xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 994829e695c075d38294d186338a2fd00a3a2c4e)
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 = 0;
230 
231 	if (!sh->flow_tbls)
232 		err = mlx5_alloc_table_hash_list(priv);
233 	else
234 		DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
235 			(void *)sh->flow_tbls);
236 	if (err)
237 		return err;
238 	/* Create tags hash list table. */
239 	snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
240 	sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
241 	if (!sh->tag_table) {
242 		DRV_LOG(ERR, "tags with hash creation failed.");
243 		err = ENOMEM;
244 		goto error;
245 	}
246 	snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
247 	sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ);
248 	if (!sh->modify_cmds) {
249 		DRV_LOG(ERR, "hdr modify hash creation failed");
250 		err = ENOMEM;
251 		goto error;
252 	}
253 	snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
254 	sh->encaps_decaps = mlx5_hlist_create(s,
255 					      MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ);
256 	if (!sh->encaps_decaps) {
257 		DRV_LOG(ERR, "encap decap hash creation failed");
258 		err = ENOMEM;
259 		goto error;
260 	}
261 #ifdef HAVE_MLX5DV_DR
262 	void *domain;
263 
264 	if (sh->dv_refcnt) {
265 		/* Shared DV/DR structures is already initialized. */
266 		sh->dv_refcnt++;
267 		priv->dr_shared = 1;
268 		return 0;
269 	}
270 	/* Reference counter is zero, we should initialize structures. */
271 	domain = mlx5_glue->dr_create_domain(sh->ctx,
272 					     MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
273 	if (!domain) {
274 		DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
275 		err = errno;
276 		goto error;
277 	}
278 	sh->rx_domain = domain;
279 	domain = mlx5_glue->dr_create_domain(sh->ctx,
280 					     MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
281 	if (!domain) {
282 		DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
283 		err = errno;
284 		goto error;
285 	}
286 	pthread_mutex_init(&sh->dv_mutex, NULL);
287 	sh->tx_domain = domain;
288 #ifdef HAVE_MLX5DV_DR_ESWITCH
289 	if (priv->config.dv_esw_en) {
290 		domain  = mlx5_glue->dr_create_domain
291 			(sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
292 		if (!domain) {
293 			DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
294 			err = errno;
295 			goto error;
296 		}
297 		sh->fdb_domain = domain;
298 		sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
299 	}
300 #endif
301 	if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
302 		mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
303 		mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
304 		if (sh->fdb_domain)
305 			mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
306 	}
307 	sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
308 #endif /* HAVE_MLX5DV_DR */
309 	sh->dv_refcnt++;
310 	priv->dr_shared = 1;
311 	return 0;
312 error:
313 	/* Rollback the created objects. */
314 	if (sh->rx_domain) {
315 		mlx5_glue->dr_destroy_domain(sh->rx_domain);
316 		sh->rx_domain = NULL;
317 	}
318 	if (sh->tx_domain) {
319 		mlx5_glue->dr_destroy_domain(sh->tx_domain);
320 		sh->tx_domain = NULL;
321 	}
322 	if (sh->fdb_domain) {
323 		mlx5_glue->dr_destroy_domain(sh->fdb_domain);
324 		sh->fdb_domain = NULL;
325 	}
326 	if (sh->esw_drop_action) {
327 		mlx5_glue->destroy_flow_action(sh->esw_drop_action);
328 		sh->esw_drop_action = NULL;
329 	}
330 	if (sh->pop_vlan_action) {
331 		mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
332 		sh->pop_vlan_action = NULL;
333 	}
334 	if (sh->encaps_decaps) {
335 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
336 		sh->encaps_decaps = NULL;
337 	}
338 	if (sh->modify_cmds) {
339 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
340 		sh->modify_cmds = NULL;
341 	}
342 	if (sh->tag_table) {
343 		/* tags should be destroyed with flow before. */
344 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
345 		sh->tag_table = NULL;
346 	}
347 	mlx5_free_table_hash_list(priv);
348 	return err;
349 }
350 
351 /**
352  * Destroy DR related data within private structure.
353  *
354  * @param[in] priv
355  *   Pointer to the private device data structure.
356  */
357 void
358 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
359 {
360 	struct mlx5_dev_ctx_shared *sh;
361 
362 	if (!priv->dr_shared)
363 		return;
364 	priv->dr_shared = 0;
365 	sh = priv->sh;
366 	MLX5_ASSERT(sh);
367 #ifdef HAVE_MLX5DV_DR
368 	MLX5_ASSERT(sh->dv_refcnt);
369 	if (sh->dv_refcnt && --sh->dv_refcnt)
370 		return;
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->encaps_decaps) {
396 		mlx5_hlist_destroy(sh->encaps_decaps, NULL, NULL);
397 		sh->encaps_decaps = NULL;
398 	}
399 	if (sh->modify_cmds) {
400 		mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
401 		sh->modify_cmds = NULL;
402 	}
403 	if (sh->tag_table) {
404 		/* tags should be destroyed with flow before. */
405 		mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
406 		sh->tag_table = NULL;
407 	}
408 	mlx5_free_table_hash_list(priv);
409 }
410 
411 /**
412  * Initialize shared data between primary and secondary process.
413  *
414  * A memzone is reserved by primary process and secondary processes attach to
415  * the memzone.
416  *
417  * @return
418  *   0 on success, a negative errno value otherwise and rte_errno is set.
419  */
420 static int
421 mlx5_init_shared_data(void)
422 {
423 	const struct rte_memzone *mz;
424 	int ret = 0;
425 
426 	rte_spinlock_lock(&mlx5_shared_data_lock);
427 	if (mlx5_shared_data == NULL) {
428 		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
429 			/* Allocate shared memory. */
430 			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
431 						 sizeof(*mlx5_shared_data),
432 						 SOCKET_ID_ANY, 0);
433 			if (mz == NULL) {
434 				DRV_LOG(ERR,
435 					"Cannot allocate mlx5 shared data");
436 				ret = -rte_errno;
437 				goto error;
438 			}
439 			mlx5_shared_data = mz->addr;
440 			memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
441 			rte_spinlock_init(&mlx5_shared_data->lock);
442 		} else {
443 			/* Lookup allocated shared memory. */
444 			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
445 			if (mz == NULL) {
446 				DRV_LOG(ERR,
447 					"Cannot attach mlx5 shared data");
448 				ret = -rte_errno;
449 				goto error;
450 			}
451 			mlx5_shared_data = mz->addr;
452 			memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
453 		}
454 	}
455 error:
456 	rte_spinlock_unlock(&mlx5_shared_data_lock);
457 	return ret;
458 }
459 
460 /**
461  * PMD global initialization.
462  *
463  * Independent from individual device, this function initializes global
464  * per-PMD data structures distinguishing primary and secondary processes.
465  * Hence, each initialization is called once per a process.
466  *
467  * @return
468  *   0 on success, a negative errno value otherwise and rte_errno is set.
469  */
470 static int
471 mlx5_init_once(void)
472 {
473 	struct mlx5_shared_data *sd;
474 	struct mlx5_local_data *ld = &mlx5_local_data;
475 	int ret = 0;
476 
477 	if (mlx5_init_shared_data())
478 		return -rte_errno;
479 	sd = mlx5_shared_data;
480 	MLX5_ASSERT(sd);
481 	rte_spinlock_lock(&sd->lock);
482 	switch (rte_eal_process_type()) {
483 	case RTE_PROC_PRIMARY:
484 		if (sd->init_done)
485 			break;
486 		LIST_INIT(&sd->mem_event_cb_list);
487 		rte_rwlock_init(&sd->mem_event_rwlock);
488 		rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
489 						mlx5_mr_mem_event_cb, NULL);
490 		ret = mlx5_mp_init_primary(MLX5_MP_NAME,
491 					   mlx5_mp_os_primary_handle);
492 		if (ret)
493 			goto out;
494 		sd->init_done = true;
495 		break;
496 	case RTE_PROC_SECONDARY:
497 		if (ld->init_done)
498 			break;
499 		ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
500 					     mlx5_mp_os_secondary_handle);
501 		if (ret)
502 			goto out;
503 		++sd->secondary_cnt;
504 		ld->init_done = true;
505 		break;
506 	default:
507 		break;
508 	}
509 out:
510 	rte_spinlock_unlock(&sd->lock);
511 	return ret;
512 }
513 
514 /**
515  * Create the Tx queue DevX/Verbs object.
516  *
517  * @param dev
518  *   Pointer to Ethernet device.
519  * @param idx
520  *   Queue index in DPDK Tx queue array.
521  *
522  * @return
523  *   0 on success, a negative errno value otherwise and rte_errno is set.
524  */
525 static int
526 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
527 {
528 	struct mlx5_priv *priv = dev->data->dev_private;
529 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
530 	struct mlx5_txq_ctrl *txq_ctrl =
531 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
532 
533 	if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
534 		return mlx5_txq_devx_obj_new(dev, idx);
535 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
536 	if (!priv->config.dv_esw_en)
537 		return mlx5_txq_devx_obj_new(dev, idx);
538 #endif
539 	return mlx5_txq_ibv_obj_new(dev, idx);
540 }
541 
542 /**
543  * Release an Tx DevX/verbs queue object.
544  *
545  * @param txq_obj
546  *   DevX/Verbs Tx queue object.
547  */
548 static void
549 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
550 {
551 	if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
552 		mlx5_txq_devx_obj_release(txq_obj);
553 		return;
554 	}
555 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
556 	if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
557 		mlx5_txq_devx_obj_release(txq_obj);
558 		return;
559 	}
560 #endif
561 	mlx5_txq_ibv_obj_release(txq_obj);
562 }
563 
564 /**
565  * DV flow counter mode detect and config.
566  *
567  * @param dev
568  *   Pointer to rte_eth_dev structure.
