xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 16a7dbc4f69006cc1c96ca2a2c6d3e3c51a2ff50)
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] __rte_unused;
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 		goto error;
237 	/* The resources below are only valid with DV support. */
238 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
239 	/* Create tags hash list table. */
240 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
241 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
242 					  MLX5_HLIST_WRITE_MOST,
243 					  flow_dv_tag_create_cb, NULL,
244 					  flow_dv_tag_remove_cb);
245 	if (!sh->tag_table) {
246 		DRV_LOG(ERR, "tags with hash creation failed.");
247 		err = ENOMEM;
248 		goto error;
249 	}
250 	sh->tag_table->ctx = sh;
251 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
252 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
253 					    0, MLX5_HLIST_WRITE_MOST |
254 					    MLX5_HLIST_DIRECT_KEY,
255 					    flow_dv_modify_create_cb,
256 					    flow_dv_modify_match_cb,
257 					    flow_dv_modify_remove_cb);
258 	if (!sh->modify_cmds) {
259 		DRV_LOG(ERR, "hdr modify hash creation failed");
260 		err = ENOMEM;
261 		goto error;
262 	}
263 	sh->modify_cmds->ctx = sh;
264 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
265 	sh->encaps_decaps = mlx5_hlist_create(s,
266 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
267 					      0, 0, NULL, NULL, NULL);
268 	if (!sh->encaps_decaps) {
269 		DRV_LOG(ERR, "encap decap hash creation failed");
270 		err = ENOMEM;
271 		goto error;
272 	}
273 #endif
274 #ifdef HAVE_MLX5DV_DR
275 	void *domain;
276 
277 	/* Reference counter is zero, we should initialize structures. */
278 	domain = mlx5_glue->dr_create_domain(sh->ctx,
279 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
280 	if (!domain) {
281 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
282 		err = errno;
283 		goto error;
284 	}
285 	sh->rx_domain = domain;
286 	domain = mlx5_glue->dr_create_domain(sh->ctx,
287 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
288 	if (!domain) {
289 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
290 		err = errno;
291 		goto error;
292 	}
293 	pthread_mutex_init(&sh->dv_mutex, NULL);
294 	sh->tx_domain = domain;
295 #ifdef HAVE_MLX5DV_DR_ESWITCH
296 	if (priv->config.dv_esw_en) {
297 		domain  = mlx5_glue->dr_create_domain
298 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
299 		if (!domain) {
300 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
301 			err = errno;
302 			goto error;
303 		}
304 		sh->fdb_domain = domain;
305 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
306 	}
307 #endif
308 	if (!sh->tunnel_hub)
309 		err = mlx5_alloc_tunnel_hub(sh);
310 	if (err) {
311 		DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
312 		goto error;
313 	}
314 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
315 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
316 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
317 		if (sh->fdb_domain)
318 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
319 	}
320 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
321 #endif /* HAVE_MLX5DV_DR */
322 	sh->default_miss_action =
323 			mlx5_glue->dr_create_flow_action_default_miss();
324 	if (!sh->default_miss_action)
325 		DRV_LOG(WARNING, "Default miss action is not supported.");
326 	return 0;
327 error:
328 	/* Rollback the created objects. */
329 	if (sh->rx_domain) {
330 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
331 		sh->rx_domain = NULL;
332 	}
333 	if (sh->tx_domain) {
334 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
335 		sh->tx_domain = NULL;
336 	}
337 	if (sh->fdb_domain) {
338 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
339 		sh->fdb_domain = NULL;
340 	}
341 	if (sh->esw_drop_action) {
342 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
343 		sh->esw_drop_action = NULL;
344 	}
345 	if (sh->pop_vlan_action) {
346 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
347 		sh->pop_vlan_action = NULL;
348 	}
349 	if (sh->encaps_decaps) {
350 		mlx5_hlist_destroy(sh->encaps_decaps);
351 		sh->encaps_decaps = NULL;
352 	}
353 	if (sh->modify_cmds) {
354 		mlx5_hlist_destroy(sh->modify_cmds);
355 		sh->modify_cmds = NULL;
356 	}
357 	if (sh->tag_table) {
358 		/* tags should be destroyed with flow before. */
359 		mlx5_hlist_destroy(sh->tag_table);
360 		sh->tag_table = NULL;
361 	}
362 	if (sh->tunnel_hub) {
363 		mlx5_release_tunnel_hub(sh, priv->dev_port);
364 		sh->tunnel_hub = NULL;
365 	}
366 	mlx5_free_table_hash_list(priv);
367 	return err;
368 }
369 
370 /**
371  * Destroy DR related data within private structure.
372  *
373  * @param[in] priv
374  *   Pointer to the private device data structure.
375  */
376 void
377 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
378 {
379 	struct mlx5_dev_ctx_shared *sh = priv->sh;
380 
381 	MLX5_ASSERT(sh && sh->refcnt);
382 	if (sh->refcnt > 1)
383 		return;
384 #ifdef HAVE_MLX5DV_DR
385 	if (sh->rx_domain) {
386 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
387 		sh->rx_domain = NULL;
388 	}
389 	if (sh->tx_domain) {
390 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
391 		sh->tx_domain = NULL;
392 	}
393 #ifdef HAVE_MLX5DV_DR_ESWITCH
394 	if (sh->fdb_domain) {
395 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
396 		sh->fdb_domain = NULL;
397 	}
398 	if (sh->esw_drop_action) {
399 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
400 		sh->esw_drop_action = NULL;
401 	}
402 #endif
403 	if (sh->pop_vlan_action) {
404 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
405 		sh->pop_vlan_action = NULL;
406 	}
407 	pthread_mutex_destroy(&sh->dv_mutex);
408 #endif /* HAVE_MLX5DV_DR */
409 	if (sh->default_miss_action)
410 		mlx5_glue->destroy_flow_action
411 				(sh->default_miss_action);
412 	if (sh->encaps_decaps) {
413 		mlx5_hlist_destroy(sh->encaps_decaps);
414 		sh->encaps_decaps = NULL;
415 	}
416 	if (sh->modify_cmds) {
417 		mlx5_hlist_destroy(sh->modify_cmds);
418 		sh->modify_cmds = NULL;
419 	}
420 	if (sh->tag_table) {
421 		/* tags should be destroyed with flow before. */
422 		mlx5_hlist_destroy(sh->tag_table);
423 		sh->tag_table = NULL;
424 	}
425 	if (sh->tunnel_hub) {
426 		mlx5_release_tunnel_hub(sh, priv->dev_port);
427 		sh->tunnel_hub = NULL;
428 	}
429 	mlx5_free_table_hash_list(priv);
430 }
431 
432 /**
433  * Initialize shared data between primary and secondary process.
434  *
435  * A memzone is reserved by primary process and secondary processes attach to
436  * the memzone.
437  *
438  * @return
439  *   0 on success, a negative errno value otherwise and rte_errno is set.
440  */
441 static int
442 mlx5_init_shared_data(void)
443 {
444 	const struct rte_memzone *mz;
445 	int ret = 0;
446 
447 	rte_spinlock_lock(&mlx5_shared_data_lock);
448 	if (mlx5_shared_data == NULL) {
449 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
450 			/* Allocate shared memory. */
451 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
452 						 sizeof(*mlx5_shared_data),
453 						 SOCKET_ID_ANY, 0);
454 			if (mz == NULL) {
455 				DRV_LOG(ERR,
456 					"Cannot allocate mlx5 shared data");
457 				ret = -rte_errno;
458 				goto error;
459 			}
460 			mlx5_shared_data = mz->addr;
461 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
462 			rte_spinlock_init(&mlx5_shared_data->lock);
463 		} else {
464 			/* Lookup allocated shared memory. */
465 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
466 			if (mz == NULL) {
467 				DRV_LOG(ERR,
468 					"Cannot attach mlx5 shared data");
469 				ret = -rte_errno;
470 				goto error;
471 			}
472 			mlx5_shared_data = mz->addr;
473 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
474 		}
475 	}
476 error:
477 	rte_spinlock_unlock(&mlx5_shared_data_lock);
478 	return ret;
479 }
480 
481 /**
482  * PMD global initialization.
483  *
484  * Independent from individual device, this function initializes global
485  * per-PMD data structures distinguishing primary and secondary processes.
486  * Hence, each initialization is called once per a process.
487  *
488  * @return
489  *   0 on success, a negative errno value otherwise and rte_errno is set.
490  */
491 static int
492 mlx5_init_once(void)
493 {
494 	struct mlx5_shared_data *sd;
495 	struct mlx5_local_data *ld = &mlx5_local_data;
496 	int ret = 0;
497 
498 	if (mlx5_init_shared_data())
499 		return -rte_errno;
500 	sd = mlx5_shared_data;
501 	MLX5_ASSERT(sd);
502 	rte_spinlock_lock(&sd->lock);
503 	switch (rte_eal_process_type()) {
504 	case RTE_PROC_PRIMARY:
505 		if (sd->init_done)
506 			break;
507 		LIST_INIT(&sd->mem_event_cb_list);
508 		rte_rwlock_init(&sd->mem_event_rwlock);
509 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
510 						mlx5_mr_mem_event_cb, NULL);
511 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
512 					   mlx5_mp_os_primary_handle);
513 		if (ret)
514 			goto out;
515 		sd->init_done = true;
516 		break;
517 	case RTE_PROC_SECONDARY:
518 		if (ld->init_done)
519 			break;
520 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
521 					     mlx5_mp_os_secondary_handle);
522 		if (ret)
523 			goto out;
524 		++sd->secondary_cnt;
525 		ld->init_done = true;
526 		break;
527 	default:
528 		break;
529 	}
530 out:
531 	rte_spinlock_unlock(&sd->lock);
532 	return ret;
533 }
534 
535 /**
536  * Create the Tx queue DevX/Verbs object.
537  *
538  * @param dev
539  *   Pointer to Ethernet device.
540  * @param idx
541  *   Queue index in DPDK Tx queue array.
542  *
543  * @return
544  *   0 on success, a negative errno value otherwise and rte_errno is set.
545  */
546 static int
547 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
548 {
549 	struct mlx5_priv *priv = dev->data->dev_private;
550 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
551 	struct mlx5_txq_ctrl *txq_ctrl =
552 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
553 
554 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
555 		return mlx5_txq_devx_obj_new(dev, idx);
556 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
557 	if (!priv->config.dv_esw_en)
558 		return mlx5_txq_devx_obj_new(dev, idx);
559 #endif
560 	return mlx5_txq_ibv_obj_new(dev, idx);
561 }
562 
563 /**
564  * Release an Tx DevX/verbs queue object.
