xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 65b3cd0dc39b9e8f94ea8453ed43d38fab7550c0)
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 <rte_ethdev_driver.h>
20 #include <rte_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_autoconf.h"
44 #include "mlx5_mr.h"
45 #include "mlx5_flow.h"
46 #include "rte_pmd_mlx5.h"
47 #include "mlx5_verbs.h"
48 #include "mlx5_nl.h"
49 #include "mlx5_devx.h"
50 
51 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
52 
53 #ifndef HAVE_IBV_MLX5_MOD_MPW
54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
56 #endif
57 
58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
60 #endif
61 
62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
63 
64 /* Spinlock for mlx5_shared_data allocation. */
65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
66 
67 /* Process local data for secondary processes. */
68 static struct mlx5_local_data mlx5_local_data;
69 
70 /**
71  * Set the completion channel file descriptor interrupt as non-blocking.
72  *
73  * @param[in] rxq_obj
74  *   Pointer to RQ channel object, which includes the channel fd
75  *
76  * @param[out] fd
77  *   The file descriptor (representing the intetrrupt) used in this channel.
78  *
79  * @return
80  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
81  */
82 int
83 mlx5_os_set_nonblock_channel_fd(int fd)
84 {
85 	int flags;
86 
87 	flags = fcntl(fd, F_GETFL);
88 	return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
89 }
90 
91 /**
92  * Get mlx5 device attributes. The glue function query_device_ex() is called
93  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
94  * device attributes from the glue out parameter.
95  *
96  * @param dev
97  *   Pointer to ibv context.
98  *
99  * @param device_attr
100  *   Pointer to mlx5 device attributes.
101  *
102  * @return
103  *   0 on success, non zero error number otherwise
104  */
105 int
106 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
107 {
108 	int err;
109 	struct ibv_device_attr_ex attr_ex;
110 	memset(device_attr, 0, sizeof(*device_attr));
111 	err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
112 	if (err)
113 		return err;
114 
115 	device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
116 	device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
117 	device_attr->max_sge = attr_ex.orig_attr.max_sge;
118 	device_attr->max_cq = attr_ex.orig_attr.max_cq;
119 	device_attr->max_qp = attr_ex.orig_attr.max_qp;
120 	device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
121 	device_attr->max_rwq_indirection_table_size =
122 		attr_ex.rss_caps.max_rwq_indirection_table_size;
123 	device_attr->max_tso = attr_ex.tso_caps.max_tso;
124 	device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
125 
126 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
127 	err = mlx5_glue->dv_query_device(ctx, &dv_attr);
128 	if (err)
129 		return err;
130 
131 	device_attr->flags = dv_attr.flags;
132 	device_attr->comp_mask = dv_attr.comp_mask;
133 #ifdef HAVE_IBV_MLX5_MOD_SWP
134 	device_attr->sw_parsing_offloads =
135 		dv_attr.sw_parsing_caps.sw_parsing_offloads;
136 #endif
137 	device_attr->min_single_stride_log_num_of_bytes =
138 		dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
139 	device_attr->max_single_stride_log_num_of_bytes =
140 		dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
141 	device_attr->min_single_wqe_log_num_of_strides =
142 		dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
143 	device_attr->max_single_wqe_log_num_of_strides =
144 		dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
145 	device_attr->stride_supported_qpts =
146 		dv_attr.striding_rq_caps.supported_qpts;
147 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
148 	device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
149 #endif
150 
151 	return err;
152 }
153 
154 /**
155  * Verbs callback to allocate a memory. This function should allocate the space
156  * according to the size provided residing inside a huge page.
157  * Please note that all allocation must respect the alignment from libmlx5
158  * (i.e. currently rte_mem_page_size()).
159  *
160  * @param[in] size
161  *   The size in bytes of the memory to allocate.
162  * @param[in] data
163  *   A pointer to the callback data.
164  *
165  * @return
166  *   Allocated buffer, NULL otherwise and rte_errno is set.
167  */
168 static void *
169 mlx5_alloc_verbs_buf(size_t size, void *data)
170 {
171 	struct mlx5_priv *priv = data;
172 	void *ret;
173 	unsigned int socket = SOCKET_ID_ANY;
174 	size_t alignment = rte_mem_page_size();
175 	if (alignment == (size_t)-1) {
176 		DRV_LOG(ERR, "Failed to get mem page size");
177 		rte_errno = ENOMEM;
178 		return NULL;
179 	}
180 
181 	if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
182 		const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
183 
184 		socket = ctrl->socket;
185 	} else if (priv->verbs_alloc_ctx.type ==
186 		   MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
187 		const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
188 
189 		socket = ctrl->socket;
190 	}
191 	MLX5_ASSERT(data != NULL);
192 	ret = mlx5_malloc(0, size, alignment, socket);
193 	if (!ret && size)
194 		rte_errno = ENOMEM;
195 	return ret;
196 }
197 
198 /**
199  * Verbs callback to free a memory.
200  *
201  * @param[in] ptr
202  *   A pointer to the memory to free.
203  * @param[in] data
204  *   A pointer to the callback data.
205  */
206 static void
207 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
208 {
209 	MLX5_ASSERT(data != NULL);
210 	mlx5_free(ptr);
211 }
212 
213 /**
214  * Initialize DR related data within private structure.
215  * Routine checks the reference counter and does actual
216  * resources creation/initialization only if counter is zero.
217  *
218  * @param[in] priv
219  *   Pointer to the private device data structure.
220  *
221  * @return
222  *   Zero on success, positive error code otherwise.
223  */
224 static int
225 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
226 {
227 	struct mlx5_dev_ctx_shared *sh = priv->sh;
228 	char s[MLX5_HLIST_NAMESIZE];
229 	int err;
230 
231 	MLX5_ASSERT(sh && sh->refcnt);
232 	if (sh->refcnt > 1)
233 		return 0;
234 	err = mlx5_alloc_table_hash_list(priv);
235 	if (err)
236 		return err;
237 	/* Create tags hash list table. */
238 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
239 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
240 	if (!sh->tag_table) {
241 		DRV_LOG(ERR, "tags with hash creation failed.");
242 		err = ENOMEM;
243 		goto error;
244 	}
245 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
246 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ);
247 	if (!sh->modify_cmds) {
248 		DRV_LOG(ERR, "hdr modify hash creation failed");
249 		err = ENOMEM;
250 		goto error;
251 	}
252 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
253 	sh->encaps_decaps = mlx5_hlist_create(s,
254 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ);
255 	if (!sh->encaps_decaps) {
256 		DRV_LOG(ERR, "encap decap hash creation failed");
257 		err = ENOMEM;
258 		goto error;
259 	}
260 #ifdef HAVE_MLX5DV_DR
261 	void *domain;
262 
263 	/* Reference counter is zero, we should initialize structures. */
264 	domain = mlx5_glue->dr_create_domain(sh->ctx,
265 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
266 	if (!domain) {
267 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
268 		err = errno;
269 		goto error;
270 	}
271 	sh->rx_domain = domain;
272 	domain = mlx5_glue->dr_create_domain(sh->ctx,
273 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
274 	if (!domain) {
275 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
276 		err = errno;
277 		goto error;
278 	}
279 	pthread_mutex_init(&sh->dv_mutex, NULL);
280 	sh->tx_domain = domain;
281 #ifdef HAVE_MLX5DV_DR_ESWITCH
282 	if (priv->config.dv_esw_en) {
283 		domain  = mlx5_glue->dr_create_domain
284 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
285 		if (!domain) {
286 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
287 			err = errno;
288 			goto error;
289 		}
290 		sh->fdb_domain = domain;
291 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
292 	}
293 #endif
294 	if (!sh->tunnel_hub)
295 		err = mlx5_alloc_tunnel_hub(sh);
296 	if (err) {
297 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
298 		goto error;
299 	}
300 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
301 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
302 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
303 		if (sh->fdb_domain)
304 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
305 	}
306 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
307 #endif /* HAVE_MLX5DV_DR */
308 	sh->default_miss_action =
309 			mlx5_glue->dr_create_flow_action_default_miss();
310 	if (!sh->default_miss_action)
311 		DRV_LOG(WARNING, "Default miss action is not supported.");
312 	return 0;
313 error:
314 	/* Rollback the created objects. */
315 	if (sh->rx_domain) {
316 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
317 		sh->rx_domain = NULL;
318 	}
319 	if (sh->tx_domain) {
320 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
321 		sh->tx_domain = NULL;
322 	}
323 	if (sh->fdb_domain) {
324 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
325 		sh->fdb_domain = NULL;
326 	}
327 	if (sh->esw_drop_action) {
328 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
329 		sh->esw_drop_action = NULL;
330 	}
331 	if (sh->pop_vlan_action) {
332 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
333 		sh->pop_vlan_action = NULL;
334 	}
335 	if (sh->encaps_decaps) {
336 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
337 		sh->encaps_decaps = NULL;
338 	}
339 	if (sh->modify_cmds) {
340 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
341 		sh->modify_cmds = NULL;
342 	}
343 	if (sh->tag_table) {
344 		/* tags should be destroyed with flow before. */
345 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
346 		sh->tag_table = NULL;
347 	}
348 	if (sh->tunnel_hub) {
349 		mlx5_release_tunnel_hub(sh, priv->dev_port);
350 		sh->tunnel_hub = NULL;
351 	}
352 	mlx5_free_table_hash_list(priv);
353 	return err;
354 }
355 
356 /**
357  * Destroy DR related data within private structure.
358  *
359  * @param[in] priv
360  *   Pointer to the private device data structure.
361  */
362 void
363 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
364 {
365 	struct mlx5_dev_ctx_shared *sh = priv->sh;
366 
367 	MLX5_ASSERT(sh && sh->refcnt);
368 	if (sh->refcnt > 1)
369 		return;
370 #ifdef HAVE_MLX5DV_DR
371 	if (sh->rx_domain) {
372 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
373 		sh->rx_domain = NULL;
374 	}
375 	if (sh->tx_domain) {
376 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
377 		sh->tx_domain = NULL;
378 	}
379 #ifdef HAVE_MLX5DV_DR_ESWITCH
380 	if (sh->fdb_domain) {
381 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
382 		sh->fdb_domain = NULL;
383 	}
384 	if (sh->esw_drop_action) {
385 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
386 		sh->esw_drop_action = NULL;
387 	}
388 #endif
389 	if (sh->pop_vlan_action) {
390 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
391 		sh->pop_vlan_action = NULL;
392 	}
393 	pthread_mutex_destroy(&sh->dv_mutex);
394 #endif /* HAVE_MLX5DV_DR */
395 	if (sh->default_miss_action)
396 		mlx5_glue->destroy_flow_action
397 				(sh->default_miss_action);
398 	if (sh->encaps_decaps) {
399 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
400 		sh->encaps_decaps = NULL;
401 	}
402 	if (sh->modify_cmds) {
403 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
404 		sh->modify_cmds = NULL;
405 	}
406 	if (sh->tag_table) {
407 		/* tags should be destroyed with flow before. */
408 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
409 		sh->tag_table = NULL;
410 	}
411 	if (sh->tunnel_hub) {
412 		mlx5_release_tunnel_hub(sh, priv->dev_port);
413 		sh->tunnel_hub = NULL;
414 	}
415 	mlx5_free_table_hash_list(priv);
416 }
417 
418 /**
419  * Initialize shared data between primary and secondary process.
