xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision e39226bde5862c7c66b3481c6cd8ff87623c24db)
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 /**
73  * Set the completion channel file descriptor interrupt as non-blocking.
74  *
75  * @param[in] rxq_obj
76  *   Pointer to RQ channel object, which includes the channel fd
77  *
78  * @param[out] fd
79  *   The file descriptor (representing the intetrrupt) used in this channel.
80  *
81  * @return
82  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
83  */
84 int
85 mlx5_os_set_nonblock_channel_fd(int fd)
86 {
87 	int flags;
88 
89 	flags = fcntl(fd, F_GETFL);
90 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
91 }
92 
93 /**
94  * Get mlx5 device attributes. The glue function query_device_ex() is called
95  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
96  * device attributes from the glue out parameter.
97  *
98  * @param dev
99  *   Pointer to ibv context.
100  *
101  * @param device_attr
102  *   Pointer to mlx5 device attributes.
103  *
104  * @return
105  *   0 on success, non zero error number otherwise
106  */
107 int
108 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
109 {
110 	int err;
111 	struct ibv_device_attr_ex attr_ex;
112 	memset(device_attr, 0, sizeof(*device_attr));
113 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
114 	if (err)
115 		return err;
116 
117 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
118 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
119 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
120 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
121 	device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
122 	device_attr->max_mr = attr_ex.orig_attr.max_mr;
123 	device_attr->max_pd = attr_ex.orig_attr.max_pd;
124 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
125 	device_attr->max_srq = attr_ex.orig_attr.max_srq;
126 	device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
127 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
128 	device_attr->max_rwq_indirection_table_size =
129 		attr_ex.rss_caps.max_rwq_indirection_table_size;
130 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
131 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
132 
133 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
134 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
135 	if (err)
136 		return err;
137 
138 	device_attr->flags = dv_attr.flags;
139 	device_attr->comp_mask = dv_attr.comp_mask;
140 #ifdef HAVE_IBV_MLX5_MOD_SWP
141 	device_attr->sw_parsing_offloads =
142 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
143 #endif
144 	device_attr->min_single_stride_log_num_of_bytes =
145 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
146 	device_attr->max_single_stride_log_num_of_bytes =
147 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
148 	device_attr->min_single_wqe_log_num_of_strides =
149 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
150 	device_attr->max_single_wqe_log_num_of_strides =
151 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
152 	device_attr->stride_supported_qpts =
153 		dv_attr.striding_rq_caps.supported_qpts;
154 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
155 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
156 #endif
157 	strlcpy(device_attr->fw_ver, attr_ex.orig_attr.fw_ver,
158 		sizeof(device_attr->fw_ver));
159 
160 	return err;
161 }
162 
163 /**
164  * Verbs callback to allocate a memory. This function should allocate the space
165  * according to the size provided residing inside a huge page.
166  * Please note that all allocation must respect the alignment from libmlx5
167  * (i.e. currently rte_mem_page_size()).
168  *
169  * @param[in] size
170  *   The size in bytes of the memory to allocate.
171  * @param[in] data
172  *   A pointer to the callback data.
173  *
174  * @return
175  *   Allocated buffer, NULL otherwise and rte_errno is set.
176  */
177 static void *
178 mlx5_alloc_verbs_buf(size_t size, void *data)
179 {
180 	struct mlx5_dev_ctx_shared *sh = data;
181 	void *ret;
182 	size_t alignment = rte_mem_page_size();
183 	if (alignment == (size_t)-1) {
184 		DRV_LOG(ERR, "Failed to get mem page size");
185 		rte_errno = ENOMEM;
186 		return NULL;
187 	}
188 
189 	MLX5_ASSERT(data != NULL);
190 	ret = mlx5_malloc(0, size, alignment, sh->numa_node);
191 	if (!ret && size)
192 		rte_errno = ENOMEM;
193 	return ret;
194 }
195 
196 /**
197  * Verbs callback to free a memory.
198  *
199  * @param[in] ptr
200  *   A pointer to the memory to free.
201  * @param[in] data
202  *   A pointer to the callback data.
203  */
204 static void
205 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
206 {
207 	MLX5_ASSERT(data != NULL);
208 	mlx5_free(ptr);
209 }
210 
211 /**
212  * Initialize DR related data within private structure.
213  * Routine checks the reference counter and does actual
214  * resources creation/initialization only if counter is zero.
215  *
216  * @param[in] priv
217  *   Pointer to the private device data structure.
218  *
219  * @return
220  *   Zero on success, positive error code otherwise.
221  */
222 static int
223 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
224 {
225 	struct mlx5_dev_ctx_shared *sh = priv->sh;
226 	char s[MLX5_HLIST_NAMESIZE] __rte_unused;
227 	int err;
228 
229 	MLX5_ASSERT(sh && sh->refcnt);
230 	if (sh->refcnt > 1)
231 		return 0;
232 	err = mlx5_alloc_table_hash_list(priv);
233 	if (err)
234 		goto error;
235 	/* The resources below are only valid with DV support. */
236 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
237 	/* Init port id action cache list. */
238 	snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
239 	mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
240 			     flow_dv_port_id_create_cb,
241 			     flow_dv_port_id_match_cb,
242 			     flow_dv_port_id_remove_cb);
243 	/* Init push vlan action cache list. */
244 	snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
245 	mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
246 			     flow_dv_push_vlan_create_cb,
247 			     flow_dv_push_vlan_match_cb,
248 			     flow_dv_push_vlan_remove_cb);
249 	/* Init sample action cache list. */
250 	snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
251 	mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
252 			     flow_dv_sample_create_cb,
253 			     flow_dv_sample_match_cb,
254 			     flow_dv_sample_remove_cb);
255 	/* Init dest array action cache list. */
256 	snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
257 	mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
258 			     flow_dv_dest_array_create_cb,
259 			     flow_dv_dest_array_match_cb,
260 			     flow_dv_dest_array_remove_cb);
261 	/* Create tags hash list table. */
262 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
263 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
264 					  MLX5_HLIST_WRITE_MOST,
265 					  flow_dv_tag_create_cb,
266 					  flow_dv_tag_match_cb,
267 					  flow_dv_tag_remove_cb);
268 	if (!sh->tag_table) {
269 		DRV_LOG(ERR, "tags with hash creation failed.");
270 		err = ENOMEM;
271 		goto error;
272 	}
273 	sh->tag_table->ctx = sh;
274 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
275 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
276 					    0, MLX5_HLIST_WRITE_MOST |
277 					    MLX5_HLIST_DIRECT_KEY,
278 					    flow_dv_modify_create_cb,
279 					    flow_dv_modify_match_cb,
280 					    flow_dv_modify_remove_cb);
281 	if (!sh->modify_cmds) {
282 		DRV_LOG(ERR, "hdr modify hash creation failed");
283 		err = ENOMEM;
284 		goto error;
285 	}
286 	sh->modify_cmds->ctx = sh;
287 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
288 	sh->encaps_decaps = mlx5_hlist_create(s,
289 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
290 					      0, MLX5_HLIST_DIRECT_KEY |
291 					      MLX5_HLIST_WRITE_MOST,
292 					      flow_dv_encap_decap_create_cb,
293 					      flow_dv_encap_decap_match_cb,
294 					      flow_dv_encap_decap_remove_cb);
295 	if (!sh->encaps_decaps) {
296 		DRV_LOG(ERR, "encap decap hash creation failed");
297 		err = ENOMEM;
298 		goto error;
299 	}
300 	sh->encaps_decaps->ctx = sh;
301 #endif
302 #ifdef HAVE_MLX5DV_DR
303 	void *domain;
304 
305 	/* Reference counter is zero, we should initialize structures. */
306 	domain = mlx5_glue->dr_create_domain(sh->ctx,
307 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
308 	if (!domain) {
309 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
310 		err = errno;
311 		goto error;
312 	}
313 	sh->rx_domain = domain;
314 	domain = mlx5_glue->dr_create_domain(sh->ctx,
315 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
316 	if (!domain) {
317 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
318 		err = errno;
319 		goto error;
320 	}
321 	sh->tx_domain = domain;
322 #ifdef HAVE_MLX5DV_DR_ESWITCH
323 	if (priv->config.dv_esw_en) {
324 		domain  = mlx5_glue->dr_create_domain
325 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
326 		if (!domain) {
327 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
328 			err = errno;
329 			goto error;
330 		}
331 		sh->fdb_domain = domain;
332 	}
333 	/*
334 	 * The drop action is just some dummy placeholder in rdma-core. It
335 	 * does not belong to domains and has no any attributes, and, can be
336 	 * shared by the entire device.
337 	 */
338 	sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
339 	if (!sh->dr_drop_action) {
340 		DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
341 		err = errno;
342 		goto error;
343 	}
344 #endif
345 	if (!sh->tunnel_hub)
346 		err = mlx5_alloc_tunnel_hub(sh);
347 	if (err) {
348 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
349 		goto error;
350 	}
351 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
352 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
353 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
354 		if (sh->fdb_domain)
355 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
356 	}
357 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
358 	if (!priv->config.allow_duplicate_pattern) {
359 #ifndef HAVE_MLX5_DR_ALLOW_DUPLICATE
360 		DRV_LOG(WARNING, "Disallow duplicate pattern is not supported - maybe old rdma-core version?");
361 #endif
362 		mlx5_glue->dr_allow_duplicate_rules(sh->rx_domain, 0);
363 		mlx5_glue->dr_allow_duplicate_rules(sh->tx_domain, 0);
364 		if (sh->fdb_domain)
365 			mlx5_glue->dr_allow_duplicate_rules(sh->fdb_domain, 0);
366 	}
367 #endif /* HAVE_MLX5DV_DR */
368 	sh->default_miss_action =
369 			mlx5_glue->dr_create_flow_action_default_miss();
370 	if (!sh->default_miss_action)
371 		DRV_LOG(WARNING, "Default miss action is not supported.");
372 	return 0;
373 error:
374 	/* Rollback the created objects. */
375 	if (sh->rx_domain) {
376 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
377 		sh->rx_domain = NULL;
378 	}
379 	if (sh->tx_domain) {
380 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
381 		sh->tx_domain = NULL;
382 	}
383 	if (sh->fdb_domain) {
384 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
385 		sh->fdb_domain = NULL;
386 	}
387 	if (sh->dr_drop_action) {
388 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
389 		sh->dr_drop_action = NULL;
390 	}
391 	if (sh->pop_vlan_action) {
392 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
393 		sh->pop_vlan_action = NULL;
394 	}
395 	if (sh->encaps_decaps) {
396 		mlx5_hlist_destroy(sh->encaps_decaps);
397 		sh->encaps_decaps = NULL;
398 	}
399 	if (sh->modify_cmds) {
400 		mlx5_hlist_destroy(sh->modify_cmds);
401 		sh->modify_cmds = NULL;
402 	}
403 	if (sh->tag_table) {
404 		/* tags should be destroyed with flow before. */
405 		mlx5_hlist_destroy(sh->tag_table);
406 		sh->tag_table = NULL;
407 	}
408 	if (sh->tunnel_hub) {
409 		mlx5_release_tunnel_hub(sh, priv->dev_port);
410 		sh->tunnel_hub = NULL;
411 	}
412 	mlx5_free_table_hash_list(priv);
413 	return err;
414 }
415 
416 /**
417  * Destroy DR related data within private structure.
418  *
419  * @param[in] priv
420  *   Pointer to the private device data structure.
