter_aso_granularity &&
1398 			log_obj_size <=
1399 			config->hca_attr.qos.log_meter_aso_max_alloc)
1400 				sh->meter_aso_en = 1;
1401 		}
1402 		if (priv->mtr_en) {
1403 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1404 			if (err) {
1405 				err = -err;
1406 				goto error;
1407 			}
1408 		}
1409 		if (config->hca_attr.flow.tunnel_header_0_1)
1410 			sh->tunnel_header_0_1 = 1;
1411 #endif
1412 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1413 		if (config->hca_attr.flow_hit_aso &&
1414 		    priv->mtr_color_reg == REG_C_3) {
1415 			sh->flow_hit_aso_en = 1;
1416 			err = mlx5_flow_aso_age_mng_init(sh);
1417 			if (err) {
1418 				err = -err;
1419 				goto error;
1420 			}
1421 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1422 		}
1423 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1424 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1425 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1426 		if (config->hca_attr.ct_offload &&
1427 		    priv->mtr_color_reg == REG_C_3) {
1428 			err = mlx5_flow_aso_ct_mng_init(sh);
1429 			if (err) {
1430 				err = -err;
1431 				goto error;
1432 			}
1433 			DRV_LOG(DEBUG, "CT ASO is supported.");
1434 			sh->ct_aso_en = 1;
1435 		}
1436 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1437 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1438 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1439 		    config->dv_flow_en) {
1440 			priv->sampler_en = 1;
1441 			DRV_LOG(DEBUG, "Sampler enabled!");
1442 		} else {
1443 			priv->sampler_en = 0;
1444 			if (!config->hca_attr.log_max_ft_sampler_num)
1445 				DRV_LOG(WARNING,
1446 					"No available register for sampler.");
1447 			else
1448 				DRV_LOG(DEBUG, "DV flow is not supported!");
1449 		}
1450 #endif
1451 	}
1452 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1453 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1454 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1455 		config->cqe_comp = 0;
1456 	}
1457 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1458 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1459 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1460 				 " format isn't supported.");
1461 		config->cqe_comp = 0;
1462 	}
1463 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1464 	    (!sh->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1465 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1466 				 " format isn't supported.");
1467 		config->cqe_comp = 0;
1468 	}
1469 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1470 			config->cqe_comp ? "" : "not ");
1471 	if (config->tx_pp) {
1472 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1473 			config->hca_attr.dev_freq_khz);
1474 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1475 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1476 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1477 			config->hca_attr.cross_channel ? "" : "not ");
1478 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1479 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1480 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1481 			config->hca_attr.non_wire_sq ? "" : "not ");
1482 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1483 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1484 			config->hca_attr.log_max_static_sq_wq);
1485 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1486 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1487 		if (!sh->devx) {
1488 			DRV_LOG(ERR, "DevX is required for packet pacing");
1489 			err = ENODEV;
1490 			goto error;
1491 		}
1492 		if (!config->hca_attr.qos.packet_pacing) {
1493 			DRV_LOG(ERR, "Packet pacing is not supported");
1494 			err = ENODEV;
1495 			goto error;
1496 		}
1497 		if (!config->hca_attr.cross_channel) {
1498 			DRV_LOG(ERR, "Cross channel operations are"
1499 				     " required for packet pacing");
1500 			err = ENODEV;
1501 			goto error;
1502 		}
1503 		if (!config->hca_attr.wqe_index_ignore) {
1504 			DRV_LOG(ERR, "WQE index ignore feature is"
1505 				     " required for packet pacing");
1506 			err = ENODEV;
1507 			goto error;
1508 		}
1509 		if (!config->hca_attr.non_wire_sq) {
1510 			DRV_LOG(ERR, "Non-wire SQ feature is"
1511 				     " required for packet pacing");
1512 			err = ENODEV;
1513 			goto error;
1514 		}
1515 		if (!config->hca_attr.log_max_static_sq_wq) {
1516 			DRV_LOG(ERR, "Static WQE SQ feature is"
1517 				     " required for packet pacing");
1518 			err = ENODEV;
1519 			goto error;
1520 		}
1521 		if (!config->hca_attr.qos.wqe_rate_pp) {
1522 			DRV_LOG(ERR, "WQE rate mode is required"
1523 				     " for packet pacing");
1524 			err = ENODEV;
1525 			goto error;
1526 		}
1527 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1528 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1529 			     " can't create queues for packet pacing");
1530 		err = ENODEV;
1531 		goto error;
1532 #endif
1533 	}
1534 	if (sh->devx) {
1535 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1536 
1537 		err = config->hca_attr.access_register_user ?
1538 			mlx5_devx_cmd_register_read
1539 				(sh->cdev->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1540 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1541 		if (!err) {
1542 			uint32_t ts_mode;
1543 
1544 			/* MTUTC register is read successfully. */
1545 			ts_mode = MLX5_GET(register_mtutc, reg,
1546 					   time_stamp_mode);
1547 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1548 				config->rt_timestamp = 1;
1549 		} else {
1550 			/* Kernel does not support register reading. */
1551 			if (config->hca_attr.dev_freq_khz ==
1552 						 (NS_PER_S / MS_PER_S))
1553 				config->rt_timestamp = 1;
1554 		}
1555 	}
1556 	/*
1557 	 * If HW has bug working with tunnel packet decapsulation and
1558 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1559 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1560 	 */
1561 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1562 		config->hw_fcs_strip = 0;
1563 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1564 		(config->hw_fcs_strip ? "" : "not "));
1565 	if (config->mprq.enabled && mprq) {
1566 		if (config->mprq.stride_num_n &&
1567 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1568 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1569 			config->mprq.stride_num_n =
1570 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1571 						mprq_min_stride_num_n),
1572 					mprq_max_stride_num_n);
1573 			DRV_LOG(WARNING,
1574 				"the number of strides"
1575 				" for Multi-Packet RQ is out of range,"
1576 				" setting default value (%u)",
1577 				1 << config->mprq.stride_num_n);
1578 		}
1579 		if (config->mprq.stride_size_n &&
1580 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1581 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1582 			config->mprq.stride_size_n =
1583 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1584 						mprq_min_stride_size_n),
1585 					mprq_max_stride_size_n);
1586 			DRV_LOG(WARNING,
1587 				"the size of a stride"
1588 				" for Multi-Packet RQ is out of range,"
1589 				" setting default value (%u)",
1590 				1 << config->mprq.stride_size_n);
1591 		}
1592 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1593 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1594 	} else if (config->mprq.enabled && !mprq) {
1595 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1596 		config->mprq.enabled = 0;
1597 	}
1598 	if (config->max_dump_files_num == 0)
1599 		config->max_dump_files_num = 128;
1600 	eth_dev = rte_eth_dev_allocate(name);
1601 	if (eth_dev == NULL) {
1602 		DRV_LOG(ERR, "can not allocate rte ethdev");
1603 		err = ENOMEM;
1604 		goto error;
1605 	}
1606 	if (priv->representor) {
1607 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1608 		eth_dev->data->representor_id = priv->representor_id;
1609 		MLX5_ETH_FOREACH_DEV(port_id, dpdk_dev) {
1610 			struct mlx5_priv *opriv =
1611 				rte_eth_devices[port_id].data->dev_private;
1612 			if (opriv &&
1613 			    opriv->master &&
1614 			    opriv->domain_id == priv->domain_id &&
1615 			    opriv->sh == priv->sh) {
1616 				eth_dev->data->backer_port_id = port_id;
1617 				break;
1618 			}
1619 		}
1620 		if (port_id >= RTE_MAX_ETHPORTS)
1621 			eth_dev->data->backer_port_id = eth_dev->data->port_id;
1622 	}
1623 	priv->mp_id.port_id = eth_dev->data->port_id;
1624 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1625 	/*
1626 	 * Store associated network device interface index. This index
1627 	 * is permanent throughout the lifetime of device. So, we may store
1628 	 * the ifindex here and use the cached value further.
