xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision e78e5408da89b6e297e84b305b698d262cd5443a)
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_common.h>
24 #include <rte_kvargs.h>
25 #include <rte_rwlock.h>
26 #include <rte_spinlock.h>
27 #include <rte_string_fns.h>
28 #include <rte_alarm.h>
29 #include <rte_eal_paging.h>
30 
31 #include <mlx5_glue.h>
32 #include <mlx5_devx_cmds.h>
33 #include <mlx5_common.h>
34 #include <mlx5_common_mp.h>
35 #include <mlx5_common_mr.h>
36 #include <mlx5_malloc.h>
37 
38 #include "mlx5_defs.h"
39 #include "mlx5.h"
40 #include "mlx5_common_os.h"
41 #include "mlx5_utils.h"
42 #include "mlx5_rxtx.h"
43 #include "mlx5_rx.h"
44 #include "mlx5_tx.h"
45 #include "mlx5_autoconf.h"
46 #include "mlx5_mr.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 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
54 
55 #ifndef HAVE_IBV_MLX5_MOD_MPW
56 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
57 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
58 #endif
59 
60 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
61 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
62 #endif
63 
64 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
65 
66 /* Spinlock for mlx5_shared_data allocation. */
67 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
68 
69 /* Process local data for secondary processes. */
70 static struct mlx5_local_data mlx5_local_data;
71 
72 /* rte flow indexed pool configuration. */
73 static struct mlx5_indexed_pool_config icfg[] = {
74 	{
75 		.size = sizeof(struct rte_flow),
76 		.trunk_size = 64,
77 		.need_lock = 1,
78 		.release_mem_en = 0,
79 		.malloc = mlx5_malloc,
80 		.free = mlx5_free,
81 		.per_core_cache = 0,
82 		.type = "ctl_flow_ipool",
83 	},
84 	{
85 		.size = sizeof(struct rte_flow),
86 		.trunk_size = 64,
87 		.grow_trunk = 3,
88 		.grow_shift = 2,
89 		.need_lock = 1,
90 		.release_mem_en = 0,
91 		.malloc = mlx5_malloc,
92 		.free = mlx5_free,
93 		.per_core_cache = 1 << 14,
94 		.type = "rte_flow_ipool",
95 	},
96 	{
97 		.size = sizeof(struct rte_flow),
98 		.trunk_size = 64,
99 		.grow_trunk = 3,
100 		.grow_shift = 2,
101 		.need_lock = 1,
102 		.release_mem_en = 0,
103 		.malloc = mlx5_malloc,
104 		.free = mlx5_free,
105 		.per_core_cache = 0,
106 		.type = "mcp_flow_ipool",
107 	},
108 };
109 
110 /**
111  * Set the completion channel file descriptor interrupt as non-blocking.
112  *
113  * @param[in] rxq_obj
114  *   Pointer to RQ channel object, which includes the channel fd
115  *
116  * @param[out] fd
117  *   The file descriptor (representing the intetrrupt) used in this channel.
118  *
119  * @return
120  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
121  */
122 int
123 mlx5_os_set_nonblock_channel_fd(int fd)
124 {
125 	int flags;
126 
127 	flags = fcntl(fd, F_GETFL);
128 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
129 }
130 
131 /**
132  * Get mlx5 device attributes. The glue function query_device_ex() is called
133  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
134  * device attributes from the glue out parameter.
135  *
136  * @param dev
137  *   Pointer to ibv context.
138  *
139  * @param device_attr
140  *   Pointer to mlx5 device attributes.
141  *
142  * @return
143  *   0 on success, non zero error number otherwise
144  */
145 int
146 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
147 {
148 	int err;
149 	struct ibv_device_attr_ex attr_ex;
150 	memset(device_attr, 0, sizeof(*device_attr));
151 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
152 	if (err)
153 		return err;
154 
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  * Verbs callback to allocate a memory. This function should allocate the space
203  * according to the size provided residing inside a huge page.
204  * Please note that all allocation must respect the alignment from libmlx5
205  * (i.e. currently rte_mem_page_size()).
206  *
207  * @param[in] size
208  *   The size in bytes of the memory to allocate.
209  * @param[in] data
210  *   A pointer to the callback data.
211  *
212  * @return
213  *   Allocated buffer, NULL otherwise and rte_errno is set.
214  */
215 static void *
216 mlx5_alloc_verbs_buf(size_t size, void *data)
217 {
218 	struct mlx5_dev_ctx_shared *sh = data;
219 	void *ret;
220 	size_t alignment = rte_mem_page_size();
221 	if (alignment == (size_t)-1) {
222 		DRV_LOG(ERR, "Failed to get mem page size");
223 		rte_errno = ENOMEM;
224 		return NULL;
225 	}
226 
227 	MLX5_ASSERT(data != NULL);
228 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
229 	if (!ret && size)
230 		rte_errno = ENOMEM;
231 	return ret;
232 }
233 
234 /**
235  * Detect misc5 support or not
236  *
237  * @param[in] priv
238  *   Device private data pointer
239  */
240 #ifdef HAVE_MLX5DV_DR
241 static void
242 __mlx5_discovery_misc5_cap(struct mlx5_priv *priv)
243 {
244 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
245 	/* Dummy VxLAN matcher to detect rdma-core misc5 cap
246 	 * Case: IPv4--->UDP--->VxLAN--->vni
247 	 */
248 	void *tbl;
249 	struct mlx5_flow_dv_match_params matcher_mask;
250 	void *match_m;
251 	void *matcher;
252 	void *headers_m;
253 	void *misc5_m;
254 	uint32_t *tunnel_header_m;
255 	struct mlx5dv_flow_matcher_attr dv_attr;
256 
257 	memset(&matcher_mask, 0, sizeof(matcher_mask));
258 	matcher_mask.size = sizeof(matcher_mask.buf);
259 	match_m = matcher_mask.buf;
260 	headers_m = MLX5_ADDR_OF(fte_match_param, match_m, outer_headers);
261 	misc5_m = MLX5_ADDR_OF(fte_match_param,
262 			       match_m, misc_parameters_5);
263 	tunnel_header_m = (uint32_t *)
264 				MLX5_ADDR_OF(fte_match_set_misc5,
265 				misc5_m, tunnel_header_1);
266 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_protocol, 0xff);
267 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 4);
268 	MLX5_SET(fte_match_set_lyr_2_4, headers_m, udp_dport, 0xffff);
269 	*tunnel_header_m = 0xffffff;
270 
271 	tbl = mlx5_glue->dr_create_flow_tbl(priv->sh->rx_domain, 1);
272 	if (!tbl) {
273 		DRV_LOG(INFO, "No SW steering support");
274 		return;
275 	}
276 	dv_attr.type = IBV_FLOW_ATTR_NORMAL,
277 	dv_attr.match_mask = (void *)&matcher_mask,
278 	dv_attr.match_criteria_enable =
279 			(1 << MLX5_MATCH_CRITERIA_ENABLE_OUTER_BIT) |
280 			(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC5_BIT);
281 	dv_attr.priority = 3;
282 #ifdef HAVE_MLX5DV_DR_ESWITCH
283 	void *misc2_m;
284 	if (priv->config.dv_esw_en) {
285 		/* FDB enabled reg_c_0 */
286 		dv_attr.match_criteria_enable |=
287 				(1 << MLX5_MATCH_CRITERIA_ENABLE_MISC2_BIT);
288 		misc2_m = MLX5_ADDR_OF(fte_match_param,
289 				       match_m, misc_parameters_2);
290 		MLX5_SET(fte_match_set_misc2, misc2_m,
291 			 metadata_reg_c_0, 0xffff);
292 	}
293 #endif
294 	matcher = mlx5_glue->dv_create_flow_matcher(priv->sh->ctx,
295 						    &dv_attr, tbl);
296 	if (matcher) {
297 		priv->sh->misc5_cap = 1;
298 		mlx5_glue->dv_destroy_flow_matcher(matcher);
299 	}
300 	mlx5_glue->dr_destroy_flow_tbl(tbl);
301 #else
302 	RTE_SET_USED(priv);
303 #endif
304 }
305 #endif
306 
307 /**
308  * Verbs callback to free a memory.
309  *
310  * @param[in] ptr
311  *   A pointer to the memory to free.
312  * @param[in] data
313  *   A pointer to the callback data.
314  */
315 static void
316 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
317 {
318 	MLX5_ASSERT(data != NULL);
319 	mlx5_free(ptr);
320 }
321 
322 /**
323  * Initialize DR related data within private structure.
324  * Routine checks the reference counter and does actual
325  * resources creation/initialization only if counter is zero.
326  *
327  * @param[in] priv
328  *   Pointer to the private device data structure.
329  *
330  * @return
331  *   Zero on success, positive error code otherwise.
332  */
333 static int
334 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
335 {
336 	struct mlx5_dev_ctx_shared *sh = priv->sh;
337 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
338 	int err;
339 
340 	MLX5_ASSERT(sh && sh->refcnt);
341 	if (sh->refcnt > 1)
342 		return 0;
343 	err = mlx5_alloc_table_hash_list(priv);
344 	if (err)
345 		goto error;
346 	/* The resources below are only valid with DV support. */
347 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
348 	/* Init port id action mlx5 list. */
349 	snprintf(s, sizeof(s), "%s_port_id_action_list", sh->ibdev_name);
350 	mlx5_list_create(&sh->port_id_action_list, s, 0, sh,
351 			     flow_dv_port_id_create_cb,
352 			     flow_dv_port_id_match_cb,
353 			     flow_dv_port_id_remove_cb);
354 	/* Init push vlan action mlx5 list. */
355 	snprintf(s, sizeof(s), "%s_push_vlan_action_list", sh->ibdev_name);
356 	mlx5_list_create(&sh->push_vlan_action_list, s, 0, sh,
357 			     flow_dv_push_vlan_create_cb,
358 			     flow_dv_push_vlan_match_cb,
359 			     flow_dv_push_vlan_remove_cb);
360 	/* Init sample action mlx5 list. */
361 	snprintf(s, sizeof(s), "%s_sample_action_list", sh->ibdev_name);
362 	mlx5_list_create(&sh->sample_action_list, s, 0, sh,
363 			     flow_dv_sample_create_cb,
364 			     flow_dv_sample_match_cb,
365 			     flow_dv_sample_remove_cb);
366 	/* Init dest array action mlx5 list. */
367 	snprintf(s, sizeof(s), "%s_dest_array_list", sh->ibdev_name);
368 	mlx5_list_create(&sh->dest_array_list, s, 0, sh,
369 			     flow_dv_dest_array_create_cb,
370 			     flow_dv_dest_array_match_cb,
371 			     flow_dv_dest_array_remove_cb);
372 	/* Create tags hash list table. */
373 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
374 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
375 					  MLX5_HLIST_WRITE_MOST,
376 					  flow_dv_tag_create_cb,
377 					  flow_dv_tag_match_cb,
378 					  flow_dv_tag_remove_cb);
379 	if (!sh->tag_table) {
380 		DRV_LOG(ERR, "tags with hash creation failed.");
381 		err = ENOMEM;
382 		goto error;
383 	}
384 	sh->tag_table->ctx = sh;
385 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
386 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
387 					    0, MLX5_HLIST_WRITE_MOST |
388 					    MLX5_HLIST_DIRECT_KEY,
389 					    flow_dv_modify_create_cb,
390 					    flow_dv_modify_match_cb,
391 					    flow_dv_modify_remove_cb);
392 	if (!sh->modify_cmds) {
393 		DRV_LOG(ERR, "hdr modify hash creation failed");
394 		err = ENOMEM;
395 		goto error;
396 	}
397 	sh->modify_cmds->ctx = sh;
398 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
399 	sh->encaps_decaps = mlx5_hlist_create(s,
400 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
401 					      0, MLX5_HLIST_DIRECT_KEY |
402 					      MLX5_HLIST_WRITE_MOST,
403 					      flow_dv_encap_decap_create_cb,
404 					      flow_dv_encap_decap_match_cb,
405 					      flow_dv_encap_decap_remove_cb);
406 	if (!sh->encaps_decaps) {
407 		DRV_LOG(ERR, "encap decap hash creation failed");
408 		err = ENOMEM;
409 		goto error;
410 	}
411 	sh->encaps_decaps->ctx = sh;
412 #endif
413 #ifdef HAVE_MLX5DV_DR
414 	void *domain;
415 
416 	/* Reference counter is zero, we should initialize structures. */
417 	domain = mlx5_glue->dr_create_domain(sh->ctx,
418 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
419 	if (!domain) {
420 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
421 		err = errno;
422 		goto error;
423 	}
424 	sh->rx_domain = domain;
425 	domain = mlx5_glue->dr_create_domain(sh->ctx,
426 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
427 	if (!domain) {
428 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
429 		err = errno;
430 		goto error;
431 	}
432 	sh->tx_domain = domain;
433 #ifdef HAVE_MLX5DV_DR_ESWITCH
434 	if (priv->config.dv_esw_en) {
435 		domain  = mlx5_glue->dr_create_domain
436 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
437 		if (!domain) {
438 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
439 			err = errno;
440 			goto error;
441 		}
442 		sh->fdb_domain = domain;
443 	}
444 	/*
445 	 * The drop action is just some dummy placeholder in rdma-core. It
446 	 * does not belong to domains and has no any attributes, and, can be
447 	 * shared by the entire device.