421  */
422 void
423 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
424 {
425 	struct mlx5_dev_ctx_shared *sh = priv->sh;
426 
427 	MLX5_ASSERT(sh && sh->refcnt);
428 	if (sh->refcnt > 1)
429 		return;
430 #ifdef HAVE_MLX5DV_DR
431 	if (sh->rx_domain) {
432 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
433 		sh->rx_domain = NULL;
434 	}
435 	if (sh->tx_domain) {
436 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
437 		sh->tx_domain = NULL;
438 	}
439 #ifdef HAVE_MLX5DV_DR_ESWITCH
440 	if (sh->fdb_domain) {
441 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
442 		sh->fdb_domain = NULL;
443 	}
444 	if (sh->dr_drop_action) {
445 		mlx5_glue->destroy_flow_action(sh->dr_drop_action);
446 		sh->dr_drop_action = NULL;
447 	}
448 #endif
449 	if (sh->pop_vlan_action) {
450 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
451 		sh->pop_vlan_action = NULL;
452 	}
453 #endif /* HAVE_MLX5DV_DR */
454 	if (sh->default_miss_action)
455 		mlx5_glue->destroy_flow_action
456 				(sh->default_miss_action);
457 	if (sh->encaps_decaps) {
458 		mlx5_hlist_destroy(sh->encaps_decaps);
459 		sh->encaps_decaps = NULL;
460 	}
461 	if (sh->modify_cmds) {
462 		mlx5_hlist_destroy(sh->modify_cmds);
463 		sh->modify_cmds = NULL;
464 	}
465 	if (sh->tag_table) {
466 		/* tags should be destroyed with flow before. */
467 		mlx5_hlist_destroy(sh->tag_table);
468 		sh->tag_table = NULL;
469 	}
470 	if (sh->tunnel_hub) {
471 		mlx5_release_tunnel_hub(sh, priv->dev_port);
472 		sh->tunnel_hub = NULL;
473 	}
474 	mlx5_cache_list_destroy(&sh->port_id_action_list);
475 	mlx5_cache_list_destroy(&sh->push_vlan_action_list);
476 	mlx5_free_table_hash_list(priv);
477 }
478 
479 /**
480  * Initialize shared data between primary and secondary process.
481  *
482  * A memzone is reserved by primary process and secondary processes attach to
483  * the memzone.
484  *
485  * @return
486  *   0 on success, a negative errno value otherwise and rte_errno is set.
487  */
488 static int
489 mlx5_init_shared_data(void)
490 {
491 	const struct rte_memzone *mz;
492 	int ret = 0;
493 
494 	rte_spinlock_lock(&mlx5_shared_data_lock);
495 	if (mlx5_shared_data == NULL) {
496 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
497 			/* Allocate shared memory. */
498 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
499 						 sizeof(*mlx5_shared_data),
500 						 SOCKET_ID_ANY, 0);
501 			if (mz == NULL) {
502 				DRV_LOG(ERR,
503 					"Cannot allocate mlx5 shared data");
504 				ret = -rte_errno;
505 				goto error;
506 			}
507 			mlx5_shared_data = mz->addr;
508 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
509 			rte_spinlock_init(&mlx5_shared_data->lock);
510 		} else {
511 			/* Lookup allocated shared memory. */
512 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
513 			if (mz == NULL) {
514 				DRV_LOG(ERR,
515 					"Cannot attach mlx5 shared data");
516 				ret = -rte_errno;
517 				goto error;
518 			}
519 			mlx5_shared_data = mz->addr;
520 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
521 		}
522 	}
523 error:
524 	rte_spinlock_unlock(&mlx5_shared_data_lock);
525 	return ret;
526 }
527 
528 /**
529  * PMD global initialization.
530  *
531  * Independent from individual device, this function initializes global
532  * per-PMD data structures distinguishing primary and secondary processes.
533  * Hence, each initialization is called once per a process.
534  *
535  * @return
536  *   0 on success, a negative errno value otherwise and rte_errno is set.
537  */
538 static int
539 mlx5_init_once(void)
540 {
541 	struct mlx5_shared_data *sd;
542 	struct mlx5_local_data *ld = &mlx5_local_data;
543 	int ret = 0;
544 
545 	if (mlx5_init_shared_data())
546 		return -rte_errno;
547 	sd = mlx5_shared_data;
548 	MLX5_ASSERT(sd);
549 	rte_spinlock_lock(&sd->lock);
550 	switch (rte_eal_process_type()) {
551 	case RTE_PROC_PRIMARY:
552 		if (sd->init_done)
553 			break;
554 		LIST_INIT(&sd->mem_event_cb_list);
555 		rte_rwlock_init(&sd->mem_event_rwlock);
556 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
557 						mlx5_mr_mem_event_cb, NULL);
558 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
559 					   mlx5_mp_os_primary_handle);
560 		if (ret)
561 			goto out;
562 		sd->init_done = true;
563 		break;
564 	case RTE_PROC_SECONDARY:
565 		if (ld->init_done)
566 			break;
567 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
568 					     mlx5_mp_os_secondary_handle);
569 		if (ret)
570 			goto out;
571 		++sd->secondary_cnt;
572 		ld->init_done = true;
573 		break;
574 	default:
575 		break;
576 	}
577 out:
578 	rte_spinlock_unlock(&sd->lock);
579 	return ret;
580 }
581 
582 /**
583  * Create the Tx queue DevX/Verbs object.
584  *
585  * @param dev
586  *   Pointer to Ethernet device.
587  * @param idx
588  *   Queue index in DPDK Tx queue array.
589  *
590  * @return
591  *   0 on success, a negative errno value otherwise and rte_errno is set.
592  */
593 static int
594 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
595 {
596 	struct mlx5_priv *priv = dev->data->dev_private;
597 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
598 	struct mlx5_txq_ctrl *txq_ctrl =
599 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
600 
601 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
602 		return mlx5_txq_devx_obj_new(dev, idx);
603 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
604 	if (!priv->config.dv_esw_en)
605 		return mlx5_txq_devx_obj_new(dev, idx);
606 #endif
607 	return mlx5_txq_ibv_obj_new(dev, idx);
608 }
609 
610 /**
611  * Release an Tx DevX/verbs queue object.
612  *
613  * @param txq_obj
614  *   DevX/Verbs Tx queue object.
615  */
616 static void
617 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
618 {
619 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
620 		mlx5_txq_devx_obj_release(txq_obj);
621 		return;
622 	}
623 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
624 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
625 		mlx5_txq_devx_obj_release(txq_obj);
626 		return;
627 	}
628 #endif
629 	mlx5_txq_ibv_obj_release(txq_obj);
630 }
631 
632 /**
633  * DV flow counter mode detect and config.
634  *
635  * @param dev
636  *   Pointer to rte_eth_dev structure.
637  *
638  */
639 static void
640 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
641 {
642 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
643 	struct mlx5_priv *priv = dev->data->dev_private;
644 	struct mlx5_dev_ctx_shared *sh = priv->sh;
645 	bool fallback;
646 
647 #ifndef HAVE_IBV_DEVX_ASYNC
648 	fallback = true;
649 #else
650 	fallback = false;
651 	if (!priv->config.devx || !priv->config.dv_flow_en ||
652 	    !priv->config.hca_attr.flow_counters_dump ||
653 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
654 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
655 		fallback = true;
656 #endif
657 	if (fallback)
658 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
659 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
660 			priv->config.hca_attr.flow_counters_dump,
661 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
662 	/* Initialize fallback mode only on the port initializes sh. */
663 	if (sh->refcnt == 1)
664 		sh->cmng.counter_fallback = fallback;
665 	else if (fallback != sh->cmng.counter_fallback)
666 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
667 			"with others:%d.", PORT_ID(priv), fallback);
668 #endif
669 }
670 
671 static void
672 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
673 {
674 	struct mlx5_priv *priv = dev->data->dev_private;
675 	void *ctx = priv->sh->ctx;
676 
677 	priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
678 	if (!priv->q_counters) {
679 		struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
680 		struct ibv_wq *wq;
681 
682 		DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
683 			"by DevX - fall-back to use the kernel driver global "
684 			"queue counter.", dev->data->port_id);
685 		/* Create WQ by kernel and query its queue counter ID. */
686 		if (cq) {
687 			wq = mlx5_glue->create_wq(ctx,
688 						  &(struct ibv_wq_init_attr){
689 						    .wq_type = IBV_WQT_RQ,
690 						    .max_wr = 1,
691 						    .max_sge = 1,
692 						    .pd = priv->sh->pd,
693 						    .cq = cq,
694 						});
695 			if (wq) {
696 				/* Counter is assigned only on RDY state. */
697 				int ret = mlx5_glue->modify_wq(wq,
698 						 &(struct ibv_wq_attr){
699 						 .attr_mask = IBV_WQ_ATTR_STATE,
700 						 .wq_state = IBV_WQS_RDY,
701 						});
702 
703 				if (ret == 0)
704 					mlx5_devx_cmd_wq_query(wq,
705 							 &priv->counter_set_id);
706 				claim_zero(mlx5_glue->destroy_wq(wq));
707 			}
708 			claim_zero(mlx5_glue->destroy_cq(cq));
709 		}
710 	} else {
711 		priv->counter_set_id = priv->q_counters->id;
712 	}
713 	if (priv->counter_set_id == 0)
714 		DRV_LOG(INFO, "Part of the port %d statistics will not be "
715 			"available.", dev->data->port_id);
716 }
717 
718 /**
719  * Check if representor spawn info match devargs.
720  *
721  * @param spawn
722  *   Verbs device parameters (name, port, switch_info) to spawn.
723  * @param eth_da
724  *   Device devargs to probe.
725  *
726  * @return
727  *   Match result.
728  */
729 static bool
730 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
731 		       struct rte_eth_devargs *eth_da)
732 {
733 	struct mlx5_switch_info *switch_info = &spawn->info;
734 	unsigned int p, f;
735 	uint16_t id;
736 	uint16_t repr_id = mlx5_representor_id_encode(switch_info,
737 						      eth_da->type);
738 
739 	switch (eth_da->type) {
740 	case RTE_ETH_REPRESENTOR_SF:
741 		if (!(spawn->info.port_name == -1 &&
742 		      switch_info->name_type ==
743 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
744 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
745 			rte_errno = EBUSY;
746 			return false;
747 		}
748 		break;
749 	case RTE_ETH_REPRESENTOR_VF:
750 		/* Allows HPF representor index -1 as exception. */
751 		if (!(spawn->info.port_name == -1 &&
752 		      switch_info->name_type ==
753 				MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
754 		    switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
755 			rte_errno = EBUSY;
756 			return false;
757 		}
758 		break;
759 	case RTE_ETH_REPRESENTOR_NONE:
760 		rte_errno = EBUSY;
761 		return false;
762 	default:
763 		rte_errno = ENOTSUP;
764 		DRV_LOG(ERR, "unsupported representor type");
765 		return false;
766 	}
767 	/* Check representor ID: */
768 	for (p = 0; p < eth_da->nb_ports; ++p) {
769 		if (spawn->pf_bond < 0) {
770 			/* For non-LAG mode, allow and ignore pf. */
771 			switch_info->pf_num = eth_da->ports[p];
772 			repr_id = mlx5_representor_id_encode(switch_info,
773 							     eth_da->type);
774 		}
775 		for (f = 0; f < eth_da->nb_representor_ports; ++f) {
776 			id = MLX5_REPRESENTOR_ID
777 				(eth_da->ports[p], eth_da->type,
778 				 eth_da->representor_ports[f]);
779 			if (repr_id == id)
780 				return true;
781 		}
782 	}
783 	rte_errno = EBUSY;
784 	return false;
785 }
786 
787 
788 /**
789  * Spawn an Ethernet device from Verbs information.
790  *
791  * @param dpdk_dev
792  *   Backing DPDK device.
793  * @param spawn
794  *   Verbs device parameters (name, port, switch_info) to spawn.
795  * @param config
796  *   Device configuration parameters.
797  * @param config
798  *   Device arguments.
799  *
800  * @return
801  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
802  *   is set. The following errors are defined:
803  *
804  *   EBUSY: device is not supposed to be spawned.