565  *
566  * @param txq_obj
567  *   DevX/Verbs Tx queue object.
568  */
569 static void
570 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
571 {
572 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
573 		mlx5_txq_devx_obj_release(txq_obj);
574 		return;
575 	}
576 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
577 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
578 		mlx5_txq_devx_obj_release(txq_obj);
579 		return;
580 	}
581 #endif
582 	mlx5_txq_ibv_obj_release(txq_obj);
583 }
584 
585 /**
586  * DV flow counter mode detect and config.
587  *
588  * @param dev
589  *   Pointer to rte_eth_dev structure.
590  *
591  */
592 static void
593 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
594 {
595 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
596 	struct mlx5_priv *priv = dev->data->dev_private;
597 	struct mlx5_dev_ctx_shared *sh = priv->sh;
598 	bool fallback;
599 
600 #ifndef HAVE_IBV_DEVX_ASYNC
601 	fallback = true;
602 #else
603 	fallback = false;
604 	if (!priv->config.devx || !priv->config.dv_flow_en ||
605 	    !priv->config.hca_attr.flow_counters_dump ||
606 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
607 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
608 		fallback = true;
609 #endif
610 	if (fallback)
611 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
612 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
613 			priv->config.hca_attr.flow_counters_dump,
614 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
615 	/* Initialize fallback mode only on the port initializes sh. */
616 	if (sh->refcnt == 1)
617 		sh->cmng.counter_fallback = fallback;
618 	else if (fallback != sh->cmng.counter_fallback)
619 		DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
620 			"with others:%d.", PORT_ID(priv), fallback);
621 #endif
622 }
623 
624 /**
625  * Spawn an Ethernet device from Verbs information.
626  *
627  * @param dpdk_dev
628  *   Backing DPDK device.
629  * @param spawn
630  *   Verbs device parameters (name, port, switch_info) to spawn.
631  * @param config
632  *   Device configuration parameters.
633  *
634  * @return
635  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
636  *   is set. The following errors are defined:
637  *
638  *   EBUSY: device is not supposed to be spawned.
639  *   EEXIST: device is already spawned
640  */
641 static struct rte_eth_dev *
642 mlx5_dev_spawn(struct rte_device *dpdk_dev,
643 	       struct mlx5_dev_spawn_data *spawn,
644 	       struct mlx5_dev_config *config)
645 {
646 	const struct mlx5_switch_info *switch_info = &spawn->info;
647 	struct mlx5_dev_ctx_shared *sh = NULL;
648 	struct ibv_port_attr port_attr;
649 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
650 	struct rte_eth_dev *eth_dev = NULL;
651 	struct mlx5_priv *priv = NULL;
652 	int err = 0;
653 	unsigned int hw_padding = 0;
654 	unsigned int mps;
655 	unsigned int cqe_comp;
656 	unsigned int cqe_pad = 0;
657 	unsigned int tunnel_en = 0;
658 	unsigned int mpls_en = 0;
659 	unsigned int swp = 0;
660 	unsigned int mprq = 0;
661 	unsigned int mprq_min_stride_size_n = 0;
662 	unsigned int mprq_max_stride_size_n = 0;
663 	unsigned int mprq_min_stride_num_n = 0;
664 	unsigned int mprq_max_stride_num_n = 0;
665 	struct rte_ether_addr mac;
666 	char name[RTE_ETH_NAME_MAX_LEN];
667 	int own_domain_id = 0;
668 	uint16_t port_id;
669 	unsigned int i;
670 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
671 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
672 #endif
673 
674 	/* Determine if this port representor is supposed to be spawned. */
675 	if (switch_info->representor && dpdk_dev->devargs) {
676 		struct rte_eth_devargs eth_da;
677 
678 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
679 		if (err) {
680 			rte_errno = -err;
681 			DRV_LOG(ERR, "failed to process device arguments: %s",
682 				strerror(rte_errno));
683 			return NULL;
684 		}
685 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
686 			if (eth_da.representor_ports[i] ==
687 			    (uint16_t)switch_info->port_name)
688 				break;
689 		if (i == eth_da.nb_representor_ports) {
690 			rte_errno = EBUSY;
691 			return NULL;
692 		}
693 	}
694 	/* Build device name. */
695 	if (spawn->pf_bond <  0) {
696 		/* Single device. */
697 		if (!switch_info->representor)
698 			strlcpy(name, dpdk_dev->name, sizeof(name));
699 		else
700 			snprintf(name, sizeof(name), "%s_representor_%u",
701 				 dpdk_dev->name, switch_info->port_name);
702 	} else {
703 		/* Bonding device. */
704 		if (!switch_info->representor)
705 			snprintf(name, sizeof(name), "%s_%s",
706 				 dpdk_dev->name,
707 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
708 		else
709 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
710 				 dpdk_dev->name,
711 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
712 				 switch_info->port_name);
713 	}
714 	/* check if the device is already spawned */
715 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
716 		rte_errno = EEXIST;
717 		return NULL;
718 	}
719 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
720 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
721 		struct mlx5_mp_id mp_id;
722 
723 		eth_dev = rte_eth_dev_attach_secondary(name);
724 		if (eth_dev == NULL) {
725 			DRV_LOG(ERR, "can not attach rte ethdev");
726 			rte_errno = ENOMEM;
727 			return NULL;
728 		}
729 		eth_dev->device = dpdk_dev;
730 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
731 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
732 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
733 		err = mlx5_proc_priv_init(eth_dev);
734 		if (err)
735 			return NULL;
736 		mp_id.port_id = eth_dev->data->port_id;
737 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
738 		/* Receive command fd from primary process */
739 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
740 		if (err < 0)
741 			goto err_secondary;
742 		/* Remap UAR for Tx queues. */
743 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
744 		if (err)
745 			goto err_secondary;
746 		/*
747 		 * Ethdev pointer is still required as input since
748 		 * the primary device is not accessible from the
749 		 * secondary process.
750 		 */
751 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
752 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
753 		return eth_dev;
754 err_secondary:
755 		mlx5_dev_close(eth_dev);
756 		return NULL;
757 	}
758 	/*
759 	 * Some parameters ("tx_db_nc" in particularly) are needed in
760 	 * advance to create dv/verbs device context. We proceed the
761 	 * devargs here to get ones, and later proceed devargs again
762 	 * to override some hardware settings.
763 	 */
764 	err = mlx5_args(config, dpdk_dev->devargs);
765 	if (err) {
766 		err = rte_errno;
767 		DRV_LOG(ERR, "failed to process device arguments: %s",
768 			strerror(rte_errno));
769 		goto error;
770 	}
771 	if (config->dv_miss_info) {
772 		if (switch_info->master || switch_info->representor)
773 			config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
774 	}
775 	mlx5_malloc_mem_select(config->sys_mem_en);
776 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
777 	if (!sh)
778 		return NULL;
779 	config->devx = sh->devx;
780 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
781 	config->dest_tir = 1;
782 #endif
783 #ifdef HAVE_IBV_MLX5_MOD_SWP
784 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
785 #endif
786 	/*
787 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
788 	 * as all ConnectX-5 devices.
789 	 */
790 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
791 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
792 #endif
793 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
794 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
795 #endif
796 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
797 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
798 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
799 			DRV_LOG(DEBUG, "enhanced MPW is supported");
800 			mps = MLX5_MPW_ENHANCED;
801 		} else {
802 			DRV_LOG(DEBUG, "MPW is supported");
803 			mps = MLX5_MPW;
804 		}
805 	} else {
806 		DRV_LOG(DEBUG, "MPW isn't supported");
807 		mps = MLX5_MPW_DISABLED;
808 	}
809 #ifdef HAVE_IBV_MLX5_MOD_SWP
810 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
811 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
812 	DRV_LOG(DEBUG, "SWP support: %u", swp);
813 #endif
814 	config->swp = !!swp;
815 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
816 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
817 		struct mlx5dv_striding_rq_caps mprq_caps =
818 			dv_attr.striding_rq_caps;
819 
820 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
821 			mprq_caps.min_single_stride_log_num_of_bytes);
822 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
823 			mprq_caps.max_single_stride_log_num_of_bytes);
824 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
825 			mprq_caps.min_single_wqe_log_num_of_strides);
826 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
827 			mprq_caps.max_single_wqe_log_num_of_strides);
828 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
829 			mprq_caps.supported_qpts);
830 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
831 		mprq = 1;
832 		mprq_min_stride_size_n =
833 			mprq_caps.min_single_stride_log_num_of_bytes;
834 		mprq_max_stride_size_n =
835 			mprq_caps.max_single_stride_log_num_of_bytes;
836 		mprq_min_stride_num_n =
837 			mprq_caps.min_single_wqe_log_num_of_strides;
838 		mprq_max_stride_num_n =
839 			mprq_caps.max_single_wqe_log_num_of_strides;
840 	}
841 #endif
842 	if (RTE_CACHE_LINE_SIZE == 128 &&
843 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
844 		cqe_comp = 0;
845 	else
846 		cqe_comp = 1;
847 	config->cqe_comp = cqe_comp;
848 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
849 	/* Whether device supports 128B Rx CQE padding. */
850 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
851 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
852 #endif
853 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
854 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
855 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
856 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
857 			     (dv_attr.tunnel_offloads_caps &
858 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
859 			     (dv_attr.tunnel_offloads_caps &
860 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
861 	}
862 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
863 		tunnel_en ? "" : "not ");
864 #else
865 	DRV_LOG(WARNING,
866 		"tunnel offloading disabled due to old OFED/rdma-core version");
867 #endif
868 	config->tunnel_en = tunnel_en;
869 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
870 	mpls_en = ((dv_attr.tunnel_offloads_caps &
871 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
872 		   (dv_attr.tunnel_offloads_caps &
873 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
874 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
875 		mpls_en ? "" : "not ");
876 #else
877 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
878 		" old OFED/rdma-core version or firmware configuration");
879 #endif
880 	config->mpls_en = mpls_en;
881 	/* Check port status. */
882 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
883 	if (err) {
884 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
885 		goto error;
886 	}
887 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
888 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
889 		err = EINVAL;
890 		goto error;
891 	}
892 	if (port_attr.state != IBV_PORT_ACTIVE)
893 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
894 			mlx5_glue->port_state_str(port_attr.state),
895 			port_attr.state);
896 	/* Allocate private eth device data. */
897 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
898 			   sizeof(*priv),
899 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
900 	if (priv == NULL) {
901 		DRV_LOG(ERR, "priv allocation failure");
902 		err = ENOMEM;
903 		goto error;
904 	}
905 	priv->sh = sh;
906 	priv->dev_port = spawn->phys_port;
907 	priv->pci_dev = spawn->pci_dev;
908 	priv->mtu = RTE_ETHER_MTU;
909 	priv->mp_id.port_id = port_id;
910 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
911 	/* Some internal functions rely on Netlink sockets, open them now. */
912 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
913 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
914 	priv->representor = !!switch_info->representor;
915 	priv->master = !!switch_info->master;
916 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
917 	priv->vport_meta_tag = 0;
918 	priv->vport_meta_mask = 0;
919 	priv->pf_bond = spawn->pf_bond;
920 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
921 	/*
922 	 * The DevX port query API is implemented. E-Switch may use
923 	 * either vport or reg_c[0] metadata register to match on
924 	 * vport index. The engaged part of metadata register is
925 	 * defined by mask.