420  *
421  * A memzone is reserved by primary process and secondary processes attach to
422  * the memzone.
423  *
424  * @return
425  *   0 on success, a negative errno value otherwise and rte_errno is set.
426  */
427 static int
428 mlx5_init_shared_data(void)
429 {
430 	const struct rte_memzone *mz;
431 	int ret = 0;
432 
433 	rte_spinlock_lock(&mlx5_shared_data_lock);
434 	if (mlx5_shared_data == NULL) {
435 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
436 			/* Allocate shared memory. */
437 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
438 						 sizeof(*mlx5_shared_data),
439 						 SOCKET_ID_ANY, 0);
440 			if (mz == NULL) {
441 				DRV_LOG(ERR,
442 					"Cannot allocate mlx5 shared data");
443 				ret = -rte_errno;
444 				goto error;
445 			}
446 			mlx5_shared_data = mz->addr;
447 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
448 			rte_spinlock_init(&mlx5_shared_data->lock);
449 		} else {
450 			/* Lookup allocated shared memory. */
451 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
452 			if (mz == NULL) {
453 				DRV_LOG(ERR,
454 					"Cannot attach mlx5 shared data");
455 				ret = -rte_errno;
456 				goto error;
457 			}
458 			mlx5_shared_data = mz->addr;
459 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
460 		}
461 	}
462 error:
463 	rte_spinlock_unlock(&mlx5_shared_data_lock);
464 	return ret;
465 }
466 
467 /**
468  * PMD global initialization.
469  *
470  * Independent from individual device, this function initializes global
471  * per-PMD data structures distinguishing primary and secondary processes.
472  * Hence, each initialization is called once per a process.
473  *
474  * @return
475  *   0 on success, a negative errno value otherwise and rte_errno is set.
476  */
477 static int
478 mlx5_init_once(void)
479 {
480 	struct mlx5_shared_data *sd;
481 	struct mlx5_local_data *ld = &mlx5_local_data;
482 	int ret = 0;
483 
484 	if (mlx5_init_shared_data())
485 		return -rte_errno;
486 	sd = mlx5_shared_data;
487 	MLX5_ASSERT(sd);
488 	rte_spinlock_lock(&sd->lock);
489 	switch (rte_eal_process_type()) {
490 	case RTE_PROC_PRIMARY:
491 		if (sd->init_done)
492 			break;
493 		LIST_INIT(&sd->mem_event_cb_list);
494 		rte_rwlock_init(&sd->mem_event_rwlock);
495 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
496 						mlx5_mr_mem_event_cb, NULL);
497 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
498 					   mlx5_mp_os_primary_handle);
499 		if (ret)
500 			goto out;
501 		sd->init_done = true;
502 		break;
503 	case RTE_PROC_SECONDARY:
504 		if (ld->init_done)
505 			break;
506 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
507 					     mlx5_mp_os_secondary_handle);
508 		if (ret)
509 			goto out;
510 		++sd->secondary_cnt;
511 		ld->init_done = true;
512 		break;
513 	default:
514 		break;
515 	}
516 out:
517 	rte_spinlock_unlock(&sd->lock);
518 	return ret;
519 }
520 
521 /**
522  * Create the Tx queue DevX/Verbs object.
523  *
524  * @param dev
525  *   Pointer to Ethernet device.
526  * @param idx
527  *   Queue index in DPDK Tx queue array.
528  *
529  * @return
530  *   0 on success, a negative errno value otherwise and rte_errno is set.
531  */
532 static int
533 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
534 {
535 	struct mlx5_priv *priv = dev->data->dev_private;
536 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
537 	struct mlx5_txq_ctrl *txq_ctrl =
538 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
539 
540 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
541 		return mlx5_txq_devx_obj_new(dev, idx);
542 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
543 	if (!priv->config.dv_esw_en)
544 		return mlx5_txq_devx_obj_new(dev, idx);
545 #endif
546 	return mlx5_txq_ibv_obj_new(dev, idx);
547 }
548 
549 /**
550  * Release an Tx DevX/verbs queue object.
551  *
552  * @param txq_obj
553  *   DevX/Verbs Tx queue object.
554  */
555 static void
556 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
557 {
558 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
559 		mlx5_txq_devx_obj_release(txq_obj);
560 		return;
561 	}
562 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
563 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
564 		mlx5_txq_devx_obj_release(txq_obj);
565 		return;
566 	}
567 #endif
568 	mlx5_txq_ibv_obj_release(txq_obj);
569 }
570 
571 /**
572  * DV flow counter mode detect and config.
573  *
574  * @param dev
575  *   Pointer to rte_eth_dev structure.
576  *
577  */
578 static void
579 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
580 {
581 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
582 	struct mlx5_priv *priv = dev->data->dev_private;
583 	struct mlx5_dev_ctx_shared *sh = priv->sh;
584 	bool fallback;
585 
586 #ifndef HAVE_IBV_DEVX_ASYNC
587 	fallback = true;
588 #else
589 	fallback = false;
590 	if (!priv->config.devx || !priv->config.dv_flow_en ||
591 	    !priv->config.hca_attr.flow_counters_dump ||
592 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
593 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
594 		fallback = true;
595 #endif
596 	if (fallback)
597 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
598 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
599 			priv->config.hca_attr.flow_counters_dump,
600 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
601 	/* Initialize fallback mode only on the port initializes sh. */
602 	if (sh->refcnt == 1)
603 		sh->cmng.counter_fallback = fallback;
604 	else if (fallback != sh->cmng.counter_fallback)
605 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
606 			"with others:%d.", PORT_ID(priv), fallback);
607 #endif
608 }
609 
610 /**
611  * Spawn an Ethernet device from Verbs information.
612  *
613  * @param dpdk_dev
614  *   Backing DPDK device.
615  * @param spawn
616  *   Verbs device parameters (name, port, switch_info) to spawn.
617  * @param config
618  *   Device configuration parameters.
619  *
620  * @return
621  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
622  *   is set. The following errors are defined:
623  *
624  *   EBUSY: device is not supposed to be spawned.
625  *   EEXIST: device is already spawned
626  */
627 static struct rte_eth_dev *
628 mlx5_dev_spawn(struct rte_device *dpdk_dev,
629 	       struct mlx5_dev_spawn_data *spawn,
630 	       struct mlx5_dev_config *config)
631 {
632 	const struct mlx5_switch_info *switch_info = &spawn->info;
633 	struct mlx5_dev_ctx_shared *sh = NULL;
634 	struct ibv_port_attr port_attr;
635 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
636 	struct rte_eth_dev *eth_dev = NULL;
637 	struct mlx5_priv *priv = NULL;
638 	int err = 0;
639 	unsigned int hw_padding = 0;
640 	unsigned int mps;
641 	unsigned int cqe_comp;
642 	unsigned int cqe_pad = 0;
643 	unsigned int tunnel_en = 0;
644 	unsigned int mpls_en = 0;
645 	unsigned int swp = 0;
646 	unsigned int mprq = 0;
647 	unsigned int mprq_min_stride_size_n = 0;
648 	unsigned int mprq_max_stride_size_n = 0;
649 	unsigned int mprq_min_stride_num_n = 0;
650 	unsigned int mprq_max_stride_num_n = 0;
651 	struct rte_ether_addr mac;
652 	char name[RTE_ETH_NAME_MAX_LEN];
653 	int own_domain_id = 0;
654 	uint16_t port_id;
655 	unsigned int i;
656 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
657 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
658 #endif
659 
660 	/* Determine if this port representor is supposed to be spawned. */
661 	if (switch_info->representor && dpdk_dev->devargs) {
662 		struct rte_eth_devargs eth_da;
663 
664 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
665 		if (err) {
666 			rte_errno = -err;
667 			DRV_LOG(ERR, "failed to process device arguments: %s",
668 				strerror(rte_errno));
669 			return NULL;
670 		}
671 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
672 			if (eth_da.representor_ports[i] ==
673 			    (uint16_t)switch_info->port_name)
674 				break;
675 		if (i == eth_da.nb_representor_ports) {
676 			rte_errno = EBUSY;
677 			return NULL;
678 		}
679 	}
680 	/* Build device name. */
681 	if (spawn->pf_bond <  0) {
682 		/* Single device. */
683 		if (!switch_info->representor)
684 			strlcpy(name, dpdk_dev->name, sizeof(name));
685 		else
686 			snprintf(name, sizeof(name), "%s_representor_%u",
687 				 dpdk_dev->name, switch_info->port_name);
688 	} else {
689 		/* Bonding device. */
690 		if (!switch_info->representor)
691 			snprintf(name, sizeof(name), "%s_%s",
692 				 dpdk_dev->name,
693 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
694 		else
695 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
696 				 dpdk_dev->name,
697 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
698 				 switch_info->port_name);
699 	}
700 	/* check if the device is already spawned */
701 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
702 		rte_errno = EEXIST;
703 		return NULL;
704 	}
705 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
706 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
707 		struct mlx5_mp_id mp_id;
708 
709 		eth_dev = rte_eth_dev_attach_secondary(name);
710 		if (eth_dev == NULL) {
711 			DRV_LOG(ERR, "can not attach rte ethdev");
712 			rte_errno = ENOMEM;
713 			return NULL;
714 		}
715 		eth_dev->device = dpdk_dev;
716 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
717 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
718 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
719 		err = mlx5_proc_priv_init(eth_dev);
720 		if (err)
721 			return NULL;
722 		mp_id.port_id = eth_dev->data->port_id;
723 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
724 		/* Receive command fd from primary process */
725 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
726 		if (err < 0)
727 			goto err_secondary;
728 		/* Remap UAR for Tx queues. */
729 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
730 		if (err)
731 			goto err_secondary;
732 		/*
733 		 * Ethdev pointer is still required as input since
734 		 * the primary device is not accessible from the
735 		 * secondary process.
736 		 */
737 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
738 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
739 		return eth_dev;
740 err_secondary:
741 		mlx5_dev_close(eth_dev);
742 		return NULL;
743 	}
744 	/*
745 	 * Some parameters ("tx_db_nc" in particularly) are needed in
746 	 * advance to create dv/verbs device context. We proceed the
747 	 * devargs here to get ones, and later proceed devargs again
748 	 * to override some hardware settings.