805  *   EEXIST: device is already spawned
806  */
807 static struct rte_eth_dev *
808 mlx5_dev_spawn(struct rte_device *dpdk_dev,
809 	       struct mlx5_dev_spawn_data *spawn,
810 	       struct mlx5_dev_config *config,
811 	       struct rte_eth_devargs *eth_da)
812 {
813 	const struct mlx5_switch_info *switch_info = &spawn->info;
814 	struct mlx5_dev_ctx_shared *sh = NULL;
815 	struct ibv_port_attr port_attr;
816 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
817 	struct rte_eth_dev *eth_dev = NULL;
818 	struct mlx5_priv *priv = NULL;
819 	int err = 0;
820 	unsigned int hw_padding = 0;
821 	unsigned int mps;
822 	unsigned int tunnel_en = 0;
823 	unsigned int mpls_en = 0;
824 	unsigned int swp = 0;
825 	unsigned int mprq = 0;
826 	unsigned int mprq_min_stride_size_n = 0;
827 	unsigned int mprq_max_stride_size_n = 0;
828 	unsigned int mprq_min_stride_num_n = 0;
829 	unsigned int mprq_max_stride_num_n = 0;
830 	struct rte_ether_addr mac;
831 	char name[RTE_ETH_NAME_MAX_LEN];
832 	int own_domain_id = 0;
833 	uint16_t port_id;
834 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
835 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
836 #endif
837 
838 	/* Determine if this port representor is supposed to be spawned. */
839 	if (switch_info->representor && dpdk_dev->devargs &&
840 	    !mlx5_representor_match(spawn, eth_da))
841 		return NULL;
842 	/* Build device name. */
843 	if (spawn->pf_bond < 0) {
844 		/* Single device. */
845 		if (!switch_info->representor)
846 			strlcpy(name, dpdk_dev->name, sizeof(name));
847 		else
848 			err = snprintf(name, sizeof(name), "%s_representor_%s%u",
849 				 dpdk_dev->name,
850 				 switch_info->name_type ==
851 				 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
852 				 switch_info->port_name);
853 	} else {
854 		/* Bonding device. */
855 		if (!switch_info->representor) {
856 			err = snprintf(name, sizeof(name), "%s_%s",
857 				 dpdk_dev->name,
858 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
859 		} else {
860 			err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
861 				dpdk_dev->name,
862 				mlx5_os_get_dev_device_name(spawn->phys_dev),
863 				switch_info->ctrl_num,
864 				switch_info->pf_num,
865 				switch_info->name_type ==
866 				MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
867 				switch_info->port_name);
868 		}
869 	}
870 	if (err >= (int)sizeof(name))
871 		DRV_LOG(WARNING, "device name overflow %s", name);
872 	/* check if the device is already spawned */
873 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
874 		rte_errno = EEXIST;
875 		return NULL;
876 	}
877 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
878 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
879 		struct mlx5_mp_id mp_id;
880 
881 		eth_dev = rte_eth_dev_attach_secondary(name);
882 		if (eth_dev == NULL) {
883 			DRV_LOG(ERR, "can not attach rte ethdev");
884 			rte_errno = ENOMEM;
885 			return NULL;
886 		}
887 		eth_dev->device = dpdk_dev;
888 		eth_dev->dev_ops = &mlx5_dev_sec_ops;
889 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
890 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
891 		err = mlx5_proc_priv_init(eth_dev);
892 		if (err)
893 			return NULL;
894 		mp_id.port_id = eth_dev->data->port_id;
895 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
896 		/* Receive command fd from primary process */
897 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
898 		if (err < 0)
899 			goto err_secondary;
900 		/* Remap UAR for Tx queues. */
901 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
902 		if (err)
903 			goto err_secondary;
904 		/*
905 		 * Ethdev pointer is still required as input since
906 		 * the primary device is not accessible from the
907 		 * secondary process.
908 		 */
909 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
910 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
911 		return eth_dev;
912 err_secondary:
913 		mlx5_dev_close(eth_dev);
914 		return NULL;
915 	}
916 	/*
917 	 * Some parameters ("tx_db_nc" in particularly) are needed in
918 	 * advance to create dv/verbs device context. We proceed the
919 	 * devargs here to get ones, and later proceed devargs again
920 	 * to override some hardware settings.
921 	 */
922 	err = mlx5_args(config, dpdk_dev->devargs);
923 	if (err) {
924 		err = rte_errno;
925 		DRV_LOG(ERR, "failed to process device arguments: %s",
926 			strerror(rte_errno));
927 		goto error;
928 	}
929 	if (config->dv_miss_info) {
930 		if (switch_info->master || switch_info->representor)
931 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
932 	}
933 	mlx5_malloc_mem_select(config->sys_mem_en);
934 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
935 	if (!sh)
936 		return NULL;
937 	config->devx = sh->devx;
938 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
939 	config->dest_tir = 1;
940 #endif
941 #ifdef HAVE_IBV_MLX5_MOD_SWP
942 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
943 #endif
944 	/*
945 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
946 	 * as all ConnectX-5 devices.
947 	 */
948 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
949 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
950 #endif
951 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
952 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
953 #endif
954 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
955 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
956 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
957 			DRV_LOG(DEBUG, "enhanced MPW is supported");
958 			mps = MLX5_MPW_ENHANCED;
959 		} else {
960 			DRV_LOG(DEBUG, "MPW is supported");
961 			mps = MLX5_MPW;
962 		}
963 	} else {
964 		DRV_LOG(DEBUG, "MPW isn't supported");
965 		mps = MLX5_MPW_DISABLED;
966 	}
967 #ifdef HAVE_IBV_MLX5_MOD_SWP
968 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
969 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
970 	DRV_LOG(DEBUG, "SWP support: %u", swp);
971 #endif
972 	config->swp = !!swp;
973 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
974 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
975 		struct mlx5dv_striding_rq_caps mprq_caps =
976 			dv_attr.striding_rq_caps;
977 
978 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
979 			mprq_caps.min_single_stride_log_num_of_bytes);
980 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
981 			mprq_caps.max_single_stride_log_num_of_bytes);
982 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
983 			mprq_caps.min_single_wqe_log_num_of_strides);
984 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
985 			mprq_caps.max_single_wqe_log_num_of_strides);
986 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
987 			mprq_caps.supported_qpts);
988 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
989 		mprq = 1;
990 		mprq_min_stride_size_n =
991 			mprq_caps.min_single_stride_log_num_of_bytes;
992 		mprq_max_stride_size_n =
993 			mprq_caps.max_single_stride_log_num_of_bytes;
994 		mprq_min_stride_num_n =
995 			mprq_caps.min_single_wqe_log_num_of_strides;
996 		mprq_max_stride_num_n =
997 			mprq_caps.max_single_wqe_log_num_of_strides;
998 	}
999 #endif
1000 	/* Rx CQE compression is enabled by default. */
1001 	config->cqe_comp = 1;
1002 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1003 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1004 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
1005 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1006 			     (dv_attr.tunnel_offloads_caps &
1007 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
1008 			     (dv_attr.tunnel_offloads_caps &
1009 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
1010 	}
1011 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1012 		tunnel_en ? "" : "not ");
1013 #else
1014 	DRV_LOG(WARNING,
1015 		"tunnel offloading disabled due to old OFED/rdma-core version");
1016 #endif
1017 	config->tunnel_en = tunnel_en;
1018 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1019 	mpls_en = ((dv_attr.tunnel_offloads_caps &
1020 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1021 		   (dv_attr.tunnel_offloads_caps &
1022 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1023 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1024 		mpls_en ? "" : "not ");
1025 #else
1026 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1027 		" old OFED/rdma-core version or firmware configuration");
1028 #endif
1029 	config->mpls_en = mpls_en;
1030 	/* Check port status. */
1031 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1032 	if (err) {
1033 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
1034 		goto error;
1035 	}
1036 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1037 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
1038 		err = EINVAL;
1039 		goto error;
1040 	}
1041 	if (port_attr.state != IBV_PORT_ACTIVE)
1042 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1043 			mlx5_glue->port_state_str(port_attr.state),
1044 			port_attr.state);
1045 	/* Allocate private eth device data. */
1046 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1047 			   sizeof(*priv),
1048 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1049 	if (priv == NULL) {
1050 		DRV_LOG(ERR, "priv allocation failure");
1051 		err = ENOMEM;
1052 		goto error;
1053 	}
1054 	priv->sh = sh;
1055 	priv->dev_port = spawn->phys_port;
1056 	priv->pci_dev = spawn->pci_dev;
1057 	priv->mtu = RTE_ETHER_MTU;
1058 	/* Some internal functions rely on Netlink sockets, open them now. */
1059 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1060 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
1061 	priv->representor = !!switch_info->representor;
1062 	priv->master = !!switch_info->master;
1063 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1064 	priv->vport_meta_tag = 0;
1065 	priv->vport_meta_mask = 0;
1066 	priv->pf_bond = spawn->pf_bond;
1067 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1068 	/*
1069 	 * The DevX port query API is implemented. E-Switch may use
1070 	 * either vport or reg_c[0] metadata register to match on
1071 	 * vport index. The engaged part of metadata register is
1072 	 * defined by mask.
1073 	 */
1074 	if (switch_info->representor || switch_info->master) {
1075 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1076 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1077 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1078 						 &devx_port);
1079 		if (err) {
1080 			DRV_LOG(WARNING,
1081 				"can't query devx port %d on device %s",
1082 				spawn->phys_port,
1083 				mlx5_os_get_dev_device_name(spawn->phys_dev));
1084 			devx_port.comp_mask = 0;
1085 		}
1086 	}
1087 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1088 		priv->vport_meta_tag = devx_port.reg_c_0.value;
1089 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
1090 		if (!priv->vport_meta_mask) {
1091 			DRV_LOG(ERR, "vport zero mask for port %d"
1092 				     " on bonding device %s",
1093 				     spawn->phys_port,
1094 				     mlx5_os_get_dev_device_name
1095 							(spawn->phys_dev));
1096 			err = ENOTSUP;
1097 			goto error;
1098 		}
1099 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1100 			DRV_LOG(ERR, "invalid vport tag for port %d"
1101 				     " on bonding device %s",
1102 				     spawn->phys_port,
1103 				     mlx5_os_get_dev_device_name
1104 							(spawn->phys_dev));
1105 			err = ENOTSUP;
1106 			goto error;
1107 		}
1108 	}
1109 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1110 		priv->vport_id = devx_port.vport_num;
1111 	} else if (spawn->pf_bond >= 0 &&
1112 		   (switch_info->representor || switch_info->master)) {
1113 		DRV_LOG(ERR, "can't deduce vport index for port %d"
1114 			     " on bonding device %s",
1115 			     spawn->phys_port,
1116 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
1117 		err = ENOTSUP;
1118 		goto error;
1119 	} else {
1120 		/* Suppose vport index in compatible way. */
1121 		priv->vport_id = switch_info->representor ?
1122 				 switch_info->port_name + 1 : -1;
1123 	}
1124 #else
1125 	/*
1126 	 * Kernel/rdma_core support single E-Switch per PF configurations
1127 	 * only and vport_id field contains the vport index for
1128 	 * associated VF, which is deduced from representor port name.
1129 	 * For example, let's have the IB device port 10, it has
1130 	 * attached network device eth0, which has port name attribute
1131 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
1132 	 * as 3 (2+1). This assigning schema should be changed if the
1133 	 * multiple E-Switch instances per PF configurations or/and PCI
1134 	 * subfunctions are added.
1135 	 */
1136 	priv->vport_id = switch_info->representor ?
1137 			 switch_info->port_name + 1 : -1;
1138 #endif
1139 	priv->representor_id = mlx5_representor_id_encode(switch_info,
1140 							  eth_da->type);
1141 	/*
1142 	 * Look for sibling devices in order to reuse their switch domain
1143 	 * if any, otherwise allocate one.