926 	 */
927 	if (switch_info->representor || switch_info->master) {
928 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
929 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
930 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
931 						 &devx_port);
932 		if (err) {
933 			DRV_LOG(WARNING,
934 				"can't query devx port %d on device %s",
935 				spawn->phys_port,
936 				mlx5_os_get_dev_device_name(spawn->phys_dev));
937 			devx_port.comp_mask = 0;
938 		}
939 	}
940 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
941 		priv->vport_meta_tag = devx_port.reg_c_0.value;
942 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
943 		if (!priv->vport_meta_mask) {
944 			DRV_LOG(ERR, "vport zero mask for port %d"
945 				     " on bonding device %s",
946 				     spawn->phys_port,
947 				     mlx5_os_get_dev_device_name
948 							(spawn->phys_dev));
949 			err = ENOTSUP;
950 			goto error;
951 		}
952 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
953 			DRV_LOG(ERR, "invalid vport tag for port %d"
954 				     " on bonding device %s",
955 				     spawn->phys_port,
956 				     mlx5_os_get_dev_device_name
957 							(spawn->phys_dev));
958 			err = ENOTSUP;
959 			goto error;
960 		}
961 	}
962 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
963 		priv->vport_id = devx_port.vport_num;
964 	} else if (spawn->pf_bond >= 0) {
965 		DRV_LOG(ERR, "can't deduce vport index for port %d"
966 			     " on bonding device %s",
967 			     spawn->phys_port,
968 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
969 		err = ENOTSUP;
970 		goto error;
971 	} else {
972 		/* Suppose vport index in compatible way. */
973 		priv->vport_id = switch_info->representor ?
974 				 switch_info->port_name + 1 : -1;
975 	}
976 #else
977 	/*
978 	 * Kernel/rdma_core support single E-Switch per PF configurations
979 	 * only and vport_id field contains the vport index for
980 	 * associated VF, which is deduced from representor port name.
981 	 * For example, let's have the IB device port 10, it has
982 	 * attached network device eth0, which has port name attribute
983 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
984 	 * as 3 (2+1). This assigning schema should be changed if the
985 	 * multiple E-Switch instances per PF configurations or/and PCI
986 	 * subfunctions are added.
987 	 */
988 	priv->vport_id = switch_info->representor ?
989 			 switch_info->port_name + 1 : -1;
990 #endif
991 	/* representor_id field keeps the unmodified VF index. */
992 	priv->representor_id = switch_info->representor ?
993 			       switch_info->port_name : -1;
994 	/*
995 	 * Look for sibling devices in order to reuse their switch domain
996 	 * if any, otherwise allocate one.
997 	 */
998 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
999 		const struct mlx5_priv *opriv =
1000 			rte_eth_devices[port_id].data->dev_private;
1001 
1002 		if (!opriv ||
1003 		    opriv->sh != priv->sh ||
1004 			opriv->domain_id ==
1005 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1006 			continue;
1007 		priv->domain_id = opriv->domain_id;
1008 		break;
1009 	}
1010 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1011 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
1012 		if (err) {
1013 			err = rte_errno;
1014 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
1015 				strerror(rte_errno));
1016 			goto error;
1017 		}
1018 		own_domain_id = 1;
1019 	}
1020 	/* Override some values set by hardware configuration. */
1021 	mlx5_args(config, dpdk_dev->devargs);
1022 	err = mlx5_dev_check_sibling_config(priv, config);
1023 	if (err)
1024 		goto error;
1025 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1026 			    IBV_DEVICE_RAW_IP_CSUM);
1027 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1028 		(config->hw_csum ? "" : "not "));
1029 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1030 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1031 	DRV_LOG(DEBUG, "counters are not supported");
1032 #endif
1033 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1034 	if (config->dv_flow_en) {
1035 		DRV_LOG(WARNING, "DV flow is not supported");
1036 		config->dv_flow_en = 0;
1037 	}
1038 #endif
1039 	config->ind_table_max_size =
1040 		sh->device_attr.max_rwq_indirection_table_size;
1041 	/*
1042 	 * Remove this check once DPDK supports larger/variable
1043 	 * indirection tables.
1044 	 */
1045 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1046 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1047 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1048 		config->ind_table_max_size);
1049 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1050 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1051 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1052 		(config->hw_vlan_strip ? "" : "not "));
1053 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1054 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1055 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1056 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1057 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1058 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1059 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1060 #endif
1061 	if (config->hw_padding && !hw_padding) {
1062 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1063 		config->hw_padding = 0;
1064 	} else if (config->hw_padding) {
1065 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1066 	}
1067 	config->tso = (sh->device_attr.max_tso > 0 &&
1068 		      (sh->device_attr.tso_supported_qpts &
1069 		       (1 << IBV_QPT_RAW_PACKET)));
1070 	if (config->tso)
1071 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1072 	/*
1073 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1074 	 * by default.
1075 	 */
1076 	if (config->mps == MLX5_ARG_UNSET)
1077 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1078 							  MLX5_MPW_DISABLED;
1079 	else
1080 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1081 	DRV_LOG(INFO, "%sMPS is %s",
1082 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1083 		config->mps == MLX5_MPW ? "legacy " : "",
1084 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1085 	if (config->cqe_comp && !cqe_comp) {
1086 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1087 		config->cqe_comp = 0;
1088 	}
1089 	if (config->cqe_pad && !cqe_pad) {
1090 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1091 		config->cqe_pad = 0;
1092 	} else if (config->cqe_pad) {
1093 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1094 	}
1095 	if (config->devx) {
1096 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1097 		if (err) {
1098 			err = -err;
1099 			goto error;
1100 		}
1101 		/* Check relax ordering support. */
1102 		if (config->hca_attr.relaxed_ordering_write &&
1103 		    config->hca_attr.relaxed_ordering_read  &&
1104 		    !haswell_broadwell_cpu)
1105 			sh->cmng.relaxed_ordering = 1;
1106 		/* Check for LRO support. */
1107 		if (config->dest_tir && config->hca_attr.lro_cap &&
1108 		    config->dv_flow_en) {
1109 			/* TBD check tunnel lro caps. */
1110 			config->lro.supported = config->hca_attr.lro_cap;
1111 			DRV_LOG(DEBUG, "Device supports LRO");
1112 			/*
1113 			 * If LRO timeout is not configured by application,
1114 			 * use the minimal supported value.
1115 			 */
1116 			if (!config->lro.timeout)
1117 				config->lro.timeout =
1118 				config->hca_attr.lro_timer_supported_periods[0];
1119 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1120 				config->lro.timeout);
1121 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1122 				"required for coalescing is %d bytes",
1123 				config->hca_attr.lro_min_mss_size);
1124 		}
1125 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1126 		if (config->hca_attr.qos.sup &&
1127 		    config->hca_attr.qos.srtcm_sup &&
1128 		    config->dv_flow_en) {
1129 			uint8_t reg_c_mask =
1130 				config->hca_attr.qos.flow_meter_reg_c_ids;
1131 			/*
1132 			 * Meter needs two REG_C's for color match and pre-sfx
1133 			 * flow match. Here get the REG_C for color match.
1134 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1135 			 */
1136 			reg_c_mask &= 0xfc;
1137 			if (__builtin_popcount(reg_c_mask) < 1) {
1138 				priv->mtr_en = 0;
1139 				DRV_LOG(WARNING, "No available register for"
1140 					" meter.");
1141 			} else {
1142 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1143 						      REG_C_0;
1144 				priv->mtr_en = 1;
1145 				priv->mtr_reg_share =
1146 				      config->hca_attr.qos.flow_meter_reg_share;
1147 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1148 					priv->mtr_color_reg);
1149 			}
1150 		}
1151 #endif
1152 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1153 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1154 		    config->dv_flow_en) {
1155 			priv->sampler_en = 1;
1156 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1157 		} else {
1158 			priv->sampler_en = 0;
1159 			if (!config->hca_attr.log_max_ft_sampler_num)
1160 				DRV_LOG(WARNING, "No available register for"
1161 						" Sampler.");
1162 			else
1163 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1164 		}
1165 #endif
1166 	}
1167 	if (config->tx_pp) {
1168 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1169 			config->hca_attr.dev_freq_khz);
1170 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1171 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1172 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1173 			config->hca_attr.cross_channel ? "" : "not ");
1174 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1175 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1176 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1177 			config->hca_attr.non_wire_sq ? "" : "not ");
1178 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1179 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1180 			config->hca_attr.log_max_static_sq_wq);
1181 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1182 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1183 		if (!config->devx) {
1184 			DRV_LOG(ERR, "DevX is required for packet pacing");
1185 			err = ENODEV;
1186 			goto error;
1187 		}
1188 		if (!config->hca_attr.qos.packet_pacing) {
1189 			DRV_LOG(ERR, "Packet pacing is not supported");
1190 			err = ENODEV;
1191 			goto error;
1192 		}
1193 		if (!config->hca_attr.cross_channel) {
1194 			DRV_LOG(ERR, "Cross channel operations are"
1195 				     " required for packet pacing");
1196 			err = ENODEV;
1197 			goto error;
1198 		}
1199 		if (!config->hca_attr.wqe_index_ignore) {
1200 			DRV_LOG(ERR, "WQE index ignore feature is"
1201 				     " required for packet pacing");
1202 			err = ENODEV;
1203 			goto error;
1204 		}
1205 		if (!config->hca_attr.non_wire_sq) {
1206 			DRV_LOG(ERR, "Non-wire SQ feature is"
1207 				     " required for packet pacing");
1208 			err = ENODEV;
1209 			goto error;
1210 		}
1211 		if (!config->hca_attr.log_max_static_sq_wq) {
1212 			DRV_LOG(ERR, "Static WQE SQ feature is"
1213 				     " required for packet pacing");
1214 			err = ENODEV;
1215 			goto error;
1216 		}
1217 		if (!config->hca_attr.qos.wqe_rate_pp) {
1218 			DRV_LOG(ERR, "WQE rate mode is required"
1219 				     " for packet pacing");
1220 			err = ENODEV;
1221 			goto error;
1222 		}
1223 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1224 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1225 			     " can't create queues for packet pacing");
1226 		err = ENODEV;
1227 		goto error;
1228 #endif
1229 	}
1230 	if (config->devx) {
1231 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1232 
1233 		err = config->hca_attr.access_register_user ?
