xref: /dpdk/drivers/net/mlx5/linux/mlx5_os.c (revision 3ec73abeed5216b18493deb2cd67970be5a3f047)
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  * Spawn an Ethernet device from Verbs information.
566  *
567  * @param dpdk_dev
568  *   Backing DPDK device.
569  * @param spawn
570  *   Verbs device parameters (name, port, switch_info) to spawn.
571  * @param config
572  *   Device configuration parameters.
573  *
574  * @return
575  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
576  *   is set. The following errors are defined:
577  *
578  *   EBUSY: device is not supposed to be spawned.
579  *   EEXIST: device is already spawned
580  */
581 static struct rte_eth_dev *
582 mlx5_dev_spawn(struct rte_device *dpdk_dev,
583 	       struct mlx5_dev_spawn_data *spawn,
584 	       struct mlx5_dev_config *config)
585 {
586 	const struct mlx5_switch_info *switch_info = &spawn->info;
587 	struct mlx5_dev_ctx_shared *sh = NULL;
588 	struct ibv_port_attr port_attr;
589 	struct mlx5dv_context dv_attr = { .comp_mask = 0 };
590 	struct rte_eth_dev *eth_dev = NULL;
591 	struct mlx5_priv *priv = NULL;
592 	int err = 0;
593 	unsigned int hw_padding = 0;
594 	unsigned int mps;
595 	unsigned int cqe_comp;
596 	unsigned int cqe_pad = 0;
597 	unsigned int tunnel_en = 0;
598 	unsigned int mpls_en = 0;
599 	unsigned int swp = 0;
600 	unsigned int mprq = 0;
601 	unsigned int mprq_min_stride_size_n = 0;
602 	unsigned int mprq_max_stride_size_n = 0;
603 	unsigned int mprq_min_stride_num_n = 0;
604 	unsigned int mprq_max_stride_num_n = 0;
605 	struct rte_ether_addr mac;
606 	char name[RTE_ETH_NAME_MAX_LEN];
607 	int own_domain_id = 0;
608 	uint16_t port_id;
609 	unsigned int i;
610 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
611 	struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
612 #endif
613 
614 	/* Determine if this port representor is supposed to be spawned. */
615 	if (switch_info->representor && dpdk_dev->devargs) {
616 		struct rte_eth_devargs eth_da;
617 
618 		err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
619 		if (err) {
620 			rte_errno = -err;
621 			DRV_LOG(ERR, "failed to process device arguments: %s",
622 				strerror(rte_errno));
623 			return NULL;
624 		}
625 		for (i = 0; i < eth_da.nb_representor_ports; ++i)
626 			if (eth_da.representor_ports[i] ==
627 			    (uint16_t)switch_info->port_name)
628 				break;
629 		if (i == eth_da.nb_representor_ports) {
630 			rte_errno = EBUSY;
631 			return NULL;
632 		}
633 	}
634 	/* Build device name. */
635 	if (spawn->pf_bond <  0) {
636 		/* Single device. */
637 		if (!switch_info->representor)
638 			strlcpy(name, dpdk_dev->name, sizeof(name));
639 		else
640 			snprintf(name, sizeof(name), "%s_representor_%u",
641 				 dpdk_dev->name, switch_info->port_name);
642 	} else {
643 		/* Bonding device. */
644 		if (!switch_info->representor)
645 			snprintf(name, sizeof(name), "%s_%s",
646 				 dpdk_dev->name,
647 				 mlx5_os_get_dev_device_name(spawn->phys_dev));
648 		else
649 			snprintf(name, sizeof(name), "%s_%s_representor_%u",
650 				 dpdk_dev->name,
651 				 mlx5_os_get_dev_device_name(spawn->phys_dev),
652 				 switch_info->port_name);
653 	}
654 	/* check if the device is already spawned */
655 	if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
656 		rte_errno = EEXIST;
657 		return NULL;
658 	}
659 	DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
660 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
661 		struct mlx5_mp_id mp_id;
662 
663 		eth_dev = rte_eth_dev_attach_secondary(name);
664 		if (eth_dev == NULL) {
665 			DRV_LOG(ERR, "can not attach rte ethdev");
666 			rte_errno = ENOMEM;
667 			return NULL;
668 		}
669 		eth_dev->device = dpdk_dev;
670 		eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
671 		eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
672 		eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
673 		err = mlx5_proc_priv_init(eth_dev);
674 		if (err)
675 			return NULL;
676 		mp_id.port_id = eth_dev->data->port_id;
677 		strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
678 		/* Receive command fd from primary process */
679 		err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
680 		if (err < 0)
681 			goto err_secondary;
682 		/* Remap UAR for Tx queues. */
683 		err = mlx5_tx_uar_init_secondary(eth_dev, err);
684 		if (err)
685 			goto err_secondary;
686 		/*
687 		 * Ethdev pointer is still required as input since
688 		 * the primary device is not accessible from the
689 		 * secondary process.
690 		 */
691 		eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
692 		eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
693 		return eth_dev;
694 err_secondary:
695 		mlx5_dev_close(eth_dev);
696 		return NULL;
697 	}
698 	/*
699 	 * Some parameters ("tx_db_nc" in particularly) are needed in
700 	 * advance to create dv/verbs device context. We proceed the
701 	 * devargs here to get ones, and later proceed devargs again
702 	 * to override some hardware settings.
703 	 */
704 	err = mlx5_args(config, dpdk_dev->devargs);
705 	if (err) {
706 		err = rte_errno;
707 		DRV_LOG(ERR, "failed to process device arguments: %s",
708 			strerror(rte_errno));
709 		goto error;
710 	}
711 	mlx5_malloc_mem_select(config->sys_mem_en);
712 	sh = mlx5_alloc_shared_dev_ctx(spawn, config);
713 	if (!sh)
714 		return NULL;
715 	config->devx = sh->devx;
716 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
717 	config->dest_tir = 1;
718 #endif
719 #ifdef HAVE_IBV_MLX5_MOD_SWP
720 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
721 #endif
722 	/*
723 	 * Multi-packet send is supported by ConnectX-4 Lx PF as well
724 	 * as all ConnectX-5 devices.
725 	 */
726 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
727 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
728 #endif
729 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
730 	dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
731 #endif
732 	mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
733 	if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
734 		if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
735 			DRV_LOG(DEBUG, "enhanced MPW is supported");
736 			mps = MLX5_MPW_ENHANCED;
737 		} else {
738 			DRV_LOG(DEBUG, "MPW is supported");
739 			mps = MLX5_MPW;
740 		}
741 	} else {
742 		DRV_LOG(DEBUG, "MPW isn't supported");
743 		mps = MLX5_MPW_DISABLED;
744 	}
745 #ifdef HAVE_IBV_MLX5_MOD_SWP
746 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
747 		swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
748 	DRV_LOG(DEBUG, "SWP support: %u", swp);
749 #endif
750 	config->swp = !!swp;
751 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
752 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
753 		struct mlx5dv_striding_rq_caps mprq_caps =
754 			dv_attr.striding_rq_caps;
755 
756 		DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
757 			mprq_caps.min_single_stride_log_num_of_bytes);
758 		DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
759 			mprq_caps.max_single_stride_log_num_of_bytes);
760 		DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
761 			mprq_caps.min_single_wqe_log_num_of_strides);
762 		DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
763 			mprq_caps.max_single_wqe_log_num_of_strides);
764 		DRV_LOG(DEBUG, "\tsupported_qpts: %d",
765 			mprq_caps.supported_qpts);
766 		DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
767 		mprq = 1;
768 		mprq_min_stride_size_n =
769 			mprq_caps.min_single_stride_log_num_of_bytes;
770 		mprq_max_stride_size_n =
771 			mprq_caps.max_single_stride_log_num_of_bytes;
772 		mprq_min_stride_num_n =
773 			mprq_caps.min_single_wqe_log_num_of_strides;
774 		mprq_max_stride_num_n =
775 			mprq_caps.max_single_wqe_log_num_of_strides;
776 	}
777 #endif
778 	if (RTE_CACHE_LINE_SIZE == 128 &&
779 	    !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
780 		cqe_comp = 0;
781 	else
782 		cqe_comp = 1;
783 	config->cqe_comp = cqe_comp;
784 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
785 	/* Whether device supports 128B Rx CQE padding. */
786 	cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
787 		  (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
788 #endif
789 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
790 	if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
791 		tunnel_en = ((dv_attr.tunnel_offloads_caps &
792 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
793 			     (dv_attr.tunnel_offloads_caps &
794 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
795 			     (dv_attr.tunnel_offloads_caps &
796 			      MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
797 	}
798 	DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
799 		tunnel_en ? "" : "not ");
800 #else
801 	DRV_LOG(WARNING,
802 		"tunnel offloading disabled due to old OFED/rdma-core version");
803 #endif
804 	config->tunnel_en = tunnel_en;
805 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
806 	mpls_en = ((dv_attr.tunnel_offloads_caps &
807 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
808 		   (dv_attr.tunnel_offloads_caps &
809 		    MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
810 	DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
811 		mpls_en ? "" : "not ");
812 #else
813 	DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
814 		" old OFED/rdma-core version or firmware configuration");
815 #endif
816 	config->mpls_en = mpls_en;
817 	/* Check port status. */
818 	err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
819 	if (err) {
820 		DRV_LOG(ERR, "port query failed: %s", strerror(err));
821 		goto error;
822 	}
823 	if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
824 		DRV_LOG(ERR, "port is not configured in Ethernet mode");
825 		err = EINVAL;
826 		goto error;
827 	}
828 	if (port_attr.state != IBV_PORT_ACTIVE)
829 		DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
830 			mlx5_glue->port_state_str(port_attr.state),
831 			port_attr.state);
832 	/* Allocate private eth device data. */
833 	priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
834 			   sizeof(*priv),
835 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
836 	if (priv == NULL) {
837 		DRV_LOG(ERR, "priv allocation failure");
838 		err = ENOMEM;
839 		goto error;
840 	}
841 	priv->sh = sh;
842 	priv->dev_port = spawn->phys_port;
843 	priv->pci_dev = spawn->pci_dev;
844 	priv->mtu = RTE_ETHER_MTU;
845 	priv->mp_id.port_id = port_id;
846 	strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
847 	/* Some internal functions rely on Netlink sockets, open them now. */
848 	priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
849 	priv->nl_socket_route =	mlx5_nl_init(NETLINK_ROUTE);
850 	priv->representor = !!switch_info->representor;
851 	priv->master = !!switch_info->master;
852 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
853 	priv->vport_meta_tag = 0;
854 	priv->vport_meta_mask = 0;
855 	priv->pf_bond = spawn->pf_bond;
856 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
857 	/*
858 	 * The DevX port query API is implemented. E-Switch may use
859 	 * either vport or reg_c[0] metadata register to match on
860 	 * vport index. The engaged part of metadata register is
861 	 * defined by mask.