448 	 */
449 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
450 	if (!sh->dr_drop_action) {
451 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
452 		err = errno;
453 		goto error;
454 	}
455 #endif
456 	if (!sh->tunnel_hub)
457 		err = mlx5_alloc_tunnel_hub(sh);
458 	if (err) {
459 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
460 		goto error;
461 	}
462 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
463 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
464 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
465 		if (sh->fdb_domain)
466 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
467 	}
468 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
469 	if (!priv->config.allow_duplicate_pattern) {
470 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
471 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
472 #endif
473 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
474 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
475 		if (sh->fdb_domain)
476 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
477 	}
478 
479 	__mlx5_discovery_misc5_cap(priv);
480 #endif /* HAVE_MLX5DV_DR */
481 	sh->default_miss_action =
482 			mlx5_glue->dr_create_flow_action_default_miss();
483 	if (!sh->default_miss_action)
484 		DRV_LOG(WARNING, "Default miss action is not supported.");
485 	return 0;
486 error:
487 	/* Rollback the created objects. */
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 	if (sh->fdb_domain) {
497 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
498 		sh->fdb_domain = NULL;
499 	}
500 	if (sh->dr_drop_action) {
501 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
502 		sh->dr_drop_action = NULL;
503 	}
504 	if (sh->pop_vlan_action) {
505 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
506 		sh->pop_vlan_action = NULL;
507 	}
508 	if (sh->encaps_decaps) {
509 		mlx5_hlist_destroy(sh->encaps_decaps);
510 		sh->encaps_decaps = NULL;
511 	}
512 	if (sh->modify_cmds) {
513 		mlx5_hlist_destroy(sh->modify_cmds);
514 		sh->modify_cmds = NULL;
515 	}
516 	if (sh->tag_table) {
517 		/* tags should be destroyed with flow before. */
518 		mlx5_hlist_destroy(sh->tag_table);
519 		sh->tag_table = NULL;
520 	}
521 	if (sh->tunnel_hub) {
522 		mlx5_release_tunnel_hub(sh, priv->dev_port);
523 		sh->tunnel_hub = NULL;
524 	}
525 	mlx5_free_table_hash_list(priv);
526 	return err;
527 }
528 
529 /**
530  * Destroy DR related data within private structure.
531  *
532  * @param[in] priv
533  *   Pointer to the private device data structure.
534  */
535 void
536 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
537 {
538 	struct mlx5_dev_ctx_shared *sh = priv->sh;
539 
540 	MLX5_ASSERT(sh && sh->refcnt);
541 	if (sh->refcnt > 1)
542 		return;
543 #ifdef HAVE_MLX5DV_DR
544 	if (sh->rx_domain) {
545 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
546 		sh->rx_domain = NULL;
547 	}
548 	if (sh->tx_domain) {
549 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
550 		sh->tx_domain = NULL;
551 	}
552 #ifdef HAVE_MLX5DV_DR_ESWITCH
553 	if (sh->fdb_domain) {
554 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
555 		sh->fdb_domain = NULL;
556 	}
557 	if (sh->dr_drop_action) {
558 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
559 		sh->dr_drop_action = NULL;
560 	}
561 #endif
562 	if (sh->pop_vlan_action) {
563 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
564 		sh->pop_vlan_action = NULL;
565 	}
566 #endif /* HAVE_MLX5DV_DR */
567 	if (sh->default_miss_action)
568 		mlx5_glue->destroy_flow_action
569 				(sh->default_miss_action);
570 	if (sh->encaps_decaps) {
571 		mlx5_hlist_destroy(sh->encaps_decaps);
572 		sh->encaps_decaps = NULL;
573 	}
574 	if (sh->modify_cmds) {
575 		mlx5_hlist_destroy(sh->modify_cmds);
576 		sh->modify_cmds = NULL;
577 	}
578 	if (sh->tag_table) {
579 		/* tags should be destroyed with flow before. */
580 		mlx5_hlist_destroy(sh->tag_table);
581 		sh->tag_table = NULL;
582 	}
583 	if (sh->tunnel_hub) {
584 		mlx5_release_tunnel_hub(sh, priv->dev_port);
585 		sh->tunnel_hub = NULL;
586 	}
587 	mlx5_list_destroy(&sh->port_id_action_list);
588 	mlx5_list_destroy(&sh->push_vlan_action_list);
589 	mlx5_free_table_hash_list(priv);
590 }
591 
592 /**
593  * Initialize shared data between primary and secondary process.
594  *
595  * A memzone is reserved by primary process and secondary processes attach to
596  * the memzone.
597  *
598  * @return
599  *   0 on success, a negative errno value otherwise and rte_errno is set.
600  */
601 static int
602 mlx5_init_shared_data(void)
603 {
604 	const struct rte_memzone *mz;
605 	int ret = 0;
606 
607 	rte_spinlock_lock(&mlx5_shared_data_lock);
608 	if (mlx5_shared_data == NULL) {
609 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
610 			/* Allocate shared memory. */
611 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
612 						 sizeof(*mlx5_shared_data),
613 						 SOCKET_ID_ANY, 0);
614 			if (mz == NULL) {
615 				DRV_LOG(ERR,
616 					"Cannot allocate mlx5 shared data");
617 				ret = -rte_errno;
618 				goto error;
619 			}
620 			mlx5_shared_data = mz->addr;
621 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
622 			rte_spinlock_init(&mlx5_shared_data->lock);
623 		} else {
624 			/* Lookup allocated shared memory. */
625 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
626 			if (mz == NULL) {
627 				DRV_LOG(ERR,
628 					"Cannot attach mlx5 shared data");
629 				ret = -rte_errno;
630 				goto error;
631 			}
632 			mlx5_shared_data = mz->addr;
633 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
634 		}
635 	}
636 error:
637 	rte_spinlock_unlock(&mlx5_shared_data_lock);
638 	return ret;
639 }
640 
641 /**
642  * PMD global initialization.
643  *
644  * Independent from individual device, this function initializes global
645  * per-PMD data structures distinguishing primary and secondary processes.
646  * Hence, each initialization is called once per a process.
647  *
648  * @return
649  *   0 on success, a negative errno value otherwise and rte_errno is set.
650  */
651 static int
652 mlx5_init_once(void)
653 {
654 	struct mlx5_shared_data *sd;
655 	struct mlx5_local_data *ld = &mlx5_local_data;
656 	int ret = 0;
657 
658 	if (mlx5_init_shared_data())
659 		return -rte_errno;
660 	sd = mlx5_shared_data;
661 	MLX5_ASSERT(sd);
662 	rte_spinlock_lock(&sd->lock);
663 	switch (rte_eal_process_type()) {
664 	case RTE_PROC_PRIMARY:
665 		if (sd->init_done)
666 			break;
667 		LIST_INIT(&sd->mem_event_cb_list);
668 		rte_rwlock_init(&sd->mem_event_rwlock);
669 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
670 						mlx5_mr_mem_event_cb, NULL);
671 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
672 					   mlx5_mp_os_primary_handle);
673 		if (ret)
674 			goto out;
675 		sd->init_done = true;
676 		break;
677 	case RTE_PROC_SECONDARY:
678 		if (ld->init_done)
679 			break;
680 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
681 					     mlx5_mp_os_secondary_handle);
682 		if (ret)
683 			goto out;
684 		++sd->secondary_cnt;
685 		ld->init_done = true;
686 		break;
687 	default:
688 		break;
689 	}
690 out:
691 	rte_spinlock_unlock(&sd->lock);
692 	return ret;
693 }
694 
695 /**
696  * Create the Tx queue DevX/Verbs object.
697  *
698  * @param dev
699  *   Pointer to Ethernet device.
700  * @param idx
701  *   Queue index in DPDK Tx queue array.
702  *
703  * @return
704  *   0 on success, a negative errno value otherwise and rte_errno is set.
705  */
706 static int
707 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
708 {
709 	struct mlx5_priv *priv = dev->data->dev_private;
710 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
711 	struct mlx5_txq_ctrl *txq_ctrl =
712 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
713 
714 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
715 		return mlx5_txq_devx_obj_new(dev, idx);
716 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
717 	if (!priv->config.dv_esw_en)
718 		return mlx5_txq_devx_obj_new(dev, idx);
719 #endif
720 	return mlx5_txq_ibv_obj_new(dev, idx);
721 }
722 
723 /**
724  * Release an Tx DevX/verbs queue object.
725  *
726  * @param txq_obj
727  *   DevX/Verbs Tx queue object.
728  */
729 static void
730 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
731 {
732 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
733 		mlx5_txq_devx_obj_release(txq_obj);
734 		return;
735 	}
736 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
737 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
738 		mlx5_txq_devx_obj_release(txq_obj);
739 		return;
740 	}
741 #endif
742 	mlx5_txq_ibv_obj_release(txq_obj);
743 }
744 
745 /**
746  * DV flow counter mode detect and config.
747  *
748  * @param dev
749  *   Pointer to rte_eth_dev structure.
750  *
751  */
752 static void
753 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
754 {
755 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
756 	struct mlx5_priv *priv = dev->data->dev_private;
757 	struct mlx5_dev_ctx_shared *sh = priv->sh;
758 	bool fallback;
759 
760 #ifndef HAVE_IBV_DEVX_ASYNC
761 	fallback = true;
762 #else
763 	fallback = false;
764 	if (!priv->config.devx || !priv->config.dv_flow_en ||
765 	    !priv->config.hca_attr.flow_counters_dump ||
766 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
767 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
768 		fallback = true;
769 #endif
770 	if (fallback)
771 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
772 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
773 			priv->config.hca_attr.flow_counters_dump,
774 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
775 	/* Initialize fallback mode only on the port initializes sh. */
776 	if (sh->refcnt == 1)
777 		sh->cmng.counter_fallback = fallback;
778 	else if (fallback != sh->cmng.counter_fallback)
779 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
780 			"with others:%d.", PORT_ID(priv), fallback);
781 #endif
782 }
783 
784 static void
785 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
786 {
787 	struct mlx5_priv *priv = dev->data->dev_private;
788 	void *ctx = priv->sh->ctx;
789 
790 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
791 	if (!priv->q_counters) {
792 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
793 		struct ibv_wq *wq;
794 
795 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
796 			"by DevX - fall-back to use the kernel driver global "
797 			"queue counter.", dev->data->port_id);
798 		/* Create WQ by kernel and query its queue counter ID. */
799 		if (cq) {
800 			wq = mlx5_glue->create_wq(ctx,
801 						  &(struct ibv_wq_init_attr){
802 						    .wq_type = IBV_WQT_RQ,
803 						    .max_wr = 1,
804 						    .max_sge = 1,
805 						    .pd = priv->sh->pd,
806 						    .cq = cq,
807 						});
808 			if (wq) {
809 				/* Counter is assigned only on RDY state. */
810 				int ret = mlx5_glue->modify_wq(wq,
811 						 &(struct ibv_wq_attr){
812 						 .attr_mask = IBV_WQ_ATTR_STATE,
813 						 .wq_state = IBV_WQS_RDY,
814 						});
815 
816 				if (ret == 0)
817 					mlx5_devx_cmd_wq_query(wq,
818 							 &priv->counter_set_id);
819 				claim_zero(mlx5_glue->destroy_wq(wq));
820 			}
821 			claim_zero(mlx5_glue->destroy_cq(cq));
822 		}
823 	} else {
824 		priv->counter_set_id = priv->q_counters->id;
825 	}
826 	if (priv->counter_set_id == 0)
827 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
828 			"available.", dev->data->port_id);
829 }
830 
831 /**
832  * Check if representor spawn info match devargs.