1629 	 */
1630 	MLX5_ASSERT(spawn->ifindex);
1631 	priv->if_index = spawn->ifindex;
1632 	eth_dev->data->dev_private = priv;
1633 	priv->dev_data = eth_dev->data;
1634 	eth_dev->data->mac_addrs = priv->mac;
1635 	eth_dev->device = dpdk_dev;
1636 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1637 	/* Configure the first MAC address by default. */
1638 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1639 		DRV_LOG(ERR,
1640 			"port %u cannot get MAC address, is mlx5_en"
1641 			" loaded? (errno: %s)",
1642 			eth_dev->data->port_id, strerror(rte_errno));
1643 		err = ENODEV;
1644 		goto error;
1645 	}
1646 	DRV_LOG(INFO,
1647 		"port %u MAC address is " RTE_ETHER_ADDR_PRT_FMT,
1648 		eth_dev->data->port_id, RTE_ETHER_ADDR_BYTES(&mac));
1649 #ifdef RTE_LIBRTE_MLX5_DEBUG
1650 	{
1651 		char ifname[MLX5_NAMESIZE];
1652 
1653 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1654 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1655 				eth_dev->data->port_id, ifname);
1656 		else
1657 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1658 				eth_dev->data->port_id);
1659 	}
1660 #endif
1661 	/* Get actual MTU if possible. */
1662 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1663 	if (err) {
1664 		err = rte_errno;
1665 		goto error;
1666 	}
1667 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1668 		priv->mtu);
1669 	/* Initialize burst functions to prevent crashes before link-up. */
1670 	eth_dev->rx_pkt_burst = removed_rx_burst;
1671 	eth_dev->tx_pkt_burst = removed_tx_burst;
1672 	eth_dev->dev_ops = &mlx5_dev_ops;
1673 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1674 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1675 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1676 	/* Register MAC address. */
1677 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1678 	if (config->vf && config->vf_nl_en)
1679 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1680 				      mlx5_ifindex(eth_dev),
1681 				      eth_dev->data->mac_addrs,
1682 				      MLX5_MAX_MAC_ADDRESSES);
1683 	priv->ctrl_flows = 0;
1684 	rte_spinlock_init(&priv->flow_list_lock);
1685 	TAILQ_INIT(&priv->flow_meters);
1686 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1687 	if (!priv->mtr_profile_tbl)
1688 		goto error;
1689 	/* Bring Ethernet device up. */
1690 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1691 		eth_dev->data->port_id);
1692 	mlx5_set_link_up(eth_dev);
1693 	/*
1694 	 * Even though the interrupt handler is not installed yet,
1695 	 * interrupts will still trigger on the async_fd from
1696 	 * Verbs context returned by ibv_open_device().
1697 	 */
1698 	mlx5_link_update(eth_dev, 0);
1699 #ifdef HAVE_MLX5DV_DR_ESWITCH
1700 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1701 	      (switch_info->representor || switch_info->master)))
1702 		config->dv_esw_en = 0;
1703 #else
1704 	config->dv_esw_en = 0;
1705 #endif
1706 	/* Detect minimal data bytes to inline. */
1707 	mlx5_set_min_inline(spawn, config);
1708 	/* Store device configuration on private structure. */
1709 	priv->config = *config;
1710 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1711 		icfg[i].release_mem_en = !!config->reclaim_mode;
1712 		if (config->reclaim_mode)
1713 			icfg[i].per_core_cache = 0;
1714 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1715 		if (!priv->flows[i])
1716 			goto error;
1717 	}
1718 	/* Create context for virtual machine VLAN workaround. */
1719 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1720 	if (config->dv_flow_en) {
1721 		err = mlx5_alloc_shared_dr(priv);
1722 		if (err)
1723 			goto error;
1724 	}
1725 	if (sh->devx && config->dv_flow_en && config->dest_tir) {
1726 		priv->obj_ops = devx_obj_ops;
1727 		priv->obj_ops.drop_action_create =
1728 						ibv_obj_ops.drop_action_create;
1729 		priv->obj_ops.drop_action_destroy =
1730 						ibv_obj_ops.drop_action_destroy;
1731 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1732 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1733 #else
1734 		if (config->dv_esw_en)
1735 			priv->obj_ops.txq_obj_modify =
1736 						ibv_obj_ops.txq_obj_modify;
1737 #endif
1738 		/* Use specific wrappers for Tx object. */
1739 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1740 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1741 		mlx5_queue_counter_id_prepare(eth_dev);
1742 		priv->obj_ops.lb_dummy_queue_create =
1743 					mlx5_rxq_ibv_obj_dummy_lb_create;
1744 		priv->obj_ops.lb_dummy_queue_release =
1745 					mlx5_rxq_ibv_obj_dummy_lb_release;
1746 	} else {
1747 		priv->obj_ops = ibv_obj_ops;
1748 	}
1749 	if (config->tx_pp &&
1750 	    (priv->config.dv_esw_en ||
1751 	     priv->obj_ops.txq_obj_new != mlx5_os_txq_obj_new)) {
1752 		/*
1753 		 * HAVE_MLX5DV_DEVX_UAR_OFFSET is required to support
1754 		 * packet pacing and already checked above.
1755 		 * Hence, we should only make sure the SQs will be created
1756 		 * with DevX, not with Verbs.
1757 		 * Verbs allocates the SQ UAR on its own and it can't be shared
1758 		 * with Clock Queue UAR as required for Tx scheduling.