1234 			mlx5_devx_cmd_register_read
1235 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1236 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1237 		if (!err) {
1238 			uint32_t ts_mode;
1239 
1240 			/* MTUTC register is read successfully. */
1241 			ts_mode = MLX5_GET(register_mtutc, reg,
1242 					   time_stamp_mode);
1243 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1244 				config->rt_timestamp = 1;
1245 		} else {
1246 			/* Kernel does not support register reading. */
1247 			if (config->hca_attr.dev_freq_khz ==
1248 						 (NS_PER_S / MS_PER_S))
1249 				config->rt_timestamp = 1;
1250 		}
1251 	}
1252 	/*
1253 	 * If HW has bug working with tunnel packet decapsulation and
1254 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1255 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1256 	 */
1257 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1258 		config->hw_fcs_strip = 0;
1259 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1260 		(config->hw_fcs_strip ? "" : "not "));
1261 	if (config->mprq.enabled && mprq) {
1262 		if (config->mprq.stride_num_n &&
1263 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1264 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1265 			config->mprq.stride_num_n =
1266 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1267 						mprq_min_stride_num_n),
1268 					mprq_max_stride_num_n);
1269 			DRV_LOG(WARNING,
1270 				"the number of strides"
1271 				" for Multi-Packet RQ is out of range,"
1272 				" setting default value (%u)",
1273 				1 << config->mprq.stride_num_n);
1274 		}
1275 		if (config->mprq.stride_size_n &&
1276 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1277 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1278 			config->mprq.stride_size_n =
1279 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1280 						mprq_min_stride_size_n),
1281 					mprq_max_stride_size_n);
1282 			DRV_LOG(WARNING,
1283 				"the size of a stride"
1284 				" for Multi-Packet RQ is out of range,"
1285 				" setting default value (%u)",
1286 				1 << config->mprq.stride_size_n);
1287 		}
1288 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1289 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1290 	} else if (config->mprq.enabled && !mprq) {
1291 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1292 		config->mprq.enabled = 0;
1293 	}
1294 	if (config->max_dump_files_num == 0)
1295 		config->max_dump_files_num = 128;
1296 	eth_dev = rte_eth_dev_allocate(name);
1297 	if (eth_dev == NULL) {
1298 		DRV_LOG(ERR, "can not allocate rte ethdev");
1299 		err = ENOMEM;
1300 		goto error;
1301 	}
1302 	if (priv->representor) {
1303 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1304 		eth_dev->data->representor_id = priv->representor_id;
1305 	}
1306 	/*
1307 	 * Store associated network device interface index. This index
1308 	 * is permanent throughout the lifetime of device. So, we may store
1309 	 * the ifindex here and use the cached value further.
1310 	 */
1311 	MLX5_ASSERT(spawn->ifindex);
1312 	priv->if_index = spawn->ifindex;
1313 	if (priv->pf_bond >= 0 && priv->master) {
1314 		/* Get bond interface info */
1315 		err = mlx5_sysfs_bond_info(priv->if_index,
1316 				     &priv->bond_ifindex,
1317 				     priv->bond_name);
1318 		if (err)
1319 			DRV_LOG(ERR, "unable to get bond info: %s",
1320 				strerror(rte_errno));
1321 		else
1322 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1323 				priv->if_index, priv->bond_ifindex,
1324 				priv->bond_name);
1325 	}
1326 	eth_dev->data->dev_private = priv;
1327 	priv->dev_data = eth_dev->data;
1328 	eth_dev->data->mac_addrs = priv->mac;
1329 	eth_dev->device = dpdk_dev;
1330 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1331 	/* Configure the first MAC address by default. */
1332 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1333 		DRV_LOG(ERR,
1334 			"port %u cannot get MAC address, is mlx5_en"
1335 			" loaded? (errno: %s)",
1336 			eth_dev->data->port_id, strerror(rte_errno));
1337 		err = ENODEV;
1338 		goto error;
1339 	}
1340 	DRV_LOG(INFO,
1341 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1342 		eth_dev->data->port_id,
1343 		mac.addr_bytes[0], mac.addr_bytes[1],
1344 		mac.addr_bytes[2], mac.addr_bytes[3],
1345 		mac.addr_bytes[4], mac.addr_bytes[5]);
1346 #ifdef RTE_LIBRTE_MLX5_DEBUG
1347 	{
1348 		char ifname[IF_NAMESIZE];
1349 
1350 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1351 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1352 				eth_dev->data->port_id, ifname);
1353 		else
1354 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1355 				eth_dev->data->port_id);
1356 	}
1357 #endif
1358 	/* Get actual MTU if possible. */
1359 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1360 	if (err) {
1361 		err = rte_errno;
1362 		goto error;
1363 	}
1364 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1365 		priv->mtu);
1366 	/* Initialize burst functions to prevent crashes before link-up. */
1367 	eth_dev->rx_pkt_burst = removed_rx_burst;
1368 	eth_dev->tx_pkt_burst = removed_tx_burst;
1369 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1370 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1371 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1372 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1373 	/* Register MAC address. */
1374 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1375 	if (config->vf && config->vf_nl_en)
1376 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1377 				      mlx5_ifindex(eth_dev),
1378 				      eth_dev->data->mac_addrs,
1379 				      MLX5_MAX_MAC_ADDRESSES);
1380 	priv->flows = 0;
1381 	priv->ctrl_flows = 0;
1382 	rte_spinlock_init(&priv->flow_list_lock);
1383 	TAILQ_INIT(&priv->flow_meters);
1384 	TAILQ_INIT(&priv->flow_meter_profiles);
1385 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1386 	mlx5_glue->dv_set_context_attr(sh->ctx,
1387 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1388 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1389 				.alloc = &mlx5_alloc_verbs_buf,
1390 				.free = &mlx5_free_verbs_buf,
1391 				.data = priv,
1392 			}));
1393 	/* Bring Ethernet device up. */
1394 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1395 		eth_dev->data->port_id);
1396 	mlx5_set_link_up(eth_dev);
1397 	/*
1398 	 * Even though the interrupt handler is not installed yet,
1399 	 * interrupts will still trigger on the async_fd from
1400 	 * Verbs context returned by ibv_open_device().
1401 	 */
1402 	mlx5_link_update(eth_dev, 0);
1403 #ifdef HAVE_MLX5DV_DR_ESWITCH
1404 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1405 	      (switch_info->representor || switch_info->master)))
1406 		config->dv_esw_en = 0;
1407 #else
1408 	config->dv_esw_en = 0;
1409 #endif
1410 	/* Detect minimal data bytes to inline. */
1411 	mlx5_set_min_inline(spawn, config);
1412 	/* Store device configuration on private structure. */
1413 	priv->config = *config;
1414 	/* Create context for virtual machine VLAN workaround. */
1415 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1416 	if (config->dv_flow_en) {
1417 		err = mlx5_alloc_shared_dr(priv);
1418 		if (err)
1419 			goto error;
1420 	}
1421 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1422 		priv->obj_ops = devx_obj_ops;
1423 		priv->obj_ops.drop_action_create =
1424 						ibv_obj_ops.drop_action_create;
1425 		priv->obj_ops.drop_action_destroy =
1426 						ibv_obj_ops.drop_action_destroy;
1427 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1428 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1429 #else
1430 		if (config->dv_esw_en)
1431 			priv->obj_ops.txq_obj_modify =
1432 						ibv_obj_ops.txq_obj_modify;
1433 #endif
1434 		/* Use specific wrappers for Tx object. */
1435 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1436 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1437 
1438 	} else {
1439 		priv->obj_ops = ibv_obj_ops;
1440 	}
1441 	priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1442 	if (!priv->drop_queue.hrxq)
1443 		goto error;
1444 	/* Supported Verbs flow priority number detection. */
1445 	err = mlx5_flow_discover_priorities(eth_dev);
1446 	if (err < 0) {
1447 		err = -err;
1448 		goto error;
1449 	}
1450 	priv->config.flow_prio = err;
1451 	if (!priv->config.dv_esw_en &&
1452 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1453 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1454 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1455 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1456 	}
1457 	mlx5_set_metadata_mask(eth_dev);
1458 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1459 	    !priv->sh->dv_regc0_mask) {
1460 		DRV_LOG(ERR, "metadata mode %u is not supported "
1461 			     "(no metadata reg_c[0] is available)",
1462 			     priv->config.dv_xmeta_en);
1463 			err = ENOTSUP;
1464 			goto error;
1465 	}
1466 	/* Query availability of metadata reg_c's. */
1467 	err = mlx5_flow_discover_mreg_c(eth_dev);
1468 	if (err < 0) {
1469 		err = -err;
1470 		goto error;
1471 	}
1472 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1473 		DRV_LOG(DEBUG,
1474 			"port %u extensive metadata register is not supported",
1475 			eth_dev->data->port_id);
1476 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1477 			DRV_LOG(ERR, "metadata mode %u is not supported "
1478 				     "(no metadata registers available)",
1479 				     priv->config.dv_xmeta_en);
1480 			err = ENOTSUP;
1481 			goto error;
1482 		}
1483 	}
1484 	if (priv->config.dv_flow_en &&
1485 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1486 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1487 	    priv->sh->dv_regc0_mask) {
1488 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1489 						      MLX5_FLOW_MREG_HTABLE_SZ,
1490 						      0, 0,
1491 						      NULL, NULL, NULL);
1492 		if (!priv->mreg_cp_tbl) {
1493 			err = ENOMEM;
1494 			goto error;
1495 		}
1496 	}
1497 	mlx5_flow_counter_mode_config(eth_dev);
1498 	return eth_dev;
1499 error:
1500 	if (priv) {
1501 		if (priv->mreg_cp_tbl)
1502 			mlx5_hlist_destroy(priv->mreg_cp_tbl);
1503 		if (priv->sh)
1504 			mlx5_os_free_shared_dr(priv);
1505 		if (priv->nl_socket_route >= 0)
1506 			close(priv->nl_socket_route);
1507 		if (priv->nl_socket_rdma >= 0)
1508 			close(priv->nl_socket_rdma);
1509 		if (priv->vmwa_context)
1510 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1511 		if (eth_dev && priv->drop_queue.hrxq)
1512 			mlx5_drop_action_destroy(eth_dev);
1513 		if (own_domain_id)
1514 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1515 		mlx5_free(priv);
1516 		if (eth_dev != NULL)
1517 			eth_dev->data->dev_private = NULL;
1518 	}
1519 	if (eth_dev != NULL) {
1520 		/* mac_addrs must not be freed alone because part of
1521 		 * dev_private
1522 		 **/
1523 		eth_dev->data->mac_addrs = NULL;
1524 		rte_eth_dev_release_port(eth_dev);
1525 	}
1526 	if (sh)
1527 		mlx5_free_shared_dev_ctx(sh);
1528 	MLX5_ASSERT(err > 0);
1529 	rte_errno = err;
1530 	return NULL;
1531 }
1532 
1533 /**
1534  * Comparison callback to sort device data.