749 	 */
750 	err = mlx5_args(config, dpdk_dev->devargs);
751 	if (err) {
752 		err = rte_errno;
753 		DRV_LOG(ERR, "failed to process device arguments: %s",
754 			strerror(rte_errno));
755 		goto error;
756 	}
757 	if (config->dv_miss_info) {
758 		if (switch_info->master || switch_info->representor)
759 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
760 	}
761 	mlx5_malloc_mem_select(config->sys_mem_en);
762 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
763 	if (!sh)
764 		return NULL;
765 	config->devx = sh->devx;
766 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
767 	config->dest_tir = 1;
768 #endif
769 #ifdef HAVE_IBV_MLX5_MOD_SWP
770 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
771 #endif
772 	/*
773 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
774 	 * as all ConnectX-5 devices.
775 	 */
776 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
777 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
778 #endif
779 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
780 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
781 #endif
782 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
783 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
784 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
785 			DRV_LOG(DEBUG, "enhanced MPW is supported");
786 			mps = MLX5_MPW_ENHANCED;
787 		} else {
788 			DRV_LOG(DEBUG, "MPW is supported");
789 			mps = MLX5_MPW;
790 		}
791 	} else {
792 		DRV_LOG(DEBUG, "MPW isn't supported");
793 		mps = MLX5_MPW_DISABLED;
794 	}
795 #ifdef HAVE_IBV_MLX5_MOD_SWP
796 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
797 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
798 	DRV_LOG(DEBUG, "SWP support: %u", swp);
799 #endif
800 	config->swp = !!swp;
801 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
802 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
803 		struct mlx5dv_striding_rq_caps mprq_caps =
804 			dv_attr.striding_rq_caps;
805 
806 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
807 			mprq_caps.min_single_stride_log_num_of_bytes);
808 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
809 			mprq_caps.max_single_stride_log_num_of_bytes);
810 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
811 			mprq_caps.min_single_wqe_log_num_of_strides);
812 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
813 			mprq_caps.max_single_wqe_log_num_of_strides);
814 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
815 			mprq_caps.supported_qpts);
816 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
817 		mprq = 1;
818 		mprq_min_stride_size_n =
819 			mprq_caps.min_single_stride_log_num_of_bytes;
820 		mprq_max_stride_size_n =
821 			mprq_caps.max_single_stride_log_num_of_bytes;
822 		mprq_min_stride_num_n =
823 			mprq_caps.min_single_wqe_log_num_of_strides;
824 		mprq_max_stride_num_n =
825 			mprq_caps.max_single_wqe_log_num_of_strides;
826 	}
827 #endif
828 	if (RTE_CACHE_LINE_SIZE == 128 &&
829 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
830 		cqe_comp = 0;
831 	else
832 		cqe_comp = 1;
833 	config->cqe_comp = cqe_comp;
834 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
835 	/* Whether device supports 128B Rx CQE padding. */
836 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
837 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
838 #endif
839 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
840 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
841 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
842 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
843 			     (dv_attr.tunnel_offloads_caps &
844 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
845 			     (dv_attr.tunnel_offloads_caps &
846 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
847 	}
848 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
849 		tunnel_en ? "" : "not ");
850 #else
851 	DRV_LOG(WARNING,
852 		"tunnel offloading disabled due to old OFED/rdma-core version");
853 #endif
854 	config->tunnel_en = tunnel_en;
855 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
856 	mpls_en = ((dv_attr.tunnel_offloads_caps &
857 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
858 		   (dv_attr.tunnel_offloads_caps &
859 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
860 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
861 		mpls_en ? "" : "not ");
862 #else
863 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
864 		" old OFED/rdma-core version or firmware configuration");
865 #endif
866 	config->mpls_en = mpls_en;
867 	/* Check port status. */
868 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
869 	if (err) {
870 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
871 		goto error;
872 	}
873 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
874 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
875 		err = EINVAL;
876 		goto error;
877 	}
878 	if (port_attr.state != IBV_PORT_ACTIVE)
879 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
880 			mlx5_glue->port_state_str(port_attr.state),
881 			port_attr.state);
882 	/* Allocate private eth device data. */
883 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
884 			   sizeof(*priv),
885 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
886 	if (priv == NULL) {
887 		DRV_LOG(ERR, "priv allocation failure");
888 		err = ENOMEM;
889 		goto error;
890 	}
891 	priv->sh = sh;
892 	priv->dev_port = spawn->phys_port;
893 	priv->pci_dev = spawn->pci_dev;
894 	priv->mtu = RTE_ETHER_MTU;
895 	priv->mp_id.port_id = port_id;
896 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
897 	/* Some internal functions rely on Netlink sockets, open them now. */
898 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
899 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
900 	priv->representor = !!switch_info->representor;
901 	priv->master = !!switch_info->master;
902 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
903 	priv->vport_meta_tag = 0;
904 	priv->vport_meta_mask = 0;
905 	priv->pf_bond = spawn->pf_bond;
906 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
907 	/*
908 	 * The DevX port query API is implemented. E-Switch may use
909 	 * either vport or reg_c[0] metadata register to match on
910 	 * vport index. The engaged part of metadata register is
911 	 * defined by mask.
912 	 */
913 	if (switch_info->representor || switch_info->master) {
914 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
915 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
916 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
917 						 &devx_port);
918 		if (err) {
919 			DRV_LOG(WARNING,
920 				"can't query devx port %d on device %s",
921 				spawn->phys_port,
922 				mlx5_os_get_dev_device_name(spawn->phys_dev));
923 			devx_port.comp_mask = 0;
924 		}
925 	}
926 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
927 		priv->vport_meta_tag = devx_port.reg_c_0.value;
928 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
929 		if (!priv->vport_meta_mask) {
930 			DRV_LOG(ERR, "vport zero mask for port %d"
931 				     " on bonding device %s",
932 				     spawn->phys_port,
933 				     mlx5_os_get_dev_device_name
934 							(spawn->phys_dev));
935 			err = ENOTSUP;
936 			goto error;
937 		}
938 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
939 			DRV_LOG(ERR, "invalid vport tag for port %d"
940 				     " on bonding device %s",
941 				     spawn->phys_port,
942 				     mlx5_os_get_dev_device_name
943 							(spawn->phys_dev));
944 			err = ENOTSUP;
945 			goto error;
946 		}
947 	}
948 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
949 		priv->vport_id = devx_port.vport_num;
950 	} else if (spawn->pf_bond >= 0) {
951 		DRV_LOG(ERR, "can't deduce vport index for port %d"
952 			     " on bonding device %s",
953 			     spawn->phys_port,
954 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
955 		err = ENOTSUP;
956 		goto error;
957 	} else {
958 		/* Suppose vport index in compatible way. */
959 		priv->vport_id = switch_info->representor ?
960 				 switch_info->port_name + 1 : -1;
961 	}
962 #else
963 	/*
964 	 * Kernel/rdma_core support single E-Switch per PF configurations
965 	 * only and vport_id field contains the vport index for
966 	 * associated VF, which is deduced from representor port name.
967 	 * For example, let's have the IB device port 10, it has
968 	 * attached network device eth0, which has port name attribute
969 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
970 	 * as 3 (2+1). This assigning schema should be changed if the
971 	 * multiple E-Switch instances per PF configurations or/and PCI
972 	 * subfunctions are added.
973 	 */
974 	priv->vport_id = switch_info->representor ?
975 			 switch_info->port_name + 1 : -1;
976 #endif
977 	/* representor_id field keeps the unmodified VF index. */
978 	priv->representor_id = switch_info->representor ?
979 			       switch_info->port_name : -1;
980 	/*
981 	 * Look for sibling devices in order to reuse their switch domain
982 	 * if any, otherwise allocate one.
983 	 */
984 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
985 		const struct mlx5_priv *opriv =
986 			rte_eth_devices[port_id].data->dev_private;
987 
988 		if (!opriv ||
989 		    opriv->sh != priv->sh ||
990 			opriv->domain_id ==
991 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
992 			continue;
993 		priv->domain_id = opriv->domain_id;
994 		break;
995 	}
996 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
997 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
998 		if (err) {
999 			err = rte_errno;
1000 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1001 				strerror(rte_errno));
1002 			goto error;
1003 		}
1004 		own_domain_id = 1;
1005 	}
1006 	/* Override some values set by hardware configuration. */
1007 	mlx5_args(config, dpdk_dev->devargs);
1008 	err = mlx5_dev_check_sibling_config(priv, config);
1009 	if (err)
1010 		goto error;
1011 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1012 			    IBV_DEVICE_RAW_IP_CSUM);
1013 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1014 		(config->hw_csum ? "" : "not "));
1015 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1016 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1017 	DRV_LOG(DEBUG, "counters are not supported");
1018 #endif
1019 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1020 	if (config->dv_flow_en) {
1021 		DRV_LOG(WARNING, "DV flow is not supported");
1022 		config->dv_flow_en = 0;
1023 	}
1024 #endif
1025 	config->ind_table_max_size =
1026 		sh->device_attr.max_rwq_indirection_table_size;
1027 	/*
1028 	 * Remove this check once DPDK supports larger/variable
1029 	 * indirection tables.
1030 	 */
1031 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1032 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1033 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1034 		config->ind_table_max_size);
1035 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1036 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1037 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1038 		(config->hw_vlan_strip ? "" : "not "));
1039 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1040 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1041 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1042 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1043 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1044 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1045 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1046 #endif
1047 	if (config->hw_padding && !hw_padding) {
1048 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1049 		config->hw_padding = 0;
1050 	} else if (config->hw_padding) {
1051 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1052 	}
1053 	config->tso = (sh->device_attr.max_tso > 0 &&
1054 		      (sh->device_attr.tso_supported_qpts &
1055 		       (1 << IBV_QPT_RAW_PACKET)));
1056 	if (config->tso)
1057 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1058 	/*
1059 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1060 	 * by default.
1061 	 */
1062 	if (config->mps == MLX5_ARG_UNSET)
1063 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1064 							  MLX5_MPW_DISABLED;
1065 	else
1066 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1067 	DRV_LOG(INFO, "%sMPS is %s",
1068 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1069 		config->mps == MLX5_MPW ? "legacy " : "",
1070 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1071 	if (config->cqe_comp && !cqe_comp) {
1072 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1073 		config->cqe_comp = 0;
1074 	}
1075 	if (config->cqe_pad && !cqe_pad) {
1076 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1077 		config->cqe_pad = 0;
1078 	} else if (config->cqe_pad) {
1079 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1080 	}
1081 	if (config->devx) {
1082 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1083 		if (err) {
1084 			err = -err;
1085 			goto error;
1086 		}
1087 		/* Check relax ordering support. */
1088 		if (config->hca_attr.relaxed_ordering_write &&
1089 		    config->hca_attr.relaxed_ordering_read  &&
1090 		    !haswell_broadwell_cpu)
1091 			sh->cmng.relaxed_ordering = 1;
1092 		/* Check for LRO support. */
1093 		if (config->dest_tir && config->hca_attr.lro_cap &&
1094 		    config->dv_flow_en) {
1095 			/* TBD check tunnel lro caps. */
1096 			config->lro.supported = config->hca_attr.lro_cap;
1097 			DRV_LOG(DEBUG, "Device supports LRO");
1098 			/*
1099 			 * If LRO timeout is not configured by application,
1100 			 * use the minimal supported value.