862 	 */
863 	if (switch_info->representor || switch_info->master) {
864 		devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
865 				      MLX5DV_DEVX_PORT_MATCH_REG_C_0;
866 		err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
867 						 &devx_port);
868 		if (err) {
869 			DRV_LOG(WARNING,
870 				"can't query devx port %d on device %s",
871 				spawn->phys_port,
872 				mlx5_os_get_dev_device_name(spawn->phys_dev));
873 			devx_port.comp_mask = 0;
874 		}
875 	}
876 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
877 		priv->vport_meta_tag = devx_port.reg_c_0.value;
878 		priv->vport_meta_mask = devx_port.reg_c_0.mask;
879 		if (!priv->vport_meta_mask) {
880 			DRV_LOG(ERR, "vport zero mask for port %d"
881 				     " on bonding device %s",
882 				     spawn->phys_port,
883 				     mlx5_os_get_dev_device_name
884 							(spawn->phys_dev));
885 			err = ENOTSUP;
886 			goto error;
887 		}
888 		if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
889 			DRV_LOG(ERR, "invalid vport tag for port %d"
890 				     " on bonding device %s",
891 				     spawn->phys_port,
892 				     mlx5_os_get_dev_device_name
893 							(spawn->phys_dev));
894 			err = ENOTSUP;
895 			goto error;
896 		}
897 	}
898 	if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
899 		priv->vport_id = devx_port.vport_num;
900 	} else if (spawn->pf_bond >= 0) {
901 		DRV_LOG(ERR, "can't deduce vport index for port %d"
902 			     " on bonding device %s",
903 			     spawn->phys_port,
904 			     mlx5_os_get_dev_device_name(spawn->phys_dev));
905 		err = ENOTSUP;
906 		goto error;
907 	} else {
908 		/* Suppose vport index in compatible way. */
909 		priv->vport_id = switch_info->representor ?
910 				 switch_info->port_name + 1 : -1;
911 	}
912 #else
913 	/*
914 	 * Kernel/rdma_core support single E-Switch per PF configurations
915 	 * only and vport_id field contains the vport index for
916 	 * associated VF, which is deduced from representor port name.
917 	 * For example, let's have the IB device port 10, it has
918 	 * attached network device eth0, which has port name attribute
919 	 * pf0vf2, we can deduce the VF number as 2, and set vport index
920 	 * as 3 (2+1). This assigning schema should be changed if the
921 	 * multiple E-Switch instances per PF configurations or/and PCI
922 	 * subfunctions are added.
923 	 */
924 	priv->vport_id = switch_info->representor ?
925 			 switch_info->port_name + 1 : -1;
926 #endif
927 	/* representor_id field keeps the unmodified VF index. */
928 	priv->representor_id = switch_info->representor ?
929 			       switch_info->port_name : -1;
930 	/*
931 	 * Look for sibling devices in order to reuse their switch domain
932 	 * if any, otherwise allocate one.
933 	 */
934 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
935 		const struct mlx5_priv *opriv =
936 			rte_eth_devices[port_id].data->dev_private;
937 
938 		if (!opriv ||
939 		    opriv->sh != priv->sh ||
940 			opriv->domain_id ==
941 			RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
942 			continue;
943 		priv->domain_id = opriv->domain_id;
944 		break;
945 	}
946 	if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
947 		err = rte_eth_switch_domain_alloc(&priv->domain_id);
948 		if (err) {
949 			err = rte_errno;
950 			DRV_LOG(ERR, "unable to allocate switch domain: %s",
951 				strerror(rte_errno));
952 			goto error;
953 		}
954 		own_domain_id = 1;
955 	}
956 	/* Override some values set by hardware configuration. */
957 	mlx5_args(config, dpdk_dev->devargs);
958 	err = mlx5_dev_check_sibling_config(priv, config);
959 	if (err)
960 		goto error;
961 	config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
962 			    IBV_DEVICE_RAW_IP_CSUM);
963 	DRV_LOG(DEBUG, "checksum offloading is %ssupported",
964 		(config->hw_csum ? "" : "not "));
965 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
966 	!defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
967 	DRV_LOG(DEBUG, "counters are not supported");
968 #endif
969 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
970 	if (config->dv_flow_en) {
971 		DRV_LOG(WARNING, "DV flow is not supported");
972 		config->dv_flow_en = 0;
973 	}
974 #endif
975 	config->ind_table_max_size =
976 		sh->device_attr.max_rwq_indirection_table_size;
977 	/*
978 	 * Remove this check once DPDK supports larger/variable
979 	 * indirection tables.
980 	 */
981 	if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
982 		config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
983 	DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
984 		config->ind_table_max_size);
985 	config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
986 				  IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
987 	DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
988 		(config->hw_vlan_strip ? "" : "not "));
989 	config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
990 				 IBV_RAW_PACKET_CAP_SCATTER_FCS);
991 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
992 	hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
993 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
994 	hw_padding = !!(sh->device_attr.device_cap_flags_ex &
995 			IBV_DEVICE_PCI_WRITE_END_PADDING);
996 #endif
997 	if (config->hw_padding && !hw_padding) {
998 		DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
999 		config->hw_padding = 0;
1000 	} else if (config->hw_padding) {
1001 		DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1002 	}
1003 	config->tso = (sh->device_attr.max_tso > 0 &&
1004 		      (sh->device_attr.tso_supported_qpts &
1005 		       (1 << IBV_QPT_RAW_PACKET)));
1006 	if (config->tso)
1007 		config->tso_max_payload_sz = sh->device_attr.max_tso;
1008 	/*
1009 	 * MPW is disabled by default, while the Enhanced MPW is enabled
1010 	 * by default.
1011 	 */
1012 	if (config->mps == MLX5_ARG_UNSET)
1013 		config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1014 							  MLX5_MPW_DISABLED;
1015 	else
1016 		config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1017 	DRV_LOG(INFO, "%sMPS is %s",
1018 		config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1019 		config->mps == MLX5_MPW ? "legacy " : "",
1020 		config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1021 	if (config->cqe_comp && !cqe_comp) {
1022 		DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1023 		config->cqe_comp = 0;
1024 	}
1025 	if (config->cqe_pad && !cqe_pad) {
1026 		DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1027 		config->cqe_pad = 0;
1028 	} else if (config->cqe_pad) {
1029 		DRV_LOG(INFO, "Rx CQE padding is enabled");
1030 	}
1031 	if (config->devx) {
1032 		priv->counter_fallback = 0;
1033 		err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1034 		if (err) {
1035 			err = -err;
1036 			goto error;
1037 		}
1038 		if (!config->hca_attr.flow_counters_dump)
1039 			priv->counter_fallback = 1;
1040 #ifndef HAVE_IBV_DEVX_ASYNC
1041 		priv->counter_fallback = 1;
1042 #endif
1043 		if (priv->counter_fallback)
1044 			DRV_LOG(INFO, "Use fall-back DV counter management");
1045 		/* Check for LRO support. */
1046 		if (config->dest_tir && config->hca_attr.lro_cap &&
1047 		    config->dv_flow_en) {
1048 			/* TBD check tunnel lro caps. */
1049 			config->lro.supported = config->hca_attr.lro_cap;
1050 			DRV_LOG(DEBUG, "Device supports LRO");
1051 			/*
1052 			 * If LRO timeout is not configured by application,
1053 			 * use the minimal supported value.
1054 			 */
1055 			if (!config->lro.timeout)
1056 				config->lro.timeout =
1057 				config->hca_attr.lro_timer_supported_periods[0];
1058 			DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1059 				config->lro.timeout);
1060 		}
1061 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1062 		if (config->hca_attr.qos.sup &&
1063 		    config->hca_attr.qos.srtcm_sup &&
1064 		    config->dv_flow_en) {
1065 			uint8_t reg_c_mask =
1066 				config->hca_attr.qos.flow_meter_reg_c_ids;
1067 			/*
1068 			 * Meter needs two REG_C's for color match and pre-sfx
1069 			 * flow match. Here get the REG_C for color match.
1070 			 * REG_C_0 and REG_C_1 is reserved for metadata feature.