1535  *
1536  * This is meant to be used with qsort().
1537  *
1538  * @param a[in]
1539  *   Pointer to pointer to first data object.
1540  * @param b[in]
1541  *   Pointer to pointer to second data object.
1542  *
1543  * @return
1544  *   0 if both objects are equal, less than 0 if the first argument is less
1545  *   than the second, greater than 0 otherwise.
1546  */
1547 static int
1548 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1549 {
1550 	const struct mlx5_switch_info *si_a =
1551 		&((const struct mlx5_dev_spawn_data *)a)->info;
1552 	const struct mlx5_switch_info *si_b =
1553 		&((const struct mlx5_dev_spawn_data *)b)->info;
1554 	int ret;
1555 
1556 	/* Master device first. */
1557 	ret = si_b->master - si_a->master;
1558 	if (ret)
1559 		return ret;
1560 	/* Then representor devices. */
1561 	ret = si_b->representor - si_a->representor;
1562 	if (ret)
1563 		return ret;
1564 	/* Unidentified devices come last in no specific order. */
1565 	if (!si_a->representor)
1566 		return 0;
1567 	/* Order representors by name. */
1568 	return si_a->port_name - si_b->port_name;
1569 }
1570 
1571 /**
1572  * Match PCI information for possible slaves of bonding device.
1573  *
1574  * @param[in] ibv_dev
1575  *   Pointer to Infiniband device structure.
1576  * @param[in] pci_dev
1577  *   Pointer to PCI device structure to match PCI address.
1578  * @param[in] nl_rdma
1579  *   Netlink RDMA group socket handle.
1580  *
1581  * @return
1582  *   negative value if no bonding device found, otherwise
1583  *   positive index of slave PF in bonding.
1584  */
1585 static int
1586 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1587 			   const struct rte_pci_device *pci_dev,
1588 			   int nl_rdma)
1589 {
1590 	char ifname[IF_NAMESIZE + 1];
1591 	unsigned int ifindex;
1592 	unsigned int np, i;
1593 	FILE *file = NULL;
1594 	int pf = -1;
1595 
1596 	/*
1597 	 * Try to get master device name. If something goes
1598 	 * wrong suppose the lack of kernel support and no
1599 	 * bonding devices.
1600 	 */
1601 	if (nl_rdma < 0)
1602 		return -1;
1603 	if (!strstr(ibv_dev->name, "bond"))
1604 		return -1;
1605 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1606 	if (!np)
1607 		return -1;
1608 	/*
1609 	 * The Master device might not be on the predefined
1610 	 * port (not on port index 1, it is not garanted),
1611 	 * we have to scan all Infiniband device port and
1612 	 * find master.
1613 	 */
1614 	for (i = 1; i <= np; ++i) {
1615 		/* Check whether Infiniband port is populated. */
1616 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1617 		if (!ifindex)
1618 			continue;
1619 		if (!if_indextoname(ifindex, ifname))
1620 			continue;
1621 		/* Try to read bonding slave names from sysfs. */
1622 		MKSTR(slaves,
1623 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1624 		file = fopen(slaves, "r");
1625 		if (file)
1626 			break;
1627 	}
1628 	if (!file)
1629 		return -1;
1630 	/* Use safe format to check maximal buffer length. */
1631 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1632 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1633 		char tmp_str[IF_NAMESIZE + 32];
1634 		struct rte_pci_addr pci_addr;
1635 		struct mlx5_switch_info	info;
1636 
1637 		/* Process slave interface names in the loop. */
1638 		snprintf(tmp_str, sizeof(tmp_str),
1639 			 "/sys/class/net/%s", ifname);
1640 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1641 			DRV_LOG(WARNING, "can not get PCI address"
1642 					 " for netdev \"%s\"", ifname);
1643 			continue;
1644 		}
1645 		if (pci_dev->addr.domain != pci_addr.domain ||
1646 		    pci_dev->addr.bus != pci_addr.bus ||
1647 		    pci_dev->addr.devid != pci_addr.devid ||
1648 		    pci_dev->addr.function != pci_addr.function)
1649 			continue;
1650 		/* Slave interface PCI address match found. */
1651 		fclose(file);
1652 		snprintf(tmp_str, sizeof(tmp_str),
1653 			 "/sys/class/net/%s/phys_port_name", ifname);
1654 		file = fopen(tmp_str, "rb");
1655 		if (!file)
1656 			break;
1657 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1658 		if (fscanf(file, "%32s", tmp_str) == 1)
1659 			mlx5_translate_port_name(tmp_str, &info);
1660 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1661 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1662 			pf = info.port_name;
1663 		break;
1664 	}
1665 	if (file)
1666 		fclose(file);
1667 	return pf;
1668 }
1669 
1670 /**
1671  * DPDK callback to register a PCI device.
1672  *
1673  * This function spawns Ethernet devices out of a given PCI device.
1674  *
1675  * @param[in] pci_drv
1676  *   PCI driver structure (mlx5_driver).
1677  * @param[in] pci_dev
1678  *   PCI device information.
1679  *
1680  * @return
1681  *   0 on success, a negative errno value otherwise and rte_errno is set.
1682  */
1683 int
1684 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1685 		  struct rte_pci_device *pci_dev)
1686 {
1687 	struct ibv_device **ibv_list;
1688 	/*
1689 	 * Number of found IB Devices matching with requested PCI BDF.
1690 	 * nd != 1 means there are multiple IB devices over the same
1691 	 * PCI device and we have representors and master.
1692 	 */
1693 	unsigned int nd = 0;
1694 	/*
1695 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1696 	 * we have the single multiport IB device, and there may be
1697 	 * representors attached to some of found ports.
1698 	 */
1699 	unsigned int np = 0;
1700 	/*
1701 	 * Number of DPDK ethernet devices to Spawn - either over
1702 	 * multiple IB devices or multiple ports of single IB device.
1703 	 * Actually this is the number of iterations to spawn.
1704 	 */
1705 	unsigned int ns = 0;
1706 	/*
1707 	 * Bonding device
1708 	 *   < 0 - no bonding device (single one)
1709 	 *  >= 0 - bonding device (value is slave PF index)
1710 	 */
1711 	int bd = -1;
1712 	struct mlx5_dev_spawn_data *list = NULL;
1713 	struct mlx5_dev_config dev_config;
1714 	unsigned int dev_config_vf;
1715 	int ret;
1716 
1717 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1718 		mlx5_pmd_socket_init();
1719 	ret = mlx5_init_once();
1720 	if (ret) {
1721 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1722 			strerror(rte_errno));
1723 		return -rte_errno;
1724 	}
1725 	errno = 0;
1726 	ibv_list = mlx5_glue->get_device_list(&ret);
1727 	if (!ibv_list) {
1728 		rte_errno = errno ? errno : ENOSYS;
1729 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1730 		return -rte_errno;
1731 	}
1732 	/*
1733 	 * First scan the list of all Infiniband devices to find
1734 	 * matching ones, gathering into the list.
1735 	 */
1736 	struct ibv_device *ibv_match[ret + 1];
1737 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1738 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1739 	unsigned int i;
1740 
1741 	while (ret-- > 0) {
1742 		struct rte_pci_addr pci_addr;
1743 
1744 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1745 		bd = mlx5_device_bond_pci_match
1746 				(ibv_list[ret], pci_dev, nl_rdma);
1747 		if (bd >= 0) {
1748 			/*
1749 			 * Bonding device detected. Only one match is allowed,
1750 			 * the bonding is supported over multi-port IB device,
1751 			 * there should be no matches on representor PCI
1752 			 * functions or non VF LAG bonding devices with
1753 			 * specified address.
1754 			 */
1755 			if (nd) {
1756 				DRV_LOG(ERR,
1757 					"multiple PCI match on bonding device"
1758 					"\"%s\" found", ibv_list[ret]->name);
1759 				rte_errno = ENOENT;
1760 				ret = -rte_errno;
1761 				goto exit;
1762 			}
1763 			DRV_LOG(INFO, "PCI information matches for"
1764 				      " slave %d bonding device \"%s\"",
1765 				      bd, ibv_list[ret]->name);
1766 			ibv_match[nd++] = ibv_list[ret];
1767 			break;
1768 		}
1769 		if (mlx5_dev_to_pci_addr
1770 			(ibv_list[ret]->ibdev_path, &pci_addr))
1771 			continue;
1772 		if (pci_dev->addr.domain != pci_addr.domain ||
1773 		    pci_dev->addr.bus != pci_addr.bus ||
1774 		    pci_dev->addr.devid != pci_addr.devid ||
1775 		    pci_dev->addr.function != pci_addr.function)
1776 			continue;
1777 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1778 			ibv_list[ret]->name);
1779 		ibv_match[nd++] = ibv_list[ret];
1780 	}
1781 	ibv_match[nd] = NULL;
1782 	if (!nd) {
1783 		/* No device matches, just complain and bail out. */
1784 		DRV_LOG(WARNING,
1785 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1786 			" are kernel drivers loaded?",
1787 			pci_dev->addr.domain, pci_dev->addr.bus,
1788 			pci_dev->addr.devid, pci_dev->addr.function);
1789 		rte_errno = ENOENT;
1790 		ret = -rte_errno;
1791 		goto exit;
1792 	}
1793 	if (nd == 1) {
1794 		/*
1795 		 * Found single matching device may have multiple ports.