833  *
834  * @param spawn
835  *   Verbs device parameters (name, port, switch_info) to spawn.
836  * @param eth_da
837  *   Device devargs to probe.
838  *
839  * @return
840  *   Match result.
841  */
842 static bool
843 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
844 		       struct rte_eth_devargs *eth_da)
845 {
846 	struct mlx5_switch_info *switch_info = &spawn->info;
847 	unsigned int p, f;
848 	uint16_t id;
849 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
850 						      eth_da->type);
851 
852 	switch (eth_da->type) {
853 	case RTE_ETH_REPRESENTOR_SF:
854 		if (!(spawn->info.port_name == -1 &&
855 		      switch_info->name_type ==
856 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
857 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
858 			rte_errno = EBUSY;
859 			return false;
860 		}
861 		break;
862 	case RTE_ETH_REPRESENTOR_VF:
863 		/* Allows HPF representor index -1 as exception. */
864 		if (!(spawn->info.port_name == -1 &&
865 		      switch_info->name_type ==
866 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
867 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
868 			rte_errno = EBUSY;
869 			return false;
870 		}
871 		break;
872 	case RTE_ETH_REPRESENTOR_NONE:
873 		rte_errno = EBUSY;
874 		return false;
875 	default:
876 		rte_errno = ENOTSUP;
877 		DRV_LOG(ERR, "unsupported representor type");
878 		return false;
879 	}
880 	/* Check representor ID: */
881 	for (p = 0; p < eth_da->nb_ports; ++p) {
882 		if (spawn->pf_bond < 0) {
883 			/* For non-LAG mode, allow and ignore pf. */
884 			switch_info->pf_num = eth_da->ports[p];
885 			repr_id = mlx5_representor_id_encode(switch_info,
886 							     eth_da->type);
887 		}
888 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
889 			id = MLX5_REPRESENTOR_ID
890 				(eth_da->ports[p], eth_da->type,
891 				 eth_da->representor_ports[f]);
892 			if (repr_id == id)
893 				return true;
894 		}
895 	}
896 	rte_errno = EBUSY;
897 	return false;
898 }
899 
900 
901 /**
902  * Spawn an Ethernet device from Verbs information.
903  *
904  * @param dpdk_dev
905  *   Backing DPDK device.
906  * @param spawn
907  *   Verbs device parameters (name, port, switch_info) to spawn.
908  * @param config
909  *   Device configuration parameters.
910  * @param config
911  *   Device arguments.
912  *
913  * @return
914  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
915  *   is set. The following errors are defined:
916  *
917  *   EBUSY: device is not supposed to be spawned.
918  *   EEXIST: device is already spawned
919  */
920 static struct rte_eth_dev *
921 mlx5_dev_spawn(struct rte_device *dpdk_dev,
922 	       struct mlx5_dev_spawn_data *spawn,
923 	       struct mlx5_dev_config *config,
924 	       struct rte_eth_devargs *eth_da)
925 {
926 	const struct mlx5_switch_info *switch_info = &spawn->info;
927 	struct mlx5_dev_ctx_shared *sh = NULL;
928 	struct ibv_port_attr port_attr;
929 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
930 	struct rte_eth_dev *eth_dev = NULL;
931 	struct mlx5_priv *priv = NULL;
932 	int err = 0;
933 	unsigned int hw_padding = 0;
934 	unsigned int mps;
935 	unsigned int tunnel_en = 0;
936 	unsigned int mpls_en = 0;
937 	unsigned int swp = 0;
938 	unsigned int mprq = 0;
939 	unsigned int mprq_min_stride_size_n = 0;
940 	unsigned int mprq_max_stride_size_n = 0;
941 	unsigned int mprq_min_stride_num_n = 0;
942 	unsigned int mprq_max_stride_num_n = 0;
943 	struct rte_ether_addr mac;
944 	char name[RTE_ETH_NAME_MAX_LEN];
945 	int own_domain_id = 0;
946 	uint16_t port_id;
947 	struct mlx5_port_info vport_info = { .query_flags = 0 };
948 	int i;
949 
950 	/* Determine if this port representor is supposed to be spawned. */
951 	if (switch_info->representor && dpdk_dev->devargs &&
952 	    !mlx5_representor_match(spawn, eth_da))
953 		return NULL;
954 	/* Build device name. */
955 	if (spawn->pf_bond < 0) {
956 		/* Single device. */
957 		if (!switch_info->representor)
958 			strlcpy(name, dpdk_dev->name, sizeof(name));
959 		else
960 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
961 				 dpdk_dev->name,
962 				 switch_info->name_type ==
963 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
964 				 switch_info->port_name);
965 	} else {
966 		/* Bonding device. */
967 		if (!switch_info->representor) {
968 			err = snprintf(name, sizeof(name), "%s_%s",
969 				 dpdk_dev->name,
970 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
971 		} else {
972 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
973 				dpdk_dev->name,
974 				mlx5_os_get_dev_device_name(spawn->phys_dev),
975 				switch_info->ctrl_num,
976 				switch_info->pf_num,
977 				switch_info->name_type ==
978 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
979 				switch_info->port_name);
980 		}
981 	}
982 	if (err >= (int)sizeof(name))
983 		DRV_LOG(WARNING, "device name overflow %s", name);
984 	/* check if the device is already spawned */
985 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
986 		rte_errno = EEXIST;
987 		return NULL;
988 	}
989 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
990 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
991 		struct mlx5_mp_id mp_id;
992 
993 		eth_dev = rte_eth_dev_attach_secondary(name);
994 		if (eth_dev == NULL) {
995 			DRV_LOG(ERR, "can not attach rte ethdev");
996 			rte_errno = ENOMEM;
997 			return NULL;
998 		}
999 		eth_dev->device = dpdk_dev;
1000 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
1001 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1002 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1003 		err = mlx5_proc_priv_init(eth_dev);
1004 		if (err)
1005 			return NULL;
1006 		mp_id.port_id = eth_dev->data->port_id;
1007 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1008 		/* Receive command fd from primary process */
1009 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
1010 		if (err < 0)
1011 			goto err_secondary;
1012 		/* Remap UAR for Tx queues. */
1013 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
1014 		if (err)
1015 			goto err_secondary;
1016 		/*
1017 		 * Ethdev pointer is still required as input since
1018 		 * the primary device is not accessible from the
1019 		 * secondary process.
1020 		 */
1021 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1022 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1023 		return eth_dev;
1024 err_secondary:
1025 		mlx5_dev_close(eth_dev);
1026 		return NULL;
1027 	}
1028 	/*
1029 	 * Some parameters ("tx_db_nc" in particularly) are needed in
1030 	 * advance to create dv/verbs device context. We proceed the
1031 	 * devargs here to get ones, and later proceed devargs again
1032 	 * to override some hardware settings.
1033 	 */
1034 	err = mlx5_args(config, dpdk_dev->devargs);
1035 	if (err) {
1036 		err = rte_errno;
1037 		DRV_LOG(ERR, "failed to process device arguments: %s",
1038 			strerror(rte_errno));
1039 		goto error;
1040 	}
1041 	if (config->dv_miss_info) {
1042 		if (switch_info->master || switch_info->representor)
1043 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
1044 	}
1045 	mlx5_malloc_mem_select(config->sys_mem_en);
1046 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
1047 	if (!sh)
1048 		return NULL;
1049 	config->devx = sh->devx;
1050 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1051 	config->dest_tir = 1;
1052 #endif
1053 #ifdef HAVE_IBV_MLX5_MOD_SWP
1054 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1055 #endif
1056 	/*
1057 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
1058 	 * as all ConnectX-5 devices.
1059 	 */
1060 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1061 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1062 #endif
1063 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1064 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1065 #endif
1066 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1067 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1068 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1069 			DRV_LOG(DEBUG, "enhanced MPW is supported");
1070 			mps = MLX5_MPW_ENHANCED;
1071 		} else {
1072 			DRV_LOG(DEBUG, "MPW is supported");
1073 			mps = MLX5_MPW;
1074 		}
1075 	} else {
1076 		DRV_LOG(DEBUG, "MPW isn't supported");
1077 		mps = MLX5_MPW_DISABLED;
1078 	}
1079 #ifdef HAVE_IBV_MLX5_MOD_SWP
1080 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1081 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1082 	DRV_LOG(DEBUG, "SWP support: %u", swp);
1083 #endif
1084 	config->swp = !!swp;
1085 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1086 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1087 		struct mlx5dv_striding_rq_caps mprq_caps =
1088 			dv_attr.striding_rq_caps;
1089 
1090 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1091 			mprq_caps.min_single_stride_log_num_of_bytes);
1092 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1093 			mprq_caps.max_single_stride_log_num_of_bytes);
1094 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1095 			mprq_caps.min_single_wqe_log_num_of_strides);
1096 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1097 			mprq_caps.max_single_wqe_log_num_of_strides);
1098 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1099 			mprq_caps.supported_qpts);
1100 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1101 		mprq = 1;
1102 		mprq_min_stride_size_n =
1103 			mprq_caps.min_single_stride_log_num_of_bytes;
1104 		mprq_max_stride_size_n =
1105 			mprq_caps.max_single_stride_log_num_of_bytes;
1106 		mprq_min_stride_num_n =
1107 			mprq_caps.min_single_wqe_log_num_of_strides;
1108 		mprq_max_stride_num_n =
1109 			mprq_caps.max_single_wqe_log_num_of_strides;
1110 	}
1111 #endif
1112 	/* Rx CQE compression is enabled by default. */
1113 	config->cqe_comp = 1;
1114 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1115 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1116 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1117 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1118 			     (dv_attr.tunnel_offloads_caps &
1119 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
1120 			     (dv_attr.tunnel_offloads_caps &
1121 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1122 	}
1123 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1124 		tunnel_en ? "" : "not ");
1125 #else
1126 	DRV_LOG(WARNING,
1127 		"tunnel offloading disabled due to old OFED/rdma-core version");
1128 #endif
1129 	config->tunnel_en = tunnel_en;
1130 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1131 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1132 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1133 		   (dv_attr.tunnel_offloads_caps &
1134 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1135 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1136 		mpls_en ? "" : "not ");
1137 #else
1138 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1139 		" old OFED/rdma-core version or firmware configuration");
1140 #endif
1141 	config->mpls_en = mpls_en;
1142 	/* Check port status. */
1143 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1144 	if (err) {
1145 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1146 		goto error;
1147 	}
1148 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1149 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1150 		err = EINVAL;
1151 		goto error;
1152 	}
1153 	if (port_attr.state != IBV_PORT_ACTIVE)
1154 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1155 			mlx5_glue->port_state_str(port_attr.state),
1156 			port_attr.state);
1157 	/* Allocate private eth device data. */
1158 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1159 			   sizeof(*priv),
1160 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1161 	if (priv == NULL) {
1162 		DRV_LOG(ERR, "priv allocation failure");
1163 		err = ENOMEM;
1164 		goto error;
1165 	}
1166 	priv->sh = sh;
1167 	priv->dev_port = spawn->phys_port;
1168 	priv->pci_dev = spawn->pci_dev;
1169 	priv->mtu = RTE_ETHER_MTU;
1170 	/* Some internal functions rely on Netlink sockets, open them now. */
1171 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1172 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1173 	priv->representor = !!switch_info->representor;
1174 	priv->master = !!switch_info->master;
1175 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1176 	priv->vport_meta_tag = 0;
1177 	priv->vport_meta_mask = 0;
1178 	priv->pf_bond = spawn->pf_bond;
1179 	/*
1180 	 * If we have E-Switch we should determine the vport attributes.
1181 	 * E-Switch may use either source vport field or reg_c[0] metadata
1182 	 * register to match on vport index. The engaged part of metadata
1183 	 * register is defined by mask.