569  *
570  */
571 static void
572 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
573 {
574 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
575 	struct mlx5_priv *priv = dev->data->dev_private;
576 
577 	/* If devx is not supported or not DV mode, counters are not working. */
578 	if (!priv->config.devx || !priv->config.dv_flow_en)
579 		return;
580 #ifndef HAVE_IBV_DEVX_ASYNC
581 	priv->counter_fallback = 1;
582 #else
583 	priv->counter_fallback = 0;
584 	if (!priv->config.hca_attr.flow_counters_dump ||
585 	    !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
586 	    (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
587 		priv->counter_fallback = 1;
588 #endif
589 	if (priv->counter_fallback)
590 		DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
591 			"counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
592 			priv->config.hca_attr.flow_counters_dump,
593 			priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
594 #endif
595 }
596 
597 /**
598  * Spawn an Ethernet device from Verbs information.
599  *
600  * @param dpdk_dev
601  *   Backing DPDK device.
602  * @param spawn
603  *   Verbs device parameters (name, port, switch_info) to spawn.
604  * @param config
605  *   Device configuration parameters.
606  *
607  * @return
608  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
609  *   is set. The following errors are defined:
610  *
611  *   EBUSY: device is not supposed to be spawned.
612  *   EEXIST: device is already spawned
613  */
614 static struct rte_eth_dev *
615 mlx5_dev_spawn(struct rte_device *dpdk_dev,
616 	       struct mlx5_dev_spawn_data *spawn,
617 	       struct mlx5_dev_config *config)
618 {
619 	const struct mlx5_switch_info *switch_info = &spawn->info;
620 	struct mlx5_dev_ctx_shared *sh = NULL;
621 	struct ibv_port_attr port_attr;
622 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
623 	struct rte_eth_dev *eth_dev = NULL;
624 	struct mlx5_priv *priv = NULL;
625 	int err = 0;
626 	unsigned int hw_padding = 0;
627 	unsigned int mps;
628 	unsigned int cqe_comp;
629 	unsigned int cqe_pad = 0;
630 	unsigned int tunnel_en = 0;
631 	unsigned int mpls_en = 0;
632 	unsigned int swp = 0;
633 	unsigned int mprq = 0;
634 	unsigned int mprq_min_stride_size_n = 0;
635 	unsigned int mprq_max_stride_size_n = 0;
636 	unsigned int mprq_min_stride_num_n = 0;
637 	unsigned int mprq_max_stride_num_n = 0;
638 	struct rte_ether_addr mac;
639 	char name[RTE_ETH_NAME_MAX_LEN];
640 	int own_domain_id = 0;
641 	uint16_t port_id;
642 	unsigned int i;
643 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
644 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
645 #endif
646 
647 	/* Determine if this port representor is supposed to be spawned. */
648 	if (switch_info->representor && dpdk_dev->devargs) {
649 		struct rte_eth_devargs eth_da;
650 
651 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
652 		if (err) {
653 			rte_errno = -err;
654 			DRV_LOG(ERR, "failed to process device arguments: %s",
655 				strerror(rte_errno));
656 			return NULL;
657 		}
658 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
659 			if (eth_da.representor_ports[i] ==
660 			    (uint16_t)switch_info->port_name)
661 				break;
662 		if (i == eth_da.nb_representor_ports) {
663 			rte_errno = EBUSY;
664 			return NULL;
665 		}
666 	}
667 	/* Build device name. */
668 	if (spawn->pf_bond <  0) {
669 		/* Single device. */
670 		if (!switch_info->representor)
671 			strlcpy(name, dpdk_dev->name, sizeof(name));
672 		else
673 			snprintf(name, sizeof(name), "%s_representor_%u",
674 				 dpdk_dev->name, switch_info->port_name);
675 	} else {
676 		/* Bonding device. */
677 		if (!switch_info->representor)
678 			snprintf(name, sizeof(name), "%s_%s",
679 				 dpdk_dev->name,
680 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
681 		else
682 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
683 				 dpdk_dev->name,
684 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
685 				 switch_info->port_name);
686 	}
687 	/* check if the device is already spawned */
688 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
689 		rte_errno = EEXIST;
690 		return NULL;
691 	}
692 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
693 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
694 		struct mlx5_mp_id mp_id;
695 
696 		eth_dev = rte_eth_dev_attach_secondary(name);
697 		if (eth_dev == NULL) {
698 			DRV_LOG(ERR, "can not attach rte ethdev");
699 			rte_errno = ENOMEM;
700 			return NULL;
701 		}
702 		eth_dev->device = dpdk_dev;
703 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
704 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
705 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
706 		err = mlx5_proc_priv_init(eth_dev);
707 		if (err)
708 			return NULL;
709 		mp_id.port_id = eth_dev->data->port_id;
710 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
711 		/* Receive command fd from primary process */
712 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
713 		if (err < 0)
714 			goto err_secondary;
715 		/* Remap UAR for Tx queues. */
716 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
717 		if (err)
718 			goto err_secondary;
719 		/*
720 		 * Ethdev pointer is still required as input since
721 		 * the primary device is not accessible from the
722 		 * secondary process.
723 		 */
724 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
725 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
726 		return eth_dev;
727 err_secondary:
728 		mlx5_dev_close(eth_dev);
729 		return NULL;
730 	}
731 	/*
732 	 * Some parameters ("tx_db_nc" in particularly) are needed in
733 	 * advance to create dv/verbs device context. We proceed the
734 	 * devargs here to get ones, and later proceed devargs again
735 	 * to override some hardware settings.
736 	 */
737 	err = mlx5_args(config, dpdk_dev->devargs);
738 	if (err) {
739 		err = rte_errno;
740 		DRV_LOG(ERR, "failed to process device arguments: %s",
741 			strerror(rte_errno));
742 		goto error;
743 	}
744 	mlx5_malloc_mem_select(config->sys_mem_en);
745 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
746 	if (!sh)
747 		return NULL;
748 	config->devx = sh->devx;
749 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
750 	config->dest_tir = 1;
751 #endif
752 #ifdef HAVE_IBV_MLX5_MOD_SWP
753 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
754 #endif
755 	/*
756 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
757 	 * as all ConnectX-5 devices.
758 	 */
759 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
760 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
761 #endif
762 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
763 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
764 #endif
765 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
766 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
767 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
768 			DRV_LOG(DEBUG, "enhanced MPW is supported");
769 			mps = MLX5_MPW_ENHANCED;
770 		} else {
771 			DRV_LOG(DEBUG, "MPW is supported");
772 			mps = MLX5_MPW;
773 		}
774 	} else {
775 		DRV_LOG(DEBUG, "MPW isn't supported");
776 		mps = MLX5_MPW_DISABLED;
777 	}
778 #ifdef HAVE_IBV_MLX5_MOD_SWP
779 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
780 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
781 	DRV_LOG(DEBUG, "SWP support: %u", swp);
782 #endif
783 	config->swp = !!swp;
784 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
785 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
786 		struct mlx5dv_striding_rq_caps mprq_caps =
787 			dv_attr.striding_rq_caps;
788 
789 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
790 			mprq_caps.min_single_stride_log_num_of_bytes);
791 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
792 			mprq_caps.max_single_stride_log_num_of_bytes);
793 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
794 			mprq_caps.min_single_wqe_log_num_of_strides);
795 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
796 			mprq_caps.max_single_wqe_log_num_of_strides);
797 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
798 			mprq_caps.supported_qpts);
799 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
800 		mprq = 1;
801 		mprq_min_stride_size_n =
802 			mprq_caps.min_single_stride_log_num_of_bytes;
803 		mprq_max_stride_size_n =
804 			mprq_caps.max_single_stride_log_num_of_bytes;
805 		mprq_min_stride_num_n =
806 			mprq_caps.min_single_wqe_log_num_of_strides;
807 		mprq_max_stride_num_n =
808 			mprq_caps.max_single_wqe_log_num_of_strides;
809 	}
810 #endif
811 	if (RTE_CACHE_LINE_SIZE == 128 &&
812 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
813 		cqe_comp = 0;
814 	else
815 		cqe_comp = 1;
816 	config->cqe_comp = cqe_comp;
817 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
818 	/* Whether device supports 128B Rx CQE padding. */
819 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
820 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
821 #endif
822 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
823 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
824 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
825 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
826 			     (dv_attr.tunnel_offloads_caps &
827 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
828 			     (dv_attr.tunnel_offloads_caps &
829 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
830 	}
831 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
832 		tunnel_en ? "" : "not ");
833 #else
834 	DRV_LOG(WARNING,
835 		"tunnel offloading disabled due to old OFED/rdma-core version");
836 #endif
837 	config->tunnel_en = tunnel_en;
838 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
839 	mpls_en = ((dv_attr.tunnel_offloads_caps &
840 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
841 		   (dv_attr.tunnel_offloads_caps &
842 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
843 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
844 		mpls_en ? "" : "not ");
845 #else
846 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
847 		" old OFED/rdma-core version or firmware configuration");
848 #endif
849 	config->mpls_en = mpls_en;
850 	/* Check port status. */
851 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
852 	if (err) {
853 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
854 		goto error;
855 	}
856 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
857 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
858 		err = EINVAL;
859 		goto error;
860 	}
861 	if (port_attr.state != IBV_PORT_ACTIVE)
862 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
863 			mlx5_glue->port_state_str(port_attr.state),
864 			port_attr.state);
865 	/* Allocate private eth device data. */
866 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
867 			   sizeof(*priv),
868 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
869 	if (priv == NULL) {
870 		DRV_LOG(ERR, "priv allocation failure");
871 		err = ENOMEM;
872 		goto error;
873 	}
874 	priv->sh = sh;
875 	priv->dev_port = spawn->phys_port;
876 	priv->pci_dev = spawn->pci_dev;
877 	priv->mtu = RTE_ETHER_MTU;
878 	priv->mp_id.port_id = port_id;
879 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
880 	/* Some internal functions rely on Netlink sockets, open them now. */
881 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
882 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
883 	priv->representor = !!switch_info->representor;
884 	priv->master = !!switch_info->master;
885 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
886 	priv->vport_meta_tag = 0;
887 	priv->vport_meta_mask = 0;
888 	priv->pf_bond = spawn->pf_bond;
889 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
890 	/*
891 	 * The DevX port query API is implemented. E-Switch may use
892 	 * either vport or reg_c[0] metadata register to match on
893 	 * vport index. The engaged part of metadata register is
894 	 * defined by mask.