1796 		 * Each port may be representor, we have to check the port
1797 		 * number and check the representors existence.
1798 		 */
1799 		if (nl_rdma >= 0)
1800 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1801 		if (!np)
1802 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1803 					 " ports number", ibv_match[0]->name);
1804 		if (bd >= 0 && !np) {
1805 			DRV_LOG(ERR, "can not get ports"
1806 				     " for bonding device");
1807 			rte_errno = ENOENT;
1808 			ret = -rte_errno;
1809 			goto exit;
1810 		}
1811 	}
1812 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1813 	if (bd >= 0) {
1814 		/*
1815 		 * This may happen if there is VF LAG kernel support and
1816 		 * application is compiled with older rdma_core library.
1817 		 */
1818 		DRV_LOG(ERR,
1819 			"No kernel/verbs support for VF LAG bonding found.");
1820 		rte_errno = ENOTSUP;
1821 		ret = -rte_errno;
1822 		goto exit;
1823 	}
1824 #endif
1825 	/*
1826 	 * Now we can determine the maximal
1827 	 * amount of devices to be spawned.
1828 	 */
1829 	list = mlx5_malloc(MLX5_MEM_ZERO,
1830 			   sizeof(struct mlx5_dev_spawn_data) *
1831 			   (np ? np : nd),
1832 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1833 	if (!list) {
1834 		DRV_LOG(ERR, "spawn data array allocation failure");
1835 		rte_errno = ENOMEM;
1836 		ret = -rte_errno;
1837 		goto exit;
1838 	}
1839 	if (bd >= 0 || np > 1) {
1840 		/*
1841 		 * Single IB device with multiple ports found,
1842 		 * it may be E-Switch master device and representors.
1843 		 * We have to perform identification through the ports.
1844 		 */
1845 		MLX5_ASSERT(nl_rdma >= 0);
1846 		MLX5_ASSERT(ns == 0);
1847 		MLX5_ASSERT(nd == 1);
1848 		MLX5_ASSERT(np);
1849 		for (i = 1; i <= np; ++i) {
1850 			list[ns].max_port = np;
1851 			list[ns].phys_port = i;
1852 			list[ns].phys_dev = ibv_match[0];
1853 			list[ns].eth_dev = NULL;
1854 			list[ns].pci_dev = pci_dev;
1855 			list[ns].pf_bond = bd;
1856 			list[ns].ifindex = mlx5_nl_ifindex
1857 				(nl_rdma,
1858 				mlx5_os_get_dev_device_name
1859 						(list[ns].phys_dev), i);
1860 			if (!list[ns].ifindex) {
1861 				/*
1862 				 * No network interface index found for the
1863 				 * specified port, it means there is no
1864 				 * representor on this port. It's OK,
1865 				 * there can be disabled ports, for example
1866 				 * if sriov_numvfs < sriov_totalvfs.
1867 				 */
1868 				continue;
1869 			}
1870 			ret = -1;
1871 			if (nl_route >= 0)
1872 				ret = mlx5_nl_switch_info
1873 					       (nl_route,
1874 						list[ns].ifindex,
1875 						&list[ns].info);
1876 			if (ret || (!list[ns].info.representor &&
1877 				    !list[ns].info.master)) {
1878 				/*
1879 				 * We failed to recognize representors with
1880 				 * Netlink, let's try to perform the task
1881 				 * with sysfs.
1882 				 */
1883 				ret =  mlx5_sysfs_switch_info
1884 						(list[ns].ifindex,
1885 						 &list[ns].info);
1886 			}
1887 			if (!ret && bd >= 0) {
1888 				switch (list[ns].info.name_type) {
1889 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1890 					if (list[ns].info.port_name == bd)
1891 						ns++;
1892 					break;
1893 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1894 					/* Fallthrough */
1895 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1896 					if (list[ns].info.pf_num == bd)
1897 						ns++;
1898 					break;
1899 				default:
1900 					break;
1901 				}
1902 				continue;
1903 			}
1904 			if (!ret && (list[ns].info.representor ^
1905 				     list[ns].info.master))
1906 				ns++;
1907 		}
1908 		if (!ns) {
1909 			DRV_LOG(ERR,
1910 				"unable to recognize master/representors"
1911 				" on the IB device with multiple ports");
1912 			rte_errno = ENOENT;
1913 			ret = -rte_errno;
1914 			goto exit;
1915 		}
1916 	} else {
1917 		/*
1918 		 * The existence of several matching entries (nd > 1) means
1919 		 * port representors have been instantiated. No existing Verbs
1920 		 * call nor sysfs entries can tell them apart, this can only
1921 		 * be done through Netlink calls assuming kernel drivers are
1922 		 * recent enough to support them.
1923 		 *
1924 		 * In the event of identification failure through Netlink,
1925 		 * try again through sysfs, then:
1926 		 *
1927 		 * 1. A single IB device matches (nd == 1) with single
1928 		 *    port (np=0/1) and is not a representor, assume
1929 		 *    no switch support.
1930 		 *
1931 		 * 2. Otherwise no safe assumptions can be made;
1932 		 *    complain louder and bail out.
1933 		 */
1934 		for (i = 0; i != nd; ++i) {
1935 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1936 			list[ns].max_port = 1;
1937 			list[ns].phys_port = 1;
1938 			list[ns].phys_dev = ibv_match[i];
1939 			list[ns].eth_dev = NULL;
1940 			list[ns].pci_dev = pci_dev;
1941 			list[ns].pf_bond = -1;
1942 			list[ns].ifindex = 0;
1943 			if (nl_rdma >= 0)
1944 				list[ns].ifindex = mlx5_nl_ifindex
1945 				(nl_rdma,
1946 				mlx5_os_get_dev_device_name
1947 						(list[ns].phys_dev), 1);
1948 			if (!list[ns].ifindex) {
1949 				char ifname[IF_NAMESIZE];
1950 
1951 				/*
1952 				 * Netlink failed, it may happen with old
1953 				 * ib_core kernel driver (before 4.16).
1954 				 * We can assume there is old driver because
1955 				 * here we are processing single ports IB
1956 				 * devices. Let's try sysfs to retrieve
1957 				 * the ifindex. The method works for
1958 				 * master device only.
1959 				 */
1960 				if (nd > 1) {
1961 					/*
1962 					 * Multiple devices found, assume
1963 					 * representors, can not distinguish
1964 					 * master/representor and retrieve
1965 					 * ifindex via sysfs.
1966 					 */
1967 					continue;
1968 				}
1969 				ret = mlx5_get_ifname_sysfs
1970 					(ibv_match[i]->ibdev_path, ifname);
1971 				if (!ret)
1972 					list[ns].ifindex =
1973 						if_nametoindex(ifname);
1974 				if (!list[ns].ifindex) {
1975 					/*
1976 					 * No network interface index found
1977 					 * for the specified device, it means
1978 					 * there it is neither representor
1979 					 * nor master.
1980 					 */
1981 					continue;
1982 				}
1983 			}
1984 			ret = -1;
1985 			if (nl_route >= 0)
1986 				ret = mlx5_nl_switch_info
1987 					       (nl_route,
1988 						list[ns].ifindex,
1989 						&list[ns].info);
1990 			if (ret || (!list[ns].info.representor &&
1991 				    !list[ns].info.master)) {
1992 				/*
1993 				 * We failed to recognize representors with
1994 				 * Netlink, let's try to perform the task
1995 				 * with sysfs.
1996 				 */
1997 				ret =  mlx5_sysfs_switch_info
1998 						(list[ns].ifindex,
1999 						 &list[ns].info);
2000 			}
2001 			if (!ret && (list[ns].info.representor ^
2002 				     list[ns].info.master)) {
2003 				ns++;
2004 			} else if ((nd == 1) &&
2005 				   !list[ns].info.representor &&
2006 				   !list[ns].info.master) {
2007 				/*
2008 				 * Single IB device with
2009 				 * one physical port and
2010 				 * attached network device.
2011 				 * May be SRIOV is not enabled
2012 				 * or there is no representors.
2013 				 */
2014 				DRV_LOG(INFO, "no E-Switch support detected");
2015 				ns++;
2016 				break;
2017 			}
2018 		}
2019 		if (!ns) {
2020 			DRV_LOG(ERR,
2021 				"unable to recognize master/representors"
2022 				" on the multiple IB devices");
2023 			rte_errno = ENOENT;
2024 			ret = -rte_errno;
2025 			goto exit;
2026 		}
2027 	}
2028 	MLX5_ASSERT(ns);
2029 	/*
2030 	 * Sort list to probe devices in natural order for users convenience
2031 	 * (i.e. master first, then representors from lowest to highest ID).