1071 			 */
1072 			reg_c_mask &= 0xfc;
1073 			if (__builtin_popcount(reg_c_mask) < 1) {
1074 				priv->mtr_en = 0;
1075 				DRV_LOG(WARNING, "No available register for"
1076 					" meter.");
1077 			} else {
1078 				priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1079 						      REG_C_0;
1080 				priv->mtr_en = 1;
1081 				priv->mtr_reg_share =
1082 				      config->hca_attr.qos.flow_meter_reg_share;
1083 				DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1084 					priv->mtr_color_reg);
1085 			}
1086 		}
1087 #endif
1088 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1089 		if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1090 		    config->dv_flow_en) {
1091 			priv->sampler_en = 1;
1092 			DRV_LOG(DEBUG, "The Sampler enabled!\n");
1093 		} else {
1094 			priv->sampler_en = 0;
1095 			if (!config->hca_attr.log_max_ft_sampler_num)
1096 				DRV_LOG(WARNING, "No available register for"
1097 						" Sampler.");
1098 			else
1099 				DRV_LOG(DEBUG, "DV flow is not supported!\n");
1100 		}
1101 #endif
1102 	}
1103 	if (config->tx_pp) {
1104 		DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1105 			config->hca_attr.dev_freq_khz);
1106 		DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1107 			config->hca_attr.qos.packet_pacing ? "" : "not ");
1108 		DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1109 			config->hca_attr.cross_channel ? "" : "not ");
1110 		DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1111 			config->hca_attr.wqe_index_ignore ? "" : "not ");
1112 		DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1113 			config->hca_attr.non_wire_sq ? "" : "not ");
1114 		DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1115 			config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1116 			config->hca_attr.log_max_static_sq_wq);
1117 		DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1118 			config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1119 		if (!config->devx) {
1120 			DRV_LOG(ERR, "DevX is required for packet pacing");
1121 			err = ENODEV;
1122 			goto error;
1123 		}
1124 		if (!config->hca_attr.qos.packet_pacing) {
1125 			DRV_LOG(ERR, "Packet pacing is not supported");
1126 			err = ENODEV;
1127 			goto error;
1128 		}
1129 		if (!config->hca_attr.cross_channel) {
1130 			DRV_LOG(ERR, "Cross channel operations are"
1131 				     " required for packet pacing");
1132 			err = ENODEV;
1133 			goto error;
1134 		}
1135 		if (!config->hca_attr.wqe_index_ignore) {
1136 			DRV_LOG(ERR, "WQE index ignore feature is"
1137 				     " required for packet pacing");
1138 			err = ENODEV;
1139 			goto error;
1140 		}
1141 		if (!config->hca_attr.non_wire_sq) {
1142 			DRV_LOG(ERR, "Non-wire SQ feature is"
1143 				     " required for packet pacing");
1144 			err = ENODEV;
1145 			goto error;
1146 		}
1147 		if (!config->hca_attr.log_max_static_sq_wq) {
1148 			DRV_LOG(ERR, "Static WQE SQ feature is"
1149 				     " required for packet pacing");
1150 			err = ENODEV;
1151 			goto error;
1152 		}
1153 		if (!config->hca_attr.qos.wqe_rate_pp) {
1154 			DRV_LOG(ERR, "WQE rate mode is required"
1155 				     " for packet pacing");
1156 			err = ENODEV;
1157 			goto error;
1158 		}
1159 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1160 		DRV_LOG(ERR, "DevX does not provide UAR offset,"
1161 			     " can't create queues for packet pacing");
1162 		err = ENODEV;
1163 		goto error;
1164 #endif
1165 	}
1166 	if (config->devx) {
1167 		uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1168 
1169 		err = config->hca_attr.access_register_user ?
1170 			mlx5_devx_cmd_register_read
1171 				(sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1172 				reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1173 		if (!err) {
1174 			uint32_t ts_mode;
1175 
1176 			/* MTUTC register is read successfully. */
1177 			ts_mode = MLX5_GET(register_mtutc, reg,
1178 					   time_stamp_mode);
1179 			if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1180 				config->rt_timestamp = 1;
1181 		} else {
1182 			/* Kernel does not support register reading. */
1183 			if (config->hca_attr.dev_freq_khz ==
1184 						 (NS_PER_S / MS_PER_S))
1185 				config->rt_timestamp = 1;
1186 		}
1187 	}
1188 	/*
1189 	 * If HW has bug working with tunnel packet decapsulation and
1190 	 * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1191 	 * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1192 	 */
1193 	if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1194 		config->hw_fcs_strip = 0;
1195 	DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1196 		(config->hw_fcs_strip ? "" : "not "));
1197 	if (config->mprq.enabled && mprq) {
1198 		if (config->mprq.stride_num_n &&
1199 		    (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1200 		     config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1201 			config->mprq.stride_num_n =
1202 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1203 						mprq_min_stride_num_n),
1204 					mprq_max_stride_num_n);
1205 			DRV_LOG(WARNING,
1206 				"the number of strides"
1207 				" for Multi-Packet RQ is out of range,"
1208 				" setting default value (%u)",
1209 				1 << config->mprq.stride_num_n);
1210 		}
1211 		if (config->mprq.stride_size_n &&
1212 		    (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1213 		     config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1214 			config->mprq.stride_size_n =
1215 				RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1216 						mprq_min_stride_size_n),
1217 					mprq_max_stride_size_n);
1218 			DRV_LOG(WARNING,
1219 				"the size of a stride"
1220 				" for Multi-Packet RQ is out of range,"
1221 				" setting default value (%u)",
1222 				1 << config->mprq.stride_size_n);
1223 		}
1224 		config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1225 		config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1226 	} else if (config->mprq.enabled && !mprq) {
1227 		DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1228 		config->mprq.enabled = 0;
1229 	}
1230 	if (config->max_dump_files_num == 0)
1231 		config->max_dump_files_num = 128;
1232 	eth_dev = rte_eth_dev_allocate(name);
1233 	if (eth_dev == NULL) {
1234 		DRV_LOG(ERR, "can not allocate rte ethdev");
1235 		err = ENOMEM;
1236 		goto error;
1237 	}
1238 	if (priv->representor) {
1239 		eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1240 		eth_dev->data->representor_id = priv->representor_id;
1241 	}
1242 	/*
1243 	 * Store associated network device interface index. This index
1244 	 * is permanent throughout the lifetime of device. So, we may store
1245 	 * the ifindex here and use the cached value further.
1246 	 */
1247 	MLX5_ASSERT(spawn->ifindex);
1248 	priv->if_index = spawn->ifindex;
1249 	if (priv->pf_bond >= 0 && priv->master) {
1250 		/* Get bond interface info */
1251 		err = mlx5_sysfs_bond_info(priv->if_index,
1252 				     &priv->bond_ifindex,
1253 				     priv->bond_name);
1254 		if (err)
1255 			DRV_LOG(ERR, "unable to get bond info: %s",
1256 				strerror(rte_errno));
1257 		else
1258 			DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1259 				priv->if_index, priv->bond_ifindex,
1260 				priv->bond_name);
1261 	}
1262 	eth_dev->data->dev_private = priv;
1263 	priv->dev_data = eth_dev->data;
1264 	eth_dev->data->mac_addrs = priv->mac;
1265 	eth_dev->device = dpdk_dev;
1266 	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1267 	/* Configure the first MAC address by default. */
1268 	if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1269 		DRV_LOG(ERR,
1270 			"port %u cannot get MAC address, is mlx5_en"
1271 			" loaded? (errno: %s)",
1272 			eth_dev->data->port_id, strerror(rte_errno));
1273 		err = ENODEV;
1274 		goto error;
1275 	}
1276 	DRV_LOG(INFO,
1277 		"port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1278 		eth_dev->data->port_id,
1279 		mac.addr_bytes[0], mac.addr_bytes[1],
1280 		mac.addr_bytes[2], mac.addr_bytes[3],
1281 		mac.addr_bytes[4], mac.addr_bytes[5]);
1282 #ifdef RTE_LIBRTE_MLX5_DEBUG
1283 	{
1284 		char ifname[IF_NAMESIZE];
1285 
1286 		if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1287 			DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1288 				eth_dev->data->port_id, ifname);
1289 		else
1290 			DRV_LOG(DEBUG, "port %u ifname is unknown",
1291 				eth_dev->data->port_id);
1292 	}
1293 #endif
1294 	/* Get actual MTU if possible. */
1295 	err = mlx5_get_mtu(eth_dev, &priv->mtu);
1296 	if (err) {
1297 		err = rte_errno;
1298 		goto error;
1299 	}
1300 	DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1301 		priv->mtu);
1302 	/* Initialize burst functions to prevent crashes before link-up. */
1303 	eth_dev->rx_pkt_burst = removed_rx_burst;
1304 	eth_dev->tx_pkt_burst = removed_tx_burst;
1305 	eth_dev->dev_ops = &mlx5_os_dev_ops;
1306 	eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1307 	eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1308 	eth_dev->rx_queue_count = mlx5_rx_queue_count;
1309 	/* Register MAC address. */
1310 	claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1311 	if (config->vf && config->vf_nl_en)
1312 		mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1313 				      mlx5_ifindex(eth_dev),
1314 				      eth_dev->data->mac_addrs,
1315 				      MLX5_MAX_MAC_ADDRESSES);
1316 	priv->flows = 0;
1317 	priv->ctrl_flows = 0;
1318 	TAILQ_INIT(&priv->flow_meters);
1319 	TAILQ_INIT(&priv->flow_meter_profiles);
1320 	/* Hint libmlx5 to use PMD allocator for data plane resources */
1321 	mlx5_glue->dv_set_context_attr(sh->ctx,
1322 			MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1323 			(void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1324 				.alloc = &mlx5_alloc_verbs_buf,
1325 				.free = &mlx5_free_verbs_buf,
1326 				.data = priv,
1327 			}));
1328 	/* Bring Ethernet device up. */
1329 	DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1330 		eth_dev->data->port_id);
1331 	mlx5_set_link_up(eth_dev);
1332 	/*
1333 	 * Even though the interrupt handler is not installed yet,
1334 	 * interrupts will still trigger on the async_fd from
1335 	 * Verbs context returned by ibv_open_device().