895 	 */
896 	if (switch_info->representor || switch_info->master) {
897 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
898 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
899 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
900 						 &devx_port);
901 		if (err) {
902 			DRV_LOG(WARNING,
903 				"can't query devx port %d on device %s",
904 				spawn->phys_port,
905 				mlx5_os_get_dev_device_name(spawn->phys_dev));
906 			devx_port.comp_mask = 0;
907 		}
908 	}
909 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
910 		priv->vport_meta_tag = devx_port.reg_c_0.value;
911 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
912 		if (!priv->vport_meta_mask) {
913 			DRV_LOG(ERR, "vport zero mask for port %d"
914 				     " on bonding device %s",
915 				     spawn->phys_port,
916 				     mlx5_os_get_dev_device_name
917 							(spawn->phys_dev));
918 			err = ENOTSUP;
919 			goto error;
920 		}
921 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
922 			DRV_LOG(ERR, "invalid vport tag for port %d"
923 				     " on bonding device %s",
924 				     spawn->phys_port,
925 				     mlx5_os_get_dev_device_name
926 							(spawn->phys_dev));
927 			err = ENOTSUP;
928 			goto error;
929 		}
930 	}
931 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
932 		priv->vport_id = devx_port.vport_num;
933 	} else if (spawn->pf_bond >= 0) {
934 		DRV_LOG(ERR, "can't deduce vport index for port %d"
935 			     " on bonding device %s",
936 			     spawn->phys_port,
937 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
938 		err = ENOTSUP;
939 		goto error;
940 	} else {
941 		/* Suppose vport index in compatible way. */
942 		priv->vport_id = switch_info->representor ?
943 				 switch_info->port_name + 1 : -1;
944 	}
945 #else
946 	/*
947 	 * Kernel/rdma_core support single E-Switch per PF configurations
948 	 * only and vport_id field contains the vport index for
949 	 * associated VF, which is deduced from representor port name.
950 	 * For example, let's have the IB device port 10, it has
951 	 * attached network device eth0, which has port name attribute
952 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
953 	 * as 3 (2+1). This assigning schema should be changed if the
954 	 * multiple E-Switch instances per PF configurations or/and PCI
955 	 * subfunctions are added.
956 	 */
957 	priv->vport_id = switch_info->representor ?
958 			 switch_info->port_name + 1 : -1;
959 #endif
960 	/* representor_id field keeps the unmodified VF index. */
961 	priv->representor_id = switch_info->representor ?
962 			       switch_info->port_name : -1;
963 	/*
964 	 * Look for sibling devices in order to reuse their switch domain
965 	 * if any, otherwise allocate one.
966 	 */
967 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
968 		const struct mlx5_priv *opriv =
969 			rte_eth_devices[port_id].data->dev_private;
970 
971 		if (!opriv ||
972 		    opriv->sh != priv->sh ||
973 			opriv->domain_id ==
974 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
975 			continue;
976 		priv->domain_id = opriv->domain_id;
977 		break;
978 	}
979 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
980 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
981 		if (err) {
982 			err = rte_errno;
983 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
984 				strerror(rte_errno));
985 			goto error;
986 		}
987 		own_domain_id = 1;
988 	}
989 	/* Override some values set by hardware configuration. */
990 	mlx5_args(config, dpdk_dev->devargs);
991 	err = mlx5_dev_check_sibling_config(priv, config);
992 	if (err)
993 		goto error;
994 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
995 			    IBV_DEVICE_RAW_IP_CSUM);
996 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
997 		(config->hw_csum ? "" : "not "));
998 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
999 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1000 	DRV_LOG(DEBUG, "counters are not supported");
1001 #endif
1002 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1003 	if (config->dv_flow_en) {
1004 		DRV_LOG(WARNING, "DV flow is not supported");
1005 		config->dv_flow_en = 0;
1006 	}
1007 #endif
1008 	config->ind_table_max_size =
1009 		sh->device_attr.max_rwq_indirection_table_size;
1010 	/*
1011 	 * Remove this check once DPDK supports larger/variable
1012 	 * indirection tables.
1013 	 */
1014 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1015 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1016 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1017 		config->ind_table_max_size);
1018 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1019 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1020 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1021 		(config->hw_vlan_strip ? "" : "not "));
1022 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1023 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
1024 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1025 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1026 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1027 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1028 			IBV_DEVICE_PCI_WRITE_END_PADDING);
1029 #endif
1030 	if (config->hw_padding && !hw_padding) {
1031 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1032 		config->hw_padding = 0;
1033 	} else if (config->hw_padding) {
1034 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1035 	}
1036 	config->tso = (sh->device_attr.max_tso > 0 &&
1037 		      (sh->device_attr.tso_supported_qpts &
1038 		       (1 << IBV_QPT_RAW_PACKET)));
1039 	if (config->tso)
1040 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1041 	/*
1042 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1043 	 * by default.
1044 	 */
1045 	if (config->mps == MLX5_ARG_UNSET)
1046 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1047 							  MLX5_MPW_DISABLED;
1048 	else
1049 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1050 	DRV_LOG(INFO, "%sMPS is %s",
1051 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1052 		config->mps == MLX5_MPW ? "legacy " : "",
1053 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1054 	if (config->cqe_comp && !cqe_comp) {
1055 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1056 		config->cqe_comp = 0;
1057 	}
1058 	if (config->cqe_pad && !cqe_pad) {
1059 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1060 		config->cqe_pad = 0;
1061 	} else if (config->cqe_pad) {
1062 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1063 	}
1064 	if (config->devx) {
1065 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1066 		if (err) {
1067 			err = -err;
1068 			goto error;
1069 		}
1070 		/* Check for LRO support. */
1071 		if (config->dest_tir && config->hca_attr.lro_cap &&
1072 		    config->dv_flow_en) {
1073 			/* TBD check tunnel lro caps. */
1074 			config->lro.supported = config->hca_attr.lro_cap;
1075 			DRV_LOG(DEBUG, "Device supports LRO");
1076 			/*
1077 			 * If LRO timeout is not configured by application,
1078 			 * use the minimal supported value.
1079 			 */
1080 			if (!config->lro.timeout)
1081 				config->lro.timeout =
1082 				config->hca_attr.lro_timer_supported_periods[0];
1083 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1084 				config->lro.timeout);
1085 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1086 				"required for coalescing is %d bytes",
1087 				config->hca_attr.lro_min_mss_size);
1088 		}
1089 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1090 		if (config->hca_attr.qos.sup &&
1091 		    config->hca_attr.qos.srtcm_sup &&
1092 		    config->dv_flow_en) {
1093 			uint8_t reg_c_mask =
1094 				config->hca_attr.qos.flow_meter_reg_c_ids;
1095 			/*
1096 			 * Meter needs two REG_C's for color match and pre-sfx
1097 			 * flow match. Here get the REG_C for color match.
1098 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1099 			 */
1100 			reg_c_mask &= 0xfc;
1101 			if (__builtin_popcount(reg_c_mask) < 1) {
1102 				priv->mtr_en = 0;
1103 				DRV_LOG(WARNING, "No available register for"
1104 					" meter.");
1105 			} else {
1106 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1107 						      REG_C_0;
1108 				priv->mtr_en = 1;
1109 				priv->mtr_reg_share =
1110 				      config->hca_attr.qos.flow_meter_reg_share;
1111 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1112 					priv->mtr_color_reg);
1113 			}
1114 		}
1115 #endif
1116 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1117 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1118 		    config->dv_flow_en) {
1119 			priv->sampler_en = 1;
1120 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1121 		} else {
1122 			priv->sampler_en = 0;
1123 			if (!config->hca_attr.log_max_ft_sampler_num)
1124 				DRV_LOG(WARNING, "No available register for"
1125 						" Sampler.");
1126 			else
1127 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1128 		}
1129 #endif
1130 	}
1131 	if (config->tx_pp) {
1132 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1133 			config->hca_attr.dev_freq_khz);
1134 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1135 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1136 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1137 			config->hca_attr.cross_channel ? "" : "not ");
1138 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1139 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1140 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1141 			config->hca_attr.non_wire_sq ? "" : "not ");
1142 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1143 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1144 			config->hca_attr.log_max_static_sq_wq);
1145 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1146 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1147 		if (!config->devx) {
1148 			DRV_LOG(ERR, "DevX is required for packet pacing");
1149 			err = ENODEV;
1150 			goto error;
1151 		}
1152 		if (!config->hca_attr.qos.packet_pacing) {
1153 			DRV_LOG(ERR, "Packet pacing is not supported");
1154 			err = ENODEV;
1155 			goto error;
1156 		}
1157 		if (!config->hca_attr.cross_channel) {
1158 			DRV_LOG(ERR, "Cross channel operations are"
1159 				     " required for packet pacing");
1160 			err = ENODEV;
1161 			goto error;
1162 		}
1163 		if (!config->hca_attr.wqe_index_ignore) {
1164 			DRV_LOG(ERR, "WQE index ignore feature is"
1165 				     " required for packet pacing");
1166 			err = ENODEV;
1167 			goto error;
1168 		}
1169 		if (!config->hca_attr.non_wire_sq) {
1170 			DRV_LOG(ERR, "Non-wire SQ feature is"
1171 				     " required for packet pacing");
1172 			err = ENODEV;
1173 			goto error;
1174 		}
1175 		if (!config->hca_attr.log_max_static_sq_wq) {
1176 			DRV_LOG(ERR, "Static WQE SQ feature is"
1177 				     " required for packet pacing");
1178 			err = ENODEV;
1179 			goto error;
1180 		}
1181 		if (!config->hca_attr.qos.wqe_rate_pp) {
1182 			DRV_LOG(ERR, "WQE rate mode is required"
1183 				     " for packet pacing");
1184 			err = ENODEV;
1185 			goto error;
1186 		}
1187 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1188 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1189 			     " can't create queues for packet pacing");
1190 		err = ENODEV;
1191 		goto error;
1192 #endif
1193 	}
1194 	if (config->devx) {
1195 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1196 
1197 		err = config->hca_attr.access_register_user ?