2032 	 */
2033 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2034 	/* Device specific configuration. */
2035 	switch (pci_dev->id.device_id) {
2036 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2037 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2038 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2039 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2040 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2041 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2042 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2043 		dev_config_vf = 1;
2044 		break;
2045 	default:
2046 		dev_config_vf = 0;
2047 		break;
2048 	}
2049 	for (i = 0; i != ns; ++i) {
2050 		uint32_t restore;
2051 
2052 		/* Default configuration. */
2053 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2054 		dev_config.vf = dev_config_vf;
2055 		dev_config.mps = MLX5_ARG_UNSET;
2056 		dev_config.dbnc = MLX5_ARG_UNSET;
2057 		dev_config.rx_vec_en = 1;
2058 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2059 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2060 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2061 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2062 		dev_config.vf_nl_en = 1;
2063 		dev_config.mr_ext_memseg_en = 1;
2064 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2065 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2066 		dev_config.dv_esw_en = 1;
2067 		dev_config.dv_flow_en = 1;
2068 		dev_config.decap_en = 1;
2069 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2070 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2071 						 &list[i],
2072 						 &dev_config);
2073 		if (!list[i].eth_dev) {
2074 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2075 				break;
2076 			/* Device is disabled or already spawned. Ignore it. */
2077 			continue;
2078 		}
2079 		restore = list[i].eth_dev->data->dev_flags;
2080 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2081 		/* Restore non-PCI flags cleared by the above call. */
2082 		list[i].eth_dev->data->dev_flags |= restore;
2083 		rte_eth_dev_probing_finish(list[i].eth_dev);
2084 	}
2085 	if (i != ns) {
2086 		DRV_LOG(ERR,
2087 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2088 			" encountering an error: %s",
2089 			pci_dev->addr.domain, pci_dev->addr.bus,
2090 			pci_dev->addr.devid, pci_dev->addr.function,
2091 			strerror(rte_errno));
2092 		ret = -rte_errno;
2093 		/* Roll back. */
2094 		while (i--) {
2095 			if (!list[i].eth_dev)
2096 				continue;
2097 			mlx5_dev_close(list[i].eth_dev);
2098 			/* mac_addrs must not be freed because in dev_private */
2099 			list[i].eth_dev->data->mac_addrs = NULL;
2100 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2101 		}
2102 		/* Restore original error. */
2103 		rte_errno = -ret;
2104 	} else {
2105 		ret = 0;
2106 	}
2107 exit:
2108 	/*
2109 	 * Do the routine cleanup:
2110 	 * - close opened Netlink sockets
2111 	 * - free allocated spawn data array
2112 	 * - free the Infiniband device list
2113 	 */
2114 	if (nl_rdma >= 0)
2115 		close(nl_rdma);
2116 	if (nl_route >= 0)
2117 		close(nl_route);
2118 	if (list)
2119 		mlx5_free(list);
2120 	MLX5_ASSERT(ibv_list);
2121 	mlx5_glue->free_device_list(ibv_list);
2122 	return ret;
2123 }
2124 
2125 static int
2126 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2127 {
2128 	char *env;
2129 	int value;
2130 
2131 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2132 	/* Get environment variable to store. */
2133 	env = getenv(MLX5_SHUT_UP_BF);
2134 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2135 	if (config->dbnc == MLX5_ARG_UNSET)
2136 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2137 	else
2138 		setenv(MLX5_SHUT_UP_BF,
2139 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2140 	return value;
2141 }
2142 
2143 static void
2144 mlx5_restore_doorbell_mapping_env(int value)
2145 {
2146 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2147 	/* Restore the original environment variable state. */
2148 	if (value == MLX5_ARG_UNSET)
2149 		unsetenv(MLX5_SHUT_UP_BF);
2150 	else
2151 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2152 }
2153 
2154 /**
2155  * Extract pdn of PD object using DV API.
2156  *
2157  * @param[in] pd
2158  *   Pointer to the verbs PD object.
2159  * @param[out] pdn
2160  *   Pointer to the PD object number variable.
2161  *
2162  * @return
2163  *   0 on success, error value otherwise.
2164  */
2165 int
2166 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2167 {
2168 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2169 	struct mlx5dv_obj obj;
2170 	struct mlx5dv_pd pd_info;
2171 	int ret = 0;
2172 
2173 	obj.pd.in = pd;
2174 	obj.pd.out = &pd_info;
2175 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2176 	if (ret) {
2177 		DRV_LOG(DEBUG, "Fail to get PD object info");
2178 		return ret;
2179 	}
2180 	*pdn = pd_info.pdn;
2181 	return 0;
2182 #else
2183 	(void)pd;
2184 	(void)pdn;
2185 	return -ENOTSUP;
2186 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2187 }
2188 
2189 /**
2190  * Function API to open IB device.
2191  *
2192  * This function calls the Linux glue APIs to open a device.
2193  *
2194  * @param[in] spawn
2195  *   Pointer to the IB device attributes (name, port, etc).
2196  * @param[out] config
2197  *   Pointer to device configuration structure.
2198  * @param[out] sh
2199  *   Pointer to shared context structure.
2200  *
2201  * @return
2202  *   0 on success, a positive error value otherwise.
2203  */
2204 int
2205 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2206 		     const struct mlx5_dev_config *config,
2207 		     struct mlx5_dev_ctx_shared *sh)
2208 {
2209 	int dbmap_env;
2210 	int err = 0;
2211 
2212 	sh->numa_node = spawn->pci_dev->device.numa_node;
2213 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2214 	/*
2215 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2216 	 * before the device creation. The rdma_core library
2217 	 * checks the variable at device creation and
2218 	 * stores the result internally.
2219 	 */
2220 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2221 	/* Try to open IB device with DV first, then usual Verbs. */
2222 	errno = 0;
2223 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2224 	if (sh->ctx) {
2225 		sh->devx = 1;
2226 		DRV_LOG(DEBUG, "DevX is supported");
2227 		/* The device is created, no need for environment. */
2228 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2229 	} else {
2230 		/* The environment variable is still configured. */
2231 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2232 		err = errno ? errno : ENODEV;
2233 		/*
2234 		 * The environment variable is not needed anymore,
2235 		 * all device creation attempts are completed.
2236 		 */
2237 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2238 		if (!sh->ctx)
2239 			return err;
2240 		DRV_LOG(DEBUG, "DevX is NOT supported");
2241 		err = 0;
2242 	}
2243 	return err;
2244 }
2245 
2246 /**
2247  * Install shared asynchronous device events handler.
2248  * This function is implemented to support event sharing
2249  * between multiple ports of single IB device.
2250  *
2251  * @param sh
2252  *   Pointer to mlx5_dev_ctx_shared object.
2253  */
2254 void
2255 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2256 {
2257 	int ret;
2258 	int flags;
2259 
2260 	sh->intr_handle.fd = -1;
2261 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2262 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2263 		    F_SETFL, flags | O_NONBLOCK);
2264 	if (ret) {
2265 		DRV_LOG(INFO, "failed to change file descriptor async event"
2266 			" queue");
2267 	} else {
2268 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2269 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2270 		if (rte_intr_callback_register(&sh->intr_handle,
2271 					mlx5_dev_interrupt_handler, sh)) {
2272 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2273 			sh->intr_handle.fd = -1;
2274 		}
2275 	}
2276 	if (sh->devx) {
2277 #ifdef HAVE_IBV_DEVX_ASYNC
2278 		sh->intr_handle_devx.fd = -1;
2279 		sh->devx_comp =
2280 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2281 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2282 		if (!devx_comp) {
2283 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2284 			return;
2285 		}
2286 		flags = fcntl(devx_comp->fd, F_GETFL);
2287 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2288 		if (ret) {
2289 			DRV_LOG(INFO, "failed to change file descriptor"
2290 				" devx comp");
2291 			return;
2292 		}
2293 		sh->intr_handle_devx.fd = devx_comp->fd;
2294 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2295 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2296 					mlx5_dev_interrupt_handler_devx, sh)) {
2297 			DRV_LOG(INFO, "Fail to install the devx shared"
2298 				" interrupt.");
2299 			sh->intr_handle_devx.fd = -1;
2300 		}
2301 #endif /* HAVE_IBV_DEVX_ASYNC */
2302 	}
2303 }
2304 
2305 /**
2306  * Uninstall shared asynchronous device events handler.
2307  * This function is implemented to support event sharing
2308  * between multiple ports of single IB device.
2309  *
2310  * @param dev
2311  *   Pointer to mlx5_dev_ctx_shared object.
2312  */
2313 void
2314 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2315 {
2316 	if (sh->intr_handle.fd >= 0)
2317 		mlx5_intr_callback_unregister(&sh->intr_handle,
2318 					      mlx5_dev_interrupt_handler, sh);
2319 #ifdef HAVE_IBV_DEVX_ASYNC
2320 	if (sh->intr_handle_devx.fd >= 0)
2321 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2322 				  mlx5_dev_interrupt_handler_devx, sh);
2323 	if (sh->devx_comp)
2324 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2325 #endif
2326 }
2327 
2328 /**
2329  * Read statistics by a named counter.
2330  *
2331  * @param[in] priv
2332  *   Pointer to the private device data structure.
2333  * @param[in] ctr_name
2334  *   Pointer to the name of the statistic counter to read
2335  * @param[out] stat
2336  *   Pointer to read statistic value.
2337  * @return
2338  *   0 on success and stat is valud, 1 if failed to read the value
2339  *   rte_errno is set.
2340  *
2341  */
2342 int
2343 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2344 		      uint64_t *stat)
2345 {
2346 	int fd;
2347 
2348 	if (priv->sh) {
2349 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2350 		      priv->sh->ibdev_path,
2351 		      priv->dev_port,
2352 		      ctr_name);
2353 		fd = open(path, O_RDONLY);
2354 		/*
2355 		 * in switchdev the file location is not per port
2356 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2357 		 */
2358 		if (fd == -1) {
2359 			MKSTR(path1, "%s/hw_counters/%s",
2360 			      priv->sh->ibdev_path,
2361 			      ctr_name);
2362 			fd = open(path1, O_RDONLY);
2363 		}
2364 		if (fd != -1) {
2365 			char buf[21] = {'\0'};
2366 			ssize_t n = read(fd, buf, sizeof(buf));
2367 
2368 			close(fd);
2369 			if (n != -1) {
2370 				*stat = strtoull(buf, NULL, 10);
2371 				return 0;
2372 			}
2373 		}
2374 	}
2375 	*stat = 0;
2376 	return 1;
2377 }
2378 
2379 /**
2380  * Set the reg_mr and dereg_mr call backs
2381  *
2382  * @param reg_mr_cb[out]
2383  *   Pointer to reg_mr func
2384  * @param dereg_mr_cb[out]
2385  *   Pointer to dereg_mr func
2386  *
2387  */
2388 void
2389 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2390 		      mlx5_dereg_mr_t *dereg_mr_cb)
2391 {
2392 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2393 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2394 }
2395 
2396 /**
2397  * Remove a MAC address from device
2398  *
2399  * @param dev
2400  *   Pointer to Ethernet device structure.
2401  * @param index
2402  *   MAC address index.