1336 	 */
1337 	mlx5_link_update(eth_dev, 0);
1338 #ifdef HAVE_MLX5DV_DR_ESWITCH
1339 	if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1340 	      (switch_info->representor || switch_info->master)))
1341 		config->dv_esw_en = 0;
1342 #else
1343 	config->dv_esw_en = 0;
1344 #endif
1345 	/* Detect minimal data bytes to inline. */
1346 	mlx5_set_min_inline(spawn, config);
1347 	/* Store device configuration on private structure. */
1348 	priv->config = *config;
1349 	/* Create context for virtual machine VLAN workaround. */
1350 	priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1351 	if (config->dv_flow_en) {
1352 		err = mlx5_alloc_shared_dr(priv);
1353 		if (err)
1354 			goto error;
1355 		/*
1356 		 * RSS id is shared with meter flow id. Meter flow id can only
1357 		 * use the 24 MSB of the register.
1358 		 */
1359 		priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1360 				     MLX5_MTR_COLOR_BITS);
1361 		if (!priv->qrss_id_pool) {
1362 			DRV_LOG(ERR, "can't create flow id pool");
1363 			err = ENOMEM;
1364 			goto error;
1365 		}
1366 	}
1367 	if (config->devx && config->dv_flow_en && config->dest_tir) {
1368 		priv->obj_ops = devx_obj_ops;
1369 		priv->obj_ops.drop_action_create =
1370 						ibv_obj_ops.drop_action_create;
1371 		priv->obj_ops.drop_action_destroy =
1372 						ibv_obj_ops.drop_action_destroy;
1373 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1374 		priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1375 #else
1376 		if (config->dv_esw_en)
1377 			priv->obj_ops.txq_obj_modify =
1378 						ibv_obj_ops.txq_obj_modify;
1379 #endif
1380 		/* Use specific wrappers for Tx object. */
1381 		priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1382 		priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1383 
1384 	} else {
1385 		priv->obj_ops = ibv_obj_ops;
1386 	}
1387 	/* Supported Verbs flow priority number detection. */
1388 	err = mlx5_flow_discover_priorities(eth_dev);
1389 	if (err < 0) {
1390 		err = -err;
1391 		goto error;
1392 	}
1393 	priv->config.flow_prio = err;
1394 	if (!priv->config.dv_esw_en &&
1395 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1396 		DRV_LOG(WARNING, "metadata mode %u is not supported "
1397 				 "(no E-Switch)", priv->config.dv_xmeta_en);
1398 		priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1399 	}
1400 	mlx5_set_metadata_mask(eth_dev);
1401 	if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1402 	    !priv->sh->dv_regc0_mask) {
1403 		DRV_LOG(ERR, "metadata mode %u is not supported "
1404 			     "(no metadata reg_c[0] is available)",
1405 			     priv->config.dv_xmeta_en);
1406 			err = ENOTSUP;
1407 			goto error;
1408 	}
1409 	/*
1410 	 * Allocate the buffer for flow creating, just once.
1411 	 * The allocation must be done before any flow creating.
1412 	 */
1413 	mlx5_flow_alloc_intermediate(eth_dev);
1414 	/* Query availability of metadata reg_c's. */
1415 	err = mlx5_flow_discover_mreg_c(eth_dev);
1416 	if (err < 0) {
1417 		err = -err;
1418 		goto error;
1419 	}
1420 	if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1421 		DRV_LOG(DEBUG,
1422 			"port %u extensive metadata register is not supported",
1423 			eth_dev->data->port_id);
1424 		if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1425 			DRV_LOG(ERR, "metadata mode %u is not supported "
1426 				     "(no metadata registers available)",
1427 				     priv->config.dv_xmeta_en);
1428 			err = ENOTSUP;
1429 			goto error;
1430 		}
1431 	}
1432 	if (priv->config.dv_flow_en &&
1433 	    priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1434 	    mlx5_flow_ext_mreg_supported(eth_dev) &&
1435 	    priv->sh->dv_regc0_mask) {
1436 		priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1437 						      MLX5_FLOW_MREG_HTABLE_SZ);
1438 		if (!priv->mreg_cp_tbl) {
1439 			err = ENOMEM;
1440 			goto error;
1441 		}
1442 	}
1443 	return eth_dev;
1444 error:
1445 	if (priv) {
1446 		if (priv->mreg_cp_tbl)
1447 			mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1448 		if (priv->sh)
1449 			mlx5_os_free_shared_dr(priv);
1450 		if (priv->nl_socket_route >= 0)
1451 			close(priv->nl_socket_route);
1452 		if (priv->nl_socket_rdma >= 0)
1453 			close(priv->nl_socket_rdma);
1454 		if (priv->vmwa_context)
1455 			mlx5_vlan_vmwa_exit(priv->vmwa_context);
1456 		if (priv->qrss_id_pool)
1457 			mlx5_flow_id_pool_release(priv->qrss_id_pool);
1458 		if (own_domain_id)
1459 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1460 		mlx5_free(priv);
1461 		if (eth_dev != NULL)
1462 			eth_dev->data->dev_private = NULL;
1463 	}
1464 	if (eth_dev != NULL) {
1465 		/* mac_addrs must not be freed alone because part of
1466 		 * dev_private
1467 		 **/
1468 		eth_dev->data->mac_addrs = NULL;
1469 		rte_eth_dev_release_port(eth_dev);
1470 	}
1471 	if (sh)
1472 		mlx5_free_shared_dev_ctx(sh);
1473 	MLX5_ASSERT(err > 0);
1474 	rte_errno = err;
1475 	return NULL;
1476 }
1477 
1478 /**
1479  * Comparison callback to sort device data.
1480  *
1481  * This is meant to be used with qsort().
1482  *
1483  * @param a[in]
1484  *   Pointer to pointer to first data object.
1485  * @param b[in]
1486  *   Pointer to pointer to second data object.
1487  *
1488  * @return
1489  *   0 if both objects are equal, less than 0 if the first argument is less
1490  *   than the second, greater than 0 otherwise.
1491  */
1492 static int
1493 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1494 {
1495 	const struct mlx5_switch_info *si_a =
1496 		&((const struct mlx5_dev_spawn_data *)a)->info;
1497 	const struct mlx5_switch_info *si_b =
1498 		&((const struct mlx5_dev_spawn_data *)b)->info;
1499 	int ret;
1500 
1501 	/* Master device first. */
1502 	ret = si_b->master - si_a->master;
1503 	if (ret)
1504 		return ret;
1505 	/* Then representor devices. */
1506 	ret = si_b->representor - si_a->representor;
1507 	if (ret)
1508 		return ret;
1509 	/* Unidentified devices come last in no specific order. */
1510 	if (!si_a->representor)
1511 		return 0;
1512 	/* Order representors by name. */
1513 	return si_a->port_name - si_b->port_name;
1514 }
1515 
1516 /**
1517  * Match PCI information for possible slaves of bonding device.
1518  *
1519  * @param[in] ibv_dev
1520  *   Pointer to Infiniband device structure.
1521  * @param[in] pci_dev
1522  *   Pointer to PCI device structure to match PCI address.
1523  * @param[in] nl_rdma
1524  *   Netlink RDMA group socket handle.
1525  *
1526  * @return
1527  *   negative value if no bonding device found, otherwise
1528  *   positive index of slave PF in bonding.
1529  */
1530 static int
1531 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1532 			   const struct rte_pci_device *pci_dev,
1533 			   int nl_rdma)
1534 {
1535 	char ifname[IF_NAMESIZE + 1];
1536 	unsigned int ifindex;
1537 	unsigned int np, i;
1538 	FILE *file = NULL;
1539 	int pf = -1;
1540 
1541 	/*
1542 	 * Try to get master device name. If something goes
1543 	 * wrong suppose the lack of kernel support and no
1544 	 * bonding devices.
1545 	 */
1546 	if (nl_rdma < 0)
1547 		return -1;
1548 	if (!strstr(ibv_dev->name, "bond"))
1549 		return -1;
1550 	np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1551 	if (!np)
1552 		return -1;
1553 	/*
1554 	 * The Master device might not be on the predefined
1555 	 * port (not on port index 1, it is not garanted),
1556 	 * we have to scan all Infiniband device port and
1557 	 * find master.
