xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 3aa279157fa0f372b8005ce47e19e2045c8e98e7)
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 relax ordering support. */
1071 		if (config->hca_attr.relaxed_ordering_write &&
1072 		    config->hca_attr.relaxed_ordering_read  &&
1073 		    !haswell_broadwell_cpu)
1074 			sh->cmng.relaxed_ordering = 1;
1075 		/* Check for LRO support. */
1076 		if (config->dest_tir && config->hca_attr.lro_cap &&
1077 		    config->dv_flow_en) {
1078 			/* TBD check tunnel lro caps. */
1079 			config->lro.supported = config->hca_attr.lro_cap;
1080 			DRV_LOG(DEBUG, "Device supports LRO");
1081 			/*
1082 			 * If LRO timeout is not configured by application,
1083 			 * use the minimal supported value.
1084 			 */
1085 			if (!config->lro.timeout)
1086 				config->lro.timeout =
1087 				config->hca_attr.lro_timer_supported_periods[0];
1088 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1089 				config->lro.timeout);
1090 			DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1091 				"required for coalescing is %d bytes",
1092 				config->hca_attr.lro_min_mss_size);
1093 		}
1094 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1095 		if (config->hca_attr.qos.sup &&
1096 		    config->hca_attr.qos.srtcm_sup &&
1097 		    config->dv_flow_en) {
1098 			uint8_t reg_c_mask =
1099 				config->hca_attr.qos.flow_meter_reg_c_ids;
1100 			/*
1101 			 * Meter needs two REG_C's for color match and pre-sfx
1102 			 * flow match. Here get the REG_C for color match.
1103 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1104 			 */
1105 			reg_c_mask &= 0xfc;
1106 			if (__builtin_popcount(reg_c_mask) < 1) {
1107 				priv->mtr_en = 0;
1108 				DRV_LOG(WARNING, "No available register for"
1109 					" meter.");
1110 			} else {
1111 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1112 						      REG_C_0;
1113 				priv->mtr_en = 1;
1114 				priv->mtr_reg_share =
1115 				      config->hca_attr.qos.flow_meter_reg_share;
1116 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1117 					priv->mtr_color_reg);
1118 			}
1119 		}
1120 #endif
1121 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1122 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1123 		    config->dv_flow_en) {
1124 			priv->sampler_en = 1;
1125 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1126 		} else {
1127 			priv->sampler_en = 0;
1128 			if (!config->hca_attr.log_max_ft_sampler_num)
1129 				DRV_LOG(WARNING, "No available register for"
1130 						" Sampler.");
1131 			else
1132 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1133 		}
1134 #endif
1135 	}
1136 	if (config->tx_pp) {
1137 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1138 			config->hca_attr.dev_freq_khz);
1139 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1140 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1141 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1142 			config->hca_attr.cross_channel ? "" : "not ");
1143 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1144 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1145 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1146 			config->hca_attr.non_wire_sq ? "" : "not ");
1147 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1148 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1149 			config->hca_attr.log_max_static_sq_wq);
1150 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1151 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1152 		if (!config->devx) {
1153 			DRV_LOG(ERR, "DevX is required for packet pacing");
1154 			err = ENODEV;
1155 			goto error;
1156 		}
1157 		if (!config->hca_attr.qos.packet_pacing) {
1158 			DRV_LOG(ERR, "Packet pacing is not supported");
1159 			err = ENODEV;
1160 			goto error;
1161 		}
1162 		if (!config->hca_attr.cross_channel) {
1163 			DRV_LOG(ERR, "Cross channel operations are"
1164 				     " required for packet pacing");
1165 			err = ENODEV;
1166 			goto error;
1167 		}
1168 		if (!config->hca_attr.wqe_index_ignore) {
1169 			DRV_LOG(ERR, "WQE index ignore feature is"
1170 				     " required for packet pacing");
1171 			err = ENODEV;
1172 			goto error;
1173 		}
1174 		if (!config->hca_attr.non_wire_sq) {
1175 			DRV_LOG(ERR, "Non-wire SQ feature is"
1176 				     " required for packet pacing");
1177 			err = ENODEV;
1178 			goto error;
1179 		}
1180 		if (!config->hca_attr.log_max_static_sq_wq) {
1181 			DRV_LOG(ERR, "Static WQE SQ feature is"
1182 				     " required for packet pacing");
1183 			err = ENODEV;
1184 			goto error;
1185 		}
1186 		if (!config->hca_attr.qos.wqe_rate_pp) {
1187 			DRV_LOG(ERR, "WQE rate mode is required"
1188 				     " for packet pacing");
1189 			err = ENODEV;
1190 			goto error;
1191 		}
1192 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1193 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1194 			     " can't create queues for packet pacing");
1195 		err = ENODEV;
1196 		goto error;
1197 #endif
1198 	}
1199 	if (config->devx) {
1200 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1201 
1202 		err = config->hca_attr.access_register_user ?
1203 			mlx5_devx_cmd_register_read
1204 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1205 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1206 		if (!err) {
1207 			uint32_t ts_mode;
1208 
1209 			/* MTUTC register is read successfully. */
1210 			ts_mode = MLX5_GET(register_mtutc, reg,
1211 					   time_stamp_mode);
1212 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1213 				config->rt_timestamp = 1;
1214 		} else {
1215 			/* Kernel does not support register reading. */
1216 			if (config->hca_attr.dev_freq_khz ==
1217 						 (NS_PER_S / MS_PER_S))
1218 				config->rt_timestamp = 1;
1219 		}
1220 	}
1221 	/*
1222 	 * If HW has bug working with tunnel packet decapsulation and
1223 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1224 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1225 	 */
1226 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1227 		config->hw_fcs_strip = 0;
1228 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1229 		(config->hw_fcs_strip ? "" : "not "));
1230 	if (config->mprq.enabled && mprq) {
1231 		if (config->mprq.stride_num_n &&
1232 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1233 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1234 			config->mprq.stride_num_n =
1235 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1236 						mprq_min_stride_num_n),
1237 					mprq_max_stride_num_n);
1238 			DRV_LOG(WARNING,
1239 				"the number of strides"
1240 				" for Multi-Packet RQ is out of range,"
1241 				" setting default value (%u)",
1242 				1 << config->mprq.stride_num_n);
1243 		}
1244 		if (config->mprq.stride_size_n &&
1245 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1246 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1247 			config->mprq.stride_size_n =
1248 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1249 						mprq_min_stride_size_n),
1250 					mprq_max_stride_size_n);
1251 			DRV_LOG(WARNING,
1252 				"the size of a stride"
1253 				" for Multi-Packet RQ is out of range,"
1254 				" setting default value (%u)",
1255 				1 << config->mprq.stride_size_n);
1256 		}
1257 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1258 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1259 	} else if (config->mprq.enabled && !mprq) {
1260 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1261 		config->mprq.enabled = 0;
1262 	}
1263 	if (config->max_dump_files_num == 0)
1264 		config->max_dump_files_num = 128;
1265 	eth_dev = rte_eth_dev_allocate(name);
1266 	if (eth_dev == NULL) {
1267 		DRV_LOG(ERR, "can not allocate rte ethdev");
1268 		err = ENOMEM;
1269 		goto error;
1270 	}
1271 	if (priv->representor) {
1272 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1273 		eth_dev->data->representor_id = priv->representor_id;
1274 	}
1275 	/*
1276 	 * Store associated network device interface index. This index
1277 	 * is permanent throughout the lifetime of device. So, we may store
1278 	 * the ifindex here and use the cached value further.