1184 	 */
1185 	if (switch_info->representor || switch_info->master) {
1186 		err = mlx5_glue->devx_port_query(sh->ctx,
1187 						 spawn->phys_port,
1188 						 &vport_info);
1189 		if (err) {
1190 			DRV_LOG(WARNING,
1191 				"can't query devx port %d on device %s",
1192 				spawn->phys_port,
1193 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1194 			vport_info.query_flags = 0;
1195 		}
1196 	}
1197 	if (vport_info.query_flags & MLX5_PORT_QUERY_REG_C0) {
1198 		priv->vport_meta_tag = vport_info.vport_meta_tag;
1199 		priv->vport_meta_mask = vport_info.vport_meta_mask;
1200 		if (!priv->vport_meta_mask) {
1201 			DRV_LOG(ERR, "vport zero mask for port %d"
1202 				     " on bonding device %s",
1203 				     spawn->phys_port,
1204 				     mlx5_os_get_dev_device_name
1205 							(spawn->phys_dev));
1206 			err = ENOTSUP;
1207 			goto error;
1208 		}
1209 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1210 			DRV_LOG(ERR, "invalid vport tag for port %d"
1211 				     " on bonding device %s",
1212 				     spawn->phys_port,
1213 				     mlx5_os_get_dev_device_name
1214 							(spawn->phys_dev));
1215 			err = ENOTSUP;
1216 			goto error;
1217 		}
1218 	}
1219 	if (vport_info.query_flags & MLX5_PORT_QUERY_VPORT) {
1220 		priv->vport_id = vport_info.vport_id;
1221 	} else if (spawn->pf_bond >= 0 &&
1222 		   (switch_info->representor || switch_info->master)) {
1223 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1224 			     " on bonding device %s",
1225 			     spawn->phys_port,
1226 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1227 		err = ENOTSUP;
1228 		goto error;
1229 	} else {
1230 		/*
1231 		 * Suppose vport index in compatible way. Kernel/rdma_core
1232 		 * support single E-Switch per PF configurations only and
1233 		 * vport_id field contains the vport index for associated VF,
1234 		 * which is deduced from representor port name.
1235 		 * For example, let's have the IB device port 10, it has
1236 		 * attached network device eth0, which has port name attribute
1237 		 * pf0vf2, we can deduce the VF number as 2, and set vport index
1238 		 * as 3 (2+1). This assigning schema should be changed if the
1239 		 * multiple E-Switch instances per PF configurations or/and PCI
1240 		 * subfunctions are added.
1241 		 */
1242 		priv->vport_id = switch_info->representor ?
1243 				 switch_info->port_name + 1 : -1;
1244 	}
1245 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1246 							  eth_da->type);
1247 	/*
1248 	 * Look for sibling devices in order to reuse their switch domain
1249 	 * if any, otherwise allocate one.
1250 	 */
1251 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1252 		const struct mlx5_priv *opriv =
1253 			rte_eth_devices[port_id].data->dev_private;
1254 
1255 		if (!opriv ||
1256 		    opriv->sh != priv->sh ||
1257 			opriv->domain_id ==
1258 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1259 			continue;
1260 		priv->domain_id = opriv->domain_id;
1261 		break;
1262 	}
1263 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1264 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1265 		if (err) {
1266 			err = rte_errno;
1267 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1268 				strerror(rte_errno));
1269 			goto error;
1270 		}
1271 		own_domain_id = 1;
1272 	}
1273 	/* Override some values set by hardware configuration. */
1274 	mlx5_args(config, dpdk_dev->devargs);
1275 	err = mlx5_dev_check_sibling_config(priv, config);
1276 	if (err)
1277 		goto error;
1278 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1279 			    IBV_DEVICE_RAW_IP_CSUM);
1280 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1281 		(config->hw_csum ? "" : "not "));
1282 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1283 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1284 	DRV_LOG(DEBUG, "counters are not supported");
1285 #endif
1286 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1287 	if (config->dv_flow_en) {
1288 		DRV_LOG(WARNING, "DV flow is not supported");
1289 		config->dv_flow_en = 0;
1290 	}
1291 #endif
1292 	config->ind_table_max_size =
1293 		sh->device_attr.max_rwq_indirection_table_size;
1294 	/*
1295 	 * Remove this check once DPDK supports larger/variable
1296 	 * indirection tables.
1297 	 */
1298 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1299 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1300 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1301 		config->ind_table_max_size);
1302 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1303 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1304 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1305 		(config->hw_vlan_strip ? "" : "not "));
1306 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1307 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1308 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1309 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1310 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1311 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1312 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1313 #endif
1314 	if (config->hw_padding && !hw_padding) {
1315 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1316 		config->hw_padding = 0;
1317 	} else if (config->hw_padding) {
1318 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1319 	}
1320 	config->tso = (sh->device_attr.max_tso > 0 &&
1321 		      (sh->device_attr.tso_supported_qpts &
1322 		       (1 << IBV_QPT_RAW_PACKET)));
1323 	if (config->tso)
1324 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1325 	/*
1326 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1327 	 * by default.
1328 	 */
1329 	if (config->mps == MLX5_ARG_UNSET)
1330 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1331 							  MLX5_MPW_DISABLED;
1332 	else
1333 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1334 	DRV_LOG(INFO, "%sMPS is %s",
1335 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1336 		config->mps == MLX5_MPW ? "legacy " : "",
1337 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1338 	if (config->devx) {
1339 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1340 		if (err) {
1341 			err = -err;
1342 			goto error;
1343 		}
1344 		/* Check relax ordering support. */
1345 		if (!haswell_broadwell_cpu) {
1346 			sh->cmng.relaxed_ordering_write =
1347 				config->hca_attr.relaxed_ordering_write;
1348 			sh->cmng.relaxed_ordering_read =
1349 				config->hca_attr.relaxed_ordering_read;
1350 		} else {
1351 			sh->cmng.relaxed_ordering_read = 0;
1352 			sh->cmng.relaxed_ordering_write = 0;
1353 		}
1354 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1355 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1356 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1357 		/* Check for LRO support. */
1358 		if (config->dest_tir && config->hca_attr.lro_cap &&
1359 		    config->dv_flow_en) {
1360 			/* TBD check tunnel lro caps. */
1361 			config->lro.supported = config->hca_attr.lro_cap;
1362 			DRV_LOG(DEBUG, "Device supports LRO");
1363 			/*
1364 			 * If LRO timeout is not configured by application,
1365 			 * use the minimal supported value.
1366 			 */
1367 			if (!config->lro.timeout)
1368 				config->lro.timeout =
1369 				config->hca_attr.lro_timer_supported_periods[0];
1370 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1371 				config->lro.timeout);
1372 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1373 				"required for coalescing is %d bytes",
1374 				config->hca_attr.lro_min_mss_size);
1375 		}
1376 #if defined(HAVE_MLX5DV_DR) && \
1377 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1378 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1379 		if (config->hca_attr.qos.sup &&
1380 		    config->hca_attr.qos.flow_meter_old &&
1381 		    config->dv_flow_en) {
1382 			uint8_t reg_c_mask =
1383 				config->hca_attr.qos.flow_meter_reg_c_ids;
1384 			/*
1385 			 * Meter needs two REG_C's for color match and pre-sfx
1386 			 * flow match. Here get the REG_C for color match.
1387 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1388 			 */
1389 			reg_c_mask &= 0xfc;
1390 			if (__builtin_popcount(reg_c_mask) < 1) {
1391 				priv->mtr_en = 0;
1392 				DRV_LOG(WARNING, "No available register for"
1393 					" meter.");
1394 			} else {
1395 				/*
1396 				 * The meter color register is used by the
1397 				 * flow-hit feature as well.
1398 				 * The flow-hit feature must use REG_C_3
1399 				 * Prefer REG_C_3 if it is available.
1400 				 */
1401 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1402 					priv->mtr_color_reg = REG_C_3;
1403 				else
1404 					priv->mtr_color_reg = ffs(reg_c_mask)
1405 							      - 1 + REG_C_0;
1406 				priv->mtr_en = 1;
1407 				priv->mtr_reg_share =
1408 				      config->hca_attr.qos.flow_meter;
1409 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1410 					priv->mtr_color_reg);
1411 			}
1412 		}
1413 		if (config->hca_attr.qos.sup &&
1414 			config->hca_attr.qos.flow_meter_aso_sup) {
1415 			uint32_t log_obj_size =
1416 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1417 			if (log_obj_size >=
1418 			config->hca_attr.qos.log_meter_aso_granularity &&
1419 			log_obj_size <=
1420 			config->hca_attr.qos.log_meter_aso_max_alloc)
1421 				sh->meter_aso_en = 1;
1422 		}
1423 		if (priv->mtr_en) {
1424 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1425 			if (err) {
1426 				err = -err;
1427 				goto error;
1428 			}
1429 		}
1430 		if (config->hca_attr.flow.tunnel_header_0_1)
1431 			sh->tunnel_header_0_1 = 1;
1432 #endif
1433 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1434 		if (config->hca_attr.flow_hit_aso &&
1435 		    priv->mtr_color_reg == REG_C_3) {
1436 			sh->flow_hit_aso_en = 1;
1437 			err = mlx5_flow_aso_age_mng_init(sh);
1438 			if (err) {
1439 				err = -err;
1440 				goto error;
1441 			}
1442 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1443 		}
1444 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1445 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1446 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1447 		if (config->hca_attr.ct_offload &&
1448 		    priv->mtr_color_reg == REG_C_3) {
1449 			err = mlx5_flow_aso_ct_mng_init(sh);
1450 			if (err) {
1451 				err = -err;
1452 				goto error;
1453 			}
1454 			DRV_LOG(DEBUG, "CT ASO is supported.");
1455 			sh->ct_aso_en = 1;
1456 		}
1457 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1458 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1459 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1460 		    config->dv_flow_en) {
1461 			priv->sampler_en = 1;
1462 			DRV_LOG(DEBUG, "Sampler enabled!");
1463 		} else {
1464 			priv->sampler_en = 0;
1465 			if (!config->hca_attr.log_max_ft_sampler_num)
1466 				DRV_LOG(WARNING,
1467 					"No available register for sampler.");
1468 			else
1469 				DRV_LOG(DEBUG, "DV flow is not supported!");
1470 		}
1471 #endif
1472 	}
1473 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1474 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1475 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1476 		config->cqe_comp = 0;
1477 	}
1478 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1479 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1480 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1481 				 " format isn't supported.");
1482 		config->cqe_comp = 0;
1483 	}
1484 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1485 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1486 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1487 				 " format isn't supported.");
1488 		config->cqe_comp = 0;
1489 	}
1490 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1491 			config->cqe_comp ? "" : "not ");
1492 	if (config->tx_pp) {
1493 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1494 			config->hca_attr.dev_freq_khz);
1495 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1496 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1497 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1498 			config->hca_attr.cross_channel ? "" : "not ");
1499 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1500 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1501 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1502 			config->hca_attr.non_wire_sq ? "" : "not ");
1503 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1504 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1505 			config->hca_attr.log_max_static_sq_wq);
1506 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1507 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1508 		if (!config->devx) {
1509 			DRV_LOG(ERR, "DevX is required for packet pacing");
1510 			err = ENODEV;
1511 			goto error;
1512 		}
1513 		if (!config->hca_attr.qos.packet_pacing) {
1514 			DRV_LOG(ERR, "Packet pacing is not supported");
1515 			err = ENODEV;
1516 			goto error;
1517 		}
1518 		if (!config->hca_attr.cross_channel) {
1519 			DRV_LOG(ERR, "Cross channel operations are"
1520 				     " required for packet pacing");
1521 			err = ENODEV;
1522 			goto error;
1523 		}
1524 		if (!config->hca_attr.wqe_index_ignore) {
1525 			DRV_LOG(ERR, "WQE index ignore feature is"
1526 				     " required for packet pacing");
1527 			err = ENODEV;
1528 			goto error;
1529 		}
1530 		if (!config->hca_attr.non_wire_sq) {
1531 			DRV_LOG(ERR, "Non-wire SQ feature is"
1532 				     " required for packet pacing");
1533 			err = ENODEV;
1534 			goto error;
1535 		}
1536 		if (!config->hca_attr.log_max_static_sq_wq) {
1537 			DRV_LOG(ERR, "Static WQE SQ feature is"
1538 				     " required for packet pacing");
1539 			err = ENODEV;
1540 			goto error;
1541 		}
1542 		if (!config->hca_attr.qos.wqe_rate_pp) {
1543 			DRV_LOG(ERR, "WQE rate mode is required"
1544 				     " for packet pacing");
1545 			err = ENODEV;
1546 			goto error;
1547 		}
1548 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1549 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1550 			     " can't create queues for packet pacing");
1551 		err = ENODEV;
1552 		goto error;
1553 #endif
1554 	}
1555 	if (config->devx) {
1556 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1557 
1558 		err = config->hca_attr.access_register_user ?