1558 	 */
1559 	for (i = 1; i <= np; ++i) {
1560 		/* Check whether Infiniband port is populated. */
1561 		ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1562 		if (!ifindex)
1563 			continue;
1564 		if (!if_indextoname(ifindex, ifname))
1565 			continue;
1566 		/* Try to read bonding slave names from sysfs. */
1567 		MKSTR(slaves,
1568 		      "/sys/class/net/%s/master/bonding/slaves", ifname);
1569 		file = fopen(slaves, "r");
1570 		if (file)
1571 			break;
1572 	}
1573 	if (!file)
1574 		return -1;
1575 	/* Use safe format to check maximal buffer length. */
1576 	MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1577 	while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1578 		char tmp_str[IF_NAMESIZE + 32];
1579 		struct rte_pci_addr pci_addr;
1580 		struct mlx5_switch_info	info;
1581 
1582 		/* Process slave interface names in the loop. */
1583 		snprintf(tmp_str, sizeof(tmp_str),
1584 			 "/sys/class/net/%s", ifname);
1585 		if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1586 			DRV_LOG(WARNING, "can not get PCI address"
1587 					 " for netdev \"%s\"", ifname);
1588 			continue;
1589 		}
1590 		if (pci_dev->addr.domain != pci_addr.domain ||
1591 		    pci_dev->addr.bus != pci_addr.bus ||
1592 		    pci_dev->addr.devid != pci_addr.devid ||
1593 		    pci_dev->addr.function != pci_addr.function)
1594 			continue;
1595 		/* Slave interface PCI address match found. */
1596 		fclose(file);
1597 		snprintf(tmp_str, sizeof(tmp_str),
1598 			 "/sys/class/net/%s/phys_port_name", ifname);
1599 		file = fopen(tmp_str, "rb");
1600 		if (!file)
1601 			break;
1602 		info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1603 		if (fscanf(file, "%32s", tmp_str) == 1)
1604 			mlx5_translate_port_name(tmp_str, &info);
1605 		if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1606 		    info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1607 			pf = info.port_name;
1608 		break;
1609 	}
1610 	if (file)
1611 		fclose(file);
1612 	return pf;
1613 }
1614 
1615 /**
1616  * DPDK callback to register a PCI device.
1617  *
1618  * This function spawns Ethernet devices out of a given PCI device.
1619  *
1620  * @param[in] pci_drv
1621  *   PCI driver structure (mlx5_driver).
1622  * @param[in] pci_dev
1623  *   PCI device information.
1624  *
1625  * @return
1626  *   0 on success, a negative errno value otherwise and rte_errno is set.
1627  */
1628 int
1629 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1630 		  struct rte_pci_device *pci_dev)
1631 {
1632 	struct ibv_device **ibv_list;
1633 	/*
1634 	 * Number of found IB Devices matching with requested PCI BDF.
1635 	 * nd != 1 means there are multiple IB devices over the same
1636 	 * PCI device and we have representors and master.
1637 	 */
1638 	unsigned int nd = 0;
1639 	/*
1640 	 * Number of found IB device Ports. nd = 1 and np = 1..n means
1641 	 * we have the single multiport IB device, and there may be
1642 	 * representors attached to some of found ports.
1643 	 */
1644 	unsigned int np = 0;
1645 	/*
1646 	 * Number of DPDK ethernet devices to Spawn - either over
1647 	 * multiple IB devices or multiple ports of single IB device.
1648 	 * Actually this is the number of iterations to spawn.
1649 	 */
1650 	unsigned int ns = 0;
1651 	/*
1652 	 * Bonding device
1653 	 *   < 0 - no bonding device (single one)
1654 	 *  >= 0 - bonding device (value is slave PF index)
1655 	 */
1656 	int bd = -1;
1657 	struct mlx5_dev_spawn_data *list = NULL;
1658 	struct mlx5_dev_config dev_config;
1659 	unsigned int dev_config_vf;
1660 	int ret;
1661 
1662 	if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1663 		mlx5_pmd_socket_init();
1664 	ret = mlx5_init_once();
1665 	if (ret) {
1666 		DRV_LOG(ERR, "unable to init PMD global data: %s",
1667 			strerror(rte_errno));
1668 		return -rte_errno;
1669 	}
1670 	errno = 0;
1671 	ibv_list = mlx5_glue->get_device_list(&ret);
1672 	if (!ibv_list) {
1673 		rte_errno = errno ? errno : ENOSYS;
1674 		DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1675 		return -rte_errno;
1676 	}
1677 	/*
1678 	 * First scan the list of all Infiniband devices to find
1679 	 * matching ones, gathering into the list.
1680 	 */
1681 	struct ibv_device *ibv_match[ret + 1];
1682 	int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1683 	int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1684 	unsigned int i;
1685 
1686 	while (ret-- > 0) {
1687 		struct rte_pci_addr pci_addr;
1688 
1689 		DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1690 		bd = mlx5_device_bond_pci_match
1691 				(ibv_list[ret], pci_dev, nl_rdma);
1692 		if (bd >= 0) {
1693 			/*
1694 			 * Bonding device detected. Only one match is allowed,
1695 			 * the bonding is supported over multi-port IB device,
1696 			 * there should be no matches on representor PCI
1697 			 * functions or non VF LAG bonding devices with
1698 			 * specified address.
1699 			 */
1700 			if (nd) {
1701 				DRV_LOG(ERR,
1702 					"multiple PCI match on bonding device"
1703 					"\"%s\" found", ibv_list[ret]->name);
1704 				rte_errno = ENOENT;
1705 				ret = -rte_errno;
1706 				goto exit;
1707 			}
1708 			DRV_LOG(INFO, "PCI information matches for"
1709 				      " slave %d bonding device \"%s\"",
1710 				      bd, ibv_list[ret]->name);
1711 			ibv_match[nd++] = ibv_list[ret];
1712 			break;
1713 		}
1714 		if (mlx5_dev_to_pci_addr
1715 			(ibv_list[ret]->ibdev_path, &pci_addr))
1716 			continue;
1717 		if (pci_dev->addr.domain != pci_addr.domain ||
1718 		    pci_dev->addr.bus != pci_addr.bus ||
1719 		    pci_dev->addr.devid != pci_addr.devid ||
1720 		    pci_dev->addr.function != pci_addr.function)
1721 			continue;
1722 		DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1723 			ibv_list[ret]->name);
1724 		ibv_match[nd++] = ibv_list[ret];
1725 	}
1726 	ibv_match[nd] = NULL;
1727 	if (!nd) {
1728 		/* No device matches, just complain and bail out. */
1729 		DRV_LOG(WARNING,
1730 			"no Verbs device matches PCI device " PCI_PRI_FMT ","
1731 			" are kernel drivers loaded?",
1732 			pci_dev->addr.domain, pci_dev->addr.bus,
1733 			pci_dev->addr.devid, pci_dev->addr.function);
1734 		rte_errno = ENOENT;
1735 		ret = -rte_errno;
1736 		goto exit;
1737 	}
1738 	if (nd == 1) {
1739 		/*
1740 		 * Found single matching device may have multiple ports.
1741 		 * Each port may be representor, we have to check the port
1742 		 * number and check the representors existence.
1743 		 */
1744 		if (nl_rdma >= 0)
1745 			np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1746 		if (!np)
1747 			DRV_LOG(WARNING, "can not get IB device \"%s\""
1748 					 " ports number", ibv_match[0]->name);
1749 		if (bd >= 0 && !np) {
1750 			DRV_LOG(ERR, "can not get ports"
1751 				     " for bonding device");
1752 			rte_errno = ENOENT;
1753 			ret = -rte_errno;
1754 			goto exit;
1755 		}
1756 	}
1757 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1758 	if (bd >= 0) {
1759 		/*
1760 		 * This may happen if there is VF LAG kernel support and
1761 		 * application is compiled with older rdma_core library.
1762 		 */
1763 		DRV_LOG(ERR,
1764 			"No kernel/verbs support for VF LAG bonding found.");
1765 		rte_errno = ENOTSUP;
1766 		ret = -rte_errno;
1767 		goto exit;
1768 	}
1769 #endif
1770 	/*
1771 	 * Now we can determine the maximal
1772 	 * amount of devices to be spawned.
1773 	 */
1774 	list = mlx5_malloc(MLX5_MEM_ZERO,
1775 			   sizeof(struct mlx5_dev_spawn_data) *
1776 			   (np ? np : nd),
1777 			   RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1778 	if (!list) {
1779 		DRV_LOG(ERR, "spawn data array allocation failure");
1780 		rte_errno = ENOMEM;
1781 		ret = -rte_errno;
1782 		goto exit;
1783 	}
1784 	if (bd >= 0 || np > 1) {
1785 		/*
1786 		 * Single IB device with multiple ports found,
1787 		 * it may be E-Switch master device and representors.
1788 		 * We have to perform identification through the ports.
1789 		 */
1790 		MLX5_ASSERT(nl_rdma >= 0);
1791 		MLX5_ASSERT(ns == 0);
1792 		MLX5_ASSERT(nd == 1);
1793 		MLX5_ASSERT(np);
1794 		for (i = 1; i <= np; ++i) {
1795 			list[ns].max_port = np;
1796 			list[ns].phys_port = i;
1797 			list[ns].phys_dev = ibv_match[0];
1798 			list[ns].eth_dev = NULL;
1799 			list[ns].pci_dev = pci_dev;
1800 			list[ns].pf_bond = bd;
1801 			list[ns].ifindex = mlx5_nl_ifindex
1802 				(nl_rdma,
1803 				mlx5_os_get_dev_device_name
1804 						(list[ns].phys_dev), i);
1805 			if (!list[ns].ifindex) {
1806 				/*
1807 				 * No network interface index found for the
1808 				 * specified port, it means there is no
1809 				 * representor on this port. It's OK,
1810 				 * there can be disabled ports, for example
1811 				 * if sriov_numvfs < sriov_totalvfs.