1759 		 */
1760 		DRV_LOG(ERR, "Verbs SQs, UAR can't be shared as required for packet pacing");
1761 		err = ENODEV;
1762 		goto error;
1763 	}
1764 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1765 	if (!priv->drop_queue.hrxq)
1766 		goto error;
1767 	/* Supported Verbs flow priority number detection. */
1768 	err = mlx5_flow_discover_priorities(eth_dev);
1769 	if (err < 0) {
1770 		err = -err;
1771 		goto error;
1772 	}
1773 	priv->config.flow_prio = err;
1774 	if (!priv->config.dv_esw_en &&
1775 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1776 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1777 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1778 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1779 	}
1780 	mlx5_set_metadata_mask(eth_dev);
1781 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1782 	    !priv->sh->dv_regc0_mask) {
1783 		DRV_LOG(ERR, "metadata mode %u is not supported "
1784 			     "(no metadata reg_c[0] is available)",
1785 			     priv->config.dv_xmeta_en);
1786 			err = ENOTSUP;
1787 			goto error;
1788 	}
1789 	priv->hrxqs = mlx5_list_create("hrxq", eth_dev, true,
1790 				       mlx5_hrxq_create_cb,
1791 				       mlx5_hrxq_match_cb,
1792 				       mlx5_hrxq_remove_cb,
1793 				       mlx5_hrxq_clone_cb,
1794 				       mlx5_hrxq_clone_free_cb);
1795 	if (!priv->hrxqs)
1796 		goto error;
1797 	rte_rwlock_init(&priv->ind_tbls_lock);
1798 	/* Query availability of metadata reg_c's. */
1799 	err = mlx5_flow_discover_mreg_c(eth_dev);
1800 	if (err < 0) {
1801 		err = -err;
1802 		goto error;
1803 	}
1804 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1805 		DRV_LOG(DEBUG,
1806 			"port %u extensive metadata register is not supported",
1807 			eth_dev->data->port_id);
1808 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1809 			DRV_LOG(ERR, "metadata mode %u is not supported "
1810 				     "(no metadata registers available)",
1811 				     priv->config.dv_xmeta_en);
1812 			err = ENOTSUP;
1813 			goto error;
1814 		}
1815 	}
1816 	if (priv->config.dv_flow_en &&
1817 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1818 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1819 	    priv->sh->dv_regc0_mask) {
1820 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1821 						      MLX5_FLOW_MREG_HTABLE_SZ,
1822 						      false, true, eth_dev,
1823 						      flow_dv_mreg_create_cb,
1824 						      flow_dv_mreg_match_cb,
1825 						      flow_dv_mreg_remove_cb,
1826 						      flow_dv_mreg_clone_cb,
1827 						    flow_dv_mreg_clone_free_cb);
1828 		if (!priv->mreg_cp_tbl) {
1829 			err = ENOMEM;
1830 			goto error;
1831 		}
1832 	}
1833 	rte_spinlock_init(&priv->shared_act_sl);
1834 	mlx5_flow_counter_mode_config(eth_dev);
1835 	mlx5_flow_drop_action_config(eth_dev);
1836 	if (priv->config.dv_flow_en)
1837 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1838 	return eth_dev;
1839 error:
1840 	if (priv) {
1841 		if (priv->mreg_cp_tbl)
1842 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1843 		if (priv->sh)
1844 			mlx5_os_free_shared_dr(priv);
1845 		if (priv->nl_socket_route >= 0)
1846 			close(priv->nl_socket_route);
1847 		if (priv->nl_socket_rdma >= 0)
1848 			close(priv->nl_socket_rdma);
1849 		if (priv->vmwa_context)
1850 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1851 		if (eth_dev && priv->drop_queue.hrxq)
1852 			mlx5_drop_action_destroy(eth_dev);
1853 		if (priv->mtr_profile_tbl)
1854 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1855 		if (own_domain_id)
1856 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1857 		if (priv->hrxqs)
1858 			mlx5_list_destroy(priv->hrxqs);
1859 		mlx5_free(priv);
1860 		if (eth_dev != NULL)
1861 			eth_dev->data->dev_private = NULL;
1862 	}
1863 	if (eth_dev != NULL) {
1864 		/* mac_addrs must not be freed alone because part of
1865 		 * dev_private
1866 		 **/
1867 		eth_dev->data->mac_addrs = NULL;
1868 		rte_eth_dev_release_port(eth_dev);
1869 	}
1870 	if (sh)
1871 		mlx5_free_shared_dev_ctx(sh);
1872 	MLX5_ASSERT(err > 0);
1873 	rte_errno = err;
1874 	return NULL;
1875 }
1876 
1877 /**
1878  * Comparison callback to sort device data.
1879  *
1880  * This is meant to be used with qsort().
1881  *
1882  * @param a[in]
1883  *   Pointer to pointer to first data object.
1884  * @param b[in]
1885  *   Pointer to pointer to second data object.
1886  *
1887  * @return
1888  *   0 if both objects are equal, less than 0 if the first argument is less
1889  *   than the second, greater than 0 otherwise.
1890  */
1891 static int
1892 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1893 {
1894 	const struct mlx5_switch_info *si_a =
1895 		&((const struct mlx5_dev_spawn_data *)a)->info;
1896 	const struct mlx5_switch_info *si_b =
1897 		&((const struct mlx5_dev_spawn_data *)b)->info;
1898 	int ret;
1899 
1900 	/* Master device first. */
1901 	ret = si_b->master - si_a->master;
1902 	if (ret)
1903 		return ret;
1904 	/* Then representor devices. */
1905 	ret = si_b->representor - si_a->representor;
1906 	if (ret)
1907 		return ret;
1908 	/* Unidentified devices come last in no specific order. */
1909 	if (!si_a->representor)
1910 		return 0;
1911 	/* Order representors by name. */
1912 	return si_a->port_name - si_b->port_name;
1913 }
1914 
1915 /**
1916  * Match PCI information for possible slaves of bonding device.
1917  *
1918  * @param[in] ibdev_name
1919  *   Name of Infiniband device.
1920  * @param[in] pci_dev
1921  *   Pointer to primary PCI address structure to match.
1922  * @param[in] nl_rdma
1923  *   Netlink RDMA group socket handle.
1924  * @param[in] owner
1925  *   Representor owner PF index.
1926  * @param[out] bond_info
1927  *   Pointer to bonding information.
1928  *
1929  * @return
1930  *   negative value if no bonding device found, otherwise
1931  *   positive index of slave PF in bonding.
1932  */
1933 static int
1934 mlx5_device_bond_pci_match(const char *ibdev_name,
1935 			   const struct rte_pci_addr *pci_dev,
1936 			   int nl_rdma, uint16_t owner,
1937 			   struct mlx5_bond_info *bond_info)
1938 {
1939 	char ifname[IF_NAMESIZE + 1];
1940 	unsigned int ifindex;
1941 	unsigned int np, i;
1942 	FILE *bond_file = NULL, *file;
1943 	int pf = -1;
1944 	int ret;
1945 
1946 	/*
1947 	 * Try to get master device name. If something goes wrong suppose
1948 	 * the lack of kernel support and no bonding devices.
1949 	 */
1950 	memset(bond_info, 0, sizeof(*bond_info));
1951 	if (nl_rdma < 0)
1952 		return -1;
1953 	if (!strstr(ibdev_name, "bond"))
1954 		return -1;
1955 	np = mlx5_nl_portnum(nl_rdma, ibdev_name);
1956 	if (!np)
1957 		return -1;
1958 	/*
1959 	 * The master device might not be on the predefined port(not on port
1960 	 * index 1, it is not guaranteed), we have to scan all Infiniband
1961 	 * device ports and find master.