1559 			mlx5_devx_cmd_register_read
1560 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1561 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1562 		if (!err) {
1563 			uint32_t ts_mode;
1564 
1565 			/* MTUTC register is read successfully. */
1566 			ts_mode = MLX5_GET(register_mtutc, reg,
1567 					   time_stamp_mode);
1568 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1569 				config->rt_timestamp = 1;
1570 		} else {
1571 			/* Kernel does not support register reading. */
1572 			if (config->hca_attr.dev_freq_khz ==
1573 						 (NS_PER_S / MS_PER_S))
1574 				config->rt_timestamp = 1;
1575 		}
1576 	}
1577 	/*
1578 	 * If HW has bug working with tunnel packet decapsulation and
1579 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1580 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1581 	 */
1582 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1583 		config->hw_fcs_strip = 0;
1584 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1585 		(config->hw_fcs_strip ? "" : "not "));
1586 	if (config->mprq.enabled && mprq) {
1587 		if (config->mprq.stride_num_n &&
1588 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1589 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1590 			config->mprq.stride_num_n =
1591 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1592 						mprq_min_stride_num_n),
1593 					mprq_max_stride_num_n);
1594 			DRV_LOG(WARNING,
1595 				"the number of strides"
1596 				" for Multi-Packet RQ is out of range,"
1597 				" setting default value (%u)",
1598 				1 << config->mprq.stride_num_n);
1599 		}
1600 		if (config->mprq.stride_size_n &&
1601 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1602 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1603 			config->mprq.stride_size_n =
1604 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1605 						mprq_min_stride_size_n),
1606 					mprq_max_stride_size_n);
1607 			DRV_LOG(WARNING,
1608 				"the size of a stride"
1609 				" for Multi-Packet RQ is out of range,"
1610 				" setting default value (%u)",
1611 				1 << config->mprq.stride_size_n);
1612 		}
1613 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1614 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1615 	} else if (config->mprq.enabled && !mprq) {
1616 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1617 		config->mprq.enabled = 0;
1618 	}
1619 	if (config->max_dump_files_num == 0)
1620 		config->max_dump_files_num = 128;
1621 	eth_dev = rte_eth_dev_allocate(name);
1622 	if (eth_dev == NULL) {
1623 		DRV_LOG(ERR, "can not allocate rte ethdev");
1624 		err = ENOMEM;
1625 		goto error;
1626 	}
1627 	if (priv->representor) {
1628 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1629 		eth_dev->data->representor_id = priv->representor_id;
1630 	}
1631 	priv->mp_id.port_id = eth_dev->data->port_id;
1632 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1633 	/*
1634 	 * Store associated network device interface index. This index
1635 	 * is permanent throughout the lifetime of device. So, we may store
1636 	 * the ifindex here and use the cached value further.
1637 	 */
1638 	MLX5_ASSERT(spawn->ifindex);
1639 	priv->if_index = spawn->ifindex;
1640 	eth_dev->data->dev_private = priv;
1641 	priv->dev_data = eth_dev->data;
1642 	eth_dev->data->mac_addrs = priv->mac;
1643 	eth_dev->device = dpdk_dev;
1644 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1645 	/* Configure the first MAC address by default. */
1646 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1647 		DRV_LOG(ERR,
1648 			"port %u cannot get MAC address, is mlx5_en"
1649 			" loaded? (errno: %s)",
1650 			eth_dev->data->port_id, strerror(rte_errno));
1651 		err = ENODEV;
1652 		goto error;
1653 	}
1654 	DRV_LOG(INFO,
1655 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1656 		eth_dev->data->port_id,
1657 		mac.addr_bytes[0], mac.addr_bytes[1],
1658 		mac.addr_bytes[2], mac.addr_bytes[3],
1659 		mac.addr_bytes[4], mac.addr_bytes[5]);
1660 #ifdef RTE_LIBRTE_MLX5_DEBUG
1661 	{
1662 		char ifname[MLX5_NAMESIZE];
1663 
1664 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1665 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1666 				eth_dev->data->port_id, ifname);
1667 		else
1668 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1669 				eth_dev->data->port_id);
1670 	}
1671 #endif
1672 	/* Get actual MTU if possible. */
1673 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1674 	if (err) {
1675 		err = rte_errno;
1676 		goto error;
1677 	}
1678 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1679 		priv->mtu);
1680 	/* Initialize burst functions to prevent crashes before link-up. */
1681 	eth_dev->rx_pkt_burst = removed_rx_burst;
1682 	eth_dev->tx_pkt_burst = removed_tx_burst;
1683 	eth_dev->dev_ops = &mlx5_dev_ops;
1684 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1685 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1686 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1687 	/* Register MAC address. */
1688 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1689 	if (config->vf && config->vf_nl_en)
1690 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1691 				      mlx5_ifindex(eth_dev),
1692 				      eth_dev->data->mac_addrs,
1693 				      MLX5_MAX_MAC_ADDRESSES);
1694 	priv->ctrl_flows = 0;
1695 	rte_spinlock_init(&priv->flow_list_lock);
1696 	TAILQ_INIT(&priv->flow_meters);
1697 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1698 	if (!priv->mtr_profile_tbl)
1699 		goto error;
1700 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1701 	mlx5_glue->dv_set_context_attr(sh->ctx,
1702 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1703 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1704 				.alloc = &mlx5_alloc_verbs_buf,
1705 				.free = &mlx5_free_verbs_buf,
1706 				.data = sh,
1707 			}));
1708 	/* Bring Ethernet device up. */
1709 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1710 		eth_dev->data->port_id);
1711 	mlx5_set_link_up(eth_dev);
1712 	/*
1713 	 * Even though the interrupt handler is not installed yet,
1714 	 * interrupts will still trigger on the async_fd from
1715 	 * Verbs context returned by ibv_open_device().
1716 	 */
1717 	mlx5_link_update(eth_dev, 0);
1718 #ifdef HAVE_MLX5DV_DR_ESWITCH
1719 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1720 	      (switch_info->representor || switch_info->master)))
1721 		config->dv_esw_en = 0;
1722 #else
1723 	config->dv_esw_en = 0;
1724 #endif
1725 	/* Detect minimal data bytes to inline. */
1726 	mlx5_set_min_inline(spawn, config);
1727 	/* Store device configuration on private structure. */
1728 	priv->config = *config;
1729 	for (i = 0; i < MLX5_FLOW_TYPE_MAXI; i++) {
1730 		icfg[i].release_mem_en = !!config->reclaim_mode;
1731 		if (config->reclaim_mode)
1732 			icfg[i].per_core_cache = 0;
1733 		priv->flows[i] = mlx5_ipool_create(&icfg[i]);
1734 		if (!priv->flows[i])
1735 			goto error;
1736 	}
1737 	/* Create context for virtual machine VLAN workaround. */
1738 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1739 	if (config->dv_flow_en) {
1740 		err = mlx5_alloc_shared_dr(priv);
1741 		if (err)
1742 			goto error;
1743 	}
1744 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1745 		priv->obj_ops = devx_obj_ops;
1746 		priv->obj_ops.drop_action_create =
1747 						ibv_obj_ops.drop_action_create;
1748 		priv->obj_ops.drop_action_destroy =
1749 						ibv_obj_ops.drop_action_destroy;
1750 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1751 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1752 #else
1753 		if (config->dv_esw_en)
1754 			priv->obj_ops.txq_obj_modify =
1755 						ibv_obj_ops.txq_obj_modify;
1756 #endif
1757 		/* Use specific wrappers for Tx object. */
1758 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1759 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1760 		mlx5_queue_counter_id_prepare(eth_dev);
1761 		priv->obj_ops.lb_dummy_queue_create =
1762 					mlx5_rxq_ibv_obj_dummy_lb_create;
1763 		priv->obj_ops.lb_dummy_queue_release =
1764 					mlx5_rxq_ibv_obj_dummy_lb_release;
1765 	} else {
1766 		priv->obj_ops = ibv_obj_ops;
1767 	}
1768 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1769 	if (!priv->drop_queue.hrxq)
1770 		goto error;
1771 	/* Supported Verbs flow priority number detection. */
1772 	err = mlx5_flow_discover_priorities(eth_dev);
1773 	if (err < 0) {
1774 		err = -err;
1775 		goto error;
1776 	}
1777 	priv->config.flow_prio = err;
1778 	if (!priv->config.dv_esw_en &&
1779 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1780 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1781 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1782 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1783 	}
1784 	mlx5_set_metadata_mask(eth_dev);
1785 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1786 	    !priv->sh->dv_regc0_mask) {
1787 		DRV_LOG(ERR, "metadata mode %u is not supported "
1788 			     "(no metadata reg_c[0] is available)",
1789 			     priv->config.dv_xmeta_en);
1790 			err = ENOTSUP;
1791 			goto error;
1792 	}
1793 	mlx5_list_create(&priv->hrxqs, "hrxq", 0, eth_dev,
1794 			     mlx5_hrxq_create_cb,
1795 			     mlx5_hrxq_match_cb,
1796 			     mlx5_hrxq_remove_cb);
1797 	/* Query availability of metadata reg_c's. */
1798 	err = mlx5_flow_discover_mreg_c(eth_dev);
1799 	if (err < 0) {
1800 		err = -err;
1801 		goto error;
1802 	}
1803 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1804 		DRV_LOG(DEBUG,
1805 			"port %u extensive metadata register is not supported",
1806 			eth_dev->data->port_id);
1807 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1808 			DRV_LOG(ERR, "metadata mode %u is not supported "
1809 				     "(no metadata registers available)",
1810 				     priv->config.dv_xmeta_en);
1811 			err = ENOTSUP;
1812 			goto error;
1813 		}
1814 	}
1815 	if (priv->config.dv_flow_en &&
1816 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1817 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1818 	    priv->sh->dv_regc0_mask) {
1819 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1820 						      MLX5_FLOW_MREG_HTABLE_SZ,
1821 						      0, 0,
1822 						      flow_dv_mreg_create_cb,
1823 						      flow_dv_mreg_match_cb,
1824 						      flow_dv_mreg_remove_cb);
1825 		if (!priv->mreg_cp_tbl) {
1826 			err = ENOMEM;
1827 			goto error;
1828 		}
1829 		priv->mreg_cp_tbl->ctx = eth_dev;
1830 	}
1831 	rte_spinlock_init(&priv->shared_act_sl);
1832 	mlx5_flow_counter_mode_config(eth_dev);
1833 	if (priv->config.dv_flow_en)
1834 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1835 	return eth_dev;
1836 error:
1837 	if (priv) {
1838 		if (priv->mreg_cp_tbl)
1839 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1840 		if (priv->sh)
1841 			mlx5_os_free_shared_dr(priv);
1842 		if (priv->nl_socket_route >= 0)
1843 			close(priv->nl_socket_route);
1844 		if (priv->nl_socket_rdma >= 0)
1845 			close(priv->nl_socket_rdma);
1846 		if (priv->vmwa_context)
1847 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1848 		if (eth_dev && priv->drop_queue.hrxq)
1849 			mlx5_drop_action_destroy(eth_dev);
1850 		if (priv->mtr_profile_tbl)
1851 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1852 		if (own_domain_id)
1853 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1854 		mlx5_list_destroy(&priv->hrxqs);
1855 		mlx5_free(priv);
1856 		if (eth_dev != NULL)
1857 			eth_dev->data->dev_private = NULL;
1858 	}
1859 	if (eth_dev != NULL) {
1860 		/* mac_addrs must not be freed alone because part of
1861 		 * dev_private
1862 		 **/
1863 		eth_dev->data->mac_addrs = NULL;
1864 		rte_eth_dev_release_port(eth_dev);
1865 	}
1866 	if (sh)
1867 		mlx5_free_shared_dev_ctx(sh);
1868 	MLX5_ASSERT(err > 0);
1869 	rte_errno = err;
1870 	return NULL;
1871 }
1872 
1873 /**
1874  * Comparison callback to sort device data.
1875  *
1876  * This is meant to be used with qsort().
1877  *
1878  * @param a[in]
1879  *   Pointer to pointer to first data object.
1880  * @param b[in]
1881  *   Pointer to pointer to second data object.
1882  *
1883  * @return
1884  *   0 if both objects are equal, less than 0 if the first argument is less
1885  *   than the second, greater than 0 otherwise.