1101 			 */
1102 			if (!config->lro.timeout)
1103 				config->lro.timeout =
1104 				config->hca_attr.lro_timer_supported_periods[0];
1105 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1106 				config->lro.timeout);
1107 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1108 				"required for coalescing is %d bytes",
1109 				config->hca_attr.lro_min_mss_size);
1110 		}
1111 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1112 		if (config->hca_attr.qos.sup &&
1113 		    config->hca_attr.qos.srtcm_sup &&
1114 		    config->dv_flow_en) {
1115 			uint8_t reg_c_mask =
1116 				config->hca_attr.qos.flow_meter_reg_c_ids;
1117 			/*
1118 			 * Meter needs two REG_C's for color match and pre-sfx
1119 			 * flow match. Here get the REG_C for color match.
1120 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1121 			 */
1122 			reg_c_mask &= 0xfc;
1123 			if (__builtin_popcount(reg_c_mask) < 1) {
1124 				priv->mtr_en = 0;
1125 				DRV_LOG(WARNING, "No available register for"
1126 					" meter.");
1127 			} else {
1128 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1129 						      REG_C_0;
1130 				priv->mtr_en = 1;
1131 				priv->mtr_reg_share =
1132 				      config->hca_attr.qos.flow_meter_reg_share;
1133 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1134 					priv->mtr_color_reg);
1135 			}
1136 		}
1137 #endif
1138 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1139 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1140 		    config->dv_flow_en) {
1141 			priv->sampler_en = 1;
1142 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1143 		} else {
1144 			priv->sampler_en = 0;
1145 			if (!config->hca_attr.log_max_ft_sampler_num)
1146 				DRV_LOG(WARNING, "No available register for"
1147 						" Sampler.");
1148 			else
1149 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1150 		}
1151 #endif
1152 	}
1153 	if (config->tx_pp) {
1154 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1155 			config->hca_attr.dev_freq_khz);
1156 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1157 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1158 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1159 			config->hca_attr.cross_channel ? "" : "not ");
1160 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1161 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1162 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1163 			config->hca_attr.non_wire_sq ? "" : "not ");
1164 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1165 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1166 			config->hca_attr.log_max_static_sq_wq);
1167 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1168 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1169 		if (!config->devx) {
1170 			DRV_LOG(ERR, "DevX is required for packet pacing");
1171 			err = ENODEV;
1172 			goto error;
1173 		}
1174 		if (!config->hca_attr.qos.packet_pacing) {
1175 			DRV_LOG(ERR, "Packet pacing is not supported");
1176 			err = ENODEV;
1177 			goto error;
1178 		}
1179 		if (!config->hca_attr.cross_channel) {
1180 			DRV_LOG(ERR, "Cross channel operations are"
1181 				     " required for packet pacing");
1182 			err = ENODEV;
1183 			goto error;
1184 		}
1185 		if (!config->hca_attr.wqe_index_ignore) {
1186 			DRV_LOG(ERR, "WQE index ignore feature is"
1187 				     " required for packet pacing");
1188 			err = ENODEV;
1189 			goto error;
1190 		}
1191 		if (!config->hca_attr.non_wire_sq) {
1192 			DRV_LOG(ERR, "Non-wire SQ feature is"
1193 				     " required for packet pacing");
1194 			err = ENODEV;
1195 			goto error;
1196 		}
1197 		if (!config->hca_attr.log_max_static_sq_wq) {
1198 			DRV_LOG(ERR, "Static WQE SQ feature is"
1199 				     " required for packet pacing");
1200 			err = ENODEV;
1201 			goto error;
1202 		}
1203 		if (!config->hca_attr.qos.wqe_rate_pp) {
1204 			DRV_LOG(ERR, "WQE rate mode is required"
1205 				     " for packet pacing");
1206 			err = ENODEV;
1207 			goto error;
1208 		}
1209 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1210 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1211 			     " can't create queues for packet pacing");
1212 		err = ENODEV;
1213 		goto error;
1214 #endif
1215 	}
1216 	if (config->devx) {
1217 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1218 
1219 		err = config->hca_attr.access_register_user ?
1220 			mlx5_devx_cmd_register_read
1221 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1222 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1223 		if (!err) {
1224 			uint32_t ts_mode;
1225 
1226 			/* MTUTC register is read successfully. */
1227 			ts_mode = MLX5_GET(register_mtutc, reg,
1228 					   time_stamp_mode);
1229 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1230 				config->rt_timestamp = 1;
1231 		} else {
1232 			/* Kernel does not support register reading. */
1233 			if (config->hca_attr.dev_freq_khz ==
1234 						 (NS_PER_S / MS_PER_S))
1235 				config->rt_timestamp = 1;
1236 		}
1237 	}
1238 	/*
1239 	 * If HW has bug working with tunnel packet decapsulation and
1240 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1241 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1242 	 */
1243 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1244 		config->hw_fcs_strip = 0;
1245 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1246 		(config->hw_fcs_strip ? "" : "not "));
1247 	if (config->mprq.enabled && mprq) {
1248 		if (config->mprq.stride_num_n &&
1249 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1250 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1251 			config->mprq.stride_num_n =
1252 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1253 						mprq_min_stride_num_n),
1254 					mprq_max_stride_num_n);
1255 			DRV_LOG(WARNING,
1256 				"the number of strides"
1257 				" for Multi-Packet RQ is out of range,"
1258 				" setting default value (%u)",
1259 				1 << config->mprq.stride_num_n);
1260 		}
1261 		if (config->mprq.stride_size_n &&
1262 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1263 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1264 			config->mprq.stride_size_n =
1265 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1266 						mprq_min_stride_size_n),
1267 					mprq_max_stride_size_n);
1268 			DRV_LOG(WARNING,
1269 				"the size of a stride"
1270 				" for Multi-Packet RQ is out of range,"
1271 				" setting default value (%u)",
1272 				1 << config->mprq.stride_size_n);
1273 		}
1274 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1275 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1276 	} else if (config->mprq.enabled && !mprq) {
1277 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1278 		config->mprq.enabled = 0;
1279 	}
1280 	if (config->max_dump_files_num == 0)
1281 		config->max_dump_files_num = 128;
1282 	eth_dev = rte_eth_dev_allocate(name);
1283 	if (eth_dev == NULL) {
1284 		DRV_LOG(ERR, "can not allocate rte ethdev");
1285 		err = ENOMEM;
1286 		goto error;
1287 	}
1288 	if (priv->representor) {
1289 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1290 		eth_dev->data->representor_id = priv->representor_id;
1291 	}
1292 	/*
1293 	 * Store associated network device interface index. This index
1294 	 * is permanent throughout the lifetime of device. So, we may store
1295 	 * the ifindex here and use the cached value further.
1296 	 */
1297 	MLX5_ASSERT(spawn->ifindex);
1298 	priv->if_index = spawn->ifindex;
1299 	if (priv->pf_bond >= 0 && priv->master) {
1300 		/* Get bond interface info */
1301 		err = mlx5_sysfs_bond_info(priv->if_index,
1302 				     &priv->bond_ifindex,
1303 				     priv->bond_name);
1304 		if (err)
1305 			DRV_LOG(ERR, "unable to get bond info: %s",
1306 				strerror(rte_errno));
1307 		else
1308 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1309 				priv->if_index, priv->bond_ifindex,
1310 				priv->bond_name);
1311 	}
1312 	eth_dev->data->dev_private = priv;
1313 	priv->dev_data = eth_dev->data;
1314 	eth_dev->data->mac_addrs = priv->mac;
1315 	eth_dev->device = dpdk_dev;
1316 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1317 	/* Configure the first MAC address by default. */
1318 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1319 		DRV_LOG(ERR,
1320 			"port %u cannot get MAC address, is mlx5_en"
1321 			" loaded? (errno: %s)",
1322 			eth_dev->data->port_id, strerror(rte_errno));
1323 		err = ENODEV;
1324 		goto error;
1325 	}
1326 	DRV_LOG(INFO,
1327 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1328 		eth_dev->data->port_id,
1329 		mac.addr_bytes[0], mac.addr_bytes[1],
1330 		mac.addr_bytes[2], mac.addr_bytes[3],
1331 		mac.addr_bytes[4], mac.addr_bytes[5]);
1332 #ifdef RTE_LIBRTE_MLX5_DEBUG
1333 	{
1334 		char ifname[IF_NAMESIZE];
1335 
1336 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1337 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1338 				eth_dev->data->port_id, ifname);
1339 		else
1340 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1341 				eth_dev->data->port_id);
1342 	}
1343 #endif
1344 	/* Get actual MTU if possible. */
1345 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1346 	if (err) {
1347 		err = rte_errno;
1348 		goto error;
1349 	}
1350 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1351 		priv->mtu);
1352 	/* Initialize burst functions to prevent crashes before link-up. */
1353 	eth_dev->rx_pkt_burst = removed_rx_burst;
1354 	eth_dev->tx_pkt_burst = removed_tx_burst;
1355 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1356 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1357 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1358 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1359 	/* Register MAC address. */
1360 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1361 	if (config->vf && config->vf_nl_en)
1362 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1363 				      mlx5_ifindex(eth_dev),
1364 				      eth_dev->data->mac_addrs,
1365 				      MLX5_MAX_MAC_ADDRESSES);
1366 	priv->flows = 0;
1367 	priv->ctrl_flows = 0;
1368 	rte_spinlock_init(&priv->flow_list_lock);
1369 	TAILQ_INIT(&priv->flow_meters);
1370 	TAILQ_INIT(&priv->flow_meter_profiles);
1371 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1372 	mlx5_glue->dv_set_context_attr(sh->ctx,
1373 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1374 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1375 				.alloc = &mlx5_alloc_verbs_buf,
1376 				.free = &mlx5_free_verbs_buf,
1377 				.data = priv,
1378 			}));
1379 	/* Bring Ethernet device up. */
1380 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1381 		eth_dev->data->port_id);
1382 	mlx5_set_link_up(eth_dev);
1383 	/*
1384 	 * Even though the interrupt handler is not installed yet,
1385 	 * interrupts will still trigger on the async_fd from
1386 	 * Verbs context returned by ibv_open_device().