2403  */
2404 void
2405 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2406 {
2407 	struct mlx5_priv *priv = dev->data->dev_private;
2408 	const int vf = priv->config.vf;
2409 
2410 	if (vf)
2411 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2412 					mlx5_ifindex(dev), priv->mac_own,
2413 					&dev->data->mac_addrs[index], index);
2414 }
2415 
2416 /**
2417  * Adds a MAC address to the device
2418  *
2419  * @param dev
2420  *   Pointer to Ethernet device structure.
2421  * @param mac_addr
2422  *   MAC address to register.
2423  * @param index
2424  *   MAC address index.
2425  *
2426  * @return
2427  *   0 on success, a negative errno value otherwise
2428  */
2429 int
2430 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2431 		     uint32_t index)
2432 {
2433 	struct mlx5_priv *priv = dev->data->dev_private;
2434 	const int vf = priv->config.vf;
2435 	int ret = 0;
2436 
2437 	if (vf)
2438 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2439 					   mlx5_ifindex(dev), priv->mac_own,
2440 					   mac, index);
2441 	return ret;
2442 }
2443 
2444 /**
2445  * Modify a VF MAC address
2446  *
2447  * @param priv
2448  *   Pointer to device private data.
2449  * @param mac_addr
2450  *   MAC address to modify into.
2451  * @param iface_idx
2452  *   Net device interface index
2453  * @param vf_index
2454  *   VF index
2455  *
2456  * @return
2457  *   0 on success, a negative errno value otherwise
2458  */
2459 int
2460 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2461 			   unsigned int iface_idx,
2462 			   struct rte_ether_addr *mac_addr,
2463 			   int vf_index)
2464 {
2465 	return mlx5_nl_vf_mac_addr_modify
2466 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2467 }
2468 
2469 /**
2470  * Set device promiscuous mode
2471  *
2472  * @param dev
2473  *   Pointer to Ethernet device structure.
2474  * @param enable
2475  *   0 - promiscuous is disabled, otherwise - enabled
2476  *
2477  * @return
2478  *   0 on success, a negative error value otherwise
2479  */
2480 int
2481 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2482 {
2483 	struct mlx5_priv *priv = dev->data->dev_private;
2484 
2485 	return mlx5_nl_promisc(priv->nl_socket_route,
2486 			       mlx5_ifindex(dev), !!enable);
2487 }
2488 
2489 /**
2490  * Set device promiscuous mode
2491  *
2492  * @param dev
2493  *   Pointer to Ethernet device structure.
2494  * @param enable
2495  *   0 - all multicase is disabled, otherwise - enabled
2496  *
2497  * @return
2498  *   0 on success, a negative error value otherwise
2499  */
2500 int
2501 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2502 {
2503 	struct mlx5_priv *priv = dev->data->dev_private;
2504 
2505 	return mlx5_nl_allmulti(priv->nl_socket_route,
2506 				mlx5_ifindex(dev), !!enable);
2507 }
2508 
2509 /**
2510  * Flush device MAC addresses
2511  *
2512  * @param dev
2513  *   Pointer to Ethernet device structure.
2514  *
2515  */
2516 void
2517 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2518 {
2519 	struct mlx5_priv *priv = dev->data->dev_private;
2520 
2521 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2522 			       dev->data->mac_addrs,
2523 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2524 }
2525 
2526 const struct eth_dev_ops mlx5_os_dev_ops = {
2527 	.dev_configure = mlx5_dev_configure,
2528 	.dev_start = mlx5_dev_start,
2529 	.dev_stop = mlx5_dev_stop,
2530 	.dev_set_link_down = mlx5_set_link_down,
2531 	.dev_set_link_up = mlx5_set_link_up,
2532 	.dev_close = mlx5_dev_close,
2533 	.promiscuous_enable = mlx5_promiscuous_enable,
2534 	.promiscuous_disable = mlx5_promiscuous_disable,
2535 	.allmulticast_enable = mlx5_allmulticast_enable,
2536 	.allmulticast_disable = mlx5_allmulticast_disable,
2537 	.link_update = mlx5_link_update,
2538 	.stats_get = mlx5_stats_get,
2539 	.stats_reset = mlx5_stats_reset,
2540 	.xstats_get = mlx5_xstats_get,
2541 	.xstats_reset = mlx5_xstats_reset,
2542 	.xstats_get_names = mlx5_xstats_get_names,
2543 	.fw_version_get = mlx5_fw_version_get,
2544 	.dev_infos_get = mlx5_dev_infos_get,
2545 	.read_clock = mlx5_txpp_read_clock,
2546 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2547 	.vlan_filter_set = mlx5_vlan_filter_set,
2548 	.rx_queue_setup = mlx5_rx_queue_setup,
2549 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2550 	.tx_queue_setup = mlx5_tx_queue_setup,
2551 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2552 	.rx_queue_release = mlx5_rx_queue_release,
2553 	.tx_queue_release = mlx5_tx_queue_release,
2554 	.rx_queue_start = mlx5_rx_queue_start,
2555 	.rx_queue_stop = mlx5_rx_queue_stop,
2556 	.tx_queue_start = mlx5_tx_queue_start,
2557 	.tx_queue_stop = mlx5_tx_queue_stop,
2558 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2559 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2560 	.mac_addr_remove = mlx5_mac_addr_remove,
2561 	.mac_addr_add = mlx5_mac_addr_add,
2562 	.mac_addr_set = mlx5_mac_addr_set,
2563 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2564 	.mtu_set = mlx5_dev_set_mtu,
2565 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2566 	.vlan_offload_set = mlx5_vlan_offload_set,
2567 	.reta_update = mlx5_dev_rss_reta_update,
2568 	.reta_query = mlx5_dev_rss_reta_query,
2569 	.rss_hash_update = mlx5_rss_hash_update,
2570 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2571 	.filter_ctrl = mlx5_dev_filter_ctrl,
2572 	.rxq_info_get = mlx5_rxq_info_get,
2573 	.txq_info_get = mlx5_txq_info_get,
2574 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2575 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2576 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2577 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2578 	.is_removed = mlx5_is_removed,
2579 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2580 	.get_module_info = mlx5_get_module_info,
2581 	.get_module_eeprom = mlx5_get_module_eeprom,
2582 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2583 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2584 	.hairpin_bind = mlx5_hairpin_bind,
2585 	.hairpin_unbind = mlx5_hairpin_unbind,
2586 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2587 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2588 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2589 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2590 };
2591 
2592 /* Available operations from secondary process. */
2593 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2594 	.stats_get = mlx5_stats_get,
2595 	.stats_reset = mlx5_stats_reset,
2596 	.xstats_get = mlx5_xstats_get,
2597 	.xstats_reset = mlx5_xstats_reset,
2598 	.xstats_get_names = mlx5_xstats_get_names,
2599 	.fw_version_get = mlx5_fw_version_get,
2600 	.dev_infos_get = mlx5_dev_infos_get,
2601 	.read_clock = mlx5_txpp_read_clock,
2602 	.rx_queue_start = mlx5_rx_queue_start,
2603 	.rx_queue_stop = mlx5_rx_queue_stop,
2604 	.tx_queue_start = mlx5_tx_queue_start,
2605 	.tx_queue_stop = mlx5_tx_queue_stop,
2606 	.rxq_info_get = mlx5_rxq_info_get,
2607 	.txq_info_get = mlx5_txq_info_get,
2608 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2609 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2610 	.get_module_info = mlx5_get_module_info,
2611 	.get_module_eeprom = mlx5_get_module_eeprom,
2612 };
2613 
2614 /* Available operations in flow isolated mode. */
2615 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2616 	.dev_configure = mlx5_dev_configure,
2617 	.dev_start = mlx5_dev_start,
2618 	.dev_stop = mlx5_dev_stop,
2619 	.dev_set_link_down = mlx5_set_link_down,
2620 	.dev_set_link_up = mlx5_set_link_up,
2621 	.dev_close = mlx5_dev_close,
2622 	.promiscuous_enable = mlx5_promiscuous_enable,
2623 	.promiscuous_disable = mlx5_promiscuous_disable,
2624 	.allmulticast_enable = mlx5_allmulticast_enable,
2625 	.allmulticast_disable = mlx5_allmulticast_disable,
2626 	.link_update = mlx5_link_update,
2627 	.stats_get = mlx5_stats_get,
2628 	.stats_reset = mlx5_stats_reset,
2629 	.xstats_get = mlx5_xstats_get,
2630 	.xstats_reset = mlx5_xstats_reset,
2631 	.xstats_get_names = mlx5_xstats_get_names,
2632 	.fw_version_get = mlx5_fw_version_get,
2633 	.dev_infos_get = mlx5_dev_infos_get,
2634 	.read_clock = mlx5_txpp_read_clock,
2635 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2636 	.vlan_filter_set = mlx5_vlan_filter_set,
2637 	.rx_queue_setup = mlx5_rx_queue_setup,
2638 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2639 	.tx_queue_setup = mlx5_tx_queue_setup,
2640 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2641 	.rx_queue_release = mlx5_rx_queue_release,
2642 	.tx_queue_release = mlx5_tx_queue_release,
2643 	.rx_queue_start = mlx5_rx_queue_start,
2644 	.rx_queue_stop = mlx5_rx_queue_stop,
2645 	.tx_queue_start = mlx5_tx_queue_start,
2646 	.tx_queue_stop = mlx5_tx_queue_stop,
2647 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2648 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2649 	.mac_addr_remove = mlx5_mac_addr_remove,
2650 	.mac_addr_add = mlx5_mac_addr_add,
2651 	.mac_addr_set = mlx5_mac_addr_set,
2652 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2653 	.mtu_set = mlx5_dev_set_mtu,
2654 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2655 	.vlan_offload_set = mlx5_vlan_offload_set,
2656 	.filter_ctrl = mlx5_dev_filter_ctrl,
2657 	.rxq_info_get = mlx5_rxq_info_get,
2658 	.txq_info_get = mlx5_txq_info_get,
2659 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2660 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2661 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2662 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2663 	.is_removed = mlx5_is_removed,
2664 	.get_module_info = mlx5_get_module_info,
2665 	.get_module_eeprom = mlx5_get_module_eeprom,
2666 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2667 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2668 	.hairpin_bind = mlx5_hairpin_bind,
2669 	.hairpin_unbind = mlx5_hairpin_unbind,
2670 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
2671 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
2672 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
2673 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
2674 };
2675