1886  */
1887 static int
1888 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1889 {
1890 	const struct mlx5_switch_info *si_a =
1891 		&((const struct mlx5_dev_spawn_data *)a)->info;
1892 	const struct mlx5_switch_info *si_b =
1893 		&((const struct mlx5_dev_spawn_data *)b)->info;
1894 	int ret;
1895 
1896 	/* Master device first. */
1897 	ret = si_b->master - si_a->master;
1898 	if (ret)
1899 		return ret;
1900 	/* Then representor devices. */
1901 	ret = si_b->representor - si_a->representor;
1902 	if (ret)
1903 		return ret;
1904 	/* Unidentified devices come last in no specific order. */
1905 	if (!si_a->representor)
1906 		return 0;
1907 	/* Order representors by name. */
1908 	return si_a->port_name - si_b->port_name;
1909 }
1910 
1911 /**
1912  * Match PCI information for possible slaves of bonding device.
1913  *
1914  * @param[in] ibv_dev
1915  *   Pointer to Infiniband device structure.
1916  * @param[in] pci_dev
1917  *   Pointer to primary PCI address structure to match.
1918  * @param[in] nl_rdma
1919  *   Netlink RDMA group socket handle.
1920  * @param[in] owner
1921  *   Rerepsentor owner PF index.
1922  * @param[out] bond_info
1923  *   Pointer to bonding information.
1924  *
1925  * @return
1926  *   negative value if no bonding device found, otherwise
1927  *   positive index of slave PF in bonding.
1928  */
1929 static int
1930 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1931 			   const struct rte_pci_addr *pci_dev,
1932 			   int nl_rdma, uint16_t owner,
1933 			   struct mlx5_bond_info *bond_info)
1934 {
1935 	char ifname[IF_NAMESIZE + 1];
1936 	unsigned int ifindex;
1937 	unsigned int np, i;
1938 	FILE *bond_file = NULL, *file;
1939 	int pf = -1;
1940 	int ret;
1941 
1942 	/*
1943 	 * Try to get master device name. If something goes
1944 	 * wrong suppose the lack of kernel support and no
1945 	 * bonding devices.
1946 	 */
1947 	memset(bond_info, 0, sizeof(*bond_info));
1948 	if (nl_rdma < 0)
1949 		return -1;
1950 	if (!strstr(ibv_dev->name, "bond"))
1951 		return -1;
1952 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1953 	if (!np)
1954 		return -1;
1955 	/*
1956 	 * The Master device might not be on the predefined
1957 	 * port (not on port index 1, it is not garanted),
1958 	 * we have to scan all Infiniband device port and
1959 	 * find master.
1960 	 */
1961 	for (i = 1; i <= np; ++i) {
1962 		/* Check whether Infiniband port is populated. */
1963 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1964 		if (!ifindex)
1965 			continue;
1966 		if (!if_indextoname(ifindex, ifname))
1967 			continue;
1968 		/* Try to read bonding slave names from sysfs. */
1969 		MKSTR(slaves,
1970 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1971 		bond_file = fopen(slaves, "r");
1972 		if (bond_file)
1973 			break;
1974 	}
1975 	if (!bond_file)
1976 		return -1;
1977 	/* Use safe format to check maximal buffer length. */
1978 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1979 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1980 		char tmp_str[IF_NAMESIZE + 32];
1981 		struct rte_pci_addr pci_addr;
1982 		struct mlx5_switch_info	info;
1983 
1984 		/* Process slave interface names in the loop. */
1985 		snprintf(tmp_str, sizeof(tmp_str),
1986 			 "/sys/class/net/%s", ifname);
1987 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1988 			DRV_LOG(WARNING, "can not get PCI address"
1989 					 " for netdev \"%s\"", ifname);
1990 			continue;
1991 		}
1992 		/* Slave interface PCI address match found. */
1993 		snprintf(tmp_str, sizeof(tmp_str),
1994 			 "/sys/class/net/%s/phys_port_name", ifname);
1995 		file = fopen(tmp_str, "rb");
1996 		if (!file)
1997 			break;
1998 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1999 		if (fscanf(file, "%32s", tmp_str) == 1)
2000 			mlx5_translate_port_name(tmp_str, &info);
2001 		fclose(file);
2002 		/* Only process PF ports. */
2003 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
2004 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2005 			continue;
2006 		/* Check max bonding member. */
2007 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
2008 			DRV_LOG(WARNING, "bonding index out of range, "
2009 				"please increase MLX5_BOND_MAX_PORTS: %s",
2010 				tmp_str);
2011 			break;
2012 		}
2013 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
2014 		if (pci_dev->domain == pci_addr.domain &&
2015 		    pci_dev->bus == pci_addr.bus &&
2016 		    pci_dev->devid == pci_addr.devid &&
2017 		    ((pci_dev->function == 0 &&
2018 		      pci_dev->function + owner == pci_addr.function) ||
2019 		     (pci_dev->function == owner &&
2020 		      pci_addr.function == owner)))
2021 			pf = info.port_name;
2022 		/* Get ifindex. */
2023 		snprintf(tmp_str, sizeof(tmp_str),
2024 			 "/sys/class/net/%s/ifindex", ifname);
2025 		file = fopen(tmp_str, "rb");
2026 		if (!file)
2027 			break;
2028 		ret = fscanf(file, "%u", &ifindex);
2029 		fclose(file);
2030 		if (ret != 1)
2031 			break;
2032 		/* Save bonding info. */
2033 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
2034 			sizeof(bond_info->ports[0].ifname));
2035 		bond_info->ports[info.port_name].pci_addr = pci_addr;
2036 		bond_info->ports[info.port_name].ifindex = ifindex;
2037 		bond_info->n_port++;
2038 	}
2039 	if (pf >= 0) {
2040 		/* Get bond interface info */
2041 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
2042 					   bond_info->ifname);
2043 		if (ret)
2044 			DRV_LOG(ERR, "unable to get bond info: %s",
2045 				strerror(rte_errno));
2046 		else
2047 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
2048 				ifindex, bond_info->ifindex, bond_info->ifname);
2049 	}
2050 	return pf;
2051 }
2052 
2053 /**
2054  * Register a PCI device within bonding.
2055  *
2056  * This function spawns Ethernet devices out of a given PCI device and
2057  * bonding owner PF index.
2058  *
2059  * @param[in] pci_dev
2060  *   PCI device information.
2061  * @param[in] req_eth_da
2062  *   Requested ethdev device argument.
2063  * @param[in] owner_id
2064  *   Requested owner PF port ID within bonding device, default to 0.
2065  *
2066  * @return
2067  *   0 on success, a negative errno value otherwise and rte_errno is set.
2068  */
2069 static int
2070 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
2071 		     struct rte_eth_devargs *req_eth_da,
2072 		     uint16_t owner_id)
2073 {
2074 	struct ibv_device **ibv_list;
2075 	/*
2076 	 * Number of found IB Devices matching with requested PCI BDF.
2077 	 * nd != 1 means there are multiple IB devices over the same
2078 	 * PCI device and we have representors and master.
2079 	 */
2080 	unsigned int nd = 0;
2081 	/*
2082 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
2083 	 * we have the single multiport IB device, and there may be
2084 	 * representors attached to some of found ports.
2085 	 */
2086 	unsigned int np = 0;
2087 	/*
2088 	 * Number of DPDK ethernet devices to Spawn - either over
2089 	 * multiple IB devices or multiple ports of single IB device.
2090 	 * Actually this is the number of iterations to spawn.
2091 	 */
2092 	unsigned int ns = 0;
2093 	/*
2094 	 * Bonding device
2095 	 *   < 0 - no bonding device (single one)
2096 	 *  >= 0 - bonding device (value is slave PF index)
2097 	 */
2098 	int bd = -1;
2099 	struct mlx5_dev_spawn_data *list = NULL;
2100 	struct mlx5_dev_config dev_config;
2101 	unsigned int dev_config_vf;
2102 	struct rte_eth_devargs eth_da = *req_eth_da;
2103 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
2104 	struct mlx5_bond_info bond_info;
2105 	int ret = -1;
2106 
2107 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
2108 		mlx5_pmd_socket_init();
2109 	ret = mlx5_init_once();
2110 	if (ret) {
2111 		DRV_LOG(ERR, "unable to init PMD global data: %s",
2112 			strerror(rte_errno));
2113 		return -rte_errno;
2114 	}
2115 	errno = 0;
2116 	ibv_list = mlx5_glue->get_device_list(&ret);
2117 	if (!ibv_list) {
2118 		rte_errno = errno ? errno : ENOSYS;
2119 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2120 		return -rte_errno;
2121 	}
2122 	/*
2123 	 * First scan the list of all Infiniband devices to find
2124 	 * matching ones, gathering into the list.
2125 	 */
2126 	struct ibv_device *ibv_match[ret + 1];
2127 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2128 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2129 	unsigned int i;
2130 
2131 	while (ret-- > 0) {
2132 		struct rte_pci_addr pci_addr;
2133 
2134 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2135 		bd = mlx5_device_bond_pci_match
2136 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2137 				 &bond_info);
2138 		if (bd >= 0) {
2139 			/*
2140 			 * Bonding device detected. Only one match is allowed,
2141 			 * the bonding is supported over multi-port IB device,
2142 			 * there should be no matches on representor PCI
2143 			 * functions or non VF LAG bonding devices with
2144 			 * specified address.
2145 			 */
2146 			if (nd) {
2147 				DRV_LOG(ERR,
2148 					"multiple PCI match on bonding device"
2149 					"\"%s\" found", ibv_list[ret]->name);
2150 				rte_errno = ENOENT;
2151 				ret = -rte_errno;
2152 				goto exit;
2153 			}
2154 			/* Amend owner pci address if owner PF ID specified. */
2155 			if (eth_da.nb_representor_ports)
2156 				owner_pci.function += owner_id;
2157 			DRV_LOG(INFO, "PCI information matches for"
2158 				      " slave %d bonding device \"%s\"",
2159 				      bd, ibv_list[ret]->name);
2160 			ibv_match[nd++] = ibv_list[ret];
2161 			break;
2162 		} else {
2163 			/* Bonding device not found. */
2164 			if (mlx5_dev_to_pci_addr
2165 				(ibv_list[ret]->ibdev_path, &pci_addr))
2166 				continue;
2167 			if (owner_pci.domain != pci_addr.domain ||
2168 			    owner_pci.bus != pci_addr.bus ||
2169 			    owner_pci.devid != pci_addr.devid ||
2170 			    owner_pci.function != pci_addr.function)
2171 				continue;
2172 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2173 				ibv_list[ret]->name);
2174 			ibv_match[nd++] = ibv_list[ret];
2175 		}
2176 	}
2177 	ibv_match[nd] = NULL;
2178 	if (!nd) {
2179 		/* No device matches, just complain and bail out. */
2180 		DRV_LOG(WARNING,
2181 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2182 			" are kernel drivers loaded?",
2183 			owner_pci.domain, owner_pci.bus,
2184 			owner_pci.devid, owner_pci.function);
2185 		rte_errno = ENOENT;
2186 		ret = -rte_errno;
2187 		goto exit;
2188 	}
2189 	if (nd == 1) {
2190 		/*
2191 		 * Found single matching device may have multiple ports.
2192 		 * Each port may be representor, we have to check the port
2193 		 * number and check the representors existence.
2194 		 */
2195 		if (nl_rdma >= 0)
2196 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2197 		if (!np)
2198 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2199 					 " ports number", ibv_match[0]->name);
2200 		if (bd >= 0 && !np) {
2201 			DRV_LOG(ERR, "can not get ports"
2202 				     " for bonding device");
2203 			rte_errno = ENOENT;
2204 			ret = -rte_errno;
2205 			goto exit;
2206 		}
2207 	}
2208 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2209 	if (bd >= 0) {
2210 		/*
2211 		 * This may happen if there is VF LAG kernel support and
2212 		 * application is compiled with older rdma_core library.
2213 		 */
2214 		DRV_LOG(ERR,
2215 			"No kernel/verbs support for VF LAG bonding found.");
2216 		rte_errno = ENOTSUP;
2217 		ret = -rte_errno;
2218 		goto exit;
2219 	}
2220 #endif
2221 	/*
2222 	 * Now we can determine the maximal
2223 	 * amount of devices to be spawned.
2224 	 */
2225 	list = mlx5_malloc(MLX5_MEM_ZERO,
2226 			   sizeof(struct mlx5_dev_spawn_data) *
2227 			   (np ? np : nd),
2228 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2229 	if (!list) {
2230 		DRV_LOG(ERR, "spawn data array allocation failure");
2231 		rte_errno = ENOMEM;
2232 		ret = -rte_errno;
2233 		goto exit;
2234 	}
2235 	if (bd >= 0 || np > 1) {
2236 		/*
2237 		 * Single IB device with multiple ports found,
2238 		 * it may be E-Switch master device and representors.