1387 	 */
1388 	mlx5_link_update(eth_dev, 0);
1389 #ifdef HAVE_MLX5DV_DR_ESWITCH
1390 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1391 	      (switch_info->representor || switch_info->master)))
1392 		config->dv_esw_en = 0;
1393 #else
1394 	config->dv_esw_en = 0;
1395 #endif
1396 	/* Detect minimal data bytes to inline. */
1397 	mlx5_set_min_inline(spawn, config);
1398 	/* Store device configuration on private structure. */
1399 	priv->config = *config;
1400 	/* Create context for virtual machine VLAN workaround. */
1401 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1402 	if (config->dv_flow_en) {
1403 		err = mlx5_alloc_shared_dr(priv);
1404 		if (err)
1405 			goto error;
1406 	}
1407 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1408 		priv->obj_ops = devx_obj_ops;
1409 		priv->obj_ops.drop_action_create =
1410 						ibv_obj_ops.drop_action_create;
1411 		priv->obj_ops.drop_action_destroy =
1412 						ibv_obj_ops.drop_action_destroy;
1413 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1414 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1415 #else
1416 		if (config->dv_esw_en)
1417 			priv->obj_ops.txq_obj_modify =
1418 						ibv_obj_ops.txq_obj_modify;
1419 #endif
1420 		/* Use specific wrappers for Tx object. */
1421 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1422 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1423 
1424 	} else {
1425 		priv->obj_ops = ibv_obj_ops;
1426 	}
1427 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1428 	if (!priv->drop_queue.hrxq)
1429 		goto error;
1430 	/* Supported Verbs flow priority number detection. */
1431 	err = mlx5_flow_discover_priorities(eth_dev);
1432 	if (err < 0) {
1433 		err = -err;
1434 		goto error;
1435 	}
1436 	priv->config.flow_prio = err;
1437 	if (!priv->config.dv_esw_en &&
1438 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1439 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1440 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1441 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1442 	}
1443 	mlx5_set_metadata_mask(eth_dev);
1444 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1445 	    !priv->sh->dv_regc0_mask) {
1446 		DRV_LOG(ERR, "metadata mode %u is not supported "
1447 			     "(no metadata reg_c[0] is available)",
1448 			     priv->config.dv_xmeta_en);
1449 			err = ENOTSUP;
1450 			goto error;
1451 	}
1452 	/* Query availability of metadata reg_c's. */
1453 	err = mlx5_flow_discover_mreg_c(eth_dev);
1454 	if (err < 0) {
1455 		err = -err;
1456 		goto error;
1457 	}
1458 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1459 		DRV_LOG(DEBUG,
1460 			"port %u extensive metadata register is not supported",
1461 			eth_dev->data->port_id);
1462 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1463 			DRV_LOG(ERR, "metadata mode %u is not supported "
1464 				     "(no metadata registers available)",
1465 				     priv->config.dv_xmeta_en);
1466 			err = ENOTSUP;
1467 			goto error;
1468 		}
1469 	}
1470 	if (priv->config.dv_flow_en &&
1471 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1472 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1473 	    priv->sh->dv_regc0_mask) {
1474 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1475 						      MLX5_FLOW_MREG_HTABLE_SZ);
1476 		if (!priv->mreg_cp_tbl) {
1477 			err = ENOMEM;
1478 			goto error;
1479 		}
1480 	}
1481 	mlx5_flow_counter_mode_config(eth_dev);
1482 	return eth_dev;
1483 error:
1484 	if (priv) {
1485 		if (priv->mreg_cp_tbl)
1486 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1487 		if (priv->sh)
1488 			mlx5_os_free_shared_dr(priv);
1489 		if (priv->nl_socket_route >= 0)
1490 			close(priv->nl_socket_route);
1491 		if (priv->nl_socket_rdma >= 0)
1492 			close(priv->nl_socket_rdma);
1493 		if (priv->vmwa_context)
1494 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1495 		if (eth_dev && priv->drop_queue.hrxq)
1496 			mlx5_drop_action_destroy(eth_dev);
1497 		if (own_domain_id)
1498 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1499 		mlx5_free(priv);
1500 		if (eth_dev != NULL)
1501 			eth_dev->data->dev_private = NULL;
1502 	}
1503 	if (eth_dev != NULL) {
1504 		/* mac_addrs must not be freed alone because part of
1505 		 * dev_private
1506 		 **/
1507 		eth_dev->data->mac_addrs = NULL;
1508 		rte_eth_dev_release_port(eth_dev);
1509 	}
1510 	if (sh)
1511 		mlx5_free_shared_dev_ctx(sh);
1512 	MLX5_ASSERT(err > 0);
1513 	rte_errno = err;
1514 	return NULL;
1515 }
1516 
1517 /**
1518  * Comparison callback to sort device data.
1519  *
1520  * This is meant to be used with qsort().
1521  *
1522  * @param a[in]
1523  *   Pointer to pointer to first data object.
1524  * @param b[in]
1525  *   Pointer to pointer to second data object.
1526  *
1527  * @return
1528  *   0 if both objects are equal, less than 0 if the first argument is less
1529  *   than the second, greater than 0 otherwise.
1530  */
1531 static int
1532 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1533 {
1534 	const struct mlx5_switch_info *si_a =
1535 		&((const struct mlx5_dev_spawn_data *)a)->info;
1536 	const struct mlx5_switch_info *si_b =
1537 		&((const struct mlx5_dev_spawn_data *)b)->info;
1538 	int ret;
1539 
1540 	/* Master device first. */
1541 	ret = si_b->master - si_a->master;
1542 	if (ret)
1543 		return ret;
1544 	/* Then representor devices. */
1545 	ret = si_b->representor - si_a->representor;
1546 	if (ret)
1547 		return ret;
1548 	/* Unidentified devices come last in no specific order. */
1549 	if (!si_a->representor)
1550 		return 0;
1551 	/* Order representors by name. */
1552 	return si_a->port_name - si_b->port_name;
1553 }
1554 
1555 /**
1556  * Match PCI information for possible slaves of bonding device.
1557  *
1558  * @param[in] ibv_dev
1559  *   Pointer to Infiniband device structure.
1560  * @param[in] pci_dev
1561  *   Pointer to PCI device structure to match PCI address.
1562  * @param[in] nl_rdma
1563  *   Netlink RDMA group socket handle.
1564  *
1565  * @return
1566  *   negative value if no bonding device found, otherwise
1567  *   positive index of slave PF in bonding.
1568  */
1569 static int
1570 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1571 			   const struct rte_pci_device *pci_dev,
1572 			   int nl_rdma)
1573 {
1574 	char ifname[IF_NAMESIZE + 1];
1575 	unsigned int ifindex;
1576 	unsigned int np, i;
1577 	FILE *file = NULL;
1578 	int pf = -1;
1579 
1580 	/*
1581 	 * Try to get master device name. If something goes
1582 	 * wrong suppose the lack of kernel support and no
1583 	 * bonding devices.
1584 	 */
1585 	if (nl_rdma < 0)
1586 		return -1;
1587 	if (!strstr(ibv_dev->name, "bond"))
1588 		return -1;
1589 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1590 	if (!np)
1591 		return -1;
1592 	/*
1593 	 * The Master device might not be on the predefined
1594 	 * port (not on port index 1, it is not garanted),
1595 	 * we have to scan all Infiniband device port and
1596 	 * find master.
1597 	 */
1598 	for (i = 1; i <= np; ++i) {
1599 		/* Check whether Infiniband port is populated. */
1600 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1601 		if (!ifindex)
1602 			continue;
1603 		if (!if_indextoname(ifindex, ifname))
1604 			continue;
1605 		/* Try to read bonding slave names from sysfs. */
1606 		MKSTR(slaves,
1607 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1608 		file = fopen(slaves, "r");
1609 		if (file)
1610 			break;
1611 	}
1612 	if (!file)
1613 		return -1;
1614 	/* Use safe format to check maximal buffer length. */
1615 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1616 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1617 		char tmp_str[IF_NAMESIZE + 32];
1618 		struct rte_pci_addr pci_addr;
1619 		struct mlx5_switch_info	info;
1620 
1621 		/* Process slave interface names in the loop. */
1622 		snprintf(tmp_str, sizeof(tmp_str),
1623 			 "/sys/class/net/%s", ifname);
1624 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1625 			DRV_LOG(WARNING, "can not get PCI address"
1626 					 " for netdev \"%s\"", ifname);
1627 			continue;
1628 		}
1629 		if (pci_dev->addr.domain != pci_addr.domain ||
1630 		    pci_dev->addr.bus != pci_addr.bus ||
1631 		    pci_dev->addr.devid != pci_addr.devid ||
1632 		    pci_dev->addr.function != pci_addr.function)
1633 			continue;
1634 		/* Slave interface PCI address match found. */
1635 		fclose(file);
1636 		snprintf(tmp_str, sizeof(tmp_str),
1637 			 "/sys/class/net/%s/phys_port_name", ifname);
1638 		file = fopen(tmp_str, "rb");
1639 		if (!file)
1640 			break;
1641 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1642 		if (fscanf(file, "%32s", tmp_str) == 1)
1643 			mlx5_translate_port_name(tmp_str, &info);
1644 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1645 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1646 			pf = info.port_name;
1647 		break;
1648 	}
1649 	if (file)
1650 		fclose(file);
1651 	return pf;
1652 }
1653 
1654 /**
1655  * DPDK callback to register a PCI device.
1656  *
1657  * This function spawns Ethernet devices out of a given PCI device.
1658  *
1659  * @param[in] pci_drv
1660  *   PCI driver structure (mlx5_driver).
1661  * @param[in] pci_dev
1662  *   PCI device information.
1663  *
1664  * @return
1665  *   0 on success, a negative errno value otherwise and rte_errno is set.
1666  */
1667 int
1668 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1669 		  struct rte_pci_device *pci_dev)
1670 {
1671 	struct ibv_device **ibv_list;
1672 	/*
1673 	 * Number of found IB Devices matching with requested PCI BDF.
1674 	 * nd != 1 means there are multiple IB devices over the same
1675 	 * PCI device and we have representors and master.
1676 	 */
1677 	unsigned int nd = 0;
1678 	/*
1679 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1680 	 * we have the single multiport IB device, and there may be
1681 	 * representors attached to some of found ports.
1682 	 */
1683 	unsigned int np = 0;
1684 	/*
1685 	 * Number of DPDK ethernet devices to Spawn - either over
1686 	 * multiple IB devices or multiple ports of single IB device.
1687 	 * Actually this is the number of iterations to spawn.
1688 	 */
1689 	unsigned int ns = 0;
1690 	/*
1691 	 * Bonding device
1692 	 *   < 0 - no bonding device (single one)
1693 	 *  >= 0 - bonding device (value is slave PF index)
1694 	 */
1695 	int bd = -1;
1696 	struct mlx5_dev_spawn_data *list = NULL;
1697 	struct mlx5_dev_config dev_config;
1698 	unsigned int dev_config_vf;
1699 	int ret;
1700 
1701 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1702 		mlx5_pmd_socket_init();
1703 	ret = mlx5_init_once();
1704 	if (ret) {
1705 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1706 			strerror(rte_errno));
1707 		return -rte_errno;
1708 	}
1709 	errno = 0;
1710 	ibv_list = mlx5_glue->get_device_list(&ret);
1711 	if (!ibv_list) {
1712 		rte_errno = errno ? errno : ENOSYS;
1713 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1714 		return -rte_errno;
1715 	}
1716 	/*
1717 	 * First scan the list of all Infiniband devices to find
1718 	 * matching ones, gathering into the list.