1279 	 */
1280 	MLX5_ASSERT(spawn->ifindex);
1281 	priv->if_index = spawn->ifindex;
1282 	if (priv->pf_bond >= 0 && priv->master) {
1283 		/* Get bond interface info */
1284 		err = mlx5_sysfs_bond_info(priv->if_index,
1285 				     &priv->bond_ifindex,
1286 				     priv->bond_name);
1287 		if (err)
1288 			DRV_LOG(ERR, "unable to get bond info: %s",
1289 				strerror(rte_errno));
1290 		else
1291 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1292 				priv->if_index, priv->bond_ifindex,
1293 				priv->bond_name);
1294 	}
1295 	eth_dev->data->dev_private = priv;
1296 	priv->dev_data = eth_dev->data;
1297 	eth_dev->data->mac_addrs = priv->mac;
1298 	eth_dev->device = dpdk_dev;
1299 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1300 	/* Configure the first MAC address by default. */
1301 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1302 		DRV_LOG(ERR,
1303 			"port %u cannot get MAC address, is mlx5_en"
1304 			" loaded? (errno: %s)",
1305 			eth_dev->data->port_id, strerror(rte_errno));
1306 		err = ENODEV;
1307 		goto error;
1308 	}
1309 	DRV_LOG(INFO,
1310 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1311 		eth_dev->data->port_id,
1312 		mac.addr_bytes[0], mac.addr_bytes[1],
1313 		mac.addr_bytes[2], mac.addr_bytes[3],
1314 		mac.addr_bytes[4], mac.addr_bytes[5]);
1315 #ifdef RTE_LIBRTE_MLX5_DEBUG
1316 	{
1317 		char ifname[IF_NAMESIZE];
1318 
1319 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1320 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1321 				eth_dev->data->port_id, ifname);
1322 		else
1323 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1324 				eth_dev->data->port_id);
1325 	}
1326 #endif
1327 	/* Get actual MTU if possible. */
1328 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1329 	if (err) {
1330 		err = rte_errno;
1331 		goto error;
1332 	}
1333 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1334 		priv->mtu);
1335 	/* Initialize burst functions to prevent crashes before link-up. */
1336 	eth_dev->rx_pkt_burst = removed_rx_burst;
1337 	eth_dev->tx_pkt_burst = removed_tx_burst;
1338 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1339 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1340 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1341 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1342 	/* Register MAC address. */
1343 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1344 	if (config->vf && config->vf_nl_en)
1345 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1346 				      mlx5_ifindex(eth_dev),
1347 				      eth_dev->data->mac_addrs,
1348 				      MLX5_MAX_MAC_ADDRESSES);
1349 	priv->flows = 0;
1350 	priv->ctrl_flows = 0;
1351 	TAILQ_INIT(&priv->flow_meters);
1352 	TAILQ_INIT(&priv->flow_meter_profiles);
1353 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1354 	mlx5_glue->dv_set_context_attr(sh->ctx,
1355 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1356 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1357 				.alloc = &mlx5_alloc_verbs_buf,
1358 				.free = &mlx5_free_verbs_buf,
1359 				.data = priv,
1360 			}));
1361 	/* Bring Ethernet device up. */
1362 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1363 		eth_dev->data->port_id);
1364 	mlx5_set_link_up(eth_dev);
1365 	/*
1366 	 * Even though the interrupt handler is not installed yet,
1367 	 * interrupts will still trigger on the async_fd from
1368 	 * Verbs context returned by ibv_open_device().
1369 	 */
1370 	mlx5_link_update(eth_dev, 0);
1371 #ifdef HAVE_MLX5DV_DR_ESWITCH
1372 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1373 	      (switch_info->representor || switch_info->master)))
1374 		config->dv_esw_en = 0;
1375 #else
1376 	config->dv_esw_en = 0;
1377 #endif
1378 	/* Detect minimal data bytes to inline. */
1379 	mlx5_set_min_inline(spawn, config);
1380 	/* Store device configuration on private structure. */
1381 	priv->config = *config;
1382 	/* Create context for virtual machine VLAN workaround. */
1383 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1384 	if (config->dv_flow_en) {
1385 		err = mlx5_alloc_shared_dr(priv);
1386 		if (err)
1387 			goto error;
1388 		/*
1389 		 * RSS id is shared with meter flow id. Meter flow id can only
1390 		 * use the 24 MSB of the register.
1391 		 */
1392 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1393 				     MLX5_MTR_COLOR_BITS);
1394 		if (!priv->qrss_id_pool) {
1395 			DRV_LOG(ERR, "can't create flow id pool");
1396 			err = ENOMEM;
1397 			goto error;
1398 		}
1399 	}
1400 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1401 		priv->obj_ops = devx_obj_ops;
1402 		priv->obj_ops.drop_action_create =
1403 						ibv_obj_ops.drop_action_create;
1404 		priv->obj_ops.drop_action_destroy =
1405 						ibv_obj_ops.drop_action_destroy;
1406 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1407 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1408 #else
1409 		if (config->dv_esw_en)
1410 			priv->obj_ops.txq_obj_modify =
1411 						ibv_obj_ops.txq_obj_modify;
1412 #endif
1413 		/* Use specific wrappers for Tx object. */
1414 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1415 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1416 
1417 	} else {
1418 		priv->obj_ops = ibv_obj_ops;
1419 	}
1420 	/* Supported Verbs flow priority number detection. */
1421 	err = mlx5_flow_discover_priorities(eth_dev);
1422 	if (err < 0) {
1423 		err = -err;
1424 		goto error;
1425 	}
1426 	priv->config.flow_prio = err;
1427 	if (!priv->config.dv_esw_en &&
1428 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1429 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1430 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1431 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1432 	}
1433 	mlx5_set_metadata_mask(eth_dev);
1434 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1435 	    !priv->sh->dv_regc0_mask) {
1436 		DRV_LOG(ERR, "metadata mode %u is not supported "
1437 			     "(no metadata reg_c[0] is available)",
1438 			     priv->config.dv_xmeta_en);
1439 			err = ENOTSUP;
1440 			goto error;
1441 	}
1442 	/*
1443 	 * Allocate the buffer for flow creating, just once.
1444 	 * The allocation must be done before any flow creating.