1144 	 */
1145 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1146 		const struct mlx5_priv *opriv =
1147 			rte_eth_devices[port_id].data->dev_private;
1148 
1149 		if (!opriv ||
1150 		    opriv->sh != priv->sh ||
1151 			opriv->domain_id ==
1152 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1153 			continue;
1154 		priv->domain_id = opriv->domain_id;
1155 		break;
1156 	}
1157 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1158 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1159 		if (err) {
1160 			err = rte_errno;
1161 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1162 				strerror(rte_errno));
1163 			goto error;
1164 		}
1165 		own_domain_id = 1;
1166 	}
1167 	/* Override some values set by hardware configuration. */
1168 	mlx5_args(config, dpdk_dev->devargs);
1169 	err = mlx5_dev_check_sibling_config(priv, config);
1170 	if (err)
1171 		goto error;
1172 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1173 			    IBV_DEVICE_RAW_IP_CSUM);
1174 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1175 		(config->hw_csum ? "" : "not "));
1176 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1177 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1178 	DRV_LOG(DEBUG, "counters are not supported");
1179 #endif
1180 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1181 	if (config->dv_flow_en) {
1182 		DRV_LOG(WARNING, "DV flow is not supported");
1183 		config->dv_flow_en = 0;
1184 	}
1185 #endif
1186 	config->ind_table_max_size =
1187 		sh->device_attr.max_rwq_indirection_table_size;
1188 	/*
1189 	 * Remove this check once DPDK supports larger/variable
1190 	 * indirection tables.
1191 	 */
1192 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1193 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1194 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1195 		config->ind_table_max_size);
1196 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1197 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1198 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1199 		(config->hw_vlan_strip ? "" : "not "));
1200 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1201 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1202 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1203 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1204 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1205 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1206 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1207 #endif
1208 	if (config->hw_padding && !hw_padding) {
1209 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1210 		config->hw_padding = 0;
1211 	} else if (config->hw_padding) {
1212 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1213 	}
1214 	config->tso = (sh->device_attr.max_tso > 0 &&
1215 		      (sh->device_attr.tso_supported_qpts &
1216 		       (1 << IBV_QPT_RAW_PACKET)));
1217 	if (config->tso)
1218 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1219 	/*
1220 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1221 	 * by default.
1222 	 */
1223 	if (config->mps == MLX5_ARG_UNSET)
1224 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1225 							  MLX5_MPW_DISABLED;
1226 	else
1227 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1228 	DRV_LOG(INFO, "%sMPS is %s",
1229 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1230 		config->mps == MLX5_MPW ? "legacy " : "",
1231 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1232 	if (config->devx) {
1233 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1234 		if (err) {
1235 			err = -err;
1236 			goto error;
1237 		}
1238 		/* Check relax ordering support. */
1239 		if (!haswell_broadwell_cpu) {
1240 			sh->cmng.relaxed_ordering_write =
1241 				config->hca_attr.relaxed_ordering_write;
1242 			sh->cmng.relaxed_ordering_read =
1243 				config->hca_attr.relaxed_ordering_read;
1244 		} else {
1245 			sh->cmng.relaxed_ordering_read = 0;
1246 			sh->cmng.relaxed_ordering_write = 0;
1247 		}
1248 		sh->rq_ts_format = config->hca_attr.rq_ts_format;
1249 		sh->sq_ts_format = config->hca_attr.sq_ts_format;
1250 		sh->qp_ts_format = config->hca_attr.qp_ts_format;
1251 		/* Check for LRO support. */
1252 		if (config->dest_tir && config->hca_attr.lro_cap &&
1253 		    config->dv_flow_en) {
1254 			/* TBD check tunnel lro caps. */
1255 			config->lro.supported = config->hca_attr.lro_cap;
1256 			DRV_LOG(DEBUG, "Device supports LRO");
1257 			/*
1258 			 * If LRO timeout is not configured by application,
1259 			 * use the minimal supported value.
1260 			 */
1261 			if (!config->lro.timeout)
1262 				config->lro.timeout =
1263 				config->hca_attr.lro_timer_supported_periods[0];
1264 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1265 				config->lro.timeout);
1266 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1267 				"required for coalescing is %d bytes",
1268 				config->hca_attr.lro_min_mss_size);
1269 		}
1270 #if defined(HAVE_MLX5DV_DR) && \
1271 	(defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1272 	 defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1273 		if (config->hca_attr.qos.sup &&
1274 		    config->hca_attr.qos.flow_meter_old &&
1275 		    config->dv_flow_en) {
1276 			uint8_t reg_c_mask =
1277 				config->hca_attr.qos.flow_meter_reg_c_ids;
1278 			/*
1279 			 * Meter needs two REG_C's for color match and pre-sfx
1280 			 * flow match. Here get the REG_C for color match.
1281 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1282 			 */
1283 			reg_c_mask &= 0xfc;
1284 			if (__builtin_popcount(reg_c_mask) < 1) {
1285 				priv->mtr_en = 0;
1286 				DRV_LOG(WARNING, "No available register for"
1287 					" meter.");
1288 			} else {
1289 				/*
1290 				 * The meter color register is used by the
1291 				 * flow-hit feature as well.
1292 				 * The flow-hit feature must use REG_C_3
1293 				 * Prefer REG_C_3 if it is available.
1294 				 */
1295 				if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1296 					priv->mtr_color_reg = REG_C_3;
1297 				else
1298 					priv->mtr_color_reg = ffs(reg_c_mask)
1299 							      - 1 + REG_C_0;
1300 				priv->mtr_en = 1;
1301 				priv->mtr_reg_share =
1302 				      config->hca_attr.qos.flow_meter;
1303 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1304 					priv->mtr_color_reg);
1305 			}
1306 		}
1307 		if (config->hca_attr.qos.sup &&
1308 			config->hca_attr.qos.flow_meter_aso_sup) {
1309 			uint32_t log_obj_size =
1310 				rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1311 			if (log_obj_size >=
1312 			config->hca_attr.qos.log_meter_aso_granularity &&
1313 			log_obj_size <=
1314 			config->hca_attr.qos.log_meter_aso_max_alloc)
1315 				sh->meter_aso_en = 1;
1316 		}
1317 		if (priv->mtr_en) {
1318 			err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1319 			if (err) {
1320 				err = -err;
1321 				goto error;
1322 			}
1323 		}
1324 #endif
1325 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1326 		if (config->hca_attr.flow_hit_aso &&
1327 		    priv->mtr_color_reg == REG_C_3) {
1328 			sh->flow_hit_aso_en = 1;
1329 			err = mlx5_flow_aso_age_mng_init(sh);
1330 			if (err) {
1331 				err = -err;
1332 				goto error;
1333 			}
1334 			DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1335 		}
1336 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1337 #if defined(HAVE_MLX5_DR_CREATE_ACTION_ASO) && \
1338 	defined(HAVE_MLX5_DR_ACTION_ASO_CT)
1339 		if (config->hca_attr.ct_offload &&
1340 		    priv->mtr_color_reg == REG_C_3) {
1341 			err = mlx5_flow_aso_ct_mng_init(sh);
1342 			if (err) {
1343 				err = -err;
1344 				goto error;
1345 			}
1346 			DRV_LOG(DEBUG, "CT ASO is supported.");
1347 			sh->ct_aso_en = 1;
1348 		}
1349 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO && HAVE_MLX5_DR_ACTION_ASO_CT */
1350 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1351 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1352 		    config->dv_flow_en) {
1353 			priv->sampler_en = 1;
1354 			DRV_LOG(DEBUG, "Sampler enabled!");
1355 		} else {
1356 			priv->sampler_en = 0;
1357 			if (!config->hca_attr.log_max_ft_sampler_num)
1358 				DRV_LOG(WARNING,
1359 					"No available register for sampler.");
1360 			else
1361 				DRV_LOG(DEBUG, "DV flow is not supported!");
1362 		}
1363 #endif
1364 	}
1365 	if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1366 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1367 		DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1368 		config->cqe_comp = 0;
1369 	}
1370 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1371 	    (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1372 		DRV_LOG(WARNING, "Flow Tag CQE compression"
1373 				 " format isn't supported.");
1374 		config->cqe_comp = 0;
1375 	}
1376 	if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1377 	    (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1378 		DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1379 				 " format isn't supported.");
1380 		config->cqe_comp = 0;
1381 	}
1382 	DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1383 			config->cqe_comp ? "" : "not ");
1384 	if (config->tx_pp) {
1385 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1386 			config->hca_attr.dev_freq_khz);
1387 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1388 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1389 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1390 			config->hca_attr.cross_channel ? "" : "not ");
1391 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1392 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1393 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1394 			config->hca_attr.non_wire_sq ? "" : "not ");
1395 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1396 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1397 			config->hca_attr.log_max_static_sq_wq);
1398 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1399 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1400 		if (!config->devx) {
1401 			DRV_LOG(ERR, "DevX is required for packet pacing");
1402 			err = ENODEV;
1403 			goto error;
1404 		}
1405 		if (!config->hca_attr.qos.packet_pacing) {
1406 			DRV_LOG(ERR, "Packet pacing is not supported");
1407 			err = ENODEV;
1408 			goto error;
1409 		}
1410 		if (!config->hca_attr.cross_channel) {
1411 			DRV_LOG(ERR, "Cross channel operations are"
1412 				     " required for packet pacing");
1413 			err = ENODEV;
1414 			goto error;
1415 		}
1416 		if (!config->hca_attr.wqe_index_ignore) {
1417 			DRV_LOG(ERR, "WQE index ignore feature is"
1418 				     " required for packet pacing");
1419 			err = ENODEV;
1420 			goto error;
1421 		}
1422 		if (!config->hca_attr.non_wire_sq) {
1423 			DRV_LOG(ERR, "Non-wire SQ feature is"
1424 				     " required for packet pacing");
1425 			err = ENODEV;
1426 			goto error;
1427 		}
1428 		if (!config->hca_attr.log_max_static_sq_wq) {
1429 			DRV_LOG(ERR, "Static WQE SQ feature is"
1430 				     " required for packet pacing");
1431 			err = ENODEV;
1432 			goto error;
1433 		}
1434 		if (!config->hca_attr.qos.wqe_rate_pp) {
1435 			DRV_LOG(ERR, "WQE rate mode is required"
1436 				     " for packet pacing");
1437 			err = ENODEV;
1438 			goto error;
1439 		}
1440 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1441 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1442 			     " can't create queues for packet pacing");
1443 		err = ENODEV;
1444 		goto error;
1445 #endif
1446 	}
1447 	if (config->devx) {
1448 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1449 
1450 		err = config->hca_attr.access_register_user ?
1451 			mlx5_devx_cmd_register_read
1452 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1453 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1454 		if (!err) {
1455 			uint32_t ts_mode;
1456 
1457 			/* MTUTC register is read successfully. */
1458 			ts_mode = MLX5_GET(register_mtutc, reg,
1459 					   time_stamp_mode);
1460 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1461 				config->rt_timestamp = 1;
1462 		} else {
1463 			/* Kernel does not support register reading. */
1464 			if (config->hca_attr.dev_freq_khz ==
1465 						 (NS_PER_S / MS_PER_S))
1466 				config->rt_timestamp = 1;
1467 		}
1468 	}
1469 	/*
1470 	 * If HW has bug working with tunnel packet decapsulation and
1471 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1472 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1473 	 */
1474 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1475 		config->hw_fcs_strip = 0;
1476 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1477 		(config->hw_fcs_strip ? "" : "not "));
1478 	if (config->mprq.enabled && mprq) {
1479 		if (config->mprq.stride_num_n &&
1480 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1481 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1482 			config->mprq.stride_num_n =
1483 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1484 						mprq_min_stride_num_n),
1485 					mprq_max_stride_num_n);
1486 			DRV_LOG(WARNING,
1487 				"the number of strides"
1488 				" for Multi-Packet RQ is out of range,"
1489 				" setting default value (%u)",
1490 				1 << config->mprq.stride_num_n);
1491 		}
1492 		if (config->mprq.stride_size_n &&
1493 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1494 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1495 			config->mprq.stride_size_n =
1496 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1497 						mprq_min_stride_size_n),
1498 					mprq_max_stride_size_n);
1499 			DRV_LOG(WARNING,
1500 				"the size of a stride"
1501 				" for Multi-Packet RQ is out of range,"
1502 				" setting default value (%u)",
1503 				1 << config->mprq.stride_size_n);
1504 		}
1505 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1506 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1507 	} else if (config->mprq.enabled && !mprq) {
1508 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1509 		config->mprq.enabled = 0;
1510 	}
1511 	if (config->max_dump_files_num == 0)
1512 		config->max_dump_files_num = 128;
1513 	eth_dev = rte_eth_dev_allocate(name);
1514 	if (eth_dev == NULL) {
1515 		DRV_LOG(ERR, "can not allocate rte ethdev");
1516 		err = ENOMEM;
1517 		goto error;
1518 	}
1519 	if (priv->representor) {
1520 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1521 		eth_dev->data->representor_id = priv->representor_id;
1522 	}
1523 	priv->mp_id.port_id = eth_dev->data->port_id;
1524 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1525 	/*
1526 	 * Store associated network device interface index. This index
1527 	 * is permanent throughout the lifetime of device. So, we may store
1528 	 * the ifindex here and use the cached value further.