1198 			mlx5_devx_cmd_register_read
1199 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1200 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1201 		if (!err) {
1202 			uint32_t ts_mode;
1203 
1204 			/* MTUTC register is read successfully. */
1205 			ts_mode = MLX5_GET(register_mtutc, reg,
1206 					   time_stamp_mode);
1207 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1208 				config->rt_timestamp = 1;
1209 		} else {
1210 			/* Kernel does not support register reading. */
1211 			if (config->hca_attr.dev_freq_khz ==
1212 						 (NS_PER_S / MS_PER_S))
1213 				config->rt_timestamp = 1;
1214 		}
1215 	}
1216 	/*
1217 	 * If HW has bug working with tunnel packet decapsulation and
1218 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1219 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1220 	 */
1221 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1222 		config->hw_fcs_strip = 0;
1223 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1224 		(config->hw_fcs_strip ? "" : "not "));
1225 	if (config->mprq.enabled && mprq) {
1226 		if (config->mprq.stride_num_n &&
1227 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1228 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1229 			config->mprq.stride_num_n =
1230 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1231 						mprq_min_stride_num_n),
1232 					mprq_max_stride_num_n);
1233 			DRV_LOG(WARNING,
1234 				"the number of strides"
1235 				" for Multi-Packet RQ is out of range,"
1236 				" setting default value (%u)",
1237 				1 << config->mprq.stride_num_n);
1238 		}
1239 		if (config->mprq.stride_size_n &&
1240 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1241 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1242 			config->mprq.stride_size_n =
1243 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1244 						mprq_min_stride_size_n),
1245 					mprq_max_stride_size_n);
1246 			DRV_LOG(WARNING,
1247 				"the size of a stride"
1248 				" for Multi-Packet RQ is out of range,"
1249 				" setting default value (%u)",
1250 				1 << config->mprq.stride_size_n);
1251 		}
1252 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1253 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1254 	} else if (config->mprq.enabled && !mprq) {
1255 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1256 		config->mprq.enabled = 0;
1257 	}
1258 	if (config->max_dump_files_num == 0)
1259 		config->max_dump_files_num = 128;
1260 	eth_dev = rte_eth_dev_allocate(name);
1261 	if (eth_dev == NULL) {
1262 		DRV_LOG(ERR, "can not allocate rte ethdev");
1263 		err = ENOMEM;
1264 		goto error;
1265 	}
1266 	if (priv->representor) {
1267 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1268 		eth_dev->data->representor_id = priv->representor_id;
1269 	}
1270 	/*
1271 	 * Store associated network device interface index. This index
1272 	 * is permanent throughout the lifetime of device. So, we may store
1273 	 * the ifindex here and use the cached value further.
1274 	 */
1275 	MLX5_ASSERT(spawn->ifindex);
1276 	priv->if_index = spawn->ifindex;
1277 	if (priv->pf_bond >= 0 && priv->master) {
1278 		/* Get bond interface info */
1279 		err = mlx5_sysfs_bond_info(priv->if_index,
1280 				     &priv->bond_ifindex,
1281 				     priv->bond_name);
1282 		if (err)
1283 			DRV_LOG(ERR, "unable to get bond info: %s",
1284 				strerror(rte_errno));
1285 		else
1286 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1287 				priv->if_index, priv->bond_ifindex,
1288 				priv->bond_name);
1289 	}
1290 	eth_dev->data->dev_private = priv;
1291 	priv->dev_data = eth_dev->data;
1292 	eth_dev->data->mac_addrs = priv->mac;
1293 	eth_dev->device = dpdk_dev;
1294 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1295 	/* Configure the first MAC address by default. */
1296 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1297 		DRV_LOG(ERR,
1298 			"port %u cannot get MAC address, is mlx5_en"
1299 			" loaded? (errno: %s)",
1300 			eth_dev->data->port_id, strerror(rte_errno));
1301 		err = ENODEV;
1302 		goto error;
1303 	}
1304 	DRV_LOG(INFO,
1305 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1306 		eth_dev->data->port_id,
1307 		mac.addr_bytes[0], mac.addr_bytes[1],
1308 		mac.addr_bytes[2], mac.addr_bytes[3],
1309 		mac.addr_bytes[4], mac.addr_bytes[5]);
1310 #ifdef RTE_LIBRTE_MLX5_DEBUG
1311 	{
1312 		char ifname[IF_NAMESIZE];
1313 
1314 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1315 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1316 				eth_dev->data->port_id, ifname);
1317 		else
1318 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1319 				eth_dev->data->port_id);
1320 	}
1321 #endif
1322 	/* Get actual MTU if possible. */
1323 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1324 	if (err) {
1325 		err = rte_errno;
1326 		goto error;
1327 	}
1328 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1329 		priv->mtu);
1330 	/* Initialize burst functions to prevent crashes before link-up. */
1331 	eth_dev->rx_pkt_burst = removed_rx_burst;
1332 	eth_dev->tx_pkt_burst = removed_tx_burst;
1333 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1334 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1335 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1336 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1337 	/* Register MAC address. */
1338 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1339 	if (config->vf && config->vf_nl_en)
1340 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1341 				      mlx5_ifindex(eth_dev),
1342 				      eth_dev->data->mac_addrs,
1343 				      MLX5_MAX_MAC_ADDRESSES);
1344 	priv->flows = 0;
1345 	priv->ctrl_flows = 0;
1346 	TAILQ_INIT(&priv->flow_meters);
1347 	TAILQ_INIT(&priv->flow_meter_profiles);
1348 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1349 	mlx5_glue->dv_set_context_attr(sh->ctx,
1350 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1351 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1352 				.alloc = &mlx5_alloc_verbs_buf,
1353 				.free = &mlx5_free_verbs_buf,
1354 				.data = priv,
1355 			}));
1356 	/* Bring Ethernet device up. */
1357 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1358 		eth_dev->data->port_id);
1359 	mlx5_set_link_up(eth_dev);
1360 	/*
1361 	 * Even though the interrupt handler is not installed yet,
1362 	 * interrupts will still trigger on the async_fd from
1363 	 * Verbs context returned by ibv_open_device().
1364 	 */
1365 	mlx5_link_update(eth_dev, 0);
1366 #ifdef HAVE_MLX5DV_DR_ESWITCH
1367 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1368 	      (switch_info->representor || switch_info->master)))
1369 		config->dv_esw_en = 0;
1370 #else
1371 	config->dv_esw_en = 0;
1372 #endif
1373 	/* Detect minimal data bytes to inline. */
1374 	mlx5_set_min_inline(spawn, config);
1375 	/* Store device configuration on private structure. */
1376 	priv->config = *config;
1377 	/* Create context for virtual machine VLAN workaround. */
1378 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1379 	if (config->dv_flow_en) {
1380 		err = mlx5_alloc_shared_dr(priv);
1381 		if (err)
1382 			goto error;
1383 		/*
1384 		 * RSS id is shared with meter flow id. Meter flow id can only
1385 		 * use the 24 MSB of the register.
1386 		 */
1387 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1388 				     MLX5_MTR_COLOR_BITS);
1389 		if (!priv->qrss_id_pool) {
1390 			DRV_LOG(ERR, "can't create flow id pool");
1391 			err = ENOMEM;
1392 			goto error;
1393 		}
1394 	}
1395 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1396 		priv->obj_ops = devx_obj_ops;
1397 		priv->obj_ops.drop_action_create =
1398 						ibv_obj_ops.drop_action_create;
1399 		priv->obj_ops.drop_action_destroy =
1400 						ibv_obj_ops.drop_action_destroy;
1401 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1402 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1403 #else
1404 		if (config->dv_esw_en)
1405 			priv->obj_ops.txq_obj_modify =
1406 						ibv_obj_ops.txq_obj_modify;
1407 #endif
1408 		/* Use specific wrappers for Tx object. */
1409 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1410 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1411 
1412 	} else {
1413 		priv->obj_ops = ibv_obj_ops;
1414 	}
1415 	/* Supported Verbs flow priority number detection. */
1416 	err = mlx5_flow_discover_priorities(eth_dev);
1417 	if (err < 0) {
1418 		err = -err;
1419 		goto error;
1420 	}
1421 	priv->config.flow_prio = err;
1422 	if (!priv->config.dv_esw_en &&
1423 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1424 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1425 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1426 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1427 	}
1428 	mlx5_set_metadata_mask(eth_dev);
1429 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1430 	    !priv->sh->dv_regc0_mask) {
1431 		DRV_LOG(ERR, "metadata mode %u is not supported "
1432 			     "(no metadata reg_c[0] is available)",
1433 			     priv->config.dv_xmeta_en);
1434 			err = ENOTSUP;
1435 			goto error;
1436 	}
1437 	/*
1438 	 * Allocate the buffer for flow creating, just once.
1439 	 * The allocation must be done before any flow creating.
1440 	 */
1441 	mlx5_flow_alloc_intermediate(eth_dev);
1442 	/* Query availability of metadata reg_c's. */
1443 	err = mlx5_flow_discover_mreg_c(eth_dev);
1444 	if (err < 0) {
1445 		err = -err;
1446 		goto error;
1447 	}
1448 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1449 		DRV_LOG(DEBUG,
1450 			"port %u extensive metadata register is not supported",
1451 			eth_dev->data->port_id);
1452 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1453 			DRV_LOG(ERR, "metadata mode %u is not supported "
1454 				     "(no metadata registers available)",
1455 				     priv->config.dv_xmeta_en);
1456 			err = ENOTSUP;
1457 			goto error;
1458 		}
1459 	}
1460 	if (priv->config.dv_flow_en &&
1461 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1462 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1463 	    priv->sh->dv_regc0_mask) {
1464 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1465 						      MLX5_FLOW_MREG_HTABLE_SZ);
1466 		if (!priv->mreg_cp_tbl) {
1467 			err = ENOMEM;
1468 			goto error;
1469 		}
1470 	}
1471 	mlx5_flow_counter_mode_config(eth_dev);
1472 	return eth_dev;
1473 error:
1474 	if (priv) {
1475 		if (priv->mreg_cp_tbl)
1476 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1477 		if (priv->sh)
1478 			mlx5_os_free_shared_dr(priv);
1479 		if (priv->nl_socket_route >= 0)
1480 			close(priv->nl_socket_route);
1481 		if (priv->nl_socket_rdma >= 0)
1482 			close(priv->nl_socket_rdma);
1483 		if (priv->vmwa_context)
1484 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1485 		if (priv->qrss_id_pool)
1486 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1487 		if (own_domain_id)
1488 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1489 		mlx5_free(priv);
1490 		if (eth_dev != NULL)
1491 			eth_dev->data->dev_private = NULL;
1492 	}
1493 	if (eth_dev != NULL) {
1494 		/* mac_addrs must not be freed alone because part of
1495 		 * dev_private
1496 		 **/
1497 		eth_dev->data->mac_addrs = NULL;
1498 		rte_eth_dev_release_port(eth_dev);
1499 	}
1500 	if (sh)
1501 		mlx5_free_shared_dev_ctx(sh);
1502 	MLX5_ASSERT(err > 0);
1503 	rte_errno = err;
1504 	return NULL;
1505 }
1506 
1507 /**
1508  * Comparison callback to sort device data.