1445 	 */
1446 	mlx5_flow_alloc_intermediate(eth_dev);
1447 	/* Query availability of metadata reg_c's. */
1448 	err = mlx5_flow_discover_mreg_c(eth_dev);
1449 	if (err < 0) {
1450 		err = -err;
1451 		goto error;
1452 	}
1453 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1454 		DRV_LOG(DEBUG,
1455 			"port %u extensive metadata register is not supported",
1456 			eth_dev->data->port_id);
1457 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1458 			DRV_LOG(ERR, "metadata mode %u is not supported "
1459 				     "(no metadata registers available)",
1460 				     priv->config.dv_xmeta_en);
1461 			err = ENOTSUP;
1462 			goto error;
1463 		}
1464 	}
1465 	if (priv->config.dv_flow_en &&
1466 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1467 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1468 	    priv->sh->dv_regc0_mask) {
1469 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1470 						      MLX5_FLOW_MREG_HTABLE_SZ);
1471 		if (!priv->mreg_cp_tbl) {
1472 			err = ENOMEM;
1473 			goto error;
1474 		}
1475 	}
1476 	mlx5_flow_counter_mode_config(eth_dev);
1477 	return eth_dev;
1478 error:
1479 	if (priv) {
1480 		if (priv->mreg_cp_tbl)
1481 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1482 		if (priv->sh)
1483 			mlx5_os_free_shared_dr(priv);
1484 		if (priv->nl_socket_route >= 0)
1485 			close(priv->nl_socket_route);
1486 		if (priv->nl_socket_rdma >= 0)
1487 			close(priv->nl_socket_rdma);
1488 		if (priv->vmwa_context)
1489 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1490 		if (priv->qrss_id_pool)
1491 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1492 		if (own_domain_id)
1493 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1494 		mlx5_free(priv);
1495 		if (eth_dev != NULL)
1496 			eth_dev->data->dev_private = NULL;
1497 	}
1498 	if (eth_dev != NULL) {
1499 		/* mac_addrs must not be freed alone because part of
1500 		 * dev_private
1501 		 **/
1502 		eth_dev->data->mac_addrs = NULL;
1503 		rte_eth_dev_release_port(eth_dev);
1504 	}
1505 	if (sh)
1506 		mlx5_free_shared_dev_ctx(sh);
1507 	MLX5_ASSERT(err > 0);
1508 	rte_errno = err;
1509 	return NULL;
1510 }
1511 
1512 /**
1513  * Comparison callback to sort device data.
1514  *
1515  * This is meant to be used with qsort().
1516  *
1517  * @param a[in]
1518  *   Pointer to pointer to first data object.
1519  * @param b[in]
1520  *   Pointer to pointer to second data object.
1521  *
1522  * @return
1523  *   0 if both objects are equal, less than 0 if the first argument is less
1524  *   than the second, greater than 0 otherwise.
1525  */
1526 static int
1527 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1528 {
1529 	const struct mlx5_switch_info *si_a =
1530 		&((const struct mlx5_dev_spawn_data *)a)->info;
1531 	const struct mlx5_switch_info *si_b =
1532 		&((const struct mlx5_dev_spawn_data *)b)->info;
1533 	int ret;
1534 
1535 	/* Master device first. */
1536 	ret = si_b->master - si_a->master;
1537 	if (ret)
1538 		return ret;
1539 	/* Then representor devices. */
1540 	ret = si_b->representor - si_a->representor;
1541 	if (ret)
1542 		return ret;
1543 	/* Unidentified devices come last in no specific order. */
1544 	if (!si_a->representor)
1545 		return 0;
1546 	/* Order representors by name. */
1547 	return si_a->port_name - si_b->port_name;
1548 }
1549 
1550 /**
1551  * Match PCI information for possible slaves of bonding device.
1552  *
1553  * @param[in] ibv_dev
1554  *   Pointer to Infiniband device structure.
1555  * @param[in] pci_dev
1556  *   Pointer to PCI device structure to match PCI address.
1557  * @param[in] nl_rdma
1558  *   Netlink RDMA group socket handle.
1559  *
1560  * @return
1561  *   negative value if no bonding device found, otherwise
1562  *   positive index of slave PF in bonding.
1563  */
1564 static int
1565 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1566 			   const struct rte_pci_device *pci_dev,
1567 			   int nl_rdma)
1568 {
1569 	char ifname[IF_NAMESIZE + 1];
1570 	unsigned int ifindex;
1571 	unsigned int np, i;
1572 	FILE *file = NULL;
1573 	int pf = -1;
1574 
1575 	/*
1576 	 * Try to get master device name. If something goes
1577 	 * wrong suppose the lack of kernel support and no
1578 	 * bonding devices.
1579 	 */
1580 	if (nl_rdma < 0)
1581 		return -1;
1582 	if (!strstr(ibv_dev->name, "bond"))
1583 		return -1;
1584 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1585 	if (!np)
1586 		return -1;
1587 	/*
1588 	 * The Master device might not be on the predefined
1589 	 * port (not on port index 1, it is not garanted),
1590 	 * we have to scan all Infiniband device port and
1591 	 * find master.
1592 	 */
1593 	for (i = 1; i <= np; ++i) {
1594 		/* Check whether Infiniband port is populated. */
1595 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1596 		if (!ifindex)
1597 			continue;
1598 		if (!if_indextoname(ifindex, ifname))
1599 			continue;
1600 		/* Try to read bonding slave names from sysfs. */
1601 		MKSTR(slaves,
1602 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1603 		file = fopen(slaves, "r");
1604 		if (file)
1605 			break;
1606 	}
1607 	if (!file)
1608 		return -1;
1609 	/* Use safe format to check maximal buffer length. */
1610 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1611 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1612 		char tmp_str[IF_NAMESIZE + 32];
1613 		struct rte_pci_addr pci_addr;
1614 		struct mlx5_switch_info	info;
1615 
1616 		/* Process slave interface names in the loop. */
1617 		snprintf(tmp_str, sizeof(tmp_str),
1618 			 "/sys/class/net/%s", ifname);
1619 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1620 			DRV_LOG(WARNING, "can not get PCI address"
1621 					 " for netdev \"%s\"", ifname);
1622 			continue;
1623 		}
1624 		if (pci_dev->addr.domain != pci_addr.domain ||
1625 		    pci_dev->addr.bus != pci_addr.bus ||
1626 		    pci_dev->addr.devid != pci_addr.devid ||
1627 		    pci_dev->addr.function != pci_addr.function)
1628 			continue;
1629 		/* Slave interface PCI address match found. */
1630 		fclose(file);
1631 		snprintf(tmp_str, sizeof(tmp_str),
1632 			 "/sys/class/net/%s/phys_port_name", ifname);
1633 		file = fopen(tmp_str, "rb");
1634 		if (!file)
1635 			break;
1636 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1637 		if (fscanf(file, "%32s", tmp_str) == 1)
1638 			mlx5_translate_port_name(tmp_str, &info);
1639 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1640 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1641 			pf = info.port_name;
1642 		break;
1643 	}
1644 	if (file)
1645 		fclose(file);
1646 	return pf;
1647 }
1648 
1649 /**
1650  * DPDK callback to register a PCI device.