2239 		 * We have to perform identification through the ports.
2240 		 */
2241 		MLX5_ASSERT(nl_rdma >= 0);
2242 		MLX5_ASSERT(ns == 0);
2243 		MLX5_ASSERT(nd == 1);
2244 		MLX5_ASSERT(np);
2245 		for (i = 1; i <= np; ++i) {
2246 			list[ns].bond_info = &bond_info;
2247 			list[ns].max_port = np;
2248 			list[ns].phys_port = i;
2249 			list[ns].phys_dev = ibv_match[0];
2250 			list[ns].eth_dev = NULL;
2251 			list[ns].pci_dev = pci_dev;
2252 			list[ns].pf_bond = bd;
2253 			list[ns].ifindex = mlx5_nl_ifindex
2254 				(nl_rdma,
2255 				mlx5_os_get_dev_device_name
2256 						(list[ns].phys_dev), i);
2257 			if (!list[ns].ifindex) {
2258 				/*
2259 				 * No network interface index found for the
2260 				 * specified port, it means there is no
2261 				 * representor on this port. It's OK,
2262 				 * there can be disabled ports, for example
2263 				 * if sriov_numvfs < sriov_totalvfs.
2264 				 */
2265 				continue;
2266 			}
2267 			ret = -1;
2268 			if (nl_route >= 0)
2269 				ret = mlx5_nl_switch_info
2270 					       (nl_route,
2271 						list[ns].ifindex,
2272 						&list[ns].info);
2273 			if (ret || (!list[ns].info.representor &&
2274 				    !list[ns].info.master)) {
2275 				/*
2276 				 * We failed to recognize representors with
2277 				 * Netlink, let's try to perform the task
2278 				 * with sysfs.
2279 				 */
2280 				ret =  mlx5_sysfs_switch_info
2281 						(list[ns].ifindex,
2282 						 &list[ns].info);
2283 			}
2284 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2285 			if (!ret && bd >= 0) {
2286 				switch (list[ns].info.name_type) {
2287 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2288 					if (list[ns].info.port_name == bd)
2289 						ns++;
2290 					break;
2291 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2292 					/* Fallthrough */
2293 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2294 					/* Fallthrough */
2295 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2296 					if (list[ns].info.pf_num == bd)
2297 						ns++;
2298 					break;
2299 				default:
2300 					break;
2301 				}
2302 				continue;
2303 			}
2304 #endif
2305 			if (!ret && (list[ns].info.representor ^
2306 				     list[ns].info.master))
2307 				ns++;
2308 		}
2309 		if (!ns) {
2310 			DRV_LOG(ERR,
2311 				"unable to recognize master/representors"
2312 				" on the IB device with multiple ports");
2313 			rte_errno = ENOENT;
2314 			ret = -rte_errno;
2315 			goto exit;
2316 		}
2317 	} else {
2318 		/*
2319 		 * The existence of several matching entries (nd > 1) means
2320 		 * port representors have been instantiated. No existing Verbs
2321 		 * call nor sysfs entries can tell them apart, this can only
2322 		 * be done through Netlink calls assuming kernel drivers are
2323 		 * recent enough to support them.
2324 		 *
2325 		 * In the event of identification failure through Netlink,
2326 		 * try again through sysfs, then:
2327 		 *
2328 		 * 1. A single IB device matches (nd == 1) with single
2329 		 *    port (np=0/1) and is not a representor, assume
2330 		 *    no switch support.
2331 		 *
2332 		 * 2. Otherwise no safe assumptions can be made;
2333 		 *    complain louder and bail out.
2334 		 */
2335 		for (i = 0; i != nd; ++i) {
2336 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2337 			list[ns].bond_info = NULL;
2338 			list[ns].max_port = 1;
2339 			list[ns].phys_port = 1;
2340 			list[ns].phys_dev = ibv_match[i];
2341 			list[ns].eth_dev = NULL;
2342 			list[ns].pci_dev = pci_dev;
2343 			list[ns].pf_bond = -1;
2344 			list[ns].ifindex = 0;
2345 			if (nl_rdma >= 0)
2346 				list[ns].ifindex = mlx5_nl_ifindex
2347 				(nl_rdma,
2348 				mlx5_os_get_dev_device_name
2349 						(list[ns].phys_dev), 1);
2350 			if (!list[ns].ifindex) {
2351 				char ifname[IF_NAMESIZE];
2352 
2353 				/*
2354 				 * Netlink failed, it may happen with old
2355 				 * ib_core kernel driver (before 4.16).
2356 				 * We can assume there is old driver because
2357 				 * here we are processing single ports IB
2358 				 * devices. Let's try sysfs to retrieve
2359 				 * the ifindex. The method works for
2360 				 * master device only.
2361 				 */
2362 				if (nd > 1) {
2363 					/*
2364 					 * Multiple devices found, assume
2365 					 * representors, can not distinguish
2366 					 * master/representor and retrieve
2367 					 * ifindex via sysfs.
2368 					 */
2369 					continue;
2370 				}
2371 				ret = mlx5_get_ifname_sysfs
2372 					(ibv_match[i]->ibdev_path, ifname);
2373 				if (!ret)
2374 					list[ns].ifindex =
2375 						if_nametoindex(ifname);
2376 				if (!list[ns].ifindex) {
2377 					/*
2378 					 * No network interface index found
2379 					 * for the specified device, it means
2380 					 * there it is neither representor
2381 					 * nor master.
2382 					 */
2383 					continue;
2384 				}
2385 			}
2386 			ret = -1;
2387 			if (nl_route >= 0)
2388 				ret = mlx5_nl_switch_info
2389 					       (nl_route,
2390 						list[ns].ifindex,
2391 						&list[ns].info);
2392 			if (ret || (!list[ns].info.representor &&
2393 				    !list[ns].info.master)) {
2394 				/*
2395 				 * We failed to recognize representors with
2396 				 * Netlink, let's try to perform the task
2397 				 * with sysfs.
2398 				 */
2399 				ret =  mlx5_sysfs_switch_info
2400 						(list[ns].ifindex,
2401 						 &list[ns].info);
2402 			}
2403 			if (!ret && (list[ns].info.representor ^
2404 				     list[ns].info.master)) {
2405 				ns++;
2406 			} else if ((nd == 1) &&
2407 				   !list[ns].info.representor &&
2408 				   !list[ns].info.master) {
2409 				/*
2410 				 * Single IB device with
2411 				 * one physical port and
2412 				 * attached network device.
2413 				 * May be SRIOV is not enabled
2414 				 * or there is no 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"
2424 				" on the multiple IB devices");
2425 			rte_errno = ENOENT;
2426 			ret = -rte_errno;
2427 			goto exit;
2428 		}
2429 		/*
2430 		 * New kernels may add the switch_id attribute for the case
2431 		 * there is no E-Switch and we wrongly recognized the
2432 		 * only device as master. Override this if there is the
2433 		 * single device with single port and new device name
2434 		 * format present.
2435 		 */
2436 		if (nd == 1 &&
2437 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2438 			list[0].info.master = 0;
2439 			list[0].info.representor = 0;
2440 		}
2441 	}
2442 	MLX5_ASSERT(ns);
2443 	/*
2444 	 * Sort list to probe devices in natural order for users convenience
2445 	 * (i.e. master first, then representors from lowest to highest ID).
2446 	 */
2447 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2448 	/* Device specific configuration. */
2449 	switch (pci_dev->id.device_id) {
2450 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2451 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2452 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2453 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2454 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2455 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2456 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2457 		dev_config_vf = 1;
2458 		break;
2459 	default:
2460 		dev_config_vf = 0;
2461 		break;
2462 	}
2463 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2464 		/* Set devargs default values. */
2465 		if (eth_da.nb_mh_controllers == 0) {
2466 			eth_da.nb_mh_controllers = 1;
2467 			eth_da.mh_controllers[0] = 0;
2468 		}
2469 		if (eth_da.nb_ports == 0 && ns > 0) {
2470 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2471 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2472 					pci_dev->device.devargs->args);
2473 			eth_da.nb_ports = 1;
2474 			eth_da.ports[0] = list[0].info.pf_num;
2475 		}
2476 		if (eth_da.nb_representor_ports == 0) {
2477 			eth_da.nb_representor_ports = 1;
2478 			eth_da.representor_ports[0] = 0;
2479 		}
2480 	}
2481 	for (i = 0; i != ns; ++i) {
2482 		uint32_t restore;
2483 
2484 		/* Default configuration. */
2485 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2486 		dev_config.vf = dev_config_vf;
2487 		dev_config.mps = MLX5_ARG_UNSET;
2488 		dev_config.dbnc = MLX5_ARG_UNSET;
2489 		dev_config.rx_vec_en = 1;
2490 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2491 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2492 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2493 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2494 		dev_config.vf_nl_en = 1;
2495 		dev_config.mr_ext_memseg_en = 1;
2496 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2497 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2498 		dev_config.dv_esw_en = 1;
2499 		dev_config.dv_flow_en = 1;
2500 		dev_config.decap_en = 1;
2501 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2502 		dev_config.allow_duplicate_pattern = 1;
2503 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2504 						 &list[i],
2505 						 &dev_config,
2506 						 &eth_da);
2507 		if (!list[i].eth_dev) {
2508 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2509 				break;
2510 			/* Device is disabled or already spawned. Ignore it. */
2511 			continue;
2512 		}
2513 		restore = list[i].eth_dev->data->dev_flags;
2514 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2515 		/* Restore non-PCI flags cleared by the above call. */
2516 		list[i].eth_dev->data->dev_flags |= restore;
2517 		rte_eth_dev_probing_finish(list[i].eth_dev);
2518 	}
2519 	if (i != ns) {
2520 		DRV_LOG(ERR,
2521 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2522 			" encountering an error: %s",
2523 			owner_pci.domain, owner_pci.bus,
2524 			owner_pci.devid, owner_pci.function,
2525 			strerror(rte_errno));
2526 		ret = -rte_errno;
2527 		/* Roll back. */
2528 		while (i--) {
2529 			if (!list[i].eth_dev)
2530 				continue;
2531 			mlx5_dev_close(list[i].eth_dev);
2532 			/* mac_addrs must not be freed because in dev_private */
2533 			list[i].eth_dev->data->mac_addrs = NULL;
2534 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2535 		}
2536 		/* Restore original error. */
2537 		rte_errno = -ret;
2538 	} else {
2539 		ret = 0;
2540 	}
2541 exit:
2542 	/*
2543 	 * Do the routine cleanup:
2544 	 * - close opened Netlink sockets
2545 	 * - free allocated spawn data array
2546 	 * - free the Infiniband device list
2547 	 */
2548 	if (nl_rdma >= 0)
2549 		close(nl_rdma);
2550 	if (nl_route >= 0)
2551 		close(nl_route);
2552 	if (list)
2553 		mlx5_free(list);
2554 	MLX5_ASSERT(ibv_list);
2555 	mlx5_glue->free_device_list(ibv_list);
2556 	return ret;
2557 }
2558 
2559 /**
2560  * DPDK callback to register a PCI device.
2561  *
2562  * This function spawns Ethernet devices out of a given PCI device.
2563  *
2564  * @param[in] pci_drv
2565  *   PCI driver structure (mlx5_driver).
2566  * @param[in] pci_dev
2567  *   PCI device information.
2568  *
2569  * @return
2570  *   0 on success, a negative errno value otherwise and rte_errno is set.