1719 	 */
1720 	struct ibv_device *ibv_match[ret + 1];
1721 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1722 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1723 	unsigned int i;
1724 
1725 	while (ret-- > 0) {
1726 		struct rte_pci_addr pci_addr;
1727 
1728 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1729 		bd = mlx5_device_bond_pci_match
1730 				(ibv_list[ret], pci_dev, nl_rdma);
1731 		if (bd >= 0) {
1732 			/*
1733 			 * Bonding device detected. Only one match is allowed,
1734 			 * the bonding is supported over multi-port IB device,
1735 			 * there should be no matches on representor PCI
1736 			 * functions or non VF LAG bonding devices with
1737 			 * specified address.
1738 			 */
1739 			if (nd) {
1740 				DRV_LOG(ERR,
1741 					"multiple PCI match on bonding device"
1742 					"\"%s\" found", ibv_list[ret]->name);
1743 				rte_errno = ENOENT;
1744 				ret = -rte_errno;
1745 				goto exit;
1746 			}
1747 			DRV_LOG(INFO, "PCI information matches for"
1748 				      " slave %d bonding device \"%s\"",
1749 				      bd, ibv_list[ret]->name);
1750 			ibv_match[nd++] = ibv_list[ret];
1751 			break;
1752 		}
1753 		if (mlx5_dev_to_pci_addr
1754 			(ibv_list[ret]->ibdev_path, &pci_addr))
1755 			continue;
1756 		if (pci_dev->addr.domain != pci_addr.domain ||
1757 		    pci_dev->addr.bus != pci_addr.bus ||
1758 		    pci_dev->addr.devid != pci_addr.devid ||
1759 		    pci_dev->addr.function != pci_addr.function)
1760 			continue;
1761 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1762 			ibv_list[ret]->name);
1763 		ibv_match[nd++] = ibv_list[ret];
1764 	}
1765 	ibv_match[nd] = NULL;
1766 	if (!nd) {
1767 		/* No device matches, just complain and bail out. */
1768 		DRV_LOG(WARNING,
1769 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1770 			" are kernel drivers loaded?",
1771 			pci_dev->addr.domain, pci_dev->addr.bus,
1772 			pci_dev->addr.devid, pci_dev->addr.function);
1773 		rte_errno = ENOENT;
1774 		ret = -rte_errno;
1775 		goto exit;
1776 	}
1777 	if (nd == 1) {
1778 		/*
1779 		 * Found single matching device may have multiple ports.
1780 		 * Each port may be representor, we have to check the port
1781 		 * number and check the representors existence.
1782 		 */
1783 		if (nl_rdma >= 0)
1784 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1785 		if (!np)
1786 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1787 					 " ports number", ibv_match[0]->name);
1788 		if (bd >= 0 && !np) {
1789 			DRV_LOG(ERR, "can not get ports"
1790 				     " for bonding device");
1791 			rte_errno = ENOENT;
1792 			ret = -rte_errno;
1793 			goto exit;
1794 		}
1795 	}
1796 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1797 	if (bd >= 0) {
1798 		/*
1799 		 * This may happen if there is VF LAG kernel support and
1800 		 * application is compiled with older rdma_core library.
1801 		 */
1802 		DRV_LOG(ERR,
1803 			"No kernel/verbs support for VF LAG bonding found.");
1804 		rte_errno = ENOTSUP;
1805 		ret = -rte_errno;
1806 		goto exit;
1807 	}
1808 #endif
1809 	/*
1810 	 * Now we can determine the maximal
1811 	 * amount of devices to be spawned.
1812 	 */
1813 	list = mlx5_malloc(MLX5_MEM_ZERO,
1814 			   sizeof(struct mlx5_dev_spawn_data) *
1815 			   (np ? np : nd),
1816 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1817 	if (!list) {
1818 		DRV_LOG(ERR, "spawn data array allocation failure");
1819 		rte_errno = ENOMEM;
1820 		ret = -rte_errno;
1821 		goto exit;
1822 	}
1823 	if (bd >= 0 || np > 1) {
1824 		/*
1825 		 * Single IB device with multiple ports found,
1826 		 * it may be E-Switch master device and representors.
1827 		 * We have to perform identification through the ports.
1828 		 */
1829 		MLX5_ASSERT(nl_rdma >= 0);
1830 		MLX5_ASSERT(ns == 0);
1831 		MLX5_ASSERT(nd == 1);
1832 		MLX5_ASSERT(np);
1833 		for (i = 1; i <= np; ++i) {
1834 			list[ns].max_port = np;
1835 			list[ns].phys_port = i;
1836 			list[ns].phys_dev = ibv_match[0];
1837 			list[ns].eth_dev = NULL;
1838 			list[ns].pci_dev = pci_dev;
1839 			list[ns].pf_bond = bd;
1840 			list[ns].ifindex = mlx5_nl_ifindex
1841 				(nl_rdma,
1842 				mlx5_os_get_dev_device_name
1843 						(list[ns].phys_dev), i);
1844 			if (!list[ns].ifindex) {
1845 				/*
1846 				 * No network interface index found for the
1847 				 * specified port, it means there is no
1848 				 * representor on this port. It's OK,
1849 				 * there can be disabled ports, for example
1850 				 * if sriov_numvfs < sriov_totalvfs.
1851 				 */
1852 				continue;
1853 			}
1854 			ret = -1;
1855 			if (nl_route >= 0)
1856 				ret = mlx5_nl_switch_info
1857 					       (nl_route,
1858 						list[ns].ifindex,
1859 						&list[ns].info);
1860 			if (ret || (!list[ns].info.representor &&
1861 				    !list[ns].info.master)) {
1862 				/*
1863 				 * We failed to recognize representors with
1864 				 * Netlink, let's try to perform the task
1865 				 * with sysfs.
1866 				 */
1867 				ret =  mlx5_sysfs_switch_info
1868 						(list[ns].ifindex,
1869 						 &list[ns].info);
1870 			}
1871 			if (!ret && bd >= 0) {
1872 				switch (list[ns].info.name_type) {
1873 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1874 					if (list[ns].info.port_name == bd)
1875 						ns++;
1876 					break;
1877 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1878 					/* Fallthrough */
1879 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1880 					if (list[ns].info.pf_num == bd)
1881 						ns++;
1882 					break;
1883 				default:
1884 					break;
1885 				}
1886 				continue;
1887 			}
1888 			if (!ret && (list[ns].info.representor ^
1889 				     list[ns].info.master))
1890 				ns++;
1891 		}
1892 		if (!ns) {
1893 			DRV_LOG(ERR,
1894 				"unable to recognize master/representors"
1895 				" on the IB device with multiple ports");
1896 			rte_errno = ENOENT;
1897 			ret = -rte_errno;
1898 			goto exit;
1899 		}
1900 	} else {
1901 		/*
1902 		 * The existence of several matching entries (nd > 1) means
1903 		 * port representors have been instantiated. No existing Verbs
1904 		 * call nor sysfs entries can tell them apart, this can only
1905 		 * be done through Netlink calls assuming kernel drivers are
1906 		 * recent enough to support them.
1907 		 *
1908 		 * In the event of identification failure through Netlink,
1909 		 * try again through sysfs, then:
1910 		 *
1911 		 * 1. A single IB device matches (nd == 1) with single
1912 		 *    port (np=0/1) and is not a representor, assume
1913 		 *    no switch support.
1914 		 *
1915 		 * 2. Otherwise no safe assumptions can be made;
1916 		 *    complain louder and bail out.
1917 		 */
1918 		for (i = 0; i != nd; ++i) {
1919 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1920 			list[ns].max_port = 1;
1921 			list[ns].phys_port = 1;
1922 			list[ns].phys_dev = ibv_match[i];
1923 			list[ns].eth_dev = NULL;
1924 			list[ns].pci_dev = pci_dev;
1925 			list[ns].pf_bond = -1;
1926 			list[ns].ifindex = 0;
1927 			if (nl_rdma >= 0)
1928 				list[ns].ifindex = mlx5_nl_ifindex
1929 				(nl_rdma,
1930 				mlx5_os_get_dev_device_name
1931 						(list[ns].phys_dev), 1);
1932 			if (!list[ns].ifindex) {
1933 				char ifname[IF_NAMESIZE];
1934 
1935 				/*
1936 				 * Netlink failed, it may happen with old
1937 				 * ib_core kernel driver (before 4.16).
1938 				 * We can assume there is old driver because
1939 				 * here we are processing single ports IB
1940 				 * devices. Let's try sysfs to retrieve
1941 				 * the ifindex. The method works for
1942 				 * master device only.
1943 				 */
1944 				if (nd > 1) {
1945 					/*
1946 					 * Multiple devices found, assume
1947 					 * representors, can not distinguish
1948 					 * master/representor and retrieve
1949 					 * ifindex via sysfs.
1950 					 */
1951 					continue;
1952 				}
1953 				ret = mlx5_get_ifname_sysfs
1954 					(ibv_match[i]->ibdev_path, ifname);
1955 				if (!ret)
1956 					list[ns].ifindex =
1957 						if_nametoindex(ifname);
1958 				if (!list[ns].ifindex) {
1959 					/*
1960 					 * No network interface index found
1961 					 * for the specified device, it means
1962 					 * there it is neither representor
1963 					 * nor master.
1964 					 */
1965 					continue;
1966 				}
1967 			}
1968 			ret = -1;
1969 			if (nl_route >= 0)
1970 				ret = mlx5_nl_switch_info
1971 					       (nl_route,
1972 						list[ns].ifindex,
1973 						&list[ns].info);
1974 			if (ret || (!list[ns].info.representor &&
1975 				    !list[ns].info.master)) {
1976 				/*
1977 				 * We failed to recognize representors with
1978 				 * Netlink, let's try to perform the task
1979 				 * with sysfs.
1980 				 */
1981 				ret =  mlx5_sysfs_switch_info
1982 						(list[ns].ifindex,
1983 						 &list[ns].info);
1984 			}
1985 			if (!ret && (list[ns].info.representor ^
1986 				     list[ns].info.master)) {
1987 				ns++;
1988 			} else if ((nd == 1) &&
1989 				   !list[ns].info.representor &&
1990 				   !list[ns].info.master) {
1991 				/*
1992 				 * Single IB device with
1993 				 * one physical port and
1994 				 * attached network device.
1995 				 * May be SRIOV is not enabled
1996 				 * or there is no representors.
1997 				 */
1998 				DRV_LOG(INFO, "no E-Switch support detected");
1999 				ns++;
2000 				break;
2001 			}
2002 		}
2003 		if (!ns) {
2004 			DRV_LOG(ERR,
2005 				"unable to recognize master/representors"
2006 				" on the multiple IB devices");
2007 			rte_errno = ENOENT;
2008 			ret = -rte_errno;
2009 			goto exit;
2010 		}
2011 	}
2012 	MLX5_ASSERT(ns);
2013 	/*
2014 	 * Sort list to probe devices in natural order for users convenience
2015 	 * (i.e. master first, then representors from lowest to highest ID).