1529 	 */
1530 	MLX5_ASSERT(spawn->ifindex);
1531 	priv->if_index = spawn->ifindex;
1532 	eth_dev->data->dev_private = priv;
1533 	priv->dev_data = eth_dev->data;
1534 	eth_dev->data->mac_addrs = priv->mac;
1535 	eth_dev->device = dpdk_dev;
1536 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1537 	/* Configure the first MAC address by default. */
1538 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1539 		DRV_LOG(ERR,
1540 			"port %u cannot get MAC address, is mlx5_en"
1541 			" loaded? (errno: %s)",
1542 			eth_dev->data->port_id, strerror(rte_errno));
1543 		err = ENODEV;
1544 		goto error;
1545 	}
1546 	DRV_LOG(INFO,
1547 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1548 		eth_dev->data->port_id,
1549 		mac.addr_bytes[0], mac.addr_bytes[1],
1550 		mac.addr_bytes[2], mac.addr_bytes[3],
1551 		mac.addr_bytes[4], mac.addr_bytes[5]);
1552 #ifdef RTE_LIBRTE_MLX5_DEBUG
1553 	{
1554 		char ifname[MLX5_NAMESIZE];
1555 
1556 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1557 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1558 				eth_dev->data->port_id, ifname);
1559 		else
1560 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1561 				eth_dev->data->port_id);
1562 	}
1563 #endif
1564 	/* Get actual MTU if possible. */
1565 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1566 	if (err) {
1567 		err = rte_errno;
1568 		goto error;
1569 	}
1570 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1571 		priv->mtu);
1572 	/* Initialize burst functions to prevent crashes before link-up. */
1573 	eth_dev->rx_pkt_burst = removed_rx_burst;
1574 	eth_dev->tx_pkt_burst = removed_tx_burst;
1575 	eth_dev->dev_ops = &mlx5_dev_ops;
1576 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1577 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1578 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1579 	/* Register MAC address. */
1580 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1581 	if (config->vf && config->vf_nl_en)
1582 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1583 				      mlx5_ifindex(eth_dev),
1584 				      eth_dev->data->mac_addrs,
1585 				      MLX5_MAX_MAC_ADDRESSES);
1586 	priv->flows = 0;
1587 	priv->ctrl_flows = 0;
1588 	rte_spinlock_init(&priv->flow_list_lock);
1589 	TAILQ_INIT(&priv->flow_meters);
1590 	priv->mtr_profile_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_PTR);
1591 	if (!priv->mtr_profile_tbl)
1592 		goto error;
1593 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1594 	mlx5_glue->dv_set_context_attr(sh->ctx,
1595 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1596 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1597 				.alloc = &mlx5_alloc_verbs_buf,
1598 				.free = &mlx5_free_verbs_buf,
1599 				.data = sh,
1600 			}));
1601 	/* Bring Ethernet device up. */
1602 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1603 		eth_dev->data->port_id);
1604 	mlx5_set_link_up(eth_dev);
1605 	/*
1606 	 * Even though the interrupt handler is not installed yet,
1607 	 * interrupts will still trigger on the async_fd from
1608 	 * Verbs context returned by ibv_open_device().
1609 	 */
1610 	mlx5_link_update(eth_dev, 0);
1611 #ifdef HAVE_MLX5DV_DR_ESWITCH
1612 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1613 	      (switch_info->representor || switch_info->master)))
1614 		config->dv_esw_en = 0;
1615 #else
1616 	config->dv_esw_en = 0;
1617 #endif
1618 	/* Detect minimal data bytes to inline. */
1619 	mlx5_set_min_inline(spawn, config);
1620 	/* Store device configuration on private structure. */
1621 	priv->config = *config;
1622 	/* Create context for virtual machine VLAN workaround. */
1623 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1624 	if (config->dv_flow_en) {
1625 		err = mlx5_alloc_shared_dr(priv);
1626 		if (err)
1627 			goto error;
1628 	}
1629 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1630 		priv->obj_ops = devx_obj_ops;
1631 		priv->obj_ops.drop_action_create =
1632 						ibv_obj_ops.drop_action_create;
1633 		priv->obj_ops.drop_action_destroy =
1634 						ibv_obj_ops.drop_action_destroy;
1635 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1636 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1637 #else
1638 		if (config->dv_esw_en)
1639 			priv->obj_ops.txq_obj_modify =
1640 						ibv_obj_ops.txq_obj_modify;
1641 #endif
1642 		/* Use specific wrappers for Tx object. */
1643 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1644 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1645 		mlx5_queue_counter_id_prepare(eth_dev);
1646 		priv->obj_ops.lb_dummy_queue_create =
1647 					mlx5_rxq_ibv_obj_dummy_lb_create;
1648 		priv->obj_ops.lb_dummy_queue_release =
1649 					mlx5_rxq_ibv_obj_dummy_lb_release;
1650 	} else {
1651 		priv->obj_ops = ibv_obj_ops;
1652 	}
1653 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1654 	if (!priv->drop_queue.hrxq)
1655 		goto error;
1656 	/* Supported Verbs flow priority number detection. */
1657 	err = mlx5_flow_discover_priorities(eth_dev);
1658 	if (err < 0) {
1659 		err = -err;
1660 		goto error;
1661 	}
1662 	priv->config.flow_prio = err;
1663 	if (!priv->config.dv_esw_en &&
1664 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1665 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1666 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1667 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1668 	}
1669 	mlx5_set_metadata_mask(eth_dev);
1670 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1671 	    !priv->sh->dv_regc0_mask) {
1672 		DRV_LOG(ERR, "metadata mode %u is not supported "
1673 			     "(no metadata reg_c[0] is available)",
1674 			     priv->config.dv_xmeta_en);
1675 			err = ENOTSUP;
1676 			goto error;
1677 	}
1678 	mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1679 			     mlx5_hrxq_create_cb,
1680 			     mlx5_hrxq_match_cb,
1681 			     mlx5_hrxq_remove_cb);
1682 	/* Query availability of metadata reg_c's. */
1683 	err = mlx5_flow_discover_mreg_c(eth_dev);
1684 	if (err < 0) {
1685 		err = -err;
1686 		goto error;
1687 	}
1688 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1689 		DRV_LOG(DEBUG,
1690 			"port %u extensive metadata register is not supported",
1691 			eth_dev->data->port_id);
1692 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1693 			DRV_LOG(ERR, "metadata mode %u is not supported "
1694 				     "(no metadata registers available)",
1695 				     priv->config.dv_xmeta_en);
1696 			err = ENOTSUP;
1697 			goto error;
1698 		}
1699 	}
1700 	if (priv->config.dv_flow_en &&
1701 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1702 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1703 	    priv->sh->dv_regc0_mask) {
1704 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1705 						      MLX5_FLOW_MREG_HTABLE_SZ,
1706 						      0, 0,
1707 						      flow_dv_mreg_create_cb,
1708 						      flow_dv_mreg_match_cb,
1709 						      flow_dv_mreg_remove_cb);
1710 		if (!priv->mreg_cp_tbl) {
1711 			err = ENOMEM;
1712 			goto error;
1713 		}
1714 		priv->mreg_cp_tbl->ctx = eth_dev;
1715 	}
1716 	rte_spinlock_init(&priv->shared_act_sl);
1717 	mlx5_flow_counter_mode_config(eth_dev);
1718 	if (priv->config.dv_flow_en)
1719 		eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1720 	return eth_dev;
1721 error:
1722 	if (priv) {
1723 		if (priv->mreg_cp_tbl)
1724 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1725 		if (priv->sh)
1726 			mlx5_os_free_shared_dr(priv);
1727 		if (priv->nl_socket_route >= 0)
1728 			close(priv->nl_socket_route);
1729 		if (priv->nl_socket_rdma >= 0)
1730 			close(priv->nl_socket_rdma);
1731 		if (priv->vmwa_context)
1732 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1733 		if (eth_dev && priv->drop_queue.hrxq)
1734 			mlx5_drop_action_destroy(eth_dev);
1735 		if (priv->mtr_profile_tbl)
1736 			mlx5_l3t_destroy(priv->mtr_profile_tbl);
1737 		if (own_domain_id)
1738 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1739 		mlx5_cache_list_destroy(&priv->hrxqs);
1740 		mlx5_free(priv);
1741 		if (eth_dev != NULL)
1742 			eth_dev->data->dev_private = NULL;
1743 	}
1744 	if (eth_dev != NULL) {
1745 		/* mac_addrs must not be freed alone because part of
1746 		 * dev_private
1747 		 **/
1748 		eth_dev->data->mac_addrs = NULL;
1749 		rte_eth_dev_release_port(eth_dev);
1750 	}
1751 	if (sh)
1752 		mlx5_free_shared_dev_ctx(sh);
1753 	MLX5_ASSERT(err > 0);
1754 	rte_errno = err;
1755 	return NULL;
1756 }
1757 
1758 /**
1759  * Comparison callback to sort device data.
1760  *
1761  * This is meant to be used with qsort().
1762  *
1763  * @param a[in]
1764  *   Pointer to pointer to first data object.
1765  * @param b[in]
1766  *   Pointer to pointer to second data object.
1767  *
1768  * @return
1769  *   0 if both objects are equal, less than 0 if the first argument is less
1770  *   than the second, greater than 0 otherwise.
1771  */
1772 static int
1773 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1774 {
1775 	const struct mlx5_switch_info *si_a =
1776 		&((const struct mlx5_dev_spawn_data *)a)->info;
1777 	const struct mlx5_switch_info *si_b =
1778 		&((const struct mlx5_dev_spawn_data *)b)->info;
1779 	int ret;
1780 
1781 	/* Master device first. */
1782 	ret = si_b->master - si_a->master;
1783 	if (ret)
1784 		return ret;
1785 	/* Then representor devices. */
1786 	ret = si_b->representor - si_a->representor;
1787 	if (ret)
1788 		return ret;
1789 	/* Unidentified devices come last in no specific order. */
1790 	if (!si_a->representor)
1791 		return 0;
1792 	/* Order representors by name. */
1793 	return si_a->port_name - si_b->port_name;
1794 }
1795 
1796 /**
1797  * Match PCI information for possible slaves of bonding device.
1798  *
1799  * @param[in] ibv_dev
1800  *   Pointer to Infiniband device structure.
1801  * @param[in] pci_dev
1802  *   Pointer to primary PCI address structure to match.
1803  * @param[in] nl_rdma
1804  *   Netlink RDMA group socket handle.
1805  * @param[in] owner
1806  *   Rerepsentor owner PF index.
1807  * @param[out] bond_info
1808  *   Pointer to bonding information.
1809  *
1810  * @return
1811  *   negative value if no bonding device found, otherwise
1812  *   positive index of slave PF in bonding.
1813  */
1814 static int
1815 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1816 			   const struct rte_pci_addr *pci_dev,
1817 			   int nl_rdma, uint16_t owner,
1818 			   struct mlx5_bond_info *bond_info)
1819 {
1820 	char ifname[IF_NAMESIZE + 1];
1821 	unsigned int ifindex;
1822 	unsigned int np, i;
1823 	FILE *bond_file = NULL, *file;
1824 	int pf = -1;
1825 	int ret;
1826 
1827 	/*
1828 	 * Try to get master device name. If something goes
1829 	 * wrong suppose the lack of kernel support and no
1830 	 * bonding devices.