1962 	 */
1963 	for (i = 1; i <= np; ++i) {
1964 		/* Check whether Infiniband port is populated. */
1965 		ifindex = mlx5_nl_ifindex(nl_rdma, ibdev_name, i);
1966 		if (!ifindex)
1967 			continue;
1968 		if (!if_indextoname(ifindex, ifname))
1969 			continue;
1970 		/* Try to read bonding slave names from sysfs. */
1971 		MKSTR(slaves,
1972 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1973 		bond_file = fopen(slaves, "r");
1974 		if (bond_file)
1975 			break;
1976 	}
1977 	if (!bond_file)
1978 		return -1;
1979 	/* Use safe format to check maximal buffer length. */
1980 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1981 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1982 		char tmp_str[IF_NAMESIZE + 32];
1983 		struct rte_pci_addr pci_addr;
1984 		struct mlx5_switch_info	info;
1985 
1986 		/* Process slave interface names in the loop. */
1987 		snprintf(tmp_str, sizeof(tmp_str),
1988 			 "/sys/class/net/%s", ifname);
1989 		if (mlx5_get_pci_addr(tmp_str, &pci_addr)) {
1990 			DRV_LOG(WARNING,
1991 				"Cannot get PCI address for netdev \"%s\".",
1992 				ifname);
1993 			continue;
1994 		}
1995 		/* Slave interface PCI address match found. */
1996 		snprintf(tmp_str, sizeof(tmp_str),
1997 			 "/sys/class/net/%s/phys_port_name", ifname);
1998 		file = fopen(tmp_str, "rb");
1999 		if (!file)
2000 			break;
2001 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2002 		if (fscanf(file, "%32s", tmp_str) == 1)
2003 			mlx5_translate_port_name(tmp_str, &info);
2004 		fclose(file);
2005 		/* Only process PF ports. */
2006 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2007 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2008 			continue;
2009 		/* Check max bonding member. */
2010 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2011 			DRV_LOG(WARNING, "bonding index out of range, "
2012 				"please increase MLX5_BOND_MAX_PORTS: %s",
2013 				tmp_str);
2014 			break;
2015 		}
2016 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
2017 		if (pci_dev->domain == pci_addr.domain &&
2018 		    pci_dev->bus == pci_addr.bus &&
2019 		    pci_dev->devid == pci_addr.devid &&
2020 		    ((pci_dev->function == 0 &&
2021 		      pci_dev->function + owner == pci_addr.function) ||
2022 		     (pci_dev->function == owner &&
2023 		      pci_addr.function == owner)))
2024 			pf = info.port_name;
2025 		/* Get ifindex. */
2026 		snprintf(tmp_str, sizeof(tmp_str),
2027 			 "/sys/class/net/%s/ifindex", ifname);
2028 		file = fopen(tmp_str, "rb");
2029 		if (!file)
2030 			break;
2031 		ret = fscanf(file, "%u", &ifindex);
2032 		fclose(file);
2033 		if (ret != 1)
2034 			break;
2035 		/* Save bonding info. */
2036 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2037 			sizeof(bond_info->ports[0].ifname));
2038 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2039 		bond_info->ports[info.port_name].ifindex = ifindex;
2040 		bond_info->n_port++;
2041 	}
2042 	if (pf >= 0) {
2043 		/* Get bond interface info */
2044 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2045 					   bond_info->ifname);
2046 		if (ret)
2047 			DRV_LOG(ERR, "unable to get bond info: %s",
2048 				strerror(rte_errno));
2049 		else
2050 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2051 				ifindex, bond_info->ifindex, bond_info->ifname);
2052 	}
2053 	return pf;
2054 }
2055 
2056 static void
2057 mlx5_os_config_default(struct mlx5_dev_config *config)
2058 {
2059 	memset(config, 0, sizeof(*config));
2060 	config->mps = MLX5_ARG_UNSET;
2061 	config->rx_vec_en = 1;
2062 	config->txq_inline_max = MLX5_ARG_UNSET;
2063 	config->txq_inline_min = MLX5_ARG_UNSET;
2064 	config->txq_inline_mpw = MLX5_ARG_UNSET;
2065 	config->txqs_inline = MLX5_ARG_UNSET;
2066 	config->vf_nl_en = 1;
2067 	config->mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2068 	config->mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2069 	config->dv_esw_en = 1;
2070 	config->dv_flow_en = 1;
2071 	config->decap_en = 1;
2072 	config->log_hp_size = MLX5_ARG_UNSET;
2073 	config->allow_duplicate_pattern = 1;
2074 }
2075 
2076 /**
2077  * Register a PCI device within bonding.
2078  *
2079  * This function spawns Ethernet devices out of a given PCI device and
2080  * bonding owner PF index.
2081  *
2082  * @param[in] cdev
2083  *   Pointer to common mlx5 device structure.
2084  * @param[in] req_eth_da
2085  *   Requested ethdev device argument.
2086  * @param[in] owner_id
2087  *   Requested owner PF port ID within bonding device, default to 0.
2088  *
2089  * @return
2090  *   0 on success, a negative errno value otherwise and rte_errno is set.
2091  */
2092 static int
2093 mlx5_os_pci_probe_pf(struct mlx5_common_device *cdev,
2094 		     struct rte_eth_devargs *req_eth_da,
2095 		     uint16_t owner_id)
2096 {
2097 	struct ibv_device **ibv_list;
2098 	/*
2099 	 * Number of found IB Devices matching with requested PCI BDF.
2100 	 * nd != 1 means there are multiple IB devices over the same
2101 	 * PCI device and we have representors and master.
2102 	 */
2103 	unsigned int nd = 0;
2104 	/*
2105 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2106 	 * we have the single multiport IB device, and there may be
2107 	 * representors attached to some of found ports.
2108 	 */
2109 	unsigned int np = 0;
2110 	/*
2111 	 * Number of DPDK ethernet devices to Spawn - either over
2112 	 * multiple IB devices or multiple ports of single IB device.
2113 	 * Actually this is the number of iterations to spawn.
2114 	 */
2115 	unsigned int ns = 0;
2116 	/*
2117 	 * Bonding device
2118 	 *   < 0 - no bonding device (single one)
2119 	 *  >= 0 - bonding device (value is slave PF index)
2120 	 */
2121 	int bd = -1;
2122 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2123 	struct mlx5_dev_spawn_data *list = NULL;
2124 	struct mlx5_dev_config dev_config;
2125 	unsigned int dev_config_vf;
2126 	struct rte_eth_devargs eth_da = *req_eth_da;
2127 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2128 	struct mlx5_bond_info bond_info;
2129 	int ret = -1;
2130 
2131 	errno = 0;
2132 	ibv_list = mlx5_glue->get_device_list(&ret);
2133 	if (!ibv_list) {
2134 		rte_errno = errno ? errno : ENOSYS;
2135 		DRV_LOG(ERR, "Cannot list devices, is ib_uverbs loaded?");
2136 		return -rte_errno;
2137 	}
2138 	/*
2139 	 * First scan the list of all Infiniband devices to find
2140 	 * matching ones, gathering into the list.
2141 	 */
2142 	struct ibv_device *ibv_match[ret + 1];
2143 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2144 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2145 	unsigned int i;
2146 
2147 	while (ret-- > 0) {
2148 		struct rte_pci_addr pci_addr;
2149 
2150 		DRV_LOG(DEBUG, "Checking device \"%s\"", ibv_list[ret]->name);
2151 		bd = mlx5_device_bond_pci_match(ibv_list[ret]->name, &owner_pci,
2152 						nl_rdma, owner_id, &bond_info);
2153 		if (bd >= 0) {
2154 			/*
2155 			 * Bonding device detected. Only one match is allowed,
2156 			 * the bonding is supported over multi-port IB device,
2157 			 * there should be no matches on representor PCI
2158 			 * functions or non VF LAG bonding devices with
2159 			 * specified address.