2016 	 */
2017 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2018 	/* Device specific configuration. */
2019 	switch (pci_dev->id.device_id) {
2020 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2021 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2022 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2023 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2024 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2025 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2026 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2027 		dev_config_vf = 1;
2028 		break;
2029 	default:
2030 		dev_config_vf = 0;
2031 		break;
2032 	}
2033 	for (i = 0; i != ns; ++i) {
2034 		uint32_t restore;
2035 
2036 		/* Default configuration. */
2037 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2038 		dev_config.vf = dev_config_vf;
2039 		dev_config.mps = MLX5_ARG_UNSET;
2040 		dev_config.dbnc = MLX5_ARG_UNSET;
2041 		dev_config.rx_vec_en = 1;
2042 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2043 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2044 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2045 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2046 		dev_config.vf_nl_en = 1;
2047 		dev_config.mr_ext_memseg_en = 1;
2048 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2049 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2050 		dev_config.dv_esw_en = 1;
2051 		dev_config.dv_flow_en = 1;
2052 		dev_config.decap_en = 1;
2053 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2054 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2055 						 &list[i],
2056 						 &dev_config);
2057 		if (!list[i].eth_dev) {
2058 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2059 				break;
2060 			/* Device is disabled or already spawned. Ignore it. */
2061 			continue;
2062 		}
2063 		restore = list[i].eth_dev->data->dev_flags;
2064 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2065 		/* Restore non-PCI flags cleared by the above call. */
2066 		list[i].eth_dev->data->dev_flags |= restore;
2067 		rte_eth_dev_probing_finish(list[i].eth_dev);
2068 	}
2069 	if (i != ns) {
2070 		DRV_LOG(ERR,
2071 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2072 			" encountering an error: %s",
2073 			pci_dev->addr.domain, pci_dev->addr.bus,
2074 			pci_dev->addr.devid, pci_dev->addr.function,
2075 			strerror(rte_errno));
2076 		ret = -rte_errno;
2077 		/* Roll back. */
2078 		while (i--) {
2079 			if (!list[i].eth_dev)
2080 				continue;
2081 			mlx5_dev_close(list[i].eth_dev);
2082 			/* mac_addrs must not be freed because in dev_private */
2083 			list[i].eth_dev->data->mac_addrs = NULL;
2084 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2085 		}
2086 		/* Restore original error. */
2087 		rte_errno = -ret;
2088 	} else {
2089 		ret = 0;
2090 	}
2091 exit:
2092 	/*
2093 	 * Do the routine cleanup:
2094 	 * - close opened Netlink sockets
2095 	 * - free allocated spawn data array
2096 	 * - free the Infiniband device list
2097 	 */
2098 	if (nl_rdma >= 0)
2099 		close(nl_rdma);
2100 	if (nl_route >= 0)
2101 		close(nl_route);
2102 	if (list)
2103 		mlx5_free(list);
2104 	MLX5_ASSERT(ibv_list);
2105 	mlx5_glue->free_device_list(ibv_list);
2106 	return ret;
2107 }
2108 
2109 static int
2110 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2111 {
2112 	char *env;
2113 	int value;
2114 
2115 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2116 	/* Get environment variable to store. */
2117 	env = getenv(MLX5_SHUT_UP_BF);
2118 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2119 	if (config->dbnc == MLX5_ARG_UNSET)
2120 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2121 	else
2122 		setenv(MLX5_SHUT_UP_BF,
2123 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2124 	return value;
2125 }
2126 
2127 static void
2128 mlx5_restore_doorbell_mapping_env(int value)
2129 {
2130 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2131 	/* Restore the original environment variable state. */
2132 	if (value == MLX5_ARG_UNSET)
2133 		unsetenv(MLX5_SHUT_UP_BF);
2134 	else
2135 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2136 }
2137 
2138 /**
2139  * Extract pdn of PD object using DV API.
2140  *
2141  * @param[in] pd
2142  *   Pointer to the verbs PD object.
2143  * @param[out] pdn
2144  *   Pointer to the PD object number variable.
2145  *
2146  * @return
2147  *   0 on success, error value otherwise.
2148  */
2149 int
2150 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2151 {
2152 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2153 	struct mlx5dv_obj obj;
2154 	struct mlx5dv_pd pd_info;
2155 	int ret = 0;
2156 
2157 	obj.pd.in = pd;
2158 	obj.pd.out = &pd_info;
2159 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2160 	if (ret) {
2161 		DRV_LOG(DEBUG, "Fail to get PD object info");
2162 		return ret;
2163 	}
2164 	*pdn = pd_info.pdn;
2165 	return 0;
2166 #else
2167 	(void)pd;
2168 	(void)pdn;
2169 	return -ENOTSUP;
2170 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2171 }
2172 
2173 /**
2174  * Function API to open IB device.
2175  *
2176  * This function calls the Linux glue APIs to open a device.
2177  *
2178  * @param[in] spawn
2179  *   Pointer to the IB device attributes (name, port, etc).
2180  * @param[out] config
2181  *   Pointer to device configuration structure.
2182  * @param[out] sh
2183  *   Pointer to shared context structure.
2184  *
2185  * @return
2186  *   0 on success, a positive error value otherwise.
2187  */
2188 int
2189 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2190 		     const struct mlx5_dev_config *config,
2191 		     struct mlx5_dev_ctx_shared *sh)
2192 {
2193 	int dbmap_env;
2194 	int err = 0;
2195 
2196 	sh->numa_node = spawn->pci_dev->device.numa_node;
2197 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2198 	/*
2199 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2200 	 * before the device creation. The rdma_core library
2201 	 * checks the variable at device creation and
2202 	 * stores the result internally.
2203 	 */
2204 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2205 	/* Try to open IB device with DV first, then usual Verbs. */
2206 	errno = 0;
2207 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2208 	if (sh->ctx) {
2209 		sh->devx = 1;
2210 		DRV_LOG(DEBUG, "DevX is supported");
2211 		/* The device is created, no need for environment. */
2212 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2213 	} else {
2214 		/* The environment variable is still configured. */
2215 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2216 		err = errno ? errno : ENODEV;
2217 		/*
2218 		 * The environment variable is not needed anymore,
2219 		 * all device creation attempts are completed.
2220 		 */
2221 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2222 		if (!sh->ctx)
2223 			return err;
2224 		DRV_LOG(DEBUG, "DevX is NOT supported");
2225 		err = 0;
2226 	}
2227 	return err;
2228 }
2229 
2230 /**
2231  * Install shared asynchronous device events handler.
2232  * This function is implemented to support event sharing
2233  * between multiple ports of single IB device.
2234  *
2235  * @param sh
2236  *   Pointer to mlx5_dev_ctx_shared object.
2237  */
2238 void
2239 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2240 {
2241 	int ret;
2242 	int flags;
2243 
2244 	sh->intr_handle.fd = -1;
2245 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2246 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2247 		    F_SETFL, flags | O_NONBLOCK);
2248 	if (ret) {
2249 		DRV_LOG(INFO, "failed to change file descriptor async event"
2250 			" queue");
2251 	} else {
2252 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2253 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2254 		if (rte_intr_callback_register(&sh->intr_handle,
2255 					mlx5_dev_interrupt_handler, sh)) {
2256 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2257 			sh->intr_handle.fd = -1;
2258 		}
2259 	}
2260 	if (sh->devx) {
2261 #ifdef HAVE_IBV_DEVX_ASYNC
2262 		sh->intr_handle_devx.fd = -1;
2263 		sh->devx_comp =
2264 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2265 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2266 		if (!devx_comp) {
2267 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2268 			return;
2269 		}
2270 		flags = fcntl(devx_comp->fd, F_GETFL);
2271 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2272 		if (ret) {
2273 			DRV_LOG(INFO, "failed to change file descriptor"
2274 				" devx comp");
2275 			return;
2276 		}
2277 		sh->intr_handle_devx.fd = devx_comp->fd;
2278 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2279 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2280 					mlx5_dev_interrupt_handler_devx, sh)) {
2281 			DRV_LOG(INFO, "Fail to install the devx shared"
2282 				" interrupt.");
2283 			sh->intr_handle_devx.fd = -1;
2284 		}
2285 #endif /* HAVE_IBV_DEVX_ASYNC */
2286 	}
2287 }
2288 
2289 /**
2290  * Uninstall shared asynchronous device events handler.
2291  * This function is implemented to support event sharing
2292  * between multiple ports of single IB device.
2293  *
2294  * @param dev
2295  *   Pointer to mlx5_dev_ctx_shared object.
2296  */
2297 void
2298 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2299 {
2300 	if (sh->intr_handle.fd >= 0)
2301 		mlx5_intr_callback_unregister(&sh->intr_handle,
2302 					      mlx5_dev_interrupt_handler, sh);
2303 #ifdef HAVE_IBV_DEVX_ASYNC
2304 	if (sh->intr_handle_devx.fd >= 0)
2305 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2306 				  mlx5_dev_interrupt_handler_devx, sh);
2307 	if (sh->devx_comp)
2308 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2309 #endif
2310 }
2311 
2312 /**
2313  * Read statistics by a named counter.
2314  *
2315  * @param[in] priv
2316  *   Pointer to the private device data structure.
2317  * @param[in] ctr_name
2318  *   Pointer to the name of the statistic counter to read
2319  * @param[out] stat
2320  *   Pointer to read statistic value.
2321  * @return
2322  *   0 on success and stat is valud, 1 if failed to read the value
2323  *   rte_errno is set.
2324  *
2325  */
2326 int
2327 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2328 		      uint64_t *stat)
2329 {
2330 	int fd;
2331 
2332 	if (priv->sh) {
2333 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2334 		      priv->sh->ibdev_path,
2335 		      priv->dev_port,
2336 		      ctr_name);
2337 		fd = open(path, O_RDONLY);
2338 		/*
2339 		 * in switchdev the file location is not per port
2340 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2341 		 */
2342 		if (fd == -1) {
2343 			MKSTR(path1, "%s/hw_counters/%s",
2344 			      priv->sh->ibdev_path,
2345 			      ctr_name);
2346 			fd = open(path1, O_RDONLY);
2347 		}
2348 		if (fd != -1) {
2349 			char buf[21] = {'\0'};
2350 			ssize_t n = read(fd, buf, sizeof(buf));
2351 
2352 			close(fd);
2353 			if (n != -1) {
2354 				*stat = strtoull(buf, NULL, 10);
2355 				return 0;
2356 			}
2357 		}
2358 	}
2359 	*stat = 0;
2360 	return 1;
2361 }
2362 
2363 /**
2364  * Set the reg_mr and dereg_mr call backs
2365  *
2366  * @param reg_mr_cb[out]
2367  *   Pointer to reg_mr func
2368  * @param dereg_mr_cb[out]
2369  *   Pointer to dereg_mr func
2370  *
2371  */
2372 void
2373 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2374 		      mlx5_dereg_mr_t *dereg_mr_cb)
2375 {
2376 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2377 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2378 }
2379 
2380 /**
2381  * Remove a MAC address from device
2382  *
2383  * @param dev
2384  *   Pointer to Ethernet device structure.