1509  *
1510  * This is meant to be used with qsort().
1511  *
1512  * @param a[in]
1513  *   Pointer to pointer to first data object.
1514  * @param b[in]
1515  *   Pointer to pointer to second data object.
1516  *
1517  * @return
1518  *   0 if both objects are equal, less than 0 if the first argument is less
1519  *   than the second, greater than 0 otherwise.
1520  */
1521 static int
1522 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1523 {
1524 	const struct mlx5_switch_info *si_a =
1525 		&((const struct mlx5_dev_spawn_data *)a)->info;
1526 	const struct mlx5_switch_info *si_b =
1527 		&((const struct mlx5_dev_spawn_data *)b)->info;
1528 	int ret;
1529 
1530 	/* Master device first. */
1531 	ret = si_b->master - si_a->master;
1532 	if (ret)
1533 		return ret;
1534 	/* Then representor devices. */
1535 	ret = si_b->representor - si_a->representor;
1536 	if (ret)
1537 		return ret;
1538 	/* Unidentified devices come last in no specific order. */
1539 	if (!si_a->representor)
1540 		return 0;
1541 	/* Order representors by name. */
1542 	return si_a->port_name - si_b->port_name;
1543 }
1544 
1545 /**
1546  * Match PCI information for possible slaves of bonding device.
1547  *
1548  * @param[in] ibv_dev
1549  *   Pointer to Infiniband device structure.
1550  * @param[in] pci_dev
1551  *   Pointer to PCI device structure to match PCI address.
1552  * @param[in] nl_rdma
1553  *   Netlink RDMA group socket handle.
1554  *
1555  * @return
1556  *   negative value if no bonding device found, otherwise
1557  *   positive index of slave PF in bonding.
1558  */
1559 static int
1560 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1561 			   const struct rte_pci_device *pci_dev,
1562 			   int nl_rdma)
1563 {
1564 	char ifname[IF_NAMESIZE + 1];
1565 	unsigned int ifindex;
1566 	unsigned int np, i;
1567 	FILE *file = NULL;
1568 	int pf = -1;
1569 
1570 	/*
1571 	 * Try to get master device name. If something goes
1572 	 * wrong suppose the lack of kernel support and no
1573 	 * bonding devices.
1574 	 */
1575 	if (nl_rdma < 0)
1576 		return -1;
1577 	if (!strstr(ibv_dev->name, "bond"))
1578 		return -1;
1579 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1580 	if (!np)
1581 		return -1;
1582 	/*
1583 	 * The Master device might not be on the predefined
1584 	 * port (not on port index 1, it is not garanted),
1585 	 * we have to scan all Infiniband device port and
1586 	 * find master.
1587 	 */
1588 	for (i = 1; i <= np; ++i) {
1589 		/* Check whether Infiniband port is populated. */
1590 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1591 		if (!ifindex)
1592 			continue;
1593 		if (!if_indextoname(ifindex, ifname))
1594 			continue;
1595 		/* Try to read bonding slave names from sysfs. */
1596 		MKSTR(slaves,
1597 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1598 		file = fopen(slaves, "r");
1599 		if (file)
1600 			break;
1601 	}
1602 	if (!file)
1603 		return -1;
1604 	/* Use safe format to check maximal buffer length. */
1605 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1606 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1607 		char tmp_str[IF_NAMESIZE + 32];
1608 		struct rte_pci_addr pci_addr;
1609 		struct mlx5_switch_info	info;
1610 
1611 		/* Process slave interface names in the loop. */
1612 		snprintf(tmp_str, sizeof(tmp_str),
1613 			 "/sys/class/net/%s", ifname);
1614 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1615 			DRV_LOG(WARNING, "can not get PCI address"
1616 					 " for netdev \"%s\"", ifname);
1617 			continue;
1618 		}
1619 		if (pci_dev->addr.domain != pci_addr.domain ||
1620 		    pci_dev->addr.bus != pci_addr.bus ||
1621 		    pci_dev->addr.devid != pci_addr.devid ||
1622 		    pci_dev->addr.function != pci_addr.function)
1623 			continue;
1624 		/* Slave interface PCI address match found. */
1625 		fclose(file);
1626 		snprintf(tmp_str, sizeof(tmp_str),
1627 			 "/sys/class/net/%s/phys_port_name", ifname);
1628 		file = fopen(tmp_str, "rb");
1629 		if (!file)
1630 			break;
1631 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1632 		if (fscanf(file, "%32s", tmp_str) == 1)
1633 			mlx5_translate_port_name(tmp_str, &info);
1634 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1635 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1636 			pf = info.port_name;
1637 		break;
1638 	}
1639 	if (file)
1640 		fclose(file);
1641 	return pf;
1642 }
1643 
1644 /**
1645  * DPDK callback to register a PCI device.
1646  *
1647  * This function spawns Ethernet devices out of a given PCI device.
1648  *
1649  * @param[in] pci_drv
1650  *   PCI driver structure (mlx5_driver).
1651  * @param[in] pci_dev
1652  *   PCI device information.
1653  *
1654  * @return
1655  *   0 on success, a negative errno value otherwise and rte_errno is set.
1656  */
1657 int
1658 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1659 		  struct rte_pci_device *pci_dev)
1660 {
1661 	struct ibv_device **ibv_list;
1662 	/*
1663 	 * Number of found IB Devices matching with requested PCI BDF.
1664 	 * nd != 1 means there are multiple IB devices over the same
1665 	 * PCI device and we have representors and master.
1666 	 */
1667 	unsigned int nd = 0;
1668 	/*
1669 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1670 	 * we have the single multiport IB device, and there may be
1671 	 * representors attached to some of found ports.
1672 	 */
1673 	unsigned int np = 0;
1674 	/*
1675 	 * Number of DPDK ethernet devices to Spawn - either over
1676 	 * multiple IB devices or multiple ports of single IB device.
1677 	 * Actually this is the number of iterations to spawn.
1678 	 */
1679 	unsigned int ns = 0;
1680 	/*
1681 	 * Bonding device
1682 	 *   < 0 - no bonding device (single one)
1683 	 *  >= 0 - bonding device (value is slave PF index)
1684 	 */
1685 	int bd = -1;
1686 	struct mlx5_dev_spawn_data *list = NULL;
1687 	struct mlx5_dev_config dev_config;
1688 	unsigned int dev_config_vf;
1689 	int ret;
1690 
1691 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1692 		mlx5_pmd_socket_init();
1693 	ret = mlx5_init_once();
1694 	if (ret) {
1695 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1696 			strerror(rte_errno));
1697 		return -rte_errno;
1698 	}
1699 	errno = 0;
1700 	ibv_list = mlx5_glue->get_device_list(&ret);
1701 	if (!ibv_list) {
1702 		rte_errno = errno ? errno : ENOSYS;
1703 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1704 		return -rte_errno;
1705 	}
1706 	/*
1707 	 * First scan the list of all Infiniband devices to find
1708 	 * matching ones, gathering into the list.
1709 	 */
1710 	struct ibv_device *ibv_match[ret + 1];
1711 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1712 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1713 	unsigned int i;
1714 
1715 	while (ret-- > 0) {
1716 		struct rte_pci_addr pci_addr;
1717 
1718 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1719 		bd = mlx5_device_bond_pci_match
1720 				(ibv_list[ret], pci_dev, nl_rdma);
1721 		if (bd >= 0) {
1722 			/*
1723 			 * Bonding device detected. Only one match is allowed,
1724 			 * the bonding is supported over multi-port IB device,
1725 			 * there should be no matches on representor PCI
1726 			 * functions or non VF LAG bonding devices with
1727 			 * specified address.
1728 			 */
1729 			if (nd) {
1730 				DRV_LOG(ERR,
1731 					"multiple PCI match on bonding device"
1732 					"\"%s\" found", ibv_list[ret]->name);
1733 				rte_errno = ENOENT;
1734 				ret = -rte_errno;
1735 				goto exit;
1736 			}
1737 			DRV_LOG(INFO, "PCI information matches for"
1738 				      " slave %d bonding device \"%s\"",
1739 				      bd, ibv_list[ret]->name);
1740 			ibv_match[nd++] = ibv_list[ret];
1741 			break;
1742 		}
1743 		if (mlx5_dev_to_pci_addr
1744 			(ibv_list[ret]->ibdev_path, &pci_addr))
1745 			continue;
1746 		if (pci_dev->addr.domain != pci_addr.domain ||
1747 		    pci_dev->addr.bus != pci_addr.bus ||
1748 		    pci_dev->addr.devid != pci_addr.devid ||
1749 		    pci_dev->addr.function != pci_addr.function)
1750 			continue;
1751 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1752 			ibv_list[ret]->name);
1753 		ibv_match[nd++] = ibv_list[ret];
1754 	}
1755 	ibv_match[nd] = NULL;
1756 	if (!nd) {
1757 		/* No device matches, just complain and bail out. */
1758 		DRV_LOG(WARNING,
1759 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1760 			" are kernel drivers loaded?",
1761 			pci_dev->addr.domain, pci_dev->addr.bus,
1762 			pci_dev->addr.devid, pci_dev->addr.function);
1763 		rte_errno = ENOENT;
1764 		ret = -rte_errno;
1765 		goto exit;
1766 	}
1767 	if (nd == 1) {
1768 		/*
1769 		 * Found single matching device may have multiple ports.