1831 	 */
1832 	memset(bond_info, 0, sizeof(*bond_info));
1833 	if (nl_rdma < 0)
1834 		return -1;
1835 	if (!strstr(ibv_dev->name, "bond"))
1836 		return -1;
1837 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1838 	if (!np)
1839 		return -1;
1840 	/*
1841 	 * The Master device might not be on the predefined
1842 	 * port (not on port index 1, it is not garanted),
1843 	 * we have to scan all Infiniband device port and
1844 	 * find master.
1845 	 */
1846 	for (i = 1; i <= np; ++i) {
1847 		/* Check whether Infiniband port is populated. */
1848 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1849 		if (!ifindex)
1850 			continue;
1851 		if (!if_indextoname(ifindex, ifname))
1852 			continue;
1853 		/* Try to read bonding slave names from sysfs. */
1854 		MKSTR(slaves,
1855 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1856 		bond_file = fopen(slaves, "r");
1857 		if (bond_file)
1858 			break;
1859 	}
1860 	if (!bond_file)
1861 		return -1;
1862 	/* Use safe format to check maximal buffer length. */
1863 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1864 	while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1865 		char tmp_str[IF_NAMESIZE + 32];
1866 		struct rte_pci_addr pci_addr;
1867 		struct mlx5_switch_info	info;
1868 
1869 		/* Process slave interface names in the loop. */
1870 		snprintf(tmp_str, sizeof(tmp_str),
1871 			 "/sys/class/net/%s", ifname);
1872 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1873 			DRV_LOG(WARNING, "can not get PCI address"
1874 					 " for netdev \"%s\"", ifname);
1875 			continue;
1876 		}
1877 		/* Slave interface PCI address match found. */
1878 		snprintf(tmp_str, sizeof(tmp_str),
1879 			 "/sys/class/net/%s/phys_port_name", ifname);
1880 		file = fopen(tmp_str, "rb");
1881 		if (!file)
1882 			break;
1883 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1884 		if (fscanf(file, "%32s", tmp_str) == 1)
1885 			mlx5_translate_port_name(tmp_str, &info);
1886 		fclose(file);
1887 		/* Only process PF ports. */
1888 		if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1889 		    info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1890 			continue;
1891 		/* Check max bonding member. */
1892 		if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1893 			DRV_LOG(WARNING, "bonding index out of range, "
1894 				"please increase MLX5_BOND_MAX_PORTS: %s",
1895 				tmp_str);
1896 			break;
1897 		}
1898 		/* Match PCI address, allows BDF0+pfx or BDFx+pfx. */
1899 		if (pci_dev->domain == pci_addr.domain &&
1900 		    pci_dev->bus == pci_addr.bus &&
1901 		    pci_dev->devid == pci_addr.devid &&
1902 		    ((pci_dev->function == 0 &&
1903 		      pci_dev->function + owner == pci_addr.function) ||
1904 		     (pci_dev->function == owner &&
1905 		      pci_addr.function == owner)))
1906 			pf = info.port_name;
1907 		/* Get ifindex. */
1908 		snprintf(tmp_str, sizeof(tmp_str),
1909 			 "/sys/class/net/%s/ifindex", ifname);
1910 		file = fopen(tmp_str, "rb");
1911 		if (!file)
1912 			break;
1913 		ret = fscanf(file, "%u", &ifindex);
1914 		fclose(file);
1915 		if (ret != 1)
1916 			break;
1917 		/* Save bonding info. */
1918 		strncpy(bond_info->ports[info.port_name].ifname, ifname,
1919 			sizeof(bond_info->ports[0].ifname));
1920 		bond_info->ports[info.port_name].pci_addr = pci_addr;
1921 		bond_info->ports[info.port_name].ifindex = ifindex;
1922 		bond_info->n_port++;
1923 	}
1924 	if (pf >= 0) {
1925 		/* Get bond interface info */
1926 		ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1927 					   bond_info->ifname);
1928 		if (ret)
1929 			DRV_LOG(ERR, "unable to get bond info: %s",
1930 				strerror(rte_errno));
1931 		else
1932 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1933 				ifindex, bond_info->ifindex, bond_info->ifname);
1934 	}
1935 	return pf;
1936 }
1937 
1938 /**
1939  * Register a PCI device within bonding.
1940  *
1941  * This function spawns Ethernet devices out of a given PCI device and
1942  * bonding owner PF index.
1943  *
1944  * @param[in] pci_dev
1945  *   PCI device information.
1946  * @param[in] req_eth_da
1947  *   Requested ethdev device argument.
1948  * @param[in] owner_id
1949  *   Requested owner PF port ID within bonding device, default to 0.
1950  *
1951  * @return
1952  *   0 on success, a negative errno value otherwise and rte_errno is set.
1953  */
1954 static int
1955 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1956 		     struct rte_eth_devargs *req_eth_da,
1957 		     uint16_t owner_id)
1958 {
1959 	struct ibv_device **ibv_list;
1960 	/*
1961 	 * Number of found IB Devices matching with requested PCI BDF.
1962 	 * nd != 1 means there are multiple IB devices over the same
1963 	 * PCI device and we have representors and master.
1964 	 */
1965 	unsigned int nd = 0;
1966 	/*
1967 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1968 	 * we have the single multiport IB device, and there may be
1969 	 * representors attached to some of found ports.
1970 	 */
1971 	unsigned int np = 0;
1972 	/*
1973 	 * Number of DPDK ethernet devices to Spawn - either over
1974 	 * multiple IB devices or multiple ports of single IB device.
1975 	 * Actually this is the number of iterations to spawn.
1976 	 */
1977 	unsigned int ns = 0;
1978 	/*
1979 	 * Bonding device
1980 	 *   < 0 - no bonding device (single one)
1981 	 *  >= 0 - bonding device (value is slave PF index)
1982 	 */
1983 	int bd = -1;
1984 	struct mlx5_dev_spawn_data *list = NULL;
1985 	struct mlx5_dev_config dev_config;
1986 	unsigned int dev_config_vf;
1987 	struct rte_eth_devargs eth_da = *req_eth_da;
1988 	struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1989 	struct mlx5_bond_info bond_info;
1990 	int ret = -1;
1991 
1992 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1993 		mlx5_pmd_socket_init();
1994 	ret = mlx5_init_once();
1995 	if (ret) {
1996 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1997 			strerror(rte_errno));
1998 		return -rte_errno;
1999 	}
2000 	errno = 0;
2001 	ibv_list = mlx5_glue->get_device_list(&ret);
2002 	if (!ibv_list) {
2003 		rte_errno = errno ? errno : ENOSYS;
2004 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2005 		return -rte_errno;
2006 	}
2007 	/*
2008 	 * First scan the list of all Infiniband devices to find
2009 	 * matching ones, gathering into the list.
2010 	 */
2011 	struct ibv_device *ibv_match[ret + 1];
2012 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2013 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2014 	unsigned int i;
2015 
2016 	while (ret-- > 0) {
2017 		struct rte_pci_addr pci_addr;
2018 
2019 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2020 		bd = mlx5_device_bond_pci_match
2021 				(ibv_list[ret], &owner_pci, nl_rdma, owner_id,
2022 				 &bond_info);
2023 		if (bd >= 0) {
2024 			/*
2025 			 * Bonding device detected. Only one match is allowed,
2026 			 * the bonding is supported over multi-port IB device,
2027 			 * there should be no matches on representor PCI
2028 			 * functions or non VF LAG bonding devices with
2029 			 * specified address.
2030 			 */
2031 			if (nd) {
2032 				DRV_LOG(ERR,
2033 					"multiple PCI match on bonding device"
2034 					"\"%s\" found", ibv_list[ret]->name);
2035 				rte_errno = ENOENT;
2036 				ret = -rte_errno;
2037 				goto exit;
2038 			}
2039 			/* Amend owner pci address if owner PF ID specified. */
2040 			if (eth_da.nb_representor_ports)
2041 				owner_pci.function += owner_id;
2042 			DRV_LOG(INFO, "PCI information matches for"
2043 				      " slave %d bonding device \"%s\"",
2044 				      bd, ibv_list[ret]->name);
2045 			ibv_match[nd++] = ibv_list[ret];
2046 			break;
2047 		} else {
2048 			/* Bonding device not found. */
2049 			if (mlx5_dev_to_pci_addr
2050 				(ibv_list[ret]->ibdev_path, &pci_addr))
2051 				continue;
2052 			if (owner_pci.domain != pci_addr.domain ||
2053 			    owner_pci.bus != pci_addr.bus ||
2054 			    owner_pci.devid != pci_addr.devid ||
2055 			    owner_pci.function != pci_addr.function)
2056 				continue;
2057 			DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2058 				ibv_list[ret]->name);
2059 			ibv_match[nd++] = ibv_list[ret];
2060 		}
2061 	}
2062 	ibv_match[nd] = NULL;
2063 	if (!nd) {
2064 		/* No device matches, just complain and bail out. */
2065 		DRV_LOG(WARNING,
2066 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
2067 			" are kernel drivers loaded?",
2068 			owner_pci.domain, owner_pci.bus,
2069 			owner_pci.devid, owner_pci.function);
2070 		rte_errno = ENOENT;
2071 		ret = -rte_errno;
2072 		goto exit;
2073 	}
2074 	if (nd == 1) {
2075 		/*
2076 		 * Found single matching device may have multiple ports.
2077 		 * Each port may be representor, we have to check the port
2078 		 * number and check the representors existence.
2079 		 */
2080 		if (nl_rdma >= 0)
2081 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2082 		if (!np)
2083 			DRV_LOG(WARNING, "can not get IB device \"%s\""
2084 					 " ports number", ibv_match[0]->name);
2085 		if (bd >= 0 && !np) {
2086 			DRV_LOG(ERR, "can not get ports"
2087 				     " for bonding device");
2088 			rte_errno = ENOENT;
2089 			ret = -rte_errno;
2090 			goto exit;
2091 		}
2092 	}
2093 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2094 	if (bd >= 0) {
2095 		/*
2096 		 * This may happen if there is VF LAG kernel support and
2097 		 * application is compiled with older rdma_core library.
2098 		 */
2099 		DRV_LOG(ERR,
2100 			"No kernel/verbs support for VF LAG bonding found.");
2101 		rte_errno = ENOTSUP;
2102 		ret = -rte_errno;
2103 		goto exit;
2104 	}
2105 #endif
2106 	/*
2107 	 * Now we can determine the maximal
2108 	 * amount of devices to be spawned.
2109 	 */
2110 	list = mlx5_malloc(MLX5_MEM_ZERO,
2111 			   sizeof(struct mlx5_dev_spawn_data) *
2112 			   (np ? np : nd),
2113 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2114 	if (!list) {
2115 		DRV_LOG(ERR, "spawn data array allocation failure");
2116 		rte_errno = ENOMEM;
2117 		ret = -rte_errno;
2118 		goto exit;
2119 	}
2120 	if (bd >= 0 || np > 1) {
2121 		/*
2122 		 * Single IB device with multiple ports found,
2123 		 * it may be E-Switch master device and representors.
2124 		 * We have to perform identification through the ports.
2125 		 */
2126 		MLX5_ASSERT(nl_rdma >= 0);
2127 		MLX5_ASSERT(ns == 0);
2128 		MLX5_ASSERT(nd == 1);
2129 		MLX5_ASSERT(np);
2130 		for (i = 1; i <= np; ++i) {
2131 			list[ns].bond_info = &bond_info;
2132 			list[ns].max_port = np;
2133 			list[ns].phys_port = i;
2134 			list[ns].phys_dev = ibv_match[0];
2135 			list[ns].eth_dev = NULL;
2136 			list[ns].pci_dev = pci_dev;
2137 			list[ns].pf_bond = bd;
2138 			list[ns].ifindex = mlx5_nl_ifindex
2139 				(nl_rdma,
2140 				mlx5_os_get_dev_device_name
2141 						(list[ns].phys_dev), i);
2142 			if (!list[ns].ifindex) {
2143 				/*
2144 				 * No network interface index found for the
2145 				 * specified port, it means there is no
2146 				 * representor on this port. It's OK,
2147 				 * there can be disabled ports, for example
2148 				 * if sriov_numvfs < sriov_totalvfs.