2160 			 */
2161 			if (nd) {
2162 				DRV_LOG(ERR,
2163 					"multiple PCI match on bonding device"
2164 					"\"%s\" found", ibv_list[ret]->name);
2165 				rte_errno = ENOENT;
2166 				ret = -rte_errno;
2167 				goto exit;
2168 			}
2169 			/* Amend owner pci address if owner PF ID specified. */
2170 			if (eth_da.nb_representor_ports)
2171 				owner_pci.function += owner_id;
2172 			DRV_LOG(INFO,
2173 				"PCI information matches for slave %d bonding device \"%s\"",
2174 				bd, ibv_list[ret]->name);
2175 			ibv_match[nd++] = ibv_list[ret];
2176 			break;
2177 		} else {
2178 			/* Bonding device not found. */
2179 			if (mlx5_get_pci_addr(ibv_list[ret]->ibdev_path,
2180 					      &pci_addr))
2181 				continue;
2182 			if (owner_pci.domain != pci_addr.domain ||
2183 			    owner_pci.bus != pci_addr.bus ||
2184 			    owner_pci.devid != pci_addr.devid ||
2185 			    owner_pci.function != pci_addr.function)
2186 				continue;
2187 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2188 				ibv_list[ret]->name);
2189 			ibv_match[nd++] = ibv_list[ret];
2190 		}
2191 	}
2192 	ibv_match[nd] = NULL;
2193 	if (!nd) {
2194 		/* No device matches, just complain and bail out. */
2195 		DRV_LOG(WARNING,
2196 			"No Verbs device matches PCI device " PCI_PRI_FMT ","
2197 			" are kernel drivers loaded?",
2198 			owner_pci.domain, owner_pci.bus,
2199 			owner_pci.devid, owner_pci.function);
2200 		rte_errno = ENOENT;
2201 		ret = -rte_errno;
2202 		goto exit;
2203 	}
2204 	if (nd == 1) {
2205 		/*
2206 		 * Found single matching device may have multiple ports.
2207 		 * Each port may be representor, we have to check the port
2208 		 * number and check the representors existence.
2209 		 */
2210 		if (nl_rdma >= 0)
2211 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2212 		if (!np)
2213 			DRV_LOG(WARNING,
2214 				"Cannot get IB device \"%s\" ports number.",
2215 				ibv_match[0]->name);
2216 		if (bd >= 0 && !np) {
2217 			DRV_LOG(ERR, "Cannot get ports for bonding device.");
2218 			rte_errno = ENOENT;
2219 			ret = -rte_errno;
2220 			goto exit;
2221 		}
2222 	}
2223 	/* Now we can determine the maximal amount of devices to be spawned. */
2224 	list = mlx5_malloc(MLX5_MEM_ZERO,
2225 			   sizeof(struct mlx5_dev_spawn_data) * (np ? np : nd),
2226 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2227 	if (!list) {
2228 		DRV_LOG(ERR, "Spawn data array allocation failure.");
2229 		rte_errno = ENOMEM;
2230 		ret = -rte_errno;
2231 		goto exit;
2232 	}
2233 	if (bd >= 0 || np > 1) {
2234 		/*
2235 		 * Single IB device with multiple ports found,
2236 		 * it may be E-Switch master device and representors.
2237 		 * We have to perform identification through the ports.
2238 		 */
2239 		MLX5_ASSERT(nl_rdma >= 0);
2240 		MLX5_ASSERT(ns == 0);
2241 		MLX5_ASSERT(nd == 1);
2242 		MLX5_ASSERT(np);
2243 		for (i = 1; i <= np; ++i) {
2244 			list[ns].bond_info = &bond_info;
2245 			list[ns].max_port = np;
2246 			list[ns].phys_port = i;
2247 			list[ns].phys_dev_name = ibv_match[0]->name;
2248 			list[ns].eth_dev = NULL;
2249 			list[ns].pci_dev = pci_dev;
2250 			list[ns].cdev = cdev;
2251 			list[ns].pf_bond = bd;
2252 			list[ns].ifindex = mlx5_nl_ifindex(nl_rdma,
2253 							   ibv_match[0]->name,
2254 							   i);
2255 			if (!list[ns].ifindex) {
2256 				/*
2257 				 * No network interface index found for the
2258 				 * specified port, it means there is no
2259 				 * representor on this port. It's OK,
2260 				 * there can be disabled ports, for example
2261 				 * if sriov_numvfs < sriov_totalvfs.
2262 				 */
2263 				continue;
2264 			}
2265 			ret = -1;
2266 			if (nl_route >= 0)
2267 				ret = mlx5_nl_switch_info(nl_route,
2268 							  list[ns].ifindex,
2269 							  &list[ns].info);
2270 			if (ret || (!list[ns].info.representor &&
2271 				    !list[ns].info.master)) {
2272 				/*
2273 				 * We failed to recognize representors with
2274 				 * Netlink, let's try to perform the task
2275 				 * with sysfs.
2276 				 */
2277 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2278 							     &list[ns].info);
2279 			}
2280 			if (!ret && bd >= 0) {
2281 				switch (list[ns].info.name_type) {
2282 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2283 					if (np == 1) {
2284 						/*
2285 						 * Force standalone bonding
2286 						 * device for ROCE LAG
2287 						 * confgiurations.
2288 						 */
2289 						list[ns].info.master = 0;
2290 						list[ns].info.representor = 0;
2291 					}
2292 					if (list[ns].info.port_name == bd)
2293 						ns++;
2294 					break;
2295 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2296 					/* Fallthrough */
2297 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2298 					/* Fallthrough */
2299 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2300 					if (list[ns].info.pf_num == bd)
2301 						ns++;
2302 					break;
2303 				default:
2304 					break;
2305 				}
2306 				continue;
2307 			}
2308 			if (!ret && (list[ns].info.representor ^
2309 				     list[ns].info.master))
2310 				ns++;
2311 		}
2312 		if (!ns) {
2313 			DRV_LOG(ERR,
2314 				"Unable to recognize master/representors on the IB device with multiple ports.");
2315 			rte_errno = ENOENT;
2316 			ret = -rte_errno;
2317 			goto exit;
2318 		}
2319 	} else {
2320 		/*
2321 		 * The existence of several matching entries (nd > 1) means
2322 		 * port representors have been instantiated. No existing Verbs
2323 		 * call nor sysfs entries can tell them apart, this can only
2324 		 * be done through Netlink calls assuming kernel drivers are
2325 		 * recent enough to support them.
2326 		 *
2327 		 * In the event of identification failure through Netlink,
2328 		 * try again through sysfs, then:
2329 		 *
2330 		 * 1. A single IB device matches (nd == 1) with single
2331 		 *    port (np=0/1) and is not a representor, assume
2332 		 *    no switch support.
2333 		 *
2334 		 * 2. Otherwise no safe assumptions can be made;
2335 		 *    complain louder and bail out.
2336 		 */
2337 		for (i = 0; i != nd; ++i) {
2338 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2339 			list[ns].bond_info = NULL;
2340 			list[ns].max_port = 1;
2341 			list[ns].phys_port = 1;
2342 			list[ns].phys_dev_name = ibv_match[i]->name;
2343 			list[ns].eth_dev = NULL;
2344 			list[ns].pci_dev = pci_dev;
2345 			list[ns].cdev = cdev;
2346 			list[ns].pf_bond = -1;
2347 			list[ns].ifindex = 0;
2348 			if (nl_rdma >= 0)
2349 				list[ns].ifindex = mlx5_nl_ifindex
2350 							    (nl_rdma,
2351 							     ibv_match[i]->name,
2352 							     1);
2353 			if (!list[ns].ifindex) {
2354 				char ifname[IF_NAMESIZE];
2355 
2356 				/*
2357 				 * Netlink failed, it may happen with old
2358 				 * ib_core kernel driver (before 4.16).