1812 				 */
1813 				continue;
1814 			}
1815 			ret = -1;
1816 			if (nl_route >= 0)
1817 				ret = mlx5_nl_switch_info
1818 					       (nl_route,
1819 						list[ns].ifindex,
1820 						&list[ns].info);
1821 			if (ret || (!list[ns].info.representor &&
1822 				    !list[ns].info.master)) {
1823 				/*
1824 				 * We failed to recognize representors with
1825 				 * Netlink, let's try to perform the task
1826 				 * with sysfs.
1827 				 */
1828 				ret =  mlx5_sysfs_switch_info
1829 						(list[ns].ifindex,
1830 						 &list[ns].info);
1831 			}
1832 			if (!ret && bd >= 0) {
1833 				switch (list[ns].info.name_type) {
1834 				case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1835 					if (list[ns].info.port_name == bd)
1836 						ns++;
1837 					break;
1838 				case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1839 					/* Fallthrough */
1840 				case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1841 					if (list[ns].info.pf_num == bd)
1842 						ns++;
1843 					break;
1844 				default:
1845 					break;
1846 				}
1847 				continue;
1848 			}
1849 			if (!ret && (list[ns].info.representor ^
1850 				     list[ns].info.master))
1851 				ns++;
1852 		}
1853 		if (!ns) {
1854 			DRV_LOG(ERR,
1855 				"unable to recognize master/representors"
1856 				" on the IB device with multiple ports");
1857 			rte_errno = ENOENT;
1858 			ret = -rte_errno;
1859 			goto exit;
1860 		}
1861 	} else {
1862 		/*
1863 		 * The existence of several matching entries (nd > 1) means
1864 		 * port representors have been instantiated. No existing Verbs
1865 		 * call nor sysfs entries can tell them apart, this can only
1866 		 * be done through Netlink calls assuming kernel drivers are
1867 		 * recent enough to support them.
1868 		 *
1869 		 * In the event of identification failure through Netlink,
1870 		 * try again through sysfs, then:
1871 		 *
1872 		 * 1. A single IB device matches (nd == 1) with single
1873 		 *    port (np=0/1) and is not a representor, assume
1874 		 *    no switch support.
1875 		 *
1876 		 * 2. Otherwise no safe assumptions can be made;
1877 		 *    complain louder and bail out.
1878 		 */
1879 		for (i = 0; i != nd; ++i) {
1880 			memset(&list[ns].info, 0, sizeof(list[ns].info));
1881 			list[ns].max_port = 1;
1882 			list[ns].phys_port = 1;
1883 			list[ns].phys_dev = ibv_match[i];
1884 			list[ns].eth_dev = NULL;
1885 			list[ns].pci_dev = pci_dev;
1886 			list[ns].pf_bond = -1;
1887 			list[ns].ifindex = 0;
1888 			if (nl_rdma >= 0)
1889 				list[ns].ifindex = mlx5_nl_ifindex
1890 				(nl_rdma,
1891 				mlx5_os_get_dev_device_name
1892 						(list[ns].phys_dev), 1);
1893 			if (!list[ns].ifindex) {
1894 				char ifname[IF_NAMESIZE];
1895 
1896 				/*
1897 				 * Netlink failed, it may happen with old
1898 				 * ib_core kernel driver (before 4.16).
1899 				 * We can assume there is old driver because
1900 				 * here we are processing single ports IB
1901 				 * devices. Let's try sysfs to retrieve
1902 				 * the ifindex. The method works for
1903 				 * master device only.
1904 				 */
1905 				if (nd > 1) {
1906 					/*
1907 					 * Multiple devices found, assume
1908 					 * representors, can not distinguish
1909 					 * master/representor and retrieve
1910 					 * ifindex via sysfs.
1911 					 */
1912 					continue;
1913 				}
1914 				ret = mlx5_get_ifname_sysfs
1915 					(ibv_match[i]->ibdev_path, ifname);
1916 				if (!ret)
1917 					list[ns].ifindex =
1918 						if_nametoindex(ifname);
1919 				if (!list[ns].ifindex) {
1920 					/*
1921 					 * No network interface index found
1922 					 * for the specified device, it means
1923 					 * there it is neither representor
1924 					 * nor master.
1925 					 */
1926 					continue;
1927 				}
1928 			}
1929 			ret = -1;
1930 			if (nl_route >= 0)
1931 				ret = mlx5_nl_switch_info
1932 					       (nl_route,
1933 						list[ns].ifindex,
1934 						&list[ns].info);
1935 			if (ret || (!list[ns].info.representor &&
1936 				    !list[ns].info.master)) {
1937 				/*
1938 				 * We failed to recognize representors with
1939 				 * Netlink, let's try to perform the task
1940 				 * with sysfs.
1941 				 */
1942 				ret =  mlx5_sysfs_switch_info
1943 						(list[ns].ifindex,
1944 						 &list[ns].info);
1945 			}
1946 			if (!ret && (list[ns].info.representor ^
1947 				     list[ns].info.master)) {
1948 				ns++;
1949 			} else if ((nd == 1) &&
1950 				   !list[ns].info.representor &&
1951 				   !list[ns].info.master) {
1952 				/*
1953 				 * Single IB device with
1954 				 * one physical port and
1955 				 * attached network device.
1956 				 * May be SRIOV is not enabled
1957 				 * or there is no representors.
1958 				 */
1959 				DRV_LOG(INFO, "no E-Switch support detected");
1960 				ns++;
1961 				break;
1962 			}
1963 		}
1964 		if (!ns) {
1965 			DRV_LOG(ERR,
1966 				"unable to recognize master/representors"
1967 				" on the multiple IB devices");
1968 			rte_errno = ENOENT;
1969 			ret = -rte_errno;
1970 			goto exit;
1971 		}
1972 	}
1973 	MLX5_ASSERT(ns);
1974 	/*
1975 	 * Sort list to probe devices in natural order for users convenience
1976 	 * (i.e. master first, then representors from lowest to highest ID).
1977 	 */
1978 	qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1979 	/* Device specific configuration. */
1980 	switch (pci_dev->id.device_id) {
1981 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1982 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1983 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1984 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1985 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1986 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1987 	case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1988 		dev_config_vf = 1;
1989 		break;
1990 	default:
1991 		dev_config_vf = 0;
1992 		break;
1993 	}
1994 	for (i = 0; i != ns; ++i) {
1995 		uint32_t restore;
1996 
1997 		/* Default configuration. */
1998 		memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1999 		dev_config.vf = dev_config_vf;
2000 		dev_config.mps = MLX5_ARG_UNSET;
2001 		dev_config.dbnc = MLX5_ARG_UNSET;
2002 		dev_config.rx_vec_en = 1;
2003 		dev_config.txq_inline_max = MLX5_ARG_UNSET;
2004 		dev_config.txq_inline_min = MLX5_ARG_UNSET;
2005 		dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2006 		dev_config.txqs_inline = MLX5_ARG_UNSET;
2007 		dev_config.vf_nl_en = 1;
2008 		dev_config.mr_ext_memseg_en = 1;
2009 		dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2010 		dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2011 		dev_config.dv_esw_en = 1;
2012 		dev_config.dv_flow_en = 1;
2013 		dev_config.decap_en = 1;
2014 		dev_config.log_hp_size = MLX5_ARG_UNSET;
2015 		list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2016 						 &list[i],
2017 						 &dev_config);
2018 		if (!list[i].eth_dev) {
2019 			if (rte_errno != EBUSY && rte_errno != EEXIST)
2020 				break;
2021 			/* Device is disabled or already spawned. Ignore it. */
2022 			continue;
2023 		}
2024 		restore = list[i].eth_dev->data->dev_flags;
2025 		rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2026 		/* Restore non-PCI flags cleared by the above call. */
2027 		list[i].eth_dev->data->dev_flags |= restore;
2028 		rte_eth_dev_probing_finish(list[i].eth_dev);
2029 	}
2030 	if (i != ns) {
2031 		DRV_LOG(ERR,
2032 			"probe of PCI device " PCI_PRI_FMT " aborted after"
2033 			" encountering an error: %s",
2034 			pci_dev->addr.domain, pci_dev->addr.bus,
2035 			pci_dev->addr.devid, pci_dev->addr.function,
2036 			strerror(rte_errno));
2037 		ret = -rte_errno;
2038 		/* Roll back. */
2039 		while (i--) {
2040 			if (!list[i].eth_dev)
2041 				continue;
2042 			mlx5_dev_close(list[i].eth_dev);
2043 			/* mac_addrs must not be freed because in dev_private */
2044 			list[i].eth_dev->data->mac_addrs = NULL;
2045 			claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2046 		}
2047 		/* Restore original error. */
2048 		rte_errno = -ret;
2049 	} else {
2050 		ret = 0;
2051 	}
2052 exit:
2053 	/*
2054 	 * Do the routine cleanup:
2055 	 * - close opened Netlink sockets
2056 	 * - free allocated spawn data array
2057 	 * - free the Infiniband device list
2058 	 */
2059 	if (nl_rdma >= 0)
2060 		close(nl_rdma);
2061 	if (nl_route >= 0)
2062 		close(nl_route);
2063 	if (list)
2064 		mlx5_free(list);
2065 	MLX5_ASSERT(ibv_list);
2066 	mlx5_glue->free_device_list(ibv_list);
2067 	return ret;
2068 }
2069 
2070 static int
2071 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2072 {
2073 	char *env;
2074 	int value;
2075 
2076 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2077 	/* Get environment variable to store. */
2078 	env = getenv(MLX5_SHUT_UP_BF);
2079 	value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2080 	if (config->dbnc == MLX5_ARG_UNSET)
2081 		setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2082 	else
2083 		setenv(MLX5_SHUT_UP_BF,
2084 		       config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2085 	return value;
2086 }
2087 
2088 static void
2089 mlx5_restore_doorbell_mapping_env(int value)
2090 {
2091 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2092 	/* Restore the original environment variable state. */
2093 	if (value == MLX5_ARG_UNSET)
2094 		unsetenv(MLX5_SHUT_UP_BF);
2095 	else
2096 		setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2097 }
2098 
2099 /**
2100  * Extract pdn of PD object using DV API.