2149 				 */
2150 				continue;
2151 			}
2152 			ret = -1;
2153 			if (nl_route >= 0)
2154 				ret = mlx5_nl_switch_info
2155 					       (nl_route,
2156 						list[ns].ifindex,
2157 						&list[ns].info);
2158 			if (ret || (!list[ns].info.representor &&
2159 				    !list[ns].info.master)) {
2160 				/*
2161 				 * We failed to recognize representors with
2162 				 * Netlink, let's try to perform the task
2163 				 * with sysfs.
2164 				 */
2165 				ret =  mlx5_sysfs_switch_info
2166 						(list[ns].ifindex,
2167 						 &list[ns].info);
2168 			}
2169 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2170 			if (!ret && bd >= 0) {
2171 				switch (list[ns].info.name_type) {
2172 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2173 					if (list[ns].info.port_name == bd)
2174 						ns++;
2175 					break;
2176 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2177 					/* Fallthrough */
2178 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2179 					/* Fallthrough */
2180 				case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2181 					if (list[ns].info.pf_num == bd)
2182 						ns++;
2183 					break;
2184 				default:
2185 					break;
2186 				}
2187 				continue;
2188 			}
2189 #endif
2190 			if (!ret && (list[ns].info.representor ^
2191 				     list[ns].info.master))
2192 				ns++;
2193 		}
2194 		if (!ns) {
2195 			DRV_LOG(ERR,
2196 				"unable to recognize master/representors"
2197 				" on the IB device with multiple ports");
2198 			rte_errno = ENOENT;
2199 			ret = -rte_errno;
2200 			goto exit;
2201 		}
2202 	} else {
2203 		/*
2204 		 * The existence of several matching entries (nd > 1) means
2205 		 * port representors have been instantiated. No existing Verbs
2206 		 * call nor sysfs entries can tell them apart, this can only
2207 		 * be done through Netlink calls assuming kernel drivers are
2208 		 * recent enough to support them.
2209 		 *
2210 		 * In the event of identification failure through Netlink,
2211 		 * try again through sysfs, then:
2212 		 *
2213 		 * 1. A single IB device matches (nd == 1) with single
2214 		 *    port (np=0/1) and is not a representor, assume
2215 		 *    no switch support.
2216 		 *
2217 		 * 2. Otherwise no safe assumptions can be made;
2218 		 *    complain louder and bail out.
2219 		 */
2220 		for (i = 0; i != nd; ++i) {
2221 			memset(&list[ns].info, 0, sizeof(list[ns].info));
2222 			list[ns].bond_info = NULL;
2223 			list[ns].max_port = 1;
2224 			list[ns].phys_port = 1;
2225 			list[ns].phys_dev = ibv_match[i];
2226 			list[ns].eth_dev = NULL;
2227 			list[ns].pci_dev = pci_dev;
2228 			list[ns].pf_bond = -1;
2229 			list[ns].ifindex = 0;
2230 			if (nl_rdma >= 0)
2231 				list[ns].ifindex = mlx5_nl_ifindex
2232 				(nl_rdma,
2233 				mlx5_os_get_dev_device_name
2234 						(list[ns].phys_dev), 1);
2235 			if (!list[ns].ifindex) {
2236 				char ifname[IF_NAMESIZE];
2237 
2238 				/*
2239 				 * Netlink failed, it may happen with old
2240 				 * ib_core kernel driver (before 4.16).
2241 				 * We can assume there is old driver because
2242 				 * here we are processing single ports IB
2243 				 * devices. Let's try sysfs to retrieve
2244 				 * the ifindex. The method works for
2245 				 * master device only.
2246 				 */
2247 				if (nd > 1) {
2248 					/*
2249 					 * Multiple devices found, assume
2250 					 * representors, can not distinguish
2251 					 * master/representor and retrieve
2252 					 * ifindex via sysfs.
2253 					 */
2254 					continue;
2255 				}
2256 				ret = mlx5_get_ifname_sysfs
2257 					(ibv_match[i]->ibdev_path, ifname);
2258 				if (!ret)
2259 					list[ns].ifindex =
2260 						if_nametoindex(ifname);
2261 				if (!list[ns].ifindex) {
2262 					/*
2263 					 * No network interface index found
2264 					 * for the specified device, it means
2265 					 * there it is neither representor
2266 					 * nor master.
2267 					 */
2268 					continue;
2269 				}
2270 			}
2271 			ret = -1;
2272 			if (nl_route >= 0)
2273 				ret = mlx5_nl_switch_info
2274 					       (nl_route,
2275 						list[ns].ifindex,
2276 						&list[ns].info);
2277 			if (ret || (!list[ns].info.representor &&
2278 				    !list[ns].info.master)) {
2279 				/*
2280 				 * We failed to recognize representors with
2281 				 * Netlink, let's try to perform the task
2282 				 * with sysfs.
2283 				 */
2284 				ret =  mlx5_sysfs_switch_info
2285 						(list[ns].ifindex,
2286 						 &list[ns].info);
2287 			}
2288 			if (!ret && (list[ns].info.representor ^
2289 				     list[ns].info.master)) {
2290 				ns++;
2291 			} else if ((nd == 1) &&
2292 				   !list[ns].info.representor &&
2293 				   !list[ns].info.master) {
2294 				/*
2295 				 * Single IB device with
2296 				 * one physical port and
2297 				 * attached network device.
2298 				 * May be SRIOV is not enabled
2299 				 * or there is no representors.
2300 				 */
2301 				DRV_LOG(INFO, "no E-Switch support detected");
2302 				ns++;
2303 				break;
2304 			}
2305 		}
2306 		if (!ns) {
2307 			DRV_LOG(ERR,
2308 				"unable to recognize master/representors"
2309 				" on the multiple IB devices");
2310 			rte_errno = ENOENT;
2311 			ret = -rte_errno;
2312 			goto exit;
2313 		}
2314 		/*
2315 		 * New kernels may add the switch_id attribute for the case
2316 		 * there is no E-Switch and we wrongly recognized the
2317 		 * only device as master. Override this if there is the
2318 		 * single device with single port and new device name
2319 		 * format present.
2320 		 */
2321 		if (nd == 1 &&
2322 		    list[0].info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK) {
2323 			list[0].info.master = 0;
2324 			list[0].info.representor = 0;
2325 		}
2326 	}
2327 	MLX5_ASSERT(ns);
2328 	/*
2329 	 * Sort list to probe devices in natural order for users convenience
2330 	 * (i.e. master first, then representors from lowest to highest ID).
2331 	 */
2332 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2333 	/* Device specific configuration. */
2334 	switch (pci_dev->id.device_id) {
2335 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2336 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2337 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2338 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2339 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2340 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2341 	case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2342 		dev_config_vf = 1;
2343 		break;
2344 	default:
2345 		dev_config_vf = 0;
2346 		break;
2347 	}
2348 	if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2349 		/* Set devargs default values. */
2350 		if (eth_da.nb_mh_controllers == 0) {
2351 			eth_da.nb_mh_controllers = 1;
2352 			eth_da.mh_controllers[0] = 0;
2353 		}
2354 		if (eth_da.nb_ports == 0 && ns > 0) {
2355 			if (list[0].pf_bond >= 0 && list[0].info.representor)
2356 				DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2357 					pci_dev->device.devargs->args);
2358 			eth_da.nb_ports = 1;
2359 			eth_da.ports[0] = list[0].info.pf_num;
2360 		}
2361 		if (eth_da.nb_representor_ports == 0) {
2362 			eth_da.nb_representor_ports = 1;
2363 			eth_da.representor_ports[0] = 0;
2364 		}
2365 	}
2366 	for (i = 0; i != ns; ++i) {
2367 		uint32_t restore;
2368 
2369 		/* Default configuration. */
2370 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2371 		dev_config.vf = dev_config_vf;
2372 		dev_config.mps = MLX5_ARG_UNSET;
2373 		dev_config.dbnc = MLX5_ARG_UNSET;
2374 		dev_config.rx_vec_en = 1;
2375 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2376 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2377 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2378 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2379 		dev_config.vf_nl_en = 1;
2380 		dev_config.mr_ext_memseg_en = 1;
2381 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2382 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2383 		dev_config.dv_esw_en = 1;
2384 		dev_config.dv_flow_en = 1;
2385 		dev_config.decap_en = 1;
2386 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2387 		dev_config.allow_duplicate_pattern = 1;
2388 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2389 						 &list[i],
2390 						 &dev_config,
2391 						 &eth_da);
2392 		if (!list[i].eth_dev) {
2393 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2394 				break;
2395 			/* Device is disabled or already spawned. Ignore it. */
2396 			continue;
2397 		}
2398 		restore = list[i].eth_dev->data->dev_flags;
2399 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2400 		/* Restore non-PCI flags cleared by the above call. */
2401 		list[i].eth_dev->data->dev_flags |= restore;
2402 		rte_eth_dev_probing_finish(list[i].eth_dev);
2403 	}
2404 	if (i != ns) {
2405 		DRV_LOG(ERR,
2406 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2407 			" encountering an error: %s",
2408 			owner_pci.domain, owner_pci.bus,
2409 			owner_pci.devid, owner_pci.function,
2410 			strerror(rte_errno));
2411 		ret = -rte_errno;
2412 		/* Roll back. */
2413 		while (i--) {
2414 			if (!list[i].eth_dev)
2415 				continue;
2416 			mlx5_dev_close(list[i].eth_dev);
2417 			/* mac_addrs must not be freed because in dev_private */
2418 			list[i].eth_dev->data->mac_addrs = NULL;
2419 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2420 		}
2421 		/* Restore original error. */
2422 		rte_errno = -ret;
2423 	} else {
2424 		ret = 0;
2425 	}
2426 exit:
2427 	/*
2428 	 * Do the routine cleanup:
2429 	 * - close opened Netlink sockets
2430 	 * - free allocated spawn data array
2431 	 * - free the Infiniband device list
2432 	 */
2433 	if (nl_rdma >= 0)
2434 		close(nl_rdma);
2435 	if (nl_route >= 0)
2436 		close(nl_route);
2437 	if (list)
2438 		mlx5_free(list);
2439 	MLX5_ASSERT(ibv_list);
2440 	mlx5_glue->free_device_list(ibv_list);
2441 	return ret;
2442 }
2443 
2444 /**
2445  * DPDK callback to register a PCI device.
2446  *
2447  * This function spawns Ethernet devices out of a given PCI device.
2448  *
2449  * @param[in] pci_drv
2450  *   PCI driver structure (mlx5_driver).
2451  * @param[in] pci_dev
2452  *   PCI device information.
2453  *
2454  * @return
2455  *   0 on success, a negative errno value otherwise and rte_errno is set.
2456  */
2457 int
2458 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2459 		  struct rte_pci_device *pci_dev)
2460 {
2461 	struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2462 	int ret = 0;
2463 	uint16_t p;
2464 
2465 	if (pci_dev->device.devargs) {
2466 		/* Parse representor information from device argument. */
2467 		if (pci_dev->device.devargs->cls_str)
2468 			ret = rte_eth_devargs_parse
2469 				(pci_dev->device.devargs->cls_str, &eth_da);
2470 		if (ret) {
2471 			DRV_LOG(ERR, "failed to parse device arguments: %s",
2472 				pci_dev->device.devargs->cls_str);
2473 			return -rte_errno;
2474 		}
2475 		if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2476 			/* Support legacy device argument */
2477 			ret = rte_eth_devargs_parse
2478 				(pci_dev->device.devargs->args, &eth_da);
2479 			if (ret) {
2480 				DRV_LOG(ERR, "failed to parse device arguments: %s",
2481 					pci_dev->device.devargs->args);
2482 				return -rte_errno;
2483 			}
2484 		}
2485 	}
2486 
2487 	if (eth_da.nb_ports > 0) {
2488 		/* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2489 		for (p = 0; p < eth_da.nb_ports; p++)
2490 			ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2491 						   eth_da.ports[p]);
2492 	} else {
2493 		ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2494 	}
2495 	return ret;
2496 }
2497 
2498 static int
2499 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2500 {
2501 	char *env;
2502 	int value;
2503 
2504 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2505 	/* Get environment variable to store. */
2506 	env = getenv(MLX5_SHUT_UP_BF);
2507 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2508 	if (config->dbnc == MLX5_ARG_UNSET)
2509 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2510 	else
2511 		setenv(MLX5_SHUT_UP_BF,
2512 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2513 	return value;
2514 }
2515 
2516 static void
2517 mlx5_restore_doorbell_mapping_env(int value)
2518 {
2519 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2520 	/* Restore the original environment variable state. */
2521 	if (value == MLX5_ARG_UNSET)
2522 		unsetenv(MLX5_SHUT_UP_BF);
2523 	else
2524 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2525 }
2526 
2527 /**
2528  * Extract pdn of PD object using DV API.