2359 				 * We can assume there is old driver because
2360 				 * here we are processing single ports IB
2361 				 * devices. Let's try sysfs to retrieve
2362 				 * the ifindex. The method works for
2363 				 * master device only.
2364 				 */
2365 				if (nd > 1) {
2366 					/*
2367 					 * Multiple devices found, assume
2368 					 * representors, can not distinguish
2369 					 * master/representor and retrieve
2370 					 * ifindex via sysfs.
2371 					 */
2372 					continue;
2373 				}
2374 				ret = mlx5_get_ifname_sysfs
2375 					(ibv_match[i]->ibdev_path, ifname);
2376 				if (!ret)
2377 					list[ns].ifindex =
2378 						if_nametoindex(ifname);
2379 				if (!list[ns].ifindex) {
2380 					/*
2381 					 * No network interface index found
2382 					 * for the specified device, it means
2383 					 * there it is neither representor
2384 					 * nor master.
2385 					 */
2386 					continue;
2387 				}
2388 			}
2389 			ret = -1;
2390 			if (nl_route >= 0)
2391 				ret = mlx5_nl_switch_info(nl_route,
2392 							  list[ns].ifindex,
2393 							  &list[ns].info);
2394 			if (ret || (!list[ns].info.representor &&
2395 				    !list[ns].info.master)) {
2396 				/*
2397 				 * We failed to recognize representors with
2398 				 * Netlink, let's try to perform the task
2399 				 * with sysfs.
2400 				 */
2401 				ret = mlx5_sysfs_switch_info(list[ns].ifindex,
2402 							     &list[ns].info);
2403 			}
2404 			if (!ret && (list[ns].info.representor ^
2405 				     list[ns].info.master)) {
2406 				ns++;
2407 			} else if ((nd == 1) &&
2408 				   !list[ns].info.representor &&
2409 				   !list[ns].info.master) {
2410 				/*
2411 				 * Single IB device with one physical port and
2412 				 * attached network device.
2413 				 * May be SRIOV is not enabled or there is no
2414 				 * representors.
2415 				 */
2416 				DRV_LOG(INFO, "No E-Switch support detected.");
2417 				ns++;
2418 				break;
2419 			}
2420 		}
2421 		if (!ns) {
2422 			DRV_LOG(ERR,
2423 				"Unable to recognize master/representors on the multiple IB devices.");
2424 			rte_errno = ENOENT;
2425 			ret = -rte_errno;
2426 			goto exit;
2427 		}
2428 		/*
2429 		 * New kernels may add the switch_id attribute for the case
2430 		 * there is no E-Switch and we wrongly recognized the only
2431 		 * device as master. Override this if there is the single
2432 		 * device with single port and new device name format present.
2433 		 */
2434 		if (nd == 1 &&
2435 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2436 			list[0].info.master = 0;
2437 			list[0].info.representor = 0;
2438 		}
2439 	}
2440 	MLX5_ASSERT(ns);
2441 	/*
2442 	 * Sort list to probe devices in natural order for users convenience
2443 	 * (i.e. master first, then representors from lowest to highest ID).
2444 	 */
2445 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2446 	/* Device specific configuration. */
2447 	switch (pci_dev->id.device_id) {
2448 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2449 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2450 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2451 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2452 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2453 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2454 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2455 		dev_config_vf = 1;
2456 		break;
2457 	default:
2458 		dev_config_vf = 0;
2459 		break;
2460 	}
2461 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2462 		/* Set devargs default values. */
2463 		if (eth_da.nb_mh_controllers == 0) {
2464 			eth_da.nb_mh_controllers = 1;
2465 			eth_da.mh_controllers[0] = 0;
2466 		}
2467 		if (eth_da.nb_ports == 0 && ns > 0) {
2468 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2469 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2470 					pci_dev->device.devargs->args);
2471 			eth_da.nb_ports = 1;
2472 			eth_da.ports[0] = list[0].info.pf_num;
2473 		}
2474 		if (eth_da.nb_representor_ports == 0) {
2475 			eth_da.nb_representor_ports = 1;
2476 			eth_da.representor_ports[0] = 0;
2477 		}
2478 	}
2479 	for (i = 0; i != ns; ++i) {
2480 		uint32_t restore;
2481 
2482 		/* Default configuration. */
2483 		mlx5_os_config_default(&dev_config);
2484 		dev_config.vf = dev_config_vf;
2485 		list[i].eth_dev = mlx5_dev_spawn(cdev->dev, &list[i],
2486 						 &dev_config, &eth_da);
2487 		if (!list[i].eth_dev) {
2488 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2489 				break;
2490 			/* Device is disabled or already spawned. Ignore it. */
2491 			continue;
2492 		}
2493 		restore = list[i].eth_dev->data->dev_flags;
2494 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2495 		/**
2496 		 * Each representor has a dedicated interrupts vector.
2497 		 * rte_eth_copy_pci_info() assigns PF interrupts handle to
2498 		 * representor eth_dev object because representor and PF
2499 		 * share the same PCI address.
2500 		 * Override representor device with a dedicated
2501 		 * interrupts handle here.
2502 		 * Representor interrupts handle is released in mlx5_dev_stop().
2503 		 */
2504 		if (list[i].info.representor) {
2505 			struct rte_intr_handle *intr_handle;
2506 			intr_handle = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
2507 						  sizeof(*intr_handle), 0,
2508 						  SOCKET_ID_ANY);
2509 			if (!intr_handle) {
2510 				DRV_LOG(ERR,
2511 					"port %u failed to allocate memory for interrupt handler "
2512 					"Rx interrupts will not be supported",
2513 					i);
2514 				rte_errno = ENOMEM;
2515 				ret = -rte_errno;
2516 				goto exit;
2517 			}
2518 			list[i].eth_dev->intr_handle = intr_handle;
2519 		}
2520 		/* Restore non-PCI flags cleared by the above call. */
2521 		list[i].eth_dev->data->dev_flags |= restore;
2522 		rte_eth_dev_probing_finish(list[i].eth_dev);
2523 	}
2524 	if (i != ns) {
2525 		DRV_LOG(ERR,
2526 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2527 			" encountering an error: %s",
2528 			owner_pci.domain, owner_pci.bus,
2529 			owner_pci.devid, owner_pci.function,
2530 			strerror(rte_errno));
2531 		ret = -rte_errno;
2532 		/* Roll back. */
2533 		while (i--) {
2534 			if (!list[i].eth_dev)
2535 				continue;
2536 			mlx5_dev_close(list[i].eth_dev);
2537 			/* mac_addrs must not be freed because in dev_private */
2538 			list[i].eth_dev->data->mac_addrs = NULL;
2539 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2540 		}
2541 		/* Restore original error. */
2542 		rte_errno = -ret;
2543 	} else {
2544 		ret = 0;
2545 	}
2546 exit:
2547 	/*
2548 	 * Do the routine cleanup:
2549 	 * - close opened Netlink sockets
2550 	 * - free allocated spawn data array
2551 	 * - free the Infiniband device list
2552 	 */
2553 	if (nl_rdma >= 0)
2554 		close(nl_rdma);
2555 	if (nl_route >= 0)
2556 		close(nl_route);
2557 	if (list)
2558 		mlx5_free(list);
2559 	MLX5_ASSERT(ibv_list);
2560 	mlx5_glue->free_device_list(ibv_list);
2561 	return ret;
2562 }
2563 
2564 static int
2565 mlx5_os_parse_eth_devargs(struct rte_device *dev,
2566 			  struct rte_eth_devargs *eth_da)
2567 {
2568 	int ret = 0;
2569 
2570 	if (dev->devargs == NULL)
2571 		return 0;
2572 	memset(eth_da, 0, sizeof(*eth_da));
2573 	/* Parse representor information first from class argument. */
2574 	if (dev->devargs->cls_str)
2575 		ret = rte_eth_devargs_parse(dev->devargs->cls_str, eth_da);
2576 	if (ret != 0) {
2577 		DRV_LOG(ERR, "failed to parse device arguments: %s",
2578 			dev->devargs->cls_str);
2579 		return -rte_errno;
2580 	}
2581 	if (eth_da->type == RTE_ETH_REPRESENTOR_NONE) {
2582 		/* Parse legacy device argument */
2583 		ret = rte_eth_devargs_parse(dev->devargs->args, eth_da);
2584 		if (ret) {
2585 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2586 				dev->devargs->args);
2587 			return -rte_errno;
2588 		}
2589 	}
2590 	return 0;
2591 }
2592 
2593 /**
2594  * Callback to register a PCI device.