1651  *
1652  * This function spawns Ethernet devices out of a given PCI device.
1653  *
1654  * @param[in] pci_drv
1655  *   PCI driver structure (mlx5_driver).
1656  * @param[in] pci_dev
1657  *   PCI device information.
1658  *
1659  * @return
1660  *   0 on success, a negative errno value otherwise and rte_errno is set.
1661  */
1662 int
1663 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1664 		  struct rte_pci_device *pci_dev)
1665 {
1666 	struct ibv_device **ibv_list;
1667 	/*
1668 	 * Number of found IB Devices matching with requested PCI BDF.
1669 	 * nd != 1 means there are multiple IB devices over the same
1670 	 * PCI device and we have representors and master.
1671 	 */
1672 	unsigned int nd = 0;
1673 	/*
1674 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1675 	 * we have the single multiport IB device, and there may be
1676 	 * representors attached to some of found ports.
1677 	 */
1678 	unsigned int np = 0;
1679 	/*
1680 	 * Number of DPDK ethernet devices to Spawn - either over
1681 	 * multiple IB devices or multiple ports of single IB device.
1682 	 * Actually this is the number of iterations to spawn.
1683 	 */
1684 	unsigned int ns = 0;
1685 	/*
1686 	 * Bonding device
1687 	 *   < 0 - no bonding device (single one)
1688 	 *  >= 0 - bonding device (value is slave PF index)
1689 	 */
1690 	int bd = -1;
1691 	struct mlx5_dev_spawn_data *list = NULL;
1692 	struct mlx5_dev_config dev_config;
1693 	unsigned int dev_config_vf;
1694 	int ret;
1695 
1696 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1697 		mlx5_pmd_socket_init();
1698 	ret = mlx5_init_once();
1699 	if (ret) {
1700 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1701 			strerror(rte_errno));
1702 		return -rte_errno;
1703 	}
1704 	errno = 0;
1705 	ibv_list = mlx5_glue->get_device_list(&ret);
1706 	if (!ibv_list) {
1707 		rte_errno = errno ? errno : ENOSYS;
1708 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1709 		return -rte_errno;
1710 	}
1711 	/*
1712 	 * First scan the list of all Infiniband devices to find
1713 	 * matching ones, gathering into the list.
1714 	 */
1715 	struct ibv_device *ibv_match[ret + 1];
1716 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1717 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1718 	unsigned int i;
1719 
1720 	while (ret-- > 0) {
1721 		struct rte_pci_addr pci_addr;
1722 
1723 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1724 		bd = mlx5_device_bond_pci_match
1725 				(ibv_list[ret], pci_dev, nl_rdma);
1726 		if (bd >= 0) {
1727 			/*
1728 			 * Bonding device detected. Only one match is allowed,
1729 			 * the bonding is supported over multi-port IB device,
1730 			 * there should be no matches on representor PCI
1731 			 * functions or non VF LAG bonding devices with
1732 			 * specified address.
1733 			 */
1734 			if (nd) {
1735 				DRV_LOG(ERR,
1736 					"multiple PCI match on bonding device"
1737 					"\"%s\" found", ibv_list[ret]->name);
1738 				rte_errno = ENOENT;
1739 				ret = -rte_errno;
1740 				goto exit;
1741 			}
1742 			DRV_LOG(INFO, "PCI information matches for"
1743 				      " slave %d bonding device \"%s\"",
1744 				      bd, ibv_list[ret]->name);
1745 			ibv_match[nd++] = ibv_list[ret];
1746 			break;
1747 		}
1748 		if (mlx5_dev_to_pci_addr
1749 			(ibv_list[ret]->ibdev_path, &pci_addr))
1750 			continue;
1751 		if (pci_dev->addr.domain != pci_addr.domain ||
1752 		    pci_dev->addr.bus != pci_addr.bus ||
1753 		    pci_dev->addr.devid != pci_addr.devid ||
1754 		    pci_dev->addr.function != pci_addr.function)
1755 			continue;
1756 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1757 			ibv_list[ret]->name);
1758 		ibv_match[nd++] = ibv_list[ret];
1759 	}
1760 	ibv_match[nd] = NULL;
1761 	if (!nd) {
1762 		/* No device matches, just complain and bail out. */
1763 		DRV_LOG(WARNING,
1764 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1765 			" are kernel drivers loaded?",
1766 			pci_dev->addr.domain, pci_dev->addr.bus,
1767 			pci_dev->addr.devid, pci_dev->addr.function);
1768 		rte_errno = ENOENT;
1769 		ret = -rte_errno;
1770 		goto exit;
1771 	}
1772 	if (nd == 1) {
1773 		/*
1774 		 * Found single matching device may have multiple ports.
1775 		 * Each port may be representor, we have to check the port
1776 		 * number and check the representors existence.
1777 		 */
1778 		if (nl_rdma >= 0)
1779 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1780 		if (!np)
1781 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1782 					 " ports number", ibv_match[0]->name);
1783 		if (bd >= 0 && !np) {
1784 			DRV_LOG(ERR, "can not get ports"
1785 				     " for bonding device");
1786 			rte_errno = ENOENT;
1787 			ret = -rte_errno;
1788 			goto exit;
1789 		}
1790 	}
1791 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1792 	if (bd >= 0) {
1793 		/*
1794 		 * This may happen if there is VF LAG kernel support and
1795 		 * application is compiled with older rdma_core library.
1796 		 */
1797 		DRV_LOG(ERR,
1798 			"No kernel/verbs support for VF LAG bonding found.");
1799 		rte_errno = ENOTSUP;
1800 		ret = -rte_errno;
1801 		goto exit;
1802 	}
1803 #endif
1804 	/*
1805 	 * Now we can determine the maximal
1806 	 * amount of devices to be spawned.
1807 	 */
1808 	list = mlx5_malloc(MLX5_MEM_ZERO,
1809 			   sizeof(struct mlx5_dev_spawn_data) *
1810 			   (np ? np : nd),
1811 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1812 	if (!list) {
1813 		DRV_LOG(ERR, "spawn data array allocation failure");
1814 		rte_errno = ENOMEM;
1815 		ret = -rte_errno;
1816 		goto exit;
1817 	}
1818 	if (bd >= 0 || np > 1) {
1819 		/*
1820 		 * Single IB device with multiple ports found,
1821 		 * it may be E-Switch master device and representors.
1822 		 * We have to perform identification through the ports.