1770 		 * Each port may be representor, we have to check the port
1771 		 * number and check the representors existence.
1772 		 */
1773 		if (nl_rdma >= 0)
1774 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1775 		if (!np)
1776 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1777 					 " ports number", ibv_match[0]->name);
1778 		if (bd >= 0 && !np) {
1779 			DRV_LOG(ERR, "can not get ports"
1780 				     " for bonding device");
1781 			rte_errno = ENOENT;
1782 			ret = -rte_errno;
1783 			goto exit;
1784 		}
1785 	}
1786 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1787 	if (bd >= 0) {
1788 		/*
1789 		 * This may happen if there is VF LAG kernel support and
1790 		 * application is compiled with older rdma_core library.
1791 		 */
1792 		DRV_LOG(ERR,
1793 			"No kernel/verbs support for VF LAG bonding found.");
1794 		rte_errno = ENOTSUP;
1795 		ret = -rte_errno;
1796 		goto exit;
1797 	}
1798 #endif
1799 	/*
1800 	 * Now we can determine the maximal
1801 	 * amount of devices to be spawned.
1802 	 */
1803 	list = mlx5_malloc(MLX5_MEM_ZERO,
1804 			   sizeof(struct mlx5_dev_spawn_data) *
1805 			   (np ? np : nd),
1806 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1807 	if (!list) {
1808 		DRV_LOG(ERR, "spawn data array allocation failure");
1809 		rte_errno = ENOMEM;
1810 		ret = -rte_errno;
1811 		goto exit;
1812 	}
1813 	if (bd >= 0 || np > 1) {
1814 		/*
1815 		 * Single IB device with multiple ports found,
1816 		 * it may be E-Switch master device and representors.
1817 		 * We have to perform identification through the ports.
1818 		 */
1819 		MLX5_ASSERT(nl_rdma >= 0);
1820 		MLX5_ASSERT(ns == 0);
1821 		MLX5_ASSERT(nd == 1);
1822 		MLX5_ASSERT(np);
1823 		for (i = 1; i <= np; ++i) {
1824 			list[ns].max_port = np;
1825 			list[ns].phys_port = i;
1826 			list[ns].phys_dev = ibv_match[0];
1827 			list[ns].eth_dev = NULL;
1828 			list[ns].pci_dev = pci_dev;
1829 			list[ns].pf_bond = bd;
1830 			list[ns].ifindex = mlx5_nl_ifindex
1831 				(nl_rdma,
1832 				mlx5_os_get_dev_device_name
1833 						(list[ns].phys_dev), i);
1834 			if (!list[ns].ifindex) {
1835 				/*
1836 				 * No network interface index found for the
1837 				 * specified port, it means there is no
1838 				 * representor on this port. It's OK,
1839 				 * there can be disabled ports, for example
1840 				 * if sriov_numvfs < sriov_totalvfs.
1841 				 */
1842 				continue;
1843 			}
1844 			ret = -1;
1845 			if (nl_route >= 0)
1846 				ret = mlx5_nl_switch_info
1847 					       (nl_route,
1848 						list[ns].ifindex,
1849 						&list[ns].info);
1850 			if (ret || (!list[ns].info.representor &&
1851 				    !list[ns].info.master)) {
1852 				/*
1853 				 * We failed to recognize representors with
1854 				 * Netlink, let's try to perform the task
1855 				 * with sysfs.
1856 				 */
1857 				ret =  mlx5_sysfs_switch_info
1858 						(list[ns].ifindex,
1859 						 &list[ns].info);
1860 			}
1861 			if (!ret && bd >= 0) {
1862 				switch (list[ns].info.name_type) {
1863 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1864 					if (list[ns].info.port_name == bd)
1865 						ns++;
1866 					break;
1867 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1868 					/* Fallthrough */
1869 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1870 					if (list[ns].info.pf_num == bd)
1871 						ns++;
1872 					break;
1873 				default:
1874 					break;
1875 				}
1876 				continue;
1877 			}
1878 			if (!ret && (list[ns].info.representor ^
1879 				     list[ns].info.master))
1880 				ns++;
1881 		}
1882 		if (!ns) {
1883 			DRV_LOG(ERR,
1884 				"unable to recognize master/representors"
1885 				" on the IB device with multiple ports");
1886 			rte_errno = ENOENT;
1887 			ret = -rte_errno;
1888 			goto exit;
1889 		}
1890 	} else {
1891 		/*
1892 		 * The existence of several matching entries (nd > 1) means
1893 		 * port representors have been instantiated. No existing Verbs
1894 		 * call nor sysfs entries can tell them apart, this can only
1895 		 * be done through Netlink calls assuming kernel drivers are
1896 		 * recent enough to support them.
1897 		 *
1898 		 * In the event of identification failure through Netlink,
1899 		 * try again through sysfs, then:
1900 		 *
1901 		 * 1. A single IB device matches (nd == 1) with single
1902 		 *    port (np=0/1) and is not a representor, assume
1903 		 *    no switch support.
1904 		 *
1905 		 * 2. Otherwise no safe assumptions can be made;
1906 		 *    complain louder and bail out.
1907 		 */
1908 		for (i = 0; i != nd; ++i) {
1909 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1910 			list[ns].max_port = 1;
1911 			list[ns].phys_port = 1;
1912 			list[ns].phys_dev = ibv_match[i];
1913 			list[ns].eth_dev = NULL;
1914 			list[ns].pci_dev = pci_dev;
1915 			list[ns].pf_bond = -1;
1916 			list[ns].ifindex = 0;
1917 			if (nl_rdma >= 0)
1918 				list[ns].ifindex = mlx5_nl_ifindex
1919 				(nl_rdma,
1920 				mlx5_os_get_dev_device_name
1921 						(list[ns].phys_dev), 1);
1922 			if (!list[ns].ifindex) {
1923 				char ifname[IF_NAMESIZE];
1924 
1925 				/*
1926 				 * Netlink failed, it may happen with old
1927 				 * ib_core kernel driver (before 4.16).
1928 				 * We can assume there is old driver because
1929 				 * here we are processing single ports IB
1930 				 * devices. Let's try sysfs to retrieve
1931 				 * the ifindex. The method works for
1932 				 * master device only.
1933 				 */
1934 				if (nd > 1) {
1935 					/*
1936 					 * Multiple devices found, assume
1937 					 * representors, can not distinguish
1938 					 * master/representor and retrieve
1939 					 * ifindex via sysfs.
1940 					 */
1941 					continue;
1942 				}
1943 				ret = mlx5_get_ifname_sysfs
1944 					(ibv_match[i]->ibdev_path, ifname);
1945 				if (!ret)
1946 					list[ns].ifindex =
1947 						if_nametoindex(ifname);
1948 				if (!list[ns].ifindex) {
1949 					/*
1950 					 * No network interface index found
1951 					 * for the specified device, it means
1952 					 * there it is neither representor
1953 					 * nor master.
1954 					 */
1955 					continue;
1956 				}
1957 			}
1958 			ret = -1;
1959 			if (nl_route >= 0)
1960 				ret = mlx5_nl_switch_info
1961 					       (nl_route,
1962 						list[ns].ifindex,
1963 						&list[ns].info);
1964 			if (ret || (!list[ns].info.representor &&
1965 				    !list[ns].info.master)) {
1966 				/*
1967 				 * We failed to recognize representors with
1968 				 * Netlink, let's try to perform the task
1969 				 * with sysfs.
1970 				 */
1971 				ret =  mlx5_sysfs_switch_info
1972 						(list[ns].ifindex,
1973 						 &list[ns].info);
1974 			}
1975 			if (!ret && (list[ns].info.representor ^
1976 				     list[ns].info.master)) {
1977 				ns++;
1978 			} else if ((nd == 1) &&
1979 				   !list[ns].info.representor &&
1980 				   !list[ns].info.master) {
1981 				/*
1982 				 * Single IB device with
1983 				 * one physical port and
1984 				 * attached network device.
1985 				 * May be SRIOV is not enabled
1986 				 * or there is no representors.
1987 				 */
1988 				DRV_LOG(INFO, "no E-Switch support detected");
1989 				ns++;
1990 				break;
1991 			}
1992 		}
1993 		if (!ns) {
1994 			DRV_LOG(ERR,
1995 				"unable to recognize master/representors"
1996 				" on the multiple IB devices");
1997 			rte_errno = ENOENT;
1998 			ret = -rte_errno;
1999 			goto exit;
2000 		}
2001 	}
2002 	MLX5_ASSERT(ns);
2003 	/*
2004 	 * Sort list to probe devices in natural order for users convenience
2005 	 * (i.e. master first, then representors from lowest to highest ID).