2385  * @param index
2386  *   MAC address index.
2387  */
2388 void
2389 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2390 {
2391 	struct mlx5_priv *priv = dev->data->dev_private;
2392 	const int vf = priv->config.vf;
2393 
2394 	if (vf)
2395 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2396 					mlx5_ifindex(dev), priv->mac_own,
2397 					&dev->data->mac_addrs[index], index);
2398 }
2399 
2400 /**
2401  * Adds a MAC address to the device
2402  *
2403  * @param dev
2404  *   Pointer to Ethernet device structure.
2405  * @param mac_addr
2406  *   MAC address to register.
2407  * @param index
2408  *   MAC address index.
2409  *
2410  * @return
2411  *   0 on success, a negative errno value otherwise
2412  */
2413 int
2414 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2415 		     uint32_t index)
2416 {
2417 	struct mlx5_priv *priv = dev->data->dev_private;
2418 	const int vf = priv->config.vf;
2419 	int ret = 0;
2420 
2421 	if (vf)
2422 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2423 					   mlx5_ifindex(dev), priv->mac_own,
2424 					   mac, index);
2425 	return ret;
2426 }
2427 
2428 /**
2429  * Modify a VF MAC address
2430  *
2431  * @param priv
2432  *   Pointer to device private data.
2433  * @param mac_addr
2434  *   MAC address to modify into.
2435  * @param iface_idx
2436  *   Net device interface index
2437  * @param vf_index
2438  *   VF index
2439  *
2440  * @return
2441  *   0 on success, a negative errno value otherwise
2442  */
2443 int
2444 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2445 			   unsigned int iface_idx,
2446 			   struct rte_ether_addr *mac_addr,
2447 			   int vf_index)
2448 {
2449 	return mlx5_nl_vf_mac_addr_modify
2450 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2451 }
2452 
2453 /**
2454  * Set device promiscuous mode
2455  *
2456  * @param dev
2457  *   Pointer to Ethernet device structure.
2458  * @param enable
2459  *   0 - promiscuous is disabled, otherwise - enabled
2460  *
2461  * @return
2462  *   0 on success, a negative error value otherwise
2463  */
2464 int
2465 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2466 {
2467 	struct mlx5_priv *priv = dev->data->dev_private;
2468 
2469 	return mlx5_nl_promisc(priv->nl_socket_route,
2470 			       mlx5_ifindex(dev), !!enable);
2471 }
2472 
2473 /**
2474  * Set device promiscuous mode
2475  *
2476  * @param dev
2477  *   Pointer to Ethernet device structure.
2478  * @param enable
2479  *   0 - all multicase is disabled, otherwise - enabled
2480  *
2481  * @return
2482  *   0 on success, a negative error value otherwise
2483  */
2484 int
2485 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2486 {
2487 	struct mlx5_priv *priv = dev->data->dev_private;
2488 
2489 	return mlx5_nl_allmulti(priv->nl_socket_route,
2490 				mlx5_ifindex(dev), !!enable);
2491 }
2492 
2493 /**
2494  * Flush device MAC addresses
2495  *
2496  * @param dev
2497  *   Pointer to Ethernet device structure.
2498  *
2499  */
2500 void
2501 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2502 {
2503 	struct mlx5_priv *priv = dev->data->dev_private;
2504 
2505 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2506 			       dev->data->mac_addrs,
2507 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2508 }
2509 
2510 const struct eth_dev_ops mlx5_os_dev_ops = {
2511 	.dev_configure = mlx5_dev_configure,
2512 	.dev_start = mlx5_dev_start,
2513 	.dev_stop = mlx5_dev_stop,
2514 	.dev_set_link_down = mlx5_set_link_down,
2515 	.dev_set_link_up = mlx5_set_link_up,
2516 	.dev_close = mlx5_dev_close,
2517 	.promiscuous_enable = mlx5_promiscuous_enable,
2518 	.promiscuous_disable = mlx5_promiscuous_disable,
2519 	.allmulticast_enable = mlx5_allmulticast_enable,
2520 	.allmulticast_disable = mlx5_allmulticast_disable,
2521 	.link_update = mlx5_link_update,
2522 	.stats_get = mlx5_stats_get,
2523 	.stats_reset = mlx5_stats_reset,
2524 	.xstats_get = mlx5_xstats_get,
2525 	.xstats_reset = mlx5_xstats_reset,
2526 	.xstats_get_names = mlx5_xstats_get_names,
2527 	.fw_version_get = mlx5_fw_version_get,
2528 	.dev_infos_get = mlx5_dev_infos_get,
2529 	.read_clock = mlx5_txpp_read_clock,
2530 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2531 	.vlan_filter_set = mlx5_vlan_filter_set,
2532 	.rx_queue_setup = mlx5_rx_queue_setup,
2533 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2534 	.tx_queue_setup = mlx5_tx_queue_setup,
2535 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2536 	.rx_queue_release = mlx5_rx_queue_release,
2537 	.tx_queue_release = mlx5_tx_queue_release,
2538 	.rx_queue_start = mlx5_rx_queue_start,
2539 	.rx_queue_stop = mlx5_rx_queue_stop,
2540 	.tx_queue_start = mlx5_tx_queue_start,
2541 	.tx_queue_stop = mlx5_tx_queue_stop,
2542 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2543 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2544 	.mac_addr_remove = mlx5_mac_addr_remove,
2545 	.mac_addr_add = mlx5_mac_addr_add,
2546 	.mac_addr_set = mlx5_mac_addr_set,
2547 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2548 	.mtu_set = mlx5_dev_set_mtu,
2549 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2550 	.vlan_offload_set = mlx5_vlan_offload_set,
2551 	.reta_update = mlx5_dev_rss_reta_update,
2552 	.reta_query = mlx5_dev_rss_reta_query,
2553 	.rss_hash_update = mlx5_rss_hash_update,
2554 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2555 	.filter_ctrl = mlx5_dev_filter_ctrl,
2556 	.rxq_info_get = mlx5_rxq_info_get,
2557 	.txq_info_get = mlx5_txq_info_get,
2558 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2559 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2560 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2561 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2562 	.is_removed = mlx5_is_removed,
2563 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2564 	.get_module_info = mlx5_get_module_info,
2565 	.get_module_eeprom = mlx5_get_module_eeprom,
2566 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2567 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2568 	.hairpin_bind = mlx5_hairpin_bind,
2569 	.hairpin_unbind = mlx5_hairpin_unbind,
2570 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2571 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2572 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2573 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2574 };
2575 
2576 /* Available operations from secondary process. */
2577 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2578 	.stats_get = mlx5_stats_get,
2579 	.stats_reset = mlx5_stats_reset,
2580 	.xstats_get = mlx5_xstats_get,
2581 	.xstats_reset = mlx5_xstats_reset,
2582 	.xstats_get_names = mlx5_xstats_get_names,
2583 	.fw_version_get = mlx5_fw_version_get,
2584 	.dev_infos_get = mlx5_dev_infos_get,
2585 	.read_clock = mlx5_txpp_read_clock,
2586 	.rx_queue_start = mlx5_rx_queue_start,
2587 	.rx_queue_stop = mlx5_rx_queue_stop,
2588 	.tx_queue_start = mlx5_tx_queue_start,
2589 	.tx_queue_stop = mlx5_tx_queue_stop,
2590 	.rxq_info_get = mlx5_rxq_info_get,
2591 	.txq_info_get = mlx5_txq_info_get,
2592 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2593 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2594 	.get_module_info = mlx5_get_module_info,
2595 	.get_module_eeprom = mlx5_get_module_eeprom,
2596 };
2597 
2598 /* Available operations in flow isolated mode. */
2599 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2600 	.dev_configure = mlx5_dev_configure,
2601 	.dev_start = mlx5_dev_start,
2602 	.dev_stop = mlx5_dev_stop,
2603 	.dev_set_link_down = mlx5_set_link_down,
2604 	.dev_set_link_up = mlx5_set_link_up,
2605 	.dev_close = mlx5_dev_close,
2606 	.promiscuous_enable = mlx5_promiscuous_enable,
2607 	.promiscuous_disable = mlx5_promiscuous_disable,
2608 	.allmulticast_enable = mlx5_allmulticast_enable,
2609 	.allmulticast_disable = mlx5_allmulticast_disable,
2610 	.link_update = mlx5_link_update,
2611 	.stats_get = mlx5_stats_get,
2612 	.stats_reset = mlx5_stats_reset,
2613 	.xstats_get = mlx5_xstats_get,
2614 	.xstats_reset = mlx5_xstats_reset,
2615 	.xstats_get_names = mlx5_xstats_get_names,
2616 	.fw_version_get = mlx5_fw_version_get,
2617 	.dev_infos_get = mlx5_dev_infos_get,
2618 	.read_clock = mlx5_txpp_read_clock,
2619 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2620 	.vlan_filter_set = mlx5_vlan_filter_set,
2621 	.rx_queue_setup = mlx5_rx_queue_setup,
2622 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2623 	.tx_queue_setup = mlx5_tx_queue_setup,
2624 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2625 	.rx_queue_release = mlx5_rx_queue_release,
2626 	.tx_queue_release = mlx5_tx_queue_release,
2627 	.rx_queue_start = mlx5_rx_queue_start,
2628 	.rx_queue_stop = mlx5_rx_queue_stop,
2629 	.tx_queue_start = mlx5_tx_queue_start,
2630 	.tx_queue_stop = mlx5_tx_queue_stop,
2631 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2632 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2633 	.mac_addr_remove = mlx5_mac_addr_remove,
2634 	.mac_addr_add = mlx5_mac_addr_add,
2635 	.mac_addr_set = mlx5_mac_addr_set,
2636 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2637 	.mtu_set = mlx5_dev_set_mtu,
2638 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2639 	.vlan_offload_set = mlx5_vlan_offload_set,
2640 	.filter_ctrl = mlx5_dev_filter_ctrl,
2641 	.rxq_info_get = mlx5_rxq_info_get,
2642 	.txq_info_get = mlx5_txq_info_get,
2643 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2644 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2645 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2646 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2647 	.is_removed = mlx5_is_removed,
2648 	.get_module_info = mlx5_get_module_info,
2649 	.get_module_eeprom = mlx5_get_module_eeprom,
2650 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2651 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2652 	.hairpin_bind = mlx5_hairpin_bind,
2653 	.hairpin_unbind = mlx5_hairpin_unbind,
2654 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2655 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2656 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2657 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2658 };
2659