1823 		 */
1824 		MLX5_ASSERT(nl_rdma >= 0);
1825 		MLX5_ASSERT(ns == 0);
1826 		MLX5_ASSERT(nd == 1);
1827 		MLX5_ASSERT(np);
1828 		for (i = 1; i <= np; ++i) {
1829 			list[ns].max_port = np;
1830 			list[ns].phys_port = i;
1831 			list[ns].phys_dev = ibv_match[0];
1832 			list[ns].eth_dev = NULL;
1833 			list[ns].pci_dev = pci_dev;
1834 			list[ns].pf_bond = bd;
1835 			list[ns].ifindex = mlx5_nl_ifindex
1836 				(nl_rdma,
1837 				mlx5_os_get_dev_device_name
1838 						(list[ns].phys_dev), i);
1839 			if (!list[ns].ifindex) {
1840 				/*
1841 				 * No network interface index found for the
1842 				 * specified port, it means there is no
1843 				 * representor on this port. It's OK,
1844 				 * there can be disabled ports, for example
1845 				 * if sriov_numvfs < sriov_totalvfs.
1846 				 */
1847 				continue;
1848 			}
1849 			ret = -1;
1850 			if (nl_route >= 0)
1851 				ret = mlx5_nl_switch_info
1852 					       (nl_route,
1853 						list[ns].ifindex,
1854 						&list[ns].info);
1855 			if (ret || (!list[ns].info.representor &&
1856 				    !list[ns].info.master)) {
1857 				/*
1858 				 * We failed to recognize representors with
1859 				 * Netlink, let's try to perform the task
1860 				 * with sysfs.
1861 				 */
1862 				ret =  mlx5_sysfs_switch_info
1863 						(list[ns].ifindex,
1864 						 &list[ns].info);
1865 			}
1866 			if (!ret && bd >= 0) {
1867 				switch (list[ns].info.name_type) {
1868 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1869 					if (list[ns].info.port_name == bd)
1870 						ns++;
1871 					break;
1872 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1873 					/* Fallthrough */
1874 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1875 					if (list[ns].info.pf_num == bd)
1876 						ns++;
1877 					break;
1878 				default:
1879 					break;
1880 				}
1881 				continue;
1882 			}
1883 			if (!ret && (list[ns].info.representor ^
1884 				     list[ns].info.master))
1885 				ns++;
1886 		}
1887 		if (!ns) {
1888 			DRV_LOG(ERR,
1889 				"unable to recognize master/representors"
1890 				" on the IB device with multiple ports");
1891 			rte_errno = ENOENT;
1892 			ret = -rte_errno;
1893 			goto exit;
1894 		}
1895 	} else {
1896 		/*
1897 		 * The existence of several matching entries (nd > 1) means
1898 		 * port representors have been instantiated. No existing Verbs
1899 		 * call nor sysfs entries can tell them apart, this can only
1900 		 * be done through Netlink calls assuming kernel drivers are
1901 		 * recent enough to support them.
1902 		 *
1903 		 * In the event of identification failure through Netlink,
1904 		 * try again through sysfs, then:
1905 		 *
1906 		 * 1. A single IB device matches (nd == 1) with single
1907 		 *    port (np=0/1) and is not a representor, assume
1908 		 *    no switch support.
1909 		 *
1910 		 * 2. Otherwise no safe assumptions can be made;
1911 		 *    complain louder and bail out.
1912 		 */
1913 		for (i = 0; i != nd; ++i) {
1914 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1915 			list[ns].max_port = 1;
1916 			list[ns].phys_port = 1;
1917 			list[ns].phys_dev = ibv_match[i];
1918 			list[ns].eth_dev = NULL;
1919 			list[ns].pci_dev = pci_dev;
1920 			list[ns].pf_bond = -1;
1921 			list[ns].ifindex = 0;
1922 			if (nl_rdma >= 0)
1923 				list[ns].ifindex = mlx5_nl_ifindex
1924 				(nl_rdma,
1925 				mlx5_os_get_dev_device_name
1926 						(list[ns].phys_dev), 1);
1927 			if (!list[ns].ifindex) {
1928 				char ifname[IF_NAMESIZE];
1929 
1930 				/*
1931 				 * Netlink failed, it may happen with old
1932 				 * ib_core kernel driver (before 4.16).
1933 				 * We can assume there is old driver because
1934 				 * here we are processing single ports IB
1935 				 * devices. Let's try sysfs to retrieve
1936 				 * the ifindex. The method works for
1937 				 * master device only.
1938 				 */
1939 				if (nd > 1) {
1940 					/*
1941 					 * Multiple devices found, assume
1942 					 * representors, can not distinguish
1943 					 * master/representor and retrieve
1944 					 * ifindex via sysfs.
1945 					 */
1946 					continue;
1947 				}
1948 				ret = mlx5_get_ifname_sysfs
1949 					(ibv_match[i]->ibdev_path, ifname);
1950 				if (!ret)
1951 					list[ns].ifindex =
1952 						if_nametoindex(ifname);
1953 				if (!list[ns].ifindex) {
1954 					/*
1955 					 * No network interface index found
1956 					 * for the specified device, it means
1957 					 * there it is neither representor
1958 					 * nor master.
1959 					 */
1960 					continue;
1961 				}
1962 			}
1963 			ret = -1;
1964 			if (nl_route >= 0)
1965 				ret = mlx5_nl_switch_info
1966 					       (nl_route,
1967 						list[ns].ifindex,
1968 						&list[ns].info);
1969 			if (ret || (!list[ns].info.representor &&
1970 				    !list[ns].info.master)) {
1971 				/*
1972 				 * We failed to recognize representors with
1973 				 * Netlink, let's try to perform the task
1974 				 * with sysfs.
1975 				 */
1976 				ret =  mlx5_sysfs_switch_info
1977 						(list[ns].ifindex,
1978 						 &list[ns].info);
1979 			}
1980 			if (!ret && (list[ns].info.representor ^
1981 				     list[ns].info.master)) {
1982 				ns++;
1983 			} else if ((nd == 1) &&
1984 				   !list[ns].info.representor &&
1985 				   !list[ns].info.master) {
1986 				/*
1987 				 * Single IB device with
1988 				 * one physical port and
1989 				 * attached network device.
1990 				 * May be SRIOV is not enabled
1991 				 * or there is no representors.
1992 				 */
1993 				DRV_LOG(INFO, "no E-Switch support detected");
1994 				ns++;
1995 				break;
1996 			}
1997 		}
1998 		if (!ns) {
1999 			DRV_LOG(ERR,
2000 				"unable to recognize master/representors"
2001 				" on the multiple IB devices");
2002 			rte_errno = ENOENT;
2003 			ret = -rte_errno;
2004 			goto exit;
2005 		}
2006 	}
2007 	MLX5_ASSERT(ns);
2008 	/*
2009 	 * Sort list to probe devices in natural order for users convenience
2010 	 * (i.e. master first, then representors from lowest to highest ID).