2006 	 */
2007 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2008 	/* Device specific configuration. */
2009 	switch (pci_dev->id.device_id) {
2010 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2011 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2012 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2013 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2014 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2015 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2016 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2017 		dev_config_vf = 1;
2018 		break;
2019 	default:
2020 		dev_config_vf = 0;
2021 		break;
2022 	}
2023 	for (i = 0; i != ns; ++i) {
2024 		uint32_t restore;
2025 
2026 		/* Default configuration. */
2027 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2028 		dev_config.vf = dev_config_vf;
2029 		dev_config.mps = MLX5_ARG_UNSET;
2030 		dev_config.dbnc = MLX5_ARG_UNSET;
2031 		dev_config.rx_vec_en = 1;
2032 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2033 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2034 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2035 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2036 		dev_config.vf_nl_en = 1;
2037 		dev_config.mr_ext_memseg_en = 1;
2038 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2039 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2040 		dev_config.dv_esw_en = 1;
2041 		dev_config.dv_flow_en = 1;
2042 		dev_config.decap_en = 1;
2043 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2044 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2045 						 &list[i],
2046 						 &dev_config);
2047 		if (!list[i].eth_dev) {
2048 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2049 				break;
2050 			/* Device is disabled or already spawned. Ignore it. */
2051 			continue;
2052 		}
2053 		restore = list[i].eth_dev->data->dev_flags;
2054 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2055 		/* Restore non-PCI flags cleared by the above call. */
2056 		list[i].eth_dev->data->dev_flags |= restore;
2057 		rte_eth_dev_probing_finish(list[i].eth_dev);
2058 	}
2059 	if (i != ns) {
2060 		DRV_LOG(ERR,
2061 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2062 			" encountering an error: %s",
2063 			pci_dev->addr.domain, pci_dev->addr.bus,
2064 			pci_dev->addr.devid, pci_dev->addr.function,
2065 			strerror(rte_errno));
2066 		ret = -rte_errno;
2067 		/* Roll back. */
2068 		while (i--) {
2069 			if (!list[i].eth_dev)
2070 				continue;
2071 			mlx5_dev_close(list[i].eth_dev);
2072 			/* mac_addrs must not be freed because in dev_private */
2073 			list[i].eth_dev->data->mac_addrs = NULL;
2074 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2075 		}
2076 		/* Restore original error. */
2077 		rte_errno = -ret;
2078 	} else {
2079 		ret = 0;
2080 	}
2081 exit:
2082 	/*
2083 	 * Do the routine cleanup:
2084 	 * - close opened Netlink sockets
2085 	 * - free allocated spawn data array
2086 	 * - free the Infiniband device list
2087 	 */
2088 	if (nl_rdma >= 0)
2089 		close(nl_rdma);
2090 	if (nl_route >= 0)
2091 		close(nl_route);
2092 	if (list)
2093 		mlx5_free(list);
2094 	MLX5_ASSERT(ibv_list);
2095 	mlx5_glue->free_device_list(ibv_list);
2096 	return ret;
2097 }
2098 
2099 static int
2100 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2101 {
2102 	char *env;
2103 	int value;
2104 
2105 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2106 	/* Get environment variable to store. */
2107 	env = getenv(MLX5_SHUT_UP_BF);
2108 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2109 	if (config->dbnc == MLX5_ARG_UNSET)
2110 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2111 	else
2112 		setenv(MLX5_SHUT_UP_BF,
2113 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2114 	return value;
2115 }
2116 
2117 static void
2118 mlx5_restore_doorbell_mapping_env(int value)
2119 {
2120 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2121 	/* Restore the original environment variable state. */
2122 	if (value == MLX5_ARG_UNSET)
2123 		unsetenv(MLX5_SHUT_UP_BF);
2124 	else
2125 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2126 }
2127 
2128 /**
2129  * Extract pdn of PD object using DV API.
2130  *
2131  * @param[in] pd
2132  *   Pointer to the verbs PD object.
2133  * @param[out] pdn
2134  *   Pointer to the PD object number variable.
2135  *
2136  * @return
2137  *   0 on success, error value otherwise.
2138  */
2139 int
2140 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2141 {
2142 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2143 	struct mlx5dv_obj obj;
2144 	struct mlx5dv_pd pd_info;
2145 	int ret = 0;
2146 
2147 	obj.pd.in = pd;
2148 	obj.pd.out = &pd_info;
2149 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2150 	if (ret) {
2151 		DRV_LOG(DEBUG, "Fail to get PD object info");
2152 		return ret;
2153 	}
2154 	*pdn = pd_info.pdn;
2155 	return 0;
2156 #else
2157 	(void)pd;
2158 	(void)pdn;
2159 	return -ENOTSUP;
2160 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2161 }
2162 
2163 /**
2164  * Function API to open IB device.
2165  *
2166  * This function calls the Linux glue APIs to open a device.
2167  *
2168  * @param[in] spawn
2169  *   Pointer to the IB device attributes (name, port, etc).
2170  * @param[out] config
2171  *   Pointer to device configuration structure.
2172  * @param[out] sh
2173  *   Pointer to shared context structure.
2174  *
2175  * @return
2176  *   0 on success, a positive error value otherwise.
2177  */
2178 int
2179 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2180 		     const struct mlx5_dev_config *config,
2181 		     struct mlx5_dev_ctx_shared *sh)
2182 {
2183 	int dbmap_env;
2184 	int err = 0;
2185 
2186 	sh->numa_node = spawn->pci_dev->device.numa_node;
2187 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2188 	/*
2189 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2190 	 * before the device creation. The rdma_core library
2191 	 * checks the variable at device creation and
2192 	 * stores the result internally.
2193 	 */
2194 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2195 	/* Try to open IB device with DV first, then usual Verbs. */
2196 	errno = 0;
2197 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2198 	if (sh->ctx) {
2199 		sh->devx = 1;
2200 		DRV_LOG(DEBUG, "DevX is supported");
2201 		/* The device is created, no need for environment. */
2202 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2203 	} else {
2204 		/* The environment variable is still configured. */
2205 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2206 		err = errno ? errno : ENODEV;
2207 		/*
2208 		 * The environment variable is not needed anymore,
2209 		 * all device creation attempts are completed.
2210 		 */
2211 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2212 		if (!sh->ctx)
2213 			return err;
2214 		DRV_LOG(DEBUG, "DevX is NOT supported");
2215 		err = 0;
2216 	}
2217 	return err;
2218 }
2219 
2220 /**
2221  * Install shared asynchronous device events handler.
2222  * This function is implemented to support event sharing
2223  * between multiple ports of single IB device.
2224  *
2225  * @param sh
2226  *   Pointer to mlx5_dev_ctx_shared object.
2227  */
2228 void
2229 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2230 {
2231 	int ret;
2232 	int flags;
2233 
2234 	sh->intr_handle.fd = -1;
2235 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2236 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2237 		    F_SETFL, flags | O_NONBLOCK);
2238 	if (ret) {
2239 		DRV_LOG(INFO, "failed to change file descriptor async event"
2240 			" queue");
2241 	} else {
2242 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2243 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2244 		if (rte_intr_callback_register(&sh->intr_handle,
2245 					mlx5_dev_interrupt_handler, sh)) {
2246 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2247 			sh->intr_handle.fd = -1;
2248 		}
2249 	}
2250 	if (sh->devx) {
2251 #ifdef HAVE_IBV_DEVX_ASYNC
2252 		sh->intr_handle_devx.fd = -1;
2253 		sh->devx_comp =
2254 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2255 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2256 		if (!devx_comp) {
2257 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2258 			return;
2259 		}
2260 		flags = fcntl(devx_comp->fd, F_GETFL);
2261 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2262 		if (ret) {
2263 			DRV_LOG(INFO, "failed to change file descriptor"
2264 				" devx comp");
2265 			return;
2266 		}
2267 		sh->intr_handle_devx.fd = devx_comp->fd;
2268 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2269 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2270 					mlx5_dev_interrupt_handler_devx, sh)) {
2271 			DRV_LOG(INFO, "Fail to install the devx shared"
2272 				" interrupt.");
2273 			sh->intr_handle_devx.fd = -1;
2274 		}
2275 #endif /* HAVE_IBV_DEVX_ASYNC */
2276 	}
2277 }
2278 
2279 /**
2280  * Uninstall shared asynchronous device events handler.
2281  * This function is implemented to support event sharing
2282  * between multiple ports of single IB device.
2283  *
2284  * @param dev
2285  *   Pointer to mlx5_dev_ctx_shared object.
2286  */
2287 void
2288 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2289 {
2290 	if (sh->intr_handle.fd >= 0)
2291 		mlx5_intr_callback_unregister(&sh->intr_handle,
2292 					      mlx5_dev_interrupt_handler, sh);
2293 #ifdef HAVE_IBV_DEVX_ASYNC
2294 	if (sh->intr_handle_devx.fd >= 0)
2295 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2296 				  mlx5_dev_interrupt_handler_devx, sh);
2297 	if (sh->devx_comp)
2298 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2299 #endif
2300 }
2301 
2302 /**
2303  * Read statistics by a named counter.
2304  *
2305  * @param[in] priv
2306  *   Pointer to the private device data structure.
2307  * @param[in] ctr_name
2308  *   Pointer to the name of the statistic counter to read
2309  * @param[out] stat
2310  *   Pointer to read statistic value.
2311  * @return
2312  *   0 on success and stat is valud, 1 if failed to read the value
2313  *   rte_errno is set.
2314  *
2315  */
2316 int
2317 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2318 		      uint64_t *stat)
2319 {
2320 	int fd;
2321 
2322 	if (priv->sh) {
2323 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2324 		      priv->sh->ibdev_path,
2325 		      priv->dev_port,
2326 		      ctr_name);
2327 		fd = open(path, O_RDONLY);
2328 		/*
2329 		 * in switchdev the file location is not per port
2330 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2331 		 */
2332 		if (fd == -1) {
2333 			MKSTR(path1, "%s/hw_counters/%s",
2334 			      priv->sh->ibdev_path,
2335 			      ctr_name);
2336 			fd = open(path1, O_RDONLY);
2337 		}
2338 		if (fd != -1) {
2339 			char buf[21] = {'\0'};
2340 			ssize_t n = read(fd, buf, sizeof(buf));
2341 
2342 			close(fd);
2343 			if (n != -1) {
2344 				*stat = strtoull(buf, NULL, 10);
2345 				return 0;
2346 			}
2347 		}
2348 	}
2349 	*stat = 0;
2350 	return 1;
2351 }
2352 
2353 /**
2354  * Set the reg_mr and dereg_mr call backs
2355  *
2356  * @param reg_mr_cb[out]
2357  *   Pointer to reg_mr func
2358  * @param dereg_mr_cb[out]
2359  *   Pointer to dereg_mr func
2360  *
2361  */
2362 void
2363 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2364 		      mlx5_dereg_mr_t *dereg_mr_cb)
2365 {
2366 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2367 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2368 }
2369 
2370 /**
2371  * Remove a MAC address from device
2372  *
2373  * @param dev
2374  *   Pointer to Ethernet device structure.