2101  *
2102  * @param[in] pd
2103  *   Pointer to the verbs PD object.
2104  * @param[out] pdn
2105  *   Pointer to the PD object number variable.
2106  *
2107  * @return
2108  *   0 on success, error value otherwise.
2109  */
2110 int
2111 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2112 {
2113 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2114 	struct mlx5dv_obj obj;
2115 	struct mlx5dv_pd pd_info;
2116 	int ret = 0;
2117 
2118 	obj.pd.in = pd;
2119 	obj.pd.out = &pd_info;
2120 	ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2121 	if (ret) {
2122 		DRV_LOG(DEBUG, "Fail to get PD object info");
2123 		return ret;
2124 	}
2125 	*pdn = pd_info.pdn;
2126 	return 0;
2127 #else
2128 	(void)pd;
2129 	(void)pdn;
2130 	return -ENOTSUP;
2131 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2132 }
2133 
2134 /**
2135  * Function API to open IB device.
2136  *
2137  * This function calls the Linux glue APIs to open a device.
2138  *
2139  * @param[in] spawn
2140  *   Pointer to the IB device attributes (name, port, etc).
2141  * @param[out] config
2142  *   Pointer to device configuration structure.
2143  * @param[out] sh
2144  *   Pointer to shared context structure.
2145  *
2146  * @return
2147  *   0 on success, a positive error value otherwise.
2148  */
2149 int
2150 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2151 		     const struct mlx5_dev_config *config,
2152 		     struct mlx5_dev_ctx_shared *sh)
2153 {
2154 	int dbmap_env;
2155 	int err = 0;
2156 
2157 	sh->numa_node = spawn->pci_dev->device.numa_node;
2158 	pthread_mutex_init(&sh->txpp.mutex, NULL);
2159 	/*
2160 	 * Configure environment variable "MLX5_BF_SHUT_UP"
2161 	 * before the device creation. The rdma_core library
2162 	 * checks the variable at device creation and
2163 	 * stores the result internally.
2164 	 */
2165 	dbmap_env = mlx5_config_doorbell_mapping_env(config);
2166 	/* Try to open IB device with DV first, then usual Verbs. */
2167 	errno = 0;
2168 	sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2169 	if (sh->ctx) {
2170 		sh->devx = 1;
2171 		DRV_LOG(DEBUG, "DevX is supported");
2172 		/* The device is created, no need for environment. */
2173 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2174 	} else {
2175 		/* The environment variable is still configured. */
2176 		sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2177 		err = errno ? errno : ENODEV;
2178 		/*
2179 		 * The environment variable is not needed anymore,
2180 		 * all device creation attempts are completed.
2181 		 */
2182 		mlx5_restore_doorbell_mapping_env(dbmap_env);
2183 		if (!sh->ctx)
2184 			return err;
2185 		DRV_LOG(DEBUG, "DevX is NOT supported");
2186 		err = 0;
2187 	}
2188 	return err;
2189 }
2190 
2191 /**
2192  * Install shared asynchronous device events handler.
2193  * This function is implemented to support event sharing
2194  * between multiple ports of single IB device.
2195  *
2196  * @param sh
2197  *   Pointer to mlx5_dev_ctx_shared object.
2198  */
2199 void
2200 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2201 {
2202 	int ret;
2203 	int flags;
2204 
2205 	sh->intr_handle.fd = -1;
2206 	flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2207 	ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2208 		    F_SETFL, flags | O_NONBLOCK);
2209 	if (ret) {
2210 		DRV_LOG(INFO, "failed to change file descriptor async event"
2211 			" queue");
2212 	} else {
2213 		sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2214 		sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2215 		if (rte_intr_callback_register(&sh->intr_handle,
2216 					mlx5_dev_interrupt_handler, sh)) {
2217 			DRV_LOG(INFO, "Fail to install the shared interrupt.");
2218 			sh->intr_handle.fd = -1;
2219 		}
2220 	}
2221 	if (sh->devx) {
2222 #ifdef HAVE_IBV_DEVX_ASYNC
2223 		sh->intr_handle_devx.fd = -1;
2224 		sh->devx_comp =
2225 			(void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2226 		struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2227 		if (!devx_comp) {
2228 			DRV_LOG(INFO, "failed to allocate devx_comp.");
2229 			return;
2230 		}
2231 		flags = fcntl(devx_comp->fd, F_GETFL);
2232 		ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2233 		if (ret) {
2234 			DRV_LOG(INFO, "failed to change file descriptor"
2235 				" devx comp");
2236 			return;
2237 		}
2238 		sh->intr_handle_devx.fd = devx_comp->fd;
2239 		sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2240 		if (rte_intr_callback_register(&sh->intr_handle_devx,
2241 					mlx5_dev_interrupt_handler_devx, sh)) {
2242 			DRV_LOG(INFO, "Fail to install the devx shared"
2243 				" interrupt.");
2244 			sh->intr_handle_devx.fd = -1;
2245 		}
2246 #endif /* HAVE_IBV_DEVX_ASYNC */
2247 	}
2248 }
2249 
2250 /**
2251  * Uninstall shared asynchronous device events handler.
2252  * This function is implemented to support event sharing
2253  * between multiple ports of single IB device.
2254  *
2255  * @param dev
2256  *   Pointer to mlx5_dev_ctx_shared object.
2257  */
2258 void
2259 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2260 {
2261 	if (sh->intr_handle.fd >= 0)
2262 		mlx5_intr_callback_unregister(&sh->intr_handle,
2263 					      mlx5_dev_interrupt_handler, sh);
2264 #ifdef HAVE_IBV_DEVX_ASYNC
2265 	if (sh->intr_handle_devx.fd >= 0)
2266 		rte_intr_callback_unregister(&sh->intr_handle_devx,
2267 				  mlx5_dev_interrupt_handler_devx, sh);
2268 	if (sh->devx_comp)
2269 		mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2270 #endif
2271 }
2272 
2273 /**
2274  * Read statistics by a named counter.
2275  *
2276  * @param[in] priv
2277  *   Pointer to the private device data structure.
2278  * @param[in] ctr_name
2279  *   Pointer to the name of the statistic counter to read
2280  * @param[out] stat
2281  *   Pointer to read statistic value.
2282  * @return
2283  *   0 on success and stat is valud, 1 if failed to read the value
2284  *   rte_errno is set.
2285  *
2286  */
2287 int
2288 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2289 		      uint64_t *stat)
2290 {
2291 	int fd;
2292 
2293 	if (priv->sh) {
2294 		MKSTR(path, "%s/ports/%d/hw_counters/%s",
2295 		      priv->sh->ibdev_path,
2296 		      priv->dev_port,
2297 		      ctr_name);
2298 		fd = open(path, O_RDONLY);
2299 		/*
2300 		 * in switchdev the file location is not per port
2301 		 * but rather in <ibdev_path>/hw_counters/<file_name>.
2302 		 */
2303 		if (fd == -1) {
2304 			MKSTR(path1, "%s/hw_counters/%s",
2305 			      priv->sh->ibdev_path,
2306 			      ctr_name);
2307 			fd = open(path1, O_RDONLY);
2308 		}
2309 		if (fd != -1) {
2310 			char buf[21] = {'\0'};
2311 			ssize_t n = read(fd, buf, sizeof(buf));
2312 
2313 			close(fd);
2314 			if (n != -1) {
2315 				*stat = strtoull(buf, NULL, 10);
2316 				return 0;
2317 			}
2318 		}
2319 	}
2320 	*stat = 0;
2321 	return 1;
2322 }
2323 
2324 /**
2325  * Set the reg_mr and dereg_mr call backs
2326  *
2327  * @param reg_mr_cb[out]
2328  *   Pointer to reg_mr func
2329  * @param dereg_mr_cb[out]
2330  *   Pointer to dereg_mr func
2331  *
2332  */
2333 void
2334 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2335 		      mlx5_dereg_mr_t *dereg_mr_cb)
2336 {
2337 	*reg_mr_cb = mlx5_verbs_ops.reg_mr;
2338 	*dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2339 }
2340 
2341 /**
2342  * Remove a MAC address from device
2343  *
2344  * @param dev
2345  *   Pointer to Ethernet device structure.
2346  * @param index
2347  *   MAC address index.
2348  */
2349 void
2350 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2351 {
2352 	struct mlx5_priv *priv = dev->data->dev_private;
2353 	const int vf = priv->config.vf;
2354 
2355 	if (vf)
2356 		mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2357 					mlx5_ifindex(dev), priv->mac_own,
2358 					&dev->data->mac_addrs[index], index);
2359 }
2360 
2361 /**
2362  * Adds a MAC address to the device
2363  *
2364  * @param dev
2365  *   Pointer to Ethernet device structure.
2366  * @param mac_addr
2367  *   MAC address to register.
2368  * @param index
2369  *   MAC address index.