2011 	 */
2012 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2013 	/* Device specific configuration. */
2014 	switch (pci_dev->id.device_id) {
2015 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2016 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2017 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2018 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2019 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2020 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2021 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
2022 		dev_config_vf = 1;
2023 		break;
2024 	default:
2025 		dev_config_vf = 0;
2026 		break;
2027 	}
2028 	for (i = 0; i != ns; ++i) {
2029 		uint32_t restore;
2030 
2031 		/* Default configuration. */
2032 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2033 		dev_config.vf = dev_config_vf;
2034 		dev_config.mps = MLX5_ARG_UNSET;
2035 		dev_config.dbnc = MLX5_ARG_UNSET;
2036 		dev_config.rx_vec_en = 1;
2037 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2038 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2039 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2040 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2041 		dev_config.vf_nl_en = 1;
2042 		dev_config.mr_ext_memseg_en = 1;
2043 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2044 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2045 		dev_config.dv_esw_en = 1;
2046 		dev_config.dv_flow_en = 1;
2047 		dev_config.decap_en = 1;
2048 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2049 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2050 						 &list[i],
2051 						 &dev_config);
2052 		if (!list[i].eth_dev) {
2053 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2054 				break;
2055 			/* Device is disabled or already spawned. Ignore it. */
2056 			continue;
2057 		}
2058 		restore = list[i].eth_dev->data->dev_flags;
2059 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2060 		/* Restore non-PCI flags cleared by the above call. */
2061 		list[i].eth_dev->data->dev_flags |= restore;
2062 		rte_eth_dev_probing_finish(list[i].eth_dev);
2063 	}
2064 	if (i != ns) {
2065 		DRV_LOG(ERR,
2066 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2067 			" encountering an error: %s",
2068 			pci_dev->addr.domain, pci_dev->addr.bus,
2069 			pci_dev->addr.devid, pci_dev->addr.function,
2070 			strerror(rte_errno));
2071 		ret = -rte_errno;
2072 		/* Roll back. */
2073 		while (i--) {
2074 			if (!list[i].eth_dev)
2075 				continue;
2076 			mlx5_dev_close(list[i].eth_dev);
2077 			/* mac_addrs must not be freed because in dev_private */
2078 			list[i].eth_dev->data->mac_addrs = NULL;
2079 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2080 		}
2081 		/* Restore original error. */
2082 		rte_errno = -ret;
2083 	} else {
2084 		ret = 0;
2085 	}
2086 exit:
2087 	/*
2088 	 * Do the routine cleanup:
2089 	 * - close opened Netlink sockets
2090 	 * - free allocated spawn data array
2091 	 * - free the Infiniband device list
2092 	 */
2093 	if (nl_rdma >= 0)
2094 		close(nl_rdma);
2095 	if (nl_route >= 0)
2096 		close(nl_route);
2097 	if (list)
2098 		mlx5_free(list);
2099 	MLX5_ASSERT(ibv_list);
2100 	mlx5_glue->free_device_list(ibv_list);
2101 	return ret;
2102 }
2103 
2104 static int
2105 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2106 {
2107 	char *env;
2108 	int value;
2109 
2110 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2111 	/* Get environment variable to store. */
2112 	env = getenv(MLX5_SHUT_UP_BF);
2113 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2114 	if (config->dbnc == MLX5_ARG_UNSET)
2115 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2116 	else
2117 		setenv(MLX5_SHUT_UP_BF,
2118 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2119 	return value;
2120 }
2121 
2122 static void
2123 mlx5_restore_doorbell_mapping_env(int value)
2124 {
2125 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2126 	/* Restore the original environment variable state. */
2127 	if (value == MLX5_ARG_UNSET)
2128 		unsetenv(MLX5_SHUT_UP_BF);
2129 	else
2130 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2131 }
2132 
2133 /**
2134  * Extract pdn of PD object using DV API.
2135  *
2136  * @param[in] pd
2137  *   Pointer to the verbs PD object.
2138  * @param[out] pdn
2139  *   Pointer to the PD object number variable.
2140  *
2141  * @return
2142  *   0 on success, error value otherwise.
2143  */
2144 int
2145 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2146 {
2147 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2148 	struct mlx5dv_obj obj;
2149 	struct mlx5dv_pd pd_info;
2150 	int ret = 0;
2151 
2152 	obj.pd.in = pd;
2153 	obj.pd.out = &pd_info;
2154 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2155 	if (ret) {
2156 		DRV_LOG(DEBUG, "Fail to get PD object info");
2157 		return ret;
2158 	}
2159 	*pdn = pd_info.pdn;
2160 	return 0;
2161 #else
2162 	(void)pd;
2163 	(void)pdn;
2164 	return -ENOTSUP;
2165 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2166 }
2167 
2168 /**
2169  * Function API to open IB device.
2170  *
2171  * This function calls the Linux glue APIs to open a device.
2172  *
2173  * @param[in] spawn
2174  *   Pointer to the IB device attributes (name, port, etc).
2175  * @param[out] config
2176  *   Pointer to device configuration structure.
2177  * @param[out] sh
2178  *   Pointer to shared context structure.
2179  *
2180  * @return
2181  *   0 on success, a positive error value otherwise.
2182  */
2183 int
2184 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2185 		     const struct mlx5_dev_config *config,
2186 		     struct mlx5_dev_ctx_shared *sh)
2187 {
2188 	int dbmap_env;
2189 	int err = 0;
2190 
2191 	sh->numa_node = spawn->pci_dev->device.numa_node;
2192 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2193 	/*
2194 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2195 	 * before the device creation. The rdma_core library
2196 	 * checks the variable at device creation and
2197 	 * stores the result internally.
2198 	 */
2199 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2200 	/* Try to open IB device with DV first, then usual Verbs. */
2201 	errno = 0;
2202 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2203 	if (sh->ctx) {
2204 		sh->devx = 1;
2205 		DRV_LOG(DEBUG, "DevX is supported");
2206 		/* The device is created, no need for environment. */
2207 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2208 	} else {
2209 		/* The environment variable is still configured. */
2210 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2211 		err = errno ? errno : ENODEV;
2212 		/*
2213 		 * The environment variable is not needed anymore,
2214 		 * all device creation attempts are completed.
2215 		 */
2216 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2217 		if (!sh->ctx)
2218 			return err;
2219 		DRV_LOG(DEBUG, "DevX is NOT supported");
2220 		err = 0;
2221 	}
2222 	return err;
2223 }
2224 
2225 /**
2226  * Install shared asynchronous device events handler.
2227  * This function is implemented to support event sharing
2228  * between multiple ports of single IB device.
2229  *
2230  * @param sh
2231  *   Pointer to mlx5_dev_ctx_shared object.