2529  *
2530  * @param[in] pd
2531  *   Pointer to the verbs PD object.
2532  * @param[out] pdn
2533  *   Pointer to the PD object number variable.
2534  *
2535  * @return
2536  *   0 on success, error value otherwise.
2537  */
2538 int
2539 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2540 {
2541 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2542 	struct mlx5dv_obj obj;
2543 	struct mlx5dv_pd pd_info;
2544 	int ret = 0;
2545 
2546 	obj.pd.in = pd;
2547 	obj.pd.out = &pd_info;
2548 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2549 	if (ret) {
2550 		DRV_LOG(DEBUG, "Fail to get PD object info");
2551 		return ret;
2552 	}
2553 	*pdn = pd_info.pdn;
2554 	return 0;
2555 #else
2556 	(void)pd;
2557 	(void)pdn;
2558 	return -ENOTSUP;
2559 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2560 }
2561 
2562 /**
2563  * Function API to open IB device.
2564  *
2565  * This function calls the Linux glue APIs to open a device.
2566  *
2567  * @param[in] spawn
2568  *   Pointer to the IB device attributes (name, port, etc).
2569  * @param[out] config
2570  *   Pointer to device configuration structure.
2571  * @param[out] sh
2572  *   Pointer to shared context structure.
2573  *
2574  * @return
2575  *   0 on success, a positive error value otherwise.
2576  */
2577 int
2578 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2579 		     const struct mlx5_dev_config *config,
2580 		     struct mlx5_dev_ctx_shared *sh)
2581 {
2582 	int dbmap_env;
2583 	int err = 0;
2584 
2585 	sh->numa_node = spawn->pci_dev->device.numa_node;
2586 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2587 	/*
2588 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2589 	 * before the device creation. The rdma_core library
2590 	 * checks the variable at device creation and
2591 	 * stores the result internally.
2592 	 */
2593 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2594 	/* Try to open IB device with DV first, then usual Verbs. */
2595 	errno = 0;
2596 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2597 	if (sh->ctx) {
2598 		sh->devx = 1;
2599 		DRV_LOG(DEBUG, "DevX is supported");
2600 		/* The device is created, no need for environment. */
2601 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2602 	} else {
2603 		/* The environment variable is still configured. */
2604 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2605 		err = errno ? errno : ENODEV;
2606 		/*
2607 		 * The environment variable is not needed anymore,
2608 		 * all device creation attempts are completed.
2609 		 */
2610 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2611 		if (!sh->ctx)
2612 			return err;
2613 		DRV_LOG(DEBUG, "DevX is NOT supported");
2614 		err = 0;
2615 	}
2616 	if (!err && sh->ctx) {
2617 		/* Hint libmlx5 to use PMD allocator for data plane resources */
2618 		mlx5_glue->dv_set_context_attr(sh->ctx,
2619 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2620 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2621 				.alloc = &mlx5_alloc_verbs_buf,
2622 				.free = &mlx5_free_verbs_buf,
2623 				.data = sh,
2624 			}));
2625 	}
2626 	return err;
2627 }
2628 
2629 /**
2630  * Install shared asynchronous device events handler.
2631  * This function is implemented to support event sharing
2632  * between multiple ports of single IB device.
2633  *
2634  * @param sh
2635  *   Pointer to mlx5_dev_ctx_shared object.
2636  */
2637 void
2638 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2639 {
2640 	int ret;
2641 	int flags;
2642 
2643 	sh->intr_handle.fd = -1;
2644 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2645 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2646 		    F_SETFL, flags | O_NONBLOCK);
2647 	if (ret) {
2648 		DRV_LOG(INFO, "failed to change file descriptor async event"
2649 			" queue");
2650 	} else {
2651 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2652 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2653 		if (rte_intr_callback_register(&sh->intr_handle,
2654 					mlx5_dev_interrupt_handler, sh)) {
2655 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2656 			sh->intr_handle.fd = -1;
2657 		}
2658 	}
2659 	if (sh->devx) {
2660 #ifdef HAVE_IBV_DEVX_ASYNC
2661 		sh->intr_handle_devx.fd = -1;
2662 		sh->devx_comp =
2663 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2664 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2665 		if (!devx_comp) {
2666 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2667 			return;
2668 		}
2669 		flags = fcntl(devx_comp->fd, F_GETFL);
2670 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2671 		if (ret) {
2672 			DRV_LOG(INFO, "failed to change file descriptor"
2673 				" devx comp");
2674 			return;
2675 		}
2676 		sh->intr_handle_devx.fd = devx_comp->fd;
2677 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2678 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2679 					mlx5_dev_interrupt_handler_devx, sh)) {
2680 			DRV_LOG(INFO, "Fail to install the devx shared"
2681 				" interrupt.");
2682 			sh->intr_handle_devx.fd = -1;
2683 		}
2684 #endif /* HAVE_IBV_DEVX_ASYNC */
2685 	}
2686 }
2687 
2688 /**
2689  * Uninstall shared asynchronous device events handler.
2690  * This function is implemented to support event sharing
2691  * between multiple ports of single IB device.
2692  *
2693  * @param dev
2694  *   Pointer to mlx5_dev_ctx_shared object.
2695  */
2696 void
2697 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2698 {
2699 	if (sh->intr_handle.fd >= 0)
2700 		mlx5_intr_callback_unregister(&sh->intr_handle,
2701 					      mlx5_dev_interrupt_handler, sh);
2702 #ifdef HAVE_IBV_DEVX_ASYNC
2703 	if (sh->intr_handle_devx.fd >= 0)
2704 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2705 				  mlx5_dev_interrupt_handler_devx, sh);
2706 	if (sh->devx_comp)
2707 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2708 #endif
2709 }
2710 
2711 /**
2712  * Read statistics by a named counter.
2713  *
2714  * @param[in] priv
2715  *   Pointer to the private device data structure.
2716  * @param[in] ctr_name
2717  *   Pointer to the name of the statistic counter to read
2718  * @param[out] stat
2719  *   Pointer to read statistic value.
2720  * @return
2721  *   0 on success and stat is valud, 1 if failed to read the value
2722  *   rte_errno is set.
2723  *
2724  */
2725 int
2726 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2727 		      uint64_t *stat)
2728 {
2729 	int fd;
2730 
2731 	if (priv->sh) {
2732 		if (priv->q_counters != NULL &&
2733 		    strcmp(ctr_name, "out_of_buffer") == 0)
2734 			return mlx5_devx_cmd_queue_counter_query
2735 					(priv->q_counters, 0, (uint32_t *)stat);
2736 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2737 		      priv->sh->ibdev_path,
2738 		      priv->dev_port,
2739 		      ctr_name);
2740 		fd = open(path, O_RDONLY);
2741 		/*
2742 		 * in switchdev the file location is not per port
2743 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2744 		 */
2745 		if (fd == -1) {
2746 			MKSTR(path1, "%s/hw_counters/%s",
2747 			      priv->sh->ibdev_path,
2748 			      ctr_name);
2749 			fd = open(path1, O_RDONLY);
2750 		}
2751 		if (fd != -1) {
2752 			char buf[21] = {'\0'};
2753 			ssize_t n = read(fd, buf, sizeof(buf));
2754 
2755 			close(fd);
2756 			if (n != -1) {
2757 				*stat = strtoull(buf, NULL, 10);
2758 				return 0;
2759 			}
2760 		}
2761 	}
2762 	*stat = 0;
2763 	return 1;
2764 }
2765 
2766 /**
2767  * Set the reg_mr and dereg_mr call backs
2768  *
2769  * @param reg_mr_cb[out]
2770  *   Pointer to reg_mr func
2771  * @param dereg_mr_cb[out]
2772  *   Pointer to dereg_mr func
2773  *
2774  */
2775 void
2776 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2777 		      mlx5_dereg_mr_t *dereg_mr_cb)
2778 {
2779 	*reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2780 	*dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2781 }
2782 
2783 /**
2784  * Remove a MAC address from device
2785  *
2786  * @param dev
2787  *   Pointer to Ethernet device structure.
2788  * @param index
2789  *   MAC address index.
2790  */
2791 void
2792 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2793 {
2794 	struct mlx5_priv *priv = dev->data->dev_private;
2795 	const int vf = priv->config.vf;
2796 
2797 	if (vf)
2798 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2799 					mlx5_ifindex(dev), priv->mac_own,
2800 					&dev->data->mac_addrs[index], index);
2801 }
2802 
2803 /**
2804  * Adds a MAC address to the device
2805  *
2806  * @param dev
2807  *   Pointer to Ethernet device structure.
2808  * @param mac_addr
2809  *   MAC address to register.
2810  * @param index
2811  *   MAC address index.
2812  *
2813  * @return
2814  *   0 on success, a negative errno value otherwise
2815  */
2816 int
2817 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2818 		     uint32_t index)
2819 {
2820 	struct mlx5_priv *priv = dev->data->dev_private;
2821 	const int vf = priv->config.vf;
2822 	int ret = 0;
2823 
2824 	if (vf)
2825 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2826 					   mlx5_ifindex(dev), priv->mac_own,
2827 					   mac, index);
2828 	return ret;
2829 }
2830 
2831 /**
2832  * Modify a VF MAC address
2833  *
2834  * @param priv
2835  *   Pointer to device private data.
2836  * @param mac_addr
2837  *   MAC address to modify into.
2838  * @param iface_idx
2839  *   Net device interface index
2840  * @param vf_index
2841  *   VF index
2842  *
2843  * @return
2844  *   0 on success, a negative errno value otherwise
2845  */
2846 int
2847 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2848 			   unsigned int iface_idx,
2849 			   struct rte_ether_addr *mac_addr,
2850 			   int vf_index)
2851 {
2852 	return mlx5_nl_vf_mac_addr_modify
2853 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2854 }
2855 
2856 /**
2857  * Set device promiscuous mode
2858  *
2859  * @param dev
2860  *   Pointer to Ethernet device structure.
2861  * @param enable
2862  *   0 - promiscuous is disabled, otherwise - enabled
2863  *
2864  * @return
2865  *   0 on success, a negative error value otherwise
2866  */
2867 int
2868 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2869 {
2870 	struct mlx5_priv *priv = dev->data->dev_private;
2871 
2872 	return mlx5_nl_promisc(priv->nl_socket_route,
2873 			       mlx5_ifindex(dev), !!enable);
2874 }
2875 
2876 /**
2877  * Set device promiscuous mode
2878  *
2879  * @param dev
2880  *   Pointer to Ethernet device structure.
2881  * @param enable
2882  *   0 - all multicase is disabled, otherwise - enabled
2883  *
2884  * @return
2885  *   0 on success, a negative error value otherwise
2886  */
2887 int
2888 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2889 {
2890 	struct mlx5_priv *priv = dev->data->dev_private;
2891 
2892 	return mlx5_nl_allmulti(priv->nl_socket_route,
2893 				mlx5_ifindex(dev), !!enable);
2894 }
2895 
2896 /**
2897  * Flush device MAC addresses
2898  *
2899  * @param dev
2900  *   Pointer to Ethernet device structure.
2901  *
2902  */
2903 void
2904 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2905 {
2906 	struct mlx5_priv *priv = dev->data->dev_private;
2907 
2908 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2909 			       dev->data->mac_addrs,
2910 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2911 }
2912