2370  *
2371  * @return
2372  *   0 on success, a negative errno value otherwise
2373  */
2374 int
2375 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2376 		     uint32_t index)
2377 {
2378 	struct mlx5_priv *priv = dev->data->dev_private;
2379 	const int vf = priv->config.vf;
2380 	int ret = 0;
2381 
2382 	if (vf)
2383 		ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2384 					   mlx5_ifindex(dev), priv->mac_own,
2385 					   mac, index);
2386 	return ret;
2387 }
2388 
2389 /**
2390  * Modify a VF MAC address
2391  *
2392  * @param priv
2393  *   Pointer to device private data.
2394  * @param mac_addr
2395  *   MAC address to modify into.
2396  * @param iface_idx
2397  *   Net device interface index
2398  * @param vf_index
2399  *   VF index
2400  *
2401  * @return
2402  *   0 on success, a negative errno value otherwise
2403  */
2404 int
2405 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2406 			   unsigned int iface_idx,
2407 			   struct rte_ether_addr *mac_addr,
2408 			   int vf_index)
2409 {
2410 	return mlx5_nl_vf_mac_addr_modify
2411 		(priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2412 }
2413 
2414 /**
2415  * Set device promiscuous mode
2416  *
2417  * @param dev
2418  *   Pointer to Ethernet device structure.
2419  * @param enable
2420  *   0 - promiscuous is disabled, otherwise - enabled
2421  *
2422  * @return
2423  *   0 on success, a negative error value otherwise
2424  */
2425 int
2426 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2427 {
2428 	struct mlx5_priv *priv = dev->data->dev_private;
2429 
2430 	return mlx5_nl_promisc(priv->nl_socket_route,
2431 			       mlx5_ifindex(dev), !!enable);
2432 }
2433 
2434 /**
2435  * Set device promiscuous mode
2436  *
2437  * @param dev
2438  *   Pointer to Ethernet device structure.
2439  * @param enable
2440  *   0 - all multicase is disabled, otherwise - enabled
2441  *
2442  * @return
2443  *   0 on success, a negative error value otherwise
2444  */
2445 int
2446 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2447 {
2448 	struct mlx5_priv *priv = dev->data->dev_private;
2449 
2450 	return mlx5_nl_allmulti(priv->nl_socket_route,
2451 				mlx5_ifindex(dev), !!enable);
2452 }
2453 
2454 /**
2455  * Flush device MAC addresses
2456  *
2457  * @param dev
2458  *   Pointer to Ethernet device structure.
2459  *
2460  */
2461 void
2462 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2463 {
2464 	struct mlx5_priv *priv = dev->data->dev_private;
2465 
2466 	mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2467 			       dev->data->mac_addrs,
2468 			       MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2469 }
2470 
2471 const struct eth_dev_ops mlx5_os_dev_ops = {
2472 	.dev_configure = mlx5_dev_configure,
2473 	.dev_start = mlx5_dev_start,
2474 	.dev_stop = mlx5_dev_stop,
2475 	.dev_set_link_down = mlx5_set_link_down,
2476 	.dev_set_link_up = mlx5_set_link_up,
2477 	.dev_close = mlx5_dev_close,
2478 	.promiscuous_enable = mlx5_promiscuous_enable,
2479 	.promiscuous_disable = mlx5_promiscuous_disable,
2480 	.allmulticast_enable = mlx5_allmulticast_enable,
2481 	.allmulticast_disable = mlx5_allmulticast_disable,
2482 	.link_update = mlx5_link_update,
2483 	.stats_get = mlx5_stats_get,
2484 	.stats_reset = mlx5_stats_reset,
2485 	.xstats_get = mlx5_xstats_get,
2486 	.xstats_reset = mlx5_xstats_reset,
2487 	.xstats_get_names = mlx5_xstats_get_names,
2488 	.fw_version_get = mlx5_fw_version_get,
2489 	.dev_infos_get = mlx5_dev_infos_get,
2490 	.read_clock = mlx5_txpp_read_clock,
2491 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2492 	.vlan_filter_set = mlx5_vlan_filter_set,
2493 	.rx_queue_setup = mlx5_rx_queue_setup,
2494 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2495 	.tx_queue_setup = mlx5_tx_queue_setup,
2496 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2497 	.rx_queue_release = mlx5_rx_queue_release,
2498 	.tx_queue_release = mlx5_tx_queue_release,
2499 	.rx_queue_start = mlx5_rx_queue_start,
2500 	.rx_queue_stop = mlx5_rx_queue_stop,
2501 	.tx_queue_start = mlx5_tx_queue_start,
2502 	.tx_queue_stop = mlx5_tx_queue_stop,
2503 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2504 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2505 	.mac_addr_remove = mlx5_mac_addr_remove,
2506 	.mac_addr_add = mlx5_mac_addr_add,
2507 	.mac_addr_set = mlx5_mac_addr_set,
2508 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2509 	.mtu_set = mlx5_dev_set_mtu,
2510 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2511 	.vlan_offload_set = mlx5_vlan_offload_set,
2512 	.reta_update = mlx5_dev_rss_reta_update,
2513 	.reta_query = mlx5_dev_rss_reta_query,
2514 	.rss_hash_update = mlx5_rss_hash_update,
2515 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
2516 	.filter_ctrl = mlx5_dev_filter_ctrl,
2517 	.rxq_info_get = mlx5_rxq_info_get,
2518 	.txq_info_get = mlx5_txq_info_get,
2519 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2520 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2521 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2522 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2523 	.is_removed = mlx5_is_removed,
2524 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2525 	.get_module_info = mlx5_get_module_info,
2526 	.get_module_eeprom = mlx5_get_module_eeprom,
2527 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2528 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2529 };
2530 
2531 /* Available operations from secondary process. */
2532 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2533 	.stats_get = mlx5_stats_get,
2534 	.stats_reset = mlx5_stats_reset,
2535 	.xstats_get = mlx5_xstats_get,
2536 	.xstats_reset = mlx5_xstats_reset,
2537 	.xstats_get_names = mlx5_xstats_get_names,
2538 	.fw_version_get = mlx5_fw_version_get,
2539 	.dev_infos_get = mlx5_dev_infos_get,
2540 	.read_clock = mlx5_txpp_read_clock,
2541 	.rx_queue_start = mlx5_rx_queue_start,
2542 	.rx_queue_stop = mlx5_rx_queue_stop,
2543 	.tx_queue_start = mlx5_tx_queue_start,
2544 	.tx_queue_stop = mlx5_tx_queue_stop,
2545 	.rxq_info_get = mlx5_rxq_info_get,
2546 	.txq_info_get = mlx5_txq_info_get,
2547 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2548 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2549 	.get_module_info = mlx5_get_module_info,
2550 	.get_module_eeprom = mlx5_get_module_eeprom,
2551 };
2552 
2553 /* Available operations in flow isolated mode. */
2554 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2555 	.dev_configure = mlx5_dev_configure,
2556 	.dev_start = mlx5_dev_start,
2557 	.dev_stop = mlx5_dev_stop,
2558 	.dev_set_link_down = mlx5_set_link_down,
2559 	.dev_set_link_up = mlx5_set_link_up,
2560 	.dev_close = mlx5_dev_close,
2561 	.promiscuous_enable = mlx5_promiscuous_enable,
2562 	.promiscuous_disable = mlx5_promiscuous_disable,
2563 	.allmulticast_enable = mlx5_allmulticast_enable,
2564 	.allmulticast_disable = mlx5_allmulticast_disable,
2565 	.link_update = mlx5_link_update,
2566 	.stats_get = mlx5_stats_get,
2567 	.stats_reset = mlx5_stats_reset,
2568 	.xstats_get = mlx5_xstats_get,
2569 	.xstats_reset = mlx5_xstats_reset,
2570 	.xstats_get_names = mlx5_xstats_get_names,
2571 	.fw_version_get = mlx5_fw_version_get,
2572 	.dev_infos_get = mlx5_dev_infos_get,
2573 	.read_clock = mlx5_txpp_read_clock,
2574 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2575 	.vlan_filter_set = mlx5_vlan_filter_set,
2576 	.rx_queue_setup = mlx5_rx_queue_setup,
2577 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2578 	.tx_queue_setup = mlx5_tx_queue_setup,
2579 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2580 	.rx_queue_release = mlx5_rx_queue_release,
2581 	.tx_queue_release = mlx5_tx_queue_release,
2582 	.rx_queue_start = mlx5_rx_queue_start,
2583 	.rx_queue_stop = mlx5_rx_queue_stop,
2584 	.tx_queue_start = mlx5_tx_queue_start,
2585 	.tx_queue_stop = mlx5_tx_queue_stop,
2586 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2587 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2588 	.mac_addr_remove = mlx5_mac_addr_remove,
2589 	.mac_addr_add = mlx5_mac_addr_add,
2590 	.mac_addr_set = mlx5_mac_addr_set,
2591 	.set_mc_addr_list = mlx5_set_mc_addr_list,
2592 	.mtu_set = mlx5_dev_set_mtu,
2593 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2594 	.vlan_offload_set = mlx5_vlan_offload_set,
2595 	.filter_ctrl = mlx5_dev_filter_ctrl,
2596 	.rxq_info_get = mlx5_rxq_info_get,
2597 	.txq_info_get = mlx5_txq_info_get,
2598 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
2599 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
2600 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
2601 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
2602 	.is_removed = mlx5_is_removed,
2603 	.get_module_info = mlx5_get_module_info,
2604 	.get_module_eeprom = mlx5_get_module_eeprom,
2605 	.hairpin_cap_get = mlx5_hairpin_cap_get,
2606 	.mtr_ops_get = mlx5_flow_meter_ops_get,
2607 };
2608