2595  *
2596  * This function spawns Ethernet devices out of a given PCI device.
2597  *
2598  * @param[in] cdev
2599  *   Pointer to common mlx5 device structure.
2600  *
2601  * @return
2602  *   0 on success, a negative errno value otherwise and rte_errno is set.
2603  */
2604 static int
2605 mlx5_os_pci_probe(struct mlx5_common_device *cdev)
2606 {
2607 	struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(cdev->dev);
2608 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2609 	int ret = 0;
2610 	uint16_t p;
2611 
2612 	ret = mlx5_os_parse_eth_devargs(cdev->dev, &eth_da);
2613 	if (ret != 0)
2614 		return ret;
2615 
2616 	if (eth_da.nb_ports > 0) {
2617 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2618 		for (p = 0; p < eth_da.nb_ports; p++) {
2619 			ret = mlx5_os_pci_probe_pf(cdev, &eth_da,
2620 						   eth_da.ports[p]);
2621 			if (ret)
2622 				break;
2623 		}
2624 		if (ret) {
2625 			DRV_LOG(ERR, "Probe of PCI device " PCI_PRI_FMT " "
2626 				"aborted due to proding failure of PF %u",
2627 				pci_dev->addr.domain, pci_dev->addr.bus,
2628 				pci_dev->addr.devid, pci_dev->addr.function,
2629 				eth_da.ports[p]);
2630 			mlx5_net_remove(cdev);
2631 		}
2632 	} else {
2633 		ret = mlx5_os_pci_probe_pf(cdev, &eth_da, 0);
2634 	}
2635 	return ret;
2636 }
2637 
2638 /* Probe a single SF device on auxiliary bus, no representor support. */
2639 static int
2640 mlx5_os_auxiliary_probe(struct mlx5_common_device *cdev)
2641 {
2642 	struct rte_eth_devargs eth_da = { .nb_ports = 0 };
2643 	struct mlx5_dev_config config;
2644 	struct mlx5_dev_spawn_data spawn = { .pf_bond = -1 };
2645 	struct rte_device *dev = cdev->dev;
2646 	struct rte_auxiliary_device *adev = RTE_DEV_TO_AUXILIARY(dev);
2647 	struct rte_eth_dev *eth_dev;
2648 	int ret = 0;
2649 
2650 	/* Parse ethdev devargs. */
2651 	ret = mlx5_os_parse_eth_devargs(dev, &eth_da);
2652 	if (ret != 0)
2653 		return ret;
2654 	/* Set default config data. */
2655 	mlx5_os_config_default(&config);
2656 	config.sf = 1;
2657 	/* Init spawn data. */
2658 	spawn.max_port = 1;
2659 	spawn.phys_port = 1;
2660 	spawn.phys_dev_name = mlx5_os_get_ctx_device_name(cdev->ctx);
2661 	ret = mlx5_auxiliary_get_ifindex(dev->name);
2662 	if (ret < 0) {
2663 		DRV_LOG(ERR, "failed to get ethdev ifindex: %s", dev->name);
2664 		return ret;
2665 	}
2666 	spawn.ifindex = ret;
2667 	spawn.cdev = cdev;
2668 	/* Spawn device. */
2669 	eth_dev = mlx5_dev_spawn(dev, &spawn, &config, &eth_da);
2670 	if (eth_dev == NULL)
2671 		return -rte_errno;
2672 	/* Post create. */
2673 	eth_dev->intr_handle = &adev->intr_handle;
2674 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2675 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
2676 		eth_dev->data->dev_flags |= RTE_ETH_DEV_INTR_RMV;
2677 		eth_dev->data->numa_node = dev->numa_node;
2678 	}
2679 	rte_eth_dev_probing_finish(eth_dev);
2680 	return 0;
2681 }
2682 
2683 /**
2684  * Net class driver callback to probe a device.
2685  *
2686  * This function probe PCI bus device(s) or a single SF on auxiliary bus.
2687  *
2688  * @param[in] cdev
2689  *   Pointer to the common mlx5 device.
2690  *
2691  * @return
2692  *   0 on success, a negative errno value otherwise and rte_errno is set.
2693  */
2694 int
2695 mlx5_os_net_probe(struct mlx5_common_device *cdev)
2696 {
2697 	int ret;
2698 
2699 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2700 		mlx5_pmd_socket_init();
2701 	ret = mlx5_init_once();
2702 	if (ret) {
2703 		DRV_LOG(ERR, "Unable to init PMD global data: %s",
2704 			strerror(rte_errno));
2705 		return -rte_errno;
2706 	}
2707 	if (mlx5_dev_is_pci(cdev->dev))
2708 		return mlx5_os_pci_probe(cdev);
2709 	else
2710 		return mlx5_os_auxiliary_probe(cdev);
2711 }
2712 
2713 /**
2714  * Install shared asynchronous device events handler.
2715  * This function is implemented to support event sharing
2716  * between multiple ports of single IB device.
2717  *
2718  * @param sh
2719  *   Pointer to mlx5_dev_ctx_shared object.