2232  */
2233 void
2234 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2235 {
2236 	int ret;
2237 	int flags;
2238 
2239 	sh->intr_handle.fd = -1;
2240 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2241 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2242 		    F_SETFL, flags | O_NONBLOCK);
2243 	if (ret) {
2244 		DRV_LOG(INFO, "failed to change file descriptor async event"
2245 			" queue");
2246 	} else {
2247 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2248 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2249 		if (rte_intr_callback_register(&sh->intr_handle,
2250 					mlx5_dev_interrupt_handler, sh)) {
2251 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2252 			sh->intr_handle.fd = -1;
2253 		}
2254 	}
2255 	if (sh->devx) {
2256 #ifdef HAVE_IBV_DEVX_ASYNC
2257 		sh->intr_handle_devx.fd = -1;
2258 		sh->devx_comp =
2259 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2260 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2261 		if (!devx_comp) {
2262 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2263 			return;
2264 		}
2265 		flags = fcntl(devx_comp->fd, F_GETFL);
2266 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2267 		if (ret) {
2268 			DRV_LOG(INFO, "failed to change file descriptor"
2269 				" devx comp");
2270 			return;
2271 		}
2272 		sh->intr_handle_devx.fd = devx_comp->fd;
2273 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2274 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2275 					mlx5_dev_interrupt_handler_devx, sh)) {
2276 			DRV_LOG(INFO, "Fail to install the devx shared"
2277 				" interrupt.");
2278 			sh->intr_handle_devx.fd = -1;
2279 		}
2280 #endif /* HAVE_IBV_DEVX_ASYNC */
2281 	}
2282 }
2283 
2284 /**
2285  * Uninstall shared asynchronous device events handler.
2286  * This function is implemented to support event sharing
2287  * between multiple ports of single IB device.
2288  *
2289  * @param dev
2290  *   Pointer to mlx5_dev_ctx_shared object.
2291  */
2292 void
2293 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2294 {
2295 	if (sh->intr_handle.fd >= 0)
2296 		mlx5_intr_callback_unregister(&sh->intr_handle,
2297 					      mlx5_dev_interrupt_handler, sh);
2298 #ifdef HAVE_IBV_DEVX_ASYNC
2299 	if (sh->intr_handle_devx.fd >= 0)
2300 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2301 				  mlx5_dev_interrupt_handler_devx, sh);
2302 	if (sh->devx_comp)
2303 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2304 #endif
2305 }
2306 
2307 /**
2308  * Read statistics by a named counter.
2309  *
2310  * @param[in] priv
2311  *   Pointer to the private device data structure.
2312  * @param[in] ctr_name
2313  *   Pointer to the name of the statistic counter to read
2314  * @param[out] stat
2315  *   Pointer to read statistic value.
2316  * @return
2317  *   0 on success and stat is valud, 1 if failed to read the value
2318  *   rte_errno is set.
2319  *
2320  */
2321 int
2322 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2323 		      uint64_t *stat)
2324 {
2325 	int fd;
2326 
2327 	if (priv->sh) {
2328 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2329 		      priv->sh->ibdev_path,
2330 		      priv->dev_port,
2331 		      ctr_name);
2332 		fd = open(path, O_RDONLY);
2333 		/*
2334 		 * in switchdev the file location is not per port
2335 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2336 		 */
2337 		if (fd == -1) {
2338 			MKSTR(path1, "%s/hw_counters/%s",
2339 			      priv->sh->ibdev_path,
2340 			      ctr_name);
2341 			fd = open(path1, O_RDONLY);
2342 		}
2343 		if (fd != -1) {
2344 			char buf[21] = {'\0'};
2345 			ssize_t n = read(fd, buf, sizeof(buf));
2346 
2347 			close(fd);
2348 			if (n != -1) {
2349 				*stat = strtoull(buf, NULL, 10);
2350 				return 0;
2351 			}
2352 		}
2353 	}
2354 	*stat = 0;
2355 	return 1;
2356 }
2357 
2358 /**
2359  * Set the reg_mr and dereg_mr call backs
2360  *
2361  * @param reg_mr_cb[out]
2362  *   Pointer to reg_mr func
2363  * @param dereg_mr_cb[out]
2364  *   Pointer to dereg_mr func
2365  *
2366  */
2367 void
2368 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2369 		      mlx5_dereg_mr_t *dereg_mr_cb)
2370 {
2371 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2372 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2373 }
2374 
2375 /**
2376  * Remove a MAC address from device
2377  *
2378  * @param dev
2379  *   Pointer to Ethernet device structure.
2380  * @param index
2381  *   MAC address index.
2382  */
2383 void
2384 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2385 {
2386 	struct mlx5_priv *priv = dev->data->dev_private;
2387 	const int vf = priv->config.vf;
2388 
2389 	if (vf)
2390 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2391 					mlx5_ifindex(dev), priv->mac_own,
2392 					&dev->data->mac_addrs[index], index);
2393 }
2394 
2395 /**
2396  * Adds a MAC address to the device
2397  *
2398  * @param dev
2399  *   Pointer to Ethernet device structure.
2400  * @param mac_addr
2401  *   MAC address to register.
2402  * @param index
2403  *   MAC address index.
2404  *
2405  * @return
2406  *   0 on success, a negative errno value otherwise
2407  */
2408 int
2409 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2410 		     uint32_t index)
2411 {
2412 	struct mlx5_priv *priv = dev->data->dev_private;
2413 	const int vf = priv->config.vf;
2414 	int ret = 0;
2415 
2416 	if (vf)
2417 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2418 					   mlx5_ifindex(dev), priv->mac_own,
2419 					   mac, index);
2420 	return ret;
2421 }
2422 
2423 /**
2424  * Modify a VF MAC address
2425  *
2426  * @param priv
2427  *   Pointer to device private data.
2428  * @param mac_addr
2429  *   MAC address to modify into.
2430  * @param iface_idx
2431  *   Net device interface index
2432  * @param vf_index
2433  *   VF index
2434  *
2435  * @return
2436  *   0 on success, a negative errno value otherwise
2437  */
2438 int
2439 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2440 			   unsigned int iface_idx,
2441 			   struct rte_ether_addr *mac_addr,
2442 			   int vf_index)
2443 {
2444 	return mlx5_nl_vf_mac_addr_modify
2445 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2446 }
2447 
2448 /**
2449  * Set device promiscuous mode
2450  *
2451  * @param dev
2452  *   Pointer to Ethernet device structure.
2453  * @param enable
2454  *   0 - promiscuous is disabled, otherwise - enabled
2455  *
2456  * @return
2457  *   0 on success, a negative error value otherwise
2458  */
2459 int
2460 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2461 {
2462 	struct mlx5_priv *priv = dev->data->dev_private;
2463 
2464 	return mlx5_nl_promisc(priv->nl_socket_route,
2465 			       mlx5_ifindex(dev), !!enable);
2466 }
2467 
2468 /**
2469  * Set device promiscuous mode
2470  *
2471  * @param dev
2472  *   Pointer to Ethernet device structure.
2473  * @param enable
2474  *   0 - all multicase is disabled, otherwise - enabled
2475  *
2476  * @return
2477  *   0 on success, a negative error value otherwise
2478  */
2479 int
2480 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2481 {
2482 	struct mlx5_priv *priv = dev->data->dev_private;
2483 
2484 	return mlx5_nl_allmulti(priv->nl_socket_route,
2485 				mlx5_ifindex(dev), !!enable);
2486 }
2487 
2488 /**
2489  * Flush device MAC addresses
2490  *
2491  * @param dev
2492  *   Pointer to Ethernet device structure.