2571  */
2572 int
2573 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2574 		  struct rte_pci_device *pci_dev)
2575 {
2576 	struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2577 	int ret = 0;
2578 	uint16_t p;
2579 
2580 	if (pci_dev->device.devargs) {
2581 		/* Parse representor information from device argument. */
2582 		if (pci_dev->device.devargs->cls_str)
2583 			ret = rte_eth_devargs_parse
2584 				(pci_dev->device.devargs->cls_str, &eth_da);
2585 		if (ret) {
2586 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2587 				pci_dev->device.devargs->cls_str);
2588 			return -rte_errno;
2589 		}
2590 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2591 			/* Support legacy device argument */
2592 			ret = rte_eth_devargs_parse
2593 				(pci_dev->device.devargs->args, &eth_da);
2594 			if (ret) {
2595 				DRV_LOG(ERR, "failed to parse device arguments: %s",
2596 					pci_dev->device.devargs->args);
2597 				return -rte_errno;
2598 			}
2599 		}
2600 	}
2601 
2602 	if (eth_da.nb_ports > 0) {
2603 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2604 		for (p = 0; p < eth_da.nb_ports; p++)
2605 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2606 						   eth_da.ports[p]);
2607 	} else {
2608 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2609 	}
2610 	return ret;
2611 }
2612 
2613 static int
2614 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2615 {
2616 	char *env;
2617 	int value;
2618 
2619 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2620 	/* Get environment variable to store. */
2621 	env = getenv(MLX5_SHUT_UP_BF);
2622 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2623 	if (config->dbnc == MLX5_ARG_UNSET)
2624 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2625 	else
2626 		setenv(MLX5_SHUT_UP_BF,
2627 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2628 	return value;
2629 }
2630 
2631 static void
2632 mlx5_restore_doorbell_mapping_env(int value)
2633 {
2634 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2635 	/* Restore the original environment variable state. */
2636 	if (value == MLX5_ARG_UNSET)
2637 		unsetenv(MLX5_SHUT_UP_BF);
2638 	else
2639 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2640 }
2641 
2642 /**
2643  * Extract pdn of PD object using DV API.
2644  *
2645  * @param[in] pd
2646  *   Pointer to the verbs PD object.
2647  * @param[out] pdn
2648  *   Pointer to the PD object number variable.
2649  *
2650  * @return
2651  *   0 on success, error value otherwise.
2652  */
2653 int
2654 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2655 {
2656 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2657 	struct mlx5dv_obj obj;
2658 	struct mlx5dv_pd pd_info;
2659 	int ret = 0;
2660 
2661 	obj.pd.in = pd;
2662 	obj.pd.out = &pd_info;
2663 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2664 	if (ret) {
2665 		DRV_LOG(DEBUG, "Fail to get PD object info");
2666 		return ret;
2667 	}
2668 	*pdn = pd_info.pdn;
2669 	return 0;
2670 #else
2671 	(void)pd;
2672 	(void)pdn;
2673 	return -ENOTSUP;
2674 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2675 }
2676 
2677 /**
2678  * Function API to open IB device.
2679  *
2680  * This function calls the Linux glue APIs to open a device.
2681  *
2682  * @param[in] spawn
2683  *   Pointer to the IB device attributes (name, port, etc).
2684  * @param[out] config
2685  *   Pointer to device configuration structure.
2686  * @param[out] sh
2687  *   Pointer to shared context structure.
2688  *
2689  * @return
2690  *   0 on success, a positive error value otherwise.
2691  */
2692 int
2693 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2694 		     const struct mlx5_dev_config *config,
2695 		     struct mlx5_dev_ctx_shared *sh)
2696 {
2697 	int dbmap_env;
2698 	int err = 0;
2699 
2700 	sh->numa_node = spawn->pci_dev->device.numa_node;
2701 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2702 	/*
2703 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2704 	 * before the device creation. The rdma_core library
2705 	 * checks the variable at device creation and
2706 	 * stores the result internally.
2707 	 */
2708 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2709 	/* Try to open IB device with DV first, then usual Verbs. */
2710 	errno = 0;
2711 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2712 	if (sh->ctx) {
2713 		sh->devx = 1;
2714 		DRV_LOG(DEBUG, "DevX is supported");
2715 		/* The device is created, no need for environment. */
2716 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2717 	} else {
2718 		/* The environment variable is still configured. */
2719 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2720 		err = errno ? errno : ENODEV;
2721 		/*
2722 		 * The environment variable is not needed anymore,
2723 		 * all device creation attempts are completed.
2724 		 */
2725 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2726 		if (!sh->ctx)
2727 			return err;
2728 		DRV_LOG(DEBUG, "DevX is NOT supported");
2729 		err = 0;
2730 	}
2731 	if (!err && sh->ctx) {
2732 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2733 		mlx5_glue->dv_set_context_attr(sh->ctx,
2734 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2735 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2736 				.alloc = &mlx5_alloc_verbs_buf,
2737 				.free = &mlx5_free_verbs_buf,
2738 				.data = sh,
2739 			}));
2740 	}
2741 	return err;
2742 }
2743 
2744 /**
2745  * Install shared asynchronous device events handler.
2746  * This function is implemented to support event sharing
2747  * between multiple ports of single IB device.
2748  *
2749  * @param sh
2750  *   Pointer to mlx5_dev_ctx_shared object.
2751  */
2752 void
2753 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2754 {
2755 	int ret;
2756 	int flags;
2757 
2758 	sh->intr_handle.fd = -1;
2759 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2760 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2761 		    F_SETFL, flags | O_NONBLOCK);
2762 	if (ret) {
2763 		DRV_LOG(INFO, "failed to change file descriptor async event"
2764 			" queue");
2765 	} else {
2766 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2767 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2768 		if (rte_intr_callback_register(&sh->intr_handle,
2769 					mlx5_dev_interrupt_handler, sh)) {
2770 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2771 			sh->intr_handle.fd = -1;
2772 		}
2773 	}
2774 	if (sh->devx) {
2775 #ifdef HAVE_IBV_DEVX_ASYNC
2776 		sh->intr_handle_devx.fd = -1;
2777 		sh->devx_comp =
2778 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2779 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2780 		if (!devx_comp) {
2781 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2782 			return;
2783 		}
2784 		flags = fcntl(devx_comp->fd, F_GETFL);
2785 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2786 		if (ret) {
2787 			DRV_LOG(INFO, "failed to change file descriptor"
2788 				" devx comp");
2789 			return;
2790 		}
2791 		sh->intr_handle_devx.fd = devx_comp->fd;
2792 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2793 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2794 					mlx5_dev_interrupt_handler_devx, sh)) {
2795 			DRV_LOG(INFO, "Fail to install the devx shared"
2796 				" interrupt.");
2797 			sh->intr_handle_devx.fd = -1;
2798 		}
2799 #endif /* HAVE_IBV_DEVX_ASYNC */
2800 	}
2801 }
2802 
2803 /**
2804  * Uninstall shared asynchronous device events handler.
2805  * This function is implemented to support event sharing
2806  * between multiple ports of single IB device.
2807  *
2808  * @param dev
2809  *   Pointer to mlx5_dev_ctx_shared object.
2810  */
2811 void
2812 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2813 {
2814 	if (sh->intr_handle.fd >= 0)
2815 		mlx5_intr_callback_unregister(&sh->intr_handle,
2816 					      mlx5_dev_interrupt_handler, sh);
2817 #ifdef HAVE_IBV_DEVX_ASYNC
2818 	if (sh->intr_handle_devx.fd >= 0)
2819 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2820 				  mlx5_dev_interrupt_handler_devx, sh);
2821 	if (sh->devx_comp)
2822 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2823 #endif
2824 }
2825 
2826 /**
2827  * Read statistics by a named counter.
2828  *
2829  * @param[in] priv
2830  *   Pointer to the private device data structure.
2831  * @param[in] ctr_name
2832  *   Pointer to the name of the statistic counter to read
2833  * @param[out] stat
2834  *   Pointer to read statistic value.
2835  * @return
2836  *   0 on success and stat is valud, 1 if failed to read the value
2837  *   rte_errno is set.
2838  *
2839  */
2840 int
2841 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2842 		      uint64_t *stat)
2843 {
2844 	int fd;
2845 
2846 	if (priv->sh) {
2847 		if (priv->q_counters != NULL &&
2848 		    strcmp(ctr_name, "out_of_buffer") == 0)
2849 			return mlx5_devx_cmd_queue_counter_query
2850 					(priv->q_counters, 0, (uint32_t *)stat);
2851 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2852 		      priv->sh->ibdev_path,
2853 		      priv->dev_port,
2854 		      ctr_name);
2855 		fd = open(path, O_RDONLY);
2856 		/*
2857 		 * in switchdev the file location is not per port
2858 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2859 		 */
2860 		if (fd == -1) {
2861 			MKSTR(path1, "%s/hw_counters/%s",
2862 			      priv->sh->ibdev_path,
2863 			      ctr_name);
2864 			fd = open(path1, O_RDONLY);
2865 		}
2866 		if (fd != -1) {
2867 			char buf[21] = {'\0'};
2868 			ssize_t n = read(fd, buf, sizeof(buf));
2869 
2870 			close(fd);
2871 			if (n != -1) {
2872 				*stat = strtoull(buf, NULL, 10);
2873 				return 0;
2874 			}
2875 		}
2876 	}
2877 	*stat = 0;
2878 	return 1;
2879 }
2880 
2881 /**
2882  * Set the reg_mr and dereg_mr call backs
2883  *
2884  * @param reg_mr_cb[out]
2885  *   Pointer to reg_mr func
2886  * @param dereg_mr_cb[out]
2887  *   Pointer to dereg_mr func
2888  *
2889  */
2890 void
2891 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2892 		      mlx5_dereg_mr_t *dereg_mr_cb)
2893 {
2894 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2895 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2896 }
2897 
2898 /**
2899  * Remove a MAC address from device
2900  *
2901  * @param dev
2902  *   Pointer to Ethernet device structure.
2903  * @param index
2904  *   MAC address index.
2905  */
2906 void
2907 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2908 {
2909 	struct mlx5_priv *priv = dev->data->dev_private;
2910 	const int vf = priv->config.vf;
2911 
2912 	if (vf)
2913 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2914 					mlx5_ifindex(dev), priv->mac_own,
2915 					&dev->data->mac_addrs[index], index);
2916 }
2917 
2918 /**
2919  * Adds a MAC address to the device
2920  *
2921  * @param dev
2922  *   Pointer to Ethernet device structure.
2923  * @param mac_addr
2924  *   MAC address to register.
2925  * @param index
2926  *   MAC address index.
2927  *
2928  * @return
2929  *   0 on success, a negative errno value otherwise
2930  */
2931 int
2932 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2933 		     uint32_t index)
2934 {
2935 	struct mlx5_priv *priv = dev->data->dev_private;
2936 	const int vf = priv->config.vf;
2937 	int ret = 0;
2938 
2939 	if (vf)
2940 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2941 					   mlx5_ifindex(dev), priv->mac_own,
2942 					   mac, index);
2943 	return ret;
2944 }
2945 
2946 /**
2947  * Modify a VF MAC address
2948  *
2949  * @param priv
2950  *   Pointer to device private data.
2951  * @param mac_addr
2952  *   MAC address to modify into.
2953  * @param iface_idx
2954  *   Net device interface index
2955  * @param vf_index
2956  *   VF index
2957  *
2958  * @return
2959  *   0 on success, a negative errno value otherwise
2960  */
2961 int
2962 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2963 			   unsigned int iface_idx,
2964 			   struct rte_ether_addr *mac_addr,
2965 			   int vf_index)
2966 {
2967 	return mlx5_nl_vf_mac_addr_modify
2968 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2969 }
2970 
2971 /**
2972  * Set device promiscuous mode
2973  *
2974  * @param dev
2975  *   Pointer to Ethernet device structure.
2976  * @param enable
2977  *   0 - promiscuous is disabled, otherwise - enabled
2978  *
2979  * @return
2980  *   0 on success, a negative error value otherwise
2981  */
2982 int
2983 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2984 {
2985 	struct mlx5_priv *priv = dev->data->dev_private;
2986 
2987 	return mlx5_nl_promisc(priv->nl_socket_route,
2988 			       mlx5_ifindex(dev), !!enable);
2989 }
2990 
2991 /**
2992  * Set device promiscuous mode
2993  *
2994  * @param dev
2995  *   Pointer to Ethernet device structure.
2996  * @param enable
2997  *   0 - all multicase is disabled, otherwise - enabled
2998  *
2999  * @return
3000  *   0 on success, a negative error value otherwise
3001  */
3002 int
3003 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
3004 {
3005 	struct mlx5_priv *priv = dev->data->dev_private;
3006 
3007 	return mlx5_nl_allmulti(priv->nl_socket_route,
3008 				mlx5_ifindex(dev), !!enable);
3009 }
3010 
3011 /**
3012  * Flush device MAC addresses
3013  *
3014  * @param dev
3015  *   Pointer to Ethernet device structure.
3016  *
3017  */
3018 void
3019 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
3020 {
3021 	struct mlx5_priv *priv = dev->data->dev_private;
3022 
3023 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
3024 			       dev->data->mac_addrs,
3025 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
3026 }
3027