2720  */
2721 void
2722 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2723 {
2724 	int ret;
2725 	int flags;
2726 	struct ibv_context *ctx = sh->cdev->ctx;
2727 
2728 	sh->intr_handle.fd = -1;
2729 	flags = fcntl(ctx->async_fd, F_GETFL);
2730 	ret = fcntl(ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2731 	if (ret) {
2732 		DRV_LOG(INFO, "failed to change file descriptor async event"
2733 			" queue");
2734 	} else {
2735 		sh->intr_handle.fd = ctx->async_fd;
2736 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2737 		if (rte_intr_callback_register(&sh->intr_handle,
2738 					mlx5_dev_interrupt_handler, sh)) {
2739 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2740 			sh->intr_handle.fd = -1;
2741 		}
2742 	}
2743 	if (sh->devx) {
2744 #ifdef HAVE_IBV_DEVX_ASYNC
2745 		sh->intr_handle_devx.fd = -1;
2746 		sh->devx_comp = (void *)mlx5_glue->devx_create_cmd_comp(ctx);
2747 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2748 		if (!devx_comp) {
2749 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2750 			return;
2751 		}
2752 		flags = fcntl(devx_comp->fd, F_GETFL);
2753 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2754 		if (ret) {
2755 			DRV_LOG(INFO, "failed to change file descriptor"
2756 				" devx comp");
2757 			return;
2758 		}
2759 		sh->intr_handle_devx.fd = devx_comp->fd;
2760 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2761 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2762 					mlx5_dev_interrupt_handler_devx, sh)) {
2763 			DRV_LOG(INFO, "Fail to install the devx shared"
2764 				" interrupt.");
2765 			sh->intr_handle_devx.fd = -1;
2766 		}
2767 #endif /* HAVE_IBV_DEVX_ASYNC */
2768 	}
2769 }
2770 
2771 /**
2772  * Uninstall shared asynchronous device events handler.
2773  * This function is implemented to support event sharing
2774  * between multiple ports of single IB device.
2775  *
2776  * @param dev
2777  *   Pointer to mlx5_dev_ctx_shared object.
2778  */
2779 void
2780 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2781 {
2782 	if (sh->intr_handle.fd >= 0)
2783 		mlx5_intr_callback_unregister(&sh->intr_handle,
2784 					      mlx5_dev_interrupt_handler, sh);
2785 #ifdef HAVE_IBV_DEVX_ASYNC
2786 	if (sh->intr_handle_devx.fd >= 0)
2787 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2788 				  mlx5_dev_interrupt_handler_devx, sh);
2789 	if (sh->devx_comp)
2790 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2791 #endif
2792 }
2793 
2794 /**
2795  * Read statistics by a named counter.
2796  *
2797  * @param[in] priv
2798  *   Pointer to the private device data structure.
2799  * @param[in] ctr_name
2800  *   Pointer to the name of the statistic counter to read
2801  * @param[out] stat
2802  *   Pointer to read statistic value.
2803  * @return
2804  *   0 on success and stat is valud, 1 if failed to read the value
2805  *   rte_errno is set.
2806  *
2807  */
2808 int
2809 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2810 		      uint64_t *stat)
2811 {
2812 	int fd;
2813 
2814 	if (priv->sh) {
2815 		if (priv->q_counters != NULL &&
2816 		    strcmp(ctr_name, "out_of_buffer") == 0)
2817 			return mlx5_devx_cmd_queue_counter_query
2818 					(priv->q_counters, 0, (uint32_t *)stat);
2819 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2820 		      priv->sh->ibdev_path,
2821 		      priv->dev_port,
2822 		      ctr_name);
2823 		fd = open(path, O_RDONLY);
2824 		/*
2825 		 * in switchdev the file location is not per port
2826 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2827 		 */
2828 		if (fd == -1) {
2829 			MKSTR(path1, "%s/hw_counters/%s",
2830 			      priv->sh->ibdev_path,
2831 			      ctr_name);
2832 			fd = open(path1, O_RDONLY);
2833 		}
2834 		if (fd != -1) {
2835 			char buf[21] = {'\0'};
2836 			ssize_t n = read(fd, buf, sizeof(buf));
2837 
2838 			close(fd);
2839 			if (n != -1) {
2840 				*stat = strtoull(buf, NULL, 10);
2841 				return 0;
2842 			}
2843 		}
2844 	}
2845 	*stat = 0;
2846 	return 1;
2847 }
2848 
2849 /**
2850  * Remove a MAC address from device
2851  *
2852  * @param dev
2853  *   Pointer to Ethernet device structure.
2854  * @param index
2855  *   MAC address index.
2856  */
2857 void
2858 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2859 {
2860 	struct mlx5_priv *priv = dev->data->dev_private;
2861 	const int vf = priv->config.vf;
2862 
2863 	if (vf)
2864 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2865 					mlx5_ifindex(dev), priv->mac_own,
2866 					&dev->data->mac_addrs[index], index);
2867 }
2868 
2869 /**
2870  * Adds a MAC address to the device
2871  *
2872  * @param dev
2873  *   Pointer to Ethernet device structure.
2874  * @param mac_addr
2875  *   MAC address to register.
2876  * @param index
2877  *   MAC address index.
2878  *
2879  * @return
2880  *   0 on success, a negative errno value otherwise
2881  */
2882 int
2883 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2884 		     uint32_t index)
2885 {
2886 	struct mlx5_priv *priv = dev->data->dev_private;
2887 	const int vf = priv->config.vf;
2888 	int ret = 0;
2889 
2890 	if (vf)
2891 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2892 					   mlx5_ifindex(dev), priv->mac_own,
2893 					   mac, index);
2894 	return ret;
2895 }
2896 
2897 /**
2898  * Modify a VF MAC address
2899  *
2900  * @param priv
2901  *   Pointer to device private data.
2902  * @param mac_addr
2903  *   MAC address to modify into.
2904  * @param iface_idx
2905  *   Net device interface index
2906  * @param vf_index
2907  *   VF index
2908  *
2909  * @return
2910  *   0 on success, a negative errno value otherwise
2911  */
2912 int
2913 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2914 			   unsigned int iface_idx,
2915 			   struct rte_ether_addr *mac_addr,
2916 			   int vf_index)
2917 {
2918 	return mlx5_nl_vf_mac_addr_modify
2919 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2920 }
2921 
2922 /**
2923  * Set device promiscuous mode
2924  *
2925  * @param dev
2926  *   Pointer to Ethernet device structure.
2927  * @param enable
2928  *   0 - promiscuous is disabled, otherwise - enabled
2929  *
2930  * @return
2931  *   0 on success, a negative error value otherwise
2932  */
2933 int
2934 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2935 {
2936 	struct mlx5_priv *priv = dev->data->dev_private;
2937 
2938 	return mlx5_nl_promisc(priv->nl_socket_route,
2939 			       mlx5_ifindex(dev), !!enable);
2940 }
2941 
2942 /**
2943  * Set device promiscuous mode
2944  *
2945  * @param dev
2946  *   Pointer to Ethernet device structure.
2947  * @param enable
2948  *   0 - all multicase is disabled, otherwise - enabled
2949  *
2950  * @return
2951  *   0 on success, a negative error value otherwise
2952  */
2953 int
2954 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2955 {
2956 	struct mlx5_priv *priv = dev->data->dev_private;
2957 
2958 	return mlx5_nl_allmulti(priv->nl_socket_route,
2959 				mlx5_ifindex(dev), !!enable);
2960 }
2961 
2962 /**
2963  * Flush device MAC addresses
2964  *
2965  * @param dev
2966  *   Pointer to Ethernet device structure.
2967  *
2968  */
2969 void
2970 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2971 {
2972 	struct mlx5_priv *priv = dev->data->dev_private;
2973 
2974 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2975 			       dev->data->mac_addrs,
2976 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2977 }
2978