2493  *
2494  */
2495 void
2496 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2497 {
2498 	struct mlx5_priv *priv = dev->data->dev_private;
2499 
2500 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2501 			       dev->data->mac_addrs,
2502 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2503 }
2504 
2505 const struct eth_dev_ops mlx5_os_dev_ops = {
2506 	.dev_configure = mlx5_dev_configure,
2507 	.dev_start = mlx5_dev_start,
2508 	.dev_stop = mlx5_dev_stop,
2509 	.dev_set_link_down = mlx5_set_link_down,
2510 	.dev_set_link_up = mlx5_set_link_up,
2511 	.dev_close = mlx5_dev_close,
2512 	.promiscuous_enable = mlx5_promiscuous_enable,
2513 	.promiscuous_disable = mlx5_promiscuous_disable,
2514 	.allmulticast_enable = mlx5_allmulticast_enable,
2515 	.allmulticast_disable = mlx5_allmulticast_disable,
2516 	.link_update = mlx5_link_update,
2517 	.stats_get = mlx5_stats_get,
2518 	.stats_reset = mlx5_stats_reset,
2519 	.xstats_get = mlx5_xstats_get,
2520 	.xstats_reset = mlx5_xstats_reset,
2521 	.xstats_get_names = mlx5_xstats_get_names,
2522 	.fw_version_get = mlx5_fw_version_get,
2523 	.dev_infos_get = mlx5_dev_infos_get,
2524 	.read_clock = mlx5_txpp_read_clock,
2525 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2526 	.vlan_filter_set = mlx5_vlan_filter_set,
2527 	.rx_queue_setup = mlx5_rx_queue_setup,
2528 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2529 	.tx_queue_setup = mlx5_tx_queue_setup,
2530 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2531 	.rx_queue_release = mlx5_rx_queue_release,
2532 	.tx_queue_release = mlx5_tx_queue_release,
2533 	.rx_queue_start = mlx5_rx_queue_start,
2534 	.rx_queue_stop = mlx5_rx_queue_stop,
2535 	.tx_queue_start = mlx5_tx_queue_start,
2536 	.tx_queue_stop = mlx5_tx_queue_stop,
2537 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2538 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2539 	.mac_addr_remove = mlx5_mac_addr_remove,
2540 	.mac_addr_add = mlx5_mac_addr_add,
2541 	.mac_addr_set = mlx5_mac_addr_set,
2542 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2543 	.mtu_set = mlx5_dev_set_mtu,
2544 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2545 	.vlan_offload_set = mlx5_vlan_offload_set,
2546 	.reta_update = mlx5_dev_rss_reta_update,
2547 	.reta_query = mlx5_dev_rss_reta_query,
2548 	.rss_hash_update = mlx5_rss_hash_update,
2549 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2550 	.filter_ctrl = mlx5_dev_filter_ctrl,
2551 	.rxq_info_get = mlx5_rxq_info_get,
2552 	.txq_info_get = mlx5_txq_info_get,
2553 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2554 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2555 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2556 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2557 	.is_removed = mlx5_is_removed,
2558 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2559 	.get_module_info = mlx5_get_module_info,
2560 	.get_module_eeprom = mlx5_get_module_eeprom,
2561 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2562 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2563 };
2564 
2565 /* Available operations from secondary process. */
2566 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2567 	.stats_get = mlx5_stats_get,
2568 	.stats_reset = mlx5_stats_reset,
2569 	.xstats_get = mlx5_xstats_get,
2570 	.xstats_reset = mlx5_xstats_reset,
2571 	.xstats_get_names = mlx5_xstats_get_names,
2572 	.fw_version_get = mlx5_fw_version_get,
2573 	.dev_infos_get = mlx5_dev_infos_get,
2574 	.read_clock = mlx5_txpp_read_clock,
2575 	.rx_queue_start = mlx5_rx_queue_start,
2576 	.rx_queue_stop = mlx5_rx_queue_stop,
2577 	.tx_queue_start = mlx5_tx_queue_start,
2578 	.tx_queue_stop = mlx5_tx_queue_stop,
2579 	.rxq_info_get = mlx5_rxq_info_get,
2580 	.txq_info_get = mlx5_txq_info_get,
2581 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2582 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2583 	.get_module_info = mlx5_get_module_info,
2584 	.get_module_eeprom = mlx5_get_module_eeprom,
2585 };
2586 
2587 /* Available operations in flow isolated mode. */
2588 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2589 	.dev_configure = mlx5_dev_configure,
2590 	.dev_start = mlx5_dev_start,
2591 	.dev_stop = mlx5_dev_stop,
2592 	.dev_set_link_down = mlx5_set_link_down,
2593 	.dev_set_link_up = mlx5_set_link_up,
2594 	.dev_close = mlx5_dev_close,
2595 	.promiscuous_enable = mlx5_promiscuous_enable,
2596 	.promiscuous_disable = mlx5_promiscuous_disable,
2597 	.allmulticast_enable = mlx5_allmulticast_enable,
2598 	.allmulticast_disable = mlx5_allmulticast_disable,
2599 	.link_update = mlx5_link_update,
2600 	.stats_get = mlx5_stats_get,
2601 	.stats_reset = mlx5_stats_reset,
2602 	.xstats_get = mlx5_xstats_get,
2603 	.xstats_reset = mlx5_xstats_reset,
2604 	.xstats_get_names = mlx5_xstats_get_names,
2605 	.fw_version_get = mlx5_fw_version_get,
2606 	.dev_infos_get = mlx5_dev_infos_get,
2607 	.read_clock = mlx5_txpp_read_clock,
2608 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2609 	.vlan_filter_set = mlx5_vlan_filter_set,
2610 	.rx_queue_setup = mlx5_rx_queue_setup,
2611 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2612 	.tx_queue_setup = mlx5_tx_queue_setup,
2613 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2614 	.rx_queue_release = mlx5_rx_queue_release,
2615 	.tx_queue_release = mlx5_tx_queue_release,
2616 	.rx_queue_start = mlx5_rx_queue_start,
2617 	.rx_queue_stop = mlx5_rx_queue_stop,
2618 	.tx_queue_start = mlx5_tx_queue_start,
2619 	.tx_queue_stop = mlx5_tx_queue_stop,
2620 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2621 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2622 	.mac_addr_remove = mlx5_mac_addr_remove,
2623 	.mac_addr_add = mlx5_mac_addr_add,
2624 	.mac_addr_set = mlx5_mac_addr_set,
2625 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2626 	.mtu_set = mlx5_dev_set_mtu,
2627 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2628 	.vlan_offload_set = mlx5_vlan_offload_set,
2629 	.filter_ctrl = mlx5_dev_filter_ctrl,
2630 	.rxq_info_get = mlx5_rxq_info_get,
2631 	.txq_info_get = mlx5_txq_info_get,
2632 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2633 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2634 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2635 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2636 	.is_removed = mlx5_is_removed,
2637 	.get_module_info = mlx5_get_module_info,
2638 	.get_module_eeprom = mlx5_get_module_eeprom,
2639 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2640 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2641 };
2642