2375  * @param index
2376  *   MAC address index.
2377  */
2378 void
2379 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2380 {
2381 	struct mlx5_priv *priv = dev->data->dev_private;
2382 	const int vf = priv->config.vf;
2383 
2384 	if (vf)
2385 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2386 					mlx5_ifindex(dev), priv->mac_own,
2387 					&dev->data->mac_addrs[index], index);
2388 }
2389 
2390 /**
2391  * Adds a MAC address to the device
2392  *
2393  * @param dev
2394  *   Pointer to Ethernet device structure.
2395  * @param mac_addr
2396  *   MAC address to register.
2397  * @param index
2398  *   MAC address index.
2399  *
2400  * @return
2401  *   0 on success, a negative errno value otherwise
2402  */
2403 int
2404 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2405 		     uint32_t index)
2406 {
2407 	struct mlx5_priv *priv = dev->data->dev_private;
2408 	const int vf = priv->config.vf;
2409 	int ret = 0;
2410 
2411 	if (vf)
2412 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2413 					   mlx5_ifindex(dev), priv->mac_own,
2414 					   mac, index);
2415 	return ret;
2416 }
2417 
2418 /**
2419  * Modify a VF MAC address
2420  *
2421  * @param priv
2422  *   Pointer to device private data.
2423  * @param mac_addr
2424  *   MAC address to modify into.
2425  * @param iface_idx
2426  *   Net device interface index
2427  * @param vf_index
2428  *   VF index
2429  *
2430  * @return
2431  *   0 on success, a negative errno value otherwise
2432  */
2433 int
2434 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2435 			   unsigned int iface_idx,
2436 			   struct rte_ether_addr *mac_addr,
2437 			   int vf_index)
2438 {
2439 	return mlx5_nl_vf_mac_addr_modify
2440 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2441 }
2442 
2443 /**
2444  * Set device promiscuous mode
2445  *
2446  * @param dev
2447  *   Pointer to Ethernet device structure.
2448  * @param enable
2449  *   0 - promiscuous is disabled, otherwise - enabled
2450  *
2451  * @return
2452  *   0 on success, a negative error value otherwise
2453  */
2454 int
2455 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2456 {
2457 	struct mlx5_priv *priv = dev->data->dev_private;
2458 
2459 	return mlx5_nl_promisc(priv->nl_socket_route,
2460 			       mlx5_ifindex(dev), !!enable);
2461 }
2462 
2463 /**
2464  * Set device promiscuous mode
2465  *
2466  * @param dev
2467  *   Pointer to Ethernet device structure.
2468  * @param enable
2469  *   0 - all multicase is disabled, otherwise - enabled
2470  *
2471  * @return
2472  *   0 on success, a negative error value otherwise
2473  */
2474 int
2475 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2476 {
2477 	struct mlx5_priv *priv = dev->data->dev_private;
2478 
2479 	return mlx5_nl_allmulti(priv->nl_socket_route,
2480 				mlx5_ifindex(dev), !!enable);
2481 }
2482 
2483 /**
2484  * Flush device MAC addresses
2485  *
2486  * @param dev
2487  *   Pointer to Ethernet device structure.
2488  *
2489  */
2490 void
2491 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2492 {
2493 	struct mlx5_priv *priv = dev->data->dev_private;
2494 
2495 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2496 			       dev->data->mac_addrs,
2497 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2498 }
2499 
2500 const struct eth_dev_ops mlx5_os_dev_ops = {
2501 	.dev_configure = mlx5_dev_configure,
2502 	.dev_start = mlx5_dev_start,
2503 	.dev_stop = mlx5_dev_stop,
2504 	.dev_set_link_down = mlx5_set_link_down,
2505 	.dev_set_link_up = mlx5_set_link_up,
2506 	.dev_close = mlx5_dev_close,
2507 	.promiscuous_enable = mlx5_promiscuous_enable,
2508 	.promiscuous_disable = mlx5_promiscuous_disable,
2509 	.allmulticast_enable = mlx5_allmulticast_enable,
2510 	.allmulticast_disable = mlx5_allmulticast_disable,
2511 	.link_update = mlx5_link_update,
2512 	.stats_get = mlx5_stats_get,
2513 	.stats_reset = mlx5_stats_reset,
2514 	.xstats_get = mlx5_xstats_get,
2515 	.xstats_reset = mlx5_xstats_reset,
2516 	.xstats_get_names = mlx5_xstats_get_names,
2517 	.fw_version_get = mlx5_fw_version_get,
2518 	.dev_infos_get = mlx5_dev_infos_get,
2519 	.read_clock = mlx5_txpp_read_clock,
2520 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2521 	.vlan_filter_set = mlx5_vlan_filter_set,
2522 	.rx_queue_setup = mlx5_rx_queue_setup,
2523 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2524 	.tx_queue_setup = mlx5_tx_queue_setup,
2525 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2526 	.rx_queue_release = mlx5_rx_queue_release,
2527 	.tx_queue_release = mlx5_tx_queue_release,
2528 	.rx_queue_start = mlx5_rx_queue_start,
2529 	.rx_queue_stop = mlx5_rx_queue_stop,
2530 	.tx_queue_start = mlx5_tx_queue_start,
2531 	.tx_queue_stop = mlx5_tx_queue_stop,
2532 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2533 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2534 	.mac_addr_remove = mlx5_mac_addr_remove,
2535 	.mac_addr_add = mlx5_mac_addr_add,
2536 	.mac_addr_set = mlx5_mac_addr_set,
2537 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2538 	.mtu_set = mlx5_dev_set_mtu,
2539 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2540 	.vlan_offload_set = mlx5_vlan_offload_set,
2541 	.reta_update = mlx5_dev_rss_reta_update,
2542 	.reta_query = mlx5_dev_rss_reta_query,
2543 	.rss_hash_update = mlx5_rss_hash_update,
2544 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2545 	.filter_ctrl = mlx5_dev_filter_ctrl,
2546 	.rxq_info_get = mlx5_rxq_info_get,
2547 	.txq_info_get = mlx5_txq_info_get,
2548 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2549 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2550 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2551 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2552 	.is_removed = mlx5_is_removed,
2553 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2554 	.get_module_info = mlx5_get_module_info,
2555 	.get_module_eeprom = mlx5_get_module_eeprom,
2556 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2557 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2558 };
2559 
2560 /* Available operations from secondary process. */
2561 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2562 	.stats_get = mlx5_stats_get,
2563 	.stats_reset = mlx5_stats_reset,
2564 	.xstats_get = mlx5_xstats_get,
2565 	.xstats_reset = mlx5_xstats_reset,
2566 	.xstats_get_names = mlx5_xstats_get_names,
2567 	.fw_version_get = mlx5_fw_version_get,
2568 	.dev_infos_get = mlx5_dev_infos_get,
2569 	.read_clock = mlx5_txpp_read_clock,
2570 	.rx_queue_start = mlx5_rx_queue_start,
2571 	.rx_queue_stop = mlx5_rx_queue_stop,
2572 	.tx_queue_start = mlx5_tx_queue_start,
2573 	.tx_queue_stop = mlx5_tx_queue_stop,
2574 	.rxq_info_get = mlx5_rxq_info_get,
2575 	.txq_info_get = mlx5_txq_info_get,
2576 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2577 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2578 	.get_module_info = mlx5_get_module_info,
2579 	.get_module_eeprom = mlx5_get_module_eeprom,
2580 };
2581 
2582 /* Available operations in flow isolated mode. */
2583 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2584 	.dev_configure = mlx5_dev_configure,
2585 	.dev_start = mlx5_dev_start,
2586 	.dev_stop = mlx5_dev_stop,
2587 	.dev_set_link_down = mlx5_set_link_down,
2588 	.dev_set_link_up = mlx5_set_link_up,
2589 	.dev_close = mlx5_dev_close,
2590 	.promiscuous_enable = mlx5_promiscuous_enable,
2591 	.promiscuous_disable = mlx5_promiscuous_disable,
2592 	.allmulticast_enable = mlx5_allmulticast_enable,
2593 	.allmulticast_disable = mlx5_allmulticast_disable,
2594 	.link_update = mlx5_link_update,
2595 	.stats_get = mlx5_stats_get,
2596 	.stats_reset = mlx5_stats_reset,
2597 	.xstats_get = mlx5_xstats_get,
2598 	.xstats_reset = mlx5_xstats_reset,
2599 	.xstats_get_names = mlx5_xstats_get_names,
2600 	.fw_version_get = mlx5_fw_version_get,
2601 	.dev_infos_get = mlx5_dev_infos_get,
2602 	.read_clock = mlx5_txpp_read_clock,
2603 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2604 	.vlan_filter_set = mlx5_vlan_filter_set,
2605 	.rx_queue_setup = mlx5_rx_queue_setup,
2606 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2607 	.tx_queue_setup = mlx5_tx_queue_setup,
2608 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2609 	.rx_queue_release = mlx5_rx_queue_release,
2610 	.tx_queue_release = mlx5_tx_queue_release,
2611 	.rx_queue_start = mlx5_rx_queue_start,
2612 	.rx_queue_stop = mlx5_rx_queue_stop,
2613 	.tx_queue_start = mlx5_tx_queue_start,
2614 	.tx_queue_stop = mlx5_tx_queue_stop,
2615 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2616 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2617 	.mac_addr_remove = mlx5_mac_addr_remove,
2618 	.mac_addr_add = mlx5_mac_addr_add,
2619 	.mac_addr_set = mlx5_mac_addr_set,
2620 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2621 	.mtu_set = mlx5_dev_set_mtu,
2622 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2623 	.vlan_offload_set = mlx5_vlan_offload_set,
2624 	.filter_ctrl = mlx5_dev_filter_ctrl,
2625 	.rxq_info_get = mlx5_rxq_info_get,
2626 	.txq_info_get = mlx5_txq_info_get,
2627 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2628 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2629 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2630 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2631 	.is_removed = mlx5_is_removed,
2632 	.get_module_info = mlx5_get_module_info,
2633 	.get_module_eeprom = mlx5_get_module_eeprom,
2634 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2635 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2636 };
2637